On 08/06/15 17:31, Jakub Jelinek wrote:
> On Mon, Jun 08, 2015 at 02:04:12PM +0200, Tom de Vries wrote:
>> this patch fixes PR66444, a problem with -fipa-ra in reload_combine.
>>
>> The problem is that for the test-case, reload_combine combines these two
>> insns:
>
> Please work out with Vlad whether reload_cse_move2add doesn't need similar
> fix (and check other spots too).
>
>> 2015-06-08  Tom de Vries  <tom@codesourcery.com>
>>
>> 	PR rtl-optimization/66444
>> 	* postreload.c (reload_combine): Use get_call_reg_set_usage instead of
>> 	call_used_regs.
>
> LGTM.
>
>> 	* gcc.dg/pr66444.c: New test.
>
>> +int __attribute__((noinline, noclone))
>> +baz (void)
>> +{
>> +  struct S *x = (struct S *) 0xe0000000U;
>
> I'm still afraid this will not really work on s390-linux (which has only
> 31-bit pointers) and will not work on 16-bit int targets either
> (some have say 24-bit pointers etc., not really familiar with the embedded
> world).
> So, I'd suggest use a macro for the address, so you don't need to duplicate
> it,

Used a macro CONST_PTR.

> and define it to say ((struct S *) 0x8000UL), if it reproduces
> even with that change without your reload_combine fix.

Unfortunately, it didn't. So I used __SIZEOF_POINTER__ to produce a 
valid constant pointer for the 16-31 bit pointer-size cases, while still 
triggering the original problem for x86_64 -m64 case using a larger pointer.

Furthermore, I used __SIZEOF_POINTER__ to ensured a valid pointer 
constant suffix.

>
> Ok for trunk and 5.2 with that change.

Committed to trunk and backported gcc-5-branch as attached.

Thanks,
- Tom