Re: 5.1b cc optimizer problem?
From: Andrew Duane USG (duane_at_wasted.zk3.dec.com)
Date: 08/21/03
- Next message: nurp: "Re: Xserver problems"
- Previous message: Joerg-Uwe Tessmer: "Re: Problems with executing programs"
- In reply to: none: "5.1b cc optimizer problem?"
- Next in thread: Jerry Feldman: "Re: 5.1b cc optimizer problem?"
- Reply: Jerry Feldman: "Re: 5.1b cc optimizer problem?"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Date: Thu, 21 Aug 2003 12:55:23 GMT
I'll let someone from the compiler group jump in with more details, but
the compiler hasn't changed a whole lot in 5.1B, and I haven't noticed
anything like this from it, now or ever before.
One of my main jobs here is to make sure the Tru64 kernel compiles correctly
with the compiler, so I keep careful tabs on code generation, optimization,
all sorts of issues. In the last 10 years, I have had at least a dozen cases
where I was sure the compiler was doing something wrong.
I was proven wrong every time.
Optimization can uncover some source code issues that lay dormant before.
You would be amazed at what some people think they can get away with.
Even code that has worked for years and many releases can suddenly break
when a "trivial" printf is added, since register allocation now changes,
or an old return value, previously unimportant, is now critical.
Dollars to donuts this is a source code issue, not a compiler issue.
In article <vk829rmd2ioud6@corp.supernews.com> (Tim Mooney) writes:
>I use the Tru64 cc compiler exclusively for building open source
>software, even stuff that isn't performance-sensitive (where cc
>really beats a compiler like gcc, which I'm not knocking at all).
>
>Since upgrading to 5.1b from 5.1, I've noted inexplicable application
>crashes with several locally-compiled apps, and although it would be
>difficult to prove, it seems to be an optimizer problem in the compiler.
>
>Take for example pine, from UW (ftp://ftp.cac.washington.edu/pine). If
>I run version 4.53 compiled under 5.1 on my 5.1b box, it works fine. If
>I recompile 4.53 in the default way (build osf) on my 5.1b box, it segfaults
>in libc's strlen(), called from the pine function format_keymenu(). If I
>modify the necessary makefiles to remove the `-O2' and rebuild, it works
>fine.
>
>Same is true for the latest version of pine, 4.56. I've seen otherwise
>unexplainable crashes from other apps too, but most of them are larger and
>harder to be certain where the problem is.
>
>I realize that turning on optimization in the compiler can in some cases
>trigger subtle bugs in the code of an app you're compiling, but my experience
>and instincts tell me that isn't what's happening here.
- Next message: nurp: "Re: Xserver problems"
- Previous message: Joerg-Uwe Tessmer: "Re: Problems with executing programs"
- In reply to: none: "5.1b cc optimizer problem?"
- Next in thread: Jerry Feldman: "Re: 5.1b cc optimizer problem?"
- Reply: Jerry Feldman: "Re: 5.1b cc optimizer problem?"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Relevant Pages
|
