[DAGCombiner] Add vector support to (sub -1, x) -> (xor x, -1) canonicalization

Improves commutation potential

llvm-svn: 284113

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index fab67e7..d01606e 100644 (file)
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -863,6 +863,17 @@ static bool isOneConstantOrOneSplatConstant(SDValue N) {
   return false;
 }
 
+// Determines if it is a constant integer of all ones or a splatted vector of a
+// constant integer of all ones (with no undefs).
+// Do not permit build vector implicit truncation.
+static bool isAllOnesConstantOrAllOnesSplatConstant(SDValue N) {
+  unsigned BitWidth = N.getScalarValueSizeInBits();
+  if (ConstantSDNode *Splat = isConstOrConstSplat(N))
+    return Splat->isAllOnesValue() &&
+           Splat->getAPIntValue().getBitWidth() == BitWidth;
+  return false;
+}
+
 SDValue DAGCombiner::ReassociateOps(unsigned Opc, const SDLoc &DL, SDValue N0,
                                     SDValue N1) {
   EVT VT = N0.getValueType();
@@ -1938,7 +1949,7 @@ SDValue DAGCombiner::visitSUB(SDNode *N) {
   }
 
   // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
-  if (isAllOnesConstant(N0))
+  if (isAllOnesConstantOrAllOnesSplatConstant(N0))
     return DAG.getNode(ISD::XOR, DL, VT, N1, N0);
 
   // fold A-(A-B) -> B

diff --git a/llvm/test/CodeGen/X86/combine-sub.ll b/llvm/test/CodeGen/X86/combine-sub.ll
index 45ffc2c..03978e7 100644 (file)
--- a/llvm/test/CodeGen/X86/combine-sub.ll
+++ b/llvm/test/CodeGen/X86/combine-sub.ll
@@ -50,14 +50,13 @@ define <4 x i32> @combine_vec_sub_negone(<4 x i32> %x) {
 ; SSE-LABEL: combine_vec_sub_negone:
 ; SSE:       # BB#0:
 ; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
-; SSE-NEXT:    psubd %xmm0, %xmm1
-; SSE-NEXT:    movdqa %xmm1, %xmm0
+; SSE-NEXT:    pxor %xmm1, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: combine_vec_sub_negone:
 ; AVX:       # BB#0:
 ; AVX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
-; AVX-NEXT:    vpsubd %xmm0, %xmm1, %xmm0
+; AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %1 = sub <4 x i32> <i32 -1, i32 -1, i32 -1, i32 -1>, %x
   ret <4 x i32> %1