[5/9] ifcvt: Allow constants operands in noce_convert_multiple_sets.

This patch checks allows immediate then/else operands for cmovs.
We rely on,emit_conditional_move returning NULL if something unsupported
was generated.

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 | 21 +++++++++++----------
 1 file changed, 11 insertions(+), 10 deletions(-)

Robin Dapp <rdapp@linux.ibm.com> writes:
> This patch checks allows immediate then/else operands for cmovs.
> We rely on,emit_conditional_move returning NULL if something unsupported
> was generated.

It seems like this is making noce_convert_multiple_sets overlap
a lot with cond_move_process_if_block (although that uses CONSTANT_P
instead of CONST_INT_P).  How do they fit together after this patch,
i.e. which cases is each one meant to handle that the other doesn't?

Thanks,
Richard

> 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 | 21 +++++++++++----------
>  1 file changed, 11 insertions(+), 10 deletions(-)
>
> diff --git a/gcc/ifcvt.c b/gcc/ifcvt.c
> index 55205cac153..99716e5f63c 100644
> --- a/gcc/ifcvt.c
> +++ b/gcc/ifcvt.c
> @@ -3214,7 +3214,9 @@ noce_convert_multiple_sets (struct noce_if_info *if_info)
>  	 we'll end up trying to emit r4:HI = cond ? (r1:SI) : (r3:HI).
>  	 Wrap the two cmove operands into subregs if appropriate to prevent
>  	 that.  */
> -      if (GET_MODE (new_val) != GET_MODE (temp))
> +
> +      if (!CONST_INT_P (new_val)
> +         && GET_MODE (new_val) != GET_MODE (temp))
>  	{
>  	  machine_mode src_mode = GET_MODE (new_val);
>  	  machine_mode dst_mode = GET_MODE (temp);
> @@ -3225,7 +3227,8 @@ noce_convert_multiple_sets (struct noce_if_info *if_info)
>  	    }
>  	  new_val = lowpart_subreg (dst_mode, new_val, src_mode);
>  	}
> -      if (GET_MODE (old_val) != GET_MODE (temp))
> +      if (!CONST_INT_P (old_val)
> +         && GET_MODE (old_val) != GET_MODE (temp))
>  	{
>  	  machine_mode src_mode = GET_MODE (old_val);
>  	  machine_mode dst_mode = GET_MODE (temp);
> @@ -3362,9 +3365,9 @@ bb_ok_for_noce_convert_multiple_sets (basic_block test_bb, unsigned *cost)
>        if (!REG_P (dest))
>  	return false;
>  
> -      if (!(REG_P (src)
> -	   || (GET_CODE (src) == SUBREG && REG_P (SUBREG_REG (src))
> -	       && subreg_lowpart_p (src))))
> +      if (!((REG_P (src) || (CONST_INT_P (src)))
> +	    || (GET_CODE (src) == SUBREG && REG_P (SUBREG_REG (src))
> +	      && subreg_lowpart_p (src))))
>  	return false;
>  
>        /* Destination must be appropriate for a conditional write.  */
> @@ -3724,7 +3727,7 @@ 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)
> @@ -3740,10 +3743,8 @@ 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

> It seems like this is making noce_convert_multiple_sets overlap
> a lot with cond_move_process_if_block (although that uses CONSTANT_P
> instead of CONST_INT_P).  How do they fit together after this patch,
> i.e. which cases is each one meant to handle that the other doesn't?

IMHO all of icvt is currently one unholy overlap with itself and the
backend where at least I could not decipher within reasonable time which
part is supposed to cover what :) Ok, it might not be that bad...

Currently convert_multiple only handles situations without an else_bb
but that seems a rather artificial limit and I see no reason why it
could not be unified with cond_move_process_if_block at a later time.

Regards
 Robin

Coming back to this just in time for it not to be three months later,
sorry...

I still think it would be better to consolidate ifcvt a bit more,
rather than effectively duplicate bits of cond_move_process_if_block
in noce_convert_multiple_sets.  But perhaps it was a historical
mistake to have two separate routines in the first place, and I guess
making noce_convert_multiple_sets handle more cases might allow us
to get rid of cond_move_process_if_block at some point.  So let's
go ahead with this anyway.

Robin Dapp <rdapp@linux.ibm.com> writes:
> This patch checks allows immediate then/else operands for cmovs.
> We rely on,emit_conditional_move returning NULL if something unsupported
> was generated.
>
> 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 | 21 +++++++++++----------
>  1 file changed, 11 insertions(+), 10 deletions(-)
>
> diff --git a/gcc/ifcvt.c b/gcc/ifcvt.c
> index 55205cac153..99716e5f63c 100644
> --- a/gcc/ifcvt.c
> +++ b/gcc/ifcvt.c
> @@ -3214,7 +3214,9 @@ noce_convert_multiple_sets (struct noce_if_info *if_info)
>  	 we'll end up trying to emit r4:HI = cond ? (r1:SI) : (r3:HI).
>  	 Wrap the two cmove operands into subregs if appropriate to prevent
>  	 that.  */
> -      if (GET_MODE (new_val) != GET_MODE (temp))
> +
> +      if (!CONST_INT_P (new_val)
> +         && GET_MODE (new_val) != GET_MODE (temp))
>  	{
>  	  machine_mode src_mode = GET_MODE (new_val);
>  	  machine_mode dst_mode = GET_MODE (temp);
> @@ -3225,7 +3227,8 @@ noce_convert_multiple_sets (struct noce_if_info *if_info)
>  	    }
>  	  new_val = lowpart_subreg (dst_mode, new_val, src_mode);
>  	}
> -      if (GET_MODE (old_val) != GET_MODE (temp))
> +      if (!CONST_INT_P (old_val)
> +         && GET_MODE (old_val) != GET_MODE (temp))
>  	{
>  	  machine_mode src_mode = GET_MODE (old_val);
>  	  machine_mode dst_mode = GET_MODE (temp);
> @@ -3362,9 +3365,9 @@ bb_ok_for_noce_convert_multiple_sets (basic_block test_bb, unsigned *cost)
>        if (!REG_P (dest))
>  	return false;
>  
> -      if (!(REG_P (src)
> -	   || (GET_CODE (src) == SUBREG && REG_P (SUBREG_REG (src))
> -	       && subreg_lowpart_p (src))))
> +      if (!((REG_P (src) || (CONST_INT_P (src)))
> +	    || (GET_CODE (src) == SUBREG && REG_P (SUBREG_REG (src))
> +	      && subreg_lowpart_p (src))))
>  	return false;
>  
>        /* Destination must be appropriate for a conditional write.  */

CONSTANT_P (as for check_cond_move_block) would cover more cases than
CONST_INT_P.  No need for brackets around CONST_INT_P (...).

I was tempted to say we should use register_operand too, but I guess
that allows unwanted (subreg (mem...))s (as if there's any other kind).

> @@ -3724,7 +3727,7 @@ 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)

I assume this is to skip notes, but shouldn't this still include JUMP_P?

It also looks like it should be an independent patch.

> @@ -3740,10 +3743,8 @@ 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

Seems wrong to use something called noce_operand_ok in a routine
that is explicitly handling cases in which conditional execution
is possible.

More importantly, I can't see anywhere that handles the MEM_P case
specially for cond_move_*, so are you sure this is really safe?
Might be better to leave it as-is if this part is just a clean-up.

Thanks,
Richard