https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78809
--- Comment #23 from Qing Zhao <qing.zhao at oracle dot com> ---
I have an implementation for the part C of this task in my private space:
part C: for strcmp (s1, s2), strncmp (s1, s2, n):
if the result is NOT used to do simple equality test against zero, one of
"s1" or "s2" is a small constant string, n is a constant, and the Min value of
the length of the constant string and "n" is smaller than a predefined
threshold T,
inline the call by a byte-to-byte comparision sequence to avoid calling
overhead.
with this implementation, I was able to measure the performance impact of the
inlining transformation on different value of "n", n is the length of the
string need to be compared. In order to decide the following two concerns:
A. what's the default value of n.
B. on a platform that support string compare hardware insns (for exmaple,
X86), which should be done first for a call to strcmp/strncmp, the inline or
the hardware insns?
On both aarch64 and X86, I tried the following small testing case for the
performance experiments:
qinzhao@gcc116:~/Bugs/78809/const_cmp$ cat t_p.c
#include <string.h>
char array[]= "fishiiiiiiiiiiiiiiiiiiiiiiiiiiiii";
#define NUM 1000000000
int __attribute__ ((noinline))
cmp2 (const char *p)
{
int result = 0;
int i;
for (i=0; i< NUM; i++) {
result |= strcmp (p, "fishiiiii");
}
return result;
}
int result = 0;
int main()
{
for (int i = 0; i < 25; i++)
result += cmp2 (array);
return 0;
}
and the option I used was: -O -fno-tree-loop-im and the corresponding option
to enable or disable the added inlining, the following is the performance
result:
aarch64 strcmp
n= 3 4 5 6 10 20
inline 31 41 62 72 114 242
no-inline 229 229 229 229 272 333
aarch64 strncmp
n= 3 4 5 6 10 20
inline 41 62 62 76 125 250
no-inline 291 291 291 291 364 427
X86 strcmp
n= 4 5 6 10 20
inline 21 25 31 42 163
no-inline 445 461 488 529 672
X86 strncmp
n= 4 5 6 10 20
inline 21 25 28 43 77
no-inline 412 435 442 495 638
From the above, we can see:
