Strange code generation bugs in egcs-20000103 and egcs-20000110 (x86-linux)

[Karl Günter Wünsch <kgw at cswgmbh dot de>]

Hello,

I have seen the strangest of behaviours of these two snapshots yet. Let
me recap what happened:
There is a long standing code generation bug which only shows up when
not optimizing. It doesn't show up when optimizing with -O2. This bug
was reported by me mid october 1999, sorry no full bug report available
since I can't produce a decent sized example of it and the bug only
shows up very late within a quite large program of mine which I can't
explain to the extent at the moment that anyone else could understand
where the bug is really located. The problem of this bug is that a
register int variable that is used within several expressions inside a
loop (amongst which is a multiplication) - but is never assigned a value
within the loop itself - ends up being multiplied by the exact value of
the multiplication for every time the loop is executed, so this seems to
be a reload error if you ask me...

I am checking every snapshot available for this bug, because there is a
lot going on with respect to code generation and my hopes are that this
bug will disappear because someone else was able to reproduce this. Now
the time seems to have come that the buggy insn's were touched so I
better speak up now, so that the people involved know that something has
crept up.

Well to cut the long story short:
 egcs-20000103: Bug is gone! But now my whole program is running like
shambles. Almost none of the calculated floating point values are
correct.
 egcs-20000110:  Bug is here again! But the rest of the Program behaves
as it should. All floating point values are correct.

So in my opinion something in the bug-critical section of the
instruction generation has happened. I am very eager to help in this
situation and am able to give almost (except the complete project code
because of said complexity) every single intermediate output necessary
to track down this problem.

Please respont off list, since I am currently not subscribed with this
e-mail address.

regards
Karl Günter Wünsch