nikomatsakis (Dec 25 2024 at 14:24): nikomatsakis (Dec 25 2024 at 14:24):I'm curious -- if I am generating webassembly, how much should I care about generating machine code that is "optimized" for the WASM interpreter. For example, I could probably simplify my life if I store all local variables into memory, but I could also work a bit harder and store some subset of them (perhaps a large subset) in WASM local variables and the native WASM stack.
Does this matter? Can I expect wasmtime to do some kind of "Heap SSA"-like transformation and rewrite both into an equivalent form?
nikomatsakis (Dec 25 2024 at 14:25):Similarly, should I try to use memory.copy to copy "small-ish" regions of memory (e.g., 16 bytes) or is it equivalent if I just generate 4 int32 loads and then 4 int32 stores?
Chris Fallin (Dec 25 2024 at 17:53):We don’t have the equivalent of mem2reg, and wasm semantics actually make it a bit tricky to do so (any memory access could trap and the heap state post-trap is externally visible if exported; so at least stores must remain, and loads at least must bounds-check unless one can prove in-range or subsumed which we don’t do). So a good cost model is probably: operand stack and locals become SSA become registers, loads/stores remain and are more expensive than native
Chris Fallin (Dec 25 2024 at 17:57):We also don’t try to merge smaller accesses into larger, because of the same trapping semantics; I would suggest using wasm-simd to move 128 bits at a time where possible. memory.copy turns into a trampoline into the runtime IIRC so not so cheap for smaller copies
nikomatsakis (Dec 26 2024 at 12:28):Thanks!
nikomatsakis (Dec 26 2024 at 12:31):I'm familiar with binaryen; I'm wondering what other wasm-to-wasm optimizers exist-- are there any that are willing to e.g. treat memory state as "not inspectable"?
Alex Crichton (Dec 26 2024 at 16:42):I'm only aware of binaryen myself and IIRC it has opt-in flags for semantics like that to enable optimizations
Dan Gohman (Dec 26 2024 at 16:48):mem2reg on Wasm is also hard because we don't have the non-aliasing guarantees of LLVM's alloca. Any memory location is up for grabs by almost any store anywhere.
Dan Gohman (Dec 26 2024 at 16:50):Binaryen faces the same limitation. I don't think any tool will ever be able to do mem2reg well on Wasm.
nikomatsakis (Dec 27 2024 at 13:28):Interesting.
nikomatsakis (Dec 27 2024 at 13:28):Seems like something that tools could emit as "opt-in" annotations of a sort
Dan Gohman (Jan 03 2025 at 13:32):It's an interesting question. Wasm doesn't have an obvious place to attach such annotations, so I don't think it's trivial to do.
Dan Gohman (Jan 03 2025 at 13:32):I sometimes wonder if would make sense to design a Wasm-like compiler IR, which would be like Wasm plus instructions for things like "begin a stack frame", "end a stack frame", and alloca. And on top of that, some concept of provenance for pointers. Maybe also GlobalVariable and noalias concepts too. All these things could be lowered into plain Wasm, but in unlowered form would have enough nondeterminism to give a C-compiler-style optimizer room to breath.
Last updated: Jan 24 2025 at 00:11 UTC