[Bug target/62025] [4.9/4.10 Regression] Miscompilation of openssl sha512.c

[amker at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62025

--- Comment #12 from amker at gcc dot gnu.org ---
Hi,

I can reproduce the exact mis-scheduled instruction issue as in Jakub's comment
with/without the ivopt patch (204497).  Turns out gcc chooses candidate like
{&K512, 128}_loop with the patch while candidate {&K512[16], 128}_loop without
the patch.  This is caused by candidate cost change.  As far as I can see, both
candidate are good for the iv use and shouldn't generate wrong code either.  
The interesting thing is, mis-scheduled instruction is the only code difference
(apart from some offset difference as base of candidates are different) caused
by ivopt patch.  Normally middle-end patch should have larger impact on
generated code if there is.  I will have a look why different iv candidate can
result in one instruction mis-scheduled.  Maybe it's useful for investigation.

Anyway, I think 204497 is generally fine in this case.

Thanks,
bin