OK, I must have misunderstood you then. Sorry ...
It's ok :)
I've managed to reduce the problem to a simple testcase, which is appended below. It fails on RedHat gcc-2.96-69 (as of 12/2000) ...
Culprit *appears* to be the global common subexpression elimination, although this is not quite definite. I haven't tracked the problem down yet.
Are you thinking of doing more investigation on that?
However, when switching off GCSE (-fno-gcse), correct code is produced. I'm not sure I'd like to do a complete Wine build without GCSE though; this might significantly reduce code quality ...
Oh the runtime behavior should not change that much but it's still non-negligeable. I like the "disable-gcse-if-you-probe-that-gcc-version" approach.
I'll write a configure check to switch off gcse, using the testcase below. We should probably contact RedHat (I think Jacub Jelinek handles 2.96 bugs ...) to confirm that GCSE is really at fault.
If you want to post it with the test program, go ahead :) I never posted to gcc bug list so I'm not sure how much attention would be given to the bug. I think the mailing list should be cc'ed on the postings so everybody can follow up how it turns out.
Ciao! Francois
-
François Jacques