Add wasm_memory_discard compatibility flag

[guybedford]

Adds a wasm_memory_discard compatibility flag gating the experimental WebAssembly memory.discard sub-proposal, exposing WebAssembly.Memory.prototype.discard(byteOffset, byteLength) and the memory.discard opcode per-worker.

Wired through V8's per-context conditional-feature mechanism: a per-isolate enabled callback (SetWasmMemoryDiscardEnabledCallback) reports the flag state, consulted both when installing the JS API (InstallConditionalFeatures) and when compiling wasm modules using the opcode. The flag stays set for the isolate lifetime since V8 re-queries it at compile time.

Floats the V8 prototype as patch 0039, a port of CL 4300676 updated to the memory.discard semantics prototyped in SpiderMonkey, behind --experimental-wasm-memory-discard.

Includes a wd-test covering the API and opcode, and as a drive-by removes the now-obsolete jspiEnabledCallback declaration.

Page release goes through one funnel, and the right allocator

memory.discard returns physical pages to the OS, so where it releases them from matters. The patch routes all three entry points — the JS API, the jitted memory_discard_wrapper, and the interpreter — through BackingStore::DiscardWasmMemory rather than discarding raw addresses:

trusted_data->memory_object(mem_index)->backing_store()->DiscardWasmMemory(dst, size);

That single point then resolves the allocator the reservation actually came from, mirroring what the free path already does:

#ifdef V8_ENABLE_SANDBOX
  if (page_allocator_.expired()) return;
  auto page_allocator = page_allocator_.lock();
  DiscardWasmMemoryPages(page_allocator.get(), address, size);
#else
  DiscardWasmMemoryPages(GetPlatformPageAllocator(), address, size);
#endif

Under V8_ENABLE_SANDBOX, wasm memory is reserved from the isolate group's backing-store allocator in TryAllocateAndPartiallyCommitMemory, not the platform allocator. Reaching for GetPlatformPageAllocator() therefore decommitted through an allocator that does not own the region. That is benign on Linux — DiscardSystemPages is madvise(MADV_DONTNEED), keyed on address — but wrong on the DecommitPages + SetPermissions path used elsewhere, which mutates allocator bookkeeping.

Verification

Measured against a patched d8 with the sandbox enabled: a 64 MB discard releases exactly 67108864 bytes of RSS, and the region reads back as zeroes and stays writable. Release is precise at page granularity — discarding a single 64 KB wasm page frees exactly 64 KB, re-discarding it frees nothing, and a subsequent overlapping 32 MB discard frees exactly 32 MB − 64 KB.

One note for anyone testing this at the V8 level: d8 enables the wasm trap handler by default, which requires an 8 GB kFullGuardSize32 reservation per memory and so cannot allocate any wasm memory inside a smaller sandbox. --no-wasm-trap-handler is needed to match an embedder that does not install the handler

[ask-bonk]

APIError: Invalid Anthropic API Key

github run

[ask-bonk]

APIError: Invalid Anthropic API Key

github run

[ask-bonk]

APIError: Invalid Anthropic API Key

github run

[ask-bonk]

@guybedford Bonk workflow failed. Check the logs for details.

View workflow run · To retry, trigger Bonk again.

[codspeed-hq]

Merging this PR will not alter performance

✅ 72 untouched benchmarks
⏩ 129 skipped benchmarks¹

---

_{Comparing gbedford/wasm-memory-discard (727e68f) with main (482c2be)}

Footnotes

1. 129 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩