Re: patch to fix PR88282

[Richard Sandiford <richard.sandiford@arm.com>]

Vladimir Makarov <vmakarov@redhat.com> writes:
>    The following patch fixes
>
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88282
>
>    The patch was successfully bootstrapped and tested on 
> x86/x86-64/ppc64/aarch64.
>
>    Committed as rev. 266784.
>
> Index: ChangeLog
> ===================================================================
> --- ChangeLog	(revision 266783)
> +++ ChangeLog	(working copy)
> @@ -1,3 +1,9 @@
> +2018-12-04  Vladimir Makarov  <vmakarov@redhat.com>
> +
> +	PR target/88282
> +	* ira-costs.c (exec): Try bigger class to use smaller register
> +	move cost.
> +
>  2018-12-04  Michael Ploujnikov  <michael.ploujnikov@oracle.com>
>  
>  	PR ipa/88297
> Index: ira-costs.c
> ===================================================================
> --- ira-costs.c	(revision 266678)
> +++ ira-costs.c	(working copy)
> @@ -1314,28 +1314,50 @@ record_operand_costs (rtx_insn *insn, en
>  	  machine_mode mode = GET_MODE (SET_SRC (set));
>  	  cost_classes_t cost_classes_ptr = regno_cost_classes[regno];
>  	  enum reg_class *cost_classes = cost_classes_ptr->classes;
> -	  reg_class_t rclass, hard_reg_class, pref_class;
> +	  reg_class_t rclass, hard_reg_class, pref_class, bigger_hard_reg_class;
>  	  int cost, k;
> +	  move_table *move_costs;
>  	  bool dead_p = find_regno_note (insn, REG_DEAD, REGNO (src));
>  
>  	  ira_init_register_move_cost_if_necessary (mode);
> +	  move_costs = ira_register_move_cost[mode];
>  	  hard_reg_class = REGNO_REG_CLASS (other_regno);
> +	  bigger_hard_reg_class = ira_pressure_class_translate[hard_reg_class];
> +	  if (bigger_hard_reg_class == NO_REGS
> +	      && (other_regno == STACK_POINTER_REGNUM
> +#ifdef STATIC_CHAIN_REGNUM
> +		  || other_regno == STATIC_CHAIN_REGNUM
> +#endif
> +		  || other_regno == FRAME_POINTER_REGNUM
> +		  || other_regno == HARD_FRAME_POINTER_REGNUM))
> +	    bigger_hard_reg_class = GENERAL_REGS;
>  	  /* Target code may return any cost for mode which does not
>  	     fit the the hard reg class (e.g. DImode for AREG on
>  	     i386).  Check this and use a bigger class to get the
>  	     right cost.  */
>  	  if (! ira_hard_reg_in_set_p (other_regno, mode,
>  				       reg_class_contents[hard_reg_class]))
> -	    hard_reg_class = ira_pressure_class_translate[hard_reg_class];
> +	    hard_reg_class = bigger_hard_reg_class;
>  	  i = regno == (int) REGNO (src) ? 1 : 0;
>  	  for (k = cost_classes_ptr->num - 1; k >= 0; k--)
>  	    {
>  	      rclass = cost_classes[k];
> -	      cost = ((i == 0
> -		       ? ira_register_move_cost[mode][hard_reg_class][rclass]
> -		       : ira_register_move_cost[mode][rclass][hard_reg_class])
> -		      * frequency);
> -	      op_costs[i]->cost[k] = cost;
> +	      cost = (i == 0
> +		      ? move_costs[hard_reg_class][rclass]
> +		      : move_costs[rclass][hard_reg_class]);
> +	      /* Target code might define wrong big costs for smaller
> +		 reg classes or reg classes containing only fixed hard
> +		 regs.  Try a bigger class.  */
> +	      if (bigger_hard_reg_class != hard_reg_class)
> +		{
> +		  int cost2 = (i == 0
> +			       ? move_costs[bigger_hard_reg_class][rclass]
> +			       : move_costs[rclass][bigger_hard_reg_class]);
> +		  if (cost2 < cost)
> +		    cost = cost2;
> +		}
> +	      
> +	      op_costs[i]->cost[k] = cost * frequency;
>  	      /* If we have assigned a class to this allocno in our
>  		 first pass, add a cost to this alternative
>  		 corresponding to what we would add if this allocno

This seems like a hack to me.  I don't see any reason why
FRAME_POINTER_REGNUM etc. have to be in GENERAL_REGS, so picking
that seems arbitrary.  And is there a reason not to include
ARG_POINTER_REGNUM or PIC_OFFSET_TABLE_REGNUM?

It sounds from the comments in the PR that the backend isn't defining
its costs correctly.  If so, what specifically does it get wrong?
I think it would be better to fix the backends than accumulate more
workarounds like this.

Thanks,
Richard