From: | Arjan van de Ven <arjan(at)linux(dot)intel(dot)com> |
---|---|
To: | pgsql-hackers(at)postgresql(dot)org |
Cc: | tgl(at)sss(dot)pgh(dot)pa(dot)us, mark(dot)dilger(at)enterprisedb(dot)com, andres(at)anarazel(dot)de |
Subject: | [PATCH v2] src/port/snprintf.c: Optimize the common base=10 case in fmtint |
Date: | 2021-10-26 20:58:17 |
Message-ID: | 6e51c644-1b6d-956e-ac24-2d1b0541d532@linux.intel.com |
Views: | Raw Message | Whole Thread | Download mbox | Resend email |
Thread: | |
Lists: | pgsql-hackers |
[PATCH v2] src/port/snprintf.c: Optimize the common base=10 case in fmtint
fmtint() turns an integer into a string for a given base, and to do this
it does a divide/modulo operation iteratively.
The only possible base values are 8, 10 and 16
On just about any CPU, divides are a pretty expensive operation, generally
10x to 20x or more expensive than adds or multiplies.
By special casing the base values, the compiler (gcc or other) can (and will)
replace the divide by a multiply with 0xcccccccccccccccd (for base 10) or bitops
for base 8 and 16, yielding a lot faster code.
I considered a switch statement, but since base 10 is the most common by far,
I implemented it as a series of if/else statements with a likely() marking the 10 case.
Even though this only shows up in the database creation phase of pgbench and not so much
during the normal run time, the optimization is simple and high value enough that
in my opinion it's worth doing
diff --git a/src/port/snprintf.c b/src/port/snprintf.c
index 7c21429369..547a59d4a0 100644
--- a/src/port/snprintf.c
+++ b/src/port/snprintf.c
@@ -1076,11 +1076,31 @@ fmtint(long long value, char type, int forcesign, int leftjust,
else
{
/* make integer string */
- do
- {
- convert[sizeof(convert) - (++vallen)] = cvt[uvalue % base];
- uvalue = uvalue / base;
- } while (uvalue);
+
+ /*
+ * Special case each of the possible base values (8, 10, 16) to avoid an
+ * expensive divide operation
+ * (the compiler will use a multiply, shift or boolean ops for this)
+ */
+ if (likely(base == 10)) {
+ do
+ {
+ convert[sizeof(convert) - (++vallen)] = cvt[uvalue % 10];
+ uvalue = uvalue / 10;
+ } while (uvalue);
+ } else if (base == 16) {
+ do
+ {
+ convert[sizeof(convert) - (++vallen)] = cvt[uvalue % 16];
+ uvalue = uvalue / 16;
+ } while (uvalue);
+ } else if (base == 8) {
+ do
+ {
+ convert[sizeof(convert) - (++vallen)] = cvt[uvalue % 8];
+ uvalue = uvalue / 8;
+ } while (uvalue);
+ }
}
zeropad = Max(0, precision - vallen);
From | Date | Subject | |
---|---|---|---|
Next Message | Robert Haas | 2021-10-26 21:02:18 | Re: Next Steps with Hash Indexes |
Previous Message | Tomas Vondra | 2021-10-26 20:43:22 | Re: Feature request for adoptive indexes |