On 2023-08-25, Linus Torvalds wrote:
On Fri, 25 Aug 2023 at 17:52, Nick Desaulniers <ndesaulniers@xxxxxxxxxx> wrote:
So 2 concerns where "I'll do it in inline asm" can pessimize codegen:
1. You alluded to this, but what happens when one of these functions
is called with a constant?
This is why our headers have a lot of __builtin_constant_p()'s in them..
In this particular case, see the x86 asm/bitops.h code:
#define ffs(x) (__builtin_constant_p(x) ? __builtin_ffs(x) :
variable_ffs(x))
For the curious (like me),
__builtin_ffs
https://gcc.gnu.org/onlinedocs/gcc/Other-Builtins.html#index-_005f_005fbuiltin_005fffs says
Returns the number of leading 0-bits in x, starting at the most significant bit position. If x is 0, the result is undefined.
The hangling of 0 seems the cause that __builtin_ffs codegen is not as
well as inline asm. Clang implemented the builtin in 2008 and took the
same constraint (penalty).
GCC compiles __builtin_ctzl(x) to xorl %eax, %eax; tzcntq %rdi, %rax
on most Intel processors (AMD -march= values are unaffected). The extra
xor is due to a false dependency issue
https://gcc.gnu.org/git/?p=gcc.git&a=commit;h=73543b2286027da1de561902440b53f775a03a86
Inline asm wins here as well since we know the argument 0 is undefined.
In May 2023, https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=cc6eb8b51f9568ae0caf46b80e2a0aff050030ce
"Disable avoid_false_dep_for_bmi for atom and icelake(and later) core processors."
removed the extra xor for icelake (and later) core processors.
but this is actually quite a common pattern, and it's often not about
something like __builtin_ffs() at all.
See all the other __builtin_constant_p()'s that we have in that same
file because we basically just use different code sequences for
constants.
And that file isn't even unusual. We use it quite a lot when we care
about code generation for some particular case.
2. by providing the definition of a symbol typically provided by libc
(and then not building with -ffreestanding) pessimizes libcall
optimization.
.. and this is partly why we often avoid libgcc things, and do certain
things by hand.
The classic rule is "Don't do 64-bit divisions using the C '/' operator".
So in the kernel you have to use do_div() and friends, because the
library versions are often disgusting and might not know that 64/32 is
much much cheaper and is what you want.
And quite often we simply use other names - but then we also do *not*
build with -freestanding, because -freestanding has at least
traditionally meant that the compiler won't optimize the simple and
obvious cases (typically things like "memcpy with a constant size").
So we mix and match and pick the best option.
The kernel really doesn't care about architecture portability, because
honestly, something like "ffs()" is entirely *trivial* to get right,
compared to the real differences between architectures (eg VM and IO
differences etc).
Linus
