2016-05-04 13:31 GMT+02:00 Oleg Endo <oleg.endo@xxxxxxxxxxx>:
> On Wed, 2016-05-04 at 12:57 +0200, Aurelien Buhrig wrote:
>
>> I'm trying to generate LTO optimized code (with -O2 -flto).
>> The code is very simple:
>>
>> int *SP;
>> int popInt(void) {return *--SP;}
>> void pushInt(int v) {*SP++ = v;}
>> void add (void) { pushInt(popInt()+popInt()); }
>>
>> When the 3 global functions popInt, pushInt and add are in the same
>> compilation unit, the lto optimization works as expected and
>> generates
>> something like
>> -*(SP -2) += *(SP-1);
>> SP--;
>>
>> But when add is in a different compilation unit, gcc cannot succeed
>> in
>> doing such an optimization at link time. Worse, it does not inline
>> the
>> popInt function (but it inlines pushInt, so lto are performed).
>>
>> I tried using the gcse options which are not enabled with O2, O3,
>> adding -finline-functions and tuning inline limits without success.
>> Any hint how I could get the same behavior with separated compilation
>> unit ?
>>
>> This is a gcc 4.6 on a private target.
>
> I'd recommend trying a newer version. Many improvements have been made
> in the past 4 or 5 years.
>
> Other than that, make sure that each compilation unit is compiled with
> -flto and the linking is done with the gcc driver program (also
> specifying -flto) and not by invoking LD directly.
>
> Cheers,
> Oleg
Thanks for your reply.
lto pass is correctly done. Maybe I would try upgrading the backend to
a newer version, but it is a quite important workload...
Thanks,
Aurélien
