diff --git a/HiSprite.py b/HiSprite.py
index c86fc06..08e5d7b 100755
--- a/HiSprite.py
+++ b/HiSprite.py
@@ -203,7 +203,7 @@ class Listing(object):
 class Sprite(Listing):
     backing_store_sizes = set()
 
-    def __init__(self, pngfile, assembler, screen, xdraw=False, use_mask=False, backing_store=False, processor="any", name=""):
+    def __init__(self, pngfile, assembler, screen, xdraw=False, use_mask=False, backing_store=False, clobber=False, processor="any", name=""):
         Listing.__init__(self, assembler)
         self.screen = screen
 
@@ -213,6 +213,7 @@ class Sprite(Listing):
         self.xdraw = xdraw
         self.use_mask = use_mask
         self.backing_store = backing_store
+        self.clobber = clobber
         self.processor = processor
         if not name:
             name = os.path.splitext(pngfile)[0]
@@ -267,7 +268,8 @@ class Sprite(Listing):
         self.asm("pla")
 
     def jump65C02(self):
-        self.save_axy_65C02()
+        if not self.clobber:
+            self.save_axy_65C02()
         self.asm("ldy PARAM0")
         self.asm("ldx MOD%d_%d,y" % (self.screen.numShifts, self.screen.bitsPerPixel))
 
@@ -280,7 +282,8 @@ class Sprite(Listing):
             self.addr("%s_SHIFT%d" % (self.slug, shift))
 
     def jump6502(self):
-        self.save_axy_6502()
+        if not self.clobber:
+            self.save_axy_6502()
         self.asm("ldy PARAM0")
         self.asm("ldx MOD%d_%d,y" % (self.screen.numShifts, self.screen.bitsPerPixel))
 
@@ -320,18 +323,20 @@ class Sprite(Listing):
         
         cycleCount, optimizationCount = self.generateBlitter(colorStreams, maskStreams, cycleCount)
 
-        if self.processor == "any":
-            self.out(".ifpC02")
-            self.restore_axy_65C02()
-            self.out(".else")
-            self.restore_axy_6502()
-            self.out(".endif")
-        elif self.processor == "65C02":
-            self.restore_axy_65C02()
-        elif self.processor == "6502":
-            self.restore_axy_6502()
-        else:
-            raise RuntimeError("Processor %s not supported" % self.processor)
+        if not self.clobber:
+            if self.processor == "any":
+                self.out(".ifpC02")
+                self.restore_axy_65C02()
+                self.out(".else")
+                self.restore_axy_6502()
+                self.out(".endif")
+            elif self.processor == "65C02":
+                self.restore_axy_65C02()
+            elif self.processor == "6502":
+                self.restore_axy_6502()
+            else:
+                raise RuntimeError("Processor %s not supported" % self.processor)
+        self.out()
         self.asm("rts")
         self.comment("Cycle count: %d, Optimized %d rows." % (cycleCount,optimizationCount))
 
@@ -698,6 +703,7 @@ if __name__ == "__main__":
     parser.add_argument("-p", "--processor", default="any", choices=["any","6502", "65C02"], help="Processor type (default: %(default)s)")
     parser.add_argument("-s", "--screen", default="hgrcolor", choices=["hgrcolor","hgrbw"], help="Screen format (default: %(default)s)")
     parser.add_argument("-n", "--name", default="", help="Name for generated assembly function (default: based on image filename)")
+    parser.add_argument("-k", "--clobber", action="store_true", default=False, help="don't save the registers on the stack")
     parser.add_argument("-o", "--output-prefix", default="", help="Base name to create a set of output files. If not supplied, all code will be sent to stdout.")
     parser.add_argument("files", metavar="IMAGE", nargs="*", help="a PNG image [or a list of them]. PNG files must not have an alpha channel!")
     options, extra_args = parser.parse_known_args()
@@ -725,7 +731,7 @@ if __name__ == "__main__":
 
     for pngfile in options.files:
         try:
-            listings.append(Sprite(pngfile, assembler, screen, options.xdraw, options.mask, options.backing_store, options.processor, options.name))
+            listings.append(Sprite(pngfile, assembler, screen, options.xdraw, options.mask, options.backing_store, options.clobber, options.processor, options.name))
         except RuntimeError, e:
             print "%s: %s" % (pngfile, e)
             sys.exit(1)
diff --git a/Makefile.rob b/Makefile.rob
index a1d3d72..1abb932 100644
--- a/Makefile.rob
+++ b/Makefile.rob
@@ -29,7 +29,7 @@ colortest.dsk: HiSprite.py colortest.s bwsprite.s
 	atrcopy colortest.dsk boot -b colortest.xex --brun 6000 -f
 
 multitest-sprite-driver.s: HiSprite.py $(BWSPRITE)
-	python HiSprite.py -a mac65 -p 6502 -s hgrbw -m -b -o multitest $(BWSPRITE) $(COLORSPRITE)
+	python HiSprite.py -a mac65 -p 6502 -s hgrbw -m -b -k -o multitest $(BWSPRITE) $(COLORSPRITE)
 
 multitest.dsk: HiSprite.py multitest.s multitest-sprite-driver.s
 	atasm -omultitest.xex multitest.s -Lmultitest.var -gmultitest.lst
diff --git a/multitest.dsk b/multitest.dsk
index 4f025cf..7d5a8f3 100644
Binary files a/multitest.dsk and b/multitest.dsk differ
diff --git a/multitest.s b/multitest.s
index 6304668..56a6934 100644
--- a/multitest.s
+++ b/multitest.s
@@ -57,8 +57,10 @@ initsprites
 ; Draw sprites by looping through the list of sprites
 renderstart
     ldy #0
+    sty PARAM3
 
 renderloop
+    ldy PARAM3
     lda sprite_active,y
     bmi renderend       ; end of list if negative
     beq renderskip      ; skip if zero
@@ -74,7 +76,7 @@ renderloop
 jsrsprite
     jsr $ffff           ; wish you could JSR ($nnnn)
 renderskip
-    iny
+    inc PARAM3
     bne renderloop      ; branch always because always positive; otherwise limited to 255 sprites (haha)
 
 renderend