From: Oscar F. <of...@wa...> - 2001-10-13 04:15:33
|
=?iso-8859-1?Q?Danny_Smith?= writes: > The current default for CPU tuning of code on mingw is i386 (set in specs). > This can of course be overriden on command line, but shouldn't we bump the > default to 486 or 586? First, what would be gain? I expect a really tiny performance improvement from setting the cpu to 486. As for the 586, I don't know how much different is from the 386 tuning. If the 386 tuning predates the 586, I expect a measurable improvement. Anyone has a benchmark library? Second, what would be loss? Possibly, the 386 is the most tested tuner ever. I don't know if the 486/586 tuners could be considered less reliable on this regard. Be ready for a bunch of reports of the type "my app crashes since I updated mingw". A lot of people does not read the changelog (or even does not understand it) and can't mind how this change affects its code. On the other side, I've observed that some major libraries and apps sets the cpu to 586 in their makefiles. If someone wants to squeeze performance, he should read the fine manual, as ever. > If nothing else, should we build the toolset binaries tuned to run on at > least 486? I bet that the speed gain will be well under 5%. How many people out there are using mingw on "garbage-type" 386 machines (for teaching C, for example)? My position on this regard is even more conservative. > Please don't suggest bumping to i686 unless you send me some money to buy a > new PC. If a new computer would make you more productive as a mingw developer, you should consider making a collect ;-) -- Oscar. |