[Bug ipa/110334] [13/14 Regresssion] unused functions not eliminated before LTO streaming

["rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110334

--- Comment #20 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #19)
> It seems that the C++ FE change in comment#13 causes libreoffice to fail to
> build with
> 
> [  553s]
> /home/abuild/rpmbuild/BUILD/libreoffice-7.5.4.2/workdir/UnpackedTarball/skia/
> include/private/SkVx.h: In function 'interpret_skvm':
> [  553s]
> /home/abuild/rpmbuild/BUILD/libreoffice-7.5.4.2/workdir/UnpackedTarball/skia/
> include/private/SkVx.h:150: error: inlining failed in call to
> 'always_inline' '__ct_comp ': target specific option mismatch
> [  553s]   150 |     using VecStorage<N,T>::VecStorage;
> [  553s]       | 
> 
> I have yet to reproduce and extract a testcase though.

libreoffice has a skia import and links its Library/libskialo.so library
with LTO, combining the AVX, HSW, SSE41, SSE42, SSSE3, CRC32, SKX, ...
SkOpts_*.o objects which likely follow a very similar scheme as the firefox
copy.  SkOpts_hsw.cpp looks like

#include "src/core/SkOpts.h"

#define SK_OPTS_NS hsw
#include "src/core/SkCubicSolver.h"
#include "src/opts/SkBitmapProcState_opts.h"
#include "src/opts/SkBlitRow_opts.h"
#include "src/opts/SkRasterPipeline_opts.h"
#include "src/opts/SkSwizzler_opts.h"
#include "src/opts/SkUtils_opts.h"
#include "src/opts/SkVM_opts.h"

namespace SkOpts {
    void Init_hsw() {
        blit_row_color32     = hsw::blit_row_color32;
        blit_row_s32a_opaque = hsw::blit_row_s32a_opaque;

        S32_alpha_D32_filter_DX  = hsw::S32_alpha_D32_filter_DX;

        cubic_solver = SK_OPTS_NS::cubic_solver;

        RGBA_to_BGRA          = SK_OPTS_NS::RGBA_to_BGRA;
        RGBA_to_rgbA          = SK_OPTS_NS::RGBA_to_rgbA;
        RGBA_to_bgrA          = SK_OPTS_NS::RGBA_to_bgrA;
        gray_to_RGB1          = SK_OPTS_NS::gray_to_RGB1;
        grayA_to_RGBA         = SK_OPTS_NS::grayA_to_RGBA;
        grayA_to_rgbA         = SK_OPTS_NS::grayA_to_rgbA;
        inverted_CMYK_to_RGB1 = SK_OPTS_NS::inverted_CMYK_to_RGB1;
        inverted_CMYK_to_BGR1 = SK_OPTS_NS::inverted_CMYK_to_BGR1;

    #define M(st) stages_highp[SkRasterPipeline::st] = (StageFn)SK_OPTS_NS::st;
        SK_RASTER_PIPELINE_STAGES(M)
        just_return_highp = (StageFn)SK_OPTS_NS::just_return;
        start_pipeline_highp = SK_OPTS_NS::start_pipeline;
    #undef M

    #define M(st) stages_lowp[SkRasterPipeline::st] =
(StageFn)SK_OPTS_NS::lowp::st;
        SK_RASTER_PIPELINE_STAGES(M)
        just_return_lowp = (StageFn)SK_OPTS_NS::lowp::just_return;
        start_pipeline_lowp = SK_OPTS_NS::lowp::start_pipeline;
    #undef M

        interpret_skvm = SK_OPTS_NS::interpret_skvm;
    }
}  // namespace SkOpts

the question in the end is why we fail to process the always_inlines
early here and only discover them late.  Unfortunately the inline
diagnostic doesn't print the original TU name (that would be useful
for all late LTO diagnostics).

I'm not sure whether this is an issue with the C++ frontend change or with
the ODR issues present in Skia.  Eventually since we're dealing with
aliases(?), always_inline isn't appropriately detected or still missing
on some callees.

Unfortunately there isn't a knob to diagnose late inlined always-inline
functions.