[PATCH] Fix PR81921

[Uros Bizjak]

On Wed, Aug 23, 2017 at 11:55 AM, Richard Biener <rguenther@suse.de> wrote:
> On Wed, 23 Aug 2017, Uros Bizjak wrote:
>
>> On Wed, Aug 23, 2017 at 11:40 AM, Uros Bizjak <ubizjak@gmail.com> wrote:
>> > On Wed, Aug 23, 2017 at 11:22 AM, Richard Biener <rguenther@suse.de> wrote:
>> >>> >  /* If the machine does not have a case insn that compares the bounds,
>> >>> > Index: gcc/config/i386/i386.c
>> >>> > ===================================================================
>> >>> > --- gcc/config/i386/i386.c      (revision 251275)
>> >>> > +++ gcc/config/i386/i386.c      (working copy)
>> >>> > @@ -7369,7 +7369,8 @@ ix86_valid_target_attribute_tree (tree a
>> >>> >        || opts->x_target_flags != def->x_target_flags
>> >>> >        || option_strings[IX86_FUNCTION_SPECIFIC_ARCH]
>> >>> >        || option_strings[IX86_FUNCTION_SPECIFIC_TUNE]
>> >>> > -      || enum_opts_set.x_ix86_fpmath)
>> >>> > +      || enum_opts_set.x_ix86_fpmath
>> >>> > +      || !TARGET_64BIT_P (opts->x_ix86_isa_flags))
>> >>>
>> >>> You should use (!TARGET_64BIT_P ... && TARGET_SSE_P ...) to match the
>> >>> condition in the special fpmath processing part. (BTW: The comment in
>> >>> that part is wrong, we need SSE, not sse2 on 32-bit targets to use SSE
>> >>> math.)
>> >>
>> >> Hmpf, with the change I again run into the libgo bootstrap fail
>> >> (appearantly my reduced testcase is incomplete)
>> >>
>> >> libtool: compile:  /abuild/rguenther/obj2/./gcc/xgcc
>> >> -B/abuild/rguenther/obj2/./gcc/ -B/usr/local/x86_64-pc-linux-gnu/bin/
>> >> -B/usr/local/x86_64-pc-linux-gnu/lib/ -isystem
>> >> /usr/local/x86_64-pc-linux-gnu/include -isystem
>> >> /usr/local/x86_64-pc-linux-gnu/sys-include -m32 -DHAVE_CONFIG_H -I.
>> >> -I/space/rguenther/src/svn/trunk2/libgo -I
>> >> /space/rguenther/src/svn/trunk2/libgo/runtime
>> >> -I/space/rguenther/src/svn/trunk2/libgo/../libffi/include
>> >> -I../libffi/include -pthread -fexceptions -fnon-call-exceptions
>> >> -fplan9-extensions -fsplit-stack -Wall -Wextra -Wwrite-strings -Wcast-qual
>> >> -Werror -minline-all-stringops -D_GNU_SOURCE -D_LARGEFILE_SOURCE
>> >> -D_FILE_OFFSET_BITS=64 -I /space/rguenther/src/svn/trunk2/libgo/../libgcc
>> >> -I /space/rguenther/src/svn/trunk2/libgo/../libbacktrace -I
>> >> ../../../gcc/include -g -O2 -m32 -MT aeshash.lo -MD -MP -MF
>> >> .deps/aeshash.Tpo -c
>> >> /space/rguenther/src/svn/trunk2/libgo/runtime/aeshash.c  -fPIC -DPIC -o
>> >> .libs/aeshash.o
>> >> In file included from
>> >> /space/rguenther/src/svn/trunk2/libgo/runtime/aeshash.c:17:0:
>> >> /space/rguenther/src/svn/trunk2/libgo/runtime/aeshash.c: In function
>> >> ‘aeshashbody’:
>> >> /abuild/rguenther/obj2/./gcc/include/emmintrin.h:700:1: error: inlining
>> >> failed in call to always_inline ‘_mm_loadu_si128’: target specific option
>> >> mismatch
>> >>  _mm_loadu_si128 (__m128i_u const *__P)
>> >>  ^~~~~~~~~~~~~~~
>> >> /space/rguenther/src/svn/trunk2/libgo/runtime/aeshash.c:387:11: note:
>> >> called from here
>> >>   mseed ^= _mm_loadu_si128(aeskeysched.__values);
>> >>            ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>> >>
>> >> that appens even with making the if a if (1).  The check we run into
>> >> is somehow still
>> >>
>> >>   else if (caller_opts->x_ix86_fpmath != callee_opts->x_ix86_fpmath)
>> >>     ret = false;
>> >>
>> >> ah ...
>> >>
>> >> #ifndef __SSE__
>> >> #pragma GCC push_options
>> >> #pragma GCC target("sse")
>> >> #define __DISABLE_SSE__
>> >> #endif /* __SSE__ */
>> >>
>> >> but on x86_64 when building with -m32 we _do_ have __SSE__ set so
>> >> the intrinsics get target_option_default_node which doesn't have
>> >> fmpath=sse.  So x86_64 -m32 -march=i586 enables SSE math for
>> >> the intrinsics while x86_64 -m32 [-march=x86-64] does not...
>> >
>> > IMO target_default_node should already enable fpmath=sse. When SSE is
>> > available, fpmath should be set to SSE, unless there is fpmath option
>> > specified.
>>
>> Following part from ix86_option_override_internal looks a bit mismatched:
>>
>> --cut here--
>>   /* For all chips supporting SSE2, -mfpmath=sse performs better than
>>      fpmath=387.  The second is however default at many targets since the
>>      extra 80bit precision of temporaries is considered to be part of ABI.
>>      Overwrite the default at least for -ffast-math.
>>      TODO: -mfpmath=both seems to produce same performing code with bit
>>      smaller binaries.  It is however not clear if register allocation is
>>      ready for this setting.
>>      Also -mfpmath=387 is overall a lot more compact (bout 4-5%) than SSE
>>      codegen.  We may switch to 387 with -ffast-math for size optimized
>>      functions. */
>>   else if (fast_math_flags_set_p (&global_options)
>>        && TARGET_SSE2_P (opts->x_ix86_isa_flags))
>>     opts->x_ix86_fpmath = FPMATH_SSE;
>> --cut here--
>>
>> We probably have to set fpmath to SSE also for (!TARGET_64BIT_P ... &&
>> TARGET_SSE_P ...).
>
> The comment says we can't because 387 math is considered part of the ABI.

Then ix86_valid_target_attribute_tree needs to set fpmath in the same
way to avoid fpmath mismatch, probably using TARGET_FPMATH_DEFAULT.

Uros.