https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71016
--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
So, trying:
long int
foo (long int i)
{
return i == 0 ? 17 : __builtin_clzl (i);
}
long int
bar (long int i)
{
return i == 0 ? 17 : __builtin_popcountl (i);
}
long int
baz (long int i)
{
return i == 0 ? 17 : __builtin_ffsl (i);
}
at -O2 to see effect on various builtins. On x86_64-linux I see:
- bsrq %rdi, %rdx
- movl $17, %eax
- xorq $63, %rdx
+ bsrq %rdi, %rax
+ movl $17, %edx
+ xorq $63, %rax
testq %rdi, %rdi
- cmovne %edx, %eax
cltq
+ cmove %rdx, %rax
in foo (so not better not worse; nonzero_bits isn't able to see through the (63
- CLZ (x)) ^ 63 stuff; but the patched compiler has potential to do better, if
we during expansion on the SUBREG set promoted flag to SRP_SIGNED_AND_UNSIGNED,
perhaps we could get rid of the unneeded cltq), in bar:
movl $17, %eax
- cltq
ret
(better; something in the pro_and_epilogue pass creates set to constant
followed by sign extension and there is no combiner afterwards to fix stuff
up),
and in baz:
- xorl %edx, %edx
- movl $17, %eax
- rep; bsfq %rdi, %rdx
- addq $1, %rdx
+ xorl %eax, %eax
+ movl $17, %edx
+ rep; bsfq %rdi, %rax
+ addq $1, %rax
testq %rdi, %rdi
- cmovne %edx, %eax
cltq
+ cmove %rdx, %rax
(different, but not better or worse).
On aarch64-linux, I see:
clz x2, x0
cmp x0, 0
- mov w1, 17
- csel w0, w1, w2, eq
- sxtw x0, w0
+ mov x1, 17
+ csel x0, x2, x1, ne
in foo (better),
- mov w0, 17
- sxtw x0, w0
+ mov x0, 17
in bar (better) and
- rbit x1, x0
- cmp x0, 0
- clz x1, x1
- mov w2, 17
- csinc w0, w2, w1, eq
+ cbz x0, .L14
+ rbit x0, x0
+ clz x0, x0
+ add x0, x0, 1
sxtw x0, w0
ret
+ .align 3
+.L14:
+ mov x0, 17
+ ret
in baz (worse). Of course, only foo (with some more sensical constant instead
of 17) is real-world stuff.
