[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <8f2331db-151f-a481-23e0-ec6dd9ba6f1c@c-s.fr>
Date: Tue, 30 Jul 2019 07:31:23 +0200
From: Christophe Leroy <christophe.leroy@....fr>
To: Nathan Chancellor <natechancellor@...il.com>,
Nick Desaulniers <ndesaulniers@...gle.com>
Cc: mpe@...erman.id.au, segher@...nel.crashing.org, arnd@...db.de,
kbuild test robot <lkp@...el.com>,
Benjamin Herrenschmidt <benh@...nel.crashing.org>,
Paul Mackerras <paulus@...ba.org>,
linuxppc-dev@...ts.ozlabs.org, linux-kernel@...r.kernel.org,
clang-built-linux@...glegroups.com
Subject: Re: [PATCH] powerpc: workaround clang codegen bug in dcbz
Le 29/07/2019 à 22:32, Nathan Chancellor a écrit :
> On Mon, Jul 29, 2019 at 01:25:41PM -0700, Nick Desaulniers wrote:
>> Commit 6c5875843b87 ("powerpc: slightly improve cache helpers") exposed
>> what looks like a codegen bug in Clang's handling of `%y` output
>> template with `Z` constraint. This is resulting in panics during boot
>> for 32b powerpc builds w/ Clang, as reported by our CI.
>>
>> Add back the original code that worked behind a preprocessor check for
>> __clang__ until we can fix LLVM.
>>
>> Further, it seems that clang allnoconfig builds are unhappy with `Z`, as
>> reported by 0day bot. This is likely because Clang warns about inline
>> asm constraints when the constraint requires inlining to be semantically
>> valid.
>>
>> Link: https://bugs.llvm.org/show_bug.cgi?id=42762
>> Link: https://github.com/ClangBuiltLinux/linux/issues/593
>> Link: https://lore.kernel.org/lkml/20190721075846.GA97701@archlinux-threadripper/
>> Debugged-by: Nathan Chancellor <natechancellor@...il.com>
>> Reported-by: Nathan Chancellor <natechancellor@...il.com>
>> Reported-by: kbuild test robot <lkp@...el.com>
>> Suggested-by: Nathan Chancellor <natechancellor@...il.com>
>> Signed-off-by: Nick Desaulniers <ndesaulniers@...gle.com>
>> ---
>> Alternatively, we could just revert 6c5875843b87. It seems that GCC
>> generates the same code for these functions for out of line versions.
>> But I'm not sure how the inlined code generated would be affected.
>
> For the record:
>
> https://godbolt.org/z/z57VU7
>
> This seems consistent with what Michael found so I don't think a revert
> is entirely unreasonable.
Your example functions are too simple to show anything. The functions
takes only one parameter so of course GCC won't use two registers
allthough given the opportunity.
Christophe
>
> Either way:
>
> Reviewed-by: Nathan Chancellor <natechancellor@...il.com>
>
Powered by blists - more mailing lists