- Supposely movlhps is faster / better than unpcklpd.
- Don't forget pshufd is only available with sse2.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26956 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index d26d787..bea938a 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -2242,7 +2242,7 @@
return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
DAG.getNode(ISD::UNDEF, V1.getValueType()),
PermMask);
- } else if (X86::isPSHUFDMask(PermMask.Val)) {
+ } else if (Subtarget->hasSSE2() && X86::isPSHUFDMask(PermMask.Val)) {
if (V2.getOpcode() == ISD::UNDEF)
// Leave the VECTOR_SHUFFLE alone. It matches PSHUFD.
return SDOperand();
@@ -2375,5 +2375,6 @@
/// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
/// are assumed to be legal.
bool X86TargetLowering::isShuffleMaskLegal(SDOperand Mask) const {
- return (X86::isSplatMask(Mask.Val) || X86::isPSHUFDMask(Mask.Val));
+ return (X86::isSplatMask(Mask.Val) ||
+ (Subtarget->hasSSE2() && X86::isPSHUFDMask(Mask.Val)));
}
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index fe8bc94..7155b8e 100644
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -55,7 +55,7 @@
return X86::isSplatMask(N);
}], SHUFFLE_get_shuf_imm>;
-def UNPCKLP_splat_mask : PatLeaf<(build_vector), [{
+def MOVLHPS_splat_mask : PatLeaf<(build_vector), [{
return X86::isSplatMask(N);
}]>;
@@ -810,18 +810,22 @@
// Splat v4f32 / v4i32
def : Pat<(vector_shuffle (v4f32 VR128:$src), (undef), SHUFP_splat_mask:$sm),
- (v4f32 (SHUFPSrr VR128:$src, VR128:$src, SHUFP_splat_mask:$sm))>;
+ (v4f32 (SHUFPSrr VR128:$src, VR128:$src, SHUFP_splat_mask:$sm))>,
+ Requires<[HasSSE1]>;
def : Pat<(vector_shuffle (v4i32 VR128:$src), (undef), SHUFP_splat_mask:$sm),
- (v4i32 (SHUFPSrr VR128:$src, VR128:$src, SHUFP_splat_mask:$sm))>;
+ (v4i32 (SHUFPSrr VR128:$src, VR128:$src, SHUFP_splat_mask:$sm))>,
+ Requires<[HasSSE1]>;
// Splat v2f64 / v2i64
-def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), UNPCKLP_splat_mask:$sm),
- (v2f64 (UNPCKLPDrr VR128:$src, VR128:$src))>;
-def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), UNPCKLP_splat_mask:$sm),
- (v2i64 (UNPCKLPDrr VR128:$src, VR128:$src))>;
+def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), MOVLHPS_splat_mask:$sm),
+ (v2f64 (MOVLHPSrr VR128:$src))>, Requires<[HasSSE1]>;
+def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), MOVLHPS_splat_mask:$sm),
+ (v2i64 (MOVLHPSrr VR128:$src))>, Requires<[HasSSE1]>;
// Shuffle v4f32 / v4i32, undef. These should only match if splat cases do not.
def : Pat<(vector_shuffle (v4f32 VR128:$src), (undef), PSHUFD_shuffle_mask:$sm),
- (v4f32 (PSHUFDrr VR128:$src, PSHUFD_shuffle_mask:$sm))>;
+ (v4f32 (PSHUFDrr VR128:$src, PSHUFD_shuffle_mask:$sm))>,
+ Requires<[HasSSE2]>;
def : Pat<(vector_shuffle (v4i32 VR128:$src), (undef), PSHUFD_shuffle_mask:$sm),
- (v4i32 (PSHUFDrr VR128:$src, PSHUFD_shuffle_mask:$sm))>;
+ (v4i32 (PSHUFDrr VR128:$src, PSHUFD_shuffle_mask:$sm))>,
+ Requires<[HasSSE2]>;