Benchmark for strcasecmp
to determine whether or not there is a better implementation.
Most implementations will convert each character to tolower
without checking to see if the characters are first the same. If the characters are the same then no lowercase conversion is needed.
This code uses Google Benchmark to perform tests against several 256 character strings of varying similarity.
glibc
int
__strcasecmp (const char *s1, const char *s2 LOCALE_PARAM)
{
const unsigned char *p1 = (const unsigned char *) s1;
const unsigned char *p2 = (const unsigned char *) s2;
int result;
if (p1 == p2)
return 0;
while ((result = TOLOWER (*p1) - TOLOWER (*p2 )) == 0)
if (*p1 == '\0')
break;
return result;
}
https://github.com/bminor/glibc/blob/b92a49359f33a461db080a33940d73f47c756126/string/strcasecmp.c
The check for p1 == p2
can add an additional benefit for user-mode applications that are compiled with string pooling enabled.
Apple
int
strcasecmp(s1, s2)
const char *s1, *s2;
{
const u_char
*us1 = (const u_char *)s1,
*us2 = (const u_char *)s2;
while (tolower(*us1) == tolower(*us2 ))
if (*us1 == '\0')
return (0);
return (tolower(*us1) - tolower(*--us2));
}
https://opensource.apple.com/source/Libc/Libc-583/string/FreeBSD/strcasecmp.c.auto.html
OpenSSL
int OPENSSL_strncasecmp(const char *s1, const char *s2, size_t n)
{
int t;
size_t i;
for (i = 0; i < n; i )
if ((t = ossl_tolower(*s1) - ossl_tolower(*s2 )) != 0)
return t;
else if (*s1 == '\0')
return 0;
return 0;
}
https://github.com/openssl/openssl/blob/79c8dcf3985a7b75eac8e53eb8652728af6c5d3d/crypto/o_str.c
cURL
int Curl_strcasecompare(const char *first, const char *second)
{
while(*first && *second) {
if(Curl_raw_toupper(*first) != Curl_raw_toupper(*second))
/* get out of the loop as soon as they don't match */
return 0;
first ;
second ;
}
/* If we're here either the strings are the same or the length is different.
We can just test if the "current" character is non-zero for one and zero
for the other. Note that the characters may not be exactly the same even
if they match, we only want to compare zero-ness. */
return !*first == !*second;
}
https://github.com/curl/curl/blob/master/lib/strcase.c
Linux Kernel
int strcasecmp(const char *s1, const char *s2)
{
int c1, c2;
do {
c1 = tolower(*s1 );
c2 = tolower(*s2 );
} while (c1 == c2 && c1 != 0);
return c1 - c2;
}
https://github.com/torvalds/linux/blob/master/lib/string.c
Interestingly enough the kernel's implementations for strcasecmp
and strncasecmp
will return different negative values for string1 = "", string2 = "A"
because the quick return in strncasecmp
when !c1 || !c2
. In this case strcasecmp = -'a'
and strncasecmp = '-A'
. But perhaps it is left up to the implementation what negative value to return.
Partially derived from Linux kernel implementations for strcasecmp
and strncasecmp
.
int strcasecmp_new(const char *s1, const char *s2)
{
unsigned char c1, c2;
do {
c1 = *s1 ;
c2 = *s2 ;
if (c1 != c2) {
c1 = tolower(c1);
c2 = tolower(c2);
if (c1 != c2)
return (int)c1 - (int)c2;
}
} while (c1 != 0);
return 0;
}
CMake instructions:
cmake -S . -B build -D CMAKE_BUILD_TYPE=Release
cmake --build build --config Release
The application allows most Google Benchmark arguments including --benchmark_repetitions=#
which can specify the number of repetitions to perform.
2022-10-25T08:48:13-07:00
Running 20 repetitions
Run on (12 X 3202.02 MHz CPU s)
CPU Caches:
L1 Data 32 KiB (x6)
L1 Instruction 32 KiB (x6)
L2 Unified 256 KiB (x6)
L3 Unified 12288 KiB (x1)
two strings 100% same case = ~48% improvement
strcasecmp_bench/100% same:0_mean 300 ns 164 ns 20
strcasecmp_bench/100% same:0_median 291 ns 229 ns 20
strcasecmp_bench/100% same:0_stddev 23.5 ns 120 ns 20
strcasecmp_bench/100% same:0_cv 7.84 % 73.54 % 20
strcasecmp_new_bench/100% same:0_mean 158 ns 89.8 ns 20
strcasecmp_new_bench/100% same:0_median 159 ns 102 ns 20
strcasecmp_new_bench/100% same:0_stddev 2.66 ns 53.5 ns 20
strcasecmp_new_bench/100% same:0_cv 1.69 % 59.58 % 20
two strings 83% same case = ~32% improvement
strcasecmp_bench/83% same:1_mean 304 ns 173 ns 20
strcasecmp_bench/83% same:1_median 304 ns 220 ns 20
strcasecmp_bench/83% same:1_stddev 6.58 ns 123 ns 20
strcasecmp_bench/83% same:1_cv 2.16 % 70.95 % 20
strcasecmp_new_bench/83% same:1_mean 204 ns 121 ns 20
strcasecmp_new_bench/83% same:1_median 203 ns 120 ns 20
strcasecmp_new_bench/83% same:1_stddev 3.48 ns 73.3 ns 20
strcasecmp_new_bench/83% same:1_cv 1.71 % 60.41 % 20
two strings 75% same case = ~36% improvement
strcasecmp_bench/75% same:2_mean 321 ns 183 ns 20
strcasecmp_bench/75% same:2_median 322 ns 237 ns 20
strcasecmp_bench/75% same:2_stddev 5.37 ns 129 ns 20
strcasecmp_bench/75% same:2_cv 1.67 % 70.76 % 20
strcasecmp_new_bench/75% same:2_mean 205 ns 127 ns 20
strcasecmp_new_bench/75% same:2_median 205 ns 173 ns 20
strcasecmp_new_bench/75% same:2_stddev 4.52 ns 81.0 ns 20
strcasecmp_new_bench/75% same:2_cv 2.21 % 63.68 % 20
two strings 66% same case = ~32% improvement
strcasecmp_bench/66% same:3_mean 351 ns 197 ns 20
strcasecmp_bench/66% same:3_median 348 ns 199 ns 20
strcasecmp_bench/66% same:3_stddev 22.8 ns 37.4 ns 20
strcasecmp_bench/66% same:3_cv 6.50 % 18.92 % 20
strcasecmp_new_bench/66% same:3_mean 236 ns 138 ns 20
strcasecmp_new_bench/66% same:3_median 232 ns 141 ns 20
strcasecmp_new_bench/66% same:3_stddev 10.9 ns 57.4 ns 20
strcasecmp_new_bench/66% same:3_cv 4.60 % 41.62 % 20
two strings 50% same case = ~32% improvement
strcasecmp_bench/50% same:4_mean 384 ns 217 ns 20
strcasecmp_bench/50% same:4_median 371 ns 220 ns 20
strcasecmp_bench/50% same:4_stddev 27.9 ns 21.2 ns 20
strcasecmp_bench/50% same:4_cv 7.26 % 9.77 % 20
strcasecmp_new_bench/50% same:4_mean 259 ns 147 ns 20
strcasecmp_new_bench/50% same:4_median 259 ns 175 ns 20
strcasecmp_new_bench/50% same:4_stddev 4.08 ns 102 ns 20
strcasecmp_new_bench/50% same:4_cv 1.57 % 69.35 % 20
two strings 33% same case = ~13% improvement
strcasecmp_bench/33% same:5_mean 303 ns 169 ns 20
strcasecmp_bench/33% same:5_median 299 ns 159 ns 20
strcasecmp_bench/33% same:5_stddev 15.7 ns 54.2 ns 20
strcasecmp_bench/33% same:5_cv 5.17 % 32.01 % 20
strcasecmp_new_bench/33% same:5_mean 263 ns 156 ns 20
strcasecmp_new_bench/33% same:5_median 264 ns 166 ns 20
strcasecmp_new_bench/33% same:5_stddev 14.7 ns 64.0 ns 20
strcasecmp_new_bench/33% same:5_cv 5.57 % 41.05 % 20
two strings 25% same case = ~5% improvement
strcasecmp_bench/25% same:6_mean 287 ns 163 ns 20
strcasecmp_bench/25% same:6_median 287 ns 161 ns 20
strcasecmp_bench/25% same:6_stddev 2.99 ns 59.9 ns 20
strcasecmp_bench/25% same:6_cv 1.04 % 36.73 % 20
strcasecmp_new_bench/25% same:6_mean 269 ns 159 ns 20
strcasecmp_new_bench/25% same:6_median 268 ns 158 ns 20
strcasecmp_new_bench/25% same:6_stddev 3.46 ns 8.69 ns 20
strcasecmp_new_bench/25% same:6_cv 1.29 % 5.47 % 20
two strings 15% same case = ~2% decrease
strcasecmp_bench/15% same:7_mean 292 ns 168 ns 20
strcasecmp_bench/15% same:7_median 291 ns 169 ns 20
strcasecmp_bench/15% same:7_stddev 2.76 ns 14.9 ns 20
strcasecmp_bench/15% same:7_cv 0.95 % 8.83 % 20
strcasecmp_new_bench/15% same:7_mean 299 ns 161 ns 20
strcasecmp_new_bench/15% same:7_median 300 ns 139 ns 20
strcasecmp_new_bench/15% same:7_stddev 6.81 ns 122 ns 20
strcasecmp_new_bench/15% same:7_cv 2.28 % 75.59 % 20
two strings 0% same case = ~11% decrease
strcasecmp_bench/0% same:8_mean 297 ns 155 ns 20
strcasecmp_bench/0% same:8_median 293 ns 193 ns 20
strcasecmp_bench/0% same:8_stddev 14.1 ns 118 ns 20
strcasecmp_bench/0% same:8_cv 4.75 % 76.46 % 20
strcasecmp_new_bench/0% same:8_mean 332 ns 209 ns 20
strcasecmp_new_bench/0% same:8_median 331 ns 242 ns 20
strcasecmp_new_bench/0% same:8_stddev 4.96 ns 111 ns 20
strcasecmp_new_bench/0% same:8_cv 1.50 % 52.89 % 20
For strings that are over 25% the same case our implementation provides up to 48% performance benefits over other implementations. However if the two strings are less than 25% there can be a similar up to 11% performance decrease.
We surmise that perhaps CPU caching does not play a significant role in string case comparison, but only branches and perdictions. tolower
has its own branch to determine if the character is uppercase or lowercase which results in a lookup to a hash map.
If we can reduce the total number of branches, by skipping an extra branch when characters match, we can speed up the function in corrolation to the similarity of the two strings.
There is another way that involves reducing the number of branches and that is by using a different character hash map.
As stated previously, some implementations of tolower
first check if the character is uppercase and if so convert it to lowercase.
This is because these tolower
implementations rely on the _ctype
hash map which is a 256 character map that can be used to lookup what each character's type is - digit, upper, lower, etc.
However, it is possible to use hash maps that are dedicated to the uppercase/lowercase conversion where the results of tolower(c)
and toupper(c)
are precalculated for each of the 256 characters.