[X86] Fold vXi1 OR(KSHIFTL(X,NumElts/2),Y) -> KUNPCK

Convert shift+or bool vector patterns into CONCAT_VECTORS if we know this will be lowered to KUNPCK (which requires 16+ vector elements).

Fixes PR32547

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -42625,6 +42625,34 @@ static SDValue combineOr(SDNode *N, SelectionDAG &DAG,
   if (SDValue R = combineLogicBlendIntoPBLENDV(N, DAG, Subtarget))
     return R;
 
+  // Combine OR(X,KSHIFTL(Y,Elts/2)) -> CONCAT_VECTORS(X,Y) == KUNPCK(X,Y).
+  // Combine OR(KSHIFTL(X,Elts/2),Y) -> CONCAT_VECTORS(Y,X) == KUNPCK(Y,X).
+  // iff the upper elements of the non-shifted arg are zero.
+  // KUNPCK require 16+ bool vector elements.
+  if (N0.getOpcode() == X86ISD::KSHIFTL || N1.getOpcode() == X86ISD::KSHIFTL) {
+    unsigned NumElts = VT.getVectorNumElements();
+    unsigned HalfElts = NumElts / 2;
+    APInt UpperElts = APInt::getHighBitsSet(NumElts, HalfElts);
+    if (NumElts >= 16 && N1.getOpcode() == X86ISD::KSHIFTL &&
+        N1.getConstantOperandAPInt(1) == HalfElts &&
+        DAG.MaskedValueIsZero(N0, APInt(1, 1), UpperElts)) {
+      SDLoc dl(N);
+      return DAG.getNode(
+          ISD::CONCAT_VECTORS, dl, VT,
+          extractSubVector(N0, 0, DAG, dl, HalfElts),
+          extractSubVector(N1.getOperand(0), 0, DAG, dl, HalfElts));
+    }
+    if (NumElts >= 16 && N0.getOpcode() == X86ISD::KSHIFTL &&
+        N0.getConstantOperandAPInt(1) == HalfElts &&
+        DAG.MaskedValueIsZero(N1, APInt(1, 1), UpperElts)) {
+      SDLoc dl(N);
+      return DAG.getNode(
+          ISD::CONCAT_VECTORS, dl, VT,
+          extractSubVector(N1, 0, DAG, dl, HalfElts),
+          extractSubVector(N0.getOperand(0), 0, DAG, dl, HalfElts));
+    }
+  }
+
   // Attempt to recursively combine an OR of shuffles.
   if (VT.isVector() && (VT.getScalarSizeInBits() % 8) == 0) {
     SDValue Op(N, 0);

llvm/test/CodeGen/X86/vector-shuffle-v1.ll

@@ -892,10 +892,7 @@ define void @PR32547(<8 x float> %a, <8 x float> %b, <8 x float> %c, <8 x float>
 ; AVX512F-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
 ; AVX512F-NEXT:    vcmpltps %zmm1, %zmm0, %k0
 ; AVX512F-NEXT:    vcmpltps %zmm3, %zmm2, %k1
-; AVX512F-NEXT:    kshiftlw $8, %k0, %k0
-; AVX512F-NEXT:    kshiftlw $8, %k1, %k1
-; AVX512F-NEXT:    kshiftrw $8, %k1, %k1
-; AVX512F-NEXT:    korw %k1, %k0, %k1
+; AVX512F-NEXT:    kunpckbw %k1, %k0, %k1
 ; AVX512F-NEXT:    vxorps %xmm0, %xmm0, %xmm0
 ; AVX512F-NEXT:    vmovaps %zmm0, (%rdi) {%k1}
 ; AVX512F-NEXT:    vzeroupper
@@ -905,8 +902,7 @@ define void @PR32547(<8 x float> %a, <8 x float> %b, <8 x float> %c, <8 x float>
 ; AVX512VL:       # %bb.0: # %entry
 ; AVX512VL-NEXT:    vcmpltps %ymm1, %ymm0, %k0
 ; AVX512VL-NEXT:    vcmpltps %ymm3, %ymm2, %k1
-; AVX512VL-NEXT:    kshiftlw $8, %k0, %k0
-; AVX512VL-NEXT:    korw %k1, %k0, %k1
+; AVX512VL-NEXT:    kunpckbw %k1, %k0, %k1
 ; AVX512VL-NEXT:    vxorps %xmm0, %xmm0, %xmm0
 ; AVX512VL-NEXT:    vmovaps %zmm0, (%rdi) {%k1}
 ; AVX512VL-NEXT:    vzeroupper
@@ -916,8 +912,7 @@ define void @PR32547(<8 x float> %a, <8 x float> %b, <8 x float> %c, <8 x float>
 ; VL_BW_DQ:       # %bb.0: # %entry
 ; VL_BW_DQ-NEXT:    vcmpltps %ymm1, %ymm0, %k0
 ; VL_BW_DQ-NEXT:    vcmpltps %ymm3, %ymm2, %k1
-; VL_BW_DQ-NEXT:    kshiftlw $8, %k0, %k0
-; VL_BW_DQ-NEXT:    korw %k1, %k0, %k1
+; VL_BW_DQ-NEXT:    kunpckbw %k1, %k0, %k1
 ; VL_BW_DQ-NEXT:    vxorps %xmm0, %xmm0, %xmm0
 ; VL_BW_DQ-NEXT:    vmovaps %zmm0, (%rdi) {%k1}
 ; VL_BW_DQ-NEXT:    vzeroupper
@@ -945,10 +940,7 @@ define void @PR32547_swap(<8 x float> %a, <8 x float> %b, <8 x float> %c, <8 x f
 ; AVX512F-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
 ; AVX512F-NEXT:    vcmpltps %zmm1, %zmm0, %k0
 ; AVX512F-NEXT:    vcmpltps %zmm3, %zmm2, %k1
-; AVX512F-NEXT:    kshiftlw $8, %k0, %k0
-; AVX512F-NEXT:    kshiftlw $8, %k1, %k1
-; AVX512F-NEXT:    kshiftrw $8, %k1, %k1
-; AVX512F-NEXT:    korw %k0, %k1, %k1
+; AVX512F-NEXT:    kunpckbw %k1, %k0, %k1
 ; AVX512F-NEXT:    vxorps %xmm0, %xmm0, %xmm0
 ; AVX512F-NEXT:    vmovaps %zmm0, (%rdi) {%k1}
 ; AVX512F-NEXT:    vzeroupper
@@ -958,8 +950,7 @@ define void @PR32547_swap(<8 x float> %a, <8 x float> %b, <8 x float> %c, <8 x f
 ; AVX512VL:       # %bb.0: # %entry
 ; AVX512VL-NEXT:    vcmpltps %ymm1, %ymm0, %k0
 ; AVX512VL-NEXT:    vcmpltps %ymm3, %ymm2, %k1
-; AVX512VL-NEXT:    kshiftlw $8, %k0, %k0
-; AVX512VL-NEXT:    korw %k0, %k1, %k1
+; AVX512VL-NEXT:    kunpckbw %k1, %k0, %k1
 ; AVX512VL-NEXT:    vxorps %xmm0, %xmm0, %xmm0
 ; AVX512VL-NEXT:    vmovaps %zmm0, (%rdi) {%k1}
 ; AVX512VL-NEXT:    vzeroupper
@@ -969,8 +960,7 @@ define void @PR32547_swap(<8 x float> %a, <8 x float> %b, <8 x float> %c, <8 x f
 ; VL_BW_DQ:       # %bb.0: # %entry
 ; VL_BW_DQ-NEXT:    vcmpltps %ymm1, %ymm0, %k0
 ; VL_BW_DQ-NEXT:    vcmpltps %ymm3, %ymm2, %k1
-; VL_BW_DQ-NEXT:    kshiftlw $8, %k0, %k0
-; VL_BW_DQ-NEXT:    korw %k0, %k1, %k1
+; VL_BW_DQ-NEXT:    kunpckbw %k1, %k0, %k1
 ; VL_BW_DQ-NEXT:    vxorps %xmm0, %xmm0, %xmm0
 ; VL_BW_DQ-NEXT:    vmovaps %zmm0, (%rdi) {%k1}
 ; VL_BW_DQ-NEXT:    vzeroupper