Several bugs in llvm-like branch

[lars-t-hansen]

@syg, I think what's on gh-pages seems to be built from changes you have not pushed, so I'm filing these bugs here until current code is in the repo:

• Atomics.store needs to return v.
• The call to GetModifySetValueInBuffer in AtomicReadModifyWrite needs to pass v, not value.
• The rewrite of Atomics.compareExchange completely discarded the conversion of the expectedValue argument to a byte list in the manner that replacementValue will be subjected to, this needs to be reinstated. It needs to be done once and at the proper place in the argument validation/conversion order, which is why compareExchange was open-coded to begin with.

[syg]

@lars-t-hansen Regarding the last bullet, the intention was that conversion was to be done as part of the compareExchange semantic function (it's really a closure, unlike all the other semantic functions that are passed to AtomicReadModifyWrite). I'll reinstate the manual conversion and make the compareExchange close over the converted value.

[syg]

@lars-t-hansen Please close the issue if you think of the fix for your last bullet in the linked commit is satisfactory.

[lars-t-hansen]

@syg, the fix for compareExchange is not satisfactory because now compareExchange has different observable argument processing order than the other functions. Consider the original spec of this function:

1. Let _buffer_ be ? ValidateSharedIntegerTypedArray(_typedArray_).
1. Let _i_ be ? ValidateAtomicAccess(_typedArray_, _index_).
1. Let _expected_ be ? ToInteger(_expectedValue_).
1. Let _replacement_ be ? ToInteger(_replacedValue_).
1. Let _arrayTypeName_ be _typedArray_.[[TypedArrayName]].
1. Let _convOp_ be the conversion operation specified in Table 49 for _arrayTypeName_.
1. Let _x_ be _convOp_ (_expected_).
1. Let _elementSize_ be the Number value of the Element Size value specified in Table 49 for _arrayTypeName_.
1. Let _elementType_ be the String value of the Element Type value in Table 49 for _arrayTypeName_.
1. Let _offset_ be _typedArray_.[[ByteOffset]].
1. Let _indexedPosition_ be (_i_ × _elementSize_) + _offset_.
1. With atomic access to shared memory, do:
    1. Let _r_ be GetValueFromBuffer(_buffer_, _indexedPosition_, _elementType_).
    1. If _r_ is the same as _x_ then:
        1. Perform SetValueInBuffer(_buffer_, _indexedPosition_, _elementType_, _replacement_).
    1. Return _r_.

Here, if the index is not convertable to number or is out of range that will be signaled in step 2 before we convert any of the other arguments. The old spec for this algorithm was written the way it was written to be compatible in its observable side effect order with the other algorithms in the proposal, all of which were designed to be compatible in behavior with existing routines in ES, sadly at a cost in spec elegance for compareExchange.

[syg]

@lars-t-hansen Okay, I basically resurrected the old version.

[lars-t-hansen]

Looks good.