pcc wrote: > This PR doesn't entirely fix the problem. Another case that I'm seeing (and > which is more effectively prevented by the previous approach) is where GVN > PRE moves the zext behind a phi, and after subsequent optimization passes it > turns into a phi pointing to other phis which ultimately point to identical > (in the `isIdenticalTo` sense) zext instructions. We could address this in > CGP by trying harder to undo what GVN did: we can check that all instructions > directly/indirectly referenced by a phi are identical and then replace the > usages of the phi with a clone of one of the instructions.
Done in #141716 https://github.com/llvm/llvm-project/pull/141326 _______________________________________________ llvm-branch-commits mailing list llvm-branch-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
