[PATCH] D11686: [RegAlloc] Trace sibling copies when checking for rematerializability during spill weight calculation (PR24139)

[Quentin Colombet]

qcolombet added a comment.

Hi Rob,

I would have preferred that we use the LiveRangeEdit information to test whether or not the value is rematerializable. Indeed, we need to look through split only if we did split, i.e., we used LRE.

That being said, your fix is fine because the LRE rematerialization test is more general and it is likely that the spill cost resulting from it will be completely off.

Ultimately, I think it should be the spiller that should give the weight to spill a live range… but this is a clean-up for another day!

Anyway, thanks for catching and fixing it.

Could you add a test case using the related PR?

Cheers,
-Quentin


http://reviews.llvm.org/D11686