auto merge of #20367 : retep998/rust/master, r=alexcrichton

Also adjusted some of the FFI definitions because apparently they don't use the long pointer prefix.
Gives a free performance boost because `SRWLock` is several times faster than `CriticalRegion` on every Windows system tested.
Fixes #19962

Author: bors

src/libstd/sys/windows/condvar.rs

@@ -27,16 +27,18 @@ impl Condvar {
 
     #[inline]
     pub unsafe fn wait(&self, mutex: &Mutex) {
-        let r = ffi::SleepConditionVariableCS(self.inner.get(),
-                                              mutex::raw(mutex),
-                                              libc::INFINITE);
+        let r = ffi::SleepConditionVariableSRW(self.inner.get(),
+                                               mutex::raw(mutex),
+                                               libc::INFINITE,
+                                               0);
         debug_assert!(r != 0);
     }
 
     pub unsafe fn wait_timeout(&self, mutex: &Mutex, dur: Duration) -> bool {
-        let r = ffi::SleepConditionVariableCS(self.inner.get(),
-                                              mutex::raw(mutex),
-                                              dur.num_milliseconds() as DWORD);
+        let r = ffi::SleepConditionVariableSRW(self.inner.get(),
+                                               mutex::raw(mutex),
+                                               dur.num_milliseconds() as DWORD,
+                                               0);
         if r == 0 {
             const ERROR_TIMEOUT: DWORD = 0x5B4;
             debug_assert_eq!(os::errno() as uint, ERROR_TIMEOUT as uint);

src/libstd/sys/windows/mutex.rs

@@ -8,73 +8,51 @@
 // option. This file may not be copied, modified, or distributed
 // except according to those terms.
 
-use prelude::v1::*;
-
-use sync::atomic::{AtomicUsize, ATOMIC_USIZE_INIT, Ordering};
-use alloc::{self, heap};
-
-use libc::DWORD;
+use marker::Sync;
+use cell::UnsafeCell;
 use sys::sync as ffi;
 
-const SPIN_COUNT: DWORD = 4000;
+pub struct Mutex { inner: UnsafeCell<ffi::SRWLOCK> }
 
-pub struct Mutex { inner: AtomicUsize }
-
-pub const MUTEX_INIT: Mutex = Mutex { inner: ATOMIC_USIZE_INIT };
+pub const MUTEX_INIT: Mutex = Mutex {
+    inner: UnsafeCell { value: ffi::SRWLOCK_INIT }
+};
 
 unsafe impl Sync for Mutex {}
 
 #[inline]
-pub unsafe fn raw(m: &Mutex) -> ffi::LPCRITICAL_SECTION {
-    m.get()
+pub unsafe fn raw(m: &Mutex) -> ffi::PSRWLOCK {
+    m.inner.get()
 }
 
+// So you might be asking why we're using SRWLock instead of CriticalSection?
+//
+// 1. SRWLock is several times faster than CriticalSection according to benchmarks performed on both
+// Windows 8 and Windows 7.
+//
+// 2. CriticalSection allows recursive locking while SRWLock deadlocks. The Unix implementation
+// deadlocks so consistency is preferred. See #19962 for more details.
+//
+// 3. While CriticalSection is fair and SRWLock is not, the current Rust policy is there there are
+// no guarantees of fairness.
+
 impl Mutex {
     #[inline]
-    pub unsafe fn new() -> Mutex {
-        Mutex { inner: AtomicUsize::new(init_lock() as uint) }
-    }
+    pub unsafe fn new() -> Mutex { MUTEX_INIT }
     #[inline]
     pub unsafe fn lock(&self) {
-        ffi::EnterCriticalSection(self.get())
+        ffi::AcquireSRWLockExclusive(self.inner.get())
     }
     #[inline]
     pub unsafe fn try_lock(&self) -> bool {
-        ffi::TryEnterCriticalSection(self.get()) != 0
+        ffi::TryAcquireSRWLockExclusive(self.inner.get()) != 0
     }
     #[inline]
     pub unsafe fn unlock(&self) {
-        ffi::LeaveCriticalSection(self.get())
+        ffi::ReleaseSRWLockExclusive(self.inner.get())
     }
+    #[inline]
     pub unsafe fn destroy(&self) {
-        let lock = self.inner.swap(0, Ordering::SeqCst);
-        if lock != 0 { free_lock(lock as ffi::LPCRITICAL_SECTION) }
-    }
-
-    unsafe fn get(&self) -> ffi::LPCRITICAL_SECTION {
-        match self.inner.load(Ordering::SeqCst) {
-            0 => {}
-            n => return n as ffi::LPCRITICAL_SECTION
-        }
-        let lock = init_lock();
-        match self.inner.compare_and_swap(0, lock as uint, Ordering::SeqCst) {
-            0 => return lock as ffi::LPCRITICAL_SECTION,
-            _ => {}
-        }
-        free_lock(lock);
-        return self.inner.load(Ordering::SeqCst) as ffi::LPCRITICAL_SECTION;
+        // ...
     }
 }
-
-unsafe fn init_lock() -> ffi::LPCRITICAL_SECTION {
-    let block = heap::allocate(ffi::CRITICAL_SECTION_SIZE, 8)
-                        as ffi::LPCRITICAL_SECTION;
-    if block.is_null() { alloc::oom() }
-    ffi::InitializeCriticalSectionAndSpinCount(block, SPIN_COUNT);
-    return block;
-}
-
-unsafe fn free_lock(h: ffi::LPCRITICAL_SECTION) {
-    ffi::DeleteCriticalSection(h);
-    heap::deallocate(h as *mut _, ffi::CRITICAL_SECTION_SIZE, 8);
-}

src/libstd/sys/windows/sync.rs

@@ -8,17 +8,12 @@
 // option. This file may not be copied, modified, or distributed
 // except according to those terms.
 
-use libc::{BOOL, DWORD, c_void, LPVOID};
+use libc::{BOOL, DWORD, c_void, LPVOID, c_ulong};
 use libc::types::os::arch::extra::BOOLEAN;
 
-pub type LPCRITICAL_SECTION = *mut c_void;
-pub type LPCONDITION_VARIABLE = *mut CONDITION_VARIABLE;
-pub type LPSRWLOCK = *mut SRWLOCK;
-
-#[cfg(target_arch = "x86")]
-pub const CRITICAL_SECTION_SIZE: uint = 24;
-#[cfg(target_arch = "x86_64")]
-pub const CRITICAL_SECTION_SIZE: uint = 40;
+pub type PCONDITION_VARIABLE = *mut CONDITION_VARIABLE;
+pub type PSRWLOCK = *mut SRWLOCK;
+pub type ULONG = c_ulong;
 
 #[repr(C)]
 pub struct CONDITION_VARIABLE { pub ptr: LPVOID }
@@ -31,28 +26,19 @@ pub const CONDITION_VARIABLE_INIT: CONDITION_VARIABLE = CONDITION_VARIABLE {
 pub const SRWLOCK_INIT: SRWLOCK = SRWLOCK { ptr: 0 as *mut _ };
 
 extern "system" {
-    // critical sections
-    pub fn InitializeCriticalSectionAndSpinCount(
-                    lpCriticalSection: LPCRITICAL_SECTION,
-                    dwSpinCount: DWORD) -> BOOL;
-    pub fn DeleteCriticalSection(lpCriticalSection: LPCRITICAL_SECTION);
-    pub fn EnterCriticalSection(lpCriticalSection: LPCRITICAL_SECTION);
-    pub fn LeaveCriticalSection(lpCriticalSection: LPCRITICAL_SECTION);
-    pub fn TryEnterCriticalSection(lpCriticalSection: LPCRITICAL_SECTION) -> BOOL;
-
     // condition variables
-    pub fn SleepConditionVariableCS(ConditionVariable: LPCONDITION_VARIABLE,
-                                    CriticalSection: LPCRITICAL_SECTION,
-                                    dwMilliseconds: DWORD) -> BOOL;
-    pub fn WakeConditionVariable(ConditionVariable: LPCONDITION_VARIABLE);
-    pub fn WakeAllConditionVariable(ConditionVariable: LPCONDITION_VARIABLE);
+    pub fn SleepConditionVariableSRW(ConditionVariable: PCONDITION_VARIABLE,
+                                     SRWLock: PSRWLOCK,
+                                     dwMilliseconds: DWORD,
+                                     Flags: ULONG) -> BOOL;
+    pub fn WakeConditionVariable(ConditionVariable: PCONDITION_VARIABLE);
+    pub fn WakeAllConditionVariable(ConditionVariable: PCONDITION_VARIABLE);
 
     // slim rwlocks
-    pub fn AcquireSRWLockExclusive(SRWLock: LPSRWLOCK);
-    pub fn AcquireSRWLockShared(SRWLock: LPSRWLOCK);
-    pub fn ReleaseSRWLockExclusive(SRWLock: LPSRWLOCK);
-    pub fn ReleaseSRWLockShared(SRWLock: LPSRWLOCK);
-    pub fn TryAcquireSRWLockExclusive(SRWLock: LPSRWLOCK) -> BOOLEAN;
-    pub fn TryAcquireSRWLockShared(SRWLock: LPSRWLOCK) -> BOOLEAN;
+    pub fn AcquireSRWLockExclusive(SRWLock: PSRWLOCK);
+    pub fn AcquireSRWLockShared(SRWLock: PSRWLOCK);
+    pub fn ReleaseSRWLockExclusive(SRWLock: PSRWLOCK);
+    pub fn ReleaseSRWLockShared(SRWLock: PSRWLOCK);
+    pub fn TryAcquireSRWLockExclusive(SRWLock: PSRWLOCK) -> BOOLEAN;
+    pub fn TryAcquireSRWLockShared(SRWLock: PSRWLOCK) -> BOOLEAN;
 }
-