On 3/13/2012 11:37 PM, Marijn Haverbeke wrote:
Thanks for digging into this so deeply. This probably also explains
why the non-optimized build currently fails on Windows (#959) -- it
simply generates binaries that are too big.
Yeah, probably the same bug.
The prospects are looking a little bleak, I guess. Debugging linkers
sounds like it'll get hairy.
In this case it turned out to be not so bad once I found something that
was actually hitting a power-of-two. 65536 relocs. Hello!
While the nominal COFF section header field is a u16 for storing reloc
counts -- and overflow + corruption here is what was going wrong -- it
turns out PE/COFF has defined behavior for handling overflow: a header
bit and an extra synthetic reloc. Microsoft tools and GNU tools already
implement this and aren't bitten by it. LLVM-MC doesn't. Or didn't. I
implemented it (BFD shows how) and it all works fine. The change is here:
https://github.com/brson/llvm/commit/4119fda4c6c4df372bd6edbbb08d48c7941577b5#diff-0
With this change your code bootstraps on my machine, bots are cycling
currently. I've upgraded our LLVM to a version carrying this patch, so
if they're in a good mood, you should be good go now.
CC'ing Bigcheese from LLVM-land. I imagine he'll want this.
-Graydon
_______________________________________________
Rust-dev mailing list
[email protected]
https://mail.mozilla.org/listinfo/rust-dev
