* team.c (gomp_thread_start): Don't clear ts.static_trip here.

	* loop.c (gomp_loop_static_start, gomp_loop_dynamic_start): Clear
	ts.static_trip here.
	* work.c (gomp_work_share_start): Don't clear ts.static_trip here.

	* libgomp.h: Include ptrlock.h.
	(struct gomp_work_share): Reshuffle fields.  Add next_alloc,
	next_ws, next_free and inline_ordered_team_ids fields, change
	ordered_team_ids into pointer from flexible array member.
	(struct gomp_team_state): Add last_work_share field, remove
	work_share_generation.
	(struct gomp_team): Remove work_share_lock, generation_mask,
	oldest_live_gen, num_live_gen and init_work_shares fields, add
	work work_share_list_alloc, work_share_list_free and work_share_chunk
	fields.  Change work_shares from pointer to pointers into an array.
	(gomp_new_team): New prototype.
	(gomp_team_start): Change type of last argument.
	(gomp_new_work_share): Removed.
	(gomp_init_work_share, gomp_fini_work_share): New prototypes.
	(gomp_work_share_init_done): New static inline.
	* team.c (gomp_thread_start): Clear ts.last_work_share, don't clear
	ts.work_share_generation.
	(new_team): Removed.
	(gomp_new_team): New function.
	(free_team): Free gomp_work_share blocks chained through next_alloc,
	instead of freeing work_shares and destroying work_share_lock.
	(gomp_team_start): Change last argument from ws to team, don't create
	new team, set ts.work_share to &team->work_shares[0] and clear
	ts.last_work_share.  Don't clear ts.work_share_generation.
	(gomp_team_end): Call gomp_fini_work_share.
	* work.c (gomp_new_work_share): Removed.
	(alloc_work_share, gomp_init_work_share, gomp_fini_work_share): New
	functions.
	(free_work_share): Add team argument.  Call gomp_fini_work_share
	and then either free ws if orphaned, or put it into
	work_share_list_free list of the current team.
	(gomp_work_share_start, gomp_work_share_end,
	gomp_work_share_end_nowait): Rewritten.
	* sections.c (GOMP_sections_start): Call gomp_work_share_init_done
	after gomp_sections_init.  If HAVE_SYNC_BUILTINS, call
	gomp_iter_dynamic_next instead of the _locked variant and don't take
	lock around it, otherwise acquire it before calling
	gomp_iter_dynamic_next_locked.
	(GOMP_sections_next): If HAVE_SYNC_BUILTINS, call
	gomp_iter_dynamic_next instead of the _locked variant and don't take
	lock around it.
	(GOMP_parallel_sections_start): Call gomp_new_team instead of
	gomp_new_work_share.  Call gomp_sections_init on &team->work_shares[0].
	Adjust gomp_team_start caller.
	* loop.c (gomp_loop_static_start, gomp_loop_ordered_static_start): Call
	gomp_work_share_init_done after gomp_loop_init.  Don't unlock ws->lock.
	(gomp_loop_dynamic_start, gomp_loop_guided_start): Call
	gomp_work_share_init_done after gomp_loop_init.  If HAVE_SYNC_BUILTINS,
	don't unlock ws->lock, otherwise lock it.
	(gomp_loop_ordered_dynamic_start, gomp_loop_ordered_guided_start): Call
	gomp_work_share_init_done after gomp_loop_init.  Lock ws->lock.
	(gomp_parallel_loop_start): Call gomp_new_team instead of
	gomp_new_work_share.  Call gomp_loop_init on &team->work_shares[0].
	Adjust gomp_team_start caller.
	* single.c (GOMP_single_start, GOMP_single_copy_start): Call
	gomp_work_share_init_done if gomp_work_share_start returned true.
	Don't unlock ws->lock.
	* parallel.c (gomp_resolve_num_threads): Fix non-HAVE_SYNC_BUILTINS
	case.
	(GOMP_parallel_start): Call gomp_new_team and pass that as last
	argument to gomp_team_start.
	* config/linux/ptrlock.c: New file.
	* config/linux/ptrlock.h: New file.
	* config/posix/ptrlock.c: New file.
	* config/posix/ptrlock.h: New file.
	* Makefile.am (libgomp_la_SOURCES): Add ptrlock.c.
	* Makefile.in: Regenerated.
	* testsuite/Makefile.in: Regenerated.

	* testsuite/libgomp.c++/collapse-1.C (main): Make k private in second
	omp for loop.


git-svn-id: https://gcc.gnu.org/svn/gcc/branches/gomp-3_0-branch@133506 138bc75d-0d04-0410-961f-82ee72b054a4

16 files changed, 616 insertions, 221 deletions

diff --git a/libgomp/ChangeLog.gomp b/libgomp/ChangeLog.gomp
index 7f5987578e4..3919787141c 100644
--- a/libgomp/ChangeLog.gomp
+++ b/libgomp/ChangeLog.gomp
@@ -1,3 +1,84 @@
+2008-03-25  Jakub Jelinek  <jakub@redhat.com>
+
+	* team.c (gomp_thread_start): Don't clear ts.static_trip here.
+	* loop.c (gomp_loop_static_start, gomp_loop_dynamic_start): Clear
+	ts.static_trip here.
+	* work.c (gomp_work_share_start): Don't clear ts.static_trip here.
+
+2008-03-21  Jakub Jelinek  <jakub@redhat.com>
+
+	* libgomp.h: Include ptrlock.h.
+	(struct gomp_work_share): Reshuffle fields.  Add next_alloc,
+	next_ws, next_free and inline_ordered_team_ids fields, change
+	ordered_team_ids into pointer from flexible array member.
+	(struct gomp_team_state): Add last_work_share field, remove
+	work_share_generation.
+	(struct gomp_team): Remove work_share_lock, generation_mask,
+	oldest_live_gen, num_live_gen and init_work_shares fields, add
+	work work_share_list_alloc, work_share_list_free and work_share_chunk
+	fields.  Change work_shares from pointer to pointers into an array.
+	(gomp_new_team): New prototype.
+	(gomp_team_start): Change type of last argument.
+	(gomp_new_work_share): Removed.
+	(gomp_init_work_share, gomp_fini_work_share): New prototypes.
+	(gomp_work_share_init_done): New static inline.
+	* team.c (gomp_thread_start): Clear ts.last_work_share, don't clear
+	ts.work_share_generation.
+	(new_team): Removed.
+	(gomp_new_team): New function.
+	(free_team): Free gomp_work_share blocks chained through next_alloc,
+	instead of freeing work_shares and destroying work_share_lock.
+	(gomp_team_start): Change last argument from ws to team, don't create
+	new team, set ts.work_share to &team->work_shares[0] and clear
+	ts.last_work_share.  Don't clear ts.work_share_generation.
+	(gomp_team_end): Call gomp_fini_work_share.
+	* work.c (gomp_new_work_share): Removed.
+	(alloc_work_share, gomp_init_work_share, gomp_fini_work_share): New
+	functions.
+	(free_work_share): Add team argument.  Call gomp_fini_work_share
+	and then either free ws if orphaned, or put it into
+	work_share_list_free list of the current team.
+	(gomp_work_share_start, gomp_work_share_end,
+	gomp_work_share_end_nowait): Rewritten.
+	* sections.c (GOMP_sections_start): Call gomp_work_share_init_done
+	after gomp_sections_init.  If HAVE_SYNC_BUILTINS, call
+	gomp_iter_dynamic_next instead of the _locked variant and don't take
+	lock around it, otherwise acquire it before calling
+	gomp_iter_dynamic_next_locked.
+	(GOMP_sections_next): If HAVE_SYNC_BUILTINS, call
+	gomp_iter_dynamic_next instead of the _locked variant and don't take
+	lock around it.
+	(GOMP_parallel_sections_start): Call gomp_new_team instead of
+	gomp_new_work_share.  Call gomp_sections_init on &team->work_shares[0].
+	Adjust gomp_team_start caller.
+	* loop.c (gomp_loop_static_start, gomp_loop_ordered_static_start): Call
+	gomp_work_share_init_done after gomp_loop_init.  Don't unlock ws->lock.
+	(gomp_loop_dynamic_start, gomp_loop_guided_start): Call
+	gomp_work_share_init_done after gomp_loop_init.  If HAVE_SYNC_BUILTINS,
+	don't unlock ws->lock, otherwise lock it.
+	(gomp_loop_ordered_dynamic_start, gomp_loop_ordered_guided_start): Call
+	gomp_work_share_init_done after gomp_loop_init.  Lock ws->lock.
+	(gomp_parallel_loop_start): Call gomp_new_team instead of
+	gomp_new_work_share.  Call gomp_loop_init on &team->work_shares[0].
+	Adjust gomp_team_start caller.
+	* single.c (GOMP_single_start, GOMP_single_copy_start): Call
+	gomp_work_share_init_done if gomp_work_share_start returned true.
+	Don't unlock ws->lock.
+	* parallel.c (gomp_resolve_num_threads): Fix non-HAVE_SYNC_BUILTINS
+	case.
+	(GOMP_parallel_start): Call gomp_new_team and pass that as last
+	argument to gomp_team_start.
+	* config/linux/ptrlock.c: New file.
+	* config/linux/ptrlock.h: New file.
+	* config/posix/ptrlock.c: New file.
+	* config/posix/ptrlock.h: New file.
+	* Makefile.am (libgomp_la_SOURCES): Add ptrlock.c.
+	* Makefile.in: Regenerated.
+	* testsuite/Makefile.in: Regenerated.
+
+	* testsuite/libgomp.c++/collapse-1.C (main): Make k private in second
+	omp for loop.
+
 2008-03-19  Jakub Jelinek  <jakub@redhat.com>
 
 	* libgomp.h (gomp_active_wait_policy): Remove decl.
diff --git a/libgomp/Makefile.am b/libgomp/Makefile.am
index 0101001fe89..7b77a8dde64 100644
--- a/libgomp/Makefile.am
+++ b/libgomp/Makefile.am
@@ -31,7 +31,7 @@ libgomp_la_LDFLAGS = $(libgomp_version_info) $(libgomp_version_script)
 
 libgomp_la_SOURCES = alloc.c barrier.c critical.c env.c error.c iter.c \
 	loop.c ordered.c parallel.c sections.c single.c task.c team.c work.c \
-	lock.c mutex.c proc.c sem.c bar.c time.c fortran.c affinity.c
+	lock.c mutex.c proc.c sem.c bar.c ptrlock.c time.c fortran.c affinity.c
 
 nodist_noinst_HEADERS = libgomp_f.h
 nodist_libsubinclude_HEADERS = omp.h
diff --git a/libgomp/Makefile.in b/libgomp/Makefile.in
index 3df46c55748..20221ca68a2 100644
--- a/libgomp/Makefile.in
+++ b/libgomp/Makefile.in
@@ -82,8 +82,8 @@ LTLIBRARIES = $(toolexeclib_LTLIBRARIES)
 libgomp_la_LIBADD =
 am_libgomp_la_OBJECTS = alloc.lo barrier.lo critical.lo env.lo \
 	error.lo iter.lo loop.lo ordered.lo parallel.lo sections.lo \
-	single.lo task.lo team.lo work.lo lock.lo mutex.lo proc.lo sem.lo \
-	bar.lo time.lo fortran.lo affinity.lo
+	single.lo task.lo team.lo work.lo lock.lo mutex.lo proc.lo \
+	sem.lo bar.lo ptrlock.lo time.lo fortran.lo affinity.lo
 libgomp_la_OBJECTS = $(am_libgomp_la_OBJECTS)
 DEFAULT_INCLUDES = -I. -I$(srcdir) -I.
 depcomp = $(SHELL) $(top_srcdir)/../depcomp
@@ -191,9 +191,15 @@ MAINTAINER_MODE_TRUE = @MAINTAINER_MODE_TRUE@
 MAKEINFO = @MAKEINFO@
 NM = @NM@
 OBJEXT = @OBJEXT@
+OMP_LOCK_25_ALIGN = @OMP_LOCK_25_ALIGN@
+OMP_LOCK_25_KIND = @OMP_LOCK_25_KIND@
+OMP_LOCK_25_SIZE = @OMP_LOCK_25_SIZE@
 OMP_LOCK_ALIGN = @OMP_LOCK_ALIGN@
 OMP_LOCK_KIND = @OMP_LOCK_KIND@
 OMP_LOCK_SIZE = @OMP_LOCK_SIZE@
+OMP_NEST_LOCK_25_ALIGN = @OMP_NEST_LOCK_25_ALIGN@
+OMP_NEST_LOCK_25_KIND = @OMP_NEST_LOCK_25_KIND@
+OMP_NEST_LOCK_25_SIZE = @OMP_NEST_LOCK_25_SIZE@
 OMP_NEST_LOCK_ALIGN = @OMP_NEST_LOCK_ALIGN@
 OMP_NEST_LOCK_KIND = @OMP_NEST_LOCK_KIND@
 OMP_NEST_LOCK_SIZE = @OMP_NEST_LOCK_SIZE@
@@ -288,7 +294,7 @@ libgomp_version_info = -version-info $(libtool_VERSION)
 libgomp_la_LDFLAGS = $(libgomp_version_info) $(libgomp_version_script)
 libgomp_la_SOURCES = alloc.c barrier.c critical.c env.c error.c iter.c \
 	loop.c ordered.c parallel.c sections.c single.c task.c team.c work.c \
-	lock.c mutex.c proc.c sem.c bar.c time.c fortran.c affinity.c
+	lock.c mutex.c proc.c sem.c bar.c ptrlock.c time.c fortran.c affinity.c
 
 nodist_noinst_HEADERS = libgomp_f.h
 nodist_libsubinclude_HEADERS = omp.h
@@ -430,6 +436,7 @@ distclean-compile:
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/ordered.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/parallel.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/proc.Plo@am__quote@
+@AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/ptrlock.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/sections.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/sem.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/single.Plo@am__quote@
diff --git a/libgomp/config/linux/ptrlock.c b/libgomp/config/linux/ptrlock.c
new file mode 100644
index 00000000000..8faa1b2287d
--- /dev/null
+++ b/libgomp/config/linux/ptrlock.c
@@ -0,0 +1,70 @@
+/* Copyright (C) 2008 Free Software Foundation, Inc.
+   Contributed by Jakub Jelinek <jakub@redhat.com>.
+
+   This file is part of the GNU OpenMP Library (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU Lesser General Public License as published by
+   the Free Software Foundation; either version 2.1 of the License, or
+   (at your option) any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public License for
+   more details.
+
+   You should have received a copy of the GNU Lesser General Public License 
+   along with libgomp; see the file COPYING.LIB.  If not, write to the
+   Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston,
+   MA 02110-1301, USA.  */
+
+/* As a special exception, if you link this library with other files, some
+   of which are compiled with GCC, to produce an executable, this library
+   does not by itself cause the resulting executable to be covered by the
+   GNU General Public License.  This exception does not however invalidate
+   any other reasons why the executable file might be covered by the GNU
+   General Public License.  */
+
+/* This is a Linux specific implementation of a mutex synchronization
+   mechanism for libgomp.  This type is private to the library.  This
+   implementation uses atomic instructions and the futex syscall.  */
+
+#include <endian.h>
+#include <limits.h>
+#include "wait.h"
+
+void *
+gomp_ptrlock_get_slow (gomp_ptrlock_t *ptrlock)
+{
+  int *intptr;
+  __sync_bool_compare_and_swap (ptrlock, 1, 2);
+
+  /* futex works on ints, not pointers.
+     But a valid work share pointer will be at least
+     8 byte aligned, so it is safe to assume the low
+     32-bits of the pointer won't contain values 1 or 2.  */
+  __asm volatile ("" : "=r" (intptr) : "0" (ptrlock));
+#if __BYTE_ORDER == __BIG_ENDIAN
+  if (sizeof (*ptrlock) > sizeof (int))
+    intptr += (sizeof (*ptrlock) / sizeof (int)) - 1;
+#endif
+  do
+    do_wait (intptr, 2);
+  while (*intptr == 2);
+  __asm volatile ("" : : : "memory");
+  return *ptrlock;
+}
+
+void
+gomp_ptrlock_set_slow (gomp_ptrlock_t *ptrlock, void *ptr)
+{
+  int *intptr;
+
+  *ptrlock = ptr;
+  __asm volatile ("" : "=r" (intptr) : "0" (ptrlock));
+#if __BYTE_ORDER == __BIG_ENDIAN
+  if (sizeof (*ptrlock) > sizeof (int))
+    intptr += (sizeof (*ptrlock) / sizeof (int)) - 1;
+#endif
+  futex_wake (intptr, INT_MAX);
+}
diff --git a/libgomp/config/linux/ptrlock.h b/libgomp/config/linux/ptrlock.h
new file mode 100644
index 00000000000..bb5441676a4
--- /dev/null
+++ b/libgomp/config/linux/ptrlock.h
@@ -0,0 +1,65 @@
+/* Copyright (C) 2008 Free Software Foundation, Inc.
+   Contributed by Jakub Jelinek <jakub@redhat.com>.
+
+   This file is part of the GNU OpenMP Library (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU Lesser General Public License as published by
+   the Free Software Foundation; either version 2.1 of the License, or
+   (at your option) any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public License for
+   more details.
+
+   You should have received a copy of the GNU Lesser General Public License 
+   along with libgomp; see the file COPYING.LIB.  If not, write to the
+   Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston,
+   MA 02110-1301, USA.  */
+
+/* As a special exception, if you link this library with other files, some
+   of which are compiled with GCC, to produce an executable, this library
+   does not by itself cause the resulting executable to be covered by the
+   GNU General Public License.  This exception does not however invalidate
+   any other reasons why the executable file might be covered by the GNU
+   General Public License.  */
+
+/* This is a Linux specific implementation of a mutex synchronization
+   mechanism for libgomp.  This type is private to the library.  This
+   implementation uses atomic instructions and the futex syscall.  */
+
+#ifndef GOMP_PTRLOCK_H
+#define GOMP_PTRLOCK_H 1
+
+typedef void *gomp_ptrlock_t;
+
+static inline void gomp_ptrlock_init (gomp_ptrlock_t *ptrlock, void *ptr)
+{
+  *ptrlock = ptr;
+}
+
+extern void *gomp_ptrlock_get_slow (gomp_ptrlock_t *ptrlock);
+static inline void *gomp_ptrlock_get (gomp_ptrlock_t *ptrlock)
+{
+  if ((uintptr_t) *ptrlock > 2)
+    return *ptrlock;
+
+  if (__sync_bool_compare_and_swap (ptrlock, NULL, (uintptr_t) 1))
+    return NULL;
+
+  return gomp_ptrlock_get_slow (ptrlock);
+}
+
+extern void gomp_ptrlock_set_slow (gomp_ptrlock_t *ptrlock, void *ptr);
+static inline void gomp_ptrlock_set (gomp_ptrlock_t *ptrlock, void *ptr)
+{
+  if (!__sync_bool_compare_and_swap (ptrlock, (uintptr_t) 1, ptr))
+    gomp_ptrlock_set_slow (ptrlock, ptr);
+}
+
+static inline void gomp_ptrlock_destroy (gomp_ptrlock_t *ptrlock)
+{
+}
+
+#endif /* GOMP_PTRLOCK_H */
diff --git a/libgomp/config/posix/ptrlock.c b/libgomp/config/posix/ptrlock.c
new file mode 100644
index 00000000000..39bb64da0f9
--- /dev/null
+++ b/libgomp/config/posix/ptrlock.c
@@ -0,0 +1 @@
+/* Everything is in the header.  */
diff --git a/libgomp/config/posix/ptrlock.h b/libgomp/config/posix/ptrlock.h
new file mode 100644
index 00000000000..1271ebb227b
--- /dev/null
+++ b/libgomp/config/posix/ptrlock.h
@@ -0,0 +1,69 @@
+/* Copyright (C) 2008 Free Software Foundation, Inc.
+   Contributed by Jakub Jelinek <jakub@redhat.com>.
+
+   This file is part of the GNU OpenMP Library (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU Lesser General Public License as published by
+   the Free Software Foundation; either version 2.1 of the License, or
+   (at your option) any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public License for
+   more details.
+
+   You should have received a copy of the GNU Lesser General Public License 
+   along with libgomp; see the file COPYING.LIB.  If not, write to the
+   Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston,
+   MA 02110-1301, USA.  */
+
+/* As a special exception, if you link this library with other files, some
+   of which are compiled with GCC, to produce an executable, this library
+   does not by itself cause the resulting executable to be covered by the
+   GNU General Public License.  This exception does not however invalidate
+   any other reasons why the executable file might be covered by the GNU
+   General Public License.  */
+
+/* This is a Linux specific implementation of a mutex synchronization
+   mechanism for libgomp.  This type is private to the library.  This
+   implementation uses atomic instructions and the futex syscall.  */
+
+#ifndef GOMP_PTRLOCK_H
+#define GOMP_PTRLOCK_H 1
+
+typedef struct { void *ptr; gomp_mutex_t lock; } gomp_ptrlock_t;
+
+static inline void gomp_ptrlock_init (gomp_ptrlock_t *ptrlock, void *ptr)
+{
+  ptrlock->ptr = ptr;
+  gomp_mutex_init (&ptrlock->lock);
+}
+
+static inline void *gomp_ptrlock_get (gomp_ptrlock_t *ptrlock)
+{
+  if (ptrlock->ptr != NULL)
+    return ptrlock->ptr;
+
+  gomp_mutex_lock (&ptrlock->lock);
+  if (ptrlock->ptr != NULL)
+    {
+      gomp_mutex_unlock (&ptrlock->lock);
+      return ptrlock->ptr;
+    }
+
+  return NULL;
+}
+
+static inline void gomp_ptrlock_set (gomp_ptrlock_t *ptrlock, void *ptr)
+{
+  ptrlock->ptr = ptr;
+  gomp_mutex_unlock (&ptrlock->lock);
+}
+
+static inline void gomp_ptrlock_destroy (gomp_ptrlock_t *ptrlock)
+{
+  gomp_mutex_destroy (&ptrlock->lock);
+}
+
+#endif /* GOMP_PTRLOCK_H */
diff --git a/libgomp/libgomp.h b/libgomp/libgomp.h
index e57077c3209..95cab09a1bd 100644
--- a/libgomp/libgomp.h
+++ b/libgomp/libgomp.h
@@ -50,6 +50,7 @@
 #include "sem.h"
 #include "mutex.h"
 #include "bar.h"
+#include "ptrlock.h"
 
 
 /* This structure contains the data to control one work-sharing construct,
@@ -84,12 +85,45 @@ struct gomp_work_share
      is always 1.  */
   long incr;
 
-  /* The above fields are written once during gomp_loop_init.  Make sure
-     the following fields are in a different cache line.  */
+  /* This is a circular queue that details which threads will be allowed
+     into the ordered region and in which order.  When a thread allocates
+     iterations on which it is going to work, it also registers itself at
+     the end of the array.  When a thread reaches the ordered region, it
+     checks to see if it is the one at the head of the queue.  If not, it
+     blocks on its RELEASE semaphore.  */
+  unsigned *ordered_team_ids;
+
+  /* This is the number of threads that have registered themselves in
+     the circular queue ordered_team_ids.  */
+  unsigned ordered_num_used;
+
+  /* This is the team_id of the currently acknowledged owner of the ordered
+     section, or -1u if the ordered section has not been acknowledged by
+     any thread.  This is distinguished from the thread that is *allowed*
+     to take the section next.  */
+  unsigned ordered_owner;
+
+  /* This is the index into the circular queue ordered_team_ids of the
+     current thread that's allowed into the ordered reason.  */
+  unsigned ordered_cur;
+
+  /* This is a chain of allocated gomp_work_share blocks, valid only
+     in the first gomp_work_share struct in the block.  */
+  struct gomp_work_share *next_alloc;
+
+  /* The above fields are written once during workshare initialization,
+     or related to ordered worksharing.  Make sure the following fields
+     are in a different cache line.  */
 
   /* This lock protects the update of the following members.  */
   gomp_mutex_t lock __attribute__((aligned (64)));
 
+  /* This is the count of the number of threads that have exited the work
+     share construct.  If the construct was marked nowait, they have moved on
+     to other work; otherwise they're blocked on a barrier.  The last member
+     of the team to exit the work share construct must deallocate it.  */
+  unsigned threads_completed;
+
   union {
     /* This is the next iteration value to be allocated.  In the case of
        GFS_STATIC loops, this the iteration start point and never changes.  */
@@ -99,33 +133,19 @@ struct gomp_work_share
     void *copyprivate;
   };
 
-  /* This is the count of the number of threads that have exited the work
-     share construct.  If the construct was marked nowait, they have moved on
-     to other work; otherwise they're blocked on a barrier.  The last member
-     of the team to exit the work share construct must deallocate it.  */
-  unsigned threads_completed;
-
-  /* This is the index into the circular queue ordered_team_ids of the 
-     current thread that's allowed into the ordered reason.  */
-  unsigned ordered_cur;
-
-  /* This is the number of threads that have registered themselves in
-     the circular queue ordered_team_ids.  */
-  unsigned ordered_num_used;
+  union {
+    /* Link to gomp_work_share struct for next work sharing construct
+       encountered after this one.  */
+    gomp_ptrlock_t next_ws;
 
-  /* This is the team_id of the currently acknoledged owner of the ordered
-     section, or -1u if the ordered section has not been acknowledged by
-     any thread.  This is distinguished from the thread that is *allowed*
-     to take the section next.  */
-  unsigned ordered_owner;
+    /* gomp_work_share structs are chained in the free work share cache
+       through this.  */
+    struct gomp_work_share *next_free;
+  };
 
-  /* This is a circular queue that details which threads will be allowed
-     into the ordered region and in which order.  When a thread allocates
-     iterations on which it is going to work, it also registers itself at
-     the end of the array.  When a thread reaches the ordered region, it
-     checks to see if it is the one at the head of the queue.  If not, it
-     blocks on its RELEASE semaphore.  */
-  unsigned ordered_team_ids[];
+  /* If only few threads are in the team, ordered_team_ids can point
+     to this array which fills the padding at the end of this struct.  */
+  unsigned inline_ordered_team_ids[0];
 };
 
 /* This structure contains all of the thread-local data associated with 
@@ -139,10 +159,15 @@ struct gomp_team_state
 
   /* This is the work share construct which this thread is currently
      processing.  Recall that with NOWAIT, not all threads may be 
-     processing the same construct.  This value is NULL when there
-     is no construct being processed.  */
+     processing the same construct.  */
   struct gomp_work_share *work_share;
 
+  /* This is the previous work share construct or NULL if there wasn't any.
+     When all threads are done with the current work sharing construct,
+     the previous one can be freed.  The current one can't, as its
+     next_ws field is used.  */
+  struct gomp_work_share *last_work_share;
+
   /* This is the ID of this thread within the team.  This value is
      guaranteed to be between 0 and N-1, where N is the number of
      threads in the team.  */
@@ -154,13 +179,6 @@ struct gomp_team_state
   /* Active nesting level.  Only active parallel regions are counted.  */
   unsigned active_level;
 
-  /* The work share "generation" is a number that increases by one for
-     each work share construct encountered in the dynamic flow of the
-     program.  It is used to find the control data for the work share
-     when encountering it for the first time.  This particular number
-     reflects the generation of the work_share member of this struct.  */
-  unsigned work_share_generation;
-
   /* For GFS_RUNTIME loops that resolved to GFS_STATIC, this is the
      trip number through the loop.  So first time a particular loop
      is encountered this number is 0, the second time through the loop
@@ -209,27 +227,13 @@ struct gomp_task
 
 struct gomp_team
 {
-  /* This lock protects access to the following work shares data structures.  */
-  gomp_mutex_t work_share_lock;
-
-  /* This is a dynamically sized array containing pointers to the control
-     structs for all "live" work share constructs.  Here "live" means that
-     the construct has been encountered by at least one thread, and not
-     completed by all threads.  */
-  struct gomp_work_share **work_shares;
-
-  /* The work_shares array is indexed by "generation & generation_mask".
-     The mask will be 2**N - 1, where 2**N is the size of the array.  */
-  unsigned generation_mask;
-
-  /* These two values define the bounds of the elements of the work_shares
-     array that are currently in use.  */
-  unsigned oldest_live_gen;
-  unsigned num_live_gen;
-
   /* This is the number of threads in the current team.  */
   unsigned nthreads;
 
+  /* This is number of gomp_work_share structs that have been allocated
+     as a block last time.  */
+  unsigned work_share_chunk;
+
   /* This is the saved team state that applied to a master thread before
      the current thread was created.  */
   struct gomp_team_state prev_ts;
@@ -243,11 +247,26 @@ struct gomp_team
      of the threads in the team.  */
   gomp_sem_t **ordered_release;
 
-  struct gomp_work_share *init_work_shares[4];
+  /* List of gomp_work_share structs chained through next_free fields.
+     This is populated and taken off only by the first thread in the
+     team encountering a new work sharing construct, in a critical
+     section.  */
+  struct gomp_work_share *work_share_list_alloc;
+
+  /* List of gomp_work_share structs freed by free_work_share.  New
+     entries are atomically added to the start of the list, and
+     alloc_work_share can safely only move all but the first entry
+     to work_share_list alloc, as free_work_share can happen concurrently
+     with alloc_work_share.  */
+  struct gomp_work_share *work_share_list_free;
 
   /* This barrier is used for most synchronization of the team.  */
   gomp_barrier_t barrier;
 
+  /* Initial work shares, to avoid allocating any gomp_work_share
+     structs in the common case.  */
+  struct gomp_work_share work_shares[8];
+
   /* This array contains structures for implicit tasks.  */
   struct gomp_task implicit_task[];
 };
@@ -368,17 +387,27 @@ extern void gomp_end_task (void);
 
 /* team.c */
 
+extern struct gomp_team *gomp_new_team (unsigned);
 extern void gomp_team_start (void (*) (void *), void *, unsigned,
-			     struct gomp_work_share *);
+			     struct gomp_team *);
 extern void gomp_team_end (void);
 
 /* work.c */
 
-extern struct gomp_work_share * gomp_new_work_share (bool, unsigned);
+extern void gomp_init_work_share (struct gomp_work_share *, bool, unsigned);
+extern void gomp_fini_work_share (struct gomp_work_share *);
 extern bool gomp_work_share_start (bool);
 extern void gomp_work_share_end (void);
 extern void gomp_work_share_end_nowait (void);
 
+static inline void
+gomp_work_share_init_done (void)
+{
+  struct gomp_thread *thr = gomp_thread ();
+  if (__builtin_expect (thr->ts.last_work_share != NULL, 1))
+    gomp_ptrlock_set (&thr->ts.last_work_share->next_ws, thr->ts.work_share);
+}
+
 #ifdef HAVE_ATTRIBUTE_VISIBILITY
 # pragma GCC visibility pop
 #endif
diff --git a/libgomp/loop.c b/libgomp/loop.c
index 725e6bdf150..13ed6d44833 100644
--- a/libgomp/loop.c
+++ b/libgomp/loop.c
@@ -86,10 +86,13 @@ gomp_loop_static_start (long start, long end, long incr, long chunk_size,
 {
   struct gomp_thread *thr = gomp_thread ();
 
+  thr->ts.static_trip = 0;
   if (gomp_work_share_start (false))
-    gomp_loop_init (thr->ts.work_share, start, end, incr,
-		    GFS_STATIC, chunk_size);
-  gomp_mutex_unlock (&thr->ts.work_share->lock);
+    {
+      gomp_loop_init (thr->ts.work_share, start, end, incr,
+		      GFS_STATIC, chunk_size);
+      gomp_work_share_init_done ();
+    }
 
   return !gomp_iter_static_next (istart, iend);
 }
@@ -102,13 +105,16 @@ gomp_loop_dynamic_start (long start, long end, long incr, long chunk_size,
   bool ret;
 
   if (gomp_work_share_start (false))
-    gomp_loop_init (thr->ts.work_share, start, end, incr,
-		    GFS_DYNAMIC, chunk_size);
+    {
+      gomp_loop_init (thr->ts.work_share, start, end, incr,
+		      GFS_DYNAMIC, chunk_size);
+      gomp_work_share_init_done ();
+    }
 
 #ifdef HAVE_SYNC_BUILTINS
-  gomp_mutex_unlock (&thr->ts.work_share->lock);
   ret = gomp_iter_dynamic_next (istart, iend);
 #else
+  gomp_mutex_lock (&thr->ts.work_share->lock);
   ret = gomp_iter_dynamic_next_locked (istart, iend);
   gomp_mutex_unlock (&thr->ts.work_share->lock);
 #endif
@@ -124,13 +130,16 @@ gomp_loop_guided_start (long start, long end, long incr, long chunk_size,
   bool ret;
 
   if (gomp_work_share_start (false))
-    gomp_loop_init (thr->ts.work_share, start, end, incr,
-		    GFS_GUIDED, chunk_size);
+    {
+      gomp_loop_init (thr->ts.work_share, start, end, incr,
+		      GFS_GUIDED, chunk_size);
+      gomp_work_share_init_done ();
+    }
 
 #ifdef HAVE_SYNC_BUILTINS
-  gomp_mutex_unlock (&thr->ts.work_share->lock);
   ret = gomp_iter_guided_next (istart, iend);
 #else
+  gomp_mutex_lock (&thr->ts.work_share->lock);
   ret = gomp_iter_guided_next_locked (istart, iend);
   gomp_mutex_unlock (&thr->ts.work_share->lock);
 #endif
@@ -168,13 +177,14 @@ gomp_loop_ordered_static_start (long start, long end, long incr,
 {
   struct gomp_thread *thr = gomp_thread ();
 
+  thr->ts.static_trip = 0;
   if (gomp_work_share_start (true))
     {
       gomp_loop_init (thr->ts.work_share, start, end, incr,
 		      GFS_STATIC, chunk_size);
       gomp_ordered_static_init ();
+      gomp_work_share_init_done ();
     }
-  gomp_mutex_unlock (&thr->ts.work_share->lock);
 
   return !gomp_iter_static_next (istart, iend);
 }
@@ -187,8 +197,14 @@ gomp_loop_ordered_dynamic_start (long start, long end, long incr,
   bool ret;
 
   if (gomp_work_share_start (true))
-    gomp_loop_init (thr->ts.work_share, start, end, incr,
-		    GFS_DYNAMIC, chunk_size);
+    {
+      gomp_loop_init (thr->ts.work_share, start, end, incr,
+		      GFS_DYNAMIC, chunk_size);
+      gomp_mutex_lock (&thr->ts.work_share->lock);
+      gomp_work_share_init_done ();
+    }
+  else
+    gomp_mutex_lock (&thr->ts.work_share->lock);
 
   ret = gomp_iter_dynamic_next_locked (istart, iend);
   if (ret)
@@ -206,8 +222,14 @@ gomp_loop_ordered_guided_start (long start, long end, long incr,
   bool ret;
 
   if (gomp_work_share_start (true))
-    gomp_loop_init (thr->ts.work_share, start, end, incr,
-		    GFS_GUIDED, chunk_size);
+    {
+      gomp_loop_init (thr->ts.work_share, start, end, incr,
+		      GFS_GUIDED, chunk_size);
+      gomp_mutex_lock (&thr->ts.work_share->lock);
+      gomp_work_share_init_done ();
+    }
+  else
+    gomp_mutex_lock (&thr->ts.work_share->lock);
 
   ret = gomp_iter_guided_next_locked (istart, iend);
   if (ret)
@@ -395,12 +417,12 @@ gomp_parallel_loop_start (void (*fn) (void *), void *data,
 			  long incr, enum gomp_schedule_type sched,
 			  long chunk_size)
 {
-  struct gomp_work_share *ws;
+  struct gomp_team *team;
 
   num_threads = gomp_resolve_num_threads (num_threads, 0);
-  ws = gomp_new_work_share (false, num_threads);
-  gomp_loop_init (ws, start, end, incr, sched, chunk_size);
-  gomp_team_start (fn, data, num_threads, ws);
+  team = gomp_new_team (num_threads);
+  gomp_loop_init (&team->work_shares[0], start, end, incr, sched, chunk_size);
+  gomp_team_start (fn, data, num_threads, team);
 }
 
 void
diff --git a/libgomp/parallel.c b/libgomp/parallel.c
index 5537a49a463..0882179d165 100644
--- a/libgomp/parallel.c
+++ b/libgomp/parallel.c
@@ -94,10 +94,11 @@ gomp_resolve_num_threads (unsigned specified, unsigned count)
 #else
   gomp_mutex_lock (&gomp_remaining_threads_lock);
   num_threads = max_num_threads;
-  if (num_threads > gomp_remaining_threads_count)
-    num_threads = gomp_remaining_threads_count + 1;
+  remaining = gomp_remaining_threads_count;
+  if (num_threads > remaining)
+    num_threads = remaining + 1;
   gomp_remaining_threads_count -= num_threads - 1;
-  gomp_mutex_unlock (&gomp_remaining_threads_unlock);
+  gomp_mutex_unlock (&gomp_remaining_threads_lock);
 #endif
 
   return num_threads;
@@ -107,7 +108,7 @@ void
 GOMP_parallel_start (void (*fn) (void *), void *data, unsigned num_threads)
 {
   num_threads = gomp_resolve_num_threads (num_threads, 0);
-  gomp_team_start (fn, data, num_threads, NULL);
+  gomp_team_start (fn, data, num_threads, gomp_new_team (num_threads));
 }
 
 void
diff --git a/libgomp/sections.c b/libgomp/sections.c
index eeccbfd47f7..27625efec3e 100644
--- a/libgomp/sections.c
+++ b/libgomp/sections.c
@@ -59,14 +59,24 @@ GOMP_sections_start (unsigned count)
   long s, e, ret;
 
   if (gomp_work_share_start (false))
-    gomp_sections_init (thr->ts.work_share, count);
+    {
+      gomp_sections_init (thr->ts.work_share, count);
+      gomp_work_share_init_done ();
+    }
 
+#ifdef HAVE_SYNC_BUILTINS
+  if (gomp_iter_dynamic_next (&s, &e))
+    ret = s;
+  else
+    ret = 0;
+#else
+  gomp_mutex_lock (&thr->ts.work_share->lock);
   if (gomp_iter_dynamic_next_locked (&s, &e))
     ret = s;
   else
     ret = 0;
-
   gomp_mutex_unlock (&thr->ts.work_share->lock);
+#endif
 
   return ret;
 }
@@ -83,15 +93,23 @@ GOMP_sections_start (unsigned count)
 unsigned
 GOMP_sections_next (void)
 {
-  struct gomp_thread *thr = gomp_thread ();
   long s, e, ret;
 
+#ifdef HAVE_SYNC_BUILTINS
+  if (gomp_iter_dynamic_next (&s, &e))
+    ret = s;
+  else
+    ret = 0;
+#else
+  struct gomp_thread *thr = gomp_thread ();
+
   gomp_mutex_lock (&thr->ts.work_share->lock);
   if (gomp_iter_dynamic_next_locked (&s, &e))
     ret = s;
   else
     ret = 0;
   gomp_mutex_unlock (&thr->ts.work_share->lock);
+#endif
 
   return ret;
 }
@@ -103,12 +121,12 @@ void
 GOMP_parallel_sections_start (void (*fn) (void *), void *data,
 			      unsigned num_threads, unsigned count)
 {
-  struct gomp_work_share *ws;
+  struct gomp_team *team;
 
   num_threads = gomp_resolve_num_threads (num_threads, count);
-  ws = gomp_new_work_share (false, num_threads);
-  gomp_sections_init (ws, count);
-  gomp_team_start (fn, data, num_threads, ws);
+  team = gomp_new_team (num_threads);
+  gomp_sections_init (&team->work_shares[0], count);
+  gomp_team_start (fn, data, num_threads, team);
 }
 
 /* The GOMP_section_end* routines are called after the thread is told
diff --git a/libgomp/single.c b/libgomp/single.c
index dde05d9ceb8..439d8e70f9d 100644
--- a/libgomp/single.c
+++ b/libgomp/single.c
@@ -1,4 +1,4 @@
-/* Copyright (C) 2005 Free Software Foundation, Inc.
+/* Copyright (C) 2005, 2008 Free Software Foundation, Inc.
    Contributed by Richard Henderson <rth@redhat.com>.
 
    This file is part of the GNU OpenMP Library (libgomp).
@@ -38,7 +38,8 @@ bool
 GOMP_single_start (void)
 {
   bool ret = gomp_work_share_start (false);
-  gomp_mutex_unlock (&gomp_thread ()->ts.work_share->lock);
+  if (ret)
+    gomp_work_share_init_done ();
   gomp_work_share_end_nowait ();
   return ret;
 }
@@ -57,10 +58,12 @@ GOMP_single_copy_start (void)
   void *ret;
 
   first = gomp_work_share_start (false);
-  gomp_mutex_unlock (&thr->ts.work_share->lock);
   
   if (first)
-    ret = NULL;
+    {
+      gomp_work_share_init_done ();
+      ret = NULL;
+    }
   else
     {
       gomp_barrier_wait (&thr->ts.team->barrier);
diff --git a/libgomp/team.c b/libgomp/team.c
index f54f08e59a0..743cd7fbef1 100644
--- a/libgomp/team.c
+++ b/libgomp/team.c
@@ -117,11 +117,10 @@ gomp_thread_start (void *xdata)
 	  thr->data = NULL;
 	  thr->ts.team = NULL;
 	  thr->ts.work_share = NULL;
+	  thr->ts.last_work_share = NULL;
 	  thr->ts.team_id = 0;
 	  thr->ts.level = 0;
 	  thr->ts.active_level = 0;
-	  thr->ts.work_share_generation = 0;
-	  thr->ts.static_trip = 0;
 
 	  gomp_barrier_wait_last (&team->barrier);
 	  gomp_barrier_wait (&gomp_threads_dock);
@@ -138,22 +137,25 @@ gomp_thread_start (void *xdata)
 
 /* Create a new team data structure.  */
 
-static struct gomp_team *
-new_team (unsigned nthreads, struct gomp_work_share *work_share)
+struct gomp_team *
+gomp_new_team (unsigned nthreads)
 {
   struct gomp_team *team;
   size_t size;
+  int i;
 
   size = sizeof (*team) + nthreads * (sizeof (team->ordered_release[0])
 				      + sizeof (team->implicit_task[0]));
   team = gomp_malloc (size);
-  gomp_mutex_init (&team->work_share_lock);
 
-  team->work_shares = team->init_work_shares;
-  team->generation_mask = 3;
-  team->oldest_live_gen = work_share == NULL;
-  team->num_live_gen = work_share != NULL;
-  team->init_work_shares[0] = work_share;
+  team->work_share_chunk = 8;
+  gomp_init_work_share (&team->work_shares[0], false, nthreads);
+  team->work_shares[0].next_alloc = NULL;
+  team->work_share_list_free = NULL;
+  team->work_share_list_alloc = &team->work_shares[1];
+  for (i = 1; i < 7; i++)
+    team->work_shares[i].next_free = &team->work_shares[i + 1];
+  team->work_shares[i].next_free = NULL;
 
   team->nthreads = nthreads;
   gomp_barrier_init (&team->barrier, nthreads);
@@ -171,9 +173,17 @@ new_team (unsigned nthreads, struct gomp_work_share *work_share)
 static void
 free_team (struct gomp_team *team)
 {
-  if (__builtin_expect (team->work_shares != team->init_work_shares, 0))
-    free (team->work_shares);
-  gomp_mutex_destroy (&team->work_share_lock);
+  if (__builtin_expect (team->work_shares[0].next_alloc != NULL, 0))
+    {
+      struct gomp_work_share *ws = team->work_shares[0].next_alloc;
+      do
+	{
+	  struct gomp_work_share *next_ws = ws->next_alloc;
+	  free (ws);
+	  ws = next_ws;
+	}
+      while (ws != NULL);
+    }
   gomp_barrier_destroy (&team->barrier);
   gomp_sem_destroy (&team->master_release);
   free (team);
@@ -184,11 +194,10 @@ free_team (struct gomp_team *team)
 
 void
 gomp_team_start (void (*fn) (void *), void *data, unsigned nthreads,
-		 struct gomp_work_share *work_share)
+		 struct gomp_team *team)
 {
   struct gomp_thread_start_data *start_data;
   struct gomp_thread *thr, *nthr;
-  struct gomp_team *team;
   struct gomp_task *task;
   struct gomp_task_icv *icv;
   bool nested;
@@ -200,20 +209,18 @@ gomp_team_start (void (*fn) (void *), void *data, unsigned nthreads,
   task = thr->task;
   icv = task ? &task->icv : &gomp_global_icv;
 
-  team = new_team (nthreads, work_share);
-
   /* Always save the previous state, even if this isn't a nested team.
      In particular, we should save any work share state from an outer
      orphaned work share construct.  */
   team->prev_ts = thr->ts;
 
   thr->ts.team = team;
-  thr->ts.work_share = work_share;
   thr->ts.team_id = 0;
   ++thr->ts.level;
   if (nthreads > 1)
     ++thr->ts.active_level;
-  thr->ts.work_share_generation = 0;
+  thr->ts.work_share = &team->work_shares[0];
+  thr->ts.last_work_share = NULL;
   thr->ts.static_trip = 0;
   thr->task = &team->implicit_task[0];
   gomp_init_task (thr->task, task, icv);
@@ -258,11 +265,11 @@ gomp_team_start (void (*fn) (void *), void *data, unsigned nthreads,
 	{
 	  nthr = gomp_threads[i];
 	  nthr->ts.team = team;
-	  nthr->ts.work_share = work_share;
+	  nthr->ts.work_share = &team->work_shares[0];
+	  nthr->ts.last_work_share = NULL;
 	  nthr->ts.team_id = i;
 	  nthr->ts.level = team->prev_ts.level + 1;
 	  nthr->ts.active_level = thr->ts.active_level;
-	  nthr->ts.work_share_generation = 0;
 	  nthr->ts.static_trip = 0;
 	  nthr->task = &team->implicit_task[i];
 	  gomp_init_task (nthr->task, task, icv);
@@ -326,11 +333,11 @@ gomp_team_start (void (*fn) (void *), void *data, unsigned nthreads,
       start_data->fn = fn;
       start_data->fn_data = data;
       start_data->ts.team = team;
-      start_data->ts.work_share = work_share;
+      start_data->ts.work_share = &team->work_shares[0];
+      start_data->ts.last_work_share = NULL;
       start_data->ts.team_id = i;
       start_data->ts.level = team->prev_ts.level + 1;
       start_data->ts.active_level = thr->ts.active_level;
-      start_data->ts.work_share_generation = 0;
       start_data->ts.static_trip = 0;
       start_data->task = &team->implicit_task[i];
       gomp_init_task (start_data->task, task, icv);
@@ -382,6 +389,8 @@ gomp_team_end (void)
 
   gomp_barrier_wait (&team->barrier);
 
+  gomp_fini_work_share (thr->ts.work_share);
+
   gomp_end_task ();
   thr->ts = team->prev_ts;
 
diff --git a/libgomp/testsuite/Makefile.in b/libgomp/testsuite/Makefile.in
index aaf23dcbc6d..b869900954a 100644
--- a/libgomp/testsuite/Makefile.in
+++ b/libgomp/testsuite/Makefile.in
@@ -110,9 +110,15 @@ MAINTAINER_MODE_TRUE = @MAINTAINER_MODE_TRUE@
 MAKEINFO = @MAKEINFO@
 NM = @NM@
 OBJEXT = @OBJEXT@
+OMP_LOCK_25_ALIGN = @OMP_LOCK_25_ALIGN@
+OMP_LOCK_25_KIND = @OMP_LOCK_25_KIND@
+OMP_LOCK_25_SIZE = @OMP_LOCK_25_SIZE@
 OMP_LOCK_ALIGN = @OMP_LOCK_ALIGN@
 OMP_LOCK_KIND = @OMP_LOCK_KIND@
 OMP_LOCK_SIZE = @OMP_LOCK_SIZE@
+OMP_NEST_LOCK_25_ALIGN = @OMP_NEST_LOCK_25_ALIGN@
+OMP_NEST_LOCK_25_KIND = @OMP_NEST_LOCK_25_KIND@
+OMP_NEST_LOCK_25_SIZE = @OMP_NEST_LOCK_25_SIZE@
 OMP_NEST_LOCK_ALIGN = @OMP_NEST_LOCK_ALIGN@
 OMP_NEST_LOCK_KIND = @OMP_NEST_LOCK_KIND@
 OMP_NEST_LOCK_SIZE = @OMP_NEST_LOCK_SIZE@
diff --git a/libgomp/testsuite/libgomp.c++/collapse-1.C b/libgomp/testsuite/libgomp.c++/collapse-1.C
index 6d30a65081b..132d35cf41d 100644
--- a/libgomp/testsuite/libgomp.c++/collapse-1.C
+++ b/libgomp/testsuite/libgomp.c++/collapse-1.C
@@ -17,7 +17,7 @@ main ()
 	  a[i][j][k] = i + j * 4 + k * 16;
   #pragma omp parallel
     {
-      #pragma omp for collapse(2) reduction(|:l)
+      #pragma omp for collapse(2) reduction(|:l) private (k)
 	for (i = 0; i < 2; i++)
 	  for (j = 0; j < 2; j++)
 	    for (k = 0; k < 2; k++)
diff --git a/libgomp/work.c b/libgomp/work.c
index 1829aeaff21..d9a515e511f 100644
--- a/libgomp/work.c
+++ b/libgomp/work.c
@@ -34,35 +34,112 @@
 #include <string.h>
 
 
-/* Create a new work share structure.  */
+/* Allocate a new work share structure, preferably from current team's
+   free gomp_work_share cache.  */
 
-struct gomp_work_share *
-gomp_new_work_share (bool ordered, unsigned nthreads)
+static struct gomp_work_share *
+alloc_work_share (struct gomp_team *team)
 {
   struct gomp_work_share *ws;
-  size_t size;
+  unsigned int i;
 
-  size = offsetof (struct gomp_work_share, ordered_team_ids);
-  if (ordered)
-    size += nthreads * sizeof (ws->ordered_team_ids[0]);
+  /* This is called in a critical section.  */
+  if (team->work_share_list_alloc != NULL)
+    {
+      ws = team->work_share_list_alloc;
+      team->work_share_list_alloc = ws->next_free;
+      return ws;
+    }
 
-  ws = gomp_malloc_cleared (size);
-  gomp_mutex_init (&ws->lock);
-  ws->ordered_owner = -1;
+  ws = team->work_share_list_free;
+  /* We need atomic read from work_share_list_free,
+     as free_work_share can be called concurrently.  */
+  __asm ("" : "+r" (ws));
 
+  if (ws && ws->next_free)
+    {
+      struct gomp_work_share *next = ws->next_free;
+      ws->next_free = NULL;
+      team->work_share_list_alloc = next->next_free;
+      return next;
+    }
+
+  team->work_share_chunk *= 2;
+  ws = gomp_malloc (team->work_share_chunk * sizeof (struct gomp_work_share));
+  ws->next_alloc = team->work_shares[0].next_alloc;
+  team->work_shares[0].next_alloc = ws;
+  team->work_share_list_alloc = &ws[1];
+  for (i = 1; i < team->work_share_chunk - 1; i++)
+    ws[i].next_free = &ws[i + 1];
+  ws[i].next_free = NULL;
   return ws;
 }
 
+/* Initialize an already allocated struct gomp_work_share.
+   This shouldn't touch the next_alloc field.  */
+
+void
+gomp_init_work_share (struct gomp_work_share *ws, bool ordered,
+		      unsigned nthreads)
+{
+  gomp_mutex_init (&ws->lock);
+  if (__builtin_expect (ordered, 0))
+    {
+#define INLINE_ORDERED_TEAM_IDS_CNT \
+  ((sizeof (struct gomp_work_share) \
+    - offsetof (struct gomp_work_share, inline_ordered_team_ids)) \
+   / sizeof (((struct gomp_work_share *) 0)->inline_ordered_team_ids[0]))
+
+      if (nthreads > INLINE_ORDERED_TEAM_IDS_CNT)
+	ws->ordered_team_ids
+	  = gomp_malloc (nthreads * sizeof (*ws->ordered_team_ids));
+      else
+	ws->ordered_team_ids = ws->inline_ordered_team_ids;
+      memset (ws->ordered_team_ids, '\0',
+	      nthreads * sizeof (*ws->ordered_team_ids));
+      ws->ordered_num_used = 0;
+      ws->ordered_owner = -1;
+      ws->ordered_cur = 0;
+    }
+  else
+    ws->ordered_team_ids = NULL;
+  gomp_ptrlock_init (&ws->next_ws, NULL);
+  ws->threads_completed = 0;
+}
 
-/* Free a work share structure.  */
+/* Do any needed destruction of gomp_work_share fields before it
+   is put back into free gomp_work_share cache or freed.  */
 
-static void
-free_work_share (struct gomp_work_share *ws)
+void
+gomp_fini_work_share (struct gomp_work_share *ws)
 {
   gomp_mutex_destroy (&ws->lock);
-  free (ws);
+  if (ws->ordered_team_ids != ws->inline_ordered_team_ids)
+    free (ws->ordered_team_ids);
+  gomp_ptrlock_destroy (&ws->next_ws);
 }
 
+/* Free a work share struct, if not orphaned, put it into current
+   team's free gomp_work_share cache.  */
+
+static inline void
+free_work_share (struct gomp_team *team, struct gomp_work_share *ws)
+{
+  gomp_fini_work_share (ws);
+  if (__builtin_expect (team == NULL, 0))
+    free (ws);
+  else
+    {
+      struct gomp_work_share *next_ws;
+      do
+	{
+	  next_ws = team->work_share_list_free;
+	  ws->next_free = next_ws;
+	}
+      while (!__sync_bool_compare_and_swap (&team->work_share_list_free,
+					    next_ws, ws));
+    }
+}
 
 /* The current thread is ready to begin the next work sharing construct.
    In all cases, thr->ts.work_share is updated to point to the new
@@ -75,79 +152,34 @@ gomp_work_share_start (bool ordered)
   struct gomp_thread *thr = gomp_thread ();
   struct gomp_team *team = thr->ts.team;
   struct gomp_work_share *ws;
-  unsigned ws_index, ws_gen;
 
   /* Work sharing constructs can be orphaned.  */
   if (team == NULL)
     {
-      ws = gomp_new_work_share (ordered, 1);
+      ws = gomp_malloc (sizeof (*ws));
+      gomp_init_work_share (ws, ordered, 1);
       thr->ts.work_share = ws;
-      thr->ts.static_trip = 0;
-      gomp_mutex_lock (&ws->lock);
-      return true;
+      return ws;
     }
 
-  gomp_mutex_lock (&team->work_share_lock);
-
-  /* This thread is beginning its next generation.  */
-  ws_gen = ++thr->ts.work_share_generation;
-
-  /* If this next generation is not newer than any other generation in
-     the team, then simply reference the existing construct.  */
-  if (ws_gen - team->oldest_live_gen < team->num_live_gen)
+  ws = thr->ts.work_share;
+  thr->ts.last_work_share = ws;
+  ws = gomp_ptrlock_get (&ws->next_ws);
+  if (ws == NULL)
     {
-      ws_index = ws_gen & team->generation_mask;
-      ws = team->work_shares[ws_index];
+      /* This thread encountered a new ws first.  */
+      struct gomp_work_share *ws = alloc_work_share (team);
+      gomp_init_work_share (ws, ordered, team->nthreads);
       thr->ts.work_share = ws;
-      thr->ts.static_trip = 0;
-
-      gomp_mutex_lock (&ws->lock);
-      gomp_mutex_unlock (&team->work_share_lock);
-
-      return false;
+      return true;
     }
-
-  /* Resize the work shares queue if we've run out of space.  */
-  if (team->num_live_gen++ == team->generation_mask)
+  else
     {
-      if (team->work_shares == team->init_work_shares)
-	{
-	  team->work_shares = gomp_malloc (2 * team->num_live_gen
-					   * sizeof (*team->work_shares));
-	  memcpy (team->work_shares, team->init_work_shares,
-		  sizeof (team->init_work_shares));
-	}
-      else
-	team->work_shares = gomp_realloc (team->work_shares,
-					  2 * team->num_live_gen
-					  * sizeof (*team->work_shares));
-
-      /* Unless oldest_live_gen is zero, the sequence of live elements
-	 wraps around the end of the array.  If we do nothing, we break
-	 lookup of the existing elements.  Fix that by unwrapping the
-	 data from the front to the end.  */
-      if (team->oldest_live_gen > 0)
-	memcpy (team->work_shares + team->num_live_gen,
-		team->work_shares,
-		(team->oldest_live_gen & team->generation_mask)
-		* sizeof (*team->work_shares));
-
-      team->generation_mask = team->generation_mask * 2 + 1;
+      thr->ts.work_share = ws;
+      return false;
     }
-
-  ws_index = ws_gen & team->generation_mask;
-  ws = gomp_new_work_share (ordered, team->nthreads);
-  thr->ts.work_share = ws;
-  thr->ts.static_trip = 0;
-  team->work_shares[ws_index] = ws;
-
-  gomp_mutex_lock (&ws->lock);
-  gomp_mutex_unlock (&team->work_share_lock);
-
-  return true;
 }
 
-
 /* The current thread is done with its current work sharing construct.
    This version does imply a barrier at the end of the work-share.  */
 
@@ -156,15 +188,13 @@ gomp_work_share_end (void)
 {
   struct gomp_thread *thr = gomp_thread ();
   struct gomp_team *team = thr->ts.team;
-  struct gomp_work_share *ws = thr->ts.work_share;
   gomp_barrier_state_t bstate;
 
-  thr->ts.work_share = NULL;
-
   /* Work sharing constructs can be orphaned.  */
   if (team == NULL)
     {
-      free_work_share (ws);
+      free_work_share (NULL, thr->ts.work_share);
+      thr->ts.work_share = NULL;
       return;
     }
 
@@ -172,20 +202,14 @@ gomp_work_share_end (void)
 
   if (gomp_barrier_last_thread (bstate))
     {
-      unsigned ws_index;
-
-      ws_index = thr->ts.work_share_generation & team->generation_mask;
-      team->work_shares[ws_index] = NULL;
-      team->oldest_live_gen++;
-      team->num_live_gen = 0;
-
-      free_work_share (ws);
+      if (__builtin_expect (thr->ts.last_work_share != NULL, 1))
+	free_work_share (team, thr->ts.last_work_share);
     }
 
   gomp_barrier_wait_end (&team->barrier, bstate);
+  thr->ts.last_work_share = NULL;
 }
 
-
 /* The current thread is done with its current work sharing construct.
    This version does NOT imply a barrier at the end of the work-share.  */
 
@@ -197,15 +221,17 @@ gomp_work_share_end_nowait (void)
   struct gomp_work_share *ws = thr->ts.work_share;
   unsigned completed;
 
-  thr->ts.work_share = NULL;
-
   /* Work sharing constructs can be orphaned.  */
   if (team == NULL)
     {
-      free_work_share (ws);
+      free_work_share (NULL, ws);
+      thr->ts.work_share = NULL;
       return;
     }
 
+  if (__builtin_expect (thr->ts.last_work_share == NULL, 0))
+    return;
+
 #ifdef HAVE_SYNC_BUILTINS
   completed = __sync_add_and_fetch (&ws->threads_completed, 1);
 #else
@@ -215,18 +241,6 @@ gomp_work_share_end_nowait (void)
 #endif
 
   if (completed == team->nthreads)
-    {
-      unsigned ws_index;
-
-      gomp_mutex_lock (&team->work_share_lock);
-
-      ws_index = thr->ts.work_share_generation & team->generation_mask;
-      team->work_shares[ws_index] = NULL;
-      team->oldest_live_gen++;
-      team->num_live_gen--;
-
-      gomp_mutex_unlock (&team->work_share_lock);
-
-      free_work_share (ws);
-    }
+    free_work_share (team, thr->ts.last_work_share);
+  thr->ts.last_work_share = NULL;
 }