No Subject

[David Edelsohn]

>>>>> Mark Mitchell writes:

ark> Good, I did the same thing, with the same result.  The only other
Mark> changes were the AVOID_CCMODE_COPIES bit, and the `return 0' for a
Mark> non-void function that was previously not returning a value.  So,
Mark> perhaps something is amiss with the AVOID_CCMODE_COPIES stuff.  But I
Mark> don't see it.  David (Edelsohn), did you try applying the Aug 25
Mark> patch, but also Richard Henderson's 9/9 patch?  That patch fixed one
Mark> other bug in this stuff, which you might be stepping on.  Of course,
Mark> that wouldn't explain why you can't bootstrap with what's in CVS today.

	I went back and applied all of the intervening non-check_dbra_loop()
changes up through the present and bootstrapped again.  This succeeds, so
the culprit seems to be check_dbra_loop().  I am a bit surprised that this
would have an effect because the comments for the function focus on
REG_NONNEG useful for decrement_and_branch_until_zero which PowerPC does
not define.  I guess it is doing other transformations as well, like the
loop reversal, which are having an affect.

	Thanks for helping me disambiguate which changes were independent
so that I could test your fixes independently.  Cygnus *should* have
access to PowerPC systems to track down what is broken with those set of
changes.

Thanks, David