scop.c: fix typos in comments
[pet.git] / scan.cc
blob202e57cc5aaf5600c1330467d53e2dbce8983282
1 /*
2 * Copyright 2011 Leiden University. All rights reserved.
3 * Copyright 2012 Ecole Normale Superieure. All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 *
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above
13 * copyright notice, this list of conditions and the following
14 * disclaimer in the documentation and/or other materials provided
15 * with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY LEIDEN UNIVERSITY ''AS IS'' AND ANY
18 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL LEIDEN UNIVERSITY OR
21 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
22 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
23 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA,
24 * OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
27 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 * The views and conclusions contained in the software and documentation
30 * are those of the authors and should not be interpreted as
31 * representing official policies, either expressed or implied, of
32 * Leiden University.
33 */
35 #include <set>
36 #include <map>
37 #include <iostream>
38 #include <llvm/Support/raw_ostream.h>
39 #include <clang/AST/ASTContext.h>
40 #include <clang/AST/ASTDiagnostic.h>
41 #include <clang/AST/Expr.h>
42 #include <clang/AST/RecursiveASTVisitor.h>
44 #include <isl/id.h>
45 #include <isl/space.h>
46 #include <isl/aff.h>
47 #include <isl/set.h>
49 #include "options.h"
50 #include "scan.h"
51 #include "scop.h"
52 #include "scop_plus.h"
54 #include "config.h"
56 using namespace std;
57 using namespace clang;
59 #if defined(DECLREFEXPR_CREATE_REQUIRES_BOOL)
60 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
62 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
63 SourceLocation(), var, false, var->getInnerLocStart(),
64 var->getType(), VK_LValue);
66 #elif defined(DECLREFEXPR_CREATE_REQUIRES_SOURCELOCATION)
67 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
69 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
70 SourceLocation(), var, var->getInnerLocStart(), var->getType(),
71 VK_LValue);
73 #else
74 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
76 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
77 var, var->getInnerLocStart(), var->getType(), VK_LValue);
79 #endif
81 /* Check if the element type corresponding to the given array type
82 * has a const qualifier.
84 static bool const_base(QualType qt)
86 const Type *type = qt.getTypePtr();
88 if (type->isPointerType())
89 return const_base(type->getPointeeType());
90 if (type->isArrayType()) {
91 const ArrayType *atype;
92 type = type->getCanonicalTypeInternal().getTypePtr();
93 atype = cast<ArrayType>(type);
94 return const_base(atype->getElementType());
97 return qt.isConstQualified();
100 /* Mark "decl" as having an unknown value in "assigned_value".
102 * If no (known or unknown) value was assigned to "decl" before,
103 * then it may have been treated as a parameter before and may
104 * therefore appear in a value assigned to another variable.
105 * If so, this assignment needs to be turned into an unknown value too.
107 static void clear_assignment(map<ValueDecl *, isl_pw_aff *> &assigned_value,
108 ValueDecl *decl)
110 map<ValueDecl *, isl_pw_aff *>::iterator it;
112 it = assigned_value.find(decl);
114 assigned_value[decl] = NULL;
116 if (it == assigned_value.end())
117 return;
119 for (it = assigned_value.begin(); it != assigned_value.end(); ++it) {
120 isl_pw_aff *pa = it->second;
121 int nparam = isl_pw_aff_dim(pa, isl_dim_param);
123 for (int i = 0; i < nparam; ++i) {
124 isl_id *id;
126 if (!isl_pw_aff_has_dim_id(pa, isl_dim_param, i))
127 continue;
128 id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
129 if (isl_id_get_user(id) == decl)
130 it->second = NULL;
131 isl_id_free(id);
136 /* Look for any assignments to scalar variables in part of the parse
137 * tree and set assigned_value to NULL for each of them.
138 * Also reset assigned_value if the address of a scalar variable
139 * is being taken. As an exception, if the address is passed to a function
140 * that is declared to receive a const pointer, then assigned_value is
141 * not reset.
143 * This ensures that we won't use any previously stored value
144 * in the current subtree and its parents.
146 struct clear_assignments : RecursiveASTVisitor<clear_assignments> {
147 map<ValueDecl *, isl_pw_aff *> &assigned_value;
148 set<UnaryOperator *> skip;
150 clear_assignments(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
151 assigned_value(assigned_value) {}
153 /* Check for "address of" operators whose value is passed
154 * to a const pointer argument and add them to "skip", so that
155 * we can skip them in VisitUnaryOperator.
157 bool VisitCallExpr(CallExpr *expr) {
158 FunctionDecl *fd;
159 fd = expr->getDirectCallee();
160 if (!fd)
161 return true;
162 for (int i = 0; i < expr->getNumArgs(); ++i) {
163 Expr *arg = expr->getArg(i);
164 UnaryOperator *op;
165 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
166 ImplicitCastExpr *ice;
167 ice = cast<ImplicitCastExpr>(arg);
168 arg = ice->getSubExpr();
170 if (arg->getStmtClass() != Stmt::UnaryOperatorClass)
171 continue;
172 op = cast<UnaryOperator>(arg);
173 if (op->getOpcode() != UO_AddrOf)
174 continue;
175 if (const_base(fd->getParamDecl(i)->getType()))
176 skip.insert(op);
178 return true;
181 bool VisitUnaryOperator(UnaryOperator *expr) {
182 Expr *arg;
183 DeclRefExpr *ref;
184 ValueDecl *decl;
186 switch (expr->getOpcode()) {
187 case UO_AddrOf:
188 case UO_PostInc:
189 case UO_PostDec:
190 case UO_PreInc:
191 case UO_PreDec:
192 break;
193 default:
194 return true;
196 if (skip.find(expr) != skip.end())
197 return true;
199 arg = expr->getSubExpr();
200 if (arg->getStmtClass() != Stmt::DeclRefExprClass)
201 return true;
202 ref = cast<DeclRefExpr>(arg);
203 decl = ref->getDecl();
204 clear_assignment(assigned_value, decl);
205 return true;
208 bool VisitBinaryOperator(BinaryOperator *expr) {
209 Expr *lhs;
210 DeclRefExpr *ref;
211 ValueDecl *decl;
213 if (!expr->isAssignmentOp())
214 return true;
215 lhs = expr->getLHS();
216 if (lhs->getStmtClass() != Stmt::DeclRefExprClass)
217 return true;
218 ref = cast<DeclRefExpr>(lhs);
219 decl = ref->getDecl();
220 clear_assignment(assigned_value, decl);
221 return true;
225 /* Keep a copy of the currently assigned values.
227 * Any variable that is assigned a value inside the current scope
228 * is removed again when we leave the scope (either because it wasn't
229 * stored in the cache or because it has a different value in the cache).
231 struct assigned_value_cache {
232 map<ValueDecl *, isl_pw_aff *> &assigned_value;
233 map<ValueDecl *, isl_pw_aff *> cache;
235 assigned_value_cache(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
236 assigned_value(assigned_value), cache(assigned_value) {}
237 ~assigned_value_cache() {
238 map<ValueDecl *, isl_pw_aff *>::iterator it = cache.begin();
239 for (it = assigned_value.begin(); it != assigned_value.end();
240 ++it) {
241 if (!it->second ||
242 (cache.find(it->first) != cache.end() &&
243 cache[it->first] != it->second))
244 cache[it->first] = NULL;
246 assigned_value = cache;
250 /* Insert an expression into the collection of expressions,
251 * provided it is not already in there.
252 * The isl_pw_affs are freed in the destructor.
254 void PetScan::insert_expression(__isl_take isl_pw_aff *expr)
256 std::set<isl_pw_aff *>::iterator it;
258 if (expressions.find(expr) == expressions.end())
259 expressions.insert(expr);
260 else
261 isl_pw_aff_free(expr);
264 PetScan::~PetScan()
266 std::set<isl_pw_aff *>::iterator it;
268 for (it = expressions.begin(); it != expressions.end(); ++it)
269 isl_pw_aff_free(*it);
271 isl_union_map_free(value_bounds);
274 /* Called if we found something we (currently) cannot handle.
275 * We'll provide more informative warnings later.
277 * We only actually complain if autodetect is false.
279 void PetScan::unsupported(Stmt *stmt, const char *msg)
281 if (options->autodetect)
282 return;
284 SourceLocation loc = stmt->getLocStart();
285 DiagnosticsEngine &diag = PP.getDiagnostics();
286 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
287 msg ? msg : "unsupported");
288 DiagnosticBuilder B = diag.Report(loc, id) << stmt->getSourceRange();
291 /* Extract an integer from "expr" and store it in "v".
293 int PetScan::extract_int(IntegerLiteral *expr, isl_int *v)
295 const Type *type = expr->getType().getTypePtr();
296 int is_signed = type->hasSignedIntegerRepresentation();
298 if (is_signed) {
299 int64_t i = expr->getValue().getSExtValue();
300 isl_int_set_si(*v, i);
301 } else {
302 uint64_t i = expr->getValue().getZExtValue();
303 isl_int_set_ui(*v, i);
306 return 0;
309 /* Extract an integer from "expr" and store it in "v".
310 * Return -1 if "expr" does not (obviously) represent an integer.
312 int PetScan::extract_int(clang::ParenExpr *expr, isl_int *v)
314 return extract_int(expr->getSubExpr(), v);
317 /* Extract an integer from "expr" and store it in "v".
318 * Return -1 if "expr" does not (obviously) represent an integer.
320 int PetScan::extract_int(clang::Expr *expr, isl_int *v)
322 if (expr->getStmtClass() == Stmt::IntegerLiteralClass)
323 return extract_int(cast<IntegerLiteral>(expr), v);
324 if (expr->getStmtClass() == Stmt::ParenExprClass)
325 return extract_int(cast<ParenExpr>(expr), v);
327 unsupported(expr);
328 return -1;
331 /* Extract an affine expression from the IntegerLiteral "expr".
333 __isl_give isl_pw_aff *PetScan::extract_affine(IntegerLiteral *expr)
335 isl_space *dim = isl_space_params_alloc(ctx, 0);
336 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
337 isl_aff *aff = isl_aff_zero_on_domain(ls);
338 isl_set *dom = isl_set_universe(dim);
339 isl_int v;
341 isl_int_init(v);
342 extract_int(expr, &v);
343 aff = isl_aff_add_constant(aff, v);
344 isl_int_clear(v);
346 return isl_pw_aff_alloc(dom, aff);
349 /* Extract an affine expression from the APInt "val".
351 __isl_give isl_pw_aff *PetScan::extract_affine(const llvm::APInt &val)
353 isl_space *dim = isl_space_params_alloc(ctx, 0);
354 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
355 isl_aff *aff = isl_aff_zero_on_domain(ls);
356 isl_set *dom = isl_set_universe(dim);
357 isl_int v;
359 isl_int_init(v);
360 isl_int_set_ui(v, val.getZExtValue());
361 aff = isl_aff_add_constant(aff, v);
362 isl_int_clear(v);
364 return isl_pw_aff_alloc(dom, aff);
367 __isl_give isl_pw_aff *PetScan::extract_affine(ImplicitCastExpr *expr)
369 return extract_affine(expr->getSubExpr());
372 static unsigned get_type_size(ValueDecl *decl)
374 return decl->getASTContext().getIntWidth(decl->getType());
377 /* Bound parameter "pos" of "set" to the possible values of "decl".
379 static __isl_give isl_set *set_parameter_bounds(__isl_take isl_set *set,
380 unsigned pos, ValueDecl *decl)
382 unsigned width;
383 isl_int v;
385 isl_int_init(v);
387 width = get_type_size(decl);
388 if (decl->getType()->isUnsignedIntegerType()) {
389 set = isl_set_lower_bound_si(set, isl_dim_param, pos, 0);
390 isl_int_set_si(v, 1);
391 isl_int_mul_2exp(v, v, width);
392 isl_int_sub_ui(v, v, 1);
393 set = isl_set_upper_bound(set, isl_dim_param, pos, v);
394 } else {
395 isl_int_set_si(v, 1);
396 isl_int_mul_2exp(v, v, width - 1);
397 isl_int_sub_ui(v, v, 1);
398 set = isl_set_upper_bound(set, isl_dim_param, pos, v);
399 isl_int_neg(v, v);
400 isl_int_sub_ui(v, v, 1);
401 set = isl_set_lower_bound(set, isl_dim_param, pos, v);
404 isl_int_clear(v);
406 return set;
409 /* Extract an affine expression from the DeclRefExpr "expr".
411 * If the variable has been assigned a value, then we check whether
412 * we know what (affine) value was assigned.
413 * If so, we return this value. Otherwise we convert "expr"
414 * to an extra parameter (provided nesting_enabled is set).
416 * Otherwise, we simply return an expression that is equal
417 * to a parameter corresponding to the referenced variable.
419 __isl_give isl_pw_aff *PetScan::extract_affine(DeclRefExpr *expr)
421 ValueDecl *decl = expr->getDecl();
422 const Type *type = decl->getType().getTypePtr();
423 isl_id *id;
424 isl_space *dim;
425 isl_aff *aff;
426 isl_set *dom;
428 if (!type->isIntegerType()) {
429 unsupported(expr);
430 return NULL;
433 if (assigned_value.find(decl) != assigned_value.end()) {
434 if (assigned_value[decl])
435 return isl_pw_aff_copy(assigned_value[decl]);
436 else
437 return nested_access(expr);
440 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
441 dim = isl_space_params_alloc(ctx, 1);
443 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
445 dom = isl_set_universe(isl_space_copy(dim));
446 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
447 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
449 return isl_pw_aff_alloc(dom, aff);
452 /* Extract an affine expression from an integer division operation.
453 * In particular, if "expr" is lhs/rhs, then return
455 * lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs)
457 * The second argument (rhs) is required to be a (positive) integer constant.
459 __isl_give isl_pw_aff *PetScan::extract_affine_div(BinaryOperator *expr)
461 int is_cst;
462 isl_pw_aff *rhs, *lhs;
464 rhs = extract_affine(expr->getRHS());
465 is_cst = isl_pw_aff_is_cst(rhs);
466 if (is_cst < 0 || !is_cst) {
467 isl_pw_aff_free(rhs);
468 if (!is_cst)
469 unsupported(expr);
470 return NULL;
473 lhs = extract_affine(expr->getLHS());
475 return isl_pw_aff_tdiv_q(lhs, rhs);
478 /* Extract an affine expression from a modulo operation.
479 * In particular, if "expr" is lhs/rhs, then return
481 * lhs - rhs * (lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs))
483 * The second argument (rhs) is required to be a (positive) integer constant.
485 __isl_give isl_pw_aff *PetScan::extract_affine_mod(BinaryOperator *expr)
487 int is_cst;
488 isl_pw_aff *rhs, *lhs;
490 rhs = extract_affine(expr->getRHS());
491 is_cst = isl_pw_aff_is_cst(rhs);
492 if (is_cst < 0 || !is_cst) {
493 isl_pw_aff_free(rhs);
494 if (!is_cst)
495 unsupported(expr);
496 return NULL;
499 lhs = extract_affine(expr->getLHS());
501 return isl_pw_aff_tdiv_r(lhs, rhs);
504 /* Extract an affine expression from a multiplication operation.
505 * This is only allowed if at least one of the two arguments
506 * is a (piecewise) constant.
508 __isl_give isl_pw_aff *PetScan::extract_affine_mul(BinaryOperator *expr)
510 isl_pw_aff *lhs;
511 isl_pw_aff *rhs;
513 lhs = extract_affine(expr->getLHS());
514 rhs = extract_affine(expr->getRHS());
516 if (!isl_pw_aff_is_cst(lhs) && !isl_pw_aff_is_cst(rhs)) {
517 isl_pw_aff_free(lhs);
518 isl_pw_aff_free(rhs);
519 unsupported(expr);
520 return NULL;
523 return isl_pw_aff_mul(lhs, rhs);
526 /* Extract an affine expression from an addition or subtraction operation.
528 __isl_give isl_pw_aff *PetScan::extract_affine_add(BinaryOperator *expr)
530 isl_pw_aff *lhs;
531 isl_pw_aff *rhs;
533 lhs = extract_affine(expr->getLHS());
534 rhs = extract_affine(expr->getRHS());
536 switch (expr->getOpcode()) {
537 case BO_Add:
538 return isl_pw_aff_add(lhs, rhs);
539 case BO_Sub:
540 return isl_pw_aff_sub(lhs, rhs);
541 default:
542 isl_pw_aff_free(lhs);
543 isl_pw_aff_free(rhs);
544 return NULL;
549 /* Compute
551 * pwaff mod 2^width
553 static __isl_give isl_pw_aff *wrap(__isl_take isl_pw_aff *pwaff,
554 unsigned width)
556 isl_int mod;
558 isl_int_init(mod);
559 isl_int_set_si(mod, 1);
560 isl_int_mul_2exp(mod, mod, width);
562 pwaff = isl_pw_aff_mod(pwaff, mod);
564 isl_int_clear(mod);
566 return pwaff;
569 /* Limit the domain of "pwaff" to those elements where the function
570 * value satisfies
572 * 2^{width-1} <= pwaff < 2^{width-1}
574 static __isl_give isl_pw_aff *avoid_overflow(__isl_take isl_pw_aff *pwaff,
575 unsigned width)
577 isl_int v;
578 isl_space *space = isl_pw_aff_get_domain_space(pwaff);
579 isl_local_space *ls = isl_local_space_from_space(space);
580 isl_aff *bound;
581 isl_set *dom;
582 isl_pw_aff *b;
584 isl_int_init(v);
585 isl_int_set_si(v, 1);
586 isl_int_mul_2exp(v, v, width - 1);
588 bound = isl_aff_zero_on_domain(ls);
589 bound = isl_aff_add_constant(bound, v);
590 b = isl_pw_aff_from_aff(bound);
592 dom = isl_pw_aff_lt_set(isl_pw_aff_copy(pwaff), isl_pw_aff_copy(b));
593 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
595 b = isl_pw_aff_neg(b);
596 dom = isl_pw_aff_ge_set(isl_pw_aff_copy(pwaff), b);
597 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
599 isl_int_clear(v);
601 return pwaff;
604 /* Handle potential overflows on signed computations.
606 * If options->signed_overflow is set to PET_OVERFLOW_AVOID,
607 * the we adjust the domain of "pa" to avoid overflows.
609 __isl_give isl_pw_aff *PetScan::signed_overflow(__isl_take isl_pw_aff *pa,
610 unsigned width)
612 if (options->signed_overflow == PET_OVERFLOW_AVOID)
613 pa = avoid_overflow(pa, width);
615 return pa;
618 /* Return the piecewise affine expression "set ? 1 : 0" defined on "dom".
620 static __isl_give isl_pw_aff *indicator_function(__isl_take isl_set *set,
621 __isl_take isl_set *dom)
623 isl_pw_aff *pa;
624 pa = isl_set_indicator_function(set);
625 pa = isl_pw_aff_intersect_domain(pa, dom);
626 return pa;
629 /* Extract an affine expression from some binary operations.
630 * If the result of the expression is unsigned, then we wrap it
631 * based on the size of the type. Otherwise, we ensure that
632 * no overflow occurs.
634 __isl_give isl_pw_aff *PetScan::extract_affine(BinaryOperator *expr)
636 isl_pw_aff *res;
637 unsigned width;
639 switch (expr->getOpcode()) {
640 case BO_Add:
641 case BO_Sub:
642 res = extract_affine_add(expr);
643 break;
644 case BO_Div:
645 res = extract_affine_div(expr);
646 break;
647 case BO_Rem:
648 res = extract_affine_mod(expr);
649 break;
650 case BO_Mul:
651 res = extract_affine_mul(expr);
652 break;
653 case BO_LT:
654 case BO_LE:
655 case BO_GT:
656 case BO_GE:
657 case BO_EQ:
658 case BO_NE:
659 case BO_LAnd:
660 case BO_LOr:
661 return extract_condition(expr);
662 default:
663 unsupported(expr);
664 return NULL;
667 width = ast_context.getIntWidth(expr->getType());
668 if (expr->getType()->isUnsignedIntegerType())
669 res = wrap(res, width);
670 else
671 res = signed_overflow(res, width);
673 return res;
676 /* Extract an affine expression from a negation operation.
678 __isl_give isl_pw_aff *PetScan::extract_affine(UnaryOperator *expr)
680 if (expr->getOpcode() == UO_Minus)
681 return isl_pw_aff_neg(extract_affine(expr->getSubExpr()));
682 if (expr->getOpcode() == UO_LNot)
683 return extract_condition(expr);
685 unsupported(expr);
686 return NULL;
689 __isl_give isl_pw_aff *PetScan::extract_affine(ParenExpr *expr)
691 return extract_affine(expr->getSubExpr());
694 /* Extract an affine expression from some special function calls.
695 * In particular, we handle "min", "max", "ceild" and "floord".
696 * In case of the latter two, the second argument needs to be
697 * a (positive) integer constant.
699 __isl_give isl_pw_aff *PetScan::extract_affine(CallExpr *expr)
701 FunctionDecl *fd;
702 string name;
703 isl_pw_aff *aff1, *aff2;
705 fd = expr->getDirectCallee();
706 if (!fd) {
707 unsupported(expr);
708 return NULL;
711 name = fd->getDeclName().getAsString();
712 if (!(expr->getNumArgs() == 2 && name == "min") &&
713 !(expr->getNumArgs() == 2 && name == "max") &&
714 !(expr->getNumArgs() == 2 && name == "floord") &&
715 !(expr->getNumArgs() == 2 && name == "ceild")) {
716 unsupported(expr);
717 return NULL;
720 if (name == "min" || name == "max") {
721 aff1 = extract_affine(expr->getArg(0));
722 aff2 = extract_affine(expr->getArg(1));
724 if (name == "min")
725 aff1 = isl_pw_aff_min(aff1, aff2);
726 else
727 aff1 = isl_pw_aff_max(aff1, aff2);
728 } else if (name == "floord" || name == "ceild") {
729 isl_int v;
730 Expr *arg2 = expr->getArg(1);
732 if (arg2->getStmtClass() != Stmt::IntegerLiteralClass) {
733 unsupported(expr);
734 return NULL;
736 aff1 = extract_affine(expr->getArg(0));
737 isl_int_init(v);
738 extract_int(cast<IntegerLiteral>(arg2), &v);
739 aff1 = isl_pw_aff_scale_down(aff1, v);
740 isl_int_clear(v);
741 if (name == "floord")
742 aff1 = isl_pw_aff_floor(aff1);
743 else
744 aff1 = isl_pw_aff_ceil(aff1);
745 } else {
746 unsupported(expr);
747 return NULL;
750 return aff1;
753 /* This method is called when we come across an access that is
754 * nested in what is supposed to be an affine expression.
755 * If nesting is allowed, we return a new parameter that corresponds
756 * to this nested access. Otherwise, we simply complain.
758 * Note that we currently don't allow nested accesses themselves
759 * to contain any nested accesses, so we check if we can extract
760 * the access without any nesting and complain if we can't.
762 * The new parameter is resolved in resolve_nested.
764 isl_pw_aff *PetScan::nested_access(Expr *expr)
766 isl_id *id;
767 isl_space *dim;
768 isl_aff *aff;
769 isl_set *dom;
770 isl_map *access;
772 if (!nesting_enabled) {
773 unsupported(expr);
774 return NULL;
777 allow_nested = false;
778 access = extract_access(expr);
779 allow_nested = true;
780 if (!access) {
781 unsupported(expr);
782 return NULL;
784 isl_map_free(access);
786 id = isl_id_alloc(ctx, NULL, expr);
787 dim = isl_space_params_alloc(ctx, 1);
789 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
791 dom = isl_set_universe(isl_space_copy(dim));
792 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
793 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
795 return isl_pw_aff_alloc(dom, aff);
798 /* Affine expressions are not supposed to contain array accesses,
799 * but if nesting is allowed, we return a parameter corresponding
800 * to the array access.
802 __isl_give isl_pw_aff *PetScan::extract_affine(ArraySubscriptExpr *expr)
804 return nested_access(expr);
807 /* Extract an affine expression from a conditional operation.
809 __isl_give isl_pw_aff *PetScan::extract_affine(ConditionalOperator *expr)
811 isl_pw_aff *cond, *lhs, *rhs, *res;
813 cond = extract_condition(expr->getCond());
814 lhs = extract_affine(expr->getTrueExpr());
815 rhs = extract_affine(expr->getFalseExpr());
817 return isl_pw_aff_cond(cond, lhs, rhs);
820 /* Extract an affine expression, if possible, from "expr".
821 * Otherwise return NULL.
823 __isl_give isl_pw_aff *PetScan::extract_affine(Expr *expr)
825 switch (expr->getStmtClass()) {
826 case Stmt::ImplicitCastExprClass:
827 return extract_affine(cast<ImplicitCastExpr>(expr));
828 case Stmt::IntegerLiteralClass:
829 return extract_affine(cast<IntegerLiteral>(expr));
830 case Stmt::DeclRefExprClass:
831 return extract_affine(cast<DeclRefExpr>(expr));
832 case Stmt::BinaryOperatorClass:
833 return extract_affine(cast<BinaryOperator>(expr));
834 case Stmt::UnaryOperatorClass:
835 return extract_affine(cast<UnaryOperator>(expr));
836 case Stmt::ParenExprClass:
837 return extract_affine(cast<ParenExpr>(expr));
838 case Stmt::CallExprClass:
839 return extract_affine(cast<CallExpr>(expr));
840 case Stmt::ArraySubscriptExprClass:
841 return extract_affine(cast<ArraySubscriptExpr>(expr));
842 case Stmt::ConditionalOperatorClass:
843 return extract_affine(cast<ConditionalOperator>(expr));
844 default:
845 unsupported(expr);
847 return NULL;
850 __isl_give isl_map *PetScan::extract_access(ImplicitCastExpr *expr)
852 return extract_access(expr->getSubExpr());
855 /* Return the depth of an array of the given type.
857 static int array_depth(const Type *type)
859 if (type->isPointerType())
860 return 1 + array_depth(type->getPointeeType().getTypePtr());
861 if (type->isArrayType()) {
862 const ArrayType *atype;
863 type = type->getCanonicalTypeInternal().getTypePtr();
864 atype = cast<ArrayType>(type);
865 return 1 + array_depth(atype->getElementType().getTypePtr());
867 return 0;
870 /* Return the element type of the given array type.
872 static QualType base_type(QualType qt)
874 const Type *type = qt.getTypePtr();
876 if (type->isPointerType())
877 return base_type(type->getPointeeType());
878 if (type->isArrayType()) {
879 const ArrayType *atype;
880 type = type->getCanonicalTypeInternal().getTypePtr();
881 atype = cast<ArrayType>(type);
882 return base_type(atype->getElementType());
884 return qt;
887 /* Extract an access relation from a reference to a variable.
888 * If the variable has name "A" and its type corresponds to an
889 * array of depth d, then the returned access relation is of the
890 * form
892 * { [] -> A[i_1,...,i_d] }
894 __isl_give isl_map *PetScan::extract_access(DeclRefExpr *expr)
896 return extract_access(expr->getDecl());
899 /* Extract an access relation from a variable.
900 * If the variable has name "A" and its type corresponds to an
901 * array of depth d, then the returned access relation is of the
902 * form
904 * { [] -> A[i_1,...,i_d] }
906 __isl_give isl_map *PetScan::extract_access(ValueDecl *decl)
908 int depth = array_depth(decl->getType().getTypePtr());
909 isl_id *id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
910 isl_space *dim = isl_space_alloc(ctx, 0, 0, depth);
911 isl_map *access_rel;
913 dim = isl_space_set_tuple_id(dim, isl_dim_out, id);
915 access_rel = isl_map_universe(dim);
917 return access_rel;
920 /* Extract an access relation from an integer contant.
921 * If the value of the constant is "v", then the returned access relation
922 * is
924 * { [] -> [v] }
926 __isl_give isl_map *PetScan::extract_access(IntegerLiteral *expr)
928 return isl_map_from_range(isl_set_from_pw_aff(extract_affine(expr)));
931 /* Try and extract an access relation from the given Expr.
932 * Return NULL if it doesn't work out.
934 __isl_give isl_map *PetScan::extract_access(Expr *expr)
936 switch (expr->getStmtClass()) {
937 case Stmt::ImplicitCastExprClass:
938 return extract_access(cast<ImplicitCastExpr>(expr));
939 case Stmt::DeclRefExprClass:
940 return extract_access(cast<DeclRefExpr>(expr));
941 case Stmt::ArraySubscriptExprClass:
942 return extract_access(cast<ArraySubscriptExpr>(expr));
943 case Stmt::IntegerLiteralClass:
944 return extract_access(cast<IntegerLiteral>(expr));
945 default:
946 unsupported(expr);
948 return NULL;
951 /* Assign the affine expression "index" to the output dimension "pos" of "map",
952 * restrict the domain to those values that result in a non-negative index
953 * and return the result.
955 __isl_give isl_map *set_index(__isl_take isl_map *map, int pos,
956 __isl_take isl_pw_aff *index)
958 isl_map *index_map;
959 int len = isl_map_dim(map, isl_dim_out);
960 isl_id *id;
961 isl_set *domain;
963 domain = isl_pw_aff_nonneg_set(isl_pw_aff_copy(index));
964 index = isl_pw_aff_intersect_domain(index, domain);
965 index_map = isl_map_from_range(isl_set_from_pw_aff(index));
966 index_map = isl_map_insert_dims(index_map, isl_dim_out, 0, pos);
967 index_map = isl_map_add_dims(index_map, isl_dim_out, len - pos - 1);
968 id = isl_map_get_tuple_id(map, isl_dim_out);
969 index_map = isl_map_set_tuple_id(index_map, isl_dim_out, id);
971 map = isl_map_intersect(map, index_map);
973 return map;
976 /* Extract an access relation from the given array subscript expression.
977 * If nesting is allowed in general, then we turn it on while
978 * examining the index expression.
980 * We first extract an access relation from the base.
981 * This will result in an access relation with a range that corresponds
982 * to the array being accessed and with earlier indices filled in already.
983 * We then extract the current index and fill that in as well.
984 * The position of the current index is based on the type of base.
985 * If base is the actual array variable, then the depth of this type
986 * will be the same as the depth of the array and we will fill in
987 * the first array index.
988 * Otherwise, the depth of the base type will be smaller and we will fill
989 * in a later index.
991 __isl_give isl_map *PetScan::extract_access(ArraySubscriptExpr *expr)
993 Expr *base = expr->getBase();
994 Expr *idx = expr->getIdx();
995 isl_pw_aff *index;
996 isl_map *base_access;
997 isl_map *access;
998 int depth = array_depth(base->getType().getTypePtr());
999 int pos;
1000 bool save_nesting = nesting_enabled;
1002 nesting_enabled = allow_nested;
1004 base_access = extract_access(base);
1005 index = extract_affine(idx);
1007 nesting_enabled = save_nesting;
1009 pos = isl_map_dim(base_access, isl_dim_out) - depth;
1010 access = set_index(base_access, pos, index);
1012 return access;
1015 /* Check if "expr" calls function "minmax" with two arguments and if so
1016 * make lhs and rhs refer to these two arguments.
1018 static bool is_minmax(Expr *expr, const char *minmax, Expr *&lhs, Expr *&rhs)
1020 CallExpr *call;
1021 FunctionDecl *fd;
1022 string name;
1024 if (expr->getStmtClass() != Stmt::CallExprClass)
1025 return false;
1027 call = cast<CallExpr>(expr);
1028 fd = call->getDirectCallee();
1029 if (!fd)
1030 return false;
1032 if (call->getNumArgs() != 2)
1033 return false;
1035 name = fd->getDeclName().getAsString();
1036 if (name != minmax)
1037 return false;
1039 lhs = call->getArg(0);
1040 rhs = call->getArg(1);
1042 return true;
1045 /* Check if "expr" is of the form min(lhs, rhs) and if so make
1046 * lhs and rhs refer to the two arguments.
1048 static bool is_min(Expr *expr, Expr *&lhs, Expr *&rhs)
1050 return is_minmax(expr, "min", lhs, rhs);
1053 /* Check if "expr" is of the form max(lhs, rhs) and if so make
1054 * lhs and rhs refer to the two arguments.
1056 static bool is_max(Expr *expr, Expr *&lhs, Expr *&rhs)
1058 return is_minmax(expr, "max", lhs, rhs);
1061 /* Return "lhs && rhs", defined on the shared definition domain.
1063 static __isl_give isl_pw_aff *pw_aff_and(__isl_take isl_pw_aff *lhs,
1064 __isl_take isl_pw_aff *rhs)
1066 isl_set *cond;
1067 isl_set *dom;
1069 dom = isl_set_intersect(isl_pw_aff_domain(isl_pw_aff_copy(lhs)),
1070 isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1071 cond = isl_set_intersect(isl_pw_aff_non_zero_set(lhs),
1072 isl_pw_aff_non_zero_set(rhs));
1073 return indicator_function(cond, dom);
1076 /* Return "lhs && rhs", with shortcut semantics.
1077 * That is, if lhs is false, then the result is defined even if rhs is not.
1078 * In practice, we compute lhs ? rhs : lhs.
1080 static __isl_give isl_pw_aff *pw_aff_and_then(__isl_take isl_pw_aff *lhs,
1081 __isl_take isl_pw_aff *rhs)
1083 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), rhs, lhs);
1086 /* Return "lhs || rhs", with shortcut semantics.
1087 * That is, if lhs is true, then the result is defined even if rhs is not.
1088 * In practice, we compute lhs ? lhs : rhs.
1090 static __isl_give isl_pw_aff *pw_aff_or_else(__isl_take isl_pw_aff *lhs,
1091 __isl_take isl_pw_aff *rhs)
1093 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), lhs, rhs);
1096 /* Extract an affine expressions representing the comparison "LHS op RHS"
1097 * "comp" is the original statement that "LHS op RHS" is derived from
1098 * and is used for diagnostics.
1100 * If the comparison is of the form
1102 * a <= min(b,c)
1104 * then the expression is constructed as the conjunction of
1105 * the comparisons
1107 * a <= b and a <= c
1109 * A similar optimization is performed for max(a,b) <= c.
1110 * We do this because that will lead to simpler representations
1111 * of the expression.
1112 * If isl is ever enhanced to explicitly deal with min and max expressions,
1113 * this optimization can be removed.
1115 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperatorKind op,
1116 Expr *LHS, Expr *RHS, Stmt *comp)
1118 isl_pw_aff *lhs;
1119 isl_pw_aff *rhs;
1120 isl_pw_aff *res;
1121 isl_set *cond;
1122 isl_set *dom;
1124 if (op == BO_GT)
1125 return extract_comparison(BO_LT, RHS, LHS, comp);
1126 if (op == BO_GE)
1127 return extract_comparison(BO_LE, RHS, LHS, comp);
1129 if (op == BO_LT || op == BO_LE) {
1130 Expr *expr1, *expr2;
1131 if (is_min(RHS, expr1, expr2)) {
1132 lhs = extract_comparison(op, LHS, expr1, comp);
1133 rhs = extract_comparison(op, LHS, expr2, comp);
1134 return pw_aff_and(lhs, rhs);
1136 if (is_max(LHS, expr1, expr2)) {
1137 lhs = extract_comparison(op, expr1, RHS, comp);
1138 rhs = extract_comparison(op, expr2, RHS, comp);
1139 return pw_aff_and(lhs, rhs);
1143 lhs = extract_affine(LHS);
1144 rhs = extract_affine(RHS);
1146 dom = isl_pw_aff_domain(isl_pw_aff_copy(lhs));
1147 dom = isl_set_intersect(dom, isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1149 switch (op) {
1150 case BO_LT:
1151 cond = isl_pw_aff_lt_set(lhs, rhs);
1152 break;
1153 case BO_LE:
1154 cond = isl_pw_aff_le_set(lhs, rhs);
1155 break;
1156 case BO_EQ:
1157 cond = isl_pw_aff_eq_set(lhs, rhs);
1158 break;
1159 case BO_NE:
1160 cond = isl_pw_aff_ne_set(lhs, rhs);
1161 break;
1162 default:
1163 isl_pw_aff_free(lhs);
1164 isl_pw_aff_free(rhs);
1165 isl_set_free(dom);
1166 unsupported(comp);
1167 return NULL;
1170 cond = isl_set_coalesce(cond);
1171 res = indicator_function(cond, dom);
1173 return res;
1176 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperator *comp)
1178 return extract_comparison(comp->getOpcode(), comp->getLHS(),
1179 comp->getRHS(), comp);
1182 /* Extract an affine expression representing the negation (logical not)
1183 * of a subexpression.
1185 __isl_give isl_pw_aff *PetScan::extract_boolean(UnaryOperator *op)
1187 isl_set *set_cond, *dom;
1188 isl_pw_aff *cond, *res;
1190 cond = extract_condition(op->getSubExpr());
1192 dom = isl_pw_aff_domain(isl_pw_aff_copy(cond));
1194 set_cond = isl_pw_aff_zero_set(cond);
1196 res = indicator_function(set_cond, dom);
1198 return res;
1201 /* Extract an affine expression representing the disjunction (logical or)
1202 * or conjunction (logical and) of two subexpressions.
1204 __isl_give isl_pw_aff *PetScan::extract_boolean(BinaryOperator *comp)
1206 isl_pw_aff *lhs, *rhs;
1208 lhs = extract_condition(comp->getLHS());
1209 rhs = extract_condition(comp->getRHS());
1211 switch (comp->getOpcode()) {
1212 case BO_LAnd:
1213 return pw_aff_and_then(lhs, rhs);
1214 case BO_LOr:
1215 return pw_aff_or_else(lhs, rhs);
1216 default:
1217 isl_pw_aff_free(lhs);
1218 isl_pw_aff_free(rhs);
1221 unsupported(comp);
1222 return NULL;
1225 __isl_give isl_pw_aff *PetScan::extract_condition(UnaryOperator *expr)
1227 switch (expr->getOpcode()) {
1228 case UO_LNot:
1229 return extract_boolean(expr);
1230 default:
1231 unsupported(expr);
1232 return NULL;
1236 /* Extract the affine expression "expr != 0 ? 1 : 0".
1238 __isl_give isl_pw_aff *PetScan::extract_implicit_condition(Expr *expr)
1240 isl_pw_aff *res;
1241 isl_set *set, *dom;
1243 res = extract_affine(expr);
1245 dom = isl_pw_aff_domain(isl_pw_aff_copy(res));
1246 set = isl_pw_aff_non_zero_set(res);
1248 res = indicator_function(set, dom);
1250 return res;
1253 /* Extract an affine expression from a boolean expression.
1254 * In particular, return the expression "expr ? 1 : 0".
1256 * If the expression doesn't look like a condition, we assume it
1257 * is an affine expression and return the condition "expr != 0 ? 1 : 0".
1259 __isl_give isl_pw_aff *PetScan::extract_condition(Expr *expr)
1261 BinaryOperator *comp;
1263 if (!expr) {
1264 isl_set *u = isl_set_universe(isl_space_params_alloc(ctx, 0));
1265 return indicator_function(u, isl_set_copy(u));
1268 if (expr->getStmtClass() == Stmt::ParenExprClass)
1269 return extract_condition(cast<ParenExpr>(expr)->getSubExpr());
1271 if (expr->getStmtClass() == Stmt::UnaryOperatorClass)
1272 return extract_condition(cast<UnaryOperator>(expr));
1274 if (expr->getStmtClass() != Stmt::BinaryOperatorClass)
1275 return extract_implicit_condition(expr);
1277 comp = cast<BinaryOperator>(expr);
1278 switch (comp->getOpcode()) {
1279 case BO_LT:
1280 case BO_LE:
1281 case BO_GT:
1282 case BO_GE:
1283 case BO_EQ:
1284 case BO_NE:
1285 return extract_comparison(comp);
1286 case BO_LAnd:
1287 case BO_LOr:
1288 return extract_boolean(comp);
1289 default:
1290 return extract_implicit_condition(expr);
1294 static enum pet_op_type UnaryOperatorKind2pet_op_type(UnaryOperatorKind kind)
1296 switch (kind) {
1297 case UO_Minus:
1298 return pet_op_minus;
1299 case UO_PostInc:
1300 return pet_op_post_inc;
1301 case UO_PostDec:
1302 return pet_op_post_dec;
1303 case UO_PreInc:
1304 return pet_op_pre_inc;
1305 case UO_PreDec:
1306 return pet_op_pre_dec;
1307 default:
1308 return pet_op_last;
1312 static enum pet_op_type BinaryOperatorKind2pet_op_type(BinaryOperatorKind kind)
1314 switch (kind) {
1315 case BO_AddAssign:
1316 return pet_op_add_assign;
1317 case BO_SubAssign:
1318 return pet_op_sub_assign;
1319 case BO_MulAssign:
1320 return pet_op_mul_assign;
1321 case BO_DivAssign:
1322 return pet_op_div_assign;
1323 case BO_Assign:
1324 return pet_op_assign;
1325 case BO_Add:
1326 return pet_op_add;
1327 case BO_Sub:
1328 return pet_op_sub;
1329 case BO_Mul:
1330 return pet_op_mul;
1331 case BO_Div:
1332 return pet_op_div;
1333 case BO_Rem:
1334 return pet_op_mod;
1335 case BO_EQ:
1336 return pet_op_eq;
1337 case BO_LE:
1338 return pet_op_le;
1339 case BO_LT:
1340 return pet_op_lt;
1341 case BO_GT:
1342 return pet_op_gt;
1343 default:
1344 return pet_op_last;
1348 /* Construct a pet_expr representing a unary operator expression.
1350 struct pet_expr *PetScan::extract_expr(UnaryOperator *expr)
1352 struct pet_expr *arg;
1353 enum pet_op_type op;
1355 op = UnaryOperatorKind2pet_op_type(expr->getOpcode());
1356 if (op == pet_op_last) {
1357 unsupported(expr);
1358 return NULL;
1361 arg = extract_expr(expr->getSubExpr());
1363 if (expr->isIncrementDecrementOp() &&
1364 arg && arg->type == pet_expr_access) {
1365 mark_write(arg);
1366 arg->acc.read = 1;
1369 return pet_expr_new_unary(ctx, op, arg);
1372 /* Mark the given access pet_expr as a write.
1373 * If a scalar is being accessed, then mark its value
1374 * as unknown in assigned_value.
1376 void PetScan::mark_write(struct pet_expr *access)
1378 isl_id *id;
1379 ValueDecl *decl;
1381 if (!access)
1382 return;
1384 access->acc.write = 1;
1385 access->acc.read = 0;
1387 if (isl_map_dim(access->acc.access, isl_dim_out) != 0)
1388 return;
1390 id = isl_map_get_tuple_id(access->acc.access, isl_dim_out);
1391 decl = (ValueDecl *) isl_id_get_user(id);
1392 clear_assignment(assigned_value, decl);
1393 isl_id_free(id);
1396 /* Assign "rhs" to "lhs".
1398 * In particular, if "lhs" is a scalar variable, then mark
1399 * the variable as having been assigned. If, furthermore, "rhs"
1400 * is an affine expression, then keep track of this value in assigned_value
1401 * so that we can plug it in when we later come across the same variable.
1403 void PetScan::assign(struct pet_expr *lhs, Expr *rhs)
1405 isl_id *id;
1406 ValueDecl *decl;
1407 isl_pw_aff *pa;
1409 if (!lhs)
1410 return;
1411 if (lhs->type != pet_expr_access)
1412 return;
1413 if (isl_map_dim(lhs->acc.access, isl_dim_out) != 0)
1414 return;
1416 id = isl_map_get_tuple_id(lhs->acc.access, isl_dim_out);
1417 decl = (ValueDecl *) isl_id_get_user(id);
1418 isl_id_free(id);
1420 pa = try_extract_affine(rhs);
1421 clear_assignment(assigned_value, decl);
1422 if (!pa)
1423 return;
1424 assigned_value[decl] = pa;
1425 insert_expression(pa);
1428 /* Construct a pet_expr representing a binary operator expression.
1430 * If the top level operator is an assignment and the LHS is an access,
1431 * then we mark that access as a write. If the operator is a compound
1432 * assignment, the access is marked as both a read and a write.
1434 * If "expr" assigns something to a scalar variable, then we mark
1435 * the variable as having been assigned. If, furthermore, the expression
1436 * is affine, then keep track of this value in assigned_value
1437 * so that we can plug it in when we later come across the same variable.
1439 struct pet_expr *PetScan::extract_expr(BinaryOperator *expr)
1441 struct pet_expr *lhs, *rhs;
1442 enum pet_op_type op;
1444 op = BinaryOperatorKind2pet_op_type(expr->getOpcode());
1445 if (op == pet_op_last) {
1446 unsupported(expr);
1447 return NULL;
1450 lhs = extract_expr(expr->getLHS());
1451 rhs = extract_expr(expr->getRHS());
1453 if (expr->isAssignmentOp() && lhs && lhs->type == pet_expr_access) {
1454 mark_write(lhs);
1455 if (expr->isCompoundAssignmentOp())
1456 lhs->acc.read = 1;
1459 if (expr->getOpcode() == BO_Assign)
1460 assign(lhs, expr->getRHS());
1462 return pet_expr_new_binary(ctx, op, lhs, rhs);
1465 /* Construct a pet_scop with a single statement killing the entire
1466 * array "array".
1468 struct pet_scop *PetScan::kill(Stmt *stmt, struct pet_array *array)
1470 isl_map *access;
1471 struct pet_expr *expr;
1473 if (!array)
1474 return NULL;
1475 access = isl_map_from_range(isl_set_copy(array->extent));
1476 expr = pet_expr_kill_from_access(access);
1477 return extract(stmt, expr);
1480 /* Construct a pet_scop for a (single) variable declaration.
1482 * The scop contains the variable being declared (as an array)
1483 * and a statement killing the array.
1485 * If the variable is initialized in the AST, then the scop
1486 * also contains an assignment to the variable.
1488 struct pet_scop *PetScan::extract(DeclStmt *stmt)
1490 Decl *decl;
1491 VarDecl *vd;
1492 struct pet_expr *lhs, *rhs, *pe;
1493 struct pet_scop *scop_decl, *scop;
1494 struct pet_array *array;
1496 if (!stmt->isSingleDecl()) {
1497 unsupported(stmt);
1498 return NULL;
1501 decl = stmt->getSingleDecl();
1502 vd = cast<VarDecl>(decl);
1504 array = extract_array(ctx, vd);
1505 if (array)
1506 array->declared = 1;
1507 scop_decl = kill(stmt, array);
1508 scop_decl = pet_scop_add_array(scop_decl, array);
1510 if (!vd->getInit())
1511 return scop_decl;
1513 lhs = pet_expr_from_access(extract_access(vd));
1514 rhs = extract_expr(vd->getInit());
1516 mark_write(lhs);
1517 assign(lhs, vd->getInit());
1519 pe = pet_expr_new_binary(ctx, pet_op_assign, lhs, rhs);
1520 scop = extract(stmt, pe);
1522 scop_decl = pet_scop_prefix(scop_decl, 0);
1523 scop = pet_scop_prefix(scop, 1);
1525 scop = pet_scop_add_seq(ctx, scop_decl, scop);
1527 return scop;
1530 /* Construct a pet_expr representing a conditional operation.
1532 * We first try to extract the condition as an affine expression.
1533 * If that fails, we construct a pet_expr tree representing the condition.
1535 struct pet_expr *PetScan::extract_expr(ConditionalOperator *expr)
1537 struct pet_expr *cond, *lhs, *rhs;
1538 isl_pw_aff *pa;
1540 pa = try_extract_affine(expr->getCond());
1541 if (pa) {
1542 isl_set *test = isl_set_from_pw_aff(pa);
1543 cond = pet_expr_from_access(isl_map_from_range(test));
1544 } else
1545 cond = extract_expr(expr->getCond());
1546 lhs = extract_expr(expr->getTrueExpr());
1547 rhs = extract_expr(expr->getFalseExpr());
1549 return pet_expr_new_ternary(ctx, cond, lhs, rhs);
1552 struct pet_expr *PetScan::extract_expr(ImplicitCastExpr *expr)
1554 return extract_expr(expr->getSubExpr());
1557 /* Construct a pet_expr representing a floating point value.
1559 * If the floating point literal does not appear in a macro,
1560 * then we use the original representation in the source code
1561 * as the string representation. Otherwise, we use the pretty
1562 * printer to produce a string representation.
1564 struct pet_expr *PetScan::extract_expr(FloatingLiteral *expr)
1566 double d;
1567 string s;
1568 const LangOptions &LO = PP.getLangOpts();
1569 SourceLocation loc = expr->getLocation();
1571 if (!loc.isMacroID()) {
1572 SourceManager &SM = PP.getSourceManager();
1573 unsigned len = Lexer::MeasureTokenLength(loc, SM, LO);
1574 s = string(SM.getCharacterData(loc), len);
1575 } else {
1576 llvm::raw_string_ostream S(s);
1577 expr->printPretty(S, 0, PrintingPolicy(LO));
1578 S.str();
1580 d = expr->getValueAsApproximateDouble();
1581 return pet_expr_new_double(ctx, d, s.c_str());
1584 /* Extract an access relation from "expr" and then convert it into
1585 * a pet_expr.
1587 struct pet_expr *PetScan::extract_access_expr(Expr *expr)
1589 isl_map *access;
1590 struct pet_expr *pe;
1592 access = extract_access(expr);
1594 pe = pet_expr_from_access(access);
1596 return pe;
1599 struct pet_expr *PetScan::extract_expr(ParenExpr *expr)
1601 return extract_expr(expr->getSubExpr());
1604 /* Construct a pet_expr representing a function call.
1606 * If we are passing along a pointer to an array element
1607 * or an entire row or even higher dimensional slice of an array,
1608 * then the function being called may write into the array.
1610 * We assume here that if the function is declared to take a pointer
1611 * to a const type, then the function will perform a read
1612 * and that otherwise, it will perform a write.
1614 struct pet_expr *PetScan::extract_expr(CallExpr *expr)
1616 struct pet_expr *res = NULL;
1617 FunctionDecl *fd;
1618 string name;
1620 fd = expr->getDirectCallee();
1621 if (!fd) {
1622 unsupported(expr);
1623 return NULL;
1626 name = fd->getDeclName().getAsString();
1627 res = pet_expr_new_call(ctx, name.c_str(), expr->getNumArgs());
1628 if (!res)
1629 return NULL;
1631 for (int i = 0; i < expr->getNumArgs(); ++i) {
1632 Expr *arg = expr->getArg(i);
1633 int is_addr = 0;
1634 pet_expr *main_arg;
1636 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
1637 ImplicitCastExpr *ice = cast<ImplicitCastExpr>(arg);
1638 arg = ice->getSubExpr();
1640 if (arg->getStmtClass() == Stmt::UnaryOperatorClass) {
1641 UnaryOperator *op = cast<UnaryOperator>(arg);
1642 if (op->getOpcode() == UO_AddrOf) {
1643 is_addr = 1;
1644 arg = op->getSubExpr();
1647 res->args[i] = PetScan::extract_expr(arg);
1648 main_arg = res->args[i];
1649 if (is_addr)
1650 res->args[i] = pet_expr_new_unary(ctx,
1651 pet_op_address_of, res->args[i]);
1652 if (!res->args[i])
1653 goto error;
1654 if (arg->getStmtClass() == Stmt::ArraySubscriptExprClass &&
1655 array_depth(arg->getType().getTypePtr()) > 0)
1656 is_addr = 1;
1657 if (is_addr && main_arg->type == pet_expr_access) {
1658 ParmVarDecl *parm;
1659 if (!fd->hasPrototype()) {
1660 unsupported(expr, "prototype required");
1661 goto error;
1663 parm = fd->getParamDecl(i);
1664 if (!const_base(parm->getType()))
1665 mark_write(main_arg);
1669 return res;
1670 error:
1671 pet_expr_free(res);
1672 return NULL;
1675 /* Try and onstruct a pet_expr representing "expr".
1677 struct pet_expr *PetScan::extract_expr(Expr *expr)
1679 switch (expr->getStmtClass()) {
1680 case Stmt::UnaryOperatorClass:
1681 return extract_expr(cast<UnaryOperator>(expr));
1682 case Stmt::CompoundAssignOperatorClass:
1683 case Stmt::BinaryOperatorClass:
1684 return extract_expr(cast<BinaryOperator>(expr));
1685 case Stmt::ImplicitCastExprClass:
1686 return extract_expr(cast<ImplicitCastExpr>(expr));
1687 case Stmt::ArraySubscriptExprClass:
1688 case Stmt::DeclRefExprClass:
1689 case Stmt::IntegerLiteralClass:
1690 return extract_access_expr(expr);
1691 case Stmt::FloatingLiteralClass:
1692 return extract_expr(cast<FloatingLiteral>(expr));
1693 case Stmt::ParenExprClass:
1694 return extract_expr(cast<ParenExpr>(expr));
1695 case Stmt::ConditionalOperatorClass:
1696 return extract_expr(cast<ConditionalOperator>(expr));
1697 case Stmt::CallExprClass:
1698 return extract_expr(cast<CallExpr>(expr));
1699 default:
1700 unsupported(expr);
1702 return NULL;
1705 /* Check if the given initialization statement is an assignment.
1706 * If so, return that assignment. Otherwise return NULL.
1708 BinaryOperator *PetScan::initialization_assignment(Stmt *init)
1710 BinaryOperator *ass;
1712 if (init->getStmtClass() != Stmt::BinaryOperatorClass)
1713 return NULL;
1715 ass = cast<BinaryOperator>(init);
1716 if (ass->getOpcode() != BO_Assign)
1717 return NULL;
1719 return ass;
1722 /* Check if the given initialization statement is a declaration
1723 * of a single variable.
1724 * If so, return that declaration. Otherwise return NULL.
1726 Decl *PetScan::initialization_declaration(Stmt *init)
1728 DeclStmt *decl;
1730 if (init->getStmtClass() != Stmt::DeclStmtClass)
1731 return NULL;
1733 decl = cast<DeclStmt>(init);
1735 if (!decl->isSingleDecl())
1736 return NULL;
1738 return decl->getSingleDecl();
1741 /* Given the assignment operator in the initialization of a for loop,
1742 * extract the induction variable, i.e., the (integer)variable being
1743 * assigned.
1745 ValueDecl *PetScan::extract_induction_variable(BinaryOperator *init)
1747 Expr *lhs;
1748 DeclRefExpr *ref;
1749 ValueDecl *decl;
1750 const Type *type;
1752 lhs = init->getLHS();
1753 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1754 unsupported(init);
1755 return NULL;
1758 ref = cast<DeclRefExpr>(lhs);
1759 decl = ref->getDecl();
1760 type = decl->getType().getTypePtr();
1762 if (!type->isIntegerType()) {
1763 unsupported(lhs);
1764 return NULL;
1767 return decl;
1770 /* Given the initialization statement of a for loop and the single
1771 * declaration in this initialization statement,
1772 * extract the induction variable, i.e., the (integer) variable being
1773 * declared.
1775 VarDecl *PetScan::extract_induction_variable(Stmt *init, Decl *decl)
1777 VarDecl *vd;
1779 vd = cast<VarDecl>(decl);
1781 const QualType type = vd->getType();
1782 if (!type->isIntegerType()) {
1783 unsupported(init);
1784 return NULL;
1787 if (!vd->getInit()) {
1788 unsupported(init);
1789 return NULL;
1792 return vd;
1795 /* Check that op is of the form iv++ or iv--.
1796 * Return an affine expression "1" or "-1" accordingly.
1798 __isl_give isl_pw_aff *PetScan::extract_unary_increment(
1799 clang::UnaryOperator *op, clang::ValueDecl *iv)
1801 Expr *sub;
1802 DeclRefExpr *ref;
1803 isl_space *space;
1804 isl_aff *aff;
1806 if (!op->isIncrementDecrementOp()) {
1807 unsupported(op);
1808 return NULL;
1811 sub = op->getSubExpr();
1812 if (sub->getStmtClass() != Stmt::DeclRefExprClass) {
1813 unsupported(op);
1814 return NULL;
1817 ref = cast<DeclRefExpr>(sub);
1818 if (ref->getDecl() != iv) {
1819 unsupported(op);
1820 return NULL;
1823 space = isl_space_params_alloc(ctx, 0);
1824 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
1826 if (op->isIncrementOp())
1827 aff = isl_aff_add_constant_si(aff, 1);
1828 else
1829 aff = isl_aff_add_constant_si(aff, -1);
1831 return isl_pw_aff_from_aff(aff);
1834 /* If the isl_pw_aff on which isl_pw_aff_foreach_piece is called
1835 * has a single constant expression, then put this constant in *user.
1836 * The caller is assumed to have checked that this function will
1837 * be called exactly once.
1839 static int extract_cst(__isl_take isl_set *set, __isl_take isl_aff *aff,
1840 void *user)
1842 isl_int *inc = (isl_int *)user;
1843 int res = 0;
1845 if (isl_aff_is_cst(aff))
1846 isl_aff_get_constant(aff, inc);
1847 else
1848 res = -1;
1850 isl_set_free(set);
1851 isl_aff_free(aff);
1853 return res;
1856 /* Check if op is of the form
1858 * iv = iv + inc
1860 * and return inc as an affine expression.
1862 * We extract an affine expression from the RHS, subtract iv and return
1863 * the result.
1865 __isl_give isl_pw_aff *PetScan::extract_binary_increment(BinaryOperator *op,
1866 clang::ValueDecl *iv)
1868 Expr *lhs;
1869 DeclRefExpr *ref;
1870 isl_id *id;
1871 isl_space *dim;
1872 isl_aff *aff;
1873 isl_pw_aff *val;
1875 if (op->getOpcode() != BO_Assign) {
1876 unsupported(op);
1877 return NULL;
1880 lhs = op->getLHS();
1881 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1882 unsupported(op);
1883 return NULL;
1886 ref = cast<DeclRefExpr>(lhs);
1887 if (ref->getDecl() != iv) {
1888 unsupported(op);
1889 return NULL;
1892 val = extract_affine(op->getRHS());
1894 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
1896 dim = isl_space_params_alloc(ctx, 1);
1897 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
1898 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
1899 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
1901 val = isl_pw_aff_sub(val, isl_pw_aff_from_aff(aff));
1903 return val;
1906 /* Check that op is of the form iv += cst or iv -= cst
1907 * and return an affine expression corresponding oto cst or -cst accordingly.
1909 __isl_give isl_pw_aff *PetScan::extract_compound_increment(
1910 CompoundAssignOperator *op, clang::ValueDecl *iv)
1912 Expr *lhs;
1913 DeclRefExpr *ref;
1914 bool neg = false;
1915 isl_pw_aff *val;
1916 BinaryOperatorKind opcode;
1918 opcode = op->getOpcode();
1919 if (opcode != BO_AddAssign && opcode != BO_SubAssign) {
1920 unsupported(op);
1921 return NULL;
1923 if (opcode == BO_SubAssign)
1924 neg = true;
1926 lhs = op->getLHS();
1927 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1928 unsupported(op);
1929 return NULL;
1932 ref = cast<DeclRefExpr>(lhs);
1933 if (ref->getDecl() != iv) {
1934 unsupported(op);
1935 return NULL;
1938 val = extract_affine(op->getRHS());
1939 if (neg)
1940 val = isl_pw_aff_neg(val);
1942 return val;
1945 /* Check that the increment of the given for loop increments
1946 * (or decrements) the induction variable "iv" and return
1947 * the increment as an affine expression if successful.
1949 __isl_give isl_pw_aff *PetScan::extract_increment(clang::ForStmt *stmt,
1950 ValueDecl *iv)
1952 Stmt *inc = stmt->getInc();
1954 if (!inc) {
1955 unsupported(stmt);
1956 return NULL;
1959 if (inc->getStmtClass() == Stmt::UnaryOperatorClass)
1960 return extract_unary_increment(cast<UnaryOperator>(inc), iv);
1961 if (inc->getStmtClass() == Stmt::CompoundAssignOperatorClass)
1962 return extract_compound_increment(
1963 cast<CompoundAssignOperator>(inc), iv);
1964 if (inc->getStmtClass() == Stmt::BinaryOperatorClass)
1965 return extract_binary_increment(cast<BinaryOperator>(inc), iv);
1967 unsupported(inc);
1968 return NULL;
1971 /* Embed the given iteration domain in an extra outer loop
1972 * with induction variable "var".
1973 * If this variable appeared as a parameter in the constraints,
1974 * it is replaced by the new outermost dimension.
1976 static __isl_give isl_set *embed(__isl_take isl_set *set,
1977 __isl_take isl_id *var)
1979 int pos;
1981 set = isl_set_insert_dims(set, isl_dim_set, 0, 1);
1982 pos = isl_set_find_dim_by_id(set, isl_dim_param, var);
1983 if (pos >= 0) {
1984 set = isl_set_equate(set, isl_dim_param, pos, isl_dim_set, 0);
1985 set = isl_set_project_out(set, isl_dim_param, pos, 1);
1988 isl_id_free(var);
1989 return set;
1992 /* Return those elements in the space of "cond" that come after
1993 * (based on "sign") an element in "cond".
1995 static __isl_give isl_set *after(__isl_take isl_set *cond, int sign)
1997 isl_map *previous_to_this;
1999 if (sign > 0)
2000 previous_to_this = isl_map_lex_lt(isl_set_get_space(cond));
2001 else
2002 previous_to_this = isl_map_lex_gt(isl_set_get_space(cond));
2004 cond = isl_set_apply(cond, previous_to_this);
2006 return cond;
2009 /* Create the infinite iteration domain
2011 * { [id] : id >= 0 }
2013 * If "scop" has an affine skip of type pet_skip_later,
2014 * then remove those iterations i that have an earlier iteration
2015 * where the skip condition is satisfied, meaning that iteration i
2016 * is not executed.
2017 * Since we are dealing with a loop without loop iterator,
2018 * the skip condition cannot refer to the current loop iterator and
2019 * so effectively, the returned set is of the form
2021 * { [0]; [id] : id >= 1 and not skip }
2023 static __isl_give isl_set *infinite_domain(__isl_take isl_id *id,
2024 struct pet_scop *scop)
2026 isl_ctx *ctx = isl_id_get_ctx(id);
2027 isl_set *domain;
2028 isl_set *skip;
2030 domain = isl_set_nat_universe(isl_space_set_alloc(ctx, 0, 1));
2031 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, id);
2033 if (!pet_scop_has_affine_skip(scop, pet_skip_later))
2034 return domain;
2036 skip = pet_scop_get_skip(scop, pet_skip_later);
2037 skip = isl_set_fix_si(skip, isl_dim_set, 0, 1);
2038 skip = isl_set_params(skip);
2039 skip = embed(skip, isl_id_copy(id));
2040 skip = isl_set_intersect(skip , isl_set_copy(domain));
2041 domain = isl_set_subtract(domain, after(skip, 1));
2043 return domain;
2046 /* Create an identity mapping on the space containing "domain".
2048 static __isl_give isl_map *identity_map(__isl_keep isl_set *domain)
2050 isl_space *space;
2051 isl_map *id;
2053 space = isl_space_map_from_set(isl_set_get_space(domain));
2054 id = isl_map_identity(space);
2056 return id;
2059 /* Add a filter to "scop" that imposes that it is only executed
2060 * when "break_access" has a zero value for all previous iterations
2061 * of "domain".
2063 * The input "break_access" has a zero-dimensional domain and range.
2065 static struct pet_scop *scop_add_break(struct pet_scop *scop,
2066 __isl_take isl_map *break_access, __isl_take isl_set *domain, int sign)
2068 isl_ctx *ctx = isl_set_get_ctx(domain);
2069 isl_id *id_test;
2070 isl_map *prev;
2072 id_test = isl_map_get_tuple_id(break_access, isl_dim_out);
2073 break_access = isl_map_add_dims(break_access, isl_dim_in, 1);
2074 break_access = isl_map_add_dims(break_access, isl_dim_out, 1);
2075 break_access = isl_map_intersect_range(break_access, domain);
2076 break_access = isl_map_set_tuple_id(break_access, isl_dim_out, id_test);
2077 if (sign > 0)
2078 prev = isl_map_lex_gt_first(isl_map_get_space(break_access), 1);
2079 else
2080 prev = isl_map_lex_lt_first(isl_map_get_space(break_access), 1);
2081 break_access = isl_map_intersect(break_access, prev);
2082 scop = pet_scop_filter(scop, break_access, 0);
2083 scop = pet_scop_merge_filters(scop);
2085 return scop;
2088 /* Construct a pet_scop for an infinite loop around the given body.
2090 * We extract a pet_scop for the body and then embed it in a loop with
2091 * iteration domain
2093 * { [t] : t >= 0 }
2095 * and schedule
2097 * { [t] -> [t] }
2099 * If the body contains any break, then it is taken into
2100 * account in infinite_domain (if the skip condition is affine)
2101 * or in scop_add_break (if the skip condition is not affine).
2103 struct pet_scop *PetScan::extract_infinite_loop(Stmt *body)
2105 isl_id *id;
2106 isl_set *domain;
2107 isl_map *ident;
2108 isl_map *access;
2109 struct pet_scop *scop;
2110 bool has_var_break;
2112 scop = extract(body);
2113 if (!scop)
2114 return NULL;
2116 id = isl_id_alloc(ctx, "t", NULL);
2117 domain = infinite_domain(isl_id_copy(id), scop);
2118 ident = identity_map(domain);
2120 has_var_break = pet_scop_has_var_skip(scop, pet_skip_later);
2121 if (has_var_break)
2122 access = pet_scop_get_skip_map(scop, pet_skip_later);
2124 scop = pet_scop_embed(scop, isl_set_copy(domain),
2125 isl_map_copy(ident), ident, id);
2126 if (has_var_break)
2127 scop = scop_add_break(scop, access, domain, 1);
2128 else
2129 isl_set_free(domain);
2131 return scop;
2134 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
2136 * for (;;)
2137 * body
2140 struct pet_scop *PetScan::extract_infinite_for(ForStmt *stmt)
2142 return extract_infinite_loop(stmt->getBody());
2145 /* Create an access to a virtual array representing the result
2146 * of a condition.
2147 * Unlike other accessed data, the id of the array is NULL as
2148 * there is no ValueDecl in the program corresponding to the virtual
2149 * array.
2150 * The array starts out as a scalar, but grows along with the
2151 * statement writing to the array in pet_scop_embed.
2153 static __isl_give isl_map *create_test_access(isl_ctx *ctx, int test_nr)
2155 isl_space *dim = isl_space_alloc(ctx, 0, 0, 0);
2156 isl_id *id;
2157 char name[50];
2159 snprintf(name, sizeof(name), "__pet_test_%d", test_nr);
2160 id = isl_id_alloc(ctx, name, NULL);
2161 dim = isl_space_set_tuple_id(dim, isl_dim_out, id);
2162 return isl_map_universe(dim);
2165 /* Add an array with the given extent ("access") to the list
2166 * of arrays in "scop" and return the extended pet_scop.
2167 * The array is marked as attaining values 0 and 1 only and
2168 * as each element being assigned at most once.
2170 static struct pet_scop *scop_add_array(struct pet_scop *scop,
2171 __isl_keep isl_map *access, clang::ASTContext &ast_ctx)
2173 isl_ctx *ctx = isl_map_get_ctx(access);
2174 isl_space *dim;
2175 struct pet_array *array;
2177 if (!scop)
2178 return NULL;
2179 if (!ctx)
2180 goto error;
2182 array = isl_calloc_type(ctx, struct pet_array);
2183 if (!array)
2184 goto error;
2186 array->extent = isl_map_range(isl_map_copy(access));
2187 dim = isl_space_params_alloc(ctx, 0);
2188 array->context = isl_set_universe(dim);
2189 dim = isl_space_set_alloc(ctx, 0, 1);
2190 array->value_bounds = isl_set_universe(dim);
2191 array->value_bounds = isl_set_lower_bound_si(array->value_bounds,
2192 isl_dim_set, 0, 0);
2193 array->value_bounds = isl_set_upper_bound_si(array->value_bounds,
2194 isl_dim_set, 0, 1);
2195 array->element_type = strdup("int");
2196 array->element_size = ast_ctx.getTypeInfo(ast_ctx.IntTy).first / 8;
2197 array->uniquely_defined = 1;
2199 if (!array->extent || !array->context)
2200 array = pet_array_free(array);
2202 scop = pet_scop_add_array(scop, array);
2204 return scop;
2205 error:
2206 pet_scop_free(scop);
2207 return NULL;
2210 /* Construct a pet_scop for a while loop of the form
2212 * while (pa)
2213 * body
2215 * In particular, construct a scop for an infinite loop around body and
2216 * intersect the domain with the affine expression.
2217 * Note that this intersection may result in an empty loop.
2219 struct pet_scop *PetScan::extract_affine_while(__isl_take isl_pw_aff *pa,
2220 Stmt *body)
2222 struct pet_scop *scop;
2223 isl_set *dom;
2224 isl_set *valid;
2226 valid = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2227 dom = isl_pw_aff_non_zero_set(pa);
2228 scop = extract_infinite_loop(body);
2229 scop = pet_scop_restrict(scop, dom);
2230 scop = pet_scop_restrict_context(scop, valid);
2232 return scop;
2235 /* Construct a scop for a while, given the scops for the condition
2236 * and the body, the filter access and the iteration domain of
2237 * the while loop.
2239 * In particular, the scop for the condition is filtered to depend
2240 * on "test_access" evaluating to true for all previous iterations
2241 * of the loop, while the scop for the body is filtered to depend
2242 * on "test_access" evaluating to true for all iterations up to the
2243 * current iteration.
2245 * These filtered scops are then combined into a single scop.
2247 * "sign" is positive if the iterator increases and negative
2248 * if it decreases.
2250 static struct pet_scop *scop_add_while(struct pet_scop *scop_cond,
2251 struct pet_scop *scop_body, __isl_take isl_map *test_access,
2252 __isl_take isl_set *domain, int sign)
2254 isl_ctx *ctx = isl_set_get_ctx(domain);
2255 isl_id *id_test;
2256 isl_map *prev;
2258 id_test = isl_map_get_tuple_id(test_access, isl_dim_out);
2259 test_access = isl_map_add_dims(test_access, isl_dim_in, 1);
2260 test_access = isl_map_add_dims(test_access, isl_dim_out, 1);
2261 test_access = isl_map_intersect_range(test_access, domain);
2262 test_access = isl_map_set_tuple_id(test_access, isl_dim_out, id_test);
2263 if (sign > 0)
2264 prev = isl_map_lex_ge_first(isl_map_get_space(test_access), 1);
2265 else
2266 prev = isl_map_lex_le_first(isl_map_get_space(test_access), 1);
2267 test_access = isl_map_intersect(test_access, prev);
2268 scop_body = pet_scop_filter(scop_body, isl_map_copy(test_access), 1);
2269 if (sign > 0)
2270 prev = isl_map_lex_gt_first(isl_map_get_space(test_access), 1);
2271 else
2272 prev = isl_map_lex_lt_first(isl_map_get_space(test_access), 1);
2273 test_access = isl_map_intersect(test_access, prev);
2274 scop_cond = pet_scop_filter(scop_cond, test_access, 1);
2276 return pet_scop_add_seq(ctx, scop_cond, scop_body);
2279 /* Check if the while loop is of the form
2281 * while (affine expression)
2282 * body
2284 * If so, call extract_affine_while to construct a scop.
2286 * Otherwise, construct a generic while scop, with iteration domain
2287 * { [t] : t >= 0 }. The scop consists of two parts, one for
2288 * evaluating the condition and one for the body.
2289 * The schedule is adjusted to reflect that the condition is evaluated
2290 * before the body is executed and the body is filtered to depend
2291 * on the result of the condition evaluating to true on all iterations
2292 * up to the current iteration, while the evaluation the condition itself
2293 * is filtered to depend on the result of the condition evaluating to true
2294 * on all previous iterations.
2295 * The context of the scop representing the body is dropped
2296 * because we don't know how many times the body will be executed,
2297 * if at all.
2299 * If the body contains any break, then it is taken into
2300 * account in infinite_domain (if the skip condition is affine)
2301 * or in scop_add_break (if the skip condition is not affine).
2303 struct pet_scop *PetScan::extract(WhileStmt *stmt)
2305 Expr *cond;
2306 isl_id *id;
2307 isl_map *test_access;
2308 isl_set *domain;
2309 isl_map *ident;
2310 isl_pw_aff *pa;
2311 struct pet_scop *scop, *scop_body;
2312 bool has_var_break;
2313 isl_map *break_access;
2315 cond = stmt->getCond();
2316 if (!cond) {
2317 unsupported(stmt);
2318 return NULL;
2321 clear_assignments clear(assigned_value);
2322 clear.TraverseStmt(stmt->getBody());
2324 pa = try_extract_affine_condition(cond);
2325 if (pa)
2326 return extract_affine_while(pa, stmt->getBody());
2328 if (!allow_nested) {
2329 unsupported(stmt);
2330 return NULL;
2333 test_access = create_test_access(ctx, n_test++);
2334 scop = extract_non_affine_condition(cond, isl_map_copy(test_access));
2335 scop = scop_add_array(scop, test_access, ast_context);
2336 scop_body = extract(stmt->getBody());
2338 id = isl_id_alloc(ctx, "t", NULL);
2339 domain = infinite_domain(isl_id_copy(id), scop_body);
2340 ident = identity_map(domain);
2342 has_var_break = pet_scop_has_var_skip(scop_body, pet_skip_later);
2343 if (has_var_break)
2344 break_access = pet_scop_get_skip_map(scop_body, pet_skip_later);
2346 scop = pet_scop_prefix(scop, 0);
2347 scop = pet_scop_embed(scop, isl_set_copy(domain), isl_map_copy(ident),
2348 isl_map_copy(ident), isl_id_copy(id));
2349 scop_body = pet_scop_reset_context(scop_body);
2350 scop_body = pet_scop_prefix(scop_body, 1);
2351 scop_body = pet_scop_embed(scop_body, isl_set_copy(domain),
2352 isl_map_copy(ident), ident, id);
2354 if (has_var_break) {
2355 scop = scop_add_break(scop, isl_map_copy(break_access),
2356 isl_set_copy(domain), 1);
2357 scop_body = scop_add_break(scop_body, break_access,
2358 isl_set_copy(domain), 1);
2360 scop = scop_add_while(scop, scop_body, test_access, domain, 1);
2362 return scop;
2365 /* Check whether "cond" expresses a simple loop bound
2366 * on the only set dimension.
2367 * In particular, if "up" is set then "cond" should contain only
2368 * upper bounds on the set dimension.
2369 * Otherwise, it should contain only lower bounds.
2371 static bool is_simple_bound(__isl_keep isl_set *cond, isl_int inc)
2373 if (isl_int_is_pos(inc))
2374 return !isl_set_dim_has_any_lower_bound(cond, isl_dim_set, 0);
2375 else
2376 return !isl_set_dim_has_any_upper_bound(cond, isl_dim_set, 0);
2379 /* Extend a condition on a given iteration of a loop to one that
2380 * imposes the same condition on all previous iterations.
2381 * "domain" expresses the lower [upper] bound on the iterations
2382 * when inc is positive [negative].
2384 * In particular, we construct the condition (when inc is positive)
2386 * forall i' : (domain(i') and i' <= i) => cond(i')
2388 * which is equivalent to
2390 * not exists i' : domain(i') and i' <= i and not cond(i')
2392 * We construct this set by negating cond, applying a map
2394 * { [i'] -> [i] : domain(i') and i' <= i }
2396 * and then negating the result again.
2398 static __isl_give isl_set *valid_for_each_iteration(__isl_take isl_set *cond,
2399 __isl_take isl_set *domain, isl_int inc)
2401 isl_map *previous_to_this;
2403 if (isl_int_is_pos(inc))
2404 previous_to_this = isl_map_lex_le(isl_set_get_space(domain));
2405 else
2406 previous_to_this = isl_map_lex_ge(isl_set_get_space(domain));
2408 previous_to_this = isl_map_intersect_domain(previous_to_this, domain);
2410 cond = isl_set_complement(cond);
2411 cond = isl_set_apply(cond, previous_to_this);
2412 cond = isl_set_complement(cond);
2414 return cond;
2417 /* Construct a domain of the form
2419 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
2421 static __isl_give isl_set *strided_domain(__isl_take isl_id *id,
2422 __isl_take isl_pw_aff *init, isl_int inc)
2424 isl_aff *aff;
2425 isl_space *dim;
2426 isl_set *set;
2428 init = isl_pw_aff_insert_dims(init, isl_dim_in, 0, 1);
2429 dim = isl_pw_aff_get_domain_space(init);
2430 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2431 aff = isl_aff_add_coefficient(aff, isl_dim_in, 0, inc);
2432 init = isl_pw_aff_add(init, isl_pw_aff_from_aff(aff));
2434 dim = isl_space_set_alloc(isl_pw_aff_get_ctx(init), 1, 1);
2435 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
2436 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2437 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2439 set = isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff), init);
2441 set = isl_set_lower_bound_si(set, isl_dim_set, 0, 0);
2443 return isl_set_params(set);
2446 /* Assuming "cond" represents a bound on a loop where the loop
2447 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2448 * is possible.
2450 * Under the given assumptions, wrapping is only possible if "cond" allows
2451 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2452 * increasing iterator and 0 in case of a decreasing iterator.
2454 static bool can_wrap(__isl_keep isl_set *cond, ValueDecl *iv, isl_int inc)
2456 bool cw;
2457 isl_int limit;
2458 isl_set *test;
2460 test = isl_set_copy(cond);
2462 isl_int_init(limit);
2463 if (isl_int_is_neg(inc))
2464 isl_int_set_si(limit, 0);
2465 else {
2466 isl_int_set_si(limit, 1);
2467 isl_int_mul_2exp(limit, limit, get_type_size(iv));
2468 isl_int_sub_ui(limit, limit, 1);
2471 test = isl_set_fix(cond, isl_dim_set, 0, limit);
2472 cw = !isl_set_is_empty(test);
2473 isl_set_free(test);
2475 isl_int_clear(limit);
2477 return cw;
2480 /* Given a one-dimensional space, construct the following mapping on this
2481 * space
2483 * { [v] -> [v mod 2^width] }
2485 * where width is the number of bits used to represent the values
2486 * of the unsigned variable "iv".
2488 static __isl_give isl_map *compute_wrapping(__isl_take isl_space *dim,
2489 ValueDecl *iv)
2491 isl_int mod;
2492 isl_aff *aff;
2493 isl_map *map;
2495 isl_int_init(mod);
2496 isl_int_set_si(mod, 1);
2497 isl_int_mul_2exp(mod, mod, get_type_size(iv));
2499 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2500 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2501 aff = isl_aff_mod(aff, mod);
2503 isl_int_clear(mod);
2505 return isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2506 map = isl_map_reverse(map);
2509 /* Project out the parameter "id" from "set".
2511 static __isl_give isl_set *set_project_out_by_id(__isl_take isl_set *set,
2512 __isl_keep isl_id *id)
2514 int pos;
2516 pos = isl_set_find_dim_by_id(set, isl_dim_param, id);
2517 if (pos >= 0)
2518 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2520 return set;
2523 /* Compute the set of parameters for which "set1" is a subset of "set2".
2525 * set1 is a subset of set2 if
2527 * forall i in set1 : i in set2
2529 * or
2531 * not exists i in set1 and i not in set2
2533 * i.e.,
2535 * not exists i in set1 \ set2
2537 static __isl_give isl_set *enforce_subset(__isl_take isl_set *set1,
2538 __isl_take isl_set *set2)
2540 return isl_set_complement(isl_set_params(isl_set_subtract(set1, set2)));
2543 /* Compute the set of parameter values for which "cond" holds
2544 * on the next iteration for each element of "dom".
2546 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2547 * and then compute the set of parameters for which the result is a subset
2548 * of "cond".
2550 static __isl_give isl_set *valid_on_next(__isl_take isl_set *cond,
2551 __isl_take isl_set *dom, isl_int inc)
2553 isl_space *space;
2554 isl_aff *aff;
2555 isl_map *next;
2557 space = isl_set_get_space(dom);
2558 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2559 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2560 aff = isl_aff_add_constant(aff, inc);
2561 next = isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2563 dom = isl_set_apply(dom, next);
2565 return enforce_subset(dom, cond);
2568 /* Does "id" refer to a nested access?
2570 static bool is_nested_parameter(__isl_keep isl_id *id)
2572 return id && isl_id_get_user(id) && !isl_id_get_name(id);
2575 /* Does parameter "pos" of "space" refer to a nested access?
2577 static bool is_nested_parameter(__isl_keep isl_space *space, int pos)
2579 bool nested;
2580 isl_id *id;
2582 id = isl_space_get_dim_id(space, isl_dim_param, pos);
2583 nested = is_nested_parameter(id);
2584 isl_id_free(id);
2586 return nested;
2589 /* Does "space" involve any parameters that refer to nested
2590 * accesses, i.e., parameters with no name?
2592 static bool has_nested(__isl_keep isl_space *space)
2594 int nparam;
2596 nparam = isl_space_dim(space, isl_dim_param);
2597 for (int i = 0; i < nparam; ++i)
2598 if (is_nested_parameter(space, i))
2599 return true;
2601 return false;
2604 /* Does "pa" involve any parameters that refer to nested
2605 * accesses, i.e., parameters with no name?
2607 static bool has_nested(__isl_keep isl_pw_aff *pa)
2609 isl_space *space;
2610 bool nested;
2612 space = isl_pw_aff_get_space(pa);
2613 nested = has_nested(space);
2614 isl_space_free(space);
2616 return nested;
2619 /* Construct a pet_scop for a for statement.
2620 * The for loop is required to be of the form
2622 * for (i = init; condition; ++i)
2624 * or
2626 * for (i = init; condition; --i)
2628 * The initialization of the for loop should either be an assignment
2629 * to an integer variable, or a declaration of such a variable with
2630 * initialization.
2632 * The condition is allowed to contain nested accesses, provided
2633 * they are not being written to inside the body of the loop.
2634 * Otherwise, or if the condition is otherwise non-affine, the for loop is
2635 * essentially treated as a while loop, with iteration domain
2636 * { [i] : i >= init }.
2638 * We extract a pet_scop for the body and then embed it in a loop with
2639 * iteration domain and schedule
2641 * { [i] : i >= init and condition' }
2642 * { [i] -> [i] }
2644 * or
2646 * { [i] : i <= init and condition' }
2647 * { [i] -> [-i] }
2649 * Where condition' is equal to condition if the latter is
2650 * a simple upper [lower] bound and a condition that is extended
2651 * to apply to all previous iterations otherwise.
2653 * If the condition is non-affine, then we drop the condition from the
2654 * iteration domain and instead create a separate statement
2655 * for evaluating the condition. The body is then filtered to depend
2656 * on the result of the condition evaluating to true on all iterations
2657 * up to the current iteration, while the evaluation the condition itself
2658 * is filtered to depend on the result of the condition evaluating to true
2659 * on all previous iterations.
2660 * The context of the scop representing the body is dropped
2661 * because we don't know how many times the body will be executed,
2662 * if at all.
2664 * If the stride of the loop is not 1, then "i >= init" is replaced by
2666 * (exists a: i = init + stride * a and a >= 0)
2668 * If the loop iterator i is unsigned, then wrapping may occur.
2669 * During the computation, we work with a virtual iterator that
2670 * does not wrap. However, the condition in the code applies
2671 * to the wrapped value, so we need to change condition(i)
2672 * into condition([i % 2^width]).
2673 * After computing the virtual domain and schedule, we apply
2674 * the function { [v] -> [v % 2^width] } to the domain and the domain
2675 * of the schedule. In order not to lose any information, we also
2676 * need to intersect the domain of the schedule with the virtual domain
2677 * first, since some iterations in the wrapped domain may be scheduled
2678 * several times, typically an infinite number of times.
2679 * Note that there may be no need to perform this final wrapping
2680 * if the loop condition (after wrapping) satisfies certain conditions.
2681 * However, the is_simple_bound condition is not enough since it doesn't
2682 * check if there even is an upper bound.
2684 * If the loop condition is non-affine, then we keep the virtual
2685 * iterator in the iteration domain and instead replace all accesses
2686 * to the original iterator by the wrapping of the virtual iterator.
2688 * Wrapping on unsigned iterators can be avoided entirely if
2689 * loop condition is simple, the loop iterator is incremented
2690 * [decremented] by one and the last value before wrapping cannot
2691 * possibly satisfy the loop condition.
2693 * Before extracting a pet_scop from the body we remove all
2694 * assignments in assigned_value to variables that are assigned
2695 * somewhere in the body of the loop.
2697 * Valid parameters for a for loop are those for which the initial
2698 * value itself, the increment on each domain iteration and
2699 * the condition on both the initial value and
2700 * the result of incrementing the iterator for each iteration of the domain
2701 * can be evaluated.
2702 * If the loop condition is non-affine, then we only consider validity
2703 * of the initial value.
2705 * If the body contains any break, then we keep track of it in "skip"
2706 * (if the skip condition is affine) or it is handled in scop_add_break
2707 * (if the skip condition is not affine).
2708 * Note that the affine break condition needs to be considered with
2709 * respect to previous iterations in the virtual domain (if any)
2710 * and that the domain needs to be kept virtual if there is a non-affine
2711 * break condition.
2713 struct pet_scop *PetScan::extract_for(ForStmt *stmt)
2715 BinaryOperator *ass;
2716 Decl *decl;
2717 Stmt *init;
2718 Expr *lhs, *rhs;
2719 ValueDecl *iv;
2720 isl_space *space;
2721 isl_set *domain;
2722 isl_map *sched;
2723 isl_set *cond = NULL;
2724 isl_set *skip = NULL;
2725 isl_id *id;
2726 struct pet_scop *scop, *scop_cond = NULL;
2727 assigned_value_cache cache(assigned_value);
2728 isl_int inc;
2729 bool is_one;
2730 bool is_unsigned;
2731 bool is_simple;
2732 bool is_virtual;
2733 bool keep_virtual = false;
2734 bool has_affine_break;
2735 bool has_var_break;
2736 isl_map *wrap = NULL;
2737 isl_pw_aff *pa, *pa_inc, *init_val;
2738 isl_set *valid_init;
2739 isl_set *valid_cond;
2740 isl_set *valid_cond_init;
2741 isl_set *valid_cond_next;
2742 isl_set *valid_inc;
2743 isl_map *test_access = NULL, *break_access = NULL;
2744 int stmt_id;
2746 if (!stmt->getInit() && !stmt->getCond() && !stmt->getInc())
2747 return extract_infinite_for(stmt);
2749 init = stmt->getInit();
2750 if (!init) {
2751 unsupported(stmt);
2752 return NULL;
2754 if ((ass = initialization_assignment(init)) != NULL) {
2755 iv = extract_induction_variable(ass);
2756 if (!iv)
2757 return NULL;
2758 lhs = ass->getLHS();
2759 rhs = ass->getRHS();
2760 } else if ((decl = initialization_declaration(init)) != NULL) {
2761 VarDecl *var = extract_induction_variable(init, decl);
2762 if (!var)
2763 return NULL;
2764 iv = var;
2765 rhs = var->getInit();
2766 lhs = create_DeclRefExpr(var);
2767 } else {
2768 unsupported(stmt->getInit());
2769 return NULL;
2772 pa_inc = extract_increment(stmt, iv);
2773 if (!pa_inc)
2774 return NULL;
2776 isl_int_init(inc);
2777 if (isl_pw_aff_n_piece(pa_inc) != 1 ||
2778 isl_pw_aff_foreach_piece(pa_inc, &extract_cst, &inc) < 0) {
2779 isl_pw_aff_free(pa_inc);
2780 unsupported(stmt->getInc());
2781 isl_int_clear(inc);
2782 return NULL;
2784 valid_inc = isl_pw_aff_domain(pa_inc);
2786 is_unsigned = iv->getType()->isUnsignedIntegerType();
2788 assigned_value.erase(iv);
2789 clear_assignments clear(assigned_value);
2790 clear.TraverseStmt(stmt->getBody());
2792 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
2794 pa = try_extract_nested_condition(stmt->getCond());
2795 if (allow_nested && (!pa || has_nested(pa)))
2796 stmt_id = n_stmt++;
2798 scop = extract(stmt->getBody());
2800 has_affine_break = scop &&
2801 pet_scop_has_affine_skip(scop, pet_skip_later);
2802 if (has_affine_break) {
2803 skip = pet_scop_get_skip(scop, pet_skip_later);
2804 skip = isl_set_fix_si(skip, isl_dim_set, 0, 1);
2805 skip = isl_set_params(skip);
2807 has_var_break = scop && pet_scop_has_var_skip(scop, pet_skip_later);
2808 if (has_var_break) {
2809 break_access = pet_scop_get_skip_map(scop, pet_skip_later);
2810 keep_virtual = true;
2813 if (pa && !is_nested_allowed(pa, scop)) {
2814 isl_pw_aff_free(pa);
2815 pa = NULL;
2818 if (!allow_nested && !pa)
2819 pa = try_extract_affine_condition(stmt->getCond());
2820 valid_cond = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2821 cond = isl_pw_aff_non_zero_set(pa);
2822 if (allow_nested && !cond) {
2823 int save_n_stmt = n_stmt;
2824 test_access = create_test_access(ctx, n_test++);
2825 n_stmt = stmt_id;
2826 scop_cond = extract_non_affine_condition(stmt->getCond(),
2827 isl_map_copy(test_access));
2828 n_stmt = save_n_stmt;
2829 scop_cond = scop_add_array(scop_cond, test_access, ast_context);
2830 scop_cond = pet_scop_prefix(scop_cond, 0);
2831 scop = pet_scop_reset_context(scop);
2832 scop = pet_scop_prefix(scop, 1);
2833 keep_virtual = true;
2834 cond = isl_set_universe(isl_space_set_alloc(ctx, 0, 0));
2837 cond = embed(cond, isl_id_copy(id));
2838 skip = embed(skip, isl_id_copy(id));
2839 valid_cond = isl_set_coalesce(valid_cond);
2840 valid_cond = embed(valid_cond, isl_id_copy(id));
2841 valid_inc = embed(valid_inc, isl_id_copy(id));
2842 is_one = isl_int_is_one(inc) || isl_int_is_negone(inc);
2843 is_virtual = is_unsigned && (!is_one || can_wrap(cond, iv, inc));
2845 init_val = extract_affine(rhs);
2846 valid_cond_init = enforce_subset(
2847 isl_set_from_pw_aff(isl_pw_aff_copy(init_val)),
2848 isl_set_copy(valid_cond));
2849 if (is_one && !is_virtual) {
2850 isl_pw_aff_free(init_val);
2851 pa = extract_comparison(isl_int_is_pos(inc) ? BO_GE : BO_LE,
2852 lhs, rhs, init);
2853 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2854 valid_init = set_project_out_by_id(valid_init, id);
2855 domain = isl_pw_aff_non_zero_set(pa);
2856 } else {
2857 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(init_val));
2858 domain = strided_domain(isl_id_copy(id), init_val, inc);
2861 domain = embed(domain, isl_id_copy(id));
2862 if (is_virtual) {
2863 isl_map *rev_wrap;
2864 wrap = compute_wrapping(isl_set_get_space(cond), iv);
2865 rev_wrap = isl_map_reverse(isl_map_copy(wrap));
2866 cond = isl_set_apply(cond, isl_map_copy(rev_wrap));
2867 skip = isl_set_apply(skip, isl_map_copy(rev_wrap));
2868 valid_cond = isl_set_apply(valid_cond, isl_map_copy(rev_wrap));
2869 valid_inc = isl_set_apply(valid_inc, rev_wrap);
2871 is_simple = is_simple_bound(cond, inc);
2872 if (!is_simple) {
2873 cond = isl_set_gist(cond, isl_set_copy(domain));
2874 is_simple = is_simple_bound(cond, inc);
2876 if (!is_simple)
2877 cond = valid_for_each_iteration(cond,
2878 isl_set_copy(domain), inc);
2879 domain = isl_set_intersect(domain, cond);
2880 if (has_affine_break) {
2881 skip = isl_set_intersect(skip , isl_set_copy(domain));
2882 skip = after(skip, isl_int_sgn(inc));
2883 domain = isl_set_subtract(domain, skip);
2885 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, isl_id_copy(id));
2886 space = isl_space_from_domain(isl_set_get_space(domain));
2887 space = isl_space_add_dims(space, isl_dim_out, 1);
2888 sched = isl_map_universe(space);
2889 if (isl_int_is_pos(inc))
2890 sched = isl_map_equate(sched, isl_dim_in, 0, isl_dim_out, 0);
2891 else
2892 sched = isl_map_oppose(sched, isl_dim_in, 0, isl_dim_out, 0);
2894 valid_cond_next = valid_on_next(valid_cond, isl_set_copy(domain), inc);
2895 valid_inc = enforce_subset(isl_set_copy(domain), valid_inc);
2897 if (is_virtual && !keep_virtual) {
2898 wrap = isl_map_set_dim_id(wrap,
2899 isl_dim_out, 0, isl_id_copy(id));
2900 sched = isl_map_intersect_domain(sched, isl_set_copy(domain));
2901 domain = isl_set_apply(domain, isl_map_copy(wrap));
2902 sched = isl_map_apply_domain(sched, wrap);
2904 if (!(is_virtual && keep_virtual)) {
2905 space = isl_set_get_space(domain);
2906 wrap = isl_map_identity(isl_space_map_from_set(space));
2909 scop_cond = pet_scop_embed(scop_cond, isl_set_copy(domain),
2910 isl_map_copy(sched), isl_map_copy(wrap), isl_id_copy(id));
2911 scop = pet_scop_embed(scop, isl_set_copy(domain), sched, wrap, id);
2912 scop = resolve_nested(scop);
2913 if (has_var_break)
2914 scop = scop_add_break(scop, break_access, isl_set_copy(domain),
2915 isl_int_sgn(inc));
2916 if (test_access) {
2917 scop = scop_add_while(scop_cond, scop, test_access, domain,
2918 isl_int_sgn(inc));
2919 isl_set_free(valid_inc);
2920 } else {
2921 scop = pet_scop_restrict_context(scop, valid_inc);
2922 scop = pet_scop_restrict_context(scop, valid_cond_next);
2923 scop = pet_scop_restrict_context(scop, valid_cond_init);
2924 isl_set_free(domain);
2926 clear_assignment(assigned_value, iv);
2928 isl_int_clear(inc);
2930 scop = pet_scop_restrict_context(scop, valid_init);
2932 return scop;
2935 struct pet_scop *PetScan::extract(CompoundStmt *stmt, bool skip_declarations)
2937 return extract(stmt->children(), true, skip_declarations);
2940 /* Does parameter "pos" of "map" refer to a nested access?
2942 static bool is_nested_parameter(__isl_keep isl_map *map, int pos)
2944 bool nested;
2945 isl_id *id;
2947 id = isl_map_get_dim_id(map, isl_dim_param, pos);
2948 nested = is_nested_parameter(id);
2949 isl_id_free(id);
2951 return nested;
2954 /* How many parameters of "space" refer to nested accesses, i.e., have no name?
2956 static int n_nested_parameter(__isl_keep isl_space *space)
2958 int n = 0;
2959 int nparam;
2961 nparam = isl_space_dim(space, isl_dim_param);
2962 for (int i = 0; i < nparam; ++i)
2963 if (is_nested_parameter(space, i))
2964 ++n;
2966 return n;
2969 /* How many parameters of "map" refer to nested accesses, i.e., have no name?
2971 static int n_nested_parameter(__isl_keep isl_map *map)
2973 isl_space *space;
2974 int n;
2976 space = isl_map_get_space(map);
2977 n = n_nested_parameter(space);
2978 isl_space_free(space);
2980 return n;
2983 /* For each nested access parameter in "space",
2984 * construct a corresponding pet_expr, place it in args and
2985 * record its position in "param2pos".
2986 * "n_arg" is the number of elements that are already in args.
2987 * The position recorded in "param2pos" takes this number into account.
2988 * If the pet_expr corresponding to a parameter is identical to
2989 * the pet_expr corresponding to an earlier parameter, then these two
2990 * parameters are made to refer to the same element in args.
2992 * Return the final number of elements in args or -1 if an error has occurred.
2994 int PetScan::extract_nested(__isl_keep isl_space *space,
2995 int n_arg, struct pet_expr **args, std::map<int,int> &param2pos)
2997 int nparam;
2999 nparam = isl_space_dim(space, isl_dim_param);
3000 for (int i = 0; i < nparam; ++i) {
3001 int j;
3002 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
3003 Expr *nested;
3005 if (!is_nested_parameter(id)) {
3006 isl_id_free(id);
3007 continue;
3010 nested = (Expr *) isl_id_get_user(id);
3011 args[n_arg] = extract_expr(nested);
3012 if (!args[n_arg])
3013 return -1;
3015 for (j = 0; j < n_arg; ++j)
3016 if (pet_expr_is_equal(args[j], args[n_arg]))
3017 break;
3019 if (j < n_arg) {
3020 pet_expr_free(args[n_arg]);
3021 args[n_arg] = NULL;
3022 param2pos[i] = j;
3023 } else
3024 param2pos[i] = n_arg++;
3026 isl_id_free(id);
3029 return n_arg;
3032 /* For each nested access parameter in the access relations in "expr",
3033 * construct a corresponding pet_expr, place it in expr->args and
3034 * record its position in "param2pos".
3035 * n is the number of nested access parameters.
3037 struct pet_expr *PetScan::extract_nested(struct pet_expr *expr, int n,
3038 std::map<int,int> &param2pos)
3040 isl_space *space;
3042 expr->args = isl_calloc_array(ctx, struct pet_expr *, n);
3043 expr->n_arg = n;
3044 if (!expr->args)
3045 goto error;
3047 space = isl_map_get_space(expr->acc.access);
3048 n = extract_nested(space, 0, expr->args, param2pos);
3049 isl_space_free(space);
3051 if (n < 0)
3052 goto error;
3054 expr->n_arg = n;
3055 return expr;
3056 error:
3057 pet_expr_free(expr);
3058 return NULL;
3061 /* Look for parameters in any access relation in "expr" that
3062 * refer to nested accesses. In particular, these are
3063 * parameters with no name.
3065 * If there are any such parameters, then the domain of the access
3066 * relation, which is still [] at this point, is replaced by
3067 * [[] -> [t_1,...,t_n]], with n the number of these parameters
3068 * (after identifying identical nested accesses).
3069 * The parameters are then equated to the corresponding t dimensions
3070 * and subsequently projected out.
3071 * param2pos maps the position of the parameter to the position
3072 * of the corresponding t dimension.
3074 struct pet_expr *PetScan::resolve_nested(struct pet_expr *expr)
3076 int n;
3077 int nparam;
3078 int n_in;
3079 isl_space *dim;
3080 isl_map *map;
3081 std::map<int,int> param2pos;
3083 if (!expr)
3084 return expr;
3086 for (int i = 0; i < expr->n_arg; ++i) {
3087 expr->args[i] = resolve_nested(expr->args[i]);
3088 if (!expr->args[i]) {
3089 pet_expr_free(expr);
3090 return NULL;
3094 if (expr->type != pet_expr_access)
3095 return expr;
3097 n = n_nested_parameter(expr->acc.access);
3098 if (n == 0)
3099 return expr;
3101 expr = extract_nested(expr, n, param2pos);
3102 if (!expr)
3103 return NULL;
3105 n = expr->n_arg;
3106 nparam = isl_map_dim(expr->acc.access, isl_dim_param);
3107 n_in = isl_map_dim(expr->acc.access, isl_dim_in);
3108 dim = isl_map_get_space(expr->acc.access);
3109 dim = isl_space_domain(dim);
3110 dim = isl_space_from_domain(dim);
3111 dim = isl_space_add_dims(dim, isl_dim_out, n);
3112 map = isl_map_universe(dim);
3113 map = isl_map_domain_map(map);
3114 map = isl_map_reverse(map);
3115 expr->acc.access = isl_map_apply_domain(expr->acc.access, map);
3117 for (int i = nparam - 1; i >= 0; --i) {
3118 isl_id *id = isl_map_get_dim_id(expr->acc.access,
3119 isl_dim_param, i);
3120 if (!is_nested_parameter(id)) {
3121 isl_id_free(id);
3122 continue;
3125 expr->acc.access = isl_map_equate(expr->acc.access,
3126 isl_dim_param, i, isl_dim_in,
3127 n_in + param2pos[i]);
3128 expr->acc.access = isl_map_project_out(expr->acc.access,
3129 isl_dim_param, i, 1);
3131 isl_id_free(id);
3134 return expr;
3135 error:
3136 pet_expr_free(expr);
3137 return NULL;
3140 /* Convert a top-level pet_expr to a pet_scop with one statement.
3141 * This mainly involves resolving nested expression parameters
3142 * and setting the name of the iteration space.
3143 * The name is given by "label" if it is non-NULL. Otherwise,
3144 * it is of the form S_<n_stmt>.
3146 struct pet_scop *PetScan::extract(Stmt *stmt, struct pet_expr *expr,
3147 __isl_take isl_id *label)
3149 struct pet_stmt *ps;
3150 SourceLocation loc = stmt->getLocStart();
3151 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3153 expr = resolve_nested(expr);
3154 ps = pet_stmt_from_pet_expr(ctx, line, label, n_stmt++, expr);
3155 return pet_scop_from_pet_stmt(ctx, ps);
3158 /* Check if we can extract an affine expression from "expr".
3159 * Return the expressions as an isl_pw_aff if we can and NULL otherwise.
3160 * We turn on autodetection so that we won't generate any warnings
3161 * and turn off nesting, so that we won't accept any non-affine constructs.
3163 __isl_give isl_pw_aff *PetScan::try_extract_affine(Expr *expr)
3165 isl_pw_aff *pwaff;
3166 int save_autodetect = options->autodetect;
3167 bool save_nesting = nesting_enabled;
3169 options->autodetect = 1;
3170 nesting_enabled = false;
3172 pwaff = extract_affine(expr);
3174 options->autodetect = save_autodetect;
3175 nesting_enabled = save_nesting;
3177 return pwaff;
3180 /* Check whether "expr" is an affine expression.
3182 bool PetScan::is_affine(Expr *expr)
3184 isl_pw_aff *pwaff;
3186 pwaff = try_extract_affine(expr);
3187 isl_pw_aff_free(pwaff);
3189 return pwaff != NULL;
3192 /* Check if we can extract an affine constraint from "expr".
3193 * Return the constraint as an isl_set if we can and NULL otherwise.
3194 * We turn on autodetection so that we won't generate any warnings
3195 * and turn off nesting, so that we won't accept any non-affine constructs.
3197 __isl_give isl_pw_aff *PetScan::try_extract_affine_condition(Expr *expr)
3199 isl_pw_aff *cond;
3200 int save_autodetect = options->autodetect;
3201 bool save_nesting = nesting_enabled;
3203 options->autodetect = 1;
3204 nesting_enabled = false;
3206 cond = extract_condition(expr);
3208 options->autodetect = save_autodetect;
3209 nesting_enabled = save_nesting;
3211 return cond;
3214 /* Check whether "expr" is an affine constraint.
3216 bool PetScan::is_affine_condition(Expr *expr)
3218 isl_pw_aff *cond;
3220 cond = try_extract_affine_condition(expr);
3221 isl_pw_aff_free(cond);
3223 return cond != NULL;
3226 /* Check if we can extract a condition from "expr".
3227 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
3228 * If allow_nested is set, then the condition may involve parameters
3229 * corresponding to nested accesses.
3230 * We turn on autodetection so that we won't generate any warnings.
3232 __isl_give isl_pw_aff *PetScan::try_extract_nested_condition(Expr *expr)
3234 isl_pw_aff *cond;
3235 int save_autodetect = options->autodetect;
3236 bool save_nesting = nesting_enabled;
3238 options->autodetect = 1;
3239 nesting_enabled = allow_nested;
3240 cond = extract_condition(expr);
3242 options->autodetect = save_autodetect;
3243 nesting_enabled = save_nesting;
3245 return cond;
3248 /* If the top-level expression of "stmt" is an assignment, then
3249 * return that assignment as a BinaryOperator.
3250 * Otherwise return NULL.
3252 static BinaryOperator *top_assignment_or_null(Stmt *stmt)
3254 BinaryOperator *ass;
3256 if (!stmt)
3257 return NULL;
3258 if (stmt->getStmtClass() != Stmt::BinaryOperatorClass)
3259 return NULL;
3261 ass = cast<BinaryOperator>(stmt);
3262 if(ass->getOpcode() != BO_Assign)
3263 return NULL;
3265 return ass;
3268 /* Check if the given if statement is a conditional assignement
3269 * with a non-affine condition. If so, construct a pet_scop
3270 * corresponding to this conditional assignment. Otherwise return NULL.
3272 * In particular we check if "stmt" is of the form
3274 * if (condition)
3275 * a = f(...);
3276 * else
3277 * a = g(...);
3279 * where a is some array or scalar access.
3280 * The constructed pet_scop then corresponds to the expression
3282 * a = condition ? f(...) : g(...)
3284 * All access relations in f(...) are intersected with condition
3285 * while all access relation in g(...) are intersected with the complement.
3287 struct pet_scop *PetScan::extract_conditional_assignment(IfStmt *stmt)
3289 BinaryOperator *ass_then, *ass_else;
3290 isl_map *write_then, *write_else;
3291 isl_set *cond, *comp;
3292 isl_map *map;
3293 isl_pw_aff *pa;
3294 int equal;
3295 struct pet_expr *pe_cond, *pe_then, *pe_else, *pe, *pe_write;
3296 bool save_nesting = nesting_enabled;
3298 if (!options->detect_conditional_assignment)
3299 return NULL;
3301 ass_then = top_assignment_or_null(stmt->getThen());
3302 ass_else = top_assignment_or_null(stmt->getElse());
3304 if (!ass_then || !ass_else)
3305 return NULL;
3307 if (is_affine_condition(stmt->getCond()))
3308 return NULL;
3310 write_then = extract_access(ass_then->getLHS());
3311 write_else = extract_access(ass_else->getLHS());
3313 equal = isl_map_is_equal(write_then, write_else);
3314 isl_map_free(write_else);
3315 if (equal < 0 || !equal) {
3316 isl_map_free(write_then);
3317 return NULL;
3320 nesting_enabled = allow_nested;
3321 pa = extract_condition(stmt->getCond());
3322 nesting_enabled = save_nesting;
3323 cond = isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa));
3324 comp = isl_pw_aff_zero_set(isl_pw_aff_copy(pa));
3325 map = isl_map_from_range(isl_set_from_pw_aff(pa));
3327 pe_cond = pet_expr_from_access(map);
3329 pe_then = extract_expr(ass_then->getRHS());
3330 pe_then = pet_expr_restrict(pe_then, cond);
3331 pe_else = extract_expr(ass_else->getRHS());
3332 pe_else = pet_expr_restrict(pe_else, comp);
3334 pe = pet_expr_new_ternary(ctx, pe_cond, pe_then, pe_else);
3335 pe_write = pet_expr_from_access(write_then);
3336 if (pe_write) {
3337 pe_write->acc.write = 1;
3338 pe_write->acc.read = 0;
3340 pe = pet_expr_new_binary(ctx, pet_op_assign, pe_write, pe);
3341 return extract(stmt, pe);
3344 /* Create a pet_scop with a single statement evaluating "cond"
3345 * and writing the result to a virtual scalar, as expressed by
3346 * "access".
3348 struct pet_scop *PetScan::extract_non_affine_condition(Expr *cond,
3349 __isl_take isl_map *access)
3351 struct pet_expr *expr, *write;
3352 struct pet_stmt *ps;
3353 struct pet_scop *scop;
3354 SourceLocation loc = cond->getLocStart();
3355 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3357 write = pet_expr_from_access(access);
3358 if (write) {
3359 write->acc.write = 1;
3360 write->acc.read = 0;
3362 expr = extract_expr(cond);
3363 expr = resolve_nested(expr);
3364 expr = pet_expr_new_binary(ctx, pet_op_assign, write, expr);
3365 ps = pet_stmt_from_pet_expr(ctx, line, NULL, n_stmt++, expr);
3366 scop = pet_scop_from_pet_stmt(ctx, ps);
3367 scop = resolve_nested(scop);
3369 return scop;
3372 extern "C" {
3373 static __isl_give isl_map *embed_access(__isl_take isl_map *access,
3374 void *user);
3377 /* Apply the map pointed to by "user" to the domain of the access
3378 * relation, thereby embedding it in the range of the map.
3379 * The domain of both relations is the zero-dimensional domain.
3381 static __isl_give isl_map *embed_access(__isl_take isl_map *access, void *user)
3383 isl_map *map = (isl_map *) user;
3385 return isl_map_apply_domain(access, isl_map_copy(map));
3388 /* Apply "map" to all access relations in "expr".
3390 static struct pet_expr *embed(struct pet_expr *expr, __isl_keep isl_map *map)
3392 return pet_expr_foreach_access(expr, &embed_access, map);
3395 /* How many parameters of "set" refer to nested accesses, i.e., have no name?
3397 static int n_nested_parameter(__isl_keep isl_set *set)
3399 isl_space *space;
3400 int n;
3402 space = isl_set_get_space(set);
3403 n = n_nested_parameter(space);
3404 isl_space_free(space);
3406 return n;
3409 /* Remove all parameters from "map" that refer to nested accesses.
3411 static __isl_give isl_map *remove_nested_parameters(__isl_take isl_map *map)
3413 int nparam;
3414 isl_space *space;
3416 space = isl_map_get_space(map);
3417 nparam = isl_space_dim(space, isl_dim_param);
3418 for (int i = nparam - 1; i >= 0; --i)
3419 if (is_nested_parameter(space, i))
3420 map = isl_map_project_out(map, isl_dim_param, i, 1);
3421 isl_space_free(space);
3423 return map;
3426 extern "C" {
3427 static __isl_give isl_map *access_remove_nested_parameters(
3428 __isl_take isl_map *access, void *user);
3431 static __isl_give isl_map *access_remove_nested_parameters(
3432 __isl_take isl_map *access, void *user)
3434 return remove_nested_parameters(access);
3437 /* Remove all nested access parameters from the schedule and all
3438 * accesses of "stmt".
3439 * There is no need to remove them from the domain as these parameters
3440 * have already been removed from the domain when this function is called.
3442 static struct pet_stmt *remove_nested_parameters(struct pet_stmt *stmt)
3444 if (!stmt)
3445 return NULL;
3446 stmt->schedule = remove_nested_parameters(stmt->schedule);
3447 stmt->body = pet_expr_foreach_access(stmt->body,
3448 &access_remove_nested_parameters, NULL);
3449 if (!stmt->schedule || !stmt->body)
3450 goto error;
3451 for (int i = 0; i < stmt->n_arg; ++i) {
3452 stmt->args[i] = pet_expr_foreach_access(stmt->args[i],
3453 &access_remove_nested_parameters, NULL);
3454 if (!stmt->args[i])
3455 goto error;
3458 return stmt;
3459 error:
3460 pet_stmt_free(stmt);
3461 return NULL;
3464 /* For each nested access parameter in the domain of "stmt",
3465 * construct a corresponding pet_expr, place it before the original
3466 * elements in stmt->args and record its position in "param2pos".
3467 * n is the number of nested access parameters.
3469 struct pet_stmt *PetScan::extract_nested(struct pet_stmt *stmt, int n,
3470 std::map<int,int> &param2pos)
3472 int i;
3473 isl_space *space;
3474 int n_arg;
3475 struct pet_expr **args;
3477 n_arg = stmt->n_arg;
3478 args = isl_calloc_array(ctx, struct pet_expr *, n + n_arg);
3479 if (!args)
3480 goto error;
3482 space = isl_set_get_space(stmt->domain);
3483 n_arg = extract_nested(space, 0, args, param2pos);
3484 isl_space_free(space);
3486 if (n_arg < 0)
3487 goto error;
3489 for (i = 0; i < stmt->n_arg; ++i)
3490 args[n_arg + i] = stmt->args[i];
3491 free(stmt->args);
3492 stmt->args = args;
3493 stmt->n_arg += n_arg;
3495 return stmt;
3496 error:
3497 if (args) {
3498 for (i = 0; i < n; ++i)
3499 pet_expr_free(args[i]);
3500 free(args);
3502 pet_stmt_free(stmt);
3503 return NULL;
3506 /* Check whether any of the arguments i of "stmt" starting at position "n"
3507 * is equal to one of the first "n" arguments j.
3508 * If so, combine the constraints on arguments i and j and remove
3509 * argument i.
3511 static struct pet_stmt *remove_duplicate_arguments(struct pet_stmt *stmt, int n)
3513 int i, j;
3514 isl_map *map;
3516 if (!stmt)
3517 return NULL;
3518 if (n == 0)
3519 return stmt;
3520 if (n == stmt->n_arg)
3521 return stmt;
3523 map = isl_set_unwrap(stmt->domain);
3525 for (i = stmt->n_arg - 1; i >= n; --i) {
3526 for (j = 0; j < n; ++j)
3527 if (pet_expr_is_equal(stmt->args[i], stmt->args[j]))
3528 break;
3529 if (j >= n)
3530 continue;
3532 map = isl_map_equate(map, isl_dim_out, i, isl_dim_out, j);
3533 map = isl_map_project_out(map, isl_dim_out, i, 1);
3535 pet_expr_free(stmt->args[i]);
3536 for (j = i; j + 1 < stmt->n_arg; ++j)
3537 stmt->args[j] = stmt->args[j + 1];
3538 stmt->n_arg--;
3541 stmt->domain = isl_map_wrap(map);
3542 if (!stmt->domain)
3543 goto error;
3544 return stmt;
3545 error:
3546 pet_stmt_free(stmt);
3547 return NULL;
3550 /* Look for parameters in the iteration domain of "stmt" that
3551 * refer to nested accesses. In particular, these are
3552 * parameters with no name.
3554 * If there are any such parameters, then as many extra variables
3555 * (after identifying identical nested accesses) are inserted in the
3556 * range of the map wrapped inside the domain, before the original variables.
3557 * If the original domain is not a wrapped map, then a new wrapped
3558 * map is created with zero output dimensions.
3559 * The parameters are then equated to the corresponding output dimensions
3560 * and subsequently projected out, from the iteration domain,
3561 * the schedule and the access relations.
3562 * For each of the output dimensions, a corresponding argument
3563 * expression is inserted. Initially they are created with
3564 * a zero-dimensional domain, so they have to be embedded
3565 * in the current iteration domain.
3566 * param2pos maps the position of the parameter to the position
3567 * of the corresponding output dimension in the wrapped map.
3569 struct pet_stmt *PetScan::resolve_nested(struct pet_stmt *stmt)
3571 int n;
3572 int nparam;
3573 unsigned n_arg;
3574 isl_map *map;
3575 std::map<int,int> param2pos;
3577 if (!stmt)
3578 return NULL;
3580 n = n_nested_parameter(stmt->domain);
3581 if (n == 0)
3582 return stmt;
3584 n_arg = stmt->n_arg;
3585 stmt = extract_nested(stmt, n, param2pos);
3586 if (!stmt)
3587 return NULL;
3589 n = stmt->n_arg - n_arg;
3590 nparam = isl_set_dim(stmt->domain, isl_dim_param);
3591 if (isl_set_is_wrapping(stmt->domain))
3592 map = isl_set_unwrap(stmt->domain);
3593 else
3594 map = isl_map_from_domain(stmt->domain);
3595 map = isl_map_insert_dims(map, isl_dim_out, 0, n);
3597 for (int i = nparam - 1; i >= 0; --i) {
3598 isl_id *id;
3600 if (!is_nested_parameter(map, i))
3601 continue;
3603 id = isl_map_get_tuple_id(stmt->args[param2pos[i]]->acc.access,
3604 isl_dim_out);
3605 map = isl_map_set_dim_id(map, isl_dim_out, param2pos[i], id);
3606 map = isl_map_equate(map, isl_dim_param, i, isl_dim_out,
3607 param2pos[i]);
3608 map = isl_map_project_out(map, isl_dim_param, i, 1);
3611 stmt->domain = isl_map_wrap(map);
3613 map = isl_set_unwrap(isl_set_copy(stmt->domain));
3614 map = isl_map_from_range(isl_map_domain(map));
3615 for (int pos = 0; pos < n; ++pos)
3616 stmt->args[pos] = embed(stmt->args[pos], map);
3617 isl_map_free(map);
3619 stmt = remove_nested_parameters(stmt);
3620 stmt = remove_duplicate_arguments(stmt, n);
3622 return stmt;
3623 error:
3624 pet_stmt_free(stmt);
3625 return NULL;
3628 /* For each statement in "scop", move the parameters that correspond
3629 * to nested access into the ranges of the domains and create
3630 * corresponding argument expressions.
3632 struct pet_scop *PetScan::resolve_nested(struct pet_scop *scop)
3634 if (!scop)
3635 return NULL;
3637 for (int i = 0; i < scop->n_stmt; ++i) {
3638 scop->stmts[i] = resolve_nested(scop->stmts[i]);
3639 if (!scop->stmts[i])
3640 goto error;
3643 return scop;
3644 error:
3645 pet_scop_free(scop);
3646 return NULL;
3649 /* Given an access expression "expr", is the variable accessed by
3650 * "expr" assigned anywhere inside "scop"?
3652 static bool is_assigned(pet_expr *expr, pet_scop *scop)
3654 bool assigned = false;
3655 isl_id *id;
3657 id = isl_map_get_tuple_id(expr->acc.access, isl_dim_out);
3658 assigned = pet_scop_writes(scop, id);
3659 isl_id_free(id);
3661 return assigned;
3664 /* Are all nested access parameters in "pa" allowed given "scop".
3665 * In particular, is none of them written by anywhere inside "scop".
3667 * If "scop" has any skip conditions, then no nested access parameters
3668 * are allowed. In particular, if there is any nested access in a guard
3669 * for a piece of code containing a "continue", then we want to introduce
3670 * a separate statement for evaluating this guard so that we can express
3671 * that the result is false for all previous iterations.
3673 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff *pa, pet_scop *scop)
3675 int nparam;
3677 if (!scop)
3678 return true;
3680 nparam = isl_pw_aff_dim(pa, isl_dim_param);
3681 for (int i = 0; i < nparam; ++i) {
3682 Expr *nested;
3683 isl_id *id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
3684 pet_expr *expr;
3685 bool allowed;
3687 if (!is_nested_parameter(id)) {
3688 isl_id_free(id);
3689 continue;
3692 if (pet_scop_has_skip(scop, pet_skip_now)) {
3693 isl_id_free(id);
3694 return false;
3697 nested = (Expr *) isl_id_get_user(id);
3698 expr = extract_expr(nested);
3699 allowed = expr && expr->type == pet_expr_access &&
3700 !is_assigned(expr, scop);
3702 pet_expr_free(expr);
3703 isl_id_free(id);
3705 if (!allowed)
3706 return false;
3709 return true;
3712 /* Do we need to construct a skip condition of the given type
3713 * on an if statement, given that the if condition is non-affine?
3715 * pet_scop_filter_skip can only handle the case where the if condition
3716 * holds (the then branch) and the skip condition is universal.
3717 * In any other case, we need to construct a new skip condition.
3719 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
3720 bool have_else, enum pet_skip type)
3722 if (have_else && scop_else && pet_scop_has_skip(scop_else, type))
3723 return true;
3724 if (scop_then && pet_scop_has_skip(scop_then, type) &&
3725 !pet_scop_has_universal_skip(scop_then, type))
3726 return true;
3727 return false;
3730 /* Do we need to construct a skip condition of the given type
3731 * on an if statement, given that the if condition is affine?
3733 * There is no need to construct a new skip condition if all
3734 * the skip conditions are affine.
3736 static bool need_skip_aff(struct pet_scop *scop_then,
3737 struct pet_scop *scop_else, bool have_else, enum pet_skip type)
3739 if (scop_then && pet_scop_has_var_skip(scop_then, type))
3740 return true;
3741 if (have_else && scop_else && pet_scop_has_var_skip(scop_else, type))
3742 return true;
3743 return false;
3746 /* Do we need to construct a skip condition of the given type
3747 * on an if statement?
3749 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
3750 bool have_else, enum pet_skip type, bool affine)
3752 if (affine)
3753 return need_skip_aff(scop_then, scop_else, have_else, type);
3754 else
3755 return need_skip(scop_then, scop_else, have_else, type);
3758 /* Construct an affine expression pet_expr that is evaluates
3759 * to the constant "val".
3761 static struct pet_expr *universally(isl_ctx *ctx, int val)
3763 isl_space *space;
3764 isl_map *map;
3766 space = isl_space_alloc(ctx, 0, 0, 1);
3767 map = isl_map_universe(space);
3768 map = isl_map_fix_si(map, isl_dim_out, 0, val);
3770 return pet_expr_from_access(map);
3773 /* Construct an affine expression pet_expr that is evaluates
3774 * to the constant 1.
3776 static struct pet_expr *universally_true(isl_ctx *ctx)
3778 return universally(ctx, 1);
3781 /* Construct an affine expression pet_expr that is evaluates
3782 * to the constant 0.
3784 static struct pet_expr *universally_false(isl_ctx *ctx)
3786 return universally(ctx, 0);
3789 /* Given an access relation "test_access" for the if condition,
3790 * an access relation "skip_access" for the skip condition and
3791 * scops for the then and else branches, construct a scop for
3792 * computing "skip_access".
3794 * The computed scop contains a single statement that essentially does
3796 * skip_cond = test_cond ? skip_cond_then : skip_cond_else
3798 * If the skip conditions of the then and/or else branch are not affine,
3799 * then they need to be filtered by test_access.
3800 * If they are missing, then this means the skip condition is false.
3802 * Since we are constructing a skip condition for the if statement,
3803 * the skip conditions on the then and else branches are removed.
3805 static struct pet_scop *extract_skip(PetScan *scan,
3806 __isl_take isl_map *test_access, __isl_take isl_map *skip_access,
3807 struct pet_scop *scop_then, struct pet_scop *scop_else, bool have_else,
3808 enum pet_skip type)
3810 struct pet_expr *expr_then, *expr_else, *expr, *expr_skip;
3811 struct pet_stmt *stmt;
3812 struct pet_scop *scop;
3813 isl_ctx *ctx = scan->ctx;
3815 if (!scop_then)
3816 goto error;
3817 if (have_else && !scop_else)
3818 goto error;
3820 if (pet_scop_has_skip(scop_then, type)) {
3821 expr_then = pet_scop_get_skip_expr(scop_then, type);
3822 pet_scop_reset_skip(scop_then, type);
3823 if (!pet_expr_is_affine(expr_then))
3824 expr_then = pet_expr_filter(expr_then,
3825 isl_map_copy(test_access), 1);
3826 } else
3827 expr_then = universally_false(ctx);
3829 if (have_else && pet_scop_has_skip(scop_else, type)) {
3830 expr_else = pet_scop_get_skip_expr(scop_else, type);
3831 pet_scop_reset_skip(scop_else, type);
3832 if (!pet_expr_is_affine(expr_else))
3833 expr_else = pet_expr_filter(expr_else,
3834 isl_map_copy(test_access), 0);
3835 } else
3836 expr_else = universally_false(ctx);
3838 expr = pet_expr_from_access(test_access);
3839 expr = pet_expr_new_ternary(ctx, expr, expr_then, expr_else);
3840 expr_skip = pet_expr_from_access(isl_map_copy(skip_access));
3841 if (expr_skip) {
3842 expr_skip->acc.write = 1;
3843 expr_skip->acc.read = 0;
3845 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
3846 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, scan->n_stmt++, expr);
3848 scop = pet_scop_from_pet_stmt(ctx, stmt);
3849 scop = scop_add_array(scop, skip_access, scan->ast_context);
3850 isl_map_free(skip_access);
3852 return scop;
3853 error:
3854 isl_map_free(test_access);
3855 isl_map_free(skip_access);
3856 return NULL;
3859 /* Is scop's skip_now condition equal to its skip_later condition?
3860 * In particular, this means that it either has no skip_now condition
3861 * or both a skip_now and a skip_later condition (that are equal to each other).
3863 static bool skip_equals_skip_later(struct pet_scop *scop)
3865 int has_skip_now, has_skip_later;
3866 int equal;
3867 isl_set *skip_now, *skip_later;
3869 if (!scop)
3870 return false;
3871 has_skip_now = pet_scop_has_skip(scop, pet_skip_now);
3872 has_skip_later = pet_scop_has_skip(scop, pet_skip_later);
3873 if (has_skip_now != has_skip_later)
3874 return false;
3875 if (!has_skip_now)
3876 return true;
3878 skip_now = pet_scop_get_skip(scop, pet_skip_now);
3879 skip_later = pet_scop_get_skip(scop, pet_skip_later);
3880 equal = isl_set_is_equal(skip_now, skip_later);
3881 isl_set_free(skip_now);
3882 isl_set_free(skip_later);
3884 return equal;
3887 /* Drop the skip conditions of type pet_skip_later from scop1 and scop2.
3889 static void drop_skip_later(struct pet_scop *scop1, struct pet_scop *scop2)
3891 pet_scop_reset_skip(scop1, pet_skip_later);
3892 pet_scop_reset_skip(scop2, pet_skip_later);
3895 /* Structure that handles the construction of skip conditions.
3897 * scop_then and scop_else represent the then and else branches
3898 * of the if statement
3900 * skip[type] is true if we need to construct a skip condition of that type
3901 * equal is set if the skip conditions of types pet_skip_now and pet_skip_later
3902 * are equal to each other
3903 * access[type] is the virtual array representing the skip condition
3904 * scop[type] is a scop for computing the skip condition
3906 struct pet_skip_info {
3907 isl_ctx *ctx;
3909 bool skip[2];
3910 bool equal;
3911 isl_map *access[2];
3912 struct pet_scop *scop[2];
3914 pet_skip_info(isl_ctx *ctx) : ctx(ctx) {}
3916 operator bool() { return skip[pet_skip_now] || skip[pet_skip_later]; }
3919 /* Structure that handles the construction of skip conditions on if statements.
3921 * scop_then and scop_else represent the then and else branches
3922 * of the if statement
3924 struct pet_skip_info_if : public pet_skip_info {
3925 struct pet_scop *scop_then, *scop_else;
3926 bool have_else;
3928 pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
3929 struct pet_scop *scop_else, bool have_else, bool affine);
3930 void extract(PetScan *scan, __isl_keep isl_map *access,
3931 enum pet_skip type);
3932 void extract(PetScan *scan, __isl_keep isl_map *access);
3933 void extract(PetScan *scan, __isl_keep isl_pw_aff *cond);
3934 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
3935 int offset);
3936 struct pet_scop *add(struct pet_scop *scop, int offset);
3939 /* Initialize a pet_skip_info_if structure based on the then and else branches
3940 * and based on whether the if condition is affine or not.
3942 pet_skip_info_if::pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
3943 struct pet_scop *scop_else, bool have_else, bool affine) :
3944 pet_skip_info(ctx), scop_then(scop_then), scop_else(scop_else),
3945 have_else(have_else)
3947 skip[pet_skip_now] =
3948 need_skip(scop_then, scop_else, have_else, pet_skip_now, affine);
3949 equal = skip[pet_skip_now] && skip_equals_skip_later(scop_then) &&
3950 (!have_else || skip_equals_skip_later(scop_else));
3951 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
3952 need_skip(scop_then, scop_else, have_else, pet_skip_later, affine);
3955 /* If we need to construct a skip condition of the given type,
3956 * then do so now.
3958 * "map" represents the if condition.
3960 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_map *map,
3961 enum pet_skip type)
3963 if (!skip[type])
3964 return;
3966 access[type] = create_test_access(isl_map_get_ctx(map), scan->n_test++);
3967 scop[type] = extract_skip(scan, isl_map_copy(map),
3968 isl_map_copy(access[type]),
3969 scop_then, scop_else, have_else, type);
3972 /* Construct the required skip conditions, given the if condition "map".
3974 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_map *map)
3976 extract(scan, map, pet_skip_now);
3977 extract(scan, map, pet_skip_later);
3978 if (equal)
3979 drop_skip_later(scop_then, scop_else);
3982 /* Construct the required skip conditions, given the if condition "cond".
3984 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_pw_aff *cond)
3986 isl_set *test_set;
3987 isl_map *test;
3989 if (!skip[pet_skip_now] && !skip[pet_skip_later])
3990 return;
3992 test_set = isl_set_from_pw_aff(isl_pw_aff_copy(cond));
3993 test = isl_map_from_range(test_set);
3994 extract(scan, test);
3995 isl_map_free(test);
3998 /* Add the computed skip condition of the give type to "main" and
3999 * add the scop for computing the condition at the given offset.
4001 * If equal is set, then we only computed a skip condition for pet_skip_now,
4002 * but we also need to set it as main's pet_skip_later.
4004 struct pet_scop *pet_skip_info_if::add(struct pet_scop *main,
4005 enum pet_skip type, int offset)
4007 isl_set *skip_set;
4009 if (!skip[type])
4010 return main;
4012 skip_set = isl_map_range(access[type]);
4013 access[type] = NULL;
4014 scop[type] = pet_scop_prefix(scop[type], offset);
4015 main = pet_scop_add_par(ctx, main, scop[type]);
4016 scop[type] = NULL;
4018 if (equal)
4019 main = pet_scop_set_skip(main, pet_skip_later,
4020 isl_set_copy(skip_set));
4022 main = pet_scop_set_skip(main, type, skip_set);
4024 return main;
4027 /* Add the computed skip conditions to "main" and
4028 * add the scops for computing the conditions at the given offset.
4030 struct pet_scop *pet_skip_info_if::add(struct pet_scop *scop, int offset)
4032 scop = add(scop, pet_skip_now, offset);
4033 scop = add(scop, pet_skip_later, offset);
4035 return scop;
4038 /* Construct a pet_scop for a non-affine if statement.
4040 * We create a separate statement that writes the result
4041 * of the non-affine condition to a virtual scalar.
4042 * A constraint requiring the value of this virtual scalar to be one
4043 * is added to the iteration domains of the then branch.
4044 * Similarly, a constraint requiring the value of this virtual scalar
4045 * to be zero is added to the iteration domains of the else branch, if any.
4046 * We adjust the schedules to ensure that the virtual scalar is written
4047 * before it is read.
4049 * If there are any breaks or continues in the then and/or else
4050 * branches, then we may have to compute a new skip condition.
4051 * This is handled using a pet_skip_info_if object.
4052 * On initialization, the object checks if skip conditions need
4053 * to be computed. If so, it does so in "extract" and adds them in "add".
4055 struct pet_scop *PetScan::extract_non_affine_if(Expr *cond,
4056 struct pet_scop *scop_then, struct pet_scop *scop_else,
4057 bool have_else, int stmt_id)
4059 struct pet_scop *scop;
4060 isl_map *test_access;
4061 int save_n_stmt = n_stmt;
4063 test_access = create_test_access(ctx, n_test++);
4064 n_stmt = stmt_id;
4065 scop = extract_non_affine_condition(cond, isl_map_copy(test_access));
4066 n_stmt = save_n_stmt;
4067 scop = scop_add_array(scop, test_access, ast_context);
4069 pet_skip_info_if skip(ctx, scop_then, scop_else, have_else, false);
4070 skip.extract(this, test_access);
4072 scop = pet_scop_prefix(scop, 0);
4073 scop_then = pet_scop_prefix(scop_then, 1);
4074 scop_then = pet_scop_filter(scop_then, isl_map_copy(test_access), 1);
4075 if (have_else) {
4076 scop_else = pet_scop_prefix(scop_else, 1);
4077 scop_else = pet_scop_filter(scop_else, test_access, 0);
4078 scop_then = pet_scop_add_par(ctx, scop_then, scop_else);
4079 } else
4080 isl_map_free(test_access);
4082 scop = pet_scop_add_seq(ctx, scop, scop_then);
4084 scop = skip.add(scop, 2);
4086 return scop;
4089 /* Construct a pet_scop for an if statement.
4091 * If the condition fits the pattern of a conditional assignment,
4092 * then it is handled by extract_conditional_assignment.
4093 * Otherwise, we do the following.
4095 * If the condition is affine, then the condition is added
4096 * to the iteration domains of the then branch, while the
4097 * opposite of the condition in added to the iteration domains
4098 * of the else branch, if any.
4099 * We allow the condition to be dynamic, i.e., to refer to
4100 * scalars or array elements that may be written to outside
4101 * of the given if statement. These nested accesses are then represented
4102 * as output dimensions in the wrapping iteration domain.
4103 * If it also written _inside_ the then or else branch, then
4104 * we treat the condition as non-affine.
4105 * As explained in extract_non_affine_if, this will introduce
4106 * an extra statement.
4107 * For aesthetic reasons, we want this statement to have a statement
4108 * number that is lower than those of the then and else branches.
4109 * In order to evaluate if will need such a statement, however, we
4110 * first construct scops for the then and else branches.
4111 * We therefore reserve a statement number if we might have to
4112 * introduce such an extra statement.
4114 * If the condition is not affine, then the scop is created in
4115 * extract_non_affine_if.
4117 * If there are any breaks or continues in the then and/or else
4118 * branches, then we may have to compute a new skip condition.
4119 * This is handled using a pet_skip_info_if object.
4120 * On initialization, the object checks if skip conditions need
4121 * to be computed. If so, it does so in "extract" and adds them in "add".
4123 struct pet_scop *PetScan::extract(IfStmt *stmt)
4125 struct pet_scop *scop_then, *scop_else = NULL, *scop;
4126 isl_pw_aff *cond;
4127 int stmt_id;
4128 isl_set *set;
4129 isl_set *valid;
4131 scop = extract_conditional_assignment(stmt);
4132 if (scop)
4133 return scop;
4135 cond = try_extract_nested_condition(stmt->getCond());
4136 if (allow_nested && (!cond || has_nested(cond)))
4137 stmt_id = n_stmt++;
4140 assigned_value_cache cache(assigned_value);
4141 scop_then = extract(stmt->getThen());
4144 if (stmt->getElse()) {
4145 assigned_value_cache cache(assigned_value);
4146 scop_else = extract(stmt->getElse());
4147 if (options->autodetect) {
4148 if (scop_then && !scop_else) {
4149 partial = true;
4150 isl_pw_aff_free(cond);
4151 return scop_then;
4153 if (!scop_then && scop_else) {
4154 partial = true;
4155 isl_pw_aff_free(cond);
4156 return scop_else;
4161 if (cond &&
4162 (!is_nested_allowed(cond, scop_then) ||
4163 (stmt->getElse() && !is_nested_allowed(cond, scop_else)))) {
4164 isl_pw_aff_free(cond);
4165 cond = NULL;
4167 if (allow_nested && !cond)
4168 return extract_non_affine_if(stmt->getCond(), scop_then,
4169 scop_else, stmt->getElse(), stmt_id);
4171 if (!cond)
4172 cond = extract_condition(stmt->getCond());
4174 pet_skip_info_if skip(ctx, scop_then, scop_else, stmt->getElse(), true);
4175 skip.extract(this, cond);
4177 valid = isl_pw_aff_domain(isl_pw_aff_copy(cond));
4178 set = isl_pw_aff_non_zero_set(cond);
4179 scop = pet_scop_restrict(scop_then, isl_set_copy(set));
4181 if (stmt->getElse()) {
4182 set = isl_set_subtract(isl_set_copy(valid), set);
4183 scop_else = pet_scop_restrict(scop_else, set);
4184 scop = pet_scop_add_par(ctx, scop, scop_else);
4185 } else
4186 isl_set_free(set);
4187 scop = resolve_nested(scop);
4188 scop = pet_scop_restrict_context(scop, valid);
4190 if (skip)
4191 scop = pet_scop_prefix(scop, 0);
4192 scop = skip.add(scop, 1);
4194 return scop;
4197 /* Try and construct a pet_scop for a label statement.
4198 * We currently only allow labels on expression statements.
4200 struct pet_scop *PetScan::extract(LabelStmt *stmt)
4202 isl_id *label;
4203 Stmt *sub;
4205 sub = stmt->getSubStmt();
4206 if (!isa<Expr>(sub)) {
4207 unsupported(stmt);
4208 return NULL;
4211 label = isl_id_alloc(ctx, stmt->getName(), NULL);
4213 return extract(sub, extract_expr(cast<Expr>(sub)), label);
4216 /* Construct a pet_scop for a continue statement.
4218 * We simply create an empty scop with a universal pet_skip_now
4219 * skip condition. This skip condition will then be taken into
4220 * account by the enclosing loop construct, possibly after
4221 * being incorporated into outer skip conditions.
4223 struct pet_scop *PetScan::extract(ContinueStmt *stmt)
4225 pet_scop *scop;
4226 isl_space *space;
4227 isl_set *set;
4229 scop = pet_scop_empty(ctx);
4230 if (!scop)
4231 return NULL;
4233 space = isl_space_set_alloc(ctx, 0, 1);
4234 set = isl_set_universe(space);
4235 set = isl_set_fix_si(set, isl_dim_set, 0, 1);
4236 scop = pet_scop_set_skip(scop, pet_skip_now, set);
4238 return scop;
4241 /* Construct a pet_scop for a break statement.
4243 * We simply create an empty scop with both a universal pet_skip_now
4244 * skip condition and a universal pet_skip_later skip condition.
4245 * These skip conditions will then be taken into
4246 * account by the enclosing loop construct, possibly after
4247 * being incorporated into outer skip conditions.
4249 struct pet_scop *PetScan::extract(BreakStmt *stmt)
4251 pet_scop *scop;
4252 isl_space *space;
4253 isl_set *set;
4255 scop = pet_scop_empty(ctx);
4256 if (!scop)
4257 return NULL;
4259 space = isl_space_set_alloc(ctx, 0, 1);
4260 set = isl_set_universe(space);
4261 set = isl_set_fix_si(set, isl_dim_set, 0, 1);
4262 scop = pet_scop_set_skip(scop, pet_skip_now, isl_set_copy(set));
4263 scop = pet_scop_set_skip(scop, pet_skip_later, set);
4265 return scop;
4268 /* Try and construct a pet_scop corresponding to "stmt".
4270 * If "stmt" is a compound statement, then "skip_declarations"
4271 * indicates whether we should skip initial declarations in the
4272 * compound statement.
4274 struct pet_scop *PetScan::extract(Stmt *stmt, bool skip_declarations)
4276 if (isa<Expr>(stmt))
4277 return extract(stmt, extract_expr(cast<Expr>(stmt)));
4279 switch (stmt->getStmtClass()) {
4280 case Stmt::WhileStmtClass:
4281 return extract(cast<WhileStmt>(stmt));
4282 case Stmt::ForStmtClass:
4283 return extract_for(cast<ForStmt>(stmt));
4284 case Stmt::IfStmtClass:
4285 return extract(cast<IfStmt>(stmt));
4286 case Stmt::CompoundStmtClass:
4287 return extract(cast<CompoundStmt>(stmt), skip_declarations);
4288 case Stmt::LabelStmtClass:
4289 return extract(cast<LabelStmt>(stmt));
4290 case Stmt::ContinueStmtClass:
4291 return extract(cast<ContinueStmt>(stmt));
4292 case Stmt::BreakStmtClass:
4293 return extract(cast<BreakStmt>(stmt));
4294 case Stmt::DeclStmtClass:
4295 return extract(cast<DeclStmt>(stmt));
4296 default:
4297 unsupported(stmt);
4300 return NULL;
4303 /* Do we need to construct a skip condition of the given type
4304 * on a sequence of statements?
4306 * There is no need to construct a new skip condition if only
4307 * only of the two statements has a skip condition or if both
4308 * of their skip conditions are affine.
4310 * In principle we also don't need a new continuation variable if
4311 * the continuation of scop2 is affine, but then we would need
4312 * to allow more complicated forms of continuations.
4314 static bool need_skip_seq(struct pet_scop *scop1, struct pet_scop *scop2,
4315 enum pet_skip type)
4317 if (!scop1 || !pet_scop_has_skip(scop1, type))
4318 return false;
4319 if (!scop2 || !pet_scop_has_skip(scop2, type))
4320 return false;
4321 if (pet_scop_has_affine_skip(scop1, type) &&
4322 pet_scop_has_affine_skip(scop2, type))
4323 return false;
4324 return true;
4327 /* Construct a scop for computing the skip condition of the given type and
4328 * with access relation "skip_access" for a sequence of two scops "scop1"
4329 * and "scop2".
4331 * The computed scop contains a single statement that essentially does
4333 * skip_cond = skip_cond_1 ? 1 : skip_cond_2
4335 * or, in other words, skip_cond1 || skip_cond2.
4336 * In this expression, skip_cond_2 is filtered to reflect that it is
4337 * only evaluated when skip_cond_1 is false.
4339 * The skip condition on scop1 is not removed because it still needs
4340 * to be applied to scop2 when these two scops are combined.
4342 static struct pet_scop *extract_skip_seq(PetScan *ps,
4343 __isl_take isl_map *skip_access,
4344 struct pet_scop *scop1, struct pet_scop *scop2, enum pet_skip type)
4346 isl_map *access;
4347 struct pet_expr *expr1, *expr2, *expr, *expr_skip;
4348 struct pet_stmt *stmt;
4349 struct pet_scop *scop;
4350 isl_ctx *ctx = ps->ctx;
4352 if (!scop1 || !scop2)
4353 goto error;
4355 expr1 = pet_scop_get_skip_expr(scop1, type);
4356 expr2 = pet_scop_get_skip_expr(scop2, type);
4357 pet_scop_reset_skip(scop2, type);
4359 expr2 = pet_expr_filter(expr2, isl_map_copy(expr1->acc.access), 0);
4361 expr = universally_true(ctx);
4362 expr = pet_expr_new_ternary(ctx, expr1, expr, expr2);
4363 expr_skip = pet_expr_from_access(isl_map_copy(skip_access));
4364 if (expr_skip) {
4365 expr_skip->acc.write = 1;
4366 expr_skip->acc.read = 0;
4368 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
4369 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, ps->n_stmt++, expr);
4371 scop = pet_scop_from_pet_stmt(ctx, stmt);
4372 scop = scop_add_array(scop, skip_access, ps->ast_context);
4373 isl_map_free(skip_access);
4375 return scop;
4376 error:
4377 isl_map_free(skip_access);
4378 return NULL;
4381 /* Structure that handles the construction of skip conditions
4382 * on sequences of statements.
4384 * scop1 and scop2 represent the two statements that are combined
4386 struct pet_skip_info_seq : public pet_skip_info {
4387 struct pet_scop *scop1, *scop2;
4389 pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4390 struct pet_scop *scop2);
4391 void extract(PetScan *scan, enum pet_skip type);
4392 void extract(PetScan *scan);
4393 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
4394 int offset);
4395 struct pet_scop *add(struct pet_scop *scop, int offset);
4398 /* Initialize a pet_skip_info_seq structure based on
4399 * on the two statements that are going to be combined.
4401 pet_skip_info_seq::pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4402 struct pet_scop *scop2) : pet_skip_info(ctx), scop1(scop1), scop2(scop2)
4404 skip[pet_skip_now] = need_skip_seq(scop1, scop2, pet_skip_now);
4405 equal = skip[pet_skip_now] && skip_equals_skip_later(scop1) &&
4406 skip_equals_skip_later(scop2);
4407 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
4408 need_skip_seq(scop1, scop2, pet_skip_later);
4411 /* If we need to construct a skip condition of the given type,
4412 * then do so now.
4414 void pet_skip_info_seq::extract(PetScan *scan, enum pet_skip type)
4416 if (!skip[type])
4417 return;
4419 access[type] = create_test_access(ctx, scan->n_test++);
4420 scop[type] = extract_skip_seq(scan, isl_map_copy(access[type]),
4421 scop1, scop2, type);
4424 /* Construct the required skip conditions.
4426 void pet_skip_info_seq::extract(PetScan *scan)
4428 extract(scan, pet_skip_now);
4429 extract(scan, pet_skip_later);
4430 if (equal)
4431 drop_skip_later(scop1, scop2);
4434 /* Add the computed skip condition of the give type to "main" and
4435 * add the scop for computing the condition at the given offset (the statement
4436 * number). Within this offset, the condition is computed at position 1
4437 * to ensure that it is computed after the corresponding statement.
4439 * If equal is set, then we only computed a skip condition for pet_skip_now,
4440 * but we also need to set it as main's pet_skip_later.
4442 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *main,
4443 enum pet_skip type, int offset)
4445 isl_set *skip_set;
4447 if (!skip[type])
4448 return main;
4450 skip_set = isl_map_range(access[type]);
4451 access[type] = NULL;
4452 scop[type] = pet_scop_prefix(scop[type], 1);
4453 scop[type] = pet_scop_prefix(scop[type], offset);
4454 main = pet_scop_add_par(ctx, main, scop[type]);
4455 scop[type] = NULL;
4457 if (equal)
4458 main = pet_scop_set_skip(main, pet_skip_later,
4459 isl_set_copy(skip_set));
4461 main = pet_scop_set_skip(main, type, skip_set);
4463 return main;
4466 /* Add the computed skip conditions to "main" and
4467 * add the scops for computing the conditions at the given offset.
4469 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *scop, int offset)
4471 scop = add(scop, pet_skip_now, offset);
4472 scop = add(scop, pet_skip_later, offset);
4474 return scop;
4477 /* Extract a clone of the kill statement in "scop".
4478 * "scop" is expected to have been created from a DeclStmt
4479 * and should have the kill as its first statement.
4481 struct pet_stmt *PetScan::extract_kill(struct pet_scop *scop)
4483 struct pet_expr *kill;
4484 struct pet_stmt *stmt;
4485 isl_map *access;
4487 if (!scop)
4488 return NULL;
4489 if (scop->n_stmt < 1)
4490 isl_die(ctx, isl_error_internal,
4491 "expecting at least one statement", return NULL);
4492 stmt = scop->stmts[0];
4493 if (stmt->body->type != pet_expr_unary ||
4494 stmt->body->op != pet_op_kill)
4495 isl_die(ctx, isl_error_internal,
4496 "expecting kill statement", return NULL);
4498 access = isl_map_copy(stmt->body->args[0]->acc.access);
4499 access = isl_map_reset_tuple_id(access, isl_dim_in);
4500 kill = pet_expr_kill_from_access(access);
4501 return pet_stmt_from_pet_expr(ctx, stmt->line, NULL, n_stmt++, kill);
4504 /* Mark all arrays in "scop" as being exposed.
4506 static struct pet_scop *mark_exposed(struct pet_scop *scop)
4508 if (!scop)
4509 return NULL;
4510 for (int i = 0; i < scop->n_array; ++i)
4511 scop->arrays[i]->exposed = 1;
4512 return scop;
4515 /* Try and construct a pet_scop corresponding to (part of)
4516 * a sequence of statements.
4518 * "block" is set if the sequence respresents the children of
4519 * a compound statement.
4520 * "skip_declarations" is set if we should skip initial declarations
4521 * in the sequence of statements.
4523 * If there are any breaks or continues in the individual statements,
4524 * then we may have to compute a new skip condition.
4525 * This is handled using a pet_skip_info_seq object.
4526 * On initialization, the object checks if skip conditions need
4527 * to be computed. If so, it does so in "extract" and adds them in "add".
4529 * If "block" is set, then we need to insert kill statements at
4530 * the end of the block for any array that has been declared by
4531 * one of the statements in the sequence. Each of these declarations
4532 * results in the construction of a kill statement at the place
4533 * of the declaration, so we simply collect duplicates of
4534 * those kill statements and append these duplicates to the constructed scop.
4536 * If "block" is not set, then any array declared by one of the statements
4537 * in the sequence is marked as being exposed.
4539 struct pet_scop *PetScan::extract(StmtRange stmt_range, bool block,
4540 bool skip_declarations)
4542 pet_scop *scop;
4543 StmtIterator i;
4544 int j;
4545 bool partial_range = false;
4546 set<struct pet_stmt *> kills;
4547 set<struct pet_stmt *>::iterator it;
4549 scop = pet_scop_empty(ctx);
4550 for (i = stmt_range.first, j = 0; i != stmt_range.second; ++i, ++j) {
4551 Stmt *child = *i;
4552 struct pet_scop *scop_i;
4554 if (skip_declarations &&
4555 child->getStmtClass() == Stmt::DeclStmtClass)
4556 continue;
4558 scop_i = extract(child);
4559 if (scop && partial) {
4560 pet_scop_free(scop_i);
4561 break;
4563 pet_skip_info_seq skip(ctx, scop, scop_i);
4564 skip.extract(this);
4565 if (skip)
4566 scop_i = pet_scop_prefix(scop_i, 0);
4567 if (scop_i && child->getStmtClass() == Stmt::DeclStmtClass) {
4568 if (block)
4569 kills.insert(extract_kill(scop_i));
4570 else
4571 scop_i = mark_exposed(scop_i);
4573 scop_i = pet_scop_prefix(scop_i, j);
4574 if (options->autodetect) {
4575 if (scop_i)
4576 scop = pet_scop_add_seq(ctx, scop, scop_i);
4577 else
4578 partial_range = true;
4579 if (scop->n_stmt != 0 && !scop_i)
4580 partial = true;
4581 } else {
4582 scop = pet_scop_add_seq(ctx, scop, scop_i);
4585 scop = skip.add(scop, j);
4587 if (partial)
4588 break;
4591 for (it = kills.begin(); it != kills.end(); ++it) {
4592 pet_scop *scop_j;
4593 scop_j = pet_scop_from_pet_stmt(ctx, *it);
4594 scop_j = pet_scop_prefix(scop_j, j);
4595 scop = pet_scop_add_seq(ctx, scop, scop_j);
4598 if (scop && partial_range)
4599 partial = true;
4601 return scop;
4604 /* Return the file offset of the expansion location of "Loc".
4606 static unsigned getExpansionOffset(SourceManager &SM, SourceLocation Loc)
4608 return SM.getFileOffset(SM.getExpansionLoc(Loc));
4611 /* Check if the scop marked by the user is exactly this Stmt
4612 * or part of this Stmt.
4613 * If so, return a pet_scop corresponding to the marked region.
4614 * Otherwise, return NULL.
4616 struct pet_scop *PetScan::scan(Stmt *stmt)
4618 SourceManager &SM = PP.getSourceManager();
4619 unsigned start_off, end_off;
4621 start_off = getExpansionOffset(SM, stmt->getLocStart());
4622 end_off = getExpansionOffset(SM, stmt->getLocEnd());
4624 if (start_off > loc.end)
4625 return NULL;
4626 if (end_off < loc.start)
4627 return NULL;
4628 if (start_off >= loc.start && end_off <= loc.end) {
4629 return extract(stmt);
4632 StmtIterator start;
4633 for (start = stmt->child_begin(); start != stmt->child_end(); ++start) {
4634 Stmt *child = *start;
4635 if (!child)
4636 continue;
4637 start_off = getExpansionOffset(SM, child->getLocStart());
4638 end_off = getExpansionOffset(SM, child->getLocEnd());
4639 if (start_off < loc.start && end_off > loc.end)
4640 return scan(child);
4641 if (start_off >= loc.start)
4642 break;
4645 StmtIterator end;
4646 for (end = start; end != stmt->child_end(); ++end) {
4647 Stmt *child = *end;
4648 start_off = SM.getFileOffset(child->getLocStart());
4649 if (start_off >= loc.end)
4650 break;
4653 return extract(StmtRange(start, end), false, false);
4656 /* Set the size of index "pos" of "array" to "size".
4657 * In particular, add a constraint of the form
4659 * i_pos < size
4661 * to array->extent and a constraint of the form
4663 * size >= 0
4665 * to array->context.
4667 static struct pet_array *update_size(struct pet_array *array, int pos,
4668 __isl_take isl_pw_aff *size)
4670 isl_set *valid;
4671 isl_set *univ;
4672 isl_set *bound;
4673 isl_space *dim;
4674 isl_aff *aff;
4675 isl_pw_aff *index;
4676 isl_id *id;
4678 valid = isl_pw_aff_nonneg_set(isl_pw_aff_copy(size));
4679 array->context = isl_set_intersect(array->context, valid);
4681 dim = isl_set_get_space(array->extent);
4682 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
4683 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, pos, 1);
4684 univ = isl_set_universe(isl_aff_get_domain_space(aff));
4685 index = isl_pw_aff_alloc(univ, aff);
4687 size = isl_pw_aff_add_dims(size, isl_dim_in,
4688 isl_set_dim(array->extent, isl_dim_set));
4689 id = isl_set_get_tuple_id(array->extent);
4690 size = isl_pw_aff_set_tuple_id(size, isl_dim_in, id);
4691 bound = isl_pw_aff_lt_set(index, size);
4693 array->extent = isl_set_intersect(array->extent, bound);
4695 if (!array->context || !array->extent)
4696 goto error;
4698 return array;
4699 error:
4700 pet_array_free(array);
4701 return NULL;
4704 /* Figure out the size of the array at position "pos" and all
4705 * subsequent positions from "type" and update "array" accordingly.
4707 struct pet_array *PetScan::set_upper_bounds(struct pet_array *array,
4708 const Type *type, int pos)
4710 const ArrayType *atype;
4711 isl_pw_aff *size;
4713 if (!array)
4714 return NULL;
4716 if (type->isPointerType()) {
4717 type = type->getPointeeType().getTypePtr();
4718 return set_upper_bounds(array, type, pos + 1);
4720 if (!type->isArrayType())
4721 return array;
4723 type = type->getCanonicalTypeInternal().getTypePtr();
4724 atype = cast<ArrayType>(type);
4726 if (type->isConstantArrayType()) {
4727 const ConstantArrayType *ca = cast<ConstantArrayType>(atype);
4728 size = extract_affine(ca->getSize());
4729 array = update_size(array, pos, size);
4730 } else if (type->isVariableArrayType()) {
4731 const VariableArrayType *vla = cast<VariableArrayType>(atype);
4732 size = extract_affine(vla->getSizeExpr());
4733 array = update_size(array, pos, size);
4736 type = atype->getElementType().getTypePtr();
4738 return set_upper_bounds(array, type, pos + 1);
4741 /* Is "T" the type of a variable length array with static size?
4743 static bool is_vla_with_static_size(QualType T)
4745 const VariableArrayType *vlatype;
4747 if (!T->isVariableArrayType())
4748 return false;
4749 vlatype = cast<VariableArrayType>(T);
4750 return vlatype->getSizeModifier() == VariableArrayType::Static;
4753 /* Return the type of "decl" as an array.
4755 * In particular, if "decl" is a parameter declaration that
4756 * is a variable length array with a static size, then
4757 * return the original type (i.e., the variable length array).
4758 * Otherwise, return the type of decl.
4760 static QualType get_array_type(ValueDecl *decl)
4762 ParmVarDecl *parm;
4763 QualType T;
4765 parm = dyn_cast<ParmVarDecl>(decl);
4766 if (!parm)
4767 return decl->getType();
4769 T = parm->getOriginalType();
4770 if (!is_vla_with_static_size(T))
4771 return decl->getType();
4772 return T;
4775 /* Construct and return a pet_array corresponding to the variable "decl".
4776 * In particular, initialize array->extent to
4778 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
4780 * and then call set_upper_bounds to set the upper bounds on the indices
4781 * based on the type of the variable.
4783 struct pet_array *PetScan::extract_array(isl_ctx *ctx, ValueDecl *decl)
4785 struct pet_array *array;
4786 QualType qt = get_array_type(decl);
4787 const Type *type = qt.getTypePtr();
4788 int depth = array_depth(type);
4789 QualType base = base_type(qt);
4790 string name;
4791 isl_id *id;
4792 isl_space *dim;
4794 array = isl_calloc_type(ctx, struct pet_array);
4795 if (!array)
4796 return NULL;
4798 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
4799 dim = isl_space_set_alloc(ctx, 0, depth);
4800 dim = isl_space_set_tuple_id(dim, isl_dim_set, id);
4802 array->extent = isl_set_nat_universe(dim);
4804 dim = isl_space_params_alloc(ctx, 0);
4805 array->context = isl_set_universe(dim);
4807 array = set_upper_bounds(array, type, 0);
4808 if (!array)
4809 return NULL;
4811 name = base.getAsString();
4812 array->element_type = strdup(name.c_str());
4813 array->element_size = decl->getASTContext().getTypeInfo(base).first / 8;
4815 return array;
4818 /* Construct a list of pet_arrays, one for each array (or scalar)
4819 * accessed inside "scop", add this list to "scop" and return the result.
4821 * The context of "scop" is updated with the intersection of
4822 * the contexts of all arrays, i.e., constraints on the parameters
4823 * that ensure that the arrays have a valid (non-negative) size.
4825 struct pet_scop *PetScan::scan_arrays(struct pet_scop *scop)
4827 int i;
4828 set<ValueDecl *> arrays;
4829 set<ValueDecl *>::iterator it;
4830 int n_array;
4831 struct pet_array **scop_arrays;
4833 if (!scop)
4834 return NULL;
4836 pet_scop_collect_arrays(scop, arrays);
4837 if (arrays.size() == 0)
4838 return scop;
4840 n_array = scop->n_array;
4842 scop_arrays = isl_realloc_array(ctx, scop->arrays, struct pet_array *,
4843 n_array + arrays.size());
4844 if (!scop_arrays)
4845 goto error;
4846 scop->arrays = scop_arrays;
4848 for (it = arrays.begin(), i = 0; it != arrays.end(); ++it, ++i) {
4849 struct pet_array *array;
4850 scop->arrays[n_array + i] = array = extract_array(ctx, *it);
4851 if (!scop->arrays[n_array + i])
4852 goto error;
4853 scop->n_array++;
4854 scop->context = isl_set_intersect(scop->context,
4855 isl_set_copy(array->context));
4856 if (!scop->context)
4857 goto error;
4860 return scop;
4861 error:
4862 pet_scop_free(scop);
4863 return NULL;
4866 /* Bound all parameters in scop->context to the possible values
4867 * of the corresponding C variable.
4869 static struct pet_scop *add_parameter_bounds(struct pet_scop *scop)
4871 int n;
4873 if (!scop)
4874 return NULL;
4876 n = isl_set_dim(scop->context, isl_dim_param);
4877 for (int i = 0; i < n; ++i) {
4878 isl_id *id;
4879 ValueDecl *decl;
4881 id = isl_set_get_dim_id(scop->context, isl_dim_param, i);
4882 if (is_nested_parameter(id)) {
4883 isl_id_free(id);
4884 isl_die(isl_set_get_ctx(scop->context),
4885 isl_error_internal,
4886 "unresolved nested parameter", goto error);
4888 decl = (ValueDecl *) isl_id_get_user(id);
4889 isl_id_free(id);
4891 scop->context = set_parameter_bounds(scop->context, i, decl);
4893 if (!scop->context)
4894 goto error;
4897 return scop;
4898 error:
4899 pet_scop_free(scop);
4900 return NULL;
4903 /* Construct a pet_scop from the given function.
4905 struct pet_scop *PetScan::scan(FunctionDecl *fd)
4907 pet_scop *scop;
4908 Stmt *stmt;
4910 stmt = fd->getBody();
4912 if (options->autodetect)
4913 scop = extract(stmt, true);
4914 else
4915 scop = scan(stmt);
4916 scop = pet_scop_detect_parameter_accesses(scop);
4917 scop = scan_arrays(scop);
4918 scop = add_parameter_bounds(scop);
4919 scop = pet_scop_gist(scop, value_bounds);
4921 return scop;