Re: [PATCH, i386] Fix emitting of prefetch instructions

[Uros Bizjak <ubizjak at gmail dot com>]

On Tue, Mar 4, 2014 at 12:31 AM, Uros Bizjak <ubizjak@gmail.com> wrote:

>>> The new gcc.target/i386/prefetchwt1-1.c test currently FAILs on Solaris 9/x86:
>>>
>>> FAIL: gcc.target/i386/prefetchwt1-1.c (test for excess errors)
>>> Excess errors:
>>> /var/gcc/regression/trunk/9-gcc-gas/build/gcc/include/xmmintrin.h:1195:1: error:
>>>  inlining failed in call to always_inline '_mm_prefetch': target specific option
>>>  mismatch
>>> /vol/gcc/src/hg/trunk/local/gcc/testsuite/gcc.target/i386/prefetchwt1-1.c:12:5:
>>> error: called from here
>>>
>>> gcc.target/i386/prefetchwt1-1.c: output file does not exist
>>> UNRESOLVED: gcc.target/i386/prefetchwt1-1.c scan-assembler [ \\t]+prefetchwt1[ \
>>> \t]+
>>>
>>> This can be fixed by compiling with -msse2.
>>
>> Actually, we should take prefetch instructions out of various GCC
>> target pragmas. Patterns that emit these instructions are designed to
>> (depending on selected ISA) always emit  the most optimal prefetch
>> instruction.
>>
>> The patch also changes the compiler to emit prefetchwt1 only for
>> _MM_HINT_T1, while for _MM_HINT_T0, it still emits prefetchw. In
>> addition, the patch corrects wrong MM_HINT_T0 value.
>>
>> Patch was bootstrapped and tested on x86_64-pc-linux-gnu {,-m32}  and
>> committed to mainline SVN.
>>
>> 2014-03-03  Uros Bizjak  <ubizjak@gmail.com>
>>
>>     * config/i386/xmmintrin.h (enum _mm_hint) <_MM_HINT_ET0>: Correct
>>     hint value.
>>     (_mm_prefetch): Move out of GCC target("sse") pragma.
>>     * config/i386/prfchwintrin.h (_m_prefetchw): Move out of
>>     GCC target("prfchw") pragma.
>>     * config/i386/i386.md (prefetch): Emit prefetchwt1 only
>>     for locality <= 2.
>>     * config/i386/i386.c (ix86_option_override_internal): Enable
>>     -mprfchw with -mprefetchwt1.
>
> BTW: There are a couple of new testsuite failures:
>
> FAIL: gcc.target/i386/avx512pf-vscatterpf0dpd-1.c (test for excess errors)
> UNRESOLVED: gcc.target/i386/avx512pf-vscatterpf0dpd-1.c
> scan-assembler-times vscatterpf0dpd[ \\\\t]+[^\\n]*%ymm[0-9] 2
> UNRESOLVED: gcc.target/i386/avx512pf-vscatterpf0dpd-1.c
> scan-assembler-times vscatterpf0dpd[ \\\\t]+[^\\n]*{%k[1-7] 1
> FAIL: gcc.target/i386/avx512pf-vscatterpf0dps-1.c (test for excess errors)
> UNRESOLVED: gcc.target/i386/avx512pf-vscatterpf0dps-1.c
> scan-assembler-times vscatterpf0dps[ \\\\t]+[^\\n]*%zmm[0-9] 2
> UNRESOLVED: gcc.target/i386/avx512pf-vscatterpf0dps-1.c
> scan-assembler-times vscatterpf0dps[ \\\\t]+[^\\n]*{%k[1-7] 1
> FAIL: gcc.target/i386/avx512pf-vscatterpf0qpd-1.c (test for excess errors)
> UNRESOLVED: gcc.target/i386/avx512pf-vscatterpf0qpd-1.c
> scan-assembler-times vscatterpf0qpd[ \\\\t]+[^\\n]*%zmm[0-9] 2
> UNRESOLVED: gcc.target/i386/avx512pf-vscatterpf0qpd-1.c
> scan-assembler-times vscatterpf0qpd[ \\\\t]+[^\\n]*{%k[1-7] 1
> FAIL: gcc.target/i386/avx512pf-vscatterpf0qps-1.c (test for excess errors)
> UNRESOLVED: gcc.target/i386/avx512pf-vscatterpf0qps-1.c
> scan-assembler-times vscatterpf0qps[ \\\\t]+[^\\n]*%zmm[0-9] 2
> UNRESOLVED: gcc.target/i386/avx512pf-vscatterpf0qps-1.c
> scan-assembler-times vscatterpf0qps[ \\\\t]+[^\\n]*{%k[1-7] 1
>
> They are all:
>
> FAIL: gcc.target/i386/avx512pf-vscatterpf0dpd-1.c (test for excess errors)
> Excess errors:
> /ssd/uros/gcc-build/gcc/include/avx512pfintrin.h:108:3: error: the
> last argument must be hint 0 or 1
>
> They are due to _MM_HINT_ET0 fix, and probably show that the pattern
> was not updated when hint constants were adjusted to 2 and 3.
>
> Kirill, can you please look at this inconsistency?

Attached patch fixes these failures, and also fixes and improves error message.

Uros.

Attachment: p.diff.txt
Description: Text document