[wasm][debugger] Fix reusing buffer for debugger

[thaystg]

When the wasm linear memory has to grow it invalidates existing views into the heap. To avoid this issue make sure to always create a new view into the heap for the debugger command immediately before use.

[ghost]

Tagging subscribers to this area: @thaystg
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Module.HEAPU8.buffer can be reallocated so we need to set _debugger_heap_bytes everytime that we will call set method.

Author:	thaystg
Assignees:	-
Labels:	arch-wasm, area-Debugger-mono
Milestone:	-

[lewing]

This fix is appropriately minimal under the circumstances but it might make sense to inline the base64 decode here and avoid an intermediate copy.

[lewing]

@kg any thoughts on a way to catch these more easily in testing?

[radical]

How was this caught, or how did it fail?

[lewing]

How was this caught, or how did it fail?

caught in CTI testing, eventually if the runtime allocates more memory than the current allocation debugging will stop working (in fairly non-obvious ways)

The underlying issue is emscripten-core/emscripten#12336

[lambdageek]

The underlying issue is emscripten-core/emscripten#12336

There's a suggestion in that issue to monkeypatch wasmMemory.grow so we get our own notification about the heap. Any reason we don't want to do that, @lewing@kg ?

[lewing]

The underlying issue is emscripten-core/emscripten#12336

There's a suggestion in that issue to monkeypatch wasmMemory.grow so we get our own notification about the heap. Any reason we don't want to do that, @lewing@kg ?

it is more a question of what exactly to do with those notifications

[kg]

The main problem is that in order to do anything about this we'd need to make people stop using raw arrayviews so we could wrap the views in a type that does something when the heap is invalidated. ArrayViews are designed to fail silently and weirdly in classic JS fashion, and because JS object indexing isn't customizable we can't even make a fake array-like object.

We could do something like the root API and store the buffers in a container and make people call .value or .buffer every time, I suppose. I'm not sure that would help. In this case it would have at least. We probably just need to audit all our code for it - I thought I did in the past but I definitely did not see this line.