Fix sdisel memcpy, memset, memmove lowering:
1. Makes it possible to lower with floating point loads and stores.
2. Avoid unaligned loads / stores unless it's fast.
3. Fix some memcpy lowering logic bug related to when to optimize a
load from constant string into a constant.
4. Adjust x86 memcpy lowering threshold to make it more sane.
5. Fix x86 target hook so it uses vector and floating point memory
ops more effectively.
rdar://7774704
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@100090 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index b94f76e..bd268ec 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -1012,7 +1012,7 @@
// FIXME: These should be based on subtarget info. Plus, the values should
// be smaller when we are in optimizing for size mode.
maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
- maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
+ maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
setPrefLoopAlignment(16);
benefitFromCodePlacementOpt = true;
@@ -1074,19 +1074,27 @@
/// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
/// determining it.
EVT
-X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
- bool isSrcConst, bool isSrcStr,
+X86TargetLowering::getOptimalMemOpType(uint64_t Size,
+ unsigned DstAlign, unsigned SrcAlign,
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
// cases like PR2962. This should be removed when PR2962 is fixed.
const Function *F = DAG.getMachineFunction().getFunction();
- bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
- if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
- if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
- return MVT::v4i32;
- if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
- return MVT::v4f32;
+ if (!F->hasFnAttr(Attribute::NoImplicitFloat)) {
+ if (Size >= 16 &&
+ (Subtarget->isUnalignedMemAccessFast() ||
+ (DstAlign == 0 || DstAlign >= 16) &&
+ (SrcAlign == 0 || SrcAlign >= 16)) &&
+ Subtarget->getStackAlignment() >= 16) {
+ if (Subtarget->hasSSE2())
+ return MVT::v4i32;
+ if (Subtarget->hasSSE1())
+ return MVT::v4f32;
+ } else if (Size >= 8 &&
+ Subtarget->getStackAlignment() >= 8 &&
+ Subtarget->hasSSE2())
+ return MVT::f64;
}
if (Subtarget->is64Bit() && Size >= 8)
return MVT::i64;
diff --git a/lib/Target/X86/X86ISelLowering.h b/lib/Target/X86/X86ISelLowering.h
index 46fa3ce..569dc1f 100644
--- a/lib/Target/X86/X86ISelLowering.h
+++ b/lib/Target/X86/X86ISelLowering.h
@@ -423,8 +423,8 @@
/// and store operations as a result of memset, memcpy, and memmove
/// lowering. It returns EVT::iAny if SelectionDAG should be responsible for
/// determining it.
- virtual EVT getOptimalMemOpType(uint64_t Size, unsigned Align,
- bool isSrcConst, bool isSrcStr,
+ virtual EVT getOptimalMemOpType(uint64_t Size,
+ unsigned DstAlign, unsigned SrcAlign,
SelectionDAG &DAG) const;
/// allowsUnalignedMemoryAccesses - Returns true if the target allows