Stream: git-wasmtime

Topic: wasmtime / Issue #2712 Inconsistent fuel accounting betwe...

Wasmtime GitHub notifications bot (Mar 08 2021 at 14:39):

sunfishcode opened Issue #2712:

In WebAssembly, a br instruction with a depth equal the the current control-flow stack depth is equivalent to a return. However, in the current fuel implementation, these are handled differently:

• return consumes 0 fuel, while a br to the control-flow stack depth consumes 1.
• return updates the fuel usage field in VMInterrupts, while a br, br_if, or br_table to the control-flow stack depth does not, even though they also exit the function

Wasmtime GitHub notifications bot (Mar 08 2021 at 14:39):

sunfishcode labeled Issue #2712:

In WebAssembly, a br instruction with a depth equal the the current control-flow stack depth is equivalent to a return. However, in the current fuel implementation, these are handled differently:

• return consumes 0 fuel, while a br to the control-flow stack depth consumes 1.
• return updates the fuel usage field in VMInterrupts, while a br, br_if, or br_table to the control-flow stack depth does not, even though they also exit the function

Wasmtime GitHub notifications bot (Mar 08 2021 at 17:17):

alexcrichton commented on Issue #2712:

FWIW the current heuristic for fuel consumed per-op is pretty primitive. It's based on the lucet one but slightly adjusted.

The thinking behind return taking 0 fuel is that the cost was already incurred with the original call, but we could pretty easily just update it to cost 1 fuel as well. In general my hope is that folks don't rely on the precise amount of fuel consumed by instructions too too much because it can make changes to the code generator brittle and/or future peformance improvements somewhat more difficult.

For the second point, though, I believe that all of these branching/control-flow opcodes update fuel correctly, even if br goes to the exit of the function. Was a bug seen in the wild though?

Wasmtime GitHub notifications bot (Mar 08 2021 at 17:50):

sunfishcode commented on Issue #2712:

Ah, no, this is just from me diving into the middle of the code. Good point about br going to the exit and being handled there.

Last updated: Oct 23 2024 at 20:03 UTC