commit 54184110f6f3e5d7276d5431e739a4fcf0c3523e
Author: Owen Taylor <otaylor@huygens.home.fishsoup.net>
Date:   Mon Apr 28 21:00:54 2008 +0200

    EXA: Use a single large glyph cache pixmap
    
    Add back exaGlyphs(); the new version copies the glyph images
    onto a single large glyph pixmap and draws from their to the
    destination surface. This reduces the management of small
    offscreen areas and will allow us to avoid texture unit setup
    between each glyph.

diff --git a/exa/Makefile.am b/exa/Makefile.am
index e2f7ed3..2b3f1e4 100644
--- a/exa/Makefile.am
+++ b/exa/Makefile.am
@@ -18,6 +18,7 @@ libexa_la_SOURCES = \
 	exa.c \
 	exa.h \
 	exa_accel.c \
+	exa_glyphs.c \
 	exa_migration.c \
 	exa_offscreen.c \
 	exa_render.c \
diff --git a/exa/exa.c b/exa/exa.c
index 3a6ad98..809fb4b 100644
--- a/exa/exa.c
+++ b/exa/exa.c
@@ -739,6 +739,8 @@ exaCloseScreen(int i, ScreenPtr pScreen)
     PictureScreenPtr	ps = GetPictureScreenIfSet(pScreen);
 #endif
 
+    exaGlyphsFini(pScreen);
+
     pScreen->CreateGC = pExaScr->SavedCreateGC;
     pScreen->CloseScreen = pExaScr->SavedCloseScreen;
     pScreen->GetImage = pExaScr->SavedGetImage;
@@ -754,7 +754,9 @@
 #ifdef RENDER
     if (ps) {
 	ps->Composite = pExaScr->SavedComposite;
+	ps->Glyphs = pExaScr->SavedGlyphs;
 	ps->Trapezoids = pExaScr->SavedTrapezoids;
+	ps->Triangles = pExaScr->SavedTriangles;
 	ps->AddTraps = pExaScr->SavedAddTraps;
     }
 #endif
@@ -914,6 +918,9 @@ exaDriverInit (ScreenPtr		pScreen,
         pExaScr->SavedComposite = ps->Composite;
 	ps->Composite = exaComposite;
 
+	pExaScr->SavedGlyphs = ps->Glyphs;
+	ps->Glyphs = exaGlyphs;
+	
 	pExaScr->SavedTriangles = ps->Triangles;
 	ps->Triangles = exaTriangles;
 
@@ -973,6 +980,8 @@ exaDriverInit (ScreenPtr		pScreen,
 	}
     }
 
+    exaGlyphsInit(pScreen);
+
     LogMessage(X_INFO, "EXA(%d): Driver registered support for the following"
 	       " operations:\n", pScreen->myNum);
     assert(pScreenInfo->PrepareSolid != NULL);
diff --git a/exa/exa_glyphs.c b/exa/exa_glyphs.c
new file mode 100644
index 0000000..3fe433a
--- /dev/null
+++ b/exa/exa_glyphs.c
@@ -0,0 +1,745 @@
+/*
+ * Copyright © 2008 Red Hat, Inc.
+ * Partly based on code Copyright © 2000 SuSE, Inc.
+ *
+ * Permission to use, copy, modify, distribute, and sell this software and its
+ * documentation for any purpose is hereby granted without fee, provided that
+ * the above copyright notice appear in all copies and that both that
+ * copyright notice and this permission notice appear in supporting
+ * documentation, and that the name of Red Hat not be used in advertising or
+ * publicity pertaining to distribution of the software without specific,
+ * written prior permission.  Red Hat makes no representations about the
+ * suitability of this software for any purpose.  It is provided "as is"
+ * without express or implied warranty.
+ *
+ * Red Hat DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE, INCLUDING ALL
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS, IN NO EVENT SHALL Red Hat
+ * BE LIABLE FOR ANY SPECIAL, INDIRECT OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+ * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION
+ * OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN 
+ * CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+ *
+ * Permission to use, copy, modify, distribute, and sell this software and its
+ * documentation for any purpose is hereby granted without fee, provided that
+ * the above copyright notice appear in all copies and that both that
+ * copyright notice and this permission notice appear in supporting
+ * documentation, and that the name of SuSE not be used in advertising or
+ * publicity pertaining to distribution of the software without specific,
+ * written prior permission.  SuSE makes no representations about the
+ * suitability of this software for any purpose.  It is provided "as is"
+ * without express or implied warranty.
+ *
+ * SuSE DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE, INCLUDING ALL
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS, IN NO EVENT SHALL SuSE
+ * BE LIABLE FOR ANY SPECIAL, INDIRECT OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+ * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION
+ * OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN 
+ * CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+ *
+ * Author: Owen Taylor <otaylor@fishsoup.net>
+ * Based on code by: Keith Packard
+ */
+
+#ifdef HAVE_DIX_CONFIG_H
+#include <dix-config.h>
+#endif
+
+#include <stdlib.h>
+
+#include "exa_priv.h"
+
+#include "mipict.h"
+
+#if DEBUG_GLYPH_CACHE
+#define DBG_GLYPH_CACHE(a) ErrorF a
+#else
+#define DBG_GLYPH_CACHE(a)
+#endif
+
+/* Instructions for rendering a single glyph */
+typedef struct {
+    INT16 xSrc;
+    INT16 ySrc;
+    INT16 xDst;
+    INT16 yDst;
+    INT16 width;
+    INT16 height;
+} ExaGlyphRenderRec, *ExaGlyphRenderPtr;
+
+/* Width of the pixmaps we use for the caches; this should be less than
+ * max texture size of the driver; this may need to actually come from
+ * the driver.
+ */
+#define CACHE_PICTURE_WIDTH 1024
+
+/* Maximum number of glyphs we buffer on the stack before flushing
+ * rendering to the mask or destination surface.
+ */
+#define GLYPH_BUFFER_SIZE 256
+
+typedef struct {
+    PicturePtr source;
+    ExaGlyphRenderRec glyphs[GLYPH_BUFFER_SIZE];
+    int count;
+} ExaGlyphBuffer, *ExaGlyphBufferPtr;
+
+typedef enum {
+    ExaGlyphSuccess,    /* Glyph added to render buffer */
+    ExaGlyphFail,       /* out of memory, etc */
+    ExaGlyphNeedFlush,  /* would evict a glyph already in the buffer */
+} ExaGlyphCacheResult;
+
+void
+exaGlyphsInit(ScreenPtr pScreen)
+{
+    ExaScreenPriv(pScreen);
+    int i = 0;
+
+    memset(pExaScr->glyphCaches, 0, sizeof(pExaScr->glyphCaches));
+
+    pExaScr->glyphCaches[i].format = PICT_a8;
+    pExaScr->glyphCaches[i].glyphWidth = pExaScr->glyphCaches[i].glyphHeight = 16;
+    i++;
+    pExaScr->glyphCaches[i].format = PICT_a8;
+    pExaScr->glyphCaches[i].glyphWidth = pExaScr->glyphCaches[i].glyphHeight = 32;
+    i++;
+    pExaScr->glyphCaches[i].format = PICT_a8r8g8b8;
+    pExaScr->glyphCaches[i].glyphWidth = pExaScr->glyphCaches[i].glyphHeight = 16;
+    i++;
+    pExaScr->glyphCaches[i].format = PICT_a8r8g8b8;
+    pExaScr->glyphCaches[i].glyphWidth = pExaScr->glyphCaches[i].glyphHeight = 32;
+    i++;
+
+    assert(i == EXA_NUM_GLYPH_CACHES);
+    
+    for (i = 0; i < EXA_NUM_GLYPH_CACHES; i++) {
+	pExaScr->glyphCaches[i].columns = CACHE_PICTURE_WIDTH / pExaScr->glyphCaches[i].glyphWidth;
+	pExaScr->glyphCaches[i].size = 256;
+	pExaScr->glyphCaches[i].hashSize = 557;
+    }
+}
+
+static void
+exaUnrealizeGlyphCaches(ScreenPtr    pScreen,
+			unsigned int format)
+{
+    ExaScreenPriv(pScreen);
+    int i;
+
+    for (i = 0; i < EXA_NUM_GLYPH_CACHES; i++) {
+	ExaGlyphCachePtr cache = &pExaScr->glyphCaches[i];
+	
+	if (cache->format != format)
+	    continue;
+
+	if (cache->picture) {
+	    FreePicture ((pointer) cache->picture, (XID) 0);
+	    cache->picture = NULL;
+	}
+
+	if (cache->hashEntries) {
+	    xfree(cache->hashEntries);
+	    cache->hashEntries = NULL;
+	}
+	
+	if (cache->glyphs) {
+	    xfree(cache->glyphs);
+	    cache->glyphs = NULL;
+	}
+	cache->glyphCount = 0;
+    }
+}
+
+/* All caches for a single format share a single pixmap for glyph storage,
+ * allowing mixing glyphs of different sizes without paying a penalty
+ * for switching between source pixmaps. (Note that for a size of font
+ * right at the border between two sizes, we might be switching for almost
+ * every glyph.)
+ *
+ * This function allocates the storage pixmap, and then fills in the
+ * rest of the allocated structures for all caches with the given format.
+ */
+static Bool
+exaRealizeGlyphCaches(ScreenPtr    pScreen,
+		      unsigned int format)
+{
+    ExaScreenPriv(pScreen);
+
+    int depth = PIXMAN_FORMAT_DEPTH(format);
+    PictFormatPtr pPictFormat;
+    PixmapPtr pPixmap;
+    PicturePtr pPicture;
+    int height;
+    int i;
+    int	error;
+
+    pPictFormat = PictureMatchFormat(pScreen, depth, format);
+    if (!pPictFormat)
+	return FALSE;
+    
+    /* Compute the total vertical size needed for the format */
+
+    height = 0;
+    for (i = 0; i < EXA_NUM_GLYPH_CACHES; i++) {
+	ExaGlyphCachePtr cache = &pExaScr->glyphCaches[i];
+	int rows;
+	
+	if (cache->format != format)
+	    continue;
+
+	rows = (cache->size + cache->columns - 1) / cache->columns;
+
+	height += rows * cache->glyphHeight;
+    }
+
+    /* Now allocate the pixmap and picture */
+       
+    pPixmap = (*pScreen->CreatePixmap) (pScreen,
+					CACHE_PICTURE_WIDTH,
+					height, depth, 0);
+    if (!pPixmap)
+	return FALSE;
+
+    pPicture = CreatePicture(0, &pPixmap->drawable, pPictFormat,
+			     0, 0, serverClient, &error);
+
+    (*pScreen->DestroyPixmap) (pPixmap); /* picture holds a refcount */
+
+    if (!pPicture)
+	return FALSE;
+
+    /* And store the picture in all the caches for the format */
+    
+    for (i = 0; i < EXA_NUM_GLYPH_CACHES; i++) {
+	ExaGlyphCachePtr cache = &pExaScr->glyphCaches[i];
+	int j;
+
+	if (cache->format != format)
+	    continue;
+
+	cache->picture = pPicture;
+	cache->picture->refcnt++;
+	cache->hashEntries = xalloc(sizeof(int) * cache->hashSize);
+	cache->glyphs = xalloc(sizeof(ExaCachedGlyphRec) * cache->size);
+	cache->glyphCount = 0;
+
+	if (!cache->hashEntries || !cache->glyphs)
+	    goto bail;
+
+	for (j = 0; j < cache->hashSize; j++)
+	    cache->hashEntries[j] = -1;
+	
+	cache->evictionPosition = rand() % cache->size;
+    }
+
+    /* Each cache references the picture individually */
+    FreePicture ((pointer) pPicture, (XID) 0);
+    return TRUE;
+
+bail:
+    exaUnrealizeGlyphCaches(pScreen, format);
+    return FALSE;
+}
+
+void
+exaGlyphsFini (ScreenPtr pScreen)
+{
+    ExaScreenPriv(pScreen);
+    int i;
+
+    for (i = 0; i < EXA_NUM_GLYPH_CACHES; i++) {
+	ExaGlyphCachePtr cache = &pExaScr->glyphCaches[i];
+
+	if (cache->picture)
+	    exaUnrealizeGlyphCaches(pScreen, cache->format);
+    }
+}
+
+static int
+exaGlyphCacheHashLookup(ExaGlyphCachePtr cache,
+			GlyphPtr         pGlyph)
+{
+    int slot;
+
+    slot = (*(CARD32 *) pGlyph->sha1) % cache->hashSize;
+    
+    while (TRUE) { /* hash table can never be full */
+	int entryPos = cache->hashEntries[slot];
+	if (entryPos == -1)
+	    return -1;
+
+	if (memcmp(pGlyph->sha1, cache->glyphs[entryPos].sha1, sizeof(pGlyph->sha1)) == 0){
+	    DBG_GLYPH_CACHE((" found entry at %d\n", slot));
+	    return entryPos;
+	}
+	    
+	DBG_GLYPH_CACHE((" lookup linear probe bumpalong\n"));
+	slot--;
+	if (slot < 0)
+	    slot = cache->hashSize - 1;
+    }
+}
+
+static void
+exaGlyphCacheHashInsert(ExaGlyphCachePtr cache,
+			GlyphPtr         pGlyph,
+			int              pos)
+{
+    int slot;
+
+    memcpy(cache->glyphs[pos].sha1, pGlyph->sha1, sizeof(pGlyph->sha1));
+    
+    slot = (*(CARD32 *) pGlyph->sha1) % cache->hashSize;
+    
+    while (TRUE) { /* hash table can never be full */
+	if (cache->hashEntries[slot] == -1) {
+	    DBG_GLYPH_CACHE((" inserting entry at %d\n", slot));
+	    cache->hashEntries[slot] = pos;
+	    return;
+	}
+	    
+	slot--;
+	if (slot < 0)
+	    slot = cache->hashSize - 1;
+    }
+}
+
+static void
+exaGlyphCacheHashRemove(ExaGlyphCachePtr cache,
+			int              pos)
+{
+    int slot;
+    int emptiedSlot = -1;
+
+    slot = (*(CARD32 *) cache->glyphs[pos].sha1) % cache->hashSize;
+
+    while (TRUE) { /* hash table can never be full */
+	int entryPos = cache->hashEntries[slot];
+	
+	if (entryPos == -1)
+	    return;
+
+	if (entryPos == pos) {
+	    cache->hashEntries[slot] = -1;
+	    emptiedSlot = slot;
+	} else if (emptiedSlot != -1) {
+	    /* See if we can move this entry into the emptied slot, we can't
+	     * do that if if entry would have hashed between the current position
+	     * and the emptied slot. (taking wrapping into account). Bad positions
+	     * are:
+	     *
+	     * |   XXXXXXXXXX             |
+	     *     i         j            
+	     *                            
+	     * |XXX                   XXXX|
+	     *     j                  i
+	     *
+	     * i - slot, j - emptiedSlot
+	     *
+	     * (Knuth 6.4R)
+	     */
+	    
+	    int entrySlot = (*(CARD32 *) cache->glyphs[entryPos].sha1) % cache->hashSize;
+
+	    if (!((entrySlot >= slot && entrySlot < emptiedSlot) ||
+		  (emptiedSlot < slot && (entrySlot < emptiedSlot || entrySlot >= slot)))) 
+	    {
+		cache->hashEntries[emptiedSlot] = entryPos;
+		cache->hashEntries[slot] = -1;
+		emptiedSlot = slot;
+	    }
+	}
+	
+	slot--;
+	if (slot < 0)
+	    slot = cache->hashSize - 1;
+    }
+}
+
+static ExaGlyphCacheResult
+exaGlyphCacheBufferGlyph(ScreenPtr         pScreen,
+			 ExaGlyphCachePtr  cache,
+			 ExaGlyphBufferPtr buffer,
+			 GlyphPtr          pGlyph,
+			 int               xGlyph,
+			 int               yGlyph)
+{
+    ExaGlyphRenderPtr glyphRec;
+    int pos;
+    
+    if (buffer->source && buffer->source != cache->picture)
+	return ExaGlyphNeedFlush;
+
+    if (!cache->picture) {
+	if (!exaRealizeGlyphCaches(pScreen, cache->format))
+	    return ExaGlyphFail;
+    }
+
+    DBG_GLYPH_CACHE(("(%d,%d,%s): buffering glyph %lx\n",
+		     cache->glyphWidth, cache->glyphHeight, cache->format == PICT_a8 ? "A" : "ARGB",
+		     (long)*(CARD32 *) pGlyph->sha1));
+   
+    pos = exaGlyphCacheHashLookup(cache, pGlyph);
+    if (pos != -1) {
+	DBG_GLYPH_CACHE(("  found existing glyph at %d\n", pos));
+    } else {
+	if (cache->glyphCount < cache->size) {
+	    /* Space remaining; we fill from the start */
+	    pos = cache->glyphCount;
+	    cache->glyphCount++;
+	    DBG_GLYPH_CACHE(("  storing glyph in free space at %d\n", pos));
+
+	    exaGlyphCacheHashInsert(cache, pGlyph, pos);
+
+	} else {
+	    /* Need to evict an entry. We have to see if any glyphs
+	     * already in the output buffer were at this position in
+	     * the cache
+	     */
+	    
+	    pos = cache->evictionPosition;
+	    DBG_GLYPH_CACHE(("  evicting glyph at %d\n", pos));
+	    if (buffer->count) {
+		int x, y;
+		int i;
+		
+		x = (pos % cache->columns) * cache->glyphWidth;
+		y = (pos / cache->columns) * cache->glyphHeight;
+
+		for (i = 0; i < buffer->count; i++) {
+		    if (buffer->glyphs[i].xSrc == x && buffer->glyphs[i].ySrc == y) {
+			DBG_GLYPH_CACHE(("  must flush buffer\n"));
+			return ExaGlyphNeedFlush;
+		    }
+		}
+	    }
+
+	    /* OK, we're all set, swap in the new glyph */
+	    exaGlyphCacheHashRemove(cache, pos);
+	    exaGlyphCacheHashInsert(cache, pGlyph, pos);
+
+	    /* And pick a new eviction position */
+	    cache->evictionPosition = rand() % cache->size;
+	}
+
+	/* Now actually upload the glyph into the cache picture */
+
+	CompositePicture (PictOpSrc,
+			  GlyphPicture(pGlyph)[pScreen->myNum],
+			  None,
+			  cache->picture,
+			  0, 0,
+			  0, 0,
+			  (pos % cache->columns) * cache->glyphWidth,
+			  (pos / cache->columns) * cache->glyphHeight,
+			  pGlyph->info.width,
+			  pGlyph->info.height);
+    }
+    
+
+    buffer->source = cache->picture;
+	    
+    glyphRec = &buffer->glyphs[buffer->count];
+    glyphRec->xSrc = (pos % cache->columns) * cache->glyphWidth;
+    glyphRec->ySrc = (pos / cache->columns) * cache->glyphHeight;
+    glyphRec->xDst = xGlyph - pGlyph->info.x;
+    glyphRec->yDst = yGlyph - pGlyph->info.y;
+    glyphRec->width = pGlyph->info.width;
+    glyphRec->height = pGlyph->info.height;
+	    
+    buffer->count++;
+
+    return ExaGlyphSuccess;
+}
+
+static ExaGlyphCacheResult
+exaBufferGlyph(ScreenPtr         pScreen,
+	       ExaGlyphBufferPtr buffer,
+	       GlyphPtr          pGlyph,
+	       int               xGlyph,
+	       int               yGlyph)
+{
+    ExaScreenPriv(pScreen);
+    unsigned int format = (GlyphPicture(pGlyph)[pScreen->myNum])->format;
+    int width = pGlyph->info.width;
+    int height = pGlyph->info.width;
+    ExaGlyphRenderPtr glyphRec;
+    PicturePtr source;
+    int i;
+
+    if (buffer->count == GLYPH_BUFFER_SIZE)
+	return ExaGlyphNeedFlush;
+    
+    for (i = 0; i < EXA_NUM_GLYPH_CACHES; i++) {
+	ExaGlyphCachePtr cache = &pExaScr->glyphCaches[i];
+
+	if (format == cache->format &&
+	    width <= cache->glyphWidth &&
+	    height <= cache->glyphHeight) {
+	    ExaGlyphCacheResult result = exaGlyphCacheBufferGlyph(pScreen, &pExaScr->glyphCaches[i],
+								  buffer,
+								  pGlyph, xGlyph, yGlyph);
+	    switch (result) {
+	    case ExaGlyphFail:
+		break;
+	    case ExaGlyphSuccess:
+	    case ExaGlyphNeedFlush:
+		return result;
+	    }
+	}
+    }
+
+    /* Couldn't find the glyph in the cache, use the glyph picture directly */
+
+    source = GlyphPicture(pGlyph)[pScreen->myNum];
+    if (buffer->source && buffer->source != source)
+	return ExaGlyphNeedFlush;
+
+    buffer->source = source;
+    
+    glyphRec = &buffer->glyphs[buffer->count];
+    glyphRec->xSrc = 0;
+    glyphRec->ySrc = 0;
+    glyphRec->xDst = xGlyph - pGlyph->info.x;
+    glyphRec->yDst = yGlyph - pGlyph->info.y;
+    glyphRec->width = pGlyph->info.width;
+    glyphRec->height = pGlyph->info.height;
+
+    buffer->count++;
+
+    return ExaGlyphSuccess;
+}
+
+static void
+exaGlyphsToMask(PicturePtr        pMask,
+		ExaGlyphBufferPtr buffer)
+{
+    int i;
+
+    for (i = 0; i < buffer->count; i++) {
+	ExaGlyphRenderPtr glyphRec = &buffer->glyphs[i];
+	
+	CompositePicture (PictOpAdd,
+			  buffer->source,
+			  None,
+			  pMask,
+			  glyphRec->xSrc,
+			  glyphRec->ySrc,
+			  0, 0,
+			  glyphRec->xDst,
+			  glyphRec->yDst,
+			  glyphRec->width,
+			  glyphRec->height);
+    }
+    
+    buffer->count = 0;
+    buffer->source = NULL;
+}
+
+static void
+exaGlyphsToDst(CARD8		 op,
+	       PicturePtr	 pSrc,
+	       PicturePtr	 pDst,
+	       ExaGlyphBufferPtr buffer,
+	       INT16		 xSrc,
+	       INT16		 ySrc,
+	       INT16		 xDst,
+	       INT16		 yDst)
+{
+    int i;
+
+    for (i = 0; i < buffer->count; i++) {
+	ExaGlyphRenderPtr glyphRec = &buffer->glyphs[i];
+	
+	CompositePicture (op,
+			  pSrc,
+			  buffer->source,
+			  pDst,
+			  xSrc + glyphRec->xDst - xDst,
+			  ySrc + glyphRec->yDst - yDst,
+			  glyphRec->xSrc,
+			  glyphRec->ySrc,
+			  glyphRec->xDst,
+			  glyphRec->yDst,
+			  glyphRec->width,
+			  glyphRec->height);
+    }
+    
+    buffer->count = 0;
+    buffer->source = NULL;
+}
+
+/* Cut and paste from render/glyph.c - probably should export it instead */
+static void
+GlyphExtents (int		nlist,
+	      GlyphListPtr	list,
+	      GlyphPtr	       *glyphs,
+	      BoxPtr		extents)
+{
+    int		x1, x2, y1, y2;
+    int		n;
+    GlyphPtr	glyph;
+    int		x, y;
+    
+    x = 0;
+    y = 0;
+    extents->x1 = MAXSHORT;
+    extents->x2 = MINSHORT;
+    extents->y1 = MAXSHORT;
+    extents->y2 = MINSHORT;
+    while (nlist--)
+    {
+	x += list->xOff;
+	y += list->yOff;
+	n = list->len;
+	list++;
+	while (n--)
+	{
+	    glyph = *glyphs++;
+	    x1 = x - glyph->info.x;
+	    if (x1 < MINSHORT)
+		x1 = MINSHORT;
+	    y1 = y - glyph->info.y;
+	    if (y1 < MINSHORT)
+		y1 = MINSHORT;
+	    x2 = x1 + glyph->info.width;
+	    if (x2 > MAXSHORT)
+		x2 = MAXSHORT;
+	    y2 = y1 + glyph->info.height;
+	    if (y2 > MAXSHORT)
+		y2 = MAXSHORT;
+	    if (x1 < extents->x1)
+		extents->x1 = x1;
+	    if (x2 > extents->x2)
+		extents->x2 = x2;
+	    if (y1 < extents->y1)
+		extents->y1 = y1;
+	    if (y2 > extents->y2)
+		extents->y2 = y2;
+	    x += glyph->info.xOff;
+	    y += glyph->info.yOff;
+	}
+    }
+}
+
+#define NeedsComponent(f) (PICT_FORMAT_A(f) != 0 && PICT_FORMAT_RGB(f) != 0)
+
+void
+exaGlyphs (CARD8 	 op,
+	   PicturePtr	 pSrc,
+	   PicturePtr	 pDst,
+	   PictFormatPtr maskFormat,
+	   INT16	 xSrc,
+	   INT16	 ySrc,
+	   int		 nlist,
+	   GlyphListPtr	 list,
+	   GlyphPtr	*glyphs)
+{
+    PicturePtr	pPicture;
+    PixmapPtr   pMaskPixmap = 0;
+    PicturePtr  pMask;
+    ScreenPtr   pScreen = pDst->pDrawable->pScreen;
+    int		width = 0, height = 0;
+    int		x, y;
+    int		xDst = list->xOff, yDst = list->yOff;
+    int		n;
+    GlyphPtr	glyph;
+    int		error;
+    BoxRec	extents = {0, 0, 0, 0};
+    CARD32	component_alpha;
+    ExaGlyphBuffer buffer;
+
+    if (maskFormat)
+    {
+	GCPtr	    pGC;
+	xRectangle  rect;
+
+	GlyphExtents (nlist, list, glyphs, &extents);
+
+	if (extents.x2 <= extents.x1 || extents.y2 <= extents.y1)
+	    return;
+	width = extents.x2 - extents.x1;
+	height = extents.y2 - extents.y1;
+	pMaskPixmap = (*pScreen->CreatePixmap) (pScreen, width, height,
+						maskFormat->depth,
+						CREATE_PIXMAP_USAGE_SCRATCH);
+	if (!pMaskPixmap)
+	    return;
+	component_alpha = NeedsComponent(maskFormat->format);
+	pMask = CreatePicture (0, &pMaskPixmap->drawable,
+			       maskFormat, CPComponentAlpha, &component_alpha,
+			       serverClient, &error);
+	if (!pMask)
+	{
+	    (*pScreen->DestroyPixmap) (pMaskPixmap);
+	    return;
+	}
+	pGC = GetScratchGC (pMaskPixmap->drawable.depth, pScreen);
+	ValidateGC (&pMaskPixmap->drawable, pGC);
+	rect.x = 0;
+	rect.y = 0;
+	rect.width = width;
+	rect.height = height;
+	(*pGC->ops->PolyFillRect) (&pMaskPixmap->drawable, pGC, 1, &rect);
+	FreeScratchGC (pGC);
+	x = -extents.x1;
+	y = -extents.y1;
+    }
+    else
+    {
+	pMask = pDst;
+	x = 0;
+	y = 0;
+    }
+    buffer.count = 0;
+    buffer.source = NULL;
+    while (nlist--)
+    {
+	x += list->xOff;
+	y += list->yOff;
+	n = list->len;
+	while (n--)
+	{
+	    glyph = *glyphs++;
+	    pPicture = GlyphPicture (glyph)[pScreen->myNum];
+
+	    if (exaBufferGlyph(pScreen, &buffer, glyph, x, y) == ExaGlyphNeedFlush)
+	    {
+		if (maskFormat)
+		    exaGlyphsToMask(pMask, &buffer);
+		else
+		    exaGlyphsToDst(op, pSrc, pDst, &buffer,
+				   xSrc, ySrc, xDst, yDst);
+
+		exaBufferGlyph(pScreen, &buffer, glyph, x, y);
+	    }
+
+	    x += glyph->info.xOff;
+	    y += glyph->info.yOff;
+	}
+	list++;
+    }
+    
+    if (maskFormat)
+	exaGlyphsToMask(pMask, &buffer);
+    else
+	exaGlyphsToDst(op, pSrc, pDst, &buffer,
+		       xSrc, ySrc, xDst, yDst);
+    
+    if (maskFormat)
+    {
+	x = extents.x1;
+	y = extents.y1;
+	CompositePicture (op,
+			  pSrc,
+			  pMask,
+			  pDst,
+			  xSrc + x - xDst,
+			  ySrc + y - yDst,
+			  0, 0,
+			  x, y,
+			  width, height);
+	FreePicture ((pointer) pMask, (XID) 0);
+	(*pScreen->DestroyPixmap) (pMaskPixmap);
+    }
+}
diff --git a/exa/exa_priv.h b/exa/exa_priv.h
index 0138e4a..aaceeb8 100644
--- a/exa/exa_priv.h
+++ b/exa/exa_priv.h
@@ -61,6 +61,7 @@
 #define DEBUG_MIGRATE		0
 #define DEBUG_PIXMAP		0
 #define DEBUG_OFFSCREEN		0
+#define DEBUG_GLYPH_CACHE	0
 
 #if DEBUG_TRACE_FALL
 #define EXA_FALLBACK(x)     					\
@@ -95,6 +96,37 @@ enum ExaMigrationHeuristic {
     ExaMigrationSmart
 };
 
+typedef struct {
+    unsigned char sha1[20];
+} ExaCachedGlyphRec, *ExaCachedGlyphPtr;
+
+typedef struct {
+    /* The identity of the cache, statically configured at initialization */
+    unsigned int format;
+    int glyphWidth;
+    int glyphHeight;
+
+    int size; /* Size of cache; eventually this should be dynamically determined */
+
+    /* Hash table mapping from glyph sha1 to position in the glyph; we use
+     * open addressing with a hash table size determined based on size and large
+     * enough so that we always have a good amount of free space, so we can
+     * use linear probing. (Linear probing is preferrable to double hashing
+     * here because it allows us to easily remove entries.)
+     */
+    int *hashEntries;
+    int hashSize;
+    
+    ExaCachedGlyphPtr glyphs;
+    int glyphCount; /* Current number of glyphs */
+    
+    PicturePtr picture;   /* Where the glyphs of the cache are stored */
+    int columns;          /* Number of columns the glyphs are layed out in */
+    int evictionPosition; /* Next random position to evict a glyph */
+} ExaGlyphCacheRec, *ExaGlyphCachePtr;
+
+#define EXA_NUM_GLYPH_CACHES 4
+
 typedef void (*EnableDisableFBAccessProcPtr)(int, Bool);
 typedef struct {
     ExaDriverPtr info;
@@ -122,6 +154,8 @@ typedef struct {
     unsigned			 disableFbCount;
     Bool			 optimize_migration;
     unsigned			 offScreenCounter;
+
+    ExaGlyphCacheRec             glyphCaches[EXA_NUM_GLYPH_CACHES];
 } ExaScreenPrivRec, *ExaScreenPrivPtr;
 
 /*
@@ -432,6 +466,13 @@ exaTriangles (CARD8 op, PicturePtr pSrc, PicturePtr pDst,
 	      PictFormatPtr maskFormat, INT16 xSrc, INT16 ySrc,
 	      int ntri, xTriangle *tris);
 
+/* exa_glyph.c */
+void
+exaGlyphsInit(ScreenPtr pScreen);
+
+void
+exaGlyphsFini (ScreenPtr pScreen);
+
 void
 exaGlyphs (CARD8	op,
 	  PicturePtr	pSrc,
commit 40eb14c9482457969e0bde97c49edad536285e02
Author: Owen Taylor <otaylor@huygens.home.fishsoup.net>
Date:   Mon Apr 28 21:00:54 2008 +0200

    EXA: Add exaCompositeRects()
    
    Add a function to composite multiple independent rectangles
    from the same source to the same destination in a single
    operation: this is useful for building a glyph mask.

diff --git a/exa/exa_glyphs.c b/exa/exa_glyphs.c
index 3fe433a..55fdb01 100644
--- a/exa/exa_glyphs.c
+++ b/exa/exa_glyphs.c
@@ -56,16 +56,6 @@
 #define DBG_GLYPH_CACHE(a)
 #endif
 
-/* Instructions for rendering a single glyph */
-typedef struct {
-    INT16 xSrc;
-    INT16 ySrc;
-    INT16 xDst;
-    INT16 yDst;
-    INT16 width;
-    INT16 height;
-} ExaGlyphRenderRec, *ExaGlyphRenderPtr;
-
 /* Width of the pixmaps we use for the caches; this should be less than
  * max texture size of the driver; this may need to actually come from
  * the driver.
@@ -79,7 +69,7 @@ typedef struct {
 
 typedef struct {
     PicturePtr source;
-    ExaGlyphRenderRec glyphs[GLYPH_BUFFER_SIZE];
+    ExaCompositeRectRec rects[GLYPH_BUFFER_SIZE];
     int count;
 } ExaGlyphBuffer, *ExaGlyphBufferPtr;
 
@@ -364,7 +354,7 @@ exaGlyphCacheBufferGlyph(ScreenPtr         pScreen,
 			 int               xGlyph,
 			 int               yGlyph)
 {
-    ExaGlyphRenderPtr glyphRec;
+    ExaCompositeRectPtr rect;
     int pos;
     
     if (buffer->source && buffer->source != cache->picture)
@@ -407,7 +397,7 @@ exaGlyphCacheBufferGlyph(ScreenPtr         pScreen,
 		y = (pos / cache->columns) * cache->glyphHeight;
 
 		for (i = 0; i < buffer->count; i++) {
-		    if (buffer->glyphs[i].xSrc == x && buffer->glyphs[i].ySrc == y) {
+		    if (buffer->rects[i].xSrc == x && buffer->rects[i].ySrc == y) {
 			DBG_GLYPH_CACHE(("  must flush buffer\n"));
 			return ExaGlyphNeedFlush;
 		    }
@@ -439,13 +429,13 @@ exaGlyphCacheBufferGlyph(ScreenPtr         pScreen,
 
     buffer->source = cache->picture;
 	    
-    glyphRec = &buffer->glyphs[buffer->count];
-    glyphRec->xSrc = (pos % cache->columns) * cache->glyphWidth;
-    glyphRec->ySrc = (pos / cache->columns) * cache->glyphHeight;
-    glyphRec->xDst = xGlyph - pGlyph->info.x;
-    glyphRec->yDst = yGlyph - pGlyph->info.y;
-    glyphRec->width = pGlyph->info.width;
-    glyphRec->height = pGlyph->info.height;
+    rect = &buffer->rects[buffer->count];
+    rect->xSrc = (pos % cache->columns) * cache->glyphWidth;
+    rect->ySrc = (pos / cache->columns) * cache->glyphHeight;
+    rect->xDst = xGlyph - pGlyph->info.x;
+    rect->yDst = yGlyph - pGlyph->info.y;
+    rect->width = pGlyph->info.width;
+    rect->height = pGlyph->info.height;
 	    
     buffer->count++;
 
@@ -463,7 +453,7 @@ exaBufferGlyph(ScreenPtr         pScreen,
     unsigned int format = (GlyphPicture(pGlyph)[pScreen->myNum])->format;
     int width = pGlyph->info.width;
     int height = pGlyph->info.width;
-    ExaGlyphRenderPtr glyphRec;
+    ExaCompositeRectPtr rect;
     PicturePtr source;
     int i;
 
@@ -497,13 +487,13 @@ exaBufferGlyph(ScreenPtr         pScreen,
 
     buffer->source = source;
     
-    glyphRec = &buffer->glyphs[buffer->count];
-    glyphRec->xSrc = 0;
-    glyphRec->ySrc = 0;
-    glyphRec->xDst = xGlyph - pGlyph->info.x;
-    glyphRec->yDst = yGlyph - pGlyph->info.y;
-    glyphRec->width = pGlyph->info.width;
-    glyphRec->height = pGlyph->info.height;
+    rect = &buffer->rects[buffer->count];
+    rect->xSrc = 0;
+    rect->ySrc = 0;
+    rect->xDst = xGlyph - pGlyph->info.x;
+    rect->yDst = yGlyph - pGlyph->info.y;
+    rect->width = pGlyph->info.width;
+    rect->height = pGlyph->info.height;
 
     buffer->count++;
 
@@ -514,23 +504,8 @@ static void
 exaGlyphsToMask(PicturePtr        pMask,
 		ExaGlyphBufferPtr buffer)
 {
-    int i;
-
-    for (i = 0; i < buffer->count; i++) {
-	ExaGlyphRenderPtr glyphRec = &buffer->glyphs[i];
-	
-	CompositePicture (PictOpAdd,
-			  buffer->source,
-			  None,
-			  pMask,
-			  glyphRec->xSrc,
-			  glyphRec->ySrc,
-			  0, 0,
-			  glyphRec->xDst,
-			  glyphRec->yDst,
-			  glyphRec->width,
-			  glyphRec->height);
-    }
+    exaCompositeRects(PictOpAdd, buffer->source, pMask,
+		      buffer->count, buffer->rects);
     
     buffer->count = 0;
     buffer->source = NULL;
@@ -549,20 +524,20 @@ exaGlyphsToDst(CARD8		 op,
     int i;
 
     for (i = 0; i < buffer->count; i++) {
-	ExaGlyphRenderPtr glyphRec = &buffer->glyphs[i];
+	ExaCompositeRectPtr rect = &buffer->rects[i];
 	
 	CompositePicture (op,
 			  pSrc,
 			  buffer->source,
 			  pDst,
-			  xSrc + glyphRec->xDst - xDst,
-			  ySrc + glyphRec->yDst - yDst,
-			  glyphRec->xSrc,
-			  glyphRec->ySrc,
-			  glyphRec->xDst,
-			  glyphRec->yDst,
-			  glyphRec->width,
-			  glyphRec->height);
+			  xSrc + rect->xDst - xDst,
+			  ySrc + rect->yDst - yDst,
+			  rect->xSrc,
+			  rect->ySrc,
+			  rect->xDst,
+			  rect->yDst,
+			  rect->width,
+			  rect->height);
     }
     
     buffer->count = 0;
diff --git a/exa/exa_priv.h b/exa/exa_priv.h
index aaceeb8..0d5d0f5 100644
--- a/exa/exa_priv.h
+++ b/exa/exa_priv.h
@@ -243,6 +243,15 @@ typedef struct _ExaMigrationRec {
     RegionPtr pReg;
 } ExaMigrationRec, *ExaMigrationPtr;
 
+typedef struct {
+    INT16 xSrc;
+    INT16 ySrc;
+    INT16 xDst;
+    INT16 yDst;
+    INT16 width;
+    INT16 height;
+} ExaCompositeRectRec, *ExaCompositeRectPtr;
+
 /**
  * exaDDXDriverInit must be implemented by the DDX using EXA, and is the place
  * to set EXA options or hook in screen functions to handle using EXA as the AA.
@@ -457,6 +466,13 @@ exaComposite(CARD8	op,
 	     CARD16	height);
 
 void
+exaCompositeRects(CARD8	              op,
+		  PicturePtr	      Src,
+		  PicturePtr	      pDst,
+		  int                 nrect,
+		  ExaCompositeRectPtr rects);
+
+void
 exaTrapezoids (CARD8 op, PicturePtr pSrc, PicturePtr pDst,
                PictFormatPtr maskFormat, INT16 xSrc, INT16 ySrc,
                int ntrap, xTrapezoid *traps);
diff --git a/exa/exa_render.c b/exa/exa_render.c
index 1d7b897..43b0029 100644
--- a/exa/exa_render.c
+++ b/exa/exa_render.c
@@ -332,6 +332,228 @@ exaTryDriverSolidFill(PicturePtr	pSrc,
 }
 
 static int
+exaTryDriverCompositeRects(CARD8	       op,
+			   PicturePtr	       pSrc,
+			   PicturePtr	       pDst,
+			   int                 nrect,
+			   ExaCompositeRectPtr rects)
+{
+    ExaScreenPriv (pDst->pDrawable->pScreen);
+    int src_off_x, src_off_y, dst_off_x, dst_off_y;
+    PixmapPtr pSrcPix, pDstPix;
+    ExaPixmapPrivPtr pSrcExaPix, pDstExaPix;
+    struct _Pixmap scratch;
+    ExaMigrationRec pixmaps[2];
+
+    pSrcPix = exaGetDrawablePixmap(pSrc->pDrawable);
+    pSrcExaPix = ExaGetPixmapPriv(pSrcPix);
+
+    pDstPix = exaGetDrawablePixmap(pDst->pDrawable);
+    pDstExaPix = ExaGetPixmapPriv(pDstPix);
+
+    /* Check whether the accelerator can use these pixmaps.
+     * FIXME: If it cannot, use temporary pixmaps so that the drawing
+     * happens within limits.
+     */
+    if (pSrcExaPix->accel_blocked ||
+	pDstExaPix->accel_blocked)
+    {
+	return -1;
+    }
+
+    if (pExaScr->info->CheckComposite &&
+	!(*pExaScr->info->CheckComposite) (op, pSrc, NULL, pDst))
+    {
+	return -1;
+    }
+    
+    exaGetDrawableDeltas (pDst->pDrawable, pDstPix, &dst_off_x, &dst_off_y);
+	
+    pixmaps[0].as_dst = TRUE;
+    pixmaps[0].as_src = exaOpReadsDestination(op);
+    pixmaps[0].pPix = pDstPix;
+    pixmaps[0].pReg = NULL;
+    pixmaps[1].as_dst = FALSE;
+    pixmaps[1].as_src = TRUE;
+    pixmaps[1].pPix = pSrcPix;
+    pixmaps[1].pReg = NULL;
+    exaDoMigration(pixmaps, 2, TRUE);
+
+    pSrcPix = exaGetOffscreenPixmap (pSrc->pDrawable, &src_off_x, &src_off_y);
+    if (!exaPixmapIsOffscreen(pDstPix))
+	return 0;
+    
+    if (!pSrcPix && pExaScr->info->UploadToScratch)
+    {
+	pSrcPix = exaGetDrawablePixmap (pSrc->pDrawable);
+	if ((*pExaScr->info->UploadToScratch) (pSrcPix, &scratch))
+	    pSrcPix = &scratch;
+    }
+
+    if (!pSrcPix)
+	return 0;
+
+    if (!(*pExaScr->info->PrepareComposite) (op, pSrc, NULL, pDst, pSrcPix,
+					     NULL, pDstPix))
+	return -1;
+
+    while (nrect--)
+    {
+	INT16 xDst = rects->xDst + pDst->pDrawable->x;
+	INT16 yDst = rects->yDst + pDst->pDrawable->y;
+	INT16 xSrc = rects->xSrc + pSrc->pDrawable->x;
+	INT16 ySrc = rects->ySrc + pSrc->pDrawable->y;
+	
+	RegionRec region;
+	BoxPtr pbox;
+	int nbox;
+	
+	if (!miComputeCompositeRegion (&region, pSrc, NULL, pDst,
+				       xSrc, ySrc, 0, 0, xDst, yDst,
+				       rects->width, rects->height))
+	    goto next_rect;
+	
+	REGION_TRANSLATE(pScreen, &region, dst_off_x, dst_off_y);
+	
+	nbox = REGION_NUM_RECTS(&region);
+	pbox = REGION_RECTS(&region);
+
+	xSrc = xSrc + src_off_x - xDst - dst_off_x;
+	ySrc = ySrc + src_off_y - yDst - dst_off_y;
+	
+	while (nbox--)
+	{
+	    (*pExaScr->info->Composite) (pDstPix,
+					 pbox->x1 + xSrc,
+					 pbox->y1 + ySrc,
+					 0, 0,
+					 pbox->x1,
+					 pbox->y1,
+					 pbox->x2 - pbox->x1,
+					 pbox->y2 - pbox->y1);
+	    pbox++;
+	}
+
+    next_rect:
+	REGION_UNINIT(pDst->pDrawable->pScreen, &region);
+
+	rects++;
+    }
+    
+    (*pExaScr->info->DoneComposite) (pDstPix);
+    exaMarkSync(pDst->pDrawable->pScreen);
+	
+    return 1;
+}
+
+/**
+ * Copy a number of rectangles from source to destination in a single
+ * operation. This is specialized for building a glyph mask: we don'y
+ * have a mask argument because we don't need it for that, and we
+ * don't have he special-case fallbacks found in exaComposite() - if the
+ * driver can support it, we use the driver functionality, otherwise we
+ * fallback straight to software.
+ */
+void
+exaCompositeRects(CARD8	              op,
+		  PicturePtr	      pSrc,
+		  PicturePtr	      pDst,
+		  int                 nrect,
+		  ExaCompositeRectPtr rects)
+{
+    PixmapPtr pPixmap = exaGetDrawablePixmap(pDst->pDrawable);
+    ExaPixmapPriv(pPixmap);
+    
+    int xoff, yoff;
+    int x1 = MAXSHORT;
+    int y1 = MAXSHORT;
+    int x2 = MINSHORT;
+    int y2 = MINSHORT;
+    RegionRec region;
+    RegionPtr pending_damage;
+    BoxRec box;
+    int n;
+    ExaCompositeRectPtr r;
+    
+    /* We have to manage the damage ourselves, since CompositeRects isn't
+     * something in the screen that can be managed by the damage extension,
+     * and EXA depends on damage to track what needs to be migrated between
+     * offscreen and onscreen.
+     */
+
+    /* Compute the overall extents of the composited region - we're making
+     * the assumption here that we are compositing a bunch of glyphs that
+     * cluster closely together and damaging each glyph individually would
+     * be a loss compared to damaging the bounding box.
+     */
+    n = nrect;
+    r = rects;
+    while (n--) {
+	int rect_x2 = r->xDst + r->width;
+	int rect_y2 = r->yDst + r->width;
+
+	if (r->xDst < x1) x1 = r->xDst;
+	if (r->xDst < y1) y1 = r->xDst;
+	if (rect_x2 > x2) x2 = rect_x2;
+	if (rect_y2 > y2) y2 = rect_y2;
+	
+	r++;
+    }
+
+    if (x2 <= x1 && y2 <= y1)
+	return;
+
+    box.x1 = x1;
+    box.x2 = x2 < MAXSHORT ? x2 : MAXSHORT;
+    box.y1 = y1;
+    box.y2 = y2 < MAXSHORT ? y2 : MAXSHORT;
+    
+    /* The pixmap migration code relies on pendingDamage indicating
+     * the bounds of the current rendering, so we need to force 
+     * the actual damage into that region before we do anything, and
+     * (see use of DamagePendingRegion in exaCopyDirty)
+     */
+    
+    REGION_INIT(pScreen, &region, &box, 1);
+    
+    exaGetDrawableDeltas(pDst->pDrawable, pPixmap, &xoff, &yoff);
+
+    REGION_TRANSLATE(pScreen, &region, xoff, yoff);
+    pending_damage = DamagePendingRegion(pExaPixmap->pDamage);
+    REGION_UNION(pScreen, pending_damage, pending_damage, &region);
+    REGION_TRANSLATE(pScreen, &region, -xoff, -yoff);
+    
+    /************************************************************/
+    
+    ValidatePicture (pSrc);
+    ValidatePicture (pDst);
+    
+    if (exaTryDriverCompositeRects(op, pSrc, pDst, nrect, rects) != 1) {
+	n = nrect;
+	r = rects;
+	while (n--) {
+	    ExaCheckComposite (op, pSrc, NULL, pDst,
+			       r->xSrc, r->ySrc,
+			       0, 0,
+			       r->xDst, r->yDst,
+			       r->width, r->height);
+	    r++;
+	}
+    }
+    
+    /************************************************************/
+
+    /* Now we have to flush the damage out from pendingDamage => damage 
+     * Calling DamageDamageRegion has that effect. (We could pass
+     * in an empty region here, but we pass in the same region we
+     * use above; the effect is the same.)
+     */
+
+    DamageDamageRegion(pDst->pDrawable, &region);
+    REGION_UNINIT(pScreen, &region);
+}
+
+static int
 exaTryDriverComposite(CARD8		op,
 		      PicturePtr	pSrc,
 		      PicturePtr	pMask,
commit fcb5949928f1c27f67f40c094c3c673786574422
Author: Owen Taylor <otaylor@huygens.home.fishsoup.net>
Date:   Mon Apr 28 21:00:54 2008 +0200

    EXA: Fix overlapping glyphs in glyph cache
    
    Allocate each cache at a different vertical position in the
    per-format pixmap. Fix width/height confusion when choosing
    the cache for a glyph.

diff --git a/exa/exa_glyphs.c b/exa/exa_glyphs.c
index 55fdb01..851e439 100644
--- a/exa/exa_glyphs.c
+++ b/exa/exa_glyphs.c
@@ -173,12 +173,13 @@ exaRealizeGlyphCaches(ScreenPtr    pScreen,
     for (i = 0; i < EXA_NUM_GLYPH_CACHES; i++) {
 	ExaGlyphCachePtr cache = &pExaScr->glyphCaches[i];
 	int rows;
-	
+
 	if (cache->format != format)
 	    continue;
 
-	rows = (cache->size + cache->columns - 1) / cache->columns;
+	cache->yOffset = height;
 
+	rows = (cache->size + cache->columns - 1) / cache->columns;
 	height += rows * cache->glyphHeight;
     }
 
@@ -346,6 +347,9 @@ exaGlyphCacheHashRemove(ExaGlyphCachePtr cache,
     }
 }
 
+#define CACHE_X(pos) (((pos) % cache->columns) * cache->glyphWidth)
+#define CACHE_Y(pos) (cache->yOffset + ((pos) / cache->columns) * cache->glyphHeight)
+
 static ExaGlyphCacheResult
 exaGlyphCacheBufferGlyph(ScreenPtr         pScreen,
 			 ExaGlyphCachePtr  cache,
@@ -393,8 +397,8 @@ exaGlyphCacheBufferGlyph(ScreenPtr         pScreen,
 		int x, y;
 		int i;
 		
-		x = (pos % cache->columns) * cache->glyphWidth;
-		y = (pos / cache->columns) * cache->glyphHeight;
+		x = CACHE_X(pos);
+		y = CACHE_Y(pos);
 
 		for (i = 0; i < buffer->count; i++) {
 		    if (buffer->rects[i].xSrc == x && buffer->rects[i].ySrc == y) {
@@ -420,8 +424,8 @@ exaGlyphCacheBufferGlyph(ScreenPtr         pScreen,
 			  cache->picture,
 			  0, 0,
 			  0, 0,
-			  (pos % cache->columns) * cache->glyphWidth,
-			  (pos / cache->columns) * cache->glyphHeight,
+			  CACHE_X(pos),
+			  CACHE_Y(pos),
 			  pGlyph->info.width,
 			  pGlyph->info.height);
     }
@@ -430,8 +434,8 @@ exaGlyphCacheBufferGlyph(ScreenPtr         pScreen,
     buffer->source = cache->picture;
 	    
     rect = &buffer->rects[buffer->count];
-    rect->xSrc = (pos % cache->columns) * cache->glyphWidth;
-    rect->ySrc = (pos / cache->columns) * cache->glyphHeight;
+    rect->xSrc = CACHE_X(pos);
+    rect->ySrc = CACHE_Y(pos);
     rect->xDst = xGlyph - pGlyph->info.x;
     rect->yDst = yGlyph - pGlyph->info.y;
     rect->width = pGlyph->info.width;
@@ -442,6 +446,9 @@ exaGlyphCacheBufferGlyph(ScreenPtr         pScreen,
     return ExaGlyphSuccess;
 }
 
+#undef CACHE_X
+#undef CACHE_Y
+
 static ExaGlyphCacheResult
 exaBufferGlyph(ScreenPtr         pScreen,
 	       ExaGlyphBufferPtr buffer,
@@ -452,7 +459,7 @@ exaBufferGlyph(ScreenPtr         pScreen,
     ExaScreenPriv(pScreen);
     unsigned int format = (GlyphPicture(pGlyph)[pScreen->myNum])->format;
     int width = pGlyph->info.width;
-    int height = pGlyph->info.width;
+    int height = pGlyph->info.height;
     ExaCompositeRectPtr rect;
     PicturePtr source;
     int i;
diff --git a/exa/exa_priv.h b/exa/exa_priv.h
index 0d5d0f5..8a17f65 100644
--- a/exa/exa_priv.h
+++ b/exa/exa_priv.h
@@ -61,7 +61,7 @@
 #define DEBUG_MIGRATE		0
 #define DEBUG_PIXMAP		0
 #define DEBUG_OFFSCREEN		0
-#define DEBUG_GLYPH_CACHE	0
+#define DEBUG_GLYPH_CACHE	1
 
 #if DEBUG_TRACE_FALL
 #define EXA_FALLBACK(x)     					\
@@ -121,6 +121,7 @@ typedef struct {
     int glyphCount; /* Current number of glyphs */
     
     PicturePtr picture;   /* Where the glyphs of the cache are stored */
+    int yOffset;          /* y location within the picture where the cache starts */
     int columns;          /* Number of columns the glyphs are layed out in */
     int evictionPosition; /* Next random position to evict a glyph */
 } ExaGlyphCacheRec, *ExaGlyphCachePtr;
commit cc08c06665ffe29ad44d023d75d0f86e5338875d
Author: Owen Taylor <otaylor@huygens.home.fishsoup.net>
Date:   Mon Apr 28 21:00:55 2008 +0200

    EXA: Use UploadToScreen() for uploads to glyph cache
    
    When possible, use UploadToScreen() rather than CompositePicture()
    to upload glyphs onto the glyph cache pixmap. This avoids allocating
    offscreen memory for each glyph making management of offscreen
    areas much more efficient.

diff --git a/exa/exa_glyphs.c b/exa/exa_glyphs.c
index 27ecd4a..95ff4d8 100644
--- a/exa/exa_glyphs.c
+++ b/exa/exa_glyphs.c
@@ -347,6 +347,56 @@ exaGlyphCacheHashRemove(ExaGlyphCachePtr cache,
 #define CACHE_X(pos) (((pos) % cache->columns) * cache->glyphWidth)
 #define CACHE_Y(pos) (cache->yOffset + ((pos) / cache->columns) * cache->glyphHeight)
 
+/* The most efficient thing to way to upload the glyph to the screen
+ * is to use the UploadToScreen() driver hook; this allows us to
+ * pipeline glyph uploads and to avoid creating offscreen pixmaps for
+ * glyphs that we'll never use again.
+ */
+static Bool
+exaGlyphCacheUploadGlyph(ScreenPtr         pScreen,
+			 ExaGlyphCachePtr  cache,
+			 int               pos,
+			 GlyphPtr          pGlyph)
+{
+    ExaScreenPriv(pScreen);
+    PicturePtr pGlyphPicture = GlyphPicture(pGlyph)[pScreen->myNum];
+    PixmapPtr pGlyphPixmap = (PixmapPtr)pGlyphPicture->pDrawable;
+    ExaPixmapPriv(pGlyphPixmap);
+    PixmapPtr pCachePixmap = (PixmapPtr)cache->picture->pDrawable;
+    ExaMigrationRec pixmaps[1];
+    int cacheXoff, cacheYoff;
+
+    if (!pExaScr->info->UploadToScreen || pExaScr->swappedOut || pExaPixmap->accel_blocked)
+	return FALSE;
+
+    /* If the glyph pixmap is already uploaded, no point in doing
+     * things this way */
+    if (exaPixmapIsOffscreen(pGlyphPixmap))
+	return FALSE;
+
+    /* cache pixmap must be offscreen. */
+    pixmaps[0].as_dst = TRUE;
+    pixmaps[0].as_src = FALSE;
+    pixmaps[0].pPix = pCachePixmap;
+    pixmaps[0].pReg = NULL;
+    exaDoMigration (pixmaps, 1, TRUE);
+
+    pCachePixmap = exaGetOffscreenPixmap ((DrawablePtr)pCachePixmap, &cacheXoff, &cacheYoff);
+    if (!pCachePixmap)
+	return FALSE;
+
+    if (!pExaScr->info->UploadToScreen(pCachePixmap,
+				       CACHE_X(pos) + cacheXoff,
+				       CACHE_Y(pos) + cacheYoff,
+				       pGlyph->info.width,
+				       pGlyph->info.height,
+				       (char *)pExaPixmap->sys_ptr,
+				       pExaPixmap->sys_pitch))
+	return FALSE;
+
+    return TRUE;
+}
+
 static ExaGlyphCacheResult
 exaGlyphCacheBufferGlyph(ScreenPtr         pScreen,
 			 ExaGlyphCachePtr  cache,
@@ -413,18 +463,23 @@ exaGlyphCacheBufferGlyph(ScreenPtr         pScreen,
 	    cache->evictionPosition = rand() % cache->size;
 	}
 
-	/* Now actually upload the glyph into the cache picture */
+	/* Now actually upload the glyph into the cache picture; if
+	 * we can't do it with UploadToScreen (because the glyph is
+	 * offscreen, etc), we fall back to CompositePicture.
+	 */
+	if (!exaGlyphCacheUploadGlyph(pScreen, cache, pos, pGlyph)) {
+	    CompositePicture (PictOpSrc,
+			      GlyphPicture(pGlyph)[pScreen->myNum],
+			      None,
+			      cache->picture,
+			      0, 0,
+			      0, 0,
+			      CACHE_X(pos),
+			      CACHE_Y(pos),
+			      pGlyph->info.width,
+			      pGlyph->info.height);
+	}
 
-	CompositePicture (PictOpSrc,
-			  GlyphPicture(pGlyph)[pScreen->myNum],
-			  None,
-			  cache->picture,
-			  0, 0,
-			  0, 0,
-			  CACHE_X(pos),
-			  CACHE_Y(pos),
-			  pGlyph->info.width,
-			  pGlyph->info.height);
     }
     
 
commit e7eaac59c424a205dd106fc7d70734ff4b390f28
Author: Michel Dänzer <michel@tungstengraphics.com>
Date:   Mon Apr 28 21:00:55 2008 +0200

    EXA: Glyph cache upload tweaks.
    
    Track damage after using UploadToScreen directly.
    
    Don't waste any effort on empty glyphs.

diff --git a/exa/exa_glyphs.c b/exa/exa_glyphs.c
index 95ff4d8..b618365 100644
--- a/exa/exa_glyphs.c
+++ b/exa/exa_glyphs.c
@@ -394,6 +394,12 @@ exaGlyphCacheUploadGlyph(ScreenPtr         pScreen,
 				       pExaPixmap->sys_pitch))
 	return FALSE;
 
+    exaPixmapDirty (pCachePixmap,
+		    CACHE_X(pos) + cacheXoff,
+		    CACHE_Y(pos) + cacheYoff,
+		    CACHE_X(pos) + cacheXoff + pGlyph->info.width,
+		    CACHE_Y(pos) + cacheYoff + pGlyph->info.height);
+
     return TRUE;
 }
 
@@ -737,7 +743,8 @@ exaGlyphs (CARD8 	 op,
 	    glyph = *glyphs++;
 	    pPicture = GlyphPicture (glyph)[pScreen->myNum];
 
-	    if (exaBufferGlyph(pScreen, &buffer, glyph, x, y) == ExaGlyphNeedFlush)
+	    if (glyph->info.width > 0 && glyph->info.height > 0 &&
+		exaBufferGlyph(pScreen, &buffer, glyph, x, y) == ExaGlyphNeedFlush)
 	    {
 		if (maskFormat)
 		    exaGlyphsToMask(pMask, &buffer);
commit 8349732a6720652bfbad7874a952be73a0e8e77b
Author: Michel Dänzer <michel@tungstengraphics.com>
Date:   Mon Apr 28 21:09:35 2008 +0200

    EXA: Try to accelerate non-antialiased text via the glyph cache as well.
    
    Treat 1 bit glyphs and masks as PICT_a8 in the glyph cache. We're not able to
    accelerate them otherwise.

diff --git a/exa/exa_glyphs.c b/exa/exa_glyphs.c
index 08ec097..ff665d5 100644
--- a/exa/exa_glyphs.c
+++ b/exa/exa_glyphs.c
@@ -374,6 +374,10 @@ exaGlyphCacheUploadGlyph(ScreenPtr         pScreen,
     if (exaPixmapIsOffscreen(pGlyphPixmap))
 	return FALSE;
 
+    /* UploadToScreen only works if bpp match */
+    if (pGlyphPixmap->drawable.bitsPerPixel != pCachePixmap->drawable.bitsPerPixel)
+	return FALSE;
+
     /* cache pixmap must be offscreen. */
     pixmaps[0].as_dst = TRUE;
     pixmaps[0].as_src = FALSE;
@@ -524,6 +528,9 @@ exaBufferGlyph(ScreenPtr         pScreen,
 
     if (buffer->count == GLYPH_BUFFER_SIZE)
 	return ExaGlyphNeedFlush;
+
+    if (PICT_FORMAT_BPP(format) == 1)
+	format = PICT_a8;
     
     for (i = 0; i < EXA_NUM_GLYPH_CACHES; i++) {
 	ExaGlyphCachePtr cache = &pExaScr->glyphCaches[i];
@@ -796,6 +803,14 @@ exaGlyphs (CARD8 	 op,
 	    return;
 	width = extents.x2 - extents.x1;
 	height = extents.y2 - extents.y1;
+
+	if (maskFormat->depth == 1) {
+	    PictFormatPtr a8Format = PictureMatchFormat (pScreen, 8, PICT_a8);
+
+	    if (a8Format)
+		maskFormat = a8Format;
+	}
+
 	pMaskPixmap = (*pScreen->CreatePixmap) (pScreen, width, height,
 						maskFormat->depth,
 						CREATE_PIXMAP_USAGE_SCRATCH);

commit 29586101dc11d498b212510f8dedbfeca7f8c859
Author: Michel Dänzer <michel@tungstengraphics.com>
Date:   Sat May 24 20:01:41 2008 +0200

    EXA: Only record damage generated by rendering operations.
    
    Recording damage from other operations (e.g. creating a client damage record)
    may confuse the migration code resulting in corruption.
    
    Option "EXAOptimizeMigration" appears safe now, so enable it by default. Also
    remove it from the manpage, as it should only be necessary on request in the
    course of bug report diagnostics anymore.

diff --git a/exa/exa.c b/exa/exa.c
index 809fb4b..fc04748 100644
--- a/exa/exa.c
+++ b/exa/exa.c
@@ -261,6 +261,21 @@ exaSetFbPitch(ExaScreenPrivPtr pExaScr, ExaPixmapPrivPtr pExaPixmap,
                                      pExaScr->info->pixmapPitchAlign);
 }
 
+
+static void
+ExaDamageReport(DamagePtr pDamage, RegionPtr pReg, void *pClosure)
+{
+    PixmapPtr pPixmap = pClosure;
+    ExaPixmapPriv(pPixmap);
+    RegionPtr pDamageReg = DamageRegion(pDamage);
+
+    if (pExaPixmap->pendingDamage) {
+	REGION_UNION(pScreen, pDamageReg, pDamageReg, pReg);
+	pExaPixmap->pendingDamage = FALSE;
+    }
+}
+
+
 /**
  * exaCreatePixmap() creates a new pixmap.
  *
@@ -352,7 +367,7 @@ exaCreatePixmap(ScreenPtr pScreen, int w, int h, int depth,
     pExaPixmap->area = NULL;
 
     /* Set up damage tracking */
-    pExaPixmap->pDamage = DamageCreate (NULL, NULL, DamageReportNone, TRUE,
+    pExaPixmap->pDamage = DamageCreate (ExaDamageReport, NULL, DamageReportRawRegion, TRUE,
 					pScreen, pPixmap);
 
     if (pExaPixmap->pDamage == NULL) {
diff --git a/exa/exa_migration.c b/exa/exa_migration.c
index 5f22474..3c79f68 100644
--- a/exa/exa_migration.c
+++ b/exa/exa_migration.c
@@ -301,6 +301,9 @@ exaDoMoveInPixmap (ExaMigrationPtr migrate)
     ExaScreenPriv (pScreen);
     ExaPixmapPriv (pPixmap);
 
+    if (migrate->as_dst)
+	pExaPixmap->pendingDamage = TRUE;
+
     /* If we're VT-switched away, no touching card memory allowed. */
     if (pExaScr->swappedOut)
 	return;
@@ -369,6 +372,9 @@ exaDoMoveOutPixmap (ExaMigrationPtr migrate)
     PixmapPtr pPixmap = migrate->pPix;
     ExaPixmapPriv (pPixmap);
 
+    if (migrate->as_dst)
+	pExaPixmap->pendingDamage = TRUE;
+
     if (!pExaPixmap->area || exaPixmapIsPinned(pPixmap))
 	return;
 
diff --git a/exa/exa_priv.h b/exa/exa_priv.h
index f3b72ae..9ec2a56 100644
--- a/exa/exa_priv.h
+++ b/exa/exa_priv.h
@@ -226,6 +226,7 @@ typedef struct {
      * location.
      */
     DamagePtr	    pDamage;
+    Bool	    pendingDamage;
     /**
      * The valid regions mark the valid bits (at least, as they're derived from
      * damage, which may be overreported) of a pixmap's system and FB copies.
diff --git a/exa/exa_unaccel.c b/exa/exa_unaccel.c
index c55ef03..5a25764 100644
--- a/exa/exa_unaccel.c
+++ b/exa/exa_unaccel.c
@@ -97,12 +97,15 @@ ExaCheckPutImage (DrawablePtr pDrawable, GCPtr pGC, int depth,
 		 int x, int y, int w, int h, int leftPad, int format,
 		 char *bits)
 {
+    ExaPixmapPriv(exaGetDrawablePixmap(pDrawable));
+
     EXA_FALLBACK(("to %p (%c)\n", pDrawable, exaDrawableLocation(pDrawable)));
     if (exaGCReadsDestination(pDrawable, pGC->planemask, pGC->fillStyle,
 			      pGC->alu))
 	exaPrepareAccess (pDrawable, EXA_PREPARE_DEST);
     else
-	ExaDoPrepareAccess (pDrawable, EXA_PREPARE_DEST);
+	exaPrepareAccessReg (pDrawable, EXA_PREPARE_DEST,
+			     DamagePendingRegion(pExaPixmap->pDamage));
     fbPutImage (pDrawable, pGC, depth, x, y, w, h, leftPad, format, bits);
     exaFinishAccess (pDrawable, EXA_PREPARE_DEST);
 }
diff --git a/hw/xfree86/exa/exa.man.pre b/hw/xfree86/exa/exa.man.pre
index 14859bc..31e1cfe 100644
--- a/hw/xfree86/exa/exa.man.pre
+++ b/hw/xfree86/exa/exa.man.pre
@@ -31,12 +31,6 @@ Disables acceleration of downloading of pixmap data from the framebuffer.
 Not usable with drivers which rely on DownloadFromScreen succeeding.
 Default: No.
 .TP
-.BI "Option \*qEXAOptimizeMigration\*q \*q" boolean \*q
-Enables an additional optimization for migration of destination pixmaps. This
-may improve performance in some cases (e.g. when switching virtual desktops with
-no compositing manager) but causes corruption in others (e.g. when starting
-compiz). Default: No.
-.TP
 .BI "Option \*qMigrationHeuristic\*q \*q" anystr \*q
 Chooses an alternate pixmap migration heuristic, for debugging purposes.  The
 default is intended to be the best performing one for general use, though others
diff --git a/hw/xfree86/exa/examodule.c b/hw/xfree86/exa/examodule.c
index e18da0a..63ea8c5 100644
--- a/hw/xfree86/exa/examodule.c
+++ b/hw/xfree86/exa/examodule.c
@@ -145,7 +145,7 @@ exaDDXDriverInit(ScreenPtr pScreen)
 	pExaScr->optimize_migration =
 	    xf86ReturnOptValBool(pScreenPriv->options,
 				 EXAOPT_OPTIMIZE_MIGRATION,
-				 FALSE);
+				 TRUE);
     }
 
     if (xf86ReturnOptValBool(pScreenPriv->options,
commit 13fd2256300b61d88b840952d838f834523f5dd7
Author: Owen Taylor <otaylor@huygens.home.fishsoup.net>
Date:   Mon Apr 28 21:00:55 2008 +0200

    EXA: Clean up debug messages

diff --git a/exa/exa_glyphs.c b/exa/exa_glyphs.c
index 851e439..27ecd4a 100644
--- a/exa/exa_glyphs.c
+++ b/exa/exa_glyphs.c
@@ -260,11 +260,9 @@ exaGlyphCacheHashLookup(ExaGlyphCachePtr cache,
 	    return -1;
 
 	if (memcmp(pGlyph->sha1, cache->glyphs[entryPos].sha1, sizeof(pGlyph->sha1)) == 0){
-	    DBG_GLYPH_CACHE((" found entry at %d\n", slot));
 	    return entryPos;
 	}
 	    
-	DBG_GLYPH_CACHE((" lookup linear probe bumpalong\n"));
 	slot--;
 	if (slot < 0)
 	    slot = cache->hashSize - 1;
@@ -284,7 +282,6 @@ exaGlyphCacheHashInsert(ExaGlyphCachePtr cache,
     
     while (TRUE) { /* hash table can never be full */
 	if (cache->hashEntries[slot] == -1) {
-	    DBG_GLYPH_CACHE((" inserting entry at %d\n", slot));
 	    cache->hashEntries[slot] = pos;
 	    return;
 	}
diff --git a/exa/exa_priv.h b/exa/exa_priv.h
index 8a17f65..f3b72ae 100644
--- a/exa/exa_priv.h
+++ b/exa/exa_priv.h
@@ -61,7 +61,7 @@
 #define DEBUG_MIGRATE		0
 #define DEBUG_PIXMAP		0
 #define DEBUG_OFFSCREEN		0
-#define DEBUG_GLYPH_CACHE	1
+#define DEBUG_GLYPH_CACHE	0
 
 #if DEBUG_TRACE_FALL
 #define EXA_FALLBACK(x)     					\