diff --git a/Quake/gl_model.c b/Quake/gl_model.c
index ea07d850..2163e03a 100644
--- a/Quake/gl_model.c
+++ b/Quake/gl_model.c
@@ -943,10 +943,26 @@ void CalcSurfaceExtents (msurface_t *s)
 
 		for (j=0 ; j<2 ; j++)
 		{
-			val = v->position[0] * tex->vecs[j][0] +
-				v->position[1] * tex->vecs[j][1] +
-				v->position[2] * tex->vecs[j][2] +
-				tex->vecs[j][3];
+			/* The following calculation is sensitive to floating-point
+			 * precision.  It needs to produce the same result that the
+			 * light compiler does, because R_BuildLightMap uses surf->
+			 * extents to know the width/height of a surface's lightmap,
+			 * and incorrect rounding here manifests itself as patches
+			 * of "corrupted" looking lightmaps.
+			 * Most light compilers are win32 executables, so they use
+			 * x87 floating point.  This means the multiplies and adds
+			 * are done at 80-bit precision, and the result is rounded
+			 * down to 32-bits and stored in val.
+			 * Adding the casts to double seems to be good enough to fix
+			 * lighting glitches when Quakespasm is compiled as x86_64
+			 * and using SSE2 floating-point.  A potential trouble spot
+			 * is the hallway at the beginning of mfxsp17.  -- ericw
+			 */
+			val =	((double)v->position[0] * (double)tex->vecs[j][0]) +
+				((double)v->position[1] * (double)tex->vecs[j][1]) +
+				((double)v->position[2] * (double)tex->vecs[j][2]) +
+				(double)tex->vecs[j][3];
+
 			if (val < mins[j])
 				mins[j] = val;
 			if (val > maxs[j])