From a99e649a6b49ae072e21a5355ab8fa27469e14e3 Mon Sep 17 00:00:00 2001
From: Zeex <zeex@rocketmail.com>
Date: Sun, 22 Oct 2017 12:23:33 +0600
Subject: [PATCH] Group sequences of repeating literals into one write

Add a new "dumpn" pseudo-opcode that dumps N copies of the same value
to the stage buffer. It's basically the same as "dump" but much faster
when dumping the same value a lot of times, for example, when writing
initial values for large arrays.

On my Linux system the time it took to dump a 100000000 cell array went
down from 20 to 8 seconds in Release configuration, i.e. 2.5 times faster!

I haven't profiled further yet (Visual Studio 2017 profiler is broken,
gprof won't output anything (I'm probably doing it wrong), I might try
valgrind later).
---
 source/compiler/libpawnc.c |  6 +++-
 source/compiler/sc.h       |  1 +
 source/compiler/sc1.c      | 66 ++++++++++++++++++++++----------------
 source/compiler/sc4.c      |  9 ++++++
 source/compiler/sc6.c      | 16 ++++++++-
 5 files changed, 68 insertions(+), 30 deletions(-)

diff --git a/source/compiler/libpawnc.c b/source/compiler/libpawnc.c
index 4a94765..78be777 100644
--- a/source/compiler/libpawnc.c
+++ b/source/compiler/libpawnc.c
@@ -327,7 +327,11 @@ char *pc_readasm(void *handle, char *string, int maxchars)
  */
 void *pc_openbin(char *filename)
 {
-  return fopen(filename,"wb");
+  FILE *fbin;
+
+  fbin=fopen(filename,"wb");
+  setvbuf(fbin,NULL,_IOFBF,1UL<<20);
+  return fbin;
 }
 
 void pc_closebin(void *handle,int deletefile)
diff --git a/source/compiler/sc.h b/source/compiler/sc.h
index a26b9d2..9639725 100644
--- a/source/compiler/sc.h
+++ b/source/compiler/sc.h
@@ -639,6 +639,7 @@ SC_FUNC void ffabort(int reason);
 SC_FUNC void ffbounds(cell size);
 SC_FUNC void jumplabel(int number);
 SC_FUNC void defstorage(void);
+SC_FUNC void defcompactstorage(void);
 SC_FUNC void getfrm(void);
 SC_FUNC void modstk(int delta);
 SC_FUNC void setstk(cell value);
diff --git a/source/compiler/sc1.c b/source/compiler/sc1.c
index cff8561..5822946 100644
--- a/source/compiler/sc1.c
+++ b/source/compiler/sc1.c
@@ -414,7 +414,11 @@ char *pc_readasm(void *handle, char *string, int maxchars)
  */
 void *pc_openbin(char *filename)
 {
-  return fopen(filename,"wb");
+  FILE *fbin;
+
+  fbin=fopen(filename,"wb");
+  setvbuf(fbin,NULL,_IOFBF,1UL<<20);
+  return fbin;
 }
 
 void pc_closebin(void *handle,int deletefile)
@@ -1716,29 +1720,42 @@ static void parse(void)
  */
 static void dumplits(void)
 {
-  int j,k;
+  int i,j;
+  static const int row_len=16;
 
   if (sc_status==statSKIP)
     return;
 
-  k=0;
-  while (k<litidx){
+  i=0;
+  while (i<litidx) {
     /* should be in the data segment */
     assert(curseg==2);
-    defstorage();
-    j=16;       /* 16 values per line */
-    while (j && k<litidx){
-      outval(litq[k], FALSE);
+    j=i+1;
+    while (j<litidx && litq[j]==litq[i])
+      j++;
+    if (j-i>=row_len-1) {
+      int count=j-i;
+      defcompactstorage();
+      outval(litq[i],FALSE);
       stgwrite(" ");
-      k++;
-      j--;
-      if (j==0 || k>=litidx)
-        stgwrite("\n");         /* force a newline after 10 dumps */
-      /* Note: stgwrite() buffers a line until it is complete. It recognizes
-       * the end of line as a sequence of "\n\0", so something like "\n\t"
-       * so should not be passed to stgwrite().
-       */
-    } /* while */
+      outval(count,TRUE);
+      i+=count;
+    } else {
+      defstorage();
+      j=row_len;       /* 16 values per line */
+      while (j && i<litidx){
+        outval(litq[i],FALSE);
+        stgwrite(" ");
+        i++;
+        j--;
+        if (j==0 || i>=litidx)
+          stgwrite("\n");         /* force a newline after 10 dumps */
+        /* Note: stgwrite() buffers a line until it is complete. It recognizes
+         * the end of line as a sequence of "\n\0", so something like "\n\t"
+         * so should not be passed to stgwrite().
+         */
+      } /* while */
+    } /* if */
   } /* while */
 }
 
@@ -1748,20 +1765,13 @@ static void dumplits(void)
  */
 static void dumpzero(int count)
 {
-  int i;
-
   if (sc_status==statSKIP || count<=0)
     return;
   assert(curseg==2);
-  defstorage();
-  i=0;
-  while (count-- > 0) {
-    outval(0, FALSE);
-    i=(i+1) % 16;
-    stgwrite((i==0 || count==0) ? "\n" : " ");
-    if (i==0 && count>0)
-      defstorage();
-  } /* while */
+  defcompactstorage();
+  outval(0, FALSE);
+  stgwrite(" ");
+  outval(count, TRUE);
 }
 
 static void aligndata(int numbytes)
diff --git a/source/compiler/sc4.c b/source/compiler/sc4.c
index 4124883..c2ea105 100644
--- a/source/compiler/sc4.c
+++ b/source/compiler/sc4.c
@@ -832,6 +832,15 @@ SC_FUNC void defstorage(void)
   stgwrite("dump ");
 }
 
+/*
+ *  Same as defstorage() but for repeating values.
+ */
+
+SC_FUNC void defcompactstorage()
+{
+  stgwrite("dumpn ");
+}
+
 /*
  *  Copies frame address to primary register
  */
diff --git a/source/compiler/sc6.c b/source/compiler/sc6.c
index 4f89a63..9007477 100644
--- a/source/compiler/sc6.c
+++ b/source/compiler/sc6.c
@@ -382,7 +382,7 @@ static cell OPHANDLER_CALL parm5(FILE *fbin,char *params,cell opcode)
 static cell OPHANDLER_CALL do_dump(FILE *fbin,char *params,cell opcode)
 {
   ucell p;
-  int num = 0;
+  int num=0;
 
   while (*params!='\0') {
     p=getparam(params,&params);
@@ -395,6 +395,19 @@ static cell OPHANDLER_CALL do_dump(FILE *fbin,char *params,cell opcode)
   return num*sizeof(cell);
 }
 
+static cell OPHANDLER_CALL do_dumpn(FILE *fbin,char *params,cell opcode)
+{
+  ucell value,num,i;
+
+  value=getparam(params,&params);
+  num=getparam(params,NULL);
+  if (fbin!=NULL) {
+    for (i=0; i<num; i++)
+      write_encoded(fbin,&value,1);
+  } /* if */
+  return num*sizeof(cell);
+}
+
 static cell OPHANDLER_CALL do_call(FILE *fbin,char *params,cell opcode)
 {
   char name[sNAMEMAX+2]; /* +1 for a possible leading dot */
@@ -521,6 +534,7 @@ static OPCODE opcodelist[] = {
   {112, "dec.pri",    sIN_CSEG, parm0 },
   {115, "dec.s",      sIN_CSEG, parm1 },
   {  0, "dump",       sIN_DSEG, do_dump },
+  {  0, "dumpn",      sIN_DSEG, do_dumpn },
   { 95, "eq",         sIN_CSEG, parm0 },
   {106, "eq.c.alt",   sIN_CSEG, parm1 },
   {105, "eq.c.pri",   sIN_CSEG, parm1 },