On 2022-10-13 03:43, Hongtao Liu wrote:
> On Thu, Oct 13, 2022 at 5:45 AM J.W. Jagersma via Gcc-help
> <gcc-help@xxxxxxxxxxx> wrote:
>>
>> Consider the following (incomplete) code:
>>
>> enum simd { mmx = 1, sse = 2 /* , ... */ };
>>
>> template<simd flags>
>> void simd_func()
>> {
>> if constexpr (flags & simd::mmx)
>> {
>> // MMX code here
>> }
>> else if constexpr (flags & simd::sse)
>> {
>> // SSE code here
>> }
>> // etc ...
>> }
>>
>> The idea is to instantiate function templates with optimized SIMD routines
>> multiple times, then I can compile all my code with -march=i386, and select the
>> best implementation at runtime.
> It looks like you want to use function multi-versioning.
>>
>> However, after already spending a lot of time restructuring my code around this
>> idea, I discover that gcc refuses to compile this without having the
>> corresponding target options enabled:
>>
>> error: inlining failed in call to 'always_inline' '__m64 _mm_unpacklo_pi8(__m64, __m64)': target specific option mismatch
>>
>> That makes no sense to me. I want gcc to emit those SIMD instructions verbatim
>> in my code, regardless of what compiler options the user specified. What then
>> is the point of feature-test macros (__SSE__, etc)? I don't get it.
>>
>> Is there any option or attribute I can use to make gcc ignore these target
>> option mismatches when inlining? Or any other way to make this work?
> maybe you can try
> #pragma GCC push_options
> #pragma GCC target("sse2")
> your codes..
> #pragma GCC pop_options
>
> https://gcc.gnu.org/onlinedocs/gcc/Function-Specific-Option-Pragmas.html#Function-Specific-Option-Pragmas
Hi, and thanks,
While that would work, it would also give the compiler free reign to auto-
vectorize any code (__builtin_memcpy, etc) using those target options, right?
My goal though is to only generate SIMD code for those targets specified in the
template parameter.
Another idea I had:
template<simd flags>
consteval const char* simd_target_options()
{
// compile-time string wizardry goes here ...
return "mmx,sse,sse2"; // just to test
}
template<simd flags>
[[gnu::target(simd_target_options<flags>())]]
void simd_func() { /* ... */ }
Which also doesn't work:
error: attribute 'target' argument is not a string
To be expected I suppose, although I know some attributes do take arbitrary
expressions. What I'm really looking for is an option to automatically
propagate target options from inlined functions, but that just doesn't seem to
exist.
