Re: assembler relaxation

["nagaraju.m" <nagaraju dot m at redpinesignals dot com>]

Hi Nick,

md_convert_frag (abfd, sec, fragp)
{

      ......
      ......

   case 6:
   value = (long) (target_address - (address_of_var_part  + 4)) + 4;
   value >>= 1;
   switch (opcodep[0])
   {
       case BEZ:
       //printf("BEZ 16 value = %d\n",value);
       opcodep[2] = 0x54; //emit beq, insertng its opcode
       break;
       case BNEZ:
       //printf("BNEZ 16 value = %d\n",value);
       opcodep[2] = 0x94; //emit bne, insertng its opcode
       break;
   }
   opcodep[0] = 0x7; //insert cmp instruction instead of bez
   opcodep[3] = ( value & 0xff );//insert offset
   fixp = fix_new (fragp, fragp->fr_fix, 4, fragp->fr_symbol,
          fragp->fr_offset, 1, BFD_RELOC_RPINE_S8_PCREL);
   fixp->fx_pcrel_adjust = 2;
   fragp->fr_fix += 2;

I am sending the part of md_convert frag.

we are trying to split

bez r0,label(branch if zero)   instruction
with
cmp r0,#0 (compare)
beq label (branch if equal)  instructions.  

I am able to see that the above code is placing correct opcode at that 
place but at the end we are seeing 00000000 in few cases.


example:
***************************************************
bez r0,L1 (if L1 > 5bits)

generate:

cmp r0,#0
beq L1
*********************************************************



I am pasting the bench mark code snippet of cg_compiler_opensrc (I have 
compiled support.c file with -c option)

file: support.s
=============================================

   ldw r6,r15,#16
   ldw r5,r15,#24
   or r6,r5
   *bez r6,.L503*
   ldw r1,r15,#20
   *bez r1,.L523*
   ldw r2,r15,#12
===============================================

objdump output after (rpine-as support.s then rpine-objdump a.out)

---------------------------------------------------------------------------------------------------

    1dce:    17 c1           0000c117     ldw        r6,sp,#0x10
   1dd0:    97 a1           0000a197     ldw        r5,sp,#0x18
   1dd2:    a3 65           000065a3      or         r6,r5
   1dd4:    07 c0           0000c007     cmp        r6,#0x0
   1dd6:    00 00           
00000000*                                     -----------------> issue 
at this line*
   1dd8:    57 21           00002157     ldw        r1,sp,#0x14
   */1dda:    07 20           00002007     cmp        r1,#0x0         
                                ------------------------> these two 
instructions are generated correctly instead of bez
   1ddc:    94 9f           00009f94      beq        1d1a 
<DeclareFunc+0x56>/*
---------------------------------------------------------------------------------------------------------
  
I have gone through the target files and up to that end of the target 
file it has the (beq) opcode but at the finally it is appending zero. 
Can you suggest where to look in to it. Are there any other documents 
other then internals I need to go through.


Thanks,
Nagaraju M



Nick Clifton wrote:
> Hi Nagaraju,
>
> >    I have ported Binutils to a new port. My problem is that in a 
> > 16-bit instruction if the constant range is exceeded then I need to 
> > emit  2 instructions. I have referred other architectures and did it 
> > and my assembler is generating instructions them. While running some 
> > bench mark codes it is observed instead of generating two 
> > instructions it is generating 1-st instruction correctly and for 
> > other instruction it is appending Zero's as a result it got struck up.
>
> Without access to the source code for your new port is hard to give 
> specific advice[1] but in general what you need to do is to look at 
> the where these instructions are generated/inserted and determine why 
> the second instruction is using zeroes.
>
> Note - normally relaxation works the other way.  That is it normally 
> removes instructions rather than adding them.  Thus I would have 
> expected that the default behaviour be for the compiler (or user, if 
> we are talking about hand generated assembler source files) to 
> generate the two instruction sequence, and for the assembler to relax 
> this to just one instruction if it can show that this will work.
>
> Cheers
>   Nick
>
> [1] I am not asking for a copy of the source code - I would rather 
> that you debug the problem yourself, possibly with some help from this 
> list.  Posting code fragments and more detailed examples of the 
> problem is a good idea though.