Due to a backend bug, dbr had picked a delay slot insn for annul-true which was
not actually elegible for annul-true. When I fixed the bug, I found that
instead an insn from the target path was chosen, the restore of the return
address, as the target is an epilogue.
The original instruction, mov r4,-1 , would have been suitable as a non-anulled
delay slot insn, since r4 is a call-used register, and the function does not
return a value that would require r4 to represent, and the epilogue did not
make use of the value in r4.
I've seen this for ARC compiling
cjpeg/jcmarker.c
using the options:
-mnorm -mswap -mmul64 -mARC600 -O3 -fomit-frame-pointer
in the function write_file_header.
For obvious reasons I can't provide preprocessed source for some eighty years.
--
Summary: delared branch scheduling doesn't fully take return into
account
Product: gcc
Version: 4.2.1
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: minor
Priority: P3
Component: rtl-optimization
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: amylaar at gcc dot gnu dot org
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38452
