[X86][AVX] Shuffle->Permute+Blend if we have one v4f64/v4i64 shuffle input in place
Even on AVX1 we can pretty cheaply (VPERM2F128+VSHUFPD) permute a single v4f64/v4i64 input (on AVX2 its just a single VPERMPD), followed by a BLENDPD.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@354729 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 6ad5503..957987f 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -15024,6 +15024,12 @@
if (SDValue Op = lowerShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
return Op;
+ // If we have one input in place, then we can permute the other input and
+ // blend the result.
+ if (isShuffleMaskInputInPlace(0, Mask) || isShuffleMaskInputInPlace(1, Mask))
+ return lowerShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2, Mask,
+ Subtarget, DAG);
+
// Try to create an in-lane repeating shuffle mask and then shuffle the
// results into the target lanes.
if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute(
@@ -15128,6 +15134,12 @@
if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v4i64, Mask, V1, V2, DAG))
return V;
+ // If we have one input in place, then we can permute the other input and
+ // blend the result.
+ if (isShuffleMaskInputInPlace(0, Mask) || isShuffleMaskInputInPlace(1, Mask))
+ return lowerShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2, Mask,
+ Subtarget, DAG);
+
// Try to create an in-lane repeating shuffle mask and then shuffle the
// results into the target lanes.
if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute(
diff --git a/test/CodeGen/X86/avx512-shuffles/partial_permute.ll b/test/CodeGen/X86/avx512-shuffles/partial_permute.ll
index 8ac883d..bda7f7b 100644
--- a/test/CodeGen/X86/avx512-shuffles/partial_permute.ll
+++ b/test/CodeGen/X86/avx512-shuffles/partial_permute.ll
@@ -4331,12 +4331,11 @@
define <4 x double> @test_masked_8xdouble_to_4xdouble_perm_mem_mask7(<8 x double>* %vp, <4 x double> %vec2, <4 x double> %mask) {
; CHECK-LABEL: test_masked_8xdouble_to_4xdouble_perm_mem_mask7:
; CHECK: # %bb.0:
-; CHECK-NEXT: vmovapd (%rdi), %ymm2
-; CHECK-NEXT: vmovapd {{.*#+}} ymm3 = [0,5,2,5]
-; CHECK-NEXT: vpermi2pd 32(%rdi), %ymm2, %ymm3
-; CHECK-NEXT: vxorpd %xmm2, %xmm2, %xmm2
-; CHECK-NEXT: vcmpeqpd %ymm2, %ymm1, %k1
-; CHECK-NEXT: vmovapd %ymm3, %ymm0 {%k1}
+; CHECK-NEXT: vbroadcastsd 40(%rdi), %ymm2
+; CHECK-NEXT: vblendpd {{.*#+}} ymm2 = mem[0],ymm2[1],mem[2],ymm2[3]
+; CHECK-NEXT: vxorpd %xmm3, %xmm3, %xmm3
+; CHECK-NEXT: vcmpeqpd %ymm3, %ymm1, %k1
+; CHECK-NEXT: vmovapd %ymm2, %ymm0 {%k1}
; CHECK-NEXT: retq
%vec = load <8 x double>, <8 x double>* %vp
%shuf = shufflevector <8 x double> %vec, <8 x double> undef, <4 x i32> <i32 0, i32 5, i32 2, i32 5>
@@ -4348,12 +4347,11 @@
define <4 x double> @test_masked_z_8xdouble_to_4xdouble_perm_mem_mask7(<8 x double>* %vp, <4 x double> %mask) {
; CHECK-LABEL: test_masked_z_8xdouble_to_4xdouble_perm_mem_mask7:
; CHECK: # %bb.0:
-; CHECK-NEXT: vmovapd (%rdi), %ymm2
-; CHECK-NEXT: vmovapd {{.*#+}} ymm1 = [0,5,2,5]
-; CHECK-NEXT: vxorpd %xmm3, %xmm3, %xmm3
-; CHECK-NEXT: vcmpeqpd %ymm3, %ymm0, %k1
-; CHECK-NEXT: vpermi2pd 32(%rdi), %ymm2, %ymm1 {%k1} {z}
-; CHECK-NEXT: vmovapd %ymm1, %ymm0
+; CHECK-NEXT: vbroadcastsd 40(%rdi), %ymm1
+; CHECK-NEXT: vblendpd {{.*#+}} ymm1 = mem[0],ymm1[1],mem[2],ymm1[3]
+; CHECK-NEXT: vxorpd %xmm2, %xmm2, %xmm2
+; CHECK-NEXT: vcmpeqpd %ymm2, %ymm0, %k1
+; CHECK-NEXT: vmovapd %ymm1, %ymm0 {%k1} {z}
; CHECK-NEXT: retq
%vec = load <8 x double>, <8 x double>* %vp
%shuf = shufflevector <8 x double> %vec, <8 x double> undef, <4 x i32> <i32 0, i32 5, i32 2, i32 5>
diff --git a/test/CodeGen/X86/subvector-broadcast.ll b/test/CodeGen/X86/subvector-broadcast.ll
index 0e66f60..f1a79e6 100644
--- a/test/CodeGen/X86/subvector-broadcast.ll
+++ b/test/CodeGen/X86/subvector-broadcast.ll
@@ -1662,44 +1662,40 @@
; X32-AVX1-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-AVX1-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
; X32-AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1
-; X32-AVX1-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7]
+; X32-AVX1-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; X32-AVX1-NEXT: retl
;
; X32-AVX2-LABEL: broadcast_v8f32_v2f32_u1uu0uEu:
; X32-AVX2: # %bb.0:
; X32-AVX2-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X32-AVX2-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
-; X32-AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7]
-; X32-AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
+; X32-AVX2-NEXT: vbroadcastsd (%eax), %ymm1
+; X32-AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; X32-AVX2-NEXT: retl
;
; X32-AVX512-LABEL: broadcast_v8f32_v2f32_u1uu0uEu:
; X32-AVX512: # %bb.0:
; X32-AVX512-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X32-AVX512-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
-; X32-AVX512-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7]
-; X32-AVX512-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
+; X32-AVX512-NEXT: vbroadcastsd (%eax), %ymm1
+; X32-AVX512-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; X32-AVX512-NEXT: retl
;
; X64-AVX1-LABEL: broadcast_v8f32_v2f32_u1uu0uEu:
; X64-AVX1: # %bb.0:
; X64-AVX1-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
; X64-AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1
-; X64-AVX1-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7]
+; X64-AVX1-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; X64-AVX1-NEXT: retq
;
; X64-AVX2-LABEL: broadcast_v8f32_v2f32_u1uu0uEu:
; X64-AVX2: # %bb.0:
-; X64-AVX2-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
-; X64-AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7]
-; X64-AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
+; X64-AVX2-NEXT: vbroadcastsd (%rdi), %ymm1
+; X64-AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; X64-AVX2-NEXT: retq
;
; X64-AVX512-LABEL: broadcast_v8f32_v2f32_u1uu0uEu:
; X64-AVX512: # %bb.0:
-; X64-AVX512-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
-; X64-AVX512-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7]
-; X64-AVX512-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,3]
+; X64-AVX512-NEXT: vbroadcastsd (%rdi), %ymm1
+; X64-AVX512-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; X64-AVX512-NEXT: retq
%vec = load <2 x float>, <2 x float>* %vp
%shuf = shufflevector <2 x float> %vec, <2 x float> undef, <8 x i32> <i32 undef, i32 1, i32 undef, i32 undef, i32 0, i32 2, i32 3, i32 undef>
