Reapply address space patch after fixing an issue in MemCopyOptimizer.
Added support for address spaces and added a isVolatile field to memcpy, memmove, and memset,
e.g., llvm.memcpy.i32(i8*, i8*, i32, i32) -> llvm.memcpy.p0i8.p0i8.i32(i8*, i8*, i32, i32, i1)
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@100304 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Transforms/Scalar/MemCpyOptimizer.cpp b/lib/Transforms/Scalar/MemCpyOptimizer.cpp
index 62e2977..3b305ae 100644
--- a/lib/Transforms/Scalar/MemCpyOptimizer.cpp
+++ b/lib/Transforms/Scalar/MemCpyOptimizer.cpp
@@ -413,7 +413,6 @@
// interesting as a small compile-time optimization.
Ranges.addStore(0, SI);
- Function *MemSetF = 0;
// Now that we have full information about ranges, loop over the ranges and
// emit memset's for anything big enough to be worthwhile.
@@ -433,29 +432,40 @@
// memset block. This ensure that the memset is dominated by any addressing
// instruction needed by the start of the block.
BasicBlock::iterator InsertPt = BI;
-
- if (MemSetF == 0) {
- const Type *Ty = Type::getInt64Ty(Context);
- MemSetF = Intrinsic::getDeclaration(M, Intrinsic::memset, &Ty, 1);
- }
-
+
// Get the starting pointer of the block.
StartPtr = Range.StartPtr;
-
+
+ // Determine alignment
+ unsigned Alignment = Range.Alignment;
+ if (Alignment == 0) {
+ const Type *EltType =
+ cast<PointerType>(StartPtr->getType())->getElementType();
+ Alignment = TD->getABITypeAlignment(EltType);
+ }
+
// Cast the start ptr to be i8* as memset requires.
- const Type *i8Ptr = Type::getInt8PtrTy(Context);
- if (StartPtr->getType() != i8Ptr)
+ const PointerType* StartPTy = cast<PointerType>(StartPtr->getType());
+ const PointerType *i8Ptr = Type::getInt8PtrTy(Context,
+ StartPTy->getAddressSpace());
+ if (StartPTy!= i8Ptr)
StartPtr = new BitCastInst(StartPtr, i8Ptr, StartPtr->getName(),
InsertPt);
-
+
Value *Ops[] = {
StartPtr, ByteVal, // Start, value
// size
ConstantInt::get(Type::getInt64Ty(Context), Range.End-Range.Start),
// align
- ConstantInt::get(Type::getInt32Ty(Context), Range.Alignment)
+ ConstantInt::get(Type::getInt32Ty(Context), Alignment),
+ // volatile
+ ConstantInt::get(Type::getInt1Ty(Context), 0),
};
- Value *C = CallInst::Create(MemSetF, Ops, Ops+4, "", InsertPt);
+ const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() };
+
+ Function *MemSetF = Intrinsic::getDeclaration(M, Intrinsic::memset, Tys, 2);
+
+ Value *C = CallInst::Create(MemSetF, Ops, Ops+5, "", InsertPt);
DEBUG(dbgs() << "Replace stores:\n";
for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
dbgs() << *Range.TheStores[i];
@@ -680,16 +690,19 @@
return false;
// If all checks passed, then we can transform these memcpy's
- const Type *Ty = M->getLength()->getType();
+ const Type *ArgTys[3] = { M->getRawDest()->getType(),
+ MDep->getRawSource()->getType(),
+ M->getLength()->getType() };
Function *MemCpyFun = Intrinsic::getDeclaration(
M->getParent()->getParent()->getParent(),
- M->getIntrinsicID(), &Ty, 1);
+ M->getIntrinsicID(), ArgTys, 3);
- Value *Args[4] = {
- M->getRawDest(), MDep->getRawSource(), M->getLength(), M->getAlignmentCst()
+ Value *Args[5] = {
+ M->getRawDest(), MDep->getRawSource(), M->getLength(),
+ M->getAlignmentCst(), M->getVolatileCst()
};
- CallInst *C = CallInst::Create(MemCpyFun, Args, Args+4, "", M);
+ CallInst *C = CallInst::Create(MemCpyFun, Args, Args+5, "", M);
// If C and M don't interfere, then this is a valid transformation. If they
@@ -728,8 +741,10 @@
// If not, then we know we can transform this.
Module *Mod = M->getParent()->getParent()->getParent();
- const Type *Ty = M->getLength()->getType();
- M->setOperand(0, Intrinsic::getDeclaration(Mod, Intrinsic::memcpy, &Ty, 1));
+ const Type *ArgTys[3] = { M->getRawDest()->getType(),
+ M->getRawSource()->getType(),
+ M->getLength()->getType() };
+ M->setOperand(0,Intrinsic::getDeclaration(Mod, Intrinsic::memcpy, ArgTys, 3));
// MemDep may have over conservative information about this instruction, just
// conservatively flush it from the cache.
diff --git a/lib/Transforms/Scalar/ScalarReplAggregates.cpp b/lib/Transforms/Scalar/ScalarReplAggregates.cpp
index bbe6270..6211beb 100644
--- a/lib/Transforms/Scalar/ScalarReplAggregates.cpp
+++ b/lib/Transforms/Scalar/ScalarReplAggregates.cpp
@@ -858,8 +858,17 @@
EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getName(), MI);
// Cast the other pointer (if we have one) to BytePtrTy.
- if (OtherElt && OtherElt->getType() != BytePtrTy)
- OtherElt = new BitCastInst(OtherElt, BytePtrTy, OtherElt->getName(), MI);
+ if (OtherElt && OtherElt->getType() != BytePtrTy) {
+ // Preserve address space of OtherElt
+ const PointerType* OtherPTy = cast<PointerType>(OtherElt->getType());
+ const PointerType* PTy = cast<PointerType>(BytePtrTy);
+ if (OtherPTy->getElementType() != PTy->getElementType()) {
+ Type *NewOtherPTy = PointerType::get(PTy->getElementType(),
+ OtherPTy->getAddressSpace());
+ OtherElt = new BitCastInst(OtherElt, NewOtherPTy,
+ OtherElt->getNameStr(), MI);
+ }
+ }
unsigned EltSize = TD->getTypeAllocSize(EltTy);
@@ -870,17 +879,28 @@
SROADest ? OtherElt : EltPtr, // Src ptr
ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
// Align
- ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign)
+ ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign),
+ MI->getVolatileCst()
};
- CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
+ // In case we fold the address space overloaded memcpy of A to B
+ // with memcpy of B to C, change the function to be a memcpy of A to C.
+ const Type *Tys[] = { Ops[0]->getType(), Ops[1]->getType(),
+ Ops[2]->getType() };
+ Module *M = MI->getParent()->getParent()->getParent();
+ TheFn = Intrinsic::getDeclaration(M, MI->getIntrinsicID(), Tys, 3);
+ CallInst::Create(TheFn, Ops, Ops + 5, "", MI);
} else {
assert(isa<MemSetInst>(MI));
Value *Ops[] = {
EltPtr, MI->getOperand(2), // Dest, Value,
ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
- Zero // Align
+ Zero, // Align
+ ConstantInt::get(Type::getInt1Ty(MI->getContext()), 0) // isVolatile
};
- CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
+ const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() };
+ Module *M = MI->getParent()->getParent()->getParent();
+ TheFn = Intrinsic::getDeclaration(M, Intrinsic::memset, Tys, 2);
+ CallInst::Create(TheFn, Ops, Ops + 5, "", MI);
}
}
DeadInsts.push_back(MI);
diff --git a/lib/Transforms/Scalar/SimplifyLibCalls.cpp b/lib/Transforms/Scalar/SimplifyLibCalls.cpp
index 5941ea6..b053cfc 100644
--- a/lib/Transforms/Scalar/SimplifyLibCalls.cpp
+++ b/lib/Transforms/Scalar/SimplifyLibCalls.cpp
@@ -142,7 +142,8 @@
// We have enough information to now generate the memcpy call to do the
// concatenation for us. Make a memcpy to copy the nul byte with align = 1.
EmitMemCpy(CpyDst, Src,
- ConstantInt::get(TD->getIntPtrType(*Context), Len+1), 1, B, TD);
+ ConstantInt::get(TD->getIntPtrType(*Context), Len+1),
+ 1, false, B, TD);
}
};
@@ -383,7 +384,8 @@
CI->getOperand(3), B, TD);
else
EmitMemCpy(Dst, Src,
- ConstantInt::get(TD->getIntPtrType(*Context), Len), 1, B, TD);
+ ConstantInt::get(TD->getIntPtrType(*Context), Len),
+ 1, false, B, TD);
return Dst;
}
};
@@ -411,8 +413,8 @@
if (SrcLen == 0) {
// strncpy(x, "", y) -> memset(x, '\0', y, 1)
- EmitMemSet(Dst, ConstantInt::get(Type::getInt8Ty(*Context), '\0'), LenOp,
- B, TD);
+ EmitMemSet(Dst, ConstantInt::get(Type::getInt8Ty(*Context), '\0'),
+ LenOp, false, B, TD);
return Dst;
}
@@ -432,7 +434,8 @@
// strncpy(x, s, c) -> memcpy(x, s, c, 1) [s and c are constant]
EmitMemCpy(Dst, Src,
- ConstantInt::get(TD->getIntPtrType(*Context), Len), 1, B, TD);
+ ConstantInt::get(TD->getIntPtrType(*Context), Len),
+ 1, false, B, TD);
return Dst;
}
@@ -593,7 +596,7 @@
// memcpy(x, y, n) -> llvm.memcpy(x, y, n, 1)
EmitMemCpy(CI->getOperand(1), CI->getOperand(2),
- CI->getOperand(3), 1, B, TD);
+ CI->getOperand(3), 1, false, B, TD);
return CI->getOperand(1);
}
};
@@ -615,7 +618,7 @@
// memmove(x, y, n) -> llvm.memmove(x, y, n, 1)
EmitMemMove(CI->getOperand(1), CI->getOperand(2),
- CI->getOperand(3), 1, B, TD);
+ CI->getOperand(3), 1, false, B, TD);
return CI->getOperand(1);
}
};
@@ -637,8 +640,8 @@
// memset(p, v, n) -> llvm.memset(p, v, n, 1)
Value *Val = B.CreateIntCast(CI->getOperand(2), Type::getInt8Ty(*Context),
- false);
- EmitMemSet(CI->getOperand(1), Val, CI->getOperand(3), B, TD);
+ false);
+ EmitMemSet(CI->getOperand(1), Val, CI->getOperand(3), false, B, TD);
return CI->getOperand(1);
}
};
@@ -999,7 +1002,7 @@
// sprintf(str, fmt) -> llvm.memcpy(str, fmt, strlen(fmt)+1, 1)
EmitMemCpy(CI->getOperand(1), CI->getOperand(2), // Copy the nul byte.
ConstantInt::get(TD->getIntPtrType(*Context),
- FormatStr.size()+1), 1, B, TD);
+ FormatStr.size()+1), 1, false, B, TD);
return ConstantInt::get(CI->getType(), FormatStr.size());
}
@@ -1013,11 +1016,11 @@
// sprintf(dst, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0
if (!CI->getOperand(3)->getType()->isIntegerTy()) return 0;
Value *V = B.CreateTrunc(CI->getOperand(3),
- Type::getInt8Ty(*Context), "char");
+ Type::getInt8Ty(*Context), "char");
Value *Ptr = CastToCStr(CI->getOperand(1), B);
B.CreateStore(V, Ptr);
Ptr = B.CreateGEP(Ptr, ConstantInt::get(Type::getInt32Ty(*Context), 1),
- "nul");
+ "nul");
B.CreateStore(Constant::getNullValue(Type::getInt8Ty(*Context)), Ptr);
return ConstantInt::get(CI->getType(), 1);
@@ -1034,7 +1037,7 @@
Value *IncLen = B.CreateAdd(Len,
ConstantInt::get(Len->getType(), 1),
"leninc");
- EmitMemCpy(CI->getOperand(1), CI->getOperand(3), IncLen, 1, B, TD);
+ EmitMemCpy(CI->getOperand(1), CI->getOperand(3), IncLen, 1, false, B, TD);
// The sprintf result is the unincremented number of bytes in the string.
return B.CreateIntCast(Len, CI->getType(), false);