Re: [PATCH v3 07/29] x86: bpf_jit, use ENTRY+ENDPROC

[David Miller]

From: Jiri Slaby <jslaby@xxxxxxx>
Date: Mon, 24 Apr 2017 16:52:43 +0200

> On 04/24/2017, 04:41 PM, David Miller wrote:
>>> It cannot stay as-is simply because we want to know where the functions
>>> end to inject debuginfo properly. The code above does not warrant for
>>> any exception.
>> 
>> I totally and completely disagree.
> 
> You can disagree as you wish but there is really nothing special on the
> bpf code with respect to annotations.
> 
>>> Executing a nop takes a little and having externally-callable functions
>>> aligned can actually help performance (no, I haven't measured nor tested
>>> the code). But sure, the tool is generic, so I can introduce a local
>>> macros to avoid alignments in the functions:
>> 
>> Not for this case, it's a bunch of entry points all packed together
>> intentionally so that SKB accesses of different access sizes (which is
>> almost always the case) from BPF programs use the smallest amount of
>> I-cache as possible.
> 
> And for that reason I suggested the special macros for the code (see the
> macros in the e-mail you replied to again). So what problem do you
> actually have with the suggested solution?

If you align the entry points, then the code sequence as a whole is
are no longer densely packed.

Or do I misunderstand how your macros work?