Unreachable code in buffer.js and unsigned bitwise shifts.

[ChALkeR]

Thing to note here: A >>> B converts both arguments to unsigned int: http://www.ecma-international.org/ecma-262/6.0/#sec-unsigned-right-shift-operator.

1. buffer.js#L314 is not reachable. If argument start is a number that is less than 0, then start >>> 0 makes it a (large) positive number. Fixed.
2. buffer.js#L498 has offset < 0 check that is not reachable, because above it was either set to 0 or offset >>> 0 or length >>> 0.
3. buffer.js#L498 has length < 0 check that is reachable only through XXX legacy write(string, encoding, offset, length) - remove in v0.13 case, because else it was set to either 0, undefined, or this.length (which should be probably not less than 0 I guess). On a side note: sould that legacy path be removed in 4.0 or not? Btw, in the «legacy write(» case there is no type checking or casting for length.

Also, on many other lines, offset = offset >>> 0 and a succeeding checkInt(this, value, offset is done.

Note that -30064771000 >>> 0 === 72, which makes -30064771000 a valid offset input, which could be a bit unexpected. Also, checkInt doesn't check for offset (or offset + ext) to be not less than zero. Such check would be unreachable now (because offset is always cast to unsigned), but it could be usable once this is fixed.

var x = new Buffer(10).fill(0);
x.writeUInt32BE(0xdeadbeef, -30064771070);
x

<Buffer 00 00 de ad be ef 00 00 00 00>

Btw, with these negative values the same happens also when >> or | (or any other cast-to-int) is used: -30064771070 >> 0 === 2, -30064771070 | 0 === 0. All checks should be probably done before int casting.

Maybe it's worth splitting this into two issues — unreachable code and casting before checks.

@trevnorris

[thefourtheye]

Glad that you brought this. I was mentioning something like this to trevnorris in a comment.

I say let's not do coercing and validate the values as they are and throw error if they are invalid. It would keep the code straightforward and easily maintainable.

[trevnorris]

As I mentioned in the referenced issue, the conditional and check should be above the value coercion. In the case of .write*(). That is a bug, as it does not follow current documentation.

We shouldn't throw any more exceptions ATM so that the patch can land in v4. Otherwise it'll have to wait for v5. And there are legitimate bugs here that should be addressed which should propagate to v4. Any other changes should be referenced against documentation. Though I know in several cases where it's just not documented.

[trevnorris]

Here's one potential solution for https://github.com/nodejs/node/blob/v3.3.0/lib/buffer.js#L314:
Replace the above start = start >>> 0; with:

if (start < 0) start = 0;
if (start > 0x7fffffff) start = 0x7fffffff;

The reason for this is because in src/node_internal.h the ParseArrayIndex() function is used to extract the values. This uses Int32Value(). Though ironically we store that into a size_t.

As you can see, this mess goes deeper. We should do a full assessment of the code and how all these cases are handled. I'm the one responsible for switching to use the >>> on incoming arguments for .write*() functions because by skipping argument checks it allowed many of these operations to be done in a couple nanoseconds. Which kicks the crap out of DataView. Is also very useful for node modules that work with image processing, video, etc. to have that kind of performance. Since doing that many writes on that scale can quickly kill performance.

Addressing throwing more often, any place we currently don't throw, we won't start for the time being. First we'll have to make an assessment of how much community breakage it will introduce. Its APIs are in massive use and possibly breaking them for anything short of a security issue would be difficult.

[ChALkeR]

This is fixed only partially.