Re: [PATCH 2/6] ifcvt: Allow constants operands in noce_convert_multiple_sets.

[Jeff Law <law at redhat dot com>]

On 11/14/18 6:07 AM, Robin Dapp wrote:
> This patch checks whether the current target supports conditional moves
> with immediate then/else operands and allows noce_convert_multiple_sets
> to deal with constants subsequently.
> Also, minor refactoring is performed.
> 
> --
> 
> gcc/ChangeLog:
> 
> 2018-11-14  Robin Dapp  <rdapp@linux.ibm.com>
> 
> 	* ifcvt.c (have_const_cmov): New function.
> 	(noce_convert_multiple_sets): Allow constants if supported.
> 	(bb_ok_for_noce_convert_multiple_sets): Likewise.
> 	(check_cond_move_block): Refactor.
> ---
>  gcc/ifcvt.c | 46 ++++++++++++++++++++++++++++++++++++----------
>  1 file changed, 36 insertions(+), 10 deletions(-)
> 
> diff --git a/gcc/ifcvt.c b/gcc/ifcvt.c
> index ddf077fa051..660bb46eb1c 100644
> --- a/gcc/ifcvt.c
> +++ b/gcc/ifcvt.c
> @@ -3077,6 +3077,27 @@ bb_valid_for_noce_process_p (basic_block test_bb, rtx cond,
>    return false;
>  }
>  
> +/* Check if we have a movcc pattern that accepts constants as then/else
> +   operand (op 2/3).  */
> +static bool
> +have_const_cmov (machine_mode mode)
> +{
> +  enum insn_code icode;
> +  if ((icode = direct_optab_handler (movcc_optab, mode))
> +      != CODE_FOR_nothing)
> +    {
> +      if (insn_data[(int) icode].operand[2].predicate
> +	  && (insn_data[(int) icode].operand[2].predicate
> +	    (const1_rtx, insn_data[(int) icode].operand[2].mode)))
> +	if (insn_data[(int) icode].operand[3].predicate
> +	    && (insn_data[(int) icode].operand[3].predicate
> +	      (const1_rtx, insn_data[(int) icode].operand[3].mode)))
> +	  return true;
> +    }
> +
> +  return false;
> +}
This may ultimately be too simplistic.  There are targets where some
constants are OK, but others may not be.   By checking the predicate
like this I think you can cause over-aggressive if-conversion if the
target allows a range of integers in the expander's operand predicate,
but allows a narrower range in the actual define_insn (presumably the
expander loads them into a pseudo or somesuch in that case).

We know that over-aggressive if-conversion into conditional moves hurts
some targets.

Ideally you'd create the actual insn with the constants you want to use
and see if that's recognized as well as compute its cost.  Is that just
too painful at this point for some reason?


> @@ -3689,7 +3717,7 @@ check_cond_move_block (basic_block bb,
>      {
>        rtx set, dest, src;
>  
> -      if (!NONDEBUG_INSN_P (insn) || JUMP_P (insn))
> +      if (!active_insn_p (insn))
>  	continue;
>        set = single_set (insn);
>        if (!set)
> @@ -3705,10 +3733,8 @@ check_cond_move_block (basic_block bb,
>        if (!CONSTANT_P (src) && !register_operand (src, VOIDmode))
>  	return FALSE;
>  
> -      if (side_effects_p (src) || side_effects_p (dest))
> -	return FALSE;
> -
> -      if (may_trap_p (src) || may_trap_p (dest))
> +      /* Check for side effects and trapping.  */
> +      if (!noce_operand_ok (src) || !noce_operand_ok (dest))
>  	return FALSE;
>  
>        /* Don't try to handle this if the source register was
These two hunks are probably OK as general cleanups.  Note that
noce_operand_ok is not strictly the same as checking side_effects_p and
may_trap_p in the case of a MEM.  But you've already filtered out MEMs
before you get here.

Jeff
>