Re: [PATCH] bug.h: Work around GCC PR82365 in BUG()

From: Arnd Bergmann
Date: Wed Dec 20 2017 - 04:01:41 EST

On Tue, Dec 19, 2017 at 11:38 PM, Vineet Gupta
<Vineet.Gupta1@xxxxxxxxxxxx> wrote:
> On 12/19/2017 12:13 PM, Arnd Bergmann wrote:
>>> I suppose BUG() implies "dead end" like semantics - which ARC was lacking
>>> before ?
>> Correct. Using __builtin_trap() here avoids the 'control reaches end of
>> non-void
>> function' warnings, but then makes us run into the stack size problem that
>> I work around with the barrier_before_unreachable().
>> It would be good if you could give this a quick test to see if you get
>> sensible
>> output from the __builtin_trap();
> It does, added a BUG() arbit, hits an abort()
> ...
> ISA Extn : atomic ll64 unalign (not used)
> : mpy[opt 9] div_rem norm barrel-shift swap minmax swape
> BPU : partial match, cache:2048, Predict Table:16384
> BUG: failure at ../arch/arc/mm/tlb.c:827/arc_mmu_init()!
> Tested-by: Vineet Gupta <vgupta@xxxxxxxxxxxx>

I meant whether it prints the right registers and stack trace, but I
assume you tested that and just did not list it above.

> FWIW newer ARC gcc actually implements the builtin so we get a trap 5
> instruction now, vs., abort() calls before.
> BTW I missed reading the hunk of your changelog where this addresses the
> long standing mystery with ARC builds and numerous -Wreturn-type warnings. I
> always wondered why they were not fixed upstream already, being too lazy to
> investigate myself, and turns out this was due to this BUG() thingy. phew !

Hmm, so with the new definition of abort(),

+__weak void abort(void)
+ BUG();
+ /* if that doesn't kill us, halt */
+ panic("Oops failed to kill thread");

won't that run into an endless recursion? Or do you then override abort()
for ARC?