PetScan::extract_int: return isl_val
[pet.git] / scan.cc
blob65242340634402fe5d5ec3b12836c4deea13320c
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 <string.h>
36 #include <set>
37 #include <map>
38 #include <iostream>
39 #include <llvm/Support/raw_ostream.h>
40 #include <clang/AST/ASTContext.h>
41 #include <clang/AST/ASTDiagnostic.h>
42 #include <clang/AST/Expr.h>
43 #include <clang/AST/RecursiveASTVisitor.h>
45 #include <isl/id.h>
46 #include <isl/space.h>
47 #include <isl/aff.h>
48 #include <isl/set.h>
50 #include "options.h"
51 #include "scan.h"
52 #include "scop.h"
53 #include "scop_plus.h"
55 #include "config.h"
57 using namespace std;
58 using namespace clang;
60 #if defined(DECLREFEXPR_CREATE_REQUIRES_BOOL)
61 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
63 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
64 SourceLocation(), var, false, var->getInnerLocStart(),
65 var->getType(), VK_LValue);
67 #elif defined(DECLREFEXPR_CREATE_REQUIRES_SOURCELOCATION)
68 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
70 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
71 SourceLocation(), var, var->getInnerLocStart(), var->getType(),
72 VK_LValue);
74 #else
75 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
77 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
78 var, var->getInnerLocStart(), var->getType(), VK_LValue);
80 #endif
82 /* Check if the element type corresponding to the given array type
83 * has a const qualifier.
85 static bool const_base(QualType qt)
87 const Type *type = qt.getTypePtr();
89 if (type->isPointerType())
90 return const_base(type->getPointeeType());
91 if (type->isArrayType()) {
92 const ArrayType *atype;
93 type = type->getCanonicalTypeInternal().getTypePtr();
94 atype = cast<ArrayType>(type);
95 return const_base(atype->getElementType());
98 return qt.isConstQualified();
101 /* Mark "decl" as having an unknown value in "assigned_value".
103 * If no (known or unknown) value was assigned to "decl" before,
104 * then it may have been treated as a parameter before and may
105 * therefore appear in a value assigned to another variable.
106 * If so, this assignment needs to be turned into an unknown value too.
108 static void clear_assignment(map<ValueDecl *, isl_pw_aff *> &assigned_value,
109 ValueDecl *decl)
111 map<ValueDecl *, isl_pw_aff *>::iterator it;
113 it = assigned_value.find(decl);
115 assigned_value[decl] = NULL;
117 if (it == assigned_value.end())
118 return;
120 for (it = assigned_value.begin(); it != assigned_value.end(); ++it) {
121 isl_pw_aff *pa = it->second;
122 int nparam = isl_pw_aff_dim(pa, isl_dim_param);
124 for (int i = 0; i < nparam; ++i) {
125 isl_id *id;
127 if (!isl_pw_aff_has_dim_id(pa, isl_dim_param, i))
128 continue;
129 id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
130 if (isl_id_get_user(id) == decl)
131 it->second = NULL;
132 isl_id_free(id);
137 /* Look for any assignments to scalar variables in part of the parse
138 * tree and set assigned_value to NULL for each of them.
139 * Also reset assigned_value if the address of a scalar variable
140 * is being taken. As an exception, if the address is passed to a function
141 * that is declared to receive a const pointer, then assigned_value is
142 * not reset.
144 * This ensures that we won't use any previously stored value
145 * in the current subtree and its parents.
147 struct clear_assignments : RecursiveASTVisitor<clear_assignments> {
148 map<ValueDecl *, isl_pw_aff *> &assigned_value;
149 set<UnaryOperator *> skip;
151 clear_assignments(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
152 assigned_value(assigned_value) {}
154 /* Check for "address of" operators whose value is passed
155 * to a const pointer argument and add them to "skip", so that
156 * we can skip them in VisitUnaryOperator.
158 bool VisitCallExpr(CallExpr *expr) {
159 FunctionDecl *fd;
160 fd = expr->getDirectCallee();
161 if (!fd)
162 return true;
163 for (int i = 0; i < expr->getNumArgs(); ++i) {
164 Expr *arg = expr->getArg(i);
165 UnaryOperator *op;
166 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
167 ImplicitCastExpr *ice;
168 ice = cast<ImplicitCastExpr>(arg);
169 arg = ice->getSubExpr();
171 if (arg->getStmtClass() != Stmt::UnaryOperatorClass)
172 continue;
173 op = cast<UnaryOperator>(arg);
174 if (op->getOpcode() != UO_AddrOf)
175 continue;
176 if (const_base(fd->getParamDecl(i)->getType()))
177 skip.insert(op);
179 return true;
182 bool VisitUnaryOperator(UnaryOperator *expr) {
183 Expr *arg;
184 DeclRefExpr *ref;
185 ValueDecl *decl;
187 switch (expr->getOpcode()) {
188 case UO_AddrOf:
189 case UO_PostInc:
190 case UO_PostDec:
191 case UO_PreInc:
192 case UO_PreDec:
193 break;
194 default:
195 return true;
197 if (skip.find(expr) != skip.end())
198 return true;
200 arg = expr->getSubExpr();
201 if (arg->getStmtClass() != Stmt::DeclRefExprClass)
202 return true;
203 ref = cast<DeclRefExpr>(arg);
204 decl = ref->getDecl();
205 clear_assignment(assigned_value, decl);
206 return true;
209 bool VisitBinaryOperator(BinaryOperator *expr) {
210 Expr *lhs;
211 DeclRefExpr *ref;
212 ValueDecl *decl;
214 if (!expr->isAssignmentOp())
215 return true;
216 lhs = expr->getLHS();
217 if (lhs->getStmtClass() != Stmt::DeclRefExprClass)
218 return true;
219 ref = cast<DeclRefExpr>(lhs);
220 decl = ref->getDecl();
221 clear_assignment(assigned_value, decl);
222 return true;
226 /* Keep a copy of the currently assigned values.
228 * Any variable that is assigned a value inside the current scope
229 * is removed again when we leave the scope (either because it wasn't
230 * stored in the cache or because it has a different value in the cache).
232 struct assigned_value_cache {
233 map<ValueDecl *, isl_pw_aff *> &assigned_value;
234 map<ValueDecl *, isl_pw_aff *> cache;
236 assigned_value_cache(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
237 assigned_value(assigned_value), cache(assigned_value) {}
238 ~assigned_value_cache() {
239 map<ValueDecl *, isl_pw_aff *>::iterator it = cache.begin();
240 for (it = assigned_value.begin(); it != assigned_value.end();
241 ++it) {
242 if (!it->second ||
243 (cache.find(it->first) != cache.end() &&
244 cache[it->first] != it->second))
245 cache[it->first] = NULL;
247 assigned_value = cache;
251 /* Insert an expression into the collection of expressions,
252 * provided it is not already in there.
253 * The isl_pw_affs are freed in the destructor.
255 void PetScan::insert_expression(__isl_take isl_pw_aff *expr)
257 std::set<isl_pw_aff *>::iterator it;
259 if (expressions.find(expr) == expressions.end())
260 expressions.insert(expr);
261 else
262 isl_pw_aff_free(expr);
265 PetScan::~PetScan()
267 std::set<isl_pw_aff *>::iterator it;
269 for (it = expressions.begin(); it != expressions.end(); ++it)
270 isl_pw_aff_free(*it);
272 isl_union_map_free(value_bounds);
275 /* Called if we found something we (currently) cannot handle.
276 * We'll provide more informative warnings later.
278 * We only actually complain if autodetect is false.
280 void PetScan::unsupported(Stmt *stmt, const char *msg)
282 if (options->autodetect)
283 return;
285 SourceLocation loc = stmt->getLocStart();
286 DiagnosticsEngine &diag = PP.getDiagnostics();
287 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
288 msg ? msg : "unsupported");
289 DiagnosticBuilder B = diag.Report(loc, id) << stmt->getSourceRange();
292 /* Extract an integer from "expr".
294 __isl_give isl_val *PetScan::extract_int(isl_ctx *ctx, IntegerLiteral *expr)
296 const Type *type = expr->getType().getTypePtr();
297 int is_signed = type->hasSignedIntegerRepresentation();
299 if (is_signed) {
300 int64_t i = expr->getValue().getSExtValue();
301 return isl_val_int_from_si(ctx, i);
302 } else {
303 uint64_t i = expr->getValue().getZExtValue();
304 return isl_val_int_from_ui(ctx, i);
308 /* Extract an integer from "expr".
309 * Return NULL if "expr" does not (obviously) represent an integer.
311 __isl_give isl_val *PetScan::extract_int(clang::ParenExpr *expr)
313 return extract_int(expr->getSubExpr());
316 /* Extract an integer from "expr".
317 * Return NULL if "expr" does not (obviously) represent an integer.
319 __isl_give isl_val *PetScan::extract_int(clang::Expr *expr)
321 if (expr->getStmtClass() == Stmt::IntegerLiteralClass)
322 return extract_int(ctx, cast<IntegerLiteral>(expr));
323 if (expr->getStmtClass() == Stmt::ParenExprClass)
324 return extract_int(cast<ParenExpr>(expr));
326 unsupported(expr);
327 return NULL;
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_val *v;
340 v = extract_int(expr);
341 aff = isl_aff_add_constant_val(aff, v);
343 return isl_pw_aff_alloc(dom, aff);
346 /* Extract an affine expression from the APInt "val".
348 __isl_give isl_pw_aff *PetScan::extract_affine(const llvm::APInt &val)
350 isl_space *dim = isl_space_params_alloc(ctx, 0);
351 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
352 isl_aff *aff = isl_aff_zero_on_domain(ls);
353 isl_set *dom = isl_set_universe(dim);
354 isl_int v;
356 isl_int_init(v);
357 isl_int_set_ui(v, val.getZExtValue());
358 aff = isl_aff_add_constant(aff, v);
359 isl_int_clear(v);
361 return isl_pw_aff_alloc(dom, aff);
364 __isl_give isl_pw_aff *PetScan::extract_affine(ImplicitCastExpr *expr)
366 return extract_affine(expr->getSubExpr());
369 static unsigned get_type_size(ValueDecl *decl)
371 return decl->getASTContext().getIntWidth(decl->getType());
374 /* Bound parameter "pos" of "set" to the possible values of "decl".
376 static __isl_give isl_set *set_parameter_bounds(__isl_take isl_set *set,
377 unsigned pos, ValueDecl *decl)
379 unsigned width;
380 isl_int v;
382 isl_int_init(v);
384 width = get_type_size(decl);
385 if (decl->getType()->isUnsignedIntegerType()) {
386 set = isl_set_lower_bound_si(set, isl_dim_param, pos, 0);
387 isl_int_set_si(v, 1);
388 isl_int_mul_2exp(v, v, width);
389 isl_int_sub_ui(v, v, 1);
390 set = isl_set_upper_bound(set, isl_dim_param, pos, v);
391 } else {
392 isl_int_set_si(v, 1);
393 isl_int_mul_2exp(v, v, width - 1);
394 isl_int_sub_ui(v, v, 1);
395 set = isl_set_upper_bound(set, isl_dim_param, pos, v);
396 isl_int_neg(v, v);
397 isl_int_sub_ui(v, v, 1);
398 set = isl_set_lower_bound(set, isl_dim_param, pos, v);
401 isl_int_clear(v);
403 return set;
406 /* Extract an affine expression from the DeclRefExpr "expr".
408 * If the variable has been assigned a value, then we check whether
409 * we know what (affine) value was assigned.
410 * If so, we return this value. Otherwise we convert "expr"
411 * to an extra parameter (provided nesting_enabled is set).
413 * Otherwise, we simply return an expression that is equal
414 * to a parameter corresponding to the referenced variable.
416 __isl_give isl_pw_aff *PetScan::extract_affine(DeclRefExpr *expr)
418 ValueDecl *decl = expr->getDecl();
419 const Type *type = decl->getType().getTypePtr();
420 isl_id *id;
421 isl_space *dim;
422 isl_aff *aff;
423 isl_set *dom;
425 if (!type->isIntegerType()) {
426 unsupported(expr);
427 return NULL;
430 if (assigned_value.find(decl) != assigned_value.end()) {
431 if (assigned_value[decl])
432 return isl_pw_aff_copy(assigned_value[decl]);
433 else
434 return nested_access(expr);
437 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
438 dim = isl_space_params_alloc(ctx, 1);
440 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
442 dom = isl_set_universe(isl_space_copy(dim));
443 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
444 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
446 return isl_pw_aff_alloc(dom, aff);
449 /* Extract an affine expression from an integer division operation.
450 * In particular, if "expr" is lhs/rhs, then return
452 * lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs)
454 * The second argument (rhs) is required to be a (positive) integer constant.
456 __isl_give isl_pw_aff *PetScan::extract_affine_div(BinaryOperator *expr)
458 int is_cst;
459 isl_pw_aff *rhs, *lhs;
461 rhs = extract_affine(expr->getRHS());
462 is_cst = isl_pw_aff_is_cst(rhs);
463 if (is_cst < 0 || !is_cst) {
464 isl_pw_aff_free(rhs);
465 if (!is_cst)
466 unsupported(expr);
467 return NULL;
470 lhs = extract_affine(expr->getLHS());
472 return isl_pw_aff_tdiv_q(lhs, rhs);
475 /* Extract an affine expression from a modulo operation.
476 * In particular, if "expr" is lhs/rhs, then return
478 * lhs - rhs * (lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs))
480 * The second argument (rhs) is required to be a (positive) integer constant.
482 __isl_give isl_pw_aff *PetScan::extract_affine_mod(BinaryOperator *expr)
484 int is_cst;
485 isl_pw_aff *rhs, *lhs;
487 rhs = extract_affine(expr->getRHS());
488 is_cst = isl_pw_aff_is_cst(rhs);
489 if (is_cst < 0 || !is_cst) {
490 isl_pw_aff_free(rhs);
491 if (!is_cst)
492 unsupported(expr);
493 return NULL;
496 lhs = extract_affine(expr->getLHS());
498 return isl_pw_aff_tdiv_r(lhs, rhs);
501 /* Extract an affine expression from a multiplication operation.
502 * This is only allowed if at least one of the two arguments
503 * is a (piecewise) constant.
505 __isl_give isl_pw_aff *PetScan::extract_affine_mul(BinaryOperator *expr)
507 isl_pw_aff *lhs;
508 isl_pw_aff *rhs;
510 lhs = extract_affine(expr->getLHS());
511 rhs = extract_affine(expr->getRHS());
513 if (!isl_pw_aff_is_cst(lhs) && !isl_pw_aff_is_cst(rhs)) {
514 isl_pw_aff_free(lhs);
515 isl_pw_aff_free(rhs);
516 unsupported(expr);
517 return NULL;
520 return isl_pw_aff_mul(lhs, rhs);
523 /* Extract an affine expression from an addition or subtraction operation.
525 __isl_give isl_pw_aff *PetScan::extract_affine_add(BinaryOperator *expr)
527 isl_pw_aff *lhs;
528 isl_pw_aff *rhs;
530 lhs = extract_affine(expr->getLHS());
531 rhs = extract_affine(expr->getRHS());
533 switch (expr->getOpcode()) {
534 case BO_Add:
535 return isl_pw_aff_add(lhs, rhs);
536 case BO_Sub:
537 return isl_pw_aff_sub(lhs, rhs);
538 default:
539 isl_pw_aff_free(lhs);
540 isl_pw_aff_free(rhs);
541 return NULL;
546 /* Compute
548 * pwaff mod 2^width
550 static __isl_give isl_pw_aff *wrap(__isl_take isl_pw_aff *pwaff,
551 unsigned width)
553 isl_int mod;
555 isl_int_init(mod);
556 isl_int_set_si(mod, 1);
557 isl_int_mul_2exp(mod, mod, width);
559 pwaff = isl_pw_aff_mod(pwaff, mod);
561 isl_int_clear(mod);
563 return pwaff;
566 /* Limit the domain of "pwaff" to those elements where the function
567 * value satisfies
569 * 2^{width-1} <= pwaff < 2^{width-1}
571 static __isl_give isl_pw_aff *avoid_overflow(__isl_take isl_pw_aff *pwaff,
572 unsigned width)
574 isl_int v;
575 isl_space *space = isl_pw_aff_get_domain_space(pwaff);
576 isl_local_space *ls = isl_local_space_from_space(space);
577 isl_aff *bound;
578 isl_set *dom;
579 isl_pw_aff *b;
581 isl_int_init(v);
582 isl_int_set_si(v, 1);
583 isl_int_mul_2exp(v, v, width - 1);
585 bound = isl_aff_zero_on_domain(ls);
586 bound = isl_aff_add_constant(bound, v);
587 b = isl_pw_aff_from_aff(bound);
589 dom = isl_pw_aff_lt_set(isl_pw_aff_copy(pwaff), isl_pw_aff_copy(b));
590 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
592 b = isl_pw_aff_neg(b);
593 dom = isl_pw_aff_ge_set(isl_pw_aff_copy(pwaff), b);
594 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
596 isl_int_clear(v);
598 return pwaff;
601 /* Handle potential overflows on signed computations.
603 * If options->signed_overflow is set to PET_OVERFLOW_AVOID,
604 * the we adjust the domain of "pa" to avoid overflows.
606 __isl_give isl_pw_aff *PetScan::signed_overflow(__isl_take isl_pw_aff *pa,
607 unsigned width)
609 if (options->signed_overflow == PET_OVERFLOW_AVOID)
610 pa = avoid_overflow(pa, width);
612 return pa;
615 /* Return the piecewise affine expression "set ? 1 : 0" defined on "dom".
617 static __isl_give isl_pw_aff *indicator_function(__isl_take isl_set *set,
618 __isl_take isl_set *dom)
620 isl_pw_aff *pa;
621 pa = isl_set_indicator_function(set);
622 pa = isl_pw_aff_intersect_domain(pa, dom);
623 return pa;
626 /* Extract an affine expression from some binary operations.
627 * If the result of the expression is unsigned, then we wrap it
628 * based on the size of the type. Otherwise, we ensure that
629 * no overflow occurs.
631 __isl_give isl_pw_aff *PetScan::extract_affine(BinaryOperator *expr)
633 isl_pw_aff *res;
634 unsigned width;
636 switch (expr->getOpcode()) {
637 case BO_Add:
638 case BO_Sub:
639 res = extract_affine_add(expr);
640 break;
641 case BO_Div:
642 res = extract_affine_div(expr);
643 break;
644 case BO_Rem:
645 res = extract_affine_mod(expr);
646 break;
647 case BO_Mul:
648 res = extract_affine_mul(expr);
649 break;
650 case BO_LT:
651 case BO_LE:
652 case BO_GT:
653 case BO_GE:
654 case BO_EQ:
655 case BO_NE:
656 case BO_LAnd:
657 case BO_LOr:
658 return extract_condition(expr);
659 default:
660 unsupported(expr);
661 return NULL;
664 width = ast_context.getIntWidth(expr->getType());
665 if (expr->getType()->isUnsignedIntegerType())
666 res = wrap(res, width);
667 else
668 res = signed_overflow(res, width);
670 return res;
673 /* Extract an affine expression from a negation operation.
675 __isl_give isl_pw_aff *PetScan::extract_affine(UnaryOperator *expr)
677 if (expr->getOpcode() == UO_Minus)
678 return isl_pw_aff_neg(extract_affine(expr->getSubExpr()));
679 if (expr->getOpcode() == UO_LNot)
680 return extract_condition(expr);
682 unsupported(expr);
683 return NULL;
686 __isl_give isl_pw_aff *PetScan::extract_affine(ParenExpr *expr)
688 return extract_affine(expr->getSubExpr());
691 /* Extract an affine expression from some special function calls.
692 * In particular, we handle "min", "max", "ceild" and "floord".
693 * In case of the latter two, the second argument needs to be
694 * a (positive) integer constant.
696 __isl_give isl_pw_aff *PetScan::extract_affine(CallExpr *expr)
698 FunctionDecl *fd;
699 string name;
700 isl_pw_aff *aff1, *aff2;
702 fd = expr->getDirectCallee();
703 if (!fd) {
704 unsupported(expr);
705 return NULL;
708 name = fd->getDeclName().getAsString();
709 if (!(expr->getNumArgs() == 2 && name == "min") &&
710 !(expr->getNumArgs() == 2 && name == "max") &&
711 !(expr->getNumArgs() == 2 && name == "floord") &&
712 !(expr->getNumArgs() == 2 && name == "ceild")) {
713 unsupported(expr);
714 return NULL;
717 if (name == "min" || name == "max") {
718 aff1 = extract_affine(expr->getArg(0));
719 aff2 = extract_affine(expr->getArg(1));
721 if (name == "min")
722 aff1 = isl_pw_aff_min(aff1, aff2);
723 else
724 aff1 = isl_pw_aff_max(aff1, aff2);
725 } else if (name == "floord" || name == "ceild") {
726 isl_val *v;
727 Expr *arg2 = expr->getArg(1);
729 if (arg2->getStmtClass() != Stmt::IntegerLiteralClass) {
730 unsupported(expr);
731 return NULL;
733 aff1 = extract_affine(expr->getArg(0));
734 v = extract_int(cast<IntegerLiteral>(arg2));
735 aff1 = isl_pw_aff_scale_down_val(aff1, v);
736 if (name == "floord")
737 aff1 = isl_pw_aff_floor(aff1);
738 else
739 aff1 = isl_pw_aff_ceil(aff1);
740 } else {
741 unsupported(expr);
742 return NULL;
745 return aff1;
748 /* This method is called when we come across an access that is
749 * nested in what is supposed to be an affine expression.
750 * If nesting is allowed, we return a new parameter that corresponds
751 * to this nested access. Otherwise, we simply complain.
753 * Note that we currently don't allow nested accesses themselves
754 * to contain any nested accesses, so we check if we can extract
755 * the access without any nesting and complain if we can't.
757 * The new parameter is resolved in resolve_nested.
759 isl_pw_aff *PetScan::nested_access(Expr *expr)
761 isl_id *id;
762 isl_space *dim;
763 isl_aff *aff;
764 isl_set *dom;
765 isl_map *access;
767 if (!nesting_enabled) {
768 unsupported(expr);
769 return NULL;
772 allow_nested = false;
773 access = extract_access(expr);
774 allow_nested = true;
775 if (!access) {
776 unsupported(expr);
777 return NULL;
779 isl_map_free(access);
781 id = isl_id_alloc(ctx, NULL, expr);
782 dim = isl_space_params_alloc(ctx, 1);
784 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
786 dom = isl_set_universe(isl_space_copy(dim));
787 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
788 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
790 return isl_pw_aff_alloc(dom, aff);
793 /* Affine expressions are not supposed to contain array accesses,
794 * but if nesting is allowed, we return a parameter corresponding
795 * to the array access.
797 __isl_give isl_pw_aff *PetScan::extract_affine(ArraySubscriptExpr *expr)
799 return nested_access(expr);
802 /* Extract an affine expression from a conditional operation.
804 __isl_give isl_pw_aff *PetScan::extract_affine(ConditionalOperator *expr)
806 isl_pw_aff *cond, *lhs, *rhs, *res;
808 cond = extract_condition(expr->getCond());
809 lhs = extract_affine(expr->getTrueExpr());
810 rhs = extract_affine(expr->getFalseExpr());
812 return isl_pw_aff_cond(cond, lhs, rhs);
815 /* Extract an affine expression, if possible, from "expr".
816 * Otherwise return NULL.
818 __isl_give isl_pw_aff *PetScan::extract_affine(Expr *expr)
820 switch (expr->getStmtClass()) {
821 case Stmt::ImplicitCastExprClass:
822 return extract_affine(cast<ImplicitCastExpr>(expr));
823 case Stmt::IntegerLiteralClass:
824 return extract_affine(cast<IntegerLiteral>(expr));
825 case Stmt::DeclRefExprClass:
826 return extract_affine(cast<DeclRefExpr>(expr));
827 case Stmt::BinaryOperatorClass:
828 return extract_affine(cast<BinaryOperator>(expr));
829 case Stmt::UnaryOperatorClass:
830 return extract_affine(cast<UnaryOperator>(expr));
831 case Stmt::ParenExprClass:
832 return extract_affine(cast<ParenExpr>(expr));
833 case Stmt::CallExprClass:
834 return extract_affine(cast<CallExpr>(expr));
835 case Stmt::ArraySubscriptExprClass:
836 return extract_affine(cast<ArraySubscriptExpr>(expr));
837 case Stmt::ConditionalOperatorClass:
838 return extract_affine(cast<ConditionalOperator>(expr));
839 default:
840 unsupported(expr);
842 return NULL;
845 __isl_give isl_map *PetScan::extract_access(ImplicitCastExpr *expr)
847 return extract_access(expr->getSubExpr());
850 /* Return the depth of an array of the given type.
852 static int array_depth(const Type *type)
854 if (type->isPointerType())
855 return 1 + array_depth(type->getPointeeType().getTypePtr());
856 if (type->isArrayType()) {
857 const ArrayType *atype;
858 type = type->getCanonicalTypeInternal().getTypePtr();
859 atype = cast<ArrayType>(type);
860 return 1 + array_depth(atype->getElementType().getTypePtr());
862 return 0;
865 /* Return the element type of the given array type.
867 static QualType base_type(QualType qt)
869 const Type *type = qt.getTypePtr();
871 if (type->isPointerType())
872 return base_type(type->getPointeeType());
873 if (type->isArrayType()) {
874 const ArrayType *atype;
875 type = type->getCanonicalTypeInternal().getTypePtr();
876 atype = cast<ArrayType>(type);
877 return base_type(atype->getElementType());
879 return qt;
882 /* Extract an access relation from a reference to a variable.
883 * If the variable has name "A" and its type corresponds to an
884 * array of depth d, then the returned access relation is of the
885 * form
887 * { [] -> A[i_1,...,i_d] }
889 __isl_give isl_map *PetScan::extract_access(DeclRefExpr *expr)
891 return extract_access(expr->getDecl());
894 /* Extract an access relation from a variable.
895 * If the variable has name "A" and its type corresponds to an
896 * array of depth d, then the returned access relation is of the
897 * form
899 * { [] -> A[i_1,...,i_d] }
901 __isl_give isl_map *PetScan::extract_access(ValueDecl *decl)
903 int depth = array_depth(decl->getType().getTypePtr());
904 isl_id *id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
905 isl_space *dim = isl_space_alloc(ctx, 0, 0, depth);
906 isl_map *access_rel;
908 dim = isl_space_set_tuple_id(dim, isl_dim_out, id);
910 access_rel = isl_map_universe(dim);
912 return access_rel;
915 /* Extract an access relation from an integer contant.
916 * If the value of the constant is "v", then the returned access relation
917 * is
919 * { [] -> [v] }
921 __isl_give isl_map *PetScan::extract_access(IntegerLiteral *expr)
923 return isl_map_from_range(isl_set_from_pw_aff(extract_affine(expr)));
926 /* Try and extract an access relation from the given Expr.
927 * Return NULL if it doesn't work out.
929 __isl_give isl_map *PetScan::extract_access(Expr *expr)
931 switch (expr->getStmtClass()) {
932 case Stmt::ImplicitCastExprClass:
933 return extract_access(cast<ImplicitCastExpr>(expr));
934 case Stmt::DeclRefExprClass:
935 return extract_access(cast<DeclRefExpr>(expr));
936 case Stmt::ArraySubscriptExprClass:
937 return extract_access(cast<ArraySubscriptExpr>(expr));
938 case Stmt::IntegerLiteralClass:
939 return extract_access(cast<IntegerLiteral>(expr));
940 default:
941 unsupported(expr);
943 return NULL;
946 /* Assign the affine expression "index" to the output dimension "pos" of "map",
947 * restrict the domain to those values that result in a non-negative index
948 * and return the result.
950 __isl_give isl_map *set_index(__isl_take isl_map *map, int pos,
951 __isl_take isl_pw_aff *index)
953 isl_map *index_map;
954 int len = isl_map_dim(map, isl_dim_out);
955 isl_id *id;
956 isl_set *domain;
958 domain = isl_pw_aff_nonneg_set(isl_pw_aff_copy(index));
959 index = isl_pw_aff_intersect_domain(index, domain);
960 index_map = isl_map_from_range(isl_set_from_pw_aff(index));
961 index_map = isl_map_insert_dims(index_map, isl_dim_out, 0, pos);
962 index_map = isl_map_add_dims(index_map, isl_dim_out, len - pos - 1);
963 id = isl_map_get_tuple_id(map, isl_dim_out);
964 index_map = isl_map_set_tuple_id(index_map, isl_dim_out, id);
966 map = isl_map_intersect(map, index_map);
968 return map;
971 /* Extract an access relation from the given array subscript expression.
972 * If nesting is allowed in general, then we turn it on while
973 * examining the index expression.
975 * We first extract an access relation from the base.
976 * This will result in an access relation with a range that corresponds
977 * to the array being accessed and with earlier indices filled in already.
978 * We then extract the current index and fill that in as well.
979 * The position of the current index is based on the type of base.
980 * If base is the actual array variable, then the depth of this type
981 * will be the same as the depth of the array and we will fill in
982 * the first array index.
983 * Otherwise, the depth of the base type will be smaller and we will fill
984 * in a later index.
986 __isl_give isl_map *PetScan::extract_access(ArraySubscriptExpr *expr)
988 Expr *base = expr->getBase();
989 Expr *idx = expr->getIdx();
990 isl_pw_aff *index;
991 isl_map *base_access;
992 isl_map *access;
993 int depth = array_depth(base->getType().getTypePtr());
994 int pos;
995 bool save_nesting = nesting_enabled;
997 nesting_enabled = allow_nested;
999 base_access = extract_access(base);
1000 index = extract_affine(idx);
1002 nesting_enabled = save_nesting;
1004 pos = isl_map_dim(base_access, isl_dim_out) - depth;
1005 access = set_index(base_access, pos, index);
1007 return access;
1010 /* Check if "expr" calls function "minmax" with two arguments and if so
1011 * make lhs and rhs refer to these two arguments.
1013 static bool is_minmax(Expr *expr, const char *minmax, Expr *&lhs, Expr *&rhs)
1015 CallExpr *call;
1016 FunctionDecl *fd;
1017 string name;
1019 if (expr->getStmtClass() != Stmt::CallExprClass)
1020 return false;
1022 call = cast<CallExpr>(expr);
1023 fd = call->getDirectCallee();
1024 if (!fd)
1025 return false;
1027 if (call->getNumArgs() != 2)
1028 return false;
1030 name = fd->getDeclName().getAsString();
1031 if (name != minmax)
1032 return false;
1034 lhs = call->getArg(0);
1035 rhs = call->getArg(1);
1037 return true;
1040 /* Check if "expr" is of the form min(lhs, rhs) and if so make
1041 * lhs and rhs refer to the two arguments.
1043 static bool is_min(Expr *expr, Expr *&lhs, Expr *&rhs)
1045 return is_minmax(expr, "min", lhs, rhs);
1048 /* Check if "expr" is of the form max(lhs, rhs) and if so make
1049 * lhs and rhs refer to the two arguments.
1051 static bool is_max(Expr *expr, Expr *&lhs, Expr *&rhs)
1053 return is_minmax(expr, "max", lhs, rhs);
1056 /* Return "lhs && rhs", defined on the shared definition domain.
1058 static __isl_give isl_pw_aff *pw_aff_and(__isl_take isl_pw_aff *lhs,
1059 __isl_take isl_pw_aff *rhs)
1061 isl_set *cond;
1062 isl_set *dom;
1064 dom = isl_set_intersect(isl_pw_aff_domain(isl_pw_aff_copy(lhs)),
1065 isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1066 cond = isl_set_intersect(isl_pw_aff_non_zero_set(lhs),
1067 isl_pw_aff_non_zero_set(rhs));
1068 return indicator_function(cond, dom);
1071 /* Return "lhs && rhs", with shortcut semantics.
1072 * That is, if lhs is false, then the result is defined even if rhs is not.
1073 * In practice, we compute lhs ? rhs : lhs.
1075 static __isl_give isl_pw_aff *pw_aff_and_then(__isl_take isl_pw_aff *lhs,
1076 __isl_take isl_pw_aff *rhs)
1078 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), rhs, lhs);
1081 /* Return "lhs || rhs", with shortcut semantics.
1082 * That is, if lhs is true, then the result is defined even if rhs is not.
1083 * In practice, we compute lhs ? lhs : rhs.
1085 static __isl_give isl_pw_aff *pw_aff_or_else(__isl_take isl_pw_aff *lhs,
1086 __isl_take isl_pw_aff *rhs)
1088 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), lhs, rhs);
1091 /* Extract an affine expressions representing the comparison "LHS op RHS"
1092 * "comp" is the original statement that "LHS op RHS" is derived from
1093 * and is used for diagnostics.
1095 * If the comparison is of the form
1097 * a <= min(b,c)
1099 * then the expression is constructed as the conjunction of
1100 * the comparisons
1102 * a <= b and a <= c
1104 * A similar optimization is performed for max(a,b) <= c.
1105 * We do this because that will lead to simpler representations
1106 * of the expression.
1107 * If isl is ever enhanced to explicitly deal with min and max expressions,
1108 * this optimization can be removed.
1110 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperatorKind op,
1111 Expr *LHS, Expr *RHS, Stmt *comp)
1113 isl_pw_aff *lhs;
1114 isl_pw_aff *rhs;
1115 isl_pw_aff *res;
1116 isl_set *cond;
1117 isl_set *dom;
1119 if (op == BO_GT)
1120 return extract_comparison(BO_LT, RHS, LHS, comp);
1121 if (op == BO_GE)
1122 return extract_comparison(BO_LE, RHS, LHS, comp);
1124 if (op == BO_LT || op == BO_LE) {
1125 Expr *expr1, *expr2;
1126 if (is_min(RHS, expr1, expr2)) {
1127 lhs = extract_comparison(op, LHS, expr1, comp);
1128 rhs = extract_comparison(op, LHS, expr2, comp);
1129 return pw_aff_and(lhs, rhs);
1131 if (is_max(LHS, expr1, expr2)) {
1132 lhs = extract_comparison(op, expr1, RHS, comp);
1133 rhs = extract_comparison(op, expr2, RHS, comp);
1134 return pw_aff_and(lhs, rhs);
1138 lhs = extract_affine(LHS);
1139 rhs = extract_affine(RHS);
1141 dom = isl_pw_aff_domain(isl_pw_aff_copy(lhs));
1142 dom = isl_set_intersect(dom, isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1144 switch (op) {
1145 case BO_LT:
1146 cond = isl_pw_aff_lt_set(lhs, rhs);
1147 break;
1148 case BO_LE:
1149 cond = isl_pw_aff_le_set(lhs, rhs);
1150 break;
1151 case BO_EQ:
1152 cond = isl_pw_aff_eq_set(lhs, rhs);
1153 break;
1154 case BO_NE:
1155 cond = isl_pw_aff_ne_set(lhs, rhs);
1156 break;
1157 default:
1158 isl_pw_aff_free(lhs);
1159 isl_pw_aff_free(rhs);
1160 isl_set_free(dom);
1161 unsupported(comp);
1162 return NULL;
1165 cond = isl_set_coalesce(cond);
1166 res = indicator_function(cond, dom);
1168 return res;
1171 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperator *comp)
1173 return extract_comparison(comp->getOpcode(), comp->getLHS(),
1174 comp->getRHS(), comp);
1177 /* Extract an affine expression representing the negation (logical not)
1178 * of a subexpression.
1180 __isl_give isl_pw_aff *PetScan::extract_boolean(UnaryOperator *op)
1182 isl_set *set_cond, *dom;
1183 isl_pw_aff *cond, *res;
1185 cond = extract_condition(op->getSubExpr());
1187 dom = isl_pw_aff_domain(isl_pw_aff_copy(cond));
1189 set_cond = isl_pw_aff_zero_set(cond);
1191 res = indicator_function(set_cond, dom);
1193 return res;
1196 /* Extract an affine expression representing the disjunction (logical or)
1197 * or conjunction (logical and) of two subexpressions.
1199 __isl_give isl_pw_aff *PetScan::extract_boolean(BinaryOperator *comp)
1201 isl_pw_aff *lhs, *rhs;
1203 lhs = extract_condition(comp->getLHS());
1204 rhs = extract_condition(comp->getRHS());
1206 switch (comp->getOpcode()) {
1207 case BO_LAnd:
1208 return pw_aff_and_then(lhs, rhs);
1209 case BO_LOr:
1210 return pw_aff_or_else(lhs, rhs);
1211 default:
1212 isl_pw_aff_free(lhs);
1213 isl_pw_aff_free(rhs);
1216 unsupported(comp);
1217 return NULL;
1220 __isl_give isl_pw_aff *PetScan::extract_condition(UnaryOperator *expr)
1222 switch (expr->getOpcode()) {
1223 case UO_LNot:
1224 return extract_boolean(expr);
1225 default:
1226 unsupported(expr);
1227 return NULL;
1231 /* Extract the affine expression "expr != 0 ? 1 : 0".
1233 __isl_give isl_pw_aff *PetScan::extract_implicit_condition(Expr *expr)
1235 isl_pw_aff *res;
1236 isl_set *set, *dom;
1238 res = extract_affine(expr);
1240 dom = isl_pw_aff_domain(isl_pw_aff_copy(res));
1241 set = isl_pw_aff_non_zero_set(res);
1243 res = indicator_function(set, dom);
1245 return res;
1248 /* Extract an affine expression from a boolean expression.
1249 * In particular, return the expression "expr ? 1 : 0".
1251 * If the expression doesn't look like a condition, we assume it
1252 * is an affine expression and return the condition "expr != 0 ? 1 : 0".
1254 __isl_give isl_pw_aff *PetScan::extract_condition(Expr *expr)
1256 BinaryOperator *comp;
1258 if (!expr) {
1259 isl_set *u = isl_set_universe(isl_space_params_alloc(ctx, 0));
1260 return indicator_function(u, isl_set_copy(u));
1263 if (expr->getStmtClass() == Stmt::ParenExprClass)
1264 return extract_condition(cast<ParenExpr>(expr)->getSubExpr());
1266 if (expr->getStmtClass() == Stmt::UnaryOperatorClass)
1267 return extract_condition(cast<UnaryOperator>(expr));
1269 if (expr->getStmtClass() != Stmt::BinaryOperatorClass)
1270 return extract_implicit_condition(expr);
1272 comp = cast<BinaryOperator>(expr);
1273 switch (comp->getOpcode()) {
1274 case BO_LT:
1275 case BO_LE:
1276 case BO_GT:
1277 case BO_GE:
1278 case BO_EQ:
1279 case BO_NE:
1280 return extract_comparison(comp);
1281 case BO_LAnd:
1282 case BO_LOr:
1283 return extract_boolean(comp);
1284 default:
1285 return extract_implicit_condition(expr);
1289 static enum pet_op_type UnaryOperatorKind2pet_op_type(UnaryOperatorKind kind)
1291 switch (kind) {
1292 case UO_Minus:
1293 return pet_op_minus;
1294 case UO_PostInc:
1295 return pet_op_post_inc;
1296 case UO_PostDec:
1297 return pet_op_post_dec;
1298 case UO_PreInc:
1299 return pet_op_pre_inc;
1300 case UO_PreDec:
1301 return pet_op_pre_dec;
1302 default:
1303 return pet_op_last;
1307 static enum pet_op_type BinaryOperatorKind2pet_op_type(BinaryOperatorKind kind)
1309 switch (kind) {
1310 case BO_AddAssign:
1311 return pet_op_add_assign;
1312 case BO_SubAssign:
1313 return pet_op_sub_assign;
1314 case BO_MulAssign:
1315 return pet_op_mul_assign;
1316 case BO_DivAssign:
1317 return pet_op_div_assign;
1318 case BO_Assign:
1319 return pet_op_assign;
1320 case BO_Add:
1321 return pet_op_add;
1322 case BO_Sub:
1323 return pet_op_sub;
1324 case BO_Mul:
1325 return pet_op_mul;
1326 case BO_Div:
1327 return pet_op_div;
1328 case BO_Rem:
1329 return pet_op_mod;
1330 case BO_EQ:
1331 return pet_op_eq;
1332 case BO_LE:
1333 return pet_op_le;
1334 case BO_LT:
1335 return pet_op_lt;
1336 case BO_GT:
1337 return pet_op_gt;
1338 default:
1339 return pet_op_last;
1343 /* Construct a pet_expr representing a unary operator expression.
1345 struct pet_expr *PetScan::extract_expr(UnaryOperator *expr)
1347 struct pet_expr *arg;
1348 enum pet_op_type op;
1350 op = UnaryOperatorKind2pet_op_type(expr->getOpcode());
1351 if (op == pet_op_last) {
1352 unsupported(expr);
1353 return NULL;
1356 arg = extract_expr(expr->getSubExpr());
1358 if (expr->isIncrementDecrementOp() &&
1359 arg && arg->type == pet_expr_access) {
1360 mark_write(arg);
1361 arg->acc.read = 1;
1364 return pet_expr_new_unary(ctx, op, arg);
1367 /* Mark the given access pet_expr as a write.
1368 * If a scalar is being accessed, then mark its value
1369 * as unknown in assigned_value.
1371 void PetScan::mark_write(struct pet_expr *access)
1373 isl_id *id;
1374 ValueDecl *decl;
1376 if (!access)
1377 return;
1379 access->acc.write = 1;
1380 access->acc.read = 0;
1382 if (isl_map_dim(access->acc.access, isl_dim_out) != 0)
1383 return;
1385 id = isl_map_get_tuple_id(access->acc.access, isl_dim_out);
1386 decl = (ValueDecl *) isl_id_get_user(id);
1387 clear_assignment(assigned_value, decl);
1388 isl_id_free(id);
1391 /* Assign "rhs" to "lhs".
1393 * In particular, if "lhs" is a scalar variable, then mark
1394 * the variable as having been assigned. If, furthermore, "rhs"
1395 * is an affine expression, then keep track of this value in assigned_value
1396 * so that we can plug it in when we later come across the same variable.
1398 void PetScan::assign(struct pet_expr *lhs, Expr *rhs)
1400 isl_id *id;
1401 ValueDecl *decl;
1402 isl_pw_aff *pa;
1404 if (!lhs)
1405 return;
1406 if (lhs->type != pet_expr_access)
1407 return;
1408 if (isl_map_dim(lhs->acc.access, isl_dim_out) != 0)
1409 return;
1411 id = isl_map_get_tuple_id(lhs->acc.access, isl_dim_out);
1412 decl = (ValueDecl *) isl_id_get_user(id);
1413 isl_id_free(id);
1415 pa = try_extract_affine(rhs);
1416 clear_assignment(assigned_value, decl);
1417 if (!pa)
1418 return;
1419 assigned_value[decl] = pa;
1420 insert_expression(pa);
1423 /* Construct a pet_expr representing a binary operator expression.
1425 * If the top level operator is an assignment and the LHS is an access,
1426 * then we mark that access as a write. If the operator is a compound
1427 * assignment, the access is marked as both a read and a write.
1429 * If "expr" assigns something to a scalar variable, then we mark
1430 * the variable as having been assigned. If, furthermore, the expression
1431 * is affine, then keep track of this value in assigned_value
1432 * so that we can plug it in when we later come across the same variable.
1434 struct pet_expr *PetScan::extract_expr(BinaryOperator *expr)
1436 struct pet_expr *lhs, *rhs;
1437 enum pet_op_type op;
1439 op = BinaryOperatorKind2pet_op_type(expr->getOpcode());
1440 if (op == pet_op_last) {
1441 unsupported(expr);
1442 return NULL;
1445 lhs = extract_expr(expr->getLHS());
1446 rhs = extract_expr(expr->getRHS());
1448 if (expr->isAssignmentOp() && lhs && lhs->type == pet_expr_access) {
1449 mark_write(lhs);
1450 if (expr->isCompoundAssignmentOp())
1451 lhs->acc.read = 1;
1454 if (expr->getOpcode() == BO_Assign)
1455 assign(lhs, expr->getRHS());
1457 return pet_expr_new_binary(ctx, op, lhs, rhs);
1460 /* Construct a pet_scop with a single statement killing the entire
1461 * array "array".
1463 struct pet_scop *PetScan::kill(Stmt *stmt, struct pet_array *array)
1465 isl_map *access;
1466 struct pet_expr *expr;
1468 if (!array)
1469 return NULL;
1470 access = isl_map_from_range(isl_set_copy(array->extent));
1471 expr = pet_expr_kill_from_access(access);
1472 return extract(stmt, expr);
1475 /* Construct a pet_scop for a (single) variable declaration.
1477 * The scop contains the variable being declared (as an array)
1478 * and a statement killing the array.
1480 * If the variable is initialized in the AST, then the scop
1481 * also contains an assignment to the variable.
1483 struct pet_scop *PetScan::extract(DeclStmt *stmt)
1485 Decl *decl;
1486 VarDecl *vd;
1487 struct pet_expr *lhs, *rhs, *pe;
1488 struct pet_scop *scop_decl, *scop;
1489 struct pet_array *array;
1491 if (!stmt->isSingleDecl()) {
1492 unsupported(stmt);
1493 return NULL;
1496 decl = stmt->getSingleDecl();
1497 vd = cast<VarDecl>(decl);
1499 array = extract_array(ctx, vd);
1500 if (array)
1501 array->declared = 1;
1502 scop_decl = kill(stmt, array);
1503 scop_decl = pet_scop_add_array(scop_decl, array);
1505 if (!vd->getInit())
1506 return scop_decl;
1508 lhs = pet_expr_from_access(extract_access(vd));
1509 rhs = extract_expr(vd->getInit());
1511 mark_write(lhs);
1512 assign(lhs, vd->getInit());
1514 pe = pet_expr_new_binary(ctx, pet_op_assign, lhs, rhs);
1515 scop = extract(stmt, pe);
1517 scop_decl = pet_scop_prefix(scop_decl, 0);
1518 scop = pet_scop_prefix(scop, 1);
1520 scop = pet_scop_add_seq(ctx, scop_decl, scop);
1522 return scop;
1525 /* Construct a pet_expr representing a conditional operation.
1527 * We first try to extract the condition as an affine expression.
1528 * If that fails, we construct a pet_expr tree representing the condition.
1530 struct pet_expr *PetScan::extract_expr(ConditionalOperator *expr)
1532 struct pet_expr *cond, *lhs, *rhs;
1533 isl_pw_aff *pa;
1535 pa = try_extract_affine(expr->getCond());
1536 if (pa) {
1537 isl_set *test = isl_set_from_pw_aff(pa);
1538 cond = pet_expr_from_access(isl_map_from_range(test));
1539 } else
1540 cond = extract_expr(expr->getCond());
1541 lhs = extract_expr(expr->getTrueExpr());
1542 rhs = extract_expr(expr->getFalseExpr());
1544 return pet_expr_new_ternary(ctx, cond, lhs, rhs);
1547 struct pet_expr *PetScan::extract_expr(ImplicitCastExpr *expr)
1549 return extract_expr(expr->getSubExpr());
1552 /* Construct a pet_expr representing a floating point value.
1554 * If the floating point literal does not appear in a macro,
1555 * then we use the original representation in the source code
1556 * as the string representation. Otherwise, we use the pretty
1557 * printer to produce a string representation.
1559 struct pet_expr *PetScan::extract_expr(FloatingLiteral *expr)
1561 double d;
1562 string s;
1563 const LangOptions &LO = PP.getLangOpts();
1564 SourceLocation loc = expr->getLocation();
1566 if (!loc.isMacroID()) {
1567 SourceManager &SM = PP.getSourceManager();
1568 unsigned len = Lexer::MeasureTokenLength(loc, SM, LO);
1569 s = string(SM.getCharacterData(loc), len);
1570 } else {
1571 llvm::raw_string_ostream S(s);
1572 expr->printPretty(S, 0, PrintingPolicy(LO));
1573 S.str();
1575 d = expr->getValueAsApproximateDouble();
1576 return pet_expr_new_double(ctx, d, s.c_str());
1579 /* Extract an access relation from "expr" and then convert it into
1580 * a pet_expr.
1582 struct pet_expr *PetScan::extract_access_expr(Expr *expr)
1584 isl_map *access;
1585 struct pet_expr *pe;
1587 access = extract_access(expr);
1589 pe = pet_expr_from_access(access);
1591 return pe;
1594 struct pet_expr *PetScan::extract_expr(ParenExpr *expr)
1596 return extract_expr(expr->getSubExpr());
1599 /* Construct a pet_expr representing a function call.
1601 * If we are passing along a pointer to an array element
1602 * or an entire row or even higher dimensional slice of an array,
1603 * then the function being called may write into the array.
1605 * We assume here that if the function is declared to take a pointer
1606 * to a const type, then the function will perform a read
1607 * and that otherwise, it will perform a write.
1609 struct pet_expr *PetScan::extract_expr(CallExpr *expr)
1611 struct pet_expr *res = NULL;
1612 FunctionDecl *fd;
1613 string name;
1615 fd = expr->getDirectCallee();
1616 if (!fd) {
1617 unsupported(expr);
1618 return NULL;
1621 name = fd->getDeclName().getAsString();
1622 res = pet_expr_new_call(ctx, name.c_str(), expr->getNumArgs());
1623 if (!res)
1624 return NULL;
1626 for (int i = 0; i < expr->getNumArgs(); ++i) {
1627 Expr *arg = expr->getArg(i);
1628 int is_addr = 0;
1629 pet_expr *main_arg;
1631 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
1632 ImplicitCastExpr *ice = cast<ImplicitCastExpr>(arg);
1633 arg = ice->getSubExpr();
1635 if (arg->getStmtClass() == Stmt::UnaryOperatorClass) {
1636 UnaryOperator *op = cast<UnaryOperator>(arg);
1637 if (op->getOpcode() == UO_AddrOf) {
1638 is_addr = 1;
1639 arg = op->getSubExpr();
1642 res->args[i] = PetScan::extract_expr(arg);
1643 main_arg = res->args[i];
1644 if (is_addr)
1645 res->args[i] = pet_expr_new_unary(ctx,
1646 pet_op_address_of, res->args[i]);
1647 if (!res->args[i])
1648 goto error;
1649 if (arg->getStmtClass() == Stmt::ArraySubscriptExprClass &&
1650 array_depth(arg->getType().getTypePtr()) > 0)
1651 is_addr = 1;
1652 if (is_addr && main_arg->type == pet_expr_access) {
1653 ParmVarDecl *parm;
1654 if (!fd->hasPrototype()) {
1655 unsupported(expr, "prototype required");
1656 goto error;
1658 parm = fd->getParamDecl(i);
1659 if (!const_base(parm->getType()))
1660 mark_write(main_arg);
1664 return res;
1665 error:
1666 pet_expr_free(res);
1667 return NULL;
1670 /* Construct a pet_expr representing a (C style) cast.
1672 struct pet_expr *PetScan::extract_expr(CStyleCastExpr *expr)
1674 struct pet_expr *arg;
1675 QualType type;
1677 arg = extract_expr(expr->getSubExpr());
1678 if (!arg)
1679 return NULL;
1681 type = expr->getTypeAsWritten();
1682 return pet_expr_new_cast(ctx, type.getAsString().c_str(), arg);
1685 /* Try and onstruct a pet_expr representing "expr".
1687 struct pet_expr *PetScan::extract_expr(Expr *expr)
1689 switch (expr->getStmtClass()) {
1690 case Stmt::UnaryOperatorClass:
1691 return extract_expr(cast<UnaryOperator>(expr));
1692 case Stmt::CompoundAssignOperatorClass:
1693 case Stmt::BinaryOperatorClass:
1694 return extract_expr(cast<BinaryOperator>(expr));
1695 case Stmt::ImplicitCastExprClass:
1696 return extract_expr(cast<ImplicitCastExpr>(expr));
1697 case Stmt::ArraySubscriptExprClass:
1698 case Stmt::DeclRefExprClass:
1699 case Stmt::IntegerLiteralClass:
1700 return extract_access_expr(expr);
1701 case Stmt::FloatingLiteralClass:
1702 return extract_expr(cast<FloatingLiteral>(expr));
1703 case Stmt::ParenExprClass:
1704 return extract_expr(cast<ParenExpr>(expr));
1705 case Stmt::ConditionalOperatorClass:
1706 return extract_expr(cast<ConditionalOperator>(expr));
1707 case Stmt::CallExprClass:
1708 return extract_expr(cast<CallExpr>(expr));
1709 case Stmt::CStyleCastExprClass:
1710 return extract_expr(cast<CStyleCastExpr>(expr));
1711 default:
1712 unsupported(expr);
1714 return NULL;
1717 /* Check if the given initialization statement is an assignment.
1718 * If so, return that assignment. Otherwise return NULL.
1720 BinaryOperator *PetScan::initialization_assignment(Stmt *init)
1722 BinaryOperator *ass;
1724 if (init->getStmtClass() != Stmt::BinaryOperatorClass)
1725 return NULL;
1727 ass = cast<BinaryOperator>(init);
1728 if (ass->getOpcode() != BO_Assign)
1729 return NULL;
1731 return ass;
1734 /* Check if the given initialization statement is a declaration
1735 * of a single variable.
1736 * If so, return that declaration. Otherwise return NULL.
1738 Decl *PetScan::initialization_declaration(Stmt *init)
1740 DeclStmt *decl;
1742 if (init->getStmtClass() != Stmt::DeclStmtClass)
1743 return NULL;
1745 decl = cast<DeclStmt>(init);
1747 if (!decl->isSingleDecl())
1748 return NULL;
1750 return decl->getSingleDecl();
1753 /* Given the assignment operator in the initialization of a for loop,
1754 * extract the induction variable, i.e., the (integer)variable being
1755 * assigned.
1757 ValueDecl *PetScan::extract_induction_variable(BinaryOperator *init)
1759 Expr *lhs;
1760 DeclRefExpr *ref;
1761 ValueDecl *decl;
1762 const Type *type;
1764 lhs = init->getLHS();
1765 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1766 unsupported(init);
1767 return NULL;
1770 ref = cast<DeclRefExpr>(lhs);
1771 decl = ref->getDecl();
1772 type = decl->getType().getTypePtr();
1774 if (!type->isIntegerType()) {
1775 unsupported(lhs);
1776 return NULL;
1779 return decl;
1782 /* Given the initialization statement of a for loop and the single
1783 * declaration in this initialization statement,
1784 * extract the induction variable, i.e., the (integer) variable being
1785 * declared.
1787 VarDecl *PetScan::extract_induction_variable(Stmt *init, Decl *decl)
1789 VarDecl *vd;
1791 vd = cast<VarDecl>(decl);
1793 const QualType type = vd->getType();
1794 if (!type->isIntegerType()) {
1795 unsupported(init);
1796 return NULL;
1799 if (!vd->getInit()) {
1800 unsupported(init);
1801 return NULL;
1804 return vd;
1807 /* Check that op is of the form iv++ or iv--.
1808 * Return an affine expression "1" or "-1" accordingly.
1810 __isl_give isl_pw_aff *PetScan::extract_unary_increment(
1811 clang::UnaryOperator *op, clang::ValueDecl *iv)
1813 Expr *sub;
1814 DeclRefExpr *ref;
1815 isl_space *space;
1816 isl_aff *aff;
1818 if (!op->isIncrementDecrementOp()) {
1819 unsupported(op);
1820 return NULL;
1823 sub = op->getSubExpr();
1824 if (sub->getStmtClass() != Stmt::DeclRefExprClass) {
1825 unsupported(op);
1826 return NULL;
1829 ref = cast<DeclRefExpr>(sub);
1830 if (ref->getDecl() != iv) {
1831 unsupported(op);
1832 return NULL;
1835 space = isl_space_params_alloc(ctx, 0);
1836 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
1838 if (op->isIncrementOp())
1839 aff = isl_aff_add_constant_si(aff, 1);
1840 else
1841 aff = isl_aff_add_constant_si(aff, -1);
1843 return isl_pw_aff_from_aff(aff);
1846 /* If the isl_pw_aff on which isl_pw_aff_foreach_piece is called
1847 * has a single constant expression, then put this constant in *user.
1848 * The caller is assumed to have checked that this function will
1849 * be called exactly once.
1851 static int extract_cst(__isl_take isl_set *set, __isl_take isl_aff *aff,
1852 void *user)
1854 isl_int *inc = (isl_int *)user;
1855 int res = 0;
1857 if (isl_aff_is_cst(aff))
1858 isl_aff_get_constant(aff, inc);
1859 else
1860 res = -1;
1862 isl_set_free(set);
1863 isl_aff_free(aff);
1865 return res;
1868 /* Check if op is of the form
1870 * iv = iv + inc
1872 * and return inc as an affine expression.
1874 * We extract an affine expression from the RHS, subtract iv and return
1875 * the result.
1877 __isl_give isl_pw_aff *PetScan::extract_binary_increment(BinaryOperator *op,
1878 clang::ValueDecl *iv)
1880 Expr *lhs;
1881 DeclRefExpr *ref;
1882 isl_id *id;
1883 isl_space *dim;
1884 isl_aff *aff;
1885 isl_pw_aff *val;
1887 if (op->getOpcode() != BO_Assign) {
1888 unsupported(op);
1889 return NULL;
1892 lhs = op->getLHS();
1893 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1894 unsupported(op);
1895 return NULL;
1898 ref = cast<DeclRefExpr>(lhs);
1899 if (ref->getDecl() != iv) {
1900 unsupported(op);
1901 return NULL;
1904 val = extract_affine(op->getRHS());
1906 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
1908 dim = isl_space_params_alloc(ctx, 1);
1909 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
1910 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
1911 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
1913 val = isl_pw_aff_sub(val, isl_pw_aff_from_aff(aff));
1915 return val;
1918 /* Check that op is of the form iv += cst or iv -= cst
1919 * and return an affine expression corresponding oto cst or -cst accordingly.
1921 __isl_give isl_pw_aff *PetScan::extract_compound_increment(
1922 CompoundAssignOperator *op, clang::ValueDecl *iv)
1924 Expr *lhs;
1925 DeclRefExpr *ref;
1926 bool neg = false;
1927 isl_pw_aff *val;
1928 BinaryOperatorKind opcode;
1930 opcode = op->getOpcode();
1931 if (opcode != BO_AddAssign && opcode != BO_SubAssign) {
1932 unsupported(op);
1933 return NULL;
1935 if (opcode == BO_SubAssign)
1936 neg = true;
1938 lhs = op->getLHS();
1939 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1940 unsupported(op);
1941 return NULL;
1944 ref = cast<DeclRefExpr>(lhs);
1945 if (ref->getDecl() != iv) {
1946 unsupported(op);
1947 return NULL;
1950 val = extract_affine(op->getRHS());
1951 if (neg)
1952 val = isl_pw_aff_neg(val);
1954 return val;
1957 /* Check that the increment of the given for loop increments
1958 * (or decrements) the induction variable "iv" and return
1959 * the increment as an affine expression if successful.
1961 __isl_give isl_pw_aff *PetScan::extract_increment(clang::ForStmt *stmt,
1962 ValueDecl *iv)
1964 Stmt *inc = stmt->getInc();
1966 if (!inc) {
1967 unsupported(stmt);
1968 return NULL;
1971 if (inc->getStmtClass() == Stmt::UnaryOperatorClass)
1972 return extract_unary_increment(cast<UnaryOperator>(inc), iv);
1973 if (inc->getStmtClass() == Stmt::CompoundAssignOperatorClass)
1974 return extract_compound_increment(
1975 cast<CompoundAssignOperator>(inc), iv);
1976 if (inc->getStmtClass() == Stmt::BinaryOperatorClass)
1977 return extract_binary_increment(cast<BinaryOperator>(inc), iv);
1979 unsupported(inc);
1980 return NULL;
1983 /* Embed the given iteration domain in an extra outer loop
1984 * with induction variable "var".
1985 * If this variable appeared as a parameter in the constraints,
1986 * it is replaced by the new outermost dimension.
1988 static __isl_give isl_set *embed(__isl_take isl_set *set,
1989 __isl_take isl_id *var)
1991 int pos;
1993 set = isl_set_insert_dims(set, isl_dim_set, 0, 1);
1994 pos = isl_set_find_dim_by_id(set, isl_dim_param, var);
1995 if (pos >= 0) {
1996 set = isl_set_equate(set, isl_dim_param, pos, isl_dim_set, 0);
1997 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2000 isl_id_free(var);
2001 return set;
2004 /* Return those elements in the space of "cond" that come after
2005 * (based on "sign") an element in "cond".
2007 static __isl_give isl_set *after(__isl_take isl_set *cond, int sign)
2009 isl_map *previous_to_this;
2011 if (sign > 0)
2012 previous_to_this = isl_map_lex_lt(isl_set_get_space(cond));
2013 else
2014 previous_to_this = isl_map_lex_gt(isl_set_get_space(cond));
2016 cond = isl_set_apply(cond, previous_to_this);
2018 return cond;
2021 /* Create the infinite iteration domain
2023 * { [id] : id >= 0 }
2025 * If "scop" has an affine skip of type pet_skip_later,
2026 * then remove those iterations i that have an earlier iteration
2027 * where the skip condition is satisfied, meaning that iteration i
2028 * is not executed.
2029 * Since we are dealing with a loop without loop iterator,
2030 * the skip condition cannot refer to the current loop iterator and
2031 * so effectively, the returned set is of the form
2033 * { [0]; [id] : id >= 1 and not skip }
2035 static __isl_give isl_set *infinite_domain(__isl_take isl_id *id,
2036 struct pet_scop *scop)
2038 isl_ctx *ctx = isl_id_get_ctx(id);
2039 isl_set *domain;
2040 isl_set *skip;
2042 domain = isl_set_nat_universe(isl_space_set_alloc(ctx, 0, 1));
2043 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, id);
2045 if (!pet_scop_has_affine_skip(scop, pet_skip_later))
2046 return domain;
2048 skip = pet_scop_get_skip(scop, pet_skip_later);
2049 skip = isl_set_fix_si(skip, isl_dim_set, 0, 1);
2050 skip = isl_set_params(skip);
2051 skip = embed(skip, isl_id_copy(id));
2052 skip = isl_set_intersect(skip , isl_set_copy(domain));
2053 domain = isl_set_subtract(domain, after(skip, 1));
2055 return domain;
2058 /* Create an identity mapping on the space containing "domain".
2060 static __isl_give isl_map *identity_map(__isl_keep isl_set *domain)
2062 isl_space *space;
2063 isl_map *id;
2065 space = isl_space_map_from_set(isl_set_get_space(domain));
2066 id = isl_map_identity(space);
2068 return id;
2071 /* Add a filter to "scop" that imposes that it is only executed
2072 * when "break_access" has a zero value for all previous iterations
2073 * of "domain".
2075 * The input "break_access" has a zero-dimensional domain and range.
2077 static struct pet_scop *scop_add_break(struct pet_scop *scop,
2078 __isl_take isl_map *break_access, __isl_take isl_set *domain, int sign)
2080 isl_ctx *ctx = isl_set_get_ctx(domain);
2081 isl_id *id_test;
2082 isl_map *prev;
2084 id_test = isl_map_get_tuple_id(break_access, isl_dim_out);
2085 break_access = isl_map_add_dims(break_access, isl_dim_in, 1);
2086 break_access = isl_map_add_dims(break_access, isl_dim_out, 1);
2087 break_access = isl_map_intersect_range(break_access, domain);
2088 break_access = isl_map_set_tuple_id(break_access, isl_dim_out, id_test);
2089 if (sign > 0)
2090 prev = isl_map_lex_gt_first(isl_map_get_space(break_access), 1);
2091 else
2092 prev = isl_map_lex_lt_first(isl_map_get_space(break_access), 1);
2093 break_access = isl_map_intersect(break_access, prev);
2094 scop = pet_scop_filter(scop, break_access, 0);
2095 scop = pet_scop_merge_filters(scop);
2097 return scop;
2100 /* Construct a pet_scop for an infinite loop around the given body.
2102 * We extract a pet_scop for the body and then embed it in a loop with
2103 * iteration domain
2105 * { [t] : t >= 0 }
2107 * and schedule
2109 * { [t] -> [t] }
2111 * If the body contains any break, then it is taken into
2112 * account in infinite_domain (if the skip condition is affine)
2113 * or in scop_add_break (if the skip condition is not affine).
2115 struct pet_scop *PetScan::extract_infinite_loop(Stmt *body)
2117 isl_id *id;
2118 isl_set *domain;
2119 isl_map *ident;
2120 isl_map *access;
2121 struct pet_scop *scop;
2122 bool has_var_break;
2124 scop = extract(body);
2125 if (!scop)
2126 return NULL;
2128 id = isl_id_alloc(ctx, "t", NULL);
2129 domain = infinite_domain(isl_id_copy(id), scop);
2130 ident = identity_map(domain);
2132 has_var_break = pet_scop_has_var_skip(scop, pet_skip_later);
2133 if (has_var_break)
2134 access = pet_scop_get_skip_map(scop, pet_skip_later);
2136 scop = pet_scop_embed(scop, isl_set_copy(domain),
2137 isl_map_copy(ident), ident, id);
2138 if (has_var_break)
2139 scop = scop_add_break(scop, access, domain, 1);
2140 else
2141 isl_set_free(domain);
2143 return scop;
2146 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
2148 * for (;;)
2149 * body
2152 struct pet_scop *PetScan::extract_infinite_for(ForStmt *stmt)
2154 return extract_infinite_loop(stmt->getBody());
2157 /* Create an access to a virtual array representing the result
2158 * of a condition.
2159 * Unlike other accessed data, the id of the array is NULL as
2160 * there is no ValueDecl in the program corresponding to the virtual
2161 * array.
2162 * The array starts out as a scalar, but grows along with the
2163 * statement writing to the array in pet_scop_embed.
2165 static __isl_give isl_map *create_test_access(isl_ctx *ctx, int test_nr)
2167 isl_space *dim = isl_space_alloc(ctx, 0, 0, 0);
2168 isl_id *id;
2169 char name[50];
2171 snprintf(name, sizeof(name), "__pet_test_%d", test_nr);
2172 id = isl_id_alloc(ctx, name, NULL);
2173 dim = isl_space_set_tuple_id(dim, isl_dim_out, id);
2174 return isl_map_universe(dim);
2177 /* Add an array with the given extent ("access") to the list
2178 * of arrays in "scop" and return the extended pet_scop.
2179 * The array is marked as attaining values 0 and 1 only and
2180 * as each element being assigned at most once.
2182 static struct pet_scop *scop_add_array(struct pet_scop *scop,
2183 __isl_keep isl_map *access, clang::ASTContext &ast_ctx)
2185 isl_ctx *ctx = isl_map_get_ctx(access);
2186 isl_space *dim;
2187 struct pet_array *array;
2189 if (!scop)
2190 return NULL;
2191 if (!ctx)
2192 goto error;
2194 array = isl_calloc_type(ctx, struct pet_array);
2195 if (!array)
2196 goto error;
2198 array->extent = isl_map_range(isl_map_copy(access));
2199 dim = isl_space_params_alloc(ctx, 0);
2200 array->context = isl_set_universe(dim);
2201 dim = isl_space_set_alloc(ctx, 0, 1);
2202 array->value_bounds = isl_set_universe(dim);
2203 array->value_bounds = isl_set_lower_bound_si(array->value_bounds,
2204 isl_dim_set, 0, 0);
2205 array->value_bounds = isl_set_upper_bound_si(array->value_bounds,
2206 isl_dim_set, 0, 1);
2207 array->element_type = strdup("int");
2208 array->element_size = ast_ctx.getTypeInfo(ast_ctx.IntTy).first / 8;
2209 array->uniquely_defined = 1;
2211 if (!array->extent || !array->context)
2212 array = pet_array_free(array);
2214 scop = pet_scop_add_array(scop, array);
2216 return scop;
2217 error:
2218 pet_scop_free(scop);
2219 return NULL;
2222 /* Construct a pet_scop for a while loop of the form
2224 * while (pa)
2225 * body
2227 * In particular, construct a scop for an infinite loop around body and
2228 * intersect the domain with the affine expression.
2229 * Note that this intersection may result in an empty loop.
2231 struct pet_scop *PetScan::extract_affine_while(__isl_take isl_pw_aff *pa,
2232 Stmt *body)
2234 struct pet_scop *scop;
2235 isl_set *dom;
2236 isl_set *valid;
2238 valid = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2239 dom = isl_pw_aff_non_zero_set(pa);
2240 scop = extract_infinite_loop(body);
2241 scop = pet_scop_restrict(scop, dom);
2242 scop = pet_scop_restrict_context(scop, valid);
2244 return scop;
2247 /* Construct a scop for a while, given the scops for the condition
2248 * and the body, the filter access and the iteration domain of
2249 * the while loop.
2251 * In particular, the scop for the condition is filtered to depend
2252 * on "test_access" evaluating to true for all previous iterations
2253 * of the loop, while the scop for the body is filtered to depend
2254 * on "test_access" evaluating to true for all iterations up to the
2255 * current iteration.
2257 * These filtered scops are then combined into a single scop.
2259 * "sign" is positive if the iterator increases and negative
2260 * if it decreases.
2262 static struct pet_scop *scop_add_while(struct pet_scop *scop_cond,
2263 struct pet_scop *scop_body, __isl_take isl_map *test_access,
2264 __isl_take isl_set *domain, int sign)
2266 isl_ctx *ctx = isl_set_get_ctx(domain);
2267 isl_id *id_test;
2268 isl_map *prev;
2270 id_test = isl_map_get_tuple_id(test_access, isl_dim_out);
2271 test_access = isl_map_add_dims(test_access, isl_dim_in, 1);
2272 test_access = isl_map_add_dims(test_access, isl_dim_out, 1);
2273 test_access = isl_map_intersect_range(test_access, domain);
2274 test_access = isl_map_set_tuple_id(test_access, isl_dim_out, id_test);
2275 if (sign > 0)
2276 prev = isl_map_lex_ge_first(isl_map_get_space(test_access), 1);
2277 else
2278 prev = isl_map_lex_le_first(isl_map_get_space(test_access), 1);
2279 test_access = isl_map_intersect(test_access, prev);
2280 scop_body = pet_scop_filter(scop_body, isl_map_copy(test_access), 1);
2281 if (sign > 0)
2282 prev = isl_map_lex_gt_first(isl_map_get_space(test_access), 1);
2283 else
2284 prev = isl_map_lex_lt_first(isl_map_get_space(test_access), 1);
2285 test_access = isl_map_intersect(test_access, prev);
2286 scop_cond = pet_scop_filter(scop_cond, test_access, 1);
2288 return pet_scop_add_seq(ctx, scop_cond, scop_body);
2291 /* Check if the while loop is of the form
2293 * while (affine expression)
2294 * body
2296 * If so, call extract_affine_while to construct a scop.
2298 * Otherwise, construct a generic while scop, with iteration domain
2299 * { [t] : t >= 0 }. The scop consists of two parts, one for
2300 * evaluating the condition and one for the body.
2301 * The schedule is adjusted to reflect that the condition is evaluated
2302 * before the body is executed and the body is filtered to depend
2303 * on the result of the condition evaluating to true on all iterations
2304 * up to the current iteration, while the evaluation the condition itself
2305 * is filtered to depend on the result of the condition evaluating to true
2306 * on all previous iterations.
2307 * The context of the scop representing the body is dropped
2308 * because we don't know how many times the body will be executed,
2309 * if at all.
2311 * If the body contains any break, then it is taken into
2312 * account in infinite_domain (if the skip condition is affine)
2313 * or in scop_add_break (if the skip condition is not affine).
2315 struct pet_scop *PetScan::extract(WhileStmt *stmt)
2317 Expr *cond;
2318 isl_id *id;
2319 isl_map *test_access;
2320 isl_set *domain;
2321 isl_map *ident;
2322 isl_pw_aff *pa;
2323 struct pet_scop *scop, *scop_body;
2324 bool has_var_break;
2325 isl_map *break_access;
2327 cond = stmt->getCond();
2328 if (!cond) {
2329 unsupported(stmt);
2330 return NULL;
2333 clear_assignments clear(assigned_value);
2334 clear.TraverseStmt(stmt->getBody());
2336 pa = try_extract_affine_condition(cond);
2337 if (pa)
2338 return extract_affine_while(pa, stmt->getBody());
2340 if (!allow_nested) {
2341 unsupported(stmt);
2342 return NULL;
2345 test_access = create_test_access(ctx, n_test++);
2346 scop = extract_non_affine_condition(cond, isl_map_copy(test_access));
2347 scop = scop_add_array(scop, test_access, ast_context);
2348 scop_body = extract(stmt->getBody());
2350 id = isl_id_alloc(ctx, "t", NULL);
2351 domain = infinite_domain(isl_id_copy(id), scop_body);
2352 ident = identity_map(domain);
2354 has_var_break = pet_scop_has_var_skip(scop_body, pet_skip_later);
2355 if (has_var_break)
2356 break_access = pet_scop_get_skip_map(scop_body, pet_skip_later);
2358 scop = pet_scop_prefix(scop, 0);
2359 scop = pet_scop_embed(scop, isl_set_copy(domain), isl_map_copy(ident),
2360 isl_map_copy(ident), isl_id_copy(id));
2361 scop_body = pet_scop_reset_context(scop_body);
2362 scop_body = pet_scop_prefix(scop_body, 1);
2363 scop_body = pet_scop_embed(scop_body, isl_set_copy(domain),
2364 isl_map_copy(ident), ident, id);
2366 if (has_var_break) {
2367 scop = scop_add_break(scop, isl_map_copy(break_access),
2368 isl_set_copy(domain), 1);
2369 scop_body = scop_add_break(scop_body, break_access,
2370 isl_set_copy(domain), 1);
2372 scop = scop_add_while(scop, scop_body, test_access, domain, 1);
2374 return scop;
2377 /* Check whether "cond" expresses a simple loop bound
2378 * on the only set dimension.
2379 * In particular, if "up" is set then "cond" should contain only
2380 * upper bounds on the set dimension.
2381 * Otherwise, it should contain only lower bounds.
2383 static bool is_simple_bound(__isl_keep isl_set *cond, isl_int inc)
2385 if (isl_int_is_pos(inc))
2386 return !isl_set_dim_has_any_lower_bound(cond, isl_dim_set, 0);
2387 else
2388 return !isl_set_dim_has_any_upper_bound(cond, isl_dim_set, 0);
2391 /* Extend a condition on a given iteration of a loop to one that
2392 * imposes the same condition on all previous iterations.
2393 * "domain" expresses the lower [upper] bound on the iterations
2394 * when inc is positive [negative].
2396 * In particular, we construct the condition (when inc is positive)
2398 * forall i' : (domain(i') and i' <= i) => cond(i')
2400 * which is equivalent to
2402 * not exists i' : domain(i') and i' <= i and not cond(i')
2404 * We construct this set by negating cond, applying a map
2406 * { [i'] -> [i] : domain(i') and i' <= i }
2408 * and then negating the result again.
2410 static __isl_give isl_set *valid_for_each_iteration(__isl_take isl_set *cond,
2411 __isl_take isl_set *domain, isl_int inc)
2413 isl_map *previous_to_this;
2415 if (isl_int_is_pos(inc))
2416 previous_to_this = isl_map_lex_le(isl_set_get_space(domain));
2417 else
2418 previous_to_this = isl_map_lex_ge(isl_set_get_space(domain));
2420 previous_to_this = isl_map_intersect_domain(previous_to_this, domain);
2422 cond = isl_set_complement(cond);
2423 cond = isl_set_apply(cond, previous_to_this);
2424 cond = isl_set_complement(cond);
2426 return cond;
2429 /* Construct a domain of the form
2431 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
2433 static __isl_give isl_set *strided_domain(__isl_take isl_id *id,
2434 __isl_take isl_pw_aff *init, isl_int inc)
2436 isl_aff *aff;
2437 isl_space *dim;
2438 isl_set *set;
2440 init = isl_pw_aff_insert_dims(init, isl_dim_in, 0, 1);
2441 dim = isl_pw_aff_get_domain_space(init);
2442 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2443 aff = isl_aff_add_coefficient(aff, isl_dim_in, 0, inc);
2444 init = isl_pw_aff_add(init, isl_pw_aff_from_aff(aff));
2446 dim = isl_space_set_alloc(isl_pw_aff_get_ctx(init), 1, 1);
2447 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
2448 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2449 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2451 set = isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff), init);
2453 set = isl_set_lower_bound_si(set, isl_dim_set, 0, 0);
2455 return isl_set_params(set);
2458 /* Assuming "cond" represents a bound on a loop where the loop
2459 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2460 * is possible.
2462 * Under the given assumptions, wrapping is only possible if "cond" allows
2463 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2464 * increasing iterator and 0 in case of a decreasing iterator.
2466 static bool can_wrap(__isl_keep isl_set *cond, ValueDecl *iv, isl_int inc)
2468 bool cw;
2469 isl_int limit;
2470 isl_set *test;
2472 test = isl_set_copy(cond);
2474 isl_int_init(limit);
2475 if (isl_int_is_neg(inc))
2476 isl_int_set_si(limit, 0);
2477 else {
2478 isl_int_set_si(limit, 1);
2479 isl_int_mul_2exp(limit, limit, get_type_size(iv));
2480 isl_int_sub_ui(limit, limit, 1);
2483 test = isl_set_fix(cond, isl_dim_set, 0, limit);
2484 cw = !isl_set_is_empty(test);
2485 isl_set_free(test);
2487 isl_int_clear(limit);
2489 return cw;
2492 /* Given a one-dimensional space, construct the following mapping on this
2493 * space
2495 * { [v] -> [v mod 2^width] }
2497 * where width is the number of bits used to represent the values
2498 * of the unsigned variable "iv".
2500 static __isl_give isl_map *compute_wrapping(__isl_take isl_space *dim,
2501 ValueDecl *iv)
2503 isl_int mod;
2504 isl_aff *aff;
2505 isl_map *map;
2507 isl_int_init(mod);
2508 isl_int_set_si(mod, 1);
2509 isl_int_mul_2exp(mod, mod, get_type_size(iv));
2511 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2512 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2513 aff = isl_aff_mod(aff, mod);
2515 isl_int_clear(mod);
2517 return isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2518 map = isl_map_reverse(map);
2521 /* Project out the parameter "id" from "set".
2523 static __isl_give isl_set *set_project_out_by_id(__isl_take isl_set *set,
2524 __isl_keep isl_id *id)
2526 int pos;
2528 pos = isl_set_find_dim_by_id(set, isl_dim_param, id);
2529 if (pos >= 0)
2530 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2532 return set;
2535 /* Compute the set of parameters for which "set1" is a subset of "set2".
2537 * set1 is a subset of set2 if
2539 * forall i in set1 : i in set2
2541 * or
2543 * not exists i in set1 and i not in set2
2545 * i.e.,
2547 * not exists i in set1 \ set2
2549 static __isl_give isl_set *enforce_subset(__isl_take isl_set *set1,
2550 __isl_take isl_set *set2)
2552 return isl_set_complement(isl_set_params(isl_set_subtract(set1, set2)));
2555 /* Compute the set of parameter values for which "cond" holds
2556 * on the next iteration for each element of "dom".
2558 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2559 * and then compute the set of parameters for which the result is a subset
2560 * of "cond".
2562 static __isl_give isl_set *valid_on_next(__isl_take isl_set *cond,
2563 __isl_take isl_set *dom, isl_int inc)
2565 isl_space *space;
2566 isl_aff *aff;
2567 isl_map *next;
2569 space = isl_set_get_space(dom);
2570 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2571 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2572 aff = isl_aff_add_constant(aff, inc);
2573 next = isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2575 dom = isl_set_apply(dom, next);
2577 return enforce_subset(dom, cond);
2580 /* Does "id" refer to a nested access?
2582 static bool is_nested_parameter(__isl_keep isl_id *id)
2584 return id && isl_id_get_user(id) && !isl_id_get_name(id);
2587 /* Does parameter "pos" of "space" refer to a nested access?
2589 static bool is_nested_parameter(__isl_keep isl_space *space, int pos)
2591 bool nested;
2592 isl_id *id;
2594 id = isl_space_get_dim_id(space, isl_dim_param, pos);
2595 nested = is_nested_parameter(id);
2596 isl_id_free(id);
2598 return nested;
2601 /* Does "space" involve any parameters that refer to nested
2602 * accesses, i.e., parameters with no name?
2604 static bool has_nested(__isl_keep isl_space *space)
2606 int nparam;
2608 nparam = isl_space_dim(space, isl_dim_param);
2609 for (int i = 0; i < nparam; ++i)
2610 if (is_nested_parameter(space, i))
2611 return true;
2613 return false;
2616 /* Does "pa" involve any parameters that refer to nested
2617 * accesses, i.e., parameters with no name?
2619 static bool has_nested(__isl_keep isl_pw_aff *pa)
2621 isl_space *space;
2622 bool nested;
2624 space = isl_pw_aff_get_space(pa);
2625 nested = has_nested(space);
2626 isl_space_free(space);
2628 return nested;
2631 /* Construct a pet_scop for a for statement.
2632 * The for loop is required to be of the form
2634 * for (i = init; condition; ++i)
2636 * or
2638 * for (i = init; condition; --i)
2640 * The initialization of the for loop should either be an assignment
2641 * to an integer variable, or a declaration of such a variable with
2642 * initialization.
2644 * The condition is allowed to contain nested accesses, provided
2645 * they are not being written to inside the body of the loop.
2646 * Otherwise, or if the condition is otherwise non-affine, the for loop is
2647 * essentially treated as a while loop, with iteration domain
2648 * { [i] : i >= init }.
2650 * We extract a pet_scop for the body and then embed it in a loop with
2651 * iteration domain and schedule
2653 * { [i] : i >= init and condition' }
2654 * { [i] -> [i] }
2656 * or
2658 * { [i] : i <= init and condition' }
2659 * { [i] -> [-i] }
2661 * Where condition' is equal to condition if the latter is
2662 * a simple upper [lower] bound and a condition that is extended
2663 * to apply to all previous iterations otherwise.
2665 * If the condition is non-affine, then we drop the condition from the
2666 * iteration domain and instead create a separate statement
2667 * for evaluating the condition. The body is then filtered to depend
2668 * on the result of the condition evaluating to true on all iterations
2669 * up to the current iteration, while the evaluation the condition itself
2670 * is filtered to depend on the result of the condition evaluating to true
2671 * on all previous iterations.
2672 * The context of the scop representing the body is dropped
2673 * because we don't know how many times the body will be executed,
2674 * if at all.
2676 * If the stride of the loop is not 1, then "i >= init" is replaced by
2678 * (exists a: i = init + stride * a and a >= 0)
2680 * If the loop iterator i is unsigned, then wrapping may occur.
2681 * During the computation, we work with a virtual iterator that
2682 * does not wrap. However, the condition in the code applies
2683 * to the wrapped value, so we need to change condition(i)
2684 * into condition([i % 2^width]).
2685 * After computing the virtual domain and schedule, we apply
2686 * the function { [v] -> [v % 2^width] } to the domain and the domain
2687 * of the schedule. In order not to lose any information, we also
2688 * need to intersect the domain of the schedule with the virtual domain
2689 * first, since some iterations in the wrapped domain may be scheduled
2690 * several times, typically an infinite number of times.
2691 * Note that there may be no need to perform this final wrapping
2692 * if the loop condition (after wrapping) satisfies certain conditions.
2693 * However, the is_simple_bound condition is not enough since it doesn't
2694 * check if there even is an upper bound.
2696 * If the loop condition is non-affine, then we keep the virtual
2697 * iterator in the iteration domain and instead replace all accesses
2698 * to the original iterator by the wrapping of the virtual iterator.
2700 * Wrapping on unsigned iterators can be avoided entirely if
2701 * loop condition is simple, the loop iterator is incremented
2702 * [decremented] by one and the last value before wrapping cannot
2703 * possibly satisfy the loop condition.
2705 * Before extracting a pet_scop from the body we remove all
2706 * assignments in assigned_value to variables that are assigned
2707 * somewhere in the body of the loop.
2709 * Valid parameters for a for loop are those for which the initial
2710 * value itself, the increment on each domain iteration and
2711 * the condition on both the initial value and
2712 * the result of incrementing the iterator for each iteration of the domain
2713 * can be evaluated.
2714 * If the loop condition is non-affine, then we only consider validity
2715 * of the initial value.
2717 * If the body contains any break, then we keep track of it in "skip"
2718 * (if the skip condition is affine) or it is handled in scop_add_break
2719 * (if the skip condition is not affine).
2720 * Note that the affine break condition needs to be considered with
2721 * respect to previous iterations in the virtual domain (if any)
2722 * and that the domain needs to be kept virtual if there is a non-affine
2723 * break condition.
2725 struct pet_scop *PetScan::extract_for(ForStmt *stmt)
2727 BinaryOperator *ass;
2728 Decl *decl;
2729 Stmt *init;
2730 Expr *lhs, *rhs;
2731 ValueDecl *iv;
2732 isl_space *space;
2733 isl_set *domain;
2734 isl_map *sched;
2735 isl_set *cond = NULL;
2736 isl_set *skip = NULL;
2737 isl_id *id;
2738 struct pet_scop *scop, *scop_cond = NULL;
2739 assigned_value_cache cache(assigned_value);
2740 isl_int inc;
2741 bool is_one;
2742 bool is_unsigned;
2743 bool is_simple;
2744 bool is_virtual;
2745 bool keep_virtual = false;
2746 bool has_affine_break;
2747 bool has_var_break;
2748 isl_map *wrap = NULL;
2749 isl_pw_aff *pa, *pa_inc, *init_val;
2750 isl_set *valid_init;
2751 isl_set *valid_cond;
2752 isl_set *valid_cond_init;
2753 isl_set *valid_cond_next;
2754 isl_set *valid_inc;
2755 isl_map *test_access = NULL, *break_access = NULL;
2756 int stmt_id;
2758 if (!stmt->getInit() && !stmt->getCond() && !stmt->getInc())
2759 return extract_infinite_for(stmt);
2761 init = stmt->getInit();
2762 if (!init) {
2763 unsupported(stmt);
2764 return NULL;
2766 if ((ass = initialization_assignment(init)) != NULL) {
2767 iv = extract_induction_variable(ass);
2768 if (!iv)
2769 return NULL;
2770 lhs = ass->getLHS();
2771 rhs = ass->getRHS();
2772 } else if ((decl = initialization_declaration(init)) != NULL) {
2773 VarDecl *var = extract_induction_variable(init, decl);
2774 if (!var)
2775 return NULL;
2776 iv = var;
2777 rhs = var->getInit();
2778 lhs = create_DeclRefExpr(var);
2779 } else {
2780 unsupported(stmt->getInit());
2781 return NULL;
2784 pa_inc = extract_increment(stmt, iv);
2785 if (!pa_inc)
2786 return NULL;
2788 isl_int_init(inc);
2789 if (isl_pw_aff_n_piece(pa_inc) != 1 ||
2790 isl_pw_aff_foreach_piece(pa_inc, &extract_cst, &inc) < 0) {
2791 isl_pw_aff_free(pa_inc);
2792 unsupported(stmt->getInc());
2793 isl_int_clear(inc);
2794 return NULL;
2796 valid_inc = isl_pw_aff_domain(pa_inc);
2798 is_unsigned = iv->getType()->isUnsignedIntegerType();
2800 assigned_value.erase(iv);
2801 clear_assignments clear(assigned_value);
2802 clear.TraverseStmt(stmt->getBody());
2804 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
2806 pa = try_extract_nested_condition(stmt->getCond());
2807 if (allow_nested && (!pa || has_nested(pa)))
2808 stmt_id = n_stmt++;
2810 scop = extract(stmt->getBody());
2812 has_affine_break = scop &&
2813 pet_scop_has_affine_skip(scop, pet_skip_later);
2814 if (has_affine_break) {
2815 skip = pet_scop_get_skip(scop, pet_skip_later);
2816 skip = isl_set_fix_si(skip, isl_dim_set, 0, 1);
2817 skip = isl_set_params(skip);
2819 has_var_break = scop && pet_scop_has_var_skip(scop, pet_skip_later);
2820 if (has_var_break) {
2821 break_access = pet_scop_get_skip_map(scop, pet_skip_later);
2822 keep_virtual = true;
2825 if (pa && !is_nested_allowed(pa, scop)) {
2826 isl_pw_aff_free(pa);
2827 pa = NULL;
2830 if (!allow_nested && !pa)
2831 pa = try_extract_affine_condition(stmt->getCond());
2832 valid_cond = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2833 cond = isl_pw_aff_non_zero_set(pa);
2834 if (allow_nested && !cond) {
2835 int save_n_stmt = n_stmt;
2836 test_access = create_test_access(ctx, n_test++);
2837 n_stmt = stmt_id;
2838 scop_cond = extract_non_affine_condition(stmt->getCond(),
2839 isl_map_copy(test_access));
2840 n_stmt = save_n_stmt;
2841 scop_cond = scop_add_array(scop_cond, test_access, ast_context);
2842 scop_cond = pet_scop_prefix(scop_cond, 0);
2843 scop = pet_scop_reset_context(scop);
2844 scop = pet_scop_prefix(scop, 1);
2845 keep_virtual = true;
2846 cond = isl_set_universe(isl_space_set_alloc(ctx, 0, 0));
2849 cond = embed(cond, isl_id_copy(id));
2850 skip = embed(skip, isl_id_copy(id));
2851 valid_cond = isl_set_coalesce(valid_cond);
2852 valid_cond = embed(valid_cond, isl_id_copy(id));
2853 valid_inc = embed(valid_inc, isl_id_copy(id));
2854 is_one = isl_int_is_one(inc) || isl_int_is_negone(inc);
2855 is_virtual = is_unsigned && (!is_one || can_wrap(cond, iv, inc));
2857 init_val = extract_affine(rhs);
2858 valid_cond_init = enforce_subset(
2859 isl_set_from_pw_aff(isl_pw_aff_copy(init_val)),
2860 isl_set_copy(valid_cond));
2861 if (is_one && !is_virtual) {
2862 isl_pw_aff_free(init_val);
2863 pa = extract_comparison(isl_int_is_pos(inc) ? BO_GE : BO_LE,
2864 lhs, rhs, init);
2865 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2866 valid_init = set_project_out_by_id(valid_init, id);
2867 domain = isl_pw_aff_non_zero_set(pa);
2868 } else {
2869 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(init_val));
2870 domain = strided_domain(isl_id_copy(id), init_val, inc);
2873 domain = embed(domain, isl_id_copy(id));
2874 if (is_virtual) {
2875 isl_map *rev_wrap;
2876 wrap = compute_wrapping(isl_set_get_space(cond), iv);
2877 rev_wrap = isl_map_reverse(isl_map_copy(wrap));
2878 cond = isl_set_apply(cond, isl_map_copy(rev_wrap));
2879 skip = isl_set_apply(skip, isl_map_copy(rev_wrap));
2880 valid_cond = isl_set_apply(valid_cond, isl_map_copy(rev_wrap));
2881 valid_inc = isl_set_apply(valid_inc, rev_wrap);
2883 is_simple = is_simple_bound(cond, inc);
2884 if (!is_simple) {
2885 cond = isl_set_gist(cond, isl_set_copy(domain));
2886 is_simple = is_simple_bound(cond, inc);
2888 if (!is_simple)
2889 cond = valid_for_each_iteration(cond,
2890 isl_set_copy(domain), inc);
2891 domain = isl_set_intersect(domain, cond);
2892 if (has_affine_break) {
2893 skip = isl_set_intersect(skip , isl_set_copy(domain));
2894 skip = after(skip, isl_int_sgn(inc));
2895 domain = isl_set_subtract(domain, skip);
2897 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, isl_id_copy(id));
2898 space = isl_space_from_domain(isl_set_get_space(domain));
2899 space = isl_space_add_dims(space, isl_dim_out, 1);
2900 sched = isl_map_universe(space);
2901 if (isl_int_is_pos(inc))
2902 sched = isl_map_equate(sched, isl_dim_in, 0, isl_dim_out, 0);
2903 else
2904 sched = isl_map_oppose(sched, isl_dim_in, 0, isl_dim_out, 0);
2906 valid_cond_next = valid_on_next(valid_cond, isl_set_copy(domain), inc);
2907 valid_inc = enforce_subset(isl_set_copy(domain), valid_inc);
2909 if (is_virtual && !keep_virtual) {
2910 wrap = isl_map_set_dim_id(wrap,
2911 isl_dim_out, 0, isl_id_copy(id));
2912 sched = isl_map_intersect_domain(sched, isl_set_copy(domain));
2913 domain = isl_set_apply(domain, isl_map_copy(wrap));
2914 sched = isl_map_apply_domain(sched, wrap);
2916 if (!(is_virtual && keep_virtual)) {
2917 space = isl_set_get_space(domain);
2918 wrap = isl_map_identity(isl_space_map_from_set(space));
2921 scop_cond = pet_scop_embed(scop_cond, isl_set_copy(domain),
2922 isl_map_copy(sched), isl_map_copy(wrap), isl_id_copy(id));
2923 scop = pet_scop_embed(scop, isl_set_copy(domain), sched, wrap, id);
2924 scop = resolve_nested(scop);
2925 if (has_var_break)
2926 scop = scop_add_break(scop, break_access, isl_set_copy(domain),
2927 isl_int_sgn(inc));
2928 if (test_access) {
2929 scop = scop_add_while(scop_cond, scop, test_access, domain,
2930 isl_int_sgn(inc));
2931 isl_set_free(valid_inc);
2932 } else {
2933 scop = pet_scop_restrict_context(scop, valid_inc);
2934 scop = pet_scop_restrict_context(scop, valid_cond_next);
2935 scop = pet_scop_restrict_context(scop, valid_cond_init);
2936 isl_set_free(domain);
2938 clear_assignment(assigned_value, iv);
2940 isl_int_clear(inc);
2942 scop = pet_scop_restrict_context(scop, valid_init);
2944 return scop;
2947 struct pet_scop *PetScan::extract(CompoundStmt *stmt, bool skip_declarations)
2949 return extract(stmt->children(), true, skip_declarations);
2952 /* Does parameter "pos" of "map" refer to a nested access?
2954 static bool is_nested_parameter(__isl_keep isl_map *map, int pos)
2956 bool nested;
2957 isl_id *id;
2959 id = isl_map_get_dim_id(map, isl_dim_param, pos);
2960 nested = is_nested_parameter(id);
2961 isl_id_free(id);
2963 return nested;
2966 /* How many parameters of "space" refer to nested accesses, i.e., have no name?
2968 static int n_nested_parameter(__isl_keep isl_space *space)
2970 int n = 0;
2971 int nparam;
2973 nparam = isl_space_dim(space, isl_dim_param);
2974 for (int i = 0; i < nparam; ++i)
2975 if (is_nested_parameter(space, i))
2976 ++n;
2978 return n;
2981 /* How many parameters of "map" refer to nested accesses, i.e., have no name?
2983 static int n_nested_parameter(__isl_keep isl_map *map)
2985 isl_space *space;
2986 int n;
2988 space = isl_map_get_space(map);
2989 n = n_nested_parameter(space);
2990 isl_space_free(space);
2992 return n;
2995 /* For each nested access parameter in "space",
2996 * construct a corresponding pet_expr, place it in args and
2997 * record its position in "param2pos".
2998 * "n_arg" is the number of elements that are already in args.
2999 * The position recorded in "param2pos" takes this number into account.
3000 * If the pet_expr corresponding to a parameter is identical to
3001 * the pet_expr corresponding to an earlier parameter, then these two
3002 * parameters are made to refer to the same element in args.
3004 * Return the final number of elements in args or -1 if an error has occurred.
3006 int PetScan::extract_nested(__isl_keep isl_space *space,
3007 int n_arg, struct pet_expr **args, std::map<int,int> &param2pos)
3009 int nparam;
3011 nparam = isl_space_dim(space, isl_dim_param);
3012 for (int i = 0; i < nparam; ++i) {
3013 int j;
3014 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
3015 Expr *nested;
3017 if (!is_nested_parameter(id)) {
3018 isl_id_free(id);
3019 continue;
3022 nested = (Expr *) isl_id_get_user(id);
3023 args[n_arg] = extract_expr(nested);
3024 if (!args[n_arg])
3025 return -1;
3027 for (j = 0; j < n_arg; ++j)
3028 if (pet_expr_is_equal(args[j], args[n_arg]))
3029 break;
3031 if (j < n_arg) {
3032 pet_expr_free(args[n_arg]);
3033 args[n_arg] = NULL;
3034 param2pos[i] = j;
3035 } else
3036 param2pos[i] = n_arg++;
3038 isl_id_free(id);
3041 return n_arg;
3044 /* For each nested access parameter in the access relations in "expr",
3045 * construct a corresponding pet_expr, place it in expr->args and
3046 * record its position in "param2pos".
3047 * n is the number of nested access parameters.
3049 struct pet_expr *PetScan::extract_nested(struct pet_expr *expr, int n,
3050 std::map<int,int> &param2pos)
3052 isl_space *space;
3054 expr->args = isl_calloc_array(ctx, struct pet_expr *, n);
3055 expr->n_arg = n;
3056 if (!expr->args)
3057 goto error;
3059 space = isl_map_get_space(expr->acc.access);
3060 n = extract_nested(space, 0, expr->args, param2pos);
3061 isl_space_free(space);
3063 if (n < 0)
3064 goto error;
3066 expr->n_arg = n;
3067 return expr;
3068 error:
3069 pet_expr_free(expr);
3070 return NULL;
3073 /* Look for parameters in any access relation in "expr" that
3074 * refer to nested accesses. In particular, these are
3075 * parameters with no name.
3077 * If there are any such parameters, then the domain of the access
3078 * relation, which is still [] at this point, is replaced by
3079 * [[] -> [t_1,...,t_n]], with n the number of these parameters
3080 * (after identifying identical nested accesses).
3081 * The parameters are then equated to the corresponding t dimensions
3082 * and subsequently projected out.
3083 * param2pos maps the position of the parameter to the position
3084 * of the corresponding t dimension.
3086 struct pet_expr *PetScan::resolve_nested(struct pet_expr *expr)
3088 int n;
3089 int nparam;
3090 int n_in;
3091 isl_space *dim;
3092 isl_map *map;
3093 std::map<int,int> param2pos;
3095 if (!expr)
3096 return expr;
3098 for (int i = 0; i < expr->n_arg; ++i) {
3099 expr->args[i] = resolve_nested(expr->args[i]);
3100 if (!expr->args[i]) {
3101 pet_expr_free(expr);
3102 return NULL;
3106 if (expr->type != pet_expr_access)
3107 return expr;
3109 n = n_nested_parameter(expr->acc.access);
3110 if (n == 0)
3111 return expr;
3113 expr = extract_nested(expr, n, param2pos);
3114 if (!expr)
3115 return NULL;
3117 n = expr->n_arg;
3118 nparam = isl_map_dim(expr->acc.access, isl_dim_param);
3119 n_in = isl_map_dim(expr->acc.access, isl_dim_in);
3120 dim = isl_map_get_space(expr->acc.access);
3121 dim = isl_space_domain(dim);
3122 dim = isl_space_from_domain(dim);
3123 dim = isl_space_add_dims(dim, isl_dim_out, n);
3124 map = isl_map_universe(dim);
3125 map = isl_map_domain_map(map);
3126 map = isl_map_reverse(map);
3127 expr->acc.access = isl_map_apply_domain(expr->acc.access, map);
3129 for (int i = nparam - 1; i >= 0; --i) {
3130 isl_id *id = isl_map_get_dim_id(expr->acc.access,
3131 isl_dim_param, i);
3132 if (!is_nested_parameter(id)) {
3133 isl_id_free(id);
3134 continue;
3137 expr->acc.access = isl_map_equate(expr->acc.access,
3138 isl_dim_param, i, isl_dim_in,
3139 n_in + param2pos[i]);
3140 expr->acc.access = isl_map_project_out(expr->acc.access,
3141 isl_dim_param, i, 1);
3143 isl_id_free(id);
3146 return expr;
3147 error:
3148 pet_expr_free(expr);
3149 return NULL;
3152 /* Return the file offset of the expansion location of "Loc".
3154 static unsigned getExpansionOffset(SourceManager &SM, SourceLocation Loc)
3156 return SM.getFileOffset(SM.getExpansionLoc(Loc));
3159 #ifdef HAVE_FINDLOCATIONAFTERTOKEN
3161 /* Return a SourceLocation for the location after the first semicolon
3162 * after "loc". If Lexer::findLocationAfterToken is available, we simply
3163 * call it and also skip trailing spaces and newline.
3165 static SourceLocation location_after_semi(SourceLocation loc, SourceManager &SM,
3166 const LangOptions &LO)
3168 return Lexer::findLocationAfterToken(loc, tok::semi, SM, LO, true);
3171 #else
3173 /* Return a SourceLocation for the location after the first semicolon
3174 * after "loc". If Lexer::findLocationAfterToken is not available,
3175 * we look in the underlying character data for the first semicolon.
3177 static SourceLocation location_after_semi(SourceLocation loc, SourceManager &SM,
3178 const LangOptions &LO)
3180 const char *semi;
3181 const char *s = SM.getCharacterData(loc);
3183 semi = strchr(s, ';');
3184 if (!semi)
3185 return SourceLocation();
3186 return loc.getFileLocWithOffset(semi + 1 - s);
3189 #endif
3191 /* Convert a top-level pet_expr to a pet_scop with one statement.
3192 * This mainly involves resolving nested expression parameters
3193 * and setting the name of the iteration space.
3194 * The name is given by "label" if it is non-NULL. Otherwise,
3195 * it is of the form S_<n_stmt>.
3196 * start and end of the pet_scop are derived from those of "stmt".
3198 struct pet_scop *PetScan::extract(Stmt *stmt, struct pet_expr *expr,
3199 __isl_take isl_id *label)
3201 struct pet_stmt *ps;
3202 struct pet_scop *scop;
3203 SourceLocation loc = stmt->getLocStart();
3204 SourceManager &SM = PP.getSourceManager();
3205 const LangOptions &LO = PP.getLangOpts();
3206 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3207 unsigned start, end;
3209 expr = resolve_nested(expr);
3210 ps = pet_stmt_from_pet_expr(ctx, line, label, n_stmt++, expr);
3211 scop = pet_scop_from_pet_stmt(ctx, ps);
3213 start = getExpansionOffset(SM, loc);
3214 loc = stmt->getLocEnd();
3215 loc = location_after_semi(loc, SM, LO);
3216 end = getExpansionOffset(SM, loc);
3218 scop = pet_scop_update_start_end(scop, start, end);
3219 return scop;
3222 /* Check if we can extract an affine expression from "expr".
3223 * Return the expressions as an isl_pw_aff if we can and NULL otherwise.
3224 * We turn on autodetection so that we won't generate any warnings
3225 * and turn off nesting, so that we won't accept any non-affine constructs.
3227 __isl_give isl_pw_aff *PetScan::try_extract_affine(Expr *expr)
3229 isl_pw_aff *pwaff;
3230 int save_autodetect = options->autodetect;
3231 bool save_nesting = nesting_enabled;
3233 options->autodetect = 1;
3234 nesting_enabled = false;
3236 pwaff = extract_affine(expr);
3238 options->autodetect = save_autodetect;
3239 nesting_enabled = save_nesting;
3241 return pwaff;
3244 /* Check whether "expr" is an affine expression.
3246 bool PetScan::is_affine(Expr *expr)
3248 isl_pw_aff *pwaff;
3250 pwaff = try_extract_affine(expr);
3251 isl_pw_aff_free(pwaff);
3253 return pwaff != NULL;
3256 /* Check if we can extract an affine constraint from "expr".
3257 * Return the constraint as an isl_set if we can and NULL otherwise.
3258 * We turn on autodetection so that we won't generate any warnings
3259 * and turn off nesting, so that we won't accept any non-affine constructs.
3261 __isl_give isl_pw_aff *PetScan::try_extract_affine_condition(Expr *expr)
3263 isl_pw_aff *cond;
3264 int save_autodetect = options->autodetect;
3265 bool save_nesting = nesting_enabled;
3267 options->autodetect = 1;
3268 nesting_enabled = false;
3270 cond = extract_condition(expr);
3272 options->autodetect = save_autodetect;
3273 nesting_enabled = save_nesting;
3275 return cond;
3278 /* Check whether "expr" is an affine constraint.
3280 bool PetScan::is_affine_condition(Expr *expr)
3282 isl_pw_aff *cond;
3284 cond = try_extract_affine_condition(expr);
3285 isl_pw_aff_free(cond);
3287 return cond != NULL;
3290 /* Check if we can extract a condition from "expr".
3291 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
3292 * If allow_nested is set, then the condition may involve parameters
3293 * corresponding to nested accesses.
3294 * We turn on autodetection so that we won't generate any warnings.
3296 __isl_give isl_pw_aff *PetScan::try_extract_nested_condition(Expr *expr)
3298 isl_pw_aff *cond;
3299 int save_autodetect = options->autodetect;
3300 bool save_nesting = nesting_enabled;
3302 options->autodetect = 1;
3303 nesting_enabled = allow_nested;
3304 cond = extract_condition(expr);
3306 options->autodetect = save_autodetect;
3307 nesting_enabled = save_nesting;
3309 return cond;
3312 /* If the top-level expression of "stmt" is an assignment, then
3313 * return that assignment as a BinaryOperator.
3314 * Otherwise return NULL.
3316 static BinaryOperator *top_assignment_or_null(Stmt *stmt)
3318 BinaryOperator *ass;
3320 if (!stmt)
3321 return NULL;
3322 if (stmt->getStmtClass() != Stmt::BinaryOperatorClass)
3323 return NULL;
3325 ass = cast<BinaryOperator>(stmt);
3326 if(ass->getOpcode() != BO_Assign)
3327 return NULL;
3329 return ass;
3332 /* Check if the given if statement is a conditional assignement
3333 * with a non-affine condition. If so, construct a pet_scop
3334 * corresponding to this conditional assignment. Otherwise return NULL.
3336 * In particular we check if "stmt" is of the form
3338 * if (condition)
3339 * a = f(...);
3340 * else
3341 * a = g(...);
3343 * where a is some array or scalar access.
3344 * The constructed pet_scop then corresponds to the expression
3346 * a = condition ? f(...) : g(...)
3348 * All access relations in f(...) are intersected with condition
3349 * while all access relation in g(...) are intersected with the complement.
3351 struct pet_scop *PetScan::extract_conditional_assignment(IfStmt *stmt)
3353 BinaryOperator *ass_then, *ass_else;
3354 isl_map *write_then, *write_else;
3355 isl_set *cond, *comp;
3356 isl_map *map;
3357 isl_pw_aff *pa;
3358 int equal;
3359 struct pet_expr *pe_cond, *pe_then, *pe_else, *pe, *pe_write;
3360 bool save_nesting = nesting_enabled;
3362 if (!options->detect_conditional_assignment)
3363 return NULL;
3365 ass_then = top_assignment_or_null(stmt->getThen());
3366 ass_else = top_assignment_or_null(stmt->getElse());
3368 if (!ass_then || !ass_else)
3369 return NULL;
3371 if (is_affine_condition(stmt->getCond()))
3372 return NULL;
3374 write_then = extract_access(ass_then->getLHS());
3375 write_else = extract_access(ass_else->getLHS());
3377 equal = isl_map_is_equal(write_then, write_else);
3378 isl_map_free(write_else);
3379 if (equal < 0 || !equal) {
3380 isl_map_free(write_then);
3381 return NULL;
3384 nesting_enabled = allow_nested;
3385 pa = extract_condition(stmt->getCond());
3386 nesting_enabled = save_nesting;
3387 cond = isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa));
3388 comp = isl_pw_aff_zero_set(isl_pw_aff_copy(pa));
3389 map = isl_map_from_range(isl_set_from_pw_aff(pa));
3391 pe_cond = pet_expr_from_access(map);
3393 pe_then = extract_expr(ass_then->getRHS());
3394 pe_then = pet_expr_restrict(pe_then, cond);
3395 pe_else = extract_expr(ass_else->getRHS());
3396 pe_else = pet_expr_restrict(pe_else, comp);
3398 pe = pet_expr_new_ternary(ctx, pe_cond, pe_then, pe_else);
3399 pe_write = pet_expr_from_access(write_then);
3400 if (pe_write) {
3401 pe_write->acc.write = 1;
3402 pe_write->acc.read = 0;
3404 pe = pet_expr_new_binary(ctx, pet_op_assign, pe_write, pe);
3405 return extract(stmt, pe);
3408 /* Create a pet_scop with a single statement evaluating "cond"
3409 * and writing the result to a virtual scalar, as expressed by
3410 * "access".
3412 struct pet_scop *PetScan::extract_non_affine_condition(Expr *cond,
3413 __isl_take isl_map *access)
3415 struct pet_expr *expr, *write;
3416 struct pet_stmt *ps;
3417 struct pet_scop *scop;
3418 SourceLocation loc = cond->getLocStart();
3419 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3421 write = pet_expr_from_access(access);
3422 if (write) {
3423 write->acc.write = 1;
3424 write->acc.read = 0;
3426 expr = extract_expr(cond);
3427 expr = resolve_nested(expr);
3428 expr = pet_expr_new_binary(ctx, pet_op_assign, write, expr);
3429 ps = pet_stmt_from_pet_expr(ctx, line, NULL, n_stmt++, expr);
3430 scop = pet_scop_from_pet_stmt(ctx, ps);
3431 scop = resolve_nested(scop);
3433 return scop;
3436 extern "C" {
3437 static __isl_give isl_map *embed_access(__isl_take isl_map *access,
3438 void *user);
3441 /* Apply the map pointed to by "user" to the domain of the access
3442 * relation, thereby embedding it in the range of the map.
3443 * The domain of both relations is the zero-dimensional domain.
3445 static __isl_give isl_map *embed_access(__isl_take isl_map *access, void *user)
3447 isl_map *map = (isl_map *) user;
3449 return isl_map_apply_domain(access, isl_map_copy(map));
3452 /* Apply "map" to all access relations in "expr".
3454 static struct pet_expr *embed(struct pet_expr *expr, __isl_keep isl_map *map)
3456 return pet_expr_foreach_access(expr, &embed_access, map);
3459 /* How many parameters of "set" refer to nested accesses, i.e., have no name?
3461 static int n_nested_parameter(__isl_keep isl_set *set)
3463 isl_space *space;
3464 int n;
3466 space = isl_set_get_space(set);
3467 n = n_nested_parameter(space);
3468 isl_space_free(space);
3470 return n;
3473 /* Remove all parameters from "map" that refer to nested accesses.
3475 static __isl_give isl_map *remove_nested_parameters(__isl_take isl_map *map)
3477 int nparam;
3478 isl_space *space;
3480 space = isl_map_get_space(map);
3481 nparam = isl_space_dim(space, isl_dim_param);
3482 for (int i = nparam - 1; i >= 0; --i)
3483 if (is_nested_parameter(space, i))
3484 map = isl_map_project_out(map, isl_dim_param, i, 1);
3485 isl_space_free(space);
3487 return map;
3490 extern "C" {
3491 static __isl_give isl_map *access_remove_nested_parameters(
3492 __isl_take isl_map *access, void *user);
3495 static __isl_give isl_map *access_remove_nested_parameters(
3496 __isl_take isl_map *access, void *user)
3498 return remove_nested_parameters(access);
3501 /* Remove all nested access parameters from the schedule and all
3502 * accesses of "stmt".
3503 * There is no need to remove them from the domain as these parameters
3504 * have already been removed from the domain when this function is called.
3506 static struct pet_stmt *remove_nested_parameters(struct pet_stmt *stmt)
3508 if (!stmt)
3509 return NULL;
3510 stmt->schedule = remove_nested_parameters(stmt->schedule);
3511 stmt->body = pet_expr_foreach_access(stmt->body,
3512 &access_remove_nested_parameters, NULL);
3513 if (!stmt->schedule || !stmt->body)
3514 goto error;
3515 for (int i = 0; i < stmt->n_arg; ++i) {
3516 stmt->args[i] = pet_expr_foreach_access(stmt->args[i],
3517 &access_remove_nested_parameters, NULL);
3518 if (!stmt->args[i])
3519 goto error;
3522 return stmt;
3523 error:
3524 pet_stmt_free(stmt);
3525 return NULL;
3528 /* For each nested access parameter in the domain of "stmt",
3529 * construct a corresponding pet_expr, place it before the original
3530 * elements in stmt->args and record its position in "param2pos".
3531 * n is the number of nested access parameters.
3533 struct pet_stmt *PetScan::extract_nested(struct pet_stmt *stmt, int n,
3534 std::map<int,int> &param2pos)
3536 int i;
3537 isl_space *space;
3538 int n_arg;
3539 struct pet_expr **args;
3541 n_arg = stmt->n_arg;
3542 args = isl_calloc_array(ctx, struct pet_expr *, n + n_arg);
3543 if (!args)
3544 goto error;
3546 space = isl_set_get_space(stmt->domain);
3547 n_arg = extract_nested(space, 0, args, param2pos);
3548 isl_space_free(space);
3550 if (n_arg < 0)
3551 goto error;
3553 for (i = 0; i < stmt->n_arg; ++i)
3554 args[n_arg + i] = stmt->args[i];
3555 free(stmt->args);
3556 stmt->args = args;
3557 stmt->n_arg += n_arg;
3559 return stmt;
3560 error:
3561 if (args) {
3562 for (i = 0; i < n; ++i)
3563 pet_expr_free(args[i]);
3564 free(args);
3566 pet_stmt_free(stmt);
3567 return NULL;
3570 /* Check whether any of the arguments i of "stmt" starting at position "n"
3571 * is equal to one of the first "n" arguments j.
3572 * If so, combine the constraints on arguments i and j and remove
3573 * argument i.
3575 static struct pet_stmt *remove_duplicate_arguments(struct pet_stmt *stmt, int n)
3577 int i, j;
3578 isl_map *map;
3580 if (!stmt)
3581 return NULL;
3582 if (n == 0)
3583 return stmt;
3584 if (n == stmt->n_arg)
3585 return stmt;
3587 map = isl_set_unwrap(stmt->domain);
3589 for (i = stmt->n_arg - 1; i >= n; --i) {
3590 for (j = 0; j < n; ++j)
3591 if (pet_expr_is_equal(stmt->args[i], stmt->args[j]))
3592 break;
3593 if (j >= n)
3594 continue;
3596 map = isl_map_equate(map, isl_dim_out, i, isl_dim_out, j);
3597 map = isl_map_project_out(map, isl_dim_out, i, 1);
3599 pet_expr_free(stmt->args[i]);
3600 for (j = i; j + 1 < stmt->n_arg; ++j)
3601 stmt->args[j] = stmt->args[j + 1];
3602 stmt->n_arg--;
3605 stmt->domain = isl_map_wrap(map);
3606 if (!stmt->domain)
3607 goto error;
3608 return stmt;
3609 error:
3610 pet_stmt_free(stmt);
3611 return NULL;
3614 /* Look for parameters in the iteration domain of "stmt" that
3615 * refer to nested accesses. In particular, these are
3616 * parameters with no name.
3618 * If there are any such parameters, then as many extra variables
3619 * (after identifying identical nested accesses) are inserted in the
3620 * range of the map wrapped inside the domain, before the original variables.
3621 * If the original domain is not a wrapped map, then a new wrapped
3622 * map is created with zero output dimensions.
3623 * The parameters are then equated to the corresponding output dimensions
3624 * and subsequently projected out, from the iteration domain,
3625 * the schedule and the access relations.
3626 * For each of the output dimensions, a corresponding argument
3627 * expression is inserted. Initially they are created with
3628 * a zero-dimensional domain, so they have to be embedded
3629 * in the current iteration domain.
3630 * param2pos maps the position of the parameter to the position
3631 * of the corresponding output dimension in the wrapped map.
3633 struct pet_stmt *PetScan::resolve_nested(struct pet_stmt *stmt)
3635 int n;
3636 int nparam;
3637 unsigned n_arg;
3638 isl_map *map;
3639 std::map<int,int> param2pos;
3641 if (!stmt)
3642 return NULL;
3644 n = n_nested_parameter(stmt->domain);
3645 if (n == 0)
3646 return stmt;
3648 n_arg = stmt->n_arg;
3649 stmt = extract_nested(stmt, n, param2pos);
3650 if (!stmt)
3651 return NULL;
3653 n = stmt->n_arg - n_arg;
3654 nparam = isl_set_dim(stmt->domain, isl_dim_param);
3655 if (isl_set_is_wrapping(stmt->domain))
3656 map = isl_set_unwrap(stmt->domain);
3657 else
3658 map = isl_map_from_domain(stmt->domain);
3659 map = isl_map_insert_dims(map, isl_dim_out, 0, n);
3661 for (int i = nparam - 1; i >= 0; --i) {
3662 isl_id *id;
3664 if (!is_nested_parameter(map, i))
3665 continue;
3667 id = isl_map_get_tuple_id(stmt->args[param2pos[i]]->acc.access,
3668 isl_dim_out);
3669 map = isl_map_set_dim_id(map, isl_dim_out, param2pos[i], id);
3670 map = isl_map_equate(map, isl_dim_param, i, isl_dim_out,
3671 param2pos[i]);
3672 map = isl_map_project_out(map, isl_dim_param, i, 1);
3675 stmt->domain = isl_map_wrap(map);
3677 map = isl_set_unwrap(isl_set_copy(stmt->domain));
3678 map = isl_map_from_range(isl_map_domain(map));
3679 for (int pos = 0; pos < n; ++pos)
3680 stmt->args[pos] = embed(stmt->args[pos], map);
3681 isl_map_free(map);
3683 stmt = remove_nested_parameters(stmt);
3684 stmt = remove_duplicate_arguments(stmt, n);
3686 return stmt;
3687 error:
3688 pet_stmt_free(stmt);
3689 return NULL;
3692 /* For each statement in "scop", move the parameters that correspond
3693 * to nested access into the ranges of the domains and create
3694 * corresponding argument expressions.
3696 struct pet_scop *PetScan::resolve_nested(struct pet_scop *scop)
3698 if (!scop)
3699 return NULL;
3701 for (int i = 0; i < scop->n_stmt; ++i) {
3702 scop->stmts[i] = resolve_nested(scop->stmts[i]);
3703 if (!scop->stmts[i])
3704 goto error;
3707 return scop;
3708 error:
3709 pet_scop_free(scop);
3710 return NULL;
3713 /* Given an access expression "expr", is the variable accessed by
3714 * "expr" assigned anywhere inside "scop"?
3716 static bool is_assigned(pet_expr *expr, pet_scop *scop)
3718 bool assigned = false;
3719 isl_id *id;
3721 id = isl_map_get_tuple_id(expr->acc.access, isl_dim_out);
3722 assigned = pet_scop_writes(scop, id);
3723 isl_id_free(id);
3725 return assigned;
3728 /* Are all nested access parameters in "pa" allowed given "scop".
3729 * In particular, is none of them written by anywhere inside "scop".
3731 * If "scop" has any skip conditions, then no nested access parameters
3732 * are allowed. In particular, if there is any nested access in a guard
3733 * for a piece of code containing a "continue", then we want to introduce
3734 * a separate statement for evaluating this guard so that we can express
3735 * that the result is false for all previous iterations.
3737 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff *pa, pet_scop *scop)
3739 int nparam;
3741 if (!scop)
3742 return true;
3744 nparam = isl_pw_aff_dim(pa, isl_dim_param);
3745 for (int i = 0; i < nparam; ++i) {
3746 Expr *nested;
3747 isl_id *id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
3748 pet_expr *expr;
3749 bool allowed;
3751 if (!is_nested_parameter(id)) {
3752 isl_id_free(id);
3753 continue;
3756 if (pet_scop_has_skip(scop, pet_skip_now)) {
3757 isl_id_free(id);
3758 return false;
3761 nested = (Expr *) isl_id_get_user(id);
3762 expr = extract_expr(nested);
3763 allowed = expr && expr->type == pet_expr_access &&
3764 !is_assigned(expr, scop);
3766 pet_expr_free(expr);
3767 isl_id_free(id);
3769 if (!allowed)
3770 return false;
3773 return true;
3776 /* Do we need to construct a skip condition of the given type
3777 * on an if statement, given that the if condition is non-affine?
3779 * pet_scop_filter_skip can only handle the case where the if condition
3780 * holds (the then branch) and the skip condition is universal.
3781 * In any other case, we need to construct a new skip condition.
3783 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
3784 bool have_else, enum pet_skip type)
3786 if (have_else && scop_else && pet_scop_has_skip(scop_else, type))
3787 return true;
3788 if (scop_then && pet_scop_has_skip(scop_then, type) &&
3789 !pet_scop_has_universal_skip(scop_then, type))
3790 return true;
3791 return false;
3794 /* Do we need to construct a skip condition of the given type
3795 * on an if statement, given that the if condition is affine?
3797 * There is no need to construct a new skip condition if all
3798 * the skip conditions are affine.
3800 static bool need_skip_aff(struct pet_scop *scop_then,
3801 struct pet_scop *scop_else, bool have_else, enum pet_skip type)
3803 if (scop_then && pet_scop_has_var_skip(scop_then, type))
3804 return true;
3805 if (have_else && scop_else && pet_scop_has_var_skip(scop_else, type))
3806 return true;
3807 return false;
3810 /* Do we need to construct a skip condition of the given type
3811 * on an if statement?
3813 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
3814 bool have_else, enum pet_skip type, bool affine)
3816 if (affine)
3817 return need_skip_aff(scop_then, scop_else, have_else, type);
3818 else
3819 return need_skip(scop_then, scop_else, have_else, type);
3822 /* Construct an affine expression pet_expr that is evaluates
3823 * to the constant "val".
3825 static struct pet_expr *universally(isl_ctx *ctx, int val)
3827 isl_space *space;
3828 isl_map *map;
3830 space = isl_space_alloc(ctx, 0, 0, 1);
3831 map = isl_map_universe(space);
3832 map = isl_map_fix_si(map, isl_dim_out, 0, val);
3834 return pet_expr_from_access(map);
3837 /* Construct an affine expression pet_expr that is evaluates
3838 * to the constant 1.
3840 static struct pet_expr *universally_true(isl_ctx *ctx)
3842 return universally(ctx, 1);
3845 /* Construct an affine expression pet_expr that is evaluates
3846 * to the constant 0.
3848 static struct pet_expr *universally_false(isl_ctx *ctx)
3850 return universally(ctx, 0);
3853 /* Given an access relation "test_access" for the if condition,
3854 * an access relation "skip_access" for the skip condition and
3855 * scops for the then and else branches, construct a scop for
3856 * computing "skip_access".
3858 * The computed scop contains a single statement that essentially does
3860 * skip_cond = test_cond ? skip_cond_then : skip_cond_else
3862 * If the skip conditions of the then and/or else branch are not affine,
3863 * then they need to be filtered by test_access.
3864 * If they are missing, then this means the skip condition is false.
3866 * Since we are constructing a skip condition for the if statement,
3867 * the skip conditions on the then and else branches are removed.
3869 static struct pet_scop *extract_skip(PetScan *scan,
3870 __isl_take isl_map *test_access, __isl_take isl_map *skip_access,
3871 struct pet_scop *scop_then, struct pet_scop *scop_else, bool have_else,
3872 enum pet_skip type)
3874 struct pet_expr *expr_then, *expr_else, *expr, *expr_skip;
3875 struct pet_stmt *stmt;
3876 struct pet_scop *scop;
3877 isl_ctx *ctx = scan->ctx;
3879 if (!scop_then)
3880 goto error;
3881 if (have_else && !scop_else)
3882 goto error;
3884 if (pet_scop_has_skip(scop_then, type)) {
3885 expr_then = pet_scop_get_skip_expr(scop_then, type);
3886 pet_scop_reset_skip(scop_then, type);
3887 if (!pet_expr_is_affine(expr_then))
3888 expr_then = pet_expr_filter(expr_then,
3889 isl_map_copy(test_access), 1);
3890 } else
3891 expr_then = universally_false(ctx);
3893 if (have_else && pet_scop_has_skip(scop_else, type)) {
3894 expr_else = pet_scop_get_skip_expr(scop_else, type);
3895 pet_scop_reset_skip(scop_else, type);
3896 if (!pet_expr_is_affine(expr_else))
3897 expr_else = pet_expr_filter(expr_else,
3898 isl_map_copy(test_access), 0);
3899 } else
3900 expr_else = universally_false(ctx);
3902 expr = pet_expr_from_access(test_access);
3903 expr = pet_expr_new_ternary(ctx, expr, expr_then, expr_else);
3904 expr_skip = pet_expr_from_access(isl_map_copy(skip_access));
3905 if (expr_skip) {
3906 expr_skip->acc.write = 1;
3907 expr_skip->acc.read = 0;
3909 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
3910 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, scan->n_stmt++, expr);
3912 scop = pet_scop_from_pet_stmt(ctx, stmt);
3913 scop = scop_add_array(scop, skip_access, scan->ast_context);
3914 isl_map_free(skip_access);
3916 return scop;
3917 error:
3918 isl_map_free(test_access);
3919 isl_map_free(skip_access);
3920 return NULL;
3923 /* Is scop's skip_now condition equal to its skip_later condition?
3924 * In particular, this means that it either has no skip_now condition
3925 * or both a skip_now and a skip_later condition (that are equal to each other).
3927 static bool skip_equals_skip_later(struct pet_scop *scop)
3929 int has_skip_now, has_skip_later;
3930 int equal;
3931 isl_set *skip_now, *skip_later;
3933 if (!scop)
3934 return false;
3935 has_skip_now = pet_scop_has_skip(scop, pet_skip_now);
3936 has_skip_later = pet_scop_has_skip(scop, pet_skip_later);
3937 if (has_skip_now != has_skip_later)
3938 return false;
3939 if (!has_skip_now)
3940 return true;
3942 skip_now = pet_scop_get_skip(scop, pet_skip_now);
3943 skip_later = pet_scop_get_skip(scop, pet_skip_later);
3944 equal = isl_set_is_equal(skip_now, skip_later);
3945 isl_set_free(skip_now);
3946 isl_set_free(skip_later);
3948 return equal;
3951 /* Drop the skip conditions of type pet_skip_later from scop1 and scop2.
3953 static void drop_skip_later(struct pet_scop *scop1, struct pet_scop *scop2)
3955 pet_scop_reset_skip(scop1, pet_skip_later);
3956 pet_scop_reset_skip(scop2, pet_skip_later);
3959 /* Structure that handles the construction of skip conditions.
3961 * scop_then and scop_else represent the then and else branches
3962 * of the if statement
3964 * skip[type] is true if we need to construct a skip condition of that type
3965 * equal is set if the skip conditions of types pet_skip_now and pet_skip_later
3966 * are equal to each other
3967 * access[type] is the virtual array representing the skip condition
3968 * scop[type] is a scop for computing the skip condition
3970 struct pet_skip_info {
3971 isl_ctx *ctx;
3973 bool skip[2];
3974 bool equal;
3975 isl_map *access[2];
3976 struct pet_scop *scop[2];
3978 pet_skip_info(isl_ctx *ctx) : ctx(ctx) {}
3980 operator bool() { return skip[pet_skip_now] || skip[pet_skip_later]; }
3983 /* Structure that handles the construction of skip conditions on if statements.
3985 * scop_then and scop_else represent the then and else branches
3986 * of the if statement
3988 struct pet_skip_info_if : public pet_skip_info {
3989 struct pet_scop *scop_then, *scop_else;
3990 bool have_else;
3992 pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
3993 struct pet_scop *scop_else, bool have_else, bool affine);
3994 void extract(PetScan *scan, __isl_keep isl_map *access,
3995 enum pet_skip type);
3996 void extract(PetScan *scan, __isl_keep isl_map *access);
3997 void extract(PetScan *scan, __isl_keep isl_pw_aff *cond);
3998 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
3999 int offset);
4000 struct pet_scop *add(struct pet_scop *scop, int offset);
4003 /* Initialize a pet_skip_info_if structure based on the then and else branches
4004 * and based on whether the if condition is affine or not.
4006 pet_skip_info_if::pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
4007 struct pet_scop *scop_else, bool have_else, bool affine) :
4008 pet_skip_info(ctx), scop_then(scop_then), scop_else(scop_else),
4009 have_else(have_else)
4011 skip[pet_skip_now] =
4012 need_skip(scop_then, scop_else, have_else, pet_skip_now, affine);
4013 equal = skip[pet_skip_now] && skip_equals_skip_later(scop_then) &&
4014 (!have_else || skip_equals_skip_later(scop_else));
4015 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
4016 need_skip(scop_then, scop_else, have_else, pet_skip_later, affine);
4019 /* If we need to construct a skip condition of the given type,
4020 * then do so now.
4022 * "map" represents the if condition.
4024 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_map *map,
4025 enum pet_skip type)
4027 if (!skip[type])
4028 return;
4030 access[type] = create_test_access(isl_map_get_ctx(map), scan->n_test++);
4031 scop[type] = extract_skip(scan, isl_map_copy(map),
4032 isl_map_copy(access[type]),
4033 scop_then, scop_else, have_else, type);
4036 /* Construct the required skip conditions, given the if condition "map".
4038 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_map *map)
4040 extract(scan, map, pet_skip_now);
4041 extract(scan, map, pet_skip_later);
4042 if (equal)
4043 drop_skip_later(scop_then, scop_else);
4046 /* Construct the required skip conditions, given the if condition "cond".
4048 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_pw_aff *cond)
4050 isl_set *test_set;
4051 isl_map *test;
4053 if (!skip[pet_skip_now] && !skip[pet_skip_later])
4054 return;
4056 test_set = isl_set_from_pw_aff(isl_pw_aff_copy(cond));
4057 test = isl_map_from_range(test_set);
4058 extract(scan, test);
4059 isl_map_free(test);
4062 /* Add the computed skip condition of the give type to "main" and
4063 * add the scop for computing the condition at the given offset.
4065 * If equal is set, then we only computed a skip condition for pet_skip_now,
4066 * but we also need to set it as main's pet_skip_later.
4068 struct pet_scop *pet_skip_info_if::add(struct pet_scop *main,
4069 enum pet_skip type, int offset)
4071 isl_set *skip_set;
4073 if (!skip[type])
4074 return main;
4076 skip_set = isl_map_range(access[type]);
4077 access[type] = NULL;
4078 scop[type] = pet_scop_prefix(scop[type], offset);
4079 main = pet_scop_add_par(ctx, main, scop[type]);
4080 scop[type] = NULL;
4082 if (equal)
4083 main = pet_scop_set_skip(main, pet_skip_later,
4084 isl_set_copy(skip_set));
4086 main = pet_scop_set_skip(main, type, skip_set);
4088 return main;
4091 /* Add the computed skip conditions to "main" and
4092 * add the scops for computing the conditions at the given offset.
4094 struct pet_scop *pet_skip_info_if::add(struct pet_scop *scop, int offset)
4096 scop = add(scop, pet_skip_now, offset);
4097 scop = add(scop, pet_skip_later, offset);
4099 return scop;
4102 /* Construct a pet_scop for a non-affine if statement.
4104 * We create a separate statement that writes the result
4105 * of the non-affine condition to a virtual scalar.
4106 * A constraint requiring the value of this virtual scalar to be one
4107 * is added to the iteration domains of the then branch.
4108 * Similarly, a constraint requiring the value of this virtual scalar
4109 * to be zero is added to the iteration domains of the else branch, if any.
4110 * We adjust the schedules to ensure that the virtual scalar is written
4111 * before it is read.
4113 * If there are any breaks or continues in the then and/or else
4114 * branches, then we may have to compute a new skip condition.
4115 * This is handled using a pet_skip_info_if object.
4116 * On initialization, the object checks if skip conditions need
4117 * to be computed. If so, it does so in "extract" and adds them in "add".
4119 struct pet_scop *PetScan::extract_non_affine_if(Expr *cond,
4120 struct pet_scop *scop_then, struct pet_scop *scop_else,
4121 bool have_else, int stmt_id)
4123 struct pet_scop *scop;
4124 isl_map *test_access;
4125 int save_n_stmt = n_stmt;
4127 test_access = create_test_access(ctx, n_test++);
4128 n_stmt = stmt_id;
4129 scop = extract_non_affine_condition(cond, isl_map_copy(test_access));
4130 n_stmt = save_n_stmt;
4131 scop = scop_add_array(scop, test_access, ast_context);
4133 pet_skip_info_if skip(ctx, scop_then, scop_else, have_else, false);
4134 skip.extract(this, test_access);
4136 scop = pet_scop_prefix(scop, 0);
4137 scop_then = pet_scop_prefix(scop_then, 1);
4138 scop_then = pet_scop_filter(scop_then, isl_map_copy(test_access), 1);
4139 if (have_else) {
4140 scop_else = pet_scop_prefix(scop_else, 1);
4141 scop_else = pet_scop_filter(scop_else, test_access, 0);
4142 scop_then = pet_scop_add_par(ctx, scop_then, scop_else);
4143 } else
4144 isl_map_free(test_access);
4146 scop = pet_scop_add_seq(ctx, scop, scop_then);
4148 scop = skip.add(scop, 2);
4150 return scop;
4153 /* Construct a pet_scop for an if statement.
4155 * If the condition fits the pattern of a conditional assignment,
4156 * then it is handled by extract_conditional_assignment.
4157 * Otherwise, we do the following.
4159 * If the condition is affine, then the condition is added
4160 * to the iteration domains of the then branch, while the
4161 * opposite of the condition in added to the iteration domains
4162 * of the else branch, if any.
4163 * We allow the condition to be dynamic, i.e., to refer to
4164 * scalars or array elements that may be written to outside
4165 * of the given if statement. These nested accesses are then represented
4166 * as output dimensions in the wrapping iteration domain.
4167 * If it also written _inside_ the then or else branch, then
4168 * we treat the condition as non-affine.
4169 * As explained in extract_non_affine_if, this will introduce
4170 * an extra statement.
4171 * For aesthetic reasons, we want this statement to have a statement
4172 * number that is lower than those of the then and else branches.
4173 * In order to evaluate if will need such a statement, however, we
4174 * first construct scops for the then and else branches.
4175 * We therefore reserve a statement number if we might have to
4176 * introduce such an extra statement.
4178 * If the condition is not affine, then the scop is created in
4179 * extract_non_affine_if.
4181 * If there are any breaks or continues in the then and/or else
4182 * branches, then we may have to compute a new skip condition.
4183 * This is handled using a pet_skip_info_if object.
4184 * On initialization, the object checks if skip conditions need
4185 * to be computed. If so, it does so in "extract" and adds them in "add".
4187 struct pet_scop *PetScan::extract(IfStmt *stmt)
4189 struct pet_scop *scop_then, *scop_else = NULL, *scop;
4190 isl_pw_aff *cond;
4191 int stmt_id;
4192 isl_set *set;
4193 isl_set *valid;
4195 scop = extract_conditional_assignment(stmt);
4196 if (scop)
4197 return scop;
4199 cond = try_extract_nested_condition(stmt->getCond());
4200 if (allow_nested && (!cond || has_nested(cond)))
4201 stmt_id = n_stmt++;
4204 assigned_value_cache cache(assigned_value);
4205 scop_then = extract(stmt->getThen());
4208 if (stmt->getElse()) {
4209 assigned_value_cache cache(assigned_value);
4210 scop_else = extract(stmt->getElse());
4211 if (options->autodetect) {
4212 if (scop_then && !scop_else) {
4213 partial = true;
4214 isl_pw_aff_free(cond);
4215 return scop_then;
4217 if (!scop_then && scop_else) {
4218 partial = true;
4219 isl_pw_aff_free(cond);
4220 return scop_else;
4225 if (cond &&
4226 (!is_nested_allowed(cond, scop_then) ||
4227 (stmt->getElse() && !is_nested_allowed(cond, scop_else)))) {
4228 isl_pw_aff_free(cond);
4229 cond = NULL;
4231 if (allow_nested && !cond)
4232 return extract_non_affine_if(stmt->getCond(), scop_then,
4233 scop_else, stmt->getElse(), stmt_id);
4235 if (!cond)
4236 cond = extract_condition(stmt->getCond());
4238 pet_skip_info_if skip(ctx, scop_then, scop_else, stmt->getElse(), true);
4239 skip.extract(this, cond);
4241 valid = isl_pw_aff_domain(isl_pw_aff_copy(cond));
4242 set = isl_pw_aff_non_zero_set(cond);
4243 scop = pet_scop_restrict(scop_then, isl_set_copy(set));
4245 if (stmt->getElse()) {
4246 set = isl_set_subtract(isl_set_copy(valid), set);
4247 scop_else = pet_scop_restrict(scop_else, set);
4248 scop = pet_scop_add_par(ctx, scop, scop_else);
4249 } else
4250 isl_set_free(set);
4251 scop = resolve_nested(scop);
4252 scop = pet_scop_restrict_context(scop, valid);
4254 if (skip)
4255 scop = pet_scop_prefix(scop, 0);
4256 scop = skip.add(scop, 1);
4258 return scop;
4261 /* Try and construct a pet_scop for a label statement.
4262 * We currently only allow labels on expression statements.
4264 struct pet_scop *PetScan::extract(LabelStmt *stmt)
4266 isl_id *label;
4267 Stmt *sub;
4269 sub = stmt->getSubStmt();
4270 if (!isa<Expr>(sub)) {
4271 unsupported(stmt);
4272 return NULL;
4275 label = isl_id_alloc(ctx, stmt->getName(), NULL);
4277 return extract(sub, extract_expr(cast<Expr>(sub)), label);
4280 /* Construct a pet_scop for a continue statement.
4282 * We simply create an empty scop with a universal pet_skip_now
4283 * skip condition. This skip condition will then be taken into
4284 * account by the enclosing loop construct, possibly after
4285 * being incorporated into outer skip conditions.
4287 struct pet_scop *PetScan::extract(ContinueStmt *stmt)
4289 pet_scop *scop;
4290 isl_space *space;
4291 isl_set *set;
4293 scop = pet_scop_empty(ctx);
4294 if (!scop)
4295 return NULL;
4297 space = isl_space_set_alloc(ctx, 0, 1);
4298 set = isl_set_universe(space);
4299 set = isl_set_fix_si(set, isl_dim_set, 0, 1);
4300 scop = pet_scop_set_skip(scop, pet_skip_now, set);
4302 return scop;
4305 /* Construct a pet_scop for a break statement.
4307 * We simply create an empty scop with both a universal pet_skip_now
4308 * skip condition and a universal pet_skip_later skip condition.
4309 * These skip conditions will then be taken into
4310 * account by the enclosing loop construct, possibly after
4311 * being incorporated into outer skip conditions.
4313 struct pet_scop *PetScan::extract(BreakStmt *stmt)
4315 pet_scop *scop;
4316 isl_space *space;
4317 isl_set *set;
4319 scop = pet_scop_empty(ctx);
4320 if (!scop)
4321 return NULL;
4323 space = isl_space_set_alloc(ctx, 0, 1);
4324 set = isl_set_universe(space);
4325 set = isl_set_fix_si(set, isl_dim_set, 0, 1);
4326 scop = pet_scop_set_skip(scop, pet_skip_now, isl_set_copy(set));
4327 scop = pet_scop_set_skip(scop, pet_skip_later, set);
4329 return scop;
4332 /* Try and construct a pet_scop corresponding to "stmt".
4334 * If "stmt" is a compound statement, then "skip_declarations"
4335 * indicates whether we should skip initial declarations in the
4336 * compound statement.
4338 * If the constructed pet_scop is not a (possibly) partial representation
4339 * of "stmt", we update start and end of the pet_scop to those of "stmt".
4340 * In particular, if skip_declarations, then we may have skipped declarations
4341 * inside "stmt" and so the pet_scop may not represent the entire "stmt".
4342 * Note that this function may be called with "stmt" referring to the entire
4343 * body of the function, including the outer braces. In such cases,
4344 * skip_declarations will be set and the braces will not be taken into
4345 * account in scop->start and scop->end.
4347 struct pet_scop *PetScan::extract(Stmt *stmt, bool skip_declarations)
4349 struct pet_scop *scop;
4350 unsigned start, end;
4351 SourceLocation loc;
4352 SourceManager &SM = PP.getSourceManager();
4354 if (isa<Expr>(stmt))
4355 return extract(stmt, extract_expr(cast<Expr>(stmt)));
4357 switch (stmt->getStmtClass()) {
4358 case Stmt::WhileStmtClass:
4359 scop = extract(cast<WhileStmt>(stmt));
4360 break;
4361 case Stmt::ForStmtClass:
4362 scop = extract_for(cast<ForStmt>(stmt));
4363 break;
4364 case Stmt::IfStmtClass:
4365 scop = extract(cast<IfStmt>(stmt));
4366 break;
4367 case Stmt::CompoundStmtClass:
4368 scop = extract(cast<CompoundStmt>(stmt), skip_declarations);
4369 break;
4370 case Stmt::LabelStmtClass:
4371 scop = extract(cast<LabelStmt>(stmt));
4372 break;
4373 case Stmt::ContinueStmtClass:
4374 scop = extract(cast<ContinueStmt>(stmt));
4375 break;
4376 case Stmt::BreakStmtClass:
4377 scop = extract(cast<BreakStmt>(stmt));
4378 break;
4379 case Stmt::DeclStmtClass:
4380 scop = extract(cast<DeclStmt>(stmt));
4381 break;
4382 default:
4383 unsupported(stmt);
4384 return NULL;
4387 if (partial || skip_declarations)
4388 return scop;
4390 start = getExpansionOffset(SM, stmt->getLocStart());
4391 loc = PP.getLocForEndOfToken(stmt->getLocEnd());
4392 end = getExpansionOffset(SM, loc);
4393 scop = pet_scop_update_start_end(scop, start, end);
4395 return scop;
4398 /* Do we need to construct a skip condition of the given type
4399 * on a sequence of statements?
4401 * There is no need to construct a new skip condition if only
4402 * only of the two statements has a skip condition or if both
4403 * of their skip conditions are affine.
4405 * In principle we also don't need a new continuation variable if
4406 * the continuation of scop2 is affine, but then we would need
4407 * to allow more complicated forms of continuations.
4409 static bool need_skip_seq(struct pet_scop *scop1, struct pet_scop *scop2,
4410 enum pet_skip type)
4412 if (!scop1 || !pet_scop_has_skip(scop1, type))
4413 return false;
4414 if (!scop2 || !pet_scop_has_skip(scop2, type))
4415 return false;
4416 if (pet_scop_has_affine_skip(scop1, type) &&
4417 pet_scop_has_affine_skip(scop2, type))
4418 return false;
4419 return true;
4422 /* Construct a scop for computing the skip condition of the given type and
4423 * with access relation "skip_access" for a sequence of two scops "scop1"
4424 * and "scop2".
4426 * The computed scop contains a single statement that essentially does
4428 * skip_cond = skip_cond_1 ? 1 : skip_cond_2
4430 * or, in other words, skip_cond1 || skip_cond2.
4431 * In this expression, skip_cond_2 is filtered to reflect that it is
4432 * only evaluated when skip_cond_1 is false.
4434 * The skip condition on scop1 is not removed because it still needs
4435 * to be applied to scop2 when these two scops are combined.
4437 static struct pet_scop *extract_skip_seq(PetScan *ps,
4438 __isl_take isl_map *skip_access,
4439 struct pet_scop *scop1, struct pet_scop *scop2, enum pet_skip type)
4441 isl_map *access;
4442 struct pet_expr *expr1, *expr2, *expr, *expr_skip;
4443 struct pet_stmt *stmt;
4444 struct pet_scop *scop;
4445 isl_ctx *ctx = ps->ctx;
4447 if (!scop1 || !scop2)
4448 goto error;
4450 expr1 = pet_scop_get_skip_expr(scop1, type);
4451 expr2 = pet_scop_get_skip_expr(scop2, type);
4452 pet_scop_reset_skip(scop2, type);
4454 expr2 = pet_expr_filter(expr2, isl_map_copy(expr1->acc.access), 0);
4456 expr = universally_true(ctx);
4457 expr = pet_expr_new_ternary(ctx, expr1, expr, expr2);
4458 expr_skip = pet_expr_from_access(isl_map_copy(skip_access));
4459 if (expr_skip) {
4460 expr_skip->acc.write = 1;
4461 expr_skip->acc.read = 0;
4463 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
4464 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, ps->n_stmt++, expr);
4466 scop = pet_scop_from_pet_stmt(ctx, stmt);
4467 scop = scop_add_array(scop, skip_access, ps->ast_context);
4468 isl_map_free(skip_access);
4470 return scop;
4471 error:
4472 isl_map_free(skip_access);
4473 return NULL;
4476 /* Structure that handles the construction of skip conditions
4477 * on sequences of statements.
4479 * scop1 and scop2 represent the two statements that are combined
4481 struct pet_skip_info_seq : public pet_skip_info {
4482 struct pet_scop *scop1, *scop2;
4484 pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4485 struct pet_scop *scop2);
4486 void extract(PetScan *scan, enum pet_skip type);
4487 void extract(PetScan *scan);
4488 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
4489 int offset);
4490 struct pet_scop *add(struct pet_scop *scop, int offset);
4493 /* Initialize a pet_skip_info_seq structure based on
4494 * on the two statements that are going to be combined.
4496 pet_skip_info_seq::pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4497 struct pet_scop *scop2) : pet_skip_info(ctx), scop1(scop1), scop2(scop2)
4499 skip[pet_skip_now] = need_skip_seq(scop1, scop2, pet_skip_now);
4500 equal = skip[pet_skip_now] && skip_equals_skip_later(scop1) &&
4501 skip_equals_skip_later(scop2);
4502 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
4503 need_skip_seq(scop1, scop2, pet_skip_later);
4506 /* If we need to construct a skip condition of the given type,
4507 * then do so now.
4509 void pet_skip_info_seq::extract(PetScan *scan, enum pet_skip type)
4511 if (!skip[type])
4512 return;
4514 access[type] = create_test_access(ctx, scan->n_test++);
4515 scop[type] = extract_skip_seq(scan, isl_map_copy(access[type]),
4516 scop1, scop2, type);
4519 /* Construct the required skip conditions.
4521 void pet_skip_info_seq::extract(PetScan *scan)
4523 extract(scan, pet_skip_now);
4524 extract(scan, pet_skip_later);
4525 if (equal)
4526 drop_skip_later(scop1, scop2);
4529 /* Add the computed skip condition of the give type to "main" and
4530 * add the scop for computing the condition at the given offset (the statement
4531 * number). Within this offset, the condition is computed at position 1
4532 * to ensure that it is computed after the corresponding statement.
4534 * If equal is set, then we only computed a skip condition for pet_skip_now,
4535 * but we also need to set it as main's pet_skip_later.
4537 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *main,
4538 enum pet_skip type, int offset)
4540 isl_set *skip_set;
4542 if (!skip[type])
4543 return main;
4545 skip_set = isl_map_range(access[type]);
4546 access[type] = NULL;
4547 scop[type] = pet_scop_prefix(scop[type], 1);
4548 scop[type] = pet_scop_prefix(scop[type], offset);
4549 main = pet_scop_add_par(ctx, main, scop[type]);
4550 scop[type] = NULL;
4552 if (equal)
4553 main = pet_scop_set_skip(main, pet_skip_later,
4554 isl_set_copy(skip_set));
4556 main = pet_scop_set_skip(main, type, skip_set);
4558 return main;
4561 /* Add the computed skip conditions to "main" and
4562 * add the scops for computing the conditions at the given offset.
4564 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *scop, int offset)
4566 scop = add(scop, pet_skip_now, offset);
4567 scop = add(scop, pet_skip_later, offset);
4569 return scop;
4572 /* Extract a clone of the kill statement in "scop".
4573 * "scop" is expected to have been created from a DeclStmt
4574 * and should have the kill as its first statement.
4576 struct pet_stmt *PetScan::extract_kill(struct pet_scop *scop)
4578 struct pet_expr *kill;
4579 struct pet_stmt *stmt;
4580 isl_map *access;
4582 if (!scop)
4583 return NULL;
4584 if (scop->n_stmt < 1)
4585 isl_die(ctx, isl_error_internal,
4586 "expecting at least one statement", return NULL);
4587 stmt = scop->stmts[0];
4588 if (stmt->body->type != pet_expr_unary ||
4589 stmt->body->op != pet_op_kill)
4590 isl_die(ctx, isl_error_internal,
4591 "expecting kill statement", return NULL);
4593 access = isl_map_copy(stmt->body->args[0]->acc.access);
4594 access = isl_map_reset_tuple_id(access, isl_dim_in);
4595 kill = pet_expr_kill_from_access(access);
4596 return pet_stmt_from_pet_expr(ctx, stmt->line, NULL, n_stmt++, kill);
4599 /* Mark all arrays in "scop" as being exposed.
4601 static struct pet_scop *mark_exposed(struct pet_scop *scop)
4603 if (!scop)
4604 return NULL;
4605 for (int i = 0; i < scop->n_array; ++i)
4606 scop->arrays[i]->exposed = 1;
4607 return scop;
4610 /* Try and construct a pet_scop corresponding to (part of)
4611 * a sequence of statements.
4613 * "block" is set if the sequence respresents the children of
4614 * a compound statement.
4615 * "skip_declarations" is set if we should skip initial declarations
4616 * in the sequence of statements.
4618 * If there are any breaks or continues in the individual statements,
4619 * then we may have to compute a new skip condition.
4620 * This is handled using a pet_skip_info_seq object.
4621 * On initialization, the object checks if skip conditions need
4622 * to be computed. If so, it does so in "extract" and adds them in "add".
4624 * If "block" is set, then we need to insert kill statements at
4625 * the end of the block for any array that has been declared by
4626 * one of the statements in the sequence. Each of these declarations
4627 * results in the construction of a kill statement at the place
4628 * of the declaration, so we simply collect duplicates of
4629 * those kill statements and append these duplicates to the constructed scop.
4631 * If "block" is not set, then any array declared by one of the statements
4632 * in the sequence is marked as being exposed.
4634 struct pet_scop *PetScan::extract(StmtRange stmt_range, bool block,
4635 bool skip_declarations)
4637 pet_scop *scop;
4638 StmtIterator i;
4639 int j;
4640 bool partial_range = false;
4641 set<struct pet_stmt *> kills;
4642 set<struct pet_stmt *>::iterator it;
4644 scop = pet_scop_empty(ctx);
4645 for (i = stmt_range.first, j = 0; i != stmt_range.second; ++i, ++j) {
4646 Stmt *child = *i;
4647 struct pet_scop *scop_i;
4649 if (skip_declarations &&
4650 child->getStmtClass() == Stmt::DeclStmtClass)
4651 continue;
4653 scop_i = extract(child);
4654 if (scop && partial) {
4655 pet_scop_free(scop_i);
4656 break;
4658 pet_skip_info_seq skip(ctx, scop, scop_i);
4659 skip.extract(this);
4660 if (skip)
4661 scop_i = pet_scop_prefix(scop_i, 0);
4662 if (scop_i && child->getStmtClass() == Stmt::DeclStmtClass) {
4663 if (block)
4664 kills.insert(extract_kill(scop_i));
4665 else
4666 scop_i = mark_exposed(scop_i);
4668 scop_i = pet_scop_prefix(scop_i, j);
4669 if (options->autodetect) {
4670 if (scop_i)
4671 scop = pet_scop_add_seq(ctx, scop, scop_i);
4672 else
4673 partial_range = true;
4674 if (scop->n_stmt != 0 && !scop_i)
4675 partial = true;
4676 } else {
4677 scop = pet_scop_add_seq(ctx, scop, scop_i);
4680 scop = skip.add(scop, j);
4682 if (partial)
4683 break;
4686 for (it = kills.begin(); it != kills.end(); ++it) {
4687 pet_scop *scop_j;
4688 scop_j = pet_scop_from_pet_stmt(ctx, *it);
4689 scop_j = pet_scop_prefix(scop_j, j);
4690 scop = pet_scop_add_seq(ctx, scop, scop_j);
4693 if (scop && partial_range)
4694 partial = true;
4696 return scop;
4699 /* Check if the scop marked by the user is exactly this Stmt
4700 * or part of this Stmt.
4701 * If so, return a pet_scop corresponding to the marked region.
4702 * Otherwise, return NULL.
4704 struct pet_scop *PetScan::scan(Stmt *stmt)
4706 SourceManager &SM = PP.getSourceManager();
4707 unsigned start_off, end_off;
4709 start_off = getExpansionOffset(SM, stmt->getLocStart());
4710 end_off = getExpansionOffset(SM, stmt->getLocEnd());
4712 if (start_off > loc.end)
4713 return NULL;
4714 if (end_off < loc.start)
4715 return NULL;
4716 if (start_off >= loc.start && end_off <= loc.end) {
4717 return extract(stmt);
4720 StmtIterator start;
4721 for (start = stmt->child_begin(); start != stmt->child_end(); ++start) {
4722 Stmt *child = *start;
4723 if (!child)
4724 continue;
4725 start_off = getExpansionOffset(SM, child->getLocStart());
4726 end_off = getExpansionOffset(SM, child->getLocEnd());
4727 if (start_off < loc.start && end_off > loc.end)
4728 return scan(child);
4729 if (start_off >= loc.start)
4730 break;
4733 StmtIterator end;
4734 for (end = start; end != stmt->child_end(); ++end) {
4735 Stmt *child = *end;
4736 start_off = SM.getFileOffset(child->getLocStart());
4737 if (start_off >= loc.end)
4738 break;
4741 return extract(StmtRange(start, end), false, false);
4744 /* Set the size of index "pos" of "array" to "size".
4745 * In particular, add a constraint of the form
4747 * i_pos < size
4749 * to array->extent and a constraint of the form
4751 * size >= 0
4753 * to array->context.
4755 static struct pet_array *update_size(struct pet_array *array, int pos,
4756 __isl_take isl_pw_aff *size)
4758 isl_set *valid;
4759 isl_set *univ;
4760 isl_set *bound;
4761 isl_space *dim;
4762 isl_aff *aff;
4763 isl_pw_aff *index;
4764 isl_id *id;
4766 valid = isl_pw_aff_nonneg_set(isl_pw_aff_copy(size));
4767 array->context = isl_set_intersect(array->context, valid);
4769 dim = isl_set_get_space(array->extent);
4770 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
4771 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, pos, 1);
4772 univ = isl_set_universe(isl_aff_get_domain_space(aff));
4773 index = isl_pw_aff_alloc(univ, aff);
4775 size = isl_pw_aff_add_dims(size, isl_dim_in,
4776 isl_set_dim(array->extent, isl_dim_set));
4777 id = isl_set_get_tuple_id(array->extent);
4778 size = isl_pw_aff_set_tuple_id(size, isl_dim_in, id);
4779 bound = isl_pw_aff_lt_set(index, size);
4781 array->extent = isl_set_intersect(array->extent, bound);
4783 if (!array->context || !array->extent)
4784 goto error;
4786 return array;
4787 error:
4788 pet_array_free(array);
4789 return NULL;
4792 /* Figure out the size of the array at position "pos" and all
4793 * subsequent positions from "type" and update "array" accordingly.
4795 struct pet_array *PetScan::set_upper_bounds(struct pet_array *array,
4796 const Type *type, int pos)
4798 const ArrayType *atype;
4799 isl_pw_aff *size;
4801 if (!array)
4802 return NULL;
4804 if (type->isPointerType()) {
4805 type = type->getPointeeType().getTypePtr();
4806 return set_upper_bounds(array, type, pos + 1);
4808 if (!type->isArrayType())
4809 return array;
4811 type = type->getCanonicalTypeInternal().getTypePtr();
4812 atype = cast<ArrayType>(type);
4814 if (type->isConstantArrayType()) {
4815 const ConstantArrayType *ca = cast<ConstantArrayType>(atype);
4816 size = extract_affine(ca->getSize());
4817 array = update_size(array, pos, size);
4818 } else if (type->isVariableArrayType()) {
4819 const VariableArrayType *vla = cast<VariableArrayType>(atype);
4820 size = extract_affine(vla->getSizeExpr());
4821 array = update_size(array, pos, size);
4824 type = atype->getElementType().getTypePtr();
4826 return set_upper_bounds(array, type, pos + 1);
4829 /* Is "T" the type of a variable length array with static size?
4831 static bool is_vla_with_static_size(QualType T)
4833 const VariableArrayType *vlatype;
4835 if (!T->isVariableArrayType())
4836 return false;
4837 vlatype = cast<VariableArrayType>(T);
4838 return vlatype->getSizeModifier() == VariableArrayType::Static;
4841 /* Return the type of "decl" as an array.
4843 * In particular, if "decl" is a parameter declaration that
4844 * is a variable length array with a static size, then
4845 * return the original type (i.e., the variable length array).
4846 * Otherwise, return the type of decl.
4848 static QualType get_array_type(ValueDecl *decl)
4850 ParmVarDecl *parm;
4851 QualType T;
4853 parm = dyn_cast<ParmVarDecl>(decl);
4854 if (!parm)
4855 return decl->getType();
4857 T = parm->getOriginalType();
4858 if (!is_vla_with_static_size(T))
4859 return decl->getType();
4860 return T;
4863 /* Construct and return a pet_array corresponding to the variable "decl".
4864 * In particular, initialize array->extent to
4866 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
4868 * and then call set_upper_bounds to set the upper bounds on the indices
4869 * based on the type of the variable.
4871 struct pet_array *PetScan::extract_array(isl_ctx *ctx, ValueDecl *decl)
4873 struct pet_array *array;
4874 QualType qt = get_array_type(decl);
4875 const Type *type = qt.getTypePtr();
4876 int depth = array_depth(type);
4877 QualType base = base_type(qt);
4878 string name;
4879 isl_id *id;
4880 isl_space *dim;
4882 array = isl_calloc_type(ctx, struct pet_array);
4883 if (!array)
4884 return NULL;
4886 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
4887 dim = isl_space_set_alloc(ctx, 0, depth);
4888 dim = isl_space_set_tuple_id(dim, isl_dim_set, id);
4890 array->extent = isl_set_nat_universe(dim);
4892 dim = isl_space_params_alloc(ctx, 0);
4893 array->context = isl_set_universe(dim);
4895 array = set_upper_bounds(array, type, 0);
4896 if (!array)
4897 return NULL;
4899 name = base.getAsString();
4900 array->element_type = strdup(name.c_str());
4901 array->element_size = decl->getASTContext().getTypeInfo(base).first / 8;
4903 return array;
4906 /* Construct a list of pet_arrays, one for each array (or scalar)
4907 * accessed inside "scop", add this list to "scop" and return the result.
4909 * The context of "scop" is updated with the intersection of
4910 * the contexts of all arrays, i.e., constraints on the parameters
4911 * that ensure that the arrays have a valid (non-negative) size.
4913 struct pet_scop *PetScan::scan_arrays(struct pet_scop *scop)
4915 int i;
4916 set<ValueDecl *> arrays;
4917 set<ValueDecl *>::iterator it;
4918 int n_array;
4919 struct pet_array **scop_arrays;
4921 if (!scop)
4922 return NULL;
4924 pet_scop_collect_arrays(scop, arrays);
4925 if (arrays.size() == 0)
4926 return scop;
4928 n_array = scop->n_array;
4930 scop_arrays = isl_realloc_array(ctx, scop->arrays, struct pet_array *,
4931 n_array + arrays.size());
4932 if (!scop_arrays)
4933 goto error;
4934 scop->arrays = scop_arrays;
4936 for (it = arrays.begin(), i = 0; it != arrays.end(); ++it, ++i) {
4937 struct pet_array *array;
4938 scop->arrays[n_array + i] = array = extract_array(ctx, *it);
4939 if (!scop->arrays[n_array + i])
4940 goto error;
4941 scop->n_array++;
4942 scop->context = isl_set_intersect(scop->context,
4943 isl_set_copy(array->context));
4944 if (!scop->context)
4945 goto error;
4948 return scop;
4949 error:
4950 pet_scop_free(scop);
4951 return NULL;
4954 /* Bound all parameters in scop->context to the possible values
4955 * of the corresponding C variable.
4957 static struct pet_scop *add_parameter_bounds(struct pet_scop *scop)
4959 int n;
4961 if (!scop)
4962 return NULL;
4964 n = isl_set_dim(scop->context, isl_dim_param);
4965 for (int i = 0; i < n; ++i) {
4966 isl_id *id;
4967 ValueDecl *decl;
4969 id = isl_set_get_dim_id(scop->context, isl_dim_param, i);
4970 if (is_nested_parameter(id)) {
4971 isl_id_free(id);
4972 isl_die(isl_set_get_ctx(scop->context),
4973 isl_error_internal,
4974 "unresolved nested parameter", goto error);
4976 decl = (ValueDecl *) isl_id_get_user(id);
4977 isl_id_free(id);
4979 scop->context = set_parameter_bounds(scop->context, i, decl);
4981 if (!scop->context)
4982 goto error;
4985 return scop;
4986 error:
4987 pet_scop_free(scop);
4988 return NULL;
4991 /* Construct a pet_scop from the given function.
4993 * If the scop was delimited by scop and endscop pragmas, then we override
4994 * the file offsets by those derived from the pragmas.
4996 struct pet_scop *PetScan::scan(FunctionDecl *fd)
4998 pet_scop *scop;
4999 Stmt *stmt;
5001 stmt = fd->getBody();
5003 if (options->autodetect)
5004 scop = extract(stmt, true);
5005 else {
5006 scop = scan(stmt);
5007 scop = pet_scop_update_start_end(scop, loc.start, loc.end);
5009 scop = pet_scop_detect_parameter_accesses(scop);
5010 scop = scan_arrays(scop);
5011 scop = add_parameter_bounds(scop);
5012 scop = pet_scop_gist(scop, value_bounds);
5014 return scop;