perf(ecs): guarded store follows forwarding edge, inline typeof/typed-array/subclass fast paths (wolf-ecs -16.5% / -20.9%)

changelog.d/8876-ecs-forwarded-store-inline-tiers.md

@@ -0,0 +1,20 @@
+Array and ECS performance. Guarded property-receiver element stores follow one
+growth-forwarding edge inline (the read tier already did), so a field that
+kept a pre-grow forwarding stub (`this.ents[id] = arch`) no longer pays the
+out-of-line extend helper and allocator resolver on every store; the raw-f64
+downgrade note is gated on the header word already loaded. `typeof x ===
+"number"` decides the definitely-Number cases inline, deferring INT32/class
+refs, raw typed-array pointers and the Web Streams id band to the classifier.
+Inline dynamic typed-array reads brand off the `GC_TYPE_TYPED_ARRAY` header
+instead of the evictable 64-slot kind cache. `js_array_get_f64` and the
+typed-feedback array-read fallback route object-backed Array-subclass
+receivers to their dense fast read before the tracked resolver and the by-name
+key path, and an `Any`-typed key that is an integer array index takes the
+inline numeric read tiers (`a[b[i]]`). Object-backed Array subclasses keep
+validated prototype-override reads, and compact guarded specializations are
+pre-statepoint inlined by lowered IR size.
+
+wolf-ecs (noctjs/ecs-benchmark) on the Mac mini reference box, versus the
+previous retained build: add/remove -18.5% (0.556 → 0.453 ms/op), entity-cycle
+-23.1% (0.499 → 0.384 ms/op); each step 11/11 paired wins, semantics probes
+byte-identical to Node.

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

@@ -998,6 +998,8 @@ pub(super) fn compile_closure(
         current_block: 0,
         discard_expr_value: false,
         discard_this_expr: false,
+        truthy_call_result_requested: false,
+        pending_truthy_call_result: None,
         func_names,
         strings,
         loop_targets: Vec::new(),
@@ -1211,6 +1213,7 @@ pub(super) fn compile_closure(
         was_unrolled: false,
         ic_site_counter: ic_base,
         ic_globals: Vec::new(),
+        property_get_ic_override: None,
         typed_parse_rodata: Vec::new(),
         buffer_data_slots: HashMap::new(),
         buffer_view_slots: HashMap::new(),

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

@@ -778,6 +778,8 @@ pub(super) fn compile_module_entry(
             current_block: 0,
             discard_expr_value: false,
             discard_this_expr: false,
+            truthy_call_result_requested: false,
+            pending_truthy_call_result: None,
             func_names,
             strings,
             loop_targets: Vec::new(),
@@ -982,6 +984,7 @@ pub(super) fn compile_module_entry(
             was_unrolled: hir.init_was_unrolled,
             ic_site_counter: ic_base,
             ic_globals: Vec::new(),
+            property_get_ic_override: None,
             typed_parse_rodata: Vec::new(),
             buffer_data_slots: HashMap::new(),
             buffer_view_slots: HashMap::new(),
@@ -1490,6 +1493,8 @@ pub(super) fn compile_module_entry(
             current_block: 0,
             discard_expr_value: false,
             discard_this_expr: false,
+            truthy_call_result_requested: false,
+            pending_truthy_call_result: None,
             func_names,
             strings,
             loop_targets: Vec::new(),
@@ -1694,6 +1699,7 @@ pub(super) fn compile_module_entry(
             was_unrolled: hir.init_was_unrolled,
             ic_site_counter: ic_base,
             ic_globals: Vec::new(),
+            property_get_ic_override: None,
             typed_parse_rodata: Vec::new(),
             buffer_data_slots: HashMap::new(),
             buffer_view_slots: HashMap::new(),

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

@@ -1032,6 +1032,8 @@ pub(super) fn compile_function(
         current_block: 0,
         discard_expr_value: false,
         discard_this_expr: false,
+        truthy_call_result_requested: false,
+        pending_truthy_call_result: None,
         func_names,
         strings,
         loop_targets: Vec::new(),
@@ -1232,6 +1234,7 @@ pub(super) fn compile_function(
         was_unrolled: f.was_unrolled,
         ic_site_counter: ic_base,
         ic_globals: Vec::new(),
+        property_get_ic_override: None,
         typed_parse_rodata: Vec::new(),
         buffer_data_slots: HashMap::new(),
         buffer_view_slots: HashMap::new(),

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

@@ -0,0 +1,176 @@
+//! Guarded omitted-argument/false-field indexed method versioning.
+
+use crate::{compile_module, CompileOptions};
+use perry_hir::types::Type;
+use perry_hir::{Class, ClassField, CompareOp, Expr, Function, Module, Param, Stmt};
+
+const ROWS: u32 = 30;
+const INDEX: u32 = 31;
+const DEFER: u32 = 32;
+
+fn param(id: u32, name: &str) -> Param {
+    Param {
+        id,
+        name: name.to_string(),
+        ty: Type::Any,
+        default: None,
+        decorators: Vec::new(),
+        is_rest: false,
+        arguments_object: None,
+    }
+}
+
+fn default_get() -> Expr {
+    Expr::PropertyGet {
+        object: Box::new(Expr::This),
+        property: "DEFAULT_DEFER".to_string(),
+        byte_offset: 0,
+    }
+}
+
+fn candidate_method() -> Function {
+    let mut defer = param(DEFER, "defer");
+    defer.default = Some(default_get());
+    Function {
+        id: 40,
+        name: "update".to_string(),
+        type_params: Vec::new(),
+        params: vec![param(ROWS, "rows"), param(INDEX, "index"), defer],
+        return_type: Type::Any,
+        body: vec![
+            Stmt::If {
+                condition: Expr::Compare {
+                    op: CompareOp::Eq,
+                    left: Box::new(Expr::LocalGet(DEFER)),
+                    right: Box::new(Expr::Undefined),
+                },
+                then_branch: vec![Stmt::Expr(Expr::LocalSet(DEFER, Box::new(default_get())))],
+                else_branch: None,
+            },
+            // Nominates the existing nonnegative-index family.
+            Stmt::Expr(Expr::IndexGet {
+                object: Box::new(Expr::LocalGet(ROWS)),
+                index: Box::new(Expr::LocalGet(INDEX)),
+            }),
+            Stmt::If {
+                condition: Expr::LocalGet(DEFER),
+                then_branch: vec![Stmt::Return(Some(Expr::Integer(1)))],
+                else_branch: Some(vec![Stmt::Return(Some(Expr::Integer(2)))]),
+            },
+        ],
+        is_async: false,
+        is_generator: false,
+        is_strict: true,
+        is_exported: false,
+        captures: Vec::new(),
+        decorators: Vec::new(),
+        was_plain_async: false,
+        was_unrolled: false,
+    }
+}
+
+fn fixture(method: Function) -> Module {
+    let class = Class {
+        id: 41,
+        name: "Store".to_string(),
+        type_params: Vec::new(),
+        extends: None,
+        extends_name: None,
+        native_extends: None,
+        extends_expr: None,
+        heritage_lexically_shadowed: false,
+        fields: vec![ClassField {
+            name: "DEFAULT_DEFER".to_string(),
+            key_expr: None,
+            ty: Type::Any,
+            init: Some(Expr::Bool(false)),
+            is_private: false,
+            is_readonly: false,
+            decorators: Vec::new(),
+        }],
+        constructor: None,
+        methods: vec![method],
+        getters: Vec::new(),
+        setters: Vec::new(),
+        static_accessor_names: Vec::new(),
+        static_accessor_fn_ids: Vec::new(),
+        computed_members: Vec::new(),
+        static_fields: Vec::new(),
+        static_methods: Vec::new(),
+        decorators: Vec::new(),
+        is_exported: false,
+        aliases: Vec::new(),
+        is_nested: false,
+        alloc_width_hint: 0,
+        specialized_from: None,
+    };
+    let mut module = Module::new("guarded_falsy_default_method.ts");
+    module.classes = vec![class];
+    module
+}
+
+fn emit(method: Function) -> String {
+    let opts = CompileOptions {
+        emit_ir_only: true,
+        output_type: "executable".to_string(),
+        ..Default::default()
+    };
+    String::from_utf8(compile_module(&fixture(method), opts).expect("fixture compiles"))
+        .expect("LLVM IR is UTF-8")
+}
+
+fn function_body<'a>(ir: &'a str, marker: &str) -> &'a str {
+    let start = ir
+        .match_indices("define ")
+        .find(|(index, _)| {
+            let end = ir[*index..]
+                .find('\n')
+                .map(|offset| index + offset)
+                .unwrap_or(ir.len());
+            ir[*index..end].contains(marker)
+        })
+        .map(|(index, _)| index)
+        .unwrap_or_else(|| panic!("missing function containing {marker}:\n{ir}"));
+    let end = ir[start..]
+        .find("\n}")
+        .map(|offset| start + offset)
+        .expect("function terminator");
+    &ir[start..end]
+}
+
+#[test]
+fn wrapper_proves_live_false_field_and_clone_erases_default_and_branch() {
+    let ir = emit(candidate_method());
+    let base = "perry_method_guarded_falsy_default_method_ts__Store__update";
+    let index = format!("{base}$idx_u31_{INDEX}");
+    let specialized = format!("{index}$default_false2");
+    let wrapper = function_body(&ir, &format!("@{base}("));
+    let ordinary = function_body(&ir, &format!("@{index}("));
+    let false_default = function_body(&ir, &format!("@{specialized}("));
+
+    assert!(wrapper.contains(&crate::nanbox::TAG_UNDEFINED_I64.to_string()));
+    assert!(wrapper.contains(&crate::nanbox::TAG_FALSE_I64.to_string()));
+    assert!(wrapper.contains("load i32, ptr @perry_class_shape_id_"));
+    assert!(wrapper.contains(&format!("@{specialized}(")));
+    assert!(wrapper.contains(&format!("@{index}(")));
+    assert!(ordinary.contains("@js_is_truthy("), "{ordinary}");
+    assert!(
+        !false_default.contains("@js_is_truthy("),
+        "the guarded clone retained the known-false condition:\n{false_default}"
+    );
+    assert!(
+        !false_default.contains("class_field_get"),
+        "the guarded clone reevaluated its already-proved default field:\n{false_default}"
+    );
+}
+
+#[test]
+fn arbitrary_parameter_use_rejects_the_clone() {
+    let mut method = candidate_method();
+    method.body.push(Stmt::Return(Some(Expr::LocalGet(DEFER))));
+    let ir = emit(method);
+    assert!(
+        !ir.contains("$default_false"),
+        "a parameter whose actual false value remains observable was specialized"
+    );
+}

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

@@ -358,6 +358,20 @@ pub(super) fn apply_pshape_inline_policy(
     }
 }
 
+/// Maximum pre-optimization LLVM IR body size admitted to the native-roots
+/// pre-statepoint inliner for a guarded specialization.
+///
+/// HIR statement count is deliberately not used here: one source statement
+/// can lower to a large property/index dispatch lattice.  Sixteen KiB admits
+/// compact exact-receiver and nonnegative-index leaves while rejecting bodies
+/// such as mutation-heavy ECS transitions by nearly an order of magnitude.
+pub(super) const GUARDED_SPECIALIZATION_PREINLINE_MAX_IR_BYTES: usize = 16 * 1024;
+
+#[inline]
+pub(super) fn guarded_specialization_fits_preinline_budget(ir_bytes: usize) -> bool {
+    ir_bytes <= GUARDED_SPECIALIZATION_PREINLINE_MAX_IR_BYTES
+}
+
 /// Maximum total (module-wide) direct call sites a function may have and still
 /// be hinted. This is the anti-bloat backstop: the raised `-inlinehint-threshold`
 /// lifts LLVM's ceiling for a hinted callee at *every* one of its call sites, so
@@ -1470,6 +1484,23 @@ mod sanitize_tests {
     }
 }
 
+#[cfg(test)]
+mod guarded_specialization_preinline_tests {
+    use super::{
+        guarded_specialization_fits_preinline_budget, GUARDED_SPECIALIZATION_PREINLINE_MAX_IR_BYTES,
+    };
+
+    #[test]
+    fn generated_ir_budget_is_inclusive_and_bounded() {
+        assert!(guarded_specialization_fits_preinline_budget(
+            GUARDED_SPECIALIZATION_PREINLINE_MAX_IR_BYTES
+        ));
+        assert!(!guarded_specialization_fits_preinline_budget(
+            GUARDED_SPECIALIZATION_PREINLINE_MAX_IR_BYTES + 1
+        ));
+    }
+}
+
 #[cfg(test)]
 mod resolve_target_triple_tests {
     use super::resolve_target_triple;