Map.clear and Set.clear keep the bucket table - #100

What

Map.clear and Set.clear zero the bucket table in place and keep its length. They shrank it to length zero.

Why

Every get, has, set and delete computes hash % buckets.length. After a clear that was a remainder by zero:

• the compiler does uint % uint in floating point (frem), so the result was NaN;
• NaN was then converted to an array index (fptosi), which is undefined.

It gave right answers only by luck. On x86-64 the conversion yields index 0, shrinking an array zero-fills its storage and keeps the block, so index 0 read as an empty bucket; and the first insert after a clear resizes before it writes. Where the optimizer can see the length is zero it emits a trap instead: the same lookup on a plain emptied array crashed under the JIT.

Change

• src/generics/lib.generics.ts: clear() in Map and in Set. The next insert still finds no room in the entries and resizes both tables, as before.
• tests/map_clear_reuse.ts (new): lookups, delete and iteration on a cleared map and set, refilling past the first size, clear after deletes, clear twice.

For the reviewer

• The new test passed before the change too. No wrong result could be produced on this platform, so it guards the behaviour and does not demonstrate the fault. The evidence for the fault is the generated IR and the trap on the plain-array probe.
• A cleared map keeps its bucket table until the next insert, which replaces it with a small one.
• Not changed here: uint % uint compiled as a floating-point remainder is in the compiler. It is on the path of every Map and Set operation.

Testing

• Release, Windows x64: the whole suite passes for gc, rc, none and own, compiled and under the JIT.
• Not run locally: debug builds, Linux, x86 and wasm.

🤖 Generated with Claude Code