# HG changeset patch
# Parent f20811cd2fde8357b5c51307decee8c877c1e16a
Import libcilkrts Build 4467 (PR target/68945)
@@ -54,7 +54,7 @@ GENERAL_FLAGS = -I$(top_srcdir)/include
# Enable Intel Cilk Plus extension
GENERAL_FLAGS += -fcilkplus
-# Always generate unwind tables
+#Always generate unwind tables
GENERAL_FLAGS += -funwind-tables
AM_CFLAGS = $(XCFLAGS) $(GENERAL_FLAGS) -std=c99
@@ -1,14 +1,16 @@
-Intel(R) Cilk(TM) Plus runtime library
+Intel(R) Cilk(TM) Plus Runtime Library
Index:
-1. BUILDING
-2. USING
-3. DOXYGEN DOCUMENTATION
-4. QUESTIONS OR BUGS
-5. CONTRIBUTIONS
+1. BUILDING WITH AUTOMAKE
+2. BUILDING WITH CMAKE
+3. INSTALLING TO VXWORKS
+4. USING
+5. DOXYGEN DOCUMENTATION
+6. QUESTIONS OR BUGS
+7. CONTRIBUTIONS
#
-# 1. BUILDING:
+# 1. BUILDING WITH AUTOMAKE:
#
To distribute applications that use the Intel Cilk Plus language
@@ -40,22 +42,87 @@ configure script:
% ./configure --prefix=/your/path/to/lib
-It is also possible to use CMake if the above method does not apply
-well in your environment. Instruction is available in CMakeLists.txt.
+#
+# 2. BUILDING WITH CMAKE:
+#
+
+To distribute applications that use the Intel Cilk Plus language
+extensions to non-development systems, you need to build the Intel
+Cilk Plus runtime library and distribute it with your application.
+This instruction describes the build process using CMake*, which
+supports Linux*, Windows*, and OS X*. It is fine to use this process
+to build a Linux library, but it is highly recommended to use the
+more mature build process described above when building on Linux.
+
+You need the CMake tool and a C/C++ compiler that supports the Intel
+Cilk Plus language extensions, and the requirements for each operating
+systems are:
+
+Common:
+ CMake 3.0.0 or later
+ Make tools such as make (Linux, OS X) or nmake (Windows)
+Linux:
+ GCC* 4.9.2 or later, or Intel(R) C++ Compiler v12.1 or later
+Windows:
+ Intel C++ Compiler v12.1 or later
+ Visual Studio* 2010 or later
+OS X:
+ Cilk-enabled branch of Clang*/LLVM* (http://cilkplus.github.io),
+ or Intel C++ Compiler v12.1 or later
+
+The common steps to build the libraries are 1) invoke cmake with
+appropriate options, 2) invoke a make tool available on the system.
+The following examples show build processes on OS X and Windows.
+
+OS X:
+ % mkdir ./build && cd ./build
+ % cmake -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ \
+ -DCMAKE_INSTALL_PREFIX=./install ..
+ % make && make install
+
+Windows:
+ % mkdir .\build && cd .\build
+ % cmake -G "NMake Makefiles" -DCMAKE_C_COMPILER=icl \
+ -DCMAKE_CXX_COMPILER=icl -DCMAKE_INSTALL_PREFIX=.\install ..
+ % nmake && nmake install
#
-# 2. USING:
+# 3. INSTALLING TO VXWORKS OS
+#
+
+For Windows host, run VxWorks_Install.bat.
+For Linux host, run VxWorks_Install.sh.
+
+You may need to give environment variable WIND_BASE to indicate VxWorks
+installation path.
+
+Create a VSB project, and you will see a layer named "CILKPLUS_KERNEL".
+Enable it and build the project.
+
+Create a VIP project with ICC and add component INCLUDE_CILKPLUS, then you
+will get support of Intel Cilk Plus features in VxWorks.
+
+#
+# 4. USING:
#
The Intel(R) C++ Compiler will automatically try to bring in the
Intel Cilk Plus runtime in any program that uses the relevant
-features. GCC requires explicit linking of both the library and
-its dependencies (libpthread, libdl). For example:
+features. GCC and Clang requires an explicit compiler option,
+-fcilkplus, to enable Intel Cilk Plus language extensions.
+For example,
+
+% gcc -fcilkplus -o foo.exe foo.c
+% clang -fcilkplus -o foo.exe foo.c
+
+Older GCC versions (e.g., 4.8 cilkplus branch) requires explicit linking
+of both the library and its dependencies (libpthread, libdl).
+For example:
% gcc foo.c -lcilkrts -lpthread -ldl
#
-# 3. DOXYGEN DOCUMENTATION:
+# 5. DOXYGEN DOCUMENTATION:
#
The library source has Doxygen markup. Generate HTML documentation
@@ -64,7 +131,7 @@ based on the markup by changing director
% doxygen doxygen.cfg
#
-# 4. QUESTIONS OR BUGS:
+# 6. QUESTIONS OR BUGS:
#
Issues with the Intel Cilk Plus runtime can be addressed in the Intel
@@ -72,7 +139,7 @@ Cilk Plus forums:
http://software.intel.com/en-us/forums/intel-cilk-plus/
#
-# 5. CONTRIBUTIONS:
+# 7. CONTRIBUTIONS:
#
The Intel Cilk Plus runtime library is dual licensed. The upstream copy
@@ -85,6 +152,13 @@ contributed to the upstream version via
Thanks to Tobias Burnus for showing us the magic to make gcc and g++
automatically include the Cilk Plus runtime.
+Thanks to Eric Olson for sharing his patch for Raspberry Pi* with us.
+
+Thanks to Rainer Orth for submitting patches for exception handling and
+enabling Cilk Plus on the SPARC* architecture.
+
------------------------
Intel and Cilk are trademarks of Intel Corporation in the U.S. and/or
other countries.
+
+*Other names and brands may be claimed as the property of others.
@@ -69,6 +69,10 @@ AC_PROG_CXX
AC_CONFIG_FILES([Makefile libcilkrts.spec])
AC_FUNC_ALLOCA
+# Check for dl functions
+AC_CHECK_LIB(dl, dladdr,
+ [AC_DEFINE(HAVE_DLADDR, 1, [Define if you have dladdr()])])
+
# Check whether the target supports protected visibility.
save_CFLAGS="$CFLAGS"
CFLAGS="$CFLAGS -Werror"
@@ -141,11 +145,7 @@ esac
# contains information on what's needed
case "${target}" in
- x86_64-*-*)
- config_dir="x86"
- ;;
-
- i?86-*-*)
+ i?86-*-* | x86_64-*-*)
config_dir="x86"
;;
@@ -153,6 +153,10 @@ case "${target}" in
config_dir="arm"
;;
+ sparc*-*-*)
+ config_dir="sparc"
+ ;;
+
*)
config_dir="generic"
;;
@@ -44,12 +44,12 @@
# Disable Cilk Runtime library for unsupported architectures.
case "${target}" in
- x86_64-*-*)
- ;;
- i?86-*-*)
+ i?86-*-* | x86_64-*-*)
;;
arm-*-*)
;;
+ sparc*-*-*)
+ ;;
*-*-*)
UNSUPPORTED=1
;;
@@ -106,9 +106,6 @@ size_t __cilkrts_get_stack_size(void);
CILK_EXPORT __CILKRTS_NOTHROW
void __cilkrts_dump_stats(void);
-CILK_EXPORT __CILKRTS_NOTHROW
-int __cilkrts_irml_version(void);
-
struct __cilk_tbb_unwatch_thunk;
struct __cilk_tbb_stack_op_thunk;
@@ -49,10 +49,10 @@
#define VERSION_MAJOR 2
#define VERSION_MINOR 0
-#define VERSION_BUILD 4420
+#define VERSION_BUILD 4467
#define VERSION_REV 0
-#define VERSION_STRING "2,0,4420,0"
-#define VERSION_HASH "3b2d6aa9059c"
+#define VERSION_STRING "2,0,4467,0"
+#define VERSION_HASH "b7e54d87bd17"
#define VERSION_BRANCH "eng"
-#define TBB_REV_NUMBER "14788"
-#define VERSION_YEAR "2015"
+#define TBB_REV_NUMBER ""
+#define VERSION_YEAR "2016"
@@ -49,4 +49,4 @@
#
# It was automatically generated by cilkrts/include/internal/Makefile
-CILK_REVISION = 4420
+CILK_REVISION = 4467
@@ -709,7 +709,6 @@ CILK_API_TBB_RETCODE
return 0; /* Success! */
}
-
// This function must be called only within a continuation, within the stack
// frame of the continuation itself.
CILK_API_INT __cilkrts_synched(void)
@@ -220,7 +220,7 @@ NORETURN cilk_fiber_sysdep::run()
// enough extra space from the top of the stack we are
// switching to for any temporaries required for this run()
// function.
- JMPBUF_SP(m_resume_jmpbuf) = m_stack_base - frame_size;
+ JMPBUF_SP(m_resume_jmpbuf) = CILK_ADJUST_SP(m_stack_base - frame_size);
// GCC doesn't allow us to call __builtin_longjmp in the same function
// that calls __builtin_setjmp, so it's been moved into it's own
@@ -73,9 +73,12 @@
*
* A value of 0 means no debugging.
* Higher values generate more debugging output.
+ *
*/
+
+#ifndef FIBER_DEBUG
#define FIBER_DEBUG 0
-
+#endif
/**
* @brief Flag for validating reference counts.
*
@@ -47,6 +47,14 @@
* for your assistance in helping us improve Cilk Plus.
**************************************************************************/
+// __atomic_* intrinsics are available since GCC 4.7.
+#define HAVE_ATOMIC_INTRINSICS defined(__GNUC__) && \
+ (__GNUC__ * 10 + __GNUC_MINOR__ >= 47)
+
+// GCC before 4.4 does not implement __sync_synchronize properly
+#define HAVE_SYNC_INTRINSICS defined(__GNUC__) && \
+ (__GNUC__ * 10 + __GNUC_MINOR__ >= 44)
+
/*
* void __cilkrts_fence(void)
*
@@ -60,5 +68,12 @@
* the CPUID instruction).
*/
-// COMMON_SYSDEP void __cilkrts_fence(void); ///< MFENCE instruction
-# define __cilkrts_fence() __asm__ __volatile__ ("mcr p15,0,%[t],c7,c10,4\n" :: [t] "r" (0) : "memory");
+#if HAVE_ATOMIC_INTRINSICS
+# define __cilkrts_fence() __atomic_thread_fence(__ATOMIC_SEQ_CST)
+#elif HAVE_SYNC_INTRINSICS
+# define __cilkrts_fence() __sync_synchronize()
+#else
+# define __cilkrts_fence()
+// Leaving this code just in case.
+//# define __cilkrts_fence() __asm__ __volatile__ ("mcr p15,0,%[t],c7,c10,4\n" :: [t] "r" (0) : "memory");
+#endif
@@ -60,7 +60,9 @@
#include <stdint.h>
#include <typeinfo>
+#ifndef DEBUG_EXCEPTIONS
#define DEBUG_EXCEPTIONS 0
+#endif
struct pending_exception_info
{
@@ -579,7 +579,6 @@ global_state_t* cilkg_init_global_state(
g->stack_size = cilkos_validate_stack_size(g->stack_size);
g->failure_to_allocate_stack = 0;
-
return g;
}
@@ -215,7 +215,6 @@ struct global_state_t { /* COMMON_PORTAB
/// Global fiber pool
cilk_fiber_pool fiber_pool;
-
/**
* @brief Track whether the runtime has failed to allocate a
* stack.
@@ -107,6 +107,18 @@
*/
#define SP(SF) JMPBUF_SP((SF)->ctx)
+/**
+ * @brief Some architecture-dependent stack adjustment.
+ */
+#if defined(__sparcv9)
+ // Subtract sparc v9 stack bias so the actual stack starts at the
+ // allocated area.
+# define CILK_ADJUST_SP(SP) ((SP) - 2047)
+# define CILK_UNADJUST_SP(SP) ((SP) + 2047)
+#else
+# define CILK_ADJUST_SP(SP) (SP)
+# define CILK_UNADJUST_SP(SP) (SP)
+#endif
__CILKRTS_BEGIN_EXTERN_C
@@ -120,7 +132,7 @@
*/
inline char *__cilkrts_get_sp(__cilkrts_stack_frame *sf)
{
- return (char *)SP(sf);
+ return (char *)CILK_UNADJUST_SP(SP(sf));
}
/**
@@ -71,7 +71,6 @@ CILKABI0
__cilkrts_hyperobject_dealloc;
__cilkrts_hyperobject_noop_destroy;
__cilkrts_init;
- __cilkrts_irml_version;
__cilkrts_leave_frame;
__cilkrts_metacall;
__cilkrts_rethrow;
@@ -360,6 +360,13 @@ struct local_state /* COMMON_PORTABLE *
unsigned int steal_failure_count;
/**
+ * Record indicating that the worker stole work at least once.
+ *
+ * [local read/write]
+ */
+ int has_stolen;
+
+ /**
* 1 if work was stolen from another worker. When true, this will flag
* setup_for_execution_pedigree to increment the pedigree when we resume
* execution to match the increment that would have been done on a return
@@ -33,7 +33,6 @@
___cilkrts_hyperobject_dealloc
___cilkrts_hyperobject_noop_destroy
___cilkrts_init
-
___cilkrts_leave_frame
___cilkrts_metacall
___cilkrts_resume
@@ -432,7 +432,9 @@ COMMON_SYSDEP void __cilkrts_idle(void)
#elif defined(__MIC__)
_mm_delay_32(1024);
#elif defined(__linux__) || \
- defined(__APPLE__)
+ defined(__APPLE__) || \
+ defined(__CYGWIN__)
+
usleep(10000);
#else
# error "Unsupported architecture"
@@ -452,6 +454,7 @@ COMMON_SYSDEP void __cilkrts_yield(void)
{
#if defined(__ANDROID__) || \
defined(__APPLE__) || \
+ defined(__CYGWIN__) || \
defined(__FreeBSD__) || \
defined(__VXWORKS__) || \
(defined(__sun__) && defined(__svr4__))
@@ -262,7 +262,7 @@ char * walk_pedigree_nodes(char *p, cons
if (pnode->parent)
{
p = walk_pedigree_nodes(p, pnode->parent);
- p += cilk_snprintf_s(p, PEDIGREE_BUFF_SIZE, "%s", (char *) "_");
+ p += cilk_snprintf_s(p, PEDIGREE_BUFF_SIZE, "%s", (char *)"_");
}
return p + cilk_snprintf_l(p, PEDIGREE_BUFF_SIZE, "%" PRIu64, pnode->rank);
}
@@ -1789,20 +1789,27 @@ static full_frame* check_for_work(__cilk
if (NULL == ff) {
// Punish the worker for failing to steal.
// No quantum for you!
- if (w->l->steal_failure_count > 30000) {
- // Punish more if the worker has been doing unsuccessful steals
- // for a long time. After return from the idle state, it will
- // be given a grace period to react quickly.
+ unsigned int max_fails = w->g->max_steal_failures << 1;
+ if (w->l->has_stolen == 0 &&
+ w->l->steal_failure_count % max_fails == max_fails - 1) {
+ // Idle briefly if the worker has never stolen anything for
+ // the given grace period
__cilkrts_idle();
- w->l->steal_failure_count -= 300;
} else {
__cilkrts_yield();
}
w->l->steal_failure_count++;
+ if (w->l->steal_failure_count > (max_fails << 8)) {
+ // Reset the flag after certain amount of failures
+ // - This will reduce cpu time in top-level synched regions
+ // - max_fails can be controlled by user (CILK_STEAL_FAILURES)
+ w->l->has_stolen = 0;
+ }
} else {
// Reset steal_failure_count since there is obviously still work to
// be done.
w->l->steal_failure_count = 0;
+ w->l->has_stolen = 1;
}
}
return ff;
@@ -2912,6 +2919,7 @@ void __cilkrts_restore_stealing(
w->l->stats = NULL;
#endif
w->l->steal_failure_count = 0;
+ w->l->has_stolen = 0;
w->l->work_stolen = 0;
@@ -85,7 +85,10 @@
* Print debugging messages and assertions for parallel reducers. 0 is
* no debugging. A higher value generates more output.
*/
+
+#ifndef REDPAR_DEBUG
#define REDPAR_DEBUG 0
+#endif
/**
* @brief Lock the worker mutex to allow exclusive access to the
@@ -465,7 +465,7 @@ char* sysdep_reset_jump_buffers_for_resu
CILK_ASSERT(fiber);
void* sp = (void*)get_sp_for_executing_sf(cilk_fiber_get_stack_base(fiber), ff, sf);
- SP(sf) = sp;
+ SP(sf) = CILK_ADJUST_SP(sp);
/* Debugging: make sure stack is accessible. */
((volatile char *)sp)[-1];
@@ -495,7 +495,7 @@ NORETURN sysdep_longjmp_to_sf(char* new_
#endif
// Set the stack pointer.
- SP(sf) = new_sp;
+ SP(sf) = CILK_ADJUST_SP(new_sp);
#ifdef RESTORE_X86_FP_STATE
if (CILK_FRAME_VERSION_VALUE(sf->flags) >= 1) {
@@ -568,7 +568,7 @@ static const char *get_runtime_path ()
{
// dladdr is a glibc extension. If it's available, use it to find the path
// for libcilkrts.so
-#ifdef _GNU_SOURCE
+#if HAVE_DLADDR
Dl_info info;
if (0 != dladdr(dummy_function, &info))
return info.dli_fname;
@@ -689,7 +689,6 @@ static void write_version_file (global_s
// ==================
// System cores: 8
// Cilk workers requested: 8
- // Thread creator: Private
fprintf(fp, "\nThread information\n");
fprintf(fp, "==================\n");
@@ -699,11 +698,6 @@ static void write_version_file (global_s
fprintf(fp, "System cores: %d\n", (int)sysconf(_SC_NPROCESSORS_ONLN));
#endif
fprintf(fp, "Cilk workers requested: %d\n", n);
-#if (PARALLEL_THREAD_CREATE)
- fprintf(fp, "Thread creator: Private (parallel)\n");
-#else
- fprintf(fp, "Thread creator: Private\n");
-#endif
if (fp != stderr && fp != stdout) fclose(fp);
else fflush(fp); // flush the handle buffer if it is stdout or stderr.