[ASDAlexander77]

Summary

Comparing an any gave wrong answers in four ways, and if (a) on an any could throw. These are the two any bugs noted in #416, both of which turned out to be broader.

	before	after
const u: any = undefined; u == undefined	false	true
const n: any = null; n == null	false	true
let t: any = "text"; t == "text"	false	true
two anys holding equal text built at run time	unequal	equal
if (u) / !u with u: any = undefined	threw "Can't cast from any type"	false / true

Causes and fixes

1. Compared with undefined, the answer was fixed at compile time. UndefLogicHelper returned == false and != true whatever the any held. It now hands an any to the run-time comparison (AnyCompareOp), as it already did for a class or an interface.

2. The other operand wasn't boxed. LogicOp boxed the other operand only when its type differed from the any's. By then the operands are already lowered, and an any, a string and null are all just ptr, so no box was made:

   • the string itself went into the comparison as if it were a box;
   • its "tag" was read from whatever the string's bytes pointed at;
   • for null, the tag was read from address 8.

   The operand is now boxed by its TypeScript type, on the stack. The box only lives for the comparison. A heap box would never have been freed under rc: 96 MB for 2M comparisons at -O0. With the stack box it's 3.9 MB.

3. undefined and null are answered before any box is opened. A null any pointer counts as undefined, which is also what a zeroed any slot is. Otherwise the tag's kind decides:

   • loosely (==, !=), undefined and null equal each other and nothing else;
   • strictly (===, !==), each equals only itself;
   • an ordering comparison stays false.

4. Two string boxes compared their pointer bytes. Equal text built at run time came out unequal. They're compared with strcmp now.

Truthiness: ___unbox<boolean> threw for an any holding undefined or null, and it's what if (a) and !a use. It now returns false for both, as JavaScript does.

Tests

00any_equality_undefined.ts covers:

• undefined and null under each operator;
• truthiness;
• a let that is reassigned;
• equal text built at run time;
• undefined against null;
• a 100k-iteration comparison loop.

Without the change it fails in every model, compiled and under JIT.

Results

Release ctest	3311/3311
DefaultLib gc (compile + JIT)	158/158
DefaultLib rc (compile + JIT)	3 of 316 fail: the known sleep/weakref_basic link errors

Not in this PR

• Under rc, a JIT run of anyValue == 7 fails with "Symbols not found: ??_7type_info@@6B@". The comparison unboxes through ___unbox<si32>, whose fallback throw needs MSVC's type_info vtable, and the rc JIT doesn't resolve it. This was already the case before this PR.

🤖 Generated with Claude Code