support (C style) casts in statements
[pet.git] / scan.cc
blob20c3066e224ad172dec78c4a5190e7cf8ba0269e
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 /* Construct a pet_expr representing a (C style) cast.
1677 struct pet_expr *PetScan::extract_expr(CStyleCastExpr *expr)
1679 struct pet_expr *arg;
1680 QualType type;
1682 arg = extract_expr(expr->getSubExpr());
1683 if (!arg)
1684 return NULL;
1686 type = expr->getTypeAsWritten();
1687 return pet_expr_new_cast(ctx, type.getAsString().c_str(), arg);
1690 /* Try and onstruct a pet_expr representing "expr".
1692 struct pet_expr *PetScan::extract_expr(Expr *expr)
1694 switch (expr->getStmtClass()) {
1695 case Stmt::UnaryOperatorClass:
1696 return extract_expr(cast<UnaryOperator>(expr));
1697 case Stmt::CompoundAssignOperatorClass:
1698 case Stmt::BinaryOperatorClass:
1699 return extract_expr(cast<BinaryOperator>(expr));
1700 case Stmt::ImplicitCastExprClass:
1701 return extract_expr(cast<ImplicitCastExpr>(expr));
1702 case Stmt::ArraySubscriptExprClass:
1703 case Stmt::DeclRefExprClass:
1704 case Stmt::IntegerLiteralClass:
1705 return extract_access_expr(expr);
1706 case Stmt::FloatingLiteralClass:
1707 return extract_expr(cast<FloatingLiteral>(expr));
1708 case Stmt::ParenExprClass:
1709 return extract_expr(cast<ParenExpr>(expr));
1710 case Stmt::ConditionalOperatorClass:
1711 return extract_expr(cast<ConditionalOperator>(expr));
1712 case Stmt::CallExprClass:
1713 return extract_expr(cast<CallExpr>(expr));
1714 case Stmt::CStyleCastExprClass:
1715 return extract_expr(cast<CStyleCastExpr>(expr));
1716 default:
1717 unsupported(expr);
1719 return NULL;
1722 /* Check if the given initialization statement is an assignment.
1723 * If so, return that assignment. Otherwise return NULL.
1725 BinaryOperator *PetScan::initialization_assignment(Stmt *init)
1727 BinaryOperator *ass;
1729 if (init->getStmtClass() != Stmt::BinaryOperatorClass)
1730 return NULL;
1732 ass = cast<BinaryOperator>(init);
1733 if (ass->getOpcode() != BO_Assign)
1734 return NULL;
1736 return ass;
1739 /* Check if the given initialization statement is a declaration
1740 * of a single variable.
1741 * If so, return that declaration. Otherwise return NULL.
1743 Decl *PetScan::initialization_declaration(Stmt *init)
1745 DeclStmt *decl;
1747 if (init->getStmtClass() != Stmt::DeclStmtClass)
1748 return NULL;
1750 decl = cast<DeclStmt>(init);
1752 if (!decl->isSingleDecl())
1753 return NULL;
1755 return decl->getSingleDecl();
1758 /* Given the assignment operator in the initialization of a for loop,
1759 * extract the induction variable, i.e., the (integer)variable being
1760 * assigned.
1762 ValueDecl *PetScan::extract_induction_variable(BinaryOperator *init)
1764 Expr *lhs;
1765 DeclRefExpr *ref;
1766 ValueDecl *decl;
1767 const Type *type;
1769 lhs = init->getLHS();
1770 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1771 unsupported(init);
1772 return NULL;
1775 ref = cast<DeclRefExpr>(lhs);
1776 decl = ref->getDecl();
1777 type = decl->getType().getTypePtr();
1779 if (!type->isIntegerType()) {
1780 unsupported(lhs);
1781 return NULL;
1784 return decl;
1787 /* Given the initialization statement of a for loop and the single
1788 * declaration in this initialization statement,
1789 * extract the induction variable, i.e., the (integer) variable being
1790 * declared.
1792 VarDecl *PetScan::extract_induction_variable(Stmt *init, Decl *decl)
1794 VarDecl *vd;
1796 vd = cast<VarDecl>(decl);
1798 const QualType type = vd->getType();
1799 if (!type->isIntegerType()) {
1800 unsupported(init);
1801 return NULL;
1804 if (!vd->getInit()) {
1805 unsupported(init);
1806 return NULL;
1809 return vd;
1812 /* Check that op is of the form iv++ or iv--.
1813 * Return an affine expression "1" or "-1" accordingly.
1815 __isl_give isl_pw_aff *PetScan::extract_unary_increment(
1816 clang::UnaryOperator *op, clang::ValueDecl *iv)
1818 Expr *sub;
1819 DeclRefExpr *ref;
1820 isl_space *space;
1821 isl_aff *aff;
1823 if (!op->isIncrementDecrementOp()) {
1824 unsupported(op);
1825 return NULL;
1828 sub = op->getSubExpr();
1829 if (sub->getStmtClass() != Stmt::DeclRefExprClass) {
1830 unsupported(op);
1831 return NULL;
1834 ref = cast<DeclRefExpr>(sub);
1835 if (ref->getDecl() != iv) {
1836 unsupported(op);
1837 return NULL;
1840 space = isl_space_params_alloc(ctx, 0);
1841 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
1843 if (op->isIncrementOp())
1844 aff = isl_aff_add_constant_si(aff, 1);
1845 else
1846 aff = isl_aff_add_constant_si(aff, -1);
1848 return isl_pw_aff_from_aff(aff);
1851 /* If the isl_pw_aff on which isl_pw_aff_foreach_piece is called
1852 * has a single constant expression, then put this constant in *user.
1853 * The caller is assumed to have checked that this function will
1854 * be called exactly once.
1856 static int extract_cst(__isl_take isl_set *set, __isl_take isl_aff *aff,
1857 void *user)
1859 isl_int *inc = (isl_int *)user;
1860 int res = 0;
1862 if (isl_aff_is_cst(aff))
1863 isl_aff_get_constant(aff, inc);
1864 else
1865 res = -1;
1867 isl_set_free(set);
1868 isl_aff_free(aff);
1870 return res;
1873 /* Check if op is of the form
1875 * iv = iv + inc
1877 * and return inc as an affine expression.
1879 * We extract an affine expression from the RHS, subtract iv and return
1880 * the result.
1882 __isl_give isl_pw_aff *PetScan::extract_binary_increment(BinaryOperator *op,
1883 clang::ValueDecl *iv)
1885 Expr *lhs;
1886 DeclRefExpr *ref;
1887 isl_id *id;
1888 isl_space *dim;
1889 isl_aff *aff;
1890 isl_pw_aff *val;
1892 if (op->getOpcode() != BO_Assign) {
1893 unsupported(op);
1894 return NULL;
1897 lhs = op->getLHS();
1898 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1899 unsupported(op);
1900 return NULL;
1903 ref = cast<DeclRefExpr>(lhs);
1904 if (ref->getDecl() != iv) {
1905 unsupported(op);
1906 return NULL;
1909 val = extract_affine(op->getRHS());
1911 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
1913 dim = isl_space_params_alloc(ctx, 1);
1914 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
1915 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
1916 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
1918 val = isl_pw_aff_sub(val, isl_pw_aff_from_aff(aff));
1920 return val;
1923 /* Check that op is of the form iv += cst or iv -= cst
1924 * and return an affine expression corresponding oto cst or -cst accordingly.
1926 __isl_give isl_pw_aff *PetScan::extract_compound_increment(
1927 CompoundAssignOperator *op, clang::ValueDecl *iv)
1929 Expr *lhs;
1930 DeclRefExpr *ref;
1931 bool neg = false;
1932 isl_pw_aff *val;
1933 BinaryOperatorKind opcode;
1935 opcode = op->getOpcode();
1936 if (opcode != BO_AddAssign && opcode != BO_SubAssign) {
1937 unsupported(op);
1938 return NULL;
1940 if (opcode == BO_SubAssign)
1941 neg = true;
1943 lhs = op->getLHS();
1944 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1945 unsupported(op);
1946 return NULL;
1949 ref = cast<DeclRefExpr>(lhs);
1950 if (ref->getDecl() != iv) {
1951 unsupported(op);
1952 return NULL;
1955 val = extract_affine(op->getRHS());
1956 if (neg)
1957 val = isl_pw_aff_neg(val);
1959 return val;
1962 /* Check that the increment of the given for loop increments
1963 * (or decrements) the induction variable "iv" and return
1964 * the increment as an affine expression if successful.
1966 __isl_give isl_pw_aff *PetScan::extract_increment(clang::ForStmt *stmt,
1967 ValueDecl *iv)
1969 Stmt *inc = stmt->getInc();
1971 if (!inc) {
1972 unsupported(stmt);
1973 return NULL;
1976 if (inc->getStmtClass() == Stmt::UnaryOperatorClass)
1977 return extract_unary_increment(cast<UnaryOperator>(inc), iv);
1978 if (inc->getStmtClass() == Stmt::CompoundAssignOperatorClass)
1979 return extract_compound_increment(
1980 cast<CompoundAssignOperator>(inc), iv);
1981 if (inc->getStmtClass() == Stmt::BinaryOperatorClass)
1982 return extract_binary_increment(cast<BinaryOperator>(inc), iv);
1984 unsupported(inc);
1985 return NULL;
1988 /* Embed the given iteration domain in an extra outer loop
1989 * with induction variable "var".
1990 * If this variable appeared as a parameter in the constraints,
1991 * it is replaced by the new outermost dimension.
1993 static __isl_give isl_set *embed(__isl_take isl_set *set,
1994 __isl_take isl_id *var)
1996 int pos;
1998 set = isl_set_insert_dims(set, isl_dim_set, 0, 1);
1999 pos = isl_set_find_dim_by_id(set, isl_dim_param, var);
2000 if (pos >= 0) {
2001 set = isl_set_equate(set, isl_dim_param, pos, isl_dim_set, 0);
2002 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2005 isl_id_free(var);
2006 return set;
2009 /* Return those elements in the space of "cond" that come after
2010 * (based on "sign") an element in "cond".
2012 static __isl_give isl_set *after(__isl_take isl_set *cond, int sign)
2014 isl_map *previous_to_this;
2016 if (sign > 0)
2017 previous_to_this = isl_map_lex_lt(isl_set_get_space(cond));
2018 else
2019 previous_to_this = isl_map_lex_gt(isl_set_get_space(cond));
2021 cond = isl_set_apply(cond, previous_to_this);
2023 return cond;
2026 /* Create the infinite iteration domain
2028 * { [id] : id >= 0 }
2030 * If "scop" has an affine skip of type pet_skip_later,
2031 * then remove those iterations i that have an earlier iteration
2032 * where the skip condition is satisfied, meaning that iteration i
2033 * is not executed.
2034 * Since we are dealing with a loop without loop iterator,
2035 * the skip condition cannot refer to the current loop iterator and
2036 * so effectively, the returned set is of the form
2038 * { [0]; [id] : id >= 1 and not skip }
2040 static __isl_give isl_set *infinite_domain(__isl_take isl_id *id,
2041 struct pet_scop *scop)
2043 isl_ctx *ctx = isl_id_get_ctx(id);
2044 isl_set *domain;
2045 isl_set *skip;
2047 domain = isl_set_nat_universe(isl_space_set_alloc(ctx, 0, 1));
2048 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, id);
2050 if (!pet_scop_has_affine_skip(scop, pet_skip_later))
2051 return domain;
2053 skip = pet_scop_get_skip(scop, pet_skip_later);
2054 skip = isl_set_fix_si(skip, isl_dim_set, 0, 1);
2055 skip = isl_set_params(skip);
2056 skip = embed(skip, isl_id_copy(id));
2057 skip = isl_set_intersect(skip , isl_set_copy(domain));
2058 domain = isl_set_subtract(domain, after(skip, 1));
2060 return domain;
2063 /* Create an identity mapping on the space containing "domain".
2065 static __isl_give isl_map *identity_map(__isl_keep isl_set *domain)
2067 isl_space *space;
2068 isl_map *id;
2070 space = isl_space_map_from_set(isl_set_get_space(domain));
2071 id = isl_map_identity(space);
2073 return id;
2076 /* Add a filter to "scop" that imposes that it is only executed
2077 * when "break_access" has a zero value for all previous iterations
2078 * of "domain".
2080 * The input "break_access" has a zero-dimensional domain and range.
2082 static struct pet_scop *scop_add_break(struct pet_scop *scop,
2083 __isl_take isl_map *break_access, __isl_take isl_set *domain, int sign)
2085 isl_ctx *ctx = isl_set_get_ctx(domain);
2086 isl_id *id_test;
2087 isl_map *prev;
2089 id_test = isl_map_get_tuple_id(break_access, isl_dim_out);
2090 break_access = isl_map_add_dims(break_access, isl_dim_in, 1);
2091 break_access = isl_map_add_dims(break_access, isl_dim_out, 1);
2092 break_access = isl_map_intersect_range(break_access, domain);
2093 break_access = isl_map_set_tuple_id(break_access, isl_dim_out, id_test);
2094 if (sign > 0)
2095 prev = isl_map_lex_gt_first(isl_map_get_space(break_access), 1);
2096 else
2097 prev = isl_map_lex_lt_first(isl_map_get_space(break_access), 1);
2098 break_access = isl_map_intersect(break_access, prev);
2099 scop = pet_scop_filter(scop, break_access, 0);
2100 scop = pet_scop_merge_filters(scop);
2102 return scop;
2105 /* Construct a pet_scop for an infinite loop around the given body.
2107 * We extract a pet_scop for the body and then embed it in a loop with
2108 * iteration domain
2110 * { [t] : t >= 0 }
2112 * and schedule
2114 * { [t] -> [t] }
2116 * If the body contains any break, then it is taken into
2117 * account in infinite_domain (if the skip condition is affine)
2118 * or in scop_add_break (if the skip condition is not affine).
2120 struct pet_scop *PetScan::extract_infinite_loop(Stmt *body)
2122 isl_id *id;
2123 isl_set *domain;
2124 isl_map *ident;
2125 isl_map *access;
2126 struct pet_scop *scop;
2127 bool has_var_break;
2129 scop = extract(body);
2130 if (!scop)
2131 return NULL;
2133 id = isl_id_alloc(ctx, "t", NULL);
2134 domain = infinite_domain(isl_id_copy(id), scop);
2135 ident = identity_map(domain);
2137 has_var_break = pet_scop_has_var_skip(scop, pet_skip_later);
2138 if (has_var_break)
2139 access = pet_scop_get_skip_map(scop, pet_skip_later);
2141 scop = pet_scop_embed(scop, isl_set_copy(domain),
2142 isl_map_copy(ident), ident, id);
2143 if (has_var_break)
2144 scop = scop_add_break(scop, access, domain, 1);
2145 else
2146 isl_set_free(domain);
2148 return scop;
2151 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
2153 * for (;;)
2154 * body
2157 struct pet_scop *PetScan::extract_infinite_for(ForStmt *stmt)
2159 return extract_infinite_loop(stmt->getBody());
2162 /* Create an access to a virtual array representing the result
2163 * of a condition.
2164 * Unlike other accessed data, the id of the array is NULL as
2165 * there is no ValueDecl in the program corresponding to the virtual
2166 * array.
2167 * The array starts out as a scalar, but grows along with the
2168 * statement writing to the array in pet_scop_embed.
2170 static __isl_give isl_map *create_test_access(isl_ctx *ctx, int test_nr)
2172 isl_space *dim = isl_space_alloc(ctx, 0, 0, 0);
2173 isl_id *id;
2174 char name[50];
2176 snprintf(name, sizeof(name), "__pet_test_%d", test_nr);
2177 id = isl_id_alloc(ctx, name, NULL);
2178 dim = isl_space_set_tuple_id(dim, isl_dim_out, id);
2179 return isl_map_universe(dim);
2182 /* Add an array with the given extent ("access") to the list
2183 * of arrays in "scop" and return the extended pet_scop.
2184 * The array is marked as attaining values 0 and 1 only and
2185 * as each element being assigned at most once.
2187 static struct pet_scop *scop_add_array(struct pet_scop *scop,
2188 __isl_keep isl_map *access, clang::ASTContext &ast_ctx)
2190 isl_ctx *ctx = isl_map_get_ctx(access);
2191 isl_space *dim;
2192 struct pet_array *array;
2194 if (!scop)
2195 return NULL;
2196 if (!ctx)
2197 goto error;
2199 array = isl_calloc_type(ctx, struct pet_array);
2200 if (!array)
2201 goto error;
2203 array->extent = isl_map_range(isl_map_copy(access));
2204 dim = isl_space_params_alloc(ctx, 0);
2205 array->context = isl_set_universe(dim);
2206 dim = isl_space_set_alloc(ctx, 0, 1);
2207 array->value_bounds = isl_set_universe(dim);
2208 array->value_bounds = isl_set_lower_bound_si(array->value_bounds,
2209 isl_dim_set, 0, 0);
2210 array->value_bounds = isl_set_upper_bound_si(array->value_bounds,
2211 isl_dim_set, 0, 1);
2212 array->element_type = strdup("int");
2213 array->element_size = ast_ctx.getTypeInfo(ast_ctx.IntTy).first / 8;
2214 array->uniquely_defined = 1;
2216 if (!array->extent || !array->context)
2217 array = pet_array_free(array);
2219 scop = pet_scop_add_array(scop, array);
2221 return scop;
2222 error:
2223 pet_scop_free(scop);
2224 return NULL;
2227 /* Construct a pet_scop for a while loop of the form
2229 * while (pa)
2230 * body
2232 * In particular, construct a scop for an infinite loop around body and
2233 * intersect the domain with the affine expression.
2234 * Note that this intersection may result in an empty loop.
2236 struct pet_scop *PetScan::extract_affine_while(__isl_take isl_pw_aff *pa,
2237 Stmt *body)
2239 struct pet_scop *scop;
2240 isl_set *dom;
2241 isl_set *valid;
2243 valid = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2244 dom = isl_pw_aff_non_zero_set(pa);
2245 scop = extract_infinite_loop(body);
2246 scop = pet_scop_restrict(scop, dom);
2247 scop = pet_scop_restrict_context(scop, valid);
2249 return scop;
2252 /* Construct a scop for a while, given the scops for the condition
2253 * and the body, the filter access and the iteration domain of
2254 * the while loop.
2256 * In particular, the scop for the condition is filtered to depend
2257 * on "test_access" evaluating to true for all previous iterations
2258 * of the loop, while the scop for the body is filtered to depend
2259 * on "test_access" evaluating to true for all iterations up to the
2260 * current iteration.
2262 * These filtered scops are then combined into a single scop.
2264 * "sign" is positive if the iterator increases and negative
2265 * if it decreases.
2267 static struct pet_scop *scop_add_while(struct pet_scop *scop_cond,
2268 struct pet_scop *scop_body, __isl_take isl_map *test_access,
2269 __isl_take isl_set *domain, int sign)
2271 isl_ctx *ctx = isl_set_get_ctx(domain);
2272 isl_id *id_test;
2273 isl_map *prev;
2275 id_test = isl_map_get_tuple_id(test_access, isl_dim_out);
2276 test_access = isl_map_add_dims(test_access, isl_dim_in, 1);
2277 test_access = isl_map_add_dims(test_access, isl_dim_out, 1);
2278 test_access = isl_map_intersect_range(test_access, domain);
2279 test_access = isl_map_set_tuple_id(test_access, isl_dim_out, id_test);
2280 if (sign > 0)
2281 prev = isl_map_lex_ge_first(isl_map_get_space(test_access), 1);
2282 else
2283 prev = isl_map_lex_le_first(isl_map_get_space(test_access), 1);
2284 test_access = isl_map_intersect(test_access, prev);
2285 scop_body = pet_scop_filter(scop_body, isl_map_copy(test_access), 1);
2286 if (sign > 0)
2287 prev = isl_map_lex_gt_first(isl_map_get_space(test_access), 1);
2288 else
2289 prev = isl_map_lex_lt_first(isl_map_get_space(test_access), 1);
2290 test_access = isl_map_intersect(test_access, prev);
2291 scop_cond = pet_scop_filter(scop_cond, test_access, 1);
2293 return pet_scop_add_seq(ctx, scop_cond, scop_body);
2296 /* Check if the while loop is of the form
2298 * while (affine expression)
2299 * body
2301 * If so, call extract_affine_while to construct a scop.
2303 * Otherwise, construct a generic while scop, with iteration domain
2304 * { [t] : t >= 0 }. The scop consists of two parts, one for
2305 * evaluating the condition and one for the body.
2306 * The schedule is adjusted to reflect that the condition is evaluated
2307 * before the body is executed and the body is filtered to depend
2308 * on the result of the condition evaluating to true on all iterations
2309 * up to the current iteration, while the evaluation the condition itself
2310 * is filtered to depend on the result of the condition evaluating to true
2311 * on all previous iterations.
2312 * The context of the scop representing the body is dropped
2313 * because we don't know how many times the body will be executed,
2314 * if at all.
2316 * If the body contains any break, then it is taken into
2317 * account in infinite_domain (if the skip condition is affine)
2318 * or in scop_add_break (if the skip condition is not affine).
2320 struct pet_scop *PetScan::extract(WhileStmt *stmt)
2322 Expr *cond;
2323 isl_id *id;
2324 isl_map *test_access;
2325 isl_set *domain;
2326 isl_map *ident;
2327 isl_pw_aff *pa;
2328 struct pet_scop *scop, *scop_body;
2329 bool has_var_break;
2330 isl_map *break_access;
2332 cond = stmt->getCond();
2333 if (!cond) {
2334 unsupported(stmt);
2335 return NULL;
2338 clear_assignments clear(assigned_value);
2339 clear.TraverseStmt(stmt->getBody());
2341 pa = try_extract_affine_condition(cond);
2342 if (pa)
2343 return extract_affine_while(pa, stmt->getBody());
2345 if (!allow_nested) {
2346 unsupported(stmt);
2347 return NULL;
2350 test_access = create_test_access(ctx, n_test++);
2351 scop = extract_non_affine_condition(cond, isl_map_copy(test_access));
2352 scop = scop_add_array(scop, test_access, ast_context);
2353 scop_body = extract(stmt->getBody());
2355 id = isl_id_alloc(ctx, "t", NULL);
2356 domain = infinite_domain(isl_id_copy(id), scop_body);
2357 ident = identity_map(domain);
2359 has_var_break = pet_scop_has_var_skip(scop_body, pet_skip_later);
2360 if (has_var_break)
2361 break_access = pet_scop_get_skip_map(scop_body, pet_skip_later);
2363 scop = pet_scop_prefix(scop, 0);
2364 scop = pet_scop_embed(scop, isl_set_copy(domain), isl_map_copy(ident),
2365 isl_map_copy(ident), isl_id_copy(id));
2366 scop_body = pet_scop_reset_context(scop_body);
2367 scop_body = pet_scop_prefix(scop_body, 1);
2368 scop_body = pet_scop_embed(scop_body, isl_set_copy(domain),
2369 isl_map_copy(ident), ident, id);
2371 if (has_var_break) {
2372 scop = scop_add_break(scop, isl_map_copy(break_access),
2373 isl_set_copy(domain), 1);
2374 scop_body = scop_add_break(scop_body, break_access,
2375 isl_set_copy(domain), 1);
2377 scop = scop_add_while(scop, scop_body, test_access, domain, 1);
2379 return scop;
2382 /* Check whether "cond" expresses a simple loop bound
2383 * on the only set dimension.
2384 * In particular, if "up" is set then "cond" should contain only
2385 * upper bounds on the set dimension.
2386 * Otherwise, it should contain only lower bounds.
2388 static bool is_simple_bound(__isl_keep isl_set *cond, isl_int inc)
2390 if (isl_int_is_pos(inc))
2391 return !isl_set_dim_has_any_lower_bound(cond, isl_dim_set, 0);
2392 else
2393 return !isl_set_dim_has_any_upper_bound(cond, isl_dim_set, 0);
2396 /* Extend a condition on a given iteration of a loop to one that
2397 * imposes the same condition on all previous iterations.
2398 * "domain" expresses the lower [upper] bound on the iterations
2399 * when inc is positive [negative].
2401 * In particular, we construct the condition (when inc is positive)
2403 * forall i' : (domain(i') and i' <= i) => cond(i')
2405 * which is equivalent to
2407 * not exists i' : domain(i') and i' <= i and not cond(i')
2409 * We construct this set by negating cond, applying a map
2411 * { [i'] -> [i] : domain(i') and i' <= i }
2413 * and then negating the result again.
2415 static __isl_give isl_set *valid_for_each_iteration(__isl_take isl_set *cond,
2416 __isl_take isl_set *domain, isl_int inc)
2418 isl_map *previous_to_this;
2420 if (isl_int_is_pos(inc))
2421 previous_to_this = isl_map_lex_le(isl_set_get_space(domain));
2422 else
2423 previous_to_this = isl_map_lex_ge(isl_set_get_space(domain));
2425 previous_to_this = isl_map_intersect_domain(previous_to_this, domain);
2427 cond = isl_set_complement(cond);
2428 cond = isl_set_apply(cond, previous_to_this);
2429 cond = isl_set_complement(cond);
2431 return cond;
2434 /* Construct a domain of the form
2436 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
2438 static __isl_give isl_set *strided_domain(__isl_take isl_id *id,
2439 __isl_take isl_pw_aff *init, isl_int inc)
2441 isl_aff *aff;
2442 isl_space *dim;
2443 isl_set *set;
2445 init = isl_pw_aff_insert_dims(init, isl_dim_in, 0, 1);
2446 dim = isl_pw_aff_get_domain_space(init);
2447 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2448 aff = isl_aff_add_coefficient(aff, isl_dim_in, 0, inc);
2449 init = isl_pw_aff_add(init, isl_pw_aff_from_aff(aff));
2451 dim = isl_space_set_alloc(isl_pw_aff_get_ctx(init), 1, 1);
2452 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
2453 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2454 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2456 set = isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff), init);
2458 set = isl_set_lower_bound_si(set, isl_dim_set, 0, 0);
2460 return isl_set_params(set);
2463 /* Assuming "cond" represents a bound on a loop where the loop
2464 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2465 * is possible.
2467 * Under the given assumptions, wrapping is only possible if "cond" allows
2468 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2469 * increasing iterator and 0 in case of a decreasing iterator.
2471 static bool can_wrap(__isl_keep isl_set *cond, ValueDecl *iv, isl_int inc)
2473 bool cw;
2474 isl_int limit;
2475 isl_set *test;
2477 test = isl_set_copy(cond);
2479 isl_int_init(limit);
2480 if (isl_int_is_neg(inc))
2481 isl_int_set_si(limit, 0);
2482 else {
2483 isl_int_set_si(limit, 1);
2484 isl_int_mul_2exp(limit, limit, get_type_size(iv));
2485 isl_int_sub_ui(limit, limit, 1);
2488 test = isl_set_fix(cond, isl_dim_set, 0, limit);
2489 cw = !isl_set_is_empty(test);
2490 isl_set_free(test);
2492 isl_int_clear(limit);
2494 return cw;
2497 /* Given a one-dimensional space, construct the following mapping on this
2498 * space
2500 * { [v] -> [v mod 2^width] }
2502 * where width is the number of bits used to represent the values
2503 * of the unsigned variable "iv".
2505 static __isl_give isl_map *compute_wrapping(__isl_take isl_space *dim,
2506 ValueDecl *iv)
2508 isl_int mod;
2509 isl_aff *aff;
2510 isl_map *map;
2512 isl_int_init(mod);
2513 isl_int_set_si(mod, 1);
2514 isl_int_mul_2exp(mod, mod, get_type_size(iv));
2516 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2517 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2518 aff = isl_aff_mod(aff, mod);
2520 isl_int_clear(mod);
2522 return isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2523 map = isl_map_reverse(map);
2526 /* Project out the parameter "id" from "set".
2528 static __isl_give isl_set *set_project_out_by_id(__isl_take isl_set *set,
2529 __isl_keep isl_id *id)
2531 int pos;
2533 pos = isl_set_find_dim_by_id(set, isl_dim_param, id);
2534 if (pos >= 0)
2535 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2537 return set;
2540 /* Compute the set of parameters for which "set1" is a subset of "set2".
2542 * set1 is a subset of set2 if
2544 * forall i in set1 : i in set2
2546 * or
2548 * not exists i in set1 and i not in set2
2550 * i.e.,
2552 * not exists i in set1 \ set2
2554 static __isl_give isl_set *enforce_subset(__isl_take isl_set *set1,
2555 __isl_take isl_set *set2)
2557 return isl_set_complement(isl_set_params(isl_set_subtract(set1, set2)));
2560 /* Compute the set of parameter values for which "cond" holds
2561 * on the next iteration for each element of "dom".
2563 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2564 * and then compute the set of parameters for which the result is a subset
2565 * of "cond".
2567 static __isl_give isl_set *valid_on_next(__isl_take isl_set *cond,
2568 __isl_take isl_set *dom, isl_int inc)
2570 isl_space *space;
2571 isl_aff *aff;
2572 isl_map *next;
2574 space = isl_set_get_space(dom);
2575 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2576 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2577 aff = isl_aff_add_constant(aff, inc);
2578 next = isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2580 dom = isl_set_apply(dom, next);
2582 return enforce_subset(dom, cond);
2585 /* Does "id" refer to a nested access?
2587 static bool is_nested_parameter(__isl_keep isl_id *id)
2589 return id && isl_id_get_user(id) && !isl_id_get_name(id);
2592 /* Does parameter "pos" of "space" refer to a nested access?
2594 static bool is_nested_parameter(__isl_keep isl_space *space, int pos)
2596 bool nested;
2597 isl_id *id;
2599 id = isl_space_get_dim_id(space, isl_dim_param, pos);
2600 nested = is_nested_parameter(id);
2601 isl_id_free(id);
2603 return nested;
2606 /* Does "space" involve any parameters that refer to nested
2607 * accesses, i.e., parameters with no name?
2609 static bool has_nested(__isl_keep isl_space *space)
2611 int nparam;
2613 nparam = isl_space_dim(space, isl_dim_param);
2614 for (int i = 0; i < nparam; ++i)
2615 if (is_nested_parameter(space, i))
2616 return true;
2618 return false;
2621 /* Does "pa" involve any parameters that refer to nested
2622 * accesses, i.e., parameters with no name?
2624 static bool has_nested(__isl_keep isl_pw_aff *pa)
2626 isl_space *space;
2627 bool nested;
2629 space = isl_pw_aff_get_space(pa);
2630 nested = has_nested(space);
2631 isl_space_free(space);
2633 return nested;
2636 /* Construct a pet_scop for a for statement.
2637 * The for loop is required to be of the form
2639 * for (i = init; condition; ++i)
2641 * or
2643 * for (i = init; condition; --i)
2645 * The initialization of the for loop should either be an assignment
2646 * to an integer variable, or a declaration of such a variable with
2647 * initialization.
2649 * The condition is allowed to contain nested accesses, provided
2650 * they are not being written to inside the body of the loop.
2651 * Otherwise, or if the condition is otherwise non-affine, the for loop is
2652 * essentially treated as a while loop, with iteration domain
2653 * { [i] : i >= init }.
2655 * We extract a pet_scop for the body and then embed it in a loop with
2656 * iteration domain and schedule
2658 * { [i] : i >= init and condition' }
2659 * { [i] -> [i] }
2661 * or
2663 * { [i] : i <= init and condition' }
2664 * { [i] -> [-i] }
2666 * Where condition' is equal to condition if the latter is
2667 * a simple upper [lower] bound and a condition that is extended
2668 * to apply to all previous iterations otherwise.
2670 * If the condition is non-affine, then we drop the condition from the
2671 * iteration domain and instead create a separate statement
2672 * for evaluating the condition. The body is then filtered to depend
2673 * on the result of the condition evaluating to true on all iterations
2674 * up to the current iteration, while the evaluation the condition itself
2675 * is filtered to depend on the result of the condition evaluating to true
2676 * on all previous iterations.
2677 * The context of the scop representing the body is dropped
2678 * because we don't know how many times the body will be executed,
2679 * if at all.
2681 * If the stride of the loop is not 1, then "i >= init" is replaced by
2683 * (exists a: i = init + stride * a and a >= 0)
2685 * If the loop iterator i is unsigned, then wrapping may occur.
2686 * During the computation, we work with a virtual iterator that
2687 * does not wrap. However, the condition in the code applies
2688 * to the wrapped value, so we need to change condition(i)
2689 * into condition([i % 2^width]).
2690 * After computing the virtual domain and schedule, we apply
2691 * the function { [v] -> [v % 2^width] } to the domain and the domain
2692 * of the schedule. In order not to lose any information, we also
2693 * need to intersect the domain of the schedule with the virtual domain
2694 * first, since some iterations in the wrapped domain may be scheduled
2695 * several times, typically an infinite number of times.
2696 * Note that there may be no need to perform this final wrapping
2697 * if the loop condition (after wrapping) satisfies certain conditions.
2698 * However, the is_simple_bound condition is not enough since it doesn't
2699 * check if there even is an upper bound.
2701 * If the loop condition is non-affine, then we keep the virtual
2702 * iterator in the iteration domain and instead replace all accesses
2703 * to the original iterator by the wrapping of the virtual iterator.
2705 * Wrapping on unsigned iterators can be avoided entirely if
2706 * loop condition is simple, the loop iterator is incremented
2707 * [decremented] by one and the last value before wrapping cannot
2708 * possibly satisfy the loop condition.
2710 * Before extracting a pet_scop from the body we remove all
2711 * assignments in assigned_value to variables that are assigned
2712 * somewhere in the body of the loop.
2714 * Valid parameters for a for loop are those for which the initial
2715 * value itself, the increment on each domain iteration and
2716 * the condition on both the initial value and
2717 * the result of incrementing the iterator for each iteration of the domain
2718 * can be evaluated.
2719 * If the loop condition is non-affine, then we only consider validity
2720 * of the initial value.
2722 * If the body contains any break, then we keep track of it in "skip"
2723 * (if the skip condition is affine) or it is handled in scop_add_break
2724 * (if the skip condition is not affine).
2725 * Note that the affine break condition needs to be considered with
2726 * respect to previous iterations in the virtual domain (if any)
2727 * and that the domain needs to be kept virtual if there is a non-affine
2728 * break condition.
2730 struct pet_scop *PetScan::extract_for(ForStmt *stmt)
2732 BinaryOperator *ass;
2733 Decl *decl;
2734 Stmt *init;
2735 Expr *lhs, *rhs;
2736 ValueDecl *iv;
2737 isl_space *space;
2738 isl_set *domain;
2739 isl_map *sched;
2740 isl_set *cond = NULL;
2741 isl_set *skip = NULL;
2742 isl_id *id;
2743 struct pet_scop *scop, *scop_cond = NULL;
2744 assigned_value_cache cache(assigned_value);
2745 isl_int inc;
2746 bool is_one;
2747 bool is_unsigned;
2748 bool is_simple;
2749 bool is_virtual;
2750 bool keep_virtual = false;
2751 bool has_affine_break;
2752 bool has_var_break;
2753 isl_map *wrap = NULL;
2754 isl_pw_aff *pa, *pa_inc, *init_val;
2755 isl_set *valid_init;
2756 isl_set *valid_cond;
2757 isl_set *valid_cond_init;
2758 isl_set *valid_cond_next;
2759 isl_set *valid_inc;
2760 isl_map *test_access = NULL, *break_access = NULL;
2761 int stmt_id;
2763 if (!stmt->getInit() && !stmt->getCond() && !stmt->getInc())
2764 return extract_infinite_for(stmt);
2766 init = stmt->getInit();
2767 if (!init) {
2768 unsupported(stmt);
2769 return NULL;
2771 if ((ass = initialization_assignment(init)) != NULL) {
2772 iv = extract_induction_variable(ass);
2773 if (!iv)
2774 return NULL;
2775 lhs = ass->getLHS();
2776 rhs = ass->getRHS();
2777 } else if ((decl = initialization_declaration(init)) != NULL) {
2778 VarDecl *var = extract_induction_variable(init, decl);
2779 if (!var)
2780 return NULL;
2781 iv = var;
2782 rhs = var->getInit();
2783 lhs = create_DeclRefExpr(var);
2784 } else {
2785 unsupported(stmt->getInit());
2786 return NULL;
2789 pa_inc = extract_increment(stmt, iv);
2790 if (!pa_inc)
2791 return NULL;
2793 isl_int_init(inc);
2794 if (isl_pw_aff_n_piece(pa_inc) != 1 ||
2795 isl_pw_aff_foreach_piece(pa_inc, &extract_cst, &inc) < 0) {
2796 isl_pw_aff_free(pa_inc);
2797 unsupported(stmt->getInc());
2798 isl_int_clear(inc);
2799 return NULL;
2801 valid_inc = isl_pw_aff_domain(pa_inc);
2803 is_unsigned = iv->getType()->isUnsignedIntegerType();
2805 assigned_value.erase(iv);
2806 clear_assignments clear(assigned_value);
2807 clear.TraverseStmt(stmt->getBody());
2809 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
2811 pa = try_extract_nested_condition(stmt->getCond());
2812 if (allow_nested && (!pa || has_nested(pa)))
2813 stmt_id = n_stmt++;
2815 scop = extract(stmt->getBody());
2817 has_affine_break = scop &&
2818 pet_scop_has_affine_skip(scop, pet_skip_later);
2819 if (has_affine_break) {
2820 skip = pet_scop_get_skip(scop, pet_skip_later);
2821 skip = isl_set_fix_si(skip, isl_dim_set, 0, 1);
2822 skip = isl_set_params(skip);
2824 has_var_break = scop && pet_scop_has_var_skip(scop, pet_skip_later);
2825 if (has_var_break) {
2826 break_access = pet_scop_get_skip_map(scop, pet_skip_later);
2827 keep_virtual = true;
2830 if (pa && !is_nested_allowed(pa, scop)) {
2831 isl_pw_aff_free(pa);
2832 pa = NULL;
2835 if (!allow_nested && !pa)
2836 pa = try_extract_affine_condition(stmt->getCond());
2837 valid_cond = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2838 cond = isl_pw_aff_non_zero_set(pa);
2839 if (allow_nested && !cond) {
2840 int save_n_stmt = n_stmt;
2841 test_access = create_test_access(ctx, n_test++);
2842 n_stmt = stmt_id;
2843 scop_cond = extract_non_affine_condition(stmt->getCond(),
2844 isl_map_copy(test_access));
2845 n_stmt = save_n_stmt;
2846 scop_cond = scop_add_array(scop_cond, test_access, ast_context);
2847 scop_cond = pet_scop_prefix(scop_cond, 0);
2848 scop = pet_scop_reset_context(scop);
2849 scop = pet_scop_prefix(scop, 1);
2850 keep_virtual = true;
2851 cond = isl_set_universe(isl_space_set_alloc(ctx, 0, 0));
2854 cond = embed(cond, isl_id_copy(id));
2855 skip = embed(skip, isl_id_copy(id));
2856 valid_cond = isl_set_coalesce(valid_cond);
2857 valid_cond = embed(valid_cond, isl_id_copy(id));
2858 valid_inc = embed(valid_inc, isl_id_copy(id));
2859 is_one = isl_int_is_one(inc) || isl_int_is_negone(inc);
2860 is_virtual = is_unsigned && (!is_one || can_wrap(cond, iv, inc));
2862 init_val = extract_affine(rhs);
2863 valid_cond_init = enforce_subset(
2864 isl_set_from_pw_aff(isl_pw_aff_copy(init_val)),
2865 isl_set_copy(valid_cond));
2866 if (is_one && !is_virtual) {
2867 isl_pw_aff_free(init_val);
2868 pa = extract_comparison(isl_int_is_pos(inc) ? BO_GE : BO_LE,
2869 lhs, rhs, init);
2870 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2871 valid_init = set_project_out_by_id(valid_init, id);
2872 domain = isl_pw_aff_non_zero_set(pa);
2873 } else {
2874 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(init_val));
2875 domain = strided_domain(isl_id_copy(id), init_val, inc);
2878 domain = embed(domain, isl_id_copy(id));
2879 if (is_virtual) {
2880 isl_map *rev_wrap;
2881 wrap = compute_wrapping(isl_set_get_space(cond), iv);
2882 rev_wrap = isl_map_reverse(isl_map_copy(wrap));
2883 cond = isl_set_apply(cond, isl_map_copy(rev_wrap));
2884 skip = isl_set_apply(skip, isl_map_copy(rev_wrap));
2885 valid_cond = isl_set_apply(valid_cond, isl_map_copy(rev_wrap));
2886 valid_inc = isl_set_apply(valid_inc, rev_wrap);
2888 is_simple = is_simple_bound(cond, inc);
2889 if (!is_simple) {
2890 cond = isl_set_gist(cond, isl_set_copy(domain));
2891 is_simple = is_simple_bound(cond, inc);
2893 if (!is_simple)
2894 cond = valid_for_each_iteration(cond,
2895 isl_set_copy(domain), inc);
2896 domain = isl_set_intersect(domain, cond);
2897 if (has_affine_break) {
2898 skip = isl_set_intersect(skip , isl_set_copy(domain));
2899 skip = after(skip, isl_int_sgn(inc));
2900 domain = isl_set_subtract(domain, skip);
2902 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, isl_id_copy(id));
2903 space = isl_space_from_domain(isl_set_get_space(domain));
2904 space = isl_space_add_dims(space, isl_dim_out, 1);
2905 sched = isl_map_universe(space);
2906 if (isl_int_is_pos(inc))
2907 sched = isl_map_equate(sched, isl_dim_in, 0, isl_dim_out, 0);
2908 else
2909 sched = isl_map_oppose(sched, isl_dim_in, 0, isl_dim_out, 0);
2911 valid_cond_next = valid_on_next(valid_cond, isl_set_copy(domain), inc);
2912 valid_inc = enforce_subset(isl_set_copy(domain), valid_inc);
2914 if (is_virtual && !keep_virtual) {
2915 wrap = isl_map_set_dim_id(wrap,
2916 isl_dim_out, 0, isl_id_copy(id));
2917 sched = isl_map_intersect_domain(sched, isl_set_copy(domain));
2918 domain = isl_set_apply(domain, isl_map_copy(wrap));
2919 sched = isl_map_apply_domain(sched, wrap);
2921 if (!(is_virtual && keep_virtual)) {
2922 space = isl_set_get_space(domain);
2923 wrap = isl_map_identity(isl_space_map_from_set(space));
2926 scop_cond = pet_scop_embed(scop_cond, isl_set_copy(domain),
2927 isl_map_copy(sched), isl_map_copy(wrap), isl_id_copy(id));
2928 scop = pet_scop_embed(scop, isl_set_copy(domain), sched, wrap, id);
2929 scop = resolve_nested(scop);
2930 if (has_var_break)
2931 scop = scop_add_break(scop, break_access, isl_set_copy(domain),
2932 isl_int_sgn(inc));
2933 if (test_access) {
2934 scop = scop_add_while(scop_cond, scop, test_access, domain,
2935 isl_int_sgn(inc));
2936 isl_set_free(valid_inc);
2937 } else {
2938 scop = pet_scop_restrict_context(scop, valid_inc);
2939 scop = pet_scop_restrict_context(scop, valid_cond_next);
2940 scop = pet_scop_restrict_context(scop, valid_cond_init);
2941 isl_set_free(domain);
2943 clear_assignment(assigned_value, iv);
2945 isl_int_clear(inc);
2947 scop = pet_scop_restrict_context(scop, valid_init);
2949 return scop;
2952 struct pet_scop *PetScan::extract(CompoundStmt *stmt, bool skip_declarations)
2954 return extract(stmt->children(), true, skip_declarations);
2957 /* Does parameter "pos" of "map" refer to a nested access?
2959 static bool is_nested_parameter(__isl_keep isl_map *map, int pos)
2961 bool nested;
2962 isl_id *id;
2964 id = isl_map_get_dim_id(map, isl_dim_param, pos);
2965 nested = is_nested_parameter(id);
2966 isl_id_free(id);
2968 return nested;
2971 /* How many parameters of "space" refer to nested accesses, i.e., have no name?
2973 static int n_nested_parameter(__isl_keep isl_space *space)
2975 int n = 0;
2976 int nparam;
2978 nparam = isl_space_dim(space, isl_dim_param);
2979 for (int i = 0; i < nparam; ++i)
2980 if (is_nested_parameter(space, i))
2981 ++n;
2983 return n;
2986 /* How many parameters of "map" refer to nested accesses, i.e., have no name?
2988 static int n_nested_parameter(__isl_keep isl_map *map)
2990 isl_space *space;
2991 int n;
2993 space = isl_map_get_space(map);
2994 n = n_nested_parameter(space);
2995 isl_space_free(space);
2997 return n;
3000 /* For each nested access parameter in "space",
3001 * construct a corresponding pet_expr, place it in args and
3002 * record its position in "param2pos".
3003 * "n_arg" is the number of elements that are already in args.
3004 * The position recorded in "param2pos" takes this number into account.
3005 * If the pet_expr corresponding to a parameter is identical to
3006 * the pet_expr corresponding to an earlier parameter, then these two
3007 * parameters are made to refer to the same element in args.
3009 * Return the final number of elements in args or -1 if an error has occurred.
3011 int PetScan::extract_nested(__isl_keep isl_space *space,
3012 int n_arg, struct pet_expr **args, std::map<int,int> &param2pos)
3014 int nparam;
3016 nparam = isl_space_dim(space, isl_dim_param);
3017 for (int i = 0; i < nparam; ++i) {
3018 int j;
3019 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
3020 Expr *nested;
3022 if (!is_nested_parameter(id)) {
3023 isl_id_free(id);
3024 continue;
3027 nested = (Expr *) isl_id_get_user(id);
3028 args[n_arg] = extract_expr(nested);
3029 if (!args[n_arg])
3030 return -1;
3032 for (j = 0; j < n_arg; ++j)
3033 if (pet_expr_is_equal(args[j], args[n_arg]))
3034 break;
3036 if (j < n_arg) {
3037 pet_expr_free(args[n_arg]);
3038 args[n_arg] = NULL;
3039 param2pos[i] = j;
3040 } else
3041 param2pos[i] = n_arg++;
3043 isl_id_free(id);
3046 return n_arg;
3049 /* For each nested access parameter in the access relations in "expr",
3050 * construct a corresponding pet_expr, place it in expr->args and
3051 * record its position in "param2pos".
3052 * n is the number of nested access parameters.
3054 struct pet_expr *PetScan::extract_nested(struct pet_expr *expr, int n,
3055 std::map<int,int> &param2pos)
3057 isl_space *space;
3059 expr->args = isl_calloc_array(ctx, struct pet_expr *, n);
3060 expr->n_arg = n;
3061 if (!expr->args)
3062 goto error;
3064 space = isl_map_get_space(expr->acc.access);
3065 n = extract_nested(space, 0, expr->args, param2pos);
3066 isl_space_free(space);
3068 if (n < 0)
3069 goto error;
3071 expr->n_arg = n;
3072 return expr;
3073 error:
3074 pet_expr_free(expr);
3075 return NULL;
3078 /* Look for parameters in any access relation in "expr" that
3079 * refer to nested accesses. In particular, these are
3080 * parameters with no name.
3082 * If there are any such parameters, then the domain of the access
3083 * relation, which is still [] at this point, is replaced by
3084 * [[] -> [t_1,...,t_n]], with n the number of these parameters
3085 * (after identifying identical nested accesses).
3086 * The parameters are then equated to the corresponding t dimensions
3087 * and subsequently projected out.
3088 * param2pos maps the position of the parameter to the position
3089 * of the corresponding t dimension.
3091 struct pet_expr *PetScan::resolve_nested(struct pet_expr *expr)
3093 int n;
3094 int nparam;
3095 int n_in;
3096 isl_space *dim;
3097 isl_map *map;
3098 std::map<int,int> param2pos;
3100 if (!expr)
3101 return expr;
3103 for (int i = 0; i < expr->n_arg; ++i) {
3104 expr->args[i] = resolve_nested(expr->args[i]);
3105 if (!expr->args[i]) {
3106 pet_expr_free(expr);
3107 return NULL;
3111 if (expr->type != pet_expr_access)
3112 return expr;
3114 n = n_nested_parameter(expr->acc.access);
3115 if (n == 0)
3116 return expr;
3118 expr = extract_nested(expr, n, param2pos);
3119 if (!expr)
3120 return NULL;
3122 n = expr->n_arg;
3123 nparam = isl_map_dim(expr->acc.access, isl_dim_param);
3124 n_in = isl_map_dim(expr->acc.access, isl_dim_in);
3125 dim = isl_map_get_space(expr->acc.access);
3126 dim = isl_space_domain(dim);
3127 dim = isl_space_from_domain(dim);
3128 dim = isl_space_add_dims(dim, isl_dim_out, n);
3129 map = isl_map_universe(dim);
3130 map = isl_map_domain_map(map);
3131 map = isl_map_reverse(map);
3132 expr->acc.access = isl_map_apply_domain(expr->acc.access, map);
3134 for (int i = nparam - 1; i >= 0; --i) {
3135 isl_id *id = isl_map_get_dim_id(expr->acc.access,
3136 isl_dim_param, i);
3137 if (!is_nested_parameter(id)) {
3138 isl_id_free(id);
3139 continue;
3142 expr->acc.access = isl_map_equate(expr->acc.access,
3143 isl_dim_param, i, isl_dim_in,
3144 n_in + param2pos[i]);
3145 expr->acc.access = isl_map_project_out(expr->acc.access,
3146 isl_dim_param, i, 1);
3148 isl_id_free(id);
3151 return expr;
3152 error:
3153 pet_expr_free(expr);
3154 return NULL;
3157 /* Convert a top-level pet_expr to a pet_scop with one statement.
3158 * This mainly involves resolving nested expression parameters
3159 * and setting the name of the iteration space.
3160 * The name is given by "label" if it is non-NULL. Otherwise,
3161 * it is of the form S_<n_stmt>.
3163 struct pet_scop *PetScan::extract(Stmt *stmt, struct pet_expr *expr,
3164 __isl_take isl_id *label)
3166 struct pet_stmt *ps;
3167 SourceLocation loc = stmt->getLocStart();
3168 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3170 expr = resolve_nested(expr);
3171 ps = pet_stmt_from_pet_expr(ctx, line, label, n_stmt++, expr);
3172 return pet_scop_from_pet_stmt(ctx, ps);
3175 /* Check if we can extract an affine expression from "expr".
3176 * Return the expressions as an isl_pw_aff if we can and NULL otherwise.
3177 * We turn on autodetection so that we won't generate any warnings
3178 * and turn off nesting, so that we won't accept any non-affine constructs.
3180 __isl_give isl_pw_aff *PetScan::try_extract_affine(Expr *expr)
3182 isl_pw_aff *pwaff;
3183 int save_autodetect = options->autodetect;
3184 bool save_nesting = nesting_enabled;
3186 options->autodetect = 1;
3187 nesting_enabled = false;
3189 pwaff = extract_affine(expr);
3191 options->autodetect = save_autodetect;
3192 nesting_enabled = save_nesting;
3194 return pwaff;
3197 /* Check whether "expr" is an affine expression.
3199 bool PetScan::is_affine(Expr *expr)
3201 isl_pw_aff *pwaff;
3203 pwaff = try_extract_affine(expr);
3204 isl_pw_aff_free(pwaff);
3206 return pwaff != NULL;
3209 /* Check if we can extract an affine constraint from "expr".
3210 * Return the constraint as an isl_set if we can and NULL otherwise.
3211 * We turn on autodetection so that we won't generate any warnings
3212 * and turn off nesting, so that we won't accept any non-affine constructs.
3214 __isl_give isl_pw_aff *PetScan::try_extract_affine_condition(Expr *expr)
3216 isl_pw_aff *cond;
3217 int save_autodetect = options->autodetect;
3218 bool save_nesting = nesting_enabled;
3220 options->autodetect = 1;
3221 nesting_enabled = false;
3223 cond = extract_condition(expr);
3225 options->autodetect = save_autodetect;
3226 nesting_enabled = save_nesting;
3228 return cond;
3231 /* Check whether "expr" is an affine constraint.
3233 bool PetScan::is_affine_condition(Expr *expr)
3235 isl_pw_aff *cond;
3237 cond = try_extract_affine_condition(expr);
3238 isl_pw_aff_free(cond);
3240 return cond != NULL;
3243 /* Check if we can extract a condition from "expr".
3244 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
3245 * If allow_nested is set, then the condition may involve parameters
3246 * corresponding to nested accesses.
3247 * We turn on autodetection so that we won't generate any warnings.
3249 __isl_give isl_pw_aff *PetScan::try_extract_nested_condition(Expr *expr)
3251 isl_pw_aff *cond;
3252 int save_autodetect = options->autodetect;
3253 bool save_nesting = nesting_enabled;
3255 options->autodetect = 1;
3256 nesting_enabled = allow_nested;
3257 cond = extract_condition(expr);
3259 options->autodetect = save_autodetect;
3260 nesting_enabled = save_nesting;
3262 return cond;
3265 /* If the top-level expression of "stmt" is an assignment, then
3266 * return that assignment as a BinaryOperator.
3267 * Otherwise return NULL.
3269 static BinaryOperator *top_assignment_or_null(Stmt *stmt)
3271 BinaryOperator *ass;
3273 if (!stmt)
3274 return NULL;
3275 if (stmt->getStmtClass() != Stmt::BinaryOperatorClass)
3276 return NULL;
3278 ass = cast<BinaryOperator>(stmt);
3279 if(ass->getOpcode() != BO_Assign)
3280 return NULL;
3282 return ass;
3285 /* Check if the given if statement is a conditional assignement
3286 * with a non-affine condition. If so, construct a pet_scop
3287 * corresponding to this conditional assignment. Otherwise return NULL.
3289 * In particular we check if "stmt" is of the form
3291 * if (condition)
3292 * a = f(...);
3293 * else
3294 * a = g(...);
3296 * where a is some array or scalar access.
3297 * The constructed pet_scop then corresponds to the expression
3299 * a = condition ? f(...) : g(...)
3301 * All access relations in f(...) are intersected with condition
3302 * while all access relation in g(...) are intersected with the complement.
3304 struct pet_scop *PetScan::extract_conditional_assignment(IfStmt *stmt)
3306 BinaryOperator *ass_then, *ass_else;
3307 isl_map *write_then, *write_else;
3308 isl_set *cond, *comp;
3309 isl_map *map;
3310 isl_pw_aff *pa;
3311 int equal;
3312 struct pet_expr *pe_cond, *pe_then, *pe_else, *pe, *pe_write;
3313 bool save_nesting = nesting_enabled;
3315 if (!options->detect_conditional_assignment)
3316 return NULL;
3318 ass_then = top_assignment_or_null(stmt->getThen());
3319 ass_else = top_assignment_or_null(stmt->getElse());
3321 if (!ass_then || !ass_else)
3322 return NULL;
3324 if (is_affine_condition(stmt->getCond()))
3325 return NULL;
3327 write_then = extract_access(ass_then->getLHS());
3328 write_else = extract_access(ass_else->getLHS());
3330 equal = isl_map_is_equal(write_then, write_else);
3331 isl_map_free(write_else);
3332 if (equal < 0 || !equal) {
3333 isl_map_free(write_then);
3334 return NULL;
3337 nesting_enabled = allow_nested;
3338 pa = extract_condition(stmt->getCond());
3339 nesting_enabled = save_nesting;
3340 cond = isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa));
3341 comp = isl_pw_aff_zero_set(isl_pw_aff_copy(pa));
3342 map = isl_map_from_range(isl_set_from_pw_aff(pa));
3344 pe_cond = pet_expr_from_access(map);
3346 pe_then = extract_expr(ass_then->getRHS());
3347 pe_then = pet_expr_restrict(pe_then, cond);
3348 pe_else = extract_expr(ass_else->getRHS());
3349 pe_else = pet_expr_restrict(pe_else, comp);
3351 pe = pet_expr_new_ternary(ctx, pe_cond, pe_then, pe_else);
3352 pe_write = pet_expr_from_access(write_then);
3353 if (pe_write) {
3354 pe_write->acc.write = 1;
3355 pe_write->acc.read = 0;
3357 pe = pet_expr_new_binary(ctx, pet_op_assign, pe_write, pe);
3358 return extract(stmt, pe);
3361 /* Create a pet_scop with a single statement evaluating "cond"
3362 * and writing the result to a virtual scalar, as expressed by
3363 * "access".
3365 struct pet_scop *PetScan::extract_non_affine_condition(Expr *cond,
3366 __isl_take isl_map *access)
3368 struct pet_expr *expr, *write;
3369 struct pet_stmt *ps;
3370 struct pet_scop *scop;
3371 SourceLocation loc = cond->getLocStart();
3372 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3374 write = pet_expr_from_access(access);
3375 if (write) {
3376 write->acc.write = 1;
3377 write->acc.read = 0;
3379 expr = extract_expr(cond);
3380 expr = resolve_nested(expr);
3381 expr = pet_expr_new_binary(ctx, pet_op_assign, write, expr);
3382 ps = pet_stmt_from_pet_expr(ctx, line, NULL, n_stmt++, expr);
3383 scop = pet_scop_from_pet_stmt(ctx, ps);
3384 scop = resolve_nested(scop);
3386 return scop;
3389 extern "C" {
3390 static __isl_give isl_map *embed_access(__isl_take isl_map *access,
3391 void *user);
3394 /* Apply the map pointed to by "user" to the domain of the access
3395 * relation, thereby embedding it in the range of the map.
3396 * The domain of both relations is the zero-dimensional domain.
3398 static __isl_give isl_map *embed_access(__isl_take isl_map *access, void *user)
3400 isl_map *map = (isl_map *) user;
3402 return isl_map_apply_domain(access, isl_map_copy(map));
3405 /* Apply "map" to all access relations in "expr".
3407 static struct pet_expr *embed(struct pet_expr *expr, __isl_keep isl_map *map)
3409 return pet_expr_foreach_access(expr, &embed_access, map);
3412 /* How many parameters of "set" refer to nested accesses, i.e., have no name?
3414 static int n_nested_parameter(__isl_keep isl_set *set)
3416 isl_space *space;
3417 int n;
3419 space = isl_set_get_space(set);
3420 n = n_nested_parameter(space);
3421 isl_space_free(space);
3423 return n;
3426 /* Remove all parameters from "map" that refer to nested accesses.
3428 static __isl_give isl_map *remove_nested_parameters(__isl_take isl_map *map)
3430 int nparam;
3431 isl_space *space;
3433 space = isl_map_get_space(map);
3434 nparam = isl_space_dim(space, isl_dim_param);
3435 for (int i = nparam - 1; i >= 0; --i)
3436 if (is_nested_parameter(space, i))
3437 map = isl_map_project_out(map, isl_dim_param, i, 1);
3438 isl_space_free(space);
3440 return map;
3443 extern "C" {
3444 static __isl_give isl_map *access_remove_nested_parameters(
3445 __isl_take isl_map *access, void *user);
3448 static __isl_give isl_map *access_remove_nested_parameters(
3449 __isl_take isl_map *access, void *user)
3451 return remove_nested_parameters(access);
3454 /* Remove all nested access parameters from the schedule and all
3455 * accesses of "stmt".
3456 * There is no need to remove them from the domain as these parameters
3457 * have already been removed from the domain when this function is called.
3459 static struct pet_stmt *remove_nested_parameters(struct pet_stmt *stmt)
3461 if (!stmt)
3462 return NULL;
3463 stmt->schedule = remove_nested_parameters(stmt->schedule);
3464 stmt->body = pet_expr_foreach_access(stmt->body,
3465 &access_remove_nested_parameters, NULL);
3466 if (!stmt->schedule || !stmt->body)
3467 goto error;
3468 for (int i = 0; i < stmt->n_arg; ++i) {
3469 stmt->args[i] = pet_expr_foreach_access(stmt->args[i],
3470 &access_remove_nested_parameters, NULL);
3471 if (!stmt->args[i])
3472 goto error;
3475 return stmt;
3476 error:
3477 pet_stmt_free(stmt);
3478 return NULL;
3481 /* For each nested access parameter in the domain of "stmt",
3482 * construct a corresponding pet_expr, place it before the original
3483 * elements in stmt->args and record its position in "param2pos".
3484 * n is the number of nested access parameters.
3486 struct pet_stmt *PetScan::extract_nested(struct pet_stmt *stmt, int n,
3487 std::map<int,int> &param2pos)
3489 int i;
3490 isl_space *space;
3491 int n_arg;
3492 struct pet_expr **args;
3494 n_arg = stmt->n_arg;
3495 args = isl_calloc_array(ctx, struct pet_expr *, n + n_arg);
3496 if (!args)
3497 goto error;
3499 space = isl_set_get_space(stmt->domain);
3500 n_arg = extract_nested(space, 0, args, param2pos);
3501 isl_space_free(space);
3503 if (n_arg < 0)
3504 goto error;
3506 for (i = 0; i < stmt->n_arg; ++i)
3507 args[n_arg + i] = stmt->args[i];
3508 free(stmt->args);
3509 stmt->args = args;
3510 stmt->n_arg += n_arg;
3512 return stmt;
3513 error:
3514 if (args) {
3515 for (i = 0; i < n; ++i)
3516 pet_expr_free(args[i]);
3517 free(args);
3519 pet_stmt_free(stmt);
3520 return NULL;
3523 /* Check whether any of the arguments i of "stmt" starting at position "n"
3524 * is equal to one of the first "n" arguments j.
3525 * If so, combine the constraints on arguments i and j and remove
3526 * argument i.
3528 static struct pet_stmt *remove_duplicate_arguments(struct pet_stmt *stmt, int n)
3530 int i, j;
3531 isl_map *map;
3533 if (!stmt)
3534 return NULL;
3535 if (n == 0)
3536 return stmt;
3537 if (n == stmt->n_arg)
3538 return stmt;
3540 map = isl_set_unwrap(stmt->domain);
3542 for (i = stmt->n_arg - 1; i >= n; --i) {
3543 for (j = 0; j < n; ++j)
3544 if (pet_expr_is_equal(stmt->args[i], stmt->args[j]))
3545 break;
3546 if (j >= n)
3547 continue;
3549 map = isl_map_equate(map, isl_dim_out, i, isl_dim_out, j);
3550 map = isl_map_project_out(map, isl_dim_out, i, 1);
3552 pet_expr_free(stmt->args[i]);
3553 for (j = i; j + 1 < stmt->n_arg; ++j)
3554 stmt->args[j] = stmt->args[j + 1];
3555 stmt->n_arg--;
3558 stmt->domain = isl_map_wrap(map);
3559 if (!stmt->domain)
3560 goto error;
3561 return stmt;
3562 error:
3563 pet_stmt_free(stmt);
3564 return NULL;
3567 /* Look for parameters in the iteration domain of "stmt" that
3568 * refer to nested accesses. In particular, these are
3569 * parameters with no name.
3571 * If there are any such parameters, then as many extra variables
3572 * (after identifying identical nested accesses) are inserted in the
3573 * range of the map wrapped inside the domain, before the original variables.
3574 * If the original domain is not a wrapped map, then a new wrapped
3575 * map is created with zero output dimensions.
3576 * The parameters are then equated to the corresponding output dimensions
3577 * and subsequently projected out, from the iteration domain,
3578 * the schedule and the access relations.
3579 * For each of the output dimensions, a corresponding argument
3580 * expression is inserted. Initially they are created with
3581 * a zero-dimensional domain, so they have to be embedded
3582 * in the current iteration domain.
3583 * param2pos maps the position of the parameter to the position
3584 * of the corresponding output dimension in the wrapped map.
3586 struct pet_stmt *PetScan::resolve_nested(struct pet_stmt *stmt)
3588 int n;
3589 int nparam;
3590 unsigned n_arg;
3591 isl_map *map;
3592 std::map<int,int> param2pos;
3594 if (!stmt)
3595 return NULL;
3597 n = n_nested_parameter(stmt->domain);
3598 if (n == 0)
3599 return stmt;
3601 n_arg = stmt->n_arg;
3602 stmt = extract_nested(stmt, n, param2pos);
3603 if (!stmt)
3604 return NULL;
3606 n = stmt->n_arg - n_arg;
3607 nparam = isl_set_dim(stmt->domain, isl_dim_param);
3608 if (isl_set_is_wrapping(stmt->domain))
3609 map = isl_set_unwrap(stmt->domain);
3610 else
3611 map = isl_map_from_domain(stmt->domain);
3612 map = isl_map_insert_dims(map, isl_dim_out, 0, n);
3614 for (int i = nparam - 1; i >= 0; --i) {
3615 isl_id *id;
3617 if (!is_nested_parameter(map, i))
3618 continue;
3620 id = isl_map_get_tuple_id(stmt->args[param2pos[i]]->acc.access,
3621 isl_dim_out);
3622 map = isl_map_set_dim_id(map, isl_dim_out, param2pos[i], id);
3623 map = isl_map_equate(map, isl_dim_param, i, isl_dim_out,
3624 param2pos[i]);
3625 map = isl_map_project_out(map, isl_dim_param, i, 1);
3628 stmt->domain = isl_map_wrap(map);
3630 map = isl_set_unwrap(isl_set_copy(stmt->domain));
3631 map = isl_map_from_range(isl_map_domain(map));
3632 for (int pos = 0; pos < n; ++pos)
3633 stmt->args[pos] = embed(stmt->args[pos], map);
3634 isl_map_free(map);
3636 stmt = remove_nested_parameters(stmt);
3637 stmt = remove_duplicate_arguments(stmt, n);
3639 return stmt;
3640 error:
3641 pet_stmt_free(stmt);
3642 return NULL;
3645 /* For each statement in "scop", move the parameters that correspond
3646 * to nested access into the ranges of the domains and create
3647 * corresponding argument expressions.
3649 struct pet_scop *PetScan::resolve_nested(struct pet_scop *scop)
3651 if (!scop)
3652 return NULL;
3654 for (int i = 0; i < scop->n_stmt; ++i) {
3655 scop->stmts[i] = resolve_nested(scop->stmts[i]);
3656 if (!scop->stmts[i])
3657 goto error;
3660 return scop;
3661 error:
3662 pet_scop_free(scop);
3663 return NULL;
3666 /* Given an access expression "expr", is the variable accessed by
3667 * "expr" assigned anywhere inside "scop"?
3669 static bool is_assigned(pet_expr *expr, pet_scop *scop)
3671 bool assigned = false;
3672 isl_id *id;
3674 id = isl_map_get_tuple_id(expr->acc.access, isl_dim_out);
3675 assigned = pet_scop_writes(scop, id);
3676 isl_id_free(id);
3678 return assigned;
3681 /* Are all nested access parameters in "pa" allowed given "scop".
3682 * In particular, is none of them written by anywhere inside "scop".
3684 * If "scop" has any skip conditions, then no nested access parameters
3685 * are allowed. In particular, if there is any nested access in a guard
3686 * for a piece of code containing a "continue", then we want to introduce
3687 * a separate statement for evaluating this guard so that we can express
3688 * that the result is false for all previous iterations.
3690 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff *pa, pet_scop *scop)
3692 int nparam;
3694 if (!scop)
3695 return true;
3697 nparam = isl_pw_aff_dim(pa, isl_dim_param);
3698 for (int i = 0; i < nparam; ++i) {
3699 Expr *nested;
3700 isl_id *id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
3701 pet_expr *expr;
3702 bool allowed;
3704 if (!is_nested_parameter(id)) {
3705 isl_id_free(id);
3706 continue;
3709 if (pet_scop_has_skip(scop, pet_skip_now)) {
3710 isl_id_free(id);
3711 return false;
3714 nested = (Expr *) isl_id_get_user(id);
3715 expr = extract_expr(nested);
3716 allowed = expr && expr->type == pet_expr_access &&
3717 !is_assigned(expr, scop);
3719 pet_expr_free(expr);
3720 isl_id_free(id);
3722 if (!allowed)
3723 return false;
3726 return true;
3729 /* Do we need to construct a skip condition of the given type
3730 * on an if statement, given that the if condition is non-affine?
3732 * pet_scop_filter_skip can only handle the case where the if condition
3733 * holds (the then branch) and the skip condition is universal.
3734 * In any other case, we need to construct a new skip condition.
3736 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
3737 bool have_else, enum pet_skip type)
3739 if (have_else && scop_else && pet_scop_has_skip(scop_else, type))
3740 return true;
3741 if (scop_then && pet_scop_has_skip(scop_then, type) &&
3742 !pet_scop_has_universal_skip(scop_then, type))
3743 return true;
3744 return false;
3747 /* Do we need to construct a skip condition of the given type
3748 * on an if statement, given that the if condition is affine?
3750 * There is no need to construct a new skip condition if all
3751 * the skip conditions are affine.
3753 static bool need_skip_aff(struct pet_scop *scop_then,
3754 struct pet_scop *scop_else, bool have_else, enum pet_skip type)
3756 if (scop_then && pet_scop_has_var_skip(scop_then, type))
3757 return true;
3758 if (have_else && scop_else && pet_scop_has_var_skip(scop_else, type))
3759 return true;
3760 return false;
3763 /* Do we need to construct a skip condition of the given type
3764 * on an if statement?
3766 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
3767 bool have_else, enum pet_skip type, bool affine)
3769 if (affine)
3770 return need_skip_aff(scop_then, scop_else, have_else, type);
3771 else
3772 return need_skip(scop_then, scop_else, have_else, type);
3775 /* Construct an affine expression pet_expr that is evaluates
3776 * to the constant "val".
3778 static struct pet_expr *universally(isl_ctx *ctx, int val)
3780 isl_space *space;
3781 isl_map *map;
3783 space = isl_space_alloc(ctx, 0, 0, 1);
3784 map = isl_map_universe(space);
3785 map = isl_map_fix_si(map, isl_dim_out, 0, val);
3787 return pet_expr_from_access(map);
3790 /* Construct an affine expression pet_expr that is evaluates
3791 * to the constant 1.
3793 static struct pet_expr *universally_true(isl_ctx *ctx)
3795 return universally(ctx, 1);
3798 /* Construct an affine expression pet_expr that is evaluates
3799 * to the constant 0.
3801 static struct pet_expr *universally_false(isl_ctx *ctx)
3803 return universally(ctx, 0);
3806 /* Given an access relation "test_access" for the if condition,
3807 * an access relation "skip_access" for the skip condition and
3808 * scops for the then and else branches, construct a scop for
3809 * computing "skip_access".
3811 * The computed scop contains a single statement that essentially does
3813 * skip_cond = test_cond ? skip_cond_then : skip_cond_else
3815 * If the skip conditions of the then and/or else branch are not affine,
3816 * then they need to be filtered by test_access.
3817 * If they are missing, then this means the skip condition is false.
3819 * Since we are constructing a skip condition for the if statement,
3820 * the skip conditions on the then and else branches are removed.
3822 static struct pet_scop *extract_skip(PetScan *scan,
3823 __isl_take isl_map *test_access, __isl_take isl_map *skip_access,
3824 struct pet_scop *scop_then, struct pet_scop *scop_else, bool have_else,
3825 enum pet_skip type)
3827 struct pet_expr *expr_then, *expr_else, *expr, *expr_skip;
3828 struct pet_stmt *stmt;
3829 struct pet_scop *scop;
3830 isl_ctx *ctx = scan->ctx;
3832 if (!scop_then)
3833 goto error;
3834 if (have_else && !scop_else)
3835 goto error;
3837 if (pet_scop_has_skip(scop_then, type)) {
3838 expr_then = pet_scop_get_skip_expr(scop_then, type);
3839 pet_scop_reset_skip(scop_then, type);
3840 if (!pet_expr_is_affine(expr_then))
3841 expr_then = pet_expr_filter(expr_then,
3842 isl_map_copy(test_access), 1);
3843 } else
3844 expr_then = universally_false(ctx);
3846 if (have_else && pet_scop_has_skip(scop_else, type)) {
3847 expr_else = pet_scop_get_skip_expr(scop_else, type);
3848 pet_scop_reset_skip(scop_else, type);
3849 if (!pet_expr_is_affine(expr_else))
3850 expr_else = pet_expr_filter(expr_else,
3851 isl_map_copy(test_access), 0);
3852 } else
3853 expr_else = universally_false(ctx);
3855 expr = pet_expr_from_access(test_access);
3856 expr = pet_expr_new_ternary(ctx, expr, expr_then, expr_else);
3857 expr_skip = pet_expr_from_access(isl_map_copy(skip_access));
3858 if (expr_skip) {
3859 expr_skip->acc.write = 1;
3860 expr_skip->acc.read = 0;
3862 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
3863 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, scan->n_stmt++, expr);
3865 scop = pet_scop_from_pet_stmt(ctx, stmt);
3866 scop = scop_add_array(scop, skip_access, scan->ast_context);
3867 isl_map_free(skip_access);
3869 return scop;
3870 error:
3871 isl_map_free(test_access);
3872 isl_map_free(skip_access);
3873 return NULL;
3876 /* Is scop's skip_now condition equal to its skip_later condition?
3877 * In particular, this means that it either has no skip_now condition
3878 * or both a skip_now and a skip_later condition (that are equal to each other).
3880 static bool skip_equals_skip_later(struct pet_scop *scop)
3882 int has_skip_now, has_skip_later;
3883 int equal;
3884 isl_set *skip_now, *skip_later;
3886 if (!scop)
3887 return false;
3888 has_skip_now = pet_scop_has_skip(scop, pet_skip_now);
3889 has_skip_later = pet_scop_has_skip(scop, pet_skip_later);
3890 if (has_skip_now != has_skip_later)
3891 return false;
3892 if (!has_skip_now)
3893 return true;
3895 skip_now = pet_scop_get_skip(scop, pet_skip_now);
3896 skip_later = pet_scop_get_skip(scop, pet_skip_later);
3897 equal = isl_set_is_equal(skip_now, skip_later);
3898 isl_set_free(skip_now);
3899 isl_set_free(skip_later);
3901 return equal;
3904 /* Drop the skip conditions of type pet_skip_later from scop1 and scop2.
3906 static void drop_skip_later(struct pet_scop *scop1, struct pet_scop *scop2)
3908 pet_scop_reset_skip(scop1, pet_skip_later);
3909 pet_scop_reset_skip(scop2, pet_skip_later);
3912 /* Structure that handles the construction of skip conditions.
3914 * scop_then and scop_else represent the then and else branches
3915 * of the if statement
3917 * skip[type] is true if we need to construct a skip condition of that type
3918 * equal is set if the skip conditions of types pet_skip_now and pet_skip_later
3919 * are equal to each other
3920 * access[type] is the virtual array representing the skip condition
3921 * scop[type] is a scop for computing the skip condition
3923 struct pet_skip_info {
3924 isl_ctx *ctx;
3926 bool skip[2];
3927 bool equal;
3928 isl_map *access[2];
3929 struct pet_scop *scop[2];
3931 pet_skip_info(isl_ctx *ctx) : ctx(ctx) {}
3933 operator bool() { return skip[pet_skip_now] || skip[pet_skip_later]; }
3936 /* Structure that handles the construction of skip conditions on if statements.
3938 * scop_then and scop_else represent the then and else branches
3939 * of the if statement
3941 struct pet_skip_info_if : public pet_skip_info {
3942 struct pet_scop *scop_then, *scop_else;
3943 bool have_else;
3945 pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
3946 struct pet_scop *scop_else, bool have_else, bool affine);
3947 void extract(PetScan *scan, __isl_keep isl_map *access,
3948 enum pet_skip type);
3949 void extract(PetScan *scan, __isl_keep isl_map *access);
3950 void extract(PetScan *scan, __isl_keep isl_pw_aff *cond);
3951 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
3952 int offset);
3953 struct pet_scop *add(struct pet_scop *scop, int offset);
3956 /* Initialize a pet_skip_info_if structure based on the then and else branches
3957 * and based on whether the if condition is affine or not.
3959 pet_skip_info_if::pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
3960 struct pet_scop *scop_else, bool have_else, bool affine) :
3961 pet_skip_info(ctx), scop_then(scop_then), scop_else(scop_else),
3962 have_else(have_else)
3964 skip[pet_skip_now] =
3965 need_skip(scop_then, scop_else, have_else, pet_skip_now, affine);
3966 equal = skip[pet_skip_now] && skip_equals_skip_later(scop_then) &&
3967 (!have_else || skip_equals_skip_later(scop_else));
3968 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
3969 need_skip(scop_then, scop_else, have_else, pet_skip_later, affine);
3972 /* If we need to construct a skip condition of the given type,
3973 * then do so now.
3975 * "map" represents the if condition.
3977 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_map *map,
3978 enum pet_skip type)
3980 if (!skip[type])
3981 return;
3983 access[type] = create_test_access(isl_map_get_ctx(map), scan->n_test++);
3984 scop[type] = extract_skip(scan, isl_map_copy(map),
3985 isl_map_copy(access[type]),
3986 scop_then, scop_else, have_else, type);
3989 /* Construct the required skip conditions, given the if condition "map".
3991 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_map *map)
3993 extract(scan, map, pet_skip_now);
3994 extract(scan, map, pet_skip_later);
3995 if (equal)
3996 drop_skip_later(scop_then, scop_else);
3999 /* Construct the required skip conditions, given the if condition "cond".
4001 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_pw_aff *cond)
4003 isl_set *test_set;
4004 isl_map *test;
4006 if (!skip[pet_skip_now] && !skip[pet_skip_later])
4007 return;
4009 test_set = isl_set_from_pw_aff(isl_pw_aff_copy(cond));
4010 test = isl_map_from_range(test_set);
4011 extract(scan, test);
4012 isl_map_free(test);
4015 /* Add the computed skip condition of the give type to "main" and
4016 * add the scop for computing the condition at the given offset.
4018 * If equal is set, then we only computed a skip condition for pet_skip_now,
4019 * but we also need to set it as main's pet_skip_later.
4021 struct pet_scop *pet_skip_info_if::add(struct pet_scop *main,
4022 enum pet_skip type, int offset)
4024 isl_set *skip_set;
4026 if (!skip[type])
4027 return main;
4029 skip_set = isl_map_range(access[type]);
4030 access[type] = NULL;
4031 scop[type] = pet_scop_prefix(scop[type], offset);
4032 main = pet_scop_add_par(ctx, main, scop[type]);
4033 scop[type] = NULL;
4035 if (equal)
4036 main = pet_scop_set_skip(main, pet_skip_later,
4037 isl_set_copy(skip_set));
4039 main = pet_scop_set_skip(main, type, skip_set);
4041 return main;
4044 /* Add the computed skip conditions to "main" and
4045 * add the scops for computing the conditions at the given offset.
4047 struct pet_scop *pet_skip_info_if::add(struct pet_scop *scop, int offset)
4049 scop = add(scop, pet_skip_now, offset);
4050 scop = add(scop, pet_skip_later, offset);
4052 return scop;
4055 /* Construct a pet_scop for a non-affine if statement.
4057 * We create a separate statement that writes the result
4058 * of the non-affine condition to a virtual scalar.
4059 * A constraint requiring the value of this virtual scalar to be one
4060 * is added to the iteration domains of the then branch.
4061 * Similarly, a constraint requiring the value of this virtual scalar
4062 * to be zero is added to the iteration domains of the else branch, if any.
4063 * We adjust the schedules to ensure that the virtual scalar is written
4064 * before it is read.
4066 * If there are any breaks or continues in the then and/or else
4067 * branches, then we may have to compute a new skip condition.
4068 * This is handled using a pet_skip_info_if object.
4069 * On initialization, the object checks if skip conditions need
4070 * to be computed. If so, it does so in "extract" and adds them in "add".
4072 struct pet_scop *PetScan::extract_non_affine_if(Expr *cond,
4073 struct pet_scop *scop_then, struct pet_scop *scop_else,
4074 bool have_else, int stmt_id)
4076 struct pet_scop *scop;
4077 isl_map *test_access;
4078 int save_n_stmt = n_stmt;
4080 test_access = create_test_access(ctx, n_test++);
4081 n_stmt = stmt_id;
4082 scop = extract_non_affine_condition(cond, isl_map_copy(test_access));
4083 n_stmt = save_n_stmt;
4084 scop = scop_add_array(scop, test_access, ast_context);
4086 pet_skip_info_if skip(ctx, scop_then, scop_else, have_else, false);
4087 skip.extract(this, test_access);
4089 scop = pet_scop_prefix(scop, 0);
4090 scop_then = pet_scop_prefix(scop_then, 1);
4091 scop_then = pet_scop_filter(scop_then, isl_map_copy(test_access), 1);
4092 if (have_else) {
4093 scop_else = pet_scop_prefix(scop_else, 1);
4094 scop_else = pet_scop_filter(scop_else, test_access, 0);
4095 scop_then = pet_scop_add_par(ctx, scop_then, scop_else);
4096 } else
4097 isl_map_free(test_access);
4099 scop = pet_scop_add_seq(ctx, scop, scop_then);
4101 scop = skip.add(scop, 2);
4103 return scop;
4106 /* Construct a pet_scop for an if statement.
4108 * If the condition fits the pattern of a conditional assignment,
4109 * then it is handled by extract_conditional_assignment.
4110 * Otherwise, we do the following.
4112 * If the condition is affine, then the condition is added
4113 * to the iteration domains of the then branch, while the
4114 * opposite of the condition in added to the iteration domains
4115 * of the else branch, if any.
4116 * We allow the condition to be dynamic, i.e., to refer to
4117 * scalars or array elements that may be written to outside
4118 * of the given if statement. These nested accesses are then represented
4119 * as output dimensions in the wrapping iteration domain.
4120 * If it also written _inside_ the then or else branch, then
4121 * we treat the condition as non-affine.
4122 * As explained in extract_non_affine_if, this will introduce
4123 * an extra statement.
4124 * For aesthetic reasons, we want this statement to have a statement
4125 * number that is lower than those of the then and else branches.
4126 * In order to evaluate if will need such a statement, however, we
4127 * first construct scops for the then and else branches.
4128 * We therefore reserve a statement number if we might have to
4129 * introduce such an extra statement.
4131 * If the condition is not affine, then the scop is created in
4132 * extract_non_affine_if.
4134 * If there are any breaks or continues in the then and/or else
4135 * branches, then we may have to compute a new skip condition.
4136 * This is handled using a pet_skip_info_if object.
4137 * On initialization, the object checks if skip conditions need
4138 * to be computed. If so, it does so in "extract" and adds them in "add".
4140 struct pet_scop *PetScan::extract(IfStmt *stmt)
4142 struct pet_scop *scop_then, *scop_else = NULL, *scop;
4143 isl_pw_aff *cond;
4144 int stmt_id;
4145 isl_set *set;
4146 isl_set *valid;
4148 scop = extract_conditional_assignment(stmt);
4149 if (scop)
4150 return scop;
4152 cond = try_extract_nested_condition(stmt->getCond());
4153 if (allow_nested && (!cond || has_nested(cond)))
4154 stmt_id = n_stmt++;
4157 assigned_value_cache cache(assigned_value);
4158 scop_then = extract(stmt->getThen());
4161 if (stmt->getElse()) {
4162 assigned_value_cache cache(assigned_value);
4163 scop_else = extract(stmt->getElse());
4164 if (options->autodetect) {
4165 if (scop_then && !scop_else) {
4166 partial = true;
4167 isl_pw_aff_free(cond);
4168 return scop_then;
4170 if (!scop_then && scop_else) {
4171 partial = true;
4172 isl_pw_aff_free(cond);
4173 return scop_else;
4178 if (cond &&
4179 (!is_nested_allowed(cond, scop_then) ||
4180 (stmt->getElse() && !is_nested_allowed(cond, scop_else)))) {
4181 isl_pw_aff_free(cond);
4182 cond = NULL;
4184 if (allow_nested && !cond)
4185 return extract_non_affine_if(stmt->getCond(), scop_then,
4186 scop_else, stmt->getElse(), stmt_id);
4188 if (!cond)
4189 cond = extract_condition(stmt->getCond());
4191 pet_skip_info_if skip(ctx, scop_then, scop_else, stmt->getElse(), true);
4192 skip.extract(this, cond);
4194 valid = isl_pw_aff_domain(isl_pw_aff_copy(cond));
4195 set = isl_pw_aff_non_zero_set(cond);
4196 scop = pet_scop_restrict(scop_then, isl_set_copy(set));
4198 if (stmt->getElse()) {
4199 set = isl_set_subtract(isl_set_copy(valid), set);
4200 scop_else = pet_scop_restrict(scop_else, set);
4201 scop = pet_scop_add_par(ctx, scop, scop_else);
4202 } else
4203 isl_set_free(set);
4204 scop = resolve_nested(scop);
4205 scop = pet_scop_restrict_context(scop, valid);
4207 if (skip)
4208 scop = pet_scop_prefix(scop, 0);
4209 scop = skip.add(scop, 1);
4211 return scop;
4214 /* Try and construct a pet_scop for a label statement.
4215 * We currently only allow labels on expression statements.
4217 struct pet_scop *PetScan::extract(LabelStmt *stmt)
4219 isl_id *label;
4220 Stmt *sub;
4222 sub = stmt->getSubStmt();
4223 if (!isa<Expr>(sub)) {
4224 unsupported(stmt);
4225 return NULL;
4228 label = isl_id_alloc(ctx, stmt->getName(), NULL);
4230 return extract(sub, extract_expr(cast<Expr>(sub)), label);
4233 /* Construct a pet_scop for a continue statement.
4235 * We simply create an empty scop with a universal pet_skip_now
4236 * skip condition. This skip condition will then be taken into
4237 * account by the enclosing loop construct, possibly after
4238 * being incorporated into outer skip conditions.
4240 struct pet_scop *PetScan::extract(ContinueStmt *stmt)
4242 pet_scop *scop;
4243 isl_space *space;
4244 isl_set *set;
4246 scop = pet_scop_empty(ctx);
4247 if (!scop)
4248 return NULL;
4250 space = isl_space_set_alloc(ctx, 0, 1);
4251 set = isl_set_universe(space);
4252 set = isl_set_fix_si(set, isl_dim_set, 0, 1);
4253 scop = pet_scop_set_skip(scop, pet_skip_now, set);
4255 return scop;
4258 /* Construct a pet_scop for a break statement.
4260 * We simply create an empty scop with both a universal pet_skip_now
4261 * skip condition and a universal pet_skip_later skip condition.
4262 * These skip conditions will then be taken into
4263 * account by the enclosing loop construct, possibly after
4264 * being incorporated into outer skip conditions.
4266 struct pet_scop *PetScan::extract(BreakStmt *stmt)
4268 pet_scop *scop;
4269 isl_space *space;
4270 isl_set *set;
4272 scop = pet_scop_empty(ctx);
4273 if (!scop)
4274 return NULL;
4276 space = isl_space_set_alloc(ctx, 0, 1);
4277 set = isl_set_universe(space);
4278 set = isl_set_fix_si(set, isl_dim_set, 0, 1);
4279 scop = pet_scop_set_skip(scop, pet_skip_now, isl_set_copy(set));
4280 scop = pet_scop_set_skip(scop, pet_skip_later, set);
4282 return scop;
4285 /* Try and construct a pet_scop corresponding to "stmt".
4287 * If "stmt" is a compound statement, then "skip_declarations"
4288 * indicates whether we should skip initial declarations in the
4289 * compound statement.
4291 struct pet_scop *PetScan::extract(Stmt *stmt, bool skip_declarations)
4293 if (isa<Expr>(stmt))
4294 return extract(stmt, extract_expr(cast<Expr>(stmt)));
4296 switch (stmt->getStmtClass()) {
4297 case Stmt::WhileStmtClass:
4298 return extract(cast<WhileStmt>(stmt));
4299 case Stmt::ForStmtClass:
4300 return extract_for(cast<ForStmt>(stmt));
4301 case Stmt::IfStmtClass:
4302 return extract(cast<IfStmt>(stmt));
4303 case Stmt::CompoundStmtClass:
4304 return extract(cast<CompoundStmt>(stmt), skip_declarations);
4305 case Stmt::LabelStmtClass:
4306 return extract(cast<LabelStmt>(stmt));
4307 case Stmt::ContinueStmtClass:
4308 return extract(cast<ContinueStmt>(stmt));
4309 case Stmt::BreakStmtClass:
4310 return extract(cast<BreakStmt>(stmt));
4311 case Stmt::DeclStmtClass:
4312 return extract(cast<DeclStmt>(stmt));
4313 default:
4314 unsupported(stmt);
4317 return NULL;
4320 /* Do we need to construct a skip condition of the given type
4321 * on a sequence of statements?
4323 * There is no need to construct a new skip condition if only
4324 * only of the two statements has a skip condition or if both
4325 * of their skip conditions are affine.
4327 * In principle we also don't need a new continuation variable if
4328 * the continuation of scop2 is affine, but then we would need
4329 * to allow more complicated forms of continuations.
4331 static bool need_skip_seq(struct pet_scop *scop1, struct pet_scop *scop2,
4332 enum pet_skip type)
4334 if (!scop1 || !pet_scop_has_skip(scop1, type))
4335 return false;
4336 if (!scop2 || !pet_scop_has_skip(scop2, type))
4337 return false;
4338 if (pet_scop_has_affine_skip(scop1, type) &&
4339 pet_scop_has_affine_skip(scop2, type))
4340 return false;
4341 return true;
4344 /* Construct a scop for computing the skip condition of the given type and
4345 * with access relation "skip_access" for a sequence of two scops "scop1"
4346 * and "scop2".
4348 * The computed scop contains a single statement that essentially does
4350 * skip_cond = skip_cond_1 ? 1 : skip_cond_2
4352 * or, in other words, skip_cond1 || skip_cond2.
4353 * In this expression, skip_cond_2 is filtered to reflect that it is
4354 * only evaluated when skip_cond_1 is false.
4356 * The skip condition on scop1 is not removed because it still needs
4357 * to be applied to scop2 when these two scops are combined.
4359 static struct pet_scop *extract_skip_seq(PetScan *ps,
4360 __isl_take isl_map *skip_access,
4361 struct pet_scop *scop1, struct pet_scop *scop2, enum pet_skip type)
4363 isl_map *access;
4364 struct pet_expr *expr1, *expr2, *expr, *expr_skip;
4365 struct pet_stmt *stmt;
4366 struct pet_scop *scop;
4367 isl_ctx *ctx = ps->ctx;
4369 if (!scop1 || !scop2)
4370 goto error;
4372 expr1 = pet_scop_get_skip_expr(scop1, type);
4373 expr2 = pet_scop_get_skip_expr(scop2, type);
4374 pet_scop_reset_skip(scop2, type);
4376 expr2 = pet_expr_filter(expr2, isl_map_copy(expr1->acc.access), 0);
4378 expr = universally_true(ctx);
4379 expr = pet_expr_new_ternary(ctx, expr1, expr, expr2);
4380 expr_skip = pet_expr_from_access(isl_map_copy(skip_access));
4381 if (expr_skip) {
4382 expr_skip->acc.write = 1;
4383 expr_skip->acc.read = 0;
4385 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
4386 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, ps->n_stmt++, expr);
4388 scop = pet_scop_from_pet_stmt(ctx, stmt);
4389 scop = scop_add_array(scop, skip_access, ps->ast_context);
4390 isl_map_free(skip_access);
4392 return scop;
4393 error:
4394 isl_map_free(skip_access);
4395 return NULL;
4398 /* Structure that handles the construction of skip conditions
4399 * on sequences of statements.
4401 * scop1 and scop2 represent the two statements that are combined
4403 struct pet_skip_info_seq : public pet_skip_info {
4404 struct pet_scop *scop1, *scop2;
4406 pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4407 struct pet_scop *scop2);
4408 void extract(PetScan *scan, enum pet_skip type);
4409 void extract(PetScan *scan);
4410 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
4411 int offset);
4412 struct pet_scop *add(struct pet_scop *scop, int offset);
4415 /* Initialize a pet_skip_info_seq structure based on
4416 * on the two statements that are going to be combined.
4418 pet_skip_info_seq::pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4419 struct pet_scop *scop2) : pet_skip_info(ctx), scop1(scop1), scop2(scop2)
4421 skip[pet_skip_now] = need_skip_seq(scop1, scop2, pet_skip_now);
4422 equal = skip[pet_skip_now] && skip_equals_skip_later(scop1) &&
4423 skip_equals_skip_later(scop2);
4424 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
4425 need_skip_seq(scop1, scop2, pet_skip_later);
4428 /* If we need to construct a skip condition of the given type,
4429 * then do so now.
4431 void pet_skip_info_seq::extract(PetScan *scan, enum pet_skip type)
4433 if (!skip[type])
4434 return;
4436 access[type] = create_test_access(ctx, scan->n_test++);
4437 scop[type] = extract_skip_seq(scan, isl_map_copy(access[type]),
4438 scop1, scop2, type);
4441 /* Construct the required skip conditions.
4443 void pet_skip_info_seq::extract(PetScan *scan)
4445 extract(scan, pet_skip_now);
4446 extract(scan, pet_skip_later);
4447 if (equal)
4448 drop_skip_later(scop1, scop2);
4451 /* Add the computed skip condition of the give type to "main" and
4452 * add the scop for computing the condition at the given offset (the statement
4453 * number). Within this offset, the condition is computed at position 1
4454 * to ensure that it is computed after the corresponding statement.
4456 * If equal is set, then we only computed a skip condition for pet_skip_now,
4457 * but we also need to set it as main's pet_skip_later.
4459 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *main,
4460 enum pet_skip type, int offset)
4462 isl_set *skip_set;
4464 if (!skip[type])
4465 return main;
4467 skip_set = isl_map_range(access[type]);
4468 access[type] = NULL;
4469 scop[type] = pet_scop_prefix(scop[type], 1);
4470 scop[type] = pet_scop_prefix(scop[type], offset);
4471 main = pet_scop_add_par(ctx, main, scop[type]);
4472 scop[type] = NULL;
4474 if (equal)
4475 main = pet_scop_set_skip(main, pet_skip_later,
4476 isl_set_copy(skip_set));
4478 main = pet_scop_set_skip(main, type, skip_set);
4480 return main;
4483 /* Add the computed skip conditions to "main" and
4484 * add the scops for computing the conditions at the given offset.
4486 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *scop, int offset)
4488 scop = add(scop, pet_skip_now, offset);
4489 scop = add(scop, pet_skip_later, offset);
4491 return scop;
4494 /* Extract a clone of the kill statement in "scop".
4495 * "scop" is expected to have been created from a DeclStmt
4496 * and should have the kill as its first statement.
4498 struct pet_stmt *PetScan::extract_kill(struct pet_scop *scop)
4500 struct pet_expr *kill;
4501 struct pet_stmt *stmt;
4502 isl_map *access;
4504 if (!scop)
4505 return NULL;
4506 if (scop->n_stmt < 1)
4507 isl_die(ctx, isl_error_internal,
4508 "expecting at least one statement", return NULL);
4509 stmt = scop->stmts[0];
4510 if (stmt->body->type != pet_expr_unary ||
4511 stmt->body->op != pet_op_kill)
4512 isl_die(ctx, isl_error_internal,
4513 "expecting kill statement", return NULL);
4515 access = isl_map_copy(stmt->body->args[0]->acc.access);
4516 access = isl_map_reset_tuple_id(access, isl_dim_in);
4517 kill = pet_expr_kill_from_access(access);
4518 return pet_stmt_from_pet_expr(ctx, stmt->line, NULL, n_stmt++, kill);
4521 /* Mark all arrays in "scop" as being exposed.
4523 static struct pet_scop *mark_exposed(struct pet_scop *scop)
4525 if (!scop)
4526 return NULL;
4527 for (int i = 0; i < scop->n_array; ++i)
4528 scop->arrays[i]->exposed = 1;
4529 return scop;
4532 /* Try and construct a pet_scop corresponding to (part of)
4533 * a sequence of statements.
4535 * "block" is set if the sequence respresents the children of
4536 * a compound statement.
4537 * "skip_declarations" is set if we should skip initial declarations
4538 * in the sequence of statements.
4540 * If there are any breaks or continues in the individual statements,
4541 * then we may have to compute a new skip condition.
4542 * This is handled using a pet_skip_info_seq object.
4543 * On initialization, the object checks if skip conditions need
4544 * to be computed. If so, it does so in "extract" and adds them in "add".
4546 * If "block" is set, then we need to insert kill statements at
4547 * the end of the block for any array that has been declared by
4548 * one of the statements in the sequence. Each of these declarations
4549 * results in the construction of a kill statement at the place
4550 * of the declaration, so we simply collect duplicates of
4551 * those kill statements and append these duplicates to the constructed scop.
4553 * If "block" is not set, then any array declared by one of the statements
4554 * in the sequence is marked as being exposed.
4556 struct pet_scop *PetScan::extract(StmtRange stmt_range, bool block,
4557 bool skip_declarations)
4559 pet_scop *scop;
4560 StmtIterator i;
4561 int j;
4562 bool partial_range = false;
4563 set<struct pet_stmt *> kills;
4564 set<struct pet_stmt *>::iterator it;
4566 scop = pet_scop_empty(ctx);
4567 for (i = stmt_range.first, j = 0; i != stmt_range.second; ++i, ++j) {
4568 Stmt *child = *i;
4569 struct pet_scop *scop_i;
4571 if (skip_declarations &&
4572 child->getStmtClass() == Stmt::DeclStmtClass)
4573 continue;
4575 scop_i = extract(child);
4576 if (scop && partial) {
4577 pet_scop_free(scop_i);
4578 break;
4580 pet_skip_info_seq skip(ctx, scop, scop_i);
4581 skip.extract(this);
4582 if (skip)
4583 scop_i = pet_scop_prefix(scop_i, 0);
4584 if (scop_i && child->getStmtClass() == Stmt::DeclStmtClass) {
4585 if (block)
4586 kills.insert(extract_kill(scop_i));
4587 else
4588 scop_i = mark_exposed(scop_i);
4590 scop_i = pet_scop_prefix(scop_i, j);
4591 if (options->autodetect) {
4592 if (scop_i)
4593 scop = pet_scop_add_seq(ctx, scop, scop_i);
4594 else
4595 partial_range = true;
4596 if (scop->n_stmt != 0 && !scop_i)
4597 partial = true;
4598 } else {
4599 scop = pet_scop_add_seq(ctx, scop, scop_i);
4602 scop = skip.add(scop, j);
4604 if (partial)
4605 break;
4608 for (it = kills.begin(); it != kills.end(); ++it) {
4609 pet_scop *scop_j;
4610 scop_j = pet_scop_from_pet_stmt(ctx, *it);
4611 scop_j = pet_scop_prefix(scop_j, j);
4612 scop = pet_scop_add_seq(ctx, scop, scop_j);
4615 if (scop && partial_range)
4616 partial = true;
4618 return scop;
4621 /* Return the file offset of the expansion location of "Loc".
4623 static unsigned getExpansionOffset(SourceManager &SM, SourceLocation Loc)
4625 return SM.getFileOffset(SM.getExpansionLoc(Loc));
4628 /* Check if the scop marked by the user is exactly this Stmt
4629 * or part of this Stmt.
4630 * If so, return a pet_scop corresponding to the marked region.
4631 * Otherwise, return NULL.
4633 struct pet_scop *PetScan::scan(Stmt *stmt)
4635 SourceManager &SM = PP.getSourceManager();
4636 unsigned start_off, end_off;
4638 start_off = getExpansionOffset(SM, stmt->getLocStart());
4639 end_off = getExpansionOffset(SM, stmt->getLocEnd());
4641 if (start_off > loc.end)
4642 return NULL;
4643 if (end_off < loc.start)
4644 return NULL;
4645 if (start_off >= loc.start && end_off <= loc.end) {
4646 return extract(stmt);
4649 StmtIterator start;
4650 for (start = stmt->child_begin(); start != stmt->child_end(); ++start) {
4651 Stmt *child = *start;
4652 if (!child)
4653 continue;
4654 start_off = getExpansionOffset(SM, child->getLocStart());
4655 end_off = getExpansionOffset(SM, child->getLocEnd());
4656 if (start_off < loc.start && end_off > loc.end)
4657 return scan(child);
4658 if (start_off >= loc.start)
4659 break;
4662 StmtIterator end;
4663 for (end = start; end != stmt->child_end(); ++end) {
4664 Stmt *child = *end;
4665 start_off = SM.getFileOffset(child->getLocStart());
4666 if (start_off >= loc.end)
4667 break;
4670 return extract(StmtRange(start, end), false, false);
4673 /* Set the size of index "pos" of "array" to "size".
4674 * In particular, add a constraint of the form
4676 * i_pos < size
4678 * to array->extent and a constraint of the form
4680 * size >= 0
4682 * to array->context.
4684 static struct pet_array *update_size(struct pet_array *array, int pos,
4685 __isl_take isl_pw_aff *size)
4687 isl_set *valid;
4688 isl_set *univ;
4689 isl_set *bound;
4690 isl_space *dim;
4691 isl_aff *aff;
4692 isl_pw_aff *index;
4693 isl_id *id;
4695 valid = isl_pw_aff_nonneg_set(isl_pw_aff_copy(size));
4696 array->context = isl_set_intersect(array->context, valid);
4698 dim = isl_set_get_space(array->extent);
4699 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
4700 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, pos, 1);
4701 univ = isl_set_universe(isl_aff_get_domain_space(aff));
4702 index = isl_pw_aff_alloc(univ, aff);
4704 size = isl_pw_aff_add_dims(size, isl_dim_in,
4705 isl_set_dim(array->extent, isl_dim_set));
4706 id = isl_set_get_tuple_id(array->extent);
4707 size = isl_pw_aff_set_tuple_id(size, isl_dim_in, id);
4708 bound = isl_pw_aff_lt_set(index, size);
4710 array->extent = isl_set_intersect(array->extent, bound);
4712 if (!array->context || !array->extent)
4713 goto error;
4715 return array;
4716 error:
4717 pet_array_free(array);
4718 return NULL;
4721 /* Figure out the size of the array at position "pos" and all
4722 * subsequent positions from "type" and update "array" accordingly.
4724 struct pet_array *PetScan::set_upper_bounds(struct pet_array *array,
4725 const Type *type, int pos)
4727 const ArrayType *atype;
4728 isl_pw_aff *size;
4730 if (!array)
4731 return NULL;
4733 if (type->isPointerType()) {
4734 type = type->getPointeeType().getTypePtr();
4735 return set_upper_bounds(array, type, pos + 1);
4737 if (!type->isArrayType())
4738 return array;
4740 type = type->getCanonicalTypeInternal().getTypePtr();
4741 atype = cast<ArrayType>(type);
4743 if (type->isConstantArrayType()) {
4744 const ConstantArrayType *ca = cast<ConstantArrayType>(atype);
4745 size = extract_affine(ca->getSize());
4746 array = update_size(array, pos, size);
4747 } else if (type->isVariableArrayType()) {
4748 const VariableArrayType *vla = cast<VariableArrayType>(atype);
4749 size = extract_affine(vla->getSizeExpr());
4750 array = update_size(array, pos, size);
4753 type = atype->getElementType().getTypePtr();
4755 return set_upper_bounds(array, type, pos + 1);
4758 /* Is "T" the type of a variable length array with static size?
4760 static bool is_vla_with_static_size(QualType T)
4762 const VariableArrayType *vlatype;
4764 if (!T->isVariableArrayType())
4765 return false;
4766 vlatype = cast<VariableArrayType>(T);
4767 return vlatype->getSizeModifier() == VariableArrayType::Static;
4770 /* Return the type of "decl" as an array.
4772 * In particular, if "decl" is a parameter declaration that
4773 * is a variable length array with a static size, then
4774 * return the original type (i.e., the variable length array).
4775 * Otherwise, return the type of decl.
4777 static QualType get_array_type(ValueDecl *decl)
4779 ParmVarDecl *parm;
4780 QualType T;
4782 parm = dyn_cast<ParmVarDecl>(decl);
4783 if (!parm)
4784 return decl->getType();
4786 T = parm->getOriginalType();
4787 if (!is_vla_with_static_size(T))
4788 return decl->getType();
4789 return T;
4792 /* Construct and return a pet_array corresponding to the variable "decl".
4793 * In particular, initialize array->extent to
4795 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
4797 * and then call set_upper_bounds to set the upper bounds on the indices
4798 * based on the type of the variable.
4800 struct pet_array *PetScan::extract_array(isl_ctx *ctx, ValueDecl *decl)
4802 struct pet_array *array;
4803 QualType qt = get_array_type(decl);
4804 const Type *type = qt.getTypePtr();
4805 int depth = array_depth(type);
4806 QualType base = base_type(qt);
4807 string name;
4808 isl_id *id;
4809 isl_space *dim;
4811 array = isl_calloc_type(ctx, struct pet_array);
4812 if (!array)
4813 return NULL;
4815 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
4816 dim = isl_space_set_alloc(ctx, 0, depth);
4817 dim = isl_space_set_tuple_id(dim, isl_dim_set, id);
4819 array->extent = isl_set_nat_universe(dim);
4821 dim = isl_space_params_alloc(ctx, 0);
4822 array->context = isl_set_universe(dim);
4824 array = set_upper_bounds(array, type, 0);
4825 if (!array)
4826 return NULL;
4828 name = base.getAsString();
4829 array->element_type = strdup(name.c_str());
4830 array->element_size = decl->getASTContext().getTypeInfo(base).first / 8;
4832 return array;
4835 /* Construct a list of pet_arrays, one for each array (or scalar)
4836 * accessed inside "scop", add this list to "scop" and return the result.
4838 * The context of "scop" is updated with the intersection of
4839 * the contexts of all arrays, i.e., constraints on the parameters
4840 * that ensure that the arrays have a valid (non-negative) size.
4842 struct pet_scop *PetScan::scan_arrays(struct pet_scop *scop)
4844 int i;
4845 set<ValueDecl *> arrays;
4846 set<ValueDecl *>::iterator it;
4847 int n_array;
4848 struct pet_array **scop_arrays;
4850 if (!scop)
4851 return NULL;
4853 pet_scop_collect_arrays(scop, arrays);
4854 if (arrays.size() == 0)
4855 return scop;
4857 n_array = scop->n_array;
4859 scop_arrays = isl_realloc_array(ctx, scop->arrays, struct pet_array *,
4860 n_array + arrays.size());
4861 if (!scop_arrays)
4862 goto error;
4863 scop->arrays = scop_arrays;
4865 for (it = arrays.begin(), i = 0; it != arrays.end(); ++it, ++i) {
4866 struct pet_array *array;
4867 scop->arrays[n_array + i] = array = extract_array(ctx, *it);
4868 if (!scop->arrays[n_array + i])
4869 goto error;
4870 scop->n_array++;
4871 scop->context = isl_set_intersect(scop->context,
4872 isl_set_copy(array->context));
4873 if (!scop->context)
4874 goto error;
4877 return scop;
4878 error:
4879 pet_scop_free(scop);
4880 return NULL;
4883 /* Bound all parameters in scop->context to the possible values
4884 * of the corresponding C variable.
4886 static struct pet_scop *add_parameter_bounds(struct pet_scop *scop)
4888 int n;
4890 if (!scop)
4891 return NULL;
4893 n = isl_set_dim(scop->context, isl_dim_param);
4894 for (int i = 0; i < n; ++i) {
4895 isl_id *id;
4896 ValueDecl *decl;
4898 id = isl_set_get_dim_id(scop->context, isl_dim_param, i);
4899 if (is_nested_parameter(id)) {
4900 isl_id_free(id);
4901 isl_die(isl_set_get_ctx(scop->context),
4902 isl_error_internal,
4903 "unresolved nested parameter", goto error);
4905 decl = (ValueDecl *) isl_id_get_user(id);
4906 isl_id_free(id);
4908 scop->context = set_parameter_bounds(scop->context, i, decl);
4910 if (!scop->context)
4911 goto error;
4914 return scop;
4915 error:
4916 pet_scop_free(scop);
4917 return NULL;
4920 /* Construct a pet_scop from the given function.
4922 struct pet_scop *PetScan::scan(FunctionDecl *fd)
4924 pet_scop *scop;
4925 Stmt *stmt;
4927 stmt = fd->getBody();
4929 if (options->autodetect)
4930 scop = extract(stmt, true);
4931 else
4932 scop = scan(stmt);
4933 scop = pet_scop_detect_parameter_accesses(scop);
4934 scop = scan_arrays(scop);
4935 scop = add_parameter_bounds(scop);
4936 scop = pet_scop_gist(scop, value_bounds);
4938 return scop;