Auto merge of #22487 - emilio:single-layout-thread-pool, r=jdm

style: Make Servo use a single thread-pool for layout-related tasks per process.

Instead of per-document. This also allows to reuse this thread-pool if needed
for other stuff, like parallel CSS parsing (#22478), and to share more code with
Gecko, which is always nice.

<!-- Reviewable:start -->
---
This change is [<img src="https://reviewable.io/review_button.svg" height="34" align="absmiddle" alt="Reviewable"/>](https://reviewable.io/reviews/servo/servo/22487)
<!-- Reviewable:end -->

components/style/Cargo.toml

@@ -16,8 +16,7 @@ path = "lib.rs"
 doctest = false
 
 [features]
-gecko = ["num_cpus",
-         "style_traits/gecko", "fallible/known_system_malloc"]
+gecko = ["style_traits/gecko", "fallible/known_system_malloc"]
 use_bindgen = ["bindgen", "regex", "toml"]
 servo = ["serde", "style_traits/servo", "servo_atoms", "servo_config", "html5ever",
          "cssparser/serde", "encoding_rs", "malloc_size_of/servo", "arrayvec/use_union",
@@ -47,7 +46,7 @@ log = "0.4"
 malloc_size_of = { path = "../malloc_size_of" }
 malloc_size_of_derive = { path = "../malloc_size_of_derive" }
 matches = "0.1"
-num_cpus = {version = "1.1.0", optional = true}
+num_cpus = {version = "1.1.0"}
 num-integer = "0.1"
 num-traits = "0.2"
 num-derive = "0.2"

components/style/gecko/mod.rs

@@ -10,7 +10,6 @@ mod non_ts_pseudo_class_list;
 pub mod arc_types;
 pub mod conversions;
 pub mod data;
-pub mod global_style_data;
 pub mod media_features;
 pub mod media_queries;
 pub mod pseudo_element;

components/style/global_style_data.rs

@@ -5,15 +5,13 @@
 //! Global style data
 
 use crate::context::StyleSystemOptions;
+#[cfg(feature = "gecko")]
 use crate::gecko_bindings::bindings;
 use crate::parallel::STYLE_THREAD_STACK_SIZE_KB;
 use crate::shared_lock::SharedRwLock;
 use crate::thread_state;
-use num_cpus;
 use rayon;
-use std::cmp;
 use std::env;
-use std::ffi::CString;
 
 /// Global style data
 pub struct GlobalStyleData {
@@ -37,20 +35,22 @@ fn thread_name(index: usize) -> String {
     format!("StyleThread#{}", index)
 }
 
-fn thread_startup(index: usize) {
+fn thread_startup(_index: usize) {
     thread_state::initialize_layout_worker_thread();
+    #[cfg(feature = "gecko")]
     unsafe {
+        use std::ffi::CString;
+
         bindings::Gecko_SetJemallocThreadLocalArena(true);
-    }
-    let name = thread_name(index);
-    let name = CString::new(name).unwrap();
-    unsafe {
+        let name = thread_name(_index);
+        let name = CString::new(name).unwrap();
         // Gecko_RegisterProfilerThread copies the passed name here.
         bindings::Gecko_RegisterProfilerThread(name.as_ptr());
     }
 }
 
 fn thread_shutdown(_: usize) {
+    #[cfg(feature = "gecko")]
     unsafe {
         bindings::Gecko_UnregisterProfilerThread();
         bindings::Gecko_SetJemallocThreadLocalArena(false);
@@ -64,12 +64,26 @@ lazy_static! {
             .map(|s| s.parse::<usize>().expect("invalid STYLO_THREADS value"));
         let mut num_threads = match stylo_threads {
             Ok(num) => num,
-            // The default heuristic is num_virtual_cores * .75. This gives us
-            // three threads on a hyper-threaded dual core, and six threads on
-            // a hyper-threaded quad core. The performance benefit of additional
-            // threads seems to level off at around six, so we cap it there on
-            // many-core machines (see bug 1431285 comment 14).
-            _ => cmp::min(cmp::max(num_cpus::get() * 3 / 4, 1), 6),
+            #[cfg(feature = "servo")]
+            _ => {
+                // We always set this pref on startup, before layout or script
+                // have had a chance of accessing (and thus creating) the
+                // thread-pool.
+                use servo_config::prefs::PREFS;
+                PREFS.get("layout.threads").as_u64().unwrap() as usize
+            }
+            #[cfg(feature = "gecko")]
+            _ => {
+                // The default heuristic is num_virtual_cores * .75. This gives
+                // us three threads on a hyper-threaded dual core, and six
+                // threads on a hyper-threaded quad core. The performance
+                // benefit of additional threads seems to level off at around
+                // six, so we cap it there on many-core machines
+                // (see bug 1431285 comment 14).
+                use num_cpus;
+                use std::cmp;
+                cmp::min(cmp::max(num_cpus::get() * 3 / 4, 1), 6)
+            }
         };
 
         // If num_threads is one, there's no point in creating a thread pool, so

components/style/lib.rs

@@ -133,6 +133,7 @@ pub mod font_metrics;
 #[cfg(feature = "gecko")]
 #[allow(unsafe_code)]
 pub mod gecko_bindings;
+pub mod global_style_data;
 pub mod hash;
 pub mod invalidation;
 #[allow(missing_docs)] // TODO.