[avx] Add patterns for combining vextractf128 + vmovaps/vmovups/vmobdqu to

vextractf128 with 128-bit mem dest.

Combines

	vextractf128 $0, %ymm0, %xmm0
	vmovaps %xmm0, (%rdi)

to

    vextractf128 $0, %ymm0, (%rdi)

rdar://11082570


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@153139 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 106 insertions, 0 deletions

diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index 02205bb2c7..df426279ea 100644
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -7309,6 +7309,23 @@ def VEXTRACTF128mr : AVXAIi8<0x19, MRMDestMem, (outs),
           []>, VEX;
 }
 
+// Extract and store.
+let Predicates = [HasAVX] in {
+  def : Pat<(alignedstore (int_x86_avx_vextractf128_ps_256 VR256:$src1, imm:$src2), addr:$dst),
+          (VEXTRACTF128mr addr:$dst, VR256:$src1, imm:$src2)>;
+  def : Pat<(alignedstore (int_x86_avx_vextractf128_pd_256 VR256:$src1, imm:$src2), addr:$dst),
+          (VEXTRACTF128mr addr:$dst, VR256:$src1, imm:$src2)>;
+  def : Pat<(alignedstore (int_x86_avx_vextractf128_si_256 VR256:$src1, imm:$src2), addr:$dst),
+          (VEXTRACTF128mr addr:$dst, VR256:$src1, imm:$src2)>;
+
+  def : Pat<(int_x86_sse_storeu_ps addr:$dst, (int_x86_avx_vextractf128_ps_256 VR256:$src1, imm:$src2)),
+          (VEXTRACTF128mr addr:$dst, VR256:$src1, imm:$src2)>;
+  def : Pat<(int_x86_sse2_storeu_pd addr:$dst, (int_x86_avx_vextractf128_pd_256 VR256:$src1, imm:$src2)),
+          (VEXTRACTF128mr addr:$dst, VR256:$src1, imm:$src2)>;
+  def : Pat<(int_x86_sse2_storeu_dq addr:$dst, (bc_v16i8 (int_x86_avx_vextractf128_si_256 VR256:$src1, imm:$src2))),
+          (VEXTRACTF128mr addr:$dst, VR256:$src1, imm:$src2)>;
+}
+
 // AVX1 patterns
 let Predicates = [HasAVX] in {
 def : Pat<(int_x86_avx_vextractf128_pd_256 VR256:$src1, imm:$src2),
diff --git a/test/CodeGen/X86/avx-vextractf128.ll b/test/CodeGen/X86/avx-vextractf128.ll
index 6d3fbfd011..fe0f6caed3 100644
--- a/test/CodeGen/X86/avx-vextractf128.ll
+++ b/test/CodeGen/X86/avx-vextractf128.ll
@@ -1,5 +1,6 @@
 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=corei7-avx -mattr=+avx | FileCheck %s
 
+; CHECK: @A
 ; CHECK-NOT: vunpck
 ; CHECK: vextractf128 $1
 define <8 x float> @A(<8 x float> %a) nounwind uwtable readnone ssp {
@@ -8,6 +9,7 @@ entry:
   ret <8 x float> %shuffle
 }
 
+; CHECK: @B
 ; CHECK-NOT: vunpck
 ; CHECK: vextractf128 $1
 define <4 x double> @B(<4 x double> %a) nounwind uwtable readnone ssp {
@@ -15,3 +17,90 @@ entry:
   %shuffle = shufflevector <4 x double> %a, <4 x double> undef, <4 x i32> <i32 2, i32 3, i32 4, i32 4>
   ret <4 x double> %shuffle
 }
+
+; CHECK: @t0
+; CHECK-NOT: vextractf128 $0, %ymm0, %xmm0
+; CHECK-NOT: vmovaps %xmm0, (%rdi)
+; CHECK: vextractf128 $0, %ymm0, (%rdi)
+define void @t0(float* nocapture %addr, <8 x float> %a) nounwind uwtable ssp {
+entry:
+  %0 = tail call <4 x float> @llvm.x86.avx.vextractf128.ps.256(<8 x float> %a, i8 0)
+  %1 = bitcast float* %addr to <4 x float>*
+  store <4 x float> %0, <4 x float>* %1, align 16
+  ret void
+}
+
+declare <4 x float> @llvm.x86.avx.vextractf128.ps.256(<8 x float>, i8) nounwind readnone
+
+; CHECK: @t1
+; CHECK-NOT: vextractf128 $0, %ymm0, %xmm0
+; CHECK-NOT: vmovups %xmm0, (%rdi)
+; CHECK: vextractf128 $0, %ymm0, (%rdi)
+define void @t1(float* %addr, <8 x float> %a) nounwind uwtable ssp {
+entry:
+  %0 = tail call <4 x float> @llvm.x86.avx.vextractf128.ps.256(<8 x float> %a, i8 0)
+  %1 = bitcast float* %addr to i8*
+  tail call void @llvm.x86.sse.storeu.ps(i8* %1, <4 x float> %0)
+  ret void
+}
+
+declare void @llvm.x86.sse.storeu.ps(i8*, <4 x float>) nounwind
+
+; CHECK: @t2
+; CHECK-NOT: vextractf128 $0, %ymm0, %xmm0
+; CHECK-NOT: vmovaps %xmm0, (%rdi)
+; CHECK: vextractf128 $0, %ymm0, (%rdi)
+define void @t2(double* nocapture %addr, <4 x double> %a) nounwind uwtable ssp {
+entry:
+  %0 = tail call <2 x double> @llvm.x86.avx.vextractf128.pd.256(<4 x double> %a, i8 0)
+  %1 = bitcast double* %addr to <2 x double>*
+  store <2 x double> %0, <2 x double>* %1, align 16
+  ret void
+}
+
+declare <2 x double> @llvm.x86.avx.vextractf128.pd.256(<4 x double>, i8) nounwind readnone
+
+; CHECK: @t3
+; CHECK-NOT: vextractf128 $0, %ymm0, %xmm0
+; CHECK-NOT: vmovups %xmm0, (%rdi)
+; CHECK: vextractf128 $0, %ymm0, (%rdi)
+define void @t3(double* %addr, <4 x double> %a) nounwind uwtable ssp {
+entry:
+  %0 = tail call <2 x double> @llvm.x86.avx.vextractf128.pd.256(<4 x double> %a, i8 0)
+  %1 = bitcast double* %addr to i8*
+  tail call void @llvm.x86.sse2.storeu.pd(i8* %1, <2 x double> %0)
+  ret void
+}
+
+declare void @llvm.x86.sse2.storeu.pd(i8*, <2 x double>) nounwind
+
+; CHECK: @t4
+; CHECK-NOT: vextractf128 $0, %ymm0, %xmm0
+; CHECK-NOT: vmovaps %xmm0, (%rdi)
+; CHECK: vextractf128 $0, %ymm0, (%rdi)
+define void @t4(<2 x i64>* nocapture %addr, <4 x i64> %a) nounwind uwtable ssp {
+entry:
+  %0 = bitcast <4 x i64> %a to <8 x i32>
+  %1 = tail call <4 x i32> @llvm.x86.avx.vextractf128.si.256(<8 x i32> %0, i8 0)
+  %2 = bitcast <4 x i32> %1 to <2 x i64>
+  store <2 x i64> %2, <2 x i64>* %addr, align 16
+  ret void
+}
+
+declare <4 x i32> @llvm.x86.avx.vextractf128.si.256(<8 x i32>, i8) nounwind readnone
+
+; CHECK: @t5
+; CHECK-NOT: vextractf128 $0, %ymm0, %xmm0
+; CHECK-NOT: vmovdqu %xmm0, (%rdi)
+; CHECK: vextractf128 $0, %ymm0, (%rdi)
+define void @t5(<2 x i64>* %addr, <4 x i64> %a) nounwind uwtable ssp {
+entry:
+  %0 = bitcast <4 x i64> %a to <8 x i32>
+  %1 = tail call <4 x i32> @llvm.x86.avx.vextractf128.si.256(<8 x i32> %0, i8 0)
+  %2 = bitcast <2 x i64>* %addr to i8*
+  %3 = bitcast <4 x i32> %1 to <16 x i8>
+  tail call void @llvm.x86.sse2.storeu.dq(i8* %2, <16 x i8> %3)
+  ret void
+}
+
+declare void @llvm.x86.sse2.storeu.dq(i8*, <16 x i8>) nounwind