Re: [PATCH] lower-bitint: Force some arrays corresponding to large/huge _BitInt SSA_NAMEs to BLKmode

[Richard Biener <rguenther@suse.de>]

On Thu, 18 Jan 2024, Jakub Jelinek wrote:

> On Thu, Jan 18, 2024 at 01:34:53PM +0100, Richard Biener wrote:
> > So - if we simply do
> > 
> >       /* If the input and output modes are both the same, we are done.  */
> >       if (mode == GET_MODE (op0) || (mode == BLKmode && MEM_P (op0))
> >         ;
> > 
> > ?  After all if we want BLKmode we want a MEM, and if we have one
> > we should be done already.  V_C_E isn't supposed to do any value
> > transform.
> 
> We'd need to make sure that op0 is actually MEM, which is not the case.
> 
> The following patch seems to work though, the discover_nonconstant_array_refs_r
> part helps for -O2, the expr.cc part is needed at all optimization levels.
> 
> 2024-01-18  Jakub Jelinek  <jakub@redhat.com>
> 
> 	* cfgexpand.cc (discover_nonconstant_array_refs_r): Force non-BLKmode
> 	VAR_DECLs referenced in BLKmode VIEW_CONVERT_EXPRs into memory.
> 	* expr.cc (expand_expr_real_1) <case VIEW_CONVERT_EXPR>: If mode is
> 	BLKmode and type a BITINT_TYPE from a non-BLKmode VAR_P with MEM
> 	DECL_RTL, use adjust_address to BLKmode for it instead of expand_expr.
> 
> --- gcc/cfgexpand.cc.jj	2024-01-16 11:45:16.159326506 +0100
> +++ gcc/cfgexpand.cc	2024-01-18 13:41:54.579551282 +0100
> @@ -6380,11 +6380,16 @@ discover_nonconstant_array_refs_r (tree
>    /* References of size POLY_INT_CST to a fixed-size object must go
>       through memory.  It's more efficient to force that here than
>       to create temporary slots on the fly.
> -     RTL expansion expectes TARGET_MEM_REF to always address actual memory.  */
> +     RTL expansion expectes TARGET_MEM_REF to always address actual memory.
> +     Also, force to stack non-BLKmode vars accessed through VIEW_CONVERT_EXPR
> +     to BLKmode BITINT_TYPEs.  */
>    else if (TREE_CODE (t) == TARGET_MEM_REF
>  	   || (TREE_CODE (t) == MEM_REF
>  	       && TYPE_SIZE (TREE_TYPE (t))
> -	       && POLY_INT_CST_P (TYPE_SIZE (TREE_TYPE (t)))))
> +	       && POLY_INT_CST_P (TYPE_SIZE (TREE_TYPE (t))))
> +	   || (TREE_CODE (t) == VIEW_CONVERT_EXPR
> +	       && TREE_CODE (TREE_TYPE (t)) == BITINT_TYPE
> +	       && TYPE_MODE (TREE_TYPE (t)) == BLKmode))

I'm still not getting what's special about BITINT_TYPE here so
shouldn't that apply to all BLKmode V_C_E?  But sure we can for
now just handle BITINT_TYPE.

That hunk looks OK to me.

>      {
>        tree base = get_base_address (t);
>        if (base
> --- gcc/expr.cc.jj	2024-01-12 10:07:58.194851657 +0100
> +++ gcc/expr.cc	2024-01-18 13:38:19.677556646 +0100
> @@ -12382,6 +12382,17 @@ expand_expr_real_1 (tree exp, rtx target
>  	  }
>        }
>  
> +      /* Ensure non-BLKmode array VAR_DECLs VCEd to BLKmode BITINT_TYPE
> +	 aren't promoted to registers.  */
> +      if (op0 == NULL_RTX
> +	  && mode == BLKmode
> +	  && TREE_CODE (type) == BITINT_TYPE
> +	  && VAR_P (treeop0)
> +	  && DECL_MODE (treeop0) != BLKmode
> +	  && DECL_RTL_SET_P (treeop0)
> +	  && MEM_P (DECL_RTL (treeop0)))
> +	op0 = adjust_address (DECL_RTL (treeop0), BLKmode, 0);
> +
>        if (!op0)
>  	op0 = expand_expr_real (treeop0, NULL_RTX, VOIDmode, modifier,
>  				NULL, inner_reference_p);

So we're now sure we have MEM_P (op0) after expand_expr_real,
even without this change, right?  What's wrong with the
suggestion to use

  if (mode == GET_MODE (op0) || (mode == BLKmode && MEM_P (op0))

thus not run into any of the special-casing?  We're doing just

          op0 = adjust_address (op0, mode, 0);

then which is basically what you do above?

> 
> 	Jakub
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)