[PATCH] D48366: [X86] Adding a check against i64 inputs in combineScalarToVector

[Mikhail Dvoretckii via Phabricator via llvm-commits]

mike.dvoretsky created this revision.
mike.dvoretsky added a reviewer: craig.topper.
Herald added subscribers: llvm-commits, hiraditya.

This patch adds checks to prevent creating scalar_to_vector DAG nodes with i64 inputs when folding scalar masking, fixing PR37879.


Repository:
  rL LLVM

https://reviews.llvm.org/D48366

Files:
  llvm/lib/Target/X86/X86ISelLowering.cpp
  llvm/test/CodeGen/X86/pr37879.ll


Index: llvm/test/CodeGen/X86/pr37879.ll
===================================================================
--- llvm/test/CodeGen/X86/pr37879.ll
+++ 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
Index: llvm/lib/Target/X86/X86ISelLowering.cpp
===================================================================
--- llvm/lib/Target/X86/X86ISelLowering.cpp
+++ llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -39124,7 +39124,8 @@
       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::i16 &&
+            Mask.getOperand(0).getValueType() != MVT::i64)
           Mask = Mask.getOperand(0);
         return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), MVT::v1i1, Mask);
       }
@@ -39138,11 +39139,12 @@
       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::i16 &&
+            Mask.getOperand(0).getValueType() != MVT::i64)
           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 an i16 or an i64. scalar_to_vector
+        // fails to select for those types, so the combine should be aborted.
+        if (Mask.getValueType() == MVT::i16 || Mask.getValueType() == MVT::i64)
           return SDValue();
         return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), MVT::v1i1, Mask);
       }


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D48366.152081.patch
Type: text/x-patch
Size: 2420 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180620/1f67c23d/attachment-0001.bin>