Added some macros to make the optimized drawing code more readable, and applied the same optimization to water drawing.

draw.h

@@ -5,6 +5,13 @@
 #include "display.h"
 #include <inline_c.h>
 
+// Macros for quickly blitting RGB and UV values with a single copy
+// This is faster than copying each value individually
+#define setColorFast(pr, r)         *((u_int*)(pr)) = *((u_int*)(r))
+#define setUVFast(pu, u)            *((u_short*)(pu)) = *((u_short*)(u))
+
+static u_int color_white = ((255 << 24) | (255 << 16) | (255 << 8) | 255);
+
 static INLINE void draw_triangle_lit(SVECTOR *verts, u_long *ot)
 {
     // Draw a single triangle
@@ -234,32 +241,31 @@ static INLINE void draw_quadstrip_textured(const ps1bsp_vertex_t *vertices, cons
         POLY_GT4 *poly = (POLY_GT4*)mem_prim(sizeof(POLY_GT4));
 
         // Fill out the quad's data fields in struct order, to optimize data access
-        // RGB and UV fields are blitted as a single value, which is faster than copying each value individually
         // First vertex and texture CLUT
-        *((u_int*)&poly->r0) = *((u_int*)&v0->r);
+        setColorFast(&poly->r0, &v0->r);
         gte_stsxy0(&poly->x0);
-        *((u_short*)&poly->u0) = *((u_short*)&v0->u);
+        setUVFast(&poly->u0, &v0->u);
         poly->clut = quake_clut;
 
         // Second vertex and texture page
-        *((u_int*)&poly->r1) = *((u_int*)&v1->r);
+        setColorFast(&poly->r1, &v1->r);
         gte_stsxy1(&poly->x1);
-        *((u_short*)&poly->u1) = *((u_short*)&v1->u);
+        setUVFast(&poly->u1, &v1->u);
         poly->tpage = tpage;
 
         // Third vertex
-        *((u_int*)&poly->r2) = *((u_int*)&v2->r);
+        setColorFast(&poly->r2, &v2->r);
         gte_stsxy2(&poly->x2);
-        *((u_short*)&poly->u2) = *((u_short*)&v2->u);
+        setUVFast(&poly->u2, &v2->u);
 
         // Transform the fourth vertex to complete the quad
         gte_ldv0(&vertices[v3->index]);
         gte_rtps();
 
         // Fourth vertex
-        *((u_int*)&poly->r3) = *((u_int*)&v3->r);
+        setColorFast(&poly->r3, &v3->r);
         gte_stsxy(&poly->x3);
-        *((u_short*)&poly->u3) = *((u_short*)&v3->u);
+        setUVFast(&poly->u3, &v3->u);
 
         setPolyGT4(poly);
         addPrim(ot, poly);
@@ -299,23 +305,33 @@ static INLINE void draw_quadstrip_water(const ps1bsp_vertex_t *vertices, const p
 
         // Draw a flat-shaded textured quad
         POLY_FT4 *poly = (POLY_FT4*)mem_prim(sizeof(POLY_FT4));
-        setPolyFT4(poly);
-        setSemiTrans(poly, semiTrans);
-        gte_stsxy3_ft3(poly);
+        
+        // Fill out the quad's data fields in struct order, to optimize data access
+        // First vertex and texture CLUT
+        setColorFast(&poly->r0, &color_white);
+        gte_stsxy0(&poly->x0);
+        setUVFast(&poly->u0, &v0->u);
+        poly->clut = quake_clut;
+
+        // Second vertex and texture page
+        gte_stsxy1(&poly->x1);
+        setUVFast(&poly->u1, &v1->u);
+        poly->tpage = tpage;
+
+        // Third vertex
+        gte_stsxy2(&poly->x2);
+        setUVFast(&poly->u2, &v2->u);
 
         // Transform the fourth vertex to complete the quad
         gte_ldv0(&vertices[v3->index]);
         gte_rtps();
-        gte_stsxy(&poly->x3);
-
-        // Texture UVs
-        setUV4(poly, v0->u, v0->v, v1->u, v1->v, v2->u, v2->v, v3->u, v3->v);
-        poly->clut = water_clut;
-        poly->tpage = tpage;
 
-        // Unlit
-        poly->r0 = poly->g0 = poly->b0 = 255;
+        // Fourth vertex
+        gte_stsxy(&poly->x3);
+        setUVFast(&poly->u3, &v3->u);
 
+        setPolyFT4(poly);
+        setSemiTrans(poly, semiTrans);
         addPrim(ot, poly);
         ++polyCount;
     }