--- Comment #27 from Jeffrey A. Law <law at redhat dot com> ---
Yea, we went through similar issues with Alex's work recently.  After that work
landed I went through all the BZs that I could find with autoinc
{pre,post}{inc,dec} that BZ could find in the hopes that his work would fix
some of them.

Sadly there was only 1 that was fixed.


So one concern if you try to expose autoincrement addressing in gimple is that
the RTL optimizers aren't prepared to handle until combine or later.  So
there's probably a fair amount of downstream work to do.

rth and I sketched out a pass that replace optimize_related_values and autoinc
passes using the concepts of PRE and SLSR many years ago.  It felt much cleaner
that what we've been doing and likely more powerful.  But we never went forward
with an implementation.

I got the sense from that recent BZ review that we're generally missing a fair
number of opportunities, but didn't dig into all of them to gain an
understanding of why we're doing so bad.

Reply via email to