[PATCH] [RFC] x86: avoid -mtune=atom with clang

[PATCH] [RFC] x86: avoid -mtune=atom with clang

[Arnd Bergmann]

From: Arnd Bergmann <arnd@arndb.de>

The clang optimization for atom results in a large number of objtool
warnings like

  drivers/video/hdmi.o: warning: objtool: hdmi_infoframe_check()+0x74: unreachable instruction
  lib/crypto/curve25519.o: warning: objtool: init_module()+0x12: unreachable instruction
  drivers/clk/bcm/clk-iproc-armpll.o: warning: objtool: iproc_armpll_setup()+0x10c: unreachable instruction
  drivers/video/fbdev/core/cfbfillrect.o: warning: objtool: cfb_fillrect()+0xa2: unreachable instruction
  drivers/clk/bcm/clk-iproc-pll.o: warning: objtool: iproc_pll_clk_setup()+0x36d: unreachable instruction
  drivers/clk/bcm/clk-iproc-asiu.o: warning: objtool: iproc_asiu_setup()+0x2cf: unreachable instruction

and in rare cases problems with the register allocator:

  arch/x86/crypto/curve25519-x86_64.c:610:3: error: inline assembly requires more registers than available

  error: ran out of registers during register allocation

All of those can be avoided by changing the -mtune= option for
clang, leaving the -march= option unchanged. It's really a bug
that should be fxied in llvm, but it's fairly clear that the
atom optimizations are not well exercised in llvm, so not
using them is probably the safe choice regardless.

Link: https://github.com/ClangBuiltLinux/linux/issues/1483
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
---
 arch/x86/Makefile        | 4 ++++
 arch/x86/Makefile_32.cpu | 4 ++++
 2 files changed, 8 insertions(+)

diff --git a/arch/x86/Makefile b/arch/x86/Makefile
index 7488cfbbd2f6..4c6f92b32385 100644
--- a/arch/x86/Makefile
+++ b/arch/x86/Makefile
@@ -120,7 +120,11 @@ else
         cflags-$(CONFIG_MK8)		+= -march=k8
         cflags-$(CONFIG_MPSC)		+= -march=nocona
         cflags-$(CONFIG_MCORE2)		+= -march=core2
+ifdef CONFIG_CC_IS_CLANG
+        cflags-$(CONFIG_MATOM)		+= -march=atom -mtune=generic
+else
         cflags-$(CONFIG_MATOM)		+= -march=atom
+endif
         cflags-$(CONFIG_GENERIC_CPU)	+= -mtune=generic
         KBUILD_CFLAGS += $(cflags-y)
 
diff --git a/arch/x86/Makefile_32.cpu b/arch/x86/Makefile_32.cpu
index 94834c4b5e5e..a17b089f367c 100644
--- a/arch/x86/Makefile_32.cpu
+++ b/arch/x86/Makefile_32.cpu
@@ -33,8 +33,12 @@ cflags-$(CONFIG_MCYRIXIII)	+= $(call cc-option,-march=c3,-march=i486) $(align)
 cflags-$(CONFIG_MVIAC3_2)	+= $(call cc-option,-march=c3-2,-march=i686)
 cflags-$(CONFIG_MVIAC7)		+= -march=i686
 cflags-$(CONFIG_MCORE2)		+= -march=i686 $(call tune,core2)
+ifdef CONFIG_CC_IS_CLANG
+cflags-$(CONFIG_MATOM)		+= -march=atom -mtune=generic
+else
 cflags-$(CONFIG_MATOM)		+= $(call cc-option,-march=atom,$(call cc-option,-march=core2,-march=i686)) \
 	$(call cc-option,-mtune=atom,$(call cc-option,-mtune=generic))
+endif
 
 # AMD Elan support
 cflags-$(CONFIG_MELAN)		+= -march=i486
-- 
2.29.2

Re: [PATCH] [RFC] x86: avoid -mtune=atom with clang

[Miguel Ojeda]

On Tue, Oct 19, 2021 at 7:11 PM Arnd Bergmann <arnd@kernel.org> wrote:
>
> +ifdef CONFIG_CC_IS_CLANG
> +        cflags-$(CONFIG_MATOM)         += -march=atom -mtune=generic
> +else
>          cflags-$(CONFIG_MATOM)         += -march=atom
> +endif

Is it possible to factor it out? Also, a comment with the LLVM version
we last checked would be good. Something like...

    cflags-$(CONFIG_MATOM)         += -march=atom

    # Atom optimizations do not seem well exercised in, at least, LLVM <= 13,
    # to the point of getting problems with register allocation.
    ifdef CONFIG_CC_IS_CLANG
            cflags-$(CONFIG_MATOM)         += -mtune=generic
    endif

Cheers,
Miguel