[llvm] [GVN] Permit load PRE to happen in more cases (PR #79324)

David Sherwood via llvm-commits llvm-commits at lists.llvm.org
Tue Feb 20 00:31:20 PST 2024


david-arm wrote:

> Haven't looked at the implementation yet, but I can already tell you that this causes [large compile-time regressions](https://llvm-compile-time-tracker.com/compare.php?from=6c1dbd5359c4336d03b11faeaea8459b421f2c5c&to=fbf36d485f0f808cdfd322658c5c52edb26ec6a6&stat=instructions:u).

Thanks for this useful info @nikic!

I can have a look at limiting the number of instructions we walk backwards to see if that helps - we can probably get away with a walk limit to 20 or 25 instructions. I could also limit the number of predecessors we're willing to consider. However, I guess there will still be some cost of doing more analysis to get the required heuristics.

https://github.com/llvm/llvm-project/pull/79324


More information about the llvm-commits mailing list