apple
Include try and await expressions which are parents of a freestanding expression macro in lexicalContext
This modifies the behavior of MacroExpansionContext.lexicalContext such that any preceding try or await expressions are included.
Motivation: The #expect macro in Swift Testing currently cannot propagate try or await placed before the freestanding expression macro to expressions in its expanded code. See Forum discussion. This change would allow us to fix that.
Resolves rdar://109470248
Just spitballing—could there be some way to alter the macro expansion itself so as to be insensitive to any try or await that precedes it? For example, would wrapping the current expansion in an immediately executed closure allow for this to work:
/* Would `try` need to be hoisted here too (?) */ { Testing.__checkValue(try f(), expression: .__fromSyntaxNode("try f()"), comments: [], isRequired: false, sourceLocation: Testing.SourceLocation.__here()).__expected() }()
One imagines that there could exist other syntax which impact the 'hygiene' of the #expect macro that could benefit from a more clever/defensive expansion.
Just spitballing—could there be some way to alter the macro expansion itself so as to be insensitive to any
tryorawaitthat precedes it? For example, would wrapping the current expansion in an immediately executed closure allow for this to work:/* Would `try` need to be hoisted here too (?) */ { Testing.__checkValue(try f(), expression: .__fromSyntaxNode("try f()"), comments: [], isRequired: false, sourceLocation: Testing.SourceLocation.__here()).__expected() }()One imagines that there could exist other syntax which impact the 'hygiene' of the
#expectmacro that could benefit from a more clever/defensive expansion.
No, that wouldn't work unfortunately. All that does is add another spot where we might need try or await, because the call to Testing.__checkValue() would still need them.
No, that wouldn't work unfortunately. All that does is add another spot where we might need
tryorawait
I wasn't under the impression that the specific text I sketched would work (in fact, I was 100% certain it wouldn't)—rather, that was a freehand sketch to try to illustrate a direction that I'm wondering whether you all have explored. The crux of the question is whether it's been concluded that there exists no solution within the confines of what macros support today that's possible with more finagling of the specific macro expansion emitted.
Indeed, as you say, the approach I have in mind is about ways to deliberately add extra trys—this would (potentially—or perhaps I am naive to think it) allow a syntactically valid expansion that unconditionally emits a try without regard to whether there is an outer one—although obviously this could have other consequences that are undesirable that I haven't worked through. Specifically, consider the following:
// This wrapper function allows one to write an extraneous `try` because it unconditionally throws.
func _wrapper<T>(_ f: @autoclosure () throws -> T) throws -> T { try f() }
func foo() throws -> Int { 42 }
func bar() -> Int { 42 }
try bar() // Warning: no calls to throwing functions...
// Look! With `_wrapper`, I can (and must) write an extra `try` and it's fine, without checking whether `bar` throws.
try _wrapper(bar())
// And it's fine to have as many outer `try`s as I want.
try { try _wrapper(bar()) }()
try { try { try _wrapper(bar()) }()
I know the idea is half-baked. However, with more elbow grease, is it fully bakeable or are we truly stuck...
(Part of the reason I ask is, while swift-syntax API changes are outside the scope of Swift Evolution, those changes that add/change what macros are capable of doing are considered to be language features that need review.)
The problem is ultimately that we can only see the syntax nodes inside the macro. Unconditionally adding a try or await ahead of the expression would change the semantics of the expression and require the calling context to be throwing and/or async.
You'll note we do unconditionally add try and await when we initialize suite types, because in the context where we do so, we're already async throws so it doesn't matter.
Yeah, I see your point.
Going in the other direction, how confident can we be that await and try are the only or even most of the cases now that have this contextual issue for freestanding macros?
Should we instead consider some sort of "whole-statement freestanding macro" that always provides the entire statement as read-only context, such that the parens don't delimit what's visible to the macro but only marks the span that will be expanded?
Mutating member function calls on value types also don't compile correctly and they are syntactically identical to non-mutating ones. There is no readily available solution, but those are relatively rare so it's not as big of a problem.
New kinds of macros are beyond the purview of Swift Testing—you may want to bring your idea to the forums to discuss with other stakeholders. 🙂