Issue 6255070: [google/gcc-4_6] Backport r171031 from upstream to fix ICE on PowerPC64

Can't Edit
Can't Publish+Mail
Start Review

Created:
12 years, 1 month ago by dougkwan

Modified:
12 years, 1 month ago

Reviewers:
Diego Novillo

CC:
gcc-patches_gcc.gnu.org

Base URL:
svn+ssh://gcc.gnu.org/svn/gcc/branches/google/gcc-4_6/

Visibility:
Public.

Patch Set 1 #

Created: 12 years, 1 month ago

Download [raw] [tar.bz2]

	Unified diffs	Side-by-side diffs	Stats (+27 lines, -4 lines)			Patch
M	contrib/testsuite-management/powerpc64-grtev2-linux-gnu.xfail	View	1 chunk	+0 lines, -2 lines	0 comments	Download
M	gcc/config/rs6000/predicates.md	View	2 chunks	+18 lines, -1 line	0 comments	Download
M	gcc/config/rs6000/rs6000.c	View	1 chunk	+8 lines, -0 lines	0 comments	Download
M	gcc/config/rs6000/vsx.md	View	1 chunk	+1 line, -1 line	0 comments	Download

Messages

Total messages: 3

Expand All Messages | Collapse All Messages

dougkwan

Hi Diego, This is a backport of this patch to fix an ICE on PowerPC64. ...

12 years, 1 month ago (2012-05-31 23:02:37 UTC) #1

Hi Diego,

	This is a backport of this patch to fix an ICE on PowerPC64.

	http://gcc.gnu.org/ml/gcc-patches/2011-03/msg00829.html

-Doug


2012-05-31   Doug Kwan  <dougkwan@google.com>

	Backport r171031 from upstream trunk in gcc/.
	2011-03-16  Alan Modra  <amodra@gmail.com>

		PR target/45844
		* config/rs6000/rs6000.c (rs6000_legitimize_reload_address):
		Don't create invalid offset address for vsx splat insn.
		* config/rs6000/predicates.md (splat_input_operand): New.
		* config/rs6000/vsx.md (vsx_splat_*): Use it.

	* contrib/testsuite-management/powerpc64-grtev2-linux-gnu.xfail:
	Remove expected failures because now the bug is fixed by the above.

Index: contrib/testsuite-management/powerpc64-grtev2-linux-gnu.xfail
===================================================================
--- contrib/testsuite-management/powerpc64-grtev2-linux-gnu.xfail	(revision
188083)
+++ contrib/testsuite-management/powerpc64-grtev2-linux-gnu.xfail	(working
copy)
@@ -69,8 +69,6 @@ FAIL: gcc.target/powerpc/pr47755-2.c execution tes
 
 # *** gfortran:
 XPASS: gfortran.dg/nint_2.f90  -O0  execution test
-FAIL: gfortran.dg/vect/pr45714-b.f  -O  (internal compiler error)
-FAIL: gfortran.dg/vect/pr45714-b.f  -O  (test for excess errors)
 FAIL: gfortran.dg/vect/fast-math-pr38968.f90 execution test
 
 # *** g++:
Index: gcc/config/rs6000/predicates.md
===================================================================
--- gcc/config/rs6000/predicates.md	(revision 188083)
+++ gcc/config/rs6000/predicates.md	(working copy)
@@ -1,5 +1,5 @@
 ;; Predicate definitions for POWER and PowerPC.
-;; Copyright (C) 2005, 2006, 2007, 2008, 2009, 2010
+;; Copyright (C) 2005, 2006, 2007, 2008, 2009, 2010, 2011
 ;; Free Software Foundation, Inc.
 ;;
 ;; This file is part of GCC.
@@ -871,6 +871,23 @@
   return 0;
 })
 
+;; Return 1 if this operand is a valid input for a vsx_splat insn.
+(define_predicate "splat_input_operand"
+  (match_code "label_ref,symbol_ref,const,high,reg,subreg,mem,
+	       const_double,const_vector,const_int,plus")
+{
+  if (MEM_P (op))
+    {
+      if (mode == DFmode)
+	mode = V2DFmode;
+      else if (mode == DImode)
+	mode = V2DImode;
+      else
+	gcc_unreachable ();        
+    }
+  return input_operand (op, mode);
+})
+
 ;; Return true if OP is an invalid SUBREG operation on the e500.
 (define_predicate "rs6000_nonimmediate_operand"
   (match_code "reg,subreg,mem")
Index: gcc/config/rs6000/rs6000.c
===================================================================
--- gcc/config/rs6000/rs6000.c	(revision 188083)
+++ gcc/config/rs6000/rs6000.c	(working copy)
@@ -6703,6 +6703,14 @@ rs6000_legitimize_reload_address (rtx x, enum mach
 {
   bool reg_offset_p = reg_offset_addressing_ok_p (mode);
 
+  /* Nasty hack for vsx_splat_V2DF/V2DI load from mem, which takes a
+     DFmode/DImode MEM.  */
+  if (reg_offset_p
+      && opnum == 1
+      && ((mode == DFmode && recog_data.operand_mode[0] == V2DFmode)
+	  || (mode == DImode && recog_data.operand_mode[0] == V2DImode)))
+    reg_offset_p = false;
+
   /* We must recognize output that we have already generated ourselves.  */
   if (GET_CODE (x) == PLUS
       && GET_CODE (XEXP (x, 0)) == PLUS
Index: gcc/config/rs6000/vsx.md
===================================================================
--- gcc/config/rs6000/vsx.md	(revision 188083)
+++ gcc/config/rs6000/vsx.md	(working copy)
@@ -1076,7 +1076,7 @@
 (define_insn "vsx_splat_<mode>"
   [(set (match_operand:VSX_D 0 "vsx_register_operand" "=wd,wd,wd,?wa,?wa,?wa")
 	(vec_duplicate:VSX_D
-	 (match_operand:<VS_scalar> 1 "input_operand" "ws,f,Z,wa,wa,Z")))]
+	 (match_operand:<VS_scalar> 1 "splat_input_operand" "ws,f,Z,wa,wa,Z")))]
   "VECTOR_MEM_VSX_P (<MODE>mode)"
   "@
    xxpermdi %x0,%x1,%x1,0

--
This patch is available for review at http://codereview.appspot.com/6255070

dougkwan

ping? On Thu, May 31, 2012 at 4:02 PM, Doug Kwan <dougkwan@google.com> wrote: > Hi ...

12 years, 1 month ago (2012-06-01 19:53:20 UTC) #2

ping?

On Thu, May 31, 2012 at 4:02 PM, Doug Kwan <dougkwan@google.com> wrote:
> Hi Diego,
>
>        This is a backport of this patch to fix an ICE on PowerPC64.
>
>        http://gcc.gnu.org/ml/gcc-patches/2011-03/msg00829.html
>
> -Doug
>
>
> 2012-05-31   Doug Kwan  <dougkwan@google.com>
>
>        Backport r171031 from upstream trunk in gcc/.
>        2011-03-16  Alan Modra  <amodra@gmail.com>
>
>                PR target/45844
>                * config/rs6000/rs6000.c (rs6000_legitimize_reload_address):
>                Don't create invalid offset address for vsx splat insn.
>                * config/rs6000/predicates.md (splat_input_operand): New.
>                * config/rs6000/vsx.md (vsx_splat_*): Use it.
>
>        * contrib/testsuite-management/powerpc64-grtev2-linux-gnu.xfail:
>        Remove expected failures because now the bug is fixed by the above.
>
> Index: contrib/testsuite-management/powerpc64-grtev2-linux-gnu.xfail
> ===================================================================
> --- contrib/testsuite-management/powerpc64-grtev2-linux-gnu.xfail      
(revision 188083)
> +++ contrib/testsuite-management/powerpc64-grtev2-linux-gnu.xfail      
(working copy)
> @@ -69,8 +69,6 @@ FAIL: gcc.target/powerpc/pr47755-2.c execution tes
>
>  # *** gfortran:
>  XPASS: gfortran.dg/nint_2.f90  -O0  execution test
> -FAIL: gfortran.dg/vect/pr45714-b.f  -O  (internal compiler error)
> -FAIL: gfortran.dg/vect/pr45714-b.f  -O  (test for excess errors)
>  FAIL: gfortran.dg/vect/fast-math-pr38968.f90 execution test
>
>  # *** g++:
> Index: gcc/config/rs6000/predicates.md
> ===================================================================
> --- gcc/config/rs6000/predicates.md     (revision 188083)
> +++ gcc/config/rs6000/predicates.md     (working copy)
> @@ -1,5 +1,5 @@
>  ;; Predicate definitions for POWER and PowerPC.
> -;; Copyright (C) 2005, 2006, 2007, 2008, 2009, 2010
> +;; Copyright (C) 2005, 2006, 2007, 2008, 2009, 2010, 2011
>  ;; Free Software Foundation, Inc.
>  ;;
>  ;; This file is part of GCC.
> @@ -871,6 +871,23 @@
>   return 0;
>  })
>
> +;; Return 1 if this operand is a valid input for a vsx_splat insn.
> +(define_predicate "splat_input_operand"
> +  (match_code "label_ref,symbol_ref,const,high,reg,subreg,mem,
> +              const_double,const_vector,const_int,plus")
> +{
> +  if (MEM_P (op))
> +    {
> +      if (mode == DFmode)
> +       mode = V2DFmode;
> +      else if (mode == DImode)
> +       mode = V2DImode;
> +      else
> +       gcc_unreachable ();
> +    }
> +  return input_operand (op, mode);
> +})
> +
>  ;; Return true if OP is an invalid SUBREG operation on the e500.
>  (define_predicate "rs6000_nonimmediate_operand"
>   (match_code "reg,subreg,mem")
> Index: gcc/config/rs6000/rs6000.c
> ===================================================================
> --- gcc/config/rs6000/rs6000.c  (revision 188083)
> +++ gcc/config/rs6000/rs6000.c  (working copy)
> @@ -6703,6 +6703,14 @@ rs6000_legitimize_reload_address (rtx x, enum mach
>  {
>   bool reg_offset_p = reg_offset_addressing_ok_p (mode);
>
> +  /* Nasty hack for vsx_splat_V2DF/V2DI load from mem, which takes a
> +     DFmode/DImode MEM.  */
> +  if (reg_offset_p
> +      && opnum == 1
> +      && ((mode == DFmode && recog_data.operand_mode[0] == V2DFmode)
> +         || (mode == DImode && recog_data.operand_mode[0] == V2DImode)))
> +    reg_offset_p = false;
> +
>   /* We must recognize output that we have already generated ourselves.  */
>   if (GET_CODE (x) == PLUS
>       && GET_CODE (XEXP (x, 0)) == PLUS
> Index: gcc/config/rs6000/vsx.md
> ===================================================================
> --- gcc/config/rs6000/vsx.md    (revision 188083)
> +++ gcc/config/rs6000/vsx.md    (working copy)
> @@ -1076,7 +1076,7 @@
>  (define_insn "vsx_splat_<mode>"
>   [(set (match_operand:VSX_D 0 "vsx_register_operand"
"=wd,wd,wd,?wa,?wa,?wa")
>        (vec_duplicate:VSX_D
> -        (match_operand:<VS_scalar> 1 "input_operand" "ws,f,Z,wa,wa,Z")))]
> +        (match_operand:<VS_scalar> 1 "splat_input_operand"
"ws,f,Z,wa,wa,Z")))]
>   "VECTOR_MEM_VSX_P (<MODE>mode)"
>   "@
>    xxpermdi %x0,%x1,%x1,0
>
> --
> This patch is available for review at http://codereview.appspot.com/6255070

Diego Novillo

12 years, 1 month ago (2012-06-01 20:00:28 UTC) #3

On 12-05-31 19:02 , Doug Kwan wrote:

> 2012-05-31   Doug Kwan<dougkwan@google.com>
>
> 	Backport r171031 from upstream trunk in gcc/.
> 	2011-03-16  Alan Modra<amodra@gmail.com>
>
> 		PR target/45844
> 		* config/rs6000/rs6000.c (rs6000_legitimize_reload_address):
> 		Don't create invalid offset address for vsx splat insn.
> 		* config/rs6000/predicates.md (splat_input_operand): New.
> 		* config/rs6000/vsx.md (vsx_splat_*): Use it.
>
> 	* contrib/testsuite-management/powerpc64-grtev2-linux-gnu.xfail:
> 	Remove expected failures because now the bug is fixed by the above.

OK.


Diego.

Expand All Messages | Collapse All Messages