Undid some of the rearrangement that made the code not want to run on real hardware.

Doesn't make any difference for the optimizations made so far.

draw.h

@@ -373,8 +373,8 @@ static INLINE void draw_quadstrip_textured(const ps1bsp_vertex_t *vertices, cons
 typedef struct _TMPVERT
 {
     int16_t vx, vy, vz, vpad;
-    uint8_t u, v;
     uint8_t r, g, b, cpad;
+    uint8_t u, v;
     uint16_t pad;
 } TMPVERT;
 
@@ -420,15 +420,15 @@ typedef struct _TMPVERT
 	"sh	$12, 0( %0 );"		\
 	"sh	$13, 2( %0 );"		\
 	"sh	$14, 4( %0 );"		\
-	"lbu	$12, 8( %1 );"	\
-	"lbu	$13, 9( %1 );"	\
-    "lbu	$14, 10( %1 );"	\
+	"lbu	$12, 12( %1 );"	\
+	"lbu	$13, 13( %1 );"	\
+    "lbu	$14, 8( %1 );"	\
 	"mtc2	$12, $9;"		\
 	"mtc2	$13, $10;"		\
     "mtc2	$14, $11;"		\
-	"lbu	$12, 8( %2 );"	\
-	"lbu	$13, 9( %2 );"	\
-    "lbu	$14, 10( %2 );"	\
+	"lbu	$12, 12( %2 );"	\
+	"lbu	$13, 13( %2 );"	\
+    "lbu	$14, 8( %2 );"	\
 	"ctc2	$12, $21;"		\
 	"ctc2	$13, $22;"		\
     "ctc2	$14, $23;"		\
@@ -438,11 +438,11 @@ typedef struct _TMPVERT
 	"mfc2	$12, $9;"		\
 	"mfc2	$13, $10;"		\
     "mfc2	$14, $11;"		\
-	"sb	$12, 8( %0 );"		\
-	"sb	$13, 9( %0 );"		\
+	"sb	$12, 12( %0 );"		\
+	"sb	$13, 13( %0 );"		\
+    "sb	$14, 8( %0 );"		\
+    "sb	$14, 9( %0 );"		\
     "sb	$14, 10( %0 );"		\
-    "sb	$14, 11( %0 );"		\
-    "sb	$14, 12( %0 );"		\
 	:						\
 	: "r"(dst), "r"(src0), "r"(src1)	\
 	: "$12", "$13", "$14", "memory" )