diff --git a/dlls/wined3d/device.c b/dlls/wined3d/device.c
index 1ce2287a871..8c98962e7b5 100644
--- a/dlls/wined3d/device.c
+++ b/dlls/wined3d/device.c
@@ -2669,7 +2669,7 @@ static HRESULT  WINAPI  IWineD3DDeviceImpl_SetTransform(IWineD3DDevice *iface, W
     /* Handle recording of state blocks */
     if (This->isRecordingState) {
         TRACE("Recording... not performing anything\n");
-        This->updateStateBlock->changed.transform[d3dts] = TRUE;
+        This->updateStateBlock->changed.transform[d3dts >> 5] |= 1 << (d3dts & 0x1f);
         This->updateStateBlock->transforms[d3dts] = *lpmatrix;
         return WINED3D_OK;
     }
@@ -4752,10 +4752,15 @@ static HRESULT WINAPI IWineD3DDeviceImpl_EndStateBlock(IWineD3DDevice *iface, IW
             object->num_contained_render_states++;
         }
     }
-    for(i = 1; i <= HIGHEST_TRANSFORMSTATE; i++) {
-        if(object->changed.transform[i]) {
-            object->contained_transform_states[object->num_contained_transform_states] = i;
-            object->num_contained_transform_states++;
+
+    for (i = 0; i <= HIGHEST_TRANSFORMSTATE >> 5; ++i)
+    {
+        DWORD map = object->changed.transform[i];
+        for (j = 0; map; map >>= 1, ++j)
+        {
+            if (!(map & 1)) continue;
+
+            object->contained_transform_states[object->num_contained_transform_states++] = (i << 5) | j;
         }
     }
     for(i = 0; i < GL_LIMITS(vshader_constantsF); i++) {
diff --git a/dlls/wined3d/stateblock.c b/dlls/wined3d/stateblock.c
index 1ffdabe1933..d5206c7349f 100644
--- a/dlls/wined3d/stateblock.c
+++ b/dlls/wined3d/stateblock.c
@@ -89,7 +89,7 @@ static void stateblock_savedstates_copy(IWineD3DStateBlock* iface, SAVEDSTATES *
     dest->streamSource = source->streamSource;
     dest->streamFreq = source->streamFreq;
     dest->textures = source->textures;
-    memcpy(dest->transform, source->transform, bsize * (HIGHEST_TRANSFORMSTATE + 1));
+    memcpy(dest->transform, source->transform, sizeof(source->transform));
     memcpy(dest->renderState, source->renderState, bsize * (WINEHIGHEST_RENDER_STATE + 1));
     memcpy(dest->textureState, source->textureState, bsize * MAX_TEXTURES * (WINED3D_HIGHEST_TEXTURE_STATE + 1));
     memcpy(dest->samplerState, source->samplerState, bsize * MAX_COMBINED_SAMPLERS * (WINED3D_HIGHEST_SAMPLER_STATE + 1));
@@ -104,6 +104,13 @@ static void stateblock_savedstates_copy(IWineD3DStateBlock* iface, SAVEDSTATES *
     memcpy(dest->vertexShaderConstantsF, source->vertexShaderConstantsF, bsize * GL_LIMITS(vshader_constantsF));
 }
 
+static inline void stateblock_set_bits(DWORD *map, UINT map_size)
+{
+    DWORD mask = (1 << (map_size & 0x1f)) - 1;
+    memset(map, 0xff, (map_size >> 5) * sizeof(*map));
+    if (mask) map[map_size >> 5] = mask;
+}
+
 /** Set all members of a stateblock savedstate to the given value */
 void stateblock_savedstates_set(
     IWineD3DStateBlock* iface,
@@ -126,7 +133,8 @@ void stateblock_savedstates_set(
     states->streamSource = value ? 0xffff : 0;
     states->streamFreq = value ? 0xffff : 0;
     states->textures = value ? 0xfffff : 0;
-    memset(states->transform, value, bsize * (HIGHEST_TRANSFORMSTATE + 1));
+    if (value) stateblock_set_bits(states->transform, HIGHEST_TRANSFORMSTATE + 1);
+    else memset(states->transform, 0, sizeof(states->transform));
     memset(states->renderState, value, bsize * (WINEHIGHEST_RENDER_STATE + 1));
     memset(states->textureState, value, bsize * MAX_TEXTURES * (WINED3D_HIGHEST_TEXTURE_STATE + 1));
     memset(states->samplerState, value, bsize * MAX_COMBINED_SAMPLERS * (WINED3D_HIGHEST_SAMPLER_STATE + 1));
diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
index 90f8ee147a5..b9740776146 100644
--- a/dlls/wined3d/wined3d_private.h
+++ b/dlls/wined3d/wined3d_private.h
@@ -1758,10 +1758,10 @@ extern const IWineD3DVertexDeclarationVtbl IWineD3DVertexDeclaration_Vtbl;
 /*   Note: Very long winded but gl Lists are not flexible enough */
 /*   to resolve everything we need, so doing it manually for now */
 typedef struct SAVEDSTATES {
+    DWORD transform[(HIGHEST_TRANSFORMSTATE >> 5) + 1];
     WORD streamSource;              /* MAX_STREAMS, 16 */
     WORD streamFreq;                /* MAX_STREAMS, 16 */
     DWORD textures;                 /* MAX_COMBINED_SAMPLERS, 20 */
-    BOOL transform[HIGHEST_TRANSFORMSTATE + 1];
     BOOL renderState[WINEHIGHEST_RENDER_STATE + 1];
     BOOL textureState[MAX_TEXTURES][WINED3D_HIGHEST_TEXTURE_STATE + 1];
     BOOL samplerState[MAX_COMBINED_SAMPLERS][WINED3D_HIGHEST_SAMPLER_STATE + 1];