Performance implications of mapping vs bounds-checked allocations

[taralx]

Since one of the intended uses of this is for ArrayBuffers, I am wondering about the performance (and predictability) implications of using mapping operations on buffers whose sizes are not a multiple of a page size (and indeed may be smaller than a page). What is supposed to happen if you access past the end of a mapped buffer?

At one point I had envisioned a set of memory instructions that accepted an object pointer and performed bounds checks (WebAssembly/gc#109). Do you imagine that these mapping instructions would be able to take the place of those?

[dtig]

Sorry for the late reply, I'm catching up after being out on vacation.

The proposal in this repository doesn't explicitly state this, but the proposed operators would take size in multiples of page size, and not as a number of bytes. My original thought was that accessing past the end of a mapped buffer should be an out of bounds check, but other options were mentioned in the design issue here - WebAssembly/design#1439.

By the mapping instruction do you mean the addition of fat pointers of fully dynamic memories?

[taralx]

Ah, it sounds like the answer is "no, these are page-sized objects". Mapping instructions are memory.map/memory.unmap.