}
break;
- case 'x':
+ case 'x':
+ // This fixes the poorly named crc32 intrinsics
+ if (Name.compare(5, 13, "x86.sse42.crc", 13) == 0) {
+ const char* NewFnName = NULL;
+ if (Name.compare(18, 2, "32", 2) == 0) {
+ if (Name.compare(20, 2, ".8") == 0 && Name.length() == 22) {
+ NewFnName = "llvm.x86.sse42.crc32.32.8";
+ } else if (Name.compare(20, 3, ".16") == 0 && Name.length() == 23) {
+ NewFnName = "llvm.x86.sse42.crc32.32.16";
+ } else if (Name.compare(20, 3, ".32") == 0 && Name.length() == 23) {
+ NewFnName = "llvm.x86.sse42.crc32.32.32";
+ }
+ }
+ else if (Name.compare(18, 2, "64", 2) == 0) {
+ if (Name.compare(20, 2, ".8") == 0 && Name.length() == 22) {
+ NewFnName = "llvm.x86.sse42.crc32.64.8";
+ } else if (Name.compare(20, 3, ".64") == 0 && Name.length() == 23) {
+ NewFnName = "llvm.x86.sse42.crc32.64.64";
+ }
+ }
+ if (NewFnName) {
+ F->setName(NewFnName);
+ NewFn = F;
+ return true;
+ }
+ }
+
// This fixes all MMX shift intrinsic instructions to take a
// x86_mmx instead of a v1i64, v2i32, v4i16, or v8i8.
if (Name.compare(5, 8, "x86.mmx.", 8) == 0) {
// or 0.
NewFn = 0;
return true;
+ } else if (Name.compare(5, 16, "x86.sse.loadu.ps", 16) == 0 ||
+ Name.compare(5, 17, "x86.sse2.loadu.dq", 17) == 0 ||
+ Name.compare(5, 17, "x86.sse2.loadu.pd", 17) == 0) {
+ // Calls to these instructions are transformed into unaligned loads.
+ NewFn = 0;
+ return true;
+ } else if (Name.compare(5, 16, "x86.sse.movnt.ps", 16) == 0 ||
+ Name.compare(5, 17, "x86.sse2.movnt.dq", 17) == 0 ||
+ Name.compare(5, 17, "x86.sse2.movnt.pd", 17) == 0 ||
+ Name.compare(5, 17, "x86.sse2.movnt.i", 16) == 0) {
+ // Calls to these instructions are transformed into nontemporal stores.
+ NewFn = 0;
+ return true;
} else if (Name.compare(5, 17, "x86.ssse3.pshuf.w", 17) == 0) {
// This is an SSE/MMX instruction.
const Type *X86_MMXTy = VectorType::getX86_MMXTy(FTy->getContext());
// Remove upgraded instruction.
CI->eraseFromParent();
-
+
+ } else 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.
+ const Type *VecTy = CI->getType();
+ const 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())
+ 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 {
llvm_unreachable("Unknown function for CallInst upgrade.");
}
break;
}
-#if 0
- case Intrinsic::x86_mmx_cvtsi32_si64: {
- // The return type needs to be changed.
- Value *Operands[1];
- Operands[0] = CI->getArgOperand(0);
- ConstructNewCallInst(NewFn, CI, Operands, 1);
- break;
- }
- case Intrinsic::x86_mmx_cvtsi64_si32: {
- Value *Operands[1];
-
- // Cast the operand to the X86 MMX type.
- Operands[0] = new BitCastInst(CI->getArgOperand(0),
- NewFn->getFunctionType()->getParamType(0),
- "upgraded.", CI);
-
- ConstructNewCallInst(NewFn, CI, Operands, 1);
- break;
- }
- case Intrinsic::x86_mmx_vec_init_b:
- case Intrinsic::x86_mmx_vec_init_w:
- case Intrinsic::x86_mmx_vec_init_d: {
- // The return type needs to be changed.
- Value *Operands[8];
- unsigned NumOps = 0;
-
- switch (NewFn->getIntrinsicID()) {
- default: break;
- case Intrinsic::x86_mmx_vec_init_b: NumOps = 8; break;
- case Intrinsic::x86_mmx_vec_init_w: NumOps = 4; break;
- case Intrinsic::x86_mmx_vec_init_d: NumOps = 2; break;
- }
-
- switch (NewFn->getIntrinsicID()) {
- default: break;
- case Intrinsic::x86_mmx_vec_init_b:
- Operands[7] = CI->getArgOperand(7);
- Operands[6] = CI->getArgOperand(6);
- Operands[5] = CI->getArgOperand(5);
- Operands[4] = CI->getArgOperand(4);
- // FALLTHRU
- case Intrinsic::x86_mmx_vec_init_w:
- Operands[3] = CI->getArgOperand(3);
- Operands[2] = CI->getArgOperand(2);
- // FALLTHRU
- case Intrinsic::x86_mmx_vec_init_d:
- Operands[1] = CI->getArgOperand(1);
- Operands[0] = CI->getArgOperand(0);
- break;
- }
-
- ConstructNewCallInst(NewFn, CI, Operands, NumOps);
- break;
- }
- case Intrinsic::x86_mmx_vec_ext_d: {
- Value *Operands[2];
-
- // Cast the operand to the X86 MMX type.
- Operands[0] = new BitCastInst(CI->getArgOperand(0),
- NewFn->getFunctionType()->getParamType(0),
- "upgraded.", CI);
- Operands[1] = CI->getArgOperand(1);
-
- ConstructNewCallInst(NewFn, CI, Operands, 2);
- break;
- }
-#endif
-
case Intrinsic::ctlz:
case Intrinsic::ctpop:
case Intrinsic::cttz: {
projects / oota-llvm.git / blobdiff