More performance improvements

data/dynos.cpp.h

@@ -291,44 +291,20 @@ public:
     inline bool Empty() const { return mCount == 0; }
 
 public:
-    bool operator==(const char *aString) const {
+    bool OPTIMIZE_O3 operator==(const char *aString) const {
-        if (strlen(aString) != mCount) return false;
+        return !strcmp(mBuffer, aString);
-        for (u8 i = 0; i != mCount; ++i) {
-            if (aString[i] != mBuffer[i]) {
-                return false;
-            }
-        }
-        return true;
     }
 
-    bool operator==(const String &aOther) const {
+    bool OPTIMIZE_O3 operator==(const String &aOther) const {
-        if (aOther.mCount != mCount) return false;
+        return !strcmp(mBuffer, aOther.mBuffer);
-        for (u8 i = 0; i != mCount; ++i) {
-            if (aOther.mBuffer[i] != mBuffer[i]) {
-                return false;
-            }
-        }
-        return true;
     }
 
-    bool operator!=(const char *aString) const {
+    bool OPTIMIZE_O3 operator!=(const char *aString) const {
-        if (strlen(aString) != mCount) return true;
+        return strcmp(mBuffer, aString);
-        for (u8 i = 0; i != mCount; ++i) {
-            if (aString[i] != mBuffer[i]) {
-                return true;
-            }
-        }
-        return false;
     }
 
-    bool operator!=(const String &aOther) const {
+    bool OPTIMIZE_O3 operator!=(const String &aOther) const {
-        if (aOther.mCount != mCount) return true;
+        return strcmp(mBuffer, aOther.mBuffer);
-        for (u8 i = 0; i != mCount; ++i) {
-            if (aOther.mBuffer[i] != mBuffer[i]) {
-                return true;
-            }
-        }
-        return false;
     }
 
 public:

src/pc/djui/djui_font.c

@@ -12,18 +12,20 @@ static Vtx djui_font_normal_vertices[] = {
     {{{ 0,   0, 0}, 0, { 512, 256}, { 0xff, 0xff, 0xff, 0xff }}},
 };
 
-const Gfx dl_font_normal_display_list[] = {
+const Gfx dl_font_normal_display_list_begin[] = {
     gsDPPipeSync(),
     gsSPClearGeometryMode(G_LIGHTING),
     gsDPSetCombineMode(G_CC_FADEA, G_CC_FADEA),
     gsDPSetRenderMode(G_RM_XLU_SURF, G_RM_XLU_SURF2),
     gsDPSetTextureFilter(G_TF_POINT),
     gsSPTexture(0xFFFF, 0xFFFF, 0, G_TX_RENDERTILE, G_ON),
-    gsDPSetTile(G_IM_FMT_IA, G_IM_SIZ_16b, 0, 0, G_TX_LOADTILE, 0, G_TX_WRAP | G_TX_NOMIRROR, 3, G_TX_NOLOD, G_TX_WRAP | G_TX_NOMIRROR, 4, G_TX_NOLOD),
-    gsDPLoadSync(),
-    gsDPLoadBlock(G_TX_LOADTILE, 0, 0, ((16 * 8 + G_IM_SIZ_4b_INCR) >> G_IM_SIZ_4b_SHIFT) - 1, CALC_DXT(16, G_IM_SIZ_4b_BYTES)),
     gsDPSetTile(G_IM_FMT_IA, G_IM_SIZ_4b, 1, 0, G_TX_RENDERTILE, 0, G_TX_WRAP | G_TX_NOMIRROR, 3, G_TX_NOLOD, G_TX_WRAP | G_TX_NOMIRROR, 4, G_TX_NOLOD),
     gsDPSetTileSize(0, 0, 0, (16 - 1) << G_TEXTURE_IMAGE_FRAC, (8 - 1) << G_TEXTURE_IMAGE_FRAC),
+    gsSPEndDisplayList(),
+};
+
+const Gfx dl_font_normal_display_list[] = {
+    gsDPLoadBlock(G_TX_LOADTILE, 0, 0, ((16 * 8 + G_IM_SIZ_4b_INCR) >> G_IM_SIZ_4b_SHIFT) - 1, CALC_DXT(16, G_IM_SIZ_4b_BYTES)),
     gsSPVertex(djui_font_normal_vertices, 4, 0),
     gsSPExecuteDjui(G_TEXCLIP_DJUI),
     gsSP2Triangles(0,  1,  2, 0x0, 0,  2,  3, 0x0),
@@ -54,7 +56,7 @@ static const struct DjuiFont sDjuiFontNormal = {
     .lineHeight           = 0.8125f,
     .defaultFontScale     = 32.0f,
     .rotatedUV            = true,
-    .textBeginDisplayList = NULL,
+    .textBeginDisplayList = dl_font_normal_display_list_begin,
     .render_char          = djui_font_normal_render_char,
     .char_width           = djui_font_normal_char_width,
 };

src/pc/gfx/gfx_pc.c

@@ -1866,7 +1866,7 @@ static uint32_t sDjuiOverrideW       = 0;
 static uint32_t sDjuiOverrideH       = 0;
 static uint32_t sDjuiOverrideB       = 0;
 
-static void djui_gfx_dp_execute_clipping(void) {
+static void OPTIMIZE_O3 djui_gfx_dp_execute_clipping(void) {
     if (!sDjuiClip) { return; }
     sDjuiClip = false;
 
@@ -1941,7 +1941,7 @@ static void djui_gfx_dp_execute_clipping(void) {
     }
 }
 
-static void djui_gfx_dp_execute_override(void) {
+static void OPTIMIZE_O3 djui_gfx_dp_execute_override(void) {
     if (!sDjuiOverride) { return; }
     sDjuiOverride = false;
 
@@ -1974,14 +1974,14 @@ static void djui_gfx_dp_execute_override(void) {
     rdp.textures_changed[1] = true;
 }
 
-static void djui_gfx_dp_execute_djui(uint32_t opcode) {
+static void OPTIMIZE_O3 djui_gfx_dp_execute_djui(uint32_t opcode) {
     switch (opcode) {
         case G_TEXOVERRIDE_DJUI: djui_gfx_dp_execute_override(); break;
         case G_TEXCLIP_DJUI:     djui_gfx_dp_execute_clipping(); break;
     }
 }
 
-static void djui_gfx_dp_set_clipping(bool rotatedUV, uint32_t x1, uint32_t y1, uint32_t x2, uint32_t y2) {
+static void OPTIMIZE_O3 djui_gfx_dp_set_clipping(bool rotatedUV, uint32_t x1, uint32_t y1, uint32_t x2, uint32_t y2) {
     sDjuiClipRotatedUV = rotatedUV;
     sDjuiClipX1 = x1;
     sDjuiClipY1 = y1;
@@ -1990,7 +1990,7 @@ static void djui_gfx_dp_set_clipping(bool rotatedUV, uint32_t x1, uint32_t y1, u
     sDjuiClip   = true;
 }
 
-static void djui_gfx_dp_set_override(void* texture, uint32_t w, uint32_t h, uint32_t b) {
+static void OPTIMIZE_O3 djui_gfx_dp_set_override(void* texture, uint32_t w, uint32_t h, uint32_t b) {
     sDjuiOverrideTexture = texture;
     sDjuiOverrideW = w;
     sDjuiOverrideH = h;
@@ -1998,7 +1998,7 @@ static void djui_gfx_dp_set_override(void* texture, uint32_t w, uint32_t h, uint
     sDjuiOverride  = (texture != NULL);
 }
 
-void djui_gfx_run_dl(Gfx* cmd) {
+void OPTIMIZE_O3 djui_gfx_run_dl(Gfx* cmd) {
     uint32_t opcode = cmd->words.w0 >> 24;
     switch (opcode) {
         case G_TEXCLIP_DJUI: