* config/rs6000/rs6000.c (rs6000_mangle_fundamental_type): Mangle
[official-gcc.git] / gcc / tree-data-ref.c
blob6f34adacd8afe492941e94f32bdd2326d9017198
1 /* Data references and dependences detectors.
2 Copyright (C) 2003, 2004, 2005 Free Software Foundation, Inc.
3 Contributed by Sebastian Pop <s.pop@laposte.net>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 2, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
20 02110-1301, USA. */
22 /* This pass walks a given loop structure searching for array
23 references. The information about the array accesses is recorded
24 in DATA_REFERENCE structures.
26 The basic test for determining the dependences is:
27 given two access functions chrec1 and chrec2 to a same array, and
28 x and y two vectors from the iteration domain, the same element of
29 the array is accessed twice at iterations x and y if and only if:
30 | chrec1 (x) == chrec2 (y).
32 The goals of this analysis are:
34 - to determine the independence: the relation between two
35 independent accesses is qualified with the chrec_known (this
36 information allows a loop parallelization),
38 - when two data references access the same data, to qualify the
39 dependence relation with classic dependence representations:
41 - distance vectors
42 - direction vectors
43 - loop carried level dependence
44 - polyhedron dependence
45 or with the chains of recurrences based representation,
47 - to define a knowledge base for storing the data dependence
48 information,
50 - to define an interface to access this data.
53 Definitions:
55 - subscript: given two array accesses a subscript is the tuple
56 composed of the access functions for a given dimension. Example:
57 Given A[f1][f2][f3] and B[g1][g2][g3], there are three subscripts:
58 (f1, g1), (f2, g2), (f3, g3).
60 - Diophantine equation: an equation whose coefficients and
61 solutions are integer constants, for example the equation
62 | 3*x + 2*y = 1
63 has an integer solution x = 1 and y = -1.
65 References:
67 - "Advanced Compilation for High Performance Computing" by Randy
68 Allen and Ken Kennedy.
69 http://citeseer.ist.psu.edu/goff91practical.html
71 - "Loop Transformations for Restructuring Compilers - The Foundations"
72 by Utpal Banerjee.
77 #include "config.h"
78 #include "system.h"
79 #include "coretypes.h"
80 #include "tm.h"
81 #include "ggc.h"
82 #include "tree.h"
84 /* These RTL headers are needed for basic-block.h. */
85 #include "rtl.h"
86 #include "basic-block.h"
87 #include "diagnostic.h"
88 #include "tree-flow.h"
89 #include "tree-dump.h"
90 #include "timevar.h"
91 #include "cfgloop.h"
92 #include "tree-chrec.h"
93 #include "tree-data-ref.h"
94 #include "tree-scalar-evolution.h"
95 #include "tree-pass.h"
97 static tree object_analysis (tree, tree, bool, struct data_reference **,
98 tree *, tree *, tree *, tree *, tree *,
99 struct ptr_info_def **, subvar_t *);
100 static struct data_reference * init_data_ref (tree, tree, tree, tree, bool,
101 tree, tree, tree, tree, tree,
102 struct ptr_info_def *,
103 enum data_ref_type);
105 /* Determine if PTR and DECL may alias, the result is put in ALIASED.
106 Return FALSE if there is no type memory tag for PTR.
108 static bool
109 ptr_decl_may_alias_p (tree ptr, tree decl,
110 struct data_reference *ptr_dr,
111 bool *aliased)
113 tree tag;
115 gcc_assert (TREE_CODE (ptr) == SSA_NAME && DECL_P (decl));
117 tag = get_var_ann (SSA_NAME_VAR (ptr))->type_mem_tag;
118 if (!tag)
119 tag = DR_MEMTAG (ptr_dr);
120 if (!tag)
121 return false;
123 *aliased = is_aliased_with (tag, decl);
124 return true;
128 /* Determine if two pointers may alias, the result is put in ALIASED.
129 Return FALSE if there is no type memory tag for one of the pointers.
131 static bool
132 ptr_ptr_may_alias_p (tree ptr_a, tree ptr_b,
133 struct data_reference *dra,
134 struct data_reference *drb,
135 bool *aliased)
137 tree tag_a, tag_b;
139 tag_a = get_var_ann (SSA_NAME_VAR (ptr_a))->type_mem_tag;
140 if (!tag_a)
141 tag_a = DR_MEMTAG (dra);
142 if (!tag_a)
143 return false;
144 tag_b = get_var_ann (SSA_NAME_VAR (ptr_b))->type_mem_tag;
145 if (!tag_b)
146 tag_b = DR_MEMTAG (drb);
147 if (!tag_b)
148 return false;
149 *aliased = (tag_a == tag_b);
150 return true;
154 /* Determine if BASE_A and BASE_B may alias, the result is put in ALIASED.
155 Return FALSE if there is no type memory tag for one of the symbols.
157 static bool
158 may_alias_p (tree base_a, tree base_b,
159 struct data_reference *dra,
160 struct data_reference *drb,
161 bool *aliased)
163 if (TREE_CODE (base_a) == ADDR_EXPR || TREE_CODE (base_b) == ADDR_EXPR)
165 if (TREE_CODE (base_a) == ADDR_EXPR && TREE_CODE (base_b) == ADDR_EXPR)
167 *aliased = (TREE_OPERAND (base_a, 0) == TREE_OPERAND (base_b, 0));
168 return true;
170 if (TREE_CODE (base_a) == ADDR_EXPR)
171 return ptr_decl_may_alias_p (base_b, TREE_OPERAND (base_a, 0), drb,
172 aliased);
173 else
174 return ptr_decl_may_alias_p (base_a, TREE_OPERAND (base_b, 0), dra,
175 aliased);
178 return ptr_ptr_may_alias_p (base_a, base_b, dra, drb, aliased);
182 /* Determine if a pointer (BASE_A) and a record/union access (BASE_B)
183 are not aliased. Return TRUE if they differ. */
184 static bool
185 record_ptr_differ_p (struct data_reference *dra,
186 struct data_reference *drb)
188 bool aliased;
189 tree base_a = DR_BASE_OBJECT (dra);
190 tree base_b = DR_BASE_OBJECT (drb);
192 if (TREE_CODE (base_b) != COMPONENT_REF)
193 return false;
195 /* Peel COMPONENT_REFs to get to the base. Do not peel INDIRECT_REFs.
196 For a.b.c.d[i] we will get a, and for a.b->c.d[i] we will get a.b.
197 Probably will be unnecessary with struct alias analysis. */
198 while (TREE_CODE (base_b) == COMPONENT_REF)
199 base_b = TREE_OPERAND (base_b, 0);
200 /* Compare a record/union access (b.c[i] or p->c[i]) and a pointer
201 ((*q)[i]). */
202 if (TREE_CODE (base_a) == INDIRECT_REF
203 && ((TREE_CODE (base_b) == VAR_DECL
204 && (ptr_decl_may_alias_p (TREE_OPERAND (base_a, 0), base_b, dra,
205 &aliased)
206 && !aliased))
207 || (TREE_CODE (base_b) == INDIRECT_REF
208 && (ptr_ptr_may_alias_p (TREE_OPERAND (base_a, 0),
209 TREE_OPERAND (base_b, 0), dra, drb,
210 &aliased)
211 && !aliased))))
212 return true;
213 else
214 return false;
218 /* Determine if an array access (BASE_A) and a record/union access (BASE_B)
219 are not aliased. Return TRUE if they differ. */
220 static bool
221 record_array_differ_p (struct data_reference *dra,
222 struct data_reference *drb)
224 bool aliased;
225 tree base_a = DR_BASE_OBJECT (dra);
226 tree base_b = DR_BASE_OBJECT (drb);
228 if (TREE_CODE (base_b) != COMPONENT_REF)
229 return false;
231 /* Peel COMPONENT_REFs to get to the base. Do not peel INDIRECT_REFs.
232 For a.b.c.d[i] we will get a, and for a.b->c.d[i] we will get a.b.
233 Probably will be unnecessary with struct alias analysis. */
234 while (TREE_CODE (base_b) == COMPONENT_REF)
235 base_b = TREE_OPERAND (base_b, 0);
237 /* Compare a record/union access (b.c[i] or p->c[i]) and an array access
238 (a[i]). In case of p->c[i] use alias analysis to verify that p is not
239 pointing to a. */
240 if (TREE_CODE (base_a) == VAR_DECL
241 && (TREE_CODE (base_b) == VAR_DECL
242 || (TREE_CODE (base_b) == INDIRECT_REF
243 && (ptr_decl_may_alias_p (TREE_OPERAND (base_b, 0), base_a, drb,
244 &aliased)
245 && !aliased))))
246 return true;
247 else
248 return false;
252 /* Determine if an array access (BASE_A) and a pointer (BASE_B)
253 are not aliased. Return TRUE if they differ. */
254 static bool
255 array_ptr_differ_p (tree base_a, tree base_b,
256 struct data_reference *drb)
258 bool aliased;
260 /* In case one of the bases is a pointer (a[i] and (*p)[i]), we check with the
261 help of alias analysis that p is not pointing to a. */
262 if (TREE_CODE (base_a) == VAR_DECL && TREE_CODE (base_b) == INDIRECT_REF
263 && (ptr_decl_may_alias_p (TREE_OPERAND (base_b, 0), base_a, drb, &aliased)
264 && !aliased))
265 return true;
266 else
267 return false;
271 /* This is the simplest data dependence test: determines whether the
272 data references A and B access the same array/region. Returns
273 false when the property is not computable at compile time.
274 Otherwise return true, and DIFFER_P will record the result. This
275 utility will not be necessary when alias_sets_conflict_p will be
276 less conservative. */
278 static bool
279 base_object_differ_p (struct data_reference *a,
280 struct data_reference *b,
281 bool *differ_p)
283 tree base_a = DR_BASE_OBJECT (a);
284 tree base_b = DR_BASE_OBJECT (b);
285 bool aliased;
287 if (!base_a || !base_b)
288 return false;
290 /* Determine if same base. Example: for the array accesses
291 a[i], b[i] or pointer accesses *a, *b, bases are a, b. */
292 if (base_a == base_b)
294 *differ_p = false;
295 return true;
298 /* For pointer based accesses, (*p)[i], (*q)[j], the bases are (*p)
299 and (*q) */
300 if (TREE_CODE (base_a) == INDIRECT_REF && TREE_CODE (base_b) == INDIRECT_REF
301 && TREE_OPERAND (base_a, 0) == TREE_OPERAND (base_b, 0))
303 *differ_p = false;
304 return true;
307 /* Record/union based accesses - s.a[i], t.b[j]. bases are s.a,t.b. */
308 if (TREE_CODE (base_a) == COMPONENT_REF && TREE_CODE (base_b) == COMPONENT_REF
309 && TREE_OPERAND (base_a, 0) == TREE_OPERAND (base_b, 0)
310 && TREE_OPERAND (base_a, 1) == TREE_OPERAND (base_b, 1))
312 *differ_p = false;
313 return true;
317 /* Determine if different bases. */
319 /* At this point we know that base_a != base_b. However, pointer
320 accesses of the form x=(*p) and y=(*q), whose bases are p and q,
321 may still be pointing to the same base. In SSAed GIMPLE p and q will
322 be SSA_NAMES in this case. Therefore, here we check if they are
323 really two different declarations. */
324 if (TREE_CODE (base_a) == VAR_DECL && TREE_CODE (base_b) == VAR_DECL)
326 *differ_p = true;
327 return true;
330 /* In case one of the bases is a pointer (a[i] and (*p)[i]), we check with the
331 help of alias analysis that p is not pointing to a. */
332 if (array_ptr_differ_p (base_a, base_b, b)
333 || array_ptr_differ_p (base_b, base_a, a))
335 *differ_p = true;
336 return true;
339 /* If the bases are pointers ((*q)[i] and (*p)[i]), we check with the
340 help of alias analysis they don't point to the same bases. */
341 if (TREE_CODE (base_a) == INDIRECT_REF && TREE_CODE (base_b) == INDIRECT_REF
342 && (may_alias_p (TREE_OPERAND (base_a, 0), TREE_OPERAND (base_b, 0), a, b,
343 &aliased)
344 && !aliased))
346 *differ_p = true;
347 return true;
350 /* Compare two record/union bases s.a and t.b: s != t or (a != b and
351 s and t are not unions). */
352 if (TREE_CODE (base_a) == COMPONENT_REF && TREE_CODE (base_b) == COMPONENT_REF
353 && ((TREE_CODE (TREE_OPERAND (base_a, 0)) == VAR_DECL
354 && TREE_CODE (TREE_OPERAND (base_b, 0)) == VAR_DECL
355 && TREE_OPERAND (base_a, 0) != TREE_OPERAND (base_b, 0))
356 || (TREE_CODE (TREE_TYPE (TREE_OPERAND (base_a, 0))) == RECORD_TYPE
357 && TREE_CODE (TREE_TYPE (TREE_OPERAND (base_b, 0))) == RECORD_TYPE
358 && TREE_OPERAND (base_a, 1) != TREE_OPERAND (base_b, 1))))
360 *differ_p = true;
361 return true;
364 /* Compare a record/union access (b.c[i] or p->c[i]) and a pointer
365 ((*q)[i]). */
366 if (record_ptr_differ_p (a, b) || record_ptr_differ_p (b, a))
368 *differ_p = true;
369 return true;
372 /* Compare a record/union access (b.c[i] or p->c[i]) and an array access
373 (a[i]). In case of p->c[i] use alias analysis to verify that p is not
374 pointing to a. */
375 if (record_array_differ_p (a, b) || record_array_differ_p (b, a))
377 *differ_p = true;
378 return true;
381 return false;
384 /* Function base_addr_differ_p.
386 This is the simplest data dependence test: determines whether the
387 data references DRA and DRB access the same array/region. Returns
388 false when the property is not computable at compile time.
389 Otherwise return true, and DIFFER_P will record the result.
391 The algorithm:
392 1. if (both DRA and DRB are represented as arrays)
393 compare DRA.BASE_OBJECT and DRB.BASE_OBJECT
394 2. else if (both DRA and DRB are represented as pointers)
395 try to prove that DRA.FIRST_LOCATION == DRB.FIRST_LOCATION
396 3. else if (DRA and DRB are represented differently or 2. fails)
397 only try to prove that the bases are surely different
401 static bool
402 base_addr_differ_p (struct data_reference *dra,
403 struct data_reference *drb,
404 bool *differ_p)
406 tree addr_a = DR_BASE_ADDRESS (dra);
407 tree addr_b = DR_BASE_ADDRESS (drb);
408 tree type_a, type_b;
409 bool aliased;
411 if (!addr_a || !addr_b)
412 return false;
414 type_a = TREE_TYPE (addr_a);
415 type_b = TREE_TYPE (addr_b);
417 gcc_assert (POINTER_TYPE_P (type_a) && POINTER_TYPE_P (type_b));
419 /* 1. if (both DRA and DRB are represented as arrays)
420 compare DRA.BASE_OBJECT and DRB.BASE_OBJECT. */
421 if (DR_TYPE (dra) == ARRAY_REF_TYPE && DR_TYPE (drb) == ARRAY_REF_TYPE)
422 return base_object_differ_p (dra, drb, differ_p);
425 /* 2. else if (both DRA and DRB are represented as pointers)
426 try to prove that DRA.FIRST_LOCATION == DRB.FIRST_LOCATION. */
427 /* If base addresses are the same, we check the offsets, since the access of
428 the data-ref is described by {base addr + offset} and its access function,
429 i.e., in order to decide whether the bases of data-refs are the same we
430 compare both base addresses and offsets. */
431 if (DR_TYPE (dra) == POINTER_REF_TYPE && DR_TYPE (drb) == POINTER_REF_TYPE
432 && (addr_a == addr_b
433 || (TREE_CODE (addr_a) == ADDR_EXPR && TREE_CODE (addr_b) == ADDR_EXPR
434 && TREE_OPERAND (addr_a, 0) == TREE_OPERAND (addr_b, 0))))
436 /* Compare offsets. */
437 tree offset_a = DR_OFFSET (dra);
438 tree offset_b = DR_OFFSET (drb);
440 STRIP_NOPS (offset_a);
441 STRIP_NOPS (offset_b);
443 /* FORNOW: we only compare offsets that are MULT_EXPR, i.e., we don't handle
444 PLUS_EXPR. */
445 if ((offset_a == offset_b)
446 || (TREE_CODE (offset_a) == MULT_EXPR
447 && TREE_CODE (offset_b) == MULT_EXPR
448 && TREE_OPERAND (offset_a, 0) == TREE_OPERAND (offset_b, 0)
449 && TREE_OPERAND (offset_a, 1) == TREE_OPERAND (offset_b, 1)))
451 *differ_p = false;
452 return true;
456 /* 3. else if (DRA and DRB are represented differently or 2. fails)
457 only try to prove that the bases are surely different. */
459 /* Apply alias analysis. */
460 if (may_alias_p (addr_a, addr_b, dra, drb, &aliased) && !aliased)
462 *differ_p = true;
463 return true;
466 /* An instruction writing through a restricted pointer is "independent" of any
467 instruction reading or writing through a different pointer, in the same
468 block/scope. */
469 else if ((TYPE_RESTRICT (type_a) && !DR_IS_READ (dra))
470 || (TYPE_RESTRICT (type_b) && !DR_IS_READ (drb)))
472 *differ_p = true;
473 return true;
475 return false;
479 /* Returns true iff A divides B. */
481 static inline bool
482 tree_fold_divides_p (tree a,
483 tree b)
485 /* Determines whether (A == gcd (A, B)). */
486 return tree_int_cst_equal (a, tree_fold_gcd (a, b));
489 /* Compute the greatest common denominator of two numbers using
490 Euclid's algorithm. */
492 static int
493 gcd (int a, int b)
496 int x, y, z;
498 x = abs (a);
499 y = abs (b);
501 while (x>0)
503 z = y % x;
504 y = x;
505 x = z;
508 return (y);
511 /* Returns true iff A divides B. */
513 static inline bool
514 int_divides_p (int a, int b)
516 return ((b % a) == 0);
521 /* Dump into FILE all the data references from DATAREFS. */
523 void
524 dump_data_references (FILE *file,
525 varray_type datarefs)
527 unsigned int i;
529 for (i = 0; i < VARRAY_ACTIVE_SIZE (datarefs); i++)
530 dump_data_reference (file, VARRAY_GENERIC_PTR (datarefs, i));
533 /* Dump into FILE all the dependence relations from DDR. */
535 void
536 dump_data_dependence_relations (FILE *file,
537 varray_type ddr)
539 unsigned int i;
541 for (i = 0; i < VARRAY_ACTIVE_SIZE (ddr); i++)
542 dump_data_dependence_relation (file, VARRAY_GENERIC_PTR (ddr, i));
545 /* Dump function for a DATA_REFERENCE structure. */
547 void
548 dump_data_reference (FILE *outf,
549 struct data_reference *dr)
551 unsigned int i;
553 fprintf (outf, "(Data Ref: \n stmt: ");
554 print_generic_stmt (outf, DR_STMT (dr), 0);
555 fprintf (outf, " ref: ");
556 print_generic_stmt (outf, DR_REF (dr), 0);
557 fprintf (outf, " base_name: ");
558 print_generic_stmt (outf, DR_BASE_OBJECT (dr), 0);
560 for (i = 0; i < DR_NUM_DIMENSIONS (dr); i++)
562 fprintf (outf, " Access function %d: ", i);
563 print_generic_stmt (outf, DR_ACCESS_FN (dr, i), 0);
565 fprintf (outf, ")\n");
568 /* Dump function for a SUBSCRIPT structure. */
570 void
571 dump_subscript (FILE *outf, struct subscript *subscript)
573 tree chrec = SUB_CONFLICTS_IN_A (subscript);
575 fprintf (outf, "\n (subscript \n");
576 fprintf (outf, " iterations_that_access_an_element_twice_in_A: ");
577 print_generic_stmt (outf, chrec, 0);
578 if (chrec == chrec_known)
579 fprintf (outf, " (no dependence)\n");
580 else if (chrec_contains_undetermined (chrec))
581 fprintf (outf, " (don't know)\n");
582 else
584 tree last_iteration = SUB_LAST_CONFLICT (subscript);
585 fprintf (outf, " last_conflict: ");
586 print_generic_stmt (outf, last_iteration, 0);
589 chrec = SUB_CONFLICTS_IN_B (subscript);
590 fprintf (outf, " iterations_that_access_an_element_twice_in_B: ");
591 print_generic_stmt (outf, chrec, 0);
592 if (chrec == chrec_known)
593 fprintf (outf, " (no dependence)\n");
594 else if (chrec_contains_undetermined (chrec))
595 fprintf (outf, " (don't know)\n");
596 else
598 tree last_iteration = SUB_LAST_CONFLICT (subscript);
599 fprintf (outf, " last_conflict: ");
600 print_generic_stmt (outf, last_iteration, 0);
603 fprintf (outf, " (Subscript distance: ");
604 print_generic_stmt (outf, SUB_DISTANCE (subscript), 0);
605 fprintf (outf, " )\n");
606 fprintf (outf, " )\n");
609 /* Dump function for a DATA_DEPENDENCE_RELATION structure. */
611 void
612 dump_data_dependence_relation (FILE *outf,
613 struct data_dependence_relation *ddr)
615 struct data_reference *dra, *drb;
617 dra = DDR_A (ddr);
618 drb = DDR_B (ddr);
619 fprintf (outf, "(Data Dep: \n");
620 if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
621 fprintf (outf, " (don't know)\n");
623 else if (DDR_ARE_DEPENDENT (ddr) == chrec_known)
624 fprintf (outf, " (no dependence)\n");
626 else if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
628 unsigned int i;
630 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
632 fprintf (outf, " access_fn_A: ");
633 print_generic_stmt (outf, DR_ACCESS_FN (dra, i), 0);
634 fprintf (outf, " access_fn_B: ");
635 print_generic_stmt (outf, DR_ACCESS_FN (drb, i), 0);
636 dump_subscript (outf, DDR_SUBSCRIPT (ddr, i));
639 for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++)
641 fprintf (outf, " distance_vector: ");
642 print_lambda_vector (outf, DDR_DIST_VECT (ddr, i),
643 DDR_SIZE_VECT (ddr));
646 for (i = 0; i < DDR_NUM_DIR_VECTS (ddr); i++)
648 fprintf (outf, " direction_vector: ");
649 print_lambda_vector (outf, DDR_DIR_VECT (ddr, i),
650 DDR_SIZE_VECT (ddr));
654 fprintf (outf, ")\n");
659 /* Dump function for a DATA_DEPENDENCE_DIRECTION structure. */
661 void
662 dump_data_dependence_direction (FILE *file,
663 enum data_dependence_direction dir)
665 switch (dir)
667 case dir_positive:
668 fprintf (file, "+");
669 break;
671 case dir_negative:
672 fprintf (file, "-");
673 break;
675 case dir_equal:
676 fprintf (file, "=");
677 break;
679 case dir_positive_or_negative:
680 fprintf (file, "+-");
681 break;
683 case dir_positive_or_equal:
684 fprintf (file, "+=");
685 break;
687 case dir_negative_or_equal:
688 fprintf (file, "-=");
689 break;
691 case dir_star:
692 fprintf (file, "*");
693 break;
695 default:
696 break;
700 /* Dumps the distance and direction vectors in FILE. DDRS contains
701 the dependence relations, and VECT_SIZE is the size of the
702 dependence vectors, or in other words the number of loops in the
703 considered nest. */
705 void
706 dump_dist_dir_vectors (FILE *file, varray_type ddrs)
708 unsigned int i, j;
710 for (i = 0; i < VARRAY_ACTIVE_SIZE (ddrs); i++)
712 struct data_dependence_relation *ddr =
713 (struct data_dependence_relation *)
714 VARRAY_GENERIC_PTR (ddrs, i);
715 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE
716 && DDR_AFFINE_P (ddr))
718 for (j = 0; j < DDR_NUM_DIST_VECTS (ddr); j++)
720 fprintf (file, "DISTANCE_V (");
721 print_lambda_vector (file, DDR_DIST_VECT (ddr, j),
722 DDR_SIZE_VECT (ddr));
723 fprintf (file, ")\n");
726 for (j = 0; j < DDR_NUM_DIR_VECTS (ddr); j++)
728 fprintf (file, "DIRECTION_V (");
729 print_lambda_vector (file, DDR_DIR_VECT (ddr, j),
730 DDR_SIZE_VECT (ddr));
731 fprintf (file, ")\n");
735 fprintf (file, "\n\n");
738 /* Dumps the data dependence relations DDRS in FILE. */
740 void
741 dump_ddrs (FILE *file, varray_type ddrs)
743 unsigned int i;
745 for (i = 0; i < VARRAY_ACTIVE_SIZE (ddrs); i++)
747 struct data_dependence_relation *ddr =
748 (struct data_dependence_relation *)
749 VARRAY_GENERIC_PTR (ddrs, i);
750 dump_data_dependence_relation (file, ddr);
752 fprintf (file, "\n\n");
757 /* Estimate the number of iterations from the size of the data and the
758 access functions. */
760 static void
761 estimate_niter_from_size_of_data (struct loop *loop,
762 tree opnd0,
763 tree access_fn,
764 tree stmt)
766 tree estimation = NULL_TREE;
767 tree array_size, data_size, element_size;
768 tree init, step;
770 init = initial_condition (access_fn);
771 step = evolution_part_in_loop_num (access_fn, loop->num);
773 array_size = TYPE_SIZE (TREE_TYPE (opnd0));
774 element_size = TYPE_SIZE (TREE_TYPE (TREE_TYPE (opnd0)));
775 if (array_size == NULL_TREE
776 || TREE_CODE (array_size) != INTEGER_CST
777 || TREE_CODE (element_size) != INTEGER_CST)
778 return;
780 data_size = fold_build2 (EXACT_DIV_EXPR, integer_type_node,
781 array_size, element_size);
783 if (init != NULL_TREE
784 && step != NULL_TREE
785 && TREE_CODE (init) == INTEGER_CST
786 && TREE_CODE (step) == INTEGER_CST)
788 tree i_plus_s = fold_build2 (PLUS_EXPR, integer_type_node, init, step);
789 tree sign = fold_binary (GT_EXPR, boolean_type_node, i_plus_s, init);
791 if (sign == boolean_true_node)
792 estimation = fold_build2 (CEIL_DIV_EXPR, integer_type_node,
793 fold_build2 (MINUS_EXPR, integer_type_node,
794 data_size, init), step);
796 /* When the step is negative, as in PR23386: (init = 3, step =
797 0ffffffff, data_size = 100), we have to compute the
798 estimation as ceil_div (init, 0 - step) + 1. */
799 else if (sign == boolean_false_node)
800 estimation =
801 fold_build2 (PLUS_EXPR, integer_type_node,
802 fold_build2 (CEIL_DIV_EXPR, integer_type_node,
803 init,
804 fold_build2 (MINUS_EXPR, unsigned_type_node,
805 integer_zero_node, step)),
806 integer_one_node);
808 if (estimation)
809 record_estimate (loop, estimation, boolean_true_node, stmt);
813 /* Given an ARRAY_REF node REF, records its access functions.
814 Example: given A[i][3], record in ACCESS_FNS the opnd1 function,
815 i.e. the constant "3", then recursively call the function on opnd0,
816 i.e. the ARRAY_REF "A[i]".
817 If ESTIMATE_ONLY is true, we just set the estimated number of loop
818 iterations, we don't store the access function.
819 The function returns the base name: "A". */
821 static tree
822 analyze_array_indexes (struct loop *loop,
823 VEC(tree,heap) **access_fns,
824 tree ref, tree stmt,
825 bool estimate_only)
827 tree opnd0, opnd1;
828 tree access_fn;
830 opnd0 = TREE_OPERAND (ref, 0);
831 opnd1 = TREE_OPERAND (ref, 1);
833 /* The detection of the evolution function for this data access is
834 postponed until the dependence test. This lazy strategy avoids
835 the computation of access functions that are of no interest for
836 the optimizers. */
837 access_fn = instantiate_parameters
838 (loop, analyze_scalar_evolution (loop, opnd1));
840 if (estimate_only
841 && chrec_contains_undetermined (loop->estimated_nb_iterations))
842 estimate_niter_from_size_of_data (loop, opnd0, access_fn, stmt);
844 if (!estimate_only)
845 VEC_safe_push (tree, heap, *access_fns, access_fn);
847 /* Recursively record other array access functions. */
848 if (TREE_CODE (opnd0) == ARRAY_REF)
849 return analyze_array_indexes (loop, access_fns, opnd0, stmt, estimate_only);
851 /* Return the base name of the data access. */
852 else
853 return opnd0;
856 /* For an array reference REF contained in STMT, attempt to bound the
857 number of iterations in the loop containing STMT */
859 void
860 estimate_iters_using_array (tree stmt, tree ref)
862 analyze_array_indexes (loop_containing_stmt (stmt), NULL, ref, stmt,
863 true);
866 /* For a data reference REF contained in the statement STMT, initialize
867 a DATA_REFERENCE structure, and return it. IS_READ flag has to be
868 set to true when REF is in the right hand side of an
869 assignment. */
871 struct data_reference *
872 analyze_array (tree stmt, tree ref, bool is_read)
874 struct data_reference *res;
875 VEC(tree,heap) *acc_fns;
877 if (dump_file && (dump_flags & TDF_DETAILS))
879 fprintf (dump_file, "(analyze_array \n");
880 fprintf (dump_file, " (ref = ");
881 print_generic_stmt (dump_file, ref, 0);
882 fprintf (dump_file, ")\n");
885 res = XNEW (struct data_reference);
887 DR_STMT (res) = stmt;
888 DR_REF (res) = ref;
889 acc_fns = VEC_alloc (tree, heap, 3);
890 DR_BASE_OBJECT (res) = analyze_array_indexes
891 (loop_containing_stmt (stmt), &acc_fns, ref, stmt, false);
892 DR_TYPE (res) = ARRAY_REF_TYPE;
893 DR_SET_ACCESS_FNS (res, acc_fns);
894 DR_IS_READ (res) = is_read;
895 DR_BASE_ADDRESS (res) = NULL_TREE;
896 DR_OFFSET (res) = NULL_TREE;
897 DR_INIT (res) = NULL_TREE;
898 DR_STEP (res) = NULL_TREE;
899 DR_OFFSET_MISALIGNMENT (res) = NULL_TREE;
900 DR_MEMTAG (res) = NULL_TREE;
901 DR_PTR_INFO (res) = NULL;
903 if (dump_file && (dump_flags & TDF_DETAILS))
904 fprintf (dump_file, ")\n");
906 return res;
910 /* Analyze an indirect memory reference, REF, that comes from STMT.
911 IS_READ is true if this is an indirect load, and false if it is
912 an indirect store.
913 Return a new data reference structure representing the indirect_ref, or
914 NULL if we cannot describe the access function. */
916 static struct data_reference *
917 analyze_indirect_ref (tree stmt, tree ref, bool is_read)
919 struct loop *loop = loop_containing_stmt (stmt);
920 tree ptr_ref = TREE_OPERAND (ref, 0);
921 tree access_fn = analyze_scalar_evolution (loop, ptr_ref);
922 tree init = initial_condition_in_loop_num (access_fn, loop->num);
923 tree base_address = NULL_TREE, evolution, step = NULL_TREE;
924 struct ptr_info_def *ptr_info = NULL;
926 if (TREE_CODE (ptr_ref) == SSA_NAME)
927 ptr_info = SSA_NAME_PTR_INFO (ptr_ref);
929 STRIP_NOPS (init);
930 if (access_fn == chrec_dont_know || !init || init == chrec_dont_know)
932 if (dump_file && (dump_flags & TDF_DETAILS))
934 fprintf (dump_file, "\nBad access function of ptr: ");
935 print_generic_expr (dump_file, ref, TDF_SLIM);
936 fprintf (dump_file, "\n");
938 return NULL;
941 if (dump_file && (dump_flags & TDF_DETAILS))
943 fprintf (dump_file, "\nAccess function of ptr: ");
944 print_generic_expr (dump_file, access_fn, TDF_SLIM);
945 fprintf (dump_file, "\n");
948 if (!expr_invariant_in_loop_p (loop, init))
950 if (dump_file && (dump_flags & TDF_DETAILS))
951 fprintf (dump_file, "\ninitial condition is not loop invariant.\n");
953 else
955 base_address = init;
956 evolution = evolution_part_in_loop_num (access_fn, loop->num);
957 if (evolution != chrec_dont_know)
959 if (!evolution)
960 step = ssize_int (0);
961 else
963 if (TREE_CODE (evolution) == INTEGER_CST)
964 step = fold_convert (ssizetype, evolution);
965 else
966 if (dump_file && (dump_flags & TDF_DETAILS))
967 fprintf (dump_file, "\nnon constant step for ptr access.\n");
970 else
971 if (dump_file && (dump_flags & TDF_DETAILS))
972 fprintf (dump_file, "\nunknown evolution of ptr.\n");
974 return init_data_ref (stmt, ref, NULL_TREE, access_fn, is_read, base_address,
975 NULL_TREE, step, NULL_TREE, NULL_TREE,
976 ptr_info, POINTER_REF_TYPE);
979 /* For a data reference REF contained in the statement STMT, initialize
980 a DATA_REFERENCE structure, and return it. */
982 struct data_reference *
983 init_data_ref (tree stmt,
984 tree ref,
985 tree base,
986 tree access_fn,
987 bool is_read,
988 tree base_address,
989 tree init_offset,
990 tree step,
991 tree misalign,
992 tree memtag,
993 struct ptr_info_def *ptr_info,
994 enum data_ref_type type)
996 struct data_reference *res;
997 VEC(tree,heap) *acc_fns;
999 if (dump_file && (dump_flags & TDF_DETAILS))
1001 fprintf (dump_file, "(init_data_ref \n");
1002 fprintf (dump_file, " (ref = ");
1003 print_generic_stmt (dump_file, ref, 0);
1004 fprintf (dump_file, ")\n");
1007 res = XNEW (struct data_reference);
1009 DR_STMT (res) = stmt;
1010 DR_REF (res) = ref;
1011 DR_BASE_OBJECT (res) = base;
1012 DR_TYPE (res) = type;
1013 acc_fns = VEC_alloc (tree, heap, 3);
1014 DR_SET_ACCESS_FNS (res, acc_fns);
1015 VEC_quick_push (tree, DR_ACCESS_FNS (res), access_fn);
1016 DR_IS_READ (res) = is_read;
1017 DR_BASE_ADDRESS (res) = base_address;
1018 DR_OFFSET (res) = init_offset;
1019 DR_INIT (res) = NULL_TREE;
1020 DR_STEP (res) = step;
1021 DR_OFFSET_MISALIGNMENT (res) = misalign;
1022 DR_MEMTAG (res) = memtag;
1023 DR_PTR_INFO (res) = ptr_info;
1025 if (dump_file && (dump_flags & TDF_DETAILS))
1026 fprintf (dump_file, ")\n");
1028 return res;
1033 /* Function strip_conversions
1035 Strip conversions that don't narrow the mode. */
1037 static tree
1038 strip_conversion (tree expr)
1040 tree to, ti, oprnd0;
1042 while (TREE_CODE (expr) == NOP_EXPR || TREE_CODE (expr) == CONVERT_EXPR)
1044 to = TREE_TYPE (expr);
1045 oprnd0 = TREE_OPERAND (expr, 0);
1046 ti = TREE_TYPE (oprnd0);
1048 if (!INTEGRAL_TYPE_P (to) || !INTEGRAL_TYPE_P (ti))
1049 return NULL_TREE;
1050 if (GET_MODE_SIZE (TYPE_MODE (to)) < GET_MODE_SIZE (TYPE_MODE (ti)))
1051 return NULL_TREE;
1053 expr = oprnd0;
1055 return expr;
1059 /* Function analyze_offset_expr
1061 Given an offset expression EXPR received from get_inner_reference, analyze
1062 it and create an expression for INITIAL_OFFSET by substituting the variables
1063 of EXPR with initial_condition of the corresponding access_fn in the loop.
1064 E.g.,
1065 for i
1066 for (j = 3; j < N; j++)
1067 a[j].b[i][j] = 0;
1069 For a[j].b[i][j], EXPR will be 'i * C_i + j * C_j + C'. 'i' cannot be
1070 substituted, since its access_fn in the inner loop is i. 'j' will be
1071 substituted with 3. An INITIAL_OFFSET will be 'i * C_i + C`', where
1072 C` = 3 * C_j + C.
1074 Compute MISALIGN (the misalignment of the data reference initial access from
1075 its base). Misalignment can be calculated only if all the variables can be
1076 substituted with constants, otherwise, we record maximum possible alignment
1077 in ALIGNED_TO. In the above example, since 'i' cannot be substituted, MISALIGN
1078 will be NULL_TREE, and the biggest divider of C_i (a power of 2) will be
1079 recorded in ALIGNED_TO.
1081 STEP is an evolution of the data reference in this loop in bytes.
1082 In the above example, STEP is C_j.
1084 Return FALSE, if the analysis fails, e.g., there is no access_fn for a
1085 variable. In this case, all the outputs (INITIAL_OFFSET, MISALIGN, ALIGNED_TO
1086 and STEP) are NULL_TREEs. Otherwise, return TRUE.
1090 static bool
1091 analyze_offset_expr (tree expr,
1092 struct loop *loop,
1093 tree *initial_offset,
1094 tree *misalign,
1095 tree *aligned_to,
1096 tree *step)
1098 tree oprnd0;
1099 tree oprnd1;
1100 tree left_offset = ssize_int (0);
1101 tree right_offset = ssize_int (0);
1102 tree left_misalign = ssize_int (0);
1103 tree right_misalign = ssize_int (0);
1104 tree left_step = ssize_int (0);
1105 tree right_step = ssize_int (0);
1106 enum tree_code code;
1107 tree init, evolution;
1108 tree left_aligned_to = NULL_TREE, right_aligned_to = NULL_TREE;
1110 *step = NULL_TREE;
1111 *misalign = NULL_TREE;
1112 *aligned_to = NULL_TREE;
1113 *initial_offset = NULL_TREE;
1115 /* Strip conversions that don't narrow the mode. */
1116 expr = strip_conversion (expr);
1117 if (!expr)
1118 return false;
1120 /* Stop conditions:
1121 1. Constant. */
1122 if (TREE_CODE (expr) == INTEGER_CST)
1124 *initial_offset = fold_convert (ssizetype, expr);
1125 *misalign = fold_convert (ssizetype, expr);
1126 *step = ssize_int (0);
1127 return true;
1130 /* 2. Variable. Try to substitute with initial_condition of the corresponding
1131 access_fn in the current loop. */
1132 if (SSA_VAR_P (expr))
1134 tree access_fn = analyze_scalar_evolution (loop, expr);
1136 if (access_fn == chrec_dont_know)
1137 /* No access_fn. */
1138 return false;
1140 init = initial_condition_in_loop_num (access_fn, loop->num);
1141 if (!expr_invariant_in_loop_p (loop, init))
1142 /* Not enough information: may be not loop invariant.
1143 E.g., for a[b[i]], we get a[D], where D=b[i]. EXPR is D, its
1144 initial_condition is D, but it depends on i - loop's induction
1145 variable. */
1146 return false;
1148 evolution = evolution_part_in_loop_num (access_fn, loop->num);
1149 if (evolution && TREE_CODE (evolution) != INTEGER_CST)
1150 /* Evolution is not constant. */
1151 return false;
1153 if (TREE_CODE (init) == INTEGER_CST)
1154 *misalign = fold_convert (ssizetype, init);
1155 else
1156 /* Not constant, misalignment cannot be calculated. */
1157 *misalign = NULL_TREE;
1159 *initial_offset = fold_convert (ssizetype, init);
1161 *step = evolution ? fold_convert (ssizetype, evolution) : ssize_int (0);
1162 return true;
1165 /* Recursive computation. */
1166 if (!BINARY_CLASS_P (expr))
1168 /* We expect to get binary expressions (PLUS/MINUS and MULT). */
1169 if (dump_file && (dump_flags & TDF_DETAILS))
1171 fprintf (dump_file, "\nNot binary expression ");
1172 print_generic_expr (dump_file, expr, TDF_SLIM);
1173 fprintf (dump_file, "\n");
1175 return false;
1177 oprnd0 = TREE_OPERAND (expr, 0);
1178 oprnd1 = TREE_OPERAND (expr, 1);
1180 if (!analyze_offset_expr (oprnd0, loop, &left_offset, &left_misalign,
1181 &left_aligned_to, &left_step)
1182 || !analyze_offset_expr (oprnd1, loop, &right_offset, &right_misalign,
1183 &right_aligned_to, &right_step))
1184 return false;
1186 /* The type of the operation: plus, minus or mult. */
1187 code = TREE_CODE (expr);
1188 switch (code)
1190 case MULT_EXPR:
1191 if (TREE_CODE (right_offset) != INTEGER_CST)
1192 /* RIGHT_OFFSET can be not constant. For example, for arrays of variable
1193 sized types.
1194 FORNOW: We don't support such cases. */
1195 return false;
1197 /* Strip conversions that don't narrow the mode. */
1198 left_offset = strip_conversion (left_offset);
1199 if (!left_offset)
1200 return false;
1201 /* Misalignment computation. */
1202 if (SSA_VAR_P (left_offset))
1204 /* If the left side contains variables that can't be substituted with
1205 constants, the misalignment is unknown. However, if the right side
1206 is a multiple of some alignment, we know that the expression is
1207 aligned to it. Therefore, we record such maximum possible value.
1209 *misalign = NULL_TREE;
1210 *aligned_to = ssize_int (highest_pow2_factor (right_offset));
1212 else
1214 /* The left operand was successfully substituted with constant. */
1215 if (left_misalign)
1217 /* In case of EXPR '(i * C1 + j) * C2', LEFT_MISALIGN is
1218 NULL_TREE. */
1219 *misalign = size_binop (code, left_misalign, right_misalign);
1220 if (left_aligned_to && right_aligned_to)
1221 *aligned_to = size_binop (MIN_EXPR, left_aligned_to,
1222 right_aligned_to);
1223 else
1224 *aligned_to = left_aligned_to ?
1225 left_aligned_to : right_aligned_to;
1227 else
1228 *misalign = NULL_TREE;
1231 /* Step calculation. */
1232 /* Multiply the step by the right operand. */
1233 *step = size_binop (MULT_EXPR, left_step, right_offset);
1234 break;
1236 case PLUS_EXPR:
1237 case MINUS_EXPR:
1238 /* Combine the recursive calculations for step and misalignment. */
1239 *step = size_binop (code, left_step, right_step);
1241 /* Unknown alignment. */
1242 if ((!left_misalign && !left_aligned_to)
1243 || (!right_misalign && !right_aligned_to))
1245 *misalign = NULL_TREE;
1246 *aligned_to = NULL_TREE;
1247 break;
1250 if (left_misalign && right_misalign)
1251 *misalign = size_binop (code, left_misalign, right_misalign);
1252 else
1253 *misalign = left_misalign ? left_misalign : right_misalign;
1255 if (left_aligned_to && right_aligned_to)
1256 *aligned_to = size_binop (MIN_EXPR, left_aligned_to, right_aligned_to);
1257 else
1258 *aligned_to = left_aligned_to ? left_aligned_to : right_aligned_to;
1260 break;
1262 default:
1263 gcc_unreachable ();
1266 /* Compute offset. */
1267 *initial_offset = fold_convert (ssizetype,
1268 fold_build2 (code, TREE_TYPE (left_offset),
1269 left_offset,
1270 right_offset));
1271 return true;
1274 /* Function address_analysis
1276 Return the BASE of the address expression EXPR.
1277 Also compute the OFFSET from BASE, MISALIGN and STEP.
1279 Input:
1280 EXPR - the address expression that is being analyzed
1281 STMT - the statement that contains EXPR or its original memory reference
1282 IS_READ - TRUE if STMT reads from EXPR, FALSE if writes to EXPR
1283 DR - data_reference struct for the original memory reference
1285 Output:
1286 BASE (returned value) - the base of the data reference EXPR.
1287 INITIAL_OFFSET - initial offset of EXPR from BASE (an expression)
1288 MISALIGN - offset of EXPR from BASE in bytes (a constant) or NULL_TREE if the
1289 computation is impossible
1290 ALIGNED_TO - maximum alignment of EXPR or NULL_TREE if MISALIGN can be
1291 calculated (doesn't depend on variables)
1292 STEP - evolution of EXPR in the loop
1294 If something unexpected is encountered (an unsupported form of data-ref),
1295 then NULL_TREE is returned.
1298 static tree
1299 address_analysis (tree expr, tree stmt, bool is_read, struct data_reference *dr,
1300 tree *offset, tree *misalign, tree *aligned_to, tree *step)
1302 tree oprnd0, oprnd1, base_address, offset_expr, base_addr0, base_addr1;
1303 tree address_offset = ssize_int (0), address_misalign = ssize_int (0);
1304 tree dummy, address_aligned_to = NULL_TREE;
1305 struct ptr_info_def *dummy1;
1306 subvar_t dummy2;
1308 switch (TREE_CODE (expr))
1310 case PLUS_EXPR:
1311 case MINUS_EXPR:
1312 /* EXPR is of form {base +/- offset} (or {offset +/- base}). */
1313 oprnd0 = TREE_OPERAND (expr, 0);
1314 oprnd1 = TREE_OPERAND (expr, 1);
1316 STRIP_NOPS (oprnd0);
1317 STRIP_NOPS (oprnd1);
1319 /* Recursively try to find the base of the address contained in EXPR.
1320 For offset, the returned base will be NULL. */
1321 base_addr0 = address_analysis (oprnd0, stmt, is_read, dr, &address_offset,
1322 &address_misalign, &address_aligned_to,
1323 step);
1325 base_addr1 = address_analysis (oprnd1, stmt, is_read, dr, &address_offset,
1326 &address_misalign, &address_aligned_to,
1327 step);
1329 /* We support cases where only one of the operands contains an
1330 address. */
1331 if ((base_addr0 && base_addr1) || (!base_addr0 && !base_addr1))
1333 if (dump_file && (dump_flags & TDF_DETAILS))
1335 fprintf (dump_file,
1336 "\neither more than one address or no addresses in expr ");
1337 print_generic_expr (dump_file, expr, TDF_SLIM);
1338 fprintf (dump_file, "\n");
1340 return NULL_TREE;
1343 /* To revert STRIP_NOPS. */
1344 oprnd0 = TREE_OPERAND (expr, 0);
1345 oprnd1 = TREE_OPERAND (expr, 1);
1347 offset_expr = base_addr0 ?
1348 fold_convert (ssizetype, oprnd1) : fold_convert (ssizetype, oprnd0);
1350 /* EXPR is of form {base +/- offset} (or {offset +/- base}). If offset is
1351 a number, we can add it to the misalignment value calculated for base,
1352 otherwise, misalignment is NULL. */
1353 if (TREE_CODE (offset_expr) == INTEGER_CST && address_misalign)
1355 *misalign = size_binop (TREE_CODE (expr), address_misalign,
1356 offset_expr);
1357 *aligned_to = address_aligned_to;
1359 else
1361 *misalign = NULL_TREE;
1362 *aligned_to = NULL_TREE;
1365 /* Combine offset (from EXPR {base + offset}) with the offset calculated
1366 for base. */
1367 *offset = size_binop (TREE_CODE (expr), address_offset, offset_expr);
1368 return base_addr0 ? base_addr0 : base_addr1;
1370 case ADDR_EXPR:
1371 base_address = object_analysis (TREE_OPERAND (expr, 0), stmt, is_read,
1372 &dr, offset, misalign, aligned_to, step,
1373 &dummy, &dummy1, &dummy2);
1374 return base_address;
1376 case SSA_NAME:
1377 if (!POINTER_TYPE_P (TREE_TYPE (expr)))
1379 if (dump_file && (dump_flags & TDF_DETAILS))
1381 fprintf (dump_file, "\nnot pointer SSA_NAME ");
1382 print_generic_expr (dump_file, expr, TDF_SLIM);
1383 fprintf (dump_file, "\n");
1385 return NULL_TREE;
1387 *aligned_to = ssize_int (TYPE_ALIGN_UNIT (TREE_TYPE (TREE_TYPE (expr))));
1388 *misalign = ssize_int (0);
1389 *offset = ssize_int (0);
1390 *step = ssize_int (0);
1391 return expr;
1393 default:
1394 return NULL_TREE;
1399 /* Function object_analysis
1401 Create a data-reference structure DR for MEMREF.
1402 Return the BASE of the data reference MEMREF if the analysis is possible.
1403 Also compute the INITIAL_OFFSET from BASE, MISALIGN and STEP.
1404 E.g., for EXPR a.b[i] + 4B, BASE is a, and OFFSET is the overall offset
1405 'a.b[i] + 4B' from a (can be an expression), MISALIGN is an OFFSET
1406 instantiated with initial_conditions of access_functions of variables,
1407 and STEP is the evolution of the DR_REF in this loop.
1409 Function get_inner_reference is used for the above in case of ARRAY_REF and
1410 COMPONENT_REF.
1412 The structure of the function is as follows:
1413 Part 1:
1414 Case 1. For handled_component_p refs
1415 1.1 build data-reference structure for MEMREF
1416 1.2 call get_inner_reference
1417 1.2.1 analyze offset expr received from get_inner_reference
1418 (fall through with BASE)
1419 Case 2. For declarations
1420 2.1 set MEMTAG
1421 Case 3. For INDIRECT_REFs
1422 3.1 build data-reference structure for MEMREF
1423 3.2 analyze evolution and initial condition of MEMREF
1424 3.3 set data-reference structure for MEMREF
1425 3.4 call address_analysis to analyze INIT of the access function
1426 3.5 extract memory tag
1428 Part 2:
1429 Combine the results of object and address analysis to calculate
1430 INITIAL_OFFSET, STEP and misalignment info.
1432 Input:
1433 MEMREF - the memory reference that is being analyzed
1434 STMT - the statement that contains MEMREF
1435 IS_READ - TRUE if STMT reads from MEMREF, FALSE if writes to MEMREF
1437 Output:
1438 BASE_ADDRESS (returned value) - the base address of the data reference MEMREF
1439 E.g, if MEMREF is a.b[k].c[i][j] the returned
1440 base is &a.
1441 DR - data_reference struct for MEMREF
1442 INITIAL_OFFSET - initial offset of MEMREF from BASE (an expression)
1443 MISALIGN - offset of MEMREF from BASE in bytes (a constant) modulo alignment of
1444 ALIGNMENT or NULL_TREE if the computation is impossible
1445 ALIGNED_TO - maximum alignment of EXPR or NULL_TREE if MISALIGN can be
1446 calculated (doesn't depend on variables)
1447 STEP - evolution of the DR_REF in the loop
1448 MEMTAG - memory tag for aliasing purposes
1449 PTR_INFO - NULL or points-to aliasing info from a pointer SSA_NAME
1450 SUBVARS - Sub-variables of the variable
1452 If the analysis of MEMREF evolution in the loop fails, NULL_TREE is returned,
1453 but DR can be created anyway.
1457 static tree
1458 object_analysis (tree memref, tree stmt, bool is_read,
1459 struct data_reference **dr, tree *offset, tree *misalign,
1460 tree *aligned_to, tree *step, tree *memtag,
1461 struct ptr_info_def **ptr_info, subvar_t *subvars)
1463 tree base = NULL_TREE, base_address = NULL_TREE;
1464 tree object_offset = ssize_int (0), object_misalign = ssize_int (0);
1465 tree object_step = ssize_int (0), address_step = ssize_int (0);
1466 tree address_offset = ssize_int (0), address_misalign = ssize_int (0);
1467 HOST_WIDE_INT pbitsize, pbitpos;
1468 tree poffset, bit_pos_in_bytes;
1469 enum machine_mode pmode;
1470 int punsignedp, pvolatilep;
1471 tree ptr_step = ssize_int (0), ptr_init = NULL_TREE;
1472 struct loop *loop = loop_containing_stmt (stmt);
1473 struct data_reference *ptr_dr = NULL;
1474 tree object_aligned_to = NULL_TREE, address_aligned_to = NULL_TREE;
1476 *ptr_info = NULL;
1478 /* Part 1: */
1479 /* Case 1. handled_component_p refs. */
1480 if (handled_component_p (memref))
1482 /* 1.1 build data-reference structure for MEMREF. */
1483 /* TODO: handle COMPONENT_REFs. */
1484 if (!(*dr))
1486 if (TREE_CODE (memref) == ARRAY_REF)
1487 *dr = analyze_array (stmt, memref, is_read);
1488 else
1490 /* FORNOW. */
1491 if (dump_file && (dump_flags & TDF_DETAILS))
1493 fprintf (dump_file, "\ndata-ref of unsupported type ");
1494 print_generic_expr (dump_file, memref, TDF_SLIM);
1495 fprintf (dump_file, "\n");
1497 return NULL_TREE;
1501 /* 1.2 call get_inner_reference. */
1502 /* Find the base and the offset from it. */
1503 base = get_inner_reference (memref, &pbitsize, &pbitpos, &poffset,
1504 &pmode, &punsignedp, &pvolatilep, false);
1505 if (!base)
1507 if (dump_file && (dump_flags & TDF_DETAILS))
1509 fprintf (dump_file, "\nfailed to get inner ref for ");
1510 print_generic_expr (dump_file, memref, TDF_SLIM);
1511 fprintf (dump_file, "\n");
1513 return NULL_TREE;
1516 /* 1.2.1 analyze offset expr received from get_inner_reference. */
1517 if (poffset
1518 && !analyze_offset_expr (poffset, loop, &object_offset,
1519 &object_misalign, &object_aligned_to,
1520 &object_step))
1522 if (dump_file && (dump_flags & TDF_DETAILS))
1524 fprintf (dump_file, "\nfailed to compute offset or step for ");
1525 print_generic_expr (dump_file, memref, TDF_SLIM);
1526 fprintf (dump_file, "\n");
1528 return NULL_TREE;
1531 /* Add bit position to OFFSET and MISALIGN. */
1533 bit_pos_in_bytes = ssize_int (pbitpos/BITS_PER_UNIT);
1534 /* Check that there is no remainder in bits. */
1535 if (pbitpos%BITS_PER_UNIT)
1537 if (dump_file && (dump_flags & TDF_DETAILS))
1538 fprintf (dump_file, "\nbit offset alignment.\n");
1539 return NULL_TREE;
1541 object_offset = size_binop (PLUS_EXPR, bit_pos_in_bytes, object_offset);
1542 if (object_misalign)
1543 object_misalign = size_binop (PLUS_EXPR, object_misalign,
1544 bit_pos_in_bytes);
1546 memref = base; /* To continue analysis of BASE. */
1547 /* fall through */
1550 /* Part 1: Case 2. Declarations. */
1551 if (DECL_P (memref))
1553 /* We expect to get a decl only if we already have a DR. */
1554 if (!(*dr))
1556 if (dump_file && (dump_flags & TDF_DETAILS))
1558 fprintf (dump_file, "\nunhandled decl ");
1559 print_generic_expr (dump_file, memref, TDF_SLIM);
1560 fprintf (dump_file, "\n");
1562 return NULL_TREE;
1565 /* TODO: if during the analysis of INDIRECT_REF we get to an object, put
1566 the object in BASE_OBJECT field if we can prove that this is O.K.,
1567 i.e., the data-ref access is bounded by the bounds of the BASE_OBJECT.
1568 (e.g., if the object is an array base 'a', where 'a[N]', we must prove
1569 that every access with 'p' (the original INDIRECT_REF based on '&a')
1570 in the loop is within the array boundaries - from a[0] to a[N-1]).
1571 Otherwise, our alias analysis can be incorrect.
1572 Even if an access function based on BASE_OBJECT can't be build, update
1573 BASE_OBJECT field to enable us to prove that two data-refs are
1574 different (without access function, distance analysis is impossible).
1576 if (SSA_VAR_P (memref) && var_can_have_subvars (memref))
1577 *subvars = get_subvars_for_var (memref);
1578 base_address = build_fold_addr_expr (memref);
1579 /* 2.1 set MEMTAG. */
1580 *memtag = memref;
1583 /* Part 1: Case 3. INDIRECT_REFs. */
1584 else if (TREE_CODE (memref) == INDIRECT_REF)
1586 tree ptr_ref = TREE_OPERAND (memref, 0);
1587 if (TREE_CODE (ptr_ref) == SSA_NAME)
1588 *ptr_info = SSA_NAME_PTR_INFO (ptr_ref);
1590 /* 3.1 build data-reference structure for MEMREF. */
1591 ptr_dr = analyze_indirect_ref (stmt, memref, is_read);
1592 if (!ptr_dr)
1594 if (dump_file && (dump_flags & TDF_DETAILS))
1596 fprintf (dump_file, "\nfailed to create dr for ");
1597 print_generic_expr (dump_file, memref, TDF_SLIM);
1598 fprintf (dump_file, "\n");
1600 return NULL_TREE;
1603 /* 3.2 analyze evolution and initial condition of MEMREF. */
1604 ptr_step = DR_STEP (ptr_dr);
1605 ptr_init = DR_BASE_ADDRESS (ptr_dr);
1606 if (!ptr_init || !ptr_step || !POINTER_TYPE_P (TREE_TYPE (ptr_init)))
1608 *dr = (*dr) ? *dr : ptr_dr;
1609 if (dump_file && (dump_flags & TDF_DETAILS))
1611 fprintf (dump_file, "\nbad pointer access ");
1612 print_generic_expr (dump_file, memref, TDF_SLIM);
1613 fprintf (dump_file, "\n");
1615 return NULL_TREE;
1618 if (integer_zerop (ptr_step) && !(*dr))
1620 if (dump_file && (dump_flags & TDF_DETAILS))
1621 fprintf (dump_file, "\nptr is loop invariant.\n");
1622 *dr = ptr_dr;
1623 return NULL_TREE;
1625 /* If there exists DR for MEMREF, we are analyzing the base of
1626 handled component (PTR_INIT), which not necessary has evolution in
1627 the loop. */
1629 object_step = size_binop (PLUS_EXPR, object_step, ptr_step);
1631 /* 3.3 set data-reference structure for MEMREF. */
1632 if (!*dr)
1633 *dr = ptr_dr;
1635 /* 3.4 call address_analysis to analyze INIT of the access
1636 function. */
1637 base_address = address_analysis (ptr_init, stmt, is_read, *dr,
1638 &address_offset, &address_misalign,
1639 &address_aligned_to, &address_step);
1640 if (!base_address)
1642 if (dump_file && (dump_flags & TDF_DETAILS))
1644 fprintf (dump_file, "\nfailed to analyze address ");
1645 print_generic_expr (dump_file, ptr_init, TDF_SLIM);
1646 fprintf (dump_file, "\n");
1648 return NULL_TREE;
1651 /* 3.5 extract memory tag. */
1652 switch (TREE_CODE (base_address))
1654 case SSA_NAME:
1655 *memtag = get_var_ann (SSA_NAME_VAR (base_address))->type_mem_tag;
1656 if (!(*memtag) && TREE_CODE (TREE_OPERAND (memref, 0)) == SSA_NAME)
1657 *memtag = get_var_ann (
1658 SSA_NAME_VAR (TREE_OPERAND (memref, 0)))->type_mem_tag;
1659 break;
1660 case ADDR_EXPR:
1661 *memtag = TREE_OPERAND (base_address, 0);
1662 break;
1663 default:
1664 if (dump_file && (dump_flags & TDF_DETAILS))
1666 fprintf (dump_file, "\nno memtag for ");
1667 print_generic_expr (dump_file, memref, TDF_SLIM);
1668 fprintf (dump_file, "\n");
1670 *memtag = NULL_TREE;
1671 break;
1675 if (!base_address)
1677 /* MEMREF cannot be analyzed. */
1678 if (dump_file && (dump_flags & TDF_DETAILS))
1680 fprintf (dump_file, "\ndata-ref of unsupported type ");
1681 print_generic_expr (dump_file, memref, TDF_SLIM);
1682 fprintf (dump_file, "\n");
1684 return NULL_TREE;
1687 if (SSA_VAR_P (*memtag) && var_can_have_subvars (*memtag))
1688 *subvars = get_subvars_for_var (*memtag);
1690 /* Part 2: Combine the results of object and address analysis to calculate
1691 INITIAL_OFFSET, STEP and misalignment info. */
1692 *offset = size_binop (PLUS_EXPR, object_offset, address_offset);
1694 if ((!object_misalign && !object_aligned_to)
1695 || (!address_misalign && !address_aligned_to))
1697 *misalign = NULL_TREE;
1698 *aligned_to = NULL_TREE;
1700 else
1702 if (object_misalign && address_misalign)
1703 *misalign = size_binop (PLUS_EXPR, object_misalign, address_misalign);
1704 else
1705 *misalign = object_misalign ? object_misalign : address_misalign;
1706 if (object_aligned_to && address_aligned_to)
1707 *aligned_to = size_binop (MIN_EXPR, object_aligned_to,
1708 address_aligned_to);
1709 else
1710 *aligned_to = object_aligned_to ?
1711 object_aligned_to : address_aligned_to;
1713 *step = size_binop (PLUS_EXPR, object_step, address_step);
1715 return base_address;
1718 /* Function analyze_offset.
1720 Extract INVARIANT and CONSTANT parts from OFFSET.
1723 static void
1724 analyze_offset (tree offset, tree *invariant, tree *constant)
1726 tree op0, op1, constant_0, constant_1, invariant_0, invariant_1;
1727 enum tree_code code = TREE_CODE (offset);
1729 *invariant = NULL_TREE;
1730 *constant = NULL_TREE;
1732 /* Not PLUS/MINUS expression - recursion stop condition. */
1733 if (code != PLUS_EXPR && code != MINUS_EXPR)
1735 if (TREE_CODE (offset) == INTEGER_CST)
1736 *constant = offset;
1737 else
1738 *invariant = offset;
1739 return;
1742 op0 = TREE_OPERAND (offset, 0);
1743 op1 = TREE_OPERAND (offset, 1);
1745 /* Recursive call with the operands. */
1746 analyze_offset (op0, &invariant_0, &constant_0);
1747 analyze_offset (op1, &invariant_1, &constant_1);
1749 /* Combine the results. */
1750 *constant = constant_0 ? constant_0 : constant_1;
1751 if (invariant_0 && invariant_1)
1752 *invariant =
1753 fold_build2 (code, TREE_TYPE (invariant_0), invariant_0, invariant_1);
1754 else
1755 *invariant = invariant_0 ? invariant_0 : invariant_1;
1759 /* Function create_data_ref.
1761 Create a data-reference structure for MEMREF. Set its DR_BASE_ADDRESS,
1762 DR_OFFSET, DR_INIT, DR_STEP, DR_OFFSET_MISALIGNMENT, DR_ALIGNED_TO,
1763 DR_MEMTAG, and DR_POINTSTO_INFO fields.
1765 Input:
1766 MEMREF - the memory reference that is being analyzed
1767 STMT - the statement that contains MEMREF
1768 IS_READ - TRUE if STMT reads from MEMREF, FALSE if writes to MEMREF
1770 Output:
1771 DR (returned value) - data_reference struct for MEMREF
1774 static struct data_reference *
1775 create_data_ref (tree memref, tree stmt, bool is_read)
1777 struct data_reference *dr = NULL;
1778 tree base_address, offset, step, misalign, memtag;
1779 struct loop *loop = loop_containing_stmt (stmt);
1780 tree invariant = NULL_TREE, constant = NULL_TREE;
1781 tree type_size, init_cond;
1782 struct ptr_info_def *ptr_info;
1783 subvar_t subvars = NULL;
1784 tree aligned_to;
1786 if (!memref)
1787 return NULL;
1789 base_address = object_analysis (memref, stmt, is_read, &dr, &offset,
1790 &misalign, &aligned_to, &step, &memtag,
1791 &ptr_info, &subvars);
1792 if (!dr || !base_address)
1794 if (dump_file && (dump_flags & TDF_DETAILS))
1796 fprintf (dump_file, "\ncreate_data_ref: failed to create a dr for ");
1797 print_generic_expr (dump_file, memref, TDF_SLIM);
1798 fprintf (dump_file, "\n");
1800 return NULL;
1803 DR_BASE_ADDRESS (dr) = base_address;
1804 DR_OFFSET (dr) = offset;
1805 DR_INIT (dr) = ssize_int (0);
1806 DR_STEP (dr) = step;
1807 DR_OFFSET_MISALIGNMENT (dr) = misalign;
1808 DR_ALIGNED_TO (dr) = aligned_to;
1809 DR_MEMTAG (dr) = memtag;
1810 DR_PTR_INFO (dr) = ptr_info;
1811 DR_SUBVARS (dr) = subvars;
1813 type_size = fold_convert (ssizetype, TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr))));
1815 /* Change the access function for INIDIRECT_REFs, according to
1816 DR_BASE_ADDRESS. Analyze OFFSET calculated in object_analysis. OFFSET is
1817 an expression that can contain loop invariant expressions and constants.
1818 We put the constant part in the initial condition of the access function
1819 (for data dependence tests), and in DR_INIT of the data-ref. The loop
1820 invariant part is put in DR_OFFSET.
1821 The evolution part of the access function is STEP calculated in
1822 object_analysis divided by the size of data type.
1824 if (!DR_BASE_OBJECT (dr))
1826 tree access_fn;
1827 tree new_step;
1829 /* Extract CONSTANT and INVARIANT from OFFSET, and put them in DR_INIT and
1830 DR_OFFSET fields of DR. */
1831 analyze_offset (offset, &invariant, &constant);
1832 if (constant)
1834 DR_INIT (dr) = fold_convert (ssizetype, constant);
1835 init_cond = fold_build2 (TRUNC_DIV_EXPR, TREE_TYPE (constant),
1836 constant, type_size);
1838 else
1839 DR_INIT (dr) = init_cond = ssize_int (0);;
1841 if (invariant)
1842 DR_OFFSET (dr) = invariant;
1843 else
1844 DR_OFFSET (dr) = ssize_int (0);
1846 /* Update access function. */
1847 access_fn = DR_ACCESS_FN (dr, 0);
1848 new_step = size_binop (TRUNC_DIV_EXPR,
1849 fold_convert (ssizetype, step), type_size);
1851 access_fn = chrec_replace_initial_condition (access_fn, init_cond);
1852 access_fn = reset_evolution_in_loop (loop->num, access_fn, new_step);
1854 VEC_replace (tree, DR_ACCESS_FNS (dr), 0, access_fn);
1857 if (dump_file && (dump_flags & TDF_DETAILS))
1859 struct ptr_info_def *pi = DR_PTR_INFO (dr);
1861 fprintf (dump_file, "\nCreated dr for ");
1862 print_generic_expr (dump_file, memref, TDF_SLIM);
1863 fprintf (dump_file, "\n\tbase_address: ");
1864 print_generic_expr (dump_file, DR_BASE_ADDRESS (dr), TDF_SLIM);
1865 fprintf (dump_file, "\n\toffset from base address: ");
1866 print_generic_expr (dump_file, DR_OFFSET (dr), TDF_SLIM);
1867 fprintf (dump_file, "\n\tconstant offset from base address: ");
1868 print_generic_expr (dump_file, DR_INIT (dr), TDF_SLIM);
1869 fprintf (dump_file, "\n\tbase_object: ");
1870 print_generic_expr (dump_file, DR_BASE_OBJECT (dr), TDF_SLIM);
1871 fprintf (dump_file, "\n\tstep: ");
1872 print_generic_expr (dump_file, DR_STEP (dr), TDF_SLIM);
1873 fprintf (dump_file, "B\n\tmisalignment from base: ");
1874 print_generic_expr (dump_file, DR_OFFSET_MISALIGNMENT (dr), TDF_SLIM);
1875 if (DR_OFFSET_MISALIGNMENT (dr))
1876 fprintf (dump_file, "B");
1877 if (DR_ALIGNED_TO (dr))
1879 fprintf (dump_file, "\n\taligned to: ");
1880 print_generic_expr (dump_file, DR_ALIGNED_TO (dr), TDF_SLIM);
1882 fprintf (dump_file, "\n\tmemtag: ");
1883 print_generic_expr (dump_file, DR_MEMTAG (dr), TDF_SLIM);
1884 fprintf (dump_file, "\n");
1885 if (pi && pi->name_mem_tag)
1887 fprintf (dump_file, "\n\tnametag: ");
1888 print_generic_expr (dump_file, pi->name_mem_tag, TDF_SLIM);
1889 fprintf (dump_file, "\n");
1892 return dr;
1896 /* Returns true when all the functions of a tree_vec CHREC are the
1897 same. */
1899 static bool
1900 all_chrecs_equal_p (tree chrec)
1902 int j;
1904 for (j = 0; j < TREE_VEC_LENGTH (chrec) - 1; j++)
1906 tree chrec_j = TREE_VEC_ELT (chrec, j);
1907 tree chrec_j_1 = TREE_VEC_ELT (chrec, j + 1);
1908 if (!integer_zerop
1909 (chrec_fold_minus
1910 (integer_type_node, chrec_j, chrec_j_1)))
1911 return false;
1913 return true;
1916 /* Determine for each subscript in the data dependence relation DDR
1917 the distance. */
1919 void
1920 compute_subscript_distance (struct data_dependence_relation *ddr)
1922 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
1924 unsigned int i;
1926 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
1928 tree conflicts_a, conflicts_b, difference;
1929 struct subscript *subscript;
1931 subscript = DDR_SUBSCRIPT (ddr, i);
1932 conflicts_a = SUB_CONFLICTS_IN_A (subscript);
1933 conflicts_b = SUB_CONFLICTS_IN_B (subscript);
1935 if (TREE_CODE (conflicts_a) == TREE_VEC)
1937 if (!all_chrecs_equal_p (conflicts_a))
1939 SUB_DISTANCE (subscript) = chrec_dont_know;
1940 return;
1942 else
1943 conflicts_a = TREE_VEC_ELT (conflicts_a, 0);
1946 if (TREE_CODE (conflicts_b) == TREE_VEC)
1948 if (!all_chrecs_equal_p (conflicts_b))
1950 SUB_DISTANCE (subscript) = chrec_dont_know;
1951 return;
1953 else
1954 conflicts_b = TREE_VEC_ELT (conflicts_b, 0);
1957 difference = chrec_fold_minus
1958 (integer_type_node, conflicts_b, conflicts_a);
1960 if (evolution_function_is_constant_p (difference))
1961 SUB_DISTANCE (subscript) = difference;
1963 else
1964 SUB_DISTANCE (subscript) = chrec_dont_know;
1969 /* Initialize a ddr. */
1971 struct data_dependence_relation *
1972 initialize_data_dependence_relation (struct data_reference *a,
1973 struct data_reference *b)
1975 struct data_dependence_relation *res;
1976 bool differ_p;
1977 unsigned int i;
1979 res = XNEW (struct data_dependence_relation);
1980 DDR_A (res) = a;
1981 DDR_B (res) = b;
1983 if (a == NULL || b == NULL)
1985 DDR_ARE_DEPENDENT (res) = chrec_dont_know;
1986 return res;
1989 /* When A and B are arrays and their dimensions differ, we directly
1990 initialize the relation to "there is no dependence": chrec_known. */
1991 if (DR_BASE_OBJECT (a) && DR_BASE_OBJECT (b)
1992 && DR_NUM_DIMENSIONS (a) != DR_NUM_DIMENSIONS (b))
1994 DDR_ARE_DEPENDENT (res) = chrec_known;
1995 return res;
1998 /* Compare the bases of the data-refs. */
1999 if (!base_addr_differ_p (a, b, &differ_p))
2001 /* Can't determine whether the data-refs access the same memory
2002 region. */
2003 DDR_ARE_DEPENDENT (res) = chrec_dont_know;
2004 return res;
2006 if (differ_p)
2008 DDR_ARE_DEPENDENT (res) = chrec_known;
2009 return res;
2012 DDR_AFFINE_P (res) = true;
2013 DDR_ARE_DEPENDENT (res) = NULL_TREE;
2014 DDR_SUBSCRIPTS_VECTOR_INIT (res, DR_NUM_DIMENSIONS (a));
2015 DDR_SIZE_VECT (res) = 0;
2016 DDR_DIR_VECTS (res) = NULL;
2017 DDR_DIST_VECTS (res) = NULL;
2019 for (i = 0; i < DR_NUM_DIMENSIONS (a); i++)
2021 struct subscript *subscript;
2023 subscript = XNEW (struct subscript);
2024 SUB_CONFLICTS_IN_A (subscript) = chrec_dont_know;
2025 SUB_CONFLICTS_IN_B (subscript) = chrec_dont_know;
2026 SUB_LAST_CONFLICT (subscript) = chrec_dont_know;
2027 SUB_DISTANCE (subscript) = chrec_dont_know;
2028 VARRAY_PUSH_GENERIC_PTR (DDR_SUBSCRIPTS (res), subscript);
2031 return res;
2034 /* Set DDR_ARE_DEPENDENT to CHREC and finalize the subscript overlap
2035 description. */
2037 static inline void
2038 finalize_ddr_dependent (struct data_dependence_relation *ddr,
2039 tree chrec)
2041 if (dump_file && (dump_flags & TDF_DETAILS))
2043 fprintf (dump_file, "(dependence classified: ");
2044 print_generic_expr (dump_file, chrec, 0);
2045 fprintf (dump_file, ")\n");
2048 DDR_ARE_DEPENDENT (ddr) = chrec;
2049 varray_clear (DDR_SUBSCRIPTS (ddr));
2052 /* The dependence relation DDR cannot be represented by a distance
2053 vector. */
2055 static inline void
2056 non_affine_dependence_relation (struct data_dependence_relation *ddr)
2058 if (dump_file && (dump_flags & TDF_DETAILS))
2059 fprintf (dump_file, "(Dependence relation cannot be represented by distance vector.) \n");
2061 DDR_AFFINE_P (ddr) = false;
2066 /* This section contains the classic Banerjee tests. */
2068 /* Returns true iff CHREC_A and CHREC_B are not dependent on any index
2069 variables, i.e., if the ZIV (Zero Index Variable) test is true. */
2071 static inline bool
2072 ziv_subscript_p (tree chrec_a,
2073 tree chrec_b)
2075 return (evolution_function_is_constant_p (chrec_a)
2076 && evolution_function_is_constant_p (chrec_b));
2079 /* Returns true iff CHREC_A and CHREC_B are dependent on an index
2080 variable, i.e., if the SIV (Single Index Variable) test is true. */
2082 static bool
2083 siv_subscript_p (tree chrec_a,
2084 tree chrec_b)
2086 if ((evolution_function_is_constant_p (chrec_a)
2087 && evolution_function_is_univariate_p (chrec_b))
2088 || (evolution_function_is_constant_p (chrec_b)
2089 && evolution_function_is_univariate_p (chrec_a)))
2090 return true;
2092 if (evolution_function_is_univariate_p (chrec_a)
2093 && evolution_function_is_univariate_p (chrec_b))
2095 switch (TREE_CODE (chrec_a))
2097 case POLYNOMIAL_CHREC:
2098 switch (TREE_CODE (chrec_b))
2100 case POLYNOMIAL_CHREC:
2101 if (CHREC_VARIABLE (chrec_a) != CHREC_VARIABLE (chrec_b))
2102 return false;
2104 default:
2105 return true;
2108 default:
2109 return true;
2113 return false;
2116 /* Analyze a ZIV (Zero Index Variable) subscript. *OVERLAPS_A and
2117 *OVERLAPS_B are initialized to the functions that describe the
2118 relation between the elements accessed twice by CHREC_A and
2119 CHREC_B. For k >= 0, the following property is verified:
2121 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
2123 static void
2124 analyze_ziv_subscript (tree chrec_a,
2125 tree chrec_b,
2126 tree *overlaps_a,
2127 tree *overlaps_b,
2128 tree *last_conflicts)
2130 tree difference;
2132 if (dump_file && (dump_flags & TDF_DETAILS))
2133 fprintf (dump_file, "(analyze_ziv_subscript \n");
2135 difference = chrec_fold_minus (integer_type_node, chrec_a, chrec_b);
2137 switch (TREE_CODE (difference))
2139 case INTEGER_CST:
2140 if (integer_zerop (difference))
2142 /* The difference is equal to zero: the accessed index
2143 overlaps for each iteration in the loop. */
2144 *overlaps_a = integer_zero_node;
2145 *overlaps_b = integer_zero_node;
2146 *last_conflicts = chrec_dont_know;
2148 else
2150 /* The accesses do not overlap. */
2151 *overlaps_a = chrec_known;
2152 *overlaps_b = chrec_known;
2153 *last_conflicts = integer_zero_node;
2155 break;
2157 default:
2158 /* We're not sure whether the indexes overlap. For the moment,
2159 conservatively answer "don't know". */
2160 *overlaps_a = chrec_dont_know;
2161 *overlaps_b = chrec_dont_know;
2162 *last_conflicts = chrec_dont_know;
2163 break;
2166 if (dump_file && (dump_flags & TDF_DETAILS))
2167 fprintf (dump_file, ")\n");
2170 /* Get the real or estimated number of iterations for LOOPNUM, whichever is
2171 available. Return the number of iterations as a tree, or NULL_TREE if
2172 we don't know. */
2174 static tree
2175 get_number_of_iters_for_loop (int loopnum)
2177 tree numiter = number_of_iterations_in_loop (current_loops->parray[loopnum]);
2179 if (TREE_CODE (numiter) != INTEGER_CST)
2180 numiter = current_loops->parray[loopnum]->estimated_nb_iterations;
2181 if (chrec_contains_undetermined (numiter))
2182 return NULL_TREE;
2183 return numiter;
2186 /* Analyze a SIV (Single Index Variable) subscript where CHREC_A is a
2187 constant, and CHREC_B is an affine function. *OVERLAPS_A and
2188 *OVERLAPS_B are initialized to the functions that describe the
2189 relation between the elements accessed twice by CHREC_A and
2190 CHREC_B. For k >= 0, the following property is verified:
2192 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
2194 static void
2195 analyze_siv_subscript_cst_affine (tree chrec_a,
2196 tree chrec_b,
2197 tree *overlaps_a,
2198 tree *overlaps_b,
2199 tree *last_conflicts)
2201 bool value0, value1, value2;
2202 tree difference = chrec_fold_minus
2203 (integer_type_node, CHREC_LEFT (chrec_b), chrec_a);
2205 if (!chrec_is_positive (initial_condition (difference), &value0))
2207 *overlaps_a = chrec_dont_know;
2208 *overlaps_b = chrec_dont_know;
2209 *last_conflicts = chrec_dont_know;
2210 return;
2212 else
2214 if (value0 == false)
2216 if (!chrec_is_positive (CHREC_RIGHT (chrec_b), &value1))
2218 *overlaps_a = chrec_dont_know;
2219 *overlaps_b = chrec_dont_know;
2220 *last_conflicts = chrec_dont_know;
2221 return;
2223 else
2225 if (value1 == true)
2227 /* Example:
2228 chrec_a = 12
2229 chrec_b = {10, +, 1}
2232 if (tree_fold_divides_p (CHREC_RIGHT (chrec_b), difference))
2234 tree numiter;
2235 int loopnum = CHREC_VARIABLE (chrec_b);
2237 *overlaps_a = integer_zero_node;
2238 *overlaps_b = fold_build2 (EXACT_DIV_EXPR, integer_type_node,
2239 fold_build1 (ABS_EXPR,
2240 integer_type_node,
2241 difference),
2242 CHREC_RIGHT (chrec_b));
2243 *last_conflicts = integer_one_node;
2246 /* Perform weak-zero siv test to see if overlap is
2247 outside the loop bounds. */
2248 numiter = get_number_of_iters_for_loop (loopnum);
2250 if (numiter != NULL_TREE
2251 && TREE_CODE (*overlaps_b) == INTEGER_CST
2252 && tree_int_cst_lt (numiter, *overlaps_b))
2254 *overlaps_a = chrec_known;
2255 *overlaps_b = chrec_known;
2256 *last_conflicts = integer_zero_node;
2257 return;
2259 return;
2262 /* When the step does not divide the difference, there are
2263 no overlaps. */
2264 else
2266 *overlaps_a = chrec_known;
2267 *overlaps_b = chrec_known;
2268 *last_conflicts = integer_zero_node;
2269 return;
2273 else
2275 /* Example:
2276 chrec_a = 12
2277 chrec_b = {10, +, -1}
2279 In this case, chrec_a will not overlap with chrec_b. */
2280 *overlaps_a = chrec_known;
2281 *overlaps_b = chrec_known;
2282 *last_conflicts = integer_zero_node;
2283 return;
2287 else
2289 if (!chrec_is_positive (CHREC_RIGHT (chrec_b), &value2))
2291 *overlaps_a = chrec_dont_know;
2292 *overlaps_b = chrec_dont_know;
2293 *last_conflicts = chrec_dont_know;
2294 return;
2296 else
2298 if (value2 == false)
2300 /* Example:
2301 chrec_a = 3
2302 chrec_b = {10, +, -1}
2304 if (tree_fold_divides_p (CHREC_RIGHT (chrec_b), difference))
2306 tree numiter;
2307 int loopnum = CHREC_VARIABLE (chrec_b);
2309 *overlaps_a = integer_zero_node;
2310 *overlaps_b = fold_build2 (EXACT_DIV_EXPR,
2311 integer_type_node, difference,
2312 CHREC_RIGHT (chrec_b));
2313 *last_conflicts = integer_one_node;
2315 /* Perform weak-zero siv test to see if overlap is
2316 outside the loop bounds. */
2317 numiter = get_number_of_iters_for_loop (loopnum);
2319 if (numiter != NULL_TREE
2320 && TREE_CODE (*overlaps_b) == INTEGER_CST
2321 && tree_int_cst_lt (numiter, *overlaps_b))
2323 *overlaps_a = chrec_known;
2324 *overlaps_b = chrec_known;
2325 *last_conflicts = integer_zero_node;
2326 return;
2328 return;
2331 /* When the step does not divide the difference, there
2332 are no overlaps. */
2333 else
2335 *overlaps_a = chrec_known;
2336 *overlaps_b = chrec_known;
2337 *last_conflicts = integer_zero_node;
2338 return;
2341 else
2343 /* Example:
2344 chrec_a = 3
2345 chrec_b = {4, +, 1}
2347 In this case, chrec_a will not overlap with chrec_b. */
2348 *overlaps_a = chrec_known;
2349 *overlaps_b = chrec_known;
2350 *last_conflicts = integer_zero_node;
2351 return;
2358 /* Helper recursive function for initializing the matrix A. Returns
2359 the initial value of CHREC. */
2361 static int
2362 initialize_matrix_A (lambda_matrix A, tree chrec, unsigned index, int mult)
2364 gcc_assert (chrec);
2366 if (TREE_CODE (chrec) != POLYNOMIAL_CHREC)
2367 return int_cst_value (chrec);
2369 A[index][0] = mult * int_cst_value (CHREC_RIGHT (chrec));
2370 return initialize_matrix_A (A, CHREC_LEFT (chrec), index + 1, mult);
2373 #define FLOOR_DIV(x,y) ((x) / (y))
2375 /* Solves the special case of the Diophantine equation:
2376 | {0, +, STEP_A}_x (OVERLAPS_A) = {0, +, STEP_B}_y (OVERLAPS_B)
2378 Computes the descriptions OVERLAPS_A and OVERLAPS_B. NITER is the
2379 number of iterations that loops X and Y run. The overlaps will be
2380 constructed as evolutions in dimension DIM. */
2382 static void
2383 compute_overlap_steps_for_affine_univar (int niter, int step_a, int step_b,
2384 tree *overlaps_a, tree *overlaps_b,
2385 tree *last_conflicts, int dim)
2387 if (((step_a > 0 && step_b > 0)
2388 || (step_a < 0 && step_b < 0)))
2390 int step_overlaps_a, step_overlaps_b;
2391 int gcd_steps_a_b, last_conflict, tau2;
2393 gcd_steps_a_b = gcd (step_a, step_b);
2394 step_overlaps_a = step_b / gcd_steps_a_b;
2395 step_overlaps_b = step_a / gcd_steps_a_b;
2397 tau2 = FLOOR_DIV (niter, step_overlaps_a);
2398 tau2 = MIN (tau2, FLOOR_DIV (niter, step_overlaps_b));
2399 last_conflict = tau2;
2401 *overlaps_a = build_polynomial_chrec
2402 (dim, integer_zero_node,
2403 build_int_cst (NULL_TREE, step_overlaps_a));
2404 *overlaps_b = build_polynomial_chrec
2405 (dim, integer_zero_node,
2406 build_int_cst (NULL_TREE, step_overlaps_b));
2407 *last_conflicts = build_int_cst (NULL_TREE, last_conflict);
2410 else
2412 *overlaps_a = integer_zero_node;
2413 *overlaps_b = integer_zero_node;
2414 *last_conflicts = integer_zero_node;
2419 /* Solves the special case of a Diophantine equation where CHREC_A is
2420 an affine bivariate function, and CHREC_B is an affine univariate
2421 function. For example,
2423 | {{0, +, 1}_x, +, 1335}_y = {0, +, 1336}_z
2425 has the following overlapping functions:
2427 | x (t, u, v) = {{0, +, 1336}_t, +, 1}_v
2428 | y (t, u, v) = {{0, +, 1336}_u, +, 1}_v
2429 | z (t, u, v) = {{{0, +, 1}_t, +, 1335}_u, +, 1}_v
2431 FORNOW: This is a specialized implementation for a case occurring in
2432 a common benchmark. Implement the general algorithm. */
2434 static void
2435 compute_overlap_steps_for_affine_1_2 (tree chrec_a, tree chrec_b,
2436 tree *overlaps_a, tree *overlaps_b,
2437 tree *last_conflicts)
2439 bool xz_p, yz_p, xyz_p;
2440 int step_x, step_y, step_z;
2441 int niter_x, niter_y, niter_z, niter;
2442 tree numiter_x, numiter_y, numiter_z;
2443 tree overlaps_a_xz, overlaps_b_xz, last_conflicts_xz;
2444 tree overlaps_a_yz, overlaps_b_yz, last_conflicts_yz;
2445 tree overlaps_a_xyz, overlaps_b_xyz, last_conflicts_xyz;
2447 step_x = int_cst_value (CHREC_RIGHT (CHREC_LEFT (chrec_a)));
2448 step_y = int_cst_value (CHREC_RIGHT (chrec_a));
2449 step_z = int_cst_value (CHREC_RIGHT (chrec_b));
2451 numiter_x = get_number_of_iters_for_loop (CHREC_VARIABLE (CHREC_LEFT (chrec_a)));
2452 numiter_y = get_number_of_iters_for_loop (CHREC_VARIABLE (chrec_a));
2453 numiter_z = get_number_of_iters_for_loop (CHREC_VARIABLE (chrec_b));
2455 if (numiter_x == NULL_TREE || numiter_y == NULL_TREE
2456 || numiter_z == NULL_TREE)
2458 *overlaps_a = chrec_dont_know;
2459 *overlaps_b = chrec_dont_know;
2460 *last_conflicts = chrec_dont_know;
2461 return;
2464 niter_x = int_cst_value (numiter_x);
2465 niter_y = int_cst_value (numiter_y);
2466 niter_z = int_cst_value (numiter_z);
2468 niter = MIN (niter_x, niter_z);
2469 compute_overlap_steps_for_affine_univar (niter, step_x, step_z,
2470 &overlaps_a_xz,
2471 &overlaps_b_xz,
2472 &last_conflicts_xz, 1);
2473 niter = MIN (niter_y, niter_z);
2474 compute_overlap_steps_for_affine_univar (niter, step_y, step_z,
2475 &overlaps_a_yz,
2476 &overlaps_b_yz,
2477 &last_conflicts_yz, 2);
2478 niter = MIN (niter_x, niter_z);
2479 niter = MIN (niter_y, niter);
2480 compute_overlap_steps_for_affine_univar (niter, step_x + step_y, step_z,
2481 &overlaps_a_xyz,
2482 &overlaps_b_xyz,
2483 &last_conflicts_xyz, 3);
2485 xz_p = !integer_zerop (last_conflicts_xz);
2486 yz_p = !integer_zerop (last_conflicts_yz);
2487 xyz_p = !integer_zerop (last_conflicts_xyz);
2489 if (xz_p || yz_p || xyz_p)
2491 *overlaps_a = make_tree_vec (2);
2492 TREE_VEC_ELT (*overlaps_a, 0) = integer_zero_node;
2493 TREE_VEC_ELT (*overlaps_a, 1) = integer_zero_node;
2494 *overlaps_b = integer_zero_node;
2495 if (xz_p)
2497 TREE_VEC_ELT (*overlaps_a, 0) =
2498 chrec_fold_plus (integer_type_node, TREE_VEC_ELT (*overlaps_a, 0),
2499 overlaps_a_xz);
2500 *overlaps_b =
2501 chrec_fold_plus (integer_type_node, *overlaps_b, overlaps_b_xz);
2502 *last_conflicts = last_conflicts_xz;
2504 if (yz_p)
2506 TREE_VEC_ELT (*overlaps_a, 1) =
2507 chrec_fold_plus (integer_type_node, TREE_VEC_ELT (*overlaps_a, 1),
2508 overlaps_a_yz);
2509 *overlaps_b =
2510 chrec_fold_plus (integer_type_node, *overlaps_b, overlaps_b_yz);
2511 *last_conflicts = last_conflicts_yz;
2513 if (xyz_p)
2515 TREE_VEC_ELT (*overlaps_a, 0) =
2516 chrec_fold_plus (integer_type_node, TREE_VEC_ELT (*overlaps_a, 0),
2517 overlaps_a_xyz);
2518 TREE_VEC_ELT (*overlaps_a, 1) =
2519 chrec_fold_plus (integer_type_node, TREE_VEC_ELT (*overlaps_a, 1),
2520 overlaps_a_xyz);
2521 *overlaps_b =
2522 chrec_fold_plus (integer_type_node, *overlaps_b, overlaps_b_xyz);
2523 *last_conflicts = last_conflicts_xyz;
2526 else
2528 *overlaps_a = integer_zero_node;
2529 *overlaps_b = integer_zero_node;
2530 *last_conflicts = integer_zero_node;
2534 /* Determines the overlapping elements due to accesses CHREC_A and
2535 CHREC_B, that are affine functions. This is a part of the
2536 subscript analyzer. */
2538 static void
2539 analyze_subscript_affine_affine (tree chrec_a,
2540 tree chrec_b,
2541 tree *overlaps_a,
2542 tree *overlaps_b,
2543 tree *last_conflicts)
2545 unsigned nb_vars_a, nb_vars_b, dim;
2546 int init_a, init_b, gamma, gcd_alpha_beta;
2547 int tau1, tau2;
2548 lambda_matrix A, U, S;
2549 tree difference = chrec_fold_minus (integer_type_node, chrec_a, chrec_b);
2551 if (integer_zerop (difference))
2553 /* The difference is equal to zero: the accessed index
2554 overlaps for each iteration in the loop. */
2555 *overlaps_a = integer_zero_node;
2556 *overlaps_b = integer_zero_node;
2557 *last_conflicts = chrec_dont_know;
2558 return;
2560 if (dump_file && (dump_flags & TDF_DETAILS))
2561 fprintf (dump_file, "(analyze_subscript_affine_affine \n");
2563 /* For determining the initial intersection, we have to solve a
2564 Diophantine equation. This is the most time consuming part.
2566 For answering to the question: "Is there a dependence?" we have
2567 to prove that there exists a solution to the Diophantine
2568 equation, and that the solution is in the iteration domain,
2569 i.e. the solution is positive or zero, and that the solution
2570 happens before the upper bound loop.nb_iterations. Otherwise
2571 there is no dependence. This function outputs a description of
2572 the iterations that hold the intersections. */
2575 nb_vars_a = nb_vars_in_chrec (chrec_a);
2576 nb_vars_b = nb_vars_in_chrec (chrec_b);
2578 dim = nb_vars_a + nb_vars_b;
2579 U = lambda_matrix_new (dim, dim);
2580 A = lambda_matrix_new (dim, 1);
2581 S = lambda_matrix_new (dim, 1);
2583 init_a = initialize_matrix_A (A, chrec_a, 0, 1);
2584 init_b = initialize_matrix_A (A, chrec_b, nb_vars_a, -1);
2585 gamma = init_b - init_a;
2587 /* Don't do all the hard work of solving the Diophantine equation
2588 when we already know the solution: for example,
2589 | {3, +, 1}_1
2590 | {3, +, 4}_2
2591 | gamma = 3 - 3 = 0.
2592 Then the first overlap occurs during the first iterations:
2593 | {3, +, 1}_1 ({0, +, 4}_x) = {3, +, 4}_2 ({0, +, 1}_x)
2595 if (gamma == 0)
2597 if (nb_vars_a == 1 && nb_vars_b == 1)
2599 int step_a, step_b;
2600 int niter, niter_a, niter_b;
2601 tree numiter_a, numiter_b;
2603 numiter_a = get_number_of_iters_for_loop (CHREC_VARIABLE (chrec_a));
2604 numiter_b = get_number_of_iters_for_loop (CHREC_VARIABLE (chrec_b));
2605 if (numiter_a == NULL_TREE || numiter_b == NULL_TREE)
2607 *overlaps_a = chrec_dont_know;
2608 *overlaps_b = chrec_dont_know;
2609 *last_conflicts = chrec_dont_know;
2610 return;
2613 niter_a = int_cst_value (numiter_a);
2614 niter_b = int_cst_value (numiter_b);
2615 niter = MIN (niter_a, niter_b);
2617 step_a = int_cst_value (CHREC_RIGHT (chrec_a));
2618 step_b = int_cst_value (CHREC_RIGHT (chrec_b));
2620 compute_overlap_steps_for_affine_univar (niter, step_a, step_b,
2621 overlaps_a, overlaps_b,
2622 last_conflicts, 1);
2625 else if (nb_vars_a == 2 && nb_vars_b == 1)
2626 compute_overlap_steps_for_affine_1_2
2627 (chrec_a, chrec_b, overlaps_a, overlaps_b, last_conflicts);
2629 else if (nb_vars_a == 1 && nb_vars_b == 2)
2630 compute_overlap_steps_for_affine_1_2
2631 (chrec_b, chrec_a, overlaps_b, overlaps_a, last_conflicts);
2633 else
2635 *overlaps_a = chrec_dont_know;
2636 *overlaps_b = chrec_dont_know;
2637 *last_conflicts = chrec_dont_know;
2639 return;
2642 /* U.A = S */
2643 lambda_matrix_right_hermite (A, dim, 1, S, U);
2645 if (S[0][0] < 0)
2647 S[0][0] *= -1;
2648 lambda_matrix_row_negate (U, dim, 0);
2650 gcd_alpha_beta = S[0][0];
2652 /* The classic "gcd-test". */
2653 if (!int_divides_p (gcd_alpha_beta, gamma))
2655 /* The "gcd-test" has determined that there is no integer
2656 solution, i.e. there is no dependence. */
2657 *overlaps_a = chrec_known;
2658 *overlaps_b = chrec_known;
2659 *last_conflicts = integer_zero_node;
2662 /* Both access functions are univariate. This includes SIV and MIV cases. */
2663 else if (nb_vars_a == 1 && nb_vars_b == 1)
2665 /* Both functions should have the same evolution sign. */
2666 if (((A[0][0] > 0 && -A[1][0] > 0)
2667 || (A[0][0] < 0 && -A[1][0] < 0)))
2669 /* The solutions are given by:
2671 | [GAMMA/GCD_ALPHA_BETA t].[u11 u12] = [x0]
2672 | [u21 u22] [y0]
2674 For a given integer t. Using the following variables,
2676 | i0 = u11 * gamma / gcd_alpha_beta
2677 | j0 = u12 * gamma / gcd_alpha_beta
2678 | i1 = u21
2679 | j1 = u22
2681 the solutions are:
2683 | x0 = i0 + i1 * t,
2684 | y0 = j0 + j1 * t. */
2686 int i0, j0, i1, j1;
2688 /* X0 and Y0 are the first iterations for which there is a
2689 dependence. X0, Y0 are two solutions of the Diophantine
2690 equation: chrec_a (X0) = chrec_b (Y0). */
2691 int x0, y0;
2692 int niter, niter_a, niter_b;
2693 tree numiter_a, numiter_b;
2695 numiter_a = get_number_of_iters_for_loop (CHREC_VARIABLE (chrec_a));
2696 numiter_b = get_number_of_iters_for_loop (CHREC_VARIABLE (chrec_b));
2698 if (numiter_a == NULL_TREE || numiter_b == NULL_TREE)
2700 *overlaps_a = chrec_dont_know;
2701 *overlaps_b = chrec_dont_know;
2702 *last_conflicts = chrec_dont_know;
2703 return;
2706 niter_a = int_cst_value (numiter_a);
2707 niter_b = int_cst_value (numiter_b);
2708 niter = MIN (niter_a, niter_b);
2710 i0 = U[0][0] * gamma / gcd_alpha_beta;
2711 j0 = U[0][1] * gamma / gcd_alpha_beta;
2712 i1 = U[1][0];
2713 j1 = U[1][1];
2715 if ((i1 == 0 && i0 < 0)
2716 || (j1 == 0 && j0 < 0))
2718 /* There is no solution.
2719 FIXME: The case "i0 > nb_iterations, j0 > nb_iterations"
2720 falls in here, but for the moment we don't look at the
2721 upper bound of the iteration domain. */
2722 *overlaps_a = chrec_known;
2723 *overlaps_b = chrec_known;
2724 *last_conflicts = integer_zero_node;
2727 else
2729 if (i1 > 0)
2731 tau1 = CEIL (-i0, i1);
2732 tau2 = FLOOR_DIV (niter - i0, i1);
2734 if (j1 > 0)
2736 int last_conflict, min_multiple;
2737 tau1 = MAX (tau1, CEIL (-j0, j1));
2738 tau2 = MIN (tau2, FLOOR_DIV (niter - j0, j1));
2740 x0 = i1 * tau1 + i0;
2741 y0 = j1 * tau1 + j0;
2743 /* At this point (x0, y0) is one of the
2744 solutions to the Diophantine equation. The
2745 next step has to compute the smallest
2746 positive solution: the first conflicts. */
2747 min_multiple = MIN (x0 / i1, y0 / j1);
2748 x0 -= i1 * min_multiple;
2749 y0 -= j1 * min_multiple;
2751 tau1 = (x0 - i0)/i1;
2752 last_conflict = tau2 - tau1;
2754 /* If the overlap occurs outside of the bounds of the
2755 loop, there is no dependence. */
2756 if (x0 > niter || y0 > niter)
2759 *overlaps_a = chrec_known;
2760 *overlaps_b = chrec_known;
2761 *last_conflicts = integer_zero_node;
2763 else
2765 *overlaps_a = build_polynomial_chrec
2767 build_int_cst (NULL_TREE, x0),
2768 build_int_cst (NULL_TREE, i1));
2769 *overlaps_b = build_polynomial_chrec
2771 build_int_cst (NULL_TREE, y0),
2772 build_int_cst (NULL_TREE, j1));
2773 *last_conflicts = build_int_cst (NULL_TREE, last_conflict);
2776 else
2778 /* FIXME: For the moment, the upper bound of the
2779 iteration domain for j is not checked. */
2780 *overlaps_a = chrec_dont_know;
2781 *overlaps_b = chrec_dont_know;
2782 *last_conflicts = chrec_dont_know;
2786 else
2788 /* FIXME: For the moment, the upper bound of the
2789 iteration domain for i is not checked. */
2790 *overlaps_a = chrec_dont_know;
2791 *overlaps_b = chrec_dont_know;
2792 *last_conflicts = chrec_dont_know;
2796 else
2798 *overlaps_a = chrec_dont_know;
2799 *overlaps_b = chrec_dont_know;
2800 *last_conflicts = chrec_dont_know;
2804 else
2806 *overlaps_a = chrec_dont_know;
2807 *overlaps_b = chrec_dont_know;
2808 *last_conflicts = chrec_dont_know;
2812 if (dump_file && (dump_flags & TDF_DETAILS))
2814 fprintf (dump_file, " (overlaps_a = ");
2815 print_generic_expr (dump_file, *overlaps_a, 0);
2816 fprintf (dump_file, ")\n (overlaps_b = ");
2817 print_generic_expr (dump_file, *overlaps_b, 0);
2818 fprintf (dump_file, ")\n");
2821 if (dump_file && (dump_flags & TDF_DETAILS))
2822 fprintf (dump_file, ")\n");
2825 /* Analyze a SIV (Single Index Variable) subscript. *OVERLAPS_A and
2826 *OVERLAPS_B are initialized to the functions that describe the
2827 relation between the elements accessed twice by CHREC_A and
2828 CHREC_B. For k >= 0, the following property is verified:
2830 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
2832 static void
2833 analyze_siv_subscript (tree chrec_a,
2834 tree chrec_b,
2835 tree *overlaps_a,
2836 tree *overlaps_b,
2837 tree *last_conflicts)
2839 if (dump_file && (dump_flags & TDF_DETAILS))
2840 fprintf (dump_file, "(analyze_siv_subscript \n");
2842 if (evolution_function_is_constant_p (chrec_a)
2843 && evolution_function_is_affine_p (chrec_b))
2844 analyze_siv_subscript_cst_affine (chrec_a, chrec_b,
2845 overlaps_a, overlaps_b, last_conflicts);
2847 else if (evolution_function_is_affine_p (chrec_a)
2848 && evolution_function_is_constant_p (chrec_b))
2849 analyze_siv_subscript_cst_affine (chrec_b, chrec_a,
2850 overlaps_b, overlaps_a, last_conflicts);
2852 else if (evolution_function_is_affine_p (chrec_a)
2853 && evolution_function_is_affine_p (chrec_b))
2854 analyze_subscript_affine_affine (chrec_a, chrec_b,
2855 overlaps_a, overlaps_b, last_conflicts);
2856 else
2858 *overlaps_a = chrec_dont_know;
2859 *overlaps_b = chrec_dont_know;
2860 *last_conflicts = chrec_dont_know;
2863 if (dump_file && (dump_flags & TDF_DETAILS))
2864 fprintf (dump_file, ")\n");
2867 /* Return true when the evolution steps of an affine CHREC divide the
2868 constant CST. */
2870 static bool
2871 chrec_steps_divide_constant_p (tree chrec,
2872 tree cst)
2874 switch (TREE_CODE (chrec))
2876 case POLYNOMIAL_CHREC:
2877 return (tree_fold_divides_p (CHREC_RIGHT (chrec), cst)
2878 && chrec_steps_divide_constant_p (CHREC_LEFT (chrec), cst));
2880 default:
2881 /* On the initial condition, return true. */
2882 return true;
2886 /* Analyze a MIV (Multiple Index Variable) subscript. *OVERLAPS_A and
2887 *OVERLAPS_B are initialized to the functions that describe the
2888 relation between the elements accessed twice by CHREC_A and
2889 CHREC_B. For k >= 0, the following property is verified:
2891 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
2893 static void
2894 analyze_miv_subscript (tree chrec_a,
2895 tree chrec_b,
2896 tree *overlaps_a,
2897 tree *overlaps_b,
2898 tree *last_conflicts)
2900 /* FIXME: This is a MIV subscript, not yet handled.
2901 Example: (A[{1, +, 1}_1] vs. A[{1, +, 1}_2]) that comes from
2902 (A[i] vs. A[j]).
2904 In the SIV test we had to solve a Diophantine equation with two
2905 variables. In the MIV case we have to solve a Diophantine
2906 equation with 2*n variables (if the subscript uses n IVs).
2908 tree difference;
2910 if (dump_file && (dump_flags & TDF_DETAILS))
2911 fprintf (dump_file, "(analyze_miv_subscript \n");
2913 difference = chrec_fold_minus (integer_type_node, chrec_a, chrec_b);
2915 if (chrec_zerop (difference))
2917 /* Access functions are the same: all the elements are accessed
2918 in the same order. */
2919 *overlaps_a = integer_zero_node;
2920 *overlaps_b = integer_zero_node;
2921 *last_conflicts = get_number_of_iters_for_loop (CHREC_VARIABLE (chrec_a));
2925 else if (evolution_function_is_constant_p (difference)
2926 /* For the moment, the following is verified:
2927 evolution_function_is_affine_multivariate_p (chrec_a) */
2928 && !chrec_steps_divide_constant_p (chrec_a, difference))
2930 /* testsuite/.../ssa-chrec-33.c
2931 {{21, +, 2}_1, +, -2}_2 vs. {{20, +, 2}_1, +, -2}_2
2933 The difference is 1, and the evolution steps are equal to 2,
2934 consequently there are no overlapping elements. */
2935 *overlaps_a = chrec_known;
2936 *overlaps_b = chrec_known;
2937 *last_conflicts = integer_zero_node;
2940 else if (evolution_function_is_affine_multivariate_p (chrec_a)
2941 && evolution_function_is_affine_multivariate_p (chrec_b))
2943 /* testsuite/.../ssa-chrec-35.c
2944 {0, +, 1}_2 vs. {0, +, 1}_3
2945 the overlapping elements are respectively located at iterations:
2946 {0, +, 1}_x and {0, +, 1}_x,
2947 in other words, we have the equality:
2948 {0, +, 1}_2 ({0, +, 1}_x) = {0, +, 1}_3 ({0, +, 1}_x)
2950 Other examples:
2951 {{0, +, 1}_1, +, 2}_2 ({0, +, 1}_x, {0, +, 1}_y) =
2952 {0, +, 1}_1 ({{0, +, 1}_x, +, 2}_y)
2954 {{0, +, 2}_1, +, 3}_2 ({0, +, 1}_y, {0, +, 1}_x) =
2955 {{0, +, 3}_1, +, 2}_2 ({0, +, 1}_x, {0, +, 1}_y)
2957 analyze_subscript_affine_affine (chrec_a, chrec_b,
2958 overlaps_a, overlaps_b, last_conflicts);
2961 else
2963 /* When the analysis is too difficult, answer "don't know". */
2964 *overlaps_a = chrec_dont_know;
2965 *overlaps_b = chrec_dont_know;
2966 *last_conflicts = chrec_dont_know;
2969 if (dump_file && (dump_flags & TDF_DETAILS))
2970 fprintf (dump_file, ")\n");
2973 /* Determines the iterations for which CHREC_A is equal to CHREC_B.
2974 OVERLAP_ITERATIONS_A and OVERLAP_ITERATIONS_B are initialized with
2975 two functions that describe the iterations that contain conflicting
2976 elements.
2978 Remark: For an integer k >= 0, the following equality is true:
2980 CHREC_A (OVERLAP_ITERATIONS_A (k)) == CHREC_B (OVERLAP_ITERATIONS_B (k)).
2983 static void
2984 analyze_overlapping_iterations (tree chrec_a,
2985 tree chrec_b,
2986 tree *overlap_iterations_a,
2987 tree *overlap_iterations_b,
2988 tree *last_conflicts)
2990 if (dump_file && (dump_flags & TDF_DETAILS))
2992 fprintf (dump_file, "(analyze_overlapping_iterations \n");
2993 fprintf (dump_file, " (chrec_a = ");
2994 print_generic_expr (dump_file, chrec_a, 0);
2995 fprintf (dump_file, ")\n chrec_b = ");
2996 print_generic_expr (dump_file, chrec_b, 0);
2997 fprintf (dump_file, ")\n");
3000 if (chrec_a == NULL_TREE
3001 || chrec_b == NULL_TREE
3002 || chrec_contains_undetermined (chrec_a)
3003 || chrec_contains_undetermined (chrec_b)
3004 || chrec_contains_symbols (chrec_a)
3005 || chrec_contains_symbols (chrec_b))
3007 *overlap_iterations_a = chrec_dont_know;
3008 *overlap_iterations_b = chrec_dont_know;
3011 else if (ziv_subscript_p (chrec_a, chrec_b))
3012 analyze_ziv_subscript (chrec_a, chrec_b,
3013 overlap_iterations_a, overlap_iterations_b,
3014 last_conflicts);
3016 else if (siv_subscript_p (chrec_a, chrec_b))
3017 analyze_siv_subscript (chrec_a, chrec_b,
3018 overlap_iterations_a, overlap_iterations_b,
3019 last_conflicts);
3021 else
3022 analyze_miv_subscript (chrec_a, chrec_b,
3023 overlap_iterations_a, overlap_iterations_b,
3024 last_conflicts);
3026 if (dump_file && (dump_flags & TDF_DETAILS))
3028 fprintf (dump_file, " (overlap_iterations_a = ");
3029 print_generic_expr (dump_file, *overlap_iterations_a, 0);
3030 fprintf (dump_file, ")\n (overlap_iterations_b = ");
3031 print_generic_expr (dump_file, *overlap_iterations_b, 0);
3032 fprintf (dump_file, ")\n");
3038 /* This section contains the affine functions dependences detector. */
3040 /* Computes the conflicting iterations, and initialize DDR. */
3042 static void
3043 subscript_dependence_tester (struct data_dependence_relation *ddr)
3045 unsigned int i;
3046 struct data_reference *dra = DDR_A (ddr);
3047 struct data_reference *drb = DDR_B (ddr);
3048 tree last_conflicts;
3050 if (dump_file && (dump_flags & TDF_DETAILS))
3051 fprintf (dump_file, "(subscript_dependence_tester \n");
3053 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
3055 tree overlaps_a, overlaps_b;
3056 struct subscript *subscript = DDR_SUBSCRIPT (ddr, i);
3058 analyze_overlapping_iterations (DR_ACCESS_FN (dra, i),
3059 DR_ACCESS_FN (drb, i),
3060 &overlaps_a, &overlaps_b,
3061 &last_conflicts);
3063 if (chrec_contains_undetermined (overlaps_a)
3064 || chrec_contains_undetermined (overlaps_b))
3066 finalize_ddr_dependent (ddr, chrec_dont_know);
3067 break;
3070 else if (overlaps_a == chrec_known
3071 || overlaps_b == chrec_known)
3073 finalize_ddr_dependent (ddr, chrec_known);
3074 break;
3077 else
3079 SUB_CONFLICTS_IN_A (subscript) = overlaps_a;
3080 SUB_CONFLICTS_IN_B (subscript) = overlaps_b;
3081 SUB_LAST_CONFLICT (subscript) = last_conflicts;
3085 if (dump_file && (dump_flags & TDF_DETAILS))
3086 fprintf (dump_file, ")\n");
3089 /* Compute the classic per loop distance vector.
3091 DDR is the data dependence relation to build a vector from.
3092 NB_LOOPS is the total number of loops we are considering.
3093 FIRST_LOOP_DEPTH is the loop->depth of the first loop in the analyzed
3094 loop nest.
3095 Return FALSE when fail to represent the data dependence as a distance
3096 vector.
3097 Return TRUE otherwise. */
3099 static bool
3100 build_classic_dist_vector (struct data_dependence_relation *ddr,
3101 int nb_loops, int first_loop_depth)
3103 unsigned i;
3104 lambda_vector dist_v, init_v;
3105 bool init_b = false;
3107 DDR_SIZE_VECT (ddr) = nb_loops;
3108 dist_v = lambda_vector_new (nb_loops);
3109 init_v = lambda_vector_new (nb_loops);
3110 lambda_vector_clear (dist_v, nb_loops);
3111 lambda_vector_clear (init_v, nb_loops);
3113 if (DDR_ARE_DEPENDENT (ddr) != NULL_TREE)
3114 return true;
3116 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
3118 tree access_fn_a, access_fn_b;
3119 struct subscript *subscript = DDR_SUBSCRIPT (ddr, i);
3121 if (chrec_contains_undetermined (SUB_DISTANCE (subscript)))
3123 non_affine_dependence_relation (ddr);
3124 return true;
3127 access_fn_a = DR_ACCESS_FN (DDR_A (ddr), i);
3128 access_fn_b = DR_ACCESS_FN (DDR_B (ddr), i);
3130 if (TREE_CODE (access_fn_a) == POLYNOMIAL_CHREC
3131 && TREE_CODE (access_fn_b) == POLYNOMIAL_CHREC)
3133 int dist, loop_nb, loop_depth;
3134 int loop_nb_a = CHREC_VARIABLE (access_fn_a);
3135 int loop_nb_b = CHREC_VARIABLE (access_fn_b);
3136 struct loop *loop_a = current_loops->parray[loop_nb_a];
3137 struct loop *loop_b = current_loops->parray[loop_nb_b];
3139 /* If the loop for either variable is at a lower depth than
3140 the first_loop's depth, then we can't possibly have a
3141 dependency at this level of the loop. */
3143 if (loop_a->depth < first_loop_depth
3144 || loop_b->depth < first_loop_depth)
3145 return false;
3147 if (loop_nb_a != loop_nb_b
3148 && !flow_loop_nested_p (loop_a, loop_b)
3149 && !flow_loop_nested_p (loop_b, loop_a))
3151 /* Example: when there are two consecutive loops,
3153 | loop_1
3154 | A[{0, +, 1}_1]
3155 | endloop_1
3156 | loop_2
3157 | A[{0, +, 1}_2]
3158 | endloop_2
3160 the dependence relation cannot be captured by the
3161 distance abstraction. */
3162 non_affine_dependence_relation (ddr);
3163 return true;
3166 /* The dependence is carried by the outermost loop. Example:
3167 | loop_1
3168 | A[{4, +, 1}_1]
3169 | loop_2
3170 | A[{5, +, 1}_2]
3171 | endloop_2
3172 | endloop_1
3173 In this case, the dependence is carried by loop_1. */
3174 loop_nb = loop_nb_a < loop_nb_b ? loop_nb_a : loop_nb_b;
3175 loop_depth = current_loops->parray[loop_nb]->depth - first_loop_depth;
3177 /* If the loop number is still greater than the number of
3178 loops we've been asked to analyze, or negative,
3179 something is borked. */
3180 gcc_assert (loop_depth >= 0);
3181 gcc_assert (loop_depth < nb_loops);
3182 if (chrec_contains_undetermined (SUB_DISTANCE (subscript)))
3184 non_affine_dependence_relation (ddr);
3185 return true;
3188 dist = int_cst_value (SUB_DISTANCE (subscript));
3190 /* This is the subscript coupling test.
3191 | loop i = 0, N, 1
3192 | T[i+1][i] = ...
3193 | ... = T[i][i]
3194 | endloop
3195 There is no dependence. */
3196 if (init_v[loop_depth] != 0
3197 && dist_v[loop_depth] != dist)
3199 finalize_ddr_dependent (ddr, chrec_known);
3200 return true;
3203 dist_v[loop_depth] = dist;
3204 init_v[loop_depth] = 1;
3205 init_b = true;
3209 /* Save the distance vector if we initialized one. */
3210 if (init_b)
3212 lambda_vector save_v;
3214 /* Verify a basic constraint: classic distance vectors should always
3215 be lexicographically positive. */
3216 if (!lambda_vector_lexico_pos (dist_v, DDR_SIZE_VECT (ddr)))
3218 if (DDR_SIZE_VECT (ddr) == 1)
3219 /* This one is simple to fix, and can be fixed.
3220 Multidimensional arrays cannot be fixed that simply. */
3221 lambda_vector_negate (dist_v, dist_v, DDR_SIZE_VECT (ddr));
3222 else
3223 /* This is not valid: we need the delta test for properly
3224 fixing all this. */
3225 return false;
3228 save_v = lambda_vector_new (DDR_SIZE_VECT (ddr));
3229 lambda_vector_copy (dist_v, save_v, DDR_SIZE_VECT (ddr));
3230 VEC_safe_push (lambda_vector, heap, DDR_DIST_VECTS (ddr), save_v);
3232 /* There is nothing more to do when there are no outer loops. */
3233 if (DDR_SIZE_VECT (ddr) == 1)
3234 goto classic_dist_done;
3237 /* There is a distance of 1 on all the outer loops:
3239 Example: there is a dependence of distance 1 on loop_1 for the array A.
3240 | loop_1
3241 | A[5] = ...
3242 | endloop
3245 struct loop *lca, *loop_a, *loop_b;
3246 struct data_reference *a = DDR_A (ddr);
3247 struct data_reference *b = DDR_B (ddr);
3248 int lca_depth;
3249 loop_a = loop_containing_stmt (DR_STMT (a));
3250 loop_b = loop_containing_stmt (DR_STMT (b));
3252 /* Get the common ancestor loop. */
3253 lca = find_common_loop (loop_a, loop_b);
3254 lca_depth = lca->depth - first_loop_depth;
3256 gcc_assert (lca_depth >= 0);
3257 gcc_assert (lca_depth < nb_loops);
3259 /* For each outer loop where init_v is not set, the accesses are
3260 in dependence of distance 1 in the loop. */
3261 while (lca->depth != 0)
3263 /* If we're considering just a sub-nest, then don't record
3264 any information on the outer loops. */
3265 if (lca_depth < 0)
3266 break;
3268 gcc_assert (lca_depth < nb_loops);
3270 /* If we haven't yet determined a distance for this outer
3271 loop, push a new distance vector composed of the previous
3272 distance, and a distance of 1 for this outer loop.
3273 Example:
3275 | loop_1
3276 | loop_2
3277 | A[10]
3278 | endloop_2
3279 | endloop_1
3281 Saved vectors are of the form (dist_in_1, dist_in_2).
3282 First, we save (0, 1), then we have to save (1, 0). */
3283 if (init_v[lca_depth] == 0)
3285 lambda_vector save_v = lambda_vector_new (DDR_SIZE_VECT (ddr));
3287 lambda_vector_copy (dist_v, save_v, DDR_SIZE_VECT (ddr));
3288 save_v[lca_depth] = 1;
3289 VEC_safe_push (lambda_vector, heap, DDR_DIST_VECTS (ddr), save_v);
3292 lca = lca->outer;
3293 lca_depth = lca->depth - first_loop_depth;
3297 classic_dist_done:;
3299 if (dump_file && (dump_flags & TDF_DETAILS))
3301 fprintf (dump_file, "(build_classic_dist_vector\n");
3303 for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++)
3305 fprintf (dump_file, " dist_vector = (");
3306 print_lambda_vector (dump_file, DDR_DIST_VECT (ddr, i),
3307 DDR_SIZE_VECT (ddr));
3308 fprintf (dump_file, " )\n");
3310 fprintf (dump_file, ")\n");
3313 return true;
3316 /* Compute the classic per loop direction vector.
3318 DDR is the data dependence relation to build a vector from.
3319 NB_LOOPS is the total number of loops we are considering.
3320 FIRST_LOOP_DEPTH is the loop->depth of the first loop in the analyzed
3321 loop nest.
3322 Return FALSE if the dependence relation is outside of the loop nest
3323 at FIRST_LOOP_DEPTH.
3324 Return TRUE otherwise. */
3326 static bool
3327 build_classic_dir_vector (struct data_dependence_relation *ddr,
3328 int nb_loops, int first_loop_depth)
3330 unsigned i;
3331 lambda_vector dir_v, init_v;
3332 bool init_b = false;
3334 dir_v = lambda_vector_new (nb_loops);
3335 init_v = lambda_vector_new (nb_loops);
3336 lambda_vector_clear (dir_v, nb_loops);
3337 lambda_vector_clear (init_v, nb_loops);
3339 DDR_SIZE_VECT (ddr) = nb_loops;
3341 if (DDR_ARE_DEPENDENT (ddr) != NULL_TREE)
3342 return true;
3344 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
3346 tree access_fn_a, access_fn_b;
3347 struct subscript *subscript = DDR_SUBSCRIPT (ddr, i);
3349 if (chrec_contains_undetermined (SUB_DISTANCE (subscript)))
3351 non_affine_dependence_relation (ddr);
3352 return true;
3355 access_fn_a = DR_ACCESS_FN (DDR_A (ddr), i);
3356 access_fn_b = DR_ACCESS_FN (DDR_B (ddr), i);
3357 if (TREE_CODE (access_fn_a) == POLYNOMIAL_CHREC
3358 && TREE_CODE (access_fn_b) == POLYNOMIAL_CHREC)
3360 int dist, loop_nb, loop_depth;
3361 enum data_dependence_direction dir = dir_star;
3362 int loop_nb_a = CHREC_VARIABLE (access_fn_a);
3363 int loop_nb_b = CHREC_VARIABLE (access_fn_b);
3364 struct loop *loop_a = current_loops->parray[loop_nb_a];
3365 struct loop *loop_b = current_loops->parray[loop_nb_b];
3367 /* If the loop for either variable is at a lower depth than
3368 the first_loop's depth, then we can't possibly have a
3369 dependency at this level of the loop. */
3371 if (loop_a->depth < first_loop_depth
3372 || loop_b->depth < first_loop_depth)
3373 return false;
3375 if (loop_nb_a != loop_nb_b
3376 && !flow_loop_nested_p (loop_a, loop_b)
3377 && !flow_loop_nested_p (loop_b, loop_a))
3379 /* Example: when there are two consecutive loops,
3381 | loop_1
3382 | A[{0, +, 1}_1]
3383 | endloop_1
3384 | loop_2
3385 | A[{0, +, 1}_2]
3386 | endloop_2
3388 the dependence relation cannot be captured by the
3389 distance abstraction. */
3390 non_affine_dependence_relation (ddr);
3391 return true;
3394 /* The dependence is carried by the outermost loop. Example:
3395 | loop_1
3396 | A[{4, +, 1}_1]
3397 | loop_2
3398 | A[{5, +, 1}_2]
3399 | endloop_2
3400 | endloop_1
3401 In this case, the dependence is carried by loop_1. */
3402 loop_nb = loop_nb_a < loop_nb_b ? loop_nb_a : loop_nb_b;
3403 loop_depth = current_loops->parray[loop_nb]->depth - first_loop_depth;
3405 /* If the loop number is still greater than the number of
3406 loops we've been asked to analyze, or negative,
3407 something is borked. */
3408 gcc_assert (loop_depth >= 0);
3409 gcc_assert (loop_depth < nb_loops);
3411 if (chrec_contains_undetermined (SUB_DISTANCE (subscript)))
3413 non_affine_dependence_relation (ddr);
3414 return true;
3417 dist = int_cst_value (SUB_DISTANCE (subscript));
3419 if (dist == 0)
3420 dir = dir_equal;
3421 else if (dist > 0)
3422 dir = dir_positive;
3423 else if (dist < 0)
3424 dir = dir_negative;
3426 /* This is the subscript coupling test.
3427 | loop i = 0, N, 1
3428 | T[i+1][i] = ...
3429 | ... = T[i][i]
3430 | endloop
3431 There is no dependence. */
3432 if (init_v[loop_depth] != 0
3433 && dir != dir_star
3434 && (enum data_dependence_direction) dir_v[loop_depth] != dir
3435 && (enum data_dependence_direction) dir_v[loop_depth] != dir_star)
3437 finalize_ddr_dependent (ddr, chrec_known);
3438 return true;
3441 dir_v[loop_depth] = dir;
3442 init_v[loop_depth] = 1;
3443 init_b = true;
3447 /* Save the direction vector if we initialized one. */
3448 if (init_b)
3450 lambda_vector save_v = lambda_vector_new (DDR_SIZE_VECT (ddr));
3452 lambda_vector_copy (dir_v, save_v, DDR_SIZE_VECT (ddr));
3453 VEC_safe_push (lambda_vector, heap, DDR_DIR_VECTS (ddr), save_v);
3456 /* There is a distance of 1 on all the outer loops:
3458 Example: there is a dependence of distance 1 on loop_1 for the array A.
3459 | loop_1
3460 | A[5] = ...
3461 | endloop
3464 struct loop *lca, *loop_a, *loop_b;
3465 struct data_reference *a = DDR_A (ddr);
3466 struct data_reference *b = DDR_B (ddr);
3467 int lca_depth;
3468 loop_a = loop_containing_stmt (DR_STMT (a));
3469 loop_b = loop_containing_stmt (DR_STMT (b));
3471 /* Get the common ancestor loop. */
3472 lca = find_common_loop (loop_a, loop_b);
3473 lca_depth = lca->depth - first_loop_depth;
3475 gcc_assert (lca_depth >= 0);
3476 gcc_assert (lca_depth < nb_loops);
3478 while (lca->depth != 0)
3480 /* If we're considering just a sub-nest, then don't record
3481 any information on the outer loops. */
3482 if (lca_depth < 0)
3483 break;
3485 gcc_assert (lca_depth < nb_loops);
3487 if (init_v[lca_depth] == 0)
3489 lambda_vector save_v = lambda_vector_new (DDR_SIZE_VECT (ddr));
3491 lambda_vector_copy (dir_v, save_v, DDR_SIZE_VECT (ddr));
3492 save_v[lca_depth] = dir_positive;
3493 VEC_safe_push (lambda_vector, heap, DDR_DIR_VECTS (ddr), save_v);
3496 lca = lca->outer;
3497 lca_depth = lca->depth - first_loop_depth;
3502 return true;
3505 /* Returns true when all the access functions of A are affine or
3506 constant. */
3508 static bool
3509 access_functions_are_affine_or_constant_p (struct data_reference *a)
3511 unsigned int i;
3512 VEC(tree,heap) **fns = DR_ACCESS_FNS_ADDR (a);
3513 tree t;
3515 for (i = 0; VEC_iterate (tree, *fns, i, t); i++)
3516 if (!evolution_function_is_constant_p (t)
3517 && !evolution_function_is_affine_multivariate_p (t))
3518 return false;
3520 return true;
3523 /* This computes the affine dependence relation between A and B.
3524 CHREC_KNOWN is used for representing the independence between two
3525 accesses, while CHREC_DONT_KNOW is used for representing the unknown
3526 relation.
3528 Note that it is possible to stop the computation of the dependence
3529 relation the first time we detect a CHREC_KNOWN element for a given
3530 subscript. */
3532 void
3533 compute_affine_dependence (struct data_dependence_relation *ddr)
3535 struct data_reference *dra = DDR_A (ddr);
3536 struct data_reference *drb = DDR_B (ddr);
3538 if (dump_file && (dump_flags & TDF_DETAILS))
3540 fprintf (dump_file, "(compute_affine_dependence\n");
3541 fprintf (dump_file, " (stmt_a = \n");
3542 print_generic_expr (dump_file, DR_STMT (dra), 0);
3543 fprintf (dump_file, ")\n (stmt_b = \n");
3544 print_generic_expr (dump_file, DR_STMT (drb), 0);
3545 fprintf (dump_file, ")\n");
3548 /* Analyze only when the dependence relation is not yet known. */
3549 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
3551 if (access_functions_are_affine_or_constant_p (dra)
3552 && access_functions_are_affine_or_constant_p (drb))
3553 subscript_dependence_tester (ddr);
3555 /* As a last case, if the dependence cannot be determined, or if
3556 the dependence is considered too difficult to determine, answer
3557 "don't know". */
3558 else
3559 finalize_ddr_dependent (ddr, chrec_dont_know);
3562 if (dump_file && (dump_flags & TDF_DETAILS))
3563 fprintf (dump_file, ")\n");
3566 /* This computes the dependence relation for the same data
3567 reference into DDR. */
3569 static void
3570 compute_self_dependence (struct data_dependence_relation *ddr)
3572 unsigned int i;
3574 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
3576 struct subscript *subscript = DDR_SUBSCRIPT (ddr, i);
3578 /* The accessed index overlaps for each iteration. */
3579 SUB_CONFLICTS_IN_A (subscript) = integer_zero_node;
3580 SUB_CONFLICTS_IN_B (subscript) = integer_zero_node;
3581 SUB_LAST_CONFLICT (subscript) = chrec_dont_know;
3586 typedef struct data_dependence_relation *ddr_p;
3587 DEF_VEC_P(ddr_p);
3588 DEF_VEC_ALLOC_P(ddr_p,heap);
3590 /* Compute a subset of the data dependence relation graph. Don't
3591 compute read-read and self relations if
3592 COMPUTE_SELF_AND_READ_READ_DEPENDENCES is FALSE, and avoid the computation
3593 of the opposite relation, i.e. when AB has been computed, don't compute BA.
3594 DATAREFS contains a list of data references, and the result is set
3595 in DEPENDENCE_RELATIONS. */
3597 static void
3598 compute_all_dependences (varray_type datarefs,
3599 bool compute_self_and_read_read_dependences,
3600 VEC(ddr_p,heap) **dependence_relations)
3602 unsigned int i, j, N;
3604 N = VARRAY_ACTIVE_SIZE (datarefs);
3606 /* Note that we specifically skip i == j because it's a self dependence, and
3607 use compute_self_dependence below. */
3609 for (i = 0; i < N; i++)
3610 for (j = i + 1; j < N; j++)
3612 struct data_reference *a, *b;
3613 struct data_dependence_relation *ddr;
3615 a = VARRAY_GENERIC_PTR (datarefs, i);
3616 b = VARRAY_GENERIC_PTR (datarefs, j);
3617 if (DR_IS_READ (a) && DR_IS_READ (b)
3618 && !compute_self_and_read_read_dependences)
3619 continue;
3620 ddr = initialize_data_dependence_relation (a, b);
3622 VEC_safe_push (ddr_p, heap, *dependence_relations, ddr);
3623 compute_affine_dependence (ddr);
3624 compute_subscript_distance (ddr);
3626 if (!compute_self_and_read_read_dependences)
3627 return;
3629 /* Compute self dependence relation of each dataref to itself. */
3631 for (i = 0; i < N; i++)
3633 struct data_reference *a, *b;
3634 struct data_dependence_relation *ddr;
3636 a = VARRAY_GENERIC_PTR (datarefs, i);
3637 b = VARRAY_GENERIC_PTR (datarefs, i);
3638 ddr = initialize_data_dependence_relation (a, b);
3640 VEC_safe_push (ddr_p, heap, *dependence_relations, ddr);
3641 compute_self_dependence (ddr);
3642 compute_subscript_distance (ddr);
3646 /* Search the data references in LOOP, and record the information into
3647 DATAREFS. Returns chrec_dont_know when failing to analyze a
3648 difficult case, returns NULL_TREE otherwise.
3650 TODO: This function should be made smarter so that it can handle address
3651 arithmetic as if they were array accesses, etc. */
3653 tree
3654 find_data_references_in_loop (struct loop *loop, varray_type *datarefs)
3656 basic_block bb, *bbs;
3657 unsigned int i;
3658 block_stmt_iterator bsi;
3659 struct data_reference *dr;
3661 bbs = get_loop_body (loop);
3663 for (i = 0; i < loop->num_nodes; i++)
3665 bb = bbs[i];
3667 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
3669 tree stmt = bsi_stmt (bsi);
3671 /* ASM_EXPR and CALL_EXPR may embed arbitrary side effects.
3672 Calls have side-effects, except those to const or pure
3673 functions. */
3674 if ((TREE_CODE (stmt) == CALL_EXPR
3675 && !(call_expr_flags (stmt) & (ECF_CONST | ECF_PURE)))
3676 || (TREE_CODE (stmt) == ASM_EXPR
3677 && ASM_VOLATILE_P (stmt)))
3678 goto insert_dont_know_node;
3680 if (ZERO_SSA_OPERANDS (stmt, SSA_OP_ALL_VIRTUALS))
3681 continue;
3683 switch (TREE_CODE (stmt))
3685 case MODIFY_EXPR:
3687 bool one_inserted = false;
3688 tree opnd0 = TREE_OPERAND (stmt, 0);
3689 tree opnd1 = TREE_OPERAND (stmt, 1);
3691 if (TREE_CODE (opnd0) == ARRAY_REF
3692 || TREE_CODE (opnd0) == INDIRECT_REF)
3694 dr = create_data_ref (opnd0, stmt, false);
3695 if (dr)
3697 VARRAY_PUSH_GENERIC_PTR (*datarefs, dr);
3698 one_inserted = true;
3702 if (TREE_CODE (opnd1) == ARRAY_REF
3703 || TREE_CODE (opnd1) == INDIRECT_REF)
3705 dr = create_data_ref (opnd1, stmt, true);
3706 if (dr)
3708 VARRAY_PUSH_GENERIC_PTR (*datarefs, dr);
3709 one_inserted = true;
3713 if (!one_inserted)
3714 goto insert_dont_know_node;
3716 break;
3719 case CALL_EXPR:
3721 tree args;
3722 bool one_inserted = false;
3724 for (args = TREE_OPERAND (stmt, 1); args;
3725 args = TREE_CHAIN (args))
3726 if (TREE_CODE (TREE_VALUE (args)) == ARRAY_REF
3727 || TREE_CODE (TREE_VALUE (args)) == INDIRECT_REF)
3729 dr = create_data_ref (TREE_VALUE (args), stmt, true);
3730 if (dr)
3732 VARRAY_PUSH_GENERIC_PTR (*datarefs, dr);
3733 one_inserted = true;
3737 if (!one_inserted)
3738 goto insert_dont_know_node;
3740 break;
3743 default:
3745 struct data_reference *res;
3747 insert_dont_know_node:;
3748 res = XNEW (struct data_reference);
3749 DR_STMT (res) = NULL_TREE;
3750 DR_REF (res) = NULL_TREE;
3751 DR_BASE_OBJECT (res) = NULL;
3752 DR_TYPE (res) = ARRAY_REF_TYPE;
3753 DR_SET_ACCESS_FNS (res, NULL);
3754 DR_BASE_OBJECT (res) = NULL;
3755 DR_IS_READ (res) = false;
3756 DR_BASE_ADDRESS (res) = NULL_TREE;
3757 DR_OFFSET (res) = NULL_TREE;
3758 DR_INIT (res) = NULL_TREE;
3759 DR_STEP (res) = NULL_TREE;
3760 DR_OFFSET_MISALIGNMENT (res) = NULL_TREE;
3761 DR_MEMTAG (res) = NULL_TREE;
3762 DR_PTR_INFO (res) = NULL;
3763 VARRAY_PUSH_GENERIC_PTR (*datarefs, res);
3765 free (bbs);
3766 return chrec_dont_know;
3770 /* When there are no defs in the loop, the loop is parallel. */
3771 if (!ZERO_SSA_OPERANDS (stmt, SSA_OP_VIRTUAL_DEFS))
3772 loop->parallel_p = false;
3776 free (bbs);
3778 return NULL_TREE;
3783 /* This section contains all the entry points. */
3785 /* Given a loop nest LOOP, the following vectors are returned:
3786 *DATAREFS is initialized to all the array elements contained in this loop,
3787 *DEPENDENCE_RELATIONS contains the relations between the data references.
3788 Compute read-read and self relations if
3789 COMPUTE_SELF_AND_READ_READ_DEPENDENCES is TRUE. */
3791 void
3792 compute_data_dependences_for_loop (struct loop *loop,
3793 bool compute_self_and_read_read_dependences,
3794 varray_type *datarefs,
3795 varray_type *dependence_relations)
3797 unsigned int i, nb_loops;
3798 VEC(ddr_p,heap) *allrelations;
3799 struct data_dependence_relation *ddr;
3800 struct loop *loop_nest = loop;
3802 while (loop_nest && loop_nest->outer && loop_nest->outer->outer)
3803 loop_nest = loop_nest->outer;
3805 nb_loops = loop_nest->level;
3807 /* If one of the data references is not computable, give up without
3808 spending time to compute other dependences. */
3809 if (find_data_references_in_loop (loop, datarefs) == chrec_dont_know)
3811 struct data_dependence_relation *ddr;
3813 /* Insert a single relation into dependence_relations:
3814 chrec_dont_know. */
3815 ddr = initialize_data_dependence_relation (NULL, NULL);
3816 VARRAY_PUSH_GENERIC_PTR (*dependence_relations, ddr);
3817 build_classic_dist_vector (ddr, nb_loops, loop->depth);
3818 build_classic_dir_vector (ddr, nb_loops, loop->depth);
3819 return;
3822 allrelations = NULL;
3823 compute_all_dependences (*datarefs, compute_self_and_read_read_dependences,
3824 &allrelations);
3826 for (i = 0; VEC_iterate (ddr_p, allrelations, i, ddr); i++)
3828 if (build_classic_dist_vector (ddr, nb_loops, loop_nest->depth))
3830 VARRAY_PUSH_GENERIC_PTR (*dependence_relations, ddr);
3831 build_classic_dir_vector (ddr, nb_loops, loop_nest->depth);
3836 /* Entry point (for testing only). Analyze all the data references
3837 and the dependence relations.
3839 The data references are computed first.
3841 A relation on these nodes is represented by a complete graph. Some
3842 of the relations could be of no interest, thus the relations can be
3843 computed on demand.
3845 In the following function we compute all the relations. This is
3846 just a first implementation that is here for:
3847 - for showing how to ask for the dependence relations,
3848 - for the debugging the whole dependence graph,
3849 - for the dejagnu testcases and maintenance.
3851 It is possible to ask only for a part of the graph, avoiding to
3852 compute the whole dependence graph. The computed dependences are
3853 stored in a knowledge base (KB) such that later queries don't
3854 recompute the same information. The implementation of this KB is
3855 transparent to the optimizer, and thus the KB can be changed with a
3856 more efficient implementation, or the KB could be disabled. */
3858 void
3859 analyze_all_data_dependences (struct loops *loops)
3861 unsigned int i;
3862 varray_type datarefs;
3863 varray_type dependence_relations;
3864 int nb_data_refs = 10;
3866 VARRAY_GENERIC_PTR_INIT (datarefs, nb_data_refs, "datarefs");
3867 VARRAY_GENERIC_PTR_INIT (dependence_relations,
3868 nb_data_refs * nb_data_refs,
3869 "dependence_relations");
3871 /* Compute DDs on the whole function. */
3872 compute_data_dependences_for_loop (loops->parray[0], false,
3873 &datarefs, &dependence_relations);
3875 if (dump_file)
3877 dump_data_dependence_relations (dump_file, dependence_relations);
3878 fprintf (dump_file, "\n\n");
3880 if (dump_flags & TDF_DETAILS)
3881 dump_dist_dir_vectors (dump_file, dependence_relations);
3883 if (dump_flags & TDF_STATS)
3885 unsigned nb_top_relations = 0;
3886 unsigned nb_bot_relations = 0;
3887 unsigned nb_basename_differ = 0;
3888 unsigned nb_chrec_relations = 0;
3890 for (i = 0; i < VARRAY_ACTIVE_SIZE (dependence_relations); i++)
3892 struct data_dependence_relation *ddr;
3893 ddr = VARRAY_GENERIC_PTR (dependence_relations, i);
3895 if (chrec_contains_undetermined (DDR_ARE_DEPENDENT (ddr)))
3896 nb_top_relations++;
3898 else if (DDR_ARE_DEPENDENT (ddr) == chrec_known)
3900 struct data_reference *a = DDR_A (ddr);
3901 struct data_reference *b = DDR_B (ddr);
3902 bool differ_p;
3904 if ((DR_BASE_OBJECT (a) && DR_BASE_OBJECT (b)
3905 && DR_NUM_DIMENSIONS (a) != DR_NUM_DIMENSIONS (b))
3906 || (base_object_differ_p (a, b, &differ_p)
3907 && differ_p))
3908 nb_basename_differ++;
3909 else
3910 nb_bot_relations++;
3913 else
3914 nb_chrec_relations++;
3917 gather_stats_on_scev_database ();
3921 free_dependence_relations (dependence_relations);
3922 free_data_refs (datarefs);
3925 /* Free the memory used by a data dependence relation DDR. */
3927 void
3928 free_dependence_relation (struct data_dependence_relation *ddr)
3930 if (ddr == NULL)
3931 return;
3933 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE && DDR_SUBSCRIPTS (ddr))
3934 varray_clear (DDR_SUBSCRIPTS (ddr));
3935 free (ddr);
3938 /* Free the memory used by the data dependence relations from
3939 DEPENDENCE_RELATIONS. */
3941 void
3942 free_dependence_relations (varray_type dependence_relations)
3944 unsigned int i;
3945 if (dependence_relations == NULL)
3946 return;
3948 for (i = 0; i < VARRAY_ACTIVE_SIZE (dependence_relations); i++)
3949 free_dependence_relation (VARRAY_GENERIC_PTR (dependence_relations, i));
3950 varray_clear (dependence_relations);
3953 /* Free the memory used by the data references from DATAREFS. */
3955 void
3956 free_data_refs (varray_type datarefs)
3958 unsigned int i;
3960 if (datarefs == NULL)
3961 return;
3963 for (i = 0; i < VARRAY_ACTIVE_SIZE (datarefs); i++)
3965 struct data_reference *dr = (struct data_reference *)
3966 VARRAY_GENERIC_PTR (datarefs, i);
3967 if (dr)
3969 DR_FREE_ACCESS_FNS (dr);
3970 free (dr);
3973 varray_clear (datarefs);