1442534 - ARM64: Improve code generation for atomics by avoiding fences

Status: NEW

(Core :: JavaScript: WebAssembly, enhancement, P3)

Description

[Lars T Hansen [:lth]]

Currently we generate an ARM-like fence + op + fence instruction sequence for ARM64 seq_cst atomics; this is safe, but we can do better.  Specifically, the canonical reference [1] has optimized sequences that avoid fences by using only the acquire and release operations, even for seq_cst operation.

[1] http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html

Comment 1

[Lars T Hansen [:lth]]

The optimized atomics are now about to become legal following an update to the memory model: https://github.com/tc39/ecma262/pull/1511. The executive summary is that the old memory model disallowed the use of release-acquire atomic ops for seq_cst operation in JS; the new model is slightly weaker and allows them, and the change is motivated by the desire to allow them on ARM64. (Apparently Chrome already uses weaker atomics here.)

Comment 2

[Lars T Hansen [:lth]]

ARMv8.1 has CAS primitives, which will help.

Also see https://github.com/WebAssembly/threads/blob/master/papers/oopsla2019.pdf for up-to-date guidance.

Comment 3

[Lars T Hansen [:lth]]

Note that if we update the JIT here we must also update the generated JS atomic operations, see jit/GenerateAtomicOperations.py.