size: compile the URL-shape probes only under url-engine

changelog.d/7435-gate-url-shape-probes.md

@@ -0,0 +1 @@
+size: the URL-shape probes on generic value paths (`ToString`, `+` coercion, `JSON.stringify`, `Object.fromEntries`, by-name field set) are now compiled only under `url-engine` — they were pinning the URL parser into every binary. Hello world drops 33,064 bytes (4,708,736 → 4,675,672).

crates/perry-runtime/src/json/replacer.rs

@@ -274,6 +274,7 @@ unsafe fn dispatch_pointer_with_replacer(
     // the URL, so the generic object walk below would trip the circular-
     // structure detector. The href is a plain string, so the emit is identical
     // for compact and pretty walks. See `write_url_href_json`.
+    #[cfg(feature = "url-engine")]
     if crate::url::is_url_object_shape(ptr as *mut crate::ObjectHeader) {
         super::stringify::write_url_href_json(ptr as *mut crate::ObjectHeader, buf);
         return;
@@ -823,6 +824,7 @@ pub(crate) unsafe fn stringify_value_pretty(
         // be walked as a plain object (its `searchParams` back-reference trips
         // the circular-structure detector). Mirrors the compact-path branch in
         // `stringify_object_inner`; see `write_url_href_json`.
+        #[cfg(feature = "url-engine")]
         if crate::url::is_url_object_shape(ptr as *mut crate::ObjectHeader) {
             super::stringify::write_url_href_json(ptr as *mut crate::ObjectHeader, buf);
             return;

crates/perry-runtime/src/object/field_set_by_name/tail.rs

@@ -314,6 +314,11 @@ pub(super) fn set_field_by_name_object_tail(
             }
         }
 
+        // Binary size: this whole arm statically references eight
+        // `js_url_set_*` entry points, which is what kept the URL parser in
+        // every binary. Gated with the other URL shape probes; see
+        // `value/to_string.rs`.
+        #[cfg(feature = "url-engine")]
         if !plan_fast
             && !key.is_null()
             && (key as usize) > 0x10000

crates/perry-runtime/src/object/object_ops/from_entries.rs

@@ -59,6 +59,7 @@ unsafe fn object_from_entries_has_iterator(value: f64, raw: i64, gc_type: Option
         | Some(crate::gc::GC_TYPE_SET) => return true,
         Some(crate::gc::GC_TYPE_OBJECT) => {
             let obj = raw as *mut ObjectHeader;
+            #[cfg(feature = "url-engine")]
             if crate::url::try_read_as_search_params(obj).is_some() {
                 return true;
             }
@@ -107,6 +108,9 @@ unsafe fn object_from_entries_materialize_entries(entries_value: f64) -> *mut Ar
         return crate::map::js_map_entries(raw as *const crate::map::MapHeader);
     }
 
+    // The whole arm exists to serve `URLSearchParams`, so it goes with the
+    // feature rather than leaving `obj` bound for a body that isn't compiled.
+    #[cfg(feature = "url-engine")]
     if gc_type == Some(crate::gc::GC_TYPE_OBJECT) {
         let obj = raw as *mut ObjectHeader;
         if crate::url::try_read_as_search_params(obj).is_some() {

crates/perry-runtime/src/value/dynamic_arith.rs

@@ -276,17 +276,24 @@ unsafe fn to_primitive_default_for_add(value: f64) -> f64 {
     // registry ids, not heap `ObjectHeader`s, so the shape probe would
     // dereference unmapped memory.
     if !crate::value::addr_class::is_handle_band(ptr) {
-        let boxed =
-            f64::from_bits(crate::value::POINTER_TAG | ((ptr as u64) & crate::value::POINTER_MASK));
-        let href = crate::url::url_class::js_url_href_if_url(boxed);
-        if href.to_bits() != crate::value::TAG_UNDEFINED {
-            let s = js_jsvalue_to_string(href);
-            return crate::value::js_nanbox_string(s as i64);
-        }
-        let obj = ptr as *mut crate::object::ObjectHeader;
-        if crate::url::try_read_as_search_params(obj).is_some() {
-            let s = crate::url::search_params::js_url_search_params_to_string(obj);
-            return crate::value::js_nanbox_string(s as i64);
+        // See the matching note in `value/to_string.rs`: shape probes on a
+        // generic path, kept out of non-URL binaries so the parser can strip.
+        // `boxed` is bound inside the gate: it feeds only these probes.
+        #[cfg(feature = "url-engine")]
+        {
+            let boxed = f64::from_bits(
+                crate::value::POINTER_TAG | ((ptr as u64) & crate::value::POINTER_MASK),
+            );
+            let href = crate::url::url_class::js_url_href_if_url(boxed);
+            if href.to_bits() != crate::value::TAG_UNDEFINED {
+                let s = js_jsvalue_to_string(href);
+                return crate::value::js_nanbox_string(s as i64);
+            }
+            let obj = ptr as *mut crate::object::ObjectHeader;
+            if crate::url::try_read_as_search_params(obj).is_some() {
+                let s = crate::url::search_params::js_url_search_params_to_string(obj);
+                return crate::value::js_nanbox_string(s as i64);
+            }
         }
     }
 

crates/perry-runtime/src/value/to_string.rs

@@ -1163,17 +1163,29 @@ pub extern "C" fn js_jsvalue_to_string(value: f64) -> *mut crate::string::String
             // heap `ObjectHeader`s, so the shape check would dereference
             // unmapped memory.
             if !crate::value::addr_class::is_handle_band(ptr as usize) {
-                let boxed = f64::from_bits(POINTER_TAG | ((ptr as u64) & POINTER_MASK));
-                let url_href = crate::url::url_class::js_url_href_if_url(boxed);
-                if url_href.to_bits() != crate::value::TAG_UNDEFINED {
-                    return js_jsvalue_to_string(url_href);
-                }
-                if crate::url::try_read_as_search_params(ptr as *mut crate::object::ObjectHeader)
-                    .is_some()
+                // Binary size: these are SHAPE probes on a generic path, so the
+                // static reference keeps the whole URL class + parser alive in
+                // every binary even though the runtime check can never pass
+                // without `url-engine`. `uses_url` (zero-false-negative by
+                // construction) is what turns that feature on, so a program
+                // with no URL API cannot own a URL or URLSearchParams here.
+                // `boxed` is bound inside the gate: it feeds only these probes.
+                #[cfg(feature = "url-engine")]
                 {
-                    return crate::url::search_params::js_url_search_params_to_string(
+                    let boxed = f64::from_bits(POINTER_TAG | ((ptr as u64) & POINTER_MASK));
+                    let url_href = crate::url::url_class::js_url_href_if_url(boxed);
+                    if url_href.to_bits() != crate::value::TAG_UNDEFINED {
+                        return js_jsvalue_to_string(url_href);
+                    }
+                    if crate::url::try_read_as_search_params(
                         ptr as *mut crate::object::ObjectHeader,
-                    );
+                    )
+                    .is_some()
+                    {
+                        return crate::url::search_params::js_url_search_params_to_string(
+                            ptr as *mut crate::object::ObjectHeader,
+                        );
+                    }
                 }
             }
             // OrdinaryToPrimitive(obj, "string"): the object has no

test-files/test_gap_url_shape_probe_gate.ts

@@ -0,0 +1,19 @@
+// The generic value paths (`ToString`, `+` coercion, `JSON.stringify`,
+// `Object.fromEntries`, and the by-name field setter) each carry a runtime
+// "is this object URL-shaped?" probe. Those probes are compiled only under
+// `url-engine`, because their static references otherwise pin the whole URL
+// parser into every binary. This file uses the URL API, so `uses_url` turns
+// the feature on and every probe below must behave exactly as before.
+
+const u = new URL("https://user:pw@example.com:8443/a/b?x=1&y=2#frag");
+console.log(String(u));
+console.log("" + u);
+console.log(u.hostname, u.port, u.pathname, u.search, u.hash);
+u.pathname = "/changed";
+u.search = "?z=9";
+u.href = "https://other.example/zzz?q=1";
+console.log(u.href, u.hostname);
+const sp = new URLSearchParams("a=1&b=2");
+console.log(String(sp), "" + sp);
+console.log(JSON.stringify({ u, sp: String(sp) }));
+console.log(JSON.stringify(Object.fromEntries(sp)));