http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55731
Bug #: 55731
Summary: Issue with complete innermost loop unrolling
(cunrolli)
Classification: Unclassified
Product: gcc
Version: 4.8.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: ysrum...@gmail.com
CC: hubi...@ucw.cz, izamya...@gmail.com
I attached 2 test-cases extracted from important benchmark at which clang and
icc outperform gcc for x86 target (atom). For 1st test-case (t.c) cunrolli
phase does not perform complete loop unrolling with the following message (test
was compiled with -O3 -funroll-loops options):
Loop size: 23
Estimated size after unrolling: 33
Not unrolling loop 1: size would grow.
but it is unrolled by cunroll phase:
Loop size: 24
Estimated size after unrolling: 32
Unrolled loop 1 completely (duplicated 2 times).
I wonder why this loop was not unrolled by cunrolli? We lost a lot of
optimizations for unrolled loop such as Constant (address) Propagation, Dead
code elimination etc. and got non-optimal binaries.
For comparsion I added another test (t2.c) with successfull complete unrolling
by cunrolli, at which we can see that all assignments to local array 'b' were
properly propagated and deleted but we don't have such transformations for 1st
test-case.
