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 LclVarsInfo XX |
9 | XX XX |
10 | XX The variables to be used by the code generator. XX |
11 | XX XX |
12 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
13 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
14 | */ |
15 | |
16 | #include "jitpch.h" |
17 | #ifdef _MSC_VER |
18 | #pragma hdrstop |
19 | #endif |
20 | #include "emit.h" |
21 | |
22 | #include "register_arg_convention.h" |
23 | |
24 | /*****************************************************************************/ |
25 | |
26 | #ifdef DEBUG |
27 | #if DOUBLE_ALIGN |
28 | /* static */ |
29 | unsigned Compiler::s_lvaDoubleAlignedProcsCount = 0; |
30 | #endif |
31 | #endif |
32 | |
33 | /*****************************************************************************/ |
34 | |
35 | void Compiler::lvaInit() |
36 | { |
37 | /* We haven't allocated stack variables yet */ |
38 | lvaRefCountState = RCS_INVALID; |
39 | |
40 | lvaGenericsContextUseCount = 0; |
41 | |
42 | lvaTrackedToVarNum = nullptr; |
43 | |
44 | lvaTrackedFixed = false; // false: We can still add new tracked variables |
45 | |
46 | lvaDoneFrameLayout = NO_FRAME_LAYOUT; |
47 | #if !FEATURE_EH_FUNCLETS |
48 | lvaShadowSPslotsVar = BAD_VAR_NUM; |
49 | #endif // !FEATURE_EH_FUNCLETS |
50 | lvaInlinedPInvokeFrameVar = BAD_VAR_NUM; |
51 | lvaReversePInvokeFrameVar = BAD_VAR_NUM; |
52 | #if FEATURE_FIXED_OUT_ARGS |
53 | lvaPInvokeFrameRegSaveVar = BAD_VAR_NUM; |
54 | lvaOutgoingArgSpaceVar = BAD_VAR_NUM; |
55 | lvaOutgoingArgSpaceSize = PhasedVar<unsigned>(); |
56 | #endif // FEATURE_FIXED_OUT_ARGS |
57 | #ifdef _TARGET_ARM_ |
58 | lvaPromotedStructAssemblyScratchVar = BAD_VAR_NUM; |
59 | #endif // _TARGET_ARM_ |
60 | #ifdef JIT32_GCENCODER |
61 | lvaLocAllocSPvar = BAD_VAR_NUM; |
62 | #endif // JIT32_GCENCODER |
63 | lvaNewObjArrayArgs = BAD_VAR_NUM; |
64 | lvaGSSecurityCookie = BAD_VAR_NUM; |
65 | #ifdef _TARGET_X86_ |
66 | lvaVarargsBaseOfStkArgs = BAD_VAR_NUM; |
67 | #endif // _TARGET_X86_ |
68 | lvaVarargsHandleArg = BAD_VAR_NUM; |
69 | lvaSecurityObject = BAD_VAR_NUM; |
70 | lvaStubArgumentVar = BAD_VAR_NUM; |
71 | lvaArg0Var = BAD_VAR_NUM; |
72 | lvaMonAcquired = BAD_VAR_NUM; |
73 | |
74 | lvaInlineeReturnSpillTemp = BAD_VAR_NUM; |
75 | |
76 | gsShadowVarInfo = nullptr; |
77 | #if FEATURE_EH_FUNCLETS |
78 | lvaPSPSym = BAD_VAR_NUM; |
79 | #endif |
80 | #if FEATURE_SIMD |
81 | lvaSIMDInitTempVarNum = BAD_VAR_NUM; |
82 | #endif // FEATURE_SIMD |
83 | lvaCurEpoch = 0; |
84 | |
85 | structPromotionHelper = new (this, CMK_Generic) StructPromotionHelper(this); |
86 | } |
87 | |
88 | /*****************************************************************************/ |
89 | |
90 | void Compiler::lvaInitTypeRef() |
91 | { |
92 | |
93 | /* x86 args look something like this: |
94 | [this ptr] [hidden return buffer] [declared arguments]* [generic context] [var arg cookie] |
95 | |
96 | x64 is closer to the native ABI: |
97 | [this ptr] [hidden return buffer] [generic context] [var arg cookie] [declared arguments]* |
98 | (Note: prior to .NET Framework 4.5.1 for Windows 8.1 (but not .NET Framework 4.5.1 "downlevel"), |
99 | the "hidden return buffer" came before the "this ptr". Now, the "this ptr" comes first. This |
100 | is different from the C++ order, where the "hidden return buffer" always comes first.) |
101 | |
102 | ARM and ARM64 are the same as the current x64 convention: |
103 | [this ptr] [hidden return buffer] [generic context] [var arg cookie] [declared arguments]* |
104 | |
105 | Key difference: |
106 | The var arg cookie and generic context are swapped with respect to the user arguments |
107 | */ |
108 | |
109 | /* Set compArgsCount and compLocalsCount */ |
110 | |
111 | info.compArgsCount = info.compMethodInfo->args.numArgs; |
112 | |
113 | // Is there a 'this' pointer |
114 | |
115 | if (!info.compIsStatic) |
116 | { |
117 | info.compArgsCount++; |
118 | } |
119 | else |
120 | { |
121 | info.compThisArg = BAD_VAR_NUM; |
122 | } |
123 | |
124 | info.compILargsCount = info.compArgsCount; |
125 | |
126 | #ifdef FEATURE_SIMD |
127 | if (featureSIMD && (info.compRetNativeType == TYP_STRUCT)) |
128 | { |
129 | var_types structType = impNormStructType(info.compMethodInfo->args.retTypeClass); |
130 | info.compRetType = structType; |
131 | } |
132 | #endif // FEATURE_SIMD |
133 | |
134 | // Are we returning a struct using a return buffer argument? |
135 | // |
136 | const bool hasRetBuffArg = impMethodInfo_hasRetBuffArg(info.compMethodInfo); |
137 | |
138 | // Possibly change the compRetNativeType from TYP_STRUCT to a "primitive" type |
139 | // when we are returning a struct by value and it fits in one register |
140 | // |
141 | if (!hasRetBuffArg && varTypeIsStruct(info.compRetNativeType)) |
142 | { |
143 | CORINFO_CLASS_HANDLE retClsHnd = info.compMethodInfo->args.retTypeClass; |
144 | |
145 | Compiler::structPassingKind howToReturnStruct; |
146 | var_types returnType = getReturnTypeForStruct(retClsHnd, &howToReturnStruct); |
147 | |
148 | // We can safely widen the return type for enclosed structs. |
149 | if ((howToReturnStruct == SPK_PrimitiveType) || (howToReturnStruct == SPK_EnclosingType)) |
150 | { |
151 | assert(returnType != TYP_UNKNOWN); |
152 | assert(!varTypeIsStruct(returnType)); |
153 | |
154 | info.compRetNativeType = returnType; |
155 | |
156 | // ToDo: Refactor this common code sequence into its own method as it is used 4+ times |
157 | if ((returnType == TYP_LONG) && (compLongUsed == false)) |
158 | { |
159 | compLongUsed = true; |
160 | } |
161 | else if (((returnType == TYP_FLOAT) || (returnType == TYP_DOUBLE)) && (compFloatingPointUsed == false)) |
162 | { |
163 | compFloatingPointUsed = true; |
164 | } |
165 | } |
166 | } |
167 | |
168 | // Do we have a RetBuffArg? |
169 | |
170 | if (hasRetBuffArg) |
171 | { |
172 | info.compArgsCount++; |
173 | } |
174 | else |
175 | { |
176 | info.compRetBuffArg = BAD_VAR_NUM; |
177 | } |
178 | |
179 | /* There is a 'hidden' cookie pushed last when the |
180 | calling convention is varargs */ |
181 | |
182 | if (info.compIsVarArgs) |
183 | { |
184 | info.compArgsCount++; |
185 | } |
186 | |
187 | // Is there an extra parameter used to pass instantiation info to |
188 | // shared generic methods and shared generic struct instance methods? |
189 | if (info.compMethodInfo->args.callConv & CORINFO_CALLCONV_PARAMTYPE) |
190 | { |
191 | info.compArgsCount++; |
192 | } |
193 | else |
194 | { |
195 | info.compTypeCtxtArg = BAD_VAR_NUM; |
196 | } |
197 | |
198 | lvaCount = info.compLocalsCount = info.compArgsCount + info.compMethodInfo->locals.numArgs; |
199 | |
200 | info.compILlocalsCount = info.compILargsCount + info.compMethodInfo->locals.numArgs; |
201 | |
202 | /* Now allocate the variable descriptor table */ |
203 | |
204 | if (compIsForInlining()) |
205 | { |
206 | lvaTable = impInlineInfo->InlinerCompiler->lvaTable; |
207 | lvaCount = impInlineInfo->InlinerCompiler->lvaCount; |
208 | lvaTableCnt = impInlineInfo->InlinerCompiler->lvaTableCnt; |
209 | |
210 | // No more stuff needs to be done. |
211 | return; |
212 | } |
213 | |
214 | lvaTableCnt = lvaCount * 2; |
215 | |
216 | if (lvaTableCnt < 16) |
217 | { |
218 | lvaTableCnt = 16; |
219 | } |
220 | |
221 | lvaTable = getAllocator(CMK_LvaTable).allocate<LclVarDsc>(lvaTableCnt); |
222 | size_t tableSize = lvaTableCnt * sizeof(*lvaTable); |
223 | memset(lvaTable, 0, tableSize); |
224 | for (unsigned i = 0; i < lvaTableCnt; i++) |
225 | { |
226 | new (&lvaTable[i], jitstd::placement_t()) LclVarDsc(); // call the constructor. |
227 | } |
228 | |
229 | //------------------------------------------------------------------------- |
230 | // Count the arguments and initialize the respective lvaTable[] entries |
231 | // |
232 | // First the implicit arguments |
233 | //------------------------------------------------------------------------- |
234 | |
235 | InitVarDscInfo varDscInfo; |
236 | varDscInfo.Init(lvaTable, hasRetBuffArg); |
237 | |
238 | lvaInitArgs(&varDscInfo); |
239 | |
240 | //------------------------------------------------------------------------- |
241 | // Finally the local variables |
242 | //------------------------------------------------------------------------- |
243 | |
244 | unsigned varNum = varDscInfo.varNum; |
245 | LclVarDsc* varDsc = varDscInfo.varDsc; |
246 | CORINFO_ARG_LIST_HANDLE localsSig = info.compMethodInfo->locals.args; |
247 | |
248 | for (unsigned i = 0; i < info.compMethodInfo->locals.numArgs; |
249 | i++, varNum++, varDsc++, localsSig = info.compCompHnd->getArgNext(localsSig)) |
250 | { |
251 | CORINFO_CLASS_HANDLE typeHnd; |
252 | CorInfoTypeWithMod corInfoTypeWithMod = |
253 | info.compCompHnd->getArgType(&info.compMethodInfo->locals, localsSig, &typeHnd); |
254 | CorInfoType corInfoType = strip(corInfoTypeWithMod); |
255 | |
256 | lvaInitVarDsc(varDsc, varNum, corInfoType, typeHnd, localsSig, &info.compMethodInfo->locals); |
257 | |
258 | varDsc->lvPinned = ((corInfoTypeWithMod & CORINFO_TYPE_MOD_PINNED) != 0); |
259 | varDsc->lvOnFrame = true; // The final home for this local variable might be our local stack frame |
260 | |
261 | if (corInfoType == CORINFO_TYPE_CLASS) |
262 | { |
263 | CORINFO_CLASS_HANDLE clsHnd = info.compCompHnd->getArgClass(&info.compMethodInfo->locals, localsSig); |
264 | lvaSetClass(varNum, clsHnd); |
265 | } |
266 | } |
267 | |
268 | if ( // If there already exist unsafe buffers, don't mark more structs as unsafe |
269 | // as that will cause them to be placed along with the real unsafe buffers, |
270 | // unnecessarily exposing them to overruns. This can affect GS tests which |
271 | // intentionally do buffer-overruns. |
272 | !getNeedsGSSecurityCookie() && |
273 | // GS checks require the stack to be re-ordered, which can't be done with EnC |
274 | !opts.compDbgEnC && compStressCompile(STRESS_UNSAFE_BUFFER_CHECKS, 25)) |
275 | { |
276 | setNeedsGSSecurityCookie(); |
277 | compGSReorderStackLayout = true; |
278 | |
279 | for (unsigned i = 0; i < lvaCount; i++) |
280 | { |
281 | if ((lvaTable[i].lvType == TYP_STRUCT) && compStressCompile(STRESS_GENERIC_VARN, 60)) |
282 | { |
283 | lvaTable[i].lvIsUnsafeBuffer = true; |
284 | } |
285 | } |
286 | } |
287 | |
288 | if (getNeedsGSSecurityCookie()) |
289 | { |
290 | // Ensure that there will be at least one stack variable since |
291 | // we require that the GSCookie does not have a 0 stack offset. |
292 | unsigned dummy = lvaGrabTempWithImplicitUse(false DEBUGARG("GSCookie dummy" )); |
293 | lvaTable[dummy].lvType = TYP_INT; |
294 | } |
295 | |
296 | // Allocate the lvaOutgoingArgSpaceVar now because we can run into problems in the |
297 | // emitter when the varNum is greater that 32767 (see emitLclVarAddr::initLclVarAddr) |
298 | lvaAllocOutgoingArgSpaceVar(); |
299 | |
300 | #ifdef DEBUG |
301 | if (verbose) |
302 | { |
303 | lvaTableDump(INITIAL_FRAME_LAYOUT); |
304 | } |
305 | #endif |
306 | } |
307 | |
308 | /*****************************************************************************/ |
309 | void Compiler::lvaInitArgs(InitVarDscInfo* varDscInfo) |
310 | { |
311 | compArgSize = 0; |
312 | |
313 | #if defined(_TARGET_ARM_) && defined(PROFILING_SUPPORTED) |
314 | // Prespill all argument regs on to stack in case of Arm when under profiler. |
315 | if (compIsProfilerHookNeeded()) |
316 | { |
317 | codeGen->regSet.rsMaskPreSpillRegArg |= RBM_ARG_REGS; |
318 | } |
319 | #endif |
320 | |
321 | //---------------------------------------------------------------------- |
322 | |
323 | /* Is there a "this" pointer ? */ |
324 | lvaInitThisPtr(varDscInfo); |
325 | |
326 | /* If we have a hidden return-buffer parameter, that comes here */ |
327 | lvaInitRetBuffArg(varDscInfo); |
328 | |
329 | //====================================================================== |
330 | |
331 | #if USER_ARGS_COME_LAST |
332 | //@GENERICS: final instantiation-info argument for shared generic methods |
333 | // and shared generic struct instance methods |
334 | lvaInitGenericsCtxt(varDscInfo); |
335 | |
336 | /* If the method is varargs, process the varargs cookie */ |
337 | lvaInitVarArgsHandle(varDscInfo); |
338 | #endif |
339 | |
340 | //------------------------------------------------------------------------- |
341 | // Now walk the function signature for the explicit user arguments |
342 | //------------------------------------------------------------------------- |
343 | lvaInitUserArgs(varDscInfo); |
344 | |
345 | #if !USER_ARGS_COME_LAST |
346 | //@GENERICS: final instantiation-info argument for shared generic methods |
347 | // and shared generic struct instance methods |
348 | lvaInitGenericsCtxt(varDscInfo); |
349 | |
350 | /* If the method is varargs, process the varargs cookie */ |
351 | lvaInitVarArgsHandle(varDscInfo); |
352 | #endif |
353 | |
354 | //---------------------------------------------------------------------- |
355 | |
356 | // We have set info.compArgsCount in compCompile() |
357 | noway_assert(varDscInfo->varNum == info.compArgsCount); |
358 | assert(varDscInfo->intRegArgNum <= MAX_REG_ARG); |
359 | |
360 | codeGen->intRegState.rsCalleeRegArgCount = varDscInfo->intRegArgNum; |
361 | codeGen->floatRegState.rsCalleeRegArgCount = varDscInfo->floatRegArgNum; |
362 | |
363 | #if FEATURE_FASTTAILCALL |
364 | // Save the stack usage information |
365 | // We can get register usage information using codeGen->intRegState and |
366 | // codeGen->floatRegState |
367 | info.compArgStackSize = varDscInfo->stackArgSize; |
368 | #endif // FEATURE_FASTTAILCALL |
369 | |
370 | // The total argument size must be aligned. |
371 | noway_assert((compArgSize % TARGET_POINTER_SIZE) == 0); |
372 | |
373 | #ifdef _TARGET_X86_ |
374 | /* We can not pass more than 2^16 dwords as arguments as the "ret" |
375 | instruction can only pop 2^16 arguments. Could be handled correctly |
376 | but it will be very difficult for fully interruptible code */ |
377 | |
378 | if (compArgSize != (size_t)(unsigned short)compArgSize) |
379 | NO_WAY("Too many arguments for the \"ret\" instruction to pop" ); |
380 | #endif |
381 | } |
382 | |
383 | /*****************************************************************************/ |
384 | void Compiler::lvaInitThisPtr(InitVarDscInfo* varDscInfo) |
385 | { |
386 | LclVarDsc* varDsc = varDscInfo->varDsc; |
387 | if (!info.compIsStatic) |
388 | { |
389 | varDsc->lvIsParam = 1; |
390 | varDsc->lvIsPtr = 1; |
391 | |
392 | lvaArg0Var = info.compThisArg = varDscInfo->varNum; |
393 | noway_assert(info.compThisArg == 0); |
394 | |
395 | if (eeIsValueClass(info.compClassHnd)) |
396 | { |
397 | varDsc->lvType = TYP_BYREF; |
398 | #ifdef FEATURE_SIMD |
399 | if (featureSIMD) |
400 | { |
401 | var_types simdBaseType = TYP_UNKNOWN; |
402 | var_types type = impNormStructType(info.compClassHnd, nullptr, nullptr, &simdBaseType); |
403 | if (simdBaseType != TYP_UNKNOWN) |
404 | { |
405 | assert(varTypeIsSIMD(type)); |
406 | varDsc->lvSIMDType = true; |
407 | varDsc->lvBaseType = simdBaseType; |
408 | varDsc->lvExactSize = genTypeSize(type); |
409 | } |
410 | } |
411 | #endif // FEATURE_SIMD |
412 | } |
413 | else |
414 | { |
415 | varDsc->lvType = TYP_REF; |
416 | lvaSetClass(varDscInfo->varNum, info.compClassHnd); |
417 | } |
418 | |
419 | if (tiVerificationNeeded) |
420 | { |
421 | varDsc->lvVerTypeInfo = verMakeTypeInfo(info.compClassHnd); |
422 | |
423 | if (varDsc->lvVerTypeInfo.IsValueClass()) |
424 | { |
425 | varDsc->lvVerTypeInfo.MakeByRef(); |
426 | } |
427 | } |
428 | else |
429 | { |
430 | varDsc->lvVerTypeInfo = typeInfo(); |
431 | } |
432 | |
433 | // Mark the 'this' pointer for the method |
434 | varDsc->lvVerTypeInfo.SetIsThisPtr(); |
435 | |
436 | varDsc->lvIsRegArg = 1; |
437 | noway_assert(varDscInfo->intRegArgNum == 0); |
438 | |
439 | varDsc->lvArgReg = genMapRegArgNumToRegNum(varDscInfo->allocRegArg(TYP_INT), varDsc->TypeGet()); |
440 | #if FEATURE_MULTIREG_ARGS |
441 | varDsc->lvOtherArgReg = REG_NA; |
442 | #endif |
443 | varDsc->lvOnFrame = true; // The final home for this incoming register might be our local stack frame |
444 | |
445 | #ifdef DEBUG |
446 | if (verbose) |
447 | { |
448 | printf("'this' passed in register %s\n" , getRegName(varDsc->lvArgReg)); |
449 | } |
450 | #endif |
451 | compArgSize += TARGET_POINTER_SIZE; |
452 | |
453 | varDscInfo->varNum++; |
454 | varDscInfo->varDsc++; |
455 | } |
456 | } |
457 | |
458 | /*****************************************************************************/ |
459 | void Compiler::lvaInitRetBuffArg(InitVarDscInfo* varDscInfo) |
460 | { |
461 | LclVarDsc* varDsc = varDscInfo->varDsc; |
462 | bool hasRetBuffArg = impMethodInfo_hasRetBuffArg(info.compMethodInfo); |
463 | |
464 | // These two should always match |
465 | noway_assert(hasRetBuffArg == varDscInfo->hasRetBufArg); |
466 | |
467 | if (hasRetBuffArg) |
468 | { |
469 | info.compRetBuffArg = varDscInfo->varNum; |
470 | varDsc->lvType = TYP_BYREF; |
471 | varDsc->lvIsParam = 1; |
472 | varDsc->lvIsRegArg = 1; |
473 | |
474 | if (hasFixedRetBuffReg()) |
475 | { |
476 | varDsc->lvArgReg = theFixedRetBuffReg(); |
477 | } |
478 | else |
479 | { |
480 | unsigned retBuffArgNum = varDscInfo->allocRegArg(TYP_INT); |
481 | varDsc->lvArgReg = genMapIntRegArgNumToRegNum(retBuffArgNum); |
482 | } |
483 | |
484 | #if FEATURE_MULTIREG_ARGS |
485 | varDsc->lvOtherArgReg = REG_NA; |
486 | #endif |
487 | varDsc->lvOnFrame = true; // The final home for this incoming register might be our local stack frame |
488 | |
489 | info.compRetBuffDefStack = 0; |
490 | if (info.compRetType == TYP_STRUCT) |
491 | { |
492 | CORINFO_SIG_INFO sigInfo; |
493 | info.compCompHnd->getMethodSig(info.compMethodHnd, &sigInfo); |
494 | assert(JITtype2varType(sigInfo.retType) == info.compRetType); // Else shouldn't have a ret buff. |
495 | |
496 | info.compRetBuffDefStack = |
497 | (info.compCompHnd->isStructRequiringStackAllocRetBuf(sigInfo.retTypeClass) == TRUE); |
498 | if (info.compRetBuffDefStack) |
499 | { |
500 | // If we're assured that the ret buff argument points into a callers stack, we will type it as |
501 | // "TYP_I_IMPL" |
502 | // (native int/unmanaged pointer) so that it's not tracked as a GC ref. |
503 | varDsc->lvType = TYP_I_IMPL; |
504 | } |
505 | } |
506 | #ifdef FEATURE_SIMD |
507 | else if (featureSIMD && varTypeIsSIMD(info.compRetType)) |
508 | { |
509 | varDsc->lvSIMDType = true; |
510 | varDsc->lvBaseType = |
511 | getBaseTypeAndSizeOfSIMDType(info.compMethodInfo->args.retTypeClass, &varDsc->lvExactSize); |
512 | assert(varDsc->lvBaseType != TYP_UNKNOWN); |
513 | } |
514 | #endif // FEATURE_SIMD |
515 | |
516 | assert(isValidIntArgReg(varDsc->lvArgReg)); |
517 | |
518 | #ifdef DEBUG |
519 | if (verbose) |
520 | { |
521 | printf("'__retBuf' passed in register %s\n" , getRegName(varDsc->lvArgReg)); |
522 | } |
523 | #endif |
524 | |
525 | /* Update the total argument size, count and varDsc */ |
526 | |
527 | compArgSize += TARGET_POINTER_SIZE; |
528 | varDscInfo->varNum++; |
529 | varDscInfo->varDsc++; |
530 | } |
531 | } |
532 | |
533 | /*****************************************************************************/ |
534 | void Compiler::lvaInitUserArgs(InitVarDscInfo* varDscInfo) |
535 | { |
536 | //------------------------------------------------------------------------- |
537 | // Walk the function signature for the explicit arguments |
538 | //------------------------------------------------------------------------- |
539 | |
540 | #if defined(_TARGET_X86_) |
541 | // Only (some of) the implicit args are enregistered for varargs |
542 | varDscInfo->maxIntRegArgNum = info.compIsVarArgs ? varDscInfo->intRegArgNum : MAX_REG_ARG; |
543 | #elif defined(_TARGET_AMD64_) && !defined(UNIX_AMD64_ABI) |
544 | // On System V type environment the float registers are not indexed together with the int ones. |
545 | varDscInfo->floatRegArgNum = varDscInfo->intRegArgNum; |
546 | #endif // _TARGET_* |
547 | |
548 | CORINFO_ARG_LIST_HANDLE argLst = info.compMethodInfo->args.args; |
549 | |
550 | const unsigned argSigLen = info.compMethodInfo->args.numArgs; |
551 | |
552 | regMaskTP doubleAlignMask = RBM_NONE; |
553 | for (unsigned i = 0; i < argSigLen; |
554 | i++, varDscInfo->varNum++, varDscInfo->varDsc++, argLst = info.compCompHnd->getArgNext(argLst)) |
555 | { |
556 | LclVarDsc* varDsc = varDscInfo->varDsc; |
557 | CORINFO_CLASS_HANDLE typeHnd = nullptr; |
558 | |
559 | CorInfoTypeWithMod corInfoType = info.compCompHnd->getArgType(&info.compMethodInfo->args, argLst, &typeHnd); |
560 | varDsc->lvIsParam = 1; |
561 | |
562 | lvaInitVarDsc(varDsc, varDscInfo->varNum, strip(corInfoType), typeHnd, argLst, &info.compMethodInfo->args); |
563 | |
564 | if (strip(corInfoType) == CORINFO_TYPE_CLASS) |
565 | { |
566 | CORINFO_CLASS_HANDLE clsHnd = info.compCompHnd->getArgClass(&info.compMethodInfo->args, argLst); |
567 | lvaSetClass(varDscInfo->varNum, clsHnd); |
568 | } |
569 | |
570 | // For ARM, ARM64, and AMD64 varargs, all arguments go in integer registers |
571 | var_types argType = mangleVarArgsType(varDsc->TypeGet()); |
572 | var_types origArgType = argType; |
573 | // ARM softfp calling convention should affect only the floating point arguments. |
574 | // Otherwise there appear too many surplus pre-spills and other memory operations |
575 | // with the associated locations . |
576 | bool isSoftFPPreSpill = opts.compUseSoftFP && varTypeIsFloating(varDsc->TypeGet()); |
577 | unsigned argSize = eeGetArgSize(argLst, &info.compMethodInfo->args); |
578 | unsigned cSlots = argSize / TARGET_POINTER_SIZE; // the total number of slots of this argument |
579 | bool isHfaArg = false; |
580 | var_types hfaType = TYP_UNDEF; |
581 | |
582 | #if defined(_TARGET_ARM64_) && defined(_TARGET_UNIX_) |
583 | // Native varargs on arm64 unix use the regular calling convention. |
584 | if (!opts.compUseSoftFP) |
585 | #else |
586 | // Methods that use VarArg or SoftFP cannot have HFA arguments |
587 | if (!info.compIsVarArgs && !opts.compUseSoftFP) |
588 | #endif // defined(_TARGET_ARM64_) && defined(_TARGET_UNIX_) |
589 | { |
590 | // If the argType is a struct, then check if it is an HFA |
591 | if (varTypeIsStruct(argType)) |
592 | { |
593 | hfaType = GetHfaType(typeHnd); // set to float or double if it is an HFA, otherwise TYP_UNDEF |
594 | isHfaArg = varTypeIsFloating(hfaType); |
595 | } |
596 | } |
597 | else if (info.compIsVarArgs) |
598 | { |
599 | #ifdef _TARGET_UNIX_ |
600 | // Currently native varargs is not implemented on non windows targets. |
601 | // |
602 | // Note that some targets like Arm64 Unix should not need much work as |
603 | // the ABI is the same. While other targets may only need small changes |
604 | // such as amd64 Unix, which just expects RAX to pass numFPArguments. |
605 | NYI("InitUserArgs for Vararg callee is not yet implemented on non Windows targets." ); |
606 | #endif |
607 | } |
608 | |
609 | if (isHfaArg) |
610 | { |
611 | // We have an HFA argument, so from here on out treat the type as a float or double. |
612 | // The orginal struct type is available by using origArgType |
613 | // We also update the cSlots to be the number of float/double fields in the HFA |
614 | argType = hfaType; |
615 | cSlots = varDsc->lvHfaSlots(); |
616 | } |
617 | // The number of slots that must be enregistered if we are to consider this argument enregistered. |
618 | // This is normally the same as cSlots, since we normally either enregister the entire object, |
619 | // or none of it. For structs on ARM, however, we only need to enregister a single slot to consider |
620 | // it enregistered, as long as we can split the rest onto the stack. |
621 | unsigned cSlotsToEnregister = cSlots; |
622 | |
623 | #if defined(_TARGET_ARM64_) && FEATURE_ARG_SPLIT |
624 | |
625 | // On arm64 Windows we will need to properly handle the case where a >8byte <=16byte |
626 | // struct is split between register r7 and virtual stack slot s[0] |
627 | // We will only do this for calls to vararg methods on Windows Arm64 |
628 | // |
629 | // !!This does not affect the normal arm64 calling convention or Unix Arm64!! |
630 | if (this->info.compIsVarArgs && argType == TYP_STRUCT) |
631 | { |
632 | if (varDscInfo->canEnreg(TYP_INT, 1) && // The beginning of the struct can go in a register |
633 | !varDscInfo->canEnreg(TYP_INT, cSlots)) // The end of the struct can't fit in a register |
634 | { |
635 | cSlotsToEnregister = 1; // Force the split |
636 | } |
637 | } |
638 | |
639 | #endif // defined(_TARGET_ARM64_) && FEATURE_ARG_SPLIT |
640 | |
641 | #ifdef _TARGET_ARM_ |
642 | // On ARM we pass the first 4 words of integer arguments and non-HFA structs in registers. |
643 | // But we pre-spill user arguments in varargs methods and structs. |
644 | // |
645 | unsigned cAlign; |
646 | bool preSpill = info.compIsVarArgs || isSoftFPPreSpill; |
647 | |
648 | switch (origArgType) |
649 | { |
650 | case TYP_STRUCT: |
651 | assert(varDsc->lvSize() == argSize); |
652 | cAlign = varDsc->lvStructDoubleAlign ? 2 : 1; |
653 | |
654 | // HFA arguments go on the stack frame. They don't get spilled in the prolog like struct |
655 | // arguments passed in the integer registers but get homed immediately after the prolog. |
656 | if (!isHfaArg) |
657 | { |
658 | // TODO-Arm32-Windows: vararg struct should be forced to split like |
659 | // ARM64 above. |
660 | cSlotsToEnregister = 1; // HFAs must be totally enregistered or not, but other structs can be split. |
661 | preSpill = true; |
662 | } |
663 | break; |
664 | |
665 | case TYP_DOUBLE: |
666 | case TYP_LONG: |
667 | cAlign = 2; |
668 | break; |
669 | |
670 | default: |
671 | cAlign = 1; |
672 | break; |
673 | } |
674 | |
675 | if (isRegParamType(argType)) |
676 | { |
677 | compArgSize += varDscInfo->alignReg(argType, cAlign) * REGSIZE_BYTES; |
678 | } |
679 | |
680 | if (argType == TYP_STRUCT) |
681 | { |
682 | // Are we going to split the struct between registers and stack? We can do that as long as |
683 | // no floating-point arguments have been put on the stack. |
684 | // |
685 | // From the ARM Procedure Call Standard: |
686 | // Rule C.5: "If the NCRN is less than r4 **and** the NSAA is equal to the SP," |
687 | // then split the argument between registers and stack. Implication: if something |
688 | // has already been spilled to the stack, then anything that would normally be |
689 | // split between the core registers and the stack will be put on the stack. |
690 | // Anything that follows will also be on the stack. However, if something from |
691 | // floating point regs has been spilled to the stack, we can still use r0-r3 until they are full. |
692 | |
693 | if (varDscInfo->canEnreg(TYP_INT, 1) && // The beginning of the struct can go in a register |
694 | !varDscInfo->canEnreg(TYP_INT, cSlots) && // The end of the struct can't fit in a register |
695 | varDscInfo->existAnyFloatStackArgs()) // There's at least one stack-based FP arg already |
696 | { |
697 | varDscInfo->setAllRegArgUsed(TYP_INT); // Prevent all future use of integer registers |
698 | preSpill = false; // This struct won't be prespilled, since it will go on the stack |
699 | } |
700 | } |
701 | |
702 | if (preSpill) |
703 | { |
704 | for (unsigned ix = 0; ix < cSlots; ix++) |
705 | { |
706 | if (!varDscInfo->canEnreg(TYP_INT, ix + 1)) |
707 | { |
708 | break; |
709 | } |
710 | regMaskTP regMask = genMapArgNumToRegMask(varDscInfo->regArgNum(TYP_INT) + ix, TYP_INT); |
711 | if (cAlign == 2) |
712 | { |
713 | doubleAlignMask |= regMask; |
714 | } |
715 | codeGen->regSet.rsMaskPreSpillRegArg |= regMask; |
716 | } |
717 | } |
718 | #else // !_TARGET_ARM_ |
719 | #if defined(UNIX_AMD64_ABI) |
720 | SYSTEMV_AMD64_CORINFO_STRUCT_REG_PASSING_DESCRIPTOR structDesc; |
721 | if (varTypeIsStruct(argType)) |
722 | { |
723 | assert(typeHnd != nullptr); |
724 | eeGetSystemVAmd64PassStructInRegisterDescriptor(typeHnd, &structDesc); |
725 | if (structDesc.passedInRegisters) |
726 | { |
727 | unsigned intRegCount = 0; |
728 | unsigned floatRegCount = 0; |
729 | |
730 | for (unsigned int i = 0; i < structDesc.eightByteCount; i++) |
731 | { |
732 | if (structDesc.IsIntegralSlot(i)) |
733 | { |
734 | intRegCount++; |
735 | } |
736 | else if (structDesc.IsSseSlot(i)) |
737 | { |
738 | floatRegCount++; |
739 | } |
740 | else |
741 | { |
742 | assert(false && "Invalid eightbyte classification type." ); |
743 | break; |
744 | } |
745 | } |
746 | |
747 | if (intRegCount != 0 && !varDscInfo->canEnreg(TYP_INT, intRegCount)) |
748 | { |
749 | structDesc.passedInRegisters = false; // No register to enregister the eightbytes. |
750 | } |
751 | |
752 | if (floatRegCount != 0 && !varDscInfo->canEnreg(TYP_FLOAT, floatRegCount)) |
753 | { |
754 | structDesc.passedInRegisters = false; // No register to enregister the eightbytes. |
755 | } |
756 | } |
757 | } |
758 | #endif // UNIX_AMD64_ABI |
759 | #endif // !_TARGET_ARM_ |
760 | |
761 | // The final home for this incoming register might be our local stack frame. |
762 | // For System V platforms the final home will always be on the local stack frame. |
763 | varDsc->lvOnFrame = true; |
764 | |
765 | bool canPassArgInRegisters = false; |
766 | |
767 | #if defined(UNIX_AMD64_ABI) |
768 | if (varTypeIsStruct(argType)) |
769 | { |
770 | canPassArgInRegisters = structDesc.passedInRegisters; |
771 | } |
772 | else |
773 | #endif // defined(UNIX_AMD64_ABI) |
774 | { |
775 | canPassArgInRegisters = varDscInfo->canEnreg(argType, cSlotsToEnregister); |
776 | } |
777 | |
778 | if (canPassArgInRegisters) |
779 | { |
780 | /* Another register argument */ |
781 | |
782 | // Allocate the registers we need. allocRegArg() returns the first argument register number of the set. |
783 | // For non-HFA structs, we still "try" to enregister the whole thing; it will just max out if splitting |
784 | // to the stack happens. |
785 | unsigned firstAllocatedRegArgNum = 0; |
786 | |
787 | #if FEATURE_MULTIREG_ARGS |
788 | varDsc->lvOtherArgReg = REG_NA; |
789 | #endif // FEATURE_MULTIREG_ARGS |
790 | |
791 | #if defined(UNIX_AMD64_ABI) |
792 | unsigned secondAllocatedRegArgNum = 0; |
793 | var_types firstEightByteType = TYP_UNDEF; |
794 | var_types secondEightByteType = TYP_UNDEF; |
795 | |
796 | if (varTypeIsStruct(argType)) |
797 | { |
798 | if (structDesc.eightByteCount >= 1) |
799 | { |
800 | firstEightByteType = GetEightByteType(structDesc, 0); |
801 | firstAllocatedRegArgNum = varDscInfo->allocRegArg(firstEightByteType, 1); |
802 | } |
803 | } |
804 | else |
805 | #endif // defined(UNIX_AMD64_ABI) |
806 | { |
807 | firstAllocatedRegArgNum = varDscInfo->allocRegArg(argType, cSlots); |
808 | } |
809 | |
810 | if (isHfaArg) |
811 | { |
812 | // We need to save the fact that this HFA is enregistered |
813 | varDsc->lvSetIsHfa(); |
814 | varDsc->lvSetIsHfaRegArg(); |
815 | varDsc->SetHfaType(hfaType); |
816 | varDsc->lvIsMultiRegArg = (varDsc->lvHfaSlots() > 1); |
817 | } |
818 | |
819 | varDsc->lvIsRegArg = 1; |
820 | |
821 | #if FEATURE_MULTIREG_ARGS |
822 | if (varTypeIsStruct(argType)) |
823 | { |
824 | #if defined(UNIX_AMD64_ABI) |
825 | varDsc->lvArgReg = genMapRegArgNumToRegNum(firstAllocatedRegArgNum, firstEightByteType); |
826 | |
827 | // If there is a second eightbyte, get a register for it too and map the arg to the reg number. |
828 | if (structDesc.eightByteCount >= 2) |
829 | { |
830 | secondEightByteType = GetEightByteType(structDesc, 1); |
831 | secondAllocatedRegArgNum = varDscInfo->allocRegArg(secondEightByteType, 1); |
832 | } |
833 | |
834 | if (secondEightByteType != TYP_UNDEF) |
835 | { |
836 | varDsc->lvOtherArgReg = genMapRegArgNumToRegNum(secondAllocatedRegArgNum, secondEightByteType); |
837 | } |
838 | #else // ARM32 or ARM64 |
839 | varDsc->lvArgReg = genMapRegArgNumToRegNum(firstAllocatedRegArgNum, TYP_I_IMPL); |
840 | #ifdef _TARGET_ARM64_ |
841 | if (cSlots == 2) |
842 | { |
843 | varDsc->lvOtherArgReg = genMapRegArgNumToRegNum(firstAllocatedRegArgNum + 1, TYP_I_IMPL); |
844 | } |
845 | #endif // _TARGET_ARM64_ |
846 | #endif // defined(UNIX_AMD64_ABI) |
847 | } |
848 | else |
849 | #endif // FEATURE_MULTIREG_ARGS |
850 | { |
851 | varDsc->lvArgReg = genMapRegArgNumToRegNum(firstAllocatedRegArgNum, argType); |
852 | } |
853 | |
854 | #ifdef _TARGET_ARM_ |
855 | if (varDsc->TypeGet() == TYP_LONG) |
856 | { |
857 | varDsc->lvOtherReg = genMapRegArgNumToRegNum(firstAllocatedRegArgNum + 1, TYP_INT); |
858 | } |
859 | #endif // _TARGET_ARM_ |
860 | |
861 | #ifdef DEBUG |
862 | if (verbose) |
863 | { |
864 | printf("Arg #%u passed in register(s) " , varDscInfo->varNum); |
865 | bool isFloat = false; |
866 | #if defined(UNIX_AMD64_ABI) |
867 | if (varTypeIsStruct(argType) && (structDesc.eightByteCount >= 1)) |
868 | { |
869 | isFloat = varTypeIsFloating(firstEightByteType); |
870 | } |
871 | else |
872 | #else |
873 | { |
874 | isFloat = varTypeIsFloating(argType); |
875 | } |
876 | #endif // !UNIX_AMD64_ABI |
877 | |
878 | #if defined(UNIX_AMD64_ABI) |
879 | if (varTypeIsStruct(argType)) |
880 | { |
881 | // Print both registers, just to be clear |
882 | if (firstEightByteType == TYP_UNDEF) |
883 | { |
884 | printf("firstEightByte: <not used>" ); |
885 | } |
886 | else |
887 | { |
888 | printf("firstEightByte: %s" , |
889 | getRegName(genMapRegArgNumToRegNum(firstAllocatedRegArgNum, firstEightByteType), |
890 | isFloat)); |
891 | } |
892 | |
893 | if (secondEightByteType == TYP_UNDEF) |
894 | { |
895 | printf(", secondEightByte: <not used>" ); |
896 | } |
897 | else |
898 | { |
899 | printf(", secondEightByte: %s" , |
900 | getRegName(genMapRegArgNumToRegNum(secondAllocatedRegArgNum, secondEightByteType), |
901 | varTypeIsFloating(secondEightByteType))); |
902 | } |
903 | } |
904 | else |
905 | #endif // defined(UNIX_AMD64_ABI) |
906 | { |
907 | isFloat = varTypeIsFloating(argType); |
908 | unsigned regArgNum = genMapRegNumToRegArgNum(varDsc->lvArgReg, argType); |
909 | |
910 | for (unsigned ix = 0; ix < cSlots; ix++, regArgNum++) |
911 | { |
912 | if (ix > 0) |
913 | { |
914 | printf("," ); |
915 | } |
916 | |
917 | if (!isFloat && (regArgNum >= varDscInfo->maxIntRegArgNum)) // a struct has been split between |
918 | // registers and stack |
919 | { |
920 | printf(" stack slots:%d" , cSlots - ix); |
921 | break; |
922 | } |
923 | |
924 | #ifdef _TARGET_ARM_ |
925 | if (isFloat) |
926 | { |
927 | // Print register size prefix |
928 | if (argType == TYP_DOUBLE) |
929 | { |
930 | // Print both registers, just to be clear |
931 | printf("%s/%s" , getRegName(genMapRegArgNumToRegNum(regArgNum, argType), isFloat), |
932 | getRegName(genMapRegArgNumToRegNum(regArgNum + 1, argType), isFloat)); |
933 | |
934 | // doubles take 2 slots |
935 | assert(ix + 1 < cSlots); |
936 | ++ix; |
937 | ++regArgNum; |
938 | } |
939 | else |
940 | { |
941 | printf("%s" , getRegName(genMapRegArgNumToRegNum(regArgNum, argType), isFloat)); |
942 | } |
943 | } |
944 | else |
945 | #endif // _TARGET_ARM_ |
946 | { |
947 | printf("%s" , getRegName(genMapRegArgNumToRegNum(regArgNum, argType), isFloat)); |
948 | } |
949 | } |
950 | } |
951 | printf("\n" ); |
952 | } |
953 | #endif // DEBUG |
954 | } // end if (canPassArgInRegisters) |
955 | else |
956 | { |
957 | #if defined(_TARGET_ARM_) |
958 | varDscInfo->setAllRegArgUsed(argType); |
959 | if (varTypeIsFloating(argType)) |
960 | { |
961 | varDscInfo->setAnyFloatStackArgs(); |
962 | } |
963 | |
964 | #elif defined(_TARGET_ARM64_) |
965 | |
966 | // If we needed to use the stack in order to pass this argument then |
967 | // record the fact that we have used up any remaining registers of this 'type' |
968 | // This prevents any 'backfilling' from occuring on ARM64 |
969 | // |
970 | varDscInfo->setAllRegArgUsed(argType); |
971 | |
972 | #endif // _TARGET_XXX_ |
973 | |
974 | #if FEATURE_FASTTAILCALL |
975 | varDscInfo->stackArgSize += roundUp(argSize, TARGET_POINTER_SIZE); |
976 | #endif // FEATURE_FASTTAILCALL |
977 | } |
978 | |
979 | #ifdef UNIX_AMD64_ABI |
980 | // The arg size is returning the number of bytes of the argument. For a struct it could return a size not a |
981 | // multiple of TARGET_POINTER_SIZE. The stack allocated space should always be multiple of TARGET_POINTER_SIZE, |
982 | // so round it up. |
983 | compArgSize += roundUp(argSize, TARGET_POINTER_SIZE); |
984 | #else // !UNIX_AMD64_ABI |
985 | compArgSize += argSize; |
986 | #endif // !UNIX_AMD64_ABI |
987 | if (info.compIsVarArgs || isHfaArg || isSoftFPPreSpill) |
988 | { |
989 | #if defined(_TARGET_X86_) |
990 | varDsc->lvStkOffs = compArgSize; |
991 | #else // !_TARGET_X86_ |
992 | // TODO-CQ: We shouldn't have to go as far as to declare these |
993 | // address-exposed -- DoNotEnregister should suffice. |
994 | lvaSetVarAddrExposed(varDscInfo->varNum); |
995 | #endif // !_TARGET_X86_ |
996 | } |
997 | } // for each user arg |
998 | |
999 | #ifdef _TARGET_ARM_ |
1000 | if (doubleAlignMask != RBM_NONE) |
1001 | { |
1002 | assert(RBM_ARG_REGS == 0xF); |
1003 | assert((doubleAlignMask & RBM_ARG_REGS) == doubleAlignMask); |
1004 | if (doubleAlignMask != RBM_NONE && doubleAlignMask != RBM_ARG_REGS) |
1005 | { |
1006 | // doubleAlignMask can only be 0011 and/or 1100 as 'double aligned types' can |
1007 | // begin at r0 or r2. |
1008 | assert(doubleAlignMask == 0x3 || doubleAlignMask == 0xC /* || 0xF is if'ed out */); |
1009 | |
1010 | // Now if doubleAlignMask is 0011 i.e., {r0,r1} and we prespill r2 or r3 |
1011 | // but not both, then the stack would be misaligned for r0. So spill both |
1012 | // r2 and r3. |
1013 | // |
1014 | // ; +0 --- caller SP double aligned ---- |
1015 | // ; -4 r2 r3 |
1016 | // ; -8 r1 r1 |
1017 | // ; -c r0 r0 <-- misaligned. |
1018 | // ; callee saved regs |
1019 | if (doubleAlignMask == 0x3 && doubleAlignMask != codeGen->regSet.rsMaskPreSpillRegArg) |
1020 | { |
1021 | codeGen->regSet.rsMaskPreSpillAlign = |
1022 | (~codeGen->regSet.rsMaskPreSpillRegArg & ~doubleAlignMask) & RBM_ARG_REGS; |
1023 | } |
1024 | } |
1025 | } |
1026 | #endif // _TARGET_ARM_ |
1027 | } |
1028 | |
1029 | /*****************************************************************************/ |
1030 | void Compiler::lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo) |
1031 | { |
1032 | //@GENERICS: final instantiation-info argument for shared generic methods |
1033 | // and shared generic struct instance methods |
1034 | if (info.compMethodInfo->args.callConv & CORINFO_CALLCONV_PARAMTYPE) |
1035 | { |
1036 | info.compTypeCtxtArg = varDscInfo->varNum; |
1037 | |
1038 | LclVarDsc* varDsc = varDscInfo->varDsc; |
1039 | varDsc->lvIsParam = 1; |
1040 | varDsc->lvType = TYP_I_IMPL; |
1041 | |
1042 | if (varDscInfo->canEnreg(TYP_I_IMPL)) |
1043 | { |
1044 | /* Another register argument */ |
1045 | |
1046 | varDsc->lvIsRegArg = 1; |
1047 | varDsc->lvArgReg = genMapRegArgNumToRegNum(varDscInfo->regArgNum(TYP_INT), varDsc->TypeGet()); |
1048 | #if FEATURE_MULTIREG_ARGS |
1049 | varDsc->lvOtherArgReg = REG_NA; |
1050 | #endif |
1051 | varDsc->lvOnFrame = true; // The final home for this incoming register might be our local stack frame |
1052 | |
1053 | varDscInfo->intRegArgNum++; |
1054 | |
1055 | #ifdef DEBUG |
1056 | if (verbose) |
1057 | { |
1058 | printf("'GenCtxt' passed in register %s\n" , getRegName(varDsc->lvArgReg)); |
1059 | } |
1060 | #endif |
1061 | } |
1062 | else |
1063 | { |
1064 | // We need to mark these as being on the stack, as this is not done elsewhere in the case that canEnreg |
1065 | // returns false. |
1066 | varDsc->lvOnFrame = true; |
1067 | #if FEATURE_FASTTAILCALL |
1068 | varDscInfo->stackArgSize += TARGET_POINTER_SIZE; |
1069 | #endif // FEATURE_FASTTAILCALL |
1070 | } |
1071 | |
1072 | compArgSize += TARGET_POINTER_SIZE; |
1073 | |
1074 | #if defined(_TARGET_X86_) |
1075 | if (info.compIsVarArgs) |
1076 | varDsc->lvStkOffs = compArgSize; |
1077 | #endif // _TARGET_X86_ |
1078 | |
1079 | varDscInfo->varNum++; |
1080 | varDscInfo->varDsc++; |
1081 | } |
1082 | } |
1083 | |
1084 | /*****************************************************************************/ |
1085 | void Compiler::lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo) |
1086 | { |
1087 | if (info.compIsVarArgs) |
1088 | { |
1089 | lvaVarargsHandleArg = varDscInfo->varNum; |
1090 | |
1091 | LclVarDsc* varDsc = varDscInfo->varDsc; |
1092 | varDsc->lvType = TYP_I_IMPL; |
1093 | varDsc->lvIsParam = 1; |
1094 | // Make sure this lives in the stack -- address may be reported to the VM. |
1095 | // TODO-CQ: This should probably be: |
1096 | // lvaSetVarDoNotEnregister(varDscInfo->varNum DEBUGARG(DNER_VMNeedsStackAddr)); |
1097 | // But that causes problems, so, for expedience, I switched back to this heavyweight |
1098 | // hammer. But I think it should be possible to switch; it may just work now |
1099 | // that other problems are fixed. |
1100 | lvaSetVarAddrExposed(varDscInfo->varNum); |
1101 | |
1102 | if (varDscInfo->canEnreg(TYP_I_IMPL)) |
1103 | { |
1104 | /* Another register argument */ |
1105 | |
1106 | unsigned varArgHndArgNum = varDscInfo->allocRegArg(TYP_I_IMPL); |
1107 | |
1108 | varDsc->lvIsRegArg = 1; |
1109 | varDsc->lvArgReg = genMapRegArgNumToRegNum(varArgHndArgNum, TYP_I_IMPL); |
1110 | #if FEATURE_MULTIREG_ARGS |
1111 | varDsc->lvOtherArgReg = REG_NA; |
1112 | #endif |
1113 | varDsc->lvOnFrame = true; // The final home for this incoming register might be our local stack frame |
1114 | #ifdef _TARGET_ARM_ |
1115 | // This has to be spilled right in front of the real arguments and we have |
1116 | // to pre-spill all the argument registers explicitly because we only have |
1117 | // have symbols for the declared ones, not any potential variadic ones. |
1118 | for (unsigned ix = varArgHndArgNum; ix < ArrLen(intArgMasks); ix++) |
1119 | { |
1120 | codeGen->regSet.rsMaskPreSpillRegArg |= intArgMasks[ix]; |
1121 | } |
1122 | #endif // _TARGET_ARM_ |
1123 | |
1124 | #ifdef DEBUG |
1125 | if (verbose) |
1126 | { |
1127 | printf("'VarArgHnd' passed in register %s\n" , getRegName(varDsc->lvArgReg)); |
1128 | } |
1129 | #endif // DEBUG |
1130 | } |
1131 | else |
1132 | { |
1133 | // We need to mark these as being on the stack, as this is not done elsewhere in the case that canEnreg |
1134 | // returns false. |
1135 | varDsc->lvOnFrame = true; |
1136 | #if FEATURE_FASTTAILCALL |
1137 | varDscInfo->stackArgSize += TARGET_POINTER_SIZE; |
1138 | #endif // FEATURE_FASTTAILCALL |
1139 | } |
1140 | |
1141 | /* Update the total argument size, count and varDsc */ |
1142 | |
1143 | compArgSize += TARGET_POINTER_SIZE; |
1144 | |
1145 | varDscInfo->varNum++; |
1146 | varDscInfo->varDsc++; |
1147 | |
1148 | #if defined(_TARGET_X86_) |
1149 | varDsc->lvStkOffs = compArgSize; |
1150 | |
1151 | // Allocate a temp to point at the beginning of the args |
1152 | |
1153 | lvaVarargsBaseOfStkArgs = lvaGrabTemp(false DEBUGARG("Varargs BaseOfStkArgs" )); |
1154 | lvaTable[lvaVarargsBaseOfStkArgs].lvType = TYP_I_IMPL; |
1155 | |
1156 | #endif // _TARGET_X86_ |
1157 | } |
1158 | } |
1159 | |
1160 | /*****************************************************************************/ |
1161 | void Compiler::lvaInitVarDsc(LclVarDsc* varDsc, |
1162 | unsigned varNum, |
1163 | CorInfoType corInfoType, |
1164 | CORINFO_CLASS_HANDLE typeHnd, |
1165 | CORINFO_ARG_LIST_HANDLE varList, |
1166 | CORINFO_SIG_INFO* varSig) |
1167 | { |
1168 | noway_assert(varDsc == &lvaTable[varNum]); |
1169 | |
1170 | switch (corInfoType) |
1171 | { |
1172 | // Mark types that looks like a pointer for doing shadow-copying of |
1173 | // parameters if we have an unsafe buffer. |
1174 | // Note that this does not handle structs with pointer fields. Instead, |
1175 | // we rely on using the assign-groups/equivalence-groups in |
1176 | // gsFindVulnerableParams() to determine if a buffer-struct contains a |
1177 | // pointer. We could do better by having the EE determine this for us. |
1178 | // Note that we want to keep buffers without pointers at lower memory |
1179 | // addresses than buffers with pointers. |
1180 | case CORINFO_TYPE_PTR: |
1181 | case CORINFO_TYPE_BYREF: |
1182 | case CORINFO_TYPE_CLASS: |
1183 | case CORINFO_TYPE_STRING: |
1184 | case CORINFO_TYPE_VAR: |
1185 | case CORINFO_TYPE_REFANY: |
1186 | varDsc->lvIsPtr = 1; |
1187 | break; |
1188 | default: |
1189 | break; |
1190 | } |
1191 | |
1192 | var_types type = JITtype2varType(corInfoType); |
1193 | if (varTypeIsFloating(type)) |
1194 | { |
1195 | compFloatingPointUsed = true; |
1196 | } |
1197 | |
1198 | if (tiVerificationNeeded) |
1199 | { |
1200 | varDsc->lvVerTypeInfo = verParseArgSigToTypeInfo(varSig, varList); |
1201 | } |
1202 | |
1203 | if (tiVerificationNeeded) |
1204 | { |
1205 | if (varDsc->lvIsParam) |
1206 | { |
1207 | // For an incoming ValueType we better be able to have the full type information |
1208 | // so that we can layout the parameter offsets correctly |
1209 | |
1210 | if (varTypeIsStruct(type) && varDsc->lvVerTypeInfo.IsDead()) |
1211 | { |
1212 | BADCODE("invalid ValueType parameter" ); |
1213 | } |
1214 | |
1215 | // For an incoming reference type we need to verify that the actual type is |
1216 | // a reference type and not a valuetype. |
1217 | |
1218 | if (type == TYP_REF && |
1219 | !(varDsc->lvVerTypeInfo.IsType(TI_REF) || varDsc->lvVerTypeInfo.IsUnboxedGenericTypeVar())) |
1220 | { |
1221 | BADCODE("parameter type mismatch" ); |
1222 | } |
1223 | } |
1224 | |
1225 | // Disallow byrefs to byref like objects (ArgTypeHandle) |
1226 | // techncally we could get away with just not setting them |
1227 | if (varDsc->lvVerTypeInfo.IsByRef() && verIsByRefLike(DereferenceByRef(varDsc->lvVerTypeInfo))) |
1228 | { |
1229 | varDsc->lvVerTypeInfo = typeInfo(); |
1230 | } |
1231 | |
1232 | // we don't want the EE to assert in lvaSetStruct on bad sigs, so change |
1233 | // the JIT type to avoid even trying to call back |
1234 | if (varTypeIsStruct(type) && varDsc->lvVerTypeInfo.IsDead()) |
1235 | { |
1236 | type = TYP_VOID; |
1237 | } |
1238 | } |
1239 | |
1240 | if (typeHnd) |
1241 | { |
1242 | unsigned cFlags = info.compCompHnd->getClassAttribs(typeHnd); |
1243 | |
1244 | // We can get typeHnds for primitive types, these are value types which only contain |
1245 | // a primitive. We will need the typeHnd to distinguish them, so we store it here. |
1246 | if ((cFlags & CORINFO_FLG_VALUECLASS) && !varTypeIsStruct(type)) |
1247 | { |
1248 | if (tiVerificationNeeded == false) |
1249 | { |
1250 | // printf("This is a struct that the JIT will treat as a primitive\n"); |
1251 | varDsc->lvVerTypeInfo = verMakeTypeInfo(typeHnd); |
1252 | } |
1253 | } |
1254 | |
1255 | varDsc->lvOverlappingFields = StructHasOverlappingFields(cFlags); |
1256 | } |
1257 | |
1258 | if (varTypeIsGC(type)) |
1259 | { |
1260 | varDsc->lvStructGcCount = 1; |
1261 | } |
1262 | |
1263 | // Set the lvType (before this point it is TYP_UNDEF). |
1264 | if ((varTypeIsStruct(type))) |
1265 | { |
1266 | lvaSetStruct(varNum, typeHnd, typeHnd != nullptr, !tiVerificationNeeded); |
1267 | if (info.compIsVarArgs) |
1268 | { |
1269 | lvaSetStructUsedAsVarArg(varNum); |
1270 | } |
1271 | } |
1272 | else |
1273 | { |
1274 | varDsc->lvType = type; |
1275 | } |
1276 | |
1277 | #if OPT_BOOL_OPS |
1278 | if (type == TYP_BOOL) |
1279 | { |
1280 | varDsc->lvIsBoolean = true; |
1281 | } |
1282 | #endif |
1283 | |
1284 | #ifdef DEBUG |
1285 | varDsc->lvStkOffs = BAD_STK_OFFS; |
1286 | #endif |
1287 | |
1288 | #if FEATURE_MULTIREG_ARGS |
1289 | varDsc->lvOtherArgReg = REG_NA; |
1290 | #endif // FEATURE_MULTIREG_ARGS |
1291 | } |
1292 | |
1293 | /***************************************************************************** |
1294 | * Returns our internal varNum for a given IL variable. |
1295 | * Asserts assume it is called after lvaTable[] has been set up. |
1296 | */ |
1297 | |
1298 | unsigned Compiler::compMapILvarNum(unsigned ILvarNum) |
1299 | { |
1300 | noway_assert(ILvarNum < info.compILlocalsCount || ILvarNum > unsigned(ICorDebugInfo::UNKNOWN_ILNUM)); |
1301 | |
1302 | unsigned varNum; |
1303 | |
1304 | if (ILvarNum == (unsigned)ICorDebugInfo::VARARGS_HND_ILNUM) |
1305 | { |
1306 | // The varargs cookie is the last argument in lvaTable[] |
1307 | noway_assert(info.compIsVarArgs); |
1308 | |
1309 | varNum = lvaVarargsHandleArg; |
1310 | noway_assert(lvaTable[varNum].lvIsParam); |
1311 | } |
1312 | else if (ILvarNum == (unsigned)ICorDebugInfo::RETBUF_ILNUM) |
1313 | { |
1314 | noway_assert(info.compRetBuffArg != BAD_VAR_NUM); |
1315 | varNum = info.compRetBuffArg; |
1316 | } |
1317 | else if (ILvarNum == (unsigned)ICorDebugInfo::TYPECTXT_ILNUM) |
1318 | { |
1319 | noway_assert(info.compTypeCtxtArg >= 0); |
1320 | varNum = unsigned(info.compTypeCtxtArg); |
1321 | } |
1322 | else if (ILvarNum < info.compILargsCount) |
1323 | { |
1324 | // Parameter |
1325 | varNum = compMapILargNum(ILvarNum); |
1326 | noway_assert(lvaTable[varNum].lvIsParam); |
1327 | } |
1328 | else if (ILvarNum < info.compILlocalsCount) |
1329 | { |
1330 | // Local variable |
1331 | unsigned lclNum = ILvarNum - info.compILargsCount; |
1332 | varNum = info.compArgsCount + lclNum; |
1333 | noway_assert(!lvaTable[varNum].lvIsParam); |
1334 | } |
1335 | else |
1336 | { |
1337 | unreached(); |
1338 | } |
1339 | |
1340 | noway_assert(varNum < info.compLocalsCount); |
1341 | return varNum; |
1342 | } |
1343 | |
1344 | /***************************************************************************** |
1345 | * Returns the IL variable number given our internal varNum. |
1346 | * Special return values are VARG_ILNUM, RETBUF_ILNUM, TYPECTXT_ILNUM. |
1347 | * |
1348 | * Returns UNKNOWN_ILNUM if it can't be mapped. |
1349 | */ |
1350 | |
1351 | unsigned Compiler::compMap2ILvarNum(unsigned varNum) |
1352 | { |
1353 | if (compIsForInlining()) |
1354 | { |
1355 | return impInlineInfo->InlinerCompiler->compMap2ILvarNum(varNum); |
1356 | } |
1357 | |
1358 | noway_assert(varNum < lvaCount); |
1359 | |
1360 | if (varNum == info.compRetBuffArg) |
1361 | { |
1362 | return (unsigned)ICorDebugInfo::RETBUF_ILNUM; |
1363 | } |
1364 | |
1365 | // Is this a varargs function? |
1366 | if (info.compIsVarArgs && varNum == lvaVarargsHandleArg) |
1367 | { |
1368 | return (unsigned)ICorDebugInfo::VARARGS_HND_ILNUM; |
1369 | } |
1370 | |
1371 | // We create an extra argument for the type context parameter |
1372 | // needed for shared generic code. |
1373 | if ((info.compMethodInfo->args.callConv & CORINFO_CALLCONV_PARAMTYPE) && varNum == (unsigned)info.compTypeCtxtArg) |
1374 | { |
1375 | return (unsigned)ICorDebugInfo::TYPECTXT_ILNUM; |
1376 | } |
1377 | |
1378 | #if FEATURE_FIXED_OUT_ARGS |
1379 | if (varNum == lvaOutgoingArgSpaceVar) |
1380 | { |
1381 | return (unsigned)ICorDebugInfo::UNKNOWN_ILNUM; // Cannot be mapped |
1382 | } |
1383 | #endif // FEATURE_FIXED_OUT_ARGS |
1384 | |
1385 | // Now mutate varNum to remove extra parameters from the count. |
1386 | if ((info.compMethodInfo->args.callConv & CORINFO_CALLCONV_PARAMTYPE) && varNum > (unsigned)info.compTypeCtxtArg) |
1387 | { |
1388 | varNum--; |
1389 | } |
1390 | |
1391 | if (info.compIsVarArgs && varNum > lvaVarargsHandleArg) |
1392 | { |
1393 | varNum--; |
1394 | } |
1395 | |
1396 | /* Is there a hidden argument for the return buffer. |
1397 | Note that this code works because if the RetBuffArg is not present, |
1398 | compRetBuffArg will be BAD_VAR_NUM */ |
1399 | if (info.compRetBuffArg != BAD_VAR_NUM && varNum > info.compRetBuffArg) |
1400 | { |
1401 | varNum--; |
1402 | } |
1403 | |
1404 | if (varNum >= info.compLocalsCount) |
1405 | { |
1406 | return (unsigned)ICorDebugInfo::UNKNOWN_ILNUM; // Cannot be mapped |
1407 | } |
1408 | |
1409 | return varNum; |
1410 | } |
1411 | |
1412 | /***************************************************************************** |
1413 | * Returns true if variable "varNum" may be address-exposed. |
1414 | */ |
1415 | |
1416 | bool Compiler::lvaVarAddrExposed(unsigned varNum) |
1417 | { |
1418 | noway_assert(varNum < lvaCount); |
1419 | LclVarDsc* varDsc = &lvaTable[varNum]; |
1420 | |
1421 | return varDsc->lvAddrExposed; |
1422 | } |
1423 | |
1424 | /***************************************************************************** |
1425 | * Returns true iff variable "varNum" should not be enregistered (or one of several reasons). |
1426 | */ |
1427 | |
1428 | bool Compiler::lvaVarDoNotEnregister(unsigned varNum) |
1429 | { |
1430 | noway_assert(varNum < lvaCount); |
1431 | LclVarDsc* varDsc = &lvaTable[varNum]; |
1432 | |
1433 | return varDsc->lvDoNotEnregister; |
1434 | } |
1435 | |
1436 | /***************************************************************************** |
1437 | * Returns the handle to the class of the local variable varNum |
1438 | */ |
1439 | |
1440 | CORINFO_CLASS_HANDLE Compiler::lvaGetStruct(unsigned varNum) |
1441 | { |
1442 | noway_assert(varNum < lvaCount); |
1443 | LclVarDsc* varDsc = &lvaTable[varNum]; |
1444 | |
1445 | return varDsc->lvVerTypeInfo.GetClassHandleForValueClass(); |
1446 | } |
1447 | |
1448 | //-------------------------------------------------------------------------------------------- |
1449 | // lvaFieldOffsetCmp - a static compare function passed to qsort() by Compiler::StructPromotionHelper; |
1450 | // compares fields' offsets. |
1451 | // |
1452 | // Arguments: |
1453 | // field1 - pointer to the first field; |
1454 | // field2 - pointer to the second field. |
1455 | // |
1456 | // Return value: |
1457 | // 0 if the fields' offsets are equal, 1 if the first field has bigger offset, -1 otherwise. |
1458 | // |
1459 | int __cdecl Compiler::lvaFieldOffsetCmp(const void* field1, const void* field2) |
1460 | { |
1461 | lvaStructFieldInfo* pFieldInfo1 = (lvaStructFieldInfo*)field1; |
1462 | lvaStructFieldInfo* pFieldInfo2 = (lvaStructFieldInfo*)field2; |
1463 | |
1464 | if (pFieldInfo1->fldOffset == pFieldInfo2->fldOffset) |
1465 | { |
1466 | return 0; |
1467 | } |
1468 | else |
1469 | { |
1470 | return (pFieldInfo1->fldOffset > pFieldInfo2->fldOffset) ? +1 : -1; |
1471 | } |
1472 | } |
1473 | |
1474 | //------------------------------------------------------------------------ |
1475 | // StructPromotionHelper constructor. |
1476 | // |
1477 | // Arguments: |
1478 | // compiler - pointer to a compiler to get access to an allocator, compHandle etc. |
1479 | // |
1480 | Compiler::StructPromotionHelper::StructPromotionHelper(Compiler* compiler) |
1481 | : compiler(compiler) |
1482 | , structPromotionInfo() |
1483 | #ifdef _TARGET_ARM_ |
1484 | , requiresScratchVar(false) |
1485 | #endif // _TARGET_ARM_ |
1486 | #ifdef DEBUG |
1487 | , retypedFieldsMap(compiler->getAllocator(CMK_DebugOnly)) |
1488 | #endif // DEBUG |
1489 | { |
1490 | } |
1491 | |
1492 | #ifdef _TARGET_ARM_ |
1493 | //-------------------------------------------------------------------------------------------- |
1494 | // GetRequiresScratchVar - do we need a stack area to assemble small fields in order to place them in a register. |
1495 | // |
1496 | // Return value: |
1497 | // true if there was a small promoted variable and scratch var is required . |
1498 | // |
1499 | bool Compiler::StructPromotionHelper::GetRequiresScratchVar() |
1500 | { |
1501 | return requiresScratchVar; |
1502 | } |
1503 | |
1504 | #endif // _TARGET_ARM_ |
1505 | |
1506 | //-------------------------------------------------------------------------------------------- |
1507 | // TryPromoteStructVar - promote struct var if it is possible and profitable. |
1508 | // |
1509 | // Arguments: |
1510 | // lclNum - struct number to try. |
1511 | // |
1512 | // Return value: |
1513 | // true if the struct var was promoted. |
1514 | // |
1515 | bool Compiler::StructPromotionHelper::TryPromoteStructVar(unsigned lclNum) |
1516 | { |
1517 | if (CanPromoteStructVar(lclNum)) |
1518 | { |
1519 | #if 0 |
1520 | // Often-useful debugging code: if you've narrowed down a struct-promotion problem to a single |
1521 | // method, this allows you to select a subset of the vars to promote (by 1-based ordinal number). |
1522 | static int structPromoVarNum = 0; |
1523 | structPromoVarNum++; |
1524 | if (atoi(getenv("structpromovarnumlo" )) <= structPromoVarNum && structPromoVarNum <= atoi(getenv("structpromovarnumhi" ))) |
1525 | #endif // 0 |
1526 | if (ShouldPromoteStructVar(lclNum)) |
1527 | { |
1528 | PromoteStructVar(lclNum); |
1529 | return true; |
1530 | } |
1531 | } |
1532 | return false; |
1533 | } |
1534 | |
1535 | #ifdef DEBUG |
1536 | //-------------------------------------------------------------------------------------------- |
1537 | // CheckRetypedAsScalar - check that the fldType for this fieldHnd was retyped as requested type. |
1538 | // |
1539 | // Arguments: |
1540 | // fieldHnd - the field handle; |
1541 | // requestedType - as which type the field was accessed; |
1542 | // |
1543 | // Notes: |
1544 | // For example it can happen when such struct A { struct B { long c } } is compiled and we access A.B.c, |
1545 | // it could look like "GT_FIELD struct B.c -> ADDR -> GT_FIELD struct A.B -> ADDR -> LCL_VAR A" , but |
1546 | // "GT_FIELD struct A.B -> ADDR -> LCL_VAR A" can be promoted to "LCL_VAR long A.B" and then |
1547 | // there is type mistmatch between "GT_FIELD struct B.c" and "LCL_VAR long A.B". |
1548 | // |
1549 | void Compiler::StructPromotionHelper::CheckRetypedAsScalar(CORINFO_FIELD_HANDLE fieldHnd, var_types requestedType) |
1550 | { |
1551 | assert(retypedFieldsMap.Lookup(fieldHnd)); |
1552 | assert(retypedFieldsMap[fieldHnd] == requestedType); |
1553 | } |
1554 | #endif // DEBUG |
1555 | |
1556 | //-------------------------------------------------------------------------------------------- |
1557 | // CanPromoteStructType - checks if the struct type can be promoted. |
1558 | // |
1559 | // Arguments: |
1560 | // typeHnd - struct handle to check. |
1561 | // |
1562 | // Return value: |
1563 | // true if the struct type can be promoted. |
1564 | // |
1565 | // Notes: |
1566 | // The last analyzed type is memorized to skip the check if we ask about the same time again next. |
1567 | // However, it was not found profitable to memorize all analyzed types in a map. |
1568 | // |
1569 | // The check initializes only nessasary fields in lvaStructPromotionInfo, |
1570 | // so if the promotion is rejected early than most fields will be uninitialized. |
1571 | // |
1572 | bool Compiler::StructPromotionHelper::CanPromoteStructType(CORINFO_CLASS_HANDLE typeHnd) |
1573 | { |
1574 | if (!compiler->eeIsValueClass(typeHnd)) |
1575 | { |
1576 | // TODO-ObjectStackAllocation: Enable promotion of fields of stack-allocated objects. |
1577 | return false; |
1578 | } |
1579 | |
1580 | if (structPromotionInfo.typeHnd == typeHnd) |
1581 | { |
1582 | // Asking for the same type of struct as the last time. |
1583 | // Nothing need to be done. |
1584 | // Fall through ... |
1585 | return structPromotionInfo.canPromote; |
1586 | } |
1587 | |
1588 | // Analyze this type from scratch. |
1589 | structPromotionInfo = lvaStructPromotionInfo(typeHnd); |
1590 | |
1591 | // sizeof(double) represents the size of the largest primitive type that we can struct promote. |
1592 | // In the future this may be changing to XMM_REGSIZE_BYTES. |
1593 | // Note: MaxOffset is used below to declare a local array, and therefore must be a compile-time constant. |
1594 | CLANG_FORMAT_COMMENT_ANCHOR; |
1595 | #if defined(FEATURE_SIMD) |
1596 | #if defined(_TARGET_XARCH_) |
1597 | // This will allow promotion of 4 Vector<T> fields on AVX2 or Vector256<T> on AVX, |
1598 | // or 8 Vector<T>/Vector128<T> fields on SSE2. |
1599 | const int MaxOffset = MAX_NumOfFieldsInPromotableStruct * YMM_REGSIZE_BYTES; |
1600 | #elif defined(_TARGET_ARM64_) |
1601 | const int MaxOffset = MAX_NumOfFieldsInPromotableStruct * FP_REGSIZE_BYTES; |
1602 | #endif // defined(_TARGET_XARCH_) || defined(_TARGET_ARM64_) |
1603 | #else // !FEATURE_SIMD |
1604 | const int MaxOffset = MAX_NumOfFieldsInPromotableStruct * sizeof(double); |
1605 | #endif // !FEATURE_SIMD |
1606 | |
1607 | assert((BYTE)MaxOffset == MaxOffset); // because lvaStructFieldInfo.fldOffset is byte-sized |
1608 | assert((BYTE)MAX_NumOfFieldsInPromotableStruct == |
1609 | MAX_NumOfFieldsInPromotableStruct); // because lvaStructFieldInfo.fieldCnt is byte-sized |
1610 | |
1611 | bool containsGCpointers = false; |
1612 | |
1613 | COMP_HANDLE compHandle = compiler->info.compCompHnd; |
1614 | |
1615 | unsigned structSize = compHandle->getClassSize(typeHnd); |
1616 | if (structSize > MaxOffset) |
1617 | { |
1618 | return false; // struct is too large |
1619 | } |
1620 | |
1621 | unsigned fieldCnt = compHandle->getClassNumInstanceFields(typeHnd); |
1622 | if (fieldCnt == 0 || fieldCnt > MAX_NumOfFieldsInPromotableStruct) |
1623 | { |
1624 | return false; // struct must have between 1 and MAX_NumOfFieldsInPromotableStruct fields |
1625 | } |
1626 | |
1627 | structPromotionInfo.fieldCnt = (unsigned char)fieldCnt; |
1628 | DWORD typeFlags = compHandle->getClassAttribs(typeHnd); |
1629 | |
1630 | bool overlappingFields = StructHasOverlappingFields(typeFlags); |
1631 | if (overlappingFields) |
1632 | { |
1633 | return false; |
1634 | } |
1635 | |
1636 | // Don't struct promote if we have an CUSTOMLAYOUT flag on an HFA type |
1637 | if (StructHasCustomLayout(typeFlags) && compiler->IsHfa(typeHnd)) |
1638 | { |
1639 | return false; |
1640 | } |
1641 | |
1642 | #ifdef _TARGET_ARM_ |
1643 | // On ARM, we have a requirement on the struct alignment; see below. |
1644 | unsigned structAlignment = roundUp(compHandle->getClassAlignmentRequirement(typeHnd), TARGET_POINTER_SIZE); |
1645 | #endif // _TARGET_ARM_ |
1646 | |
1647 | unsigned fieldsSize = 0; |
1648 | |
1649 | for (BYTE ordinal = 0; ordinal < fieldCnt; ++ordinal) |
1650 | { |
1651 | CORINFO_FIELD_HANDLE fieldHnd = compHandle->getFieldInClass(typeHnd, ordinal); |
1652 | structPromotionInfo.fields[ordinal] = GetFieldInfo(fieldHnd, ordinal); |
1653 | const lvaStructFieldInfo& fieldInfo = structPromotionInfo.fields[ordinal]; |
1654 | |
1655 | noway_assert(fieldInfo.fldOffset < structSize); |
1656 | |
1657 | if (fieldInfo.fldSize == 0) |
1658 | { |
1659 | // Not a scalar type. |
1660 | return false; |
1661 | } |
1662 | |
1663 | if ((fieldInfo.fldOffset % fieldInfo.fldSize) != 0) |
1664 | { |
1665 | // The code in Compiler::genPushArgList that reconstitutes |
1666 | // struct values on the stack from promoted fields expects |
1667 | // those fields to be at their natural alignment. |
1668 | return false; |
1669 | } |
1670 | |
1671 | if (varTypeIsGC(fieldInfo.fldType)) |
1672 | { |
1673 | containsGCpointers = true; |
1674 | } |
1675 | |
1676 | // The end offset for this field should never be larger than our structSize. |
1677 | noway_assert(fieldInfo.fldOffset + fieldInfo.fldSize <= structSize); |
1678 | |
1679 | fieldsSize += fieldInfo.fldSize; |
1680 | |
1681 | #ifdef _TARGET_ARM_ |
1682 | // On ARM, for struct types that don't use explicit layout, the alignment of the struct is |
1683 | // at least the max alignment of its fields. We take advantage of this invariant in struct promotion, |
1684 | // so verify it here. |
1685 | if (fieldInfo.fldSize > structAlignment) |
1686 | { |
1687 | // Don't promote vars whose struct types violates the invariant. (Alignment == size for primitives.) |
1688 | return false; |
1689 | } |
1690 | // If we have any small fields we will allocate a single PromotedStructScratch local var for the method. |
1691 | // This is a stack area that we use to assemble the small fields in order to place them in a register |
1692 | // argument. |
1693 | // |
1694 | if (fieldInfo.fldSize < TARGET_POINTER_SIZE) |
1695 | { |
1696 | requiresScratchVar = true; |
1697 | } |
1698 | #endif // _TARGET_ARM_ |
1699 | } |
1700 | |
1701 | // If we saw any GC pointer or by-ref fields above then CORINFO_FLG_CONTAINS_GC_PTR or |
1702 | // CORINFO_FLG_CONTAINS_STACK_PTR has to be set! |
1703 | noway_assert((containsGCpointers == false) || |
1704 | ((typeFlags & (CORINFO_FLG_CONTAINS_GC_PTR | CORINFO_FLG_CONTAINS_STACK_PTR)) != 0)); |
1705 | |
1706 | // If we have "Custom Layout" then we might have an explicit Size attribute |
1707 | // Managed C++ uses this for its structs, such C++ types will not contain GC pointers. |
1708 | // |
1709 | // The current VM implementation also incorrectly sets the CORINFO_FLG_CUSTOMLAYOUT |
1710 | // whenever a managed value class contains any GC pointers. |
1711 | // (See the comment for VMFLAG_NOT_TIGHTLY_PACKED in class.h) |
1712 | // |
1713 | // It is important to struct promote managed value classes that have GC pointers |
1714 | // So we compute the correct value for "CustomLayout" here |
1715 | // |
1716 | if (StructHasCustomLayout(typeFlags) && ((typeFlags & CORINFO_FLG_CONTAINS_GC_PTR) == 0)) |
1717 | { |
1718 | structPromotionInfo.customLayout = true; |
1719 | } |
1720 | |
1721 | // Check if this promoted struct contains any holes. |
1722 | assert(!overlappingFields); |
1723 | if (fieldsSize != structSize) |
1724 | { |
1725 | // If sizes do not match it means we have an overlapping fields or holes. |
1726 | // Overlapping fields were rejected early, so here it can mean only holes. |
1727 | structPromotionInfo.containsHoles = true; |
1728 | } |
1729 | |
1730 | // Cool, this struct is promotable. |
1731 | |
1732 | structPromotionInfo.canPromote = true; |
1733 | return true; |
1734 | } |
1735 | |
1736 | //-------------------------------------------------------------------------------------------- |
1737 | // CanPromoteStructVar - checks if the struct can be promoted. |
1738 | // |
1739 | // Arguments: |
1740 | // lclNum - struct number to check. |
1741 | // |
1742 | // Return value: |
1743 | // true if the struct var can be promoted. |
1744 | // |
1745 | bool Compiler::StructPromotionHelper::CanPromoteStructVar(unsigned lclNum) |
1746 | { |
1747 | LclVarDsc* varDsc = compiler->lvaGetDesc(lclNum); |
1748 | |
1749 | assert(varTypeIsStruct(varDsc)); |
1750 | assert(!varDsc->lvPromoted); // Don't ask again :) |
1751 | |
1752 | // If this lclVar is used in a SIMD intrinsic, then we don't want to struct promote it. |
1753 | // Note, however, that SIMD lclVars that are NOT used in a SIMD intrinsic may be |
1754 | // profitably promoted. |
1755 | if (varDsc->lvIsUsedInSIMDIntrinsic()) |
1756 | { |
1757 | JITDUMP(" struct promotion of V%02u is disabled because lvIsUsedInSIMDIntrinsic()\n" , lclNum); |
1758 | return false; |
1759 | } |
1760 | |
1761 | // Reject struct promotion of parameters when -GS stack reordering is enabled |
1762 | // as we could introduce shadow copies of them. |
1763 | if (varDsc->lvIsParam && compiler->compGSReorderStackLayout) |
1764 | { |
1765 | JITDUMP(" struct promotion of V%02u is disabled because lvIsParam and compGSReorderStackLayout\n" , lclNum); |
1766 | return false; |
1767 | } |
1768 | |
1769 | // Explicitly check for HFA reg args and reject them for promotion here. |
1770 | // Promoting HFA args will fire an assert in lvaAssignFrameOffsets |
1771 | // when the HFA reg arg is struct promoted. |
1772 | // |
1773 | // TODO-PERF - Allow struct promotion for HFA register arguments |
1774 | if (varDsc->lvIsHfaRegArg()) |
1775 | { |
1776 | JITDUMP(" struct promotion of V%02u is disabled because lvIsHfaRegArg()\n" , lclNum); |
1777 | return false; |
1778 | } |
1779 | |
1780 | #if !FEATURE_MULTIREG_STRUCT_PROMOTE |
1781 | if (varDsc->lvIsMultiRegArg) |
1782 | { |
1783 | JITDUMP(" struct promotion of V%02u is disabled because lvIsMultiRegArg\n" , lclNum); |
1784 | return false; |
1785 | } |
1786 | #endif |
1787 | |
1788 | if (varDsc->lvIsMultiRegRet) |
1789 | { |
1790 | JITDUMP(" struct promotion of V%02u is disabled because lvIsMultiRegRet\n" , lclNum); |
1791 | return false; |
1792 | } |
1793 | |
1794 | CORINFO_CLASS_HANDLE typeHnd = varDsc->lvVerTypeInfo.GetClassHandle(); |
1795 | return CanPromoteStructType(typeHnd); |
1796 | } |
1797 | |
1798 | //-------------------------------------------------------------------------------------------- |
1799 | // ShouldPromoteStructVar - Should a struct var be promoted if it can be promoted? |
1800 | // This routine mainly performs profitability checks. Right now it also has |
1801 | // some correctness checks due to limitations of down-stream phases. |
1802 | // |
1803 | // Arguments: |
1804 | // lclNum - struct local number; |
1805 | // |
1806 | // Return value: |
1807 | // true if the struct should be promoted. |
1808 | // |
1809 | bool Compiler::StructPromotionHelper::ShouldPromoteStructVar(unsigned lclNum) |
1810 | { |
1811 | assert(lclNum < compiler->lvaCount); |
1812 | |
1813 | LclVarDsc* varDsc = &compiler->lvaTable[lclNum]; |
1814 | assert(varTypeIsStruct(varDsc)); |
1815 | assert(varDsc->lvVerTypeInfo.GetClassHandle() == structPromotionInfo.typeHnd); |
1816 | assert(structPromotionInfo.canPromote); |
1817 | |
1818 | bool shouldPromote = true; |
1819 | |
1820 | // We *can* promote; *should* we promote? |
1821 | // We should only do so if promotion has potential savings. One source of savings |
1822 | // is if a field of the struct is accessed, since this access will be turned into |
1823 | // an access of the corresponding promoted field variable. Even if there are no |
1824 | // field accesses, but only block-level operations on the whole struct, if the struct |
1825 | // has only one or two fields, then doing those block operations field-wise is probably faster |
1826 | // than doing a whole-variable block operation (e.g., a hardware "copy loop" on x86). |
1827 | // Struct promotion also provides the following benefits: reduce stack frame size, |
1828 | // reduce the need for zero init of stack frame and fine grained constant/copy prop. |
1829 | // Asm diffs indicate that promoting structs up to 3 fields is a net size win. |
1830 | // So if no fields are accessed independently, and there are four or more fields, |
1831 | // then do not promote. |
1832 | // |
1833 | // TODO: Ideally we would want to consider the impact of whether the struct is |
1834 | // passed as a parameter or assigned the return value of a call. Because once promoted, |
1835 | // struct copying is done by field by field assignment instead of a more efficient |
1836 | // rep.stos or xmm reg based copy. |
1837 | if (structPromotionInfo.fieldCnt > 3 && !varDsc->lvFieldAccessed) |
1838 | { |
1839 | JITDUMP("Not promoting promotable struct local V%02u: #fields = %d, fieldAccessed = %d.\n" , lclNum, |
1840 | structPromotionInfo.fieldCnt, varDsc->lvFieldAccessed); |
1841 | shouldPromote = false; |
1842 | } |
1843 | #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_) || defined(_TARGET_ARM_) |
1844 | // TODO-PERF - Only do this when the LclVar is used in an argument context |
1845 | // TODO-ARM64 - HFA support should also eliminate the need for this. |
1846 | // TODO-ARM32 - HFA support should also eliminate the need for this. |
1847 | // TODO-LSRA - Currently doesn't support the passing of floating point LCL_VARS in the integer registers |
1848 | // |
1849 | // For now we currently don't promote structs with a single float field |
1850 | // Promoting it can cause us to shuffle it back and forth between the int and |
1851 | // the float regs when it is used as a argument, which is very expensive for XARCH |
1852 | // |
1853 | else if ((structPromotionInfo.fieldCnt == 1) && varTypeIsFloating(structPromotionInfo.fields[0].fldType)) |
1854 | { |
1855 | JITDUMP("Not promoting promotable struct local V%02u: #fields = %d because it is a struct with " |
1856 | "single float field.\n" , |
1857 | lclNum, structPromotionInfo.fieldCnt); |
1858 | shouldPromote = false; |
1859 | } |
1860 | #endif // _TARGET_AMD64_ || _TARGET_ARM64_ || _TARGET_ARM_ |
1861 | else if (varDsc->lvIsParam && !compiler->lvaIsImplicitByRefLocal(lclNum)) |
1862 | { |
1863 | #if FEATURE_MULTIREG_STRUCT_PROMOTE |
1864 | // Is this a variable holding a value with exactly two fields passed in |
1865 | // multiple registers? |
1866 | if ((structPromotionInfo.fieldCnt != 2) && compiler->lvaIsMultiregStruct(varDsc, compiler->info.compIsVarArgs)) |
1867 | { |
1868 | JITDUMP("Not promoting multireg struct local V%02u, because lvIsParam is true and #fields != 2\n" , lclNum); |
1869 | shouldPromote = false; |
1870 | } |
1871 | else |
1872 | #endif // !FEATURE_MULTIREG_STRUCT_PROMOTE |
1873 | |
1874 | // TODO-PERF - Implement struct promotion for incoming multireg structs |
1875 | // Currently it hits assert(lvFieldCnt==1) in lclvar.cpp line 4417 |
1876 | // Also the implementation of jmp uses the 4 byte move to store |
1877 | // byte parameters to the stack, so that if we have a byte field |
1878 | // with something else occupying the same 4-byte slot, it will |
1879 | // overwrite other fields. |
1880 | if (structPromotionInfo.fieldCnt != 1) |
1881 | { |
1882 | JITDUMP("Not promoting promotable struct local V%02u, because lvIsParam is true and #fields = " |
1883 | "%d.\n" , |
1884 | lclNum, structPromotionInfo.fieldCnt); |
1885 | shouldPromote = false; |
1886 | } |
1887 | } |
1888 | |
1889 | // |
1890 | // If the lvRefCnt is zero and we have a struct promoted parameter we can end up with an extra store of |
1891 | // the the incoming register into the stack frame slot. |
1892 | // In that case, we would like to avoid promortion. |
1893 | // However we haven't yet computed the lvRefCnt values so we can't do that. |
1894 | // |
1895 | CLANG_FORMAT_COMMENT_ANCHOR; |
1896 | |
1897 | return shouldPromote; |
1898 | } |
1899 | |
1900 | //-------------------------------------------------------------------------------------------- |
1901 | // SortStructFields - sort the fields according to the increasing order of the field offset. |
1902 | // |
1903 | // Notes: |
1904 | // This is needed because the fields need to be pushed on stack (when referenced as a struct) in offset order. |
1905 | // |
1906 | void Compiler::StructPromotionHelper::SortStructFields() |
1907 | { |
1908 | assert(!structPromotionInfo.fieldsSorted); |
1909 | qsort(structPromotionInfo.fields, structPromotionInfo.fieldCnt, sizeof(*structPromotionInfo.fields), |
1910 | lvaFieldOffsetCmp); |
1911 | structPromotionInfo.fieldsSorted = true; |
1912 | } |
1913 | |
1914 | //-------------------------------------------------------------------------------------------- |
1915 | // GetFieldInfo - get struct field information. |
1916 | // Arguments: |
1917 | // fieldHnd - field handle to get info for; |
1918 | // ordinal - field ordinal. |
1919 | // |
1920 | // Return value: |
1921 | // field information. |
1922 | // |
1923 | Compiler::lvaStructFieldInfo Compiler::StructPromotionHelper::GetFieldInfo(CORINFO_FIELD_HANDLE fieldHnd, BYTE ordinal) |
1924 | { |
1925 | lvaStructFieldInfo fieldInfo; |
1926 | fieldInfo.fldHnd = fieldHnd; |
1927 | |
1928 | unsigned fldOffset = compiler->info.compCompHnd->getFieldOffset(fieldInfo.fldHnd); |
1929 | fieldInfo.fldOffset = (BYTE)fldOffset; |
1930 | |
1931 | fieldInfo.fldOrdinal = ordinal; |
1932 | CorInfoType corType = compiler->info.compCompHnd->getFieldType(fieldInfo.fldHnd, &fieldInfo.fldTypeHnd); |
1933 | fieldInfo.fldType = JITtype2varType(corType); |
1934 | fieldInfo.fldSize = genTypeSize(fieldInfo.fldType); |
1935 | |
1936 | #ifdef FEATURE_SIMD |
1937 | // Check to see if this is a SIMD type. |
1938 | // We will only check this if we have already found a SIMD type, which will be true if |
1939 | // we have encountered any SIMD intrinsics. |
1940 | if (compiler->usesSIMDTypes() && (fieldInfo.fldSize == 0) && compiler->isSIMDorHWSIMDClass(fieldInfo.fldTypeHnd)) |
1941 | { |
1942 | unsigned simdSize; |
1943 | var_types simdBaseType = compiler->getBaseTypeAndSizeOfSIMDType(fieldInfo.fldTypeHnd, &simdSize); |
1944 | if (simdBaseType != TYP_UNKNOWN) |
1945 | { |
1946 | fieldInfo.fldType = compiler->getSIMDTypeForSize(simdSize); |
1947 | fieldInfo.fldSize = simdSize; |
1948 | #ifdef DEBUG |
1949 | retypedFieldsMap.Set(fieldInfo.fldHnd, fieldInfo.fldType); |
1950 | #endif // DEBUG |
1951 | } |
1952 | } |
1953 | #endif // FEATURE_SIMD |
1954 | |
1955 | if (fieldInfo.fldSize == 0) |
1956 | { |
1957 | TryPromoteStructField(fieldInfo); |
1958 | } |
1959 | |
1960 | return fieldInfo; |
1961 | } |
1962 | |
1963 | //-------------------------------------------------------------------------------------------- |
1964 | // TryPromoteStructField - checks that this struct's field is a struct that can be promoted as scalar type |
1965 | // aligned at its natural boundary. Promotes the field as a scalar if the check succeeded. |
1966 | // |
1967 | // Arguments: |
1968 | // fieldInfo - information about the field in the outer struct. |
1969 | // |
1970 | // Return value: |
1971 | // true if the internal struct was promoted. |
1972 | // |
1973 | bool Compiler::StructPromotionHelper::TryPromoteStructField(lvaStructFieldInfo& fieldInfo) |
1974 | { |
1975 | // Size of TYP_BLK, TYP_FUNC, TYP_VOID and TYP_STRUCT is zero. |
1976 | // Early out if field type is other than TYP_STRUCT. |
1977 | // This is a defensive check as we don't expect a struct to have |
1978 | // fields of TYP_BLK, TYP_FUNC or TYP_VOID. |
1979 | if (fieldInfo.fldType != TYP_STRUCT) |
1980 | { |
1981 | return false; |
1982 | } |
1983 | |
1984 | COMP_HANDLE compHandle = compiler->info.compCompHnd; |
1985 | |
1986 | // Do not promote if the struct field in turn has more than one field. |
1987 | if (compHandle->getClassNumInstanceFields(fieldInfo.fldTypeHnd) != 1) |
1988 | { |
1989 | return false; |
1990 | } |
1991 | |
1992 | COMP_HANDLE compHandl = compiler->info.compCompHnd; |
1993 | |
1994 | // Do not promote if the single field is not aligned at its natural boundary within |
1995 | // the struct field. |
1996 | CORINFO_FIELD_HANDLE innerFieldHndl = compHandle->getFieldInClass(fieldInfo.fldTypeHnd, 0); |
1997 | unsigned innerFieldOffset = compHandle->getFieldOffset(innerFieldHndl); |
1998 | if (innerFieldOffset != 0) |
1999 | { |
2000 | return false; |
2001 | } |
2002 | |
2003 | CorInfoType fieldCorType = compHandle->getFieldType(innerFieldHndl); |
2004 | var_types fieldVarType = JITtype2varType(fieldCorType); |
2005 | unsigned fieldSize = genTypeSize(fieldVarType); |
2006 | |
2007 | // Do not promote if either not a primitive type or size equal to ptr size on |
2008 | // target or a struct containing a single floating-point field. |
2009 | // |
2010 | // TODO-PERF: Structs containing a single floating-point field on Amd64 |
2011 | // need to be passed in integer registers. Right now LSRA doesn't support |
2012 | // passing of floating-point LCL_VARS in integer registers. Enabling promotion |
2013 | // of such structs results in an assert in lsra right now. |
2014 | // |
2015 | // TODO-PERF: Right now promotion is confined to struct containing a ptr sized |
2016 | // field (int/uint/ref/byref on 32-bits and long/ulong/ref/byref on 64-bits). |
2017 | // Though this would serve the purpose of promoting Span<T> containing ByReference<T>, |
2018 | // this can be extended to other primitive types as long as they are aligned at their |
2019 | // natural boundary. |
2020 | // |
2021 | // TODO-CQ: Right now we only promote an actual SIMD typed field, which would cause |
2022 | // a nested SIMD type to fail promotion. |
2023 | if (fieldSize == 0 || fieldSize != TARGET_POINTER_SIZE || varTypeIsFloating(fieldVarType)) |
2024 | { |
2025 | JITDUMP("Promotion blocked: struct contains struct field with one field," |
2026 | " but that field has invalid size or type" ); |
2027 | return false; |
2028 | } |
2029 | |
2030 | // Insist this wrapped field occupy all of its parent storage. |
2031 | unsigned innerStructSize = compHandle->getClassSize(fieldInfo.fldTypeHnd); |
2032 | |
2033 | if (fieldSize != innerStructSize) |
2034 | { |
2035 | JITDUMP("Promotion blocked: struct contains struct field with one field," |
2036 | " but that field is not the same size as its parent." ); |
2037 | return false; |
2038 | } |
2039 | |
2040 | // Retype the field as the type of the single field of the struct. |
2041 | // This is a hack that allows us to promote such fields before we support recursive struct promotion |
2042 | // (tracked by #10019). |
2043 | fieldInfo.fldType = fieldVarType; |
2044 | fieldInfo.fldSize = fieldSize; |
2045 | #ifdef DEBUG |
2046 | retypedFieldsMap.Set(fieldInfo.fldHnd, fieldInfo.fldType); |
2047 | #endif // DEBUG |
2048 | return true; |
2049 | } |
2050 | |
2051 | //-------------------------------------------------------------------------------------------- |
2052 | // PromoteStructVar - promote struct variable. |
2053 | // |
2054 | // Arguments: |
2055 | // lclNum - struct local number; |
2056 | // |
2057 | void Compiler::StructPromotionHelper::PromoteStructVar(unsigned lclNum) |
2058 | { |
2059 | LclVarDsc* varDsc = &compiler->lvaTable[lclNum]; |
2060 | |
2061 | // We should never see a reg-sized non-field-addressed struct here. |
2062 | assert(!varDsc->lvRegStruct); |
2063 | |
2064 | assert(varDsc->lvVerTypeInfo.GetClassHandle() == structPromotionInfo.typeHnd); |
2065 | assert(structPromotionInfo.canPromote); |
2066 | |
2067 | varDsc->lvFieldCnt = structPromotionInfo.fieldCnt; |
2068 | varDsc->lvFieldLclStart = compiler->lvaCount; |
2069 | varDsc->lvPromoted = true; |
2070 | varDsc->lvContainsHoles = structPromotionInfo.containsHoles; |
2071 | varDsc->lvCustomLayout = structPromotionInfo.customLayout; |
2072 | |
2073 | #ifdef DEBUG |
2074 | // Don't change the source to a TYP_BLK either. |
2075 | varDsc->lvKeepType = 1; |
2076 | #endif |
2077 | |
2078 | #ifdef DEBUG |
2079 | if (compiler->verbose) |
2080 | { |
2081 | printf("\nPromoting struct local V%02u (%s):" , lclNum, |
2082 | compiler->eeGetClassName(varDsc->lvVerTypeInfo.GetClassHandle())); |
2083 | } |
2084 | #endif |
2085 | |
2086 | if (!structPromotionInfo.fieldsSorted) |
2087 | { |
2088 | SortStructFields(); |
2089 | } |
2090 | |
2091 | for (unsigned index = 0; index < structPromotionInfo.fieldCnt; ++index) |
2092 | { |
2093 | const lvaStructFieldInfo* pFieldInfo = &structPromotionInfo.fields[index]; |
2094 | |
2095 | if (varTypeIsFloating(pFieldInfo->fldType) || varTypeIsSIMD(pFieldInfo->fldType)) |
2096 | { |
2097 | // Whenever we promote a struct that contains a floating point field |
2098 | // it's possible we transition from a method that originally only had integer |
2099 | // local vars to start having FP. We have to communicate this through this flag |
2100 | // since LSRA later on will use this flag to determine whether or not to track FP register sets. |
2101 | compiler->compFloatingPointUsed = true; |
2102 | } |
2103 | |
2104 | // Now grab the temp for the field local. |
2105 | |
2106 | #ifdef DEBUG |
2107 | char buf[200]; |
2108 | sprintf_s(buf, sizeof(buf), "%s V%02u.%s (fldOffset=0x%x)" , "field" , lclNum, |
2109 | compiler->eeGetFieldName(pFieldInfo->fldHnd), pFieldInfo->fldOffset); |
2110 | |
2111 | // We need to copy 'buf' as lvaGrabTemp() below caches a copy to its argument. |
2112 | size_t len = strlen(buf) + 1; |
2113 | char* bufp = compiler->getAllocator(CMK_DebugOnly).allocate<char>(len); |
2114 | strcpy_s(bufp, len, buf); |
2115 | |
2116 | if (index > 0) |
2117 | { |
2118 | noway_assert(pFieldInfo->fldOffset > (pFieldInfo - 1)->fldOffset); |
2119 | } |
2120 | #endif |
2121 | |
2122 | // Lifetime of field locals might span multiple BBs, so they must be long lifetime temps. |
2123 | unsigned varNum = compiler->lvaGrabTemp(false DEBUGARG(bufp)); |
2124 | |
2125 | varDsc = &compiler->lvaTable[lclNum]; // lvaGrabTemp can reallocate the lvaTable |
2126 | |
2127 | LclVarDsc* fieldVarDsc = &compiler->lvaTable[varNum]; |
2128 | fieldVarDsc->lvType = pFieldInfo->fldType; |
2129 | fieldVarDsc->lvExactSize = pFieldInfo->fldSize; |
2130 | fieldVarDsc->lvIsStructField = true; |
2131 | fieldVarDsc->lvFieldHnd = pFieldInfo->fldHnd; |
2132 | fieldVarDsc->lvFldOffset = pFieldInfo->fldOffset; |
2133 | fieldVarDsc->lvFldOrdinal = pFieldInfo->fldOrdinal; |
2134 | fieldVarDsc->lvParentLcl = lclNum; |
2135 | fieldVarDsc->lvIsParam = varDsc->lvIsParam; |
2136 | #if defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_) |
2137 | // Do we have a parameter that can be enregistered? |
2138 | // |
2139 | if (varDsc->lvIsRegArg) |
2140 | { |
2141 | fieldVarDsc->lvIsRegArg = true; |
2142 | fieldVarDsc->lvArgReg = varDsc->lvArgReg; |
2143 | #if FEATURE_MULTIREG_ARGS && defined(FEATURE_SIMD) |
2144 | if (varTypeIsSIMD(fieldVarDsc) && !compiler->lvaIsImplicitByRefLocal(lclNum)) |
2145 | { |
2146 | // This field is a SIMD type, and will be considered to be passed in multiple registers |
2147 | // if the parent struct was. Note that this code relies on the fact that if there is |
2148 | // a SIMD field of an enregisterable struct, it is the only field. |
2149 | // We will assert that, in case future changes are made to the ABI. |
2150 | assert(varDsc->lvFieldCnt == 1); |
2151 | fieldVarDsc->lvOtherArgReg = varDsc->lvOtherArgReg; |
2152 | } |
2153 | #endif // FEATURE_MULTIREG_ARGS && defined(FEATURE_SIMD) |
2154 | } |
2155 | #endif |
2156 | |
2157 | #ifdef FEATURE_SIMD |
2158 | if (varTypeIsSIMD(pFieldInfo->fldType)) |
2159 | { |
2160 | // Set size to zero so that lvaSetStruct will appropriately set the SIMD-relevant fields. |
2161 | fieldVarDsc->lvExactSize = 0; |
2162 | compiler->lvaSetStruct(varNum, pFieldInfo->fldTypeHnd, false, true); |
2163 | } |
2164 | #endif // FEATURE_SIMD |
2165 | |
2166 | #ifdef DEBUG |
2167 | // This temporary should not be converted to a double in stress mode, |
2168 | // because we introduce assigns to it after the stress conversion |
2169 | fieldVarDsc->lvKeepType = 1; |
2170 | #endif |
2171 | } |
2172 | } |
2173 | |
2174 | #if !defined(_TARGET_64BIT_) |
2175 | //------------------------------------------------------------------------ |
2176 | // lvaPromoteLongVars: "Struct promote" all register candidate longs as if they are structs of two ints. |
2177 | // |
2178 | // Arguments: |
2179 | // None. |
2180 | // |
2181 | // Return Value: |
2182 | // None. |
2183 | // |
2184 | void Compiler::lvaPromoteLongVars() |
2185 | { |
2186 | if ((opts.compFlags & CLFLG_REGVAR) == 0) |
2187 | { |
2188 | return; |
2189 | } |
2190 | |
2191 | // The lvaTable might grow as we grab temps. Make a local copy here. |
2192 | unsigned startLvaCount = lvaCount; |
2193 | for (unsigned lclNum = 0; lclNum < startLvaCount; lclNum++) |
2194 | { |
2195 | LclVarDsc* varDsc = &lvaTable[lclNum]; |
2196 | if (!varTypeIsLong(varDsc) || varDsc->lvDoNotEnregister || varDsc->lvIsMultiRegArgOrRet() || |
2197 | (varDsc->lvRefCnt() == 0) || varDsc->lvIsStructField || (fgNoStructPromotion && varDsc->lvIsParam)) |
2198 | { |
2199 | continue; |
2200 | } |
2201 | |
2202 | varDsc->lvFieldCnt = 2; |
2203 | varDsc->lvFieldLclStart = lvaCount; |
2204 | varDsc->lvPromoted = true; |
2205 | varDsc->lvContainsHoles = false; |
2206 | |
2207 | #ifdef DEBUG |
2208 | if (verbose) |
2209 | { |
2210 | printf("\nPromoting long local V%02u:" , lclNum); |
2211 | } |
2212 | #endif |
2213 | |
2214 | bool isParam = varDsc->lvIsParam; |
2215 | |
2216 | for (unsigned index = 0; index < 2; ++index) |
2217 | { |
2218 | // Grab the temp for the field local. |
2219 | CLANG_FORMAT_COMMENT_ANCHOR; |
2220 | |
2221 | #ifdef DEBUG |
2222 | char buf[200]; |
2223 | sprintf_s(buf, sizeof(buf), "%s V%02u.%s (fldOffset=0x%x)" , "field" , lclNum, index == 0 ? "lo" : "hi" , |
2224 | index * 4); |
2225 | |
2226 | // We need to copy 'buf' as lvaGrabTemp() below caches a copy to its argument. |
2227 | size_t len = strlen(buf) + 1; |
2228 | char* bufp = getAllocator(CMK_DebugOnly).allocate<char>(len); |
2229 | strcpy_s(bufp, len, buf); |
2230 | #endif |
2231 | |
2232 | unsigned varNum = lvaGrabTemp(false DEBUGARG(bufp)); // Lifetime of field locals might span multiple BBs, so |
2233 | // they are long lifetime temps. |
2234 | |
2235 | LclVarDsc* fieldVarDsc = &lvaTable[varNum]; |
2236 | fieldVarDsc->lvType = TYP_INT; |
2237 | fieldVarDsc->lvExactSize = genTypeSize(TYP_INT); |
2238 | fieldVarDsc->lvIsStructField = true; |
2239 | fieldVarDsc->lvFldOffset = (unsigned char)(index * genTypeSize(TYP_INT)); |
2240 | fieldVarDsc->lvFldOrdinal = (unsigned char)index; |
2241 | fieldVarDsc->lvParentLcl = lclNum; |
2242 | // Currently we do not support enregistering incoming promoted aggregates with more than one field. |
2243 | if (isParam) |
2244 | { |
2245 | fieldVarDsc->lvIsParam = true; |
2246 | lvaSetVarDoNotEnregister(varNum DEBUGARG(DNER_LongParamField)); |
2247 | } |
2248 | } |
2249 | } |
2250 | |
2251 | #ifdef DEBUG |
2252 | if (verbose) |
2253 | { |
2254 | printf("\nlvaTable after lvaPromoteLongVars\n" ); |
2255 | lvaTableDump(); |
2256 | } |
2257 | #endif // DEBUG |
2258 | } |
2259 | #endif // !defined(_TARGET_64BIT_) |
2260 | |
2261 | //-------------------------------------------------------------------------------------------- |
2262 | // lvaGetFieldLocal - returns the local var index for a promoted field in a promoted struct var. |
2263 | // |
2264 | // Arguments: |
2265 | // varDsc - the promoted struct var descriptor; |
2266 | // fldOffset - field offset in the struct. |
2267 | // |
2268 | // Return value: |
2269 | // the index of the local that represents this field. |
2270 | // |
2271 | unsigned Compiler::lvaGetFieldLocal(const LclVarDsc* varDsc, unsigned int fldOffset) |
2272 | { |
2273 | noway_assert(varTypeIsStruct(varDsc)); |
2274 | noway_assert(varDsc->lvPromoted); |
2275 | |
2276 | for (unsigned i = varDsc->lvFieldLclStart; i < varDsc->lvFieldLclStart + varDsc->lvFieldCnt; ++i) |
2277 | { |
2278 | noway_assert(lvaTable[i].lvIsStructField); |
2279 | noway_assert(lvaTable[i].lvParentLcl == (unsigned)(varDsc - lvaTable)); |
2280 | if (lvaTable[i].lvFldOffset == fldOffset) |
2281 | { |
2282 | return i; |
2283 | } |
2284 | } |
2285 | |
2286 | // This is the not-found error return path, the caller should check for BAD_VAR_NUM |
2287 | return BAD_VAR_NUM; |
2288 | } |
2289 | |
2290 | /***************************************************************************** |
2291 | * |
2292 | * Set the local var "varNum" as address-exposed. |
2293 | * If this is a promoted struct, label it's fields the same way. |
2294 | */ |
2295 | |
2296 | void Compiler::lvaSetVarAddrExposed(unsigned varNum) |
2297 | { |
2298 | noway_assert(varNum < lvaCount); |
2299 | |
2300 | LclVarDsc* varDsc = &lvaTable[varNum]; |
2301 | |
2302 | varDsc->lvAddrExposed = 1; |
2303 | |
2304 | if (varDsc->lvPromoted) |
2305 | { |
2306 | noway_assert(varTypeIsStruct(varDsc)); |
2307 | |
2308 | for (unsigned i = varDsc->lvFieldLclStart; i < varDsc->lvFieldLclStart + varDsc->lvFieldCnt; ++i) |
2309 | { |
2310 | noway_assert(lvaTable[i].lvIsStructField); |
2311 | lvaTable[i].lvAddrExposed = 1; // Make field local as address-exposed. |
2312 | lvaSetVarDoNotEnregister(i DEBUGARG(DNER_AddrExposed)); |
2313 | } |
2314 | } |
2315 | |
2316 | lvaSetVarDoNotEnregister(varNum DEBUGARG(DNER_AddrExposed)); |
2317 | } |
2318 | |
2319 | /***************************************************************************** |
2320 | * |
2321 | * Record that the local var "varNum" should not be enregistered (for one of several reasons.) |
2322 | */ |
2323 | |
2324 | void Compiler::lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregisterReason reason)) |
2325 | { |
2326 | noway_assert(varNum < lvaCount); |
2327 | LclVarDsc* varDsc = &lvaTable[varNum]; |
2328 | varDsc->lvDoNotEnregister = 1; |
2329 | |
2330 | #ifdef DEBUG |
2331 | if (verbose) |
2332 | { |
2333 | printf("\nLocal V%02u should not be enregistered because: " , varNum); |
2334 | } |
2335 | switch (reason) |
2336 | { |
2337 | case DNER_AddrExposed: |
2338 | JITDUMP("it is address exposed\n" ); |
2339 | assert(varDsc->lvAddrExposed); |
2340 | break; |
2341 | case DNER_IsStruct: |
2342 | JITDUMP("it is a struct\n" ); |
2343 | assert(varTypeIsStruct(varDsc)); |
2344 | break; |
2345 | case DNER_IsStructArg: |
2346 | JITDUMP("it is a struct arg\n" ); |
2347 | assert(varTypeIsStruct(varDsc)); |
2348 | break; |
2349 | case DNER_BlockOp: |
2350 | JITDUMP("written in a block op\n" ); |
2351 | varDsc->lvLclBlockOpAddr = 1; |
2352 | break; |
2353 | case DNER_LocalField: |
2354 | JITDUMP("was accessed as a local field\n" ); |
2355 | varDsc->lvLclFieldExpr = 1; |
2356 | break; |
2357 | case DNER_VMNeedsStackAddr: |
2358 | JITDUMP("needs stack addr\n" ); |
2359 | varDsc->lvVMNeedsStackAddr = 1; |
2360 | break; |
2361 | case DNER_LiveInOutOfHandler: |
2362 | JITDUMP("live in/out of a handler\n" ); |
2363 | varDsc->lvLiveInOutOfHndlr = 1; |
2364 | break; |
2365 | case DNER_LiveAcrossUnmanagedCall: |
2366 | JITDUMP("live across unmanaged call\n" ); |
2367 | varDsc->lvLiveAcrossUCall = 1; |
2368 | break; |
2369 | case DNER_DepField: |
2370 | JITDUMP("field of a dependently promoted struct\n" ); |
2371 | assert(varDsc->lvIsStructField && (lvaGetParentPromotionType(varNum) != PROMOTION_TYPE_INDEPENDENT)); |
2372 | break; |
2373 | case DNER_NoRegVars: |
2374 | JITDUMP("opts.compFlags & CLFLG_REGVAR is not set\n" ); |
2375 | assert((opts.compFlags & CLFLG_REGVAR) == 0); |
2376 | break; |
2377 | case DNER_MinOptsGC: |
2378 | JITDUMP("It is a GC Ref and we are compiling MinOpts\n" ); |
2379 | assert(!JitConfig.JitMinOptsTrackGCrefs() && varTypeIsGC(varDsc->TypeGet())); |
2380 | break; |
2381 | #ifdef JIT32_GCENCODER |
2382 | case DNER_PinningRef: |
2383 | JITDUMP("pinning ref\n" ); |
2384 | assert(varDsc->lvPinned); |
2385 | break; |
2386 | #endif |
2387 | #if !defined(_TARGET_64BIT_) |
2388 | case DNER_LongParamField: |
2389 | JITDUMP("it is a decomposed field of a long parameter\n" ); |
2390 | break; |
2391 | #endif |
2392 | default: |
2393 | unreached(); |
2394 | break; |
2395 | } |
2396 | #endif |
2397 | } |
2398 | |
2399 | // Returns true if this local var is a multireg struct. |
2400 | // TODO-Throughput: This does a lookup on the class handle, and in the outgoing arg context |
2401 | // this information is already available on the fgArgTabEntry, and shouldn't need to be |
2402 | // recomputed. |
2403 | // |
2404 | bool Compiler::lvaIsMultiregStruct(LclVarDsc* varDsc, bool isVarArg) |
2405 | { |
2406 | if (varTypeIsStruct(varDsc->TypeGet())) |
2407 | { |
2408 | CORINFO_CLASS_HANDLE clsHnd = varDsc->lvVerTypeInfo.GetClassHandleForValueClass(); |
2409 | structPassingKind howToPassStruct; |
2410 | |
2411 | var_types type = getArgTypeForStruct(clsHnd, &howToPassStruct, isVarArg, varDsc->lvExactSize); |
2412 | |
2413 | if (howToPassStruct == SPK_ByValueAsHfa) |
2414 | { |
2415 | assert(type == TYP_STRUCT); |
2416 | return true; |
2417 | } |
2418 | |
2419 | #if defined(UNIX_AMD64_ABI) || defined(_TARGET_ARM64_) |
2420 | if (howToPassStruct == SPK_ByValue) |
2421 | { |
2422 | assert(type == TYP_STRUCT); |
2423 | return true; |
2424 | } |
2425 | #endif |
2426 | } |
2427 | return false; |
2428 | } |
2429 | |
2430 | /***************************************************************************** |
2431 | * Set the lvClass for a local variable of a struct type */ |
2432 | |
2433 | void Compiler::lvaSetStruct(unsigned varNum, CORINFO_CLASS_HANDLE typeHnd, bool unsafeValueClsCheck, bool setTypeInfo) |
2434 | { |
2435 | noway_assert(varNum < lvaCount); |
2436 | |
2437 | LclVarDsc* varDsc = &lvaTable[varNum]; |
2438 | if (setTypeInfo) |
2439 | { |
2440 | varDsc->lvVerTypeInfo = typeInfo(TI_STRUCT, typeHnd); |
2441 | } |
2442 | |
2443 | // Set the type and associated info if we haven't already set it. |
2444 | var_types structType = varDsc->lvType; |
2445 | if (varDsc->lvType == TYP_UNDEF) |
2446 | { |
2447 | varDsc->lvType = TYP_STRUCT; |
2448 | } |
2449 | if (varDsc->lvExactSize == 0) |
2450 | { |
2451 | BOOL isValueClass = info.compCompHnd->isValueClass(typeHnd); |
2452 | |
2453 | if (isValueClass) |
2454 | { |
2455 | varDsc->lvExactSize = info.compCompHnd->getClassSize(typeHnd); |
2456 | } |
2457 | else |
2458 | { |
2459 | varDsc->lvExactSize = info.compCompHnd->getHeapClassSize(typeHnd); |
2460 | } |
2461 | |
2462 | size_t lvSize = varDsc->lvSize(); |
2463 | assert((lvSize % TARGET_POINTER_SIZE) == |
2464 | 0); // The struct needs to be a multiple of TARGET_POINTER_SIZE bytes for getClassGClayout() to be valid. |
2465 | varDsc->lvGcLayout = getAllocator(CMK_LvaTable).allocate<BYTE>(lvSize / TARGET_POINTER_SIZE); |
2466 | unsigned numGCVars; |
2467 | var_types simdBaseType = TYP_UNKNOWN; |
2468 | if (isValueClass) |
2469 | { |
2470 | varDsc->lvType = impNormStructType(typeHnd, varDsc->lvGcLayout, &numGCVars, &simdBaseType); |
2471 | } |
2472 | else |
2473 | { |
2474 | numGCVars = info.compCompHnd->getClassGClayout(typeHnd, varDsc->lvGcLayout); |
2475 | } |
2476 | |
2477 | // We only save the count of GC vars in a struct up to 7. |
2478 | if (numGCVars >= 8) |
2479 | { |
2480 | numGCVars = 7; |
2481 | } |
2482 | varDsc->lvStructGcCount = numGCVars; |
2483 | |
2484 | if (isValueClass) |
2485 | { |
2486 | #if FEATURE_SIMD |
2487 | if (simdBaseType != TYP_UNKNOWN) |
2488 | { |
2489 | assert(varTypeIsSIMD(varDsc)); |
2490 | varDsc->lvSIMDType = true; |
2491 | varDsc->lvBaseType = simdBaseType; |
2492 | } |
2493 | #endif // FEATURE_SIMD |
2494 | #ifdef FEATURE_HFA |
2495 | // for structs that are small enough, we check and set lvIsHfa and lvHfaTypeIsFloat |
2496 | if (varDsc->lvExactSize <= MAX_PASS_MULTIREG_BYTES) |
2497 | { |
2498 | var_types hfaType = GetHfaType(typeHnd); // set to float or double if it is an HFA, otherwise TYP_UNDEF |
2499 | if (varTypeIsFloating(hfaType)) |
2500 | { |
2501 | varDsc->_lvIsHfa = true; |
2502 | varDsc->lvSetHfaTypeIsFloat(hfaType == TYP_FLOAT); |
2503 | |
2504 | // hfa variables can never contain GC pointers |
2505 | assert(varDsc->lvStructGcCount == 0); |
2506 | // The size of this struct should be evenly divisible by 4 or 8 |
2507 | assert((varDsc->lvExactSize % genTypeSize(hfaType)) == 0); |
2508 | // The number of elements in the HFA should fit into our MAX_ARG_REG_COUNT limit |
2509 | assert((varDsc->lvExactSize / genTypeSize(hfaType)) <= MAX_ARG_REG_COUNT); |
2510 | } |
2511 | } |
2512 | #endif // FEATURE_HFA |
2513 | } |
2514 | } |
2515 | else |
2516 | { |
2517 | #if FEATURE_SIMD |
2518 | assert(!varTypeIsSIMD(varDsc) || (varDsc->lvBaseType != TYP_UNKNOWN)); |
2519 | #endif // FEATURE_SIMD |
2520 | } |
2521 | |
2522 | #ifndef _TARGET_64BIT_ |
2523 | BOOL fDoubleAlignHint = FALSE; |
2524 | #ifdef _TARGET_X86_ |
2525 | fDoubleAlignHint = TRUE; |
2526 | #endif |
2527 | |
2528 | if (info.compCompHnd->getClassAlignmentRequirement(typeHnd, fDoubleAlignHint) == 8) |
2529 | { |
2530 | #ifdef DEBUG |
2531 | if (verbose) |
2532 | { |
2533 | printf("Marking struct in V%02i with double align flag\n" , varNum); |
2534 | } |
2535 | #endif |
2536 | varDsc->lvStructDoubleAlign = 1; |
2537 | } |
2538 | #endif // not _TARGET_64BIT_ |
2539 | |
2540 | unsigned classAttribs = info.compCompHnd->getClassAttribs(typeHnd); |
2541 | |
2542 | varDsc->lvOverlappingFields = StructHasOverlappingFields(classAttribs); |
2543 | |
2544 | // Check whether this local is an unsafe value type and requires GS cookie protection. |
2545 | // GS checks require the stack to be re-ordered, which can't be done with EnC. |
2546 | if (unsafeValueClsCheck && (classAttribs & CORINFO_FLG_UNSAFE_VALUECLASS) && !opts.compDbgEnC) |
2547 | { |
2548 | setNeedsGSSecurityCookie(); |
2549 | compGSReorderStackLayout = true; |
2550 | varDsc->lvIsUnsafeBuffer = true; |
2551 | } |
2552 | } |
2553 | |
2554 | //------------------------------------------------------------------------ |
2555 | // lvaSetStructUsedAsVarArg: update hfa information for vararg struct args |
2556 | // |
2557 | // Arguments: |
2558 | // varNum -- number of the variable |
2559 | // |
2560 | // Notes: |
2561 | // This only affects arm64 varargs on windows where we need to pass |
2562 | // hfa arguments as if they are not HFAs. |
2563 | // |
2564 | // This function should only be called if the struct is used in a varargs |
2565 | // method. |
2566 | |
2567 | void Compiler::lvaSetStructUsedAsVarArg(unsigned varNum) |
2568 | { |
2569 | #ifdef FEATURE_HFA |
2570 | #if defined(_TARGET_WINDOWS_) && defined(_TARGET_ARM64_) |
2571 | LclVarDsc* varDsc = &lvaTable[varNum]; |
2572 | // For varargs methods incoming and outgoing arguments should not be treated |
2573 | // as HFA. |
2574 | varDsc->_lvIsHfa = false; |
2575 | varDsc->_lvHfaTypeIsFloat = false; |
2576 | #endif // defined(_TARGET_WINDOWS_) && defined(_TARGET_ARM64_) |
2577 | #endif // FEATURE_HFA |
2578 | } |
2579 | |
2580 | //------------------------------------------------------------------------ |
2581 | // lvaSetClass: set class information for a local var. |
2582 | // |
2583 | // Arguments: |
2584 | // varNum -- number of the variable |
2585 | // clsHnd -- class handle to use in set or update |
2586 | // isExact -- true if class is known exactly |
2587 | // |
2588 | // Notes: |
2589 | // varNum must not already have a ref class handle. |
2590 | |
2591 | void Compiler::lvaSetClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact) |
2592 | { |
2593 | noway_assert(varNum < lvaCount); |
2594 | |
2595 | // If we are just importing, we cannot reliably track local ref types, |
2596 | // since the jit maps CORINFO_TYPE_VAR to TYP_REF. |
2597 | if (compIsForImportOnly()) |
2598 | { |
2599 | return; |
2600 | } |
2601 | |
2602 | // Else we should have a type handle. |
2603 | assert(clsHnd != nullptr); |
2604 | |
2605 | LclVarDsc* varDsc = &lvaTable[varNum]; |
2606 | assert(varDsc->lvType == TYP_REF); |
2607 | |
2608 | // We shoud not have any ref type information for this var. |
2609 | assert(varDsc->lvClassHnd == nullptr); |
2610 | assert(!varDsc->lvClassIsExact); |
2611 | |
2612 | JITDUMP("\nlvaSetClass: setting class for V%02i to (%p) %s %s\n" , varNum, dspPtr(clsHnd), |
2613 | info.compCompHnd->getClassName(clsHnd), isExact ? " [exact]" : "" ); |
2614 | |
2615 | varDsc->lvClassHnd = clsHnd; |
2616 | varDsc->lvClassIsExact = isExact; |
2617 | } |
2618 | |
2619 | //------------------------------------------------------------------------ |
2620 | // lvaSetClass: set class information for a local var from a tree or stack type |
2621 | // |
2622 | // Arguments: |
2623 | // varNum -- number of the variable. Must be a single def local |
2624 | // tree -- tree establishing the variable's value |
2625 | // stackHnd -- handle for the type from the evaluation stack |
2626 | // |
2627 | // Notes: |
2628 | // Preferentially uses the tree's type, when available. Since not all |
2629 | // tree kinds can track ref types, the stack type is used as a |
2630 | // fallback. |
2631 | |
2632 | void Compiler::lvaSetClass(unsigned varNum, GenTree* tree, CORINFO_CLASS_HANDLE stackHnd) |
2633 | { |
2634 | bool isExact = false; |
2635 | bool isNonNull = false; |
2636 | CORINFO_CLASS_HANDLE clsHnd = gtGetClassHandle(tree, &isExact, &isNonNull); |
2637 | |
2638 | if (clsHnd != nullptr) |
2639 | { |
2640 | lvaSetClass(varNum, clsHnd, isExact); |
2641 | } |
2642 | else if (stackHnd != nullptr) |
2643 | { |
2644 | lvaSetClass(varNum, stackHnd); |
2645 | } |
2646 | } |
2647 | |
2648 | //------------------------------------------------------------------------ |
2649 | // lvaUpdateClass: update class information for a local var. |
2650 | // |
2651 | // Arguments: |
2652 | // varNum -- number of the variable |
2653 | // clsHnd -- class handle to use in set or update |
2654 | // isExact -- true if class is known exactly |
2655 | // |
2656 | // Notes: |
2657 | // |
2658 | // This method models the type update rule for an assignment. |
2659 | // |
2660 | // Updates currently should only happen for single-def user args or |
2661 | // locals, when we are processing the expression actually being |
2662 | // used to initialize the local (or inlined arg). The update will |
2663 | // change the local from the declared type to the type of the |
2664 | // initial value. |
2665 | // |
2666 | // These updates should always *improve* what we know about the |
2667 | // type, that is making an inexact type exact, or changing a type |
2668 | // to some subtype. However the jit lacks precise type information |
2669 | // for shared code, so ensuring this is so is currently not |
2670 | // possible. |
2671 | |
2672 | void Compiler::lvaUpdateClass(unsigned varNum, CORINFO_CLASS_HANDLE clsHnd, bool isExact) |
2673 | { |
2674 | assert(varNum < lvaCount); |
2675 | |
2676 | // If we are just importing, we cannot reliably track local ref types, |
2677 | // since the jit maps CORINFO_TYPE_VAR to TYP_REF. |
2678 | if (compIsForImportOnly()) |
2679 | { |
2680 | return; |
2681 | } |
2682 | |
2683 | // Else we should have a class handle to consider |
2684 | assert(clsHnd != nullptr); |
2685 | |
2686 | LclVarDsc* varDsc = &lvaTable[varNum]; |
2687 | assert(varDsc->lvType == TYP_REF); |
2688 | |
2689 | // We should already have a class |
2690 | assert(varDsc->lvClassHnd != nullptr); |
2691 | |
2692 | // We should only be updating classes for single-def locals. |
2693 | assert(varDsc->lvSingleDef); |
2694 | |
2695 | // Now see if we should update. |
2696 | // |
2697 | // New information may not always be "better" so do some |
2698 | // simple analysis to decide if the update is worthwhile. |
2699 | const bool isNewClass = (clsHnd != varDsc->lvClassHnd); |
2700 | bool shouldUpdate = false; |
2701 | |
2702 | // Are we attempting to update the class? Only check this when we have |
2703 | // an new type and the existing class is inexact... we should not be |
2704 | // updating exact classes. |
2705 | if (!varDsc->lvClassIsExact && isNewClass) |
2706 | { |
2707 | // Todo: improve this analysis by adding a new jit interface method |
2708 | DWORD newAttrs = info.compCompHnd->getClassAttribs(clsHnd); |
2709 | DWORD oldAttrs = info.compCompHnd->getClassAttribs(varDsc->lvClassHnd); |
2710 | |
2711 | // Avoid funny things with __Canon by only merging if both shared or both unshared |
2712 | if ((newAttrs & CORINFO_FLG_SHAREDINST) == (oldAttrs & CORINFO_FLG_SHAREDINST)) |
2713 | { |
2714 | // If we merge types and we get back the old class, the new class is more |
2715 | // specific and we should update to it. |
2716 | CORINFO_CLASS_HANDLE mergeClass = info.compCompHnd->mergeClasses(clsHnd, varDsc->lvClassHnd); |
2717 | |
2718 | if (mergeClass == varDsc->lvClassHnd) |
2719 | { |
2720 | shouldUpdate = true; |
2721 | } |
2722 | } |
2723 | else if ((newAttrs & CORINFO_FLG_SHAREDINST) == 0) |
2724 | { |
2725 | // Update if we go from shared to unshared |
2726 | shouldUpdate = true; |
2727 | } |
2728 | } |
2729 | // Else are we attempting to update exactness? |
2730 | else if (isExact && !varDsc->lvClassIsExact && !isNewClass) |
2731 | { |
2732 | shouldUpdate = true; |
2733 | } |
2734 | |
2735 | #if DEBUG |
2736 | if (isNewClass || (isExact != varDsc->lvClassIsExact)) |
2737 | { |
2738 | JITDUMP("\nlvaUpdateClass:%s Updating class for V%02u" , shouldUpdate ? "" : " NOT" , varNum); |
2739 | JITDUMP(" from(%p) %s%s" , dspPtr(varDsc->lvClassHnd), info.compCompHnd->getClassName(varDsc->lvClassHnd), |
2740 | varDsc->lvClassIsExact ? " [exact]" : "" ); |
2741 | JITDUMP(" to(%p) %s%s\n" , dspPtr(clsHnd), info.compCompHnd->getClassName(clsHnd), isExact ? " [exact]" : "" ); |
2742 | } |
2743 | #endif // DEBUG |
2744 | |
2745 | if (shouldUpdate) |
2746 | { |
2747 | varDsc->lvClassHnd = clsHnd; |
2748 | varDsc->lvClassIsExact = isExact; |
2749 | |
2750 | #if DEBUG |
2751 | // Note we've modified the type... |
2752 | varDsc->lvClassInfoUpdated = true; |
2753 | #endif // DEBUG |
2754 | } |
2755 | |
2756 | return; |
2757 | } |
2758 | |
2759 | //------------------------------------------------------------------------ |
2760 | // lvaUpdateClass: Uupdate class information for a local var from a tree |
2761 | // or stack type |
2762 | // |
2763 | // Arguments: |
2764 | // varNum -- number of the variable. Must be a single def local |
2765 | // tree -- tree establishing the variable's value |
2766 | // stackHnd -- handle for the type from the evaluation stack |
2767 | // |
2768 | // Notes: |
2769 | // Preferentially uses the tree's type, when available. Since not all |
2770 | // tree kinds can track ref types, the stack type is used as a |
2771 | // fallback. |
2772 | |
2773 | void Compiler::lvaUpdateClass(unsigned varNum, GenTree* tree, CORINFO_CLASS_HANDLE stackHnd) |
2774 | { |
2775 | bool isExact = false; |
2776 | bool isNonNull = false; |
2777 | CORINFO_CLASS_HANDLE clsHnd = gtGetClassHandle(tree, &isExact, &isNonNull); |
2778 | |
2779 | if (clsHnd != nullptr) |
2780 | { |
2781 | lvaUpdateClass(varNum, clsHnd, isExact); |
2782 | } |
2783 | else if (stackHnd != nullptr) |
2784 | { |
2785 | lvaUpdateClass(varNum, stackHnd); |
2786 | } |
2787 | } |
2788 | |
2789 | /***************************************************************************** |
2790 | * Returns the array of BYTEs containing the GC layout information |
2791 | */ |
2792 | |
2793 | BYTE* Compiler::lvaGetGcLayout(unsigned varNum) |
2794 | { |
2795 | assert(varTypeIsStruct(lvaTable[varNum].lvType) && (lvaTable[varNum].lvExactSize >= TARGET_POINTER_SIZE)); |
2796 | |
2797 | return lvaTable[varNum].lvGcLayout; |
2798 | } |
2799 | |
2800 | //------------------------------------------------------------------------ |
2801 | // lvaLclSize: returns size of a local variable, in bytes |
2802 | // |
2803 | // Arguments: |
2804 | // varNum -- variable to query |
2805 | // |
2806 | // Returns: |
2807 | // Number of bytes needed on the frame for such a local. |
2808 | |
2809 | unsigned Compiler::lvaLclSize(unsigned varNum) |
2810 | { |
2811 | assert(varNum < lvaCount); |
2812 | |
2813 | var_types varType = lvaTable[varNum].TypeGet(); |
2814 | |
2815 | switch (varType) |
2816 | { |
2817 | case TYP_STRUCT: |
2818 | case TYP_BLK: |
2819 | return lvaTable[varNum].lvSize(); |
2820 | |
2821 | case TYP_LCLBLK: |
2822 | #if FEATURE_FIXED_OUT_ARGS |
2823 | // Note that this operation performs a read of a PhasedVar |
2824 | noway_assert(varNum == lvaOutgoingArgSpaceVar); |
2825 | return lvaOutgoingArgSpaceSize; |
2826 | #else // FEATURE_FIXED_OUT_ARGS |
2827 | assert(!"Unknown size" ); |
2828 | NO_WAY("Target doesn't support TYP_LCLBLK" ); |
2829 | |
2830 | // Keep prefast happy |
2831 | __fallthrough; |
2832 | |
2833 | #endif // FEATURE_FIXED_OUT_ARGS |
2834 | |
2835 | default: // This must be a primitive var. Fall out of switch statement |
2836 | break; |
2837 | } |
2838 | #ifdef _TARGET_64BIT_ |
2839 | // We only need this Quirk for _TARGET_64BIT_ |
2840 | if (lvaTable[varNum].lvQuirkToLong) |
2841 | { |
2842 | noway_assert(lvaTable[varNum].lvAddrExposed); |
2843 | return genTypeStSz(TYP_LONG) * sizeof(int); // return 8 (2 * 4) |
2844 | } |
2845 | #endif |
2846 | return genTypeStSz(varType) * sizeof(int); |
2847 | } |
2848 | |
2849 | // |
2850 | // Return the exact width of local variable "varNum" -- the number of bytes |
2851 | // you'd need to copy in order to overwrite the value. |
2852 | // |
2853 | unsigned Compiler::lvaLclExactSize(unsigned varNum) |
2854 | { |
2855 | assert(varNum < lvaCount); |
2856 | |
2857 | var_types varType = lvaTable[varNum].TypeGet(); |
2858 | |
2859 | switch (varType) |
2860 | { |
2861 | case TYP_STRUCT: |
2862 | case TYP_BLK: |
2863 | return lvaTable[varNum].lvExactSize; |
2864 | |
2865 | case TYP_LCLBLK: |
2866 | #if FEATURE_FIXED_OUT_ARGS |
2867 | // Note that this operation performs a read of a PhasedVar |
2868 | noway_assert(lvaOutgoingArgSpaceSize >= 0); |
2869 | noway_assert(varNum == lvaOutgoingArgSpaceVar); |
2870 | return lvaOutgoingArgSpaceSize; |
2871 | |
2872 | #else // FEATURE_FIXED_OUT_ARGS |
2873 | assert(!"Unknown size" ); |
2874 | NO_WAY("Target doesn't support TYP_LCLBLK" ); |
2875 | |
2876 | // Keep prefast happy |
2877 | __fallthrough; |
2878 | |
2879 | #endif // FEATURE_FIXED_OUT_ARGS |
2880 | |
2881 | default: // This must be a primitive var. Fall out of switch statement |
2882 | break; |
2883 | } |
2884 | |
2885 | return genTypeSize(varType); |
2886 | } |
2887 | |
2888 | // getCalledCount -- get the value used to normalized weights for this method |
2889 | // if we don't have profile data then getCalledCount will return BB_UNITY_WEIGHT (100) |
2890 | // otherwise it returns the number of times that profile data says the method was called. |
2891 | // |
2892 | BasicBlock::weight_t BasicBlock::getCalledCount(Compiler* comp) |
2893 | { |
2894 | // when we don't have profile data then fgCalledCount will be BB_UNITY_WEIGHT (100) |
2895 | BasicBlock::weight_t calledCount = comp->fgCalledCount; |
2896 | |
2897 | // If we haven't yet reach the place where we setup fgCalledCount it could still be zero |
2898 | // so return a reasonable value to use until we set it. |
2899 | // |
2900 | if (calledCount == 0) |
2901 | { |
2902 | if (comp->fgIsUsingProfileWeights()) |
2903 | { |
2904 | // When we use profile data block counts we have exact counts, |
2905 | // not multiples of BB_UNITY_WEIGHT (100) |
2906 | calledCount = 1; |
2907 | } |
2908 | else |
2909 | { |
2910 | calledCount = comp->fgFirstBB->bbWeight; |
2911 | |
2912 | if (calledCount == 0) |
2913 | { |
2914 | calledCount = BB_UNITY_WEIGHT; |
2915 | } |
2916 | } |
2917 | } |
2918 | return calledCount; |
2919 | } |
2920 | |
2921 | // getBBWeight -- get the normalized weight of this block |
2922 | BasicBlock::weight_t BasicBlock::getBBWeight(Compiler* comp) |
2923 | { |
2924 | if (this->bbWeight == 0) |
2925 | { |
2926 | return 0; |
2927 | } |
2928 | else |
2929 | { |
2930 | weight_t calledCount = getCalledCount(comp); |
2931 | |
2932 | // Normalize the bbWeights by multiplying by BB_UNITY_WEIGHT and dividing by the calledCount. |
2933 | // |
2934 | // 1. For methods that do not have IBC data the called weight will always be 100 (BB_UNITY_WEIGHT) |
2935 | // and the entry point bbWeight value is almost always 100 (BB_UNITY_WEIGHT) |
2936 | // 2. For methods that do have IBC data the called weight is the actual number of calls |
2937 | // from the IBC data and the entry point bbWeight value is almost always the actual |
2938 | // number of calls from the IBC data. |
2939 | // |
2940 | // "almost always" - except for the rare case where a loop backedge jumps to BB01 |
2941 | // |
2942 | // We also perform a rounding operation by adding half of the 'calledCount' before performing |
2943 | // the division. |
2944 | // |
2945 | // Thus for both cases we will return 100 (BB_UNITY_WEIGHT) for the entry point BasicBlock |
2946 | // |
2947 | // Note that with a 100 (BB_UNITY_WEIGHT) values between 1 and 99 represent decimal fractions. |
2948 | // (i.e. 33 represents 33% and 75 represents 75%, and values greater than 100 require |
2949 | // some kind of loop backedge) |
2950 | // |
2951 | |
2952 | if (this->bbWeight < (BB_MAX_WEIGHT / BB_UNITY_WEIGHT)) |
2953 | { |
2954 | // Calculate the result using unsigned arithmetic |
2955 | weight_t result = ((this->bbWeight * BB_UNITY_WEIGHT) + (calledCount / 2)) / calledCount; |
2956 | |
2957 | // We don't allow a value of zero, as that would imply rarely run |
2958 | return max(1, result); |
2959 | } |
2960 | else |
2961 | { |
2962 | // Calculate the full result using floating point |
2963 | double fullResult = ((double)this->bbWeight * (double)BB_UNITY_WEIGHT) / (double)calledCount; |
2964 | |
2965 | if (fullResult < (double)BB_MAX_WEIGHT) |
2966 | { |
2967 | // Add 0.5 and truncate to unsigned |
2968 | return (weight_t)(fullResult + 0.5); |
2969 | } |
2970 | else |
2971 | { |
2972 | return BB_MAX_WEIGHT; |
2973 | } |
2974 | } |
2975 | } |
2976 | } |
2977 | |
2978 | /***************************************************************************** |
2979 | * |
2980 | * Compare function passed to qsort() by Compiler::lclVars.lvaSortByRefCount(). |
2981 | * when generating SMALL_CODE. |
2982 | * Return positive if dsc2 has a higher ref count |
2983 | * Return negative if dsc1 has a higher ref count |
2984 | * Return zero if the ref counts are the same |
2985 | */ |
2986 | |
2987 | /* static */ |
2988 | int __cdecl Compiler::RefCntCmp(const void* op1, const void* op2) |
2989 | { |
2990 | LclVarDsc* dsc1 = *(LclVarDsc**)op1; |
2991 | LclVarDsc* dsc2 = *(LclVarDsc**)op2; |
2992 | |
2993 | /* Make sure we preference tracked variables over untracked variables */ |
2994 | |
2995 | if (dsc1->lvTracked != dsc2->lvTracked) |
2996 | { |
2997 | return (dsc2->lvTracked) ? +1 : -1; |
2998 | } |
2999 | |
3000 | unsigned weight1 = dsc1->lvRefCnt(); |
3001 | unsigned weight2 = dsc2->lvRefCnt(); |
3002 | |
3003 | #ifndef _TARGET_ARM_ |
3004 | // ARM-TODO: this was disabled for ARM under !FEATURE_FP_REGALLOC; it was probably a left-over from |
3005 | // legacy backend. It should be enabled and verified. |
3006 | |
3007 | /* Force integer candidates to sort above float candidates */ |
3008 | |
3009 | bool isFloat1 = isFloatRegType(dsc1->lvType); |
3010 | bool isFloat2 = isFloatRegType(dsc2->lvType); |
3011 | |
3012 | if (isFloat1 != isFloat2) |
3013 | { |
3014 | if (weight2 && isFloat1) |
3015 | { |
3016 | return +1; |
3017 | } |
3018 | if (weight1 && isFloat2) |
3019 | { |
3020 | return -1; |
3021 | } |
3022 | } |
3023 | #endif |
3024 | |
3025 | int diff = weight2 - weight1; |
3026 | |
3027 | if (diff != 0) |
3028 | { |
3029 | return diff; |
3030 | } |
3031 | |
3032 | /* The unweighted ref counts were the same */ |
3033 | /* If the weighted ref counts are different then use their difference */ |
3034 | diff = dsc2->lvRefCntWtd() - dsc1->lvRefCntWtd(); |
3035 | |
3036 | if (diff != 0) |
3037 | { |
3038 | return diff; |
3039 | } |
3040 | |
3041 | /* We have equal ref counts and weighted ref counts */ |
3042 | |
3043 | /* Break the tie by: */ |
3044 | /* Increasing the weight by 2 if we are a register arg */ |
3045 | /* Increasing the weight by 0.5 if we are a GC type */ |
3046 | /* Increasing the weight by 0.5 if we were enregistered in the previous pass */ |
3047 | |
3048 | if (weight1) |
3049 | { |
3050 | if (dsc1->lvIsRegArg) |
3051 | { |
3052 | weight2 += 2 * BB_UNITY_WEIGHT; |
3053 | } |
3054 | |
3055 | if (varTypeIsGC(dsc1->TypeGet())) |
3056 | { |
3057 | weight1 += BB_UNITY_WEIGHT / 2; |
3058 | } |
3059 | |
3060 | if (dsc1->lvRegister) |
3061 | { |
3062 | weight1 += BB_UNITY_WEIGHT / 2; |
3063 | } |
3064 | } |
3065 | |
3066 | if (weight2) |
3067 | { |
3068 | if (dsc2->lvIsRegArg) |
3069 | { |
3070 | weight2 += 2 * BB_UNITY_WEIGHT; |
3071 | } |
3072 | |
3073 | if (varTypeIsGC(dsc2->TypeGet())) |
3074 | { |
3075 | weight2 += BB_UNITY_WEIGHT / 2; |
3076 | } |
3077 | |
3078 | if (dsc2->lvRegister) |
3079 | { |
3080 | weight2 += BB_UNITY_WEIGHT / 2; |
3081 | } |
3082 | } |
3083 | |
3084 | diff = weight2 - weight1; |
3085 | |
3086 | if (diff != 0) |
3087 | { |
3088 | return diff; |
3089 | } |
3090 | |
3091 | /* To achieve a Stable Sort we use the LclNum (by way of the pointer address) */ |
3092 | |
3093 | if (dsc1 < dsc2) |
3094 | { |
3095 | return -1; |
3096 | } |
3097 | if (dsc1 > dsc2) |
3098 | { |
3099 | return +1; |
3100 | } |
3101 | |
3102 | return 0; |
3103 | } |
3104 | |
3105 | /***************************************************************************** |
3106 | * |
3107 | * Compare function passed to qsort() by Compiler::lclVars.lvaSortByRefCount(). |
3108 | * when not generating SMALL_CODE. |
3109 | * Return positive if dsc2 has a higher weighted ref count |
3110 | * Return negative if dsc1 has a higher weighted ref count |
3111 | * Return zero if the ref counts are the same |
3112 | */ |
3113 | |
3114 | /* static */ |
3115 | int __cdecl Compiler::WtdRefCntCmp(const void* op1, const void* op2) |
3116 | { |
3117 | LclVarDsc* dsc1 = *(LclVarDsc**)op1; |
3118 | LclVarDsc* dsc2 = *(LclVarDsc**)op2; |
3119 | |
3120 | /* Make sure we preference tracked variables over untracked variables */ |
3121 | |
3122 | if (dsc1->lvTracked != dsc2->lvTracked) |
3123 | { |
3124 | return (dsc2->lvTracked) ? +1 : -1; |
3125 | } |
3126 | |
3127 | unsigned weight1 = dsc1->lvRefCntWtd(); |
3128 | unsigned weight2 = dsc2->lvRefCntWtd(); |
3129 | |
3130 | #ifndef _TARGET_ARM_ |
3131 | // ARM-TODO: this was disabled for ARM under !FEATURE_FP_REGALLOC; it was probably a left-over from |
3132 | // legacy backend. It should be enabled and verified. |
3133 | |
3134 | /* Force integer candidates to sort above float candidates */ |
3135 | |
3136 | bool isFloat1 = isFloatRegType(dsc1->lvType); |
3137 | bool isFloat2 = isFloatRegType(dsc2->lvType); |
3138 | |
3139 | if (isFloat1 != isFloat2) |
3140 | { |
3141 | if (weight2 && isFloat1) |
3142 | { |
3143 | return +1; |
3144 | } |
3145 | if (weight1 && isFloat2) |
3146 | { |
3147 | return -1; |
3148 | } |
3149 | } |
3150 | #endif |
3151 | |
3152 | if (weight1 && dsc1->lvIsRegArg) |
3153 | { |
3154 | weight1 += 2 * BB_UNITY_WEIGHT; |
3155 | } |
3156 | |
3157 | if (weight2 && dsc2->lvIsRegArg) |
3158 | { |
3159 | weight2 += 2 * BB_UNITY_WEIGHT; |
3160 | } |
3161 | |
3162 | if (weight2 > weight1) |
3163 | { |
3164 | return 1; |
3165 | } |
3166 | else if (weight2 < weight1) |
3167 | { |
3168 | return -1; |
3169 | } |
3170 | |
3171 | // Otherwise, we have equal weighted ref counts. |
3172 | |
3173 | /* If the unweighted ref counts are different then use their difference */ |
3174 | int diff = (int)dsc2->lvRefCnt() - (int)dsc1->lvRefCnt(); |
3175 | |
3176 | if (diff != 0) |
3177 | { |
3178 | return diff; |
3179 | } |
3180 | |
3181 | /* If one is a GC type and the other is not the GC type wins */ |
3182 | if (varTypeIsGC(dsc1->TypeGet()) != varTypeIsGC(dsc2->TypeGet())) |
3183 | { |
3184 | if (varTypeIsGC(dsc1->TypeGet())) |
3185 | { |
3186 | diff = -1; |
3187 | } |
3188 | else |
3189 | { |
3190 | diff = +1; |
3191 | } |
3192 | |
3193 | return diff; |
3194 | } |
3195 | |
3196 | /* If one was enregistered in the previous pass then it wins */ |
3197 | if (dsc1->lvRegister != dsc2->lvRegister) |
3198 | { |
3199 | if (dsc1->lvRegister) |
3200 | { |
3201 | diff = -1; |
3202 | } |
3203 | else |
3204 | { |
3205 | diff = +1; |
3206 | } |
3207 | |
3208 | return diff; |
3209 | } |
3210 | |
3211 | /* We have a tie! */ |
3212 | |
3213 | /* To achieve a Stable Sort we use the LclNum (by way of the pointer address) */ |
3214 | |
3215 | if (dsc1 < dsc2) |
3216 | { |
3217 | return -1; |
3218 | } |
3219 | if (dsc1 > dsc2) |
3220 | { |
3221 | return +1; |
3222 | } |
3223 | |
3224 | return 0; |
3225 | } |
3226 | |
3227 | /***************************************************************************** |
3228 | * |
3229 | * Sort the local variable table by refcount and assign tracking indices. |
3230 | */ |
3231 | |
3232 | void Compiler::lvaSortOnly() |
3233 | { |
3234 | /* Now sort the variable table by ref-count */ |
3235 | |
3236 | qsort(lvaRefSorted, lvaCount, sizeof(*lvaRefSorted), (compCodeOpt() == SMALL_CODE) ? RefCntCmp : WtdRefCntCmp); |
3237 | lvaDumpRefCounts(); |
3238 | } |
3239 | |
3240 | void Compiler::lvaDumpRefCounts() |
3241 | { |
3242 | #ifdef DEBUG |
3243 | |
3244 | if (verbose && lvaCount) |
3245 | { |
3246 | printf("refCnt table for '%s':\n" , info.compMethodName); |
3247 | |
3248 | for (unsigned lclNum = 0; lclNum < lvaCount; lclNum++) |
3249 | { |
3250 | unsigned refCnt = lvaRefSorted[lclNum]->lvRefCnt(); |
3251 | if (refCnt == 0) |
3252 | { |
3253 | break; |
3254 | } |
3255 | unsigned refCntWtd = lvaRefSorted[lclNum]->lvRefCntWtd(); |
3256 | |
3257 | printf(" " ); |
3258 | gtDispLclVar((unsigned)(lvaRefSorted[lclNum] - lvaTable)); |
3259 | printf(" [%6s]: refCnt = %4u, refCntWtd = %6s" , varTypeName(lvaRefSorted[lclNum]->TypeGet()), refCnt, |
3260 | refCntWtd2str(refCntWtd)); |
3261 | printf("\n" ); |
3262 | } |
3263 | |
3264 | printf("\n" ); |
3265 | } |
3266 | |
3267 | #endif |
3268 | } |
3269 | |
3270 | /***************************************************************************** |
3271 | * |
3272 | * Sort the local variable table by refcount and assign tracking indices. |
3273 | */ |
3274 | |
3275 | void Compiler::lvaSortByRefCount() |
3276 | { |
3277 | lvaTrackedCount = 0; |
3278 | lvaTrackedCountInSizeTUnits = 0; |
3279 | |
3280 | #ifdef DEBUG |
3281 | VarSetOps::AssignNoCopy(this, lvaTrackedVars, VarSetOps::MakeEmpty(this)); |
3282 | #endif |
3283 | |
3284 | if (lvaCount == 0) |
3285 | { |
3286 | return; |
3287 | } |
3288 | |
3289 | unsigned lclNum; |
3290 | LclVarDsc* varDsc; |
3291 | |
3292 | LclVarDsc** refTab; |
3293 | |
3294 | /* We'll sort the variables by ref count - allocate the sorted table */ |
3295 | |
3296 | lvaRefSorted = refTab = new (this, CMK_LvaTable) LclVarDsc*[lvaCount]; |
3297 | |
3298 | /* Fill in the table used for sorting */ |
3299 | |
3300 | for (lclNum = 0, varDsc = lvaTable; lclNum < lvaCount; lclNum++, varDsc++) |
3301 | { |
3302 | /* Append this variable to the table for sorting */ |
3303 | |
3304 | *refTab++ = varDsc; |
3305 | |
3306 | /* For now assume we'll be able to track all locals */ |
3307 | |
3308 | varDsc->lvTracked = 1; |
3309 | |
3310 | /* If the ref count is zero */ |
3311 | if (varDsc->lvRefCnt() == 0) |
3312 | { |
3313 | /* Zero ref count, make this untracked */ |
3314 | varDsc->lvTracked = 0; |
3315 | varDsc->setLvRefCntWtd(0); |
3316 | } |
3317 | |
3318 | #if !defined(_TARGET_64BIT_) |
3319 | if (varTypeIsLong(varDsc) && varDsc->lvPromoted) |
3320 | { |
3321 | varDsc->lvTracked = 0; |
3322 | } |
3323 | #endif // !defined(_TARGET_64BIT_) |
3324 | |
3325 | // Variables that are address-exposed, and all struct locals, are never enregistered, or tracked. |
3326 | // (The struct may be promoted, and its field variables enregistered/tracked, or the VM may "normalize" |
3327 | // its type so that its not seen by the JIT as a struct.) |
3328 | // Pinned variables may not be tracked (a condition of the GCInfo representation) |
3329 | // or enregistered, on x86 -- it is believed that we can enregister pinned (more properly, "pinning") |
3330 | // references when using the general GC encoding. |
3331 | if (varDsc->lvAddrExposed) |
3332 | { |
3333 | varDsc->lvTracked = 0; |
3334 | assert(varDsc->lvType != TYP_STRUCT || |
3335 | varDsc->lvDoNotEnregister); // For structs, should have set this when we set lvAddrExposed. |
3336 | } |
3337 | else if (varTypeIsStruct(varDsc)) |
3338 | { |
3339 | // Promoted structs will never be considered for enregistration anyway, |
3340 | // and the DoNotEnregister flag was used to indicate whether promotion was |
3341 | // independent or dependent. |
3342 | if (varDsc->lvPromoted) |
3343 | { |
3344 | varDsc->lvTracked = 0; |
3345 | } |
3346 | else if ((varDsc->lvType == TYP_STRUCT) && !varDsc->lvRegStruct) |
3347 | { |
3348 | lvaSetVarDoNotEnregister(lclNum DEBUGARG(DNER_IsStruct)); |
3349 | } |
3350 | } |
3351 | else if (varDsc->lvIsStructField && (lvaGetParentPromotionType(lclNum) != PROMOTION_TYPE_INDEPENDENT)) |
3352 | { |
3353 | // SSA must exclude struct fields that are not independently promoted |
3354 | // as dependent fields could be assigned using a CopyBlock |
3355 | // resulting in a single node causing multiple SSA definitions |
3356 | // which isn't currently supported by SSA |
3357 | // |
3358 | // TODO-CQ: Consider using lvLclBlockOpAddr and only marking these LclVars |
3359 | // untracked when a blockOp is used to assign the struct. |
3360 | // |
3361 | varDsc->lvTracked = 0; // so, don't mark as tracked |
3362 | lvaSetVarDoNotEnregister(lclNum DEBUGARG(DNER_DepField)); |
3363 | } |
3364 | else if (varDsc->lvPinned) |
3365 | { |
3366 | varDsc->lvTracked = 0; |
3367 | #ifdef JIT32_GCENCODER |
3368 | lvaSetVarDoNotEnregister(lclNum DEBUGARG(DNER_PinningRef)); |
3369 | #endif |
3370 | } |
3371 | else if (opts.MinOpts() && !JitConfig.JitMinOptsTrackGCrefs() && varTypeIsGC(varDsc->TypeGet())) |
3372 | { |
3373 | varDsc->lvTracked = 0; |
3374 | lvaSetVarDoNotEnregister(lclNum DEBUGARG(DNER_MinOptsGC)); |
3375 | } |
3376 | else if ((opts.compFlags & CLFLG_REGVAR) == 0) |
3377 | { |
3378 | lvaSetVarDoNotEnregister(lclNum DEBUGARG(DNER_NoRegVars)); |
3379 | } |
3380 | #if defined(JIT32_GCENCODER) && defined(WIN64EXCEPTIONS) |
3381 | else if (lvaIsOriginalThisArg(lclNum) && (info.compMethodInfo->options & CORINFO_GENERICS_CTXT_FROM_THIS) != 0) |
3382 | { |
3383 | // For x86/Linux, we need to track "this". |
3384 | // However we cannot have it in tracked variables, so we set "this" pointer always untracked |
3385 | varDsc->lvTracked = 0; |
3386 | } |
3387 | #endif |
3388 | |
3389 | // Are we not optimizing and we have exception handlers? |
3390 | // if so mark all args and locals "do not enregister". |
3391 | // |
3392 | if (opts.MinOpts() && compHndBBtabCount > 0) |
3393 | { |
3394 | lvaSetVarDoNotEnregister(lclNum DEBUGARG(DNER_LiveInOutOfHandler)); |
3395 | continue; |
3396 | } |
3397 | |
3398 | var_types type = genActualType(varDsc->TypeGet()); |
3399 | |
3400 | switch (type) |
3401 | { |
3402 | #if CPU_HAS_FP_SUPPORT |
3403 | case TYP_FLOAT: |
3404 | case TYP_DOUBLE: |
3405 | #endif |
3406 | case TYP_INT: |
3407 | case TYP_LONG: |
3408 | case TYP_REF: |
3409 | case TYP_BYREF: |
3410 | #ifdef FEATURE_SIMD |
3411 | case TYP_SIMD8: |
3412 | case TYP_SIMD12: |
3413 | case TYP_SIMD16: |
3414 | case TYP_SIMD32: |
3415 | #endif // FEATURE_SIMD |
3416 | case TYP_STRUCT: |
3417 | break; |
3418 | |
3419 | case TYP_UNDEF: |
3420 | case TYP_UNKNOWN: |
3421 | noway_assert(!"lvType not set correctly" ); |
3422 | varDsc->lvType = TYP_INT; |
3423 | |
3424 | __fallthrough; |
3425 | |
3426 | default: |
3427 | varDsc->lvTracked = 0; |
3428 | } |
3429 | } |
3430 | |
3431 | /* Now sort the variable table by ref-count */ |
3432 | |
3433 | lvaSortOnly(); |
3434 | |
3435 | /* Decide which variables will be worth tracking */ |
3436 | |
3437 | if (lvaCount > lclMAX_TRACKED) |
3438 | { |
3439 | /* Mark all variables past the first 'lclMAX_TRACKED' as untracked */ |
3440 | |
3441 | for (lclNum = lclMAX_TRACKED; lclNum < lvaCount; lclNum++) |
3442 | { |
3443 | lvaRefSorted[lclNum]->lvTracked = 0; |
3444 | } |
3445 | } |
3446 | |
3447 | if (lvaTrackedToVarNum == nullptr) |
3448 | { |
3449 | lvaTrackedToVarNum = new (getAllocator(CMK_LvaTable)) unsigned[lclMAX_TRACKED]; |
3450 | } |
3451 | |
3452 | #ifdef DEBUG |
3453 | // Re-Initialize to -1 for safety in debug build. |
3454 | memset(lvaTrackedToVarNum, -1, lclMAX_TRACKED * sizeof(unsigned)); |
3455 | #endif |
3456 | |
3457 | /* Assign indices to all the variables we've decided to track */ |
3458 | |
3459 | for (lclNum = 0; lclNum < min(lvaCount, lclMAX_TRACKED); lclNum++) |
3460 | { |
3461 | varDsc = lvaRefSorted[lclNum]; |
3462 | if (varDsc->lvTracked) |
3463 | { |
3464 | noway_assert(varDsc->lvRefCnt() > 0); |
3465 | |
3466 | /* This variable will be tracked - assign it an index */ |
3467 | |
3468 | lvaTrackedToVarNum[lvaTrackedCount] = (unsigned)(varDsc - lvaTable); // The type of varDsc and lvaTable |
3469 | // is LclVarDsc. Subtraction will give us |
3470 | // the index. |
3471 | varDsc->lvVarIndex = lvaTrackedCount++; |
3472 | } |
3473 | } |
3474 | |
3475 | // We have a new epoch, and also cache the tracked var count in terms of size_t's sufficient to hold that many bits. |
3476 | lvaCurEpoch++; |
3477 | lvaTrackedCountInSizeTUnits = |
3478 | roundUp((unsigned)lvaTrackedCount, (unsigned)(sizeof(size_t) * 8)) / unsigned(sizeof(size_t) * 8); |
3479 | |
3480 | #ifdef DEBUG |
3481 | VarSetOps::AssignNoCopy(this, lvaTrackedVars, VarSetOps::MakeFull(this)); |
3482 | #endif |
3483 | } |
3484 | |
3485 | #if ASSERTION_PROP |
3486 | /***************************************************************************** |
3487 | * |
3488 | * This is called by lvaMarkLclRefs to disqualify a variable from being |
3489 | * considered by optAddCopies() |
3490 | */ |
3491 | void LclVarDsc::lvaDisqualifyVar() |
3492 | { |
3493 | this->lvDisqualify = true; |
3494 | this->lvSingleDef = false; |
3495 | this->lvDefStmt = nullptr; |
3496 | } |
3497 | #endif // ASSERTION_PROP |
3498 | |
3499 | /********************************************************************************** |
3500 | * Get stack size of the varDsc. |
3501 | */ |
3502 | size_t LclVarDsc::lvArgStackSize() const |
3503 | { |
3504 | // Make sure this will have a stack size |
3505 | assert(!this->lvIsRegArg); |
3506 | |
3507 | size_t stackSize = 0; |
3508 | if (varTypeIsStruct(this)) |
3509 | { |
3510 | #if defined(WINDOWS_AMD64_ABI) |
3511 | // Structs are either passed by reference or can be passed by value using one pointer |
3512 | stackSize = TARGET_POINTER_SIZE; |
3513 | #elif defined(_TARGET_ARM64_) || defined(UNIX_AMD64_ABI) |
3514 | // lvSize performs a roundup. |
3515 | stackSize = this->lvSize(); |
3516 | |
3517 | #if defined(_TARGET_ARM64_) |
3518 | if ((stackSize > TARGET_POINTER_SIZE * 2) && (!this->lvIsHfa())) |
3519 | { |
3520 | // If the size is greater than 16 bytes then it will |
3521 | // be passed by reference. |
3522 | stackSize = TARGET_POINTER_SIZE; |
3523 | } |
3524 | #endif // defined(_TARGET_ARM64_) |
3525 | |
3526 | #else // !_TARGET_ARM64_ !WINDOWS_AMD64_ABI !UNIX_AMD64_ABI |
3527 | |
3528 | NYI("Unsupported target." ); |
3529 | unreached(); |
3530 | |
3531 | #endif // !_TARGET_ARM64_ !WINDOWS_AMD64_ABI !UNIX_AMD64_ABI |
3532 | } |
3533 | else |
3534 | { |
3535 | stackSize = TARGET_POINTER_SIZE; |
3536 | } |
3537 | |
3538 | return stackSize; |
3539 | } |
3540 | |
3541 | /********************************************************************************** |
3542 | * Get type of a variable when passed as an argument. |
3543 | */ |
3544 | var_types LclVarDsc::lvaArgType() |
3545 | { |
3546 | var_types type = TypeGet(); |
3547 | |
3548 | #ifdef _TARGET_AMD64_ |
3549 | #ifdef UNIX_AMD64_ABI |
3550 | if (type == TYP_STRUCT) |
3551 | { |
3552 | NYI("lvaArgType" ); |
3553 | } |
3554 | #else //! UNIX_AMD64_ABI |
3555 | if (type == TYP_STRUCT) |
3556 | { |
3557 | switch (lvExactSize) |
3558 | { |
3559 | case 1: |
3560 | type = TYP_BYTE; |
3561 | break; |
3562 | case 2: |
3563 | type = TYP_SHORT; |
3564 | break; |
3565 | case 4: |
3566 | type = TYP_INT; |
3567 | break; |
3568 | case 8: |
3569 | switch (*lvGcLayout) |
3570 | { |
3571 | case TYPE_GC_NONE: |
3572 | type = TYP_I_IMPL; |
3573 | break; |
3574 | |
3575 | case TYPE_GC_REF: |
3576 | type = TYP_REF; |
3577 | break; |
3578 | |
3579 | case TYPE_GC_BYREF: |
3580 | type = TYP_BYREF; |
3581 | break; |
3582 | |
3583 | default: |
3584 | unreached(); |
3585 | } |
3586 | break; |
3587 | |
3588 | default: |
3589 | type = TYP_BYREF; |
3590 | break; |
3591 | } |
3592 | } |
3593 | #endif // !UNIX_AMD64_ABI |
3594 | #elif defined(_TARGET_ARM64_) |
3595 | if (type == TYP_STRUCT) |
3596 | { |
3597 | NYI("lvaArgType" ); |
3598 | } |
3599 | #elif defined(_TARGET_X86_) |
3600 | // Nothing to do; use the type as is. |
3601 | #else |
3602 | NYI("lvaArgType" ); |
3603 | #endif //_TARGET_AMD64_ |
3604 | |
3605 | return type; |
3606 | } |
3607 | |
3608 | //------------------------------------------------------------------------ |
3609 | // lvaMarkLclRefs: increment local var references counts and more |
3610 | // |
3611 | // Arguments: |
3612 | // tree - some node in a tree |
3613 | // block - block that the tree node belongs to |
3614 | // stmt - stmt that the tree node belongs to |
3615 | // isRecompute - true if we should just recompute counts |
3616 | // |
3617 | // Notes: |
3618 | // Invoked via the MarkLocalVarsVisitor |
3619 | // |
3620 | // Primarily increments the regular and weighted local var ref |
3621 | // counts for any local referred to directly by tree. |
3622 | // |
3623 | // Also: |
3624 | // |
3625 | // Accounts for implicit references to frame list root for |
3626 | // pinvokes that will be expanded later. |
3627 | // |
3628 | // Determines if locals of TYP_BOOL can safely be considered |
3629 | // to hold only 0 or 1 or may have a broader range of true values. |
3630 | // |
3631 | // Does some setup work for assertion prop, noting locals that are |
3632 | // eligible for assertion prop, single defs, and tracking which blocks |
3633 | // hold uses. |
3634 | // |
3635 | // In checked builds: |
3636 | // |
3637 | // Verifies that local accesses are consistenly typed. |
3638 | // Verifies that casts remain in bounds. |
3639 | |
3640 | void Compiler::lvaMarkLclRefs(GenTree* tree, BasicBlock* block, GenTreeStmt* stmt, bool isRecompute) |
3641 | { |
3642 | const BasicBlock::weight_t weight = block->getBBWeight(this); |
3643 | |
3644 | /* Is this a call to unmanaged code ? */ |
3645 | if (tree->gtOper == GT_CALL && tree->gtFlags & GTF_CALL_UNMANAGED) |
3646 | { |
3647 | assert((!opts.ShouldUsePInvokeHelpers()) || (info.compLvFrameListRoot == BAD_VAR_NUM)); |
3648 | if (!opts.ShouldUsePInvokeHelpers()) |
3649 | { |
3650 | /* Get the special variable descriptor */ |
3651 | |
3652 | unsigned lclNum = info.compLvFrameListRoot; |
3653 | |
3654 | noway_assert(lclNum <= lvaCount); |
3655 | LclVarDsc* varDsc = lvaTable + lclNum; |
3656 | |
3657 | /* Increment the ref counts twice */ |
3658 | varDsc->incRefCnts(weight, this); |
3659 | varDsc->incRefCnts(weight, this); |
3660 | } |
3661 | } |
3662 | |
3663 | if (!isRecompute) |
3664 | { |
3665 | /* Is this an assigment? */ |
3666 | |
3667 | if (tree->OperIs(GT_ASG)) |
3668 | { |
3669 | GenTree* op1 = tree->gtOp.gtOp1; |
3670 | GenTree* op2 = tree->gtOp.gtOp2; |
3671 | |
3672 | #if OPT_BOOL_OPS |
3673 | |
3674 | /* Is this an assignment to a local variable? */ |
3675 | |
3676 | if (op1->gtOper == GT_LCL_VAR && op2->gtType != TYP_BOOL) |
3677 | { |
3678 | /* Only simple assignments allowed for booleans */ |
3679 | |
3680 | if (tree->gtOper != GT_ASG) |
3681 | { |
3682 | goto NOT_BOOL; |
3683 | } |
3684 | |
3685 | /* Is the RHS clearly a boolean value? */ |
3686 | |
3687 | switch (op2->gtOper) |
3688 | { |
3689 | unsigned lclNum; |
3690 | |
3691 | case GT_CNS_INT: |
3692 | |
3693 | if (op2->gtIntCon.gtIconVal == 0) |
3694 | { |
3695 | break; |
3696 | } |
3697 | if (op2->gtIntCon.gtIconVal == 1) |
3698 | { |
3699 | break; |
3700 | } |
3701 | |
3702 | // Not 0 or 1, fall through .... |
3703 | __fallthrough; |
3704 | |
3705 | default: |
3706 | |
3707 | if (op2->OperIsCompare()) |
3708 | { |
3709 | break; |
3710 | } |
3711 | |
3712 | NOT_BOOL: |
3713 | |
3714 | lclNum = op1->gtLclVarCommon.gtLclNum; |
3715 | noway_assert(lclNum < lvaCount); |
3716 | |
3717 | lvaTable[lclNum].lvIsBoolean = false; |
3718 | break; |
3719 | } |
3720 | } |
3721 | #endif |
3722 | } |
3723 | } |
3724 | |
3725 | if ((tree->gtOper != GT_LCL_VAR) && (tree->gtOper != GT_LCL_FLD)) |
3726 | { |
3727 | return; |
3728 | } |
3729 | |
3730 | /* This must be a local variable reference */ |
3731 | |
3732 | assert((tree->gtOper == GT_LCL_VAR) || (tree->gtOper == GT_LCL_FLD)); |
3733 | unsigned lclNum = tree->gtLclVarCommon.gtLclNum; |
3734 | |
3735 | noway_assert(lclNum < lvaCount); |
3736 | LclVarDsc* varDsc = lvaTable + lclNum; |
3737 | |
3738 | /* Increment the reference counts */ |
3739 | |
3740 | varDsc->incRefCnts(weight, this); |
3741 | |
3742 | if (!isRecompute) |
3743 | { |
3744 | if (lvaVarAddrExposed(lclNum)) |
3745 | { |
3746 | varDsc->lvIsBoolean = false; |
3747 | } |
3748 | |
3749 | if (tree->gtOper == GT_LCL_FLD) |
3750 | { |
3751 | #if ASSERTION_PROP |
3752 | // variables that have uses inside a GT_LCL_FLD |
3753 | // cause problems, so we will disqualify them here |
3754 | varDsc->lvaDisqualifyVar(); |
3755 | #endif // ASSERTION_PROP |
3756 | return; |
3757 | } |
3758 | |
3759 | #if ASSERTION_PROP |
3760 | if (fgDomsComputed && IsDominatedByExceptionalEntry(block)) |
3761 | { |
3762 | SetVolatileHint(varDsc); |
3763 | } |
3764 | |
3765 | /* Record if the variable has a single def or not */ |
3766 | |
3767 | if (!varDsc->lvDisqualify) // If this variable is already disqualified we can skip this |
3768 | { |
3769 | if (tree->gtFlags & GTF_VAR_DEF) // Is this is a def of our variable |
3770 | { |
3771 | /* |
3772 | If we have one of these cases: |
3773 | 1. We have already seen a definition (i.e lvSingleDef is true) |
3774 | 2. or info.CompInitMem is true (thus this would be the second definition) |
3775 | 3. or we have an assignment inside QMARK-COLON trees |
3776 | 4. or we have an update form of assignment (i.e. +=, -=, *=) |
3777 | Then we must disqualify this variable for use in optAddCopies() |
3778 | |
3779 | Note that all parameters start out with lvSingleDef set to true |
3780 | */ |
3781 | if ((varDsc->lvSingleDef == true) || (info.compInitMem == true) || (tree->gtFlags & GTF_COLON_COND) || |
3782 | (tree->gtFlags & GTF_VAR_USEASG)) |
3783 | { |
3784 | varDsc->lvaDisqualifyVar(); |
3785 | } |
3786 | else |
3787 | { |
3788 | varDsc->lvSingleDef = true; |
3789 | varDsc->lvDefStmt = stmt; |
3790 | } |
3791 | } |
3792 | else // otherwise this is a ref of our variable |
3793 | { |
3794 | if (BlockSetOps::MayBeUninit(varDsc->lvRefBlks)) |
3795 | { |
3796 | // Lazy initialization |
3797 | BlockSetOps::AssignNoCopy(this, varDsc->lvRefBlks, BlockSetOps::MakeEmpty(this)); |
3798 | } |
3799 | BlockSetOps::AddElemD(this, varDsc->lvRefBlks, block->bbNum); |
3800 | } |
3801 | } |
3802 | #endif // ASSERTION_PROP |
3803 | |
3804 | bool allowStructs = false; |
3805 | #ifdef UNIX_AMD64_ABI |
3806 | // On System V the type of the var could be a struct type. |
3807 | allowStructs = varTypeIsStruct(varDsc); |
3808 | #endif // UNIX_AMD64_ABI |
3809 | |
3810 | /* Variables must be used as the same type throughout the method */ |
3811 | noway_assert(tiVerificationNeeded || varDsc->lvType == TYP_UNDEF || tree->gtType == TYP_UNKNOWN || |
3812 | allowStructs || genActualType(varDsc->TypeGet()) == genActualType(tree->gtType) || |
3813 | (tree->gtType == TYP_BYREF && varDsc->TypeGet() == TYP_I_IMPL) || |
3814 | (tree->gtType == TYP_I_IMPL && varDsc->TypeGet() == TYP_BYREF) || (tree->gtFlags & GTF_VAR_CAST) || |
3815 | varTypeIsFloating(varDsc->TypeGet()) && varTypeIsFloating(tree->gtType)); |
3816 | |
3817 | /* Remember the type of the reference */ |
3818 | |
3819 | if (tree->gtType == TYP_UNKNOWN || varDsc->lvType == TYP_UNDEF) |
3820 | { |
3821 | varDsc->lvType = tree->gtType; |
3822 | noway_assert(genActualType(varDsc->TypeGet()) == tree->gtType); // no truncation |
3823 | } |
3824 | |
3825 | #ifdef DEBUG |
3826 | if (tree->gtFlags & GTF_VAR_CAST) |
3827 | { |
3828 | // it should never be bigger than the variable slot |
3829 | |
3830 | // Trees don't store the full information about structs |
3831 | // so we can't check them. |
3832 | if (tree->TypeGet() != TYP_STRUCT) |
3833 | { |
3834 | unsigned treeSize = genTypeSize(tree->TypeGet()); |
3835 | unsigned varSize = genTypeSize(varDsc->TypeGet()); |
3836 | if (varDsc->TypeGet() == TYP_STRUCT) |
3837 | { |
3838 | varSize = varDsc->lvSize(); |
3839 | } |
3840 | |
3841 | assert(treeSize <= varSize); |
3842 | } |
3843 | } |
3844 | #endif |
3845 | } |
3846 | } |
3847 | |
3848 | //------------------------------------------------------------------------ |
3849 | // IsDominatedByExceptionalEntry: Check is the block dominated by an exception entry block. |
3850 | // |
3851 | // Arguments: |
3852 | // block - the checking block. |
3853 | // |
3854 | bool Compiler::IsDominatedByExceptionalEntry(BasicBlock* block) |
3855 | { |
3856 | assert(fgDomsComputed); |
3857 | return block->IsDominatedByExceptionalEntryFlag(); |
3858 | } |
3859 | |
3860 | //------------------------------------------------------------------------ |
3861 | // SetVolatileHint: Set a local var's volatile hint. |
3862 | // |
3863 | // Arguments: |
3864 | // varDsc - the local variable that needs the hint. |
3865 | // |
3866 | void Compiler::SetVolatileHint(LclVarDsc* varDsc) |
3867 | { |
3868 | varDsc->lvVolatileHint = true; |
3869 | } |
3870 | |
3871 | //------------------------------------------------------------------------ |
3872 | // lvaMarkLocalVars: update local var ref counts for IR in a basic block |
3873 | // |
3874 | // Arguments: |
3875 | // block - the block in question |
3876 | // isRecompute - true if counts are being recomputed |
3877 | // |
3878 | // Notes: |
3879 | // Invokes lvaMarkLclRefs on each tree node for each |
3880 | // statement in the block. |
3881 | |
3882 | void Compiler::lvaMarkLocalVars(BasicBlock* block, bool isRecompute) |
3883 | { |
3884 | class MarkLocalVarsVisitor final : public GenTreeVisitor<MarkLocalVarsVisitor> |
3885 | { |
3886 | private: |
3887 | BasicBlock* m_block; |
3888 | GenTreeStmt* m_stmt; |
3889 | bool m_isRecompute; |
3890 | |
3891 | public: |
3892 | enum |
3893 | { |
3894 | DoPreOrder = true, |
3895 | }; |
3896 | |
3897 | MarkLocalVarsVisitor(Compiler* compiler, BasicBlock* block, GenTreeStmt* stmt, bool isRecompute) |
3898 | : GenTreeVisitor<MarkLocalVarsVisitor>(compiler), m_block(block), m_stmt(stmt), m_isRecompute(isRecompute) |
3899 | { |
3900 | } |
3901 | |
3902 | Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user) |
3903 | { |
3904 | m_compiler->lvaMarkLclRefs(*use, m_block, m_stmt, m_isRecompute); |
3905 | return WALK_CONTINUE; |
3906 | } |
3907 | }; |
3908 | |
3909 | JITDUMP("\n*** %s local variables in block " FMT_BB " (weight=%s)\n" , isRecompute ? "recomputing" : "marking" , |
3910 | block->bbNum, refCntWtd2str(block->getBBWeight(this))); |
3911 | |
3912 | for (GenTreeStmt* stmt = block->FirstNonPhiDef(); stmt != nullptr; stmt = stmt->getNextStmt()) |
3913 | { |
3914 | MarkLocalVarsVisitor visitor(this, block, stmt, isRecompute); |
3915 | DISPTREE(stmt); |
3916 | visitor.WalkTree(&stmt->gtStmtExpr, nullptr); |
3917 | } |
3918 | } |
3919 | |
3920 | //------------------------------------------------------------------------ |
3921 | // lvaMarkLocalVars: enable normal ref counting, compute initial counts, sort locals table |
3922 | // |
3923 | // Notes: |
3924 | // Now behaves differently in minopts / debug. Instead of actually inspecting |
3925 | // the IR and counting references, the jit assumes all locals are referenced |
3926 | // and does not sort the locals table. |
3927 | // |
3928 | // Also, when optimizing, lays the groundwork for assertion prop and more. |
3929 | // See details in lvaMarkLclRefs. |
3930 | |
3931 | void Compiler::lvaMarkLocalVars() |
3932 | { |
3933 | |
3934 | JITDUMP("\n*************** In lvaMarkLocalVars()" ); |
3935 | |
3936 | // If we have direct pinvokes, verify the frame list root local was set up properly |
3937 | if (info.compCallUnmanaged != 0) |
3938 | { |
3939 | assert((!opts.ShouldUsePInvokeHelpers()) || (info.compLvFrameListRoot == BAD_VAR_NUM)); |
3940 | if (!opts.ShouldUsePInvokeHelpers()) |
3941 | { |
3942 | noway_assert(info.compLvFrameListRoot >= info.compLocalsCount && info.compLvFrameListRoot < lvaCount); |
3943 | } |
3944 | } |
3945 | |
3946 | #if !FEATURE_EH_FUNCLETS |
3947 | |
3948 | // Grab space for exception handling |
3949 | |
3950 | if (ehNeedsShadowSPslots()) |
3951 | { |
3952 | // The first slot is reserved for ICodeManager::FixContext(ppEndRegion) |
3953 | // ie. the offset of the end-of-last-executed-filter |
3954 | unsigned slotsNeeded = 1; |
3955 | |
3956 | unsigned handlerNestingLevel = ehMaxHndNestingCount; |
3957 | |
3958 | if (opts.compDbgEnC && (handlerNestingLevel < (unsigned)MAX_EnC_HANDLER_NESTING_LEVEL)) |
3959 | handlerNestingLevel = (unsigned)MAX_EnC_HANDLER_NESTING_LEVEL; |
3960 | |
3961 | slotsNeeded += handlerNestingLevel; |
3962 | |
3963 | // For a filter (which can be active at the same time as a catch/finally handler) |
3964 | slotsNeeded++; |
3965 | // For zero-termination of the shadow-Stack-pointer chain |
3966 | slotsNeeded++; |
3967 | |
3968 | lvaShadowSPslotsVar = lvaGrabTempWithImplicitUse(false DEBUGARG("lvaShadowSPslotsVar" )); |
3969 | LclVarDsc* shadowSPslotsVar = &lvaTable[lvaShadowSPslotsVar]; |
3970 | shadowSPslotsVar->lvType = TYP_BLK; |
3971 | shadowSPslotsVar->lvExactSize = (slotsNeeded * TARGET_POINTER_SIZE); |
3972 | } |
3973 | |
3974 | #endif // !FEATURE_EH_FUNCLETS |
3975 | |
3976 | // PSPSym and LocAllocSPvar are not used by the CoreRT ABI |
3977 | if (!IsTargetAbi(CORINFO_CORERT_ABI)) |
3978 | { |
3979 | #if FEATURE_EH_FUNCLETS |
3980 | if (ehNeedsPSPSym()) |
3981 | { |
3982 | lvaPSPSym = lvaGrabTempWithImplicitUse(false DEBUGARG("PSPSym" )); |
3983 | LclVarDsc* lclPSPSym = &lvaTable[lvaPSPSym]; |
3984 | lclPSPSym->lvType = TYP_I_IMPL; |
3985 | } |
3986 | #endif // FEATURE_EH_FUNCLETS |
3987 | |
3988 | #ifdef JIT32_GCENCODER |
3989 | // LocAllocSPvar is only required by the implicit frame layout expected by the VM on x86. Whether |
3990 | // a function contains a Localloc is conveyed in the GC information, in the InfoHdrSmall.localloc |
3991 | // field. The function must have an EBP frame. Then, the VM finds the LocAllocSP slot by assuming |
3992 | // the following stack layout: |
3993 | // |
3994 | // -- higher addresses -- |
3995 | // saved EBP <-- EBP points here |
3996 | // other callee-saved registers // InfoHdrSmall.savedRegsCountExclFP specifies this size |
3997 | // optional GS cookie // InfoHdrSmall.security is 1 if this exists |
3998 | // LocAllocSP slot |
3999 | // -- lower addresses -- |
4000 | // |
4001 | // See also eetwain.cpp::GetLocallocSPOffset() and its callers. |
4002 | if (compLocallocUsed) |
4003 | { |
4004 | lvaLocAllocSPvar = lvaGrabTempWithImplicitUse(false DEBUGARG("LocAllocSPvar" )); |
4005 | LclVarDsc* locAllocSPvar = &lvaTable[lvaLocAllocSPvar]; |
4006 | locAllocSPvar->lvType = TYP_I_IMPL; |
4007 | } |
4008 | #endif // JIT32_GCENCODER |
4009 | } |
4010 | |
4011 | // Ref counting is now enabled normally. |
4012 | lvaRefCountState = RCS_NORMAL; |
4013 | |
4014 | #if defined(DEBUG) |
4015 | const bool setSlotNumbers = true; |
4016 | #else |
4017 | const bool setSlotNumbers = opts.compScopeInfo && (info.compVarScopesCount > 0); |
4018 | #endif // defined(DEBUG) |
4019 | |
4020 | const bool isRecompute = false; |
4021 | lvaComputeRefCounts(isRecompute, setSlotNumbers); |
4022 | |
4023 | // If we're not optimizing, we're done. |
4024 | if (opts.OptimizationDisabled()) |
4025 | { |
4026 | return; |
4027 | } |
4028 | |
4029 | #if ASSERTION_PROP |
4030 | assert(opts.OptimizationEnabled()); |
4031 | |
4032 | // Note: optAddCopies() depends on lvaRefBlks, which is set in lvaMarkLocalVars(BasicBlock*), called above. |
4033 | optAddCopies(); |
4034 | #endif |
4035 | |
4036 | if (lvaKeepAliveAndReportThis()) |
4037 | { |
4038 | lvaTable[0].lvImplicitlyReferenced = 1; |
4039 | // This isn't strictly needed as we will make a copy of the param-type-arg |
4040 | // in the prolog. However, this ensures that the LclVarDsc corresponding to |
4041 | // info.compTypeCtxtArg is valid. |
4042 | } |
4043 | else if (lvaReportParamTypeArg()) |
4044 | { |
4045 | lvaTable[info.compTypeCtxtArg].lvImplicitlyReferenced = 1; |
4046 | } |
4047 | |
4048 | lvaSortByRefCount(); |
4049 | } |
4050 | |
4051 | //------------------------------------------------------------------------ |
4052 | // lvaComputeRefCounts: compute ref counts for locals |
4053 | // |
4054 | // Arguments: |
4055 | // isRecompute -- true if we just want ref counts and no other side effects; |
4056 | // false means to also look for true boolean locals, lay |
4057 | // groundwork for assertion prop, check type consistency, etc. |
4058 | // See lvaMarkLclRefs for details on what else goes on. |
4059 | // setSlotNumbers -- true if local slot numbers should be assigned. |
4060 | // |
4061 | // Notes: |
4062 | // Some implicit references are given actual counts or weight bumps here |
4063 | // to match pre-existing behavior. |
4064 | // |
4065 | // In fast-jitting modes where we don't ref count locals, this bypasses |
4066 | // actual counting, and makes all locals implicitly referenced on first |
4067 | // compute. It asserts all locals are implicitly referenced on recompute. |
4068 | |
4069 | void Compiler::lvaComputeRefCounts(bool isRecompute, bool setSlotNumbers) |
4070 | { |
4071 | JITDUMP("\n*** lvaComputeRefCounts ***\n" ); |
4072 | unsigned lclNum = 0; |
4073 | LclVarDsc* varDsc = nullptr; |
4074 | |
4075 | // Fast path for minopts and debug codegen. |
4076 | // |
4077 | // On first compute: mark all locals as implicitly referenced and untracked. |
4078 | // On recompute: do nothing. |
4079 | if (opts.OptimizationDisabled()) |
4080 | { |
4081 | if (isRecompute) |
4082 | { |
4083 | |
4084 | #if defined(DEBUG) |
4085 | // All local vars should be marked as implicitly referenced |
4086 | // and not tracked. |
4087 | for (lclNum = 0, varDsc = lvaTable; lclNum < lvaCount; lclNum++, varDsc++) |
4088 | { |
4089 | const bool isSpecialVarargsParam = varDsc->lvIsParam && raIsVarargsStackArg(lclNum); |
4090 | |
4091 | if (isSpecialVarargsParam) |
4092 | { |
4093 | assert(varDsc->lvRefCnt() == 0); |
4094 | } |
4095 | else |
4096 | { |
4097 | assert(varDsc->lvImplicitlyReferenced); |
4098 | } |
4099 | |
4100 | assert(!varDsc->lvTracked); |
4101 | } |
4102 | #endif // defined (DEBUG) |
4103 | |
4104 | return; |
4105 | } |
4106 | |
4107 | // First compute. |
4108 | for (lclNum = 0, varDsc = lvaTable; lclNum < lvaCount; lclNum++, varDsc++) |
4109 | { |
4110 | // Using lvImplicitlyReferenced here ensures that we can't |
4111 | // accidentally make locals be unreferenced later by decrementing |
4112 | // the ref count to zero. |
4113 | // |
4114 | // If, in minopts/debug, we really want to allow locals to become |
4115 | // unreferenced later, we'll have to explicitly clear this bit. |
4116 | varDsc->setLvRefCnt(0); |
4117 | varDsc->setLvRefCntWtd(0); |
4118 | |
4119 | // Special case for some varargs params ... these must |
4120 | // remain unreferenced. |
4121 | const bool isSpecialVarargsParam = varDsc->lvIsParam && raIsVarargsStackArg(lclNum); |
4122 | |
4123 | if (!isSpecialVarargsParam) |
4124 | { |
4125 | varDsc->lvImplicitlyReferenced = 1; |
4126 | } |
4127 | |
4128 | varDsc->lvTracked = 0; |
4129 | |
4130 | if (setSlotNumbers) |
4131 | { |
4132 | varDsc->lvSlotNum = lclNum; |
4133 | } |
4134 | |
4135 | // Assert that it's ok to bypass the type repair logic in lvaMarkLclRefs |
4136 | assert((varDsc->lvType != TYP_UNDEF) && (varDsc->lvType != TYP_VOID) && (varDsc->lvType != TYP_UNKNOWN)); |
4137 | } |
4138 | |
4139 | lvaCurEpoch++; |
4140 | lvaTrackedCount = 0; |
4141 | lvaTrackedCountInSizeTUnits = 0; |
4142 | return; |
4143 | } |
4144 | |
4145 | // Slower path we take when optimizing, to get accurate counts. |
4146 | // |
4147 | // First, reset all explicit ref counts and weights. |
4148 | for (lclNum = 0, varDsc = lvaTable; lclNum < lvaCount; lclNum++, varDsc++) |
4149 | { |
4150 | varDsc->setLvRefCnt(0); |
4151 | varDsc->setLvRefCntWtd(BB_ZERO_WEIGHT); |
4152 | |
4153 | if (setSlotNumbers) |
4154 | { |
4155 | varDsc->lvSlotNum = lclNum; |
4156 | } |
4157 | |
4158 | // Set initial value for lvSingleDef for explicit and implicit |
4159 | // argument locals as they are "defined" on entry. |
4160 | varDsc->lvSingleDef = varDsc->lvIsParam; |
4161 | } |
4162 | |
4163 | JITDUMP("\n*** lvaComputeRefCounts -- explicit counts ***\n" ); |
4164 | |
4165 | // Second, account for all explicit local variable references |
4166 | for (BasicBlock* block = fgFirstBB; block; block = block->bbNext) |
4167 | { |
4168 | if (block->IsLIR()) |
4169 | { |
4170 | assert(isRecompute); |
4171 | |
4172 | const BasicBlock::weight_t weight = block->getBBWeight(this); |
4173 | for (GenTree* node : LIR::AsRange(block).NonPhiNodes()) |
4174 | { |
4175 | switch (node->OperGet()) |
4176 | { |
4177 | case GT_LCL_VAR: |
4178 | case GT_LCL_FLD: |
4179 | case GT_LCL_VAR_ADDR: |
4180 | case GT_LCL_FLD_ADDR: |
4181 | case GT_STORE_LCL_VAR: |
4182 | case GT_STORE_LCL_FLD: |
4183 | { |
4184 | const unsigned lclNum = node->AsLclVarCommon()->gtLclNum; |
4185 | lvaTable[lclNum].incRefCnts(weight, this); |
4186 | break; |
4187 | } |
4188 | |
4189 | default: |
4190 | break; |
4191 | } |
4192 | } |
4193 | } |
4194 | else |
4195 | { |
4196 | lvaMarkLocalVars(block, isRecompute); |
4197 | } |
4198 | } |
4199 | |
4200 | JITDUMP("\n*** lvaComputeRefCounts -- implicit counts ***\n" ); |
4201 | |
4202 | // Third, bump ref counts for some implicit prolog references |
4203 | for (lclNum = 0, varDsc = lvaTable; lclNum < lvaCount; lclNum++, varDsc++) |
4204 | { |
4205 | // Todo: review justification for these count bumps. |
4206 | if (varDsc->lvIsRegArg) |
4207 | { |
4208 | if ((lclNum < info.compArgsCount) && (varDsc->lvRefCnt() > 0)) |
4209 | { |
4210 | // Fix 388376 ARM JitStress WP7 |
4211 | varDsc->incRefCnts(BB_UNITY_WEIGHT, this); |
4212 | varDsc->incRefCnts(BB_UNITY_WEIGHT, this); |
4213 | } |
4214 | |
4215 | // Ref count bump that was in lvaPromoteStructVar |
4216 | // |
4217 | // This was formerly done during RCS_EARLY counting, |
4218 | // and we did not used to reset counts like we do now. |
4219 | if (varDsc->lvIsStructField) |
4220 | { |
4221 | varDsc->incRefCnts(BB_UNITY_WEIGHT, this); |
4222 | } |
4223 | } |
4224 | |
4225 | // If we have JMP, all arguments must have a location |
4226 | // even if we don't use them inside the method |
4227 | if (compJmpOpUsed && varDsc->lvIsParam && (varDsc->lvRefCnt() == 0)) |
4228 | { |
4229 | // except when we have varargs and the argument is |
4230 | // passed on the stack. In that case, it's important |
4231 | // for the ref count to be zero, so that we don't attempt |
4232 | // to track them for GC info (which is not possible since we |
4233 | // don't know their offset in the stack). See the assert at the |
4234 | // end of raMarkStkVars and bug #28949 for more info. |
4235 | if (!raIsVarargsStackArg(lclNum)) |
4236 | { |
4237 | varDsc->lvImplicitlyReferenced = 1; |
4238 | } |
4239 | } |
4240 | } |
4241 | } |
4242 | |
4243 | void Compiler::lvaAllocOutgoingArgSpaceVar() |
4244 | { |
4245 | #if FEATURE_FIXED_OUT_ARGS |
4246 | |
4247 | // Setup the outgoing argument region, in case we end up using it later |
4248 | |
4249 | if (lvaOutgoingArgSpaceVar == BAD_VAR_NUM) |
4250 | { |
4251 | lvaOutgoingArgSpaceVar = lvaGrabTemp(false DEBUGARG("OutgoingArgSpace" )); |
4252 | |
4253 | lvaTable[lvaOutgoingArgSpaceVar].lvType = TYP_LCLBLK; |
4254 | lvaTable[lvaOutgoingArgSpaceVar].lvImplicitlyReferenced = 1; |
4255 | } |
4256 | |
4257 | noway_assert(lvaOutgoingArgSpaceVar >= info.compLocalsCount && lvaOutgoingArgSpaceVar < lvaCount); |
4258 | |
4259 | #endif // FEATURE_FIXED_OUT_ARGS |
4260 | } |
4261 | |
4262 | inline void Compiler::lvaIncrementFrameSize(unsigned size) |
4263 | { |
4264 | if (size > MAX_FrameSize || compLclFrameSize + size > MAX_FrameSize) |
4265 | { |
4266 | BADCODE("Frame size overflow" ); |
4267 | } |
4268 | |
4269 | compLclFrameSize += size; |
4270 | } |
4271 | |
4272 | /**************************************************************************** |
4273 | * |
4274 | * Return true if absolute offsets of temps are larger than vars, or in other |
4275 | * words, did we allocate temps before of after vars. The /GS buffer overrun |
4276 | * checks want temps to be at low stack addresses than buffers |
4277 | */ |
4278 | bool Compiler::lvaTempsHaveLargerOffsetThanVars() |
4279 | { |
4280 | #ifdef _TARGET_ARM_ |
4281 | // We never want to place the temps with larger offsets for ARM |
4282 | return false; |
4283 | #else |
4284 | if (compGSReorderStackLayout) |
4285 | { |
4286 | return codeGen->isFramePointerUsed(); |
4287 | } |
4288 | else |
4289 | { |
4290 | return true; |
4291 | } |
4292 | #endif |
4293 | } |
4294 | |
4295 | /**************************************************************************** |
4296 | * |
4297 | * Return an upper bound estimate for the size of the compiler spill temps |
4298 | * |
4299 | */ |
4300 | unsigned Compiler::lvaGetMaxSpillTempSize() |
4301 | { |
4302 | unsigned result = 0; |
4303 | |
4304 | if (lvaDoneFrameLayout >= REGALLOC_FRAME_LAYOUT) |
4305 | { |
4306 | result = codeGen->regSet.tmpGetTotalSize(); |
4307 | } |
4308 | else |
4309 | { |
4310 | result = MAX_SPILL_TEMP_SIZE; |
4311 | } |
4312 | return result; |
4313 | } |
4314 | |
4315 | // clang-format off |
4316 | /***************************************************************************** |
4317 | * |
4318 | * Compute stack frame offsets for arguments, locals and optionally temps. |
4319 | * |
4320 | * The frame is laid out as follows for x86: |
4321 | * |
4322 | * ESP frames |
4323 | * |
4324 | * | | |
4325 | * |-----------------------| |
4326 | * | incoming | |
4327 | * | arguments | |
4328 | * |-----------------------| <---- Virtual '0' |
4329 | * | return address | |
4330 | * +=======================+ |
4331 | * |Callee saved registers | |
4332 | * |-----------------------| |
4333 | * | Temps | |
4334 | * |-----------------------| |
4335 | * | Variables | |
4336 | * |-----------------------| <---- Ambient ESP |
4337 | * | Arguments for the | |
4338 | * ~ next function ~ |
4339 | * | | |
4340 | * | | | |
4341 | * | | Stack grows | |
4342 | * | downward |
4343 | * V |
4344 | * |
4345 | * |
4346 | * EBP frames |
4347 | * |
4348 | * | | |
4349 | * |-----------------------| |
4350 | * | incoming | |
4351 | * | arguments | |
4352 | * |-----------------------| <---- Virtual '0' |
4353 | * | return address | |
4354 | * +=======================+ |
4355 | * | incoming EBP | |
4356 | * |-----------------------| <---- EBP |
4357 | * |Callee saved registers | |
4358 | * |-----------------------| |
4359 | * | security object | |
4360 | * |-----------------------| |
4361 | * | ParamTypeArg | |
4362 | * |-----------------------| |
4363 | * | Last-executed-filter | |
4364 | * |-----------------------| |
4365 | * | | |
4366 | * ~ Shadow SPs ~ |
4367 | * | | |
4368 | * |-----------------------| |
4369 | * | | |
4370 | * ~ Variables ~ |
4371 | * | | |
4372 | * ~-----------------------| |
4373 | * | Temps | |
4374 | * |-----------------------| |
4375 | * | localloc | |
4376 | * |-----------------------| <---- Ambient ESP |
4377 | * | Arguments for the | |
4378 | * | next function ~ |
4379 | * | | |
4380 | * | | | |
4381 | * | | Stack grows | |
4382 | * | downward |
4383 | * V |
4384 | * |
4385 | * |
4386 | * The frame is laid out as follows for x64: |
4387 | * |
4388 | * RSP frames |
4389 | * | | |
4390 | * |-----------------------| |
4391 | * | incoming | |
4392 | * | arguments | |
4393 | * |-----------------------| |
4394 | * | 4 fixed incoming | |
4395 | * | argument slots | |
4396 | * |-----------------------| <---- Caller's SP & Virtual '0' |
4397 | * | return address | |
4398 | * +=======================+ |
4399 | * | Callee saved Int regs | |
4400 | * ------------------------- |
4401 | * | Padding | <---- this padding (0 or 8 bytes) is to ensure flt registers are saved at a mem location aligned at 16-bytes |
4402 | * | | so that we can save 128-bit callee saved xmm regs using performant "movaps" instruction instead of "movups" |
4403 | * ------------------------- |
4404 | * | Callee saved Flt regs | <----- entire 128-bits of callee saved xmm registers are stored here |
4405 | * |-----------------------| |
4406 | * | Temps | |
4407 | * |-----------------------| |
4408 | * | Variables | |
4409 | * |-----------------------| |
4410 | * | Arguments for the | |
4411 | * ~ next function ~ |
4412 | * | | |
4413 | * |-----------------------| |
4414 | * | 4 fixed outgoing | |
4415 | * | argument slots | |
4416 | * |-----------------------| <---- Ambient RSP |
4417 | * | | | |
4418 | * ~ | Stack grows ~ |
4419 | * | | downward | |
4420 | * V |
4421 | * |
4422 | * |
4423 | * RBP frames |
4424 | * | | |
4425 | * |-----------------------| |
4426 | * | incoming | |
4427 | * | arguments | |
4428 | * |-----------------------| |
4429 | * | 4 fixed incoming | |
4430 | * | argument slots | |
4431 | * |-----------------------| <---- Caller's SP & Virtual '0' |
4432 | * | return address | |
4433 | * +=======================+ |
4434 | * | Callee saved Int regs | |
4435 | * ------------------------- |
4436 | * | Padding | |
4437 | * ------------------------- |
4438 | * | Callee saved Flt regs | |
4439 | * |-----------------------| |
4440 | * | security object | |
4441 | * |-----------------------| |
4442 | * | ParamTypeArg | |
4443 | * |-----------------------| |
4444 | * | | |
4445 | * | | |
4446 | * ~ Variables ~ |
4447 | * | | |
4448 | * | | |
4449 | * |-----------------------| |
4450 | * | Temps | |
4451 | * |-----------------------| |
4452 | * | | |
4453 | * ~ localloc ~ // not in frames with EH |
4454 | * | | |
4455 | * |-----------------------| |
4456 | * | PSPSym | // only in frames with EH (thus no localloc) |
4457 | * | | |
4458 | * |-----------------------| <---- RBP in localloc frames (max 240 bytes from Initial-SP) |
4459 | * | Arguments for the | |
4460 | * ~ next function ~ |
4461 | * | | |
4462 | * |-----------------------| |
4463 | * | 4 fixed outgoing | |
4464 | * | argument slots | |
4465 | * |-----------------------| <---- Ambient RSP (before localloc, this is Initial-SP) |
4466 | * | | | |
4467 | * ~ | Stack grows ~ |
4468 | * | | downward | |
4469 | * V |
4470 | * |
4471 | * |
4472 | * The frame is laid out as follows for ARM (this is a general picture; details may differ for different conditions): |
4473 | * |
4474 | * SP frames |
4475 | * | | |
4476 | * |-----------------------| |
4477 | * | incoming | |
4478 | * | arguments | |
4479 | * +=======================+ <---- Caller's SP |
4480 | * | Pre-spill registers | |
4481 | * |-----------------------| <---- Virtual '0' |
4482 | * |Callee saved registers | |
4483 | * |-----------------------| |
4484 | * ~ possible double align ~ |
4485 | * |-----------------------| |
4486 | * | security object | |
4487 | * |-----------------------| |
4488 | * | ParamTypeArg | |
4489 | * |-----------------------| |
4490 | * | possible GS cookie | |
4491 | * |-----------------------| |
4492 | * | Variables | |
4493 | * |-----------------------| |
4494 | * | possible GS cookie | |
4495 | * |-----------------------| |
4496 | * | Temps | |
4497 | * |-----------------------| |
4498 | * | Stub Argument Var | |
4499 | * |-----------------------| |
4500 | * |Inlined PInvoke Frame V| |
4501 | * |-----------------------| |
4502 | * ~ possible double align ~ |
4503 | * |-----------------------| |
4504 | * | Arguments for the | |
4505 | * ~ next function ~ |
4506 | * | | |
4507 | * |-----------------------| <---- Ambient SP |
4508 | * | | | |
4509 | * ~ | Stack grows ~ |
4510 | * | | downward | |
4511 | * V |
4512 | * |
4513 | * |
4514 | * FP / R11 frames |
4515 | * | | |
4516 | * |-----------------------| |
4517 | * | incoming | |
4518 | * | arguments | |
4519 | * +=======================+ <---- Caller's SP |
4520 | * | Pre-spill registers | |
4521 | * |-----------------------| <---- Virtual '0' |
4522 | * |Callee saved registers | |
4523 | * |-----------------------| |
4524 | * | PSPSym | // Only for frames with EH, which means FP-based frames |
4525 | * |-----------------------| |
4526 | * ~ possible double align ~ |
4527 | * |-----------------------| |
4528 | * | security object | |
4529 | * |-----------------------| |
4530 | * | ParamTypeArg | |
4531 | * |-----------------------| |
4532 | * | possible GS cookie | |
4533 | * |-----------------------| |
4534 | * | Variables | |
4535 | * |-----------------------| |
4536 | * | possible GS cookie | |
4537 | * |-----------------------| |
4538 | * | Temps | |
4539 | * |-----------------------| |
4540 | * | Stub Argument Var | |
4541 | * |-----------------------| |
4542 | * |Inlined PInvoke Frame V| |
4543 | * |-----------------------| |
4544 | * ~ possible double align ~ |
4545 | * |-----------------------| |
4546 | * | localloc | |
4547 | * |-----------------------| |
4548 | * | Arguments for the | |
4549 | * ~ next function ~ |
4550 | * | | |
4551 | * |-----------------------| <---- Ambient SP |
4552 | * | | | |
4553 | * ~ | Stack grows ~ |
4554 | * | | downward | |
4555 | * V |
4556 | * |
4557 | * |
4558 | * The frame is laid out as follows for ARM64 (this is a general picture; details may differ for different conditions): |
4559 | * TODO-ARM64-NYI: this is preliminary (copied from ARM and modified), and needs to be reviewed. |
4560 | * NOTE: SP must be 16-byte aligned, so there may be alignment slots in the frame. |
4561 | * We will often save and establish a frame pointer to create better ETW stack walks. |
4562 | * |
4563 | * SP frames |
4564 | * | | |
4565 | * |-----------------------| |
4566 | * | incoming | |
4567 | * | arguments | |
4568 | * +=======================+ <---- Caller's SP |
4569 | * | homed | // this is only needed if reg argument need to be homed, e.g., for varargs |
4570 | * | register arguments | |
4571 | * |-----------------------| <---- Virtual '0' |
4572 | * |Callee saved registers | |
4573 | * | except fp/lr | |
4574 | * |-----------------------| |
4575 | * | security object | |
4576 | * |-----------------------| |
4577 | * | ParamTypeArg | |
4578 | * |-----------------------| |
4579 | * | possible GS cookie | |
4580 | * |-----------------------| |
4581 | * | Variables | |
4582 | * |-----------------------| |
4583 | * | possible GS cookie | |
4584 | * |-----------------------| |
4585 | * | Temps | |
4586 | * |-----------------------| |
4587 | * | Stub Argument Var | |
4588 | * |-----------------------| |
4589 | * |Inlined PInvoke Frame V| |
4590 | * |-----------------------| |
4591 | * | Saved LR | |
4592 | * |-----------------------| |
4593 | * | Saved FP | <---- Frame pointer |
4594 | * |-----------------------| |
4595 | * | Stack arguments for | |
4596 | * | the next function | |
4597 | * |-----------------------| <---- SP |
4598 | * | | | |
4599 | * ~ | Stack grows ~ |
4600 | * | | downward | |
4601 | * V |
4602 | * |
4603 | * |
4604 | * FP (R29 / x29) frames |
4605 | * | | |
4606 | * |-----------------------| |
4607 | * | incoming | |
4608 | * | arguments | |
4609 | * +=======================+ <---- Caller's SP |
4610 | * | optional homed | // this is only needed if reg argument need to be homed, e.g., for varargs |
4611 | * | register arguments | |
4612 | * |-----------------------| <---- Virtual '0' |
4613 | * |Callee saved registers | |
4614 | * | except fp/lr | |
4615 | * |-----------------------| |
4616 | * | PSPSym | // Only for frames with EH, which requires FP-based frames |
4617 | * |-----------------------| |
4618 | * | security object | |
4619 | * |-----------------------| |
4620 | * | ParamTypeArg | |
4621 | * |-----------------------| |
4622 | * | possible GS cookie | |
4623 | * |-----------------------| |
4624 | * | Variables | |
4625 | * |-----------------------| |
4626 | * | possible GS cookie | |
4627 | * |-----------------------| |
4628 | * | Temps | |
4629 | * |-----------------------| |
4630 | * | Stub Argument Var | |
4631 | * |-----------------------| |
4632 | * |Inlined PInvoke Frame V| |
4633 | * |-----------------------| |
4634 | * | Saved LR | |
4635 | * |-----------------------| |
4636 | * | Saved FP | <---- Frame pointer |
4637 | * |-----------------------| |
4638 | * ~ localloc ~ |
4639 | * |-----------------------| |
4640 | * | Stack arguments for | |
4641 | * | the next function | |
4642 | * |-----------------------| <---- Ambient SP |
4643 | * | | | |
4644 | * ~ | Stack grows ~ |
4645 | * | | downward | |
4646 | * V |
4647 | * |
4648 | * |
4649 | * Doing this all in one pass is 'hard'. So instead we do it in 2 basic passes: |
4650 | * 1. Assign all the offsets relative to the Virtual '0'. Offsets above (the |
4651 | * incoming arguments) are positive. Offsets below (everything else) are |
4652 | * negative. This pass also calcuates the total frame size (between Caller's |
4653 | * SP/return address and the Ambient SP). |
4654 | * 2. Figure out where to place the frame pointer, and then adjust the offsets |
4655 | * as needed for the final stack size and whether the offset is frame pointer |
4656 | * relative or stack pointer relative. |
4657 | * |
4658 | */ |
4659 | // clang-format on |
4660 | |
4661 | void Compiler::lvaAssignFrameOffsets(FrameLayoutState curState) |
4662 | { |
4663 | noway_assert((lvaDoneFrameLayout < curState) || (curState == REGALLOC_FRAME_LAYOUT)); |
4664 | |
4665 | lvaDoneFrameLayout = curState; |
4666 | |
4667 | #ifdef DEBUG |
4668 | if (verbose) |
4669 | { |
4670 | |
4671 | printf("*************** In lvaAssignFrameOffsets" ); |
4672 | if (curState == INITIAL_FRAME_LAYOUT) |
4673 | { |
4674 | printf("(INITIAL_FRAME_LAYOUT)" ); |
4675 | } |
4676 | else if (curState == PRE_REGALLOC_FRAME_LAYOUT) |
4677 | { |
4678 | printf("(PRE_REGALLOC_FRAME_LAYOUT)" ); |
4679 | } |
4680 | else if (curState == REGALLOC_FRAME_LAYOUT) |
4681 | { |
4682 | printf("(REGALLOC_FRAME_LAYOUT)" ); |
4683 | } |
4684 | else if (curState == TENTATIVE_FRAME_LAYOUT) |
4685 | { |
4686 | printf("(TENTATIVE_FRAME_LAYOUT)" ); |
4687 | } |
4688 | else if (curState == FINAL_FRAME_LAYOUT) |
4689 | { |
4690 | printf("(FINAL_FRAME_LAYOUT)" ); |
4691 | } |
4692 | else |
4693 | { |
4694 | printf("(UNKNOWN)" ); |
4695 | unreached(); |
4696 | } |
4697 | printf("\n" ); |
4698 | } |
4699 | #endif |
4700 | |
4701 | #if FEATURE_FIXED_OUT_ARGS |
4702 | assert(lvaOutgoingArgSpaceVar != BAD_VAR_NUM); |
4703 | #endif // FEATURE_FIXED_OUT_ARGS |
4704 | |
4705 | /*------------------------------------------------------------------------- |
4706 | * |
4707 | * First process the arguments. |
4708 | * |
4709 | *------------------------------------------------------------------------- |
4710 | */ |
4711 | |
4712 | lvaAssignVirtualFrameOffsetsToArgs(); |
4713 | |
4714 | /*------------------------------------------------------------------------- |
4715 | * |
4716 | * Now compute stack offsets for any variables that don't live in registers |
4717 | * |
4718 | *------------------------------------------------------------------------- |
4719 | */ |
4720 | |
4721 | lvaAssignVirtualFrameOffsetsToLocals(); |
4722 | |
4723 | lvaAlignFrame(); |
4724 | |
4725 | /*------------------------------------------------------------------------- |
4726 | * |
4727 | * Now patch the offsets |
4728 | * |
4729 | *------------------------------------------------------------------------- |
4730 | */ |
4731 | |
4732 | lvaFixVirtualFrameOffsets(); |
4733 | |
4734 | // Modify the stack offset for fields of promoted structs. |
4735 | lvaAssignFrameOffsetsToPromotedStructs(); |
4736 | |
4737 | /*------------------------------------------------------------------------- |
4738 | * |
4739 | * Finalize |
4740 | * |
4741 | *------------------------------------------------------------------------- |
4742 | */ |
4743 | |
4744 | // If it's not the final frame layout, then it's just an estimate. This means |
4745 | // we're allowed to once again write to these variables, even if we've read |
4746 | // from them to make tentative code generation or frame layout decisions. |
4747 | if (curState < FINAL_FRAME_LAYOUT) |
4748 | { |
4749 | codeGen->resetFramePointerUsedWritePhase(); |
4750 | } |
4751 | } |
4752 | |
4753 | /***************************************************************************** |
4754 | * lvaFixVirtualFrameOffsets() : Now that everything has a virtual offset, |
4755 | * determine the final value for the frame pointer (if needed) and then |
4756 | * adjust all the offsets appropriately. |
4757 | * |
4758 | * This routine fixes virtual offset to be relative to frame pointer or SP |
4759 | * based on whether varDsc->lvFramePointerBased is true or false respectively. |
4760 | */ |
4761 | void Compiler::lvaFixVirtualFrameOffsets() |
4762 | { |
4763 | LclVarDsc* varDsc; |
4764 | |
4765 | #if FEATURE_EH_FUNCLETS && defined(_TARGET_AMD64_) |
4766 | if (lvaPSPSym != BAD_VAR_NUM) |
4767 | { |
4768 | // We need to fix the offset of the PSPSym so there is no padding between it and the outgoing argument space. |
4769 | // Without this code, lvaAlignFrame might have put the padding lower than the PSPSym, which would be between |
4770 | // the PSPSym and the outgoing argument space. |
4771 | varDsc = &lvaTable[lvaPSPSym]; |
4772 | assert(varDsc->lvFramePointerBased); // We always access it RBP-relative. |
4773 | assert(!varDsc->lvMustInit); // It is never "must init". |
4774 | varDsc->lvStkOffs = codeGen->genCallerSPtoInitialSPdelta() + lvaLclSize(lvaOutgoingArgSpaceVar); |
4775 | } |
4776 | #endif |
4777 | |
4778 | // The delta to be added to virtual offset to adjust it relative to frame pointer or SP |
4779 | int delta = 0; |
4780 | |
4781 | #ifdef _TARGET_XARCH_ |
4782 | delta += REGSIZE_BYTES; // pushed PC (return address) for x86/x64 |
4783 | |
4784 | if (codeGen->doubleAlignOrFramePointerUsed()) |
4785 | { |
4786 | delta += REGSIZE_BYTES; // pushed EBP (frame pointer) |
4787 | } |
4788 | #endif |
4789 | |
4790 | if (!codeGen->isFramePointerUsed()) |
4791 | { |
4792 | // pushed registers, return address, and padding |
4793 | delta += codeGen->genTotalFrameSize(); |
4794 | } |
4795 | #if defined(_TARGET_ARM_) |
4796 | else |
4797 | { |
4798 | // We set FP to be after LR, FP |
4799 | delta += 2 * REGSIZE_BYTES; |
4800 | } |
4801 | #elif defined(_TARGET_AMD64_) || defined(_TARGET_ARM64_) |
4802 | else |
4803 | { |
4804 | // FP is used. |
4805 | delta += codeGen->genTotalFrameSize() - codeGen->genSPtoFPdelta(); |
4806 | } |
4807 | #endif //_TARGET_AMD64_ |
4808 | |
4809 | unsigned lclNum; |
4810 | for (lclNum = 0, varDsc = lvaTable; lclNum < lvaCount; lclNum++, varDsc++) |
4811 | { |
4812 | bool doAssignStkOffs = true; |
4813 | |
4814 | // Can't be relative to EBP unless we have an EBP |
4815 | noway_assert(!varDsc->lvFramePointerBased || codeGen->doubleAlignOrFramePointerUsed()); |
4816 | |
4817 | // Is this a non-param promoted struct field? |
4818 | // if so then set doAssignStkOffs to false. |
4819 | // |
4820 | if (varDsc->lvIsStructField && !varDsc->lvIsParam) |
4821 | { |
4822 | LclVarDsc* parentvarDsc = &lvaTable[varDsc->lvParentLcl]; |
4823 | lvaPromotionType promotionType = lvaGetPromotionType(parentvarDsc); |
4824 | |
4825 | if (promotionType == PROMOTION_TYPE_DEPENDENT) |
4826 | { |
4827 | doAssignStkOffs = false; // Assigned later in lvaAssignFrameOffsetsToPromotedStructs() |
4828 | } |
4829 | } |
4830 | |
4831 | if (!varDsc->lvOnFrame) |
4832 | { |
4833 | if (!varDsc->lvIsParam |
4834 | #if !defined(_TARGET_AMD64_) |
4835 | || (varDsc->lvIsRegArg |
4836 | #if defined(_TARGET_ARM_) && defined(PROFILING_SUPPORTED) |
4837 | && compIsProfilerHookNeeded() && |
4838 | !lvaIsPreSpilled(lclNum, codeGen->regSet.rsMaskPreSpillRegs(false)) // We need assign stack offsets |
4839 | // for prespilled arguments |
4840 | #endif |
4841 | ) |
4842 | #endif // !defined(_TARGET_AMD64_) |
4843 | ) |
4844 | { |
4845 | doAssignStkOffs = false; // Not on frame or an incomming stack arg |
4846 | } |
4847 | } |
4848 | |
4849 | if (doAssignStkOffs) |
4850 | { |
4851 | varDsc->lvStkOffs += delta; |
4852 | |
4853 | #if DOUBLE_ALIGN |
4854 | if (genDoubleAlign() && !codeGen->isFramePointerUsed()) |
4855 | { |
4856 | if (varDsc->lvFramePointerBased) |
4857 | { |
4858 | varDsc->lvStkOffs -= delta; |
4859 | |
4860 | // We need to re-adjust the offsets of the parameters so they are EBP |
4861 | // relative rather than stack/frame pointer relative |
4862 | |
4863 | varDsc->lvStkOffs += (2 * TARGET_POINTER_SIZE); // return address and pushed EBP |
4864 | |
4865 | noway_assert(varDsc->lvStkOffs >= FIRST_ARG_STACK_OFFS); |
4866 | } |
4867 | } |
4868 | #endif |
4869 | // On System V environments the stkOffs could be 0 for params passed in registers. |
4870 | assert(codeGen->isFramePointerUsed() || |
4871 | varDsc->lvStkOffs >= 0); // Only EBP relative references can have negative offsets |
4872 | } |
4873 | } |
4874 | |
4875 | assert(codeGen->regSet.tmpAllFree()); |
4876 | for (TempDsc* temp = codeGen->regSet.tmpListBeg(); temp != nullptr; temp = codeGen->regSet.tmpListNxt(temp)) |
4877 | { |
4878 | temp->tdAdjustTempOffs(delta); |
4879 | } |
4880 | |
4881 | lvaCachedGenericContextArgOffs += delta; |
4882 | |
4883 | #if FEATURE_FIXED_OUT_ARGS |
4884 | |
4885 | if (lvaOutgoingArgSpaceVar != BAD_VAR_NUM) |
4886 | { |
4887 | varDsc = &lvaTable[lvaOutgoingArgSpaceVar]; |
4888 | varDsc->lvStkOffs = 0; |
4889 | varDsc->lvFramePointerBased = false; |
4890 | varDsc->lvMustInit = false; |
4891 | } |
4892 | |
4893 | #endif // FEATURE_FIXED_OUT_ARGS |
4894 | } |
4895 | |
4896 | #ifdef _TARGET_ARM_ |
4897 | bool Compiler::lvaIsPreSpilled(unsigned lclNum, regMaskTP preSpillMask) |
4898 | { |
4899 | const LclVarDsc& desc = lvaTable[lclNum]; |
4900 | return desc.lvIsRegArg && (preSpillMask & genRegMask(desc.lvArgReg)); |
4901 | } |
4902 | #endif // _TARGET_ARM_ |
4903 | |
4904 | /***************************************************************************** |
4905 | * lvaUpdateArgsWithInitialReg() : For each argument variable descriptor, update |
4906 | * its current register with the initial register as assigned by LSRA. |
4907 | */ |
4908 | void Compiler::lvaUpdateArgsWithInitialReg() |
4909 | { |
4910 | if (!compLSRADone) |
4911 | { |
4912 | return; |
4913 | } |
4914 | |
4915 | for (unsigned lclNum = 0; lclNum < info.compArgsCount; lclNum++) |
4916 | { |
4917 | LclVarDsc* varDsc = lvaTable + lclNum; |
4918 | |
4919 | if (varDsc->lvPromotedStruct()) |
4920 | { |
4921 | noway_assert(varDsc->lvFieldCnt == 1); // We only handle one field here |
4922 | |
4923 | unsigned fieldVarNum = varDsc->lvFieldLclStart; |
4924 | varDsc = lvaTable + fieldVarNum; |
4925 | } |
4926 | |
4927 | noway_assert(varDsc->lvIsParam); |
4928 | |
4929 | if (varDsc->lvIsRegCandidate()) |
4930 | { |
4931 | varDsc->lvRegNum = varDsc->lvArgInitReg; |
4932 | } |
4933 | } |
4934 | } |
4935 | |
4936 | /***************************************************************************** |
4937 | * lvaAssignVirtualFrameOffsetsToArgs() : Assign virtual stack offsets to the |
4938 | * arguments, and implicit arguments (this ptr, return buffer, generics, |
4939 | * and varargs). |
4940 | */ |
4941 | void Compiler::lvaAssignVirtualFrameOffsetsToArgs() |
4942 | { |
4943 | unsigned lclNum = 0; |
4944 | int argOffs = 0; |
4945 | #ifdef UNIX_AMD64_ABI |
4946 | int callerArgOffset = 0; |
4947 | #endif // UNIX_AMD64_ABI |
4948 | |
4949 | /* |
4950 | Assign stack offsets to arguments (in reverse order of passing). |
4951 | |
4952 | This means that if we pass arguments left->right, we start at |
4953 | the end of the list and work backwards, for right->left we start |
4954 | with the first argument and move forward. |
4955 | |
4956 | This is all relative to our Virtual '0' |
4957 | */ |
4958 | |
4959 | if (Target::g_tgtArgOrder == Target::ARG_ORDER_L2R) |
4960 | { |
4961 | argOffs = compArgSize; |
4962 | } |
4963 | |
4964 | /* Update the argOffs to reflect arguments that are passed in registers */ |
4965 | |
4966 | noway_assert(codeGen->intRegState.rsCalleeRegArgCount <= MAX_REG_ARG); |
4967 | noway_assert(compArgSize >= codeGen->intRegState.rsCalleeRegArgCount * REGSIZE_BYTES); |
4968 | |
4969 | #ifdef _TARGET_X86_ |
4970 | argOffs -= codeGen->intRegState.rsCalleeRegArgCount * REGSIZE_BYTES; |
4971 | #endif |
4972 | |
4973 | // Update the arg initial register locations. |
4974 | lvaUpdateArgsWithInitialReg(); |
4975 | |
4976 | /* Is there a "this" argument? */ |
4977 | |
4978 | if (!info.compIsStatic) |
4979 | { |
4980 | noway_assert(lclNum == info.compThisArg); |
4981 | #ifndef _TARGET_X86_ |
4982 | argOffs = |
4983 | lvaAssignVirtualFrameOffsetToArg(lclNum, REGSIZE_BYTES, argOffs UNIX_AMD64_ABI_ONLY_ARG(&callerArgOffset)); |
4984 | #endif // _TARGET_X86_ |
4985 | lclNum++; |
4986 | } |
4987 | |
4988 | /* if we have a hidden buffer parameter, that comes here */ |
4989 | |
4990 | if (info.compRetBuffArg != BAD_VAR_NUM) |
4991 | { |
4992 | noway_assert(lclNum == info.compRetBuffArg); |
4993 | noway_assert(lvaTable[lclNum].lvIsRegArg); |
4994 | #ifndef _TARGET_X86_ |
4995 | argOffs = |
4996 | lvaAssignVirtualFrameOffsetToArg(lclNum, REGSIZE_BYTES, argOffs UNIX_AMD64_ABI_ONLY_ARG(&callerArgOffset)); |
4997 | #endif // _TARGET_X86_ |
4998 | lclNum++; |
4999 | } |
5000 | |
5001 | #if USER_ARGS_COME_LAST |
5002 | |
5003 | //@GENERICS: extra argument for instantiation info |
5004 | if (info.compMethodInfo->args.callConv & CORINFO_CALLCONV_PARAMTYPE) |
5005 | { |
5006 | noway_assert(lclNum == (unsigned)info.compTypeCtxtArg); |
5007 | argOffs = lvaAssignVirtualFrameOffsetToArg(lclNum++, REGSIZE_BYTES, |
5008 | argOffs UNIX_AMD64_ABI_ONLY_ARG(&callerArgOffset)); |
5009 | } |
5010 | |
5011 | if (info.compIsVarArgs) |
5012 | { |
5013 | argOffs = lvaAssignVirtualFrameOffsetToArg(lclNum++, REGSIZE_BYTES, |
5014 | argOffs UNIX_AMD64_ABI_ONLY_ARG(&callerArgOffset)); |
5015 | } |
5016 | |
5017 | #endif // USER_ARGS_COME_LAST |
5018 | |
5019 | CORINFO_ARG_LIST_HANDLE argLst = info.compMethodInfo->args.args; |
5020 | unsigned argSigLen = info.compMethodInfo->args.numArgs; |
5021 | |
5022 | #ifdef _TARGET_ARM_ |
5023 | // |
5024 | // struct_n { int; int; ... n times }; |
5025 | // |
5026 | // Consider signature: |
5027 | // |
5028 | // Foo (float a,double b,float c,double d,float e,double f,float g,double h, |
5029 | // float i,double j,float k,double l,struct_3 m) { } |
5030 | // |
5031 | // Basically the signature is: (all float regs full, 1 double, struct_3); |
5032 | // |
5033 | // The double argument occurs before pre spill in the argument iteration and |
5034 | // computes an argOffset of 0. struct_3 offset becomes 8. This is wrong. |
5035 | // Because struct_3 is prespilled and double occurs after prespill. |
5036 | // The correct offsets are double = 16 (aligned stk), struct_3 = 0..12, |
5037 | // Offset 12 will be skipped for double alignment of double. |
5038 | // |
5039 | // Another example is (struct_2, all float regs full, double, struct_2); |
5040 | // Here, notice the order is similarly messed up because of 2 pre-spilled |
5041 | // struct_2. |
5042 | // |
5043 | // Succinctly, |
5044 | // ARG_INDEX(i) > ARG_INDEX(j) DOES NOT IMPLY |ARG_OFFSET(i)| > |ARG_OFFSET(j)| |
5045 | // |
5046 | // Therefore, we'll do a two pass offset calculation, one that considers pre-spill |
5047 | // and the next, stack args. |
5048 | // |
5049 | |
5050 | unsigned argLcls = 0; |
5051 | |
5052 | // Take care of pre spill registers first. |
5053 | regMaskTP preSpillMask = codeGen->regSet.rsMaskPreSpillRegs(false); |
5054 | regMaskTP tempMask = RBM_NONE; |
5055 | for (unsigned i = 0, preSpillLclNum = lclNum; i < argSigLen; ++i, ++preSpillLclNum) |
5056 | { |
5057 | if (lvaIsPreSpilled(preSpillLclNum, preSpillMask)) |
5058 | { |
5059 | unsigned argSize = eeGetArgSize(argLst, &info.compMethodInfo->args); |
5060 | argOffs = lvaAssignVirtualFrameOffsetToArg(preSpillLclNum, argSize, argOffs); |
5061 | argLcls++; |
5062 | |
5063 | // Early out if we can. If size is 8 and base reg is 2, then the mask is 0x1100 |
5064 | tempMask |= ((((1 << (roundUp(argSize, TARGET_POINTER_SIZE) / REGSIZE_BYTES))) - 1) |
5065 | << lvaTable[preSpillLclNum].lvArgReg); |
5066 | if (tempMask == preSpillMask) |
5067 | { |
5068 | // We won't encounter more pre-spilled registers, |
5069 | // so don't bother iterating further. |
5070 | break; |
5071 | } |
5072 | } |
5073 | argLst = info.compCompHnd->getArgNext(argLst); |
5074 | } |
5075 | |
5076 | // Take care of non pre-spilled stack arguments. |
5077 | argLst = info.compMethodInfo->args.args; |
5078 | for (unsigned i = 0, stkLclNum = lclNum; i < argSigLen; ++i, ++stkLclNum) |
5079 | { |
5080 | if (!lvaIsPreSpilled(stkLclNum, preSpillMask)) |
5081 | { |
5082 | argOffs = |
5083 | lvaAssignVirtualFrameOffsetToArg(stkLclNum, eeGetArgSize(argLst, &info.compMethodInfo->args), argOffs); |
5084 | argLcls++; |
5085 | } |
5086 | argLst = info.compCompHnd->getArgNext(argLst); |
5087 | } |
5088 | |
5089 | lclNum += argLcls; |
5090 | #else // !_TARGET_ARM_ |
5091 | for (unsigned i = 0; i < argSigLen; i++) |
5092 | { |
5093 | unsigned argumentSize = eeGetArgSize(argLst, &info.compMethodInfo->args); |
5094 | |
5095 | #ifdef UNIX_AMD64_ABI |
5096 | // On the stack frame the homed arg always takes a full number of slots |
5097 | // for proper stack alignment. Make sure the real struct size is properly rounded up. |
5098 | argumentSize = roundUp(argumentSize, TARGET_POINTER_SIZE); |
5099 | #endif // UNIX_AMD64_ABI |
5100 | |
5101 | argOffs = |
5102 | lvaAssignVirtualFrameOffsetToArg(lclNum++, argumentSize, argOffs UNIX_AMD64_ABI_ONLY_ARG(&callerArgOffset)); |
5103 | argLst = info.compCompHnd->getArgNext(argLst); |
5104 | } |
5105 | #endif // !_TARGET_ARM_ |
5106 | |
5107 | #if !USER_ARGS_COME_LAST |
5108 | |
5109 | //@GENERICS: extra argument for instantiation info |
5110 | if (info.compMethodInfo->args.callConv & CORINFO_CALLCONV_PARAMTYPE) |
5111 | { |
5112 | noway_assert(lclNum == (unsigned)info.compTypeCtxtArg); |
5113 | argOffs = lvaAssignVirtualFrameOffsetToArg(lclNum++, REGSIZE_BYTES, |
5114 | argOffs UNIX_AMD64_ABI_ONLY_ARG(&callerArgOffset)); |
5115 | } |
5116 | |
5117 | if (info.compIsVarArgs) |
5118 | { |
5119 | argOffs = lvaAssignVirtualFrameOffsetToArg(lclNum++, REGSIZE_BYTES, |
5120 | argOffs UNIX_AMD64_ABI_ONLY_ARG(&callerArgOffset)); |
5121 | } |
5122 | |
5123 | #endif // USER_ARGS_COME_LAST |
5124 | } |
5125 | |
5126 | #ifdef UNIX_AMD64_ABI |
5127 | // |
5128 | // lvaAssignVirtualFrameOffsetToArg() : Assign virtual stack offsets to an |
5129 | // individual argument, and return the offset for the next argument. |
5130 | // Note: This method only calculates the initial offset of the stack passed/spilled arguments |
5131 | // (if any - the RA might decide to spill(home on the stack) register passed arguments, if rarely used.) |
5132 | // The final offset is calculated in lvaFixVirtualFrameOffsets method. It accounts for FP existance, |
5133 | // ret address slot, stack frame padding, alloca instructions, etc. |
5134 | // Note: This is the implementation for UNIX_AMD64 System V platforms. |
5135 | // |
5136 | int Compiler::lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, |
5137 | unsigned argSize, |
5138 | int argOffs UNIX_AMD64_ABI_ONLY_ARG(int* callerArgOffset)) |
5139 | { |
5140 | noway_assert(lclNum < info.compArgsCount); |
5141 | noway_assert(argSize); |
5142 | |
5143 | if (Target::g_tgtArgOrder == Target::ARG_ORDER_L2R) |
5144 | { |
5145 | argOffs -= argSize; |
5146 | } |
5147 | |
5148 | unsigned fieldVarNum = BAD_VAR_NUM; |
5149 | |
5150 | noway_assert(lclNum < lvaCount); |
5151 | LclVarDsc* varDsc = lvaTable + lclNum; |
5152 | |
5153 | if (varDsc->lvPromotedStruct()) |
5154 | { |
5155 | noway_assert(varDsc->lvFieldCnt == 1); // We only handle one field here |
5156 | fieldVarNum = varDsc->lvFieldLclStart; |
5157 | |
5158 | lvaPromotionType promotionType = lvaGetPromotionType(varDsc); |
5159 | |
5160 | if (promotionType == PROMOTION_TYPE_INDEPENDENT) |
5161 | { |
5162 | lclNum = fieldVarNum; |
5163 | noway_assert(lclNum < lvaCount); |
5164 | varDsc = lvaTable + lclNum; |
5165 | assert(varDsc->lvIsStructField); |
5166 | } |
5167 | } |
5168 | |
5169 | noway_assert(varDsc->lvIsParam); |
5170 | |
5171 | if (varDsc->lvIsRegArg) |
5172 | { |
5173 | // Argument is passed in a register, don't count it |
5174 | // when updating the current offset on the stack. |
5175 | |
5176 | if (varDsc->lvOnFrame) |
5177 | { |
5178 | // The offset for args needs to be set only for the stack homed arguments for System V. |
5179 | varDsc->lvStkOffs = argOffs; |
5180 | } |
5181 | else |
5182 | { |
5183 | varDsc->lvStkOffs = 0; |
5184 | } |
5185 | } |
5186 | else |
5187 | { |
5188 | // For Windows AMD64 there are 4 slots for the register passed arguments on the top of the caller's stack. |
5189 | // This is where they are always homed. So, they can be accessed with positive offset. |
5190 | // On System V platforms, if the RA decides to home a register passed arg on the stack, it creates a stack |
5191 | // location on the callee stack (like any other local var.) In such a case, the register passed, stack homed |
5192 | // arguments are accessed using negative offsets and the stack passed arguments are accessed using positive |
5193 | // offset (from the caller's stack.) |
5194 | // For System V platforms if there is no frame pointer the caller stack parameter offset should include the |
5195 | // callee allocated space. If frame register is used, the callee allocated space should not be included for |
5196 | // accessing the caller stack parameters. The last two requirements are met in lvaFixVirtualFrameOffsets |
5197 | // method, which fixes the offsets, based on frame pointer existence, existence of alloca instructions, ret |
5198 | // address pushed, ets. |
5199 | |
5200 | varDsc->lvStkOffs = *callerArgOffset; |
5201 | // Structs passed on stack could be of size less than TARGET_POINTER_SIZE. |
5202 | // Make sure they get at least TARGET_POINTER_SIZE on the stack - this is required for alignment. |
5203 | if (argSize > TARGET_POINTER_SIZE) |
5204 | { |
5205 | *callerArgOffset += (int)roundUp(argSize, TARGET_POINTER_SIZE); |
5206 | } |
5207 | else |
5208 | { |
5209 | *callerArgOffset += TARGET_POINTER_SIZE; |
5210 | } |
5211 | } |
5212 | |
5213 | // For struct promoted parameters we need to set the offsets for both LclVars. |
5214 | // |
5215 | // For a dependent promoted struct we also assign the struct fields stack offset |
5216 | if (varDsc->lvPromotedStruct()) |
5217 | { |
5218 | lvaPromotionType promotionType = lvaGetPromotionType(varDsc); |
5219 | |
5220 | if (promotionType == PROMOTION_TYPE_DEPENDENT) |
5221 | { |
5222 | noway_assert(varDsc->lvFieldCnt == 1); // We only handle one field here |
5223 | |
5224 | assert(fieldVarNum == varDsc->lvFieldLclStart); |
5225 | lvaTable[fieldVarNum].lvStkOffs = varDsc->lvStkOffs; |
5226 | } |
5227 | } |
5228 | // For an independent promoted struct field we also assign the parent struct stack offset |
5229 | else if (varDsc->lvIsStructField) |
5230 | { |
5231 | noway_assert(varDsc->lvParentLcl < lvaCount); |
5232 | lvaTable[varDsc->lvParentLcl].lvStkOffs = varDsc->lvStkOffs; |
5233 | } |
5234 | |
5235 | if (Target::g_tgtArgOrder == Target::ARG_ORDER_R2L && !varDsc->lvIsRegArg) |
5236 | { |
5237 | argOffs += argSize; |
5238 | } |
5239 | |
5240 | return argOffs; |
5241 | } |
5242 | |
5243 | #else // !UNIX_AMD64_ABI |
5244 | |
5245 | // |
5246 | // lvaAssignVirtualFrameOffsetToArg() : Assign virtual stack offsets to an |
5247 | // individual argument, and return the offset for the next argument. |
5248 | // Note: This method only calculates the initial offset of the stack passed/spilled arguments |
5249 | // (if any - the RA might decide to spill(home on the stack) register passed arguments, if rarely used.) |
5250 | // The final offset is calculated in lvaFixVirtualFrameOffsets method. It accounts for FP existance, |
5251 | // ret address slot, stack frame padding, alloca instructions, etc. |
5252 | // Note: This implementation for all the platforms but UNIX_AMD64 OSs (System V 64 bit.) |
5253 | int Compiler::lvaAssignVirtualFrameOffsetToArg(unsigned lclNum, |
5254 | unsigned argSize, |
5255 | int argOffs UNIX_AMD64_ABI_ONLY_ARG(int* callerArgOffset)) |
5256 | { |
5257 | noway_assert(lclNum < info.compArgsCount); |
5258 | noway_assert(argSize); |
5259 | |
5260 | if (Target::g_tgtArgOrder == Target::ARG_ORDER_L2R) |
5261 | { |
5262 | argOffs -= argSize; |
5263 | } |
5264 | |
5265 | unsigned fieldVarNum = BAD_VAR_NUM; |
5266 | |
5267 | noway_assert(lclNum < lvaCount); |
5268 | LclVarDsc* varDsc = lvaTable + lclNum; |
5269 | |
5270 | if (varDsc->lvPromotedStruct()) |
5271 | { |
5272 | noway_assert(varDsc->lvFieldCnt == 1); // We only handle one field here |
5273 | fieldVarNum = varDsc->lvFieldLclStart; |
5274 | |
5275 | lvaPromotionType promotionType = lvaGetPromotionType(varDsc); |
5276 | |
5277 | if (promotionType == PROMOTION_TYPE_INDEPENDENT) |
5278 | { |
5279 | lclNum = fieldVarNum; |
5280 | noway_assert(lclNum < lvaCount); |
5281 | varDsc = lvaTable + lclNum; |
5282 | assert(varDsc->lvIsStructField); |
5283 | } |
5284 | } |
5285 | |
5286 | noway_assert(varDsc->lvIsParam); |
5287 | |
5288 | if (varDsc->lvIsRegArg) |
5289 | { |
5290 | /* Argument is passed in a register, don't count it |
5291 | * when updating the current offset on the stack */ |
5292 | CLANG_FORMAT_COMMENT_ANCHOR; |
5293 | |
5294 | #if !defined(_TARGET_ARMARCH_) |
5295 | #if DEBUG |
5296 | // TODO: Remove this noway_assert and replace occurrences of TARGET_POINTER_SIZE with argSize |
5297 | // Also investigate why we are incrementing argOffs for X86 as this seems incorrect |
5298 | // |
5299 | noway_assert(argSize == TARGET_POINTER_SIZE); |
5300 | #endif // DEBUG |
5301 | #endif |
5302 | |
5303 | #if defined(_TARGET_X86_) |
5304 | argOffs += TARGET_POINTER_SIZE; |
5305 | #elif defined(_TARGET_AMD64_) |
5306 | // Register arguments on AMD64 also takes stack space. (in the backing store) |
5307 | varDsc->lvStkOffs = argOffs; |
5308 | argOffs += TARGET_POINTER_SIZE; |
5309 | #elif defined(_TARGET_ARM64_) |
5310 | // Register arguments on ARM64 only take stack space when they have a frame home. |
5311 | // Unless on windows and in a vararg method. |
5312 | #if FEATURE_ARG_SPLIT |
5313 | if (this->info.compIsVarArgs) |
5314 | { |
5315 | if (varDsc->lvType == TYP_STRUCT && varDsc->lvOtherArgReg >= MAX_REG_ARG && varDsc->lvOtherArgReg != REG_NA) |
5316 | { |
5317 | // This is a split struct. It will account for an extra (8 bytes) |
5318 | // of alignment. |
5319 | varDsc->lvStkOffs += TARGET_POINTER_SIZE; |
5320 | argOffs += TARGET_POINTER_SIZE; |
5321 | } |
5322 | } |
5323 | #endif // FEATURE_ARG_SPLIT |
5324 | |
5325 | #elif defined(_TARGET_ARM_) |
5326 | // On ARM we spill the registers in codeGen->regSet.rsMaskPreSpillRegArg |
5327 | // in the prolog, so we have to fill in lvStkOffs here |
5328 | // |
5329 | regMaskTP regMask = genRegMask(varDsc->lvArgReg); |
5330 | if (codeGen->regSet.rsMaskPreSpillRegArg & regMask) |
5331 | { |
5332 | // Signature: void foo(struct_8, int, struct_4) |
5333 | // ------- CALLER SP ------- |
5334 | // r3 struct_4 |
5335 | // r2 int - not prespilled, but added for alignment. argOffs should skip this. |
5336 | // r1 struct_8 |
5337 | // r0 struct_8 |
5338 | // ------------------------- |
5339 | // If we added alignment we need to fix argOffs for all registers above alignment. |
5340 | if (codeGen->regSet.rsMaskPreSpillAlign != RBM_NONE) |
5341 | { |
5342 | assert(genCountBits(codeGen->regSet.rsMaskPreSpillAlign) == 1); |
5343 | // Is register beyond the alignment pos? |
5344 | if (regMask > codeGen->regSet.rsMaskPreSpillAlign) |
5345 | { |
5346 | // Increment argOffs just once for the _first_ register after alignment pos |
5347 | // in the prespill mask. |
5348 | if (!BitsBetween(codeGen->regSet.rsMaskPreSpillRegArg, regMask, |
5349 | codeGen->regSet.rsMaskPreSpillAlign)) |
5350 | { |
5351 | argOffs += TARGET_POINTER_SIZE; |
5352 | } |
5353 | } |
5354 | } |
5355 | |
5356 | switch (varDsc->lvType) |
5357 | { |
5358 | case TYP_STRUCT: |
5359 | if (!varDsc->lvStructDoubleAlign) |
5360 | { |
5361 | break; |
5362 | } |
5363 | __fallthrough; |
5364 | |
5365 | case TYP_DOUBLE: |
5366 | case TYP_LONG: |
5367 | { |
5368 | // |
5369 | // Let's assign offsets to arg1, a double in r2. argOffs has to be 4 not 8. |
5370 | // |
5371 | // ------- CALLER SP ------- |
5372 | // r3 |
5373 | // r2 double -- argOffs = 4, but it doesn't need to be skipped, because there is no skipping. |
5374 | // r1 VACookie -- argOffs = 0 |
5375 | // ------------------------- |
5376 | // |
5377 | // Consider argOffs as if it accounts for number of prespilled registers before the current |
5378 | // register. In the above example, for r2, it is r1 that is prespilled, but since r1 is |
5379 | // accounted for by argOffs being 4, there should have been no skipping. Instead, if we didn't |
5380 | // assign r1 to any variable, then argOffs would still be 0 which implies it is not accounting |
5381 | // for r1, equivalently r1 is skipped. |
5382 | // |
5383 | // If prevRegsSize is unaccounted for by a corresponding argOffs, we must have skipped a register. |
5384 | int prevRegsSize = |
5385 | genCountBits(codeGen->regSet.rsMaskPreSpillRegArg & (regMask - 1)) * TARGET_POINTER_SIZE; |
5386 | if (argOffs < prevRegsSize) |
5387 | { |
5388 | // We must align up the argOffset to a multiple of 8 to account for skipped registers. |
5389 | argOffs = roundUp((unsigned)argOffs, 2 * TARGET_POINTER_SIZE); |
5390 | } |
5391 | // We should've skipped only a single register. |
5392 | assert(argOffs == prevRegsSize); |
5393 | } |
5394 | break; |
5395 | |
5396 | default: |
5397 | // No alignment of argOffs required |
5398 | break; |
5399 | } |
5400 | varDsc->lvStkOffs = argOffs; |
5401 | argOffs += argSize; |
5402 | } |
5403 | #else // _TARGET_* |
5404 | #error Unsupported or unset target architecture |
5405 | #endif // _TARGET_* |
5406 | } |
5407 | else |
5408 | { |
5409 | #if defined(_TARGET_ARM_) |
5410 | // Dev11 Bug 42817: incorrect codegen for DrawFlatCheckBox causes A/V in WinForms |
5411 | // |
5412 | // Here we have method with a signature (int a1, struct a2, struct a3, int a4, int a5). |
5413 | // Struct parameter 'a2' is 16-bytes with no alignment requirements; |
5414 | // it uses r1,r2,r3 and [OutArg+0] when passed. |
5415 | // Struct parameter 'a3' is 16-bytes that is required to be double aligned; |
5416 | // the caller skips [OutArg+4] and starts the argument at [OutArg+8]. |
5417 | // Thus the caller generates the correct code to pass the arguments. |
5418 | // When generating code to receive the arguments we set codeGen->regSet.rsMaskPreSpillRegArg to [r1,r2,r3] |
5419 | // and spill these three registers as the first instruction in the prolog. |
5420 | // Then when we layout the arguments' stack offsets we have an argOffs 0 which |
5421 | // points at the location that we spilled r1 into the stack. For this first |
5422 | // struct we take the lvIsRegArg path above with "codeGen->regSet.rsMaskPreSpillRegArg &" matching. |
5423 | // Next when we calculate the argOffs for the second 16-byte struct we have an argOffs |
5424 | // of 16, which appears to be aligned properly so we don't skip a stack slot. |
5425 | // |
5426 | // To fix this we must recover the actual OutArg offset by subtracting off the |
5427 | // sizeof of the PreSpill register args. |
5428 | // Then we align this offset to a multiple of 8 and add back the sizeof |
5429 | // of the PreSpill register args. |
5430 | // |
5431 | // Dev11 Bug 71767: failure of assert(sizeofPreSpillRegArgs <= argOffs) |
5432 | // |
5433 | // We have a method with 'this' passed in r0, RetBuf arg in r1, VarArgs cookie |
5434 | // in r2. The first user arg is a 144 byte struct with double alignment required, |
5435 | // r3 is skipped, and the struct is passed on the stack. However, 'r3' is added |
5436 | // to the codeGen->regSet.rsMaskPreSpillRegArg mask by the VarArgs cookie code, since we need to |
5437 | // home all the potential varargs arguments in registers, even if we don't have |
5438 | // signature type information for the variadic arguments. However, due to alignment, |
5439 | // we have skipped a register that doesn't have a corresponding symbol. Make up |
5440 | // for that by increasing argOffs here. |
5441 | // |
5442 | |
5443 | int sizeofPreSpillRegArgs = genCountBits(codeGen->regSet.rsMaskPreSpillRegs(true)) * REGSIZE_BYTES; |
5444 | |
5445 | if (argOffs < sizeofPreSpillRegArgs) |
5446 | { |
5447 | // This can only happen if we skipped the last register spot because current stk arg |
5448 | // is a struct requiring alignment or a pre-spill alignment was required because the |
5449 | // first reg arg needed alignment. |
5450 | // |
5451 | // Example 1: First Stk Argument requiring alignment in vararg case (same as above comment.) |
5452 | // Signature (int a0, int a1, int a2, struct {long} a3, ...) |
5453 | // |
5454 | // stk arg a3 --> argOffs here will be 12 (r0-r2) but pre-spill will be 16. |
5455 | // ---- Caller SP ---- |
5456 | // r3 --> Stack slot is skipped in this case. |
5457 | // r2 int a2 |
5458 | // r1 int a1 |
5459 | // r0 int a0 |
5460 | // |
5461 | // Example 2: First Reg Argument requiring alignment in no-vararg case. |
5462 | // Signature (struct {long} a0, struct {int} a1, int a2, int a3) |
5463 | // |
5464 | // stk arg --> argOffs here will be 12 {r0-r2} but pre-spill will be 16. |
5465 | // ---- Caller SP ---- |
5466 | // r3 int a2 --> pushed (not pre-spilled) for alignment of a0 by lvaInitUserArgs. |
5467 | // r2 struct { int } a1 |
5468 | // r0-r1 struct { long } a0 |
5469 | CLANG_FORMAT_COMMENT_ANCHOR; |
5470 | |
5471 | #ifdef PROFILING_SUPPORTED |
5472 | // On Arm under profiler, r0-r3 are always prespilled on stack. |
5473 | // It is possible to have methods that accept only HFAs as parameters e.g. Signature(struct hfa1, struct |
5474 | // hfa2), in which case hfa1 and hfa2 will be en-registered in co-processor registers and will have an |
5475 | // argument offset less than size of preSpill. |
5476 | // |
5477 | // For this reason the following conditions are asserted when not under profiler. |
5478 | if (!compIsProfilerHookNeeded()) |
5479 | #endif |
5480 | { |
5481 | bool cond = ((info.compIsVarArgs || opts.compUseSoftFP) && |
5482 | // Does cur stk arg require double alignment? |
5483 | ((varDsc->lvType == TYP_STRUCT && varDsc->lvStructDoubleAlign) || |
5484 | (varDsc->lvType == TYP_DOUBLE) || (varDsc->lvType == TYP_LONG))) || |
5485 | // Did first reg arg require alignment? |
5486 | (codeGen->regSet.rsMaskPreSpillAlign & genRegMask(REG_ARG_LAST)); |
5487 | |
5488 | noway_assert(cond); |
5489 | noway_assert(sizeofPreSpillRegArgs <= |
5490 | argOffs + TARGET_POINTER_SIZE); // at most one register of alignment |
5491 | } |
5492 | argOffs = sizeofPreSpillRegArgs; |
5493 | } |
5494 | |
5495 | noway_assert(argOffs >= sizeofPreSpillRegArgs); |
5496 | int argOffsWithoutPreSpillRegArgs = argOffs - sizeofPreSpillRegArgs; |
5497 | |
5498 | switch (varDsc->lvType) |
5499 | { |
5500 | case TYP_STRUCT: |
5501 | if (!varDsc->lvStructDoubleAlign) |
5502 | break; |
5503 | |
5504 | __fallthrough; |
5505 | |
5506 | case TYP_DOUBLE: |
5507 | case TYP_LONG: |
5508 | // We must align up the argOffset to a multiple of 8 |
5509 | argOffs = |
5510 | roundUp((unsigned)argOffsWithoutPreSpillRegArgs, 2 * TARGET_POINTER_SIZE) + sizeofPreSpillRegArgs; |
5511 | break; |
5512 | |
5513 | default: |
5514 | // No alignment of argOffs required |
5515 | break; |
5516 | } |
5517 | #endif // _TARGET_ARM_ |
5518 | |
5519 | varDsc->lvStkOffs = argOffs; |
5520 | } |
5521 | |
5522 | // For struct promoted parameters we need to set the offsets for both LclVars. |
5523 | // |
5524 | // For a dependent promoted struct we also assign the struct fields stack offset |
5525 | CLANG_FORMAT_COMMENT_ANCHOR; |
5526 | |
5527 | #if !defined(_TARGET_64BIT_) |
5528 | if ((varDsc->TypeGet() == TYP_LONG) && varDsc->lvPromoted) |
5529 | { |
5530 | noway_assert(varDsc->lvFieldCnt == 2); |
5531 | fieldVarNum = varDsc->lvFieldLclStart; |
5532 | lvaTable[fieldVarNum].lvStkOffs = varDsc->lvStkOffs; |
5533 | lvaTable[fieldVarNum + 1].lvStkOffs = varDsc->lvStkOffs + genTypeSize(TYP_INT); |
5534 | } |
5535 | else |
5536 | #endif // !defined(_TARGET_64BIT_) |
5537 | if (varDsc->lvPromotedStruct()) |
5538 | { |
5539 | lvaPromotionType promotionType = lvaGetPromotionType(varDsc); |
5540 | |
5541 | if (promotionType == PROMOTION_TYPE_DEPENDENT) |
5542 | { |
5543 | noway_assert(varDsc->lvFieldCnt == 1); // We only handle one field here |
5544 | |
5545 | assert(fieldVarNum == varDsc->lvFieldLclStart); |
5546 | lvaTable[fieldVarNum].lvStkOffs = varDsc->lvStkOffs; |
5547 | } |
5548 | } |
5549 | // For an independent promoted struct field we also assign the parent struct stack offset |
5550 | else if (varDsc->lvIsStructField) |
5551 | { |
5552 | noway_assert(varDsc->lvParentLcl < lvaCount); |
5553 | lvaTable[varDsc->lvParentLcl].lvStkOffs = varDsc->lvStkOffs; |
5554 | } |
5555 | |
5556 | if (Target::g_tgtArgOrder == Target::ARG_ORDER_R2L && !varDsc->lvIsRegArg) |
5557 | { |
5558 | argOffs += argSize; |
5559 | } |
5560 | |
5561 | return argOffs; |
5562 | } |
5563 | #endif // !UNIX_AMD64_ABI |
5564 | |
5565 | /***************************************************************************** |
5566 | * lvaAssignVirtualFrameOffsetsToLocals() : Assign virtual stack offsets to |
5567 | * locals, temps, and anything else. These will all be negative offsets |
5568 | * (stack grows down) relative to the virtual '0'/return address |
5569 | */ |
5570 | void Compiler::lvaAssignVirtualFrameOffsetsToLocals() |
5571 | { |
5572 | int stkOffs = 0; |
5573 | // codeGen->isFramePointerUsed is set in regalloc phase. Initialize it to a guess for pre-regalloc layout. |
5574 | if (lvaDoneFrameLayout <= PRE_REGALLOC_FRAME_LAYOUT) |
5575 | { |
5576 | codeGen->setFramePointerUsed(codeGen->isFramePointerRequired()); |
5577 | } |
5578 | |
5579 | #ifdef _TARGET_XARCH_ |
5580 | // On x86/amd64, the return address has already been pushed by the call instruction in the caller. |
5581 | stkOffs -= TARGET_POINTER_SIZE; // return address; |
5582 | |
5583 | // TODO-AMD64-CQ: for X64 eventually this should be pushed with all the other |
5584 | // calleeregs. When you fix this, you'll also need to fix |
5585 | // the assert at the bottom of this method |
5586 | if (codeGen->doubleAlignOrFramePointerUsed()) |
5587 | { |
5588 | stkOffs -= REGSIZE_BYTES; |
5589 | } |
5590 | #endif //_TARGET_XARCH_ |
5591 | |
5592 | int preSpillSize = 0; |
5593 | bool mustDoubleAlign = false; |
5594 | |
5595 | #ifdef _TARGET_ARM_ |
5596 | mustDoubleAlign = true; |
5597 | preSpillSize = genCountBits(codeGen->regSet.rsMaskPreSpillRegs(true)) * REGSIZE_BYTES; |
5598 | #else // !_TARGET_ARM_ |
5599 | #if DOUBLE_ALIGN |
5600 | if (genDoubleAlign()) |
5601 | { |
5602 | mustDoubleAlign = true; // X86 only |
5603 | } |
5604 | #endif |
5605 | #endif // !_TARGET_ARM_ |
5606 | |
5607 | #ifdef _TARGET_ARM64_ |
5608 | // If the frame pointer is used, then we'll save FP/LR at the bottom of the stack. |
5609 | // Otherwise, we won't store FP, and we'll store LR at the top, with the other callee-save |
5610 | // registers (if any). |
5611 | |
5612 | int initialStkOffs = 0; |
5613 | if (info.compIsVarArgs) |
5614 | { |
5615 | // For varargs we always save all of the integer register arguments |
5616 | // so that they are contiguous with the incoming stack arguments. |
5617 | initialStkOffs = MAX_REG_ARG * REGSIZE_BYTES; |
5618 | stkOffs -= initialStkOffs; |
5619 | } |
5620 | |
5621 | if (isFramePointerUsed()) |
5622 | { |
5623 | // Subtract off FP and LR. |
5624 | assert(compCalleeRegsPushed >= 2); |
5625 | stkOffs -= (compCalleeRegsPushed - 2) * REGSIZE_BYTES; |
5626 | } |
5627 | else |
5628 | { |
5629 | stkOffs -= compCalleeRegsPushed * REGSIZE_BYTES; |
5630 | } |
5631 | |
5632 | #else // !_TARGET_ARM64_ |
5633 | stkOffs -= compCalleeRegsPushed * REGSIZE_BYTES; |
5634 | #endif // !_TARGET_ARM64_ |
5635 | |
5636 | compLclFrameSize = 0; |
5637 | |
5638 | #ifdef _TARGET_AMD64_ |
5639 | // In case of Amd64 compCalleeRegsPushed includes float regs (Xmm6-xmm15) that |
5640 | // need to be pushed. But Amd64 doesn't support push/pop of xmm registers. |
5641 | // Instead we need to allocate space for them on the stack and save them in prolog. |
5642 | // Therefore, we consider xmm registers being saved while computing stack offsets |
5643 | // but space for xmm registers is considered part of compLclFrameSize. |
5644 | // Notes |
5645 | // 1) We need to save the entire 128-bits of xmm register to stack, since amd64 |
5646 | // prolog unwind codes allow encoding of an instruction that stores the entire xmm reg |
5647 | // at an offset relative to SP |
5648 | // 2) We adjust frame size so that SP is aligned at 16-bytes after pushing integer registers. |
5649 | // This means while saving the first xmm register to its allocated stack location we might |
5650 | // have to skip 8-bytes. The reason for padding is to use efficient "movaps" to save/restore |
5651 | // xmm registers to/from stack to match Jit64 codegen. Without the aligning on 16-byte |
5652 | // boundary we would have to use movups when offset turns out unaligned. Movaps is more |
5653 | // performant than movups. |
5654 | unsigned calleeFPRegsSavedSize = genCountBits(compCalleeFPRegsSavedMask) * XMM_REGSIZE_BYTES; |
5655 | if (calleeFPRegsSavedSize > 0 && ((stkOffs % XMM_REGSIZE_BYTES) != 0)) |
5656 | { |
5657 | // Take care of alignment |
5658 | int alignPad = (int)AlignmentPad((unsigned)-stkOffs, XMM_REGSIZE_BYTES); |
5659 | stkOffs -= alignPad; |
5660 | lvaIncrementFrameSize(alignPad); |
5661 | } |
5662 | |
5663 | stkOffs -= calleeFPRegsSavedSize; |
5664 | lvaIncrementFrameSize(calleeFPRegsSavedSize); |
5665 | |
5666 | // Quirk for VS debug-launch scenario to work |
5667 | if (compVSQuirkStackPaddingNeeded > 0) |
5668 | { |
5669 | #ifdef DEBUG |
5670 | if (verbose) |
5671 | { |
5672 | printf("\nAdding VS quirk stack padding of %d bytes between save-reg area and locals\n" , |
5673 | compVSQuirkStackPaddingNeeded); |
5674 | } |
5675 | #endif // DEBUG |
5676 | |
5677 | stkOffs -= compVSQuirkStackPaddingNeeded; |
5678 | lvaIncrementFrameSize(compVSQuirkStackPaddingNeeded); |
5679 | } |
5680 | #endif //_TARGET_AMD64_ |
5681 | |
5682 | #if FEATURE_EH_FUNCLETS && defined(_TARGET_ARMARCH_) |
5683 | if (lvaPSPSym != BAD_VAR_NUM) |
5684 | { |
5685 | // On ARM/ARM64, if we need a PSPSym, allocate it first, before anything else, including |
5686 | // padding (so we can avoid computing the same padding in the funclet |
5687 | // frame). Note that there is no special padding requirement for the PSPSym. |
5688 | noway_assert(codeGen->isFramePointerUsed()); // We need an explicit frame pointer |
5689 | stkOffs = lvaAllocLocalAndSetVirtualOffset(lvaPSPSym, TARGET_POINTER_SIZE, stkOffs); |
5690 | } |
5691 | #endif // FEATURE_EH_FUNCLETS && defined(_TARGET_ARMARCH_) |
5692 | |
5693 | if (mustDoubleAlign) |
5694 | { |
5695 | if (lvaDoneFrameLayout != FINAL_FRAME_LAYOUT) |
5696 | { |
5697 | // Allocate a pointer sized stack slot, since we may need to double align here |
5698 | // when lvaDoneFrameLayout == FINAL_FRAME_LAYOUT |
5699 | // |
5700 | lvaIncrementFrameSize(TARGET_POINTER_SIZE); |
5701 | stkOffs -= TARGET_POINTER_SIZE; |
5702 | |
5703 | // If we have any TYP_LONG, TYP_DOUBLE or double aligned structs |
5704 | // then we need to allocate a second pointer sized stack slot, |
5705 | // since we may need to double align that LclVar when we see it |
5706 | // in the loop below. We will just always do this so that the |
5707 | // offsets that we calculate for the stack frame will always |
5708 | // be greater (or equal) to what they can be in the final layout. |
5709 | // |
5710 | lvaIncrementFrameSize(TARGET_POINTER_SIZE); |
5711 | stkOffs -= TARGET_POINTER_SIZE; |
5712 | } |
5713 | else // FINAL_FRAME_LAYOUT |
5714 | { |
5715 | if (((stkOffs + preSpillSize) % (2 * TARGET_POINTER_SIZE)) != 0) |
5716 | { |
5717 | lvaIncrementFrameSize(TARGET_POINTER_SIZE); |
5718 | stkOffs -= TARGET_POINTER_SIZE; |
5719 | } |
5720 | // We should now have a double-aligned (stkOffs+preSpillSize) |
5721 | noway_assert(((stkOffs + preSpillSize) % (2 * TARGET_POINTER_SIZE)) == 0); |
5722 | } |
5723 | } |
5724 | |
5725 | if (lvaMonAcquired != BAD_VAR_NUM) |
5726 | { |
5727 | // This var must go first, in what is called the 'frame header' for EnC so that it is |
5728 | // preserved when remapping occurs. See vm\eetwain.cpp for detailed comment specifying frame |
5729 | // layout requirements for EnC to work. |
5730 | stkOffs = lvaAllocLocalAndSetVirtualOffset(lvaMonAcquired, lvaLclSize(lvaMonAcquired), stkOffs); |
5731 | } |
5732 | |
5733 | if (opts.compNeedSecurityCheck) |
5734 | { |
5735 | #ifdef JIT32_GCENCODER |
5736 | /* This can't work without an explicit frame, so make sure */ |
5737 | noway_assert(codeGen->isFramePointerUsed()); |
5738 | #endif |
5739 | stkOffs = lvaAllocLocalAndSetVirtualOffset(lvaSecurityObject, TARGET_POINTER_SIZE, stkOffs); |
5740 | } |
5741 | |
5742 | #ifdef JIT32_GCENCODER |
5743 | if (lvaLocAllocSPvar != BAD_VAR_NUM) |
5744 | { |
5745 | noway_assert(codeGen->isFramePointerUsed()); // else offsets of locals of frameless methods will be incorrect |
5746 | stkOffs = lvaAllocLocalAndSetVirtualOffset(lvaLocAllocSPvar, TARGET_POINTER_SIZE, stkOffs); |
5747 | } |
5748 | #endif // JIT32_GCENCODER |
5749 | |
5750 | if (lvaReportParamTypeArg()) |
5751 | { |
5752 | #ifdef JIT32_GCENCODER |
5753 | noway_assert(codeGen->isFramePointerUsed()); |
5754 | #endif |
5755 | // For CORINFO_CALLCONV_PARAMTYPE (if needed) |
5756 | lvaIncrementFrameSize(TARGET_POINTER_SIZE); |
5757 | stkOffs -= TARGET_POINTER_SIZE; |
5758 | lvaCachedGenericContextArgOffs = stkOffs; |
5759 | } |
5760 | #ifndef JIT32_GCENCODER |
5761 | else if (lvaKeepAliveAndReportThis()) |
5762 | { |
5763 | // When "this" is also used as generic context arg. |
5764 | lvaIncrementFrameSize(TARGET_POINTER_SIZE); |
5765 | stkOffs -= TARGET_POINTER_SIZE; |
5766 | lvaCachedGenericContextArgOffs = stkOffs; |
5767 | } |
5768 | #endif |
5769 | |
5770 | #if !FEATURE_EH_FUNCLETS |
5771 | /* If we need space for slots for shadow SP, reserve it now */ |
5772 | if (ehNeedsShadowSPslots()) |
5773 | { |
5774 | noway_assert(codeGen->isFramePointerUsed()); // else offsets of locals of frameless methods will be incorrect |
5775 | if (!lvaReportParamTypeArg()) |
5776 | { |
5777 | #ifndef JIT32_GCENCODER |
5778 | if (!lvaKeepAliveAndReportThis()) |
5779 | #endif |
5780 | { |
5781 | // In order to keep the gc info encoding smaller, the VM assumes that all methods with EH |
5782 | // have also saved space for a ParamTypeArg, so we need to do that here |
5783 | lvaIncrementFrameSize(TARGET_POINTER_SIZE); |
5784 | stkOffs -= TARGET_POINTER_SIZE; |
5785 | } |
5786 | } |
5787 | stkOffs = lvaAllocLocalAndSetVirtualOffset(lvaShadowSPslotsVar, lvaLclSize(lvaShadowSPslotsVar), stkOffs); |
5788 | } |
5789 | #endif // !FEATURE_EH_FUNCLETS |
5790 | |
5791 | if (compGSReorderStackLayout) |
5792 | { |
5793 | assert(getNeedsGSSecurityCookie()); |
5794 | stkOffs = lvaAllocLocalAndSetVirtualOffset(lvaGSSecurityCookie, lvaLclSize(lvaGSSecurityCookie), stkOffs); |
5795 | } |
5796 | |
5797 | /* |
5798 | If we're supposed to track lifetimes of pointer temps, we'll |
5799 | assign frame offsets in the following order: |
5800 | |
5801 | non-pointer local variables (also untracked pointer variables) |
5802 | pointer local variables |
5803 | pointer temps |
5804 | non-pointer temps |
5805 | */ |
5806 | |
5807 | enum Allocation |
5808 | { |
5809 | ALLOC_NON_PTRS = 0x1, // assign offsets to non-ptr |
5810 | ALLOC_PTRS = 0x2, // Second pass, assign offsets to tracked ptrs |
5811 | ALLOC_UNSAFE_BUFFERS = 0x4, |
5812 | ALLOC_UNSAFE_BUFFERS_WITH_PTRS = 0x8 |
5813 | }; |
5814 | UINT alloc_order[5]; |
5815 | |
5816 | unsigned int cur = 0; |
5817 | |
5818 | if (compGSReorderStackLayout) |
5819 | { |
5820 | noway_assert(getNeedsGSSecurityCookie()); |
5821 | |
5822 | if (codeGen->isFramePointerUsed()) |
5823 | { |
5824 | alloc_order[cur++] = ALLOC_UNSAFE_BUFFERS; |
5825 | alloc_order[cur++] = ALLOC_UNSAFE_BUFFERS_WITH_PTRS; |
5826 | } |
5827 | } |
5828 | |
5829 | bool tempsAllocated = false; |
5830 | |
5831 | if (lvaTempsHaveLargerOffsetThanVars() && !codeGen->isFramePointerUsed()) |
5832 | { |
5833 | // Because we want the temps to have a larger offset than locals |
5834 | // and we're not using a frame pointer, we have to place the temps |
5835 | // above the vars. Otherwise we place them after the vars (at the |
5836 | // bottom of the frame). |
5837 | noway_assert(!tempsAllocated); |
5838 | stkOffs = lvaAllocateTemps(stkOffs, mustDoubleAlign); |
5839 | tempsAllocated = true; |
5840 | } |
5841 | |
5842 | alloc_order[cur++] = ALLOC_NON_PTRS; |
5843 | |
5844 | if (opts.compDbgEnC) |
5845 | { |
5846 | /* We will use just one pass, and assign offsets to all variables */ |
5847 | alloc_order[cur - 1] |= ALLOC_PTRS; |
5848 | noway_assert(compGSReorderStackLayout == false); |
5849 | } |
5850 | else |
5851 | { |
5852 | alloc_order[cur++] = ALLOC_PTRS; |
5853 | } |
5854 | |
5855 | if (!codeGen->isFramePointerUsed() && compGSReorderStackLayout) |
5856 | { |
5857 | alloc_order[cur++] = ALLOC_UNSAFE_BUFFERS_WITH_PTRS; |
5858 | alloc_order[cur++] = ALLOC_UNSAFE_BUFFERS; |
5859 | } |
5860 | |
5861 | alloc_order[cur] = 0; |
5862 | |
5863 | noway_assert(cur < _countof(alloc_order)); |
5864 | |
5865 | // Force first pass to happen |
5866 | UINT assignMore = 0xFFFFFFFF; |
5867 | bool have_LclVarDoubleAlign = false; |
5868 | |
5869 | for (cur = 0; alloc_order[cur]; cur++) |
5870 | { |
5871 | if ((assignMore & alloc_order[cur]) == 0) |
5872 | { |
5873 | continue; |
5874 | } |
5875 | |
5876 | assignMore = 0; |
5877 | |
5878 | unsigned lclNum; |
5879 | LclVarDsc* varDsc; |
5880 | |
5881 | for (lclNum = 0, varDsc = lvaTable; lclNum < lvaCount; lclNum++, varDsc++) |
5882 | { |
5883 | /* Ignore field locals of the promotion type PROMOTION_TYPE_FIELD_DEPENDENT. |
5884 | In other words, we will not calculate the "base" address of the struct local if |
5885 | the promotion type is PROMOTION_TYPE_FIELD_DEPENDENT. |
5886 | */ |
5887 | if (lvaIsFieldOfDependentlyPromotedStruct(varDsc)) |
5888 | { |
5889 | continue; |
5890 | } |
5891 | |
5892 | #if FEATURE_FIXED_OUT_ARGS |
5893 | // The scratch mem is used for the outgoing arguments, and it must be absolutely last |
5894 | if (lclNum == lvaOutgoingArgSpaceVar) |
5895 | { |
5896 | continue; |
5897 | } |
5898 | #endif |
5899 | |
5900 | bool allocateOnFrame = varDsc->lvOnFrame; |
5901 | |
5902 | if (varDsc->lvRegister && (lvaDoneFrameLayout == REGALLOC_FRAME_LAYOUT) && |
5903 | ((varDsc->TypeGet() != TYP_LONG) || (varDsc->lvOtherReg != REG_STK))) |
5904 | { |
5905 | allocateOnFrame = false; |
5906 | } |
5907 | |
5908 | /* Ignore variables that are not on the stack frame */ |
5909 | |
5910 | if (!allocateOnFrame) |
5911 | { |
5912 | /* For EnC, all variables have to be allocated space on the |
5913 | stack, even though they may actually be enregistered. This |
5914 | way, the frame layout can be directly inferred from the |
5915 | locals-sig. |
5916 | */ |
5917 | |
5918 | if (!opts.compDbgEnC) |
5919 | { |
5920 | continue; |
5921 | } |
5922 | else if (lclNum >= info.compLocalsCount) |
5923 | { // ignore temps for EnC |
5924 | continue; |
5925 | } |
5926 | } |
5927 | else if (lvaGSSecurityCookie == lclNum && getNeedsGSSecurityCookie()) |
5928 | { |
5929 | continue; // This is allocated outside of this loop. |
5930 | } |
5931 | |
5932 | // These need to be located as the very first variables (highest memory address) |
5933 | // and so they have already been assigned an offset |
5934 | if ( |
5935 | #if FEATURE_EH_FUNCLETS |
5936 | lclNum == lvaPSPSym || |
5937 | #else |
5938 | lclNum == lvaShadowSPslotsVar || |
5939 | #endif // FEATURE_EH_FUNCLETS |
5940 | #ifdef JIT32_GCENCODER |
5941 | lclNum == lvaLocAllocSPvar || |
5942 | #endif // JIT32_GCENCODER |
5943 | lclNum == lvaSecurityObject) |
5944 | { |
5945 | assert(varDsc->lvStkOffs != BAD_STK_OFFS); |
5946 | continue; |
5947 | } |
5948 | |
5949 | if (lclNum == lvaMonAcquired) |
5950 | { |
5951 | continue; |
5952 | } |
5953 | |
5954 | // This should be low on the stack. Hence, it will be assigned later. |
5955 | if (lclNum == lvaStubArgumentVar) |
5956 | { |
5957 | #ifdef JIT32_GCENCODER |
5958 | noway_assert(codeGen->isFramePointerUsed()); |
5959 | #endif |
5960 | continue; |
5961 | } |
5962 | |
5963 | // This should be low on the stack. Hence, it will be assigned later. |
5964 | if (lclNum == lvaInlinedPInvokeFrameVar) |
5965 | { |
5966 | noway_assert(codeGen->isFramePointerUsed()); |
5967 | continue; |
5968 | } |
5969 | |
5970 | if (varDsc->lvIsParam) |
5971 | { |
5972 | #if defined(_TARGET_AMD64_) && !defined(UNIX_AMD64_ABI) |
5973 | |
5974 | // On Windows AMD64 we can use the caller-reserved stack area that is already setup |
5975 | assert(varDsc->lvStkOffs != BAD_STK_OFFS); |
5976 | continue; |
5977 | |
5978 | #else // !_TARGET_AMD64_ |
5979 | |
5980 | // A register argument that is not enregistered ends up as |
5981 | // a local variable which will need stack frame space. |
5982 | // |
5983 | if (!varDsc->lvIsRegArg) |
5984 | { |
5985 | continue; |
5986 | } |
5987 | |
5988 | #ifdef _TARGET_ARM64_ |
5989 | if (info.compIsVarArgs && varDsc->lvArgReg != theFixedRetBuffArgNum()) |
5990 | { |
5991 | // Stack offset to varargs (parameters) should point to home area which will be preallocated. |
5992 | varDsc->lvStkOffs = |
5993 | -initialStkOffs + genMapIntRegNumToRegArgNum(varDsc->GetArgReg()) * REGSIZE_BYTES; |
5994 | continue; |
5995 | } |
5996 | |
5997 | #endif |
5998 | |
5999 | #ifdef _TARGET_ARM_ |
6000 | // On ARM we spill the registers in codeGen->regSet.rsMaskPreSpillRegArg |
6001 | // in the prolog, thus they don't need stack frame space. |
6002 | // |
6003 | if ((codeGen->regSet.rsMaskPreSpillRegs(false) & genRegMask(varDsc->lvArgReg)) != 0) |
6004 | { |
6005 | assert(varDsc->lvStkOffs != BAD_STK_OFFS); |
6006 | continue; |
6007 | } |
6008 | #endif |
6009 | |
6010 | #endif // !_TARGET_AMD64_ |
6011 | } |
6012 | |
6013 | /* Make sure the type is appropriate */ |
6014 | |
6015 | if (varDsc->lvIsUnsafeBuffer && compGSReorderStackLayout) |
6016 | { |
6017 | if (varDsc->lvIsPtr) |
6018 | { |
6019 | if ((alloc_order[cur] & ALLOC_UNSAFE_BUFFERS_WITH_PTRS) == 0) |
6020 | { |
6021 | assignMore |= ALLOC_UNSAFE_BUFFERS_WITH_PTRS; |
6022 | continue; |
6023 | } |
6024 | } |
6025 | else |
6026 | { |
6027 | if ((alloc_order[cur] & ALLOC_UNSAFE_BUFFERS) == 0) |
6028 | { |
6029 | assignMore |= ALLOC_UNSAFE_BUFFERS; |
6030 | continue; |
6031 | } |
6032 | } |
6033 | } |
6034 | else if (varTypeIsGC(varDsc->TypeGet()) && varDsc->lvTracked) |
6035 | { |
6036 | if ((alloc_order[cur] & ALLOC_PTRS) == 0) |
6037 | { |
6038 | assignMore |= ALLOC_PTRS; |
6039 | continue; |
6040 | } |
6041 | } |
6042 | else |
6043 | { |
6044 | if ((alloc_order[cur] & ALLOC_NON_PTRS) == 0) |
6045 | { |
6046 | assignMore |= ALLOC_NON_PTRS; |
6047 | continue; |
6048 | } |
6049 | } |
6050 | |
6051 | /* Need to align the offset? */ |
6052 | |
6053 | if (mustDoubleAlign && (varDsc->lvType == TYP_DOUBLE // Align doubles for ARM and x86 |
6054 | #ifdef _TARGET_ARM_ |
6055 | || varDsc->lvType == TYP_LONG // Align longs for ARM |
6056 | #endif |
6057 | #ifndef _TARGET_64BIT_ |
6058 | || varDsc->lvStructDoubleAlign // Align when lvStructDoubleAlign is true |
6059 | #endif // !_TARGET_64BIT_ |
6060 | )) |
6061 | { |
6062 | noway_assert((compLclFrameSize % TARGET_POINTER_SIZE) == 0); |
6063 | |
6064 | if ((lvaDoneFrameLayout != FINAL_FRAME_LAYOUT) && !have_LclVarDoubleAlign) |
6065 | { |
6066 | // If this is the first TYP_LONG, TYP_DOUBLE or double aligned struct |
6067 | // then we have seen in this loop then we allocate a pointer sized |
6068 | // stack slot since we may need to double align this LclVar |
6069 | // when lvaDoneFrameLayout == FINAL_FRAME_LAYOUT |
6070 | // |
6071 | lvaIncrementFrameSize(TARGET_POINTER_SIZE); |
6072 | stkOffs -= TARGET_POINTER_SIZE; |
6073 | } |
6074 | else |
6075 | { |
6076 | if (((stkOffs + preSpillSize) % (2 * TARGET_POINTER_SIZE)) != 0) |
6077 | { |
6078 | lvaIncrementFrameSize(TARGET_POINTER_SIZE); |
6079 | stkOffs -= TARGET_POINTER_SIZE; |
6080 | } |
6081 | |
6082 | // We should now have a double-aligned (stkOffs+preSpillSize) |
6083 | noway_assert(((stkOffs + preSpillSize) % (2 * TARGET_POINTER_SIZE)) == 0); |
6084 | } |
6085 | |
6086 | // Remember that we had to double align a LclVar |
6087 | have_LclVarDoubleAlign = true; |
6088 | } |
6089 | |
6090 | // Reserve the stack space for this variable |
6091 | stkOffs = lvaAllocLocalAndSetVirtualOffset(lclNum, lvaLclSize(lclNum), stkOffs); |
6092 | #ifdef _TARGET_ARM64_ |
6093 | // If we have an incoming register argument that has a struct promoted field |
6094 | // then we need to copy the lvStkOff (the stack home) from the reg arg to the field lclvar |
6095 | // |
6096 | if (varDsc->lvIsRegArg && varDsc->lvPromotedStruct()) |
6097 | { |
6098 | noway_assert(varDsc->lvFieldCnt == 1); // We only handle one field here |
6099 | |
6100 | unsigned fieldVarNum = varDsc->lvFieldLclStart; |
6101 | lvaTable[fieldVarNum].lvStkOffs = varDsc->lvStkOffs; |
6102 | } |
6103 | #endif // _TARGET_ARM64_ |
6104 | #ifdef _TARGET_ARM_ |
6105 | // If we have an incoming register argument that has a promoted long |
6106 | // then we need to copy the lvStkOff (the stack home) from the reg arg to the field lclvar |
6107 | // |
6108 | if (varDsc->lvIsRegArg && varDsc->lvPromoted) |
6109 | { |
6110 | assert(varTypeIsLong(varDsc) && (varDsc->lvFieldCnt == 2)); |
6111 | |
6112 | unsigned fieldVarNum = varDsc->lvFieldLclStart; |
6113 | lvaTable[fieldVarNum].lvStkOffs = varDsc->lvStkOffs; |
6114 | lvaTable[fieldVarNum + 1].lvStkOffs = varDsc->lvStkOffs + 4; |
6115 | } |
6116 | #endif // _TARGET_ARM_ |
6117 | } |
6118 | } |
6119 | |
6120 | if (getNeedsGSSecurityCookie() && !compGSReorderStackLayout) |
6121 | { |
6122 | // LOCALLOC used, but we have no unsafe buffer. Allocated cookie last, close to localloc buffer. |
6123 | stkOffs = lvaAllocLocalAndSetVirtualOffset(lvaGSSecurityCookie, lvaLclSize(lvaGSSecurityCookie), stkOffs); |
6124 | } |
6125 | |
6126 | if (tempsAllocated == false) |
6127 | { |
6128 | /*------------------------------------------------------------------------- |
6129 | * |
6130 | * Now the temps |
6131 | * |
6132 | *------------------------------------------------------------------------- |
6133 | */ |
6134 | stkOffs = lvaAllocateTemps(stkOffs, mustDoubleAlign); |
6135 | } |
6136 | |
6137 | /*------------------------------------------------------------------------- |
6138 | * |
6139 | * Now do some final stuff |
6140 | * |
6141 | *------------------------------------------------------------------------- |
6142 | */ |
6143 | |
6144 | // lvaInlinedPInvokeFrameVar and lvaStubArgumentVar need to be assigned last |
6145 | // Important: The stack walker depends on lvaStubArgumentVar immediately |
6146 | // following lvaInlinedPInvokeFrameVar in the frame. |
6147 | |
6148 | if (lvaStubArgumentVar != BAD_VAR_NUM) |
6149 | { |
6150 | #ifdef JIT32_GCENCODER |
6151 | noway_assert(codeGen->isFramePointerUsed()); |
6152 | #endif |
6153 | stkOffs = lvaAllocLocalAndSetVirtualOffset(lvaStubArgumentVar, lvaLclSize(lvaStubArgumentVar), stkOffs); |
6154 | } |
6155 | |
6156 | if (lvaInlinedPInvokeFrameVar != BAD_VAR_NUM) |
6157 | { |
6158 | noway_assert(codeGen->isFramePointerUsed()); |
6159 | stkOffs = |
6160 | lvaAllocLocalAndSetVirtualOffset(lvaInlinedPInvokeFrameVar, lvaLclSize(lvaInlinedPInvokeFrameVar), stkOffs); |
6161 | } |
6162 | |
6163 | if (mustDoubleAlign) |
6164 | { |
6165 | if (lvaDoneFrameLayout != FINAL_FRAME_LAYOUT) |
6166 | { |
6167 | // Allocate a pointer sized stack slot, since we may need to double align here |
6168 | // when lvaDoneFrameLayout == FINAL_FRAME_LAYOUT |
6169 | // |
6170 | lvaIncrementFrameSize(TARGET_POINTER_SIZE); |
6171 | stkOffs -= TARGET_POINTER_SIZE; |
6172 | |
6173 | if (have_LclVarDoubleAlign) |
6174 | { |
6175 | // If we have any TYP_LONG, TYP_DOUBLE or double aligned structs |
6176 | // the we need to allocate a second pointer sized stack slot, |
6177 | // since we may need to double align the last LclVar that we saw |
6178 | // in the loop above. We do this so that the offsets that we |
6179 | // calculate for the stack frame are always greater than they will |
6180 | // be in the final layout. |
6181 | // |
6182 | lvaIncrementFrameSize(TARGET_POINTER_SIZE); |
6183 | stkOffs -= TARGET_POINTER_SIZE; |
6184 | } |
6185 | } |
6186 | else // FINAL_FRAME_LAYOUT |
6187 | { |
6188 | if (((stkOffs + preSpillSize) % (2 * TARGET_POINTER_SIZE)) != 0) |
6189 | { |
6190 | lvaIncrementFrameSize(TARGET_POINTER_SIZE); |
6191 | stkOffs -= TARGET_POINTER_SIZE; |
6192 | } |
6193 | // We should now have a double-aligned (stkOffs+preSpillSize) |
6194 | noway_assert(((stkOffs + preSpillSize) % (2 * TARGET_POINTER_SIZE)) == 0); |
6195 | } |
6196 | } |
6197 | |
6198 | #if FEATURE_EH_FUNCLETS && defined(_TARGET_AMD64_) |
6199 | if (lvaPSPSym != BAD_VAR_NUM) |
6200 | { |
6201 | // On AMD64, if we need a PSPSym, allocate it last, immediately above the outgoing argument |
6202 | // space. Any padding will be higher on the stack than this |
6203 | // (including the padding added by lvaAlignFrame()). |
6204 | noway_assert(codeGen->isFramePointerUsed()); // We need an explicit frame pointer |
6205 | stkOffs = lvaAllocLocalAndSetVirtualOffset(lvaPSPSym, TARGET_POINTER_SIZE, stkOffs); |
6206 | } |
6207 | #endif // FEATURE_EH_FUNCLETS && defined(_TARGET_AMD64_) |
6208 | |
6209 | #ifdef _TARGET_ARM64_ |
6210 | if (isFramePointerUsed()) |
6211 | { |
6212 | // Create space for saving FP and LR. |
6213 | stkOffs -= 2 * REGSIZE_BYTES; |
6214 | } |
6215 | #endif // _TARGET_ARM64_ |
6216 | |
6217 | #if FEATURE_FIXED_OUT_ARGS |
6218 | if (lvaOutgoingArgSpaceSize > 0) |
6219 | { |
6220 | #if defined(_TARGET_AMD64_) && !defined(UNIX_AMD64_ABI) // No 4 slots for outgoing params on System V. |
6221 | noway_assert(lvaOutgoingArgSpaceSize >= (4 * TARGET_POINTER_SIZE)); |
6222 | #endif |
6223 | noway_assert((lvaOutgoingArgSpaceSize % TARGET_POINTER_SIZE) == 0); |
6224 | |
6225 | // Give it a value so we can avoid asserts in CHK builds. |
6226 | // Since this will always use an SP relative offset of zero |
6227 | // at the end of lvaFixVirtualFrameOffsets, it will be set to absolute '0' |
6228 | |
6229 | stkOffs = lvaAllocLocalAndSetVirtualOffset(lvaOutgoingArgSpaceVar, lvaLclSize(lvaOutgoingArgSpaceVar), stkOffs); |
6230 | } |
6231 | #endif // FEATURE_FIXED_OUT_ARGS |
6232 | |
6233 | // compLclFrameSize equals our negated virtual stack offset minus the pushed registers and return address |
6234 | // and the pushed frame pointer register which for some strange reason isn't part of 'compCalleeRegsPushed'. |
6235 | int pushedCount = compCalleeRegsPushed; |
6236 | |
6237 | #ifdef _TARGET_ARM64_ |
6238 | if (info.compIsVarArgs) |
6239 | { |
6240 | pushedCount += MAX_REG_ARG; |
6241 | } |
6242 | #endif |
6243 | |
6244 | #ifdef _TARGET_XARCH_ |
6245 | if (codeGen->doubleAlignOrFramePointerUsed()) |
6246 | { |
6247 | pushedCount += 1; // pushed EBP (frame pointer) |
6248 | } |
6249 | pushedCount += 1; // pushed PC (return address) |
6250 | #endif |
6251 | |
6252 | noway_assert(compLclFrameSize == (unsigned)-(stkOffs + (pushedCount * (int)TARGET_POINTER_SIZE))); |
6253 | } |
6254 | |
6255 | int Compiler::lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs) |
6256 | { |
6257 | noway_assert(lclNum != BAD_VAR_NUM); |
6258 | |
6259 | #ifdef _TARGET_64BIT_ |
6260 | // Before final frame layout, assume the worst case, that every >=8 byte local will need |
6261 | // maximum padding to be aligned. This is because we generate code based on the stack offset |
6262 | // computed during tentative frame layout. These offsets cannot get bigger during final |
6263 | // frame layout, as that would possibly require different code generation (for example, |
6264 | // using a 4-byte offset instead of a 1-byte offset in an instruction). The offsets can get |
6265 | // smaller. It is possible there is different alignment at the point locals are allocated |
6266 | // between tentative and final frame layout which would introduce padding between locals |
6267 | // and thus increase the offset (from the stack pointer) of one of the locals. Hence the |
6268 | // need to assume the worst alignment before final frame layout. |
6269 | // We could probably improve this by sorting all the objects by alignment, |
6270 | // such that all 8 byte objects are together, 4 byte objects are together, etc., which |
6271 | // would require at most one alignment padding per group. |
6272 | // |
6273 | // TYP_SIMD structs locals have alignment preference given by getSIMDTypeAlignment() for |
6274 | // better performance. |
6275 | if ((size >= 8) && ((lvaDoneFrameLayout != FINAL_FRAME_LAYOUT) || ((stkOffs % 8) != 0) |
6276 | #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES |
6277 | || lclVarIsSIMDType(lclNum) |
6278 | #endif |
6279 | )) |
6280 | { |
6281 | // Note that stack offsets are negative or equal to zero |
6282 | assert(stkOffs <= 0); |
6283 | |
6284 | // alignment padding |
6285 | unsigned pad = 0; |
6286 | #if defined(FEATURE_SIMD) && ALIGN_SIMD_TYPES |
6287 | if (lclVarIsSIMDType(lclNum) && !lvaIsImplicitByRefLocal(lclNum)) |
6288 | { |
6289 | int alignment = getSIMDTypeAlignment(lvaTable[lclNum].lvType); |
6290 | |
6291 | if (stkOffs % alignment != 0) |
6292 | { |
6293 | if (lvaDoneFrameLayout != FINAL_FRAME_LAYOUT) |
6294 | { |
6295 | pad = alignment - 1; |
6296 | // Note that all the objects will probably be misaligned, but we'll fix that in final layout. |
6297 | } |
6298 | else |
6299 | { |
6300 | pad = alignment + (stkOffs % alignment); // +1 to +(alignment-1) bytes |
6301 | } |
6302 | } |
6303 | } |
6304 | else |
6305 | #endif // FEATURE_SIMD && ALIGN_SIMD_TYPES |
6306 | { |
6307 | if (lvaDoneFrameLayout != FINAL_FRAME_LAYOUT) |
6308 | { |
6309 | pad = 7; |
6310 | // Note that all the objects will probably be misaligned, but we'll fix that in final layout. |
6311 | } |
6312 | else |
6313 | { |
6314 | pad = 8 + (stkOffs % 8); // +1 to +7 bytes |
6315 | } |
6316 | } |
6317 | // Will the pad ever be anything except 4? Do we put smaller-than-4-sized objects on the stack? |
6318 | lvaIncrementFrameSize(pad); |
6319 | stkOffs -= pad; |
6320 | |
6321 | #ifdef DEBUG |
6322 | if (verbose) |
6323 | { |
6324 | printf("Pad " ); |
6325 | gtDispLclVar(lclNum, /*pad*/ false); |
6326 | printf(", size=%d, stkOffs=%c0x%x, pad=%d\n" , size, stkOffs < 0 ? '-' : '+', |
6327 | stkOffs < 0 ? -stkOffs : stkOffs, pad); |
6328 | } |
6329 | #endif |
6330 | } |
6331 | #endif // _TARGET_64BIT_ |
6332 | |
6333 | /* Reserve space on the stack by bumping the frame size */ |
6334 | |
6335 | lvaIncrementFrameSize(size); |
6336 | stkOffs -= size; |
6337 | lvaTable[lclNum].lvStkOffs = stkOffs; |
6338 | |
6339 | #ifdef DEBUG |
6340 | if (verbose) |
6341 | { |
6342 | printf("Assign " ); |
6343 | gtDispLclVar(lclNum, /*pad*/ false); |
6344 | printf(", size=%d, stkOffs=%c0x%x\n" , size, stkOffs < 0 ? '-' : '+', stkOffs < 0 ? -stkOffs : stkOffs); |
6345 | } |
6346 | #endif |
6347 | |
6348 | return stkOffs; |
6349 | } |
6350 | |
6351 | #ifdef _TARGET_AMD64_ |
6352 | /***************************************************************************** |
6353 | * lvaIsCalleeSavedIntRegCountEven() : returns true if the number of integer registers |
6354 | * pushed onto stack is even including RBP if used as frame pointer |
6355 | * |
6356 | * Note that this excludes return address (PC) pushed by caller. To know whether |
6357 | * the SP offset after pushing integer registers is aligned, we need to take |
6358 | * negation of this routine. |
6359 | */ |
6360 | bool Compiler::lvaIsCalleeSavedIntRegCountEven() |
6361 | { |
6362 | unsigned regsPushed = compCalleeRegsPushed + (codeGen->isFramePointerUsed() ? 1 : 0); |
6363 | return (regsPushed % (16 / REGSIZE_BYTES)) == 0; |
6364 | } |
6365 | #endif //_TARGET_AMD64_ |
6366 | |
6367 | /***************************************************************************** |
6368 | * lvaAlignFrame() : After allocating everything on the frame, reserve any |
6369 | * extra space needed to keep the frame aligned |
6370 | */ |
6371 | void Compiler::lvaAlignFrame() |
6372 | { |
6373 | #if defined(_TARGET_AMD64_) |
6374 | |
6375 | // Leaf frames do not need full alignment, but the unwind info is smaller if we |
6376 | // are at least 8 byte aligned (and we assert as much) |
6377 | if ((compLclFrameSize % 8) != 0) |
6378 | { |
6379 | lvaIncrementFrameSize(8 - (compLclFrameSize % 8)); |
6380 | } |
6381 | else if (lvaDoneFrameLayout != FINAL_FRAME_LAYOUT) |
6382 | { |
6383 | // If we are not doing final layout, we don't know the exact value of compLclFrameSize |
6384 | // and thus do not know how much we will need to add in order to be aligned. |
6385 | // We add 8 so compLclFrameSize is still a multiple of 8. |
6386 | lvaIncrementFrameSize(8); |
6387 | } |
6388 | assert((compLclFrameSize % 8) == 0); |
6389 | |
6390 | // Ensure that the stack is always 16-byte aligned by grabbing an unused QWORD |
6391 | // if needed, but off by 8 because of the return value. |
6392 | // And don't forget that compCalleeRegsPused does *not* include RBP if we are |
6393 | // using it as the frame pointer. |
6394 | // |
6395 | bool regPushedCountAligned = lvaIsCalleeSavedIntRegCountEven(); |
6396 | bool lclFrameSizeAligned = (compLclFrameSize % 16) == 0; |
6397 | |
6398 | // If this isn't the final frame layout, assume we have to push an extra QWORD |
6399 | // Just so the offsets are true upper limits. |
6400 | CLANG_FORMAT_COMMENT_ANCHOR; |
6401 | |
6402 | #ifdef UNIX_AMD64_ABI |
6403 | // The compNeedToAlignFrame flag is indicating if there is a need to align the frame. |
6404 | // On AMD64-Windows, if there are calls, 4 slots for the outgoing ars are allocated, except for |
6405 | // FastTailCall. This slots makes the frame size non-zero, so alignment logic will be called. |
6406 | // On AMD64-Unix, there are no such slots. There is a possibility to have calls in the method with frame size of 0. |
6407 | // The frame alignment logic won't kick in. This flags takes care of the AMD64-Unix case by remembering that there |
6408 | // are calls and making sure the frame alignment logic is executed. |
6409 | bool stackNeedsAlignment = (compLclFrameSize != 0 || opts.compNeedToAlignFrame); |
6410 | #else // !UNIX_AMD64_ABI |
6411 | bool stackNeedsAlignment = compLclFrameSize != 0; |
6412 | #endif // !UNIX_AMD64_ABI |
6413 | if ((!codeGen->isFramePointerUsed() && (lvaDoneFrameLayout != FINAL_FRAME_LAYOUT)) || |
6414 | (stackNeedsAlignment && (regPushedCountAligned == lclFrameSizeAligned))) |
6415 | { |
6416 | lvaIncrementFrameSize(REGSIZE_BYTES); |
6417 | } |
6418 | |
6419 | #elif defined(_TARGET_ARM64_) |
6420 | |
6421 | // The stack on ARM64 must be 16 byte aligned. |
6422 | |
6423 | // First, align up to 8. |
6424 | if ((compLclFrameSize % 8) != 0) |
6425 | { |
6426 | lvaIncrementFrameSize(8 - (compLclFrameSize % 8)); |
6427 | } |
6428 | else if (lvaDoneFrameLayout != FINAL_FRAME_LAYOUT) |
6429 | { |
6430 | // If we are not doing final layout, we don't know the exact value of compLclFrameSize |
6431 | // and thus do not know how much we will need to add in order to be aligned. |
6432 | // We add 8 so compLclFrameSize is still a multiple of 8. |
6433 | lvaIncrementFrameSize(8); |
6434 | } |
6435 | assert((compLclFrameSize % 8) == 0); |
6436 | |
6437 | // Ensure that the stack is always 16-byte aligned by grabbing an unused QWORD |
6438 | // if needed. |
6439 | bool regPushedCountAligned = (compCalleeRegsPushed % (16 / REGSIZE_BYTES)) == 0; |
6440 | bool lclFrameSizeAligned = (compLclFrameSize % 16) == 0; |
6441 | |
6442 | // If this isn't the final frame layout, assume we have to push an extra QWORD |
6443 | // Just so the offsets are true upper limits. |
6444 | if ((lvaDoneFrameLayout != FINAL_FRAME_LAYOUT) || (regPushedCountAligned != lclFrameSizeAligned)) |
6445 | { |
6446 | lvaIncrementFrameSize(REGSIZE_BYTES); |
6447 | } |
6448 | |
6449 | #elif defined(_TARGET_ARM_) |
6450 | |
6451 | // Ensure that stack offsets will be double-aligned by grabbing an unused DWORD if needed. |
6452 | // |
6453 | bool lclFrameSizeAligned = (compLclFrameSize % sizeof(double)) == 0; |
6454 | bool regPushedCountAligned = ((compCalleeRegsPushed + genCountBits(codeGen->regSet.rsMaskPreSpillRegs(true))) % |
6455 | (sizeof(double) / TARGET_POINTER_SIZE)) == 0; |
6456 | |
6457 | if (regPushedCountAligned != lclFrameSizeAligned) |
6458 | { |
6459 | lvaIncrementFrameSize(TARGET_POINTER_SIZE); |
6460 | } |
6461 | |
6462 | #elif defined(_TARGET_X86_) |
6463 | |
6464 | #if DOUBLE_ALIGN |
6465 | if (genDoubleAlign()) |
6466 | { |
6467 | // Double Frame Alignement for x86 is handled in Compiler::lvaAssignVirtualFrameOffsetsToLocals() |
6468 | |
6469 | if (compLclFrameSize == 0) |
6470 | { |
6471 | // This can only happen with JitStress=1 or JitDoubleAlign=2 |
6472 | lvaIncrementFrameSize(TARGET_POINTER_SIZE); |
6473 | } |
6474 | } |
6475 | #endif |
6476 | |
6477 | if (STACK_ALIGN > REGSIZE_BYTES) |
6478 | { |
6479 | if (lvaDoneFrameLayout != FINAL_FRAME_LAYOUT) |
6480 | { |
6481 | // If we are not doing final layout, we don't know the exact value of compLclFrameSize |
6482 | // and thus do not know how much we will need to add in order to be aligned. |
6483 | // We add the maximum pad that we could ever have (which is 12) |
6484 | lvaIncrementFrameSize(STACK_ALIGN - REGSIZE_BYTES); |
6485 | } |
6486 | |
6487 | // Align the stack with STACK_ALIGN value. |
6488 | int adjustFrameSize = compLclFrameSize; |
6489 | #if defined(UNIX_X86_ABI) |
6490 | bool isEbpPushed = codeGen->isFramePointerUsed(); |
6491 | #if DOUBLE_ALIGN |
6492 | isEbpPushed |= genDoubleAlign(); |
6493 | #endif |
6494 | // we need to consider spilled register(s) plus return address and/or EBP |
6495 | int adjustCount = compCalleeRegsPushed + 1 + (isEbpPushed ? 1 : 0); |
6496 | adjustFrameSize += (adjustCount * REGSIZE_BYTES) % STACK_ALIGN; |
6497 | #endif |
6498 | if ((adjustFrameSize % STACK_ALIGN) != 0) |
6499 | { |
6500 | lvaIncrementFrameSize(STACK_ALIGN - (adjustFrameSize % STACK_ALIGN)); |
6501 | } |
6502 | } |
6503 | |
6504 | #else |
6505 | NYI("TARGET specific lvaAlignFrame" ); |
6506 | #endif // !_TARGET_AMD64_ |
6507 | } |
6508 | |
6509 | /***************************************************************************** |
6510 | * lvaAssignFrameOffsetsToPromotedStructs() : Assign offsets to fields |
6511 | * within a promoted struct (worker for lvaAssignFrameOffsets). |
6512 | */ |
6513 | void Compiler::lvaAssignFrameOffsetsToPromotedStructs() |
6514 | { |
6515 | LclVarDsc* varDsc = lvaTable; |
6516 | for (unsigned lclNum = 0; lclNum < lvaCount; lclNum++, varDsc++) |
6517 | { |
6518 | // For promoted struct fields that are params, we will |
6519 | // assign their offsets in lvaAssignVirtualFrameOffsetToArg(). |
6520 | // This is not true for the System V systems since there is no |
6521 | // outgoing args space. Assign the dependently promoted fields properly. |
6522 | // |
6523 | if (varDsc->lvIsStructField |
6524 | #ifndef UNIX_AMD64_ABI |
6525 | #if !defined(_TARGET_ARM_) |
6526 | // ARM: lo/hi parts of a promoted long arg need to be updated. |
6527 | |
6528 | // For System V platforms there is no outgoing args space. |
6529 | // A register passed struct arg is homed on the stack in a separate local var. |
6530 | // The offset of these structs is already calculated in lvaAssignVirtualFrameOffsetToArg methos. |
6531 | // Make sure the code below is not executed for these structs and the offset is not changed. |
6532 | && !varDsc->lvIsParam |
6533 | #endif // !defined(_TARGET_ARM_) |
6534 | #endif // !UNIX_AMD64_ABI |
6535 | ) |
6536 | { |
6537 | LclVarDsc* parentvarDsc = &lvaTable[varDsc->lvParentLcl]; |
6538 | lvaPromotionType promotionType = lvaGetPromotionType(parentvarDsc); |
6539 | |
6540 | if (promotionType == PROMOTION_TYPE_INDEPENDENT) |
6541 | { |
6542 | // The stack offset for these field locals must have been calculated |
6543 | // by the normal frame offset assignment. |
6544 | continue; |
6545 | } |
6546 | else |
6547 | { |
6548 | noway_assert(promotionType == PROMOTION_TYPE_DEPENDENT); |
6549 | noway_assert(varDsc->lvOnFrame); |
6550 | if (parentvarDsc->lvOnFrame) |
6551 | { |
6552 | varDsc->lvStkOffs = parentvarDsc->lvStkOffs + varDsc->lvFldOffset; |
6553 | } |
6554 | else |
6555 | { |
6556 | varDsc->lvOnFrame = false; |
6557 | noway_assert(varDsc->lvRefCnt() == 0); |
6558 | } |
6559 | } |
6560 | } |
6561 | } |
6562 | } |
6563 | |
6564 | /***************************************************************************** |
6565 | * lvaAllocateTemps() : Assign virtual offsets to temps (always negative). |
6566 | */ |
6567 | int Compiler::lvaAllocateTemps(int stkOffs, bool mustDoubleAlign) |
6568 | { |
6569 | unsigned spillTempSize = 0; |
6570 | |
6571 | if (lvaDoneFrameLayout == FINAL_FRAME_LAYOUT) |
6572 | { |
6573 | int preSpillSize = 0; |
6574 | #ifdef _TARGET_ARM_ |
6575 | preSpillSize = genCountBits(codeGen->regSet.rsMaskPreSpillRegs(true)) * TARGET_POINTER_SIZE; |
6576 | #endif |
6577 | bool assignDone; |
6578 | bool assignNptr; |
6579 | bool assignPtrs = true; |
6580 | |
6581 | /* Allocate temps */ |
6582 | |
6583 | if (TRACK_GC_TEMP_LIFETIMES) |
6584 | { |
6585 | /* first pointers, then non-pointers in second pass */ |
6586 | assignNptr = false; |
6587 | assignDone = false; |
6588 | } |
6589 | else |
6590 | { |
6591 | /* Pointers and non-pointers together in single pass */ |
6592 | assignNptr = true; |
6593 | assignDone = true; |
6594 | } |
6595 | |
6596 | assert(codeGen->regSet.tmpAllFree()); |
6597 | |
6598 | AGAIN2: |
6599 | |
6600 | for (TempDsc* temp = codeGen->regSet.tmpListBeg(); temp != nullptr; temp = codeGen->regSet.tmpListNxt(temp)) |
6601 | { |
6602 | var_types tempType = temp->tdTempType(); |
6603 | unsigned size; |
6604 | |
6605 | /* Make sure the type is appropriate */ |
6606 | |
6607 | if (!assignPtrs && varTypeIsGC(tempType)) |
6608 | { |
6609 | continue; |
6610 | } |
6611 | if (!assignNptr && !varTypeIsGC(tempType)) |
6612 | { |
6613 | continue; |
6614 | } |
6615 | |
6616 | size = temp->tdTempSize(); |
6617 | |
6618 | /* Figure out and record the stack offset of the temp */ |
6619 | |
6620 | /* Need to align the offset? */ |
6621 | CLANG_FORMAT_COMMENT_ANCHOR; |
6622 | |
6623 | #ifdef _TARGET_64BIT_ |
6624 | if (varTypeIsGC(tempType) && ((stkOffs % TARGET_POINTER_SIZE) != 0)) |
6625 | { |
6626 | // Calculate 'pad' as the number of bytes to align up 'stkOffs' to be a multiple of TARGET_POINTER_SIZE |
6627 | // In practice this is really just a fancy way of writing 4. (as all stack locations are at least 4-byte |
6628 | // aligned). Note stkOffs is always negative, so (stkOffs % TARGET_POINTER_SIZE) yields a negative |
6629 | // value. |
6630 | // |
6631 | int alignPad = (int)AlignmentPad((unsigned)-stkOffs, TARGET_POINTER_SIZE); |
6632 | |
6633 | spillTempSize += alignPad; |
6634 | lvaIncrementFrameSize(alignPad); |
6635 | stkOffs -= alignPad; |
6636 | |
6637 | noway_assert((stkOffs % TARGET_POINTER_SIZE) == 0); |
6638 | } |
6639 | #endif |
6640 | |
6641 | if (mustDoubleAlign && (tempType == TYP_DOUBLE)) // Align doubles for x86 and ARM |
6642 | { |
6643 | noway_assert((compLclFrameSize % TARGET_POINTER_SIZE) == 0); |
6644 | |
6645 | if (((stkOffs + preSpillSize) % (2 * TARGET_POINTER_SIZE)) != 0) |
6646 | { |
6647 | spillTempSize += TARGET_POINTER_SIZE; |
6648 | lvaIncrementFrameSize(TARGET_POINTER_SIZE); |
6649 | stkOffs -= TARGET_POINTER_SIZE; |
6650 | } |
6651 | // We should now have a double-aligned (stkOffs+preSpillSize) |
6652 | noway_assert(((stkOffs + preSpillSize) % (2 * TARGET_POINTER_SIZE)) == 0); |
6653 | } |
6654 | |
6655 | spillTempSize += size; |
6656 | lvaIncrementFrameSize(size); |
6657 | stkOffs -= size; |
6658 | temp->tdSetTempOffs(stkOffs); |
6659 | } |
6660 | #ifdef _TARGET_ARM_ |
6661 | // Only required for the ARM platform that we have an accurate estimate for the spillTempSize |
6662 | noway_assert(spillTempSize <= lvaGetMaxSpillTempSize()); |
6663 | #endif |
6664 | |
6665 | /* If we've only assigned some temps, go back and do the rest now */ |
6666 | |
6667 | if (!assignDone) |
6668 | { |
6669 | assignNptr = !assignNptr; |
6670 | assignPtrs = !assignPtrs; |
6671 | assignDone = true; |
6672 | |
6673 | goto AGAIN2; |
6674 | } |
6675 | } |
6676 | else // We haven't run codegen, so there are no Spill temps yet! |
6677 | { |
6678 | unsigned size = lvaGetMaxSpillTempSize(); |
6679 | |
6680 | lvaIncrementFrameSize(size); |
6681 | stkOffs -= size; |
6682 | } |
6683 | |
6684 | return stkOffs; |
6685 | } |
6686 | |
6687 | #ifdef DEBUG |
6688 | |
6689 | /***************************************************************************** |
6690 | * |
6691 | * Dump the register a local is in right now. It is only the current location, since the location changes and it |
6692 | * is updated throughout code generation based on LSRA register assignments. |
6693 | */ |
6694 | |
6695 | void Compiler::lvaDumpRegLocation(unsigned lclNum) |
6696 | { |
6697 | LclVarDsc* varDsc = lvaTable + lclNum; |
6698 | |
6699 | #ifdef _TARGET_ARM_ |
6700 | if (varDsc->TypeGet() == TYP_DOUBLE) |
6701 | { |
6702 | // The assigned registers are `lvRegNum:RegNext(lvRegNum)` |
6703 | printf("%3s:%-3s " , getRegName(varDsc->lvRegNum), getRegName(REG_NEXT(varDsc->lvRegNum))); |
6704 | } |
6705 | else |
6706 | #endif // _TARGET_ARM_ |
6707 | { |
6708 | printf("%3s " , getRegName(varDsc->lvRegNum)); |
6709 | } |
6710 | } |
6711 | |
6712 | /***************************************************************************** |
6713 | * |
6714 | * Dump the frame location assigned to a local. |
6715 | * It's the home location, even though the variable doesn't always live |
6716 | * in its home location. |
6717 | */ |
6718 | |
6719 | void Compiler::lvaDumpFrameLocation(unsigned lclNum) |
6720 | { |
6721 | int offset; |
6722 | regNumber baseReg; |
6723 | |
6724 | #ifdef _TARGET_ARM_ |
6725 | offset = lvaFrameAddress(lclNum, compLocallocUsed, &baseReg, 0, /* isFloatUsage */ false); |
6726 | #else |
6727 | bool EBPbased; |
6728 | offset = lvaFrameAddress(lclNum, &EBPbased); |
6729 | baseReg = EBPbased ? REG_FPBASE : REG_SPBASE; |
6730 | #endif |
6731 | |
6732 | printf("[%2s%1s0x%02X] " , getRegName(baseReg), (offset < 0 ? "-" : "+" ), (offset < 0 ? -offset : offset)); |
6733 | } |
6734 | |
6735 | /***************************************************************************** |
6736 | * |
6737 | * dump a single lvaTable entry |
6738 | */ |
6739 | |
6740 | void Compiler::lvaDumpEntry(unsigned lclNum, FrameLayoutState curState, size_t refCntWtdWidth) |
6741 | { |
6742 | LclVarDsc* varDsc = lvaTable + lclNum; |
6743 | var_types type = varDsc->TypeGet(); |
6744 | |
6745 | if (curState == INITIAL_FRAME_LAYOUT) |
6746 | { |
6747 | printf("; " ); |
6748 | gtDispLclVar(lclNum); |
6749 | |
6750 | printf(" %7s " , varTypeName(type)); |
6751 | if (genTypeSize(type) == 0) |
6752 | { |
6753 | #if FEATURE_FIXED_OUT_ARGS |
6754 | if (lclNum == lvaOutgoingArgSpaceVar) |
6755 | { |
6756 | // Since lvaOutgoingArgSpaceSize is a PhasedVar we can't read it for Dumping until |
6757 | // after we set it to something. |
6758 | if (lvaOutgoingArgSpaceSize.HasFinalValue()) |
6759 | { |
6760 | // A PhasedVar<T> can't be directly used as an arg to a variadic function |
6761 | unsigned value = lvaOutgoingArgSpaceSize; |
6762 | printf("(%2d) " , value); |
6763 | } |
6764 | else |
6765 | { |
6766 | printf("(na) " ); // The value hasn't yet been determined |
6767 | } |
6768 | } |
6769 | else |
6770 | #endif // FEATURE_FIXED_OUT_ARGS |
6771 | { |
6772 | printf("(%2d) " , lvaLclSize(lclNum)); |
6773 | } |
6774 | } |
6775 | } |
6776 | else |
6777 | { |
6778 | if (varDsc->lvRefCnt() == 0) |
6779 | { |
6780 | // Print this with a special indicator that the variable is unused. Even though the |
6781 | // variable itself is unused, it might be a struct that is promoted, so seeing it |
6782 | // can be useful when looking at the promoted struct fields. It's also weird to see |
6783 | // missing var numbers if these aren't printed. |
6784 | printf(";* " ); |
6785 | } |
6786 | #if FEATURE_FIXED_OUT_ARGS |
6787 | // Since lvaOutgoingArgSpaceSize is a PhasedVar we can't read it for Dumping until |
6788 | // after we set it to something. |
6789 | else if ((lclNum == lvaOutgoingArgSpaceVar) && lvaOutgoingArgSpaceSize.HasFinalValue() && |
6790 | (lvaOutgoingArgSpaceSize == 0)) |
6791 | { |
6792 | // Similar to above; print this anyway. |
6793 | printf(";# " ); |
6794 | } |
6795 | #endif // FEATURE_FIXED_OUT_ARGS |
6796 | else |
6797 | { |
6798 | printf("; " ); |
6799 | } |
6800 | |
6801 | gtDispLclVar(lclNum); |
6802 | |
6803 | printf("[V%02u" , lclNum); |
6804 | if (varDsc->lvTracked) |
6805 | { |
6806 | printf(",T%02u]" , varDsc->lvVarIndex); |
6807 | } |
6808 | else |
6809 | { |
6810 | printf(" ]" ); |
6811 | } |
6812 | |
6813 | printf(" (%3u,%*s)" , varDsc->lvRefCnt(), (int)refCntWtdWidth, refCntWtd2str(varDsc->lvRefCntWtd())); |
6814 | |
6815 | printf(" %7s " , varTypeName(type)); |
6816 | if (genTypeSize(type) == 0) |
6817 | { |
6818 | printf("(%2d) " , lvaLclSize(lclNum)); |
6819 | } |
6820 | else |
6821 | { |
6822 | printf(" -> " ); |
6823 | } |
6824 | |
6825 | // The register or stack location field is 11 characters wide. |
6826 | if (varDsc->lvRefCnt() == 0) |
6827 | { |
6828 | printf("zero-ref " ); |
6829 | } |
6830 | else if (varDsc->lvRegister != 0) |
6831 | { |
6832 | // It's always a register, and always in the same register. |
6833 | lvaDumpRegLocation(lclNum); |
6834 | } |
6835 | else if (varDsc->lvOnFrame == 0) |
6836 | { |
6837 | printf("registers " ); |
6838 | } |
6839 | else |
6840 | { |
6841 | // For RyuJIT backend, it might be in a register part of the time, but it will definitely have a stack home |
6842 | // location. Otherwise, it's always on the stack. |
6843 | if (lvaDoneFrameLayout != NO_FRAME_LAYOUT) |
6844 | { |
6845 | lvaDumpFrameLocation(lclNum); |
6846 | } |
6847 | } |
6848 | } |
6849 | |
6850 | if (varDsc->lvIsHfaRegArg()) |
6851 | { |
6852 | if (varDsc->lvHfaTypeIsFloat()) |
6853 | { |
6854 | printf(" (enregistered HFA: float) " ); |
6855 | } |
6856 | else |
6857 | { |
6858 | printf(" (enregistered HFA: double)" ); |
6859 | } |
6860 | } |
6861 | |
6862 | if (varDsc->lvDoNotEnregister) |
6863 | { |
6864 | printf(" do-not-enreg[" ); |
6865 | if (varDsc->lvAddrExposed) |
6866 | { |
6867 | printf("X" ); |
6868 | } |
6869 | if (varTypeIsStruct(varDsc)) |
6870 | { |
6871 | printf("S" ); |
6872 | } |
6873 | if (varDsc->lvVMNeedsStackAddr) |
6874 | { |
6875 | printf("V" ); |
6876 | } |
6877 | if (varDsc->lvLiveInOutOfHndlr) |
6878 | { |
6879 | printf("H" ); |
6880 | } |
6881 | if (varDsc->lvLclFieldExpr) |
6882 | { |
6883 | printf("F" ); |
6884 | } |
6885 | if (varDsc->lvLclBlockOpAddr) |
6886 | { |
6887 | printf("B" ); |
6888 | } |
6889 | if (varDsc->lvLiveAcrossUCall) |
6890 | { |
6891 | printf("U" ); |
6892 | } |
6893 | if (varDsc->lvIsMultiRegArg) |
6894 | { |
6895 | printf("A" ); |
6896 | } |
6897 | if (varDsc->lvIsMultiRegRet) |
6898 | { |
6899 | printf("R" ); |
6900 | } |
6901 | #ifdef JIT32_GCENCODER |
6902 | if (varDsc->lvPinned) |
6903 | printf("P" ); |
6904 | #endif // JIT32_GCENCODER |
6905 | printf("]" ); |
6906 | } |
6907 | |
6908 | if (varDsc->lvIsMultiRegArg) |
6909 | { |
6910 | printf(" multireg-arg" ); |
6911 | } |
6912 | if (varDsc->lvIsMultiRegRet) |
6913 | { |
6914 | printf(" multireg-ret" ); |
6915 | } |
6916 | if (varDsc->lvMustInit) |
6917 | { |
6918 | printf(" must-init" ); |
6919 | } |
6920 | if (varDsc->lvAddrExposed) |
6921 | { |
6922 | printf(" addr-exposed" ); |
6923 | } |
6924 | if (varDsc->lvHasLdAddrOp) |
6925 | { |
6926 | printf(" ld-addr-op" ); |
6927 | } |
6928 | if (varDsc->lvVerTypeInfo.IsThisPtr()) |
6929 | { |
6930 | printf(" this" ); |
6931 | } |
6932 | if (varDsc->lvPinned) |
6933 | { |
6934 | printf(" pinned" ); |
6935 | } |
6936 | if (varDsc->lvStackByref) |
6937 | { |
6938 | printf(" stack-byref" ); |
6939 | } |
6940 | if (varDsc->lvClassHnd != nullptr) |
6941 | { |
6942 | printf(" class-hnd" ); |
6943 | } |
6944 | if (varDsc->lvClassIsExact) |
6945 | { |
6946 | printf(" exact" ); |
6947 | } |
6948 | #ifndef _TARGET_64BIT_ |
6949 | if (varDsc->lvStructDoubleAlign) |
6950 | printf(" double-align" ); |
6951 | #endif // !_TARGET_64BIT_ |
6952 | if (varDsc->lvOverlappingFields) |
6953 | { |
6954 | printf(" overlapping-fields" ); |
6955 | } |
6956 | |
6957 | if (compGSReorderStackLayout && !varDsc->lvRegister) |
6958 | { |
6959 | if (varDsc->lvIsPtr) |
6960 | { |
6961 | printf(" ptr" ); |
6962 | } |
6963 | if (varDsc->lvIsUnsafeBuffer) |
6964 | { |
6965 | printf(" unsafe-buffer" ); |
6966 | } |
6967 | } |
6968 | if (varDsc->lvIsStructField) |
6969 | { |
6970 | LclVarDsc* parentvarDsc = &lvaTable[varDsc->lvParentLcl]; |
6971 | #if !defined(_TARGET_64BIT_) |
6972 | if (varTypeIsLong(parentvarDsc)) |
6973 | { |
6974 | bool isLo = (lclNum == parentvarDsc->lvFieldLclStart); |
6975 | printf(" V%02u.%s(offs=0x%02x)" , varDsc->lvParentLcl, isLo ? "lo" : "hi" , isLo ? 0 : genTypeSize(TYP_INT)); |
6976 | } |
6977 | else |
6978 | #endif // !defined(_TARGET_64BIT_) |
6979 | { |
6980 | CORINFO_CLASS_HANDLE typeHnd = parentvarDsc->lvVerTypeInfo.GetClassHandle(); |
6981 | CORINFO_FIELD_HANDLE fldHnd = info.compCompHnd->getFieldInClass(typeHnd, varDsc->lvFldOrdinal); |
6982 | |
6983 | printf(" V%02u.%s(offs=0x%02x)" , varDsc->lvParentLcl, eeGetFieldName(fldHnd), varDsc->lvFldOffset); |
6984 | |
6985 | lvaPromotionType promotionType = lvaGetPromotionType(parentvarDsc); |
6986 | // We should never have lvIsStructField set if it is a reg-sized non-field-addressed struct. |
6987 | assert(!varDsc->lvRegStruct); |
6988 | switch (promotionType) |
6989 | { |
6990 | case PROMOTION_TYPE_NONE: |
6991 | printf(" P-NONE" ); |
6992 | break; |
6993 | case PROMOTION_TYPE_DEPENDENT: |
6994 | printf(" P-DEP" ); |
6995 | break; |
6996 | case PROMOTION_TYPE_INDEPENDENT: |
6997 | printf(" P-INDEP" ); |
6998 | break; |
6999 | } |
7000 | } |
7001 | } |
7002 | |
7003 | if (varDsc->lvReason != nullptr) |
7004 | { |
7005 | printf(" \"%s\"" , varDsc->lvReason); |
7006 | } |
7007 | |
7008 | printf("\n" ); |
7009 | } |
7010 | |
7011 | /***************************************************************************** |
7012 | * |
7013 | * dump the lvaTable |
7014 | */ |
7015 | |
7016 | void Compiler::lvaTableDump(FrameLayoutState curState) |
7017 | { |
7018 | if (curState == NO_FRAME_LAYOUT) |
7019 | { |
7020 | curState = lvaDoneFrameLayout; |
7021 | if (curState == NO_FRAME_LAYOUT) |
7022 | { |
7023 | // Still no layout? Could be a bug, but just display the initial layout |
7024 | curState = INITIAL_FRAME_LAYOUT; |
7025 | } |
7026 | } |
7027 | |
7028 | if (curState == INITIAL_FRAME_LAYOUT) |
7029 | { |
7030 | printf("; Initial" ); |
7031 | } |
7032 | else if (curState == PRE_REGALLOC_FRAME_LAYOUT) |
7033 | { |
7034 | printf("; Pre-RegAlloc" ); |
7035 | } |
7036 | else if (curState == REGALLOC_FRAME_LAYOUT) |
7037 | { |
7038 | printf("; RegAlloc" ); |
7039 | } |
7040 | else if (curState == TENTATIVE_FRAME_LAYOUT) |
7041 | { |
7042 | printf("; Tentative" ); |
7043 | } |
7044 | else if (curState == FINAL_FRAME_LAYOUT) |
7045 | { |
7046 | printf("; Final" ); |
7047 | } |
7048 | else |
7049 | { |
7050 | printf("UNKNOWN FrameLayoutState!" ); |
7051 | unreached(); |
7052 | } |
7053 | |
7054 | printf(" local variable assignments\n" ); |
7055 | printf(";\n" ); |
7056 | |
7057 | unsigned lclNum; |
7058 | LclVarDsc* varDsc; |
7059 | |
7060 | // Figure out some sizes, to help line things up |
7061 | |
7062 | size_t refCntWtdWidth = 6; // Use 6 as the minimum width |
7063 | |
7064 | if (curState != INITIAL_FRAME_LAYOUT) // don't need this info for INITIAL_FRAME_LAYOUT |
7065 | { |
7066 | for (lclNum = 0, varDsc = lvaTable; lclNum < lvaCount; lclNum++, varDsc++) |
7067 | { |
7068 | size_t width = strlen(refCntWtd2str(varDsc->lvRefCntWtd())); |
7069 | if (width > refCntWtdWidth) |
7070 | { |
7071 | refCntWtdWidth = width; |
7072 | } |
7073 | } |
7074 | } |
7075 | |
7076 | // Do the actual output |
7077 | |
7078 | for (lclNum = 0, varDsc = lvaTable; lclNum < lvaCount; lclNum++, varDsc++) |
7079 | { |
7080 | lvaDumpEntry(lclNum, curState, refCntWtdWidth); |
7081 | } |
7082 | |
7083 | //------------------------------------------------------------------------- |
7084 | // Display the code-gen temps |
7085 | |
7086 | assert(codeGen->regSet.tmpAllFree()); |
7087 | for (TempDsc* temp = codeGen->regSet.tmpListBeg(); temp != nullptr; temp = codeGen->regSet.tmpListNxt(temp)) |
7088 | { |
7089 | printf("; TEMP_%02u %26s%*s%7s -> " , -temp->tdTempNum(), " " , refCntWtdWidth, " " , |
7090 | varTypeName(temp->tdTempType())); |
7091 | int offset = temp->tdTempOffs(); |
7092 | printf(" [%2s%1s0x%02X]\n" , isFramePointerUsed() ? STR_FPBASE : STR_SPBASE, (offset < 0 ? "-" : "+" ), |
7093 | (offset < 0 ? -offset : offset)); |
7094 | } |
7095 | |
7096 | if (curState >= TENTATIVE_FRAME_LAYOUT) |
7097 | { |
7098 | printf(";\n" ); |
7099 | printf("; Lcl frame size = %d\n" , compLclFrameSize); |
7100 | } |
7101 | } |
7102 | #endif // DEBUG |
7103 | |
7104 | /***************************************************************************** |
7105 | * |
7106 | * Conservatively estimate the layout of the stack frame. |
7107 | * |
7108 | * This function is only used before final frame layout. It conservatively estimates the |
7109 | * number of callee-saved registers that must be saved, then calls lvaAssignFrameOffsets(). |
7110 | * To do final frame layout, the callee-saved registers are known precisely, so |
7111 | * lvaAssignFrameOffsets() is called directly. |
7112 | * |
7113 | * Returns the (conservative, that is, overly large) estimated size of the frame, |
7114 | * including the callee-saved registers. This is only used by the emitter during code |
7115 | * generation when estimating the size of the offset of instructions accessing temps, |
7116 | * and only if temps have a larger offset than variables. |
7117 | */ |
7118 | |
7119 | unsigned Compiler::lvaFrameSize(FrameLayoutState curState) |
7120 | { |
7121 | assert(curState < FINAL_FRAME_LAYOUT); |
7122 | |
7123 | unsigned result; |
7124 | |
7125 | /* Layout the stack frame conservatively. |
7126 | Assume all callee-saved registers are spilled to stack */ |
7127 | |
7128 | compCalleeRegsPushed = CNT_CALLEE_SAVED; |
7129 | |
7130 | #if defined(_TARGET_ARMARCH_) |
7131 | if (compFloatingPointUsed) |
7132 | compCalleeRegsPushed += CNT_CALLEE_SAVED_FLOAT; |
7133 | |
7134 | compCalleeRegsPushed++; // we always push LR. See genPushCalleeSavedRegisters |
7135 | #elif defined(_TARGET_AMD64_) |
7136 | if (compFloatingPointUsed) |
7137 | { |
7138 | compCalleeFPRegsSavedMask = RBM_FLT_CALLEE_SAVED; |
7139 | } |
7140 | else |
7141 | { |
7142 | compCalleeFPRegsSavedMask = RBM_NONE; |
7143 | } |
7144 | #endif |
7145 | |
7146 | #if DOUBLE_ALIGN |
7147 | if (genDoubleAlign()) |
7148 | { |
7149 | // X86 only - account for extra 4-byte pad that may be created by "and esp, -8" instruction |
7150 | compCalleeRegsPushed++; |
7151 | } |
7152 | #endif |
7153 | |
7154 | #ifdef _TARGET_XARCH_ |
7155 | // Since FP/EBP is included in the SAVED_REG_MAXSZ we need to |
7156 | // subtract 1 register if codeGen->isFramePointerUsed() is true. |
7157 | if (codeGen->isFramePointerUsed()) |
7158 | { |
7159 | compCalleeRegsPushed--; |
7160 | } |
7161 | #endif |
7162 | |
7163 | lvaAssignFrameOffsets(curState); |
7164 | |
7165 | unsigned calleeSavedRegMaxSz = CALLEE_SAVED_REG_MAXSZ; |
7166 | #if defined(_TARGET_ARMARCH_) |
7167 | if (compFloatingPointUsed) |
7168 | { |
7169 | calleeSavedRegMaxSz += CALLEE_SAVED_FLOAT_MAXSZ; |
7170 | } |
7171 | calleeSavedRegMaxSz += REGSIZE_BYTES; // we always push LR. See genPushCalleeSavedRegisters |
7172 | #endif |
7173 | |
7174 | result = compLclFrameSize + calleeSavedRegMaxSz; |
7175 | return result; |
7176 | } |
7177 | |
7178 | //------------------------------------------------------------------------ |
7179 | // lvaGetSPRelativeOffset: Given a variable, return the offset of that |
7180 | // variable in the frame from the stack pointer. This number will be positive, |
7181 | // since the stack pointer must be at a lower address than everything on the |
7182 | // stack. |
7183 | // |
7184 | // This can't be called for localloc functions, since the stack pointer |
7185 | // varies, and thus there is no fixed offset to a variable from the stack pointer. |
7186 | // |
7187 | // Arguments: |
7188 | // varNum - the variable number |
7189 | // |
7190 | // Return Value: |
7191 | // The offset. |
7192 | |
7193 | int Compiler::lvaGetSPRelativeOffset(unsigned varNum) |
7194 | { |
7195 | assert(!compLocallocUsed); |
7196 | assert(lvaDoneFrameLayout == FINAL_FRAME_LAYOUT); |
7197 | assert(varNum < lvaCount); |
7198 | const LclVarDsc* varDsc = lvaTable + varNum; |
7199 | assert(varDsc->lvOnFrame); |
7200 | int spRelativeOffset; |
7201 | |
7202 | if (varDsc->lvFramePointerBased) |
7203 | { |
7204 | // The stack offset is relative to the frame pointer, so convert it to be |
7205 | // relative to the stack pointer (which makes no sense for localloc functions). |
7206 | spRelativeOffset = varDsc->lvStkOffs + codeGen->genSPtoFPdelta(); |
7207 | } |
7208 | else |
7209 | { |
7210 | spRelativeOffset = varDsc->lvStkOffs; |
7211 | } |
7212 | |
7213 | assert(spRelativeOffset >= 0); |
7214 | return spRelativeOffset; |
7215 | } |
7216 | |
7217 | /***************************************************************************** |
7218 | * |
7219 | * Return the caller-SP-relative stack offset of a local/parameter. |
7220 | * Requires the local to be on the stack and frame layout to be complete. |
7221 | */ |
7222 | |
7223 | int Compiler::lvaGetCallerSPRelativeOffset(unsigned varNum) |
7224 | { |
7225 | assert(lvaDoneFrameLayout == FINAL_FRAME_LAYOUT); |
7226 | assert(varNum < lvaCount); |
7227 | LclVarDsc* varDsc = lvaTable + varNum; |
7228 | assert(varDsc->lvOnFrame); |
7229 | |
7230 | return lvaToCallerSPRelativeOffset(varDsc->lvStkOffs, varDsc->lvFramePointerBased); |
7231 | } |
7232 | |
7233 | int Compiler::lvaToCallerSPRelativeOffset(int offset, bool isFpBased) |
7234 | { |
7235 | assert(lvaDoneFrameLayout == FINAL_FRAME_LAYOUT); |
7236 | |
7237 | if (isFpBased) |
7238 | { |
7239 | offset += codeGen->genCallerSPtoFPdelta(); |
7240 | } |
7241 | else |
7242 | { |
7243 | offset += codeGen->genCallerSPtoInitialSPdelta(); |
7244 | } |
7245 | |
7246 | return offset; |
7247 | } |
7248 | |
7249 | /***************************************************************************** |
7250 | * |
7251 | * Return the Initial-SP-relative stack offset of a local/parameter. |
7252 | * Requires the local to be on the stack and frame layout to be complete. |
7253 | */ |
7254 | |
7255 | int Compiler::lvaGetInitialSPRelativeOffset(unsigned varNum) |
7256 | { |
7257 | assert(lvaDoneFrameLayout == FINAL_FRAME_LAYOUT); |
7258 | assert(varNum < lvaCount); |
7259 | LclVarDsc* varDsc = lvaTable + varNum; |
7260 | assert(varDsc->lvOnFrame); |
7261 | |
7262 | return lvaToInitialSPRelativeOffset(varDsc->lvStkOffs, varDsc->lvFramePointerBased); |
7263 | } |
7264 | |
7265 | // Given a local variable offset, and whether that offset is frame-pointer based, return its offset from Initial-SP. |
7266 | // This is used, for example, to figure out the offset of the frame pointer from Initial-SP. |
7267 | int Compiler::lvaToInitialSPRelativeOffset(unsigned offset, bool isFpBased) |
7268 | { |
7269 | assert(lvaDoneFrameLayout == FINAL_FRAME_LAYOUT); |
7270 | #ifdef _TARGET_AMD64_ |
7271 | if (isFpBased) |
7272 | { |
7273 | // Currently, the frame starts by pushing ebp, ebp points to the saved ebp |
7274 | // (so we have ebp pointer chaining). Add the fixed-size frame size plus the |
7275 | // size of the callee-saved regs (not including ebp itself) to find Initial-SP. |
7276 | |
7277 | assert(codeGen->isFramePointerUsed()); |
7278 | offset += codeGen->genSPtoFPdelta(); |
7279 | } |
7280 | else |
7281 | { |
7282 | // The offset is correct already! |
7283 | } |
7284 | #else // !_TARGET_AMD64_ |
7285 | NYI("lvaToInitialSPRelativeOffset" ); |
7286 | #endif // !_TARGET_AMD64_ |
7287 | |
7288 | return offset; |
7289 | } |
7290 | |
7291 | /*****************************************************************************/ |
7292 | |
7293 | #ifdef DEBUG |
7294 | /***************************************************************************** |
7295 | * Pick a padding size at "random" for the local. |
7296 | * 0 means that it should not be converted to a GT_LCL_FLD |
7297 | */ |
7298 | |
7299 | static unsigned LCL_FLD_PADDING(unsigned lclNum) |
7300 | { |
7301 | // Convert every 2nd variable |
7302 | if (lclNum % 2) |
7303 | { |
7304 | return 0; |
7305 | } |
7306 | |
7307 | // Pick a padding size at "random" |
7308 | unsigned size = lclNum % 7; |
7309 | |
7310 | return size; |
7311 | } |
7312 | |
7313 | /***************************************************************************** |
7314 | * |
7315 | * Callback for fgWalkAllTreesPre() |
7316 | * Convert as many GT_LCL_VAR's to GT_LCL_FLD's |
7317 | */ |
7318 | |
7319 | /* static */ |
7320 | /* |
7321 | The stress mode does 2 passes. |
7322 | |
7323 | In the first pass we will mark the locals where we CAN't apply the stress mode. |
7324 | In the second pass we will do the appropiate morphing wherever we've not determined we can't do it. |
7325 | */ |
7326 | Compiler::fgWalkResult Compiler::lvaStressLclFldCB(GenTree** pTree, fgWalkData* data) |
7327 | { |
7328 | GenTree* tree = *pTree; |
7329 | genTreeOps oper = tree->OperGet(); |
7330 | GenTree* lcl; |
7331 | |
7332 | switch (oper) |
7333 | { |
7334 | case GT_LCL_VAR: |
7335 | lcl = tree; |
7336 | break; |
7337 | |
7338 | case GT_ADDR: |
7339 | if (tree->gtOp.gtOp1->gtOper != GT_LCL_VAR) |
7340 | { |
7341 | return WALK_CONTINUE; |
7342 | } |
7343 | lcl = tree->gtOp.gtOp1; |
7344 | break; |
7345 | |
7346 | default: |
7347 | return WALK_CONTINUE; |
7348 | } |
7349 | |
7350 | Compiler* pComp = ((lvaStressLclFldArgs*)data->pCallbackData)->m_pCompiler; |
7351 | bool bFirstPass = ((lvaStressLclFldArgs*)data->pCallbackData)->m_bFirstPass; |
7352 | noway_assert(lcl->gtOper == GT_LCL_VAR); |
7353 | unsigned lclNum = lcl->gtLclVarCommon.gtLclNum; |
7354 | var_types type = lcl->TypeGet(); |
7355 | LclVarDsc* varDsc = &pComp->lvaTable[lclNum]; |
7356 | |
7357 | if (varDsc->lvNoLclFldStress) |
7358 | { |
7359 | // Already determined we can't do anything for this var |
7360 | return WALK_SKIP_SUBTREES; |
7361 | } |
7362 | |
7363 | if (bFirstPass) |
7364 | { |
7365 | // Ignore arguments and temps |
7366 | if (varDsc->lvIsParam || lclNum >= pComp->info.compLocalsCount) |
7367 | { |
7368 | varDsc->lvNoLclFldStress = true; |
7369 | return WALK_SKIP_SUBTREES; |
7370 | } |
7371 | |
7372 | // Fix for lcl_fld stress mode |
7373 | if (varDsc->lvKeepType) |
7374 | { |
7375 | varDsc->lvNoLclFldStress = true; |
7376 | return WALK_SKIP_SUBTREES; |
7377 | } |
7378 | |
7379 | // Can't have GC ptrs in TYP_BLK. |
7380 | if (!varTypeIsArithmetic(type)) |
7381 | { |
7382 | varDsc->lvNoLclFldStress = true; |
7383 | return WALK_SKIP_SUBTREES; |
7384 | } |
7385 | |
7386 | // Weed out "small" types like TYP_BYTE as we don't mark the GT_LCL_VAR |
7387 | // node with the accurate small type. If we bash lvaTable[].lvType, |
7388 | // then there will be no indication that it was ever a small type. |
7389 | var_types varType = varDsc->TypeGet(); |
7390 | if (varType != TYP_BLK && genTypeSize(varType) != genTypeSize(genActualType(varType))) |
7391 | { |
7392 | varDsc->lvNoLclFldStress = true; |
7393 | return WALK_SKIP_SUBTREES; |
7394 | } |
7395 | |
7396 | // Offset some of the local variable by a "random" non-zero amount |
7397 | unsigned padding = LCL_FLD_PADDING(lclNum); |
7398 | if (padding == 0) |
7399 | { |
7400 | varDsc->lvNoLclFldStress = true; |
7401 | return WALK_SKIP_SUBTREES; |
7402 | } |
7403 | } |
7404 | else |
7405 | { |
7406 | // Do the morphing |
7407 | noway_assert(varDsc->lvType == lcl->gtType || varDsc->lvType == TYP_BLK); |
7408 | var_types varType = varDsc->TypeGet(); |
7409 | |
7410 | // Calculate padding |
7411 | unsigned padding = LCL_FLD_PADDING(lclNum); |
7412 | |
7413 | #ifdef _TARGET_ARMARCH_ |
7414 | // We need to support alignment requirements to access memory on ARM ARCH |
7415 | unsigned alignment = 1; |
7416 | pComp->codeGen->InferOpSizeAlign(lcl, &alignment); |
7417 | alignment = roundUp(alignment, TARGET_POINTER_SIZE); |
7418 | padding = roundUp(padding, alignment); |
7419 | #endif // _TARGET_ARMARCH_ |
7420 | |
7421 | // Change the variable to a TYP_BLK |
7422 | if (varType != TYP_BLK) |
7423 | { |
7424 | varDsc->lvExactSize = roundUp(padding + pComp->lvaLclSize(lclNum), TARGET_POINTER_SIZE); |
7425 | varDsc->lvType = TYP_BLK; |
7426 | pComp->lvaSetVarAddrExposed(lclNum); |
7427 | } |
7428 | |
7429 | tree->gtFlags |= GTF_GLOB_REF; |
7430 | |
7431 | /* Now morph the tree appropriately */ |
7432 | if (oper == GT_LCL_VAR) |
7433 | { |
7434 | /* Change lclVar(lclNum) to lclFld(lclNum,padding) */ |
7435 | |
7436 | tree->ChangeOper(GT_LCL_FLD); |
7437 | tree->gtLclFld.gtLclOffs = padding; |
7438 | } |
7439 | else |
7440 | { |
7441 | /* Change addr(lclVar) to addr(lclVar)+padding */ |
7442 | |
7443 | noway_assert(oper == GT_ADDR); |
7444 | GenTree* paddingTree = pComp->gtNewIconNode(padding); |
7445 | GenTree* newAddr = pComp->gtNewOperNode(GT_ADD, tree->gtType, tree, paddingTree); |
7446 | |
7447 | *pTree = newAddr; |
7448 | |
7449 | lcl->gtType = TYP_BLK; |
7450 | } |
7451 | } |
7452 | |
7453 | return WALK_SKIP_SUBTREES; |
7454 | } |
7455 | |
7456 | /*****************************************************************************/ |
7457 | |
7458 | void Compiler::lvaStressLclFld() |
7459 | { |
7460 | if (!compStressCompile(STRESS_LCL_FLDS, 5)) |
7461 | { |
7462 | return; |
7463 | } |
7464 | |
7465 | lvaStressLclFldArgs Args; |
7466 | Args.m_pCompiler = this; |
7467 | Args.m_bFirstPass = true; |
7468 | |
7469 | // Do First pass |
7470 | fgWalkAllTreesPre(lvaStressLclFldCB, &Args); |
7471 | |
7472 | // Second pass |
7473 | Args.m_bFirstPass = false; |
7474 | fgWalkAllTreesPre(lvaStressLclFldCB, &Args); |
7475 | } |
7476 | |
7477 | #endif // DEBUG |
7478 | |
7479 | /***************************************************************************** |
7480 | * |
7481 | * A little routine that displays a local variable bitset. |
7482 | * 'set' is mask of variables that have to be displayed |
7483 | * 'allVars' is the complete set of interesting variables (blank space is |
7484 | * inserted if its corresponding bit is not in 'set'). |
7485 | */ |
7486 | |
7487 | #ifdef DEBUG |
7488 | void Compiler::lvaDispVarSet(VARSET_VALARG_TP set) |
7489 | { |
7490 | VARSET_TP allVars(VarSetOps::MakeEmpty(this)); |
7491 | lvaDispVarSet(set, allVars); |
7492 | } |
7493 | |
7494 | void Compiler::lvaDispVarSet(VARSET_VALARG_TP set, VARSET_VALARG_TP allVars) |
7495 | { |
7496 | printf("{" ); |
7497 | |
7498 | bool needSpace = false; |
7499 | |
7500 | for (unsigned index = 0; index < lvaTrackedCount; index++) |
7501 | { |
7502 | if (VarSetOps::IsMember(this, set, index)) |
7503 | { |
7504 | unsigned lclNum; |
7505 | LclVarDsc* varDsc; |
7506 | |
7507 | /* Look for the matching variable */ |
7508 | |
7509 | for (lclNum = 0, varDsc = lvaTable; lclNum < lvaCount; lclNum++, varDsc++) |
7510 | { |
7511 | if ((varDsc->lvVarIndex == index) && varDsc->lvTracked) |
7512 | { |
7513 | break; |
7514 | } |
7515 | } |
7516 | |
7517 | if (needSpace) |
7518 | { |
7519 | printf(" " ); |
7520 | } |
7521 | else |
7522 | { |
7523 | needSpace = true; |
7524 | } |
7525 | |
7526 | printf("V%02u" , lclNum); |
7527 | } |
7528 | else if (VarSetOps::IsMember(this, allVars, index)) |
7529 | { |
7530 | if (needSpace) |
7531 | { |
7532 | printf(" " ); |
7533 | } |
7534 | else |
7535 | { |
7536 | needSpace = true; |
7537 | } |
7538 | |
7539 | printf(" " ); |
7540 | } |
7541 | } |
7542 | |
7543 | printf("}" ); |
7544 | } |
7545 | |
7546 | #endif // DEBUG |
7547 | |