On Wed, Oct 5, 2016 at 4:06 PM, Kyrill Tkachov
<kyrylo.tkac...@foss.arm.com> wrote:
> Hi all,
>
> I encountered a wrong-code issue with my WIP store merging pass when it was
> trying to encode HFmode constants.
> I am using native_encode_real to write the constants to a byte array and
> it's breaking on big-endian.
> For a 2-byte constant it ended up writing bytes at offsets 3 and 2 rather
> than 1 and 0.
>
> The fix in this patch makes the logic in native_encode_real match the logic
> in native_interpret_real,
> I just copied the logic across.
>
> I don't have a testcase against clean trunk that demonstrates the issue but
> with my store merging patch
Might be sth simple as
unsigned short foo (void)
{
unsigned short bar;
typedef HFmode alias_HF __attribute__((may_alias));
*(ailas_HFmode *)bar = HFmode constant;
return bar;
}
which should get a bogus integer representation from FRE.
> the testcase gcc.target/aarch64/advsimd-intrinsics/vldX.c starts failing
> without this patch because
> adjacent 16-bit float constants are not being merged correctly.
>
> Bootstrapped and tested on aarch64-none-linux-gnu.
> As this patch only affects the big-endian code path I also tested it on
> aarch64_be-none-elf.
>
> Ok for trunk? Do you think this needs to be backported?
>
> Thanks,
> Kyrill
>
> 2016-10-05 Kyrylo Tkachov <kyrylo.tkac...@arm.com>
>
> * fold-const.c (native_encode_real): Fix logic for selecting offset
> to write to when BYTES_BIG_ENDIAN.
