[X86][AVX2] Add support for combining target shuffles to VPERMD/VPERMPS

llvm-svn: 283080

2 files changed, 32 insertions, 24 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 674aa257db1..62c116d0b5d 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -25444,12 +25444,32 @@ static bool combineX86ShuffleChain(ArrayRef<SDValue> Inputs, SDValue Root,
   if (Depth < 2)
     return false;
 
-  if (is128BitLaneCrossingShuffleMask(MaskVT, Mask))
-    return false;
-
   bool MaskContainsZeros =
       any_of(Mask, [](int M) { return M == SM_SentinelZero; });
 
+  if (is128BitLaneCrossingShuffleMask(MaskVT, Mask)) {
+    // If we have a single input lane-crossing shuffle with 32-bit scalars then
+    // lower to VPERMD/VPERMPS.
+    if (UnaryShuffle && (Depth >= 3 || HasVariableMask) && !MaskContainsZeros &&
+        Subtarget.hasAVX2() && (MaskVT == MVT::v8f32 || MaskVT == MVT::v8i32)) {
+      SDValue VPermIdx[8];
+      for (int i = 0; i < 8; ++i)
+        VPermIdx[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
+                                  : DAG.getConstant(Mask[i], DL, MVT::i32);
+
+      SDValue VPermMask = DAG.getBuildVector(MVT::v8i32, DL, VPermIdx);
+      DCI.AddToWorklist(VPermMask.getNode());
+      Res = DAG.getBitcast(MaskVT, V1);
+      DCI.AddToWorklist(Res.getNode());
+      Res = DAG.getNode(X86ISD::VPERMV, DL, MaskVT, VPermMask, Res);
+      DCI.AddToWorklist(Res.getNode());
+      DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Res),
+                    /*AddTo*/ true);
+      return true;
+    }
+    return false;
+  }
+
   // If we have a single input shuffle with different shuffle patterns in the
   // the 128-bit lanes use the variable mask to VPERMILPS.
   // TODO Combine other mask types at higher depths.
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx2.ll b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx2.ll
index 72eff32ac03..3465429e163 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx2.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx2.ll
@@ -92,18 +92,14 @@ define <4 x i64> @combine_permq_pshufb_as_vperm2i128(<4 x i64> %a0) {
 define <8 x i32> @combine_as_vpermd(<8 x i32> %a0) {
 ; X32-LABEL: combine_as_vpermd:
 ; X32:       # BB#0:
-; X32-NEXT:    vmovdqa {{.*#+}} ymm1 = <4,u,u,5,u,u,0,7>
-; X32-NEXT:    vpermd %ymm0, %ymm1, %ymm1
-; X32-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
-; X32-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6,7]
+; X32-NEXT:    vmovdqa {{.*#+}} ymm1 = [4,5,4,5,6,7,0,7]
+; X32-NEXT:    vpermd %ymm0, %ymm1, %ymm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: combine_as_vpermd:
 ; X64:       # BB#0:
-; X64-NEXT:    vmovdqa {{.*#+}} ymm1 = <4,u,u,5,u,u,0,7>
-; X64-NEXT:    vpermd %ymm0, %ymm1, %ymm1
-; X64-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
-; X64-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4,5],ymm1[6,7]
+; X64-NEXT:    vmovdqa {{.*#+}} ymm1 = [4,5,4,5,6,7,0,7]
+; X64-NEXT:    vpermd %ymm0, %ymm1, %ymm0
 ; X64-NEXT:    retq
   %1 = shufflevector <8 x i32> %a0, <8 x i32> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 0, i32 1, i32 2, i32 3>
   %2 = tail call <8 x i32> @llvm.x86.avx2.permd(<8 x i32> %a0, <8 x i32> <i32 5, i32 4, i32 3, i32 2, i32 1, i32 0, i32 7, i32 6>)
@@ -114,25 +110,17 @@ define <8 x i32> @combine_as_vpermd(<8 x i32> %a0) {
 define <8 x float> @combine_as_vpermps(<8 x float> %a0) {
 ; X32-LABEL: combine_as_vpermps:
 ; X32:       # BB#0:
-; X32-NEXT:    vpermilps {{.*#+}} ymm1 = ymm0[1,0,3,2,5,4,7,6]
-; X32-NEXT:    vmovaps {{.*#+}} ymm2 = <u,4,u,5,u,u,4,7>
-; X32-NEXT:    vpermps %ymm0, %ymm2, %ymm0
-; X32-NEXT:    vmovaps {{.*#+}} ymm2 = <7,u,6,u,0,1,u,u>
-; X32-NEXT:    vpermps %ymm1, %ymm2, %ymm1
-; X32-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4,5],ymm0[6,7]
+; X32-NEXT:    vmovaps {{.*#+}} ymm1 = <6,4,7,5,1,u,4,7>
+; X32-NEXT:    vpermps %ymm0, %ymm1, %ymm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: combine_as_vpermps:
 ; X64:       # BB#0:
-; X64-NEXT:    vpermilps {{.*#+}} ymm1 = ymm0[1,0,3,2,5,4,7,6]
-; X64-NEXT:    vmovaps {{.*#+}} ymm2 = <u,4,u,5,u,u,4,7>
-; X64-NEXT:    vpermps %ymm0, %ymm2, %ymm0
-; X64-NEXT:    vmovaps {{.*#+}} ymm2 = <7,u,6,u,0,1,u,u>
-; X64-NEXT:    vpermps %ymm1, %ymm2, %ymm1
-; X64-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4,5],ymm0[6,7]
+; X64-NEXT:    vmovaps {{.*#+}} ymm1 = <6,4,7,5,1,u,4,7>
+; X64-NEXT:    vpermps %ymm0, %ymm1, %ymm0
 ; X64-NEXT:    retq
   %1 = shufflevector <8 x float> %a0, <8 x float> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
-  %2 = tail call <8 x float> @llvm.x86.avx2.permps(<8 x float> %a0, <8 x i32> <i32 1, i32 0, i32 3, i32 2, i32 5, i32 4, i32 7, i32 6>)
+  %2 = tail call <8 x float> @llvm.x86.avx2.permps(<8 x float> %a0, <8 x i32> <i32 1, i32 undef, i32 3, i32 2, i32 5, i32 4, i32 7, i32 6>)
   %3 = shufflevector <8 x float> %1, <8 x float> %2, <8 x i32> <i32 15, i32 0, i32 14, i32 1, i32 8, i32 9, i32 4, i32 3>
   ret <8 x float> %3
 }