On 4/26/19 10:43 AM, Alexander Monakov wrote:
> On Fri, 26 Apr 2019, Martin Liška wrote:
>
>> I'm suggesting to adjust that to:
>> - -Os will keep using rep-scasb as -Os means optimize for size
>> no matter what speed impact is
>
> I'm not sure it's a good choice, the inline sequence is
>
> xorl %eax, %eax
> orq $-1, %rcx
> repnz scasb
> notq %rcx
> decq %rcx
>
> compared to simply
>
> call strlen
>
> it's not even shorter.
Then I'm very open to adjust that even for -Os to 'call strlen'.
>
>> - otherwise use call to strlen
>> - when -minline-all-stringops is enabled and -O2+ is used, then
>> expand to 4B loop for all possible alignments
>
> But it's not OK to use misaligned loads, because overreading past
> end of string may cross a page boundary and cause a segfault.
Yes, the code is handling that correctly by doing a prologue, see bellow:
> What does
> your patch generate for the testcase at -O2 -minline-all-stringops?
$ cat strlen-O2.c
unsigned foo_1B (char *ptr)
{
return __builtin_strlen (ptr);
}
unsigned foo_4B (char *ptr)
{
return __builtin_strlen (__builtin_assume_aligned (ptr, 4));
}
$ gcc strlen-O2.c -O2 -S -o/dev/stdout -minline-all-stringops
...
foo_1B:
.LFB0:
.cfi_startproc
movq %rdi, %rdx
movq %rdi, %rax
andl $3, %edx
je .L10
cmpq $2, %rdx
je .L5
ja .L4
cmpb $0, (%rdi)
je .L3
leaq 1(%rdi), %rax
.L5:
cmpb $0, (%rax)
je .L3
addq $1, %rax
.L4:
cmpb $0, (%rax)
je .L3
addq $1, %rax
.p2align 4,,10
.p2align 3
.L10:
movl (%rax), %ecx
addq $4, %rax
leal -16843009(%rcx), %edx
notl %ecx
andl %ecx, %edx
andl $-2139062144, %edx
je .L10
movl %edx, %ecx
shrl $16, %ecx
testl $32896, %edx
cmove %ecx, %edx
leaq 2(%rax), %rcx
cmove %rcx, %rax
movl %edx, %esi
addb %dl, %sil
sbbq $3, %rax
.L3:
subq %rdi, %rax
ret
...
foo_4B:
.LFB1:
.cfi_startproc
movq %rdi, %rax
.L16:
movl (%rax), %ecx
addq $4, %rax
leal -16843009(%rcx), %edx
notl %ecx
andl %ecx, %edx
andl $-2139062144, %edx
je .L16
movl %edx, %ecx
shrl $16, %ecx
testl $32896, %edx
cmove %ecx, %edx
leaq 2(%rax), %rcx
cmove %rcx, %rax
movl %edx, %esi
addb %dl, %sil
sbbq $3, %rax
subq %rdi, %rax
ret
Martin
>
> Alexander
>
