Patchwork [Fortran] PR 55806 - Inefficient ANY with array constructors

login
register
mail settings
Submitter Thomas Koenig
Date Jan. 1, 2013, 8:18 p.m.
Message ID <50E344A6.1000802@netcologne.de>
Download mbox | patch
Permalink /patch/208961/
State New
Headers show

Comments

Thomas Koenig - Jan. 1, 2013, 8:18 p.m.
Hello world,

the attached patch replaces ANY(a, b, c) with a .or. b .or c,
leading to reduced execution time.  It also handles ALL, PRODUCT
and SUM.

This fixes a bug noted by Michael Metcalf.

Regression-tested.  OK for trunk?

	Thomas

2013-01-01  Thomas Koenig  <tkoenig@gcc.gnu.org>

         PR fortran/55806
         * frontend-passes.c (optimize_reduction):  New function,
         including prototype.
         (callback_reduction):  Likewise.
         (gfc_run_passes):  Also run optimize_reduction.
         (copy_walk_reduction_arg):  New function.
         (dummy_code_callback):  New function.

2013-01-01  Thomas Koenig  <tkoenig@gcc.gnu.org>

         PR fortran/55806
         * gfortran.dg/array_constructor_40.f90:  New test.
Thomas Koenig - Jan. 6, 2013, 4:33 p.m.
Ping?


http://gcc.gnu.org/ml/fortran/2013-01/msg00000.html

> Hello world,
>
> the attached patch replaces ANY(a, b, c) with a .or. b .or c,
> leading to reduced execution time.  It also handles ALL, PRODUCT
> and SUM.
>
> This fixes a bug noted by Michael Metcalf.
>
> Regression-tested.  OK for trunk?
Thomas Koenig - Jan. 8, 2013, 11:16 p.m.
Ping**2?

This was submitted before the review, so I think it should still be OK.

> Ping?
>
>
> http://gcc.gnu.org/ml/fortran/2013-01/msg00000.html
>
>> Hello world,
>>
>> the attached patch replaces ANY(a, b, c) with a .or. b .or c,
>> leading to reduced execution time.  It also handles ALL, PRODUCT
>> and SUM.
>>
>> This fixes a bug noted by Michael Metcalf.
>>
>> Regression-tested.  OK for trunk?
>
>
Mikael Morin - Jan. 11, 2013, 7:53 p.m.
Le 01/01/2013 21:18, Thomas Koenig a écrit :
> Hello world,
>
> the attached patch replaces ANY(a, b, c) with a .or. b .or c,
> leading to reduced execution time. It also handles ALL, PRODUCT
> and SUM.
>
> This fixes a bug noted by Michael Metcalf.
>
> Regression-tested. OK for trunk?
>
A few comments below.

Mikael

> Index: frontend-passes.c
> ===================================================================
> --- frontend-passes.c	(Revision 194760)
> +++ frontend-passes.c	(Arbeitskopie)
> @@ -180,7 +183,172 @@ optimize_expr (gfc_expr **e, int *walk_subtrees AT
>    return 0;
>  }
>
> +/* Auxiliary function to handle the arguments to reduction intrnisics.
> +   If the function is a scalar, just copy it; otherwise Returns the new
> +   element, the old one can be freed.  */
>
> +static gfc_expr *
> +copy_walk_reduction_arg (gfc_expr *e, gfc_expr *fn)
> +{
> +  gfc_expr *fcn;
> +  const char *new_name;
> +  gfc_actual_arglist *actual_arglist;
> +
> +  if (e->rank == 0 || e->expr_type == EXPR_FUNCTION)
> +    fcn = gfc_copy_expr (e);
> +  else
> +    {
> +      fcn = gfc_get_expr ();
> +      fcn->expr_type = EXPR_FUNCTION;
> +      fcn->value.function.isym = fn->value.function.isym;
> +      actual_arglist = gfc_get_actual_arglist ();
> +      actual_arglist->expr = gfc_copy_expr (e);
> +      actual_arglist->next = gfc_get_actual_arglist ();
Another one is needed.  I get a segmentation fault with SUM.

[...]
> +
> +/* Callback function for optimzation of reductions to scalars.  Transform
> +   ANY ([f1,f2,f3, ...]) to f1 .or. f2 .or. f3 .or. ..., with ANY,
> +   SUM and PRODUCT correspondingly.  Handly only the simple cases without
> +   MASK and DIM.  */
> +
> +static int
> +callback_reduction (gfc_expr **e, int *walk_subtrees ATTRIBUTE_UNUSED,
> +		    void *data ATTRIBUTE_UNUSED)
> +{
> +  gfc_expr *fn, *arg;
> +  gfc_intrinsic_op op;
> +  gfc_isym_id id;
> +  gfc_actual_arglist *a;
> +  gfc_actual_arglist *dim;
> +  gfc_constructor *c;
> +  gfc_expr *res, *new_expr;
> +
> +  fn = *e;
> +
> +  if (fn->rank != 0 || fn->expr_type != EXPR_FUNCTION
> +      || fn->value.function.isym == NULL)
> +    return 0;
> +
> +  id = fn->value.function.isym->id;
> +
> +  if (id != GFC_ISYM_SUM && id != GFC_ISYM_PRODUCT
> +      && id != GFC_ISYM_ANY && id != GFC_ISYM_ALL)
> +    return 0;
> +
> +  a = fn->value.function.actual;
> +
> +  /* Don't handle MASK or DIM.  */
> +
> +  dim = a->next;
> +
> +  if (dim != NULL)
> +    {
Minor, but I think you can assume dim != NULL.  Same for mask.

> +      gfc_actual_arglist *mask;
> +
> +      if (dim->expr != NULL)
> +	return 0;
> +
> +      mask = dim->next;
> +      if (mask != NULL)
> +	if ( mask->expr != NULL)
> +	  return 0;
> +    }
> +
> +  arg = a->expr;
> +
> +  if (arg->expr_type != EXPR_ARRAY)
> +    return 0;
> +
> +  switch (id)
> +    {
> +    case GFC_ISYM_SUM:
> +      op = INTRINSIC_PLUS;
> +      break;
> +
> +    case GFC_ISYM_PRODUCT:
> +      op = INTRINSIC_TIMES;
> +      break;
> +
> +    case GFC_ISYM_ANY:
> +      op = INTRINSIC_OR;
> +      break;
> +
> +    case GFC_ISYM_ALL:
> +      op = INTRINSIC_AND;
> +      break;
> +
> +    default:
> +      return 0;
> +    }
> +
> +  c = gfc_constructor_first (arg->value.constructor);
> +
> +  if (c == NULL)
> +    return 0;
> +
> +  res = copy_walk_reduction_arg (c->expr, fn);
> +
> +  c = gfc_constructor_next (c);
> +  while (c)
> +    {
> +      new_expr = gfc_get_expr ();
> +      new_expr->ts = fn->ts;
> +      new_expr->expr_type = EXPR_OP;
> +      new_expr->rank = fn->rank;
> +      new_expr->where = fn->where;
> +      new_expr->value.op.op = op;
> +      new_expr->value.op.op1 = res;
> +      new_expr->value.op.op2 = copy_walk_reduction_arg (c->expr, fn);
> +      res = new_expr;
> +      c = gfc_constructor_next (c);
> +    }
> +
> +  gfc_simplify_expr (res, 0);
> +  *e = res;
> +  gfc_free_expr (fn);
> +
> +  /* We changed things from under the expression walker.  Walking the
> +     old tree would mess up things, so let's not do that.  */
> +  return 1;
I think this prevents any further reduction optimization. The following 
variant of your test case doesn't avoid the temporary:

     do i=1,3
        if (any([abs(a(i,1) - b(i,1)) > acc,  &
             (j==i+1,j=3,8)])) cycle
        if (any([abs(a(i,2) - b(i,2)) > acc, &
             abs(a(i,3) - b(i,3)) > acc, lo(i,:)])) cycle
        c = c + i
     end do

Patch

Index: frontend-passes.c
===================================================================
--- frontend-passes.c	(Revision 194760)
+++ frontend-passes.c	(Arbeitskopie)
@@ -40,6 +40,8 @@  static bool optimize_lexical_comparison (gfc_expr
 static void optimize_minmaxloc (gfc_expr **);
 static bool is_empty_string (gfc_expr *e);
 static void doloop_warn (gfc_namespace *);
+static void optimize_reduction (gfc_namespace *);
+static int callback_reduction (gfc_expr **, int *, void *);
 
 /* How deep we are inside an argument list.  */
 
@@ -107,6 +109,7 @@  gfc_run_passes (gfc_namespace *ns)
       expr_array = XNEWVEC(gfc_expr **, expr_size);
 
       optimize_namespace (ns);
+      optimize_reduction (ns);
       if (gfc_option.dump_fortran_optimized)
 	gfc_dump_parse_tree (ns, stdout);
 
@@ -180,7 +183,172 @@  optimize_expr (gfc_expr **e, int *walk_subtrees AT
   return 0;
 }
 
+/* Auxiliary function to handle the arguments to reduction intrnisics.
+   If the function is a scalar, just copy it; otherwise Returns the new
+   element, the old one can be freed.  */
 
+static gfc_expr *
+copy_walk_reduction_arg (gfc_expr *e, gfc_expr *fn)
+{
+  gfc_expr *fcn;
+  const char *new_name;
+  gfc_actual_arglist *actual_arglist;
+
+  if (e->rank == 0 || e->expr_type == EXPR_FUNCTION)
+    fcn = gfc_copy_expr (e);
+  else
+    {
+      fcn = gfc_get_expr ();
+      fcn->expr_type = EXPR_FUNCTION;
+      fcn->value.function.isym = fn->value.function.isym;
+      actual_arglist = gfc_get_actual_arglist ();
+      actual_arglist->expr = gfc_copy_expr (e);
+      actual_arglist->next = gfc_get_actual_arglist ();
+      fcn->value.function.actual = actual_arglist;
+      fcn->ts = fn->ts;
+
+      switch (fn->value.function.isym->id)
+	{
+	case GFC_ISYM_SUM:
+	  new_name = "__internal_sum";
+	  break;
+	  
+	case GFC_ISYM_PRODUCT:
+	  new_name = "__internal_product";
+	  break;
+	  
+	case GFC_ISYM_ANY:
+	  new_name = "__internal_any";
+	  break;
+
+	case GFC_ISYM_ALL:
+	  new_name = "__internal_all";
+	  break;
+
+	default:
+	  abort ();
+	}
+
+      gfc_get_sym_tree (new_name, current_ns, &fcn->symtree, false);
+      fcn->symtree->n.sym->attr.flavor = FL_PROCEDURE;
+      fcn->symtree->n.sym->attr.function = 1;
+      fcn->symtree->n.sym->attr.elemental = 1;
+      fcn->symtree->n.sym->attr.referenced = 1;
+      fcn->symtree->n.sym->attr.access = ACCESS_PRIVATE;
+      gfc_commit_symbol (fcn->symtree->n.sym);
+    }
+
+  (void) gfc_expr_walker (&fcn, callback_reduction, NULL);
+
+  return fcn;
+}
+
+/* Callback function for optimzation of reductions to scalars.  Transform
+   ANY ([f1,f2,f3, ...]) to f1 .or. f2 .or. f3 .or. ..., with ANY,
+   SUM and PRODUCT correspondingly.  Handly only the simple cases without
+   MASK and DIM.  */
+
+static int
+callback_reduction (gfc_expr **e, int *walk_subtrees ATTRIBUTE_UNUSED,
+		    void *data ATTRIBUTE_UNUSED)
+{
+  gfc_expr *fn, *arg;
+  gfc_intrinsic_op op;
+  gfc_isym_id id;
+  gfc_actual_arglist *a;
+  gfc_actual_arglist *dim;
+  gfc_constructor *c;
+  gfc_expr *res, *new_expr;
+
+  fn = *e;
+
+  if (fn->rank != 0 || fn->expr_type != EXPR_FUNCTION
+      || fn->value.function.isym == NULL)
+    return 0;
+
+  id = fn->value.function.isym->id;
+
+  if (id != GFC_ISYM_SUM && id != GFC_ISYM_PRODUCT
+      && id != GFC_ISYM_ANY && id != GFC_ISYM_ALL)
+    return 0;
+
+  a = fn->value.function.actual;
+
+  /* Don't handle MASK or DIM.  */
+
+  dim = a->next;
+
+  if (dim != NULL)
+    {
+      gfc_actual_arglist *mask;
+
+      if (dim->expr != NULL)
+	return 0;
+      
+      mask = dim->next;
+      if (mask != NULL)
+	if ( mask->expr != NULL)
+	  return 0;
+    }
+
+  arg = a->expr;
+
+  if (arg->expr_type != EXPR_ARRAY)
+    return 0;
+
+  switch (id)
+    {
+    case GFC_ISYM_SUM:
+      op = INTRINSIC_PLUS;
+      break;
+
+    case GFC_ISYM_PRODUCT:
+      op = INTRINSIC_TIMES;
+      break;
+
+    case GFC_ISYM_ANY:
+      op = INTRINSIC_OR;
+      break;
+
+    case GFC_ISYM_ALL:
+      op = INTRINSIC_AND;
+      break;
+
+    default:
+      return 0;
+    }
+
+  c = gfc_constructor_first (arg->value.constructor);
+
+  if (c == NULL)
+    return 0;
+
+  res = copy_walk_reduction_arg (c->expr, fn);
+
+  c = gfc_constructor_next (c);
+  while (c)
+    {
+      new_expr = gfc_get_expr ();
+      new_expr->ts = fn->ts;
+      new_expr->expr_type = EXPR_OP;
+      new_expr->rank = fn->rank;
+      new_expr->where = fn->where;
+      new_expr->value.op.op = op;
+      new_expr->value.op.op1 = res;
+      new_expr->value.op.op2 = copy_walk_reduction_arg (c->expr, fn);
+      res = new_expr;
+      c = gfc_constructor_next (c);
+    }
+
+  gfc_simplify_expr (res, 0);
+  *e = res;
+  gfc_free_expr (fn);
+
+  /* We changed things from under the expression walker.  Walking the
+     old tree would mess up things, so let's not do that.  */
+  return 1;
+}
+
 /* Callback function for common function elimination, called from cfe_expr_0.
    Put all eligible function expressions into expr_array.  */
 
@@ -484,6 +652,16 @@  dummy_expr_callback (gfc_expr **e ATTRIBUTE_UNUSED
   return 0;
 }
 
+/* Dummy function for code callback, for use when we really
+   don't want to do anything.  */
+static int
+dummy_code_callback (gfc_code **e ATTRIBUTE_UNUSED,
+		     int *walk_subtrees ATTRIBUTE_UNUSED,
+		     void *data ATTRIBUTE_UNUSED)
+{
+  return 0;
+}
+
 /* Code callback function for converting
    do while(a)
    end do
@@ -639,6 +817,20 @@  optimize_namespace (gfc_namespace *ns)
     }
 }
 
+static void
+optimize_reduction (gfc_namespace *ns)
+{
+  current_ns = ns;
+  gfc_code_walker (&ns->code, dummy_code_callback, callback_reduction, NULL);
+
+/* BLOCKs are handled in the expression walker below.  */
+  for (ns = ns->contained; ns; ns = ns->sibling)
+    {
+      if (ns->code == NULL || ns->code->op != EXEC_BLOCK)
+	optimize_reduction (ns);
+    }
+}
+
 /* Replace code like
    a = matmul(b,c) + d
    with