1 /* Heuristics and transform for loop blocking and strip mining on
2 polyhedral representation.
4 Copyright (C) 2009-2015 Free Software Foundation, Inc.
5 Contributed by Sebastian Pop <sebastian.pop@amd.com> and
6 Pranav Garg <pranav.garg2107@gmail.com>.
8 This file is part of GCC.
10 GCC is free software; you can redistribute it and/or modify
11 it under the terms of the GNU General Public License as published by
12 the Free Software Foundation; either version 3, or (at your option)
15 GCC is distributed in the hope that it will be useful,
16 but WITHOUT ANY WARRANTY; without even the implied warranty of
17 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 GNU General Public License for more details.
20 You should have received a copy of the GNU General Public License
21 along with GCC; see the file COPYING3. If not see
22 <http://www.gnu.org/licenses/>. */
27 #include <isl/constraint.h>
30 #include <isl/union_map.h>
31 #include <isl/constraint.h>
35 #include "coretypes.h"
39 #include "double-int.h"
47 #include "fold-const.h"
50 #include "hard-reg-set.h"
53 #include "dominance.h"
54 #include "basic-block.h"
55 #include "tree-ssa-alias.h"
56 #include "internal-fn.h"
57 #include "gimple-expr.h"
60 #include "gimple-iterator.h"
61 #include "tree-ssa-loop.h"
64 #include "tree-chrec.h"
65 #include "tree-data-ref.h"
69 #include "graphite-poly.h"
72 /* Strip mines with a factor STRIDE the scattering (time) dimension
73 around PBB at depth TIME_DEPTH.
75 The following example comes from the wiki page:
76 http://gcc.gnu.org/wiki/Graphite/Strip_mine
78 The strip mine of a loop with a tile of 64 can be obtained with a
79 scattering function as follows:
81 $ cat ./albert_strip_mine.cloog
85 # parameter {n | n >= 0}
92 1 # Number of statements:
105 1 # Scattering functions
116 #the output of CLooG is like this:
117 #$ cloog ./albert_strip_mine.cloog
118 # for (NEW=0;NEW<=floord(n,64);NEW++) {
119 # for (OLD=max(64*NEW,0);OLD<=min(64*NEW+63,n);OLD++) {
126 pbb_strip_mine_time_depth (poly_bb_p pbb, int time_depth, int stride)
131 /* STRIP is the dimension that iterates with stride STRIDE. */
132 /* ITER is the dimension that enumerates single iterations inside
133 one strip that has at most STRIDE iterations. */
137 pbb->transformed = isl_map_insert_dims (pbb->transformed, isl_dim_out,
140 /* Lower bound of the striped loop. */
141 d = isl_map_get_space (pbb->transformed);
142 c = isl_inequality_alloc (isl_local_space_from_space (d));
143 c = isl_constraint_set_coefficient_si (c, isl_dim_out, strip, -stride);
144 c = isl_constraint_set_coefficient_si (c, isl_dim_out, iter, 1);
145 pbb->transformed = isl_map_add_constraint (pbb->transformed, c);
147 /* Upper bound of the striped loop. */
148 d = isl_map_get_space (pbb->transformed);
149 c = isl_inequality_alloc (isl_local_space_from_space (d));
150 c = isl_constraint_set_coefficient_si (c, isl_dim_out, strip, stride);
151 c = isl_constraint_set_coefficient_si (c, isl_dim_out, iter, -1);
152 c = isl_constraint_set_constant_si (c, stride - 1);
153 pbb->transformed = isl_map_add_constraint (pbb->transformed, c);
155 /* Static scheduling for ITER level.
156 This is mandatory to keep the 2d + 1 canonical scheduling format. */
157 d = isl_map_get_space (pbb->transformed);
158 c = isl_equality_alloc (isl_local_space_from_space (d));
159 c = isl_constraint_set_coefficient_si (c, isl_dim_out, strip + 1, 1);
160 pbb->transformed = isl_map_add_constraint (pbb->transformed, c);
163 /* Returns true when strip mining with STRIDE of the loop LST is
167 lst_strip_mine_profitable_p (lst_p lst, int stride)
169 mpz_t niter, strip_stride;
172 gcc_assert (LST_LOOP_P (lst));
173 mpz_init (strip_stride);
176 mpz_set_si (strip_stride, stride);
177 lst_niter_for_loop (lst, niter);
178 res = (mpz_cmp (niter, strip_stride) > 0);
180 mpz_clear (strip_stride);
185 /* Strip-mines all the loops of LST with STRIDE. Return the number of
186 loops strip-mined. */
189 lst_do_strip_mine_loop (lst_p lst, int depth, int stride)
198 if (LST_LOOP_P (lst))
202 FOR_EACH_VEC_ELT (LST_SEQ (lst), i, l)
203 res += lst_do_strip_mine_loop (l, depth, stride);
209 pbb_strip_mine_time_depth (pbb, psct_dynamic_dim (pbb, depth), stride);
213 /* Strip-mines all the loops of LST with STRIDE. When STRIDE is zero,
214 read the stride from the PARAM_LOOP_BLOCK_TILE_SIZE. Return the
215 number of strip-mined loops.
217 Strip mining transforms a loop
219 | for (i = 0; i < N; i++)
222 into the following loop nest:
224 | for (k = 0; k < N; k += STRIDE)
225 | for (j = 0; j < STRIDE; j++)
230 lst_do_strip_mine (lst_p lst, int stride)
238 stride = PARAM_VALUE (PARAM_LOOP_BLOCK_TILE_SIZE);
241 || !LST_LOOP_P (lst))
244 FOR_EACH_VEC_ELT (LST_SEQ (lst), i, l)
245 res += lst_do_strip_mine (l, stride);
247 depth = lst_depth (lst);
249 && lst_strip_mine_profitable_p (lst, stride))
251 res += lst_do_strip_mine_loop (lst, lst_depth (lst), stride);
252 lst_add_loop_under_loop (lst);
258 /* Strip mines all the loops in SCOP. Returns the number of
259 strip-mined loops. */
262 scop_do_strip_mine (scop_p scop, int stride)
264 return lst_do_strip_mine (SCOP_TRANSFORMED_SCHEDULE (scop), stride);
267 /* Loop blocks all the loops in SCOP. Returns true when we manage to
271 scop_do_block (scop_p scop)
273 store_scattering (scop);
275 /* If we don't strip mine at least two loops, or not interchange
276 loops, the strip mine alone will not be profitable, and the
277 transform is not a loop blocking: so revert the transform. */
278 if (lst_do_strip_mine (SCOP_TRANSFORMED_SCHEDULE (scop), 0) < 2
279 || scop_do_interchange (scop) == 0)
281 restore_scattering (scop);
285 if (dump_file && (dump_flags & TDF_DETAILS))
286 fprintf (dump_file, "SCoP will be loop blocked.\n");