[gforth] / gforth / engine / main.c  

gforth: gforth/engine/main.c

Diff for /gforth/engine/main.c between version 1.48 and 1.145

version 1.48, Mon Dec 24 20:39:30 2001 UTC version 1.145, Sat Jan 22 16:39:59 2005 UTC
Line 1 
Line 1 
 /* command line interpretation, image loading etc. for Gforth  /* command line interpretation, image loading etc. for Gforth
   
   
   Copyright (C) 1995,1996,1997,1998,2000 Free Software Foundation, Inc.    Copyright (C) 1995,1996,1997,1998,2000,2003,2004 Free Software Foundation, Inc.
   
   This file is part of Gforth.    This file is part of Gforth.
   
Line 21 
Line 21 
 */  */
   
 #include "config.h"  #include "config.h"
   #include "forth.h"
 #include <errno.h>  #include <errno.h>
 #include <ctype.h>  #include <ctype.h>
 #include <stdio.h>  #include <stdio.h>
Line 34 
Line 35 
 #include <fcntl.h>  #include <fcntl.h>
 #include <assert.h>  #include <assert.h>
 #include <stdlib.h>  #include <stdlib.h>
   #include <signal.h>
 #ifndef STANDALONE  #ifndef STANDALONE
 #if HAVE_SYS_MMAN_H  #if HAVE_SYS_MMAN_H
 #include <sys/mman.h>  #include <sys/mman.h>
 #endif  #endif
 #endif  #endif
 #include "forth.h"  
 #include "io.h"  #include "io.h"
 #include "getopt.h"  #include "getopt.h"
 #ifdef STANDALONE  #ifdef STANDALONE
 #include <systypes.h>  #include <systypes.h>
 #endif  #endif
   
   typedef enum prim_num {
   /* definitions of N_execute etc. */
   #include PRIM_NUM_I
     N_START_SUPER
   } PrimNum;
   
   /* global variables for engine.c
      We put them here because engine.c is compiled several times in
      different ways for the same engine. */
   Cell *SP;
   Float *FP;
   Address UP=NULL;
   
   #ifdef HAS_FFCALL
   Cell *RP;
   Address LP;
   
   #include <callback.h>
   
   va_alist clist;
   
   void engine_callback(Xt* fcall, void * alist)
   {
     /* save global valiables */
     Cell *rp = RP;
     Cell *sp = SP;
     Float *fp = FP;
     Address lp = LP;
   
     clist = (va_alist)alist;
   
     engine(fcall, sp, rp, fp, lp);
   
     /* restore global variables */
     RP = rp;
     SP = sp;
     FP = fp;
     LP = lp;
   }
   #endif
   
   #ifdef GFORTH_DEBUGGING
   /* define some VM registers as global variables, so they survive exceptions;
      global register variables are not up to the task (according to the
      GNU C manual) */
   Xt *saved_ip;
   Cell *rp;
   #endif
   
   #ifdef NO_IP
   Label next_code;
   #endif
   
   #ifdef HAS_FILE
   char* fileattr[6]={"rb","rb","r+b","r+b","wb","wb"};
   char* pfileattr[6]={"r","r","r+","r+","w","w"};
   
   #ifndef O_BINARY
   #define O_BINARY 0
   #endif
   #ifndef O_TEXT
   #define O_TEXT 0
   #endif
   
   int ufileattr[6]= {
     O_RDONLY|O_BINARY, O_RDONLY|O_BINARY,
     O_RDWR  |O_BINARY, O_RDWR  |O_BINARY,
     O_WRONLY|O_BINARY, O_WRONLY|O_BINARY };
   #endif
   /* end global vars for engine.c */
   
 #define PRIM_VERSION 1  #define PRIM_VERSION 1
 /* increment this whenever the primitives change in an incompatible way */  /* increment this whenever the primitives change in an incompatible way */
   
Line 57 
Line 129 
 jmp_buf throw_jmp_buf;  jmp_buf throw_jmp_buf;
 #endif  #endif
   
 #if defined(DIRECT_THREADED)  #if defined(DOUBLY_INDIRECT)
 #  define CA(n) (symbols[(n)])  #  define CFA(n)        ({Cell _n = (n); ((Cell)(((_n & 0x4000) ? symbols : xts)+(_n&~0x4000UL)));})
 #else  #else
 #  define CA(n) ((Cell)(symbols+(n)))  #  define CFA(n)        ((Cell)(symbols+((n)&~0x4000UL)))
 #endif  #endif
   
 #define maxaligned(n)   (typeof(n))((((Cell)n)+sizeof(Float)-1)&-sizeof(Float))  #define maxaligned(n)   (typeof(n))((((Cell)n)+sizeof(Float)-1)&-sizeof(Float))
Line 81 
Line 153 
 int optind = 1;  int optind = 1;
 #endif  #endif
   
   #define CODE_BLOCK_SIZE (4096*1024) /* !! overflow handling for -native */
 Address code_area=0;  Address code_area=0;
 Address code_here=0; /* does for code-area what HERE does for the dictionary */  Cell code_area_size = CODE_BLOCK_SIZE;
   Address code_here=NULL+CODE_BLOCK_SIZE; /* does for code-area what HERE
                                              does for the dictionary */
   Address start_flush=NULL; /* start of unflushed code */
   Cell last_jump=0; /* if the last prim was compiled without jump, this
                        is it's number, otherwise this contains 0 */
   
   static int no_super=0;   /* true if compile_prim should not fuse prims */
   static int no_dynamic=NO_DYNAMIC_DEFAULT; /* if true, no code is generated
                                                dynamically */
   static int print_metrics=0; /* if true, print metrics on exit */
   static int static_super_number = 10000000; /* number of ss used if available */
   #define MAX_STATE 4 /* maximum number of states */
   static int maxstates = MAX_STATE; /* number of states for stack caching */
   static int ss_greedy = 0; /* if true: use greedy, not optimal ss selection */
   static int diag = 0; /* if true: print diagnostic informations */
   static int relocs = 0;
   static int nonrelocs = 0;
   
 #ifdef HAS_DEBUG  #ifdef HAS_DEBUG
 static int debug=0;  int debug=0;
   # define debugp(x...) if (debug) fprintf(x);
 #else  #else
 # define debug 0  
 # define perror(x...)  # define perror(x...)
 # define fprintf(x...)  # define fprintf(x...)
   # define debugp(x...)
 #endif  #endif
   
 ImageHeader *gforth_header;  ImageHeader *gforth_header;
 Label *vm_prims;  Label *vm_prims;
   #ifdef DOUBLY_INDIRECT
   Label *xts; /* same content as vm_prims, but should only be used for xts */
   #endif
   
   #ifndef NO_DYNAMIC
   #define MAX_IMMARGS 2
   
   typedef struct {
     Label start; /* NULL if not relocatable */
     Cell length; /* only includes the jump iff superend is true*/
     Cell restlength; /* length of the rest (i.e., the jump or (on superend) 0) */
     char superend; /* true if primitive ends superinstruction, i.e.,
                        unconditional branch, execute, etc. */
     Cell nimmargs;
     struct immarg {
       Cell offset; /* offset of immarg within prim */
       char rel;    /* true if immarg is relative */
     } immargs[MAX_IMMARGS];
   } PrimInfo;
   
   PrimInfo *priminfos;
   PrimInfo **decomp_prims;
   
   const char const* const prim_names[]={
   #include PRIM_NAMES_I
   };
   
   static int is_relocatable(int p)
   {
     return !no_dynamic && priminfos[p].start != NULL;
   }
   #else /* defined(NO_DYNAMIC) */
   static int is_relocatable(int p)
   {
     return 0;
   }
   #endif /* defined(NO_DYNAMIC) */
   
 #ifdef MEMCMP_AS_SUBROUTINE  #ifdef MEMCMP_AS_SUBROUTINE
 int gforth_memcmp(const char * s1, const char * s2, size_t n)  int gforth_memcmp(const char * s1, const char * s2, size_t n)
Line 102 
Line 230 
 }  }
 #endif  #endif
   
   static Cell max(Cell a, Cell b)
   {
     return a>b?a:b;
   }
   
   static Cell min(Cell a, Cell b)
   {
     return a<b?a:b;
   }
   
 /* image file format:  /* image file format:
  *  "#! binary-path -i\n" (e.g., "#! /usr/local/bin/gforth-0.4.0 -i\n")   *  "#! binary-path -i\n" (e.g., "#! /usr/local/bin/gforth-0.4.0 -i\n")
  *   padding to a multiple of 8   *   padding to a multiple of 8
  *   magic: "Gforth2x" means format 0.4,   *   magic: "Gforth3x" means format 0.6,
  *              where x is a byte with   *              where x is a byte with
  *              bit 7:   reserved = 0   *              bit 7:   reserved = 0
  *              bit 6:5: address unit size 2^n octets   *              bit 6:5: address unit size 2^n octets
Line 126 
Line 264 
  * If the word =CF(DODOES), it's a DOES> CFA   * If the word =CF(DODOES), it's a DOES> CFA
  * If the word =CF(DOESJUMP), it's a DOES JUMP (2 Cells after DOES>,   * If the word =CF(DOESJUMP), it's a DOES JUMP (2 Cells after DOES>,
  *                                      possibly containing a jump to dodoes)   *                                      possibly containing a jump to dodoes)
  * If the word is <CF(DOESJUMP), it's a primitive   * If the word is <CF(DOESJUMP) and bit 14 is set, it's the xt of a primitive
    * If the word is <CF(DOESJUMP) and bit 14 is clear,
    *                                        it's the threaded code of a primitive
    * bits 13..9 of a primitive token state which group the primitive belongs to,
    * bits 8..0 of a primitive token index into the group
  */   */
   
 void relocate(Cell *image, const char *bitstring,  Cell groups[32] = {
               int size, int base, Label symbols[])    0,
     0
   #undef GROUP
   #undef GROUPADD
   #define GROUPADD(n) +n
   #define GROUP(x, n) , 0
   #include PRIM_GRP_I
   #undef GROUP
   #undef GROUPADD
   #define GROUP(x, n)
   #define GROUPADD(n)
   };
   
   unsigned char *branch_targets(Cell *image, const unsigned char *bitstring,
                                 int size, Cell base)
        /* produce a bitmask marking all the branch targets */
 {  {
   int i=0, j, k, steps=(size/sizeof(Cell))/RELINFOBITS;    int i=0, j, k, steps=(((size-1)/sizeof(Cell))/RELINFOBITS)+1;
     Cell token;
     unsigned char bits;
     unsigned char *result=malloc(steps);
   
     memset(result, 0, steps);
     for(k=0; k<steps; k++) {
       for(j=0, bits=bitstring[k]; j<RELINFOBITS; j++, i++, bits<<=1) {
         if(bits & (1U << (RELINFOBITS-1))) {
           assert(i*sizeof(Cell) < size);
           token=image[i];
           if (token>=base) { /* relocatable address */
             UCell bitnum=(token-base)/sizeof(Cell);
             result[bitnum/RELINFOBITS] |= 1U << ((~bitnum)&(RELINFOBITS-1));
           }
         }
       }
     }
     return result;
   }
   
   void relocate(Cell *image, const unsigned char *bitstring,
                 int size, Cell base, Label symbols[])
   {
     int i=0, j, k, steps=(((size-1)/sizeof(Cell))/RELINFOBITS)+1;
   Cell token;    Cell token;
   char bits;    char bits;
   Cell max_symbols;    Cell max_symbols;
   /*    /*
    * A virtial start address that's the real start address minus     * A virtual start address that's the real start address minus
    * the one in the image     * the one in the image
    */     */
   Cell *start = (Cell * ) (((void *) image) - ((void *) base));    Cell *start = (Cell * ) (((void *) image) - ((void *) base));
     unsigned char *targets = branch_targets(image, bitstring, size, base);
   
     /* group index into table */
     if(groups[31]==0) {
       int groupsum=0;
       for(i=0; i<32; i++) {
         groupsum += groups[i];
         groups[i] = groupsum;
         /* printf("group[%d]=%d\n",i,groupsum); */
       }
       i=0;
     }
   
 /* printf("relocating to %x[%x] start=%x base=%x\n", image, size, start, base); */  /* printf("relocating to %x[%x] start=%x base=%x\n", image, size, start, base); */
   
   for (max_symbols=DOESJUMP+1; symbols[max_symbols]!=0; max_symbols++)    for (max_symbols=0; symbols[max_symbols]!=0; max_symbols++)
     ;      ;
   max_symbols--;    max_symbols--;
   size/=sizeof(Cell);  
   
   for(k=0; k<=steps; k++) {    for(k=0; k<steps; k++) {
     for(j=0, bits=bitstring[k]; j<RELINFOBITS; j++, i++, bits<<=1) {      for(j=0, bits=bitstring[k]; j<RELINFOBITS; j++, i++, bits<<=1) {
       /*      fprintf(stderr,"relocate: image[%d]\n", i);*/        /*      fprintf(stderr,"relocate: image[%d]\n", i);*/
       if((i < size) && (bits & (1U << (RELINFOBITS-1)))) {        if(bits & (1U << (RELINFOBITS-1))) {
           assert(i*sizeof(Cell) < size);
         /* fprintf(stderr,"relocate: image[%d]=%d of %d\n", i, image[i], size/sizeof(Cell)); */          /* fprintf(stderr,"relocate: image[%d]=%d of %d\n", i, image[i], size/sizeof(Cell)); */
         token=image[i];          token=image[i];
         if(token<0)          if(token<0) {
           switch(token)            int group = (-token & 0x3E00) >> 9;
             {            if(group == 0) {
               switch(token|0x4000) {
             case CF_NIL      : image[i]=0; break;              case CF_NIL      : image[i]=0; break;
 #if !defined(DOUBLY_INDIRECT)  #if !defined(DOUBLY_INDIRECT)
             case CF(DOCOL)   :              case CF(DOCOL)   :
Line 167 
Line 360 
             case CF(DOUSER)  :              case CF(DOUSER)  :
             case CF(DODEFER) :              case CF(DODEFER) :
             case CF(DOFIELD) : MAKE_CF(image+i,symbols[CF(token)]); break;              case CF(DOFIELD) : MAKE_CF(image+i,symbols[CF(token)]); break;
             case CF(DOESJUMP): MAKE_DOES_HANDLER(image+i); break;              case CF(DOESJUMP): image[i]=0; break;
 #endif /* !defined(DOUBLY_INDIRECT) */  #endif /* !defined(DOUBLY_INDIRECT) */
             case CF(DODOES)  :              case CF(DODOES)  :
               MAKE_DOES_CF(image+i,(Xt *)(image[i+1]+((Cell)start)));                MAKE_DOES_CF(image+i,(Xt *)(image[i+1]+((Cell)start)));
               break;                break;
             default          :              default          : /* backward compatibility */
 /*            printf("Code field generation image[%x]:=CA(%x)\n",  /*            printf("Code field generation image[%x]:=CFA(%x)\n",
                      i, CF(image[i])); */                       i, CF(image[i])); */
               if (CF(token)<max_symbols)                if (CF((token | 0x4000))<max_symbols) {
                 image[i]=(Cell)CA(CF(token));                  image[i]=(Cell)CFA(CF(token));
               else  #ifdef DIRECT_THREADED
                 fprintf(stderr,"Primitive %d used in this image at $%lx is not implemented by this\n engine (%s); executing this code will crash.\n",CF(token),(long)&image[i],VERSION);                  if ((token & 0x4000) == 0) { /* threade code, no CFA */
                     if (targets[k] & (1U<<(RELINFOBITS-1-j)))
                       compile_prim1(0);
                     compile_prim1(&image[i]);
             }              }
         else {  #endif
           // if base is > 0: 0 is a null reference so don't adjust                } else
                   fprintf(stderr,"Primitive %ld used in this image at $%lx (offset $%x) is not implemented by this\n engine (%s); executing this code will crash.\n",(long)CF(token),(long)&image[i], i, PACKAGE_VERSION);
               }
             } else {
               int tok = -token & 0x1FF;
               if (tok < (groups[group+1]-groups[group])) {
   #if defined(DOUBLY_INDIRECT)
                 image[i]=(Cell)CFA(((groups[group]+tok) | (CF(token) & 0x4000)));
   #else
                 image[i]=(Cell)CFA((groups[group]+tok));
   #endif
   #ifdef DIRECT_THREADED
                 if ((token & 0x4000) == 0) { /* threade code, no CFA */
                   if (targets[k] & (1U<<(RELINFOBITS-1-j)))
                     compile_prim1(0);
                   compile_prim1(&image[i]);
                 }
   #endif
               } else
                 fprintf(stderr,"Primitive %lx, %d of group %d used in this image at $%lx (offset $%x) is not implemented by this\n engine (%s); executing this code will crash.\n", (long)-token, tok, group, (long)&image[i],i,PACKAGE_VERSION);
             }
           } else {
             /* if base is > 0: 0 is a null reference so don't adjust*/
           if (token>=base) {            if (token>=base) {
             image[i]+=(Cell)start;              image[i]+=(Cell)start;
           }            }
Line 189 
Line 407 
       }        }
     }      }
   }    }
     free(targets);
     finish_code();
   ((ImageHeader*)(image))->base = (Address) image;    ((ImageHeader*)(image))->base = (Address) image;
 }  }
   
Line 225 
Line 445 
     exit(1);      exit(1);
   }    }
   r = (Address)((((Cell)r)+(sizeof(Float)-1))&(-sizeof(Float)));    r = (Address)((((Cell)r)+(sizeof(Float)-1))&(-sizeof(Float)));
   if (debug)    debugp(stderr, "malloc succeeds, address=$%lx\n", (long)r);
     fprintf(stderr, "malloc succeeds, address=$%lx\n", (long)r);  
   return r;    return r;
 }  }
   
Line 234 
Line 453 
 void after_alloc(Address r, Cell size)  void after_alloc(Address r, Cell size)
 {  {
   if (r != (Address)-1) {    if (r != (Address)-1) {
     if (debug)      debugp(stderr, "success, address=$%lx\n", (long) r);
       fprintf(stderr, "success, address=$%lx\n", (long) r);  
     if (pagesize != 1)      if (pagesize != 1)
       next_address = (Address)(((((Cell)r)+size-1)&-pagesize)+2*pagesize); /* leave one page unmapped */        next_address = (Address)(((((Cell)r)+size-1)&-pagesize)+2*pagesize); /* leave one page unmapped */
   } else {    } else {
     if (debug)      debugp(stderr, "failed: %s\n", strerror(errno));
       fprintf(stderr, "failed: %s\n", strerror(errno));  
   }    }
 }  }
   
Line 253 
Line 470 
 #ifndef MAP_PRIVATE  #ifndef MAP_PRIVATE
 # define MAP_PRIVATE 0  # define MAP_PRIVATE 0
 #endif  #endif
   #if !defined(MAP_ANON) && defined(MAP_ANONYMOUS)
   # define MAP_ANON MAP_ANONYMOUS
   #endif
   
 #if defined(HAVE_MMAP)  #if defined(HAVE_MMAP)
 static Address alloc_mmap(Cell size)  static Address alloc_mmap(Cell size)
Line 260 
Line 480 
   Address r;    Address r;
   
 #if defined(MAP_ANON)  #if defined(MAP_ANON)
   if (debug)    debugp(stderr,"try mmap($%lx, $%lx, ..., MAP_ANON, ...); ", (long)next_address, (long)size);
     fprintf(stderr,"try mmap($%lx, $%lx, ..., MAP_ANON, ...); ", (long)next_address, (long)size);  
   r = mmap(next_address, size, PROT_EXEC|PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE, -1, 0);    r = mmap(next_address, size, PROT_EXEC|PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE, -1, 0);
 #else /* !defined(MAP_ANON) */  #else /* !defined(MAP_ANON) */
   /* Ultrix (at least) does not define MAP_FILE and MAP_PRIVATE (both are    /* Ultrix (at least) does not define MAP_FILE and MAP_PRIVATE (both are
Line 272 
Line 491 
     dev_zero = open("/dev/zero", O_RDONLY);      dev_zero = open("/dev/zero", O_RDONLY);
   if (dev_zero == -1) {    if (dev_zero == -1) {
     r = MAP_FAILED;      r = MAP_FAILED;
     if (debug)      debugp(stderr, "open(\"/dev/zero\"...) failed (%s), no mmap; ",
       fprintf(stderr, "open(\"/dev/zero\"...) failed (%s), no mmap; ",  
               strerror(errno));                strerror(errno));
   } else {    } else {
     if (debug)      debugp(stderr,"try mmap($%lx, $%lx, ..., MAP_FILE, dev_zero, ...); ", (long)next_address, (long)size);
       fprintf(stderr,"try mmap($%lx, $%lx, ..., MAP_FILE, dev_zero, ...); ", (long)next_address, (long)size);  
     r=mmap(next_address, size, PROT_EXEC|PROT_READ|PROT_WRITE, MAP_FILE|MAP_PRIVATE, dev_zero, 0);      r=mmap(next_address, size, PROT_EXEC|PROT_READ|PROT_WRITE, MAP_FILE|MAP_PRIVATE, dev_zero, 0);
   }    }
 #endif /* !defined(MAP_ANON) */  #endif /* !defined(MAP_ANON) */
Line 292 
Line 509 
   Address r;    Address r;
   
   r=alloc_mmap(size);    r=alloc_mmap(size);
   if (r!=MAP_FAILED)    if (r!=(Address)MAP_FAILED)
     return r;      return r;
 #endif /* HAVE_MMAP */  #endif /* HAVE_MMAP */
   /* use malloc as fallback */    /* use malloc as fallback */
   return verbose_malloc(size);    return verbose_malloc(size);
 }  }
   
 #if (defined(mips) && !defined(INDIRECT_THREADED))  
 /* the 256MB jump restriction on the MIPS architecture makes the  
    combination of direct threading and mmap unsafe. */  
 #define mips_dict_alloc 1  
 #define dict_alloc(size) verbose_malloc(size)  
 #else  
 #define dict_alloc(size) my_alloc(size)  
 #endif  
   
 Address dict_alloc_read(FILE *file, Cell imagesize, Cell dictsize, Cell offset)  Address dict_alloc_read(FILE *file, Cell imagesize, Cell dictsize, Cell offset)
 {  {
   Address image = MAP_FAILED;    Address image = MAP_FAILED;
   
 #if defined(HAVE_MMAP) && !defined(mips_dict_alloc)  #if defined(HAVE_MMAP)
   if (offset==0) {    if (offset==0) {
     image=alloc_mmap(dictsize);      image=alloc_mmap(dictsize);
     if (debug)      debugp(stderr,"try mmap($%lx, $%lx, ..., MAP_FIXED|MAP_FILE, imagefile, 0); ", (long)image, (long)imagesize);
       fprintf(stderr,"try mmap($%lx, $%lx, ..., MAP_FIXED|MAP_FILE, imagefile, 0); ", (long)image, (long)imagesize);  
     image = mmap(image, imagesize, PROT_EXEC|PROT_READ|PROT_WRITE, MAP_FIXED|MAP_FILE|MAP_PRIVATE, fileno(file), 0);      image = mmap(image, imagesize, PROT_EXEC|PROT_READ|PROT_WRITE, MAP_FIXED|MAP_FILE|MAP_PRIVATE, fileno(file), 0);
     after_alloc(image,dictsize);      after_alloc(image,dictsize);
   }    }
 #endif /* defined(MAP_ANON) && !defined(mips_dict_alloc) */  #endif /* defined(HAVE_MMAP) */
   if (image == MAP_FAILED) {    if (image == (Address)MAP_FAILED) {
     image = dict_alloc(dictsize+offset)+offset;      image = my_alloc(dictsize+offset)+offset;
     rewind(file);  /* fseek(imagefile,0L,SEEK_SET); */      rewind(file);  /* fseek(imagefile,0L,SEEK_SET); */
     fread(image, 1, imagesize, file);      fread(image, 1, imagesize, file);
   }    }
Line 360 
Line 567 
   header->fp_stack_base=my_alloc(fsize);    header->fp_stack_base=my_alloc(fsize);
   header->return_stack_base=my_alloc(rsize);    header->return_stack_base=my_alloc(rsize);
   header->locals_stack_base=my_alloc(lsize);    header->locals_stack_base=my_alloc(lsize);
   code_here = code_area = my_alloc(dictsize);  
 }  }
   
 #warning You can ignore the warnings about clobbered variables in go_forth  #warning You can ignore the warnings about clobbered variables in go_forth
Line 399 
Line 605 
     signal_data_stack[7]=throw_code;      signal_data_stack[7]=throw_code;
   
 #ifdef GFORTH_DEBUGGING  #ifdef GFORTH_DEBUGGING
     /* fprintf(stderr,"\nrp=%ld\n",(long)rp); */      debugp(stderr,"\ncaught signal, throwing exception %d, ip=%p rp=%p\n",
                 throw_code, saved_ip, rp);
     if (rp <= orig_rp0 && rp > (Cell *)(image_header->return_stack_base+5)) {      if (rp <= orig_rp0 && rp > (Cell *)(image_header->return_stack_base+5)) {
       /* no rstack overflow or underflow */        /* no rstack overflow or underflow */
       rp0 = rp;        rp0 = rp;
       *--rp0 = (Cell)ip;        *--rp0 = (Cell)saved_ip;
     }      }
     else /* I love non-syntactic ifdefs :-) */      else /* I love non-syntactic ifdefs :-) */
 #endif  
     rp0 = signal_return_stack+8;      rp0 = signal_return_stack+8;
   #else  /* !defined(GFORTH_DEBUGGING) */
       debugp(stderr,"\ncaught signal, throwing exception %d\n", throw_code);
         rp0 = signal_return_stack+8;
   #endif /* !defined(GFORTH_DEBUGGING) */
     /* fprintf(stderr, "rp=$%x\n",rp0);*/      /* fprintf(stderr, "rp=$%x\n",rp0);*/
   
     return((int)(Cell)engine(image_header->throw_entry, signal_data_stack+7,      return((int)(Cell)engine(image_header->throw_entry, signal_data_stack+7,
Line 418 
Line 628 
   return((int)(Cell)engine(ip0,sp0,rp0,fp0,lp0));    return((int)(Cell)engine(ip0,sp0,rp0,fp0,lp0));
 }  }
   
   
 #ifndef INCLUDE_IMAGE  #ifndef INCLUDE_IMAGE
 void print_sizes(Cell sizebyte)  void print_sizes(Cell sizebyte)
      /* print size information */       /* print size information */
Line 432 
Line 641 
           1 << ((sizebyte >> 5) & 3));            1 << ((sizebyte >> 5) & 3));
 }  }
   
 typedef struct {  /* static superinstruction stuff */
   Label start;  
   Cell length; /* excluding the jump */  struct cost { /* super_info might be a more accurate name */
   char super_end; /* true if primitive ends superinstruction, i.e.,    char loads;       /* number of stack loads */
                      unconditional branch, execute, etc. */    char stores;      /* number of stack stores */
 } PrimInfo;    char updates;     /* number of stack pointer updates */
     char branch;      /* is it a branch (SET_IP) */
     unsigned char state_in;    /* state on entry */
     unsigned char state_out;   /* state on exit */
     unsigned char imm_ops;     /* number of immediate operands */
     short offset;     /* offset into super2 table */
     unsigned char length;      /* number of components */
   };
   
   PrimNum super2[] = {
   #include SUPER2_I
   };
   
   struct cost super_costs[] = {
   #include COSTS_I
   };
   
   struct super_state {
     struct super_state *next;
     PrimNum super;
   };
   
   #define HASH_SIZE 256
   
   struct super_table_entry {
     struct super_table_entry *next;
     PrimNum *start;
     short length;
     struct super_state *ss_list; /* list of supers */
   } *super_table[HASH_SIZE];
   int max_super=2;
   
   struct super_state *state_transitions=NULL;
   
   int hash_super(PrimNum *start, int length)
   {
     int i, r;
   
     for (i=0, r=0; i<length; i++) {
       r <<= 1;
       r += start[i];
     }
     return r & (HASH_SIZE-1);
   }
   
   struct super_state **lookup_super(PrimNum *start, int length)
   {
     int hash=hash_super(start,length);
     struct super_table_entry *p = super_table[hash];
   
     /* assert(length >= 2); */
     for (; p!=NULL; p = p->next) {
       if (length == p->length &&
           memcmp((char *)p->start, (char *)start, length*sizeof(PrimNum))==0)
         return &(p->ss_list);
     }
     return NULL;
   }
   
   void prepare_super_table()
   {
     int i;
     int nsupers = 0;
   
     for (i=0; i<sizeof(super_costs)/sizeof(super_costs[0]); i++) {
       struct cost *c = &super_costs[i];
       if ((c->length < 2 || nsupers < static_super_number) &&
           c->state_in < maxstates && c->state_out < maxstates) {
         struct super_state **ss_listp= lookup_super(super2+c->offset, c->length);
         struct super_state *ss = malloc(sizeof(struct super_state));
         ss->super= i;
         if (c->offset==N_noop && i != N_noop) {
           if (is_relocatable(i)) {
             ss->next = state_transitions;
             state_transitions = ss;
           }
         } else if (ss_listp != NULL) {
           ss->next = *ss_listp;
           *ss_listp = ss;
         } else {
           int hash = hash_super(super2+c->offset, c->length);
           struct super_table_entry **p = &super_table[hash];
           struct super_table_entry *e = malloc(sizeof(struct super_table_entry));
           ss->next = NULL;
           e->next = *p;
           e->start = super2 + c->offset;
           e->length = c->length;
           e->ss_list = ss;
           *p = e;
         }
         if (c->length > max_super)
           max_super = c->length;
         if (c->length >= 2)
           nsupers++;
       }
     }
     debugp(stderr, "Using %d static superinsts\n", nsupers);
   }
   
   /* dynamic replication/superinstruction stuff */
   
   #ifndef NO_DYNAMIC
   int compare_priminfo_length(const void *_a, const void *_b)
   {
     PrimInfo **a = (PrimInfo **)_a;
     PrimInfo **b = (PrimInfo **)_b;
     Cell diff = (*a)->length - (*b)->length;
     if (diff)
       return diff;
     else /* break ties by start address; thus the decompiler produces
             the earliest primitive with the same code (e.g. noop instead
             of (char) and @ instead of >code-address */
       return (*b)->start - (*a)->start;
   }
   #endif /* !defined(NO_DYNAMIC) */
   
   static char MAYBE_UNUSED superend[]={
   #include PRIM_SUPEREND_I
   };
   
 PrimInfo *priminfos;  
 Cell npriminfos=0;  Cell npriminfos=0;
   
   int compare_labels(const void *pa, const void *pb)
   {
     Label a = *(Label *)pa;
     Label b = *(Label *)pb;
     return a-b;
   }
   
   Label bsearch_next(Label key, Label *a, UCell n)
        /* a is sorted; return the label >=key that is the closest in a;
           return NULL if there is no label in a >=key */
   {
     int mid = (n-1)/2;
     if (n<1)
       return NULL;
     if (n == 1) {
       if (a[0] < key)
         return NULL;
       else
         return a[0];
     }
     if (a[mid] < key)
       return bsearch_next(key, a+mid+1, n-mid-1);
     else
       return bsearch_next(key, a, mid+1);
   }
   
 void check_prims(Label symbols1[])  void check_prims(Label symbols1[])
 {  {
 #ifndef DOUBLY_INDIRECT  
   int i;    int i;
   Label *symbols2=engine2(0,0,0,0,0);  #ifndef NO_DYNAMIC
   char superend[]={    Label *symbols2, *symbols3, *ends1, *ends1j, *ends1jsorted;
 #include "prim_superend.i"    int nends1j;
   };  #endif
   
   for (i=DOESJUMP+1; symbols1[i+1]!=0; i++)    if (debug)
   #ifdef __VERSION__
       fprintf(stderr, "Compiled with gcc-" __VERSION__ "\n");
   #else
   #define xstr(s) str(s)
   #define str(s) #s
     fprintf(stderr, "Compiled with gcc-" xstr(__GNUC__) "." xstr(__GNUC_MINOR__) "\n");
   #endif
     for (i=0; symbols1[i]!=0; i++)
     ;      ;
   priminfos = calloc(i,sizeof(PrimInfo));  
   npriminfos = i;    npriminfos = i;
   for (i=DOESJUMP+1; symbols1[i+1]!=0; i++) {  
     int prim_len=symbols1[i+1]-symbols1[i];  #ifndef NO_DYNAMIC
     if (no_dynamic)
       return;
     symbols2=engine2(0,0,0,0,0);
   #if NO_IP
     symbols3=engine3(0,0,0,0,0);
   #else
     symbols3=symbols1;
   #endif
     ends1 = symbols1+i+1;
     ends1j =   ends1+i;
     nends1j = i+1;
     ends1jsorted = (Label *)alloca(nends1j*sizeof(Label));
     memcpy(ends1jsorted,ends1j,nends1j*sizeof(Label));
     qsort(ends1jsorted, nends1j, sizeof(Label), compare_labels);
   
     priminfos = calloc(i,sizeof(PrimInfo));
     for (i=0; symbols1[i]!=0; i++) {
       int prim_len = ends1[i]-symbols1[i];
     PrimInfo *pi=&priminfos[i];      PrimInfo *pi=&priminfos[i];
     int j;      int j=0;
     pi->super_end = superend[i-DOESJUMP-1];      char *s1 = (char *)symbols1[i];
     for (j=prim_len-3; ; j--) {      char *s2 = (char *)symbols2[i];
       if (((*(Cell *)(symbols1[i]+j)) & 0xfff8ff) == 0xfc60ff) {      char *s3 = (char *)symbols3[i];
         /* jmp -4(reg), i.e., the NEXT jump */      Label endlabel = bsearch_next(symbols1[i]+1,ends1jsorted,nends1j);
         prim_len = j;  
         if (pi->super_end)      pi->start = s1;
           prim_len += 3; /* include the jump */      pi->superend = superend[i]|no_super;
       if (pi->superend)
         pi->length = endlabel-symbols1[i];
       else
         pi->length = prim_len;
       pi->restlength = endlabel - symbols1[i] - pi->length;
       pi->nimmargs = 0;
       relocs++;
       debugp(stderr, "%-15s %3d %p %p %p len=%3ld restlen=%2ld s-end=%1d",
                 prim_names[i], i, s1, s2, s3, (long)(pi->length), (long)(pi->restlength), pi->superend);
       if (endlabel == NULL) {
         pi->start = NULL; /* not relocatable */
         if (pi->length<0) pi->length=100;
         debugp(stderr,"\n   non_reloc: no J label > start found\n");
         relocs--;
         nonrelocs++;
         continue;
       }
       if (ends1[i] > endlabel && !pi->superend) {
         pi->start = NULL; /* not relocatable */
         pi->length = endlabel-symbols1[i];
         debugp(stderr,"\n   non_reloc: there is a J label before the K label (restlength<0)\n");
         relocs--;
         nonrelocs++;
         continue;
       }
       if (ends1[i] < pi->start && !pi->superend) {
         pi->start = NULL; /* not relocatable */
         pi->length = endlabel-symbols1[i];
         debugp(stderr,"\n   non_reloc: K label before I label (length<0)\n");
         relocs--;
         nonrelocs++;
         continue;
       }
       assert(pi->length>=0);
       assert(pi->restlength >=0);
       while (j<(pi->length+pi->restlength)) {
         if (s1[j]==s3[j]) {
           if (s1[j] != s2[j]) {
             pi->start = NULL; /* not relocatable */
             debugp(stderr,"\n   non_reloc: engine1!=engine2 offset %3d",j);
             /* assert(j<prim_len); */
             relocs--;
             nonrelocs++;
         break;          break;
       }        }
       if (j==0) { /* NEXT jump not found, e.g., execute */          j++;
         if (!pi->super_end && debug)        } else {
           fprintf(stderr, "NEXT jump not found for primitive %d, making it super_end\n", i);          struct immarg *ia=&pi->immargs[pi->nimmargs];
         pi->super_end = 1;  
           pi->nimmargs++;
           ia->offset=j;
           if ((~*(Cell *)&(s1[j]))==*(Cell *)&(s3[j])) {
             ia->rel=0;
             debugp(stderr,"\n   absolute immarg: offset %3d",j);
           } else if ((&(s1[j]))+(*(Cell *)&(s1[j]))+4 ==
                      symbols1[DOESJUMP+1]) {
             ia->rel=1;
             debugp(stderr,"\n   relative immarg: offset %3d",j);
           } else {
             pi->start = NULL; /* not relocatable */
             debugp(stderr,"\n   non_reloc: engine1!=engine3 offset %3d",j);
             /* assert(j<prim_len);*/
             relocs--;
             nonrelocs++;
         break;          break;
       }        }
           j+=4;
     }      }
     /* fprintf(stderr,"checking primitive %d: memcmp(%p, %p, %d)\n",  
        i, symbols1[i], symbols2[i], prim_len);*/  
     if (memcmp(symbols1[i],symbols2[i],prim_len)!=0) {  
       if (debug)  
         fprintf(stderr,"Primitive %d not relocatable: memcmp(%p, %p, %d)\n",  
                 i, symbols1[i], symbols2[i], prim_len);  
     } else {  
       pi->start = symbols1[i];  
       pi->length = prim_len;  
       if (debug)  
         fprintf(stderr,"Primitive %d relocatable: start %p, length %ld, super_end %d\n",  
                 i, pi->start, pi->length, pi->super_end);  
     }      }
       debugp(stderr,"\n");
   }    }
     decomp_prims = calloc(i,sizeof(PrimInfo *));
     for (i=DOESJUMP+1; i<npriminfos; i++)
       decomp_prims[i] = &(priminfos[i]);
     qsort(decomp_prims+DOESJUMP+1, npriminfos-DOESJUMP-1, sizeof(PrimInfo *),
           compare_priminfo_length);
 #endif  #endif
 }  }
   
 Label compile_prim(Label prim)  void flush_to_here(void)
 {  {
   int i;  #ifndef NO_DYNAMIC
     if (start_flush)
       FLUSH_ICACHE(start_flush, code_here-start_flush);
     start_flush=code_here;
   #endif
   }
   
   #ifndef NO_DYNAMIC
   void append_jump(void)
   {
     if (last_jump) {
       PrimInfo *pi = &priminfos[last_jump];
   
       memcpy(code_here, pi->start+pi->length, pi->restlength);
       code_here += pi->restlength;
       last_jump=0;
     }
   }
   
   /* Gforth remembers all code blocks in this list.  On forgetting (by
   executing a marker) the code blocks are not freed (because Gforth does
   not remember how they were allocated; hmm, remembering that might be
   easier and cleaner).  Instead, code_here etc. are reset to the old
   value, and the "forgotten" code blocks are reused when they are
   needed. */
   
   struct code_block_list {
     struct code_block_list *next;
     Address block;
     Cell size;
   } *code_block_list=NULL, **next_code_blockp=&code_block_list;
   
   Address append_prim(Cell p)
   {
     PrimInfo *pi = &priminfos[p];
   Address old_code_here=code_here;    Address old_code_here=code_here;
   static Address last_jump=0;  
   
   for (i=0; ; i++) {    if (code_area+code_area_size < code_here+pi->length+pi->restlength) {
     if (i>=npriminfos) { /* not a relocatable prim */      struct code_block_list *p;
       if (last_jump) { /* make sure the last sequence is complete */      append_jump();
         memcpy(code_here, last_jump, 3);      flush_to_here();
         code_here += 3;      if (*next_code_blockp == NULL) {
         code_here = start_flush = code_area = my_alloc(code_area_size);
         p = (struct code_block_list *)malloc(sizeof(struct code_block_list));
         *next_code_blockp = p;
         p->next = NULL;
         p->block = code_here;
         p->size = code_area_size;
       } else {
         p = *next_code_blockp;
         code_here = start_flush = code_area = p->block;
       }
       old_code_here = code_here;
       next_code_blockp = &(p->next);
     }
     memcpy(code_here, pi->start, pi->length);
     code_here += pi->length;
     return old_code_here;
   }
   #endif
   
   int forget_dyncode(Address code)
   {
   #ifdef NO_DYNAMIC
     return -1;
   #else
     struct code_block_list *p, **pp;
   
     for (pp=&code_block_list, p=*pp; p!=NULL; pp=&(p->next), p=*pp) {
       if (code >= p->block && code < p->block+p->size) {
         next_code_blockp = &(p->next);
         code_here = start_flush = code;
         code_area = p->block;
         last_jump = 0;          last_jump = 0;
         return -1;
       }        }
       return prim;  
     }      }
     if (priminfos[i].start==prim)    return -no_dynamic;
   #endif /* !defined(NO_DYNAMIC) */
   }
   
   long dyncodesize(void)
   {
   #ifndef NO_DYNAMIC
     struct code_block_list *p;
     long size=0;
     for (p=code_block_list; p!=NULL; p=p->next) {
       if (code_here >= p->block && code_here < p->block+p->size)
         return size + (code_here - p->block);
       else
         size += p->size;
     }
   #endif /* !defined(NO_DYNAMIC) */
     return 0;
   }
   
   Label decompile_code(Label _code)
   {
   #ifdef NO_DYNAMIC
     return _code;
   #else /* !defined(NO_DYNAMIC) */
     Cell i;
     struct code_block_list *p;
     Address code=_code;
   
     /* first, check if we are in code at all */
     for (p = code_block_list;; p = p->next) {
       if (p == NULL)
         return code;
       if (code >= p->block && code < p->block+p->size)
       break;        break;
   }    }
   memcpy(code_here, (Address)prim, priminfos[i].length);    /* reverse order because NOOP might match other prims */
   code_here += priminfos[i].length;    for (i=npriminfos-1; i>DOESJUMP; i--) {
   last_jump = (priminfos[i].super_end) ? 0 : (prim+priminfos[i].length);      PrimInfo *pi=decomp_prims[i];
   return (Label)old_code_here;      if (pi->start==code || (pi->start && memcmp(code,pi->start,pi->length)==0))
         return vm_prims[super2[super_costs[pi-priminfos].offset]];
       /* return pi->start;*/
     }
     return code;
   #endif /* !defined(NO_DYNAMIC) */
   }
   
   #ifdef NO_IP
   int nbranchinfos=0;
   
   struct branchinfo {
     Label **targetpp; /* **(bi->targetpp) is the target */
     Cell *addressptr; /* store the target here */
   } branchinfos[100000];
   
   int ndoesexecinfos=0;
   struct doesexecinfo {
     int branchinfo; /* fix the targetptr of branchinfos[...->branchinfo] */
     Label *targetp; /*target for branch (because this is not in threaded code)*/
     Cell *xt; /* cfa of word whose does-code needs calling */
   } doesexecinfos[10000];
   
   void set_rel_target(Cell *source, Label target)
   {
     *source = ((Cell)target)-(((Cell)source)+4);
   }
   
   void register_branchinfo(Label source, Cell *targetpp)
   {
     struct branchinfo *bi = &(branchinfos[nbranchinfos]);
     bi->targetpp = (Label **)targetpp;
     bi->addressptr = (Cell *)source;
     nbranchinfos++;
   }
   
   Address compile_prim1arg(PrimNum p, Cell **argp)
   {
     Address old_code_here=append_prim(p);
   
     assert(vm_prims[p]==priminfos[p].start);
     *argp = (Cell*)(old_code_here+priminfos[p].immargs[0].offset);
     return old_code_here;
   }
   
   Address compile_call2(Cell *targetpp, Cell **next_code_targetp)
   {
     PrimInfo *pi = &priminfos[N_call2];
     Address old_code_here = append_prim(N_call2);
   
     *next_code_targetp = (Cell *)(old_code_here + pi->immargs[0].offset);
     register_branchinfo(old_code_here + pi->immargs[1].offset, targetpp);
     return old_code_here;
   }
   #endif
   
   void finish_code(void)
   {
   #ifdef NO_IP
     Cell i;
   
     compile_prim1(NULL);
     for (i=0; i<ndoesexecinfos; i++) {
       struct doesexecinfo *dei = &doesexecinfos[i];
       dei->targetp = (Label *)DOES_CODE1((dei->xt));
       branchinfos[dei->branchinfo].targetpp = &(dei->targetp);
     }
     ndoesexecinfos = 0;
     for (i=0; i<nbranchinfos; i++) {
       struct branchinfo *bi=&branchinfos[i];
       set_rel_target(bi->addressptr, **(bi->targetpp));
     }
     nbranchinfos = 0;
   #else
     compile_prim1(NULL);
   #endif
     flush_to_here();
   }
   
   #ifdef NO_IP
   Cell compile_prim_dyn(PrimNum p, Cell *tcp)
        /* compile prim #p dynamically (mod flags etc.) and return start
           address of generated code for putting it into the threaded
           code. This function is only called if all the associated
           inline arguments of p are already in place (at tcp[1] etc.) */
   {
     PrimInfo *pi=&priminfos[p];
     Cell *next_code_target=NULL;
     Address codeaddr;
     Address primstart;
   
     assert(p<npriminfos);
     if (p==N_execute || p==N_perform || p==N_lit_perform) {
       codeaddr = compile_prim1arg(N_set_next_code, &next_code_target);
       primstart = append_prim(p);
       goto other_prim;
     } else if (p==N_call) {
       codeaddr = compile_call2(tcp+1, &next_code_target);
     } else if (p==N_does_exec) {
       struct doesexecinfo *dei = &doesexecinfos[ndoesexecinfos++];
       Cell *arg;
       codeaddr = compile_prim1arg(N_lit,&arg);
       *arg = (Cell)PFA(tcp[1]);
       /* we cannot determine the callee now (last_start[1] may be a
          forward reference), so just register an arbitrary target, and
          register in dei that we need to fix this before resolving
          branches */
       dei->branchinfo = nbranchinfos;
       dei->xt = (Cell *)(tcp[1]);
       compile_call2(0, &next_code_target);
     } else if (!is_relocatable(p)) {
       Cell *branch_target;
       codeaddr = compile_prim1arg(N_set_next_code, &next_code_target);
       compile_prim1arg(N_branch,&branch_target);
       set_rel_target(branch_target,vm_prims[p]);
     } else {
       unsigned j;
   
       codeaddr = primstart = append_prim(p);
     other_prim:
       for (j=0; j<pi->nimmargs; j++) {
         struct immarg *ia = &(pi->immargs[j]);
         Cell *argp = tcp + pi->nimmargs - j;
         Cell argval = *argp; /* !! specific to prims */
         if (ia->rel) { /* !! assumption: relative refs are branches */
           register_branchinfo(primstart + ia->offset, argp);
         } else /* plain argument */
           *(Cell *)(primstart + ia->offset) = argval;
       }
     }
     if (next_code_target!=NULL)
       *next_code_target = (Cell)code_here;
     return (Cell)codeaddr;
   }
   #else /* !defined(NO_IP) */
   Cell compile_prim_dyn(PrimNum p, Cell *tcp)
        /* compile prim #p dynamically (mod flags etc.) and return start
           address of generated code for putting it into the threaded code */
   {
     Cell static_prim = (Cell)vm_prims[p];
   #if defined(NO_DYNAMIC)
     return static_prim;
   #else /* !defined(NO_DYNAMIC) */
     Address old_code_here;
   
     if (no_dynamic)
       return static_prim;
     if (p>=npriminfos || !is_relocatable(p)) {
       append_jump();
       return static_prim;
     }
     old_code_here = append_prim(p);
     last_jump = (priminfos[p].superend) ? 0 : p;
     return (Cell)old_code_here;
   #endif  /* !defined(NO_DYNAMIC) */
   }
   #endif /* !defined(NO_IP) */
   
   #ifndef NO_DYNAMIC
   int cost_codesize(int prim)
   {
     return priminfos[prim].length;
   }
   #endif
   
   int cost_ls(int prim)
   {
     struct cost *c = super_costs+prim;
   
     return c->loads + c->stores;
   }
   
   int cost_lsu(int prim)
   {
     struct cost *c = super_costs+prim;
   
     return c->loads + c->stores + c->updates;
   }
   
   int cost_nexts(int prim)
   {
     return 1;
   }
   
   typedef int Costfunc(int);
   Costfunc *ss_cost =  /* cost function for optimize_bb */
   #ifdef NO_DYNAMIC
   cost_lsu;
   #else
   cost_codesize;
   #endif
   
   struct {
     Costfunc *costfunc;
     char *metricname;
     long sum;
   } cost_sums[] = {
   #ifndef NO_DYNAMIC
     { cost_codesize, "codesize", 0 },
   #endif
     { cost_ls,       "ls",       0 },
     { cost_lsu,      "lsu",      0 },
     { cost_nexts,    "nexts",    0 }
   };
   
   #define MAX_BB 128 /* maximum number of instructions in BB */
   #define INF_COST 1000000 /* infinite cost */
   #define CANONICAL_STATE 0
   
   struct waypoint {
     int cost;     /* the cost from here to the end */
     PrimNum inst; /* the inst used from here to the next waypoint */
     char relocatable; /* the last non-transition was relocatable */
     char no_transition; /* don't use the next transition (relocatability)
                          * or this transition (does not change state) */
   };
   
   void init_waypoints(struct waypoint ws[])
   {
     int k;
   
     for (k=0; k<maxstates; k++)
       ws[k].cost=INF_COST;
   }
   
   void transitions(struct waypoint inst[], struct waypoint trans[])
   {
     int k;
     struct super_state *l;
   
     for (k=0; k<maxstates; k++) {
       trans[k] = inst[k];
       trans[k].no_transition = 1;
     }
     for (l = state_transitions; l != NULL; l = l->next) {
       PrimNum s = l->super;
       int jcost;
       struct cost *c=super_costs+s;
       struct waypoint *wi=&(trans[c->state_in]);
       struct waypoint *wo=&(inst[c->state_out]);
       if (wo->cost == INF_COST)
         continue;
       jcost = wo->cost + ss_cost(s);
       if (jcost <= wi->cost) {
         wi->cost = jcost;
         wi->inst = s;
         wi->relocatable = wo->relocatable;
         wi->no_transition = 0;
         /* if (ss_greedy) wi->cost = wo->cost ? */
       }
     }
   }
   
   /* use dynamic programming to find the shortest paths within the basic
      block origs[0..ninsts-1] and rewrite the instructions pointed to by
      instps to use it */
   void optimize_rewrite(Cell *instps[], PrimNum origs[], int ninsts)
   {
     int i,j;
     static struct waypoint inst[MAX_BB+1][MAX_STATE];  /* before instruction*/
     static struct waypoint trans[MAX_BB+1][MAX_STATE]; /* before transition */
     int nextdyn, nextstate, no_transition;
   
     init_waypoints(inst[ninsts]);
     inst[ninsts][CANONICAL_STATE].cost=0;
     transitions(inst[ninsts],trans[ninsts]);
     for (i=ninsts-1; i>=0; i--) {
       init_waypoints(inst[i]);
       for (j=1; j<=max_super && i+j<=ninsts; j++) {
         struct super_state **superp = lookup_super(origs+i, j);
         if (superp!=NULL) {
           struct super_state *supers = *superp;
           for (; supers!=NULL; supers = supers->next) {
             PrimNum s = supers->super;
             int jcost;
             struct cost *c=super_costs+s;
             struct waypoint *wi=&(inst[i][c->state_in]);
             struct waypoint *wo=&(trans[i+j][c->state_out]);
             int no_transition = wo->no_transition;
             if (!(is_relocatable(s)) && !wo->relocatable) {
               wo=&(inst[i+j][c->state_out]);
               no_transition=1;
             }
             if (wo->cost == INF_COST)
               continue;
             jcost = wo->cost + ss_cost(s);
             if (jcost <= wi->cost) {
               wi->cost = jcost;
               wi->inst = s;
               wi->relocatable = is_relocatable(s);
               wi->no_transition = no_transition;
               /* if (ss_greedy) wi->cost = wo->cost ? */
             }
           }
         }
       }
       transitions(inst[i],trans[i]);
     }
     /* now rewrite the instructions */
     nextdyn=0;
     nextstate=CANONICAL_STATE;
     no_transition = ((!trans[0][nextstate].relocatable)
                      ||trans[0][nextstate].no_transition);
     for (i=0; i<ninsts; i++) {
       Cell tc=0, tc2;
       if (i==nextdyn) {
         if (!no_transition) {
           /* process trans */
           PrimNum p = trans[i][nextstate].inst;
           struct cost *c = super_costs+p;
           assert(trans[i][nextstate].cost != INF_COST);
           assert(c->state_in==nextstate);
           tc = compile_prim_dyn(p,NULL);
           nextstate = c->state_out;
         }
         {
           /* process inst */
           PrimNum p = inst[i][nextstate].inst;
           struct cost *c=super_costs+p;
           assert(c->state_in==nextstate);
           assert(inst[i][nextstate].cost != INF_COST);
   #if defined(GFORTH_DEBUGGING)
           assert(p == origs[i]);
   #endif
           tc2 = compile_prim_dyn(p,instps[i]);
           if (no_transition || !is_relocatable(p))
             /* !! actually what we care about is if and where
              * compile_prim_dyn() puts NEXTs */
             tc=tc2;
           no_transition = inst[i][nextstate].no_transition;
           nextstate = c->state_out;
           nextdyn += c->length;
         }
       } else {
   #if defined(GFORTH_DEBUGGING)
         assert(0);
   #endif
         tc=0;
         /* tc= (Cell)vm_prims[inst[i][CANONICAL_STATE].inst]; */
       }
       *(instps[i]) = tc;
     }
     if (!no_transition) {
       PrimNum p = trans[i][nextstate].inst;
       struct cost *c = super_costs+p;
       assert(c->state_in==nextstate);
       assert(trans[i][nextstate].cost != INF_COST);
       assert(i==nextdyn);
       (void)compile_prim_dyn(p,NULL);
       nextstate = c->state_out;
     }
     assert(nextstate==CANONICAL_STATE);
   }
   
   /* compile *start, possibly rewriting it into a static and/or dynamic
      superinstruction */
   void compile_prim1(Cell *start)
   {
   #if defined(DOUBLY_INDIRECT)
     Label prim;
   
     if (start==NULL)
       return;
     prim = (Label)*start;
     if (prim<((Label)(xts+DOESJUMP)) || prim>((Label)(xts+npriminfos))) {
       fprintf(stderr,"compile_prim encountered xt %p\n", prim);
       *start=(Cell)prim;
       return;
     } else {
       *start = (Cell)(prim-((Label)xts)+((Label)vm_prims));
       return;
     }
   #elif defined(INDIRECT_THREADED)
     return;
   #else /* !(defined(DOUBLY_INDIRECT) || defined(INDIRECT_THREADED)) */
     /* !! does not work, for unknown reasons; but something like this is
        probably needed to ensure that we don't call compile_prim_dyn
        before the inline arguments are there */
     static Cell *instps[MAX_BB];
     static PrimNum origs[MAX_BB];
     static int ninsts=0;
     PrimNum prim_num;
   
     if (start==NULL || ninsts >= MAX_BB ||
         (ninsts>0 && superend[origs[ninsts-1]])) {
       /* after bb, or at the start of the next bb */
       optimize_rewrite(instps,origs,ninsts);
       /* fprintf(stderr,"optimize_rewrite(...,%d)\n",ninsts); */
       ninsts=0;
       if (start==NULL)
         return;
     }
     prim_num = ((Xt)*start)-vm_prims;
     if(prim_num >= npriminfos) {
       optimize_rewrite(instps,origs,ninsts);
       /* fprintf(stderr,"optimize_rewrite(...,%d)\n",ninsts);*/
       ninsts=0;
       return;
     }
     assert(ninsts<MAX_BB);
     instps[ninsts] = start;
     origs[ninsts] = prim_num;
     ninsts++;
   #endif /* !(defined(DOUBLY_INDIRECT) || defined(INDIRECT_THREADED)) */
 }  }
   
 Address loader(FILE *imagefile, char* filename)  Address loader(FILE *imagefile, char* filename)
Line 546 
Line 1474 
   
   vm_prims = engine(0,0,0,0,0);    vm_prims = engine(0,0,0,0,0);
   check_prims(vm_prims);    check_prims(vm_prims);
     prepare_super_table();
 #ifndef DOUBLY_INDIRECT  #ifndef DOUBLY_INDIRECT
   #ifdef PRINT_SUPER_LENGTHS
     print_super_lengths();
   #endif
   check_sum = checksum(vm_prims);    check_sum = checksum(vm_prims);
 #else /* defined(DOUBLY_INDIRECT) */  #else /* defined(DOUBLY_INDIRECT) */
   check_sum = (UCell)vm_prims;    check_sum = (UCell)vm_prims;
Line 554 
Line 1486 
   
   do {    do {
     if(fread(magic,sizeof(Char),8,imagefile) < 8) {      if(fread(magic,sizeof(Char),8,imagefile) < 8) {
       fprintf(stderr,"%s: image %s doesn't seem to be a Gforth (>=0.4) image.\n",        fprintf(stderr,"%s: image %s doesn't seem to be a Gforth (>=0.6) image.\n",
               progname, filename);                progname, filename);
       exit(1);        exit(1);
     }      }
     preamblesize+=8;      preamblesize+=8;
   } while(memcmp(magic,"Gforth2",7));    } while(memcmp(magic,"Gforth3",7));
   magic7 = magic[7];    magic7 = magic[7];
   if (debug) {    if (debug) {
     magic[7]='\0';      magic[7]='\0';
Line 587 
Line 1519 
 #elif PAGESIZE  #elif PAGESIZE
   pagesize=PAGESIZE; /* in limits.h according to Gallmeister's POSIX.4 book */    pagesize=PAGESIZE; /* in limits.h according to Gallmeister's POSIX.4 book */
 #endif  #endif
   if (debug)    debugp(stderr,"pagesize=%ld\n",(unsigned long) pagesize);
     fprintf(stderr,"pagesize=%ld\n",(unsigned long) pagesize);  
   
   image = dict_alloc_read(imagefile, preamblesize+header.image_size,    image = dict_alloc_read(imagefile, preamblesize+header.image_size,
                           preamblesize+dictsize, data_offset);                            preamblesize+dictsize, data_offset);
   imp=image+preamblesize;    imp=image+preamblesize;
     alloc_stacks((ImageHeader *)imp);
   if (clear_dictionary)    if (clear_dictionary)
     memset(imp+header.image_size, 0, dictsize-header.image_size);      memset(imp+header.image_size, 0, dictsize-header.image_size);
   if(header.base==0 || header.base  == 0x100) {    if(header.base==0 || header.base  == (Address)0x100) {
     Cell reloc_size=((header.image_size-1)/sizeof(Cell))/8+1;      Cell reloc_size=((header.image_size-1)/sizeof(Cell))/8+1;
     char reloc_bits[reloc_size];      char reloc_bits[reloc_size];
     fseek(imagefile, preamblesize+header.image_size, SEEK_SET);      fseek(imagefile, preamblesize+header.image_size, SEEK_SET);
     fread(reloc_bits, 1, reloc_size, imagefile);      fread(reloc_bits, 1, reloc_size, imagefile);
     relocate((Cell *)imp, reloc_bits, header.image_size, header.base, vm_prims);      relocate((Cell *)imp, reloc_bits, header.image_size, (Cell)header.base, vm_prims);
 #if 0  #if 0
     { /* let's see what the relocator did */      { /* let's see what the relocator did */
       FILE *snapshot=fopen("snapshot.fi","wb");        FILE *snapshot=fopen("snapshot.fi","wb");
Line 621 
Line 1553 
             progname, (unsigned long)(header.checksum),(unsigned long)check_sum);              progname, (unsigned long)(header.checksum),(unsigned long)check_sum);
     exit(1);      exit(1);
   }    }
   #ifdef DOUBLY_INDIRECT
     ((ImageHeader *)imp)->xt_base = xts;
   #endif
   fclose(imagefile);    fclose(imagefile);
   
   alloc_stacks((ImageHeader *)imp);    /* unnecessary, except maybe for CODE words */
     /* FLUSH_ICACHE(imp, header.image_size);*/
   CACHE_FLUSH(imp, header.image_size);  
   
   return imp;    return imp;
 }  }
   
 /* index of last '/' or '\' in file, 0 if there is none. !! Hmm, could  /* pointer to last '/' or '\' in file, 0 if there is none. */
    be implemented with strrchr and the separator should be  char *onlypath(char *filename)
    OS-dependent */  
 int onlypath(char *file)  
 {  {
   int i;    return strrchr(filename, DIRSEP);
   i=strlen(file);  
   while (i) {  
     if (file[i]=='\\' || file[i]=='/') break;  
     i--;  
   }  
   return i;  
 }  }
   
 FILE *openimage(char *fullfilename)  FILE *openimage(char *fullfilename)
Line 662 
Line 1588 
   char fullfilename[dirlen+strlen(imagename)+2];    char fullfilename[dirlen+strlen(imagename)+2];
   
   memcpy(fullfilename, path, dirlen);    memcpy(fullfilename, path, dirlen);
   if (fullfilename[dirlen-1]!='/')    if (fullfilename[dirlen-1]!=DIRSEP)
     fullfilename[dirlen++]='/';      fullfilename[dirlen++]=DIRSEP;
   strcpy(fullfilename+dirlen,imagename);    strcpy(fullfilename+dirlen,imagename);
   return openimage(fullfilename);    return openimage(fullfilename);
 }  }
Line 673 
Line 1599 
   FILE * image_file=NULL;    FILE * image_file=NULL;
   char *origpath=path;    char *origpath=path;
   
   if(strchr(imagename, '/')==NULL) {    if(strchr(imagename, DIRSEP)==NULL) {
     /* first check the directory where the exe file is in !! 01may97jaw */      /* first check the directory where the exe file is in !! 01may97jaw */
     if (onlypath(progname))      if (onlypath(progname))
       image_file=checkimage(progname, onlypath(progname), imagename);        image_file=checkimage(progname, onlypath(progname)-progname, imagename);
     if (!image_file)      if (!image_file)
       do {        do {
         char *pend=strchr(path, PATHSEP);          char *pend=strchr(path, PATHSEP);
Line 735 
Line 1661 
   return n*m;    return n*m;
 }  }
   
   enum {
     ss_number = 256,
     ss_states,
     ss_min_codesize,
     ss_min_ls,
     ss_min_lsu,
     ss_min_nexts,
   };
   
 void gforth_args(int argc, char ** argv, char ** path, char ** imagename)  void gforth_args(int argc, char ** argv, char ** path, char ** imagename)
 {  {
   int c;    int c;
Line 759 
Line 1694 
       {"clear-dictionary", no_argument, &clear_dictionary, 1},        {"clear-dictionary", no_argument, &clear_dictionary, 1},
       {"die-on-signal", no_argument, &die_on_signal, 1},        {"die-on-signal", no_argument, &die_on_signal, 1},
       {"debug", no_argument, &debug, 1},        {"debug", no_argument, &debug, 1},
         {"diag", no_argument, &diag, 1},
         {"no-super", no_argument, &no_super, 1},
         {"no-dynamic", no_argument, &no_dynamic, 1},
         {"dynamic", no_argument, &no_dynamic, 0},
         {"print-metrics", no_argument, &print_metrics, 1},
         {"ss-number", required_argument, NULL, ss_number},
         {"ss-states", required_argument, NULL, ss_states},
   #ifndef NO_DYNAMIC
         {"ss-min-codesize", no_argument, NULL, ss_min_codesize},
   #endif
         {"ss-min-ls",       no_argument, NULL, ss_min_ls},
         {"ss-min-lsu",      no_argument, NULL, ss_min_lsu},
         {"ss-min-nexts",    no_argument, NULL, ss_min_nexts},
         {"ss-greedy",       no_argument, &ss_greedy, 1},
       {0,0,0,0}        {0,0,0,0}
       /* no-init-file, no-rc? */        /* no-init-file, no-rc? */
     };      };
Line 781 
Line 1730 
     case 'c': clear_dictionary = 1; break;      case 'c': clear_dictionary = 1; break;
     case 's': die_on_signal = 1; break;      case 's': die_on_signal = 1; break;
     case 'x': debug = 1; break;      case 'x': debug = 1; break;
     case 'v': fprintf(stderr, "gforth %s\n", VERSION); exit(0);      case 'v': fputs(PACKAGE_STRING"\n", stderr); exit(0);
       case ss_number: static_super_number = atoi(optarg); break;
       case ss_states: maxstates = max(min(atoi(optarg),MAX_STATE),1); break;
   #ifndef NO_DYNAMIC
       case ss_min_codesize: ss_cost = cost_codesize; break;
   #endif
       case ss_min_ls:       ss_cost = cost_ls;       break;
       case ss_min_lsu:      ss_cost = cost_lsu;      break;
       case ss_min_nexts:    ss_cost = cost_nexts;    break;
     case 'h':      case 'h':
       fprintf(stderr, "Usage: %s [engine options] ['--'] [image arguments]\n\        fprintf(stderr, "Usage: %s [engine options] ['--'] [image arguments]\n\
 Engine Options:\n\  Engine Options:\n\
Line 789 
Line 1746 
   --clear-dictionary                Initialize the dictionary with 0 bytes\n\    --clear-dictionary                Initialize the dictionary with 0 bytes\n\
   -d SIZE, --data-stack-size=SIZE   Specify data stack size\n\    -d SIZE, --data-stack-size=SIZE   Specify data stack size\n\
   --debug                           Print debugging information during startup\n\    --debug                           Print debugging information during startup\n\
     --diag                            Print diagnostic information during startup\n\
   --die-on-signal                   exit instead of CATCHing some signals\n\    --die-on-signal                   exit instead of CATCHing some signals\n\
     --dynamic                         use dynamic native code\n\
   -f SIZE, --fp-stack-size=SIZE     Specify floating point stack size\n\    -f SIZE, --fp-stack-size=SIZE     Specify floating point stack size\n\
   -h, --help                        Print this message and exit\n\    -h, --help                        Print this message and exit\n\
   -i FILE, --image-file=FILE        Use image FILE instead of `gforth.fi'\n\    -i FILE, --image-file=FILE        Use image FILE instead of `gforth.fi'\n\
   -l SIZE, --locals-stack-size=SIZE Specify locals stack size\n\    -l SIZE, --locals-stack-size=SIZE Specify locals stack size\n\
   -m SIZE, --dictionary-size=SIZE   Specify Forth dictionary size\n\    -m SIZE, --dictionary-size=SIZE   Specify Forth dictionary size\n\
     --no-dynamic                      Use only statically compiled primitives\n\
   --no-offset-im                    Load image at normal position\n\    --no-offset-im                    Load image at normal position\n\
     --no-super                        No dynamically formed superinstructions\n\
   --offset-image                    Load image at a different position\n\    --offset-image                    Load image at a different position\n\
   -p PATH, --path=PATH              Search path for finding image and sources\n\    -p PATH, --path=PATH              Search path for finding image and sources\n\
     --print-metrics                   Print some code generation metrics on exit\n\
   -r SIZE, --return-stack-size=SIZE Specify return stack size\n\    -r SIZE, --return-stack-size=SIZE Specify return stack size\n\
   -v, --version                     Print version and exit\n\    --ss-greedy                       greedy, not optimal superinst selection\n\
     --ss-min-codesize                 select superinsts for smallest native code\n\
     --ss-min-ls                       minimize loads and stores\n\
     --ss-min-lsu                      minimize loads, stores, and pointer updates\n\
     --ss-min-nexts                    minimize the number of static superinsts\n\
     --ss-number=N                     use N static superinsts (default max)\n\
     --ss-states=N                     N states for stack caching (default max)\n\
     -v, --version                     Print engine version and exit\n\
 SIZE arguments consist of an integer followed by a unit. The unit can be\n\  SIZE arguments consist of an integer followed by a unit. The unit can be\n\
   `b' (byte), `e' (element; default), `k' (KB), `M' (MB), `G' (GB) or `T' (TB).\n",    `b' (byte), `e' (element; default), `k' (KB), `M' (MB), `G' (GB) or `T' (TB).\n",
               argv[0]);                argv[0]);
Line 810 
Line 1779 
 }  }
 #endif  #endif
   
   void print_diag()
   {
   
   #if !defined(HAVE_GETRUSAGE) || !defined(HAS_FFCALL)
     fprintf(stderr, "*** missing functionality ***\n"
   #ifndef HAVE_GETRUSAGE
             "    no getrusage -> CPUTIME broken\n"
   #endif
   #ifndef HAS_FFCALL
             "    no ffcall -> only old-style foreign function calls (no fflib.fs)\n"
   #endif
             );
   #endif
     if((relocs < nonrelocs) ||
   #if defined(BUGGY_LL_CMP) || defined(BUGGY_LL_MUL) || defined(BUGGY_LL_DIV) || defined(BUGGY_LL_ADD) || defined(BUGGY_LL_SHIFT) || defined(BUGGY_LL_D2F) || defined(BUGGY_LL_F2D)
        1
   #else
        0
   #endif
        )
       debugp(stderr, "relocs: %d:%d\n", relocs, nonrelocs);
       fprintf(stderr, "*** performance problems ***\n%s"
   #if defined(BUGGY_LL_CMP) || defined(BUGGY_LL_MUL) || defined(BUGGY_LL_DIV) || defined(BUGGY_LL_ADD) || defined(BUGGY_LL_SHIFT) || defined(BUGGY_LL_D2F) || defined(BUGGY_LL_F2D)
               "    double-cell integer type buggy ->\n        "
   #ifdef BUGGY_LL_CMP
               "CMP, "
   #endif
   #ifdef BUGGY_LL_MUL
               "MUL, "
   #endif
   #ifdef BUGGY_LL_DIV
               "DIV, "
   #endif
   #ifdef BUGGY_LL_ADD
               "ADD, "
   #endif
   #ifdef BUGGY_LL_SHIFT
               "SHIFT, "
   #endif
   #ifdef BUGGY_LL_D2F
               "D2F, "
   #endif
   #ifdef BUGGY_LL_F2D
               "F2D, "
   #endif
               "\b\b slow\n"
   #endif
   #ifndef FORCE_REG
               "    automatic register allocation: performance degradation possible\n"
   #endif
   #if !defined(FORCE_REG) || defined(BUGGY_LONG_LONG)
               "*** Suggested remedy: try ./configure"
   #ifndef FORCE_REG
               " --enable-force-reg"
   #endif
   #ifdef BUGGY_LONG_LONG
               " --enable-force-ll"
   #endif
               "\n"
   #endif
               ,
               (relocs < nonrelocs) ? "    gcc PR 15242 -> no dynamic code generation (use gcc-2.95 instead)\n" : "");
   }
   
 #ifdef INCLUDE_IMAGE  #ifdef INCLUDE_IMAGE
 extern Cell image[];  extern Cell image[];
 extern const char reloc_bits[];  extern const char reloc_bits[];
Line 829 
Line 1862 
 #endif  #endif
   int retvalue;    int retvalue;
   
 #if defined(i386) && defined(ALIGNMENT_CHECK) && !defined(DIRECT_THREADED)  #if defined(i386) && defined(ALIGNMENT_CHECK)
   /* turn on alignment checks on the 486.    /* turn on alignment checks on the 486.
    * on the 386 this should have no effect. */     * on the 386 this should have no effect. */
   __asm__("pushfl; popl %eax; orl $0x40000, %eax; pushl %eax; popfl;");    __asm__("pushfl; popl %eax; orl $0x40000, %eax; pushl %eax; popfl;");
Line 851 
Line 1884 
   
 #ifdef HAS_OS  #ifdef HAS_OS
   gforth_args(argc, argv, &path, &imagename);    gforth_args(argc, argv, &path, &imagename);
 #endif  #ifndef NO_DYNAMIC
     if (no_dynamic && ss_cost == cost_codesize) {
       ss_cost = cost_nexts;
       cost_sums[0] = cost_sums[1]; /* don't use cost_codesize for print-metrics */
       debugp(stderr, "--no-dynamic conflicts with --ss-min-codesize, reverting to --ss-min-nexts\n");
     }
   #endif /* !defined(NO_DYNAMIC) */
   #endif /* defined(HAS_OS) */
   
 #ifdef INCLUDE_IMAGE  #ifdef INCLUDE_IMAGE
   set_stack_sizes((ImageHeader *)image);    set_stack_sizes((ImageHeader *)image);
Line 865 
Line 1905 
 #endif  #endif
   gforth_header=(ImageHeader *)image; /* used in SIGSEGV handler */    gforth_header=(ImageHeader *)image; /* used in SIGSEGV handler */
   
     if (diag)
       print_diag();
   {    {
     char path2[strlen(path)+1];      char path2[strlen(path)+1];
     char *p1, *p2;      char *p1, *p2;
Line 886 
Line 1928 
         *p2 = *p1;          *p2 = *p1;
     *p2='\0';      *p2='\0';
     retvalue = go_forth(image, 4, environ);      retvalue = go_forth(image, 4, environ);
   #ifdef SIGPIPE
       bsd_signal(SIGPIPE, SIG_IGN);
   #endif
 #ifdef VM_PROFILING  #ifdef VM_PROFILING
     vm_print_profile(stderr);      vm_print_profile(stderr);
 #endif  #endif
     deprep_terminal();      deprep_terminal();
   }    }
     if (print_metrics) {
       int i;
       fprintf(stderr, "code size = %8ld\n", dyncodesize());
       for (i=0; i<sizeof(cost_sums)/sizeof(cost_sums[0]); i++)
         fprintf(stderr, "metric %8s: %8ld\n",
                 cost_sums[i].metricname, cost_sums[i].sum);
     }
   return retvalue;    return retvalue;
 }  }

Generate output suitable for use with a patch program
Legend:
Removed from v.1.48  
changed lines
  Added in v.1.145
CVS Admin

Powered by ViewCVS 1.0-dev 		(Powered by ViewCVS)

ViewCVS and CVS Help