projects / platform / upstream / llvm.git / commit
? search:
summary | shortlog | log | commit | commitdiff | tree
(parent: c3bc72c) | patch
llvm-reduce: Don't write out IR to score IR complexity
authorMatt Arsenault <Matthew.Arsenault@amd.com>
Sat, 8 Oct 2022 17:21:33 +0000 (10:21 -0700)
committerMatt Arsenault <arsenm2@gmail.com>
Thu, 13 Oct 2022 00:25:23 +0000 (17:25 -0700)
commitc23ac22f0ed9b76f127794534b3dfd5557389aed
tree31ab85100d9d3c640fc21ac62422797d8e672531tree | snapshot
parentc3bc72ccb23eb13aa6efba81adf68a0518eebc52commit | diff
llvm-reduce: Don't write out IR to score IR complexity

In a testcase I'm working on, the old write out and count IR lines
was taking about 200-300ms per iteration. This drops it out of the
profile.

This doesn't account for everything, but it doesn't seem to matter.
We should probably try to account for metadata and constantexpr tree
depths.
llvm/tools/llvm-reduce/ReducerWorkItem.cpp diff | blob | history
llvm/tools/llvm-reduce/ReducerWorkItem.h diff | blob | history
Domain: System / Toolchain;