markused,gen: fix direct str()call on a smart-casted interface variable

[rilaaax]

Fixes #27713

Problem

Calling .str() directly on an interface variable that's been smart-cast to a primitive type inside a match branch (without rebinding it to a new variable first) had two separate bugs:

1. It crashed the C compiler instead of producing a working binary:

cc: error: call to undeclared function 'builtin__str_intp'
cc: error: use of undeclared identifier 'StrIntpData'

2. Once that crash was fixed, the printed result was wrong: &42 instead of 42

Minimal repro:

interface Value {}

fn get_str(v Value) string {
	return match v {
		int { v.str() }
		else { '' }
	}
}

fn main() {
	println(get_str(42))
}

Note: rebinding first (x := v; x.str()) avoids both issues, which was the
first clue pointing at cgen/dead-code-elimination rather than the checker.

Cause

Interfaces box non-pointer-kind variants (like int) behind an allocated pointer in C (_int is stored as int*). For a direct .str() call on a smart-cast identifier, cgen resolves the receiver using the raw internal smartcast type, which still carries that pointer flag — unlike the "exposed" V-level type (plain int) that the checker and node.left_type use.

This raw, pointer-flagged type causes two independent problems:

1. markused/walker.v decides whether to keep str_intp (the string interpolation runtime) based on node.left_type (plain int, no pointer flag), so it never learns that cgen's pointer-flagged code path will actually need str_intp, and strips it from the final C output.
2. gen_expr_to_string() in str.v treats any pointer-flagged type as "this is a real reference the user wants printed with &" — correct for genuine pointers/structs stored by reference, but not for a primitive value that only looks like a pointer because of interface boxing.

Fix

• markused/walker.v: when registering .str() usage, also check whether the receiver is a smart-cast identifier whose raw smartcast type is a pointer, and register that type too, so str_intp/StrIntpData aren't stripped when cgen will actually need them.
• str.v: detect the same case (an interface variable smart-cast to a primitive/non struct-like type) and use an empty ref prefix instead of &, leaving all other pointer/struct/reference printing behavior unchanged.

Test

Regression test added.

[GGRei]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 58be88136c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve real pointer prefixes in interface .str()

When an interface is smart-cast to an actual pointer primitive, for example an empty-interface value containing &int and a match v { &int { v.str() } } arm, this predicate is also true because typ.deref() resolves to int. The code then suppresses ref_str, so direct .str() loses the pointer-style prefix (42 instead of the real pointer receiver formatting such as &42) and no longer matches rebinding before .str(). Please distinguish boxed value smartcasts from genuine pointer variants before dropping the & prefix.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Keep aggregate smartcasts out of primitive handling

For grouped interface match arms such as A, B { v.str() }, the smartcast target is represented as an .aggregate, but .aggregate is not excluded here even though the surrounding smartcast code treats it like structs/arrays rather than scalars. That makes is_interface_smartcast_boxed_primitive true for aggregate arms containing struct variants and drops the leading &, so direct .str() in those arms prints A{}/B{} instead of the existing interface-smartcast struct formatting (&A{}/&B{}) and differs from single-variant arms. Please include .aggregate in the non-primitive set.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Avoid retaining str_intp for enum smartcasts

For an interface match arm whose variant is an enum, e.g. match v { MyEnum { v.str() } }, cgen handles the pointer-flagged receiver in the enum branch and emits the enum str call directly rather than builtin__str_intp. This unconditional raw-pointer mark therefore keeps str_intp/StrIntpData in the generated C even though the call does not need them, regressing skip-unused output size for enum smartcasts. Please restrict the extra mark to the raw smartcast kinds whose .str() code path actually emits string interpolation support.

Useful? React with 👍 / 👎.