This is the mail archive of the
gcc-bugs@gcc.gnu.org
mailing list for the GCC project.
[Bug tree-optimization/21485] [5/6/7 Regression] missed load PRE, PRE makes i?86/7 suck
- From: "law at redhat dot com" <gcc-bugzilla at gcc dot gnu dot org>
- To: gcc-bugs at gcc dot gnu dot org
- Date: Tue, 07 Feb 2017 23:03:39 +0000
- Subject: [Bug tree-optimization/21485] [5/6/7 Regression] missed load PRE, PRE makes i?86/7 suck
- Auto-submitted: auto-generated
- References: <bug-21485-4@http.gcc.gnu.org/bugzilla/>
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=21485
Jeffrey A. Law <law at redhat dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |law at redhat dot com
--- Comment #62 from Jeffrey A. Law <law at redhat dot com> ---
So one thing that seems to ever so slightly improve this code is to realize
that as we come around to the top of the loop the test _14 < j_20 can be
rewritten as _14 != j_20 (we already know that _14 <= j_20).
That's quite a surprise as all that seems to do is allow propagation of j_20
for _14 in a later PHI node and we do less use work in PRE -- so it seems like
a step backwards at this point.
But then sink comes along and moves two address computations and a load into a
more control dependent block. LIM4 later pulls the two address computations
totally out of the loop.
But that all seems to be relatively accidental improvements based on PRE not
seeing a transformation.