Wasmtime/Cranelift: carry "FP to SP offset" in exception data, and use it in stackwalk. (#11500)

* Wasmtime/Cranelift: carry "FP to SP offset" in exception data, and use it in stackwalk.

Currently

Wasmtime/Cranelift: carry "FP to SP offset" in exception data, and use it in stackwalk. (#11500)

* Wasmtime/Cranelift: carry "FP to SP offset" in exception data, and use it in stackwalk.

Currently Wasmtime unwinds stack frames to look for exception handlers
by walking frames one-by-one, following the FP chain as usual, and
assuming that *these frames are contiguous*: that is, that the SP in
any given frame (bottom of that frame) is immediately above the FP of
the next lower frame, plus the FP/return address pair (e.g. 16 bytes).
This allows us to get the SP for any given frame in addition to FP. We
need SP for two reasons:

- To look up dynamic context, to match Wasm tag instances for handlers
against the thrown tag;
- To actually set SP when we resume, if we do resume to a handler in
this frame.

This logic *almost but not quite* worked: I had forgotten that in our
tail-call ABI, we need to clean up incoming stack args in the
callee (because only the final callee in a parade of tail-calling
functions that reuse the same stack frame location knows how many args
it has, not the original caller). This implies that there is an
"incoming args area" *above* the FP/return address pair. Thus, frames
are not necessarily contiguous by the above definition.

In #11489 we see a case where a function of signature `(func)`
tail-calls one of `(func (param i32 i32 i32 i32 i32))`, which on
x86-64 (with four arg registers left for Wasm) is sufficient to create
incoming stack args, which then trips up the unwinder, reading a bogus
vmctx and segfaulting.

The most reasonable solution seems to be to embed the SP-to-FP offset
in the exception metadata itself, so from only the FP (which is
totally robust -- we rely on the FP chain for multiple kinds of
stack-walking) we can get the SP, allowing us to read dynamic context
and to reset SP during resume.

This PR does just that. Technically, in our ABI, the SP-to-FP offset
is constant for an entire function, but it was simpler in the
exception metadata to encode this per callsite instead (there is no
other notion of "per-function" data, only "per-callsite", so it would
be a separate binary search).

Fixes #11489.

prtest:full

* Review feedback.

show more ...