[PATCH] D19261: X86 _comi_ intrinsics - Fixed lowering

[David Kreitzer via llvm-commits]

DavidKreitzer added a comment.

Hi Elena,

You didn't answer my question about the vpmovzx changes. Are they somehow related to the comi changes?

Thanks,
-Dave


================
Comment at: ../lib/Target/X86/X86ISelLowering.cpp:17496
@@ -17483,3 +17495,3 @@
       return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
     }
     case COMI_RM: { // Comparison intrinsics with Sae
----------------
This still isn't quite correct. NaN comparisons will give the wrong results for the [u]comilt and [u]comile intrinsics. They are supposed to behave like the C "<" and "<=" operators, which return false for comparisons involving NaN.

In order to compute _mm_comilt_ss(A, B) with just one SETcc, you need to generate code just like _mm_comigt_ss(B, A).

Did you try running the ex26.c test that I sent you?



Repository:
  rL LLVM

http://reviews.llvm.org/D19261