[PATCH v2] elf: Add optimization barrier for __ehdr_start and _end

[Florian Weimer]

* Jakub Jelinek:

> On Mon, Jun 16, 2025 at 09:42:54AM +0300, Alexander Monakov wrote:
>> 
>> 
>> On Mon, 16 Jun 2025, Jakub Jelinek wrote:
>> 
>> > On Mon, Jun 16, 2025 at 09:23:49AM +0300, Alexander Monakov wrote:
>> > > But you are applying it after the assignment into _dl_rtld_map, which is too
>> > > late.  As written, it doesn't prevent vectorization.  I think Jakub's suggestion
>> > > was to use an asm to implement a non-transparent copy of the value (before use).
>> > 
>> > That is indeed what I meant initially, but I think "+g" on the memory is
>> > actually better, it does prevent vectorization of the stores because one
>> > address has to be stored before the other one, but in a way which will allow
>> > on some architectures e.g. direct store of the __ehdr_start address to the
>> > memory.
>> 
>> The patch is
>> 
>>    _dl_rtld_map.l_map_start = (ElfW(Addr)) &__ehdr_start;
>> +  asm ("" : "+g" (_dl_rtld_map.l_map_start));
>>    _dl_rtld_map.l_map_end = (ElfW(Addr)) _end;
>> +  asm ("" : "+g" (_dl_rtld_map.l_map_end));
>> 
>> Computing the address of _end is independent of storing into l_map_start, so
>> this is functionally identical to
>> 
>>    _dl_rtld_map.l_map_start = (ElfW(Addr)) &__ehdr_start;
>>    _dl_rtld_map.l_map_end = (ElfW(Addr)) _end;
>> +  asm ("" : "+g" (_dl_rtld_map.l_map_start));
>> +  asm ("" : "+g" (_dl_rtld_map.l_map_end));
>> 
>> and hence vectorization of stores is not prevented.
>
> So just make it
>    _dl_rtld_map.l_map_start = (ElfW(Addr)) &__ehdr_start;
> +  asm ("" : "=m" (_dl_rtld_map) : "m" (_dl_rtld_map));
>    _dl_rtld_map.l_map_end = (ElfW(Addr)) _end;
> +  asm ("" : "=m" (_dl_rtld_map) : "m" (_dl_rtld_map));

This still does not prevent loading &__ehdr_start and &_end into a
vector register if there is sufficient register pressure, I think.

Thanks,
Florian