[PATCH] D26799: [X86][AVX512] Instructions fixups

[Elena Demikhovsky via llvm-commits]

delena added inline comments.


================
Comment at: lib/Target/X86/X86InstrAVX512.td:6431
 
-defm VCVTDQ2PD : avx512_cvtdq2pd<0xE6, "vcvtdq2pd", sint_to_fp, X86cvtdq2pd>, XS,
-                                EVEX_CD8<32, CD8VH>;
+defm VCVTDQ2PD : avx512_cvtdq2pd<0xE6, "vcvtdq2pd", sint_to_fp, X86cvtdq2pd, i128mem>,
+                                XS, EVEX_CD8<32, CD8VH>;
----------------
coby wrote:
> craig.topper wrote:
> > Shouldn't this be i64mem too? The only difference between VCVTDQ2PD and VCVTUDQ2PD is that one is signed and the other is unsigned. They should have the same load size.
> Note we are altering (in this case - preserving) the AVX512(f,vl) variants alone, which are to load from an xmm-width memory locations (according to formal documentation).
> quad-word instances are restricted to AVX/SSE.
> OTOH, VCVTUDQ2PD, is indeed spec'd to load from a quad-word.
I also think that it should be i64mem for XMM-destination form and i128mem for YMM-destination form.
So, in this case I agree with Craig. The bug is in documentation.


Repository:
  rL LLVM

https://reviews.llvm.org/D26799