From f27394249ef3e3b219ac73a7b2259509ab5b6dbf Mon Sep 17 00:00:00 2001
From: "BADIM-PC\\Vadim" <grigoruk@gmail.com>
Date: Fri, 17 Nov 2017 23:37:37 +0300
Subject: [PATCH] blit api performance optimization

---
 src/tic.c   | 28 +++++++++++-----------------
 src/tools.c | 24 ++----------------------
 src/tools.h | 25 +++++++++++++++++++++++--
 3 files changed, 36 insertions(+), 41 deletions(-)

diff --git a/src/tic.c b/src/tic.c
index 99c1a11..5e0ab7c 100644
--- a/src/tic.c
+++ b/src/tic.c
@@ -1647,27 +1647,21 @@ static void api_blit(tic_mem* tic, u32* out, tic_scanline scanline)
 
 	memset4(&out[0 * TIC80_FULLWIDTH], pal[tic->ram.vram.vars.border], TIC80_FULLWIDTH*Top);
 
-	for(s32 r = 0; r < TIC80_HEIGHT; r++)
+	u32* rowPtr = out + Top*TIC80_FULLWIDTH;
+
+	for(s32 r = 0, y = tic->ram.vram.vars.offset.y; r < TIC80_HEIGHT; r++, y++, rowPtr += TIC80_FULLWIDTH)
 	{
-		memset4(&out[(r+Top) * TIC80_FULLWIDTH], pal[tic->ram.vram.vars.border], Left);
-		memset4(&out[(r+Top) * TIC80_FULLWIDTH + Left], pal[tic->ram.vram.vars.bg], TIC80_WIDTH);
+		memset4(rowPtr, pal[tic->ram.vram.vars.border], Left);
+		memset4(rowPtr + Left, pal[tic->ram.vram.vars.bg], TIC80_WIDTH);
 
-		{
-			s32 y = r + tic->ram.vram.vars.offset.y;
+		u32* colPtr = rowPtr + Left;
 
-			if(y < 0 || y >= TIC80_HEIGHT) continue;
-			
-			for(s32 c = 0; c < TIC80_WIDTH; c++)
-			{
-				s32 x = c + tic->ram.vram.vars.offset.x;
+		if(y >= 0 && y < TIC80_HEIGHT)
+			for(s32 c = 0, x = tic->ram.vram.vars.offset.x, index = y * TIC80_WIDTH + x; c < TIC80_WIDTH; c++, colPtr++, x++, index++)
+				if(x >= 0 && x < TIC80_WIDTH)
+					*colPtr = pal[tic_tool_peek4(tic->ram.vram.screen.data, index)];
 
-				if(x < 0 || x >= TIC80_WIDTH) continue;
-
-				out[(c + Left) + (r+Top) * TIC80_FULLWIDTH] = pal[tic_tool_peek4(tic->ram.vram.screen.data, x + y * TIC80_WIDTH)];
-			}			
-		}
-
-		memset4(&out[(r+Top) * TIC80_FULLWIDTH + (TIC80_FULLWIDTH-Right)], pal[tic->ram.vram.vars.border], Right);
+		memset4(rowPtr + (TIC80_FULLWIDTH-Right), pal[tic->ram.vram.vars.border], Right);
 
 		if(scanline && (r < TIC80_HEIGHT-1))
 		{
diff --git a/src/tools.c b/src/tools.c
index 05c7ea8..b4dab40 100644
--- a/src/tools.c
+++ b/src/tools.c
@@ -25,28 +25,8 @@
 
 #include <string.h>
 
-void tic_tool_poke4(void* addr, u32 index, u8 value)
-{
-	u8* val = (u8*)addr + (index >> 1);
-
-	if(index & 1)
-	{
-		*val &= 0x0f;
-		*val |= (value << 4);
-	}
-	else
-	{
-		*val &= 0xf0;
-		*val |= value & 0x0f;
-	}
-}
-
-u8 tic_tool_peek4(const void* addr, u32 index)
-{
-	u8 val = ((u8*)addr)[index >> 1];
-
-	return index & 1 ? val >> 4 : val & 0xf;
-}
+extern void tic_tool_poke4(void* addr, u32 index, u8 value);
+extern u8 tic_tool_peek4(const void* addr, u32 index);
 
 s32 tic_tool_get_pattern_id(const tic_track* track, s32 frame, s32 channel)
 {
diff --git a/src/tools.h b/src/tools.h
index bfa327b..50e5361 100644
--- a/src/tools.h
+++ b/src/tools.h
@@ -24,8 +24,29 @@
 
 #include "tic.h"
 
-void tic_tool_poke4(void* addr, u32 index, u8 value);
-u8 tic_tool_peek4(const void* addr, u32 index);
+inline void tic_tool_poke4(void* addr, u32 index, u8 value)
+{
+	u8* val = (u8*)addr + (index >> 1);
+
+	if(index & 1)
+	{
+		*val &= 0x0f;
+		*val |= (value << 4);
+	}
+	else
+	{
+		*val &= 0xf0;
+		*val |= value & 0x0f;
+	}
+}
+
+inline u8 tic_tool_peek4(const void* addr, u32 index)
+{
+	u8 val = ((u8*)addr)[index >> 1];
+
+	return index & 1 ? val >> 4 : val & 0xf;
+}
+
 bool tic_tool_parse_note(const char* noteStr, s32* note, s32* octave);
 s32 tic_tool_get_pattern_id(const tic_track* track, s32 frame, s32 channel);
 void tic_tool_set_pattern_id(tic_track* track, s32 frame, s32 channel, s32 id);