Re: [PATCH 1/6] x86: last-insn recording should be per-section

[Jan Beulich <jbeulich@suse.com>]

On 24.11.2023 10:03, Jan Beulich wrote:
> Otherwise intermediate section switches result in inconsistent behavior.
> Note, however, that intermediate sub-section switches will continue to
> result in inconsistent or even inappropriate behavior.
> 
> While there also add recording of state to s_insn().
> ---
> i386_cons_align() qualifying the recording on SEC_CODE seems suspicious,
> too: md_assemble() isn't constrained to emitting instructions to only
> text sections.
> 
> For sub-sections the best I can think of that we could do is warn when
> now_subseg is nonzero wherever we already warn based on last_insn state
> (and similarly bail rather than inserting any code). That would have
> not really intended effects in subsequent patches, though: We'd then
> generally suppress optimizations outside of subseg 0, and we'd also
> prefix plain old NOPs there (in i386_generate_nops()).

With some prereq tweaking of (at least) obj-elf.c, perhaps we could leverage
md_elf_section_change_hook() to deal with subsections. The prereq tweaking
would be needed as from my initial investigations I'm getting the impression
that not all section switches actually properly invoke that hook.

Jan