[backport from gcc-4.8/trunk r191363 <mikpe@it.uu.se>]

From: Richard Sandiford <rdsandiford at googlemail dot com>
Subject: Re: Finish up PR rtl-optimization/44194
Date: Sun, 16 Sep 2012 11:52:56 +0100
List-Archive: <http://gcc.gnu.org/ml/gcc-patches/>

Eric Botcazou <ebotcazou@adacore.com> writes:
> This is the PR about the useless spilling to memory of structures that are 
> returned in registers.  It was essentially addressed last year by Easwaran with 
> an enhancement of the RTL DSE pass, but Easwaran also noted that we still spill 
> to memory in the simplest cases, e.g. gcc.dg/pr44194-1.c, because expand_call
> creates a temporary on the stack to store the value returned in registers...
>
> The attached patch solves this problem by copying the value into pseudos 
> instead by means of emit_group_move_into_temps.  This is sufficient to get rid 
> of the remaining memory accesses for gcc.dg/pr44194-1.c on x86-64 for example,
> but not on strict-alignment platforms like SPARC64.

Thanks for doing this.  It'll obviously help n32 and n64 long doubles too.

I hit one problem building libgfortran for mips64-linux-gnu.
The calls.c change was:

> Index: calls.c
> ===================================================================
> --- calls.c	(revision 191198)
> +++ calls.c	(working copy)
> @@ -3272,16 +3272,8 @@ expand_call (tree exp, rtx target, int i
>        else if (GET_CODE (valreg) == PARALLEL)
>  	{
>  	  if (target == 0)
> -	    {
> -	      /* This will only be assigned once, so it can be readonly.  */
> -	      tree nt = build_qualified_type (rettype,
> -					      (TYPE_QUALS (rettype)
> -					       | TYPE_QUAL_CONST));
> -
> -	      target = assign_temp (nt, 1, 1);
> -	    }
> -
> -	  if (! rtx_equal_p (target, valreg))
> +	    target = emit_group_move_into_temps (valreg);
> +	  else if (!rtx_equal_p (target, valreg))
>  	    emit_group_store (target, valreg, rettype,
>  			      int_size_in_bytes (rettype));
>  
>  	  /* We can not support sibling calls for this case.  */
>  	  sibcall_failure = 1;

But if we're trying to use a sibcall, we go through this loop twice,
and the second iteration has to cope with a PARALLEL target created
by the first.  How about the patch below?  Tested on mips64-linux-gnu,
where a full testrun looked good.

In some ways it's a bit silly to emit anything at all in the first iteration,
given that we then go on to set sibcall_failure.  It's not the kind of loop
we can just continue out of though.  Also, your patch probably means that
we only need to set sibcall_failure for the emit_group_store case,
although I've not tested that.

Richard

gcc/

2012-09-16  Richard Sandiford  <rdsandiford@googlemail.com>

	* calls.c (expand_call): Use emit_group_move for PARALLEL->PARALLEL
	moves.

--- gcc-4.7.2/gcc/calls.c.~1~	2012-10-21 16:34:07.000000000 +0200
+++ gcc-4.7.2/gcc/calls.c	2012-10-21 16:34:54.000000000 +0200
@@ -3231,7 +3231,13 @@ expand_call (tree exp, rtx target, int i
 	{
 	  if (target == 0)
 	    target = emit_group_move_into_temps (valreg);
-	  else if (!rtx_equal_p (target, valreg))
+	  else if (rtx_equal_p (target, valreg))
+	    ;
+	  else if (GET_CODE (target) == PARALLEL)
+	    /* Handle the result of a emit_group_move_into_temps
+	       call in the previous pass.  */
+	    emit_group_move (target, valreg);
+	  else
 	    emit_group_store (target, valreg, rettype,
 			      int_size_in_bytes (rettype));