From 5401bae5230c59e887bbc987da85a360e8e21973 Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Sun, 20 Nov 2016 15:24:38 +0000
Subject: [PATCH] [X86][AVX512] Add support for VBMI VPERMV3 target shuffle
 combines

llvm-svn: 287496
---
 llvm/lib/Target/X86/X86ISelLowering.cpp            |  4 ++-
 .../X86/vector-shuffle-combining-avx512vbmi.ll     | 32 ++++++++++++++++++++++
 2 files changed, 35 insertions(+), 1 deletion(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 793176e..6560fe3 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -25984,7 +25984,9 @@ static bool combineX86ShuffleChain(ArrayRef<SDValue> Inputs, SDValue Root,
           (MaskVT == MVT::v4f64 || MaskVT == MVT::v4i64 ||
            MaskVT == MVT::v8f32 || MaskVT == MVT::v8i32)) ||
          (Subtarget.hasBWI() && MaskVT == MVT::v32i16) ||
-         (Subtarget.hasBWI() && Subtarget.hasVLX() && MaskVT == MVT::v16i16))) {
+         (Subtarget.hasBWI() && Subtarget.hasVLX() && MaskVT == MVT::v16i16) ||
+         (Subtarget.hasVBMI() && MaskVT == MVT::v64i8) ||
+         (Subtarget.hasVBMI() && Subtarget.hasVLX() && MaskVT == MVT::v32i8))) {
       MVT VPermMaskSVT = MVT::getIntegerVT(MaskEltSizeInBits);
       MVT VPermMaskVT = MVT::getVectorVT(VPermMaskSVT, NumMaskElts);
       SDValue VPermMask = getConstVector(Mask, VPermMaskVT, DAG, DL, true);
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512vbmi.ll b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512vbmi.ll
index d589a92..849c277 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512vbmi.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512vbmi.ll
@@ -125,3 +125,35 @@ define <16 x i8> @combine_vpermt2var_vpermi2var_16i8_as_vperm2(<16 x i8> %x0, <1
   %res1 = call <16 x i8> @llvm.x86.avx512.maskz.vpermt2var.qi.128(<16 x i8> <i8 0, i8 17, i8 2, i8 18, i8 4, i8 19, i8 6, i8 21, i8 8, i8 23, i8 10, i8 25, i8 12, i8 27, i8 14, i8 29>, <16 x i8> %res0, <16 x i8> %res0, i16 -1)
   ret <16 x i8> %res1
 }
+define <32 x i8> @combine_vpermi2var_32i8_as_vperm2(<32 x i8> %x0, <32 x i8> %x1) {
+; X32-LABEL: combine_vpermi2var_32i8_as_vperm2:
+; X32:       # BB#0:
+; X32-NEXT:    vmovdqu8 {{.*#+}} ymm2 = [0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19]
+; X32-NEXT:    vpermt2b %ymm1, %ymm2, %ymm0
+; X32-NEXT:    retl
+;
+; X64-LABEL: combine_vpermi2var_32i8_as_vperm2:
+; X64:       # BB#0:
+; X64-NEXT:    vmovdqu8 {{.*#+}} ymm2 = [0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19]
+; X64-NEXT:    vpermt2b %ymm1, %ymm2, %ymm0
+; X64-NEXT:    retq
+  %res0 = shufflevector <32 x i8> %x0, <32 x i8> %x1, <32 x i32> <i32 0, i32 32, i32 1, i32 33, i32 2, i32 34, i32 3, i32 35, i32 4, i32 36, i32 5, i32 37, i32 6, i32 38, i32 7, i32 39, i32 16, i32 48, i32 17, i32 49, i32 18, i32 50, i32 19, i32 51, i32 20, i32 52, i32 21, i32 53, i32 22, i32 54, i32 23, i32 55>
+  %res1 = call <32 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.256(<32 x i8> %res0, <32 x i8> <i8 0, i8 32, i8 2, i8 30, i8 4, i8 28, i8 6, i8 26, i8 8, i8 28, i8 10, i8 26, i8 12, i8 24, i8 14, i8 22, i8 0, i8 32, i8 2, i8 30, i8 4, i8 28, i8 6, i8 26, i8 8, i8 28, i8 10, i8 26, i8 12, i8 24, i8 14, i8 22>, <32 x i8> %x1, i32 -1)
+  ret <32 x i8> %res1
+}
+define <64 x i8> @combine_vpermi2var_64i8_as_vperm2(<64 x i8> %x0, <64 x i8> %x1) {
+; X32-LABEL: combine_vpermi2var_64i8_as_vperm2:
+; X32:       # BB#0:
+; X32-NEXT:    vmovdqu8 {{.*#+}} zmm2 = [0,80,1,70,2,54,3,49,4,36,5,23,6,18,7,5,0,90,1,100,2,110,3,120,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19]
+; X32-NEXT:    vpermt2b %zmm1, %zmm2, %zmm0
+; X32-NEXT:    retl
+;
+; X64-LABEL: combine_vpermi2var_64i8_as_vperm2:
+; X64:       # BB#0:
+; X64-NEXT:    vmovdqu8 {{.*#+}} zmm2 = [0,80,1,70,2,54,3,49,4,36,5,23,6,18,7,5,0,90,1,100,2,110,3,120,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19,0,32,1,23,2,22,3,21,4,22,5,21,6,20,7,19]
+; X64-NEXT:    vpermt2b %zmm1, %zmm2, %zmm0
+; X64-NEXT:    retq
+  %res0 = shufflevector <64 x i8> %x0, <64 x i8> %x1, <64 x i32> <i32 0, i32 64, i32 1, i32 65, i32 2, i32 66, i32 3, i32 67, i32 4, i32 68, i32 5, i32 69, i32 6, i32 70, i32 7, i32 71, i32 16, i32 80, i32 17, i32 81, i32 18, i32 82, i32 19, i32 83, i32 20, i32 84, i32 21, i32 85, i32 22, i32 86, i32 23, i32 87, i32 32, i32 96, i32 33, i32 97, i32 34, i32 98, i32 35, i32 99, i32 36, i32 100, i32 37, i32 101, i32 38, i32 102, i32 39, i32 103, i32 48, i32 112, i32 49, i32 113, i32 50, i32 114, i32 51, i32 115, i32 52, i32 116, i32 53, i32 117, i32 54, i32 118, i32 55, i32 119>
+  %res1 = call <64 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.512(<64 x i8> %res0, <64 x i8> <i8 0, i8 80, i8 2, i8 70, i8 4, i8 60, i8 6, i8 50, i8 8, i8 40, i8 10, i8 30, i8 12, i8 20, i8 14, i8 10, i8 0, i8 90, i8 2, i8 100, i8 4, i8 110, i8 6, i8 120, i8 8, i8 28, i8 10, i8 26, i8 12, i8 24, i8 14, i8 22, i8 0, i8 32, i8 2, i8 30, i8 4, i8 28, i8 6, i8 26, i8 8, i8 28, i8 10, i8 26, i8 12, i8 24, i8 14, i8 22, i8 0, i8 32, i8 2, i8 30, i8 4, i8 28, i8 6, i8 26, i8 8, i8 28, i8 10, i8 26, i8 12, i8 24, i8 14, i8 22>, <64 x i8> %x1, i64 -1)
+  ret <64 x i8> %res1
+}
-- 
2.7.4