Re: [PATCH 1/2] LoongArch: bfd: Remove elf_seg_map condition in loongarch_elf_relax_section

[Xi Ruoyao <xry111@xry111.site>]

Hi Jinyang,

Any progress on this?  During my attempt trying to balance relaxation
and scheduling better in GCC I found the elf_seg_map condition is
preventing relaxation on *every* shared library (when we are linking a
shared library the elf_seg_map of .text is NULL).

On a modern system most code paths are in shared libraries, so it's
really bad not to perform the relaxation on them.  Esp. now we are
disabling explicit relocs for relaxation, so if we don't relax shared
libraries we are likely regressing the overall performance of the
system.

Sorry for disturbing you during the national holiday, feel free to defer
the reply until the holiday ends!

On Thu, 2023-07-13 at 16:18 +0800, Jinyang He wrote:
> On 2023-07-13 15:19, Jinyang He wrote:
> 
> > Hi,
> > 
> > 
> > We should pay attention to them, the bug in relaxation. Relaxation may 
> > not work correctly because of the bug1 reported by the second patch 
> > and the bug2 reported in the my previous email, although it works well 
> > in most of the time. Those 2 bugs are just what we've found, but there 
> > may be hidden problems.
> > I also tested the bug2 on RISCV and although it couldn't be resolved, 
> > it reported an error.
> > 
> > Test is,
> > $ cat c.s
> >   .section ".another.text", "x"
> >   .fill 1, 4, 0
> >   call fn2
> >   call fn2
> > fn1:
> >   nop
> > 
> > .text
> > fn2:
> >   .fill 4, 4, 0
> >   call fn1
> > 
> > $ ./gas/as-new  -march=rv32ic_zicsr -mno-arch-attr  -o c.o c.s
> > $ ./ld/ld-new c.o -o c -e 0 --section-start=.another.text=0x1000000 
> > -Ttext 0x1100000
> > c.o: in function `fn2':
> > (.text+0x10): relocation truncated to fit: R_RISCV_JAL against `fn1'
> > 
> > So, I'll find the reason why reloc_bits_pcrel20_s2() not reports the
> > overfollow error, and add error report which I think it is more 
> > important than these two patches so that we can get feedback when we
> > link.
> > 
> 
> Now, I find that the sign-overflow-check has bug, include 
> reloc_bits_pcrel20_s2(), reloc_bits_b26() and (others, I did not test 
> them). And I think the "mask" should be
>      bfd_signed_vma mask = ((bfd_signed_vma)0x1 << (howto->bitsize - 
> (sign ? 1 : 0)) - 1;
> 
> The test about reloc_bits_b26() is,
> 
> $ cat d.s
>          .section ".another.text", "x"
>          b       sym1
> 
>          .text
>          nop
> sym1:
>          nop
> 
> $ ./gas/as-new -mabi=lp64d -o d.o d.s
> $ ./ld/ld-new d.o -o d -e 0 --section-start=.another.text=0x130000000 
> -Ttext 0x120000000
> 
> I'll do fix and send v2 after you test spec2006. Or other responses?
> 
> 
> Thanks!
> 
> 
> > 
> > On 2023-07-13 11:13, mengqinggang wrote:
> > > Hi,
> > > 
> > > I will apply this two patches to see if there are any errors in 
> > > specc2006.
> > > 
> > > 在 2023/7/12 下午4:08, Jinyang He 写道:
> > > > Hi,
> > > > 
> > > > 
> > > > I'm not fimiliar with the process of output binary. I just noted
> > > > that it calls lang_relax_sections() before calling 
> > > > _bfd_elf_map_sections_to_segments() in ldelf_map_segments(). Thus I 
> > > > think the condition, elf_seg_map (info->output_bfd) == NULL, always 
> > > > be true when first time it calls loongarch_elf_relax_section(). 
> > > > Actually we can get right symbol value when relaxation, but ...
> > > > 
> > > > You reminder me that the symbol value will be reduced due to 
> > > > relaxation, which like the instruction pc is reduced. Then we cannot 
> > > > get right symbol value if the symbol in relaxable section. And I
> > > > write a test,
> > > > 
> > > > $ cat c.s
> > > >     .section ".another.text", "x"
> > > >     nop
> > > >     la.local $a0, sym1
> > > > sym2:
> > > >     nop
> > > > 
> > > >     .text
> > > > sym1:
> > > >     nop
> > > >     nop
> > > >     nop
> > > >     la.local $a0, sym2
> > > > 
> > > > $ ./gas/as-new -mabi=lp64d -o c.o c.s
> > > > $ ./ld/ld-new c.o -o c -e 0 
> > > > --section-start=.another.text=0x120000000 -Ttext 0x120200000
> > > > 
> > > > Disassembly of section .another.text:
> > > > 
> > > > 0000000120000000 <sym2-0x8>:
> > > >    120000000:    03400000     andi            $zero, $zero, 0x0
> > > >    120000004:    18ffffe4     pcaddi          $a0, 524287(0x7ffff)
> > > > 
> > > > 0000000120000008 <sym2>:
> > > >    120000008:    03400000     andi            $zero, $zero, 0x0
> > > > 
> > > > Disassembly of section .text:
> > > > 
> > > > 0000000120200000 <sym1>:
> > > >    120200000:    03400000     andi            $zero, $zero, 0x0
> > > >    120200004:    03400000     andi            $zero, $zero, 0x0
> > > >    120200008:    03400000     andi            $zero, $zero, 0x0
> > > >    12020000c:    18ffffe4     pcaddi          $a0, 524287(0x7ffff)
> > > > 
> > > > Due to the bug repaired by Patch2 and the bug reported by this test. 
> > > > And I also did not see that condition in other archs. I'm not sure 
> > > > that the relaxation spec2006 error is caused by that condition or 
> > > > those bugs.
> > > > 
> > > > 
> > > > On 2023-07-12 14:35, mengqinggang wrote:
> > > > > Hi,
> > > > > 
> > > > > Without this condition, the value of 'symbol - pc' may be 
> > > > > differences between
> > > > > loongarch_elf_relax_section and loongarch_elf_relocate_section.
> > > > > And it will cause some spec2006 test errors.
> > > > > 
> > > > > 在 2023/7/11 下午4:49, Jinyang He 写道:
> > > > > > When I tried "./ld-new sth." the relax does not work because the
> > > > > > elf_seg_map (info->output_bfd) was NULL in 
> > > > > > loongarch_elf_relax_section.
> > > > > > There is no need for the segment info, we can get the address through
> > > > > > the section info. Thus, remove this condition.
> > > > > > 
> > > > > > bfd/ChangeLog:
> > > > > > 
> > > > > >     * elfnn-loongarch.c (loongarch_elf_relax_section): Remove
> > > > > >     elf_seg_map condition.
> > > > > > 
> > > > > > ld/ChangeLog:
> > > > > > 
> > > > > >     * testsuite/ld-loongarch-elf/relax.exp: Update trigger method.
> > > > > >     * testsuite/ld-loongarch-elf/relax.dd: New test.
> > > > > > ---
> > > > > >   bfd/elfnn-loongarch.c                   |  1 -
> > > > > >   ld/testsuite/ld-loongarch-elf/relax.dd  |  6 +++++
> > > > > >   ld/testsuite/ld-loongarch-elf/relax.exp | 35 
> > > > > > +++++++------------------
> > > > > >   3 files changed, 16 insertions(+), 26 deletions(-)
> > > > > >   create mode 100644 ld/testsuite/ld-loongarch-elf/relax.dd
> > > > > > 
> > > > > > diff --git a/bfd/elfnn-loongarch.c b/bfd/elfnn-loongarch.c
> > > > > > index d3d8419d8..01f349a24 100644
> > > > > > --- a/bfd/elfnn-loongarch.c
> > > > > > +++ b/bfd/elfnn-loongarch.c
> > > > > > @@ -3845,7 +3845,6 @@ loongarch_elf_relax_section (bfd *abfd, 
> > > > > > asection *sec,
> > > > > >         || sec->sec_flg0
> > > > > >         || (sec->flags & SEC_RELOC) == 0
> > > > > >         || sec->reloc_count == 0
> > > > > > -      || elf_seg_map (info->output_bfd) == NULL
> > > > > >         || (info->disable_target_specific_optimizations
> > > > > >         && info->relax_pass == 0)
> > > > > >         /* The exp_seg_relro_adjust is enum phase_enum (0x4),
> > > > > > diff --git a/ld/testsuite/ld-loongarch-elf/relax.dd 
> > > > > > b/ld/testsuite/ld-loongarch-elf/relax.dd
> > > > > > new file mode 100644
> > > > > > index 000000000..e266beb8f
> > > > > > --- /dev/null
> > > > > > +++ b/ld/testsuite/ld-loongarch-elf/relax.dd
> > > > > > @@ -0,0 +1,6 @@
> > > > > > +#...
> > > > > > +.*pcaddi.*
> > > > > > +.*ld.w.*
> > > > > > +.*pcaddi.*
> > > > > > +.*ld.w.*
> > > > > > +#pass
> > > > > > diff --git a/ld/testsuite/ld-loongarch-elf/relax.exp 
> > > > > > b/ld/testsuite/ld-loongarch-elf/relax.exp
> > > > > > index 7ff876d79..d2bb967c6 100644
> > > > > > --- a/ld/testsuite/ld-loongarch-elf/relax.exp
> > > > > > +++ b/ld/testsuite/ld-loongarch-elf/relax.exp
> > > > > > @@ -21,31 +21,6 @@
> > > > > >     if [istarget loongarch64-*-*] {
> > > > > >   -  if [isbuild loongarch64-*-*] {
> > > > > > -    set testname "loongarch relax build"
> > > > > > -    set pre_builds [list \
> > > > > > -      [list \
> > > > > > -    "$testname" \
> > > > > > -    "" \
> > > > > > -    "" \
> > > > > > -    {relax.s} \
> > > > > > -    {} \
> > > > > > -    "relax" \
> > > > > > -      ] \
> > > > > > -    ]
> > > > > > -
> > > > > > -    run_cc_link_tests $pre_builds
> > > > > > -
> > > > > > -    if [file exist "tmpdir/relax"] {
> > > > > > -      set objdump_output [run_host_cmd "objdump" "-d tmpdir/relax"]
> > > > > > -      if { [ regexp ".*pcaddi.*pcaddi.*" $objdump_output] } {
> > > > > > -    pass "loongarch relax"
> > > > > > -      } {
> > > > > > -    fail "loongarch relax"
> > > > > > -      }
> > > > > > -    }
> > > > > > -  }
> > > > > > -
> > > > > >     run_ld_link_tests \
> > > > > >         [list \
> > > > > >         [list \
> > > > > > @@ -58,6 +33,16 @@ if [istarget loongarch64-*-*] {
> > > > > >             ] \
> > > > > >             "relax-align" \
> > > > > >         ] \
> > > > > > +      [list \
> > > > > > +          "relax" \
> > > > > > +          "-e 0 -Ttext 0x10000 -Tdata 0x20000" "" \
> > > > > > +          "" \
> > > > > > +          {relax.s} \
> > > > > > +          [list \
> > > > > > +          [list objdump -d relax.dd] \
> > > > > > +          ] \
> > > > > > +          "relax" \
> > > > > > +      ] \
> > > > > >         ]
> > > > > >       set objdump_flags "-s -j .data"
> 

-- 
Xi Ruoyao <xry111@xry111.site>
School of Aerospace Science and Technology, Xidian University