* Added autoconf support for cacheline_aligned and SMP_CACHE_BYTES.

git-svn-id: svn://rakshasa.no/libtorrent/trunk/rtorrent@1115 e378c898-3ddf-0310-93e7-cc216c733640

configure.ac

@@ -45,6 +45,9 @@ AC_LANG_POP(C++)
 AC_DEFINE(HAVE_CONFIG_H, 1, true if config.h was included)
 AC_DEFINE(USER_AGENT, [std::string(PACKAGE "/" VERSION "/") + torrent::version()], Http user agent)
 
+AC_CHECK_FUNC(posix_memalign)
+TORRENT_CHECK_CACHELINE()
+
 CC_ATTRIBUTE_UNUSED(
 	AC_DEFINE([__UNUSED], [__attribute__((unused))], [Wrapper around unused attribute]),
 	AC_DEFINE([__UNUSED], [], [Null-wrapper if unused attribute is unsupported])

rak/allocators.h

@@ -39,20 +39,21 @@
 #ifndef RAK_ALLOCATORS_H
 #define RAK_ALLOCATORS_H
 
+#include <cstddef>
 #include <limits>
 #include <stdlib.h>
+#include <sys/types.h>
 
 namespace rak {
 
 template <class T = void*>
 class cacheline_allocator {
 public:
-  typedef std::allocator<T> base_type;
-
   typedef size_t size_type;
   typedef ptrdiff_t difference_type;
   typedef T* pointer;
   typedef const T* const_pointer;
+  typedef const void* const_void_pointer;
   typedef T& reference;
   typedef const T& const_reference;
   typedef T value_type;
@@ -72,10 +73,11 @@ public:
 
   size_type max_size () const throw() { return std::numeric_limits<size_t>::max() / sizeof(T); }
 
-  pointer allocate(size_type num, std::allocator<void>::const_pointer hint = 0) {
+  pointer allocate(size_type num, const_void_pointer hint = 0) { return alloc_size(num*sizeof(T)); }
+
+  static pointer alloc_size(size_type size) {
     pointer ptr = NULL;
-    posix_memalign((void**)&ptr, L1_CACHE_BYTES, num*sizeof(T));
-  
+    posix_memalign((void**)&ptr, LT_SMP_CACHE_BYTES, size);
     return ptr;
   }
 
@@ -102,11 +104,6 @@ bool operator!= (const cacheline_allocator<T1>&, const cacheline_allocator<T2>&)
 //
 
 template <typename T>
-void* operator new(size_t s, rak::cacheline_allocator<T> a) {
-  typename rak::cacheline_allocator<T>::pointer ptr = NULL;
-  posix_memalign((void**)&ptr, L1_CACHE_BYTES, s);
-  
-  return ptr;
-}
+void* operator new(size_t s, rak::cacheline_allocator<T> a) { return a.alloc_size(s); }
 
 #endif // namespace rak

scripts/common.m4

@@ -181,6 +181,26 @@ AC_DEFUN([TORRENT_CHECK_MADVISE], [
   ])
 ])
 
+AC_DEFUN([TORRENT_CHECK_CACHELINE], [
+  AC_MSG_CHECKING(for cacheline)
+
+  AC_COMPILE_IFELSE(
+    [[#include <stdlib.h>
+          #include <linux/cache.h>
+          void* vptr __cacheline_aligned;
+          void f() { posix_memalign(&vptr, SMP_CACHE_BYTES, 42); }
+    ]],
+    [
+      AC_MSG_RESULT(found builtin)
+      AC_DEFINE(LT_SMP_CACHE_BYTES, SMP_CACHE_BYTES, Largest L1 cache size we know of, should work on all archs.)
+      AC_DEFINE(lt_cacheline_aligned, __cacheline_aligned, LibTorrent defined cacheline aligned.)
+    ], [
+      AC_MSG_RESULT(using default 128 bytes)
+      AC_DEFINE(LT_SMP_CACHE_BYTES, 128, Largest L1 cache size we know of, should work on all archs.)
+      AC_DEFINE(lt_cacheline_aligned, __attribute__((__aligned__(LT_SMP_CACHE_BYTES))), LibTorrent defined cacheline aligned.)
+  ])
+])
+
 AC_DEFUN([TORRENT_CHECK_EXECINFO], [
   AC_MSG_CHECKING(for execinfo.h)
 

src/core/poll_manager_select.cc

@@ -40,6 +40,7 @@
 #include <stdexcept>
 #include <unistd.h>
 #include <sys/time.h>
+#include <rak/allocators.h>
 #include <torrent/exceptions.h>
 #include <torrent/poll_select.h>
 #include <torrent/torrent.h>
@@ -53,9 +54,7 @@ PollManagerSelect::PollManagerSelect(torrent::Poll* p) : PollManager(p) {
 #if defined USE_VARIABLE_FDSET
   m_setSize = (m_poll->open_max() + 7) / 8;
 
-  char* buffer;
-  posix_memalign((void**)&buffer, L1_CACHE_BYTES, 3 * m_setSize);
-
+  char* buffer = rak::cacheline_allocator<char>::alloc_size(3 * m_setSize);
   std::memset(buffer, 0, 3 * m_setSize);
 
   m_readSet = (fd_set*)buffer;

src/core/poll_manager_select.h

@@ -45,7 +45,7 @@ namespace torrent {
 
 namespace core {
 
-class __cacheline_aligned PollManagerSelect : public PollManager {
+class lt_cacheline_aligned PollManagerSelect : public PollManager {
 public:
   static PollManagerSelect* create(int maxOpenSockets);
   ~PollManagerSelect();

src/thread_base.cc

@@ -47,9 +47,9 @@
 
 // Temporarly injected into config.h.
 /* temp hack */
-//#define __cacheline_aligned __attribute__((__aligned__(128)))
+//#define lt_cacheline_aligned __attribute__((__aligned__(128)))
 
-class __cacheline_aligned thread_queue_hack {
+class lt_cacheline_aligned thread_queue_hack {
 public:
   typedef ThreadBase::thread_base_func value_type;
   typedef ThreadBase::thread_base_func* iterator;

src/thread_main.h

@@ -42,7 +42,7 @@
 // Check if cacheline aligned with inheritance ends up taking two
 // cachelines.
 
-class __cacheline_aligned ThreadMain : public ThreadBase {
+class lt_cacheline_aligned ThreadMain : public ThreadBase {
 public:
   ThreadMain() {}
   ~ThreadMain();

src/thread_worker.h

@@ -44,7 +44,7 @@
 // Check if cacheline aligned with inheritance ends up taking two
 // cachelines.
 
-class __cacheline_aligned ThreadWorker : public ThreadBase {
+class lt_cacheline_aligned ThreadWorker : public ThreadBase {
 public:
   ThreadWorker();
   ~ThreadWorker();