[PATCH v2] RISC-V: Add no-op relax pass 0 to update section VMAs before shortening [PR34979]
Kito Cheng
kito.cheng@gmail.com
Thu Nov 19 09:07:29 GMT 2026
Hi Zhuo:
Thanks for the patch, v2 is LGTM,
although I am little suppress that could be fixed by adding a dummy
relax pass :P
Verified on my local, pushed to trunk :)
Zhuo Zhi <xeonacid@archlinuxcn.org> 於 2026年11月11日週三 下午9:33寫道:
>
> When ldelf_map_segments invokes lang_relax_sections (true) after either
> bfd_elf_discard_info edits .eh_frame or bfd_elf_map_sections_to_segments
> increases elf_program_header_size (growing SIZEOF_HEADERS),
> lang_size_sections_1 walks the linker script top-down and updates each
> output section's VMA immediately before invoking bfd_relax_section on
> its input sections.
>
> As a result, when relaxing an earlier output section (such as .text)
> that calls a symbol in a later output section, the earlier output
> section already has its new VMA (shifted forward by the SIZEOF_HEADERS
> increase), whereas the later output section still holds its stale VMA
> from the previous iteration.
>
> Although _bfd_riscv_relax_call and _bfd_riscv_relax_jal already reserve
> max_alignment slack for alignment-induced section movement, max_alignment
> can be smaller than the SIZEOF_HEADERS growth (for example, max_alignment
> is 8 bytes while 2 * sizeof(Elf64_Phdr) is 112 bytes). The temporary
> stale-VMA discrepancy causes _bfd_riscv_relax_call to underestimate the
> forward PC-relative offset and irreversibly shorten auipc+jalr into jal,
> which then overflows R_RISCV_JAL once the target output section's VMA is
> updated:
>
> relocation truncated to fit: R_RISCV_JAL against symbol `...'
>
> Per Alan Modra's suggestion, prepend a no-op "layout" pass at pass 0 in
> bfd_elfNN_riscv_init_relax_passes so that all output section VMAs are
> recalculated consistently before the "shorten" pass runs, without adding
> an extra sizing pass to other targets.
>
> bfd/
> PR ld/34979
> * elfnn-riscv.c (struct riscv_relax_pass): Document NULL select.
> (riscv_relax_pass_layout): New no-op relax pass.
> (bfd_elfNN_riscv_init_relax_passes): Register
> riscv_relax_pass_layout as pass 0.
> (_bfd_riscv_relax_section): Return early when pass->select is
> NULL.
>
> ld/
> PR ld/34979
> * testsuite/ld-riscv-elf/pr34979.s: New test source.
> * testsuite/ld-riscv-elf/pr34979-rv32.d: New test.
> * testsuite/ld-riscv-elf/pr34979-rv64.d: New test.
> * testsuite/ld-riscv-elf/ld-riscv-elf.exp: Run pr34979-rv32 and
> pr34979-rv64.
>
> Signed-off-by: Zhuo Zhi <xeonacid@archlinuxcn.org>
> ---
> Thanks Alan and Kito for the feedback!
>
> Alan: _bfd_riscv_relax_call and _bfd_riscv_relax_jal already subtract
> max_alignment (the maximum alignment across output sections) as slack
> for alignment-induced section movement. The failure in PR34979 happened
> because max_alignment was only 8 bytes whereas SIZEOF_HEADERS grew by
> 112 bytes (2 * sizeof(Elf64_Phdr)), creating a temporary stale-VMA gap
> much larger than max_alignment even though the true distance between the
> sections never changed. Following your suggestion, v2 drops the generic
> ld/ldlang.c change and instead adds a no-op pass 0 on RISC-V.
>
> Changes in v2:
> - Rebased onto master.
> - Dropped the generic ld/ldlang.c change; instead prepended a no-op
> "layout" pass at pass 0 in bfd_elfNN_riscv_init_relax_passes as
> suggested by Alan Modra.
> - Added RV32 and RV64 PIE regression tests (pr34979-rv32, pr34979-rv64)
> as requested by Kito Cheng.
>
> bfd/elfnn-riscv.c | 11 ++++++++++-
> ld/testsuite/ld-riscv-elf/ld-riscv-elf.exp | 2 ++
> ld/testsuite/ld-riscv-elf/pr34979-rv32.d | 18 ++++++++++++++++++
> ld/testsuite/ld-riscv-elf/pr34979-rv64.d | 18 ++++++++++++++++++
> ld/testsuite/ld-riscv-elf/pr34979.s | 10 ++++++++++
> 5 files changed, 58 insertions(+), 1 deletion(-)
> create mode 100644 ld/testsuite/ld-riscv-elf/pr34979-rv32.d
> create mode 100644 ld/testsuite/ld-riscv-elf/pr34979-rv64.d
> create mode 100644 ld/testsuite/ld-riscv-elf/pr34979.s
>
> diff --git a/bfd/elfnn-riscv.c b/bfd/elfnn-riscv.c
> index 17c714ab2c..78b62e6f33 100644
> --- a/bfd/elfnn-riscv.c
> +++ b/bfd/elfnn-riscv.c
> @@ -5760,7 +5760,9 @@ typedef struct riscv_relax_pass
>
> /* Decide how to relax RELOCS[*I]. Set *FUNC to the relax function, or
> leave it NULL to skip the reloc. May step *I over a paired
> - R_RISCV_RELAX. Return false on error. */
> + R_RISCV_RELAX. Return false on error. If SELECT itself is NULL,
> + nothing is relaxed in this pass so that section addresses are
> + recalculated first. */
> bool (*select) (bfd *, asection *, struct bfd_link_info *,
> Elf_Internal_Rela *, unsigned int *, relax_func_t *);
>
> @@ -5855,6 +5857,11 @@ riscv_relax_select_align (bfd *abfd ATTRIBUTE_UNUSED,
> return true;
> }
>
> +/* No-op pass to recalculate output section VMAs (e.g. after .eh_frame
> + edits or SIZEOF_HEADERS changes) before shortening instructions. */
> +static const riscv_relax_pass riscv_relax_pass_layout =
> + { "layout", false, NULL, NULL, NULL };
> +
> static const riscv_relax_pass riscv_relax_pass_shorten =
> { "shorten", false, riscv_relax_select_shorten, NULL, NULL };
>
> @@ -5894,6 +5901,7 @@ bfd_elfNN_riscv_init_relax_passes (struct bfd_link_info *info)
> htab->num_relax_passes = 0;
> htab->cur_relax_pass = -1;
>
> + riscv_add_relax_pass (htab, &riscv_relax_pass_layout);
> riscv_add_relax_pass (htab, &riscv_relax_pass_shorten);
> riscv_add_relax_pass (htab, &riscv_relax_pass_align);
>
> @@ -5932,6 +5940,7 @@ _bfd_riscv_relax_section (bfd *abfd, asection *sec,
> pass = htab->relax_passes[info->relax_pass];
>
> if (bfd_link_relocatable (info)
> + || pass->select == NULL
> || sec->sec_flg0
> || sec->reloc_count == 0
> || (sec->flags & SEC_RELOC) == 0
> diff --git a/ld/testsuite/ld-riscv-elf/ld-riscv-elf.exp b/ld/testsuite/ld-riscv-elf/ld-riscv-elf.exp
> index ad12f8d63b..cba0923c65 100644
> --- a/ld/testsuite/ld-riscv-elf/ld-riscv-elf.exp
> +++ b/ld/testsuite/ld-riscv-elf/ld-riscv-elf.exp
> @@ -286,6 +286,8 @@ if [istarget "riscv*-*-*"] {
>
> run_dump_test "pie-bind-locally-rv32"
> run_dump_test "pie-bind-locally-rv64"
> + run_dump_test "pr34979-rv32"
> + run_dump_test "pr34979-rv64"
>
> run_dump_test "discard-exe"
> run_dump_test "discard-pie"
> diff --git a/ld/testsuite/ld-riscv-elf/pr34979-rv32.d b/ld/testsuite/ld-riscv-elf/pr34979-rv32.d
> new file mode 100644
> index 0000000000..3d8b9d7106
> --- /dev/null
> +++ b/ld/testsuite/ld-riscv-elf/pr34979-rv32.d
> @@ -0,0 +1,18 @@
> +#source: pr34979.s
> +#as: -march=rv32i -mabi=ilp32 -mrelax
> +#ld: -m[riscv_choose_ilp32_emul] -pie --relax
> +#objdump: -d -Mno-aliases
> +
> +.*:[ ]+file format .*
> +
> +Disassembly of section \.text:
> +
> +0+[0-9a-f]+ <_start>:
> +[ ]+[0-9a-f]+:[ ]+[0-9a-f]+[ ]+auipc[ ]+ra,0x100
> +[ ]+[0-9a-f]+:[ ]+[0-9a-f]+[ ]+jalr[ ]+ra,4\(ra\) # [0-9a-f]+ <target>
> + \.\.\.
> +
> +Disassembly of section __lcxx_override:
> +
> +0+[0-9a-f]+ <target>:
> +[ ]+[0-9a-f]+:[ ]+[0-9a-f]+[ ]+jalr[ ]+zero,0\(ra\)
> diff --git a/ld/testsuite/ld-riscv-elf/pr34979-rv64.d b/ld/testsuite/ld-riscv-elf/pr34979-rv64.d
> new file mode 100644
> index 0000000000..bedc784ac7
> --- /dev/null
> +++ b/ld/testsuite/ld-riscv-elf/pr34979-rv64.d
> @@ -0,0 +1,18 @@
> +#source: pr34979.s
> +#as: -march=rv64i -mabi=lp64 -mrelax
> +#ld: -m[riscv_choose_lp64_emul] -pie --relax
> +#objdump: -d -Mno-aliases
> +
> +.*:[ ]+file format .*
> +
> +Disassembly of section \.text:
> +
> +0+[0-9a-f]+ <_start>:
> +[ ]+[0-9a-f]+:[ ]+[0-9a-f]+[ ]+auipc[ ]+ra,0x100
> +[ ]+[0-9a-f]+:[ ]+[0-9a-f]+[ ]+jalr[ ]+ra,4\(ra\) # [0-9a-f]+ <target>
> + \.\.\.
> +
> +Disassembly of section __lcxx_override:
> +
> +0+[0-9a-f]+ <target>:
> +[ ]+[0-9a-f]+:[ ]+[0-9a-f]+[ ]+jalr[ ]+zero,0\(ra\)
> diff --git a/ld/testsuite/ld-riscv-elf/pr34979.s b/ld/testsuite/ld-riscv-elf/pr34979.s
> new file mode 100644
> index 0000000000..1f4bffa101
> --- /dev/null
> +++ b/ld/testsuite/ld-riscv-elf/pr34979.s
> @@ -0,0 +1,10 @@
> + .text
> + .globl _start
> +_start:
> + call target
> + .space 0xffffc
> +
> + .section __lcxx_override,"ax",@progbits
> + .globl target
> +target:
> + ret
> --
> 2.56.0
>
More information about the Binutils
mailing list