Re: static_branch/jump_label vs branch merging

[Ard Biesheuvel]

On Thu, 8 Apr 2021 at 18:53, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> Hi!
>
> Given code like:
>
> DEFINE_STATIC_KEY_FALSE(sched_schedstats);
>
> #define   schedstat_enabled()           static_branch_unlikely(&sched_schedstats)
> #define   schedstat_set(var, val)       do { if (schedstat_enabled()) { var = (val); } } while (0)
> #define __schedstat_set(var, val)       do { var = (val); } while (0)
>
> void foo(void)
> {
>         struct task_struct *p = current;
>
>         schedstat_set(p->se.statistics.wait_start,  0);
>         schedstat_set(p->se.statistics.sleep_start, 0);
>         schedstat_set(p->se.statistics.block_start, 0);
> }
>
> Where the static_branch_unlikely() ends up being:
>
> static __always_inline bool arch_static_branch(struct static_key * const key, const bool branch)
> {
>         asm_volatile_goto("1:"
>                 ".byte " __stringify(BYTES_NOP5) "\n\t"
>                 ".pushsection __jump_table,  \"aw\" \n\t"
>                 _ASM_ALIGN "\n\t"
>                 ".long 1b - ., %l[l_yes] - . \n\t"
>                 _ASM_PTR "%c0 + %c1 - .\n\t"
>                 ".popsection \n\t"
>                 : :  "i" (key), "i" (branch) : : l_yes);
>
>         return false;
> l_yes:
>         return true;
> }
>
> The compiler gives us code like:
>
> 000000000000a290 <foo>:
>     a290:       65 48 8b 04 25 00 00 00 00      mov    %gs:0x0,%rax     a295: R_X86_64_32S      current_task
>     a299:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)
>     a29e:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)
>     a2a3:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)
>     a2a8:       c3                      retq
>     a2a9:       48 c7 80 28 01 00 00 00 00 00 00        movq   $0x0,0x128(%rax)
>     a2b4:       eb e8                   jmp    a29e <foo+0xe>
>     a2b6:       48 c7 80 58 01 00 00 00 00 00 00        movq   $0x0,0x158(%rax)
>     a2c1:       eb e0                   jmp    a2a3 <foo+0x13>
>     a2c3:       48 c7 80 70 01 00 00 00 00 00 00        movq   $0x0,0x170(%rax)
>     a2ce:       c3                      retq
>
>
> Now, in this case I can easily rewrite foo like:
>
> void foo2(void)
> {
>         struct task_struct *p = current;
>
>         if (schedstat_enabled()) {
>                 __schedstat_set(p->se.statistics.wait_start,  0);
>                 __schedstat_set(p->se.statistics.sleep_start, 0);
>                 __schedstat_set(p->se.statistics.block_start, 0);
>         }
> }
>
> Which gives the far more reasonable:
>
> 000000000000a2d0 <foo2>:
>     a2d0:       65 48 8b 04 25 00 00 00 00      mov    %gs:0x0,%rax     a2d5: R_X86_64_32S      current_task
>     a2d9:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)
>     a2de:       c3                      retq
>     a2df:       48 c7 80 28 01 00 00 00 00 00 00        movq   $0x0,0x128(%rax)
>     a2ea:       48 c7 80 58 01 00 00 00 00 00 00        movq   $0x0,0x158(%rax)
>     a2f5:       48 c7 80 70 01 00 00 00 00 00 00        movq   $0x0,0x170(%rax)
>     a300:       c3                      retq
>
> But I've found a few sites where this isn't so trivial.
>
> Is there *any* way in which we can have the compiler recognise that the
> asm_goto only depends on its arguments and have it merge the branches
> itself?
>
> I do realize that asm-goto being volatile this is a fairly huge ask, but
> I figured I should at least raise the issue, if only to raise awareness.
>

Wouldn't that require the compiler to interpret the contents of the asm() block?