mjit.c: copy inline cache values to MJIT worker

on VM_CHECK_INTS. Letting MJIT worker directly see inline cache which
may be being updated could result in inconsistent IC index and serial.

mjit_worker.c: request the copy job after dequeue, and receive the
result synchronously.

tool/ruby_vm/views/_mjit_compile_ivar.erb: use the copied IC

mjit_compile.c: change the interface to pass is_entries
mjit.h: ditto

=== Optcarrot Benchmark ===
Thankfully this didn't have major performance regression.

$ 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 65263) +JIT [x86_64-linux]
after: ruby 2.6.0dev (2018-10-21 trunk 65263) +JIT [x86_64-linux]
last_commit=mjit.c: copy inline cache values to MJIT worker
Calculating -------------------------------------
                             before       after
Optcarrot Lan_Master.nes     85.421      85.454 fps

Comparison:
             Optcarrot Lan_Master.nes
                   after:        85.5 fps
                  before:        85.4 fps - 1.00x  slower

git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@65275 b2dd03c8-39d4-4d8f-98ff-823fe69b080e

mjit.c

@@ -20,6 +20,19 @@
 #include "constant.h"
 #include "id_table.h"
 
+/* Copy ISeq's states so that race condition does not happen on compilation. */
+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);
+
+    CRITICAL_SECTION_START(3, "in MJIT copy job wait");
+    job->finish_p = TRUE;
+    rb_native_cond_broadcast(&mjit_worker_wakeup);
+    CRITICAL_SECTION_FINISH(3, "in MJIT copy job wait");
+}
+
 extern int rb_thread_create_mjit_thread(void (*worker_func)(void));
 
 /* Return an unique file name in /tmp with PREFIX and SUFFIX and

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);
+extern int mjit_compile(FILE *f, const struct rb_iseq_constant_body *body, const char *funcname, 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);

mjit_compile.c

@@ -34,6 +34,8 @@ 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. */
+    union iseq_inline_storage_entry *is_entries;
 };
 
 /* Storage to keep data which is consistent in each conditional branch.
@@ -195,7 +197,7 @@ 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)
+mjit_compile(FILE *f, const struct rb_iseq_constant_body *body, const char *funcname, union iseq_inline_storage_entry *is_entries)
 {
     struct compile_status status;
     status.success = TRUE;
@@ -204,6 +206,7 @@ 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.is_entries = is_entries;
 
     /* For performance, we verify stack size only on compilation time (mjit_compile.inc.erb) without --jit-debug */
     if (!mjit_opts.debug) {

mjit_worker.c

@@ -76,6 +76,7 @@
 #include "mjit.h"
 #include "gc.h"
 #include "ruby_assert.h"
+#include "ruby/debug.h"
 #include "ruby/thread.h"
 
 #ifdef _WIN32
@@ -1028,7 +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)
+convert_unit_to_func(struct rb_mjit_unit *unit, 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;
@@ -1096,7 +1097,7 @@ convert_unit_to_func(struct rb_mjit_unit *unit)
         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);
+    success = mjit_compile(f, unit->iseq->body, funcname, is_entries);
 
     /* release blocking mjit_gc_start_hook */
     CRITICAL_SECTION_START(3, "after mjit_compile to wakeup client for GC");
@@ -1160,6 +1161,14 @@ convert_unit_to_func(struct rb_mjit_unit *unit)
     return (mjit_func_t)func;
 }
 
+struct mjit_copy_job {
+    const struct rb_iseq_constant_body *body;
+    union iseq_inline_storage_entry *is_entries;
+    int finish_p;
+};
+
+static void mjit_copy_job_handler(void *data);
+
 /* 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. */
@@ -1195,7 +1204,27 @@ mjit_worker(void)
         CRITICAL_SECTION_FINISH(3, "in worker dequeue");
 
         if (node) {
-            mjit_func_t func = convert_unit_to_func(node->unit);
+            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");
+            }
+
+            /* JIT compile */
+            func = convert_unit_to_func(node->unit, job.is_entries);
 
             CRITICAL_SECTION_START(3, "in jit func replace");
             if (node->unit->iseq) { /* Check whether GCed or not */
@@ -1212,6 +1241,9 @@ mjit_worker(void)
                 compact_all_jit_code();
             }
 #endif
+            if (job.is_entries != NULL) {
+                free(job.is_entries);
+            }
         }
     }
 

tool/ruby_vm/views/_mjit_compile_ivar.erb

@@ -13,17 +13,19 @@
 % insn.opes.each_with_index do |ope, i|
     MAYBE_UNUSED(<%= ope.fetch(:decl) %>) = (<%= ope.fetch(:type) %>)operands[<%= i %>];
 % end
+% # compiler: Use copied IC to avoid race condition
+    IC ic_copy = &(status->is_entries + ((union iseq_inline_storage_entry *)ic - body->is_entries))->cache;
 %
-% # compiler: Consider cfp->self as T_OBJECT if ic->ic_serial is set
+% # compiler: Consider cfp->self as T_OBJECT if ic_copy->ic_serial is set
-    if (ic->ic_serial) {
+    if (ic_copy->ic_serial) {
 % # JIT: optimize away motion of sp and pc. This path does not call rb_warning() and so it's always leaf and not `handles_sp`.
 % # <%= render 'mjit_compile_pc_and_sp', locals: { insn: insn } -%>
 %
 % # JIT: prepare vm_getivar's arguments and variables
         fprintf(f, "{\n");
         fprintf(f, "    VALUE obj = GET_SELF();\n");
-        fprintf(f, "    const rb_serial_t ic_serial = (rb_serial_t)%"PRI_SERIALT_PREFIX"u;\n", ic->ic_serial);
+        fprintf(f, "    const rb_serial_t ic_serial = (rb_serial_t)%"PRI_SERIALT_PREFIX"u;\n", ic_copy->ic_serial);
-        fprintf(f, "    const st_index_t index = %"PRIuSIZE";\n", ic->ic_value.index);
+        fprintf(f, "    const st_index_t index = %"PRIuSIZE";\n", ic_copy->ic_value.index);
 % if insn.name == 'setinstancevariable'
         fprintf(f, "    VALUE val = stack[%d];\n", b->stack_size - 1);
 % end