Re: [PATCH] arm64: fix unreachable code issue with cmpxchg
From: Will Deacon
Date: Tue Sep 10 2019 - 09:24:23 EST
On Tue, Sep 10, 2019 at 10:04:24AM +0200, Arnd Bergmann wrote:
> On Tue, Sep 10, 2019 at 9:46 AM Will Deacon <will@xxxxxxxxxx> wrote:
> > On Mon, Sep 09, 2019 at 10:21:35PM +0200, Arnd Bergmann wrote:
> > > On arm64 build with clang, sometimes the __cmpxchg_mb is not inlined
> > > when CONFIG_OPTIMIZE_INLINING is set.
> > Hmm. Given that CONFIG_OPTIMIZE_INLINING has also been shown to break
> > assignment of local 'register' variables on GCC, perhaps we should just
> > disable that option for arm64 (at least) since we don't have any toolchains
> > that seem to like it very much! I'd certainly prefer that over playing
> > whack-a-mole with __always_inline.
> Right, but I can also see good reasons to keep going:
> - In theory, CONFIG_OPTIMIZE_INLINING is the right thing to do -- the compilers
> also make some particularly bad decisions around inlining when each inline
> turns into an __always_inline, as has been the case in Linux for a long time.
> I think in most cases, we get better object code with CONFIG_OPTIMIZE_INLINING
> and in the cases where this is worse, it may be better to fix the compiler.
> The new "asm_inline" macro should also help with that.
Sure, in theory, but it looks like there isn't a single arm64 compiler out
there which gets it right.
> - The x86 folks have apparently whacked most of the moles already, see this
> commit from 2008
> commit 3f9b5cc018566ad9562df0648395649aebdbc5e0
> Author: Ingo Molnar <mingo@xxxxxxx>
> Date: Fri Jul 18 16:30:05 2008 +0200
> x86: re-enable OPTIMIZE_INLINING
> re-enable OPTIMIZE_INLINING more widely. Jeff Dike fixed the remaining
> outstanding issue in this commit:
> | commit 4f81c5350b44bcc501ab6f8a089b16d064b4d2f6
> | Author: Jeff Dike <jdike@xxxxxxxxxxx>
> | Date: Mon Jul 7 13:36:56 2008 -0400
> | [UML] fix gcc ICEs and unresolved externs
> | This patch reintroduces unit-at-a-time for gcc >= 4.0,
> bringing back the
> | possibility of Uli's crash. If that happens, we'll debug it.
> it's still default-off and thus opt-in.
This appears to be fixing an ICE, whereas the issue reported recently for
arm64 gcc was silent miscompilation of atomics in some cases. Unfortunately,
I can't seem to find the thread :/ Mark, you were on that one too, right?
> - The inlining decisions of gcc and clang are already very different, and
> the bugs we are finding around that are much more common than
> the difference between CONFIG_OPTIMIZE_INLINING=y/n on a
> given compiler.
Sorry, not sure that you're getting at here.
Anyway, the second version of your patch looks fine, but I would still
prefer to go the extra mile and disable CONFIG_OPTIMIZE_INLINING altogether
given that I don't think it's a safe option to enable for us.