[codex] Lower loop-bound array indices as i32

PERF_RUN_LOG.md

@@ -234,3 +234,70 @@
   - `benchmarks/baseline.json` is stale for this Linux environment; compare was run with `--warn-only` and the before/after comparison above uses the captured local third-cycle results.
   - This follow-up is intended as a stacked draft PR on top of the monomorphic array guard fast-cache PR.
 - PR: https://github.com/PerryTS/perry/pull/5309
+
+## 2026-06-17 - I32 lowering for loop-bound numeric array indices
+
+- Start revision: `966729232`
+- Branch: `codex/perry-i32-array-index-lowering`
+- Worker assignment: single Codex pass in this worktree
+- Benchmark environment: Linux `/usr/bin/time`; local `node` cannot execute `.ts` benchmark inputs, so Node columns and correctness comparisons were skipped by the harness
+- Baseline commands:
+  - `cargo build --release`
+  - `target/release/perry compile --no-cache benchmarks/suite/16_matrix_multiply.ts -o /tmp/perry-matrix-guard-precheck-final2 --quiet`
+  - `for i in 1 2 3 4 5; do /usr/bin/time -f "sample=$i wall=%e rss_kb=%M" /tmp/perry-matrix-guard-precheck-final2; done`
+  - `PERRY_TYPED_FEEDBACK_TRACE=/tmp/perry-matrix-guard-precheck-final2-trace.json /usr/bin/time -f "wall=%e rss_kb=%M" /tmp/perry-matrix-guard-precheck-final2 && jq '.guards' /tmp/perry-matrix-guard-precheck-final2-trace.json`
+  - `perf stat -e cycles,instructions,branches,branch-misses /tmp/perry-matrix-guard-precheck-final2`
+  - `benchmarks/quick.sh`
+  - `benchmarks/compare.sh --quick --runs 3 --warn-only --json-out /tmp/perry-guard-precheck-final2.json`
+- Baseline results:
+  - direct matrix binary: 400ms, 398ms, 398ms, 385ms, 386ms; checksum always `41079519680`
+  - final trace run: `matrix_multiply:395`, checksum `41079519680`, wall 0.42s, RSS 31500KB, 33,619,968 numeric array index-get guard passes, 65,536 numeric array index-set guard passes, 0 get/set guard failures; push guard retained 39 fallback calls
+  - `perf stat` direct matrix binary: 1,443,394,074 cycles, 7,034,084,638 instructions, 1,568,434,556 branches, 241,348 branch-misses, 0.4222s elapsed
+  - compare quick medians: loop_overhead 76ms/18880KB, fibonacci 266ms/18764KB, math_intensive 55ms/19092KB, nested_loops 225ms/23204KB, factorial 95ms/18764KB
+  - quick: fibonacci 254ms/18MB, math_intensive 73ms/18MB, nested_loops 229ms/22MB, factorial 97ms/18MB, matrix_multiply 407ms/30MB
+- Selected gap and evidence:
+  - After numeric array guard pre-classification, `matrix_multiply` remained the slowest `quick.sh` case at 407ms.
+  - LLVM trace for `benchmarks/suite/16_matrix_multiply.ts` still lowered hot computed get indices such as `i * size + k` through `sitofp`/`fmul`/`fadd`/`fptosi` before calling the typed-feedback numeric array get guard.
+  - Loop-bound analysis already proved and hoisted `size` as an i32 loop bound for `i < size` and `k < size`, but that trusted bound was not visible to the existing i32 expression lowering used by index expressions.
+- Change:
+  - Reused or inserted an i32 slot for local loop bounds classified by the `i < n` loop-bound path and kept that slot visible while lowering the loop body.
+  - Used the existing `can_lower_expr_as_i32` / `lower_expr_as_i32` machinery for known-array computed get indices when the index expression is fully backed by trusted i32 slots, integer locals, or constants.
+  - Preserved the typed-feedback numeric array get guard and fallback path; the final i32 index is converted back to double only for the guard's double index argument.
+  - Added an IR regression test covering `xs[i * size + 1]` inside `for (let i = 0; i < size; i++)`, asserting guarded fallback emission plus `mul i32`/`add i32` and no `fmul double` for that computed index.
+- Post-change benchmark commands:
+  - `cargo build --release`
+  - `target/release/perry compile --no-cache benchmarks/suite/16_matrix_multiply.ts -o /tmp/perry-matrix-i32-index-proto --trace llvm --quiet`
+  - `rg -n "js_typed_feedback_numeric_array_index_get_guard|fmul double|mul i32|add i32|sitofp i32" .perry-trace/llvm/_16_matrix_multiply_ts.ll`
+  - `for i in 1 2 3 4 5; do /usr/bin/time -f "sample=$i wall=%e rss_kb=%M" /tmp/perry-matrix-i32-index-proto; done`
+  - `PERRY_TYPED_FEEDBACK_TRACE=/tmp/perry-matrix-i32-index-proto-trace.json /usr/bin/time -f "wall=%e rss_kb=%M" /tmp/perry-matrix-i32-index-proto && jq '.guards' /tmp/perry-matrix-i32-index-proto-trace.json`
+  - `perf stat -e cycles,instructions,branches,branch-misses /tmp/perry-matrix-i32-index-proto`
+  - `benchmarks/quick.sh`
+  - `benchmarks/compare.sh --quick --runs 3 --warn-only --json-out /tmp/perry-i32-index-proto.json`
+  - `for i in 1 2 3 4 5 6 7 8 9 10; do /usr/bin/time -f "sample=$i wall=%e rss_kb=%M" /tmp/perry-matrix-i32-index-proto; done`
+- Post-change results:
+  - LLVM trace confirmed the two hot matmul numeric-array get indices now use `mul i32` and `add i32` before `call i32 @js_typed_feedback_numeric_array_index_get_guard`; remaining `sitofp i32` values feed the guard's double index argument.
+  - direct matrix binary first sample set: 400ms, 393ms, 388ms, 403ms, 393ms; checksum always `41079519680`
+  - direct matrix binary 10-sample set: 397ms, 396ms, 390ms, 392ms, 392ms, 393ms, 383ms, 389ms, 386ms, 384ms; checksum always `41079519680`
+  - trace run: `matrix_multiply:397`, checksum `41079519680`, wall 0.42s, RSS 31440KB, 33,619,968 numeric array index-get guard passes, 65,536 numeric array index-set guard passes, 0 get/set guard failures; push guard retained 39 fallback calls
+  - `perf stat` direct matrix binary: 1,456,553,467 cycles, 7,017,622,346 instructions, 1,568,480,860 branches, 249,497 branch-misses, 0.4217s elapsed
+  - quick: fibonacci 251ms/18MB, math_intensive 71ms/18MB, nested_loops 202ms/22MB, factorial 99ms/18MB, matrix_multiply 387ms/30MB
+  - compare quick medians: loop_overhead 56ms/18784KB, fibonacci 248ms/18896KB, math_intensive 55ms/18900KB, nested_loops 214ms/23268KB, factorial 78ms/18776KB
+- Measured impact:
+  - `16_matrix_multiply` direct median: 398ms -> 391ms, 1.8% faster
+  - `16_matrix_multiply` quick: 407ms -> 387ms, 4.9% faster
+  - Direct matrix binary instructions: 7.034B -> 7.018B, 0.2% fewer
+  - Direct matrix binary cycles: 1.443B -> 1.457B, 0.9% more in the single perf sample; branch misses also rose from 241K to 249K, so counter impact is mixed despite lower wall-time samples
+  - `10_nested_loops` compare median: 225ms -> 214ms, 4.9% faster
+- Verification:
+  - `cargo fmt --check`
+  - `git diff --check`
+  - `cargo test -p perry-codegen --test typed_feedback`
+  - `cargo test -p perry-codegen --test typed_shape_descriptors`
+  - `PERRY_BIN=target/release/perry python3 tests/test_typed_feedback_runtime_evidence.py`
+  - `tests/test_benchmark_output_verifier.sh`
+  - `cargo build --release`
+  - Typed-feedback trace confirmed get/set guard pass counts and zero get/set failures match the pre-change trace.
+- Notes:
+  - `benchmarks/baseline.json` is stale for this Linux environment; compare was run with `--warn-only`, and the before/after comparison above uses the captured local fourth-cycle baseline.
+  - This is a smaller cleanup than the preceding guard-cache work. The keeper signal is the consistent matrix wall-time reduction plus removal of double arithmetic from the hottest generated get-index chains; perf counters should be watched on future runs.
+- PR: https://github.com/PerryTS/perry/pull/5310

crates/perry-codegen/src/expr/index_get.rs

@@ -35,7 +35,7 @@ use crate::types::{DOUBLE, I1, I16, I32, I64, I8, PTR};
 use super::arrays_finds::lower_buffer_index_get_i32;
 #[allow(unused_imports)]
 use super::{
-    buffer_access_materialization_reason, buffer_alias_metadata_suffix,
+    buffer_access_materialization_reason, buffer_alias_metadata_suffix, can_lower_expr_as_i32,
     emit_layout_note_slot_on_block, emit_shadow_slot_clear, emit_shadow_slot_update_for_expr,
     emit_string_literal_global, emit_typed_feedback_register_site, emit_v8_export_call,
     emit_v8_member_method_call, emit_write_barrier, emit_write_barrier_slot_on_block,
@@ -847,8 +847,30 @@ pub(crate) fn lower(ctx: &mut FnCtx<'_>, expr: &Expr) -> Result<String> {
                 }
 
                 let arr_box = lower_expr(ctx, object)?;
-                let idx_double = lower_expr(ctx, index)?;
-                let idx_i32 = ctx.block().fptosi(DOUBLE, &idx_double, I32);
+                let i32_slots = ctx.i32_counter_slots.clone();
+                let flat_const_arrays = ctx.flat_const_arrays.clone();
+                let array_row_aliases = ctx.array_row_aliases.clone();
+                let integer_locals = ctx.integer_locals.clone();
+                let use_i32_index = can_lower_expr_as_i32(
+                    index,
+                    &i32_slots,
+                    &flat_const_arrays,
+                    &array_row_aliases,
+                    &integer_locals,
+                    ctx.clamp3_functions,
+                    ctx.clamp_u8_functions,
+                    ctx.integer_returning_functions,
+                    ctx.i32_identity_functions,
+                );
+                let (idx_double, idx_i32) = if use_i32_index {
+                    let idx_i32 = lower_expr_as_i32(ctx, index)?;
+                    let idx_double = ctx.block().sitofp(I32, &idx_i32, DOUBLE);
+                    (idx_double, idx_i32)
+                } else {
+                    let idx_double = lower_expr(ctx, index)?;
+                    let idx_i32 = ctx.block().fptosi(DOUBLE, &idx_double, I32);
+                    (idx_double, idx_i32)
+                };
                 if !require_numeric_layout
                     && !matches!(index.as_ref(), Expr::Integer(_) | Expr::Number(_))
                 {

crates/perry-codegen/src/stmt/loops.rs

@@ -386,6 +386,7 @@ pub(crate) fn lower_for(
     // site having done so already).  Only the site that inserted should
     // remove it at loop exit to avoid disturbing a pre-existing slot.
     let local_bound_counter_i32_was_fresh: bool;
+    let local_bound_bound_i32_was_fresh: bool;
     let i32_local_bound_slot: Option<String> =
         if let Some((counter_id, bound_id, _op)) = local_bound_classification {
             // Allocate a parallel i32 slot for the counter if not already
@@ -411,18 +412,28 @@ pub(crate) fn lower_for(
             local_bound_counter_i32_was_fresh = fresh;
             // Hoist `fptosi(n)` to a fresh i32 alloca before the cond block
             // so LLVM sees a loop-invariant integer bound — critical for
-            // SCEV / LoopVectorizer to recognize the induction variable.
-            if let Some(bound_slot) = ctx.locals.get(&bound_id).cloned() {
+            // SCEV / LoopVectorizer to recognize the induction variable. Also
+            // expose that slot while lowering the loop body so integer index
+            // expressions like `i * n + k` can reuse the same trusted bound
+            // instead of rebuilding the index through double arithmetic.
+            if let Some(existing) = ctx.i32_counter_slots.get(&bound_id).cloned() {
+                local_bound_bound_i32_was_fresh = false;
+                Some(existing)
+            } else if let Some(bound_slot) = ctx.locals.get(&bound_id).cloned() {
                 let bound_dbl = ctx.block().load(DOUBLE, &bound_slot);
                 let bound_i32 = ctx.block().fptosi(DOUBLE, &bound_dbl, I32);
                 let slot = ctx.func.alloca_entry(I32);
                 ctx.block().store(I32, &bound_i32, &slot);
+                ctx.i32_counter_slots.insert(bound_id, slot.clone());
+                local_bound_bound_i32_was_fresh = true;
                 Some(slot)
             } else {
+                local_bound_bound_i32_was_fresh = false;
                 None
             }
         } else {
             local_bound_counter_i32_was_fresh = false;
+            local_bound_bound_i32_was_fresh = false;
             None
         };
     // Issue #168 follow-up: when neither the `arr.length` hoist nor the static
@@ -718,6 +729,11 @@ pub(crate) fn lower_for(
             ctx.i32_counter_slots.remove(&counter_id);
         }
     }
+    if local_bound_bound_i32_was_fresh {
+        if let Some((_, bound_id, _)) = local_bound_classification {
+            ctx.i32_counter_slots.remove(&bound_id);
+        }
+    }
     let _ = i32_local_bound_slot;
     // Same cleanup for the runtime-guarded `any`-bound path.
     if let Some(dyn_bound) = dynamic_i32_bound {

crates/perry-codegen/tests/typed_feedback.rs

@@ -1,5 +1,8 @@
 use perry_codegen::{compile_module, AppMetadata, CompileOptions};
-use perry_hir::{BinaryOp, Class, ClassField, Expr, Function, Module, ModuleInitKind, Param, Stmt};
+use perry_hir::{
+    BinaryOp, Class, ClassField, CompareOp, Expr, Function, Module, ModuleInitKind, Param, Stmt,
+    UpdateOp,
+};
 use perry_types::{FunctionType, Type};
 
 /// Serializes env-mutating tests so a concurrent test never observes a
@@ -547,3 +550,51 @@ fn typed_feedback_guards_computed_numeric_array_index_hot_path() {
     assert!(!ir.contains("call double @js_array_numeric_get_f64_unboxed"));
     assert!(ir.contains("load double"));
 }
+
+#[test]
+fn typed_feedback_guards_computed_numeric_array_index_uses_i32_loop_bound() {
+    let array_ty = Type::Array(Box::new(Type::Number));
+    let ir = ir_for(module(
+        "typed_feedback_loop_bound_computed_array.ts",
+        vec![param(1, "xs", array_ty), param(2, "size", Type::Number)],
+        Type::Number,
+        vec![Stmt::For {
+            init: Some(Box::new(Stmt::Let {
+                id: 3,
+                name: "i".to_string(),
+                ty: Type::Number,
+                mutable: true,
+                init: Some(Expr::Integer(0)),
+            })),
+            condition: Some(Expr::Compare {
+                op: CompareOp::Lt,
+                left: Box::new(Expr::LocalGet(3)),
+                right: Box::new(Expr::LocalGet(2)),
+            }),
+            update: Some(Expr::Update {
+                id: 3,
+                op: UpdateOp::Increment,
+                prefix: false,
+            }),
+            body: vec![Stmt::Return(Some(Expr::IndexGet {
+                object: Box::new(Expr::LocalGet(1)),
+                index: Box::new(Expr::Binary {
+                    op: BinaryOp::Add,
+                    left: Box::new(Expr::Binary {
+                        op: BinaryOp::Mul,
+                        left: Box::new(Expr::LocalGet(3)),
+                        right: Box::new(Expr::LocalGet(2)),
+                    }),
+                    right: Box::new(Expr::Integer(1)),
+                }),
+            }))],
+        }],
+    ));
+
+    assert!(ir.contains("call i32 @js_typed_feedback_numeric_array_index_get_guard"));
+    assert!(ir.contains("call double @js_typed_feedback_array_index_get_fallback_boxed"));
+    assert!(ir.contains("mul i32"), "{ir}");
+    assert!(ir.contains("add i32"), "{ir}");
+    assert!(!ir.contains("fmul double"), "{ir}");
+    assert!(!ir.contains("call double @js_array_numeric_get_f64_unboxed"));
+}