Making generic ld testcases pass on more targets

Making generic ld testcases pass on more targets

[Alan Modra]

There are a number of ld testcases that fail on various targets, all due
to relocation problems.  The typical testcase does something like

 .long foo

and expects this to work regardless of the size of addresses.  On 64-bit
targets, the default linker script might locate foo somewhere above 4G,
which will mean a reloc overflow.  Also, some 64-bit targets don't
support 32-bit relocations in shared libraries.  At the other extreme, a
16-bit target might not even have 32-bit relocs.

So it would be nice to have a means of creating a suitably sized
address, if only for the ld testsuite.  One possibility is to define a
symbol (or trickery a la md_parse_name) which could be tested.  eg.

 .ifle __address_size - 32
  .ifle __address_size - 16
   .word foo
  .else
   .long foo
  .endif
 .else
  .quad foo
 .endif

Another way is to define a new pseudo, .dc.a say, that magically does
the above.  Easily hacked together like the following, but I'd like to
solicit opinion on which approach is better (or something else entirely)
before I commit.

	* read.c (potable): Add "dc.a".
	(cons_worker): Handle dc.a.

Index: gas/read.c
===================================================================
RCS file: /cvs/src/src/gas/read.c,v
retrieving revision 1.88
diff -u -p -r1.88 read.c
--- gas/read.c	31 Jan 2005 14:26:08 -0000	1.88
+++ gas/read.c	16 Feb 2005 11:41:01 -0000
@@ -263,6 +263,9 @@ static const pseudo_typeS potable[] = {
   {"common.s", s_mri_common, 1},
   {"data", s_data, 0},
   {"dc", cons, 2},
+#ifdef BFD_ASSEMBLER
+  {"dc.a", cons, 0},
+#endif
   {"dc.b", cons, 1},
   {"dc.d", float_cons, 'd'},
   {"dc.l", cons, 4},
@@ -3339,6 +3342,18 @@ cons_worker (register int nbytes,	/* 1=.
       return;
     }
 
+#ifdef BFD_ASSEMBLER
+  if (nbytes == 0)
+    {
+      /* Choose smallest of 1, 2, 4, 8 bytes that is large enough to
+	 contain an address.  */
+      nbytes = (stdoutput->arch_info->bits_per_address - 1) / 8;
+      nbytes |= nbytes >> 1;
+      nbytes |= nbytes >> 2;
+      nbytes += 1;
+    }
+#endif
+
 #ifdef md_cons_align
   md_cons_align (nbytes);
 #endif

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre

Re: Making generic ld testcases pass on more targets

[Nick Clifton]

Hi Alan,

> +#ifdef BFD_ASSEMBLER
> +  {"dc.a", cons, 0},
> +#endif

I like this, but why make the pseudo specific to BFD_ASSEMBLERs ?


> +#ifdef BFD_ASSEMBLER
> +  if (nbytes == 0)
> +    {
> +      /* Choose smallest of 1, 2, 4, 8 bytes that is large enough to
> +	 contain an address.  */
> +      nbytes = (stdoutput->arch_info->bits_per_address - 1) / 8;

It might make sense to allow for individual targets to override this 
computation with their own value if they so wish.

Cheers
   Nick

Re: Making generic ld testcases pass on more targets

[Alan Modra]

On Wed, Feb 16, 2005 at 12:14:09PM +0000, Nick Clifton wrote:
> >+#ifdef BFD_ASSEMBLER
> >+  {"dc.a", cons, 0},
> >+#endif
> 
> I like this, but why make the pseudo specific to BFD_ASSEMBLERs ?

Because only BFD_ASSEMBLER has arch_info.

> >+      nbytes = (stdoutput->arch_info->bits_per_address - 1) / 8;
> 
> It might make sense to allow for individual targets to override this 
> computation with their own value if they so wish.

Yes, that would allow non-BFD_ASSEMBLER to use dc.a too.

Another approach occurred to me also, a variation on using
__address_size.  Instead of the assembler providing the symbol, we could
have the test harness feed a suitable --defsym to the assembler, chosen
by target name.

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre

Re: Making generic ld testcases pass on more targets

[Nick Clifton]

Hi Alan,

>>>+#ifdef BFD_ASSEMBLER
>>>+  {"dc.a", cons, 0},
>>>+#endif
>>
>>I like this, but why make the pseudo specific to BFD_ASSEMBLERs ?
>
> Because only BFD_ASSEMBLER has arch_info.

Ah - but if we want to use the pseudo in ld testcases then it would need 
to be present for non BFD assemblers as well.

>>>+      nbytes = (stdoutput->arch_info->bits_per_address - 1) / 8;
>>
>>It might make sense to allow for individual targets to override this 
>>computation with their own value if they so wish.
> 
> Yes, that would allow non-BFD_ASSEMBLER to use dc.a too.

OK - this would be my preferred solution.

> Another approach occurred to me also, a variation on using
> __address_size.  Instead of the assembler providing the symbol, we could
> have the test harness feed a suitable --defsym to the assembler, chosen
> by target name.

The problem here is remembering to do this when creating a new target.

Plus it is a duplication of information.  As you have pointed out the 
address size is already in the arch_info structure (for BFD assemblers) 
so why force the developer to put it into the testsuite as well.

Plus it introduces target specific information into a file that might 
otherwise be generic.

Cheers
   Nick

Re: Making generic ld testcases pass on more targets

[Hans-Peter Nilsson]

On Wed, 16 Feb 2005, Alan Modra wrote:
> Another way is to define a new pseudo, .dc.a say, that magically does
> the above.

I do like the idea (in contrast to the --defsym suggestion).

But please make the name less collision-prone.  How about
".gnu.dc.a"?

While you're at it, some fixed-size relocs would be nice too:
".gnu.dc.1" .. ".gnu.dc.4" (and 8 for 64-bitters).

brgds, H-P

Re: Making generic ld testcases pass on more targets

[Hans-Peter Nilsson]

On Wed, 16 Feb 2005, Hans-Peter Nilsson wrote:
> While you're at it, some fixed-size relocs would be nice too:

s/relocs/pseudos/

brgds, H-P

Re: Making generic ld testcases pass on more targets

[Alan Modra]

On Wed, Feb 16, 2005 at 08:31:25PM -0500, Hans-Peter Nilsson wrote:
> But please make the name less collision-prone.  How about
> ".gnu.dc.a"?

Hmm.  Do you know of some other assembler that uses dc.a?

> While you're at it, some fixed-size relocs would be nice too:
> ".gnu.dc.1" .. ".gnu.dc.4" (and 8 for 64-bitters).

You already have them.  Lots of 'em.  See read.c potable.

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre

Re: Making generic ld testcases pass on more targets

[Hans-Peter Nilsson]

On Thu, 17 Feb 2005, Alan Modra wrote:

> On Wed, Feb 16, 2005 at 08:31:25PM -0500, Hans-Peter Nilsson wrote:
> > But please make the name less collision-prone.  How about
> > ".gnu.dc.a"?
>
> Hmm.  Do you know of some other assembler that uses dc.a?

Nope.  But I know of some other that use .dc.C (C=b,w).

It just seems painlessly safer to use a unique prefix, now that
we don't *have* to cast *any* doubt whatsoever of colliding with
any pre-existing convention.  I mean, why not?

> > While you're at it, some fixed-size relocs would be nice too:
> > ".gnu.dc.1" .. ".gnu.dc.4" (and 8 for 64-bitters).
>
> You already have them.  Lots of 'em.  See read.c potable.

But still only with the "dc." prefix and similar.

brgds, H-P

Re: Making generic ld testcases pass on more targets

[Alan Modra]

On Wed, Feb 16, 2005 at 10:37:44PM -0500, Hans-Peter Nilsson wrote:
> On Thu, 17 Feb 2005, Alan Modra wrote:
> 
> > On Wed, Feb 16, 2005 at 08:31:25PM -0500, Hans-Peter Nilsson wrote:
> > > But please make the name less collision-prone.  How about
> > > ".gnu.dc.a"?
> >
> > Hmm.  Do you know of some other assembler that uses dc.a?
> 
> Nope.  But I know of some other that use .dc.C (C=b,w).

Sure.  We do too.

> It just seems painlessly safer to use a unique prefix, now that
> we don't *have* to cast *any* doubt whatsoever of colliding with
> any pre-existing convention.  I mean, why not?

I think the GNU tools have evolved far enough that it's reasonable to
let other competing toolchains worry about being compatible with us.  ;)

> > > While you're at it, some fixed-size relocs would be nice too:
> > > ".gnu.dc.1" .. ".gnu.dc.4" (and 8 for 64-bitters).
> >
> > You already have them.  Lots of 'em.  See read.c potable.
> 
> But still only with the "dc." prefix and similar.

Which are perfectly good to use.

gas/
	* read.c (address_bytes): New function.
	(TC_ADDRESS_BYTES): Default for BSD_ASSEMBLER to address_bytes.
	(potable): Add "dc.a".
	(cons_worker): Handle "dc.a".
	* doc/internals.texi (TC_ADDRESS_BYTES): Document.
ld/testsuite/
	* ld-elf/exclude1.s: Use ".dc.a".
	* ld-elfvsb/hidden2.s: Likewise.

Index: gas/read.c
===================================================================
RCS file: /cvs/src/src/gas/read.c,v
retrieving revision 1.90
diff -u -p -r1.90 read.c
--- gas/read.c	19 Feb 2005 08:56:08 -0000	1.90
+++ gas/read.c	21 Feb 2005 06:41:47 -0000
@@ -243,6 +243,24 @@ read_begin (void)
     lex_type['?'] = 3;
 }
 \f
+#ifndef TC_ADDRESS_BYTES
+#ifdef BFD_ASSEMBLER
+#define TC_ADDRESS_BYTES address_bytes
+
+static inline int
+address_bytes (void)
+{
+  /* Choose smallest of 1, 2, 4, 8 bytes that is large enough to
+     contain an address.  */
+  int n = (stdoutput->arch_info->bits_per_address - 1) / 8;
+  n |= n >> 1;
+  n |= n >> 2;
+  n += 1;
+  return n;
+}
+#endif
+#endif
+
 /* Set up pseudo-op tables.  */
 
 static struct hash_control *po_hash;
@@ -263,6 +281,9 @@ static const pseudo_typeS potable[] = {
   {"common.s", s_mri_common, 1},
   {"data", s_data, 0},
   {"dc", cons, 2},
+#ifdef TC_ADDRESS_BYTES
+  {"dc.a", cons, 0},
+#endif
   {"dc.b", cons, 1},
   {"dc.d", float_cons, 'd'},
   {"dc.l", cons, 4},
@@ -3335,6 +3356,11 @@ cons_worker (register int nbytes,	/* 1=.
       return;
     }
 
+#ifdef TC_ADDRESS_BYTES
+  if (nbytes == 0)
+    nbytes = TC_ADDRESS_BYTES ();
+#endif
+
 #ifdef md_cons_align
   md_cons_align (nbytes);
 #endif
Index: gas/doc/internals.texi
===================================================================
RCS file: /cvs/src/src/gas/doc/internals.texi,v
retrieving revision 1.41
diff -u -p -r1.41 internals.texi
--- gas/doc/internals.texi	19 Feb 2005 08:56:08 -0000	1.41
+++ gas/doc/internals.texi	21 Feb 2005 06:59:14 -0000
@@ -1114,6 +1114,11 @@ pseudo-op.
 @cindex TC_CONS_FIX_NEW
 You may define this macro to generate a fixup for a data allocation pseudo-op.
 
+@item TC_ADDRESS_BYTES
+@cindex TC_ADDRESS_BYTES
+Define this macro to specify the number of bytes used to store an address.
+Used to implement @code{dc.a}.  The target must have a reloc for this size.
+
 @item TC_INIT_FIX_DATA (@var{fixp})
 @cindex TC_INIT_FIX_DATA
 A C statement to initialize the target specific fields of fixup @var{fixp}.
Index: ld/testsuite/ld-elf/exclude1.s
===================================================================
RCS file: /cvs/src/src/ld/testsuite/ld-elf/exclude1.s,v
retrieving revision 1.2
diff -u -p -r1.2 exclude1.s
--- ld/testsuite/ld-elf/exclude1.s	19 Oct 2004 17:14:38 -0000	1.2
+++ ld/testsuite/ld-elf/exclude1.s	21 Feb 2005 06:14:07 -0000
@@ -1,4 +1,4 @@
 	.globl include_sym
 	.data
 include_sym:
-	.long	exclude_sym
+	.dc.a	exclude_sym
Index: ld/testsuite/ld-elfvsb/hidden2.s
===================================================================
RCS file: /cvs/src/src/ld/testsuite/ld-elfvsb/hidden2.s,v
retrieving revision 1.2
diff -u -p -r1.2 hidden2.s
--- ld/testsuite/ld-elfvsb/hidden2.s	11 Feb 2005 14:25:13 -0000	1.2
+++ ld/testsuite/ld-elfvsb/hidden2.s	21 Feb 2005 06:14:08 -0000
@@ -1,4 +1,4 @@
 	.data
 	.hidden foo
 	.global foo
-	.long foo
+	.dc.a foo

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre