craig.topper added a comment.
Hi @aemerson, I'm not opposed to adding it back. But the clang policy for
vector builtins has always been that we won't support all the builtins that gcc
does and to encourage the use of the _mm_* wrappers which are guaranteed to
work in both compilers. It
aemerson added a comment.
Hi Craig,
The `__builtin_ia32_cvtdq2ps` builtin seems to be supported by gcc, and this
change breaks code which has been using this without problems for years. Can we
restore it?
Amara
Repository:
rC Clang
https://reviews.llvm.org/D46863
This revision was automatically updated to reflect the committed changes.
Closed by commit rC332882: [X86] Use __builtin_convertvector to implement some
of the packed integer to… (authored by ctopper, committed by ).
Changed prior to commit:
https://reviews.llvm.org/D46863?vs=146750=147853#toc
DavidKreitzer accepted this revision.
DavidKreitzer added a comment.
This revision is now accepted and ready to land.
I had actually done that as part of my initial review, so LGTM.
Repository:
rC Clang
https://reviews.llvm.org/D46863
___
craig.topper added a comment.
So I think we've covered the whether this is ok to do questions. If someone can
double check signed/unsigned and vector element sizes are all correct and
approve this that would be great.
Repository:
rC Clang
https://reviews.llvm.org/D46863
DavidKreitzer added a comment.
This looks good to me, Craig. I am not worried about the constant folding
issue, as I think constant folding these conversion intrinsics (assuming
round-to-nearest) is a perfectly valid optimization in the absence of
FENV_ACCESS. (FWIW, we don't do this constant