[X86] Use VPERMI2B for v16i8 shuffles if we have VBMI+VLX and would have otherwise used two PSHUFBs ORed together.

llvm-svn: 321249

2 files changed, 38 insertions, 20 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index f17bd583f1a..5ae3be7d018 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -11953,6 +11953,19 @@ static int canLowerByDroppingEvenElements(ArrayRef<int> Mask,
   return 0;
 }
 
+static SDValue lowerVectorShuffleWithPERMV(const SDLoc &DL, MVT VT,
+                                           ArrayRef<int> Mask, SDValue V1,
+                                           SDValue V2, SelectionDAG &DAG) {
+  MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
+  MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
+
+  SDValue MaskNode = getConstVector(Mask, MaskVecVT, DAG, DL, true);
+  if (V2.isUndef())
+    return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
+
+  return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
+}
+
 /// \brief Generic lowering of v16i8 shuffles.
 ///
 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
@@ -12143,6 +12156,10 @@ static SDValue lowerV16I8VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask,
       if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(
               DL, MVT::v16i8, V1, V2, Mask, DAG))
         return Unpack;
+
+      // If we have VBMI we can use one VPERM instead of multiple PSHUFBs.
+      if (Subtarget.hasVBMI() && Subtarget.hasVLX())
+        return lowerVectorShuffleWithPERMV(DL, MVT::v16i8, Mask, V1, V2, DAG);
     }
 
     return PSHUFB;
@@ -13043,19 +13060,6 @@ static SDValue lowerVectorShuffleWithSHUFPD(const SDLoc &DL, MVT VT,
                      DAG.getConstant(Immediate, DL, MVT::i8));
 }
 
-static SDValue lowerVectorShuffleWithPERMV(const SDLoc &DL, MVT VT,
-                                           ArrayRef<int> Mask, SDValue V1,
-                                           SDValue V2, SelectionDAG &DAG) {
-  MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
-  MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
-
-  SDValue MaskNode = getConstVector(Mask, MaskVecVT, DAG, DL, true);
-  if (V2.isUndef())
-    return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
-
-  return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
-}
-
 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
 ///
 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-128-v16.ll b/llvm/test/CodeGen/X86/vector-shuffle-128-v16.ll
index 6aa7c0739ee..2f5a2b11611 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-128-v16.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-128-v16.ll
@@ -5,7 +5,8 @@
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX1OR2 --check-prefix=AVX1
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX1OR2 --check-prefix=AVX2OR512VL --check-prefix=AVX2 --check-prefix=AVX2-SLOW
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2,+fast-variable-shuffle | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX1OR2 --check-prefix=AVX2OR512VL --check-prefix=AVX2 --check-prefix=AVX2-FAST
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+avx512bw,+fast-variable-shuffle | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX2OR512VL --check-prefix=AVX512VL
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+avx512bw,+fast-variable-shuffle | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX2OR512VL --check-prefix=AVX512VL --check-prefix=AVX512VLBW
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+avx512bw,+avx512vbmi,+fast-variable-shuffle | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX2OR512VL --check-prefix=AVX512VL --check-prefix=AVX512VLVBMI
 
 define <16 x i8> @shuffle_v16i8_00_00_00_00_00_00_00_00_00_00_00_00_00_00_00_00(<16 x i8> %a, <16 x i8> %b) {
 ; SSE2-LABEL: shuffle_v16i8_00_00_00_00_00_00_00_00_00_00_00_00_00_00_00_00:
@@ -1222,12 +1223,25 @@ define <16 x i8> @shuffle_v16i8_uu_10_02_07_22_14_07_02_18_03_01_14_18_09_11_00(
 ; SSE41-NEXT:    por %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX-LABEL: shuffle_v16i8_uu_10_02_07_22_14_07_02_18_03_01_14_18_09_11_00:
-; AVX:       # %bb.0: # %entry
-; AVX-NEXT:    vpshufb {{.*#+}} xmm1 = xmm1[u],zero,zero,zero,xmm1[6],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[2],zero,zero,zero
-; AVX-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[u,10,2,7],zero,xmm0[14,7,2],zero,xmm0[3,1,14],zero,xmm0[9,11,0]
-; AVX-NEXT:    vpor %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    retq
+; AVX1OR2-LABEL: shuffle_v16i8_uu_10_02_07_22_14_07_02_18_03_01_14_18_09_11_00:
+; AVX1OR2:       # %bb.0: # %entry
+; AVX1OR2-NEXT:    vpshufb {{.*#+}} xmm1 = xmm1[u],zero,zero,zero,xmm1[6],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[2],zero,zero,zero
+; AVX1OR2-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[u,10,2,7],zero,xmm0[14,7,2],zero,xmm0[3,1,14],zero,xmm0[9,11,0]
+; AVX1OR2-NEXT:    vpor %xmm1, %xmm0, %xmm0
+; AVX1OR2-NEXT:    retq
+;
+; AVX512VLBW-LABEL: shuffle_v16i8_uu_10_02_07_22_14_07_02_18_03_01_14_18_09_11_00:
+; AVX512VLBW:       # %bb.0: # %entry
+; AVX512VLBW-NEXT:    vpshufb {{.*#+}} xmm1 = xmm1[u],zero,zero,zero,xmm1[6],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[2],zero,zero,zero
+; AVX512VLBW-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[u,10,2,7],zero,xmm0[14,7,2],zero,xmm0[3,1,14],zero,xmm0[9,11,0]
+; AVX512VLBW-NEXT:    vpor %xmm1, %xmm0, %xmm0
+; AVX512VLBW-NEXT:    retq
+;
+; AVX512VLVBMI-LABEL: shuffle_v16i8_uu_10_02_07_22_14_07_02_18_03_01_14_18_09_11_00:
+; AVX512VLVBMI:       # %bb.0: # %entry
+; AVX512VLVBMI-NEXT:    vmovdqa {{.*#+}} xmm2 = <u,10,2,7,22,14,7,2,18,3,1,14,18,9,11,0>
+; AVX512VLVBMI-NEXT:    vpermt2b %xmm1, %xmm2, %xmm0
+; AVX512VLVBMI-NEXT:    retq
 entry:
   %shuffle = shufflevector <16 x i8> %a, <16 x i8> %b, <16 x i32> <i32 undef, i32 10, i32 2, i32 7, i32 22, i32 14, i32 7, i32 2, i32 18, i32 3, i32 1, i32 14, i32 18, i32 9, i32 11, i32 0>