arm/msac: Trim C functions from release build

changed milestone to %1.4.1

added ARM performance labels

requested review from @mstorsjo

This fails to build on arm64 with -Denable_asm=false -Dtrim_dsp=true but has a green pipeline.

Hmm, I guess we're lacking CI coverage for some combinations of architectures and -Dtrim_dsp=true, that would probably be good to fill.

marked this merge request as draft

added 1 commit

• ba300a81 - arm64/msac: Trim C functions, saves 1280 bytes

Compare with previous version

added 1 commit

• 6a98be72 - arm32/msac: Trim C functions, saves 1028 bytes

Compare with previous version

changed title from Draft: arm64/msac: Trim C functions, saves 1280 bytes to Draft: arm/msac: Trim C functions from release build

Can you clarify how this relates to !1614 (merged)?

The functions that are enclosed within an ifdef here, they are all non-static, so I guess that means that the compiler can't omit them. But if we can ifdef them out altogether, that means nobody is using them in extern form, so can't we just make them static, and they should be trimmed out like all other ones?

Unfortunately we cannot just make them static. The msac DSP functions are the only ones that are specialized using #define's instead of function pointers, except for x86 where there is a function pointer added to MsacContext to pick between SSE2 and AVX2 at run time. See lines 51-59 and 71-92 of src/msac.h

This relates to !1614 (merged) in that the specialization condition #if defined(__ARM_NEON) || defined(__APPLE__) || defined(_WIN32) || ARCH_AARCH64 needs to match. Otherwise we could trim the C functions and cause a linking error.

Thanks, that makes sense.

@gramner, I presume we could/should do something similar for x86 as well?

Maybe this could be generalized to something non-arch-specific instead of hardcoding every arch combination?

Possibly something along the lines of

#define TRIM_SYMBOL_ADAPT (defined(dav1d_msac_decode_symbol_adapt4) && \
                           defined(dav1d_msac_decode_symbol_adapt8 && \
                           defined(dav1d_msac_decode_symbol_adapt16))
#define TRIM_BOOL_ADAPT defined(dav1d_msac_decode_bool_adapt)

etc. at https://code.videolan.org/videolan/dav1d/-/blob/master/src/msac.h?ref_type=heads#L70

and then ifdef each C function under

#if !(TRIM_DSP && TRIM_X)

I will address this in another MR. I would like to move all of the architectures to the same place and then we can simplify the #define's.

added 22 commits

• 6a98be72...b5b394cd - 20 commits from branch videolan:master
• b9f53330 - arm64/msac: Trim C functions, saves 1392 bytes
• 84b9c6bf - arm32/msac: Trim C functions, saves 1028 bytes

Compare with previous version

marked this merge request as ready

added 1 commit

• 0fff614a - arm32/msac: Trim C functions, saves 1024 bytes

Compare with previous version

approved this merge request

merged