On 18/11/13 08:37, Yvan Roux wrote:
So, the validation is ok with this patch, I'm just not able to say if
the original performance issue is still fixed with it. Could you
check it Kyrylo ?
Hi Yvan,
I'll run the benchmark today to confirm the performance, but from compiling some
code sequences that exhibited the bad behaviour in the past, I see that this
patch still fixes the issues. store_minmaxsi is not generated when optimising
for speed.
Thanks,
Kyrill
Yvan
2013-11-17 Yvan Roux <yvan.r...@linaro.org>
* config/arm/arm.md (store_minmaxsi): Use only when
optimize_function_for_size_p.
On 15 November 2013 15:59, Yvan Roux <yvan.r...@linaro.org> wrote:
Sometimes 4 will be needed, since both original register values may
remain live.
Indeed.
However, I'm inclined to agree that while it should be possible to
decide at the *function* level whether or not an insn is valid, doing so
at the block level is probably unsafe.
Ok, so the attached patch should fix the issue, its validation is ongoing.