Psossible problem with gas on MIPS

Psossible problem with gas on MIPS

[Koundinya.K]

Hi,

Here is a a small test case that I could derive out of the large piece of 
assembler code that assembling

$ cat li.s

..set noreorder
li $2,7
beqz $7, noerror
nop


assembling the above with the native assembler 

$ /usr/bin/as -EB -KPIC li.s -o li.o

$ objdump -r li.o

li.o:     file format elf32-tradbigmips

RELOCATION RECORDS FOR [.text]:
OFFSET   TYPE              VALUE 
00000004 R_MIPS_PC16       noerror


Assembling the same using gas

$as -KPIC li.s -o li_gas.o

li.s: Assembler messages:
li.s:3: Error: Can not represent BFD_RELOC_16_PCREL_S2 relocation in this 
object file format

Could this be a possible bug in gas ?. I am looking at the source too. Could 
someone having access to MIPS boxes try this test case and verify ?


Thanks

koundinya

Re: Possible problem with gas on MIPS

[Ulf Carlsson]

> RELOCATION RECORDS FOR [.text]:
> OFFSET   TYPE              VALUE 
> 00000004 R_MIPS_PC16       noerror

How is this supposed to work?  R_MIPS_PC16 doesn't shift out the lower two
bits of the offset, so the branch instructions can not possibly use this
relocation type.  However can do this with the R_MIPS_GNU_REL16_S2 relocation
which is a Cygnus extension.

> li.s: Assembler messages:
> li.s:3: Error: Can not represent BFD_RELOC_16_PCREL_S2 relocation in this 
> object file format

I tried it and it goes through the assembler if you use -membedded-pic.  I
have not tested it any further though.

Ulf

Re: Possible problem with gas on MIPS

[Koundinya.K]

Hi,
 On the tons of assembler code I have I can see that R_MIPS_PC16 type of 
relocation is used by objects modules to contain external references from
branch opcodes.

I would suspect that the shift for R_MIPS_PC16 has not been documented in 
the ABI or elsewhere and has just been missed out.


Can't  R_MIPS_PC16 be used for BFD_RELOC_16_PCREL_S2 relocations?
This should solve the problem which would also allow object modules to 
contain external references from branch opcodes. Also there is not that much 
heavy use of R_MIPS_PC16.

koundinya




-> > RELOCATION RECORDS FOR [.text]:
-> > OFFSET   TYPE              VALUE 
-> > 00000004 R_MIPS_PC16       noerror
-> 
-> How is this supposed to work?  R_MIPS_PC16 doesn't shift out the lower two
-> bits of the offset, so the branch instructions can not possibly use this
-> relocation type.  However can do this with the R_MIPS_GNU_REL16_S2 relocation
-> which is a Cygnus extension.
-> 
-> > li.s: Assembler messages:
-> > li.s:3: Error: Can not represent BFD_RELOC_16_PCREL_S2 relocation in this 
-> > object file format
-> 
-> I tried it and it goes through the assembler if you use -membedded-pic.  I
-> have not tested it any further though.
-> 
-> Ulf
-> 

Re: Possible problem with gas on MIPS

[Ulf Carlsson]

> I would suspect that the shift for R_MIPS_PC16 has not been documented in 
> the ABI or elsewhere and has just been missed out.

Yes, you're actually right.  It has been missed out.  We do shift the offsets
right by 2 when we link objects including R_MIPS_PC16 with gld.  This is 
however nothing we want to generate with gas.  IRIX as is actually reluctant
to assemble branches to symbols as well.  I think IRIX ld pukes on R_MIPS_PC16
whenever it gets the chance.

Ulf

Re: Possible problem with gas on MIPS

[Koundinya.K]

O.K, here is a tiny patch that I applied to get my assembler stuff compiling 
for the same issue.

Index: elf32-mips.c
===================================================================
RCS file: /gnu/cvsroot/binutils-000213/bfd/elf32-mips.c,v
retrieving revision 1.2
diff -u -p -b -r1.2 elf32-mips.c
--- elf32-mips.c        2000/06/08 08:23:45     1.2
+++ elf32-mips.c        2000/06/08 10:13:49
@@ -1843,6 +1843,7 @@ static CONST struct elf_reloc_map mips_r
   { BFD_RELOC_MIPS_LITERAL, R_MIPS_LITERAL },
   { BFD_RELOC_MIPS_GOT16, R_MIPS_GOT16 },
   { BFD_RELOC_16_PCREL, R_MIPS_PC16 },
+  { BFD_RELOC_16_PCREL_S2, R_MIPS_PC16 },
   { BFD_RELOC_MIPS_CALL16, R_MIPS_CALL16 },
   { BFD_RELOC_MIPS_GPREL32, R_MIPS_GPREL32 },
   { BFD_RELOC_MIPS_GOT_HI16, R_MIPS_GOT_HI16 },




koundinya


-> > I would suspect that the shift for R_MIPS_PC16 has not been documented in 
-> > the ABI or elsewhere and has just been missed out.
-> 
-> Yes, you're actually right.  It has been missed out.  We do shift the offsets
-> right by 2 when we link objects including R_MIPS_PC16 with gld.  This is 
-> however nothing we want to generate with gas.  IRIX as is actually reluctant
-> to assemble branches to symbols as well.  I think IRIX ld pukes on R_MIPS_PC16
-> whenever it gets the chance.
-> 
-> Ulf
-> 

Re: Psossible problem with gas on MIPS

[Maciej W. Rozycki]

On Thu, 8 Jun 2000, Koundinya.K wrote:

> li.s: Assembler messages:
> li.s:3: Error: Can not represent BFD_RELOC_16_PCREL_S2 relocation in this 
> object file format
> 
> Could this be a possible bug in gas ?. I am looking at the source too. Could 
> someone having access to MIPS boxes try this test case and verify ?

  I recall myself discussing this issue here back in October.  The
conclusion was that due to a limited range and thus a questionable
usefulness relocations for branches are not defined by the MIPS ELF ABI. 
I prepared a patch to map BFD_RELOC_16_PCREL_S2 to R_MIPS_PC16 then but it
got rejected as not being ABI-compliant. 

 A R_MIPS_PC16 relocation that the vendor-supplied as generates in not
appropriate for branches.  That's beacuse R_MIPS_PC16 is defined as a
simple difference between two addresses in the ABI while for branches you
need a shifted left by two difference.

-- 
+  Maciej W. Rozycki, Technical University of Gdansk, Poland   +
+--------------------------------------------------------------+
+        e-mail: macro@ds2.pg.gda.pl, PGP key available        +

Re: Possible problem with gas on MIPS

[Ian Lance Taylor]

   Date: Thu, 08 Jun 2000 17:21:14 +0530
   From: "Koundinya.K" <kk@ddeorg.soft.net>

   O.K, here is a tiny patch that I applied to get my assembler stuff compiling 
   for the same issue.

   Index: elf32-mips.c
   ===================================================================
   RCS file: /gnu/cvsroot/binutils-000213/bfd/elf32-mips.c,v
   retrieving revision 1.2
   diff -u -p -b -r1.2 elf32-mips.c
   --- elf32-mips.c        2000/06/08 08:23:45     1.2
   +++ elf32-mips.c        2000/06/08 10:13:49
   @@ -1843,6 +1843,7 @@ static CONST struct elf_reloc_map mips_r
      { BFD_RELOC_MIPS_LITERAL, R_MIPS_LITERAL },
      { BFD_RELOC_MIPS_GOT16, R_MIPS_GOT16 },
      { BFD_RELOC_16_PCREL, R_MIPS_PC16 },
   +  { BFD_RELOC_16_PCREL_S2, R_MIPS_PC16 },
      { BFD_RELOC_MIPS_CALL16, R_MIPS_CALL16 },
      { BFD_RELOC_MIPS_GPREL32, R_MIPS_GPREL32 },
      { BFD_RELOC_MIPS_GOT_HI16, R_MIPS_GOT_HI16 },

Well, this patch can't be right.  Either BFD_RELOC_16_PCREL or
BFD_RELOC_16_PCREL_S2 can be translated to R_MIPS_PC16, but not both.
The two BFD relocations are significantly different; we can't possibly
translate them to the same R_MIPS relocation.

Which one is correct depends upon the definition of R_MIPS_PC16.

Ian