On 28/04/16 09:47, Maxim Kuvyrkov wrote:
>> On Apr 27, 2016, at 7:26 PM, Szabolcs Nagy <szabolcs.n...@arm.com> wrote:
>>
>> with -mfentry, by default the user only has to
>> implement the fentry call (linux wants nops there, but
>> e.g. glibc could use -pg -mfentry for profiling on
>> aarch64 and the target specific details are easier to
>> document for an -m option than for something general).
>
> I don't understand your point here, could you elaborate, please?
>
if we only provide -mfentry then
- the kernel can use it (they have tools to nop patch the binary),
- others who don't want to fiddle with nops, just have the call,
can also use it (e.g. user-space profiling cannot really use
something that needs binary patching in case the user prefers
-pg -mfentry over the current -pg behaviour).
- it's target specific, so the magic abi of the fentry call can
be documented by the target according to the specific instruction
sequence that is used. (with nop-padding there are psabi and
compiler optimization interactions that may be hard to document
in a generic way and letting the user figure it out may cause
problems later in compiler development.. but i'm just speculating
based on the powerpc toc handling and ipa-ra findings.)
>> the nop-padding is more general, but the size and
>> layout of nops and the call abi will be target
>> specific and the user will most likely need to modify
>> the binary (to get the right sequence) which needs
>> additional tooling. i don't know who might use it
>> other than linux (which already has tools to deal with
>> -mfentry).
>
> Right, but this tooling will require minimal (if any) changes
> to be adapted to nop-pad approach. If I remember correctly,
> recent versions of GCC and kernel for x86_64 generate NOPs,
> not the call sequence in the prologs when -mfentry is used.
i'm trying to find where this happens in the kernel, but
i only see scripts/recordmcount.{c,pl} which are based on
nop patching the fentry/mcount call sites.
without such call sites the tools have to be implemented
differently and the way the kernel records the call site
positions might not match the prolog-pad recording.
