1 /* Interchange heuristics and transform for loop interchange on
2 polyhedral representation.
4 Copyright (C) 2009-2013 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"
42 #include "tree-chrec.h"
43 #include "tree-data-ref.h"
44 #include "tree-scalar-evolution.h"
48 #include "graphite-poly.h"
50 /* XXX isl rewrite following comment */
51 /* Builds a linear expression, of dimension DIM, representing PDR's
54 L = r_{n}*r_{n-1}*...*r_{1}*s_{0} + ... + r_{n}*s_{n-1} + s_{n}.
56 For an array A[10][20] with two subscript locations s0 and s1, the
57 linear memory access is 20 * s0 + s1: a stride of 1 in subscript s0
58 corresponds to a memory stride of 20.
60 OFFSET is a number of dimensions to prepend before the
61 subscript dimensions: s_0, s_1, ..., s_n.
63 Thus, the final linear expression has the following format:
64 0 .. 0_{offset} | 0 .. 0_{nit} | 0 .. 0_{gd} | 0 | c_0 c_1 ... c_n
65 where the expression itself is:
66 c_0 * s_0 + c_1 * s_1 + ... c_n * s_n. */
68 static isl_constraint
*
69 build_linearized_memory_access (isl_map
*map
, poly_dr_p pdr
)
72 isl_local_space
*ls
= isl_local_space_from_space (isl_map_get_space (map
));
73 unsigned offset
, nsubs
;
75 isl_int size
, subsize
;
77 res
= isl_equality_alloc (ls
);
79 isl_int_set_ui (size
, 1);
80 isl_int_init (subsize
);
81 isl_int_set_ui (subsize
, 1);
83 nsubs
= isl_set_dim (pdr
->extent
, isl_dim_set
);
84 /* -1 for the already included L dimension. */
85 offset
= isl_map_dim (map
, isl_dim_out
) - 1 - nsubs
;
86 res
= isl_constraint_set_coefficient_si (res
, isl_dim_out
, offset
+ nsubs
, -1);
87 /* Go through all subscripts from last to first. First dimension
88 is the alias set, ignore it. */
89 for (i
= nsubs
- 1; i
>= 1; i
--)
94 res
= isl_constraint_set_coefficient (res
, isl_dim_out
, offset
+ i
, size
);
96 dc
= isl_set_get_space (pdr
->extent
);
97 aff
= isl_aff_zero_on_domain (isl_local_space_from_space (dc
));
98 aff
= isl_aff_set_coefficient_si (aff
, isl_dim_in
, i
, 1);
99 isl_set_max (pdr
->extent
, aff
, &subsize
);
101 isl_int_mul (size
, size
, subsize
);
104 isl_int_clear (subsize
);
105 isl_int_clear (size
);
110 /* Set STRIDE to the stride of PDR in memory by advancing by one in
111 the loop at DEPTH. */
114 pdr_stride_in_loop (mpz_t stride
, graphite_dim_t depth
, poly_dr_p pdr
)
116 poly_bb_p pbb
= PDR_PBB (pdr
);
121 isl_constraint
*lma
, *c
;
123 graphite_dim_t time_depth
;
126 /* XXX isl rewrite following comments. */
127 /* Builds a partial difference equations and inserts them
128 into pointset powerset polyhedron P. Polyhedron is assumed
129 to have the format: T|I|T'|I'|G|S|S'|l1|l2.
131 TIME_DEPTH is the time dimension w.r.t. which we are
133 OFFSET represents the number of dimensions between
134 columns t_{time_depth} and t'_{time_depth}.
135 DIM_SCTR is the number of scattering dimensions. It is
136 essentially the dimensionality of the T vector.
138 The following equations are inserted into the polyhedron P:
141 | t_{time_depth-1} = t'_{time_depth-1}
142 | t_{time_depth} = t'_{time_depth} + 1
143 | t_{time_depth+1} = t'_{time_depth + 1}
145 | t_{dim_sctr} = t'_{dim_sctr}. */
147 /* Add the equality: t_{time_depth} = t'_{time_depth} + 1.
148 This is the core part of this alogrithm, since this
149 constraint asks for the memory access stride (difference)
150 between two consecutive points in time dimensions. */
155 | t_{time_depth-1} = t'_{time_depth-1}
156 | t_{time_depth+1} = t'_{time_depth+1}
158 | t_{dim_sctr} = t'_{dim_sctr}
160 This means that all the time dimensions are equal except for
161 time_depth, where the constraint is t_{depth} = t'_{depth} + 1
162 step. More to this: we should be careful not to add equalities
163 to the 'coupled' dimensions, which happens when the one dimension
164 is stripmined dimension, and the other dimension corresponds
165 to the point loop inside stripmined dimension. */
167 /* pdr->accesses: [P1..nb_param,I1..nb_domain]->[a,S1..nb_subscript]
168 ??? [P] not used for PDRs?
169 pdr->extent: [a,S1..nb_subscript]
170 pbb->domain: [P1..nb_param,I1..nb_domain]
171 pbb->transformed: [P1..nb_param,I1..nb_domain]->[T1..Tnb_sctr]
172 [T] includes local vars (currently unused)
174 First we create [P,I] -> [T,a,S]. */
176 map
= isl_map_flat_range_product (isl_map_copy (pbb
->transformed
),
177 isl_map_copy (pdr
->accesses
));
178 /* Add a dimension for L: [P,I] -> [T,a,S,L].*/
179 map
= isl_map_add_dims (map
, isl_dim_out
, 1);
180 /* Build a constraint for "lma[S] - L == 0", effectively calculating
181 L in terms of subscripts. */
182 lma
= build_linearized_memory_access (map
, pdr
);
183 /* And add it to the map, so we now have:
184 [P,I] -> [T,a,S,L] : lma([S]) == L. */
185 map
= isl_map_add_constraint (map
, lma
);
187 /* Then we create [P,I,P',I'] -> [T,a,S,L,T',a',S',L']. */
188 map
= isl_map_flat_product (map
, isl_map_copy (map
));
190 /* Now add the equality T[time_depth] == T'[time_depth]+1. This will
191 force L' to be the linear address at T[time_depth] + 1. */
192 time_depth
= psct_dynamic_dim (pbb
, depth
);
193 /* Length of [a,S] plus [L] ... */
194 offset
= 1 + isl_map_dim (pdr
->accesses
, isl_dim_out
);
196 offset
+= isl_map_dim (pbb
->transformed
, isl_dim_out
);
198 c
= isl_equality_alloc (isl_local_space_from_space (isl_map_get_space (map
)));
199 c
= isl_constraint_set_coefficient_si (c
, isl_dim_out
, time_depth
, 1);
200 c
= isl_constraint_set_coefficient_si (c
, isl_dim_out
,
201 offset
+ time_depth
, -1);
202 c
= isl_constraint_set_constant_si (c
, 1);
203 map
= isl_map_add_constraint (map
, c
);
205 /* Now we equate most of the T/T' elements (making PITaSL nearly
206 the same is (PITaSL)', except for one dimension, namely for 'depth'
207 (an index into [I]), after translating to index into [T]. Take care
208 to not produce an empty map, which indicates we wanted to equate
209 two dimensions that are already coupled via the above time_depth
210 dimension. Happens with strip mining where several scatter dimension
211 are interdependend. */
213 nt
= pbb_nb_scattering_transform (pbb
) + pbb_nb_local_vars (pbb
);
214 for (i
= 0; i
< nt
; i
++)
217 isl_map
*temp
= isl_map_equate (isl_map_copy (map
),
219 isl_dim_out
, offset
+ i
);
220 if (isl_map_is_empty (temp
))
229 /* Now maximize the expression L' - L. */
230 set
= isl_map_range (map
);
231 dc
= isl_set_get_space (set
);
232 aff
= isl_aff_zero_on_domain (isl_local_space_from_space (dc
));
233 aff
= isl_aff_set_coefficient_si (aff
, isl_dim_in
, offset
- 1, -1);
234 aff
= isl_aff_set_coefficient_si (aff
, isl_dim_in
, offset
+ offset
- 1, 1);
235 isl_int_init (islstride
);
236 isl_set_max (set
, aff
, &islstride
);
237 isl_int_get_gmp (islstride
, stride
);
238 isl_int_clear (islstride
);
242 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
244 gmp_fprintf (dump_file
, "\nStride in BB_%d, DR_%d, depth %d: %Zd ",
245 pbb_index (pbb
), PDR_ID (pdr
), (int) depth
, stride
);
249 /* Sets STRIDES to the sum of all the strides of the data references
250 accessed in LOOP at DEPTH. */
253 memory_strides_in_loop_1 (lst_p loop
, graphite_dim_t depth
, mpz_t strides
)
263 FOR_EACH_VEC_ELT (LST_SEQ (loop
), j
, l
)
265 memory_strides_in_loop_1 (l
, depth
, strides
);
267 FOR_EACH_VEC_ELT (PBB_DRS (LST_PBB (l
)), i
, pdr
)
269 pdr_stride_in_loop (s
, depth
, pdr
);
270 mpz_set_si (n
, PDR_NB_REFS (pdr
));
272 mpz_add (strides
, strides
, s
);
279 /* Sets STRIDES to the sum of all the strides of the data references
280 accessed in LOOP at DEPTH. */
283 memory_strides_in_loop (lst_p loop
, graphite_dim_t depth
, mpz_t strides
)
285 if (mpz_cmp_si (loop
->memory_strides
, -1) == 0)
287 mpz_set_si (strides
, 0);
288 memory_strides_in_loop_1 (loop
, depth
, strides
);
291 mpz_set (strides
, loop
->memory_strides
);
294 /* Return true when the interchange of loops LOOP1 and LOOP2 is
307 | for (i = 0; i < N; i++)
308 | for (j = 0; j < N; j++)
314 The data access A[j][i] is described like this:
322 | 0 0 0 0 -1 0 100 >= 0
323 | 0 0 0 0 0 -1 100 >= 0
325 The linearized memory access L to A[100][100] is:
330 TODO: the shown format is not valid as it does not show the fact
331 that the iteration domain "i j" is transformed using the scattering.
333 Next, to measure the impact of iterating once in loop "i", we build
334 a maximization problem: first, we add to DR accesses the dimensions
335 k, s2, s3, L1 = 100 * s0 + s1, L2, and D1: this is the polyhedron P1.
336 L1 and L2 are the linearized memory access functions.
338 | i j N a s0 s1 k s2 s3 L1 L2 D1 1
339 | 0 0 0 1 0 0 0 0 0 0 0 0 -5 = 0 alias = 5
340 | 0 -1 0 0 1 0 0 0 0 0 0 0 0 = 0 s0 = j
341 |-2 0 0 0 0 1 0 0 0 0 0 0 0 = 0 s1 = 2 * i
342 | 0 0 0 0 1 0 0 0 0 0 0 0 0 >= 0
343 | 0 0 0 0 0 1 0 0 0 0 0 0 0 >= 0
344 | 0 0 0 0 -1 0 0 0 0 0 0 0 100 >= 0
345 | 0 0 0 0 0 -1 0 0 0 0 0 0 100 >= 0
346 | 0 0 0 0 100 1 0 0 0 -1 0 0 0 = 0 L1 = 100 * s0 + s1
348 Then, we generate the polyhedron P2 by interchanging the dimensions
349 (s0, s2), (s1, s3), (L1, L2), (k, i)
351 | i j N a s0 s1 k s2 s3 L1 L2 D1 1
352 | 0 0 0 1 0 0 0 0 0 0 0 0 -5 = 0 alias = 5
353 | 0 -1 0 0 0 0 0 1 0 0 0 0 0 = 0 s2 = j
354 | 0 0 0 0 0 0 -2 0 1 0 0 0 0 = 0 s3 = 2 * k
355 | 0 0 0 0 0 0 0 1 0 0 0 0 0 >= 0
356 | 0 0 0 0 0 0 0 0 1 0 0 0 0 >= 0
357 | 0 0 0 0 0 0 0 -1 0 0 0 0 100 >= 0
358 | 0 0 0 0 0 0 0 0 -1 0 0 0 100 >= 0
359 | 0 0 0 0 0 0 0 100 1 0 -1 0 0 = 0 L2 = 100 * s2 + s3
361 then we add to P2 the equality k = i + 1:
363 |-1 0 0 0 0 0 1 0 0 0 0 0 -1 = 0 k = i + 1
365 and finally we maximize the expression "D1 = max (P1 inter P2, L2 - L1)".
367 Similarly, to determine the impact of one iteration on loop "j", we
368 interchange (k, j), we add "k = j + 1", and we compute D2 the
369 maximal value of the difference.
371 Finally, the profitability test is D1 < D2: if in the outer loop
372 the strides are smaller than in the inner loop, then it is
373 profitable to interchange the loops at DEPTH1 and DEPTH2. */
376 lst_interchange_profitable_p (lst_p nest
, int depth1
, int depth2
)
381 gcc_assert (depth1
< depth2
);
386 memory_strides_in_loop (nest
, depth1
, d1
);
387 memory_strides_in_loop (nest
, depth2
, d2
);
389 res
= mpz_cmp (d1
, d2
) < 0;
397 /* Interchanges the loops at DEPTH1 and DEPTH2 of the original
398 scattering and assigns the resulting polyhedron to the transformed
402 pbb_interchange_loop_depths (graphite_dim_t depth1
, graphite_dim_t depth2
,
406 unsigned dim1
= psct_dynamic_dim (pbb
, depth1
);
407 unsigned dim2
= psct_dynamic_dim (pbb
, depth2
);
408 isl_space
*d
= isl_map_get_space (pbb
->transformed
);
409 isl_space
*d1
= isl_space_range (d
);
410 unsigned n
= isl_space_dim (d1
, isl_dim_out
);
411 isl_space
*d2
= isl_space_add_dims (d1
, isl_dim_in
, n
);
412 isl_map
*x
= isl_map_universe (d2
);
414 x
= isl_map_equate (x
, isl_dim_in
, dim1
, isl_dim_out
, dim2
);
415 x
= isl_map_equate (x
, isl_dim_in
, dim2
, isl_dim_out
, dim1
);
417 for (i
= 0; i
< n
; i
++)
418 if (i
!= dim1
&& i
!= dim2
)
419 x
= isl_map_equate (x
, isl_dim_in
, i
, isl_dim_out
, i
);
421 pbb
->transformed
= isl_map_apply_range (pbb
->transformed
, x
);
424 /* Apply the interchange of loops at depths DEPTH1 and DEPTH2 to all
425 the statements below LST. */
428 lst_apply_interchange (lst_p lst
, int depth1
, int depth2
)
433 if (LST_LOOP_P (lst
))
438 FOR_EACH_VEC_ELT (LST_SEQ (lst
), i
, l
)
439 lst_apply_interchange (l
, depth1
, depth2
);
442 pbb_interchange_loop_depths (depth1
, depth2
, LST_PBB (lst
));
445 /* Return true when the nest starting at LOOP1 and ending on LOOP2 is
446 perfect: i.e. there are no sequence of statements. */
449 lst_perfectly_nested_p (lst_p loop1
, lst_p loop2
)
454 if (!LST_LOOP_P (loop1
))
457 return LST_SEQ (loop1
).length () == 1
458 && lst_perfectly_nested_p (LST_SEQ (loop1
)[0], loop2
);
461 /* Transform the loop nest between LOOP1 and LOOP2 into a perfect
462 nest. To continue the naming tradition, this function is called
463 after perfect_nestify. NEST is set to the perfectly nested loop
464 that is created. BEFORE/AFTER are set to the loops distributed
465 before/after the loop NEST. */
468 lst_perfect_nestify (lst_p loop1
, lst_p loop2
, lst_p
*before
,
469 lst_p
*nest
, lst_p
*after
)
471 poly_bb_p first
, last
;
473 gcc_assert (loop1
&& loop2
475 && LST_LOOP_P (loop1
) && LST_LOOP_P (loop2
));
477 first
= LST_PBB (lst_find_first_pbb (loop2
));
478 last
= LST_PBB (lst_find_last_pbb (loop2
));
480 *before
= copy_lst (loop1
);
481 *nest
= copy_lst (loop1
);
482 *after
= copy_lst (loop1
);
484 lst_remove_all_before_including_pbb (*before
, first
, false);
485 lst_remove_all_before_including_pbb (*after
, last
, true);
487 lst_remove_all_before_excluding_pbb (*nest
, first
, true);
488 lst_remove_all_before_excluding_pbb (*nest
, last
, false);
490 if (lst_empty_p (*before
))
495 if (lst_empty_p (*after
))
500 if (lst_empty_p (*nest
))
507 /* Try to interchange LOOP1 with LOOP2 for all the statements of the
508 body of LOOP2. LOOP1 contains LOOP2. Return true if it did the
512 lst_try_interchange_loops (scop_p scop
, lst_p loop1
, lst_p loop2
)
514 int depth1
= lst_depth (loop1
);
515 int depth2
= lst_depth (loop2
);
518 lst_p before
= NULL
, nest
= NULL
, after
= NULL
;
520 if (!lst_perfectly_nested_p (loop1
, loop2
))
521 lst_perfect_nestify (loop1
, loop2
, &before
, &nest
, &after
);
523 if (!lst_interchange_profitable_p (loop2
, depth1
, depth2
))
526 lst_apply_interchange (loop2
, depth1
, depth2
);
528 /* Sync the transformed LST information and the PBB scatterings
529 before using the scatterings in the data dependence analysis. */
530 if (before
|| nest
|| after
)
532 transformed
= lst_substitute_3 (SCOP_TRANSFORMED_SCHEDULE (scop
), loop1
,
533 before
, nest
, after
);
534 lst_update_scattering (transformed
);
535 free_lst (transformed
);
538 if (graphite_legal_transform (scop
))
540 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
542 "Loops at depths %d and %d will be interchanged.\n",
545 /* Transform the SCOP_TRANSFORMED_SCHEDULE of the SCOP. */
546 lst_insert_in_sequence (before
, loop1
, true);
547 lst_insert_in_sequence (after
, loop1
, false);
551 lst_replace (loop1
, nest
);
558 /* Undo the transform. */
562 lst_apply_interchange (loop2
, depth2
, depth1
);
566 /* Selects the inner loop in LST_SEQ (INNER_FATHER) to be interchanged
567 with the loop OUTER in LST_SEQ (OUTER_FATHER). */
570 lst_interchange_select_inner (scop_p scop
, lst_p outer_father
, int outer
,
576 gcc_assert (outer_father
577 && LST_LOOP_P (outer_father
)
578 && LST_LOOP_P (LST_SEQ (outer_father
)[outer
])
580 && LST_LOOP_P (inner_father
));
582 loop1
= LST_SEQ (outer_father
)[outer
];
584 FOR_EACH_VEC_ELT (LST_SEQ (inner_father
), inner
, loop2
)
585 if (LST_LOOP_P (loop2
)
586 && (lst_try_interchange_loops (scop
, loop1
, loop2
)
587 || lst_interchange_select_inner (scop
, outer_father
, outer
, loop2
)))
593 /* Interchanges all the loops of LOOP and the loops of its body that
594 are considered profitable to interchange. Return the number of
595 interchanged loops. OUTER is the index in LST_SEQ (LOOP) that
596 points to the next outer loop to be considered for interchange. */
599 lst_interchange_select_outer (scop_p scop
, lst_p loop
, int outer
)
606 if (!loop
|| !LST_LOOP_P (loop
))
609 father
= LST_LOOP_FATHER (loop
);
612 while (lst_interchange_select_inner (scop
, father
, outer
, loop
))
615 loop
= LST_SEQ (father
)[outer
];
619 if (LST_LOOP_P (loop
))
620 FOR_EACH_VEC_ELT (LST_SEQ (loop
), i
, l
)
622 res
+= lst_interchange_select_outer (scop
, l
, i
);
627 /* Interchanges all the loop depths that are considered profitable for
628 SCOP. Return the number of interchanged loops. */
631 scop_do_interchange (scop_p scop
)
633 int res
= lst_interchange_select_outer
634 (scop
, SCOP_TRANSFORMED_SCHEDULE (scop
), 0);
636 lst_update_scattering (SCOP_TRANSFORMED_SCHEDULE (scop
));