summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authork0kubun <k0kubun@b2dd03c8-39d4-4d8f-98ff-823fe69b080e>2018-10-21 14:23:24 +0000
committerk0kubun <k0kubun@b2dd03c8-39d4-4d8f-98ff-823fe69b080e>2018-10-21 14:23:24 +0000
commit13df05acfa02aeccac89e93f8789933b23004113 (patch)
treeaf70b90f8ac5c364aaee1a5ea69899c413a03065
parentbca2cdcc0b1d4630fd1cb36ecb5b8dc57cdd8e41 (diff)
mjit.c: copy call cache values to MJIT worker
same as r65275 but for call cache. === Optcarrot Benchmark === $ benchmark-driver benchmark.yml --rbenv 'before::before --disable-gems --jit;after::after --disable-gems --jit' -v --repeat-count 24 before: ruby 2.6.0dev (2018-10-21 trunk 65277) +JIT [x86_64-linux] after: ruby 2.6.0dev (2018-10-21 trunk 65277) +JIT [x86_64-linux] last_commit=mjit.c: copy call cache values to MJIT worker Calculating ------------------------------------- before after Optcarrot Lan_Master.nes 85.372 85.359 fps Comparison: Optcarrot Lan_Master.nes before: 85.4 fps after: 85.4 fps - 1.00x slower git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@65279 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
-rw-r--r--mjit.c7
-rw-r--r--mjit.h2
-rw-r--r--mjit_compile.c6
-rw-r--r--mjit_worker.c49
-rw-r--r--tool/ruby_vm/views/_mjit_compile_send.erb13
5 files changed, 49 insertions, 28 deletions
diff --git a/mjit.c b/mjit.c
index f2b2581..215057a 100644
--- a/mjit.c
+++ b/mjit.c
@@ -25,7 +25,12 @@ static void
mjit_copy_job_handler(void *data)
{
struct mjit_copy_job *job = (struct mjit_copy_job *)data;
- memcpy(job->is_entries, job->body->is_entries, sizeof(union iseq_inline_storage_entry) * job->body->is_size);
+ if (job->cc_entries) {
+ memcpy(job->cc_entries, job->body->cc_entries, sizeof(struct rb_call_cache) * (job->body->ci_size + job->body->ci_kw_size));
+ }
+ if (job->is_entries) {
+ memcpy(job->is_entries, job->body->is_entries, sizeof(union iseq_inline_storage_entry) * job->body->is_size);
+ }
CRITICAL_SECTION_START(3, "in MJIT copy job wait");
job->finish_p = TRUE;
diff --git a/mjit.h b/mjit.h
index 9209389..325fe55 100644
--- a/mjit.h
+++ b/mjit.h
@@ -64,7 +64,7 @@ extern void mjit_add_iseq_to_process(const rb_iseq_t *iseq);
extern VALUE mjit_wait_call(rb_execution_context_t *ec, struct rb_iseq_constant_body *body);
RUBY_SYMBOL_EXPORT_END
-extern int mjit_compile(FILE *f, const struct rb_iseq_constant_body *body, const char *funcname, union iseq_inline_storage_entry *is_entries);
+extern int mjit_compile(FILE *f, const struct rb_iseq_constant_body *body, const char *funcname, struct rb_call_cache *cc_entries, union iseq_inline_storage_entry *is_entries);
extern void mjit_init(struct mjit_options *opts);
extern void mjit_finish(void);
extern void mjit_gc_start_hook(void);
diff --git a/mjit_compile.c b/mjit_compile.c
index 7cf3a47..07e417e 100644
--- a/mjit_compile.c
+++ b/mjit_compile.c
@@ -34,8 +34,9 @@ struct compile_status {
/* If TRUE, JIT-ed code will use local variables to store pushed values instead of
using VM's stack and moving stack pointer. */
int local_stack_p;
- /* Safely-accessible is_entries copied from main thread. */
+ /* Safely-accessible cache entries copied from main thread. */
union iseq_inline_storage_entry *is_entries;
+ struct rb_call_cache *cc_entries;
};
/* Storage to keep data which is consistent in each conditional branch.
@@ -197,7 +198,7 @@ compile_cancel_handler(FILE *f, const struct rb_iseq_constant_body *body, struct
/* Compile ISeq to C code in F. It returns 1 if it succeeds to compile. */
int
-mjit_compile(FILE *f, const struct rb_iseq_constant_body *body, const char *funcname, union iseq_inline_storage_entry *is_entries)
+mjit_compile(FILE *f, const struct rb_iseq_constant_body *body, const char *funcname, struct rb_call_cache *cc_entries, union iseq_inline_storage_entry *is_entries)
{
struct compile_status status;
status.success = TRUE;
@@ -206,6 +207,7 @@ mjit_compile(FILE *f, const struct rb_iseq_constant_body *body, const char *func
if (status.stack_size_for_pos == NULL)
return FALSE;
memset(status.stack_size_for_pos, NOT_COMPILED_STACK_SIZE, sizeof(int) * body->iseq_size);
+ status.cc_entries = cc_entries;
status.is_entries = is_entries;
/* For performance, we verify stack size only on compilation time (mjit_compile.inc.erb) without --jit-debug */
diff --git a/mjit_worker.c b/mjit_worker.c
index 0dcd253..2234dcd 100644
--- a/mjit_worker.c
+++ b/mjit_worker.c
@@ -1029,7 +1029,7 @@ compile_prelude(FILE *f)
/* Compile ISeq in UNIT and return function pointer of JIT-ed code.
It may return NOT_COMPILED_JIT_ISEQ_FUNC if something went wrong. */
static mjit_func_t
-convert_unit_to_func(struct rb_mjit_unit *unit, union iseq_inline_storage_entry *is_entries)
+convert_unit_to_func(struct rb_mjit_unit *unit, struct rb_call_cache *cc_entries, union iseq_inline_storage_entry *is_entries)
{
char c_file_buff[MAXPATHLEN], *c_file = c_file_buff, *so_file, funcname[35]; /* TODO: reconsider `35` */
int success;
@@ -1097,7 +1097,7 @@ convert_unit_to_func(struct rb_mjit_unit *unit, union iseq_inline_storage_entry
verbose(2, "start compilation: %s@%s:%d -> %s", label, path, lineno, c_file);
fprintf(f, "/* %s@%s:%d */\n\n", label, path, lineno);
}
- success = mjit_compile(f, unit->iseq->body, funcname, is_entries);
+ success = mjit_compile(f, unit->iseq->body, funcname, cc_entries, is_entries);
/* release blocking mjit_gc_start_hook */
CRITICAL_SECTION_START(3, "after mjit_compile to wakeup client for GC");
@@ -1163,12 +1163,29 @@ convert_unit_to_func(struct rb_mjit_unit *unit, union iseq_inline_storage_entry
struct mjit_copy_job {
const struct rb_iseq_constant_body *body;
+ struct rb_call_cache *cc_entries;
union iseq_inline_storage_entry *is_entries;
int finish_p;
};
static void mjit_copy_job_handler(void *data);
+/* We're lazily copying cache values from main thread because these cache values
+ could be different between ones on enqueue timing and ones on dequeue timing. */
+static void
+copy_cache_from_main_thread(struct mjit_copy_job *job)
+{
+ job->finish_p = FALSE;
+
+ rb_postponed_job_register(0, mjit_copy_job_handler, (void *)job);
+ CRITICAL_SECTION_START(3, "in MJIT copy job wait");
+ while (!job->finish_p) {
+ rb_native_cond_wait(&mjit_worker_wakeup, &mjit_engine_mutex);
+ verbose(3, "Getting wakeup from client");
+ }
+ CRITICAL_SECTION_FINISH(3, "in MJIT copy job wait");
+}
+
/* The function implementing a worker. It is executed in a separate
thread by rb_thread_create_mjit_thread. It compiles precompiled header
and then compiles requested ISeqs. */
@@ -1207,24 +1224,21 @@ mjit_worker(void)
mjit_func_t func;
struct mjit_copy_job job;
- /* Copy ISeq's inline caches from main thread. */
- job.is_entries = NULL;
job.body = node->unit->iseq->body;
- if (job.body->is_size > 0) {
- job.is_entries = malloc(sizeof(union iseq_inline_storage_entry) * job.body->is_size);
- job.finish_p = FALSE;
-
- rb_postponed_job_register(0, mjit_copy_job_handler, (void *)&job);
- CRITICAL_SECTION_START(3, "in MJIT copy job wait");
- while (!job.finish_p) {
- rb_native_cond_wait(&mjit_worker_wakeup, &mjit_engine_mutex);
- verbose(3, "Getting wakeup from client");
- }
- CRITICAL_SECTION_FINISH(3, "in MJIT copy job wait");
+ job.cc_entries = NULL;
+ if (job.body->ci_size > 0 || job.body->ci_kw_size > 0)
+ job.cc_entries = alloca(sizeof(struct rb_call_cache) * (job.body->ci_size + job.body->ci_kw_size));
+ job.is_entries = NULL;
+ if (job.body->is_size > 0)
+ job.is_entries = alloca(sizeof(union iseq_inline_storage_entry) * job.body->is_size);
+
+ /* Copy ISeq's inline caches values to avoid race condition. */
+ if (job.cc_entries != NULL || job.is_entries != NULL) {
+ copy_cache_from_main_thread(&job);
}
/* JIT compile */
- func = convert_unit_to_func(node->unit, job.is_entries);
+ func = convert_unit_to_func(node->unit, job.cc_entries, job.is_entries);
CRITICAL_SECTION_START(3, "in jit func replace");
if (node->unit->iseq) { /* Check whether GCed or not */
@@ -1241,9 +1255,6 @@ mjit_worker(void)
compact_all_jit_code();
}
#endif
- if (job.is_entries != NULL) {
- free(job.is_entries);
- }
}
}
diff --git a/tool/ruby_vm/views/_mjit_compile_send.erb b/tool/ruby_vm/views/_mjit_compile_send.erb
index 4104a76..4ba97ac 100644
--- a/tool/ruby_vm/views/_mjit_compile_send.erb
+++ b/tool/ruby_vm/views/_mjit_compile_send.erb
@@ -13,6 +13,9 @@
% insn.opes.each_with_index do |ope, i|
MAYBE_UNUSED(<%= ope.fetch(:decl) %>) = (<%= ope.fetch(:type) %>)operands[<%= i %>];
% end
+% # compiler: Use copied cc to avoid race condition
+ CALL_CACHE cc_copy = status->cc_entries + (cc - body->cc_entries);
+ cc_copy = cc;
%
if (has_valid_method_type(cc)) {
const rb_iseq_t *iseq;
@@ -21,7 +24,7 @@
argc += ((ci->flag & VM_CALL_ARGS_BLOCKARG) ? 1 : 0);
% end
- if (cc->me->def->type == VM_METHOD_TYPE_ISEQ && inlinable_iseq_p(ci, cc, iseq = rb_iseq_check(cc->me->def->body.iseq.iseqptr))) { /* CC_SET_FASTPATH in vm_callee_setup_arg */
+ if (cc_copy->me->def->type == VM_METHOD_TYPE_ISEQ && inlinable_iseq_p(ci, cc, iseq = rb_iseq_check(cc_copy->me->def->body.iseq.iseqptr))) { /* CC_SET_FASTPATH in vm_callee_setup_arg */
int param_size = iseq->body->param.size; /* TODO: check calling->argc for argument_arity_error */
fprintf(f, "{\n");
@@ -31,8 +34,8 @@
}
% # JIT: Invalidate call cache if it requires vm_search_method. This allows to inline some of following things.
- fprintf(f, " if (UNLIKELY(GET_GLOBAL_METHOD_STATE() != %"PRI_SERIALT_PREFIX"u ||\n", cc->method_state);
- fprintf(f, " RCLASS_SERIAL(CLASS_OF(stack[%d])) != %"PRI_SERIALT_PREFIX"u)) {\n", b->stack_size - 1 - argc, cc->class_serial);
+ fprintf(f, " if (UNLIKELY(GET_GLOBAL_METHOD_STATE() != %"PRI_SERIALT_PREFIX"u ||\n", cc_copy->method_state);
+ fprintf(f, " RCLASS_SERIAL(CLASS_OF(stack[%d])) != %"PRI_SERIALT_PREFIX"u)) {\n", b->stack_size - 1 - argc, cc_copy->class_serial);
fprintf(f, " reg_cfp->pc = original_body_iseq + %d;\n", pos);
fprintf(f, " reg_cfp->sp = (VALUE *)reg_cfp->bp + %d;\n", b->stack_size + 1);
fprintf(f, " goto cancel;\n");
@@ -52,11 +55,11 @@
fprintf(f, " calling.argc = %d;\n", ci->orig_argc);
fprintf(f, " calling.recv = stack[%d];\n", b->stack_size - 1 - argc);
-% # JIT: Special CALL_METHOD. Bypass cc->call and inline vm_call_iseq_setup_normal for vm_call_iseq_setup_func FASTPATH.
+% # JIT: Special CALL_METHOD. Bypass cc_copy->call and inline vm_call_iseq_setup_normal for vm_call_iseq_setup_func FASTPATH.
fprintf(f, " {\n");
fprintf(f, " VALUE v;\n");
fprintf(f, " vm_call_iseq_setup_normal(ec, reg_cfp, &calling, (const rb_callable_method_entry_t *)0x%"PRIxVALUE", 0, %d, %d);\n",
- (VALUE)cc->me, param_size, iseq->body->local_table_size); /* rb_simple_iseq_p checks rb_simple_iseq_p, which ensures has_opt == FALSE */
+ (VALUE)cc_copy->me, param_size, iseq->body->local_table_size); /* rb_simple_iseq_p checks rb_simple_iseq_p, which ensures has_opt == FALSE */
if (iseq->body->catch_except_p) {
fprintf(f, " VM_ENV_FLAGS_SET(ec->cfp->ep, VM_FRAME_FLAG_FINISH);\n");
fprintf(f, " v = vm_exec(ec, TRUE);\n");