TODO

3 files changed, 63 insertions, 59 deletions

diff --git a/TODO b/TODO
index df2c513..8bbeafb 100644
--- a/TODO
+++ b/TODO
@@ -95,11 +95,11 @@ should be encoded with the earlier version in preference to the newer
 one.
 
 - Consider changing labels to be integers instead of strings. This
-will make it possible in many cases for code to be stored in static
-const arrays.
+  will make it possible in many cases for code to be stored in static
+  const arrays.
 
 - Storing the encoding in the info field. Maybe rename info to
-encoding or details.
+  encoding or details.
 
 
 Feature checking
diff --git a/iterjit.c b/iterjit.c
index d4d6689..c63fbaa 100644
--- a/iterjit.c
+++ b/iterjit.c
@@ -1,13 +1,61 @@
 /*
-  
+  Flow:
+  - outer loop:
+	- generates outer loop
+	- src/mask/dest have pre-scanline setups called
+	- body is generated by destination iter
+	- src/mask/dest have post-scanline finalizers called
+
+  - destination iter:
+	- for each alignment loop, it generates an inner loop
+	- body of inner loop is generated by combiner
+	- combiner returns pixels or an indication that 
+	     no change is required
+	- destination iter writes back if necessary
+	- advances pointer
+
+  - combiner drives generation of pixels
+  - if there is a mask, then
+	- tell mask iter to load n pixels
+	- generate code to check if all pixels are 0
+	- if it is, then
+		- depending on operator
+			- return 0
+			- return "no change needed"
+	  else:
+		- tell source iter to load n pixels
+		- add code to multiply them together
+		- depending on operator
+			- add code to check if a result is 0xff
+			- if it is, then
+				- return source
+			  else:
+				- tell destination to read a pixel
+				- generate code to combine with source
+				- return result
+	  tell mask iter to advance
+    else:
+          - tell source iter to load n pixels
+	  - depending on operator 
+		- add code to check if source is 0xff
+		- if it is, then
+			- return source
+		  else:
+			- tell destination to read pixel
+			- generate code to combine with source
+			- return result
+    tell source iter to advance
+
+  -=-
+    
   Iterator based JIT compiler
 
   There is a new structure that can be either source or
   destination. It contains function pointers that know how to generate
   pixels. The function pointers:
-
+    
   Source iterators:
-
+    
   - begin():		generates code that runs before the kernel
   - begin_line():	generates code that runs before each scanline
   - get_one_pixel():	   generates code that reads one pixel
@@ -24,7 +72,7 @@
   format_16_a8
   format_8_a8r8g8b8
   and possibly others
-
+    
   The iterator is supposed to return pixels in that format. It will
   never be asked to fetch more pixels than the format supports. Ie.,
   get_sixteen_pixels() will never be called if the format is
@@ -50,52 +98,6 @@
 
   There may also be setup()/setup_line() methods required.
 
-  Flow:
-  - outer loop:
-	- generates outer loop
-	- src/mask/dest have pre-scanline setups called
-	- body is generated by destination iter
-	- src/mask/dest have post-scanline finalizers called
-
-  - destination iter:
-	- for each alignment loop, it generates an inner loop
-	- body of inner loop is generated by combiner
-	- combiner returns pixels or an indication that 
-	     no change is required
-	- destination iter writes back if necessary
-	- advances pointer
-
-  - combiner drives generation of pixels
-  - if there is a mask, then
-	- tell mask iter to load n pixels
-	- generate code to check if all pixels are 0
-	- if it is, then
-		- depending on operator
-			- return 0
-			- return "no change needed"
-	  else:
-		- tell source iter to load n pixels
-		- add code to multiply them together
-		- depending on operator
-			- add code to check if a result is 0xff
-			- if it is, then
-				- return source
-			  else:
-				- tell destination to read a pixel
-				- generate code to combine with source
-				- return result
-	  tell mask iter to advance
-    else:
-          - tell source iter to load n pixels
-	  - depending on operator 
-		- add code to check if source is 0xff
-		- if it is, then
-			- return source
-		  else:
-			- tell destination to read pixel
-			- generate code to combine with source
-			- return result
-    tell source iter to advance
 */
 
 #include <stdio.h>
diff --git a/simplex86.c b/simplex86.c
index 9c51f17..f77e812 100644
--- a/simplex86.c
+++ b/simplex86.c
@@ -1912,6 +1912,7 @@ emit (fragment_t *frag, const variant_t *variant, const op_t ops[4])
 
     w = r = x = b = FALSE;
 
+    /* Emit opcode */
     if (variant->encoding != E_ANNOTATE && variant->encoding != E_D)
     {
         if (variant->encoding == E_O)
@@ -1927,8 +1928,9 @@ emit (fragment_t *frag, const variant_t *variant, const op_t ops[4])
 
         c = emit_opcode (c, opc);
     }
-    
-    vex_regno = 0x0000;
+
+    /* Emit operands */
+    vex_regno = 0x0;
     switch (variant->encoding)
     {
     case E_RM:
@@ -2053,10 +2055,10 @@ emit (fragment_t *frag, const variant_t *variant, const op_t ops[4])
     }
     else
     {
-        uint8_t rex = 0x40 | (w << 3) | (r << 2) | (x << 1) | (b << 0);
+        uint8_t rex = (w << 3) | (r << 2) | (x << 1) | (b << 0);
 
-        if (rex != 0x40)
-            *p++ = rex;
+        if (rex)
+            *p++ = 0x40 | rex;
     }
 
     /* Emit immediates */
@@ -2091,7 +2093,7 @@ emit (fragment_t *frag, const variant_t *variant, const op_t ops[4])
 	add_annotation (&frag->annotations, &ann);
     }
 
-    array_append (&frag->code, p - prefixes + c - code, &d);
+    array_append (&frag->code, (p - prefixes) + (c - code), &d);
     memcpy (d, prefixes, p - prefixes);
     memcpy (d + (p - prefixes), code, c - code);
 }