On Tue, May 05, 2020 at 11:13:53AM -0700, Nick Desaulniers wrote:
> On Tue, May 5, 2020 at 2:36 AM Peter Zijlstra <pet...@infradead.org> wrote:
> >
> >
> > HJ, Nick,
> >
> > Any chance any of you can see a way to make your respective compilers
> > not emit utter junk for this?
> >
> > On Mon, May 04, 2020 at 10:14:45PM +0200, Peter Zijlstra wrote:
> >
> > > https://godbolt.org/z/SDRG2q
>
> Woah, a godbolt link! Now we're speaking the same language. What were
> you expecting?
Given the output for x86-64 clang (trunk)
bar: # @bar
movl %edi, .L_x$local(%rip)
retq
ponies: # @ponies
movq .Lfoo$local(%rip), %rax
testq %rax, %rax
movl $__static_call_nop, %ecx
cmovneq %rax, %rcx
jmpq *%rcx # TAILCALL
__static_call_nop: # @__static_call_nop
retq
_x:
.L_x$local:
.long 0 # 0x0
foo:
.Lfoo$local:
.zero 8
I was hoping for:
bar: # @bar
movl %edi, .L_x$local(%rip)
retq
ponies: # @ponies
movq .Lfoo$local(%rip), %rax
testq %rax, %rax
jz 1f
jmpq *%rcx # TAILCALL
1:
retq
That avoids the indirect call (possible retpoline) and does an immediate
return.
So it does 2 things different:
- it realizes the NULL case is a constant and uses an
immediate call and avoids the indirect call/jmp.
- it realizes __static_call_nop() is a no-op and avoids the call
entirely and does an immediate return.
> Us to remove the conditional check that a volatile read
> wasn't NULL?
No, obviously the load is required, and the READ_ONCE() is so that the
compiler will not emit 2 different loads (just for giggles).
That is:
tmp1 = name.func;
if (!tmp) {
tmp2 = name.func;
tmp2(args);
}
is a valid translation of:
if (!name.func)
name.func(args)
and allows for a NULL dereference (as noted by Rasmus).
What I did do want, per the above, is to avoid the indirect (tail) call.
Because indirect jmp/call are evil and expensive.
> I am simultaneously impressed
> and disgusted by this btw, cool stuff.
Yes, it's nasty, esp the casting of a function pointer like that is
gruesome.
