Skip statically dead typeof branches at MLIR generation

[ASDAlexander77]

Problem

In a generic specialised for a concrete type, a typeof check whose answer is known still generated its branch. When that branch uses the value as the narrowed type, it emits a cast the value cannot take:

class Array1 {
    static [Symbol.hasInstance](instance) {
        if (typeof instance === "string")
            print("str: ", instance);   // specialised for array<any>: array -> string cast
        return false;
    }
}
print([] instanceof Array1);

In debug builds the cast reached llvm_unreachable("review usage") at CastLogicHelper.h:815; in release builds that is undefined behaviour and the test only passes because the branch never runs. This is why 01symbol and 00funcs_generic_with_typeof fail on a local debug build but never on CI (all workflows build Release).

Two separate causes:

1. typeof x === "name" was not folded. For a statically typed x, typeof emits a TypeDescriptor, whose run-time name (typeOfAsString of its type) is compared with the literal by content, so the result is known during MLIR generation.
2. if added narrowing before deciding to skip a branch. Even with a constant condition, checkSafeCast emitted the narrowing cast into the skipped branch. Unused, it was removed as dead code, but under --di (which the test runner always passes) the narrowed variable's ts.DebugVariable kept it alive into LLVM lowering.

Fix

• foldStaticTypeOfCompare (MLIRGenImpl.h): ==, ===, != and !== between a TypeDescriptor and a string literal, in either order, produce a boolean literal, which if already uses to skip a branch. any, union tags and stored typeof values keep their run-time check. Names are compared exactly as at run time (e.g. an integer literal is still "s32"), so no existing result changes.
• mlirGen(IfStatement): narrowing is added only to a branch that is generated, for both then and else.

Not covered: ?:, &&/|| and while also narrow but always generate both branches, so a typeof-guarded ?: in an array-specialised generic can still hit the same cast. That was already the case before this change.

Tests

• New 00typeof_static_fold.ts (compile, jit, rc/none corpus): a generic describe<T> called with a string, an empty array, a string array, a number and a boolean, plus typeof on an any. It crashes the compiler without the fix.
• 01symbol and 00funcs_generic_with_typeof produce the same output as the release build.
• Full Windows debug suite: 2722/2727. The 5 remaining failures are pre-existing and unrelated (00owned_array_ops under rc, gc-shared-auto, gc-defaultlib-collector); 12 tests that failed before now pass.

🤖 Generated with Claude Code