Fixes for loongarch compilers

[jinboson]

Happy New Year, Michael @mr-c.

We rewrited the PR #1247 (can be closed now) and seperated it into two patches. We did two things on them.

1）The first thing is to fix a typo for lasx implemention in x86 avx2 that modified simde_mm256_bslli_epi128 to simde_mm256_bsrli_epi128, very sorry for this typo error.
2）The second thing is to work around compiler errors in use-cases that users would pass variables for shift count for slli/srli/srai instrutions, which can not support for loongarch at present.

We hope you could take some time to review this for us, if you have any questions please feel free to let us know, thank you Michael， thank you.

[mr-c]

Why are you changing this? For simde_mm_slli_epi16 we require imm8 to be a constant value.

See 1715c16#diff-d5773576112e65731ce8f01f5beee4f665e6fcffd1017ee76c583fa0988a40b4R6554-R6555

Same for simde_mm_slli_epi32, simde_mm_slli_epi64, simde_mm_srli_epi16, simde_mm_srli_epi32 and simde_mm_srli_epi64

[jinboson]

Why are you changing this?

That's because we found that end-users would use _mm_slli_epi16 like below :

int shift;
computed the shift in some time...
then use the shfit to do : __mm_slli_epi16(src_vec, shift)

At compile time the shift was unknow, in this case loongarch compilers would emit errors in the build process. Actually, we really found the strange use-cases in google's AV1 Codec Library aom project. So in order to avoid build failure on loongarch we made these modifications. What do you think for these use-cases, Michael ?

[mr-c]

Wacky. Sure, it doesn't bother me either way

[jinboson]

#include <stdio.h>
#include "/data/home/Jin/data/community/simde-loong64/simde/x86/sse2.h"

int main()
{
        simde__m128i vec0;
        const int imm8 = 1;     //emulate use-cases that end-users pass variable for shift count.
        simde_mm_slli_epi16(vec0, imm8);
	return 0;
}

I built above demo on real loongarch hardware with simde git master, and the compiler output errors:

Same for simde_mm_slli_epi32, simde_mm_slli_epi64, simde_mm_srli_epi16, simde_mm_srli_epi32 and simde_mm_srli_epi64

[jinboson]

#include <stdio.h>
#include "/data/home/Jin/data/community/simde-loong64/simde/x86/sse2.h"

int main()
{
        simde__m128i vec0;
        //const int imm8 = 1;     //emulate use-cases that end-users pass variable for shift count.
        simde_mm_slli_epi16(vec0, 1);
	return 0;
}

If user pass immediate values for shift count just like above demo, it's ok for loongarch compiler.

[jinboson]

So, at present, we need to load imm8 into vectors and then use sll version instruction instead. If in the feature, the loongarch compiler can accept a variable for shift count for slli instruction then we can solve this dilemma.

[mr-c]

if the LASX_NATIVE versions don't require the use of immediate values, then please move those optimized implementaions into the function definitions instead of using macros.

[mr-c]

@jinboson if SIMDe is important to the LOONGARCH ecosystem, it would be helpful if I had some assistance with #1177 and similar ; I'm the only maintainer these days..

[jinboson]

@jinboson if SIMDe is important to the LOONGARCH ecosystem, it would be helpful if I had some assistance with #1177 and similar ; I'm the only maintainer these days..

Understood, thank you @mr-c , I have posted the test result based on the real loongarch hardware and made explanations for you. Thanks for you reply!

[jinboson]

if the LASX_NATIVE versions don't require the use of immediate values, then please move those optimized implementaions into the function definitions instead of using macros.

It depends on how users use them. We do have slli version instructions that require the use of immediate values, but users must pass an immediate value rather than a variable value for shift count when using them. Now the problem we are taking about is that those slli version instructions that require the use of immediate values are passed variable values for shift count in some strange use-cases, so we need workround for those implementations using macros.

[jinboson]

move those optimized implementaions into the function definitions instead of using macros.

Done by the latest update, thank you Michael.