Starting to optimize
authorHeikki Levanto <heikki@indexdata.dk>
Wed, 25 Aug 1999 18:09:23 +0000 (18:09 +0000)
committerHeikki Levanto <heikki@indexdata.dk>
Wed, 25 Aug 1999 18:09:23 +0000 (18:09 +0000)
isamc/isamd-p.h
isamc/isamd.c
isamc/merge-d.c

index aeb0a4c..c27a78e 100644 (file)
@@ -1,4 +1,4 @@
-/* $Id: isamd-p.h,v 1.5 1999-08-20 12:25:58 heikki Exp $
+/* $Id: isamd-p.h,v 1.6 1999-08-25 18:09:23 heikki Exp $
  * Copyright (c) 1995-1996, Index Data.
  * See the file LICENSE for details.
  * Heikki Levanto
@@ -68,15 +68,16 @@ typedef struct ISAMD_DIFF_s *ISAMD_DIFF;
 struct ISAMD_PP_s {
     char *buf;   /* buffer for read/write operations */
     ISAMD_BLOCK_SIZE offset; /* position for next read/write */
-    ISAMD_BLOCK_SIZE size;   /* size of actual pointer data */
+    ISAMD_BLOCK_SIZE size;   /* size of actual data */
     int cat;  /* category of this block */
     int pos;  /* block number of this block */
     int next; /* number of the next block */
     int diffs; /* either block or offset (in head) of start of diffs */
+               /* will not be used in the improved version! */
     ISAMD is;
-    void *decodeClientData;
+    void *decodeClientData;  /* delta-encoder's own data */
     ISAMD_DIFF diffinfo;
-    char *diffbuf; /* buffer for the diff block, only when reading */
+    char *diffbuf; /* buffer for the diff block */
     int numKeys;
 };
 
@@ -104,7 +105,10 @@ void isamd_free_diffs(ISAMD_PP pp);
 
 /*
  * $Log: isamd-p.h,v $
- * Revision 1.5  1999-08-20 12:25:58  heikki
+ * Revision 1.6  1999-08-25 18:09:23  heikki
+ * Starting to optimize
+ *
+ * Revision 1.5  1999/08/20 12:25:58  heikki
  * Statistics in isamd
  *
  * Revision 1.4  1999/07/21 14:24:50  heikki
index 6dcf031..af83ebd 100644 (file)
@@ -1,7 +1,7 @@
 /*
  * Copyright (c) 1995-1998, Index Data.
  * See the file LICENSE for details.
- * $Id: isamd.c,v 1.10 1999-08-24 13:17:42 heikki Exp $ 
+ * $Id: isamd.c,v 1.11 1999-08-25 18:09:24 heikki Exp $ 
  *
  * Isamd - isam with diffs 
  * Programmed by: Heikki Levanto
@@ -27,7 +27,7 @@ static void init_fc (ISAMD is, int cat);
 
 #define ISAMD_FREELIST_CHUNK 1
 
-#define SMALL_TEST 0
+#define SMALL_TEST 1
 
 ISAMD_M isamd_getmethod (ISAMD_M me)
 {
@@ -717,7 +717,10 @@ void isamd_pp_dump (ISAMD is, ISAMD_P ipos)
 
 /*
  * $Log: isamd.c,v $
- * Revision 1.10  1999-08-24 13:17:42  heikki
+ * Revision 1.11  1999-08-25 18:09:24  heikki
+ * Starting to optimize
+ *
+ * Revision 1.10  1999/08/24 13:17:42  heikki
  * Block sizes, comments
  *
  * Revision 1.9  1999/08/20 12:25:58  heikki
index 980f6d7..196f0e3 100644 (file)
@@ -3,7 +3,7 @@
  * See the file LICENSE for details.
  * Heikki Levanto
  *
- * $Id: merge-d.c,v 1.17 1999-08-24 13:17:42 heikki Exp $
+ * $Id: merge-d.c,v 1.18 1999-08-25 18:09:24 heikki Exp $
  *
  * missing
  *
@@ -24,6 +24,8 @@
  *  - On allocating more blocks (in append), check the order of blocks, and
  *    if needed, swap them. 
  *  - In merge, merge also with the input data.
+ *  - Write a routine to save/load indexes into a block, save only as many 
+ *    bytes as needed (size, diff, diffindexes)
  *
  * bugs
  *
  * may get an extra seek in read, but save merges.
  */
 
+
+#define NEW_ISAM_D 0  /* not yet ready to delete the old one! */
+
 #include <stdlib.h>
 #include <assert.h>
 #include <string.h>
@@ -133,6 +138,8 @@ static char *hexdump(unsigned char *p, int len, char *buff) {
   return buff;
 }
 
+#ifndef NEW_ISAM_D
+/* The next many lines are the old ISAM_D. Works, but not optimal */
 
 static int separateDiffBlock(ISAMD_PP pp)
 {
@@ -155,7 +162,7 @@ static int separateDiffBlock(ISAMD_PP pp)
 
 static void getDiffInfo(ISAMD_PP pp, int diffidx)
 { /* builds the diff info structures from a diffblock */
-   int maxinfos = pp->is->method->filecat[pp->cat].bsize / 5 +1;
+   int maxinfos = pp->is->method->filecat[pp->cat].bsize / 5 +2;
     /* Each diff takes at least 5 bytes. Probably more, but this is safe */
    int i=1;  /* [0] is used for the main data */
    int diffsz= maxinfos * sizeof(struct ISAMD_DIFF_s);
@@ -979,9 +986,721 @@ ISAMD_P isamd_append (ISAMD is, ISAMD_P ipos, ISAMD_I data)
 } /*  isamd_append */
 
 
-/*
- * $Log: merge-d.c,v $
- * Revision 1.17  1999-08-24 13:17:42  heikki
+
+#else /* NEW_ISAM_D */
+/***************************************************************
+ ***************************************************************
+ ***************************************************************
+ ***************************************************************
+ ***************************************************************
+ ***************************************************************
+ ***************************************************************
+ ***************************************************************
+ ***************************************************************
+ ***************************************************************
+ ***************************************************************
+ ***************************************************************
+ ***************************************************************
+ ***************************************************************
+ ***************************************************************
+ ***************************************************************
+ ***************************************************************
+ ***************************************************************/
+
+
+/***************************************************************
+ * General support routines
+ ***************************************************************/
+
+static void isamd_reduceblock(ISAMD_PP pp)
+/* takes a large block, and reduces its category if possible */
+/* Presumably the first block in an isam-list */
+{
+   if (pp->pos)
+      return; /* existing block, do not touch */
+      /* TODO: Probably we may touch anyway? */
+   if (pp->is->method->debug > 5)  
+     logf(LOG_LOG,"isamd_reduce: start p=%d c=%d sz=%d",
+       pp->pos, pp->cat, pp->size); 
+   while ( ( pp->cat > 0 ) && (!pp->next) && 
+           (pp->offset < pp->is->method->filecat[pp->cat-1].bsize ) )
+      pp->cat--;
+   pp->pos = isamd_alloc_block(pp->is, pp->cat);
+   if (pp->is->method->debug > 5) 
+     logf(LOG_LOG,"isamd_reduce:  got  p=%d c=%d sz=%d",
+       pp->pos, pp->cat, pp->size);    
+} /* reduceblock */
+
+
+static int save_first_pp ( ISAMD_PP firstpp)
+{
+   isamd_buildfirstblock(firstpp);
+   isamd_write_block(firstpp->is,firstpp->cat,firstpp->pos,firstpp->buf);
+   return isamd_addr(firstpp->pos,firstpp->cat);
+}
+
+static void save_last_pp (ISAMD_PP pp)
+{
+   pp->next = 0;/* just to be sure */
+   isamd_buildlaterblock(pp);
+   isamd_write_block(pp->is,pp->cat,pp->pos,pp->buf);
+}
+
+#ifdef UNUSED
+static int save_both_pps (ISAMD_PP firstpp, ISAMD_PP pp)
+{
+   /* order of things: Better to save firstpp first, if there are just two */
+   /* blocks, but last if there are blocks in between, as these have already */
+   /* been saved... optimise later (that's why this is in its own func...*/
+   int retval = save_first_pp(firstpp);
+   if (firstpp!=pp){ 
+      save_last_pp(pp);
+      isamd_pp_close(pp);
+   }
+   isamd_pp_close(firstpp);
+   return retval;
+} /* save_both_pps */
+#endif
+
+
+
+/***************************************************************
+ * Diffblock handling
+ ***************************************************************/
+
+void isamd_free_diffs(ISAMD_PP pp)
+{
+  int i;
+  if (pp->is->method->debug > 5)
+     logf(LOG_LOG,"isamd_free_diffs: pp=%p di=%p", pp, pp->diffinfo);
+  if (!pp->diffinfo) 
+    return;
+  for (i=1;pp->diffinfo[i].decodeData;i++) 
+  {
+      if (pp->is->method->debug > 8)
+         logf(LOG_LOG,"isamd_free_diffs [%d]=%p",i, 
+                       pp->diffinfo[i].decodeData);
+      (*pp->is->method->code_stop)(ISAMD_DECODE,pp->diffinfo[i].decodeData);
+  } 
+  xfree(pp->diffinfo);
+  if (pp->diffbuf != pp->buf)
+    xfree (pp->diffbuf);  
+  pp->diffbuf=0;
+  pp->diffinfo=0;
+} /* isamd_free_diffs */
+
+
+static void getDiffInfo(ISAMD_PP pp, int diffidx)
+{ /* builds the diff info structures from a diffblock */
+   int maxinfos = pp->is->method->filecat[pp->cat].bsize / 5 +1;
+    /* Each diff takes at least 5 bytes. Probably more, but this is safe */
+   int i=2;  /* [0] is used for the main data, [1] for merge inputs */
+   int diffsz= maxinfos * sizeof(struct ISAMD_DIFF_s);
+
+   pp->diffinfo = xmalloc( diffsz ); 
+   memset(pp->diffinfo,'\0',diffsz);
+   if (pp->is->method->debug > 5)   
+      logf(LOG_LOG,"isamd_getDiffInfo: %d (%d:%d), ix=%d mx=%d",
+         isamd_addr(pp->pos, pp->cat), pp->cat, pp->pos, diffidx,maxinfos);
+   assert(pp->diffbuf);
+   
+   pp->diffinfo[0].maxidx=-1; /* mark as special */
+   pp->diffinfo[1].maxidx=-1; /* mark as special */
+
+   while (i<maxinfos) 
+   {  
+      if ( diffidx+sizeof(int) > pp->is->method->filecat[pp->cat].bsize )
+      {
+         if (pp->is->method->debug > 5)
+           logf(LOG_LOG,"isamd_getDiffInfo:Near end (no room for len) at ix=%d n=%d",
+               diffidx, i);
+         return; /* whole block done */
+      }
+      memcpy( &pp->diffinfo[i].maxidx, &pp->diffbuf[diffidx], sizeof(int) );
+
+      if (pp->is->method->debug > 5)
+        logf(LOG_LOG,"isamd_getDiffInfo: max=%d ix=%d dbuf=%p",
+          pp->diffinfo[i].maxidx, diffidx, pp->diffbuf);
+
+      if ( (pp->is->method->debug > 0) &&
+         (pp->diffinfo[i].maxidx > pp->is->method->filecat[pp->cat].bsize) )
+      { /* bug-hunting, this fails on some long runs that log too much */
+         logf(LOG_LOG,"Bad MaxIx!!! %s:%d: diffidx=%d", 
+                       __FILE__,__LINE__, diffidx);
+         logf(LOG_LOG,"i=%d maxix=%d bsz=%d", i, pp->diffinfo[i].maxidx,
+                       pp->is->method->filecat[pp->cat].bsize);
+         logf(LOG_LOG,"pp=%d=%d:%d  pp->nx=%d=%d:%d",
+                       isamd_addr(pp->pos,pp->cat), pp->pos, pp->cat,
+                       pp->next, isamd_type(pp->next), isamd_block(pp->next) );                      
+      }
+      assert(pp->diffinfo[i].maxidx <= pp->is->method->filecat[pp->cat].bsize+1);
+
+      if (0==pp->diffinfo[i].maxidx)
+      {
+         if (pp->is->method->debug > 5)  //!!! 4
+           logf(LOG_LOG,"isamd_getDiffInfo:End mark at ix=%d n=%d",
+               diffidx, i);
+         return; /* end marker */
+      }
+      diffidx += sizeof(int);
+      pp->diffinfo[i].decodeData = (*pp->is->method->code_start)(ISAMD_DECODE);
+      pp->diffinfo[i].diffidx = diffidx;
+      if (pp->is->method->debug > 5)
+        logf(LOG_LOG,"isamd_getDiff[%d]:%d-%d %s",
+          i,diffidx-sizeof(int),pp->diffinfo[i].maxidx,
+          hexdump((char *)&pp->diffbuf[diffidx-4],8,0) );
+      diffidx=pp->diffinfo[i].maxidx;
+      if ( diffidx > pp->is->method->filecat[pp->cat].bsize )
+        return; /* whole block done */
+      ++i;
+   }
+   assert (!"too many diff sequences in the block");
+}
+
+/***************************************************************
+ * Main block operations
+ ***************************************************************/
+
+
+static ISAMD_PP get_new_main_block( ISAMD_PP firstpp, ISAMD_PP pp)
+{  /* allocates a new block for the main data, and links it in */
+  int newblock;
+  if (0 == firstpp->next) 
+  {  /* special case, pp not yet allocated. */
+     /*Started as largest size, that's fine */
+     pp->pos = isamd_alloc_block(pp->is,pp->cat);
+     firstpp->next = isamd_addr(pp->pos,pp->cat);
+     if (pp->is->method->debug >3) 
+        logf(LOG_LOG,"isamd_build: Alloc 1. dblock  p=%d=%d:%d",
+           isamd_addr(pp->pos,pp->cat), pp->cat, pp->pos);
+  }
+  newblock=isamd_alloc_block(pp->is,pp->cat);
+  pp->next=isamd_addr(pp->cat,newblock);
+  isamd_buildlaterblock(pp);
+  isamd_write_block(pp->is,pp->cat,pp->pos,pp->buf);
+  if (pp->is->method->debug >3) 
+     logf(LOG_LOG,"isamd_build: Alloc nxt %d=%d:%d -> %d=%d:%d",
+        isamd_addr(pp->pos,pp->cat), pp->cat, pp->pos,
+        isamd_addr(newblock,pp->cat), pp->cat, newblock);
+  pp->next=0;
+  pp->pos=newblock;
+  pp->size=pp->offset=ISAMD_BLOCK_OFFSET_N;  
+  return pp;
+} /* get_new_main_block */
+
+
+static ISAMD_PP  append_main_item(ISAMD_PP firstpp, 
+                                  ISAMD_PP pp, 
+                                  struct it_key *i_key)
+{  /* appends one item in the main data block, allocates new if needed */
+   char *i_item= (char *) i_key;  /* same as char */
+   char *i_ptr=i_item;
+   char codebuff[128];
+   char *c_ptr = codebuff;
+   int codelen;
+   char hexbuff[64];
+
+   int maxsize = pp->is->method->filecat[pp->is->max_cat].bsize; 
+
+   c_ptr=codebuff;
+   i_ptr=i_item;
+   (*pp->is->method->code_item)(ISAMD_ENCODE, pp->decodeClientData,
+                                &c_ptr, &i_ptr);
+   codelen = c_ptr - codebuff;
+   assert ( (codelen<128) && (codelen>0));
+   if (pp->is->method->debug >7)
+      logf(LOG_LOG,"isamd:build: coded into %s  (nk=%d)",
+          hexdump(codebuff, c_ptr-codebuff,hexbuff), firstpp->numKeys+1);
+
+   if (pp->offset + codelen > maxsize )
+   { /* oops, block full - get a new one */
+      pp =  get_new_main_block( firstpp, pp );
+      /* reset encoging and code again */
+      (*pp->is->method->code_reset)(pp->decodeClientData);
+      c_ptr=codebuff;
+      i_ptr=i_item;
+      (*pp->is->method->code_item)(ISAMD_ENCODE, pp->decodeClientData, 
+                                   &c_ptr, &i_ptr);
+      codelen = c_ptr - codebuff;
+      assert ( (codelen<128) && (codelen>0));
+      if (pp->is->method->debug >7)
+         logf(LOG_LOG,"isamd:build: recoded into %s  (nk=%d)",
+             hexdump(codebuff, c_ptr-codebuff,hexbuff), firstpp->numKeys+1);
+   } /* block full */    
+
+   assert (pp->offset + codelen <= maxsize );
+   
+   /* write the data into pp, now we must have room */ 
+   memcpy(&(pp->buf[pp->offset]),codebuff,codelen);
+   pp->offset += codelen;
+   pp->size += codelen;
+   firstpp->numKeys++;
+   /* clear the next 4 bytes in block, to avoid confusions with diff lens */
+   /* dirty, it should not be done here, but something slips somewhere, and */
+   /* I hope this fixes it...  - Heikki */
+   codelen = pp->offset;
+   while ( (codelen < maxsize ) && (codelen <= pp->offset+4) )
+     pp->buf[codelen++] = '\0';
+   return pp;    
+} /* append_main_item */
+
+
+/***************************************************************
+ * Merge
+ ***************************************************************/
+
+static int merge ( ISAMD_PP firstpp,      /* first pp (with diffs) */
+                   struct it_key *p_key,  /* the data item that didn't fit*/
+                   ISAMD_I data)          /* more input data comes here */
+{
+  int diffidx;  
+  int killblk=0;
+  struct it_key r_key;
+  char * r_ptr;
+  int r_more = 1;
+  ISAMD_PP pp;
+  ISAMD_PP readpp=firstpp;
+  int retval=0;
+  int diffcat = firstpp->cat;  /* keep the category of the diffblock even */
+                               /* if it is going to be empty now. */
+                               /* Alternative: Make it the minimal, and */
+                               /* resize later. Saves disk, but will lead */
+                               /* into bad seeks. */
+  
+  ++(readpp->is->files[0].no_merges);
+     
+  /* set up diffs as they should be for reading */
+  diffidx = ISAMD_BLOCK_OFFSET_1; 
+  readpp->diffbuf=readpp->buf;
+  getDiffInfo(readpp,diffidx);
+  
+  if (readpp->is->method->debug >4) 
+      logf(LOG_LOG,"isamd_merge: f=%d=%d:%d n=%d=%d:%d",
+        isamd_addr(firstpp->pos,firstpp->cat), firstpp->cat, firstpp->pos,
+        firstpp->next, isamd_type(firstpp->next), isamd_block(firstpp->next));  
+
+  /* release our data block. Do before reading, when pos is stable ! */
+  killblk=firstpp->pos;
+  if (killblk)
+  {
+      isamd_release_block(firstpp->is, firstpp->cat, killblk);
+      if (readpp->is->method->debug >3)   
+          logf(LOG_LOG,"isamd_merge: released old firstblock %d (%d:%d)",
+              isamd_addr(killblk,firstpp->cat), firstpp->cat, killblk );
+  }
+  
+  /* force the read to reload the first data block at first try */
+  readpp->offset=readpp->size+1;
+
+
+  r_ptr= (char *) &r_key;
+  r_more = isamd_read_item( readpp, &r_ptr);
+  if (!r_more)  
+  { /* oops, all data has been deleted! what to do??? */
+    /* never mind, we have at least one more delta to add to the block */
+    /* pray that is not a delete as well... */
+    r_key.sysno = 0;
+    r_key.seqno = 0;
+     if (readpp->is->method->debug >5) 
+         logf(LOG_LOG,"isamd_merge:all data has been deleted (nk=%d) ",
+            readpp->numKeys);
+    //assert (readpp->numKeys == 0);  /* no longer true! */
+  }
+
+
+  /* set up the new blocks for simple writing */
+  firstpp=isamd_pp_open(readpp->is,isamd_addr(0, diffcat));
+  firstpp->pos=isamd_alloc_block(firstpp->is,diffcat);
+  
+  pp=isamd_pp_open(readpp->is,isamd_addr(0,readpp->is->max_cat) );
+  
+  while (r_more)
+  {
+     if (readpp->is->method->debug >6) 
+         logf(LOG_LOG,"isamd_merge: got key %d.%d",
+           r_key.sysno, r_key.seqno );
+     pp= append_main_item(firstpp, pp, &r_key);
+
+     if ( (readpp->pos != killblk ) && (0!=readpp->pos) )
+     {  /* pos can get to 0 at end of main seq, if still diffs left...*/
+        if (readpp->is->method->debug >3)  
+            logf(LOG_LOG,"isamd_merge: released block %d (%d:%d) now %d=%d:%d",
+                isamd_addr(killblk,readpp->cat), readpp->cat, killblk,
+                isamd_addr(readpp->pos,readpp->cat),readpp->cat, readpp->pos );
+        isamd_release_block(readpp->is, readpp->cat, readpp->pos);
+        killblk=readpp->pos;
+     }
+
+     /* (try to) read next item */
+     r_ptr= (char *) &r_key;
+     r_more = isamd_read_item( readpp, &r_ptr);
+
+  } /* while read */
+  
+  
+  firstpp->diffs=0; 
+
+
+  isamd_reduceblock(pp);  /* reduce size if possible */
+  save_last_pp(pp);
+  if (readpp->is->method->debug >4) 
+      logf(LOG_LOG,"isamd_merge: saved last block %d=%d:%d",
+            isamd_addr(pp->pos,pp->cat), pp->cat, pp->pos);
+  isamd_pp_close(pp);
+
+  if (readpp->is->method->debug >5) 
+        logf(LOG_LOG,"isamd_merge: closing readpp %d=%d:%d di=%p",
+              isamd_addr(readpp->pos,readpp->cat), readpp->cat, readpp->pos,
+              readpp->diffinfo);
+  isamd_pp_close(readpp); /* pos is 0 by now, at eof. close works anyway */
+
+  if (readpp->is->method->debug >2)  
+      logf(LOG_LOG,"isamd_merge: merge ret f=%d=%d:%d pp=%d=%d:%d",
+            isamd_addr(firstpp->pos,pp->cat), firstpp->cat, firstpp->pos,
+            isamd_addr(pp->pos,pp->cat), pp->cat, pp->pos);
+
+  retval = isamd_addr(firstpp->pos, firstpp->cat);
+  isamd_pp_close(firstpp);
+
+  return retval; 
+  
+} /* merge */
+
+
+
+
+/***************************************************************
+ * Read with merge
+ ***************************************************************/
+
+/* Reads one item and corrects for the diffs, if any */
+/* return 1 for ok, 0 for eof */
+int isamd_read_item_merge (
+                     ISAMD_PP pp, 
+                     char **dst,
+                     struct it_key *p_key,  /* the data item that didn't fit*/
+                     ISAMD_I data)          /* more input data comes here */
+{                    /* The last two args can be null for ordinary reads */
+  char *keyptr;
+  char *codeptr;
+  char *codestart;
+  int winner=0; /* which diff holds the day */
+  int i; /* looping diffs */
+  int cmp;
+  int retry=1;
+  if (pp->diffs==0)  /* no diffs, just read the thing */
+     return isamd_read_main_item(pp,dst);
+
+  if (!pp->diffinfo)
+    getDiffInfo(pp, pp->offset);
+
+  if (p_key)  
+    pp->diffinfo[1].key = *p_key;  /* the key merge could not handle */
+  else
+    pp->diffinfo[1].key.sysno=0;
+
+  if (data)
+    pp->diffinfo[1].maxidx=-1;  /* signal we have diffs to read */
+  else
+    pp->diffinfo[1].maxidx=0;
+  pp->size=pp->offset=pp->is->method->filecat[pp->cat].bsize; 
+     /* this forces a read of the next block at first read */
+        
+  while (retry)
+
+  {
+     retry=0;
+     if (0==pp->diffinfo[0].key.sysno) 
+     { /* 0 is special case, main data. */
+        keyptr=(char*) &(pp->diffinfo[0].key);
+        pp->diffinfo[0].mode = ! isamd_read_main_item(pp,&keyptr);
+        if (pp->is->method->debug > 7)
+          logf(LOG_LOG,"isamd_read_item: read main %d.%d (%x.%x)",
+            pp->diffinfo[0].key.sysno, pp->diffinfo[0].key.seqno,
+            pp->diffinfo[0].key.sysno, pp->diffinfo[0].key.seqno);
+     } /* get main data */
+     
+     if ( (0==pp->diffinfo[1].key.sysno) && (-1==pp->diffinfo[1].maxidx) );
+     { /* 1 is another special case, the input data at merge */
+        keyptr = (char *) &pp->diffinfo[1].key;
+        i = (*data->read_item)(data->clientData, &keyptr, &pp->diffinfo[1].mode); 
+        if (!i) 
+        {  /* did not get it */
+           pp->diffinfo[1].key.sysno=0;
+           pp->diffinfo[1].maxidx=-2; /* stop trying */
+        }
+        if (pp->is->method->debug >6)
+           logf(LOG_LOG,"merge: read diff m=%d %d.%d (%x.%x)",
+              pp->diffinfo[1].mode, 
+              pp->diffinfo[1].key.sysno, pp->diffinfo[1].key.seqno,
+              pp->diffinfo[1].key.sysno, pp->diffinfo[1].key.seqno );        
+     } /* get input data */
+     
+     winner = 0;
+     for (i=1; (!retry) && (pp->diffinfo[i].decodeData); i++)
+     {
+        if (pp->is->method->debug > 8)
+          logf(LOG_LOG,"isamd_read_item: considering d%d %d.%d ix=%d mx=%d",
+               i, pp->diffinfo[i].key.sysno, pp->diffinfo[i].key.seqno,
+                  pp->diffinfo[i].diffidx,   pp->diffinfo[i].maxidx);
+            
+        if ( (0==pp->diffinfo[i].key.sysno) &&
+             (pp->diffinfo[i].diffidx < pp->diffinfo[i].maxidx))
+        {/* read a new one, if possible */
+           codeptr= codestart = &(pp->diffbuf[pp->diffinfo[i].diffidx]);
+           keyptr=(char *)&(pp->diffinfo[i].key);
+           (*pp->is->method->code_item)(ISAMD_DECODE,
+                 pp->diffinfo[i].decodeData, &keyptr, &codeptr);
+           pp->diffinfo[i].diffidx += codeptr-codestart;
+           pp->diffinfo[i].mode = pp->diffinfo[i].key.seqno & 1;
+           pp->diffinfo[i].key.seqno = pp->diffinfo[i].key.seqno >>1 ;
+           if (pp->is->method->debug > 7)
+             logf(LOG_LOG,"isamd_read_item: read diff[%d] %d.%d (%x.%x)",i,
+               pp->diffinfo[i].key.sysno, pp->diffinfo[i].key.seqno,
+               pp->diffinfo[i].key.sysno, pp->diffinfo[i].key.seqno);
+        } 
+        if ( 0!= pp->diffinfo[i].key.sysno)
+        { /* got a key, compare */
+          cmp=key_compare(&pp->diffinfo[i].key, &pp->diffinfo[winner].key);
+          if (0==pp->diffinfo[winner].key.sysno)
+            cmp=-1; /* end of main sequence, take all diffs */
+          if (cmp<0)
+          {
+             if (pp->is->method->debug > 8)
+               logf(LOG_LOG,"isamd_read_item: ins %d<%d %d.%d (%x.%x) < %d.%d (%x.%x)",
+                 i, winner,
+                 pp->diffinfo[i].key.sysno, pp->diffinfo[i].key.seqno,
+                 pp->diffinfo[i].key.sysno, pp->diffinfo[i].key.seqno,
+                 pp->diffinfo[winner].key.sysno, pp->diffinfo[winner].key.seqno,
+                 pp->diffinfo[winner].key.sysno, pp->diffinfo[winner].key.seqno);
+             if (pp->diffinfo[i].mode)  /* insert diff, should always be */
+               winner = i;
+             else
+               assert(!"delete diff for nonexisting item");  
+               /* is an assert too steep here? Not really.*/
+          } /* earlier key */
+          else if (cmp==0)
+          {
+             if (!pp->diffinfo[i].mode) /* delete diff. should always be */
+             {
+                if (pp->is->method->debug > 8)
+                  logf(LOG_LOG,"isamd_read_item: del %d at%d %d.%d (%x.%x)",
+                    i, winner,
+                    pp->diffinfo[i].key.sysno, pp->diffinfo[i].key.seqno,
+                    pp->diffinfo[i].key.sysno, pp->diffinfo[i].key.seqno);
+                pp->diffinfo[winner].key.sysno=0; /* delete it */
+             }
+             else
+                if (pp->is->method->debug > 2)
+                  logf(LOG_LOG,"isamd_read_item: duplicate ins %d at%d %d.%d (%x.%x)",
+                    i, winner,
+                    pp->diffinfo[i].key.sysno, pp->diffinfo[i].key.seqno,
+                    pp->diffinfo[i].key.sysno, pp->diffinfo[i].key.seqno);
+                /* skip the insert, since we already have it in the base */
+                /* Should we fail an assertion here??? */
+             pp->diffinfo[i].key.sysno=0; /* done with the delete */
+             retry=1; /* start all over again */
+          } /* matching key */
+          /* else it is a later key, its turn will come */
+        } /* got a key */
+     } /* for each diff */
+  } /* not retry */
+
+  if ( pp->diffinfo[winner].key.sysno)
+  {
+    if (pp->is->method->debug > 7)
+      logf(LOG_LOG,"isamd_read_item: got %d  %d.%d (%x.%x)",
+        winner,
+        pp->diffinfo[winner].key.sysno, pp->diffinfo[winner].key.seqno,
+        pp->diffinfo[winner].key.sysno, pp->diffinfo[winner].key.seqno);
+    memcpy(*dst, &pp->diffinfo[winner].key, sizeof(struct it_key) );
+    *dst += sizeof(struct it_key);
+    pp->diffinfo[winner].key.sysno=0; /* used that one up */
+    cmp= 1;
+  } 
+  else 
+  {
+    if (pp->is->method->debug > 7)
+      logf(LOG_LOG,"isamd_read_item: eof w=%d  %d.%d (%x.%x)",
+        winner,
+        pp->diffinfo[winner].key.sysno, pp->diffinfo[winner].key.seqno,
+        pp->diffinfo[winner].key.sysno, pp->diffinfo[winner].key.seqno);
+    assert(winner==0); /* if nothing found, nothing comes from a diff */
+    cmp= 0; /* eof */
+  }
+  return cmp;
+   
+} /* isamd_read_item */
+
+
+int isamd_read_item (ISAMD_PP pp, char **dst)
+{
+  return isamd_read_item_merge(pp,dst,0,0);
+}
+
+
+/***************************************************************
+ * Appending diffs 
+ ***************************************************************/
+
+
+
+static int append_diffs(ISAMD is, ISAMD_P ipos, ISAMD_I data)
+{
+   struct it_key i_key;    /* one input item */
+   char *i_item = (char *) &i_key;  /* same as chars */
+   char *i_ptr=i_item;
+   int i_more =1;
+   int i_mode;     /* 0 for delete, 1 for insert */ 
+
+   ISAMD_PP firstpp;
+   char hexbuff[64];
+   int diffidx=0;
+   int maxsize=0;
+   int difflenidx;
+   char codebuff[128];
+   char *c_ptr = codebuff;
+   int codelen;
+   int merge_rc;
+   int retval=0;
+
+   if (0==ipos)
+   {
+       firstpp=isamd_pp_open(is, isamd_addr(0,0) );
+       firstpp->size=firstpp->offset=ISAMD_BLOCK_OFFSET_1;
+         /* create in smallest category, will expand later */
+       ++(is->files[0].no_fbuilds);
+   } 
+   else
+   {
+       firstpp=isamd_pp_open(is, ipos);
+       ++(is->files[0].no_appds);
+   }
+
+   if (is->method->debug >2) 
+      logf(LOG_LOG,"isamd_appd: Start ipos=%d=%d:%d n=%d=%d:%d nk=%d",
+        ipos, isamd_type(ipos), isamd_block(ipos),
+        firstpp->next, isamd_type(firstpp->next), isamd_block(firstpp->diffs),
+        firstpp->numKeys);
+   maxsize = is->method->filecat[firstpp->cat].bsize; 
+   
+   difflenidx = diffidx = firstpp->size;
+   
+   diffidx+=sizeof(int);  /* difflen will be stored here */
+   
+   /* read first input */
+   i_ptr = i_item;
+   i_more = (*data->read_item)(data->clientData, &i_ptr, &i_mode); 
+
+   if (is->method->debug >6)
+      logf(LOG_LOG,"isamd_appd: start with m=%d %s",
+         i_mode, hexdump(i_item,i_ptr-i_item,hexbuff) );
+
+   while (i_more)
+   {     
+      /* store the mode bit inside key */
+      assert( ((i_key.seqno<<1)>>1) == i_key.seqno); /* can spare the bit */
+      i_key.seqno = i_key.seqno * 2 + i_mode;  
+
+      c_ptr=codebuff;
+      i_ptr=i_item;
+      (*is->method->code_item)(ISAMD_ENCODE, firstpp->decodeClientData, 
+                               &c_ptr, &i_ptr);
+      codelen = c_ptr - codebuff;
+      assert ( (codelen<128) && (codelen>0));
+      if (is->method->debug >7)
+         logf(LOG_LOG,"isamd_appd: coded into %d: %s (nk=%d) (ix=%d)",
+             codelen, hexdump(codebuff, codelen,hexbuff), 
+             firstpp->numKeys,diffidx);
+
+      if (diffidx + codelen > maxsize )
+      { /* block full */
+         if (firstpp->cat < firstpp->is->max_cat)
+         { /* just increase the block size */
+             if (firstpp->pos > 0)  /* free the old block if allocated */
+                 isamd_release_block(is, firstpp->cat, firstpp->pos);
+             ++firstpp->cat;
+             maxsize = is->method->filecat[firstpp->cat].bsize; 
+             firstpp->pos=0; /* need to allocate it when saving */             
+             if (is->method->debug >3)
+                logf(LOG_LOG,"isamd_appd: increased diff block to %d (%d)",
+                   firstpp->cat, maxsize);
+         }
+         else 
+         { /* max size already - can't help, need to merge it */
+             merge_rc = merge (firstpp, &i_key, data);
+             if (0!=merge_rc)
+               return merge_rc;  /* merge handled them all ! */
+             assert(!"merge returned zero ??");
+         } /* need to merge */
+      } /* block full */
+      
+      assert ( diffidx+codelen <= maxsize );
+      
+      /* save the diff */ 
+      memcpy(&(firstpp->buf[diffidx]),codebuff,codelen);
+      diffidx += codelen;
+      firstpp->size += codelen;
+      firstpp->offset +=codelen;
+      
+      if (i_mode)
+        firstpp->numKeys++; /* insert diff */
+      else
+        firstpp->numKeys--; /* delete diff */ 
+
+      /* update length of this diff run */
+      memcpy(&(firstpp->buf[difflenidx]),&diffidx,sizeof(diffidx));
+      
+      /* (try to) read the next input */
+      i_ptr = i_item;
+      i_more = (*data->read_item)(data->clientData, &i_ptr, &i_mode); 
+      if ( (i_more) && (is->method->debug >6) )
+         logf(LOG_LOG,"isamd_appd: got m=%d %s",
+            i_mode, hexdump(i_item,i_ptr-i_item,hexbuff) );
+   } /* more loop */
+
+   /* clear the next difflen, if room for such */
+   difflenidx = diffidx;
+   while ( (difflenidx-diffidx<=sizeof(int)) && (difflenidx<maxsize))
+     firstpp->buf[difflenidx++]='\0';
+
+   if (0==firstpp->pos)  /* need to (re)alloc the block */
+      firstpp->pos = isamd_alloc_block(is, firstpp->cat);
+
+   retval = save_first_pp( firstpp );
+   isamd_pp_close(firstpp);
+    
+   return retval;
+} /* append_diffs */
+
+
+
+
+/*************************************************************
+ * isamd_append itself, Sweet, isn't it
+ *************************************************************/
+
+ISAMD_P isamd_append (ISAMD is, ISAMD_P ipos, ISAMD_I data)
+{
+   return append_diffs(is,ipos,data);
+} /*  isamd_append */
+
+
+
+
+
+#endif /* NEW_ISAM_D */
+
+
+
+/*
+ * $Log: merge-d.c,v $
+ * Revision 1.18  1999-08-25 18:09:24  heikki
+ * Starting to optimize
+ *
+ * Revision 1.17  1999/08/24 13:17:42  heikki
  * Block sizes, comments
  *
  * Revision 1.16  1999/08/24 10:12:02  heikki