Auto merge of #115388 - Zoxc:sharded-lock, r=SparrowLii

Add optimized lock methods for `Sharded` and refactor `Lock`

This adds methods to `Sharded` which pick a shard and also locks it. These branch on parallelism just once instead of twice, improving performance.

Benchmark for `cfg(parallel_compiler)` and 1 thread:
<table><tr><td rowspan="2">Benchmark</td><td colspan="1"><b>Before</b></th><td colspan="2"><b>After</b></th></tr><tr><td align="right">Time</td><td align="right">Time</td><td align="right">%</th></tr><tr><td>🟣 <b>clap</b>:check</td><td align="right">1.6461s</td><td align="right">1.6345s</td><td align="right"> -0.70%</td></tr><tr><td>🟣 <b>hyper</b>:check</td><td align="right">0.2414s</td><td align="right">0.2394s</td><td align="right"> -0.83%</td></tr><tr><td>🟣 <b>regex</b>:check</td><td align="right">0.9205s</td><td align="right">0.9143s</td><td align="right"> -0.67%</td></tr><tr><td>🟣 <b>syn</b>:check</td><td align="right">1.4981s</td><td align="right">1.4869s</td><td align="right"> -0.75%</td></tr><tr><td>🟣 <b>syntex_syntax</b>:check</td><td align="right">5.7629s</td><td align="right">5.7256s</td><td align="right"> -0.65%</td></tr><tr><td>Total</td><td align="right">10.0690s</td><td align="right">10.0008s</td><td align="right"> -0.68%</td></tr><tr><td>Summary</td><td align="right">1.0000s</td><td align="right">0.9928s</td><td align="right"> -0.72%</td></tr></table>

cc `@SparrowLii`

1 files changed, 187 insertions, 188 deletions

diff --git a/compiler/rustc_data_structures/src/sync/lock.rs b/compiler/rustc_data_structures/src/sync/lock.rs
index 62cd1b993de..339aebbf81a 100644
--- a/compiler/rustc_data_structures/src/sync/lock.rs
+++ b/compiler/rustc_data_structures/src/sync/lock.rs
@@ -3,224 +3,229 @@
 //!
 //! When `cfg(parallel_compiler)` is not set, the lock is instead a wrapper around `RefCell`.
 
-#[cfg(not(parallel_compiler))]
-use std::cell::RefCell;
-#[cfg(parallel_compiler)]
-use {
-    crate::cold_path,
-    crate::sync::DynSend,
-    crate::sync::DynSync,
-    parking_lot::lock_api::RawMutex,
-    std::cell::Cell,
-    std::cell::UnsafeCell,
-    std::fmt,
-    std::intrinsics::{likely, unlikely},
-    std::marker::PhantomData,
-    std::mem::ManuallyDrop,
-    std::ops::{Deref, DerefMut},
-};
+#![allow(dead_code)]
 
-#[cfg(not(parallel_compiler))]
-pub use std::cell::RefMut as LockGuard;
+use std::fmt;
 
+#[cfg(parallel_compiler)]
+pub use maybe_sync::*;
 #[cfg(not(parallel_compiler))]
-#[derive(Debug)]
-pub struct Lock<T>(RefCell<T>);
+pub use no_sync::*;
 
-#[cfg(not(parallel_compiler))]
-impl<T> Lock<T> {
-    #[inline(always)]
-    pub fn new(inner: T) -> Self {
-        Lock(RefCell::new(inner))
-    }
+#[derive(Clone, Copy, PartialEq)]
+pub enum Mode {
+    NoSync,
+    Sync,
+}
 
-    #[inline(always)]
-    pub fn into_inner(self) -> T {
-        self.0.into_inner()
+mod maybe_sync {
+    use super::Mode;
+    use crate::sync::mode;
+    #[cfg(parallel_compiler)]
+    use crate::sync::{DynSend, DynSync};
+    use parking_lot::lock_api::RawMutex as _;
+    use parking_lot::RawMutex;
+    use std::cell::Cell;
+    use std::cell::UnsafeCell;
+    use std::intrinsics::unlikely;
+    use std::marker::PhantomData;
+    use std::mem::ManuallyDrop;
+    use std::ops::{Deref, DerefMut};
+
+    /// A guard holding mutable access to a `Lock` which is in a locked state.
+    #[must_use = "if unused the Lock will immediately unlock"]
+    pub struct LockGuard<'a, T> {
+        lock: &'a Lock<T>,
+        marker: PhantomData<&'a mut T>,
+
+        /// The syncronization mode of the lock. This is explicitly passed to let LLVM relate it
+        /// to the original lock operation.
+        mode: Mode,
     }
 
-    #[inline(always)]
-    pub fn get_mut(&mut self) -> &mut T {
-        self.0.get_mut()
+    impl<'a, T: 'a> Deref for LockGuard<'a, T> {
+        type Target = T;
+        #[inline]
+        fn deref(&self) -> &T {
+            // SAFETY: We have shared access to the mutable access owned by this type,
+            // so we can give out a shared reference.
+            unsafe { &*self.lock.data.get() }
+        }
     }
 
-    #[inline(always)]
-    pub fn try_lock(&self) -> Option<LockGuard<'_, T>> {
-        self.0.try_borrow_mut().ok()
+    impl<'a, T: 'a> DerefMut for LockGuard<'a, T> {
+        #[inline]
+        fn deref_mut(&mut self) -> &mut T {
+            // SAFETY: We have mutable access to the data so we can give out a mutable reference.
+            unsafe { &mut *self.lock.data.get() }
+        }
     }
 
-    #[inline(always)]
-    #[track_caller]
-    pub fn lock(&self) -> LockGuard<'_, T> {
-        self.0.borrow_mut()
+    impl<'a, T: 'a> Drop for LockGuard<'a, T> {
+        #[inline]
+        fn drop(&mut self) {
+            // SAFETY (union access): We get `self.mode` from the lock operation so it is consistent
+            // with the `lock.mode` state. This means we access the right union fields.
+            match self.mode {
+                Mode::NoSync => {
+                    let cell = unsafe { &self.lock.mode_union.no_sync };
+                    debug_assert_eq!(cell.get(), true);
+                    cell.set(false);
+                }
+                // SAFETY (unlock): We know that the lock is locked as this type is a proof of that.
+                Mode::Sync => unsafe { self.lock.mode_union.sync.unlock() },
+            }
+        }
     }
-}
 
-/// A guard holding mutable access to a `Lock` which is in a locked state.
-#[cfg(parallel_compiler)]
-#[must_use = "if unused the Lock will immediately unlock"]
-pub struct LockGuard<'a, T> {
-    lock: &'a Lock<T>,
-    marker: PhantomData<&'a mut T>,
-}
+    union ModeUnion {
+        /// Indicates if the cell is locked. Only used if `Lock.mode` is `NoSync`.
+        no_sync: ManuallyDrop<Cell<bool>>,
 
-#[cfg(parallel_compiler)]
-impl<'a, T: 'a> Deref for LockGuard<'a, T> {
-    type Target = T;
-    #[inline]
-    fn deref(&self) -> &T {
-        // SAFETY: We have shared access to the mutable access owned by this type,
-        // so we can give out a shared reference.
-        unsafe { &*self.lock.data.get() }
+        /// A lock implementation that's only used if `Lock.mode` is `Sync`.
+        sync: ManuallyDrop<RawMutex>,
     }
-}
 
-#[cfg(parallel_compiler)]
-impl<'a, T: 'a> DerefMut for LockGuard<'a, T> {
-    #[inline]
-    fn deref_mut(&mut self) -> &mut T {
-        // SAFETY: We have mutable access to the data so we can give out a mutable reference.
-        unsafe { &mut *self.lock.data.get() }
-    }
-}
+    /// The value representing a locked state for the `Cell`.
+    const LOCKED: bool = true;
 
-#[cfg(parallel_compiler)]
-impl<'a, T: 'a> Drop for LockGuard<'a, T> {
-    #[inline]
-    fn drop(&mut self) {
-        // SAFETY: We know that the lock is in a locked
-        // state because it is a invariant of this type.
-        unsafe { self.lock.raw.unlock() };
-    }
-}
+    /// A lock which only uses synchronization if `might_be_dyn_thread_safe` is true.
+    /// It implements `DynSend` and `DynSync` instead of the typical `Send` and `Sync`.
+    pub struct Lock<T> {
+        /// Indicates if synchronization is used via `mode_union.sync` if it's `Sync`, or if a
+        /// not thread safe cell is used via `mode_union.no_sync` if it's `NoSync`.
+        /// This is set on initialization and never changed.
+        mode: Mode,
 
-#[cfg(parallel_compiler)]
-union LockRawUnion {
-    /// Indicates if the cell is locked. Only used if `LockRaw.sync` is false.
-    cell: ManuallyDrop<Cell<bool>>,
+        mode_union: ModeUnion,
+        data: UnsafeCell<T>,
+    }
 
-    /// A lock implementation that's only used if `LockRaw.sync` is true.
-    lock: ManuallyDrop<parking_lot::RawMutex>,
-}
+    impl<T> Lock<T> {
+        #[inline(always)]
+        pub fn new(inner: T) -> Self {
+            let (mode, mode_union) = if unlikely(mode::might_be_dyn_thread_safe()) {
+                // Create the lock with synchronization enabled using the `RawMutex` type.
+                (Mode::Sync, ModeUnion { sync: ManuallyDrop::new(RawMutex::INIT) })
+            } else {
+                // Create the lock with synchronization disabled.
+                (Mode::NoSync, ModeUnion { no_sync: ManuallyDrop::new(Cell::new(!LOCKED)) })
+            };
+            Lock { mode, mode_union, data: UnsafeCell::new(inner) }
+        }
 
-/// A raw lock which only uses synchronization if `might_be_dyn_thread_safe` is true.
-/// It contains no associated data and is used in the implementation of `Lock` which does have such data.
-///
-/// A manual implementation of a tagged union is used with the `sync` field and the `LockRawUnion` instead
-/// of using enums as it results in better code generation.
-#[cfg(parallel_compiler)]
-struct LockRaw {
-    /// Indicates if synchronization is used via `opt.lock` if true,
-    /// or if a non-thread safe cell is used via `opt.cell`. This is set on initialization and never changed.
-    sync: bool,
-    opt: LockRawUnion,
-}
+        #[inline(always)]
+        pub fn into_inner(self) -> T {
+            self.data.into_inner()
+        }
 
-#[cfg(parallel_compiler)]
-impl LockRaw {
-    fn new() -> Self {
-        if unlikely(super::mode::might_be_dyn_thread_safe()) {
-            // Create the lock with synchronization enabled using the `RawMutex` type.
-            LockRaw {
-                sync: true,
-                opt: LockRawUnion { lock: ManuallyDrop::new(parking_lot::RawMutex::INIT) },
-            }
-        } else {
-            // Create the lock with synchronization disabled.
-            LockRaw { sync: false, opt: LockRawUnion { cell: ManuallyDrop::new(Cell::new(false)) } }
+        #[inline(always)]
+        pub fn get_mut(&mut self) -> &mut T {
+            self.data.get_mut()
         }
-    }
 
-    #[inline(always)]
-    fn try_lock(&self) -> bool {
-        // SAFETY: This is safe since the union fields are used in accordance with `self.sync`.
-        unsafe {
-            if likely(!self.sync) {
-                if self.opt.cell.get() {
-                    false
-                } else {
-                    self.opt.cell.set(true);
-                    true
+        #[inline(always)]
+        pub fn try_lock(&self) -> Option<LockGuard<'_, T>> {
+            let mode = self.mode;
+            // SAFETY: This is safe since the union fields are used in accordance with `self.mode`.
+            match mode {
+                Mode::NoSync => {
+                    let cell = unsafe { &self.mode_union.no_sync };
+                    let was_unlocked = cell.get() != LOCKED;
+                    if was_unlocked {
+                        cell.set(LOCKED);
+                    }
+                    was_unlocked
                 }
-            } else {
-                self.opt.lock.try_lock()
+                Mode::Sync => unsafe { self.mode_union.sync.try_lock() },
             }
+            .then(|| LockGuard { lock: self, marker: PhantomData, mode })
         }
-    }
 
-    #[inline(always)]
-    fn lock(&self) {
-        if super::ERROR_CHECKING {
-            // We're in the debugging mode, so assert that the lock is not held so we
-            // get a panic instead of waiting for the lock.
-            assert_eq!(self.try_lock(), true, "lock must not be hold");
-        } else {
-            // SAFETY: This is safe since the union fields are used in accordance with `self.sync`.
+        /// This acquires the lock assuming syncronization is in a specific mode.
+        ///
+        /// Safety
+        /// This method must only be called with `Mode::Sync` if `might_be_dyn_thread_safe` was
+        /// true on lock creation.
+        #[inline(always)]
+        #[track_caller]
+        pub unsafe fn lock_assume(&self, mode: Mode) -> LockGuard<'_, T> {
+            #[inline(never)]
+            #[track_caller]
+            #[cold]
+            fn lock_held() -> ! {
+                panic!("lock was already held")
+            }
+
+            // SAFETY: This is safe since the union fields are used in accordance with `mode`
+            // which also must match `self.mode` due to the safety precondition.
             unsafe {
-                if likely(!self.sync) {
-                    if unlikely(self.opt.cell.replace(true)) {
-                        cold_path(|| panic!("lock was already held"))
+                match mode {
+                    Mode::NoSync => {
+                        if unlikely(self.mode_union.no_sync.replace(LOCKED) == LOCKED) {
+                            lock_held()
+                        }
                     }
-                } else {
-                    self.opt.lock.lock();
+                    Mode::Sync => self.mode_union.sync.lock(),
                 }
             }
+            LockGuard { lock: self, marker: PhantomData, mode }
         }
-    }
 
-    /// This unlocks the lock.
-    ///
-    /// Safety
-    /// This method may only be called if the lock is currently held.
-    #[inline(always)]
-    unsafe fn unlock(&self) {
-        // SAFETY: The union use is safe since the union fields are used in accordance with
-        // `self.sync` and the `unlock` method precondition is upheld by the caller.
-        unsafe {
-            if likely(!self.sync) {
-                debug_assert_eq!(self.opt.cell.get(), true);
-                self.opt.cell.set(false);
-            } else {
-                self.opt.lock.unlock();
-            }
+        #[inline(always)]
+        #[track_caller]
+        pub fn lock(&self) -> LockGuard<'_, T> {
+            unsafe { self.lock_assume(self.mode) }
         }
     }
-}
 
-/// A lock which only uses synchronization if `might_be_dyn_thread_safe` is true.
-/// It implements `DynSend` and `DynSync` instead of the typical `Send` and `Sync`.
-#[cfg(parallel_compiler)]
-pub struct Lock<T> {
-    raw: LockRaw,
-    data: UnsafeCell<T>,
+    #[cfg(parallel_compiler)]
+    unsafe impl<T: DynSend> DynSend for Lock<T> {}
+    #[cfg(parallel_compiler)]
+    unsafe impl<T: DynSend> DynSync for Lock<T> {}
 }
 
-#[cfg(parallel_compiler)]
-impl<T> Lock<T> {
-    #[inline(always)]
-    pub fn new(inner: T) -> Self {
-        Lock { raw: LockRaw::new(), data: UnsafeCell::new(inner) }
-    }
+mod no_sync {
+    use super::Mode;
+    use std::cell::RefCell;
 
-    #[inline(always)]
-    pub fn into_inner(self) -> T {
-        self.data.into_inner()
-    }
+    pub use std::cell::RefMut as LockGuard;
 
-    #[inline(always)]
-    pub fn get_mut(&mut self) -> &mut T {
-        self.data.get_mut()
-    }
+    pub struct Lock<T>(RefCell<T>);
 
-    #[inline(always)]
-    pub fn try_lock(&self) -> Option<LockGuard<'_, T>> {
-        if self.raw.try_lock() { Some(LockGuard { lock: self, marker: PhantomData }) } else { None }
-    }
+    impl<T> Lock<T> {
+        #[inline(always)]
+        pub fn new(inner: T) -> Self {
+            Lock(RefCell::new(inner))
+        }
 
-    #[inline(always)]
-    pub fn lock(&self) -> LockGuard<'_, T> {
-        self.raw.lock();
-        LockGuard { lock: self, marker: PhantomData }
+        #[inline(always)]
+        pub fn into_inner(self) -> T {
+            self.0.into_inner()
+        }
+
+        #[inline(always)]
+        pub fn get_mut(&mut self) -> &mut T {
+            self.0.get_mut()
+        }
+
+        #[inline(always)]
+        pub fn try_lock(&self) -> Option<LockGuard<'_, T>> {
+            self.0.try_borrow_mut().ok()
+        }
+
+        #[inline(always)]
+        #[track_caller]
+        // This is unsafe to match the API for the `parallel_compiler` case.
+        pub unsafe fn lock_assume(&self, _mode: Mode) -> LockGuard<'_, T> {
+            self.0.borrow_mut()
+        }
+
+        #[inline(always)]
+        #[track_caller]
+        pub fn lock(&self) -> LockGuard<'_, T> {
+            self.0.borrow_mut()
+        }
     }
 }
 
@@ -244,12 +249,13 @@ impl<T> Lock<T> {
     }
 }
 
-#[cfg(parallel_compiler)]
-unsafe impl<T: DynSend> DynSend for Lock<T> {}
-#[cfg(parallel_compiler)]
-unsafe impl<T: DynSend> DynSync for Lock<T> {}
+impl<T: Default> Default for Lock<T> {
+    #[inline]
+    fn default() -> Self {
+        Lock::new(T::default())
+    }
+}
 
-#[cfg(parallel_compiler)]
 impl<T: fmt::Debug> fmt::Debug for Lock<T> {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
         match self.try_lock() {
@@ -267,10 +273,3 @@ impl<T: fmt::Debug> fmt::Debug for Lock<T> {
         }
     }
 }
-
-impl<T: Default> Default for Lock<T> {
-    #[inline]
-    fn default() -> Self {
-        Lock::new(T::default())
-    }
-}