You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Attempt to have the try-delegate label after delegate{l} (failing tests)
[Currently](https://github.com/WebAssembly/exception-handling/blob/main/proposals/exception-handling/Exceptions-formal-overview.md) `(try ... delegate l)`
reduces to `(label_n{} ( delegate{l} ... end ) end)`, so by putting a label outside (i.e., before) the administrative `delegate{l}`.
An idea proposed in past [unresolved](WebAssembly#205 (comment))
[discussions](https://github.com/WebAssembly/exception-handling/pull/143/files#r812476148) of WebAssembly#205 and WebAssembly#143,
is to simplify and improve the formalism by instead putting the delegate label inside (i.e., after) the `delegate{l}`.
So instead to reduce to `(delegate{l} ( label_n{} ... end ) end)`.
TL;DR
-----
I can't seem to make it work.
This PR explored an approach to implement this idea, in the
[formal overview file ](https://github.com/WebAssembly/exception-handling/blob/main/proposals/exception-handling/Exceptions-formal-overview.md),
as well as in the interpreter, but had failing tests which I wasn't able to fix. Perhaps I'm overseeing some other solution or approach,
or there is some mistake in the interpreter implementation and/or argumentation below.
Details
-------
I think the problem is as follows.
With the current [definition of block contexts](https://webassembly.github.io/exception-handling/core/exec/runtime.html#block-contexts),
the instruction sequence `B^l[ delegate{l} T[val^n (throw a)] end ]` not only is ambiguous, but doesn't work with a try-catch label located
outside of the try-catch.
Take for example the following possible reduction.
```
(try (try (throw x) delegate 0) catch x end)
↪ (label_0{}
(catch{a_x ε}
(delegate{0}
(label_0{}
(throw a_x) end) end) end) end)
```
The intention for this delegate is to throw inside the handler `catch{a_x ε}` and be caught there.
However, a possible `B^0` for the reduction of `delegate{0}` is `B^0 = [_] catch{a_x ε}`,
in which case the reduction rule gives the following.
```
↪ (label_0{} (throw a_x) end)
↪ (throw a_x)
```
The issue here seems to be that there is no label between the `delegate{l}` and the `catch{...}`.
Perhaps there is a different change we can easily make it work, for example changing control contexts or block contexts?
Failing tests
.............
We can observe the above wrong behaviour also in the interpreter tests, although this could be fixable somehow.
In particular, the first commit of this PR has the formal overview changes also implemented in the execution steps of the interpreter (in `interpreter/exec/eval.ml`):
- The reduction of `try ... delegate l` puts the `label{}` after the `delegate{l}`.
- The reduction of `delegate{l}` does not pattern match for an initial label.
I tried to minimise a failing test from `test/core/try_delegate.wast` in the file `test/core/try_delegate_minimal_fail.wast`.
To reproduce the failure build the interpreter and run the above test file as follows,
for example from a Linux terminal in the base directory of the repository:
```
cd interpreter
make
./wasm ../test/core/try_delegate_minimal_fail.wast
```
See also comments in the test file.
Copy file name to clipboardExpand all lines: proposals/exception-handling/Exceptions-formal-overview.md
+8-7Lines changed: 8 additions & 7 deletions
Original file line number
Diff line number
Diff line change
@@ -180,21 +180,22 @@ catch T[val^n (throw a)] end ↪ val^n (throw a)
180
180
181
181
182
182
F; val^n (try bt instr* delegate l)
183
-
↪ F; label_m{} (delegate{l} val^n instr* end) end
183
+
↪ F; delegate{l} (label_m{} val^n instr* end) end
184
184
(if expand_F(bt) = [t1^n]→[t2^m])
185
185
186
186
delegate{l} val* end ↪ val*
187
187
188
-
label_m{} B^l[ delegate{l} T[val^n (throw a)] end ] end
188
+
B^l[ delegate{l} T[val^n (throw a)] end ]
189
189
↪ val^n (throw a)
190
190
```
191
191
192
192
Note that the last reduction step above is similar to the reduction of `br l`[1], if we look at the entire `delegate{l}...end` as the `br l`, but also doing a throw after it breaks.
193
193
194
-
There is a subtle difference though. The instruction `br l` searches for the `l+1`th surrounding block and breaks out after that block. Because `delegate{l}` is always wrapped in its own `label_n{} ... end`[2], with the same lookup as for `br l` we end up breaking inside the `l+1`th surrounding block, and throwing there. So if that `l+1`th surrounding block is a try, we end up throwing in its "try code", and thus correctly getting delegated to that try's catches.
194
+
There is a subtle difference though. The instruction `br l` searches for the `l+1` surrounding block and breaks out after that block.
195
+
On the contrary, `delegate{l}` should "break and throw" _inside_ the `l+1` label, so we should break one label earlier.
196
+
This is why the reduction step for `delegate{l}` has one label less.
195
197
196
198
-[1][The execution step for `br l`](https://webassembly.github.io/spec/core/exec/instructions.html#xref-syntax-instructions-syntax-instr-control-mathsf-br-l)
197
-
-[2] The label that always wraps `delegate{l}...end` can be thought of as "level -1" and cannot be referred to by the delegate's label index `l`.
0 commit comments