https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61743
Bug ID: 61743
Summary: Complete unroll is not happened for loops with short
upper bound
Product: gcc
Version: 4.10.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ysrumyan at gmail dot com
We discovered significant performance regression on one important benchmark
from eembc2.0 suite after r211625. It turned out that complete unroll didn't
happen.
I prepared simple reproducer that exhibits the problem. If we compile it with
'-Dbtype=int' both innermost loops are unrolled completely by latest trunk
compiler:
test.c:20:5: note: loop with 8 iterations completely unrolled
test.c:11:5: note: loop with 8 iterations completely unrolled
but if '-Dbtype=e_u8' unroll does not happen.
Note also that for this particular test-case compiler built before r211625
performs curoll only for one loop.
