[PATCH] D27415: [ELF] - Replace MergeOutputSection with synthetic input section MergeSection.

[Rui Ueyama via llvm-commits]

On Fri, Jan 27, 2017 at 3:02 AM, George Rimar via Phabricator <
reviews at reviews.llvm.org> wrote:

> grimar added a comment.
>
> In https://reviews.llvm.org/D27415#657878, @ruiu wrote:
>
> > I probably do not understand what you are trying to solve.
> >
> > Currently, LLD merges two mergeable input sections if they have the same
> name, types and section flags. But, you are saying that that is not always
> correct, right?
> >
> > Can you briefly describe the exact semantics this patch is trying to
> implement, and then why you think that is the better behavior?
>
>
> I am trying to solve issue I am demonstrating at
> https://reviews.llvm.org/D29217 page.
> It has:
>
>   .section .aaa.1,"a"
>   .byte 11
>   .section .aaa.2,"a"
>   .byte 22
>
>   .section .bbb.1,"aMS", at progbits,1
>   .asciz "foo"
>   .section .bbb.2,"aMS", at progbits,1
>   .asciz "foo"
>
>   .section .ccc.1,"a"
>   .byte 33
>   .section .ccc.2,"a"
>   .byte 44
>
> It has also symbols assignments to A and B that should mark the start/end
> of .ccc :
>
>   .rodata : { *(.aaa.*) *(.bbb.*) A = .; *(.ccc.*) B = .; }
>
> **1. LLD currently (clean head revision) do: **
>
> It creates 2 output .rodata sections:
> .rodata (.aaa.1 .aaa.2 .ccc.1 .ccc.2)
> .rodata (.bbb.1 .bbb.2)
>

OK, so my understanding is this.

Assume you have four mergeable string input sections .foo.1, .foo.2, .bar.1
and .bar.2, where .foo.{1,2} are writable but .bar.{1,2} are not. Also
assume you have this linker script.

  .rodata : { *(.foo.*) *(.bar.*); }

LLD currently creates two .rodata sections for those inputs. One for
.foo.{1,2} and the other for .bar.{1,2}. This is because mergeable sections
are merged based on section attributes.

Is this correct? What do you think is a correct behavior?


> And because of that it assigns values of A and B wrong currently. I think
> the cleanest way to fix is implement synthetic merge section instead of
> MergeOutputSection.
> (this is what this patch was about, right ?)
> That way output should be single .rodata section which consist of [ .aaa.1
> .aaa.2, synthetic section holding <.bbb.1 and .bbb.2>, .ccc.1, .ccc.2]. And
> symbols will be assigned correctly then.
> Output will probably be equal to bfd/gold, btw.
>
> **2. Latest diff of this patch do:**
>
> It creates synthetic sections for mergeable sections early now. Before
> passing them to script. After that we have next input sections available to
> work with on linerscript side:
>
> [.aaa.1 .aaa.2 **SynteticMergeSection**(.bbb.1)
> **SynteticMergeSection**(.bbb.2) .ccc.1 .ccc.2]
>
> Since synthetic merge sections are already created, script just places all
> above in a single .rodata section. We loose string merging optimization
> here, because have 2 pre-created SynteticMergeSections.
> There is no way to merge 2 synthetic mergable sections together on this
> step (I do not think we want to implement this too).
> And we can not create single .bbb section for holding .bbb.1 and .bbb.2
> early before proccessing linkerscript commands, because we do not know it
> will want to put them together.
>
> **3. Patch I want to do basing on what diff2 of this patch already did:**
>
> I want to create synthetic merge sections later. For linkerscript step it
> should be done after script prepares input sections list to create output
> sections. That was done in diff2.
> That way we have inputs:
> [.aaa.1 .aaa.2 **MergeInputSection(.bbb.1)** **MergeInputSection(.bbb.2)**
> .ccc.1 .ccc.2]
>
> Script will create single SynteticMergeSection for mergable input sections
> ind convert inputs to:
> [.aaa.1 .aaa.2 **SynteticMergeSection(.bbb.1, .bbb.2)** .ccc.1 .ccc.2]
>
> And we will end up with single .rodata with all optimizations working.
>
>
> https://reviews.llvm.org/D27415
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170127/ce6b8100/attachment.html>