erichkeane wrote:

I realize I suggested this location for this, but the additional complexity 
that causes is problematic.  It would be wonderful to have done this in 
LowerToLLVM, as that minimizes the impact here, but all the bookkeeping and 
rewrites that ended up happening here are unfortunate.

I wonder if this would be better in cir-lowering-prepare?  It wouldn't save us 
as much memory (as we'd end up creating another CIR type), but that seems 
acceptable to me.  

While we should still do the >8 elements check from Classic Codegen, I wonder 
if that is a pretty terrible heuristic for CIR.  Since we end up having to 
'copy' the array (though perhaps we can 'steal' the old storage of the array 
when creating the new one?), I wonder if it should be percentage-based.  But 
thats just 'shower thoughts' for the future, our initial patch should still do 
the 8 elements check.

https://github.com/llvm/llvm-project/pull/205918
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to