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 Utils.cpp XX |
9 | XX XX |
10 | XX Has miscellaneous utility functions XX |
11 | XX XX |
12 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
13 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
14 | */ |
15 | |
16 | #include "jitpch.h" |
17 | #ifdef _MSC_VER |
18 | #pragma hdrstop |
19 | #endif |
20 | |
21 | #include "opcode.h" |
22 | |
23 | /*****************************************************************************/ |
24 | // Define the string platform name based on compilation #ifdefs. This is the |
25 | // same code for all platforms, hence it is here instead of in the targetXXX.cpp |
26 | // files. |
27 | |
28 | #ifdef _TARGET_UNIX_ |
29 | // Should we distinguish Mac? Can we? |
30 | // Should we distinguish flavors of Unix? Can we? |
31 | const char* Target::g_tgtPlatformName = "Unix" ; |
32 | #else // !_TARGET_UNIX_ |
33 | const char* Target::g_tgtPlatformName = "Windows" ; |
34 | #endif // !_TARGET_UNIX_ |
35 | |
36 | /*****************************************************************************/ |
37 | |
38 | #define DECLARE_DATA |
39 | |
40 | // clang-format off |
41 | extern |
42 | const signed char opcodeSizes[] = |
43 | { |
44 | #define InlineNone_size 0 |
45 | #define ShortInlineVar_size 1 |
46 | #define InlineVar_size 2 |
47 | #define ShortInlineI_size 1 |
48 | #define InlineI_size 4 |
49 | #define InlineI8_size 8 |
50 | #define ShortInlineR_size 4 |
51 | #define InlineR_size 8 |
52 | #define ShortInlineBrTarget_size 1 |
53 | #define InlineBrTarget_size 4 |
54 | #define InlineMethod_size 4 |
55 | #define InlineField_size 4 |
56 | #define InlineType_size 4 |
57 | #define InlineString_size 4 |
58 | #define InlineSig_size 4 |
59 | #define InlineRVA_size 4 |
60 | #define InlineTok_size 4 |
61 | #define InlineSwitch_size 0 // for now |
62 | #define InlinePhi_size 0 // for now |
63 | #define InlineVarTok_size 0 // remove |
64 | |
65 | #define OPDEF(name,string,pop,push,oprType,opcType,l,s1,s2,ctrl) oprType ## _size , |
66 | #include "opcode.def" |
67 | #undef OPDEF |
68 | |
69 | #undef InlineNone_size |
70 | #undef ShortInlineVar_size |
71 | #undef InlineVar_size |
72 | #undef ShortInlineI_size |
73 | #undef InlineI_size |
74 | #undef InlineI8_size |
75 | #undef ShortInlineR_size |
76 | #undef InlineR_size |
77 | #undef ShortInlineBrTarget_size |
78 | #undef InlineBrTarget_size |
79 | #undef InlineMethod_size |
80 | #undef InlineField_size |
81 | #undef InlineType_size |
82 | #undef InlineString_size |
83 | #undef InlineSig_size |
84 | #undef InlineRVA_size |
85 | #undef InlineTok_size |
86 | #undef InlineSwitch_size |
87 | #undef InlinePhi_size |
88 | }; |
89 | // clang-format on |
90 | |
91 | const BYTE varTypeClassification[] = { |
92 | #define DEF_TP(tn, nm, jitType, verType, sz, sze, asze, st, al, tf, howUsed) tf, |
93 | #include "typelist.h" |
94 | #undef DEF_TP |
95 | }; |
96 | |
97 | /*****************************************************************************/ |
98 | /*****************************************************************************/ |
99 | #ifdef DEBUG |
100 | extern const char* const opcodeNames[] = { |
101 | #define OPDEF(name, string, pop, push, oprType, opcType, l, s1, s2, ctrl) string, |
102 | #include "opcode.def" |
103 | #undef OPDEF |
104 | }; |
105 | |
106 | extern const BYTE opcodeArgKinds[] = { |
107 | #define OPDEF(name, string, pop, push, oprType, opcType, l, s1, s2, ctrl) (BYTE) oprType, |
108 | #include "opcode.def" |
109 | #undef OPDEF |
110 | }; |
111 | #endif |
112 | |
113 | /*****************************************************************************/ |
114 | |
115 | const char* varTypeName(var_types vt) |
116 | { |
117 | static const char* const varTypeNames[] = { |
118 | #define DEF_TP(tn, nm, jitType, verType, sz, sze, asze, st, al, tf, howUsed) nm, |
119 | #include "typelist.h" |
120 | #undef DEF_TP |
121 | }; |
122 | |
123 | assert((unsigned)vt < _countof(varTypeNames)); |
124 | |
125 | return varTypeNames[vt]; |
126 | } |
127 | |
128 | #if defined(DEBUG) || defined(LATE_DISASM) |
129 | /***************************************************************************** |
130 | * |
131 | * Return the name of the given register. |
132 | */ |
133 | |
134 | const char* getRegName(regNumber reg, bool isFloat) |
135 | { |
136 | // Special-case REG_NA; it's not in the regNames array, but we might want to print it. |
137 | if (reg == REG_NA) |
138 | { |
139 | return "NA" ; |
140 | } |
141 | |
142 | #if defined(_TARGET_ARM64_) |
143 | static const char* const regNames[] = { |
144 | #define REGDEF(name, rnum, mask, xname, wname) xname, |
145 | #include "register.h" |
146 | }; |
147 | assert(reg < ArrLen(regNames)); |
148 | return regNames[reg]; |
149 | #else |
150 | static const char* const regNames[] = { |
151 | #define REGDEF(name, rnum, mask, sname) sname, |
152 | #include "register.h" |
153 | }; |
154 | assert(reg < ArrLen(regNames)); |
155 | return regNames[reg]; |
156 | #endif |
157 | } |
158 | |
159 | const char* getRegName(unsigned reg, |
160 | bool isFloat) // this is for gcencode.cpp and disasm.cpp that dont use the regNumber type |
161 | { |
162 | return getRegName((regNumber)reg, isFloat); |
163 | } |
164 | #endif // defined(DEBUG) || defined(LATE_DISASM) |
165 | |
166 | #if defined(DEBUG) |
167 | |
168 | const char* getRegNameFloat(regNumber reg, var_types type) |
169 | { |
170 | #ifdef _TARGET_ARM_ |
171 | assert(genIsValidFloatReg(reg)); |
172 | if (type == TYP_FLOAT) |
173 | return getRegName(reg); |
174 | else |
175 | { |
176 | const char* regName; |
177 | |
178 | switch (reg) |
179 | { |
180 | default: |
181 | assert(!"Bad double register" ); |
182 | regName = "d??" ; |
183 | break; |
184 | case REG_F0: |
185 | regName = "d0" ; |
186 | break; |
187 | case REG_F2: |
188 | regName = "d2" ; |
189 | break; |
190 | case REG_F4: |
191 | regName = "d4" ; |
192 | break; |
193 | case REG_F6: |
194 | regName = "d6" ; |
195 | break; |
196 | case REG_F8: |
197 | regName = "d8" ; |
198 | break; |
199 | case REG_F10: |
200 | regName = "d10" ; |
201 | break; |
202 | case REG_F12: |
203 | regName = "d12" ; |
204 | break; |
205 | case REG_F14: |
206 | regName = "d14" ; |
207 | break; |
208 | case REG_F16: |
209 | regName = "d16" ; |
210 | break; |
211 | case REG_F18: |
212 | regName = "d18" ; |
213 | break; |
214 | case REG_F20: |
215 | regName = "d20" ; |
216 | break; |
217 | case REG_F22: |
218 | regName = "d22" ; |
219 | break; |
220 | case REG_F24: |
221 | regName = "d24" ; |
222 | break; |
223 | case REG_F26: |
224 | regName = "d26" ; |
225 | break; |
226 | case REG_F28: |
227 | regName = "d28" ; |
228 | break; |
229 | case REG_F30: |
230 | regName = "d30" ; |
231 | break; |
232 | } |
233 | return regName; |
234 | } |
235 | |
236 | #elif defined(_TARGET_ARM64_) |
237 | |
238 | static const char* regNamesFloat[] = { |
239 | #define REGDEF(name, rnum, mask, xname, wname) xname, |
240 | #include "register.h" |
241 | }; |
242 | assert((unsigned)reg < ArrLen(regNamesFloat)); |
243 | |
244 | return regNamesFloat[reg]; |
245 | |
246 | #else |
247 | static const char* regNamesFloat[] = { |
248 | #define REGDEF(name, rnum, mask, sname) "x" sname, |
249 | #include "register.h" |
250 | }; |
251 | #ifdef FEATURE_SIMD |
252 | static const char* regNamesYMM[] = { |
253 | #define REGDEF(name, rnum, mask, sname) "y" sname, |
254 | #include "register.h" |
255 | }; |
256 | #endif // FEATURE_SIMD |
257 | assert((unsigned)reg < ArrLen(regNamesFloat)); |
258 | |
259 | #ifdef FEATURE_SIMD |
260 | if (type == TYP_SIMD32) |
261 | { |
262 | return regNamesYMM[reg]; |
263 | } |
264 | #endif // FEATURE_SIMD |
265 | |
266 | return regNamesFloat[reg]; |
267 | #endif |
268 | } |
269 | |
270 | /***************************************************************************** |
271 | * |
272 | * Displays a register set. |
273 | * TODO-ARM64-Cleanup: don't allow ip0, ip1 as part of a range. |
274 | */ |
275 | |
276 | void dspRegMask(regMaskTP regMask, size_t minSiz) |
277 | { |
278 | const char* sep = "" ; |
279 | |
280 | printf("[" ); |
281 | |
282 | bool inRegRange = false; |
283 | regNumber regPrev = REG_NA; |
284 | regNumber regHead = REG_NA; // When we start a range, remember the first register of the range, so we don't use |
285 | // range notation if the range contains just a single register. |
286 | for (regNumber regNum = REG_INT_FIRST; regNum <= REG_INT_LAST; regNum = REG_NEXT(regNum)) |
287 | { |
288 | regMaskTP regBit = genRegMask(regNum); |
289 | |
290 | if ((regMask & regBit) != 0) |
291 | { |
292 | // We have a register to display. It gets displayed now if: |
293 | // 1. This is the first register to display of a new range of registers (possibly because |
294 | // no register has ever been displayed). |
295 | // 2. This is the last register of an acceptable range (either the last integer register, |
296 | // or the last of a range that is displayed with range notation). |
297 | if (!inRegRange) |
298 | { |
299 | // It's the first register of a potential range. |
300 | const char* nam = getRegName(regNum); |
301 | printf("%s%s" , sep, nam); |
302 | minSiz -= strlen(sep) + strlen(nam); |
303 | |
304 | // By default, we're not starting a potential register range. |
305 | sep = " " ; |
306 | |
307 | // What kind of separator should we use for this range (if it is indeed going to be a range)? |
308 | CLANG_FORMAT_COMMENT_ANCHOR; |
309 | |
310 | #if defined(_TARGET_AMD64_) |
311 | // For AMD64, create ranges for int registers R8 through R15, but not the "old" registers. |
312 | if (regNum >= REG_R8) |
313 | { |
314 | regHead = regNum; |
315 | inRegRange = true; |
316 | sep = "-" ; |
317 | } |
318 | #elif defined(_TARGET_ARM64_) |
319 | // R17 and R28 can't be the start of a range, since the range would include TEB or FP |
320 | if ((regNum < REG_R17) || ((REG_R19 <= regNum) && (regNum < REG_R28))) |
321 | { |
322 | regHead = regNum; |
323 | inRegRange = true; |
324 | sep = "-" ; |
325 | } |
326 | #elif defined(_TARGET_ARM_) |
327 | if (regNum < REG_R12) |
328 | { |
329 | regHead = regNum; |
330 | inRegRange = true; |
331 | sep = "-" ; |
332 | } |
333 | #elif defined(_TARGET_X86_) |
334 | // No register ranges |
335 | #else // _TARGET_* |
336 | #error Unsupported or unset target architecture |
337 | #endif // _TARGET_* |
338 | } |
339 | |
340 | #if defined(_TARGET_ARM64_) |
341 | // We've already printed a register. Is this the end of a range? |
342 | else if ((regNum == REG_INT_LAST) || (regNum == REG_R17) // last register before TEB |
343 | || (regNum == REG_R28)) // last register before FP |
344 | #else // _TARGET_ARM64_ |
345 | // We've already printed a register. Is this the end of a range? |
346 | else if (regNum == REG_INT_LAST) |
347 | #endif // _TARGET_ARM64_ |
348 | { |
349 | const char* nam = getRegName(regNum); |
350 | printf("%s%s" , sep, nam); |
351 | minSiz -= strlen(sep) + strlen(nam); |
352 | inRegRange = false; // No longer in the middle of a register range |
353 | regHead = REG_NA; |
354 | sep = " " ; |
355 | } |
356 | } |
357 | else // ((regMask & regBit) == 0) |
358 | { |
359 | if (inRegRange) |
360 | { |
361 | assert(regHead != REG_NA); |
362 | if (regPrev != regHead) |
363 | { |
364 | // Close out the previous range, if it included more than one register. |
365 | const char* nam = getRegName(regPrev); |
366 | printf("%s%s" , sep, nam); |
367 | minSiz -= strlen(sep) + strlen(nam); |
368 | } |
369 | sep = " " ; |
370 | inRegRange = false; |
371 | regHead = REG_NA; |
372 | } |
373 | } |
374 | |
375 | if (regBit > regMask) |
376 | { |
377 | break; |
378 | } |
379 | |
380 | regPrev = regNum; |
381 | } |
382 | |
383 | if (strlen(sep) > 0) |
384 | { |
385 | // We've already printed something. |
386 | sep = " " ; |
387 | } |
388 | inRegRange = false; |
389 | regPrev = REG_NA; |
390 | regHead = REG_NA; |
391 | for (regNumber regNum = REG_FP_FIRST; regNum <= REG_FP_LAST; regNum = REG_NEXT(regNum)) |
392 | { |
393 | regMaskTP regBit = genRegMask(regNum); |
394 | |
395 | if (regMask & regBit) |
396 | { |
397 | if (!inRegRange || (regNum == REG_FP_LAST)) |
398 | { |
399 | const char* nam = getRegName(regNum); |
400 | printf("%s%s" , sep, nam); |
401 | minSiz -= strlen(sep) + strlen(nam); |
402 | sep = "-" ; |
403 | regHead = regNum; |
404 | } |
405 | inRegRange = true; |
406 | } |
407 | else |
408 | { |
409 | if (inRegRange) |
410 | { |
411 | if (regPrev != regHead) |
412 | { |
413 | const char* nam = getRegName(regPrev); |
414 | printf("%s%s" , sep, nam); |
415 | minSiz -= (strlen(sep) + strlen(nam)); |
416 | } |
417 | sep = " " ; |
418 | } |
419 | inRegRange = false; |
420 | } |
421 | |
422 | if (regBit > regMask) |
423 | { |
424 | break; |
425 | } |
426 | |
427 | regPrev = regNum; |
428 | } |
429 | |
430 | printf("]" ); |
431 | |
432 | while ((int)minSiz > 0) |
433 | { |
434 | printf(" " ); |
435 | minSiz--; |
436 | } |
437 | } |
438 | |
439 | //------------------------------------------------------------------------ |
440 | // dumpILBytes: Helper for dumpSingleInstr() to dump hex bytes of an IL stream, |
441 | // aligning up to a minimum alignment width. |
442 | // |
443 | // Arguments: |
444 | // codeAddr - Pointer to IL byte stream to display. |
445 | // codeSize - Number of bytes of IL byte stream to display. |
446 | // alignSize - Pad out to this many characters, if fewer than this were written. |
447 | // |
448 | void dumpILBytes(const BYTE* const codeAddr, |
449 | unsigned codeSize, |
450 | unsigned alignSize) // number of characters to write, for alignment |
451 | { |
452 | for (IL_OFFSET offs = 0; offs < codeSize; ++offs) |
453 | { |
454 | printf(" %02x" , *(codeAddr + offs)); |
455 | } |
456 | |
457 | unsigned charsWritten = 3 * codeSize; |
458 | for (unsigned i = charsWritten; i < alignSize; i++) |
459 | { |
460 | printf(" " ); |
461 | } |
462 | } |
463 | |
464 | //------------------------------------------------------------------------ |
465 | // dumpSingleInstr: Display a single IL instruction. |
466 | // |
467 | // Arguments: |
468 | // codeAddr - Base pointer to a stream of IL instructions. |
469 | // offs - Offset from codeAddr of the IL instruction to display. |
470 | // prefix - Optional string to prefix the IL instruction with (if nullptr, no prefix is output). |
471 | // |
472 | // Return Value: |
473 | // Size of the displayed IL instruction in the instruction stream, in bytes. (Add this to 'offs' to |
474 | // get to the next instruction.) |
475 | // |
476 | unsigned dumpSingleInstr(const BYTE* const codeAddr, IL_OFFSET offs, const char* prefix) |
477 | { |
478 | const BYTE* opcodePtr = codeAddr + offs; |
479 | const BYTE* startOpcodePtr = opcodePtr; |
480 | const unsigned ALIGN_WIDTH = 3 * 6; // assume 3 characters * (1 byte opcode + 4 bytes data + 1 prefix byte) for |
481 | // most things |
482 | |
483 | if (prefix != nullptr) |
484 | { |
485 | printf("%s" , prefix); |
486 | } |
487 | |
488 | OPCODE opcode = (OPCODE)getU1LittleEndian(opcodePtr); |
489 | opcodePtr += sizeof(__int8); |
490 | |
491 | DECODE_OPCODE: |
492 | |
493 | if (opcode >= CEE_COUNT) |
494 | { |
495 | printf("\nIllegal opcode: %02X\n" , (int)opcode); |
496 | return (IL_OFFSET)(opcodePtr - startOpcodePtr); |
497 | } |
498 | |
499 | /* Get the size of additional parameters */ |
500 | |
501 | size_t sz = opcodeSizes[opcode]; |
502 | unsigned argKind = opcodeArgKinds[opcode]; |
503 | |
504 | /* See what kind of an opcode we have, then */ |
505 | |
506 | switch (opcode) |
507 | { |
508 | case CEE_PREFIX1: |
509 | opcode = OPCODE(getU1LittleEndian(opcodePtr) + 256); |
510 | opcodePtr += sizeof(__int8); |
511 | goto DECODE_OPCODE; |
512 | |
513 | default: |
514 | { |
515 | __int64 iOp; |
516 | double dOp; |
517 | int jOp; |
518 | DWORD jOp2; |
519 | |
520 | switch (argKind) |
521 | { |
522 | case InlineNone: |
523 | dumpILBytes(startOpcodePtr, (unsigned)(opcodePtr - startOpcodePtr), ALIGN_WIDTH); |
524 | printf(" %-12s" , opcodeNames[opcode]); |
525 | break; |
526 | |
527 | case ShortInlineVar: |
528 | iOp = getU1LittleEndian(opcodePtr); |
529 | goto INT_OP; |
530 | case ShortInlineI: |
531 | iOp = getI1LittleEndian(opcodePtr); |
532 | goto INT_OP; |
533 | case InlineVar: |
534 | iOp = getU2LittleEndian(opcodePtr); |
535 | goto INT_OP; |
536 | case InlineTok: |
537 | case InlineMethod: |
538 | case InlineField: |
539 | case InlineType: |
540 | case InlineString: |
541 | case InlineSig: |
542 | case InlineI: |
543 | iOp = getI4LittleEndian(opcodePtr); |
544 | goto INT_OP; |
545 | case InlineI8: |
546 | iOp = getU4LittleEndian(opcodePtr); |
547 | iOp |= (__int64)getU4LittleEndian(opcodePtr + 4) << 32; |
548 | goto INT_OP; |
549 | |
550 | INT_OP: |
551 | dumpILBytes(startOpcodePtr, (unsigned)((opcodePtr - startOpcodePtr) + sz), ALIGN_WIDTH); |
552 | printf(" %-12s 0x%X" , opcodeNames[opcode], iOp); |
553 | break; |
554 | |
555 | case ShortInlineR: |
556 | dOp = getR4LittleEndian(opcodePtr); |
557 | goto FLT_OP; |
558 | case InlineR: |
559 | dOp = getR8LittleEndian(opcodePtr); |
560 | goto FLT_OP; |
561 | |
562 | FLT_OP: |
563 | dumpILBytes(startOpcodePtr, (unsigned)((opcodePtr - startOpcodePtr) + sz), ALIGN_WIDTH); |
564 | printf(" %-12s %f" , opcodeNames[opcode], dOp); |
565 | break; |
566 | |
567 | case ShortInlineBrTarget: |
568 | jOp = getI1LittleEndian(opcodePtr); |
569 | goto JMP_OP; |
570 | case InlineBrTarget: |
571 | jOp = getI4LittleEndian(opcodePtr); |
572 | goto JMP_OP; |
573 | |
574 | JMP_OP: |
575 | dumpILBytes(startOpcodePtr, (unsigned)((opcodePtr - startOpcodePtr) + sz), ALIGN_WIDTH); |
576 | printf(" %-12s %d (IL_%04x)" , opcodeNames[opcode], jOp, (int)(opcodePtr + sz - codeAddr) + jOp); |
577 | break; |
578 | |
579 | case InlineSwitch: |
580 | jOp2 = getU4LittleEndian(opcodePtr); |
581 | opcodePtr += 4; |
582 | opcodePtr += jOp2 * 4; // Jump over the table |
583 | dumpILBytes(startOpcodePtr, (unsigned)(opcodePtr - startOpcodePtr), ALIGN_WIDTH); |
584 | printf(" %-12s" , opcodeNames[opcode]); |
585 | break; |
586 | |
587 | case InlinePhi: |
588 | jOp2 = getU1LittleEndian(opcodePtr); |
589 | opcodePtr += 1; |
590 | opcodePtr += jOp2 * 2; // Jump over the table |
591 | dumpILBytes(startOpcodePtr, (unsigned)(opcodePtr - startOpcodePtr), ALIGN_WIDTH); |
592 | printf(" %-12s" , opcodeNames[opcode]); |
593 | break; |
594 | |
595 | default: |
596 | assert(!"Bad argKind" ); |
597 | } |
598 | |
599 | opcodePtr += sz; |
600 | break; |
601 | } |
602 | } |
603 | |
604 | printf("\n" ); |
605 | return (IL_OFFSET)(opcodePtr - startOpcodePtr); |
606 | } |
607 | |
608 | //------------------------------------------------------------------------ |
609 | // dumpILRange: Display a range of IL instructions from an IL instruction stream. |
610 | // |
611 | // Arguments: |
612 | // codeAddr - Pointer to IL byte stream to display. |
613 | // codeSize - Number of bytes of IL byte stream to display. |
614 | // |
615 | void dumpILRange(const BYTE* const codeAddr, unsigned codeSize) // in bytes |
616 | { |
617 | for (IL_OFFSET offs = 0; offs < codeSize;) |
618 | { |
619 | char prefix[100]; |
620 | sprintf_s(prefix, _countof(prefix), "IL_%04x " , offs); |
621 | unsigned codeBytesDumped = dumpSingleInstr(codeAddr, offs, prefix); |
622 | offs += codeBytesDumped; |
623 | } |
624 | } |
625 | |
626 | /***************************************************************************** |
627 | * |
628 | * Display a variable set. |
629 | */ |
630 | const char* genES2str(BitVecTraits* traits, EXPSET_TP set) |
631 | { |
632 | const int bufSize = 17; |
633 | static char num1[bufSize]; |
634 | |
635 | static char num2[bufSize]; |
636 | |
637 | static char* nump = num1; |
638 | |
639 | char* temp = nump; |
640 | |
641 | nump = (nump == num1) ? num2 : num1; |
642 | |
643 | sprintf_s(temp, bufSize, "%s" , BitVecOps::ToString(traits, set)); |
644 | |
645 | return temp; |
646 | } |
647 | |
648 | const char* refCntWtd2str(unsigned refCntWtd) |
649 | { |
650 | const int bufSize = 17; |
651 | static char num1[bufSize]; |
652 | |
653 | static char num2[bufSize]; |
654 | |
655 | static char* nump = num1; |
656 | |
657 | char* temp = nump; |
658 | |
659 | nump = (nump == num1) ? num2 : num1; |
660 | |
661 | if (refCntWtd == BB_MAX_WEIGHT) |
662 | { |
663 | sprintf_s(temp, bufSize, "MAX " ); |
664 | } |
665 | else |
666 | { |
667 | unsigned valueInt = refCntWtd / BB_UNITY_WEIGHT; |
668 | unsigned valueFrac = refCntWtd % BB_UNITY_WEIGHT; |
669 | |
670 | if (valueFrac == 0) |
671 | { |
672 | sprintf_s(temp, bufSize, "%u " , valueInt); |
673 | } |
674 | else |
675 | { |
676 | sprintf_s(temp, bufSize, "%u.%02u" , valueInt, (valueFrac * 100 / BB_UNITY_WEIGHT)); |
677 | } |
678 | } |
679 | return temp; |
680 | } |
681 | |
682 | #endif // DEBUG |
683 | |
684 | #if defined(DEBUG) || defined(INLINE_DATA) |
685 | |
686 | //------------------------------------------------------------------------ |
687 | // Contains: check if the range includes a particular method |
688 | // |
689 | // Arguments: |
690 | // info -- jit interface pointer |
691 | // method -- method handle for the method of interest |
692 | |
693 | bool ConfigMethodRange::Contains(ICorJitInfo* info, CORINFO_METHOD_HANDLE method) |
694 | { |
695 | _ASSERT(m_inited == 1); |
696 | |
697 | // No ranges specified means all methods included. |
698 | if (m_lastRange == 0) |
699 | { |
700 | return true; |
701 | } |
702 | |
703 | // Check the hash. Note we can't use the cached hash here since |
704 | // we may not be asking about the method currently being jitted. |
705 | const unsigned hash = info->getMethodHash(method); |
706 | |
707 | for (unsigned i = 0; i < m_lastRange; i++) |
708 | { |
709 | if ((m_ranges[i].m_low <= hash) && (hash <= m_ranges[i].m_high)) |
710 | { |
711 | return true; |
712 | } |
713 | } |
714 | |
715 | return false; |
716 | } |
717 | |
718 | //------------------------------------------------------------------------ |
719 | // InitRanges: parse the range string and set up the range info |
720 | // |
721 | // Arguments: |
722 | // rangeStr -- string to parse (may be nullptr) |
723 | // capacity -- number ranges to allocate in the range array |
724 | // |
725 | // Notes: |
726 | // Does some internal error checking; clients can use Error() |
727 | // to determine if the range string couldn't be fully parsed |
728 | // because of bad characters or too many entries, or had values |
729 | // that were too large to represent. |
730 | |
731 | void ConfigMethodRange::InitRanges(const wchar_t* rangeStr, unsigned capacity) |
732 | { |
733 | // Make sure that the memory was zero initialized |
734 | assert(m_inited == 0 || m_inited == 1); |
735 | assert(m_entries == 0); |
736 | assert(m_ranges == nullptr); |
737 | assert(m_lastRange == 0); |
738 | |
739 | // Flag any crazy-looking requests |
740 | assert(capacity < 100000); |
741 | |
742 | if (rangeStr == nullptr) |
743 | { |
744 | m_inited = 1; |
745 | return; |
746 | } |
747 | |
748 | // Allocate some persistent memory |
749 | ICorJitHost* jitHost = g_jitHost; |
750 | m_ranges = (Range*)jitHost->allocateMemory(capacity * sizeof(Range)); |
751 | m_entries = capacity; |
752 | |
753 | const wchar_t* p = rangeStr; |
754 | unsigned lastRange = 0; |
755 | bool setHighPart = false; |
756 | |
757 | while ((*p != 0) && (lastRange < m_entries)) |
758 | { |
759 | while (*p == L' ') |
760 | { |
761 | p++; |
762 | } |
763 | |
764 | int i = 0; |
765 | |
766 | while (L'0' <= *p && *p <= L'9') |
767 | { |
768 | int j = 10 * i + ((*p++) - L'0'); |
769 | |
770 | // Check for overflow |
771 | if ((m_badChar != 0) && (j <= i)) |
772 | { |
773 | m_badChar = (p - rangeStr) + 1; |
774 | } |
775 | |
776 | i = j; |
777 | } |
778 | |
779 | // Was this the high part of a low-high pair? |
780 | if (setHighPart) |
781 | { |
782 | // Yep, set it and move to the next range |
783 | m_ranges[lastRange].m_high = i; |
784 | |
785 | // Sanity check that range is proper |
786 | if ((m_badChar != 0) && (m_ranges[lastRange].m_high < m_ranges[lastRange].m_low)) |
787 | { |
788 | m_badChar = (p - rangeStr) + 1; |
789 | } |
790 | |
791 | lastRange++; |
792 | setHighPart = false; |
793 | continue; |
794 | } |
795 | |
796 | // Must have been looking for the low part of a range |
797 | m_ranges[lastRange].m_low = i; |
798 | |
799 | while (*p == L' ') |
800 | { |
801 | p++; |
802 | } |
803 | |
804 | // Was that the low part of a low-high pair? |
805 | if (*p == L'-') |
806 | { |
807 | // Yep, skip the dash and set high part next time around. |
808 | p++; |
809 | setHighPart = true; |
810 | continue; |
811 | } |
812 | |
813 | // Else we have a point range, so set high = low |
814 | m_ranges[lastRange].m_high = i; |
815 | lastRange++; |
816 | } |
817 | |
818 | // If we didn't parse the full range string, note index of the the |
819 | // first bad char. |
820 | if ((m_badChar != 0) && (*p != 0)) |
821 | { |
822 | m_badChar = (p - rangeStr) + 1; |
823 | } |
824 | |
825 | // Finish off any remaining open range |
826 | if (setHighPart) |
827 | { |
828 | m_ranges[lastRange].m_high = UINT_MAX; |
829 | lastRange++; |
830 | } |
831 | |
832 | assert(lastRange <= m_entries); |
833 | m_lastRange = lastRange; |
834 | m_inited = 1; |
835 | } |
836 | |
837 | #endif // defined(DEBUG) || defined(INLINE_DATA) |
838 | |
839 | #if CALL_ARG_STATS || COUNT_BASIC_BLOCKS || COUNT_LOOPS || EMITTER_STATS || MEASURE_NODE_SIZE || MEASURE_MEM_ALLOC |
840 | |
841 | /***************************************************************************** |
842 | * Histogram class. |
843 | */ |
844 | |
845 | Histogram::Histogram(const unsigned* const sizeTable) : m_sizeTable(sizeTable) |
846 | { |
847 | unsigned sizeCount = 0; |
848 | do |
849 | { |
850 | sizeCount++; |
851 | } while ((sizeTable[sizeCount] != 0) && (sizeCount < 1000)); |
852 | |
853 | assert(sizeCount < HISTOGRAM_MAX_SIZE_COUNT - 1); |
854 | |
855 | m_sizeCount = sizeCount; |
856 | |
857 | memset(m_counts, 0, (m_sizeCount + 1) * sizeof(*m_counts)); |
858 | } |
859 | |
860 | void Histogram::dump(FILE* output) |
861 | { |
862 | unsigned t = 0; |
863 | for (unsigned i = 0; i < m_sizeCount; i++) |
864 | { |
865 | t += m_counts[i]; |
866 | } |
867 | |
868 | for (unsigned c = 0, i = 0; i <= m_sizeCount; i++) |
869 | { |
870 | if (i == m_sizeCount) |
871 | { |
872 | if (m_counts[i] == 0) |
873 | { |
874 | break; |
875 | } |
876 | |
877 | fprintf(output, " > %7u" , m_sizeTable[i - 1]); |
878 | } |
879 | else |
880 | { |
881 | if (i == 0) |
882 | { |
883 | fprintf(output, " <= " ); |
884 | } |
885 | else |
886 | { |
887 | fprintf(output, "%7u .. " , m_sizeTable[i - 1] + 1); |
888 | } |
889 | |
890 | fprintf(output, "%7u" , m_sizeTable[i]); |
891 | } |
892 | |
893 | c += m_counts[i]; |
894 | |
895 | fprintf(output, " ===> %7u count (%3u%% of total)\n" , m_counts[i], (int)(100.0 * c / t)); |
896 | } |
897 | } |
898 | |
899 | void Histogram::record(unsigned size) |
900 | { |
901 | unsigned i; |
902 | for (i = 0; i < m_sizeCount; i++) |
903 | { |
904 | if (m_sizeTable[i] >= size) |
905 | { |
906 | break; |
907 | } |
908 | } |
909 | |
910 | m_counts[i]++; |
911 | } |
912 | |
913 | #endif // CALL_ARG_STATS || COUNT_BASIC_BLOCKS || COUNT_LOOPS || EMITTER_STATS || MEASURE_NODE_SIZE |
914 | |
915 | /***************************************************************************** |
916 | * Fixed bit vector class |
917 | */ |
918 | |
919 | // bitChunkSize() - Returns number of bits in a bitVect chunk |
920 | inline UINT FixedBitVect::bitChunkSize() |
921 | { |
922 | return sizeof(UINT) * 8; |
923 | } |
924 | |
925 | // bitNumToBit() - Returns a bit mask of the given bit number |
926 | inline UINT FixedBitVect::bitNumToBit(UINT bitNum) |
927 | { |
928 | assert(bitNum < bitChunkSize()); |
929 | assert(bitChunkSize() <= sizeof(int) * 8); |
930 | |
931 | return 1 << bitNum; |
932 | } |
933 | |
934 | // bitVectInit() - Initializes a bit vector of a given size |
935 | FixedBitVect* FixedBitVect::bitVectInit(UINT size, Compiler* comp) |
936 | { |
937 | UINT bitVectMemSize, numberOfChunks; |
938 | FixedBitVect* bv; |
939 | |
940 | assert(size != 0); |
941 | |
942 | numberOfChunks = (size - 1) / bitChunkSize() + 1; |
943 | bitVectMemSize = numberOfChunks * (bitChunkSize() / 8); // size in bytes |
944 | |
945 | assert(bitVectMemSize * bitChunkSize() >= size); |
946 | |
947 | bv = (FixedBitVect*)comp->getAllocator(CMK_FixedBitVect).allocate<char>(sizeof(FixedBitVect) + bitVectMemSize); |
948 | memset(bv->bitVect, 0, bitVectMemSize); |
949 | |
950 | bv->bitVectSize = size; |
951 | |
952 | return bv; |
953 | } |
954 | |
955 | // bitVectSet() - Sets the given bit |
956 | void FixedBitVect::bitVectSet(UINT bitNum) |
957 | { |
958 | UINT index; |
959 | |
960 | assert(bitNum <= bitVectSize); |
961 | |
962 | index = bitNum / bitChunkSize(); |
963 | bitNum -= index * bitChunkSize(); |
964 | |
965 | bitVect[index] |= bitNumToBit(bitNum); |
966 | } |
967 | |
968 | // bitVectTest() - Tests the given bit |
969 | bool FixedBitVect::bitVectTest(UINT bitNum) |
970 | { |
971 | UINT index; |
972 | |
973 | assert(bitNum <= bitVectSize); |
974 | |
975 | index = bitNum / bitChunkSize(); |
976 | bitNum -= index * bitChunkSize(); |
977 | |
978 | return (bitVect[index] & bitNumToBit(bitNum)) != 0; |
979 | } |
980 | |
981 | // bitVectOr() - Or in the given bit vector |
982 | void FixedBitVect::bitVectOr(FixedBitVect* bv) |
983 | { |
984 | UINT bitChunkCnt = (bitVectSize - 1) / bitChunkSize() + 1; |
985 | |
986 | assert(bitVectSize == bv->bitVectSize); |
987 | |
988 | // Or each chunks |
989 | for (UINT i = 0; i < bitChunkCnt; i++) |
990 | { |
991 | bitVect[i] |= bv->bitVect[i]; |
992 | } |
993 | } |
994 | |
995 | // bitVectAnd() - And with passed in bit vector |
996 | void FixedBitVect::bitVectAnd(FixedBitVect& bv) |
997 | { |
998 | UINT bitChunkCnt = (bitVectSize - 1) / bitChunkSize() + 1; |
999 | |
1000 | assert(bitVectSize == bv.bitVectSize); |
1001 | |
1002 | // And each chunks |
1003 | for (UINT i = 0; i < bitChunkCnt; i++) |
1004 | { |
1005 | bitVect[i] &= bv.bitVect[i]; |
1006 | } |
1007 | } |
1008 | |
1009 | // bitVectGetFirst() - Find the first bit on and return bit num, |
1010 | // Return -1 if no bits found. |
1011 | UINT FixedBitVect::bitVectGetFirst() |
1012 | { |
1013 | return bitVectGetNext((UINT)-1); |
1014 | } |
1015 | |
1016 | // bitVectGetNext() - Find the next bit on given previous position and return bit num. |
1017 | // Return -1 if no bits found. |
1018 | UINT FixedBitVect::bitVectGetNext(UINT bitNumPrev) |
1019 | { |
1020 | UINT bitNum = (UINT)-1; |
1021 | UINT index; |
1022 | UINT bitMask; |
1023 | UINT bitChunkCnt = (bitVectSize - 1) / bitChunkSize() + 1; |
1024 | UINT i; |
1025 | |
1026 | if (bitNumPrev == (UINT)-1) |
1027 | { |
1028 | index = 0; |
1029 | bitMask = (UINT)-1; |
1030 | } |
1031 | else |
1032 | { |
1033 | UINT bit; |
1034 | |
1035 | index = bitNumPrev / bitChunkSize(); |
1036 | bitNumPrev -= index * bitChunkSize(); |
1037 | bit = bitNumToBit(bitNumPrev); |
1038 | bitMask = ~(bit | (bit - 1)); |
1039 | } |
1040 | |
1041 | // Find first bit |
1042 | for (i = index; i < bitChunkCnt; i++) |
1043 | { |
1044 | UINT bitChunk = bitVect[i] & bitMask; |
1045 | |
1046 | if (bitChunk != 0) |
1047 | { |
1048 | BitScanForward((ULONG*)&bitNum, bitChunk); |
1049 | break; |
1050 | } |
1051 | |
1052 | bitMask = 0xFFFFFFFF; |
1053 | } |
1054 | |
1055 | // Empty bit vector? |
1056 | if (bitNum == (UINT)-1) |
1057 | { |
1058 | return (UINT)-1; |
1059 | } |
1060 | |
1061 | bitNum += i * bitChunkSize(); |
1062 | |
1063 | assert(bitNum <= bitVectSize); |
1064 | |
1065 | return bitNum; |
1066 | } |
1067 | |
1068 | // bitVectGetNextAndClear() - Find the first bit on, clear it and return it. |
1069 | // Return -1 if no bits found. |
1070 | UINT FixedBitVect::bitVectGetNextAndClear() |
1071 | { |
1072 | UINT bitNum = (UINT)-1; |
1073 | UINT bitChunkCnt = (bitVectSize - 1) / bitChunkSize() + 1; |
1074 | UINT i; |
1075 | |
1076 | // Find first bit |
1077 | for (i = 0; i < bitChunkCnt; i++) |
1078 | { |
1079 | if (bitVect[i] != 0) |
1080 | { |
1081 | BitScanForward((ULONG*)&bitNum, bitVect[i]); |
1082 | break; |
1083 | } |
1084 | } |
1085 | |
1086 | // Empty bit vector? |
1087 | if (bitNum == (UINT)-1) |
1088 | { |
1089 | return (UINT)-1; |
1090 | } |
1091 | |
1092 | // Clear the bit in the right chunk |
1093 | bitVect[i] &= ~bitNumToBit(bitNum); |
1094 | |
1095 | bitNum += i * bitChunkSize(); |
1096 | |
1097 | assert(bitNum <= bitVectSize); |
1098 | |
1099 | return bitNum; |
1100 | } |
1101 | |
1102 | int SimpleSprintf_s(__in_ecount(cbBufSize - (pWriteStart - pBufStart)) char* pWriteStart, |
1103 | __in_ecount(cbBufSize) char* pBufStart, |
1104 | size_t cbBufSize, |
1105 | __in_z const char* fmt, |
1106 | ...) |
1107 | { |
1108 | assert(fmt); |
1109 | assert(pBufStart); |
1110 | assert(pWriteStart); |
1111 | assert((size_t)pBufStart <= (size_t)pWriteStart); |
1112 | int ret; |
1113 | |
1114 | // compute the space left in the buffer. |
1115 | if ((pBufStart + cbBufSize) < pWriteStart) |
1116 | { |
1117 | NO_WAY("pWriteStart is past end of buffer" ); |
1118 | } |
1119 | size_t cbSpaceLeft = (size_t)((pBufStart + cbBufSize) - pWriteStart); |
1120 | va_list args; |
1121 | va_start(args, fmt); |
1122 | ret = vsprintf_s(pWriteStart, cbSpaceLeft, const_cast<char*>(fmt), args); |
1123 | va_end(args); |
1124 | if (ret < 0) |
1125 | { |
1126 | NO_WAY("vsprintf_s failed." ); |
1127 | } |
1128 | return ret; |
1129 | } |
1130 | |
1131 | #ifdef DEBUG |
1132 | |
1133 | void hexDump(FILE* dmpf, const char* name, BYTE* addr, size_t size) |
1134 | { |
1135 | if (!size) |
1136 | { |
1137 | return; |
1138 | } |
1139 | |
1140 | assert(addr); |
1141 | |
1142 | fprintf(dmpf, "Hex dump of %s:\n" , name); |
1143 | |
1144 | for (unsigned i = 0; i < size; i++) |
1145 | { |
1146 | if ((i % 16) == 0) |
1147 | { |
1148 | fprintf(dmpf, "\n %04X: " , i); |
1149 | } |
1150 | |
1151 | fprintf(dmpf, "%02X " , *addr++); |
1152 | } |
1153 | |
1154 | fprintf(dmpf, "\n\n" ); |
1155 | } |
1156 | |
1157 | #endif // DEBUG |
1158 | |
1159 | void HelperCallProperties::init() |
1160 | { |
1161 | for (CorInfoHelpFunc helper = CORINFO_HELP_UNDEF; // initialize helper |
1162 | (helper < CORINFO_HELP_COUNT); // test helper for loop exit |
1163 | helper = CorInfoHelpFunc(int(helper) + 1)) // update helper to next |
1164 | { |
1165 | // Generally you want initialize these to their most typical/safest result |
1166 | // |
1167 | bool isPure = false; // true if the result only depends upon input args and not any global state |
1168 | bool noThrow = false; // true if the helper will never throw |
1169 | bool nonNullReturn = false; // true if the result will never be null or zero |
1170 | bool isAllocator = false; // true if the result is usually a newly created heap item, or may throw OutOfMemory |
1171 | bool mutatesHeap = false; // true if any previous heap objects [are|can be] modified |
1172 | bool mayRunCctor = false; // true if the helper call may cause a static constructor to be run. |
1173 | |
1174 | switch (helper) |
1175 | { |
1176 | // Arithmetic helpers that cannot throw |
1177 | case CORINFO_HELP_LLSH: |
1178 | case CORINFO_HELP_LRSH: |
1179 | case CORINFO_HELP_LRSZ: |
1180 | case CORINFO_HELP_LMUL: |
1181 | case CORINFO_HELP_LNG2DBL: |
1182 | case CORINFO_HELP_ULNG2DBL: |
1183 | case CORINFO_HELP_DBL2INT: |
1184 | case CORINFO_HELP_DBL2LNG: |
1185 | case CORINFO_HELP_DBL2UINT: |
1186 | case CORINFO_HELP_DBL2ULNG: |
1187 | case CORINFO_HELP_FLTREM: |
1188 | case CORINFO_HELP_DBLREM: |
1189 | case CORINFO_HELP_FLTROUND: |
1190 | case CORINFO_HELP_DBLROUND: |
1191 | |
1192 | isPure = true; |
1193 | noThrow = true; |
1194 | break; |
1195 | |
1196 | // Arithmetic helpers that *can* throw. |
1197 | |
1198 | // This (or these) are not pure, in that they have "VM side effects"...but they don't mutate the heap. |
1199 | case CORINFO_HELP_ENDCATCH: |
1200 | |
1201 | noThrow = true; |
1202 | break; |
1203 | |
1204 | // Arithmetic helpers that may throw |
1205 | case CORINFO_HELP_LMOD: // Mods throw div-by zero, and signed mods have problems with the smallest integer |
1206 | // mod -1, |
1207 | case CORINFO_HELP_MOD: // which is not representable as a positive integer. |
1208 | case CORINFO_HELP_UMOD: |
1209 | case CORINFO_HELP_ULMOD: |
1210 | |
1211 | case CORINFO_HELP_UDIV: // Divs throw divide-by-zero. |
1212 | case CORINFO_HELP_DIV: |
1213 | case CORINFO_HELP_LDIV: |
1214 | case CORINFO_HELP_ULDIV: |
1215 | |
1216 | case CORINFO_HELP_LMUL_OVF: |
1217 | case CORINFO_HELP_ULMUL_OVF: |
1218 | case CORINFO_HELP_DBL2INT_OVF: |
1219 | case CORINFO_HELP_DBL2LNG_OVF: |
1220 | case CORINFO_HELP_DBL2UINT_OVF: |
1221 | case CORINFO_HELP_DBL2ULNG_OVF: |
1222 | |
1223 | isPure = true; |
1224 | break; |
1225 | |
1226 | // Heap Allocation helpers, these all never return null |
1227 | case CORINFO_HELP_NEWSFAST: |
1228 | case CORINFO_HELP_NEWSFAST_ALIGN8: |
1229 | case CORINFO_HELP_NEWSFAST_ALIGN8_VC: |
1230 | case CORINFO_HELP_NEW_CROSSCONTEXT: |
1231 | case CORINFO_HELP_NEWFAST: |
1232 | case CORINFO_HELP_NEWSFAST_FINALIZE: |
1233 | case CORINFO_HELP_NEWSFAST_ALIGN8_FINALIZE: |
1234 | case CORINFO_HELP_READYTORUN_NEW: |
1235 | case CORINFO_HELP_BOX: |
1236 | |
1237 | isAllocator = true; |
1238 | nonNullReturn = true; |
1239 | noThrow = true; // only can throw OutOfMemory |
1240 | break; |
1241 | |
1242 | // These allocation helpers do some checks on the size (and lower bound) inputs, |
1243 | // and can throw exceptions other than OOM. |
1244 | case CORINFO_HELP_NEWARR_1_VC: |
1245 | case CORINFO_HELP_NEWARR_1_ALIGN8: |
1246 | case CORINFO_HELP_NEW_MDARR: |
1247 | case CORINFO_HELP_NEWARR_1_DIRECT: |
1248 | case CORINFO_HELP_NEWARR_1_OBJ: |
1249 | case CORINFO_HELP_NEWARR_1_R2R_DIRECT: |
1250 | case CORINFO_HELP_READYTORUN_NEWARR_1: |
1251 | |
1252 | isAllocator = true; |
1253 | nonNullReturn = true; |
1254 | break; |
1255 | |
1256 | // Heap Allocation helpers that are also pure |
1257 | case CORINFO_HELP_STRCNS: |
1258 | |
1259 | isPure = true; |
1260 | isAllocator = true; |
1261 | nonNullReturn = true; |
1262 | noThrow = true; // only can throw OutOfMemory |
1263 | break; |
1264 | |
1265 | case CORINFO_HELP_BOX_NULLABLE: |
1266 | // Box Nullable is not a 'pure' function |
1267 | // It has a Byref argument that it reads the contents of. |
1268 | // |
1269 | // So two calls to Box Nullable that pass the same address (with the same Value Number) |
1270 | // will produce different results when the contents of the memory pointed to by the Byref changes |
1271 | // |
1272 | isAllocator = true; |
1273 | noThrow = true; // only can throw OutOfMemory |
1274 | break; |
1275 | |
1276 | case CORINFO_HELP_RUNTIMEHANDLE_METHOD: |
1277 | case CORINFO_HELP_RUNTIMEHANDLE_CLASS: |
1278 | case CORINFO_HELP_RUNTIMEHANDLE_METHOD_LOG: |
1279 | case CORINFO_HELP_RUNTIMEHANDLE_CLASS_LOG: |
1280 | // logging helpers are not technically pure but can be optimized away |
1281 | isPure = true; |
1282 | noThrow = true; |
1283 | nonNullReturn = true; |
1284 | break; |
1285 | |
1286 | // type casting helpers |
1287 | case CORINFO_HELP_ISINSTANCEOFINTERFACE: |
1288 | case CORINFO_HELP_ISINSTANCEOFARRAY: |
1289 | case CORINFO_HELP_ISINSTANCEOFCLASS: |
1290 | case CORINFO_HELP_ISINSTANCEOFANY: |
1291 | case CORINFO_HELP_READYTORUN_ISINSTANCEOF: |
1292 | case CORINFO_HELP_TYPEHANDLE_TO_RUNTIMETYPE: |
1293 | case CORINFO_HELP_TYPEHANDLE_TO_RUNTIMETYPEHANDLE: |
1294 | |
1295 | isPure = true; |
1296 | noThrow = true; // These return null for a failing cast |
1297 | break; |
1298 | |
1299 | case CORINFO_HELP_ARE_TYPES_EQUIVALENT: |
1300 | |
1301 | isPure = true; |
1302 | noThrow = true; |
1303 | break; |
1304 | |
1305 | // type casting helpers that throw |
1306 | case CORINFO_HELP_CHKCASTINTERFACE: |
1307 | case CORINFO_HELP_CHKCASTARRAY: |
1308 | case CORINFO_HELP_CHKCASTCLASS: |
1309 | case CORINFO_HELP_CHKCASTANY: |
1310 | case CORINFO_HELP_CHKCASTCLASS_SPECIAL: |
1311 | case CORINFO_HELP_READYTORUN_CHKCAST: |
1312 | |
1313 | // These throw for a failing cast |
1314 | // But if given a null input arg will return null |
1315 | isPure = true; |
1316 | break; |
1317 | |
1318 | // helpers returning addresses, these can also throw |
1319 | case CORINFO_HELP_UNBOX: |
1320 | case CORINFO_HELP_GETREFANY: |
1321 | case CORINFO_HELP_LDELEMA_REF: |
1322 | |
1323 | isPure = true; |
1324 | break; |
1325 | |
1326 | // helpers that return internal handle |
1327 | case CORINFO_HELP_GETCLASSFROMMETHODPARAM: |
1328 | case CORINFO_HELP_GETSYNCFROMCLASSHANDLE: |
1329 | |
1330 | isPure = true; |
1331 | noThrow = true; |
1332 | break; |
1333 | |
1334 | // Helpers that load the base address for static variables. |
1335 | // We divide these between those that may and may not invoke |
1336 | // static class constructors. |
1337 | case CORINFO_HELP_GETSHARED_GCSTATIC_BASE: |
1338 | case CORINFO_HELP_GETSHARED_NONGCSTATIC_BASE: |
1339 | case CORINFO_HELP_GETSHARED_GCSTATIC_BASE_DYNAMICCLASS: |
1340 | case CORINFO_HELP_GETSHARED_NONGCSTATIC_BASE_DYNAMICCLASS: |
1341 | case CORINFO_HELP_GETGENERICS_GCTHREADSTATIC_BASE: |
1342 | case CORINFO_HELP_GETGENERICS_NONGCTHREADSTATIC_BASE: |
1343 | case CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE: |
1344 | case CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE: |
1345 | case CORINFO_HELP_CLASSINIT_SHARED_DYNAMICCLASS: |
1346 | case CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_DYNAMICCLASS: |
1347 | case CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_DYNAMICCLASS: |
1348 | case CORINFO_HELP_GETSTATICFIELDADDR_CONTEXT: |
1349 | case CORINFO_HELP_GETSTATICFIELDADDR_TLS: |
1350 | case CORINFO_HELP_GETGENERICS_GCSTATIC_BASE: |
1351 | case CORINFO_HELP_GETGENERICS_NONGCSTATIC_BASE: |
1352 | case CORINFO_HELP_READYTORUN_STATIC_BASE: |
1353 | case CORINFO_HELP_READYTORUN_GENERIC_STATIC_BASE: |
1354 | |
1355 | // These may invoke static class constructors |
1356 | // These can throw InvalidProgram exception if the class can not be constructed |
1357 | // |
1358 | isPure = true; |
1359 | nonNullReturn = true; |
1360 | mayRunCctor = true; |
1361 | break; |
1362 | |
1363 | case CORINFO_HELP_GETSHARED_GCSTATIC_BASE_NOCTOR: |
1364 | case CORINFO_HELP_GETSHARED_NONGCSTATIC_BASE_NOCTOR: |
1365 | case CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR: |
1366 | case CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR: |
1367 | |
1368 | // These do not invoke static class constructors |
1369 | // |
1370 | isPure = true; |
1371 | noThrow = true; |
1372 | nonNullReturn = true; |
1373 | break; |
1374 | |
1375 | // GC Write barrier support |
1376 | // TODO-ARM64-Bug?: Can these throw or not? |
1377 | case CORINFO_HELP_ASSIGN_REF: |
1378 | case CORINFO_HELP_CHECKED_ASSIGN_REF: |
1379 | case CORINFO_HELP_ASSIGN_REF_ENSURE_NONHEAP: |
1380 | case CORINFO_HELP_ASSIGN_BYREF: |
1381 | case CORINFO_HELP_ASSIGN_STRUCT: |
1382 | |
1383 | mutatesHeap = true; |
1384 | break; |
1385 | |
1386 | // Accessing fields (write) |
1387 | case CORINFO_HELP_SETFIELD32: |
1388 | case CORINFO_HELP_SETFIELD64: |
1389 | case CORINFO_HELP_SETFIELDOBJ: |
1390 | case CORINFO_HELP_SETFIELDSTRUCT: |
1391 | case CORINFO_HELP_SETFIELDFLOAT: |
1392 | case CORINFO_HELP_SETFIELDDOUBLE: |
1393 | case CORINFO_HELP_ARRADDR_ST: |
1394 | |
1395 | mutatesHeap = true; |
1396 | break; |
1397 | |
1398 | // These helper calls always throw an exception |
1399 | case CORINFO_HELP_OVERFLOW: |
1400 | case CORINFO_HELP_VERIFICATION: |
1401 | case CORINFO_HELP_RNGCHKFAIL: |
1402 | case CORINFO_HELP_THROWDIVZERO: |
1403 | case CORINFO_HELP_THROWNULLREF: |
1404 | case CORINFO_HELP_THROW: |
1405 | case CORINFO_HELP_RETHROW: |
1406 | case CORINFO_HELP_THROW_ARGUMENTEXCEPTION: |
1407 | case CORINFO_HELP_THROW_ARGUMENTOUTOFRANGEEXCEPTION: |
1408 | case CORINFO_HELP_THROW_NOT_IMPLEMENTED: |
1409 | case CORINFO_HELP_THROW_PLATFORM_NOT_SUPPORTED: |
1410 | case CORINFO_HELP_THROW_TYPE_NOT_SUPPORTED: |
1411 | |
1412 | break; |
1413 | |
1414 | // These helper calls may throw an exception |
1415 | case CORINFO_HELP_METHOD_ACCESS_CHECK: |
1416 | case CORINFO_HELP_FIELD_ACCESS_CHECK: |
1417 | case CORINFO_HELP_CLASS_ACCESS_CHECK: |
1418 | case CORINFO_HELP_DELEGATE_SECURITY_CHECK: |
1419 | case CORINFO_HELP_MON_EXIT_STATIC: |
1420 | |
1421 | break; |
1422 | |
1423 | // This is a debugging aid; it simply returns a constant address. |
1424 | case CORINFO_HELP_LOOP_CLONE_CHOICE_ADDR: |
1425 | isPure = true; |
1426 | noThrow = true; |
1427 | break; |
1428 | |
1429 | case CORINFO_HELP_DBG_IS_JUST_MY_CODE: |
1430 | case CORINFO_HELP_BBT_FCN_ENTER: |
1431 | case CORINFO_HELP_POLL_GC: |
1432 | case CORINFO_HELP_MON_ENTER: |
1433 | case CORINFO_HELP_MON_EXIT: |
1434 | case CORINFO_HELP_MON_ENTER_STATIC: |
1435 | case CORINFO_HELP_JIT_REVERSE_PINVOKE_ENTER: |
1436 | case CORINFO_HELP_JIT_REVERSE_PINVOKE_EXIT: |
1437 | case CORINFO_HELP_SECURITY_PROLOG: |
1438 | case CORINFO_HELP_SECURITY_PROLOG_FRAMED: |
1439 | case CORINFO_HELP_VERIFICATION_RUNTIME_CHECK: |
1440 | case CORINFO_HELP_GETFIELDADDR: |
1441 | case CORINFO_HELP_INIT_PINVOKE_FRAME: |
1442 | case CORINFO_HELP_JIT_PINVOKE_BEGIN: |
1443 | case CORINFO_HELP_JIT_PINVOKE_END: |
1444 | case CORINFO_HELP_GETCURRENTMANAGEDTHREADID: |
1445 | |
1446 | noThrow = true; |
1447 | break; |
1448 | |
1449 | // Not sure how to handle optimization involving the rest of these helpers |
1450 | default: |
1451 | |
1452 | // The most pessimistic results are returned for these helpers |
1453 | mutatesHeap = true; |
1454 | break; |
1455 | } |
1456 | |
1457 | m_isPure[helper] = isPure; |
1458 | m_noThrow[helper] = noThrow; |
1459 | m_nonNullReturn[helper] = nonNullReturn; |
1460 | m_isAllocator[helper] = isAllocator; |
1461 | m_mutatesHeap[helper] = mutatesHeap; |
1462 | m_mayRunCctor[helper] = mayRunCctor; |
1463 | } |
1464 | } |
1465 | |
1466 | //============================================================================= |
1467 | // AssemblyNamesList2 |
1468 | //============================================================================= |
1469 | // The string should be of the form |
1470 | // MyAssembly |
1471 | // MyAssembly;mscorlib;System |
1472 | // MyAssembly;mscorlib System |
1473 | |
1474 | AssemblyNamesList2::AssemblyNamesList2(const wchar_t* list, HostAllocator alloc) : m_alloc(alloc) |
1475 | { |
1476 | WCHAR prevChar = '?'; // dummy |
1477 | LPWSTR nameStart = nullptr; // start of the name currently being processed. nullptr if no current name |
1478 | AssemblyName** ppPrevLink = &m_pNames; |
1479 | |
1480 | for (LPWSTR listWalk = const_cast<LPWSTR>(list); prevChar != '\0'; prevChar = *listWalk, listWalk++) |
1481 | { |
1482 | WCHAR curChar = *listWalk; |
1483 | |
1484 | if (iswspace(curChar) || curChar == W(';') || curChar == W('\0')) |
1485 | { |
1486 | // |
1487 | // Found white-space |
1488 | // |
1489 | |
1490 | if (nameStart) |
1491 | { |
1492 | // Found the end of the current name; add a new assembly name to the list. |
1493 | |
1494 | AssemblyName* newName = new (m_alloc) AssemblyName(); |
1495 | |
1496 | // Null out the current character so we can do zero-terminated string work; we'll restore it later. |
1497 | *listWalk = W('\0'); |
1498 | |
1499 | // How much space do we need? |
1500 | int convertedNameLenBytes = |
1501 | WszWideCharToMultiByte(CP_UTF8, 0, nameStart, -1, nullptr, 0, nullptr, nullptr); |
1502 | newName->m_assemblyName = new (m_alloc) char[convertedNameLenBytes]; // convertedNameLenBytes includes |
1503 | // the trailing null character |
1504 | if (WszWideCharToMultiByte(CP_UTF8, 0, nameStart, -1, newName->m_assemblyName, convertedNameLenBytes, |
1505 | nullptr, nullptr) != 0) |
1506 | { |
1507 | *ppPrevLink = newName; |
1508 | ppPrevLink = &newName->m_next; |
1509 | } |
1510 | else |
1511 | { |
1512 | // Failed to convert the string. Ignore this string (and leak the memory). |
1513 | } |
1514 | |
1515 | nameStart = nullptr; |
1516 | |
1517 | // Restore the current character. |
1518 | *listWalk = curChar; |
1519 | } |
1520 | } |
1521 | else if (!nameStart) |
1522 | { |
1523 | // |
1524 | // Found the start of a new name |
1525 | // |
1526 | |
1527 | nameStart = listWalk; |
1528 | } |
1529 | } |
1530 | |
1531 | assert(nameStart == nullptr); // cannot be in the middle of a name |
1532 | *ppPrevLink = nullptr; // Terminate the last element of the list. |
1533 | } |
1534 | |
1535 | AssemblyNamesList2::~AssemblyNamesList2() |
1536 | { |
1537 | for (AssemblyName* pName = m_pNames; pName != nullptr; /**/) |
1538 | { |
1539 | AssemblyName* cur = pName; |
1540 | pName = pName->m_next; |
1541 | |
1542 | m_alloc.deallocate(cur->m_assemblyName); |
1543 | m_alloc.deallocate(cur); |
1544 | } |
1545 | } |
1546 | |
1547 | bool AssemblyNamesList2::IsInList(const char* assemblyName) |
1548 | { |
1549 | for (AssemblyName* pName = m_pNames; pName != nullptr; pName = pName->m_next) |
1550 | { |
1551 | if (_stricmp(pName->m_assemblyName, assemblyName) == 0) |
1552 | { |
1553 | return true; |
1554 | } |
1555 | } |
1556 | |
1557 | return false; |
1558 | } |
1559 | |
1560 | #ifdef FEATURE_JIT_METHOD_PERF |
1561 | CycleCount::CycleCount() : cps(CycleTimer::CyclesPerSecond()) |
1562 | { |
1563 | } |
1564 | |
1565 | bool CycleCount::GetCycles(unsigned __int64* time) |
1566 | { |
1567 | return CycleTimer::GetThreadCyclesS(time); |
1568 | } |
1569 | |
1570 | bool CycleCount::Start() |
1571 | { |
1572 | return GetCycles(&beginCycles); |
1573 | } |
1574 | |
1575 | double CycleCount::ElapsedTime() |
1576 | { |
1577 | unsigned __int64 nowCycles; |
1578 | (void)GetCycles(&nowCycles); |
1579 | return ((double)(nowCycles - beginCycles) / cps) * 1000.0; |
1580 | } |
1581 | |
1582 | bool PerfCounter::Start() |
1583 | { |
1584 | bool result = QueryPerformanceFrequency(&beg) != 0; |
1585 | if (!result) |
1586 | { |
1587 | return result; |
1588 | } |
1589 | freq = (double)beg.QuadPart / 1000.0; |
1590 | (void)QueryPerformanceCounter(&beg); |
1591 | return result; |
1592 | } |
1593 | |
1594 | // Return elapsed time from Start() in millis. |
1595 | double PerfCounter::ElapsedTime() |
1596 | { |
1597 | LARGE_INTEGER li; |
1598 | (void)QueryPerformanceCounter(&li); |
1599 | return (double)(li.QuadPart - beg.QuadPart) / freq; |
1600 | } |
1601 | |
1602 | #endif |
1603 | |
1604 | #ifdef DEBUG |
1605 | |
1606 | /***************************************************************************** |
1607 | * Return the number of digits in a number of the given base (default base 10). |
1608 | * Used when outputting strings. |
1609 | */ |
1610 | unsigned CountDigits(unsigned num, unsigned base /* = 10 */) |
1611 | { |
1612 | assert(2 <= base && base <= 16); // sanity check |
1613 | unsigned count = 1; |
1614 | while (num >= base) |
1615 | { |
1616 | num /= base; |
1617 | ++count; |
1618 | } |
1619 | return count; |
1620 | } |
1621 | |
1622 | #endif // DEBUG |
1623 | |
1624 | double FloatingPointUtils::convertUInt64ToDouble(unsigned __int64 uIntVal) |
1625 | { |
1626 | __int64 s64 = uIntVal; |
1627 | double d; |
1628 | if (s64 < 0) |
1629 | { |
1630 | #if defined(_TARGET_XARCH_) |
1631 | // RyuJIT codegen and clang (or gcc) may produce different results for casting uint64 to |
1632 | // double, and the clang result is more accurate. For example, |
1633 | // 1) (double)0x84595161401484A0UL --> 43e08b2a2c280290 (RyuJIT codegen or VC++) |
1634 | // 2) (double)0x84595161401484A0UL --> 43e08b2a2c280291 (clang or gcc) |
1635 | // If the folding optimization below is implemented by simple casting of (double)uint64_val |
1636 | // and it is compiled by clang, casting result can be inconsistent, depending on whether |
1637 | // the folding optimization is triggered or the codegen generates instructions for casting. // |
1638 | // The current solution is to force the same math as the codegen does, so that casting |
1639 | // result is always consistent. |
1640 | |
1641 | // d = (double)(int64_t)uint64 + 0x1p64 |
1642 | uint64_t adjHex = 0x43F0000000000000UL; |
1643 | d = (double)s64 + *(double*)&adjHex; |
1644 | #else |
1645 | d = (double)uIntVal; |
1646 | #endif |
1647 | } |
1648 | else |
1649 | { |
1650 | d = (double)uIntVal; |
1651 | } |
1652 | return d; |
1653 | } |
1654 | |
1655 | float FloatingPointUtils::convertUInt64ToFloat(unsigned __int64 u64) |
1656 | { |
1657 | double d = convertUInt64ToDouble(u64); |
1658 | return (float)d; |
1659 | } |
1660 | |
1661 | unsigned __int64 FloatingPointUtils::convertDoubleToUInt64(double d) |
1662 | { |
1663 | unsigned __int64 u64; |
1664 | if (d >= 0.0) |
1665 | { |
1666 | // Work around a C++ issue where it doesn't properly convert large positive doubles |
1667 | const double two63 = 2147483648.0 * 4294967296.0; |
1668 | if (d < two63) |
1669 | { |
1670 | u64 = UINT64(d); |
1671 | } |
1672 | else |
1673 | { |
1674 | // subtract 0x8000000000000000, do the convert then add it back again |
1675 | u64 = INT64(d - two63) + I64(0x8000000000000000); |
1676 | } |
1677 | return u64; |
1678 | } |
1679 | |
1680 | #ifdef _TARGET_XARCH_ |
1681 | |
1682 | // While the Ecma spec does not specifically call this out, |
1683 | // the case of conversion from negative double to unsigned integer is |
1684 | // effectively an overflow and therefore the result is unspecified. |
1685 | // With MSVC for x86/x64, such a conversion results in the bit-equivalent |
1686 | // unsigned value of the conversion to integer. Other compilers convert |
1687 | // negative doubles to zero when the target is unsigned. |
1688 | // To make the behavior consistent across OS's on TARGET_XARCH, |
1689 | // this double cast is needed to conform MSVC behavior. |
1690 | |
1691 | u64 = UINT64(INT64(d)); |
1692 | #else |
1693 | u64 = UINT64(d); |
1694 | #endif // _TARGET_XARCH_ |
1695 | |
1696 | return u64; |
1697 | } |
1698 | |
1699 | // Rounds a double-precision floating-point value to the nearest integer, |
1700 | // and rounds midpoint values to the nearest even number. |
1701 | double FloatingPointUtils::round(double x) |
1702 | { |
1703 | // ************************************************************************************ |
1704 | // IMPORTANT: Do not change this implementation without also updating Math.Round(double), |
1705 | // MathF.Round(float), and FloatingPointUtils::round(float) |
1706 | // ************************************************************************************ |
1707 | |
1708 | // If the number has no fractional part do nothing |
1709 | // This shortcut is necessary to workaround precision loss in borderline cases on some platforms |
1710 | |
1711 | if (x == (double)((INT64)x)) |
1712 | { |
1713 | return x; |
1714 | } |
1715 | |
1716 | // We had a number that was equally close to 2 integers. |
1717 | // We need to return the even one. |
1718 | |
1719 | double flrTempVal = floor(x + 0.5); |
1720 | |
1721 | if ((x == (floor(x) + 0.5)) && (fmod(flrTempVal, 2.0) != 0)) |
1722 | { |
1723 | flrTempVal -= 1.0; |
1724 | } |
1725 | |
1726 | return _copysign(flrTempVal, x); |
1727 | } |
1728 | |
1729 | // Windows x86 and Windows ARM/ARM64 may not define _copysignf() but they do define _copysign(). |
1730 | // We will redirect the macro to this other functions if the macro is not defined for the platform. |
1731 | // This has the side effect of a possible implicit upcasting for arguments passed in and an explicit |
1732 | // downcasting for the _copysign() call. |
1733 | #if (defined(_TARGET_X86_) || defined(_TARGET_ARM_) || defined(_TARGET_ARM64_)) && !defined(FEATURE_PAL) |
1734 | |
1735 | #if !defined(_copysignf) |
1736 | #define _copysignf (float)_copysign |
1737 | #endif |
1738 | |
1739 | #endif |
1740 | |
1741 | // Rounds a single-precision floating-point value to the nearest integer, |
1742 | // and rounds midpoint values to the nearest even number. |
1743 | float FloatingPointUtils::round(float x) |
1744 | { |
1745 | // ************************************************************************************ |
1746 | // IMPORTANT: Do not change this implementation without also updating MathF.Round(float), |
1747 | // Math.Round(double), and FloatingPointUtils::round(double) |
1748 | // ************************************************************************************ |
1749 | |
1750 | // If the number has no fractional part do nothing |
1751 | // This shortcut is necessary to workaround precision loss in borderline cases on some platforms |
1752 | |
1753 | if (x == (float)((INT32)x)) |
1754 | { |
1755 | return x; |
1756 | } |
1757 | |
1758 | // We had a number that was equally close to 2 integers. |
1759 | // We need to return the even one. |
1760 | |
1761 | float flrTempVal = floorf(x + 0.5f); |
1762 | |
1763 | if ((x == (floorf(x) + 0.5f)) && (fmodf(flrTempVal, 2.0f) != 0)) |
1764 | { |
1765 | flrTempVal -= 1.0f; |
1766 | } |
1767 | |
1768 | return _copysignf(flrTempVal, x); |
1769 | } |
1770 | |
1771 | namespace MagicDivide |
1772 | { |
1773 | template <int TableBase = 0, int TableSize, typename Magic> |
1774 | static const Magic* TryGetMagic(const Magic (&table)[TableSize], typename Magic::DivisorType index) |
1775 | { |
1776 | if ((index < TableBase) || (TableBase + TableSize <= index)) |
1777 | { |
1778 | return nullptr; |
1779 | } |
1780 | |
1781 | const Magic* p = &table[index - TableBase]; |
1782 | |
1783 | if (p->magic == 0) |
1784 | { |
1785 | return nullptr; |
1786 | } |
1787 | |
1788 | return p; |
1789 | }; |
1790 | |
1791 | template <typename T> |
1792 | struct UnsignedMagic |
1793 | { |
1794 | typedef T DivisorType; |
1795 | |
1796 | T magic; |
1797 | bool add; |
1798 | int shift; |
1799 | }; |
1800 | |
1801 | template <typename T> |
1802 | const UnsignedMagic<T>* TryGetUnsignedMagic(T divisor) |
1803 | { |
1804 | return nullptr; |
1805 | } |
1806 | |
1807 | template <> |
1808 | const UnsignedMagic<uint32_t>* TryGetUnsignedMagic(uint32_t divisor) |
1809 | { |
1810 | static const UnsignedMagic<uint32_t> table[]{ |
1811 | {0xaaaaaaab, false, 1}, // 3 |
1812 | {}, |
1813 | {0xcccccccd, false, 2}, // 5 |
1814 | {0xaaaaaaab, false, 2}, // 6 |
1815 | {0x24924925, true, 3}, // 7 |
1816 | {}, |
1817 | {0x38e38e39, false, 1}, // 9 |
1818 | {0xcccccccd, false, 3}, // 10 |
1819 | {0xba2e8ba3, false, 3}, // 11 |
1820 | {0xaaaaaaab, false, 3}, // 12 |
1821 | }; |
1822 | |
1823 | return TryGetMagic<3>(table, divisor); |
1824 | } |
1825 | |
1826 | template <> |
1827 | const UnsignedMagic<uint64_t>* TryGetUnsignedMagic(uint64_t divisor) |
1828 | { |
1829 | static const UnsignedMagic<uint64_t> table[]{ |
1830 | {0xaaaaaaaaaaaaaaab, false, 1}, // 3 |
1831 | {}, |
1832 | {0xcccccccccccccccd, false, 2}, // 5 |
1833 | {0xaaaaaaaaaaaaaaab, false, 2}, // 6 |
1834 | {0x2492492492492493, true, 3}, // 7 |
1835 | {}, |
1836 | {0xe38e38e38e38e38f, false, 3}, // 9 |
1837 | {0xcccccccccccccccd, false, 3}, // 10 |
1838 | {0x2e8ba2e8ba2e8ba3, false, 1}, // 11 |
1839 | {0xaaaaaaaaaaaaaaab, false, 3}, // 12 |
1840 | }; |
1841 | |
1842 | return TryGetMagic<3>(table, divisor); |
1843 | } |
1844 | |
1845 | //------------------------------------------------------------------------ |
1846 | // GetUnsignedMagic: Generates a magic number and shift amount for the magic |
1847 | // number unsigned division optimization. |
1848 | // |
1849 | // Arguments: |
1850 | // d - The divisor |
1851 | // add - Pointer to a flag indicating the kind of code to generate |
1852 | // shift - Pointer to the shift value to be returned |
1853 | // |
1854 | // Returns: |
1855 | // The magic number. |
1856 | // |
1857 | // Notes: |
1858 | // This code is adapted from _The_PowerPC_Compiler_Writer's_Guide_, pages 57-58. |
1859 | // The paper is based on "Division by invariant integers using multiplication" |
1860 | // by Torbjorn Granlund and Peter L. Montgomery in PLDI 94 |
1861 | |
1862 | template <typename T> |
1863 | T GetUnsignedMagic(T d, bool* add /*out*/, int* shift /*out*/) |
1864 | { |
1865 | assert((d >= 3) && !isPow2(d)); |
1866 | |
1867 | const UnsignedMagic<T>* magic = TryGetUnsignedMagic(d); |
1868 | |
1869 | if (magic != nullptr) |
1870 | { |
1871 | *shift = magic->shift; |
1872 | *add = magic->add; |
1873 | return magic->magic; |
1874 | } |
1875 | |
1876 | typedef typename jitstd::make_signed<T>::type ST; |
1877 | |
1878 | const unsigned bits = sizeof(T) * 8; |
1879 | const unsigned bitsMinus1 = bits - 1; |
1880 | const T twoNMinus1 = T(1) << bitsMinus1; |
1881 | |
1882 | *add = false; |
1883 | const T nc = -ST(1) - -ST(d) % ST(d); |
1884 | unsigned p = bitsMinus1; |
1885 | T q1 = twoNMinus1 / nc; |
1886 | T r1 = twoNMinus1 - (q1 * nc); |
1887 | T q2 = (twoNMinus1 - 1) / d; |
1888 | T r2 = (twoNMinus1 - 1) - (q2 * d); |
1889 | T delta; |
1890 | |
1891 | do |
1892 | { |
1893 | p++; |
1894 | |
1895 | if (r1 >= (nc - r1)) |
1896 | { |
1897 | q1 = 2 * q1 + 1; |
1898 | r1 = 2 * r1 - nc; |
1899 | } |
1900 | else |
1901 | { |
1902 | q1 = 2 * q1; |
1903 | r1 = 2 * r1; |
1904 | } |
1905 | |
1906 | if ((r2 + 1) >= (d - r2)) |
1907 | { |
1908 | if (q2 >= (twoNMinus1 - 1)) |
1909 | { |
1910 | *add = true; |
1911 | } |
1912 | |
1913 | q2 = 2 * q2 + 1; |
1914 | r2 = 2 * r2 + 1 - d; |
1915 | } |
1916 | else |
1917 | { |
1918 | if (q2 >= twoNMinus1) |
1919 | { |
1920 | *add = true; |
1921 | } |
1922 | |
1923 | q2 = 2 * q2; |
1924 | r2 = 2 * r2 + 1; |
1925 | } |
1926 | |
1927 | delta = d - 1 - r2; |
1928 | |
1929 | } while ((p < (bits * 2)) && ((q1 < delta) || ((q1 == delta) && (r1 == 0)))); |
1930 | |
1931 | *shift = p - bits; // resulting shift |
1932 | return q2 + 1; // resulting magic number |
1933 | } |
1934 | |
1935 | uint32_t GetUnsigned32Magic(uint32_t d, bool* add /*out*/, int* shift /*out*/) |
1936 | { |
1937 | return GetUnsignedMagic<uint32_t>(d, add, shift); |
1938 | } |
1939 | |
1940 | #ifdef _TARGET_64BIT_ |
1941 | uint64_t GetUnsigned64Magic(uint64_t d, bool* add /*out*/, int* shift /*out*/) |
1942 | { |
1943 | return GetUnsignedMagic<uint64_t>(d, add, shift); |
1944 | } |
1945 | #endif |
1946 | |
1947 | template <typename T> |
1948 | struct SignedMagic |
1949 | { |
1950 | typedef T DivisorType; |
1951 | |
1952 | T magic; |
1953 | int shift; |
1954 | }; |
1955 | |
1956 | template <typename T> |
1957 | const SignedMagic<T>* TryGetSignedMagic(T divisor) |
1958 | { |
1959 | return nullptr; |
1960 | } |
1961 | |
1962 | template <> |
1963 | const SignedMagic<int32_t>* TryGetSignedMagic(int32_t divisor) |
1964 | { |
1965 | static const SignedMagic<int32_t> table[]{ |
1966 | {0x55555556, 0}, // 3 |
1967 | {}, |
1968 | {0x66666667, 1}, // 5 |
1969 | {0x2aaaaaab, 0}, // 6 |
1970 | {0x92492493, 2}, // 7 |
1971 | {}, |
1972 | {0x38e38e39, 1}, // 9 |
1973 | {0x66666667, 2}, // 10 |
1974 | {0x2e8ba2e9, 1}, // 11 |
1975 | {0x2aaaaaab, 1}, // 12 |
1976 | }; |
1977 | |
1978 | return TryGetMagic<3>(table, divisor); |
1979 | } |
1980 | |
1981 | template <> |
1982 | const SignedMagic<int64_t>* TryGetSignedMagic(int64_t divisor) |
1983 | { |
1984 | static const SignedMagic<int64_t> table[]{ |
1985 | {0x5555555555555556, 0}, // 3 |
1986 | {}, |
1987 | {0x6666666666666667, 1}, // 5 |
1988 | {0x2aaaaaaaaaaaaaab, 0}, // 6 |
1989 | {0x4924924924924925, 1}, // 7 |
1990 | {}, |
1991 | {0x1c71c71c71c71c72, 0}, // 9 |
1992 | {0x6666666666666667, 2}, // 10 |
1993 | {0x2e8ba2e8ba2e8ba3, 1}, // 11 |
1994 | {0x2aaaaaaaaaaaaaab, 1}, // 12 |
1995 | }; |
1996 | |
1997 | return TryGetMagic<3>(table, divisor); |
1998 | } |
1999 | |
2000 | //------------------------------------------------------------------------ |
2001 | // GetSignedMagic: Generates a magic number and shift amount for |
2002 | // the magic number division optimization. |
2003 | // |
2004 | // Arguments: |
2005 | // denom - The denominator |
2006 | // shift - Pointer to the shift value to be returned |
2007 | // |
2008 | // Returns: |
2009 | // The magic number. |
2010 | // |
2011 | // Notes: |
2012 | // This code is previously from UTC where it notes it was taken from |
2013 | // _The_PowerPC_Compiler_Writer's_Guide_, pages 57-58. The paper is based on |
2014 | // is "Division by invariant integers using multiplication" by Torbjorn Granlund |
2015 | // and Peter L. Montgomery in PLDI 94 |
2016 | |
2017 | template <typename T> |
2018 | T GetSignedMagic(T denom, int* shift /*out*/) |
2019 | { |
2020 | const SignedMagic<T>* magic = TryGetSignedMagic(denom); |
2021 | |
2022 | if (magic != nullptr) |
2023 | { |
2024 | *shift = magic->shift; |
2025 | return magic->magic; |
2026 | } |
2027 | |
2028 | const int bits = sizeof(T) * 8; |
2029 | const int bits_minus_1 = bits - 1; |
2030 | |
2031 | typedef typename jitstd::make_unsigned<T>::type UT; |
2032 | |
2033 | const UT two_nminus1 = UT(1) << bits_minus_1; |
2034 | |
2035 | int p; |
2036 | UT absDenom; |
2037 | UT absNc; |
2038 | UT delta; |
2039 | UT q1; |
2040 | UT r1; |
2041 | UT r2; |
2042 | UT q2; |
2043 | UT t; |
2044 | T result_magic; |
2045 | int iters = 0; |
2046 | |
2047 | absDenom = abs(denom); |
2048 | t = two_nminus1 + ((unsigned int)denom >> 31); |
2049 | absNc = t - 1 - (t % absDenom); // absolute value of nc |
2050 | p = bits_minus_1; // initialize p |
2051 | q1 = two_nminus1 / absNc; // initialize q1 = 2^p / abs(nc) |
2052 | r1 = two_nminus1 - (q1 * absNc); // initialize r1 = rem(2^p, abs(nc)) |
2053 | q2 = two_nminus1 / absDenom; // initialize q1 = 2^p / abs(denom) |
2054 | r2 = two_nminus1 - (q2 * absDenom); // initialize r1 = rem(2^p, abs(denom)) |
2055 | |
2056 | do |
2057 | { |
2058 | iters++; |
2059 | p++; |
2060 | q1 *= 2; // update q1 = 2^p / abs(nc) |
2061 | r1 *= 2; // update r1 = rem(2^p / abs(nc)) |
2062 | |
2063 | if (r1 >= absNc) |
2064 | { // must be unsigned comparison |
2065 | q1++; |
2066 | r1 -= absNc; |
2067 | } |
2068 | |
2069 | q2 *= 2; // update q2 = 2^p / abs(denom) |
2070 | r2 *= 2; // update r2 = rem(2^p / abs(denom)) |
2071 | |
2072 | if (r2 >= absDenom) |
2073 | { // must be unsigned comparison |
2074 | q2++; |
2075 | r2 -= absDenom; |
2076 | } |
2077 | |
2078 | delta = absDenom - r2; |
2079 | } while (q1 < delta || (q1 == delta && r1 == 0)); |
2080 | |
2081 | result_magic = q2 + 1; // resulting magic number |
2082 | if (denom < 0) |
2083 | { |
2084 | result_magic = -result_magic; |
2085 | } |
2086 | *shift = p - bits; // resulting shift |
2087 | |
2088 | return result_magic; |
2089 | } |
2090 | |
2091 | int32_t GetSigned32Magic(int32_t d, int* shift /*out*/) |
2092 | { |
2093 | return GetSignedMagic<int32_t>(d, shift); |
2094 | } |
2095 | |
2096 | #ifdef _TARGET_64BIT_ |
2097 | int64_t GetSigned64Magic(int64_t d, int* shift /*out*/) |
2098 | { |
2099 | return GetSignedMagic<int64_t>(d, shift); |
2100 | } |
2101 | #endif |
2102 | } |
2103 | |