scop.c: expr_extract_context: don't assume access is an affine expression
[pet.git] / scan.cc
blob9a8e9890441ddb787f07570a8cb648be13c68cf5
1 /*
2 * Copyright 2011 Leiden University. All rights reserved.
3 * Copyright 2012 Ecole Normale Superieure. All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 *
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above
13 * copyright notice, this list of conditions and the following
14 * disclaimer in the documentation and/or other materials provided
15 * with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY LEIDEN UNIVERSITY ''AS IS'' AND ANY
18 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL LEIDEN UNIVERSITY OR
21 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
22 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
23 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA,
24 * OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
27 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 * The views and conclusions contained in the software and documentation
30 * are those of the authors and should not be interpreted as
31 * representing official policies, either expressed or implied, of
32 * Leiden University.
33 */
35 #include <set>
36 #include <map>
37 #include <iostream>
38 #include <clang/AST/ASTDiagnostic.h>
39 #include <clang/AST/Expr.h>
40 #include <clang/AST/RecursiveASTVisitor.h>
42 #include <isl/id.h>
43 #include <isl/space.h>
44 #include <isl/aff.h>
45 #include <isl/set.h>
47 #include "scan.h"
48 #include "scop.h"
49 #include "scop_plus.h"
51 #include "config.h"
53 using namespace std;
54 using namespace clang;
56 #if defined(DECLREFEXPR_CREATE_REQUIRES_BOOL)
57 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
59 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
60 SourceLocation(), var, false, var->getInnerLocStart(),
61 var->getType(), VK_LValue);
63 #elif defined(DECLREFEXPR_CREATE_REQUIRES_SOURCELOCATION)
64 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
66 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
67 SourceLocation(), var, var->getInnerLocStart(), var->getType(),
68 VK_LValue);
70 #else
71 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
73 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
74 var, var->getInnerLocStart(), var->getType(), VK_LValue);
76 #endif
78 /* Check if the element type corresponding to the given array type
79 * has a const qualifier.
81 static bool const_base(QualType qt)
83 const Type *type = qt.getTypePtr();
85 if (type->isPointerType())
86 return const_base(type->getPointeeType());
87 if (type->isArrayType()) {
88 const ArrayType *atype;
89 type = type->getCanonicalTypeInternal().getTypePtr();
90 atype = cast<ArrayType>(type);
91 return const_base(atype->getElementType());
94 return qt.isConstQualified();
97 /* Mark "decl" as having an unknown value in "assigned_value".
99 * If no (known or unknown) value was assigned to "decl" before,
100 * then it may have been treated as a parameter before and may
101 * therefore appear in a value assigned to another variable.
102 * If so, this assignment needs to be turned into an unknown value too.
104 static void clear_assignment(map<ValueDecl *, isl_pw_aff *> &assigned_value,
105 ValueDecl *decl)
107 map<ValueDecl *, isl_pw_aff *>::iterator it;
109 it = assigned_value.find(decl);
111 assigned_value[decl] = NULL;
113 if (it == assigned_value.end())
114 return;
116 for (it = assigned_value.begin(); it != assigned_value.end(); ++it) {
117 isl_pw_aff *pa = it->second;
118 int nparam = isl_pw_aff_dim(pa, isl_dim_param);
120 for (int i = 0; i < nparam; ++i) {
121 isl_id *id;
123 if (!isl_pw_aff_has_dim_id(pa, isl_dim_param, i))
124 continue;
125 id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
126 if (isl_id_get_user(id) == decl)
127 it->second = NULL;
128 isl_id_free(id);
133 /* Look for any assignments to scalar variables in part of the parse
134 * tree and set assigned_value to NULL for each of them.
135 * Also reset assigned_value if the address of a scalar variable
136 * is being taken. As an exception, if the address is passed to a function
137 * that is declared to receive a const pointer, then assigned_value is
138 * not reset.
140 * This ensures that we won't use any previously stored value
141 * in the current subtree and its parents.
143 struct clear_assignments : RecursiveASTVisitor<clear_assignments> {
144 map<ValueDecl *, isl_pw_aff *> &assigned_value;
145 set<UnaryOperator *> skip;
147 clear_assignments(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
148 assigned_value(assigned_value) {}
150 /* Check for "address of" operators whose value is passed
151 * to a const pointer argument and add them to "skip", so that
152 * we can skip them in VisitUnaryOperator.
154 bool VisitCallExpr(CallExpr *expr) {
155 FunctionDecl *fd;
156 fd = expr->getDirectCallee();
157 if (!fd)
158 return true;
159 for (int i = 0; i < expr->getNumArgs(); ++i) {
160 Expr *arg = expr->getArg(i);
161 UnaryOperator *op;
162 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
163 ImplicitCastExpr *ice;
164 ice = cast<ImplicitCastExpr>(arg);
165 arg = ice->getSubExpr();
167 if (arg->getStmtClass() != Stmt::UnaryOperatorClass)
168 continue;
169 op = cast<UnaryOperator>(arg);
170 if (op->getOpcode() != UO_AddrOf)
171 continue;
172 if (const_base(fd->getParamDecl(i)->getType()))
173 skip.insert(op);
175 return true;
178 bool VisitUnaryOperator(UnaryOperator *expr) {
179 Expr *arg;
180 DeclRefExpr *ref;
181 ValueDecl *decl;
183 if (expr->getOpcode() != UO_AddrOf)
184 return true;
185 if (skip.find(expr) != skip.end())
186 return true;
188 arg = expr->getSubExpr();
189 if (arg->getStmtClass() != Stmt::DeclRefExprClass)
190 return true;
191 ref = cast<DeclRefExpr>(arg);
192 decl = ref->getDecl();
193 clear_assignment(assigned_value, decl);
194 return true;
197 bool VisitBinaryOperator(BinaryOperator *expr) {
198 Expr *lhs;
199 DeclRefExpr *ref;
200 ValueDecl *decl;
202 if (!expr->isAssignmentOp())
203 return true;
204 lhs = expr->getLHS();
205 if (lhs->getStmtClass() != Stmt::DeclRefExprClass)
206 return true;
207 ref = cast<DeclRefExpr>(lhs);
208 decl = ref->getDecl();
209 clear_assignment(assigned_value, decl);
210 return true;
214 /* Keep a copy of the currently assigned values.
216 * Any variable that is assigned a value inside the current scope
217 * is removed again when we leave the scope (either because it wasn't
218 * stored in the cache or because it has a different value in the cache).
220 struct assigned_value_cache {
221 map<ValueDecl *, isl_pw_aff *> &assigned_value;
222 map<ValueDecl *, isl_pw_aff *> cache;
224 assigned_value_cache(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
225 assigned_value(assigned_value), cache(assigned_value) {}
226 ~assigned_value_cache() {
227 map<ValueDecl *, isl_pw_aff *>::iterator it = cache.begin();
228 for (it = assigned_value.begin(); it != assigned_value.end();
229 ++it) {
230 if (!it->second ||
231 (cache.find(it->first) != cache.end() &&
232 cache[it->first] != it->second))
233 cache[it->first] = NULL;
235 assigned_value = cache;
239 /* Insert an expression into the collection of expressions,
240 * provided it is not already in there.
241 * The isl_pw_affs are freed in the destructor.
243 void PetScan::insert_expression(__isl_take isl_pw_aff *expr)
245 std::set<isl_pw_aff *>::iterator it;
247 if (expressions.find(expr) == expressions.end())
248 expressions.insert(expr);
249 else
250 isl_pw_aff_free(expr);
253 PetScan::~PetScan()
255 std::set<isl_pw_aff *>::iterator it;
257 for (it = expressions.begin(); it != expressions.end(); ++it)
258 isl_pw_aff_free(*it);
260 isl_union_map_free(value_bounds);
263 /* Called if we found something we (currently) cannot handle.
264 * We'll provide more informative warnings later.
266 * We only actually complain if autodetect is false.
268 void PetScan::unsupported(Stmt *stmt, const char *msg)
270 if (autodetect)
271 return;
273 SourceLocation loc = stmt->getLocStart();
274 DiagnosticsEngine &diag = PP.getDiagnostics();
275 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
276 msg ? msg : "unsupported");
277 DiagnosticBuilder B = diag.Report(loc, id) << stmt->getSourceRange();
280 /* Extract an integer from "expr" and store it in "v".
282 int PetScan::extract_int(IntegerLiteral *expr, isl_int *v)
284 const Type *type = expr->getType().getTypePtr();
285 int is_signed = type->hasSignedIntegerRepresentation();
287 if (is_signed) {
288 int64_t i = expr->getValue().getSExtValue();
289 isl_int_set_si(*v, i);
290 } else {
291 uint64_t i = expr->getValue().getZExtValue();
292 isl_int_set_ui(*v, i);
295 return 0;
298 /* Extract an integer from "expr" and store it in "v".
299 * Return -1 if "expr" does not (obviously) represent an integer.
301 int PetScan::extract_int(clang::ParenExpr *expr, isl_int *v)
303 return extract_int(expr->getSubExpr(), v);
306 /* Extract an integer from "expr" and store it in "v".
307 * Return -1 if "expr" does not (obviously) represent an integer.
309 int PetScan::extract_int(clang::Expr *expr, isl_int *v)
311 if (expr->getStmtClass() == Stmt::IntegerLiteralClass)
312 return extract_int(cast<IntegerLiteral>(expr), v);
313 if (expr->getStmtClass() == Stmt::ParenExprClass)
314 return extract_int(cast<ParenExpr>(expr), v);
316 unsupported(expr);
317 return -1;
320 /* Extract an affine expression from the IntegerLiteral "expr".
322 __isl_give isl_pw_aff *PetScan::extract_affine(IntegerLiteral *expr)
324 isl_space *dim = isl_space_params_alloc(ctx, 0);
325 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
326 isl_aff *aff = isl_aff_zero_on_domain(ls);
327 isl_set *dom = isl_set_universe(dim);
328 isl_int v;
330 isl_int_init(v);
331 extract_int(expr, &v);
332 aff = isl_aff_add_constant(aff, v);
333 isl_int_clear(v);
335 return isl_pw_aff_alloc(dom, aff);
338 /* Extract an affine expression from the APInt "val".
340 __isl_give isl_pw_aff *PetScan::extract_affine(const llvm::APInt &val)
342 isl_space *dim = isl_space_params_alloc(ctx, 0);
343 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
344 isl_aff *aff = isl_aff_zero_on_domain(ls);
345 isl_set *dom = isl_set_universe(dim);
346 isl_int v;
348 isl_int_init(v);
349 isl_int_set_ui(v, val.getZExtValue());
350 aff = isl_aff_add_constant(aff, v);
351 isl_int_clear(v);
353 return isl_pw_aff_alloc(dom, aff);
356 __isl_give isl_pw_aff *PetScan::extract_affine(ImplicitCastExpr *expr)
358 return extract_affine(expr->getSubExpr());
361 static unsigned get_type_size(ValueDecl *decl)
363 return decl->getASTContext().getIntWidth(decl->getType());
366 /* Bound parameter "pos" of "set" to the possible values of "decl".
368 static __isl_give isl_set *set_parameter_bounds(__isl_take isl_set *set,
369 unsigned pos, ValueDecl *decl)
371 unsigned width;
372 isl_int v;
374 isl_int_init(v);
376 width = get_type_size(decl);
377 if (decl->getType()->isUnsignedIntegerType()) {
378 set = isl_set_lower_bound_si(set, isl_dim_param, pos, 0);
379 isl_int_set_si(v, 1);
380 isl_int_mul_2exp(v, v, width);
381 isl_int_sub_ui(v, v, 1);
382 set = isl_set_upper_bound(set, isl_dim_param, pos, v);
383 } else {
384 isl_int_set_si(v, 1);
385 isl_int_mul_2exp(v, v, width - 1);
386 isl_int_sub_ui(v, v, 1);
387 set = isl_set_upper_bound(set, isl_dim_param, pos, v);
388 isl_int_neg(v, v);
389 isl_int_sub_ui(v, v, 1);
390 set = isl_set_lower_bound(set, isl_dim_param, pos, v);
393 isl_int_clear(v);
395 return set;
398 /* Extract an affine expression from the DeclRefExpr "expr".
400 * If the variable has been assigned a value, then we check whether
401 * we know what (affine) value was assigned.
402 * If so, we return this value. Otherwise we convert "expr"
403 * to an extra parameter (provided nesting_enabled is set).
405 * Otherwise, we simply return an expression that is equal
406 * to a parameter corresponding to the referenced variable.
408 __isl_give isl_pw_aff *PetScan::extract_affine(DeclRefExpr *expr)
410 ValueDecl *decl = expr->getDecl();
411 const Type *type = decl->getType().getTypePtr();
412 isl_id *id;
413 isl_space *dim;
414 isl_aff *aff;
415 isl_set *dom;
417 if (!type->isIntegerType()) {
418 unsupported(expr);
419 return NULL;
422 if (assigned_value.find(decl) != assigned_value.end()) {
423 if (assigned_value[decl])
424 return isl_pw_aff_copy(assigned_value[decl]);
425 else
426 return nested_access(expr);
429 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
430 dim = isl_space_params_alloc(ctx, 1);
432 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
434 dom = isl_set_universe(isl_space_copy(dim));
435 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
436 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
438 return isl_pw_aff_alloc(dom, aff);
441 /* Extract an affine expression from an integer division operation.
442 * In particular, if "expr" is lhs/rhs, then return
444 * lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs)
446 * The second argument (rhs) is required to be a (positive) integer constant.
448 __isl_give isl_pw_aff *PetScan::extract_affine_div(BinaryOperator *expr)
450 Expr *rhs_expr;
451 isl_pw_aff *lhs, *lhs_f, *lhs_c;
452 isl_pw_aff *res;
453 isl_int v;
454 isl_set *cond;
456 rhs_expr = expr->getRHS();
457 isl_int_init(v);
458 if (extract_int(rhs_expr, &v) < 0) {
459 isl_int_clear(v);
460 return NULL;
463 lhs = extract_affine(expr->getLHS());
464 cond = isl_pw_aff_nonneg_set(isl_pw_aff_copy(lhs));
466 lhs = isl_pw_aff_scale_down(lhs, v);
467 isl_int_clear(v);
469 lhs_f = isl_pw_aff_floor(isl_pw_aff_copy(lhs));
470 lhs_c = isl_pw_aff_ceil(lhs);
471 res = isl_pw_aff_cond(isl_set_indicator_function(cond), lhs_f, lhs_c);
473 return res;
476 /* Extract an affine expression from a modulo operation.
477 * In particular, if "expr" is lhs/rhs, then return
479 * lhs - rhs * (lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs))
481 * The second argument (rhs) is required to be a (positive) integer constant.
483 __isl_give isl_pw_aff *PetScan::extract_affine_mod(BinaryOperator *expr)
485 Expr *rhs_expr;
486 isl_pw_aff *lhs, *lhs_f, *lhs_c;
487 isl_pw_aff *res;
488 isl_int v;
489 isl_set *cond;
491 rhs_expr = expr->getRHS();
492 if (rhs_expr->getStmtClass() != Stmt::IntegerLiteralClass) {
493 unsupported(expr);
494 return NULL;
497 lhs = extract_affine(expr->getLHS());
498 cond = isl_pw_aff_nonneg_set(isl_pw_aff_copy(lhs));
500 isl_int_init(v);
501 extract_int(cast<IntegerLiteral>(rhs_expr), &v);
502 res = isl_pw_aff_scale_down(isl_pw_aff_copy(lhs), v);
504 lhs_f = isl_pw_aff_floor(isl_pw_aff_copy(res));
505 lhs_c = isl_pw_aff_ceil(res);
506 res = isl_pw_aff_cond(isl_set_indicator_function(cond), lhs_f, lhs_c);
508 res = isl_pw_aff_scale(res, v);
509 isl_int_clear(v);
511 res = isl_pw_aff_sub(lhs, res);
513 return res;
516 /* Extract an affine expression from a multiplication operation.
517 * This is only allowed if at least one of the two arguments
518 * is a (piecewise) constant.
520 __isl_give isl_pw_aff *PetScan::extract_affine_mul(BinaryOperator *expr)
522 isl_pw_aff *lhs;
523 isl_pw_aff *rhs;
525 lhs = extract_affine(expr->getLHS());
526 rhs = extract_affine(expr->getRHS());
528 if (!isl_pw_aff_is_cst(lhs) && !isl_pw_aff_is_cst(rhs)) {
529 isl_pw_aff_free(lhs);
530 isl_pw_aff_free(rhs);
531 unsupported(expr);
532 return NULL;
535 return isl_pw_aff_mul(lhs, rhs);
538 /* Extract an affine expression from an addition or subtraction operation.
540 __isl_give isl_pw_aff *PetScan::extract_affine_add(BinaryOperator *expr)
542 isl_pw_aff *lhs;
543 isl_pw_aff *rhs;
545 lhs = extract_affine(expr->getLHS());
546 rhs = extract_affine(expr->getRHS());
548 switch (expr->getOpcode()) {
549 case BO_Add:
550 return isl_pw_aff_add(lhs, rhs);
551 case BO_Sub:
552 return isl_pw_aff_sub(lhs, rhs);
553 default:
554 isl_pw_aff_free(lhs);
555 isl_pw_aff_free(rhs);
556 return NULL;
561 /* Compute
563 * pwaff mod 2^width
565 static __isl_give isl_pw_aff *wrap(__isl_take isl_pw_aff *pwaff,
566 unsigned width)
568 isl_int mod;
570 isl_int_init(mod);
571 isl_int_set_si(mod, 1);
572 isl_int_mul_2exp(mod, mod, width);
574 pwaff = isl_pw_aff_mod(pwaff, mod);
576 isl_int_clear(mod);
578 return pwaff;
581 /* Limit the domain of "pwaff" to those elements where the function
582 * value satisfies
584 * 2^{width-1} <= pwaff < 2^{width-1}
586 static __isl_give isl_pw_aff *avoid_overflow(__isl_take isl_pw_aff *pwaff,
587 unsigned width)
589 isl_int v;
590 isl_space *space = isl_pw_aff_get_domain_space(pwaff);
591 isl_local_space *ls = isl_local_space_from_space(space);
592 isl_aff *bound;
593 isl_set *dom;
594 isl_pw_aff *b;
596 isl_int_init(v);
597 isl_int_set_si(v, 1);
598 isl_int_mul_2exp(v, v, width - 1);
600 bound = isl_aff_zero_on_domain(ls);
601 bound = isl_aff_add_constant(bound, v);
602 b = isl_pw_aff_from_aff(bound);
604 dom = isl_pw_aff_lt_set(isl_pw_aff_copy(pwaff), isl_pw_aff_copy(b));
605 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
607 b = isl_pw_aff_neg(b);
608 dom = isl_pw_aff_ge_set(isl_pw_aff_copy(pwaff), b);
609 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
611 isl_int_clear(v);
613 return pwaff;
616 /* Return the piecewise affine expression "set ? 1 : 0" defined on "dom".
618 static __isl_give isl_pw_aff *indicator_function(__isl_take isl_set *set,
619 __isl_take isl_set *dom)
621 isl_pw_aff *pa;
622 pa = isl_set_indicator_function(set);
623 pa = isl_pw_aff_intersect_domain(pa, dom);
624 return pa;
627 /* Extract an affine expression from some binary operations.
628 * If the result of the expression is unsigned, then we wrap it
629 * based on the size of the type. Otherwise, we ensure that
630 * no overflow occurs.
632 __isl_give isl_pw_aff *PetScan::extract_affine(BinaryOperator *expr)
634 isl_pw_aff *res;
635 unsigned width;
637 switch (expr->getOpcode()) {
638 case BO_Add:
639 case BO_Sub:
640 res = extract_affine_add(expr);
641 break;
642 case BO_Div:
643 res = extract_affine_div(expr);
644 break;
645 case BO_Rem:
646 res = extract_affine_mod(expr);
647 break;
648 case BO_Mul:
649 res = extract_affine_mul(expr);
650 break;
651 case BO_LT:
652 case BO_LE:
653 case BO_GT:
654 case BO_GE:
655 case BO_EQ:
656 case BO_NE:
657 case BO_LAnd:
658 case BO_LOr:
659 return extract_condition(expr);
660 default:
661 unsupported(expr);
662 return NULL;
665 width = ast_context.getIntWidth(expr->getType());
666 if (expr->getType()->isUnsignedIntegerType())
667 res = wrap(res, width);
668 else
669 res = avoid_overflow(res, width);
671 return res;
674 /* Extract an affine expression from a negation operation.
676 __isl_give isl_pw_aff *PetScan::extract_affine(UnaryOperator *expr)
678 if (expr->getOpcode() == UO_Minus)
679 return isl_pw_aff_neg(extract_affine(expr->getSubExpr()));
680 if (expr->getOpcode() == UO_LNot)
681 return extract_condition(expr);
683 unsupported(expr);
684 return NULL;
687 __isl_give isl_pw_aff *PetScan::extract_affine(ParenExpr *expr)
689 return extract_affine(expr->getSubExpr());
692 /* Extract an affine expression from some special function calls.
693 * In particular, we handle "min", "max", "ceild" and "floord".
694 * In case of the latter two, the second argument needs to be
695 * a (positive) integer constant.
697 __isl_give isl_pw_aff *PetScan::extract_affine(CallExpr *expr)
699 FunctionDecl *fd;
700 string name;
701 isl_pw_aff *aff1, *aff2;
703 fd = expr->getDirectCallee();
704 if (!fd) {
705 unsupported(expr);
706 return NULL;
709 name = fd->getDeclName().getAsString();
710 if (!(expr->getNumArgs() == 2 && name == "min") &&
711 !(expr->getNumArgs() == 2 && name == "max") &&
712 !(expr->getNumArgs() == 2 && name == "floord") &&
713 !(expr->getNumArgs() == 2 && name == "ceild")) {
714 unsupported(expr);
715 return NULL;
718 if (name == "min" || name == "max") {
719 aff1 = extract_affine(expr->getArg(0));
720 aff2 = extract_affine(expr->getArg(1));
722 if (name == "min")
723 aff1 = isl_pw_aff_min(aff1, aff2);
724 else
725 aff1 = isl_pw_aff_max(aff1, aff2);
726 } else if (name == "floord" || name == "ceild") {
727 isl_int v;
728 Expr *arg2 = expr->getArg(1);
730 if (arg2->getStmtClass() != Stmt::IntegerLiteralClass) {
731 unsupported(expr);
732 return NULL;
734 aff1 = extract_affine(expr->getArg(0));
735 isl_int_init(v);
736 extract_int(cast<IntegerLiteral>(arg2), &v);
737 aff1 = isl_pw_aff_scale_down(aff1, v);
738 isl_int_clear(v);
739 if (name == "floord")
740 aff1 = isl_pw_aff_floor(aff1);
741 else
742 aff1 = isl_pw_aff_ceil(aff1);
743 } else {
744 unsupported(expr);
745 return NULL;
748 return aff1;
752 /* This method is called when we come across an access that is
753 * nested in what is supposed to be an affine expression.
754 * If nesting is allowed, we return a new parameter that corresponds
755 * to this nested access. Otherwise, we simply complain.
757 * 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;
766 if (!nesting_enabled) {
767 unsupported(expr);
768 return NULL;
771 id = isl_id_alloc(ctx, NULL, expr);
772 dim = isl_space_params_alloc(ctx, 1);
774 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
776 dom = isl_set_universe(isl_space_copy(dim));
777 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
778 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
780 return isl_pw_aff_alloc(dom, aff);
783 /* Affine expressions are not supposed to contain array accesses,
784 * but if nesting is allowed, we return a parameter corresponding
785 * to the array access.
787 __isl_give isl_pw_aff *PetScan::extract_affine(ArraySubscriptExpr *expr)
789 return nested_access(expr);
792 /* Extract an affine expression from a conditional operation.
794 __isl_give isl_pw_aff *PetScan::extract_affine(ConditionalOperator *expr)
796 isl_pw_aff *cond, *lhs, *rhs, *res;
798 cond = extract_condition(expr->getCond());
799 lhs = extract_affine(expr->getTrueExpr());
800 rhs = extract_affine(expr->getFalseExpr());
802 return isl_pw_aff_cond(cond, lhs, rhs);
805 /* Extract an affine expression, if possible, from "expr".
806 * Otherwise return NULL.
808 __isl_give isl_pw_aff *PetScan::extract_affine(Expr *expr)
810 switch (expr->getStmtClass()) {
811 case Stmt::ImplicitCastExprClass:
812 return extract_affine(cast<ImplicitCastExpr>(expr));
813 case Stmt::IntegerLiteralClass:
814 return extract_affine(cast<IntegerLiteral>(expr));
815 case Stmt::DeclRefExprClass:
816 return extract_affine(cast<DeclRefExpr>(expr));
817 case Stmt::BinaryOperatorClass:
818 return extract_affine(cast<BinaryOperator>(expr));
819 case Stmt::UnaryOperatorClass:
820 return extract_affine(cast<UnaryOperator>(expr));
821 case Stmt::ParenExprClass:
822 return extract_affine(cast<ParenExpr>(expr));
823 case Stmt::CallExprClass:
824 return extract_affine(cast<CallExpr>(expr));
825 case Stmt::ArraySubscriptExprClass:
826 return extract_affine(cast<ArraySubscriptExpr>(expr));
827 case Stmt::ConditionalOperatorClass:
828 return extract_affine(cast<ConditionalOperator>(expr));
829 default:
830 unsupported(expr);
832 return NULL;
835 __isl_give isl_map *PetScan::extract_access(ImplicitCastExpr *expr)
837 return extract_access(expr->getSubExpr());
840 /* Return the depth of an array of the given type.
842 static int array_depth(const Type *type)
844 if (type->isPointerType())
845 return 1 + array_depth(type->getPointeeType().getTypePtr());
846 if (type->isArrayType()) {
847 const ArrayType *atype;
848 type = type->getCanonicalTypeInternal().getTypePtr();
849 atype = cast<ArrayType>(type);
850 return 1 + array_depth(atype->getElementType().getTypePtr());
852 return 0;
855 /* Return the element type of the given array type.
857 static QualType base_type(QualType qt)
859 const Type *type = qt.getTypePtr();
861 if (type->isPointerType())
862 return base_type(type->getPointeeType());
863 if (type->isArrayType()) {
864 const ArrayType *atype;
865 type = type->getCanonicalTypeInternal().getTypePtr();
866 atype = cast<ArrayType>(type);
867 return base_type(atype->getElementType());
869 return qt;
872 /* Extract an access relation from a reference to a variable.
873 * If the variable has name "A" and its type corresponds to an
874 * array of depth d, then the returned access relation is of the
875 * form
877 * { [] -> A[i_1,...,i_d] }
879 __isl_give isl_map *PetScan::extract_access(DeclRefExpr *expr)
881 ValueDecl *decl = expr->getDecl();
882 int depth = array_depth(decl->getType().getTypePtr());
883 isl_id *id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
884 isl_space *dim = isl_space_alloc(ctx, 0, 0, depth);
885 isl_map *access_rel;
887 dim = isl_space_set_tuple_id(dim, isl_dim_out, id);
889 access_rel = isl_map_universe(dim);
891 return access_rel;
894 /* Extract an access relation from an integer contant.
895 * If the value of the constant is "v", then the returned access relation
896 * is
898 * { [] -> [v] }
900 __isl_give isl_map *PetScan::extract_access(IntegerLiteral *expr)
902 return isl_map_from_range(isl_set_from_pw_aff(extract_affine(expr)));
905 /* Try and extract an access relation from the given Expr.
906 * Return NULL if it doesn't work out.
908 __isl_give isl_map *PetScan::extract_access(Expr *expr)
910 switch (expr->getStmtClass()) {
911 case Stmt::ImplicitCastExprClass:
912 return extract_access(cast<ImplicitCastExpr>(expr));
913 case Stmt::DeclRefExprClass:
914 return extract_access(cast<DeclRefExpr>(expr));
915 case Stmt::ArraySubscriptExprClass:
916 return extract_access(cast<ArraySubscriptExpr>(expr));
917 default:
918 unsupported(expr);
920 return NULL;
923 /* Assign the affine expression "index" to the output dimension "pos" of "map"
924 * and return the result.
926 __isl_give isl_map *set_index(__isl_take isl_map *map, int pos,
927 __isl_take isl_pw_aff *index)
929 isl_map *index_map;
930 int len = isl_map_dim(map, isl_dim_out);
931 isl_id *id;
933 index_map = isl_map_from_range(isl_set_from_pw_aff(index));
934 index_map = isl_map_insert_dims(index_map, isl_dim_out, 0, pos);
935 index_map = isl_map_add_dims(index_map, isl_dim_out, len - pos - 1);
936 id = isl_map_get_tuple_id(map, isl_dim_out);
937 index_map = isl_map_set_tuple_id(index_map, isl_dim_out, id);
939 map = isl_map_intersect(map, index_map);
941 return map;
944 /* Extract an access relation from the given array subscript expression.
945 * If nesting is allowed in general, then we turn it on while
946 * examining the index expression.
948 * We first extract an access relation from the base.
949 * This will result in an access relation with a range that corresponds
950 * to the array being accessed and with earlier indices filled in already.
951 * We then extract the current index and fill that in as well.
952 * The position of the current index is based on the type of base.
953 * If base is the actual array variable, then the depth of this type
954 * will be the same as the depth of the array and we will fill in
955 * the first array index.
956 * Otherwise, the depth of the base type will be smaller and we will fill
957 * in a later index.
959 __isl_give isl_map *PetScan::extract_access(ArraySubscriptExpr *expr)
961 Expr *base = expr->getBase();
962 Expr *idx = expr->getIdx();
963 isl_pw_aff *index;
964 isl_map *base_access;
965 isl_map *access;
966 int depth = array_depth(base->getType().getTypePtr());
967 int pos;
968 bool save_nesting = nesting_enabled;
970 nesting_enabled = allow_nested;
972 base_access = extract_access(base);
973 index = extract_affine(idx);
975 nesting_enabled = save_nesting;
977 pos = isl_map_dim(base_access, isl_dim_out) - depth;
978 access = set_index(base_access, pos, index);
980 return access;
983 /* Check if "expr" calls function "minmax" with two arguments and if so
984 * make lhs and rhs refer to these two arguments.
986 static bool is_minmax(Expr *expr, const char *minmax, Expr *&lhs, Expr *&rhs)
988 CallExpr *call;
989 FunctionDecl *fd;
990 string name;
992 if (expr->getStmtClass() != Stmt::CallExprClass)
993 return false;
995 call = cast<CallExpr>(expr);
996 fd = call->getDirectCallee();
997 if (!fd)
998 return false;
1000 if (call->getNumArgs() != 2)
1001 return false;
1003 name = fd->getDeclName().getAsString();
1004 if (name != minmax)
1005 return false;
1007 lhs = call->getArg(0);
1008 rhs = call->getArg(1);
1010 return true;
1013 /* Check if "expr" is of the form min(lhs, rhs) and if so make
1014 * lhs and rhs refer to the two arguments.
1016 static bool is_min(Expr *expr, Expr *&lhs, Expr *&rhs)
1018 return is_minmax(expr, "min", lhs, rhs);
1021 /* Check if "expr" is of the form max(lhs, rhs) and if so make
1022 * lhs and rhs refer to the two arguments.
1024 static bool is_max(Expr *expr, Expr *&lhs, Expr *&rhs)
1026 return is_minmax(expr, "max", lhs, rhs);
1029 /* Return "lhs && rhs", defined on the shared definition domain.
1031 static __isl_give isl_pw_aff *pw_aff_and(__isl_take isl_pw_aff *lhs,
1032 __isl_take isl_pw_aff *rhs)
1034 isl_set *cond;
1035 isl_set *dom;
1037 dom = isl_set_intersect(isl_pw_aff_domain(isl_pw_aff_copy(lhs)),
1038 isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1039 cond = isl_set_intersect(isl_pw_aff_non_zero_set(lhs),
1040 isl_pw_aff_non_zero_set(rhs));
1041 return indicator_function(cond, dom);
1044 /* Return "lhs && rhs", with shortcut semantics.
1045 * That is, if lhs is false, then the result is defined even if rhs is not.
1046 * In practice, we compute lhs ? rhs : lhs.
1048 static __isl_give isl_pw_aff *pw_aff_and_then(__isl_take isl_pw_aff *lhs,
1049 __isl_take isl_pw_aff *rhs)
1051 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), rhs, lhs);
1054 /* Return "lhs || rhs", with shortcut semantics.
1055 * That is, if lhs is true, then the result is defined even if rhs is not.
1056 * In practice, we compute lhs ? lhs : rhs.
1058 static __isl_give isl_pw_aff *pw_aff_or_else(__isl_take isl_pw_aff *lhs,
1059 __isl_take isl_pw_aff *rhs)
1061 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), lhs, rhs);
1064 /* Extract an affine expressions representing the comparison "LHS op RHS"
1065 * "comp" is the original statement that "LHS op RHS" is derived from
1066 * and is used for diagnostics.
1068 * If the comparison is of the form
1070 * a <= min(b,c)
1072 * then the expression is constructed as the conjunction of
1073 * the comparisons
1075 * a <= b and a <= c
1077 * A similar optimization is performed for max(a,b) <= c.
1078 * We do this because that will lead to simpler representations
1079 * of the expression.
1080 * If isl is ever enhanced to explicitly deal with min and max expressions,
1081 * this optimization can be removed.
1083 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperatorKind op,
1084 Expr *LHS, Expr *RHS, Stmt *comp)
1086 isl_pw_aff *lhs;
1087 isl_pw_aff *rhs;
1088 isl_pw_aff *res;
1089 isl_set *cond;
1090 isl_set *dom;
1092 if (op == BO_GT)
1093 return extract_comparison(BO_LT, RHS, LHS, comp);
1094 if (op == BO_GE)
1095 return extract_comparison(BO_LE, RHS, LHS, comp);
1097 if (op == BO_LT || op == BO_LE) {
1098 Expr *expr1, *expr2;
1099 if (is_min(RHS, expr1, expr2)) {
1100 lhs = extract_comparison(op, LHS, expr1, comp);
1101 rhs = extract_comparison(op, LHS, expr2, comp);
1102 return pw_aff_and(lhs, rhs);
1104 if (is_max(LHS, expr1, expr2)) {
1105 lhs = extract_comparison(op, expr1, RHS, comp);
1106 rhs = extract_comparison(op, expr2, RHS, comp);
1107 return pw_aff_and(lhs, rhs);
1111 lhs = extract_affine(LHS);
1112 rhs = extract_affine(RHS);
1114 dom = isl_pw_aff_domain(isl_pw_aff_copy(lhs));
1115 dom = isl_set_intersect(dom, isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1117 switch (op) {
1118 case BO_LT:
1119 cond = isl_pw_aff_lt_set(lhs, rhs);
1120 break;
1121 case BO_LE:
1122 cond = isl_pw_aff_le_set(lhs, rhs);
1123 break;
1124 case BO_EQ:
1125 cond = isl_pw_aff_eq_set(lhs, rhs);
1126 break;
1127 case BO_NE:
1128 cond = isl_pw_aff_ne_set(lhs, rhs);
1129 break;
1130 default:
1131 isl_pw_aff_free(lhs);
1132 isl_pw_aff_free(rhs);
1133 isl_set_free(dom);
1134 unsupported(comp);
1135 return NULL;
1138 cond = isl_set_coalesce(cond);
1139 res = indicator_function(cond, dom);
1141 return res;
1144 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperator *comp)
1146 return extract_comparison(comp->getOpcode(), comp->getLHS(),
1147 comp->getRHS(), comp);
1150 /* Extract an affine expression representing the negation (logical not)
1151 * of a subexpression.
1153 __isl_give isl_pw_aff *PetScan::extract_boolean(UnaryOperator *op)
1155 isl_set *set_cond, *dom;
1156 isl_pw_aff *cond, *res;
1158 cond = extract_condition(op->getSubExpr());
1160 dom = isl_pw_aff_domain(isl_pw_aff_copy(cond));
1162 set_cond = isl_pw_aff_zero_set(cond);
1164 res = indicator_function(set_cond, dom);
1166 return res;
1169 /* Extract an affine expression representing the disjunction (logical or)
1170 * or conjunction (logical and) of two subexpressions.
1172 __isl_give isl_pw_aff *PetScan::extract_boolean(BinaryOperator *comp)
1174 isl_pw_aff *lhs, *rhs;
1176 lhs = extract_condition(comp->getLHS());
1177 rhs = extract_condition(comp->getRHS());
1179 switch (comp->getOpcode()) {
1180 case BO_LAnd:
1181 return pw_aff_and_then(lhs, rhs);
1182 case BO_LOr:
1183 return pw_aff_or_else(lhs, rhs);
1184 default:
1185 isl_pw_aff_free(lhs);
1186 isl_pw_aff_free(rhs);
1189 unsupported(comp);
1190 return NULL;
1193 __isl_give isl_pw_aff *PetScan::extract_condition(UnaryOperator *expr)
1195 switch (expr->getOpcode()) {
1196 case UO_LNot:
1197 return extract_boolean(expr);
1198 default:
1199 unsupported(expr);
1200 return NULL;
1204 /* Extract the affine expression "expr != 0 ? 1 : 0".
1206 __isl_give isl_pw_aff *PetScan::extract_implicit_condition(Expr *expr)
1208 isl_pw_aff *res;
1209 isl_set *set, *dom;
1211 res = extract_affine(expr);
1213 dom = isl_pw_aff_domain(isl_pw_aff_copy(res));
1214 set = isl_pw_aff_non_zero_set(res);
1216 res = indicator_function(set, dom);
1218 return res;
1221 /* Extract an affine expression from a boolean expression.
1222 * In particular, return the expression "expr ? 1 : 0".
1224 * If the expression doesn't look like a condition, we assume it
1225 * is an affine expression and return the condition "expr != 0 ? 1 : 0".
1227 __isl_give isl_pw_aff *PetScan::extract_condition(Expr *expr)
1229 BinaryOperator *comp;
1231 if (!expr) {
1232 isl_set *u = isl_set_universe(isl_space_params_alloc(ctx, 0));
1233 return indicator_function(u, isl_set_copy(u));
1236 if (expr->getStmtClass() == Stmt::ParenExprClass)
1237 return extract_condition(cast<ParenExpr>(expr)->getSubExpr());
1239 if (expr->getStmtClass() == Stmt::UnaryOperatorClass)
1240 return extract_condition(cast<UnaryOperator>(expr));
1242 if (expr->getStmtClass() != Stmt::BinaryOperatorClass)
1243 return extract_implicit_condition(expr);
1245 comp = cast<BinaryOperator>(expr);
1246 switch (comp->getOpcode()) {
1247 case BO_LT:
1248 case BO_LE:
1249 case BO_GT:
1250 case BO_GE:
1251 case BO_EQ:
1252 case BO_NE:
1253 return extract_comparison(comp);
1254 case BO_LAnd:
1255 case BO_LOr:
1256 return extract_boolean(comp);
1257 default:
1258 return extract_implicit_condition(expr);
1262 static enum pet_op_type UnaryOperatorKind2pet_op_type(UnaryOperatorKind kind)
1264 switch (kind) {
1265 case UO_Minus:
1266 return pet_op_minus;
1267 default:
1268 return pet_op_last;
1272 static enum pet_op_type BinaryOperatorKind2pet_op_type(BinaryOperatorKind kind)
1274 switch (kind) {
1275 case BO_AddAssign:
1276 return pet_op_add_assign;
1277 case BO_SubAssign:
1278 return pet_op_sub_assign;
1279 case BO_MulAssign:
1280 return pet_op_mul_assign;
1281 case BO_DivAssign:
1282 return pet_op_div_assign;
1283 case BO_Assign:
1284 return pet_op_assign;
1285 case BO_Add:
1286 return pet_op_add;
1287 case BO_Sub:
1288 return pet_op_sub;
1289 case BO_Mul:
1290 return pet_op_mul;
1291 case BO_Div:
1292 return pet_op_div;
1293 case BO_EQ:
1294 return pet_op_eq;
1295 case BO_LE:
1296 return pet_op_le;
1297 case BO_LT:
1298 return pet_op_lt;
1299 case BO_GT:
1300 return pet_op_gt;
1301 default:
1302 return pet_op_last;
1306 /* Construct a pet_expr representing a unary operator expression.
1308 struct pet_expr *PetScan::extract_expr(UnaryOperator *expr)
1310 struct pet_expr *arg;
1311 enum pet_op_type op;
1313 op = UnaryOperatorKind2pet_op_type(expr->getOpcode());
1314 if (op == pet_op_last) {
1315 unsupported(expr);
1316 return NULL;
1319 arg = extract_expr(expr->getSubExpr());
1321 return pet_expr_new_unary(ctx, op, arg);
1324 /* Mark the given access pet_expr as a write.
1325 * If a scalar is being accessed, then mark its value
1326 * as unknown in assigned_value.
1328 void PetScan::mark_write(struct pet_expr *access)
1330 isl_id *id;
1331 ValueDecl *decl;
1333 access->acc.write = 1;
1334 access->acc.read = 0;
1336 if (isl_map_dim(access->acc.access, isl_dim_out) != 0)
1337 return;
1339 id = isl_map_get_tuple_id(access->acc.access, isl_dim_out);
1340 decl = (ValueDecl *) isl_id_get_user(id);
1341 clear_assignment(assigned_value, decl);
1342 isl_id_free(id);
1345 /* Construct a pet_expr representing a binary operator expression.
1347 * If the top level operator is an assignment and the LHS is an access,
1348 * then we mark that access as a write. If the operator is a compound
1349 * assignment, the access is marked as both a read and a write.
1351 * If "expr" assigns something to a scalar variable, then we mark
1352 * the variable as having been assigned. If, furthermore, the expression
1353 * is affine, then keep track of this value in assigned_value
1354 * so that we can plug it in when we later come across the same variable.
1356 struct pet_expr *PetScan::extract_expr(BinaryOperator *expr)
1358 struct pet_expr *lhs, *rhs;
1359 enum pet_op_type op;
1361 op = BinaryOperatorKind2pet_op_type(expr->getOpcode());
1362 if (op == pet_op_last) {
1363 unsupported(expr);
1364 return NULL;
1367 lhs = extract_expr(expr->getLHS());
1368 rhs = extract_expr(expr->getRHS());
1370 if (expr->isAssignmentOp() && lhs && lhs->type == pet_expr_access) {
1371 mark_write(lhs);
1372 if (expr->isCompoundAssignmentOp())
1373 lhs->acc.read = 1;
1376 if (expr->getOpcode() == BO_Assign &&
1377 lhs && lhs->type == pet_expr_access &&
1378 isl_map_dim(lhs->acc.access, isl_dim_out) == 0) {
1379 isl_id *id = isl_map_get_tuple_id(lhs->acc.access, isl_dim_out);
1380 ValueDecl *decl = (ValueDecl *) isl_id_get_user(id);
1381 Expr *rhs = expr->getRHS();
1382 isl_pw_aff *pa = try_extract_affine(rhs);
1383 clear_assignment(assigned_value, decl);
1384 if (pa) {
1385 assigned_value[decl] = pa;
1386 insert_expression(pa);
1388 isl_id_free(id);
1391 return pet_expr_new_binary(ctx, op, lhs, rhs);
1394 /* Construct a pet_expr representing a conditional operation.
1396 * We first try to extract the condition as an affine expression.
1397 * If that fails, we construct a pet_expr tree representing the condition.
1399 struct pet_expr *PetScan::extract_expr(ConditionalOperator *expr)
1401 struct pet_expr *cond, *lhs, *rhs;
1402 isl_pw_aff *pa;
1404 pa = try_extract_affine(expr->getCond());
1405 if (pa) {
1406 isl_set *test = isl_set_from_pw_aff(pa);
1407 cond = pet_expr_from_access(isl_map_from_range(test));
1408 } else
1409 cond = extract_expr(expr->getCond());
1410 lhs = extract_expr(expr->getTrueExpr());
1411 rhs = extract_expr(expr->getFalseExpr());
1413 return pet_expr_new_ternary(ctx, cond, lhs, rhs);
1416 struct pet_expr *PetScan::extract_expr(ImplicitCastExpr *expr)
1418 return extract_expr(expr->getSubExpr());
1421 /* Construct a pet_expr representing a floating point value.
1423 struct pet_expr *PetScan::extract_expr(FloatingLiteral *expr)
1425 return pet_expr_new_double(ctx, expr->getValueAsApproximateDouble());
1428 /* Extract an access relation from "expr" and then convert it into
1429 * a pet_expr.
1431 struct pet_expr *PetScan::extract_access_expr(Expr *expr)
1433 isl_map *access;
1434 struct pet_expr *pe;
1436 switch (expr->getStmtClass()) {
1437 case Stmt::ArraySubscriptExprClass:
1438 access = extract_access(cast<ArraySubscriptExpr>(expr));
1439 break;
1440 case Stmt::DeclRefExprClass:
1441 access = extract_access(cast<DeclRefExpr>(expr));
1442 break;
1443 case Stmt::IntegerLiteralClass:
1444 access = extract_access(cast<IntegerLiteral>(expr));
1445 break;
1446 default:
1447 unsupported(expr);
1448 return NULL;
1451 pe = pet_expr_from_access(access);
1453 return pe;
1456 struct pet_expr *PetScan::extract_expr(ParenExpr *expr)
1458 return extract_expr(expr->getSubExpr());
1461 /* Construct a pet_expr representing a function call.
1463 * If we are passing along a pointer to an array element
1464 * or an entire row or even higher dimensional slice of an array,
1465 * then the function being called may write into the array.
1467 * We assume here that if the function is declared to take a pointer
1468 * to a const type, then the function will perform a read
1469 * and that otherwise, it will perform a write.
1471 struct pet_expr *PetScan::extract_expr(CallExpr *expr)
1473 struct pet_expr *res = NULL;
1474 FunctionDecl *fd;
1475 string name;
1477 fd = expr->getDirectCallee();
1478 if (!fd) {
1479 unsupported(expr);
1480 return NULL;
1483 name = fd->getDeclName().getAsString();
1484 res = pet_expr_new_call(ctx, name.c_str(), expr->getNumArgs());
1485 if (!res)
1486 return NULL;
1488 for (int i = 0; i < expr->getNumArgs(); ++i) {
1489 Expr *arg = expr->getArg(i);
1490 int is_addr = 0;
1491 pet_expr *main_arg;
1493 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
1494 ImplicitCastExpr *ice = cast<ImplicitCastExpr>(arg);
1495 arg = ice->getSubExpr();
1497 if (arg->getStmtClass() == Stmt::UnaryOperatorClass) {
1498 UnaryOperator *op = cast<UnaryOperator>(arg);
1499 if (op->getOpcode() == UO_AddrOf) {
1500 is_addr = 1;
1501 arg = op->getSubExpr();
1504 res->args[i] = PetScan::extract_expr(arg);
1505 main_arg = res->args[i];
1506 if (is_addr)
1507 res->args[i] = pet_expr_new_unary(ctx,
1508 pet_op_address_of, res->args[i]);
1509 if (!res->args[i])
1510 goto error;
1511 if (arg->getStmtClass() == Stmt::ArraySubscriptExprClass &&
1512 array_depth(arg->getType().getTypePtr()) > 0)
1513 is_addr = 1;
1514 if (is_addr && main_arg->type == pet_expr_access) {
1515 ParmVarDecl *parm;
1516 if (!fd->hasPrototype()) {
1517 unsupported(expr, "prototype required");
1518 goto error;
1520 parm = fd->getParamDecl(i);
1521 if (!const_base(parm->getType()))
1522 mark_write(main_arg);
1526 return res;
1527 error:
1528 pet_expr_free(res);
1529 return NULL;
1532 /* Try and onstruct a pet_expr representing "expr".
1534 struct pet_expr *PetScan::extract_expr(Expr *expr)
1536 switch (expr->getStmtClass()) {
1537 case Stmt::UnaryOperatorClass:
1538 return extract_expr(cast<UnaryOperator>(expr));
1539 case Stmt::CompoundAssignOperatorClass:
1540 case Stmt::BinaryOperatorClass:
1541 return extract_expr(cast<BinaryOperator>(expr));
1542 case Stmt::ImplicitCastExprClass:
1543 return extract_expr(cast<ImplicitCastExpr>(expr));
1544 case Stmt::ArraySubscriptExprClass:
1545 case Stmt::DeclRefExprClass:
1546 case Stmt::IntegerLiteralClass:
1547 return extract_access_expr(expr);
1548 case Stmt::FloatingLiteralClass:
1549 return extract_expr(cast<FloatingLiteral>(expr));
1550 case Stmt::ParenExprClass:
1551 return extract_expr(cast<ParenExpr>(expr));
1552 case Stmt::ConditionalOperatorClass:
1553 return extract_expr(cast<ConditionalOperator>(expr));
1554 case Stmt::CallExprClass:
1555 return extract_expr(cast<CallExpr>(expr));
1556 default:
1557 unsupported(expr);
1559 return NULL;
1562 /* Check if the given initialization statement is an assignment.
1563 * If so, return that assignment. Otherwise return NULL.
1565 BinaryOperator *PetScan::initialization_assignment(Stmt *init)
1567 BinaryOperator *ass;
1569 if (init->getStmtClass() != Stmt::BinaryOperatorClass)
1570 return NULL;
1572 ass = cast<BinaryOperator>(init);
1573 if (ass->getOpcode() != BO_Assign)
1574 return NULL;
1576 return ass;
1579 /* Check if the given initialization statement is a declaration
1580 * of a single variable.
1581 * If so, return that declaration. Otherwise return NULL.
1583 Decl *PetScan::initialization_declaration(Stmt *init)
1585 DeclStmt *decl;
1587 if (init->getStmtClass() != Stmt::DeclStmtClass)
1588 return NULL;
1590 decl = cast<DeclStmt>(init);
1592 if (!decl->isSingleDecl())
1593 return NULL;
1595 return decl->getSingleDecl();
1598 /* Given the assignment operator in the initialization of a for loop,
1599 * extract the induction variable, i.e., the (integer)variable being
1600 * assigned.
1602 ValueDecl *PetScan::extract_induction_variable(BinaryOperator *init)
1604 Expr *lhs;
1605 DeclRefExpr *ref;
1606 ValueDecl *decl;
1607 const Type *type;
1609 lhs = init->getLHS();
1610 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1611 unsupported(init);
1612 return NULL;
1615 ref = cast<DeclRefExpr>(lhs);
1616 decl = ref->getDecl();
1617 type = decl->getType().getTypePtr();
1619 if (!type->isIntegerType()) {
1620 unsupported(lhs);
1621 return NULL;
1624 return decl;
1627 /* Given the initialization statement of a for loop and the single
1628 * declaration in this initialization statement,
1629 * extract the induction variable, i.e., the (integer) variable being
1630 * declared.
1632 VarDecl *PetScan::extract_induction_variable(Stmt *init, Decl *decl)
1634 VarDecl *vd;
1636 vd = cast<VarDecl>(decl);
1638 const QualType type = vd->getType();
1639 if (!type->isIntegerType()) {
1640 unsupported(init);
1641 return NULL;
1644 if (!vd->getInit()) {
1645 unsupported(init);
1646 return NULL;
1649 return vd;
1652 /* Check that op is of the form iv++ or iv--.
1653 * Return an affine expression "1" or "-1" accordingly.
1655 __isl_give isl_pw_aff *PetScan::extract_unary_increment(
1656 clang::UnaryOperator *op, clang::ValueDecl *iv)
1658 Expr *sub;
1659 DeclRefExpr *ref;
1660 isl_space *space;
1661 isl_aff *aff;
1663 if (!op->isIncrementDecrementOp()) {
1664 unsupported(op);
1665 return NULL;
1668 sub = op->getSubExpr();
1669 if (sub->getStmtClass() != Stmt::DeclRefExprClass) {
1670 unsupported(op);
1671 return NULL;
1674 ref = cast<DeclRefExpr>(sub);
1675 if (ref->getDecl() != iv) {
1676 unsupported(op);
1677 return NULL;
1680 space = isl_space_params_alloc(ctx, 0);
1681 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
1683 if (op->isIncrementOp())
1684 aff = isl_aff_add_constant_si(aff, 1);
1685 else
1686 aff = isl_aff_add_constant_si(aff, -1);
1688 return isl_pw_aff_from_aff(aff);
1691 /* If the isl_pw_aff on which isl_pw_aff_foreach_piece is called
1692 * has a single constant expression, then put this constant in *user.
1693 * The caller is assumed to have checked that this function will
1694 * be called exactly once.
1696 static int extract_cst(__isl_take isl_set *set, __isl_take isl_aff *aff,
1697 void *user)
1699 isl_int *inc = (isl_int *)user;
1700 int res = 0;
1702 if (isl_aff_is_cst(aff))
1703 isl_aff_get_constant(aff, inc);
1704 else
1705 res = -1;
1707 isl_set_free(set);
1708 isl_aff_free(aff);
1710 return res;
1713 /* Check if op is of the form
1715 * iv = iv + inc
1717 * and return inc as an affine expression.
1719 * We extract an affine expression from the RHS, subtract iv and return
1720 * the result.
1722 __isl_give isl_pw_aff *PetScan::extract_binary_increment(BinaryOperator *op,
1723 clang::ValueDecl *iv)
1725 Expr *lhs;
1726 DeclRefExpr *ref;
1727 isl_id *id;
1728 isl_space *dim;
1729 isl_aff *aff;
1730 isl_pw_aff *val;
1732 if (op->getOpcode() != BO_Assign) {
1733 unsupported(op);
1734 return NULL;
1737 lhs = op->getLHS();
1738 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1739 unsupported(op);
1740 return NULL;
1743 ref = cast<DeclRefExpr>(lhs);
1744 if (ref->getDecl() != iv) {
1745 unsupported(op);
1746 return NULL;
1749 val = extract_affine(op->getRHS());
1751 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
1753 dim = isl_space_params_alloc(ctx, 1);
1754 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
1755 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
1756 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
1758 val = isl_pw_aff_sub(val, isl_pw_aff_from_aff(aff));
1760 return val;
1763 /* Check that op is of the form iv += cst or iv -= cst
1764 * and return an affine expression corresponding oto cst or -cst accordingly.
1766 __isl_give isl_pw_aff *PetScan::extract_compound_increment(
1767 CompoundAssignOperator *op, clang::ValueDecl *iv)
1769 Expr *lhs;
1770 DeclRefExpr *ref;
1771 bool neg = false;
1772 isl_pw_aff *val;
1773 BinaryOperatorKind opcode;
1775 opcode = op->getOpcode();
1776 if (opcode != BO_AddAssign && opcode != BO_SubAssign) {
1777 unsupported(op);
1778 return NULL;
1780 if (opcode == BO_SubAssign)
1781 neg = true;
1783 lhs = op->getLHS();
1784 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1785 unsupported(op);
1786 return NULL;
1789 ref = cast<DeclRefExpr>(lhs);
1790 if (ref->getDecl() != iv) {
1791 unsupported(op);
1792 return NULL;
1795 val = extract_affine(op->getRHS());
1796 if (neg)
1797 val = isl_pw_aff_neg(val);
1799 return val;
1802 /* Check that the increment of the given for loop increments
1803 * (or decrements) the induction variable "iv" and return
1804 * the increment as an affine expression if successful.
1806 __isl_give isl_pw_aff *PetScan::extract_increment(clang::ForStmt *stmt,
1807 ValueDecl *iv)
1809 Stmt *inc = stmt->getInc();
1811 if (!inc) {
1812 unsupported(stmt);
1813 return NULL;
1816 if (inc->getStmtClass() == Stmt::UnaryOperatorClass)
1817 return extract_unary_increment(cast<UnaryOperator>(inc), iv);
1818 if (inc->getStmtClass() == Stmt::CompoundAssignOperatorClass)
1819 return extract_compound_increment(
1820 cast<CompoundAssignOperator>(inc), iv);
1821 if (inc->getStmtClass() == Stmt::BinaryOperatorClass)
1822 return extract_binary_increment(cast<BinaryOperator>(inc), iv);
1824 unsupported(inc);
1825 return NULL;
1828 /* Embed the given iteration domain in an extra outer loop
1829 * with induction variable "var".
1830 * If this variable appeared as a parameter in the constraints,
1831 * it is replaced by the new outermost dimension.
1833 static __isl_give isl_set *embed(__isl_take isl_set *set,
1834 __isl_take isl_id *var)
1836 int pos;
1838 set = isl_set_insert_dims(set, isl_dim_set, 0, 1);
1839 pos = isl_set_find_dim_by_id(set, isl_dim_param, var);
1840 if (pos >= 0) {
1841 set = isl_set_equate(set, isl_dim_param, pos, isl_dim_set, 0);
1842 set = isl_set_project_out(set, isl_dim_param, pos, 1);
1845 isl_id_free(var);
1846 return set;
1849 /* Construct a pet_scop for an infinite loop around the given body.
1851 * We extract a pet_scop for the body and then embed it in a loop with
1852 * iteration domain
1854 * { [t] : t >= 0 }
1856 * and schedule
1858 * { [t] -> [t] }
1860 struct pet_scop *PetScan::extract_infinite_loop(Stmt *body)
1862 isl_id *id;
1863 isl_space *dim;
1864 isl_set *domain;
1865 isl_map *sched;
1866 struct pet_scop *scop;
1868 scop = extract(body);
1869 if (!scop)
1870 return NULL;
1872 id = isl_id_alloc(ctx, "t", NULL);
1873 domain = isl_set_nat_universe(isl_space_set_alloc(ctx, 0, 1));
1874 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, isl_id_copy(id));
1875 dim = isl_space_from_domain(isl_set_get_space(domain));
1876 dim = isl_space_add_dims(dim, isl_dim_out, 1);
1877 sched = isl_map_universe(dim);
1878 sched = isl_map_equate(sched, isl_dim_in, 0, isl_dim_out, 0);
1879 scop = pet_scop_embed(scop, domain, sched, id);
1881 return scop;
1884 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
1886 * for (;;)
1887 * body
1890 struct pet_scop *PetScan::extract_infinite_for(ForStmt *stmt)
1892 return extract_infinite_loop(stmt->getBody());
1895 /* Check if the while loop is of the form
1897 * while (1)
1898 * body
1900 * If so, construct a scop for an infinite loop around body.
1901 * Otherwise, fail.
1903 struct pet_scop *PetScan::extract(WhileStmt *stmt)
1905 Expr *cond;
1906 isl_set *set;
1907 int is_universe;
1909 cond = stmt->getCond();
1910 if (!cond) {
1911 unsupported(stmt);
1912 return NULL;
1915 set = isl_pw_aff_non_zero_set(extract_condition(cond));
1916 is_universe = isl_set_plain_is_universe(set);
1917 isl_set_free(set);
1919 if (!is_universe) {
1920 unsupported(stmt);
1921 return NULL;
1924 return extract_infinite_loop(stmt->getBody());
1927 /* Check whether "cond" expresses a simple loop bound
1928 * on the only set dimension.
1929 * In particular, if "up" is set then "cond" should contain only
1930 * upper bounds on the set dimension.
1931 * Otherwise, it should contain only lower bounds.
1933 static bool is_simple_bound(__isl_keep isl_set *cond, isl_int inc)
1935 if (isl_int_is_pos(inc))
1936 return !isl_set_dim_has_lower_bound(cond, isl_dim_set, 0);
1937 else
1938 return !isl_set_dim_has_upper_bound(cond, isl_dim_set, 0);
1941 /* Extend a condition on a given iteration of a loop to one that
1942 * imposes the same condition on all previous iterations.
1943 * "domain" expresses the lower [upper] bound on the iterations
1944 * when inc is positive [negative].
1946 * In particular, we construct the condition (when inc is positive)
1948 * forall i' : (domain(i') and i' <= i) => cond(i')
1950 * which is equivalent to
1952 * not exists i' : domain(i') and i' <= i and not cond(i')
1954 * We construct this set by negating cond, applying a map
1956 * { [i'] -> [i] : domain(i') and i' <= i }
1958 * and then negating the result again.
1960 static __isl_give isl_set *valid_for_each_iteration(__isl_take isl_set *cond,
1961 __isl_take isl_set *domain, isl_int inc)
1963 isl_map *previous_to_this;
1965 if (isl_int_is_pos(inc))
1966 previous_to_this = isl_map_lex_le(isl_set_get_space(domain));
1967 else
1968 previous_to_this = isl_map_lex_ge(isl_set_get_space(domain));
1970 previous_to_this = isl_map_intersect_domain(previous_to_this, domain);
1972 cond = isl_set_complement(cond);
1973 cond = isl_set_apply(cond, previous_to_this);
1974 cond = isl_set_complement(cond);
1976 return cond;
1979 /* Construct a domain of the form
1981 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
1983 static __isl_give isl_set *strided_domain(__isl_take isl_id *id,
1984 __isl_take isl_pw_aff *init, isl_int inc)
1986 isl_aff *aff;
1987 isl_space *dim;
1988 isl_set *set;
1990 init = isl_pw_aff_insert_dims(init, isl_dim_in, 0, 1);
1991 dim = isl_pw_aff_get_domain_space(init);
1992 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
1993 aff = isl_aff_add_coefficient(aff, isl_dim_in, 0, inc);
1994 init = isl_pw_aff_add(init, isl_pw_aff_from_aff(aff));
1996 dim = isl_space_set_alloc(isl_pw_aff_get_ctx(init), 1, 1);
1997 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
1998 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
1999 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2001 set = isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff), init);
2003 set = isl_set_lower_bound_si(set, isl_dim_set, 0, 0);
2005 return isl_set_params(set);
2008 /* Assuming "cond" represents a bound on a loop where the loop
2009 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2010 * is possible.
2012 * Under the given assumptions, wrapping is only possible if "cond" allows
2013 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2014 * increasing iterator and 0 in case of a decreasing iterator.
2016 static bool can_wrap(__isl_keep isl_set *cond, ValueDecl *iv, isl_int inc)
2018 bool cw;
2019 isl_int limit;
2020 isl_set *test;
2022 test = isl_set_copy(cond);
2024 isl_int_init(limit);
2025 if (isl_int_is_neg(inc))
2026 isl_int_set_si(limit, 0);
2027 else {
2028 isl_int_set_si(limit, 1);
2029 isl_int_mul_2exp(limit, limit, get_type_size(iv));
2030 isl_int_sub_ui(limit, limit, 1);
2033 test = isl_set_fix(cond, isl_dim_set, 0, limit);
2034 cw = !isl_set_is_empty(test);
2035 isl_set_free(test);
2037 isl_int_clear(limit);
2039 return cw;
2042 /* Given a one-dimensional space, construct the following mapping on this
2043 * space
2045 * { [v] -> [v mod 2^width] }
2047 * where width is the number of bits used to represent the values
2048 * of the unsigned variable "iv".
2050 static __isl_give isl_map *compute_wrapping(__isl_take isl_space *dim,
2051 ValueDecl *iv)
2053 isl_int mod;
2054 isl_aff *aff;
2055 isl_map *map;
2057 isl_int_init(mod);
2058 isl_int_set_si(mod, 1);
2059 isl_int_mul_2exp(mod, mod, get_type_size(iv));
2061 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2062 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2063 aff = isl_aff_mod(aff, mod);
2065 isl_int_clear(mod);
2067 return isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2068 map = isl_map_reverse(map);
2071 /* Project out the parameter "id" from "set".
2073 static __isl_give isl_set *set_project_out_by_id(__isl_take isl_set *set,
2074 __isl_keep isl_id *id)
2076 int pos;
2078 pos = isl_set_find_dim_by_id(set, isl_dim_param, id);
2079 if (pos >= 0)
2080 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2082 return set;
2085 /* Compute the set of parameters for which "set1" is a subset of "set2".
2087 * set1 is a subset of set2 if
2089 * forall i in set1 : i in set2
2091 * or
2093 * not exists i in set1 and i not in set2
2095 * i.e.,
2097 * not exists i in set1 \ set2
2099 static __isl_give isl_set *enforce_subset(__isl_take isl_set *set1,
2100 __isl_take isl_set *set2)
2102 return isl_set_complement(isl_set_params(isl_set_subtract(set1, set2)));
2105 /* Compute the set of parameter values for which "cond" holds
2106 * on the next iteration for each element of "dom".
2108 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2109 * and then compute the set of parameters for which the result is a subset
2110 * of "cond".
2112 static __isl_give isl_set *valid_on_next(__isl_take isl_set *cond,
2113 __isl_take isl_set *dom, isl_int inc)
2115 isl_space *space;
2116 isl_aff *aff;
2117 isl_map *next;
2119 space = isl_set_get_space(dom);
2120 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2121 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2122 aff = isl_aff_add_constant(aff, inc);
2123 next = isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2125 dom = isl_set_apply(dom, next);
2127 return enforce_subset(dom, cond);
2130 /* Construct a pet_scop for a for statement.
2131 * The for loop is required to be of the form
2133 * for (i = init; condition; ++i)
2135 * or
2137 * for (i = init; condition; --i)
2139 * The initialization of the for loop should either be an assignment
2140 * to an integer variable, or a declaration of such a variable with
2141 * initialization.
2143 * The condition is allowed to contain nested accesses, provided
2144 * they are not being written to inside the body of the loop.
2146 * We extract a pet_scop for the body and then embed it in a loop with
2147 * iteration domain and schedule
2149 * { [i] : i >= init and condition' }
2150 * { [i] -> [i] }
2152 * or
2154 * { [i] : i <= init and condition' }
2155 * { [i] -> [-i] }
2157 * Where condition' is equal to condition if the latter is
2158 * a simple upper [lower] bound and a condition that is extended
2159 * to apply to all previous iterations otherwise.
2161 * If the stride of the loop is not 1, then "i >= init" is replaced by
2163 * (exists a: i = init + stride * a and a >= 0)
2165 * If the loop iterator i is unsigned, then wrapping may occur.
2166 * During the computation, we work with a virtual iterator that
2167 * does not wrap. However, the condition in the code applies
2168 * to the wrapped value, so we need to change condition(i)
2169 * into condition([i % 2^width]).
2170 * After computing the virtual domain and schedule, we apply
2171 * the function { [v] -> [v % 2^width] } to the domain and the domain
2172 * of the schedule. In order not to lose any information, we also
2173 * need to intersect the domain of the schedule with the virtual domain
2174 * first, since some iterations in the wrapped domain may be scheduled
2175 * several times, typically an infinite number of times.
2176 * Note that there is no need to perform this final wrapping
2177 * if the loop condition (after wrapping) is simple.
2179 * Wrapping on unsigned iterators can be avoided entirely if
2180 * loop condition is simple, the loop iterator is incremented
2181 * [decremented] by one and the last value before wrapping cannot
2182 * possibly satisfy the loop condition.
2184 * Before extracting a pet_scop from the body we remove all
2185 * assignments in assigned_value to variables that are assigned
2186 * somewhere in the body of the loop.
2188 * Valid parameters for a for loop are those for which the initial
2189 * value itself, the increment on each domain iteration and
2190 * the condition on both the initial value and
2191 * the result of incrementing the iterator for each iteration of the domain
2192 * can be evaluated.
2194 struct pet_scop *PetScan::extract_for(ForStmt *stmt)
2196 BinaryOperator *ass;
2197 Decl *decl;
2198 Stmt *init;
2199 Expr *lhs, *rhs;
2200 ValueDecl *iv;
2201 isl_space *dim;
2202 isl_set *domain;
2203 isl_map *sched;
2204 isl_set *cond = NULL;
2205 isl_id *id;
2206 struct pet_scop *scop;
2207 assigned_value_cache cache(assigned_value);
2208 isl_int inc;
2209 bool is_one;
2210 bool is_unsigned;
2211 bool is_simple;
2212 bool is_virtual;
2213 isl_map *wrap = NULL;
2214 isl_pw_aff *pa, *pa_inc, *init_val;
2215 isl_set *valid_init;
2216 isl_set *valid_cond;
2217 isl_set *valid_cond_init;
2218 isl_set *valid_cond_next;
2219 isl_set *valid_inc;
2221 if (!stmt->getInit() && !stmt->getCond() && !stmt->getInc())
2222 return extract_infinite_for(stmt);
2224 init = stmt->getInit();
2225 if (!init) {
2226 unsupported(stmt);
2227 return NULL;
2229 if ((ass = initialization_assignment(init)) != NULL) {
2230 iv = extract_induction_variable(ass);
2231 if (!iv)
2232 return NULL;
2233 lhs = ass->getLHS();
2234 rhs = ass->getRHS();
2235 } else if ((decl = initialization_declaration(init)) != NULL) {
2236 VarDecl *var = extract_induction_variable(init, decl);
2237 if (!var)
2238 return NULL;
2239 iv = var;
2240 rhs = var->getInit();
2241 lhs = create_DeclRefExpr(var);
2242 } else {
2243 unsupported(stmt->getInit());
2244 return NULL;
2247 pa_inc = extract_increment(stmt, iv);
2248 if (!pa_inc)
2249 return NULL;
2251 isl_int_init(inc);
2252 if (isl_pw_aff_n_piece(pa_inc) != 1 ||
2253 isl_pw_aff_foreach_piece(pa_inc, &extract_cst, &inc) < 0) {
2254 isl_pw_aff_free(pa_inc);
2255 unsupported(stmt->getInc());
2256 isl_int_clear(inc);
2257 return NULL;
2259 valid_inc = isl_pw_aff_domain(pa_inc);
2261 is_unsigned = iv->getType()->isUnsignedIntegerType();
2263 assigned_value.erase(iv);
2264 clear_assignments clear(assigned_value);
2265 clear.TraverseStmt(stmt->getBody());
2267 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
2269 scop = extract(stmt->getBody());
2271 pa = try_extract_nested_condition(stmt->getCond());
2272 if (pa && !is_nested_allowed(pa, scop)) {
2273 isl_pw_aff_free(pa);
2274 pa = NULL;
2277 if (!pa)
2278 pa = extract_condition(stmt->getCond());
2279 valid_cond = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2280 cond = isl_pw_aff_non_zero_set(pa);
2281 cond = embed(cond, isl_id_copy(id));
2282 valid_cond = isl_set_coalesce(valid_cond);
2283 valid_cond = embed(valid_cond, isl_id_copy(id));
2284 valid_inc = embed(valid_inc, isl_id_copy(id));
2285 is_one = isl_int_is_one(inc) || isl_int_is_negone(inc);
2286 is_virtual = is_unsigned && (!is_one || can_wrap(cond, iv, inc));
2288 init_val = extract_affine(rhs);
2289 valid_cond_init = enforce_subset(
2290 isl_set_from_pw_aff(isl_pw_aff_copy(init_val)),
2291 isl_set_copy(valid_cond));
2292 if (is_one && !is_virtual) {
2293 isl_pw_aff_free(init_val);
2294 pa = extract_comparison(isl_int_is_pos(inc) ? BO_GE : BO_LE,
2295 lhs, rhs, init);
2296 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2297 valid_init = set_project_out_by_id(valid_init, id);
2298 domain = isl_pw_aff_non_zero_set(pa);
2299 } else {
2300 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(init_val));
2301 domain = strided_domain(isl_id_copy(id), init_val, inc);
2304 domain = embed(domain, isl_id_copy(id));
2305 if (is_virtual) {
2306 isl_map *rev_wrap;
2307 wrap = compute_wrapping(isl_set_get_space(cond), iv);
2308 rev_wrap = isl_map_reverse(isl_map_copy(wrap));
2309 cond = isl_set_apply(cond, isl_map_copy(rev_wrap));
2310 valid_cond = isl_set_apply(valid_cond, isl_map_copy(rev_wrap));
2311 valid_inc = isl_set_apply(valid_inc, rev_wrap);
2313 cond = isl_set_gist(cond, isl_set_copy(domain));
2314 is_simple = is_simple_bound(cond, inc);
2315 if (!is_simple)
2316 cond = valid_for_each_iteration(cond,
2317 isl_set_copy(domain), inc);
2318 domain = isl_set_intersect(domain, cond);
2319 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, isl_id_copy(id));
2320 dim = isl_space_from_domain(isl_set_get_space(domain));
2321 dim = isl_space_add_dims(dim, isl_dim_out, 1);
2322 sched = isl_map_universe(dim);
2323 if (isl_int_is_pos(inc))
2324 sched = isl_map_equate(sched, isl_dim_in, 0, isl_dim_out, 0);
2325 else
2326 sched = isl_map_oppose(sched, isl_dim_in, 0, isl_dim_out, 0);
2328 valid_cond_next = valid_on_next(valid_cond, isl_set_copy(domain), inc);
2329 valid_inc = enforce_subset(isl_set_copy(domain), valid_inc);
2331 if (is_virtual && !is_simple) {
2332 wrap = isl_map_set_dim_id(wrap,
2333 isl_dim_out, 0, isl_id_copy(id));
2334 sched = isl_map_intersect_domain(sched, isl_set_copy(domain));
2335 domain = isl_set_apply(domain, isl_map_copy(wrap));
2336 sched = isl_map_apply_domain(sched, wrap);
2337 } else
2338 isl_map_free(wrap);
2340 scop = pet_scop_embed(scop, domain, sched, id);
2341 scop = resolve_nested(scop);
2342 clear_assignment(assigned_value, iv);
2344 isl_int_clear(inc);
2346 scop = pet_scop_restrict_context(scop, valid_init);
2347 scop = pet_scop_restrict_context(scop, valid_inc);
2348 scop = pet_scop_restrict_context(scop, valid_cond_next);
2349 scop = pet_scop_restrict_context(scop, valid_cond_init);
2351 return scop;
2354 struct pet_scop *PetScan::extract(CompoundStmt *stmt)
2356 return extract(stmt->children());
2359 /* Does "id" refer to a nested access?
2361 static bool is_nested_parameter(__isl_keep isl_id *id)
2363 return id && isl_id_get_user(id) && !isl_id_get_name(id);
2366 /* Does parameter "pos" of "space" refer to a nested access?
2368 static bool is_nested_parameter(__isl_keep isl_space *space, int pos)
2370 bool nested;
2371 isl_id *id;
2373 id = isl_space_get_dim_id(space, isl_dim_param, pos);
2374 nested = is_nested_parameter(id);
2375 isl_id_free(id);
2377 return nested;
2380 /* Does parameter "pos" of "map" refer to a nested access?
2382 static bool is_nested_parameter(__isl_keep isl_map *map, int pos)
2384 bool nested;
2385 isl_id *id;
2387 id = isl_map_get_dim_id(map, isl_dim_param, pos);
2388 nested = is_nested_parameter(id);
2389 isl_id_free(id);
2391 return nested;
2394 /* How many parameters of "space" refer to nested accesses, i.e., have no name?
2396 static int n_nested_parameter(__isl_keep isl_space *space)
2398 int n = 0;
2399 int nparam;
2401 nparam = isl_space_dim(space, isl_dim_param);
2402 for (int i = 0; i < nparam; ++i)
2403 if (is_nested_parameter(space, i))
2404 ++n;
2406 return n;
2409 /* How many parameters of "map" refer to nested accesses, i.e., have no name?
2411 static int n_nested_parameter(__isl_keep isl_map *map)
2413 isl_space *space;
2414 int n;
2416 space = isl_map_get_space(map);
2417 n = n_nested_parameter(space);
2418 isl_space_free(space);
2420 return n;
2423 /* For each nested access parameter in "space",
2424 * construct a corresponding pet_expr, place it in args and
2425 * record its position in "param2pos".
2426 * "n_arg" is the number of elements that are already in args.
2427 * The position recorded in "param2pos" takes this number into account.
2428 * If the pet_expr corresponding to a parameter is identical to
2429 * the pet_expr corresponding to an earlier parameter, then these two
2430 * parameters are made to refer to the same element in args.
2432 * Return the final number of elements in args or -1 if an error has occurred.
2434 int PetScan::extract_nested(__isl_keep isl_space *space,
2435 int n_arg, struct pet_expr **args, std::map<int,int> &param2pos)
2437 int nparam;
2439 nparam = isl_space_dim(space, isl_dim_param);
2440 for (int i = 0; i < nparam; ++i) {
2441 int j;
2442 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
2443 Expr *nested;
2445 if (!is_nested_parameter(id)) {
2446 isl_id_free(id);
2447 continue;
2450 nested = (Expr *) isl_id_get_user(id);
2451 args[n_arg] = extract_expr(nested);
2452 if (!args[n_arg])
2453 return -1;
2455 for (j = 0; j < n_arg; ++j)
2456 if (pet_expr_is_equal(args[j], args[n_arg]))
2457 break;
2459 if (j < n_arg) {
2460 pet_expr_free(args[n_arg]);
2461 args[n_arg] = NULL;
2462 param2pos[i] = j;
2463 } else
2464 param2pos[i] = n_arg++;
2466 isl_id_free(id);
2469 return n_arg;
2472 /* For each nested access parameter in the access relations in "expr",
2473 * construct a corresponding pet_expr, place it in expr->args and
2474 * record its position in "param2pos".
2475 * n is the number of nested access parameters.
2477 struct pet_expr *PetScan::extract_nested(struct pet_expr *expr, int n,
2478 std::map<int,int> &param2pos)
2480 isl_space *space;
2482 expr->args = isl_calloc_array(ctx, struct pet_expr *, n);
2483 expr->n_arg = n;
2484 if (!expr->args)
2485 goto error;
2487 space = isl_map_get_space(expr->acc.access);
2488 n = extract_nested(space, 0, expr->args, param2pos);
2489 isl_space_free(space);
2491 if (n < 0)
2492 goto error;
2494 expr->n_arg = n;
2495 return expr;
2496 error:
2497 pet_expr_free(expr);
2498 return NULL;
2501 /* Look for parameters in any access relation in "expr" that
2502 * refer to nested accesses. In particular, these are
2503 * parameters with no name.
2505 * If there are any such parameters, then the domain of the access
2506 * relation, which is still [] at this point, is replaced by
2507 * [[] -> [t_1,...,t_n]], with n the number of these parameters
2508 * (after identifying identical nested accesses).
2509 * The parameters are then equated to the corresponding t dimensions
2510 * and subsequently projected out.
2511 * param2pos maps the position of the parameter to the position
2512 * of the corresponding t dimension.
2514 struct pet_expr *PetScan::resolve_nested(struct pet_expr *expr)
2516 int n;
2517 int nparam;
2518 int n_in;
2519 isl_space *dim;
2520 isl_map *map;
2521 std::map<int,int> param2pos;
2523 if (!expr)
2524 return expr;
2526 for (int i = 0; i < expr->n_arg; ++i) {
2527 expr->args[i] = resolve_nested(expr->args[i]);
2528 if (!expr->args[i]) {
2529 pet_expr_free(expr);
2530 return NULL;
2534 if (expr->type != pet_expr_access)
2535 return expr;
2537 n = n_nested_parameter(expr->acc.access);
2538 if (n == 0)
2539 return expr;
2541 expr = extract_nested(expr, n, param2pos);
2542 if (!expr)
2543 return NULL;
2545 n = expr->n_arg;
2546 nparam = isl_map_dim(expr->acc.access, isl_dim_param);
2547 n_in = isl_map_dim(expr->acc.access, isl_dim_in);
2548 dim = isl_map_get_space(expr->acc.access);
2549 dim = isl_space_domain(dim);
2550 dim = isl_space_from_domain(dim);
2551 dim = isl_space_add_dims(dim, isl_dim_out, n);
2552 map = isl_map_universe(dim);
2553 map = isl_map_domain_map(map);
2554 map = isl_map_reverse(map);
2555 expr->acc.access = isl_map_apply_domain(expr->acc.access, map);
2557 for (int i = nparam - 1; i >= 0; --i) {
2558 isl_id *id = isl_map_get_dim_id(expr->acc.access,
2559 isl_dim_param, i);
2560 if (!is_nested_parameter(id)) {
2561 isl_id_free(id);
2562 continue;
2565 expr->acc.access = isl_map_equate(expr->acc.access,
2566 isl_dim_param, i, isl_dim_in,
2567 n_in + param2pos[i]);
2568 expr->acc.access = isl_map_project_out(expr->acc.access,
2569 isl_dim_param, i, 1);
2571 isl_id_free(id);
2574 return expr;
2575 error:
2576 pet_expr_free(expr);
2577 return NULL;
2580 /* Convert a top-level pet_expr to a pet_scop with one statement.
2581 * This mainly involves resolving nested expression parameters
2582 * and setting the name of the iteration space.
2583 * The name is given by "label" if it is non-NULL. Otherwise,
2584 * it is of the form S_<n_stmt>.
2586 struct pet_scop *PetScan::extract(Stmt *stmt, struct pet_expr *expr,
2587 __isl_take isl_id *label)
2589 struct pet_stmt *ps;
2590 SourceLocation loc = stmt->getLocStart();
2591 int line = PP.getSourceManager().getExpansionLineNumber(loc);
2593 expr = resolve_nested(expr);
2594 ps = pet_stmt_from_pet_expr(ctx, line, label, n_stmt++, expr);
2595 return pet_scop_from_pet_stmt(ctx, ps);
2598 /* Check if we can extract an affine expression from "expr".
2599 * Return the expressions as an isl_pw_aff if we can and NULL otherwise.
2600 * We turn on autodetection so that we won't generate any warnings
2601 * and turn off nesting, so that we won't accept any non-affine constructs.
2603 __isl_give isl_pw_aff *PetScan::try_extract_affine(Expr *expr)
2605 isl_pw_aff *pwaff;
2606 int save_autodetect = autodetect;
2607 bool save_nesting = nesting_enabled;
2609 autodetect = 1;
2610 nesting_enabled = false;
2612 pwaff = extract_affine(expr);
2614 autodetect = save_autodetect;
2615 nesting_enabled = save_nesting;
2617 return pwaff;
2620 /* Check whether "expr" is an affine expression.
2622 bool PetScan::is_affine(Expr *expr)
2624 isl_pw_aff *pwaff;
2626 pwaff = try_extract_affine(expr);
2627 isl_pw_aff_free(pwaff);
2629 return pwaff != NULL;
2632 /* Check whether "expr" is an affine constraint.
2633 * We turn on autodetection so that we won't generate any warnings
2634 * and turn off nesting, so that we won't accept any non-affine constructs.
2636 bool PetScan::is_affine_condition(Expr *expr)
2638 isl_pw_aff *cond;
2639 int save_autodetect = autodetect;
2640 bool save_nesting = nesting_enabled;
2642 autodetect = 1;
2643 nesting_enabled = false;
2645 cond = extract_condition(expr);
2646 isl_pw_aff_free(cond);
2648 autodetect = save_autodetect;
2649 nesting_enabled = save_nesting;
2651 return cond != NULL;
2654 /* Check if we can extract a condition from "expr".
2655 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
2656 * If allow_nested is set, then the condition may involve parameters
2657 * corresponding to nested accesses.
2658 * We turn on autodetection so that we won't generate any warnings.
2660 __isl_give isl_pw_aff *PetScan::try_extract_nested_condition(Expr *expr)
2662 isl_pw_aff *cond;
2663 int save_autodetect = autodetect;
2664 bool save_nesting = nesting_enabled;
2666 autodetect = 1;
2667 nesting_enabled = allow_nested;
2668 cond = extract_condition(expr);
2670 autodetect = save_autodetect;
2671 nesting_enabled = save_nesting;
2673 return cond;
2676 /* If the top-level expression of "stmt" is an assignment, then
2677 * return that assignment as a BinaryOperator.
2678 * Otherwise return NULL.
2680 static BinaryOperator *top_assignment_or_null(Stmt *stmt)
2682 BinaryOperator *ass;
2684 if (!stmt)
2685 return NULL;
2686 if (stmt->getStmtClass() != Stmt::BinaryOperatorClass)
2687 return NULL;
2689 ass = cast<BinaryOperator>(stmt);
2690 if(ass->getOpcode() != BO_Assign)
2691 return NULL;
2693 return ass;
2696 /* Check if the given if statement is a conditional assignement
2697 * with a non-affine condition. If so, construct a pet_scop
2698 * corresponding to this conditional assignment. Otherwise return NULL.
2700 * In particular we check if "stmt" is of the form
2702 * if (condition)
2703 * a = f(...);
2704 * else
2705 * a = g(...);
2707 * where a is some array or scalar access.
2708 * The constructed pet_scop then corresponds to the expression
2710 * a = condition ? f(...) : g(...)
2712 * All access relations in f(...) are intersected with condition
2713 * while all access relation in g(...) are intersected with the complement.
2715 struct pet_scop *PetScan::extract_conditional_assignment(IfStmt *stmt)
2717 BinaryOperator *ass_then, *ass_else;
2718 isl_map *write_then, *write_else;
2719 isl_set *cond, *comp;
2720 isl_map *map;
2721 isl_pw_aff *pa;
2722 int equal;
2723 struct pet_expr *pe_cond, *pe_then, *pe_else, *pe, *pe_write;
2724 bool save_nesting = nesting_enabled;
2726 ass_then = top_assignment_or_null(stmt->getThen());
2727 ass_else = top_assignment_or_null(stmt->getElse());
2729 if (!ass_then || !ass_else)
2730 return NULL;
2732 if (is_affine_condition(stmt->getCond()))
2733 return NULL;
2735 write_then = extract_access(ass_then->getLHS());
2736 write_else = extract_access(ass_else->getLHS());
2738 equal = isl_map_is_equal(write_then, write_else);
2739 isl_map_free(write_else);
2740 if (equal < 0 || !equal) {
2741 isl_map_free(write_then);
2742 return NULL;
2745 nesting_enabled = allow_nested;
2746 pa = extract_condition(stmt->getCond());
2747 nesting_enabled = save_nesting;
2748 cond = isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa));
2749 comp = isl_pw_aff_zero_set(isl_pw_aff_copy(pa));
2750 map = isl_map_from_range(isl_set_from_pw_aff(pa));
2752 pe_cond = pet_expr_from_access(map);
2754 pe_then = extract_expr(ass_then->getRHS());
2755 pe_then = pet_expr_restrict(pe_then, cond);
2756 pe_else = extract_expr(ass_else->getRHS());
2757 pe_else = pet_expr_restrict(pe_else, comp);
2759 pe = pet_expr_new_ternary(ctx, pe_cond, pe_then, pe_else);
2760 pe_write = pet_expr_from_access(write_then);
2761 if (pe_write) {
2762 pe_write->acc.write = 1;
2763 pe_write->acc.read = 0;
2765 pe = pet_expr_new_binary(ctx, pet_op_assign, pe_write, pe);
2766 return extract(stmt, pe);
2769 /* Create an access to a virtual array representing the result
2770 * of a condition.
2771 * Unlike other accessed data, the id of the array is NULL as
2772 * there is no ValueDecl in the program corresponding to the virtual
2773 * array.
2774 * The array starts out as a scalar, but grows along with the
2775 * statement writing to the array in pet_scop_embed.
2777 static __isl_give isl_map *create_test_access(isl_ctx *ctx, int test_nr)
2779 isl_space *dim = isl_space_alloc(ctx, 0, 0, 0);
2780 isl_id *id;
2781 char name[50];
2783 snprintf(name, sizeof(name), "__pet_test_%d", test_nr);
2784 id = isl_id_alloc(ctx, name, NULL);
2785 dim = isl_space_set_tuple_id(dim, isl_dim_out, id);
2786 return isl_map_universe(dim);
2789 /* Create a pet_scop with a single statement evaluating "cond"
2790 * and writing the result to a virtual scalar, as expressed by
2791 * "access".
2793 struct pet_scop *PetScan::extract_non_affine_condition(Expr *cond,
2794 __isl_take isl_map *access)
2796 struct pet_expr *expr, *write;
2797 struct pet_stmt *ps;
2798 SourceLocation loc = cond->getLocStart();
2799 int line = PP.getSourceManager().getExpansionLineNumber(loc);
2801 write = pet_expr_from_access(access);
2802 if (write) {
2803 write->acc.write = 1;
2804 write->acc.read = 0;
2806 expr = extract_expr(cond);
2807 expr = pet_expr_new_binary(ctx, pet_op_assign, write, expr);
2808 ps = pet_stmt_from_pet_expr(ctx, line, NULL, n_stmt++, expr);
2809 return pet_scop_from_pet_stmt(ctx, ps);
2812 /* Add an array with the given extent ("access") to the list
2813 * of arrays in "scop" and return the extended pet_scop.
2814 * The array is marked as attaining values 0 and 1 only.
2816 static struct pet_scop *scop_add_array(struct pet_scop *scop,
2817 __isl_keep isl_map *access, clang::ASTContext &ast_ctx)
2819 isl_ctx *ctx = isl_map_get_ctx(access);
2820 isl_space *dim;
2821 struct pet_array **arrays;
2822 struct pet_array *array;
2824 if (!scop)
2825 return NULL;
2826 if (!ctx)
2827 goto error;
2829 arrays = isl_realloc_array(ctx, scop->arrays, struct pet_array *,
2830 scop->n_array + 1);
2831 if (!arrays)
2832 goto error;
2833 scop->arrays = arrays;
2835 array = isl_calloc_type(ctx, struct pet_array);
2836 if (!array)
2837 goto error;
2839 array->extent = isl_map_range(isl_map_copy(access));
2840 dim = isl_space_params_alloc(ctx, 0);
2841 array->context = isl_set_universe(dim);
2842 dim = isl_space_set_alloc(ctx, 0, 1);
2843 array->value_bounds = isl_set_universe(dim);
2844 array->value_bounds = isl_set_lower_bound_si(array->value_bounds,
2845 isl_dim_set, 0, 0);
2846 array->value_bounds = isl_set_upper_bound_si(array->value_bounds,
2847 isl_dim_set, 0, 1);
2848 array->element_type = strdup("int");
2849 array->element_size = ast_ctx.getTypeInfo(ast_ctx.IntTy).first / 8;
2851 scop->arrays[scop->n_array] = array;
2852 scop->n_array++;
2854 if (!array->extent || !array->context)
2855 goto error;
2857 return scop;
2858 error:
2859 pet_scop_free(scop);
2860 return NULL;
2863 extern "C" {
2864 static __isl_give isl_map *embed_access(__isl_take isl_map *access,
2865 void *user);
2868 /* Apply the map pointed to by "user" to the domain of the access
2869 * relation, thereby embedding it in the range of the map.
2870 * The domain of both relations is the zero-dimensional domain.
2872 static __isl_give isl_map *embed_access(__isl_take isl_map *access, void *user)
2874 isl_map *map = (isl_map *) user;
2876 return isl_map_apply_domain(access, isl_map_copy(map));
2879 /* Apply "map" to all access relations in "expr".
2881 static struct pet_expr *embed(struct pet_expr *expr, __isl_keep isl_map *map)
2883 return pet_expr_foreach_access(expr, &embed_access, map);
2886 /* How many parameters of "set" refer to nested accesses, i.e., have no name?
2888 static int n_nested_parameter(__isl_keep isl_set *set)
2890 isl_space *space;
2891 int n;
2893 space = isl_set_get_space(set);
2894 n = n_nested_parameter(space);
2895 isl_space_free(space);
2897 return n;
2900 /* Remove all parameters from "map" that refer to nested accesses.
2902 static __isl_give isl_map *remove_nested_parameters(__isl_take isl_map *map)
2904 int nparam;
2905 isl_space *space;
2907 space = isl_map_get_space(map);
2908 nparam = isl_space_dim(space, isl_dim_param);
2909 for (int i = nparam - 1; i >= 0; --i)
2910 if (is_nested_parameter(space, i))
2911 map = isl_map_project_out(map, isl_dim_param, i, 1);
2912 isl_space_free(space);
2914 return map;
2917 extern "C" {
2918 static __isl_give isl_map *access_remove_nested_parameters(
2919 __isl_take isl_map *access, void *user);
2922 static __isl_give isl_map *access_remove_nested_parameters(
2923 __isl_take isl_map *access, void *user)
2925 return remove_nested_parameters(access);
2928 /* Remove all nested access parameters from the schedule and all
2929 * accesses of "stmt".
2930 * There is no need to remove them from the domain as these parameters
2931 * have already been removed from the domain when this function is called.
2933 static struct pet_stmt *remove_nested_parameters(struct pet_stmt *stmt)
2935 if (!stmt)
2936 return NULL;
2937 stmt->schedule = remove_nested_parameters(stmt->schedule);
2938 stmt->body = pet_expr_foreach_access(stmt->body,
2939 &access_remove_nested_parameters, NULL);
2940 if (!stmt->schedule || !stmt->body)
2941 goto error;
2942 for (int i = 0; i < stmt->n_arg; ++i) {
2943 stmt->args[i] = pet_expr_foreach_access(stmt->args[i],
2944 &access_remove_nested_parameters, NULL);
2945 if (!stmt->args[i])
2946 goto error;
2949 return stmt;
2950 error:
2951 pet_stmt_free(stmt);
2952 return NULL;
2955 /* For each nested access parameter in the domain of "stmt",
2956 * construct a corresponding pet_expr, place it in stmt->args and
2957 * record its position in "param2pos".
2958 * n is the number of nested access parameters.
2960 struct pet_stmt *PetScan::extract_nested(struct pet_stmt *stmt, int n,
2961 std::map<int,int> &param2pos)
2963 isl_space *space;
2964 unsigned n_arg;
2965 struct pet_expr **args;
2967 n_arg = stmt->n_arg;
2968 args = isl_realloc_array(ctx, stmt->args, struct pet_expr *, n_arg + n);
2969 if (!args)
2970 goto error;
2971 stmt->args = args;
2972 stmt->n_arg += n;
2974 space = isl_set_get_space(stmt->domain);
2975 n = extract_nested(space, n_arg, stmt->args, param2pos);
2976 isl_space_free(space);
2978 if (n < 0)
2979 goto error;
2981 stmt->n_arg = n;
2982 return stmt;
2983 error:
2984 pet_stmt_free(stmt);
2985 return NULL;
2988 /* Look for parameters in the iteration domain of "stmt" that
2989 * refer to nested accesses. In particular, these are
2990 * parameters with no name.
2992 * If there are any such parameters, then as many extra variables
2993 * (after identifying identical nested accesses) are added to the
2994 * range of the map wrapped inside the domain.
2995 * If the original domain is not a wrapped map, then a new wrapped
2996 * map is created with zero output dimensions.
2997 * The parameters are then equated to the corresponding output dimensions
2998 * and subsequently projected out, from the iteration domain,
2999 * the schedule and the access relations.
3000 * For each of the output dimensions, a corresponding argument
3001 * expression is added. Initially they are created with
3002 * a zero-dimensional domain, so they have to be embedded
3003 * in the current iteration domain.
3004 * param2pos maps the position of the parameter to the position
3005 * of the corresponding output dimension in the wrapped map.
3007 struct pet_stmt *PetScan::resolve_nested(struct pet_stmt *stmt)
3009 int n;
3010 int nparam;
3011 unsigned n_arg;
3012 isl_map *map;
3013 std::map<int,int> param2pos;
3015 if (!stmt)
3016 return NULL;
3018 n = n_nested_parameter(stmt->domain);
3019 if (n == 0)
3020 return stmt;
3022 n_arg = stmt->n_arg;
3023 stmt = extract_nested(stmt, n, param2pos);
3024 if (!stmt)
3025 return NULL;
3027 n = stmt->n_arg - n_arg;
3028 nparam = isl_set_dim(stmt->domain, isl_dim_param);
3029 if (isl_set_is_wrapping(stmt->domain))
3030 map = isl_set_unwrap(stmt->domain);
3031 else
3032 map = isl_map_from_domain(stmt->domain);
3033 map = isl_map_add_dims(map, isl_dim_out, n);
3035 for (int i = nparam - 1; i >= 0; --i) {
3036 isl_id *id;
3038 if (!is_nested_parameter(map, i))
3039 continue;
3041 id = isl_map_get_tuple_id(stmt->args[param2pos[i]]->acc.access,
3042 isl_dim_out);
3043 map = isl_map_set_dim_id(map, isl_dim_out, param2pos[i], id);
3044 map = isl_map_equate(map, isl_dim_param, i, isl_dim_out,
3045 param2pos[i]);
3046 map = isl_map_project_out(map, isl_dim_param, i, 1);
3049 stmt->domain = isl_map_wrap(map);
3051 map = isl_set_unwrap(isl_set_copy(stmt->domain));
3052 map = isl_map_from_range(isl_map_domain(map));
3053 for (int pos = n_arg; pos < stmt->n_arg; ++pos)
3054 stmt->args[pos] = embed(stmt->args[pos], map);
3055 isl_map_free(map);
3057 stmt = remove_nested_parameters(stmt);
3059 return stmt;
3060 error:
3061 pet_stmt_free(stmt);
3062 return NULL;
3065 /* For each statement in "scop", move the parameters that correspond
3066 * to nested access into the ranges of the domains and create
3067 * corresponding argument expressions.
3069 struct pet_scop *PetScan::resolve_nested(struct pet_scop *scop)
3071 if (!scop)
3072 return NULL;
3074 for (int i = 0; i < scop->n_stmt; ++i) {
3075 scop->stmts[i] = resolve_nested(scop->stmts[i]);
3076 if (!scop->stmts[i])
3077 goto error;
3080 return scop;
3081 error:
3082 pet_scop_free(scop);
3083 return NULL;
3086 /* Does "space" involve any parameters that refer to nested
3087 * accesses, i.e., parameters with no name?
3089 static bool has_nested(__isl_keep isl_space *space)
3091 int nparam;
3093 nparam = isl_space_dim(space, isl_dim_param);
3094 for (int i = 0; i < nparam; ++i)
3095 if (is_nested_parameter(space, i))
3096 return true;
3098 return false;
3101 /* Does "pa" involve any parameters that refer to nested
3102 * accesses, i.e., parameters with no name?
3104 static bool has_nested(__isl_keep isl_pw_aff *pa)
3106 isl_space *space;
3107 bool nested;
3109 space = isl_pw_aff_get_space(pa);
3110 nested = has_nested(space);
3111 isl_space_free(space);
3113 return nested;
3116 /* Given an access expression "expr", is the variable accessed by
3117 * "expr" assigned anywhere inside "scop"?
3119 static bool is_assigned(pet_expr *expr, pet_scop *scop)
3121 bool assigned = false;
3122 isl_id *id;
3124 id = isl_map_get_tuple_id(expr->acc.access, isl_dim_out);
3125 assigned = pet_scop_writes(scop, id);
3126 isl_id_free(id);
3128 return assigned;
3131 /* Are all nested access parameters in "pa" allowed given "scop".
3132 * In particular, is none of them written by anywhere inside "scop".
3134 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff *pa, pet_scop *scop)
3136 int nparam;
3138 nparam = isl_pw_aff_dim(pa, isl_dim_param);
3139 for (int i = 0; i < nparam; ++i) {
3140 Expr *nested;
3141 isl_id *id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
3142 pet_expr *expr;
3143 bool allowed;
3145 if (!is_nested_parameter(id)) {
3146 isl_id_free(id);
3147 continue;
3150 nested = (Expr *) isl_id_get_user(id);
3151 expr = extract_expr(nested);
3152 allowed = expr && expr->type == pet_expr_access &&
3153 !is_assigned(expr, scop);
3155 pet_expr_free(expr);
3156 isl_id_free(id);
3158 if (!allowed)
3159 return false;
3162 return true;
3165 /* Construct a pet_scop for an if statement.
3167 * If the condition fits the pattern of a conditional assignment,
3168 * then it is handled by extract_conditional_assignment.
3169 * Otherwise, we do the following.
3171 * If the condition is affine, then the condition is added
3172 * to the iteration domains of the then branch, while the
3173 * opposite of the condition in added to the iteration domains
3174 * of the else branch, if any.
3175 * We allow the condition to be dynamic, i.e., to refer to
3176 * scalars or array elements that may be written to outside
3177 * of the given if statement. These nested accesses are then represented
3178 * as output dimensions in the wrapping iteration domain.
3179 * If it also written _inside_ the then or else branch, then
3180 * we treat the condition as non-affine.
3181 * As explained below, this will introduce an extra statement.
3182 * For aesthetic reasons, we want this statement to have a statement
3183 * number that is lower than those of the then and else branches.
3184 * In order to evaluate if will need such a statement, however, we
3185 * first construct scops for the then and else branches.
3186 * We therefore reserve a statement number if we might have to
3187 * introduce such an extra statement.
3189 * If the condition is not affine, then we create a separate
3190 * statement that writes the result of the condition to a virtual scalar.
3191 * A constraint requiring the value of this virtual scalar to be one
3192 * is added to the iteration domains of the then branch.
3193 * Similarly, a constraint requiring the value of this virtual scalar
3194 * to be zero is added to the iteration domains of the else branch, if any.
3195 * We adjust the schedules to ensure that the virtual scalar is written
3196 * before it is read.
3198 struct pet_scop *PetScan::extract(IfStmt *stmt)
3200 struct pet_scop *scop_then, *scop_else, *scop;
3201 assigned_value_cache cache(assigned_value);
3202 isl_map *test_access = NULL;
3203 isl_pw_aff *cond;
3204 int stmt_id;
3206 scop = extract_conditional_assignment(stmt);
3207 if (scop)
3208 return scop;
3210 cond = try_extract_nested_condition(stmt->getCond());
3211 if (allow_nested && (!cond || has_nested(cond)))
3212 stmt_id = n_stmt++;
3214 scop_then = extract(stmt->getThen());
3216 if (stmt->getElse()) {
3217 scop_else = extract(stmt->getElse());
3218 if (autodetect) {
3219 if (scop_then && !scop_else) {
3220 partial = true;
3221 isl_pw_aff_free(cond);
3222 return scop_then;
3224 if (!scop_then && scop_else) {
3225 partial = true;
3226 isl_pw_aff_free(cond);
3227 return scop_else;
3232 if (cond &&
3233 (!is_nested_allowed(cond, scop_then) ||
3234 (stmt->getElse() && !is_nested_allowed(cond, scop_else)))) {
3235 isl_pw_aff_free(cond);
3236 cond = NULL;
3238 if (allow_nested && !cond) {
3239 int save_n_stmt = n_stmt;
3240 test_access = create_test_access(ctx, n_test++);
3241 n_stmt = stmt_id;
3242 scop = extract_non_affine_condition(stmt->getCond(),
3243 isl_map_copy(test_access));
3244 n_stmt = save_n_stmt;
3245 scop = scop_add_array(scop, test_access, ast_context);
3246 if (!scop) {
3247 pet_scop_free(scop_then);
3248 pet_scop_free(scop_else);
3249 isl_map_free(test_access);
3250 return NULL;
3254 if (!scop) {
3255 isl_set *set;
3256 isl_set *valid;
3258 if (!cond)
3259 cond = extract_condition(stmt->getCond());
3260 valid = isl_pw_aff_domain(isl_pw_aff_copy(cond));
3261 set = isl_pw_aff_non_zero_set(cond);
3262 scop = pet_scop_restrict(scop_then, isl_set_copy(set));
3264 if (stmt->getElse()) {
3265 set = isl_set_subtract(isl_set_copy(valid), set);
3266 scop_else = pet_scop_restrict(scop_else, set);
3267 scop = pet_scop_add(ctx, scop, scop_else);
3268 } else
3269 isl_set_free(set);
3270 scop = resolve_nested(scop);
3271 scop = pet_scop_restrict_context(scop, valid);
3272 } else {
3273 scop = pet_scop_prefix(scop, 0);
3274 scop_then = pet_scop_prefix(scop_then, 1);
3275 scop_then = pet_scop_filter(scop_then,
3276 isl_map_copy(test_access), 1);
3277 scop = pet_scop_add(ctx, scop, scop_then);
3278 if (stmt->getElse()) {
3279 scop_else = pet_scop_prefix(scop_else, 1);
3280 scop_else = pet_scop_filter(scop_else, test_access, 0);
3281 scop = pet_scop_add(ctx, scop, scop_else);
3282 } else
3283 isl_map_free(test_access);
3286 return scop;
3289 /* Try and construct a pet_scop for a label statement.
3290 * We currently only allow labels on expression statements.
3292 struct pet_scop *PetScan::extract(LabelStmt *stmt)
3294 isl_id *label;
3295 Stmt *sub;
3297 sub = stmt->getSubStmt();
3298 if (!isa<Expr>(sub)) {
3299 unsupported(stmt);
3300 return NULL;
3303 label = isl_id_alloc(ctx, stmt->getName(), NULL);
3305 return extract(sub, extract_expr(cast<Expr>(sub)), label);
3308 /* Try and construct a pet_scop corresponding to "stmt".
3310 struct pet_scop *PetScan::extract(Stmt *stmt)
3312 if (isa<Expr>(stmt))
3313 return extract(stmt, extract_expr(cast<Expr>(stmt)));
3315 switch (stmt->getStmtClass()) {
3316 case Stmt::WhileStmtClass:
3317 return extract(cast<WhileStmt>(stmt));
3318 case Stmt::ForStmtClass:
3319 return extract_for(cast<ForStmt>(stmt));
3320 case Stmt::IfStmtClass:
3321 return extract(cast<IfStmt>(stmt));
3322 case Stmt::CompoundStmtClass:
3323 return extract(cast<CompoundStmt>(stmt));
3324 case Stmt::LabelStmtClass:
3325 return extract(cast<LabelStmt>(stmt));
3326 default:
3327 unsupported(stmt);
3330 return NULL;
3333 /* Try and construct a pet_scop corresponding to (part of)
3334 * a sequence of statements.
3336 struct pet_scop *PetScan::extract(StmtRange stmt_range)
3338 pet_scop *scop;
3339 StmtIterator i;
3340 int j;
3341 bool partial_range = false;
3343 scop = pet_scop_empty(ctx);
3344 for (i = stmt_range.first, j = 0; i != stmt_range.second; ++i, ++j) {
3345 Stmt *child = *i;
3346 struct pet_scop *scop_i;
3347 scop_i = extract(child);
3348 if (scop && partial) {
3349 pet_scop_free(scop_i);
3350 break;
3352 scop_i = pet_scop_prefix(scop_i, j);
3353 if (autodetect) {
3354 if (scop_i)
3355 scop = pet_scop_add(ctx, scop, scop_i);
3356 else
3357 partial_range = true;
3358 if (scop->n_stmt != 0 && !scop_i)
3359 partial = true;
3360 } else {
3361 scop = pet_scop_add(ctx, scop, scop_i);
3363 if (partial)
3364 break;
3367 if (scop && partial_range)
3368 partial = true;
3370 return scop;
3373 /* Check if the scop marked by the user is exactly this Stmt
3374 * or part of this Stmt.
3375 * If so, return a pet_scop corresponding to the marked region.
3376 * Otherwise, return NULL.
3378 struct pet_scop *PetScan::scan(Stmt *stmt)
3380 SourceManager &SM = PP.getSourceManager();
3381 unsigned start_off, end_off;
3383 start_off = SM.getFileOffset(stmt->getLocStart());
3384 end_off = SM.getFileOffset(stmt->getLocEnd());
3386 if (start_off > loc.end)
3387 return NULL;
3388 if (end_off < loc.start)
3389 return NULL;
3390 if (start_off >= loc.start && end_off <= loc.end) {
3391 return extract(stmt);
3394 StmtIterator start;
3395 for (start = stmt->child_begin(); start != stmt->child_end(); ++start) {
3396 Stmt *child = *start;
3397 if (!child)
3398 continue;
3399 start_off = SM.getFileOffset(child->getLocStart());
3400 end_off = SM.getFileOffset(child->getLocEnd());
3401 if (start_off < loc.start && end_off > loc.end)
3402 return scan(child);
3403 if (start_off >= loc.start)
3404 break;
3407 StmtIterator end;
3408 for (end = start; end != stmt->child_end(); ++end) {
3409 Stmt *child = *end;
3410 start_off = SM.getFileOffset(child->getLocStart());
3411 if (start_off >= loc.end)
3412 break;
3415 return extract(StmtRange(start, end));
3418 /* Set the size of index "pos" of "array" to "size".
3419 * In particular, add a constraint of the form
3421 * i_pos < size
3423 * to array->extent and a constraint of the form
3425 * size >= 0
3427 * to array->context.
3429 static struct pet_array *update_size(struct pet_array *array, int pos,
3430 __isl_take isl_pw_aff *size)
3432 isl_set *valid;
3433 isl_set *univ;
3434 isl_set *bound;
3435 isl_space *dim;
3436 isl_aff *aff;
3437 isl_pw_aff *index;
3438 isl_id *id;
3440 valid = isl_pw_aff_nonneg_set(isl_pw_aff_copy(size));
3441 array->context = isl_set_intersect(array->context, valid);
3443 dim = isl_set_get_space(array->extent);
3444 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
3445 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, pos, 1);
3446 univ = isl_set_universe(isl_aff_get_domain_space(aff));
3447 index = isl_pw_aff_alloc(univ, aff);
3449 size = isl_pw_aff_add_dims(size, isl_dim_in,
3450 isl_set_dim(array->extent, isl_dim_set));
3451 id = isl_set_get_tuple_id(array->extent);
3452 size = isl_pw_aff_set_tuple_id(size, isl_dim_in, id);
3453 bound = isl_pw_aff_lt_set(index, size);
3455 array->extent = isl_set_intersect(array->extent, bound);
3457 if (!array->context || !array->extent)
3458 goto error;
3460 return array;
3461 error:
3462 pet_array_free(array);
3463 return NULL;
3466 /* Figure out the size of the array at position "pos" and all
3467 * subsequent positions from "type" and update "array" accordingly.
3469 struct pet_array *PetScan::set_upper_bounds(struct pet_array *array,
3470 const Type *type, int pos)
3472 const ArrayType *atype;
3473 isl_pw_aff *size;
3475 if (!array)
3476 return NULL;
3478 if (type->isPointerType()) {
3479 type = type->getPointeeType().getTypePtr();
3480 return set_upper_bounds(array, type, pos + 1);
3482 if (!type->isArrayType())
3483 return array;
3485 type = type->getCanonicalTypeInternal().getTypePtr();
3486 atype = cast<ArrayType>(type);
3488 if (type->isConstantArrayType()) {
3489 const ConstantArrayType *ca = cast<ConstantArrayType>(atype);
3490 size = extract_affine(ca->getSize());
3491 array = update_size(array, pos, size);
3492 } else if (type->isVariableArrayType()) {
3493 const VariableArrayType *vla = cast<VariableArrayType>(atype);
3494 size = extract_affine(vla->getSizeExpr());
3495 array = update_size(array, pos, size);
3498 type = atype->getElementType().getTypePtr();
3500 return set_upper_bounds(array, type, pos + 1);
3503 /* Construct and return a pet_array corresponding to the variable "decl".
3504 * In particular, initialize array->extent to
3506 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
3508 * and then call set_upper_bounds to set the upper bounds on the indices
3509 * based on the type of the variable.
3511 struct pet_array *PetScan::extract_array(isl_ctx *ctx, ValueDecl *decl)
3513 struct pet_array *array;
3514 QualType qt = decl->getType();
3515 const Type *type = qt.getTypePtr();
3516 int depth = array_depth(type);
3517 QualType base = base_type(qt);
3518 string name;
3519 isl_id *id;
3520 isl_space *dim;
3522 array = isl_calloc_type(ctx, struct pet_array);
3523 if (!array)
3524 return NULL;
3526 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
3527 dim = isl_space_set_alloc(ctx, 0, depth);
3528 dim = isl_space_set_tuple_id(dim, isl_dim_set, id);
3530 array->extent = isl_set_nat_universe(dim);
3532 dim = isl_space_params_alloc(ctx, 0);
3533 array->context = isl_set_universe(dim);
3535 array = set_upper_bounds(array, type, 0);
3536 if (!array)
3537 return NULL;
3539 name = base.getAsString();
3540 array->element_type = strdup(name.c_str());
3541 array->element_size = decl->getASTContext().getTypeInfo(base).first / 8;
3543 return array;
3546 /* Construct a list of pet_arrays, one for each array (or scalar)
3547 * accessed inside "scop" add this list to "scop" and return the result.
3549 * The context of "scop" is updated with the intesection of
3550 * the contexts of all arrays, i.e., constraints on the parameters
3551 * that ensure that the arrays have a valid (non-negative) size.
3553 struct pet_scop *PetScan::scan_arrays(struct pet_scop *scop)
3555 int i;
3556 set<ValueDecl *> arrays;
3557 set<ValueDecl *>::iterator it;
3558 int n_array;
3559 struct pet_array **scop_arrays;
3561 if (!scop)
3562 return NULL;
3564 pet_scop_collect_arrays(scop, arrays);
3565 if (arrays.size() == 0)
3566 return scop;
3568 n_array = scop->n_array;
3570 scop_arrays = isl_realloc_array(ctx, scop->arrays, struct pet_array *,
3571 n_array + arrays.size());
3572 if (!scop_arrays)
3573 goto error;
3574 scop->arrays = scop_arrays;
3576 for (it = arrays.begin(), i = 0; it != arrays.end(); ++it, ++i) {
3577 struct pet_array *array;
3578 scop->arrays[n_array + i] = array = extract_array(ctx, *it);
3579 if (!scop->arrays[n_array + i])
3580 goto error;
3581 scop->n_array++;
3582 scop->context = isl_set_intersect(scop->context,
3583 isl_set_copy(array->context));
3584 if (!scop->context)
3585 goto error;
3588 return scop;
3589 error:
3590 pet_scop_free(scop);
3591 return NULL;
3594 /* Bound all parameters in scop->context to the possible values
3595 * of the corresponding C variable.
3597 static struct pet_scop *add_parameter_bounds(struct pet_scop *scop)
3599 int n;
3601 if (!scop)
3602 return NULL;
3604 n = isl_set_dim(scop->context, isl_dim_param);
3605 for (int i = 0; i < n; ++i) {
3606 isl_id *id;
3607 ValueDecl *decl;
3609 id = isl_set_get_dim_id(scop->context, isl_dim_param, i);
3610 decl = (ValueDecl *) isl_id_get_user(id);
3611 isl_id_free(id);
3613 scop->context = set_parameter_bounds(scop->context, i, decl);
3615 if (!scop->context)
3616 goto error;
3619 return scop;
3620 error:
3621 pet_scop_free(scop);
3622 return NULL;
3625 /* Construct a pet_scop from the given function.
3627 struct pet_scop *PetScan::scan(FunctionDecl *fd)
3629 pet_scop *scop;
3630 Stmt *stmt;
3632 stmt = fd->getBody();
3634 if (autodetect)
3635 scop = extract(stmt);
3636 else
3637 scop = scan(stmt);
3638 scop = pet_scop_detect_parameter_accesses(scop);
3639 scop = scan_arrays(scop);
3640 scop = add_parameter_bounds(scop);
3641 scop = pet_scop_gist(scop, value_bounds);
3643 return scop;