[llvm-branch-commits] [clang] [llvm] [GVN] Simple GVN-based hoisting of scalars (PR #210337)

[Nikita Popov via llvm-branch-commits]

nikic wrote:

If I remember correctly, SimplifyCFG hoisting was capable of handling the motivating case, but enabling the hoisting earlier led to a regression on some other SPEC benchmark because something was unrolled too early. So the "GVN" hoisting was not actually useful, it was only useful insofar as it ran in a slightly different pipeline position. Is this still the case, or am I thinking about a different issue?

https://github.com/llvm/llvm-project/pull/210337