Constraints between operands

Constraints between operands

[Will Newton]

I'm using cgen to write a binutils port for a processor. I've come
across a problem I haven't been able to solve yet with expressing a
constraint between a number of fields in an instruction.

The instruction is encoded like this:

Bit 24: Data Unit Specifier
Bits 23-19: Destination Register
Bites 18-14: Source Register

(other bits omitted for simplicity)

The data unit bit specifies with set of registers this instruction
operates on. I have modelled this as a multi field for each register.
However I need to ensure that the assembler only allows destination and
source registers from the same unit to be specified and gives an error
otherwise, e.g.:

ADD D0.1,D0.2 ; Data unit 0 for both regs, OK
ADD D0.1,D1.2 ; Data unit mismatch, error!

Is there a way to express this with cgen? I have tried using (error) to
notify the user about this but cgen_rtx_error is not defined by opcodes
and I don't really know if it's possible to define it appropriately.

Re: Constraints between operands

[Frank Ch. Eigler]

Hi -

On Thu, Sep 08, 2005 at 02:55:16PM +0100, Will Newton wrote:
> [...]
> I'm using cgen to write a binutils port for a processor. I've come
> across a problem I haven't been able to solve yet with expressing a
> constraint between a number of fields in an instruction.
> [...]
> ADD D0.1,D0.2 ; Data unit 0 for both regs, OK
> ADD D0.1,D1.2 ; Data unit mismatch, error!
> [...]

One way may be to write a custom operand parser for the second D slot,
which would enforce this constraint.  It would signal a parse error.


- FChE

RE: Constraints between operands

[Will Newton]

 
> I was going to say have either a special "parse" or "insert" handler.
> I seem to recall other instances where I needed to validate 
> two operands against each other, but I can't find one at the moment.
> At any rate, you're right, the parse handler isn't passed 
> sufficient info.
> 
> If I can't find an existing example I think we need to extend cgen.
> e.g. pass the fields struct to either or both of the parse 
> and insert handlers.
> 
> Comments folks?

One thing that would seem useful is to be able to parse things easily
extending the existing infrastructure. I have come across two minor
instances where it could be useful to do this.

1. One of our addressing modes looks like this:

GETD  Reg1, [Reg2+#0x20++] ; Reg2+offset, post-increment Reg2

It is not possible to parse this using the builtin integer parser, it
barfs on the trailing ++. I implemented a parse handler for the integer
value and ended up using strtol, it would be nice if I could perhaps
pass an end pointer or length argument to cgen_parse_unsigned_integer to
tell it where to stop parsing and use that instead. Maybe this isn't
possible, it's not that big an issue.

2. To parse a register in a custom handler I ended up doing this:

  CGEN_OPERAND oper = meta_cgen_operand_table[opindex];
  int i;
  CGEN_KEYWORD *reg_names = NULL;

  for (i = 0; i < HW_MAX; i++)
    {
      if (meta_cgen_hw_table[i].type == oper.hw_type)
        {
          reg_names = (CGEN_KEYWORD *)meta_cgen_hw_table[i].asm_data;
        }
    }

  if (reg_names == NULL)
    return "internal error";

  errmsg = cgen_parse_keyword (cd, strp, reg_names, valuep);

Which is a little bit of a roundabout way to get something that perhaps
I should already have.

RE: Constraints between operands

[Doug Evans]

Will Newton writes:
 > 
 > > > [...]
 > > > I'm using cgen to write a binutils port for a processor. I've come 
 > > > across a problem I haven't been able to solve yet with expressing a 
 > > > constraint between a number of fields in an instruction.
 > > > [...]
 > > > ADD D0.1,D0.2 ; Data unit 0 for both regs, OK ADD D0.1,D1.2 ; Data 
 > > > unit mismatch, error!
 > > > [...]
 > > 
 > > One way may be to write a custom operand parser for the 
 > > second D slot, which would enforce this constraint.  It would 
 > > signal a parse error.

Another alternative I've seen is to specify a special operand at the
end whose sole purpose is to validate the insn.  It's at the end so it
has a view of the entire insn.

 > I could not see an easy way to reference a previous field. The prototype
 > of parse handlers is like:
 > 
 > static const char *
 > parse_hi16 (cd, strp, opindex, valuep)
 >      CGEN_CPU_DESC cd;                  /* CPU description */
 >      const char **strp;                 /* Current position in input
 > text */
 >      int opindex;                       /* ??? */
 >      unsigned long *valuep;             /* Result */
 > 
 > In order to find what the previous operand's data unit was the only way
 > I can see is to rewind strp and parse the input text, which may or may
 > not work and seems quite nasty. Or is there an easier way?

I was going to say have either a special "parse" or "insert" handler.
I seem to recall other instances where I needed to validate two
operands against each other, but I can't find one at the moment.
At any rate, you're right, the parse handler isn't passed sufficient info.

If I can't find an existing example I think we need to extend cgen.
e.g. pass the fields struct to either or both of the parse and insert
handlers.

Comments folks?

RE: Constraints between operands

[Will Newton]

> > [...]
> > I'm using cgen to write a binutils port for a processor. I've come 
> > across a problem I haven't been able to solve yet with expressing a 
> > constraint between a number of fields in an instruction.
> > [...]
> > ADD D0.1,D0.2 ; Data unit 0 for both regs, OK ADD D0.1,D1.2 ; Data 
> > unit mismatch, error!
> > [...]
> 
> One way may be to write a custom operand parser for the 
> second D slot, which would enforce this constraint.  It would 
> signal a parse error.

I could not see an easy way to reference a previous field. The prototype
of parse handlers is like:

static const char *
parse_hi16 (cd, strp, opindex, valuep)
     CGEN_CPU_DESC cd;                  /* CPU description */
     const char **strp;                 /* Current position in input
text */
     int opindex;                       /* ??? */
     unsigned long *valuep;             /* Result */

In order to find what the previous operand's data unit was the only way
I can see is to rewind strp and parse the input text, which may or may
not work and seems quite nasty. Or is there an easier way?