What's the deal with jump tables?

[sethp]

From what I can tell, we're down to just a pair of crossrefs back into Flash from RAM along the interrupt service path. One in _start_trap_rust:

and one in esp_hal_common::interrupt::riscv::vectored::handle_interrupts:

Both of these appear to be LLVM "jump tables": a mapping from some numerical value to a "next address" that compresses a whole bunch of repeated branch instructions into a single (offset) lookup + jump pair. Unfortunately, they're in .rodata, which is a good place for them in general, but we'd like to find a way to move these two somewhere else.

[sethp]

Seems like the three broad categories of options I've found are:

1. Turn off jump tables, either
   • globally (works with -C llvm-args=--max-jump-table-size=0, by way of rustc -C llvm-args='--help-list-hidden' | grep -- 'jump-table'),
   • locally (if we can find a way to set the no-jump-tables attribute for just the pair of functions in question here)
2. Ask LLVM to emit the jump tables in a way we can match on them in a linker script.
   • The linker's core verbs seem to be "map this input section to that output section" with everything else being set dressing for applying attributes (VMA, LMA, flags, etc.) to those sections; unfortunately, the jump tables seem to be mixed in to the .rodata section (along with, for example, our 256KB of background image data), so moving them wholesale seems to be a non-starter.
   • There's maybe a -ffunction-sections equivalent flag or similar that puts every function (and, it seems, every function's jump table) into its own section ; it's not clear to me that helps though, because the wildcard patterns only allow file-level exclusions (not section-level).
3. Writing a linker wrapper that peeks inside of all the object files, identifies the jump tables,
   • This is both complicated and slow (the latter, I suspect, being why linker scripts don't express that level of detail),
   • It's written in an unfortunate "why would you try to do this" tone (because I'm out of better ideas, of course), but there's still some useful context in https://stackoverflow.com/questions/29229281/gcc-linker-move-a-symbol-in-a-specified-section
   • There's some useful prior art for building and integrating a general linker wrapper here: https://github.com/knurling-rs/flip-link , but it's not useful much beyond that (it generates linker scripts rather than doing any ELF manipulation itself)

I asked for feedback on these ideas over at https://discourse.llvm.org/t/moving-jump-tables-sections-on-embedded-systems/70226

[sethp]

Oh, stray thought:

the wildcard patterns only allow file-level exclusions (not section-level).

Maybe we should be trying to get the jump tables into their own object file?

[sethp]

We should probably also give 1 a whack to get a better sense for how much (or little) it helps with our timing issues.