AsmJs/WAsm calling convention issues on xplat

[Penguinwizzard]

When attempting to pass a floating-point argument to an asmjs/wasm function on xplat, the caller and callee don't agree on the calling convention. This causes correctness issues, as we end up often using the wrong value. This occurs in JITted code.

Repro:
Remove these lines from test/runtest.py (link):

not_compile_flags = set(['-simdjs']) \  
    if sys.platform != 'win32' else None

Run runtests.py -d AsmJSFloat.

Note: There's a separate bug filed related to removing the lines hiding this issue from runtest.py at #2412.

Repro file:
Here's a separate file by @Cellule that reproduces the issue with -maic:0

function AsmModuleDouble() {
        "use asm";

        function add(x) {
                    x = +x;
                    return +sub(1.5);
                }
        function sub(x) {
                    x = +x;
                    return +(x);
                }
        return {
                    add: add
                };
}

var asmModuleDouble = AsmModuleDouble();     // produces AOT-compiled version
print(asmModuleDouble.add(2, 1))

[jianchun]

@Penguinwizzard Indeed, didn't get to complete Simd when I worked on xplat JIT. That's why Simd tests were disabled on xplat by that line.

[Cellule]

From what I was able to gather this happens because LowererMDArch::LowerCallArgs calls in LowererMDArch::GetArgSlotOpnd which tries to use xmm registers for float argument when we are supposed to put them on the stack on xplat

[Cellule]

@jianchun The problem is almost all asm.js test have that flag even though simd is not used (I am not entirely sure the reason why, but I think there's a good one).

[Penguinwizzard]

@jianchun If Simd doesn't work on xplat, should we disable building Simd/not recognize the flag/error out on use on xplat? It looks like a lot of tests just fail with "SIMD is undefined", which seems fine - those tests can just be excluded.

[jianchun]

@Penguinwizzard @Cellule At the time I felt Simd was close to completion, but I was not exactly sure of the calling conventions. Left Simd code in build so it would not diverge further (Simd code kept introducing build failures at the time).

[MikeHolman]

Is wrong calling convention still an issue?

[Cellule]

@MikeHolman AFAIK it hasn't been fixed yet, but maybe @obastemur has more info on this

[MikeHolman]

Then can we increase the priority on this issue? It is essentially blocking our wasm fuzzing and generally is a pretty high impact issue for wasm/asm.js code.

[obastemur]

#2716 is merged and should fix this issue.

[Penguinwizzard]

Problems still occur; remove exclude_xplat on most wasm test cases to repro.

[obastemur]

@Penguinwizzard the problem shouldn't be calling conventions. At least it was fixed for asm.js and others. About WASM, we have couple of disable stuff (SIMD) for xplat. Maybe we should close this issue and open a new one : Enable WASM on xplat ?

[Cellule]

This issue is also hitting in asm.js not just wasm.
One of the problem is how we make calls in jitted code which is shared by both asm.js and wasm

[Cellule]

Closing in favor of #3561