-Wframe-larger-than instances in drivers/gpu/drm/amd/display/dc/dml with ARCH=x86-64

[ms178]

I see similar warnings as posted in #1752 with my x86-64-build of a 6.0.6 Kernel with some patches (more details and config here):

LD [M]  drivers/gpu/drm/amd/amdgpu/amdgpu.o
ld.lld: warning: <unknown>:0:0: stack frame size (2808) exceeds limit (2048) in function 'dml31_ModeSupportAndSystemConfigurationFull'
1638/2808 (58.33%) spills, 1170/2808 (41.67%) variables
ld.lld: warning: <unknown>:0:0: stack frame size (2808) exceeds limit (2048) in function 'dml314_ModeSupportAndSystemConfigurationFull'
1638/2808 (58.33%) spills, 1170/2808 (41.67%) variables

That was seen with FullLTO, my custom LLVM-Toolchain from my repo and clang version 16.0.0 (c97035c49d941e5b196a938f0393f811d1adbd57).

[nathanchance]

I would be curious to know if the same configuration without LTO shows the same warnings. Based on the variables to spill ratio, it is possible that LTO does more inlining and blows out the stack.

[ms178]

Without LTO, I've just noticed these two:

drivers/gpu/drm/amd/amdgpu/../display/dc/dml/dcn31/display_mode_vba_31.c:3767:6: warning: stack frame size (2680) exceeds limit (2048) in 'dml31_ModeSupportAndSystemConfigurationFull' [-Wframe-larger-than]
void dml31_ModeSupportAndSystemConfigurationFull(struct display_mode_lib *mode_lib)
     ^
1387/2680 (51.75%) spills, 1293/2680 (48.25%) variables
 
drivers/gpu/drm/amd/amdgpu/../display/dc/dml/dcn314/display_mode_vba_314.c:3890:6: warning: stack frame size (2584) exceeds limit (2048) in 'dml314_ModeSupportAndSystemConfigurationFull' [-Wframe-larger-than]
void dml314_ModeSupportAndSystemConfigurationFull(struct display_mode_lib *mode_lib)
     ^
1355/2584 (52.44%) spills, 1229/2584 (47.56%) variables

[ms178]

This is what I get with Kernel 6.1 and Clang-16 (23e3e107dceb1fd8c1fb81d1229f6aab09a1fce2):

 LD [M]  drivers/gpu/drm/amd/amdgpu/amdgpu.o
ld.lld: warning: <unknown>:0:0: stack frame size (3032) exceeds limit (2048) in function 'dml31_ModeSupportAndSystemConfigurationFull'
2055/3032 (67.78%) spills, 977/3032 (32.22%) variables
ld.lld: warning: <unknown>:0:0: stack frame size (3000) exceeds limit (2048) in function 'dml314_ModeSupportAndSystemConfigurationFull'
2051/3000 (68.37%) spills, 949/3000 (31.63%) variables
ld.lld: warning: <unknown>:0:0: stack frame size (2072) exceeds limit (2048) in function 'dml32_CalculateVMRowAndSwath'
1554/2072 (75.00%) spills, 518/2072 (25.00%) variables

[dileks]

With Linux 6.2-rc4 and LLVM 15.0.3 I still see:

ld.lld: warning: stack frame size (2088) exceeds limit (2048) in function 'dml314_ModeSupportAndSystemConfigurationFull'
ld.lld: warning: stack frame size (2088) exceeds limit (2048) in function 'dml31_ModeSupportAndSystemConfigurationFull'

kernel-config: ThinLTO + KCFI

[nathanchance]

This series might address those warnings:

https://lore.kernel.org/20230110154325.3350916-1-Rodrigo.Siqueira@amd.com/

[dileks]

@nathanchance

Thanks for the link.

The patch-series applies cleanly on top of Linux 6.2-rc4.

[dileks]

@nathanchance @rodrigosiqueira

A big thank-you pointing me to the link and Rodrigo Siqueira Rodrigo.Siqueira@amd.com for the patch-set.

I did a new build including this patch-set and this fixes the warnings for me.

Here and formally (patch-series on top of Linux v6.2-rc4):

Tested-by: Sedat Dilek sedat.dilek@gmail.com # LLVM-15.0.3 x86-64

[ms178]

I can confirm that the warnings are gone with that patchset and clang version 16.0.0 (60b98979241151f8d6ec8af9ebc4da421f4a0258), using FullLTO and x86-64 on my modified Kernel 6.1.7.

Tested-by: Marcus Seyfarth m.seyfarth@gmail.com

[dileks]

Unfortunately - patchset applied on top of Linux v6.2-rc5+ - with LLVM 16.0.0-rc1 (stage1-only) I see again:

27616:ld.lld: warning: drivers/gpu/drm/amd/amdgpu/../display/dc/dml/dcn314/display_mode_vba_314.c:3890:0: stack frame size (2056) exceeds limit (2048) in function 'dml314_ModeSupportAndSystemConfigurationFull'
27622:ld.lld: warning: drivers/gpu/drm/amd/amdgpu/../display/dc/dml/dcn31/display_mode_vba_31.c:3791:0: stack frame size (2120) exceeds limit (2048) in function 'dml31_ModeSupportAndSystemConfigurationFull'

config-6.2.0-rc5-2-amd64-clang16-kcfi.txt

[nathanchance]

I suspect this series that was just sent out will resolve this issue.

[nickdesaulniers]

Also, clang-18 has some improvements around stack usage.
https://reviews.llvm.org/rGe698695fbbf62e6676f8907665187f2d2c4d814b

Since it's not clear to me from this thread what fixed this where (and I can no longer reproduce w/ clang-18), closing.

Please reopen with more info if you can reproduce with clang-18+ (containing https://reviews.llvm.org/rGe698695fbbf62e6676f8907665187f2d2c4d814b)

[ms178]

@nickdesaulniers I've been carrying the following series with my Kernel that addresses this issue to some degree: https://lists.freedesktop.org/archives/amd-gfx/2023-January/088203.html

Will there be a backport of your patch for LLVM/Clang-17 or a point release? As we are still in the RC phase, it might be still possible to land this for the Clang-17 release?

[nickdesaulniers]

https://reviews.llvm.org/rGe698695fbbf62e6676f8907665187f2d2c4d814b has already been reverted once.

I relanded it without any hint to whatever was the matter with it the first time (possibly nothing; it may have been exposing an older issue that's since been found+fixed). So I'm still holding my breath that it will stick and not get reverted a second time. As such, I consider it too high risk to backport; I'd like it to bake longer to ensure it sticks.