lots of optimization, new features

7 files changed, 268 insertions, 107 deletions

diff --git a/Makefile b/Makefile
index 476b8f7..00c3bcf 100644
--- a/Makefile
+++ b/Makefile
@@ -3,8 +3,8 @@ CC = gcc
 LD = ld
 INCLUDES = -Idrivers/include -Ikernel/include -Ilibc/include -Iapps/
 
-CFLAGS := -std=gnu99 -ffreestanding -fno-stack-protector -nostdlib -Wall -Wextra -m32 $(INCLUDES) -g -O1 -mtune=generic -msse
-CFLAGS += -fexpensive-optimizations -ftree-loop-vectorize -finline-functions -fomit-frame-pointer -ftree-vectorize
+CFLAGS = -std=gnu99 -ffreestanding -fno-stack-protector -nostdlib -Wall -Wextra -m32 $(INCLUDES) -g
+OPTFLAGS = -fexpensive-optimizations -ftree-loop-vectorize -finline-functions -fomit-frame-pointer -ftree-vectorize -O1 -mtune=generic
 
 QEMU = qemu-system-i386
 BOCHS = bochs
@@ -37,13 +37,19 @@ iso: kappa.bin
 kappa.bin: $(OBJ) $(SOURCES) Makefile
 	@$(LD) -T kernel/linker.ld -o kappa.bin -melf_i386 $(OBJ)
 	@echo "LD $@"
+
+drivers/gfx.o: drivers/gfx.c Makefile
+	@echo "CC $<"
+	@$(CC) $(CFLAGS) -O3 -msse -c $< -o $@
+
 %.o: %.c Makefile
-	@$(CC) $(CFLAGS) -c $< -o $@
 	@echo "CC $<"
+	@$(CC) $(CFLAGS) $(OPTFLAGS) -c $< -o $@
+
 
 %.o: %.S Makefile
-	@$(AS) $(ASFLAGS) -c $< -o $@
 	@echo "AS $<"
+	@$(AS) $(ASFLAGS) -c $< -o $@
 clean:
 	@echo "Cleaning build directory..."
 	@rm -f $(OBJ) kappa.iso kappa.bin
diff --git a/drivers/gfx-as.S b/drivers/gfx-as.S
index 70009c1..51c1db8 100644
--- a/drivers/gfx-as.S
+++ b/drivers/gfx-as.S
@@ -17,11 +17,73 @@ gfx_clear:
         addl    %eax, %edx
         cmpl    %edx, %eax
         jnb     .L2
-.L6:
+.L1:
         movl    %ebx, (%eax)
         addl    %ecx, %eax
         cmpl    %eax, %edx
-        ja      .L6
+        ja      .L1
 .L2:
         popl    %ebx
         ret
+
+        .global gfx_hline
+gfx_hline:
+        pushl   %ebx
+        movl    8(%esp), %eax
+        movl    12(%esp), %ecx
+        cmpl    %eax, %ecx
+        jge     .L3
+        xchgl   %eax, %ecx
+.L3:
+        movzwl  fb_stride, %edx
+        movzbl  fb_bpp, %ebx
+        imull   16(%esp), %edx
+        addl    framebuffer, %edx
+        imull   %ebx, %ecx
+        imull   %ebx, %eax
+        addl    %edx, %eax
+        addl    %ecx, %edx
+        movl    _gfx_fgcol, %ecx
+        cmpl    %edx, %eax
+        jnb     .L5
+.L4:
+        movl    %ecx, (%eax)
+        addl    %ebx, %eax
+        cmpl    %eax, %edx
+        ja      .L4
+.L5:
+        popl    %ebx
+        ret
+
+        .global gfx_vline
+gfx_vline:
+        pushl   %esi
+        pushl   %ebx
+        movl    12(%esp), %eax
+        movl    16(%esp), %edx
+        cmpl    %eax, %edx
+        jge     .L6
+        xchgl   %eax, %edx
+.L6:
+        movzwl  fb_stride, %ebx
+        movzbl  fb_bpp, %ecx
+        imull   20(%esp), %ecx
+        movl    framebuffer, %esi
+        imull   %ebx, %edx
+        imull   %ebx, %eax
+        addl    %ecx, %eax
+        addl    %edx, %ecx
+        leal    (%esi,%ecx), %edx
+        addl    %esi, %eax
+        movl    _gfx_fgcol, %ecx
+        cmpl    %edx, %eax
+        jnb     .L8
+.L7:
+        movl    %ecx, (%eax)
+        addl    %ebx, %eax
+        cmpl    %eax, %edx
+        ja      .L7
+.L8:
+        popl    %ebx
+        popl    %esi
+        ret
diff --git a/drivers/gfx.c b/drivers/gfx.c
index 8dc9d74..e0c5f17 100644
--- a/drivers/gfx.c
+++ b/drivers/gfx.c
@@ -15,6 +15,7 @@ uint16_t fb_height;
 
 /* this is BYTES per pixel */
 uint8_t  fb_bpp;
+uint16_t fb_stride;
 const uint8_t *gfx_bpp = &fb_bpp;
 
 const uint16_t *gfx_width = &fb_width;
@@ -73,19 +74,18 @@ void gfx_reset(void)
 
 void gfx_drawchar(int x, int y, char c)
 {
-    int stride = fb_bpp * fb_width;
-    uint8_t *line_addr = framebuffer + (x * fb_bpp) + (y * stride);
+    uint8_t *line_addr = framebuffer + (x * fb_bpp) + (y * fb_stride);
+    const uint32_t bg = _gfx_bgcol;
+    const uint32_t fg = _gfx_fgcol;
     for(int i = 0; i < FONT_HEIGHT; ++i)
     {
-        uint32_t line_buf[8] = {_gfx_bgcol};
-        uint8_t mask = 0x80;
-        for(int j = 0; j < 8; ++j, mask >>= 1)
+        uint8_t mask_table[8] = {128, 64, 32, 16, 8, 4, 2, 1};
+        for(int j = 0; j < 8; ++j)
         {
-            if(gfx_font[(int)c][i] & mask)
-                line_buf[j] = _gfx_fgcol;
+            if(gfx_font[c][i] & mask_table[j])
+                ((uint32_t*)line_addr)[j] = fg;
         }
-        memcpy(line_addr, line_buf, sizeof(line_buf));
-        line_addr += stride;
+        line_addr += fb_stride;
     }
 }
 
@@ -126,12 +126,67 @@ void gfx_puts(const char* str)
     }
 }
 
+/* implemented in assembly */
+#if 0
+void gfx_hline(int x1, int x2, int y)
+{
+    /* make sure x1 is to the left of x2 */
+    if(x2 < x1)
+    {
+        int temp = x1;
+        x1 = x2;
+        x2 = temp;
+    }
+
+    uint8_t *base = framebuffer + y * fb_stride;
+
+    uint8_t *dest = base + x1 * fb_bpp;
+    uint8_t *stop = base + x2 * fb_bpp;
+    const uint32_t col = _gfx_fgcol;
+    while(dest < stop)
+    {
+        *(uint32_t*)dest = col;
+        dest += fb_bpp;
+    }
+}
+
+void gfx_vline(int y1, int y2, int x)
+{
+    /* make sure y1 is above y2 */
+    if(y2 < y1)
+    {
+        int temp = y1;
+        y1 = y2;
+        y2 = temp;
+    }
+
+    uint8_t *dest = framebuffer + y1 * fb_stride + x * fb_bpp;
+    uint8_t *stop = framebuffer + y2 * fb_stride + x * fb_bpp;
+    const uint32_t col = _gfx_fgcol;
+    while(dest < stop)
+    {
+        *(uint32_t*)dest = col;
+        dest += fb_stride;
+    }
+}
+#endif
+
+#define SWAP(x, y, tmp) (tmp=x;x=y;y=tmp;)
+void gfx_fillrect(int x, int y, int w, int h)
+{
+    for(int i = 0; i < h; ++i)
+    {
+        gfx_hline(x, x + w, y + i);
+    }
+}
+
 bool gfx_init(struct vbe_info_t *vbe_mode_info)
 {
     framebuffer = (uint8_t*)vbe_mode_info->physbase;
     fb_width = vbe_mode_info->Xres;
     fb_height = vbe_mode_info->Yres;
     fb_bpp = vbe_mode_info->bpp / 8;
+    fb_stride = fb_bpp * fb_width;
     if(fb_bpp != 4)
     {
         printf("WARNING: BPP != 32, falling back to text mode...\n");
diff --git a/drivers/include/gfx.h b/drivers/include/gfx.h
index 42ada32..561eb3e 100644
--- a/drivers/include/gfx.h
+++ b/drivers/include/gfx.h
@@ -54,6 +54,12 @@ void gfx_set_background(uint32_t);
 
 uint32_t gfx_get_background(void);
 
+void gfx_hline(int x1, int x2, int y);
+
+void gfx_vline(int y1, int y2, int x);
+
+void gfx_fillrect(int x1, int y1, int w, int h);
+
 extern const uint16_t *gfx_width, *gfx_height;
 
 /* this is _BYTES_ per pixel, NOT BITS per pixel! */
diff --git a/kernel/irq-as.S b/kernel/irq-as.S
index 12de4b6..8840299 100644
--- a/kernel/irq-as.S
+++ b/kernel/irq-as.S
@@ -44,96 +44,96 @@ irq_stub:
 
 _irq0:
         cli
-        push $0
-        push $32
+        pushl $0
+        pushl $32
         jmp irq_stub
 
 _irq1:
         cli
-        push $0
-        push $33
+        pushl $0
+        pushl $33
         jmp irq_stub
 
 _irq2:
         cli
-        push $0
-        push $34
+        pushl $0
+        pushl $34
         jmp irq_stub
 
 _irq3:
         cli
-        push $0
-        push $35
+        pushl $0
+        pushl $35
         jmp irq_stub
 
 _irq4:
         cli
-        push $0
-        push $36
+        pushl $0
+        pushl $36
         jmp irq_stub
 
 _irq5:
         cli
-        push $0
-        push $37
+        pushl $0
+        pushl $37
         jmp irq_stub
 
 _irq6:
         cli
-        push $0
-        push $38
+        pushl $0
+        pushl $38
         jmp irq_stub
 
 _irq7:
         cli
-        push $0
-        push $39
+        pushl $0
+        pushl $39
         jmp irq_stub
 
 _irq8:
         cli
-        push $0
-        push $40
+        pushl $0
+        pushl $40
         jmp irq_stub
 
 _irq9:
         cli
-        push $0
-        push $41
+        pushl $0
+        pushl $41
         jmp irq_stub
 
 _irq10:
         cli
-        push $0
-        push $42
+        pushl $0
+        pushl $42
         jmp irq_stub
 
 _irq11:
         cli
-        push $0
-        push $43
+        pushl $0
+        pushl $43
         jmp irq_stub
 
 _irq12:
         cli
-        push $0
-        push $44
+        pushl $0
+        pushl $44
         jmp irq_stub
 
 _irq13:
         cli
-        push $0
-        push $45
+        pushl $0
+        pushl $45
         jmp irq_stub
 
 _irq14:
         cli
-        push $0
-        push $46
+        pushl $0
+        pushl $46
         jmp irq_stub
 
 _irq15:
         cli
-        push $0
-        push $47
+        pushl $0
+        pushl $47
         jmp irq_stub
diff --git a/kernel/isr-as.S b/kernel/isr-as.S
index 464ec5b..858ce47 100644
--- a/kernel/isr-as.S
+++ b/kernel/isr-as.S
@@ -60,155 +60,155 @@ isr_stub:
 
 _isr0:
     cli
-    push $0
-    push $0
+    pushl $0
+    pushl $0
     jmp isr_stub
 _isr1:
     cli
-    push $0
-    push $1
+    pushl $0
+    pushl $1
     jmp isr_stub
 _isr2:
     cli
-    push $0
-    push $2
+    pushl $0
+    pushl $2
     jmp isr_stub
 _isr3:
     cli
-    push $0
-    push $3
+    pushl $0
+    pushl $3
     jmp isr_stub
 _isr4:
     cli
-    push $0
-    push $4
+    pushl $0
+    pushl $4
     jmp isr_stub
 _isr5:
     cli
-    push $0
-    push $5
+    pushl $0
+    pushl $5
     jmp isr_stub
 _isr6:
     cli
-    push $0
-    push $6
+    pushl $0
+    pushl $6
     jmp isr_stub
 _isr7:
     cli
-    push $0
-    push $7
+    pushl $0
+    pushl $7
     jmp isr_stub
 _isr8:
     cli
-    push $8
+    pushl $8
     jmp isr_stub
 _isr9:
     cli
-    push $0
-    push $9
+    pushl $0
+    pushl $9
     jmp isr_stub
 _isr10:
     cli
-    push $10
+    pushl $10
     jmp isr_stub
 _isr11:
     cli
-    push $11
+    pushl $11
     jmp isr_stub
 _isr12:
     cli
-    push $12
+    pushl $12
     jmp isr_stub
 _isr13:
     cli
-    push $13
+    pushl $13
     jmp isr_stub
 _isr14:
     cli
-    push $14
+    pushl $14
     jmp isr_stub
 _isr15:
     cli
-    push $0
-    push $15
+    pushl $0
+    pushl $15
     jmp isr_stub
 _isr16:
     cli
-    push $0
-    push $16
+    pushl $0
+    pushl $16
     jmp isr_stub
 _isr17:
     cli
-    push $0
-    push $17
+    pushl $0
+    pushl $17
     jmp isr_stub
 _isr18:
     cli
-    push $0
-    push $18
+    pushl $0
+    pushl $18
     jmp isr_stub
 _isr19:
     cli
-    push $0
-    push $19
+    pushl $0
+    pushl $19
     jmp isr_stub
 _isr20:
     cli
-    push $0
-    push $20
+    pushl $0
+    pushl $20
     jmp isr_stub
 _isr21:
     cli
-    push $0
-    push $21
+    pushl $0
+    pushl $21
     jmp isr_stub
 _isr22:
     cli
-    push $0
-    push $22
+    pushl $0
+    pushl $22
     jmp isr_stub
 _isr23:
     cli
-    push $0
-    push $23
+    pushl $0
+    pushl $23
     jmp isr_stub
 _isr24:
     cli
-    push $0
-    push $24
+    pushl $0
+    pushl $24
     jmp isr_stub
 _isr25:
     cli
-    push $0
-    push $25
+    pushl $0
+    pushl $25
     jmp isr_stub
 _isr26:
     cli
-    push $0
-    push $26
+    pushl $0
+    pushl $26
     jmp isr_stub
 _isr27:
     cli
-    push $0
-    push $27
+    pushl $0
+    pushl $27
     jmp isr_stub
 _isr28:
     cli
-    push $0
-    push $28
+    pushl $0
+    pushl $28
     jmp isr_stub
 _isr29:
     cli
-    push $0
-    push $29
+    pushl $0
+    pushl $29
     jmp isr_stub
 _isr30:
     cli
-    push $0
-    push $30
+    pushl $0
+    pushl $30
     jmp isr_stub
 _isr31:
     cli
-    push $0
-    push $31
+    pushl $0
+    pushl $31
     jmp isr_stub
diff --git a/kernel/main.c b/kernel/main.c
index 927a497..211be18 100644
--- a/kernel/main.c
+++ b/kernel/main.c
@@ -19,7 +19,7 @@ void gpf(struct regs_t regs)
 {
     printf("General protection fault!\n");
     printf("EIP before fault: 0x%x\n", regs.eip);
-    printf("Error code: %d\n", regs.err_code);
+    printf("Error code: 0x%x\n", regs.err_code);
     panic("GPF");
 }
 
@@ -78,6 +78,7 @@ void main(struct multiboot_info_t *hdr, uint32_t magic)
             gfx_drawpixel(rx, ry);
         }
         int endpix = *current_tick;
+
         int startfill = *current_tick;
         for(int i=0;i<1000;++i)
         {
@@ -97,11 +98,43 @@ void main(struct multiboot_info_t *hdr, uint32_t magic)
             gfx_drawchar(rx, ry, rand()%127+1);
         }
         int endtext = *current_tick;
+        gfx_clear();
+
+        int starthline = *current_tick;
+        for(int i=0;i<1000000;++i)
+        {
+            gfx_hline(rand() % *gfx_width, rand() % *gfx_width, rand() % *gfx_height);
+        }
+        int endhline = *current_tick;
+
+        gfx_clear();
+
+        int startvline = *current_tick;
+        for(int i=0;i<1000000;++i)
+        {
+            gfx_vline(rand() % *gfx_height, rand() % *gfx_height, rand() % *gfx_width);
+        }
+        int endvline = *current_tick;
+
+        int startrect = *current_tick;
+        for(int i=0;i<1000;++i)
+        {
+            int x = rand() % *gfx_width;
+            int y = rand() % *gfx_height;
+            int w = rand() % (*gfx_width - x);
+            int h = rand() % (*gfx_height - y);
+            gfx_fillrect(x, y, w, h);
+        }
+        int endrect = *current_tick;
+
         gfx_reset();
         printf("--- Graphics benchmark results ---\n");
         printf("Ticks for 1,000,000 random pixels: %d\n", endpix-startpix);
         printf("Ticks for 1,000 random fills:      %d\n", endfill-startfill);
         printf("Ticks for 1,000,000 random chars:  %d\n", endtext-starttext);
+        printf("Ticks for 1,000,000 random hlines: %d\n", endhline-starthline);
+        printf("Ticks for 1,000,000 random vlines: %d\n", endvline-startvline);
+        printf("Ticks for 1,000 random rects:      %d\n", endrect-startrect);
         printf("Ticks per second:                  %d\n", HZ);
         printf("Resolution: %dx%dx%d\n", *gfx_width, *gfx_height, *gfx_bpp * 8);
     }
@@ -119,5 +152,4 @@ void main(struct multiboot_info_t *hdr, uint32_t magic)
         ps2_set_leds(PS2_CAPS_LOCK);
         timer_delay(HZ/4);
     }
-    while(1);
 }