This is the mail archive of the
gcc-bugs@gcc.gnu.org
mailing list for the GCC project.
[Bug rtl-optimization/34503] Issues with constant/copy propagation implementation in gcse.c
- From: "steven at gcc dot gnu dot org" <gcc-bugzilla at gcc dot gnu dot org>
- To: gcc-bugs at gcc dot gnu dot org
- Date: 17 Dec 2007 08:30:50 -0000
- Subject: [Bug rtl-optimization/34503] Issues with constant/copy propagation implementation in gcse.c
- References: <bug-34503-280@http.gcc.gnu.org/bugzilla/>
- Reply-to: gcc-bugzilla at gcc dot gnu dot org
------- Comment #3 from steven at gcc dot gnu dot org 2007-12-17 08:30 -------
The global pass uses inefficient local dataflow properties. Because there is a
local cprop pass before the global one, some constants/copies do not have to be
recorded. This results in a smaller set of expressions to compute AVAIL for.
Some experiments I did, showed that the set can be reduced by more than an
order of magnitude for large functions (for example try the idea on
c-common.c).
A register R in defined a load-immediate instruction (i.e. R<-const) in basic
block B should not be recorded if the register is not in LIVE_OUT(B)After local
cprop, the constant is propagated as far as possible within B, and the
definition of R does not reach any uses outside B. Thus, not recording the
register value cannot result in missed optimizations.
Likewise, if R1 is defined in a copy instruction R1<-R2 in basic block B, and
R1 is not in LIVE_OUT(B), and R2 is also not in LIVE_OUT(B), then recording the
copy cannot result in missed optimizations, because R2 will have been
propagated locally as far as possible already.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34503