#include "llvm/AutoUpgrade.h"
#include "llvm/Constants.h"
#include "llvm/Function.h"
+#include "llvm/LLVMContext.h"
#include "llvm/Module.h"
-#include "llvm/Instructions.h"
-#include "llvm/Intrinsics.h"
+#include "llvm/IntrinsicInst.h"
#include "llvm/ADT/SmallVector.h"
+#include "llvm/Support/CallSite.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/IRBuilder.h"
#include <cstring>
using namespace llvm;
static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
assert(F && "Illegal to upgrade a non-existent Function.");
- // Get the Function's name.
- const std::string& Name = F->getName();
-
- // Convenience
- const FunctionType *FTy = F->getFunctionType();
-
// Quickly eliminate it, if it's not a candidate.
- if (Name.length() <= 8 || Name[0] != 'l' || Name[1] != 'l' ||
- Name[2] != 'v' || Name[3] != 'm' || Name[4] != '.')
+ StringRef Name = F->getName();
+ if (Name.size() <= 8 || !Name.startswith("llvm."))
return false;
+ Name = Name.substr(5); // Strip off "llvm."
+ FunctionType *FTy = F->getFunctionType();
Module *M = F->getParent();
- switch (Name[5]) {
+
+ switch (Name[0]) {
default: break;
- case 'a':
- // This upgrades the llvm.atomic.lcs, llvm.atomic.las, and llvm.atomic.lss
- // to their new function name
- if (Name.compare(5,8,"atomic.l",8) == 0) {
- if (Name.compare(12,3,"lcs",3) == 0) {
- std::string::size_type delim = Name.find('.',12);
- F->setName("llvm.atomic.cmp.swap"+Name.substr(delim));
- NewFn = F;
- return true;
- }
- else if (Name.compare(12,3,"las",3) == 0) {
- std::string::size_type delim = Name.find('.',12);
- F->setName("llvm.atomic.load.add"+Name.substr(delim));
- NewFn = F;
- return true;
- }
- else if (Name.compare(12,3,"lss",3) == 0) {
- std::string::size_type delim = Name.find('.',12);
- F->setName("llvm.atomic.load.sub"+Name.substr(delim));
- NewFn = F;
- return true;
- }
- }
- break;
- case 'b':
- // This upgrades the name of the llvm.bswap intrinsic function to only use
- // a single type name for overloading. We only care about the old format
- // 'llvm.bswap.i*.i*', so check for 'bswap.' and then for there being
- // a '.' after 'bswap.'
- if (Name.compare(5,6,"bswap.",6) == 0) {
- std::string::size_type delim = Name.find('.',11);
-
- if (delim != std::string::npos) {
- // Construct the new name as 'llvm.bswap' + '.i*'
- F->setName(Name.substr(0,10)+Name.substr(delim));
- NewFn = F;
- return true;
- }
- }
- break;
+ case 'p':
+ // This upgrades the llvm.prefetch intrinsic to accept one more parameter,
+ // which is a instruction / data cache identifier. The old version only
+ // implicitly accepted the data version.
+ if (Name == "prefetch") {
+ // Don't do anything if it has the correct number of arguments already
+ if (FTy->getNumParams() == 4)
+ break;
- case 'c':
- // We only want to fix the 'llvm.ct*' intrinsics which do not have the
- // correct return type, so we check for the name, and then check if the
- // return type does not match the parameter type.
- if ( (Name.compare(5,5,"ctpop",5) == 0 ||
- Name.compare(5,4,"ctlz",4) == 0 ||
- Name.compare(5,4,"cttz",4) == 0) &&
- FTy->getReturnType() != FTy->getParamType(0)) {
- // We first need to change the name of the old (bad) intrinsic, because
- // its type is incorrect, but we cannot overload that name. We
- // arbitrarily unique it here allowing us to construct a correctly named
+ assert(FTy->getNumParams() == 3 && "old prefetch takes 3 args!");
+ // We first need to change the name of the old (bad) intrinsic, because
+ // its type is incorrect, but we cannot overload that name. We
+ // arbitrarily unique it here allowing us to construct a correctly named
// and typed function below.
+ std::string NameTmp = F->getName();
F->setName("");
-
- // Now construct the new intrinsic with the correct name and type. We
- // leave the old function around in order to query its type, whatever it
- // may be, and correctly convert up to the new type.
- NewFn = cast<Function>(M->getOrInsertFunction(Name,
- FTy->getParamType(0),
+ NewFn = cast<Function>(M->getOrInsertFunction(NameTmp,
+ FTy->getReturnType(),
FTy->getParamType(0),
- (Type *)0));
+ FTy->getParamType(1),
+ FTy->getParamType(2),
+ FTy->getParamType(2),
+ (Type*)0));
return true;
}
- break;
-
- case 'p':
- // This upgrades the llvm.part.select overloaded intrinsic names to only
- // use one type specifier in the name. We only care about the old format
- // 'llvm.part.select.i*.i*', and solve as above with bswap.
- if (Name.compare(5,12,"part.select.",12) == 0) {
- std::string::size_type delim = Name.find('.',17);
-
- if (delim != std::string::npos) {
- // Construct a new name as 'llvm.part.select' + '.i*'
- F->setName(Name.substr(0,16)+Name.substr(delim));
- NewFn = F;
- return true;
- }
- break;
- }
-
- // This upgrades the llvm.part.set intrinsics similarly as above, however
- // we care about 'llvm.part.set.i*.i*.i*', but only the first two types
- // must match. There is an additional type specifier after these two
- // matching types that we must retain when upgrading. Thus, we require
- // finding 2 periods, not just one, after the intrinsic name.
- if (Name.compare(5,9,"part.set.",9) == 0) {
- std::string::size_type delim = Name.find('.',14);
-
- if (delim != std::string::npos &&
- Name.find('.',delim+1) != std::string::npos) {
- // Construct a new name as 'llvm.part.select' + '.i*.i*'
- F->setName(Name.substr(0,13)+Name.substr(delim));
- NewFn = F;
- return true;
- }
- break;
- }
break;
- case 'x':
- // This fixes all MMX shift intrinsic instructions to take a
- // v1i64 instead of a v2i32 as the second parameter.
- if (Name.compare(5,10,"x86.mmx.ps",10) == 0 &&
- (Name.compare(13,4,"psll", 4) == 0 ||
- Name.compare(13,4,"psra", 4) == 0 ||
- Name.compare(13,4,"psrl", 4) == 0) && Name[17] != 'i') {
-
- const llvm::Type *VT = VectorType::get(IntegerType::get(64), 1);
-
- // We don't have to do anything if the parameter already has
- // the correct type.
- if (FTy->getParamType(1) == VT)
- break;
-
- // We first need to change the name of the old (bad) intrinsic, because
- // its type is incorrect, but we cannot overload that name. We
- // arbitrarily unique it here allowing us to construct a correctly named
- // and typed function below.
- F->setName("");
+ case 'x': {
+ const char *NewFnName = NULL;
+ // This fixes the poorly named crc32 intrinsics.
+ if (Name == "x86.sse42.crc32.8")
+ NewFnName = "llvm.x86.sse42.crc32.32.8";
+ else if (Name == "x86.sse42.crc32.16")
+ NewFnName = "llvm.x86.sse42.crc32.32.16";
+ else if (Name == "x86.sse42.crc32.32")
+ NewFnName = "llvm.x86.sse42.crc32.32.32";
+ else if (Name == "x86.sse42.crc64.8")
+ NewFnName = "llvm.x86.sse42.crc32.64.8";
+ else if (Name == "x86.sse42.crc64.64")
+ NewFnName = "llvm.x86.sse42.crc32.64.64";
+
+ if (NewFnName) {
+ F->setName(NewFnName);
+ NewFn = F;
+ return true;
+ }
- assert(FTy->getNumParams() == 2 && "MMX shift intrinsics take 2 args!");
-
- // Now construct the new intrinsic with the correct name and type. We
- // leave the old function around in order to query its type, whatever it
- // may be, and correctly convert up to the new type.
- NewFn = cast<Function>(M->getOrInsertFunction(Name,
- FTy->getReturnType(),
- FTy->getParamType(0),
- VT,
- (Type *)0));
+ // Calls to these instructions are transformed into unaligned loads.
+ if (Name == "x86.sse.loadu.ps" || Name == "x86.sse2.loadu.dq" ||
+ Name == "x86.sse2.loadu.pd")
return true;
- } else if (Name.compare(5,17,"x86.sse2.loadh.pd",17) == 0 ||
- Name.compare(5,17,"x86.sse2.loadl.pd",17) == 0 ||
- Name.compare(5,16,"x86.sse2.movl.dq",16) == 0 ||
- Name.compare(5,15,"x86.sse2.movs.d",15) == 0 ||
- Name.compare(5,16,"x86.sse2.shuf.pd",16) == 0 ||
- Name.compare(5,18,"x86.sse2.unpckh.pd",18) == 0 ||
- Name.compare(5,18,"x86.sse2.unpckl.pd",18) == 0 ||
- Name.compare(5,20,"x86.sse2.punpckh.qdq",20) == 0 ||
- Name.compare(5,20,"x86.sse2.punpckl.qdq",20) == 0) {
- // Calls to these intrinsics are transformed into ShuffleVector's.
- NewFn = 0;
+
+ // Calls to these instructions are transformed into nontemporal stores.
+ if (Name == "x86.sse.movnt.ps" || Name == "x86.sse2.movnt.dq" ||
+ Name == "x86.sse2.movnt.pd" || Name == "x86.sse2.movnt.i")
return true;
- }
break;
}
+ }
// This may not belong here. This function is effectively being overloaded
// to both detect an intrinsic which needs upgrading, and to provide the
// Upgrade intrinsic attributes. This does not change the function.
if (NewFn)
F = NewFn;
- if (unsigned id = F->getIntrinsicID(true))
- F->setParamAttrs(Intrinsic::getParamAttrs((Intrinsic::ID)id));
+ if (unsigned id = F->getIntrinsicID())
+ F->setAttributes(Intrinsic::getAttributes((Intrinsic::ID)id));
return Upgraded;
}
+bool llvm::UpgradeGlobalVariable(GlobalVariable *GV) {
+ // Nothing to do yet.
+ return false;
+}
+
// UpgradeIntrinsicCall - Upgrade a call to an old intrinsic to be a call the
// upgraded intrinsic. All argument and return casting must be provided in
// order to seamlessly integrate with existing context.
void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
Function *F = CI->getCalledFunction();
+ LLVMContext &C = CI->getContext();
+ ImmutableCallSite CS(CI);
+
assert(F && "CallInst has no function associated with it.");
if (!NewFn) {
- bool isLoadH = false, isLoadL = false, isMovL = false;
- bool isMovSD = false, isShufPD = false;
- bool isUnpckhPD = false, isUnpcklPD = false;
- bool isPunpckhQPD = false, isPunpcklQPD = false;
- if (strcmp(F->getNameStart(), "llvm.x86.sse2.loadh.pd") == 0)
- isLoadH = true;
- else if (strcmp(F->getNameStart(), "llvm.x86.sse2.loadl.pd") == 0)
- isLoadL = true;
- else if (strcmp(F->getNameStart(), "llvm.x86.sse2.movl.dq") == 0)
- isMovL = true;
- else if (strcmp(F->getNameStart(), "llvm.x86.sse2.movs.d") == 0)
- isMovSD = true;
- else if (strcmp(F->getNameStart(), "llvm.x86.sse2.shuf.pd") == 0)
- isShufPD = true;
- else if (strcmp(F->getNameStart(), "llvm.x86.sse2.unpckh.pd") == 0)
- isUnpckhPD = true;
- else if (strcmp(F->getNameStart(), "llvm.x86.sse2.unpckl.pd") == 0)
- isUnpcklPD = true;
- else if (strcmp(F->getNameStart(), "llvm.x86.sse2.punpckh.qdq") == 0)
- isPunpckhQPD = true;
- else if (strcmp(F->getNameStart(), "llvm.x86.sse2.punpckl.qdq") == 0)
- isPunpcklQPD = true;
-
- if (isLoadH || isLoadL || isMovL || isMovSD || isShufPD ||
- isUnpckhPD || isUnpcklPD || isPunpckhQPD || isPunpcklQPD) {
- std::vector<Constant*> Idxs;
- Value *Op0 = CI->getOperand(1);
- ShuffleVectorInst *SI = NULL;
- if (isLoadH || isLoadL) {
- Value *Op1 = UndefValue::get(Op0->getType());
- Value *Addr = new BitCastInst(CI->getOperand(2),
- PointerType::getUnqual(Type::DoubleTy),
- "upgraded.", CI);
- Value *Load = new LoadInst(Addr, "upgraded.", false, 8, CI);
- Value *Idx = ConstantInt::get(Type::Int32Ty, 0);
- Op1 = InsertElementInst::Create(Op1, Load, Idx, "upgraded.", CI);
-
- if (isLoadH) {
- Idxs.push_back(ConstantInt::get(Type::Int32Ty, 0));
- Idxs.push_back(ConstantInt::get(Type::Int32Ty, 2));
- } else {
- Idxs.push_back(ConstantInt::get(Type::Int32Ty, 2));
- Idxs.push_back(ConstantInt::get(Type::Int32Ty, 1));
- }
- Value *Mask = ConstantVector::get(Idxs);
- SI = new ShuffleVectorInst(Op0, Op1, Mask, "upgraded.", CI);
- } else if (isMovL) {
- Constant *Zero = ConstantInt::get(Type::Int32Ty, 0);
- Idxs.push_back(Zero);
- Idxs.push_back(Zero);
- Idxs.push_back(Zero);
- Idxs.push_back(Zero);
- Value *ZeroV = ConstantVector::get(Idxs);
-
- Idxs.clear();
- Idxs.push_back(ConstantInt::get(Type::Int32Ty, 4));
- Idxs.push_back(ConstantInt::get(Type::Int32Ty, 5));
- Idxs.push_back(ConstantInt::get(Type::Int32Ty, 2));
- Idxs.push_back(ConstantInt::get(Type::Int32Ty, 3));
- Value *Mask = ConstantVector::get(Idxs);
- SI = new ShuffleVectorInst(ZeroV, Op0, Mask, "upgraded.", CI);
- } else if (isMovSD ||
- isUnpckhPD || isUnpcklPD || isPunpckhQPD || isPunpcklQPD) {
- Value *Op1 = CI->getOperand(2);
- if (isMovSD) {
- Idxs.push_back(ConstantInt::get(Type::Int32Ty, 2));
- Idxs.push_back(ConstantInt::get(Type::Int32Ty, 1));
- } else if (isUnpckhPD || isPunpckhQPD) {
- Idxs.push_back(ConstantInt::get(Type::Int32Ty, 1));
- Idxs.push_back(ConstantInt::get(Type::Int32Ty, 3));
- } else {
- Idxs.push_back(ConstantInt::get(Type::Int32Ty, 0));
- Idxs.push_back(ConstantInt::get(Type::Int32Ty, 2));
- }
- Value *Mask = ConstantVector::get(Idxs);
- SI = new ShuffleVectorInst(Op0, Op1, Mask, "upgraded.", CI);
- } else if (isShufPD) {
- Value *Op1 = CI->getOperand(2);
- unsigned MaskVal = cast<ConstantInt>(CI->getOperand(3))->getZExtValue();
- Idxs.push_back(ConstantInt::get(Type::Int32Ty, MaskVal & 1));
- Idxs.push_back(ConstantInt::get(Type::Int32Ty, ((MaskVal >> 1) & 1)+2));
- Value *Mask = ConstantVector::get(Idxs);
- SI = new ShuffleVectorInst(Op0, Op1, Mask, "upgraded.", CI);
- }
-
- assert(SI && "Unexpected!");
-
- // Handle any uses of the old CallInst.
+ if (F->getName() == "llvm.x86.sse.loadu.ps" ||
+ F->getName() == "llvm.x86.sse2.loadu.dq" ||
+ F->getName() == "llvm.x86.sse2.loadu.pd") {
+ // Convert to a native, unaligned load.
+ Type *VecTy = CI->getType();
+ Type *IntTy = IntegerType::get(C, 128);
+ IRBuilder<> Builder(C);
+ Builder.SetInsertPoint(CI->getParent(), CI);
+
+ Value *BC = Builder.CreateBitCast(CI->getArgOperand(0),
+ PointerType::getUnqual(IntTy),
+ "cast");
+ LoadInst *LI = Builder.CreateLoad(BC, CI->getName());
+ LI->setAlignment(1); // Unaligned load.
+ BC = Builder.CreateBitCast(LI, VecTy, "new.cast");
+
+ // Fix up all the uses with our new load.
if (!CI->use_empty())
- // Replace all uses of the old call with the new cast which has the
- // correct type.
- CI->replaceAllUsesWith(SI);
-
- // Clean up the old call now that it has been completely upgraded.
+ CI->replaceAllUsesWith(BC);
+
+ // Remove intrinsic.
+ CI->eraseFromParent();
+ } else if (F->getName() == "llvm.x86.sse.movnt.ps" ||
+ F->getName() == "llvm.x86.sse2.movnt.dq" ||
+ F->getName() == "llvm.x86.sse2.movnt.pd" ||
+ F->getName() == "llvm.x86.sse2.movnt.i") {
+ IRBuilder<> Builder(C);
+ Builder.SetInsertPoint(CI->getParent(), CI);
+
+ Module *M = F->getParent();
+ SmallVector<Value *, 1> Elts;
+ Elts.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
+ MDNode *Node = MDNode::get(C, Elts);
+
+ Value *Arg0 = CI->getArgOperand(0);
+ Value *Arg1 = CI->getArgOperand(1);
+
+ // Convert the type of the pointer to a pointer to the stored type.
+ Value *BC = Builder.CreateBitCast(Arg0,
+ PointerType::getUnqual(Arg1->getType()),
+ "cast");
+ StoreInst *SI = Builder.CreateStore(Arg1, BC);
+ SI->setMetadata(M->getMDKindID("nontemporal"), Node);
+ SI->setAlignment(16);
+
+ // Remove intrinsic.
CI->eraseFromParent();
} else {
- assert(0 && "Unknown function for CallInst upgrade.");
+ llvm_unreachable("Unknown function for CallInst upgrade.");
}
return;
}
switch (NewFn->getIntrinsicID()) {
- default: assert(0 && "Unknown function for CallInst upgrade.");
- case Intrinsic::x86_mmx_psll_d:
- case Intrinsic::x86_mmx_psll_q:
- case Intrinsic::x86_mmx_psll_w:
- case Intrinsic::x86_mmx_psra_d:
- case Intrinsic::x86_mmx_psra_w:
- case Intrinsic::x86_mmx_psrl_d:
- case Intrinsic::x86_mmx_psrl_q:
- case Intrinsic::x86_mmx_psrl_w: {
- Value *Operands[2];
-
- Operands[0] = CI->getOperand(1);
-
- // Cast the second parameter to the correct type.
- BitCastInst *BC = new BitCastInst(CI->getOperand(2),
- NewFn->getFunctionType()->getParamType(1),
- "upgraded.", CI);
- Operands[1] = BC;
-
- // Construct a new CallInst
- CallInst *NewCI = CallInst::Create(NewFn, Operands, Operands+2,
- "upgraded."+CI->getName(), CI);
+ case Intrinsic::prefetch: {
+ IRBuilder<> Builder(C);
+ Builder.SetInsertPoint(CI->getParent(), CI);
+ llvm::Type *I32Ty = llvm::Type::getInt32Ty(CI->getContext());
+
+ // Add the extra "data cache" argument
+ Value *Operands[4] = { CI->getArgOperand(0), CI->getArgOperand(1),
+ CI->getArgOperand(2),
+ llvm::ConstantInt::get(I32Ty, 1) };
+ CallInst *NewCI = CallInst::Create(NewFn, Operands,
+ CI->getName(), CI);
NewCI->setTailCall(CI->isTailCall());
NewCI->setCallingConv(CI->getCallingConv());
-
// Handle any uses of the old CallInst.
if (!CI->use_empty())
- // Replace all uses of the old call with the new cast which has the
+ // Replace all uses of the old call with the new cast which has the
// correct type.
CI->replaceAllUsesWith(NewCI);
-
- // Clean up the old call now that it has been completely upgraded.
- CI->eraseFromParent();
- break;
- }
- case Intrinsic::ctlz:
- case Intrinsic::ctpop:
- case Intrinsic::cttz: {
- // Build a small vector of the 1..(N-1) operands, which are the
- // parameters.
- SmallVector<Value*, 8> Operands(CI->op_begin()+1, CI->op_end());
-
- // Construct a new CallInst
- CallInst *NewCI = CallInst::Create(NewFn, Operands.begin(), Operands.end(),
- "upgraded."+CI->getName(), CI);
- NewCI->setTailCall(CI->isTailCall());
- NewCI->setCallingConv(CI->getCallingConv());
-
- // Handle any uses of the old CallInst.
- if (!CI->use_empty()) {
- // Check for sign extend parameter attributes on the return values.
- bool SrcSExt = NewFn->getParamAttrs().paramHasAttr(0, ParamAttr::SExt);
- bool DestSExt = F->getParamAttrs().paramHasAttr(0, ParamAttr::SExt);
-
- // Construct an appropriate cast from the new return type to the old.
- CastInst *RetCast = CastInst::Create(
- CastInst::getCastOpcode(NewCI, SrcSExt,
- F->getReturnType(),
- DestSExt),
- NewCI, F->getReturnType(),
- NewCI->getName(), CI);
- NewCI->moveBefore(RetCast);
-
- // Replace all uses of the old call with the new cast which has the
- // correct type.
- CI->replaceAllUsesWith(RetCast);
- }
// Clean up the old call now that it has been completely upgraded.
CI->eraseFromParent();
+ break;
}
- break;
}
}
assert(F && "Illegal attempt to upgrade a non-existent intrinsic.");
// Upgrade the function and check if it is a totaly new function.
- Function* NewFn;
+ Function *NewFn;
if (UpgradeIntrinsicFunction(F, NewFn)) {
if (NewFn != F) {
// Replace all uses to the old function with the new one if necessary.
for (Value::use_iterator UI = F->use_begin(), UE = F->use_end();
UI != UE; ) {
- if (CallInst* CI = dyn_cast<CallInst>(*UI++))
+ if (CallInst *CI = dyn_cast<CallInst>(*UI++))
UpgradeIntrinsicCall(CI, NewFn);
}
// Remove old function, no longer used, from the module.
}
}
}
+
+/// This function strips all debug info intrinsics, except for llvm.dbg.declare.
+/// If an llvm.dbg.declare intrinsic is invalid, then this function simply
+/// strips that use.
+void llvm::CheckDebugInfoIntrinsics(Module *M) {
+ if (Function *FuncStart = M->getFunction("llvm.dbg.func.start")) {
+ while (!FuncStart->use_empty())
+ cast<CallInst>(FuncStart->use_back())->eraseFromParent();
+ FuncStart->eraseFromParent();
+ }
+
+ if (Function *StopPoint = M->getFunction("llvm.dbg.stoppoint")) {
+ while (!StopPoint->use_empty())
+ cast<CallInst>(StopPoint->use_back())->eraseFromParent();
+ StopPoint->eraseFromParent();
+ }
+
+ if (Function *RegionStart = M->getFunction("llvm.dbg.region.start")) {
+ while (!RegionStart->use_empty())
+ cast<CallInst>(RegionStart->use_back())->eraseFromParent();
+ RegionStart->eraseFromParent();
+ }
+
+ if (Function *RegionEnd = M->getFunction("llvm.dbg.region.end")) {
+ while (!RegionEnd->use_empty())
+ cast<CallInst>(RegionEnd->use_back())->eraseFromParent();
+ RegionEnd->eraseFromParent();
+ }
+
+ if (Function *Declare = M->getFunction("llvm.dbg.declare")) {
+ if (!Declare->use_empty()) {
+ DbgDeclareInst *DDI = cast<DbgDeclareInst>(Declare->use_back());
+ if (!isa<MDNode>(DDI->getArgOperand(0)) ||
+ !isa<MDNode>(DDI->getArgOperand(1))) {
+ while (!Declare->use_empty()) {
+ CallInst *CI = cast<CallInst>(Declare->use_back());
+ CI->eraseFromParent();
+ }
+ Declare->eraseFromParent();
+ }
+ }
+ }
+}