From c143db3b1032042193c152790bcefe34365e6d6c Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Thu, 18 Jun 2020 17:41:15 +0100
Subject: [PATCH] [X86][SSE] combineHorizontalPredicateResult - improve
 all_of(X == 0) for vXi64 on pre-SSE41 targets

Without SSE41 we don't have the PCMPEQQ instruction, making cmp-with-zero reductions more complicated than necessary. We can compare as vXi32 (PCMPEQD) and tweak the MOVMSK comparison to test upper/lower DWORD comparisons.

This pre-fixes something that occurs with null tests for vectors of (64-bit) pointers such as in PR35129.
---
 llvm/lib/Target/X86/X86ISelLowering.cpp         | 19 +++++++++++++
 llvm/test/CodeGen/X86/vector-reduce-and-bool.ll | 36 ++++++++-----------------
 2 files changed, 30 insertions(+), 25 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index cd58423..e2a9231 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -38504,6 +38504,25 @@ static SDValue combineHorizontalPredicateResult(SDNode *Extract,
       EVT MovmskVT = EVT::getIntegerVT(*DAG.getContext(), NumElts);
       Movmsk = DAG.getBitcast(MovmskVT, Match);
     } else {
+      // For all_of(setcc(vec,0,eq)) - avoid vXi64 comparisons if we don't have
+      // PCMPEQQ (SSE41+), use PCMPEQD instead.
+      if (BinOp == ISD::AND && !Subtarget.hasSSE41() &&
+          Match.getOpcode() == ISD::SETCC &&
+          ISD::isBuildVectorAllZeros(Match.getOperand(1).getNode()) &&
+          cast<CondCodeSDNode>(Match.getOperand(2))->get() ==
+              ISD::CondCode::SETEQ) {
+        SDValue Vec = Match.getOperand(0);
+        if (Vec.getValueType().getScalarType() == MVT::i64 &&
+            (2 * NumElts) <= MaxElts) {
+          NumElts *= 2;
+          EVT CmpVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
+          MatchVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, NumElts);
+          Match = DAG.getSetCC(
+              DL, MatchVT, DAG.getBitcast(CmpVT, Match.getOperand(0)),
+              DAG.getBitcast(CmpVT, Match.getOperand(1)), ISD::CondCode::SETEQ);
+        }
+      }
+
       // Use combineBitcastvxi1 to create the MOVMSK.
       while (NumElts > MaxElts) {
         SDValue Lo, Hi;
diff --git a/llvm/test/CodeGen/X86/vector-reduce-and-bool.ll b/llvm/test/CodeGen/X86/vector-reduce-and-bool.ll
index ce6a424..4c34ee8 100644
--- a/llvm/test/CodeGen/X86/vector-reduce-and-bool.ll
+++ b/llvm/test/CodeGen/X86/vector-reduce-and-bool.ll
@@ -858,10 +858,8 @@ define i1 @icmp_v2i64_v2i1(<2 x i64>) {
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    pxor %xmm1, %xmm1
 ; SSE2-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,0,3,2]
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    movmskpd %xmm0, %eax
-; SSE2-NEXT:    cmpb $3, %al
+; SSE2-NEXT:    movmskps %xmm1, %eax
+; SSE2-NEXT:    cmpb $15, %al
 ; SSE2-NEXT:    sete %al
 ; SSE2-NEXT:    retq
 ;
@@ -1088,14 +1086,11 @@ define i1 @icmp_v4i64_v4i1(<4 x i64>) {
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    pxor %xmm2, %xmm2
 ; SSE2-NEXT:    pcmpeqd %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[1,0,3,2]
-; SSE2-NEXT:    pand %xmm1, %xmm3
 ; SSE2-NEXT:    pcmpeqd %xmm2, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,0,3,2]
-; SSE2-NEXT:    pand %xmm0, %xmm1
-; SSE2-NEXT:    packssdw %xmm3, %xmm1
-; SSE2-NEXT:    movmskps %xmm1, %eax
-; SSE2-NEXT:    cmpb $15, %al
+; SSE2-NEXT:    packssdw %xmm1, %xmm0
+; SSE2-NEXT:    packsswb %xmm0, %xmm0
+; SSE2-NEXT:    pmovmskb %xmm0, %eax
+; SSE2-NEXT:    cmpb $-1, %al
 ; SSE2-NEXT:    sete %al
 ; SSE2-NEXT:    retq
 ;
@@ -1383,23 +1378,14 @@ define i1 @icmp_v8i64_v8i1(<8 x i64>) {
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    pxor %xmm4, %xmm4
 ; SSE2-NEXT:    pcmpeqd %xmm4, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[1,0,3,2]
-; SSE2-NEXT:    pand %xmm3, %xmm5
 ; SSE2-NEXT:    pcmpeqd %xmm4, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[1,0,3,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    packssdw %xmm5, %xmm3
+; SSE2-NEXT:    packssdw %xmm3, %xmm2
 ; SSE2-NEXT:    pcmpeqd %xmm4, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,0,3,2]
-; SSE2-NEXT:    pand %xmm1, %xmm2
 ; SSE2-NEXT:    pcmpeqd %xmm4, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,0,3,2]
-; SSE2-NEXT:    pand %xmm0, %xmm1
-; SSE2-NEXT:    packssdw %xmm2, %xmm1
-; SSE2-NEXT:    packssdw %xmm3, %xmm1
-; SSE2-NEXT:    packsswb %xmm1, %xmm1
-; SSE2-NEXT:    pmovmskb %xmm1, %eax
-; SSE2-NEXT:    cmpb $-1, %al
+; SSE2-NEXT:    packssdw %xmm1, %xmm0
+; SSE2-NEXT:    packsswb %xmm2, %xmm0
+; SSE2-NEXT:    pmovmskb %xmm0, %eax
+; SSE2-NEXT:    cmpw $-1, %ax
 ; SSE2-NEXT:    sete %al
 ; SSE2-NEXT:    retq
 ;
-- 
2.7.4