Annotation of gforth/engine/main.c, revision 1.171

1.1       anton       1: /* command line interpretation, image loading etc. for Gforth
                      2: 
                      3: 
1.170     anton       4:   Copyright (C) 1995,1996,1997,1998,2000,2003,2004,2005,2006 Free Software Foundation, Inc.
1.1       anton       5: 
                      6:   This file is part of Gforth.
                      7: 
                      8:   Gforth is free software; you can redistribute it and/or
                      9:   modify it under the terms of the GNU General Public License
                     10:   as published by the Free Software Foundation; either version 2
                     11:   of the License, or (at your option) any later version.
                     12: 
                     13:   This program is distributed in the hope that it will be useful,
                     14:   but WITHOUT ANY WARRANTY; without even the implied warranty of
                     15:   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
                     16:   GNU General Public License for more details.
                     17: 
                     18:   You should have received a copy of the GNU General Public License
                     19:   along with this program; if not, write to the Free Software
1.40      anton      20:   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111, USA.
1.1       anton      21: */
                     22: 
                     23: #include "config.h"
1.82      anton      24: #include "forth.h"
1.1       anton      25: #include <errno.h>
                     26: #include <ctype.h>
                     27: #include <stdio.h>
1.2       pazsan     28: #include <unistd.h>
1.1       anton      29: #include <string.h>
                     30: #include <math.h>
                     31: #include <sys/types.h>
1.32      pazsan     32: #ifndef STANDALONE
1.1       anton      33: #include <sys/stat.h>
1.32      pazsan     34: #endif
1.1       anton      35: #include <fcntl.h>
                     36: #include <assert.h>
                     37: #include <stdlib.h>
1.102     anton      38: #include <signal.h>
1.11      pazsan     39: #ifndef STANDALONE
1.1       anton      40: #if HAVE_SYS_MMAN_H
                     41: #include <sys/mman.h>
                     42: #endif
1.11      pazsan     43: #endif
1.1       anton      44: #include "io.h"
                     45: #include "getopt.h"
1.11      pazsan     46: #ifdef STANDALONE
                     47: #include <systypes.h>
                     48: #endif
1.1       anton      49: 
1.121     anton      50: typedef enum prim_num {
1.119     anton      51: /* definitions of N_execute etc. */
1.126     anton      52: #include PRIM_NUM_I
1.119     anton      53:   N_START_SUPER
1.121     anton      54: } PrimNum;
1.119     anton      55: 
1.79      anton      56: /* global variables for engine.c 
                     57:    We put them here because engine.c is compiled several times in
                     58:    different ways for the same engine. */
1.161     pazsan     59: Cell *gforth_SP;
                     60: Float *gforth_FP;
                     61: Address gforth_UP=NULL;
1.79      anton      62: 
1.115     pazsan     63: #ifdef HAS_FFCALL
1.161     pazsan     64: Cell *gforth_RP;
                     65: Address gforth_LP;
1.115     pazsan     66: 
                     67: #include <callback.h>
                     68: 
1.161     pazsan     69: va_alist gforth_clist;
1.115     pazsan     70: 
1.161     pazsan     71: void gforth_callback(Xt* fcall, void * alist)
1.115     pazsan     72: {
1.140     pazsan     73:   /* save global valiables */
1.161     pazsan     74:   Cell *rp = gforth_RP;
                     75:   Cell *sp = gforth_SP;
                     76:   Float *fp = gforth_FP;
                     77:   Address lp = gforth_LP;
1.168     pazsan     78:   va_alist clist = gforth_clist;
1.140     pazsan     79: 
1.161     pazsan     80:   gforth_clist = (va_alist)alist;
1.140     pazsan     81: 
1.161     pazsan     82:   gforth_engine(fcall, sp, rp, fp, lp);
1.140     pazsan     83: 
                     84:   /* restore global variables */
1.161     pazsan     85:   gforth_RP = rp;
                     86:   gforth_SP = sp;
                     87:   gforth_FP = fp;
                     88:   gforth_LP = lp;
1.168     pazsan     89:   gforth_clist = clist;
1.115     pazsan     90: }
                     91: #endif
                     92: 
1.153     pazsan     93: #ifdef HAS_LIBFFI
1.161     pazsan     94: Cell *gforth_RP;
                     95: Address gforth_LP;
1.153     pazsan     96: 
                     97: #include <ffi.h>
                     98: 
1.164     pazsan     99: void ** gforth_clist;
                    100: void * gforth_ritem;
1.153     pazsan    101: 
1.162     pazsan    102: void gforth_callback(ffi_cif * cif, void * resp, void ** args, void * ip)
1.153     pazsan    103: {
1.161     pazsan    104:   Cell *rp = gforth_RP;
                    105:   Cell *sp = gforth_SP;
                    106:   Float *fp = gforth_FP;
                    107:   Address lp = gforth_LP;
1.168     pazsan    108:   void ** clist = gforth_clist;
                    109:   void * ritem = gforth_ritem;
1.153     pazsan    110: 
1.164     pazsan    111:   gforth_clist = args;
                    112:   gforth_ritem = resp;
1.153     pazsan    113: 
1.164     pazsan    114:   gforth_engine((Xt *)ip, sp, rp, fp, lp);
1.153     pazsan    115: 
                    116:   /* restore global variables */
1.161     pazsan    117:   gforth_RP = rp;
                    118:   gforth_SP = sp;
                    119:   gforth_FP = fp;
                    120:   gforth_LP = lp;
1.168     pazsan    121:   gforth_clist = clist;
                    122:   gforth_ritem = ritem;
1.153     pazsan    123: }
                    124: #endif
                    125: 
1.79      anton     126: #ifdef GFORTH_DEBUGGING
                    127: /* define some VM registers as global variables, so they survive exceptions;
                    128:    global register variables are not up to the task (according to the 
                    129:    GNU C manual) */
                    130: Xt *saved_ip;
                    131: Cell *rp;
                    132: #endif
                    133: 
                    134: #ifdef NO_IP
                    135: Label next_code;
                    136: #endif
                    137: 
                    138: #ifdef HAS_FILE
                    139: char* fileattr[6]={"rb","rb","r+b","r+b","wb","wb"};
                    140: char* pfileattr[6]={"r","r","r+","r+","w","w"};
                    141: 
                    142: #ifndef O_BINARY
                    143: #define O_BINARY 0
                    144: #endif
                    145: #ifndef O_TEXT
                    146: #define O_TEXT 0
                    147: #endif
                    148: 
                    149: int ufileattr[6]= {
                    150:   O_RDONLY|O_BINARY, O_RDONLY|O_BINARY,
                    151:   O_RDWR  |O_BINARY, O_RDWR  |O_BINARY,
                    152:   O_WRONLY|O_BINARY, O_WRONLY|O_BINARY };
                    153: #endif
                    154: /* end global vars for engine.c */
                    155: 
1.1       anton     156: #define PRIM_VERSION 1
                    157: /* increment this whenever the primitives change in an incompatible way */
                    158: 
1.14      pazsan    159: #ifndef DEFAULTPATH
1.39      anton     160: #  define DEFAULTPATH "."
1.14      pazsan    161: #endif
                    162: 
1.1       anton     163: #ifdef MSDOS
                    164: jmp_buf throw_jmp_buf;
                    165: #endif
                    166: 
1.56      anton     167: #if defined(DOUBLY_INDIRECT)
                    168: #  define CFA(n)       ({Cell _n = (n); ((Cell)(((_n & 0x4000) ? symbols : xts)+(_n&~0x4000UL)));})
1.1       anton     169: #else
1.56      anton     170: #  define CFA(n)       ((Cell)(symbols+((n)&~0x4000UL)))
1.1       anton     171: #endif
                    172: 
                    173: #define maxaligned(n)  (typeof(n))((((Cell)n)+sizeof(Float)-1)&-sizeof(Float))
                    174: 
                    175: static UCell dictsize=0;
                    176: static UCell dsize=0;
                    177: static UCell rsize=0;
                    178: static UCell fsize=0;
                    179: static UCell lsize=0;
                    180: int offset_image=0;
1.4       anton     181: int die_on_signal=0;
1.169     anton     182: int ignore_async_signals=0;
1.13      pazsan    183: #ifndef INCLUDE_IMAGE
1.1       anton     184: static int clear_dictionary=0;
1.24      anton     185: UCell pagesize=1;
1.22      pazsan    186: char *progname;
                    187: #else
                    188: char *progname = "gforth";
                    189: int optind = 1;
1.13      pazsan    190: #endif
1.31      pazsan    191: 
1.167     anton     192: #define CODE_BLOCK_SIZE (512*1024) /* !! overflow handling for -native */
1.48      anton     193: Address code_area=0;
1.73      anton     194: Cell code_area_size = CODE_BLOCK_SIZE;
1.75      anton     195: Address code_here=NULL+CODE_BLOCK_SIZE; /* does for code-area what HERE
                    196:                                           does for the dictionary */
1.100     anton     197: Address start_flush=NULL; /* start of unflushed code */
1.74      anton     198: Cell last_jump=0; /* if the last prim was compiled without jump, this
                    199:                      is it's number, otherwise this contains 0 */
1.48      anton     200: 
1.60      anton     201: static int no_super=0;   /* true if compile_prim should not fuse prims */
1.81      anton     202: static int no_dynamic=NO_DYNAMIC_DEFAULT; /* if true, no code is generated
                    203:                                             dynamically */
1.110     anton     204: static int print_metrics=0; /* if true, print metrics on exit */
1.171   ! anton     205: static int static_super_number = 0; /* number of ss used if available */
        !           206:                                     /* disabled because of tpa */
1.152     anton     207: #define MAX_STATE 9 /* maximum number of states */
1.125     anton     208: static int maxstates = MAX_STATE; /* number of states for stack caching */
1.110     anton     209: static int ss_greedy = 0; /* if true: use greedy, not optimal ss selection */
1.144     pazsan    210: static int diag = 0; /* if true: print diagnostic informations */
1.158     anton     211: static int tpa_noequiv = 0;     /* if true: no state equivalence checking */
                    212: static int tpa_noautomaton = 0; /* if true: no tree parsing automaton */
                    213: static int tpa_trace = 0; /* if true: data for line graph of new states etc. */
1.144     pazsan    214: static int relocs = 0;
                    215: static int nonrelocs = 0;
1.60      anton     216: 
1.30      pazsan    217: #ifdef HAS_DEBUG
1.68      anton     218: int debug=0;
1.144     pazsan    219: # define debugp(x...) if (debug) fprintf(x);
1.31      pazsan    220: #else
                    221: # define perror(x...)
                    222: # define fprintf(x...)
1.144     pazsan    223: # define debugp(x...)
1.30      pazsan    224: #endif
1.31      pazsan    225: 
1.24      anton     226: ImageHeader *gforth_header;
1.43      anton     227: Label *vm_prims;
1.53      anton     228: #ifdef DOUBLY_INDIRECT
                    229: Label *xts; /* same content as vm_prims, but should only be used for xts */
                    230: #endif
1.1       anton     231: 
1.125     anton     232: #ifndef NO_DYNAMIC
                    233: #define MAX_IMMARGS 2
                    234: 
                    235: typedef struct {
                    236:   Label start; /* NULL if not relocatable */
                    237:   Cell length; /* only includes the jump iff superend is true*/
                    238:   Cell restlength; /* length of the rest (i.e., the jump or (on superend) 0) */
                    239:   char superend; /* true if primitive ends superinstruction, i.e.,
                    240:                      unconditional branch, execute, etc. */
                    241:   Cell nimmargs;
                    242:   struct immarg {
                    243:     Cell offset; /* offset of immarg within prim */
                    244:     char rel;    /* true if immarg is relative */
                    245:   } immargs[MAX_IMMARGS];
                    246: } PrimInfo;
                    247: 
                    248: PrimInfo *priminfos;
                    249: PrimInfo **decomp_prims;
                    250: 
1.139     anton     251: const char const* const prim_names[]={
                    252: #include PRIM_NAMES_I
                    253: };
                    254: 
1.148     anton     255: void init_ss_cost(void);
                    256: 
1.125     anton     257: static int is_relocatable(int p)
                    258: {
                    259:   return !no_dynamic && priminfos[p].start != NULL;
                    260: }
                    261: #else /* defined(NO_DYNAMIC) */
                    262: static int is_relocatable(int p)
                    263: {
                    264:   return 0;
                    265: }
                    266: #endif /* defined(NO_DYNAMIC) */
                    267: 
1.30      pazsan    268: #ifdef MEMCMP_AS_SUBROUTINE
                    269: int gforth_memcmp(const char * s1, const char * s2, size_t n)
                    270: {
                    271:   return memcmp(s1, s2, n);
                    272: }
                    273: #endif
                    274: 
1.125     anton     275: static Cell max(Cell a, Cell b)
                    276: {
                    277:   return a>b?a:b;
                    278: }
                    279: 
                    280: static Cell min(Cell a, Cell b)
                    281: {
                    282:   return a<b?a:b;
                    283: }
                    284: 
1.1       anton     285: /* image file format:
1.15      pazsan    286:  *  "#! binary-path -i\n" (e.g., "#! /usr/local/bin/gforth-0.4.0 -i\n")
1.1       anton     287:  *   padding to a multiple of 8
1.84      anton     288:  *   magic: "Gforth3x" means format 0.6,
1.15      pazsan    289:  *              where x is a byte with
                    290:  *              bit 7:   reserved = 0
                    291:  *              bit 6:5: address unit size 2^n octets
                    292:  *              bit 4:3: character size 2^n octets
                    293:  *              bit 2:1: cell size 2^n octets
                    294:  *              bit 0:   endian, big=0, little=1.
                    295:  *  The magic are always 8 octets, no matter what the native AU/character size is
1.1       anton     296:  *  padding to max alignment (no padding necessary on current machines)
1.24      anton     297:  *  ImageHeader structure (see forth.h)
1.1       anton     298:  *  data (size in ImageHeader.image_size)
                    299:  *  tags ((if relocatable, 1 bit/data cell)
                    300:  *
                    301:  * tag==1 means that the corresponding word is an address;
                    302:  * If the word is >=0, the address is within the image;
                    303:  * addresses within the image are given relative to the start of the image.
                    304:  * If the word =-1 (CF_NIL), the address is NIL,
                    305:  * If the word is <CF_NIL and >CF(DODOES), it's a CFA (:, Create, ...)
                    306:  * If the word =CF(DODOES), it's a DOES> CFA
                    307:  * If the word =CF(DOESJUMP), it's a DOES JUMP (2 Cells after DOES>,
                    308:  *                                     possibly containing a jump to dodoes)
1.51      anton     309:  * If the word is <CF(DOESJUMP) and bit 14 is set, it's the xt of a primitive
                    310:  * If the word is <CF(DOESJUMP) and bit 14 is clear, 
                    311:  *                                        it's the threaded code of a primitive
1.85      pazsan    312:  * bits 13..9 of a primitive token state which group the primitive belongs to,
                    313:  * bits 8..0 of a primitive token index into the group
1.1       anton     314:  */
                    315: 
1.115     pazsan    316: Cell groups[32] = {
1.85      pazsan    317:   0,
1.121     anton     318:   0
1.90      anton     319: #undef GROUP
1.115     pazsan    320: #undef GROUPADD
                    321: #define GROUPADD(n) +n
                    322: #define GROUP(x, n) , 0
1.126     anton     323: #include PRIM_GRP_I
1.90      anton     324: #undef GROUP
1.115     pazsan    325: #undef GROUPADD
1.85      pazsan    326: #define GROUP(x, n)
1.115     pazsan    327: #define GROUPADD(n)
1.85      pazsan    328: };
                    329: 
1.161     pazsan    330: static unsigned char *branch_targets(Cell *image, const unsigned char *bitstring,
1.125     anton     331:                              int size, Cell base)
                    332:      /* produce a bitmask marking all the branch targets */
                    333: {
1.130     anton     334:   int i=0, j, k, steps=(((size-1)/sizeof(Cell))/RELINFOBITS)+1;
1.125     anton     335:   Cell token;
                    336:   unsigned char bits;
1.130     anton     337:   unsigned char *result=malloc(steps);
                    338: 
                    339:   memset(result, 0, steps);
                    340:   for(k=0; k<steps; k++) {
1.125     anton     341:     for(j=0, bits=bitstring[k]; j<RELINFOBITS; j++, i++, bits<<=1) {
1.130     anton     342:       if(bits & (1U << (RELINFOBITS-1))) {
                    343:        assert(i*sizeof(Cell) < size);
1.125     anton     344:         token=image[i];
                    345:        if (token>=base) { /* relocatable address */
                    346:          UCell bitnum=(token-base)/sizeof(Cell);
1.154     anton     347:          if (bitnum/RELINFOBITS < (UCell)steps)
                    348:            result[bitnum/RELINFOBITS] |= 1U << ((~bitnum)&(RELINFOBITS-1));
1.125     anton     349:        }
                    350:       }
                    351:     }
                    352:   }
                    353:   return result;
                    354: }
                    355: 
1.162     pazsan    356: void gforth_relocate(Cell *image, const Char *bitstring, 
                    357:                     UCell size, Cell base, Label symbols[])
1.1       anton     358: {
1.130     anton     359:   int i=0, j, k, steps=(((size-1)/sizeof(Cell))/RELINFOBITS)+1;
1.11      pazsan    360:   Cell token;
1.1       anton     361:   char bits;
1.37      anton     362:   Cell max_symbols;
1.46      jwilke    363:   /* 
1.85      pazsan    364:    * A virtual start address that's the real start address minus 
1.46      jwilke    365:    * the one in the image 
                    366:    */
1.45      jwilke    367:   Cell *start = (Cell * ) (((void *) image) - ((void *) base));
1.125     anton     368:   unsigned char *targets = branch_targets(image, bitstring, size, base);
1.1       anton     369: 
1.85      pazsan    370:   /* group index into table */
1.115     pazsan    371:   if(groups[31]==0) {
                    372:     int groupsum=0;
                    373:     for(i=0; i<32; i++) {
                    374:       groupsum += groups[i];
                    375:       groups[i] = groupsum;
                    376:       /* printf("group[%d]=%d\n",i,groupsum); */
                    377:     }
                    378:     i=0;
                    379:   }
1.46      jwilke    380:   
                    381: /* printf("relocating to %x[%x] start=%x base=%x\n", image, size, start, base); */
1.37      anton     382:   
1.121     anton     383:   for (max_symbols=0; symbols[max_symbols]!=0; max_symbols++)
1.37      anton     384:     ;
1.47      anton     385:   max_symbols--;
1.35      pazsan    386: 
1.130     anton     387:   for(k=0; k<steps; k++) {
1.13      pazsan    388:     for(j=0, bits=bitstring[k]; j<RELINFOBITS; j++, i++, bits<<=1) {
1.1       anton     389:       /*      fprintf(stderr,"relocate: image[%d]\n", i);*/
1.130     anton     390:       if(bits & (1U << (RELINFOBITS-1))) {
                    391:        assert(i*sizeof(Cell) < size);
1.35      pazsan    392:        /* fprintf(stderr,"relocate: image[%d]=%d of %d\n", i, image[i], size/sizeof(Cell)); */
1.45      jwilke    393:         token=image[i];
1.85      pazsan    394:        if(token<0) {
                    395:          int group = (-token & 0x3E00) >> 9;
                    396:          if(group == 0) {
                    397:            switch(token|0x4000) {
1.1       anton     398:            case CF_NIL      : image[i]=0; break;
                    399: #if !defined(DOUBLY_INDIRECT)
                    400:            case CF(DOCOL)   :
                    401:            case CF(DOVAR)   :
                    402:            case CF(DOCON)   :
                    403:            case CF(DOUSER)  : 
                    404:            case CF(DODEFER) : 
1.11      pazsan    405:            case CF(DOFIELD) : MAKE_CF(image+i,symbols[CF(token)]); break;
1.92      anton     406:            case CF(DOESJUMP): image[i]=0; break;
1.1       anton     407: #endif /* !defined(DOUBLY_INDIRECT) */
                    408:            case CF(DODOES)  :
1.45      jwilke    409:              MAKE_DOES_CF(image+i,(Xt *)(image[i+1]+((Cell)start)));
1.1       anton     410:              break;
1.85      pazsan    411:            default          : /* backward compatibility */
1.56      anton     412: /*           printf("Code field generation image[%x]:=CFA(%x)\n",
1.1       anton     413:                     i, CF(image[i])); */
1.55      anton     414:              if (CF((token | 0x4000))<max_symbols) {
1.56      anton     415:                image[i]=(Cell)CFA(CF(token));
                    416: #ifdef DIRECT_THREADED
1.125     anton     417:                if ((token & 0x4000) == 0) { /* threade code, no CFA */
                    418:                  if (targets[k] & (1U<<(RELINFOBITS-1-j)))
                    419:                    compile_prim1(0);
1.70      anton     420:                  compile_prim1(&image[i]);
1.125     anton     421:                }
1.56      anton     422: #endif
1.55      anton     423:              } else
1.115     pazsan    424:                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);
1.1       anton     425:            }
1.85      pazsan    426:          } else {
                    427:            int tok = -token & 0x1FF;
                    428:            if (tok < (groups[group+1]-groups[group])) {
                    429: #if defined(DOUBLY_INDIRECT)
                    430:              image[i]=(Cell)CFA(((groups[group]+tok) | (CF(token) & 0x4000)));
                    431: #else
                    432:              image[i]=(Cell)CFA((groups[group]+tok));
                    433: #endif
                    434: #ifdef DIRECT_THREADED
1.125     anton     435:              if ((token & 0x4000) == 0) { /* threade code, no CFA */
                    436:                if (targets[k] & (1U<<(RELINFOBITS-1-j)))
                    437:                  compile_prim1(0);
1.85      pazsan    438:                compile_prim1(&image[i]);
1.125     anton     439:              }
1.85      pazsan    440: #endif
                    441:            } else
1.115     pazsan    442:              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);
1.85      pazsan    443:          }
                    444:        } else {
1.101     anton     445:           /* if base is > 0: 0 is a null reference so don't adjust*/
1.45      jwilke    446:           if (token>=base) {
                    447:             image[i]+=(Cell)start;
                    448:           }
1.46      jwilke    449:         }
1.1       anton     450:       }
                    451:     }
1.31      pazsan    452:   }
1.125     anton     453:   free(targets);
1.70      anton     454:   finish_code();
1.26      jwilke    455:   ((ImageHeader*)(image))->base = (Address) image;
1.1       anton     456: }
                    457: 
1.162     pazsan    458: #ifndef DOUBLY_INDIRECT
1.161     pazsan    459: static UCell checksum(Label symbols[])
1.1       anton     460: {
                    461:   UCell r=PRIM_VERSION;
                    462:   Cell i;
                    463: 
                    464:   for (i=DOCOL; i<=DOESJUMP; i++) {
                    465:     r ^= (UCell)(symbols[i]);
                    466:     r = (r << 5) | (r >> (8*sizeof(Cell)-5));
                    467:   }
                    468: #ifdef DIRECT_THREADED
                    469:   /* we have to consider all the primitives */
                    470:   for (; symbols[i]!=(Label)0; i++) {
                    471:     r ^= (UCell)(symbols[i]);
                    472:     r = (r << 5) | (r >> (8*sizeof(Cell)-5));
                    473:   }
                    474: #else
                    475:   /* in indirect threaded code all primitives are accessed through the
                    476:      symbols table, so we just have to put the base address of symbols
                    477:      in the checksum */
                    478:   r ^= (UCell)symbols;
                    479: #endif
                    480:   return r;
                    481: }
1.162     pazsan    482: #endif
1.1       anton     483: 
1.161     pazsan    484: static Address verbose_malloc(Cell size)
1.3       anton     485: {
                    486:   Address r;
                    487:   /* leave a little room (64B) for stack underflows */
                    488:   if ((r = malloc(size+64))==NULL) {
                    489:     perror(progname);
                    490:     exit(1);
                    491:   }
                    492:   r = (Address)((((Cell)r)+(sizeof(Float)-1))&(-sizeof(Float)));
1.144     pazsan    493:   debugp(stderr, "malloc succeeds, address=$%lx\n", (long)r);
1.3       anton     494:   return r;
                    495: }
                    496: 
1.33      anton     497: static Address next_address=0;
1.161     pazsan    498: static void after_alloc(Address r, Cell size)
1.33      anton     499: {
                    500:   if (r != (Address)-1) {
1.144     pazsan    501:     debugp(stderr, "success, address=$%lx\n", (long) r);
1.33      anton     502:     if (pagesize != 1)
                    503:       next_address = (Address)(((((Cell)r)+size-1)&-pagesize)+2*pagesize); /* leave one page unmapped */
                    504:   } else {
1.144     pazsan    505:     debugp(stderr, "failed: %s\n", strerror(errno));
1.33      anton     506:   }
                    507: }
                    508: 
1.34      anton     509: #ifndef MAP_FAILED
                    510: #define MAP_FAILED ((Address) -1)
                    511: #endif
                    512: #ifndef MAP_FILE
                    513: # define MAP_FILE 0
                    514: #endif
                    515: #ifndef MAP_PRIVATE
                    516: # define MAP_PRIVATE 0
                    517: #endif
1.91      anton     518: #if !defined(MAP_ANON) && defined(MAP_ANONYMOUS)
                    519: # define MAP_ANON MAP_ANONYMOUS
                    520: #endif
1.34      anton     521: 
                    522: #if defined(HAVE_MMAP)
                    523: static Address alloc_mmap(Cell size)
1.1       anton     524: {
                    525:   Address r;
                    526: 
                    527: #if defined(MAP_ANON)
1.144     pazsan    528:   debugp(stderr,"try mmap($%lx, $%lx, ..., MAP_ANON, ...); ", (long)next_address, (long)size);
1.34      anton     529:   r = mmap(next_address, size, PROT_EXEC|PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE, -1, 0);
1.1       anton     530: #else /* !defined(MAP_ANON) */
1.17      anton     531:   /* Ultrix (at least) does not define MAP_FILE and MAP_PRIVATE (both are
                    532:      apparently defaults) */
1.1       anton     533:   static int dev_zero=-1;
                    534: 
                    535:   if (dev_zero == -1)
                    536:     dev_zero = open("/dev/zero", O_RDONLY);
                    537:   if (dev_zero == -1) {
1.34      anton     538:     r = MAP_FAILED;
1.144     pazsan    539:     debugp(stderr, "open(\"/dev/zero\"...) failed (%s), no mmap; ", 
1.1       anton     540:              strerror(errno));
                    541:   } else {
1.144     pazsan    542:     debugp(stderr,"try mmap($%lx, $%lx, ..., MAP_FILE, dev_zero, ...); ", (long)next_address, (long)size);
1.1       anton     543:     r=mmap(next_address, size, PROT_EXEC|PROT_READ|PROT_WRITE, MAP_FILE|MAP_PRIVATE, dev_zero, 0);
                    544:   }
                    545: #endif /* !defined(MAP_ANON) */
1.34      anton     546:   after_alloc(r, size);
                    547:   return r;  
                    548: }
                    549: #endif
                    550: 
1.161     pazsan    551: Address gforth_alloc(Cell size)
1.34      anton     552: {
                    553: #if HAVE_MMAP
                    554:   Address r;
                    555: 
                    556:   r=alloc_mmap(size);
1.117     anton     557:   if (r!=(Address)MAP_FAILED)
1.1       anton     558:     return r;
                    559: #endif /* HAVE_MMAP */
1.3       anton     560:   /* use malloc as fallback */
                    561:   return verbose_malloc(size);
1.1       anton     562: }
                    563: 
1.161     pazsan    564: static Address dict_alloc_read(FILE *file, Cell imagesize, Cell dictsize, Cell offset)
1.33      anton     565: {
1.34      anton     566:   Address image = MAP_FAILED;
1.33      anton     567: 
1.56      anton     568: #if defined(HAVE_MMAP)
1.33      anton     569:   if (offset==0) {
1.34      anton     570:     image=alloc_mmap(dictsize);
1.150     anton     571:     if (image != (Address)MAP_FAILED) {
                    572:       Address image1;
                    573:       debugp(stderr,"try mmap($%lx, $%lx, ..., MAP_FIXED|MAP_FILE, imagefile, 0); ", (long)image, (long)imagesize);
                    574:       image1 = mmap(image, imagesize, PROT_EXEC|PROT_READ|PROT_WRITE, MAP_FIXED|MAP_FILE|MAP_PRIVATE, fileno(file), 0);
                    575:       after_alloc(image1,dictsize);
                    576:       if (image1 == (Address)MAP_FAILED)
                    577:        goto read_image;
                    578:     }
1.33      anton     579:   }
1.56      anton     580: #endif /* defined(HAVE_MMAP) */
1.117     anton     581:   if (image == (Address)MAP_FAILED) {
1.161     pazsan    582:     image = gforth_alloc(dictsize+offset)+offset;
1.149     anton     583:   read_image:
1.33      anton     584:     rewind(file);  /* fseek(imagefile,0L,SEEK_SET); */
1.34      anton     585:     fread(image, 1, imagesize, file);
1.33      anton     586:   }
                    587:   return image;
                    588: }
                    589: 
1.10      pazsan    590: void set_stack_sizes(ImageHeader * header)
                    591: {
                    592:   if (dictsize==0)
                    593:     dictsize = header->dict_size;
                    594:   if (dsize==0)
                    595:     dsize = header->data_stack_size;
                    596:   if (rsize==0)
                    597:     rsize = header->return_stack_size;
                    598:   if (fsize==0)
                    599:     fsize = header->fp_stack_size;
                    600:   if (lsize==0)
                    601:     lsize = header->locals_stack_size;
                    602:   dictsize=maxaligned(dictsize);
                    603:   dsize=maxaligned(dsize);
                    604:   rsize=maxaligned(rsize);
                    605:   lsize=maxaligned(lsize);
                    606:   fsize=maxaligned(fsize);
                    607: }
                    608: 
                    609: void alloc_stacks(ImageHeader * header)
                    610: {
                    611:   header->dict_size=dictsize;
                    612:   header->data_stack_size=dsize;
                    613:   header->fp_stack_size=fsize;
                    614:   header->return_stack_size=rsize;
                    615:   header->locals_stack_size=lsize;
                    616: 
1.161     pazsan    617:   header->data_stack_base=gforth_alloc(dsize);
                    618:   header->fp_stack_base=gforth_alloc(fsize);
                    619:   header->return_stack_base=gforth_alloc(rsize);
                    620:   header->locals_stack_base=gforth_alloc(lsize);
1.10      pazsan    621: }
                    622: 
1.161     pazsan    623: #warning You can ignore the warnings about clobbered variables in gforth_go
                    624: int gforth_go(Address image, int stack, Cell *entries)
1.11      pazsan    625: {
1.38      anton     626:   volatile ImageHeader *image_header = (ImageHeader *)image;
1.18      anton     627:   Cell *sp0=(Cell*)(image_header->data_stack_base + dsize);
1.44      pazsan    628:   Cell *rp0=(Cell *)(image_header->return_stack_base + rsize);
1.18      anton     629:   Float *fp0=(Float *)(image_header->fp_stack_base + fsize);
1.44      pazsan    630: #ifdef GFORTH_DEBUGGING
1.38      anton     631:   volatile Cell *orig_rp0=rp0;
1.44      pazsan    632: #endif
1.18      anton     633:   Address lp0=image_header->locals_stack_base + lsize;
                    634:   Xt *ip0=(Xt *)(image_header->boot_entry);
1.13      pazsan    635: #ifdef SYSSIGNALS
1.11      pazsan    636:   int throw_code;
1.13      pazsan    637: #endif
1.11      pazsan    638: 
                    639:   /* ensure that the cached elements (if any) are accessible */
1.151     anton     640: #if !(defined(GFORTH_DEBUGGING) || defined(INDIRECT_THREADED) || defined(DOUBLY_INDIRECT) || defined(VM_PROFILING))
                    641:   sp0 -= 8; /* make stuff below bottom accessible for stack caching */
                    642: #endif
1.41      anton     643:   IF_fpTOS(fp0--);
1.11      pazsan    644:   
                    645:   for(;stack>0;stack--)
1.18      anton     646:     *--sp0=entries[stack-1];
1.11      pazsan    647: 
1.30      pazsan    648: #ifdef SYSSIGNALS
1.11      pazsan    649:   get_winsize();
                    650:    
                    651:   install_signal_handlers(); /* right place? */
                    652:   
                    653:   if ((throw_code=setjmp(throw_jmp_buf))) {
1.152     anton     654:     static Cell signal_data_stack[24];
                    655:     static Cell signal_return_stack[16];
1.11      pazsan    656:     static Float signal_fp_stack[1];
1.13      pazsan    657: 
1.152     anton     658:     signal_data_stack[15]=throw_code;
1.18      anton     659: 
                    660: #ifdef GFORTH_DEBUGGING
1.144     pazsan    661:     debugp(stderr,"\ncaught signal, throwing exception %d, ip=%p rp=%p\n",
1.97      anton     662:              throw_code, saved_ip, rp);
1.38      anton     663:     if (rp <= orig_rp0 && rp > (Cell *)(image_header->return_stack_base+5)) {
1.18      anton     664:       /* no rstack overflow or underflow */
                    665:       rp0 = rp;
1.63      anton     666:       *--rp0 = (Cell)saved_ip;
1.18      anton     667:     }
                    668:     else /* I love non-syntactic ifdefs :-) */
1.152     anton     669:       rp0 = signal_return_stack+16;
1.97      anton     670: #else  /* !defined(GFORTH_DEBUGGING) */
1.144     pazsan    671:     debugp(stderr,"\ncaught signal, throwing exception %d\n", throw_code);
1.152     anton     672:       rp0 = signal_return_stack+16;
1.97      anton     673: #endif /* !defined(GFORTH_DEBUGGING) */
1.25      anton     674:     /* fprintf(stderr, "rp=$%x\n",rp0);*/
1.11      pazsan    675:     
1.164     pazsan    676:     return((int)(Cell)gforth_engine(image_header->throw_entry, signal_data_stack+15,
1.18      anton     677:                       rp0, signal_fp_stack, 0));
1.11      pazsan    678:   }
1.13      pazsan    679: #endif
1.11      pazsan    680: 
1.164     pazsan    681:   return((int)(Cell)gforth_engine(ip0,sp0,rp0,fp0,lp0));
1.11      pazsan    682: }
                    683: 
1.30      pazsan    684: #ifndef INCLUDE_IMAGE
1.161     pazsan    685: static void print_sizes(Cell sizebyte)
1.21      anton     686:      /* print size information */
                    687: {
                    688:   static char* endianstring[]= { "   big","little" };
                    689:   
                    690:   fprintf(stderr,"%s endian, cell=%d bytes, char=%d bytes, au=%d bytes\n",
                    691:          endianstring[sizebyte & 1],
                    692:          1 << ((sizebyte >> 1) & 3),
                    693:          1 << ((sizebyte >> 3) & 3),
                    694:          1 << ((sizebyte >> 5) & 3));
                    695: }
                    696: 
1.106     anton     697: /* static superinstruction stuff */
                    698: 
1.141     anton     699: struct cost { /* super_info might be a more accurate name */
1.106     anton     700:   char loads;       /* number of stack loads */
                    701:   char stores;      /* number of stack stores */
                    702:   char updates;     /* number of stack pointer updates */
1.123     anton     703:   char branch;     /* is it a branch (SET_IP) */
1.125     anton     704:   unsigned char state_in;    /* state on entry */
                    705:   unsigned char state_out;   /* state on exit */
1.142     anton     706:   unsigned char imm_ops;     /* number of immediate operands */
1.123     anton     707:   short offset;     /* offset into super2 table */
1.125     anton     708:   unsigned char length;      /* number of components */
1.106     anton     709: };
                    710: 
1.121     anton     711: PrimNum super2[] = {
1.126     anton     712: #include SUPER2_I
1.106     anton     713: };
                    714: 
                    715: struct cost super_costs[] = {
1.126     anton     716: #include COSTS_I
1.106     anton     717: };
                    718: 
1.125     anton     719: struct super_state {
                    720:   struct super_state *next;
                    721:   PrimNum super;
                    722: };
                    723: 
1.106     anton     724: #define HASH_SIZE 256
                    725: 
                    726: struct super_table_entry {
                    727:   struct super_table_entry *next;
1.121     anton     728:   PrimNum *start;
1.106     anton     729:   short length;
1.125     anton     730:   struct super_state *ss_list; /* list of supers */
1.106     anton     731: } *super_table[HASH_SIZE];
                    732: int max_super=2;
                    733: 
1.125     anton     734: struct super_state *state_transitions=NULL;
                    735: 
1.161     pazsan    736: static int hash_super(PrimNum *start, int length)
1.106     anton     737: {
                    738:   int i, r;
                    739:   
                    740:   for (i=0, r=0; i<length; i++) {
                    741:     r <<= 1;
                    742:     r += start[i];
                    743:   }
                    744:   return r & (HASH_SIZE-1);
                    745: }
                    746: 
1.161     pazsan    747: static struct super_state **lookup_super(PrimNum *start, int length)
1.106     anton     748: {
                    749:   int hash=hash_super(start,length);
                    750:   struct super_table_entry *p = super_table[hash];
                    751: 
1.125     anton     752:   /* assert(length >= 2); */
1.106     anton     753:   for (; p!=NULL; p = p->next) {
                    754:     if (length == p->length &&
1.121     anton     755:        memcmp((char *)p->start, (char *)start, length*sizeof(PrimNum))==0)
1.125     anton     756:       return &(p->ss_list);
1.106     anton     757:   }
1.125     anton     758:   return NULL;
1.106     anton     759: }
                    760: 
1.161     pazsan    761: static void prepare_super_table()
1.106     anton     762: {
                    763:   int i;
1.109     anton     764:   int nsupers = 0;
1.106     anton     765: 
                    766:   for (i=0; i<sizeof(super_costs)/sizeof(super_costs[0]); i++) {
                    767:     struct cost *c = &super_costs[i];
1.125     anton     768:     if ((c->length < 2 || nsupers < static_super_number) &&
                    769:        c->state_in < maxstates && c->state_out < maxstates) {
                    770:       struct super_state **ss_listp= lookup_super(super2+c->offset, c->length);
                    771:       struct super_state *ss = malloc(sizeof(struct super_state));
                    772:       ss->super= i;
                    773:       if (c->offset==N_noop && i != N_noop) {
                    774:        if (is_relocatable(i)) {
                    775:          ss->next = state_transitions;
                    776:          state_transitions = ss;
                    777:        }
                    778:       } else if (ss_listp != NULL) {
                    779:        ss->next = *ss_listp;
                    780:        *ss_listp = ss;
                    781:       } else {
                    782:        int hash = hash_super(super2+c->offset, c->length);
                    783:        struct super_table_entry **p = &super_table[hash];
                    784:        struct super_table_entry *e = malloc(sizeof(struct super_table_entry));
                    785:        ss->next = NULL;
                    786:        e->next = *p;
                    787:        e->start = super2 + c->offset;
                    788:        e->length = c->length;
                    789:        e->ss_list = ss;
                    790:        *p = e;
                    791:       }
1.106     anton     792:       if (c->length > max_super)
                    793:        max_super = c->length;
1.125     anton     794:       if (c->length >= 2)
                    795:        nsupers++;
1.106     anton     796:     }
                    797:   }
1.144     pazsan    798:   debugp(stderr, "Using %d static superinsts\n", nsupers);
1.106     anton     799: }
                    800: 
                    801: /* dynamic replication/superinstruction stuff */
                    802: 
1.69      anton     803: #ifndef NO_DYNAMIC
1.161     pazsan    804: static int compare_priminfo_length(const void *_a, const void *_b)
1.76      anton     805: {
1.90      anton     806:   PrimInfo **a = (PrimInfo **)_a;
                    807:   PrimInfo **b = (PrimInfo **)_b;
1.77      anton     808:   Cell diff = (*a)->length - (*b)->length;
                    809:   if (diff)
                    810:     return diff;
                    811:   else /* break ties by start address; thus the decompiler produces
                    812:           the earliest primitive with the same code (e.g. noop instead
                    813:           of (char) and @ instead of >code-address */
                    814:     return (*b)->start - (*a)->start;
1.76      anton     815: }
1.112     anton     816: #endif /* !defined(NO_DYNAMIC) */
1.76      anton     817: 
1.125     anton     818: static char MAYBE_UNUSED superend[]={
1.126     anton     819: #include PRIM_SUPEREND_I
1.106     anton     820: };
1.107     anton     821: 
                    822: Cell npriminfos=0;
1.76      anton     823: 
1.146     anton     824: Label goto_start;
                    825: Cell goto_len;
                    826: 
1.162     pazsan    827: #ifndef NO_DYNAMIC
1.161     pazsan    828: static int compare_labels(const void *pa, const void *pb)
1.113     anton     829: {
1.114     anton     830:   Label a = *(Label *)pa;
                    831:   Label b = *(Label *)pb;
                    832:   return a-b;
                    833: }
1.162     pazsan    834: #endif
1.113     anton     835: 
1.161     pazsan    836: static Label bsearch_next(Label key, Label *a, UCell n)
1.114     anton     837:      /* a is sorted; return the label >=key that is the closest in a;
                    838:         return NULL if there is no label in a >=key */
                    839: {
                    840:   int mid = (n-1)/2;
                    841:   if (n<1)
                    842:     return NULL;
                    843:   if (n == 1) {
                    844:     if (a[0] < key)
                    845:       return NULL;
                    846:     else
                    847:       return a[0];
                    848:   }
                    849:   if (a[mid] < key)
                    850:     return bsearch_next(key, a+mid+1, n-mid-1);
                    851:   else
                    852:     return bsearch_next(key, a, mid+1);
1.113     anton     853: }
                    854: 
1.161     pazsan    855: static void check_prims(Label symbols1[])
1.47      anton     856: {
                    857:   int i;
1.90      anton     858: #ifndef NO_DYNAMIC
1.146     anton     859:   Label *symbols2, *symbols3, *ends1, *ends1j, *ends1jsorted, *goto_p;
1.119     anton     860:   int nends1j;
1.90      anton     861: #endif
1.47      anton     862: 
1.66      anton     863:   if (debug)
                    864: #ifdef __VERSION__
                    865:     fprintf(stderr, "Compiled with gcc-" __VERSION__ "\n");
                    866: #else
                    867: #define xstr(s) str(s)
                    868: #define str(s) #s
                    869:   fprintf(stderr, "Compiled with gcc-" xstr(__GNUC__) "." xstr(__GNUC_MINOR__) "\n"); 
                    870: #endif
1.121     anton     871:   for (i=0; symbols1[i]!=0; i++)
1.47      anton     872:     ;
1.55      anton     873:   npriminfos = i;
1.70      anton     874:   
                    875: #ifndef NO_DYNAMIC
1.66      anton     876:   if (no_dynamic)
                    877:     return;
1.164     pazsan    878:   symbols2=gforth_engine2(0,0,0,0,0);
1.70      anton     879: #if NO_IP
1.164     pazsan    880:   symbols3=gforth_engine3(0,0,0,0,0);
1.70      anton     881: #else
                    882:   symbols3=symbols1;
                    883: #endif
1.121     anton     884:   ends1 = symbols1+i+1;
1.119     anton     885:   ends1j =   ends1+i;
1.146     anton     886:   goto_p = ends1j+i+1; /* goto_p[0]==before; ...[1]==after;*/
1.121     anton     887:   nends1j = i+1;
1.119     anton     888:   ends1jsorted = (Label *)alloca(nends1j*sizeof(Label));
                    889:   memcpy(ends1jsorted,ends1j,nends1j*sizeof(Label));
                    890:   qsort(ends1jsorted, nends1j, sizeof(Label), compare_labels);
1.146     anton     891: 
                    892:   /* check whether the "goto *" is relocatable */
                    893:   goto_len = goto_p[1]-goto_p[0];
                    894:   debugp(stderr, "goto * %p %p len=%ld\n",
                    895:         goto_p[0],symbols2[goto_p-symbols1],goto_len);
                    896:   if (memcmp(goto_p[0],symbols2[goto_p-symbols1],goto_len)!=0) { /* unequal */
                    897:     no_dynamic=1;
                    898:     debugp(stderr,"  not relocatable, disabling dynamic code generation\n");
1.148     anton     899:     init_ss_cost();
1.146     anton     900:     return;
                    901:   }
                    902:   goto_start = goto_p[0];
1.113     anton     903:   
1.47      anton     904:   priminfos = calloc(i,sizeof(PrimInfo));
1.121     anton     905:   for (i=0; symbols1[i]!=0; i++) {
1.70      anton     906:     int prim_len = ends1[i]-symbols1[i];
1.47      anton     907:     PrimInfo *pi=&priminfos[i];
1.154     anton     908:     struct cost *sc=&super_costs[i];
1.70      anton     909:     int j=0;
                    910:     char *s1 = (char *)symbols1[i];
                    911:     char *s2 = (char *)symbols2[i];
                    912:     char *s3 = (char *)symbols3[i];
1.119     anton     913:     Label endlabel = bsearch_next(symbols1[i]+1,ends1jsorted,nends1j);
1.70      anton     914: 
                    915:     pi->start = s1;
1.121     anton     916:     pi->superend = superend[i]|no_super;
1.147     anton     917:     pi->length = prim_len;
1.113     anton     918:     pi->restlength = endlabel - symbols1[i] - pi->length;
1.70      anton     919:     pi->nimmargs = 0;
1.144     pazsan    920:     relocs++;
1.154     anton     921:     debugp(stderr, "%-15s %d-%d %4d %p %p %p len=%3ld rest=%2ld send=%1d",
                    922:           prim_names[i], sc->state_in, sc->state_out,
                    923:           i, s1, s2, s3, (long)(pi->length), (long)(pi->restlength),
                    924:           pi->superend);
1.114     anton     925:     if (endlabel == NULL) {
                    926:       pi->start = NULL; /* not relocatable */
1.122     anton     927:       if (pi->length<0) pi->length=100;
1.144     pazsan    928:       debugp(stderr,"\n   non_reloc: no J label > start found\n");
                    929:       relocs--;
                    930:       nonrelocs++;
1.114     anton     931:       continue;
                    932:     }
                    933:     if (ends1[i] > endlabel && !pi->superend) {
1.113     anton     934:       pi->start = NULL; /* not relocatable */
1.122     anton     935:       pi->length = endlabel-symbols1[i];
1.144     pazsan    936:       debugp(stderr,"\n   non_reloc: there is a J label before the K label (restlength<0)\n");
                    937:       relocs--;
                    938:       nonrelocs++;
1.113     anton     939:       continue;
                    940:     }
1.114     anton     941:     if (ends1[i] < pi->start && !pi->superend) {
1.113     anton     942:       pi->start = NULL; /* not relocatable */
1.122     anton     943:       pi->length = endlabel-symbols1[i];
1.144     pazsan    944:       debugp(stderr,"\n   non_reloc: K label before I label (length<0)\n");
                    945:       relocs--;
                    946:       nonrelocs++;
1.113     anton     947:       continue;
                    948:     }
1.138     anton     949:     assert(pi->length>=0);
1.113     anton     950:     assert(pi->restlength >=0);
1.74      anton     951:     while (j<(pi->length+pi->restlength)) {
1.70      anton     952:       if (s1[j]==s3[j]) {
                    953:        if (s1[j] != s2[j]) {
                    954:          pi->start = NULL; /* not relocatable */
1.144     pazsan    955:          debugp(stderr,"\n   non_reloc: engine1!=engine2 offset %3d",j);
1.74      anton     956:          /* assert(j<prim_len); */
1.144     pazsan    957:          relocs--;
                    958:          nonrelocs++;
1.70      anton     959:          break;
                    960:        }
                    961:        j++;
                    962:       } else {
                    963:        struct immarg *ia=&pi->immargs[pi->nimmargs];
                    964: 
                    965:        pi->nimmargs++;
                    966:        ia->offset=j;
                    967:        if ((~*(Cell *)&(s1[j]))==*(Cell *)&(s3[j])) {
                    968:          ia->rel=0;
1.144     pazsan    969:          debugp(stderr,"\n   absolute immarg: offset %3d",j);
1.70      anton     970:        } else if ((&(s1[j]))+(*(Cell *)&(s1[j]))+4 ==
                    971:                   symbols1[DOESJUMP+1]) {
                    972:          ia->rel=1;
1.144     pazsan    973:          debugp(stderr,"\n   relative immarg: offset %3d",j);
1.70      anton     974:        } else {
                    975:          pi->start = NULL; /* not relocatable */
1.144     pazsan    976:          debugp(stderr,"\n   non_reloc: engine1!=engine3 offset %3d",j);
1.74      anton     977:          /* assert(j<prim_len);*/
1.144     pazsan    978:          relocs--;
                    979:          nonrelocs++;
1.70      anton     980:          break;
                    981:        }
                    982:        j+=4;
1.47      anton     983:       }
                    984:     }
1.144     pazsan    985:     debugp(stderr,"\n");
1.70      anton     986:   }
1.76      anton     987:   decomp_prims = calloc(i,sizeof(PrimInfo *));
                    988:   for (i=DOESJUMP+1; i<npriminfos; i++)
                    989:     decomp_prims[i] = &(priminfos[i]);
                    990:   qsort(decomp_prims+DOESJUMP+1, npriminfos-DOESJUMP-1, sizeof(PrimInfo *),
                    991:        compare_priminfo_length);
1.70      anton     992: #endif
                    993: }
                    994: 
1.161     pazsan    995: static void flush_to_here(void)
1.74      anton     996: {
1.93      anton     997: #ifndef NO_DYNAMIC
1.100     anton     998:   if (start_flush)
                    999:     FLUSH_ICACHE(start_flush, code_here-start_flush);
1.74      anton    1000:   start_flush=code_here;
1.93      anton    1001: #endif
1.74      anton    1002: }
                   1003: 
1.93      anton    1004: #ifndef NO_DYNAMIC
1.161     pazsan   1005: static void append_jump(void)
1.74      anton    1006: {
                   1007:   if (last_jump) {
                   1008:     PrimInfo *pi = &priminfos[last_jump];
                   1009:     
                   1010:     memcpy(code_here, pi->start+pi->length, pi->restlength);
                   1011:     code_here += pi->restlength;
1.147     anton    1012:     memcpy(code_here, goto_start, goto_len);
                   1013:     code_here += goto_len;
1.74      anton    1014:     last_jump=0;
                   1015:   }
                   1016: }
                   1017: 
1.75      anton    1018: /* Gforth remembers all code blocks in this list.  On forgetting (by
                   1019: executing a marker) the code blocks are not freed (because Gforth does
                   1020: not remember how they were allocated; hmm, remembering that might be
                   1021: easier and cleaner).  Instead, code_here etc. are reset to the old
                   1022: value, and the "forgotten" code blocks are reused when they are
                   1023: needed. */
                   1024: 
                   1025: struct code_block_list {
                   1026:   struct code_block_list *next;
                   1027:   Address block;
                   1028:   Cell size;
                   1029: } *code_block_list=NULL, **next_code_blockp=&code_block_list;
                   1030: 
1.161     pazsan   1031: static Address append_prim(Cell p)
1.74      anton    1032: {
                   1033:   PrimInfo *pi = &priminfos[p];
                   1034:   Address old_code_here = code_here;
                   1035: 
1.159     anton    1036:   if (code_area+code_area_size < code_here+pi->length+pi->restlength+goto_len) {
1.75      anton    1037:     struct code_block_list *p;
1.74      anton    1038:     append_jump();
1.93      anton    1039:     flush_to_here();
1.75      anton    1040:     if (*next_code_blockp == NULL) {
1.161     pazsan   1041:       code_here = start_flush = code_area = gforth_alloc(code_area_size);
1.75      anton    1042:       p = (struct code_block_list *)malloc(sizeof(struct code_block_list));
                   1043:       *next_code_blockp = p;
                   1044:       p->next = NULL;
                   1045:       p->block = code_here;
                   1046:       p->size = code_area_size;
                   1047:     } else {
                   1048:       p = *next_code_blockp;
                   1049:       code_here = start_flush = code_area = p->block;
                   1050:     }
1.74      anton    1051:     old_code_here = code_here;
1.75      anton    1052:     next_code_blockp = &(p->next);
1.74      anton    1053:   }
                   1054:   memcpy(code_here, pi->start, pi->length);
                   1055:   code_here += pi->length;
                   1056:   return old_code_here;
                   1057: }
                   1058: #endif
1.75      anton    1059: 
                   1060: int forget_dyncode(Address code)
                   1061: {
                   1062: #ifdef NO_DYNAMIC
                   1063:   return -1;
                   1064: #else
                   1065:   struct code_block_list *p, **pp;
                   1066: 
                   1067:   for (pp=&code_block_list, p=*pp; p!=NULL; pp=&(p->next), p=*pp) {
                   1068:     if (code >= p->block && code < p->block+p->size) {
                   1069:       next_code_blockp = &(p->next);
                   1070:       code_here = start_flush = code;
                   1071:       code_area = p->block;
                   1072:       last_jump = 0;
                   1073:       return -1;
                   1074:     }
                   1075:   }
1.78      anton    1076:   return -no_dynamic;
1.75      anton    1077: #endif /* !defined(NO_DYNAMIC) */
                   1078: }
                   1079: 
1.161     pazsan   1080: static long dyncodesize(void)
1.104     anton    1081: {
                   1082: #ifndef NO_DYNAMIC
1.106     anton    1083:   struct code_block_list *p;
1.104     anton    1084:   long size=0;
                   1085:   for (p=code_block_list; p!=NULL; p=p->next) {
                   1086:     if (code_here >= p->block && code_here < p->block+p->size)
                   1087:       return size + (code_here - p->block);
                   1088:     else
                   1089:       size += p->size;
                   1090:   }
                   1091: #endif /* !defined(NO_DYNAMIC) */
                   1092:   return 0;
                   1093: }
                   1094: 
1.90      anton    1095: Label decompile_code(Label _code)
1.75      anton    1096: {
1.76      anton    1097: #ifdef NO_DYNAMIC
1.90      anton    1098:   return _code;
1.76      anton    1099: #else /* !defined(NO_DYNAMIC) */
                   1100:   Cell i;
1.77      anton    1101:   struct code_block_list *p;
1.90      anton    1102:   Address code=_code;
1.76      anton    1103: 
1.77      anton    1104:   /* first, check if we are in code at all */
                   1105:   for (p = code_block_list;; p = p->next) {
                   1106:     if (p == NULL)
                   1107:       return code;
                   1108:     if (code >= p->block && code < p->block+p->size)
                   1109:       break;
                   1110:   }
1.76      anton    1111:   /* reverse order because NOOP might match other prims */
                   1112:   for (i=npriminfos-1; i>DOESJUMP; i--) {
                   1113:     PrimInfo *pi=decomp_prims[i];
                   1114:     if (pi->start==code || (pi->start && memcmp(code,pi->start,pi->length)==0))
1.121     anton    1115:       return vm_prims[super2[super_costs[pi-priminfos].offset]];
1.118     anton    1116:     /* return pi->start;*/
1.76      anton    1117:   }
                   1118:   return code;
                   1119: #endif /* !defined(NO_DYNAMIC) */
1.75      anton    1120: }
1.74      anton    1121: 
1.70      anton    1122: #ifdef NO_IP
                   1123: int nbranchinfos=0;
                   1124: 
                   1125: struct branchinfo {
1.136     anton    1126:   Label **targetpp; /* **(bi->targetpp) is the target */
1.70      anton    1127:   Cell *addressptr; /* store the target here */
                   1128: } branchinfos[100000];
                   1129: 
                   1130: int ndoesexecinfos=0;
                   1131: struct doesexecinfo {
                   1132:   int branchinfo; /* fix the targetptr of branchinfos[...->branchinfo] */
1.136     anton    1133:   Label *targetp; /*target for branch (because this is not in threaded code)*/
1.70      anton    1134:   Cell *xt; /* cfa of word whose does-code needs calling */
                   1135: } doesexecinfos[10000];
                   1136: 
1.161     pazsan   1137: static void set_rel_target(Cell *source, Label target)
1.70      anton    1138: {
                   1139:   *source = ((Cell)target)-(((Cell)source)+4);
                   1140: }
                   1141: 
1.161     pazsan   1142: static void register_branchinfo(Label source, Cell *targetpp)
1.70      anton    1143: {
                   1144:   struct branchinfo *bi = &(branchinfos[nbranchinfos]);
1.136     anton    1145:   bi->targetpp = (Label **)targetpp;
1.70      anton    1146:   bi->addressptr = (Cell *)source;
                   1147:   nbranchinfos++;
                   1148: }
                   1149: 
1.161     pazsan   1150: static Address compile_prim1arg(PrimNum p, Cell **argp)
1.70      anton    1151: {
1.133     anton    1152:   Address old_code_here=append_prim(p);
1.70      anton    1153: 
1.74      anton    1154:   assert(vm_prims[p]==priminfos[p].start);
1.133     anton    1155:   *argp = (Cell*)(old_code_here+priminfos[p].immargs[0].offset);
                   1156:   return old_code_here;
1.70      anton    1157: }
                   1158: 
1.161     pazsan   1159: static Address compile_call2(Cell *targetpp, Cell **next_code_targetp)
1.70      anton    1160: {
1.73      anton    1161:   PrimInfo *pi = &priminfos[N_call2];
1.74      anton    1162:   Address old_code_here = append_prim(N_call2);
1.70      anton    1163: 
1.134     anton    1164:   *next_code_targetp = (Cell *)(old_code_here + pi->immargs[0].offset);
1.136     anton    1165:   register_branchinfo(old_code_here + pi->immargs[1].offset, targetpp);
1.134     anton    1166:   return old_code_here;
1.70      anton    1167: }
                   1168: #endif
                   1169: 
                   1170: void finish_code(void)
                   1171: {
                   1172: #ifdef NO_IP
                   1173:   Cell i;
                   1174: 
                   1175:   compile_prim1(NULL);
                   1176:   for (i=0; i<ndoesexecinfos; i++) {
                   1177:     struct doesexecinfo *dei = &doesexecinfos[i];
1.136     anton    1178:     dei->targetp = (Label *)DOES_CODE1((dei->xt));
                   1179:     branchinfos[dei->branchinfo].targetpp = &(dei->targetp);
1.70      anton    1180:   }
                   1181:   ndoesexecinfos = 0;
                   1182:   for (i=0; i<nbranchinfos; i++) {
                   1183:     struct branchinfo *bi=&branchinfos[i];
1.136     anton    1184:     set_rel_target(bi->addressptr, **(bi->targetpp));
1.70      anton    1185:   }
                   1186:   nbranchinfos = 0;
1.128     anton    1187: #else
                   1188:   compile_prim1(NULL);
1.48      anton    1189: #endif
1.93      anton    1190:   flush_to_here();
1.48      anton    1191: }
                   1192: 
1.162     pazsan   1193: #if !(defined(DOUBLY_INDIRECT) || defined(INDIRECT_THREADED))
1.128     anton    1194: #ifdef NO_IP
1.161     pazsan   1195: static Cell compile_prim_dyn(PrimNum p, Cell *tcp)
1.128     anton    1196:      /* compile prim #p dynamically (mod flags etc.) and return start
                   1197:        address of generated code for putting it into the threaded
                   1198:        code. This function is only called if all the associated
                   1199:        inline arguments of p are already in place (at tcp[1] etc.) */
                   1200: {
                   1201:   PrimInfo *pi=&priminfos[p];
                   1202:   Cell *next_code_target=NULL;
1.135     anton    1203:   Address codeaddr;
                   1204:   Address primstart;
1.128     anton    1205:   
                   1206:   assert(p<npriminfos);
                   1207:   if (p==N_execute || p==N_perform || p==N_lit_perform) {
1.134     anton    1208:     codeaddr = compile_prim1arg(N_set_next_code, &next_code_target);
1.135     anton    1209:     primstart = append_prim(p);
                   1210:     goto other_prim;
                   1211:   } else if (p==N_call) {
1.136     anton    1212:     codeaddr = compile_call2(tcp+1, &next_code_target);
1.128     anton    1213:   } else if (p==N_does_exec) {
                   1214:     struct doesexecinfo *dei = &doesexecinfos[ndoesexecinfos++];
1.133     anton    1215:     Cell *arg;
                   1216:     codeaddr = compile_prim1arg(N_lit,&arg);
                   1217:     *arg = (Cell)PFA(tcp[1]);
1.128     anton    1218:     /* we cannot determine the callee now (last_start[1] may be a
                   1219:        forward reference), so just register an arbitrary target, and
                   1220:        register in dei that we need to fix this before resolving
                   1221:        branches */
                   1222:     dei->branchinfo = nbranchinfos;
                   1223:     dei->xt = (Cell *)(tcp[1]);
1.134     anton    1224:     compile_call2(0, &next_code_target);
1.128     anton    1225:   } else if (!is_relocatable(p)) {
1.133     anton    1226:     Cell *branch_target;
                   1227:     codeaddr = compile_prim1arg(N_set_next_code, &next_code_target);
                   1228:     compile_prim1arg(N_branch,&branch_target);
                   1229:     set_rel_target(branch_target,vm_prims[p]);
1.128     anton    1230:   } else {
                   1231:     unsigned j;
1.135     anton    1232: 
                   1233:     codeaddr = primstart = append_prim(p);
                   1234:   other_prim:
1.128     anton    1235:     for (j=0; j<pi->nimmargs; j++) {
                   1236:       struct immarg *ia = &(pi->immargs[j]);
1.136     anton    1237:       Cell *argp = tcp + pi->nimmargs - j;
                   1238:       Cell argval = *argp; /* !! specific to prims */
1.128     anton    1239:       if (ia->rel) { /* !! assumption: relative refs are branches */
1.136     anton    1240:        register_branchinfo(primstart + ia->offset, argp);
1.128     anton    1241:       } else /* plain argument */
1.135     anton    1242:        *(Cell *)(primstart + ia->offset) = argval;
1.128     anton    1243:     }
                   1244:   }
                   1245:   if (next_code_target!=NULL)
                   1246:     *next_code_target = (Cell)code_here;
1.135     anton    1247:   return (Cell)codeaddr;
1.128     anton    1248: }
                   1249: #else /* !defined(NO_IP) */
1.161     pazsan   1250: static Cell compile_prim_dyn(PrimNum p, Cell *tcp)
1.128     anton    1251:      /* compile prim #p dynamically (mod flags etc.) and return start
                   1252:         address of generated code for putting it into the threaded code */
1.108     anton    1253: {
1.121     anton    1254:   Cell static_prim = (Cell)vm_prims[p];
1.108     anton    1255: #if defined(NO_DYNAMIC)
                   1256:   return static_prim;
                   1257: #else /* !defined(NO_DYNAMIC) */
                   1258:   Address old_code_here;
                   1259: 
                   1260:   if (no_dynamic)
                   1261:     return static_prim;
1.125     anton    1262:   if (p>=npriminfos || !is_relocatable(p)) {
1.108     anton    1263:     append_jump();
                   1264:     return static_prim;
                   1265:   }
                   1266:   old_code_here = append_prim(p);
1.147     anton    1267:   last_jump = p;
                   1268:   if (priminfos[p].superend)
                   1269:     append_jump();
1.108     anton    1270:   return (Cell)old_code_here;
                   1271: #endif  /* !defined(NO_DYNAMIC) */
                   1272: }
1.128     anton    1273: #endif /* !defined(NO_IP) */
1.162     pazsan   1274: #endif
1.70      anton    1275: 
1.109     anton    1276: #ifndef NO_DYNAMIC
1.161     pazsan   1277: static int cost_codesize(int prim)
1.109     anton    1278: {
1.121     anton    1279:   return priminfos[prim].length;
1.109     anton    1280: }
                   1281: #endif
                   1282: 
1.161     pazsan   1283: static int cost_ls(int prim)
1.109     anton    1284: {
                   1285:   struct cost *c = super_costs+prim;
                   1286: 
                   1287:   return c->loads + c->stores;
                   1288: }
                   1289: 
1.161     pazsan   1290: static int cost_lsu(int prim)
1.109     anton    1291: {
                   1292:   struct cost *c = super_costs+prim;
                   1293: 
                   1294:   return c->loads + c->stores + c->updates;
                   1295: }
                   1296: 
1.161     pazsan   1297: static int cost_nexts(int prim)
1.109     anton    1298: {
                   1299:   return 1;
                   1300: }
                   1301: 
                   1302: typedef int Costfunc(int);
                   1303: Costfunc *ss_cost =  /* cost function for optimize_bb */
                   1304: #ifdef NO_DYNAMIC
                   1305: cost_lsu;
                   1306: #else
                   1307: cost_codesize;
                   1308: #endif
                   1309: 
1.110     anton    1310: struct {
                   1311:   Costfunc *costfunc;
                   1312:   char *metricname;
                   1313:   long sum;
                   1314: } cost_sums[] = {
                   1315: #ifndef NO_DYNAMIC
                   1316:   { cost_codesize, "codesize", 0 },
                   1317: #endif
                   1318:   { cost_ls,       "ls",       0 },
                   1319:   { cost_lsu,      "lsu",      0 },
                   1320:   { cost_nexts,    "nexts",    0 }
                   1321: };
                   1322: 
1.148     anton    1323: #ifndef NO_DYNAMIC
                   1324: void init_ss_cost(void) {
                   1325:   if (no_dynamic && ss_cost == cost_codesize) {
                   1326:     ss_cost = cost_nexts;
                   1327:     cost_sums[0] = cost_sums[1]; /* don't use cost_codesize for print-metrics */
                   1328:     debugp(stderr, "--no-dynamic conflicts with --ss-min-codesize, reverting to --ss-min-nexts\n");
                   1329:   }
                   1330: }
                   1331: #endif
                   1332: 
1.106     anton    1333: #define MAX_BB 128 /* maximum number of instructions in BB */
1.125     anton    1334: #define INF_COST 1000000 /* infinite cost */
                   1335: #define CANONICAL_STATE 0
                   1336: 
                   1337: struct waypoint {
                   1338:   int cost;     /* the cost from here to the end */
                   1339:   PrimNum inst; /* the inst used from here to the next waypoint */
                   1340:   char relocatable; /* the last non-transition was relocatable */
                   1341:   char no_transition; /* don't use the next transition (relocatability)
                   1342:                       * or this transition (does not change state) */
                   1343: };
                   1344: 
1.156     anton    1345: struct tpa_state { /* tree parsing automaton (like) state */
1.155     anton    1346:   /* labeling is back-to-front */
                   1347:   struct waypoint *inst;  /* in front of instruction */
                   1348:   struct waypoint *trans; /* in front of instruction and transition */
                   1349: }; 
                   1350: 
1.156     anton    1351: struct tpa_state *termstate = NULL; /* initialized in loader() */
1.155     anton    1352: 
1.158     anton    1353: /* statistics about tree parsing (lazyburg) stuff */
                   1354: long lb_basic_blocks = 0;
                   1355: long lb_labeler_steps = 0;
                   1356: long lb_labeler_automaton = 0;
                   1357: long lb_labeler_dynprog = 0;
                   1358: long lb_newstate_equiv = 0;
                   1359: long lb_newstate_new = 0;
                   1360: long lb_applicable_base_rules = 0;
                   1361: long lb_applicable_chain_rules = 0;
                   1362: 
1.162     pazsan   1363: #if !(defined(DOUBLY_INDIRECT) || defined(INDIRECT_THREADED))
1.161     pazsan   1364: static void init_waypoints(struct waypoint ws[])
1.125     anton    1365: {
                   1366:   int k;
                   1367: 
                   1368:   for (k=0; k<maxstates; k++)
                   1369:     ws[k].cost=INF_COST;
                   1370: }
1.106     anton    1371: 
1.161     pazsan   1372: static struct tpa_state *empty_tpa_state()
1.155     anton    1373: {
1.156     anton    1374:   struct tpa_state *s = malloc(sizeof(struct tpa_state));
1.155     anton    1375: 
1.157     anton    1376:   s->inst  = calloc(maxstates,sizeof(struct waypoint));
1.155     anton    1377:   init_waypoints(s->inst);
1.157     anton    1378:   s->trans = calloc(maxstates,sizeof(struct waypoint));
1.155     anton    1379:   /* init_waypoints(s->trans);*/
                   1380:   return s;
                   1381: }
                   1382: 
1.161     pazsan   1383: static void transitions(struct tpa_state *t)
1.107     anton    1384: {
1.125     anton    1385:   int k;
                   1386:   struct super_state *l;
                   1387:   
                   1388:   for (k=0; k<maxstates; k++) {
1.155     anton    1389:     t->trans[k] = t->inst[k];
                   1390:     t->trans[k].no_transition = 1;
1.125     anton    1391:   }
                   1392:   for (l = state_transitions; l != NULL; l = l->next) {
                   1393:     PrimNum s = l->super;
                   1394:     int jcost;
                   1395:     struct cost *c=super_costs+s;
1.155     anton    1396:     struct waypoint *wi=&(t->trans[c->state_in]);
                   1397:     struct waypoint *wo=&(t->inst[c->state_out]);
1.158     anton    1398:     lb_applicable_chain_rules++;
1.125     anton    1399:     if (wo->cost == INF_COST)
                   1400:       continue;
                   1401:     jcost = wo->cost + ss_cost(s);
                   1402:     if (jcost <= wi->cost) {
                   1403:       wi->cost = jcost;
                   1404:       wi->inst = s;
                   1405:       wi->relocatable = wo->relocatable;
                   1406:       wi->no_transition = 0;
                   1407:       /* if (ss_greedy) wi->cost = wo->cost ? */
                   1408:     }
                   1409:   }
                   1410: }
1.107     anton    1411: 
1.161     pazsan   1412: static struct tpa_state *make_termstate()
1.155     anton    1413: {
1.157     anton    1414:   struct tpa_state *s = empty_tpa_state();
1.155     anton    1415: 
                   1416:   s->inst[CANONICAL_STATE].cost = 0;
                   1417:   transitions(s);
                   1418:   return s;
                   1419: }
1.162     pazsan   1420: #endif
1.155     anton    1421: 
1.156     anton    1422: #define TPA_SIZE 16384
                   1423: 
                   1424: struct tpa_entry {
                   1425:   struct tpa_entry *next;
                   1426:   PrimNum inst;
                   1427:   struct tpa_state *state_behind;  /* note: brack-to-front labeling */
                   1428:   struct tpa_state *state_infront; /* note: brack-to-front labeling */
                   1429: } *tpa_table[TPA_SIZE];
                   1430: 
1.162     pazsan   1431: #if !(defined(DOUBLY_INDIRECT) || defined(INDIRECT_THREADED))
1.161     pazsan   1432: static Cell hash_tpa(PrimNum p, struct tpa_state *t)
1.156     anton    1433: {
                   1434:   UCell it = (UCell )t;
                   1435:   return (p+it+(it>>14))&(TPA_SIZE-1);
                   1436: }
                   1437: 
1.161     pazsan   1438: static struct tpa_state **lookup_tpa(PrimNum p, struct tpa_state *t2)
1.156     anton    1439: {
                   1440:   int hash=hash_tpa(p, t2);
                   1441:   struct tpa_entry *te = tpa_table[hash];
                   1442: 
1.158     anton    1443:   if (tpa_noautomaton) {
                   1444:     static struct tpa_state *t;
                   1445:     t = NULL;
                   1446:     return &t;
                   1447:   }
1.156     anton    1448:   for (; te!=NULL; te = te->next) {
                   1449:     if (p == te->inst && t2 == te->state_behind)
                   1450:       return &(te->state_infront);
                   1451:   }
                   1452:   te = (struct tpa_entry *)malloc(sizeof(struct tpa_entry));
                   1453:   te->next = tpa_table[hash];
                   1454:   te->inst = p;
                   1455:   te->state_behind = t2;
                   1456:   te->state_infront = NULL;
                   1457:   tpa_table[hash] = te;
                   1458:   return &(te->state_infront);
                   1459: }
                   1460: 
1.161     pazsan   1461: static void tpa_state_normalize(struct tpa_state *t)
1.157     anton    1462: {
                   1463:   /* normalize so cost of canonical state=0; this may result in
                   1464:      negative states for some states */
                   1465:   int d = t->inst[CANONICAL_STATE].cost;
                   1466:   int i;
                   1467: 
                   1468:   for (i=0; i<maxstates; i++) {
                   1469:     if (t->inst[i].cost != INF_COST)
                   1470:       t->inst[i].cost -= d;
                   1471:     if (t->trans[i].cost != INF_COST)
                   1472:       t->trans[i].cost -= d;
                   1473:   }
                   1474: }
                   1475: 
1.161     pazsan   1476: static int tpa_state_equivalent(struct tpa_state *t1, struct tpa_state *t2)
1.157     anton    1477: {
                   1478:   return (memcmp(t1->inst, t2->inst, maxstates*sizeof(struct waypoint)) == 0 &&
                   1479:          memcmp(t1->trans,t2->trans,maxstates*sizeof(struct waypoint)) == 0);
                   1480: }
1.162     pazsan   1481: #endif
1.157     anton    1482: 
                   1483: struct tpa_state_entry {
                   1484:   struct tpa_state_entry *next;
                   1485:   struct tpa_state *state;
                   1486: } *tpa_state_table[TPA_SIZE];
                   1487: 
1.163     pazsan   1488: #if !(defined(DOUBLY_INDIRECT) || defined(INDIRECT_THREADED))
1.161     pazsan   1489: static Cell hash_tpa_state(struct tpa_state *t)
1.157     anton    1490: {
                   1491:   int *ti = (int *)(t->inst);
                   1492:   int *tt = (int *)(t->trans);
                   1493:   int r=0;
                   1494:   int i;
                   1495: 
                   1496:   for (i=0; ti+i < (int *)(t->inst+maxstates); i++)
                   1497:     r += ti[i]+tt[i];
                   1498:   return (r+(r>>14)+(r>>22)) & (TPA_SIZE-1);
                   1499: }
                   1500: 
1.161     pazsan   1501: static struct tpa_state *lookup_tpa_state(struct tpa_state *t)
1.157     anton    1502: {
                   1503:   Cell hash = hash_tpa_state(t);
                   1504:   struct tpa_state_entry *te = tpa_state_table[hash];
                   1505:   struct tpa_state_entry *tn;
                   1506: 
1.158     anton    1507:   if (!tpa_noequiv) {
                   1508:     for (; te!=NULL; te = te->next) {
                   1509:       if (tpa_state_equivalent(t, te->state)) {
                   1510:        lb_newstate_equiv++;
                   1511:        free(t->inst);
                   1512:        free(t->trans);
                   1513:        free(t);
                   1514:        return te->state;
                   1515:       }
1.157     anton    1516:     }
1.158     anton    1517:     tn = (struct tpa_state_entry *)malloc(sizeof(struct tpa_state_entry));
                   1518:     tn->next = te;
                   1519:     tn->state = t;
                   1520:     tpa_state_table[hash] = tn;
                   1521:   }
                   1522:   lb_newstate_new++;
                   1523:   if (tpa_trace)
                   1524:     fprintf(stderr, "%ld %ld lb_states\n", lb_labeler_steps, lb_newstate_new);
1.157     anton    1525:   return t;
                   1526: }
                   1527: 
1.125     anton    1528: /* use dynamic programming to find the shortest paths within the basic
                   1529:    block origs[0..ninsts-1] and rewrite the instructions pointed to by
                   1530:    instps to use it */
1.161     pazsan   1531: static void optimize_rewrite(Cell *instps[], PrimNum origs[], int ninsts)
1.125     anton    1532: {
                   1533:   int i,j;
1.156     anton    1534:   struct tpa_state *ts[ninsts+1];
1.125     anton    1535:   int nextdyn, nextstate, no_transition;
                   1536:   
1.158     anton    1537:   lb_basic_blocks++;
1.155     anton    1538:   ts[ninsts] = termstate;
1.107     anton    1539:   for (i=ninsts-1; i>=0; i--) {
1.156     anton    1540:     struct tpa_state **tp = lookup_tpa(origs[i],ts[i+1]);
                   1541:     struct tpa_state *t = *tp;
1.158     anton    1542:     lb_labeler_steps++;
                   1543:     if (t) {
1.156     anton    1544:       ts[i] = t;
1.158     anton    1545:       lb_labeler_automaton++;
                   1546:     }
1.156     anton    1547:     else {
1.158     anton    1548:       lb_labeler_dynprog++;
1.156     anton    1549:       ts[i] = empty_tpa_state();
                   1550:       for (j=1; j<=max_super && i+j<=ninsts; j++) {
                   1551:        struct super_state **superp = lookup_super(origs+i, j);
                   1552:        if (superp!=NULL) {
                   1553:          struct super_state *supers = *superp;
                   1554:          for (; supers!=NULL; supers = supers->next) {
                   1555:            PrimNum s = supers->super;
                   1556:            int jcost;
                   1557:            struct cost *c=super_costs+s;
                   1558:            struct waypoint *wi=&(ts[i]->inst[c->state_in]);
                   1559:            struct waypoint *wo=&(ts[i+j]->trans[c->state_out]);
                   1560:            int no_transition = wo->no_transition;
1.158     anton    1561:            lb_applicable_base_rules++;
1.156     anton    1562:            if (!(is_relocatable(s)) && !wo->relocatable) {
                   1563:              wo=&(ts[i+j]->inst[c->state_out]);
                   1564:              no_transition=1;
                   1565:            }
                   1566:            if (wo->cost == INF_COST) 
                   1567:              continue;
                   1568:            jcost = wo->cost + ss_cost(s);
                   1569:            if (jcost <= wi->cost) {
                   1570:              wi->cost = jcost;
                   1571:              wi->inst = s;
                   1572:              wi->relocatable = is_relocatable(s);
                   1573:              wi->no_transition = no_transition;
                   1574:              /* if (ss_greedy) wi->cost = wo->cost ? */
                   1575:            }
1.125     anton    1576:          }
1.107     anton    1577:        }
                   1578:       }
1.156     anton    1579:       transitions(ts[i]);
1.157     anton    1580:       tpa_state_normalize(ts[i]);
                   1581:       *tp = ts[i] = lookup_tpa_state(ts[i]);
1.158     anton    1582:       if (tpa_trace)
                   1583:        fprintf(stderr, "%ld %ld lb_table_entries\n", lb_labeler_steps, lb_labeler_dynprog);
1.107     anton    1584:     }
1.125     anton    1585:   }
                   1586:   /* now rewrite the instructions */
                   1587:   nextdyn=0;
                   1588:   nextstate=CANONICAL_STATE;
1.155     anton    1589:   no_transition = ((!ts[0]->trans[nextstate].relocatable) 
                   1590:                   ||ts[0]->trans[nextstate].no_transition);
1.125     anton    1591:   for (i=0; i<ninsts; i++) {
                   1592:     Cell tc=0, tc2;
                   1593:     if (i==nextdyn) {
                   1594:       if (!no_transition) {
                   1595:        /* process trans */
1.155     anton    1596:        PrimNum p = ts[i]->trans[nextstate].inst;
1.125     anton    1597:        struct cost *c = super_costs+p;
1.155     anton    1598:        assert(ts[i]->trans[nextstate].cost != INF_COST);
1.125     anton    1599:        assert(c->state_in==nextstate);
1.128     anton    1600:        tc = compile_prim_dyn(p,NULL);
1.125     anton    1601:        nextstate = c->state_out;
                   1602:       }
                   1603:       {
                   1604:        /* process inst */
1.155     anton    1605:        PrimNum p = ts[i]->inst[nextstate].inst;
1.125     anton    1606:        struct cost *c=super_costs+p;
                   1607:        assert(c->state_in==nextstate);
1.155     anton    1608:        assert(ts[i]->inst[nextstate].cost != INF_COST);
1.125     anton    1609: #if defined(GFORTH_DEBUGGING)
                   1610:        assert(p == origs[i]);
                   1611: #endif
1.128     anton    1612:        tc2 = compile_prim_dyn(p,instps[i]);
1.125     anton    1613:        if (no_transition || !is_relocatable(p))
                   1614:          /* !! actually what we care about is if and where
                   1615:           * compile_prim_dyn() puts NEXTs */
                   1616:          tc=tc2;
1.155     anton    1617:        no_transition = ts[i]->inst[nextstate].no_transition;
1.125     anton    1618:        nextstate = c->state_out;
                   1619:        nextdyn += c->length;
                   1620:       }
                   1621:     } else {
                   1622: #if defined(GFORTH_DEBUGGING)
                   1623:       assert(0);
                   1624: #endif
                   1625:       tc=0;
1.155     anton    1626:       /* tc= (Cell)vm_prims[ts[i]->inst[CANONICAL_STATE].inst]; */
1.125     anton    1627:     }
                   1628:     *(instps[i]) = tc;
                   1629:   }      
                   1630:   if (!no_transition) {
1.155     anton    1631:     PrimNum p = ts[i]->trans[nextstate].inst;
1.125     anton    1632:     struct cost *c = super_costs+p;
                   1633:     assert(c->state_in==nextstate);
1.155     anton    1634:     assert(ts[i]->trans[nextstate].cost != INF_COST);
1.125     anton    1635:     assert(i==nextdyn);
1.128     anton    1636:     (void)compile_prim_dyn(p,NULL);
1.125     anton    1637:     nextstate = c->state_out;
1.107     anton    1638:   }
1.125     anton    1639:   assert(nextstate==CANONICAL_STATE);
1.107     anton    1640: }
1.162     pazsan   1641: #endif
1.107     anton    1642: 
1.105     anton    1643: /* compile *start, possibly rewriting it into a static and/or dynamic
                   1644:    superinstruction */
                   1645: void compile_prim1(Cell *start)
1.70      anton    1646: {
1.108     anton    1647: #if defined(DOUBLY_INDIRECT)
1.125     anton    1648:   Label prim;
                   1649: 
                   1650:   if (start==NULL)
                   1651:     return;
                   1652:   prim = (Label)*start;
1.108     anton    1653:   if (prim<((Label)(xts+DOESJUMP)) || prim>((Label)(xts+npriminfos))) {
                   1654:     fprintf(stderr,"compile_prim encountered xt %p\n", prim);
                   1655:     *start=(Cell)prim;
                   1656:     return;
                   1657:   } else {
                   1658:     *start = (Cell)(prim-((Label)xts)+((Label)vm_prims));
                   1659:     return;
                   1660:   }
                   1661: #elif defined(INDIRECT_THREADED)
                   1662:   return;
1.112     anton    1663: #else /* !(defined(DOUBLY_INDIRECT) || defined(INDIRECT_THREADED)) */
1.128     anton    1664:   /* !! does not work, for unknown reasons; but something like this is
                   1665:      probably needed to ensure that we don't call compile_prim_dyn
                   1666:      before the inline arguments are there */
                   1667:   static Cell *instps[MAX_BB];
                   1668:   static PrimNum origs[MAX_BB];
                   1669:   static int ninsts=0;
                   1670:   PrimNum prim_num;
                   1671: 
                   1672:   if (start==NULL || ninsts >= MAX_BB ||
                   1673:       (ninsts>0 && superend[origs[ninsts-1]])) {
                   1674:     /* after bb, or at the start of the next bb */
                   1675:     optimize_rewrite(instps,origs,ninsts);
                   1676:     /* fprintf(stderr,"optimize_rewrite(...,%d)\n",ninsts); */
                   1677:     ninsts=0;
                   1678:     if (start==NULL)
                   1679:       return;
                   1680:   }
                   1681:   prim_num = ((Xt)*start)-vm_prims;
                   1682:   if(prim_num >= npriminfos) {
                   1683:     optimize_rewrite(instps,origs,ninsts);
1.129     anton    1684:     /* fprintf(stderr,"optimize_rewrite(...,%d)\n",ninsts);*/
1.128     anton    1685:     ninsts=0;
                   1686:     return;
                   1687:   }    
                   1688:   assert(ninsts<MAX_BB);
                   1689:   instps[ninsts] = start;
                   1690:   origs[ninsts] = prim_num;
                   1691:   ninsts++;
1.112     anton    1692: #endif /* !(defined(DOUBLY_INDIRECT) || defined(INDIRECT_THREADED)) */
1.47      anton    1693: }
                   1694: 
1.161     pazsan   1695: Address gforth_loader(FILE *imagefile, char* filename)
1.1       anton    1696: /* returns the address of the image proper (after the preamble) */
                   1697: {
                   1698:   ImageHeader header;
                   1699:   Address image;
                   1700:   Address imp; /* image+preamble */
1.17      anton    1701:   Char magic[8];
                   1702:   char magic7; /* size byte of magic number */
1.1       anton    1703:   Cell preamblesize=0;
1.6       pazsan   1704:   Cell data_offset = offset_image ? 56*sizeof(Cell) : 0;
1.1       anton    1705:   UCell check_sum;
1.15      pazsan   1706:   Cell ausize = ((RELINFOBITS ==  8) ? 0 :
                   1707:                 (RELINFOBITS == 16) ? 1 :
                   1708:                 (RELINFOBITS == 32) ? 2 : 3);
                   1709:   Cell charsize = ((sizeof(Char) == 1) ? 0 :
                   1710:                   (sizeof(Char) == 2) ? 1 :
                   1711:                   (sizeof(Char) == 4) ? 2 : 3) + ausize;
                   1712:   Cell cellsize = ((sizeof(Cell) == 1) ? 0 :
                   1713:                   (sizeof(Cell) == 2) ? 1 :
                   1714:                   (sizeof(Cell) == 4) ? 2 : 3) + ausize;
1.21      anton    1715:   Cell sizebyte = (ausize << 5) + (charsize << 3) + (cellsize << 1) +
                   1716: #ifdef WORDS_BIGENDIAN
                   1717:        0
                   1718: #else
                   1719:        1
                   1720: #endif
                   1721:     ;
1.1       anton    1722: 
1.164     pazsan   1723:   vm_prims = gforth_engine(0,0,0,0,0);
1.47      anton    1724:   check_prims(vm_prims);
1.106     anton    1725:   prepare_super_table();
1.1       anton    1726: #ifndef DOUBLY_INDIRECT
1.59      anton    1727: #ifdef PRINT_SUPER_LENGTHS
                   1728:   print_super_lengths();
                   1729: #endif
1.43      anton    1730:   check_sum = checksum(vm_prims);
1.1       anton    1731: #else /* defined(DOUBLY_INDIRECT) */
1.43      anton    1732:   check_sum = (UCell)vm_prims;
1.1       anton    1733: #endif /* defined(DOUBLY_INDIRECT) */
1.155     anton    1734: #if !(defined(DOUBLY_INDIRECT) || defined(INDIRECT_THREADED))
                   1735:   termstate = make_termstate();
                   1736: #endif /* !(defined(DOUBLY_INDIRECT) || defined(INDIRECT_THREADED)) */
1.10      pazsan   1737:   
                   1738:   do {
                   1739:     if(fread(magic,sizeof(Char),8,imagefile) < 8) {
1.84      anton    1740:       fprintf(stderr,"%s: image %s doesn't seem to be a Gforth (>=0.6) image.\n",
1.10      pazsan   1741:              progname, filename);
                   1742:       exit(1);
1.1       anton    1743:     }
1.10      pazsan   1744:     preamblesize+=8;
1.84      anton    1745:   } while(memcmp(magic,"Gforth3",7));
1.17      anton    1746:   magic7 = magic[7];
1.1       anton    1747:   if (debug) {
1.17      anton    1748:     magic[7]='\0';
1.21      anton    1749:     fprintf(stderr,"Magic found: %s ", magic);
                   1750:     print_sizes(magic7);
1.1       anton    1751:   }
                   1752: 
1.21      anton    1753:   if (magic7 != sizebyte)
                   1754:     {
                   1755:       fprintf(stderr,"This image is:         ");
                   1756:       print_sizes(magic7);
                   1757:       fprintf(stderr,"whereas the machine is ");
                   1758:       print_sizes(sizebyte);
1.1       anton    1759:       exit(-2);
                   1760:     };
                   1761: 
                   1762:   fread((void *)&header,sizeof(ImageHeader),1,imagefile);
1.10      pazsan   1763: 
                   1764:   set_stack_sizes(&header);
1.1       anton    1765:   
                   1766: #if HAVE_GETPAGESIZE
                   1767:   pagesize=getpagesize(); /* Linux/GNU libc offers this */
                   1768: #elif HAVE_SYSCONF && defined(_SC_PAGESIZE)
                   1769:   pagesize=sysconf(_SC_PAGESIZE); /* POSIX.4 */
                   1770: #elif PAGESIZE
                   1771:   pagesize=PAGESIZE; /* in limits.h according to Gallmeister's POSIX.4 book */
                   1772: #endif
1.144     pazsan   1773:   debugp(stderr,"pagesize=%ld\n",(unsigned long) pagesize);
1.1       anton    1774: 
1.34      anton    1775:   image = dict_alloc_read(imagefile, preamblesize+header.image_size,
                   1776:                          preamblesize+dictsize, data_offset);
1.33      anton    1777:   imp=image+preamblesize;
1.57      anton    1778:   alloc_stacks((ImageHeader *)imp);
1.1       anton    1779:   if (clear_dictionary)
1.33      anton    1780:     memset(imp+header.image_size, 0, dictsize-header.image_size);
1.90      anton    1781:   if(header.base==0 || header.base  == (Address)0x100) {
1.1       anton    1782:     Cell reloc_size=((header.image_size-1)/sizeof(Cell))/8+1;
1.162     pazsan   1783:     Char reloc_bits[reloc_size];
1.33      anton    1784:     fseek(imagefile, preamblesize+header.image_size, SEEK_SET);
1.10      pazsan   1785:     fread(reloc_bits, 1, reloc_size, imagefile);
1.161     pazsan   1786:     gforth_relocate((Cell *)imp, reloc_bits, header.image_size, (Cell)header.base, vm_prims);
1.1       anton    1787: #if 0
                   1788:     { /* let's see what the relocator did */
                   1789:       FILE *snapshot=fopen("snapshot.fi","wb");
                   1790:       fwrite(image,1,imagesize,snapshot);
                   1791:       fclose(snapshot);
                   1792:     }
                   1793: #endif
1.46      jwilke   1794:   }
                   1795:   else if(header.base!=imp) {
                   1796:     fprintf(stderr,"%s: Cannot load nonrelocatable image (compiled for address $%lx) at address $%lx\n",
                   1797:            progname, (unsigned long)header.base, (unsigned long)imp);
                   1798:     exit(1);
1.1       anton    1799:   }
                   1800:   if (header.checksum==0)
                   1801:     ((ImageHeader *)imp)->checksum=check_sum;
                   1802:   else if (header.checksum != check_sum) {
                   1803:     fprintf(stderr,"%s: Checksum of image ($%lx) does not match the executable ($%lx)\n",
                   1804:            progname, (unsigned long)(header.checksum),(unsigned long)check_sum);
                   1805:     exit(1);
                   1806:   }
1.53      anton    1807: #ifdef DOUBLY_INDIRECT
                   1808:   ((ImageHeader *)imp)->xt_base = xts;
                   1809: #endif
1.1       anton    1810:   fclose(imagefile);
                   1811: 
1.56      anton    1812:   /* unnecessary, except maybe for CODE words */
                   1813:   /* FLUSH_ICACHE(imp, header.image_size);*/
1.1       anton    1814: 
                   1815:   return imp;
                   1816: }
                   1817: 
1.72      anton    1818: /* pointer to last '/' or '\' in file, 0 if there is none. */
1.161     pazsan   1819: static char *onlypath(char *filename)
1.10      pazsan   1820: {
1.72      anton    1821:   return strrchr(filename, DIRSEP);
1.1       anton    1822: }
                   1823: 
1.161     pazsan   1824: static FILE *openimage(char *fullfilename)
1.10      pazsan   1825: {
                   1826:   FILE *image_file;
1.162     pazsan   1827:   char * expfilename = tilde_cstr((Char *)fullfilename, strlen(fullfilename), 1);
1.10      pazsan   1828: 
1.28      anton    1829:   image_file=fopen(expfilename,"rb");
1.1       anton    1830:   if (image_file!=NULL && debug)
1.28      anton    1831:     fprintf(stderr, "Opened image file: %s\n", expfilename);
1.10      pazsan   1832:   return image_file;
1.1       anton    1833: }
                   1834: 
1.28      anton    1835: /* try to open image file concat(path[0:len],imagename) */
1.161     pazsan   1836: static FILE *checkimage(char *path, int len, char *imagename)
1.10      pazsan   1837: {
                   1838:   int dirlen=len;
1.162     pazsan   1839:   char fullfilename[dirlen+strlen((char *)imagename)+2];
1.10      pazsan   1840: 
1.1       anton    1841:   memcpy(fullfilename, path, dirlen);
1.71      pazsan   1842:   if (fullfilename[dirlen-1]!=DIRSEP)
                   1843:     fullfilename[dirlen++]=DIRSEP;
1.1       anton    1844:   strcpy(fullfilename+dirlen,imagename);
1.10      pazsan   1845:   return openimage(fullfilename);
1.1       anton    1846: }
                   1847: 
1.161     pazsan   1848: static FILE * open_image_file(char * imagename, char * path)
1.1       anton    1849: {
1.10      pazsan   1850:   FILE * image_file=NULL;
1.28      anton    1851:   char *origpath=path;
1.10      pazsan   1852:   
1.71      pazsan   1853:   if(strchr(imagename, DIRSEP)==NULL) {
1.10      pazsan   1854:     /* first check the directory where the exe file is in !! 01may97jaw */
                   1855:     if (onlypath(progname))
1.72      anton    1856:       image_file=checkimage(progname, onlypath(progname)-progname, imagename);
1.10      pazsan   1857:     if (!image_file)
                   1858:       do {
                   1859:        char *pend=strchr(path, PATHSEP);
                   1860:        if (pend==NULL)
                   1861:          pend=path+strlen(path);
                   1862:        if (strlen(path)==0) break;
                   1863:        image_file=checkimage(path, pend-path, imagename);
                   1864:        path=pend+(*pend==PATHSEP);
                   1865:       } while (image_file==NULL);
                   1866:   } else {
                   1867:     image_file=openimage(imagename);
                   1868:   }
1.1       anton    1869: 
1.10      pazsan   1870:   if (!image_file) {
                   1871:     fprintf(stderr,"%s: cannot open image file %s in path %s for reading\n",
1.28      anton    1872:            progname, imagename, origpath);
1.10      pazsan   1873:     exit(1);
1.7       anton    1874:   }
                   1875: 
1.10      pazsan   1876:   return image_file;
                   1877: }
1.11      pazsan   1878: #endif
                   1879: 
                   1880: #ifdef HAS_OS
1.161     pazsan   1881: static UCell convsize(char *s, UCell elemsize)
1.11      pazsan   1882: /* converts s of the format [0-9]+[bekMGT]? (e.g. 25k) into the number
                   1883:    of bytes.  the letter at the end indicates the unit, where e stands
                   1884:    for the element size. default is e */
                   1885: {
                   1886:   char *endp;
                   1887:   UCell n,m;
                   1888: 
                   1889:   m = elemsize;
                   1890:   n = strtoul(s,&endp,0);
                   1891:   if (endp!=NULL) {
                   1892:     if (strcmp(endp,"b")==0)
                   1893:       m=1;
                   1894:     else if (strcmp(endp,"k")==0)
                   1895:       m=1024;
                   1896:     else if (strcmp(endp,"M")==0)
                   1897:       m=1024*1024;
                   1898:     else if (strcmp(endp,"G")==0)
                   1899:       m=1024*1024*1024;
                   1900:     else if (strcmp(endp,"T")==0) {
                   1901: #if (SIZEOF_CHAR_P > 4)
1.24      anton    1902:       m=1024L*1024*1024*1024;
1.11      pazsan   1903: #else
                   1904:       fprintf(stderr,"%s: size specification \"%s\" too large for this machine\n", progname, endp);
                   1905:       exit(1);
                   1906: #endif
                   1907:     } else if (strcmp(endp,"e")!=0 && strcmp(endp,"")!=0) {
                   1908:       fprintf(stderr,"%s: cannot grok size specification %s: invalid unit \"%s\"\n", progname, s, endp);
                   1909:       exit(1);
                   1910:     }
                   1911:   }
                   1912:   return n*m;
                   1913: }
1.10      pazsan   1914: 
1.109     anton    1915: enum {
                   1916:   ss_number = 256,
1.125     anton    1917:   ss_states,
1.109     anton    1918:   ss_min_codesize,
                   1919:   ss_min_ls,
                   1920:   ss_min_lsu,
                   1921:   ss_min_nexts,
                   1922: };
                   1923: 
1.10      pazsan   1924: void gforth_args(int argc, char ** argv, char ** path, char ** imagename)
                   1925: {
                   1926:   int c;
                   1927: 
1.1       anton    1928:   opterr=0;
                   1929:   while (1) {
                   1930:     int option_index=0;
                   1931:     static struct option opts[] = {
1.29      anton    1932:       {"appl-image", required_argument, NULL, 'a'},
1.1       anton    1933:       {"image-file", required_argument, NULL, 'i'},
                   1934:       {"dictionary-size", required_argument, NULL, 'm'},
                   1935:       {"data-stack-size", required_argument, NULL, 'd'},
                   1936:       {"return-stack-size", required_argument, NULL, 'r'},
                   1937:       {"fp-stack-size", required_argument, NULL, 'f'},
                   1938:       {"locals-stack-size", required_argument, NULL, 'l'},
                   1939:       {"path", required_argument, NULL, 'p'},
                   1940:       {"version", no_argument, NULL, 'v'},
                   1941:       {"help", no_argument, NULL, 'h'},
                   1942:       /* put something != 0 into offset_image */
                   1943:       {"offset-image", no_argument, &offset_image, 1},
                   1944:       {"no-offset-im", no_argument, &offset_image, 0},
                   1945:       {"clear-dictionary", no_argument, &clear_dictionary, 1},
1.4       anton    1946:       {"die-on-signal", no_argument, &die_on_signal, 1},
1.169     anton    1947:       {"ignore-async-signals", no_argument, &ignore_async_signals, 1},
1.1       anton    1948:       {"debug", no_argument, &debug, 1},
1.144     pazsan   1949:       {"diag", no_argument, &diag, 1},
1.60      anton    1950:       {"no-super", no_argument, &no_super, 1},
                   1951:       {"no-dynamic", no_argument, &no_dynamic, 1},
1.66      anton    1952:       {"dynamic", no_argument, &no_dynamic, 0},
1.110     anton    1953:       {"print-metrics", no_argument, &print_metrics, 1},
1.109     anton    1954:       {"ss-number", required_argument, NULL, ss_number},
1.125     anton    1955:       {"ss-states", required_argument, NULL, ss_states},
1.109     anton    1956: #ifndef NO_DYNAMIC
                   1957:       {"ss-min-codesize", no_argument, NULL, ss_min_codesize},
                   1958: #endif
                   1959:       {"ss-min-ls",       no_argument, NULL, ss_min_ls},
                   1960:       {"ss-min-lsu",      no_argument, NULL, ss_min_lsu},
                   1961:       {"ss-min-nexts",    no_argument, NULL, ss_min_nexts},
1.110     anton    1962:       {"ss-greedy",       no_argument, &ss_greedy, 1},
1.158     anton    1963:       {"tpa-noequiv",     no_argument, &tpa_noequiv, 1},
                   1964:       {"tpa-noautomaton", no_argument, &tpa_noautomaton, 1},
                   1965:       {"tpa-trace",      no_argument, &tpa_trace, 1},
1.1       anton    1966:       {0,0,0,0}
                   1967:       /* no-init-file, no-rc? */
                   1968:     };
                   1969:     
1.36      pazsan   1970:     c = getopt_long(argc, argv, "+i:m:d:r:f:l:p:vhoncsx", opts, &option_index);
1.1       anton    1971:     
                   1972:     switch (c) {
1.29      anton    1973:     case EOF: return;
                   1974:     case '?': optind--; return;
                   1975:     case 'a': *imagename = optarg; return;
1.10      pazsan   1976:     case 'i': *imagename = optarg; break;
1.1       anton    1977:     case 'm': dictsize = convsize(optarg,sizeof(Cell)); break;
                   1978:     case 'd': dsize = convsize(optarg,sizeof(Cell)); break;
                   1979:     case 'r': rsize = convsize(optarg,sizeof(Cell)); break;
                   1980:     case 'f': fsize = convsize(optarg,sizeof(Float)); break;
                   1981:     case 'l': lsize = convsize(optarg,sizeof(Cell)); break;
1.10      pazsan   1982:     case 'p': *path = optarg; break;
1.36      pazsan   1983:     case 'o': offset_image = 1; break;
                   1984:     case 'n': offset_image = 0; break;
                   1985:     case 'c': clear_dictionary = 1; break;
                   1986:     case 's': die_on_signal = 1; break;
                   1987:     case 'x': debug = 1; break;
1.83      anton    1988:     case 'v': fputs(PACKAGE_STRING"\n", stderr); exit(0);
1.109     anton    1989:     case ss_number: static_super_number = atoi(optarg); break;
1.125     anton    1990:     case ss_states: maxstates = max(min(atoi(optarg),MAX_STATE),1); break;
1.109     anton    1991: #ifndef NO_DYNAMIC
                   1992:     case ss_min_codesize: ss_cost = cost_codesize; break;
                   1993: #endif
                   1994:     case ss_min_ls:       ss_cost = cost_ls;       break;
                   1995:     case ss_min_lsu:      ss_cost = cost_lsu;      break;
                   1996:     case ss_min_nexts:    ss_cost = cost_nexts;    break;
1.1       anton    1997:     case 'h': 
1.29      anton    1998:       fprintf(stderr, "Usage: %s [engine options] ['--'] [image arguments]\n\
1.1       anton    1999: Engine Options:\n\
1.29      anton    2000:   --appl-image FILE                equivalent to '--image-file=FILE --'\n\
1.10      pazsan   2001:   --clear-dictionary               Initialize the dictionary with 0 bytes\n\
                   2002:   -d SIZE, --data-stack-size=SIZE   Specify data stack size\n\
                   2003:   --debug                          Print debugging information during startup\n\
1.144     pazsan   2004:   --diag                           Print diagnostic information during startup\n\
1.169     anton    2005:   --die-on-signal                  exit instead of THROWing some signals\n\
1.66      anton    2006:   --dynamic                        use dynamic native code\n\
1.10      pazsan   2007:   -f SIZE, --fp-stack-size=SIZE            Specify floating point stack size\n\
                   2008:   -h, --help                       Print this message and exit\n\
1.169     anton    2009:   --ignore-async-signals            ignore instead of THROWing async. signals\n\
1.10      pazsan   2010:   -i FILE, --image-file=FILE       Use image FILE instead of `gforth.fi'\n\
                   2011:   -l SIZE, --locals-stack-size=SIZE Specify locals stack size\n\
                   2012:   -m SIZE, --dictionary-size=SIZE   Specify Forth dictionary size\n\
1.60      anton    2013:   --no-dynamic                     Use only statically compiled primitives\n\
1.10      pazsan   2014:   --no-offset-im                   Load image at normal position\n\
1.60      anton    2015:   --no-super                        No dynamically formed superinstructions\n\
1.10      pazsan   2016:   --offset-image                   Load image at a different position\n\
                   2017:   -p PATH, --path=PATH             Search path for finding image and sources\n\
1.110     anton    2018:   --print-metrics                  Print some code generation metrics on exit\n\
1.10      pazsan   2019:   -r SIZE, --return-stack-size=SIZE Specify return stack size\n\
1.111     anton    2020:   --ss-greedy                       greedy, not optimal superinst selection\n\
                   2021:   --ss-min-codesize                 select superinsts for smallest native code\n\
                   2022:   --ss-min-ls                       minimize loads and stores\n\
                   2023:   --ss-min-lsu                      minimize loads, stores, and pointer updates\n\
                   2024:   --ss-min-nexts                    minimize the number of static superinsts\n\
                   2025:   --ss-number=N                     use N static superinsts (default max)\n\
1.125     anton    2026:   --ss-states=N                     N states for stack caching (default max)\n\
1.158     anton    2027:   --tpa-noequiv                     automaton without state equivalence\n\
                   2028:   --tpa-noautomaton                 dynamic programming only\n\
                   2029:   --tpa-trace                       report new states etc.\n\
1.66      anton    2030:   -v, --version                            Print engine version and exit\n\
1.1       anton    2031: SIZE arguments consist of an integer followed by a unit. The unit can be\n\
1.10      pazsan   2032:   `b' (byte), `e' (element; default), `k' (KB), `M' (MB), `G' (GB) or `T' (TB).\n",
                   2033:              argv[0]);
                   2034:       optind--;
                   2035:       return;
1.1       anton    2036:     }
                   2037:   }
1.10      pazsan   2038: }
1.11      pazsan   2039: #endif
1.10      pazsan   2040: 
1.161     pazsan   2041: static void print_diag()
1.144     pazsan   2042: {
                   2043: 
1.153     pazsan   2044: #if !defined(HAVE_GETRUSAGE) || (!defined(HAS_FFCALL) && !defined(HAS_LIBFFI))
1.145     pazsan   2045:   fprintf(stderr, "*** missing functionality ***\n"
1.144     pazsan   2046: #ifndef HAVE_GETRUSAGE
                   2047:          "    no getrusage -> CPUTIME broken\n"
                   2048: #endif
1.153     pazsan   2049: #if !defined(HAS_FFCALL) && !defined(HAS_LIBFFI)
1.144     pazsan   2050:          "    no ffcall -> only old-style foreign function calls (no fflib.fs)\n"
                   2051: #endif
                   2052:          );
                   2053: #endif
                   2054:   if((relocs < nonrelocs) ||
                   2055: #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)
                   2056:      1
                   2057: #else
                   2058:      0
                   2059: #endif
                   2060:      )
                   2061:     debugp(stderr, "relocs: %d:%d\n", relocs, nonrelocs);
1.165     pazsan   2062:     fprintf(stderr, "*** %sperformance problems ***\n%s",
                   2063: #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) || !defined(FORCE_REG) || defined(BUGGY_LONG_LONG)
                   2064:            "",
                   2065: #else
                   2066:            "no ",
                   2067: #endif
1.144     pazsan   2068: #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)
                   2069:            "    double-cell integer type buggy ->\n        "
                   2070: #ifdef BUGGY_LL_CMP
                   2071:            "CMP, "
                   2072: #endif
                   2073: #ifdef BUGGY_LL_MUL
                   2074:            "MUL, "
                   2075: #endif
                   2076: #ifdef BUGGY_LL_DIV
                   2077:            "DIV, "
                   2078: #endif
                   2079: #ifdef BUGGY_LL_ADD
                   2080:            "ADD, "
                   2081: #endif
                   2082: #ifdef BUGGY_LL_SHIFT
                   2083:            "SHIFT, "
                   2084: #endif
                   2085: #ifdef BUGGY_LL_D2F
                   2086:            "D2F, "
                   2087: #endif
                   2088: #ifdef BUGGY_LL_F2D
                   2089:            "F2D, "
                   2090: #endif
                   2091:            "\b\b slow\n"
1.145     pazsan   2092: #endif
                   2093: #ifndef FORCE_REG
                   2094:            "    automatic register allocation: performance degradation possible\n"
                   2095: #endif
                   2096: #if !defined(FORCE_REG) || defined(BUGGY_LONG_LONG)
                   2097:            "*** Suggested remedy: try ./configure"
                   2098: #ifndef FORCE_REG
                   2099:            " --enable-force-reg"
                   2100: #endif
                   2101: #ifdef BUGGY_LONG_LONG
                   2102:            " --enable-force-ll"
                   2103: #endif
                   2104:            "\n"
1.166     pazsan   2105: #else
                   2106:            ""
1.144     pazsan   2107: #endif
                   2108:            ,
                   2109:            (relocs < nonrelocs) ? "    gcc PR 15242 -> no dynamic code generation (use gcc-2.95 instead)\n" : "");
                   2110: }
                   2111: 
1.10      pazsan   2112: #ifdef INCLUDE_IMAGE
                   2113: extern Cell image[];
                   2114: extern const char reloc_bits[];
                   2115: #endif
1.67      pazsan   2116: 
1.10      pazsan   2117: int main(int argc, char **argv, char **env)
                   2118: {
1.30      pazsan   2119: #ifdef HAS_OS
1.10      pazsan   2120:   char *path = getenv("GFORTHPATH") ? : DEFAULTPATH;
1.30      pazsan   2121: #else
                   2122:   char *path = DEFAULTPATH;
                   2123: #endif
1.13      pazsan   2124: #ifndef INCLUDE_IMAGE
1.10      pazsan   2125:   char *imagename="gforth.fi";
                   2126:   FILE *image_file;
                   2127:   Address image;
                   2128: #endif
                   2129:   int retvalue;
                   2130:          
1.56      anton    2131: #if defined(i386) && defined(ALIGNMENT_CHECK)
1.10      pazsan   2132:   /* turn on alignment checks on the 486.
                   2133:    * on the 386 this should have no effect. */
                   2134:   __asm__("pushfl; popl %eax; orl $0x40000, %eax; pushl %eax; popfl;");
                   2135:   /* this is unusable with Linux' libc.4.6.27, because this library is
                   2136:      not alignment-clean; we would have to replace some library
                   2137:      functions (e.g., memcpy) to make it work. Also GCC doesn't try to keep
                   2138:      the stack FP-aligned. */
                   2139: #endif
                   2140: 
                   2141:   /* buffering of the user output device */
1.11      pazsan   2142: #ifdef _IONBF
1.10      pazsan   2143:   if (isatty(fileno(stdout))) {
                   2144:     fflush(stdout);
                   2145:     setvbuf(stdout,NULL,_IONBF,0);
1.1       anton    2146:   }
1.11      pazsan   2147: #endif
1.1       anton    2148: 
1.10      pazsan   2149:   progname = argv[0];
                   2150: 
1.11      pazsan   2151: #ifdef HAS_OS
1.10      pazsan   2152:   gforth_args(argc, argv, &path, &imagename);
1.109     anton    2153: #ifndef NO_DYNAMIC
1.148     anton    2154:   init_ss_cost();
1.109     anton    2155: #endif /* !defined(NO_DYNAMIC) */
                   2156: #endif /* defined(HAS_OS) */
1.10      pazsan   2157: 
                   2158: #ifdef INCLUDE_IMAGE
                   2159:   set_stack_sizes((ImageHeader *)image);
1.22      pazsan   2160:   if(((ImageHeader *)image)->base != image)
1.161     pazsan   2161:     gforth_relocate(image, reloc_bits, ((ImageHeader *)image)->image_size,
1.164     pazsan   2162:                    (Label*)gforth_engine(0, 0, 0, 0, 0));
1.10      pazsan   2163:   alloc_stacks((ImageHeader *)image);
                   2164: #else
                   2165:   image_file = open_image_file(imagename, path);
1.161     pazsan   2166:   image = gforth_loader(image_file, imagename);
1.10      pazsan   2167: #endif
1.24      anton    2168:   gforth_header=(ImageHeader *)image; /* used in SIGSEGV handler */
1.1       anton    2169: 
1.144     pazsan   2170:   if (diag)
                   2171:     print_diag();
1.1       anton    2172:   {
1.10      pazsan   2173:     char path2[strlen(path)+1];
1.1       anton    2174:     char *p1, *p2;
                   2175:     Cell environ[]= {
                   2176:       (Cell)argc-(optind-1),
                   2177:       (Cell)(argv+(optind-1)),
1.10      pazsan   2178:       (Cell)strlen(path),
1.1       anton    2179:       (Cell)path2};
                   2180:     argv[optind-1] = progname;
                   2181:     /*
                   2182:        for (i=0; i<environ[0]; i++)
                   2183:        printf("%s\n", ((char **)(environ[1]))[i]);
                   2184:        */
                   2185:     /* make path OS-independent by replacing path separators with NUL */
1.10      pazsan   2186:     for (p1=path, p2=path2; *p1!='\0'; p1++, p2++)
1.1       anton    2187:       if (*p1==PATHSEP)
                   2188:        *p2 = '\0';
                   2189:       else
                   2190:        *p2 = *p1;
                   2191:     *p2='\0';
1.161     pazsan   2192:     retvalue = gforth_go(image, 4, environ);
1.102     anton    2193: #ifdef SIGPIPE
                   2194:     bsd_signal(SIGPIPE, SIG_IGN);
                   2195: #endif
1.42      anton    2196: #ifdef VM_PROFILING
                   2197:     vm_print_profile(stderr);
                   2198: #endif
1.1       anton    2199:     deprep_terminal();
1.104     anton    2200:   }
1.110     anton    2201:   if (print_metrics) {
                   2202:     int i;
                   2203:     fprintf(stderr, "code size = %8ld\n", dyncodesize());
                   2204:     for (i=0; i<sizeof(cost_sums)/sizeof(cost_sums[0]); i++)
                   2205:       fprintf(stderr, "metric %8s: %8ld\n",
                   2206:              cost_sums[i].metricname, cost_sums[i].sum);
1.158     anton    2207:     fprintf(stderr,"lb_basic_blocks = %ld\n", lb_basic_blocks);
                   2208:     fprintf(stderr,"lb_labeler_steps = %ld\n", lb_labeler_steps);
                   2209:     fprintf(stderr,"lb_labeler_automaton = %ld\n", lb_labeler_automaton);
                   2210:     fprintf(stderr,"lb_labeler_dynprog = %ld\n", lb_labeler_dynprog);
                   2211:     fprintf(stderr,"lb_newstate_equiv = %ld\n", lb_newstate_equiv);
                   2212:     fprintf(stderr,"lb_newstate_new = %ld\n", lb_newstate_new);
                   2213:     fprintf(stderr,"lb_applicable_base_rules = %ld\n", lb_applicable_base_rules);
                   2214:     fprintf(stderr,"lb_applicable_chain_rules = %ld\n", lb_applicable_chain_rules);
                   2215:   }
                   2216:   if (tpa_trace) {
                   2217:     fprintf(stderr, "%ld %ld lb_states\n", lb_labeler_steps, lb_newstate_new);
                   2218:     fprintf(stderr, "%ld %ld lb_table_entries\n", lb_labeler_steps, lb_labeler_dynprog);
1.1       anton    2219:   }
1.13      pazsan   2220:   return retvalue;
1.1       anton    2221: }

FreeBSD-CVSweb <freebsd-cvsweb@FreeBSD.org>