gistfile1.diff

diff --git a/Makefile.wii b/Makefile.wii
index 89364c3..c6a6c38 100644
--- a/Makefile.wii
+++ b/Makefile.wii
@@ -14,6 +14,8 @@ CC = $(DEVKITPPC)/bin/powerpc-eabi-gcc
 CXX = $(DEVKITPPC)/bin/powerpc-eabi-g++
 LD = $(DEVKITPPC)/bin/powerpc-eabi-ld
 
+FREETYPECONFIG = $(DEVKITPRO)/portlibs/ppc/bin/freetype-config
+
 DOL_TARGET := retroarch.dol
 ELF_TARGET := retroarch.elf
 
@@ -21,10 +23,10 @@ INCLUDE := -I. -I$(DEVKITPRO)/libogc/include
 LIBDIRS := -L$(DEVKITPRO)/libogc/lib/wii -L.
 
 MACHDEP := -DGEKKO -DHW_RVL -mrvl -mcpu=750 -meabi -mhard-float
-CFLAGS += -Wall -std=gnu99 $(MACHDEP) $(INCLUDE)
-CXXFLAGS += -Wall -std=gnu99 $(MACHDEP) $(INCLUDE)
-LDFLAGS := $(MACHDEP)
-LIBS := -lfat -lretro -lwiiuse -logc -lbte -lfreetype
+CFLAGS += -Wall -std=gnu99 $(MACHDEP) $(INCLUDE) $(shell $(FREETYPECONFIG) --cflags)
+CXXFLAGS += -Wall -std=gnu99 $(MACHDEP) $(INCLUDE) $(shell $(FREETYPECONFIG) --cflags)
+LDFLAGS := $(MACHDEP) $(shell $(FREETYPECONFIG) --libs)
+LIBS := -lfat -lretro -lwiiuse -logc -lbte
 
 OBJ = console/griffin/griffin.o console/font.bmpobj console/rzlib/rzlib.o
 
diff --git a/wii/video.c b/wii/video.c
index 338f8b0..ac45565 100644
--- a/wii/video.c
+++ b/wii/video.c
@@ -25,7 +25,7 @@
 
 // All very hardcoded for now.
 
-static void *g_framebuf[3];
+static void *g_framebuf[2];
 static unsigned g_vi_framebuf;
 static unsigned g_render_framebuf;
 
@@ -45,7 +45,7 @@ static size_t display_list_size;
 static void retrace_callback(u32 retrace_count)
 {
    (void)retrace_count;
-   VIDEO_SetNextFramebuffer(g_framebuf[g_vi_framebuf % 3]);
+   VIDEO_SetNextFramebuffer(g_framebuf[g_vi_framebuf % 2]);
    VIDEO_Flush();
    if (g_vi_framebuf < g_render_framebuf)
       g_vi_framebuf++;
@@ -54,7 +54,7 @@ static void retrace_callback(u32 retrace_count)
 static void setup_video_mode(GXRModeObj *mode)
 {
    VIDEO_Configure(mode);
-   for (unsigned i = 0; i < 3; i++)
+   for (unsigned i = 0; i < 2; i++)
    {
       g_framebuf[i] = MEM_K0_TO_K1(SYS_AllocateFramebuffer(mode));
       VIDEO_ClearFrameBuffer(mode, g_framebuf[i], COLOR_BLACK);
@@ -171,7 +171,7 @@ void wii_video_deinit(void)
    VIDEO_SetBlack(true);
    VIDEO_Flush();
 
-   for (unsigned i = 0; i < 3; i++)
+   for (unsigned i = 0; i < 2; i++)
       free(MEM_K1_TO_K0(g_framebuf[i]));
 }
 
@@ -207,7 +207,7 @@ static void *wii_init(const video_info_t *video,
    src += pitch; \
 }
 
-static void update_texture(const uint32_t *src,
+static void update_texture2(const uint32_t *src,
       unsigned width, unsigned height, unsigned pitch)
 {
    pitch >>= 2;
@@ -225,12 +225,97 @@ static void update_texture(const uint32_t *src,
       BLIT_LINE(4)
       BLIT_LINE(6)
    }
-   
+
    init_texture(width << 1, height);
    DCFlushRange(g_tex.data, sizeof(g_tex.data));
    GX_InvalidateTexAll();
 }
 
+static void update_texture(const uint32_t *src,
+      unsigned width, unsigned height, unsigned pitch)
+{
+   register u32 tmp0, tmp1, tmp2, tmp3, line2, line2b, line3, line3b, line4, line4b, line5;
+   register u32 ormask=0x80008000u;
+   register u32 *dst = g_tex.data;
+   width &= ~3;
+   height &= ~3;
+
+   __asm__ __volatile__ (
+      "     srwi     %[width],%[width],2\n"
+      "     srwi     %[height],%[height],2\n"
+      "     subi     %[tmp3],%[dst],4\n"
+      "     mr       %[dst],%[tmp3]\n"
+      "     subi     %[dst],%[dst],4\n"
+      "     mr       %[line2],%[pitch]\n"
+      "     addi     %[line2b],%[line2],4\n"
+      "     mulli    %[line3],%[pitch],2\n"
+      "     addi     %[line3b],%[line3],4\n"
+      "     mulli    %[line4],%[pitch],3\n"
+      "     addi     %[line4b],%[line4],4\n"
+      "     mulli    %[line5],%[pitch],4\n"
+
+      "2:   mtctr    %[width]\n"
+      "     mr       %[tmp0],%[src]\n"
+
+      "1:   lwz      %[tmp1],0(%[src])\n"
+      "     or       %[tmp1],%[tmp1],%[ormask]\n"
+      "     stwu     %[tmp1],8(%[dst])\n"
+      "     lwz      %[tmp2],4(%[src])\n"
+      "     or       %[tmp2],%[tmp2],%[ormask]\n"
+      "     stwu     %[tmp2],8(%[tmp3])\n"
+
+      "     lwzx     %[tmp1],%[line2],%[src]\n"
+      "     or       %[tmp1],%[tmp1],%[ormask]\n"
+      "     stwu     %[tmp1],8(%[dst])\n"
+      "     lwzx     %[tmp2],%[line2b],%[src]\n"
+      "     or       %[tmp2],%[tmp2],%[ormask]\n"
+      "     stwu     %[tmp2],8(%[tmp3])\n"
+
+      "     lwzx     %[tmp1],%[line3],%[src]\n"
+      "     or       %[tmp1],%[tmp1],%[ormask]\n"
+      "     stwu     %[tmp1],8(%[dst])\n"
+      "     lwzx     %[tmp2],%[line3b],%[src]\n"
+      "     or       %[tmp2],%[tmp2],%[ormask]\n"
+      "     stwu     %[tmp2],8(%[tmp3])\n"
+
+      "     lwzx     %[tmp1],%[line4],%[src]\n"
+      "     or       %[tmp1],%[tmp1],%[ormask]\n"
+      "     stwu     %[tmp1],8(%[dst])\n"
+      "     lwzx     %[tmp2],%[line4b],%[src]\n"
+      "     or       %[tmp2],%[tmp2],%[ormask]\n"
+      "     stwu     %[tmp2],8(%[tmp3])\n"
+
+      "     addi     %[src],%[src],8\n"
+      "     bdnz     1b\n"
+      "     add      %[src],%[tmp0],%[line5]\n"
+      "     subic.   %[height],%[height],1\n"
+      "     bne      2b"
+      :
+         [tmp0] "=&b" (tmp0),
+         [tmp1] "=&b" (tmp1),
+         [tmp2] "=&b" (tmp2),
+         [tmp3] "=&b" (tmp3),
+         [line2] "=&b" (line2),
+         [line2b] "=&b" (line2b),
+         [line3] "=&b" (line3),
+         [line3b] "=&b" (line3b),
+         [line4] "=&b" (line4),
+         [line4b] "=&b" (line4b),
+         [line5] "=&b" (line5),
+         [dst] "+b" (dst)
+      :
+         [src] "b" (src),
+         [width] "b" (width),
+         [height] "b" (height),
+         [pitch] "b" (pitch),
+         [ormask] "b" (ormask)
+   );
+
+   init_texture(width, height);
+   DCFlushRange(g_tex.data, sizeof(g_tex.data));
+   GX_InvalidateTexAll();
+}
+
 static bool wii_frame(void *data, const void *frame,
       unsigned width, unsigned height, unsigned pitch,
       const char *msg)
@@ -242,7 +327,7 @@ static bool wii_frame(void *data, const void *frame,
    GX_CallDispList(display_list, display_list_size);
    GX_DrawDone();
 
-   GX_CopyDisp(g_framebuf[g_render_framebuf % 3], GX_TRUE);
+   GX_CopyDisp(g_framebuf[g_render_framebuf % 2], GX_TRUE);
    GX_Flush();
 
    g_render_framebuf++;