On Tue, 21 Jul 2009, Krzysztof Oledzki wrote:

Perfect.

And interestingly, the "fno-strict-overflow" kernel is actually much closer to the "fnone" kernel than to the "fwrapv" one. I have some silly scripts based on 'objdump -d' plus a lot of stupid sed scripting to remove the trivial differences due to instruction addresses, and then doing a 'diff -u' between the munged disassembly of the kernels gives me:

[torvalds@nehalem fno-strict-overflow]$ wc -l fnone-to-fno-strict-overflow
fwrapv-to-fno-strict-overflow 39309 fnone-to-fno-strict-overflow 91423 fwrapv-to-fno-strict-overflow 130732 total

ie the diff from the kernel with no flags is less than twice the size of the diff from fwrapv.

Still - it's almost 40kB of diffs of disassembly, so I'm not going to guarantee that I can make any sense of it and find the compiler problem in it. But I'll try. And send you test-patches to see if I can pinpoint the code that causes the problem.

Thank you.

Great. This is all about as perfect as could be asked for. Now it's just a question of trying to find the right code generation difference...

Linus