From: Paolo Bonzini <pbonzini@redhat.com>
To: Dmitry Vyukov <dvyukov@google.com>
Cc: "Radim Krčmář" <rkrcmar@redhat.com>,
"David Hildenbrand" <david@redhat.com>,
LKML <linux-kernel@vger.kernel.org>,
"KVM list" <kvm@vger.kernel.org>,
llvmlinux@lists.linuxfoundation.org,
"Alexander Potapenko" <glider@google.com>,
andreyknvl <andreyknvl@google.com>,
"Michael Davidson" <md@google.com>,
"Greg Hackmann" <ghackmann@google.com>,
"Nick Desaulniers" <ndesaulniers@google.com>
Subject: Re: "KVM: x86: generalize guest_cpuid_has_ helpers" breaks clang
Date: Tue, 12 Sep 2017 18:33:24 +0200 [thread overview]
Message-ID: <74260214-ade2-48d8-9b18-d00e31b11421@redhat.com> (raw)
In-Reply-To: <CACT4Y+bRNVv9gcTcK+XCZROawa5pmiDBpb9fAQUhHm1Vej5Tfw@mail.gmail.com>
On 12/09/2017 18:16, Dmitry Vyukov wrote:
> On Tue, Sep 12, 2017 at 6:03 PM, Paolo Bonzini <pbonzini@redhat.com> wrote:
>> On 12/09/2017 17:54, Dmitry Vyukov wrote:
>>>> I guess clang still eliminates dead branches. Clang optimizer does
>>>> know that these are constant, it just does not allow build
>>>> success/failure nor runtime behavior depend on optimization level and
>>>> compiler version. I.e. with gcc you can get build failure with only
>>>> some compiler flags and/or compiler versions. Clang gives stable
>>>> result. But the optimizer does use constant propagation, etc during
>>>> optimization.
>>
>> I can reproduce it:
>>
>> $ cat f.c
>> int bad_code();
>>
>> static inline void __attribute__((always_inline)) f(int x)
>> {
>> if (!__builtin_constant_p(x))
>> bad_code();
>> }
>>
>> int main()
>> {
>> f(0);
>> f(1);
>> f(100);
>> }
>>
>> $ clang --version
>> clang version 4.0.0 (tags/RELEASE_400/final)
>> $ clang f.c -O2 -c -o f.o
>> $ nm f.o
>> U bad_code
>> 0000000000000000 T main
>>
>> $ gcc f.c -O2 -c -o f.o
>> $ nm f.o
>> 0000000000000000 T main
>>
>> ... but I don't know, it seems very weird. The purpose of
>> __builtin_constant_p is to be resolved only relatively late in the
>> optimization pipeline, and it has been like this for at least 15 years
>> in GCC.
>>
>> The docs say what to expect:
>>
>> You may use this built-in function in either a macro or an inline
>> function. However, if you use it in an inlined function and pass an
>> argument of the function as the argument to the built-in, GCC never
>> returns 1 when you call the inline function with a string constant or
>> compound literal (see Compound Literals) and does not return 1 when
>> you pass a constant numeric value to the inline function **unless you
>> specify the -O option**.
>>
>> (emphasis mine).
>
>
> Yes, I know. This difference was surprising for me and lots of other
> people as well. But this is a fundamental position for clang and is
> related to some implementation choices. Namely, C/C++ frontend needs
> to know values of compile-time const expressions in order to verify
> correctness and generate middle-end representation. But for gcc's
> meaning of __builtin_constant_p, its value only becomes known deep
> inside of middle-end. Which kinda creates a cycle. In gcc it's all
> somehow mixed together (front-end/middle-end) and somehow works. Can't
> possibly work for clang with strict separation between front-end and
> middle-end.
This is nonsense, GCC is also separating front-end and middle-end. The
front-end only ever produces a 0 value for __builtin_constant_p if an
integer constant expression is syntactically required.
When entering the middle-end a __builtin_constant_p with non-constant
argument is lowered to a builtin function when optimization is on, or 0
when optimization is off.
The middle-end knows about __builtin_constant_p and can fold it to 1
when the argument is a constant. At some point, GCC decides it's had
enough and changes all remaining calls to return 0. There's no reason
why LLVM couldn't have such a builtin.
> I proposed to introduce another builtin that returns a value that is
> constant from optimizer point of view (e.g. it can eliminate dead code
> on branches), but is not constant from language/front-end point of
> view (e.g. you can't declare a stack array using the value as size).
> It should do in such cases and should be implementable in clang. But
> we don't have it yet, and again it's not __builtin_constant_p, because
> gcc's __builtin_constant_p returns a compile-time constant.
I think this has to be fixed at the include/linux/ level. I'm okay with
warning instead of erroring, so maybe add WARN_IF_NONCONSTANT() and make
it do nothing (or live with the warning) on clang?
Paolo
next prev parent reply other threads:[~2017-09-12 16:33 UTC|newest]
Thread overview: 9+ messages / expand[flat|nested] mbox.gz Atom feed top
2017-09-12 14:42 Dmitry Vyukov
2017-09-12 15:18 ` Radim Krčmář
2017-09-12 15:51 ` Dmitry Vyukov
2017-09-12 15:54 ` Dmitry Vyukov
2017-09-12 16:03 ` Paolo Bonzini
2017-09-12 16:16 ` Dmitry Vyukov
2017-09-12 16:33 ` Paolo Bonzini [this message]
2017-09-12 17:33 ` Dmitry Vyukov
2017-09-13 10:58 ` Radim Krčmář
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=74260214-ade2-48d8-9b18-d00e31b11421@redhat.com \
--to=pbonzini@redhat.com \
--cc=andreyknvl@google.com \
--cc=david@redhat.com \
--cc=dvyukov@google.com \
--cc=ghackmann@google.com \
--cc=glider@google.com \
--cc=kvm@vger.kernel.org \
--cc=linux-kernel@vger.kernel.org \
--cc=llvmlinux@lists.linuxfoundation.org \
--cc=md@google.com \
--cc=ndesaulniers@google.com \
--cc=rkrcmar@redhat.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®