[ld]: Checks for address space overflow will reject any avr program.

[ld]: Checks for address space overflow will reject any avr program.

[Georg-Johann Lay]

Hi Tristan.

Your addition to ld/ldlang.c::lang_check_section_addresses() leads
to rejection of any program by ld, for example

int x;

int main (void)
{
     return 0;
}

And compiled with

$ avr-gcc-7 main.c -Os -mmcu=atmega8 -save-temps

/INSTALL/bin/../lib/gcc/avr/7.0.1/../../../../avr/bin/ld: section .data 
VMA wraps around address space
/INSTALL/bin/../lib/gcc/avr/7.0.1/../../../../avr/bin/ld: section .bss 
VMA wraps around address space
collect2: error: ld returned 1 exit status

The generated asm is (effectively):

	.file	"main.c"
	.section	.text.startup,"ax",@progbits
.global	main
	.type	main, @function
main:
	ldi r25,0
	ldi r24,0
	ret
	.size	main, .-main
	.comm	x,2,1
	.ident	"GCC: (GNU) 7.0.1 20170314 (experimental) [trunk revision 244001]"


I didn't actually check whether it's due to your change, but as
you are active there, you'll likely know the solution.

avr is using virtual addresses in order to linearize the address
space.


If there is no obvious solution to fix this for avr, it would be
highly appreciated to have a configure option so that respective
portions of ld sources can be deactivated.

If you need more information to reproduce the issue, please let
me know.


Johann

Re: [ld]: Checks for address space overflow will reject any avr program.

[Tristan Gingold]

[...]

> I didn't actually check whether it's due to your change, but as
> you are active there, you'll likely know the solution.
> 
> avr is using virtual addresses in order to linearize the address
> space.

Yes, I suppose that section offset is larger than address space.

I think I have an idea for this.

Tristan.

Re: [ld]: Checks for address space overflow will reject any avr program.

[Tristan Gingold]

> On 14 Mar 2017, at 12:16, Tristan Gingold <gingold@adacore.com> wrote:
> 
> [...]
> 
>> I didn't actually check whether it's due to your change, but as
>> you are active there, you'll likely know the solution.
>> 
>> avr is using virtual addresses in order to linearize the address
>> space.
> 
> Yes, I suppose that section offset is larger than address space.
> 
> I think I have an idea for this.

Does this work for you ?

Tristan.

diff --git a/ld/ldlang.c b/ld/ldlang.c
index a0638ea..cf7aadc 100644
--- a/ld/ldlang.c
+++ b/ld/ldlang.c
@@ -4782,16 +4782,23 @@ lang_check_section_addresses (void)
   addr_mask = (addr_mask << 1) + 1;
   for (s = link_info.output_bfd->sections; s != NULL; s = s->next)
     {
-      s_end = (s->vma + s->size) & addr_mask;
-      if (s_end != 0 && s_end < s->vma)
-	einfo (_("%X%P: section %s VMA wraps around address space\n"),
-	       s->name);
-      else
+      /* Only for allocated sections and if already in the address space.
+	 This handles targets like avr which have I+D small spaces linearized
+	 in the only one ELF address space.  */
+      if ((s->flags & SEC_ALLOC) != 0
+	  && (s->vma & addr_mask) == s->vma)
 	{
-	  s_end = (s->lma + s->size) & addr_mask;
-	  if (s_end != 0 && s_end < s->lma)
-	    einfo (_("%X%P: section %s LMA wraps around address space\n"),
+	  s_end = (s->vma + s->size) & addr_mask;
+	  if (s_end != 0 && s_end < s->vma)
+	    einfo (_("%X%P: section %s VMA wraps around address space\n"),
 		   s->name);
+	  else
+	    {
+	      s_end = (s->lma + s->size) & addr_mask;
+	      if (s_end != 0 && s_end < s->lma)
+		einfo (_("%X%P: section %s LMA wraps around address space\n"),
+		       s->name);
+	    }
 	}
     }
 

Re: [ld]: Checks for address space overflow will reject any avr program.

[Georg-Johann Lay]

On 14.03.2017 12:52, Tristan Gingold wrote:
>
>> On 14 Mar 2017, at 12:16, Tristan Gingold <gingold@adacore.com> wrote:
>>
>> [...]
>>
>>> I didn't actually check whether it's due to your change, but as
>>> you are active there, you'll likely know the solution.
>>>
>>> avr is using virtual addresses in order to linearize the address
>>> space.
>>
>> Yes, I suppose that section offset is larger than address space.
>>
>> I think I have an idea for this.
>
> Does this work for you ?
>
> Tristan.
>
> diff --git a/ld/ldlang.c b/ld/ldlang.c
> index a0638ea..cf7aadc 100644
> --- a/ld/ldlang.c
> +++ b/ld/ldlang.c
> @@ -4782,16 +4782,23 @@ lang_check_section_addresses (void)
>    addr_mask = (addr_mask << 1) + 1;
>    for (s = link_info.output_bfd->sections; s != NULL; s = s->next)
>      {
> -      s_end = (s->vma + s->size) & addr_mask;
> -      if (s_end != 0 && s_end < s->vma)
> -	einfo (_("%X%P: section %s VMA wraps around address space\n"),
> -	       s->name);
> -      else
> +      /* Only for allocated sections and if already in the address space.
> +	 This handles targets like avr which have I+D small spaces linearized
> +	 in the only one ELF address space.  */
> +      if ((s->flags & SEC_ALLOC) != 0
> +	  && (s->vma & addr_mask) == s->vma)
>  	{
> -	  s_end = (s->lma + s->size) & addr_mask;
> -	  if (s_end != 0 && s_end < s->lma)
> -	    einfo (_("%X%P: section %s LMA wraps around address space\n"),
> +	  s_end = (s->vma + s->size) & addr_mask;
> +	  if (s_end != 0 && s_end < s->vma)
> +	    einfo (_("%X%P: section %s VMA wraps around address space\n"),
>  		   s->name);
> +	  else
> +	    {
> +	      s_end = (s->lma + s->size) & addr_mask;
> +	      if (s_end != 0 && s_end < s->lma)
> +		einfo (_("%X%P: section %s LMA wraps around address space\n"),
> +		       s->name);
> +	    }
>  	}
>      }
>

I tried this patch with some simple programs and it worked for them.

Thanks for the fast fix.

Johann