parse.c: extract_double: fix return type
[pet.git] / scan.cc
blob2f3de081e482b25e476da637454626d054823280
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 <clang/AST/ASTContext.h>
39 #include <clang/AST/ASTDiagnostic.h>
40 #include <clang/AST/Expr.h>
41 #include <clang/AST/RecursiveASTVisitor.h>
43 #include <isl/id.h>
44 #include <isl/space.h>
45 #include <isl/aff.h>
46 #include <isl/set.h>
48 #include "options.h"
49 #include "scan.h"
50 #include "scop.h"
51 #include "scop_plus.h"
53 #include "config.h"
55 using namespace std;
56 using namespace clang;
58 #if defined(DECLREFEXPR_CREATE_REQUIRES_BOOL)
59 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
61 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
62 SourceLocation(), var, false, var->getInnerLocStart(),
63 var->getType(), VK_LValue);
65 #elif defined(DECLREFEXPR_CREATE_REQUIRES_SOURCELOCATION)
66 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
68 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
69 SourceLocation(), var, var->getInnerLocStart(), var->getType(),
70 VK_LValue);
72 #else
73 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
75 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
76 var, var->getInnerLocStart(), var->getType(), VK_LValue);
78 #endif
80 /* Check if the element type corresponding to the given array type
81 * has a const qualifier.
83 static bool const_base(QualType qt)
85 const Type *type = qt.getTypePtr();
87 if (type->isPointerType())
88 return const_base(type->getPointeeType());
89 if (type->isArrayType()) {
90 const ArrayType *atype;
91 type = type->getCanonicalTypeInternal().getTypePtr();
92 atype = cast<ArrayType>(type);
93 return const_base(atype->getElementType());
96 return qt.isConstQualified();
99 /* Mark "decl" as having an unknown value in "assigned_value".
101 * If no (known or unknown) value was assigned to "decl" before,
102 * then it may have been treated as a parameter before and may
103 * therefore appear in a value assigned to another variable.
104 * If so, this assignment needs to be turned into an unknown value too.
106 static void clear_assignment(map<ValueDecl *, isl_pw_aff *> &assigned_value,
107 ValueDecl *decl)
109 map<ValueDecl *, isl_pw_aff *>::iterator it;
111 it = assigned_value.find(decl);
113 assigned_value[decl] = NULL;
115 if (it == assigned_value.end())
116 return;
118 for (it = assigned_value.begin(); it != assigned_value.end(); ++it) {
119 isl_pw_aff *pa = it->second;
120 int nparam = isl_pw_aff_dim(pa, isl_dim_param);
122 for (int i = 0; i < nparam; ++i) {
123 isl_id *id;
125 if (!isl_pw_aff_has_dim_id(pa, isl_dim_param, i))
126 continue;
127 id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
128 if (isl_id_get_user(id) == decl)
129 it->second = NULL;
130 isl_id_free(id);
135 /* Look for any assignments to scalar variables in part of the parse
136 * tree and set assigned_value to NULL for each of them.
137 * Also reset assigned_value if the address of a scalar variable
138 * is being taken. As an exception, if the address is passed to a function
139 * that is declared to receive a const pointer, then assigned_value is
140 * not reset.
142 * This ensures that we won't use any previously stored value
143 * in the current subtree and its parents.
145 struct clear_assignments : RecursiveASTVisitor<clear_assignments> {
146 map<ValueDecl *, isl_pw_aff *> &assigned_value;
147 set<UnaryOperator *> skip;
149 clear_assignments(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
150 assigned_value(assigned_value) {}
152 /* Check for "address of" operators whose value is passed
153 * to a const pointer argument and add them to "skip", so that
154 * we can skip them in VisitUnaryOperator.
156 bool VisitCallExpr(CallExpr *expr) {
157 FunctionDecl *fd;
158 fd = expr->getDirectCallee();
159 if (!fd)
160 return true;
161 for (int i = 0; i < expr->getNumArgs(); ++i) {
162 Expr *arg = expr->getArg(i);
163 UnaryOperator *op;
164 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
165 ImplicitCastExpr *ice;
166 ice = cast<ImplicitCastExpr>(arg);
167 arg = ice->getSubExpr();
169 if (arg->getStmtClass() != Stmt::UnaryOperatorClass)
170 continue;
171 op = cast<UnaryOperator>(arg);
172 if (op->getOpcode() != UO_AddrOf)
173 continue;
174 if (const_base(fd->getParamDecl(i)->getType()))
175 skip.insert(op);
177 return true;
180 bool VisitUnaryOperator(UnaryOperator *expr) {
181 Expr *arg;
182 DeclRefExpr *ref;
183 ValueDecl *decl;
185 switch (expr->getOpcode()) {
186 case UO_AddrOf:
187 case UO_PostInc:
188 case UO_PostDec:
189 case UO_PreInc:
190 case UO_PreDec:
191 break;
192 default:
193 return true;
195 if (skip.find(expr) != skip.end())
196 return true;
198 arg = expr->getSubExpr();
199 if (arg->getStmtClass() != Stmt::DeclRefExprClass)
200 return true;
201 ref = cast<DeclRefExpr>(arg);
202 decl = ref->getDecl();
203 clear_assignment(assigned_value, decl);
204 return true;
207 bool VisitBinaryOperator(BinaryOperator *expr) {
208 Expr *lhs;
209 DeclRefExpr *ref;
210 ValueDecl *decl;
212 if (!expr->isAssignmentOp())
213 return true;
214 lhs = expr->getLHS();
215 if (lhs->getStmtClass() != Stmt::DeclRefExprClass)
216 return true;
217 ref = cast<DeclRefExpr>(lhs);
218 decl = ref->getDecl();
219 clear_assignment(assigned_value, decl);
220 return true;
224 /* Keep a copy of the currently assigned values.
226 * Any variable that is assigned a value inside the current scope
227 * is removed again when we leave the scope (either because it wasn't
228 * stored in the cache or because it has a different value in the cache).
230 struct assigned_value_cache {
231 map<ValueDecl *, isl_pw_aff *> &assigned_value;
232 map<ValueDecl *, isl_pw_aff *> cache;
234 assigned_value_cache(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
235 assigned_value(assigned_value), cache(assigned_value) {}
236 ~assigned_value_cache() {
237 map<ValueDecl *, isl_pw_aff *>::iterator it = cache.begin();
238 for (it = assigned_value.begin(); it != assigned_value.end();
239 ++it) {
240 if (!it->second ||
241 (cache.find(it->first) != cache.end() &&
242 cache[it->first] != it->second))
243 cache[it->first] = NULL;
245 assigned_value = cache;
249 /* Insert an expression into the collection of expressions,
250 * provided it is not already in there.
251 * The isl_pw_affs are freed in the destructor.
253 void PetScan::insert_expression(__isl_take isl_pw_aff *expr)
255 std::set<isl_pw_aff *>::iterator it;
257 if (expressions.find(expr) == expressions.end())
258 expressions.insert(expr);
259 else
260 isl_pw_aff_free(expr);
263 PetScan::~PetScan()
265 std::set<isl_pw_aff *>::iterator it;
267 for (it = expressions.begin(); it != expressions.end(); ++it)
268 isl_pw_aff_free(*it);
270 isl_union_map_free(value_bounds);
273 /* Called if we found something we (currently) cannot handle.
274 * We'll provide more informative warnings later.
276 * We only actually complain if autodetect is false.
278 void PetScan::unsupported(Stmt *stmt, const char *msg)
280 if (options->autodetect)
281 return;
283 SourceLocation loc = stmt->getLocStart();
284 DiagnosticsEngine &diag = PP.getDiagnostics();
285 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
286 msg ? msg : "unsupported");
287 DiagnosticBuilder B = diag.Report(loc, id) << stmt->getSourceRange();
290 /* Extract an integer from "expr" and store it in "v".
292 int PetScan::extract_int(IntegerLiteral *expr, isl_int *v)
294 const Type *type = expr->getType().getTypePtr();
295 int is_signed = type->hasSignedIntegerRepresentation();
297 if (is_signed) {
298 int64_t i = expr->getValue().getSExtValue();
299 isl_int_set_si(*v, i);
300 } else {
301 uint64_t i = expr->getValue().getZExtValue();
302 isl_int_set_ui(*v, i);
305 return 0;
308 /* Extract an integer from "expr" and store it in "v".
309 * Return -1 if "expr" does not (obviously) represent an integer.
311 int PetScan::extract_int(clang::ParenExpr *expr, isl_int *v)
313 return extract_int(expr->getSubExpr(), v);
316 /* Extract an integer from "expr" and store it in "v".
317 * Return -1 if "expr" does not (obviously) represent an integer.
319 int PetScan::extract_int(clang::Expr *expr, isl_int *v)
321 if (expr->getStmtClass() == Stmt::IntegerLiteralClass)
322 return extract_int(cast<IntegerLiteral>(expr), v);
323 if (expr->getStmtClass() == Stmt::ParenExprClass)
324 return extract_int(cast<ParenExpr>(expr), v);
326 unsupported(expr);
327 return -1;
330 /* Extract an affine expression from the IntegerLiteral "expr".
332 __isl_give isl_pw_aff *PetScan::extract_affine(IntegerLiteral *expr)
334 isl_space *dim = isl_space_params_alloc(ctx, 0);
335 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
336 isl_aff *aff = isl_aff_zero_on_domain(ls);
337 isl_set *dom = isl_set_universe(dim);
338 isl_int v;
340 isl_int_init(v);
341 extract_int(expr, &v);
342 aff = isl_aff_add_constant(aff, v);
343 isl_int_clear(v);
345 return isl_pw_aff_alloc(dom, aff);
348 /* Extract an affine expression from the APInt "val".
350 __isl_give isl_pw_aff *PetScan::extract_affine(const llvm::APInt &val)
352 isl_space *dim = isl_space_params_alloc(ctx, 0);
353 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
354 isl_aff *aff = isl_aff_zero_on_domain(ls);
355 isl_set *dom = isl_set_universe(dim);
356 isl_int v;
358 isl_int_init(v);
359 isl_int_set_ui(v, val.getZExtValue());
360 aff = isl_aff_add_constant(aff, v);
361 isl_int_clear(v);
363 return isl_pw_aff_alloc(dom, aff);
366 __isl_give isl_pw_aff *PetScan::extract_affine(ImplicitCastExpr *expr)
368 return extract_affine(expr->getSubExpr());
371 static unsigned get_type_size(ValueDecl *decl)
373 return decl->getASTContext().getIntWidth(decl->getType());
376 /* Bound parameter "pos" of "set" to the possible values of "decl".
378 static __isl_give isl_set *set_parameter_bounds(__isl_take isl_set *set,
379 unsigned pos, ValueDecl *decl)
381 unsigned width;
382 isl_int v;
384 isl_int_init(v);
386 width = get_type_size(decl);
387 if (decl->getType()->isUnsignedIntegerType()) {
388 set = isl_set_lower_bound_si(set, isl_dim_param, pos, 0);
389 isl_int_set_si(v, 1);
390 isl_int_mul_2exp(v, v, width);
391 isl_int_sub_ui(v, v, 1);
392 set = isl_set_upper_bound(set, isl_dim_param, pos, v);
393 } else {
394 isl_int_set_si(v, 1);
395 isl_int_mul_2exp(v, v, width - 1);
396 isl_int_sub_ui(v, v, 1);
397 set = isl_set_upper_bound(set, isl_dim_param, pos, v);
398 isl_int_neg(v, v);
399 isl_int_sub_ui(v, v, 1);
400 set = isl_set_lower_bound(set, isl_dim_param, pos, v);
403 isl_int_clear(v);
405 return set;
408 /* Extract an affine expression from the DeclRefExpr "expr".
410 * If the variable has been assigned a value, then we check whether
411 * we know what (affine) value was assigned.
412 * If so, we return this value. Otherwise we convert "expr"
413 * to an extra parameter (provided nesting_enabled is set).
415 * Otherwise, we simply return an expression that is equal
416 * to a parameter corresponding to the referenced variable.
418 __isl_give isl_pw_aff *PetScan::extract_affine(DeclRefExpr *expr)
420 ValueDecl *decl = expr->getDecl();
421 const Type *type = decl->getType().getTypePtr();
422 isl_id *id;
423 isl_space *dim;
424 isl_aff *aff;
425 isl_set *dom;
427 if (!type->isIntegerType()) {
428 unsupported(expr);
429 return NULL;
432 if (assigned_value.find(decl) != assigned_value.end()) {
433 if (assigned_value[decl])
434 return isl_pw_aff_copy(assigned_value[decl]);
435 else
436 return nested_access(expr);
439 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
440 dim = isl_space_params_alloc(ctx, 1);
442 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
444 dom = isl_set_universe(isl_space_copy(dim));
445 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
446 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
448 return isl_pw_aff_alloc(dom, aff);
451 /* Extract an affine expression from an integer division operation.
452 * In particular, if "expr" is lhs/rhs, then return
454 * lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs)
456 * The second argument (rhs) is required to be a (positive) integer constant.
458 __isl_give isl_pw_aff *PetScan::extract_affine_div(BinaryOperator *expr)
460 int is_cst;
461 isl_pw_aff *rhs, *lhs;
463 rhs = extract_affine(expr->getRHS());
464 is_cst = isl_pw_aff_is_cst(rhs);
465 if (is_cst < 0 || !is_cst) {
466 isl_pw_aff_free(rhs);
467 if (!is_cst)
468 unsupported(expr);
469 return NULL;
472 lhs = extract_affine(expr->getLHS());
474 return isl_pw_aff_tdiv_q(lhs, rhs);
477 /* Extract an affine expression from a modulo operation.
478 * In particular, if "expr" is lhs/rhs, then return
480 * lhs - rhs * (lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs))
482 * The second argument (rhs) is required to be a (positive) integer constant.
484 __isl_give isl_pw_aff *PetScan::extract_affine_mod(BinaryOperator *expr)
486 int is_cst;
487 isl_pw_aff *rhs, *lhs;
489 rhs = extract_affine(expr->getRHS());
490 is_cst = isl_pw_aff_is_cst(rhs);
491 if (is_cst < 0 || !is_cst) {
492 isl_pw_aff_free(rhs);
493 if (!is_cst)
494 unsupported(expr);
495 return NULL;
498 lhs = extract_affine(expr->getLHS());
500 return isl_pw_aff_tdiv_r(lhs, rhs);
503 /* Extract an affine expression from a multiplication operation.
504 * This is only allowed if at least one of the two arguments
505 * is a (piecewise) constant.
507 __isl_give isl_pw_aff *PetScan::extract_affine_mul(BinaryOperator *expr)
509 isl_pw_aff *lhs;
510 isl_pw_aff *rhs;
512 lhs = extract_affine(expr->getLHS());
513 rhs = extract_affine(expr->getRHS());
515 if (!isl_pw_aff_is_cst(lhs) && !isl_pw_aff_is_cst(rhs)) {
516 isl_pw_aff_free(lhs);
517 isl_pw_aff_free(rhs);
518 unsupported(expr);
519 return NULL;
522 return isl_pw_aff_mul(lhs, rhs);
525 /* Extract an affine expression from an addition or subtraction operation.
527 __isl_give isl_pw_aff *PetScan::extract_affine_add(BinaryOperator *expr)
529 isl_pw_aff *lhs;
530 isl_pw_aff *rhs;
532 lhs = extract_affine(expr->getLHS());
533 rhs = extract_affine(expr->getRHS());
535 switch (expr->getOpcode()) {
536 case BO_Add:
537 return isl_pw_aff_add(lhs, rhs);
538 case BO_Sub:
539 return isl_pw_aff_sub(lhs, rhs);
540 default:
541 isl_pw_aff_free(lhs);
542 isl_pw_aff_free(rhs);
543 return NULL;
548 /* Compute
550 * pwaff mod 2^width
552 static __isl_give isl_pw_aff *wrap(__isl_take isl_pw_aff *pwaff,
553 unsigned width)
555 isl_int mod;
557 isl_int_init(mod);
558 isl_int_set_si(mod, 1);
559 isl_int_mul_2exp(mod, mod, width);
561 pwaff = isl_pw_aff_mod(pwaff, mod);
563 isl_int_clear(mod);
565 return pwaff;
568 /* Limit the domain of "pwaff" to those elements where the function
569 * value satisfies
571 * 2^{width-1} <= pwaff < 2^{width-1}
573 static __isl_give isl_pw_aff *avoid_overflow(__isl_take isl_pw_aff *pwaff,
574 unsigned width)
576 isl_int v;
577 isl_space *space = isl_pw_aff_get_domain_space(pwaff);
578 isl_local_space *ls = isl_local_space_from_space(space);
579 isl_aff *bound;
580 isl_set *dom;
581 isl_pw_aff *b;
583 isl_int_init(v);
584 isl_int_set_si(v, 1);
585 isl_int_mul_2exp(v, v, width - 1);
587 bound = isl_aff_zero_on_domain(ls);
588 bound = isl_aff_add_constant(bound, v);
589 b = isl_pw_aff_from_aff(bound);
591 dom = isl_pw_aff_lt_set(isl_pw_aff_copy(pwaff), isl_pw_aff_copy(b));
592 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
594 b = isl_pw_aff_neg(b);
595 dom = isl_pw_aff_ge_set(isl_pw_aff_copy(pwaff), b);
596 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
598 isl_int_clear(v);
600 return pwaff;
603 /* Handle potential overflows on signed computations.
605 * If options->signed_overflow is set to PET_OVERFLOW_AVOID,
606 * the we adjust the domain of "pa" to avoid overflows.
608 __isl_give isl_pw_aff *PetScan::signed_overflow(__isl_take isl_pw_aff *pa,
609 unsigned width)
611 if (options->signed_overflow == PET_OVERFLOW_AVOID)
612 pa = avoid_overflow(pa, width);
614 return pa;
617 /* Return the piecewise affine expression "set ? 1 : 0" defined on "dom".
619 static __isl_give isl_pw_aff *indicator_function(__isl_take isl_set *set,
620 __isl_take isl_set *dom)
622 isl_pw_aff *pa;
623 pa = isl_set_indicator_function(set);
624 pa = isl_pw_aff_intersect_domain(pa, dom);
625 return pa;
628 /* Extract an affine expression from some binary operations.
629 * If the result of the expression is unsigned, then we wrap it
630 * based on the size of the type. Otherwise, we ensure that
631 * no overflow occurs.
633 __isl_give isl_pw_aff *PetScan::extract_affine(BinaryOperator *expr)
635 isl_pw_aff *res;
636 unsigned width;
638 switch (expr->getOpcode()) {
639 case BO_Add:
640 case BO_Sub:
641 res = extract_affine_add(expr);
642 break;
643 case BO_Div:
644 res = extract_affine_div(expr);
645 break;
646 case BO_Rem:
647 res = extract_affine_mod(expr);
648 break;
649 case BO_Mul:
650 res = extract_affine_mul(expr);
651 break;
652 case BO_LT:
653 case BO_LE:
654 case BO_GT:
655 case BO_GE:
656 case BO_EQ:
657 case BO_NE:
658 case BO_LAnd:
659 case BO_LOr:
660 return extract_condition(expr);
661 default:
662 unsupported(expr);
663 return NULL;
666 width = ast_context.getIntWidth(expr->getType());
667 if (expr->getType()->isUnsignedIntegerType())
668 res = wrap(res, width);
669 else
670 res = signed_overflow(res, width);
672 return res;
675 /* Extract an affine expression from a negation operation.
677 __isl_give isl_pw_aff *PetScan::extract_affine(UnaryOperator *expr)
679 if (expr->getOpcode() == UO_Minus)
680 return isl_pw_aff_neg(extract_affine(expr->getSubExpr()));
681 if (expr->getOpcode() == UO_LNot)
682 return extract_condition(expr);
684 unsupported(expr);
685 return NULL;
688 __isl_give isl_pw_aff *PetScan::extract_affine(ParenExpr *expr)
690 return extract_affine(expr->getSubExpr());
693 /* Extract an affine expression from some special function calls.
694 * In particular, we handle "min", "max", "ceild" and "floord".
695 * In case of the latter two, the second argument needs to be
696 * a (positive) integer constant.
698 __isl_give isl_pw_aff *PetScan::extract_affine(CallExpr *expr)
700 FunctionDecl *fd;
701 string name;
702 isl_pw_aff *aff1, *aff2;
704 fd = expr->getDirectCallee();
705 if (!fd) {
706 unsupported(expr);
707 return NULL;
710 name = fd->getDeclName().getAsString();
711 if (!(expr->getNumArgs() == 2 && name == "min") &&
712 !(expr->getNumArgs() == 2 && name == "max") &&
713 !(expr->getNumArgs() == 2 && name == "floord") &&
714 !(expr->getNumArgs() == 2 && name == "ceild")) {
715 unsupported(expr);
716 return NULL;
719 if (name == "min" || name == "max") {
720 aff1 = extract_affine(expr->getArg(0));
721 aff2 = extract_affine(expr->getArg(1));
723 if (name == "min")
724 aff1 = isl_pw_aff_min(aff1, aff2);
725 else
726 aff1 = isl_pw_aff_max(aff1, aff2);
727 } else if (name == "floord" || name == "ceild") {
728 isl_int v;
729 Expr *arg2 = expr->getArg(1);
731 if (arg2->getStmtClass() != Stmt::IntegerLiteralClass) {
732 unsupported(expr);
733 return NULL;
735 aff1 = extract_affine(expr->getArg(0));
736 isl_int_init(v);
737 extract_int(cast<IntegerLiteral>(arg2), &v);
738 aff1 = isl_pw_aff_scale_down(aff1, v);
739 isl_int_clear(v);
740 if (name == "floord")
741 aff1 = isl_pw_aff_floor(aff1);
742 else
743 aff1 = isl_pw_aff_ceil(aff1);
744 } else {
745 unsupported(expr);
746 return NULL;
749 return aff1;
752 /* This method is called when we come across an access that is
753 * nested in what is supposed to be an affine expression.
754 * If nesting is allowed, we return a new parameter that corresponds
755 * to this nested access. Otherwise, we simply complain.
757 * Note that we currently don't allow nested accesses themselves
758 * to contain any nested accesses, so we check if we can extract
759 * the access without any nesting and complain if we can't.
761 * The new parameter is resolved in resolve_nested.
763 isl_pw_aff *PetScan::nested_access(Expr *expr)
765 isl_id *id;
766 isl_space *dim;
767 isl_aff *aff;
768 isl_set *dom;
769 isl_map *access;
771 if (!nesting_enabled) {
772 unsupported(expr);
773 return NULL;
776 allow_nested = false;
777 access = extract_access(expr);
778 allow_nested = true;
779 if (!access) {
780 unsupported(expr);
781 return NULL;
783 isl_map_free(access);
785 id = isl_id_alloc(ctx, NULL, expr);
786 dim = isl_space_params_alloc(ctx, 1);
788 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
790 dom = isl_set_universe(isl_space_copy(dim));
791 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
792 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
794 return isl_pw_aff_alloc(dom, aff);
797 /* Affine expressions are not supposed to contain array accesses,
798 * but if nesting is allowed, we return a parameter corresponding
799 * to the array access.
801 __isl_give isl_pw_aff *PetScan::extract_affine(ArraySubscriptExpr *expr)
803 return nested_access(expr);
806 /* Extract an affine expression from a conditional operation.
808 __isl_give isl_pw_aff *PetScan::extract_affine(ConditionalOperator *expr)
810 isl_pw_aff *cond, *lhs, *rhs, *res;
812 cond = extract_condition(expr->getCond());
813 lhs = extract_affine(expr->getTrueExpr());
814 rhs = extract_affine(expr->getFalseExpr());
816 return isl_pw_aff_cond(cond, lhs, rhs);
819 /* Extract an affine expression, if possible, from "expr".
820 * Otherwise return NULL.
822 __isl_give isl_pw_aff *PetScan::extract_affine(Expr *expr)
824 switch (expr->getStmtClass()) {
825 case Stmt::ImplicitCastExprClass:
826 return extract_affine(cast<ImplicitCastExpr>(expr));
827 case Stmt::IntegerLiteralClass:
828 return extract_affine(cast<IntegerLiteral>(expr));
829 case Stmt::DeclRefExprClass:
830 return extract_affine(cast<DeclRefExpr>(expr));
831 case Stmt::BinaryOperatorClass:
832 return extract_affine(cast<BinaryOperator>(expr));
833 case Stmt::UnaryOperatorClass:
834 return extract_affine(cast<UnaryOperator>(expr));
835 case Stmt::ParenExprClass:
836 return extract_affine(cast<ParenExpr>(expr));
837 case Stmt::CallExprClass:
838 return extract_affine(cast<CallExpr>(expr));
839 case Stmt::ArraySubscriptExprClass:
840 return extract_affine(cast<ArraySubscriptExpr>(expr));
841 case Stmt::ConditionalOperatorClass:
842 return extract_affine(cast<ConditionalOperator>(expr));
843 default:
844 unsupported(expr);
846 return NULL;
849 __isl_give isl_map *PetScan::extract_access(ImplicitCastExpr *expr)
851 return extract_access(expr->getSubExpr());
854 /* Return the depth of an array of the given type.
856 static int array_depth(const Type *type)
858 if (type->isPointerType())
859 return 1 + array_depth(type->getPointeeType().getTypePtr());
860 if (type->isArrayType()) {
861 const ArrayType *atype;
862 type = type->getCanonicalTypeInternal().getTypePtr();
863 atype = cast<ArrayType>(type);
864 return 1 + array_depth(atype->getElementType().getTypePtr());
866 return 0;
869 /* Return the element type of the given array type.
871 static QualType base_type(QualType qt)
873 const Type *type = qt.getTypePtr();
875 if (type->isPointerType())
876 return base_type(type->getPointeeType());
877 if (type->isArrayType()) {
878 const ArrayType *atype;
879 type = type->getCanonicalTypeInternal().getTypePtr();
880 atype = cast<ArrayType>(type);
881 return base_type(atype->getElementType());
883 return qt;
886 /* Extract an access relation from a reference to a variable.
887 * If the variable has name "A" and its type corresponds to an
888 * array of depth d, then the returned access relation is of the
889 * form
891 * { [] -> A[i_1,...,i_d] }
893 __isl_give isl_map *PetScan::extract_access(DeclRefExpr *expr)
895 return extract_access(expr->getDecl());
898 /* Extract an access relation from a variable.
899 * If the variable has name "A" and its type corresponds to an
900 * array of depth d, then the returned access relation is of the
901 * form
903 * { [] -> A[i_1,...,i_d] }
905 __isl_give isl_map *PetScan::extract_access(ValueDecl *decl)
907 int depth = array_depth(decl->getType().getTypePtr());
908 isl_id *id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
909 isl_space *dim = isl_space_alloc(ctx, 0, 0, depth);
910 isl_map *access_rel;
912 dim = isl_space_set_tuple_id(dim, isl_dim_out, id);
914 access_rel = isl_map_universe(dim);
916 return access_rel;
919 /* Extract an access relation from an integer contant.
920 * If the value of the constant is "v", then the returned access relation
921 * is
923 * { [] -> [v] }
925 __isl_give isl_map *PetScan::extract_access(IntegerLiteral *expr)
927 return isl_map_from_range(isl_set_from_pw_aff(extract_affine(expr)));
930 /* Try and extract an access relation from the given Expr.
931 * Return NULL if it doesn't work out.
933 __isl_give isl_map *PetScan::extract_access(Expr *expr)
935 switch (expr->getStmtClass()) {
936 case Stmt::ImplicitCastExprClass:
937 return extract_access(cast<ImplicitCastExpr>(expr));
938 case Stmt::DeclRefExprClass:
939 return extract_access(cast<DeclRefExpr>(expr));
940 case Stmt::ArraySubscriptExprClass:
941 return extract_access(cast<ArraySubscriptExpr>(expr));
942 case Stmt::IntegerLiteralClass:
943 return extract_access(cast<IntegerLiteral>(expr));
944 default:
945 unsupported(expr);
947 return NULL;
950 /* Assign the affine expression "index" to the output dimension "pos" of "map",
951 * restrict the domain to those values that result in a non-negative index
952 * and return the result.
954 __isl_give isl_map *set_index(__isl_take isl_map *map, int pos,
955 __isl_take isl_pw_aff *index)
957 isl_map *index_map;
958 int len = isl_map_dim(map, isl_dim_out);
959 isl_id *id;
960 isl_set *domain;
962 domain = isl_pw_aff_nonneg_set(isl_pw_aff_copy(index));
963 index = isl_pw_aff_intersect_domain(index, domain);
964 index_map = isl_map_from_range(isl_set_from_pw_aff(index));
965 index_map = isl_map_insert_dims(index_map, isl_dim_out, 0, pos);
966 index_map = isl_map_add_dims(index_map, isl_dim_out, len - pos - 1);
967 id = isl_map_get_tuple_id(map, isl_dim_out);
968 index_map = isl_map_set_tuple_id(index_map, isl_dim_out, id);
970 map = isl_map_intersect(map, index_map);
972 return map;
975 /* Extract an access relation from the given array subscript expression.
976 * If nesting is allowed in general, then we turn it on while
977 * examining the index expression.
979 * We first extract an access relation from the base.
980 * This will result in an access relation with a range that corresponds
981 * to the array being accessed and with earlier indices filled in already.
982 * We then extract the current index and fill that in as well.
983 * The position of the current index is based on the type of base.
984 * If base is the actual array variable, then the depth of this type
985 * will be the same as the depth of the array and we will fill in
986 * the first array index.
987 * Otherwise, the depth of the base type will be smaller and we will fill
988 * in a later index.
990 __isl_give isl_map *PetScan::extract_access(ArraySubscriptExpr *expr)
992 Expr *base = expr->getBase();
993 Expr *idx = expr->getIdx();
994 isl_pw_aff *index;
995 isl_map *base_access;
996 isl_map *access;
997 int depth = array_depth(base->getType().getTypePtr());
998 int pos;
999 bool save_nesting = nesting_enabled;
1001 nesting_enabled = allow_nested;
1003 base_access = extract_access(base);
1004 index = extract_affine(idx);
1006 nesting_enabled = save_nesting;
1008 pos = isl_map_dim(base_access, isl_dim_out) - depth;
1009 access = set_index(base_access, pos, index);
1011 return access;
1014 /* Check if "expr" calls function "minmax" with two arguments and if so
1015 * make lhs and rhs refer to these two arguments.
1017 static bool is_minmax(Expr *expr, const char *minmax, Expr *&lhs, Expr *&rhs)
1019 CallExpr *call;
1020 FunctionDecl *fd;
1021 string name;
1023 if (expr->getStmtClass() != Stmt::CallExprClass)
1024 return false;
1026 call = cast<CallExpr>(expr);
1027 fd = call->getDirectCallee();
1028 if (!fd)
1029 return false;
1031 if (call->getNumArgs() != 2)
1032 return false;
1034 name = fd->getDeclName().getAsString();
1035 if (name != minmax)
1036 return false;
1038 lhs = call->getArg(0);
1039 rhs = call->getArg(1);
1041 return true;
1044 /* Check if "expr" is of the form min(lhs, rhs) and if so make
1045 * lhs and rhs refer to the two arguments.
1047 static bool is_min(Expr *expr, Expr *&lhs, Expr *&rhs)
1049 return is_minmax(expr, "min", lhs, rhs);
1052 /* Check if "expr" is of the form max(lhs, rhs) and if so make
1053 * lhs and rhs refer to the two arguments.
1055 static bool is_max(Expr *expr, Expr *&lhs, Expr *&rhs)
1057 return is_minmax(expr, "max", lhs, rhs);
1060 /* Return "lhs && rhs", defined on the shared definition domain.
1062 static __isl_give isl_pw_aff *pw_aff_and(__isl_take isl_pw_aff *lhs,
1063 __isl_take isl_pw_aff *rhs)
1065 isl_set *cond;
1066 isl_set *dom;
1068 dom = isl_set_intersect(isl_pw_aff_domain(isl_pw_aff_copy(lhs)),
1069 isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1070 cond = isl_set_intersect(isl_pw_aff_non_zero_set(lhs),
1071 isl_pw_aff_non_zero_set(rhs));
1072 return indicator_function(cond, dom);
1075 /* Return "lhs && rhs", with shortcut semantics.
1076 * That is, if lhs is false, then the result is defined even if rhs is not.
1077 * In practice, we compute lhs ? rhs : lhs.
1079 static __isl_give isl_pw_aff *pw_aff_and_then(__isl_take isl_pw_aff *lhs,
1080 __isl_take isl_pw_aff *rhs)
1082 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), rhs, lhs);
1085 /* Return "lhs || rhs", with shortcut semantics.
1086 * That is, if lhs is true, then the result is defined even if rhs is not.
1087 * In practice, we compute lhs ? lhs : rhs.
1089 static __isl_give isl_pw_aff *pw_aff_or_else(__isl_take isl_pw_aff *lhs,
1090 __isl_take isl_pw_aff *rhs)
1092 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), lhs, rhs);
1095 /* Extract an affine expressions representing the comparison "LHS op RHS"
1096 * "comp" is the original statement that "LHS op RHS" is derived from
1097 * and is used for diagnostics.
1099 * If the comparison is of the form
1101 * a <= min(b,c)
1103 * then the expression is constructed as the conjunction of
1104 * the comparisons
1106 * a <= b and a <= c
1108 * A similar optimization is performed for max(a,b) <= c.
1109 * We do this because that will lead to simpler representations
1110 * of the expression.
1111 * If isl is ever enhanced to explicitly deal with min and max expressions,
1112 * this optimization can be removed.
1114 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperatorKind op,
1115 Expr *LHS, Expr *RHS, Stmt *comp)
1117 isl_pw_aff *lhs;
1118 isl_pw_aff *rhs;
1119 isl_pw_aff *res;
1120 isl_set *cond;
1121 isl_set *dom;
1123 if (op == BO_GT)
1124 return extract_comparison(BO_LT, RHS, LHS, comp);
1125 if (op == BO_GE)
1126 return extract_comparison(BO_LE, RHS, LHS, comp);
1128 if (op == BO_LT || op == BO_LE) {
1129 Expr *expr1, *expr2;
1130 if (is_min(RHS, expr1, expr2)) {
1131 lhs = extract_comparison(op, LHS, expr1, comp);
1132 rhs = extract_comparison(op, LHS, expr2, comp);
1133 return pw_aff_and(lhs, rhs);
1135 if (is_max(LHS, expr1, expr2)) {
1136 lhs = extract_comparison(op, expr1, RHS, comp);
1137 rhs = extract_comparison(op, expr2, RHS, comp);
1138 return pw_aff_and(lhs, rhs);
1142 lhs = extract_affine(LHS);
1143 rhs = extract_affine(RHS);
1145 dom = isl_pw_aff_domain(isl_pw_aff_copy(lhs));
1146 dom = isl_set_intersect(dom, isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1148 switch (op) {
1149 case BO_LT:
1150 cond = isl_pw_aff_lt_set(lhs, rhs);
1151 break;
1152 case BO_LE:
1153 cond = isl_pw_aff_le_set(lhs, rhs);
1154 break;
1155 case BO_EQ:
1156 cond = isl_pw_aff_eq_set(lhs, rhs);
1157 break;
1158 case BO_NE:
1159 cond = isl_pw_aff_ne_set(lhs, rhs);
1160 break;
1161 default:
1162 isl_pw_aff_free(lhs);
1163 isl_pw_aff_free(rhs);
1164 isl_set_free(dom);
1165 unsupported(comp);
1166 return NULL;
1169 cond = isl_set_coalesce(cond);
1170 res = indicator_function(cond, dom);
1172 return res;
1175 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperator *comp)
1177 return extract_comparison(comp->getOpcode(), comp->getLHS(),
1178 comp->getRHS(), comp);
1181 /* Extract an affine expression representing the negation (logical not)
1182 * of a subexpression.
1184 __isl_give isl_pw_aff *PetScan::extract_boolean(UnaryOperator *op)
1186 isl_set *set_cond, *dom;
1187 isl_pw_aff *cond, *res;
1189 cond = extract_condition(op->getSubExpr());
1191 dom = isl_pw_aff_domain(isl_pw_aff_copy(cond));
1193 set_cond = isl_pw_aff_zero_set(cond);
1195 res = indicator_function(set_cond, dom);
1197 return res;
1200 /* Extract an affine expression representing the disjunction (logical or)
1201 * or conjunction (logical and) of two subexpressions.
1203 __isl_give isl_pw_aff *PetScan::extract_boolean(BinaryOperator *comp)
1205 isl_pw_aff *lhs, *rhs;
1207 lhs = extract_condition(comp->getLHS());
1208 rhs = extract_condition(comp->getRHS());
1210 switch (comp->getOpcode()) {
1211 case BO_LAnd:
1212 return pw_aff_and_then(lhs, rhs);
1213 case BO_LOr:
1214 return pw_aff_or_else(lhs, rhs);
1215 default:
1216 isl_pw_aff_free(lhs);
1217 isl_pw_aff_free(rhs);
1220 unsupported(comp);
1221 return NULL;
1224 __isl_give isl_pw_aff *PetScan::extract_condition(UnaryOperator *expr)
1226 switch (expr->getOpcode()) {
1227 case UO_LNot:
1228 return extract_boolean(expr);
1229 default:
1230 unsupported(expr);
1231 return NULL;
1235 /* Extract the affine expression "expr != 0 ? 1 : 0".
1237 __isl_give isl_pw_aff *PetScan::extract_implicit_condition(Expr *expr)
1239 isl_pw_aff *res;
1240 isl_set *set, *dom;
1242 res = extract_affine(expr);
1244 dom = isl_pw_aff_domain(isl_pw_aff_copy(res));
1245 set = isl_pw_aff_non_zero_set(res);
1247 res = indicator_function(set, dom);
1249 return res;
1252 /* Extract an affine expression from a boolean expression.
1253 * In particular, return the expression "expr ? 1 : 0".
1255 * If the expression doesn't look like a condition, we assume it
1256 * is an affine expression and return the condition "expr != 0 ? 1 : 0".
1258 __isl_give isl_pw_aff *PetScan::extract_condition(Expr *expr)
1260 BinaryOperator *comp;
1262 if (!expr) {
1263 isl_set *u = isl_set_universe(isl_space_params_alloc(ctx, 0));
1264 return indicator_function(u, isl_set_copy(u));
1267 if (expr->getStmtClass() == Stmt::ParenExprClass)
1268 return extract_condition(cast<ParenExpr>(expr)->getSubExpr());
1270 if (expr->getStmtClass() == Stmt::UnaryOperatorClass)
1271 return extract_condition(cast<UnaryOperator>(expr));
1273 if (expr->getStmtClass() != Stmt::BinaryOperatorClass)
1274 return extract_implicit_condition(expr);
1276 comp = cast<BinaryOperator>(expr);
1277 switch (comp->getOpcode()) {
1278 case BO_LT:
1279 case BO_LE:
1280 case BO_GT:
1281 case BO_GE:
1282 case BO_EQ:
1283 case BO_NE:
1284 return extract_comparison(comp);
1285 case BO_LAnd:
1286 case BO_LOr:
1287 return extract_boolean(comp);
1288 default:
1289 return extract_implicit_condition(expr);
1293 static enum pet_op_type UnaryOperatorKind2pet_op_type(UnaryOperatorKind kind)
1295 switch (kind) {
1296 case UO_Minus:
1297 return pet_op_minus;
1298 case UO_PostInc:
1299 return pet_op_post_inc;
1300 case UO_PostDec:
1301 return pet_op_post_dec;
1302 case UO_PreInc:
1303 return pet_op_pre_inc;
1304 case UO_PreDec:
1305 return pet_op_pre_dec;
1306 default:
1307 return pet_op_last;
1311 static enum pet_op_type BinaryOperatorKind2pet_op_type(BinaryOperatorKind kind)
1313 switch (kind) {
1314 case BO_AddAssign:
1315 return pet_op_add_assign;
1316 case BO_SubAssign:
1317 return pet_op_sub_assign;
1318 case BO_MulAssign:
1319 return pet_op_mul_assign;
1320 case BO_DivAssign:
1321 return pet_op_div_assign;
1322 case BO_Assign:
1323 return pet_op_assign;
1324 case BO_Add:
1325 return pet_op_add;
1326 case BO_Sub:
1327 return pet_op_sub;
1328 case BO_Mul:
1329 return pet_op_mul;
1330 case BO_Div:
1331 return pet_op_div;
1332 case BO_Rem:
1333 return pet_op_mod;
1334 case BO_EQ:
1335 return pet_op_eq;
1336 case BO_LE:
1337 return pet_op_le;
1338 case BO_LT:
1339 return pet_op_lt;
1340 case BO_GT:
1341 return pet_op_gt;
1342 default:
1343 return pet_op_last;
1347 /* Construct a pet_expr representing a unary operator expression.
1349 struct pet_expr *PetScan::extract_expr(UnaryOperator *expr)
1351 struct pet_expr *arg;
1352 enum pet_op_type op;
1354 op = UnaryOperatorKind2pet_op_type(expr->getOpcode());
1355 if (op == pet_op_last) {
1356 unsupported(expr);
1357 return NULL;
1360 arg = extract_expr(expr->getSubExpr());
1362 if (expr->isIncrementDecrementOp() &&
1363 arg && arg->type == pet_expr_access) {
1364 mark_write(arg);
1365 arg->acc.read = 1;
1368 return pet_expr_new_unary(ctx, op, arg);
1371 /* Mark the given access pet_expr as a write.
1372 * If a scalar is being accessed, then mark its value
1373 * as unknown in assigned_value.
1375 void PetScan::mark_write(struct pet_expr *access)
1377 isl_id *id;
1378 ValueDecl *decl;
1380 if (!access)
1381 return;
1383 access->acc.write = 1;
1384 access->acc.read = 0;
1386 if (isl_map_dim(access->acc.access, isl_dim_out) != 0)
1387 return;
1389 id = isl_map_get_tuple_id(access->acc.access, isl_dim_out);
1390 decl = (ValueDecl *) isl_id_get_user(id);
1391 clear_assignment(assigned_value, decl);
1392 isl_id_free(id);
1395 /* Assign "rhs" to "lhs".
1397 * In particular, if "lhs" is a scalar variable, then mark
1398 * the variable as having been assigned. If, furthermore, "rhs"
1399 * is an affine expression, then keep track of this value in assigned_value
1400 * so that we can plug it in when we later come across the same variable.
1402 void PetScan::assign(struct pet_expr *lhs, Expr *rhs)
1404 isl_id *id;
1405 ValueDecl *decl;
1406 isl_pw_aff *pa;
1408 if (!lhs)
1409 return;
1410 if (lhs->type != pet_expr_access)
1411 return;
1412 if (isl_map_dim(lhs->acc.access, isl_dim_out) != 0)
1413 return;
1415 id = isl_map_get_tuple_id(lhs->acc.access, isl_dim_out);
1416 decl = (ValueDecl *) isl_id_get_user(id);
1417 isl_id_free(id);
1419 pa = try_extract_affine(rhs);
1420 clear_assignment(assigned_value, decl);
1421 if (!pa)
1422 return;
1423 assigned_value[decl] = pa;
1424 insert_expression(pa);
1427 /* Construct a pet_expr representing a binary operator expression.
1429 * If the top level operator is an assignment and the LHS is an access,
1430 * then we mark that access as a write. If the operator is a compound
1431 * assignment, the access is marked as both a read and a write.
1433 * If "expr" assigns something to a scalar variable, then we mark
1434 * the variable as having been assigned. If, furthermore, the expression
1435 * is affine, then keep track of this value in assigned_value
1436 * so that we can plug it in when we later come across the same variable.
1438 struct pet_expr *PetScan::extract_expr(BinaryOperator *expr)
1440 struct pet_expr *lhs, *rhs;
1441 enum pet_op_type op;
1443 op = BinaryOperatorKind2pet_op_type(expr->getOpcode());
1444 if (op == pet_op_last) {
1445 unsupported(expr);
1446 return NULL;
1449 lhs = extract_expr(expr->getLHS());
1450 rhs = extract_expr(expr->getRHS());
1452 if (expr->isAssignmentOp() && lhs && lhs->type == pet_expr_access) {
1453 mark_write(lhs);
1454 if (expr->isCompoundAssignmentOp())
1455 lhs->acc.read = 1;
1458 if (expr->getOpcode() == BO_Assign)
1459 assign(lhs, expr->getRHS());
1461 return pet_expr_new_binary(ctx, op, lhs, rhs);
1464 /* Construct a pet_scop with a single statement killing the entire
1465 * array "array".
1467 struct pet_scop *PetScan::kill(Stmt *stmt, struct pet_array *array)
1469 isl_map *access;
1470 struct pet_expr *expr;
1472 if (!array)
1473 return NULL;
1474 access = isl_map_from_range(isl_set_copy(array->extent));
1475 expr = pet_expr_kill_from_access(access);
1476 return extract(stmt, expr);
1479 /* Construct a pet_scop for a (single) variable declaration.
1481 * The scop contains the variable being declared (as an array)
1482 * and a statement killing the array.
1484 * If the variable is initialized in the AST, then the scop
1485 * also contains an assignment to the variable.
1487 struct pet_scop *PetScan::extract(DeclStmt *stmt)
1489 Decl *decl;
1490 VarDecl *vd;
1491 struct pet_expr *lhs, *rhs, *pe;
1492 struct pet_scop *scop_decl, *scop;
1493 struct pet_array *array;
1495 if (!stmt->isSingleDecl()) {
1496 unsupported(stmt);
1497 return NULL;
1500 decl = stmt->getSingleDecl();
1501 vd = cast<VarDecl>(decl);
1503 array = extract_array(ctx, vd);
1504 if (array)
1505 array->declared = 1;
1506 scop_decl = kill(stmt, array);
1507 scop_decl = pet_scop_add_array(scop_decl, array);
1509 if (!vd->getInit())
1510 return scop_decl;
1512 lhs = pet_expr_from_access(extract_access(vd));
1513 rhs = extract_expr(vd->getInit());
1515 mark_write(lhs);
1516 assign(lhs, vd->getInit());
1518 pe = pet_expr_new_binary(ctx, pet_op_assign, lhs, rhs);
1519 scop = extract(stmt, pe);
1521 scop_decl = pet_scop_prefix(scop_decl, 0);
1522 scop = pet_scop_prefix(scop, 1);
1524 scop = pet_scop_add_seq(ctx, scop_decl, scop);
1526 return scop;
1529 /* Construct a pet_expr representing a conditional operation.
1531 * We first try to extract the condition as an affine expression.
1532 * If that fails, we construct a pet_expr tree representing the condition.
1534 struct pet_expr *PetScan::extract_expr(ConditionalOperator *expr)
1536 struct pet_expr *cond, *lhs, *rhs;
1537 isl_pw_aff *pa;
1539 pa = try_extract_affine(expr->getCond());
1540 if (pa) {
1541 isl_set *test = isl_set_from_pw_aff(pa);
1542 cond = pet_expr_from_access(isl_map_from_range(test));
1543 } else
1544 cond = extract_expr(expr->getCond());
1545 lhs = extract_expr(expr->getTrueExpr());
1546 rhs = extract_expr(expr->getFalseExpr());
1548 return pet_expr_new_ternary(ctx, cond, lhs, rhs);
1551 struct pet_expr *PetScan::extract_expr(ImplicitCastExpr *expr)
1553 return extract_expr(expr->getSubExpr());
1556 /* Construct a pet_expr representing a floating point value.
1558 struct pet_expr *PetScan::extract_expr(FloatingLiteral *expr)
1560 return pet_expr_new_double(ctx, expr->getValueAsApproximateDouble());
1563 /* Extract an access relation from "expr" and then convert it into
1564 * a pet_expr.
1566 struct pet_expr *PetScan::extract_access_expr(Expr *expr)
1568 isl_map *access;
1569 struct pet_expr *pe;
1571 access = extract_access(expr);
1573 pe = pet_expr_from_access(access);
1575 return pe;
1578 struct pet_expr *PetScan::extract_expr(ParenExpr *expr)
1580 return extract_expr(expr->getSubExpr());
1583 /* Construct a pet_expr representing a function call.
1585 * If we are passing along a pointer to an array element
1586 * or an entire row or even higher dimensional slice of an array,
1587 * then the function being called may write into the array.
1589 * We assume here that if the function is declared to take a pointer
1590 * to a const type, then the function will perform a read
1591 * and that otherwise, it will perform a write.
1593 struct pet_expr *PetScan::extract_expr(CallExpr *expr)
1595 struct pet_expr *res = NULL;
1596 FunctionDecl *fd;
1597 string name;
1599 fd = expr->getDirectCallee();
1600 if (!fd) {
1601 unsupported(expr);
1602 return NULL;
1605 name = fd->getDeclName().getAsString();
1606 res = pet_expr_new_call(ctx, name.c_str(), expr->getNumArgs());
1607 if (!res)
1608 return NULL;
1610 for (int i = 0; i < expr->getNumArgs(); ++i) {
1611 Expr *arg = expr->getArg(i);
1612 int is_addr = 0;
1613 pet_expr *main_arg;
1615 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
1616 ImplicitCastExpr *ice = cast<ImplicitCastExpr>(arg);
1617 arg = ice->getSubExpr();
1619 if (arg->getStmtClass() == Stmt::UnaryOperatorClass) {
1620 UnaryOperator *op = cast<UnaryOperator>(arg);
1621 if (op->getOpcode() == UO_AddrOf) {
1622 is_addr = 1;
1623 arg = op->getSubExpr();
1626 res->args[i] = PetScan::extract_expr(arg);
1627 main_arg = res->args[i];
1628 if (is_addr)
1629 res->args[i] = pet_expr_new_unary(ctx,
1630 pet_op_address_of, res->args[i]);
1631 if (!res->args[i])
1632 goto error;
1633 if (arg->getStmtClass() == Stmt::ArraySubscriptExprClass &&
1634 array_depth(arg->getType().getTypePtr()) > 0)
1635 is_addr = 1;
1636 if (is_addr && main_arg->type == pet_expr_access) {
1637 ParmVarDecl *parm;
1638 if (!fd->hasPrototype()) {
1639 unsupported(expr, "prototype required");
1640 goto error;
1642 parm = fd->getParamDecl(i);
1643 if (!const_base(parm->getType()))
1644 mark_write(main_arg);
1648 return res;
1649 error:
1650 pet_expr_free(res);
1651 return NULL;
1654 /* Try and onstruct a pet_expr representing "expr".
1656 struct pet_expr *PetScan::extract_expr(Expr *expr)
1658 switch (expr->getStmtClass()) {
1659 case Stmt::UnaryOperatorClass:
1660 return extract_expr(cast<UnaryOperator>(expr));
1661 case Stmt::CompoundAssignOperatorClass:
1662 case Stmt::BinaryOperatorClass:
1663 return extract_expr(cast<BinaryOperator>(expr));
1664 case Stmt::ImplicitCastExprClass:
1665 return extract_expr(cast<ImplicitCastExpr>(expr));
1666 case Stmt::ArraySubscriptExprClass:
1667 case Stmt::DeclRefExprClass:
1668 case Stmt::IntegerLiteralClass:
1669 return extract_access_expr(expr);
1670 case Stmt::FloatingLiteralClass:
1671 return extract_expr(cast<FloatingLiteral>(expr));
1672 case Stmt::ParenExprClass:
1673 return extract_expr(cast<ParenExpr>(expr));
1674 case Stmt::ConditionalOperatorClass:
1675 return extract_expr(cast<ConditionalOperator>(expr));
1676 case Stmt::CallExprClass:
1677 return extract_expr(cast<CallExpr>(expr));
1678 default:
1679 unsupported(expr);
1681 return NULL;
1684 /* Check if the given initialization statement is an assignment.
1685 * If so, return that assignment. Otherwise return NULL.
1687 BinaryOperator *PetScan::initialization_assignment(Stmt *init)
1689 BinaryOperator *ass;
1691 if (init->getStmtClass() != Stmt::BinaryOperatorClass)
1692 return NULL;
1694 ass = cast<BinaryOperator>(init);
1695 if (ass->getOpcode() != BO_Assign)
1696 return NULL;
1698 return ass;
1701 /* Check if the given initialization statement is a declaration
1702 * of a single variable.
1703 * If so, return that declaration. Otherwise return NULL.
1705 Decl *PetScan::initialization_declaration(Stmt *init)
1707 DeclStmt *decl;
1709 if (init->getStmtClass() != Stmt::DeclStmtClass)
1710 return NULL;
1712 decl = cast<DeclStmt>(init);
1714 if (!decl->isSingleDecl())
1715 return NULL;
1717 return decl->getSingleDecl();
1720 /* Given the assignment operator in the initialization of a for loop,
1721 * extract the induction variable, i.e., the (integer)variable being
1722 * assigned.
1724 ValueDecl *PetScan::extract_induction_variable(BinaryOperator *init)
1726 Expr *lhs;
1727 DeclRefExpr *ref;
1728 ValueDecl *decl;
1729 const Type *type;
1731 lhs = init->getLHS();
1732 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1733 unsupported(init);
1734 return NULL;
1737 ref = cast<DeclRefExpr>(lhs);
1738 decl = ref->getDecl();
1739 type = decl->getType().getTypePtr();
1741 if (!type->isIntegerType()) {
1742 unsupported(lhs);
1743 return NULL;
1746 return decl;
1749 /* Given the initialization statement of a for loop and the single
1750 * declaration in this initialization statement,
1751 * extract the induction variable, i.e., the (integer) variable being
1752 * declared.
1754 VarDecl *PetScan::extract_induction_variable(Stmt *init, Decl *decl)
1756 VarDecl *vd;
1758 vd = cast<VarDecl>(decl);
1760 const QualType type = vd->getType();
1761 if (!type->isIntegerType()) {
1762 unsupported(init);
1763 return NULL;
1766 if (!vd->getInit()) {
1767 unsupported(init);
1768 return NULL;
1771 return vd;
1774 /* Check that op is of the form iv++ or iv--.
1775 * Return an affine expression "1" or "-1" accordingly.
1777 __isl_give isl_pw_aff *PetScan::extract_unary_increment(
1778 clang::UnaryOperator *op, clang::ValueDecl *iv)
1780 Expr *sub;
1781 DeclRefExpr *ref;
1782 isl_space *space;
1783 isl_aff *aff;
1785 if (!op->isIncrementDecrementOp()) {
1786 unsupported(op);
1787 return NULL;
1790 sub = op->getSubExpr();
1791 if (sub->getStmtClass() != Stmt::DeclRefExprClass) {
1792 unsupported(op);
1793 return NULL;
1796 ref = cast<DeclRefExpr>(sub);
1797 if (ref->getDecl() != iv) {
1798 unsupported(op);
1799 return NULL;
1802 space = isl_space_params_alloc(ctx, 0);
1803 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
1805 if (op->isIncrementOp())
1806 aff = isl_aff_add_constant_si(aff, 1);
1807 else
1808 aff = isl_aff_add_constant_si(aff, -1);
1810 return isl_pw_aff_from_aff(aff);
1813 /* If the isl_pw_aff on which isl_pw_aff_foreach_piece is called
1814 * has a single constant expression, then put this constant in *user.
1815 * The caller is assumed to have checked that this function will
1816 * be called exactly once.
1818 static int extract_cst(__isl_take isl_set *set, __isl_take isl_aff *aff,
1819 void *user)
1821 isl_int *inc = (isl_int *)user;
1822 int res = 0;
1824 if (isl_aff_is_cst(aff))
1825 isl_aff_get_constant(aff, inc);
1826 else
1827 res = -1;
1829 isl_set_free(set);
1830 isl_aff_free(aff);
1832 return res;
1835 /* Check if op is of the form
1837 * iv = iv + inc
1839 * and return inc as an affine expression.
1841 * We extract an affine expression from the RHS, subtract iv and return
1842 * the result.
1844 __isl_give isl_pw_aff *PetScan::extract_binary_increment(BinaryOperator *op,
1845 clang::ValueDecl *iv)
1847 Expr *lhs;
1848 DeclRefExpr *ref;
1849 isl_id *id;
1850 isl_space *dim;
1851 isl_aff *aff;
1852 isl_pw_aff *val;
1854 if (op->getOpcode() != BO_Assign) {
1855 unsupported(op);
1856 return NULL;
1859 lhs = op->getLHS();
1860 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1861 unsupported(op);
1862 return NULL;
1865 ref = cast<DeclRefExpr>(lhs);
1866 if (ref->getDecl() != iv) {
1867 unsupported(op);
1868 return NULL;
1871 val = extract_affine(op->getRHS());
1873 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
1875 dim = isl_space_params_alloc(ctx, 1);
1876 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
1877 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
1878 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
1880 val = isl_pw_aff_sub(val, isl_pw_aff_from_aff(aff));
1882 return val;
1885 /* Check that op is of the form iv += cst or iv -= cst
1886 * and return an affine expression corresponding oto cst or -cst accordingly.
1888 __isl_give isl_pw_aff *PetScan::extract_compound_increment(
1889 CompoundAssignOperator *op, clang::ValueDecl *iv)
1891 Expr *lhs;
1892 DeclRefExpr *ref;
1893 bool neg = false;
1894 isl_pw_aff *val;
1895 BinaryOperatorKind opcode;
1897 opcode = op->getOpcode();
1898 if (opcode != BO_AddAssign && opcode != BO_SubAssign) {
1899 unsupported(op);
1900 return NULL;
1902 if (opcode == BO_SubAssign)
1903 neg = true;
1905 lhs = op->getLHS();
1906 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1907 unsupported(op);
1908 return NULL;
1911 ref = cast<DeclRefExpr>(lhs);
1912 if (ref->getDecl() != iv) {
1913 unsupported(op);
1914 return NULL;
1917 val = extract_affine(op->getRHS());
1918 if (neg)
1919 val = isl_pw_aff_neg(val);
1921 return val;
1924 /* Check that the increment of the given for loop increments
1925 * (or decrements) the induction variable "iv" and return
1926 * the increment as an affine expression if successful.
1928 __isl_give isl_pw_aff *PetScan::extract_increment(clang::ForStmt *stmt,
1929 ValueDecl *iv)
1931 Stmt *inc = stmt->getInc();
1933 if (!inc) {
1934 unsupported(stmt);
1935 return NULL;
1938 if (inc->getStmtClass() == Stmt::UnaryOperatorClass)
1939 return extract_unary_increment(cast<UnaryOperator>(inc), iv);
1940 if (inc->getStmtClass() == Stmt::CompoundAssignOperatorClass)
1941 return extract_compound_increment(
1942 cast<CompoundAssignOperator>(inc), iv);
1943 if (inc->getStmtClass() == Stmt::BinaryOperatorClass)
1944 return extract_binary_increment(cast<BinaryOperator>(inc), iv);
1946 unsupported(inc);
1947 return NULL;
1950 /* Embed the given iteration domain in an extra outer loop
1951 * with induction variable "var".
1952 * If this variable appeared as a parameter in the constraints,
1953 * it is replaced by the new outermost dimension.
1955 static __isl_give isl_set *embed(__isl_take isl_set *set,
1956 __isl_take isl_id *var)
1958 int pos;
1960 set = isl_set_insert_dims(set, isl_dim_set, 0, 1);
1961 pos = isl_set_find_dim_by_id(set, isl_dim_param, var);
1962 if (pos >= 0) {
1963 set = isl_set_equate(set, isl_dim_param, pos, isl_dim_set, 0);
1964 set = isl_set_project_out(set, isl_dim_param, pos, 1);
1967 isl_id_free(var);
1968 return set;
1971 /* Return those elements in the space of "cond" that come after
1972 * (based on "sign") an element in "cond".
1974 static __isl_give isl_set *after(__isl_take isl_set *cond, int sign)
1976 isl_map *previous_to_this;
1978 if (sign > 0)
1979 previous_to_this = isl_map_lex_lt(isl_set_get_space(cond));
1980 else
1981 previous_to_this = isl_map_lex_gt(isl_set_get_space(cond));
1983 cond = isl_set_apply(cond, previous_to_this);
1985 return cond;
1988 /* Create the infinite iteration domain
1990 * { [id] : id >= 0 }
1992 * If "scop" has an affine skip of type pet_skip_later,
1993 * then remove those iterations i that have an earlier iteration
1994 * where the skip condition is satisfied, meaning that iteration i
1995 * is not executed.
1996 * Since we are dealing with a loop without loop iterator,
1997 * the skip condition cannot refer to the current loop iterator and
1998 * so effectively, the returned set is of the form
2000 * { [0]; [id] : id >= 1 and not skip }
2002 static __isl_give isl_set *infinite_domain(__isl_take isl_id *id,
2003 struct pet_scop *scop)
2005 isl_ctx *ctx = isl_id_get_ctx(id);
2006 isl_set *domain;
2007 isl_set *skip;
2009 domain = isl_set_nat_universe(isl_space_set_alloc(ctx, 0, 1));
2010 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, id);
2012 if (!pet_scop_has_affine_skip(scop, pet_skip_later))
2013 return domain;
2015 skip = pet_scop_get_skip(scop, pet_skip_later);
2016 skip = isl_set_fix_si(skip, isl_dim_set, 0, 1);
2017 skip = isl_set_params(skip);
2018 skip = embed(skip, isl_id_copy(id));
2019 skip = isl_set_intersect(skip , isl_set_copy(domain));
2020 domain = isl_set_subtract(domain, after(skip, 1));
2022 return domain;
2025 /* Create an identity mapping on the space containing "domain".
2027 static __isl_give isl_map *identity_map(__isl_keep isl_set *domain)
2029 isl_space *space;
2030 isl_map *id;
2032 space = isl_space_map_from_set(isl_set_get_space(domain));
2033 id = isl_map_identity(space);
2035 return id;
2038 /* Add a filter to "scop" that imposes that it is only executed
2039 * when "break_access" has a zero value for all previous iterations
2040 * of "domain".
2042 * The input "break_access" has a zero-dimensional domain and range.
2044 static struct pet_scop *scop_add_break(struct pet_scop *scop,
2045 __isl_take isl_map *break_access, __isl_take isl_set *domain, int sign)
2047 isl_ctx *ctx = isl_set_get_ctx(domain);
2048 isl_id *id_test;
2049 isl_map *prev;
2051 id_test = isl_map_get_tuple_id(break_access, isl_dim_out);
2052 break_access = isl_map_add_dims(break_access, isl_dim_in, 1);
2053 break_access = isl_map_add_dims(break_access, isl_dim_out, 1);
2054 break_access = isl_map_intersect_range(break_access, domain);
2055 break_access = isl_map_set_tuple_id(break_access, isl_dim_out, id_test);
2056 if (sign > 0)
2057 prev = isl_map_lex_gt_first(isl_map_get_space(break_access), 1);
2058 else
2059 prev = isl_map_lex_lt_first(isl_map_get_space(break_access), 1);
2060 break_access = isl_map_intersect(break_access, prev);
2061 scop = pet_scop_filter(scop, break_access, 0);
2062 scop = pet_scop_merge_filters(scop);
2064 return scop;
2067 /* Construct a pet_scop for an infinite loop around the given body.
2069 * We extract a pet_scop for the body and then embed it in a loop with
2070 * iteration domain
2072 * { [t] : t >= 0 }
2074 * and schedule
2076 * { [t] -> [t] }
2078 * If the body contains any break, then it is taken into
2079 * account in infinite_domain (if the skip condition is affine)
2080 * or in scop_add_break (if the skip condition is not affine).
2082 struct pet_scop *PetScan::extract_infinite_loop(Stmt *body)
2084 isl_id *id;
2085 isl_set *domain;
2086 isl_map *ident;
2087 isl_map *access;
2088 struct pet_scop *scop;
2089 bool has_var_break;
2091 scop = extract(body);
2092 if (!scop)
2093 return NULL;
2095 id = isl_id_alloc(ctx, "t", NULL);
2096 domain = infinite_domain(isl_id_copy(id), scop);
2097 ident = identity_map(domain);
2099 has_var_break = pet_scop_has_var_skip(scop, pet_skip_later);
2100 if (has_var_break)
2101 access = pet_scop_get_skip_map(scop, pet_skip_later);
2103 scop = pet_scop_embed(scop, isl_set_copy(domain),
2104 isl_map_copy(ident), ident, id);
2105 if (has_var_break)
2106 scop = scop_add_break(scop, access, domain, 1);
2107 else
2108 isl_set_free(domain);
2110 return scop;
2113 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
2115 * for (;;)
2116 * body
2119 struct pet_scop *PetScan::extract_infinite_for(ForStmt *stmt)
2121 return extract_infinite_loop(stmt->getBody());
2124 /* Create an access to a virtual array representing the result
2125 * of a condition.
2126 * Unlike other accessed data, the id of the array is NULL as
2127 * there is no ValueDecl in the program corresponding to the virtual
2128 * array.
2129 * The array starts out as a scalar, but grows along with the
2130 * statement writing to the array in pet_scop_embed.
2132 static __isl_give isl_map *create_test_access(isl_ctx *ctx, int test_nr)
2134 isl_space *dim = isl_space_alloc(ctx, 0, 0, 0);
2135 isl_id *id;
2136 char name[50];
2138 snprintf(name, sizeof(name), "__pet_test_%d", test_nr);
2139 id = isl_id_alloc(ctx, name, NULL);
2140 dim = isl_space_set_tuple_id(dim, isl_dim_out, id);
2141 return isl_map_universe(dim);
2144 /* Add an array with the given extent ("access") to the list
2145 * of arrays in "scop" and return the extended pet_scop.
2146 * The array is marked as attaining values 0 and 1 only and
2147 * as each element being assigned at most once.
2149 static struct pet_scop *scop_add_array(struct pet_scop *scop,
2150 __isl_keep isl_map *access, clang::ASTContext &ast_ctx)
2152 isl_ctx *ctx = isl_map_get_ctx(access);
2153 isl_space *dim;
2154 struct pet_array *array;
2156 if (!scop)
2157 return NULL;
2158 if (!ctx)
2159 goto error;
2161 array = isl_calloc_type(ctx, struct pet_array);
2162 if (!array)
2163 goto error;
2165 array->extent = isl_map_range(isl_map_copy(access));
2166 dim = isl_space_params_alloc(ctx, 0);
2167 array->context = isl_set_universe(dim);
2168 dim = isl_space_set_alloc(ctx, 0, 1);
2169 array->value_bounds = isl_set_universe(dim);
2170 array->value_bounds = isl_set_lower_bound_si(array->value_bounds,
2171 isl_dim_set, 0, 0);
2172 array->value_bounds = isl_set_upper_bound_si(array->value_bounds,
2173 isl_dim_set, 0, 1);
2174 array->element_type = strdup("int");
2175 array->element_size = ast_ctx.getTypeInfo(ast_ctx.IntTy).first / 8;
2176 array->uniquely_defined = 1;
2178 if (!array->extent || !array->context)
2179 array = pet_array_free(array);
2181 scop = pet_scop_add_array(scop, array);
2183 return scop;
2184 error:
2185 pet_scop_free(scop);
2186 return NULL;
2189 /* Construct a pet_scop for a while loop of the form
2191 * while (pa)
2192 * body
2194 * In particular, construct a scop for an infinite loop around body and
2195 * intersect the domain with the affine expression.
2196 * Note that this intersection may result in an empty loop.
2198 struct pet_scop *PetScan::extract_affine_while(__isl_take isl_pw_aff *pa,
2199 Stmt *body)
2201 struct pet_scop *scop;
2202 isl_set *dom;
2203 isl_set *valid;
2205 valid = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2206 dom = isl_pw_aff_non_zero_set(pa);
2207 scop = extract_infinite_loop(body);
2208 scop = pet_scop_restrict(scop, dom);
2209 scop = pet_scop_restrict_context(scop, valid);
2211 return scop;
2214 /* Construct a scop for a while, given the scops for the condition
2215 * and the body, the filter access and the iteration domain of
2216 * the while loop.
2218 * In particular, the scop for the condition is filtered to depend
2219 * on "test_access" evaluating to true for all previous iterations
2220 * of the loop, while the scop for the body is filtered to depend
2221 * on "test_access" evaluating to true for all iterations up to the
2222 * current iteration.
2224 * These filtered scops are then combined into a single scop.
2226 * "sign" is positive if the iterator increases and negative
2227 * if it decreases.
2229 static struct pet_scop *scop_add_while(struct pet_scop *scop_cond,
2230 struct pet_scop *scop_body, __isl_take isl_map *test_access,
2231 __isl_take isl_set *domain, int sign)
2233 isl_ctx *ctx = isl_set_get_ctx(domain);
2234 isl_id *id_test;
2235 isl_map *prev;
2237 id_test = isl_map_get_tuple_id(test_access, isl_dim_out);
2238 test_access = isl_map_add_dims(test_access, isl_dim_in, 1);
2239 test_access = isl_map_add_dims(test_access, isl_dim_out, 1);
2240 test_access = isl_map_intersect_range(test_access, domain);
2241 test_access = isl_map_set_tuple_id(test_access, isl_dim_out, id_test);
2242 if (sign > 0)
2243 prev = isl_map_lex_ge_first(isl_map_get_space(test_access), 1);
2244 else
2245 prev = isl_map_lex_le_first(isl_map_get_space(test_access), 1);
2246 test_access = isl_map_intersect(test_access, prev);
2247 scop_body = pet_scop_filter(scop_body, isl_map_copy(test_access), 1);
2248 if (sign > 0)
2249 prev = isl_map_lex_gt_first(isl_map_get_space(test_access), 1);
2250 else
2251 prev = isl_map_lex_lt_first(isl_map_get_space(test_access), 1);
2252 test_access = isl_map_intersect(test_access, prev);
2253 scop_cond = pet_scop_filter(scop_cond, test_access, 1);
2255 return pet_scop_add_seq(ctx, scop_cond, scop_body);
2258 /* Check if the while loop is of the form
2260 * while (affine expression)
2261 * body
2263 * If so, call extract_affine_while to construct a scop.
2265 * Otherwise, construct a generic while scop, with iteration domain
2266 * { [t] : t >= 0 }. The scop consists of two parts, one for
2267 * evaluating the condition and one for the body.
2268 * The schedule is adjusted to reflect that the condition is evaluated
2269 * before the body is executed and the body is filtered to depend
2270 * on the result of the condition evaluating to true on all iterations
2271 * up to the current iteration, while the evaluation the condition itself
2272 * is filtered to depend on the result of the condition evaluating to true
2273 * on all previous iterations.
2274 * The context of the scop representing the body is dropped
2275 * because we don't know how many times the body will be executed,
2276 * if at all.
2278 * If the body contains any break, then it is taken into
2279 * account in infinite_domain (if the skip condition is affine)
2280 * or in scop_add_break (if the skip condition is not affine).
2282 struct pet_scop *PetScan::extract(WhileStmt *stmt)
2284 Expr *cond;
2285 isl_id *id;
2286 isl_map *test_access;
2287 isl_set *domain;
2288 isl_map *ident;
2289 isl_pw_aff *pa;
2290 struct pet_scop *scop, *scop_body;
2291 bool has_var_break;
2292 isl_map *break_access;
2294 cond = stmt->getCond();
2295 if (!cond) {
2296 unsupported(stmt);
2297 return NULL;
2300 clear_assignments clear(assigned_value);
2301 clear.TraverseStmt(stmt->getBody());
2303 pa = try_extract_affine_condition(cond);
2304 if (pa)
2305 return extract_affine_while(pa, stmt->getBody());
2307 if (!allow_nested) {
2308 unsupported(stmt);
2309 return NULL;
2312 test_access = create_test_access(ctx, n_test++);
2313 scop = extract_non_affine_condition(cond, isl_map_copy(test_access));
2314 scop = scop_add_array(scop, test_access, ast_context);
2315 scop_body = extract(stmt->getBody());
2317 id = isl_id_alloc(ctx, "t", NULL);
2318 domain = infinite_domain(isl_id_copy(id), scop_body);
2319 ident = identity_map(domain);
2321 has_var_break = pet_scop_has_var_skip(scop_body, pet_skip_later);
2322 if (has_var_break)
2323 break_access = pet_scop_get_skip_map(scop_body, pet_skip_later);
2325 scop = pet_scop_prefix(scop, 0);
2326 scop = pet_scop_embed(scop, isl_set_copy(domain), isl_map_copy(ident),
2327 isl_map_copy(ident), isl_id_copy(id));
2328 scop_body = pet_scop_reset_context(scop_body);
2329 scop_body = pet_scop_prefix(scop_body, 1);
2330 scop_body = pet_scop_embed(scop_body, isl_set_copy(domain),
2331 isl_map_copy(ident), ident, id);
2333 if (has_var_break) {
2334 scop = scop_add_break(scop, isl_map_copy(break_access),
2335 isl_set_copy(domain), 1);
2336 scop_body = scop_add_break(scop_body, break_access,
2337 isl_set_copy(domain), 1);
2339 scop = scop_add_while(scop, scop_body, test_access, domain, 1);
2341 return scop;
2344 /* Check whether "cond" expresses a simple loop bound
2345 * on the only set dimension.
2346 * In particular, if "up" is set then "cond" should contain only
2347 * upper bounds on the set dimension.
2348 * Otherwise, it should contain only lower bounds.
2350 static bool is_simple_bound(__isl_keep isl_set *cond, isl_int inc)
2352 if (isl_int_is_pos(inc))
2353 return !isl_set_dim_has_any_lower_bound(cond, isl_dim_set, 0);
2354 else
2355 return !isl_set_dim_has_any_upper_bound(cond, isl_dim_set, 0);
2358 /* Extend a condition on a given iteration of a loop to one that
2359 * imposes the same condition on all previous iterations.
2360 * "domain" expresses the lower [upper] bound on the iterations
2361 * when inc is positive [negative].
2363 * In particular, we construct the condition (when inc is positive)
2365 * forall i' : (domain(i') and i' <= i) => cond(i')
2367 * which is equivalent to
2369 * not exists i' : domain(i') and i' <= i and not cond(i')
2371 * We construct this set by negating cond, applying a map
2373 * { [i'] -> [i] : domain(i') and i' <= i }
2375 * and then negating the result again.
2377 static __isl_give isl_set *valid_for_each_iteration(__isl_take isl_set *cond,
2378 __isl_take isl_set *domain, isl_int inc)
2380 isl_map *previous_to_this;
2382 if (isl_int_is_pos(inc))
2383 previous_to_this = isl_map_lex_le(isl_set_get_space(domain));
2384 else
2385 previous_to_this = isl_map_lex_ge(isl_set_get_space(domain));
2387 previous_to_this = isl_map_intersect_domain(previous_to_this, domain);
2389 cond = isl_set_complement(cond);
2390 cond = isl_set_apply(cond, previous_to_this);
2391 cond = isl_set_complement(cond);
2393 return cond;
2396 /* Construct a domain of the form
2398 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
2400 static __isl_give isl_set *strided_domain(__isl_take isl_id *id,
2401 __isl_take isl_pw_aff *init, isl_int inc)
2403 isl_aff *aff;
2404 isl_space *dim;
2405 isl_set *set;
2407 init = isl_pw_aff_insert_dims(init, isl_dim_in, 0, 1);
2408 dim = isl_pw_aff_get_domain_space(init);
2409 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2410 aff = isl_aff_add_coefficient(aff, isl_dim_in, 0, inc);
2411 init = isl_pw_aff_add(init, isl_pw_aff_from_aff(aff));
2413 dim = isl_space_set_alloc(isl_pw_aff_get_ctx(init), 1, 1);
2414 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
2415 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2416 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2418 set = isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff), init);
2420 set = isl_set_lower_bound_si(set, isl_dim_set, 0, 0);
2422 return isl_set_params(set);
2425 /* Assuming "cond" represents a bound on a loop where the loop
2426 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2427 * is possible.
2429 * Under the given assumptions, wrapping is only possible if "cond" allows
2430 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2431 * increasing iterator and 0 in case of a decreasing iterator.
2433 static bool can_wrap(__isl_keep isl_set *cond, ValueDecl *iv, isl_int inc)
2435 bool cw;
2436 isl_int limit;
2437 isl_set *test;
2439 test = isl_set_copy(cond);
2441 isl_int_init(limit);
2442 if (isl_int_is_neg(inc))
2443 isl_int_set_si(limit, 0);
2444 else {
2445 isl_int_set_si(limit, 1);
2446 isl_int_mul_2exp(limit, limit, get_type_size(iv));
2447 isl_int_sub_ui(limit, limit, 1);
2450 test = isl_set_fix(cond, isl_dim_set, 0, limit);
2451 cw = !isl_set_is_empty(test);
2452 isl_set_free(test);
2454 isl_int_clear(limit);
2456 return cw;
2459 /* Given a one-dimensional space, construct the following mapping on this
2460 * space
2462 * { [v] -> [v mod 2^width] }
2464 * where width is the number of bits used to represent the values
2465 * of the unsigned variable "iv".
2467 static __isl_give isl_map *compute_wrapping(__isl_take isl_space *dim,
2468 ValueDecl *iv)
2470 isl_int mod;
2471 isl_aff *aff;
2472 isl_map *map;
2474 isl_int_init(mod);
2475 isl_int_set_si(mod, 1);
2476 isl_int_mul_2exp(mod, mod, get_type_size(iv));
2478 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2479 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2480 aff = isl_aff_mod(aff, mod);
2482 isl_int_clear(mod);
2484 return isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2485 map = isl_map_reverse(map);
2488 /* Project out the parameter "id" from "set".
2490 static __isl_give isl_set *set_project_out_by_id(__isl_take isl_set *set,
2491 __isl_keep isl_id *id)
2493 int pos;
2495 pos = isl_set_find_dim_by_id(set, isl_dim_param, id);
2496 if (pos >= 0)
2497 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2499 return set;
2502 /* Compute the set of parameters for which "set1" is a subset of "set2".
2504 * set1 is a subset of set2 if
2506 * forall i in set1 : i in set2
2508 * or
2510 * not exists i in set1 and i not in set2
2512 * i.e.,
2514 * not exists i in set1 \ set2
2516 static __isl_give isl_set *enforce_subset(__isl_take isl_set *set1,
2517 __isl_take isl_set *set2)
2519 return isl_set_complement(isl_set_params(isl_set_subtract(set1, set2)));
2522 /* Compute the set of parameter values for which "cond" holds
2523 * on the next iteration for each element of "dom".
2525 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2526 * and then compute the set of parameters for which the result is a subset
2527 * of "cond".
2529 static __isl_give isl_set *valid_on_next(__isl_take isl_set *cond,
2530 __isl_take isl_set *dom, isl_int inc)
2532 isl_space *space;
2533 isl_aff *aff;
2534 isl_map *next;
2536 space = isl_set_get_space(dom);
2537 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2538 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2539 aff = isl_aff_add_constant(aff, inc);
2540 next = isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2542 dom = isl_set_apply(dom, next);
2544 return enforce_subset(dom, cond);
2547 /* Does "id" refer to a nested access?
2549 static bool is_nested_parameter(__isl_keep isl_id *id)
2551 return id && isl_id_get_user(id) && !isl_id_get_name(id);
2554 /* Does parameter "pos" of "space" refer to a nested access?
2556 static bool is_nested_parameter(__isl_keep isl_space *space, int pos)
2558 bool nested;
2559 isl_id *id;
2561 id = isl_space_get_dim_id(space, isl_dim_param, pos);
2562 nested = is_nested_parameter(id);
2563 isl_id_free(id);
2565 return nested;
2568 /* Does "space" involve any parameters that refer to nested
2569 * accesses, i.e., parameters with no name?
2571 static bool has_nested(__isl_keep isl_space *space)
2573 int nparam;
2575 nparam = isl_space_dim(space, isl_dim_param);
2576 for (int i = 0; i < nparam; ++i)
2577 if (is_nested_parameter(space, i))
2578 return true;
2580 return false;
2583 /* Does "pa" involve any parameters that refer to nested
2584 * accesses, i.e., parameters with no name?
2586 static bool has_nested(__isl_keep isl_pw_aff *pa)
2588 isl_space *space;
2589 bool nested;
2591 space = isl_pw_aff_get_space(pa);
2592 nested = has_nested(space);
2593 isl_space_free(space);
2595 return nested;
2598 /* Construct a pet_scop for a for statement.
2599 * The for loop is required to be of the form
2601 * for (i = init; condition; ++i)
2603 * or
2605 * for (i = init; condition; --i)
2607 * The initialization of the for loop should either be an assignment
2608 * to an integer variable, or a declaration of such a variable with
2609 * initialization.
2611 * The condition is allowed to contain nested accesses, provided
2612 * they are not being written to inside the body of the loop.
2613 * Otherwise, or if the condition is otherwise non-affine, the for loop is
2614 * essentially treated as a while loop, with iteration domain
2615 * { [i] : i >= init }.
2617 * We extract a pet_scop for the body and then embed it in a loop with
2618 * iteration domain and schedule
2620 * { [i] : i >= init and condition' }
2621 * { [i] -> [i] }
2623 * or
2625 * { [i] : i <= init and condition' }
2626 * { [i] -> [-i] }
2628 * Where condition' is equal to condition if the latter is
2629 * a simple upper [lower] bound and a condition that is extended
2630 * to apply to all previous iterations otherwise.
2632 * If the condition is non-affine, then we drop the condition from the
2633 * iteration domain and instead create a separate statement
2634 * for evaluating the condition. The body is then filtered to depend
2635 * on the result of the condition evaluating to true on all iterations
2636 * up to the current iteration, while the evaluation the condition itself
2637 * is filtered to depend on the result of the condition evaluating to true
2638 * on all previous iterations.
2639 * The context of the scop representing the body is dropped
2640 * because we don't know how many times the body will be executed,
2641 * if at all.
2643 * If the stride of the loop is not 1, then "i >= init" is replaced by
2645 * (exists a: i = init + stride * a and a >= 0)
2647 * If the loop iterator i is unsigned, then wrapping may occur.
2648 * During the computation, we work with a virtual iterator that
2649 * does not wrap. However, the condition in the code applies
2650 * to the wrapped value, so we need to change condition(i)
2651 * into condition([i % 2^width]).
2652 * After computing the virtual domain and schedule, we apply
2653 * the function { [v] -> [v % 2^width] } to the domain and the domain
2654 * of the schedule. In order not to lose any information, we also
2655 * need to intersect the domain of the schedule with the virtual domain
2656 * first, since some iterations in the wrapped domain may be scheduled
2657 * several times, typically an infinite number of times.
2658 * Note that there may be no need to perform this final wrapping
2659 * if the loop condition (after wrapping) satisfies certain conditions.
2660 * However, the is_simple_bound condition is not enough since it doesn't
2661 * check if there even is an upper bound.
2663 * If the loop condition is non-affine, then we keep the virtual
2664 * iterator in the iteration domain and instead replace all accesses
2665 * to the original iterator by the wrapping of the virtual iterator.
2667 * Wrapping on unsigned iterators can be avoided entirely if
2668 * loop condition is simple, the loop iterator is incremented
2669 * [decremented] by one and the last value before wrapping cannot
2670 * possibly satisfy the loop condition.
2672 * Before extracting a pet_scop from the body we remove all
2673 * assignments in assigned_value to variables that are assigned
2674 * somewhere in the body of the loop.
2676 * Valid parameters for a for loop are those for which the initial
2677 * value itself, the increment on each domain iteration and
2678 * the condition on both the initial value and
2679 * the result of incrementing the iterator for each iteration of the domain
2680 * can be evaluated.
2681 * If the loop condition is non-affine, then we only consider validity
2682 * of the initial value.
2684 * If the body contains any break, then we keep track of it in "skip"
2685 * (if the skip condition is affine) or it is handled in scop_add_break
2686 * (if the skip condition is not affine).
2687 * Note that the affine break condition needs to be considered with
2688 * respect to previous iterations in the virtual domain (if any)
2689 * and that the domain needs to be kept virtual if there is a non-affine
2690 * break condition.
2692 struct pet_scop *PetScan::extract_for(ForStmt *stmt)
2694 BinaryOperator *ass;
2695 Decl *decl;
2696 Stmt *init;
2697 Expr *lhs, *rhs;
2698 ValueDecl *iv;
2699 isl_space *space;
2700 isl_set *domain;
2701 isl_map *sched;
2702 isl_set *cond = NULL;
2703 isl_set *skip = NULL;
2704 isl_id *id;
2705 struct pet_scop *scop, *scop_cond = NULL;
2706 assigned_value_cache cache(assigned_value);
2707 isl_int inc;
2708 bool is_one;
2709 bool is_unsigned;
2710 bool is_simple;
2711 bool is_virtual;
2712 bool keep_virtual = false;
2713 bool has_affine_break;
2714 bool has_var_break;
2715 isl_map *wrap = NULL;
2716 isl_pw_aff *pa, *pa_inc, *init_val;
2717 isl_set *valid_init;
2718 isl_set *valid_cond;
2719 isl_set *valid_cond_init;
2720 isl_set *valid_cond_next;
2721 isl_set *valid_inc;
2722 isl_map *test_access = NULL, *break_access = NULL;
2723 int stmt_id;
2725 if (!stmt->getInit() && !stmt->getCond() && !stmt->getInc())
2726 return extract_infinite_for(stmt);
2728 init = stmt->getInit();
2729 if (!init) {
2730 unsupported(stmt);
2731 return NULL;
2733 if ((ass = initialization_assignment(init)) != NULL) {
2734 iv = extract_induction_variable(ass);
2735 if (!iv)
2736 return NULL;
2737 lhs = ass->getLHS();
2738 rhs = ass->getRHS();
2739 } else if ((decl = initialization_declaration(init)) != NULL) {
2740 VarDecl *var = extract_induction_variable(init, decl);
2741 if (!var)
2742 return NULL;
2743 iv = var;
2744 rhs = var->getInit();
2745 lhs = create_DeclRefExpr(var);
2746 } else {
2747 unsupported(stmt->getInit());
2748 return NULL;
2751 pa_inc = extract_increment(stmt, iv);
2752 if (!pa_inc)
2753 return NULL;
2755 isl_int_init(inc);
2756 if (isl_pw_aff_n_piece(pa_inc) != 1 ||
2757 isl_pw_aff_foreach_piece(pa_inc, &extract_cst, &inc) < 0) {
2758 isl_pw_aff_free(pa_inc);
2759 unsupported(stmt->getInc());
2760 isl_int_clear(inc);
2761 return NULL;
2763 valid_inc = isl_pw_aff_domain(pa_inc);
2765 is_unsigned = iv->getType()->isUnsignedIntegerType();
2767 assigned_value.erase(iv);
2768 clear_assignments clear(assigned_value);
2769 clear.TraverseStmt(stmt->getBody());
2771 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
2773 pa = try_extract_nested_condition(stmt->getCond());
2774 if (allow_nested && (!pa || has_nested(pa)))
2775 stmt_id = n_stmt++;
2777 scop = extract(stmt->getBody());
2779 has_affine_break = scop &&
2780 pet_scop_has_affine_skip(scop, pet_skip_later);
2781 if (has_affine_break) {
2782 skip = pet_scop_get_skip(scop, pet_skip_later);
2783 skip = isl_set_fix_si(skip, isl_dim_set, 0, 1);
2784 skip = isl_set_params(skip);
2786 has_var_break = scop && pet_scop_has_var_skip(scop, pet_skip_later);
2787 if (has_var_break) {
2788 break_access = pet_scop_get_skip_map(scop, pet_skip_later);
2789 keep_virtual = true;
2792 if (pa && !is_nested_allowed(pa, scop)) {
2793 isl_pw_aff_free(pa);
2794 pa = NULL;
2797 if (!allow_nested && !pa)
2798 pa = try_extract_affine_condition(stmt->getCond());
2799 valid_cond = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2800 cond = isl_pw_aff_non_zero_set(pa);
2801 if (allow_nested && !cond) {
2802 int save_n_stmt = n_stmt;
2803 test_access = create_test_access(ctx, n_test++);
2804 n_stmt = stmt_id;
2805 scop_cond = extract_non_affine_condition(stmt->getCond(),
2806 isl_map_copy(test_access));
2807 n_stmt = save_n_stmt;
2808 scop_cond = scop_add_array(scop_cond, test_access, ast_context);
2809 scop_cond = pet_scop_prefix(scop_cond, 0);
2810 scop = pet_scop_reset_context(scop);
2811 scop = pet_scop_prefix(scop, 1);
2812 keep_virtual = true;
2813 cond = isl_set_universe(isl_space_set_alloc(ctx, 0, 0));
2816 cond = embed(cond, isl_id_copy(id));
2817 skip = embed(skip, isl_id_copy(id));
2818 valid_cond = isl_set_coalesce(valid_cond);
2819 valid_cond = embed(valid_cond, isl_id_copy(id));
2820 valid_inc = embed(valid_inc, isl_id_copy(id));
2821 is_one = isl_int_is_one(inc) || isl_int_is_negone(inc);
2822 is_virtual = is_unsigned && (!is_one || can_wrap(cond, iv, inc));
2824 init_val = extract_affine(rhs);
2825 valid_cond_init = enforce_subset(
2826 isl_set_from_pw_aff(isl_pw_aff_copy(init_val)),
2827 isl_set_copy(valid_cond));
2828 if (is_one && !is_virtual) {
2829 isl_pw_aff_free(init_val);
2830 pa = extract_comparison(isl_int_is_pos(inc) ? BO_GE : BO_LE,
2831 lhs, rhs, init);
2832 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2833 valid_init = set_project_out_by_id(valid_init, id);
2834 domain = isl_pw_aff_non_zero_set(pa);
2835 } else {
2836 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(init_val));
2837 domain = strided_domain(isl_id_copy(id), init_val, inc);
2840 domain = embed(domain, isl_id_copy(id));
2841 if (is_virtual) {
2842 isl_map *rev_wrap;
2843 wrap = compute_wrapping(isl_set_get_space(cond), iv);
2844 rev_wrap = isl_map_reverse(isl_map_copy(wrap));
2845 cond = isl_set_apply(cond, isl_map_copy(rev_wrap));
2846 skip = isl_set_apply(skip, isl_map_copy(rev_wrap));
2847 valid_cond = isl_set_apply(valid_cond, isl_map_copy(rev_wrap));
2848 valid_inc = isl_set_apply(valid_inc, rev_wrap);
2850 is_simple = is_simple_bound(cond, inc);
2851 if (!is_simple) {
2852 cond = isl_set_gist(cond, isl_set_copy(domain));
2853 is_simple = is_simple_bound(cond, inc);
2855 if (!is_simple)
2856 cond = valid_for_each_iteration(cond,
2857 isl_set_copy(domain), inc);
2858 domain = isl_set_intersect(domain, cond);
2859 if (has_affine_break) {
2860 skip = isl_set_intersect(skip , isl_set_copy(domain));
2861 skip = after(skip, isl_int_sgn(inc));
2862 domain = isl_set_subtract(domain, skip);
2864 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, isl_id_copy(id));
2865 space = isl_space_from_domain(isl_set_get_space(domain));
2866 space = isl_space_add_dims(space, isl_dim_out, 1);
2867 sched = isl_map_universe(space);
2868 if (isl_int_is_pos(inc))
2869 sched = isl_map_equate(sched, isl_dim_in, 0, isl_dim_out, 0);
2870 else
2871 sched = isl_map_oppose(sched, isl_dim_in, 0, isl_dim_out, 0);
2873 valid_cond_next = valid_on_next(valid_cond, isl_set_copy(domain), inc);
2874 valid_inc = enforce_subset(isl_set_copy(domain), valid_inc);
2876 if (is_virtual && !keep_virtual) {
2877 wrap = isl_map_set_dim_id(wrap,
2878 isl_dim_out, 0, isl_id_copy(id));
2879 sched = isl_map_intersect_domain(sched, isl_set_copy(domain));
2880 domain = isl_set_apply(domain, isl_map_copy(wrap));
2881 sched = isl_map_apply_domain(sched, wrap);
2883 if (!(is_virtual && keep_virtual)) {
2884 space = isl_set_get_space(domain);
2885 wrap = isl_map_identity(isl_space_map_from_set(space));
2888 scop_cond = pet_scop_embed(scop_cond, isl_set_copy(domain),
2889 isl_map_copy(sched), isl_map_copy(wrap), isl_id_copy(id));
2890 scop = pet_scop_embed(scop, isl_set_copy(domain), sched, wrap, id);
2891 scop = resolve_nested(scop);
2892 if (has_var_break)
2893 scop = scop_add_break(scop, break_access, isl_set_copy(domain),
2894 isl_int_sgn(inc));
2895 if (test_access) {
2896 scop = scop_add_while(scop_cond, scop, test_access, domain,
2897 isl_int_sgn(inc));
2898 isl_set_free(valid_inc);
2899 } else {
2900 scop = pet_scop_restrict_context(scop, valid_inc);
2901 scop = pet_scop_restrict_context(scop, valid_cond_next);
2902 scop = pet_scop_restrict_context(scop, valid_cond_init);
2903 isl_set_free(domain);
2905 clear_assignment(assigned_value, iv);
2907 isl_int_clear(inc);
2909 scop = pet_scop_restrict_context(scop, valid_init);
2911 return scop;
2914 struct pet_scop *PetScan::extract(CompoundStmt *stmt, bool skip_declarations)
2916 return extract(stmt->children(), true, skip_declarations);
2919 /* Does parameter "pos" of "map" refer to a nested access?
2921 static bool is_nested_parameter(__isl_keep isl_map *map, int pos)
2923 bool nested;
2924 isl_id *id;
2926 id = isl_map_get_dim_id(map, isl_dim_param, pos);
2927 nested = is_nested_parameter(id);
2928 isl_id_free(id);
2930 return nested;
2933 /* How many parameters of "space" refer to nested accesses, i.e., have no name?
2935 static int n_nested_parameter(__isl_keep isl_space *space)
2937 int n = 0;
2938 int nparam;
2940 nparam = isl_space_dim(space, isl_dim_param);
2941 for (int i = 0; i < nparam; ++i)
2942 if (is_nested_parameter(space, i))
2943 ++n;
2945 return n;
2948 /* How many parameters of "map" refer to nested accesses, i.e., have no name?
2950 static int n_nested_parameter(__isl_keep isl_map *map)
2952 isl_space *space;
2953 int n;
2955 space = isl_map_get_space(map);
2956 n = n_nested_parameter(space);
2957 isl_space_free(space);
2959 return n;
2962 /* For each nested access parameter in "space",
2963 * construct a corresponding pet_expr, place it in args and
2964 * record its position in "param2pos".
2965 * "n_arg" is the number of elements that are already in args.
2966 * The position recorded in "param2pos" takes this number into account.
2967 * If the pet_expr corresponding to a parameter is identical to
2968 * the pet_expr corresponding to an earlier parameter, then these two
2969 * parameters are made to refer to the same element in args.
2971 * Return the final number of elements in args or -1 if an error has occurred.
2973 int PetScan::extract_nested(__isl_keep isl_space *space,
2974 int n_arg, struct pet_expr **args, std::map<int,int> &param2pos)
2976 int nparam;
2978 nparam = isl_space_dim(space, isl_dim_param);
2979 for (int i = 0; i < nparam; ++i) {
2980 int j;
2981 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
2982 Expr *nested;
2984 if (!is_nested_parameter(id)) {
2985 isl_id_free(id);
2986 continue;
2989 nested = (Expr *) isl_id_get_user(id);
2990 args[n_arg] = extract_expr(nested);
2991 if (!args[n_arg])
2992 return -1;
2994 for (j = 0; j < n_arg; ++j)
2995 if (pet_expr_is_equal(args[j], args[n_arg]))
2996 break;
2998 if (j < n_arg) {
2999 pet_expr_free(args[n_arg]);
3000 args[n_arg] = NULL;
3001 param2pos[i] = j;
3002 } else
3003 param2pos[i] = n_arg++;
3005 isl_id_free(id);
3008 return n_arg;
3011 /* For each nested access parameter in the access relations in "expr",
3012 * construct a corresponding pet_expr, place it in expr->args and
3013 * record its position in "param2pos".
3014 * n is the number of nested access parameters.
3016 struct pet_expr *PetScan::extract_nested(struct pet_expr *expr, int n,
3017 std::map<int,int> &param2pos)
3019 isl_space *space;
3021 expr->args = isl_calloc_array(ctx, struct pet_expr *, n);
3022 expr->n_arg = n;
3023 if (!expr->args)
3024 goto error;
3026 space = isl_map_get_space(expr->acc.access);
3027 n = extract_nested(space, 0, expr->args, param2pos);
3028 isl_space_free(space);
3030 if (n < 0)
3031 goto error;
3033 expr->n_arg = n;
3034 return expr;
3035 error:
3036 pet_expr_free(expr);
3037 return NULL;
3040 /* Look for parameters in any access relation in "expr" that
3041 * refer to nested accesses. In particular, these are
3042 * parameters with no name.
3044 * If there are any such parameters, then the domain of the access
3045 * relation, which is still [] at this point, is replaced by
3046 * [[] -> [t_1,...,t_n]], with n the number of these parameters
3047 * (after identifying identical nested accesses).
3048 * The parameters are then equated to the corresponding t dimensions
3049 * and subsequently projected out.
3050 * param2pos maps the position of the parameter to the position
3051 * of the corresponding t dimension.
3053 struct pet_expr *PetScan::resolve_nested(struct pet_expr *expr)
3055 int n;
3056 int nparam;
3057 int n_in;
3058 isl_space *dim;
3059 isl_map *map;
3060 std::map<int,int> param2pos;
3062 if (!expr)
3063 return expr;
3065 for (int i = 0; i < expr->n_arg; ++i) {
3066 expr->args[i] = resolve_nested(expr->args[i]);
3067 if (!expr->args[i]) {
3068 pet_expr_free(expr);
3069 return NULL;
3073 if (expr->type != pet_expr_access)
3074 return expr;
3076 n = n_nested_parameter(expr->acc.access);
3077 if (n == 0)
3078 return expr;
3080 expr = extract_nested(expr, n, param2pos);
3081 if (!expr)
3082 return NULL;
3084 n = expr->n_arg;
3085 nparam = isl_map_dim(expr->acc.access, isl_dim_param);
3086 n_in = isl_map_dim(expr->acc.access, isl_dim_in);
3087 dim = isl_map_get_space(expr->acc.access);
3088 dim = isl_space_domain(dim);
3089 dim = isl_space_from_domain(dim);
3090 dim = isl_space_add_dims(dim, isl_dim_out, n);
3091 map = isl_map_universe(dim);
3092 map = isl_map_domain_map(map);
3093 map = isl_map_reverse(map);
3094 expr->acc.access = isl_map_apply_domain(expr->acc.access, map);
3096 for (int i = nparam - 1; i >= 0; --i) {
3097 isl_id *id = isl_map_get_dim_id(expr->acc.access,
3098 isl_dim_param, i);
3099 if (!is_nested_parameter(id)) {
3100 isl_id_free(id);
3101 continue;
3104 expr->acc.access = isl_map_equate(expr->acc.access,
3105 isl_dim_param, i, isl_dim_in,
3106 n_in + param2pos[i]);
3107 expr->acc.access = isl_map_project_out(expr->acc.access,
3108 isl_dim_param, i, 1);
3110 isl_id_free(id);
3113 return expr;
3114 error:
3115 pet_expr_free(expr);
3116 return NULL;
3119 /* Convert a top-level pet_expr to a pet_scop with one statement.
3120 * This mainly involves resolving nested expression parameters
3121 * and setting the name of the iteration space.
3122 * The name is given by "label" if it is non-NULL. Otherwise,
3123 * it is of the form S_<n_stmt>.
3125 struct pet_scop *PetScan::extract(Stmt *stmt, struct pet_expr *expr,
3126 __isl_take isl_id *label)
3128 struct pet_stmt *ps;
3129 SourceLocation loc = stmt->getLocStart();
3130 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3132 expr = resolve_nested(expr);
3133 ps = pet_stmt_from_pet_expr(ctx, line, label, n_stmt++, expr);
3134 return pet_scop_from_pet_stmt(ctx, ps);
3137 /* Check if we can extract an affine expression from "expr".
3138 * Return the expressions as an isl_pw_aff if we can and NULL otherwise.
3139 * We turn on autodetection so that we won't generate any warnings
3140 * and turn off nesting, so that we won't accept any non-affine constructs.
3142 __isl_give isl_pw_aff *PetScan::try_extract_affine(Expr *expr)
3144 isl_pw_aff *pwaff;
3145 int save_autodetect = options->autodetect;
3146 bool save_nesting = nesting_enabled;
3148 options->autodetect = 1;
3149 nesting_enabled = false;
3151 pwaff = extract_affine(expr);
3153 options->autodetect = save_autodetect;
3154 nesting_enabled = save_nesting;
3156 return pwaff;
3159 /* Check whether "expr" is an affine expression.
3161 bool PetScan::is_affine(Expr *expr)
3163 isl_pw_aff *pwaff;
3165 pwaff = try_extract_affine(expr);
3166 isl_pw_aff_free(pwaff);
3168 return pwaff != NULL;
3171 /* Check if we can extract an affine constraint from "expr".
3172 * Return the constraint as an isl_set if we can and NULL otherwise.
3173 * We turn on autodetection so that we won't generate any warnings
3174 * and turn off nesting, so that we won't accept any non-affine constructs.
3176 __isl_give isl_pw_aff *PetScan::try_extract_affine_condition(Expr *expr)
3178 isl_pw_aff *cond;
3179 int save_autodetect = options->autodetect;
3180 bool save_nesting = nesting_enabled;
3182 options->autodetect = 1;
3183 nesting_enabled = false;
3185 cond = extract_condition(expr);
3187 options->autodetect = save_autodetect;
3188 nesting_enabled = save_nesting;
3190 return cond;
3193 /* Check whether "expr" is an affine constraint.
3195 bool PetScan::is_affine_condition(Expr *expr)
3197 isl_pw_aff *cond;
3199 cond = try_extract_affine_condition(expr);
3200 isl_pw_aff_free(cond);
3202 return cond != NULL;
3205 /* Check if we can extract a condition from "expr".
3206 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
3207 * If allow_nested is set, then the condition may involve parameters
3208 * corresponding to nested accesses.
3209 * We turn on autodetection so that we won't generate any warnings.
3211 __isl_give isl_pw_aff *PetScan::try_extract_nested_condition(Expr *expr)
3213 isl_pw_aff *cond;
3214 int save_autodetect = options->autodetect;
3215 bool save_nesting = nesting_enabled;
3217 options->autodetect = 1;
3218 nesting_enabled = allow_nested;
3219 cond = extract_condition(expr);
3221 options->autodetect = save_autodetect;
3222 nesting_enabled = save_nesting;
3224 return cond;
3227 /* If the top-level expression of "stmt" is an assignment, then
3228 * return that assignment as a BinaryOperator.
3229 * Otherwise return NULL.
3231 static BinaryOperator *top_assignment_or_null(Stmt *stmt)
3233 BinaryOperator *ass;
3235 if (!stmt)
3236 return NULL;
3237 if (stmt->getStmtClass() != Stmt::BinaryOperatorClass)
3238 return NULL;
3240 ass = cast<BinaryOperator>(stmt);
3241 if(ass->getOpcode() != BO_Assign)
3242 return NULL;
3244 return ass;
3247 /* Check if the given if statement is a conditional assignement
3248 * with a non-affine condition. If so, construct a pet_scop
3249 * corresponding to this conditional assignment. Otherwise return NULL.
3251 * In particular we check if "stmt" is of the form
3253 * if (condition)
3254 * a = f(...);
3255 * else
3256 * a = g(...);
3258 * where a is some array or scalar access.
3259 * The constructed pet_scop then corresponds to the expression
3261 * a = condition ? f(...) : g(...)
3263 * All access relations in f(...) are intersected with condition
3264 * while all access relation in g(...) are intersected with the complement.
3266 struct pet_scop *PetScan::extract_conditional_assignment(IfStmt *stmt)
3268 BinaryOperator *ass_then, *ass_else;
3269 isl_map *write_then, *write_else;
3270 isl_set *cond, *comp;
3271 isl_map *map;
3272 isl_pw_aff *pa;
3273 int equal;
3274 struct pet_expr *pe_cond, *pe_then, *pe_else, *pe, *pe_write;
3275 bool save_nesting = nesting_enabled;
3277 if (!options->detect_conditional_assignment)
3278 return NULL;
3280 ass_then = top_assignment_or_null(stmt->getThen());
3281 ass_else = top_assignment_or_null(stmt->getElse());
3283 if (!ass_then || !ass_else)
3284 return NULL;
3286 if (is_affine_condition(stmt->getCond()))
3287 return NULL;
3289 write_then = extract_access(ass_then->getLHS());
3290 write_else = extract_access(ass_else->getLHS());
3292 equal = isl_map_is_equal(write_then, write_else);
3293 isl_map_free(write_else);
3294 if (equal < 0 || !equal) {
3295 isl_map_free(write_then);
3296 return NULL;
3299 nesting_enabled = allow_nested;
3300 pa = extract_condition(stmt->getCond());
3301 nesting_enabled = save_nesting;
3302 cond = isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa));
3303 comp = isl_pw_aff_zero_set(isl_pw_aff_copy(pa));
3304 map = isl_map_from_range(isl_set_from_pw_aff(pa));
3306 pe_cond = pet_expr_from_access(map);
3308 pe_then = extract_expr(ass_then->getRHS());
3309 pe_then = pet_expr_restrict(pe_then, cond);
3310 pe_else = extract_expr(ass_else->getRHS());
3311 pe_else = pet_expr_restrict(pe_else, comp);
3313 pe = pet_expr_new_ternary(ctx, pe_cond, pe_then, pe_else);
3314 pe_write = pet_expr_from_access(write_then);
3315 if (pe_write) {
3316 pe_write->acc.write = 1;
3317 pe_write->acc.read = 0;
3319 pe = pet_expr_new_binary(ctx, pet_op_assign, pe_write, pe);
3320 return extract(stmt, pe);
3323 /* Create a pet_scop with a single statement evaluating "cond"
3324 * and writing the result to a virtual scalar, as expressed by
3325 * "access".
3327 struct pet_scop *PetScan::extract_non_affine_condition(Expr *cond,
3328 __isl_take isl_map *access)
3330 struct pet_expr *expr, *write;
3331 struct pet_stmt *ps;
3332 struct pet_scop *scop;
3333 SourceLocation loc = cond->getLocStart();
3334 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3336 write = pet_expr_from_access(access);
3337 if (write) {
3338 write->acc.write = 1;
3339 write->acc.read = 0;
3341 expr = extract_expr(cond);
3342 expr = resolve_nested(expr);
3343 expr = pet_expr_new_binary(ctx, pet_op_assign, write, expr);
3344 ps = pet_stmt_from_pet_expr(ctx, line, NULL, n_stmt++, expr);
3345 scop = pet_scop_from_pet_stmt(ctx, ps);
3346 scop = resolve_nested(scop);
3348 return scop;
3351 extern "C" {
3352 static __isl_give isl_map *embed_access(__isl_take isl_map *access,
3353 void *user);
3356 /* Apply the map pointed to by "user" to the domain of the access
3357 * relation, thereby embedding it in the range of the map.
3358 * The domain of both relations is the zero-dimensional domain.
3360 static __isl_give isl_map *embed_access(__isl_take isl_map *access, void *user)
3362 isl_map *map = (isl_map *) user;
3364 return isl_map_apply_domain(access, isl_map_copy(map));
3367 /* Apply "map" to all access relations in "expr".
3369 static struct pet_expr *embed(struct pet_expr *expr, __isl_keep isl_map *map)
3371 return pet_expr_foreach_access(expr, &embed_access, map);
3374 /* How many parameters of "set" refer to nested accesses, i.e., have no name?
3376 static int n_nested_parameter(__isl_keep isl_set *set)
3378 isl_space *space;
3379 int n;
3381 space = isl_set_get_space(set);
3382 n = n_nested_parameter(space);
3383 isl_space_free(space);
3385 return n;
3388 /* Remove all parameters from "map" that refer to nested accesses.
3390 static __isl_give isl_map *remove_nested_parameters(__isl_take isl_map *map)
3392 int nparam;
3393 isl_space *space;
3395 space = isl_map_get_space(map);
3396 nparam = isl_space_dim(space, isl_dim_param);
3397 for (int i = nparam - 1; i >= 0; --i)
3398 if (is_nested_parameter(space, i))
3399 map = isl_map_project_out(map, isl_dim_param, i, 1);
3400 isl_space_free(space);
3402 return map;
3405 extern "C" {
3406 static __isl_give isl_map *access_remove_nested_parameters(
3407 __isl_take isl_map *access, void *user);
3410 static __isl_give isl_map *access_remove_nested_parameters(
3411 __isl_take isl_map *access, void *user)
3413 return remove_nested_parameters(access);
3416 /* Remove all nested access parameters from the schedule and all
3417 * accesses of "stmt".
3418 * There is no need to remove them from the domain as these parameters
3419 * have already been removed from the domain when this function is called.
3421 static struct pet_stmt *remove_nested_parameters(struct pet_stmt *stmt)
3423 if (!stmt)
3424 return NULL;
3425 stmt->schedule = remove_nested_parameters(stmt->schedule);
3426 stmt->body = pet_expr_foreach_access(stmt->body,
3427 &access_remove_nested_parameters, NULL);
3428 if (!stmt->schedule || !stmt->body)
3429 goto error;
3430 for (int i = 0; i < stmt->n_arg; ++i) {
3431 stmt->args[i] = pet_expr_foreach_access(stmt->args[i],
3432 &access_remove_nested_parameters, NULL);
3433 if (!stmt->args[i])
3434 goto error;
3437 return stmt;
3438 error:
3439 pet_stmt_free(stmt);
3440 return NULL;
3443 /* For each nested access parameter in the domain of "stmt",
3444 * construct a corresponding pet_expr, place it before the original
3445 * elements in stmt->args and record its position in "param2pos".
3446 * n is the number of nested access parameters.
3448 struct pet_stmt *PetScan::extract_nested(struct pet_stmt *stmt, int n,
3449 std::map<int,int> &param2pos)
3451 int i;
3452 isl_space *space;
3453 int n_arg;
3454 struct pet_expr **args;
3456 n_arg = stmt->n_arg;
3457 args = isl_calloc_array(ctx, struct pet_expr *, n + n_arg);
3458 if (!args)
3459 goto error;
3461 space = isl_set_get_space(stmt->domain);
3462 n_arg = extract_nested(space, 0, args, param2pos);
3463 isl_space_free(space);
3465 if (n_arg < 0)
3466 goto error;
3468 for (i = 0; i < stmt->n_arg; ++i)
3469 args[n_arg + i] = stmt->args[i];
3470 free(stmt->args);
3471 stmt->args = args;
3472 stmt->n_arg += n_arg;
3474 return stmt;
3475 error:
3476 if (args) {
3477 for (i = 0; i < n; ++i)
3478 pet_expr_free(args[i]);
3479 free(args);
3481 pet_stmt_free(stmt);
3482 return NULL;
3485 /* Check whether any of the arguments i of "stmt" starting at position "n"
3486 * is equal to one of the first "n" arguments j.
3487 * If so, combine the constraints on arguments i and j and remove
3488 * argument i.
3490 static struct pet_stmt *remove_duplicate_arguments(struct pet_stmt *stmt, int n)
3492 int i, j;
3493 isl_map *map;
3495 if (!stmt)
3496 return NULL;
3497 if (n == 0)
3498 return stmt;
3499 if (n == stmt->n_arg)
3500 return stmt;
3502 map = isl_set_unwrap(stmt->domain);
3504 for (i = stmt->n_arg - 1; i >= n; --i) {
3505 for (j = 0; j < n; ++j)
3506 if (pet_expr_is_equal(stmt->args[i], stmt->args[j]))
3507 break;
3508 if (j >= n)
3509 continue;
3511 map = isl_map_equate(map, isl_dim_out, i, isl_dim_out, j);
3512 map = isl_map_project_out(map, isl_dim_out, i, 1);
3514 pet_expr_free(stmt->args[i]);
3515 for (j = i; j + 1 < stmt->n_arg; ++j)
3516 stmt->args[j] = stmt->args[j + 1];
3517 stmt->n_arg--;
3520 stmt->domain = isl_map_wrap(map);
3521 if (!stmt->domain)
3522 goto error;
3523 return stmt;
3524 error:
3525 pet_stmt_free(stmt);
3526 return NULL;
3529 /* Look for parameters in the iteration domain of "stmt" that
3530 * refer to nested accesses. In particular, these are
3531 * parameters with no name.
3533 * If there are any such parameters, then as many extra variables
3534 * (after identifying identical nested accesses) are inserted in the
3535 * range of the map wrapped inside the domain, before the original variables.
3536 * If the original domain is not a wrapped map, then a new wrapped
3537 * map is created with zero output dimensions.
3538 * The parameters are then equated to the corresponding output dimensions
3539 * and subsequently projected out, from the iteration domain,
3540 * the schedule and the access relations.
3541 * For each of the output dimensions, a corresponding argument
3542 * expression is inserted. Initially they are created with
3543 * a zero-dimensional domain, so they have to be embedded
3544 * in the current iteration domain.
3545 * param2pos maps the position of the parameter to the position
3546 * of the corresponding output dimension in the wrapped map.
3548 struct pet_stmt *PetScan::resolve_nested(struct pet_stmt *stmt)
3550 int n;
3551 int nparam;
3552 unsigned n_arg;
3553 isl_map *map;
3554 std::map<int,int> param2pos;
3556 if (!stmt)
3557 return NULL;
3559 n = n_nested_parameter(stmt->domain);
3560 if (n == 0)
3561 return stmt;
3563 n_arg = stmt->n_arg;
3564 stmt = extract_nested(stmt, n, param2pos);
3565 if (!stmt)
3566 return NULL;
3568 n = stmt->n_arg - n_arg;
3569 nparam = isl_set_dim(stmt->domain, isl_dim_param);
3570 if (isl_set_is_wrapping(stmt->domain))
3571 map = isl_set_unwrap(stmt->domain);
3572 else
3573 map = isl_map_from_domain(stmt->domain);
3574 map = isl_map_insert_dims(map, isl_dim_out, 0, n);
3576 for (int i = nparam - 1; i >= 0; --i) {
3577 isl_id *id;
3579 if (!is_nested_parameter(map, i))
3580 continue;
3582 id = isl_map_get_tuple_id(stmt->args[param2pos[i]]->acc.access,
3583 isl_dim_out);
3584 map = isl_map_set_dim_id(map, isl_dim_out, param2pos[i], id);
3585 map = isl_map_equate(map, isl_dim_param, i, isl_dim_out,
3586 param2pos[i]);
3587 map = isl_map_project_out(map, isl_dim_param, i, 1);
3590 stmt->domain = isl_map_wrap(map);
3592 map = isl_set_unwrap(isl_set_copy(stmt->domain));
3593 map = isl_map_from_range(isl_map_domain(map));
3594 for (int pos = 0; pos < n; ++pos)
3595 stmt->args[pos] = embed(stmt->args[pos], map);
3596 isl_map_free(map);
3598 stmt = remove_nested_parameters(stmt);
3599 stmt = remove_duplicate_arguments(stmt, n);
3601 return stmt;
3602 error:
3603 pet_stmt_free(stmt);
3604 return NULL;
3607 /* For each statement in "scop", move the parameters that correspond
3608 * to nested access into the ranges of the domains and create
3609 * corresponding argument expressions.
3611 struct pet_scop *PetScan::resolve_nested(struct pet_scop *scop)
3613 if (!scop)
3614 return NULL;
3616 for (int i = 0; i < scop->n_stmt; ++i) {
3617 scop->stmts[i] = resolve_nested(scop->stmts[i]);
3618 if (!scop->stmts[i])
3619 goto error;
3622 return scop;
3623 error:
3624 pet_scop_free(scop);
3625 return NULL;
3628 /* Given an access expression "expr", is the variable accessed by
3629 * "expr" assigned anywhere inside "scop"?
3631 static bool is_assigned(pet_expr *expr, pet_scop *scop)
3633 bool assigned = false;
3634 isl_id *id;
3636 id = isl_map_get_tuple_id(expr->acc.access, isl_dim_out);
3637 assigned = pet_scop_writes(scop, id);
3638 isl_id_free(id);
3640 return assigned;
3643 /* Are all nested access parameters in "pa" allowed given "scop".
3644 * In particular, is none of them written by anywhere inside "scop".
3646 * If "scop" has any skip conditions, then no nested access parameters
3647 * are allowed. In particular, if there is any nested access in a guard
3648 * for a piece of code containing a "continue", then we want to introduce
3649 * a separate statement for evaluating this guard so that we can express
3650 * that the result is false for all previous iterations.
3652 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff *pa, pet_scop *scop)
3654 int nparam;
3656 if (!scop)
3657 return true;
3659 nparam = isl_pw_aff_dim(pa, isl_dim_param);
3660 for (int i = 0; i < nparam; ++i) {
3661 Expr *nested;
3662 isl_id *id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
3663 pet_expr *expr;
3664 bool allowed;
3666 if (!is_nested_parameter(id)) {
3667 isl_id_free(id);
3668 continue;
3671 if (pet_scop_has_skip(scop, pet_skip_now)) {
3672 isl_id_free(id);
3673 return false;
3676 nested = (Expr *) isl_id_get_user(id);
3677 expr = extract_expr(nested);
3678 allowed = expr && expr->type == pet_expr_access &&
3679 !is_assigned(expr, scop);
3681 pet_expr_free(expr);
3682 isl_id_free(id);
3684 if (!allowed)
3685 return false;
3688 return true;
3691 /* Do we need to construct a skip condition of the given type
3692 * on an if statement, given that the if condition is non-affine?
3694 * pet_scop_filter_skip can only handle the case where the if condition
3695 * holds (the then branch) and the skip condition is universal.
3696 * In any other case, we need to construct a new skip condition.
3698 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
3699 bool have_else, enum pet_skip type)
3701 if (have_else && scop_else && pet_scop_has_skip(scop_else, type))
3702 return true;
3703 if (scop_then && pet_scop_has_skip(scop_then, type) &&
3704 !pet_scop_has_universal_skip(scop_then, type))
3705 return true;
3706 return false;
3709 /* Do we need to construct a skip condition of the given type
3710 * on an if statement, given that the if condition is affine?
3712 * There is no need to construct a new skip condition if all
3713 * the skip conditions are affine.
3715 static bool need_skip_aff(struct pet_scop *scop_then,
3716 struct pet_scop *scop_else, bool have_else, enum pet_skip type)
3718 if (scop_then && pet_scop_has_var_skip(scop_then, type))
3719 return true;
3720 if (have_else && scop_else && pet_scop_has_var_skip(scop_else, type))
3721 return true;
3722 return false;
3725 /* Do we need to construct a skip condition of the given type
3726 * on an if statement?
3728 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
3729 bool have_else, enum pet_skip type, bool affine)
3731 if (affine)
3732 return need_skip_aff(scop_then, scop_else, have_else, type);
3733 else
3734 return need_skip(scop_then, scop_else, have_else, type);
3737 /* Construct an affine expression pet_expr that is evaluates
3738 * to the constant "val".
3740 static struct pet_expr *universally(isl_ctx *ctx, int val)
3742 isl_space *space;
3743 isl_map *map;
3745 space = isl_space_alloc(ctx, 0, 0, 1);
3746 map = isl_map_universe(space);
3747 map = isl_map_fix_si(map, isl_dim_out, 0, val);
3749 return pet_expr_from_access(map);
3752 /* Construct an affine expression pet_expr that is evaluates
3753 * to the constant 1.
3755 static struct pet_expr *universally_true(isl_ctx *ctx)
3757 return universally(ctx, 1);
3760 /* Construct an affine expression pet_expr that is evaluates
3761 * to the constant 0.
3763 static struct pet_expr *universally_false(isl_ctx *ctx)
3765 return universally(ctx, 0);
3768 /* Given an access relation "test_access" for the if condition,
3769 * an access relation "skip_access" for the skip condition and
3770 * scops for the then and else branches, construct a scop for
3771 * computing "skip_access".
3773 * The computed scop contains a single statement that essentially does
3775 * skip_cond = test_cond ? skip_cond_then : skip_cond_else
3777 * If the skip conditions of the then and/or else branch are not affine,
3778 * then they need to be filtered by test_access.
3779 * If they are missing, then this means the skip condition is false.
3781 * Since we are constructing a skip condition for the if statement,
3782 * the skip conditions on the then and else branches are removed.
3784 static struct pet_scop *extract_skip(PetScan *scan,
3785 __isl_take isl_map *test_access, __isl_take isl_map *skip_access,
3786 struct pet_scop *scop_then, struct pet_scop *scop_else, bool have_else,
3787 enum pet_skip type)
3789 struct pet_expr *expr_then, *expr_else, *expr, *expr_skip;
3790 struct pet_stmt *stmt;
3791 struct pet_scop *scop;
3792 isl_ctx *ctx = scan->ctx;
3794 if (!scop_then)
3795 goto error;
3796 if (have_else && !scop_else)
3797 goto error;
3799 if (pet_scop_has_skip(scop_then, type)) {
3800 expr_then = pet_scop_get_skip_expr(scop_then, type);
3801 pet_scop_reset_skip(scop_then, type);
3802 if (!pet_expr_is_affine(expr_then))
3803 expr_then = pet_expr_filter(expr_then,
3804 isl_map_copy(test_access), 1);
3805 } else
3806 expr_then = universally_false(ctx);
3808 if (have_else && pet_scop_has_skip(scop_else, type)) {
3809 expr_else = pet_scop_get_skip_expr(scop_else, type);
3810 pet_scop_reset_skip(scop_else, type);
3811 if (!pet_expr_is_affine(expr_else))
3812 expr_else = pet_expr_filter(expr_else,
3813 isl_map_copy(test_access), 0);
3814 } else
3815 expr_else = universally_false(ctx);
3817 expr = pet_expr_from_access(test_access);
3818 expr = pet_expr_new_ternary(ctx, expr, expr_then, expr_else);
3819 expr_skip = pet_expr_from_access(isl_map_copy(skip_access));
3820 if (expr_skip) {
3821 expr_skip->acc.write = 1;
3822 expr_skip->acc.read = 0;
3824 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
3825 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, scan->n_stmt++, expr);
3827 scop = pet_scop_from_pet_stmt(ctx, stmt);
3828 scop = scop_add_array(scop, skip_access, scan->ast_context);
3829 isl_map_free(skip_access);
3831 return scop;
3832 error:
3833 isl_map_free(test_access);
3834 isl_map_free(skip_access);
3835 return NULL;
3838 /* Is scop's skip_now condition equal to its skip_later condition?
3839 * In particular, this means that it either has no skip_now condition
3840 * or both a skip_now and a skip_later condition (that are equal to each other).
3842 static bool skip_equals_skip_later(struct pet_scop *scop)
3844 int has_skip_now, has_skip_later;
3845 int equal;
3846 isl_set *skip_now, *skip_later;
3848 if (!scop)
3849 return false;
3850 has_skip_now = pet_scop_has_skip(scop, pet_skip_now);
3851 has_skip_later = pet_scop_has_skip(scop, pet_skip_later);
3852 if (has_skip_now != has_skip_later)
3853 return false;
3854 if (!has_skip_now)
3855 return true;
3857 skip_now = pet_scop_get_skip(scop, pet_skip_now);
3858 skip_later = pet_scop_get_skip(scop, pet_skip_later);
3859 equal = isl_set_is_equal(skip_now, skip_later);
3860 isl_set_free(skip_now);
3861 isl_set_free(skip_later);
3863 return equal;
3866 /* Drop the skip conditions of type pet_skip_later from scop1 and scop2.
3868 static void drop_skip_later(struct pet_scop *scop1, struct pet_scop *scop2)
3870 pet_scop_reset_skip(scop1, pet_skip_later);
3871 pet_scop_reset_skip(scop2, pet_skip_later);
3874 /* Structure that handles the construction of skip conditions.
3876 * scop_then and scop_else represent the then and else branches
3877 * of the if statement
3879 * skip[type] is true if we need to construct a skip condition of that type
3880 * equal is set if the skip conditions of types pet_skip_now and pet_skip_later
3881 * are equal to each other
3882 * access[type] is the virtual array representing the skip condition
3883 * scop[type] is a scop for computing the skip condition
3885 struct pet_skip_info {
3886 isl_ctx *ctx;
3888 bool skip[2];
3889 bool equal;
3890 isl_map *access[2];
3891 struct pet_scop *scop[2];
3893 pet_skip_info(isl_ctx *ctx) : ctx(ctx) {}
3895 operator bool() { return skip[pet_skip_now] || skip[pet_skip_later]; }
3898 /* Structure that handles the construction of skip conditions on if statements.
3900 * scop_then and scop_else represent the then and else branches
3901 * of the if statement
3903 struct pet_skip_info_if : public pet_skip_info {
3904 struct pet_scop *scop_then, *scop_else;
3905 bool have_else;
3907 pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
3908 struct pet_scop *scop_else, bool have_else, bool affine);
3909 void extract(PetScan *scan, __isl_keep isl_map *access,
3910 enum pet_skip type);
3911 void extract(PetScan *scan, __isl_keep isl_map *access);
3912 void extract(PetScan *scan, __isl_keep isl_pw_aff *cond);
3913 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
3914 int offset);
3915 struct pet_scop *add(struct pet_scop *scop, int offset);
3918 /* Initialize a pet_skip_info_if structure based on the then and else branches
3919 * and based on whether the if condition is affine or not.
3921 pet_skip_info_if::pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
3922 struct pet_scop *scop_else, bool have_else, bool affine) :
3923 pet_skip_info(ctx), scop_then(scop_then), scop_else(scop_else),
3924 have_else(have_else)
3926 skip[pet_skip_now] =
3927 need_skip(scop_then, scop_else, have_else, pet_skip_now, affine);
3928 equal = skip[pet_skip_now] && skip_equals_skip_later(scop_then) &&
3929 (!have_else || skip_equals_skip_later(scop_else));
3930 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
3931 need_skip(scop_then, scop_else, have_else, pet_skip_later, affine);
3934 /* If we need to construct a skip condition of the given type,
3935 * then do so now.
3937 * "map" represents the if condition.
3939 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_map *map,
3940 enum pet_skip type)
3942 if (!skip[type])
3943 return;
3945 access[type] = create_test_access(isl_map_get_ctx(map), scan->n_test++);
3946 scop[type] = extract_skip(scan, isl_map_copy(map),
3947 isl_map_copy(access[type]),
3948 scop_then, scop_else, have_else, type);
3951 /* Construct the required skip conditions, given the if condition "map".
3953 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_map *map)
3955 extract(scan, map, pet_skip_now);
3956 extract(scan, map, pet_skip_later);
3957 if (equal)
3958 drop_skip_later(scop_then, scop_else);
3961 /* Construct the required skip conditions, given the if condition "cond".
3963 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_pw_aff *cond)
3965 isl_set *test_set;
3966 isl_map *test;
3968 if (!skip[pet_skip_now] && !skip[pet_skip_later])
3969 return;
3971 test_set = isl_set_from_pw_aff(isl_pw_aff_copy(cond));
3972 test = isl_map_from_range(test_set);
3973 extract(scan, test);
3974 isl_map_free(test);
3977 /* Add the computed skip condition of the give type to "main" and
3978 * add the scop for computing the condition at the given offset.
3980 * If equal is set, then we only computed a skip condition for pet_skip_now,
3981 * but we also need to set it as main's pet_skip_later.
3983 struct pet_scop *pet_skip_info_if::add(struct pet_scop *main,
3984 enum pet_skip type, int offset)
3986 isl_set *skip_set;
3988 if (!skip[type])
3989 return main;
3991 skip_set = isl_map_range(access[type]);
3992 access[type] = NULL;
3993 scop[type] = pet_scop_prefix(scop[type], offset);
3994 main = pet_scop_add_par(ctx, main, scop[type]);
3995 scop[type] = NULL;
3997 if (equal)
3998 main = pet_scop_set_skip(main, pet_skip_later,
3999 isl_set_copy(skip_set));
4001 main = pet_scop_set_skip(main, type, skip_set);
4003 return main;
4006 /* Add the computed skip conditions to "main" and
4007 * add the scops for computing the conditions at the given offset.
4009 struct pet_scop *pet_skip_info_if::add(struct pet_scop *scop, int offset)
4011 scop = add(scop, pet_skip_now, offset);
4012 scop = add(scop, pet_skip_later, offset);
4014 return scop;
4017 /* Construct a pet_scop for a non-affine if statement.
4019 * We create a separate statement that writes the result
4020 * of the non-affine condition to a virtual scalar.
4021 * A constraint requiring the value of this virtual scalar to be one
4022 * is added to the iteration domains of the then branch.
4023 * Similarly, a constraint requiring the value of this virtual scalar
4024 * to be zero is added to the iteration domains of the else branch, if any.
4025 * We adjust the schedules to ensure that the virtual scalar is written
4026 * before it is read.
4028 * If there are any breaks or continues in the then and/or else
4029 * branches, then we may have to compute a new skip condition.
4030 * This is handled using a pet_skip_info_if object.
4031 * On initialization, the object checks if skip conditions need
4032 * to be computed. If so, it does so in "extract" and adds them in "add".
4034 struct pet_scop *PetScan::extract_non_affine_if(Expr *cond,
4035 struct pet_scop *scop_then, struct pet_scop *scop_else,
4036 bool have_else, int stmt_id)
4038 struct pet_scop *scop;
4039 isl_map *test_access;
4040 int save_n_stmt = n_stmt;
4042 test_access = create_test_access(ctx, n_test++);
4043 n_stmt = stmt_id;
4044 scop = extract_non_affine_condition(cond, isl_map_copy(test_access));
4045 n_stmt = save_n_stmt;
4046 scop = scop_add_array(scop, test_access, ast_context);
4048 pet_skip_info_if skip(ctx, scop_then, scop_else, have_else, false);
4049 skip.extract(this, test_access);
4051 scop = pet_scop_prefix(scop, 0);
4052 scop_then = pet_scop_prefix(scop_then, 1);
4053 scop_then = pet_scop_filter(scop_then, isl_map_copy(test_access), 1);
4054 if (have_else) {
4055 scop_else = pet_scop_prefix(scop_else, 1);
4056 scop_else = pet_scop_filter(scop_else, test_access, 0);
4057 scop_then = pet_scop_add_par(ctx, scop_then, scop_else);
4058 } else
4059 isl_map_free(test_access);
4061 scop = pet_scop_add_seq(ctx, scop, scop_then);
4063 scop = skip.add(scop, 2);
4065 return scop;
4068 /* Construct a pet_scop for an if statement.
4070 * If the condition fits the pattern of a conditional assignment,
4071 * then it is handled by extract_conditional_assignment.
4072 * Otherwise, we do the following.
4074 * If the condition is affine, then the condition is added
4075 * to the iteration domains of the then branch, while the
4076 * opposite of the condition in added to the iteration domains
4077 * of the else branch, if any.
4078 * We allow the condition to be dynamic, i.e., to refer to
4079 * scalars or array elements that may be written to outside
4080 * of the given if statement. These nested accesses are then represented
4081 * as output dimensions in the wrapping iteration domain.
4082 * If it also written _inside_ the then or else branch, then
4083 * we treat the condition as non-affine.
4084 * As explained in extract_non_affine_if, this will introduce
4085 * an extra statement.
4086 * For aesthetic reasons, we want this statement to have a statement
4087 * number that is lower than those of the then and else branches.
4088 * In order to evaluate if will need such a statement, however, we
4089 * first construct scops for the then and else branches.
4090 * We therefore reserve a statement number if we might have to
4091 * introduce such an extra statement.
4093 * If the condition is not affine, then the scop is created in
4094 * extract_non_affine_if.
4096 * If there are any breaks or continues in the then and/or else
4097 * branches, then we may have to compute a new skip condition.
4098 * This is handled using a pet_skip_info_if object.
4099 * On initialization, the object checks if skip conditions need
4100 * to be computed. If so, it does so in "extract" and adds them in "add".
4102 struct pet_scop *PetScan::extract(IfStmt *stmt)
4104 struct pet_scop *scop_then, *scop_else = NULL, *scop;
4105 isl_pw_aff *cond;
4106 int stmt_id;
4107 isl_set *set;
4108 isl_set *valid;
4110 scop = extract_conditional_assignment(stmt);
4111 if (scop)
4112 return scop;
4114 cond = try_extract_nested_condition(stmt->getCond());
4115 if (allow_nested && (!cond || has_nested(cond)))
4116 stmt_id = n_stmt++;
4119 assigned_value_cache cache(assigned_value);
4120 scop_then = extract(stmt->getThen());
4123 if (stmt->getElse()) {
4124 assigned_value_cache cache(assigned_value);
4125 scop_else = extract(stmt->getElse());
4126 if (options->autodetect) {
4127 if (scop_then && !scop_else) {
4128 partial = true;
4129 isl_pw_aff_free(cond);
4130 return scop_then;
4132 if (!scop_then && scop_else) {
4133 partial = true;
4134 isl_pw_aff_free(cond);
4135 return scop_else;
4140 if (cond &&
4141 (!is_nested_allowed(cond, scop_then) ||
4142 (stmt->getElse() && !is_nested_allowed(cond, scop_else)))) {
4143 isl_pw_aff_free(cond);
4144 cond = NULL;
4146 if (allow_nested && !cond)
4147 return extract_non_affine_if(stmt->getCond(), scop_then,
4148 scop_else, stmt->getElse(), stmt_id);
4150 if (!cond)
4151 cond = extract_condition(stmt->getCond());
4153 pet_skip_info_if skip(ctx, scop_then, scop_else, stmt->getElse(), true);
4154 skip.extract(this, cond);
4156 valid = isl_pw_aff_domain(isl_pw_aff_copy(cond));
4157 set = isl_pw_aff_non_zero_set(cond);
4158 scop = pet_scop_restrict(scop_then, isl_set_copy(set));
4160 if (stmt->getElse()) {
4161 set = isl_set_subtract(isl_set_copy(valid), set);
4162 scop_else = pet_scop_restrict(scop_else, set);
4163 scop = pet_scop_add_par(ctx, scop, scop_else);
4164 } else
4165 isl_set_free(set);
4166 scop = resolve_nested(scop);
4167 scop = pet_scop_restrict_context(scop, valid);
4169 if (skip)
4170 scop = pet_scop_prefix(scop, 0);
4171 scop = skip.add(scop, 1);
4173 return scop;
4176 /* Try and construct a pet_scop for a label statement.
4177 * We currently only allow labels on expression statements.
4179 struct pet_scop *PetScan::extract(LabelStmt *stmt)
4181 isl_id *label;
4182 Stmt *sub;
4184 sub = stmt->getSubStmt();
4185 if (!isa<Expr>(sub)) {
4186 unsupported(stmt);
4187 return NULL;
4190 label = isl_id_alloc(ctx, stmt->getName(), NULL);
4192 return extract(sub, extract_expr(cast<Expr>(sub)), label);
4195 /* Construct a pet_scop for a continue statement.
4197 * We simply create an empty scop with a universal pet_skip_now
4198 * skip condition. This skip condition will then be taken into
4199 * account by the enclosing loop construct, possibly after
4200 * being incorporated into outer skip conditions.
4202 struct pet_scop *PetScan::extract(ContinueStmt *stmt)
4204 pet_scop *scop;
4205 isl_space *space;
4206 isl_set *set;
4208 scop = pet_scop_empty(ctx);
4209 if (!scop)
4210 return NULL;
4212 space = isl_space_set_alloc(ctx, 0, 1);
4213 set = isl_set_universe(space);
4214 set = isl_set_fix_si(set, isl_dim_set, 0, 1);
4215 scop = pet_scop_set_skip(scop, pet_skip_now, set);
4217 return scop;
4220 /* Construct a pet_scop for a break statement.
4222 * We simply create an empty scop with both a universal pet_skip_now
4223 * skip condition and a universal pet_skip_later skip condition.
4224 * These skip conditions will then be taken into
4225 * account by the enclosing loop construct, possibly after
4226 * being incorporated into outer skip conditions.
4228 struct pet_scop *PetScan::extract(BreakStmt *stmt)
4230 pet_scop *scop;
4231 isl_space *space;
4232 isl_set *set;
4234 scop = pet_scop_empty(ctx);
4235 if (!scop)
4236 return NULL;
4238 space = isl_space_set_alloc(ctx, 0, 1);
4239 set = isl_set_universe(space);
4240 set = isl_set_fix_si(set, isl_dim_set, 0, 1);
4241 scop = pet_scop_set_skip(scop, pet_skip_now, isl_set_copy(set));
4242 scop = pet_scop_set_skip(scop, pet_skip_later, set);
4244 return scop;
4247 /* Try and construct a pet_scop corresponding to "stmt".
4249 * If "stmt" is a compound statement, then "skip_declarations"
4250 * indicates whether we should skip initial declarations in the
4251 * compound statement.
4253 struct pet_scop *PetScan::extract(Stmt *stmt, bool skip_declarations)
4255 if (isa<Expr>(stmt))
4256 return extract(stmt, extract_expr(cast<Expr>(stmt)));
4258 switch (stmt->getStmtClass()) {
4259 case Stmt::WhileStmtClass:
4260 return extract(cast<WhileStmt>(stmt));
4261 case Stmt::ForStmtClass:
4262 return extract_for(cast<ForStmt>(stmt));
4263 case Stmt::IfStmtClass:
4264 return extract(cast<IfStmt>(stmt));
4265 case Stmt::CompoundStmtClass:
4266 return extract(cast<CompoundStmt>(stmt), skip_declarations);
4267 case Stmt::LabelStmtClass:
4268 return extract(cast<LabelStmt>(stmt));
4269 case Stmt::ContinueStmtClass:
4270 return extract(cast<ContinueStmt>(stmt));
4271 case Stmt::BreakStmtClass:
4272 return extract(cast<BreakStmt>(stmt));
4273 case Stmt::DeclStmtClass:
4274 return extract(cast<DeclStmt>(stmt));
4275 default:
4276 unsupported(stmt);
4279 return NULL;
4282 /* Do we need to construct a skip condition of the given type
4283 * on a sequence of statements?
4285 * There is no need to construct a new skip condition if only
4286 * only of the two statements has a skip condition or if both
4287 * of their skip conditions are affine.
4289 * In principle we also don't need a new continuation variable if
4290 * the continuation of scop2 is affine, but then we would need
4291 * to allow more complicated forms of continuations.
4293 static bool need_skip_seq(struct pet_scop *scop1, struct pet_scop *scop2,
4294 enum pet_skip type)
4296 if (!scop1 || !pet_scop_has_skip(scop1, type))
4297 return false;
4298 if (!scop2 || !pet_scop_has_skip(scop2, type))
4299 return false;
4300 if (pet_scop_has_affine_skip(scop1, type) &&
4301 pet_scop_has_affine_skip(scop2, type))
4302 return false;
4303 return true;
4306 /* Construct a scop for computing the skip condition of the given type and
4307 * with access relation "skip_access" for a sequence of two scops "scop1"
4308 * and "scop2".
4310 * The computed scop contains a single statement that essentially does
4312 * skip_cond = skip_cond_1 ? 1 : skip_cond_2
4314 * or, in other words, skip_cond1 || skip_cond2.
4315 * In this expression, skip_cond_2 is filtered to reflect that it is
4316 * only evaluated when skip_cond_1 is false.
4318 * The skip condition on scop1 is not removed because it still needs
4319 * to be applied to scop2 when these two scops are combined.
4321 static struct pet_scop *extract_skip_seq(PetScan *ps,
4322 __isl_take isl_map *skip_access,
4323 struct pet_scop *scop1, struct pet_scop *scop2, enum pet_skip type)
4325 isl_map *access;
4326 struct pet_expr *expr1, *expr2, *expr, *expr_skip;
4327 struct pet_stmt *stmt;
4328 struct pet_scop *scop;
4329 isl_ctx *ctx = ps->ctx;
4331 if (!scop1 || !scop2)
4332 goto error;
4334 expr1 = pet_scop_get_skip_expr(scop1, type);
4335 expr2 = pet_scop_get_skip_expr(scop2, type);
4336 pet_scop_reset_skip(scop2, type);
4338 expr2 = pet_expr_filter(expr2, isl_map_copy(expr1->acc.access), 0);
4340 expr = universally_true(ctx);
4341 expr = pet_expr_new_ternary(ctx, expr1, expr, expr2);
4342 expr_skip = pet_expr_from_access(isl_map_copy(skip_access));
4343 if (expr_skip) {
4344 expr_skip->acc.write = 1;
4345 expr_skip->acc.read = 0;
4347 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
4348 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, ps->n_stmt++, expr);
4350 scop = pet_scop_from_pet_stmt(ctx, stmt);
4351 scop = scop_add_array(scop, skip_access, ps->ast_context);
4352 isl_map_free(skip_access);
4354 return scop;
4355 error:
4356 isl_map_free(skip_access);
4357 return NULL;
4360 /* Structure that handles the construction of skip conditions
4361 * on sequences of statements.
4363 * scop1 and scop2 represent the two statements that are combined
4365 struct pet_skip_info_seq : public pet_skip_info {
4366 struct pet_scop *scop1, *scop2;
4368 pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4369 struct pet_scop *scop2);
4370 void extract(PetScan *scan, enum pet_skip type);
4371 void extract(PetScan *scan);
4372 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
4373 int offset);
4374 struct pet_scop *add(struct pet_scop *scop, int offset);
4377 /* Initialize a pet_skip_info_seq structure based on
4378 * on the two statements that are going to be combined.
4380 pet_skip_info_seq::pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4381 struct pet_scop *scop2) : pet_skip_info(ctx), scop1(scop1), scop2(scop2)
4383 skip[pet_skip_now] = need_skip_seq(scop1, scop2, pet_skip_now);
4384 equal = skip[pet_skip_now] && skip_equals_skip_later(scop1) &&
4385 skip_equals_skip_later(scop2);
4386 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
4387 need_skip_seq(scop1, scop2, pet_skip_later);
4390 /* If we need to construct a skip condition of the given type,
4391 * then do so now.
4393 void pet_skip_info_seq::extract(PetScan *scan, enum pet_skip type)
4395 if (!skip[type])
4396 return;
4398 access[type] = create_test_access(ctx, scan->n_test++);
4399 scop[type] = extract_skip_seq(scan, isl_map_copy(access[type]),
4400 scop1, scop2, type);
4403 /* Construct the required skip conditions.
4405 void pet_skip_info_seq::extract(PetScan *scan)
4407 extract(scan, pet_skip_now);
4408 extract(scan, pet_skip_later);
4409 if (equal)
4410 drop_skip_later(scop1, scop2);
4413 /* Add the computed skip condition of the give type to "main" and
4414 * add the scop for computing the condition at the given offset (the statement
4415 * number). Within this offset, the condition is computed at position 1
4416 * to ensure that it is computed after the corresponding statement.
4418 * If equal is set, then we only computed a skip condition for pet_skip_now,
4419 * but we also need to set it as main's pet_skip_later.
4421 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *main,
4422 enum pet_skip type, int offset)
4424 isl_set *skip_set;
4426 if (!skip[type])
4427 return main;
4429 skip_set = isl_map_range(access[type]);
4430 access[type] = NULL;
4431 scop[type] = pet_scop_prefix(scop[type], 1);
4432 scop[type] = pet_scop_prefix(scop[type], offset);
4433 main = pet_scop_add_par(ctx, main, scop[type]);
4434 scop[type] = NULL;
4436 if (equal)
4437 main = pet_scop_set_skip(main, pet_skip_later,
4438 isl_set_copy(skip_set));
4440 main = pet_scop_set_skip(main, type, skip_set);
4442 return main;
4445 /* Add the computed skip conditions to "main" and
4446 * add the scops for computing the conditions at the given offset.
4448 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *scop, int offset)
4450 scop = add(scop, pet_skip_now, offset);
4451 scop = add(scop, pet_skip_later, offset);
4453 return scop;
4456 /* Extract a clone of the kill statement in "scop".
4457 * "scop" is expected to have been created from a DeclStmt
4458 * and should have the kill as its first statement.
4460 struct pet_stmt *PetScan::extract_kill(struct pet_scop *scop)
4462 struct pet_expr *kill;
4463 struct pet_stmt *stmt;
4464 isl_map *access;
4466 if (!scop)
4467 return NULL;
4468 if (scop->n_stmt < 1)
4469 isl_die(ctx, isl_error_internal,
4470 "expecting at least one statement", return NULL);
4471 stmt = scop->stmts[0];
4472 if (stmt->body->type != pet_expr_unary ||
4473 stmt->body->op != pet_op_kill)
4474 isl_die(ctx, isl_error_internal,
4475 "expecting kill statement", return NULL);
4477 access = isl_map_copy(stmt->body->args[0]->acc.access);
4478 access = isl_map_reset_tuple_id(access, isl_dim_in);
4479 kill = pet_expr_kill_from_access(access);
4480 return pet_stmt_from_pet_expr(ctx, stmt->line, NULL, n_stmt++, kill);
4483 /* Mark all arrays in "scop" as being exposed.
4485 static struct pet_scop *mark_exposed(struct pet_scop *scop)
4487 if (!scop)
4488 return NULL;
4489 for (int i = 0; i < scop->n_array; ++i)
4490 scop->arrays[i]->exposed = 1;
4491 return scop;
4494 /* Try and construct a pet_scop corresponding to (part of)
4495 * a sequence of statements.
4497 * "block" is set if the sequence respresents the children of
4498 * a compound statement.
4499 * "skip_declarations" is set if we should skip initial declarations
4500 * in the sequence of statements.
4502 * If there are any breaks or continues in the individual statements,
4503 * then we may have to compute a new skip condition.
4504 * This is handled using a pet_skip_info_seq object.
4505 * On initialization, the object checks if skip conditions need
4506 * to be computed. If so, it does so in "extract" and adds them in "add".
4508 * If "block" is set, then we need to insert kill statements at
4509 * the end of the block for any array that has been declared by
4510 * one of the statements in the sequence. Each of these declarations
4511 * results in the construction of a kill statement at the place
4512 * of the declaration, so we simply collect duplicates of
4513 * those kill statements and append these duplicates to the constructed scop.
4515 * If "block" is not set, then any array declared by one of the statements
4516 * in the sequence is marked as being exposed.
4518 struct pet_scop *PetScan::extract(StmtRange stmt_range, bool block,
4519 bool skip_declarations)
4521 pet_scop *scop;
4522 StmtIterator i;
4523 int j;
4524 bool partial_range = false;
4525 set<struct pet_stmt *> kills;
4526 set<struct pet_stmt *>::iterator it;
4528 scop = pet_scop_empty(ctx);
4529 for (i = stmt_range.first, j = 0; i != stmt_range.second; ++i, ++j) {
4530 Stmt *child = *i;
4531 struct pet_scop *scop_i;
4533 if (skip_declarations &&
4534 child->getStmtClass() == Stmt::DeclStmtClass)
4535 continue;
4537 scop_i = extract(child);
4538 if (scop && partial) {
4539 pet_scop_free(scop_i);
4540 break;
4542 pet_skip_info_seq skip(ctx, scop, scop_i);
4543 skip.extract(this);
4544 if (skip)
4545 scop_i = pet_scop_prefix(scop_i, 0);
4546 if (scop_i && child->getStmtClass() == Stmt::DeclStmtClass) {
4547 if (block)
4548 kills.insert(extract_kill(scop_i));
4549 else
4550 scop_i = mark_exposed(scop_i);
4552 scop_i = pet_scop_prefix(scop_i, j);
4553 if (options->autodetect) {
4554 if (scop_i)
4555 scop = pet_scop_add_seq(ctx, scop, scop_i);
4556 else
4557 partial_range = true;
4558 if (scop->n_stmt != 0 && !scop_i)
4559 partial = true;
4560 } else {
4561 scop = pet_scop_add_seq(ctx, scop, scop_i);
4564 scop = skip.add(scop, j);
4566 if (partial)
4567 break;
4570 for (it = kills.begin(); it != kills.end(); ++it) {
4571 pet_scop *scop_j;
4572 scop_j = pet_scop_from_pet_stmt(ctx, *it);
4573 scop_j = pet_scop_prefix(scop_j, j);
4574 scop = pet_scop_add_seq(ctx, scop, scop_j);
4577 if (scop && partial_range)
4578 partial = true;
4580 return scop;
4583 /* Return the file offset of the expansion location of "Loc".
4585 static unsigned getExpansionOffset(SourceManager &SM, SourceLocation Loc)
4587 return SM.getFileOffset(SM.getExpansionLoc(Loc));
4590 /* Check if the scop marked by the user is exactly this Stmt
4591 * or part of this Stmt.
4592 * If so, return a pet_scop corresponding to the marked region.
4593 * Otherwise, return NULL.
4595 struct pet_scop *PetScan::scan(Stmt *stmt)
4597 SourceManager &SM = PP.getSourceManager();
4598 unsigned start_off, end_off;
4600 start_off = getExpansionOffset(SM, stmt->getLocStart());
4601 end_off = getExpansionOffset(SM, stmt->getLocEnd());
4603 if (start_off > loc.end)
4604 return NULL;
4605 if (end_off < loc.start)
4606 return NULL;
4607 if (start_off >= loc.start && end_off <= loc.end) {
4608 return extract(stmt);
4611 StmtIterator start;
4612 for (start = stmt->child_begin(); start != stmt->child_end(); ++start) {
4613 Stmt *child = *start;
4614 if (!child)
4615 continue;
4616 start_off = getExpansionOffset(SM, child->getLocStart());
4617 end_off = getExpansionOffset(SM, child->getLocEnd());
4618 if (start_off < loc.start && end_off > loc.end)
4619 return scan(child);
4620 if (start_off >= loc.start)
4621 break;
4624 StmtIterator end;
4625 for (end = start; end != stmt->child_end(); ++end) {
4626 Stmt *child = *end;
4627 start_off = SM.getFileOffset(child->getLocStart());
4628 if (start_off >= loc.end)
4629 break;
4632 return extract(StmtRange(start, end), false, false);
4635 /* Set the size of index "pos" of "array" to "size".
4636 * In particular, add a constraint of the form
4638 * i_pos < size
4640 * to array->extent and a constraint of the form
4642 * size >= 0
4644 * to array->context.
4646 static struct pet_array *update_size(struct pet_array *array, int pos,
4647 __isl_take isl_pw_aff *size)
4649 isl_set *valid;
4650 isl_set *univ;
4651 isl_set *bound;
4652 isl_space *dim;
4653 isl_aff *aff;
4654 isl_pw_aff *index;
4655 isl_id *id;
4657 valid = isl_pw_aff_nonneg_set(isl_pw_aff_copy(size));
4658 array->context = isl_set_intersect(array->context, valid);
4660 dim = isl_set_get_space(array->extent);
4661 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
4662 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, pos, 1);
4663 univ = isl_set_universe(isl_aff_get_domain_space(aff));
4664 index = isl_pw_aff_alloc(univ, aff);
4666 size = isl_pw_aff_add_dims(size, isl_dim_in,
4667 isl_set_dim(array->extent, isl_dim_set));
4668 id = isl_set_get_tuple_id(array->extent);
4669 size = isl_pw_aff_set_tuple_id(size, isl_dim_in, id);
4670 bound = isl_pw_aff_lt_set(index, size);
4672 array->extent = isl_set_intersect(array->extent, bound);
4674 if (!array->context || !array->extent)
4675 goto error;
4677 return array;
4678 error:
4679 pet_array_free(array);
4680 return NULL;
4683 /* Figure out the size of the array at position "pos" and all
4684 * subsequent positions from "type" and update "array" accordingly.
4686 struct pet_array *PetScan::set_upper_bounds(struct pet_array *array,
4687 const Type *type, int pos)
4689 const ArrayType *atype;
4690 isl_pw_aff *size;
4692 if (!array)
4693 return NULL;
4695 if (type->isPointerType()) {
4696 type = type->getPointeeType().getTypePtr();
4697 return set_upper_bounds(array, type, pos + 1);
4699 if (!type->isArrayType())
4700 return array;
4702 type = type->getCanonicalTypeInternal().getTypePtr();
4703 atype = cast<ArrayType>(type);
4705 if (type->isConstantArrayType()) {
4706 const ConstantArrayType *ca = cast<ConstantArrayType>(atype);
4707 size = extract_affine(ca->getSize());
4708 array = update_size(array, pos, size);
4709 } else if (type->isVariableArrayType()) {
4710 const VariableArrayType *vla = cast<VariableArrayType>(atype);
4711 size = extract_affine(vla->getSizeExpr());
4712 array = update_size(array, pos, size);
4715 type = atype->getElementType().getTypePtr();
4717 return set_upper_bounds(array, type, pos + 1);
4720 /* Is "T" the type of a variable length array with static size?
4722 static bool is_vla_with_static_size(QualType T)
4724 const VariableArrayType *vlatype;
4726 if (!T->isVariableArrayType())
4727 return false;
4728 vlatype = cast<VariableArrayType>(T);
4729 return vlatype->getSizeModifier() == VariableArrayType::Static;
4732 /* Return the type of "decl" as an array.
4734 * In particular, if "decl" is a parameter declaration that
4735 * is a variable length array with a static size, then
4736 * return the original type (i.e., the variable length array).
4737 * Otherwise, return the type of decl.
4739 static QualType get_array_type(ValueDecl *decl)
4741 ParmVarDecl *parm;
4742 QualType T;
4744 parm = dyn_cast<ParmVarDecl>(decl);
4745 if (!parm)
4746 return decl->getType();
4748 T = parm->getOriginalType();
4749 if (!is_vla_with_static_size(T))
4750 return decl->getType();
4751 return T;
4754 /* Construct and return a pet_array corresponding to the variable "decl".
4755 * In particular, initialize array->extent to
4757 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
4759 * and then call set_upper_bounds to set the upper bounds on the indices
4760 * based on the type of the variable.
4762 struct pet_array *PetScan::extract_array(isl_ctx *ctx, ValueDecl *decl)
4764 struct pet_array *array;
4765 QualType qt = get_array_type(decl);
4766 const Type *type = qt.getTypePtr();
4767 int depth = array_depth(type);
4768 QualType base = base_type(qt);
4769 string name;
4770 isl_id *id;
4771 isl_space *dim;
4773 array = isl_calloc_type(ctx, struct pet_array);
4774 if (!array)
4775 return NULL;
4777 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
4778 dim = isl_space_set_alloc(ctx, 0, depth);
4779 dim = isl_space_set_tuple_id(dim, isl_dim_set, id);
4781 array->extent = isl_set_nat_universe(dim);
4783 dim = isl_space_params_alloc(ctx, 0);
4784 array->context = isl_set_universe(dim);
4786 array = set_upper_bounds(array, type, 0);
4787 if (!array)
4788 return NULL;
4790 name = base.getAsString();
4791 array->element_type = strdup(name.c_str());
4792 array->element_size = decl->getASTContext().getTypeInfo(base).first / 8;
4794 return array;
4797 /* Construct a list of pet_arrays, one for each array (or scalar)
4798 * accessed inside "scop", add this list to "scop" and return the result.
4800 * The context of "scop" is updated with the intersection of
4801 * the contexts of all arrays, i.e., constraints on the parameters
4802 * that ensure that the arrays have a valid (non-negative) size.
4804 struct pet_scop *PetScan::scan_arrays(struct pet_scop *scop)
4806 int i;
4807 set<ValueDecl *> arrays;
4808 set<ValueDecl *>::iterator it;
4809 int n_array;
4810 struct pet_array **scop_arrays;
4812 if (!scop)
4813 return NULL;
4815 pet_scop_collect_arrays(scop, arrays);
4816 if (arrays.size() == 0)
4817 return scop;
4819 n_array = scop->n_array;
4821 scop_arrays = isl_realloc_array(ctx, scop->arrays, struct pet_array *,
4822 n_array + arrays.size());
4823 if (!scop_arrays)
4824 goto error;
4825 scop->arrays = scop_arrays;
4827 for (it = arrays.begin(), i = 0; it != arrays.end(); ++it, ++i) {
4828 struct pet_array *array;
4829 scop->arrays[n_array + i] = array = extract_array(ctx, *it);
4830 if (!scop->arrays[n_array + i])
4831 goto error;
4832 scop->n_array++;
4833 scop->context = isl_set_intersect(scop->context,
4834 isl_set_copy(array->context));
4835 if (!scop->context)
4836 goto error;
4839 return scop;
4840 error:
4841 pet_scop_free(scop);
4842 return NULL;
4845 /* Bound all parameters in scop->context to the possible values
4846 * of the corresponding C variable.
4848 static struct pet_scop *add_parameter_bounds(struct pet_scop *scop)
4850 int n;
4852 if (!scop)
4853 return NULL;
4855 n = isl_set_dim(scop->context, isl_dim_param);
4856 for (int i = 0; i < n; ++i) {
4857 isl_id *id;
4858 ValueDecl *decl;
4860 id = isl_set_get_dim_id(scop->context, isl_dim_param, i);
4861 if (is_nested_parameter(id)) {
4862 isl_id_free(id);
4863 isl_die(isl_set_get_ctx(scop->context),
4864 isl_error_internal,
4865 "unresolved nested parameter", goto error);
4867 decl = (ValueDecl *) isl_id_get_user(id);
4868 isl_id_free(id);
4870 scop->context = set_parameter_bounds(scop->context, i, decl);
4872 if (!scop->context)
4873 goto error;
4876 return scop;
4877 error:
4878 pet_scop_free(scop);
4879 return NULL;
4882 /* Construct a pet_scop from the given function.
4884 struct pet_scop *PetScan::scan(FunctionDecl *fd)
4886 pet_scop *scop;
4887 Stmt *stmt;
4889 stmt = fd->getBody();
4891 if (options->autodetect)
4892 scop = extract(stmt, true);
4893 else
4894 scop = scan(stmt);
4895 scop = pet_scop_detect_parameter_accesses(scop);
4896 scop = scan_arrays(scop);
4897 scop = add_parameter_bounds(scop);
4898 scop = pet_scop_gist(scop, value_bounds);
4900 return scop;