Fix missed case of switching getConstant to getTargetConstant. Try 2.

[saugustine]

Summary

This fixes a crasher introduced by r372338.

[RKSimon]

The test file doesn't look like it needs GlobalIsel - rename it llvm/test/CodeGen/X86/isel-blendi-gettargetconstant.ll

llvm/test/CodeGen/X86/GlobalISel/isel-blendi-gettargetconstant.ll

1 ↗

(On Diff #221077)

RUN: llc < %s -mtriple=x86_64-linux-gnu -mattr=sse4.2 | FileCheck %s

And then run the file through the utils/update_llc_test_checks.py script

3 ↗

(On Diff #221077)

Remove #0

12 ↗

(On Diff #221077)

Remove this

[arsenm]

llvm/test/CodeGen/X86/GlobalISel/isel-blendi-gettargetconstant.ll

1 ↗

(On Diff #221077)

This is not a GlobalISel test, and should not go in the GlobalISel directory

[saugustine]

Addressing comments.

[saugustine]

llvm/test/CodeGen/X86/GlobalISel/isel-blendi-gettargetconstant.ll

1 ↗

(On Diff #221077)

Where should it go?

I'm inclined just to drop the test case altogether. Most of the other code paths that needed the switch from getConstant to getTargetConstant don't have test cases either.

[craig.topper]

llvm/test/CodeGen/X86/GlobalISel/isel-blendi-gettargetconstant.ll

1 ↗

(On Diff #221077)

This shouldn't be in the GlobalISel directory. It doesn't enable GlobalIsel.

1 ↗

(On Diff #221077)

It can just go in test/CodeGen/X86. That's where all the SelectionDAG based tests are. Only tests that enable the non-default global instruction selection should be in the GlobalISel subdirectory.

[saugustine]

Move test case out of GlobalISel.

[echristo]

Looks much better, but then again I approved the last one ;)

Let's let Craig or Simon give an ack too.

[saugustine]

We've been blocked on this for a day now, so going ahead and checking it in. Will wait and watch for problems.

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

if (EltVT == MVT::f32) {
         // a vector, we prefer to generate a blend with immediate rather
         // than an insertps. Blends are simpler operations in hardware and so
         // will always have equal or better performance than insertps.
         // But if optimizing for size and there's a load folding opportunity,
         // generate insertps because blendps does not have a 32-bit memory
         // operand form.
         N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
         return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1,
-                           DAG.getConstant(1, dl, MVT::i8));
+                           DAG.getTargetConstant(1, dl, MVT::i8));
       }
       // Create this as a scalar to vector..
       N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
       return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1,
                          DAG.getTargetConstant(IdxVal << 4, dl, MVT::i8));
     }
 
     // PINSR* works with constant index.

llvm/trunk/test/CodeGen/X86/isel-blendi-gettargetconstant.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-linux-gnu -mattr=sse4.2 | FileCheck %s
+
+define void @csrot_(float* %0) {
+; CHECK-LABEL: csrot_:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; CHECK-NEXT:    xorps %xmm0, %xmm1
+; CHECK-NEXT:    blendps {{.*#+}} xmm1 = xmm1[0],mem[1,2,3]
+; CHECK-NEXT:    movlps %xmm1, (%rax)
+; CHECK-NEXT:    retq
+1:
+  %2 = load float, float* %0, align 4
+  %3 = fsub float -0.000000e+00, %2
+  %4 = insertelement <2 x float> <float undef, float -0.000000e+00>, float %3, i32 0
+  store <2 x float> %4, <2 x float>* undef, align 8
+  ret void
+}