[X86][AVX] combineBitcastvxi1 - peek through bitops to determine size of original...

We were only testing for direct SETCC results - this allows us to peek through AND/OR/XOR combinations of the comparison results as well.

There's a missing SEXT(PACKSS) fold that I need to investigate for v8i1 cases before I can enable it there as well.

llvm-svn: 361716

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 170e3cf..8b6edaa 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -34126,6 +34126,21 @@ static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
                      EltNo);
 }
 
+// Helper to peek through bitops/setcc to determine size of source vector.
+// Allows combineBitcastvxi1 to determine what size vector generated a <X x i1>.
+static bool checkBitcastSrcVectorSize(SDValue Src, unsigned Size) {
+  switch (Src.getOpcode()) {
+  case ISD::SETCC:
+    return Src.getOperand(0).getValueSizeInBits() == Size;
+  case ISD::AND:
+  case ISD::XOR:
+  case ISD::OR:
+    return checkBitcastSrcVectorSize(Src.getOperand(0), Size) &&
+           checkBitcastSrcVectorSize(Src.getOperand(1), Size);
+  }
+  return false;
+}
+
 // Try to match patterns such as
 // (i16 bitcast (v16i1 x))
 // ->
@@ -34174,10 +34189,8 @@ static SDValue combineBitcastvxi1(SelectionDAG &DAG, EVT VT, SDValue Src,
     SExtVT = MVT::v4i32;
     // For cases such as (i4 bitcast (v4i1 setcc v4i64 v1, v2))
     // sign-extend to a 256-bit operation to avoid truncation.
-    if (Src.getOpcode() == ISD::SETCC && Subtarget.hasAVX() &&
-        Src.getOperand(0).getValueType().is256BitVector()) {
+    if (Subtarget.hasAVX() && checkBitcastSrcVectorSize(Src, 256))
       SExtVT = MVT::v4i64;
-    }
     break;
   case MVT::v8i1:
     SExtVT = MVT::v8i16;
@@ -34186,6 +34199,7 @@ static SDValue combineBitcastvxi1(SelectionDAG &DAG, EVT VT, SDValue Src,
     // If the setcc operand is 128-bit, prefer sign-extending to 128-bit over
     // 256-bit because the shuffle is cheaper than sign extending the result of
     // the compare.
+    // TODO : use checkBitcastSrcVectorSize
     if (Src.getOpcode() == ISD::SETCC && Subtarget.hasAVX() &&
         (Src.getOperand(0).getValueType().is256BitVector() ||
          Src.getOperand(0).getValueType().is512BitVector())) {

diff --git a/llvm/test/CodeGen/X86/bitcast-and-setcc-256.ll b/llvm/test/CodeGen/X86/bitcast-and-setcc-256.ll
index 85ae7c0..b982cde 100644 (file)
--- a/llvm/test/CodeGen/X86/bitcast-and-setcc-256.ll
+++ b/llvm/test/CodeGen/X86/bitcast-and-setcc-256.ll
@@ -55,18 +55,18 @@ define i4 @v4i64(<4 x i64> %a, <4 x i64> %b, <4 x i64> %c, <4 x i64> %d) {
 ;
 ; AVX1-LABEL: v4i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm4
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm4
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm5
+; AVX1-NEXT:    vpcmpgtq %xmm4, %xmm5, %xmm4
 ; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtq %xmm3, %xmm2, %xmm1
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm1
-; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm3
-; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm1
+; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm4
+; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm4, %xmm1
 ; AVX1-NEXT:    vpcmpgtq %xmm3, %xmm2, %xmm2
-; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpackssdw %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vmovmskps %xmm0, %eax
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm2, %ymm1
+; AVX1-NEXT:    vandpd %ymm1, %ymm0, %ymm0
+; AVX1-NEXT:    vmovmskpd %ymm0, %eax
 ; AVX1-NEXT:    # kill: def $al killed $al killed $eax
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
@@ -76,9 +76,7 @@ define i4 @v4i64(<4 x i64> %a, <4 x i64> %b, <4 x i64> %c, <4 x i64> %d) {
 ; AVX2-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    vpcmpgtq %ymm3, %ymm2, %ymm1
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT:    vpackssdw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vmovmskps %xmm0, %eax
+; AVX2-NEXT:    vmovmskpd %ymm0, %eax
 ; AVX2-NEXT:    # kill: def $al killed $al killed $eax
 ; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq
@@ -126,9 +124,7 @@ define i4 @v4f64(<4 x double> %a, <4 x double> %b, <4 x double> %c, <4 x double>
 ; AVX12-NEXT:    vcmpltpd %ymm0, %ymm1, %ymm0
 ; AVX12-NEXT:    vcmpltpd %ymm2, %ymm3, %ymm1
 ; AVX12-NEXT:    vandpd %ymm1, %ymm0, %ymm0
-; AVX12-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX12-NEXT:    vpackssdw %xmm1, %xmm0, %xmm0
-; AVX12-NEXT:    vmovmskps %xmm0, %eax
+; AVX12-NEXT:    vmovmskpd %ymm0, %eax
 ; AVX12-NEXT:    # kill: def $al killed $al killed $eax
 ; AVX12-NEXT:    vzeroupper
 ; AVX12-NEXT:    retq