reflect: implement MakeChan

[0pcom]

reflect.MakeChan currently panics with "unimplemented". This implements it.

It completes a small gap: MakeSlice, MakeMap and New are all supported, so code that builds a channel reflectively is the odd one out — and it is reached by generic container and codec libraries that construct channel types from a reflect.Type.

Verified by building TinyGo and running the compiler test suite.

[dgryski]

Please add a test for this new function. For reflect, this can either be in testdata/reflect.go or reflect/value_test.go.

[0pcom]

Added TestTinyMakeChan in src/reflect/value_test.go, and rebased on dev.

It covers buffered and unbuffered creation (kind, Cap, Len, and a real send/receive through Interface(), since Value.Send/Value.Recv are still unimplemented), plus the three panic paths: non-channel type, negative buffer size, and a unidirectional channel type.

On verification: CI has not run on this PR, so I checked the test's expectations against the host reflect package, where it passes — that pins the behaviour to upstream semantics. I could not run it against these sources under a released TinyGo binary (the musl build gets in the way), so the implementation itself is still only verified by reading.

[dgryski]

LGTM. The tests are currently failing on wasi because #5550 is not yet merged (which ideally should be "soon"). We can either hold off on merging this until that one is in, or update the tests that panic to not run on wasi yet.

[0pcom]

Took the second option, since #5550 is still open with review discussion on it and I would rather not have this wait on an unknown.

The three panic cases now skip on wasip1, with a TODO pointing at #5550 so the skip goes when that lands. The buffered and unbuffered cases — which are the substance, covering Cap, Len, and a real send and receive through Interface() — still run everywhere.

If you would rather hold this until #5550 is in, that last commit is the only thing to drop; the rest is unchanged from your LGTM.

[dgryski]

t.Skip() doesn't work on wasi because it requires the same underlying features as panic/recover. We need to t.Log(); return or wait for the panic/recover PR to land.

[0pcom]

Swapped for t.Log(); return — thanks, I had not realised t.Skip needs the same machinery it was standing in for.

The two substantive cases, buffered and unbuffered, still run everywhere; only the three panic cases return early on wasip1, with the TODO still pointing at #5550 so they come back when it lands.

[deadprogram]

Looks like the tests are still failing on this...

[dgryski]

Yes, the runtime.GOOS == "wasip1" check needs to also check for GOOS == "wasip2"

[0pcom]

Fixed, and it turned out to be the CI failure rather than a cosmetic gap — thanks for spotting it.

tinygo-test-wasip2-fast was failing with panic: reflect: call of MakeChan on int Value and a wasm trap at value_test.go:1056. That is the first of the three panic cases: under wasip2 the GOOS == "wasip1" guard did not fire, the case ran, recover does not work on wasm yet, so the panic escaped and took the test binary with it.

I checked on runtime.GOARCH == "wasm" rather than adding wasip2 to a GOOS list, because the limitation is wasm's rather than any one platform's: wasip1, wasip2 and wasm all set the tinygo.wasm build tag, and arch_tinygowasm.go defines GOARCH as wasm for it, so one condition covers all three of the wasm test targets — including tinygo-test-wasm, which the explicit GOOS pair would still have missed.

One thing worth flagging for anyone reading along, because it briefly convinced me I had it wrong: targets/wasip2.json says "goarch": "arm". That is what is handed to the Go toolchain for package selection; runtime.GOARCH is the constant above and reports wasm.

Happy to switch to the explicit wasip1 || wasip2 || js list instead if you would rather it read the same way as the other checks in the tree.