New patch: [AArch64] [BE] [1/2] Make large opaque integer modes endianness-safe.
David Sherwood
david.sherwood@arm.com
Thu Nov 27 15:03:00 GMT 2014
> On 18 November 2014 10:14, David Sherwood <david.sherwood@arm.com> wrote:
> > Hi Christophe,
> >
> > Ah sorry. My mistake - it fixes this in bugzilla:
> >
> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59810
>
> I did look at that PR, but since it has no testcase attached, I was unsure.
> And I am still not :-)
> PR 59810 is "[AArch64] LDn/STn implementations are not ABI-conformant
> for bigendian."
> but the advsimd-intrinsics/vldX.c and vldX_lane.c now PASS with Alan's
> patches on aarch64_be, so I thought Alan's patches solve PR59810.
>
> What am I missing?
Hi Christophe,
I think probably this is our fault for making our lives way too difficult and
artificially splitting all these patches up. :)
Alan's patch:
https://gcc.gnu.org/ml/gcc-patches/2014-10/msg00952.html
fixes some issues on aarch64_be, but also causes regressions. For example,
====
Tests that now fail, but worked before:
aarch64_be-elf-aem: gcc.dg/vect/slp-perm-8.c -flto -ffat-lto-objects execution test
aarch64_be-elf-aem: gcc.dg/vect/slp-perm-8.c execution test
aarch64_be-elf-aem: gcc.dg/vect/vect-over-widen-1-big-array.c -flto -ffat-lto-objects execution test
...
Tests that now work, but didn't before:
aarch64_be-elf-aem: gcc.dg/vect/fast-math-vect-complex-3.c execution test
aarch64_be-elf-aem: gcc.dg/vect/if-cvt-stores-vect-ifcvt-18.c execution test
aarch64_be-elf-aem: gcc.dg/vect/no-scevccp-outer-10a.c execution test
...
====
His patch is only half of the story and must be applied at the same time as the
"[AArch64] [BE] [1/2] Make large opaque integer modes endianness-safe."
patch. With both patches applied the result looks much healthier:
====
# Comparing 1 common sum files
## /bin/sh ./src/gcc/contrib/compare_tests /tmp/gxx-sum1.10051 /tmp/gxx-sum2.10051
Tests that now work, but didn't before:
aarch64_be-elf-aem: gcc.dg/torture/pr52028.c -O3 -fomit-frame-pointer execution test
aarch64_be-elf-aem: gcc.dg/torture/pr52028.c -O3 -fomit-frame-pointer -funroll-all-loops -finline-functions execution test
aarch64_be-elf-aem: gcc.dg/torture/pr52028.c -O3 -fomit-frame-pointer -funroll-loops execution test
...
====
with no new regressions. After applying both patches the aarch64_be gcc testsuite is
on a parity with the aarch64 testsuite. Furthermore, after applying both of these patches:
"[AArch64] [BE] [1/2] Make large opaque integer modes endianness-safe"
"[AArch64] [BE] Fix vector load/stores to not use ld1/st1"
it then becomes safe for us to remove the CCMC macro, which is the cause of
unnecessary spills to the stack for certain auto-vectorised code. So really I
suppose when I posted my second patch
"[AArch64] [BE] [2/2] Make large opaque integer modes endianness-safe"
I should have really just called this
"[AArch64] [BE] Remove CCMC for aarch64"
in order to make it clear exactly what the purpose of these patches is.
Kind Regards,
David Sherwood.
More information about the Gcc-patches
mailing list