Re: [AArch64] Fix predicate and constraint mismatch in logical atomic operations

[Michael Collison <michael.collison@linaro.org>]

Segher,

The problem is the "CONST_INT 0", not a large constant. This constant is 
not accepted by the predicate, but is accepted by the constraint.

On 09/25/2014 03:12 AM, Segher Boessenkool wrote:
> On Wed, Sep 24, 2014 at 09:17:23PM -0700, Andrew Pinski wrote:
>> On Wed, Sep 24, 2014 at 9:13 PM, Michael Collison
>> <michael.collison@linaro.org> wrote:
>>> I have that attached to the bug report at the URL provided. I will work on a
>>> testcase if you think it is warranted.
>> Yes it is almost always warranted.
>>
>> https://gcc.gnu.org/contribute.html#patches
>>
>> Testcases   If you cannot follow the recommendations of the GCC coding
>> conventions about testcases, you should include a justification for
>> why adequate testcases cannot be added.
>>
>> See the last part of that sentence.  You don't have any justification
>> on why you are not including testcases.
> It is very hard to make a reliable testcase for such problems, because
> they only happen when register allocation is under pressure.
>
> The problem is not that "n" allows more than your predicate does.  The
> predicate allows registers too, so the compiler happily made a register
> contain some big const.  Now RA comes along, is out of registers but hey,
> there is this "n", let's just put the big constant there!  Carnage.
>
> So this is hard to test for; you can add some (big) code that exposed the
> problem, but in a few months time that won't trigger the problem anymore
> because earlier stages in the compiler will have generated slightly
> different code.
>
> It also does nothing to catch similar problems in other patterns.
>
>
> Segher

-- 
Michael Collison
Linaro Toolchain Working Group
michael.collison@linaro.org