Math.imul not lowered — 'expression MathImul not yet supported'

[proggeramlug]

Repro

const r = Math.imul(0x01000193, 0x811c9dc5);
console.log('imul result:', r);

Actual (Perry 0.5.30)

Error compiling module '…': lowering entry of module '…': lowering init
statements of module '…': perry-codegen Phase 2: expression MathImul not
yet supported

Expected

Compile successfully; print the 32-bit-wrap product (e.g. -2110866647).

Notes

Math.imul(a, b) is the canonical spelling for 32-bit-wrap multiplication in portable JS/TS — hash functions (FNV-1a-32, MurmurHash3, xxhash32, CityHash32), PRNGs (PCG, xorshift*), and most CRC32 implementations all use it. Without it, TS that targets Node and Perry has to branch, or hand-roll the 16-bit hi/lo split:

function imul32(a: number, b: number): number {
  const aHi = (a >>> 16) & 0xffff, aLo = a & 0xffff;
  const bHi = (b >>> 16) & 0xffff, bLo = b & 0xffff;
  return ((aLo * bLo) + (((aHi * bLo + aLo * bHi) << 16) >>> 0)) | 0;
}

— which is what the honest_bench/workloads/3_image_convolution/perry/image_conv.ts and honest_bench/workloads/1_json_pipeline/perry/json_pipeline.ts files do as a workaround.

The underlying primitive should be cheap to lower: fptosi a → trunc to i32 → fptosi b → trunc to i32 → mul i32 → sitofp.

[proggeramlug]

Fixed in v0.5.34.

Root cause: Math.imul(a, b) reached HIR as Expr::MathImul (the HIR lowering matches the builtin and constructs the node), but the LLVM codegen in crates/perry-codegen/src/expr.rs had no match arm — every call fell through to the catch-all which errored with Phase 2: expression MathImul not yet supported. The JS / WASM / Glance emitters all had arms; the LLVM backend was the only gap.

Fix: 5-instruction inline sequence — fptosi DOUBLE→I64 → trunc I64→I32 → mul i32 → sitofp I32→DOUBLE. This matches the JS ToInt32 sequence (wrap-to-i64, take the low 32 bits) for every finite double, and LLVM mul i32 wraps without nsw/nuw. No runtime helper needed.

Verified against Node:

• Math.imul(0x01000193, 0x811c9dc5) → 84696351 ✓ (the issue's illustrative -2110866647 was for a different argument pair)
• Math.imul(-1, -1) → 1 ✓
• Math.imul(0xffffffff, 5) → -5 ✓
• Math.imul(3, 0x7fffffff) → 2147483645 ✓

Unblocks FNV-1a-32 / MurmurHash3 / xxhash32 / CRC32 / PCG / xorshift* in user TS without the 16-bit hi/lo workaround.