dart:ffi interaction with dynamic

[dcharkes]

While not efficient, dart:ffi should be available with dynamic (except for trampolines).

Pointer<Int8> p = allocate();
dynamic q = p;
q.store(123);
q.free();

TODO: test whether it works and write tests for this functionality.

[sjindel-google]

We will lock out these methods on dynamic for now so that switching to extension methods will be a non-breaking change (against the runtime behaviour).

Note: we need to determine whether "existing" code which passes the type arguments to load will be broken.

[sjindel-google]

This task is mentioned in code which checks that type arguments to load and store are statically known, which is a separate additional issue.

[sjindel-google]

Another thing to check is that the trampolines and other ops that work on Pointers but aren't instance methods still work when the Pointer class is implemented (instead of extended).

[sjindel-google]

@mraleph @dcharkes
This could be another substantial task to work on next: it requires re-writing the dynamic invocation forwarders in the VM.

[sjindel-google]

Not sure what the exact scope of this issue is, but we need to also disallow dynamic invocations of asFunction and lookupFunction.

[dcharkes]

Not sure what the exact scope of this issue is, but we need to also disallow dynamic invocations of asFunction and lookupFunction.

That would be trivial with extension methods.

If we ship before extension methods land, we need to prohibit this. However, this is work that we would be obsolete once we switch to extension methods, so I'd give this low priority.

[sjindel-google]

We can give it low priority as long as the dynamic invocations don't make the VM crash.

[sjindel-google]

@dcharkes Assigning myself since suggested that we should remove dynamic invocations of Pointer.load and Pointer.store altogether in https://dart-review.googlesource.com/c/sdk/+/107513

[dcharkes]

This is part of the CL that addresses Pointer optimizations (#38172), and is currently in review.

[mkustermann]

We already track extension methods in #37773 and do not plan support for dynamic load/store atm, right? So can we close this issue?

[sjindel-google]

Yes, this issue is obsolete.

[dcharkes]

The CL that rules out invocations on dynamic hasn't landed yet. It should land today.