LU-JOHN wrote: > It is not obvious to judge whether this patch is profitable. Can you share > some performance data on SPEC/LLVM test-suite/your motivating case?
This PR was inspired by code generated for Triton. A reduced example can be found in function testA in jump-threading-live-on-exit.ll. I've changed the block searching to go backward from uses to the def instead of forward from the def. This is more efficient since a forward search is not guaranteed to hit a use, but a backward search must hit the def. I've also added a limit to ensure that the search gives up if too many blocks are searched. I tested this with a 3-stage build of LLVM. This change increased the number of threaded branches from` 186726 to 197086, +10360. The number of blocks that could be threaded and required backward searches from the uses was 331442. The number of blocks searched from the uses ranged from 0 to 25 (the search limit). The average of number of blocks searched was 5.21. https://github.com/llvm/llvm-project/pull/135079 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
