[Mesa-dev] [PATCH] meson: enable -fmerge-all-constants by default

[Konstantin Kharlamov]

FWIW I just realized it also enables -fmerge-constants which is enabled 
at all optimize levels, and built/compared it with fmerge-all-constants. 
The difference is still noticeable. If anybody interested, I can 
pretty-format it, and send here.

On 25.07.2018 14:03, Konstantin Kharlamov wrote:
> This is a standard non-conforming optimization, which, quoting,
> "…considers e.g. even constant initialized arrays or initialized
> constant variables with integral or floating-point types. Languages like
> C or C++ require each variable, including multiple instances of the same
> variable in recursive calls, to have distinct locations, so using this
> option results in non-conforming behavior."
> 
> Having the same or distinct addresses of constants in recursive calls is
> too obscure behavior to make any use of. More over, until 4 April 2018
> clang have had this option enabled by default¹, and nobody ever noticed
> until kernel build fail this year (having a low-level hackery in kernel
> surprises no one), resulting in PR disabling the option by default.
> 
> As for Mesa: I personally have been using the option for release builds
> since forever, never had any problems.
> 
> Real world significance, sizes of binaries in bytes:
> 
> default                                      | -fmerge-all-constants                        | diff
> -----------------------------------------------------------------------------------------------------
>   8048024 <-> bellagio/libomx_mesa.so         |  7867800 <-> bellagio/libomx_mesa.so         | -180224
> 10209120 <-> d3d/d3dadapter9.so.1.0.0        | 10016608 <-> d3d/d3dadapter9.so.1.0.0        | -192512
> 14464936 <-> dri/i915_dri.so                 | 14383016 <-> dri/i915_dri.so                 | -81920
> 16894344 <-> dri/kms_swrast_dri.so           | 16644488 <-> dri/kms_swrast_dri.so           | -249856
>   8077560 <-> dri/nouveau_drv_video.so        |  7893240 <-> dri/nouveau_drv_video.so        | -184320
>   3882872 <-> gallium-pipe/pipe_nouveau.so    |  3870584 <-> gallium-pipe/pipe_nouveau.so    | -12288
>   2511160 <-> gallium-pipe/pipe_r300.so       |  2502968 <-> gallium-pipe/pipe_r300.so       | -8192
>   2818520 <-> gallium-pipe/pipe_r600.so       |  2810328 <-> gallium-pipe/pipe_r600.so       | -8192
>   4123056 <-> gallium-pipe/pipe_radeonsi.so   |  3955120 <-> gallium-pipe/pipe_radeonsi.so   | -167936
>   2413936 <-> gallium-pipe/pipe_swrast.so     |  2401648 <-> gallium-pipe/pipe_swrast.so     | -12288
>   2701360 <-> gallium-pipe/pipe_vmwgfx.so     |  2689072 <-> gallium-pipe/pipe_vmwgfx.so     | -12288
>    582496 <-> libGLX_mesa.so.0.0.0            |   578400 <-> libGLX_mesa.so.0.0.0            | -4096
> 23451656 <-> libMesaOpenCL.so.1.0.0          | 23447560 <-> libMesaOpenCL.so.1.0.0          | -4096
> 10370344 <-> libOSMesa.so.8.0.0              | 10300712 <-> libOSMesa.so.8.0.0              | -69632
>   4766032 <-> libvulkan_intel.so              |  4745552 <-> libvulkan_intel.so              | -20480
>   2761376 <-> libvulkan_radeon.so             |  2753184 <-> libvulkan_radeon.so             | -8192
>   4703776 <-> libxatracker.so.2.4.0           |  4687392 <-> libxatracker.so.2.4.0           | -16384
>   8510392 <-> vdpau/libvdpau_nouveau.so.1.0.0 |  8326072 <-> vdpau/libvdpau_nouveau.so.1.0.0 | -184320
> 
> For consistency it's enabled for both debug and release builds.
> 
> 1: https://reviews.llvm.org/D45289
> 
> Signed-off-by: Konstantin Kharlamov <Hi-Angel at yandex.ru>
> ---
>   meson.build | 5 +++++
>   1 file changed, 5 insertions(+)
> 
> diff --git a/meson.build b/meson.build
> index e05645cbf3..36016cf959 100644
> --- a/meson.build
> +++ b/meson.build
> @@ -794,6 +794,11 @@ foreach a : ['-Wall', '-fno-math-errno', '-fno-trapping-math',
>     endif
>   endforeach
>   
> +if ['gcc', 'clang'].contains(cc.get_id())
> +  c_args += '-fmerge-all-constants'
> +  cpp_args += '-fmerge-all-constants'
> +endif
> +
>   # For some reason, the test for -Wno-foo always succeeds with gcc, even if the
>   # option is not supported. Hence, check for -Wfoo instead.
>   
>