Update StarlingMonkey

[andreiltd]

I will keep it as a draft until we have StarlingMonkey release.

[andreiltd]

The latest failures looks a bit scary and they only happens for weval build, but so far I was able to rule out that this is related to partial evaluation by just patching weval and skipping evaluation step -- I got the same errors.

The simplest test (integers) failure looks like import dispatch is just wrong for weval build, which may suggest some codegen assumptions in the splicer has changed -- the wasi-sdk jump is pretty significant: from v20 when weval was last used to v30 now, that is llvm v16 to v21. I will keep digging.

[andreiltd]

Hm... OK, interesting -- it looks like splicer only works on the module that was optimized with wasm-opt and we skip this step for weval engine because we cannot weval wasm-opted code. I verified this by adding WASM_OPT=OFF to non-weval engine build and I got the same test failures.

[andreiltd]

So it looks like indeed wasm-opt is producing the codegen that splicer expects for coreabi_get_import and skipping wasm-opt break the assumptions:

Before wasm-opt:

global.get 1         -> load Runtime.cx via global
local.set 0            -> this overwrites idx parameter,
...
global.get 4           -> this gives 1
i32.const 3020       -> this gives 1+3020=3021
i32.add

After wasm-opt -O3:

i32.const 2570476    -> constant folded memory address (no global.get 1, no local.set 0)
i32.load                     -> direct load of cx
i32.const 3021           -> the split global.get 1 + i32.const 3020 is folded into single const

So the fix is to handle both: store idx so that it doesn't get overwritten and then check if global index is referenced and include that in final base calculation:

ab1d874

It seems a bit fragile, but we should soon switch to wit-dylib world so we don't have to maintain that code for very long (hopefully).

This also suggests there are some further optimization opportunities for weval build: we could run the wasm-opt after wevaling. The main drawback is that we would have to run wasm-opt for each componentize command which will make the command time much longer.

[andreiltd]

Another fix was to increase the RETRY count in tests because macos runners were crashing under the added weval load.

[andreiltd]

I think next steps to integrate the changes should be:

• remove libspidermonkey release - the release didn't pick up the weval artifacts so SM is building spidermonkey from source every time,
• remove the libspidermonkey tag
• merge SM release PR
• bump SM submodule
• review and merge this PR

cc: @tschneidereit @vados-cosmonic