The functions addSaturate() and subSaturate() are defined on the 8x16 and 16x8
integer SIMD types only.
Theee are no 32x4 variants defined in SIMD.js because current hardware doesn't
support it directly.
This is the right shift function that the SIMD.js spec requires. The old
shiftRightArithmeticByScalar() and shiftRightLogicalByScalar() functions will go
away.
These functions perform an arithmetic shift for signed types and a logical
shift for unsigned types.
Add support to Odin and Ion too, at least for the Int32x4 variant.
Fix a bug in ShiftRightArithmetic when the underlying Elem type is unsigned.
Need to cast to a signed type to so it sign-extends to int when shifting.
Implement both logical and arithmetic right shift implementations for both
signed and unsigned integers. This is needed to test our current implementation
where both signed and unsigned SIMD types have the two
shiftRightArithmeticByScalar() and shiftRightLogicalByScalar() functions.
Soon, shiftRightArithmeticByScalar() and shiftRightLogicalByScalar() will be
replaced by a single shiftRightByScalar() function whose behavior is dependent
on the signedness of the underlying type.
- Add new types UInt8x16, UInt16x8, UInt32x4 with all the boilerplate.
- Add corresponding conversion and bitcast functions to existing types.
- Add tests/ecma_7/SIMD tests for all new functions. Mostly boilerplate.
- Add full type coverage in ToSource.js. Fix existing SIMD types that were
broken.
- Use shared test vectors for all the 32x4 integer binary-op test cases.
- Fix a bug in the 32-bit multiplication reference implementation for Int32x4
and Uint32x4. A simple 'a*b' double multiplication loses precision to rounding
for some inputs.
The lists TypeDescriptorMethods and TypedObjectMethods are identical for all
SIMD types, so save a bit of memory and source code by sharing a single list.
In the current SIMD.js spec, value conversions are only defined between types
with the same number of lanes:
Float32x4.fromInt32x4()
Float32x4.fromUInt32x4()
Int32x4.fromFloat32x4()
UInt32x4.fromFloat32x4()
Remove existing conversion operators between vectors with different numbers of
lanes:
Int32x4.fromFloat64x2()
Float32x4.fromFloat64x2()
Float64x2.fromInt32x4()
Float64x2.fromFloat32x4()
Add a static assertion to FuncConvert.
clang-cl attempts to emulate MSVC's handling of __VA_ARGS__, but doesn't
quite get it right:
https://llvm.org/bugs/show_bug.cgi?id=25875
As a result of this, compiling files that #include MacroAssembler.h with
clang-cl result in fallbacks to MSVC. Since falling back to MSVC is
non-ideal (and also causes problems around things like linking function
template instantiations), let's disable MacroAssembler.h's macro magic
for the time being. Ideally, the problem will get fixed upstream
soon (even though it looks somewhat complicated); in the meantime,
fixing this issue lets forward progress be made when compiling Gecko
with clang-cl.
