From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Fri, 19 Jun 2020 09:27:09 +0000 (+0100)
Subject: [X86][SSE] combineSetCCMOVMSK - fold MOVMSK(SHUFFLE(X,u)) -> MOVMSK(X)
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=cad2038700b204ba437b269dcba5b995f0c385ab;p=platform%2Fupstream%2Fllvm.git

[X86][SSE] combineSetCCMOVMSK - fold MOVMSK(SHUFFLE(X,u)) -> MOVMSK(X)

If we're permuting ALL the elements of a single vector, then for allof/anyof MOVMSK tests we can avoid the shuffle entirely.
---

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index ed936af..cd58423 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -5441,6 +5441,14 @@ static bool isAnyZero(ArrayRef<int> Mask) {
   return llvm::any_of(Mask, [](int M) { return M == SM_SentinelZero; });
 }
 
+/// Return true if the value of any element in Mask is the zero or undef
+/// sentinel values.
+static bool isAnyZeroOrUndef(ArrayRef<int> Mask) {
+  return llvm::any_of(Mask, [](int M) {
+    return M == SM_SentinelZero || M == SM_SentinelUndef;
+  });
+}
+
 /// Return true if Val is undef or if its value falls within the
 /// specified range (L, H].
 static bool isUndefOrInRange(int Val, int Low, int Hi) {
@@ -40590,6 +40598,31 @@ static SDValue combineSetCCMOVMSK(SDValue EFLAGS, X86::CondCode &CC,
     }
   }
 
+  // MOVMSK(SHUFFLE(X,u)) -> MOVMSK(X) iff every element is referenced.
+  SmallVector<int, 32> ShuffleMask;
+  SmallVector<SDValue, 2> ShuffleInputs;
+  if (NumElts == CmpBits &&
+      getTargetShuffleInputs(peekThroughBitcasts(Vec), ShuffleInputs,
+                             ShuffleMask, DAG) &&
+      ShuffleInputs.size() == 1 && !isAnyZeroOrUndef(ShuffleMask) &&
+      ShuffleInputs[0].getValueSizeInBits() == VecVT.getSizeInBits()) {
+    unsigned NumShuffleElts = ShuffleMask.size();
+    APInt DemandedElts = APInt::getNullValue(NumShuffleElts);
+    for (int M : ShuffleMask) {
+      assert(0 <= M && M < (int)NumShuffleElts && "Bad unary shuffle index");
+      DemandedElts.setBit(M);
+    }
+    if (DemandedElts.isAllOnesValue()) {
+      SDLoc DL(EFLAGS);
+      SDValue Result = DAG.getBitcast(VecVT, ShuffleInputs[0]);
+      Result = DAG.getNode(X86ISD::MOVMSK, DL, MVT::i32, Result);
+      Result =
+          DAG.getZExtOrTrunc(Result, DL, EFLAGS.getOperand(0).getValueType());
+      return DAG.getNode(X86ISD::CMP, DL, MVT::i32, Result,
+                         EFLAGS.getOperand(1));
+    }
+  }
+
   return SDValue();
 }
 
diff --git a/llvm/test/CodeGen/X86/movmsk-cmp.ll b/llvm/test/CodeGen/X86/movmsk-cmp.ll
index e0e8ac0..204bcc97 100644
--- a/llvm/test/CodeGen/X86/movmsk-cmp.ll
+++ b/llvm/test/CodeGen/X86/movmsk-cmp.ll
@@ -530,7 +530,6 @@ define i1 @allones_v32i16_sign(<32 x i16> %arg) {
 ; AVX2-LABEL: allones_v32i16_sign:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpacksswb %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vpmovmskb %ymm0, %eax
 ; AVX2-NEXT:    cmpl $-1, %eax
 ; AVX2-NEXT:    sete %al
@@ -598,7 +597,6 @@ define i1 @allzeros_v32i16_sign(<32 x i16> %arg) {
 ; AVX2-LABEL: allzeros_v32i16_sign:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpacksswb %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vpmovmskb %ymm0, %eax
 ; AVX2-NEXT:    testl %eax, %eax
 ; AVX2-NEXT:    sete %al
@@ -1008,7 +1006,6 @@ define i1 @allones_v8i64_sign(<8 x i64> %arg) {
 ; AVX2-LABEL: allones_v8i64_sign:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpackssdw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vmovmskps %ymm0, %eax
 ; AVX2-NEXT:    cmpb $-1, %al
 ; AVX2-NEXT:    sete %al
@@ -1068,7 +1065,6 @@ define i1 @allzeros_v8i64_sign(<8 x i64> %arg) {
 ; AVX2-LABEL: allzeros_v8i64_sign:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpackssdw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vmovmskps %ymm0, %eax
 ; AVX2-NEXT:    testb %al, %al
 ; AVX2-NEXT:    sete %al
@@ -1642,7 +1638,6 @@ define i1 @allones_v32i16_and1(<32 x i16> %arg) {
 ; AVX2-NEXT:    vpsllw $15, %ymm1, %ymm1
 ; AVX2-NEXT:    vpsllw $15, %ymm0, %ymm0
 ; AVX2-NEXT:    vpacksswb %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vpmovmskb %ymm0, %eax
 ; AVX2-NEXT:    cmpl $-1, %eax
 ; AVX2-NEXT:    sete %al
@@ -1722,7 +1717,6 @@ define i1 @allzeros_v32i16_and1(<32 x i16> %arg) {
 ; AVX2-NEXT:    vpsllw $15, %ymm1, %ymm1
 ; AVX2-NEXT:    vpsllw $15, %ymm0, %ymm0
 ; AVX2-NEXT:    vpacksswb %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vpmovmskb %ymm0, %eax
 ; AVX2-NEXT:    testl %eax, %eax
 ; AVX2-NEXT:    sete %al
@@ -2391,7 +2385,6 @@ define i1 @allones_v8i64_and1(<8 x i64> %arg) {
 ; AVX2-NEXT:    vpsllq $63, %ymm1, %ymm1
 ; AVX2-NEXT:    vpsllq $63, %ymm0, %ymm0
 ; AVX2-NEXT:    vpackssdw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vmovmskps %ymm0, %eax
 ; AVX2-NEXT:    cmpb $-1, %al
 ; AVX2-NEXT:    sete %al
@@ -2461,7 +2454,6 @@ define i1 @allzeros_v8i64_and1(<8 x i64> %arg) {
 ; AVX2-NEXT:    vpsllq $63, %ymm1, %ymm1
 ; AVX2-NEXT:    vpsllq $63, %ymm0, %ymm0
 ; AVX2-NEXT:    vpackssdw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vmovmskps %ymm0, %eax
 ; AVX2-NEXT:    testb %al, %al
 ; AVX2-NEXT:    sete %al
@@ -3035,7 +3027,6 @@ define i1 @allones_v32i16_and4(<32 x i16> %arg) {
 ; AVX2-NEXT:    vpsllw $13, %ymm1, %ymm1
 ; AVX2-NEXT:    vpsllw $13, %ymm0, %ymm0
 ; AVX2-NEXT:    vpacksswb %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vpmovmskb %ymm0, %eax
 ; AVX2-NEXT:    cmpl $-1, %eax
 ; AVX2-NEXT:    sete %al
@@ -3115,7 +3106,6 @@ define i1 @allzeros_v32i16_and4(<32 x i16> %arg) {
 ; AVX2-NEXT:    vpsllw $13, %ymm1, %ymm1
 ; AVX2-NEXT:    vpsllw $13, %ymm0, %ymm0
 ; AVX2-NEXT:    vpacksswb %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vpmovmskb %ymm0, %eax
 ; AVX2-NEXT:    testl %eax, %eax
 ; AVX2-NEXT:    sete %al
@@ -3784,7 +3774,6 @@ define i1 @allones_v8i64_and4(<8 x i64> %arg) {
 ; AVX2-NEXT:    vpsllq $61, %ymm1, %ymm1
 ; AVX2-NEXT:    vpsllq $61, %ymm0, %ymm0
 ; AVX2-NEXT:    vpackssdw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vmovmskps %ymm0, %eax
 ; AVX2-NEXT:    cmpb $-1, %al
 ; AVX2-NEXT:    sete %al
@@ -3854,7 +3843,6 @@ define i1 @allzeros_v8i64_and4(<8 x i64> %arg) {
 ; AVX2-NEXT:    vpsllq $61, %ymm1, %ymm1
 ; AVX2-NEXT:    vpsllq $61, %ymm0, %ymm0
 ; AVX2-NEXT:    vpackssdw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vmovmskps %ymm0, %eax
 ; AVX2-NEXT:    testb %al, %al
 ; AVX2-NEXT:    sete %al
diff --git a/llvm/test/CodeGen/X86/vector-reduce-and-bool.ll b/llvm/test/CodeGen/X86/vector-reduce-and-bool.ll
index f3a1b44..ce6a424 100644
--- a/llvm/test/CodeGen/X86/vector-reduce-and-bool.ll
+++ b/llvm/test/CodeGen/X86/vector-reduce-and-bool.ll
@@ -1443,7 +1443,6 @@ define i1 @icmp_v8i64_v8i1(<8 x i64>) {
 ; AVX2-NEXT:    vpcmpeqq %ymm2, %ymm1, %ymm1
 ; AVX2-NEXT:    vpcmpeqq %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpackssdw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vmovmskps %ymm0, %eax
 ; AVX2-NEXT:    cmpb $-1, %al
 ; AVX2-NEXT:    sete %al
@@ -1582,7 +1581,6 @@ define i1 @icmp_v32i16_v32i1(<32 x i16>) {
 ; AVX2-NEXT:    vpcmpeqw %ymm2, %ymm1, %ymm1
 ; AVX2-NEXT:    vpcmpeqw %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpacksswb %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vpmovmskb %ymm0, %eax
 ; AVX2-NEXT:    cmpl $-1, %eax
 ; AVX2-NEXT:    sete %al
diff --git a/llvm/test/CodeGen/X86/vector-reduce-or-bool.ll b/llvm/test/CodeGen/X86/vector-reduce-or-bool.ll
index e08ec78..3162d25 100644
--- a/llvm/test/CodeGen/X86/vector-reduce-or-bool.ll
+++ b/llvm/test/CodeGen/X86/vector-reduce-or-bool.ll
@@ -1420,7 +1420,6 @@ define i1 @icmp_v8i64_v8i1(<8 x i64>) {
 ; AVX2-NEXT:    vpcmpeqq %ymm2, %ymm1, %ymm1
 ; AVX2-NEXT:    vpcmpeqq %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpackssdw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vmovmskps %ymm0, %eax
 ; AVX2-NEXT:    testb %al, %al
 ; AVX2-NEXT:    setne %al
@@ -1557,7 +1556,6 @@ define i1 @icmp_v32i16_v32i1(<32 x i16>) {
 ; AVX2-NEXT:    vpcmpeqw %ymm2, %ymm1, %ymm1
 ; AVX2-NEXT:    vpcmpeqw %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpacksswb %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vpmovmskb %ymm0, %eax
 ; AVX2-NEXT:    testl %eax, %eax
 ; AVX2-NEXT:    setne %al