- Avoid using floating point stores to implement memset unless the value is zero.
- Do not try to infer GV alignment unless its type is sized. It's not possible to infer alignment if it has opaque type.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@100118 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/include/llvm/Target/TargetLowering.h b/include/llvm/Target/TargetLowering.h
index 23506c9..ad6fcef 100644
--- a/include/llvm/Target/TargetLowering.h
+++ b/include/llvm/Target/TargetLowering.h
@@ -638,7 +638,7 @@
/// determining it.
virtual EVT getOptimalMemOpType(uint64_t Size,
unsigned DstAlign, unsigned SrcAlign,
- SelectionDAG &DAG) const {
+ bool SafeToUseFP, SelectionDAG &DAG) const {
return MVT::Other;
}
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 25bf861..4fcfeec 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -3195,9 +3195,9 @@
/// is below the threshold. It returns the types of the sequence of
/// memory ops to perform memset / memcpy by reference.
static bool FindOptimalMemOpLowering(std::vector<EVT> &MemOps,
- SDValue Dst, SDValue Src,
unsigned Limit, uint64_t Size,
unsigned DstAlign, unsigned SrcAlign,
+ bool SafeToUseFP,
SelectionDAG &DAG,
const TargetLowering &TLI) {
assert((SrcAlign == 0 || SrcAlign >= DstAlign) &&
@@ -3207,7 +3207,7 @@
// the inferred alignment of the source. 'DstAlign', on the other hand, is the
// specified alignment of the memory operation. If it is zero, that means
// it's possible to change the alignment of the destination.
- EVT VT = TLI.getOptimalMemOpType(Size, DstAlign, SrcAlign, DAG);
+ EVT VT = TLI.getOptimalMemOpType(Size, DstAlign, SrcAlign, SafeToUseFP, DAG);
if (VT == MVT::Other) {
VT = TLI.getPointerTy();
@@ -3285,9 +3285,9 @@
std::string Str;
bool CopyFromStr = isMemSrcFromString(Src, Str);
bool isZeroStr = CopyFromStr && Str.empty();
- if (!FindOptimalMemOpLowering(MemOps, Dst, Src, Limit, Size,
+ if (!FindOptimalMemOpLowering(MemOps, Limit, Size,
(DstAlignCanChange ? 0 : Align),
- (isZeroStr ? 0 : SrcAlign), DAG, TLI))
+ (isZeroStr ? 0 : SrcAlign), true, DAG, TLI))
return SDValue();
if (DstAlignCanChange) {
@@ -3369,9 +3369,9 @@
if (Align > SrcAlign)
SrcAlign = Align;
- if (!FindOptimalMemOpLowering(MemOps, Dst, Src, Limit, Size,
+ if (!FindOptimalMemOpLowering(MemOps, Limit, Size,
(DstAlignCanChange ? 0 : Align),
- SrcAlign, DAG, TLI))
+ SrcAlign, true, DAG, TLI))
return SDValue();
if (DstAlignCanChange) {
@@ -3436,9 +3436,11 @@
FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Dst);
if (FI && !MFI->isFixedObjectIndex(FI->getIndex()))
DstAlignCanChange = true;
- if (!FindOptimalMemOpLowering(MemOps, Dst, Src, TLI.getMaxStoresPerMemset(),
+ bool IsZero = isa<ConstantSDNode>(Src) &&
+ cast<ConstantSDNode>(Src)->isNullValue();
+ if (!FindOptimalMemOpLowering(MemOps, TLI.getMaxStoresPerMemset(),
Size, (DstAlignCanChange ? 0 : Align), 0,
- DAG, TLI))
+ IsZero, DAG, TLI))
return SDValue();
if (DstAlignCanChange) {
@@ -6150,8 +6152,10 @@
unsigned Align = GV->getAlignment();
if (!Align) {
if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) {
- const TargetData *TD = TLI.getTargetData();
- Align = TD->getPreferredAlignment(GVar);
+ if (GV->getType()->getElementType()->isSized()) {
+ const TargetData *TD = TLI.getTargetData();
+ Align = TD->getPreferredAlignment(GVar);
+ }
}
}
return MinAlign(Align, GVOffset);
diff --git a/lib/Target/PowerPC/PPCISelLowering.cpp b/lib/Target/PowerPC/PPCISelLowering.cpp
index d00fbff..d043ec6 100644
--- a/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -5541,6 +5541,7 @@
EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
unsigned DstAlign, unsigned SrcAlign,
+ bool SafeToUseFP,
SelectionDAG &DAG) const {
if (this->PPCSubTarget.isPPC64()) {
return MVT::i64;
diff --git a/lib/Target/PowerPC/PPCISelLowering.h b/lib/Target/PowerPC/PPCISelLowering.h
index 2d5daef..bb2d84e 100644
--- a/lib/Target/PowerPC/PPCISelLowering.h
+++ b/lib/Target/PowerPC/PPCISelLowering.h
@@ -349,7 +349,7 @@
virtual EVT getOptimalMemOpType(uint64_t Size,
unsigned DstAlign, unsigned SrcAlign,
- SelectionDAG &DAG) const;
+ bool SafeToUseFP, SelectionDAG &DAG) const;
/// getFunctionAlignment - Return the Log2 alignment of this function.
virtual unsigned getFunctionAlignment(const Function *F) const;
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index bd268ec..5d6e571 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -1076,6 +1076,7 @@
EVT
X86TargetLowering::getOptimalMemOpType(uint64_t Size,
unsigned DstAlign, unsigned SrcAlign,
+ bool SafeToUseFP,
SelectionDAG &DAG) const {
// FIXME: This turns off use of xmm stores for memset/memcpy on targets like
// linux. This is because the stack realignment code can't handle certain
@@ -1089,9 +1090,10 @@
Subtarget->getStackAlignment() >= 16) {
if (Subtarget->hasSSE2())
return MVT::v4i32;
- if (Subtarget->hasSSE1())
+ if (SafeToUseFP && Subtarget->hasSSE1())
return MVT::v4f32;
- } else if (Size >= 8 &&
+ } else if (SafeToUseFP &&
+ Size >= 8 &&
Subtarget->getStackAlignment() >= 8 &&
Subtarget->hasSSE2())
return MVT::f64;
diff --git a/lib/Target/X86/X86ISelLowering.h b/lib/Target/X86/X86ISelLowering.h
index 569dc1f..cc7e964 100644
--- a/lib/Target/X86/X86ISelLowering.h
+++ b/lib/Target/X86/X86ISelLowering.h
@@ -425,7 +425,7 @@
/// determining it.
virtual EVT getOptimalMemOpType(uint64_t Size,
unsigned DstAlign, unsigned SrcAlign,
- SelectionDAG &DAG) const;
+ bool SafeToUseFP, SelectionDAG &DAG) const;
/// allowsUnalignedMemoryAccesses - Returns true if the target allows
/// unaligned memory accesses. of the specified type.
diff --git a/test/CodeGen/X86/memset-2.ll b/test/CodeGen/X86/memset-2.ll
index e2eba76..702632c 100644
--- a/test/CodeGen/X86/memset-2.ll
+++ b/test/CodeGen/X86/memset-2.ll
@@ -4,10 +4,18 @@
declare void @llvm.memset.i32(i8*, i8, i32, i32) nounwind
-define fastcc void @t() nounwind {
+define fastcc void @t1() nounwind {
entry:
-; CHECK: t:
+; CHECK: t1:
; CHECK: call memset
call void @llvm.memset.i32( i8* null, i8 0, i32 188, i32 1 ) nounwind
unreachable
}
+
+define fastcc void @t2(i8 signext %c) nounwind {
+entry:
+; CHECK: t2:
+; CHECK: call memset
+ call void @llvm.memset.i32( i8* undef, i8 %c, i32 76, i32 1 ) nounwind
+ unreachable
+}