[PATCH] ld: Recalculate section addresses before relaxing when need_layout is set [PR34979]
Kito Cheng
kito.cheng@gmail.com
Sun Nov 8 15:36:15 GMT 2026
Could you add test case to the patch?
Zhuo Zhi <xeonacid@archlinuxcn.org>於 2026年11月8日週日,21:34寫道:
> Before entering lang_relax_sections, elf_ldlang_before_allocation may
> set need_layout = true when either:
> - bfd_elf_discard_info shrinks or removes .eh_frame entries, or
> - bfd_elf_map_sections_to_segments updates elf_program_header_size
> (causing SIZEOF_HEADERS to change).
>
> However, lang_relax_sections previously only checked need_layout after
> all relaxation passes finished, or inside the relaxation loop after
> individual passes. When relaxation starts with need_layout == true,
> lang_size_sections_1 walks the linker script top-down and updates each
> output section's VMA right before invoking bfd_relax_section on its
> input sections.
>
> As a result, when relaxing an earlier output section (such as .text)
> that references a symbol in a later output section, the earlier output
> section already has its new VMA (reflecting the updated SIZEOF_HEADERS
> or .eh_frame size), whereas the later output section still holds its
> stale VMA from before elf_ldlang_before_allocation.
>
> On RISC-V, this causes _bfd_riscv_relax_call in pass 0 to underestimate
> the forward distance from .text to a later output section (e.g. when
> SIZEOF_HEADERS grows by 2 * sizeof(Elf64_Phdr) = 112 bytes after adding
> PT_TLS / PT_GNU_RELRO segments) and relax an auipc+jalr pair into jal.
> Once the later output section's VMA is updated, the actual distance
> exceeds the signed 21-bit range of R_RISCV_JAL, resulting in:
>
> relocation truncated to fit: R_RISCV_JAL against symbol `...'
>
> Recalculate section addresses prior to the relaxation loop when
> need_layout is set so that all output sections start relaxation with
> consistent VMAs.
>
> PR ld/34979
> * ldlang.c (lang_relax_sections): If need_layout is true,
> recalculate section addresses before running relaxation passes.
>
> Signed-off-by: Zhuo Zhi <xeonacid@archlinuxcn.org>
> ---
> ld/ldlang.c | 14 ++++++++++++++
> 1 file changed, 14 insertions(+)
>
> diff --git a/ld/ldlang.c b/ld/ldlang.c
> index b04ba113aa..757f04a79a 100644
> --- a/ld/ldlang.c
> +++ b/ld/ldlang.c
> @@ -8249,6 +8249,20 @@ lang_relax_sections (bool need_layout)
> /* NB: Also enable relaxation to layout sections for DT_RELR. */
> if (RELAXATION_ENABLED || link_info.enable_dt_relr)
> {
> + /* If section layout has changed since sections were last sized
> + (for example, when bfd_elf_discard_info edits .eh_frame or
> + bfd_elf_map_sections_to_segments updates program_header_size),
> + recalculate section addresses before relaxing. Otherwise,
> + lang_size_sections_1 updates output section VMAs top-down
> + while invoking bfd_relax_section, causing earlier sections to
> + see the new VMAs while later sections still have stale VMAs. */
> + if (need_layout)
> + {
> + lang_do_assignments (lang_assigning_phase_enum);
> + lang_reset_memory_regions ();
> + lang_size_sections (NULL, false);
> + }
> +
> /* We may need more than one relaxation pass. */
> int i = link_info.relax_pass;
>
> --
> 2.56.0
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://sourceware.org/pipermail/binutils/attachments/20261108/603cd49a/attachment.htm>
More information about the Binutils
mailing list