On Wed, Jul 10, 2024 at 8:36 AM Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx> wrote: > > But, tbh, I'd make this stricter. I'd dictate that within a subprogram > all no_csr patterns *should* end with the same stack offset and I > would check for that (so what I mentioned before that instead of min > or max we need assignment and check for equality if we already > assigned it once). No. Compiler has full right to use different slots in nocsr region for spills/fills. At every nocsr callsite there could be a different number of live registers too.
