static_branch/jump_label vs branch merging

[Peter Zijlstra <peterz@infradead.org>]

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.