permissible numerical optimizations (Thomas M. Breuel)
Tue, 14 Jul 1992 16:21:47 GMT

          From comp.compilers

Related articles
Common subexpression analysis (summary) (1992-06-26)
Re: Common subexpression analysis (summary) (1992-07-13)
permissible numerical optimizations (1992-07-14)
Re: permissible numerical optimizations (1992-07-16)
| List of all articles for this month |

Newsgroups: comp.compilers
From: (Thomas M. Breuel)
Organization: IDIAP (Institut Dalle Molle d'Intelligence Artificielle Perceptive)
Date: Tue, 14 Jul 1992 16:21:47 GMT
Keywords: optimize
References: 92-06-135 92-07-028

I'm not sure whether those optimizations that have been proposed by
various people in this newsgroup are "legal". It would seem to me that
they could result in worse accuracy than the original expression.

I have no trouble with optimizers modifying my numerical code to give me
higher accuracy or range (i.e., eliminate overflows as in "2*x-x" ==>
"x"), but giving me less accurate answers is generally not acceptable.

More generally, I'm curious to know: what kinds of liberties and
prohibitions do numerical optimizers operate under in various languages
and compilers (in particular for FORTRAN)?

[The F77 standard allows any mathematically equivalent operation, which is
so broad that real optimizers do much less.]

Post a followup to this message

Return to the comp.compilers page.
Search the comp.compilers archives again.