This is the mail archive of the
gcc-bugs@gcc.gnu.org
mailing list for the GCC project.
[Bug rtl-optimization/68695] [6 Regression] Performance regression related to ssa patch / ifcvt
- From: "law at redhat dot com" <gcc-bugzilla at gcc dot gnu dot org>
- To: gcc-bugs at gcc dot gnu dot org
- Date: Fri, 25 Mar 2016 23:28:25 +0000
- Subject: [Bug rtl-optimization/68695] [6 Regression] Performance regression related to ssa patch / ifcvt
- Auto-submitted: auto-generated
- References: <bug-68695-4 at http dot gcc dot gnu dot org/bugzilla/>
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68695
Jeffrey A. Law <law at redhat dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Priority|P1 |P2
CC| |law at redhat dot com
--- Comment #24 from Jeffrey A. Law <law at redhat dot com> ---
I think one could easily argue here that the only reason we got good code
before Alex's change is because the copyrename pass was buggy.
Given something like this:
# i_1 = PHI <x_3(D)(2), a_5(D)(3)>
Where x & a are parameters and i is a local, all of type int. The copyrename
pass would ignore the fact that x & a were promoted to DImode while i was
SImode
Alex's work exposed this as problematical in general and the suggested fix was
to not coalesce with the promoted modes were different.
I don't think this deserves to be a release blocker, but I do think we can
continue to look at either improving things on the RTL side, or finding a safe
set of cases where we can coalesce even when the promotions are different.