X-pop3-spooler: POP3MAIL 2.1.0 b 4 980420 -bs- Message-Id: Date: Thu, 3 Sep 98 08:57 From: strasbur AT chkw386 DOT ch DOT pwr DOT wroc DOT pl (Krzysztof Strasburger) To: pcg AT goof DOT com, strasbur AT chkw386 DOT ch DOT pwr DOT wroc DOT pl Subject: Re: pgcc-1.1a - first impression Cc: beastium-list AT desk DOT nl Sender: Marc Lehmann Status: RO X-Status: A Content-Length: 2074 Lines: 42 >> The first thing I noticed is code bloat :-(, but this is rather egcs >Does -fno-exceptions help? No. Anyway, my executables are linked with the old libgcc.a (2.7.2.3). >> [no speed improvement on pentium] >Thats, erhm, bad. But did you try -O6, as the most effective optimization >on pentiums is enabled with -O5 and higher only. -O6 and -O5 give identical code. It is really fast. Great! Previous pgcc gave nothing besides code bloat. These results have been obtained on Pentium 133. Pgcc uses the old instruction scheduler (no Haifa). Optimization pgcc-1.0.3a pgcc-1.1a level executable execution executable execution size times size times -O2 46632 39.39s 48072 39.75s -O3 46968 38.94s 48488 39.29s -O4 48056 38.97s 49384 39.15s -O5 52888 39.60s 54152 37.89s -Os - - 47080 39.24s gcc-2.7.2.3 -m486 -O2 -malign-double 45472 42.42s pgcc-1.1a -Os without double aligning 46664 39.42s >Having said this, I have no idea what could have happened to pentium opts, >as I can't remember changing them significantly ;( But the performance >problems in egcs/pgcc I talked about in my earlier mail is mainly related to >fpu-intensive programs... Well, optimizations below -O5 give really slower code than with pgcc-1.0.3a. >Could you try without the *align-double options? (just a guess) And maybe >you could post your benchmark data with -O2,3,6? See above. There is no big speed difference between executables with and without double aligning, because almost all data are allocated by malloc() and properly aligned. I found that this program is pretty immune against optimizations - maybe because the most critical code is split in short functions. The 10% speedup against the old gcc is very good result. >Thanks, Thanks for the compiler, I couldn't post these tests without it! Krzysztof