Re: [PATCH] Reduce ARM linking time

[Jie Zhang <jie at codesourcery dot com>]

On 03/20/2010 12:05 AM, Jie Zhang wrote:
> We encountered an problem with ARM ld that the linking time of a large
> application was very long, about 1.5 hour.

Forgot to mention that with this patch, that linking time of that 
particular application is reduced to about 30 seconds.

> The root cause of this issue is ARM uses a double link list to track
> sections with ARM specific section data allocated. Below is the related
> comment:
>
> /* Unfortunately we need to keep a list of sections for which
> an _arm_elf_section_data structure has been allocated. This
> is because it is possible for functions like elf32_arm_write_section
> to be called on a section which has had an elf_data_structure
> allocated for it (and so the used_by_bfd field is valid) but
> for which the ARM extended version of this structure - the
> _arm_elf_section_data structure - has not been allocated. */
>
> When ARM specific section data for a particular section is needed, the
> *whole* list is traversed to verify if that section is in the list, i.e.
> it has ARM specific section data allocated. For large applications with
> tens of thousands of sections, the traverse time becomes extremely long.
> This was used to fix:
>
> http://sourceware.org/bugzilla/show_bug.cgi?id=1147
>
> This patch uses is_arm_elf, as Paul and Dan suggested, to decide if this
> section has an _arm_elf_section_data. We also need to set mapcount to -1
> in elf32_arm_write_section to distinguish the two cases: no mapping
> symbols (mapcount == 0) vs mapping symbols have been consumed (mapcount
> == 1). Otherwise, the bug mentioned in
>
> http://sourceware.org/ml/binutils/2010-02/msg00018.html
>
> would come out and bite us again.
>
> Ld regression testing with arm-none-linux-gnueabi is OK.
>
> OK to commit?


--
Jie Zhang
CodeSourcery
(650) 331-3385 x735