File:  [gforth] / gforth / engine / threaded.h
Revision 1.40: download - view: text, annotated - select for diffs
Sat Nov 1 22:19:31 2008 UTC (14 years, 3 months ago) by anton
Branches: MAIN
CVS tags: v0-7-0, HEAD
updated copyright years

/* This file defines a number of threading schemes.

  Copyright (C) 1995, 1996,1997,1999,2003,2004,2005,2007,2008 Free Software Foundation, Inc.

  This file is part of Gforth.

  Gforth is free software; you can redistribute it and/or
  modify it under the terms of the GNU General Public License
  as published by the Free Software Foundation, either version 3
  of the License, or (at your option) any later version.

  This program is distributed in the hope that it will be useful,
  but WITHOUT ANY WARRANTY; without even the implied warranty of
  GNU General Public License for more details.

  You should have received a copy of the GNU General Public License
  along with this program; if not, see

  This files defines macros for threading. Many sets of macros are
  defined. Functionally they have only one difference: Some implement
  direct threading, some indirect threading. The other differences are
  just variations to help GCC generate faster code for various

  (Well, to tell the truth, there actually is another functional
  difference in some pathological cases: e.g., a '!' stores into the
  cell where the next executed word comes from; or, the next word
  executed comes from the top-of-stack. These differences are one of
  the reasons why GCC cannot produce the right variation by itself. We
  chose disallowing such practices and using the added implementation
  freedom to achieve a significant speedup, because these practices
  are not common in Forth (I have never heard of or seen anyone using
  them), and it is easy to circumvent problems: A control flow change
  will flush any prefetched words; you may want to do a "0
  drop" before that to write back the top-of-stack cache.)

  These macro sets are used in the following ways: After translation
  to C a typical primitive looks like

  other declarations
  main part of the primitive
  store results to stack

  DEF_CA and all the NEXT_P* together must implement NEXT; In the main
  part the instruction pointer can be read with IP, changed with
  INC_IP(const_inc), and the cell right behind the presently executing
  word (i.e. the value of *IP) is accessed with NEXT_INST.

  If a primitive does not fall through the main part, it has to do the
  rest by itself. If it changes ip, it has to redo NEXT_P0 (perhaps we
  should define a macro SET_IP).

  Some primitives (execute, dodefer) do not end with NEXT, but with
  EXEC(.). If NEXT_P0 has been called earlier, it has to perform
  "ip=IP;" to ensure that ip has the right value (NEXT_P0 may change

  Finally, there is NEXT1_P1 and NEXT1_P2, which are parts of EXEC
  (EXEC(XT) could be defined as "cfa=XT; NEXT1_P1; NEXT1_P2;" (is this
  true?)) and are used for making docol faster.

  We can define the ways in which these macros are used with a regular

  For a primitive

  DEF_CA NEXT_P0 ( IP | INC_IP | NEXT_INST | ip=...; NEXT_P0 ) * ( NEXT_P1 NEXT_P2 | EXEC(...) )

  For a run-time routine, e.g., docol:
  PFA1(cfa) ( NEXT_P0 NEXT | cfa=...; NEXT1_P1; NEXT1_P2 | EXEC(...) )

  This comment does not yet describe all the dependences that the
  macros have to satisfy.

  To organize the former ifdef chaos, each path is separated
  This gives a quite impressive number of paths, but you clearly
  find things that go together.

  It should be possible to organize the whole thing in a way that
  contains less redundancy and allows a simpler description.


#if !defined(GCC_PR15242_WORKAROUND)
#if __GNUC__ == 3
/* various gcc-3.x version have problems (including PR15242) that are
   solved with this workaround */
#define GCC_PR15242_WORKAROUND 1
/* other gcc versions are better off without the workaround for
   primitives that are not relocatable */
#define GCC_PR15242_WORKAROUND 0

#define DO_GOTO goto before_goto
#define DO_GOTO goto *real_ca

#ifndef GOTO_ALIGN
#define GOTO_ALIGN

#define GOTO(target) do {(real_ca=(target));} while(0)
#define NEXT_P2 do {NEXT_P1_5; DO_GOTO;} while(0)
#define EXEC(XT) do { real_ca=EXEC1(XT); DO_GOTO;} while (0)
#define VM_JUMP(target) do {GOTO(target);} while (0)
#define NEXT do {DEF_CA NEXT_P1; NEXT_P2;} while(0)
before_goto: goto *real_ca; after_goto:
#define FIRST_NEXT do {DEF_CA NEXT_P1; FIRST_NEXT_P2;} while(0)

# ifndef DEBUG_DITC
#  define DEBUG_DITC 0
# endif
/* define to 1 if you want to check consistency */
#  define NEXT_P0	do {cfa1=cfa; cfa=*ip;} while(0)
#  define CFA		cfa1
#  define MORE_VARS     Xt cfa1;
#  define IP		(ip)
#  define SET_IP(p)	do {ip=(p); cfa=*ip;} while(0)
#  define NEXT_INST	(cfa)
#  define INC_IP(const_inc)	do {cfa=IP[const_inc]; ip+=(const_inc);} while(0)
#  define DEF_CA	Label MAYBE_UNUSED ca;
#  define NEXT_P1	do {\
  if (DEBUG_DITC && (cfa<=vm_prims+DOESJUMP || cfa>=vm_prims+npriminfos)) \
    fprintf(stderr,"NEXT encountered prim %p at ip=%p\n", cfa, ip); \
  ip++;} while(0)
#  define NEXT_P1_5	do {ca=**cfa; GOTO(ca);} while(0)
#  define EXEC1(XT)	({DEF_CA cfa=(XT);\
  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; ca;})

#elif defined(NO_IP)

#define NEXT_P0
#  define CFA		cfa
#define SET_IP(target)	assert(0)
#define INC_IP(n)	((void)0)
#define DEF_CA
#define NEXT_P1
#define NEXT_P1_5		do {goto *next_code;} while(0)
/* set next_code to the return address before performing EXEC */
/* original: */
/* #define EXEC1(XT)	do {cfa=(XT); goto **cfa;} while(0) */
/* fake, to make syntax check work */
#define EXEC1(XT)	({cfa=(XT); *cfa;})

#else  /* !defined(DOUBLY_INDIRECT) && !defined(NO_IP) */

#if defined(DIRECT_THREADED)

/* 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

   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));

/* #warning direct threading scheme 8: cfa dead, i386 hack */
#  define NEXT_P0
#  define CFA		cfa
#  define IP		(ip)
#  define SET_IP(p)	do {ip=(p); NEXT_P0;} while(0)
#  define NEXT_INST	(*IP)
#  define INC_IP(const_inc)	do { ip+=(const_inc);} while(0)
#  define DEF_CA
#  define NEXT_P1	(ip++)
#  define NEXT_P1_5	do {KILLS GOTO(*(ip-1));} while(0)
#  define EXEC1(XT)	({cfa=(XT); *cfa;})

/* direct threaded */
/* indirect THREADED  */

/* #warning indirect threading scheme 8: low latency,cisc */
#  define NEXT_P0
#  define CFA		cfa
#  define IP		(ip)
#  define SET_IP(p)	do {ip=(p); NEXT_P0;} while(0)
#  define NEXT_INST	(*ip)
#  define INC_IP(const_inc)	do {ip+=(const_inc);} while(0)
#  define DEF_CA
#  define NEXT_P1
#  define NEXT_P1_5	do {cfa=*ip++; GOTO(*cfa);} while(0)
#  define EXEC1(XT)	({cfa=(XT); *cfa;})

/* indirect threaded */

#endif /* !defined(DOUBLY_INDIRECT) && !defined(NO_IP) */

FreeBSD-CVSweb <>