[X86][AVX2] combineSetCCMOVMSK - handle all_of patterns for PMOVMSKB(…

…PACKSSBW(LO(X), HI(X))) In the sign splat case, we can fold PMOVMSKB(PACKSSBW(LO(X), HI(X))) -> PMOVMSKB(BITCAST_v32i8(X)) without introducing a signmask + comparison (which unlike for any_of won't fold into a single TEST).
llvm · Jun 7, 2020 · ce677ef · ce677ef
1 parent a25f5cd
commit ce677ef
Show file tree

Hide file tree

Showing 3 changed files with 11 additions and 12 deletions.
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -40297,13 +40297,13 @@ static SDValue combineSetCCMOVMSK(SDValue EFLAGS, X86::CondCode &CC,
   // sign bits prior to the comparison with zero unless we know that
   // the vXi16 splats the sign bit down to the lower i8 half.
   // TODO: Handle all_of patterns.
-  if (IsAnyOf && Vec.getOpcode() == X86ISD::PACKSS && VecVT == MVT::v16i8) {
+  if (Vec.getOpcode() == X86ISD::PACKSS && VecVT == MVT::v16i8) {
     SDValue VecOp0 = Vec.getOperand(0);
     SDValue VecOp1 = Vec.getOperand(1);
     bool SignExt0 = DAG.ComputeNumSignBits(VecOp0) > 8;
     bool SignExt1 = DAG.ComputeNumSignBits(VecOp1) > 8;
     // PMOVMSKB(PACKSSBW(X, undef)) -> PMOVMSKB(BITCAST_v16i8(X)) & 0xAAAA.
-    if (CmpBits == 8 && VecOp1.isUndef()) {
+    if (IsAnyOf && CmpBits == 8 && VecOp1.isUndef()) {
       SDLoc DL(EFLAGS);
       SDValue Result = DAG.getBitcast(MVT::v16i8, VecOp0);
       Result = DAG.getNode(X86ISD::MOVMSK, DL, MVT::i32, Result);
@@ -40322,16 +40322,19 @@ static SDValue combineSetCCMOVMSK(SDValue EFLAGS, X86::CondCode &CC,
         VecOp1.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
         VecOp0.getOperand(0) == VecOp1.getOperand(0) &&
         VecOp0.getConstantOperandAPInt(1) == 0 &&
-        VecOp1.getConstantOperandAPInt(1) == 8) {
+        VecOp1.getConstantOperandAPInt(1) == 8 &&
+        (IsAnyOf || (SignExt0 && SignExt1))) {
       SDLoc DL(EFLAGS);
       SDValue Result = DAG.getBitcast(MVT::v32i8, VecOp0.getOperand(0));
       Result = DAG.getNode(X86ISD::MOVMSK, DL, MVT::i32, Result);
+      unsigned CmpMask = IsAnyOf ? 0 : 0xFFFFFFFF;
       if (!SignExt0 || !SignExt1) {
+        assert(IsAnyOf && "Only perform v16i16 signmasks for any_of patterns");
         Result = DAG.getNode(ISD::AND, DL, MVT::i32, Result,
                              DAG.getConstant(0xAAAAAAAA, DL, MVT::i32));
       }
       return DAG.getNode(X86ISD::CMP, DL, MVT::i32, Result,
-                         DAG.getConstant(0, DL, MVT::i32));
+                         DAG.getConstant(CmpMask, DL, MVT::i32));
     }
   }
 

diff --git a/llvm/test/CodeGen/X86/vector-compare-all_of.ll b/llvm/test/CodeGen/X86/vector-compare-all_of.ll
@@ -1304,10 +1304,8 @@ define i1 @bool_reduction_v16i16(<16 x i16> %x, <16 x i16> %y) {
 ; AVX2-LABEL: bool_reduction_v16i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT:    vpacksswb %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpmovmskb %xmm0, %eax
-; AVX2-NEXT:    cmpw $-1, %ax
+; AVX2-NEXT:    vpmovmskb %ymm0, %eax
+; AVX2-NEXT:    cmpl $-1, %eax
 ; AVX2-NEXT:    sete %al
 ; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq

diff --git a/llvm/test/CodeGen/X86/vector-reduce-and-bool.ll b/llvm/test/CodeGen/X86/vector-reduce-and-bool.ll
@@ -1269,10 +1269,8 @@ define i1 @icmp_v16i16_v16i1(<16 x i16>) {
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX2-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT:    vpacksswb %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpmovmskb %xmm0, %eax
-; AVX2-NEXT:    cmpw $-1, %ax
+; AVX2-NEXT:    vpmovmskb %ymm0, %eax
+; AVX2-NEXT:    cmpl $-1, %eax
 ; AVX2-NEXT:    sete %al
 ; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq