The MEMSET() macro is now an order of magnitude *slower* than glibc.

It has been removed and relegated to the annals of SCM history.

7 files changed, 59 insertions, 89 deletions

diff --git a/include/libast.h b/include/libast.h
index 8d243f1..cf22050 100644
--- a/include/libast.h
+++ b/include/libast.h
@@ -1257,76 +1257,6 @@ extern int re_exec();
 # define GC_DUMP()                              NOP
 #endif
 
-/* Fast memset() macro contributed by vendu */
-#if !defined(SIZEOF_LONG) || (SIZEOF_LONG == 8)
-/** UNDOCUMENTED */
-# define MEMSET_LONG() (l |= l<<32)
-#else
-/** UNDOCUMENTED */
-# define MEMSET_LONG() NOP
-#endif
-
-/**
- * @def MEMSET(s, c, count)
- * Initialize a memory region to a particular value.
- *
- * This macro is a replacement for the libc function memset().  It
- * initializes the memory region pointed to by @a s to the value
- * specified by @a c.  The size of the memory region is specified by
- * @a count.  Note that @a c must be a byte (char) value.
- *
- * This macro has been optimized to set as many bytes simultaneously as
- * the architecture can handle, so it should offer superior
- * performance to libc's memset() function.
- *
- * @param s     A pointer to the memory region to initialize.
- * @param c     The value to which all bytes in the block will be
- *              set.
- * @param count The size, in bytes, of the memory region.
- * @see @link DOXGRP_MEM Memory Management Subsystem @endlink
- * @ingroup DOXGRP_MEM
- */
-#define MEMSET(s, c, count) do { \
-    char *end = (char *)(s) + (count); \
-    long l; \
-    long *l_dest = (long *)(s); \
-    char *c_dest; \
- \
-    if (!(s)) { \
-        break; \
-    } \
-    /* areas of less than 4 * sizeof(long) are set in 1-byte chunks. */ \
-    if (((unsigned long) count) >= 4 * sizeof(long)) { \
-        /* fill l with c. */ \
-        l = (c) | (c)<<8; \
-        l |= l<<16; \
-        MEMSET_LONG(); \
- \
-        /* fill in 1-byte chunks until boundary of long is reached. */ \
-        if ((unsigned long)l_dest & (unsigned long)(sizeof(long) -1)) { \
-            c_dest = (char *)l_dest; \
-            while ((unsigned long)c_dest & (unsigned long)(sizeof(long) -1)) { \
-                *(c_dest++) = (c); \
-            } \
-            l_dest = (long *)c_dest; \
-        } \
- \
-        /* fill in long-size chunks as long as possible. */ \
-        while (((unsigned long) (end - (char *)l_dest)) >= sizeof(long)) { \
-            *(l_dest++) = l; \
-        } \
-    } \
- \
-    /* fill the tail in 1-byte chunks. */ \
-    if ((char *)l_dest < end) { \
-        c_dest = (char *)l_dest; \
-        *(c_dest++) = (c); \
-        while (c_dest < end) { \
-            *(c_dest++) = (c); \
-        } \
-    } \
-  } while (0)
-
 
 
 /******************************* STRINGS GOOP *********************************/
diff --git a/src/array.c b/src/array.c
index 55fc532..f32a7b0 100644
--- a/src/array.c
+++ b/src/array.c
@@ -683,7 +683,7 @@ spif_array_insert_at(spif_array_t self, spif_obj_t obj, spif_listidx_t idx)
     } else if (left < 0) {
         /* NULL out the new gap in the list. */
         left = -left;
-        MEMSET(self->items + (idx - left), 0, sizeof(spif_obj_t) * left);
+        memset(self->items + (idx - left), 0, sizeof(spif_obj_t) * left);
     }
     self->items[idx] = obj;
     self->len++;
diff --git a/src/conf.c b/src/conf.c
index 49fe9b2..b684517 100644
--- a/src/conf.c
+++ b/src/conf.c
@@ -74,7 +74,7 @@ spifconf_init_subsystem(void)
     ctx_cnt = 20;
     ctx_idx = 0;
     context = (ctx_t *) MALLOC(sizeof(ctx_t) * ctx_cnt);
-    MEMSET(context, 0, sizeof(ctx_t) * ctx_cnt);
+    memset(context, 0, sizeof(ctx_t) * ctx_cnt);
     context[0].name = (spif_charptr_t) STRDUP("null");
     context[0].handler = parse_null;
 
@@ -82,19 +82,19 @@ spifconf_init_subsystem(void)
     ctx_state_cnt = 20;
     ctx_state_idx = 0;
     ctx_state = (ctx_state_t *) MALLOC(sizeof(ctx_state_t) * ctx_state_cnt);
-    MEMSET(ctx_state, 0, sizeof(ctx_state_t) * ctx_state_cnt);
+    memset(ctx_state, 0, sizeof(ctx_state_t) * ctx_state_cnt);
 
     /* Initialize the file state stack */
     fstate_cnt = 10;
     fstate_idx = 0;
     fstate = (fstate_t *) MALLOC(sizeof(fstate_t) * fstate_cnt);
-    MEMSET(fstate, 0, sizeof(fstate_t) * fstate_cnt);
+    memset(fstate, 0, sizeof(fstate_t) * fstate_cnt);
 
     /* Initialize the builtin function table */
     builtin_cnt = 10;
     builtin_idx = 0;
     builtins = (spifconf_func_t *) MALLOC(sizeof(spifconf_func_t) * builtin_cnt);
-    MEMSET(builtins, 0, sizeof(spifconf_func_t) * builtin_cnt);
+    memset(builtins, 0, sizeof(spifconf_func_t) * builtin_cnt);
 
     /* Register the omni-present builtin functions */
     spifconf_register_builtin("appname", builtin_appname);
@@ -203,7 +203,7 @@ spifconf_new_var(void)
     spifconf_var_t *v;
 
     v = (spifconf_var_t *) MALLOC(sizeof(spifconf_var_t));
-    MEMSET(v, 0, sizeof(spifconf_var_t));
+    memset(v, 0, sizeof(spifconf_var_t));
     return v;
 }
 
diff --git a/src/pthreads.c b/src/pthreads.c
index 2fe76d6..5948e16 100644
--- a/src/pthreads.c
+++ b/src/pthreads.c
@@ -343,7 +343,7 @@ spif_pthreads_tls_calloc(spif_pthreads_t self, size_t count, size_t size)
 
     handle = spif_pthreads_tls_malloc(self, count * size);
     data = spif_pthreads_tls_get(self, handle);
-    MEMSET(data, 0, count * size);
+    memset(data, 0, count * size);
     return handle;
 }
 
diff --git a/test/perf.c b/test/perf.c
index 038dcd0..c0e9f11 100644
--- a/test/perf.c
+++ b/test/perf.c
@@ -680,7 +680,7 @@ perf_url(void)
     spif_charptr_t tmp3 = "/path/to/some/file.jpg";
     spif_charptr_t tmp4 = "pop3://dummy:moo@pop.nowhere.com:110";
 
-    PERF_SET_REPS(100);
+    PERF_SET_REPS(10);
 
     PERF_BEGIN("spif_url_new_from_ptr() function");
     PERF_TEST(testurl = spif_url_new_from_ptr(tmp1); spif_url_del(testurl););
diff --git a/test/perf.h b/test/perf.h
index 625455a..fff2f3a 100644
--- a/test/perf.h
+++ b/test/perf.h
@@ -24,6 +24,56 @@
 #ifndef _LIBAST_PERF_H_
 #  define _LIBAST_PERF_H_
 
+/* "Fast" memset() macro contributed by vendu.  Now an order of magnitude
+ * slower than glibc, so it's only included here for performance testing
+ * in the future.  It very likely should not be used.
+ */
+#if !defined(SIZEOF_LONG) || (SIZEOF_LONG == 8)
+# define MEMSET_LONG() (l |= l<<32)
+#else
+# define MEMSET_LONG() NOP
+#endif
+#define MEMSET(s, c, count) do { \
+    char *end = (char *)(s) + (count); \
+    long l; \
+    long *l_dest = (long *)(s); \
+    char *c_dest; \
+ \
+    if (!(s)) { \
+        break; \
+    } \
+    /* areas of less than 4 * sizeof(long) are set in 1-byte chunks. */ \
+    if (((unsigned long) count) >= 4 * sizeof(long)) { \
+        /* fill l with c. */ \
+        l = (c) | (c)<<8; \
+        l |= l<<16; \
+        MEMSET_LONG(); \
+ \
+        /* fill in 1-byte chunks until boundary of long is reached. */ \
+        if ((unsigned long)l_dest & (unsigned long)(sizeof(long) -1)) { \
+            c_dest = (char *)l_dest; \
+            while ((unsigned long)c_dest & (unsigned long)(sizeof(long) -1)) { \
+                *(c_dest++) = (c); \
+            } \
+            l_dest = (long *)c_dest; \
+        } \
+ \
+        /* fill in long-size chunks as long as possible. */ \
+        while (((unsigned long) (end - (char *)l_dest)) >= sizeof(long)) { \
+            *(l_dest++) = l; \
+        } \
+    } \
+ \
+    /* fill the tail in 1-byte chunks. */ \
+    if ((char *)l_dest < end) { \
+        c_dest = (char *)l_dest; \
+        *(c_dest++) = (c); \
+        while (c_dest < end) { \
+            *(c_dest++) = (c); \
+        } \
+    } \
+  } while (0)
+
 #  define PERF_SET_REPS(c)     do {rep_cnt = (c) * rep_mult; if (rep_cnt < (c)) {rep_cnt = (size_t) -1;}} while (0)
 #  define TDIFF(t1, t2)        (((t2).tv_sec + ((double) (t2).tv_usec / 1000000.0)) \
                                  - ((t1).tv_sec + ((double) (t1).tv_usec / 1000000.0)))
diff --git a/test/test.c b/test/test.c
index 1fdff41..16d2a18 100644
--- a/test/test.c
+++ b/test/test.c
@@ -60,16 +60,6 @@ test_macros(void)
     unsigned long sl1 = 0x98765432, sl2 = 0xffeeddff;
     void *vp1 = &sc1, *vp2 = &sc2;
 
-    TEST_BEGIN("MEMSET() macro");
-    MEMSET(memset_test, '!', CONST_STRLEN(memset_test));
-    TEST_FAIL_IF(strcmp((char *) memset_test, "!!!!!!!!!!!!!!!!!!!!!!!!!!"));
-    MEMSET(memset_test + 3, '*', 14);
-    TEST_FAIL_IF(strcmp((char *) memset_test, "!!!**************!!!!!!!!!"));
-    MEMSET(memset_test, '&', 0 );
-    TEST_FAIL_IF(strcmp((char *) memset_test, "!!!**************!!!!!!!!!"));
-    MEMSET((spif_charptr_t) NULL, '_', CONST_STRLEN(memset_test));
-    TEST_PASS();
-
     TEST_BEGIN("SWAP() macro");
     SWAP(sc1, sc2);
     SWAP(si1, si2);
@@ -2709,7 +2699,7 @@ test_hash_functions(void)
                     && (key_length % 4)) {
                     continue;
                 }
-                MEMSET(buff, 0, sizeof(buff));
+                memset(buff, 0, sizeof(buff));
 
                 if (hash_func == spifhash_jenkins32) {
                     ref_hash = hash_func(pbuff, key_length / 4, 1);