RKSimon added inline comments.
================
Comment at: emmintrin.h:1607
+///
+/// This intrinsic corresponds to the <c> VMOVSD / MOVSD </c> instruction.
+///
----------------
kromanova wrote:
> kromanova wrote:
> > kromanova wrote:
> > > probinson wrote:
> > > > should this be VMOVQ/MOVQ instead?
> > > Probably yes. Let me know if you have a different opinion.
> > >
> > > If I use this intrinsic by itself, clang generates VMOVSD instruction.
> > > It happens because the default domain is chooses to generate smaller
> > > instruction code.
> > > I got confused because I couldn't find Intel's documentation about
> > > _mm_loadu_si64, so I just wrote a test like the one below and looked what
> > > instructions got generated.
> > >
> > > ```
> > > __m128i foo22 (void const * __a)
> > > {
> > > return _mm_loadu_si64 (__a);
> > > }
> > > ```
> > >
> > > However, if I change the test and use an intrisic to add 2 64-bit
> > > integers after the load intrinsics, I can see that VMOVQ instruction gets
> > > generated.
> > >
> > > ```
> > > __m128d foo44 (double const * __a)
> > > {
> > > __m128i first = _mm_loadu_si64 (__a);
> > > __m128i second = _mm_loadu_si64 (__a);
> > > return _mm_add_epi64(first, second);
> > >
> > > }
> > > ```
> > >
> > > So, as you see clang could generate either VMOVSD/MOVSD or VMOVSQ/MOVSQ.
> > > I think it makes sense to change the documentation as Paul suggested:
> > >
> > > /// This intrinsic corresponds to the VMOVSQ/MOVSQ.
> > >
> > > Or, alternatively, we could list all the instructions that correspond to
> > > this intrinsics:
> > >
> > > /// This intrinsic corresponds to the VMOVSQ/MOVSQ/VMOVSD/MOVSD.
> > >
> > >
> > It will be interesting to hear Asaf Badoug opinion, since he added this
> > intrisic. He probably has access to Intel's documentation for this
> > intrinsic too (which I wasn't able to find online).
> There is a similar situation for one intrisic just a few lines above, namely
> _mm_loadu_pd. It could generate either VMOVUPD / MOVUPD or VMOVUPS/MOVUPS
> instructions.
> I have actually asked Simon question about it offline just a couple of days
> ago.
>
> I decided to kept referring to VMOVUPD / MOVUPD as a corresponding
> instruction for _mm_loadu_pd. However, if we end up doing things differently
> for _mm_loadu_si64, we need to do a similar change to _mm_loadu_pd (and
> probably to some other intrinsics).
It should be VMOVQ/MOVQ (note NOT VMOVSQ/MOVSQ!). Whatever the domain fixup
code does to it, that was the original intent of the code and matches what
other compilers says it will (probably) be.
Repository:
rL LLVM
https://reviews.llvm.org/D28503
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
