[X86] Changing the check for valid inputs in combineScalarToVector

Changing the logic of scalar mask folding to check for valid input types rather
than against invalid ones, making it more robust and fixing PR37879.

Differential Revision: https://reviews.llvm.org/D48366

llvm-svn: 335323

2 files changed, 15 insertions, 6 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 314a944858b..27a0c51e3fd 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -39229,7 +39229,8 @@ static SDValue combineScalarToVector(SDNode *N, SelectionDAG &DAG) {
       if (C->getAPIntValue().isOneValue()) {
         SDValue Mask = Src.getOperand(0);
         if (Mask.getOpcode() == ISD::TRUNCATE &&
-            Mask.getOperand(0).getValueType() != MVT::i16)
+            (Mask.getOperand(0).getValueType() == MVT::i8 ||
+             Mask.getOperand(0).getValueType() == MVT::i32))
           Mask = Mask.getOperand(0);
         return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), MVT::v1i1, Mask);
       }
@@ -39243,11 +39244,11 @@ static SDValue combineScalarToVector(SDNode *N, SelectionDAG &DAG) {
       if (C->getAPIntValue().isOneValue()) {
         SDValue Mask = Src.getOperand(0).getOperand(0);
         if (Mask.getOpcode() == ISD::TRUNCATE &&
-            Mask.getOperand(0).getValueType() != MVT::i16)
+            (Mask.getOperand(0).getValueType() == MVT::i8 ||
+             Mask.getOperand(0).getValueType() == MVT::i32))
           Mask = Mask.getOperand(0);
-        // Check if the initial value is an i16. scalar_to_vector fails to
-        // select for that type, so the combine should be aborted.
-        if (Mask.getValueType() == MVT::i16)
+        // Check if the initial value is of a legal type for scalar_to_vector.
+        if (Mask.getValueType() != MVT::i8 && Mask.getValueType() != MVT::i32)
           return SDValue();
         return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), MVT::v1i1, Mask);
       }
diff --git a/llvm/test/CodeGen/X86/pr37879.ll b/llvm/test/CodeGen/X86/pr37879.ll
index 9beb6dbae00..033cad8ea0c 100644
--- a/llvm/test/CodeGen/X86/pr37879.ll
+++ b/llvm/test/CodeGen/X86/pr37879.ll
@@ -1,7 +1,15 @@
-; XFAIL: *
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -O3 < %s -mtriple=x86_64-apple-darwin -mattr=+avx512bw | FileCheck %s
 
 define double @foo(i32** nocapture readonly) #0 {
+; CHECK-LABEL: foo:
+; CHECK:       ## %bb.0:
+; CHECK-NEXT:    movq (%rax), %rax
+; CHECK-NEXT:    vcvtsi2sdq %rax, %xmm0, %xmm1
+; CHECK-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
+; CHECK-NEXT:    kmovd %eax, %k1
+; CHECK-NEXT:    vmovsd %xmm1, %xmm0, %xmm0 {%k1}
+; CHECK-NEXT:    retq
   %2 = load i64, i64* undef, align 8
   %3 = and i64 %2, 1
   %4 = icmp eq i64 %3, 0