From 810572205c9702e0c42f76e68287e0d97273945a Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper@intel.com>
Date: Fri, 9 Feb 2018 23:32:27 +0000
Subject: [PATCH] [X86] Teach lower1BitVectorShuffle to recognize shuffles that
 are just filling upper elements with zero. Replace with insert_subvector.

There's still some extra kshifts in one of the modified test cases here, but hopefully that's only a DAG combine away.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@324782 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/X86/X86ISelLowering.cpp | 31 ++++++++++++++++++++++++++++++-
 test/CodeGen/X86/avx512-mask-op.ll | 22 ++++++++--------------
 2 files changed, 38 insertions(+), 15 deletions(-)

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index f5000385010..3c0ad7e125f 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -14414,8 +14414,36 @@ static SDValue lower512BitVectorShuffle(const SDLoc &DL, ArrayRef<int> Mask,
 // vector, shuffle and then truncate it back.
 static SDValue lower1BitVectorShuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                       MVT VT, SDValue V1, SDValue V2,
+                                      const APInt &Zeroable,
                                       const X86Subtarget &Subtarget,
                                       SelectionDAG &DAG) {
+  unsigned NumElts = Mask.size();
+
+  // Try to recognize shuffles that are just padding a subvector with zeros.
+  unsigned SubvecElts = 0;
+  for (int i = 0; i != (int)NumElts; ++i) {
+    if (Mask[i] >= 0 && Mask[i] != i)
+      break;
+
+    ++SubvecElts;
+  }
+  assert(SubvecElts != NumElts && "Identity shuffle?");
+
+  // Clip to a power 2.
+  SubvecElts = PowerOf2Floor(SubvecElts);
+
+  // Make sure the number of zeroable bits in the top at least covers the bits
+  // not covered by the subvector.
+  if (Zeroable.countLeadingOnes() >= (NumElts - SubvecElts)) {
+    MVT ExtractVT = MVT::getVectorVT(MVT::i1, SubvecElts);
+    SDValue Extract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ExtractVT,
+                                  V1, DAG.getIntPtrConstant(0, DL));
+    return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
+                       getZeroVector(VT, Subtarget, DAG, DL),
+                       Extract, DAG.getIntPtrConstant(0, DL));
+  }
+
+
   assert(Subtarget.hasAVX512() &&
          "Cannot lower 512-bit vectors w/o basic ISA!");
   MVT ExtVT;
@@ -14624,7 +14652,8 @@ static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget &Subtarget,
                                     DAG);
 
   if (Is1BitVector)
-    return lower1BitVectorShuffle(DL, Mask, VT, V1, V2, Subtarget, DAG);
+    return lower1BitVectorShuffle(DL, Mask, VT, V1, V2, Zeroable, Subtarget,
+                                  DAG);
 
   llvm_unreachable("Unimplemented!");
 }
diff --git a/test/CodeGen/X86/avx512-mask-op.ll b/test/CodeGen/X86/avx512-mask-op.ll
index 40d91356843..44e0a74db2e 100644
--- a/test/CodeGen/X86/avx512-mask-op.ll
+++ b/test/CodeGen/X86/avx512-mask-op.ll
@@ -2757,19 +2757,16 @@ define <8 x i64> @mask_widening(<2 x i64> %a, <2 x i64> %b, <2 x i64> %c, <2 x i
 ; KNL-NEXT:    ## kill: def $xmm0 killed $xmm0 def $zmm0
 ; KNL-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
 ; KNL-NEXT:    kshiftlw $12, %k0, %k0
-; KNL-NEXT:    kshiftrw $12, %k0, %k1
-; KNL-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
-; KNL-NEXT:    vmovdqa %ymm0, %ymm0
-; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k1
+; KNL-NEXT:    kshiftrw $12, %k0, %k0
+; KNL-NEXT:    kshiftlw $8, %k0, %k0
+; KNL-NEXT:    kshiftrw $8, %k0, %k1
 ; KNL-NEXT:    vpblendmd %zmm5, %zmm4, %zmm0 {%k1}
 ; KNL-NEXT:    retq
 ;
 ; SKX-LABEL: mask_widening:
 ; SKX:       ## %bb.0: ## %entry
 ; SKX-NEXT:    vpcmpeqd %xmm1, %xmm0, %k0
-; SKX-NEXT:    vpmovm2d %k0, %zmm0
-; SKX-NEXT:    vmovdqa %ymm0, %ymm0
-; SKX-NEXT:    vpmovd2m %zmm0, %k1
+; SKX-NEXT:    kmovb %k0, %k1
 ; SKX-NEXT:    vpblendmd %zmm5, %zmm4, %zmm0 {%k1}
 ; SKX-NEXT:    retq
 ;
@@ -2779,10 +2776,9 @@ define <8 x i64> @mask_widening(<2 x i64> %a, <2 x i64> %b, <2 x i64> %c, <2 x i
 ; AVX512BW-NEXT:    ## kill: def $xmm0 killed $xmm0 def $zmm0
 ; AVX512BW-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
 ; AVX512BW-NEXT:    kshiftlw $12, %k0, %k0
-; AVX512BW-NEXT:    kshiftrw $12, %k0, %k1
-; AVX512BW-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
-; AVX512BW-NEXT:    vmovdqa %ymm0, %ymm0
-; AVX512BW-NEXT:    vptestmd %zmm0, %zmm0, %k1
+; AVX512BW-NEXT:    kshiftrw $12, %k0, %k0
+; AVX512BW-NEXT:    kshiftlw $8, %k0, %k0
+; AVX512BW-NEXT:    kshiftrw $8, %k0, %k1
 ; AVX512BW-NEXT:    vpblendmd %zmm5, %zmm4, %zmm0 {%k1}
 ; AVX512BW-NEXT:    retq
 ;
@@ -2793,9 +2789,7 @@ define <8 x i64> @mask_widening(<2 x i64> %a, <2 x i64> %b, <2 x i64> %c, <2 x i
 ; AVX512DQ-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
 ; AVX512DQ-NEXT:    kshiftlb $4, %k0, %k0
 ; AVX512DQ-NEXT:    kshiftrb $4, %k0, %k0
-; AVX512DQ-NEXT:    vpmovm2d %k0, %zmm0
-; AVX512DQ-NEXT:    vmovdqa %ymm0, %ymm0
-; AVX512DQ-NEXT:    vpmovd2m %zmm0, %k1
+; AVX512DQ-NEXT:    kmovb %k0, %k1
 ; AVX512DQ-NEXT:    vpblendmd %zmm5, %zmm4, %zmm0 {%k1}
 ; AVX512DQ-NEXT:    retq
 entry:
-- 
2.11.0