Il-Capitano wrote:

I made the changes needed to determine whether we need to emit any CFI in the 
first place, or not. Saving LR to the stack is also handled now.

This patch seems to be the first use of `.cfi_rel_offset` in the MIR pipeline, 
so https://github.com/llvm/llvm-project/pull/239555 is also needed to actually 
compile something that makes use of that CFI directive.

The final remaining issue I think is that the current approach walks the 
machine function for each of its outlining candidates, so the current algorithm 
is quadratic in the function size, which is not ideal. This can be fixed by 
only calculating the necessary block information once, for the first outlining 
candidate of the function, and reusing that same information in the future. I'm 
not sure what is the best way to do this. If there's precedent for this kind of 
caching in other MIR passes, I'd be happy to hear about it. In the mean time 
I'll try to come up with something.

https://github.com/llvm/llvm-project/pull/236836
_______________________________________________
llvm-branch-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits

Reply via email to