[PATCH 2/6] aarch64: Fix .unreq of predefined aliases

[Matthieu Longo]

On 03/03/2026 12:38, Alice Carlotti wrote:
> Attempting to free a static reg_entry is invalid, so use
> insert_reg_alias to create a dynamic copy of the content.
> 
> 
> diff --git a/gas/config/tc-aarch64.c b/gas/config/tc-aarch64.c
> index 5e92007007310b9f12f1f0051890839216666c17..7af05abbb5c777de1e7e3fa9e9f910a9719a152e 100644
> --- a/gas/config/tc-aarch64.c
> +++ b/gas/config/tc-aarch64.c
> @@ -1599,7 +1599,7 @@ parse_simd_vector_with_bit_index (char **ccp, struct vector_type_el *typeinfo)
>   /* Directives: register aliases.  */
>   
>   static reg_entry *
> -insert_reg_alias (char *str, int number, aarch64_reg_type type)
> +insert_reg_alias (const char *str, int number, aarch64_reg_type type)
>   {
>     reg_entry *new;
>     const char *name;
> @@ -10589,8 +10589,17 @@ md_begin (void)
>   			aarch64_sys_regs_sr + i);
>   
>     for (i = 0; i < ARRAY_SIZE (reg_names); i++)
> -    checked_hash_insert (aarch64_reg_hsh, reg_names[i].name,
> -			 reg_names + i);
> +    {
> +      if (reg_names[i].builtin)
> +	checked_hash_insert (aarch64_reg_hsh, reg_names[i].name,
> +			     reg_names + i);
> +      else
> +	/* Aliases need to be safely removable, so create a copy of
> +	   the reg_entry data.  */
> +	insert_reg_alias (reg_names[i].name, reg_names[i].number,
> +			  reg_names[i].type);
> +    }
> +
>   
>     for (i = 0; i < ARRAY_SIZE (nzcv_names); i++)
>       checked_hash_insert (aarch64_nzcv_hsh, nzcv_names[i].template,
> diff --git a/gas/testsuite/gas/aarch64/register_aliases.s b/gas/testsuite/gas/aarch64/register_aliases.s
> index 856be5699ce434e41d52923020dcbc615208f983..10ee6d189170f6692cfe8e5f2c725fe138c72827 100644
> --- a/gas/testsuite/gas/aarch64/register_aliases.s
> +++ b/gas/testsuite/gas/aarch64/register_aliases.s
> @@ -10,3 +10,9 @@
>   	str 	IP0, [fp]
>   	ldr	IP1, [fp]
>   	str	zero, [x0]
> +
> +	# Removing builtin register aliases should not segfault.
> +	.unreq lr
> +	.unreq fp
> +	.unreq ip0
> +	.unreq ip1

I agree on the code fix, but not the testing.
Please look at patch 4 from my patch series that has more complete tests.

Matthieu