perf(codegen): CPU-aware default for native LLVM unit workers (#8583)

changelog.d/8586-rs4gc-budget-assert.md

@@ -0,0 +1,5 @@
+### Changed
+
+- `PERRY_LL_PREOPT_OPTNONE_INSTRS` is removed (#8583). It stamped `optnone` before `rewrite-statepoints-for-gc`, which makes the pass manager skip `mem2reg`/`sccp` while RS4GC still runs, so a demoted function's root allocas were never promoted and the collector never saw them. The cap defaulted to 0, so no shipped build was affected; a test now pins that an `optnone` function loses every root under the rewrite.
+- `PERRY_LL_RS4GC_MAX_INSTRS` (default 1.5 Mi): after `rewrite-statepoints-for-gc`, a function whose body exceeds the per-function budget fails its codegen unit with the function's name and its sizes before and after the rewrite, instead of entering an optimizer pipeline that is super-linear on statepoint relocation fan-out and would not finish. This is an assertion, not a fallback — no function is ever demoted and the requested optimization level applies to every function. `<n>` raises it, `warn:<n>` only warns, `0` disables. Both caches key on it.
+- `PERRY_CODEGEN_UNIT_TIMINGS` now reports, per codegen unit, the widest function by estimated IR before LLVM starts, and after compile the instruction totals and widest function before and after RS4GC, the growth factor, and rewrite/optimize/emit times.

changelog.d/8587-root-spill.md

@@ -0,0 +1,3 @@
+### Changed
+
+- Native GC-root spilling (#8583): a function whose estimated statepoint relocation count — `live_root_slots × safepoints` — exceeds `PERRY_ROOT_SPILL_RELOCATIONS` (default 4,000,000) keeps its GC roots in a heap shadow frame instead of native statepoints. `rewrite-statepoints-for-gc` adds one relocation per live root per safepoint, so a minified-bundle entry function (measured: 795 root slots × ~106k safepoints, grown 439k → 6.5M instructions under RS4GC) drove the `-Os` middle-end super-linear and did not finish; the same unit optimizes in ~5s once that one function is spilled. The function is still compiled at the requested optimization level — only its root representation changes — and its roots stay precise: the runtime already scans shadow-frame and stack-map roots in one walk, and the frame pointer is kept so the FP-chain walker steps over the spilled frame. Each spilled function is reported at default verbosity. `PERRY_ROOT_SPILL_RELOCATIONS=0` disables spilling (every function on native statepoints, the previous behavior).

changelog.d/8590-native-unit-workers.md

@@ -0,0 +1,3 @@
+### Changed
+
+- The default number of concurrent native LLVM codegen-unit workers is now CPU-aware on non-Windows: half the machine's logical CPUs, clamped to `[2, 8]`, instead of a hard-coded `2` (#8583). The `2` default (#8017) was chosen for Windows pagefile pressure and applied everywhere; on a large real bundle (the Claude Code `cli.js` lowers to ~84 units) it left most cores idle while dozens of ~7-minute units ran two at a time. With the giant entry function's roots spilled (#8583) no unit carries an unbounded RS4GC fan-out, so per-unit peak RSS is a bounded ~1-2 GiB and the two-worker cap — not memory — was the wall. Windows keeps the conservative `2`. `PERRY_CODEGEN_UNIT_JOBS` still overrides on every platform.

crates/perry-codegen/src/codegen/closure.rs

@@ -612,6 +612,12 @@ pub(super) fn compile_closure(
         // which spans the body.
         let capture_root_slots =
             u32::from(captures_this || enclosing_class.is_some()) + u32::from(captures_new_target);
+        crate::codegen::helpers::maybe_spill_roots_to_shadow_frame(
+            lf,
+            &llvm_name,
+            m.len() + capture_root_slots as usize,
+            body,
+        );
         lf.enable_shadow_frame(m.len() as u32 + capture_root_slots);
         m
     } else {

crates/perry-codegen/src/codegen/function.rs

@@ -554,6 +554,12 @@ pub(super) fn compile_function(
             cross_module.flat_const_arrays.keys().copied().collect();
         let m =
             crate::collectors::collect_pointer_typed_locals(&f.params, &f.body, &flat_const_ids);
+        crate::codegen::helpers::maybe_spill_roots_to_shadow_frame(
+            lf,
+            &llvm_name,
+            m.len(),
+            &f.body,
+        );
         lf.enable_shadow_frame(m.len() as u32);
         m
     } else {

crates/perry-codegen/src/codegen/helpers.rs

@@ -357,6 +357,74 @@ pub(crate) fn inline_hot_small_max_call_sites() -> u32 {
     })
 }
 
+/// #8583: statepoint relocation estimate above which a function keeps its GC
+/// roots in a shadow frame instead of native statepoints.
+///
+/// `rewrite-statepoints-for-gc` adds one relocation per GC value live across
+/// each safepoint, so the optimizer's post-rewrite cost scales with
+/// `live_roots × safepoints`. Past a point that fan-out makes the `-Os`/`-O3`
+/// middle-end super-linear and the compile does not finish (the Claude Code
+/// bundle's 68 MB entry body measured 795 root slots × ~106k safepoints ≈ 8.4e7
+/// and grew 439k → 6.5M instructions under RS4GC; without RS4GC the same unit
+/// optimized at `-Os` in ~5s). Real functions sit orders of magnitude below
+/// this: hundreds of call sites times tens of slots is ~1e4–1e5. The default
+/// is set well under the measured pathological point and well over ordinary
+/// code, and the post-RS4GC instruction-budget assertion (#8583, inprocess.rs)
+/// backstops any function the estimate misses.
+///
+/// `PERRY_ROOT_SPILL_RELOCATIONS=<n>` overrides it; `0` disables spilling
+/// (every function stays on native statepoints, the pre-#8583 behavior).
+const DEFAULT_ROOT_SPILL_RELOCATIONS: usize = 4_000_000;
+
+fn root_spill_relocation_threshold() -> usize {
+    std::env::var("PERRY_ROOT_SPILL_RELOCATIONS")
+        .ok()
+        .and_then(|v| v.trim().parse::<usize>().ok())
+        .unwrap_or(DEFAULT_ROOT_SPILL_RELOCATIONS)
+}
+
+/// The relocation estimate for a function with `slot_count` GC-root slots and
+/// a body containing `safepoint_sites` call-like expressions. Saturating so a
+/// pathological product cannot wrap.
+pub(crate) fn root_relocation_estimate(slot_count: usize, safepoint_sites: usize) -> usize {
+    slot_count.saturating_mul(safepoint_sites)
+}
+
+/// Decide whether `func` should spill its roots to the shadow frame, and if so
+/// mark it (BEFORE its `enable_*_shadow_frame` call) and report it. Only
+/// meaningful under native stack-map roots — the shadow frame is already the
+/// lowering otherwise. Reporting is at default verbosity because #8421 requires
+/// that a change to how a function is compiled is never silent; the message
+/// states that the optimization level is unchanged.
+pub(super) fn maybe_spill_roots_to_shadow_frame(
+    func: &mut crate::function::LlFunction,
+    fn_name: &str,
+    slot_count: usize,
+    body: &[perry_hir::Stmt],
+) {
+    if !native_stack_roots_enabled() {
+        return;
+    }
+    let threshold = root_spill_relocation_threshold();
+    if threshold == 0 {
+        return;
+    }
+    let sites = crate::collectors::count_safepoint_sites(body);
+    let estimate = root_relocation_estimate(slot_count, sites);
+    if estimate <= threshold {
+        return;
+    }
+    func.request_shadow_frame_spill();
+    eprintln!(
+        "perry: `{fn_name}` keeps its {slot_count} GC roots in a shadow frame instead of \
+         statepoints: an estimated {estimate} relocations ({slot_count} roots × {sites} \
+         safepoints) would make rewrite-statepoints-for-gc fan-out super-linear in the \
+         optimizer (> {threshold}). The function is still compiled at the requested \
+         optimization level; only its GC-root representation changes, and its roots stay \
+         precise (#8583). Override with PERRY_ROOT_SPILL_RELOCATIONS."
+    );
+}
+
 pub(super) fn enable_module_init_shadow_frame(
     func: &mut crate::function::LlFunction,
     stmts: &[perry_hir::Stmt],
@@ -368,6 +436,10 @@ pub(super) fn enable_module_init_shadow_frame(
 
     let shadow_slot_map =
         crate::collectors::collect_pointer_typed_locals(&[], stmts, flat_const_ids);
+    // #8583: the module-entry body is the minified-bundle IIFE — the function
+    // that fans out catastrophically under RS4GC. Decide its root lowering
+    // before the frame is built.
+    maybe_spill_roots_to_shadow_frame(func, "main", shadow_slot_map.len(), stmts);
     func.enable_post_init_shadow_frame(shadow_slot_map.len() as u32);
     let shadow_slot_clears_after_stmt =
         crate::collectors::collect_shadow_slot_clear_points(stmts, &shadow_slot_map);

crates/perry-codegen/src/codegen/method.rs

@@ -319,6 +319,12 @@ pub(super) fn compile_method(
             &method.body,
             &flat_const_ids,
         );
+        crate::codegen::helpers::maybe_spill_roots_to_shadow_frame(
+            lf,
+            &llvm_name,
+            m.len() + 1,
+            &method.body,
+        );
         lf.enable_shadow_frame(m.len() as u32 + 1);
         m
     } else {
@@ -1392,6 +1398,12 @@ pub(super) fn compile_static_method(
             cross_module.flat_const_arrays.keys().copied().collect();
         let m =
             crate::collectors::collect_pointer_typed_locals(&f.params, &f.body, &flat_const_ids);
+        crate::codegen::helpers::maybe_spill_roots_to_shadow_frame(
+            lf,
+            &llvm_name,
+            m.len() + 1,
+            &f.body,
+        );
         lf.enable_shadow_frame(m.len() as u32 + 1);
         m
     } else {

crates/perry-codegen/src/collectors/mod.rs

@@ -42,6 +42,7 @@ mod ptr_shape_report;
 mod ptr_shape_returns;
 mod refs;
 mod repsel_benefit;
+mod safepoint_sites;
 mod scalar_method_dispatch;
 mod scalar_methods;
 mod shadow_slots;
@@ -93,6 +94,7 @@ pub(crate) use ptr_shape_returns::collect_exported_return_shapes;
 pub(crate) use refs::{
     collect_let_ids, collect_ref_ids_in_expr, collect_ref_ids_in_stmts, is_clamp_call,
 };
+pub(crate) use safepoint_sites::count_safepoint_sites;
 pub(crate) use scalar_method_dispatch::{
     collect_module_dispatch_facts, mark_unstable_scalar_method_receivers, ModuleDispatchFacts,
 };

crates/perry-codegen/src/collectors/safepoint_sites.rs

@@ -0,0 +1,219 @@
+//! Count the GC safepoints in a function body (#8583).
+//!
+//! `rewrite-statepoints-for-gc` inserts, at every safepoint, one relocation
+//! per GC value live across it — so the optimizer's post-rewrite work grows
+//! with `live_roots × safepoints`. A function whose product is large enough
+//! makes the `-Os`/`-O3` middle-end super-linear: the 68 MB minified entry
+//! body of the Claude Code bundle measured 795 root slots × ~106k safepoints
+//! and grew 439k → 6.5M instructions under RS4GC, and a single `-Os` pass on
+//! the result did not finish in practical time (#8583).
+//! `codegen/helpers::maybe_spill_roots_to_shadow_frame` multiplies this count
+//! by the function's root-slot count and, past a threshold, keeps that
+//! function's roots in a shadow frame instead of statepoints.
+//!
+//! A safepoint is any call-like expression: a call can re-enter the runtime
+//! and collect. The count is an over-approximation biased toward spilling —
+//! a false positive is a shadow frame on a function that would have been fine
+//! (cheap; the shadow lowering is the pre-#7370 default), while a false
+//! negative would let relocation fan-out reach the optimizer. Nested closures
+//! are NOT counted: each compiles to its own `LlFunction` with its own frame,
+//! so its safepoints belong to it (`walk_expr_children` does not descend into
+//! a closure's body, only its parameter defaults).
+
+use perry_hir::{Expr, Stmt};
+
+/// Total call-like expressions reachable from `stmts` without descending into
+/// nested closures.
+pub fn count_safepoint_sites(stmts: &[Stmt]) -> usize {
+    let mut n = 0usize;
+    for s in stmts {
+        count_in_stmt(s, &mut n);
+    }
+    n
+}
+
+/// A call-like expression is a potential safepoint: anything whose lowering
+/// emits a call that can re-enter the runtime. Nodes not listed contribute
+/// nothing themselves but are still recursed into, so adding a new call
+/// variant can only make the estimate more conservative (a possible
+/// under-count that the post-RS4GC instruction-budget assertion backstops),
+/// never wrong in a way that hides a fan-out.
+fn is_safepoint(e: &Expr) -> bool {
+    matches!(
+        e,
+        Expr::Call { .. }
+            | Expr::CallSpread { .. }
+            | Expr::NativeMethodCall { .. }
+            | Expr::StaticMethodCall { .. }
+            | Expr::SuperCall(_)
+            | Expr::SuperCallSpread(_)
+            | Expr::SuperMethodCall { .. }
+            | Expr::SuperMethodCallSpread { .. }
+            | Expr::ObjectSuperMethodCall { .. }
+            | Expr::New { .. }
+            | Expr::NewDynamic { .. }
+            | Expr::NewDynamicSpread { .. }
+            | Expr::Await(_)
+            | Expr::Yield { .. }
+            | Expr::AsyncFirstCall { .. }
+    )
+}
+
+fn count_in_expr(e: &Expr, n: &mut usize) {
+    if is_safepoint(e) {
+        *n += 1;
+    }
+    // Generic recursion into direct sub-expressions. `walk_expr_children` does
+    // not descend into a closure's statement body (only its param defaults),
+    // which is exactly the boundary we want: a nested closure is a separate
+    // frame and its safepoints are not this function's.
+    perry_hir::walker::walk_expr_children(e, &mut |child| count_in_expr(child, n));
+}
+
+fn count_in_stmt(s: &Stmt, n: &mut usize) {
+    match s {
+        Stmt::Let { init: Some(e), .. }
+        | Stmt::Expr(e)
+        | Stmt::Throw(e)
+        | Stmt::Return(Some(e)) => count_in_expr(e, n),
+        Stmt::Let { init: None, .. } | Stmt::Return(None) => {}
+        Stmt::If {
+            condition,
+            then_branch,
+            else_branch,
+        } => {
+            count_in_expr(condition, n);
+            for st in then_branch {
+                count_in_stmt(st, n);
+            }
+            if let Some(else_branch) = else_branch {
+                for st in else_branch {
+                    count_in_stmt(st, n);
+                }
+            }
+        }
+        Stmt::While { condition, body } | Stmt::DoWhile { body, condition } => {
+            count_in_expr(condition, n);
+            for st in body {
+                count_in_stmt(st, n);
+            }
+        }
+        Stmt::For {
+            init,
+            condition,
+            update,
+            body,
+        } => {
+            if let Some(init) = init {
+                count_in_stmt(init, n);
+            }
+            if let Some(condition) = condition {
+                count_in_expr(condition, n);
+            }
+            if let Some(update) = update {
+                count_in_expr(update, n);
+            }
+            for st in body {
+                count_in_stmt(st, n);
+            }
+        }
+        Stmt::Labeled { body, .. } => count_in_stmt(body, n),
+        Stmt::Try {
+            body,
+            catch,
+            finally,
+        } => {
+            for st in body {
+                count_in_stmt(st, n);
+            }
+            if let Some(catch) = catch {
+                for st in &catch.body {
+                    count_in_stmt(st, n);
+                }
+            }
+            if let Some(finally) = finally {
+                for st in finally {
+                    count_in_stmt(st, n);
+                }
+            }
+        }
+        Stmt::Switch {
+            discriminant,
+            cases,
+        } => {
+            count_in_expr(discriminant, n);
+            for c in cases {
+                if let Some(t) = &c.test {
+                    count_in_expr(t, n);
+                }
+                for st in &c.body {
+                    count_in_stmt(st, n);
+                }
+            }
+        }
+        // No expression children.
+        Stmt::Break
+        | Stmt::Continue
+        | Stmt::LabeledBreak(_)
+        | Stmt::LabeledContinue(_)
+        | Stmt::PreallocateBoxes(_)
+        | Stmt::PreallocateTdzBoxes(_)
+        | Stmt::ReleaseBoxes(_) => {}
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::count_safepoint_sites;
+    use perry_hir::types::Type;
+    use perry_hir::{Expr, Stmt};
+
+    fn call(args: Vec<Expr>) -> Expr {
+        Expr::Call {
+            callee: Box::new(Expr::Undefined),
+            args,
+            type_args: vec![],
+            byte_offset: 0,
+        }
+    }
+
+    fn empty_closure(body: Vec<Stmt>) -> Expr {
+        Expr::Closure {
+            func_id: 0,
+            params: vec![],
+            return_type: Type::Any,
+            body,
+            captures: vec![],
+            mutable_captures: vec![],
+            captures_this: false,
+            captures_new_target: false,
+            enclosing_class: None,
+            is_arrow: false,
+            is_async: false,
+            is_generator: false,
+            is_strict: false,
+        }
+    }
+
+    #[test]
+    fn counts_calls_across_control_flow_but_not_into_closures() {
+        let stmts = vec![
+            Stmt::Expr(call(vec![])),
+            Stmt::While {
+                condition: Expr::Bool(true),
+                body: vec![Stmt::Expr(call(vec![]))],
+            },
+            // A call buried in a nested closure body must NOT be counted.
+            Stmt::Expr(empty_closure(vec![Stmt::Expr(call(vec![]))])),
+            Stmt::Return(Some(call(vec![]))),
+        ];
+        assert_eq!(count_safepoint_sites(&stmts), 3);
+    }
+
+    #[test]
+    fn call_arguments_are_themselves_safepoints() {
+        // f(g(), h()) is three calls.
+        let nested = call(vec![call(vec![]), call(vec![])]);
+        assert_eq!(count_safepoint_sites(&[Stmt::Expr(nested)]), 3);
+    }
+}