From: Yusuke ENDOH <mame@...>
Date: 2008-11-18T00:44:53+09:00
Subject: [ruby-dev:37106] [Feature:1.9] speed up continuation in 1.9

遠藤です。

1.9 の継続は 1.8 に比べて極端に遅いようです。


$ time ruby18 -e 'i = 0; callcc {|c| $c = c }; i += 1; $c.call if i < 1000000'
real    0m1.060s
user    0m1.050s
sys     0m0.010s

$ time ruby19 -rcontinuation -e 'i = 0; callcc {|c| $c = c }; i += 1;
$c.call if i < 1000000'
real    1m57.022s
user    1m56.780s
sys     0m0.180s


capture や call の際、VM のスタックを常に丸ごとコピーしているのが
原因で、必要なところ (先頭の sp 部分と終端の cfp 部分) だけコピー
するようにしたら、1.8 並に速くなりました。


$ time ./ruby.fast-cont -rcontinuation -e 'i = 0; callcc {|c| $c = c
}; i += 1; $c.call if i < 1000000'

real    0m0.660s
user    0m0.660s
sys     0m0.000s


私の環境で test-all が通ることは確認しています。
とくに異論がなければコミットしようと思います。


Index: cont.c
===================================================================
--- cont.c	(revision 20241)
+++ cont.c	(working copy)
@@ -14,6 +14,8 @@
 #include "gc.h"
 #include "eval_intern.h"

+#define CAPTURE_JUST_VALID_VM_STACK 1
+
 enum context_type {
     CONTINUATION_CONTEXT = 0,
     FIBER_CONTEXT = 1,
@@ -25,6 +27,10 @@
     VALUE self;
     VALUE value;
     VALUE *vm_stack;
+#ifdef CAPTURE_JUST_VALID_VM_STACK
+    int vm_stack_slen;  /* length of stack (head of th->stack) */
+    int vm_stack_clen;  /* length of control frames (tail of th->stack) */
+#endif
     VALUE *machine_stack;
     VALUE *machine_stack_src;
 #ifdef __ia64
@@ -75,8 +81,13 @@
 	rb_thread_mark(&cont->saved_thread);

 	if (cont->vm_stack) {
+#ifdef CAPTURE_JUST_VALID_VM_STACK
 	    rb_gc_mark_locations(cont->vm_stack,
-				 cont->vm_stack + cont->saved_thread.stack_size);
+				 cont->vm_stack + cont->vm_stack_slen + cont->vm_stack_clen);
+#elif
+	    rb_gc_mark_localtion(cont->vm_stack,
+				 cont->vm_stack, cont->saved_thread.stack_size);
+#endif
 	}

 	if (cont->machine_stack) {
@@ -247,8 +258,16 @@
     contval = cont->self;
     sth = &cont->saved_thread;

+#ifdef CAPTURE_JUST_VALID_VM_STACK
+    cont->vm_stack_slen = th->cfp->sp + th->mark_stack_len - th->stack;
+    cont->vm_stack_clen = th->stack + th->stack_size - (VALUE*)th->cfp;
+    cont->vm_stack = ALLOC_N(VALUE, cont->vm_stack_slen + cont->vm_stack_clen);
+    MEMCPY(cont->vm_stack, th->stack, VALUE, cont->vm_stack_slen);
+    MEMCPY(cont->vm_stack + cont->vm_stack_slen, (VALUE*)th->cfp,
VALUE, cont->vm_stack_clen);
+#elif
     cont->vm_stack = ALLOC_N(VALUE, th->stack_size);
     MEMCPY(cont->vm_stack, th->stack, VALUE, th->stack_size);
+#endif
     sth->stack = 0;

     cont_save_machine_stack(th, cont);
@@ -288,7 +307,13 @@
 	    th->stack_size = fcont->saved_thread.stack_size;
 	    th->stack = fcont->saved_thread.stack;
 	}
+#ifdef CAPTURE_JUST_VALID_VM_STACK
+	MEMCPY(th->stack, cont->vm_stack, VALUE, cont->vm_stack_slen);
+	MEMCPY(th->stack + sth->stack_size - cont->vm_stack_clen,
+	       cont->vm_stack + cont->vm_stack_slen, VALUE, cont->vm_stack_clen);
+#elif
 	MEMCPY(th->stack, cont->vm_stack, VALUE, sth->stack_size);
+#endif
     }
     else {
 	/* fiber */

-- 
Yusuke ENDOH <mame@tsg.ne.jp>