Re: x86 bootstrap failure

[kenner at vlsi1 dot ultra dot nyu dot edu (Richard Kenner)]

    GCC seems to be more sensative to individual characteristics of
    architectures.  It may be good that GCC is starting to take advantage
    of individual architecture features.  Howevr, it is making it much
    more difficult to test patches because successful bootstrap no longer
    is indicative that the patch is safe.

    I do not know what we can do to improve the stability of the CVS
    sources.  Requiring bootstrap tests on many architectures is impractical
    for a number of reasons.

This is not a new problem and indeed I think things have been getting
*better* in this area, with the addition of things like RTL checking and
other validity tests.

Perhaps the only practical approach would be *unit testing*, where each
optimzer pass is fed lots of different types of synthetic RTL as a test
instead of relying on actual machines.

This is, of course, very difficult to do, but I think is the only complete
way to deal with the problem.

Luckily, we have people testing enough architectures often enough that things
don't stay broken for long, but the case we had here where two completely
unrelated patches installed by two different people within a 24-hour period
each broke a different machine's bootstrap and didn't affect the one in which
they were tested, is indeed a confusing situation, especially since there's
some evidence there's also a *third* unrelated problem (the IA32 cfgbuild.c
comparison failure).