RE: [PATCH] drm/amdgpu: drop the long-double-128 powerpc check/hack
From: "Deucher, Alexander" <Alexander.Deucher@amd.com>
Date: 2023-02-01 19:27:17
Also in:
amd-gfx, stable
[Public]
-----Original Message-----
From: amd-gfx <redacted> On Behalf Of
Daniel Kolesa
Sent: Wednesday, February 1, 2023 1:46 PM
To: linuxppc-dev@lists.ozlabs.org
Cc: dan@danny.cz; stable@vger.kernel.org;
tpearson@raptorengineering.com; amd-gfx@lists.freedesktop.org;
alexdeucher@gmail.com; torvalds@linux-foundation.org
Subject: [PATCH] drm/amdgpu: drop the long-double-128 powerpc
check/hack
Commit c653c591789b ("drm/amdgpu: Re-enable DCN for 64-bit powerpc")
introduced this check as a workaround for the driver not building with
toolchains that default to 64-bit long double.
The reason things worked on 128-bit-long-double toolchains and not
otherwise was however largely accidental. The real issue was that some files
containing floating point code were compiled without -mhard-float, while
others were compiled with -mhard-float.
The PowerPC compilers tag object files that use long doubles with a special
ABI tag in order to differentiate 64-bit long double, IBM long double and IEEE
128-bit long double. When no long double is used in a source file, the file
does not receive the ABI tag.
Since only regular doubles are used in the AMDGPU source, there is no ABI
tag on the soft-float object files with 128-bit-ldbl compilers, and therefore no
error. With 64-bit long double, the double and long double types are equal,
and an ABI tag is introduced.
Of course, this resulted in the real bug, which was mixing of hard and soft
float object files, getting hidden, which makes this check technically
incorrect.
Since then, work has been done to ensure that all float code is separately
compiled. This was also necessary in order to enable
AArch64 support in the display stack, as AArch64 does not have any soft-float
ABI, and all code that does not explicitly conatain floats is compiled with -
mgeneral-regs-only, which prevents float-using code from being compiled at
all. That means AArch64 support will from now on always safeguard against
such cases happening ever again.
In mainline, this work is now fully done, so this check is fully redundant and
does not do anything except preventing AMDGPU DC from being built on
systems such as those using musl libc. The last piece of work to enable this
was commit c92b7fe0d92a
("drm/amd/display: move remaining FPU code to dml folder") and this has
since been backported to 6.1 stable (in 6.1.7).
Relevant issue: https://gitlab.freedesktop.org/drm/amd/-/issues/2288
Signed-off-by: Daniel Kolesa <redacted>Acked-by: Alex Deucher <alexander.deucher@amd.com>
quoted hunk ↗ jump to hunk
--- arch/powerpc/Kconfig | 4 ---- drivers/gpu/drm/amd/display/Kconfig | 2 +- 2 files changed, 1 insertion(+), 5 deletions(-)diff --git a/arch/powerpc/Kconfig b/arch/powerpc/Kconfig indexb8c4ac56b..267805072 100644--- a/arch/powerpc/Kconfig +++ b/arch/powerpc/Kconfig@@ -289,10 +289,6 @@ config PPC # Please keep this list sorted alphabetically. # -config PPC_LONG_DOUBLE_128 - depends on PPC64 && ALTIVEC - def_bool $(success,test "$(shell,echo __LONG_DOUBLE_128__ |$(CC) -E -P -)" = 1) - config PPC_BARRIER_NOSPEC bool default ydiff --git a/drivers/gpu/drm/amd/display/Kconfigb/drivers/gpu/drm/amd/display/Kconfig index 2efe93f74..94645b6ef 100644--- a/drivers/gpu/drm/amd/display/Kconfig +++ b/drivers/gpu/drm/amd/display/Kconfig@@ -8,7 +8,7 @@ config DRM_AMD_DC depends on BROKEN || !CC_IS_CLANG || X86_64 || SPARC64 ||ARM64 select SND_HDA_COMPONENT if SND_HDA_CORE # !CC_IS_CLANG: https://github.com/ClangBuiltLinux/linux/issues/1752 - select DRM_AMD_DC_DCN if (X86 || PPC_LONG_DOUBLE_128 || (ARM64 && KERNEL_MODE_NEON && !CC_IS_CLANG)) + select DRM_AMD_DC_DCN if (X86 || PPC64 || (ARM64 && KERNEL_MODE_NEON +&& !CC_IS_CLANG)) help Choose this option if you want to use the new display engine support for AMDGPU. This adds required support for Vega and -- 2.34.1