1 /* Interchange heuristics and transform for loop interchange on
2 polyhedral representation.
4 Copyright (C) 2009-2014 Free Software Foundation, Inc.
5 Contributed by Sebastian Pop <sebastian.pop@amd.com> and
6 Harsha Jagasia <harsha.jagasia@amd.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/>. */
30 #include <isl/union_map.h>
32 #include <cloog/cloog.h>
33 #include <cloog/isl/domain.h>
37 #include "coretypes.h"
39 #include "basic-block.h"
40 #include "tree-ssa-alias.h"
41 #include "internal-fn.h"
42 #include "gimple-expr.h"
45 #include "gimple-iterator.h"
46 #include "tree-ssa-loop.h"
49 #include "tree-chrec.h"
50 #include "tree-data-ref.h"
51 #include "tree-scalar-evolution.h"
55 #include "graphite-poly.h"
57 /* XXX isl rewrite following comment */
58 /* Builds a linear expression, of dimension DIM, representing PDR's
61 L = r_{n}*r_{n-1}*...*r_{1}*s_{0} + ... + r_{n}*s_{n-1} + s_{n}.
63 For an array A[10][20] with two subscript locations s0 and s1, the
64 linear memory access is 20 * s0 + s1: a stride of 1 in subscript s0
65 corresponds to a memory stride of 20.
67 OFFSET is a number of dimensions to prepend before the
68 subscript dimensions: s_0, s_1, ..., s_n.
70 Thus, the final linear expression has the following format:
71 0 .. 0_{offset} | 0 .. 0_{nit} | 0 .. 0_{gd} | 0 | c_0 c_1 ... c_n
72 where the expression itself is:
73 c_0 * s_0 + c_1 * s_1 + ... c_n * s_n. */
75 static isl_constraint
*
76 build_linearized_memory_access (isl_map
*map
, poly_dr_p pdr
)
79 isl_local_space
*ls
= isl_local_space_from_space (isl_map_get_space (map
));
80 unsigned offset
, nsubs
;
82 isl_int size
, subsize
;
84 res
= isl_equality_alloc (ls
);
86 isl_int_set_ui (size
, 1);
87 isl_int_init (subsize
);
88 isl_int_set_ui (subsize
, 1);
90 nsubs
= isl_set_dim (pdr
->extent
, isl_dim_set
);
91 /* -1 for the already included L dimension. */
92 offset
= isl_map_dim (map
, isl_dim_out
) - 1 - nsubs
;
93 res
= isl_constraint_set_coefficient_si (res
, isl_dim_out
, offset
+ nsubs
, -1);
94 /* Go through all subscripts from last to first. First dimension
95 is the alias set, ignore it. */
96 for (i
= nsubs
- 1; i
>= 1; i
--)
101 res
= isl_constraint_set_coefficient (res
, isl_dim_out
, offset
+ i
, size
);
103 dc
= isl_set_get_space (pdr
->extent
);
104 aff
= isl_aff_zero_on_domain (isl_local_space_from_space (dc
));
105 aff
= isl_aff_set_coefficient_si (aff
, isl_dim_in
, i
, 1);
106 isl_set_max (pdr
->extent
, aff
, &subsize
);
108 isl_int_mul (size
, size
, subsize
);
111 isl_int_clear (subsize
);
112 isl_int_clear (size
);
117 /* Set STRIDE to the stride of PDR in memory by advancing by one in
118 the loop at DEPTH. */
121 pdr_stride_in_loop (mpz_t stride
, graphite_dim_t depth
, poly_dr_p pdr
)
123 poly_bb_p pbb
= PDR_PBB (pdr
);
128 isl_constraint
*lma
, *c
;
130 graphite_dim_t time_depth
;
133 /* XXX isl rewrite following comments. */
134 /* Builds a partial difference equations and inserts them
135 into pointset powerset polyhedron P. Polyhedron is assumed
136 to have the format: T|I|T'|I'|G|S|S'|l1|l2.
138 TIME_DEPTH is the time dimension w.r.t. which we are
140 OFFSET represents the number of dimensions between
141 columns t_{time_depth} and t'_{time_depth}.
142 DIM_SCTR is the number of scattering dimensions. It is
143 essentially the dimensionality of the T vector.
145 The following equations are inserted into the polyhedron P:
148 | t_{time_depth-1} = t'_{time_depth-1}
149 | t_{time_depth} = t'_{time_depth} + 1
150 | t_{time_depth+1} = t'_{time_depth + 1}
152 | t_{dim_sctr} = t'_{dim_sctr}. */
154 /* Add the equality: t_{time_depth} = t'_{time_depth} + 1.
155 This is the core part of this alogrithm, since this
156 constraint asks for the memory access stride (difference)
157 between two consecutive points in time dimensions. */
162 | t_{time_depth-1} = t'_{time_depth-1}
163 | t_{time_depth+1} = t'_{time_depth+1}
165 | t_{dim_sctr} = t'_{dim_sctr}
167 This means that all the time dimensions are equal except for
168 time_depth, where the constraint is t_{depth} = t'_{depth} + 1
169 step. More to this: we should be careful not to add equalities
170 to the 'coupled' dimensions, which happens when the one dimension
171 is stripmined dimension, and the other dimension corresponds
172 to the point loop inside stripmined dimension. */
174 /* pdr->accesses: [P1..nb_param,I1..nb_domain]->[a,S1..nb_subscript]
175 ??? [P] not used for PDRs?
176 pdr->extent: [a,S1..nb_subscript]
177 pbb->domain: [P1..nb_param,I1..nb_domain]
178 pbb->transformed: [P1..nb_param,I1..nb_domain]->[T1..Tnb_sctr]
179 [T] includes local vars (currently unused)
181 First we create [P,I] -> [T,a,S]. */
183 map
= isl_map_flat_range_product (isl_map_copy (pbb
->transformed
),
184 isl_map_copy (pdr
->accesses
));
185 /* Add a dimension for L: [P,I] -> [T,a,S,L].*/
186 map
= isl_map_add_dims (map
, isl_dim_out
, 1);
187 /* Build a constraint for "lma[S] - L == 0", effectively calculating
188 L in terms of subscripts. */
189 lma
= build_linearized_memory_access (map
, pdr
);
190 /* And add it to the map, so we now have:
191 [P,I] -> [T,a,S,L] : lma([S]) == L. */
192 map
= isl_map_add_constraint (map
, lma
);
194 /* Then we create [P,I,P',I'] -> [T,a,S,L,T',a',S',L']. */
195 map
= isl_map_flat_product (map
, isl_map_copy (map
));
197 /* Now add the equality T[time_depth] == T'[time_depth]+1. This will
198 force L' to be the linear address at T[time_depth] + 1. */
199 time_depth
= psct_dynamic_dim (pbb
, depth
);
200 /* Length of [a,S] plus [L] ... */
201 offset
= 1 + isl_map_dim (pdr
->accesses
, isl_dim_out
);
203 offset
+= isl_map_dim (pbb
->transformed
, isl_dim_out
);
205 c
= isl_equality_alloc (isl_local_space_from_space (isl_map_get_space (map
)));
206 c
= isl_constraint_set_coefficient_si (c
, isl_dim_out
, time_depth
, 1);
207 c
= isl_constraint_set_coefficient_si (c
, isl_dim_out
,
208 offset
+ time_depth
, -1);
209 c
= isl_constraint_set_constant_si (c
, 1);
210 map
= isl_map_add_constraint (map
, c
);
212 /* Now we equate most of the T/T' elements (making PITaSL nearly
213 the same is (PITaSL)', except for one dimension, namely for 'depth'
214 (an index into [I]), after translating to index into [T]. Take care
215 to not produce an empty map, which indicates we wanted to equate
216 two dimensions that are already coupled via the above time_depth
217 dimension. Happens with strip mining where several scatter dimension
218 are interdependend. */
220 nt
= pbb_nb_scattering_transform (pbb
) + pbb_nb_local_vars (pbb
);
221 for (i
= 0; i
< nt
; i
++)
224 isl_map
*temp
= isl_map_equate (isl_map_copy (map
),
226 isl_dim_out
, offset
+ i
);
227 if (isl_map_is_empty (temp
))
236 /* Now maximize the expression L' - L. */
237 set
= isl_map_range (map
);
238 dc
= isl_set_get_space (set
);
239 aff
= isl_aff_zero_on_domain (isl_local_space_from_space (dc
));
240 aff
= isl_aff_set_coefficient_si (aff
, isl_dim_in
, offset
- 1, -1);
241 aff
= isl_aff_set_coefficient_si (aff
, isl_dim_in
, offset
+ offset
- 1, 1);
242 isl_int_init (islstride
);
243 isl_set_max (set
, aff
, &islstride
);
244 isl_int_get_gmp (islstride
, stride
);
245 isl_int_clear (islstride
);
249 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
251 gmp_fprintf (dump_file
, "\nStride in BB_%d, DR_%d, depth %d: %Zd ",
252 pbb_index (pbb
), PDR_ID (pdr
), (int) depth
, stride
);
256 /* Sets STRIDES to the sum of all the strides of the data references
257 accessed in LOOP at DEPTH. */
260 memory_strides_in_loop_1 (lst_p loop
, graphite_dim_t depth
, mpz_t strides
)
270 FOR_EACH_VEC_ELT (LST_SEQ (loop
), j
, l
)
272 memory_strides_in_loop_1 (l
, depth
, strides
);
274 FOR_EACH_VEC_ELT (PBB_DRS (LST_PBB (l
)), i
, pdr
)
276 pdr_stride_in_loop (s
, depth
, pdr
);
277 mpz_set_si (n
, PDR_NB_REFS (pdr
));
279 mpz_add (strides
, strides
, s
);
286 /* Sets STRIDES to the sum of all the strides of the data references
287 accessed in LOOP at DEPTH. */
290 memory_strides_in_loop (lst_p loop
, graphite_dim_t depth
, mpz_t strides
)
292 if (mpz_cmp_si (loop
->memory_strides
, -1) == 0)
294 mpz_set_si (strides
, 0);
295 memory_strides_in_loop_1 (loop
, depth
, strides
);
298 mpz_set (strides
, loop
->memory_strides
);
301 /* Return true when the interchange of loops LOOP1 and LOOP2 is
314 | for (i = 0; i < N; i++)
315 | for (j = 0; j < N; j++)
321 The data access A[j][i] is described like this:
329 | 0 0 0 0 -1 0 100 >= 0
330 | 0 0 0 0 0 -1 100 >= 0
332 The linearized memory access L to A[100][100] is:
337 TODO: the shown format is not valid as it does not show the fact
338 that the iteration domain "i j" is transformed using the scattering.
340 Next, to measure the impact of iterating once in loop "i", we build
341 a maximization problem: first, we add to DR accesses the dimensions
342 k, s2, s3, L1 = 100 * s0 + s1, L2, and D1: this is the polyhedron P1.
343 L1 and L2 are the linearized memory access functions.
345 | i j N a s0 s1 k s2 s3 L1 L2 D1 1
346 | 0 0 0 1 0 0 0 0 0 0 0 0 -5 = 0 alias = 5
347 | 0 -1 0 0 1 0 0 0 0 0 0 0 0 = 0 s0 = j
348 |-2 0 0 0 0 1 0 0 0 0 0 0 0 = 0 s1 = 2 * i
349 | 0 0 0 0 1 0 0 0 0 0 0 0 0 >= 0
350 | 0 0 0 0 0 1 0 0 0 0 0 0 0 >= 0
351 | 0 0 0 0 -1 0 0 0 0 0 0 0 100 >= 0
352 | 0 0 0 0 0 -1 0 0 0 0 0 0 100 >= 0
353 | 0 0 0 0 100 1 0 0 0 -1 0 0 0 = 0 L1 = 100 * s0 + s1
355 Then, we generate the polyhedron P2 by interchanging the dimensions
356 (s0, s2), (s1, s3), (L1, L2), (k, i)
358 | i j N a s0 s1 k s2 s3 L1 L2 D1 1
359 | 0 0 0 1 0 0 0 0 0 0 0 0 -5 = 0 alias = 5
360 | 0 -1 0 0 0 0 0 1 0 0 0 0 0 = 0 s2 = j
361 | 0 0 0 0 0 0 -2 0 1 0 0 0 0 = 0 s3 = 2 * k
362 | 0 0 0 0 0 0 0 1 0 0 0 0 0 >= 0
363 | 0 0 0 0 0 0 0 0 1 0 0 0 0 >= 0
364 | 0 0 0 0 0 0 0 -1 0 0 0 0 100 >= 0
365 | 0 0 0 0 0 0 0 0 -1 0 0 0 100 >= 0
366 | 0 0 0 0 0 0 0 100 1 0 -1 0 0 = 0 L2 = 100 * s2 + s3
368 then we add to P2 the equality k = i + 1:
370 |-1 0 0 0 0 0 1 0 0 0 0 0 -1 = 0 k = i + 1
372 and finally we maximize the expression "D1 = max (P1 inter P2, L2 - L1)".
374 Similarly, to determine the impact of one iteration on loop "j", we
375 interchange (k, j), we add "k = j + 1", and we compute D2 the
376 maximal value of the difference.
378 Finally, the profitability test is D1 < D2: if in the outer loop
379 the strides are smaller than in the inner loop, then it is
380 profitable to interchange the loops at DEPTH1 and DEPTH2. */
383 lst_interchange_profitable_p (lst_p nest
, int depth1
, int depth2
)
388 gcc_assert (depth1
< depth2
);
393 memory_strides_in_loop (nest
, depth1
, d1
);
394 memory_strides_in_loop (nest
, depth2
, d2
);
396 res
= mpz_cmp (d1
, d2
) < 0;
404 /* Interchanges the loops at DEPTH1 and DEPTH2 of the original
405 scattering and assigns the resulting polyhedron to the transformed
409 pbb_interchange_loop_depths (graphite_dim_t depth1
, graphite_dim_t depth2
,
413 unsigned dim1
= psct_dynamic_dim (pbb
, depth1
);
414 unsigned dim2
= psct_dynamic_dim (pbb
, depth2
);
415 isl_space
*d
= isl_map_get_space (pbb
->transformed
);
416 isl_space
*d1
= isl_space_range (d
);
417 unsigned n
= isl_space_dim (d1
, isl_dim_out
);
418 isl_space
*d2
= isl_space_add_dims (d1
, isl_dim_in
, n
);
419 isl_map
*x
= isl_map_universe (d2
);
421 x
= isl_map_equate (x
, isl_dim_in
, dim1
, isl_dim_out
, dim2
);
422 x
= isl_map_equate (x
, isl_dim_in
, dim2
, isl_dim_out
, dim1
);
424 for (i
= 0; i
< n
; i
++)
425 if (i
!= dim1
&& i
!= dim2
)
426 x
= isl_map_equate (x
, isl_dim_in
, i
, isl_dim_out
, i
);
428 pbb
->transformed
= isl_map_apply_range (pbb
->transformed
, x
);
431 /* Apply the interchange of loops at depths DEPTH1 and DEPTH2 to all
432 the statements below LST. */
435 lst_apply_interchange (lst_p lst
, int depth1
, int depth2
)
440 if (LST_LOOP_P (lst
))
445 FOR_EACH_VEC_ELT (LST_SEQ (lst
), i
, l
)
446 lst_apply_interchange (l
, depth1
, depth2
);
449 pbb_interchange_loop_depths (depth1
, depth2
, LST_PBB (lst
));
452 /* Return true when the nest starting at LOOP1 and ending on LOOP2 is
453 perfect: i.e. there are no sequence of statements. */
456 lst_perfectly_nested_p (lst_p loop1
, lst_p loop2
)
461 if (!LST_LOOP_P (loop1
))
464 return LST_SEQ (loop1
).length () == 1
465 && lst_perfectly_nested_p (LST_SEQ (loop1
)[0], loop2
);
468 /* Transform the loop nest between LOOP1 and LOOP2 into a perfect
469 nest. To continue the naming tradition, this function is called
470 after perfect_nestify. NEST is set to the perfectly nested loop
471 that is created. BEFORE/AFTER are set to the loops distributed
472 before/after the loop NEST. */
475 lst_perfect_nestify (lst_p loop1
, lst_p loop2
, lst_p
*before
,
476 lst_p
*nest
, lst_p
*after
)
478 poly_bb_p first
, last
;
480 gcc_assert (loop1
&& loop2
482 && LST_LOOP_P (loop1
) && LST_LOOP_P (loop2
));
484 first
= LST_PBB (lst_find_first_pbb (loop2
));
485 last
= LST_PBB (lst_find_last_pbb (loop2
));
487 *before
= copy_lst (loop1
);
488 *nest
= copy_lst (loop1
);
489 *after
= copy_lst (loop1
);
491 lst_remove_all_before_including_pbb (*before
, first
, false);
492 lst_remove_all_before_including_pbb (*after
, last
, true);
494 lst_remove_all_before_excluding_pbb (*nest
, first
, true);
495 lst_remove_all_before_excluding_pbb (*nest
, last
, false);
497 if (lst_empty_p (*before
))
502 if (lst_empty_p (*after
))
507 if (lst_empty_p (*nest
))
514 /* Try to interchange LOOP1 with LOOP2 for all the statements of the
515 body of LOOP2. LOOP1 contains LOOP2. Return true if it did the
519 lst_try_interchange_loops (scop_p scop
, lst_p loop1
, lst_p loop2
)
521 int depth1
= lst_depth (loop1
);
522 int depth2
= lst_depth (loop2
);
525 lst_p before
= NULL
, nest
= NULL
, after
= NULL
;
527 if (!lst_perfectly_nested_p (loop1
, loop2
))
528 lst_perfect_nestify (loop1
, loop2
, &before
, &nest
, &after
);
530 if (!lst_interchange_profitable_p (loop2
, depth1
, depth2
))
533 lst_apply_interchange (loop2
, depth1
, depth2
);
535 /* Sync the transformed LST information and the PBB scatterings
536 before using the scatterings in the data dependence analysis. */
537 if (before
|| nest
|| after
)
539 transformed
= lst_substitute_3 (SCOP_TRANSFORMED_SCHEDULE (scop
), loop1
,
540 before
, nest
, after
);
541 lst_update_scattering (transformed
);
542 free_lst (transformed
);
545 if (graphite_legal_transform (scop
))
547 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
549 "Loops at depths %d and %d will be interchanged.\n",
552 /* Transform the SCOP_TRANSFORMED_SCHEDULE of the SCOP. */
553 lst_insert_in_sequence (before
, loop1
, true);
554 lst_insert_in_sequence (after
, loop1
, false);
558 lst_replace (loop1
, nest
);
565 /* Undo the transform. */
569 lst_apply_interchange (loop2
, depth2
, depth1
);
573 /* Selects the inner loop in LST_SEQ (INNER_FATHER) to be interchanged
574 with the loop OUTER in LST_SEQ (OUTER_FATHER). */
577 lst_interchange_select_inner (scop_p scop
, lst_p outer_father
, int outer
,
583 gcc_assert (outer_father
584 && LST_LOOP_P (outer_father
)
585 && LST_LOOP_P (LST_SEQ (outer_father
)[outer
])
587 && LST_LOOP_P (inner_father
));
589 loop1
= LST_SEQ (outer_father
)[outer
];
591 FOR_EACH_VEC_ELT (LST_SEQ (inner_father
), inner
, loop2
)
592 if (LST_LOOP_P (loop2
)
593 && (lst_try_interchange_loops (scop
, loop1
, loop2
)
594 || lst_interchange_select_inner (scop
, outer_father
, outer
, loop2
)))
600 /* Interchanges all the loops of LOOP and the loops of its body that
601 are considered profitable to interchange. Return the number of
602 interchanged loops. OUTER is the index in LST_SEQ (LOOP) that
603 points to the next outer loop to be considered for interchange. */
606 lst_interchange_select_outer (scop_p scop
, lst_p loop
, int outer
)
613 if (!loop
|| !LST_LOOP_P (loop
))
616 father
= LST_LOOP_FATHER (loop
);
619 while (lst_interchange_select_inner (scop
, father
, outer
, loop
))
622 loop
= LST_SEQ (father
)[outer
];
626 if (LST_LOOP_P (loop
))
627 FOR_EACH_VEC_ELT (LST_SEQ (loop
), i
, l
)
629 res
+= lst_interchange_select_outer (scop
, l
, i
);
634 /* Interchanges all the loop depths that are considered profitable for
635 SCOP. Return the number of interchanged loops. */
638 scop_do_interchange (scop_p scop
)
640 int res
= lst_interchange_select_outer
641 (scop
, SCOP_TRANSFORMED_SCHEDULE (scop
), 0);
643 lst_update_scattering (SCOP_TRANSFORMED_SCHEDULE (scop
));