On 3/6/18 11:45 AM, Sitsofe Wheeler wrote:
>> BTW, another (and probably more problematic) issue is that the
>> compiler may well have support for sse/avx/avx512, while the
>> machine it's run on does not. If we enable the instruction set
>> based on that, we could be seeing illegal instructions outside
>> just the memcpy test. And that would be a concern, since it would
>> prevent fio from running at all.
>>
>> I think we have to make it something like the below because of that.
>
> If someone forces SSE on then by all means we should add it to the
> CFLAGS but if they haven't and we can compile a program which uses
> headers/instructions that are only available when SSE is enabled then
> can't we deduce that the user is happy for those instructions to be in
> their code?
Right, _if_ they turn it on. The initial patch did it unconditionally,
this followup patch adds the option to turn it on, while defaulting to
not enabling sse. We can't add -msse if gcc decides to enable SSE
memcpy, for instance, which will then fail with SIGILL if the machine
doesn't support it.
> [...]
>> +##########################################
>> +# sse probe
>> +sse="no"
>> +cat > $TMPC << EOF
>> +#include <xmmintrin.h>
>> +#include <immintrin.h>
>> +int main(int argc, char **argv)
>> +{
>> + __m128 val;
>> + float const *src = NULL;
>> + float *dst = NULL;
>> + val = _mm_load_ps(src);
>> + _mm_store_ps(dst, val);
>> + return 0;
>> +}
>
> We could try and use __builtin_ia32_addps (from
> https://gcc.gnu.org/onlinedocs/gcc/x86-Built-in-Functions.html#x86-Built-in-Functions
Sure
>> +EOF
>> +if test "$enable_sse" = "yes" && compile_prog "-msse" "" "sse"; then
>> + sse="yes"
>> +fi
>> +print_config "SSE (compiler)" "$sse"
>
> Here we wouldn't explicitly add -msse because we would have already
> added it to the global CFLAGS if forced or depended on the user's
> particular compiler flags if not.
It's only enabled IFF the global flag (--enable-sse) is set AND the
configure test passes.
--
Jens Axboe
--
To unsubscribe from this list: send the line "unsubscribe fio" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
