Re: .reloc improvement

["Maciej W. Rozycki" <macro at codesourcery dot com>]

Hi Alan,

On Tue, 6 Sep 2011, Maciej W. Rozycki wrote:

> On Mon, 22 Aug 2011, Alan Modra wrote:
> 
> > >  Please note that on REL MIPS targets local symbols have to be retained in 
> > > several cases where crucial information is lost if a relocation against 
> > > such a symbol is converted to one against the containing section's symbol.  
> > > This applies to all the PC-relative relocations, where the relocatable 
> > > field is too narrow to hold arbitrary addends, and also, more recently, to 
> > > microMIPS targets, where linker relaxation has to know local symbol 
> > > (label) addresses to perform branch displacement recalculations and to 
> > > make the LUI/ADDIU->ADDIUPC relaxation.  The latter is a linker's internal 
> > > limitation and may possibly be lifted (possibly by using the RELA 
> > > representation internally even on REL targets), but the former is an 
> > > inherent problem of the object file format used.
> > > 
> > >  Just making sure these issues are not missed -- I can't have a look at 
> > > your change at the moment and see what exact implications it has, sorry.
> > 
> > The implication of my change is that the programmer will need to be
> > careful in choosing symbols used with .reloc.  While that isn't ideal,
> > .reloc is such a low-level interface that you need to know what you're
> > doing if you use it.
> 
>  Problem is on REL targets some MIPS relocation types can never ever be 
> safely converted to refer to a section symbol + addend instead of the 
> intended symbol.  You'd therefore make .reloc useless for these types.
> 
>  Yes, REL shouldn't have been chosen for the MIPS ABI in the first place, 
> with the complication observed here being the tip of an iceberg only, but 
> there you go.  The choice was later fixed with the new ABIs made for 
> 64-bit MIPS processors, but the old ABI still remains for 32-bit ones.
> 
> > Alternatively, I'm quite willing to disable the symbol to section symbol 
> > conversion for REL targets.
> 
>  Good idea, that sounds like a plan to me.  That could be made platform 
> specific if that made anybody's life easier.

 Where are we with this problem?  Your change regressed this program (an 
excerpt from a test case I prepared a while ago and was about to integrate 
with our test suite):

$ cat reloc.s
	.text

	.fill	0x1000000

	.set	micromips
foo:
	nop32
	.reloc	., R_MICROMIPS_PC23_S2, 0f
	.fill	0
	addiu	$2, $pc, 8
	nop32
0:

from this:

$ mips-sde-elf-as -32 -o reloc.o reloc.s
$ mips-sde-elf-objdump -dr reloc.o

reloc.o:     file format elf32-tradbigmips


Disassembly of section .text:

00000000 <foo-0x1000000>:
	...

01000000 <foo>:
 1000000:	0000 0000 	nop
 1000004:	7900 0002 	addiu	v0,$pc,8
			1000004: R_MICROMIPS_PC23_S2	.L1
 1000008:	0000 0000 	nop

to this:

$ mips-sde-elf-as -o reloc.o reloc.s
reloc.s: Assembler messages:
reloc.s:8: Error: relocation overflow

  Maciej