[X86] Lower vselect with constant condition to vector_shuffle even with AVX512 instructions.

Summary:
It's better to use our shuffle lowering code to handle these than loading an immediate into a k-register.

It really feels like this should be a DAG combine optimization rather than a lowering operation, but that's a problem for another day.

Reviewers: RKSimon, delena, zvi

Reviewed By: delena

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D38932

llvm-svn: 315849

4 files changed, 14 insertions, 33 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index c08d79663fa..f155de8ef02 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -14077,16 +14077,16 @@ SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
       ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
     return SDValue();
 
-  // If this VSELECT has a vector if i1 as a mask, it will be directly matched
-  // with patterns on the mask registers on AVX-512.
-  if (Op->getOperand(0).getValueType().getScalarSizeInBits() == 1)
-    return Op;
-
   // Try to lower this to a blend-style vector shuffle. This can handle all
   // constant condition cases.
   if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
     return BlendOp;
 
+  // If this VSELECT has a vector if i1 as a mask, it will be directly matched
+  // with patterns on the mask registers on AVX-512.
+  if (Op->getOperand(0).getValueType().getScalarSizeInBits() == 1)
+    return Op;
+
   // Variable blends are only legal from SSE4.1 onward.
   if (!Subtarget.hasSSE41())
     return SDValue();
diff --git a/llvm/test/CodeGen/X86/avx512-mask-op.ll b/llvm/test/CodeGen/X86/avx512-mask-op.ll
index d44c24071bb..30aa4f44503 100644
--- a/llvm/test/CodeGen/X86/avx512-mask-op.ll
+++ b/llvm/test/CodeGen/X86/avx512-mask-op.ll
@@ -1700,16 +1700,12 @@ define <64 x i8> @test_build_vec_v64i1(<64 x i8> %x) {
 ;
 ; SKX-LABEL: test_build_vec_v64i1:
 ; SKX:       ## BB#0:
-; SKX-NEXT:    movabsq $6432645796886517060, %rax ## imm = 0x5945594549549544
-; SKX-NEXT:    kmovq %rax, %k1
-; SKX-NEXT:    vmovdqu8 %zmm0, %zmm0 {%k1} {z}
+; SKX-NEXT:    vpshufb {{.*#+}} zmm0 = zero,zero,zmm0[2],zero,zero,zero,zmm0[6],zero,zmm0[8],zero,zmm0[10],zero,zmm0[12],zero,zero,zmm0[15],zero,zero,zmm0[18],zero,zmm0[20],zero,zmm0[22],zero,zmm0[24],zero,zero,zmm0[27],zero,zero,zmm0[30],zero,zmm0[32],zero,zmm0[34],zero,zero,zero,zmm0[38],zero,zmm0[40],zero,zero,zmm0[43,44],zero,zmm0[46],zero,zmm0[48],zero,zmm0[50],zero,zero,zero,zmm0[54],zero,zmm0[56],zero,zero,zmm0[59,60],zero,zmm0[62],zero
 ; SKX-NEXT:    retq
 ;
 ; AVX512BW-LABEL: test_build_vec_v64i1:
 ; AVX512BW:       ## BB#0:
-; AVX512BW-NEXT:    movabsq $6432645796886517060, %rax ## imm = 0x5945594549549544
-; AVX512BW-NEXT:    kmovq %rax, %k1
-; AVX512BW-NEXT:    vmovdqu8 %zmm0, %zmm0 {%k1} {z}
+; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm0 = zero,zero,zmm0[2],zero,zero,zero,zmm0[6],zero,zmm0[8],zero,zmm0[10],zero,zmm0[12],zero,zero,zmm0[15],zero,zero,zmm0[18],zero,zmm0[20],zero,zmm0[22],zero,zmm0[24],zero,zero,zmm0[27],zero,zero,zmm0[30],zero,zmm0[32],zero,zmm0[34],zero,zero,zero,zmm0[38],zero,zmm0[40],zero,zero,zmm0[43,44],zero,zmm0[46],zero,zmm0[48],zero,zmm0[50],zero,zero,zero,zmm0[54],zero,zmm0[56],zero,zero,zmm0[59,60],zero,zmm0[62],zero
 ; AVX512BW-NEXT:    retq
 ;
 ; AVX512DQ-LABEL: test_build_vec_v64i1:
diff --git a/llvm/test/CodeGen/X86/avx512-schedule.ll b/llvm/test/CodeGen/X86/avx512-schedule.ll
index 3c9ea6acbd1..a546dd0eb24 100755
--- a/llvm/test/CodeGen/X86/avx512-schedule.ll
+++ b/llvm/test/CodeGen/X86/avx512-schedule.ll
@@ -5307,10 +5307,7 @@ define <32 x i16> @test_build_vec_v32i1(<32 x i16> %x) {
 define <64 x i8> @test_build_vec_v64i1(<64 x i8> %x) {
 ; CHECK-LABEL: test_build_vec_v64i1:
 ; CHECK:       # BB#0:
-; CHECK-NEXT:    movabsq $6432645796886517060, %rax # imm = 0x5945594549549544
-; CHECK-NEXT:    # sched: [1:0.25]
-; CHECK-NEXT:    kmovq %rax, %k1 # sched: [1:1.00]
-; CHECK-NEXT:    vmovdqu8 %zmm0, %zmm0 {%k1} {z}
+; CHECK-NEXT:    vpshufb {{.*#+}} zmm0 = zero,zero,zmm0[2],zero,zero,zero,zmm0[6],zero,zmm0[8],zero,zmm0[10],zero,zmm0[12],zero,zero,zmm0[15],zero,zero,zmm0[18],zero,zmm0[20],zero,zmm0[22],zero,zmm0[24],zero,zero,zmm0[27],zero,zero,zmm0[30],zero,zmm0[32],zero,zmm0[34],zero,zero,zero,zmm0[38],zero,zmm0[40],zero,zero,zmm0[43,44],zero,zmm0[46],zero,zmm0[48],zero,zmm0[50],zero,zero,zero,zmm0[54],zero,zmm0[56],zero,zero,zmm0[59,60],zero,zmm0[62],zero sched: [8:1.00]
 ; CHECK-NEXT:    retq # sched: [7:1.00]
   %ret = select <64 x i1> <i1 false, i1 false, i1 true, i1 false, i1 false, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 true, i1 false, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 true, i1 false, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 true, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 true, i1 true, i1 false, i1 true, i1 false>, <64 x i8> %x, <64 x i8> zeroinitializer
   ret <64 x i8> %ret
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx2.ll b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx2.ll
index f0c7ae38b6b..e7ad4aca204 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx2.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx2.ll
@@ -985,15 +985,9 @@ define internal fastcc <8 x float> @PR34577(<8 x float> %inp0, <8 x float> %inp1
 ;
 ; X32-AVX512-LABEL: PR34577:
 ; X32-AVX512:       # BB#0: # %entry
-; X32-AVX512-NEXT:    vmovaps {{.*#+}} ymm3 = <1,u,u,u,2,u,5,0>
-; X32-AVX512-NEXT:    vpermps %ymm0, %ymm3, %ymm0
-; X32-AVX512-NEXT:    vmovaps {{.*#+}} ymm3 = <u,2,3,5,u,5,u,u>
-; X32-AVX512-NEXT:    vpermps %ymm2, %ymm3, %ymm2
-; X32-AVX512-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0],ymm2[1,2,3],ymm0[4],ymm2[5],ymm0[6,7]
-; X32-AVX512-NEXT:    vxorps %xmm2, %xmm2, %xmm2
-; X32-AVX512-NEXT:    movb $86, %al
-; X32-AVX512-NEXT:    kmovw %eax, %k1
-; X32-AVX512-NEXT:    vblendmps %zmm0, %zmm2, %zmm0 {%k1}
+; X32-AVX512-NEXT:    vmovapd {{.*#+}} ymm2 = <1,u,u,u,2,u,5,0>
+; X32-AVX512-NEXT:    vpermps %ymm0, %ymm2, %ymm0
+; X32-AVX512-NEXT:    vxorpd %xmm2, %xmm2, %xmm2
 ; X32-AVX512-NEXT:    vblendpd {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3]
 ; X32-AVX512-NEXT:    vmovapd {{.*#+}} ymm2 = <u,u,7,2,u,u,3,2>
 ; X32-AVX512-NEXT:    vpermps %ymm1, %ymm2, %ymm1
@@ -1012,15 +1006,9 @@ define internal fastcc <8 x float> @PR34577(<8 x float> %inp0, <8 x float> %inp1
 ;
 ; X64-AVX512-LABEL: PR34577:
 ; X64-AVX512:       # BB#0: # %entry
-; X64-AVX512-NEXT:    vmovaps {{.*#+}} ymm3 = <1,u,u,u,2,u,5,0>
-; X64-AVX512-NEXT:    vpermps %ymm0, %ymm3, %ymm0
-; X64-AVX512-NEXT:    vmovaps {{.*#+}} ymm3 = <u,2,3,5,u,5,u,u>
-; X64-AVX512-NEXT:    vpermps %ymm2, %ymm3, %ymm2
-; X64-AVX512-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0],ymm2[1,2,3],ymm0[4],ymm2[5],ymm0[6,7]
-; X64-AVX512-NEXT:    vxorps %xmm2, %xmm2, %xmm2
-; X64-AVX512-NEXT:    movb $86, %al
-; X64-AVX512-NEXT:    kmovw %eax, %k1
-; X64-AVX512-NEXT:    vblendmps %zmm0, %zmm2, %zmm0 {%k1}
+; X64-AVX512-NEXT:    vmovapd {{.*#+}} ymm2 = <1,u,u,u,2,u,5,0>
+; X64-AVX512-NEXT:    vpermps %ymm0, %ymm2, %ymm0
+; X64-AVX512-NEXT:    vxorpd %xmm2, %xmm2, %xmm2
 ; X64-AVX512-NEXT:    vblendpd {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3]
 ; X64-AVX512-NEXT:    vmovapd {{.*#+}} ymm2 = <u,u,7,2,u,u,3,2>
 ; X64-AVX512-NEXT:    vpermps %ymm1, %ymm2, %ymm1