[RFC] Removing DefinedSynthetic

[Rafael Avila de Espindola via llvm-commits]

On February 22, 2017 1:20:44 PM EST, Rui Ueyama <ruiu at google.com> wrote:
>On Wed, Feb 22, 2017 at 9:43 AM, Rafael Avila de Espindola <
>rafael.espindola at gmail.com> wrote:
>
>> I am looking at https://bugs.llvm.org/show_bug.cgi?id=32031 and there
>> are a couple of ways to interpret it.
>>
>> One way is that we are just computing symbol values too early. When
>we
>> do it some output sections still don't have their address set and
>some
>> input sections still don't know what output section they will be in.
>>
>> The other way to look at it is just that our representation of symbol
>> values in the linker script is wrong. I started experimenting with
>using
>> a ExprValue struct that is similar in idea to a MCValue. It is what
>Expr
>> returns and it contains a PointerUnion of an input or output section
>and
>> the offset. This allows us to avoid having to check the OutSec field
>> before it is set and solves the problem.
>>
>> But that got me thinking, can we remove DefinedSynthetic completely?
>> With the addition of a few dummy input sections and potentially a
>couple
>> more conversions to SyntheticSection I think we can.
>>
>
>Regular Defined symbols are defined relative to input sections because
>they
>are both read from input files. Regular symbols get their addresses
>when
>the associated input sections get addresses.
>
>DefinedSynthetic is being used to define linker-synthesized symbols
>that
>are relative to output sections. _end or _edata are a few examples.
>
>So, if you remove DefinedSynthetic, how do you represent _end, for
>example?

By attaching it to a empty synthetic section at the end of the corresponding output section.

Cheers,
Rafael

-- 
Sent from my Android device with K-9 Mail. Please excuse my brevity.