diff --git a/source/src/physics.cpp b/source/src/physics.cpp
index 50411c8d..6e39274c 100644
--- a/source/src/physics.cpp
+++ b/source/src/physics.cpp
@@ -176,7 +176,7 @@ bool collide(physent *d, bool spawn, float drop, float rise)
     const float eyeheight = d->eyeheight;
     const float playerheight = eyeheight + d->aboveeye;
 
-    for(int x = x1; x<=x2; x++) for(int y = y1; y<=y2; y++)     // collide with map
+    for(int y = y1; y<=y2; y++) for(int x = x1; x<=x2; x++)     // collide with map
     {
         if(OUTBORD(x,y)) return false;
         sqr *s = S(x,y);
diff --git a/source/src/rendercubes.cpp b/source/src/rendercubes.cpp
index 4a0a916e..7845e473 100644
--- a/source/src/rendercubes.cpp
+++ b/source/src/rendercubes.cpp
@@ -144,7 +144,7 @@ void finishstrips() { stripend(verts); }
 sqr sbright, sdark;
 VARP(lighterror, 1, 4, 100);
 
-void render_flat(int wtex, int x, int y, int size, int h, sqr *l1, sqr *l2, sqr *l3, sqr *l4, bool isceil)  // floor/ceil quads
+void render_flat(int wtex, int x, int y, int size, int h, sqr *l1, sqr *l4, sqr *l3, sqr *l2, bool isceil)  // floor/ceil quads
 {
     if(showm) { l3 = l1 = &sbright; l4 = l2 = &sdark; }
 
@@ -156,7 +156,7 @@ void render_flat(int wtex, int x, int y, int size, int h, sqr *l1, sqr *l2, sqr
     float xo = xf*x;
     float yo = yf*y;
 
-    bool first = !floorstrip || y!=oy+size || striptex!=wtex || h!=oh || x!=ox;
+    bool first = !floorstrip || x!=ox+size || striptex!=wtex || h!=oh || y!=oy;
 
     if(first)       // start strip here
     {
@@ -164,17 +164,17 @@ void render_flat(int wtex, int x, int y, int size, int h, sqr *l1, sqr *l2, sqr
         firstindex = verts.length();
         striptex = wtex;
         oh = h;
-        ox = x;
+        oy = y;
         floorstrip = true;
         if(isceil)
         {
-            vert(x+size, y, h, l2, xo+xs, yo);
-            vert(x,      y, h, l1, xo, yo);
+            vert(x, y,      h, l1, xo, yo);
+            vert(x, y+size, h, l2, xo, yo+ys);
         }
         else
         {
-            vert(x,      y, h, l1, xo,    yo);
-            vert(x+size, y, h, l2, xo+xs, yo);
+            vert(x, y+size, h, l2, xo, yo+ys);
+            vert(x, y,      h, l1, xo, yo);
         }
         ol3r = l1->r;
         ol3g = l1->g;
@@ -208,20 +208,20 @@ void render_flat(int wtex, int x, int y, int size, int h, sqr *l1, sqr *l2, sqr
 
     if(isceil)
     {
+        vert(x+size, y,      h, l4, xo+xs, yo);
         vert(x+size, y+size, h, l3, xo+xs, yo+ys);
-        vert(x,      y+size, h, l4, xo,    yo+ys); 
     }
     else
     {
-        vert(x,      y+size, h, l4, xo,    yo+ys);
-        vert(x+size, y+size, h, l3, xo+xs, yo+ys); 
+        vert(x+size, y+size, h, l3, xo+xs, yo+ys);
+        vert(x+size, y,      h, l4, xo+xs, yo);
     }
 
-    oy = y;
+    ox = x;
     nquads++;
 }
 
-void render_flatdelta(int wtex, int x, int y, int size, float h1, float h2, float h3, float h4, sqr *l1, sqr *l2, sqr *l3, sqr *l4, bool isceil)  // floor/ceil quads on a slope
+void render_flatdelta(int wtex, int x, int y, int size, float h1, float h4, float h3, float h2, sqr *l1, sqr *l4, sqr *l3, sqr *l2, bool isceil)  // floor/ceil quads on a slope
 {
     if(showm) { l3 = l1 = &sbright; l4 = l2 = &sdark; }
 
@@ -233,24 +233,24 @@ void render_flatdelta(int wtex, int x, int y, int size, float h1, float h2, floa
     float xo = xf*x;
     float yo = yf*y;
 
-    bool first = !deltastrip || y!=oy+size || striptex!=wtex || x!=ox; 
+    bool first = !deltastrip || x!=ox+size || striptex!=wtex || y!=oy; 
 
     if(first) 
     {
         stripend(verts);
         firstindex = verts.length();
         striptex = wtex;
-        ox = x;
+        oy = y;
         deltastrip = true;
         if(isceil)
         {
-            vert(x+size, y, h2, l2, xo+xs, yo);
-            vert(x,      y, h1, l1, xo,    yo);
+            vert(x, y,      h1, l1, xo, yo);
+            vert(x, y+size, h2, l2, xo, yo+ys);
         }
         else
         {
-            vert(x,      y, h1, l1, xo,    yo);
-            vert(x+size, y, h2, l2, xo+xs, yo);
+            vert(x, y+size, h2, l2, xo, yo+ys);
+            vert(x, y,      h1, l1, xo, yo);
         }
         ol3r = l1->r;
         ol3g = l1->g;
@@ -262,16 +262,16 @@ void render_flatdelta(int wtex, int x, int y, int size, float h1, float h2, floa
 
     if(isceil)
     {
+        vert(x+size, y,      h4, l4, xo+xs, yo);
         vert(x+size, y+size, h3, l3, xo+xs, yo+ys); 
-        vert(x,      y+size, h4, l4, xo,    yo+ys);
     }
     else
     {
-        vert(x,      y+size, h4, l4, xo,    yo+ys);
-        vert(x+size, y+size, h3, l3, xo+xs, yo+ys); 
+        vert(x+size, y+size, h3, l3, xo+xs, yo+ys);
+        vert(x+size, y,      h4, l4, xo+xs, yo);
     }
 
-    oy = y;
+    ox = x;
     nquads++;
 }
 
@@ -422,7 +422,7 @@ void rendershadow_tris(int x, int y, bool topleft, sqr *h1, sqr *h2)
 
 static void rendershadow_flat(int x, int y, int h) // floor quads
 {
-    bool first = !floorstrip || y!=oy+1 || h!=oh || x!=ox;
+    bool first = !floorstrip || x!=ox+1 || h!=oh || y!=oy;
 
     if(first)       // start strip here
     {
@@ -430,41 +430,41 @@ static void rendershadow_flat(int x, int y, int h) // floor quads
         firstindex = shadowverts.length();
         striptex = DEFAULT_FLOOR;
         oh = h;
-        ox = x;
+        oy = y;
         floorstrip = true;
-        shadowvert(x,   y, h);
-        shadowvert(x+1, y, h);
+        shadowvert(x, y+1, h);
+        shadowvert(x, y,   h);
     }
     else        // continue strip
     {
         shadowverts.setsizenodelete(shadowverts.length()-2);
     }
 
-    shadowvert(x,   y+1, h);
     shadowvert(x+1, y+1, h);
+    shadowvert(x+1, y,   h);
 
-    oy = y;
+    ox = x;
 }
 
-static void rendershadow_flatdelta(int x, int y, float h1, float h2, float h3, float h4)  // floor quads on a slope
+static void rendershadow_flatdelta(int x, int y, float h1, float h4, float h3, float h2)  // floor quads on a slope
 {
-    bool first = !deltastrip || y!=oy+1 || x!=ox;
+    bool first = !deltastrip || x!=ox+1 || y!=oy;
 
     if(first)
     {
         stripend(shadowverts);
         firstindex = shadowverts.length();
         striptex = DEFAULT_FLOOR;
-        ox = x;
+        oy = y;
         deltastrip = true;
-        shadowvert(x,   y, h1);
-        shadowvert(x+1, y, h2);
+        shadowvert(x, y+1, h2);
+        shadowvert(x, y,   h1);
     }
 
-    shadowvert(x,   y+1, h4);
     shadowvert(x+1, y+1, h3);
+    shadowvert(x+1, y,   h4);
 
-    oy = y;
+    ox = x;
 }
 
 void rendershadow(int x, int y, int xs, int ys, const vec &texgenS, const vec &texgenT)
@@ -479,7 +479,7 @@ void rendershadow(int x, int y, int xs, int ys, const vec &texgenS, const vec &t
     #define df(x) s->floor-(x->vdelta/4.0f)
 
     sqr *w = wmip[0];
-    for(int xx = x; xx<xs; xx++) for(int yy = y; yy<ys; yy++)
+    for(int yy = y; yy<ys; yy++) for(int xx = x; xx<xs; xx++)
     {
         sqr *s = SW(w,xx,yy);
         if(s->type==SPACE || s->type==CHF)
diff --git a/source/src/rndmap.cpp b/source/src/rndmap.cpp
index ea082fdb..592f3edb 100644
--- a/source/src/rndmap.cpp
+++ b/source/src/rndmap.cpp
@@ -59,7 +59,7 @@ void perlinarea(block &b, int scale, int seed, int psize)
     srand(seed);
     seed = rnd(10000);
     if(!scale) scale = 10;
-    for(int x = b.x; x<=b.x+b.xs; x++) for(int y = b.y; y<=b.y+b.ys; y++)
+    for(int y = b.y; y<=b.y+b.ys; y++) for(int x = b.x; x<=b.x+b.xs; x++)
     {
         sqr *s = S(x,y);
         if(!SOLID(s) && x!=b.x+b.xs && y!=b.y+b.ys) s->type = FHF; 
diff --git a/source/src/world.cpp b/source/src/world.cpp
index 423be36c..0cc8c791 100644
--- a/source/src/world.cpp
+++ b/source/src/world.cpp
@@ -29,7 +29,7 @@ void remip(const block &b, int level)
     if(s.y&1) { s.y--; s.ys++; }
     s.xs = (s.xs+1)&~1;
     s.ys = (s.ys+1)&~1;
-    for(int x = s.x; x<s.x+s.xs; x+=2) for(int y = s.y; y<s.y+s.ys; y+=2)
+    for(int y = s.y; y<s.y+s.ys; y+=2) for(int x = s.x; x<s.x+s.xs; x+=2)
     {
         sqr *o[4];
         o[0] = SWS(w,x,y,wfactor);                               // the 4 constituent cubes
diff --git a/source/src/world.h b/source/src/world.h
index 51870417..6c3c2249 100644
--- a/source/src/world.h
+++ b/source/src/world.h
@@ -20,6 +20,7 @@ struct sqr
     char occluded;              // true when occluded
     uchar utex;                 // upper wall tex id
     uchar tag;                  // used by triggers
+    uchar reserved[2];
 };
 
 enum                            // hardcoded texture numbers
diff --git a/source/src/worldio.cpp b/source/src/worldio.cpp
index 34da1eae..cceb73a7 100644
--- a/source/src/worldio.cpp
+++ b/source/src/worldio.cpp
@@ -49,7 +49,7 @@ inline bool nhf(sqr *s) { return s->type!=FHF && s->type!=CHF; }
 
 void voptimize()        // reset vdeltas on non-hf cubes
 {
-    loop(x, ssize) loop(y, ssize)
+    loop(y, ssize) loop(x, ssize)
     {
         sqr *s = S(x, y);
         if(x && y) { if(nhf(s) && nhf(S(x-1, y)) && nhf(S(x-1, y-1)) && nhf(S(x, y-1))) s->vdelta = 0; }
@@ -90,7 +90,7 @@ void toptimize() // FIXME: only does 2x2, make atleast for 4x4 also
 {
     bool wf[4], uf[4];
     sqr *s[4];
-    for(int x = 2; x<ssize-4; x += 2) for(int y = 2; y<ssize-4; y += 2)
+    for(int y = 2; y<ssize-4; y += 2) for(int x = 2; x<ssize-4; x += 2)
     {
         s[0] = S(x,y);
         int wt = s[0]->wtex, ut = s[0]->utex;
@@ -382,7 +382,7 @@ bool load_world(char *mname)        // still supports all map formats that have
             {
                 if(type<0 || type>=MAXTYPE)
                 {
-                    conoutf("while reading map at %d: type %d out of range", k, type); 
+                    conoutf("while reading map at %d: type %d out of range", k, type);
                     gzclose(f);
                     f = NULL;
                     k--;
diff --git a/source/src/worldlight.cpp b/source/src/worldlight.cpp
index 702eebf4..13502677 100644
--- a/source/src/worldlight.cpp
+++ b/source/src/worldlight.cpp
@@ -316,14 +316,14 @@ block *blockcopy(const block &s)
     block *b = (block *)new uchar[sizeof(block)+s.xs*s.ys*sizeof(sqr)];
     *b = s;
     sqr *q = (sqr *)(b+1);
-    for(int x = s.x; x<s.xs+s.x; x++) for(int y = s.y; y<s.ys+s.y; y++) *q++ = *S(x,y);
+    for(int y = s.y; y<s.ys+s.y; y++) for(int x = s.x; x<s.xs+s.x; x++) *q++ = *S(x,y);
     return b;
 }
 
 void blockpaste(const block &b)
 {
     const sqr *q = (const sqr *)((&b)+1);
-    for(int x = b.x; x<b.xs+b.x; x++) for(int y = b.y; y<b.ys+b.y; y++) *S(x,y) = *q++;
+    for(int y = b.y; y<b.ys+b.y; y++) for(int x = b.x; x<b.xs+b.x; x++) *S(x,y) = *q++;
     remipmore(b);
 }
 
diff --git a/source/src/worldrender.cpp b/source/src/worldrender.cpp
index 2321b660..768c939e 100644
--- a/source/src/worldrender.cpp
+++ b/source/src/worldrender.cpp
@@ -118,7 +118,7 @@ void render_seg_new(float vx, float vy, float vh, int mip, int x, int y, int xs,
     int ry = vyy+lodbot;
 
     float fsize = (float)(1<<mip);
-    for(int ox = x; ox<xs; ox++) for(int oy = y; oy<ys; oy++)       // first collect occlusion information for this block
+    for(int oy = y; oy<ys; oy++) for(int ox = x; ox<xs; ox++)       // first collect occlusion information for this block
     {
         SWS(w,ox,oy,mfactor)->occluded = isoccluded(camera1->o.x, camera1->o.y, (float)(ox<<mip), (float)(oy<<mip), fsize);
     }
@@ -140,8 +140,8 @@ void render_seg_new(float vx, float vy, float vh, int mip, int x, int y, int xs,
     // also deferred, and render them recursively. Anything left (perfect mips and higher lods) we
     // render here.
 
-    #define LOOPH {for(int xx = x; xx<xs; xx++) for(int yy = y; yy<ys; yy++) { \
-                  sqr *s = SWS(w,xx,yy,mfactor); if(s->occluded==1) continue; \
+    #define LOOPH {for(int yy = y; yy<ys; yy++) for(int xx = x; xx<xs; xx++) { \
+                  sqr *s = SWS(w,xx,yy,mfactor); if(s->occluded) continue; \
                   if(s->defer && !s->occluded && mip && xx>=lx && xx<rx && yy>=ly && yy<ry)
     #define LOOPD sqr *t = SWS(s,1,0,mfactor); \
                   sqr *u = SWS(s,1,1,mfactor); \
@@ -149,10 +149,10 @@ void render_seg_new(float vx, float vy, float vh, int mip, int x, int y, int xs,
 
     LOOPH // floors
         {
-            int start = yy;
+            int start = xx;
             sqr *next;
-            while(yy<ys-1 && (next = SWS(w,xx,yy+1,mfactor))->defer && !next->occluded) yy++;    // collect 2xN rect of lower mip
-            render_seg_new(vx, vy, vh, mip-1, xx*2, start*2, xx*2+2, yy*2+2);
+            while(xx<xs-1 && (next = SWS(w,xx+1,yy,mfactor))->defer && !next->occluded) xx++;    // collect 2xN rect of lower mip
+            render_seg_new(vx, vy, vh, mip-1, start*2, yy*2, xx*2+2, yy*2+2);
             continue;
         }
         stats[mip]++;