Re: Funny? (Herman Rubin)
18 Apr 1997 01:33:20 -0400

          From comp.compilers

Related articles
Funny? (JUKKA) (1997-04-13)
Re: Funny? danwang@atomic.CS.Princeton.EDU (1997-04-16)
Re: Funny? (Paul David Fox) (1997-04-16)
Re: Funny? (1997-04-18)
Re: Funny? (William D Clinger) (1997-04-18)
Re: Funny? (1997-04-18)
| List of all articles for this month |

From: (Herman Rubin)
Newsgroups: comp.compilers
Date: 18 Apr 1997 01:33:20 -0400
Organization: Purdue University Statistics Department
References: 97-04-067 97-04-089
Keywords: optimize, performance

Daniel Wang <danwang@atomic.CS.Princeton.EDU> wrote:
><> writes:

>> I just created a Visual C++ program under Windows 95 which runs
>> slower when it is optimised for speed. And which runs faster when
>> it is a debug version without any optimisation and lot of extra
>> debug code.

>Just a guess, but sounds like the optimizer is unrolling some loop so that
>the "optimized" code no longer fits in the instruction cache, so you're
>paying a cache hit on every loop. The debug version is probably faster since
>it fits in the cache.

I remember a long time ago when I tried the main loop of a program
which had been written in assembler on the various compilers for a CDC
6500. Following Knuth's article on Fortran optimization, the best
that any conceivable Fortran compiler could do was 45%; to do better
would have required register management across blocks and conditional
calls. The non-optimizing compilers got about 33%, with half the loss
due to not being able to use assembler instructions for one small code
segment. The optimizing compilers got 23% efficiency; they made too
many wrong assumptions.
Herman Rubin, Dept. of Statistics, Purdue Univ., West Lafayette IN47907-1399 Phone: (765)494-6054 FAX: (765)494-0558

Post a followup to this message

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