The testcase from PR target/27827 shows another problem, this time with
-ffast-math. The runtime performance of -ffast-math code drops for ~30%.
The problem could be traced down to reassociation tree pass, because the
performance jumps back when "flag_unsafe_math_optimizations" switch is disabled
by changing every occurence in tree-ssa-reassoc.c with
(flag_unsafe_math_optimizations && 0).
To see the problem, -funsafe-math-optimizations should be added to MMFLAGS in
target/27827 example Makefile:
MM4FLAGS = $(GMMFLAGS) -funsafe-math-optimizations
Current mainline gcc produces code with following results:
-O -mfpmath=387
ALGORITHM NB REPS TIME MFLOPS
========= ===== ===== ========== ==========
atlasmm 60 1000 0.260 1663.04
-O -msse2 -mfpmath=sse
ALGORITHM NB REPS TIME MFLOPS
========= ===== ===== ========== ==========
atlasmm 60 1000 0.229 1890.47
gcc with disabled reassoc pass for floating point values:
-O -mfpmath=387
ALGORITHM NB REPS TIME MFLOPS
========= ===== ===== ========== ==========
atlasmm 60 1000 0.162 2664.87
-O -msse2 -mfpmath=sse
ALGORITHM NB REPS TIME MFLOPS
========= ===== ===== ========== ==========
atlasmm 60 1000 0.164 2633.15
--
Summary: reassociation pass produces ~30% slower matrix
multiplication code
Product: gcc
Version: 4.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: uros at kss-loka dot si
GCC build triplet: i686-pc-linux-gnu
GCC host triplet: i686-pc-linux-gnu
GCC target triplet: i686-pc-linux-gnu
OtherBugsDependingO 27827
nThis:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27855
