pattern problem with register assignment

[Jean-Marc Saffroy]

On 02/23/2011 09:52 PM, Christian Grössler wrote:
> Hello,
> 
> I have a problem with register allocation. Our architecture has some
> pointer registers "pX" (24bit)
> and some data registers "dX" (32bit). Since pointers are only 24bit,
> we're using PSImode for them.
> 
> There are restrictions in the "add" opcode, we can do
> 
> pX = add(pX,<imm>)
> pX = add(pX,dX)
> pX = add(dX,pX)
> dX = add(pX,pX)
> dX = add(pX,<imm>)
> dX = add(dX,dX)
> dX = add(dX,<imm>)
> dX = add(pX,dX)
> dX = add(dX,pX)
> 
> pX is allowed in at most 2 of the registers involved, but not in all 3
> of them.
> The pattern to add 2 PSImode values looks like this:
> 
> (define_insn "*addpsi3"
>   [(set (match_operand:PSI 0 "p_d_operand"                  
> "=a,d0d3,?d0d9 m,?d0d9 m")
>         (plus:PSI (match_operand:PSI 1 "p_d_general_operand"
> "%a,d0d3,d0d9 a m,d0d9 a i")
>           (match_operand:PSI 2 "p_d_general_operand" "i d0d9 m,d0d3 m
> i,d0d9 i a,d0d9 a m")))]
>   ""
>   "%0=add(%1,%2)"
> )
> 
> It worked fine in the gcc version from 2 years ago, but I'm updating the
> port to current gcc, and
> I get a testsuite failure (one of many :-)) in
> gcc.c-torture/compile/20080812-1.c:
> 
> 20080812-1.c: In function 'foo':
> 20080812-1.c:21:1: error: insn does not satisfy its constraints:
> (insn 49 76 77 3 (set (reg:PSI 6 p2 [144])
>         (plus:PSI (reg:PSI 5 p1 [orig:145 ivtmp.1 ] [145])
>             (reg:PSI 7 p3))) 20080812-1.c:15 193 {*addpsi3}
>      (nil))
> 20080812-1.c:21:1: internal compiler error: in
> reload_cse_simplify_operands, at postreload.c:401
> Please submit a full bug report,
> with preprocessed source if appropriate.
> See <http://gcc.gnu.org/bugs.html> for instructions.
> 
> 
> It seems that gcc wants to create an instruction like "p2=add(p1,p3)".
> How can I tell him not to do that?
> I tried to fiddle with the "p_d_general_operand" predication and use a
> modified one for operand 2, but
> at the time the constraint is called, I only see pseudo registers, and
> don't know in which hard register
> they will appear at the end.
> 
> I'm using a gcc snapshot from Jan-19-2011.
> 
> regards,
> chris
> 
> 

FWIW, I faced a similar challenge for my private port (ie. define add
for different types of registers), and I got gcc to work by defining a
single pattern for all adds, with predicates that don't discriminate on
register classes: the register contraints do the job of selecting the
proper combinations.

But there is probably more than one way to do it.

Cheers,
JM