glyph plotter; Use repeated inclusion of header

To avoid having to duplicate a fairly large block of code several
times, use repeated inclusion of a header with some macros to
generate optimised glyph plotters.

3 files changed, 270 insertions, 138 deletions

diff --git a/platform/win32/libmupdf.vcproj b/platform/win32/libmupdf.vcproj
index 0fdf568d..3f7cfb69 100644
--- a/platform/win32/libmupdf.vcproj
+++ b/platform/win32/libmupdf.vcproj
@@ -971,6 +971,10 @@
 				>
 			</File>
 			<File
+				RelativePath="..\..\source\fitz\paint-glyph.h"
+				>
+			</File>
+			<File
 				RelativePath="..\..\source\fitz\path.c"
 				>
 			</File>
diff --git a/source/fitz/draw-paint.c b/source/fitz/draw-paint.c
index b2083bd6..33a72450 100644
--- a/source/fitz/draw-paint.c
+++ b/source/fitz/draw-paint.c
@@ -1155,6 +1155,7 @@ intermediate_run:
 	}
 }
 
+#if 0
 static inline void
 fz_paint_glyph_alpha_N(unsigned char *colorbv, int n, int span, unsigned char *dp, fz_glyph *glyph, int w, int h, int skip_x, int skip_y)
 {
@@ -1291,139 +1292,33 @@ intermediate_run:
 		dp += span;
 	}
 }
+#endif
 
-static inline void
-fz_paint_glyph_solid_N(unsigned char *colorbv, int n, int span, unsigned char *dp, fz_glyph *glyph, int w, int h, int skip_x, int skip_y)
-{
-	while (h--)
-	{
-		int skip_xx, ww, len, extend;
-		unsigned char *runp;
-		unsigned char *ddp = dp;
-		int offset = ((int *)(glyph->data))[skip_y++];
-		if (offset >= 0)
-		{
-			int eol = 0;
-			runp = &glyph->data[offset];
-			extend = 0;
-			ww = w;
-			skip_xx = skip_x;
-			while (skip_xx)
-			{
-				int v = *runp++;
-				switch (v & 3)
-				{
-				case 0: /* Extend */
-					extend = v>>2;
-					len = 0;
-					break;
-				case 1: /* Transparent */
-					len = (v>>2) + 1 + (extend<<6);
-					extend = 0;
-					if (len > skip_xx)
-					{
-						len -= skip_xx;
-						goto transparent_run;
-					}
-					break;
-				case 2: /* Solid */
-					eol = v & 4;
-					len = (v>>3) + 1 + (extend<<5);
-					extend = 0;
-					if (len > skip_xx)
-					{
-						len -= skip_xx;
-						goto solid_run;
-					}
-					break;
-				default: /* Intermediate */
-					eol = v & 4;
-					len = (v>>3) + 1 + (extend<<5);
-					extend = 0;
-					if (len > skip_xx)
-					{
-						runp += skip_xx;
-						len -= skip_xx;
-						goto intermediate_run;
-					}
-					runp += len;
-					break;
-				}
-				if (eol)
-				{
-					ww = 0;
-					break;
-				}
-				skip_xx -= len;
-			}
-			while (ww > 0)
-			{
-				int v = *runp++;
-				switch(v & 3)
-				{
-				case 0: /* Extend */
-					extend = v>>2;
-					break;
-				case 1: /* Transparent */
-					len = (v>>2) + 1 + (extend<<6);
-					extend = 0;
-transparent_run:
-					if (len > ww)
-						len = ww;
-					ww -= len;
-					ddp += len * n;
-					break;
-				case 2: /* Solid */
-					eol = v & 4;
-					len = (v>>3) + 1 + (extend<<5);
-					extend = 0;
-solid_run:
-					if (len > ww)
-						len = ww;
-					ww -= len;
-					do
-					{
-						int k = 0;
-						do
-						{
-							*ddp++ = colorbv[k++];
-						}
-						while (k != n);
-					}
-					while (--len);
-					break;
-				default: /* Intermediate */
-					eol = v & 4;
-					len = (v>>3) + 1 + (extend<<5);
-					extend = 0;
-intermediate_run:
-					if (len > ww)
-						len = ww;
-					ww -= len;
-					do
-					{
-						int k = 0;
-						int a = *runp++;
-						a = FZ_EXPAND(a);
-						do
-						{
-							*ddp = FZ_BLEND(colorbv[k++], *ddp, a);
-							ddp++;
-						}
-						while (k != n-1);
-						*ddp = FZ_BLEND(0xFF, *ddp, a);
-						ddp++;
-					}
-					while (--len);
-					break;
-				}
-				if (eol)
-					break;
-			}
-		}
-		dp += span;
-	}
-}
+#define N 2
+#include "paint-glyph.h"
+
+#define N 4
+#include "paint-glyph.h"
+
+#define N 5
+#include "paint-glyph.h"
+
+#include "paint-glyph.h"
+
+#define ALPHA
+#define N 2
+#include "paint-glyph.h"
+
+#define ALPHA
+#define N 4
+#include "paint-glyph.h"
+
+#define ALPHA
+#define N 5
+#include "paint-glyph.h"
+
+#define ALPHA
+#include "paint-glyph.h"
 
 static inline void
 fz_paint_glyph_alpha(unsigned char *colorbv, int n, int span, unsigned char *dp, fz_glyph *glyph, int w, int h, int skip_x, int skip_y)
@@ -1431,13 +1326,13 @@ fz_paint_glyph_alpha(unsigned char *colorbv, int n, int span, unsigned char *dp,
 	switch (n)
 	{
 	case 2:
-		fz_paint_glyph_alpha_N(colorbv, 2, span, dp, glyph, w, h, skip_x, skip_y);
+		fz_paint_glyph_alpha_2(colorbv, span, dp, glyph, w, h, skip_x, skip_y);
 		break;
 	case 4:
-		fz_paint_glyph_alpha_N(colorbv, 4, span, dp, glyph, w, h, skip_x, skip_y);
+		fz_paint_glyph_alpha_4(colorbv, span, dp, glyph, w, h, skip_x, skip_y);
 		break;
 	case 5:
-		fz_paint_glyph_alpha_N(colorbv, 5, span, dp, glyph, w, h, skip_x, skip_y);
+		fz_paint_glyph_alpha_5(colorbv, span, dp, glyph, w, h, skip_x, skip_y);
 		break;
 	default:
 		fz_paint_glyph_alpha_N(colorbv, n, span, dp, glyph, w, h, skip_x, skip_y);
@@ -1451,13 +1346,13 @@ fz_paint_glyph_solid(unsigned char *colorbv, int n, int span, unsigned char *dp,
 	switch (n)
 	{
 	case 2:
-		fz_paint_glyph_solid_N(colorbv, 2, span, dp, glyph, w, h, skip_x, skip_y);
+		fz_paint_glyph_solid_2(colorbv, span, dp, glyph, w, h, skip_x, skip_y);
 		break;
 	case 4:
-		fz_paint_glyph_solid_N(colorbv, 4, span, dp, glyph, w, h, skip_x, skip_y);
+		fz_paint_glyph_solid_4(colorbv, span, dp, glyph, w, h, skip_x, skip_y);
 		break;
 	case 5:
-		fz_paint_glyph_solid_N(colorbv, 5, span, dp, glyph, w, h, skip_x, skip_y);
+		fz_paint_glyph_solid_5(colorbv, span, dp, glyph, w, h, skip_x, skip_y);
 		break;
 	default:
 		fz_paint_glyph_solid_N(colorbv, n, span, dp, glyph, w, h, skip_x, skip_y);
diff --git a/source/fitz/paint-glyph.h b/source/fitz/paint-glyph.h
new file mode 100644
index 00000000..055693b2
--- /dev/null
+++ b/source/fitz/paint-glyph.h
@@ -0,0 +1,233 @@
+/*
+	This file is #included by draw-paint.c multiple times to
+	produce optimised plotters.
+*/
+
+#ifdef ALPHA
+#define NAME alpha
+#else
+#define NAME solid
+#endif
+
+#define FUNCTION_NAMER(NAME,N) fz_paint_glyph_##NAME##_##N
+#define FUNCTION_NAME(NAME,N) FUNCTION_NAMER(NAME,N)
+
+static inline void
+FUNCTION_NAME(NAME,N)(unsigned char *colorbv,
+#ifndef N
+				int n,
+#endif
+				int span, unsigned char *dp, fz_glyph *glyph, int w, int h, int skip_x, int skip_y)
+{
+#ifdef N
+	const int n = N;
+#endif
+#ifdef ALPHA
+	int sa = FZ_EXPAND(colorbv[n-1]);
+#endif
+	while (h--)
+	{
+		int skip_xx, ww, len, extend;
+		unsigned char *runp;
+		unsigned char *ddp = dp;
+		int offset = ((int *)(glyph->data))[skip_y++];
+		if (offset >= 0)
+		{
+			int eol = 0;
+			runp = &glyph->data[offset];
+			extend = 0;
+			ww = w;
+			skip_xx = skip_x;
+			while (skip_xx)
+			{
+				int v = *runp++;
+				switch (v & 3)
+				{
+				case 0: /* Extend */
+					extend = v>>2;
+					len = 0;
+					break;
+				case 1: /* Transparent */
+					len = (v>>2) + 1 + (extend<<6);
+					extend = 0;
+					if (len > skip_xx)
+					{
+						len -= skip_xx;
+						goto transparent_run;
+					}
+					break;
+				case 2: /* Solid */
+					eol = v & 4;
+					len = (v>>3) + 1 + (extend<<5);
+					extend = 0;
+					if (len > skip_xx)
+					{
+						len -= skip_xx;
+						goto solid_run;
+					}
+					break;
+				default: /* Intermediate */
+					eol = v & 4;
+					len = (v>>3) + 1 + (extend<<5);
+					extend = 0;
+					if (len > skip_xx)
+					{
+						runp += skip_xx;
+						len -= skip_xx;
+						goto intermediate_run;
+					}
+					runp += len;
+					break;
+				}
+				if (eol)
+				{
+					ww = 0;
+					break;
+				}
+				skip_xx -= len;
+			}
+			while (ww > 0)
+			{
+				int v = *runp++;
+				switch(v & 3)
+				{
+				case 0: /* Extend */
+					extend = v>>2;
+					break;
+				case 1: /* Transparent */
+					len = (v>>2) + 1 + (extend<<6);
+					extend = 0;
+transparent_run:
+					if (len > ww)
+						len = ww;
+					ww -= len;
+					ddp += len * n;
+					break;
+				case 2: /* Solid */
+					eol = v & 4;
+					len = (v>>3) + 1 + (extend<<5);
+					extend = 0;
+solid_run:
+					if (len > ww)
+						len = ww;
+					ww -= len;
+					do
+					{
+#ifdef ALPHA
+#if defined(N) && N == 2
+						ddp[0] = FZ_BLEND(colorbv[0], ddp[0], sa);
+						ddp[1] = FZ_BLEND(0xFF, ddp[1], sa);
+						ddp += 2;
+#elif defined(N) && N == 4
+						ddp[0] = FZ_BLEND(colorbv[0], ddp[0], sa);
+						ddp[1] = FZ_BLEND(colorbv[1], ddp[1], sa);
+						ddp[2] = FZ_BLEND(colorbv[2], ddp[2], sa);
+						ddp[3] = FZ_BLEND(0xFF, ddp[3], sa);
+						ddp += 4;
+#elif defined(N) && N == 5
+						ddp[0] = FZ_BLEND(colorbv[0], ddp[0], sa);
+						ddp[1] = FZ_BLEND(colorbv[1], ddp[1], sa);
+						ddp[2] = FZ_BLEND(colorbv[2], ddp[2], sa);
+						ddp[3] = FZ_BLEND(colorbv[3], ddp[3], sa);
+						ddp[4] = FZ_BLEND(0xFF, ddp[4], sa);
+						ddp += 5;
+#else
+						int k = 0;
+						do
+						{
+							*ddp = FZ_BLEND(colorbv[k++], *ddp, sa);
+							ddp++;
+						}
+						while (k != n-1);
+						*ddp = FZ_BLEND(0xFF, *ddp, sa);
+						ddp++;
+#endif
+#else
+#if defined(N) && N == 2
+						ddp[0] = colorbv[0];
+						ddp[1] = colorbv[1];
+						ddp += 2;
+#elif defined(N) && N == 4
+						*(uint32_t *)ddp = *(uint32_t*)colorbv;
+						ddp += 4;
+#elif defined(N) && N == 5
+						ddp[0] = colorbv[0];
+						ddp[1] = colorbv[1];
+						ddp[2] = colorbv[2];
+						ddp[3] = colorbv[3];
+						ddp[4] = colorbv[4];
+						ddp += 5;
+#else
+						int k = 0;
+						do
+						{
+							*ddp++ = colorbv[k++];
+						}
+						while (k != n);
+#endif
+#endif
+					}
+					while (--len);
+					break;
+				default: /* Intermediate */
+					eol = v & 4;
+					len = (v>>3) + 1 + (extend<<5);
+					extend = 0;
+intermediate_run:
+					if (len > ww)
+						len = ww;
+					ww -= len;
+					do
+					{
+						int k = 0;
+						int a = *runp++;
+#ifdef ALPHA
+						a = FZ_COMBINE(sa, FZ_EXPAND(a));
+#else
+						a = FZ_EXPAND(a);
+#endif
+						(void)k;
+#if defined(N) && N == 2
+						ddp[0] = FZ_BLEND(colorbv[0], ddp[0], a);
+						ddp[1] = FZ_BLEND(0xFF, ddp[1], a);
+						ddp += 2;
+#elif defined(N) && N == 4
+						ddp[0] = FZ_BLEND(colorbv[0], ddp[0], a);
+						ddp[1] = FZ_BLEND(colorbv[1], ddp[1], a);
+						ddp[2] = FZ_BLEND(colorbv[2], ddp[2], a);
+						ddp[3] = FZ_BLEND(0xFF, ddp[3], a);
+						ddp += 4;
+#elif defined(N) && N == 5
+						ddp[0] = FZ_BLEND(colorbv[0], ddp[0], a);
+						ddp[1] = FZ_BLEND(colorbv[1], ddp[1], a);
+						ddp[2] = FZ_BLEND(colorbv[2], ddp[2], a);
+						ddp[3] = FZ_BLEND(colorbv[3], ddp[3], a);
+						ddp[4] = FZ_BLEND(0xFF, ddp[4], a);
+						ddp += 5;
+#else
+						do
+						{
+							*ddp = FZ_BLEND(colorbv[k++], *ddp, a);
+							ddp++;
+						}
+						while (k != n-1);
+						*ddp = FZ_BLEND(0xFF, *ddp, a);
+						ddp++;
+#endif
+					}
+					while (--len);
+					break;
+				}
+				if (eol)
+					break;
+			}
+		}
+		dp += span;
+	}
+}
+
+#undef NAME
+#undef ALPHA
+#undef N
+#undef FUNCTION_NAMER
+#undef FUNCTION_NAME