brnstnd@kramden.acf.nyu.edu (Dan Bernstein) (11/20/90)
In article <1990Nov18.033622.1517@murdoch.acc.Virginia.EDU> gl8f@astsun9.astro.virginia.edu (Greg Lindahl) writes: > In article <14780:Nov1605:10:4490@kramden.acf.nyu.edu> brnstnd@kramden.acf.nyu.edu (Dan Bernstein) writes: > [ after insulting gcc, claiming it can't get strength reduction > correctly ] As a matter of fact, my ``insults'' were correct. Michael Meissner posted an article pointing out that gcc converts multiplication into a subroutine call on the Sparc, then fails to recognize the call during optimization. Greg, f*ck off. (Pardon my French, but the level of Greg's rhetoric is getting rather sickening.) > >All you wanted was real examples for real optimizers. I had thought that > >you were listening when other people posted real examples. Apparently > >not. Now I've posted the examples for you. > And your examples are wrong. Please, before you post any more > examples, check them for careless errors. My examples are right, and I most certainly did check for mistakes. > Stop misrepresenting mine. No existing compiler does the necessary > analysis. No compiler can do an anlysis as strict as the FORTRAN > anti-aliasing requirement. See my previous postings for the example > that breaks your technique. You never replied to my example. I haven't seen any such example. What was the message-id? > I guess you > find non-local analysis easier than analysis that doesn't have to go > outside of the basic block. No, I don't. If you understood the aliasing detection mechanism I've been talking about, you'd know that it does not require non-local or interprocedural analysis. (I think Jim finally accepts this fact.) If you're willing to do some non-local analysis, you can decide partitions on the basis of global use of variables. But the basic method doesn't depend on that. > And, finally, I guess you don't remember that I suggested that > comparing the starting addresses of arrays could lead to significant > optimizations quite a while ago Lots of people have suggested this, so you're right: I don't remember you in particular saying it. (The sorting thread started because I had pointed out the same run-time solution to Jim, and he didn't think comparing the addresses of n arrays could be done in linear time.) Anyway, a compile-time solution is always better than a run-time solution. ---Dan