1 //===-- AutoUpgrade.cpp - Implement auto-upgrade helper functions ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the auto-upgrade helper functions
12 //===----------------------------------------------------------------------===//
14 #include "llvm/AutoUpgrade.h"
15 #include "llvm/Constants.h"
16 #include "llvm/Function.h"
17 #include "llvm/LLVMContext.h"
18 #include "llvm/Module.h"
19 #include "llvm/IntrinsicInst.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/Support/CallSite.h"
22 #include "llvm/Support/ErrorHandling.h"
23 #include "llvm/Support/IRBuilder.h"
28 static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
29 assert(F && "Illegal to upgrade a non-existent Function.");
31 // Get the Function's name.
32 const std::string& Name = F->getName();
35 const FunctionType *FTy = F->getFunctionType();
37 // Quickly eliminate it, if it's not a candidate.
38 if (Name.length() <= 8 || Name[0] != 'l' || Name[1] != 'l' ||
39 Name[2] != 'v' || Name[3] != 'm' || Name[4] != '.')
42 Module *M = F->getParent();
46 // This upgrades the llvm.atomic.lcs, llvm.atomic.las, llvm.atomic.lss,
47 // and atomics with default address spaces to their new names to their new
48 // function name (e.g. llvm.atomic.add.i32 => llvm.atomic.add.i32.p0i32)
49 if (Name.compare(5,7,"atomic.",7) == 0) {
50 if (Name.compare(12,3,"lcs",3) == 0) {
51 std::string::size_type delim = Name.find('.',12);
52 F->setName("llvm.atomic.cmp.swap" + Name.substr(delim) +
53 ".p0" + Name.substr(delim+1));
57 else if (Name.compare(12,3,"las",3) == 0) {
58 std::string::size_type delim = Name.find('.',12);
59 F->setName("llvm.atomic.load.add"+Name.substr(delim)
60 + ".p0" + Name.substr(delim+1));
64 else if (Name.compare(12,3,"lss",3) == 0) {
65 std::string::size_type delim = Name.find('.',12);
66 F->setName("llvm.atomic.load.sub"+Name.substr(delim)
67 + ".p0" + Name.substr(delim+1));
71 else if (Name.rfind(".p") == std::string::npos) {
72 // We don't have an address space qualifier so this has be upgraded
73 // to the new name. Copy the type name at the end of the intrinsic
75 std::string::size_type delim = Name.find_last_of('.');
76 assert(delim != std::string::npos && "can not find type");
77 F->setName(Name + ".p0" + Name.substr(delim+1));
81 } else if (Name.compare(5, 9, "arm.neon.", 9) == 0) {
82 if (Name.compare(14, 7, "vmovls.", 7) == 0 ||
83 Name.compare(14, 7, "vmovlu.", 7) == 0) {
84 // Calls to these are transformed into IR without intrinsics.
88 // Old versions of NEON ld/st intrinsics are missing alignment arguments.
89 bool isVLd = (Name.compare(14, 3, "vld", 3) == 0);
90 bool isVSt = (Name.compare(14, 3, "vst", 3) == 0);
92 unsigned NumVecs = Name.at(17) - '0';
93 if (NumVecs == 0 || NumVecs > 4)
95 bool isLaneOp = (Name.compare(18, 5, "lane.", 5) == 0);
96 if (!isLaneOp && Name.at(18) != '.')
98 unsigned ExpectedArgs = 2; // for the address and alignment
99 if (isVSt || isLaneOp)
100 ExpectedArgs += NumVecs;
102 ExpectedArgs += 1; // for the lane number
103 unsigned NumP = FTy->getNumParams();
104 if (NumP != ExpectedArgs - 1)
107 // Change the name of the old (bad) intrinsic, because
108 // its type is incorrect, but we cannot overload that name.
111 // One argument is missing: add the alignment argument.
112 std::vector<const Type*> NewParams;
113 for (unsigned p = 0; p < NumP; ++p)
114 NewParams.push_back(FTy->getParamType(p));
115 NewParams.push_back(Type::getInt32Ty(F->getContext()));
116 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(),
118 NewFn = cast<Function>(M->getOrInsertFunction(Name, NewFTy));
124 // This upgrades the name of the llvm.bswap intrinsic function to only use
125 // a single type name for overloading. We only care about the old format
126 // 'llvm.bswap.i*.i*', so check for 'bswap.' and then for there being
127 // a '.' after 'bswap.'
128 if (Name.compare(5,6,"bswap.",6) == 0) {
129 std::string::size_type delim = Name.find('.',11);
131 if (delim != std::string::npos) {
132 // Construct the new name as 'llvm.bswap' + '.i*'
133 F->setName(Name.substr(0,10)+Name.substr(delim));
141 // We only want to fix the 'llvm.ct*' intrinsics which do not have the
142 // correct return type, so we check for the name, and then check if the
143 // return type does not match the parameter type.
144 if ( (Name.compare(5,5,"ctpop",5) == 0 ||
145 Name.compare(5,4,"ctlz",4) == 0 ||
146 Name.compare(5,4,"cttz",4) == 0) &&
147 FTy->getReturnType() != FTy->getParamType(0)) {
148 // We first need to change the name of the old (bad) intrinsic, because
149 // its type is incorrect, but we cannot overload that name. We
150 // arbitrarily unique it here allowing us to construct a correctly named
151 // and typed function below.
154 // Now construct the new intrinsic with the correct name and type. We
155 // leave the old function around in order to query its type, whatever it
156 // may be, and correctly convert up to the new type.
157 NewFn = cast<Function>(M->getOrInsertFunction(Name,
158 FTy->getParamType(0),
159 FTy->getParamType(0),
166 // The old llvm.eh.selector.i32 is equivalent to the new llvm.eh.selector.
167 if (Name.compare("llvm.eh.selector.i32") == 0) {
168 F->setName("llvm.eh.selector");
172 // The old llvm.eh.typeid.for.i32 is equivalent to llvm.eh.typeid.for.
173 if (Name.compare("llvm.eh.typeid.for.i32") == 0) {
174 F->setName("llvm.eh.typeid.for");
178 // Convert the old llvm.eh.selector.i64 to a call to llvm.eh.selector.
179 if (Name.compare("llvm.eh.selector.i64") == 0) {
180 NewFn = Intrinsic::getDeclaration(M, Intrinsic::eh_selector);
183 // Convert the old llvm.eh.typeid.for.i64 to a call to llvm.eh.typeid.for.
184 if (Name.compare("llvm.eh.typeid.for.i64") == 0) {
185 NewFn = Intrinsic::getDeclaration(M, Intrinsic::eh_typeid_for);
191 // This upgrades the llvm.memcpy, llvm.memmove, and llvm.memset to the
192 // new format that allows overloading the pointer for different address
193 // space (e.g., llvm.memcpy.i16 => llvm.memcpy.p0i8.p0i8.i16)
194 const char* NewFnName = NULL;
195 if (Name.compare(5,8,"memcpy.i",8) == 0) {
197 NewFnName = "llvm.memcpy.p0i8.p0i8.i8";
198 else if (Name.compare(13,2,"16") == 0)
199 NewFnName = "llvm.memcpy.p0i8.p0i8.i16";
200 else if (Name.compare(13,2,"32") == 0)
201 NewFnName = "llvm.memcpy.p0i8.p0i8.i32";
202 else if (Name.compare(13,2,"64") == 0)
203 NewFnName = "llvm.memcpy.p0i8.p0i8.i64";
204 } else if (Name.compare(5,9,"memmove.i",9) == 0) {
206 NewFnName = "llvm.memmove.p0i8.p0i8.i8";
207 else if (Name.compare(14,2,"16") == 0)
208 NewFnName = "llvm.memmove.p0i8.p0i8.i16";
209 else if (Name.compare(14,2,"32") == 0)
210 NewFnName = "llvm.memmove.p0i8.p0i8.i32";
211 else if (Name.compare(14,2,"64") == 0)
212 NewFnName = "llvm.memmove.p0i8.p0i8.i64";
214 else if (Name.compare(5,8,"memset.i",8) == 0) {
216 NewFnName = "llvm.memset.p0i8.i8";
217 else if (Name.compare(13,2,"16") == 0)
218 NewFnName = "llvm.memset.p0i8.i16";
219 else if (Name.compare(13,2,"32") == 0)
220 NewFnName = "llvm.memset.p0i8.i32";
221 else if (Name.compare(13,2,"64") == 0)
222 NewFnName = "llvm.memset.p0i8.i64";
225 NewFn = cast<Function>(M->getOrInsertFunction(NewFnName,
226 FTy->getReturnType(),
227 FTy->getParamType(0),
228 FTy->getParamType(1),
229 FTy->getParamType(2),
230 FTy->getParamType(3),
231 Type::getInt1Ty(F->getContext()),
238 // This upgrades the llvm.part.select overloaded intrinsic names to only
239 // use one type specifier in the name. We only care about the old format
240 // 'llvm.part.select.i*.i*', and solve as above with bswap.
241 if (Name.compare(5,12,"part.select.",12) == 0) {
242 std::string::size_type delim = Name.find('.',17);
244 if (delim != std::string::npos) {
245 // Construct a new name as 'llvm.part.select' + '.i*'
246 F->setName(Name.substr(0,16)+Name.substr(delim));
253 // This upgrades the llvm.part.set intrinsics similarly as above, however
254 // we care about 'llvm.part.set.i*.i*.i*', but only the first two types
255 // must match. There is an additional type specifier after these two
256 // matching types that we must retain when upgrading. Thus, we require
257 // finding 2 periods, not just one, after the intrinsic name.
258 if (Name.compare(5,9,"part.set.",9) == 0) {
259 std::string::size_type delim = Name.find('.',14);
261 if (delim != std::string::npos &&
262 Name.find('.',delim+1) != std::string::npos) {
263 // Construct a new name as 'llvm.part.select' + '.i*.i*'
264 F->setName(Name.substr(0,13)+Name.substr(delim));
273 // This fixes all MMX shift intrinsic instructions to take a
274 // v1i64 instead of a v2i32 as the second parameter.
275 if (Name.compare(5,10,"x86.mmx.ps",10) == 0 &&
276 (Name.compare(13,4,"psll", 4) == 0 ||
277 Name.compare(13,4,"psra", 4) == 0 ||
278 Name.compare(13,4,"psrl", 4) == 0) && Name[17] != 'i') {
280 const llvm::Type *VT =
281 VectorType::get(IntegerType::get(FTy->getContext(), 64), 1);
283 // We don't have to do anything if the parameter already has
285 if (FTy->getParamType(1) == VT)
288 // We first need to change the name of the old (bad) intrinsic, because
289 // its type is incorrect, but we cannot overload that name. We
290 // arbitrarily unique it here allowing us to construct a correctly named
291 // and typed function below.
294 assert(FTy->getNumParams() == 2 && "MMX shift intrinsics take 2 args!");
296 // Now construct the new intrinsic with the correct name and type. We
297 // leave the old function around in order to query its type, whatever it
298 // may be, and correctly convert up to the new type.
299 NewFn = cast<Function>(M->getOrInsertFunction(Name,
300 FTy->getReturnType(),
301 FTy->getParamType(0),
305 } else if (Name.compare(5,17,"x86.sse2.loadh.pd",17) == 0 ||
306 Name.compare(5,17,"x86.sse2.loadl.pd",17) == 0 ||
307 Name.compare(5,16,"x86.sse2.movl.dq",16) == 0 ||
308 Name.compare(5,15,"x86.sse2.movs.d",15) == 0 ||
309 Name.compare(5,16,"x86.sse2.shuf.pd",16) == 0 ||
310 Name.compare(5,18,"x86.sse2.unpckh.pd",18) == 0 ||
311 Name.compare(5,18,"x86.sse2.unpckl.pd",18) == 0 ||
312 Name.compare(5,20,"x86.sse2.punpckh.qdq",20) == 0 ||
313 Name.compare(5,20,"x86.sse2.punpckl.qdq",20) == 0) {
314 // Calls to these intrinsics are transformed into ShuffleVector's.
317 } else if (Name.compare(5, 16, "x86.sse41.pmulld", 16) == 0) {
318 // Calls to these intrinsics are transformed into vector multiplies.
321 } else if (Name.compare(5, 18, "x86.ssse3.palign.r", 18) == 0 ||
322 Name.compare(5, 22, "x86.ssse3.palign.r.128", 22) == 0) {
323 // Calls to these intrinsics are transformed into vector shuffles, shifts,
332 // This may not belong here. This function is effectively being overloaded
333 // to both detect an intrinsic which needs upgrading, and to provide the
334 // upgraded form of the intrinsic. We should perhaps have two separate
335 // functions for this.
339 bool llvm::UpgradeIntrinsicFunction(Function *F, Function *&NewFn) {
341 bool Upgraded = UpgradeIntrinsicFunction1(F, NewFn);
343 // Upgrade intrinsic attributes. This does not change the function.
346 if (unsigned id = F->getIntrinsicID())
347 F->setAttributes(Intrinsic::getAttributes((Intrinsic::ID)id));
351 // UpgradeIntrinsicCall - Upgrade a call to an old intrinsic to be a call the
352 // upgraded intrinsic. All argument and return casting must be provided in
353 // order to seamlessly integrate with existing context.
354 void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
355 Function *F = CI->getCalledFunction();
356 LLVMContext &C = CI->getContext();
357 ImmutableCallSite CS(CI);
359 assert(F && "CallInst has no function associated with it.");
362 // Get the Function's name.
363 const std::string& Name = F->getName();
365 // Upgrade ARM NEON intrinsics.
366 if (Name.compare(5, 9, "arm.neon.", 9) == 0) {
368 if (Name.compare(14, 7, "vmovls.", 7) == 0) {
369 NewI = new SExtInst(CI->getArgOperand(0), CI->getType(),
370 "upgraded." + CI->getName(), CI);
371 } else if (Name.compare(14, 7, "vmovlu.", 7) == 0) {
372 NewI = new ZExtInst(CI->getArgOperand(0), CI->getType(),
373 "upgraded." + CI->getName(), CI);
375 llvm_unreachable("Unknown arm.neon function for CallInst upgrade.");
377 // Replace any uses of the old CallInst.
378 if (!CI->use_empty())
379 CI->replaceAllUsesWith(NewI);
380 CI->eraseFromParent();
384 bool isLoadH = false, isLoadL = false, isMovL = false;
385 bool isMovSD = false, isShufPD = false;
386 bool isUnpckhPD = false, isUnpcklPD = false;
387 bool isPunpckhQPD = false, isPunpcklQPD = false;
388 if (F->getName() == "llvm.x86.sse2.loadh.pd")
390 else if (F->getName() == "llvm.x86.sse2.loadl.pd")
392 else if (F->getName() == "llvm.x86.sse2.movl.dq")
394 else if (F->getName() == "llvm.x86.sse2.movs.d")
396 else if (F->getName() == "llvm.x86.sse2.shuf.pd")
398 else if (F->getName() == "llvm.x86.sse2.unpckh.pd")
400 else if (F->getName() == "llvm.x86.sse2.unpckl.pd")
402 else if (F->getName() == "llvm.x86.sse2.punpckh.qdq")
404 else if (F->getName() == "llvm.x86.sse2.punpckl.qdq")
407 if (isLoadH || isLoadL || isMovL || isMovSD || isShufPD ||
408 isUnpckhPD || isUnpcklPD || isPunpckhQPD || isPunpcklQPD) {
409 std::vector<Constant*> Idxs;
410 Value *Op0 = CI->getArgOperand(0);
411 ShuffleVectorInst *SI = NULL;
412 if (isLoadH || isLoadL) {
413 Value *Op1 = UndefValue::get(Op0->getType());
414 Value *Addr = new BitCastInst(CI->getArgOperand(1),
415 Type::getDoublePtrTy(C),
417 Value *Load = new LoadInst(Addr, "upgraded.", false, 8, CI);
418 Value *Idx = ConstantInt::get(Type::getInt32Ty(C), 0);
419 Op1 = InsertElementInst::Create(Op1, Load, Idx, "upgraded.", CI);
422 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 0));
423 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
425 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
426 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
428 Value *Mask = ConstantVector::get(Idxs);
429 SI = new ShuffleVectorInst(Op0, Op1, Mask, "upgraded.", CI);
431 Constant *Zero = ConstantInt::get(Type::getInt32Ty(C), 0);
432 Idxs.push_back(Zero);
433 Idxs.push_back(Zero);
434 Idxs.push_back(Zero);
435 Idxs.push_back(Zero);
436 Value *ZeroV = ConstantVector::get(Idxs);
439 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 4));
440 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 5));
441 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
442 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 3));
443 Value *Mask = ConstantVector::get(Idxs);
444 SI = new ShuffleVectorInst(ZeroV, Op0, Mask, "upgraded.", CI);
445 } else if (isMovSD ||
446 isUnpckhPD || isUnpcklPD || isPunpckhQPD || isPunpcklQPD) {
447 Value *Op1 = CI->getArgOperand(1);
449 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
450 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
451 } else if (isUnpckhPD || isPunpckhQPD) {
452 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
453 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 3));
455 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 0));
456 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
458 Value *Mask = ConstantVector::get(Idxs);
459 SI = new ShuffleVectorInst(Op0, Op1, Mask, "upgraded.", CI);
460 } else if (isShufPD) {
461 Value *Op1 = CI->getArgOperand(1);
463 cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
464 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), MaskVal & 1));
465 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C),
466 ((MaskVal >> 1) & 1)+2));
467 Value *Mask = ConstantVector::get(Idxs);
468 SI = new ShuffleVectorInst(Op0, Op1, Mask, "upgraded.", CI);
471 assert(SI && "Unexpected!");
473 // Handle any uses of the old CallInst.
474 if (!CI->use_empty())
475 // Replace all uses of the old call with the new cast which has the
477 CI->replaceAllUsesWith(SI);
479 // Clean up the old call now that it has been completely upgraded.
480 CI->eraseFromParent();
481 } else if (F->getName() == "llvm.x86.sse41.pmulld") {
482 // Upgrade this set of intrinsics into vector multiplies.
483 Instruction *Mul = BinaryOperator::CreateMul(CI->getArgOperand(0),
484 CI->getArgOperand(1),
487 // Fix up all the uses with our new multiply.
488 if (!CI->use_empty())
489 CI->replaceAllUsesWith(Mul);
491 // Remove upgraded multiply.
492 CI->eraseFromParent();
493 } else if (F->getName() == "llvm.x86.ssse3.palign.r") {
494 Value *Op1 = CI->getArgOperand(0);
495 Value *Op2 = CI->getArgOperand(1);
496 Value *Op3 = CI->getArgOperand(2);
497 unsigned shiftVal = cast<ConstantInt>(Op3)->getZExtValue();
499 IRBuilder<> Builder(C);
500 Builder.SetInsertPoint(CI->getParent(), CI);
502 // If palignr is shifting the pair of input vectors less than 9 bytes,
503 // emit a shuffle instruction.
505 const Type *IntTy = Type::getInt32Ty(C);
506 const Type *EltTy = Type::getInt8Ty(C);
507 const Type *VecTy = VectorType::get(EltTy, 8);
509 Op2 = Builder.CreateBitCast(Op2, VecTy);
510 Op1 = Builder.CreateBitCast(Op1, VecTy);
512 llvm::SmallVector<llvm::Constant*, 8> Indices;
513 for (unsigned i = 0; i != 8; ++i)
514 Indices.push_back(ConstantInt::get(IntTy, shiftVal + i));
516 Value *SV = ConstantVector::get(Indices.begin(), Indices.size());
517 Rep = Builder.CreateShuffleVector(Op2, Op1, SV, "palignr");
518 Rep = Builder.CreateBitCast(Rep, F->getReturnType());
521 // If palignr is shifting the pair of input vectors more than 8 but less
522 // than 16 bytes, emit a logical right shift of the destination.
523 else if (shiftVal < 16) {
524 // MMX has these as 1 x i64 vectors for some odd optimization reasons.
525 const Type *EltTy = Type::getInt64Ty(C);
526 const Type *VecTy = VectorType::get(EltTy, 1);
528 Op1 = Builder.CreateBitCast(Op1, VecTy, "cast");
529 Op2 = ConstantInt::get(VecTy, (shiftVal-8) * 8);
531 // create i32 constant
533 Intrinsic::getDeclaration(F->getParent(), Intrinsic::x86_mmx_psrl_q);
534 Rep = Builder.CreateCall2(I, Op1, Op2, "palignr");
537 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
539 Rep = Constant::getNullValue(F->getReturnType());
542 // Replace any uses with our new instruction.
543 if (!CI->use_empty())
544 CI->replaceAllUsesWith(Rep);
546 // Remove upgraded instruction.
547 CI->eraseFromParent();
549 } else if (F->getName() == "llvm.x86.ssse3.palign.r.128") {
550 Value *Op1 = CI->getArgOperand(0);
551 Value *Op2 = CI->getArgOperand(1);
552 Value *Op3 = CI->getArgOperand(2);
553 unsigned shiftVal = cast<ConstantInt>(Op3)->getZExtValue();
555 IRBuilder<> Builder(C);
556 Builder.SetInsertPoint(CI->getParent(), CI);
558 // If palignr is shifting the pair of input vectors less than 17 bytes,
559 // emit a shuffle instruction.
560 if (shiftVal <= 16) {
561 const Type *IntTy = Type::getInt32Ty(C);
562 const Type *EltTy = Type::getInt8Ty(C);
563 const Type *VecTy = VectorType::get(EltTy, 16);
565 Op2 = Builder.CreateBitCast(Op2, VecTy);
566 Op1 = Builder.CreateBitCast(Op1, VecTy);
568 llvm::SmallVector<llvm::Constant*, 16> Indices;
569 for (unsigned i = 0; i != 16; ++i)
570 Indices.push_back(ConstantInt::get(IntTy, shiftVal + i));
572 Value *SV = ConstantVector::get(Indices.begin(), Indices.size());
573 Rep = Builder.CreateShuffleVector(Op2, Op1, SV, "palignr");
574 Rep = Builder.CreateBitCast(Rep, F->getReturnType());
577 // If palignr is shifting the pair of input vectors more than 16 but less
578 // than 32 bytes, emit a logical right shift of the destination.
579 else if (shiftVal < 32) {
580 const Type *EltTy = Type::getInt64Ty(C);
581 const Type *VecTy = VectorType::get(EltTy, 2);
582 const Type *IntTy = Type::getInt32Ty(C);
584 Op1 = Builder.CreateBitCast(Op1, VecTy, "cast");
585 Op2 = ConstantInt::get(IntTy, (shiftVal-16) * 8);
587 // create i32 constant
589 Intrinsic::getDeclaration(F->getParent(), Intrinsic::x86_sse2_psrl_dq);
590 Rep = Builder.CreateCall2(I, Op1, Op2, "palignr");
593 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
595 Rep = Constant::getNullValue(F->getReturnType());
598 // Replace any uses with our new instruction.
599 if (!CI->use_empty())
600 CI->replaceAllUsesWith(Rep);
602 // Remove upgraded instruction.
603 CI->eraseFromParent();
606 llvm_unreachable("Unknown function for CallInst upgrade.");
611 switch (NewFn->getIntrinsicID()) {
612 default: llvm_unreachable("Unknown function for CallInst upgrade.");
613 case Intrinsic::arm_neon_vld1:
614 case Intrinsic::arm_neon_vld2:
615 case Intrinsic::arm_neon_vld3:
616 case Intrinsic::arm_neon_vld4:
617 case Intrinsic::arm_neon_vst1:
618 case Intrinsic::arm_neon_vst2:
619 case Intrinsic::arm_neon_vst3:
620 case Intrinsic::arm_neon_vst4:
621 case Intrinsic::arm_neon_vld2lane:
622 case Intrinsic::arm_neon_vld3lane:
623 case Intrinsic::arm_neon_vld4lane:
624 case Intrinsic::arm_neon_vst2lane:
625 case Intrinsic::arm_neon_vst3lane:
626 case Intrinsic::arm_neon_vst4lane: {
627 // Add a default alignment argument of 1.
628 SmallVector<Value*, 8> Operands(CS.arg_begin(), CS.arg_end());
629 Operands.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
630 CallInst *NewCI = CallInst::Create(NewFn, Operands.begin(), Operands.end(),
632 NewCI->setTailCall(CI->isTailCall());
633 NewCI->setCallingConv(CI->getCallingConv());
635 // Handle any uses of the old CallInst.
636 if (!CI->use_empty())
637 // Replace all uses of the old call with the new cast which has the
639 CI->replaceAllUsesWith(NewCI);
641 // Clean up the old call now that it has been completely upgraded.
642 CI->eraseFromParent();
646 case Intrinsic::x86_mmx_psll_d:
647 case Intrinsic::x86_mmx_psll_q:
648 case Intrinsic::x86_mmx_psll_w:
649 case Intrinsic::x86_mmx_psra_d:
650 case Intrinsic::x86_mmx_psra_w:
651 case Intrinsic::x86_mmx_psrl_d:
652 case Intrinsic::x86_mmx_psrl_q:
653 case Intrinsic::x86_mmx_psrl_w: {
656 Operands[0] = CI->getArgOperand(0);
658 // Cast the second parameter to the correct type.
659 BitCastInst *BC = new BitCastInst(CI->getArgOperand(1),
660 NewFn->getFunctionType()->getParamType(1),
664 // Construct a new CallInst
665 CallInst *NewCI = CallInst::Create(NewFn, Operands, Operands+2,
666 "upgraded."+CI->getName(), CI);
667 NewCI->setTailCall(CI->isTailCall());
668 NewCI->setCallingConv(CI->getCallingConv());
670 // Handle any uses of the old CallInst.
671 if (!CI->use_empty())
672 // Replace all uses of the old call with the new cast which has the
674 CI->replaceAllUsesWith(NewCI);
676 // Clean up the old call now that it has been completely upgraded.
677 CI->eraseFromParent();
680 case Intrinsic::ctlz:
681 case Intrinsic::ctpop:
682 case Intrinsic::cttz: {
683 // Build a small vector of the original arguments.
684 SmallVector<Value*, 8> Operands(CS.arg_begin(), CS.arg_end());
686 // Construct a new CallInst
687 CallInst *NewCI = CallInst::Create(NewFn, Operands.begin(), Operands.end(),
688 "upgraded."+CI->getName(), CI);
689 NewCI->setTailCall(CI->isTailCall());
690 NewCI->setCallingConv(CI->getCallingConv());
692 // Handle any uses of the old CallInst.
693 if (!CI->use_empty()) {
694 // Check for sign extend parameter attributes on the return values.
695 bool SrcSExt = NewFn->getAttributes().paramHasAttr(0, Attribute::SExt);
696 bool DestSExt = F->getAttributes().paramHasAttr(0, Attribute::SExt);
698 // Construct an appropriate cast from the new return type to the old.
699 CastInst *RetCast = CastInst::Create(
700 CastInst::getCastOpcode(NewCI, SrcSExt,
703 NewCI, F->getReturnType(),
704 NewCI->getName(), CI);
705 NewCI->moveBefore(RetCast);
707 // Replace all uses of the old call with the new cast which has the
709 CI->replaceAllUsesWith(RetCast);
712 // Clean up the old call now that it has been completely upgraded.
713 CI->eraseFromParent();
716 case Intrinsic::eh_selector:
717 case Intrinsic::eh_typeid_for: {
718 // Only the return type changed.
719 SmallVector<Value*, 8> Operands(CS.arg_begin(), CS.arg_end());
720 CallInst *NewCI = CallInst::Create(NewFn, Operands.begin(), Operands.end(),
721 "upgraded." + CI->getName(), CI);
722 NewCI->setTailCall(CI->isTailCall());
723 NewCI->setCallingConv(CI->getCallingConv());
725 // Handle any uses of the old CallInst.
726 if (!CI->use_empty()) {
727 // Construct an appropriate cast from the new return type to the old.
729 CastInst::Create(CastInst::getCastOpcode(NewCI, true,
730 F->getReturnType(), true),
731 NewCI, F->getReturnType(), NewCI->getName(), CI);
732 CI->replaceAllUsesWith(RetCast);
734 CI->eraseFromParent();
737 case Intrinsic::memcpy:
738 case Intrinsic::memmove:
739 case Intrinsic::memset: {
741 const llvm::Type *I1Ty = llvm::Type::getInt1Ty(CI->getContext());
742 Value *Operands[5] = { CI->getArgOperand(0), CI->getArgOperand(1),
743 CI->getArgOperand(2), CI->getArgOperand(3),
744 llvm::ConstantInt::get(I1Ty, 0) };
745 CallInst *NewCI = CallInst::Create(NewFn, Operands, Operands+5,
747 NewCI->setTailCall(CI->isTailCall());
748 NewCI->setCallingConv(CI->getCallingConv());
749 // Handle any uses of the old CallInst.
750 if (!CI->use_empty())
751 // Replace all uses of the old call with the new cast which has the
753 CI->replaceAllUsesWith(NewCI);
755 // Clean up the old call now that it has been completely upgraded.
756 CI->eraseFromParent();
762 // This tests each Function to determine if it needs upgrading. When we find
763 // one we are interested in, we then upgrade all calls to reflect the new
765 void llvm::UpgradeCallsToIntrinsic(Function* F) {
766 assert(F && "Illegal attempt to upgrade a non-existent intrinsic.");
768 // Upgrade the function and check if it is a totaly new function.
770 if (UpgradeIntrinsicFunction(F, NewFn)) {
772 // Replace all uses to the old function with the new one if necessary.
773 for (Value::use_iterator UI = F->use_begin(), UE = F->use_end();
775 if (CallInst* CI = dyn_cast<CallInst>(*UI++))
776 UpgradeIntrinsicCall(CI, NewFn);
778 // Remove old function, no longer used, from the module.
779 F->eraseFromParent();
784 /// This function strips all debug info intrinsics, except for llvm.dbg.declare.
785 /// If an llvm.dbg.declare intrinsic is invalid, then this function simply
787 void llvm::CheckDebugInfoIntrinsics(Module *M) {
790 if (Function *FuncStart = M->getFunction("llvm.dbg.func.start")) {
791 while (!FuncStart->use_empty()) {
792 CallInst *CI = cast<CallInst>(FuncStart->use_back());
793 CI->eraseFromParent();
795 FuncStart->eraseFromParent();
798 if (Function *StopPoint = M->getFunction("llvm.dbg.stoppoint")) {
799 while (!StopPoint->use_empty()) {
800 CallInst *CI = cast<CallInst>(StopPoint->use_back());
801 CI->eraseFromParent();
803 StopPoint->eraseFromParent();
806 if (Function *RegionStart = M->getFunction("llvm.dbg.region.start")) {
807 while (!RegionStart->use_empty()) {
808 CallInst *CI = cast<CallInst>(RegionStart->use_back());
809 CI->eraseFromParent();
811 RegionStart->eraseFromParent();
814 if (Function *RegionEnd = M->getFunction("llvm.dbg.region.end")) {
815 while (!RegionEnd->use_empty()) {
816 CallInst *CI = cast<CallInst>(RegionEnd->use_back());
817 CI->eraseFromParent();
819 RegionEnd->eraseFromParent();
822 if (Function *Declare = M->getFunction("llvm.dbg.declare")) {
823 if (!Declare->use_empty()) {
824 DbgDeclareInst *DDI = cast<DbgDeclareInst>(Declare->use_back());
825 if (!isa<MDNode>(DDI->getArgOperand(0)) ||
826 !isa<MDNode>(DDI->getArgOperand(1))) {
827 while (!Declare->use_empty()) {
828 CallInst *CI = cast<CallInst>(Declare->use_back());
829 CI->eraseFromParent();
831 Declare->eraseFromParent();