Re: Tonight's snapshot

[craig at jcb-sc dot com]

>In a previous message either you or someone else said that the X86 mis-compile
>of texinfo may slide. My $0.02 says lets not let it slide. there is nothing
>in texinfo that is so radically different to what other programs do that we
>should let it go. If texinfo miscompiles, will TeX itself? X11? GIMP? Where
>exactly does the failure domain stop? I really hope that we can fix whatever
>the problem is, as I believe that whatever the bug is will affect a lot more
>than just texinfo.

texinfo is distributed with gcc itself, so that makes it sort of "special",
in that it's really embarrassing when a compiler won't, in essence,
compile itself, on a very popular target.

That being said, it's true that, in general, any bug we know that causes
mis-compilation of one piece of code could affect any other.

I assume the gcc people who've investigated the bug's failure mode have
concluded that it's unlikely to impact lots of other packages -- especially
assuming those other packages have been tested against the prerelease version
of the compiler.

Also, there are *always* bugs in a compiler -- in fact, many of them, in
a compiler built the way gcc is built (which is the way most compilers
are built these days, I believe -- gcc being on the more-robust side of
what I've personally seen).

>In a sort of related vein, this shows a deficiency (possibly an unavoidable
>one) in the way regression tests are done. While the current test suite is
>excelent for testing specific compiler features and bugs in a one-by-one
>fashion, we have little or no practical way of testing the relibaility of
>the compiler in a larger application where the environment is more complex.
>The only real way to do such testing is to compile larger applications and
>run THEIR test cases. Lets consider the compiling of texinfo just such
>a case, and deep the compiler "broken" if it mis-compiles it.

I believe that's essentially what we ask the larger community of gcc
developers, testers, etc. to do during the prerelease period, even
earlier if they are able.

        tq vm, (burley)