Part of it is due to equivalence set bitmaps becoming huge (and their
inefficient computation). equivalences are only used when trying to
simplify comparisons (but during iteration). So the only used
equivalence bitmaps are for operands of those. equivalences are
generated only by asserts and SSA name copies, so they can be computed
up-front and shouldn't change (apart from handling of non-executable
edges in PHI nodes) during iteration. At least equivalence processing
can be ignored for SSA name chains not leading up to a condition operand.
Is there a test case for this PR? I've seen this behavior in other PRs
but I can't find them. It'd be nice to have this bug linked to such
cases (and to the "GCC is slow" meta-bug).
I filed this bug when working on PR56113.