[DAGCombiner] Canonicalize the order of a chain of INSERT_SUBVECTORs.

Based on similar code for INSERT_VECTOR_ELT.

llvm-svn: 294110

2 files changed, 40 insertions, 22 deletions

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index d38a9fca019..aad07d8ae3c 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -14547,15 +14547,35 @@ SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
     return DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N), VT, N0.getOperand(0),
                        N1, N2);
 
+  if (!isa<ConstantSDNode>(N2))
+    return SDValue();
+
+  unsigned InsIdx = cast<ConstantSDNode>(N2)->getZExtValue();
+
+  // Canonicalize insert_subvector dag nodes.
+  // Example:
+  // (insert_subvector (insert_subvector A, Idx0), Idx1)
+  // -> (insert_subvector (insert_subvector A, Idx1), Idx0)
+  if (N0.getOpcode() == ISD::INSERT_SUBVECTOR && N0.hasOneUse() &&
+      N1.getValueType() == N0.getOperand(1).getValueType() &&
+      isa<ConstantSDNode>(N0.getOperand(2))) {
+    unsigned OtherIdx = cast<ConstantSDNode>(N0.getOperand(2))->getZExtValue();
+    if (InsIdx < OtherIdx) {
+      // Swap nodes.
+      SDValue NewOp = DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N), VT,
+                                  N0.getOperand(0), N1, N2);
+      AddToWorklist(NewOp.getNode());
+      return DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N0.getNode()),
+                         VT, NewOp, N0.getOperand(1), N0.getOperand(2));
+    }
+  }
+
   if (N0.getValueType() != N1.getValueType())
     return SDValue();
 
   // If the input vector is a concatenation, and the insert replaces
   // one of the halves, we can optimize into a single concat_vectors.
-  if (N0.getOpcode() == ISD::CONCAT_VECTORS && N0->getNumOperands() == 2 &&
-      isa<ConstantSDNode>(N2)) {
-    unsigned InsIdx = cast<ConstantSDNode>(N2)->getZExtValue();
-
+  if (N0.getOpcode() == ISD::CONCAT_VECTORS && N0->getNumOperands() == 2) {
     // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
     // (concat_vectors Z, Y)
     if (InsIdx == 0)
diff --git a/llvm/test/CodeGen/X86/avx512-mask-op.ll b/llvm/test/CodeGen/X86/avx512-mask-op.ll
index 41fb19f38e0..6bd28b18b23 100644
--- a/llvm/test/CodeGen/X86/avx512-mask-op.ll
+++ b/llvm/test/CodeGen/X86/avx512-mask-op.ll
@@ -509,18 +509,13 @@ define <64 x i8> @test16(i64 %x) {
 ; KNL-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
 ; KNL-NEXT:    kmovw (%rsp), %k1
 ; KNL-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
-; KNL-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
+; KNL-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k2} {z}
 ; KNL-NEXT:    vpmovdb %zmm0, %xmm0
-; KNL-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k2} {z}
+; KNL-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
 ; KNL-NEXT:    vpmovdb %zmm1, %xmm1
-; KNL-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
 ; KNL-NEXT:    movl $1, %eax
-; KNL-NEXT:    vpinsrb $5, %eax, %xmm0, %xmm0
-; KNL-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
-; KNL-NEXT:    vpsllw $7, %ymm0, %ymm0
-; KNL-NEXT:    vpand {{.*}}(%rip), %ymm0, %ymm0
-; KNL-NEXT:    vpxor %ymm1, %ymm1, %ymm1
-; KNL-NEXT:    vpcmpgtb %ymm0, %ymm1, %ymm0
+; KNL-NEXT:    vpinsrb $5, %eax, %xmm1, %xmm1
+; KNL-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
 ; KNL-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
 ; KNL-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
 ; KNL-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
@@ -528,6 +523,10 @@ define <64 x i8> @test16(i64 %x) {
 ; KNL-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z}
 ; KNL-NEXT:    vpmovdb %zmm2, %xmm2
 ; KNL-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
+; KNL-NEXT:    vpsllw $7, %ymm0, %ymm0
+; KNL-NEXT:    vpand {{.*}}(%rip), %ymm0, %ymm0
+; KNL-NEXT:    vpxor %ymm2, %ymm2, %ymm2
+; KNL-NEXT:    vpcmpgtb %ymm0, %ymm2, %ymm0
 ; KNL-NEXT:    movq %rbp, %rsp
 ; KNL-NEXT:    popq %rbp
 ; KNL-NEXT:    retq
@@ -573,20 +572,15 @@ define <64 x i8> @test17(i64 %x, i32 %y, i32 %z) {
 ; KNL-NEXT:    movl %edi, {{[0-9]+}}(%rsp)
 ; KNL-NEXT:    kmovw (%rsp), %k1
 ; KNL-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
-; KNL-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
+; KNL-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k2} {z}
 ; KNL-NEXT:    vpmovdb %zmm0, %xmm0
-; KNL-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k2} {z}
+; KNL-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
 ; KNL-NEXT:    vpmovdb %zmm1, %xmm1
-; KNL-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
 ; KNL-NEXT:    xorl %eax, %eax
 ; KNL-NEXT:    cmpl %edx, %esi
 ; KNL-NEXT:    setg %al
-; KNL-NEXT:    vpinsrb $5, %eax, %xmm0, %xmm0
-; KNL-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
-; KNL-NEXT:    vpsllw $7, %ymm0, %ymm0
-; KNL-NEXT:    vpand {{.*}}(%rip), %ymm0, %ymm0
-; KNL-NEXT:    vpxor %ymm1, %ymm1, %ymm1
-; KNL-NEXT:    vpcmpgtb %ymm0, %ymm1, %ymm0
+; KNL-NEXT:    vpinsrb $5, %eax, %xmm1, %xmm1
+; KNL-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
 ; KNL-NEXT:    kmovw {{[0-9]+}}(%rsp), %k1
 ; KNL-NEXT:    kmovw {{[0-9]+}}(%rsp), %k2
 ; KNL-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
@@ -594,6 +588,10 @@ define <64 x i8> @test17(i64 %x, i32 %y, i32 %z) {
 ; KNL-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z}
 ; KNL-NEXT:    vpmovdb %zmm2, %xmm2
 ; KNL-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
+; KNL-NEXT:    vpsllw $7, %ymm0, %ymm0
+; KNL-NEXT:    vpand {{.*}}(%rip), %ymm0, %ymm0
+; KNL-NEXT:    vpxor %ymm2, %ymm2, %ymm2
+; KNL-NEXT:    vpcmpgtb %ymm0, %ymm2, %ymm0
 ; KNL-NEXT:    movq %rbp, %rsp
 ; KNL-NEXT:    popq %rbp
 ; KNL-NEXT:    retq