This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Best way to compute cost of a sequence of gimple stmt


On Wed, Jun 11, 2014 at 10:27 AM, Thomas Preud'homme
<thomas.preudhomme@arm.com> wrote:
>> From: Richard Biener [mailto:richard.guenther@gmail.com]
>> Sent: Wednesday, June 11, 2014 4:09 PM
>
>> >
>> > Oh I see. Doing it there would mean instead of two independent
>> > operations you'd do the best combination possible, is that right?
>>
>> Yes (but probably it's not worth the trouble).
>
> I understood that.
>
>>
>> > I'm tempted to use a simple heuristic such as comparing the
>> > number of loads before and after, adding one if the load is
>> > unaligned. So in the above example, supposing that there is
>> > some computation done around x[0] before the return line,
>> > we'd have 2 loads before Vs 2 x is unaligned and we would
>> > cancel the optimization. If x is aligned the optimization would
>> > proceed.
>> >
>> > Do you thing this approach is also too much trouble or would
>> > not work?
>>
>> I'm not sure.  For noop-loads I'd keep them unconditionally, even if
>> unaligned.  I'd disable unaligned-load + bswap for now.  People
>> interested and sitting on such a target should do the measurements
>> and decide if it's worth the trouble (is arm affected?).
>
> Yes it is.
>
>>
>> But I see that the code currently does not limit itself to single-use
>> chains and thus may end up keeping the whole original code life
>> by unrelated uses.  So a good thing would be to impose proper
>> restrictions here.  For example, in find_bswap_or_nop_1 do
>>
>>   if (TREE_CODE (rhs1) != SSA_NAME
>>      || !has_single_use (rhs1))
>
> But then the example in gcc.dg/optimize-bswapdi-2.c would not
> work for instance. Same for swap32_b in gcc.dg/optimize-bswapsi-1.c
>
> To make it work you'd need to check that there is no use outside the
> sets of statements that form the bitwise OR operation you are
> considering.

Yes, of course (also for generic shuffles that may have duplicate entries).

Richard.

> Best regards,
>
> Thomas
>
>


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]