--- a/Changes	Thu Jul 31 20:57:39 2003 +0900
+++ b/Changes	Tue Aug 05 13:00:16 2003 +0900
@@ -3220,3 +3220,79 @@
 
 それぞれの利点と欠点を考察すればCWは通るんじゃない?
 (本気?)
+
+Mon Aug  4 12:48:40 JST 2003
+
+converted call
+480
+8.070u 0.010s 0:08.22 98.2%     0+0k 0+2io 0pf+0w
+call
+480
+6.590u 0.000s 0:06.67 98.8%     0+0k 0+0io 0pf+0w
+
+だいたい10%ぐらい遅いなぁ。(なんで?) げ、-O6 だと十倍違う...
+
+げ、powerpc の subroutine 内のstaticが通らんじゃん。
+
+サブルーチンを三段に直しました。
+
+% time ./a.out 0
+720
+9.590u 0.010s 0:09.64 99.5%     0+0k 0+0io 0pf+0w
+% time ./a.out 1
+719
+12.610u 0.020s 0:12.71 99.3%    0+0k 0+0io 0pf+0w
+% time ./a.out 2
+720
+6.310u 0.020s 0:06.41 98.7%     0+0k 0+0io 0pf+0w
+% vi test/conv1.c
+% gcc -O6 test/conv1.c
+% ./a.out 0
+720
+% time ./a.out 0
+720
+1.990u 0.020s 0:02.03 99.0%     0+0k 0+2io 0pf+0w
+% gcc  test/conv1.c
+% time ./a.out 0
+720
+7.530u 0.000s 0:07.64 98.5%     0+0k 0+0io 0pf+0w
+% gcc -O2 test/conv1.c
+% time ./a.out 0
+720
+3.520u 0.010s 0:03.55 99.4%     0+0k 0+0io 0pf+0w
+% 
+
+ま、こんなものかな。-O2 に負けるなよって気もするが。
+
+琉大は40Mbps
+
+Intel PC (on PowerPC) の場合
+
+[root@localhost ~/device]# time ./a.out 2                                       
+470                                                                             
+0.660u 0.010s 0:02.37 28.2%     0+0k 0+0io 92pf+0w                              
+[root@localhost ~/device]# time ./a.out 3                                       
+720                                                                             
+0.920u 0.020s 0:01.74 54.0%     0+0k 0+0io 92pf+0w                              
+[root@localhost ~/device]# gcc test/conv1.c                                     
+[root@localhost ~/device]# ./a.out 0                                            
+                                                                                
+[root@localhost ~/device]# time ./a.out 0                                       
+720                                                                             
+1.310u 0.030s 0:01.39 96.4%     0+0k 0+0io 92pf+0w                              
+[root@localhost ~/device]# gcc -O test/conv1.c                                  
+[root@localhost ~/device]# time ./a.out 0                                       
+720                                                                             
+1.130u 0.030s 0:01.16 100.0%    0+0k 0+0io 92pf+0w                              
+[root@localhost ~/device]# gcc -O4 test/conv1.c                                 
+[root@localhost ~/device]# time ./a.out 0                                       
+720                                                                             
+0.870u 0.000s 0:00.87 100.0%    0+0k 0+0io 92pf+0w                              
+[root@localhost ~/device]# gcc -O6 test/conv1.c                                 
+[root@localhost ~/device]# time ./a.out 0                                       
+720                                                                             
+0.850u 0.020s 0:00.88 98.8%     0+0k 0+0io 92pf+0w                              
+[root@localhost ~/device]#                                                      
+
+PowerPC の場合は、PIC symbol がやっぱり遅いね。
+こいつをなんとかするだけでだいぶ違うかも知れない。

--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/test/conv1.c	Tue Aug 05 13:00:16 2003 +0900
@@ -0,0 +1,219 @@
+#include "stdio.h"
+
+static int loop;
+
+#ifdef __micro_c__
+#define CC_ONLY 0
+#else
+#define CC_ONLY 1
+#endif
+
+/* classical function call case (0) */
+
+f0(int i) {
+    int k,j;
+    k = 3+i;
+    j = g0(i+3);
+    return k+4+j;
+}
+
+g0(int i) {
+    return h0(i+4)+i;
+}
+
+h0(int i) {
+    return i+4;
+}
+
+#if !CC_ONLY
+
+/* straight conversion case (1) */
+
+typedef char *stack;
+
+struct cont_interface { // General Return Continuation
+    code (*ret)();
+};
+
+code f(int i,stack sp) {
+    int k,j;
+    k = 3+i;
+    goto f_g0(i,k,sp);
+}
+
+struct f_g0_interface {  // Specialized Return Continuation
+    code (*ret)();
+    int i_,k_,j_;
+};
+
+code f_g1(int j,stack sp);
+
+code f_g0(int i,int k,stack sp) { // Caller
+    struct f_g0_interface *c = 
+	(struct f_g0_interface *)(sp -= sizeof(struct f_g0_interface));
+
+    c->ret = f_g1;
+    c->k_ = k;
+    c->i_ = i;
+
+    goto g(i+3,sp);
+}
+
+code f_g1(int j,stack sp) {  // Continuation 
+    struct f_g0_interface *c = sp;
+    int k = c->k_;
+    sp += sizeof(struct f_g0_interface);
+    goto (( (struct cont_interface *)sp)->ret)(k+4+j,sp);
+}
+
+code g(int i,stack sp) { // Caller
+    struct f_g0_interface *c = 
+	(struct f_g0_interface *)(sp -= sizeof(struct f_g0_interface));
+
+    c->ret = g_h1;
+    c->i_ = i;
+
+    goto h(i+3,sp);
+}
+
+code g_h1(int j,stack sp) {  // Continuation 
+    struct f_g0_interface *c = sp;
+    int i = c->i_;
+    sp += sizeof(struct f_g0_interface);
+    goto (( (struct cont_interface *)sp)->ret)(j+i,sp);
+}
+
+code h(int i,stack sp) {
+    goto (( (struct cont_interface *)sp)->ret)(i+4,sp);
+}
+
+struct main_continuation { // General Return Continuation
+    code (*ret)();
+    code (*main_ret)();
+    void *env;
+};
+
+code main_return(int i,stack sp) {
+    if (loop-->0)
+	goto f(233,sp);
+    printf("%d\n",i);
+    goto (( (struct main_continuation *)sp)->main_ret)(0),
+           ((struct main_continuation *)sp)->env;
+}
+
+/* little optimzation without stack continuation (2) */
+
+code f2(int i,char *sp) {
+    int k,j;
+    k = 3+i;
+    goto g2(i,k,i+3,sp);
+}
+
+code g2(int i,int k,int j,char *sp) {
+    j = j+4;
+    goto h2(i,k+4+j,sp);
+}
+
+code h2_1(int i,int k,int j,char *sp) {
+    goto main_return2(i+j,sp);
+}
+
+code h2(int i,int k,char *sp) {
+    goto h2_1(i,k,i+4,sp);
+}
+
+code main_return2(int i,stack sp) {
+    if (loop-->0)
+	goto f2(233,sp);
+    printf("%d\n",i);
+    goto (( (struct main_continuation *)sp)->main_ret)(0),
+           ((struct main_continuation *)sp)->env;
+}
+
+/* little optimizaed case (3) */
+
+code f2_1(int i,char *sp) {
+    int k,j;
+    k = 3+i;
+    goto g2_1(k,i+3,sp);
+}
+
+code g2_1(int k,int i,char *sp) {
+    goto h2_11(k,i+4,sp);
+}
+
+code f2_0_1(int k,int j,char *sp);
+code h2_1_1(int i,int k,int j,char *sp) {
+    goto f2_0_1(k,i+j,sp);
+}
+
+code h2_11(int i,int k,char *sp) {
+    goto h2_1_1(i,k,i+4,sp);
+}
+
+code f2_0_1(int k,int j,char *sp) {
+    goto (( (struct cont_interface *)sp)->ret)(k+4+j,sp);
+}
+
+code main_return2_1(int i,stack sp) {
+    if (loop-->0)
+        goto f2_1(233,sp);
+    printf("%d\n",i);
+    goto (( (struct main_continuation *)sp)->main_ret)(0),
+           ((struct main_continuation *)sp)->env;
+}
+
+#define STACK_SIZE 2048
+stack main_stack[STACK_SIZE];
+#define stack_last (&main_stack[STACK_SIZE])
+
+#endif
+
+#define LOOP_COUNT 10000000
+
+main(int ac,char *av[])
+{
+#if !CC_ONLY
+    struct main_continuation *cont;
+    stack sp = stack_last;
+#endif
+    int sw;
+    int j;
+    if (ac!=2) exit(0);
+    sw = atoi(av[1]);
+
+    if (sw==0) {
+	for(loop=0;loop<LOOP_COUNT;loop++) {
+	   j = f0(233);
+	}
+	printf("%d\n",j);
+#if !CC_ONLY
+    } else if (sw==1) {
+	loop = LOOP_COUNT;
+	sp -= sizeof(*cont);
+	cont = (struct main_continuation *)sp;
+	cont->ret = main_return;
+	cont->main_ret = return;
+	cont->env = environment;
+	goto f(233,sp);
+    } else if (sw==2) {
+	loop = LOOP_COUNT;
+	sp -= sizeof(*cont);
+	cont = (struct main_continuation *)sp;
+	cont->ret = main_return2;
+	cont->main_ret = return;
+	cont->env = environment;
+	goto f2(233,sp);
+    } else if (sw==3) {
+	loop = LOOP_COUNT;
+	sp -= sizeof(*cont);
+	cont = (struct main_continuation *)sp;
+	cont->ret = main_return2_1;
+	cont->main_ret = return;
+	cont->env = environment;
+	goto f2_1(233,sp);
+#endif
+    }
+}
+
+/* end */