[PATCH] Avoid tree NRV on is_gimple_reg_type result types (PR c/35739)

[PATCH] Avoid tree NRV on is_gimple_reg_type result types (PR c/35739)

[Jakub Jelinek]

Hi!

If result is aggregate_value_p, but also a GIMPLE reg (e.g. when ABI
says it should be passed into memory), doing NRV optimization can result
in non-gimple - <result> isn't valid GIMPLE reg in that case, so NRV
can create non-GIMPLE like <result> = <result> + 3;
Also, in some cases it can hurt optimizations, as <result> must live
in memory wherever it will be replaced, but without NRV optimization
it is only stored at the end of function before returning to the <result>
memory slot.

Bootstrapped/regtested on x86_64-linux/trunk, bootstrap/regtest on 4
linux arches/4.3 is ongoing.

Ok for trunk/4.3?

2008-04-15  Jakub Jelinek  <jakub@redhat.com>

	PR c/35739
	* tree-nrv.c (tree_nrv): Don't optimize if result_type is GIMPLE
	reg type.

	* gcc.dg/dfp/pr35739.c: New test.

--- gcc/tree-nrv.c.jj	2008-04-04 15:12:00.000000000 +0200
+++ gcc/tree-nrv.c	2008-04-15 20:00:07.000000000 +0200
@@ -1,5 +1,5 @@
 /* Language independent return value optimizations
-   Copyright (C) 2004, 2005, 2007 Free Software Foundation, Inc.
+   Copyright (C) 2004, 2005, 2007, 2008 Free Software Foundation, Inc.
 
 This file is part of GCC.
 
@@ -115,6 +115,11 @@ tree_nrv (void)
   if (!aggregate_value_p (result, current_function_decl))
     return 0;
 
+  /* If a GIMPLE type is returned in memory, finalize_nrv_r might create
+     non-GIMPLE.  */
+  if (is_gimple_reg_type (result_type))
+    return 0;
+
   /* Look through each block for assignments to the RESULT_DECL.  */
   FOR_EACH_BB (bb)
     {
--- gcc/testsuite/gcc.dg/dfp/pr35739.c.jj	2008-04-15 20:02:55.000000000 +0200
+++ gcc/testsuite/gcc.dg/dfp/pr35739.c	2008-04-15 20:18:15.000000000 +0200
@@ -0,0 +1,16 @@
+/* PR c/35739 */
+/* { dg-do compile { target *-*-linux* } } */
+/* { dg-options "-O -fpreprocessed -fmudflap" } */
+
+_Decimal128
+foo (int n, ...)
+{
+  int i;
+  _Decimal128 j = 0;
+  __builtin_va_list ap;
+  __builtin_va_start (ap, n);
+  for (i = 0; i < n; i++)
+    j += __builtin_va_arg (ap, _Decimal128);
+  __builtin_va_end (ap);
+  return j;
+}

	Jakub

Re: [PATCH] Avoid tree NRV on is_gimple_reg_type result types (PR c/35739)

[Richard Guenther]

On Wed, Apr 16, 2008 at 12:59 PM, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!
>
>  If result is aggregate_value_p, but also a GIMPLE reg (e.g. when ABI
>  says it should be passed into memory), doing NRV optimization can result
>  in non-gimple - <result> isn't valid GIMPLE reg in that case, so NRV
>  can create non-GIMPLE like <result> = <result> + 3;
>  Also, in some cases it can hurt optimizations, as <result> must live
>  in memory wherever it will be replaced, but without NRV optimization
>  it is only stored at the end of function before returning to the <result>
>  memory slot.
>
>  Bootstrapped/regtested on x86_64-linux/trunk, bootstrap/regtest on 4
>  linux arches/4.3 is ongoing.
>
>  Ok for trunk/4.3?

Ok.

Thanks,
Richard.

>  2008-04-15  Jakub Jelinek  <jakub@redhat.com>
>
>         PR c/35739
>         * tree-nrv.c (tree_nrv): Don't optimize if result_type is GIMPLE
>         reg type.
>
>         * gcc.dg/dfp/pr35739.c: New test.
>
>  --- gcc/tree-nrv.c.jj   2008-04-04 15:12:00.000000000 +0200
>  +++ gcc/tree-nrv.c      2008-04-15 20:00:07.000000000 +0200
>  @@ -1,5 +1,5 @@
>   /* Language independent return value optimizations
>  -   Copyright (C) 2004, 2005, 2007 Free Software Foundation, Inc.
>  +   Copyright (C) 2004, 2005, 2007, 2008 Free Software Foundation, Inc.
>
>   This file is part of GCC.
>
>  @@ -115,6 +115,11 @@ tree_nrv (void)
>    if (!aggregate_value_p (result, current_function_decl))
>      return 0;
>
>  +  /* If a GIMPLE type is returned in memory, finalize_nrv_r might create
>  +     non-GIMPLE.  */
>  +  if (is_gimple_reg_type (result_type))
>  +    return 0;
>  +
>    /* Look through each block for assignments to the RESULT_DECL.  */
>    FOR_EACH_BB (bb)
>      {
>  --- gcc/testsuite/gcc.dg/dfp/pr35739.c.jj       2008-04-15 20:02:55.000000000 +0200
>  +++ gcc/testsuite/gcc.dg/dfp/pr35739.c  2008-04-15 20:18:15.000000000 +0200
>  @@ -0,0 +1,16 @@
>  +/* PR c/35739 */
>  +/* { dg-do compile { target *-*-linux* } } */
>  +/* { dg-options "-O -fpreprocessed -fmudflap" } */
>  +
>  +_Decimal128
>  +foo (int n, ...)
>  +{
>  +  int i;
>  +  _Decimal128 j = 0;
>  +  __builtin_va_list ap;
>  +  __builtin_va_start (ap, n);
>  +  for (i = 0; i < n; i++)
>  +    j += __builtin_va_arg (ap, _Decimal128);
>  +  __builtin_va_end (ap);
>  +  return j;
>  +}
>
>         Jakub
>