optimization/858: Internal compiler error in scan_rtx_reg, at regrename.c:334

[Richard Henderson]

On Tue, Nov 21, 2000 at 08:04:34PM -0000, igor@txc.com wrote:
> bibli1.c: In function `minim00':
> bibli1.c:2742: Internal compiler error in scan_rtx_reg, at regrename.c:334
> Please submit a full bug report.
> See <URL: http://www.gnu.org/software/gcc/bugs.html > for instructions.

The alternative within sparc's movdf_insn_sp32 that we matched uses
an EXTRA_CONSTRAINT, from which we can derive no class information.

I suspect this usage of EXTRA_CONSTRAINT to be incorrect, or at least
sub-optimal, since c/859 appears to suffer from this as well.  However,
I'm not certain how I'd rearrange things for this pattern either, so
I'll work around the regrename crash for now.



r~


        * regrename.c (scan_rtx_reg): Terminate the chain rather than 
        abort on mark_read with NO_REGS.

Index: regrename.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/regrename.c,v
retrieving revision 1.16
diff -c -p -d -r1.16 regrename.c
*** regrename.c	2000/11/15 11:06:13	1.16
--- regrename.c	2000/11/22 02:37:50
*************** scan_rtx_reg (insn, loc, class, action, 
*** 325,350 ****
  
  	  if (regno + nregs <= this_regno
  	      || this_regno + this_nregs <= regno)
! 	    p = &this->next_chain;
! 	  else if (action == mark_read)
  	    {
  	      if (! exact_match)
  		abort ();
- 	      if (class == NO_REGS)
- 		abort ();
  
! 	      this = (struct du_chain *)
! 		obstack_alloc (&rename_obstack, sizeof (struct du_chain));
! 	      this->next_use = *p;
! 	      this->next_chain = (*p)->next_chain;
! 	      this->loc = loc;
! 	      this->insn = insn;
! 	      this->class = class;
! 	      this->need_caller_save_reg = 0;
! 	      *p = this;
! 	      return;
  	    }
! 	  else if (action != terminate_overlapping_read || ! exact_match)
  	    {
  	      struct du_chain *next = this->next_chain;
  
--- 325,360 ----
  
  	  if (regno + nregs <= this_regno
  	      || this_regno + this_nregs <= regno)
! 	    {
! 	      p = &this->next_chain;
! 	      continue;
! 	    }
! 
! 	  if (action == mark_read)
  	    {
  	      if (! exact_match)
  		abort ();
  
! 	      /* ??? Class NO_REGS can happen if the md file makes use of 
! 		 EXTRA_CONSTRAINTS to match registers.  Which is arguably
! 		 wrong, but there we are.  Since we know not what this may
! 		 be replaced with, terminate the chain.  */
! 	      if (class != NO_REGS)
! 		{
! 		  this = (struct du_chain *)
! 		    obstack_alloc (&rename_obstack, sizeof (struct du_chain));
! 		  this->next_use = *p;
! 		  this->next_chain = (*p)->next_chain;
! 		  this->loc = loc;
! 		  this->insn = insn;
! 		  this->class = class;
! 		  this->need_caller_save_reg = 0;
! 		  *p = this;
! 		  return;
! 		}
  	    }
! 
! 	  if (action != terminate_overlapping_read || ! exact_match)
  	    {
  	      struct du_chain *next = this->next_chain;