This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
RE: [PATCH, MIPS] Calling convention differs depending on the presence of MSA
Sameera Deshpande <Sameera.Deshpande@imgtec.com> writes:
> Hi Matthew,
>
> Please find attached updated patch as per our offline discussion.
>
> I have disabled return in registers for all vector float types, and
> updated the test case accordingly.
>
> Ok for trunk?
Hi Sameera,
Sorry for the slow response. I'd like to explain my thinking here
as I believe your change is right but you are actually ending up
fixing more issues than you expected!
There are 3 potential vector modes that are affected by your change
rather than just 2. V2SF (paired single), V4SF and V2DF (MSA).
The V4SF and V2DF changes are the ones needed for MSA as the test
case shows. These become valid modes with -mmsa so the float
vector types no longer get assigned BLKmode. (As an aside I think
it is a legacy bug that this function was ever defined in terms of
modes for o32).
The V2SF is an interesting one as it would not historically have
been a valid mode for o32 until o32 FP64 was re-invented. Now that
o32 FP64 is a compatible ABI extension we(I) have in fact allowed an
incompatible ABI change through as a vector of 2 floats will be
returned in registers for o32 FP64 and not the other o32 variants.
As an aside... The integer vector types up to 16-bytes get returned
in registers also potentially by chance because...
1) We have TImode support in the backend
2) TImode is 16-bytes and hence the type->mode logic will fall back
to looking for a wider integer mode for integer vectors and find
TImode
3) The mips_hard_regno_mode_ok_p logic does not limit the size of
a mode allowed in GPRs as long as the first register is even.
4) TImode is not BLKmode so is not forced to memory here.
I'm doubtful this is by design but it is what it is. Let it
hereby be part of the MIPS o32 ABI definition!
Are you sure you need the dg-skip-if on the test case? Perhaps just
-O0. The test does not need vectorization as it is explicitly
Vectorized but it may break the scan-assembler at -O0. Also please
can you apply the normal code style to the tests?
v4f32
fcmpOeqVector4 (v4f32 a, v4f32 b)
{
return a + b;
}
As this is an ABI fix, just wait a day or so in case Catherine has
any comment otherwise OK to commit.
Thanks,
Matthew
>
> - Thanks and regards,
> Sameera D.
> ________________________________________
> From: Sameera Deshpande
> Sent: 08 February 2017 14:10:52
> To: Matthew Fortune
> Cc: gcc-patches@gcc.gnu.org
> Subject: [PATCH, MIPS] Calling convention differs depending on the
> presence of MSA
>
> Hi Matthew,
>
> Please find attached the patch to fix the calling convention issue,
> where argument and result passing convention differed for MSA and non-
> MSA variants.
>
> The implementation of TARGET_RETURN_IN_MEMORY is altered to block V4SF
> to be returned in registers.
>
> Ok for trunk?
>
> - Thanks and regards,
> Sameera D.
>
>
> Changelog:
> gcc/
> * config/mips/mips.c (mips_return_in_memory) : Restrict V4SFmode to
> be returned in registers.
>
> gcc/testsuite/
> * gcc.target/mips/msa-fp-cc.c : New testcase.