This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: Revision 107218 changed addressing mode generation
Richard Guenther wrote:
I have been resistant to change this canonicalization simply because
the current one seems to work well with the tree loop optimizers and
date-dependence analysis nowadays, thus also my hint at doing
such change during stage1 instead of stage4 (after all this "problem"
now exists since 4.1 ...?)
Well, it's apparently been found by two people before me, just not
fixed. I wouldn't put the word "problem" in quotes, as it is a real
performance regression for some code. I got back some benchmark results
from Robin Getz who tested my patch against the cryptopp library.
Here's a selection of the results from tests where gcc-4.3 regressed
against 4.1 on the Blackfin:
test 4.1 (trunk) 4.3 (trunk) 4.3 (patch)
DES/CTR 162.4 213.4 186.3
DES-XEX3/CTR 189.8 244.4 217.0
DES-EDE3/CTR 424.4 579.4 497.1
ECDHC over GF(2^n) 233 40.23 57.84 57.61
Key Agreement
ECDHC over GF(2^n) 233 39.92 57.84 57.84
Key-Pair Generation
Two-Track-MAC 42.0 59.0 41.6
Blowfish/CTR 110.9 106.1 99.5
RC5 (r=16) 60.1 76.6 76.4
MARS/CTR 95.4 107.9 103.4
CAST-256/CTR 280.7 314.1 262.9
[This is in cycles; lower is better]
Results overall:
results-4.1 (trunk) :Throughput Geometric Average: 8.02
results-4.3 (trunk) :Throughput Geometric Average: 10.14
results-4.3 (patch) :Throughput Geometric Average: 10.83
From Robin: "I looked through most of the tests - and didn't notice
anything that got worse with the patch - it was either better or the same."
Bernd
--
This footer brought to you by insane German lawmakers.
Analog Devices GmbH Wilhelm-Wagenfeld-Str. 6 80807 Muenchen
Sitz der Gesellschaft Muenchen, Registergericht Muenchen HRB 40368
Geschaeftsfuehrer Thomas Wessel, William A. Martin, Margaret Seif