fix(codegen): #168 — specialize i32 loop counter for number-typed bounds

[proggeramlug]

Summary

Closes #168.

Perry already specializes loop counters to i32 when the bound is arr.length (the classify_for_length_hoist peephole). This PR extends that optimization to the equally-common i < n shape where n is a number-typed function parameter or local — the pattern that blocks LoopVectorizer on Buffer-read and other intrinsic-heavy hot paths.

What changed

Single file: crates/perry-codegen/src/stmt.rs (+142 lines, no other files).

New helper — classify_for_local_bound(cond, ctx):

• Accepts Compare { op: Lt|Le, left: LocalGet(i), right: LocalGet(n) }
• Requires i to be in integer_locals (initialized from integer literal, only mutated via Update)
• Requires n to be either also in integer_locals, or a number-typed (Type::Number | Type::Int32) non-boxed, non-global slot (covers n: number parameters and simple let locals typed number)

In lower_for — new parallel block alongside the arr.length path:

1. Allocates a parallel i32 alloca for the counter (if not already present from the Let site) and initializes it via fptosi of the current double value
2. Emits fptosi(n) once before the cond block into a fresh i32 alloca (the key: loop-invariant integer bound in a register LLVM/SCEV can see)
3. Uses icmp slt i32 %i, %n.i32 (or icmp sle for <=) instead of fcmp olt double
4. The existing Expr::Update lowering already keeps the i32 slot in sync with add i32 1 per iteration
5. Removes the i32 counter slot after loop exit (only if this path was the one that allocated it)

Correctness

• Counter (i): integer_locals guarantees it is initialized from an integer literal and only ever modified by Update ++/--. The i32 slot round-trips exactly.
• Bound (n): when n is in integer_locals the same guarantee holds. When n is a number-typed parameter slot Perry trusts the TypeScript type annotation (the same trust-types contract already applied throughout codegen). A non-integer float bound (e.g. foo(3.7)) would observe at most one fewer iteration — a trade-off within Perry's existing contract.

Test

Functional correctness verified locally:

sumA (const bound, N=1000) = 499500
sumB (param bound, N=1000) = 499500
match: YES

The variant B sum now equals variant A — previously the param-bound loop ran correctly but without i32 specialization; now it emits icmp slt i32 in the cond block, giving LLVM a modelable integer induction variable.

Gap suite: stable at 23/28 (no regressions). All perry-codegen, perry-hir, perry-transform, perry-runtime, and perry crate tests pass (72/72).

Before / after (representative LLVM IR shape)

Before (variant B cond block):

%ctr_dbl = load double, ptr %i_slot
%bound_dbl = load double, ptr %n_slot
%cmp = fcmp olt double %ctr_dbl, %bound_dbl
br i1 %cmp, label %for.body, label %for.exit

After (variant B cond block):

%ctr_i32 = load i32, ptr %i_i32_slot
%bound_i32 = load i32, ptr %n_i32_slot   ; hoisted fptosi(n) from loop head
%cmp = icmp slt i32 %ctr_i32, %bound_i32
br i1 %cmp, label %for.body, label %for.exit

The icmp slt i32 gives LLVM's SCEV a clean integer induction variable, which is the prerequisite for LoopVectorizer to widen the loop body.

Notes for maintainer

• No version bump, no CLAUDE.md entry, no Recent Changes entry per the external-contributor policy in CLAUDE.md.
• The Le case (i <= n) is handled via icmp_sle (consistent with the existing classify_for_length_hoist extension).
• The classify_for_local_bound guard deliberately does not fire when hoist_classification.is_some() — the two peepholes are mutually exclusive.

https://claude.ai/code/session_019Y542k6QyAWV6JF2an8DXU

---

Generated by Claude Code