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, 5, "vmovl", 5) == 0 ||
83 Name.compare(14, 5, "vaddl", 5) == 0 ||
84 Name.compare(14, 5, "vsubl", 5) == 0 ||
85 Name.compare(14, 5, "vaddw", 5) == 0 ||
86 Name.compare(14, 5, "vsubw", 5) == 0 ||
87 Name.compare(14, 5, "vmull", 5) == 0 ||
88 Name.compare(14, 5, "vmlal", 5) == 0 ||
89 Name.compare(14, 5, "vmlsl", 5) == 0 ||
90 Name.compare(14, 5, "vabdl", 5) == 0 ||
91 Name.compare(14, 5, "vabal", 5) == 0) &&
92 (Name.compare(19, 2, "s.", 2) == 0 ||
93 Name.compare(19, 2, "u.", 2) == 0)) ||
95 (Name.compare(14, 4, "vaba", 4) == 0 &&
96 (Name.compare(18, 2, "s.", 2) == 0 ||
97 Name.compare(18, 2, "u.", 2) == 0)) ||
99 (Name.compare(14, 6, "vmovn.", 6) == 0)) {
101 // Calls to these are transformed into IR without intrinsics.
105 // Old versions of NEON ld/st intrinsics are missing alignment arguments.
106 bool isVLd = (Name.compare(14, 3, "vld", 3) == 0);
107 bool isVSt = (Name.compare(14, 3, "vst", 3) == 0);
108 if (isVLd || isVSt) {
109 unsigned NumVecs = Name.at(17) - '0';
110 if (NumVecs == 0 || NumVecs > 4)
112 bool isLaneOp = (Name.compare(18, 5, "lane.", 5) == 0);
113 if (!isLaneOp && Name.at(18) != '.')
115 unsigned ExpectedArgs = 2; // for the address and alignment
116 if (isVSt || isLaneOp)
117 ExpectedArgs += NumVecs;
119 ExpectedArgs += 1; // for the lane number
120 unsigned NumP = FTy->getNumParams();
121 if (NumP != ExpectedArgs - 1)
124 // Change the name of the old (bad) intrinsic, because
125 // its type is incorrect, but we cannot overload that name.
128 // One argument is missing: add the alignment argument.
129 std::vector<const Type*> NewParams;
130 for (unsigned p = 0; p < NumP; ++p)
131 NewParams.push_back(FTy->getParamType(p));
132 NewParams.push_back(Type::getInt32Ty(F->getContext()));
133 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(),
135 NewFn = cast<Function>(M->getOrInsertFunction(Name, NewFTy));
141 // This upgrades the name of the llvm.bswap intrinsic function to only use
142 // a single type name for overloading. We only care about the old format
143 // 'llvm.bswap.i*.i*', so check for 'bswap.' and then for there being
144 // a '.' after 'bswap.'
145 if (Name.compare(5,6,"bswap.",6) == 0) {
146 std::string::size_type delim = Name.find('.',11);
148 if (delim != std::string::npos) {
149 // Construct the new name as 'llvm.bswap' + '.i*'
150 F->setName(Name.substr(0,10)+Name.substr(delim));
158 // We only want to fix the 'llvm.ct*' intrinsics which do not have the
159 // correct return type, so we check for the name, and then check if the
160 // return type does not match the parameter type.
161 if ( (Name.compare(5,5,"ctpop",5) == 0 ||
162 Name.compare(5,4,"ctlz",4) == 0 ||
163 Name.compare(5,4,"cttz",4) == 0) &&
164 FTy->getReturnType() != FTy->getParamType(0)) {
165 // We first need to change the name of the old (bad) intrinsic, because
166 // its type is incorrect, but we cannot overload that name. We
167 // arbitrarily unique it here allowing us to construct a correctly named
168 // and typed function below.
171 // Now construct the new intrinsic with the correct name and type. We
172 // leave the old function around in order to query its type, whatever it
173 // may be, and correctly convert up to the new type.
174 NewFn = cast<Function>(M->getOrInsertFunction(Name,
175 FTy->getParamType(0),
176 FTy->getParamType(0),
183 // The old llvm.eh.selector.i32 is equivalent to the new llvm.eh.selector.
184 if (Name.compare("llvm.eh.selector.i32") == 0) {
185 F->setName("llvm.eh.selector");
189 // The old llvm.eh.typeid.for.i32 is equivalent to llvm.eh.typeid.for.
190 if (Name.compare("llvm.eh.typeid.for.i32") == 0) {
191 F->setName("llvm.eh.typeid.for");
195 // Convert the old llvm.eh.selector.i64 to a call to llvm.eh.selector.
196 if (Name.compare("llvm.eh.selector.i64") == 0) {
197 NewFn = Intrinsic::getDeclaration(M, Intrinsic::eh_selector);
200 // Convert the old llvm.eh.typeid.for.i64 to a call to llvm.eh.typeid.for.
201 if (Name.compare("llvm.eh.typeid.for.i64") == 0) {
202 NewFn = Intrinsic::getDeclaration(M, Intrinsic::eh_typeid_for);
208 // This upgrades the llvm.memcpy, llvm.memmove, and llvm.memset to the
209 // new format that allows overloading the pointer for different address
210 // space (e.g., llvm.memcpy.i16 => llvm.memcpy.p0i8.p0i8.i16)
211 const char* NewFnName = NULL;
212 if (Name.compare(5,8,"memcpy.i",8) == 0) {
214 NewFnName = "llvm.memcpy.p0i8.p0i8.i8";
215 else if (Name.compare(13,2,"16") == 0)
216 NewFnName = "llvm.memcpy.p0i8.p0i8.i16";
217 else if (Name.compare(13,2,"32") == 0)
218 NewFnName = "llvm.memcpy.p0i8.p0i8.i32";
219 else if (Name.compare(13,2,"64") == 0)
220 NewFnName = "llvm.memcpy.p0i8.p0i8.i64";
221 } else if (Name.compare(5,9,"memmove.i",9) == 0) {
223 NewFnName = "llvm.memmove.p0i8.p0i8.i8";
224 else if (Name.compare(14,2,"16") == 0)
225 NewFnName = "llvm.memmove.p0i8.p0i8.i16";
226 else if (Name.compare(14,2,"32") == 0)
227 NewFnName = "llvm.memmove.p0i8.p0i8.i32";
228 else if (Name.compare(14,2,"64") == 0)
229 NewFnName = "llvm.memmove.p0i8.p0i8.i64";
231 else if (Name.compare(5,8,"memset.i",8) == 0) {
233 NewFnName = "llvm.memset.p0i8.i8";
234 else if (Name.compare(13,2,"16") == 0)
235 NewFnName = "llvm.memset.p0i8.i16";
236 else if (Name.compare(13,2,"32") == 0)
237 NewFnName = "llvm.memset.p0i8.i32";
238 else if (Name.compare(13,2,"64") == 0)
239 NewFnName = "llvm.memset.p0i8.i64";
242 NewFn = cast<Function>(M->getOrInsertFunction(NewFnName,
243 FTy->getReturnType(),
244 FTy->getParamType(0),
245 FTy->getParamType(1),
246 FTy->getParamType(2),
247 FTy->getParamType(3),
248 Type::getInt1Ty(F->getContext()),
255 // This upgrades the llvm.part.select overloaded intrinsic names to only
256 // use one type specifier in the name. We only care about the old format
257 // 'llvm.part.select.i*.i*', and solve as above with bswap.
258 if (Name.compare(5,12,"part.select.",12) == 0) {
259 std::string::size_type delim = Name.find('.',17);
261 if (delim != std::string::npos) {
262 // Construct a new name as 'llvm.part.select' + '.i*'
263 F->setName(Name.substr(0,16)+Name.substr(delim));
270 // This upgrades the llvm.part.set intrinsics similarly as above, however
271 // we care about 'llvm.part.set.i*.i*.i*', but only the first two types
272 // must match. There is an additional type specifier after these two
273 // matching types that we must retain when upgrading. Thus, we require
274 // finding 2 periods, not just one, after the intrinsic name.
275 if (Name.compare(5,9,"part.set.",9) == 0) {
276 std::string::size_type delim = Name.find('.',14);
278 if (delim != std::string::npos &&
279 Name.find('.',delim+1) != std::string::npos) {
280 // Construct a new name as 'llvm.part.select' + '.i*.i*'
281 F->setName(Name.substr(0,13)+Name.substr(delim));
290 // This fixes all MMX shift intrinsic instructions to take a
291 // v1i64 instead of a v2i32 as the second parameter.
292 if (Name.compare(5,10,"x86.mmx.ps",10) == 0 &&
293 (Name.compare(13,4,"psll", 4) == 0 ||
294 Name.compare(13,4,"psra", 4) == 0 ||
295 Name.compare(13,4,"psrl", 4) == 0) && Name[17] != 'i') {
297 const llvm::Type *VT =
298 VectorType::get(IntegerType::get(FTy->getContext(), 64), 1);
300 // We don't have to do anything if the parameter already has
302 if (FTy->getParamType(1) == VT)
305 // We first need to change the name of the old (bad) intrinsic, because
306 // its type is incorrect, but we cannot overload that name. We
307 // arbitrarily unique it here allowing us to construct a correctly named
308 // and typed function below.
311 assert(FTy->getNumParams() == 2 && "MMX shift intrinsics take 2 args!");
313 // Now construct the new intrinsic with the correct name and type. We
314 // leave the old function around in order to query its type, whatever it
315 // may be, and correctly convert up to the new type.
316 NewFn = cast<Function>(M->getOrInsertFunction(Name,
317 FTy->getReturnType(),
318 FTy->getParamType(0),
322 } else if (Name.compare(5,17,"x86.sse2.loadh.pd",17) == 0 ||
323 Name.compare(5,17,"x86.sse2.loadl.pd",17) == 0 ||
324 Name.compare(5,16,"x86.sse2.movl.dq",16) == 0 ||
325 Name.compare(5,15,"x86.sse2.movs.d",15) == 0 ||
326 Name.compare(5,16,"x86.sse2.shuf.pd",16) == 0 ||
327 Name.compare(5,18,"x86.sse2.unpckh.pd",18) == 0 ||
328 Name.compare(5,18,"x86.sse2.unpckl.pd",18) == 0 ||
329 Name.compare(5,20,"x86.sse2.punpckh.qdq",20) == 0 ||
330 Name.compare(5,20,"x86.sse2.punpckl.qdq",20) == 0) {
331 // Calls to these intrinsics are transformed into ShuffleVector's.
334 } else if (Name.compare(5, 16, "x86.sse41.pmulld", 16) == 0) {
335 // Calls to these intrinsics are transformed into vector multiplies.
338 } else if (Name.compare(5, 18, "x86.ssse3.palign.r", 18) == 0 ||
339 Name.compare(5, 22, "x86.ssse3.palign.r.128", 22) == 0) {
340 // Calls to these intrinsics are transformed into vector shuffles, shifts,
349 // This may not belong here. This function is effectively being overloaded
350 // to both detect an intrinsic which needs upgrading, and to provide the
351 // upgraded form of the intrinsic. We should perhaps have two separate
352 // functions for this.
356 bool llvm::UpgradeIntrinsicFunction(Function *F, Function *&NewFn) {
358 bool Upgraded = UpgradeIntrinsicFunction1(F, NewFn);
360 // Upgrade intrinsic attributes. This does not change the function.
363 if (unsigned id = F->getIntrinsicID())
364 F->setAttributes(Intrinsic::getAttributes((Intrinsic::ID)id));
368 bool llvm::UpgradeGlobalVariable(GlobalVariable *GV) {
369 StringRef Name(GV->getName());
371 // We are only upgrading one symbol here.
372 if (Name == ".llvm.eh.catch.all.value") {
373 GV->setName("llvm.eh.catch.all.value");
380 /// ExtendNEONArgs - For NEON "long" and "wide" operations, where the results
381 /// have vector elements twice as big as one or both source operands, do the
382 /// sign- or zero-extension that used to be handled by intrinsics. The
383 /// extended values are returned via V0 and V1.
384 static void ExtendNEONArgs(CallInst *CI, Value *Arg0, Value *Arg1,
385 Value *&V0, Value *&V1) {
386 Function *F = CI->getCalledFunction();
387 const std::string& Name = F->getName();
388 bool isLong = (Name.at(18) == 'l');
389 bool isSigned = (Name.at(19) == 's');
393 V0 = new SExtInst(Arg0, CI->getType(), "", CI);
396 V1 = new SExtInst(Arg1, CI->getType(), "", CI);
399 V0 = new ZExtInst(Arg0, CI->getType(), "", CI);
402 V1 = new ZExtInst(Arg1, CI->getType(), "", CI);
406 /// CallVABD - As part of expanding a call to one of the old NEON vabdl, vaba,
407 /// or vabal intrinsics, construct a call to a vabd intrinsic. Examine the
408 /// name of the old intrinsic to determine whether to use a signed or unsigned
409 /// vabd intrinsic. Get the type from the old call instruction, adjusted for
410 /// half-size vector elements if the old intrinsic was vabdl or vabal.
411 static Instruction *CallVABD(CallInst *CI, Value *Arg0, Value *Arg1) {
412 Function *F = CI->getCalledFunction();
413 const std::string& Name = F->getName();
414 bool isLong = (Name.at(18) == 'l');
415 bool isSigned = (Name.at(isLong ? 19 : 18) == 's');
419 intID = Intrinsic::arm_neon_vabds;
421 intID = Intrinsic::arm_neon_vabdu;
423 const Type *Ty = CI->getType();
425 Ty = VectorType::getTruncatedElementVectorType(cast<const VectorType>(Ty));
427 Function *VABD = Intrinsic::getDeclaration(F->getParent(), intID, &Ty, 1);
431 return CallInst::Create(VABD, Operands, Operands+2,
432 "upgraded."+CI->getName(), CI);
435 // UpgradeIntrinsicCall - Upgrade a call to an old intrinsic to be a call the
436 // upgraded intrinsic. All argument and return casting must be provided in
437 // order to seamlessly integrate with existing context.
438 void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
439 Function *F = CI->getCalledFunction();
440 LLVMContext &C = CI->getContext();
441 ImmutableCallSite CS(CI);
443 assert(F && "CallInst has no function associated with it.");
446 // Get the Function's name.
447 const std::string& Name = F->getName();
449 // Upgrade ARM NEON intrinsics.
450 if (Name.compare(5, 9, "arm.neon.", 9) == 0) {
453 if (Name.compare(14, 7, "vmovls.", 7) == 0) {
454 NewI = new SExtInst(CI->getArgOperand(0), CI->getType(),
455 "upgraded." + CI->getName(), CI);
456 } else if (Name.compare(14, 7, "vmovlu.", 7) == 0) {
457 NewI = new ZExtInst(CI->getArgOperand(0), CI->getType(),
458 "upgraded." + CI->getName(), CI);
459 } else if (Name.compare(14, 4, "vadd", 4) == 0) {
460 ExtendNEONArgs(CI, CI->getArgOperand(0), CI->getArgOperand(1), V0, V1);
461 NewI = BinaryOperator::CreateAdd(V0, V1, "upgraded."+CI->getName(), CI);
462 } else if (Name.compare(14, 4, "vsub", 4) == 0) {
463 ExtendNEONArgs(CI, CI->getArgOperand(0), CI->getArgOperand(1), V0, V1);
464 NewI = BinaryOperator::CreateSub(V0, V1,"upgraded."+CI->getName(),CI);
465 } else if (Name.compare(14, 4, "vmul", 4) == 0) {
466 ExtendNEONArgs(CI, CI->getArgOperand(0), CI->getArgOperand(1), V0, V1);
467 NewI = BinaryOperator::CreateMul(V0, V1,"upgraded."+CI->getName(),CI);
468 } else if (Name.compare(14, 4, "vmla", 4) == 0) {
469 ExtendNEONArgs(CI, CI->getArgOperand(1), CI->getArgOperand(2), V0, V1);
470 Instruction *MulI = BinaryOperator::CreateMul(V0, V1, "", CI);
471 NewI = BinaryOperator::CreateAdd(CI->getArgOperand(0), MulI,
472 "upgraded."+CI->getName(), CI);
473 } else if (Name.compare(14, 4, "vmls", 4) == 0) {
474 ExtendNEONArgs(CI, CI->getArgOperand(1), CI->getArgOperand(2), V0, V1);
475 Instruction *MulI = BinaryOperator::CreateMul(V0, V1, "", CI);
476 NewI = BinaryOperator::CreateSub(CI->getArgOperand(0), MulI,
477 "upgraded."+CI->getName(), CI);
478 } else if (Name.compare(14, 4, "vabd", 4) == 0) {
479 NewI = CallVABD(CI, CI->getArgOperand(0), CI->getArgOperand(1));
480 NewI = new ZExtInst(NewI, CI->getType(), "upgraded."+CI->getName(), CI);
481 } else if (Name.compare(14, 4, "vaba", 4) == 0) {
482 NewI = CallVABD(CI, CI->getArgOperand(1), CI->getArgOperand(2));
483 if (Name.at(18) == 'l')
484 NewI = new ZExtInst(NewI, CI->getType(), "", CI);
485 NewI = BinaryOperator::CreateAdd(CI->getArgOperand(0), NewI,
486 "upgraded."+CI->getName(), CI);
487 } else if (Name.compare(14, 6, "vmovn.", 6) == 0) {
488 NewI = new TruncInst(CI->getArgOperand(0), CI->getType(),
489 "upgraded." + CI->getName(), CI);
491 llvm_unreachable("Unknown arm.neon function for CallInst upgrade.");
493 // Replace any uses of the old CallInst.
494 if (!CI->use_empty())
495 CI->replaceAllUsesWith(NewI);
496 CI->eraseFromParent();
500 bool isLoadH = false, isLoadL = false, isMovL = false;
501 bool isMovSD = false, isShufPD = false;
502 bool isUnpckhPD = false, isUnpcklPD = false;
503 bool isPunpckhQPD = false, isPunpcklQPD = false;
504 if (F->getName() == "llvm.x86.sse2.loadh.pd")
506 else if (F->getName() == "llvm.x86.sse2.loadl.pd")
508 else if (F->getName() == "llvm.x86.sse2.movl.dq")
510 else if (F->getName() == "llvm.x86.sse2.movs.d")
512 else if (F->getName() == "llvm.x86.sse2.shuf.pd")
514 else if (F->getName() == "llvm.x86.sse2.unpckh.pd")
516 else if (F->getName() == "llvm.x86.sse2.unpckl.pd")
518 else if (F->getName() == "llvm.x86.sse2.punpckh.qdq")
520 else if (F->getName() == "llvm.x86.sse2.punpckl.qdq")
523 if (isLoadH || isLoadL || isMovL || isMovSD || isShufPD ||
524 isUnpckhPD || isUnpcklPD || isPunpckhQPD || isPunpcklQPD) {
525 std::vector<Constant*> Idxs;
526 Value *Op0 = CI->getArgOperand(0);
527 ShuffleVectorInst *SI = NULL;
528 if (isLoadH || isLoadL) {
529 Value *Op1 = UndefValue::get(Op0->getType());
530 Value *Addr = new BitCastInst(CI->getArgOperand(1),
531 Type::getDoublePtrTy(C),
533 Value *Load = new LoadInst(Addr, "upgraded.", false, 8, CI);
534 Value *Idx = ConstantInt::get(Type::getInt32Ty(C), 0);
535 Op1 = InsertElementInst::Create(Op1, Load, Idx, "upgraded.", CI);
538 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 0));
539 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
541 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
542 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
544 Value *Mask = ConstantVector::get(Idxs);
545 SI = new ShuffleVectorInst(Op0, Op1, Mask, "upgraded.", CI);
547 Constant *Zero = ConstantInt::get(Type::getInt32Ty(C), 0);
548 Idxs.push_back(Zero);
549 Idxs.push_back(Zero);
550 Idxs.push_back(Zero);
551 Idxs.push_back(Zero);
552 Value *ZeroV = ConstantVector::get(Idxs);
555 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 4));
556 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 5));
557 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
558 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 3));
559 Value *Mask = ConstantVector::get(Idxs);
560 SI = new ShuffleVectorInst(ZeroV, Op0, Mask, "upgraded.", CI);
561 } else if (isMovSD ||
562 isUnpckhPD || isUnpcklPD || isPunpckhQPD || isPunpcklQPD) {
563 Value *Op1 = CI->getArgOperand(1);
565 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
566 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
567 } else if (isUnpckhPD || isPunpckhQPD) {
568 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
569 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 3));
571 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 0));
572 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
574 Value *Mask = ConstantVector::get(Idxs);
575 SI = new ShuffleVectorInst(Op0, Op1, Mask, "upgraded.", CI);
576 } else if (isShufPD) {
577 Value *Op1 = CI->getArgOperand(1);
579 cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
580 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), MaskVal & 1));
581 Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C),
582 ((MaskVal >> 1) & 1)+2));
583 Value *Mask = ConstantVector::get(Idxs);
584 SI = new ShuffleVectorInst(Op0, Op1, Mask, "upgraded.", CI);
587 assert(SI && "Unexpected!");
589 // Handle any uses of the old CallInst.
590 if (!CI->use_empty())
591 // Replace all uses of the old call with the new cast which has the
593 CI->replaceAllUsesWith(SI);
595 // Clean up the old call now that it has been completely upgraded.
596 CI->eraseFromParent();
597 } else if (F->getName() == "llvm.x86.sse41.pmulld") {
598 // Upgrade this set of intrinsics into vector multiplies.
599 Instruction *Mul = BinaryOperator::CreateMul(CI->getArgOperand(0),
600 CI->getArgOperand(1),
603 // Fix up all the uses with our new multiply.
604 if (!CI->use_empty())
605 CI->replaceAllUsesWith(Mul);
607 // Remove upgraded multiply.
608 CI->eraseFromParent();
609 } else if (F->getName() == "llvm.x86.ssse3.palign.r") {
610 Value *Op1 = CI->getArgOperand(0);
611 Value *Op2 = CI->getArgOperand(1);
612 Value *Op3 = CI->getArgOperand(2);
613 unsigned shiftVal = cast<ConstantInt>(Op3)->getZExtValue();
615 IRBuilder<> Builder(C);
616 Builder.SetInsertPoint(CI->getParent(), CI);
618 // If palignr is shifting the pair of input vectors less than 9 bytes,
619 // emit a shuffle instruction.
621 const Type *IntTy = Type::getInt32Ty(C);
622 const Type *EltTy = Type::getInt8Ty(C);
623 const Type *VecTy = VectorType::get(EltTy, 8);
625 Op2 = Builder.CreateBitCast(Op2, VecTy);
626 Op1 = Builder.CreateBitCast(Op1, VecTy);
628 llvm::SmallVector<llvm::Constant*, 8> Indices;
629 for (unsigned i = 0; i != 8; ++i)
630 Indices.push_back(ConstantInt::get(IntTy, shiftVal + i));
632 Value *SV = ConstantVector::get(Indices.begin(), Indices.size());
633 Rep = Builder.CreateShuffleVector(Op2, Op1, SV, "palignr");
634 Rep = Builder.CreateBitCast(Rep, F->getReturnType());
637 // If palignr is shifting the pair of input vectors more than 8 but less
638 // than 16 bytes, emit a logical right shift of the destination.
639 else if (shiftVal < 16) {
640 // MMX has these as 1 x i64 vectors for some odd optimization reasons.
641 const Type *EltTy = Type::getInt64Ty(C);
642 const Type *VecTy = VectorType::get(EltTy, 1);
644 Op1 = Builder.CreateBitCast(Op1, VecTy, "cast");
645 Op2 = ConstantInt::get(VecTy, (shiftVal-8) * 8);
647 // create i32 constant
649 Intrinsic::getDeclaration(F->getParent(), Intrinsic::x86_mmx_psrl_q);
650 Rep = Builder.CreateCall2(I, Op1, Op2, "palignr");
653 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
655 Rep = Constant::getNullValue(F->getReturnType());
658 // Replace any uses with our new instruction.
659 if (!CI->use_empty())
660 CI->replaceAllUsesWith(Rep);
662 // Remove upgraded instruction.
663 CI->eraseFromParent();
665 } else if (F->getName() == "llvm.x86.ssse3.palign.r.128") {
666 Value *Op1 = CI->getArgOperand(0);
667 Value *Op2 = CI->getArgOperand(1);
668 Value *Op3 = CI->getArgOperand(2);
669 unsigned shiftVal = cast<ConstantInt>(Op3)->getZExtValue();
671 IRBuilder<> Builder(C);
672 Builder.SetInsertPoint(CI->getParent(), CI);
674 // If palignr is shifting the pair of input vectors less than 17 bytes,
675 // emit a shuffle instruction.
676 if (shiftVal <= 16) {
677 const Type *IntTy = Type::getInt32Ty(C);
678 const Type *EltTy = Type::getInt8Ty(C);
679 const Type *VecTy = VectorType::get(EltTy, 16);
681 Op2 = Builder.CreateBitCast(Op2, VecTy);
682 Op1 = Builder.CreateBitCast(Op1, VecTy);
684 llvm::SmallVector<llvm::Constant*, 16> Indices;
685 for (unsigned i = 0; i != 16; ++i)
686 Indices.push_back(ConstantInt::get(IntTy, shiftVal + i));
688 Value *SV = ConstantVector::get(Indices.begin(), Indices.size());
689 Rep = Builder.CreateShuffleVector(Op2, Op1, SV, "palignr");
690 Rep = Builder.CreateBitCast(Rep, F->getReturnType());
693 // If palignr is shifting the pair of input vectors more than 16 but less
694 // than 32 bytes, emit a logical right shift of the destination.
695 else if (shiftVal < 32) {
696 const Type *EltTy = Type::getInt64Ty(C);
697 const Type *VecTy = VectorType::get(EltTy, 2);
698 const Type *IntTy = Type::getInt32Ty(C);
700 Op1 = Builder.CreateBitCast(Op1, VecTy, "cast");
701 Op2 = ConstantInt::get(IntTy, (shiftVal-16) * 8);
703 // create i32 constant
705 Intrinsic::getDeclaration(F->getParent(), Intrinsic::x86_sse2_psrl_dq);
706 Rep = Builder.CreateCall2(I, Op1, Op2, "palignr");
709 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
711 Rep = Constant::getNullValue(F->getReturnType());
714 // Replace any uses with our new instruction.
715 if (!CI->use_empty())
716 CI->replaceAllUsesWith(Rep);
718 // Remove upgraded instruction.
719 CI->eraseFromParent();
722 llvm_unreachable("Unknown function for CallInst upgrade.");
727 switch (NewFn->getIntrinsicID()) {
728 default: llvm_unreachable("Unknown function for CallInst upgrade.");
729 case Intrinsic::arm_neon_vld1:
730 case Intrinsic::arm_neon_vld2:
731 case Intrinsic::arm_neon_vld3:
732 case Intrinsic::arm_neon_vld4:
733 case Intrinsic::arm_neon_vst1:
734 case Intrinsic::arm_neon_vst2:
735 case Intrinsic::arm_neon_vst3:
736 case Intrinsic::arm_neon_vst4:
737 case Intrinsic::arm_neon_vld2lane:
738 case Intrinsic::arm_neon_vld3lane:
739 case Intrinsic::arm_neon_vld4lane:
740 case Intrinsic::arm_neon_vst2lane:
741 case Intrinsic::arm_neon_vst3lane:
742 case Intrinsic::arm_neon_vst4lane: {
743 // Add a default alignment argument of 1.
744 SmallVector<Value*, 8> Operands(CS.arg_begin(), CS.arg_end());
745 Operands.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
746 CallInst *NewCI = CallInst::Create(NewFn, Operands.begin(), Operands.end(),
748 NewCI->setTailCall(CI->isTailCall());
749 NewCI->setCallingConv(CI->getCallingConv());
751 // Handle any uses of the old CallInst.
752 if (!CI->use_empty())
753 // Replace all uses of the old call with the new cast which has the
755 CI->replaceAllUsesWith(NewCI);
757 // Clean up the old call now that it has been completely upgraded.
758 CI->eraseFromParent();
762 case Intrinsic::x86_mmx_psll_d:
763 case Intrinsic::x86_mmx_psll_q:
764 case Intrinsic::x86_mmx_psll_w:
765 case Intrinsic::x86_mmx_psra_d:
766 case Intrinsic::x86_mmx_psra_w:
767 case Intrinsic::x86_mmx_psrl_d:
768 case Intrinsic::x86_mmx_psrl_q:
769 case Intrinsic::x86_mmx_psrl_w: {
772 Operands[0] = CI->getArgOperand(0);
774 // Cast the second parameter to the correct type.
775 BitCastInst *BC = new BitCastInst(CI->getArgOperand(1),
776 NewFn->getFunctionType()->getParamType(1),
780 // Construct a new CallInst
781 CallInst *NewCI = CallInst::Create(NewFn, Operands, Operands+2,
782 "upgraded."+CI->getName(), CI);
783 NewCI->setTailCall(CI->isTailCall());
784 NewCI->setCallingConv(CI->getCallingConv());
786 // Handle any uses of the old CallInst.
787 if (!CI->use_empty())
788 // Replace all uses of the old call with the new cast which has the
790 CI->replaceAllUsesWith(NewCI);
792 // Clean up the old call now that it has been completely upgraded.
793 CI->eraseFromParent();
796 case Intrinsic::ctlz:
797 case Intrinsic::ctpop:
798 case Intrinsic::cttz: {
799 // Build a small vector of the original arguments.
800 SmallVector<Value*, 8> Operands(CS.arg_begin(), CS.arg_end());
802 // Construct a new CallInst
803 CallInst *NewCI = CallInst::Create(NewFn, Operands.begin(), Operands.end(),
804 "upgraded."+CI->getName(), CI);
805 NewCI->setTailCall(CI->isTailCall());
806 NewCI->setCallingConv(CI->getCallingConv());
808 // Handle any uses of the old CallInst.
809 if (!CI->use_empty()) {
810 // Check for sign extend parameter attributes on the return values.
811 bool SrcSExt = NewFn->getAttributes().paramHasAttr(0, Attribute::SExt);
812 bool DestSExt = F->getAttributes().paramHasAttr(0, Attribute::SExt);
814 // Construct an appropriate cast from the new return type to the old.
815 CastInst *RetCast = CastInst::Create(
816 CastInst::getCastOpcode(NewCI, SrcSExt,
819 NewCI, F->getReturnType(),
820 NewCI->getName(), CI);
821 NewCI->moveBefore(RetCast);
823 // Replace all uses of the old call with the new cast which has the
825 CI->replaceAllUsesWith(RetCast);
828 // Clean up the old call now that it has been completely upgraded.
829 CI->eraseFromParent();
832 case Intrinsic::eh_selector:
833 case Intrinsic::eh_typeid_for: {
834 // Only the return type changed.
835 SmallVector<Value*, 8> Operands(CS.arg_begin(), CS.arg_end());
836 CallInst *NewCI = CallInst::Create(NewFn, Operands.begin(), Operands.end(),
837 "upgraded." + CI->getName(), CI);
838 NewCI->setTailCall(CI->isTailCall());
839 NewCI->setCallingConv(CI->getCallingConv());
841 // Handle any uses of the old CallInst.
842 if (!CI->use_empty()) {
843 // Construct an appropriate cast from the new return type to the old.
845 CastInst::Create(CastInst::getCastOpcode(NewCI, true,
846 F->getReturnType(), true),
847 NewCI, F->getReturnType(), NewCI->getName(), CI);
848 CI->replaceAllUsesWith(RetCast);
850 CI->eraseFromParent();
853 case Intrinsic::memcpy:
854 case Intrinsic::memmove:
855 case Intrinsic::memset: {
857 const llvm::Type *I1Ty = llvm::Type::getInt1Ty(CI->getContext());
858 Value *Operands[5] = { CI->getArgOperand(0), CI->getArgOperand(1),
859 CI->getArgOperand(2), CI->getArgOperand(3),
860 llvm::ConstantInt::get(I1Ty, 0) };
861 CallInst *NewCI = CallInst::Create(NewFn, Operands, Operands+5,
863 NewCI->setTailCall(CI->isTailCall());
864 NewCI->setCallingConv(CI->getCallingConv());
865 // Handle any uses of the old CallInst.
866 if (!CI->use_empty())
867 // Replace all uses of the old call with the new cast which has the
869 CI->replaceAllUsesWith(NewCI);
871 // Clean up the old call now that it has been completely upgraded.
872 CI->eraseFromParent();
878 // This tests each Function to determine if it needs upgrading. When we find
879 // one we are interested in, we then upgrade all calls to reflect the new
881 void llvm::UpgradeCallsToIntrinsic(Function* F) {
882 assert(F && "Illegal attempt to upgrade a non-existent intrinsic.");
884 // Upgrade the function and check if it is a totaly new function.
886 if (UpgradeIntrinsicFunction(F, NewFn)) {
888 // Replace all uses to the old function with the new one if necessary.
889 for (Value::use_iterator UI = F->use_begin(), UE = F->use_end();
891 if (CallInst* CI = dyn_cast<CallInst>(*UI++))
892 UpgradeIntrinsicCall(CI, NewFn);
894 // Remove old function, no longer used, from the module.
895 F->eraseFromParent();
900 /// This function strips all debug info intrinsics, except for llvm.dbg.declare.
901 /// If an llvm.dbg.declare intrinsic is invalid, then this function simply
903 void llvm::CheckDebugInfoIntrinsics(Module *M) {
906 if (Function *FuncStart = M->getFunction("llvm.dbg.func.start")) {
907 while (!FuncStart->use_empty()) {
908 CallInst *CI = cast<CallInst>(FuncStart->use_back());
909 CI->eraseFromParent();
911 FuncStart->eraseFromParent();
914 if (Function *StopPoint = M->getFunction("llvm.dbg.stoppoint")) {
915 while (!StopPoint->use_empty()) {
916 CallInst *CI = cast<CallInst>(StopPoint->use_back());
917 CI->eraseFromParent();
919 StopPoint->eraseFromParent();
922 if (Function *RegionStart = M->getFunction("llvm.dbg.region.start")) {
923 while (!RegionStart->use_empty()) {
924 CallInst *CI = cast<CallInst>(RegionStart->use_back());
925 CI->eraseFromParent();
927 RegionStart->eraseFromParent();
930 if (Function *RegionEnd = M->getFunction("llvm.dbg.region.end")) {
931 while (!RegionEnd->use_empty()) {
932 CallInst *CI = cast<CallInst>(RegionEnd->use_back());
933 CI->eraseFromParent();
935 RegionEnd->eraseFromParent();
938 if (Function *Declare = M->getFunction("llvm.dbg.declare")) {
939 if (!Declare->use_empty()) {
940 DbgDeclareInst *DDI = cast<DbgDeclareInst>(Declare->use_back());
941 if (!isa<MDNode>(DDI->getArgOperand(0)) ||
942 !isa<MDNode>(DDI->getArgOperand(1))) {
943 while (!Declare->use_empty()) {
944 CallInst *CI = cast<CallInst>(Declare->use_back());
945 CI->eraseFromParent();
947 Declare->eraseFromParent();