[OPTIM] halog: add assembly version of the field lookup code
Gcc tries to be a bit too smart in these small loops and the result is
that on i386 we waste a lot of time there. By recoding these loops in
assembly, we save up to 23% total processing time on i386! The savings
on x86_64 are much lower, probably because there are more registers and
gcc has to do less tricks. However, those savings vary a lot between gcc
versions and even cause harm on some of them (eg: 4.4) because gcc does
not know how to optimize the code once inlined.
However, by recoding field_start() in C to try to match the assembly
code as much as possible, we can significantly reduce its execution
time without risking the negative impacts. Thus, the assembly version
is less interesting there but still worth being used on some compilers.
diff --git a/contrib/halog/halog.c b/contrib/halog/halog.c
index 32ca670..ad90510 100644
--- a/contrib/halog/halog.c
+++ b/contrib/halog/halog.c
@@ -135,6 +135,32 @@
/* return pointer to first char not part of current field starting at <p>. */
+
+#if defined(__i386__)
+/* this one is always faster on 32-bits */
+static inline const char *field_stop(const char *p)
+{
+ asm(
+ /* Look for spaces */
+ "4: \n\t"
+ "inc %0 \n\t"
+ "cmpb $0x20, -1(%0) \n\t"
+ "ja 4b \n\t"
+ "jz 3f \n\t"
+
+ /* we only get there for control chars 0..31. Leave if we find '\0' */
+ "cmpb $0x0, -1(%0) \n\t"
+ "jnz 4b \n\t"
+
+ /* return %0-1 = position of the last char we checked */
+ "3: \n\t"
+ "dec %0 \n\t"
+ : "=r" (p)
+ : "0" (p)
+ );
+ return p;
+}
+#else
const char *field_stop(const char *p)
{
unsigned char c;
@@ -148,6 +174,7 @@
}
return p - 1;
}
+#endif
/* return field <field> (starting from 1) in string <p>. Only consider
* contiguous spaces (or tabs) as one delimiter. May return pointer to
@@ -155,39 +182,81 @@
*/
const char *field_start(const char *p, int field)
{
+#ifndef PREFER_ASM
unsigned char c;
while (1) {
/* skip spaces */
while (1) {
- c = *p;
+ c = *(p++);
if (c > ' ')
break;
if (c == ' ')
- goto next;
+ continue;
if (!c) /* end of line */
- return p;
+ return p-1;
/* other char => new field */
break;
- next:
- p++;
}
/* start of field */
field--;
if (!field)
- return p;
+ return p-1;
/* skip this field */
while (1) {
c = *(p++);
- if (c > ' ')
- continue;
if (c == ' ')
break;
+ if (c > ' ')
+ continue;
if (c == '\0')
- return p;
+ return p - 1;
}
}
+#else
+ /* This version works optimally on i386 and x86_64 but the code above
+ * shows similar performance. However, depending on the version of GCC
+ * used, inlining rules change and it may have difficulties to make
+ * efficient use of this code at other locations and could result in
+ * worse performance (eg: gcc 4.4). You may want to experience.
+ */
+ asm(
+ /* skip spaces */
+ "1: \n\t"
+ "inc %0 \n\t"
+ "cmpb $0x20, -1(%0) \n\t"
+ "ja 2f \n\t"
+ "jz 1b \n\t"
+
+ /* we only get there for control chars 0..31. Leave if we find '\0' */
+ "cmpb $0x0, -1(%0) \n\t"
+ "jz 3f \n\t"
+
+ /* start of field at [%0-1]. Check if we need to skip more fields */
+ "2: \n\t"
+ "dec %1 \n\t"
+ "jz 3f \n\t"
+
+ /* Look for spaces */
+ "4: \n\t"
+ "inc %0 \n\t"
+ "cmpb $0x20, -1(%0) \n\t"
+ "jz 1b \n\t"
+ "ja 4b \n\t"
+
+ /* we only get there for control chars 0..31. Leave if we find '\0' */
+ "cmpb $0x0, -1(%0) \n\t"
+ "jnz 4b \n\t"
+
+ /* return %0-1 = position of the last char we checked */
+ "3: \n\t"
+ "dec %0 \n\t"
+ : "=r" (p)
+ : "r" (field), "0" (p)
+ );
+ return p;
+#endif
}
/* keep only the <bits> higher bits of <i> */