Skip to contentSkip to conversation

include/linux/compiler-clang.h: define __naked

PATCH In review Stefan Agner 4 participants 6 messages Sep 10, 2018 – Sep 13, 2018 Clang/LLVM build support

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 include/linux/compiler-clang.h: define __naked

ARM32 arch code uses the __naked attribute. This has previously been defined in include/linux/compiler-gcc.h, which is no longer included for Clang. Define __naked for Clang. Conservatively add all attributes previously used (and supported by Clang).

This fixes compile errors when building ARM32 using Clang:
arch/arm/mach-exynos/mcpm-exynos.c:193:13: error: variable has incomplete type 'void' static void __naked exynos_pm_power_up_setup(unsigned int affinity_level)

              ^
Fixes 815f0ddb346c ("include/linux/compiler*.h: make compiler-*.h mutually exclusive")
---
 include/linux/compiler-clang.h | 6 ++++++
 1 file changed, 6 insertions(+)
include/linux/compiler-clang.h +6 -0
index b1ce500fe8b3..a593e3ac0720 100644
@@ -23,6 +23,12 @@
 
 #define __no_sanitize_address __attribute__((no_sanitize("address")))
 
+/*
+ * ARM32 is currently the only user of __naked supported by Clang. Follow
+ * gcc: Do not trace naked functions and make sure they don't get inlined.
+ */
+#define __naked __attribute__((naked)) noinline notrace
+
 /*
  * Not all versions of clang implement the the type-generic versions
  * of the builtin overflow checkers. Fortunately, clang implements
Signature · 1 line
2.18.0
#1 lore raw
Quoted 14 lines · Stefan Agner

ARM32 arch code uses the __naked attribute. This has previously been defined in include/linux/compiler-gcc.h, which is no longer included for Clang. Define __naked for Clang. Conservatively add all attributes previously used (and supported by Clang).

This fixes compile errors when building ARM32 using Clang:
arch/arm/mach-exynos/mcpm-exynos.c:193:13: error: variable has incomplete type 'void' static void __naked exynos_pm_power_up_setup(unsigned int affinity_level)

              ^
Fixes 815f0ddb346c ("include/linux/compiler*.h: make compiler-*.h mutually exclusive")

include/linux/compiler-clang.h line 31

Quoted 6 lines
include/linux/compiler-clang.h +6 -0
+/*
+ * ARM32 is currently the only user of __naked supported by Clang. Follow
+ * gcc: Do not trace naked functions and make sure they don't get inlined.
+ */
+#define __naked __attribute__((naked)) noinline notrace
+

Please see patches 5 and 6 of the series that Miguel posted:

https://lore.kernel.org/lkml/20180908212459.19736-6-miguel.ojeda.sandonis at gmail.com/

I suppose we want the patch to fix clang build as soon as possible though, and follow up with the cleanup for the next merge window, right?

Arnd

#2 lore raw
Quoted 8 lines · Arnd Bergmann
Quoted 6 lines · Stefan Agner

ARM32 arch code uses the __naked attribute. This has previously been defined in include/linux/compiler-gcc.h, which is no longer included for Clang. Define __naked for Clang. Conservatively add all attributes previously used (and supported by Clang).

Quoted 7 lines
Quoted 7 lines

This fixes compile errors when building ARM32 using Clang:
arch/arm/mach-exynos/mcpm-exynos.c:193:13: error: variable has incomplete type 'void' static void __naked exynos_pm_power_up_setup(unsigned int affinity_level)

              ^
Fixes 815f0ddb346c ("include/linux/compiler*.h: make compiler-*.h mutually exclusive")

Cool, thanks for this patch! I'll admit I need to start testing more architectures with Clang.

Quoted 11 lines

include/linux/compiler-clang.h line 31

Quoted 7 lines
include/linux/compiler-clang.h +6 -0
+/*
+ * ARM32 is currently the only user of __naked supported by Clang. Follow
+ * gcc: Do not trace naked functions and make sure they don't get inlined.
+ */
+#define __naked __attribute__((naked)) noinline notrace
+

Please see patches 5 and 6 of the series that Miguel posted:

https://lore.kernel.org/lkml/20180908212459.19736-6-miguel.ojeda.sandonis at gmail.com/

Yes, I'd prefer those 2 patches.

Quoted 3 lines

I suppose we want the patch to fix clang build as soon as possible though, and follow up with the cleanup for the next merge window, right?

Can we take Miguel's two patches from above in the ARM tree, since this attribute is only used on ARM, IIUC?

Signature · 2 lines
Thanks,
~Nick Desaulniers
#3 lore raw
In reply to Arnd Bergmann

Hi Arnd, Nick, Stefan,

Quoted 24 lines · Arnd Bergmann
Quoted 14 lines · Stefan Agner

ARM32 arch code uses the __naked attribute. This has previously been defined in include/linux/compiler-gcc.h, which is no longer included for Clang. Define __naked for Clang. Conservatively add all attributes previously used (and supported by Clang).

This fixes compile errors when building ARM32 using Clang:
arch/arm/mach-exynos/mcpm-exynos.c:193:13: error: variable has incomplete type 'void' static void __naked exynos_pm_power_up_setup(unsigned int affinity_level)

              ^
Fixes 815f0ddb346c ("include/linux/compiler*.h: make compiler-*.h mutually exclusive")

include/linux/compiler-clang.h line 31

Quoted 6 lines
include/linux/compiler-clang.h +6 -0
+/*
+ * ARM32 is currently the only user of __naked supported by Clang. Follow
+ * gcc: Do not trace naked functions and make sure they don't get inlined.
+ */
+#define __naked __attribute__((naked)) noinline notrace
+

Please see patches 5 and 6 of the series that Miguel posted:

Quoted 4 lines

https://lore.kernel.org/lkml/20180908212459.19736-6-miguel.ojeda.sandonis at gmail.com/

I suppose we want the patch to fix clang build as soon as possible though, and follow up with the cleanup for the next merge window, right?

Not sure what the plans of Linus et. al. are, if they have any; but that would be a safe bet.

In case they want to speed this up and put the entire series into v4.19 (instead of the two patches), I have done a binary & objdump diff between -rc2 and v4 (based on -rc2) on all object files (with UTS_RELEASE fixed to avoid some differences).

In a x86_64 tinyconfig with gcc 7.3, the differences I found are:

$ ./compare.py linux-rc2 linux-v4
[2018-09-12 06:16:39,483] [INFO] [arch/x86/boot/compressed/piggy.o]
Binary diff (use 'bash -c "cmp
linux-rc2/arch/x86/boot/compressed/piggy.o
linux-v4/arch/x86/boot/compressed/piggy.o"' to replicate)
[2018-09-12 06:16:39,606] [INFO] [arch/x86/boot/header.o] Binary diff
(use 'bash -c "cmp linux-rc2/arch/x86/boot/header.o
linux-v4/arch/x86/boot/header.o"' to replicate)
[2018-09-12 06:16:39,659] [INFO] [arch/x86/boot/version.o] Binary diff
(use 'bash -c "cmp linux-rc2/arch/x86/boot/version.o
linux-v4/arch/x86/boot/version.o"' to replicate)
[2018-09-12 06:16:40,483] [INFO] [init/version.o] Binary diff (use
'bash -c "cmp linux-rc2/init/version.o linux-v4/init/version.o"' to
replicate)

I will do a bigger one tomorrow or so and see if there are any important differences. Regardless of what we do, I will send the __naked patches separately as well (requested by Nick on GitHub).

Cheers,
Miguel

#4 lore raw

Hi again,

Quoted 4 lines · Miguel Ojeda

I will do a bigger one tomorrow or so and see if there are any important differences. Regardless of what we do, I will send the __naked patches separately as well (requested by Nick on GitHub).

So I did a comparison with a full allyesconfig between -rc2 and -rc2 + v4. I find quite a lot binary differences, but not a single objdump -d difference (object file by object file) just by fixing UTS_RELEASE to the same value, so I guess that is a very good sign in case someone wants to pick the entire series sooner than expected. Regardless, I will send the separate __naked patches.

I should note that, at some point in one of my allyesconfig builds I got this:

  UPD     include/generated/compile.h
  CC      init/version.o
  AR      init/built-in.a
  AR      built-in.a
  LD      vmlinux.o

MODPOST vmlinux.o
ld: drivers/edac/amd64_edac.o(.text+0x200031d2): reloc against `__asan_load4_noabort': error 4
ld: final link failed: Nonrepresentable section on output Makefile:1035: recipe for target 'vmlinux' failed
make: *** [vmlinux] Error 1

Running make (without cleaning) didn't reproduce it, and the other allyesconfig builds didn't suffer any problem either. Maybe my hardware is failing, but I wanted to let you know nevertheless.

Cheers,
Miguel

#5 lore raw
Quoted 24 lines · Miguel Ojeda

Hi again,

Quoted 4 lines · Miguel Ojeda

I will do a bigger one tomorrow or so and see if there are any important differences. Regardless of what we do, I will send the __naked patches separately as well (requested by Nick on GitHub).

So I did a comparison with a full allyesconfig between -rc2 and -rc2 + v4. I find quite a lot binary differences, but not a single objdump -d difference (object file by object file) just by fixing UTS_RELEASE to the same value, so I guess that is a very good sign in case someone wants to pick the entire series sooner than expected. Regardless, I will send the separate __naked patches.

I should note that, at some point in one of my allyesconfig builds I got this:

  UPD     include/generated/compile.h
  CC      init/version.o
  AR      init/built-in.a
  AR      built-in.a
  LD      vmlinux.o
Quoted 6 lines

MODPOST vmlinux.o
ld: drivers/edac/amd64_edac.o(.text+0x200031d2): reloc against `__asan_load4_noabort': error 4
ld: final link failed: Nonrepresentable section on output Makefile:1035: recipe for target 'vmlinux' failed
make: *** [vmlinux] Error 1

+ Dmitry for the KASAN error (Andrey was already cc'ed)

Quoted 7 lines

Running make (without cleaning) didn't reproduce it, and the other allyesconfig builds didn't suffer any problem either. Maybe my hardware is failing, but I wanted to let you know nevertheless.

Cheers,
Miguel

Signature · 2 lines
Thanks,
~Nick Desaulniers