On 4/19/20 10:29 PM, Ranier Vilela wrote:
Em dom., 19 de abr. de 2020 às 16:33, Tomas Vondra
<tomas.von...@2ndquadrant.com <mailto:tomas.von...@2ndquadrant.com>>
escreveu:
On Sun, Apr 19, 2020 at 11:24:38AM -0300, Ranier Vilela wrote:
>Hi,
>strlen it is one of the low fruits that can be harvested.
>What is your opinion?
>
That assumes this actually affects/improves performance, without any
measurements proving that. Considering large number of the places you
modified are related to DDL (CreateComment, ChooseIndexColumnNames, ...)
or stuff that runs only once or infrequently (like the changes in
PostmasterMain or libpqrcv_get_senderinfo). Likewise, it seems entirely
pointless to worry about strlen() overhead e.g. in fsync_parent_path
which is probably dominated by I/O.
With code as interconnected as postgres, it is difficult to say that a
function, which calls strlen, repeatedly, would not have any gain.
Regarding the functions, I was just being consistent, trying to remove
all occurrences, even where, there is very little gain.
At least gcc 9.3 optimizes "strlen(s) == 0" to "s[0] == '\0'", even at
low optimization levels. I tried it out with https://godbolt.org/.
Maybe some of the others cases are performance improvements, I have not
checked your patch in details, but strlen() == 0 is easily handled by
the compiler.
C code:
int f1(char *str) {
return strlen(str) == 0;
}
int f2(char *str) {
return str[0] == '\0';
}
Assembly generated with default flags:
f1:
pushq %rbp
movq %rsp, %rbp
movq %rdi, -8(%rbp)
movq -8(%rbp), %rax
movzbl (%rax), %eax
testb %al, %al
sete %al
movzbl %al, %eax
popq %rbp
ret
f2:
pushq %rbp
movq %rsp, %rbp
movq %rdi, -8(%rbp)
movq -8(%rbp), %rax
movzbl (%rax), %eax
testb %al, %al
sete %al
movzbl %al, %eax
popq %rbp
ret
Assembly generated with -O2.
f1:
xorl %eax, %eax
cmpb $0, (%rdi)
sete %al
ret
f2:
xorl %eax, %eax
cmpb $0, (%rdi)
sete %al
ret
Andreas
