#include "llvm/AutoUpgrade.h"
#include "llvm/Constants.h"
#include "llvm/Function.h"
+#include "llvm/IRBuilder.h"
+#include "llvm/Instruction.h"
+#include "llvm/IntrinsicInst.h"
#include "llvm/LLVMContext.h"
#include "llvm/Module.h"
-#include "llvm/IntrinsicInst.h"
-#include "llvm/ADT/SmallVector.h"
+#include "llvm/Support/CFG.h"
#include "llvm/Support/CallSite.h"
#include "llvm/Support/ErrorHandling.h"
-#include "llvm/Support/IRBuilder.h"
#include <cstring>
using namespace llvm;
+// Upgrade the declarations of the SSE4.1 functions whose arguments have
+// changed their type from v4f32 to v2i64.
+static bool UpgradeSSE41Function(Function* F, Intrinsic::ID IID,
+ Function *&NewFn) {
+ // Check whether this is an old version of the function, which received
+ // v4f32 arguments.
+ Type *Arg0Type = F->getFunctionType()->getParamType(0);
+ if (Arg0Type != VectorType::get(Type::getFloatTy(F->getContext()), 4))
+ return false;
+
+ // Yes, it's old, replace it with new version.
+ F->setName(F->getName() + ".old");
+ NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
+ return true;
+}
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."
- Module *M = F->getParent();
- switch (Name[5]) {
+ switch (Name[0]) {
default: break;
- case 'a':
- // This upgrades the llvm.atomic.lcs, llvm.atomic.las, llvm.atomic.lss,
- // and atomics with default address spaces to their new names to their new
- // function name (e.g. llvm.atomic.add.i32 => llvm.atomic.add.i32.p0i32)
- if (Name.compare(5,7,"atomic.",7) == 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) +
- ".p0" + Name.substr(delim+1));
- 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)
- + ".p0" + Name.substr(delim+1));
- 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)
- + ".p0" + Name.substr(delim+1));
- NewFn = F;
- return true;
- }
- else if (Name.rfind(".p") == std::string::npos) {
- // We don't have an address space qualifier so this has be upgraded
- // to the new name. Copy the type name at the end of the intrinsic
- // and add to it
- std::string::size_type delim = Name.find_last_of('.');
- assert(delim != std::string::npos && "can not find type");
- F->setName(Name + ".p0" + Name.substr(delim+1));
- NewFn = F;
- return true;
- }
- } else if (Name.compare(5, 9, "arm.neon.", 9) == 0) {
- if (Name.compare(14, 7, "vmovls.", 7) == 0 ||
- Name.compare(14, 7, "vmovlu.", 7) == 0) {
- // Calls to these are transformed into IR without intrinsics.
- NewFn = 0;
- 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 '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
- // and typed function below.
- 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),
- FTy->getParamType(0),
- (Type *)0));
- return true;
- }
- break;
-
- case 'e':
- // The old llvm.eh.selector.i32 is equivalent to the new llvm.eh.selector.
- if (Name.compare("llvm.eh.selector.i32") == 0) {
- F->setName("llvm.eh.selector");
- NewFn = F;
+ case 'a': {
+ if (Name.startswith("arm.neon.vclz")) {
+ Type* args[2] = {
+ F->arg_begin()->getType(),
+ Type::getInt1Ty(F->getContext())
+ };
+ // Can't use Intrinsic::getDeclaration here as it adds a ".i1" to
+ // the end of the name. Change name from llvm.arm.neon.vclz.* to
+ // llvm.ctlz.*
+ FunctionType* fType = FunctionType::get(F->getReturnType(), args, false);
+ NewFn = Function::Create(fType, F->getLinkage(),
+ "llvm.ctlz." + Name.substr(14), F->getParent());
return true;
}
- // The old llvm.eh.typeid.for.i32 is equivalent to llvm.eh.typeid.for.
- if (Name.compare("llvm.eh.typeid.for.i32") == 0) {
- F->setName("llvm.eh.typeid.for");
- NewFn = F;
- return true;
- }
- // Convert the old llvm.eh.selector.i64 to a call to llvm.eh.selector.
- if (Name.compare("llvm.eh.selector.i64") == 0) {
- NewFn = Intrinsic::getDeclaration(M, Intrinsic::eh_selector);
- return true;
- }
- // Convert the old llvm.eh.typeid.for.i64 to a call to llvm.eh.typeid.for.
- if (Name.compare("llvm.eh.typeid.for.i64") == 0) {
- NewFn = Intrinsic::getDeclaration(M, Intrinsic::eh_typeid_for);
+ if (Name.startswith("arm.neon.vcnt")) {
+ NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctpop,
+ F->arg_begin()->getType());
return true;
}
break;
-
- case 'm': {
- // This upgrades the llvm.memcpy, llvm.memmove, and llvm.memset to the
- // new format that allows overloading the pointer for different address
- // space (e.g., llvm.memcpy.i16 => llvm.memcpy.p0i8.p0i8.i16)
- const char* NewFnName = NULL;
- if (Name.compare(5,8,"memcpy.i",8) == 0) {
- if (Name[13] == '8')
- NewFnName = "llvm.memcpy.p0i8.p0i8.i8";
- else if (Name.compare(13,2,"16") == 0)
- NewFnName = "llvm.memcpy.p0i8.p0i8.i16";
- else if (Name.compare(13,2,"32") == 0)
- NewFnName = "llvm.memcpy.p0i8.p0i8.i32";
- else if (Name.compare(13,2,"64") == 0)
- NewFnName = "llvm.memcpy.p0i8.p0i8.i64";
- } else if (Name.compare(5,9,"memmove.i",9) == 0) {
- if (Name[14] == '8')
- NewFnName = "llvm.memmove.p0i8.p0i8.i8";
- else if (Name.compare(14,2,"16") == 0)
- NewFnName = "llvm.memmove.p0i8.p0i8.i16";
- else if (Name.compare(14,2,"32") == 0)
- NewFnName = "llvm.memmove.p0i8.p0i8.i32";
- else if (Name.compare(14,2,"64") == 0)
- NewFnName = "llvm.memmove.p0i8.p0i8.i64";
- }
- else if (Name.compare(5,8,"memset.i",8) == 0) {
- if (Name[13] == '8')
- NewFnName = "llvm.memset.p0i8.i8";
- else if (Name.compare(13,2,"16") == 0)
- NewFnName = "llvm.memset.p0i8.i16";
- else if (Name.compare(13,2,"32") == 0)
- NewFnName = "llvm.memset.p0i8.i32";
- else if (Name.compare(13,2,"64") == 0)
- NewFnName = "llvm.memset.p0i8.i64";
+ }
+ case 'c': {
+ if (Name.startswith("ctlz.") && F->arg_size() == 1) {
+ F->setName(Name + ".old");
+ NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz,
+ F->arg_begin()->getType());
+ return true;
}
- if (NewFnName) {
- const FunctionType *FTy = F->getFunctionType();
- NewFn = cast<Function>(M->getOrInsertFunction(NewFnName,
- FTy->getReturnType(),
- FTy->getParamType(0),
- FTy->getParamType(1),
- FTy->getParamType(2),
- FTy->getParamType(3),
- Type::getInt1Ty(F->getContext()),
- (Type *)0));
+ if (Name.startswith("cttz.") && F->arg_size() == 1) {
+ F->setName(Name + ".old");
+ NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::cttz,
+ F->arg_begin()->getType());
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;
+ case 'x': {
+ if (Name.startswith("x86.sse2.pcmpeq.") ||
+ Name.startswith("x86.sse2.pcmpgt.") ||
+ Name.startswith("x86.avx2.pcmpeq.") ||
+ Name.startswith("x86.avx2.pcmpgt.") ||
+ Name.startswith("x86.avx.vpermil.") ||
+ Name == "x86.avx.movnt.dq.256" ||
+ Name == "x86.avx.movnt.pd.256" ||
+ Name == "x86.avx.movnt.ps.256" ||
+ (Name.startswith("x86.xop.vpcom") && F->arg_size() == 2)) {
+ NewFn = 0;
+ return true;
}
-
- // 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;
+ // SSE4.1 ptest functions may have an old signature.
+ if (Name.startswith("x86.sse41.ptest")) {
+ if (Name == "x86.sse41.ptestc")
+ return UpgradeSSE41Function(F, Intrinsic::x86_sse41_ptestc, NewFn);
+ if (Name == "x86.sse41.ptestz")
+ return UpgradeSSE41Function(F, Intrinsic::x86_sse41_ptestz, NewFn);
+ if (Name == "x86.sse41.ptestnzc")
+ return UpgradeSSE41Function(F, Intrinsic::x86_sse41_ptestnzc, NewFn);
}
-
- 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(FTy->getContext(), 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("");
-
- 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));
+ // frcz.ss/sd may need to have an argument dropped
+ if (Name.startswith("x86.xop.vfrcz.ss") && F->arg_size() == 2) {
+ F->setName(Name + ".old");
+ NewFn = Intrinsic::getDeclaration(F->getParent(),
+ Intrinsic::x86_xop_vfrcz_ss);
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;
+ }
+ if (Name.startswith("x86.xop.vfrcz.sd") && F->arg_size() == 2) {
+ F->setName(Name + ".old");
+ NewFn = Intrinsic::getDeclaration(F->getParent(),
+ Intrinsic::x86_xop_vfrcz_sd);
return true;
- } else if (Name.compare(5, 16, "x86.sse41.pmulld", 16) == 0) {
- // Calls to these intrinsics are transformed into vector multiplies.
- NewFn = 0;
+ }
+ // Fix the FMA4 intrinsics to remove the 4
+ if (Name.startswith("x86.fma4.")) {
+ F->setName("llvm.x86.fma" + Name.substr(8));
+ NewFn = F;
return true;
- } else if (Name.compare(5, 18, "x86.ssse3.palign.r", 18) == 0 ||
- Name.compare(5, 22, "x86.ssse3.palign.r.128", 22) == 0) {
- // Calls to these intrinsics are transformed into vector shuffles, shifts,
- // or 0.
- NewFn = 0;
- 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
- // upgraded form of the intrinsic. We should perhaps have two separate
+ // This may not belong here. This function is effectively being overloaded
+ // to both detect an intrinsic which needs upgrading, and to provide the
+ // upgraded form of the intrinsic. We should perhaps have two separate
// functions for this.
return false;
}
return Upgraded;
}
-// UpgradeIntrinsicCall - Upgrade a call to an old intrinsic to be a call the
-// upgraded intrinsic. All argument and return casting must be provided in
+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);
+ IRBuilder<> Builder(C);
+ Builder.SetInsertPoint(CI->getParent(), CI);
- assert(F && "CallInst has no function associated with it.");
+ assert(F && "Intrinsic call is not direct?");
if (!NewFn) {
// Get the Function's name.
- const std::string& Name = F->getName();
-
- // Upgrade ARM NEON intrinsics.
- if (Name.compare(5, 9, "arm.neon.", 9) == 0) {
- Instruction *NewI;
- if (Name.compare(14, 7, "vmovls.", 7) == 0) {
- NewI = new SExtInst(CI->getArgOperand(0), CI->getType(),
- "upgraded." + CI->getName(), CI);
- } else if (Name.compare(14, 7, "vmovlu.", 7) == 0) {
- NewI = new ZExtInst(CI->getArgOperand(0), CI->getType(),
- "upgraded." + CI->getName(), CI);
- } else {
- llvm_unreachable("Unknown arm.neon function for CallInst upgrade.");
- }
- // Replace any uses of the old CallInst.
- if (!CI->use_empty())
- CI->replaceAllUsesWith(NewI);
- CI->eraseFromParent();
- return;
- }
-
- bool isLoadH = false, isLoadL = false, isMovL = false;
- bool isMovSD = false, isShufPD = false;
- bool isUnpckhPD = false, isUnpcklPD = false;
- bool isPunpckhQPD = false, isPunpcklQPD = false;
- if (F->getName() == "llvm.x86.sse2.loadh.pd")
- isLoadH = true;
- else if (F->getName() == "llvm.x86.sse2.loadl.pd")
- isLoadL = true;
- else if (F->getName() == "llvm.x86.sse2.movl.dq")
- isMovL = true;
- else if (F->getName() == "llvm.x86.sse2.movs.d")
- isMovSD = true;
- else if (F->getName() == "llvm.x86.sse2.shuf.pd")
- isShufPD = true;
- else if (F->getName() == "llvm.x86.sse2.unpckh.pd")
- isUnpckhPD = true;
- else if (F->getName() == "llvm.x86.sse2.unpckl.pd")
- isUnpcklPD = true;
- else if (F->getName() == "llvm.x86.sse2.punpckh.qdq")
- isPunpckhQPD = true;
- else if (F->getName() == "llvm.x86.sse2.punpckl.qdq")
- isPunpcklQPD = true;
-
- if (isLoadH || isLoadL || isMovL || isMovSD || isShufPD ||
- isUnpckhPD || isUnpcklPD || isPunpckhQPD || isPunpcklQPD) {
- std::vector<Constant*> Idxs;
- Value *Op0 = CI->getArgOperand(0);
- ShuffleVectorInst *SI = NULL;
- if (isLoadH || isLoadL) {
- Value *Op1 = UndefValue::get(Op0->getType());
- Value *Addr = new BitCastInst(CI->getArgOperand(1),
- Type::getDoublePtrTy(C),
- "upgraded.", CI);
- Value *Load = new LoadInst(Addr, "upgraded.", false, 8, CI);
- Value *Idx = ConstantInt::get(Type::getInt32Ty(C), 0);
- Op1 = InsertElementInst::Create(Op1, Load, Idx, "upgraded.", CI);
-
- if (isLoadH) {
- Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 0));
- Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
- } else {
- Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
- Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
- }
- Value *Mask = ConstantVector::get(Idxs);
- SI = new ShuffleVectorInst(Op0, Op1, Mask, "upgraded.", CI);
- } else if (isMovL) {
- Constant *Zero = ConstantInt::get(Type::getInt32Ty(C), 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::getInt32Ty(C), 4));
- Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 5));
- Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
- Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 3));
- Value *Mask = ConstantVector::get(Idxs);
- SI = new ShuffleVectorInst(ZeroV, Op0, Mask, "upgraded.", CI);
- } else if (isMovSD ||
- isUnpckhPD || isUnpcklPD || isPunpckhQPD || isPunpcklQPD) {
- Value *Op1 = CI->getArgOperand(1);
- if (isMovSD) {
- Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
- Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
- } else if (isUnpckhPD || isPunpckhQPD) {
- Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
- Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 3));
- } else {
- Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 0));
- Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
- }
- Value *Mask = ConstantVector::get(Idxs);
- SI = new ShuffleVectorInst(Op0, Op1, Mask, "upgraded.", CI);
- } else if (isShufPD) {
- Value *Op1 = CI->getArgOperand(1);
- unsigned MaskVal =
- cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
- Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), MaskVal & 1));
- Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C),
- ((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 (!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->eraseFromParent();
- } else if (F->getName() == "llvm.x86.sse41.pmulld") {
- // Upgrade this set of intrinsics into vector multiplies.
- Instruction *Mul = BinaryOperator::CreateMul(CI->getArgOperand(0),
- CI->getArgOperand(1),
- CI->getName(),
- CI);
- // Fix up all the uses with our new multiply.
- if (!CI->use_empty())
- CI->replaceAllUsesWith(Mul);
-
- // Remove upgraded multiply.
- CI->eraseFromParent();
- } else if (F->getName() == "llvm.x86.ssse3.palign.r") {
- Value *Op1 = CI->getArgOperand(0);
- Value *Op2 = CI->getArgOperand(1);
- Value *Op3 = CI->getArgOperand(2);
- unsigned shiftVal = cast<ConstantInt>(Op3)->getZExtValue();
- Value *Rep;
+ StringRef Name = F->getName();
+
+ Value *Rep;
+ // Upgrade packed integer vector compares intrinsics to compare instructions
+ if (Name.startswith("llvm.x86.sse2.pcmpeq.") ||
+ Name.startswith("llvm.x86.avx2.pcmpeq.")) {
+ Rep = Builder.CreateICmpEQ(CI->getArgOperand(0), CI->getArgOperand(1),
+ "pcmpeq");
+ // need to sign extend since icmp returns vector of i1
+ Rep = Builder.CreateSExt(Rep, CI->getType(), "");
+ } else if (Name.startswith("llvm.x86.sse2.pcmpgt.") ||
+ Name.startswith("llvm.x86.avx2.pcmpgt.")) {
+ Rep = Builder.CreateICmpSGT(CI->getArgOperand(0), CI->getArgOperand(1),
+ "pcmpgt");
+ // need to sign extend since icmp returns vector of i1
+ Rep = Builder.CreateSExt(Rep, CI->getType(), "");
+ } else if (Name == "llvm.x86.avx.movnt.dq.256" ||
+ Name == "llvm.x86.avx.movnt.ps.256" ||
+ Name == "llvm.x86.avx.movnt.pd.256") {
IRBuilder<> Builder(C);
Builder.SetInsertPoint(CI->getParent(), CI);
- // If palignr is shifting the pair of input vectors less than 9 bytes,
- // emit a shuffle instruction.
- if (shiftVal <= 8) {
- const Type *IntTy = Type::getInt32Ty(C);
- const Type *EltTy = Type::getInt8Ty(C);
- const Type *VecTy = VectorType::get(EltTy, 8);
-
- Op2 = Builder.CreateBitCast(Op2, VecTy);
- Op1 = Builder.CreateBitCast(Op1, VecTy);
-
- llvm::SmallVector<llvm::Constant*, 8> Indices;
- for (unsigned i = 0; i != 8; ++i)
- Indices.push_back(ConstantInt::get(IntTy, shiftVal + i));
-
- Value *SV = ConstantVector::get(Indices.begin(), Indices.size());
- Rep = Builder.CreateShuffleVector(Op2, Op1, SV, "palignr");
- Rep = Builder.CreateBitCast(Rep, F->getReturnType());
- }
-
- // If palignr is shifting the pair of input vectors more than 8 but less
- // than 16 bytes, emit a logical right shift of the destination.
- else if (shiftVal < 16) {
- // MMX has these as 1 x i64 vectors for some odd optimization reasons.
- const Type *EltTy = Type::getInt64Ty(C);
- const Type *VecTy = VectorType::get(EltTy, 1);
+ Module *M = F->getParent();
+ SmallVector<Value *, 1> Elts;
+ Elts.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
+ MDNode *Node = MDNode::get(C, Elts);
- Op1 = Builder.CreateBitCast(Op1, VecTy, "cast");
- Op2 = ConstantInt::get(VecTy, (shiftVal-8) * 8);
-
- // create i32 constant
- Function *I =
- Intrinsic::getDeclaration(F->getParent(), Intrinsic::x86_mmx_psrl_q);
- Rep = Builder.CreateCall2(I, Op1, Op2, "palignr");
- }
+ Value *Arg0 = CI->getArgOperand(0);
+ Value *Arg1 = CI->getArgOperand(1);
- // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
- else {
- Rep = Constant::getNullValue(F->getReturnType());
- }
-
- // Replace any uses with our new instruction.
- if (!CI->use_empty())
- CI->replaceAllUsesWith(Rep);
-
- // Remove upgraded instruction.
- CI->eraseFromParent();
-
- } else if (F->getName() == "llvm.x86.ssse3.palign.r.128") {
- Value *Op1 = CI->getArgOperand(0);
- Value *Op2 = CI->getArgOperand(1);
- Value *Op3 = CI->getArgOperand(2);
- unsigned shiftVal = cast<ConstantInt>(Op3)->getZExtValue();
- Value *Rep;
- IRBuilder<> Builder(C);
- Builder.SetInsertPoint(CI->getParent(), CI);
+ // 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);
- // If palignr is shifting the pair of input vectors less than 17 bytes,
- // emit a shuffle instruction.
- if (shiftVal <= 16) {
- const Type *IntTy = Type::getInt32Ty(C);
- const Type *EltTy = Type::getInt8Ty(C);
- const Type *VecTy = VectorType::get(EltTy, 16);
-
- Op2 = Builder.CreateBitCast(Op2, VecTy);
- Op1 = Builder.CreateBitCast(Op1, VecTy);
-
- llvm::SmallVector<llvm::Constant*, 16> Indices;
- for (unsigned i = 0; i != 16; ++i)
- Indices.push_back(ConstantInt::get(IntTy, shiftVal + i));
-
- Value *SV = ConstantVector::get(Indices.begin(), Indices.size());
- Rep = Builder.CreateShuffleVector(Op2, Op1, SV, "palignr");
- Rep = Builder.CreateBitCast(Rep, F->getReturnType());
- }
-
- // If palignr is shifting the pair of input vectors more than 16 but less
- // than 32 bytes, emit a logical right shift of the destination.
- else if (shiftVal < 32) {
- const Type *EltTy = Type::getInt64Ty(C);
- const Type *VecTy = VectorType::get(EltTy, 2);
- const Type *IntTy = Type::getInt32Ty(C);
-
- Op1 = Builder.CreateBitCast(Op1, VecTy, "cast");
- Op2 = ConstantInt::get(IntTy, (shiftVal-16) * 8);
-
- // create i32 constant
- Function *I =
- Intrinsic::getDeclaration(F->getParent(), Intrinsic::x86_sse2_psrl_dq);
- Rep = Builder.CreateCall2(I, Op1, Op2, "palignr");
- }
-
- // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
- else {
- Rep = Constant::getNullValue(F->getReturnType());
- }
-
- // Replace any uses with our new instruction.
- if (!CI->use_empty())
- CI->replaceAllUsesWith(Rep);
-
- // Remove upgraded instruction.
+ // Remove intrinsic.
CI->eraseFromParent();
-
+ return;
+ } else if (Name.startswith("llvm.x86.xop.vpcom")) {
+ Intrinsic::ID intID;
+ if (Name.endswith("ub"))
+ intID = Intrinsic::x86_xop_vpcomub;
+ else if (Name.endswith("uw"))
+ intID = Intrinsic::x86_xop_vpcomuw;
+ else if (Name.endswith("ud"))
+ intID = Intrinsic::x86_xop_vpcomud;
+ else if (Name.endswith("uq"))
+ intID = Intrinsic::x86_xop_vpcomuq;
+ else if (Name.endswith("b"))
+ intID = Intrinsic::x86_xop_vpcomb;
+ else if (Name.endswith("w"))
+ intID = Intrinsic::x86_xop_vpcomw;
+ else if (Name.endswith("d"))
+ intID = Intrinsic::x86_xop_vpcomd;
+ else if (Name.endswith("q"))
+ intID = Intrinsic::x86_xop_vpcomq;
+ else
+ llvm_unreachable("Unknown suffix");
+
+ Name = Name.substr(18); // strip off "llvm.x86.xop.vpcom"
+ unsigned Imm;
+ if (Name.startswith("lt"))
+ Imm = 0;
+ else if (Name.startswith("le"))
+ Imm = 1;
+ else if (Name.startswith("gt"))
+ Imm = 2;
+ else if (Name.startswith("ge"))
+ Imm = 3;
+ else if (Name.startswith("eq"))
+ Imm = 4;
+ else if (Name.startswith("ne"))
+ Imm = 5;
+ else if (Name.startswith("true"))
+ Imm = 6;
+ else if (Name.startswith("false"))
+ Imm = 7;
+ else
+ llvm_unreachable("Unknown condition");
+
+ Function *VPCOM = Intrinsic::getDeclaration(F->getParent(), intID);
+ Rep = Builder.CreateCall3(VPCOM, CI->getArgOperand(0),
+ CI->getArgOperand(1), Builder.getInt8(Imm));
} else {
- llvm_unreachable("Unknown function for CallInst upgrade.");
+ bool PD128 = false, PD256 = false, PS128 = false, PS256 = false;
+ if (Name == "llvm.x86.avx.vpermil.pd.256")
+ PD256 = true;
+ else if (Name == "llvm.x86.avx.vpermil.pd")
+ PD128 = true;
+ else if (Name == "llvm.x86.avx.vpermil.ps.256")
+ PS256 = true;
+ else if (Name == "llvm.x86.avx.vpermil.ps")
+ PS128 = true;
+
+ if (PD256 || PD128 || PS256 || PS128) {
+ Value *Op0 = CI->getArgOperand(0);
+ unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
+ SmallVector<Constant*, 8> Idxs;
+
+ if (PD128)
+ for (unsigned i = 0; i != 2; ++i)
+ Idxs.push_back(Builder.getInt32((Imm >> i) & 0x1));
+ else if (PD256)
+ for (unsigned l = 0; l != 4; l+=2)
+ for (unsigned i = 0; i != 2; ++i)
+ Idxs.push_back(Builder.getInt32(((Imm >> (l+i)) & 0x1) + l));
+ else if (PS128)
+ for (unsigned i = 0; i != 4; ++i)
+ Idxs.push_back(Builder.getInt32((Imm >> (2 * i)) & 0x3));
+ else if (PS256)
+ for (unsigned l = 0; l != 8; l+=4)
+ for (unsigned i = 0; i != 4; ++i)
+ Idxs.push_back(Builder.getInt32(((Imm >> (2 * i)) & 0x3) + l));
+ else
+ llvm_unreachable("Unexpected function");
+
+ Rep = Builder.CreateShuffleVector(Op0, Op0, ConstantVector::get(Idxs));
+ } else {
+ llvm_unreachable("Unknown function for CallInst upgrade.");
+ }
}
+
+ CI->replaceAllUsesWith(Rep);
+ CI->eraseFromParent();
return;
}
+ std::string Name = CI->getName().str();
+ CI->setName(Name + ".old");
+
switch (NewFn->getIntrinsicID()) {
- default: llvm_unreachable("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->getArgOperand(0);
-
- // Cast the second parameter to the correct type.
- BitCastInst *BC = new BitCastInst(CI->getArgOperand(1),
- 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);
- 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
- // correct type.
- CI->replaceAllUsesWith(NewCI);
-
- // Clean up the old call now that it has been completely upgraded.
- CI->eraseFromParent();
- break;
- }
+ default:
+ llvm_unreachable("Unknown function for CallInst upgrade.");
+
case Intrinsic::ctlz:
- case Intrinsic::ctpop:
- case Intrinsic::cttz: {
- // Build a small vector of the original arguments.
- SmallVector<Value*, 8> Operands(CS.arg_begin(), CS.arg_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->getAttributes().paramHasAttr(0, Attribute::SExt);
- bool DestSExt = F->getAttributes().paramHasAttr(0, Attribute::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);
- }
+ case Intrinsic::cttz:
+ assert(CI->getNumArgOperands() == 1 &&
+ "Mismatch between function args and call args");
+ CI->replaceAllUsesWith(Builder.CreateCall2(NewFn, CI->getArgOperand(0),
+ Builder.getFalse(), Name));
+ CI->eraseFromParent();
+ return;
- // Clean up the old call now that it has been completely upgraded.
+ case Intrinsic::arm_neon_vclz: {
+ // Change name from llvm.arm.neon.vclz.* to llvm.ctlz.*
+ CI->replaceAllUsesWith(Builder.CreateCall2(NewFn, CI->getArgOperand(0),
+ Builder.getFalse(),
+ "llvm.ctlz." + Name.substr(14)));
CI->eraseFromParent();
+ return;
}
- break;
- case Intrinsic::eh_selector:
- case Intrinsic::eh_typeid_for: {
- // Only the return type changed.
- SmallVector<Value*, 8> Operands(CS.arg_begin(), CS.arg_end());
- 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()) {
- // Construct an appropriate cast from the new return type to the old.
- CastInst *RetCast =
- CastInst::Create(CastInst::getCastOpcode(NewCI, true,
- F->getReturnType(), true),
- NewCI, F->getReturnType(), NewCI->getName(), CI);
- CI->replaceAllUsesWith(RetCast);
- }
+ case Intrinsic::ctpop: {
+ CI->replaceAllUsesWith(Builder.CreateCall(NewFn, CI->getArgOperand(0)));
CI->eraseFromParent();
+ return;
}
- break;
- case Intrinsic::memcpy:
- case Intrinsic::memmove:
- case Intrinsic::memset: {
- // Add isVolatile
- const llvm::Type *I1Ty = llvm::Type::getInt1Ty(CI->getContext());
- Value *Operands[5] = { CI->getArgOperand(0), CI->getArgOperand(1),
- CI->getArgOperand(2), CI->getArgOperand(3),
- llvm::ConstantInt::get(I1Ty, 0) };
- CallInst *NewCI = CallInst::Create(NewFn, Operands, Operands+5,
- 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
- // correct type.
- CI->replaceAllUsesWith(NewCI);
-
- // Clean up the old call now that it has been completely upgraded.
+
+ case Intrinsic::x86_xop_vfrcz_ss:
+ case Intrinsic::x86_xop_vfrcz_sd:
+ CI->replaceAllUsesWith(Builder.CreateCall(NewFn, CI->getArgOperand(1),
+ Name));
CI->eraseFromParent();
- break;
+ return;
+
+ case Intrinsic::x86_sse41_ptestc:
+ case Intrinsic::x86_sse41_ptestz:
+ case Intrinsic::x86_sse41_ptestnzc: {
+ // The arguments for these intrinsics used to be v4f32, and changed
+ // to v2i64. This is purely a nop, since those are bitwise intrinsics.
+ // So, the only thing required is a bitcast for both arguments.
+ // First, check the arguments have the old type.
+ Value *Arg0 = CI->getArgOperand(0);
+ if (Arg0->getType() != VectorType::get(Type::getFloatTy(C), 4))
+ return;
+
+ // Old intrinsic, add bitcasts
+ Value *Arg1 = CI->getArgOperand(1);
+
+ Value *BC0 =
+ Builder.CreateBitCast(Arg0,
+ VectorType::get(Type::getInt64Ty(C), 2),
+ "cast");
+ Value *BC1 =
+ Builder.CreateBitCast(Arg1,
+ VectorType::get(Type::getInt64Ty(C), 2),
+ "cast");
+
+ CallInst* NewCall = Builder.CreateCall2(NewFn, BC0, BC1, Name);
+ CI->replaceAllUsesWith(NewCall);
+ CI->eraseFromParent();
+ return;
}
}
}
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()) {
- CallInst *CI = cast<CallInst>(FuncStart->use_back());
- CI->eraseFromParent();
- }
- FuncStart->eraseFromParent();
- }
-
- if (Function *StopPoint = M->getFunction("llvm.dbg.stoppoint")) {
- while (!StopPoint->use_empty()) {
- CallInst *CI = cast<CallInst>(StopPoint->use_back());
- CI->eraseFromParent();
- }
- StopPoint->eraseFromParent();
- }
-
- if (Function *RegionStart = M->getFunction("llvm.dbg.region.start")) {
- while (!RegionStart->use_empty()) {
- CallInst *CI = cast<CallInst>(RegionStart->use_back());
- CI->eraseFromParent();
- }
- RegionStart->eraseFromParent();
- }
-
- if (Function *RegionEnd = M->getFunction("llvm.dbg.region.end")) {
- while (!RegionEnd->use_empty()) {
- CallInst *CI = cast<CallInst>(RegionEnd->use_back());
- CI->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();
- }
- }
- }
-}