e6fdd7e
--- valgrind-3.2.3/cachegrind/cg_sim.c.jj	2007-01-08 02:43:10.000000000 -0500
e6fdd7e
+++ valgrind-3.2.3/cachegrind/cg_sim.c	2007-02-13 07:15:46.000000000 -0500
8f5f0f5
@@ -42,29 +42,32 @@ typedef struct {
8f5f0f5
    Int          size;                   /* bytes */
8f5f0f5
    Int          assoc;
8f5f0f5
    Int          line_size;              /* bytes */
8f5f0f5
-   Int          sets;
8f5f0f5
    Int          sets_min_1;
8f5f0f5
    Int          assoc_bits;
8f5f0f5
    Int          line_size_bits;
8f5f0f5
    Int          tag_shift;
8f5f0f5
-   Char         desc_line[128];
8f5f0f5
    UWord*       tags;
8f5f0f5
-} cache_t2;
8f5f0f5
+   Char         desc_line[128];
8f5f0f5
+} cache_t2
8f5f0f5
+#ifdef __GNUC__
8f5f0f5
+__attribute__ ((aligned (8 * sizeof (Int))))
8f5f0f5
+#endif
8f5f0f5
+;
8f5f0f5
 
8f5f0f5
 /* By this point, the size/assoc/line_size has been checked. */
8f5f0f5
 static void cachesim_initcache(cache_t config, cache_t2* c)
8f5f0f5
 {
8f5f0f5
-   Int i;
8f5f0f5
+   Int sets;
8f5f0f5
 
8f5f0f5
    c->size      = config.size;
8f5f0f5
    c->assoc     = config.assoc;
8f5f0f5
    c->line_size = config.line_size;
8f5f0f5
 
8f5f0f5
-   c->sets           = (c->size / c->line_size) / c->assoc;
8f5f0f5
-   c->sets_min_1     = c->sets - 1;
8f5f0f5
+   sets              = (c->size / c->line_size) / c->assoc;
8f5f0f5
+   c->sets_min_1     = sets - 1;
8f5f0f5
    c->assoc_bits     = VG_(log2)(c->assoc);
8f5f0f5
    c->line_size_bits = VG_(log2)(c->line_size);
8f5f0f5
-   c->tag_shift      = c->line_size_bits + VG_(log2)(c->sets);
8f5f0f5
+   c->tag_shift      = c->line_size_bits + VG_(log2)(sets);
8f5f0f5
 
8f5f0f5
    if (c->assoc == 1) {
8f5f0f5
       VG_(sprintf)(c->desc_line, "%d B, %d B, direct-mapped", 
8f5f0f5
@@ -74,10 +77,7 @@ static void cachesim_initcache(cache_t c
8f5f0f5
                                  c->size, c->line_size, c->assoc);
8f5f0f5
    }
8f5f0f5
 
8f5f0f5
-   c->tags = VG_(malloc)(sizeof(UWord) * c->sets * c->assoc);
8f5f0f5
-
8f5f0f5
-   for (i = 0; i < c->sets * c->assoc; i++)
8f5f0f5
-      c->tags[i] = 0;
8f5f0f5
+   c->tags = VG_(calloc)(sizeof(UWord) * sets, c->assoc);
8f5f0f5
 }
8f5f0f5
 
8f5f0f5
 /* This is done as a macro rather than by passing in the cache_t2 as an 
8f5f0f5
@@ -140,8 +140,7 @@ void cachesim_##L##_doref(Addr a, UChar 
8f5f0f5
       return;                                                               \
8f5f0f5
                                                                             \
8f5f0f5
    /* Second case: word straddles two lines. */                             \
8f5f0f5
-   /* Nb: this is a fast way of doing ((set1+1) % L.sets) */                \
8f5f0f5
-   } else if (((set1 + 1) & (L.sets-1)) == set2) {                          \
8f5f0f5
+   } else if (((set1 + 1) & (L.sets_min_1)) == set2) {                      \
8f5f0f5
       set = &(L.tags[set1 << L.assoc_bits]);                                \
8f5f0f5
       if (tag == set[0]) {                                                  \
8f5f0f5
          goto block2;                                                       \