From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Wed, 3 Apr 2019 17:17:13 +0000 (+0000)
Subject: [X86][AVX] combineHorizontalPredicateResult - support v16i16/v32i8 reduction on AVX1
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=9e28dddf555c686f23a0de5336e2602b58aab7d5;p=platform%2Fupstream%2Fllvm.git

[X86][AVX] combineHorizontalPredicateResult - support v16i16/v32i8 reduction on AVX1

Use getPMOVMSKB helper which splits v32i8 MOVMSK calls on pre-AVX2 targets.

llvm-svn: 357608
---

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 601c7be..0ae7b5b 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -34301,11 +34301,8 @@ static SDValue combineHorizontalPredicateResult(SDNode *Extract,
   if (Match.getScalarValueSizeInBits() != BitWidth)
     return SDValue();
 
-  // We require AVX2 for PMOVMSKB for v16i16/v32i8;
   unsigned MatchSizeInBits = Match.getValueSizeInBits();
-  if (!(MatchSizeInBits == 128 ||
-        (MatchSizeInBits == 256 &&
-         ((Subtarget.hasAVX() && BitWidth >= 32) || Subtarget.hasAVX2()))))
+  if (!(MatchSizeInBits == 128 || (MatchSizeInBits == 256 && Subtarget.hasAVX())))
     return SDValue();
 
   // Make sure this isn't a vector of 1 element. The perf win from using MOVMSK
@@ -34344,9 +34341,9 @@ static SDValue combineHorizontalPredicateResult(SDNode *Extract,
 
   // The setcc produces an i8 of 0/1, so extend that to the result width and
   // negate to get the final 0/-1 mask value.
-  SDValue BitcastLogicOp = DAG.getBitcast(MaskSrcVT, Match);
-  SDValue Movmsk = DAG.getNode(X86ISD::MOVMSK, DL, MVT::i32, BitcastLogicOp);
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+  SDValue BitcastLogicOp = DAG.getBitcast(MaskSrcVT, Match);
+  SDValue Movmsk = getPMOVMSKB(DL, BitcastLogicOp, DAG, Subtarget);
   EVT SetccVT = TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(),
                                        MVT::i32);
   SDValue Setcc = DAG.getSetCC(DL, SetccVT, Movmsk, CmpC, CondCode);
diff --git a/llvm/test/CodeGen/X86/vector-compare-all_of.ll b/llvm/test/CodeGen/X86/vector-compare-all_of.ll
index 510c515..8c45094 100644
--- a/llvm/test/CodeGen/X86/vector-compare-all_of.ll
+++ b/llvm/test/CodeGen/X86/vector-compare-all_of.ll
@@ -673,14 +673,14 @@ define i16 @test_v16i16_sext(<16 x i16> %a0, <16 x i16> %a1) {
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
 ; AVX1-NEXT:    vpcmpgtw %xmm2, %xmm3, %xmm2
 ; AVX1-NEXT:    vpcmpgtw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vmovd %xmm0, %eax
+; AVX1-NEXT:    vpmovmskb %xmm0, %eax
+; AVX1-NEXT:    vpmovmskb %xmm2, %ecx
+; AVX1-NEXT:    shll $16, %ecx
+; AVX1-NEXT:    orl %eax, %ecx
+; AVX1-NEXT:    xorl %eax, %eax
+; AVX1-NEXT:    cmpl $-1, %ecx
+; AVX1-NEXT:    sete %al
+; AVX1-NEXT:    negl %eax
 ; AVX1-NEXT:    # kill: def $ax killed $ax killed $eax
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
@@ -867,17 +867,13 @@ define i8 @test_v32i8_sext(<32 x i8> %a0, <32 x i8> %a1) {
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
 ; AVX1-NEXT:    vpcmpgtb %xmm2, %xmm3, %xmm2
 ; AVX1-NEXT:    vpcmpgtb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpextrb $0, %xmm0, %eax
-; AVX1-NEXT:    # kill: def $al killed $al killed $eax
+; AVX1-NEXT:    vpmovmskb %xmm0, %eax
+; AVX1-NEXT:    vpmovmskb %xmm2, %ecx
+; AVX1-NEXT:    shll $16, %ecx
+; AVX1-NEXT:    orl %eax, %ecx
+; AVX1-NEXT:    cmpl $-1, %ecx
+; AVX1-NEXT:    sete %al
+; AVX1-NEXT:    negb %al
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;
@@ -1555,21 +1551,17 @@ define i1 @bool_reduction_v32i8(<32 x i8> %x, <32 x i8> %y) {
 ;
 ; AVX1-LABEL: bool_reduction_v32i8:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
+; AVX1-NEXT:    vpcmpeqb %xmm2, %xmm3, %xmm2
 ; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX1-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; AVX1-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; AVX1-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpextrb $0, %xmm0, %eax
-; AVX1-NEXT:    # kill: def $al killed $al killed $eax
+; AVX1-NEXT:    vpmovmskb %xmm0, %eax
+; AVX1-NEXT:    vpmovmskb %xmm2, %ecx
+; AVX1-NEXT:    shll $16, %ecx
+; AVX1-NEXT:    orl %eax, %ecx
+; AVX1-NEXT:    cmpl $-1, %ecx
+; AVX1-NEXT:    sete %al
+; AVX1-NEXT:    negb %al
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;
diff --git a/llvm/test/CodeGen/X86/vector-compare-any_of.ll b/llvm/test/CodeGen/X86/vector-compare-any_of.ll
index a8bfc42..2236e95 100644
--- a/llvm/test/CodeGen/X86/vector-compare-any_of.ll
+++ b/llvm/test/CodeGen/X86/vector-compare-any_of.ll
@@ -611,14 +611,13 @@ define i16 @test_v16i16_sext(<16 x i16> %a0, <16 x i16> %a1) {
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
 ; AVX1-NEXT:    vpcmpgtw %xmm2, %xmm3, %xmm2
 ; AVX1-NEXT:    vpcmpgtw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpor %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX1-NEXT:    vpor %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; AVX1-NEXT:    vpor %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; AVX1-NEXT:    vpor %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vmovd %xmm0, %eax
+; AVX1-NEXT:    vpmovmskb %xmm0, %ecx
+; AVX1-NEXT:    vpmovmskb %xmm2, %edx
+; AVX1-NEXT:    shll $16, %edx
+; AVX1-NEXT:    xorl %eax, %eax
+; AVX1-NEXT:    orl %ecx, %edx
+; AVX1-NEXT:    setne %al
+; AVX1-NEXT:    negl %eax
 ; AVX1-NEXT:    # kill: def $ax killed $ax killed $eax
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
@@ -791,17 +790,12 @@ define i8 @test_v32i8_sext(<32 x i8> %a0, <32 x i8> %a1) {
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
 ; AVX1-NEXT:    vpcmpgtb %xmm2, %xmm3, %xmm2
 ; AVX1-NEXT:    vpcmpgtb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpor %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX1-NEXT:    vpor %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; AVX1-NEXT:    vpor %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; AVX1-NEXT:    vpor %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpor %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpextrb $0, %xmm0, %eax
-; AVX1-NEXT:    # kill: def $al killed $al killed $eax
+; AVX1-NEXT:    vpmovmskb %xmm0, %eax
+; AVX1-NEXT:    vpmovmskb %xmm2, %ecx
+; AVX1-NEXT:    shll $16, %ecx
+; AVX1-NEXT:    orl %eax, %ecx
+; AVX1-NEXT:    setne %al
+; AVX1-NEXT:    negb %al
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;
@@ -1490,21 +1484,16 @@ define i1 @bool_reduction_v32i8(<32 x i8> %x, <32 x i8> %y) {
 ;
 ; AVX1-LABEL: bool_reduction_v32i8:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
+; AVX1-NEXT:    vpcmpeqb %xmm2, %xmm3, %xmm2
 ; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpor %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX1-NEXT:    vpor %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; AVX1-NEXT:    vpor %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; AVX1-NEXT:    vpor %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpor %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpextrb $0, %xmm0, %eax
-; AVX1-NEXT:    # kill: def $al killed $al killed $eax
+; AVX1-NEXT:    vpmovmskb %xmm0, %eax
+; AVX1-NEXT:    vpmovmskb %xmm2, %ecx
+; AVX1-NEXT:    shll $16, %ecx
+; AVX1-NEXT:    orl %eax, %ecx
+; AVX1-NEXT:    setne %al
+; AVX1-NEXT:    negb %al
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;