On Thu, 17 Oct 2013 12:10:28 +0200
Radim Krčmář <rkrcmar@xxxxxxxxxx> wrote:
> We implemented the optimized branch selection in higher levels of api.
> That made static_keys very unintuitive, so this patch introduces another
> element to jump_table, carrying one bit that tells the underlying code
> which branch to optimize.
>
> It is now possible to select optimized branch for every jump_entry.
>
> Current side effect is 1/3 increase increase in space, we could:
> * use bitmasks and selectors on 2+ aligned code/struct.
> - aligning jump target is easy, but because it is not done by default
> and few bytes in .text are much worse that few kilos in .data,
> I chose not to
> - data is probably aligned by default on all current architectures,
> but programmer can force misalignment of static_key
> * optimize each architecture independently
> - I can't test everything and this patch shouldn't break anything, so
> others can contribute in the future
> * choose something worse, like packing or splitting
> * ignore
>
> proof: example & x86_64 disassembly: (F = ffffffff)
>
> struct static_key flexible_feature;
> noinline void jump_label_experiment(void) {
> if ( static_key_false(&flexible_feature))
> asm ("push 0xa1");
> else asm ("push 0xa0");
> if (!static_key_false(&flexible_feature))
> asm ("push 0xb0");
> else asm ("push 0xb1");
> if ( static_key_true(&flexible_feature))
> asm ("push 0xc1");
> else asm ("push 0xc0");
> if (!static_key_true(&flexible_feature))
> asm ("push 0xd0");
> else asm ("push 0xd1");
> }
>
> Disassembly of section .text: (push marked by "->")
>
> F81002000 <jump_label_experiment>:
> F81002000: e8 7b 29 75 00 callq F81754980 <__fentry__>
> F81002005: 55 push %rbp
> F81002006: 48 89 e5 mov %rsp,%rbp
> F81002009: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
> F8100200e: -> ff 34 25 a0 00 00 00 pushq 0xa0
> F81002015: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
> F8100201a: -> ff 34 25 b0 00 00 00 pushq 0xb0
> F81002021: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
> F81002026: -> ff 34 25 c1 00 00 00 pushq 0xc1
> F8100202d: 0f 1f 00 nopl (%rax)
> F81002030: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
> F81002035: -> ff 34 25 d1 00 00 00 pushq 0xd1
> F8100203c: 5d pop %rbp
> F8100203d: 0f 1f 00 nopl (%rax)
> F81002040: c3 retq
This looks exactly like what we want. I take it this is with your
patch. What was the result before the patch?
-- Steve
> F81002041: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
> F81002048: -> ff 34 25 d0 00 00 00 pushq 0xd0
> F8100204f: 5d pop %rbp
> F81002050: c3 retq
> F81002051: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
> F81002058: -> ff 34 25 c0 00 00 00 pushq 0xc0
> F8100205f: 90 nop
> F81002060: eb cb jmp F8100202d <[...]+0x2d>
> F81002062: 66 0f 1f 44 00 00 nopw 0x0(%rax,%rax,1)
> F81002068: -> ff 34 25 b1 00 00 00 pushq 0xb1
> F8100206f: 90 nop
> F81002070: eb af jmp F81002021 <[...]+0x21>
> F81002072: 66 0f 1f 44 00 00 nopw 0x0(%rax,%rax,1)
> F81002078: -> ff 34 25 a1 00 00 00 pushq 0xa1
> F8100207f: 90 nop
> F81002080: eb 93 jmp F81002015 <[...]+0x15>
> F81002082: 66 66 66 66 66 2e 0f [...]
> F81002089: 1f 84 00 00 00 00 00
>
> Contents of section .data: (relevant part of embedded __jump_table)
--
To unsubscribe from this list: send the line "unsubscribe linux-doc" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
