Re: Code quality

davidm@questor.rational.com (David Moore)
Wed, 6 Jan 1993 19:24:21 GMT

          From comp.compilers

Related articles
Code quality drw@zermelo.mit.edu (1993-01-06)
Re: Code quality preston@dawn.cs.rice.edu (1993-01-06)
Re: Code quality davidm@questor.rational.com (1993-01-06)
Re: Code quality henry@zoo.toronto.edu (1993-01-06)
Gcc, Lcc, and 2c mike@skinner.cs.uoregon.edu (Michael John Haertel) (1993-01-07)
Re: Code quality tchannon@black.demon.co.uk (1993-01-07)
Re: Code quality prener@watson.ibm.com (1993-01-07)
Re: Code quality ssimmons@convex.com (1993-01-07)
Re: Code quality bill@amber.csd.harris.com (1993-01-07)
[8 later articles]
| List of all articles for this month |
Newsgroups: comp.compilers
From: davidm@questor.rational.com (David Moore)
Organization: Rational
Date: Wed, 6 Jan 1993 19:24:21 GMT
References: 93-01-017
Keywords: optimize, comment

drw@zermelo.mit.edu (Dale R. Worley) writes:


>How important is generated code quality these days? There are a lot of
>good optimization techniques that seem to be adequate for ordinary
>programming. But they still are at least 10% or 20% worse than the ideal.
>Is there much of a market for another 10% in speed of generated code?


It seems to me that compile time is roughly exponential in the deficiency
of the generated code. So, to produce code that is 10% worse than optimal
takes twice as long as it does to produce code 20% less than optimal (if
your compiler is optimizer-bound). I suspect that programmer time required
to get the optimizer solid is also exponential.


So getting that last few percent requires a lot of resources.


Perhaps someone has collected some numbers on this? I am just making the
statement based on a gut feeling gotten from writing an optimizer.
[It varies all over the place. The Princeton/Bell Labs lcc compiler
is supposed to produce better code faster than GCC. Ken Thompson's Plan 9
compiler is supposed to be better still in both dimensions. -John]
--


Post a followup to this message

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