Re: RFC: ld: Add --text-section-ordering-file (version 2)

[Alan Modra <amodra@gmail.com>]

On Thu, Apr 25, 2024 at 02:01:01PM +0100, Nick Clifton wrote:
> Hi Guys,
> 
>   Attached is a patch to add a --section-ordering-file option to the BFD
>   linker.  It is based upon H.J.'s original patch, but extended so that
>   it will work with multiple output sections.
> 
>   There are a couple of points that I feel I should highlight:
> 
>   * The option only works in conjunction with a modified linker script.
>   
>     In particular the script must have: "INCLUDE section-ordering-file"
>     statements in it, wherever it wants section ordering to be allowed.
>     This statement can appear multiple times in a linker script,
>     although it would not make sense to have it appear more than once in
>     any given output section.  Here is an example:
>     
>       SECTIONS
>       {
>         .text : {
>           INCLUDE section-ordering-file
>           *(.text)
>         }
>         .data : {
>           INCLUDE section-ordering-file
>           *(.data)
>         }
>       }
>     
>   * H.J's original version allowed for linker script like
>     "filename(section)" syntax to be used to name sections, eg:
>     "*(.text.*)", as well as a simpler "section name regexp", eg
>     ".text.*", to be used.  This version only supports the latter
>     format.
> 
>     In addition H.J.'s syntax allowed for abbreviated section names to
>     match.  So ".t*t" would match any section starting with ".t" and
>     ending with "t" and would put it into the .text section.  In this
>     version however the output section is selected based upon matching
>     the fixed part at the start of the pattern with the output section.
>     So ".t*t" would only work if the output section was called ".t".
> 
>     To help compensate for this, and to allow arbitrary input sections
>     to be mapped to specific output sections, the output section name
>     can be provided as if it were a filename.  So .foo(.bar) would map
>     all sections called .bar to the output section .foo, but only if the
>     linker script has an output section called .foo, and only if that
>     output section declaration includes a INCLUDE section-ordering-file
>     statement.
>     
>     Perhaps an example will make things clearer.  If the above linker
>     script is used and the section ordering file contains:
> 
>       # A comment - this will be ignored.
>       .text.hot .text.cold .text.warm
>       .data.big
>       .data(.bar)
>       .text.foo*
>       .ignore(.me)
> 
>     This is roughly equivalent to a linker script that looks like this:
> 
>       SECTIONS
>       {
>         .text : {
>           *(.text.hot)
>           *(.text.cold)
>           *(.text.warm)
>           *(.text.foo*)
>           *(.text)
>         }
>         .data : {
>           *(.data.big)
>           *(.bar)
>           *(.data)
>         }
>       }
> 
>     Note - the linker will not warn about entries in the section
>     ordering file that do not match an output section name.  So in the
>     above example the ".ignore(.me)" entry will not generate a warning
>     about it not being used.
> 
>   Thoughts, comments, etc ?

This seems really clunky to me.  How about a script that augments
the default script, but looks very similar to other scripts, ie. with
a SECTIONS clause?  To take your example, use a script like:

  SECTIONS
  {
    .text : {
      *(.text.hot)
      *(.text.cold)
      *(.text.warm)
      *(.text.foo*)
    }
    .data : {
      *(.data.big)
      *(.bar)
    }
  }

Then arrange to insert the contents of each output section statement
in this script at the start of corresponding output section statement
in the default script.

-- 
Alan Modra
Australia Development Lab, IBM