Re: [PATCH] Reduce the number of expensive division instructions doneby _parse_integer()

[Linus Torvalds]

On Thu, Feb 9, 2012 at 8:58 AM, Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> So David's patch looks ok, although if this really is
> performance-critical (and I could imagine some crazy /proc or /sys
> access loads where the divide really does show up very clearly) I do
> think it could be even done more efficiently. But probably only if we
> start doing some arch-specific stuff.

Looking at the code generated, the "val >> 60" thing actually does
generate a shift, and at least on x86-64, the attached patch generates
better code.

David, does this work for you? I also removed the nasty pointer
indirection crap.

                        Linus
 lib/kstrtox.c |   18 +++++++++++++-----
 1 files changed, 13 insertions(+), 5 deletions(-)

diff --git a/lib/kstrtox.c b/lib/kstrtox.c
index 7a94c8f14e29..b1dd3e7d88cb 100644
--- a/lib/kstrtox.c
+++ b/lib/kstrtox.c
@@ -44,12 +44,13 @@ const char *_parse_integer_fixup_radix(const char *s, unsigned int *base)
  *
  * Don't you dare use this function.
  */
-unsigned int _parse_integer(const char *s, unsigned int base, unsigned long long *res)
+unsigned int _parse_integer(const char *s, unsigned int base, unsigned long long *p)
 {
+	unsigned long long res;
 	unsigned int rv;
 	int overflow;
 
-	*res = 0;
+	res = 0;
 	rv = 0;
 	overflow = 0;
 	while (*s) {
@@ -64,12 +65,19 @@ unsigned int _parse_integer(const char *s, unsigned int base, unsigned long long
 
 		if (val >= base)
 			break;
-		if (*res > div_u64(ULLONG_MAX - val, base))
-			overflow = 1;
-		*res = *res * base + val;
+		/*
+		 * Check for overflow only if we are within range of
+		 * it in the max base we support (16)
+		 */
+		if (unlikely(res & (~0ull << 60))) {
+			if (res > div_u64(ULLONG_MAX - val, base))
+				overflow = 1;
+		}
+		res = res * base + val;
 		rv++;
 		s++;
 	}
+	*p = res;
 	if (overflow)
 		rv |= KSTRTOX_OVERFLOW;
 	return rv;