https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87812
Iain Sandoe <iains at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2018-10-30
Ever confirmed|0 |1
--- Comment #2 from Iain Sandoe <iains at gcc dot gnu.org> ---
(In reply to H.J. Lu from comment #1)
> Without AVX, GCC treats
>
> typedef qi __attribute__((vector_size (32))) v32qi;
>
> as a type of 32 bytes, which is passed in memory. Clang does
>
> # clang -S -msse2 x.c -O2
> ...
> movaps .LCPI0_0(%rip), %xmm0 # xmm0 =
> [48,49,50,51,52,53,54,55,56,57,97,98,99,100,101,102]
> movaps %xmm0, %xmm1
> retq
> # clang -S -msse2 x.c -O2 -m32
> ...
> movaps .LCPI0_0, %xmm0 # xmm0 =
> [48,49,50,51,52,53,54,55,56,57,97,98,99,100,101,102]
> movaps %xmm0, %xmm1
> retl
>
> They look very odd to me. You either pass it as a vector or you don't.
> You can't have it both ways.
(when there's no native size support on the hardware) breaking objects into
pieces to pass in registers is IMO a reasonable ABI approach.
However, I am not here to argue for one or the other - what's needed is a clear
statement of the "correct" ABI ad then :
* a bug filed against the implementation(s) which don't comply.
* a plan to support the mismatch on systems "in the wild".
I don't currently have access to icc (or other compilers) so no idea what they
do.
FWIW clang -mno-sse passes in memory as per the GCC code.
