DREAMWEB: optimized consequtive movsb/w and stosb/w

5 files changed, 97 insertions, 120 deletions

diff --git a/devtools/tasmrecover/tasm/cpp.py b/devtools/tasmrecover/tasm/cpp.py
index 4e37dc8e5b..2420829869 100644
--- a/devtools/tasmrecover/tasm/cpp.py
+++ b/devtools/tasmrecover/tasm/cpp.py
@@ -403,31 +403,31 @@ namespace %s {
 		self.body += p
 
 	def _rep(self):
-		self.body += "\twhile(context.cx--) ";
+		self.body += "\twhile(context.cx--) "
 
 	def _lodsb(self):
-		self.body += "\tcontext._lodsb();\n";
+		self.body += "\tcontext._lodsb();\n"
 
 	def _lodsw(self):
-		self.body += "\tcontext._lodsw();\n";
+		self.body += "\tcontext._lodsw();\n"
 
-	def _stosb(self):
-		self.body += "\tcontext._stosb();\n";
+	def _stosb(self, n):
+		self.body += "\tcontext._stosb(%s);\n" %("" if n == 1 else n)
 
-	def _stosw(self):
-		self.body += "\tcontext._stosw();\n";
+	def _stosw(self, n):
+		self.body += "\tcontext._stosw(%s);\n" %("" if n == 1 else n)
 
-	def _movsb(self):
-		self.body += "\tcontext._movsb();\n ";
+	def _movsb(self, n):
+		self.body += "\tcontext._movsb(%s);\n " %("" if n == 1 else n)
 
-	def _movsw(self):
-		self.body += "\tcontext._movsw();\n ";
+	def _movsw(self, n):
+		self.body += "\tcontext._movsw(%s);\n " %("" if n == 1 else n)
 
 	def _stc(self):
-		self.body += "\tcontext.flags._c = true;\n ";
+		self.body += "\tcontext.flags._c = true;\n "
 
 	def _clc(self):
-		self.body += "\tcontext.flags._c = false;\n ";
+		self.body += "\tcontext.flags._c = false;\n "
 
 	def __proc(self, name, def_skip = 0):
 		try:
diff --git a/devtools/tasmrecover/tasm/op.py b/devtools/tasmrecover/tasm/op.py
index 33f79c3e5c..9baebccfc3 100644
--- a/devtools/tasmrecover/tasm/op.py
+++ b/devtools/tasmrecover/tasm/op.py
@@ -316,27 +316,27 @@ class _lodsw(baseop):
 
 class _stosw(baseop):
 	def __init__(self, arg):
-		pass
+		self.repeat = 1
 	def visit(self, visitor):
-		visitor._stosw()
+		visitor._stosw(self.repeat)
 
 class _stosb(baseop):
 	def __init__(self, arg):
-		pass
+		self.repeat = 1
 	def visit(self, visitor):
-		visitor._stosb()
+		visitor._stosb(self.repeat)
 
 class _movsw(baseop):
 	def __init__(self, arg):
-		pass
+		self.repeat = 1
 	def visit(self, visitor):
-		visitor._movsw()
+		visitor._movsw(self.repeat)
 
 class _movsb(baseop):
 	def __init__(self, arg):
-		pass
+		self.repeat = 1
 	def visit(self, visitor):
-		visitor._movsb()
+		visitor._movsb(self.repeat)
 
 class _in(baseop):
 	def __init__(self, arg):
diff --git a/devtools/tasmrecover/tasm/proc.py b/devtools/tasmrecover/tasm/proc.py
index 57f17518fb..3c4ac21372 100644
--- a/devtools/tasmrecover/tasm/proc.py
+++ b/devtools/tasmrecover/tasm/proc.py
@@ -27,6 +27,30 @@ class proc:
 				self.stmts.remove(l)
 				return
 	
+	def optimize_sequence(self, cls):
+		i = 0
+		stmts = self.stmts
+		while i < len(stmts):
+			if not isinstance(stmts[i], cls):
+				i += 1
+				continue
+			j = i + 1
+
+			while j < len(stmts):
+				if not isinstance(stmts[j], cls):
+					break
+				j = j + 1
+
+			n = j - i
+			if n > 1:
+				print "Eliminate consequtive storage instructions at %u-%u" %(i, j)
+				del stmts[i + 1:j]
+				stmts[i].repeat = n
+			else:
+				i = j
+
+		return
+	
 	def optimize(self):
 		print "optimizing..."
 		#trivial simplifications, removing last ret
@@ -72,6 +96,11 @@ class proc:
 			if not used:
 				print self.labels
 				self.remove_label(s.name)
+
+		self.optimize_sequence(op._stosb);
+		self.optimize_sequence(op._stosw);
+		self.optimize_sequence(op._movsb);
+		self.optimize_sequence(op._movsw);
 	
 	def add(self, stmt):
 		#print stmt
diff --git a/engines/dreamweb/dreamgen.cpp b/engines/dreamweb/dreamgen.cpp
index cbdb25256c..f76a3a5b58 100644
--- a/engines/dreamweb/dreamgen.cpp
+++ b/engines/dreamweb/dreamgen.cpp
@@ -4204,38 +4204,7 @@ void clearwork(Context &context) {
 	context.di = 0;
 	context.cx = (200*320)/64;
 clearloop:
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
+	context._stosw(32);
 	if (--context.cx) goto clearloop;
 }
 
@@ -4366,70 +4335,27 @@ loop124:
 	context.bh = 14;
 	context.bh = 4;
 firstbitofblock:
-	context._movsw();
- 	context._movsw();
- 	context._movsw();
- 	context._movsw();
- 	context._movsw();
- 	context._movsw();
- 	context._movsw();
- 	context._movsw();
+	context._movsw(8);
  	context._add(context.di, (320)-16);
 	context._dec(context.bh);
 	if (!context.flags.z()) goto firstbitofblock;
 	context.bh = 12;
 loop125:
-	context._movsw();
- 	context._movsw();
- 	context._movsw();
- 	context._movsw();
- 	context._movsw();
- 	context._movsw();
- 	context._movsw();
- 	context._movsw();
+	context._movsw(8);
  	context.ax = 0x0dfdf;
-	context._stosw();
-	context._stosw();
+	context._stosw(2);
 	context._add(context.di, (320)-20);
 	context._dec(context.bh);
 	if (!context.flags.z()) goto loop125;
 	context._add(context.di, 4);
 	context.ax = 0x0dfdf;
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
+	context._stosw(8);
 	context._add(context.di, (320)-16);
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
+	context._stosw(8);
 	context._add(context.di, (320)-16);
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
+	context._stosw(8);
 	context._add(context.di, (320)-16);
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
-	context._stosw();
+	context._stosw(8);
 zeroblock:
 	context.si = context.pop();
 	context.di = context.pop();
@@ -4811,9 +4737,7 @@ void fadetowhite(Context &context) {
 	while(context.cx--) 	context._stosb();
 	context.di = (0+(180*10)+32+60+(32*32)+(11*10*3)+768);
 	context.al = 0;
-	context._stosb();
-	context._stosb();
-	context._stosb();
+	context._stosb(3);
 	paltostartpal(context);
 	context.data.byte(kFadedirection) = 1;
 	context.data.byte(kFadecount) = 63;
@@ -4830,9 +4754,7 @@ void fadefromwhite(Context &context) {
 	while(context.cx--) 	context._stosb();
 	context.di = (0+(180*10)+32+60+(32*32)+(11*10*3));
 	context.al = 0;
-	context._stosb();
-	context._stosb();
-	context._stosb();
+	context._stosb(3);
 	paltoendpal(context);
 	context.data.byte(kFadedirection) = 1;
 	context.data.byte(kFadecount) = 63;
@@ -17179,8 +17101,7 @@ void clearchanges(Context &context) {
 	context.es = context.cs;
 	context.di = 8011;
 	context.al = 1;
-	context._stosb();
-	context._stosb();
+	context._stosb(2);
 	context.al = 0;
 	context._stosb();
 	context.al = 1;
diff --git a/engines/dreamweb/runtime.h b/engines/dreamweb/runtime.h
index 1fb232c187..3b97ce1930 100644
--- a/engines/dreamweb/runtime.h
+++ b/engines/dreamweb/runtime.h
@@ -449,31 +449,58 @@ public:
 		src = r;
 	}
 
+	inline void _lodsb() {
+		al = ds.byte(si++);
+	}
+
+	inline void _lodsw() {
+		ax = ds.word(si);
+		si += 2;
+	}
+
 	inline void _movsb() {
 		es.byte(di++) = ds.byte(si++);
 	}
 
+	inline void _movsb(uint size) {
+		uint8 *dst = es.ptr(di, size);
+		uint8 *src = ds.ptr(si, size);
+		memcpy(dst, src, size);
+		di += size;
+		si += size;
+	}
+
 	inline void _movsw() {
 		_movsb();
 		_movsb();
 	}
 
-	inline void _lodsb() {
-		al = ds.byte(si++);
-	}
-
-	inline void _lodsw() {
-		ax = ds.word(si);
-		si += 2;
+	inline void _movsw(uint size) {
+		_movsb(size * 2);
 	}
 
 	inline void _stosb() {
 		es.byte(di++) = al;
 	}
 
+	inline void _stosb(uint size) {
+		uint8 *dst = es.ptr(di, size);
+		memset(dst, al, size);
+		di += size;
+	}
+
 	inline void _stosw() {
-		es.word(di) = ax;
-		di += 2;
+		es.byte(di++) = al;
+		es.byte(di++) = ah;
+	}
+
+	inline void _stosw(uint size) {
+		uint8 *dst = es.ptr(di, size);
+		di += 2 * size;
+		while(size--) {
+			*dst++ = al;
+			*dst++ = ah;
+		}
 	}
 
 	inline void _xchg(uint16 &a, uint16 &b) {