re PR libfortran/32972 (performance of pack/unpack)

2008-03-21  Thomas Koenig  <tkoenig@gcc.gnu.org>

	PR libfortran/32972
	* Makefile.am:  Add new variable, i_pack_c, containing
	pack_i1.c, pack_i2.c, pack_i4.c, pack_i8.c, pack_i16.c,
	pack_r4.c, pack_r8.c, pack_r10.c, pack_r16.c, pack_c4.c,
	pack_c8.c, pack_c10.c, pack_c16.c.
	Add m4/pack.m4 to m4_files.
	Add i_pack_c to gfor_built_src.
	Add rule to generate i_pack_c from m4/pack.m4.
	* Makefile.in:  Regenerated.
	* libgfortran.h: Add prototypes for pack_i1, pack_i2, pack_i4,
	pack_i8, pack_i16, pack_r4, pack_r8, pack_c4, pack_c8,
	pack_c10, pack_c16.
	* intrinsics/pack_generic.c:  Add calls to specific
	pack functions.
	* m4/pack.m4:  New file.
	* generated/pack_i1.c:  New file.
	* generated/pack_i2.c:  New file.
	* generated/pack_i4.c:  New file.
	* generated/pack_i8.c:  New file.
	* generated/pack_i16.c:  New file.
	* generated/pack_r4.c:  New file.
	* generated/pack_r8.c:  New file.
	* generated/pack_r10.c:  New file.
	* generated/pack_r16.c:  New file.
	* generated/pack_c4.c:  New file.
	* generated/pack_c8.c:  New file.
	* generated/pack_c10.c:  New file.
	* generated/pack_c16.c:  New file.

2008-03-21  Thomas Koenig  <tkoenig@gcc.gnu.org>

	PR libfortran/32972
	* internal_pack_1.f90:  New test case.
	* internal_pack_2.f90:  New test case.
	* internal_pack_3.f90:  New test case.

From-SVN: r133427

1 files changed, 312 insertions, 0 deletions

diff --git a/libgfortran/m4/pack.m4 b/libgfortran/m4/pack.m4
new file mode 100644
index 0000000..87409a5
--- /dev/null
+++ b/libgfortran/m4/pack.m4
@@ -0,0 +1,312 @@
+`/* Specific implementation of the PACK intrinsic
+   Copyright (C) 2002, 2004, 2005, 2006, 2007, 2008 Free Software Foundation, Inc.
+   Contributed by Paul Brook <paul@nowt.org>
+
+This file is part of the GNU Fortran 95 runtime library (libgfortran).
+
+Libgfortran is free software; you can redistribute it and/or
+modify it under the terms of the GNU General Public
+License as published by the Free Software Foundation; either
+version 2 of the License, or (at your option) any later version.
+
+In addition to the permissions in the GNU General Public License, the
+Free Software Foundation gives you unlimited permission to link the
+compiled version of this file into combinations with other programs,
+and to distribute those combinations without any restriction coming
+from the use of this file.  (The General Public License restrictions
+do apply in other respects; for example, they cover modification of
+the file, and distribution when not linked into a combine
+executable.)
+
+Ligbfortran is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public
+License along with libgfortran; see the file COPYING.  If not,
+write to the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+Boston, MA 02110-1301, USA.  */
+
+#include "libgfortran.h"
+#include <stdlib.h>
+#include <assert.h>
+#include <string.h>'
+
+include(iparm.m4)dnl
+
+`#if defined (HAVE_'rtype_name`)
+
+/* PACK is specified as follows:
+
+   13.14.80 PACK (ARRAY, MASK, [VECTOR])
+
+   Description: Pack an array into an array of rank one under the
+   control of a mask.
+
+   Class: Transformational function.
+
+   Arguments:
+      ARRAY   may be of any type. It shall not be scalar.
+      MASK    shall be of type LOGICAL. It shall be conformable with ARRAY.
+      VECTOR  (optional) shall be of the same type and type parameters
+              as ARRAY. VECTOR shall have at least as many elements as
+              there are true elements in MASK. If MASK is a scalar
+              with the value true, VECTOR shall have at least as many
+              elements as there are in ARRAY.
+
+   Result Characteristics: The result is an array of rank one with the
+   same type and type parameters as ARRAY. If VECTOR is present, the
+   result size is that of VECTOR; otherwise, the result size is the
+   number /t/ of true elements in MASK unless MASK is scalar with the
+   value true, in which case the result size is the size of ARRAY.
+
+   Result Value: Element /i/ of the result is the element of ARRAY
+   that corresponds to the /i/th true element of MASK, taking elements
+   in array element order, for /i/ = 1, 2, ..., /t/. If VECTOR is
+   present and has size /n/ > /t/, element /i/ of the result has the
+   value VECTOR(/i/), for /i/ = /t/ + 1, ..., /n/.
+
+   Examples: The nonzero elements of an array M with the value
+   | 0 0 0 |
+   | 9 0 0 | may be "gathered" by the function PACK. The result of
+   | 0 0 7 |
+   PACK (M, MASK = M.NE.0) is [9,7] and the result of PACK (M, M.NE.0,
+   VECTOR = (/ 2,4,6,8,10,12 /)) is [9,7,6,8,10,12].
+
+There are two variants of the PACK intrinsic: one, where MASK is
+array valued, and the other one where MASK is scalar.  */
+
+void
+pack_'rtype_code` ('rtype` *ret, const 'rtype` *array,
+	       const gfc_array_l1 *mask, const 'rtype` *vector)
+{
+  /* r.* indicates the return array.  */
+  index_type rstride0;
+  'rtype_name` *rptr;
+  /* s.* indicates the source array.  */
+  index_type sstride[GFC_MAX_DIMENSIONS];
+  index_type sstride0;
+  const 'rtype_name` *sptr;
+  /* m.* indicates the mask array.  */
+  index_type mstride[GFC_MAX_DIMENSIONS];
+  index_type mstride0;
+  const GFC_LOGICAL_1 *mptr;
+
+  index_type count[GFC_MAX_DIMENSIONS];
+  index_type extent[GFC_MAX_DIMENSIONS];
+  int zero_sized;
+  index_type n;
+  index_type dim;
+  index_type nelem;
+  index_type total;
+  int mask_kind;
+
+  dim = GFC_DESCRIPTOR_RANK (array);
+
+  sptr = array->data;
+  mptr = mask->data;
+
+  /* Use the same loop for all logical types, by using GFC_LOGICAL_1
+     and using shifting to address size and endian issues.  */
+
+  mask_kind = GFC_DESCRIPTOR_SIZE (mask);
+
+  if (mask_kind == 1 || mask_kind == 2 || mask_kind == 4 || mask_kind == 8
+#ifdef HAVE_GFC_LOGICAL_16
+      || mask_kind == 16
+#endif
+      )
+    {
+      /*  Do not convert a NULL pointer as we use test for NULL below.  */
+      if (mptr)
+	mptr = GFOR_POINTER_TO_L1 (mptr, mask_kind);
+    }
+  else
+    runtime_error ("Funny sized logical array");
+
+  zero_sized = 0;
+  for (n = 0; n < dim; n++)
+    {
+      count[n] = 0;
+      extent[n] = array->dim[n].ubound + 1 - array->dim[n].lbound;
+      if (extent[n] <= 0)
+       zero_sized = 1;
+      sstride[n] = array->dim[n].stride;
+      mstride[n] = mask->dim[n].stride * mask_kind;
+    }
+  if (sstride[0] == 0)
+    sstride[0] = 1;
+  if (mstride[0] == 0)
+    mstride[0] = mask_kind;
+
+  if (ret->data == NULL || compile_options.bounds_check)
+    {
+      /* Count the elements, either for allocating memory or
+	 for bounds checking.  */
+
+      if (vector != NULL)
+	{
+	  /* The return array will have as many
+	     elements as there are in VECTOR.  */
+	  total = vector->dim[0].ubound + 1 - vector->dim[0].lbound;
+	}
+      else
+	{
+	  /* We have to count the true elements in MASK.  */
+
+	  /* TODO: We could speed up pack easily in the case of only
+	     few .TRUE. entries in MASK, by keeping track of where we
+	     would be in the source array during the initial traversal
+	     of MASK, and caching the pointers to those elements. Then,
+	     supposed the number of elements is small enough, we would
+	     only have to traverse the list, and copy those elements
+	     into the result array. In the case of datatypes which fit
+	     in one of the integer types we could also cache the
+	     value instead of a pointer to it.
+	     This approach might be bad from the point of view of
+	     cache behavior in the case where our cache is not big
+	     enough to hold all elements that have to be copied.  */
+
+	  const GFC_LOGICAL_1 *m = mptr;
+
+	  total = 0;
+	  if (zero_sized)
+	    m = NULL;
+
+	  while (m)
+	    {
+	      /* Test this element.  */
+	      if (*m)
+		total++;
+
+	      /* Advance to the next element.  */
+	      m += mstride[0];
+	      count[0]++;
+	      n = 0;
+	      while (count[n] == extent[n])
+		{
+		  /* When we get to the end of a dimension, reset it
+		     and increment the next dimension.  */
+		  count[n] = 0;
+		  /* We could precalculate this product, but this is a
+		     less frequently used path so probably not worth
+		     it.  */
+		  m -= mstride[n] * extent[n];
+		  n++;
+		  if (n >= dim)
+		    {
+		      /* Break out of the loop.  */
+		      m = NULL;
+		      break;
+		    }
+		  else
+		    {
+		      count[n]++;
+		      m += mstride[n];
+		    }
+		}
+	    }
+	}
+
+      if (ret->data == NULL)
+	{
+	  /* Setup the array descriptor.  */
+	  ret->dim[0].lbound = 0;
+	  ret->dim[0].ubound = total - 1;
+	  ret->dim[0].stride = 1;
+
+	  ret->offset = 0;
+	  if (total == 0)
+	    {
+	      /* In this case, nothing remains to be done.  */
+	      ret->data = internal_malloc_size (1);
+	      return;
+	    }
+	  else
+	    ret->data = internal_malloc_size (sizeof ('rtype_name`) * total);
+	}
+      else 
+	{
+	  /* We come here because of range checking.  */
+	  index_type ret_extent;
+
+	  ret_extent = ret->dim[0].ubound + 1 - ret->dim[0].lbound;
+	  if (total != ret_extent)
+	    runtime_error ("Incorrect extent in return value of PACK intrinsic;"
+			   " is %ld, should be %ld", (long int) total,
+			   (long int) ret_extent);
+	}
+    }
+
+  rstride0 = ret->dim[0].stride;
+  if (rstride0 == 0)
+    rstride0 = 1;
+  sstride0 = sstride[0];
+  mstride0 = mstride[0];
+  rptr = ret->data;
+
+  while (sptr && mptr)
+    {
+      /* Test this element.  */
+      if (*mptr)
+        {
+          /* Add it.  */
+	  *rptr = *sptr;
+          rptr += rstride0;
+        }
+      /* Advance to the next element.  */
+      sptr += sstride0;
+      mptr += mstride0;
+      count[0]++;
+      n = 0;
+      while (count[n] == extent[n])
+        {
+          /* When we get to the end of a dimension, reset it and increment
+             the next dimension.  */
+          count[n] = 0;
+          /* We could precalculate these products, but this is a less
+             frequently used path so probably not worth it.  */
+          sptr -= sstride[n] * extent[n];
+          mptr -= mstride[n] * extent[n];
+          n++;
+          if (n >= dim)
+            {
+              /* Break out of the loop.  */
+              sptr = NULL;
+              break;
+            }
+          else
+            {
+              count[n]++;
+              sptr += sstride[n];
+              mptr += mstride[n];
+            }
+        }
+    }
+
+  /* Add any remaining elements from VECTOR.  */
+  if (vector)
+    {
+      n = vector->dim[0].ubound + 1 - vector->dim[0].lbound;
+      nelem = ((rptr - ret->data) / rstride0);
+      if (n > nelem)
+        {
+          sstride0 = vector->dim[0].stride;
+          if (sstride0 == 0)
+            sstride0 = 1;
+
+          sptr = vector->data + sstride0 * nelem;
+          n -= nelem;
+          while (n--)
+            {
+	      *rptr = *sptr;
+              rptr += rstride0;
+              sptr += sstride0;
+            }
+        }
+    }
+}
+
+#endif
+'
\ No newline at end of file