[PATCH xserver 3/7] EXA: Hardcode glyph cache dimensions to allow more efficient code generation.

[Michel Dänzer]

From: Michel Dänzer <daenzer at vmware.com>

Signed-off-by: Michel Dänzer <daenzer at vmware.com>
---
 exa/exa_glyphs.c |   43 ++++++++++++++++++++-----------------------
 exa/exa_priv.h   |    3 ---
 2 files changed, 20 insertions(+), 26 deletions(-)

diff --git a/exa/exa_glyphs.c b/exa/exa_glyphs.c
index 28aa01f..5711ca7 100644
--- a/exa/exa_glyphs.c
+++ b/exa/exa_glyphs.c
@@ -60,7 +60,10 @@
  * max texture size of the driver; this may need to actually come from
  * the driver.
  */
-#define CACHE_PICTURE_WIDTH 1024
+#define CACHE_COLUMNS (1024 / 32)
+
+#define CACHE_SIZE 256
+#define HASH_SIZE 557
 
 /* Maximum number of glyphs we buffer on the stack before flushing
  * rendering to the mask or destination surface.
@@ -101,12 +104,6 @@ exaGlyphsInit(ScreenPtr pScreen)
     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
@@ -178,13 +175,13 @@ exaRealizeGlyphCaches(ScreenPtr    pScreen,
 
 	cache->yOffset = height;
 
-	rows = (cache->size + cache->columns - 1) / cache->columns;
+	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,
+					CACHE_COLUMNS * 32,
 					height, depth, 0);
     if (!pPixmap)
 	return FALSE;
@@ -209,14 +206,14 @@ exaRealizeGlyphCaches(ScreenPtr    pScreen,
 
 	cache->picture = pPicture;
 	cache->picture->refcnt++;
-	cache->hashEntries = malloc(sizeof(int) * cache->hashSize);
-	cache->glyphs = malloc(sizeof(ExaCachedGlyphRec) * cache->size);
+	cache->hashEntries = malloc(sizeof(int) * HASH_SIZE);
+	cache->glyphs = malloc(sizeof(ExaCachedGlyphRec) * CACHE_SIZE);
 	cache->glyphCount = 0;
 
 	if (!cache->hashEntries || !cache->glyphs)
 	    goto bail;
 
-	for (j = 0; j < cache->hashSize; j++)
+	for (j = 0; j < HASH_SIZE; j++)
 	    cache->hashEntries[j] = -1;
 	
 	cache->evictionPosition = 0;
@@ -251,7 +248,7 @@ exaGlyphCacheHashLookup(ExaGlyphCachePtr cache,
 {
     int slot;
 
-    slot = (*(CARD32 *) pGlyph->sha1) % cache->hashSize;
+    slot = (*(CARD32 *) pGlyph->sha1) % HASH_SIZE;
     
     while (TRUE) { /* hash table can never be full */
 	int entryPos = cache->hashEntries[slot];
@@ -264,7 +261,7 @@ exaGlyphCacheHashLookup(ExaGlyphCachePtr cache,
 	    
 	slot--;
 	if (slot < 0)
-	    slot = cache->hashSize - 1;
+	    slot = HASH_SIZE - 1;
     }
 }
 
@@ -277,7 +274,7 @@ exaGlyphCacheHashInsert(ExaGlyphCachePtr cache,
 
     memcpy(cache->glyphs[pos].sha1, pGlyph->sha1, sizeof(pGlyph->sha1));
     
-    slot = (*(CARD32 *) pGlyph->sha1) % cache->hashSize;
+    slot = (*(CARD32 *) pGlyph->sha1) % HASH_SIZE;
     
     while (TRUE) { /* hash table can never be full */
 	if (cache->hashEntries[slot] == -1) {
@@ -287,7 +284,7 @@ exaGlyphCacheHashInsert(ExaGlyphCachePtr cache,
 	    
 	slot--;
 	if (slot < 0)
-	    slot = cache->hashSize - 1;
+	    slot = HASH_SIZE - 1;
     }
 }
 
@@ -298,7 +295,7 @@ exaGlyphCacheHashRemove(ExaGlyphCachePtr cache,
     int slot;
     int emptiedSlot = -1;
 
-    slot = (*(CARD32 *) cache->glyphs[pos].sha1) % cache->hashSize;
+    slot = (*(CARD32 *) cache->glyphs[pos].sha1) % HASH_SIZE;
 
     while (TRUE) { /* hash table can never be full */
 	int entryPos = cache->hashEntries[slot];
@@ -326,7 +323,7 @@ exaGlyphCacheHashRemove(ExaGlyphCachePtr cache,
 	     * (Knuth 6.4R)
 	     */
 	    
-	    int entrySlot = (*(CARD32 *) cache->glyphs[entryPos].sha1) % cache->hashSize;
+	    int entrySlot = (*(CARD32 *) cache->glyphs[entryPos].sha1) % HASH_SIZE;
 
 	    if (!((entrySlot >= slot && entrySlot < emptiedSlot) ||
 		  (emptiedSlot < slot && (entrySlot < emptiedSlot || entrySlot >= slot)))) 
@@ -339,12 +336,12 @@ exaGlyphCacheHashRemove(ExaGlyphCachePtr cache,
 	
 	slot--;
 	if (slot < 0)
-	    slot = cache->hashSize - 1;
+	    slot = HASH_SIZE - 1;
     }
 }
 
-#define CACHE_X(pos) (((pos) % cache->columns) * cache->glyphWidth)
-#define CACHE_Y(pos) (cache->yOffset + ((pos) / cache->columns) * cache->glyphHeight)
+#define CACHE_X(pos) (((pos) & (CACHE_COLUMNS - 1)) * 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
@@ -463,7 +460,7 @@ exaGlyphCacheBufferGlyph(ScreenPtr         pScreen,
 	x = CACHE_X(pos);
 	y = CACHE_Y(pos);
     } else {
-	if (cache->glyphCount < cache->size) {
+	if (cache->glyphCount < CACHE_SIZE) {
 	    /* Space remaining; we fill from the start */
 	    pos = cache->glyphCount;
 	    x = CACHE_X(pos);
@@ -500,7 +497,7 @@ exaGlyphCacheBufferGlyph(ScreenPtr         pScreen,
 	    exaGlyphCacheHashInsert(cache, pGlyph, pos);
 
 	    /* And pick a new eviction position */
-	    cache->evictionPosition = (pos + 1) % cache->size;
+	    cache->evictionPosition = (pos + 1) & (CACHE_SIZE - 1);
 	}
 
 	exaGlyphCacheUploadGlyph(pScreen, cache, x, y, pGlyph);
diff --git a/exa/exa_priv.h b/exa/exa_priv.h
index 8c5ea15..7d2ea2c 100644
--- a/exa/exa_priv.h
+++ b/exa/exa_priv.h
@@ -115,8 +115,6 @@ typedef struct {
     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
@@ -124,7 +122,6 @@ typedef struct {
      * here because it allows us to easily remove entries.)
      */
     int *hashEntries;
-    int hashSize;
     
     ExaCachedGlyphPtr glyphs;
     int glyphCount; /* Current number of glyphs */
-- 
1.7.5.1