Skip to contentSkip to conversation

Avoid use of divides in print_size

PATCH New Nick Thompson 2 participants 2 messages May 11, 2010 Library code

lore
Share & tools
lore
#0 lore raw

PATCH v3 1/1 Avoid use of divides in print_size

Modification of print_size to avoid use of divides and especially long long divides. Keep the binary scale factor in terms of bit shifts instead. This should be faster, since the previous code gave the compiler no clues that the divides where always powers of two, preventing optimisation.

---
Since v1:
Removed misplaced brackets that resulted in round-up value of 5 rather than the intended 0.5. Now passes Timur Tabi's test.

Since v2:
Fix overflow in calculation of 0.5 rounding value for numbers with engineered factional parts representing values >4GiB

 lib/display_options.c |   14 ++++++++------
 1 files changed, 8 insertions(+), 6 deletions(-)
lib/display_options.c +8 -6
index 86df05d..a711425 100644
@@ -46,13 +46,14 @@ int display_options (void)
 void print_size(unsigned long long size, const char *s)
 {
 	unsigned long m = 0, n;
+	unsigned long long f;
 	static const char names[] = {'E', 'P', 'T', 'G', 'M', 'K'};
-	unsigned long long d = 1ULL << (10 * ARRAY_SIZE(names));
+	unsigned long d = 10 * ARRAY_SIZE(names);
 	char c = 0;
 	unsigned int i;
 
-	for (i = 0; i < ARRAY_SIZE(names); i++, d >>= 10) {
-		if (size >= d) {
+	for (i = 0; i < ARRAY_SIZE(names); i++, d -= 10) {
+		if (size >> d) {
 			c = names[i];
 			break;
 		}
@@ -63,11 +64,12 @@ void print_size(unsigned long long size, const char *s)
 		return;
 	}
 
-	n = size / d;
+	n = size >> d;
+	f = size & ((1ULL << d) - 1);
 
 	/* If there's a remainder, deal with it */
-	if(size % d) {
-		m = (10 * (size - (n * d)) + (d / 2) ) / d;
+	if (f) {
+		m = (10ULL * f + (1ULL << (d - 1))) >> d;
 
 		if (m >= 10) {
 			m -= 10;
Signature · 1 line
1.7.0.4
#1 lore raw

PATCH v3 1/1 Avoid use of divides in print_size

Quoted 8 lines · Nick Thompson

Modification of print_size to avoid use of divides and especially long long divides. Keep the binary scale factor in terms of bit shifts instead. This should be faster, since the previous code gave the compiler no clues that the divides where always powers of two, preventing optimisation.

---

I don't know if this solves the problem on ARM, but it works on PowerPC.

Signature · 2 lines
Timur Tabi
Linux kernel developer at Freescale