single-pass regalloc: Corruption (?) with throw.wast

[alexcrichton]

This currently fails on main:

$ cargo run wast /home/alex/code/wasmtime/tests/spec_testsuite/proposals/wasm-3.0/throw.wast --target x86_64-unknown-linux-gnu -Oregalloc-algorithm=single-pass -Osignals-based-traps=n -Wexceptions=y
...
Error: failed to run script file '/home/alex/code/wasmtime/tests/spec_testsuite/proposals/wasm-3.0/throw.wast'

Caused by:
    0: failed directive on /home/alex/code/wasmtime/tests/spec_testsuite/proposals/wasm-3.0/throw.wast:49
    1: error while executing at wasm backtrace:
           0:    0x120 - <unknown>!<wasm function 7>
    2: wasm trap: wasm `unreachable` instruction executed

Removing -Osignals-based-traps=n yields a panic:

$ cargo run wast /home/alex/code/wasmtime/tests/spec_testsuite/proposals/wasm-3.0/throw.wast --target x86_64-unknown-linux-gnu -Oregalloc-algorithm=single-pass -Wexceptions=y
...

thread 'main' panicked at /home/alex/code/wasmtime/crates/unwinder/src/exception_table.rs:168:25:
Wasmtime exception unwind info only supports dynamic contexts on the stack
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

[alexcrichton]

cc #11533 and @d-sonuga

[d-sonuga]

I see that the issue happens because of the following scenario:

branch inst: def v0 (fixed: p0). succs: [block1(v0)]

The single-pass relies on moving branch args (v0 in the case above) to expected spillslots of the successor blocks before branch instructions. That's not possible here because the vreg used as the branch arg is defined in the branch instruction.

Before I change that to account for this scenario, I need to know, can these vregs have any of the possible kinds of constraints?
I mean, is it possible for v0 in the above scenario to also be an any-reg constraint, or stack-only or no constraint?

[cfallin]

Ah, yes, indeed, the try_call defines its return values and is also a terminator. The retvals' constraints will either be fixed-reg or any; the latter hppens when there are more return values than the ABI puts in registers. In practice we shouldn't see stack or reg, so if it's much more work to support those on terminators too, then no need at the moment.

[cfallin]

(I'll keep this issue open until we do a release in RA2 and pull the new version over, but thanks to Demilade for addressing the actual issue in the regalloc!)