Author: Gerd Isenberg
Date: 11:06:21 07/06/03
Go up one level in this thread
On July 06, 2003 at 13:49:30, Dieter Buerssner wrote: >On July 06, 2003 at 12:23:32, Jens Kahlenberg wrote: > >>Would be interesting to know what version Dieter is using and to find out why >>his version seems to have a __builtin_abs with branching target-code. > >Gcc 3.2 from DJGPP. It defaults to the branching version. When using >-march=pentium4 it uses conditional move (also with simple_abs). I use this, to >make DOS executables. They should normally run with 386. I have actually sent my >engine to few people with no pentium. Gcc from MinGW shows the same behaviour >(no surprise). > >BTW. Using the branchless code may very well make real applications (vs. stupid >tight loop testing of abs()) slower. For example the cdq method will use eax and >edx registers always, an might make surrounding code slower, because no values >can be let in those registers. The branching method will need no additional >register at all, often, and will work with any register. Yes, a trade off, one additional register against the probability of branch misprediction. > >>Dieter >>might even (with a lot of work) optimize further by compiling a new >>(P4-targeted) gcc-compiler on his system with the existing one and perhaps abs >>will be alright then. > >I doubt it very much. The optimization of the compiler itself should nothing >have to do with the generated code. If it does, the most probable reason might >be a bug in the compiler. Actually gcc bootstrapping procedure does this test. >It compiles the compiler 3 times. First with the old compiler, than with the new >compiler, then again ... Last 2 compiles are compared. > >Regards, >Dieter
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.