Re: Why GLIBC requires compiler optimizations to build

[Carlos O'Donell <carlos@redhat.com>]

On 06/20/2018 11:22 AM, Gary Benson wrote:
> On 20 June 2018 at 13:51, Carlos O'Donell <carlos@redhat.com> wrote:
>> On 06/20/2018 06:35 AM, Gary Benson wrote:
>>> On 11 June 2018 at 22:48, H.J. Lu <hjl.tools@gmail.com> wrote:
>>>> On Mon, Jun 11, 2018 at 2:38 PM, Adhemerval Zanella <adhemerval.zanella@linaro.org> wrote:
>>>>> It has been some time since I tried to build glibc without optimization
>>>>> as an exercise and I can't really recall what exactly has prevented me
>>>>> to accomplish it.  So the question is if it is make sense to add such
>>>>> requirement and if it is the case whether it would be a useful options
>>>>> and what prevents us to do so?
>>>>
>>>> The main issue is to bootstrap ld.so to prevent dynamic relocations
>>>> before it is ready.
>>>
>>> Is it only really elf/rtld.c that specifically requires optimization then?
>>
>> No.
>>
>> The dynamic loader has many functions it needs to execute the early startup.
>>
>> So there are a lot of dependencies on other *.c files here and there, which
>> must be callable directly and not through the PLT.
> 
> Ok. Thanks for explaining this.
> 
>> The perfect solution is a refactored build system that can distinguish what
>> is going into rtld that needs optimizations and what is not. Then apply
>> optimizations only to rtld.
>>
>> Even then, for debugging purposes, I have gotten away with using gcc's
>> function attributes to mark some function as -O0 to debug them more easily,
>> otherwise it's very hard to debug and _dl_debug_printf() is your highest
>> value tool.
> 
> I did not know about __attribute__ ((optimize ())), thank you!

I couldn't live without it when debugging complex code bases where using
-O0 makes the bug go away or breaks rtld in this case... so you can use
the function attribute to alter just a function at a time.

Cheers,
Carlos.