[PATCH v1] LoongArch: Do not relax against __[start|stop]_SECNAME symbol

[PATCH v1] LoongArch: Do not relax against __[start|stop]_SECNAME symbol

[Lulu Cai]

The value of the __[start|stop]_SECNAME symbol will be set after
loongarch_elf_relax_section. If relaxation is done against the
__[start|stop]_SECNAME symbol, the correct symbol value cannot
be obtained.
---
 bfd/elfnn-loongarch.c                      |  5 ++++-
 ld/testsuite/ld-loongarch-elf/relax.exp    |  2 ++
 ld/testsuite/ld-loongarch-elf/start_stop.d | 13 +++++++++++++
 ld/testsuite/ld-loongarch-elf/start_stop.s | 10 ++++++++++
 4 files changed, 29 insertions(+), 1 deletion(-)
 create mode 100644 ld/testsuite/ld-loongarch-elf/start_stop.d
 create mode 100644 ld/testsuite/ld-loongarch-elf/start_stop.s

diff --git a/bfd/elfnn-loongarch.c b/bfd/elfnn-loongarch.c
index efa57184e44..dfb760dbf22 100644
--- a/bfd/elfnn-loongarch.c
+++ b/bfd/elfnn-loongarch.c
@@ -5636,10 +5636,13 @@ loongarch_elf_relax_section (bfd *abfd, asection *sec,
 	}
       else
 	{
+	  /* Do not relax __[start|stop]_SECNAME, since the symbol value
+	     is not set yet.  */
 	  if (h != NULL
 	      && ((h->type == STT_GNU_IFUNC
 		   && r_type != R_LARCH_CALL36)
-		  || bfd_is_abs_section (h->root.u.def.section)))
+		  || bfd_is_abs_section (h->root.u.def.section)
+		  || h->start_stop))
 	    continue;
 
 	  /* The GOT entry of tls symbols must in current execute file or
diff --git a/ld/testsuite/ld-loongarch-elf/relax.exp b/ld/testsuite/ld-loongarch-elf/relax.exp
index 81924cd35cb..5c17a2ef403 100644
--- a/ld/testsuite/ld-loongarch-elf/relax.exp
+++ b/ld/testsuite/ld-loongarch-elf/relax.exp
@@ -484,4 +484,6 @@ if [istarget loongarch64-*-*] {
 	      "uleb128" \
 	  ] \
       ]
+
+  run_dump_test "start_stop"
 }
diff --git a/ld/testsuite/ld-loongarch-elf/start_stop.d b/ld/testsuite/ld-loongarch-elf/start_stop.d
new file mode 100644
index 00000000000..b887dcc2eb7
--- /dev/null
+++ b/ld/testsuite/ld-loongarch-elf/start_stop.d
@@ -0,0 +1,13 @@
+#ld: -z norelro -pie -e0
+#objdump: -d
+
+.*:     file format .*
+
+
+Disassembly of section sec1:
+
+[0-9a-f]+ <_start>:
+ [0-9a-f]+:	[0-9a-f]+ 	pcalau12i   	\$a0, .*
+ [0-9a-f]+:	[0-9a-f]+ 	ld.d        	\$a0, \$a0, .*
+ [0-9a-f]+:	[0-9a-f]+ 	pcalau12i   	\$a0, .*
+ [0-9a-f]+:	[0-9a-f]+ 	ld.d        	\$a0, \$a0, .*
diff --git a/ld/testsuite/ld-loongarch-elf/start_stop.s b/ld/testsuite/ld-loongarch-elf/start_stop.s
new file mode 100644
index 00000000000..f50bc603abb
--- /dev/null
+++ b/ld/testsuite/ld-loongarch-elf/start_stop.s
@@ -0,0 +1,10 @@
+	.section	sec1,"ax",@progbits
+	.globl	_start
+	.type   _start, @function
+_start:
+	la.got	$a0,__start_sec1
+	la.got	$a0,__stop_sec1
+	.size   _start, .-_start
+
+	.hidden	__start_sec1
+	.hidden	__stop_sec1
-- 
2.45.2

Re: [PATCH v1] LoongArch: Do not relax against __[start|stop]_SECNAME symbol

[Lulu Cai]

Hi Nick,


   This patch fixes a relaxation bug for LoongArch and needs to be 
applied to 2.44 branch and the mainline source.

   link: https://sourceware.org/pipermail/binutils/2025-January/138943.html

   Thanks.

On 1/22/2025 5:00 PM, Lulu Cai wrote:
> The value of the __[start|stop]_SECNAME symbol will be set after
> loongarch_elf_relax_section. If relaxation is done against the
> __[start|stop]_SECNAME symbol, the correct symbol value cannot
> be obtained.
> ---
>   bfd/elfnn-loongarch.c                      |  5 ++++-
>   ld/testsuite/ld-loongarch-elf/relax.exp    |  2 ++
>   ld/testsuite/ld-loongarch-elf/start_stop.d | 13 +++++++++++++
>   ld/testsuite/ld-loongarch-elf/start_stop.s | 10 ++++++++++
>   4 files changed, 29 insertions(+), 1 deletion(-)
>   create mode 100644 ld/testsuite/ld-loongarch-elf/start_stop.d
>   create mode 100644 ld/testsuite/ld-loongarch-elf/start_stop.s
>
> diff --git a/bfd/elfnn-loongarch.c b/bfd/elfnn-loongarch.c
> index efa57184e44..dfb760dbf22 100644
> --- a/bfd/elfnn-loongarch.c
> +++ b/bfd/elfnn-loongarch.c
> @@ -5636,10 +5636,13 @@ loongarch_elf_relax_section (bfd *abfd, asection *sec,
>   	}
>         else
>   	{
> +	  /* Do not relax __[start|stop]_SECNAME, since the symbol value
> +	     is not set yet.  */
>   	  if (h != NULL
>   	      && ((h->type == STT_GNU_IFUNC
>   		   && r_type != R_LARCH_CALL36)
> -		  || bfd_is_abs_section (h->root.u.def.section)))
> +		  || bfd_is_abs_section (h->root.u.def.section)
> +		  || h->start_stop))
>   	    continue;
>   
>   	  /* The GOT entry of tls symbols must in current execute file or
> diff --git a/ld/testsuite/ld-loongarch-elf/relax.exp b/ld/testsuite/ld-loongarch-elf/relax.exp
> index 81924cd35cb..5c17a2ef403 100644
> --- a/ld/testsuite/ld-loongarch-elf/relax.exp
> +++ b/ld/testsuite/ld-loongarch-elf/relax.exp
> @@ -484,4 +484,6 @@ if [istarget loongarch64-*-*] {
>   	      "uleb128" \
>   	  ] \
>         ]
> +
> +  run_dump_test "start_stop"
>   }
> diff --git a/ld/testsuite/ld-loongarch-elf/start_stop.d b/ld/testsuite/ld-loongarch-elf/start_stop.d
> new file mode 100644
> index 00000000000..b887dcc2eb7
> --- /dev/null
> +++ b/ld/testsuite/ld-loongarch-elf/start_stop.d
> @@ -0,0 +1,13 @@
> +#ld: -z norelro -pie -e0
> +#objdump: -d
> +
> +.*:     file format .*
> +
> +
> +Disassembly of section sec1:
> +
> +[0-9a-f]+ <_start>:
> + [0-9a-f]+:	[0-9a-f]+ 	pcalau12i   	\$a0, .*
> + [0-9a-f]+:	[0-9a-f]+ 	ld.d        	\$a0, \$a0, .*
> + [0-9a-f]+:	[0-9a-f]+ 	pcalau12i   	\$a0, .*
> + [0-9a-f]+:	[0-9a-f]+ 	ld.d        	\$a0, \$a0, .*
> diff --git a/ld/testsuite/ld-loongarch-elf/start_stop.s b/ld/testsuite/ld-loongarch-elf/start_stop.s
> new file mode 100644
> index 00000000000..f50bc603abb
> --- /dev/null
> +++ b/ld/testsuite/ld-loongarch-elf/start_stop.s
> @@ -0,0 +1,10 @@
> +	.section	sec1,"ax",@progbits
> +	.globl	_start
> +	.type   _start, @function
> +_start:
> +	la.got	$a0,__start_sec1
> +	la.got	$a0,__stop_sec1
> +	.size   _start, .-_start
> +
> +	.hidden	__start_sec1
> +	.hidden	__stop_sec1

Re: [PATCH v1] LoongArch: Do not relax against __[start|stop]_SECNAME symbol

[Nick Clifton]

Hi Lulu,

>    This patch fixes a relaxation bug for LoongArch and needs to be applied to 2.44 branch and the mainline source.
> 
>    link: https://sourceware.org/pipermail/binutils/2025-January/138943.html

When I test this patch on my local machine (an x86_64 box) I get a
new linker testsuite failure:

   FAIL: ld-loongarch-elf/start_stop

The reason, according to ld.log is:

   Executing on host: sh -c {./ld-new   -L/work/sources/binutils/upstream/2.44/ld/testsuite/ld-loongarch-elf  -z norelro -pie -e0 -o tmpdir/dump tmpdir/start_stop.o  2>&1} 
/dev/null dump.tmp (timeout = 300)
   spawn [open ...]
   ./ld-new: -pie not supported
   failed with: <./ld-new: -pie not supported>, no expected output

I think that this must be a problem with my local setup, so please
go ahead and apply the fix to both the mainline and the 2.44 branch.

But if you can, please do run a quick check to make sure that the
linker test does actually pass for you...

Cheers
   Nick

Re: [PATCH v1] LoongArch: Do not relax against __[start|stop]_SECNAME symbol

[Lulu Cai]

On 1/28/2025 9:45 PM, Nick Clifton wrote:
> Hi Lulu,
>
>>    This patch fixes a relaxation bug for LoongArch and needs to be 
>> applied to 2.44 branch and the mainline source.
>>
>>    link: 
>> https://sourceware.org/pipermail/binutils/2025-January/138943.html
>
> When I test this patch on my local machine (an x86_64 box) I get a
> new linker testsuite failure:
>
>   FAIL: ld-loongarch-elf/start_stop
>
> The reason, according to ld.log is:
>
>   Executing on host: sh -c {./ld-new 
> -L/work/sources/binutils/upstream/2.44/ld/testsuite/ld-loongarch-elf 
> -z norelro -pie -e0 -o tmpdir/dump tmpdir/start_stop.o 2>&1} /dev/null 
> dump.tmp (timeout = 300)
>   spawn [open ...]
>   ./ld-new: -pie not supported
>   failed with: <./ld-new: -pie not supported>, no expected output
>
> I think that this must be a problem with my local setup, so please
> go ahead and apply the fix to both the mainline and the 2.44 branch.

Hi Nick,

   Thanks for the problem report otherwise I would have missed it.

   I have sent patch V2 to fix it by checking for loongarch64 target and 
pie support. Please apply this patch to mainline and 2.44 branch.

   Link: https://sourceware.org/pipermail/binutils/2025-January/139161.html

Thanks
   Lulu

Re: [PATCH v1] LoongArch: Do not relax against __[start|stop]_SECNAME symbol

[Nick Clifton]

Hi Lulu,

>    Thanks for the problem report otherwise I would have missed it.
> 
>    I have sent patch V2 to fix it by checking for loongarch64 target and pie support. Please apply this patch to mainline and 2.44 branch.
> 
>    Link: https://sourceware.org/pipermail/binutils/2025-January/139161.html

Thanks - I have applied this patch to the mainline and the 2.44 branch.

Cheers
   Nick