Stream: wasmtime

Topic: exception handling implementation

Joel Dice (Feb 01 2024 at 13:53):

@fitzgen (he/him) thanks for the informative presentation on the status of core wasm proposals in Wasmtime yesterday. Regarding exception handling: you mentioned there were two paths to implementing it:

• Unwind tables, with the advantage of being efficient for modules which don't use exception handling or use it rarely, but at the cost of being more difficult to implement
• Swift-style exception propagation via "normal" control flow using result variants, with the advantage of being simpler to implement, but also less efficient for code which rarely uses exceptions

I'm curious if you've considered a third option: do a wasm->wasm transformation that converts use of exception handling to normal, Swift-style control flow (i.e. transform the wasm module into one that doesn't use wasm exception handling). The upshot being that Wasmtime wouldn't need to implement exception handling at all except as a preprocessing step prior to AOT compilation. What kind of performance penalty would that incur vs. doing a Swift-style implementation in Cranelift itself? And are there any other drawbacks you would anticipate with such an approach?

fitzgen (he/him) (Feb 01 2024 at 14:03):

good question! the primary advantage of having cranelift do the "resultification" would be that we can tweak the ABI that functions use to make that work especially well. for example, IIRC swift will actually set a particular register when returning an error, and then propagating that error for callers ends up being a very simple test and branch, and doesn't have to do any kind of unpacking data from memory that was passed via return pointer or anything like that

fitzgen (he/him) (Feb 01 2024 at 14:05):

that said, I suspect that we will probably want to do the unwind tables approach, because there are other things that want similar tables (eg stack maps and such) and the primary benefit of avoiding tables is to have things "obviously correct when looking at code emitted by the compiler" but as soon as we rely on any tables, we lose that property. and yeah there is something to be said about degrees, but I think going from 0 to 1 is a much bigger decision than 1 to 2 if that makes sense

Catherine (whitequark) (Feb 13 2024 at 19:10):

I'm curious if you've considered a third option: do a wasm->wasm transformation that converts use of exception handling to normal, Swift-style control flow (i.e. transform the wasm module into one that doesn't use wasm exception handling).

I've considered implementing this (in binaryen, incidentally) since wasmtime doesn't support EH and the work required to add that support required time I do not have. I actually still have a WIP branch with the beginning of that implementation, here's a file I grabbed from it.

I've taken a slightly different approach (inspired by, but incompatible with Asyncify), which essentially has a global "unwinding" flag and all the associated exception data, and functions do branch after each invoke, but only on that flag rather than on the result. This has, I assume, several advantages:

• The very hot flag could be allocated to a global register by a sufficiently smart runtime. (I think right now wasmtime only special-cases the stack pointer? I may be out of date.)
• You do strictly less work on the non-exceptional case, and constrain the register allocator less by requiring fewer registers in returns.

I've never finished the work because the Wasm EH semantics is quite complex and has lots of edge cases that for the most part are not actually used by any frontends but still require some degree of care in a general Wasm->Wasm transformation.

For example, you can leave a catch_all without rethrow. This isn't something that a code generator lowering Itanium EHABI will ever produce, as far as I know, but it's permitted by the spec. (For those who aren't familiar: The way things usually work--and this is true for native code as well--is that a low-level exception type is allocated for the concept of a "C++ exception" as a whole, which is caught when any catch block at all is present, even catch(...), and rethrown if the C++ type doesn't match. catch_all is reserved for finally blocks in presence of foreign exceptions, e.g. Ada exception propagating from C++ code. This is relevant basically never in native EHABI, though I think this might be how JS exceptions propagate from Wasm frames, in which case it'd be quite important to get this right for finally to work.)

Another example is lexical rethrows (rethrow 1 for example), which I understand some people (the WAMR author I think?) consider a mistake. I don't quite recall all the details but it is true that it adds considerable implementation complexity. Fortunately that one doesn't seem to be relevant as clang will only ever emit rethrow 0, right at the end of a catch or catch_all.

Hope this helps. I'm still interested in having this transformation be possible, just no time to work on it any time soon. It's a bit of a nightmare to test too.

Alex Crichton (Feb 13 2024 at 19:19):

@Catherine (whitequark) question for you, is your feedback here and the complexity of implementation all based on the "old" definition of the exception-handling proposal?

Alex Crichton (Feb 13 2024 at 19:19):

w.r.t. the past october the CG decided to take the exception handling approach in a different direction I think in response to some of this feedback

Catherine (whitequark) (Feb 13 2024 at 19:46):

Oh yeah I just discovered that while I wasn't looking, the entire EH proposal was changed to something completely different. I'm really happy with this new proposal, as it can for the most part be trivially lowered to non-EH Wasm:

• the transformation would add a global/thread-local exception tag index, plus a few globals for exception arguments (up to the maximum amount discovered; in practice, this is just two);
• throw sets the tag index to some non-zero value corresponding to the thrown tag, and copies the arguments over;
• every function call (to a function that's not proven to never unwind) checks the global tag index, and if it's nonzero, returns some indeterminate value.
• try_table checks the global tag index against the tag (or just if it's non-zero) in the catch clause, and if it's a catch[_all]_ref clause, packs all the globals into a tuple, replaces the exnref with that tuple, and clears the global tag index;
• throw_ref unpacks the tuple back into the globals and returns (or jumps to the outer try_table).

This approach has only one wrinkle: it's not strictly speaking fully general as it can't handle an open world system (e.g. dynamic linking) where any newcoming module may decide to allocate an exception with an ever-higher amount of arguments. This can be handled by storing the arguments in memory instead of in globals, but I'll argue this is unnecessary: in many cases a closed-world approach is fine, and whenever it's not, you could have a parameter saying how many arguments at most should be expected. People really do not create new EH lowerings very often and the value of 2 is sufficient to handle all code I've seen personally.

(If you go with an open-world, dynamic linking approach, the dynamic linker will have to allocate tag values too. But this lets you lower EH across Wasm module boundaries, which I think is pretty cool!)

Alex Crichton (Feb 13 2024 at 19:52):

Oh nice! Wanted to confirm one way or another, but that sounds like it speaks quite well in favor of the current trajectory

Catherine (whitequark) (Feb 13 2024 at 21:03):

It's nice and simple enough that it's tempting to go and try my hand at implementing it, but I know better than that. Very happy to leave fun projects to other (newer?) contributors too :) and/or lend a hand implementing it

Last updated: Nov 22 2024 at 16:03 UTC