1 | // Licensed to the .NET Foundation under one or more agreements. |
2 | // The .NET Foundation licenses this file to you under the MIT license. |
3 | // See the LICENSE file in the project root for more information. |
4 | |
5 | /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
6 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
7 | XX XX |
8 | XX Compiler XX |
9 | XX XX |
10 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
11 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
12 | */ |
13 | #include "jitpch.h" |
14 | #ifdef _MSC_VER |
15 | #pragma hdrstop |
16 | #endif // _MSC_VER |
17 | #include "hostallocator.h" |
18 | #include "emit.h" |
19 | #include "ssabuilder.h" |
20 | #include "valuenum.h" |
21 | #include "rangecheck.h" |
22 | #include "lower.h" |
23 | #include "stacklevelsetter.h" |
24 | #include "jittelemetry.h" |
25 | |
26 | #if defined(DEBUG) |
27 | // Column settings for COMPlus_JitDumpIR. We could(should) make these programmable. |
28 | #define COLUMN_OPCODE 30 |
29 | #define COLUMN_OPERANDS (COLUMN_OPCODE + 25) |
30 | #define COLUMN_KINDS 110 |
31 | #define COLUMN_FLAGS (COLUMN_KINDS + 32) |
32 | #endif |
33 | |
34 | #if defined(DEBUG) |
35 | unsigned Compiler::jitTotalMethodCompiled = 0; |
36 | #endif // defined(DEBUG) |
37 | |
38 | #if defined(DEBUG) |
39 | LONG Compiler::jitNestingLevel = 0; |
40 | #endif // defined(DEBUG) |
41 | |
42 | #ifdef ALT_JIT |
43 | // static |
44 | bool Compiler::s_pAltJitExcludeAssembliesListInitialized = false; |
45 | AssemblyNamesList2* Compiler::s_pAltJitExcludeAssembliesList = nullptr; |
46 | #endif // ALT_JIT |
47 | |
48 | #ifdef DEBUG |
49 | // static |
50 | bool Compiler::s_pJitDisasmIncludeAssembliesListInitialized = false; |
51 | AssemblyNamesList2* Compiler::s_pJitDisasmIncludeAssembliesList = nullptr; |
52 | #endif // DEBUG |
53 | |
54 | /***************************************************************************** |
55 | * |
56 | * Little helpers to grab the current cycle counter value; this is done |
57 | * differently based on target architecture, host toolchain, etc. The |
58 | * main thing is to keep the overhead absolutely minimal; in fact, on |
59 | * x86/x64 we use RDTSC even though it's not thread-safe; GetThreadCycles |
60 | * (which is monotonous) is just too expensive. |
61 | */ |
62 | #ifdef FEATURE_JIT_METHOD_PERF |
63 | |
64 | #if defined(_HOST_X86_) || defined(_HOST_AMD64_) |
65 | |
66 | #if defined(_MSC_VER) |
67 | |
68 | #include <intrin.h> |
69 | inline bool _our_GetThreadCycles(unsigned __int64* cycleOut) |
70 | { |
71 | *cycleOut = __rdtsc(); |
72 | return true; |
73 | } |
74 | |
75 | #elif defined(__clang__) |
76 | |
77 | inline bool _our_GetThreadCycles(unsigned __int64* cycleOut) |
78 | { |
79 | uint32_t hi, lo; |
80 | __asm__ __volatile__("rdtsc" : "=a" (lo), "=d" (hi)); |
81 | *cycleOut = (static_cast<unsigned __int64>(hi) << 32) | static_cast<unsigned __int64>(lo); |
82 | return true; |
83 | } |
84 | |
85 | #else // neither _MSC_VER nor __clang__ |
86 | |
87 | // The following *might* work - might as well try. |
88 | #define _our_GetThreadCycles(cp) GetThreadCycles(cp) |
89 | |
90 | #endif |
91 | |
92 | #elif defined(_HOST_ARM_) || defined(_HOST_ARM64_) |
93 | |
94 | // If this doesn't work please see ../gc/gc.cpp for additional ARM |
95 | // info (and possible solutions). |
96 | #define _our_GetThreadCycles(cp) GetThreadCycles(cp) |
97 | |
98 | #else // not x86/x64 and not ARM |
99 | |
100 | // Don't know what this target is, but let's give it a try; if |
101 | // someone really wants to make this work, please add the right |
102 | // code here. |
103 | #define _our_GetThreadCycles(cp) GetThreadCycles(cp) |
104 | |
105 | #endif // which host OS |
106 | |
107 | #endif // FEATURE_JIT_METHOD_PERF |
108 | /*****************************************************************************/ |
109 | inline unsigned getCurTime() |
110 | { |
111 | SYSTEMTIME tim; |
112 | |
113 | GetSystemTime(&tim); |
114 | |
115 | return (((tim.wHour * 60) + tim.wMinute) * 60 + tim.wSecond) * 1000 + tim.wMilliseconds; |
116 | } |
117 | |
118 | /*****************************************************************************/ |
119 | #ifdef DEBUG |
120 | /*****************************************************************************/ |
121 | |
122 | static FILE* jitSrcFilePtr; |
123 | |
124 | static unsigned jitCurSrcLine; |
125 | |
126 | void Compiler::JitLogEE(unsigned level, const char* fmt, ...) |
127 | { |
128 | va_list args; |
129 | |
130 | if (verbose) |
131 | { |
132 | va_start(args, fmt); |
133 | vflogf(jitstdout, fmt, args); |
134 | va_end(args); |
135 | } |
136 | |
137 | va_start(args, fmt); |
138 | vlogf(level, fmt, args); |
139 | va_end(args); |
140 | } |
141 | |
142 | void Compiler::compDspSrcLinesByLineNum(unsigned line, bool seek) |
143 | { |
144 | if (!jitSrcFilePtr) |
145 | { |
146 | return; |
147 | } |
148 | |
149 | if (jitCurSrcLine == line) |
150 | { |
151 | return; |
152 | } |
153 | |
154 | if (jitCurSrcLine > line) |
155 | { |
156 | if (!seek) |
157 | { |
158 | return; |
159 | } |
160 | |
161 | if (fseek(jitSrcFilePtr, 0, SEEK_SET) != 0) |
162 | { |
163 | printf("Compiler::compDspSrcLinesByLineNum: fseek returned an error.\n" ); |
164 | } |
165 | jitCurSrcLine = 0; |
166 | } |
167 | |
168 | if (!seek) |
169 | { |
170 | printf(";\n" ); |
171 | } |
172 | |
173 | do |
174 | { |
175 | char temp[128]; |
176 | size_t llen; |
177 | |
178 | if (!fgets(temp, sizeof(temp), jitSrcFilePtr)) |
179 | { |
180 | return; |
181 | } |
182 | |
183 | if (seek) |
184 | { |
185 | continue; |
186 | } |
187 | |
188 | llen = strlen(temp); |
189 | if (llen && temp[llen - 1] == '\n') |
190 | { |
191 | temp[llen - 1] = 0; |
192 | } |
193 | |
194 | printf("; %s\n" , temp); |
195 | } while (++jitCurSrcLine < line); |
196 | |
197 | if (!seek) |
198 | { |
199 | printf(";\n" ); |
200 | } |
201 | } |
202 | |
203 | /*****************************************************************************/ |
204 | |
205 | void Compiler::compDspSrcLinesByNativeIP(UNATIVE_OFFSET curIP) |
206 | { |
207 | static IPmappingDsc* nextMappingDsc; |
208 | static unsigned lastLine; |
209 | |
210 | if (!opts.dspLines) |
211 | { |
212 | return; |
213 | } |
214 | |
215 | if (curIP == 0) |
216 | { |
217 | if (genIPmappingList) |
218 | { |
219 | nextMappingDsc = genIPmappingList; |
220 | lastLine = jitGetILoffs(nextMappingDsc->ipmdILoffsx); |
221 | |
222 | unsigned firstLine = jitGetILoffs(nextMappingDsc->ipmdILoffsx); |
223 | |
224 | unsigned earlierLine = (firstLine < 5) ? 0 : firstLine - 5; |
225 | |
226 | compDspSrcLinesByLineNum(earlierLine, true); // display previous 5 lines |
227 | compDspSrcLinesByLineNum(firstLine, false); |
228 | } |
229 | else |
230 | { |
231 | nextMappingDsc = nullptr; |
232 | } |
233 | |
234 | return; |
235 | } |
236 | |
237 | if (nextMappingDsc) |
238 | { |
239 | UNATIVE_OFFSET offset = nextMappingDsc->ipmdNativeLoc.CodeOffset(genEmitter); |
240 | |
241 | if (offset <= curIP) |
242 | { |
243 | IL_OFFSET nextOffs = jitGetILoffs(nextMappingDsc->ipmdILoffsx); |
244 | |
245 | if (lastLine < nextOffs) |
246 | { |
247 | compDspSrcLinesByLineNum(nextOffs); |
248 | } |
249 | else |
250 | { |
251 | // This offset corresponds to a previous line. Rewind to that line |
252 | |
253 | compDspSrcLinesByLineNum(nextOffs - 2, true); |
254 | compDspSrcLinesByLineNum(nextOffs); |
255 | } |
256 | |
257 | lastLine = nextOffs; |
258 | nextMappingDsc = nextMappingDsc->ipmdNext; |
259 | } |
260 | } |
261 | } |
262 | |
263 | /*****************************************************************************/ |
264 | #endif // DEBUG |
265 | |
266 | /*****************************************************************************/ |
267 | #if defined(DEBUG) || MEASURE_NODE_SIZE || MEASURE_BLOCK_SIZE || DISPLAY_SIZES || CALL_ARG_STATS |
268 | |
269 | static unsigned genMethodCnt; // total number of methods JIT'ted |
270 | unsigned genMethodICnt; // number of interruptible methods |
271 | unsigned genMethodNCnt; // number of non-interruptible methods |
272 | static unsigned = 0; |
273 | |
274 | #endif |
275 | |
276 | /*****************************************************************************/ |
277 | #if MEASURE_NODE_SIZE |
278 | NodeSizeStats genNodeSizeStats; |
279 | NodeSizeStats genNodeSizeStatsPerFunc; |
280 | |
281 | unsigned genTreeNcntHistBuckets[] = {10, 20, 30, 40, 50, 100, 200, 300, 400, 500, 1000, 5000, 10000, 0}; |
282 | Histogram genTreeNcntHist(genTreeNcntHistBuckets); |
283 | |
284 | unsigned genTreeNsizHistBuckets[] = {1000, 5000, 10000, 50000, 100000, 500000, 1000000, 0}; |
285 | Histogram genTreeNsizHist(genTreeNsizHistBuckets); |
286 | #endif // MEASURE_NODE_SIZE |
287 | |
288 | /*****************************************************************************/ |
289 | #if MEASURE_MEM_ALLOC |
290 | |
291 | unsigned memAllocHistBuckets[] = {64, 128, 192, 256, 512, 1024, 4096, 8192, 0}; |
292 | Histogram memAllocHist(memAllocHistBuckets); |
293 | unsigned memUsedHistBuckets[] = {16, 32, 64, 128, 192, 256, 512, 1024, 4096, 8192, 0}; |
294 | Histogram memUsedHist(memUsedHistBuckets); |
295 | |
296 | #endif // MEASURE_MEM_ALLOC |
297 | |
298 | /***************************************************************************** |
299 | * |
300 | * Variables to keep track of total code amounts. |
301 | */ |
302 | |
303 | #if DISPLAY_SIZES |
304 | |
305 | size_t grossVMsize; // Total IL code size |
306 | size_t grossNCsize; // Native code + data size |
307 | size_t totalNCsize; // Native code + data + GC info size (TODO-Cleanup: GC info size only accurate for JIT32_GCENCODER) |
308 | size_t gcHeaderISize; // GC header size: interruptible methods |
309 | size_t gcPtrMapISize; // GC pointer map size: interruptible methods |
310 | size_t gcHeaderNSize; // GC header size: non-interruptible methods |
311 | size_t gcPtrMapNSize; // GC pointer map size: non-interruptible methods |
312 | |
313 | #endif // DISPLAY_SIZES |
314 | |
315 | /***************************************************************************** |
316 | * |
317 | * Variables to keep track of argument counts. |
318 | */ |
319 | |
320 | #if CALL_ARG_STATS |
321 | |
322 | unsigned argTotalCalls; |
323 | unsigned argHelperCalls; |
324 | unsigned argStaticCalls; |
325 | unsigned argNonVirtualCalls; |
326 | unsigned argVirtualCalls; |
327 | |
328 | unsigned argTotalArgs; // total number of args for all calls (including objectPtr) |
329 | unsigned argTotalDWordArgs; |
330 | unsigned argTotalLongArgs; |
331 | unsigned argTotalFloatArgs; |
332 | unsigned argTotalDoubleArgs; |
333 | |
334 | unsigned argTotalRegArgs; |
335 | unsigned argTotalTemps; |
336 | unsigned argTotalLclVar; |
337 | unsigned argTotalDeferred; |
338 | unsigned argTotalConst; |
339 | |
340 | unsigned argTotalObjPtr; |
341 | unsigned argTotalGTF_ASGinArgs; |
342 | |
343 | unsigned argMaxTempsPerMethod; |
344 | |
345 | unsigned argCntBuckets[] = {0, 1, 2, 3, 4, 5, 6, 10, 0}; |
346 | Histogram argCntTable(argCntBuckets); |
347 | |
348 | unsigned argDWordCntBuckets[] = {0, 1, 2, 3, 4, 5, 6, 10, 0}; |
349 | Histogram argDWordCntTable(argDWordCntBuckets); |
350 | |
351 | unsigned argDWordLngCntBuckets[] = {0, 1, 2, 3, 4, 5, 6, 10, 0}; |
352 | Histogram argDWordLngCntTable(argDWordLngCntBuckets); |
353 | |
354 | unsigned argTempsCntBuckets[] = {0, 1, 2, 3, 4, 5, 6, 10, 0}; |
355 | Histogram argTempsCntTable(argTempsCntBuckets); |
356 | |
357 | #endif // CALL_ARG_STATS |
358 | |
359 | /***************************************************************************** |
360 | * |
361 | * Variables to keep track of basic block counts. |
362 | */ |
363 | |
364 | #if COUNT_BASIC_BLOCKS |
365 | |
366 | // -------------------------------------------------- |
367 | // Basic block count frequency table: |
368 | // -------------------------------------------------- |
369 | // <= 1 ===> 26872 count ( 56% of total) |
370 | // 2 .. 2 ===> 669 count ( 58% of total) |
371 | // 3 .. 3 ===> 4687 count ( 68% of total) |
372 | // 4 .. 5 ===> 5101 count ( 78% of total) |
373 | // 6 .. 10 ===> 5575 count ( 90% of total) |
374 | // 11 .. 20 ===> 3028 count ( 97% of total) |
375 | // 21 .. 50 ===> 1108 count ( 99% of total) |
376 | // 51 .. 100 ===> 182 count ( 99% of total) |
377 | // 101 .. 1000 ===> 34 count (100% of total) |
378 | // 1001 .. 10000 ===> 0 count (100% of total) |
379 | // -------------------------------------------------- |
380 | |
381 | unsigned bbCntBuckets[] = {1, 2, 3, 5, 10, 20, 50, 100, 1000, 10000, 0}; |
382 | Histogram bbCntTable(bbCntBuckets); |
383 | |
384 | /* Histogram for the IL opcode size of methods with a single basic block */ |
385 | |
386 | unsigned bbSizeBuckets[] = {1, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 0}; |
387 | Histogram bbOneBBSizeTable(bbSizeBuckets); |
388 | |
389 | #endif // COUNT_BASIC_BLOCKS |
390 | |
391 | /***************************************************************************** |
392 | * |
393 | * Used by optFindNaturalLoops to gather statistical information such as |
394 | * - total number of natural loops |
395 | * - number of loops with 1, 2, ... exit conditions |
396 | * - number of loops that have an iterator (for like) |
397 | * - number of loops that have a constant iterator |
398 | */ |
399 | |
400 | #if COUNT_LOOPS |
401 | |
402 | unsigned totalLoopMethods; // counts the total number of methods that have natural loops |
403 | unsigned maxLoopsPerMethod; // counts the maximum number of loops a method has |
404 | unsigned totalLoopOverflows; // # of methods that identified more loops than we can represent |
405 | unsigned totalLoopCount; // counts the total number of natural loops |
406 | unsigned totalUnnatLoopCount; // counts the total number of (not-necessarily natural) loops |
407 | unsigned totalUnnatLoopOverflows; // # of methods that identified more unnatural loops than we can represent |
408 | unsigned iterLoopCount; // counts the # of loops with an iterator (for like) |
409 | unsigned simpleTestLoopCount; // counts the # of loops with an iterator and a simple loop condition (iter < const) |
410 | unsigned constIterLoopCount; // counts the # of loops with a constant iterator (for like) |
411 | bool hasMethodLoops; // flag to keep track if we already counted a method as having loops |
412 | unsigned loopsThisMethod; // counts the number of loops in the current method |
413 | bool loopOverflowThisMethod; // True if we exceeded the max # of loops in the method. |
414 | |
415 | /* Histogram for number of loops in a method */ |
416 | |
417 | unsigned loopCountBuckets[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 0}; |
418 | Histogram loopCountTable(loopCountBuckets); |
419 | |
420 | /* Histogram for number of loop exits */ |
421 | |
422 | unsigned loopExitCountBuckets[] = {0, 1, 2, 3, 4, 5, 6, 0}; |
423 | Histogram loopExitCountTable(loopExitCountBuckets); |
424 | |
425 | #endif // COUNT_LOOPS |
426 | |
427 | //------------------------------------------------------------------------ |
428 | // getJitGCType: Given the VM's CorInfoGCType convert it to the JIT's var_types |
429 | // |
430 | // Arguments: |
431 | // gcType - an enum value that originally came from an element |
432 | // of the BYTE[] returned from getClassGClayout() |
433 | // |
434 | // Return Value: |
435 | // The corresponsing enum value from the JIT's var_types |
436 | // |
437 | // Notes: |
438 | // The gcLayout of each field of a struct is returned from getClassGClayout() |
439 | // as a BYTE[] but each BYTE element is actually a CorInfoGCType value |
440 | // Note when we 'know' that there is only one element in theis array |
441 | // the JIT will often pass the address of a single BYTE, instead of a BYTE[] |
442 | // |
443 | |
444 | var_types Compiler::getJitGCType(BYTE gcType) |
445 | { |
446 | var_types result = TYP_UNKNOWN; |
447 | CorInfoGCType corInfoType = (CorInfoGCType)gcType; |
448 | |
449 | if (corInfoType == TYPE_GC_NONE) |
450 | { |
451 | result = TYP_I_IMPL; |
452 | } |
453 | else if (corInfoType == TYPE_GC_REF) |
454 | { |
455 | result = TYP_REF; |
456 | } |
457 | else if (corInfoType == TYPE_GC_BYREF) |
458 | { |
459 | result = TYP_BYREF; |
460 | } |
461 | else |
462 | { |
463 | noway_assert(!"Bad value of 'gcType'" ); |
464 | } |
465 | return result; |
466 | } |
467 | |
468 | #if FEATURE_MULTIREG_ARGS |
469 | //--------------------------------------------------------------------------- |
470 | // getStructGcPtrsFromOp: Given a GenTree node of TYP_STRUCT that represents |
471 | // a pass by value argument, return the gcPtr layout |
472 | // for the pointers sized fields |
473 | // Arguments: |
474 | // op - the operand of TYP_STRUCT that is passed by value |
475 | // gcPtrsOut - an array of BYTES that are written by this method |
476 | // they will contain the VM's CorInfoGCType values |
477 | // for each pointer sized field |
478 | // Return Value: |
479 | // Two [or more] values are written into the gcPtrs array |
480 | // |
481 | // Note that for ARM64 there will always be exactly two pointer sized fields |
482 | |
483 | void Compiler::getStructGcPtrsFromOp(GenTree* op, BYTE* gcPtrsOut) |
484 | { |
485 | assert(op->TypeGet() == TYP_STRUCT); |
486 | |
487 | #ifdef _TARGET_ARM64_ |
488 | if (op->OperGet() == GT_OBJ) |
489 | { |
490 | CORINFO_CLASS_HANDLE objClass = op->gtObj.gtClass; |
491 | |
492 | int structSize = info.compCompHnd->getClassSize(objClass); |
493 | assert(structSize <= 2 * TARGET_POINTER_SIZE); |
494 | |
495 | BYTE gcPtrsTmp[2] = {TYPE_GC_NONE, TYPE_GC_NONE}; |
496 | |
497 | info.compCompHnd->getClassGClayout(objClass, &gcPtrsTmp[0]); |
498 | |
499 | gcPtrsOut[0] = gcPtrsTmp[0]; |
500 | gcPtrsOut[1] = gcPtrsTmp[1]; |
501 | } |
502 | else if (op->OperGet() == GT_LCL_VAR) |
503 | { |
504 | GenTreeLclVarCommon* varNode = op->AsLclVarCommon(); |
505 | unsigned varNum = varNode->gtLclNum; |
506 | assert(varNum < lvaCount); |
507 | LclVarDsc* varDsc = &lvaTable[varNum]; |
508 | |
509 | // At this point any TYP_STRUCT LclVar must be a 16-byte pass by value argument |
510 | assert(varDsc->lvSize() == 2 * TARGET_POINTER_SIZE); |
511 | |
512 | gcPtrsOut[0] = varDsc->lvGcLayout[0]; |
513 | gcPtrsOut[1] = varDsc->lvGcLayout[1]; |
514 | } |
515 | else |
516 | #endif |
517 | { |
518 | noway_assert(!"Unsupported Oper for getStructGcPtrsFromOp" ); |
519 | } |
520 | } |
521 | #endif // FEATURE_MULTIREG_ARGS |
522 | |
523 | #ifdef ARM_SOFTFP |
524 | //--------------------------------------------------------------------------- |
525 | // IsSingleFloat32Struct: |
526 | // Check if the given struct type contains only one float32 value type |
527 | // |
528 | // Arguments: |
529 | // clsHnd - the handle for the struct type |
530 | // |
531 | // Return Value: |
532 | // true if the given struct type contains only one float32 value type, |
533 | // false otherwise. |
534 | // |
535 | |
536 | bool Compiler::isSingleFloat32Struct(CORINFO_CLASS_HANDLE clsHnd) |
537 | { |
538 | for (;;) |
539 | { |
540 | // all of class chain must be of value type and must have only one field |
541 | if (!info.compCompHnd->isValueClass(clsHnd) || info.compCompHnd->getClassNumInstanceFields(clsHnd) != 1) |
542 | { |
543 | return false; |
544 | } |
545 | |
546 | CORINFO_CLASS_HANDLE* pClsHnd = &clsHnd; |
547 | CORINFO_FIELD_HANDLE fldHnd = info.compCompHnd->getFieldInClass(clsHnd, 0); |
548 | CorInfoType fieldType = info.compCompHnd->getFieldType(fldHnd, pClsHnd); |
549 | |
550 | switch (fieldType) |
551 | { |
552 | case CORINFO_TYPE_VALUECLASS: |
553 | clsHnd = *pClsHnd; |
554 | break; |
555 | |
556 | case CORINFO_TYPE_FLOAT: |
557 | return true; |
558 | |
559 | default: |
560 | return false; |
561 | } |
562 | } |
563 | } |
564 | #endif // ARM_SOFTFP |
565 | |
566 | //----------------------------------------------------------------------------- |
567 | // getPrimitiveTypeForStruct: |
568 | // Get the "primitive" type that is is used for a struct |
569 | // of size 'structSize'. |
570 | // We examine 'clsHnd' to check the GC layout of the struct and |
571 | // return TYP_REF for structs that simply wrap an object. |
572 | // If the struct is a one element HFA, we will return the |
573 | // proper floating point type. |
574 | // |
575 | // Arguments: |
576 | // structSize - the size of the struct type, cannot be zero |
577 | // clsHnd - the handle for the struct type, used when may have |
578 | // an HFA or if we need the GC layout for an object ref. |
579 | // |
580 | // Return Value: |
581 | // The primitive type (i.e. byte, short, int, long, ref, float, double) |
582 | // used to pass or return structs of this size. |
583 | // If we shouldn't use a "primitive" type then TYP_UNKNOWN is returned. |
584 | // Notes: |
585 | // For 32-bit targets (X86/ARM32) the 64-bit TYP_LONG type is not |
586 | // considered a primitive type by this method. |
587 | // So a struct that wraps a 'long' is passed and returned in the |
588 | // same way as any other 8-byte struct |
589 | // For ARM32 if we have an HFA struct that wraps a 64-bit double |
590 | // we will return TYP_DOUBLE. |
591 | // |
592 | var_types Compiler::getPrimitiveTypeForStruct(unsigned structSize, CORINFO_CLASS_HANDLE clsHnd, bool isVarArg) |
593 | { |
594 | assert(structSize != 0); |
595 | |
596 | var_types useType; |
597 | |
598 | switch (structSize) |
599 | { |
600 | case 1: |
601 | useType = TYP_BYTE; |
602 | break; |
603 | |
604 | case 2: |
605 | useType = TYP_SHORT; |
606 | break; |
607 | |
608 | #if !defined(_TARGET_XARCH_) || defined(UNIX_AMD64_ABI) |
609 | case 3: |
610 | useType = TYP_INT; |
611 | break; |
612 | |
613 | #endif // !_TARGET_XARCH_ || UNIX_AMD64_ABI |
614 | |
615 | #ifdef _TARGET_64BIT_ |
616 | case 4: |
617 | if (IsHfa(clsHnd)) |
618 | { |
619 | // A structSize of 4 with IsHfa, it must be an HFA of one float |
620 | useType = TYP_FLOAT; |
621 | } |
622 | else |
623 | { |
624 | useType = TYP_INT; |
625 | } |
626 | break; |
627 | |
628 | #if !defined(_TARGET_XARCH_) || defined(UNIX_AMD64_ABI) |
629 | case 5: |
630 | case 6: |
631 | case 7: |
632 | useType = TYP_I_IMPL; |
633 | break; |
634 | |
635 | #endif // !_TARGET_XARCH_ || UNIX_AMD64_ABI |
636 | #endif // _TARGET_64BIT_ |
637 | |
638 | case TARGET_POINTER_SIZE: |
639 | #ifdef ARM_SOFTFP |
640 | // For ARM_SOFTFP, HFA is unsupported so we need to check in another way |
641 | // This matters only for size-4 struct cause bigger structs would be processed with RetBuf |
642 | if (isSingleFloat32Struct(clsHnd)) |
643 | #else // !ARM_SOFTFP |
644 | if (IsHfa(clsHnd) |
645 | #if defined(_TARGET_WINDOWS_) && defined(_TARGET_ARM64_) |
646 | // Arm64 Windows VarArg methods arguments will not |
647 | // classify HFA types, they will need to be treated |
648 | // as if they are not HFA types. |
649 | && !isVarArg |
650 | #endif // defined(_TARGET_WINDOWS_) && defined(_TARGET_ARM64_) |
651 | ) |
652 | #endif // ARM_SOFTFP |
653 | { |
654 | #ifdef _TARGET_64BIT_ |
655 | var_types hfaType = GetHfaType(clsHnd); |
656 | |
657 | // A structSize of 8 with IsHfa, we have two possiblities: |
658 | // An HFA of one double or an HFA of two floats |
659 | // |
660 | // Check and exclude the case of an HFA of two floats |
661 | if (hfaType == TYP_DOUBLE) |
662 | { |
663 | // We have an HFA of one double |
664 | useType = TYP_DOUBLE; |
665 | } |
666 | else |
667 | { |
668 | assert(hfaType == TYP_FLOAT); |
669 | |
670 | // We have an HFA of two floats |
671 | // This should be passed or returned in two FP registers |
672 | useType = TYP_UNKNOWN; |
673 | } |
674 | #else // a 32BIT target |
675 | // A structSize of 4 with IsHfa, it must be an HFA of one float |
676 | useType = TYP_FLOAT; |
677 | #endif // _TARGET_64BIT_ |
678 | } |
679 | else |
680 | { |
681 | BYTE gcPtr = 0; |
682 | // Check if this pointer-sized struct is wrapping a GC object |
683 | info.compCompHnd->getClassGClayout(clsHnd, &gcPtr); |
684 | useType = getJitGCType(gcPtr); |
685 | } |
686 | break; |
687 | |
688 | #ifdef _TARGET_ARM_ |
689 | case 8: |
690 | if (IsHfa(clsHnd)) |
691 | { |
692 | var_types hfaType = GetHfaType(clsHnd); |
693 | |
694 | // A structSize of 8 with IsHfa, we have two possiblities: |
695 | // An HFA of one double or an HFA of two floats |
696 | // |
697 | // Check and exclude the case of an HFA of two floats |
698 | if (hfaType == TYP_DOUBLE) |
699 | { |
700 | // We have an HFA of one double |
701 | useType = TYP_DOUBLE; |
702 | } |
703 | else |
704 | { |
705 | assert(hfaType == TYP_FLOAT); |
706 | |
707 | // We have an HFA of two floats |
708 | // This should be passed or returned in two FP registers |
709 | useType = TYP_UNKNOWN; |
710 | } |
711 | } |
712 | else |
713 | { |
714 | // We don't have an HFA |
715 | useType = TYP_UNKNOWN; |
716 | } |
717 | break; |
718 | #endif // _TARGET_ARM_ |
719 | |
720 | default: |
721 | useType = TYP_UNKNOWN; |
722 | break; |
723 | } |
724 | |
725 | return useType; |
726 | } |
727 | |
728 | //----------------------------------------------------------------------------- |
729 | // getArgTypeForStruct: |
730 | // Get the type that is used to pass values of the given struct type. |
731 | // If you have already retrieved the struct size then it should be |
732 | // passed as the optional third argument, as this allows us to avoid |
733 | // an extra call to getClassSize(clsHnd) |
734 | // |
735 | // Arguments: |
736 | // clsHnd - the handle for the struct type |
737 | // wbPassStruct - An "out" argument with information about how |
738 | // the struct is to be passed |
739 | // isVarArg - is vararg, used to ignore HFA types for Arm64 windows varargs |
740 | // structSize - the size of the struct type, |
741 | // or zero if we should call getClassSize(clsHnd) |
742 | // |
743 | // Return Value: |
744 | // For wbPassStruct you can pass a 'nullptr' and nothing will be written |
745 | // or returned for that out parameter. |
746 | // When *wbPassStruct is SPK_PrimitiveType this method's return value |
747 | // is the primitive type used to pass the struct. |
748 | // When *wbPassStruct is SPK_ByReference this method's return value |
749 | // is always TYP_UNKNOWN and the struct type is passed by reference to a copy |
750 | // When *wbPassStruct is SPK_ByValue or SPK_ByValueAsHfa this method's return value |
751 | // is always TYP_STRUCT and the struct type is passed by value either |
752 | // using multiple registers or on the stack. |
753 | // |
754 | // Assumptions: |
755 | // The size must be the size of the given type. |
756 | // The given class handle must be for a value type (struct). |
757 | // |
758 | // Notes: |
759 | // About HFA types: |
760 | // When the clsHnd is a one element HFA type we return the appropriate |
761 | // floating point primitive type and *wbPassStruct is SPK_PrimitiveType |
762 | // If there are two or more elements in the HFA type then the this method's |
763 | // return value is TYP_STRUCT and *wbPassStruct is SPK_ByValueAsHfa |
764 | // |
765 | var_types Compiler::getArgTypeForStruct(CORINFO_CLASS_HANDLE clsHnd, |
766 | structPassingKind* wbPassStruct, |
767 | bool isVarArg, |
768 | unsigned structSize) |
769 | { |
770 | var_types useType = TYP_UNKNOWN; |
771 | structPassingKind howToPassStruct = SPK_Unknown; // We must change this before we return |
772 | |
773 | assert(structSize != 0); |
774 | |
775 | // Determine if we can pass the struct as a primitive type. |
776 | // Note that on x86 we never pass structs as primitive types (unless the VM unwraps them for us). |
777 | #ifndef _TARGET_X86_ |
778 | #ifdef UNIX_AMD64_ABI |
779 | |
780 | // An 8-byte struct may need to be passed in a floating point register |
781 | // So we always consult the struct "Classifier" routine |
782 | // |
783 | SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc; |
784 | eeGetSystemVAmd64PassStructInRegisterDescriptor(clsHnd, &structDesc); |
785 | |
786 | if (structDesc.passedInRegisters && (structDesc.eightByteCount != 1)) |
787 | { |
788 | // We can't pass this as a primitive type. |
789 | } |
790 | else if (structDesc.eightByteClassifications[0] == SystemVClassificationTypeSSE) |
791 | { |
792 | // If this is passed as a floating type, use that. |
793 | // Otherwise, we'll use the general case - we don't want to use the "EightByteType" |
794 | // directly, because it returns `TYP_INT` for any integral type <= 4 bytes, and |
795 | // we need to preserve small types. |
796 | useType = GetEightByteType(structDesc, 0); |
797 | } |
798 | else |
799 | #endif // UNIX_AMD64_ABI |
800 | |
801 | // The largest primitive type is 8 bytes (TYP_DOUBLE) |
802 | // so we can skip calling getPrimitiveTypeForStruct when we |
803 | // have a struct that is larger than that. |
804 | // |
805 | if (structSize <= sizeof(double)) |
806 | { |
807 | // We set the "primitive" useType based upon the structSize |
808 | // and also examine the clsHnd to see if it is an HFA of count one |
809 | useType = getPrimitiveTypeForStruct(structSize, clsHnd, isVarArg); |
810 | } |
811 | |
812 | #endif // !_TARGET_X86_ |
813 | |
814 | // Did we change this struct type into a simple "primitive" type? |
815 | // |
816 | if (useType != TYP_UNKNOWN) |
817 | { |
818 | // Yes, we should use the "primitive" type in 'useType' |
819 | howToPassStruct = SPK_PrimitiveType; |
820 | } |
821 | else // We can't replace the struct with a "primitive" type |
822 | { |
823 | // See if we can pass this struct by value, possibly in multiple registers |
824 | // or if we should pass it by reference to a copy |
825 | // |
826 | if (structSize <= MAX_PASS_MULTIREG_BYTES) |
827 | { |
828 | // Structs that are HFA's are passed by value in multiple registers |
829 | if (IsHfa(clsHnd) |
830 | #if defined(_TARGET_WINDOWS_) && defined(_TARGET_ARM64_) |
831 | && !isVarArg // Arm64 Windows VarArg methods arguments will not |
832 | // classify HFA types, they will need to be treated |
833 | // as if they are not HFA types. |
834 | #endif // defined(_TARGET_WINDOWS_) && defined(_TARGET_ARM64_) |
835 | ) |
836 | { |
837 | // HFA's of count one should have been handled by getPrimitiveTypeForStruct |
838 | assert(GetHfaCount(clsHnd) >= 2); |
839 | |
840 | // setup wbPassType and useType indicate that this is passed by value as an HFA |
841 | // using multiple registers |
842 | // (when all of the parameters registers are used, then the stack will be used) |
843 | howToPassStruct = SPK_ByValueAsHfa; |
844 | useType = TYP_STRUCT; |
845 | } |
846 | else // Not an HFA struct type |
847 | { |
848 | |
849 | #ifdef UNIX_AMD64_ABI |
850 | |
851 | // The case of (structDesc.eightByteCount == 1) should have already been handled |
852 | if ((structDesc.eightByteCount > 1) || !structDesc.passedInRegisters) |
853 | { |
854 | // setup wbPassType and useType indicate that this is passed by value in multiple registers |
855 | // (when all of the parameters registers are used, then the stack will be used) |
856 | howToPassStruct = SPK_ByValue; |
857 | useType = TYP_STRUCT; |
858 | } |
859 | else |
860 | { |
861 | assert(structDesc.eightByteCount == 0); |
862 | // Otherwise we pass this struct by reference to a copy |
863 | // setup wbPassType and useType indicate that this is passed using one register |
864 | // (by reference to a copy) |
865 | howToPassStruct = SPK_ByReference; |
866 | useType = TYP_UNKNOWN; |
867 | } |
868 | |
869 | #elif defined(_TARGET_ARM64_) |
870 | |
871 | // Structs that are pointer sized or smaller should have been handled by getPrimitiveTypeForStruct |
872 | assert(structSize > TARGET_POINTER_SIZE); |
873 | |
874 | // On ARM64 structs that are 9-16 bytes are passed by value in multiple registers |
875 | // |
876 | if (structSize <= (TARGET_POINTER_SIZE * 2)) |
877 | { |
878 | // setup wbPassType and useType indicate that this is passed by value in multiple registers |
879 | // (when all of the parameters registers are used, then the stack will be used) |
880 | howToPassStruct = SPK_ByValue; |
881 | useType = TYP_STRUCT; |
882 | } |
883 | else // a structSize that is 17-32 bytes in size |
884 | { |
885 | // Otherwise we pass this struct by reference to a copy |
886 | // setup wbPassType and useType indicate that this is passed using one register |
887 | // (by reference to a copy) |
888 | howToPassStruct = SPK_ByReference; |
889 | useType = TYP_UNKNOWN; |
890 | } |
891 | |
892 | #elif defined(_TARGET_X86_) || defined(_TARGET_ARM_) |
893 | |
894 | // Otherwise we pass this struct by value on the stack |
895 | // setup wbPassType and useType indicate that this is passed by value according to the X86/ARM32 ABI |
896 | howToPassStruct = SPK_ByValue; |
897 | useType = TYP_STRUCT; |
898 | |
899 | #else // _TARGET_XXX_ |
900 | |
901 | noway_assert(!"Unhandled TARGET in getArgTypeForStruct (with FEATURE_MULTIREG_ARGS=1)" ); |
902 | |
903 | #endif // _TARGET_XXX_ |
904 | } |
905 | } |
906 | else // (structSize > MAX_PASS_MULTIREG_BYTES) |
907 | { |
908 | // We have a (large) struct that can't be replaced with a "primitive" type |
909 | // and can't be passed in multiple registers |
910 | CLANG_FORMAT_COMMENT_ANCHOR; |
911 | |
912 | #if defined(_TARGET_X86_) || defined(_TARGET_ARM_) || defined(UNIX_AMD64_ABI) |
913 | |
914 | // Otherwise we pass this struct by value on the stack |
915 | // setup wbPassType and useType indicate that this is passed by value according to the X86/ARM32 ABI |
916 | howToPassStruct = SPK_ByValue; |
917 | useType = TYP_STRUCT; |
918 | |
919 | #elif defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_) |
920 | |
921 | // Otherwise we pass this struct by reference to a copy |
922 | // setup wbPassType and useType indicate that this is passed using one register (by reference to a copy) |
923 | howToPassStruct = SPK_ByReference; |
924 | useType = TYP_UNKNOWN; |
925 | |
926 | #else // _TARGET_XXX_ |
927 | |
928 | noway_assert(!"Unhandled TARGET in getArgTypeForStruct" ); |
929 | |
930 | #endif // _TARGET_XXX_ |
931 | } |
932 | } |
933 | |
934 | // 'howToPassStruct' must be set to one of the valid values before we return |
935 | assert(howToPassStruct != SPK_Unknown); |
936 | if (wbPassStruct != nullptr) |
937 | { |
938 | *wbPassStruct = howToPassStruct; |
939 | } |
940 | |
941 | return useType; |
942 | } |
943 | |
944 | //----------------------------------------------------------------------------- |
945 | // getReturnTypeForStruct: |
946 | // Get the type that is used to return values of the given struct type. |
947 | // If you have already retrieved the struct size then it should be |
948 | // passed as the optional third argument, as this allows us to avoid |
949 | // an extra call to getClassSize(clsHnd) |
950 | // |
951 | // Arguments: |
952 | // clsHnd - the handle for the struct type |
953 | // wbReturnStruct - An "out" argument with information about how |
954 | // the struct is to be returned |
955 | // structSize - the size of the struct type, |
956 | // or zero if we should call getClassSize(clsHnd) |
957 | // |
958 | // Return Value: |
959 | // For wbReturnStruct you can pass a 'nullptr' and nothing will be written |
960 | // or returned for that out parameter. |
961 | // When *wbReturnStruct is SPK_PrimitiveType this method's return value |
962 | // is the primitive type used to return the struct. |
963 | // When *wbReturnStruct is SPK_ByReference this method's return value |
964 | // is always TYP_UNKNOWN and the struct type is returned using a return buffer |
965 | // When *wbReturnStruct is SPK_ByValue or SPK_ByValueAsHfa this method's return value |
966 | // is always TYP_STRUCT and the struct type is returned using multiple registers. |
967 | // |
968 | // Assumptions: |
969 | // The size must be the size of the given type. |
970 | // The given class handle must be for a value type (struct). |
971 | // |
972 | // Notes: |
973 | // About HFA types: |
974 | // When the clsHnd is a one element HFA type then this method's return |
975 | // value is the appropriate floating point primitive type and |
976 | // *wbReturnStruct is SPK_PrimitiveType. |
977 | // If there are two or more elements in the HFA type and the target supports |
978 | // multireg return types then the return value is TYP_STRUCT and |
979 | // *wbReturnStruct is SPK_ByValueAsHfa. |
980 | // Additionally if there are two or more elements in the HFA type and |
981 | // the target doesn't support multreg return types then it is treated |
982 | // as if it wasn't an HFA type. |
983 | // About returning TYP_STRUCT: |
984 | // Whenever this method's return value is TYP_STRUCT it always means |
985 | // that multiple registers are used to return this struct. |
986 | // |
987 | var_types Compiler::getReturnTypeForStruct(CORINFO_CLASS_HANDLE clsHnd, |
988 | structPassingKind* wbReturnStruct /* = nullptr */, |
989 | unsigned structSize /* = 0 */) |
990 | { |
991 | var_types useType = TYP_UNKNOWN; |
992 | structPassingKind howToReturnStruct = SPK_Unknown; // We must change this before we return |
993 | bool canReturnInRegister = true; |
994 | |
995 | assert(clsHnd != NO_CLASS_HANDLE); |
996 | |
997 | if (structSize == 0) |
998 | { |
999 | structSize = info.compCompHnd->getClassSize(clsHnd); |
1000 | } |
1001 | assert(structSize > 0); |
1002 | |
1003 | #ifdef UNIX_AMD64_ABI |
1004 | // An 8-byte struct may need to be returned in a floating point register |
1005 | // So we always consult the struct "Classifier" routine |
1006 | // |
1007 | SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc; |
1008 | eeGetSystemVAmd64PassStructInRegisterDescriptor(clsHnd, &structDesc); |
1009 | |
1010 | if (structDesc.eightByteCount == 1) |
1011 | { |
1012 | assert(structSize <= sizeof(double)); |
1013 | assert(structDesc.passedInRegisters); |
1014 | |
1015 | if (structDesc.eightByteClassifications[0] == SystemVClassificationTypeSSE) |
1016 | { |
1017 | // If this is returned as a floating type, use that. |
1018 | // Otherwise, leave as TYP_UNKONWN and we'll sort things out below. |
1019 | useType = GetEightByteType(structDesc, 0); |
1020 | howToReturnStruct = SPK_PrimitiveType; |
1021 | } |
1022 | } |
1023 | else |
1024 | { |
1025 | // Return classification is not always size based... |
1026 | canReturnInRegister = structDesc.passedInRegisters; |
1027 | } |
1028 | |
1029 | #endif // UNIX_AMD64_ABI |
1030 | |
1031 | // Check for cases where a small struct is returned in a register |
1032 | // via a primitive type. |
1033 | // |
1034 | // The largest primitive type is 8 bytes (TYP_DOUBLE) |
1035 | // so we can skip calling getPrimitiveTypeForStruct when we |
1036 | // have a struct that is larger than that. |
1037 | if (canReturnInRegister && (useType == TYP_UNKNOWN) && (structSize <= sizeof(double))) |
1038 | { |
1039 | // We set the "primitive" useType based upon the structSize |
1040 | // and also examine the clsHnd to see if it is an HFA of count one |
1041 | // |
1042 | // The ABI for struct returns in varArg methods, is same as the normal case, |
1043 | // so pass false for isVararg |
1044 | useType = getPrimitiveTypeForStruct(structSize, clsHnd, /*isVararg=*/false); |
1045 | |
1046 | if (useType != TYP_UNKNOWN) |
1047 | { |
1048 | if (structSize == genTypeSize(useType)) |
1049 | { |
1050 | // Currently: 1, 2, 4, or 8 byte structs |
1051 | howToReturnStruct = SPK_PrimitiveType; |
1052 | } |
1053 | else |
1054 | { |
1055 | // Currently: 3, 5, 6, or 7 byte structs |
1056 | assert(structSize < genTypeSize(useType)); |
1057 | howToReturnStruct = SPK_EnclosingType; |
1058 | } |
1059 | } |
1060 | } |
1061 | |
1062 | #ifdef _TARGET_64BIT_ |
1063 | // Note this handles an odd case when FEATURE_MULTIREG_RET is disabled and HFAs are enabled |
1064 | // |
1065 | // getPrimitiveTypeForStruct will return TYP_UNKNOWN for a struct that is an HFA of two floats |
1066 | // because when HFA are enabled, normally we would use two FP registers to pass or return it |
1067 | // |
1068 | // But if we don't have support for multiple register return types, we have to change this. |
1069 | // Since we what we have an 8-byte struct (float + float) we change useType to TYP_I_IMPL |
1070 | // so that the struct is returned instead using an 8-byte integer register. |
1071 | // |
1072 | if ((FEATURE_MULTIREG_RET == 0) && (useType == TYP_UNKNOWN) && (structSize == (2 * sizeof(float))) && IsHfa(clsHnd)) |
1073 | { |
1074 | useType = TYP_I_IMPL; |
1075 | howToReturnStruct = SPK_PrimitiveType; |
1076 | } |
1077 | #endif |
1078 | |
1079 | // Did we change this struct type into a simple "primitive" type? |
1080 | if (useType != TYP_UNKNOWN) |
1081 | { |
1082 | // If so, we should have already set howToReturnStruct, too. |
1083 | assert(howToReturnStruct != SPK_Unknown); |
1084 | } |
1085 | else // We can't replace the struct with a "primitive" type |
1086 | { |
1087 | // See if we can return this struct by value, possibly in multiple registers |
1088 | // or if we should return it using a return buffer register |
1089 | // |
1090 | if ((FEATURE_MULTIREG_RET == 1) && (structSize <= MAX_RET_MULTIREG_BYTES)) |
1091 | { |
1092 | // Structs that are HFA's are returned in multiple registers |
1093 | if (IsHfa(clsHnd)) |
1094 | { |
1095 | // HFA's of count one should have been handled by getPrimitiveTypeForStruct |
1096 | assert(GetHfaCount(clsHnd) >= 2); |
1097 | |
1098 | // setup wbPassType and useType indicate that this is returned by value as an HFA |
1099 | // using multiple registers |
1100 | howToReturnStruct = SPK_ByValueAsHfa; |
1101 | useType = TYP_STRUCT; |
1102 | } |
1103 | else // Not an HFA struct type |
1104 | { |
1105 | |
1106 | #ifdef UNIX_AMD64_ABI |
1107 | |
1108 | // The case of (structDesc.eightByteCount == 1) should have already been handled |
1109 | if (structDesc.eightByteCount > 1) |
1110 | { |
1111 | // setup wbPassType and useType indicate that this is returned by value in multiple registers |
1112 | howToReturnStruct = SPK_ByValue; |
1113 | useType = TYP_STRUCT; |
1114 | assert(structDesc.passedInRegisters == true); |
1115 | } |
1116 | else |
1117 | { |
1118 | assert(structDesc.eightByteCount == 0); |
1119 | // Otherwise we return this struct using a return buffer |
1120 | // setup wbPassType and useType indicate that this is return using a return buffer register |
1121 | // (reference to a return buffer) |
1122 | howToReturnStruct = SPK_ByReference; |
1123 | useType = TYP_UNKNOWN; |
1124 | assert(structDesc.passedInRegisters == false); |
1125 | } |
1126 | |
1127 | #elif defined(_TARGET_ARM64_) |
1128 | |
1129 | // Structs that are pointer sized or smaller should have been handled by getPrimitiveTypeForStruct |
1130 | assert(structSize > TARGET_POINTER_SIZE); |
1131 | |
1132 | // On ARM64 structs that are 9-16 bytes are returned by value in multiple registers |
1133 | // |
1134 | if (structSize <= (TARGET_POINTER_SIZE * 2)) |
1135 | { |
1136 | // setup wbPassType and useType indicate that this is return by value in multiple registers |
1137 | howToReturnStruct = SPK_ByValue; |
1138 | useType = TYP_STRUCT; |
1139 | } |
1140 | else // a structSize that is 17-32 bytes in size |
1141 | { |
1142 | // Otherwise we return this struct using a return buffer |
1143 | // setup wbPassType and useType indicate that this is returned using a return buffer register |
1144 | // (reference to a return buffer) |
1145 | howToReturnStruct = SPK_ByReference; |
1146 | useType = TYP_UNKNOWN; |
1147 | } |
1148 | |
1149 | #elif defined(_TARGET_ARM_) || defined(_TARGET_X86_) |
1150 | |
1151 | // Otherwise we return this struct using a return buffer |
1152 | // setup wbPassType and useType indicate that this is returned using a return buffer register |
1153 | // (reference to a return buffer) |
1154 | howToReturnStruct = SPK_ByReference; |
1155 | useType = TYP_UNKNOWN; |
1156 | |
1157 | #else // _TARGET_XXX_ |
1158 | |
1159 | noway_assert(!"Unhandled TARGET in getReturnTypeForStruct (with FEATURE_MULTIREG_ARGS=1)" ); |
1160 | |
1161 | #endif // _TARGET_XXX_ |
1162 | } |
1163 | } |
1164 | else // (structSize > MAX_RET_MULTIREG_BYTES) || (FEATURE_MULTIREG_RET == 0) |
1165 | { |
1166 | // We have a (large) struct that can't be replaced with a "primitive" type |
1167 | // and can't be returned in multiple registers |
1168 | |
1169 | // We return this struct using a return buffer register |
1170 | // setup wbPassType and useType indicate that this is returned using a return buffer register |
1171 | // (reference to a return buffer) |
1172 | howToReturnStruct = SPK_ByReference; |
1173 | useType = TYP_UNKNOWN; |
1174 | } |
1175 | } |
1176 | |
1177 | // 'howToReturnStruct' must be set to one of the valid values before we return |
1178 | assert(howToReturnStruct != SPK_Unknown); |
1179 | if (wbReturnStruct != nullptr) |
1180 | { |
1181 | *wbReturnStruct = howToReturnStruct; |
1182 | } |
1183 | |
1184 | return useType; |
1185 | } |
1186 | |
1187 | /////////////////////////////////////////////////////////////////////////////// |
1188 | // |
1189 | // MEASURE_NOWAY: code to measure and rank dynamic occurences of noway_assert. |
1190 | // (Just the appearances of noway_assert, whether the assert is true or false.) |
1191 | // This might help characterize the cost of noway_assert in non-DEBUG builds, |
1192 | // or determine which noway_assert should be simple DEBUG-only asserts. |
1193 | // |
1194 | /////////////////////////////////////////////////////////////////////////////// |
1195 | |
1196 | #if MEASURE_NOWAY |
1197 | |
1198 | struct FileLine |
1199 | { |
1200 | char* m_file; |
1201 | unsigned m_line; |
1202 | char* m_condStr; |
1203 | |
1204 | FileLine() : m_file(nullptr), m_line(0), m_condStr(nullptr) |
1205 | { |
1206 | } |
1207 | |
1208 | FileLine(const char* file, unsigned line, const char* condStr) : m_line(line) |
1209 | { |
1210 | size_t newSize = (strlen(file) + 1) * sizeof(char); |
1211 | m_file = HostAllocator::getHostAllocator().allocate<char>(newSize); |
1212 | strcpy_s(m_file, newSize, file); |
1213 | |
1214 | newSize = (strlen(condStr) + 1) * sizeof(char); |
1215 | m_condStr = HostAllocator::getHostAllocator().allocate<char>(newSize); |
1216 | strcpy_s(m_condStr, newSize, condStr); |
1217 | } |
1218 | |
1219 | FileLine(const FileLine& other) |
1220 | { |
1221 | m_file = other.m_file; |
1222 | m_line = other.m_line; |
1223 | m_condStr = other.m_condStr; |
1224 | } |
1225 | |
1226 | // GetHashCode() and Equals() are needed by JitHashTable |
1227 | |
1228 | static unsigned GetHashCode(FileLine fl) |
1229 | { |
1230 | assert(fl.m_file != nullptr); |
1231 | unsigned code = fl.m_line; |
1232 | for (const char* p = fl.m_file; *p != '\0'; p++) |
1233 | { |
1234 | code += *p; |
1235 | } |
1236 | // Could also add condStr. |
1237 | return code; |
1238 | } |
1239 | |
1240 | static bool Equals(FileLine fl1, FileLine fl2) |
1241 | { |
1242 | return (fl1.m_line == fl2.m_line) && (0 == strcmp(fl1.m_file, fl2.m_file)); |
1243 | } |
1244 | }; |
1245 | |
1246 | typedef JitHashTable<FileLine, FileLine, size_t, HostAllocator> FileLineToCountMap; |
1247 | FileLineToCountMap* NowayAssertMap; |
1248 | |
1249 | void Compiler::RecordNowayAssert(const char* filename, unsigned line, const char* condStr) |
1250 | { |
1251 | if (NowayAssertMap == nullptr) |
1252 | { |
1253 | NowayAssertMap = new (HostAllocator::getHostAllocator()) FileLineToCountMap(HostAllocator::getHostAllocator()); |
1254 | } |
1255 | FileLine fl(filename, line, condStr); |
1256 | size_t* pCount = NowayAssertMap->LookupPointer(fl); |
1257 | if (pCount == nullptr) |
1258 | { |
1259 | NowayAssertMap->Set(fl, 1); |
1260 | } |
1261 | else |
1262 | { |
1263 | ++(*pCount); |
1264 | } |
1265 | } |
1266 | |
1267 | void RecordNowayAssertGlobal(const char* filename, unsigned line, const char* condStr) |
1268 | { |
1269 | if ((JitConfig.JitMeasureNowayAssert() == 1) && (JitTls::GetCompiler() != nullptr)) |
1270 | { |
1271 | JitTls::GetCompiler()->RecordNowayAssert(filename, line, condStr); |
1272 | } |
1273 | } |
1274 | |
1275 | struct NowayAssertCountMap |
1276 | { |
1277 | size_t count; |
1278 | FileLine fl; |
1279 | |
1280 | NowayAssertCountMap() : count(0) |
1281 | { |
1282 | } |
1283 | |
1284 | static int __cdecl compare(const void* elem1, const void* elem2) |
1285 | { |
1286 | NowayAssertCountMap* e1 = (NowayAssertCountMap*)elem1; |
1287 | NowayAssertCountMap* e2 = (NowayAssertCountMap*)elem2; |
1288 | return (int)((ssize_t)e2->count - (ssize_t)e1->count); // sort in descending order |
1289 | } |
1290 | }; |
1291 | |
1292 | void DisplayNowayAssertMap() |
1293 | { |
1294 | if (NowayAssertMap != nullptr) |
1295 | { |
1296 | FILE* fout; |
1297 | |
1298 | LPCWSTR strJitMeasureNowayAssertFile = JitConfig.JitMeasureNowayAssertFile(); |
1299 | if (strJitMeasureNowayAssertFile != nullptr) |
1300 | { |
1301 | fout = _wfopen(strJitMeasureNowayAssertFile, W("a" )); |
1302 | if (fout == nullptr) |
1303 | { |
1304 | fprintf(jitstdout, "Failed to open JitMeasureNowayAssertFile \"%ws\"\n" , strJitMeasureNowayAssertFile); |
1305 | return; |
1306 | } |
1307 | } |
1308 | else |
1309 | { |
1310 | fout = jitstdout; |
1311 | } |
1312 | |
1313 | // Iterate noway assert map, create sorted table by occurrence, dump it. |
1314 | unsigned count = NowayAssertMap->GetCount(); |
1315 | NowayAssertCountMap* nacp = new NowayAssertCountMap[count]; |
1316 | unsigned i = 0; |
1317 | |
1318 | for (FileLineToCountMap::KeyIterator iter = NowayAssertMap->Begin(), end = NowayAssertMap->End(); |
1319 | !iter.Equal(end); ++iter) |
1320 | { |
1321 | nacp[i].count = iter.GetValue(); |
1322 | nacp[i].fl = iter.Get(); |
1323 | ++i; |
1324 | } |
1325 | |
1326 | qsort(nacp, count, sizeof(nacp[0]), NowayAssertCountMap::compare); |
1327 | |
1328 | if (fout == jitstdout) |
1329 | { |
1330 | // Don't output the header if writing to a file, since we'll be appending to existing dumps in that case. |
1331 | fprintf(fout, "\nnoway_assert counts:\n" ); |
1332 | fprintf(fout, "count, file, line, text\n" ); |
1333 | } |
1334 | |
1335 | for (i = 0; i < count; i++) |
1336 | { |
1337 | fprintf(fout, "%u, %s, %u, \"%s\"\n" , nacp[i].count, nacp[i].fl.m_file, nacp[i].fl.m_line, |
1338 | nacp[i].fl.m_condStr); |
1339 | } |
1340 | |
1341 | if (fout != jitstdout) |
1342 | { |
1343 | fclose(fout); |
1344 | fout = nullptr; |
1345 | } |
1346 | } |
1347 | } |
1348 | |
1349 | #endif // MEASURE_NOWAY |
1350 | |
1351 | /***************************************************************************** |
1352 | * variables to keep track of how many iterations we go in a dataflow pass |
1353 | */ |
1354 | |
1355 | #if DATAFLOW_ITER |
1356 | |
1357 | unsigned CSEiterCount; // counts the # of iteration for the CSE dataflow |
1358 | unsigned CFiterCount; // counts the # of iteration for the Const Folding dataflow |
1359 | |
1360 | #endif // DATAFLOW_ITER |
1361 | |
1362 | #if MEASURE_BLOCK_SIZE |
1363 | size_t genFlowNodeSize; |
1364 | size_t genFlowNodeCnt; |
1365 | #endif // MEASURE_BLOCK_SIZE |
1366 | |
1367 | /*****************************************************************************/ |
1368 | // We keep track of methods we've already compiled. |
1369 | |
1370 | /***************************************************************************** |
1371 | * Declare the statics |
1372 | */ |
1373 | |
1374 | #ifdef DEBUG |
1375 | /* static */ |
1376 | unsigned Compiler::s_compMethodsCount = 0; // to produce unique label names |
1377 | #endif |
1378 | |
1379 | #if MEASURE_MEM_ALLOC |
1380 | /* static */ |
1381 | bool Compiler::s_dspMemStats = false; |
1382 | #endif |
1383 | |
1384 | #ifndef PROFILING_SUPPORTED |
1385 | const bool Compiler::Options::compNoPInvokeInlineCB = false; |
1386 | #endif |
1387 | |
1388 | /***************************************************************************** |
1389 | * |
1390 | * One time initialization code |
1391 | */ |
1392 | |
1393 | /* static */ |
1394 | void Compiler::compStartup() |
1395 | { |
1396 | #if DISPLAY_SIZES |
1397 | grossVMsize = grossNCsize = totalNCsize = 0; |
1398 | #endif // DISPLAY_SIZES |
1399 | |
1400 | /* Initialize the table of tree node sizes */ |
1401 | |
1402 | GenTree::InitNodeSize(); |
1403 | |
1404 | #ifdef JIT32_GCENCODER |
1405 | // Initialize the GC encoder lookup table |
1406 | |
1407 | GCInfo::gcInitEncoderLookupTable(); |
1408 | #endif |
1409 | |
1410 | /* Initialize the emitter */ |
1411 | |
1412 | emitter::emitInit(); |
1413 | |
1414 | // Static vars of ValueNumStore |
1415 | ValueNumStore::InitValueNumStoreStatics(); |
1416 | |
1417 | compDisplayStaticSizes(jitstdout); |
1418 | } |
1419 | |
1420 | /***************************************************************************** |
1421 | * |
1422 | * One time finalization code |
1423 | */ |
1424 | |
1425 | /* static */ |
1426 | void Compiler::compShutdown() |
1427 | { |
1428 | #ifdef ALT_JIT |
1429 | if (s_pAltJitExcludeAssembliesList != nullptr) |
1430 | { |
1431 | s_pAltJitExcludeAssembliesList->~AssemblyNamesList2(); // call the destructor |
1432 | s_pAltJitExcludeAssembliesList = nullptr; |
1433 | } |
1434 | #endif // ALT_JIT |
1435 | |
1436 | #ifdef DEBUG |
1437 | if (s_pJitDisasmIncludeAssembliesList != nullptr) |
1438 | { |
1439 | s_pJitDisasmIncludeAssembliesList->~AssemblyNamesList2(); // call the destructor |
1440 | s_pJitDisasmIncludeAssembliesList = nullptr; |
1441 | } |
1442 | #endif // DEBUG |
1443 | |
1444 | #if MEASURE_NOWAY |
1445 | DisplayNowayAssertMap(); |
1446 | #endif // MEASURE_NOWAY |
1447 | |
1448 | /* Shut down the emitter */ |
1449 | |
1450 | emitter::emitDone(); |
1451 | |
1452 | #if defined(DEBUG) || defined(INLINE_DATA) |
1453 | // Finish reading and/or writing inline xml |
1454 | InlineStrategy::FinalizeXml(); |
1455 | #endif // defined(DEBUG) || defined(INLINE_DATA) |
1456 | |
1457 | #if defined(DEBUG) || MEASURE_NODE_SIZE || MEASURE_BLOCK_SIZE || DISPLAY_SIZES || CALL_ARG_STATS |
1458 | if (genMethodCnt == 0) |
1459 | { |
1460 | return; |
1461 | } |
1462 | #endif |
1463 | |
1464 | #if NODEBASH_STATS |
1465 | GenTree::ReportOperBashing(jitstdout); |
1466 | #endif |
1467 | |
1468 | // Where should we write our statistics output? |
1469 | FILE* fout = jitstdout; |
1470 | |
1471 | #ifdef FEATURE_JIT_METHOD_PERF |
1472 | if (compJitTimeLogFilename != nullptr) |
1473 | { |
1474 | FILE* jitTimeLogFile = _wfopen(compJitTimeLogFilename, W("a" )); |
1475 | if (jitTimeLogFile != nullptr) |
1476 | { |
1477 | CompTimeSummaryInfo::s_compTimeSummary.Print(jitTimeLogFile); |
1478 | fclose(jitTimeLogFile); |
1479 | } |
1480 | } |
1481 | #endif // FEATURE_JIT_METHOD_PERF |
1482 | |
1483 | #if COUNT_AST_OPERS |
1484 | |
1485 | // Add up all the counts so that we can show percentages of total |
1486 | unsigned gtc = 0; |
1487 | for (unsigned op = 0; op < GT_COUNT; op++) |
1488 | gtc += GenTree::s_gtNodeCounts[op]; |
1489 | |
1490 | if (gtc > 0) |
1491 | { |
1492 | unsigned rem_total = gtc; |
1493 | unsigned rem_large = 0; |
1494 | unsigned rem_small = 0; |
1495 | |
1496 | unsigned tot_large = 0; |
1497 | unsigned tot_small = 0; |
1498 | |
1499 | fprintf(fout, "\nGenTree operator counts (approximate):\n\n" ); |
1500 | |
1501 | for (unsigned op = 0; op < GT_COUNT; op++) |
1502 | { |
1503 | unsigned siz = GenTree::s_gtTrueSizes[op]; |
1504 | unsigned cnt = GenTree::s_gtNodeCounts[op]; |
1505 | double pct = 100.0 * cnt / gtc; |
1506 | |
1507 | if (siz > TREE_NODE_SZ_SMALL) |
1508 | tot_large += cnt; |
1509 | else |
1510 | tot_small += cnt; |
1511 | |
1512 | // Let's not show anything below a threshold |
1513 | if (pct >= 0.5) |
1514 | { |
1515 | fprintf(fout, " GT_%-17s %7u (%4.1lf%%) %3u bytes each\n" , GenTree::OpName((genTreeOps)op), cnt, |
1516 | pct, siz); |
1517 | rem_total -= cnt; |
1518 | } |
1519 | else |
1520 | { |
1521 | if (siz > TREE_NODE_SZ_SMALL) |
1522 | rem_large += cnt; |
1523 | else |
1524 | rem_small += cnt; |
1525 | } |
1526 | } |
1527 | if (rem_total > 0) |
1528 | { |
1529 | fprintf(fout, " All other GT_xxx ... %7u (%4.1lf%%) ... %4.1lf%% small + %4.1lf%% large\n" , rem_total, |
1530 | 100.0 * rem_total / gtc, 100.0 * rem_small / gtc, 100.0 * rem_large / gtc); |
1531 | } |
1532 | fprintf(fout, " -----------------------------------------------------\n" ); |
1533 | fprintf(fout, " Total ....... %11u --ALL-- ... %4.1lf%% small + %4.1lf%% large\n" , gtc, |
1534 | 100.0 * tot_small / gtc, 100.0 * tot_large / gtc); |
1535 | fprintf(fout, "\n" ); |
1536 | } |
1537 | |
1538 | #endif // COUNT_AST_OPERS |
1539 | |
1540 | #if DISPLAY_SIZES |
1541 | |
1542 | if (grossVMsize && grossNCsize) |
1543 | { |
1544 | fprintf(fout, "\n" ); |
1545 | fprintf(fout, "--------------------------------------\n" ); |
1546 | fprintf(fout, "Function and GC info size stats\n" ); |
1547 | fprintf(fout, "--------------------------------------\n" ); |
1548 | |
1549 | fprintf(fout, "[%7u VM, %8u %6s %4u%%] %s\n" , grossVMsize, grossNCsize, Target::g_tgtCPUName, |
1550 | 100 * grossNCsize / grossVMsize, "Total (excluding GC info)" ); |
1551 | |
1552 | fprintf(fout, "[%7u VM, %8u %6s %4u%%] %s\n" , grossVMsize, totalNCsize, Target::g_tgtCPUName, |
1553 | 100 * totalNCsize / grossVMsize, "Total (including GC info)" ); |
1554 | |
1555 | if (gcHeaderISize || gcHeaderNSize) |
1556 | { |
1557 | fprintf(fout, "\n" ); |
1558 | |
1559 | fprintf(fout, "GC tables : [%7uI,%7uN] %7u byt (%u%% of IL, %u%% of %s).\n" , |
1560 | gcHeaderISize + gcPtrMapISize, gcHeaderNSize + gcPtrMapNSize, totalNCsize - grossNCsize, |
1561 | 100 * (totalNCsize - grossNCsize) / grossVMsize, 100 * (totalNCsize - grossNCsize) / grossNCsize, |
1562 | Target::g_tgtCPUName); |
1563 | |
1564 | fprintf(fout, "GC headers : [%7uI,%7uN] %7u byt, [%4.1fI,%4.1fN] %4.1f byt/meth\n" , gcHeaderISize, |
1565 | gcHeaderNSize, gcHeaderISize + gcHeaderNSize, (float)gcHeaderISize / (genMethodICnt + 0.001), |
1566 | (float)gcHeaderNSize / (genMethodNCnt + 0.001), |
1567 | (float)(gcHeaderISize + gcHeaderNSize) / genMethodCnt); |
1568 | |
1569 | fprintf(fout, "GC ptr maps : [%7uI,%7uN] %7u byt, [%4.1fI,%4.1fN] %4.1f byt/meth\n" , gcPtrMapISize, |
1570 | gcPtrMapNSize, gcPtrMapISize + gcPtrMapNSize, (float)gcPtrMapISize / (genMethodICnt + 0.001), |
1571 | (float)gcPtrMapNSize / (genMethodNCnt + 0.001), |
1572 | (float)(gcPtrMapISize + gcPtrMapNSize) / genMethodCnt); |
1573 | } |
1574 | else |
1575 | { |
1576 | fprintf(fout, "\n" ); |
1577 | |
1578 | fprintf(fout, "GC tables take up %u bytes (%u%% of instr, %u%% of %6s code).\n" , |
1579 | totalNCsize - grossNCsize, 100 * (totalNCsize - grossNCsize) / grossVMsize, |
1580 | 100 * (totalNCsize - grossNCsize) / grossNCsize, Target::g_tgtCPUName); |
1581 | } |
1582 | |
1583 | #ifdef DEBUG |
1584 | #if DOUBLE_ALIGN |
1585 | fprintf(fout, "%u out of %u methods generated with double-aligned stack\n" , |
1586 | Compiler::s_lvaDoubleAlignedProcsCount, genMethodCnt); |
1587 | #endif |
1588 | #endif |
1589 | } |
1590 | |
1591 | #endif // DISPLAY_SIZES |
1592 | |
1593 | #if CALL_ARG_STATS |
1594 | compDispCallArgStats(fout); |
1595 | #endif |
1596 | |
1597 | #if COUNT_BASIC_BLOCKS |
1598 | fprintf(fout, "--------------------------------------------------\n" ); |
1599 | fprintf(fout, "Basic block count frequency table:\n" ); |
1600 | fprintf(fout, "--------------------------------------------------\n" ); |
1601 | bbCntTable.dump(fout); |
1602 | fprintf(fout, "--------------------------------------------------\n" ); |
1603 | |
1604 | fprintf(fout, "\n" ); |
1605 | |
1606 | fprintf(fout, "--------------------------------------------------\n" ); |
1607 | fprintf(fout, "IL method size frequency table for methods with a single basic block:\n" ); |
1608 | fprintf(fout, "--------------------------------------------------\n" ); |
1609 | bbOneBBSizeTable.dump(fout); |
1610 | fprintf(fout, "--------------------------------------------------\n" ); |
1611 | #endif // COUNT_BASIC_BLOCKS |
1612 | |
1613 | #if COUNT_LOOPS |
1614 | |
1615 | fprintf(fout, "\n" ); |
1616 | fprintf(fout, "---------------------------------------------------\n" ); |
1617 | fprintf(fout, "Loop stats\n" ); |
1618 | fprintf(fout, "---------------------------------------------------\n" ); |
1619 | fprintf(fout, "Total number of methods with loops is %5u\n" , totalLoopMethods); |
1620 | fprintf(fout, "Total number of loops is %5u\n" , totalLoopCount); |
1621 | fprintf(fout, "Maximum number of loops per method is %5u\n" , maxLoopsPerMethod); |
1622 | fprintf(fout, "# of methods overflowing nat loop table is %5u\n" , totalLoopOverflows); |
1623 | fprintf(fout, "Total number of 'unnatural' loops is %5u\n" , totalUnnatLoopCount); |
1624 | fprintf(fout, "# of methods overflowing unnat loop limit is %5u\n" , totalUnnatLoopOverflows); |
1625 | fprintf(fout, "Total number of loops with an iterator is %5u\n" , iterLoopCount); |
1626 | fprintf(fout, "Total number of loops with a simple iterator is %5u\n" , simpleTestLoopCount); |
1627 | fprintf(fout, "Total number of loops with a constant iterator is %5u\n" , constIterLoopCount); |
1628 | |
1629 | fprintf(fout, "--------------------------------------------------\n" ); |
1630 | fprintf(fout, "Loop count frequency table:\n" ); |
1631 | fprintf(fout, "--------------------------------------------------\n" ); |
1632 | loopCountTable.dump(fout); |
1633 | fprintf(fout, "--------------------------------------------------\n" ); |
1634 | fprintf(fout, "Loop exit count frequency table:\n" ); |
1635 | fprintf(fout, "--------------------------------------------------\n" ); |
1636 | loopExitCountTable.dump(fout); |
1637 | fprintf(fout, "--------------------------------------------------\n" ); |
1638 | |
1639 | #endif // COUNT_LOOPS |
1640 | |
1641 | #if DATAFLOW_ITER |
1642 | |
1643 | fprintf(fout, "---------------------------------------------------\n" ); |
1644 | fprintf(fout, "Total number of iterations in the CSE dataflow loop is %5u\n" , CSEiterCount); |
1645 | fprintf(fout, "Total number of iterations in the CF dataflow loop is %5u\n" , CFiterCount); |
1646 | |
1647 | #endif // DATAFLOW_ITER |
1648 | |
1649 | #if MEASURE_NODE_SIZE |
1650 | |
1651 | fprintf(fout, "\n" ); |
1652 | fprintf(fout, "---------------------------------------------------\n" ); |
1653 | fprintf(fout, "GenTree node allocation stats\n" ); |
1654 | fprintf(fout, "---------------------------------------------------\n" ); |
1655 | |
1656 | fprintf(fout, "Allocated %6I64u tree nodes (%7I64u bytes total, avg %4I64u bytes per method)\n" , |
1657 | genNodeSizeStats.genTreeNodeCnt, genNodeSizeStats.genTreeNodeSize, |
1658 | genNodeSizeStats.genTreeNodeSize / genMethodCnt); |
1659 | |
1660 | fprintf(fout, "Allocated %7I64u bytes of unused tree node space (%3.2f%%)\n" , |
1661 | genNodeSizeStats.genTreeNodeSize - genNodeSizeStats.genTreeNodeActualSize, |
1662 | (float)(100 * (genNodeSizeStats.genTreeNodeSize - genNodeSizeStats.genTreeNodeActualSize)) / |
1663 | genNodeSizeStats.genTreeNodeSize); |
1664 | |
1665 | fprintf(fout, "\n" ); |
1666 | fprintf(fout, "---------------------------------------------------\n" ); |
1667 | fprintf(fout, "Distribution of per-method GenTree node counts:\n" ); |
1668 | genTreeNcntHist.dump(fout); |
1669 | |
1670 | fprintf(fout, "\n" ); |
1671 | fprintf(fout, "---------------------------------------------------\n" ); |
1672 | fprintf(fout, "Distribution of per-method GenTree node allocations (in bytes):\n" ); |
1673 | genTreeNsizHist.dump(fout); |
1674 | |
1675 | #endif // MEASURE_NODE_SIZE |
1676 | |
1677 | #if MEASURE_BLOCK_SIZE |
1678 | |
1679 | fprintf(fout, "\n" ); |
1680 | fprintf(fout, "---------------------------------------------------\n" ); |
1681 | fprintf(fout, "BasicBlock and flowList/BasicBlockList allocation stats\n" ); |
1682 | fprintf(fout, "---------------------------------------------------\n" ); |
1683 | |
1684 | fprintf(fout, "Allocated %6u basic blocks (%7u bytes total, avg %4u bytes per method)\n" , BasicBlock::s_Count, |
1685 | BasicBlock::s_Size, BasicBlock::s_Size / genMethodCnt); |
1686 | fprintf(fout, "Allocated %6u flow nodes (%7u bytes total, avg %4u bytes per method)\n" , genFlowNodeCnt, |
1687 | genFlowNodeSize, genFlowNodeSize / genMethodCnt); |
1688 | |
1689 | #endif // MEASURE_BLOCK_SIZE |
1690 | |
1691 | #if MEASURE_MEM_ALLOC |
1692 | |
1693 | if (s_dspMemStats) |
1694 | { |
1695 | fprintf(fout, "\nAll allocations:\n" ); |
1696 | ArenaAllocator::dumpAggregateMemStats(jitstdout); |
1697 | |
1698 | fprintf(fout, "\nLargest method:\n" ); |
1699 | ArenaAllocator::dumpMaxMemStats(jitstdout); |
1700 | |
1701 | fprintf(fout, "\n" ); |
1702 | fprintf(fout, "---------------------------------------------------\n" ); |
1703 | fprintf(fout, "Distribution of total memory allocated per method (in KB):\n" ); |
1704 | memAllocHist.dump(fout); |
1705 | |
1706 | fprintf(fout, "\n" ); |
1707 | fprintf(fout, "---------------------------------------------------\n" ); |
1708 | fprintf(fout, "Distribution of total memory used per method (in KB):\n" ); |
1709 | memUsedHist.dump(fout); |
1710 | } |
1711 | |
1712 | #endif // MEASURE_MEM_ALLOC |
1713 | |
1714 | #if LOOP_HOIST_STATS |
1715 | #ifdef DEBUG // Always display loop stats in retail |
1716 | if (JitConfig.DisplayLoopHoistStats() != 0) |
1717 | #endif // DEBUG |
1718 | { |
1719 | PrintAggregateLoopHoistStats(jitstdout); |
1720 | } |
1721 | #endif // LOOP_HOIST_STATS |
1722 | |
1723 | #if MEASURE_PTRTAB_SIZE |
1724 | |
1725 | fprintf(fout, "\n" ); |
1726 | fprintf(fout, "---------------------------------------------------\n" ); |
1727 | fprintf(fout, "GC pointer table stats\n" ); |
1728 | fprintf(fout, "---------------------------------------------------\n" ); |
1729 | |
1730 | fprintf(fout, "Reg pointer descriptor size (internal): %8u (avg %4u per method)\n" , GCInfo::s_gcRegPtrDscSize, |
1731 | GCInfo::s_gcRegPtrDscSize / genMethodCnt); |
1732 | |
1733 | fprintf(fout, "Total pointer table size: %8u (avg %4u per method)\n" , GCInfo::s_gcTotalPtrTabSize, |
1734 | GCInfo::s_gcTotalPtrTabSize / genMethodCnt); |
1735 | |
1736 | #endif // MEASURE_PTRTAB_SIZE |
1737 | |
1738 | #if MEASURE_NODE_SIZE || MEASURE_BLOCK_SIZE || MEASURE_PTRTAB_SIZE || DISPLAY_SIZES |
1739 | |
1740 | if (genMethodCnt != 0) |
1741 | { |
1742 | fprintf(fout, "\n" ); |
1743 | fprintf(fout, "A total of %6u methods compiled" , genMethodCnt); |
1744 | #if DISPLAY_SIZES |
1745 | if (genMethodICnt || genMethodNCnt) |
1746 | { |
1747 | fprintf(fout, " (%u interruptible, %u non-interruptible)" , genMethodICnt, genMethodNCnt); |
1748 | } |
1749 | #endif // DISPLAY_SIZES |
1750 | fprintf(fout, ".\n" ); |
1751 | } |
1752 | |
1753 | #endif // MEASURE_NODE_SIZE || MEASURE_BLOCK_SIZE || MEASURE_PTRTAB_SIZE || DISPLAY_SIZES |
1754 | |
1755 | #if EMITTER_STATS |
1756 | emitterStats(fout); |
1757 | #endif |
1758 | |
1759 | #if MEASURE_FATAL |
1760 | fprintf(fout, "\n" ); |
1761 | fprintf(fout, "---------------------------------------------------\n" ); |
1762 | fprintf(fout, "Fatal errors stats\n" ); |
1763 | fprintf(fout, "---------------------------------------------------\n" ); |
1764 | fprintf(fout, " badCode: %u\n" , fatal_badCode); |
1765 | fprintf(fout, " noWay: %u\n" , fatal_noWay); |
1766 | fprintf(fout, " NOMEM: %u\n" , fatal_NOMEM); |
1767 | fprintf(fout, " noWayAssertBody: %u\n" , fatal_noWayAssertBody); |
1768 | #ifdef DEBUG |
1769 | fprintf(fout, " noWayAssertBodyArgs: %u\n" , fatal_noWayAssertBodyArgs); |
1770 | #endif // DEBUG |
1771 | fprintf(fout, " NYI: %u\n" , fatal_NYI); |
1772 | #endif // MEASURE_FATAL |
1773 | } |
1774 | |
1775 | /***************************************************************************** |
1776 | * Display static data structure sizes. |
1777 | */ |
1778 | |
1779 | /* static */ |
1780 | void Compiler::compDisplayStaticSizes(FILE* fout) |
1781 | { |
1782 | |
1783 | #if MEASURE_NODE_SIZE |
1784 | GenTree::DumpNodeSizes(fout); |
1785 | #endif |
1786 | |
1787 | #if MEASURE_BLOCK_SIZE |
1788 | |
1789 | BasicBlock* bbDummy = nullptr; |
1790 | |
1791 | fprintf(fout, "\n" ); |
1792 | fprintf(fout, "Offset / size of bbNext = %3u / %3u\n" , offsetof(BasicBlock, bbNext), |
1793 | sizeof(bbDummy->bbNext)); |
1794 | fprintf(fout, "Offset / size of bbNum = %3u / %3u\n" , offsetof(BasicBlock, bbNum), |
1795 | sizeof(bbDummy->bbNum)); |
1796 | fprintf(fout, "Offset / size of bbPostOrderNum = %3u / %3u\n" , offsetof(BasicBlock, bbPostOrderNum), |
1797 | sizeof(bbDummy->bbPostOrderNum)); |
1798 | fprintf(fout, "Offset / size of bbRefs = %3u / %3u\n" , offsetof(BasicBlock, bbRefs), |
1799 | sizeof(bbDummy->bbRefs)); |
1800 | fprintf(fout, "Offset / size of bbFlags = %3u / %3u\n" , offsetof(BasicBlock, bbFlags), |
1801 | sizeof(bbDummy->bbFlags)); |
1802 | fprintf(fout, "Offset / size of bbWeight = %3u / %3u\n" , offsetof(BasicBlock, bbWeight), |
1803 | sizeof(bbDummy->bbWeight)); |
1804 | fprintf(fout, "Offset / size of bbJumpKind = %3u / %3u\n" , offsetof(BasicBlock, bbJumpKind), |
1805 | sizeof(bbDummy->bbJumpKind)); |
1806 | fprintf(fout, "Offset / size of bbJumpOffs = %3u / %3u\n" , offsetof(BasicBlock, bbJumpOffs), |
1807 | sizeof(bbDummy->bbJumpOffs)); |
1808 | fprintf(fout, "Offset / size of bbJumpDest = %3u / %3u\n" , offsetof(BasicBlock, bbJumpDest), |
1809 | sizeof(bbDummy->bbJumpDest)); |
1810 | fprintf(fout, "Offset / size of bbJumpSwt = %3u / %3u\n" , offsetof(BasicBlock, bbJumpSwt), |
1811 | sizeof(bbDummy->bbJumpSwt)); |
1812 | fprintf(fout, "Offset / size of bbEntryState = %3u / %3u\n" , offsetof(BasicBlock, bbEntryState), |
1813 | sizeof(bbDummy->bbEntryState)); |
1814 | fprintf(fout, "Offset / size of bbStkTempsIn = %3u / %3u\n" , offsetof(BasicBlock, bbStkTempsIn), |
1815 | sizeof(bbDummy->bbStkTempsIn)); |
1816 | fprintf(fout, "Offset / size of bbStkTempsOut = %3u / %3u\n" , offsetof(BasicBlock, bbStkTempsOut), |
1817 | sizeof(bbDummy->bbStkTempsOut)); |
1818 | fprintf(fout, "Offset / size of bbTryIndex = %3u / %3u\n" , offsetof(BasicBlock, bbTryIndex), |
1819 | sizeof(bbDummy->bbTryIndex)); |
1820 | fprintf(fout, "Offset / size of bbHndIndex = %3u / %3u\n" , offsetof(BasicBlock, bbHndIndex), |
1821 | sizeof(bbDummy->bbHndIndex)); |
1822 | fprintf(fout, "Offset / size of bbCatchTyp = %3u / %3u\n" , offsetof(BasicBlock, bbCatchTyp), |
1823 | sizeof(bbDummy->bbCatchTyp)); |
1824 | fprintf(fout, "Offset / size of bbStkDepth = %3u / %3u\n" , offsetof(BasicBlock, bbStkDepth), |
1825 | sizeof(bbDummy->bbStkDepth)); |
1826 | fprintf(fout, "Offset / size of bbFPinVars = %3u / %3u\n" , offsetof(BasicBlock, bbFPinVars), |
1827 | sizeof(bbDummy->bbFPinVars)); |
1828 | fprintf(fout, "Offset / size of bbPreds = %3u / %3u\n" , offsetof(BasicBlock, bbPreds), |
1829 | sizeof(bbDummy->bbPreds)); |
1830 | fprintf(fout, "Offset / size of bbReach = %3u / %3u\n" , offsetof(BasicBlock, bbReach), |
1831 | sizeof(bbDummy->bbReach)); |
1832 | fprintf(fout, "Offset / size of bbIDom = %3u / %3u\n" , offsetof(BasicBlock, bbIDom), |
1833 | sizeof(bbDummy->bbIDom)); |
1834 | fprintf(fout, "Offset / size of bbDfsNum = %3u / %3u\n" , offsetof(BasicBlock, bbDfsNum), |
1835 | sizeof(bbDummy->bbDfsNum)); |
1836 | fprintf(fout, "Offset / size of bbCodeOffs = %3u / %3u\n" , offsetof(BasicBlock, bbCodeOffs), |
1837 | sizeof(bbDummy->bbCodeOffs)); |
1838 | fprintf(fout, "Offset / size of bbCodeOffsEnd = %3u / %3u\n" , offsetof(BasicBlock, bbCodeOffsEnd), |
1839 | sizeof(bbDummy->bbCodeOffsEnd)); |
1840 | fprintf(fout, "Offset / size of bbVarUse = %3u / %3u\n" , offsetof(BasicBlock, bbVarUse), |
1841 | sizeof(bbDummy->bbVarUse)); |
1842 | fprintf(fout, "Offset / size of bbVarDef = %3u / %3u\n" , offsetof(BasicBlock, bbVarDef), |
1843 | sizeof(bbDummy->bbVarDef)); |
1844 | fprintf(fout, "Offset / size of bbLiveIn = %3u / %3u\n" , offsetof(BasicBlock, bbLiveIn), |
1845 | sizeof(bbDummy->bbLiveIn)); |
1846 | fprintf(fout, "Offset / size of bbLiveOut = %3u / %3u\n" , offsetof(BasicBlock, bbLiveOut), |
1847 | sizeof(bbDummy->bbLiveOut)); |
1848 | fprintf(fout, "Offset / size of bbMemorySsaPhiFunc = %3u / %3u\n" , offsetof(BasicBlock, bbMemorySsaPhiFunc), |
1849 | sizeof(bbDummy->bbMemorySsaPhiFunc)); |
1850 | fprintf(fout, "Offset / size of bbMemorySsaNumIn = %3u / %3u\n" , offsetof(BasicBlock, bbMemorySsaNumIn), |
1851 | sizeof(bbDummy->bbMemorySsaNumIn)); |
1852 | fprintf(fout, "Offset / size of bbMemorySsaNumOut = %3u / %3u\n" , offsetof(BasicBlock, bbMemorySsaNumOut), |
1853 | sizeof(bbDummy->bbMemorySsaNumOut)); |
1854 | fprintf(fout, "Offset / size of bbScope = %3u / %3u\n" , offsetof(BasicBlock, bbScope), |
1855 | sizeof(bbDummy->bbScope)); |
1856 | fprintf(fout, "Offset / size of bbCseGen = %3u / %3u\n" , offsetof(BasicBlock, bbCseGen), |
1857 | sizeof(bbDummy->bbCseGen)); |
1858 | fprintf(fout, "Offset / size of bbCseIn = %3u / %3u\n" , offsetof(BasicBlock, bbCseIn), |
1859 | sizeof(bbDummy->bbCseIn)); |
1860 | fprintf(fout, "Offset / size of bbCseOut = %3u / %3u\n" , offsetof(BasicBlock, bbCseOut), |
1861 | sizeof(bbDummy->bbCseOut)); |
1862 | |
1863 | fprintf(fout, "Offset / size of bbEmitCookie = %3u / %3u\n" , offsetof(BasicBlock, bbEmitCookie), |
1864 | sizeof(bbDummy->bbEmitCookie)); |
1865 | |
1866 | #if FEATURE_EH_FUNCLETS && defined(_TARGET_ARM_) |
1867 | fprintf(fout, "Offset / size of bbUnwindNopEmitCookie = %3u / %3u\n" , offsetof(BasicBlock, bbUnwindNopEmitCookie), |
1868 | sizeof(bbDummy->bbUnwindNopEmitCookie)); |
1869 | #endif // FEATURE_EH_FUNCLETS && defined(_TARGET_ARM_) |
1870 | |
1871 | #ifdef VERIFIER |
1872 | fprintf(fout, "Offset / size of bbStackIn = %3u / %3u\n" , offsetof(BasicBlock, bbStackIn), |
1873 | sizeof(bbDummy->bbStackIn)); |
1874 | fprintf(fout, "Offset / size of bbStackOut = %3u / %3u\n" , offsetof(BasicBlock, bbStackOut), |
1875 | sizeof(bbDummy->bbStackOut)); |
1876 | fprintf(fout, "Offset / size of bbTypesIn = %3u / %3u\n" , offsetof(BasicBlock, bbTypesIn), |
1877 | sizeof(bbDummy->bbTypesIn)); |
1878 | fprintf(fout, "Offset / size of bbTypesOut = %3u / %3u\n" , offsetof(BasicBlock, bbTypesOut), |
1879 | sizeof(bbDummy->bbTypesOut)); |
1880 | #endif // VERIFIER |
1881 | |
1882 | #ifdef DEBUG |
1883 | fprintf(fout, "Offset / size of bbLoopNum = %3u / %3u\n" , offsetof(BasicBlock, bbLoopNum), |
1884 | sizeof(bbDummy->bbLoopNum)); |
1885 | #endif // DEBUG |
1886 | |
1887 | fprintf(fout, "\n" ); |
1888 | fprintf(fout, "Size of BasicBlock = %3u\n" , sizeof(BasicBlock)); |
1889 | |
1890 | #endif // MEASURE_BLOCK_SIZE |
1891 | |
1892 | #if EMITTER_STATS |
1893 | emitterStaticStats(fout); |
1894 | #endif |
1895 | } |
1896 | |
1897 | /***************************************************************************** |
1898 | * |
1899 | * Constructor |
1900 | */ |
1901 | |
1902 | void Compiler::compInit(ArenaAllocator* pAlloc, InlineInfo* inlineInfo) |
1903 | { |
1904 | assert(pAlloc); |
1905 | compArenaAllocator = pAlloc; |
1906 | |
1907 | // Inlinee Compile object will only be allocated when needed for the 1st time. |
1908 | InlineeCompiler = nullptr; |
1909 | |
1910 | // Set the inline info. |
1911 | impInlineInfo = inlineInfo; |
1912 | |
1913 | eeInfoInitialized = false; |
1914 | |
1915 | compDoAggressiveInlining = false; |
1916 | |
1917 | if (compIsForInlining()) |
1918 | { |
1919 | m_inlineStrategy = nullptr; |
1920 | compInlineResult = inlineInfo->inlineResult; |
1921 | } |
1922 | else |
1923 | { |
1924 | m_inlineStrategy = new (this, CMK_Inlining) InlineStrategy(this); |
1925 | compInlineResult = nullptr; |
1926 | } |
1927 | |
1928 | #ifdef FEATURE_TRACELOGGING |
1929 | // Make sure JIT telemetry is initialized as soon as allocations can be made |
1930 | // but no later than a point where noway_asserts can be thrown. |
1931 | // 1. JIT telemetry could allocate some objects internally. |
1932 | // 2. NowayAsserts are tracked through telemetry. |
1933 | // Note: JIT telemetry could gather data when compiler is not fully initialized. |
1934 | // So you have to initialize the compiler variables you use for telemetry. |
1935 | assert((unsigned)PHASE_PRE_IMPORT == 0); |
1936 | previousCompletedPhase = PHASE_PRE_IMPORT; |
1937 | info.compILCodeSize = 0; |
1938 | info.compMethodHnd = nullptr; |
1939 | compJitTelemetry.Initialize(this); |
1940 | #endif |
1941 | |
1942 | #ifdef DEBUG |
1943 | bRangeAllowStress = false; |
1944 | #endif |
1945 | |
1946 | fgInit(); |
1947 | lvaInit(); |
1948 | |
1949 | if (!compIsForInlining()) |
1950 | { |
1951 | codeGen = getCodeGenerator(this); |
1952 | optInit(); |
1953 | hashBv::Init(this); |
1954 | |
1955 | compVarScopeMap = nullptr; |
1956 | |
1957 | // If this method were a real constructor for Compiler, these would |
1958 | // become method initializations. |
1959 | impPendingBlockMembers = JitExpandArray<BYTE>(getAllocator()); |
1960 | impSpillCliquePredMembers = JitExpandArray<BYTE>(getAllocator()); |
1961 | impSpillCliqueSuccMembers = JitExpandArray<BYTE>(getAllocator()); |
1962 | |
1963 | lvMemoryPerSsaData = SsaDefArray<SsaMemDef>(); |
1964 | |
1965 | // |
1966 | // Initialize all the per-method statistics gathering data structures. |
1967 | // |
1968 | |
1969 | optLoopsCloned = 0; |
1970 | |
1971 | #if LOOP_HOIST_STATS |
1972 | m_loopsConsidered = 0; |
1973 | m_curLoopHasHoistedExpression = false; |
1974 | m_loopsWithHoistedExpressions = 0; |
1975 | m_totalHoistedExpressions = 0; |
1976 | #endif // LOOP_HOIST_STATS |
1977 | #if MEASURE_NODE_SIZE |
1978 | genNodeSizeStatsPerFunc.Init(); |
1979 | #endif // MEASURE_NODE_SIZE |
1980 | } |
1981 | else |
1982 | { |
1983 | codeGen = nullptr; |
1984 | } |
1985 | |
1986 | compJmpOpUsed = false; |
1987 | compLongUsed = false; |
1988 | compTailCallUsed = false; |
1989 | compLocallocUsed = false; |
1990 | compLocallocOptimized = false; |
1991 | compQmarkRationalized = false; |
1992 | compQmarkUsed = false; |
1993 | compFloatingPointUsed = false; |
1994 | compUnsafeCastUsed = false; |
1995 | |
1996 | compNeedsGSSecurityCookie = false; |
1997 | compGSReorderStackLayout = false; |
1998 | #if STACK_PROBES |
1999 | compStackProbePrologDone = false; |
2000 | #endif |
2001 | |
2002 | compGeneratingProlog = false; |
2003 | compGeneratingEpilog = false; |
2004 | |
2005 | compLSRADone = false; |
2006 | compRationalIRForm = false; |
2007 | |
2008 | #ifdef DEBUG |
2009 | compCodeGenDone = false; |
2010 | compRegSetCheckLevel = 0; |
2011 | opts.compMinOptsIsUsed = false; |
2012 | #endif |
2013 | opts.compMinOptsIsSet = false; |
2014 | |
2015 | // Used by fgFindJumpTargets for inlining heuristics. |
2016 | opts.instrCount = 0; |
2017 | |
2018 | // Used to track when we should consider running EarlyProp |
2019 | optMethodFlags = 0; |
2020 | |
2021 | #ifdef DEBUG |
2022 | m_nodeTestData = nullptr; |
2023 | m_loopHoistCSEClass = FIRST_LOOP_HOIST_CSE_CLASS; |
2024 | #endif |
2025 | m_switchDescMap = nullptr; |
2026 | m_blockToEHPreds = nullptr; |
2027 | m_fieldSeqStore = nullptr; |
2028 | m_zeroOffsetFieldMap = nullptr; |
2029 | m_arrayInfoMap = nullptr; |
2030 | m_refAnyClass = nullptr; |
2031 | for (MemoryKind memoryKind : allMemoryKinds()) |
2032 | { |
2033 | m_memorySsaMap[memoryKind] = nullptr; |
2034 | } |
2035 | |
2036 | #ifdef DEBUG |
2037 | if (!compIsForInlining()) |
2038 | { |
2039 | compDoComponentUnitTestsOnce(); |
2040 | } |
2041 | #endif // DEBUG |
2042 | |
2043 | vnStore = nullptr; |
2044 | m_opAsgnVarDefSsaNums = nullptr; |
2045 | fgSsaPassesCompleted = 0; |
2046 | fgVNPassesCompleted = 0; |
2047 | |
2048 | // check that HelperCallProperties are initialized |
2049 | |
2050 | assert(s_helperCallProperties.IsPure(CORINFO_HELP_GETSHARED_GCSTATIC_BASE)); |
2051 | assert(!s_helperCallProperties.IsPure(CORINFO_HELP_GETFIELDOBJ)); // quick sanity check |
2052 | |
2053 | // We start with the flow graph in tree-order |
2054 | fgOrder = FGOrderTree; |
2055 | |
2056 | #ifdef FEATURE_SIMD |
2057 | m_simdHandleCache = nullptr; |
2058 | #endif // FEATURE_SIMD |
2059 | |
2060 | compUsesThrowHelper = false; |
2061 | } |
2062 | |
2063 | /***************************************************************************** |
2064 | * |
2065 | * Destructor |
2066 | */ |
2067 | |
2068 | void Compiler::compDone() |
2069 | { |
2070 | } |
2071 | |
2072 | void* Compiler::compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */ |
2073 | void** ppIndirection) /* OUT */ |
2074 | { |
2075 | void* addr; |
2076 | |
2077 | if (info.compMatchedVM) |
2078 | { |
2079 | addr = info.compCompHnd->getHelperFtn(ftnNum, ppIndirection); |
2080 | } |
2081 | else |
2082 | { |
2083 | // If we don't have a matched VM, we won't get valid results when asking for a helper function. |
2084 | addr = UlongToPtr(0xCA11CA11); // "callcall" |
2085 | } |
2086 | |
2087 | return addr; |
2088 | } |
2089 | |
2090 | unsigned Compiler::compGetTypeSize(CorInfoType cit, CORINFO_CLASS_HANDLE clsHnd) |
2091 | { |
2092 | var_types sigType = genActualType(JITtype2varType(cit)); |
2093 | unsigned sigSize; |
2094 | sigSize = genTypeSize(sigType); |
2095 | if (cit == CORINFO_TYPE_VALUECLASS) |
2096 | { |
2097 | sigSize = info.compCompHnd->getClassSize(clsHnd); |
2098 | } |
2099 | else if (cit == CORINFO_TYPE_REFANY) |
2100 | { |
2101 | sigSize = 2 * TARGET_POINTER_SIZE; |
2102 | } |
2103 | return sigSize; |
2104 | } |
2105 | |
2106 | #ifdef DEBUG |
2107 | static bool DidComponentUnitTests = false; |
2108 | |
2109 | void Compiler::compDoComponentUnitTestsOnce() |
2110 | { |
2111 | if (!JitConfig.RunComponentUnitTests()) |
2112 | { |
2113 | return; |
2114 | } |
2115 | |
2116 | if (!DidComponentUnitTests) |
2117 | { |
2118 | DidComponentUnitTests = true; |
2119 | ValueNumStore::RunTests(this); |
2120 | BitSetSupport::TestSuite(getAllocatorDebugOnly()); |
2121 | } |
2122 | } |
2123 | |
2124 | //------------------------------------------------------------------------ |
2125 | // compGetJitDefaultFill: |
2126 | // |
2127 | // Return Value: |
2128 | // An unsigned char value used to initizalize memory allocated by the JIT. |
2129 | // The default value is taken from COMPLUS_JitDefaultFill, if is not set |
2130 | // the value will be 0xdd. When JitStress is active a random value based |
2131 | // on the method hash is used. |
2132 | // |
2133 | // Notes: |
2134 | // Note that we can't use small values like zero, because we have some |
2135 | // asserts that can fire for such values. |
2136 | // |
2137 | unsigned char Compiler::compGetJitDefaultFill() |
2138 | { |
2139 | unsigned char defaultFill = (unsigned char)JitConfig.JitDefaultFill(); |
2140 | |
2141 | if ((this != nullptr) && (compStressCompile(STRESS_GENERIC_VARN, 50))) |
2142 | { |
2143 | unsigned temp; |
2144 | temp = info.compMethodHash(); |
2145 | temp = (temp >> 16) ^ temp; |
2146 | temp = (temp >> 8) ^ temp; |
2147 | temp = temp & 0xff; |
2148 | // asserts like this: assert(!IsUninitialized(stkLvl)); |
2149 | // mean that small values for defaultFill are problematic |
2150 | // so we make the value larger in that case. |
2151 | if (temp < 0x20) |
2152 | { |
2153 | temp |= 0x80; |
2154 | } |
2155 | defaultFill = (unsigned char)temp; |
2156 | } |
2157 | |
2158 | return defaultFill; |
2159 | } |
2160 | |
2161 | #endif // DEBUG |
2162 | |
2163 | /*****************************************************************************/ |
2164 | #ifdef DEBUG |
2165 | /*****************************************************************************/ |
2166 | |
2167 | VarName Compiler::compVarName(regNumber reg, bool isFloatReg) |
2168 | { |
2169 | if (isFloatReg) |
2170 | { |
2171 | assert(genIsValidFloatReg(reg)); |
2172 | } |
2173 | else |
2174 | { |
2175 | assert(genIsValidReg(reg)); |
2176 | } |
2177 | |
2178 | if ((info.compVarScopesCount > 0) && compCurBB && opts.varNames) |
2179 | { |
2180 | unsigned lclNum; |
2181 | LclVarDsc* varDsc; |
2182 | |
2183 | /* Look for the matching register */ |
2184 | for (lclNum = 0, varDsc = lvaTable; lclNum < lvaCount; lclNum++, varDsc++) |
2185 | { |
2186 | /* If the variable is not in a register, or not in the register we're looking for, quit. */ |
2187 | /* Also, if it is a compiler generated variable (i.e. slot# > info.compVarScopesCount), don't bother. */ |
2188 | if ((varDsc->lvRegister != 0) && (varDsc->lvRegNum == reg) && (varDsc->IsFloatRegType() || !isFloatReg) && |
2189 | (varDsc->lvSlotNum < info.compVarScopesCount)) |
2190 | { |
2191 | /* check if variable in that register is live */ |
2192 | if (VarSetOps::IsMember(this, compCurLife, varDsc->lvVarIndex)) |
2193 | { |
2194 | /* variable is live - find the corresponding slot */ |
2195 | VarScopeDsc* varScope = |
2196 | compFindLocalVar(varDsc->lvSlotNum, compCurBB->bbCodeOffs, compCurBB->bbCodeOffsEnd); |
2197 | if (varScope) |
2198 | { |
2199 | return varScope->vsdName; |
2200 | } |
2201 | } |
2202 | } |
2203 | } |
2204 | } |
2205 | |
2206 | return nullptr; |
2207 | } |
2208 | |
2209 | const char* Compiler::compRegVarName(regNumber reg, bool displayVar, bool isFloatReg) |
2210 | { |
2211 | |
2212 | #ifdef _TARGET_ARM_ |
2213 | isFloatReg = genIsValidFloatReg(reg); |
2214 | #endif |
2215 | |
2216 | if (displayVar && (reg != REG_NA)) |
2217 | { |
2218 | VarName varName = compVarName(reg, isFloatReg); |
2219 | |
2220 | if (varName) |
2221 | { |
2222 | const int NAME_VAR_REG_BUFFER_LEN = 4 + 256 + 1; |
2223 | static char nameVarReg[2][NAME_VAR_REG_BUFFER_LEN]; // to avoid overwriting the buffer when have 2 |
2224 | // consecutive calls before printing |
2225 | static int index = 0; // for circular index into the name array |
2226 | |
2227 | index = (index + 1) % 2; // circular reuse of index |
2228 | sprintf_s(nameVarReg[index], NAME_VAR_REG_BUFFER_LEN, "%s'%s'" , getRegName(reg, isFloatReg), |
2229 | VarNameToStr(varName)); |
2230 | |
2231 | return nameVarReg[index]; |
2232 | } |
2233 | } |
2234 | |
2235 | /* no debug info required or no variable in that register |
2236 | -> return standard name */ |
2237 | |
2238 | return getRegName(reg, isFloatReg); |
2239 | } |
2240 | |
2241 | const char* Compiler::compRegNameForSize(regNumber reg, size_t size) |
2242 | { |
2243 | if (size == 0 || size >= 4) |
2244 | { |
2245 | return compRegVarName(reg, true); |
2246 | } |
2247 | |
2248 | // clang-format off |
2249 | static |
2250 | const char * sizeNames[][2] = |
2251 | { |
2252 | { "al" , "ax" }, |
2253 | { "cl" , "cx" }, |
2254 | { "dl" , "dx" }, |
2255 | { "bl" , "bx" }, |
2256 | #ifdef _TARGET_AMD64_ |
2257 | { "spl" , "sp" }, // ESP |
2258 | { "bpl" , "bp" }, // EBP |
2259 | { "sil" , "si" }, // ESI |
2260 | { "dil" , "di" }, // EDI |
2261 | { "r8b" , "r8w" }, |
2262 | { "r9b" , "r9w" }, |
2263 | { "r10b" , "r10w" }, |
2264 | { "r11b" , "r11w" }, |
2265 | { "r12b" , "r12w" }, |
2266 | { "r13b" , "r13w" }, |
2267 | { "r14b" , "r14w" }, |
2268 | { "r15b" , "r15w" }, |
2269 | #endif // _TARGET_AMD64_ |
2270 | }; |
2271 | // clang-format on |
2272 | |
2273 | assert(isByteReg(reg)); |
2274 | assert(genRegMask(reg) & RBM_BYTE_REGS); |
2275 | assert(size == 1 || size == 2); |
2276 | |
2277 | return sizeNames[reg][size - 1]; |
2278 | } |
2279 | |
2280 | const char* Compiler::compFPregVarName(unsigned fpReg, bool displayVar) |
2281 | { |
2282 | const int NAME_VAR_REG_BUFFER_LEN = 4 + 256 + 1; |
2283 | static char nameVarReg[2][NAME_VAR_REG_BUFFER_LEN]; // to avoid overwriting the buffer when have 2 consecutive calls |
2284 | // before printing |
2285 | static int index = 0; // for circular index into the name array |
2286 | |
2287 | index = (index + 1) % 2; // circular reuse of index |
2288 | |
2289 | /* no debug info required or no variable in that register |
2290 | -> return standard name */ |
2291 | |
2292 | sprintf_s(nameVarReg[index], NAME_VAR_REG_BUFFER_LEN, "ST(%d)" , fpReg); |
2293 | return nameVarReg[index]; |
2294 | } |
2295 | |
2296 | const char* Compiler::compLocalVarName(unsigned varNum, unsigned offs) |
2297 | { |
2298 | unsigned i; |
2299 | VarScopeDsc* t; |
2300 | |
2301 | for (i = 0, t = info.compVarScopes; i < info.compVarScopesCount; i++, t++) |
2302 | { |
2303 | if (t->vsdVarNum != varNum) |
2304 | { |
2305 | continue; |
2306 | } |
2307 | |
2308 | if (offs >= t->vsdLifeBeg && offs < t->vsdLifeEnd) |
2309 | { |
2310 | return VarNameToStr(t->vsdName); |
2311 | } |
2312 | } |
2313 | |
2314 | return nullptr; |
2315 | } |
2316 | |
2317 | /*****************************************************************************/ |
2318 | #endif // DEBUG |
2319 | /*****************************************************************************/ |
2320 | |
2321 | void Compiler::compSetProcessor() |
2322 | { |
2323 | // |
2324 | // NOTE: This function needs to be kept in sync with EEJitManager::SetCpuInfo() in vm\codemap.cpp |
2325 | // |
2326 | |
2327 | const JitFlags& jitFlags = *opts.jitFlags; |
2328 | |
2329 | #if defined(_TARGET_ARM_) |
2330 | info.genCPU = CPU_ARM; |
2331 | #elif defined(_TARGET_ARM64_) |
2332 | info.genCPU = CPU_ARM64; |
2333 | #elif defined(_TARGET_AMD64_) |
2334 | info.genCPU = CPU_X64; |
2335 | #elif defined(_TARGET_X86_) |
2336 | if (jitFlags.IsSet(JitFlags::JIT_FLAG_TARGET_P4)) |
2337 | info.genCPU = CPU_X86_PENTIUM_4; |
2338 | else |
2339 | info.genCPU = CPU_X86; |
2340 | #endif |
2341 | |
2342 | // |
2343 | // Processor specific optimizations |
2344 | // |
2345 | CLANG_FORMAT_COMMENT_ANCHOR; |
2346 | |
2347 | #ifdef _TARGET_AMD64_ |
2348 | opts.compUseFCOMI = false; |
2349 | opts.compUseCMOV = true; |
2350 | #elif defined(_TARGET_X86_) |
2351 | opts.compUseFCOMI = jitFlags.IsSet(JitFlags::JIT_FLAG_USE_FCOMI); |
2352 | opts.compUseCMOV = jitFlags.IsSet(JitFlags::JIT_FLAG_USE_CMOV); |
2353 | |
2354 | #ifdef DEBUG |
2355 | if (opts.compUseFCOMI) |
2356 | opts.compUseFCOMI = !compStressCompile(STRESS_USE_FCOMI, 50); |
2357 | if (opts.compUseCMOV) |
2358 | opts.compUseCMOV = !compStressCompile(STRESS_USE_CMOV, 50); |
2359 | #endif // DEBUG |
2360 | |
2361 | #endif // _TARGET_X86_ |
2362 | |
2363 | // Instruction set flags for Intel hardware intrinsics |
2364 | #ifdef _TARGET_XARCH_ |
2365 | opts.compSupportsISA = 0; |
2366 | |
2367 | if (!jitFlags.IsSet(JitFlags::JIT_FLAG_PREJIT)) |
2368 | { |
2369 | #ifdef FEATURE_CORECLR |
2370 | if (JitConfig.EnableHWIntrinsic()) |
2371 | { |
2372 | opts.setSupportedISA(InstructionSet_Base); |
2373 | |
2374 | if (JitConfig.EnableSSE()) |
2375 | { |
2376 | opts.setSupportedISA(InstructionSet_SSE); |
2377 | #ifdef _TARGET_AMD64_ |
2378 | opts.setSupportedISA(InstructionSet_SSE_X64); |
2379 | #endif // _TARGET_AMD64_ |
2380 | |
2381 | if (JitConfig.EnableSSE2()) |
2382 | { |
2383 | opts.setSupportedISA(InstructionSet_SSE2); |
2384 | #ifdef _TARGET_AMD64_ |
2385 | opts.setSupportedISA(InstructionSet_SSE2_X64); |
2386 | #endif // _TARGET_AMD64_ |
2387 | |
2388 | if (jitFlags.IsSet(JitFlags::JIT_FLAG_USE_AES) && JitConfig.EnableAES()) |
2389 | { |
2390 | opts.setSupportedISA(InstructionSet_AES); |
2391 | } |
2392 | |
2393 | if (jitFlags.IsSet(JitFlags::JIT_FLAG_USE_PCLMULQDQ) && JitConfig.EnablePCLMULQDQ()) |
2394 | { |
2395 | opts.setSupportedISA(InstructionSet_PCLMULQDQ); |
2396 | } |
2397 | |
2398 | // We need to additionaly check that COMPlus_EnableSSE3_4 is set, as that |
2399 | // is a prexisting config flag that controls the SSE3+ ISAs |
2400 | if (jitFlags.IsSet(JitFlags::JIT_FLAG_USE_SSE3) && JitConfig.EnableSSE3() && |
2401 | JitConfig.EnableSSE3_4()) |
2402 | { |
2403 | opts.setSupportedISA(InstructionSet_SSE3); |
2404 | |
2405 | if (jitFlags.IsSet(JitFlags::JIT_FLAG_USE_SSSE3) && JitConfig.EnableSSSE3()) |
2406 | { |
2407 | opts.setSupportedISA(InstructionSet_SSSE3); |
2408 | |
2409 | if (jitFlags.IsSet(JitFlags::JIT_FLAG_USE_SSE41) && JitConfig.EnableSSE41()) |
2410 | { |
2411 | opts.setSupportedISA(InstructionSet_SSE41); |
2412 | #ifdef _TARGET_AMD64_ |
2413 | opts.setSupportedISA(InstructionSet_SSE41_X64); |
2414 | #endif // _TARGET_AMD64_ |
2415 | |
2416 | if (jitFlags.IsSet(JitFlags::JIT_FLAG_USE_SSE42) && JitConfig.EnableSSE42()) |
2417 | { |
2418 | opts.setSupportedISA(InstructionSet_SSE42); |
2419 | #ifdef _TARGET_AMD64_ |
2420 | opts.setSupportedISA(InstructionSet_SSE42_X64); |
2421 | #endif // _TARGET_AMD64_ |
2422 | |
2423 | if (jitFlags.IsSet(JitFlags::JIT_FLAG_USE_POPCNT) && JitConfig.EnablePOPCNT()) |
2424 | { |
2425 | opts.setSupportedISA(InstructionSet_POPCNT); |
2426 | #ifdef _TARGET_AMD64_ |
2427 | opts.setSupportedISA(InstructionSet_POPCNT_X64); |
2428 | #endif // _TARGET_AMD64_ |
2429 | } |
2430 | |
2431 | if (jitFlags.IsSet(JitFlags::JIT_FLAG_USE_AVX) && JitConfig.EnableAVX()) |
2432 | { |
2433 | opts.setSupportedISA(InstructionSet_AVX); |
2434 | |
2435 | if (jitFlags.IsSet(JitFlags::JIT_FLAG_USE_FMA) && JitConfig.EnableFMA()) |
2436 | { |
2437 | opts.setSupportedISA(InstructionSet_FMA); |
2438 | } |
2439 | |
2440 | if (jitFlags.IsSet(JitFlags::JIT_FLAG_USE_AVX2) && JitConfig.EnableAVX2()) |
2441 | { |
2442 | opts.setSupportedISA(InstructionSet_AVX2); |
2443 | } |
2444 | } |
2445 | } |
2446 | } |
2447 | } |
2448 | } |
2449 | } |
2450 | } |
2451 | |
2452 | if (jitFlags.IsSet(JitFlags::JIT_FLAG_USE_LZCNT) && JitConfig.EnableLZCNT()) |
2453 | { |
2454 | opts.setSupportedISA(InstructionSet_LZCNT); |
2455 | #ifdef _TARGET_AMD64_ |
2456 | opts.setSupportedISA(InstructionSet_LZCNT_X64); |
2457 | #endif // _TARGET_AMD64_ |
2458 | } |
2459 | |
2460 | // We currently need to also check that AVX is supported as that controls the support for the VEX encoding |
2461 | // in the emitter. |
2462 | if (jitFlags.IsSet(JitFlags::JIT_FLAG_USE_BMI1) && JitConfig.EnableBMI1() && |
2463 | compSupports(InstructionSet_AVX)) |
2464 | { |
2465 | opts.setSupportedISA(InstructionSet_BMI1); |
2466 | #ifdef _TARGET_AMD64_ |
2467 | opts.setSupportedISA(InstructionSet_BMI1_X64); |
2468 | #endif // _TARGET_AMD64_ |
2469 | } |
2470 | |
2471 | // We currently need to also check that AVX is supported as that controls the support for the VEX encoding |
2472 | // in the emitter. |
2473 | if (jitFlags.IsSet(JitFlags::JIT_FLAG_USE_BMI2) && JitConfig.EnableBMI2() && |
2474 | compSupports(InstructionSet_AVX)) |
2475 | { |
2476 | opts.setSupportedISA(InstructionSet_BMI2); |
2477 | #ifdef _TARGET_AMD64_ |
2478 | opts.setSupportedISA(InstructionSet_BMI2_X64); |
2479 | #endif // _TARGET_AMD64_ |
2480 | } |
2481 | } |
2482 | #else // !FEATURE_CORECLR |
2483 | // If this is not FEATURE_CORECLR, the only flags supported by the VM are AVX and AVX2. |
2484 | // Furthermore, the only two configurations supported by the desktop JIT are SSE2 and AVX2, |
2485 | // so if the latter is set, we also check all the in-between options. |
2486 | // Note that the EnableSSE2 and EnableSSE flags are only checked by HW Intrinsic code, |
2487 | // so the System.Numerics.Vector support doesn't depend on those flags. |
2488 | // However, if any of these are disabled, we will not enable AVX2. |
2489 | // |
2490 | if (jitFlags.IsSet(JitFlags::JIT_FLAG_USE_AVX) && jitFlags.IsSet(JitFlags::JIT_FLAG_USE_AVX2) && |
2491 | (JitConfig.EnableAVX2() != 0) && (JitConfig.EnableAVX() != 0) && (JitConfig.EnableSSE42() != 0) && |
2492 | (JitConfig.EnableSSE41() != 0) && (JitConfig.EnableSSSE3() != 0) && (JitConfig.EnableSSE3() != 0) && |
2493 | (JitConfig.EnableSSE2() != 0) && (JitConfig.EnableSSE() != 0) && (JitConfig.EnableSSE3_4() != 0)) |
2494 | { |
2495 | opts.setSupportedISA(InstructionSet_SSE); |
2496 | opts.setSupportedISA(InstructionSet_SSE2); |
2497 | opts.setSupportedISA(InstructionSet_SSE3); |
2498 | opts.setSupportedISA(InstructionSet_SSSE3); |
2499 | opts.setSupportedISA(InstructionSet_SSE41); |
2500 | opts.setSupportedISA(InstructionSet_SSE42); |
2501 | opts.setSupportedISA(InstructionSet_AVX); |
2502 | opts.setSupportedISA(InstructionSet_AVX2); |
2503 | } |
2504 | #endif // !FEATURE_CORECLR |
2505 | } |
2506 | |
2507 | if (!compIsForInlining()) |
2508 | { |
2509 | if (canUseVexEncoding()) |
2510 | { |
2511 | codeGen->getEmitter()->SetUseVEXEncoding(true); |
2512 | // Assume each JITted method does not contain AVX instruction at first |
2513 | codeGen->getEmitter()->SetContainsAVX(false); |
2514 | codeGen->getEmitter()->SetContains256bitAVX(false); |
2515 | } |
2516 | } |
2517 | #endif // _TARGET_XARCH_ |
2518 | |
2519 | #if defined(_TARGET_ARM64_) |
2520 | // There is no JitFlag for Base instructions handle manually |
2521 | opts.setSupportedISA(InstructionSet_Base); |
2522 | #define HARDWARE_INTRINSIC_CLASS(flag, isa) \ |
2523 | if (jitFlags.IsSet(JitFlags::flag)) \ |
2524 | opts.setSupportedISA(InstructionSet_##isa); |
2525 | #include "hwintrinsiclistArm64.h" |
2526 | |
2527 | #endif |
2528 | } |
2529 | |
2530 | #ifdef PROFILING_SUPPORTED |
2531 | // A Dummy routine to receive Enter/Leave/Tailcall profiler callbacks. |
2532 | // These are used when complus_JitEltHookEnabled=1 |
2533 | #ifdef _TARGET_AMD64_ |
2534 | void DummyProfilerELTStub(UINT_PTR ProfilerHandle, UINT_PTR callerSP) |
2535 | { |
2536 | return; |
2537 | } |
2538 | #else //! _TARGET_AMD64_ |
2539 | void DummyProfilerELTStub(UINT_PTR ProfilerHandle) |
2540 | { |
2541 | return; |
2542 | } |
2543 | #endif //!_TARGET_AMD64_ |
2544 | |
2545 | #endif // PROFILING_SUPPORTED |
2546 | |
2547 | bool Compiler::compIsFullTrust() |
2548 | { |
2549 | return (info.compCompHnd->canSkipMethodVerification(info.compMethodHnd) == CORINFO_VERIFICATION_CAN_SKIP); |
2550 | } |
2551 | |
2552 | bool Compiler::compShouldThrowOnNoway( |
2553 | #ifdef FEATURE_TRACELOGGING |
2554 | const char* filename, unsigned line |
2555 | #endif |
2556 | ) |
2557 | { |
2558 | #ifdef FEATURE_TRACELOGGING |
2559 | compJitTelemetry.NotifyNowayAssert(filename, line); |
2560 | #endif |
2561 | |
2562 | // In min opts, we don't want the noway assert to go through the exception |
2563 | // path. Instead we want it to just silently go through codegen for |
2564 | // compat reasons. |
2565 | // If we are not in full trust, we should always fire for security. |
2566 | return !opts.MinOpts() || !compIsFullTrust(); |
2567 | } |
2568 | |
2569 | // ConfigInteger does not offer an option for decimal flags. Any numbers are interpreted as hex. |
2570 | // I could add the decimal option to ConfigInteger or I could write a function to reinterpret this |
2571 | // value as the user intended. |
2572 | unsigned ReinterpretHexAsDecimal(unsigned in) |
2573 | { |
2574 | // ex: in: 0x100 returns: 100 |
2575 | unsigned result = 0; |
2576 | unsigned index = 1; |
2577 | |
2578 | // default value |
2579 | if (in == INT_MAX) |
2580 | { |
2581 | return in; |
2582 | } |
2583 | |
2584 | while (in) |
2585 | { |
2586 | unsigned digit = in % 16; |
2587 | in >>= 4; |
2588 | assert(digit < 10); |
2589 | result += digit * index; |
2590 | index *= 10; |
2591 | } |
2592 | return result; |
2593 | } |
2594 | |
2595 | void Compiler::compInitOptions(JitFlags* jitFlags) |
2596 | { |
2597 | #ifdef UNIX_AMD64_ABI |
2598 | opts.compNeedToAlignFrame = false; |
2599 | #endif // UNIX_AMD64_ABI |
2600 | memset(&opts, 0, sizeof(opts)); |
2601 | |
2602 | if (compIsForInlining()) |
2603 | { |
2604 | // The following flags are lost when inlining. (They are removed in |
2605 | // Compiler::fgInvokeInlineeCompiler().) |
2606 | assert(!jitFlags->IsSet(JitFlags::JIT_FLAG_BBOPT)); |
2607 | assert(!jitFlags->IsSet(JitFlags::JIT_FLAG_BBINSTR)); |
2608 | assert(!jitFlags->IsSet(JitFlags::JIT_FLAG_PROF_ENTERLEAVE)); |
2609 | assert(!jitFlags->IsSet(JitFlags::JIT_FLAG_DEBUG_EnC)); |
2610 | assert(!jitFlags->IsSet(JitFlags::JIT_FLAG_DEBUG_INFO)); |
2611 | |
2612 | assert(jitFlags->IsSet(JitFlags::JIT_FLAG_SKIP_VERIFICATION)); |
2613 | } |
2614 | |
2615 | opts.jitFlags = jitFlags; |
2616 | opts.compFlags = CLFLG_MAXOPT; // Default value is for full optimization |
2617 | |
2618 | if (jitFlags->IsSet(JitFlags::JIT_FLAG_DEBUG_CODE) || jitFlags->IsSet(JitFlags::JIT_FLAG_MIN_OPT) || |
2619 | jitFlags->IsSet(JitFlags::JIT_FLAG_TIER0)) |
2620 | { |
2621 | opts.compFlags = CLFLG_MINOPT; |
2622 | } |
2623 | // Don't optimize .cctors (except prejit) or if we're an inlinee |
2624 | else if (!jitFlags->IsSet(JitFlags::JIT_FLAG_PREJIT) && ((info.compFlags & FLG_CCTOR) == FLG_CCTOR) && |
2625 | !compIsForInlining()) |
2626 | { |
2627 | opts.compFlags = CLFLG_MINOPT; |
2628 | } |
2629 | |
2630 | // Default value is to generate a blend of size and speed optimizations |
2631 | // |
2632 | opts.compCodeOpt = BLENDED_CODE; |
2633 | |
2634 | // If the EE sets SIZE_OPT or if we are compiling a Class constructor |
2635 | // we will optimize for code size at the expense of speed |
2636 | // |
2637 | if (jitFlags->IsSet(JitFlags::JIT_FLAG_SIZE_OPT) || ((info.compFlags & FLG_CCTOR) == FLG_CCTOR)) |
2638 | { |
2639 | opts.compCodeOpt = SMALL_CODE; |
2640 | } |
2641 | // |
2642 | // If the EE sets SPEED_OPT we will optimize for speed at the expense of code size |
2643 | // |
2644 | else if (jitFlags->IsSet(JitFlags::JIT_FLAG_SPEED_OPT) || |
2645 | (jitFlags->IsSet(JitFlags::JIT_FLAG_TIER1) && !jitFlags->IsSet(JitFlags::JIT_FLAG_MIN_OPT))) |
2646 | { |
2647 | opts.compCodeOpt = FAST_CODE; |
2648 | assert(!jitFlags->IsSet(JitFlags::JIT_FLAG_SIZE_OPT)); |
2649 | } |
2650 | |
2651 | //------------------------------------------------------------------------- |
2652 | |
2653 | opts.compDbgCode = jitFlags->IsSet(JitFlags::JIT_FLAG_DEBUG_CODE); |
2654 | opts.compDbgInfo = jitFlags->IsSet(JitFlags::JIT_FLAG_DEBUG_INFO); |
2655 | opts.compDbgEnC = jitFlags->IsSet(JitFlags::JIT_FLAG_DEBUG_EnC); |
2656 | |
2657 | #if REGEN_SHORTCUTS || REGEN_CALLPAT |
2658 | // We never want to have debugging enabled when regenerating GC encoding patterns |
2659 | opts.compDbgCode = false; |
2660 | opts.compDbgInfo = false; |
2661 | opts.compDbgEnC = false; |
2662 | #endif |
2663 | |
2664 | compSetProcessor(); |
2665 | |
2666 | #ifdef DEBUG |
2667 | opts.dspOrder = false; |
2668 | if (compIsForInlining()) |
2669 | { |
2670 | verbose = impInlineInfo->InlinerCompiler->verbose; |
2671 | } |
2672 | else |
2673 | { |
2674 | verbose = false; |
2675 | codeGen->setVerbose(false); |
2676 | } |
2677 | verboseTrees = verbose && shouldUseVerboseTrees(); |
2678 | verboseSsa = verbose && shouldUseVerboseSsa(); |
2679 | asciiTrees = shouldDumpASCIITrees(); |
2680 | opts.dspDiffable = compIsForInlining() ? impInlineInfo->InlinerCompiler->opts.dspDiffable : false; |
2681 | #endif |
2682 | |
2683 | opts.compNeedSecurityCheck = false; |
2684 | opts.altJit = false; |
2685 | |
2686 | #if defined(LATE_DISASM) && !defined(DEBUG) |
2687 | // For non-debug builds with the late disassembler built in, we currently always do late disassembly |
2688 | // (we have no way to determine when not to, since we don't have class/method names). |
2689 | // In the DEBUG case, this is initialized to false, below. |
2690 | opts.doLateDisasm = true; |
2691 | #endif |
2692 | |
2693 | #ifdef DEBUG |
2694 | |
2695 | const JitConfigValues::MethodSet* pfAltJit; |
2696 | if (jitFlags->IsSet(JitFlags::JIT_FLAG_PREJIT)) |
2697 | { |
2698 | pfAltJit = &JitConfig.AltJitNgen(); |
2699 | } |
2700 | else |
2701 | { |
2702 | pfAltJit = &JitConfig.AltJit(); |
2703 | } |
2704 | |
2705 | #ifdef ALT_JIT |
2706 | if (pfAltJit->contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
2707 | { |
2708 | opts.altJit = true; |
2709 | } |
2710 | |
2711 | unsigned altJitLimit = ReinterpretHexAsDecimal(JitConfig.AltJitLimit()); |
2712 | if (altJitLimit > 0 && Compiler::jitTotalMethodCompiled >= altJitLimit) |
2713 | { |
2714 | opts.altJit = false; |
2715 | } |
2716 | #endif // ALT_JIT |
2717 | |
2718 | #else // !DEBUG |
2719 | |
2720 | const char* altJitVal; |
2721 | if (jitFlags->IsSet(JitFlags::JIT_FLAG_PREJIT)) |
2722 | { |
2723 | altJitVal = JitConfig.AltJitNgen().list(); |
2724 | } |
2725 | else |
2726 | { |
2727 | altJitVal = JitConfig.AltJit().list(); |
2728 | } |
2729 | |
2730 | #ifdef ALT_JIT |
2731 | // In release mode, you either get all methods or no methods. You must use "*" as the parameter, or we ignore it. |
2732 | // You don't get to give a regular expression of methods to match. |
2733 | // (Partially, this is because we haven't computed and stored the method and class name except in debug, and it |
2734 | // might be expensive to do so.) |
2735 | if ((altJitVal != nullptr) && (strcmp(altJitVal, "*" ) == 0)) |
2736 | { |
2737 | opts.altJit = true; |
2738 | } |
2739 | #endif // ALT_JIT |
2740 | |
2741 | #endif // !DEBUG |
2742 | |
2743 | #ifdef ALT_JIT |
2744 | // Take care of COMPlus_AltJitExcludeAssemblies. |
2745 | if (opts.altJit) |
2746 | { |
2747 | // First, initialize the AltJitExcludeAssemblies list, but only do it once. |
2748 | if (!s_pAltJitExcludeAssembliesListInitialized) |
2749 | { |
2750 | const wchar_t* wszAltJitExcludeAssemblyList = JitConfig.AltJitExcludeAssemblies(); |
2751 | if (wszAltJitExcludeAssemblyList != nullptr) |
2752 | { |
2753 | // NOTE: The Assembly name list is allocated in the process heap, not in the no-release heap, which is |
2754 | // reclaimed |
2755 | // for every compilation. This is ok because we only allocate once, due to the static. |
2756 | s_pAltJitExcludeAssembliesList = new (HostAllocator::getHostAllocator()) |
2757 | AssemblyNamesList2(wszAltJitExcludeAssemblyList, HostAllocator::getHostAllocator()); |
2758 | } |
2759 | s_pAltJitExcludeAssembliesListInitialized = true; |
2760 | } |
2761 | |
2762 | if (s_pAltJitExcludeAssembliesList != nullptr) |
2763 | { |
2764 | // We have an exclusion list. See if this method is in an assembly that is on the list. |
2765 | // Note that we check this for every method, since we might inline across modules, and |
2766 | // if the inlinee module is on the list, we don't want to use the altjit for it. |
2767 | const char* methodAssemblyName = info.compCompHnd->getAssemblyName( |
2768 | info.compCompHnd->getModuleAssembly(info.compCompHnd->getClassModule(info.compClassHnd))); |
2769 | if (s_pAltJitExcludeAssembliesList->IsInList(methodAssemblyName)) |
2770 | { |
2771 | opts.altJit = false; |
2772 | } |
2773 | } |
2774 | } |
2775 | #endif // ALT_JIT |
2776 | |
2777 | #ifdef DEBUG |
2778 | |
2779 | bool altJitConfig = !pfAltJit->isEmpty(); |
2780 | |
2781 | // If we have a non-empty AltJit config then we change all of these other |
2782 | // config values to refer only to the AltJit. Otherwise, a lot of COMPlus_* variables |
2783 | // would apply to both the altjit and the normal JIT, but we only care about |
2784 | // debugging the altjit if the COMPlus_AltJit configuration is set. |
2785 | // |
2786 | if (compIsForImportOnly() && (!altJitConfig || opts.altJit)) |
2787 | { |
2788 | if (JitConfig.JitImportBreak().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
2789 | { |
2790 | assert(!"JitImportBreak reached" ); |
2791 | } |
2792 | } |
2793 | |
2794 | bool verboseDump = false; |
2795 | bool dumpIR = false; |
2796 | bool dumpIRTypes = false; |
2797 | bool dumpIRLocals = false; |
2798 | bool dumpIRRegs = false; |
2799 | bool = false; |
2800 | bool dumpIRValnums = false; |
2801 | bool dumpIRCosts = false; |
2802 | bool dumpIRFlags = false; |
2803 | bool dumpIRKinds = false; |
2804 | bool dumpIRNodes = false; |
2805 | bool dumpIRNoLists = false; |
2806 | bool dumpIRNoLeafs = false; |
2807 | bool dumpIRNoStmts = false; |
2808 | bool dumpIRTrees = false; |
2809 | bool dumpIRLinear = false; |
2810 | bool dumpIRDataflow = false; |
2811 | bool = false; |
2812 | bool dumpIRExit = false; |
2813 | LPCWSTR dumpIRPhase = nullptr; |
2814 | LPCWSTR dumpIRFormat = nullptr; |
2815 | |
2816 | if (!altJitConfig || opts.altJit) |
2817 | { |
2818 | LPCWSTR dumpIRFormat = nullptr; |
2819 | |
2820 | // We should only enable 'verboseDump' when we are actually compiling a matching method |
2821 | // and not enable it when we are just considering inlining a matching method. |
2822 | // |
2823 | if (!compIsForInlining()) |
2824 | { |
2825 | if (jitFlags->IsSet(JitFlags::JIT_FLAG_PREJIT)) |
2826 | { |
2827 | if (JitConfig.NgenDump().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
2828 | { |
2829 | verboseDump = true; |
2830 | } |
2831 | unsigned ngenHashDumpVal = (unsigned)JitConfig.NgenHashDump(); |
2832 | if ((ngenHashDumpVal != (DWORD)-1) && (ngenHashDumpVal == info.compMethodHash())) |
2833 | { |
2834 | verboseDump = true; |
2835 | } |
2836 | if (JitConfig.NgenDumpIR().contains(info.compMethodName, info.compClassName, |
2837 | &info.compMethodInfo->args)) |
2838 | { |
2839 | dumpIR = true; |
2840 | } |
2841 | unsigned ngenHashDumpIRVal = (unsigned)JitConfig.NgenHashDumpIR(); |
2842 | if ((ngenHashDumpIRVal != (DWORD)-1) && (ngenHashDumpIRVal == info.compMethodHash())) |
2843 | { |
2844 | dumpIR = true; |
2845 | } |
2846 | dumpIRFormat = JitConfig.NgenDumpIRFormat(); |
2847 | dumpIRPhase = JitConfig.NgenDumpIRPhase(); |
2848 | } |
2849 | else |
2850 | { |
2851 | if (JitConfig.JitDump().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
2852 | { |
2853 | verboseDump = true; |
2854 | } |
2855 | unsigned jitHashDumpVal = (unsigned)JitConfig.JitHashDump(); |
2856 | if ((jitHashDumpVal != (DWORD)-1) && (jitHashDumpVal == info.compMethodHash())) |
2857 | { |
2858 | verboseDump = true; |
2859 | } |
2860 | if (JitConfig.JitDumpIR().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
2861 | { |
2862 | dumpIR = true; |
2863 | } |
2864 | unsigned jitHashDumpIRVal = (unsigned)JitConfig.JitHashDumpIR(); |
2865 | if ((jitHashDumpIRVal != (DWORD)-1) && (jitHashDumpIRVal == info.compMethodHash())) |
2866 | { |
2867 | dumpIR = true; |
2868 | } |
2869 | dumpIRFormat = JitConfig.JitDumpIRFormat(); |
2870 | dumpIRPhase = JitConfig.JitDumpIRPhase(); |
2871 | } |
2872 | } |
2873 | |
2874 | if (dumpIRPhase == nullptr) |
2875 | { |
2876 | dumpIRPhase = W("*" ); |
2877 | } |
2878 | |
2879 | this->dumpIRPhase = dumpIRPhase; |
2880 | |
2881 | if (dumpIRFormat != nullptr) |
2882 | { |
2883 | this->dumpIRFormat = dumpIRFormat; |
2884 | } |
2885 | |
2886 | dumpIRTrees = false; |
2887 | dumpIRLinear = true; |
2888 | if (dumpIRFormat != nullptr) |
2889 | { |
2890 | for (LPCWSTR p = dumpIRFormat; (*p != 0);) |
2891 | { |
2892 | for (; (*p != 0); p++) |
2893 | { |
2894 | if (*p != L' ') |
2895 | { |
2896 | break; |
2897 | } |
2898 | } |
2899 | |
2900 | if (*p == 0) |
2901 | { |
2902 | break; |
2903 | } |
2904 | |
2905 | static bool dumpedHelp = false; |
2906 | |
2907 | if ((*p == L'?') && (!dumpedHelp)) |
2908 | { |
2909 | printf("*******************************************************************************\n" ); |
2910 | printf("\n" ); |
2911 | dFormatIR(); |
2912 | printf("\n" ); |
2913 | printf("\n" ); |
2914 | printf("Available specifiers (comma separated):\n" ); |
2915 | printf("\n" ); |
2916 | printf("? dump out value of COMPlus_JitDumpIRFormat and this list of values\n" ); |
2917 | printf("\n" ); |
2918 | printf("linear linear IR dump (default)\n" ); |
2919 | printf("tree tree IR dump (traditional)\n" ); |
2920 | printf("mixed intermingle tree dump with linear IR dump\n" ); |
2921 | printf("\n" ); |
2922 | printf("dataflow use data flow form of linear IR dump\n" ); |
2923 | printf("structural use structural form of linear IR dump\n" ); |
2924 | printf("all implies structural, include everything\n" ); |
2925 | printf("\n" ); |
2926 | printf("kinds include tree node kinds in dump, example: \"kinds=[LEAF][LOCAL]\"\n" ); |
2927 | printf("flags include tree node flags in dump, example: \"flags=[CALL][GLOB_REF]\" \n" ); |
2928 | printf("types includes tree node types in dump, example: \".int\"\n" ); |
2929 | printf("locals include local numbers and tracking numbers in dump, example: \"(V3,T1)\"\n" ); |
2930 | printf("regs include register assignments in dump, example: \"(rdx)\"\n" ); |
2931 | printf("ssa include SSA numbers in dump, example: \"<d:3>\" or \"<u:3>\"\n" ); |
2932 | printf("valnums include Value numbers in dump, example: \"<v:$c4>\" or \"<v:$c4,$c5>\"\n" ); |
2933 | printf("\n" ); |
2934 | printf("nolist exclude GT_LIST nodes from dump\n" ); |
2935 | printf("noleafs exclude LEAF nodes from dump (fold into operations)\n" ); |
2936 | printf("nostmts exclude GT_STMTS from dump (unless required by dependencies)\n" ); |
2937 | printf("\n" ); |
2938 | printf("blkhdrs include block headers\n" ); |
2939 | printf("exit exit program after last phase dump (used with single method)\n" ); |
2940 | printf("\n" ); |
2941 | printf("*******************************************************************************\n" ); |
2942 | dumpedHelp = true; |
2943 | } |
2944 | |
2945 | if (wcsncmp(p, W("types" ), 5) == 0) |
2946 | { |
2947 | dumpIRTypes = true; |
2948 | } |
2949 | |
2950 | if (wcsncmp(p, W("locals" ), 6) == 0) |
2951 | { |
2952 | dumpIRLocals = true; |
2953 | } |
2954 | |
2955 | if (wcsncmp(p, W("regs" ), 4) == 0) |
2956 | { |
2957 | dumpIRRegs = true; |
2958 | } |
2959 | |
2960 | if (wcsncmp(p, W("ssa" ), 3) == 0) |
2961 | { |
2962 | dumpIRSsa = true; |
2963 | } |
2964 | |
2965 | if (wcsncmp(p, W("valnums" ), 7) == 0) |
2966 | { |
2967 | dumpIRValnums = true; |
2968 | } |
2969 | |
2970 | if (wcsncmp(p, W("costs" ), 5) == 0) |
2971 | { |
2972 | dumpIRCosts = true; |
2973 | } |
2974 | |
2975 | if (wcsncmp(p, W("flags" ), 5) == 0) |
2976 | { |
2977 | dumpIRFlags = true; |
2978 | } |
2979 | |
2980 | if (wcsncmp(p, W("kinds" ), 5) == 0) |
2981 | { |
2982 | dumpIRKinds = true; |
2983 | } |
2984 | |
2985 | if (wcsncmp(p, W("nodes" ), 5) == 0) |
2986 | { |
2987 | dumpIRNodes = true; |
2988 | } |
2989 | |
2990 | if (wcsncmp(p, W("exit" ), 4) == 0) |
2991 | { |
2992 | dumpIRExit = true; |
2993 | } |
2994 | |
2995 | if (wcsncmp(p, W("nolists" ), 7) == 0) |
2996 | { |
2997 | dumpIRNoLists = true; |
2998 | } |
2999 | |
3000 | if (wcsncmp(p, W("noleafs" ), 7) == 0) |
3001 | { |
3002 | dumpIRNoLeafs = true; |
3003 | } |
3004 | |
3005 | if (wcsncmp(p, W("nostmts" ), 7) == 0) |
3006 | { |
3007 | dumpIRNoStmts = true; |
3008 | } |
3009 | |
3010 | if (wcsncmp(p, W("trees" ), 5) == 0) |
3011 | { |
3012 | dumpIRTrees = true; |
3013 | dumpIRLinear = false; |
3014 | } |
3015 | |
3016 | if (wcsncmp(p, W("structural" ), 10) == 0) |
3017 | { |
3018 | dumpIRLinear = true; |
3019 | dumpIRNoStmts = false; |
3020 | dumpIRNoLeafs = false; |
3021 | dumpIRNoLists = false; |
3022 | } |
3023 | |
3024 | if (wcsncmp(p, W("all" ), 3) == 0) |
3025 | { |
3026 | dumpIRLinear = true; |
3027 | dumpIRKinds = true; |
3028 | dumpIRFlags = true; |
3029 | dumpIRTypes = true; |
3030 | dumpIRLocals = true; |
3031 | dumpIRRegs = true; |
3032 | dumpIRSsa = true; |
3033 | dumpIRValnums = true; |
3034 | dumpIRCosts = true; |
3035 | dumpIRNoStmts = false; |
3036 | dumpIRNoLeafs = false; |
3037 | dumpIRNoLists = false; |
3038 | } |
3039 | |
3040 | if (wcsncmp(p, W("linear" ), 6) == 0) |
3041 | { |
3042 | dumpIRTrees = false; |
3043 | dumpIRLinear = true; |
3044 | } |
3045 | |
3046 | if (wcsncmp(p, W("mixed" ), 5) == 0) |
3047 | { |
3048 | dumpIRTrees = true; |
3049 | dumpIRLinear = true; |
3050 | } |
3051 | |
3052 | if (wcsncmp(p, W("dataflow" ), 8) == 0) |
3053 | { |
3054 | dumpIRDataflow = true; |
3055 | dumpIRNoLeafs = true; |
3056 | dumpIRNoLists = true; |
3057 | dumpIRNoStmts = true; |
3058 | } |
3059 | |
3060 | if (wcsncmp(p, W("blkhdrs" ), 7) == 0) |
3061 | { |
3062 | dumpIRBlockHeaders = true; |
3063 | } |
3064 | |
3065 | for (; (*p != 0); p++) |
3066 | { |
3067 | if (*p == L',') |
3068 | { |
3069 | p++; |
3070 | break; |
3071 | } |
3072 | } |
3073 | } |
3074 | } |
3075 | } |
3076 | |
3077 | if (verboseDump) |
3078 | { |
3079 | verbose = true; |
3080 | } |
3081 | |
3082 | if (dumpIR) |
3083 | { |
3084 | this->dumpIR = true; |
3085 | } |
3086 | |
3087 | if (dumpIRTypes) |
3088 | { |
3089 | this->dumpIRTypes = true; |
3090 | } |
3091 | |
3092 | if (dumpIRLocals) |
3093 | { |
3094 | this->dumpIRLocals = true; |
3095 | } |
3096 | |
3097 | if (dumpIRRegs) |
3098 | { |
3099 | this->dumpIRRegs = true; |
3100 | } |
3101 | |
3102 | if (dumpIRSsa) |
3103 | { |
3104 | this->dumpIRSsa = true; |
3105 | } |
3106 | |
3107 | if (dumpIRValnums) |
3108 | { |
3109 | this->dumpIRValnums = true; |
3110 | } |
3111 | |
3112 | if (dumpIRCosts) |
3113 | { |
3114 | this->dumpIRCosts = true; |
3115 | } |
3116 | |
3117 | if (dumpIRFlags) |
3118 | { |
3119 | this->dumpIRFlags = true; |
3120 | } |
3121 | |
3122 | if (dumpIRKinds) |
3123 | { |
3124 | this->dumpIRKinds = true; |
3125 | } |
3126 | |
3127 | if (dumpIRNodes) |
3128 | { |
3129 | this->dumpIRNodes = true; |
3130 | } |
3131 | |
3132 | if (dumpIRNoLists) |
3133 | { |
3134 | this->dumpIRNoLists = true; |
3135 | } |
3136 | |
3137 | if (dumpIRNoLeafs) |
3138 | { |
3139 | this->dumpIRNoLeafs = true; |
3140 | } |
3141 | |
3142 | if (dumpIRNoLeafs && dumpIRDataflow) |
3143 | { |
3144 | this->dumpIRDataflow = true; |
3145 | } |
3146 | |
3147 | if (dumpIRNoStmts) |
3148 | { |
3149 | this->dumpIRNoStmts = true; |
3150 | } |
3151 | |
3152 | if (dumpIRTrees) |
3153 | { |
3154 | this->dumpIRTrees = true; |
3155 | } |
3156 | |
3157 | if (dumpIRLinear) |
3158 | { |
3159 | this->dumpIRLinear = true; |
3160 | } |
3161 | |
3162 | if (dumpIRBlockHeaders) |
3163 | { |
3164 | this->dumpIRBlockHeaders = true; |
3165 | } |
3166 | |
3167 | if (dumpIRExit) |
3168 | { |
3169 | this->dumpIRExit = true; |
3170 | } |
3171 | |
3172 | #endif // DEBUG |
3173 | |
3174 | #ifdef FEATURE_SIMD |
3175 | // Minimum bar for availing SIMD benefits is SSE2 on AMD64/x86. |
3176 | featureSIMD = jitFlags->IsSet(JitFlags::JIT_FLAG_FEATURE_SIMD); |
3177 | setUsesSIMDTypes(false); |
3178 | #endif // FEATURE_SIMD |
3179 | |
3180 | if (compIsForImportOnly()) |
3181 | { |
3182 | return; |
3183 | } |
3184 | |
3185 | #if FEATURE_TAILCALL_OPT |
3186 | // By default opportunistic tail call optimization is enabled. |
3187 | // Recognition is done in the importer so this must be set for |
3188 | // inlinees as well. |
3189 | opts.compTailCallOpt = true; |
3190 | #endif // FEATURE_TAILCALL_OPT |
3191 | |
3192 | if (compIsForInlining()) |
3193 | { |
3194 | return; |
3195 | } |
3196 | |
3197 | // The rest of the opts fields that we initialize here |
3198 | // should only be used when we generate code for the method |
3199 | // They should not be used when importing or inlining |
3200 | CLANG_FORMAT_COMMENT_ANCHOR; |
3201 | |
3202 | #if FEATURE_TAILCALL_OPT |
3203 | opts.compTailCallLoopOpt = true; |
3204 | #endif // FEATURE_TAILCALL_OPT |
3205 | |
3206 | opts.genFPorder = true; |
3207 | opts.genFPopt = true; |
3208 | |
3209 | opts.instrCount = 0; |
3210 | opts.lvRefCount = 0; |
3211 | |
3212 | #ifdef PROFILING_SUPPORTED |
3213 | opts.compJitELTHookEnabled = false; |
3214 | #endif // PROFILING_SUPPORTED |
3215 | |
3216 | #ifdef DEBUG |
3217 | opts.dspInstrs = false; |
3218 | opts.dspEmit = false; |
3219 | opts.dspLines = false; |
3220 | opts.varNames = false; |
3221 | opts.dmpHex = false; |
3222 | opts.disAsm = false; |
3223 | opts.disAsmSpilled = false; |
3224 | opts.disDiffable = false; |
3225 | opts.dspCode = false; |
3226 | opts.dspEHTable = false; |
3227 | opts.dspDebugInfo = false; |
3228 | opts.dspGCtbls = false; |
3229 | opts.disAsm2 = false; |
3230 | opts.dspUnwind = false; |
3231 | opts.compLongAddress = false; |
3232 | opts.optRepeat = false; |
3233 | |
3234 | #ifdef LATE_DISASM |
3235 | opts.doLateDisasm = false; |
3236 | #endif // LATE_DISASM |
3237 | |
3238 | compDebugBreak = false; |
3239 | |
3240 | // If we have a non-empty AltJit config then we change all of these other |
3241 | // config values to refer only to the AltJit. |
3242 | // |
3243 | if (!altJitConfig || opts.altJit) |
3244 | { |
3245 | if (jitFlags->IsSet(JitFlags::JIT_FLAG_PREJIT)) |
3246 | { |
3247 | if ((JitConfig.NgenOrder() & 1) == 1) |
3248 | { |
3249 | opts.dspOrder = true; |
3250 | } |
3251 | |
3252 | if (JitConfig.NgenGCDump().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
3253 | { |
3254 | opts.dspGCtbls = true; |
3255 | } |
3256 | |
3257 | if (JitConfig.NgenDisasm().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
3258 | { |
3259 | opts.disAsm = true; |
3260 | } |
3261 | if (JitConfig.NgenDisasm().contains("SPILLED" , nullptr, nullptr)) |
3262 | { |
3263 | opts.disAsmSpilled = true; |
3264 | } |
3265 | |
3266 | if (JitConfig.NgenUnwindDump().contains(info.compMethodName, info.compClassName, |
3267 | &info.compMethodInfo->args)) |
3268 | { |
3269 | opts.dspUnwind = true; |
3270 | } |
3271 | |
3272 | if (JitConfig.NgenEHDump().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
3273 | { |
3274 | opts.dspEHTable = true; |
3275 | } |
3276 | |
3277 | if (JitConfig.NgenDebugDump().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
3278 | { |
3279 | opts.dspDebugInfo = true; |
3280 | } |
3281 | } |
3282 | else |
3283 | { |
3284 | bool disEnabled = true; |
3285 | |
3286 | // Setup assembly name list for disassembly, if not already set up. |
3287 | if (!s_pJitDisasmIncludeAssembliesListInitialized) |
3288 | { |
3289 | const wchar_t* assemblyNameList = JitConfig.JitDisasmAssemblies(); |
3290 | if (assemblyNameList != nullptr) |
3291 | { |
3292 | s_pJitDisasmIncludeAssembliesList = new (HostAllocator::getHostAllocator()) |
3293 | AssemblyNamesList2(assemblyNameList, HostAllocator::getHostAllocator()); |
3294 | } |
3295 | s_pJitDisasmIncludeAssembliesListInitialized = true; |
3296 | } |
3297 | |
3298 | // If we have an assembly name list for disassembly, also check this method's assembly. |
3299 | if (s_pJitDisasmIncludeAssembliesList != nullptr) |
3300 | { |
3301 | const char* assemblyName = info.compCompHnd->getAssemblyName( |
3302 | info.compCompHnd->getModuleAssembly(info.compCompHnd->getClassModule(info.compClassHnd))); |
3303 | |
3304 | if (!s_pJitDisasmIncludeAssembliesList->IsInList(assemblyName)) |
3305 | { |
3306 | disEnabled = false; |
3307 | } |
3308 | } |
3309 | |
3310 | if (disEnabled) |
3311 | { |
3312 | if ((JitConfig.JitOrder() & 1) == 1) |
3313 | { |
3314 | opts.dspOrder = true; |
3315 | } |
3316 | |
3317 | if (JitConfig.JitGCDump().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
3318 | { |
3319 | opts.dspGCtbls = true; |
3320 | } |
3321 | |
3322 | if (JitConfig.JitDisasm().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
3323 | { |
3324 | opts.disAsm = true; |
3325 | } |
3326 | |
3327 | if (JitConfig.JitDisasm().contains("SPILLED" , nullptr, nullptr)) |
3328 | { |
3329 | opts.disAsmSpilled = true; |
3330 | } |
3331 | |
3332 | if (JitConfig.JitUnwindDump().contains(info.compMethodName, info.compClassName, |
3333 | &info.compMethodInfo->args)) |
3334 | { |
3335 | opts.dspUnwind = true; |
3336 | } |
3337 | |
3338 | if (JitConfig.JitEHDump().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
3339 | { |
3340 | opts.dspEHTable = true; |
3341 | } |
3342 | |
3343 | if (JitConfig.JitDebugDump().contains(info.compMethodName, info.compClassName, |
3344 | &info.compMethodInfo->args)) |
3345 | { |
3346 | opts.dspDebugInfo = true; |
3347 | } |
3348 | } |
3349 | } |
3350 | |
3351 | #ifdef LATE_DISASM |
3352 | if (JitConfig.JitLateDisasm().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
3353 | opts.doLateDisasm = true; |
3354 | #endif // LATE_DISASM |
3355 | |
3356 | // This one applies to both Ngen/Jit Disasm output: COMPlus_JitDiffableDasm=1 |
3357 | if (JitConfig.DiffableDasm() != 0) |
3358 | { |
3359 | opts.disDiffable = true; |
3360 | opts.dspDiffable = true; |
3361 | } |
3362 | |
3363 | if (JitConfig.JitLongAddress() != 0) |
3364 | { |
3365 | opts.compLongAddress = true; |
3366 | } |
3367 | |
3368 | if (JitConfig.JitOptRepeat().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
3369 | { |
3370 | opts.optRepeat = true; |
3371 | } |
3372 | } |
3373 | |
3374 | if (verboseDump) |
3375 | { |
3376 | opts.dspCode = true; |
3377 | opts.dspEHTable = true; |
3378 | opts.dspGCtbls = true; |
3379 | opts.disAsm2 = true; |
3380 | opts.dspUnwind = true; |
3381 | verbose = true; |
3382 | verboseTrees = shouldUseVerboseTrees(); |
3383 | verboseSsa = shouldUseVerboseSsa(); |
3384 | codeGen->setVerbose(true); |
3385 | } |
3386 | |
3387 | treesBeforeAfterMorph = (JitConfig.TreesBeforeAfterMorph() == 1); |
3388 | morphNum = 0; // Initialize the morphed-trees counting. |
3389 | |
3390 | expensiveDebugCheckLevel = JitConfig.JitExpensiveDebugCheckLevel(); |
3391 | if (expensiveDebugCheckLevel == 0) |
3392 | { |
3393 | // If we're in a stress mode that modifies the flowgraph, make 1 the default. |
3394 | if (fgStressBBProf() || compStressCompile(STRESS_DO_WHILE_LOOPS, 30)) |
3395 | { |
3396 | expensiveDebugCheckLevel = 1; |
3397 | } |
3398 | } |
3399 | |
3400 | if (verbose) |
3401 | { |
3402 | printf("****** START compiling %s (MethodHash=%08x)\n" , info.compFullName, info.compMethodHash()); |
3403 | printf("Generating code for %s %s\n" , Target::g_tgtPlatformName, Target::g_tgtCPUName); |
3404 | printf("" ); // in our logic this causes a flush |
3405 | } |
3406 | |
3407 | if (JitConfig.JitBreak().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
3408 | { |
3409 | assert(!"JitBreak reached" ); |
3410 | } |
3411 | |
3412 | unsigned jitHashBreakVal = (unsigned)JitConfig.JitHashBreak(); |
3413 | if ((jitHashBreakVal != (DWORD)-1) && (jitHashBreakVal == info.compMethodHash())) |
3414 | { |
3415 | assert(!"JitHashBreak reached" ); |
3416 | } |
3417 | |
3418 | if (verbose || |
3419 | JitConfig.JitDebugBreak().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args) || |
3420 | JitConfig.JitBreak().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
3421 | { |
3422 | compDebugBreak = true; |
3423 | } |
3424 | |
3425 | memset(compActiveStressModes, 0, sizeof(compActiveStressModes)); |
3426 | |
3427 | #endif // DEBUG |
3428 | |
3429 | //------------------------------------------------------------------------- |
3430 | |
3431 | #ifdef DEBUG |
3432 | assert(!codeGen->isGCTypeFixed()); |
3433 | opts.compGcChecks = (JitConfig.JitGCChecks() != 0) || compStressCompile(STRESS_GENERIC_VARN, 5); |
3434 | #endif |
3435 | |
3436 | #if defined(DEBUG) && defined(_TARGET_XARCH_) |
3437 | enum |
3438 | { |
3439 | STACK_CHECK_ON_RETURN = 0x1, |
3440 | STACK_CHECK_ON_CALL = 0x2, |
3441 | STACK_CHECK_ALL = 0x3 |
3442 | }; |
3443 | |
3444 | DWORD dwJitStackChecks = JitConfig.JitStackChecks(); |
3445 | if (compStressCompile(STRESS_GENERIC_VARN, 5)) |
3446 | { |
3447 | dwJitStackChecks = STACK_CHECK_ALL; |
3448 | } |
3449 | opts.compStackCheckOnRet = (dwJitStackChecks & DWORD(STACK_CHECK_ON_RETURN)) != 0; |
3450 | #if defined(_TARGET_X86_) |
3451 | opts.compStackCheckOnCall = (dwJitStackChecks & DWORD(STACK_CHECK_ON_CALL)) != 0; |
3452 | #endif // defined(_TARGET_X86_) |
3453 | #endif // defined(DEBUG) && defined(_TARGET_XARCH_) |
3454 | |
3455 | #if MEASURE_MEM_ALLOC |
3456 | s_dspMemStats = (JitConfig.DisplayMemStats() != 0); |
3457 | #endif |
3458 | |
3459 | #ifdef PROFILING_SUPPORTED |
3460 | opts.compNoPInvokeInlineCB = jitFlags->IsSet(JitFlags::JIT_FLAG_PROF_NO_PINVOKE_INLINE); |
3461 | |
3462 | // Cache the profiler handle |
3463 | if (jitFlags->IsSet(JitFlags::JIT_FLAG_PROF_ENTERLEAVE)) |
3464 | { |
3465 | BOOL hookNeeded; |
3466 | BOOL indirected; |
3467 | info.compCompHnd->GetProfilingHandle(&hookNeeded, &compProfilerMethHnd, &indirected); |
3468 | compProfilerHookNeeded = !!hookNeeded; |
3469 | compProfilerMethHndIndirected = !!indirected; |
3470 | } |
3471 | else |
3472 | { |
3473 | compProfilerHookNeeded = false; |
3474 | compProfilerMethHnd = nullptr; |
3475 | compProfilerMethHndIndirected = false; |
3476 | } |
3477 | |
3478 | // Honour COMPlus_JitELTHookEnabled only if VM has not asked us to generate profiler |
3479 | // hooks in the first place. That is, override VM only if it hasn't asked for a |
3480 | // profiler callback for this method. |
3481 | if (!compProfilerHookNeeded && (JitConfig.JitELTHookEnabled() != 0)) |
3482 | { |
3483 | opts.compJitELTHookEnabled = true; |
3484 | } |
3485 | |
3486 | // TBD: Exclude PInvoke stubs |
3487 | if (opts.compJitELTHookEnabled) |
3488 | { |
3489 | compProfilerMethHnd = (void*)DummyProfilerELTStub; |
3490 | compProfilerMethHndIndirected = false; |
3491 | } |
3492 | |
3493 | #endif // PROFILING_SUPPORTED |
3494 | |
3495 | #if FEATURE_TAILCALL_OPT |
3496 | const wchar_t* strTailCallOpt = JitConfig.TailCallOpt(); |
3497 | if (strTailCallOpt != nullptr) |
3498 | { |
3499 | opts.compTailCallOpt = (UINT)_wtoi(strTailCallOpt) != 0; |
3500 | } |
3501 | |
3502 | if (JitConfig.TailCallLoopOpt() == 0) |
3503 | { |
3504 | opts.compTailCallLoopOpt = false; |
3505 | } |
3506 | #endif |
3507 | |
3508 | opts.compScopeInfo = opts.compDbgInfo; |
3509 | |
3510 | #ifdef LATE_DISASM |
3511 | codeGen->getDisAssembler().disOpenForLateDisAsm(info.compMethodName, info.compClassName, |
3512 | info.compMethodInfo->args.pSig); |
3513 | #endif |
3514 | |
3515 | //------------------------------------------------------------------------- |
3516 | |
3517 | opts.compReloc = jitFlags->IsSet(JitFlags::JIT_FLAG_RELOC); |
3518 | |
3519 | #ifdef DEBUG |
3520 | #if defined(_TARGET_XARCH_) |
3521 | // Whether encoding of absolute addr as PC-rel offset is enabled |
3522 | opts.compEnablePCRelAddr = (JitConfig.EnablePCRelAddr() != 0); |
3523 | #endif |
3524 | #endif // DEBUG |
3525 | |
3526 | opts.compProcedureSplitting = jitFlags->IsSet(JitFlags::JIT_FLAG_PROCSPLIT); |
3527 | |
3528 | #ifdef _TARGET_ARM64_ |
3529 | // TODO-ARM64-NYI: enable hot/cold splitting |
3530 | opts.compProcedureSplitting = false; |
3531 | #endif // _TARGET_ARM64_ |
3532 | |
3533 | #ifdef DEBUG |
3534 | opts.compProcedureSplittingEH = opts.compProcedureSplitting; |
3535 | #endif // DEBUG |
3536 | |
3537 | if (opts.compProcedureSplitting) |
3538 | { |
3539 | // Note that opts.compdbgCode is true under ngen for checked assemblies! |
3540 | opts.compProcedureSplitting = !opts.compDbgCode; |
3541 | |
3542 | #ifdef DEBUG |
3543 | // JitForceProcedureSplitting is used to force procedure splitting on checked assemblies. |
3544 | // This is useful for debugging on a checked build. Note that we still only do procedure |
3545 | // splitting in the zapper. |
3546 | if (JitConfig.JitForceProcedureSplitting().contains(info.compMethodName, info.compClassName, |
3547 | &info.compMethodInfo->args)) |
3548 | { |
3549 | opts.compProcedureSplitting = true; |
3550 | } |
3551 | |
3552 | // JitNoProcedureSplitting will always disable procedure splitting. |
3553 | if (JitConfig.JitNoProcedureSplitting().contains(info.compMethodName, info.compClassName, |
3554 | &info.compMethodInfo->args)) |
3555 | { |
3556 | opts.compProcedureSplitting = false; |
3557 | } |
3558 | // |
3559 | // JitNoProcedureSplittingEH will disable procedure splitting in functions with EH. |
3560 | if (JitConfig.JitNoProcedureSplittingEH().contains(info.compMethodName, info.compClassName, |
3561 | &info.compMethodInfo->args)) |
3562 | { |
3563 | opts.compProcedureSplittingEH = false; |
3564 | } |
3565 | #endif |
3566 | } |
3567 | |
3568 | fgProfileBuffer = nullptr; |
3569 | fgProfileData_ILSizeMismatch = false; |
3570 | fgNumProfileRuns = 0; |
3571 | if (jitFlags->IsSet(JitFlags::JIT_FLAG_BBOPT)) |
3572 | { |
3573 | assert(!compIsForInlining()); |
3574 | HRESULT hr; |
3575 | hr = info.compCompHnd->getBBProfileData(info.compMethodHnd, &fgProfileBufferCount, &fgProfileBuffer, |
3576 | &fgNumProfileRuns); |
3577 | |
3578 | // a failed result that also has a non-NULL fgProfileBuffer |
3579 | // indicates that the ILSize for the method no longer matches |
3580 | // the ILSize for the method when profile data was collected. |
3581 | // |
3582 | // We will discard the IBC data in this case |
3583 | // |
3584 | if (FAILED(hr) && (fgProfileBuffer != nullptr)) |
3585 | { |
3586 | fgProfileData_ILSizeMismatch = true; |
3587 | fgProfileBuffer = nullptr; |
3588 | } |
3589 | #ifdef DEBUG |
3590 | // A successful result implies a non-NULL fgProfileBuffer |
3591 | // |
3592 | if (SUCCEEDED(hr)) |
3593 | { |
3594 | assert(fgProfileBuffer != nullptr); |
3595 | } |
3596 | |
3597 | // A failed result implies a NULL fgProfileBuffer |
3598 | // see implementation of Compiler::fgHaveProfileData() |
3599 | // |
3600 | if (FAILED(hr)) |
3601 | { |
3602 | assert(fgProfileBuffer == nullptr); |
3603 | } |
3604 | #endif |
3605 | } |
3606 | |
3607 | opts.compNeedStackProbes = false; |
3608 | |
3609 | #ifdef DEBUG |
3610 | if (JitConfig.StackProbesOverride() != 0 || compStressCompile(STRESS_GENERIC_VARN, 5)) |
3611 | { |
3612 | opts.compNeedStackProbes = true; |
3613 | } |
3614 | #endif |
3615 | |
3616 | #ifdef DEBUG |
3617 | // Now, set compMaxUncheckedOffsetForNullObject for STRESS_NULL_OBJECT_CHECK |
3618 | if (compStressCompile(STRESS_NULL_OBJECT_CHECK, 30)) |
3619 | { |
3620 | compMaxUncheckedOffsetForNullObject = (size_t)JitConfig.JitMaxUncheckedOffset(); |
3621 | if (verbose) |
3622 | { |
3623 | printf("STRESS_NULL_OBJECT_CHECK: compMaxUncheckedOffsetForNullObject=0x%X\n" , |
3624 | compMaxUncheckedOffsetForNullObject); |
3625 | } |
3626 | } |
3627 | |
3628 | if (verbose) |
3629 | { |
3630 | // If we are compiling for a specific tier, make that very obvious in the output. |
3631 | // Note that we don't expect multiple TIER flags to be set at one time, but there |
3632 | // is nothing preventing that. |
3633 | if (jitFlags->IsSet(JitFlags::JIT_FLAG_TIER0)) |
3634 | { |
3635 | printf("OPTIONS: Tier-0 compilation (set COMPlus_TieredCompilation=0 to disable)\n" ); |
3636 | } |
3637 | if (jitFlags->IsSet(JitFlags::JIT_FLAG_TIER1)) |
3638 | { |
3639 | printf("OPTIONS: Tier-1 compilation\n" ); |
3640 | } |
3641 | |
3642 | printf("OPTIONS: compCodeOpt = %s\n" , |
3643 | (opts.compCodeOpt == BLENDED_CODE) |
3644 | ? "BLENDED_CODE" |
3645 | : (opts.compCodeOpt == SMALL_CODE) ? "SMALL_CODE" |
3646 | : (opts.compCodeOpt == FAST_CODE) ? "FAST_CODE" : "UNKNOWN_CODE" ); |
3647 | |
3648 | printf("OPTIONS: compDbgCode = %s\n" , dspBool(opts.compDbgCode)); |
3649 | printf("OPTIONS: compDbgInfo = %s\n" , dspBool(opts.compDbgInfo)); |
3650 | printf("OPTIONS: compDbgEnC = %s\n" , dspBool(opts.compDbgEnC)); |
3651 | printf("OPTIONS: compProcedureSplitting = %s\n" , dspBool(opts.compProcedureSplitting)); |
3652 | printf("OPTIONS: compProcedureSplittingEH = %s\n" , dspBool(opts.compProcedureSplittingEH)); |
3653 | |
3654 | if (jitFlags->IsSet(JitFlags::JIT_FLAG_BBOPT) && fgHaveProfileData()) |
3655 | { |
3656 | printf("OPTIONS: using real profile data\n" ); |
3657 | } |
3658 | |
3659 | if (fgProfileData_ILSizeMismatch) |
3660 | { |
3661 | printf("OPTIONS: discarded IBC profile data due to mismatch in ILSize\n" ); |
3662 | } |
3663 | |
3664 | if (jitFlags->IsSet(JitFlags::JIT_FLAG_PREJIT)) |
3665 | { |
3666 | printf("OPTIONS: Jit invoked for ngen\n" ); |
3667 | } |
3668 | printf("OPTIONS: Stack probing is %s\n" , opts.compNeedStackProbes ? "ENABLED" : "DISABLED" ); |
3669 | } |
3670 | #endif |
3671 | |
3672 | opts.compGCPollType = GCPOLL_NONE; |
3673 | if (jitFlags->IsSet(JitFlags::JIT_FLAG_GCPOLL_CALLS)) |
3674 | { |
3675 | opts.compGCPollType = GCPOLL_CALL; |
3676 | } |
3677 | else if (jitFlags->IsSet(JitFlags::JIT_FLAG_GCPOLL_INLINE)) |
3678 | { |
3679 | // make sure that the EE didn't set both flags. |
3680 | assert(opts.compGCPollType == GCPOLL_NONE); |
3681 | opts.compGCPollType = GCPOLL_INLINE; |
3682 | } |
3683 | |
3684 | #ifdef PROFILING_SUPPORTED |
3685 | #ifdef UNIX_AMD64_ABI |
3686 | if (compIsProfilerHookNeeded()) |
3687 | { |
3688 | opts.compNeedToAlignFrame = true; |
3689 | } |
3690 | #endif // UNIX_AMD64_ABI |
3691 | #endif |
3692 | } |
3693 | |
3694 | #ifdef DEBUG |
3695 | |
3696 | bool Compiler::compJitHaltMethod() |
3697 | { |
3698 | /* This method returns true when we use an INS_BREAKPOINT to allow us to step into the generated native code */ |
3699 | /* Note that this these two "Jit" environment variables also work for ngen images */ |
3700 | |
3701 | if (JitConfig.JitHalt().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
3702 | { |
3703 | return true; |
3704 | } |
3705 | |
3706 | /* Use this Hash variant when there are a lot of method with the same name and different signatures */ |
3707 | |
3708 | unsigned fJitHashHaltVal = (unsigned)JitConfig.JitHashHalt(); |
3709 | if ((fJitHashHaltVal != (unsigned)-1) && (fJitHashHaltVal == info.compMethodHash())) |
3710 | { |
3711 | return true; |
3712 | } |
3713 | |
3714 | return false; |
3715 | } |
3716 | |
3717 | /***************************************************************************** |
3718 | * Should we use a "stress-mode" for the given stressArea. We have different |
3719 | * areas to allow the areas to be mixed in different combinations in |
3720 | * different methods. |
3721 | * 'weight' indicates how often (as a percentage) the area should be stressed. |
3722 | * It should reflect the usefulness:overhead ratio. |
3723 | */ |
3724 | |
3725 | const LPCWSTR Compiler::s_compStressModeNames[STRESS_COUNT + 1] = { |
3726 | #define STRESS_MODE(mode) W("STRESS_") W(#mode), |
3727 | |
3728 | STRESS_MODES |
3729 | #undef STRESS_MODE |
3730 | }; |
3731 | |
3732 | bool Compiler::compStressCompile(compStressArea stressArea, unsigned weight) |
3733 | { |
3734 | unsigned hash; |
3735 | DWORD stressLevel; |
3736 | |
3737 | if (!bRangeAllowStress) |
3738 | { |
3739 | return false; |
3740 | } |
3741 | |
3742 | if (!JitConfig.JitStressOnly().isEmpty() && |
3743 | !JitConfig.JitStressOnly().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
3744 | { |
3745 | return false; |
3746 | } |
3747 | |
3748 | bool doStress = false; |
3749 | const wchar_t* strStressModeNames; |
3750 | |
3751 | // Does user explicitly prevent using this STRESS_MODE through the command line? |
3752 | const wchar_t* strStressModeNamesNot = JitConfig.JitStressModeNamesNot(); |
3753 | if ((strStressModeNamesNot != nullptr) && |
3754 | (wcsstr(strStressModeNamesNot, s_compStressModeNames[stressArea]) != nullptr)) |
3755 | { |
3756 | doStress = false; |
3757 | goto _done; |
3758 | } |
3759 | |
3760 | // Does user explicitly set this STRESS_MODE through the command line? |
3761 | strStressModeNames = JitConfig.JitStressModeNames(); |
3762 | if (strStressModeNames != nullptr) |
3763 | { |
3764 | if (wcsstr(strStressModeNames, s_compStressModeNames[stressArea]) != nullptr) |
3765 | { |
3766 | doStress = true; |
3767 | goto _done; |
3768 | } |
3769 | |
3770 | // This stress mode name did not match anything in the stress |
3771 | // mode whitelist. If user has requested only enable mode, |
3772 | // don't allow this stress mode to turn on. |
3773 | const bool onlyEnableMode = JitConfig.JitStressModeNamesOnly() != 0; |
3774 | |
3775 | if (onlyEnableMode) |
3776 | { |
3777 | doStress = false; |
3778 | goto _done; |
3779 | } |
3780 | } |
3781 | |
3782 | // 0: No stress (Except when explicitly set in complus_JitStressModeNames) |
3783 | // !=2: Vary stress. Performance will be slightly/moderately degraded |
3784 | // 2: Check-all stress. Performance will be REALLY horrible |
3785 | stressLevel = getJitStressLevel(); |
3786 | |
3787 | assert(weight <= MAX_STRESS_WEIGHT); |
3788 | |
3789 | /* Check for boundary conditions */ |
3790 | |
3791 | if (stressLevel == 0 || weight == 0) |
3792 | { |
3793 | return false; |
3794 | } |
3795 | |
3796 | // Should we allow unlimited stress ? |
3797 | if (stressArea > STRESS_COUNT_VARN && stressLevel == 2) |
3798 | { |
3799 | return true; |
3800 | } |
3801 | |
3802 | if (weight == MAX_STRESS_WEIGHT) |
3803 | { |
3804 | doStress = true; |
3805 | goto _done; |
3806 | } |
3807 | |
3808 | // Get a hash which can be compared with 'weight' |
3809 | |
3810 | assert(stressArea != 0); |
3811 | hash = (info.compMethodHash() ^ stressArea ^ stressLevel) % MAX_STRESS_WEIGHT; |
3812 | |
3813 | assert(hash < MAX_STRESS_WEIGHT && weight <= MAX_STRESS_WEIGHT); |
3814 | doStress = (hash < weight); |
3815 | |
3816 | _done: |
3817 | |
3818 | if (doStress && !compActiveStressModes[stressArea]) |
3819 | { |
3820 | if (verbose) |
3821 | { |
3822 | printf("\n\n*** JitStress: %ws ***\n\n" , s_compStressModeNames[stressArea]); |
3823 | } |
3824 | compActiveStressModes[stressArea] = 1; |
3825 | } |
3826 | |
3827 | return doStress; |
3828 | } |
3829 | |
3830 | #endif // DEBUG |
3831 | |
3832 | void Compiler::compInitDebuggingInfo() |
3833 | { |
3834 | assert(!compIsForInlining()); |
3835 | |
3836 | #ifdef DEBUG |
3837 | if (verbose) |
3838 | { |
3839 | printf("*************** In compInitDebuggingInfo() for %s\n" , info.compFullName); |
3840 | } |
3841 | #endif |
3842 | |
3843 | /*------------------------------------------------------------------------- |
3844 | * |
3845 | * Get hold of the local variable records, if there are any |
3846 | */ |
3847 | |
3848 | info.compVarScopesCount = 0; |
3849 | |
3850 | if (opts.compScopeInfo) |
3851 | { |
3852 | eeGetVars(); |
3853 | } |
3854 | |
3855 | compInitVarScopeMap(); |
3856 | |
3857 | if (opts.compScopeInfo || opts.compDbgCode) |
3858 | { |
3859 | compInitScopeLists(); |
3860 | } |
3861 | |
3862 | if (opts.compDbgCode && (info.compVarScopesCount > 0)) |
3863 | { |
3864 | /* Create a new empty basic block. fgExtendDbgLifetimes() may add |
3865 | initialization of variables which are in scope right from the |
3866 | start of the (real) first BB (and therefore artificially marked |
3867 | as alive) into this block. |
3868 | */ |
3869 | |
3870 | fgEnsureFirstBBisScratch(); |
3871 | |
3872 | fgInsertStmtAtEnd(fgFirstBB, gtNewNothingNode()); |
3873 | |
3874 | JITDUMP("Debuggable code - Add new %s to perform initialization of variables\n" , fgFirstBB->dspToString()); |
3875 | } |
3876 | |
3877 | /*------------------------------------------------------------------------- |
3878 | * |
3879 | * Read the stmt-offsets table and the line-number table |
3880 | */ |
3881 | |
3882 | info.compStmtOffsetsImplicit = ICorDebugInfo::NO_BOUNDARIES; |
3883 | |
3884 | // We can only report debug info for EnC at places where the stack is empty. |
3885 | // Actually, at places where there are not live temps. Else, we won't be able |
3886 | // to map between the old and the new versions correctly as we won't have |
3887 | // any info for the live temps. |
3888 | |
3889 | assert(!opts.compDbgEnC || !opts.compDbgInfo || |
3890 | 0 == (info.compStmtOffsetsImplicit & ~ICorDebugInfo::STACK_EMPTY_BOUNDARIES)); |
3891 | |
3892 | info.compStmtOffsetsCount = 0; |
3893 | |
3894 | if (opts.compDbgInfo) |
3895 | { |
3896 | /* Get hold of the line# records, if there are any */ |
3897 | |
3898 | eeGetStmtOffsets(); |
3899 | |
3900 | #ifdef DEBUG |
3901 | if (verbose) |
3902 | { |
3903 | printf("info.compStmtOffsetsCount = %d\n" , info.compStmtOffsetsCount); |
3904 | printf("info.compStmtOffsetsImplicit = %04Xh" , info.compStmtOffsetsImplicit); |
3905 | |
3906 | if (info.compStmtOffsetsImplicit) |
3907 | { |
3908 | printf(" ( " ); |
3909 | if (info.compStmtOffsetsImplicit & ICorDebugInfo::STACK_EMPTY_BOUNDARIES) |
3910 | { |
3911 | printf("STACK_EMPTY " ); |
3912 | } |
3913 | if (info.compStmtOffsetsImplicit & ICorDebugInfo::NOP_BOUNDARIES) |
3914 | { |
3915 | printf("NOP " ); |
3916 | } |
3917 | if (info.compStmtOffsetsImplicit & ICorDebugInfo::CALL_SITE_BOUNDARIES) |
3918 | { |
3919 | printf("CALL_SITE " ); |
3920 | } |
3921 | printf(")" ); |
3922 | } |
3923 | printf("\n" ); |
3924 | IL_OFFSET* pOffs = info.compStmtOffsets; |
3925 | for (unsigned i = 0; i < info.compStmtOffsetsCount; i++, pOffs++) |
3926 | { |
3927 | printf("%02d) IL_%04Xh\n" , i, *pOffs); |
3928 | } |
3929 | } |
3930 | #endif |
3931 | } |
3932 | } |
3933 | |
3934 | void Compiler::compSetOptimizationLevel() |
3935 | { |
3936 | bool theMinOptsValue; |
3937 | #pragma warning(suppress : 4101) |
3938 | unsigned jitMinOpts; |
3939 | |
3940 | if (compIsForInlining()) |
3941 | { |
3942 | theMinOptsValue = impInlineInfo->InlinerCompiler->opts.MinOpts(); |
3943 | goto _SetMinOpts; |
3944 | } |
3945 | |
3946 | theMinOptsValue = false; |
3947 | |
3948 | if (opts.compFlags == CLFLG_MINOPT) |
3949 | { |
3950 | JITLOG((LL_INFO100, "CLFLG_MINOPT set for method %s\n" , info.compFullName)); |
3951 | theMinOptsValue = true; |
3952 | } |
3953 | |
3954 | #ifdef DEBUG |
3955 | jitMinOpts = JitConfig.JitMinOpts(); |
3956 | |
3957 | if (!theMinOptsValue && (jitMinOpts > 0)) |
3958 | { |
3959 | // jitTotalMethodCompiled does not include the method that is being compiled now, so make +1. |
3960 | unsigned methodCount = Compiler::jitTotalMethodCompiled + 1; |
3961 | unsigned methodCountMask = methodCount & 0xFFF; |
3962 | unsigned kind = (jitMinOpts & 0xF000000) >> 24; |
3963 | switch (kind) |
3964 | { |
3965 | default: |
3966 | if (jitMinOpts <= methodCount) |
3967 | { |
3968 | if (verbose) |
3969 | { |
3970 | printf(" Optimizations disabled by JitMinOpts and methodCount\n" ); |
3971 | } |
3972 | theMinOptsValue = true; |
3973 | } |
3974 | break; |
3975 | case 0xD: |
3976 | { |
3977 | unsigned firstMinopts = (jitMinOpts >> 12) & 0xFFF; |
3978 | unsigned secondMinopts = (jitMinOpts >> 0) & 0xFFF; |
3979 | |
3980 | if ((firstMinopts == methodCountMask) || (secondMinopts == methodCountMask)) |
3981 | { |
3982 | if (verbose) |
3983 | { |
3984 | printf("0xD: Optimizations disabled by JitMinOpts and methodCountMask\n" ); |
3985 | } |
3986 | theMinOptsValue = true; |
3987 | } |
3988 | } |
3989 | break; |
3990 | case 0xE: |
3991 | { |
3992 | unsigned startMinopts = (jitMinOpts >> 12) & 0xFFF; |
3993 | unsigned endMinopts = (jitMinOpts >> 0) & 0xFFF; |
3994 | |
3995 | if ((startMinopts <= methodCountMask) && (endMinopts >= methodCountMask)) |
3996 | { |
3997 | if (verbose) |
3998 | { |
3999 | printf("0xE: Optimizations disabled by JitMinOpts and methodCountMask\n" ); |
4000 | } |
4001 | theMinOptsValue = true; |
4002 | } |
4003 | } |
4004 | break; |
4005 | case 0xF: |
4006 | { |
4007 | unsigned bitsZero = (jitMinOpts >> 12) & 0xFFF; |
4008 | unsigned bitsOne = (jitMinOpts >> 0) & 0xFFF; |
4009 | |
4010 | if (((methodCountMask & bitsOne) == bitsOne) && ((~methodCountMask & bitsZero) == bitsZero)) |
4011 | { |
4012 | if (verbose) |
4013 | { |
4014 | printf("0xF: Optimizations disabled by JitMinOpts and methodCountMask\n" ); |
4015 | } |
4016 | theMinOptsValue = true; |
4017 | } |
4018 | } |
4019 | break; |
4020 | } |
4021 | } |
4022 | |
4023 | if (!theMinOptsValue) |
4024 | { |
4025 | if (JitConfig.JitMinOptsName().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
4026 | { |
4027 | theMinOptsValue = true; |
4028 | } |
4029 | } |
4030 | |
4031 | #if 0 |
4032 | // The code in this #if can be used to debug optimization issues according to method hash. |
4033 | // To use, uncomment, rebuild and set environment variables minoptshashlo and minoptshashhi. |
4034 | #ifdef DEBUG |
4035 | unsigned methHash = info.compMethodHash(); |
4036 | char* lostr = getenv("minoptshashlo" ); |
4037 | unsigned methHashLo = 0; |
4038 | if (lostr != nullptr) |
4039 | { |
4040 | sscanf_s(lostr, "%x" , &methHashLo); |
4041 | char* histr = getenv("minoptshashhi" ); |
4042 | unsigned methHashHi = UINT32_MAX; |
4043 | if (histr != nullptr) |
4044 | { |
4045 | sscanf_s(histr, "%x" , &methHashHi); |
4046 | if (methHash >= methHashLo && methHash <= methHashHi) |
4047 | { |
4048 | printf("MinOpts for method %s, hash = 0x%x.\n" , |
4049 | info.compFullName, info.compMethodHash()); |
4050 | printf("" ); // in our logic this causes a flush |
4051 | theMinOptsValue = true; |
4052 | } |
4053 | } |
4054 | } |
4055 | #endif |
4056 | #endif |
4057 | |
4058 | if (compStressCompile(STRESS_MIN_OPTS, 5)) |
4059 | { |
4060 | theMinOptsValue = true; |
4061 | } |
4062 | // For PREJIT we never drop down to MinOpts |
4063 | // unless unless CLFLG_MINOPT is set |
4064 | else if (!opts.jitFlags->IsSet(JitFlags::JIT_FLAG_PREJIT)) |
4065 | { |
4066 | if ((unsigned)JitConfig.JitMinOptsCodeSize() < info.compILCodeSize) |
4067 | { |
4068 | JITLOG((LL_INFO10, "IL Code Size exceeded, using MinOpts for method %s\n" , info.compFullName)); |
4069 | theMinOptsValue = true; |
4070 | } |
4071 | else if ((unsigned)JitConfig.JitMinOptsInstrCount() < opts.instrCount) |
4072 | { |
4073 | JITLOG((LL_INFO10, "IL instruction count exceeded, using MinOpts for method %s\n" , info.compFullName)); |
4074 | theMinOptsValue = true; |
4075 | } |
4076 | else if ((unsigned)JitConfig.JitMinOptsBbCount() < fgBBcount) |
4077 | { |
4078 | JITLOG((LL_INFO10, "Basic Block count exceeded, using MinOpts for method %s\n" , info.compFullName)); |
4079 | theMinOptsValue = true; |
4080 | } |
4081 | else if ((unsigned)JitConfig.JitMinOptsLvNumCount() < lvaCount) |
4082 | { |
4083 | JITLOG((LL_INFO10, "Local Variable Num count exceeded, using MinOpts for method %s\n" , info.compFullName)); |
4084 | theMinOptsValue = true; |
4085 | } |
4086 | else if ((unsigned)JitConfig.JitMinOptsLvRefCount() < opts.lvRefCount) |
4087 | { |
4088 | JITLOG((LL_INFO10, "Local Variable Ref count exceeded, using MinOpts for method %s\n" , info.compFullName)); |
4089 | theMinOptsValue = true; |
4090 | } |
4091 | if (theMinOptsValue == true) |
4092 | { |
4093 | JITLOG((LL_INFO10000, "IL Code Size,Instr %4d,%4d, Basic Block count %3d, Local Variable Num,Ref count " |
4094 | "%3d,%3d for method %s\n" , |
4095 | info.compILCodeSize, opts.instrCount, fgBBcount, lvaCount, opts.lvRefCount, info.compFullName)); |
4096 | if (JitConfig.JitBreakOnMinOpts() != 0) |
4097 | { |
4098 | assert(!"MinOpts enabled" ); |
4099 | } |
4100 | } |
4101 | } |
4102 | #else // !DEBUG |
4103 | // Retail check if we should force Minopts due to the complexity of the method |
4104 | // For PREJIT we never drop down to MinOpts |
4105 | // unless unless CLFLG_MINOPT is set |
4106 | if (!theMinOptsValue && !opts.jitFlags->IsSet(JitFlags::JIT_FLAG_PREJIT) && |
4107 | ((DEFAULT_MIN_OPTS_CODE_SIZE < info.compILCodeSize) || (DEFAULT_MIN_OPTS_INSTR_COUNT < opts.instrCount) || |
4108 | (DEFAULT_MIN_OPTS_BB_COUNT < fgBBcount) || (DEFAULT_MIN_OPTS_LV_NUM_COUNT < lvaCount) || |
4109 | (DEFAULT_MIN_OPTS_LV_REF_COUNT < opts.lvRefCount))) |
4110 | { |
4111 | theMinOptsValue = true; |
4112 | } |
4113 | #endif // DEBUG |
4114 | |
4115 | JITLOG((LL_INFO10000, |
4116 | "IL Code Size,Instr %4d,%4d, Basic Block count %3d, Local Variable Num,Ref count %3d,%3d for method %s\n" , |
4117 | info.compILCodeSize, opts.instrCount, fgBBcount, lvaCount, opts.lvRefCount, info.compFullName)); |
4118 | |
4119 | #if 0 |
4120 | // The code in this #if has been useful in debugging loop cloning issues, by |
4121 | // enabling selective enablement of the loop cloning optimization according to |
4122 | // method hash. |
4123 | #ifdef DEBUG |
4124 | if (!theMinOptsValue) |
4125 | { |
4126 | unsigned methHash = info.compMethodHash(); |
4127 | char* lostr = getenv("opthashlo" ); |
4128 | unsigned methHashLo = 0; |
4129 | if (lostr != NULL) |
4130 | { |
4131 | sscanf_s(lostr, "%x" , &methHashLo); |
4132 | // methHashLo = (unsigned(atoi(lostr)) << 2); // So we don't have to use negative numbers. |
4133 | } |
4134 | char* histr = getenv("opthashhi" ); |
4135 | unsigned methHashHi = UINT32_MAX; |
4136 | if (histr != NULL) |
4137 | { |
4138 | sscanf_s(histr, "%x" , &methHashHi); |
4139 | // methHashHi = (unsigned(atoi(histr)) << 2); // So we don't have to use negative numbers. |
4140 | } |
4141 | if (methHash < methHashLo || methHash > methHashHi) |
4142 | { |
4143 | theMinOptsValue = true; |
4144 | } |
4145 | else |
4146 | { |
4147 | printf("Doing optimization in in %s (0x%x).\n" , info.compFullName, methHash); |
4148 | } |
4149 | } |
4150 | #endif |
4151 | #endif |
4152 | |
4153 | _SetMinOpts: |
4154 | |
4155 | // Set the MinOpts value |
4156 | opts.SetMinOpts(theMinOptsValue); |
4157 | |
4158 | #ifdef DEBUG |
4159 | if (verbose && !compIsForInlining()) |
4160 | { |
4161 | printf("OPTIONS: opts.MinOpts() == %s\n" , opts.MinOpts() ? "true" : "false" ); |
4162 | } |
4163 | #endif |
4164 | |
4165 | /* Control the optimizations */ |
4166 | |
4167 | if (opts.OptimizationDisabled()) |
4168 | { |
4169 | opts.compFlags &= ~CLFLG_MAXOPT; |
4170 | opts.compFlags |= CLFLG_MINOPT; |
4171 | } |
4172 | |
4173 | if (!compIsForInlining()) |
4174 | { |
4175 | codeGen->setFramePointerRequired(false); |
4176 | codeGen->setFrameRequired(false); |
4177 | |
4178 | if (opts.OptimizationDisabled()) |
4179 | { |
4180 | codeGen->setFrameRequired(true); |
4181 | } |
4182 | |
4183 | #if !defined(_TARGET_AMD64_) |
4184 | // The VM sets JitFlags::JIT_FLAG_FRAMED for two reasons: (1) the COMPlus_JitFramed variable is set, or |
4185 | // (2) the function is marked "noinline". The reason for #2 is that people mark functions |
4186 | // noinline to ensure the show up on in a stack walk. But for AMD64, we don't need a frame |
4187 | // pointer for the frame to show up in stack walk. |
4188 | if (opts.jitFlags->IsSet(JitFlags::JIT_FLAG_FRAMED)) |
4189 | codeGen->setFrameRequired(true); |
4190 | #endif |
4191 | |
4192 | if (opts.jitFlags->IsSet(JitFlags::JIT_FLAG_RELOC)) |
4193 | { |
4194 | codeGen->genAlignLoops = false; // loop alignment not supported for prejitted code |
4195 | |
4196 | // The zapper doesn't set JitFlags::JIT_FLAG_ALIGN_LOOPS, and there is |
4197 | // no reason for it to set it as the JIT doesn't currently support loop alignment |
4198 | // for prejitted images. (The JIT doesn't know the final address of the code, hence |
4199 | // it can't align code based on unknown addresses.) |
4200 | assert(!opts.jitFlags->IsSet(JitFlags::JIT_FLAG_ALIGN_LOOPS)); |
4201 | } |
4202 | else |
4203 | { |
4204 | codeGen->genAlignLoops = opts.jitFlags->IsSet(JitFlags::JIT_FLAG_ALIGN_LOOPS); |
4205 | } |
4206 | } |
4207 | |
4208 | info.compUnwrapContextful = opts.OptimizationEnabled(); |
4209 | |
4210 | fgCanRelocateEHRegions = true; |
4211 | } |
4212 | |
4213 | #ifdef _TARGET_ARMARCH_ |
4214 | // Function compRsvdRegCheck: |
4215 | // given a curState to use for calculating the total frame size |
4216 | // it will return true if the REG_OPT_RSVD should be reserved so |
4217 | // that it can be use to form large offsets when accessing stack |
4218 | // based LclVar including both incoming and out going argument areas. |
4219 | // |
4220 | // The method advances the frame layout state to curState by calling |
4221 | // lvaFrameSize(curState). |
4222 | // |
4223 | bool Compiler::compRsvdRegCheck(FrameLayoutState curState) |
4224 | { |
4225 | // Always do the layout even if returning early. Callers might |
4226 | // depend on us to do the layout. |
4227 | unsigned frameSize = lvaFrameSize(curState); |
4228 | JITDUMP("\n" |
4229 | "compRsvdRegCheck\n" |
4230 | " frame size = %6d\n" |
4231 | " compArgSize = %6d\n" , |
4232 | frameSize, compArgSize); |
4233 | |
4234 | if (opts.MinOpts()) |
4235 | { |
4236 | // Have a recovery path in case we fail to reserve REG_OPT_RSVD and go |
4237 | // over the limit of SP and FP offset ranges due to large |
4238 | // temps. |
4239 | JITDUMP(" Returning true (MinOpts)\n\n" ); |
4240 | return true; |
4241 | } |
4242 | |
4243 | unsigned calleeSavedRegMaxSz = CALLEE_SAVED_REG_MAXSZ; |
4244 | if (compFloatingPointUsed) |
4245 | { |
4246 | calleeSavedRegMaxSz += CALLEE_SAVED_FLOAT_MAXSZ; |
4247 | } |
4248 | calleeSavedRegMaxSz += REGSIZE_BYTES; // we always push LR. See genPushCalleeSavedRegisters |
4249 | |
4250 | noway_assert(frameSize >= calleeSavedRegMaxSz); |
4251 | |
4252 | #if defined(_TARGET_ARM64_) |
4253 | |
4254 | // TODO-ARM64-CQ: update this! |
4255 | JITDUMP(" Returning true (ARM64)\n\n" ); |
4256 | return true; // just always assume we'll need it, for now |
4257 | |
4258 | #else // _TARGET_ARM_ |
4259 | |
4260 | // frame layout: |
4261 | // |
4262 | // ... high addresses ... |
4263 | // frame contents size |
4264 | // ------------------- ------------------------ |
4265 | // inArgs compArgSize (includes prespill) |
4266 | // caller SP ---> |
4267 | // prespill |
4268 | // LR REGSIZE_BYTES |
4269 | // R11 ---> R11 REGSIZE_BYTES |
4270 | // callee saved regs CALLEE_SAVED_REG_MAXSZ (32 bytes) |
4271 | // optional saved fp regs CALLEE_SAVED_FLOAT_MAXSZ (64 bytes) |
4272 | // lclSize |
4273 | // incl. TEMPS MAX_SPILL_TEMP_SIZE |
4274 | // incl. outArgs |
4275 | // SP ---> |
4276 | // ... low addresses ... |
4277 | // |
4278 | // When codeGen->isFramePointerRequired is true, R11 will be established as a frame pointer. |
4279 | // We can then use R11 to access incoming args with positive offsets, and LclVars with |
4280 | // negative offsets. |
4281 | // |
4282 | // In functions with EH, in the non-funclet (or main) region, even though we will have a |
4283 | // frame pointer, we can use SP with positive offsets to access any or all locals or arguments |
4284 | // that we can reach with SP-relative encodings. The funclet region might require the reserved |
4285 | // register, since it must use offsets from R11 to access the parent frame. |
4286 | |
4287 | unsigned maxR11PositiveEncodingOffset = compFloatingPointUsed ? 0x03FC : 0x0FFF; |
4288 | JITDUMP(" maxR11PositiveEncodingOffset = %6d\n" , maxR11PositiveEncodingOffset); |
4289 | |
4290 | // Floating point load/store instructions (VLDR/VSTR) can address up to -0x3FC from R11, but we |
4291 | // don't know if there are either no integer locals, or if we don't need large negative offsets |
4292 | // for the integer locals, so we must use the integer max negative offset, which is a |
4293 | // smaller (absolute value) number. |
4294 | unsigned maxR11NegativeEncodingOffset = 0x00FF; // This is a negative offset from R11. |
4295 | JITDUMP(" maxR11NegativeEncodingOffset = %6d\n" , maxR11NegativeEncodingOffset); |
4296 | |
4297 | // -1 because otherwise we are computing the address just beyond the last argument, which we don't need to do. |
4298 | unsigned maxR11PositiveOffset = compArgSize + (2 * REGSIZE_BYTES) - 1; |
4299 | JITDUMP(" maxR11PositiveOffset = %6d\n" , maxR11PositiveOffset); |
4300 | |
4301 | // The value is positive, but represents a negative offset from R11. |
4302 | // frameSize includes callee-saved space for R11 and LR, which are at non-negative offsets from R11 |
4303 | // (+0 and +4, respectively), so don't include those in the max possible negative offset. |
4304 | assert(frameSize >= (2 * REGSIZE_BYTES)); |
4305 | unsigned maxR11NegativeOffset = frameSize - (2 * REGSIZE_BYTES); |
4306 | JITDUMP(" maxR11NegativeOffset = %6d\n" , maxR11NegativeOffset); |
4307 | |
4308 | if (codeGen->isFramePointerRequired()) |
4309 | { |
4310 | if (maxR11NegativeOffset > maxR11NegativeEncodingOffset) |
4311 | { |
4312 | JITDUMP(" Returning true (frame required and maxR11NegativeOffset)\n\n" ); |
4313 | return true; |
4314 | } |
4315 | if (maxR11PositiveOffset > maxR11PositiveEncodingOffset) |
4316 | { |
4317 | JITDUMP(" Returning true (frame required and maxR11PositiveOffset)\n\n" ); |
4318 | return true; |
4319 | } |
4320 | } |
4321 | |
4322 | // Now consider the SP based frame case. Note that we will use SP based offsets to access the stack in R11 based |
4323 | // frames in the non-funclet main code area. |
4324 | |
4325 | unsigned maxSPPositiveEncodingOffset = compFloatingPointUsed ? 0x03FC : 0x0FFF; |
4326 | JITDUMP(" maxSPPositiveEncodingOffset = %6d\n" , maxSPPositiveEncodingOffset); |
4327 | |
4328 | // -1 because otherwise we are computing the address just beyond the last argument, which we don't need to do. |
4329 | assert(compArgSize + frameSize > 0); |
4330 | unsigned maxSPPositiveOffset = compArgSize + frameSize - 1; |
4331 | |
4332 | if (codeGen->isFramePointerUsed()) |
4333 | { |
4334 | // We have a frame pointer, so we can use it to access part of the stack, even if SP can't reach those parts. |
4335 | // We will still generate SP-relative offsets if SP can reach. |
4336 | |
4337 | // First, check that the stack between R11 and SP can be fully reached, either via negative offset from FP |
4338 | // or positive offset from SP. Don't count stored R11 or LR, which are reached from positive offsets from FP. |
4339 | |
4340 | unsigned maxSPLocalsCombinedOffset = frameSize - (2 * REGSIZE_BYTES) - 1; |
4341 | JITDUMP(" maxSPLocalsCombinedOffset = %6d\n" , maxSPLocalsCombinedOffset); |
4342 | |
4343 | if (maxSPLocalsCombinedOffset > maxSPPositiveEncodingOffset) |
4344 | { |
4345 | // Can R11 help? |
4346 | unsigned maxRemainingLocalsCombinedOffset = maxSPLocalsCombinedOffset - maxSPPositiveEncodingOffset; |
4347 | JITDUMP(" maxRemainingLocalsCombinedOffset = %6d\n" , maxRemainingLocalsCombinedOffset); |
4348 | |
4349 | if (maxRemainingLocalsCombinedOffset > maxR11NegativeEncodingOffset) |
4350 | { |
4351 | JITDUMP(" Returning true (frame pointer exists; R11 and SP can't reach entire stack between them)\n\n" ); |
4352 | return true; |
4353 | } |
4354 | |
4355 | // Otherwise, yes, we can address the remaining parts of the locals frame with negative offsets from R11. |
4356 | } |
4357 | |
4358 | // Check whether either R11 or SP can access the arguments. |
4359 | if ((maxR11PositiveOffset > maxR11PositiveEncodingOffset) && |
4360 | (maxSPPositiveOffset > maxSPPositiveEncodingOffset)) |
4361 | { |
4362 | JITDUMP(" Returning true (frame pointer exists; R11 and SP can't reach all arguments)\n\n" ); |
4363 | return true; |
4364 | } |
4365 | } |
4366 | else |
4367 | { |
4368 | if (maxSPPositiveOffset > maxSPPositiveEncodingOffset) |
4369 | { |
4370 | JITDUMP(" Returning true (no frame pointer exists; SP can't reach all of frame)\n\n" ); |
4371 | return true; |
4372 | } |
4373 | } |
4374 | |
4375 | // We won't need to reserve REG_OPT_RSVD. |
4376 | // |
4377 | JITDUMP(" Returning false\n\n" ); |
4378 | return false; |
4379 | #endif // _TARGET_ARM_ |
4380 | } |
4381 | #endif // _TARGET_ARMARCH_ |
4382 | |
4383 | void Compiler::compFunctionTraceStart() |
4384 | { |
4385 | #ifdef DEBUG |
4386 | if (compIsForInlining()) |
4387 | { |
4388 | return; |
4389 | } |
4390 | |
4391 | if ((JitConfig.JitFunctionTrace() != 0) && !opts.disDiffable) |
4392 | { |
4393 | LONG newJitNestingLevel = InterlockedIncrement(&Compiler::jitNestingLevel); |
4394 | if (newJitNestingLevel <= 0) |
4395 | { |
4396 | printf("{ Illegal nesting level %d }\n" , newJitNestingLevel); |
4397 | } |
4398 | |
4399 | for (LONG i = 0; i < newJitNestingLevel - 1; i++) |
4400 | { |
4401 | printf(" " ); |
4402 | } |
4403 | printf("{ Start Jitting %s (MethodHash=%08x)\n" , info.compFullName, |
4404 | info.compMethodHash()); /* } editor brace matching workaround for this printf */ |
4405 | } |
4406 | #endif // DEBUG |
4407 | } |
4408 | |
4409 | void Compiler::compFunctionTraceEnd(void* methodCodePtr, ULONG methodCodeSize, bool isNYI) |
4410 | { |
4411 | #ifdef DEBUG |
4412 | assert(!compIsForInlining()); |
4413 | |
4414 | if ((JitConfig.JitFunctionTrace() != 0) && !opts.disDiffable) |
4415 | { |
4416 | LONG newJitNestingLevel = InterlockedDecrement(&Compiler::jitNestingLevel); |
4417 | if (newJitNestingLevel < 0) |
4418 | { |
4419 | printf("{ Illegal nesting level %d }\n" , newJitNestingLevel); |
4420 | } |
4421 | |
4422 | for (LONG i = 0; i < newJitNestingLevel; i++) |
4423 | { |
4424 | printf(" " ); |
4425 | } |
4426 | /* { editor brace-matching workaround for following printf */ |
4427 | printf("} Jitted Entry %03x at" FMT_ADDR "method %s size %08x%s\n" , Compiler::jitTotalMethodCompiled, |
4428 | DBG_ADDR(methodCodePtr), info.compFullName, methodCodeSize, |
4429 | isNYI ? " NYI" : (compIsForImportOnly() ? " import only" : "" )); |
4430 | } |
4431 | #endif // DEBUG |
4432 | } |
4433 | |
4434 | //********************************************************************************************* |
4435 | // #Phases |
4436 | // |
4437 | // This is the most interesting 'toplevel' function in the JIT. It goes through the operations of |
4438 | // importing, morphing, optimizations and code generation. This is called from the EE through the |
4439 | // code:CILJit::compileMethod function. |
4440 | // |
4441 | // For an overview of the structure of the JIT, see: |
4442 | // https://github.com/dotnet/coreclr/blob/master/Documentation/botr/ryujit-overview.md |
4443 | // |
4444 | void Compiler::compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags) |
4445 | { |
4446 | if (compIsForInlining()) |
4447 | { |
4448 | // Notify root instance that an inline attempt is about to import IL |
4449 | impInlineRoot()->m_inlineStrategy->NoteImport(); |
4450 | } |
4451 | |
4452 | hashBv::Init(this); |
4453 | |
4454 | VarSetOps::AssignAllowUninitRhs(this, compCurLife, VarSetOps::UninitVal()); |
4455 | |
4456 | /* The temp holding the secret stub argument is used by fgImport() when importing the intrinsic. */ |
4457 | |
4458 | if (info.compPublishStubParam) |
4459 | { |
4460 | assert(lvaStubArgumentVar == BAD_VAR_NUM); |
4461 | lvaStubArgumentVar = lvaGrabTempWithImplicitUse(false DEBUGARG("stub argument" )); |
4462 | lvaTable[lvaStubArgumentVar].lvType = TYP_I_IMPL; |
4463 | } |
4464 | |
4465 | EndPhase(PHASE_PRE_IMPORT); |
4466 | |
4467 | compFunctionTraceStart(); |
4468 | |
4469 | /* Convert the instrs in each basic block to a tree based intermediate representation */ |
4470 | |
4471 | fgImport(); |
4472 | |
4473 | assert(!fgComputePredsDone); |
4474 | if (fgCheapPredsValid) |
4475 | { |
4476 | // Remove cheap predecessors before inlining and fat call transformation; |
4477 | // allowing the cheap predecessor lists to be inserted causes problems |
4478 | // with splitting existing blocks. |
4479 | fgRemovePreds(); |
4480 | } |
4481 | |
4482 | // Transform indirect calls that require control flow expansion. |
4483 | fgTransformIndirectCalls(); |
4484 | |
4485 | EndPhase(PHASE_IMPORTATION); |
4486 | |
4487 | if (compIsForInlining()) |
4488 | { |
4489 | /* Quit inlining if fgImport() failed for any reason. */ |
4490 | |
4491 | if (!compDonotInline()) |
4492 | { |
4493 | /* Filter out unimported BBs */ |
4494 | |
4495 | fgRemoveEmptyBlocks(); |
4496 | |
4497 | // Update type of return spill temp if we have gathered |
4498 | // better info when importing the inlinee, and the return |
4499 | // spill temp is single def. |
4500 | if (fgNeedReturnSpillTemp()) |
4501 | { |
4502 | CORINFO_CLASS_HANDLE retExprClassHnd = impInlineInfo->retExprClassHnd; |
4503 | if (retExprClassHnd != nullptr) |
4504 | { |
4505 | LclVarDsc* returnSpillVarDsc = lvaGetDesc(lvaInlineeReturnSpillTemp); |
4506 | |
4507 | if (returnSpillVarDsc->lvSingleDef) |
4508 | { |
4509 | lvaUpdateClass(lvaInlineeReturnSpillTemp, retExprClassHnd, |
4510 | impInlineInfo->retExprClassHndIsExact); |
4511 | } |
4512 | } |
4513 | } |
4514 | } |
4515 | |
4516 | EndPhase(PHASE_POST_IMPORT); |
4517 | |
4518 | #ifdef FEATURE_JIT_METHOD_PERF |
4519 | if (pCompJitTimer != nullptr) |
4520 | { |
4521 | #if MEASURE_CLRAPI_CALLS |
4522 | EndPhase(PHASE_CLR_API); |
4523 | #endif |
4524 | pCompJitTimer->Terminate(this, CompTimeSummaryInfo::s_compTimeSummary, false); |
4525 | } |
4526 | #endif |
4527 | |
4528 | return; |
4529 | } |
4530 | |
4531 | assert(!compDonotInline()); |
4532 | |
4533 | // Maybe the caller was not interested in generating code |
4534 | if (compIsForImportOnly()) |
4535 | { |
4536 | compFunctionTraceEnd(nullptr, 0, false); |
4537 | return; |
4538 | } |
4539 | |
4540 | #if !FEATURE_EH |
4541 | // If we aren't yet supporting EH in a compiler bring-up, remove as many EH handlers as possible, so |
4542 | // we can pass tests that contain try/catch EH, but don't actually throw any exceptions. |
4543 | fgRemoveEH(); |
4544 | #endif // !FEATURE_EH |
4545 | |
4546 | if (compileFlags->IsSet(JitFlags::JIT_FLAG_BBINSTR)) |
4547 | { |
4548 | fgInstrumentMethod(); |
4549 | } |
4550 | |
4551 | // We could allow ESP frames. Just need to reserve space for |
4552 | // pushing EBP if the method becomes an EBP-frame after an edit. |
4553 | // Note that requiring a EBP Frame disallows double alignment. Thus if we change this |
4554 | // we either have to disallow double alignment for E&C some other way or handle it in EETwain. |
4555 | |
4556 | if (opts.compDbgEnC) |
4557 | { |
4558 | codeGen->setFramePointerRequired(true); |
4559 | |
4560 | // Since we need a slots for security near ebp, its not possible |
4561 | // to do this after an Edit without shifting all the locals. |
4562 | // So we just always reserve space for these slots in case an Edit adds them |
4563 | opts.compNeedSecurityCheck = true; |
4564 | |
4565 | // We don't care about localloc right now. If we do support it, |
4566 | // EECodeManager::FixContextForEnC() needs to handle it smartly |
4567 | // in case the localloc was actually executed. |
4568 | // |
4569 | // compLocallocUsed = true; |
4570 | } |
4571 | |
4572 | EndPhase(PHASE_POST_IMPORT); |
4573 | |
4574 | /* Initialize the BlockSet epoch */ |
4575 | |
4576 | NewBasicBlockEpoch(); |
4577 | |
4578 | /* Massage the trees so that we can generate code out of them */ |
4579 | |
4580 | fgMorph(); |
4581 | EndPhase(PHASE_MORPH_END); |
4582 | |
4583 | /* GS security checks for unsafe buffers */ |
4584 | if (getNeedsGSSecurityCookie()) |
4585 | { |
4586 | #ifdef DEBUG |
4587 | if (verbose) |
4588 | { |
4589 | printf("\n*************** -GS checks for unsafe buffers \n" ); |
4590 | } |
4591 | #endif |
4592 | |
4593 | gsGSChecksInitCookie(); |
4594 | |
4595 | if (compGSReorderStackLayout) |
4596 | { |
4597 | gsCopyShadowParams(); |
4598 | } |
4599 | |
4600 | #ifdef DEBUG |
4601 | if (verbose) |
4602 | { |
4603 | fgDispBasicBlocks(true); |
4604 | printf("\n" ); |
4605 | } |
4606 | #endif |
4607 | } |
4608 | EndPhase(PHASE_GS_COOKIE); |
4609 | |
4610 | /* Compute bbNum, bbRefs and bbPreds */ |
4611 | |
4612 | JITDUMP("\nRenumbering the basic blocks for fgComputePred\n" ); |
4613 | fgRenumberBlocks(); |
4614 | |
4615 | noway_assert(!fgComputePredsDone); // This is the first time full (not cheap) preds will be computed. |
4616 | fgComputePreds(); |
4617 | EndPhase(PHASE_COMPUTE_PREDS); |
4618 | |
4619 | /* If we need to emit GC Poll calls, mark the blocks that need them now. This is conservative and can |
4620 | * be optimized later. */ |
4621 | fgMarkGCPollBlocks(); |
4622 | EndPhase(PHASE_MARK_GC_POLL_BLOCKS); |
4623 | |
4624 | /* From this point on the flowgraph information such as bbNum, |
4625 | * bbRefs or bbPreds has to be kept updated */ |
4626 | |
4627 | // Compute the block and edge weights |
4628 | fgComputeBlockAndEdgeWeights(); |
4629 | EndPhase(PHASE_COMPUTE_EDGE_WEIGHTS); |
4630 | |
4631 | #if FEATURE_EH_FUNCLETS |
4632 | |
4633 | /* Create funclets from the EH handlers. */ |
4634 | |
4635 | fgCreateFunclets(); |
4636 | EndPhase(PHASE_CREATE_FUNCLETS); |
4637 | |
4638 | #endif // FEATURE_EH_FUNCLETS |
4639 | |
4640 | if (opts.OptimizationEnabled()) |
4641 | { |
4642 | optOptimizeLayout(); |
4643 | EndPhase(PHASE_OPTIMIZE_LAYOUT); |
4644 | |
4645 | // Compute reachability sets and dominators. |
4646 | fgComputeReachability(); |
4647 | EndPhase(PHASE_COMPUTE_REACHABILITY); |
4648 | } |
4649 | |
4650 | if (opts.OptimizationEnabled()) |
4651 | { |
4652 | /* Perform loop inversion (i.e. transform "while" loops into |
4653 | "repeat" loops) and discover and classify natural loops |
4654 | (e.g. mark iterative loops as such). Also marks loop blocks |
4655 | and sets bbWeight to the loop nesting levels |
4656 | */ |
4657 | |
4658 | optOptimizeLoops(); |
4659 | EndPhase(PHASE_OPTIMIZE_LOOPS); |
4660 | |
4661 | // Clone loops with optimization opportunities, and |
4662 | // choose the one based on dynamic condition evaluation. |
4663 | optCloneLoops(); |
4664 | EndPhase(PHASE_CLONE_LOOPS); |
4665 | |
4666 | /* Unroll loops */ |
4667 | optUnrollLoops(); |
4668 | EndPhase(PHASE_UNROLL_LOOPS); |
4669 | } |
4670 | |
4671 | #ifdef DEBUG |
4672 | fgDebugCheckLinks(); |
4673 | #endif |
4674 | |
4675 | /* Create the variable table (and compute variable ref counts) */ |
4676 | |
4677 | lvaMarkLocalVars(); |
4678 | EndPhase(PHASE_MARK_LOCAL_VARS); |
4679 | |
4680 | // IMPORTANT, after this point, every place where trees are modified or cloned |
4681 | // the local variable reference counts must be updated |
4682 | // You can test the value of the following variable to see if |
4683 | // the local variable ref counts must be updated |
4684 | // |
4685 | assert(lvaLocalVarRefCounted()); |
4686 | |
4687 | if (opts.OptimizationEnabled()) |
4688 | { |
4689 | /* Optimize boolean conditions */ |
4690 | |
4691 | optOptimizeBools(); |
4692 | EndPhase(PHASE_OPTIMIZE_BOOLS); |
4693 | |
4694 | // optOptimizeBools() might have changed the number of blocks; the dominators/reachability might be bad. |
4695 | } |
4696 | |
4697 | /* Figure out the order in which operators are to be evaluated */ |
4698 | fgFindOperOrder(); |
4699 | EndPhase(PHASE_FIND_OPER_ORDER); |
4700 | |
4701 | // Weave the tree lists. Anyone who modifies the tree shapes after |
4702 | // this point is responsible for calling fgSetStmtSeq() to keep the |
4703 | // nodes properly linked. |
4704 | // This can create GC poll calls, and create new BasicBlocks (without updating dominators/reachability). |
4705 | fgSetBlockOrder(); |
4706 | EndPhase(PHASE_SET_BLOCK_ORDER); |
4707 | |
4708 | // IMPORTANT, after this point, every place where tree topology changes must redo evaluation |
4709 | // order (gtSetStmtInfo) and relink nodes (fgSetStmtSeq) if required. |
4710 | CLANG_FORMAT_COMMENT_ANCHOR; |
4711 | |
4712 | #ifdef DEBUG |
4713 | // Now we have determined the order of evaluation and the gtCosts for every node. |
4714 | // If verbose, dump the full set of trees here before the optimization phases mutate them |
4715 | // |
4716 | if (verbose) |
4717 | { |
4718 | fgDispBasicBlocks(true); // 'true' will call fgDumpTrees() after dumping the BasicBlocks |
4719 | printf("\n" ); |
4720 | } |
4721 | #endif |
4722 | |
4723 | // At this point we know if we are fully interruptible or not |
4724 | if (opts.OptimizationEnabled()) |
4725 | { |
4726 | bool doSsa = true; |
4727 | bool doEarlyProp = true; |
4728 | bool doValueNum = true; |
4729 | bool doLoopHoisting = true; |
4730 | bool doCopyProp = true; |
4731 | bool doAssertionProp = true; |
4732 | bool doRangeAnalysis = true; |
4733 | int iterations = 1; |
4734 | |
4735 | #if defined(OPT_CONFIG) |
4736 | doSsa = (JitConfig.JitDoSsa() != 0); |
4737 | doEarlyProp = doSsa && (JitConfig.JitDoEarlyProp() != 0); |
4738 | doValueNum = doSsa && (JitConfig.JitDoValueNumber() != 0); |
4739 | doLoopHoisting = doValueNum && (JitConfig.JitDoLoopHoisting() != 0); |
4740 | doCopyProp = doValueNum && (JitConfig.JitDoCopyProp() != 0); |
4741 | doAssertionProp = doValueNum && (JitConfig.JitDoAssertionProp() != 0); |
4742 | doRangeAnalysis = doAssertionProp && (JitConfig.JitDoRangeAnalysis() != 0); |
4743 | |
4744 | if (opts.optRepeat) |
4745 | { |
4746 | iterations = JitConfig.JitOptRepeatCount(); |
4747 | } |
4748 | #endif // defined(OPT_CONFIG) |
4749 | |
4750 | while (iterations > 0) |
4751 | { |
4752 | if (doSsa) |
4753 | { |
4754 | fgSsaBuild(); |
4755 | EndPhase(PHASE_BUILD_SSA); |
4756 | } |
4757 | |
4758 | if (doEarlyProp) |
4759 | { |
4760 | /* Propagate array length and rewrite getType() method call */ |
4761 | optEarlyProp(); |
4762 | EndPhase(PHASE_EARLY_PROP); |
4763 | } |
4764 | |
4765 | if (doValueNum) |
4766 | { |
4767 | fgValueNumber(); |
4768 | EndPhase(PHASE_VALUE_NUMBER); |
4769 | } |
4770 | |
4771 | if (doLoopHoisting) |
4772 | { |
4773 | /* Hoist invariant code out of loops */ |
4774 | optHoistLoopCode(); |
4775 | EndPhase(PHASE_HOIST_LOOP_CODE); |
4776 | } |
4777 | |
4778 | if (doCopyProp) |
4779 | { |
4780 | /* Perform VN based copy propagation */ |
4781 | optVnCopyProp(); |
4782 | EndPhase(PHASE_VN_COPY_PROP); |
4783 | } |
4784 | |
4785 | #if FEATURE_ANYCSE |
4786 | /* Remove common sub-expressions */ |
4787 | optOptimizeCSEs(); |
4788 | #endif // FEATURE_ANYCSE |
4789 | |
4790 | #if ASSERTION_PROP |
4791 | if (doAssertionProp) |
4792 | { |
4793 | /* Assertion propagation */ |
4794 | optAssertionPropMain(); |
4795 | EndPhase(PHASE_ASSERTION_PROP_MAIN); |
4796 | } |
4797 | |
4798 | if (doRangeAnalysis) |
4799 | { |
4800 | /* Optimize array index range checks */ |
4801 | RangeCheck rc(this); |
4802 | rc.OptimizeRangeChecks(); |
4803 | EndPhase(PHASE_OPTIMIZE_INDEX_CHECKS); |
4804 | } |
4805 | #endif // ASSERTION_PROP |
4806 | |
4807 | /* update the flowgraph if we modified it during the optimization phase*/ |
4808 | if (fgModified) |
4809 | { |
4810 | fgUpdateFlowGraph(); |
4811 | EndPhase(PHASE_UPDATE_FLOW_GRAPH); |
4812 | |
4813 | // Recompute the edge weight if we have modified the flow graph |
4814 | fgComputeEdgeWeights(); |
4815 | EndPhase(PHASE_COMPUTE_EDGE_WEIGHTS2); |
4816 | } |
4817 | |
4818 | // Iterate if requested, resetting annotations first. |
4819 | if (--iterations == 0) |
4820 | { |
4821 | break; |
4822 | } |
4823 | ResetOptAnnotations(); |
4824 | RecomputeLoopInfo(); |
4825 | } |
4826 | } |
4827 | |
4828 | #ifdef _TARGET_AMD64_ |
4829 | // Check if we need to add the Quirk for the PPP backward compat issue |
4830 | compQuirkForPPPflag = compQuirkForPPP(); |
4831 | #endif |
4832 | |
4833 | fgDetermineFirstColdBlock(); |
4834 | EndPhase(PHASE_DETERMINE_FIRST_COLD_BLOCK); |
4835 | |
4836 | #ifdef DEBUG |
4837 | fgDebugCheckLinks(compStressCompile(STRESS_REMORPH_TREES, 50)); |
4838 | |
4839 | // Stash the current estimate of the function's size if necessary. |
4840 | if (verbose) |
4841 | { |
4842 | compSizeEstimate = 0; |
4843 | compCycleEstimate = 0; |
4844 | for (BasicBlock* block = fgFirstBB; block != nullptr; block = block->bbNext) |
4845 | { |
4846 | for (GenTreeStmt* stmt = block->firstStmt(); stmt != nullptr; stmt = stmt->getNextStmt()) |
4847 | { |
4848 | compSizeEstimate += stmt->GetCostSz(); |
4849 | compCycleEstimate += stmt->GetCostEx(); |
4850 | } |
4851 | } |
4852 | } |
4853 | #endif |
4854 | |
4855 | // rationalize trees |
4856 | Rationalizer rat(this); // PHASE_RATIONALIZE |
4857 | rat.Run(); |
4858 | |
4859 | // Here we do "simple lowering". When the RyuJIT backend works for all |
4860 | // platforms, this will be part of the more general lowering phase. For now, though, we do a separate |
4861 | // pass of "final lowering." We must do this before (final) liveness analysis, because this creates |
4862 | // range check throw blocks, in which the liveness must be correct. |
4863 | fgSimpleLowering(); |
4864 | EndPhase(PHASE_SIMPLE_LOWERING); |
4865 | |
4866 | #ifdef DEBUG |
4867 | fgDebugCheckBBlist(); |
4868 | fgDebugCheckLinks(); |
4869 | #endif |
4870 | |
4871 | /* Enable this to gather statistical data such as |
4872 | * call and register argument info, flowgraph and loop info, etc. */ |
4873 | |
4874 | compJitStats(); |
4875 | |
4876 | #ifdef _TARGET_ARM_ |
4877 | if (compLocallocUsed) |
4878 | { |
4879 | // We reserve REG_SAVED_LOCALLOC_SP to store SP on entry for stack unwinding |
4880 | codeGen->regSet.rsMaskResvd |= RBM_SAVED_LOCALLOC_SP; |
4881 | } |
4882 | #endif // _TARGET_ARM_ |
4883 | |
4884 | /* Assign registers to variables, etc. */ |
4885 | |
4886 | /////////////////////////////////////////////////////////////////////////////// |
4887 | // Dominator and reachability sets are no longer valid. They haven't been |
4888 | // maintained up to here, and shouldn't be used (unless recomputed). |
4889 | /////////////////////////////////////////////////////////////////////////////// |
4890 | fgDomsComputed = false; |
4891 | |
4892 | /* Create LSRA before Lowering, this way Lowering can initialize the TreeNode Map */ |
4893 | m_pLinearScan = getLinearScanAllocator(this); |
4894 | |
4895 | /* Lower */ |
4896 | m_pLowering = new (this, CMK_LSRA) Lowering(this, m_pLinearScan); // PHASE_LOWERING |
4897 | m_pLowering->Run(); |
4898 | |
4899 | StackLevelSetter stackLevelSetter(this); // PHASE_STACK_LEVEL_SETTER |
4900 | stackLevelSetter.Run(); |
4901 | |
4902 | lvaTrackedFixed = true; // We can not add any new tracked variables after this point. |
4903 | |
4904 | /* Now that lowering is completed we can proceed to perform register allocation */ |
4905 | m_pLinearScan->doLinearScan(); |
4906 | EndPhase(PHASE_LINEAR_SCAN); |
4907 | |
4908 | // Copied from rpPredictRegUse() |
4909 | genFullPtrRegMap = (codeGen->genInterruptible || !codeGen->isFramePointerUsed()); |
4910 | |
4911 | #ifdef DEBUG |
4912 | fgDebugCheckLinks(); |
4913 | #endif |
4914 | |
4915 | /* Generate code */ |
4916 | |
4917 | codeGen->genGenerateCode(methodCodePtr, methodCodeSize); |
4918 | |
4919 | #ifdef FEATURE_JIT_METHOD_PERF |
4920 | if (pCompJitTimer) |
4921 | { |
4922 | #if MEASURE_CLRAPI_CALLS |
4923 | EndPhase(PHASE_CLR_API); |
4924 | #endif |
4925 | pCompJitTimer->Terminate(this, CompTimeSummaryInfo::s_compTimeSummary, true); |
4926 | } |
4927 | #endif |
4928 | |
4929 | RecordStateAtEndOfCompilation(); |
4930 | |
4931 | #ifdef FEATURE_TRACELOGGING |
4932 | compJitTelemetry.NotifyEndOfCompilation(); |
4933 | #endif |
4934 | |
4935 | #if defined(DEBUG) |
4936 | ++Compiler::jitTotalMethodCompiled; |
4937 | #endif // defined(DEBUG) |
4938 | |
4939 | compFunctionTraceEnd(*methodCodePtr, *methodCodeSize, false); |
4940 | JITDUMP("Method code size: %d\n" , (unsigned)(*methodCodeSize)); |
4941 | |
4942 | #if FUNC_INFO_LOGGING |
4943 | if (compJitFuncInfoFile != nullptr) |
4944 | { |
4945 | assert(!compIsForInlining()); |
4946 | #ifdef DEBUG // We only have access to info.compFullName in DEBUG builds. |
4947 | fprintf(compJitFuncInfoFile, "%s\n" , info.compFullName); |
4948 | #elif FEATURE_SIMD |
4949 | fprintf(compJitFuncInfoFile, " %s\n" , eeGetMethodFullName(info.compMethodHnd)); |
4950 | #endif |
4951 | fprintf(compJitFuncInfoFile, "" ); // in our logic this causes a flush |
4952 | } |
4953 | #endif // FUNC_INFO_LOGGING |
4954 | } |
4955 | |
4956 | //------------------------------------------------------------------------ |
4957 | // ResetOptAnnotations: Clear annotations produced during global optimizations. |
4958 | // |
4959 | // Notes: |
4960 | // The intent of this method is to clear any information typically assumed |
4961 | // to be set only once; it is used between iterations when JitOptRepeat is |
4962 | // in effect. |
4963 | |
4964 | void Compiler::ResetOptAnnotations() |
4965 | { |
4966 | assert(opts.optRepeat); |
4967 | assert(JitConfig.JitOptRepeatCount() > 0); |
4968 | fgResetForSsa(); |
4969 | vnStore = nullptr; |
4970 | m_opAsgnVarDefSsaNums = nullptr; |
4971 | m_blockToEHPreds = nullptr; |
4972 | fgSsaPassesCompleted = 0; |
4973 | fgVNPassesCompleted = 0; |
4974 | |
4975 | for (BasicBlock* block = fgFirstBB; block != nullptr; block = block->bbNext) |
4976 | { |
4977 | for (GenTreeStmt* stmt = block->firstStmt(); stmt != nullptr; stmt = stmt->getNextStmt()) |
4978 | { |
4979 | stmt->gtFlags &= ~GTF_STMT_HAS_CSE; |
4980 | |
4981 | for (GenTree* tree = stmt->gtStmt.gtStmtList; tree != nullptr; tree = tree->gtNext) |
4982 | { |
4983 | tree->ClearVN(); |
4984 | tree->ClearAssertion(); |
4985 | tree->gtCSEnum = NO_CSE; |
4986 | } |
4987 | } |
4988 | } |
4989 | } |
4990 | |
4991 | //------------------------------------------------------------------------ |
4992 | // RecomputeLoopInfo: Recompute loop annotations between opt-repeat iterations. |
4993 | // |
4994 | // Notes: |
4995 | // The intent of this method is to update loop structure annotations, and those |
4996 | // they depend on; these annotations may have become stale during optimization, |
4997 | // and need to be up-to-date before running another iteration of optimizations. |
4998 | |
4999 | void Compiler::RecomputeLoopInfo() |
5000 | { |
5001 | assert(opts.optRepeat); |
5002 | assert(JitConfig.JitOptRepeatCount() > 0); |
5003 | // Recompute reachability sets, dominators, and loops. |
5004 | optLoopCount = 0; |
5005 | fgDomsComputed = false; |
5006 | for (BasicBlock* block = fgFirstBB; block != nullptr; block = block->bbNext) |
5007 | { |
5008 | block->bbFlags &= ~BBF_LOOP_FLAGS; |
5009 | } |
5010 | fgComputeReachability(); |
5011 | // Rebuild the loop tree annotations themselves. Since this is performed as |
5012 | // part of 'optOptimizeLoops', this will also re-perform loop rotation, but |
5013 | // not other optimizations, as the others are not part of 'optOptimizeLoops'. |
5014 | optOptimizeLoops(); |
5015 | } |
5016 | |
5017 | /*****************************************************************************/ |
5018 | void Compiler::ProcessShutdownWork(ICorStaticInfo* statInfo) |
5019 | { |
5020 | } |
5021 | |
5022 | #ifdef _TARGET_AMD64_ |
5023 | // Check if we need to add the Quirk for the PPP backward compat issue. |
5024 | // This Quirk addresses a compatibility issue between the new RyuJit and the previous JIT64. |
5025 | // A backward compatibity issue called 'PPP' exists where a PInvoke call passes a 32-byte struct |
5026 | // into a native API which basically writes 48 bytes of data into the struct. |
5027 | // With the stack frame layout used by the RyuJIT the extra 16 bytes written corrupts a |
5028 | // caller saved register and this leads to an A/V in the calling method. |
5029 | // The older JIT64 jit compiler just happened to have a different stack layout and/or |
5030 | // caller saved register set so that it didn't hit the A/V in the caller. |
5031 | // By increasing the amount of stack allocted for the struct by 32 bytes we can fix this. |
5032 | // |
5033 | // Return true if we actually perform the Quirk, otherwise return false |
5034 | // |
5035 | bool Compiler::compQuirkForPPP() |
5036 | { |
5037 | if (lvaCount != 2) |
5038 | { // We require that there are exactly two locals |
5039 | return false; |
5040 | } |
5041 | |
5042 | if (compTailCallUsed) |
5043 | { // Don't try this quirk if a tail call was used |
5044 | return false; |
5045 | } |
5046 | |
5047 | bool hasOutArgs = false; |
5048 | LclVarDsc* varDscExposedStruct = nullptr; |
5049 | |
5050 | unsigned lclNum; |
5051 | LclVarDsc* varDsc; |
5052 | |
5053 | /* Look for struct locals that are address taken */ |
5054 | for (lclNum = 0, varDsc = lvaTable; lclNum < lvaCount; lclNum++, varDsc++) |
5055 | { |
5056 | if (varDsc->lvIsParam) // It can't be a parameter |
5057 | { |
5058 | continue; |
5059 | } |
5060 | |
5061 | // We require that the OutgoingArg space lclVar exists |
5062 | if (lclNum == lvaOutgoingArgSpaceVar) |
5063 | { |
5064 | hasOutArgs = true; // Record that we saw it |
5065 | continue; |
5066 | } |
5067 | |
5068 | // Look for a 32-byte address exposed Struct and record its varDsc |
5069 | if ((varDsc->TypeGet() == TYP_STRUCT) && varDsc->lvAddrExposed && (varDsc->lvExactSize == 32)) |
5070 | { |
5071 | varDscExposedStruct = varDsc; |
5072 | } |
5073 | } |
5074 | |
5075 | // We only perform the Quirk when there are two locals |
5076 | // one of them is a address exposed struct of size 32 |
5077 | // and the other is the outgoing arg space local |
5078 | // |
5079 | if (hasOutArgs && (varDscExposedStruct != nullptr)) |
5080 | { |
5081 | #ifdef DEBUG |
5082 | if (verbose) |
5083 | { |
5084 | printf("\nAdding a backwards compatibility quirk for the 'PPP' issue\n" ); |
5085 | } |
5086 | #endif // DEBUG |
5087 | |
5088 | // Increase the exact size of this struct by 32 bytes |
5089 | // This fixes the PPP backward compat issue |
5090 | varDscExposedStruct->lvExactSize += 32; |
5091 | |
5092 | // Update the GC info to indicate that the padding area does |
5093 | // not contain any GC pointers. |
5094 | // |
5095 | // The struct is now 64 bytes. |
5096 | // |
5097 | // We're on x64 so this should be 8 pointer slots. |
5098 | assert((varDscExposedStruct->lvExactSize / TARGET_POINTER_SIZE) == 8); |
5099 | |
5100 | BYTE* oldGCPtrs = varDscExposedStruct->lvGcLayout; |
5101 | BYTE* newGCPtrs = getAllocator(CMK_LvaTable).allocate<BYTE>(8); |
5102 | |
5103 | for (int i = 0; i < 4; i++) |
5104 | { |
5105 | newGCPtrs[i] = oldGCPtrs[i]; |
5106 | } |
5107 | |
5108 | for (int i = 4; i < 8; i++) |
5109 | { |
5110 | newGCPtrs[i] = TYPE_GC_NONE; |
5111 | } |
5112 | |
5113 | varDscExposedStruct->lvGcLayout = newGCPtrs; |
5114 | |
5115 | return true; |
5116 | } |
5117 | return false; |
5118 | } |
5119 | #endif // _TARGET_AMD64_ |
5120 | |
5121 | /*****************************************************************************/ |
5122 | |
5123 | #ifdef DEBUG |
5124 | void* forceFrameJIT; // used to force to frame &useful for fastchecked debugging |
5125 | |
5126 | bool Compiler::skipMethod() |
5127 | { |
5128 | static ConfigMethodRange fJitRange; |
5129 | fJitRange.EnsureInit(JitConfig.JitRange()); |
5130 | assert(!fJitRange.Error()); |
5131 | |
5132 | // Normally JitConfig.JitRange() is null, we don't want to skip |
5133 | // jitting any methods. |
5134 | // |
5135 | // So, the logic below relies on the fact that a null range string |
5136 | // passed to ConfigMethodRange represents the set of all methods. |
5137 | |
5138 | if (!fJitRange.Contains(info.compCompHnd, info.compMethodHnd)) |
5139 | { |
5140 | return true; |
5141 | } |
5142 | |
5143 | if (JitConfig.JitExclude().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
5144 | { |
5145 | return true; |
5146 | } |
5147 | |
5148 | if (!JitConfig.JitInclude().isEmpty() && |
5149 | !JitConfig.JitInclude().contains(info.compMethodName, info.compClassName, &info.compMethodInfo->args)) |
5150 | { |
5151 | return true; |
5152 | } |
5153 | |
5154 | return false; |
5155 | } |
5156 | |
5157 | #endif |
5158 | |
5159 | /*****************************************************************************/ |
5160 | |
5161 | int Compiler::compCompile(CORINFO_METHOD_HANDLE methodHnd, |
5162 | CORINFO_MODULE_HANDLE classPtr, |
5163 | COMP_HANDLE compHnd, |
5164 | CORINFO_METHOD_INFO* methodInfo, |
5165 | void** methodCodePtr, |
5166 | ULONG* methodCodeSize, |
5167 | JitFlags* compileFlags) |
5168 | { |
5169 | #ifdef FEATURE_JIT_METHOD_PERF |
5170 | static bool checkedForJitTimeLog = false; |
5171 | |
5172 | pCompJitTimer = nullptr; |
5173 | |
5174 | if (!checkedForJitTimeLog) |
5175 | { |
5176 | // Call into VM to get the config strings. FEATURE_JIT_METHOD_PERF is enabled for |
5177 | // retail builds. Do not call the regular Config helper here as it would pull |
5178 | // in a copy of the config parser into the clrjit.dll. |
5179 | InterlockedCompareExchangeT(&Compiler::compJitTimeLogFilename, compHnd->getJitTimeLogFilename(), NULL); |
5180 | |
5181 | // At a process or module boundary clear the file and start afresh. |
5182 | JitTimer::PrintCsvHeader(); |
5183 | |
5184 | checkedForJitTimeLog = true; |
5185 | } |
5186 | if ((Compiler::compJitTimeLogFilename != nullptr) || (JitTimeLogCsv() != nullptr)) |
5187 | { |
5188 | pCompJitTimer = JitTimer::Create(this, methodInfo->ILCodeSize); |
5189 | } |
5190 | #endif // FEATURE_JIT_METHOD_PERF |
5191 | |
5192 | #ifdef DEBUG |
5193 | Compiler* me = this; |
5194 | forceFrameJIT = (void*)&me; // let us see the this pointer in fastchecked build |
5195 | // set this early so we can use it without relying on random memory values |
5196 | verbose = compIsForInlining() ? impInlineInfo->InlinerCompiler->verbose : false; |
5197 | |
5198 | this->dumpIR = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIR : false; |
5199 | this->dumpIRPhase = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRPhase : nullptr; |
5200 | this->dumpIRFormat = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRFormat : nullptr; |
5201 | this->dumpIRTypes = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRTypes : false; |
5202 | this->dumpIRLocals = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRLocals : false; |
5203 | this->dumpIRRegs = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRRegs : false; |
5204 | this->dumpIRSsa = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRSsa : false; |
5205 | this->dumpIRValnums = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRValnums : false; |
5206 | this->dumpIRCosts = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRCosts : false; |
5207 | this->dumpIRFlags = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRFlags : false; |
5208 | this->dumpIRKinds = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRKinds : false; |
5209 | this->dumpIRNodes = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRNodes : false; |
5210 | this->dumpIRNoLists = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRNoLists : false; |
5211 | this->dumpIRNoLeafs = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRNoLeafs : false; |
5212 | this->dumpIRNoStmts = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRNoStmts : false; |
5213 | this->dumpIRTrees = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRTrees : false; |
5214 | this->dumpIRLinear = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRLinear : false; |
5215 | this->dumpIRDataflow = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRDataflow : false; |
5216 | this->dumpIRBlockHeaders = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRBlockHeaders : NULL; |
5217 | this->dumpIRExit = compIsForInlining() ? impInlineInfo->InlinerCompiler->dumpIRExit : NULL; |
5218 | |
5219 | #endif |
5220 | |
5221 | #if defined(DEBUG) || defined(INLINE_DATA) |
5222 | info.compMethodHashPrivate = 0; |
5223 | #endif // defined(DEBUG) || defined(INLINE_DATA) |
5224 | |
5225 | #if FUNC_INFO_LOGGING |
5226 | LPCWSTR tmpJitFuncInfoFilename = JitConfig.JitFuncInfoFile(); |
5227 | |
5228 | if (tmpJitFuncInfoFilename != nullptr) |
5229 | { |
5230 | LPCWSTR oldFuncInfoFileName = |
5231 | InterlockedCompareExchangeT(&compJitFuncInfoFilename, tmpJitFuncInfoFilename, NULL); |
5232 | if (oldFuncInfoFileName == nullptr) |
5233 | { |
5234 | assert(compJitFuncInfoFile == nullptr); |
5235 | compJitFuncInfoFile = _wfopen(compJitFuncInfoFilename, W("a" )); |
5236 | if (compJitFuncInfoFile == nullptr) |
5237 | { |
5238 | #if defined(DEBUG) && !defined(FEATURE_PAL) // no 'perror' in the PAL |
5239 | perror("Failed to open JitFuncInfoLogFile" ); |
5240 | #endif // defined(DEBUG) && !defined(FEATURE_PAL) |
5241 | } |
5242 | } |
5243 | } |
5244 | #endif // FUNC_INFO_LOGGING |
5245 | |
5246 | // if (s_compMethodsCount==0) setvbuf(jitstdout, NULL, _IONBF, 0); |
5247 | |
5248 | info.compCompHnd = compHnd; |
5249 | info.compMethodHnd = methodHnd; |
5250 | info.compMethodInfo = methodInfo; |
5251 | |
5252 | virtualStubParamInfo = new (this, CMK_Unknown) VirtualStubParamInfo(IsTargetAbi(CORINFO_CORERT_ABI)); |
5253 | |
5254 | // Do we have a matched VM? Or are we "abusing" the VM to help us do JIT work (such as using an x86 native VM |
5255 | // with an ARM-targeting "altjit"). |
5256 | info.compMatchedVM = IMAGE_FILE_MACHINE_TARGET == info.compCompHnd->getExpectedTargetArchitecture(); |
5257 | |
5258 | #if (defined(_TARGET_UNIX_) && !defined(_HOST_UNIX_)) || (!defined(_TARGET_UNIX_) && defined(_HOST_UNIX_)) |
5259 | // The host and target platforms don't match. This info isn't handled by the existing |
5260 | // getExpectedTargetArchitecture() JIT-EE interface method. |
5261 | info.compMatchedVM = false; |
5262 | #endif |
5263 | |
5264 | // If we are not compiling for a matched VM, then we are getting JIT flags that don't match our target |
5265 | // architecture. The two main examples here are an ARM targeting altjit hosted on x86 and an ARM64 |
5266 | // targeting altjit hosted on x64. (Though with cross-bitness work, the host doesn't necessarily need |
5267 | // to be of the same bitness.) In these cases, we need to fix up the JIT flags to be appropriate for |
5268 | // the target, as the VM's expected target may overlap bit flags with different meaning to our target. |
5269 | // Note that it might be better to do this immediately when setting the JIT flags in CILJit::compileMethod() |
5270 | // (when JitFlags::SetFromFlags() is called), but this is close enough. (To move this logic to |
5271 | // CILJit::compileMethod() would require moving the info.compMatchedVM computation there as well.) |
5272 | |
5273 | if (!info.compMatchedVM) |
5274 | { |
5275 | #if defined(_TARGET_ARM_) |
5276 | |
5277 | // Currently nothing needs to be done. There are no ARM flags that conflict with other flags. |
5278 | |
5279 | #endif // defined(_TARGET_ARM_) |
5280 | |
5281 | #if defined(_TARGET_ARM64_) |
5282 | |
5283 | // The x86/x64 architecture capabilities flags overlap with the ARM64 ones. Set a reasonable architecture |
5284 | // target default. Currently this is disabling all ARM64 architecture features except FP and SIMD, but this |
5285 | // should be altered to possibly enable all of them, when they are known to all work. |
5286 | |
5287 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_AES); |
5288 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_ATOMICS); |
5289 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_CRC32); |
5290 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_DCPOP); |
5291 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_DP); |
5292 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_FCMA); |
5293 | compileFlags->Set(JitFlags::JIT_FLAG_HAS_ARM64_FP); |
5294 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_FP16); |
5295 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_JSCVT); |
5296 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_LRCPC); |
5297 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_PMULL); |
5298 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_SHA1); |
5299 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_SHA256); |
5300 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_SHA512); |
5301 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_SHA3); |
5302 | compileFlags->Set(JitFlags::JIT_FLAG_HAS_ARM64_SIMD); |
5303 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_SIMD_V81); |
5304 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_SIMD_FP16); |
5305 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_SM3); |
5306 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_SM4); |
5307 | compileFlags->Clear(JitFlags::JIT_FLAG_HAS_ARM64_SVE); |
5308 | |
5309 | #endif // defined(_TARGET_ARM64_) |
5310 | } |
5311 | |
5312 | compMaxUncheckedOffsetForNullObject = eeGetEEInfo()->maxUncheckedOffsetForNullObject; |
5313 | |
5314 | // Set the context for token lookup. |
5315 | if (compIsForInlining()) |
5316 | { |
5317 | impTokenLookupContextHandle = impInlineInfo->tokenLookupContextHandle; |
5318 | |
5319 | assert(impInlineInfo->inlineCandidateInfo->clsHandle == compHnd->getMethodClass(methodHnd)); |
5320 | info.compClassHnd = impInlineInfo->inlineCandidateInfo->clsHandle; |
5321 | |
5322 | assert(impInlineInfo->inlineCandidateInfo->clsAttr == info.compCompHnd->getClassAttribs(info.compClassHnd)); |
5323 | // printf("%x != %x\n", impInlineInfo->inlineCandidateInfo->clsAttr, |
5324 | // info.compCompHnd->getClassAttribs(info.compClassHnd)); |
5325 | info.compClassAttr = impInlineInfo->inlineCandidateInfo->clsAttr; |
5326 | } |
5327 | else |
5328 | { |
5329 | impTokenLookupContextHandle = MAKE_METHODCONTEXT(info.compMethodHnd); |
5330 | |
5331 | info.compClassHnd = compHnd->getMethodClass(methodHnd); |
5332 | info.compClassAttr = info.compCompHnd->getClassAttribs(info.compClassHnd); |
5333 | } |
5334 | |
5335 | info.compProfilerCallback = false; // Assume false until we are told to hook this method. |
5336 | |
5337 | #if defined(DEBUG) || defined(LATE_DISASM) |
5338 | const char* classNamePtr; |
5339 | |
5340 | info.compMethodName = eeGetMethodName(methodHnd, &classNamePtr); |
5341 | unsigned len = (unsigned)roundUp(strlen(classNamePtr) + 1); |
5342 | info.compClassName = getAllocator(CMK_DebugOnly).allocate<char>(len); |
5343 | strcpy_s((char*)info.compClassName, len, classNamePtr); |
5344 | |
5345 | info.compFullName = eeGetMethodFullName(methodHnd); |
5346 | #endif // defined(DEBUG) || defined(LATE_DISASM) |
5347 | |
5348 | #ifdef DEBUG |
5349 | if (!compIsForInlining()) |
5350 | { |
5351 | JitTls::GetLogEnv()->setCompiler(this); |
5352 | } |
5353 | |
5354 | // Have we been told to be more selective in our Jitting? |
5355 | if (skipMethod()) |
5356 | { |
5357 | if (compIsForInlining()) |
5358 | { |
5359 | compInlineResult->NoteFatal(InlineObservation::CALLEE_MARKED_AS_SKIPPED); |
5360 | } |
5361 | return CORJIT_SKIPPED; |
5362 | } |
5363 | |
5364 | // Opt-in to jit stress based on method hash ranges. |
5365 | // |
5366 | // Note the default (with JitStressRange not set) is that all |
5367 | // methods will be subject to stress. |
5368 | static ConfigMethodRange fJitStressRange; |
5369 | fJitStressRange.EnsureInit(JitConfig.JitStressRange()); |
5370 | assert(!fJitStressRange.Error()); |
5371 | bRangeAllowStress = fJitStressRange.Contains(info.compCompHnd, info.compMethodHnd); |
5372 | |
5373 | #endif // DEBUG |
5374 | |
5375 | // Set this before the first 'BADCODE' |
5376 | // Skip verification where possible |
5377 | tiVerificationNeeded = !compileFlags->IsSet(JitFlags::JIT_FLAG_SKIP_VERIFICATION); |
5378 | |
5379 | assert(!compIsForInlining() || !tiVerificationNeeded); // Inlinees must have been verified. |
5380 | |
5381 | // assume the code is verifiable unless proven otherwise |
5382 | tiIsVerifiableCode = TRUE; |
5383 | |
5384 | tiRuntimeCalloutNeeded = false; |
5385 | |
5386 | CorInfoInstantiationVerification instVerInfo = INSTVER_GENERIC_PASSED_VERIFICATION; |
5387 | |
5388 | if (!compIsForInlining() && tiVerificationNeeded) |
5389 | { |
5390 | instVerInfo = compHnd->isInstantiationOfVerifiedGeneric(methodHnd); |
5391 | |
5392 | if (tiVerificationNeeded && (instVerInfo == INSTVER_GENERIC_FAILED_VERIFICATION)) |
5393 | { |
5394 | CorInfoCanSkipVerificationResult canSkipVerificationResult = |
5395 | info.compCompHnd->canSkipMethodVerification(info.compMethodHnd); |
5396 | |
5397 | switch (canSkipVerificationResult) |
5398 | { |
5399 | case CORINFO_VERIFICATION_CANNOT_SKIP: |
5400 | // We cannot verify concrete instantiation. |
5401 | // We can only verify the typical/open instantiation |
5402 | // The VM should throw a VerificationException instead of allowing this. |
5403 | NO_WAY("Verification of closed instantiations is not supported" ); |
5404 | break; |
5405 | |
5406 | case CORINFO_VERIFICATION_CAN_SKIP: |
5407 | // The VM should first verify the open instantiation. If unverifiable code |
5408 | // is detected, it should pass in JitFlags::JIT_FLAG_SKIP_VERIFICATION. |
5409 | assert(!"The VM should have used JitFlags::JIT_FLAG_SKIP_VERIFICATION" ); |
5410 | tiVerificationNeeded = false; |
5411 | break; |
5412 | |
5413 | case CORINFO_VERIFICATION_RUNTIME_CHECK: |
5414 | // This is a concrete generic instantiation with unverifiable code, that also |
5415 | // needs a runtime callout. |
5416 | tiVerificationNeeded = false; |
5417 | tiRuntimeCalloutNeeded = true; |
5418 | break; |
5419 | |
5420 | case CORINFO_VERIFICATION_DONT_JIT: |
5421 | // We cannot verify concrete instantiation. |
5422 | // We can only verify the typical/open instantiation |
5423 | // The VM should throw a VerificationException instead of allowing this. |
5424 | BADCODE("NGEN of unverifiable transparent code is not supported" ); |
5425 | break; |
5426 | } |
5427 | } |
5428 | |
5429 | // load any constraints for verification, noting any cycles to be rejected by the verifying importer |
5430 | if (tiVerificationNeeded) |
5431 | { |
5432 | compHnd->initConstraintsForVerification(methodHnd, &info.hasCircularClassConstraints, |
5433 | &info.hasCircularMethodConstraints); |
5434 | } |
5435 | } |
5436 | |
5437 | /* Setup an error trap */ |
5438 | |
5439 | struct Param |
5440 | { |
5441 | Compiler* pThis; |
5442 | |
5443 | CORINFO_MODULE_HANDLE classPtr; |
5444 | COMP_HANDLE compHnd; |
5445 | CORINFO_METHOD_INFO* methodInfo; |
5446 | void** methodCodePtr; |
5447 | ULONG* methodCodeSize; |
5448 | JitFlags* compileFlags; |
5449 | |
5450 | CorInfoInstantiationVerification instVerInfo; |
5451 | int result; |
5452 | } param; |
5453 | param.pThis = this; |
5454 | param.classPtr = classPtr; |
5455 | param.compHnd = compHnd; |
5456 | param.methodInfo = methodInfo; |
5457 | param.methodCodePtr = methodCodePtr; |
5458 | param.methodCodeSize = methodCodeSize; |
5459 | param.compileFlags = compileFlags; |
5460 | param.instVerInfo = instVerInfo; |
5461 | param.result = CORJIT_INTERNALERROR; |
5462 | |
5463 | setErrorTrap(compHnd, Param*, pParam, ¶m) // ERROR TRAP: Start normal block |
5464 | { |
5465 | pParam->result = pParam->pThis->compCompileHelper(pParam->classPtr, pParam->compHnd, pParam->methodInfo, |
5466 | pParam->methodCodePtr, pParam->methodCodeSize, |
5467 | pParam->compileFlags, pParam->instVerInfo); |
5468 | } |
5469 | finallyErrorTrap() // ERROR TRAP: The following block handles errors |
5470 | { |
5471 | /* Cleanup */ |
5472 | |
5473 | if (compIsForInlining()) |
5474 | { |
5475 | goto DoneCleanUp; |
5476 | } |
5477 | |
5478 | /* Tell the emitter that we're done with this function */ |
5479 | |
5480 | genEmitter->emitEndCG(); |
5481 | |
5482 | DoneCleanUp: |
5483 | compDone(); |
5484 | } |
5485 | endErrorTrap() // ERROR TRAP: End |
5486 | |
5487 | return param.result; |
5488 | } |
5489 | |
5490 | #if defined(DEBUG) || defined(INLINE_DATA) |
5491 | unsigned Compiler::Info::compMethodHash() const |
5492 | { |
5493 | if (compMethodHashPrivate == 0) |
5494 | { |
5495 | compMethodHashPrivate = compCompHnd->getMethodHash(compMethodHnd); |
5496 | } |
5497 | return compMethodHashPrivate; |
5498 | } |
5499 | #endif // defined(DEBUG) || defined(INLINE_DATA) |
5500 | |
5501 | void Compiler::compCompileFinish() |
5502 | { |
5503 | #if defined(DEBUG) || MEASURE_NODE_SIZE || MEASURE_BLOCK_SIZE || DISPLAY_SIZES || CALL_ARG_STATS |
5504 | genMethodCnt++; |
5505 | #endif |
5506 | |
5507 | #if MEASURE_MEM_ALLOC |
5508 | { |
5509 | compArenaAllocator->finishMemStats(); |
5510 | memAllocHist.record((unsigned)((compArenaAllocator->getTotalBytesAllocated() + 1023) / 1024)); |
5511 | memUsedHist.record((unsigned)((compArenaAllocator->getTotalBytesUsed() + 1023) / 1024)); |
5512 | } |
5513 | |
5514 | #ifdef DEBUG |
5515 | if (s_dspMemStats || verbose) |
5516 | { |
5517 | printf("\nAllocations for %s (MethodHash=%08x)\n" , info.compFullName, info.compMethodHash()); |
5518 | compArenaAllocator->dumpMemStats(jitstdout); |
5519 | } |
5520 | #endif // DEBUG |
5521 | #endif // MEASURE_MEM_ALLOC |
5522 | |
5523 | #if LOOP_HOIST_STATS |
5524 | AddLoopHoistStats(); |
5525 | #endif // LOOP_HOIST_STATS |
5526 | |
5527 | #if MEASURE_NODE_SIZE |
5528 | genTreeNcntHist.record(static_cast<unsigned>(genNodeSizeStatsPerFunc.genTreeNodeCnt)); |
5529 | genTreeNsizHist.record(static_cast<unsigned>(genNodeSizeStatsPerFunc.genTreeNodeSize)); |
5530 | #endif |
5531 | |
5532 | #if defined(DEBUG) |
5533 | // Small methods should fit in ArenaAllocator::getDefaultPageSize(), or else |
5534 | // we should bump up ArenaAllocator::getDefaultPageSize() |
5535 | |
5536 | if ((info.compILCodeSize <= 32) && // Is it a reasonably small method? |
5537 | (info.compNativeCodeSize < 512) && // Some trivial methods generate huge native code. eg. pushing a single huge |
5538 | // struct |
5539 | (impInlinedCodeSize <= 128) && // Is the the inlining reasonably bounded? |
5540 | // Small methods cannot meaningfully have a big number of locals |
5541 | // or arguments. We always track arguments at the start of |
5542 | // the prolog which requires memory |
5543 | (info.compLocalsCount <= 32) && (!opts.MinOpts()) && // We may have too many local variables, etc |
5544 | (getJitStressLevel() == 0) && // We need extra memory for stress |
5545 | !opts.optRepeat && // We need extra memory to repeat opts |
5546 | !compArenaAllocator->bypassHostAllocator() && // ArenaAllocator::getDefaultPageSize() is artificially low for |
5547 | // DirectAlloc |
5548 | // Factor of 2x is because data-structures are bigger under DEBUG |
5549 | (compArenaAllocator->getTotalBytesAllocated() > (2 * ArenaAllocator::getDefaultPageSize())) && |
5550 | // RyuJIT backend needs memory tuning! TODO-Cleanup: remove this case when memory tuning is complete. |
5551 | (compArenaAllocator->getTotalBytesAllocated() > (10 * ArenaAllocator::getDefaultPageSize())) && |
5552 | !verbose) // We allocate lots of memory to convert sets to strings for JitDump |
5553 | { |
5554 | genSmallMethodsNeedingExtraMemoryCnt++; |
5555 | |
5556 | // Less than 1% of all methods should run into this. |
5557 | // We cannot be more strict as there are always degenerate cases where we |
5558 | // would need extra memory (like huge structs as locals - see lvaSetStruct()). |
5559 | assert((genMethodCnt < 500) || (genSmallMethodsNeedingExtraMemoryCnt < (genMethodCnt / 100))); |
5560 | } |
5561 | #endif // DEBUG |
5562 | |
5563 | #if defined(DEBUG) || defined(INLINE_DATA) |
5564 | |
5565 | m_inlineStrategy->DumpData(); |
5566 | m_inlineStrategy->DumpXml(); |
5567 | |
5568 | #endif |
5569 | |
5570 | #ifdef DEBUG |
5571 | if (opts.dspOrder) |
5572 | { |
5573 | // mdMethodDef __stdcall CEEInfo::getMethodDefFromMethod(CORINFO_METHOD_HANDLE hMethod) |
5574 | mdMethodDef currentMethodToken = info.compCompHnd->getMethodDefFromMethod(info.compMethodHnd); |
5575 | |
5576 | unsigned profCallCount = 0; |
5577 | if (opts.jitFlags->IsSet(JitFlags::JIT_FLAG_BBOPT) && fgHaveProfileData()) |
5578 | { |
5579 | assert(fgProfileBuffer[0].ILOffset == 0); |
5580 | profCallCount = fgProfileBuffer[0].ExecutionCount; |
5581 | } |
5582 | |
5583 | static bool = false; |
5584 | if (!headerPrinted) |
5585 | { |
5586 | // clang-format off |
5587 | headerPrinted = true; |
5588 | printf(" | Profiled | Exec- | Method has | calls | Num |LclV |AProp| CSE | Reg |bytes | %3s code size | \n" , Target::g_tgtCPUName); |
5589 | printf(" mdToken | | RGN | Count | EH | FRM | LOOP | NRM | IND | BBs | Cnt | Cnt | Cnt | Alloc | IL | HOT | COLD | method name \n" ); |
5590 | printf("---------+-----+------+----------+----+-----+------+-----+-----+-----+-----+-----+-----+---------+------+-------+-------+-----------\n" ); |
5591 | // 06001234 | PRF | HOT | 219 | EH | ebp | LOOP | 15 | 6 | 12 | 17 | 12 | 8 | 28 p2 | 145 | 211 | 123 | System.Example(int) |
5592 | // clang-format on |
5593 | } |
5594 | |
5595 | printf("%08X | " , currentMethodToken); |
5596 | |
5597 | CorInfoRegionKind regionKind = info.compMethodInfo->regionKind; |
5598 | |
5599 | if (opts.altJit) |
5600 | { |
5601 | printf("ALT | " ); |
5602 | } |
5603 | else if (fgHaveProfileData()) |
5604 | { |
5605 | printf("PRF | " ); |
5606 | } |
5607 | else |
5608 | { |
5609 | printf(" | " ); |
5610 | } |
5611 | |
5612 | if (regionKind == CORINFO_REGION_NONE) |
5613 | { |
5614 | printf(" | " ); |
5615 | } |
5616 | else if (regionKind == CORINFO_REGION_HOT) |
5617 | { |
5618 | printf(" HOT | " ); |
5619 | } |
5620 | else if (regionKind == CORINFO_REGION_COLD) |
5621 | { |
5622 | printf("COLD | " ); |
5623 | } |
5624 | else if (regionKind == CORINFO_REGION_JIT) |
5625 | { |
5626 | printf(" JIT | " ); |
5627 | } |
5628 | else |
5629 | { |
5630 | printf("UNKN | " ); |
5631 | } |
5632 | |
5633 | printf("%8d | " , profCallCount); |
5634 | |
5635 | if (compHndBBtabCount > 0) |
5636 | { |
5637 | printf("EH | " ); |
5638 | } |
5639 | else |
5640 | { |
5641 | printf(" | " ); |
5642 | } |
5643 | |
5644 | if (rpFrameType == FT_EBP_FRAME) |
5645 | { |
5646 | printf("%3s | " , STR_FPBASE); |
5647 | } |
5648 | else if (rpFrameType == FT_ESP_FRAME) |
5649 | { |
5650 | printf("%3s | " , STR_SPBASE); |
5651 | } |
5652 | #if DOUBLE_ALIGN |
5653 | else if (rpFrameType == FT_DOUBLE_ALIGN_FRAME) |
5654 | { |
5655 | printf("dbl | " ); |
5656 | } |
5657 | #endif |
5658 | else // (rpFrameType == FT_NOT_SET) |
5659 | { |
5660 | printf("??? | " ); |
5661 | } |
5662 | |
5663 | if (fgHasLoops) |
5664 | { |
5665 | printf("LOOP |" ); |
5666 | } |
5667 | else |
5668 | { |
5669 | printf(" |" ); |
5670 | } |
5671 | |
5672 | printf(" %3d |" , optCallCount); |
5673 | printf(" %3d |" , optIndirectCallCount); |
5674 | printf(" %3d |" , fgBBcountAtCodegen); |
5675 | printf(" %3d |" , lvaCount); |
5676 | |
5677 | if (opts.MinOpts()) |
5678 | { |
5679 | printf(" MinOpts |" ); |
5680 | } |
5681 | else |
5682 | { |
5683 | printf(" %3d |" , optAssertionCount); |
5684 | #if FEATURE_ANYCSE |
5685 | printf(" %3d |" , optCSEcount); |
5686 | #else |
5687 | printf(" %3d |" , 0); |
5688 | #endif // FEATURE_ANYCSE |
5689 | } |
5690 | |
5691 | printf(" LSRA |" ); // TODO-Cleanup: dump some interesting LSRA stat into the order file? |
5692 | printf(" %4d |" , info.compMethodInfo->ILCodeSize); |
5693 | printf(" %5d |" , info.compTotalHotCodeSize); |
5694 | printf(" %5d |" , info.compTotalColdCodeSize); |
5695 | |
5696 | printf(" %s\n" , eeGetMethodFullName(info.compMethodHnd)); |
5697 | printf("" ); // in our logic this causes a flush |
5698 | } |
5699 | |
5700 | if (verbose) |
5701 | { |
5702 | printf("****** DONE compiling %s\n" , info.compFullName); |
5703 | printf("" ); // in our logic this causes a flush |
5704 | } |
5705 | |
5706 | // Only call _DbgBreakCheck when we are jitting, not when we are ngen-ing |
5707 | // For ngen the int3 or breakpoint instruction will be right at the |
5708 | // start of the ngen method and we will stop when we execute it. |
5709 | // |
5710 | if (!opts.jitFlags->IsSet(JitFlags::JIT_FLAG_PREJIT)) |
5711 | { |
5712 | if (compJitHaltMethod()) |
5713 | { |
5714 | #if !defined(_HOST_UNIX_) |
5715 | // TODO-UNIX: re-enable this when we have an OS that supports a pop-up dialog |
5716 | |
5717 | // Don't do an assert, but just put up the dialog box so we get just-in-time debugger |
5718 | // launching. When you hit 'retry' it will continue and naturally stop at the INT 3 |
5719 | // that the JIT put in the code |
5720 | _DbgBreakCheck(__FILE__, __LINE__, "JitHalt" ); |
5721 | #endif |
5722 | } |
5723 | } |
5724 | #endif // DEBUG |
5725 | } |
5726 | |
5727 | #ifdef PSEUDORANDOM_NOP_INSERTION |
5728 | // this is zlib adler32 checksum. source came from windows base |
5729 | |
5730 | #define BASE 65521L // largest prime smaller than 65536 |
5731 | #define NMAX 5552 |
5732 | // NMAX is the largest n such that 255n(n+1)/2 + (n+1)(BASE-1) <= 2^32-1 |
5733 | |
5734 | #define DO1(buf, i) \ |
5735 | { \ |
5736 | s1 += buf[i]; \ |
5737 | s2 += s1; \ |
5738 | } |
5739 | #define DO2(buf, i) \ |
5740 | DO1(buf, i); \ |
5741 | DO1(buf, i + 1); |
5742 | #define DO4(buf, i) \ |
5743 | DO2(buf, i); \ |
5744 | DO2(buf, i + 2); |
5745 | #define DO8(buf, i) \ |
5746 | DO4(buf, i); \ |
5747 | DO4(buf, i + 4); |
5748 | #define DO16(buf) \ |
5749 | DO8(buf, 0); \ |
5750 | DO8(buf, 8); |
5751 | |
5752 | unsigned adler32(unsigned adler, char* buf, unsigned int len) |
5753 | { |
5754 | unsigned int s1 = adler & 0xffff; |
5755 | unsigned int s2 = (adler >> 16) & 0xffff; |
5756 | int k; |
5757 | |
5758 | if (buf == NULL) |
5759 | return 1L; |
5760 | |
5761 | while (len > 0) |
5762 | { |
5763 | k = len < NMAX ? len : NMAX; |
5764 | len -= k; |
5765 | while (k >= 16) |
5766 | { |
5767 | DO16(buf); |
5768 | buf += 16; |
5769 | k -= 16; |
5770 | } |
5771 | if (k != 0) |
5772 | do |
5773 | { |
5774 | s1 += *buf++; |
5775 | s2 += s1; |
5776 | } while (--k); |
5777 | s1 %= BASE; |
5778 | s2 %= BASE; |
5779 | } |
5780 | return (s2 << 16) | s1; |
5781 | } |
5782 | #endif |
5783 | |
5784 | unsigned getMethodBodyChecksum(__in_z char* code, int size) |
5785 | { |
5786 | #ifdef PSEUDORANDOM_NOP_INSERTION |
5787 | return adler32(0, code, size); |
5788 | #else |
5789 | return 0; |
5790 | #endif |
5791 | } |
5792 | |
5793 | int Compiler::compCompileHelper(CORINFO_MODULE_HANDLE classPtr, |
5794 | COMP_HANDLE compHnd, |
5795 | CORINFO_METHOD_INFO* methodInfo, |
5796 | void** methodCodePtr, |
5797 | ULONG* methodCodeSize, |
5798 | JitFlags* compileFlags, |
5799 | CorInfoInstantiationVerification instVerInfo) |
5800 | { |
5801 | CORINFO_METHOD_HANDLE methodHnd = info.compMethodHnd; |
5802 | |
5803 | info.compCode = methodInfo->ILCode; |
5804 | info.compILCodeSize = methodInfo->ILCodeSize; |
5805 | |
5806 | if (info.compILCodeSize == 0) |
5807 | { |
5808 | BADCODE("code size is zero" ); |
5809 | } |
5810 | |
5811 | if (compIsForInlining()) |
5812 | { |
5813 | #ifdef DEBUG |
5814 | unsigned methAttr_Old = impInlineInfo->inlineCandidateInfo->methAttr; |
5815 | unsigned methAttr_New = info.compCompHnd->getMethodAttribs(info.compMethodHnd); |
5816 | unsigned flagsToIgnore = CORINFO_FLG_DONT_INLINE | CORINFO_FLG_FORCEINLINE; |
5817 | assert((methAttr_Old & (~flagsToIgnore)) == (methAttr_New & (~flagsToIgnore))); |
5818 | #endif |
5819 | |
5820 | info.compFlags = impInlineInfo->inlineCandidateInfo->methAttr; |
5821 | } |
5822 | else |
5823 | { |
5824 | info.compFlags = info.compCompHnd->getMethodAttribs(info.compMethodHnd); |
5825 | #ifdef PSEUDORANDOM_NOP_INSERTION |
5826 | info.compChecksum = getMethodBodyChecksum((char*)methodInfo->ILCode, methodInfo->ILCodeSize); |
5827 | #endif |
5828 | } |
5829 | |
5830 | // compInitOptions will set the correct verbose flag. |
5831 | |
5832 | compInitOptions(compileFlags); |
5833 | |
5834 | #ifdef ALT_JIT |
5835 | if (!compIsForInlining() && !opts.altJit) |
5836 | { |
5837 | // We're an altjit, but the COMPlus_AltJit configuration did not say to compile this method, |
5838 | // so skip it. |
5839 | return CORJIT_SKIPPED; |
5840 | } |
5841 | #endif // ALT_JIT |
5842 | |
5843 | #ifdef DEBUG |
5844 | |
5845 | if (verbose) |
5846 | { |
5847 | printf("IL to import:\n" ); |
5848 | dumpILRange(info.compCode, info.compILCodeSize); |
5849 | } |
5850 | |
5851 | #endif |
5852 | |
5853 | // Check for COMPlus_AggressiveInlining |
5854 | if (JitConfig.JitAggressiveInlining()) |
5855 | { |
5856 | compDoAggressiveInlining = true; |
5857 | } |
5858 | |
5859 | if (compDoAggressiveInlining) |
5860 | { |
5861 | info.compFlags |= CORINFO_FLG_FORCEINLINE; |
5862 | } |
5863 | |
5864 | #ifdef DEBUG |
5865 | |
5866 | // Check for ForceInline stress. |
5867 | if (compStressCompile(STRESS_FORCE_INLINE, 0)) |
5868 | { |
5869 | info.compFlags |= CORINFO_FLG_FORCEINLINE; |
5870 | } |
5871 | |
5872 | if (compIsForInlining()) |
5873 | { |
5874 | JITLOG((LL_INFO100000, "\nINLINER impTokenLookupContextHandle for %s is 0x%p.\n" , |
5875 | eeGetMethodFullName(info.compMethodHnd), dspPtr(impTokenLookupContextHandle))); |
5876 | } |
5877 | |
5878 | // Force verification if asked to do so |
5879 | if (JitConfig.JitForceVer()) |
5880 | { |
5881 | tiVerificationNeeded = (instVerInfo == INSTVER_NOT_INSTANTIATION); |
5882 | } |
5883 | |
5884 | if (tiVerificationNeeded) |
5885 | { |
5886 | JITLOG((LL_INFO10000, "tiVerificationNeeded initially set to true for %s\n" , info.compFullName)); |
5887 | } |
5888 | #endif // DEBUG |
5889 | |
5890 | /* Since tiVerificationNeeded can be turned off in the middle of |
5891 | compiling a method, and it might have caused blocks to be queued up |
5892 | for reimporting, impCanReimport can be used to check for reimporting. */ |
5893 | |
5894 | impCanReimport = (tiVerificationNeeded || compStressCompile(STRESS_CHK_REIMPORT, 15)); |
5895 | |
5896 | // Need security prolog/epilog callouts when there is a declarative security in the method. |
5897 | tiSecurityCalloutNeeded = ((info.compFlags & CORINFO_FLG_NOSECURITYWRAP) == 0); |
5898 | |
5899 | if (tiSecurityCalloutNeeded || (info.compFlags & CORINFO_FLG_SECURITYCHECK)) |
5900 | { |
5901 | // We need to allocate the security object on the stack |
5902 | // when the method being compiled has a declarative security |
5903 | // (i.e. when CORINFO_FLG_NOSECURITYWRAP is reset for the current method). |
5904 | // This is also the case when we inject a prolog and epilog in the method. |
5905 | opts.compNeedSecurityCheck = true; |
5906 | } |
5907 | |
5908 | /* Initialize set a bunch of global values */ |
5909 | |
5910 | info.compScopeHnd = classPtr; |
5911 | info.compXcptnsCount = methodInfo->EHcount; |
5912 | info.compMaxStack = methodInfo->maxStack; |
5913 | compHndBBtab = nullptr; |
5914 | compHndBBtabCount = 0; |
5915 | compHndBBtabAllocCount = 0; |
5916 | |
5917 | info.compNativeCodeSize = 0; |
5918 | info.compTotalHotCodeSize = 0; |
5919 | info.compTotalColdCodeSize = 0; |
5920 | |
5921 | #ifdef DEBUG |
5922 | compCurBB = nullptr; |
5923 | lvaTable = nullptr; |
5924 | |
5925 | // Reset node and block ID counter |
5926 | compGenTreeID = 0; |
5927 | compBasicBlockID = 0; |
5928 | #endif |
5929 | |
5930 | /* Initialize emitter */ |
5931 | |
5932 | if (!compIsForInlining()) |
5933 | { |
5934 | codeGen->getEmitter()->emitBegCG(this, compHnd); |
5935 | } |
5936 | |
5937 | info.compIsStatic = (info.compFlags & CORINFO_FLG_STATIC) != 0; |
5938 | |
5939 | info.compIsContextful = (info.compClassAttr & CORINFO_FLG_CONTEXTFUL) != 0; |
5940 | |
5941 | info.compPublishStubParam = opts.jitFlags->IsSet(JitFlags::JIT_FLAG_PUBLISH_SECRET_PARAM); |
5942 | |
5943 | switch (methodInfo->args.getCallConv()) |
5944 | { |
5945 | case CORINFO_CALLCONV_VARARG: |
5946 | case CORINFO_CALLCONV_NATIVEVARARG: |
5947 | info.compIsVarArgs = true; |
5948 | break; |
5949 | case CORINFO_CALLCONV_DEFAULT: |
5950 | info.compIsVarArgs = false; |
5951 | break; |
5952 | default: |
5953 | BADCODE("bad calling convention" ); |
5954 | } |
5955 | info.compRetNativeType = info.compRetType = JITtype2varType(methodInfo->args.retType); |
5956 | |
5957 | info.compCallUnmanaged = 0; |
5958 | info.compLvFrameListRoot = BAD_VAR_NUM; |
5959 | |
5960 | info.compInitMem = ((methodInfo->options & CORINFO_OPT_INIT_LOCALS) != 0); |
5961 | |
5962 | /* Allocate the local variable table */ |
5963 | |
5964 | lvaInitTypeRef(); |
5965 | |
5966 | if (!compIsForInlining()) |
5967 | { |
5968 | compInitDebuggingInfo(); |
5969 | } |
5970 | |
5971 | #ifdef DEBUG |
5972 | if (compIsForInlining()) |
5973 | { |
5974 | compBasicBlockID = impInlineInfo->InlinerCompiler->compBasicBlockID; |
5975 | } |
5976 | #endif |
5977 | |
5978 | const bool forceInline = !!(info.compFlags & CORINFO_FLG_FORCEINLINE); |
5979 | |
5980 | if (!compIsForInlining() && opts.jitFlags->IsSet(JitFlags::JIT_FLAG_PREJIT)) |
5981 | { |
5982 | // We're prejitting the root method. We also will analyze it as |
5983 | // a potential inline candidate. |
5984 | InlineResult prejitResult(this, methodHnd, "prejit" ); |
5985 | |
5986 | // Do the initial inline screen. |
5987 | impCanInlineIL(methodHnd, methodInfo, forceInline, &prejitResult); |
5988 | |
5989 | // Temporarily install the prejitResult as the |
5990 | // compInlineResult so it's available to fgFindJumpTargets |
5991 | // and can accumulate more observations as the IL is |
5992 | // scanned. |
5993 | // |
5994 | // We don't pass prejitResult in as a parameter to avoid |
5995 | // potential aliasing confusion -- the other call to |
5996 | // fgFindBasicBlocks may have set up compInlineResult and |
5997 | // the code in fgFindJumpTargets references that data |
5998 | // member extensively. |
5999 | assert(compInlineResult == nullptr); |
6000 | assert(impInlineInfo == nullptr); |
6001 | compInlineResult = &prejitResult; |
6002 | |
6003 | // Find the basic blocks. We must do this regardless of |
6004 | // inlineability, since we are prejitting this method. |
6005 | // |
6006 | // This will also update the status of this method as |
6007 | // an inline candidate. |
6008 | fgFindBasicBlocks(); |
6009 | |
6010 | // Undo the temporary setup. |
6011 | assert(compInlineResult == &prejitResult); |
6012 | compInlineResult = nullptr; |
6013 | |
6014 | // If still a viable, discretionary inline, assess |
6015 | // profitability. |
6016 | if (prejitResult.IsDiscretionaryCandidate()) |
6017 | { |
6018 | prejitResult.DetermineProfitability(methodInfo); |
6019 | } |
6020 | |
6021 | m_inlineStrategy->NotePrejitDecision(prejitResult); |
6022 | |
6023 | // Handle the results of the inline analysis. |
6024 | if (prejitResult.IsFailure()) |
6025 | { |
6026 | // This method is a bad inlinee according to our |
6027 | // analysis. We will let the InlineResult destructor |
6028 | // mark it as noinline in the prejit image to save the |
6029 | // jit some work. |
6030 | // |
6031 | // This decision better not be context-dependent. |
6032 | assert(prejitResult.IsNever()); |
6033 | } |
6034 | else |
6035 | { |
6036 | // This looks like a viable inline candidate. Since |
6037 | // we're not actually inlining, don't report anything. |
6038 | prejitResult.SetReported(); |
6039 | } |
6040 | } |
6041 | else |
6042 | { |
6043 | // We are jitting the root method, or inlining. |
6044 | fgFindBasicBlocks(); |
6045 | } |
6046 | |
6047 | // If we're inlining and the candidate is bad, bail out. |
6048 | if (compDonotInline()) |
6049 | { |
6050 | goto _Next; |
6051 | } |
6052 | |
6053 | compSetOptimizationLevel(); |
6054 | |
6055 | #if COUNT_BASIC_BLOCKS |
6056 | bbCntTable.record(fgBBcount); |
6057 | |
6058 | if (fgBBcount == 1) |
6059 | { |
6060 | bbOneBBSizeTable.record(methodInfo->ILCodeSize); |
6061 | } |
6062 | #endif // COUNT_BASIC_BLOCKS |
6063 | |
6064 | #ifdef DEBUG |
6065 | if (verbose) |
6066 | { |
6067 | printf("Basic block list for '%s'\n" , info.compFullName); |
6068 | fgDispBasicBlocks(); |
6069 | } |
6070 | #endif |
6071 | |
6072 | #ifdef DEBUG |
6073 | /* Give the function a unique number */ |
6074 | |
6075 | if (opts.disAsm || opts.dspEmit || verbose) |
6076 | { |
6077 | s_compMethodsCount = ~info.compMethodHash() & 0xffff; |
6078 | } |
6079 | else |
6080 | { |
6081 | s_compMethodsCount++; |
6082 | } |
6083 | #endif |
6084 | |
6085 | if (compIsForInlining()) |
6086 | { |
6087 | compInlineResult->NoteInt(InlineObservation::CALLEE_NUMBER_OF_BASIC_BLOCKS, fgBBcount); |
6088 | |
6089 | if (compInlineResult->IsFailure()) |
6090 | { |
6091 | goto _Next; |
6092 | } |
6093 | } |
6094 | |
6095 | #ifdef DEBUG |
6096 | if (JitConfig.DumpJittedMethods() == 1 && !compIsForInlining()) |
6097 | { |
6098 | printf("Compiling %4d %s::%s, IL size = %u, hsh=0x%x\n" , Compiler::jitTotalMethodCompiled, info.compClassName, |
6099 | info.compMethodName, info.compILCodeSize, info.compMethodHash()); |
6100 | } |
6101 | if (compIsForInlining()) |
6102 | { |
6103 | compGenTreeID = impInlineInfo->InlinerCompiler->compGenTreeID; |
6104 | } |
6105 | #endif |
6106 | |
6107 | compCompile(methodCodePtr, methodCodeSize, compileFlags); |
6108 | |
6109 | #ifdef DEBUG |
6110 | if (compIsForInlining()) |
6111 | { |
6112 | impInlineInfo->InlinerCompiler->compGenTreeID = compGenTreeID; |
6113 | impInlineInfo->InlinerCompiler->compBasicBlockID = compBasicBlockID; |
6114 | } |
6115 | #endif |
6116 | |
6117 | _Next: |
6118 | |
6119 | if (compDonotInline()) |
6120 | { |
6121 | // Verify we have only one inline result in play. |
6122 | assert(impInlineInfo->inlineResult == compInlineResult); |
6123 | } |
6124 | |
6125 | if (!compIsForInlining()) |
6126 | { |
6127 | compCompileFinish(); |
6128 | |
6129 | // Did we just compile for a target architecture that the VM isn't expecting? If so, the VM |
6130 | // can't used the generated code (and we better be an AltJit!). |
6131 | |
6132 | if (!info.compMatchedVM) |
6133 | { |
6134 | return CORJIT_SKIPPED; |
6135 | } |
6136 | |
6137 | #ifdef ALT_JIT |
6138 | #ifdef DEBUG |
6139 | if (JitConfig.RunAltJitCode() == 0) |
6140 | { |
6141 | return CORJIT_SKIPPED; |
6142 | } |
6143 | #endif // DEBUG |
6144 | #endif // ALT_JIT |
6145 | } |
6146 | |
6147 | /* Success! */ |
6148 | return CORJIT_OK; |
6149 | } |
6150 | |
6151 | //------------------------------------------------------------------------ |
6152 | // compFindLocalVarLinear: Linear search for variable's scope containing offset. |
6153 | // |
6154 | // Arguments: |
6155 | // varNum The variable number to search for in the array of scopes. |
6156 | // offs The offset value which should occur within the life of the variable. |
6157 | // |
6158 | // Return Value: |
6159 | // VarScopeDsc* of a matching variable that contains the offset within its life |
6160 | // begin and life end or nullptr when there is no match found. |
6161 | // |
6162 | // Description: |
6163 | // Linear search for matching variables with their life begin and end containing |
6164 | // the offset. |
6165 | // or NULL if one couldn't be found. |
6166 | // |
6167 | // Note: |
6168 | // Usually called for scope count = 4. Could be called for values upto 8. |
6169 | // |
6170 | VarScopeDsc* Compiler::compFindLocalVarLinear(unsigned varNum, unsigned offs) |
6171 | { |
6172 | for (unsigned i = 0; i < info.compVarScopesCount; i++) |
6173 | { |
6174 | VarScopeDsc* dsc = &info.compVarScopes[i]; |
6175 | if ((dsc->vsdVarNum == varNum) && (dsc->vsdLifeBeg <= offs) && (dsc->vsdLifeEnd > offs)) |
6176 | { |
6177 | return dsc; |
6178 | } |
6179 | } |
6180 | return nullptr; |
6181 | } |
6182 | |
6183 | //------------------------------------------------------------------------ |
6184 | // compFindLocalVar: Search for variable's scope containing offset. |
6185 | // |
6186 | // Arguments: |
6187 | // varNum The variable number to search for in the array of scopes. |
6188 | // offs The offset value which should occur within the life of the variable. |
6189 | // |
6190 | // Return Value: |
6191 | // VarScopeDsc* of a matching variable that contains the offset within its life |
6192 | // begin and life end. |
6193 | // or NULL if one couldn't be found. |
6194 | // |
6195 | // Description: |
6196 | // Linear search for matching variables with their life begin and end containing |
6197 | // the offset only when the scope count is < MAX_LINEAR_FIND_LCL_SCOPELIST, |
6198 | // else use the hashtable lookup. |
6199 | // |
6200 | VarScopeDsc* Compiler::compFindLocalVar(unsigned varNum, unsigned offs) |
6201 | { |
6202 | if (info.compVarScopesCount < MAX_LINEAR_FIND_LCL_SCOPELIST) |
6203 | { |
6204 | return compFindLocalVarLinear(varNum, offs); |
6205 | } |
6206 | else |
6207 | { |
6208 | VarScopeDsc* ret = compFindLocalVar(varNum, offs, offs); |
6209 | assert(ret == compFindLocalVarLinear(varNum, offs)); |
6210 | return ret; |
6211 | } |
6212 | } |
6213 | |
6214 | //------------------------------------------------------------------------ |
6215 | // compFindLocalVar: Search for variable's scope containing offset. |
6216 | // |
6217 | // Arguments: |
6218 | // varNum The variable number to search for in the array of scopes. |
6219 | // lifeBeg The life begin of the variable's scope |
6220 | // lifeEnd The life end of the variable's scope |
6221 | // |
6222 | // Return Value: |
6223 | // VarScopeDsc* of a matching variable that contains the offset within its life |
6224 | // begin and life end, or NULL if one couldn't be found. |
6225 | // |
6226 | // Description: |
6227 | // Following are the steps used: |
6228 | // 1. Index into the hashtable using varNum. |
6229 | // 2. Iterate through the linked list at index varNum to find a matching |
6230 | // var scope. |
6231 | // |
6232 | VarScopeDsc* Compiler::compFindLocalVar(unsigned varNum, unsigned lifeBeg, unsigned lifeEnd) |
6233 | { |
6234 | assert(compVarScopeMap != nullptr); |
6235 | |
6236 | VarScopeMapInfo* info; |
6237 | if (compVarScopeMap->Lookup(varNum, &info)) |
6238 | { |
6239 | VarScopeListNode* list = info->head; |
6240 | while (list != nullptr) |
6241 | { |
6242 | if ((list->data->vsdLifeBeg <= lifeBeg) && (list->data->vsdLifeEnd > lifeEnd)) |
6243 | { |
6244 | return list->data; |
6245 | } |
6246 | list = list->next; |
6247 | } |
6248 | } |
6249 | return nullptr; |
6250 | } |
6251 | |
6252 | //------------------------------------------------------------------------- |
6253 | // compInitVarScopeMap: Create a scope map so it can be looked up by varNum |
6254 | // |
6255 | // Description: |
6256 | // Map.K => Map.V :: varNum => List(ScopeDsc) |
6257 | // |
6258 | // Create a scope map that can be indexed by varNum and can be iterated |
6259 | // on it's values to look for matching scope when given an offs or |
6260 | // lifeBeg and lifeEnd. |
6261 | // |
6262 | // Notes: |
6263 | // 1. Build the map only when we think linear search is slow, i.e., |
6264 | // MAX_LINEAR_FIND_LCL_SCOPELIST is large. |
6265 | // 2. Linked list preserves original array order. |
6266 | // |
6267 | void Compiler::compInitVarScopeMap() |
6268 | { |
6269 | if (info.compVarScopesCount < MAX_LINEAR_FIND_LCL_SCOPELIST) |
6270 | { |
6271 | return; |
6272 | } |
6273 | |
6274 | assert(compVarScopeMap == nullptr); |
6275 | |
6276 | compVarScopeMap = new (getAllocator()) VarNumToScopeDscMap(getAllocator()); |
6277 | |
6278 | // 599 prime to limit huge allocations; for ex: duplicated scopes on single var. |
6279 | compVarScopeMap->Reallocate(min(info.compVarScopesCount, 599)); |
6280 | |
6281 | for (unsigned i = 0; i < info.compVarScopesCount; ++i) |
6282 | { |
6283 | unsigned varNum = info.compVarScopes[i].vsdVarNum; |
6284 | |
6285 | VarScopeListNode* node = VarScopeListNode::Create(&info.compVarScopes[i], getAllocator()); |
6286 | |
6287 | // Index by varNum and if the list exists append "node" to the "list". |
6288 | VarScopeMapInfo* info; |
6289 | if (compVarScopeMap->Lookup(varNum, &info)) |
6290 | { |
6291 | info->tail->next = node; |
6292 | info->tail = node; |
6293 | } |
6294 | // Create a new list. |
6295 | else |
6296 | { |
6297 | info = VarScopeMapInfo::Create(node, getAllocator()); |
6298 | compVarScopeMap->Set(varNum, info); |
6299 | } |
6300 | } |
6301 | } |
6302 | |
6303 | int __cdecl genCmpLocalVarLifeBeg(const void* elem1, const void* elem2) |
6304 | { |
6305 | return (*((VarScopeDsc**)elem1))->vsdLifeBeg - (*((VarScopeDsc**)elem2))->vsdLifeBeg; |
6306 | } |
6307 | |
6308 | int __cdecl genCmpLocalVarLifeEnd(const void* elem1, const void* elem2) |
6309 | { |
6310 | return (*((VarScopeDsc**)elem1))->vsdLifeEnd - (*((VarScopeDsc**)elem2))->vsdLifeEnd; |
6311 | } |
6312 | |
6313 | inline void Compiler::compInitScopeLists() |
6314 | { |
6315 | if (info.compVarScopesCount == 0) |
6316 | { |
6317 | compEnterScopeList = compExitScopeList = nullptr; |
6318 | return; |
6319 | } |
6320 | |
6321 | // Populate the 'compEnterScopeList' and 'compExitScopeList' lists |
6322 | |
6323 | compEnterScopeList = new (this, CMK_DebugInfo) VarScopeDsc*[info.compVarScopesCount]; |
6324 | compExitScopeList = new (this, CMK_DebugInfo) VarScopeDsc*[info.compVarScopesCount]; |
6325 | |
6326 | for (unsigned i = 0; i < info.compVarScopesCount; i++) |
6327 | { |
6328 | compEnterScopeList[i] = compExitScopeList[i] = &info.compVarScopes[i]; |
6329 | } |
6330 | |
6331 | qsort(compEnterScopeList, info.compVarScopesCount, sizeof(*compEnterScopeList), genCmpLocalVarLifeBeg); |
6332 | qsort(compExitScopeList, info.compVarScopesCount, sizeof(*compExitScopeList), genCmpLocalVarLifeEnd); |
6333 | } |
6334 | |
6335 | void Compiler::compResetScopeLists() |
6336 | { |
6337 | if (info.compVarScopesCount == 0) |
6338 | { |
6339 | return; |
6340 | } |
6341 | |
6342 | assert(compEnterScopeList && compExitScopeList); |
6343 | |
6344 | compNextEnterScope = compNextExitScope = 0; |
6345 | } |
6346 | |
6347 | VarScopeDsc* Compiler::compGetNextEnterScope(unsigned offs, bool scan) |
6348 | { |
6349 | assert(info.compVarScopesCount); |
6350 | assert(compEnterScopeList && compExitScopeList); |
6351 | |
6352 | if (compNextEnterScope < info.compVarScopesCount) |
6353 | { |
6354 | assert(compEnterScopeList[compNextEnterScope]); |
6355 | unsigned nextEnterOff = compEnterScopeList[compNextEnterScope]->vsdLifeBeg; |
6356 | assert(scan || (offs <= nextEnterOff)); |
6357 | |
6358 | if (!scan) |
6359 | { |
6360 | if (offs == nextEnterOff) |
6361 | { |
6362 | return compEnterScopeList[compNextEnterScope++]; |
6363 | } |
6364 | } |
6365 | else |
6366 | { |
6367 | if (nextEnterOff <= offs) |
6368 | { |
6369 | return compEnterScopeList[compNextEnterScope++]; |
6370 | } |
6371 | } |
6372 | } |
6373 | |
6374 | return nullptr; |
6375 | } |
6376 | |
6377 | VarScopeDsc* Compiler::compGetNextExitScope(unsigned offs, bool scan) |
6378 | { |
6379 | assert(info.compVarScopesCount); |
6380 | assert(compEnterScopeList && compExitScopeList); |
6381 | |
6382 | if (compNextExitScope < info.compVarScopesCount) |
6383 | { |
6384 | assert(compExitScopeList[compNextExitScope]); |
6385 | unsigned nextExitOffs = compExitScopeList[compNextExitScope]->vsdLifeEnd; |
6386 | assert(scan || (offs <= nextExitOffs)); |
6387 | |
6388 | if (!scan) |
6389 | { |
6390 | if (offs == nextExitOffs) |
6391 | { |
6392 | return compExitScopeList[compNextExitScope++]; |
6393 | } |
6394 | } |
6395 | else |
6396 | { |
6397 | if (nextExitOffs <= offs) |
6398 | { |
6399 | return compExitScopeList[compNextExitScope++]; |
6400 | } |
6401 | } |
6402 | } |
6403 | |
6404 | return nullptr; |
6405 | } |
6406 | |
6407 | // The function will call the callback functions for scopes with boundaries |
6408 | // at instrs from the current status of the scope lists to 'offset', |
6409 | // ordered by instrs. |
6410 | |
6411 | void Compiler::compProcessScopesUntil(unsigned offset, |
6412 | VARSET_TP* inScope, |
6413 | void (Compiler::*enterScopeFn)(VARSET_TP* inScope, VarScopeDsc*), |
6414 | void (Compiler::*exitScopeFn)(VARSET_TP* inScope, VarScopeDsc*)) |
6415 | { |
6416 | assert(offset != BAD_IL_OFFSET); |
6417 | assert(inScope != nullptr); |
6418 | |
6419 | bool foundExit = false, foundEnter = true; |
6420 | VarScopeDsc* scope; |
6421 | VarScopeDsc* nextExitScope = nullptr; |
6422 | VarScopeDsc* nextEnterScope = nullptr; |
6423 | unsigned offs = offset, curEnterOffs = 0; |
6424 | |
6425 | goto START_FINDING_SCOPES; |
6426 | |
6427 | // We need to determine the scopes which are open for the current block. |
6428 | // This loop walks over the missing blocks between the current and the |
6429 | // previous block, keeping the enter and exit offsets in lockstep. |
6430 | |
6431 | do |
6432 | { |
6433 | foundExit = foundEnter = false; |
6434 | |
6435 | if (nextExitScope) |
6436 | { |
6437 | (this->*exitScopeFn)(inScope, nextExitScope); |
6438 | nextExitScope = nullptr; |
6439 | foundExit = true; |
6440 | } |
6441 | |
6442 | offs = nextEnterScope ? nextEnterScope->vsdLifeBeg : offset; |
6443 | |
6444 | while ((scope = compGetNextExitScope(offs, true)) != nullptr) |
6445 | { |
6446 | foundExit = true; |
6447 | |
6448 | if (!nextEnterScope || scope->vsdLifeEnd > nextEnterScope->vsdLifeBeg) |
6449 | { |
6450 | // We overshot the last found Enter scope. Save the scope for later |
6451 | // and find an entering scope |
6452 | |
6453 | nextExitScope = scope; |
6454 | break; |
6455 | } |
6456 | |
6457 | (this->*exitScopeFn)(inScope, scope); |
6458 | } |
6459 | |
6460 | if (nextEnterScope) |
6461 | { |
6462 | (this->*enterScopeFn)(inScope, nextEnterScope); |
6463 | curEnterOffs = nextEnterScope->vsdLifeBeg; |
6464 | nextEnterScope = nullptr; |
6465 | foundEnter = true; |
6466 | } |
6467 | |
6468 | offs = nextExitScope ? nextExitScope->vsdLifeEnd : offset; |
6469 | |
6470 | START_FINDING_SCOPES: |
6471 | |
6472 | while ((scope = compGetNextEnterScope(offs, true)) != nullptr) |
6473 | { |
6474 | foundEnter = true; |
6475 | |
6476 | if ((nextExitScope && scope->vsdLifeBeg >= nextExitScope->vsdLifeEnd) || (scope->vsdLifeBeg > curEnterOffs)) |
6477 | { |
6478 | // We overshot the last found exit scope. Save the scope for later |
6479 | // and find an exiting scope |
6480 | |
6481 | nextEnterScope = scope; |
6482 | break; |
6483 | } |
6484 | |
6485 | (this->*enterScopeFn)(inScope, scope); |
6486 | |
6487 | if (!nextExitScope) |
6488 | { |
6489 | curEnterOffs = scope->vsdLifeBeg; |
6490 | } |
6491 | } |
6492 | } while (foundExit || foundEnter); |
6493 | } |
6494 | |
6495 | #if defined(DEBUG) |
6496 | |
6497 | void Compiler::compDispScopeLists() |
6498 | { |
6499 | unsigned i; |
6500 | |
6501 | printf("Local variable scopes = %d\n" , info.compVarScopesCount); |
6502 | |
6503 | if (info.compVarScopesCount) |
6504 | { |
6505 | printf(" \tVarNum \tLVNum \t Name \tBeg \tEnd\n" ); |
6506 | } |
6507 | |
6508 | printf("Sorted by enter scope:\n" ); |
6509 | for (i = 0; i < info.compVarScopesCount; i++) |
6510 | { |
6511 | VarScopeDsc* varScope = compEnterScopeList[i]; |
6512 | assert(varScope); |
6513 | printf("%2d: \t%02Xh \t%02Xh \t%10s \t%03Xh \t%03Xh" , i, varScope->vsdVarNum, varScope->vsdLVnum, |
6514 | VarNameToStr(varScope->vsdName) == nullptr ? "UNKNOWN" : VarNameToStr(varScope->vsdName), |
6515 | varScope->vsdLifeBeg, varScope->vsdLifeEnd); |
6516 | |
6517 | if (compNextEnterScope == i) |
6518 | { |
6519 | printf(" <-- next enter scope" ); |
6520 | } |
6521 | |
6522 | printf("\n" ); |
6523 | } |
6524 | |
6525 | printf("Sorted by exit scope:\n" ); |
6526 | for (i = 0; i < info.compVarScopesCount; i++) |
6527 | { |
6528 | VarScopeDsc* varScope = compExitScopeList[i]; |
6529 | assert(varScope); |
6530 | printf("%2d: \t%02Xh \t%02Xh \t%10s \t%03Xh \t%03Xh" , i, varScope->vsdVarNum, varScope->vsdLVnum, |
6531 | VarNameToStr(varScope->vsdName) == nullptr ? "UNKNOWN" : VarNameToStr(varScope->vsdName), |
6532 | varScope->vsdLifeBeg, varScope->vsdLifeEnd); |
6533 | |
6534 | if (compNextExitScope == i) |
6535 | { |
6536 | printf(" <-- next exit scope" ); |
6537 | } |
6538 | |
6539 | printf("\n" ); |
6540 | } |
6541 | } |
6542 | |
6543 | void Compiler::compDispLocalVars() |
6544 | { |
6545 | printf("info.compVarScopesCount = %d\n" , info.compVarScopesCount); |
6546 | |
6547 | if (info.compVarScopesCount > 0) |
6548 | { |
6549 | printf(" \tVarNum \tLVNum \t Name \tBeg \tEnd\n" ); |
6550 | } |
6551 | |
6552 | for (unsigned i = 0; i < info.compVarScopesCount; i++) |
6553 | { |
6554 | VarScopeDsc* varScope = &info.compVarScopes[i]; |
6555 | printf("%2d: \t%02Xh \t%02Xh \t%10s \t%03Xh \t%03Xh\n" , i, varScope->vsdVarNum, varScope->vsdLVnum, |
6556 | VarNameToStr(varScope->vsdName) == nullptr ? "UNKNOWN" : VarNameToStr(varScope->vsdName), |
6557 | varScope->vsdLifeBeg, varScope->vsdLifeEnd); |
6558 | } |
6559 | } |
6560 | |
6561 | #endif // DEBUG |
6562 | |
6563 | /*****************************************************************************/ |
6564 | |
6565 | #if MEASURE_CLRAPI_CALLS |
6566 | |
6567 | struct WrapICorJitInfo : public ICorJitInfo |
6568 | { |
6569 | //------------------------------------------------------------------------ |
6570 | // WrapICorJitInfo::makeOne: allocate an instance of WrapICorJitInfo |
6571 | // |
6572 | // Arguments: |
6573 | // alloc - the allocator to get memory from for the instance |
6574 | // compile - the compiler instance |
6575 | // compHndRef - the ICorJitInfo handle from the EE; the caller's |
6576 | // copy may be replaced with a "wrapper" instance |
6577 | // |
6578 | // Return Value: |
6579 | // If the config flags indicate that ICorJitInfo should be wrapped, |
6580 | // we return the "wrapper" instance; otherwise we return "nullptr". |
6581 | |
6582 | static WrapICorJitInfo* makeOne(ArenaAllocator* alloc, Compiler* compiler, COMP_HANDLE& compHndRef /* INOUT */) |
6583 | { |
6584 | WrapICorJitInfo* wrap = nullptr; |
6585 | |
6586 | if (JitConfig.JitEECallTimingInfo() != 0) |
6587 | { |
6588 | // It's too early to use the default allocator, so we do this |
6589 | // in two steps to be safe (the constructor doesn't need to do |
6590 | // anything except fill in the vtable pointer, so we let the |
6591 | // compiler do it). |
6592 | void* inst = alloc->allocateMemory(roundUp(sizeof(WrapICorJitInfo))); |
6593 | if (inst != nullptr) |
6594 | { |
6595 | // If you get a build error here due to 'WrapICorJitInfo' being |
6596 | // an abstract class, it's very likely that the wrapper bodies |
6597 | // in ICorJitInfo_API_wrapper.hpp are no longer in sync with |
6598 | // the EE interface; please be kind and update the header file. |
6599 | wrap = new (inst, jitstd::placement_t()) WrapICorJitInfo(); |
6600 | |
6601 | wrap->wrapComp = compiler; |
6602 | |
6603 | // Save the real handle and replace it with our wrapped version. |
6604 | wrap->wrapHnd = compHndRef; |
6605 | compHndRef = wrap; |
6606 | } |
6607 | } |
6608 | |
6609 | return wrap; |
6610 | } |
6611 | |
6612 | private: |
6613 | Compiler* wrapComp; |
6614 | COMP_HANDLE wrapHnd; // the "real thing" |
6615 | |
6616 | public: |
6617 | #include "ICorJitInfo_API_wrapper.hpp" |
6618 | }; |
6619 | |
6620 | #endif // MEASURE_CLRAPI_CALLS |
6621 | |
6622 | /*****************************************************************************/ |
6623 | |
6624 | // Compile a single method |
6625 | |
6626 | int jitNativeCode(CORINFO_METHOD_HANDLE methodHnd, |
6627 | CORINFO_MODULE_HANDLE classPtr, |
6628 | COMP_HANDLE compHnd, |
6629 | CORINFO_METHOD_INFO* methodInfo, |
6630 | void** methodCodePtr, |
6631 | ULONG* methodCodeSize, |
6632 | JitFlags* compileFlags, |
6633 | void* inlineInfoPtr) |
6634 | { |
6635 | // |
6636 | // A non-NULL inlineInfo means we are compiling the inlinee method. |
6637 | // |
6638 | InlineInfo* inlineInfo = (InlineInfo*)inlineInfoPtr; |
6639 | |
6640 | bool jitFallbackCompile = false; |
6641 | START: |
6642 | int result = CORJIT_INTERNALERROR; |
6643 | |
6644 | ArenaAllocator* pAlloc = nullptr; |
6645 | ArenaAllocator alloc; |
6646 | |
6647 | #if MEASURE_CLRAPI_CALLS |
6648 | WrapICorJitInfo* wrapCLR = nullptr; |
6649 | #endif |
6650 | |
6651 | if (inlineInfo) |
6652 | { |
6653 | // Use inliner's memory allocator when compiling the inlinee. |
6654 | pAlloc = inlineInfo->InlinerCompiler->compGetArenaAllocator(); |
6655 | } |
6656 | else |
6657 | { |
6658 | pAlloc = &alloc; |
6659 | } |
6660 | |
6661 | Compiler* pComp; |
6662 | pComp = nullptr; |
6663 | |
6664 | struct Param |
6665 | { |
6666 | Compiler* pComp; |
6667 | ArenaAllocator* pAlloc; |
6668 | bool jitFallbackCompile; |
6669 | |
6670 | CORINFO_METHOD_HANDLE methodHnd; |
6671 | CORINFO_MODULE_HANDLE classPtr; |
6672 | COMP_HANDLE compHnd; |
6673 | CORINFO_METHOD_INFO* methodInfo; |
6674 | void** methodCodePtr; |
6675 | ULONG* methodCodeSize; |
6676 | JitFlags* compileFlags; |
6677 | InlineInfo* inlineInfo; |
6678 | #if MEASURE_CLRAPI_CALLS |
6679 | WrapICorJitInfo* wrapCLR; |
6680 | #endif |
6681 | |
6682 | int result; |
6683 | } param; |
6684 | param.pComp = nullptr; |
6685 | param.pAlloc = pAlloc; |
6686 | param.jitFallbackCompile = jitFallbackCompile; |
6687 | param.methodHnd = methodHnd; |
6688 | param.classPtr = classPtr; |
6689 | param.compHnd = compHnd; |
6690 | param.methodInfo = methodInfo; |
6691 | param.methodCodePtr = methodCodePtr; |
6692 | param.methodCodeSize = methodCodeSize; |
6693 | param.compileFlags = compileFlags; |
6694 | param.inlineInfo = inlineInfo; |
6695 | #if MEASURE_CLRAPI_CALLS |
6696 | param.wrapCLR = nullptr; |
6697 | #endif |
6698 | param.result = result; |
6699 | |
6700 | setErrorTrap(compHnd, Param*, pParamOuter, ¶m) |
6701 | { |
6702 | setErrorTrap(nullptr, Param*, pParam, pParamOuter) |
6703 | { |
6704 | if (pParam->inlineInfo) |
6705 | { |
6706 | // Lazily create the inlinee compiler object |
6707 | if (pParam->inlineInfo->InlinerCompiler->InlineeCompiler == nullptr) |
6708 | { |
6709 | pParam->inlineInfo->InlinerCompiler->InlineeCompiler = |
6710 | (Compiler*)pParam->pAlloc->allocateMemory(roundUp(sizeof(*pParam->pComp))); |
6711 | } |
6712 | |
6713 | // Use the inlinee compiler object |
6714 | pParam->pComp = pParam->inlineInfo->InlinerCompiler->InlineeCompiler; |
6715 | #ifdef DEBUG |
6716 | // memset(pParam->pComp, 0xEE, sizeof(Compiler)); |
6717 | #endif |
6718 | } |
6719 | else |
6720 | { |
6721 | // Allocate create the inliner compiler object |
6722 | pParam->pComp = (Compiler*)pParam->pAlloc->allocateMemory(roundUp(sizeof(*pParam->pComp))); |
6723 | } |
6724 | |
6725 | #if MEASURE_CLRAPI_CALLS |
6726 | pParam->wrapCLR = WrapICorJitInfo::makeOne(pParam->pAlloc, pParam->pComp, pParam->compHnd); |
6727 | #endif |
6728 | |
6729 | // push this compiler on the stack (TLS) |
6730 | pParam->pComp->prevCompiler = JitTls::GetCompiler(); |
6731 | JitTls::SetCompiler(pParam->pComp); |
6732 | |
6733 | // PREFIX_ASSUME gets turned into ASSERT_CHECK and we cannot have it here |
6734 | #if defined(_PREFAST_) || defined(_PREFIX_) |
6735 | PREFIX_ASSUME(pParam->pComp != NULL); |
6736 | #else |
6737 | assert(pParam->pComp != nullptr); |
6738 | #endif |
6739 | |
6740 | pParam->pComp->compInit(pParam->pAlloc, pParam->inlineInfo); |
6741 | |
6742 | #ifdef DEBUG |
6743 | pParam->pComp->jitFallbackCompile = pParam->jitFallbackCompile; |
6744 | #endif |
6745 | |
6746 | // Now generate the code |
6747 | pParam->result = |
6748 | pParam->pComp->compCompile(pParam->methodHnd, pParam->classPtr, pParam->compHnd, pParam->methodInfo, |
6749 | pParam->methodCodePtr, pParam->methodCodeSize, pParam->compileFlags); |
6750 | } |
6751 | finallyErrorTrap() |
6752 | { |
6753 | Compiler* pCompiler = pParamOuter->pComp; |
6754 | |
6755 | // If OOM is thrown when allocating memory for a pComp, we will end up here. |
6756 | // For this case, pComp and also pCompiler will be a nullptr |
6757 | // |
6758 | if (pCompiler != nullptr) |
6759 | { |
6760 | pCompiler->info.compCode = nullptr; |
6761 | |
6762 | // pop the compiler off the TLS stack only if it was linked above |
6763 | assert(JitTls::GetCompiler() == pCompiler); |
6764 | JitTls::SetCompiler(pCompiler->prevCompiler); |
6765 | } |
6766 | |
6767 | if (pParamOuter->inlineInfo == nullptr) |
6768 | { |
6769 | // Free up the allocator we were using |
6770 | pParamOuter->pAlloc->destroy(); |
6771 | } |
6772 | } |
6773 | endErrorTrap() |
6774 | } |
6775 | impJitErrorTrap() |
6776 | { |
6777 | // If we were looking at an inlinee.... |
6778 | if (inlineInfo != nullptr) |
6779 | { |
6780 | // Note that we failed to compile the inlinee, and that |
6781 | // there's no point trying to inline it again anywhere else. |
6782 | inlineInfo->inlineResult->NoteFatal(InlineObservation::CALLEE_COMPILATION_ERROR); |
6783 | } |
6784 | param.result = __errc; |
6785 | } |
6786 | endErrorTrap() |
6787 | |
6788 | result = param.result; |
6789 | |
6790 | if (!inlineInfo && (result == CORJIT_INTERNALERROR || result == CORJIT_RECOVERABLEERROR) && !jitFallbackCompile) |
6791 | { |
6792 | // If we failed the JIT, reattempt with debuggable code. |
6793 | jitFallbackCompile = true; |
6794 | |
6795 | // Update the flags for 'safer' code generation. |
6796 | compileFlags->Set(JitFlags::JIT_FLAG_MIN_OPT); |
6797 | compileFlags->Clear(JitFlags::JIT_FLAG_SIZE_OPT); |
6798 | compileFlags->Clear(JitFlags::JIT_FLAG_SPEED_OPT); |
6799 | |
6800 | goto START; |
6801 | } |
6802 | |
6803 | return result; |
6804 | } |
6805 | |
6806 | #if defined(UNIX_AMD64_ABI) |
6807 | |
6808 | // GetTypeFromClassificationAndSizes: |
6809 | // Returns the type of the eightbyte accounting for the classification and size of the eightbyte. |
6810 | // |
6811 | // args: |
6812 | // classType: classification type |
6813 | // size: size of the eightbyte. |
6814 | // |
6815 | // static |
6816 | var_types Compiler::GetTypeFromClassificationAndSizes(SystemVClassificationType classType, int size) |
6817 | { |
6818 | var_types type = TYP_UNKNOWN; |
6819 | switch (classType) |
6820 | { |
6821 | case SystemVClassificationTypeInteger: |
6822 | if (size == 1) |
6823 | { |
6824 | type = TYP_BYTE; |
6825 | } |
6826 | else if (size <= 2) |
6827 | { |
6828 | type = TYP_SHORT; |
6829 | } |
6830 | else if (size <= 4) |
6831 | { |
6832 | type = TYP_INT; |
6833 | } |
6834 | else if (size <= 8) |
6835 | { |
6836 | type = TYP_LONG; |
6837 | } |
6838 | else |
6839 | { |
6840 | assert(false && "GetTypeFromClassificationAndSizes Invalid Integer classification type." ); |
6841 | } |
6842 | break; |
6843 | case SystemVClassificationTypeIntegerReference: |
6844 | type = TYP_REF; |
6845 | break; |
6846 | case SystemVClassificationTypeIntegerByRef: |
6847 | type = TYP_BYREF; |
6848 | break; |
6849 | case SystemVClassificationTypeSSE: |
6850 | if (size <= 4) |
6851 | { |
6852 | type = TYP_FLOAT; |
6853 | } |
6854 | else if (size <= 8) |
6855 | { |
6856 | type = TYP_DOUBLE; |
6857 | } |
6858 | else |
6859 | { |
6860 | assert(false && "GetTypeFromClassificationAndSizes Invalid SSE classification type." ); |
6861 | } |
6862 | break; |
6863 | |
6864 | default: |
6865 | assert(false && "GetTypeFromClassificationAndSizes Invalid classification type." ); |
6866 | break; |
6867 | } |
6868 | |
6869 | return type; |
6870 | } |
6871 | |
6872 | //------------------------------------------------------------------- |
6873 | // GetEightByteType: Returns the type of eightbyte slot of a struct |
6874 | // |
6875 | // Arguments: |
6876 | // structDesc - struct classification description. |
6877 | // slotNum - eightbyte slot number for the struct. |
6878 | // |
6879 | // Return Value: |
6880 | // type of the eightbyte slot of the struct |
6881 | // |
6882 | // static |
6883 | var_types Compiler::GetEightByteType(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc, |
6884 | unsigned slotNum) |
6885 | { |
6886 | var_types eightByteType = TYP_UNDEF; |
6887 | unsigned len = structDesc.eightByteSizes[slotNum]; |
6888 | |
6889 | switch (structDesc.eightByteClassifications[slotNum]) |
6890 | { |
6891 | case SystemVClassificationTypeInteger: |
6892 | // See typelist.h for jit type definition. |
6893 | // All the types of size < 4 bytes are of jit type TYP_INT. |
6894 | if (structDesc.eightByteSizes[slotNum] <= 4) |
6895 | { |
6896 | eightByteType = TYP_INT; |
6897 | } |
6898 | else if (structDesc.eightByteSizes[slotNum] <= 8) |
6899 | { |
6900 | eightByteType = TYP_LONG; |
6901 | } |
6902 | else |
6903 | { |
6904 | assert(false && "GetEightByteType Invalid Integer classification type." ); |
6905 | } |
6906 | break; |
6907 | case SystemVClassificationTypeIntegerReference: |
6908 | assert(len == REGSIZE_BYTES); |
6909 | eightByteType = TYP_REF; |
6910 | break; |
6911 | case SystemVClassificationTypeIntegerByRef: |
6912 | assert(len == REGSIZE_BYTES); |
6913 | eightByteType = TYP_BYREF; |
6914 | break; |
6915 | case SystemVClassificationTypeSSE: |
6916 | if (structDesc.eightByteSizes[slotNum] <= 4) |
6917 | { |
6918 | eightByteType = TYP_FLOAT; |
6919 | } |
6920 | else if (structDesc.eightByteSizes[slotNum] <= 8) |
6921 | { |
6922 | eightByteType = TYP_DOUBLE; |
6923 | } |
6924 | else |
6925 | { |
6926 | assert(false && "GetEightByteType Invalid SSE classification type." ); |
6927 | } |
6928 | break; |
6929 | default: |
6930 | assert(false && "GetEightByteType Invalid classification type." ); |
6931 | break; |
6932 | } |
6933 | |
6934 | return eightByteType; |
6935 | } |
6936 | |
6937 | //------------------------------------------------------------------------------------------------------ |
6938 | // GetStructTypeOffset: Gets the type, size and offset of the eightbytes of a struct for System V systems. |
6939 | // |
6940 | // Arguments: |
6941 | // 'structDesc' - struct description |
6942 | // 'type0' - out param; returns the type of the first eightbyte. |
6943 | // 'type1' - out param; returns the type of the second eightbyte. |
6944 | // 'offset0' - out param; returns the offset of the first eightbyte. |
6945 | // 'offset1' - out param; returns the offset of the second eightbyte. |
6946 | // |
6947 | // static |
6948 | void Compiler::GetStructTypeOffset(const SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR& structDesc, |
6949 | var_types* type0, |
6950 | var_types* type1, |
6951 | unsigned __int8* offset0, |
6952 | unsigned __int8* offset1) |
6953 | { |
6954 | *offset0 = structDesc.eightByteOffsets[0]; |
6955 | *offset1 = structDesc.eightByteOffsets[1]; |
6956 | |
6957 | *type0 = TYP_UNKNOWN; |
6958 | *type1 = TYP_UNKNOWN; |
6959 | |
6960 | // Set the first eightbyte data |
6961 | if (structDesc.eightByteCount >= 1) |
6962 | { |
6963 | *type0 = GetEightByteType(structDesc, 0); |
6964 | } |
6965 | |
6966 | // Set the second eight byte data |
6967 | if (structDesc.eightByteCount == 2) |
6968 | { |
6969 | *type1 = GetEightByteType(structDesc, 1); |
6970 | } |
6971 | } |
6972 | |
6973 | //------------------------------------------------------------------------------------------------------ |
6974 | // GetStructTypeOffset: Gets the type, size and offset of the eightbytes of a struct for System V systems. |
6975 | // |
6976 | // Arguments: |
6977 | // 'typeHnd' - type handle |
6978 | // 'type0' - out param; returns the type of the first eightbyte. |
6979 | // 'type1' - out param; returns the type of the second eightbyte. |
6980 | // 'offset0' - out param; returns the offset of the first eightbyte. |
6981 | // 'offset1' - out param; returns the offset of the second eightbyte. |
6982 | // |
6983 | void Compiler::GetStructTypeOffset(CORINFO_CLASS_HANDLE typeHnd, |
6984 | var_types* type0, |
6985 | var_types* type1, |
6986 | unsigned __int8* offset0, |
6987 | unsigned __int8* offset1) |
6988 | { |
6989 | SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc; |
6990 | eeGetSystemVAmd64PassStructInRegisterDescriptor(typeHnd, &structDesc); |
6991 | assert(structDesc.passedInRegisters); |
6992 | GetStructTypeOffset(structDesc, type0, type1, offset0, offset1); |
6993 | } |
6994 | |
6995 | #endif // defined(UNIX_AMD64_ABI) |
6996 | |
6997 | /*****************************************************************************/ |
6998 | /*****************************************************************************/ |
6999 | |
7000 | #ifdef DEBUG |
7001 | Compiler::NodeToIntMap* Compiler::FindReachableNodesInNodeTestData() |
7002 | { |
7003 | NodeToIntMap* reachable = new (getAllocatorDebugOnly()) NodeToIntMap(getAllocatorDebugOnly()); |
7004 | |
7005 | if (m_nodeTestData == nullptr) |
7006 | { |
7007 | return reachable; |
7008 | } |
7009 | |
7010 | // Otherwise, iterate. |
7011 | |
7012 | for (BasicBlock* block = fgFirstBB; block != nullptr; block = block->bbNext) |
7013 | { |
7014 | for (GenTree* stmt = block->FirstNonPhiDef(); stmt != nullptr; stmt = stmt->gtNext) |
7015 | { |
7016 | for (GenTree* tree = stmt->gtStmt.gtStmtList; tree; tree = tree->gtNext) |
7017 | { |
7018 | TestLabelAndNum tlAndN; |
7019 | |
7020 | // For call nodes, translate late args to what they stand for. |
7021 | if (tree->OperGet() == GT_CALL) |
7022 | { |
7023 | GenTreeCall* call = tree->AsCall(); |
7024 | GenTreeArgList* args = call->gtCallArgs; |
7025 | unsigned i = 0; |
7026 | while (args != nullptr) |
7027 | { |
7028 | GenTree* arg = args->Current(); |
7029 | if (arg->gtFlags & GTF_LATE_ARG) |
7030 | { |
7031 | // Find the corresponding late arg. |
7032 | GenTree* lateArg = call->fgArgInfo->GetArgNode(i); |
7033 | if (GetNodeTestData()->Lookup(lateArg, &tlAndN)) |
7034 | { |
7035 | reachable->Set(lateArg, 0); |
7036 | } |
7037 | } |
7038 | i++; |
7039 | args = args->Rest(); |
7040 | } |
7041 | } |
7042 | |
7043 | if (GetNodeTestData()->Lookup(tree, &tlAndN)) |
7044 | { |
7045 | reachable->Set(tree, 0); |
7046 | } |
7047 | } |
7048 | } |
7049 | } |
7050 | return reachable; |
7051 | } |
7052 | |
7053 | void Compiler::TransferTestDataToNode(GenTree* from, GenTree* to) |
7054 | { |
7055 | TestLabelAndNum tlAndN; |
7056 | // We can't currently associate multiple annotations with a single node. |
7057 | // If we need to, we can fix this... |
7058 | |
7059 | // If the table is null, don't create it just to do the lookup, which would fail... |
7060 | if (m_nodeTestData != nullptr && GetNodeTestData()->Lookup(from, &tlAndN)) |
7061 | { |
7062 | assert(!GetNodeTestData()->Lookup(to, &tlAndN)); |
7063 | // We can't currently associate multiple annotations with a single node. |
7064 | // If we need to, we can fix this... |
7065 | TestLabelAndNum tlAndNTo; |
7066 | assert(!GetNodeTestData()->Lookup(to, &tlAndNTo)); |
7067 | |
7068 | GetNodeTestData()->Remove(from); |
7069 | GetNodeTestData()->Set(to, tlAndN); |
7070 | } |
7071 | } |
7072 | |
7073 | void Compiler::CopyTestDataToCloneTree(GenTree* from, GenTree* to) |
7074 | { |
7075 | if (m_nodeTestData == nullptr) |
7076 | { |
7077 | return; |
7078 | } |
7079 | if (from == nullptr) |
7080 | { |
7081 | assert(to == nullptr); |
7082 | return; |
7083 | } |
7084 | // Otherwise... |
7085 | TestLabelAndNum tlAndN; |
7086 | if (GetNodeTestData()->Lookup(from, &tlAndN)) |
7087 | { |
7088 | // We can't currently associate multiple annotations with a single node. |
7089 | // If we need to, we can fix this... |
7090 | TestLabelAndNum tlAndNTo; |
7091 | assert(!GetNodeTestData()->Lookup(to, &tlAndNTo)); |
7092 | GetNodeTestData()->Set(to, tlAndN); |
7093 | } |
7094 | // Now recurse, in parallel on both trees. |
7095 | |
7096 | genTreeOps oper = from->OperGet(); |
7097 | unsigned kind = from->OperKind(); |
7098 | assert(oper == to->OperGet()); |
7099 | |
7100 | // Cconstant or leaf nodes have no children. |
7101 | if (kind & (GTK_CONST | GTK_LEAF)) |
7102 | { |
7103 | return; |
7104 | } |
7105 | |
7106 | // Otherwise, is it a 'simple' unary/binary operator? |
7107 | |
7108 | if (kind & GTK_SMPOP) |
7109 | { |
7110 | if (from->gtOp.gtOp1 != nullptr) |
7111 | { |
7112 | assert(to->gtOp.gtOp1 != nullptr); |
7113 | CopyTestDataToCloneTree(from->gtOp.gtOp1, to->gtOp.gtOp1); |
7114 | } |
7115 | else |
7116 | { |
7117 | assert(to->gtOp.gtOp1 == nullptr); |
7118 | } |
7119 | |
7120 | if (from->gtGetOp2IfPresent() != nullptr) |
7121 | { |
7122 | assert(to->gtGetOp2IfPresent() != nullptr); |
7123 | CopyTestDataToCloneTree(from->gtGetOp2(), to->gtGetOp2()); |
7124 | } |
7125 | else |
7126 | { |
7127 | assert(to->gtGetOp2IfPresent() == nullptr); |
7128 | } |
7129 | |
7130 | return; |
7131 | } |
7132 | |
7133 | // Otherwise, see what kind of a special operator we have here. |
7134 | |
7135 | switch (oper) |
7136 | { |
7137 | case GT_STMT: |
7138 | CopyTestDataToCloneTree(from->gtStmt.gtStmtExpr, to->gtStmt.gtStmtExpr); |
7139 | return; |
7140 | |
7141 | case GT_CALL: |
7142 | CopyTestDataToCloneTree(from->gtCall.gtCallObjp, to->gtCall.gtCallObjp); |
7143 | CopyTestDataToCloneTree(from->gtCall.gtCallArgs, to->gtCall.gtCallArgs); |
7144 | CopyTestDataToCloneTree(from->gtCall.gtCallLateArgs, to->gtCall.gtCallLateArgs); |
7145 | |
7146 | if (from->gtCall.gtCallType == CT_INDIRECT) |
7147 | { |
7148 | CopyTestDataToCloneTree(from->gtCall.gtCallCookie, to->gtCall.gtCallCookie); |
7149 | CopyTestDataToCloneTree(from->gtCall.gtCallAddr, to->gtCall.gtCallAddr); |
7150 | } |
7151 | // The other call types do not have additional GenTree arguments. |
7152 | |
7153 | return; |
7154 | |
7155 | case GT_FIELD: |
7156 | CopyTestDataToCloneTree(from->gtField.gtFldObj, to->gtField.gtFldObj); |
7157 | return; |
7158 | |
7159 | case GT_ARR_ELEM: |
7160 | assert(from->gtArrElem.gtArrRank == to->gtArrElem.gtArrRank); |
7161 | for (unsigned dim = 0; dim < from->gtArrElem.gtArrRank; dim++) |
7162 | { |
7163 | CopyTestDataToCloneTree(from->gtArrElem.gtArrInds[dim], to->gtArrElem.gtArrInds[dim]); |
7164 | } |
7165 | CopyTestDataToCloneTree(from->gtArrElem.gtArrObj, to->gtArrElem.gtArrObj); |
7166 | return; |
7167 | |
7168 | case GT_CMPXCHG: |
7169 | CopyTestDataToCloneTree(from->gtCmpXchg.gtOpLocation, to->gtCmpXchg.gtOpLocation); |
7170 | CopyTestDataToCloneTree(from->gtCmpXchg.gtOpValue, to->gtCmpXchg.gtOpValue); |
7171 | CopyTestDataToCloneTree(from->gtCmpXchg.gtOpComparand, to->gtCmpXchg.gtOpComparand); |
7172 | return; |
7173 | |
7174 | case GT_ARR_BOUNDS_CHECK: |
7175 | #ifdef FEATURE_SIMD |
7176 | case GT_SIMD_CHK: |
7177 | #endif // FEATURE_SIMD |
7178 | #ifdef FEATURE_HW_INTRINSICS |
7179 | case GT_HW_INTRINSIC_CHK: |
7180 | #endif // FEATURE_HW_INTRINSICS |
7181 | CopyTestDataToCloneTree(from->gtBoundsChk.gtIndex, to->gtBoundsChk.gtIndex); |
7182 | CopyTestDataToCloneTree(from->gtBoundsChk.gtArrLen, to->gtBoundsChk.gtArrLen); |
7183 | return; |
7184 | |
7185 | default: |
7186 | unreached(); |
7187 | } |
7188 | } |
7189 | |
7190 | #endif // DEBUG |
7191 | |
7192 | /* |
7193 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
7194 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
7195 | XX XX |
7196 | XX jvc XX |
7197 | XX XX |
7198 | XX Functions for the stand-alone version of the JIT . XX |
7199 | XX XX |
7200 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
7201 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
7202 | */ |
7203 | |
7204 | /*****************************************************************************/ |
7205 | void codeGeneratorCodeSizeBeg() |
7206 | { |
7207 | } |
7208 | /*****************************************************************************/ |
7209 | |
7210 | /***************************************************************************** |
7211 | * |
7212 | * If any temporary tables are smaller than 'genMinSize2free' we won't bother |
7213 | * freeing them. |
7214 | */ |
7215 | |
7216 | const size_t genMinSize2free = 64; |
7217 | |
7218 | /*****************************************************************************/ |
7219 | |
7220 | /***************************************************************************** |
7221 | * |
7222 | * Used for counting pointer assignments. |
7223 | */ |
7224 | |
7225 | /*****************************************************************************/ |
7226 | void codeGeneratorCodeSizeEnd() |
7227 | { |
7228 | } |
7229 | /***************************************************************************** |
7230 | * |
7231 | * Gather statistics - mainly used for the standalone |
7232 | * Enable various #ifdef's to get the information you need |
7233 | */ |
7234 | |
7235 | void Compiler::compJitStats() |
7236 | { |
7237 | #if CALL_ARG_STATS |
7238 | |
7239 | /* Method types and argument statistics */ |
7240 | compCallArgStats(); |
7241 | #endif // CALL_ARG_STATS |
7242 | } |
7243 | |
7244 | #if CALL_ARG_STATS |
7245 | |
7246 | /***************************************************************************** |
7247 | * |
7248 | * Gather statistics about method calls and arguments |
7249 | */ |
7250 | |
7251 | void Compiler::compCallArgStats() |
7252 | { |
7253 | GenTree* args; |
7254 | GenTree* argx; |
7255 | |
7256 | BasicBlock* block; |
7257 | GenTree* stmt; |
7258 | GenTree* call; |
7259 | |
7260 | unsigned argNum; |
7261 | |
7262 | unsigned argDWordNum; |
7263 | unsigned argLngNum; |
7264 | unsigned argFltNum; |
7265 | unsigned argDblNum; |
7266 | |
7267 | unsigned regArgNum; |
7268 | unsigned regArgDeferred; |
7269 | unsigned regArgTemp; |
7270 | |
7271 | unsigned regArgLclVar; |
7272 | unsigned regArgConst; |
7273 | |
7274 | unsigned argTempsThisMethod = 0; |
7275 | |
7276 | assert(fgStmtListThreaded); |
7277 | |
7278 | for (block = fgFirstBB; block; block = block->bbNext) |
7279 | { |
7280 | for (stmt = block->bbTreeList; stmt; stmt = stmt->gtNext) |
7281 | { |
7282 | assert(stmt->gtOper == GT_STMT); |
7283 | |
7284 | for (call = stmt->gtStmt.gtStmtList; call; call = call->gtNext) |
7285 | { |
7286 | if (call->gtOper != GT_CALL) |
7287 | continue; |
7288 | |
7289 | argNum = |
7290 | |
7291 | regArgNum = regArgDeferred = regArgTemp = |
7292 | |
7293 | regArgConst = regArgLclVar = |
7294 | |
7295 | argDWordNum = argLngNum = argFltNum = argDblNum = 0; |
7296 | |
7297 | argTotalCalls++; |
7298 | |
7299 | if (!call->gtCall.gtCallObjp) |
7300 | { |
7301 | if (call->gtCall.gtCallType == CT_HELPER) |
7302 | { |
7303 | argHelperCalls++; |
7304 | } |
7305 | else |
7306 | { |
7307 | argStaticCalls++; |
7308 | } |
7309 | } |
7310 | else |
7311 | { |
7312 | /* We have a 'this' pointer */ |
7313 | |
7314 | argDWordNum++; |
7315 | argNum++; |
7316 | regArgNum++; |
7317 | regArgDeferred++; |
7318 | argTotalObjPtr++; |
7319 | |
7320 | if (call->IsVirtual()) |
7321 | { |
7322 | /* virtual function */ |
7323 | argVirtualCalls++; |
7324 | } |
7325 | else |
7326 | { |
7327 | argNonVirtualCalls++; |
7328 | } |
7329 | } |
7330 | } |
7331 | } |
7332 | } |
7333 | |
7334 | argTempsCntTable.record(argTempsThisMethod); |
7335 | |
7336 | if (argMaxTempsPerMethod < argTempsThisMethod) |
7337 | { |
7338 | argMaxTempsPerMethod = argTempsThisMethod; |
7339 | } |
7340 | } |
7341 | |
7342 | /* static */ |
7343 | void Compiler::compDispCallArgStats(FILE* fout) |
7344 | { |
7345 | if (argTotalCalls == 0) |
7346 | return; |
7347 | |
7348 | fprintf(fout, "\n" ); |
7349 | fprintf(fout, "--------------------------------------------------\n" ); |
7350 | fprintf(fout, "Call stats\n" ); |
7351 | fprintf(fout, "--------------------------------------------------\n" ); |
7352 | fprintf(fout, "Total # of calls = %d, calls / method = %.3f\n\n" , argTotalCalls, |
7353 | (float)argTotalCalls / genMethodCnt); |
7354 | |
7355 | fprintf(fout, "Percentage of helper calls = %4.2f %%\n" , (float)(100 * argHelperCalls) / argTotalCalls); |
7356 | fprintf(fout, "Percentage of static calls = %4.2f %%\n" , (float)(100 * argStaticCalls) / argTotalCalls); |
7357 | fprintf(fout, "Percentage of virtual calls = %4.2f %%\n" , (float)(100 * argVirtualCalls) / argTotalCalls); |
7358 | fprintf(fout, "Percentage of non-virtual calls = %4.2f %%\n\n" , (float)(100 * argNonVirtualCalls) / argTotalCalls); |
7359 | |
7360 | fprintf(fout, "Average # of arguments per call = %.2f%%\n\n" , (float)argTotalArgs / argTotalCalls); |
7361 | |
7362 | fprintf(fout, "Percentage of DWORD arguments = %.2f %%\n" , (float)(100 * argTotalDWordArgs) / argTotalArgs); |
7363 | fprintf(fout, "Percentage of LONG arguments = %.2f %%\n" , (float)(100 * argTotalLongArgs) / argTotalArgs); |
7364 | fprintf(fout, "Percentage of FLOAT arguments = %.2f %%\n" , (float)(100 * argTotalFloatArgs) / argTotalArgs); |
7365 | fprintf(fout, "Percentage of DOUBLE arguments = %.2f %%\n\n" , (float)(100 * argTotalDoubleArgs) / argTotalArgs); |
7366 | |
7367 | if (argTotalRegArgs == 0) |
7368 | return; |
7369 | |
7370 | /* |
7371 | fprintf(fout, "Total deferred arguments = %d \n", argTotalDeferred); |
7372 | |
7373 | fprintf(fout, "Total temp arguments = %d \n\n", argTotalTemps); |
7374 | |
7375 | fprintf(fout, "Total 'this' arguments = %d \n", argTotalObjPtr); |
7376 | fprintf(fout, "Total local var arguments = %d \n", argTotalLclVar); |
7377 | fprintf(fout, "Total constant arguments = %d \n\n", argTotalConst); |
7378 | */ |
7379 | |
7380 | fprintf(fout, "\nRegister Arguments:\n\n" ); |
7381 | |
7382 | fprintf(fout, "Percentage of deferred arguments = %.2f %%\n" , (float)(100 * argTotalDeferred) / argTotalRegArgs); |
7383 | fprintf(fout, "Percentage of temp arguments = %.2f %%\n\n" , (float)(100 * argTotalTemps) / argTotalRegArgs); |
7384 | |
7385 | fprintf(fout, "Maximum # of temps per method = %d\n\n" , argMaxTempsPerMethod); |
7386 | |
7387 | fprintf(fout, "Percentage of ObjPtr arguments = %.2f %%\n" , (float)(100 * argTotalObjPtr) / argTotalRegArgs); |
7388 | // fprintf(fout, "Percentage of global arguments = %.2f %%\n", (float)(100 * argTotalDWordGlobEf) / |
7389 | // argTotalRegArgs); |
7390 | fprintf(fout, "Percentage of constant arguments = %.2f %%\n" , (float)(100 * argTotalConst) / argTotalRegArgs); |
7391 | fprintf(fout, "Percentage of lcl var arguments = %.2f %%\n\n" , (float)(100 * argTotalLclVar) / argTotalRegArgs); |
7392 | |
7393 | fprintf(fout, "--------------------------------------------------\n" ); |
7394 | fprintf(fout, "Argument count frequency table (includes ObjPtr):\n" ); |
7395 | fprintf(fout, "--------------------------------------------------\n" ); |
7396 | argCntTable.dump(fout); |
7397 | fprintf(fout, "--------------------------------------------------\n" ); |
7398 | |
7399 | fprintf(fout, "--------------------------------------------------\n" ); |
7400 | fprintf(fout, "DWORD argument count frequency table (w/o LONG):\n" ); |
7401 | fprintf(fout, "--------------------------------------------------\n" ); |
7402 | argDWordCntTable.dump(fout); |
7403 | fprintf(fout, "--------------------------------------------------\n" ); |
7404 | |
7405 | fprintf(fout, "--------------------------------------------------\n" ); |
7406 | fprintf(fout, "Temps count frequency table (per method):\n" ); |
7407 | fprintf(fout, "--------------------------------------------------\n" ); |
7408 | argTempsCntTable.dump(fout); |
7409 | fprintf(fout, "--------------------------------------------------\n" ); |
7410 | |
7411 | /* |
7412 | fprintf(fout, "--------------------------------------------------\n"); |
7413 | fprintf(fout, "DWORD argument count frequency table (w/ LONG):\n"); |
7414 | fprintf(fout, "--------------------------------------------------\n"); |
7415 | argDWordLngCntTable.dump(fout); |
7416 | fprintf(fout, "--------------------------------------------------\n"); |
7417 | */ |
7418 | } |
7419 | |
7420 | #endif // CALL_ARG_STATS |
7421 | |
7422 | // JIT time end to end, and by phases. |
7423 | |
7424 | #ifdef FEATURE_JIT_METHOD_PERF |
7425 | // Static variables |
7426 | CritSecObject CompTimeSummaryInfo::s_compTimeSummaryLock; |
7427 | CompTimeSummaryInfo CompTimeSummaryInfo::s_compTimeSummary; |
7428 | #if MEASURE_CLRAPI_CALLS |
7429 | double JitTimer::s_cyclesPerSec = CycleTimer::CyclesPerSecond(); |
7430 | #endif |
7431 | #endif // FEATURE_JIT_METHOD_PERF |
7432 | |
7433 | #if defined(FEATURE_JIT_METHOD_PERF) || DUMP_FLOWGRAPHS || defined(FEATURE_TRACELOGGING) |
7434 | const char* PhaseNames[] = { |
7435 | #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent, measureIR) string_nm, |
7436 | #include "compphases.h" |
7437 | }; |
7438 | |
7439 | const char* PhaseEnums[] = { |
7440 | #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent, measureIR) #enum_nm, |
7441 | #include "compphases.h" |
7442 | }; |
7443 | |
7444 | const LPCWSTR PhaseShortNames[] = { |
7445 | #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent, measureIR) W(short_nm), |
7446 | #include "compphases.h" |
7447 | }; |
7448 | #endif // defined(FEATURE_JIT_METHOD_PERF) || DUMP_FLOWGRAPHS |
7449 | |
7450 | #ifdef FEATURE_JIT_METHOD_PERF |
7451 | bool PhaseHasChildren[] = { |
7452 | #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent, measureIR) hasChildren, |
7453 | #include "compphases.h" |
7454 | }; |
7455 | |
7456 | int PhaseParent[] = { |
7457 | #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent, measureIR) parent, |
7458 | #include "compphases.h" |
7459 | }; |
7460 | |
7461 | bool PhaseReportsIRSize[] = { |
7462 | #define CompPhaseNameMacro(enum_nm, string_nm, short_nm, hasChildren, parent, measureIR) measureIR, |
7463 | #include "compphases.h" |
7464 | }; |
7465 | |
7466 | CompTimeInfo::CompTimeInfo(unsigned byteCodeBytes) |
7467 | : m_byteCodeBytes(byteCodeBytes) |
7468 | , m_totalCycles(0) |
7469 | , m_parentPhaseEndSlop(0) |
7470 | , m_timerFailure(false) |
7471 | #if MEASURE_CLRAPI_CALLS |
7472 | , m_allClrAPIcalls(0) |
7473 | , m_allClrAPIcycles(0) |
7474 | #endif |
7475 | { |
7476 | for (int i = 0; i < PHASE_NUMBER_OF; i++) |
7477 | { |
7478 | m_invokesByPhase[i] = 0; |
7479 | m_cyclesByPhase[i] = 0; |
7480 | #if MEASURE_CLRAPI_CALLS |
7481 | m_CLRinvokesByPhase[i] = 0; |
7482 | m_CLRcyclesByPhase[i] = 0; |
7483 | #endif |
7484 | } |
7485 | |
7486 | #if MEASURE_CLRAPI_CALLS |
7487 | assert(ARRAYSIZE(m_perClrAPIcalls) == API_ICorJitInfo_Names::API_COUNT); |
7488 | assert(ARRAYSIZE(m_perClrAPIcycles) == API_ICorJitInfo_Names::API_COUNT); |
7489 | assert(ARRAYSIZE(m_maxClrAPIcycles) == API_ICorJitInfo_Names::API_COUNT); |
7490 | for (int i = 0; i < API_ICorJitInfo_Names::API_COUNT; i++) |
7491 | { |
7492 | m_perClrAPIcalls[i] = 0; |
7493 | m_perClrAPIcycles[i] = 0; |
7494 | m_maxClrAPIcycles[i] = 0; |
7495 | } |
7496 | #endif |
7497 | } |
7498 | |
7499 | bool CompTimeSummaryInfo::IncludedInFilteredData(CompTimeInfo& info) |
7500 | { |
7501 | return false; // info.m_byteCodeBytes < 10; |
7502 | } |
7503 | |
7504 | //------------------------------------------------------------------------ |
7505 | // CompTimeSummaryInfo::AddInfo: Record timing info from one compile. |
7506 | // |
7507 | // Arguments: |
7508 | // info - The timing information to record. |
7509 | // includePhases - If "true", the per-phase info in "info" is valid, |
7510 | // which means that a "normal" compile has ended; if |
7511 | // the value is "false" we are recording the results |
7512 | // of a partial compile (typically an import-only run |
7513 | // on behalf of the inliner) in which case the phase |
7514 | // info is not valid and so we only record EE call |
7515 | // overhead. |
7516 | void CompTimeSummaryInfo::AddInfo(CompTimeInfo& info, bool includePhases) |
7517 | { |
7518 | if (info.m_timerFailure) |
7519 | { |
7520 | return; // Don't update if there was a failure. |
7521 | } |
7522 | |
7523 | CritSecHolder timeLock(s_compTimeSummaryLock); |
7524 | |
7525 | if (includePhases) |
7526 | { |
7527 | bool includeInFiltered = IncludedInFilteredData(info); |
7528 | |
7529 | m_numMethods++; |
7530 | |
7531 | // Update the totals and maxima. |
7532 | m_total.m_byteCodeBytes += info.m_byteCodeBytes; |
7533 | m_maximum.m_byteCodeBytes = max(m_maximum.m_byteCodeBytes, info.m_byteCodeBytes); |
7534 | m_total.m_totalCycles += info.m_totalCycles; |
7535 | m_maximum.m_totalCycles = max(m_maximum.m_totalCycles, info.m_totalCycles); |
7536 | |
7537 | #if MEASURE_CLRAPI_CALLS |
7538 | // Update the CLR-API values. |
7539 | m_total.m_allClrAPIcalls += info.m_allClrAPIcalls; |
7540 | m_maximum.m_allClrAPIcalls = max(m_maximum.m_allClrAPIcalls, info.m_allClrAPIcalls); |
7541 | m_total.m_allClrAPIcycles += info.m_allClrAPIcycles; |
7542 | m_maximum.m_allClrAPIcycles = max(m_maximum.m_allClrAPIcycles, info.m_allClrAPIcycles); |
7543 | #endif |
7544 | |
7545 | if (includeInFiltered) |
7546 | { |
7547 | m_numFilteredMethods++; |
7548 | m_filtered.m_byteCodeBytes += info.m_byteCodeBytes; |
7549 | m_filtered.m_totalCycles += info.m_totalCycles; |
7550 | m_filtered.m_parentPhaseEndSlop += info.m_parentPhaseEndSlop; |
7551 | } |
7552 | |
7553 | for (int i = 0; i < PHASE_NUMBER_OF; i++) |
7554 | { |
7555 | m_total.m_invokesByPhase[i] += info.m_invokesByPhase[i]; |
7556 | m_total.m_cyclesByPhase[i] += info.m_cyclesByPhase[i]; |
7557 | |
7558 | #if MEASURE_CLRAPI_CALLS |
7559 | m_total.m_CLRinvokesByPhase[i] += info.m_CLRinvokesByPhase[i]; |
7560 | m_total.m_CLRcyclesByPhase[i] += info.m_CLRcyclesByPhase[i]; |
7561 | #endif |
7562 | |
7563 | if (includeInFiltered) |
7564 | { |
7565 | m_filtered.m_invokesByPhase[i] += info.m_invokesByPhase[i]; |
7566 | m_filtered.m_cyclesByPhase[i] += info.m_cyclesByPhase[i]; |
7567 | #if MEASURE_CLRAPI_CALLS |
7568 | m_filtered.m_CLRinvokesByPhase[i] += info.m_CLRinvokesByPhase[i]; |
7569 | m_filtered.m_CLRcyclesByPhase[i] += info.m_CLRcyclesByPhase[i]; |
7570 | #endif |
7571 | } |
7572 | m_maximum.m_cyclesByPhase[i] = max(m_maximum.m_cyclesByPhase[i], info.m_cyclesByPhase[i]); |
7573 | |
7574 | #if MEASURE_CLRAPI_CALLS |
7575 | m_maximum.m_CLRcyclesByPhase[i] = max(m_maximum.m_CLRcyclesByPhase[i], info.m_CLRcyclesByPhase[i]); |
7576 | #endif |
7577 | } |
7578 | m_total.m_parentPhaseEndSlop += info.m_parentPhaseEndSlop; |
7579 | m_maximum.m_parentPhaseEndSlop = max(m_maximum.m_parentPhaseEndSlop, info.m_parentPhaseEndSlop); |
7580 | } |
7581 | #if MEASURE_CLRAPI_CALLS |
7582 | else |
7583 | { |
7584 | m_totMethods++; |
7585 | |
7586 | // Update the "global" CLR-API values. |
7587 | m_total.m_allClrAPIcalls += info.m_allClrAPIcalls; |
7588 | m_maximum.m_allClrAPIcalls = max(m_maximum.m_allClrAPIcalls, info.m_allClrAPIcalls); |
7589 | m_total.m_allClrAPIcycles += info.m_allClrAPIcycles; |
7590 | m_maximum.m_allClrAPIcycles = max(m_maximum.m_allClrAPIcycles, info.m_allClrAPIcycles); |
7591 | |
7592 | // Update the per-phase CLR-API values. |
7593 | m_total.m_invokesByPhase[PHASE_CLR_API] += info.m_allClrAPIcalls; |
7594 | m_maximum.m_invokesByPhase[PHASE_CLR_API] = |
7595 | max(m_maximum.m_perClrAPIcalls[PHASE_CLR_API], info.m_allClrAPIcalls); |
7596 | m_total.m_cyclesByPhase[PHASE_CLR_API] += info.m_allClrAPIcycles; |
7597 | m_maximum.m_cyclesByPhase[PHASE_CLR_API] = |
7598 | max(m_maximum.m_cyclesByPhase[PHASE_CLR_API], info.m_allClrAPIcycles); |
7599 | } |
7600 | |
7601 | for (int i = 0; i < API_ICorJitInfo_Names::API_COUNT; i++) |
7602 | { |
7603 | m_total.m_perClrAPIcalls[i] += info.m_perClrAPIcalls[i]; |
7604 | m_maximum.m_perClrAPIcalls[i] = max(m_maximum.m_perClrAPIcalls[i], info.m_perClrAPIcalls[i]); |
7605 | |
7606 | m_total.m_perClrAPIcycles[i] += info.m_perClrAPIcycles[i]; |
7607 | m_maximum.m_perClrAPIcycles[i] = max(m_maximum.m_perClrAPIcycles[i], info.m_perClrAPIcycles[i]); |
7608 | |
7609 | m_maximum.m_maxClrAPIcycles[i] = max(m_maximum.m_maxClrAPIcycles[i], info.m_maxClrAPIcycles[i]); |
7610 | } |
7611 | #endif |
7612 | } |
7613 | |
7614 | // Static |
7615 | LPCWSTR Compiler::compJitTimeLogFilename = nullptr; |
7616 | |
7617 | void CompTimeSummaryInfo::Print(FILE* f) |
7618 | { |
7619 | if (f == nullptr) |
7620 | { |
7621 | return; |
7622 | } |
7623 | // Otherwise... |
7624 | double countsPerSec = CycleTimer::CyclesPerSecond(); |
7625 | if (countsPerSec == 0.0) |
7626 | { |
7627 | fprintf(f, "Processor does not have a high-frequency timer.\n" ); |
7628 | return; |
7629 | } |
7630 | |
7631 | bool = (JitConfig.JitEECallTimingInfo() != 0); |
7632 | double totTime_ms = 0.0; |
7633 | |
7634 | fprintf(f, "JIT Compilation time report:\n" ); |
7635 | fprintf(f, " Compiled %d methods.\n" , m_numMethods); |
7636 | if (m_numMethods != 0) |
7637 | { |
7638 | fprintf(f, " Compiled %d bytecodes total (%d max, %8.2f avg).\n" , m_total.m_byteCodeBytes, |
7639 | m_maximum.m_byteCodeBytes, (double)m_total.m_byteCodeBytes / (double)m_numMethods); |
7640 | totTime_ms = ((double)m_total.m_totalCycles / countsPerSec) * 1000.0; |
7641 | fprintf(f, " Time: total: %10.3f Mcycles/%10.3f ms\n" , ((double)m_total.m_totalCycles / 1000000.0), |
7642 | totTime_ms); |
7643 | fprintf(f, " max: %10.3f Mcycles/%10.3f ms\n" , ((double)m_maximum.m_totalCycles) / 1000000.0, |
7644 | ((double)m_maximum.m_totalCycles / countsPerSec) * 1000.0); |
7645 | fprintf(f, " avg: %10.3f Mcycles/%10.3f ms\n" , |
7646 | ((double)m_total.m_totalCycles) / 1000000.0 / (double)m_numMethods, totTime_ms / (double)m_numMethods); |
7647 | |
7648 | const char* = "" ; |
7649 | const char* = "" ; |
7650 | #if MEASURE_CLRAPI_CALLS |
7651 | if (extraInfo) |
7652 | { |
7653 | extraHdr1 = " CLRs/meth % in CLR" ; |
7654 | extraHdr2 = "-----------------------" ; |
7655 | } |
7656 | #endif |
7657 | |
7658 | fprintf(f, "\n Total time by phases:\n" ); |
7659 | fprintf(f, " PHASE inv/meth Mcycles time (ms) %% of total max (ms)%s\n" , |
7660 | extraHdr1); |
7661 | fprintf(f, " ---------------------------------------------------------------------------------------%s\n" , |
7662 | extraHdr2); |
7663 | |
7664 | // Ensure that at least the names array and the Phases enum have the same number of entries: |
7665 | assert(_countof(PhaseNames) == PHASE_NUMBER_OF); |
7666 | for (int i = 0; i < PHASE_NUMBER_OF; i++) |
7667 | { |
7668 | double phase_tot_ms = (((double)m_total.m_cyclesByPhase[i]) / countsPerSec) * 1000.0; |
7669 | double phase_max_ms = (((double)m_maximum.m_cyclesByPhase[i]) / countsPerSec) * 1000.0; |
7670 | double phase_tot_pct = 100.0 * phase_tot_ms / totTime_ms; |
7671 | |
7672 | #if MEASURE_CLRAPI_CALLS |
7673 | // Skip showing CLR API call info if we didn't collect any |
7674 | if (i == PHASE_CLR_API && !extraInfo) |
7675 | continue; |
7676 | #endif |
7677 | |
7678 | // Indent nested phases, according to depth. |
7679 | int ancPhase = PhaseParent[i]; |
7680 | while (ancPhase != -1) |
7681 | { |
7682 | fprintf(f, " " ); |
7683 | ancPhase = PhaseParent[ancPhase]; |
7684 | } |
7685 | fprintf(f, " %-30s %6.2f %10.2f %9.3f %8.2f%% %8.3f" , PhaseNames[i], |
7686 | ((double)m_total.m_invokesByPhase[i]) / ((double)m_numMethods), |
7687 | ((double)m_total.m_cyclesByPhase[i]) / 1000000.0, phase_tot_ms, (phase_tot_ms * 100.0 / totTime_ms), |
7688 | phase_max_ms); |
7689 | |
7690 | #if MEASURE_CLRAPI_CALLS |
7691 | if (extraInfo && i != PHASE_CLR_API) |
7692 | { |
7693 | double nest_tot_ms = (((double)m_total.m_CLRcyclesByPhase[i]) / countsPerSec) * 1000.0; |
7694 | double nest_percent = nest_tot_ms * 100.0 / totTime_ms; |
7695 | double calls_per_fn = ((double)m_total.m_CLRinvokesByPhase[i]) / ((double)m_numMethods); |
7696 | |
7697 | if (nest_percent > 0.1 || calls_per_fn > 10) |
7698 | fprintf(f, " %5.1f %8.2f%%" , calls_per_fn, nest_percent); |
7699 | } |
7700 | #endif |
7701 | fprintf(f, "\n" ); |
7702 | } |
7703 | |
7704 | // Show slop if it's over a certain percentage of the total |
7705 | double pslop_pct = 100.0 * m_total.m_parentPhaseEndSlop * 1000.0 / countsPerSec / totTime_ms; |
7706 | if (pslop_pct >= 1.0) |
7707 | { |
7708 | fprintf(f, "\n 'End phase slop' should be very small (if not, there's unattributed time): %9.3f Mcycles = " |
7709 | "%3.1f%% of total.\n\n" , |
7710 | m_total.m_parentPhaseEndSlop / 1000000.0, pslop_pct); |
7711 | } |
7712 | } |
7713 | if (m_numFilteredMethods > 0) |
7714 | { |
7715 | fprintf(f, " Compiled %d methods that meet the filter requirement.\n" , m_numFilteredMethods); |
7716 | fprintf(f, " Compiled %d bytecodes total (%8.2f avg).\n" , m_filtered.m_byteCodeBytes, |
7717 | (double)m_filtered.m_byteCodeBytes / (double)m_numFilteredMethods); |
7718 | double totTime_ms = ((double)m_filtered.m_totalCycles / countsPerSec) * 1000.0; |
7719 | fprintf(f, " Time: total: %10.3f Mcycles/%10.3f ms\n" , ((double)m_filtered.m_totalCycles / 1000000.0), |
7720 | totTime_ms); |
7721 | fprintf(f, " avg: %10.3f Mcycles/%10.3f ms\n" , |
7722 | ((double)m_filtered.m_totalCycles) / 1000000.0 / (double)m_numFilteredMethods, |
7723 | totTime_ms / (double)m_numFilteredMethods); |
7724 | |
7725 | fprintf(f, " Total time by phases:\n" ); |
7726 | fprintf(f, " PHASE inv/meth Mcycles time (ms) %% of total\n" ); |
7727 | fprintf(f, " --------------------------------------------------------------------------------------\n" ); |
7728 | // Ensure that at least the names array and the Phases enum have the same number of entries: |
7729 | assert(_countof(PhaseNames) == PHASE_NUMBER_OF); |
7730 | for (int i = 0; i < PHASE_NUMBER_OF; i++) |
7731 | { |
7732 | double phase_tot_ms = (((double)m_filtered.m_cyclesByPhase[i]) / countsPerSec) * 1000.0; |
7733 | // Indent nested phases, according to depth. |
7734 | int ancPhase = PhaseParent[i]; |
7735 | while (ancPhase != -1) |
7736 | { |
7737 | fprintf(f, " " ); |
7738 | ancPhase = PhaseParent[ancPhase]; |
7739 | } |
7740 | fprintf(f, " %-30s %5.2f %10.2f %9.3f %8.2f%%\n" , PhaseNames[i], |
7741 | ((double)m_filtered.m_invokesByPhase[i]) / ((double)m_numFilteredMethods), |
7742 | ((double)m_filtered.m_cyclesByPhase[i]) / 1000000.0, phase_tot_ms, |
7743 | (phase_tot_ms * 100.0 / totTime_ms)); |
7744 | } |
7745 | |
7746 | double fslop_ms = m_filtered.m_parentPhaseEndSlop * 1000.0 / countsPerSec; |
7747 | if (fslop_ms > 1.0) |
7748 | { |
7749 | fprintf(f, "\n 'End phase slop' should be very small (if not, there's unattributed time): %9.3f Mcycles = " |
7750 | "%3.1f%% of total.\n\n" , |
7751 | m_filtered.m_parentPhaseEndSlop / 1000000.0, fslop_ms); |
7752 | } |
7753 | } |
7754 | |
7755 | #if MEASURE_CLRAPI_CALLS |
7756 | if (m_total.m_allClrAPIcalls > 0 && m_total.m_allClrAPIcycles > 0) |
7757 | { |
7758 | fprintf(f, "\n" ); |
7759 | if (m_totMethods > 0) |
7760 | fprintf(f, " Imported %u methods.\n\n" , m_numMethods + m_totMethods); |
7761 | |
7762 | fprintf(f, " CLR API # calls total time max time avg time %% " |
7763 | "of total\n" ); |
7764 | fprintf(f, " -------------------------------------------------------------------------------" ); |
7765 | fprintf(f, "---------------------\n" ); |
7766 | |
7767 | static const char* APInames[] = { |
7768 | #define DEF_CLR_API(name) #name, |
7769 | #include "ICorJitInfo_API_names.h" |
7770 | }; |
7771 | |
7772 | unsigned shownCalls = 0; |
7773 | double shownMillis = 0.0; |
7774 | #ifdef DEBUG |
7775 | unsigned checkedCalls = 0; |
7776 | double checkedMillis = 0.0; |
7777 | #endif |
7778 | |
7779 | for (unsigned pass = 0; pass < 2; pass++) |
7780 | { |
7781 | for (unsigned i = 0; i < API_ICorJitInfo_Names::API_COUNT; i++) |
7782 | { |
7783 | unsigned calls = m_total.m_perClrAPIcalls[i]; |
7784 | if (calls == 0) |
7785 | continue; |
7786 | |
7787 | unsigned __int64 cycles = m_total.m_perClrAPIcycles[i]; |
7788 | double millis = 1000.0 * cycles / countsPerSec; |
7789 | |
7790 | // Don't show the small fry to keep the results manageable |
7791 | if (millis < 0.5) |
7792 | { |
7793 | // We always show the following API because it is always called |
7794 | // exactly once for each method and its body is the simplest one |
7795 | // possible (it just returns an integer constant), and therefore |
7796 | // it can be used to measure the overhead of adding the CLR API |
7797 | // timing code. Roughly speaking, on a 3GHz x64 box the overhead |
7798 | // per call should be around 40 ns when using RDTSC, compared to |
7799 | // about 140 ns when using GetThreadCycles() under Windows. |
7800 | if (i != API_ICorJitInfo_Names::API_getExpectedTargetArchitecture) |
7801 | continue; |
7802 | } |
7803 | |
7804 | // In the first pass we just compute the totals. |
7805 | if (pass == 0) |
7806 | { |
7807 | shownCalls += m_total.m_perClrAPIcalls[i]; |
7808 | shownMillis += millis; |
7809 | continue; |
7810 | } |
7811 | |
7812 | unsigned __int32 maxcyc = m_maximum.m_maxClrAPIcycles[i]; |
7813 | double max_ms = 1000.0 * maxcyc / countsPerSec; |
7814 | |
7815 | fprintf(f, " %-40s" , APInames[i]); // API name |
7816 | fprintf(f, " %8u %9.1f ms" , calls, millis); // #calls, total time |
7817 | fprintf(f, " %8.1f ms %8.1f ns" , max_ms, 1000000.0 * millis / calls); // max, avg time |
7818 | fprintf(f, " %5.1f%%\n" , 100.0 * millis / shownMillis); // % of total |
7819 | |
7820 | #ifdef DEBUG |
7821 | checkedCalls += m_total.m_perClrAPIcalls[i]; |
7822 | checkedMillis += millis; |
7823 | #endif |
7824 | } |
7825 | } |
7826 | |
7827 | #ifdef DEBUG |
7828 | assert(checkedCalls == shownCalls); |
7829 | assert(checkedMillis == shownMillis); |
7830 | #endif |
7831 | |
7832 | if (shownCalls > 0 || shownMillis > 0) |
7833 | { |
7834 | fprintf(f, " -------------------------" ); |
7835 | fprintf(f, "---------------------------------------------------------------------------\n" ); |
7836 | fprintf(f, " Total for calls shown above %8u %10.1f ms" , shownCalls, shownMillis); |
7837 | if (totTime_ms > 0.0) |
7838 | fprintf(f, " (%4.1lf%% of overall JIT time)" , shownMillis * 100.0 / totTime_ms); |
7839 | fprintf(f, "\n" ); |
7840 | } |
7841 | fprintf(f, "\n" ); |
7842 | } |
7843 | #endif |
7844 | |
7845 | fprintf(f, "\n" ); |
7846 | } |
7847 | |
7848 | JitTimer::JitTimer(unsigned byteCodeSize) : m_info(byteCodeSize) |
7849 | { |
7850 | #if MEASURE_CLRAPI_CALLS |
7851 | m_CLRcallInvokes = 0; |
7852 | m_CLRcallCycles = 0; |
7853 | #endif |
7854 | |
7855 | #ifdef DEBUG |
7856 | m_lastPhase = (Phases)-1; |
7857 | #if MEASURE_CLRAPI_CALLS |
7858 | m_CLRcallAPInum = -1; |
7859 | #endif |
7860 | #endif |
7861 | |
7862 | unsigned __int64 threadCurCycles; |
7863 | if (_our_GetThreadCycles(&threadCurCycles)) |
7864 | { |
7865 | m_start = threadCurCycles; |
7866 | m_curPhaseStart = threadCurCycles; |
7867 | } |
7868 | } |
7869 | |
7870 | void JitTimer::EndPhase(Compiler* compiler, Phases phase) |
7871 | { |
7872 | // Otherwise... |
7873 | // We re-run some phases currently, so this following assert doesn't work. |
7874 | // assert((int)phase > (int)m_lastPhase); // We should end phases in increasing order. |
7875 | |
7876 | unsigned __int64 threadCurCycles; |
7877 | if (_our_GetThreadCycles(&threadCurCycles)) |
7878 | { |
7879 | unsigned __int64 phaseCycles = (threadCurCycles - m_curPhaseStart); |
7880 | |
7881 | // If this is not a leaf phase, the assumption is that the last subphase must have just recently ended. |
7882 | // Credit the duration to "slop", the total of which should be very small. |
7883 | if (PhaseHasChildren[phase]) |
7884 | { |
7885 | m_info.m_parentPhaseEndSlop += phaseCycles; |
7886 | } |
7887 | else |
7888 | { |
7889 | // It is a leaf phase. Credit duration to it. |
7890 | m_info.m_invokesByPhase[phase]++; |
7891 | m_info.m_cyclesByPhase[phase] += phaseCycles; |
7892 | |
7893 | #if MEASURE_CLRAPI_CALLS |
7894 | // Record the CLR API timing info as well. |
7895 | m_info.m_CLRinvokesByPhase[phase] += m_CLRcallInvokes; |
7896 | m_info.m_CLRcyclesByPhase[phase] += m_CLRcallCycles; |
7897 | #endif |
7898 | |
7899 | // Credit the phase's ancestors, if any. |
7900 | int ancPhase = PhaseParent[phase]; |
7901 | while (ancPhase != -1) |
7902 | { |
7903 | m_info.m_cyclesByPhase[ancPhase] += phaseCycles; |
7904 | ancPhase = PhaseParent[ancPhase]; |
7905 | } |
7906 | |
7907 | #if MEASURE_CLRAPI_CALLS |
7908 | const Phases lastPhase = PHASE_CLR_API; |
7909 | #else |
7910 | const Phases lastPhase = PHASE_NUMBER_OF; |
7911 | #endif |
7912 | if (phase + 1 == lastPhase) |
7913 | { |
7914 | m_info.m_totalCycles = (threadCurCycles - m_start); |
7915 | } |
7916 | else |
7917 | { |
7918 | m_curPhaseStart = threadCurCycles; |
7919 | } |
7920 | } |
7921 | |
7922 | if ((JitConfig.JitMeasureIR() != 0) && PhaseReportsIRSize[phase]) |
7923 | { |
7924 | m_info.m_nodeCountAfterPhase[phase] = compiler->fgMeasureIR(); |
7925 | } |
7926 | else |
7927 | { |
7928 | m_info.m_nodeCountAfterPhase[phase] = 0; |
7929 | } |
7930 | } |
7931 | |
7932 | #ifdef DEBUG |
7933 | m_lastPhase = phase; |
7934 | #endif |
7935 | #if MEASURE_CLRAPI_CALLS |
7936 | m_CLRcallInvokes = 0; |
7937 | m_CLRcallCycles = 0; |
7938 | #endif |
7939 | } |
7940 | |
7941 | #if MEASURE_CLRAPI_CALLS |
7942 | |
7943 | //------------------------------------------------------------------------ |
7944 | // JitTimer::CLRApiCallEnter: Start the stopwatch for an EE call. |
7945 | // |
7946 | // Arguments: |
7947 | // apix - The API index - an "enum API_ICorJitInfo_Names" value. |
7948 | // |
7949 | |
7950 | void JitTimer::CLRApiCallEnter(unsigned apix) |
7951 | { |
7952 | assert(m_CLRcallAPInum == -1); // Nested calls not allowed |
7953 | m_CLRcallAPInum = apix; |
7954 | |
7955 | // If we can't get the cycles, we'll just ignore this call |
7956 | if (!_our_GetThreadCycles(&m_CLRcallStart)) |
7957 | m_CLRcallStart = 0; |
7958 | } |
7959 | |
7960 | //------------------------------------------------------------------------ |
7961 | // JitTimer::CLRApiCallLeave: compute / record time spent in an EE call. |
7962 | // |
7963 | // Arguments: |
7964 | // apix - The API's "enum API_ICorJitInfo_Names" value; this value |
7965 | // should match the value passed to the most recent call to |
7966 | // "CLRApiCallEnter" (i.e. these must come as matched pairs), |
7967 | // and they also may not nest. |
7968 | // |
7969 | |
7970 | void JitTimer::CLRApiCallLeave(unsigned apix) |
7971 | { |
7972 | // Make sure we're actually inside a measured CLR call. |
7973 | assert(m_CLRcallAPInum != -1); |
7974 | m_CLRcallAPInum = -1; |
7975 | |
7976 | // Ignore this one if we don't have a valid starting counter. |
7977 | if (m_CLRcallStart != 0) |
7978 | { |
7979 | if (JitConfig.JitEECallTimingInfo() != 0) |
7980 | { |
7981 | unsigned __int64 threadCurCycles; |
7982 | if (_our_GetThreadCycles(&threadCurCycles)) |
7983 | { |
7984 | // Compute the cycles spent in the call. |
7985 | threadCurCycles -= m_CLRcallStart; |
7986 | |
7987 | // Add the cycles to the 'phase' and bump its use count. |
7988 | m_info.m_cyclesByPhase[PHASE_CLR_API] += threadCurCycles; |
7989 | m_info.m_invokesByPhase[PHASE_CLR_API] += 1; |
7990 | |
7991 | // Add the values to the "per API" info. |
7992 | m_info.m_allClrAPIcycles += threadCurCycles; |
7993 | m_info.m_allClrAPIcalls += 1; |
7994 | |
7995 | m_info.m_perClrAPIcalls[apix] += 1; |
7996 | m_info.m_perClrAPIcycles[apix] += threadCurCycles; |
7997 | m_info.m_maxClrAPIcycles[apix] = max(m_info.m_maxClrAPIcycles[apix], (unsigned __int32)threadCurCycles); |
7998 | |
7999 | // Subtract the cycles from the enclosing phase by bumping its start time |
8000 | m_curPhaseStart += threadCurCycles; |
8001 | |
8002 | // Update the running totals. |
8003 | m_CLRcallInvokes += 1; |
8004 | m_CLRcallCycles += threadCurCycles; |
8005 | } |
8006 | } |
8007 | |
8008 | m_CLRcallStart = 0; |
8009 | } |
8010 | |
8011 | assert(m_CLRcallAPInum != -1); // No longer in this API call. |
8012 | m_CLRcallAPInum = -1; |
8013 | } |
8014 | |
8015 | #endif // MEASURE_CLRAPI_CALLS |
8016 | |
8017 | CritSecObject JitTimer::s_csvLock; |
8018 | |
8019 | LPCWSTR Compiler::JitTimeLogCsv() |
8020 | { |
8021 | LPCWSTR jitTimeLogCsv = JitConfig.JitTimeLogCsv(); |
8022 | return jitTimeLogCsv; |
8023 | } |
8024 | |
8025 | void JitTimer::() |
8026 | { |
8027 | LPCWSTR jitTimeLogCsv = Compiler::JitTimeLogCsv(); |
8028 | if (jitTimeLogCsv == nullptr) |
8029 | { |
8030 | return; |
8031 | } |
8032 | |
8033 | CritSecHolder csvLock(s_csvLock); |
8034 | |
8035 | FILE* fp = _wfopen(jitTimeLogCsv, W("a" )); |
8036 | if (fp != nullptr) |
8037 | { |
8038 | // Seek to the end of the file s.t. `ftell` doesn't lie to us on Windows |
8039 | fseek(fp, 0, SEEK_END); |
8040 | |
8041 | // Write the header if the file is empty |
8042 | if (ftell(fp) == 0) |
8043 | { |
8044 | fprintf(fp, "\"Method Name\"," ); |
8045 | fprintf(fp, "\"Assembly or SPMI Index\"," ); |
8046 | fprintf(fp, "\"IL Bytes\"," ); |
8047 | fprintf(fp, "\"Basic Blocks\"," ); |
8048 | fprintf(fp, "\"Min Opts\"," ); |
8049 | fprintf(fp, "\"Loops Cloned\"," ); |
8050 | |
8051 | for (int i = 0; i < PHASE_NUMBER_OF; i++) |
8052 | { |
8053 | fprintf(fp, "\"%s\"," , PhaseNames[i]); |
8054 | if ((JitConfig.JitMeasureIR() != 0) && PhaseReportsIRSize[i]) |
8055 | { |
8056 | fprintf(fp, "\"Node Count After %s\"," , PhaseNames[i]); |
8057 | } |
8058 | } |
8059 | |
8060 | InlineStrategy::DumpCsvHeader(fp); |
8061 | |
8062 | fprintf(fp, "\"Executable Code Bytes\"," ); |
8063 | fprintf(fp, "\"GC Info Bytes\"," ); |
8064 | fprintf(fp, "\"Total Bytes Allocated\"," ); |
8065 | fprintf(fp, "\"Total Cycles\"," ); |
8066 | fprintf(fp, "\"CPS\"\n" ); |
8067 | } |
8068 | fclose(fp); |
8069 | } |
8070 | } |
8071 | |
8072 | extern ICorJitHost* g_jitHost; |
8073 | |
8074 | void JitTimer::PrintCsvMethodStats(Compiler* comp) |
8075 | { |
8076 | LPCWSTR jitTimeLogCsv = Compiler::JitTimeLogCsv(); |
8077 | if (jitTimeLogCsv == nullptr) |
8078 | { |
8079 | return; |
8080 | } |
8081 | |
8082 | // eeGetMethodFullName uses locks, so don't enter crit sec before this call. |
8083 | const char* methName = comp->eeGetMethodFullName(comp->info.compMethodHnd); |
8084 | |
8085 | // Try and access the SPMI index to report in the data set. |
8086 | // |
8087 | // If the jit is not hosted under SPMI this will return the |
8088 | // default value of zero. |
8089 | // |
8090 | // Query the jit host directly here instead of going via the |
8091 | // config cache, since value will change for each method. |
8092 | int index = g_jitHost->getIntConfigValue(W("SuperPMIMethodContextNumber" ), 0); |
8093 | |
8094 | CritSecHolder csvLock(s_csvLock); |
8095 | |
8096 | FILE* fp = _wfopen(jitTimeLogCsv, W("a" )); |
8097 | fprintf(fp, "\"%s\"," , methName); |
8098 | if (index != 0) |
8099 | { |
8100 | fprintf(fp, "%d," , index); |
8101 | } |
8102 | else |
8103 | { |
8104 | const char* methodAssemblyName = comp->info.compCompHnd->getAssemblyName( |
8105 | comp->info.compCompHnd->getModuleAssembly(comp->info.compCompHnd->getClassModule(comp->info.compClassHnd))); |
8106 | fprintf(fp, "\"%s\"," , methodAssemblyName); |
8107 | } |
8108 | fprintf(fp, "%u," , comp->info.compILCodeSize); |
8109 | fprintf(fp, "%u," , comp->fgBBcount); |
8110 | fprintf(fp, "%u," , comp->opts.MinOpts()); |
8111 | fprintf(fp, "%u," , comp->optLoopsCloned); |
8112 | unsigned __int64 totCycles = 0; |
8113 | for (int i = 0; i < PHASE_NUMBER_OF; i++) |
8114 | { |
8115 | if (!PhaseHasChildren[i]) |
8116 | { |
8117 | totCycles += m_info.m_cyclesByPhase[i]; |
8118 | } |
8119 | fprintf(fp, "%I64u," , m_info.m_cyclesByPhase[i]); |
8120 | |
8121 | if ((JitConfig.JitMeasureIR() != 0) && PhaseReportsIRSize[i]) |
8122 | { |
8123 | fprintf(fp, "%u," , m_info.m_nodeCountAfterPhase[i]); |
8124 | } |
8125 | } |
8126 | |
8127 | comp->m_inlineStrategy->DumpCsvData(fp); |
8128 | |
8129 | fprintf(fp, "%u," , comp->info.compNativeCodeSize); |
8130 | fprintf(fp, "%Iu," , comp->compInfoBlkSize); |
8131 | fprintf(fp, "%Iu," , comp->compGetArenaAllocator()->getTotalBytesAllocated()); |
8132 | fprintf(fp, "%I64u," , m_info.m_totalCycles); |
8133 | fprintf(fp, "%f\n" , CycleTimer::CyclesPerSecond()); |
8134 | fclose(fp); |
8135 | } |
8136 | |
8137 | // Completes the timing of the current method, and adds it to "sum". |
8138 | void JitTimer::Terminate(Compiler* comp, CompTimeSummaryInfo& sum, bool includePhases) |
8139 | { |
8140 | if (includePhases) |
8141 | { |
8142 | PrintCsvMethodStats(comp); |
8143 | } |
8144 | |
8145 | sum.AddInfo(m_info, includePhases); |
8146 | } |
8147 | #endif // FEATURE_JIT_METHOD_PERF |
8148 | |
8149 | #if LOOP_HOIST_STATS |
8150 | // Static fields. |
8151 | CritSecObject Compiler::s_loopHoistStatsLock; // Default constructor. |
8152 | unsigned Compiler::s_loopsConsidered = 0; |
8153 | unsigned Compiler::s_loopsWithHoistedExpressions = 0; |
8154 | unsigned Compiler::s_totalHoistedExpressions = 0; |
8155 | |
8156 | // static |
8157 | void Compiler::PrintAggregateLoopHoistStats(FILE* f) |
8158 | { |
8159 | fprintf(f, "\n" ); |
8160 | fprintf(f, "---------------------------------------------------\n" ); |
8161 | fprintf(f, "Loop hoisting stats\n" ); |
8162 | fprintf(f, "---------------------------------------------------\n" ); |
8163 | |
8164 | double pctWithHoisted = 0.0; |
8165 | if (s_loopsConsidered > 0) |
8166 | { |
8167 | pctWithHoisted = 100.0 * (double(s_loopsWithHoistedExpressions) / double(s_loopsConsidered)); |
8168 | } |
8169 | double exprsPerLoopWithExpr = 0.0; |
8170 | if (s_loopsWithHoistedExpressions > 0) |
8171 | { |
8172 | exprsPerLoopWithExpr = double(s_totalHoistedExpressions) / double(s_loopsWithHoistedExpressions); |
8173 | } |
8174 | fprintf(f, "Considered %d loops. Of these, we hoisted expressions out of %d (%6.2f%%).\n" , s_loopsConsidered, |
8175 | s_loopsWithHoistedExpressions, pctWithHoisted); |
8176 | fprintf(f, " A total of %d expressions were hoisted, an average of %5.2f per loop-with-hoisted-expr.\n" , |
8177 | s_totalHoistedExpressions, exprsPerLoopWithExpr); |
8178 | } |
8179 | |
8180 | void Compiler::AddLoopHoistStats() |
8181 | { |
8182 | CritSecHolder statsLock(s_loopHoistStatsLock); |
8183 | |
8184 | s_loopsConsidered += m_loopsConsidered; |
8185 | s_loopsWithHoistedExpressions += m_loopsWithHoistedExpressions; |
8186 | s_totalHoistedExpressions += m_totalHoistedExpressions; |
8187 | } |
8188 | |
8189 | void Compiler::PrintPerMethodLoopHoistStats() |
8190 | { |
8191 | double pctWithHoisted = 0.0; |
8192 | if (m_loopsConsidered > 0) |
8193 | { |
8194 | pctWithHoisted = 100.0 * (double(m_loopsWithHoistedExpressions) / double(m_loopsConsidered)); |
8195 | } |
8196 | double exprsPerLoopWithExpr = 0.0; |
8197 | if (m_loopsWithHoistedExpressions > 0) |
8198 | { |
8199 | exprsPerLoopWithExpr = double(m_totalHoistedExpressions) / double(m_loopsWithHoistedExpressions); |
8200 | } |
8201 | printf("Considered %d loops. Of these, we hoisted expressions out of %d (%5.2f%%).\n" , m_loopsConsidered, |
8202 | m_loopsWithHoistedExpressions, pctWithHoisted); |
8203 | printf(" A total of %d expressions were hoisted, an average of %5.2f per loop-with-hoisted-expr.\n" , |
8204 | m_totalHoistedExpressions, exprsPerLoopWithExpr); |
8205 | } |
8206 | #endif // LOOP_HOIST_STATS |
8207 | |
8208 | //------------------------------------------------------------------------ |
8209 | // RecordStateAtEndOfInlining: capture timing data (if enabled) after |
8210 | // inlining as completed. |
8211 | // |
8212 | // Note: |
8213 | // Records data needed for SQM and inlining data dumps. Should be |
8214 | // called after inlining is complete. (We do this after inlining |
8215 | // because this marks the last point at which the JIT is likely to |
8216 | // cause type-loading and class initialization). |
8217 | |
8218 | void Compiler::RecordStateAtEndOfInlining() |
8219 | { |
8220 | #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM) |
8221 | |
8222 | m_compCyclesAtEndOfInlining = 0; |
8223 | m_compTickCountAtEndOfInlining = 0; |
8224 | bool b = CycleTimer::GetThreadCyclesS(&m_compCyclesAtEndOfInlining); |
8225 | if (!b) |
8226 | { |
8227 | return; // We don't have a thread cycle counter. |
8228 | } |
8229 | m_compTickCountAtEndOfInlining = GetTickCount(); |
8230 | |
8231 | #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM) |
8232 | } |
8233 | |
8234 | //------------------------------------------------------------------------ |
8235 | // RecordStateAtEndOfCompilation: capture timing data (if enabled) after |
8236 | // compilation is completed. |
8237 | |
8238 | void Compiler::RecordStateAtEndOfCompilation() |
8239 | { |
8240 | #if defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM) |
8241 | |
8242 | // Common portion |
8243 | m_compCycles = 0; |
8244 | unsigned __int64 compCyclesAtEnd; |
8245 | bool b = CycleTimer::GetThreadCyclesS(&compCyclesAtEnd); |
8246 | if (!b) |
8247 | { |
8248 | return; // We don't have a thread cycle counter. |
8249 | } |
8250 | assert(compCyclesAtEnd >= m_compCyclesAtEndOfInlining); |
8251 | |
8252 | m_compCycles = compCyclesAtEnd - m_compCyclesAtEndOfInlining; |
8253 | |
8254 | #endif // defined(DEBUG) || defined(INLINE_DATA) || defined(FEATURE_CLRSQM) |
8255 | |
8256 | #ifdef FEATURE_CLRSQM |
8257 | |
8258 | // SQM only portion |
8259 | unsigned __int64 mcycles64 = m_compCycles / ((unsigned __int64)1000000); |
8260 | unsigned mcycles; |
8261 | if (mcycles64 > UINT32_MAX) |
8262 | { |
8263 | mcycles = UINT32_MAX; |
8264 | } |
8265 | else |
8266 | { |
8267 | mcycles = (unsigned)mcycles64; |
8268 | } |
8269 | |
8270 | DWORD ticksAtEnd = GetTickCount(); |
8271 | assert(ticksAtEnd >= m_compTickCountAtEndOfInlining); |
8272 | DWORD compTicks = ticksAtEnd - m_compTickCountAtEndOfInlining; |
8273 | |
8274 | if (mcycles >= 1000) |
8275 | { |
8276 | info.compCompHnd->logSQMLongJitEvent(mcycles, compTicks, info.compILCodeSize, fgBBcount, opts.MinOpts(), |
8277 | info.compMethodHnd); |
8278 | } |
8279 | |
8280 | #endif // FEATURE_CLRSQM |
8281 | } |
8282 | |
8283 | #if FUNC_INFO_LOGGING |
8284 | // static |
8285 | LPCWSTR Compiler::compJitFuncInfoFilename = nullptr; |
8286 | |
8287 | // static |
8288 | FILE* Compiler::compJitFuncInfoFile = nullptr; |
8289 | #endif // FUNC_INFO_LOGGING |
8290 | |
8291 | #ifdef DEBUG |
8292 | |
8293 | // dumpConvertedVarSet() dumps the varset bits that are tracked |
8294 | // variable indices, and we convert them to variable numbers, sort the variable numbers, and |
8295 | // print them as variable numbers. To do this, we use a temporary set indexed by |
8296 | // variable number. We can't use the "all varset" type because it is still size-limited, and might |
8297 | // not be big enough to handle all possible variable numbers. |
8298 | void dumpConvertedVarSet(Compiler* comp, VARSET_VALARG_TP vars) |
8299 | { |
8300 | BYTE* pVarNumSet; // trivial set: one byte per varNum, 0 means not in set, 1 means in set. |
8301 | |
8302 | size_t varNumSetBytes = comp->lvaCount * sizeof(BYTE); |
8303 | pVarNumSet = (BYTE*)_alloca(varNumSetBytes); |
8304 | memset(pVarNumSet, 0, varNumSetBytes); // empty the set |
8305 | |
8306 | VarSetOps::Iter iter(comp, vars); |
8307 | unsigned varIndex = 0; |
8308 | while (iter.NextElem(&varIndex)) |
8309 | { |
8310 | unsigned varNum = comp->lvaTrackedToVarNum[varIndex]; |
8311 | assert(varNum < comp->lvaCount); |
8312 | pVarNumSet[varNum] = 1; // This varNum is in the set |
8313 | } |
8314 | |
8315 | bool first = true; |
8316 | printf("{" ); |
8317 | for (size_t varNum = 0; varNum < comp->lvaCount; varNum++) |
8318 | { |
8319 | if (pVarNumSet[varNum] == 1) |
8320 | { |
8321 | if (!first) |
8322 | { |
8323 | printf(" " ); |
8324 | } |
8325 | printf("V%02u" , varNum); |
8326 | first = false; |
8327 | } |
8328 | } |
8329 | printf("}" ); |
8330 | } |
8331 | |
8332 | /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
8333 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
8334 | XX XX |
8335 | XX Debugging helpers XX |
8336 | XX XX |
8337 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
8338 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
8339 | */ |
8340 | |
8341 | /*****************************************************************************/ |
8342 | /* The following functions are intended to be called from the debugger, to dump |
8343 | * various data structures. |
8344 | * |
8345 | * The versions that start with 'c' take a Compiler* as the first argument. |
8346 | * The versions that start with 'd' use the tlsCompiler, so don't require a Compiler*. |
8347 | * |
8348 | * Summary: |
8349 | * cBlock, dBlock : Display a basic block (call fgTableDispBasicBlock()). |
8350 | * cBlocks, dBlocks : Display all the basic blocks of a function (call fgDispBasicBlocks()). |
8351 | * cBlocksV, dBlocksV : Display all the basic blocks of a function (call fgDispBasicBlocks(true)). |
8352 | * "V" means "verbose", and will dump all the trees. |
8353 | * cTree, dTree : Display a tree (call gtDispTree()). |
8354 | * cTreeLIR, dTreeLIR : Display a tree in LIR form (call gtDispLIRNode()). |
8355 | * cTrees, dTrees : Display all the trees in a function (call fgDumpTrees()). |
8356 | * cEH, dEH : Display the EH handler table (call fgDispHandlerTab()). |
8357 | * cVar, dVar : Display a local variable given its number (call lvaDumpEntry()). |
8358 | * cVarDsc, dVarDsc : Display a local variable given a LclVarDsc* (call lvaDumpEntry()). |
8359 | * cVars, dVars : Display the local variable table (call lvaTableDump()). |
8360 | * cVarsFinal, dVarsFinal : Display the local variable table (call lvaTableDump(FINAL_FRAME_LAYOUT)). |
8361 | * cBlockCheapPreds, dBlockCheapPreds : Display a block's cheap predecessors (call block->dspCheapPreds()). |
8362 | * cBlockPreds, dBlockPreds : Display a block's predecessors (call block->dspPreds()). |
8363 | * cBlockSuccs, dBlockSuccs : Display a block's successors (call block->dspSuccs(compiler)). |
8364 | * cReach, dReach : Display all block reachability (call fgDispReach()). |
8365 | * cDoms, dDoms : Display all block dominators (call fgDispDoms()). |
8366 | * cLiveness, dLiveness : Display per-block variable liveness (call fgDispBBLiveness()). |
8367 | * cCVarSet, dCVarSet : Display a "converted" VARSET_TP: the varset is assumed to be tracked variable |
8368 | * indices. These are converted to variable numbers and sorted. (Calls |
8369 | * dumpConvertedVarSet()). |
8370 | * |
8371 | * cFuncIR, dFuncIR : Display all the basic blocks of a function in linear IR form. |
8372 | * cLoopIR, dLoopIR : Display a loop in linear IR form. |
8373 | * dLoopNumIR : Display a loop (given number) in linear IR form. |
8374 | * cBlockIR, dBlockIR : Display a basic block in linear IR form. |
8375 | * cTreeIR, dTreeIR : Display a tree in linear IR form. |
8376 | * dTabStopIR : Display spaces to the next tab stop column |
8377 | * cTreeTypeIR dTreeTypeIR : Display tree type |
8378 | * cTreeKindsIR dTreeKindsIR : Display tree kinds |
8379 | * cTreeFlagsIR dTreeFlagsIR : Display tree flags |
8380 | * cOperandIR dOperandIR : Display tree operand |
8381 | * cLeafIR dLeafIR : Display tree leaf |
8382 | * cIndirIR dIndirIR : Display indir tree as [t#] or [leaf] |
8383 | * cListIR dListIR : Display tree list |
8384 | * cSsaNumIR dSsaNumIR : Display SSA number as <u|d:#> |
8385 | * cValNumIR dValNumIR : Display Value number as <v{l|c}:#{,R}> |
8386 | * cDependsIR : Display dependencies of a tree DEP(t# ...) node |
8387 | * based on child comma tree nodes |
8388 | * dFormatIR : Display dump format specified on command line |
8389 | * |
8390 | * |
8391 | * The following don't require a Compiler* to work: |
8392 | * dRegMask : Display a regMaskTP (call dspRegMask(mask)). |
8393 | */ |
8394 | |
8395 | void cBlock(Compiler* comp, BasicBlock* block) |
8396 | { |
8397 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8398 | printf("===================================================================== *Block %u\n" , sequenceNumber++); |
8399 | comp->fgTableDispBasicBlock(block); |
8400 | } |
8401 | |
8402 | void cBlocks(Compiler* comp) |
8403 | { |
8404 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8405 | printf("===================================================================== *Blocks %u\n" , sequenceNumber++); |
8406 | comp->fgDispBasicBlocks(); |
8407 | } |
8408 | |
8409 | void cBlocksV(Compiler* comp) |
8410 | { |
8411 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8412 | printf("===================================================================== *BlocksV %u\n" , sequenceNumber++); |
8413 | comp->fgDispBasicBlocks(true); |
8414 | } |
8415 | |
8416 | void cTree(Compiler* comp, GenTree* tree) |
8417 | { |
8418 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8419 | printf("===================================================================== *Tree %u\n" , sequenceNumber++); |
8420 | comp->gtDispTree(tree, nullptr, ">>>" ); |
8421 | } |
8422 | |
8423 | void cTreeLIR(Compiler* comp, GenTree* tree) |
8424 | { |
8425 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8426 | printf("===================================================================== *TreeLIR %u\n" , sequenceNumber++); |
8427 | comp->gtDispLIRNode(tree); |
8428 | } |
8429 | |
8430 | void cTrees(Compiler* comp) |
8431 | { |
8432 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8433 | printf("===================================================================== *Trees %u\n" , sequenceNumber++); |
8434 | comp->fgDumpTrees(comp->fgFirstBB, nullptr); |
8435 | } |
8436 | |
8437 | void cEH(Compiler* comp) |
8438 | { |
8439 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8440 | printf("===================================================================== *EH %u\n" , sequenceNumber++); |
8441 | comp->fgDispHandlerTab(); |
8442 | } |
8443 | |
8444 | void cVar(Compiler* comp, unsigned lclNum) |
8445 | { |
8446 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8447 | printf("===================================================================== *Var %u\n" , sequenceNumber++); |
8448 | comp->lvaDumpEntry(lclNum, Compiler::FINAL_FRAME_LAYOUT); |
8449 | } |
8450 | |
8451 | void cVarDsc(Compiler* comp, LclVarDsc* varDsc) |
8452 | { |
8453 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8454 | printf("===================================================================== *VarDsc %u\n" , sequenceNumber++); |
8455 | unsigned lclNum = (unsigned)(varDsc - comp->lvaTable); |
8456 | comp->lvaDumpEntry(lclNum, Compiler::FINAL_FRAME_LAYOUT); |
8457 | } |
8458 | |
8459 | void cVars(Compiler* comp) |
8460 | { |
8461 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8462 | printf("===================================================================== *Vars %u\n" , sequenceNumber++); |
8463 | comp->lvaTableDump(); |
8464 | } |
8465 | |
8466 | void cVarsFinal(Compiler* comp) |
8467 | { |
8468 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8469 | printf("===================================================================== *Vars %u\n" , sequenceNumber++); |
8470 | comp->lvaTableDump(Compiler::FINAL_FRAME_LAYOUT); |
8471 | } |
8472 | |
8473 | void cBlockCheapPreds(Compiler* comp, BasicBlock* block) |
8474 | { |
8475 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8476 | printf("===================================================================== *BlockCheapPreds %u\n" , |
8477 | sequenceNumber++); |
8478 | block->dspCheapPreds(); |
8479 | } |
8480 | |
8481 | void cBlockPreds(Compiler* comp, BasicBlock* block) |
8482 | { |
8483 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8484 | printf("===================================================================== *BlockPreds %u\n" , sequenceNumber++); |
8485 | block->dspPreds(); |
8486 | } |
8487 | |
8488 | void cBlockSuccs(Compiler* comp, BasicBlock* block) |
8489 | { |
8490 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8491 | printf("===================================================================== *BlockSuccs %u\n" , sequenceNumber++); |
8492 | block->dspSuccs(comp); |
8493 | } |
8494 | |
8495 | void cReach(Compiler* comp) |
8496 | { |
8497 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8498 | printf("===================================================================== *Reach %u\n" , sequenceNumber++); |
8499 | comp->fgDispReach(); |
8500 | } |
8501 | |
8502 | void cDoms(Compiler* comp) |
8503 | { |
8504 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8505 | printf("===================================================================== *Doms %u\n" , sequenceNumber++); |
8506 | comp->fgDispDoms(); |
8507 | } |
8508 | |
8509 | void cLiveness(Compiler* comp) |
8510 | { |
8511 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8512 | printf("===================================================================== *Liveness %u\n" , sequenceNumber++); |
8513 | comp->fgDispBBLiveness(); |
8514 | } |
8515 | |
8516 | void cCVarSet(Compiler* comp, VARSET_VALARG_TP vars) |
8517 | { |
8518 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8519 | printf("===================================================================== dCVarSet %u\n" , sequenceNumber++); |
8520 | dumpConvertedVarSet(comp, vars); |
8521 | printf("\n" ); // dumpConvertedVarSet() doesn't emit a trailing newline |
8522 | } |
8523 | |
8524 | void dBlock(BasicBlock* block) |
8525 | { |
8526 | cBlock(JitTls::GetCompiler(), block); |
8527 | } |
8528 | |
8529 | void dBlocks() |
8530 | { |
8531 | cBlocks(JitTls::GetCompiler()); |
8532 | } |
8533 | |
8534 | void dBlocksV() |
8535 | { |
8536 | cBlocksV(JitTls::GetCompiler()); |
8537 | } |
8538 | |
8539 | void dTree(GenTree* tree) |
8540 | { |
8541 | cTree(JitTls::GetCompiler(), tree); |
8542 | } |
8543 | |
8544 | void dTreeLIR(GenTree* tree) |
8545 | { |
8546 | cTreeLIR(JitTls::GetCompiler(), tree); |
8547 | } |
8548 | |
8549 | void dTrees() |
8550 | { |
8551 | cTrees(JitTls::GetCompiler()); |
8552 | } |
8553 | |
8554 | void dEH() |
8555 | { |
8556 | cEH(JitTls::GetCompiler()); |
8557 | } |
8558 | |
8559 | void dVar(unsigned lclNum) |
8560 | { |
8561 | cVar(JitTls::GetCompiler(), lclNum); |
8562 | } |
8563 | |
8564 | void dVarDsc(LclVarDsc* varDsc) |
8565 | { |
8566 | cVarDsc(JitTls::GetCompiler(), varDsc); |
8567 | } |
8568 | |
8569 | void dVars() |
8570 | { |
8571 | cVars(JitTls::GetCompiler()); |
8572 | } |
8573 | |
8574 | void dVarsFinal() |
8575 | { |
8576 | cVarsFinal(JitTls::GetCompiler()); |
8577 | } |
8578 | |
8579 | void dBlockPreds(BasicBlock* block) |
8580 | { |
8581 | cBlockPreds(JitTls::GetCompiler(), block); |
8582 | } |
8583 | |
8584 | void dBlockCheapPreds(BasicBlock* block) |
8585 | { |
8586 | cBlockCheapPreds(JitTls::GetCompiler(), block); |
8587 | } |
8588 | |
8589 | void dBlockSuccs(BasicBlock* block) |
8590 | { |
8591 | cBlockSuccs(JitTls::GetCompiler(), block); |
8592 | } |
8593 | |
8594 | void dReach() |
8595 | { |
8596 | cReach(JitTls::GetCompiler()); |
8597 | } |
8598 | |
8599 | void dDoms() |
8600 | { |
8601 | cDoms(JitTls::GetCompiler()); |
8602 | } |
8603 | |
8604 | void dLiveness() |
8605 | { |
8606 | cLiveness(JitTls::GetCompiler()); |
8607 | } |
8608 | |
8609 | void dCVarSet(VARSET_VALARG_TP vars) |
8610 | { |
8611 | cCVarSet(JitTls::GetCompiler(), vars); |
8612 | } |
8613 | |
8614 | void dRegMask(regMaskTP mask) |
8615 | { |
8616 | static unsigned sequenceNumber = 0; // separate calls with a number to indicate this function has been called |
8617 | printf("===================================================================== dRegMask %u\n" , sequenceNumber++); |
8618 | dspRegMask(mask); |
8619 | printf("\n" ); // dspRegMask() doesn't emit a trailing newline |
8620 | } |
8621 | |
8622 | void dBlockList(BasicBlockList* list) |
8623 | { |
8624 | printf("WorkList: " ); |
8625 | while (list != nullptr) |
8626 | { |
8627 | printf(FMT_BB " " , list->block->bbNum); |
8628 | list = list->next; |
8629 | } |
8630 | printf("\n" ); |
8631 | } |
8632 | |
8633 | // Global variables available in debug mode. That are set by debug APIs for finding |
8634 | // Trees, Stmts, and/or Blocks using id or bbNum. |
8635 | // That can be used in watch window or as a way to get address of fields for data break points. |
8636 | |
8637 | GenTree* dbTree; |
8638 | GenTreeStmt* dbStmt; |
8639 | BasicBlock* dbTreeBlock; |
8640 | BasicBlock* dbBlock; |
8641 | |
8642 | // Debug APIs for finding Trees, Stmts, and/or Blocks. |
8643 | // As a side effect, they set the debug variables above. |
8644 | |
8645 | GenTree* dFindTree(GenTree* tree, unsigned id) |
8646 | { |
8647 | GenTree* child; |
8648 | |
8649 | if (tree == nullptr) |
8650 | { |
8651 | return nullptr; |
8652 | } |
8653 | |
8654 | if (tree->gtTreeID == id) |
8655 | { |
8656 | dbTree = tree; |
8657 | return tree; |
8658 | } |
8659 | |
8660 | unsigned childCount = tree->NumChildren(); |
8661 | for (unsigned childIndex = 0; childIndex < childCount; childIndex++) |
8662 | { |
8663 | child = tree->GetChild(childIndex); |
8664 | child = dFindTree(child, id); |
8665 | if (child != nullptr) |
8666 | { |
8667 | return child; |
8668 | } |
8669 | } |
8670 | |
8671 | return nullptr; |
8672 | } |
8673 | |
8674 | GenTree* dFindTree(unsigned id) |
8675 | { |
8676 | Compiler* comp = JitTls::GetCompiler(); |
8677 | BasicBlock* block; |
8678 | GenTree* tree; |
8679 | |
8680 | dbTreeBlock = nullptr; |
8681 | dbTree = nullptr; |
8682 | |
8683 | for (block = comp->fgFirstBB; block != nullptr; block = block->bbNext) |
8684 | { |
8685 | for (GenTreeStmt* stmt = block->firstStmt(); stmt; stmt = stmt->gtNextStmt) |
8686 | { |
8687 | tree = dFindTree(stmt, id); |
8688 | if (tree != nullptr) |
8689 | { |
8690 | dbTreeBlock = block; |
8691 | return tree; |
8692 | } |
8693 | } |
8694 | } |
8695 | |
8696 | return nullptr; |
8697 | } |
8698 | |
8699 | GenTreeStmt* dFindStmt(unsigned id) |
8700 | { |
8701 | Compiler* comp = JitTls::GetCompiler(); |
8702 | BasicBlock* block; |
8703 | |
8704 | dbStmt = nullptr; |
8705 | |
8706 | unsigned stmtId = 0; |
8707 | for (block = comp->fgFirstBB; block != nullptr; block = block->bbNext) |
8708 | { |
8709 | for (GenTreeStmt* stmt = block->firstStmt(); stmt; stmt = stmt->gtNextStmt) |
8710 | { |
8711 | stmtId++; |
8712 | if (stmtId == id) |
8713 | { |
8714 | dbStmt = stmt; |
8715 | return stmt; |
8716 | } |
8717 | } |
8718 | } |
8719 | |
8720 | return nullptr; |
8721 | } |
8722 | |
8723 | BasicBlock* dFindBlock(unsigned bbNum) |
8724 | { |
8725 | Compiler* comp = JitTls::GetCompiler(); |
8726 | BasicBlock* block = nullptr; |
8727 | |
8728 | dbBlock = nullptr; |
8729 | for (block = comp->fgFirstBB; block != nullptr; block = block->bbNext) |
8730 | { |
8731 | if (block->bbNum == bbNum) |
8732 | { |
8733 | dbBlock = block; |
8734 | break; |
8735 | } |
8736 | } |
8737 | |
8738 | return block; |
8739 | } |
8740 | |
8741 | /***************************************************************************** |
8742 | * |
8743 | * COMPlus_JitDumpIR support - dump out function in linear IR form |
8744 | */ |
8745 | |
8746 | void cFuncIR(Compiler* comp) |
8747 | { |
8748 | BasicBlock* block; |
8749 | |
8750 | printf("Method %s::%s, hsh=0x%x\n" , comp->info.compClassName, comp->info.compMethodName, |
8751 | comp->info.compMethodHash()); |
8752 | |
8753 | printf("\n" ); |
8754 | |
8755 | for (block = comp->fgFirstBB; block != nullptr; block = block->bbNext) |
8756 | { |
8757 | cBlockIR(comp, block); |
8758 | } |
8759 | } |
8760 | |
8761 | /***************************************************************************** |
8762 | * |
8763 | * COMPlus_JitDumpIR support - dump out the format specifiers from COMPlus_JitDumpIRFormat |
8764 | */ |
8765 | |
8766 | void dFormatIR() |
8767 | { |
8768 | Compiler* comp = JitTls::GetCompiler(); |
8769 | |
8770 | if (comp->dumpIRFormat != nullptr) |
8771 | { |
8772 | printf("COMPlus_JitDumpIRFormat=%ls" , comp->dumpIRFormat); |
8773 | } |
8774 | } |
8775 | |
8776 | /***************************************************************************** |
8777 | * |
8778 | * COMPlus_JitDumpIR support - dump out function in linear IR form |
8779 | */ |
8780 | |
8781 | void dFuncIR() |
8782 | { |
8783 | cFuncIR(JitTls::GetCompiler()); |
8784 | } |
8785 | |
8786 | /***************************************************************************** |
8787 | * |
8788 | * COMPlus_JitDumpIR support - dump out loop in linear IR form |
8789 | */ |
8790 | |
8791 | void cLoopIR(Compiler* comp, Compiler::LoopDsc* loop) |
8792 | { |
8793 | BasicBlock* blockHead = loop->lpHead; |
8794 | BasicBlock* blockFirst = loop->lpFirst; |
8795 | BasicBlock* blockTop = loop->lpTop; |
8796 | BasicBlock* blockEntry = loop->lpEntry; |
8797 | BasicBlock* blockBottom = loop->lpBottom; |
8798 | BasicBlock* blockExit = loop->lpExit; |
8799 | BasicBlock* blockLast = blockBottom->bbNext; |
8800 | BasicBlock* block; |
8801 | |
8802 | printf("LOOP\n" ); |
8803 | printf("\n" ); |
8804 | printf("HEAD " FMT_BB "\n" , blockHead->bbNum); |
8805 | printf("FIRST " FMT_BB "\n" , blockFirst->bbNum); |
8806 | printf("TOP " FMT_BB "\n" , blockTop->bbNum); |
8807 | printf("ENTRY " FMT_BB "\n" , blockEntry->bbNum); |
8808 | if (loop->lpExitCnt == 1) |
8809 | { |
8810 | printf("EXIT " FMT_BB "\n" , blockExit->bbNum); |
8811 | } |
8812 | else |
8813 | { |
8814 | printf("EXITS %u" , loop->lpExitCnt); |
8815 | } |
8816 | printf("BOTTOM " FMT_BB "\n" , blockBottom->bbNum); |
8817 | printf("\n" ); |
8818 | |
8819 | cBlockIR(comp, blockHead); |
8820 | for (block = blockFirst; ((block != nullptr) && (block != blockLast)); block = block->bbNext) |
8821 | { |
8822 | cBlockIR(comp, block); |
8823 | } |
8824 | } |
8825 | |
8826 | /***************************************************************************** |
8827 | * |
8828 | * COMPlus_JitDumpIR support - dump out loop in linear IR form |
8829 | */ |
8830 | |
8831 | void dLoopIR(Compiler::LoopDsc* loop) |
8832 | { |
8833 | cLoopIR(JitTls::GetCompiler(), loop); |
8834 | } |
8835 | |
8836 | /***************************************************************************** |
8837 | * |
8838 | * COMPlus_JitDumpIR support - dump out loop (given loop number) in linear IR form |
8839 | */ |
8840 | |
8841 | void dLoopNumIR(unsigned loopNum) |
8842 | { |
8843 | Compiler* comp = JitTls::GetCompiler(); |
8844 | |
8845 | if (loopNum >= comp->optLoopCount) |
8846 | { |
8847 | printf("loopNum %u out of range\n" ); |
8848 | return; |
8849 | } |
8850 | |
8851 | Compiler::LoopDsc* loop = &comp->optLoopTable[loopNum]; |
8852 | cLoopIR(JitTls::GetCompiler(), loop); |
8853 | } |
8854 | |
8855 | /***************************************************************************** |
8856 | * |
8857 | * COMPlus_JitDumpIR support - dump spaces to specified tab stop |
8858 | */ |
8859 | |
8860 | int dTabStopIR(int curr, int tabstop) |
8861 | { |
8862 | int chars = 0; |
8863 | |
8864 | if (tabstop <= curr) |
8865 | { |
8866 | chars += printf(" " ); |
8867 | } |
8868 | |
8869 | for (int i = curr; i < tabstop; i++) |
8870 | { |
8871 | chars += printf(" " ); |
8872 | } |
8873 | |
8874 | return chars; |
8875 | } |
8876 | |
8877 | void cNodeIR(Compiler* comp, GenTree* tree); |
8878 | |
8879 | /***************************************************************************** |
8880 | * |
8881 | * COMPlus_JitDumpIR support - dump out block in linear IR form |
8882 | */ |
8883 | |
8884 | void cBlockIR(Compiler* comp, BasicBlock* block) |
8885 | { |
8886 | bool noStmts = comp->dumpIRNoStmts; |
8887 | bool trees = comp->dumpIRTrees; |
8888 | |
8889 | if (comp->dumpIRBlockHeaders) |
8890 | { |
8891 | block->dspBlockHeader(comp); |
8892 | } |
8893 | else |
8894 | { |
8895 | printf(FMT_BB ":\n" , block->bbNum); |
8896 | } |
8897 | |
8898 | printf("\n" ); |
8899 | |
8900 | if (!block->IsLIR()) |
8901 | { |
8902 | for (GenTreeStmt* stmt = block->firstStmt(); stmt; stmt = stmt->gtNextStmt) |
8903 | { |
8904 | // Print current stmt. |
8905 | |
8906 | if (trees) |
8907 | { |
8908 | cTree(comp, stmt); |
8909 | printf("\n" ); |
8910 | printf("=====================================================================\n" ); |
8911 | } |
8912 | |
8913 | if (comp->compRationalIRForm) |
8914 | { |
8915 | GenTree* tree; |
8916 | |
8917 | foreach_treenode_execution_order(tree, stmt) |
8918 | { |
8919 | cNodeIR(comp, tree); |
8920 | } |
8921 | } |
8922 | else |
8923 | { |
8924 | cTreeIR(comp, stmt); |
8925 | } |
8926 | |
8927 | if (!noStmts && !trees) |
8928 | { |
8929 | printf("\n" ); |
8930 | } |
8931 | } |
8932 | } |
8933 | else |
8934 | { |
8935 | for (GenTree* node = block->bbTreeList; node != nullptr; node = node->gtNext) |
8936 | { |
8937 | cNodeIR(comp, node); |
8938 | } |
8939 | } |
8940 | |
8941 | int chars = 0; |
8942 | |
8943 | chars += dTabStopIR(chars, COLUMN_OPCODE); |
8944 | |
8945 | chars += printf(" " ); |
8946 | switch (block->bbJumpKind) |
8947 | { |
8948 | case BBJ_EHFINALLYRET: |
8949 | chars += printf("BRANCH(EHFINALLYRET)" ); |
8950 | break; |
8951 | |
8952 | case BBJ_EHFILTERRET: |
8953 | chars += printf("BRANCH(EHFILTERRET)" ); |
8954 | break; |
8955 | |
8956 | case BBJ_EHCATCHRET: |
8957 | chars += printf("BRANCH(EHCATCHRETURN)" ); |
8958 | chars += dTabStopIR(chars, COLUMN_OPERANDS); |
8959 | chars += printf(" " FMT_BB, block->bbJumpDest->bbNum); |
8960 | break; |
8961 | |
8962 | case BBJ_THROW: |
8963 | chars += printf("BRANCH(THROW)" ); |
8964 | break; |
8965 | |
8966 | case BBJ_RETURN: |
8967 | chars += printf("BRANCH(RETURN)" ); |
8968 | break; |
8969 | |
8970 | case BBJ_NONE: |
8971 | // For fall-through blocks |
8972 | chars += printf("BRANCH(NONE)" ); |
8973 | break; |
8974 | |
8975 | case BBJ_ALWAYS: |
8976 | chars += printf("BRANCH(ALWAYS)" ); |
8977 | chars += dTabStopIR(chars, COLUMN_OPERANDS); |
8978 | chars += printf(" " FMT_BB, block->bbJumpDest->bbNum); |
8979 | if (block->bbFlags & BBF_KEEP_BBJ_ALWAYS) |
8980 | { |
8981 | chars += dTabStopIR(chars, COLUMN_KINDS); |
8982 | chars += printf("; [KEEP_BBJ_ALWAYS]" ); |
8983 | } |
8984 | break; |
8985 | |
8986 | case BBJ_LEAVE: |
8987 | chars += printf("BRANCH(LEAVE)" ); |
8988 | chars += dTabStopIR(chars, COLUMN_OPERANDS); |
8989 | chars += printf(" " FMT_BB, block->bbJumpDest->bbNum); |
8990 | break; |
8991 | |
8992 | case BBJ_CALLFINALLY: |
8993 | chars += printf("BRANCH(CALLFINALLY)" ); |
8994 | chars += dTabStopIR(chars, COLUMN_OPERANDS); |
8995 | chars += printf(" " FMT_BB, block->bbJumpDest->bbNum); |
8996 | break; |
8997 | |
8998 | case BBJ_COND: |
8999 | chars += printf("BRANCH(COND)" ); |
9000 | chars += dTabStopIR(chars, COLUMN_OPERANDS); |
9001 | chars += printf(" " FMT_BB, block->bbJumpDest->bbNum); |
9002 | break; |
9003 | |
9004 | case BBJ_SWITCH: |
9005 | chars += printf("BRANCH(SWITCH)" ); |
9006 | chars += dTabStopIR(chars, COLUMN_OPERANDS); |
9007 | |
9008 | unsigned jumpCnt; |
9009 | jumpCnt = block->bbJumpSwt->bbsCount; |
9010 | BasicBlock** jumpTab; |
9011 | jumpTab = block->bbJumpSwt->bbsDstTab; |
9012 | do |
9013 | { |
9014 | chars += printf("%c " FMT_BB, (jumpTab == block->bbJumpSwt->bbsDstTab) ? ' ' : ',', (*jumpTab)->bbNum); |
9015 | } while (++jumpTab, --jumpCnt); |
9016 | break; |
9017 | |
9018 | default: |
9019 | unreached(); |
9020 | break; |
9021 | } |
9022 | |
9023 | printf("\n" ); |
9024 | if (block->bbNext != nullptr) |
9025 | { |
9026 | printf("\n" ); |
9027 | } |
9028 | } |
9029 | |
9030 | /***************************************************************************** |
9031 | * |
9032 | * COMPlus_JitDumpIR support - dump out block in linear IR form |
9033 | */ |
9034 | |
9035 | void dBlockIR(BasicBlock* block) |
9036 | { |
9037 | cBlockIR(JitTls::GetCompiler(), block); |
9038 | } |
9039 | |
9040 | /***************************************************************************** |
9041 | * |
9042 | * COMPlus_JitDumpIR support - dump out tree node type for linear IR form |
9043 | */ |
9044 | |
9045 | int cTreeTypeIR(Compiler* comp, GenTree* tree) |
9046 | { |
9047 | int chars = 0; |
9048 | |
9049 | var_types type = tree->TypeGet(); |
9050 | |
9051 | const char* typeName = varTypeName(type); |
9052 | chars += printf(".%s" , typeName); |
9053 | |
9054 | return chars; |
9055 | } |
9056 | |
9057 | /***************************************************************************** |
9058 | * |
9059 | * COMPlus_JitDumpIR support - dump out tree node type for linear IR form |
9060 | */ |
9061 | |
9062 | int dTreeTypeIR(GenTree* tree) |
9063 | { |
9064 | int chars = cTreeTypeIR(JitTls::GetCompiler(), tree); |
9065 | |
9066 | return chars; |
9067 | } |
9068 | |
9069 | /***************************************************************************** |
9070 | * |
9071 | * COMPlus_JitDumpIR support - dump out tree node kind for linear IR form |
9072 | */ |
9073 | |
9074 | int cTreeKindsIR(Compiler* comp, GenTree* tree) |
9075 | { |
9076 | int chars = 0; |
9077 | |
9078 | unsigned kind = tree->OperKind(); |
9079 | |
9080 | chars += printf("kinds=" ); |
9081 | if (kind == GTK_SPECIAL) |
9082 | { |
9083 | chars += printf("[SPECIAL]" ); |
9084 | } |
9085 | if (kind & GTK_CONST) |
9086 | { |
9087 | chars += printf("[CONST]" ); |
9088 | } |
9089 | if (kind & GTK_LEAF) |
9090 | { |
9091 | chars += printf("[LEAF]" ); |
9092 | } |
9093 | if (kind & GTK_UNOP) |
9094 | { |
9095 | chars += printf("[UNOP]" ); |
9096 | } |
9097 | if (kind & GTK_BINOP) |
9098 | { |
9099 | chars += printf("[BINOP]" ); |
9100 | } |
9101 | if (kind & GTK_LOGOP) |
9102 | { |
9103 | chars += printf("[LOGOP]" ); |
9104 | } |
9105 | if (kind & GTK_COMMUTE) |
9106 | { |
9107 | chars += printf("[COMMUTE]" ); |
9108 | } |
9109 | if (kind & GTK_EXOP) |
9110 | { |
9111 | chars += printf("[EXOP]" ); |
9112 | } |
9113 | if (kind & GTK_LOCAL) |
9114 | { |
9115 | chars += printf("[LOCAL]" ); |
9116 | } |
9117 | if (kind & GTK_SMPOP) |
9118 | { |
9119 | chars += printf("[SMPOP]" ); |
9120 | } |
9121 | |
9122 | return chars; |
9123 | } |
9124 | |
9125 | /***************************************************************************** |
9126 | * |
9127 | * COMPlus_JitDumpIR support - dump out tree node kind for linear IR form |
9128 | */ |
9129 | |
9130 | int dTreeKindsIR(GenTree* tree) |
9131 | { |
9132 | int chars = cTreeKindsIR(JitTls::GetCompiler(), tree); |
9133 | |
9134 | return chars; |
9135 | } |
9136 | |
9137 | /***************************************************************************** |
9138 | * |
9139 | * COMPlus_JitDumpIR support - dump out tree node flags for linear IR form |
9140 | */ |
9141 | |
9142 | int cTreeFlagsIR(Compiler* comp, GenTree* tree) |
9143 | { |
9144 | int chars = 0; |
9145 | |
9146 | if (tree->gtFlags != 0) |
9147 | { |
9148 | chars += printf("flags=" ); |
9149 | |
9150 | // Node flags |
9151 | CLANG_FORMAT_COMMENT_ANCHOR; |
9152 | |
9153 | #if defined(DEBUG) |
9154 | #if SMALL_TREE_NODES |
9155 | if (comp->dumpIRNodes) |
9156 | { |
9157 | if (tree->gtDebugFlags & GTF_DEBUG_NODE_LARGE) |
9158 | { |
9159 | chars += printf("[NODE_LARGE]" ); |
9160 | } |
9161 | if (tree->gtDebugFlags & GTF_DEBUG_NODE_SMALL) |
9162 | { |
9163 | chars += printf("[NODE_SMALL]" ); |
9164 | } |
9165 | } |
9166 | #endif // SMALL_TREE_NODES |
9167 | if (tree->gtDebugFlags & GTF_DEBUG_NODE_MORPHED) |
9168 | { |
9169 | chars += printf("[MORPHED]" ); |
9170 | } |
9171 | #endif // defined(DEBUG) |
9172 | |
9173 | if (tree->gtFlags & GTF_COLON_COND) |
9174 | { |
9175 | chars += printf("[COLON_COND]" ); |
9176 | } |
9177 | |
9178 | // Operator flags |
9179 | |
9180 | genTreeOps op = tree->OperGet(); |
9181 | switch (op) |
9182 | { |
9183 | |
9184 | case GT_LCL_VAR: |
9185 | case GT_LCL_VAR_ADDR: |
9186 | case GT_LCL_FLD: |
9187 | case GT_LCL_FLD_ADDR: |
9188 | case GT_STORE_LCL_FLD: |
9189 | case GT_STORE_LCL_VAR: |
9190 | if (tree->gtFlags & GTF_VAR_DEF) |
9191 | { |
9192 | chars += printf("[VAR_DEF]" ); |
9193 | } |
9194 | if (tree->gtFlags & GTF_VAR_USEASG) |
9195 | { |
9196 | chars += printf("[VAR_USEASG]" ); |
9197 | } |
9198 | if (tree->gtFlags & GTF_VAR_CAST) |
9199 | { |
9200 | chars += printf("[VAR_CAST]" ); |
9201 | } |
9202 | if (tree->gtFlags & GTF_VAR_ITERATOR) |
9203 | { |
9204 | chars += printf("[VAR_ITERATOR]" ); |
9205 | } |
9206 | if (tree->gtFlags & GTF_VAR_CLONED) |
9207 | { |
9208 | chars += printf("[VAR_CLONED]" ); |
9209 | } |
9210 | if (tree->gtFlags & GTF_VAR_DEATH) |
9211 | { |
9212 | chars += printf("[VAR_DEATH]" ); |
9213 | } |
9214 | if (tree->gtFlags & GTF_VAR_ARR_INDEX) |
9215 | { |
9216 | chars += printf("[VAR_ARR_INDEX]" ); |
9217 | } |
9218 | #if defined(DEBUG) |
9219 | if (tree->gtDebugFlags & GTF_DEBUG_VAR_CSE_REF) |
9220 | { |
9221 | chars += printf("[VAR_CSE_REF]" ); |
9222 | } |
9223 | #endif |
9224 | break; |
9225 | |
9226 | case GT_NOP: |
9227 | |
9228 | if (tree->gtFlags & GTF_NOP_DEATH) |
9229 | { |
9230 | chars += printf("[NOP_DEATH]" ); |
9231 | } |
9232 | break; |
9233 | |
9234 | case GT_NO_OP: |
9235 | break; |
9236 | |
9237 | case GT_FIELD: |
9238 | if (tree->gtFlags & GTF_FLD_VOLATILE) |
9239 | { |
9240 | chars += printf("[FLD_VOLATILE]" ); |
9241 | } |
9242 | break; |
9243 | |
9244 | case GT_INDEX: |
9245 | |
9246 | if (tree->gtFlags & GTF_INX_REFARR_LAYOUT) |
9247 | { |
9248 | chars += printf("[INX_REFARR_LAYOUT]" ); |
9249 | } |
9250 | if (tree->gtFlags & GTF_INX_STRING_LAYOUT) |
9251 | { |
9252 | chars += printf("[INX_STRING_LAYOUT]" ); |
9253 | } |
9254 | __fallthrough; |
9255 | case GT_INDEX_ADDR: |
9256 | if (tree->gtFlags & GTF_INX_RNGCHK) |
9257 | { |
9258 | chars += printf("[INX_RNGCHK]" ); |
9259 | } |
9260 | break; |
9261 | |
9262 | case GT_IND: |
9263 | case GT_STOREIND: |
9264 | |
9265 | if (tree->gtFlags & GTF_IND_VOLATILE) |
9266 | { |
9267 | chars += printf("[IND_VOLATILE]" ); |
9268 | } |
9269 | if (tree->gtFlags & GTF_IND_TGTANYWHERE) |
9270 | { |
9271 | chars += printf("[IND_TGTANYWHERE]" ); |
9272 | } |
9273 | if (tree->gtFlags & GTF_IND_TLS_REF) |
9274 | { |
9275 | chars += printf("[IND_TLS_REF]" ); |
9276 | } |
9277 | if (tree->gtFlags & GTF_IND_ASG_LHS) |
9278 | { |
9279 | chars += printf("[IND_ASG_LHS]" ); |
9280 | } |
9281 | if (tree->gtFlags & GTF_IND_UNALIGNED) |
9282 | { |
9283 | chars += printf("[IND_UNALIGNED]" ); |
9284 | } |
9285 | if (tree->gtFlags & GTF_IND_INVARIANT) |
9286 | { |
9287 | chars += printf("[IND_INVARIANT]" ); |
9288 | } |
9289 | break; |
9290 | |
9291 | case GT_CLS_VAR: |
9292 | |
9293 | if (tree->gtFlags & GTF_CLS_VAR_ASG_LHS) |
9294 | { |
9295 | chars += printf("[CLS_VAR_ASG_LHS]" ); |
9296 | } |
9297 | break; |
9298 | |
9299 | case GT_ADDR: |
9300 | |
9301 | if (tree->gtFlags & GTF_ADDR_ONSTACK) |
9302 | { |
9303 | chars += printf("[ADDR_ONSTACK]" ); |
9304 | } |
9305 | break; |
9306 | |
9307 | case GT_MUL: |
9308 | #if !defined(_TARGET_64BIT_) |
9309 | case GT_MUL_LONG: |
9310 | #endif |
9311 | |
9312 | if (tree->gtFlags & GTF_MUL_64RSLT) |
9313 | { |
9314 | chars += printf("[64RSLT]" ); |
9315 | } |
9316 | if (tree->gtFlags & GTF_ADDRMODE_NO_CSE) |
9317 | { |
9318 | chars += printf("[ADDRMODE_NO_CSE]" ); |
9319 | } |
9320 | break; |
9321 | |
9322 | case GT_ADD: |
9323 | |
9324 | if (tree->gtFlags & GTF_ADDRMODE_NO_CSE) |
9325 | { |
9326 | chars += printf("[ADDRMODE_NO_CSE]" ); |
9327 | } |
9328 | break; |
9329 | |
9330 | case GT_LSH: |
9331 | |
9332 | if (tree->gtFlags & GTF_ADDRMODE_NO_CSE) |
9333 | { |
9334 | chars += printf("[ADDRMODE_NO_CSE]" ); |
9335 | } |
9336 | break; |
9337 | |
9338 | case GT_MOD: |
9339 | case GT_UMOD: |
9340 | break; |
9341 | |
9342 | case GT_EQ: |
9343 | case GT_NE: |
9344 | case GT_LT: |
9345 | case GT_LE: |
9346 | case GT_GT: |
9347 | case GT_GE: |
9348 | |
9349 | if (tree->gtFlags & GTF_RELOP_NAN_UN) |
9350 | { |
9351 | chars += printf("[RELOP_NAN_UN]" ); |
9352 | } |
9353 | if (tree->gtFlags & GTF_RELOP_JMP_USED) |
9354 | { |
9355 | chars += printf("[RELOP_JMP_USED]" ); |
9356 | } |
9357 | if (tree->gtFlags & GTF_RELOP_QMARK) |
9358 | { |
9359 | chars += printf("[RELOP_QMARK]" ); |
9360 | } |
9361 | break; |
9362 | |
9363 | case GT_QMARK: |
9364 | |
9365 | if (tree->gtFlags & GTF_QMARK_CAST_INSTOF) |
9366 | { |
9367 | chars += printf("[QMARK_CAST_INSTOF]" ); |
9368 | } |
9369 | break; |
9370 | |
9371 | case GT_BOX: |
9372 | |
9373 | if (tree->gtFlags & GTF_BOX_VALUE) |
9374 | { |
9375 | chars += printf("[BOX_VALUE]" ); |
9376 | } |
9377 | break; |
9378 | |
9379 | case GT_CNS_INT: |
9380 | |
9381 | { |
9382 | unsigned handleKind = (tree->gtFlags & GTF_ICON_HDL_MASK); |
9383 | |
9384 | switch (handleKind) |
9385 | { |
9386 | |
9387 | case GTF_ICON_SCOPE_HDL: |
9388 | |
9389 | chars += printf("[ICON_SCOPE_HDL]" ); |
9390 | break; |
9391 | |
9392 | case GTF_ICON_CLASS_HDL: |
9393 | |
9394 | chars += printf("[ICON_CLASS_HDL]" ); |
9395 | break; |
9396 | |
9397 | case GTF_ICON_METHOD_HDL: |
9398 | |
9399 | chars += printf("[ICON_METHOD_HDL]" ); |
9400 | break; |
9401 | |
9402 | case GTF_ICON_FIELD_HDL: |
9403 | |
9404 | chars += printf("[ICON_FIELD_HDL]" ); |
9405 | break; |
9406 | |
9407 | case GTF_ICON_STATIC_HDL: |
9408 | |
9409 | chars += printf("[ICON_STATIC_HDL]" ); |
9410 | break; |
9411 | |
9412 | case GTF_ICON_STR_HDL: |
9413 | |
9414 | chars += printf("[ICON_STR_HDL]" ); |
9415 | break; |
9416 | |
9417 | case GTF_ICON_PSTR_HDL: |
9418 | |
9419 | chars += printf("[ICON_PSTR_HDL]" ); |
9420 | break; |
9421 | |
9422 | case GTF_ICON_PTR_HDL: |
9423 | |
9424 | chars += printf("[ICON_PTR_HDL]" ); |
9425 | break; |
9426 | |
9427 | case GTF_ICON_VARG_HDL: |
9428 | |
9429 | chars += printf("[ICON_VARG_HDL]" ); |
9430 | break; |
9431 | |
9432 | case GTF_ICON_PINVKI_HDL: |
9433 | |
9434 | chars += printf("[ICON_PINVKI_HDL]" ); |
9435 | break; |
9436 | |
9437 | case GTF_ICON_TOKEN_HDL: |
9438 | |
9439 | chars += printf("[ICON_TOKEN_HDL]" ); |
9440 | break; |
9441 | |
9442 | case GTF_ICON_TLS_HDL: |
9443 | |
9444 | chars += printf("[ICON_TLD_HDL]" ); |
9445 | break; |
9446 | |
9447 | case GTF_ICON_FTN_ADDR: |
9448 | |
9449 | chars += printf("[ICON_FTN_ADDR]" ); |
9450 | break; |
9451 | |
9452 | case GTF_ICON_CIDMID_HDL: |
9453 | |
9454 | chars += printf("[ICON_CIDMID_HDL]" ); |
9455 | break; |
9456 | |
9457 | case GTF_ICON_BBC_PTR: |
9458 | |
9459 | chars += printf("[ICON_BBC_PTR]" ); |
9460 | break; |
9461 | |
9462 | case GTF_ICON_FIELD_OFF: |
9463 | |
9464 | chars += printf("[ICON_FIELD_OFF]" ); |
9465 | break; |
9466 | } |
9467 | } |
9468 | break; |
9469 | |
9470 | case GT_OBJ: |
9471 | case GT_STORE_OBJ: |
9472 | if (tree->AsObj()->HasGCPtr()) |
9473 | { |
9474 | chars += printf("[BLK_HASGCPTR]" ); |
9475 | } |
9476 | __fallthrough; |
9477 | |
9478 | case GT_BLK: |
9479 | case GT_DYN_BLK: |
9480 | case GT_STORE_BLK: |
9481 | case GT_STORE_DYN_BLK: |
9482 | |
9483 | if (tree->gtFlags & GTF_BLK_VOLATILE) |
9484 | { |
9485 | chars += printf("[BLK_VOLATILE]" ); |
9486 | } |
9487 | if (tree->AsBlk()->IsUnaligned()) |
9488 | { |
9489 | chars += printf("[BLK_UNALIGNED]" ); |
9490 | } |
9491 | break; |
9492 | |
9493 | case GT_CALL: |
9494 | |
9495 | if (tree->gtFlags & GTF_CALL_UNMANAGED) |
9496 | { |
9497 | chars += printf("[CALL_UNMANAGED]" ); |
9498 | } |
9499 | if (tree->gtFlags & GTF_CALL_INLINE_CANDIDATE) |
9500 | { |
9501 | chars += printf("[CALL_INLINE_CANDIDATE]" ); |
9502 | } |
9503 | if (!tree->AsCall()->IsVirtual()) |
9504 | { |
9505 | chars += printf("[CALL_NONVIRT]" ); |
9506 | } |
9507 | if (tree->AsCall()->IsVirtualVtable()) |
9508 | { |
9509 | chars += printf("[CALL_VIRT_VTABLE]" ); |
9510 | } |
9511 | if (tree->AsCall()->IsVirtualStub()) |
9512 | { |
9513 | chars += printf("[CALL_VIRT_STUB]" ); |
9514 | } |
9515 | if (tree->gtFlags & GTF_CALL_NULLCHECK) |
9516 | { |
9517 | chars += printf("[CALL_NULLCHECK]" ); |
9518 | } |
9519 | if (tree->gtFlags & GTF_CALL_POP_ARGS) |
9520 | { |
9521 | chars += printf("[CALL_POP_ARGS]" ); |
9522 | } |
9523 | if (tree->gtFlags & GTF_CALL_HOISTABLE) |
9524 | { |
9525 | chars += printf("[CALL_HOISTABLE]" ); |
9526 | } |
9527 | |
9528 | // More flags associated with calls. |
9529 | |
9530 | { |
9531 | GenTreeCall* call = tree->AsCall(); |
9532 | |
9533 | if (call->gtCallMoreFlags & GTF_CALL_M_EXPLICIT_TAILCALL) |
9534 | { |
9535 | chars += printf("[CALL_M_EXPLICIT_TAILCALL]" ); |
9536 | } |
9537 | if (call->gtCallMoreFlags & GTF_CALL_M_TAILCALL) |
9538 | { |
9539 | chars += printf("[CALL_M_TAILCALL]" ); |
9540 | } |
9541 | if (call->gtCallMoreFlags & GTF_CALL_M_VARARGS) |
9542 | { |
9543 | chars += printf("[CALL_M_VARARGS]" ); |
9544 | } |
9545 | if (call->gtCallMoreFlags & GTF_CALL_M_RETBUFFARG) |
9546 | { |
9547 | chars += printf("[CALL_M_RETBUFFARG]" ); |
9548 | } |
9549 | if (call->gtCallMoreFlags & GTF_CALL_M_DELEGATE_INV) |
9550 | { |
9551 | chars += printf("[CALL_M_DELEGATE_INV]" ); |
9552 | } |
9553 | if (call->gtCallMoreFlags & GTF_CALL_M_NOGCCHECK) |
9554 | { |
9555 | chars += printf("[CALL_M_NOGCCHECK]" ); |
9556 | } |
9557 | if (call->gtCallMoreFlags & GTF_CALL_M_SPECIAL_INTRINSIC) |
9558 | { |
9559 | chars += printf("[CALL_M_SPECIAL_INTRINSIC]" ); |
9560 | } |
9561 | |
9562 | if (call->IsUnmanaged()) |
9563 | { |
9564 | if (call->gtCallMoreFlags & GTF_CALL_M_UNMGD_THISCALL) |
9565 | { |
9566 | chars += printf("[CALL_M_UNMGD_THISCALL]" ); |
9567 | } |
9568 | } |
9569 | else if (call->IsVirtualStub()) |
9570 | { |
9571 | if (call->gtCallMoreFlags & GTF_CALL_M_VIRTSTUB_REL_INDIRECT) |
9572 | { |
9573 | chars += printf("[CALL_M_VIRTSTUB_REL_INDIRECT]" ); |
9574 | } |
9575 | } |
9576 | else if (!call->IsVirtual()) |
9577 | { |
9578 | if (call->gtCallMoreFlags & GTF_CALL_M_NONVIRT_SAME_THIS) |
9579 | { |
9580 | chars += printf("[CALL_M_NONVIRT_SAME_THIS]" ); |
9581 | } |
9582 | } |
9583 | |
9584 | if (call->gtCallMoreFlags & GTF_CALL_M_FRAME_VAR_DEATH) |
9585 | { |
9586 | chars += printf("[CALL_M_FRAME_VAR_DEATH]" ); |
9587 | } |
9588 | if (call->gtCallMoreFlags & GTF_CALL_M_TAILCALL_VIA_HELPER) |
9589 | { |
9590 | chars += printf("[CALL_M_TAILCALL_VIA_HELPER]" ); |
9591 | } |
9592 | #if FEATURE_TAILCALL_OPT |
9593 | if (call->gtCallMoreFlags & GTF_CALL_M_IMPLICIT_TAILCALL) |
9594 | { |
9595 | chars += printf("[CALL_M_IMPLICIT_TAILCALL]" ); |
9596 | } |
9597 | #endif |
9598 | if (call->gtCallMoreFlags & GTF_CALL_M_PINVOKE) |
9599 | { |
9600 | chars += printf("[CALL_M_PINVOKE]" ); |
9601 | } |
9602 | } |
9603 | break; |
9604 | |
9605 | case GT_STMT: |
9606 | |
9607 | if (tree->gtFlags & GTF_STMT_CMPADD) |
9608 | { |
9609 | chars += printf("[STMT_CMPADD]" ); |
9610 | } |
9611 | if (tree->gtFlags & GTF_STMT_HAS_CSE) |
9612 | { |
9613 | chars += printf("[STMT_HAS_CSE]" ); |
9614 | } |
9615 | break; |
9616 | |
9617 | default: |
9618 | |
9619 | { |
9620 | unsigned flags = (tree->gtFlags & (~(unsigned)(GTF_COMMON_MASK | GTF_OVERFLOW))); |
9621 | if (flags != 0) |
9622 | { |
9623 | chars += printf("[%08X]" , flags); |
9624 | } |
9625 | } |
9626 | break; |
9627 | } |
9628 | |
9629 | // Common flags. |
9630 | |
9631 | if (tree->gtFlags & GTF_ASG) |
9632 | { |
9633 | chars += printf("[ASG]" ); |
9634 | } |
9635 | if (tree->gtFlags & GTF_CALL) |
9636 | { |
9637 | chars += printf("[CALL]" ); |
9638 | } |
9639 | switch (op) |
9640 | { |
9641 | case GT_MUL: |
9642 | case GT_CAST: |
9643 | case GT_ADD: |
9644 | case GT_SUB: |
9645 | if (tree->gtFlags & GTF_OVERFLOW) |
9646 | { |
9647 | chars += printf("[OVERFLOW]" ); |
9648 | } |
9649 | break; |
9650 | default: |
9651 | break; |
9652 | } |
9653 | if (tree->gtFlags & GTF_EXCEPT) |
9654 | { |
9655 | chars += printf("[EXCEPT]" ); |
9656 | } |
9657 | if (tree->gtFlags & GTF_GLOB_REF) |
9658 | { |
9659 | chars += printf("[GLOB_REF]" ); |
9660 | } |
9661 | if (tree->gtFlags & GTF_ORDER_SIDEEFF) |
9662 | { |
9663 | chars += printf("[ORDER_SIDEEFF]" ); |
9664 | } |
9665 | if (tree->gtFlags & GTF_REVERSE_OPS) |
9666 | { |
9667 | if (op != GT_LCL_VAR) |
9668 | { |
9669 | chars += printf("[REVERSE_OPS]" ); |
9670 | } |
9671 | } |
9672 | if (tree->gtFlags & GTF_SPILLED) |
9673 | { |
9674 | chars += printf("[SPILLED_OPER]" ); |
9675 | } |
9676 | #if FEATURE_SET_FLAGS |
9677 | if (tree->gtFlags & GTF_SET_FLAGS) |
9678 | { |
9679 | if ((op != GT_IND) && (op != GT_STOREIND)) |
9680 | { |
9681 | chars += printf("[ZSF_SET_FLAGS]" ); |
9682 | } |
9683 | } |
9684 | #endif |
9685 | if (tree->gtFlags & GTF_IND_NONFAULTING) |
9686 | { |
9687 | if (tree->OperIsIndirOrArrLength()) |
9688 | { |
9689 | chars += printf("[IND_NONFAULTING]" ); |
9690 | } |
9691 | } |
9692 | if (tree->gtFlags & GTF_MAKE_CSE) |
9693 | { |
9694 | chars += printf("[MAKE_CSE]" ); |
9695 | } |
9696 | if (tree->gtFlags & GTF_DONT_CSE) |
9697 | { |
9698 | chars += printf("[DONT_CSE]" ); |
9699 | } |
9700 | if (tree->gtFlags & GTF_BOOLEAN) |
9701 | { |
9702 | chars += printf("[BOOLEAN]" ); |
9703 | } |
9704 | if (tree->gtFlags & GTF_UNSIGNED) |
9705 | { |
9706 | chars += printf("[SMALL_UNSIGNED]" ); |
9707 | } |
9708 | if (tree->gtFlags & GTF_LATE_ARG) |
9709 | { |
9710 | chars += printf("[SMALL_LATE_ARG]" ); |
9711 | } |
9712 | if (tree->gtFlags & GTF_SPILL) |
9713 | { |
9714 | chars += printf("[SPILL]" ); |
9715 | } |
9716 | if (tree->gtFlags & GTF_REUSE_REG_VAL) |
9717 | { |
9718 | if (op == GT_CNS_INT) |
9719 | { |
9720 | chars += printf("[REUSE_REG_VAL]" ); |
9721 | } |
9722 | } |
9723 | } |
9724 | |
9725 | return chars; |
9726 | } |
9727 | |
9728 | /***************************************************************************** |
9729 | * |
9730 | * COMPlus_JitDumpIR support - dump out tree node flags for linear IR form |
9731 | */ |
9732 | |
9733 | int dTreeFlagsIR(GenTree* tree) |
9734 | { |
9735 | int chars = cTreeFlagsIR(JitTls::GetCompiler(), tree); |
9736 | |
9737 | return chars; |
9738 | } |
9739 | |
9740 | /***************************************************************************** |
9741 | * |
9742 | * COMPlus_JitDumpIR support - dump out SSA number on tree node for linear IR form |
9743 | */ |
9744 | |
9745 | int cSsaNumIR(Compiler* comp, GenTree* tree) |
9746 | { |
9747 | int chars = 0; |
9748 | |
9749 | if (tree->gtLclVarCommon.HasSsaName()) |
9750 | { |
9751 | if (tree->gtFlags & GTF_VAR_USEASG) |
9752 | { |
9753 | assert(tree->gtFlags & GTF_VAR_DEF); |
9754 | chars += printf("<u:%d><d:%d>" , tree->gtLclVarCommon.gtSsaNum, comp->GetSsaNumForLocalVarDef(tree)); |
9755 | } |
9756 | else |
9757 | { |
9758 | chars += printf("<%s:%d>" , (tree->gtFlags & GTF_VAR_DEF) ? "d" : "u" , tree->gtLclVarCommon.gtSsaNum); |
9759 | } |
9760 | } |
9761 | |
9762 | return chars; |
9763 | } |
9764 | |
9765 | /***************************************************************************** |
9766 | * |
9767 | * COMPlus_JitDumpIR support - dump out SSA number on tree node for linear IR form |
9768 | */ |
9769 | |
9770 | int dSsaNumIR(GenTree* tree) |
9771 | { |
9772 | int chars = cSsaNumIR(JitTls::GetCompiler(), tree); |
9773 | |
9774 | return chars; |
9775 | } |
9776 | |
9777 | /***************************************************************************** |
9778 | * |
9779 | * COMPlus_JitDumpIR support - dump out Value Number on tree node for linear IR form |
9780 | */ |
9781 | |
9782 | int cValNumIR(Compiler* comp, GenTree* tree) |
9783 | { |
9784 | int chars = 0; |
9785 | |
9786 | if (tree->gtVNPair.GetLiberal() != ValueNumStore::NoVN) |
9787 | { |
9788 | assert(tree->gtVNPair.GetConservative() != ValueNumStore::NoVN); |
9789 | ValueNumPair vnp = tree->gtVNPair; |
9790 | ValueNum vn; |
9791 | if (vnp.BothEqual()) |
9792 | { |
9793 | chars += printf("<v:" ); |
9794 | vn = vnp.GetLiberal(); |
9795 | chars += printf(FMT_VN, vn); |
9796 | if (ValueNumStore::isReservedVN(vn)) |
9797 | { |
9798 | chars += printf("R" ); |
9799 | } |
9800 | chars += printf(">" ); |
9801 | } |
9802 | else |
9803 | { |
9804 | vn = vnp.GetLiberal(); |
9805 | chars += printf("<v:" ); |
9806 | chars += printf(FMT_VN, vn); |
9807 | if (ValueNumStore::isReservedVN(vn)) |
9808 | { |
9809 | chars += printf("R" ); |
9810 | } |
9811 | chars += printf("," ); |
9812 | vn = vnp.GetConservative(); |
9813 | chars += printf(FMT_VN, vn); |
9814 | if (ValueNumStore::isReservedVN(vn)) |
9815 | { |
9816 | chars += printf("R" ); |
9817 | } |
9818 | chars += printf(">" ); |
9819 | } |
9820 | } |
9821 | |
9822 | return chars; |
9823 | } |
9824 | |
9825 | /***************************************************************************** |
9826 | * |
9827 | * COMPlus_JitDumpIR support - dump out Value Number on tree node for linear IR form |
9828 | */ |
9829 | |
9830 | int dValNumIR(GenTree* tree) |
9831 | { |
9832 | int chars = cValNumIR(JitTls::GetCompiler(), tree); |
9833 | |
9834 | return chars; |
9835 | } |
9836 | |
9837 | /***************************************************************************** |
9838 | * |
9839 | * COMPlus_JitDumpIR support - dump out tree leaf node for linear IR form |
9840 | */ |
9841 | |
9842 | int cLeafIR(Compiler* comp, GenTree* tree) |
9843 | { |
9844 | int chars = 0; |
9845 | genTreeOps op = tree->OperGet(); |
9846 | const char* ilKind = nullptr; |
9847 | const char* ilName = nullptr; |
9848 | unsigned ilNum = 0; |
9849 | unsigned lclNum = 0; |
9850 | bool hasSsa = false; |
9851 | |
9852 | switch (op) |
9853 | { |
9854 | |
9855 | case GT_PHI_ARG: |
9856 | case GT_LCL_VAR: |
9857 | case GT_LCL_VAR_ADDR: |
9858 | case GT_STORE_LCL_VAR: |
9859 | lclNum = tree->gtLclVarCommon.gtLclNum; |
9860 | comp->gtGetLclVarNameInfo(lclNum, &ilKind, &ilName, &ilNum); |
9861 | if (ilName != nullptr) |
9862 | { |
9863 | chars += printf("%s" , ilName); |
9864 | } |
9865 | else |
9866 | { |
9867 | LclVarDsc* varDsc = comp->lvaTable + lclNum; |
9868 | chars += printf("%s%d" , ilKind, ilNum); |
9869 | if (comp->dumpIRLocals) |
9870 | { |
9871 | chars += printf("(V%02u" , lclNum); |
9872 | if (varDsc->lvTracked) |
9873 | { |
9874 | chars += printf(":T%02u" , varDsc->lvVarIndex); |
9875 | } |
9876 | if (comp->dumpIRRegs) |
9877 | { |
9878 | if (varDsc->lvRegister) |
9879 | { |
9880 | chars += printf(":%s" , getRegName(varDsc->lvRegNum)); |
9881 | } |
9882 | else |
9883 | { |
9884 | switch (tree->GetRegTag()) |
9885 | { |
9886 | case GenTree::GT_REGTAG_REG: |
9887 | chars += printf(":%s" , comp->compRegVarName(tree->gtRegNum)); |
9888 | break; |
9889 | default: |
9890 | break; |
9891 | } |
9892 | } |
9893 | } |
9894 | chars += printf(")" ); |
9895 | } |
9896 | else if (comp->dumpIRRegs) |
9897 | { |
9898 | if (varDsc->lvRegister) |
9899 | { |
9900 | chars += printf("(%s)" , getRegName(varDsc->lvRegNum)); |
9901 | } |
9902 | else |
9903 | { |
9904 | switch (tree->GetRegTag()) |
9905 | { |
9906 | case GenTree::GT_REGTAG_REG: |
9907 | chars += printf("(%s)" , comp->compRegVarName(tree->gtRegNum)); |
9908 | break; |
9909 | default: |
9910 | break; |
9911 | } |
9912 | } |
9913 | } |
9914 | } |
9915 | |
9916 | hasSsa = true; |
9917 | break; |
9918 | |
9919 | case GT_LCL_FLD: |
9920 | case GT_LCL_FLD_ADDR: |
9921 | case GT_STORE_LCL_FLD: |
9922 | |
9923 | lclNum = tree->gtLclVarCommon.gtLclNum; |
9924 | comp->gtGetLclVarNameInfo(lclNum, &ilKind, &ilName, &ilNum); |
9925 | if (ilName != nullptr) |
9926 | { |
9927 | chars += printf("%s+%u" , ilName, tree->gtLclFld.gtLclOffs); |
9928 | } |
9929 | else |
9930 | { |
9931 | chars += printf("%s%d+%u" , ilKind, ilNum, tree->gtLclFld.gtLclOffs); |
9932 | LclVarDsc* varDsc = comp->lvaTable + lclNum; |
9933 | if (comp->dumpIRLocals) |
9934 | { |
9935 | chars += printf("(V%02u" , lclNum); |
9936 | if (varDsc->lvTracked) |
9937 | { |
9938 | chars += printf(":T%02u" , varDsc->lvVarIndex); |
9939 | } |
9940 | if (comp->dumpIRRegs) |
9941 | { |
9942 | if (varDsc->lvRegister) |
9943 | { |
9944 | chars += printf(":%s" , getRegName(varDsc->lvRegNum)); |
9945 | } |
9946 | else |
9947 | { |
9948 | switch (tree->GetRegTag()) |
9949 | { |
9950 | case GenTree::GT_REGTAG_REG: |
9951 | chars += printf(":%s" , comp->compRegVarName(tree->gtRegNum)); |
9952 | break; |
9953 | default: |
9954 | break; |
9955 | } |
9956 | } |
9957 | } |
9958 | chars += printf(")" ); |
9959 | } |
9960 | else if (comp->dumpIRRegs) |
9961 | { |
9962 | if (varDsc->lvRegister) |
9963 | { |
9964 | chars += printf("(%s)" , getRegName(varDsc->lvRegNum)); |
9965 | } |
9966 | else |
9967 | { |
9968 | switch (tree->GetRegTag()) |
9969 | { |
9970 | case GenTree::GT_REGTAG_REG: |
9971 | chars += printf("(%s)" , comp->compRegVarName(tree->gtRegNum)); |
9972 | break; |
9973 | default: |
9974 | break; |
9975 | } |
9976 | } |
9977 | } |
9978 | } |
9979 | |
9980 | // TODO: We probably want to expand field sequence. |
9981 | // gtDispFieldSeq(tree->gtLclFld.gtFieldSeq); |
9982 | |
9983 | hasSsa = true; |
9984 | break; |
9985 | |
9986 | case GT_CNS_INT: |
9987 | |
9988 | if (tree->IsIconHandle()) |
9989 | { |
9990 | #if 0 |
9991 | // TODO: Commented out because sometimes the CLR throws |
9992 | // and exception when asking the names of some handles. |
9993 | // Need to investigate. |
9994 | |
9995 | const char* className; |
9996 | const char* fieldName; |
9997 | const char* methodName; |
9998 | const wchar_t* str; |
9999 | |
10000 | switch (tree->GetIconHandleFlag()) |
10001 | { |
10002 | |
10003 | case GTF_ICON_SCOPE_HDL: |
10004 | |
10005 | chars += printf("SCOPE(?)" ); |
10006 | break; |
10007 | |
10008 | case GTF_ICON_CLASS_HDL: |
10009 | |
10010 | className = comp->eeGetClassName((CORINFO_CLASS_HANDLE)tree->gtIntCon.gtIconVal); |
10011 | chars += printf("CLASS(%s)" , className); |
10012 | break; |
10013 | |
10014 | case GTF_ICON_METHOD_HDL: |
10015 | |
10016 | methodName = comp->eeGetMethodName((CORINFO_METHOD_HANDLE)tree->gtIntCon.gtIconVal, |
10017 | &className); |
10018 | chars += printf("METHOD(%s.%s)" , className, methodName); |
10019 | break; |
10020 | |
10021 | case GTF_ICON_FIELD_HDL: |
10022 | |
10023 | fieldName = comp->eeGetFieldName((CORINFO_FIELD_HANDLE)tree->gtIntCon.gtIconVal, |
10024 | &className); |
10025 | chars += printf("FIELD(%s.%s) " , className, fieldName); |
10026 | break; |
10027 | |
10028 | case GTF_ICON_STATIC_HDL: |
10029 | |
10030 | fieldName = comp->eeGetFieldName((CORINFO_FIELD_HANDLE)tree->gtIntCon.gtIconVal, |
10031 | &className); |
10032 | chars += printf("STATIC_FIELD(%s.%s)" , className, fieldName); |
10033 | break; |
10034 | |
10035 | case GTF_ICON_STR_HDL: |
10036 | |
10037 | str = comp->eeGetCPString(tree->gtIntCon.gtIconVal); |
10038 | chars += printf("\"%S\"" , str); |
10039 | break; |
10040 | |
10041 | case GTF_ICON_PSTR_HDL: |
10042 | |
10043 | chars += printf("PSTR(?)" ); |
10044 | break; |
10045 | |
10046 | case GTF_ICON_PTR_HDL: |
10047 | |
10048 | chars += printf("PTR(?)" ); |
10049 | break; |
10050 | |
10051 | case GTF_ICON_VARG_HDL: |
10052 | |
10053 | chars += printf("VARARG(?)" ); |
10054 | break; |
10055 | |
10056 | case GTF_ICON_PINVKI_HDL: |
10057 | |
10058 | chars += printf("PINVOKE(?)" ); |
10059 | break; |
10060 | |
10061 | case GTF_ICON_TOKEN_HDL: |
10062 | |
10063 | chars += printf("TOKEN(%08X)" , tree->gtIntCon.gtIconVal); |
10064 | break; |
10065 | |
10066 | case GTF_ICON_TLS_HDL: |
10067 | |
10068 | chars += printf("TLS(?)" ); |
10069 | break; |
10070 | |
10071 | case GTF_ICON_FTN_ADDR: |
10072 | |
10073 | chars += printf("FTN(?)" ); |
10074 | break; |
10075 | |
10076 | case GTF_ICON_CIDMID_HDL: |
10077 | |
10078 | chars += printf("CIDMID(?)" ); |
10079 | break; |
10080 | |
10081 | case GTF_ICON_BBC_PTR: |
10082 | |
10083 | chars += printf("BBC(?)" ); |
10084 | break; |
10085 | |
10086 | default: |
10087 | |
10088 | chars += printf("HANDLE(?)" ); |
10089 | break; |
10090 | } |
10091 | #else |
10092 | #ifdef _TARGET_64BIT_ |
10093 | if ((tree->gtIntCon.gtIconVal & 0xFFFFFFFF00000000LL) != 0) |
10094 | { |
10095 | chars += printf("HANDLE(0x%llx)" , dspPtr(tree->gtIntCon.gtIconVal)); |
10096 | } |
10097 | else |
10098 | #endif |
10099 | { |
10100 | chars += printf("HANDLE(0x%0x)" , dspPtr(tree->gtIntCon.gtIconVal)); |
10101 | } |
10102 | #endif |
10103 | } |
10104 | else |
10105 | { |
10106 | if (tree->TypeGet() == TYP_REF) |
10107 | { |
10108 | assert(tree->gtIntCon.gtIconVal == 0); |
10109 | chars += printf("null" ); |
10110 | } |
10111 | #ifdef _TARGET_64BIT_ |
10112 | else if ((tree->gtIntCon.gtIconVal & 0xFFFFFFFF00000000LL) != 0) |
10113 | { |
10114 | chars += printf("0x%llx" , tree->gtIntCon.gtIconVal); |
10115 | } |
10116 | else |
10117 | #endif |
10118 | { |
10119 | chars += printf("%ld(0x%x)" , tree->gtIntCon.gtIconVal, tree->gtIntCon.gtIconVal); |
10120 | } |
10121 | } |
10122 | break; |
10123 | |
10124 | case GT_CNS_LNG: |
10125 | |
10126 | chars += printf("CONST(LONG)" ); |
10127 | break; |
10128 | |
10129 | case GT_CNS_DBL: |
10130 | |
10131 | chars += printf("CONST(DOUBLE)" ); |
10132 | break; |
10133 | |
10134 | case GT_CNS_STR: |
10135 | |
10136 | chars += printf("CONST(STR)" ); |
10137 | break; |
10138 | |
10139 | case GT_JMP: |
10140 | |
10141 | { |
10142 | const char* methodName; |
10143 | const char* className; |
10144 | |
10145 | methodName = comp->eeGetMethodName((CORINFO_METHOD_HANDLE)tree->gtVal.gtVal1, &className); |
10146 | chars += printf(" %s.%s" , className, methodName); |
10147 | } |
10148 | break; |
10149 | |
10150 | case GT_NO_OP: |
10151 | case GT_START_NONGC: |
10152 | case GT_PROF_HOOK: |
10153 | case GT_CATCH_ARG: |
10154 | case GT_MEMORYBARRIER: |
10155 | case GT_ARGPLACE: |
10156 | case GT_PINVOKE_PROLOG: |
10157 | case GT_JMPTABLE: |
10158 | // Do nothing. |
10159 | break; |
10160 | |
10161 | case GT_RET_EXPR: |
10162 | |
10163 | chars += printf("t%d" , tree->gtRetExpr.gtInlineCandidate->gtTreeID); |
10164 | break; |
10165 | |
10166 | case GT_PHYSREG: |
10167 | |
10168 | chars += printf("%s" , getRegName(tree->gtPhysReg.gtSrcReg, varTypeIsFloating(tree))); |
10169 | break; |
10170 | |
10171 | case GT_LABEL: |
10172 | |
10173 | if (tree->gtLabel.gtLabBB) |
10174 | { |
10175 | chars += printf(FMT_BB, tree->gtLabel.gtLabBB->bbNum); |
10176 | } |
10177 | else |
10178 | { |
10179 | chars += printf("BB?" ); |
10180 | } |
10181 | break; |
10182 | |
10183 | case GT_IL_OFFSET: |
10184 | |
10185 | if (tree->gtStmt.gtStmtILoffsx == BAD_IL_OFFSET) |
10186 | { |
10187 | chars += printf("?" ); |
10188 | } |
10189 | else |
10190 | { |
10191 | chars += printf("0x%x" , jitGetILoffs(tree->gtStmt.gtStmtILoffsx)); |
10192 | } |
10193 | break; |
10194 | |
10195 | case GT_CLS_VAR: |
10196 | case GT_CLS_VAR_ADDR: |
10197 | default: |
10198 | |
10199 | if (tree->OperIsLeaf()) |
10200 | { |
10201 | chars += printf("<leaf nyi: %s>" , tree->OpName(tree->OperGet())); |
10202 | } |
10203 | |
10204 | chars += printf("t%d" , tree->gtTreeID); |
10205 | break; |
10206 | } |
10207 | |
10208 | if (comp->dumpIRTypes) |
10209 | { |
10210 | chars += cTreeTypeIR(comp, tree); |
10211 | } |
10212 | if (comp->dumpIRValnums) |
10213 | { |
10214 | chars += cValNumIR(comp, tree); |
10215 | } |
10216 | if (hasSsa && comp->dumpIRSsa) |
10217 | { |
10218 | chars += cSsaNumIR(comp, tree); |
10219 | } |
10220 | |
10221 | return chars; |
10222 | } |
10223 | |
10224 | /***************************************************************************** |
10225 | * |
10226 | * COMPlus_JitDumpIR support - dump out tree leaf node for linear IR form |
10227 | */ |
10228 | |
10229 | int dLeafIR(GenTree* tree) |
10230 | { |
10231 | int chars = cLeafIR(JitTls::GetCompiler(), tree); |
10232 | |
10233 | return chars; |
10234 | } |
10235 | |
10236 | /***************************************************************************** |
10237 | * |
10238 | * COMPlus_JitDumpIR support - dump out tree indir node for linear IR form |
10239 | */ |
10240 | |
10241 | int cIndirIR(Compiler* comp, GenTree* tree) |
10242 | { |
10243 | assert(tree->gtOper == GT_IND); |
10244 | |
10245 | int chars = 0; |
10246 | GenTree* child; |
10247 | |
10248 | chars += printf("[" ); |
10249 | child = tree->GetChild(0); |
10250 | chars += cLeafIR(comp, child); |
10251 | chars += printf("]" ); |
10252 | |
10253 | return chars; |
10254 | } |
10255 | |
10256 | /***************************************************************************** |
10257 | * |
10258 | * COMPlus_JitDumpIR support - dump out tree indir node for linear IR form |
10259 | */ |
10260 | |
10261 | int dIndirIR(GenTree* tree) |
10262 | { |
10263 | int chars = cIndirIR(JitTls::GetCompiler(), tree); |
10264 | |
10265 | return chars; |
10266 | } |
10267 | |
10268 | /***************************************************************************** |
10269 | * |
10270 | * COMPlus_JitDumpIR support - dump out tree operand node for linear IR form |
10271 | */ |
10272 | |
10273 | int cOperandIR(Compiler* comp, GenTree* operand) |
10274 | { |
10275 | int chars = 0; |
10276 | |
10277 | if (operand == nullptr) |
10278 | { |
10279 | chars += printf("t?" ); |
10280 | return chars; |
10281 | } |
10282 | |
10283 | bool dumpTypes = comp->dumpIRTypes; |
10284 | bool dumpValnums = comp->dumpIRValnums; |
10285 | bool foldIndirs = comp->dumpIRDataflow; |
10286 | bool foldLeafs = comp->dumpIRNoLeafs; |
10287 | bool foldCommas = comp->dumpIRDataflow; |
10288 | bool dumpDataflow = comp->dumpIRDataflow; |
10289 | bool foldLists = comp->dumpIRNoLists; |
10290 | bool dumpRegs = comp->dumpIRRegs; |
10291 | |
10292 | genTreeOps op = operand->OperGet(); |
10293 | |
10294 | if (foldLeafs && operand->OperIsLeaf()) |
10295 | { |
10296 | if ((op == GT_ARGPLACE) && foldLists) |
10297 | { |
10298 | return chars; |
10299 | } |
10300 | chars += cLeafIR(comp, operand); |
10301 | } |
10302 | else if (dumpDataflow && (operand->OperIs(GT_ASG) || (op == GT_STORE_LCL_VAR) || (op == GT_STORE_LCL_FLD))) |
10303 | { |
10304 | operand = operand->GetChild(0); |
10305 | chars += cOperandIR(comp, operand); |
10306 | } |
10307 | else if ((op == GT_INDEX) && foldIndirs) |
10308 | { |
10309 | chars += printf("[t%d]" , operand->gtTreeID); |
10310 | if (dumpTypes) |
10311 | { |
10312 | chars += cTreeTypeIR(comp, operand); |
10313 | } |
10314 | if (dumpValnums) |
10315 | { |
10316 | chars += cValNumIR(comp, operand); |
10317 | } |
10318 | } |
10319 | else if ((op == GT_IND) && foldIndirs) |
10320 | { |
10321 | chars += cIndirIR(comp, operand); |
10322 | if (dumpTypes) |
10323 | { |
10324 | chars += cTreeTypeIR(comp, operand); |
10325 | } |
10326 | if (dumpValnums) |
10327 | { |
10328 | chars += cValNumIR(comp, operand); |
10329 | } |
10330 | } |
10331 | else if ((op == GT_COMMA) && foldCommas) |
10332 | { |
10333 | operand = operand->GetChild(1); |
10334 | chars += cOperandIR(comp, operand); |
10335 | } |
10336 | else if ((op == GT_LIST) && foldLists) |
10337 | { |
10338 | GenTree* list = operand; |
10339 | unsigned childCount = list->NumChildren(); |
10340 | |
10341 | operand = list->GetChild(0); |
10342 | int operandChars = cOperandIR(comp, operand); |
10343 | chars += operandChars; |
10344 | if (childCount > 1) |
10345 | { |
10346 | if (operandChars > 0) |
10347 | { |
10348 | chars += printf(", " ); |
10349 | } |
10350 | operand = list->GetChild(1); |
10351 | if (operand->gtOper == GT_LIST) |
10352 | { |
10353 | chars += cListIR(comp, operand); |
10354 | } |
10355 | else |
10356 | { |
10357 | chars += cOperandIR(comp, operand); |
10358 | } |
10359 | } |
10360 | } |
10361 | else |
10362 | { |
10363 | chars += printf("t%d" , operand->gtTreeID); |
10364 | if (dumpRegs) |
10365 | { |
10366 | regNumber regNum = operand->GetReg(); |
10367 | if (regNum != REG_NA) |
10368 | { |
10369 | chars += printf("(%s)" , getRegName(regNum)); |
10370 | } |
10371 | } |
10372 | if (dumpTypes) |
10373 | { |
10374 | chars += cTreeTypeIR(comp, operand); |
10375 | } |
10376 | if (dumpValnums) |
10377 | { |
10378 | chars += cValNumIR(comp, operand); |
10379 | } |
10380 | } |
10381 | |
10382 | return chars; |
10383 | } |
10384 | |
10385 | /***************************************************************************** |
10386 | * |
10387 | * COMPlus_JitDumpIR support - dump out tree operand node for linear IR form |
10388 | */ |
10389 | |
10390 | int dOperandIR(GenTree* operand) |
10391 | { |
10392 | int chars = cOperandIR(JitTls::GetCompiler(), operand); |
10393 | |
10394 | return chars; |
10395 | } |
10396 | |
10397 | /***************************************************************************** |
10398 | * |
10399 | * COMPlus_JitDumpIR support - dump out tree list of nodes for linear IR form |
10400 | */ |
10401 | |
10402 | int cListIR(Compiler* comp, GenTree* list) |
10403 | { |
10404 | int chars = 0; |
10405 | int operandChars; |
10406 | |
10407 | assert(list->gtOper == GT_LIST); |
10408 | |
10409 | GenTree* child; |
10410 | unsigned childCount; |
10411 | |
10412 | childCount = list->NumChildren(); |
10413 | assert(childCount == 1 || childCount == 2); |
10414 | |
10415 | operandChars = 0; |
10416 | for (unsigned childIndex = 0; childIndex < childCount; childIndex++) |
10417 | { |
10418 | if ((childIndex > 0) && (operandChars > 0)) |
10419 | { |
10420 | chars += printf(", " ); |
10421 | } |
10422 | |
10423 | child = list->GetChild(childIndex); |
10424 | operandChars = cOperandIR(comp, child); |
10425 | chars += operandChars; |
10426 | } |
10427 | |
10428 | return chars; |
10429 | } |
10430 | |
10431 | /***************************************************************************** |
10432 | * |
10433 | * COMPlus_JitDumpIR support - dump out tree list of nodes for linear IR form |
10434 | */ |
10435 | |
10436 | int dListIR(GenTree* list) |
10437 | { |
10438 | int chars = cListIR(JitTls::GetCompiler(), list); |
10439 | |
10440 | return chars; |
10441 | } |
10442 | |
10443 | /***************************************************************************** |
10444 | * |
10445 | * COMPlus_JitDumpIR support - dump out tree dependencies based on comma nodes for linear IR form |
10446 | */ |
10447 | |
10448 | int cDependsIR(Compiler* comp, GenTree* comma, bool* first) |
10449 | { |
10450 | int chars = 0; |
10451 | |
10452 | assert(comma->gtOper == GT_COMMA); |
10453 | |
10454 | GenTree* child; |
10455 | |
10456 | child = comma->GetChild(0); |
10457 | if (child->gtOper == GT_COMMA) |
10458 | { |
10459 | chars += cDependsIR(comp, child, first); |
10460 | } |
10461 | else |
10462 | { |
10463 | if (!(*first)) |
10464 | { |
10465 | chars += printf(", " ); |
10466 | } |
10467 | chars += printf("t%d" , child->gtTreeID); |
10468 | *first = false; |
10469 | } |
10470 | |
10471 | child = comma->GetChild(1); |
10472 | if (child->gtOper == GT_COMMA) |
10473 | { |
10474 | chars += cDependsIR(comp, child, first); |
10475 | } |
10476 | |
10477 | return chars; |
10478 | } |
10479 | |
10480 | /***************************************************************************** |
10481 | * |
10482 | * COMPlus_JitDumpIR support - dump out tree dependencies based on comma nodes for linear IR form |
10483 | */ |
10484 | |
10485 | int dDependsIR(GenTree* comma) |
10486 | { |
10487 | int chars = 0; |
10488 | bool first = TRUE; |
10489 | |
10490 | chars = cDependsIR(JitTls::GetCompiler(), comma, &first); |
10491 | |
10492 | return chars; |
10493 | } |
10494 | |
10495 | /***************************************************************************** |
10496 | * |
10497 | * COMPlus_JitDumpIR support - dump out tree node in linear IR form |
10498 | */ |
10499 | |
10500 | void cNodeIR(Compiler* comp, GenTree* tree) |
10501 | { |
10502 | bool foldLeafs = comp->dumpIRNoLeafs; |
10503 | bool foldIndirs = comp->dumpIRDataflow; |
10504 | bool foldLists = comp->dumpIRNoLists; |
10505 | bool dataflowView = comp->dumpIRDataflow; |
10506 | bool dumpTypes = comp->dumpIRTypes; |
10507 | bool dumpValnums = comp->dumpIRValnums; |
10508 | bool noStmts = comp->dumpIRNoStmts; |
10509 | genTreeOps op = tree->OperGet(); |
10510 | unsigned childCount = tree->NumChildren(); |
10511 | GenTree* child; |
10512 | |
10513 | // What are we skipping? |
10514 | |
10515 | if (tree->OperIsLeaf()) |
10516 | { |
10517 | if (foldLeafs) |
10518 | { |
10519 | return; |
10520 | } |
10521 | } |
10522 | else if (op == GT_IND) |
10523 | { |
10524 | if (foldIndirs) |
10525 | { |
10526 | return; |
10527 | } |
10528 | } |
10529 | else if (op == GT_LIST) |
10530 | { |
10531 | if (foldLists) |
10532 | { |
10533 | return; |
10534 | } |
10535 | } |
10536 | else if (op == GT_STMT) |
10537 | { |
10538 | if (noStmts) |
10539 | { |
10540 | if (dataflowView) |
10541 | { |
10542 | child = tree->GetChild(0); |
10543 | if (child->gtOper != GT_COMMA) |
10544 | { |
10545 | return; |
10546 | } |
10547 | } |
10548 | else |
10549 | { |
10550 | return; |
10551 | } |
10552 | } |
10553 | } |
10554 | else if (op == GT_COMMA) |
10555 | { |
10556 | if (dataflowView) |
10557 | { |
10558 | return; |
10559 | } |
10560 | } |
10561 | |
10562 | bool nodeIsValue = tree->IsValue(); |
10563 | |
10564 | // Dump tree id or dataflow destination. |
10565 | |
10566 | int chars = 0; |
10567 | |
10568 | // if (comp->compRationalIRForm) |
10569 | // { |
10570 | // chars += printf("R"); |
10571 | // } |
10572 | |
10573 | chars += printf(" " ); |
10574 | if (dataflowView && tree->OperIs(GT_ASG)) |
10575 | { |
10576 | child = tree->GetChild(0); |
10577 | chars += cOperandIR(comp, child); |
10578 | } |
10579 | else if (dataflowView && ((op == GT_STORE_LCL_VAR) || (op == GT_STORE_LCL_FLD))) |
10580 | { |
10581 | chars += cLeafIR(comp, tree); |
10582 | } |
10583 | else if (dataflowView && (op == GT_STOREIND)) |
10584 | { |
10585 | child = tree->GetChild(0); |
10586 | chars += printf("[" ); |
10587 | chars += cOperandIR(comp, child); |
10588 | chars += printf("]" ); |
10589 | if (dumpTypes) |
10590 | { |
10591 | chars += cTreeTypeIR(comp, tree); |
10592 | } |
10593 | if (dumpValnums) |
10594 | { |
10595 | chars += cValNumIR(comp, tree); |
10596 | } |
10597 | } |
10598 | else if (nodeIsValue) |
10599 | { |
10600 | chars += printf("t%d" , tree->gtTreeID); |
10601 | if (comp->dumpIRRegs) |
10602 | { |
10603 | regNumber regNum = tree->GetReg(); |
10604 | if (regNum != REG_NA) |
10605 | { |
10606 | chars += printf("(%s)" , getRegName(regNum)); |
10607 | } |
10608 | } |
10609 | if (dumpTypes) |
10610 | { |
10611 | chars += cTreeTypeIR(comp, tree); |
10612 | } |
10613 | if (dumpValnums) |
10614 | { |
10615 | chars += cValNumIR(comp, tree); |
10616 | } |
10617 | } |
10618 | |
10619 | // Dump opcode and tree ID if need in dataflow view. |
10620 | |
10621 | chars += dTabStopIR(chars, COLUMN_OPCODE); |
10622 | const char* opName = tree->OpName(op); |
10623 | chars += printf(" %c %s" , nodeIsValue ? '=' : ' ', opName); |
10624 | |
10625 | if (dataflowView) |
10626 | { |
10627 | if (tree->OperIs(GT_ASG) || (op == GT_STORE_LCL_VAR) || (op == GT_STORE_LCL_FLD) || (op == GT_STOREIND)) |
10628 | { |
10629 | chars += printf("(t%d)" , tree->gtTreeID); |
10630 | } |
10631 | } |
10632 | |
10633 | // Dump modifiers for opcodes to help with readability |
10634 | |
10635 | if (op == GT_CALL) |
10636 | { |
10637 | GenTreeCall* call = tree->AsCall(); |
10638 | |
10639 | if (call->gtCallType == CT_USER_FUNC) |
10640 | { |
10641 | if (call->IsVirtualStub()) |
10642 | { |
10643 | chars += printf(":VS" ); |
10644 | } |
10645 | else if (call->IsVirtualVtable()) |
10646 | { |
10647 | chars += printf(":VT" ); |
10648 | } |
10649 | else if (call->IsVirtual()) |
10650 | { |
10651 | chars += printf(":V" ); |
10652 | } |
10653 | } |
10654 | else if (call->gtCallType == CT_HELPER) |
10655 | { |
10656 | chars += printf(":H" ); |
10657 | } |
10658 | else if (call->gtCallType == CT_INDIRECT) |
10659 | { |
10660 | chars += printf(":I" ); |
10661 | } |
10662 | else if (call->IsUnmanaged()) |
10663 | { |
10664 | chars += printf(":U" ); |
10665 | } |
10666 | else |
10667 | { |
10668 | if (call->IsVirtualStub()) |
10669 | { |
10670 | chars += printf(":XVS" ); |
10671 | } |
10672 | else if (call->IsVirtualVtable()) |
10673 | { |
10674 | chars += printf(":XVT" ); |
10675 | } |
10676 | else |
10677 | { |
10678 | chars += printf(":?" ); |
10679 | } |
10680 | } |
10681 | |
10682 | if (call->IsUnmanaged()) |
10683 | { |
10684 | if (call->gtCallMoreFlags & GTF_CALL_M_UNMGD_THISCALL) |
10685 | { |
10686 | chars += printf(":T" ); |
10687 | } |
10688 | } |
10689 | |
10690 | if (tree->gtFlags & GTF_CALL_NULLCHECK) |
10691 | { |
10692 | chars += printf(":N" ); |
10693 | } |
10694 | } |
10695 | else if (op == GT_INTRINSIC) |
10696 | { |
10697 | CorInfoIntrinsics intrin = tree->gtIntrinsic.gtIntrinsicId; |
10698 | |
10699 | chars += printf(":" ); |
10700 | switch (intrin) |
10701 | { |
10702 | case CORINFO_INTRINSIC_Sin: |
10703 | chars += printf("Sin" ); |
10704 | break; |
10705 | case CORINFO_INTRINSIC_Cos: |
10706 | chars += printf("Cos" ); |
10707 | break; |
10708 | case CORINFO_INTRINSIC_Cbrt: |
10709 | chars += printf("Cbrt" ); |
10710 | break; |
10711 | case CORINFO_INTRINSIC_Sqrt: |
10712 | chars += printf("Sqrt" ); |
10713 | break; |
10714 | case CORINFO_INTRINSIC_Cosh: |
10715 | chars += printf("Cosh" ); |
10716 | break; |
10717 | case CORINFO_INTRINSIC_Sinh: |
10718 | chars += printf("Sinh" ); |
10719 | break; |
10720 | case CORINFO_INTRINSIC_Tan: |
10721 | chars += printf("Tan" ); |
10722 | break; |
10723 | case CORINFO_INTRINSIC_Tanh: |
10724 | chars += printf("Tanh" ); |
10725 | break; |
10726 | case CORINFO_INTRINSIC_Asin: |
10727 | chars += printf("Asin" ); |
10728 | break; |
10729 | case CORINFO_INTRINSIC_Asinh: |
10730 | chars += printf("Asinh" ); |
10731 | break; |
10732 | case CORINFO_INTRINSIC_Acos: |
10733 | chars += printf("Acos" ); |
10734 | break; |
10735 | case CORINFO_INTRINSIC_Acosh: |
10736 | chars += printf("Acosh" ); |
10737 | break; |
10738 | case CORINFO_INTRINSIC_Atan: |
10739 | chars += printf("Atan" ); |
10740 | break; |
10741 | case CORINFO_INTRINSIC_Atan2: |
10742 | chars += printf("Atan2" ); |
10743 | break; |
10744 | case CORINFO_INTRINSIC_Atanh: |
10745 | chars += printf("Atanh" ); |
10746 | break; |
10747 | case CORINFO_INTRINSIC_Log10: |
10748 | chars += printf("Log10" ); |
10749 | break; |
10750 | case CORINFO_INTRINSIC_Pow: |
10751 | chars += printf("Pow" ); |
10752 | break; |
10753 | case CORINFO_INTRINSIC_Exp: |
10754 | chars += printf("Exp" ); |
10755 | break; |
10756 | case CORINFO_INTRINSIC_Ceiling: |
10757 | chars += printf("Ceiling" ); |
10758 | break; |
10759 | case CORINFO_INTRINSIC_Floor: |
10760 | chars += printf("Floor" ); |
10761 | break; |
10762 | default: |
10763 | chars += printf("unknown(%d)" , intrin); |
10764 | break; |
10765 | } |
10766 | } |
10767 | |
10768 | // Dump operands. |
10769 | |
10770 | chars += dTabStopIR(chars, COLUMN_OPERANDS); |
10771 | |
10772 | // Dump operator specific fields as operands |
10773 | |
10774 | switch (op) |
10775 | { |
10776 | default: |
10777 | break; |
10778 | case GT_FIELD: |
10779 | |
10780 | { |
10781 | const char* className = nullptr; |
10782 | const char* fieldName = comp->eeGetFieldName(tree->gtField.gtFldHnd, &className); |
10783 | |
10784 | chars += printf(" %s.%s" , className, fieldName); |
10785 | } |
10786 | break; |
10787 | |
10788 | case GT_CALL: |
10789 | |
10790 | if (tree->gtCall.gtCallType != CT_INDIRECT) |
10791 | { |
10792 | const char* methodName; |
10793 | const char* className; |
10794 | |
10795 | methodName = comp->eeGetMethodName(tree->gtCall.gtCallMethHnd, &className); |
10796 | |
10797 | chars += printf(" %s.%s" , className, methodName); |
10798 | } |
10799 | break; |
10800 | |
10801 | case GT_STORE_LCL_VAR: |
10802 | case GT_STORE_LCL_FLD: |
10803 | |
10804 | if (!dataflowView) |
10805 | { |
10806 | chars += printf(" " ); |
10807 | chars += cLeafIR(comp, tree); |
10808 | } |
10809 | break; |
10810 | |
10811 | case GT_LEA: |
10812 | |
10813 | GenTreeAddrMode* lea = tree->AsAddrMode(); |
10814 | GenTree* base = lea->Base(); |
10815 | GenTree* index = lea->Index(); |
10816 | unsigned scale = lea->gtScale; |
10817 | int offset = lea->Offset(); |
10818 | |
10819 | chars += printf(" [" ); |
10820 | if (base != nullptr) |
10821 | { |
10822 | chars += cOperandIR(comp, base); |
10823 | } |
10824 | if (index != nullptr) |
10825 | { |
10826 | if (base != nullptr) |
10827 | { |
10828 | chars += printf("+" ); |
10829 | } |
10830 | chars += cOperandIR(comp, index); |
10831 | if (scale > 1) |
10832 | { |
10833 | chars += printf("*%u" , scale); |
10834 | } |
10835 | } |
10836 | if ((offset != 0) || ((base == nullptr) && (index == nullptr))) |
10837 | { |
10838 | if ((base != nullptr) || (index != nullptr)) |
10839 | { |
10840 | chars += printf("+" ); |
10841 | } |
10842 | chars += printf("%d" , offset); |
10843 | } |
10844 | chars += printf("]" ); |
10845 | break; |
10846 | } |
10847 | |
10848 | // Dump operands. |
10849 | |
10850 | if (tree->OperIsLeaf()) |
10851 | { |
10852 | chars += printf(" " ); |
10853 | chars += cLeafIR(comp, tree); |
10854 | } |
10855 | else if (op == GT_LEA) |
10856 | { |
10857 | // Already dumped it above. |
10858 | } |
10859 | else if (op == GT_PHI) |
10860 | { |
10861 | if (tree->gtOp.gtOp1 != nullptr) |
10862 | { |
10863 | bool first = true; |
10864 | for (GenTreeArgList* args = tree->gtOp.gtOp1->AsArgList(); args != nullptr; args = args->Rest()) |
10865 | { |
10866 | child = args->Current(); |
10867 | if (!first) |
10868 | { |
10869 | chars += printf("," ); |
10870 | } |
10871 | first = false; |
10872 | chars += printf(" " ); |
10873 | chars += cOperandIR(comp, child); |
10874 | } |
10875 | } |
10876 | } |
10877 | else |
10878 | { |
10879 | bool hasComma = false; |
10880 | bool first = true; |
10881 | int operandChars = 0; |
10882 | for (unsigned childIndex = 0; childIndex < childCount; childIndex++) |
10883 | { |
10884 | child = tree->GetChild(childIndex); |
10885 | if (child == nullptr) |
10886 | { |
10887 | continue; |
10888 | } |
10889 | |
10890 | if (child->gtOper == GT_COMMA) |
10891 | { |
10892 | hasComma = true; |
10893 | } |
10894 | |
10895 | if (dataflowView && (childIndex == 0)) |
10896 | { |
10897 | if ((op == GT_ASG) || (op == GT_STOREIND)) |
10898 | { |
10899 | continue; |
10900 | } |
10901 | } |
10902 | |
10903 | if (!first) |
10904 | { |
10905 | chars += printf("," ); |
10906 | } |
10907 | |
10908 | bool isList = (child->gtOper == GT_LIST); |
10909 | if (!isList || !foldLists) |
10910 | { |
10911 | if (foldLeafs && (child->gtOper == GT_ARGPLACE)) |
10912 | { |
10913 | continue; |
10914 | } |
10915 | chars += printf(" " ); |
10916 | operandChars = cOperandIR(comp, child); |
10917 | chars += operandChars; |
10918 | if (operandChars > 0) |
10919 | { |
10920 | first = false; |
10921 | } |
10922 | } |
10923 | else |
10924 | { |
10925 | assert(isList); |
10926 | chars += printf(" " ); |
10927 | operandChars = cOperandIR(comp, child); |
10928 | chars += operandChars; |
10929 | if (operandChars > 0) |
10930 | { |
10931 | first = false; |
10932 | } |
10933 | } |
10934 | } |
10935 | |
10936 | if (dataflowView && hasComma) |
10937 | { |
10938 | chars += printf(", DEPS(" ); |
10939 | first = true; |
10940 | for (unsigned childIndex = 0; childIndex < childCount; childIndex++) |
10941 | { |
10942 | child = tree->GetChild(childIndex); |
10943 | if (child->gtOper == GT_COMMA) |
10944 | { |
10945 | chars += cDependsIR(comp, child, &first); |
10946 | } |
10947 | } |
10948 | chars += printf(")" ); |
10949 | } |
10950 | } |
10951 | |
10952 | // Dump kinds, flags, costs |
10953 | |
10954 | if (comp->dumpIRKinds || comp->dumpIRFlags || comp->dumpIRCosts) |
10955 | { |
10956 | chars += dTabStopIR(chars, COLUMN_KINDS); |
10957 | chars += printf(";" ); |
10958 | if (comp->dumpIRKinds) |
10959 | { |
10960 | chars += printf(" " ); |
10961 | chars += cTreeKindsIR(comp, tree); |
10962 | } |
10963 | if (comp->dumpIRFlags && (tree->gtFlags != 0)) |
10964 | { |
10965 | if (comp->dumpIRKinds) |
10966 | { |
10967 | chars += dTabStopIR(chars, COLUMN_FLAGS); |
10968 | } |
10969 | else |
10970 | { |
10971 | chars += printf(" " ); |
10972 | } |
10973 | chars += cTreeFlagsIR(comp, tree); |
10974 | } |
10975 | if (comp->dumpIRCosts && (tree->gtCostsInitialized)) |
10976 | { |
10977 | chars += printf(" CostEx=%d, CostSz=%d" , tree->GetCostEx(), tree->GetCostSz()); |
10978 | } |
10979 | } |
10980 | |
10981 | printf("\n" ); |
10982 | } |
10983 | |
10984 | /***************************************************************************** |
10985 | * |
10986 | * COMPlus_JitDumpIR support - dump out tree in linear IR form |
10987 | */ |
10988 | |
10989 | void cTreeIR(Compiler* comp, GenTree* tree) |
10990 | { |
10991 | bool foldLeafs = comp->dumpIRNoLeafs; |
10992 | bool foldIndirs = comp->dumpIRDataflow; |
10993 | bool foldLists = comp->dumpIRNoLists; |
10994 | bool dataflowView = comp->dumpIRDataflow; |
10995 | bool dumpTypes = comp->dumpIRTypes; |
10996 | bool dumpValnums = comp->dumpIRValnums; |
10997 | bool noStmts = comp->dumpIRNoStmts; |
10998 | genTreeOps op = tree->OperGet(); |
10999 | unsigned childCount = tree->NumChildren(); |
11000 | GenTree* child; |
11001 | |
11002 | // Recurse and dump trees that this node depends on. |
11003 | |
11004 | if (tree->OperIsLeaf()) |
11005 | { |
11006 | } |
11007 | else if (tree->OperIsBinary() && tree->IsReverseOp()) |
11008 | { |
11009 | child = tree->GetChild(1); |
11010 | cTreeIR(comp, child); |
11011 | child = tree->GetChild(0); |
11012 | cTreeIR(comp, child); |
11013 | } |
11014 | else if (op == GT_PHI) |
11015 | { |
11016 | // Don't recurse. |
11017 | } |
11018 | else |
11019 | { |
11020 | assert(!tree->IsReverseOp()); |
11021 | for (unsigned childIndex = 0; childIndex < childCount; childIndex++) |
11022 | { |
11023 | child = tree->GetChild(childIndex); |
11024 | if (child != nullptr) |
11025 | { |
11026 | cTreeIR(comp, child); |
11027 | } |
11028 | } |
11029 | } |
11030 | |
11031 | cNodeIR(comp, tree); |
11032 | } |
11033 | |
11034 | /***************************************************************************** |
11035 | * |
11036 | * COMPlus_JitDumpIR support - dump out tree in linear IR form |
11037 | */ |
11038 | |
11039 | void dTreeIR(GenTree* tree) |
11040 | { |
11041 | cTreeIR(JitTls::GetCompiler(), tree); |
11042 | } |
11043 | |
11044 | #endif // DEBUG |
11045 | |
11046 | #if VARSET_COUNTOPS |
11047 | // static |
11048 | BitSetSupport::BitSetOpCounter Compiler::m_varsetOpCounter("VarSetOpCounts.log" ); |
11049 | #endif |
11050 | #if ALLVARSET_COUNTOPS |
11051 | // static |
11052 | BitSetSupport::BitSetOpCounter Compiler::m_allvarsetOpCounter("AllVarSetOpCounts.log" ); |
11053 | #endif |
11054 | |
11055 | // static |
11056 | HelperCallProperties Compiler::s_helperCallProperties; |
11057 | |
11058 | /*****************************************************************************/ |
11059 | /*****************************************************************************/ |
11060 | |
11061 | //------------------------------------------------------------------------ |
11062 | // killGCRefs: |
11063 | // Given some tree node return does it need all GC refs to be spilled from |
11064 | // callee save registers. |
11065 | // |
11066 | // Arguments: |
11067 | // tree - the tree for which we ask about gc refs. |
11068 | // |
11069 | // Return Value: |
11070 | // true - tree kills GC refs on callee save registers |
11071 | // false - tree doesn't affect GC refs on callee save registers |
11072 | bool Compiler::killGCRefs(GenTree* tree) |
11073 | { |
11074 | if (tree->IsCall()) |
11075 | { |
11076 | GenTreeCall* call = tree->AsCall(); |
11077 | if (call->IsUnmanaged()) |
11078 | { |
11079 | return true; |
11080 | } |
11081 | |
11082 | if (call->gtCallMethHnd == eeFindHelper(CORINFO_HELP_JIT_PINVOKE_BEGIN)) |
11083 | { |
11084 | assert(opts.ShouldUsePInvokeHelpers()); |
11085 | return true; |
11086 | } |
11087 | } |
11088 | return false; |
11089 | } |
11090 | |