--- gforth/engine/threaded.h	2002/01/14 08:40:24	1.13
+++ gforth/engine/threaded.h	2003/01/07 22:38:36	1.21
@@ -91,12 +91,11 @@
 
 */
 
-/* CFA_NEXT: if NEXT uses cfa, you have to #define CFA_NEXT, to get
- * cfa declared in engine.
- */
-
 #ifdef DOUBLY_INDIRECT
-#  define CFA_NEXT
+# ifndef DEBUG_DITC
+#  define DEBUG_DITC 0
+# endif
+/* define to 1 if you want to check consistency */
 #  define NEXT_P0	({cfa=*ip;})
 #  define IP		(ip)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -104,24 +103,57 @@
 #  define INC_IP(const_inc)	({cfa=IP[const_inc]; ip+=(const_inc);})
 #  define DEF_CA	Label ca;
 #  define NEXT_P1	({\
-  if (cfa<=vm_prims+DOESJUMP || cfa>=vm_prims+sizeof(routines)) \
+  if (DEBUG_DITC && (cfa<=vm_prims+DOESJUMP || cfa>=vm_prims+npriminfos)) \
     fprintf(stderr,"NEXT encountered prim %p at ip=%p\n", cfa, ip); \
-  ip++; ca=**cfa;})
-#  define NEXT_P2	({goto *ca;})
+  ip++;})
+#  define NEXT_P2	({ca=**cfa; goto *ca;})
 #  define EXEC(XT)	({DEF_CA cfa=(XT);\
-  if (cfa>vm_prims+DOESJUMP && cfa<vm_prims+sizeof(routines)) \
-    fprintf(stderr,"EXEC encountered xt %p at ip=%p\n", cfa, ip); \
+  if (DEBUG_DITC && (cfa>vm_prims+DOESJUMP && cfa<vm_prims+npriminfos)) \
+    fprintf(stderr,"EXEC encountered xt %p at ip=%p, vm_prims=%p, xts=%p\n", cfa, ip, vm_prims, xts); \
  ca=**cfa; goto *ca;})
 
-#else /* !defined(DOUBLY_INDIRECT) */
+#elif defined(NO_IP)
+
+#define NEXT_P0
+#define SET_IP(target)	assert(0)
+#define INC_IP(n)	((void)0)
+#define DEF_CA
+#define NEXT_P1
+#define NEXT_P2		({goto *next_code;})
+/* set next_code to the return address before performing EXEC */
+#define EXEC(XT)	({cfa=(XT); goto **cfa;})
+
+#else  /* !defined(DOUBLY_INDIRECT) && !defined(NO_IP) */
 
 #if defined(DIRECT_THREADED)
 
-/* note that the "cfa dead" versions only work if GETCFA exists and works */
+/* This lets the compiler know that cfa is dead before; we place it at
+   "goto *"s that perform direct threaded dispatch (i.e., not EXECUTE
+   etc.), and thus do not reach doers, which would use cfa; the only
+   way to a doer is through EXECUTE etc., which set the cfa
+   themselves.
+
+   Some of these direct threaded schemes use "cfa" to hold the code
+   address in normal direct threaded code.  Of course we cannot use
+   KILLS there.
+
+   KILLS works by having an empty asm instruction, and claiming to the
+   compiler that it writes to cfa.
+
+   KILLS is optional.  You can write
+
+#define KILLS
+
+   and lose just a little performance.
+*/
+#define KILLS asm("":"=X"(cfa));
+
+#ifndef THREADING_SCHEME
+#define THREADING_SCHEME 6
+#endif
 
 #if THREADING_SCHEME==1
 #warning direct threading scheme 1: autoinc, long latency, cfa live
-#  define CFA_NEXT
 #  define NEXT_P0	({cfa=*ip++;})
 #  define IP		(ip-1)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -130,14 +162,11 @@
 #  define DEF_CA
 #  define NEXT_P1
 #  define NEXT_P2	({goto *cfa;})
-#  define EXEC(XT)	({cfa=(XT); goto *cfa;})
+#  define EXEC(XT)	({cfa=(XT); goto **cfa;})
 #endif
 
 #if THREADING_SCHEME==2
 #warning direct threading scheme 2: autoinc, long latency, cfa dead
-#ifndef GETCFA
-#error GETCFA must be defined for cfa dead threading
-#endif
 #  define NEXT_P0	(ip++)
 #  define IP		(ip-1)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -145,14 +174,13 @@
 #  define INC_IP(const_inc)	({ ip+=(const_inc);})
 #  define DEF_CA
 #  define NEXT_P1
-#  define NEXT_P2	({goto **(ip-1);})
-#  define EXEC(XT)	({goto *(XT);})
+#  define NEXT_P2	({KILLS goto **(ip-1);})
+#  define EXEC(XT)	({cfa=(XT); goto **cfa;})
 #endif
 
 
 #if THREADING_SCHEME==3
 #warning direct threading scheme 3: autoinc, low latency, cfa live
-#  define CFA_NEXT
 #  define NEXT_P0
 #  define IP		(ip)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -161,14 +189,11 @@
 #  define DEF_CA
 #  define NEXT_P1	({cfa=*ip++;})
 #  define NEXT_P2	({goto *cfa;})
-#  define EXEC(XT)	({cfa=(XT); goto *cfa;})
+#  define EXEC(XT)	({cfa=(XT); goto **cfa;})
 #endif
 
 #if THREADING_SCHEME==4
 #warning direct threading scheme 4: autoinc, low latency, cfa dead
-#ifndef GETCFA
-#error GETCFA must be defined for cfa dead threading
-#endif
 #  define NEXT_P0
 #  define IP		(ip)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -176,13 +201,12 @@
 #  define INC_IP(const_inc)	({ ip+=(const_inc);})
 #  define DEF_CA
 #  define NEXT_P1
-#  define NEXT_P2	({goto **(ip++);})
-#  define EXEC(XT)	({goto *(XT);})
+#  define NEXT_P2	({KILLS goto **(ip++);})
+#  define EXEC(XT)	({cfa=(XT); goto **cfa;})
 #endif
 
 #if THREADING_SCHEME==5
 #warning direct threading scheme 5: long latency, cfa live
-#  define CFA_NEXT
 #  define NEXT_P0	({cfa=*ip;})
 #  define IP		(ip)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -191,14 +215,11 @@
 #  define DEF_CA
 #  define NEXT_P1	(ip++)
 #  define NEXT_P2	({goto *cfa;})
-#  define EXEC(XT)	({cfa=(XT); goto *cfa;})
+#  define EXEC(XT)	({cfa=(XT); goto **cfa;})
 #endif
 
 #if THREADING_SCHEME==6
 #warning direct threading scheme 6: long latency, cfa dead
-#ifndef GETCFA
-#error GETCFA must be defined for cfa dead threading
-#endif
 #  define NEXT_P0
 #  define IP		(ip)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -206,14 +227,13 @@
 #  define INC_IP(const_inc)	({ip+=(const_inc);})
 #  define DEF_CA
 #  define NEXT_P1	(ip++)
-#  define NEXT_P2	({goto **(ip-1);})
-#  define EXEC(XT)	({goto *(XT);})
+#  define NEXT_P2	({KILLS goto **(ip-1);})
+#  define EXEC(XT)	({cfa=(XT); goto **cfa;})
 #endif
 
 
 #if THREADING_SCHEME==7
 #warning direct threading scheme 7: low latency, cfa live
-#  define CFA_NEXT
 #  define NEXT_P0
 #  define IP		(ip)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -222,14 +242,11 @@
 #  define DEF_CA
 #  define NEXT_P1	({cfa=*ip++;})
 #  define NEXT_P2	({goto *cfa;})
-#  define EXEC(XT)	({cfa=(XT); goto *cfa;})
+#  define EXEC(XT)	({cfa=(XT); goto **cfa;})
 #endif
 
 #if THREADING_SCHEME==8
 #warning direct threading scheme 8: cfa dead, i386 hack
-#ifndef GETCFA
-#error GETCFA must be defined for cfa dead threading
-#endif
 #  define NEXT_P0
 #  define IP		(ip)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -237,8 +254,8 @@
 #  define INC_IP(const_inc)	({ ip+=(const_inc);})
 #  define DEF_CA
 #  define NEXT_P1	(ip++)
-#  define NEXT_P2	({goto **(ip-1);})
-#  define EXEC(XT)	({goto *(XT);})
+#  define NEXT_P2	({KILLS goto **(ip-1);})
+#  define EXEC(XT)	({cfa=(XT); goto **cfa;})
 #endif
 
 #if THREADING_SCHEME==9
@@ -246,7 +263,6 @@
 /* Power uses a prepare-to-branch instruction, and the latency between
    this inst and the branch is 5 cycles on a PPC604; so we utilize this
    to do some prefetching in between */
-#  define CFA_NEXT
 #  define NEXT_P0
 #  define IP		ip
 #  define SET_IP(p)	({ip=(p); next_cfa=*ip; NEXT_P0;})
@@ -255,13 +271,12 @@
 #  define DEF_CA	
 #  define NEXT_P1	({cfa=next_cfa; ip++; next_cfa=*ip;})
 #  define NEXT_P2	({goto *cfa;})
-#  define EXEC(XT)	({cfa=(XT); goto *cfa;})
+#  define EXEC(XT)	({cfa=(XT); goto **cfa;})
 #  define MORE_VARS	Xt next_cfa;
 #endif
 
 #if THREADING_SCHEME==10
 #warning direct threading scheme 10: plain (no attempt at scheduling)
-#  define CFA_NEXT
 #  define NEXT_P0
 #  define IP		(ip)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -270,16 +285,19 @@
 #  define DEF_CA
 #  define NEXT_P1
 #  define NEXT_P2	({cfa=*ip++; goto *cfa;})
-#  define EXEC(XT)	({cfa=(XT); goto *cfa;})
+#  define EXEC(XT)	({cfa=(XT); goto **cfa;})
 #endif
 
 /* direct threaded */
 #else
 /* indirect THREADED  */
 
+#ifndef THREADING_SCHEME
+#define THREADING_SCHEME 6
+#endif
+
 #if THREADING_SCHEME==1
 #warning indirect threading scheme 1: autoinc, long latency, cisc
-#  define CFA_NEXT
 #  define NEXT_P0	({cfa=*ip++;})
 #  define IP		(ip-1)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -293,7 +311,6 @@
 
 #if THREADING_SCHEME==2
 #warning indirect threading scheme 2: autoinc, long latency
-#  define CFA_NEXT
 #  define NEXT_P0	({cfa=*ip++;})
 #  define IP		(ip-1)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -308,7 +325,6 @@
 
 #if THREADING_SCHEME==3
 #warning indirect threading scheme 3: autoinc, low latency, cisc
-#  define CFA_NEXT
 #  define NEXT_P0
 #  define IP		(ip)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -322,7 +338,6 @@
 
 #if THREADING_SCHEME==4
 #warning indirect threading scheme 4: autoinc, low latency
-#  define CFA_NEXT
 #  define NEXT_P0	({cfa=*ip++;})
 #  define IP		(ip-1)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -337,7 +352,6 @@
 
 #if THREADING_SCHEME==5
 #warning indirect threading scheme 5: long latency, cisc
-#  define CFA_NEXT
 #  define NEXT_P0	({cfa=*ip;})
 #  define IP		(ip)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -351,7 +365,6 @@
 
 #if THREADING_SCHEME==6
 #warning indirect threading scheme 6: long latency
-#  define CFA_NEXT
 #  define NEXT_P0	({cfa=*ip;})
 #  define IP		(ip)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -365,7 +378,6 @@
 
 #if THREADING_SCHEME==7
 #warning indirect threading scheme 7: low latency
-#  define CFA_NEXT
 #  define NEXT_P0	({cfa=*ip;})
 #  define IP		(ip)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -379,7 +391,6 @@
 
 #if THREADING_SCHEME==8
 #warning indirect threading scheme 8: low latency,cisc
-#  define CFA_NEXT
 #  define NEXT_P0
 #  define IP		(ip)
 #  define SET_IP(p)	({ip=(p); NEXT_P0;})
@@ -394,7 +405,7 @@
 /* indirect threaded */
 #endif
 
-#endif /* !defined(DOUBLY_INDIRECT) */
+#endif /* !defined(DOUBLY_INDIRECT) && !defined(NO_IP) */
 
 #define NEXT ({DEF_CA NEXT_P1; NEXT_P2;})
 #define IPTOS NEXT_INST