Re: [OpenACC] internal fn folding

[Thomas Schwinge <thomas at codesourcery dot com>]

Hi!

On Thu, 5 Nov 2015 10:48:02 -0500, Nathan Sidwell <nathan@acm.org> wrote:
> On 11/04/15 05:02, Bernd Schmidt wrote:
> > On 11/02/2015 02:56 PM, Nathan Sidwell wrote:
> >> On 10/28/15 14:40, Nathan Sidwell wrote:
> >>> Richard,
> >>> this patch adds folding for the new GOACC_DIM_POS and GOACC_DIM_SIZE
> >>> internal
> >>> functions.  IIUC gimple_fold_call is the right place to add this.
> >>>
> >>> The size of a compute dimension is very often a compile-time
> >>> constant.  On the
> >>> host, in particular it's 1, which means we can deduce the POS must be
> >>> zero.

> This is what I committed, using the helpers I recently added. (I realized we can 
> only get here for functions with the oacc attribute already set)

> --- gimple-fold.c	(revision 229809)
> +++ gimple-fold.c	(working copy)

> +/* Transform IFN_GOACC_DIM_SIZE and IFN_GOACC_DIM_POS internal
> +   function calls to constants, where possible.  */
> +
> +static tree
> +fold_internal_goacc_dim (const gimple *call)
> +{
> +  int axis = get_oacc_ifn_dim_arg (call);
> +  int size = get_oacc_fn_dim_size (current_function_decl, axis);
> +  bool is_pos = gimple_call_internal_fn (call) == IFN_GOACC_DIM_POS;
> +  tree result = NULL_TREE;
> +
> +  /* If the size is 1, or we only want the size and it is not dynamic,
> +     we know the answer.  */
> +  if (size == 1 || (!is_pos && size))
> +    {
> +      tree type = TREE_TYPE (gimple_call_lhs (call));
> +      result = build_int_cst (type, size - is_pos);
> +    }
> +
> +  return result;
> +}

> @@ -3106,6 +3129,10 @@ gimple_fold_call (gimple_stmt_iterator *
>  	      return true;
>  	    }
>  	  break;
> +	case IFN_GOACC_DIM_SIZE:
> +	case IFN_GOACC_DIM_POS:
> +	  result = fold_internal_goacc_dim (stmt);
> +	  break;

Merging this into gomp-4_0-branch, we'd run into a lot of regressions
(for OpenACC kernels construct), for example:

    [...]/source-gcc/gcc/testsuite/c-c++-common/goacc/kernels-loop-g.c: In function 'main._omp_fn.0':
    [...]/source-gcc/gcc/testsuite/c-c++-common/goacc/kernels-loop-g.c:8:0: internal compiler error: Segmentation fault
    0xac387f crash_signal
            [...]/source-gcc/gcc/toplev.c:336
    0x9b8399 tree_int_cst_elt_check
            [...]/source-gcc/gcc/tree.h:3129
    0x9b8399 get_oacc_fn_dim_size(tree_node*, int)
            [...]/source-gcc/gcc/omp-low.c:12630
    0x86c530 fold_internal_goacc_dim
            [...]/source-gcc/gcc/gimple-fold.c:2917
    0x86c530 gimple_fold_call
            [...]/source-gcc/gcc/gimple-fold.c:3134
    0x86dfe4 fold_stmt_1
            [...]/source-gcc/gcc/gimple-fold.c:3702
    0xbf9953 execute
            [...]/source-gcc/gcc/tree-ssa-forwprop.c:2310

The dims in gcc/omp-low.c:get_oacc_fn_dim_size don't have values set, so
the "TREE_INT_CST_LOW (TREE_VALUE (dims))" fails.  I have not analyzed
what exactly is going wrong; I just figured out that it's related to the
IFN_GOACC_DIM_POS without LHS usage that Tom introduced in
gomp-4_0-branch r228735 for OpenACC kernels to "neuter gang-single code
in gang-redundant mode",
<http://news.gmane.org/find-root.php?message_id=%3C561C1336.2050401%40mentor.com%3E>.

So, in r229948 I merged Nathan's trunk r229816 into gomp-4_0-branch with
an additional hack as indicated ("++" prefix) by the following three-way
diff:

commit b421a6415fc223866bc97f8248a1fbd0a524505e
Merge: 7a6eb6b b0ccb4e
Author: tschwinge <tschwinge@138bc75d-0d04-0410-961f-82ee72b054a4>
Date:   Sun Nov 8 13:49:46 2015 +0000

    svn merge -r 229814:229816 svn+ssh://gcc.gnu.org/svn/gcc/trunk
    
    
    git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/branches/gomp-4_0-branch@229948 138bc75d-0d04-0410-961f-82ee72b054a4

 gcc/ChangeLog     |  6 ++++++
 gcc/gimple-fold.c | 33 +++++++++++++++++++++++++++++++++
 2 files changed, 39 insertions(+)

diff --cc gcc/gimple-fold.c
index c9b9593,45840af..869c6c2
--- gcc/gimple-fold.c
+++ gcc/gimple-fold.c
@@@ -2906,6 -2907,28 +2907,34 @@@ gimple_fold_builtin (gimple_stmt_iterat
    return false;
  }
  
+ /* Transform IFN_GOACC_DIM_SIZE and IFN_GOACC_DIM_POS internal
+    function calls to constants, where possible.  */
+ 
+ static tree
+ fold_internal_goacc_dim (const gimple *call)
+ {
++  /* TODO.  There is something going wrong here, for the gang_single
++     IFN_GOACC_DIM_POS without LHS, generated in gcc/omp-low.c:lower_omp_target
++     for is_oacc_kernels (see gomp-4_0-branch r228735).  */
++  if (gimple_call_lhs (call) == NULL_TREE)
++    return NULL_TREE;
++
+   int axis = get_oacc_ifn_dim_arg (call);
+   int size = get_oacc_fn_dim_size (current_function_decl, axis);
+   bool is_pos = gimple_call_internal_fn (call) == IFN_GOACC_DIM_POS;
+   tree result = NULL_TREE;
+ 
+   /* If the size is 1, or we only want the size and it is not dynamic,
+      we know the answer.  */
+   if (size == 1 || (!is_pos && size))
+     {
+       tree type = TREE_TYPE (gimple_call_lhs (call));
+       result = build_int_cst (type, size - is_pos);
+     }
+ 
+   return result;
+ }
+ 
  /* Return true if ARG0 CODE ARG1 in infinite signed precision operation
     doesn't fit into TYPE.  The test for overflow should be regardless of
     -fwrapv, and even for unsigned types.  */


GrÃÃe
 Thomas

Attachment: signature.asc
Description: PGP signature