Bug 150760 – Wrong value recovery for DFG try/catch with a getter that throws during an IC miss

• Status: RESOLVED
• Resolution: FIXED
• Component: JavaScriptCore

[Saam Barati]

This program breaks in the DFG. Probably because we're assigning the base and the result to the same register and the IC fails.
```
function assert(b) {
    if (!b)
        throw new Error("bad value")
}
noInline(assert);

let oThrow = {
    x: 20,
    y: 40,
    z: 50,
    get f() { throw new Error("Hello World!"); }
};

let o1 = {
    x: 20,
    f: 40
};

let o2 = {
    x: 20,
    y: 50,
    get f() { return 20; }
};

function foo(f) {
    let o = f();
    try {
        o = o.f;
    } catch(e) {
        print(o); // Make this not undefined.
        assert(o === oThrow);
    }
}
noInline(foo);

let i;
let flag = false;
function f() {
    if (flag)
        return oThrow;
    if (i % 2)
        return o1;
    return o2;
}
noInline(f);
for (i = 0; i < 10000; i++) {
    foo(f);
}
flag = true;
foo(f);
```
I discovered this while solving this issue in FTL try/catch.

[Saam Barati]

What I think is happening.
The IC misses, but the operationGetByIdOptimize stores the result
of the call into the result register. Even though the C call throws,
we store the result. And then, we try to recover the base from
the register we stored the result into because the result and base
are the same register. This is a big bag of fail.

[Saam Barati]

(In reply to comment #1)
> What I think is happening.
> The IC misses, but the operationGetByIdOptimize stores the result
> of the call into the result register. Even though the C call throws,
> we store the result. And then, we try to recover the base from
> the register we stored the result into because the result and base
> are the same register. This is a big bag of fail.

I've actually realized that the problem is deeper than this.
I do think the situation I'm describing is a problem, but
I've realized that the result of the call in this example
is looking dead to DFG OSR exit when we use PhantomLocal
as the liveness preservation mechanism. I'm going
to switch the code back to using Flush so that I can
land FTL try/catch, and then I will investigate this problem
further.

[Saam Barati]

(In reply to comment #2)
> (In reply to comment #1)
> > What I think is happening.
> > The IC misses, but the operationGetByIdOptimize stores the result
> > of the call into the result register. Even though the C call throws,
> > we store the result. And then, we try to recover the base from
> > the register we stored the result into because the result and base
> > are the same register. This is a big bag of fail.
> 
> I've actually realized that the problem is deeper than this.
> I do think the situation I'm describing is a problem, but
> I've realized that the result of the call in this example
> is looking dead to DFG OSR exit when we use PhantomLocal
> as the liveness preservation mechanism. I'm going
> to switch the code back to using Flush so that I can
> land FTL try/catch, and then I will investigate this problem
> further.

Let me elaborate:
the result of the call looks dead after we do flushRegisters().

[Saam Barati]

Created attachment 264646 [details]
patch

[Geoffrey Garen]

Comment on attachment 264646 [details]
patch

r=me

[WebKit Commit Bot]

Attachment 264646 [details] did not pass style-queue:


ERROR: Source/JavaScriptCore/dfg/DFGLiveCatchVariablePreservationPhase.cpp:133:  Place brace on its own line for function definitions.  [whitespace/braces] [4]
Total errors found: 1 in 3 files


If any of these errors are false positives, please file a bug against check-webkit-style.

[Saam Barati]

landed in:
http://trac.webkit.org/changeset/191930