PetScan::scan: use expansion offset to determine if statement is inside scop
[pet.git] / scan.cc
blob3bbc81223d3f0f68059d4f04a7c057a86857b9e8
1 /*
2 * Copyright 2011 Leiden University. All rights reserved.
3 * Copyright 2012 Ecole Normale Superieure. All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 *
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above
13 * copyright notice, this list of conditions and the following
14 * disclaimer in the documentation and/or other materials provided
15 * with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY LEIDEN UNIVERSITY ''AS IS'' AND ANY
18 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL LEIDEN UNIVERSITY OR
21 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
22 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
23 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA,
24 * OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
27 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 * The views and conclusions contained in the software and documentation
30 * are those of the authors and should not be interpreted as
31 * representing official policies, either expressed or implied, of
32 * Leiden University.
33 */
35 #include <set>
36 #include <map>
37 #include <iostream>
38 #include <clang/AST/ASTContext.h>
39 #include <clang/AST/ASTDiagnostic.h>
40 #include <clang/AST/Expr.h>
41 #include <clang/AST/RecursiveASTVisitor.h>
43 #include <isl/id.h>
44 #include <isl/space.h>
45 #include <isl/aff.h>
46 #include <isl/set.h>
48 #include "options.h"
49 #include "scan.h"
50 #include "scop.h"
51 #include "scop_plus.h"
53 #include "config.h"
55 using namespace std;
56 using namespace clang;
58 #if defined(DECLREFEXPR_CREATE_REQUIRES_BOOL)
59 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
61 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
62 SourceLocation(), var, false, var->getInnerLocStart(),
63 var->getType(), VK_LValue);
65 #elif defined(DECLREFEXPR_CREATE_REQUIRES_SOURCELOCATION)
66 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
68 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
69 SourceLocation(), var, var->getInnerLocStart(), var->getType(),
70 VK_LValue);
72 #else
73 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
75 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
76 var, var->getInnerLocStart(), var->getType(), VK_LValue);
78 #endif
80 /* Check if the element type corresponding to the given array type
81 * has a const qualifier.
83 static bool const_base(QualType qt)
85 const Type *type = qt.getTypePtr();
87 if (type->isPointerType())
88 return const_base(type->getPointeeType());
89 if (type->isArrayType()) {
90 const ArrayType *atype;
91 type = type->getCanonicalTypeInternal().getTypePtr();
92 atype = cast<ArrayType>(type);
93 return const_base(atype->getElementType());
96 return qt.isConstQualified();
99 /* Mark "decl" as having an unknown value in "assigned_value".
101 * If no (known or unknown) value was assigned to "decl" before,
102 * then it may have been treated as a parameter before and may
103 * therefore appear in a value assigned to another variable.
104 * If so, this assignment needs to be turned into an unknown value too.
106 static void clear_assignment(map<ValueDecl *, isl_pw_aff *> &assigned_value,
107 ValueDecl *decl)
109 map<ValueDecl *, isl_pw_aff *>::iterator it;
111 it = assigned_value.find(decl);
113 assigned_value[decl] = NULL;
115 if (it == assigned_value.end())
116 return;
118 for (it = assigned_value.begin(); it != assigned_value.end(); ++it) {
119 isl_pw_aff *pa = it->second;
120 int nparam = isl_pw_aff_dim(pa, isl_dim_param);
122 for (int i = 0; i < nparam; ++i) {
123 isl_id *id;
125 if (!isl_pw_aff_has_dim_id(pa, isl_dim_param, i))
126 continue;
127 id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
128 if (isl_id_get_user(id) == decl)
129 it->second = NULL;
130 isl_id_free(id);
135 /* Look for any assignments to scalar variables in part of the parse
136 * tree and set assigned_value to NULL for each of them.
137 * Also reset assigned_value if the address of a scalar variable
138 * is being taken. As an exception, if the address is passed to a function
139 * that is declared to receive a const pointer, then assigned_value is
140 * not reset.
142 * This ensures that we won't use any previously stored value
143 * in the current subtree and its parents.
145 struct clear_assignments : RecursiveASTVisitor<clear_assignments> {
146 map<ValueDecl *, isl_pw_aff *> &assigned_value;
147 set<UnaryOperator *> skip;
149 clear_assignments(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
150 assigned_value(assigned_value) {}
152 /* Check for "address of" operators whose value is passed
153 * to a const pointer argument and add them to "skip", so that
154 * we can skip them in VisitUnaryOperator.
156 bool VisitCallExpr(CallExpr *expr) {
157 FunctionDecl *fd;
158 fd = expr->getDirectCallee();
159 if (!fd)
160 return true;
161 for (int i = 0; i < expr->getNumArgs(); ++i) {
162 Expr *arg = expr->getArg(i);
163 UnaryOperator *op;
164 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
165 ImplicitCastExpr *ice;
166 ice = cast<ImplicitCastExpr>(arg);
167 arg = ice->getSubExpr();
169 if (arg->getStmtClass() != Stmt::UnaryOperatorClass)
170 continue;
171 op = cast<UnaryOperator>(arg);
172 if (op->getOpcode() != UO_AddrOf)
173 continue;
174 if (const_base(fd->getParamDecl(i)->getType()))
175 skip.insert(op);
177 return true;
180 bool VisitUnaryOperator(UnaryOperator *expr) {
181 Expr *arg;
182 DeclRefExpr *ref;
183 ValueDecl *decl;
185 if (expr->getOpcode() != UO_AddrOf)
186 return true;
187 if (skip.find(expr) != skip.end())
188 return true;
190 arg = expr->getSubExpr();
191 if (arg->getStmtClass() != Stmt::DeclRefExprClass)
192 return true;
193 ref = cast<DeclRefExpr>(arg);
194 decl = ref->getDecl();
195 clear_assignment(assigned_value, decl);
196 return true;
199 bool VisitBinaryOperator(BinaryOperator *expr) {
200 Expr *lhs;
201 DeclRefExpr *ref;
202 ValueDecl *decl;
204 if (!expr->isAssignmentOp())
205 return true;
206 lhs = expr->getLHS();
207 if (lhs->getStmtClass() != Stmt::DeclRefExprClass)
208 return true;
209 ref = cast<DeclRefExpr>(lhs);
210 decl = ref->getDecl();
211 clear_assignment(assigned_value, decl);
212 return true;
216 /* Keep a copy of the currently assigned values.
218 * Any variable that is assigned a value inside the current scope
219 * is removed again when we leave the scope (either because it wasn't
220 * stored in the cache or because it has a different value in the cache).
222 struct assigned_value_cache {
223 map<ValueDecl *, isl_pw_aff *> &assigned_value;
224 map<ValueDecl *, isl_pw_aff *> cache;
226 assigned_value_cache(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
227 assigned_value(assigned_value), cache(assigned_value) {}
228 ~assigned_value_cache() {
229 map<ValueDecl *, isl_pw_aff *>::iterator it = cache.begin();
230 for (it = assigned_value.begin(); it != assigned_value.end();
231 ++it) {
232 if (!it->second ||
233 (cache.find(it->first) != cache.end() &&
234 cache[it->first] != it->second))
235 cache[it->first] = NULL;
237 assigned_value = cache;
241 /* Insert an expression into the collection of expressions,
242 * provided it is not already in there.
243 * The isl_pw_affs are freed in the destructor.
245 void PetScan::insert_expression(__isl_take isl_pw_aff *expr)
247 std::set<isl_pw_aff *>::iterator it;
249 if (expressions.find(expr) == expressions.end())
250 expressions.insert(expr);
251 else
252 isl_pw_aff_free(expr);
255 PetScan::~PetScan()
257 std::set<isl_pw_aff *>::iterator it;
259 for (it = expressions.begin(); it != expressions.end(); ++it)
260 isl_pw_aff_free(*it);
262 isl_union_map_free(value_bounds);
265 /* Called if we found something we (currently) cannot handle.
266 * We'll provide more informative warnings later.
268 * We only actually complain if autodetect is false.
270 void PetScan::unsupported(Stmt *stmt, const char *msg)
272 if (options->autodetect)
273 return;
275 SourceLocation loc = stmt->getLocStart();
276 DiagnosticsEngine &diag = PP.getDiagnostics();
277 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
278 msg ? msg : "unsupported");
279 DiagnosticBuilder B = diag.Report(loc, id) << stmt->getSourceRange();
282 /* Extract an integer from "expr" and store it in "v".
284 int PetScan::extract_int(IntegerLiteral *expr, isl_int *v)
286 const Type *type = expr->getType().getTypePtr();
287 int is_signed = type->hasSignedIntegerRepresentation();
289 if (is_signed) {
290 int64_t i = expr->getValue().getSExtValue();
291 isl_int_set_si(*v, i);
292 } else {
293 uint64_t i = expr->getValue().getZExtValue();
294 isl_int_set_ui(*v, i);
297 return 0;
300 /* Extract an integer from "expr" and store it in "v".
301 * Return -1 if "expr" does not (obviously) represent an integer.
303 int PetScan::extract_int(clang::ParenExpr *expr, isl_int *v)
305 return extract_int(expr->getSubExpr(), v);
308 /* Extract an integer from "expr" and store it in "v".
309 * Return -1 if "expr" does not (obviously) represent an integer.
311 int PetScan::extract_int(clang::Expr *expr, isl_int *v)
313 if (expr->getStmtClass() == Stmt::IntegerLiteralClass)
314 return extract_int(cast<IntegerLiteral>(expr), v);
315 if (expr->getStmtClass() == Stmt::ParenExprClass)
316 return extract_int(cast<ParenExpr>(expr), v);
318 unsupported(expr);
319 return -1;
322 /* Extract an affine expression from the IntegerLiteral "expr".
324 __isl_give isl_pw_aff *PetScan::extract_affine(IntegerLiteral *expr)
326 isl_space *dim = isl_space_params_alloc(ctx, 0);
327 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
328 isl_aff *aff = isl_aff_zero_on_domain(ls);
329 isl_set *dom = isl_set_universe(dim);
330 isl_int v;
332 isl_int_init(v);
333 extract_int(expr, &v);
334 aff = isl_aff_add_constant(aff, v);
335 isl_int_clear(v);
337 return isl_pw_aff_alloc(dom, aff);
340 /* Extract an affine expression from the APInt "val".
342 __isl_give isl_pw_aff *PetScan::extract_affine(const llvm::APInt &val)
344 isl_space *dim = isl_space_params_alloc(ctx, 0);
345 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
346 isl_aff *aff = isl_aff_zero_on_domain(ls);
347 isl_set *dom = isl_set_universe(dim);
348 isl_int v;
350 isl_int_init(v);
351 isl_int_set_ui(v, val.getZExtValue());
352 aff = isl_aff_add_constant(aff, v);
353 isl_int_clear(v);
355 return isl_pw_aff_alloc(dom, aff);
358 __isl_give isl_pw_aff *PetScan::extract_affine(ImplicitCastExpr *expr)
360 return extract_affine(expr->getSubExpr());
363 static unsigned get_type_size(ValueDecl *decl)
365 return decl->getASTContext().getIntWidth(decl->getType());
368 /* Bound parameter "pos" of "set" to the possible values of "decl".
370 static __isl_give isl_set *set_parameter_bounds(__isl_take isl_set *set,
371 unsigned pos, ValueDecl *decl)
373 unsigned width;
374 isl_int v;
376 isl_int_init(v);
378 width = get_type_size(decl);
379 if (decl->getType()->isUnsignedIntegerType()) {
380 set = isl_set_lower_bound_si(set, isl_dim_param, pos, 0);
381 isl_int_set_si(v, 1);
382 isl_int_mul_2exp(v, v, width);
383 isl_int_sub_ui(v, v, 1);
384 set = isl_set_upper_bound(set, isl_dim_param, pos, v);
385 } else {
386 isl_int_set_si(v, 1);
387 isl_int_mul_2exp(v, v, width - 1);
388 isl_int_sub_ui(v, v, 1);
389 set = isl_set_upper_bound(set, isl_dim_param, pos, v);
390 isl_int_neg(v, v);
391 isl_int_sub_ui(v, v, 1);
392 set = isl_set_lower_bound(set, isl_dim_param, pos, v);
395 isl_int_clear(v);
397 return set;
400 /* Extract an affine expression from the DeclRefExpr "expr".
402 * If the variable has been assigned a value, then we check whether
403 * we know what (affine) value was assigned.
404 * If so, we return this value. Otherwise we convert "expr"
405 * to an extra parameter (provided nesting_enabled is set).
407 * Otherwise, we simply return an expression that is equal
408 * to a parameter corresponding to the referenced variable.
410 __isl_give isl_pw_aff *PetScan::extract_affine(DeclRefExpr *expr)
412 ValueDecl *decl = expr->getDecl();
413 const Type *type = decl->getType().getTypePtr();
414 isl_id *id;
415 isl_space *dim;
416 isl_aff *aff;
417 isl_set *dom;
419 if (!type->isIntegerType()) {
420 unsupported(expr);
421 return NULL;
424 if (assigned_value.find(decl) != assigned_value.end()) {
425 if (assigned_value[decl])
426 return isl_pw_aff_copy(assigned_value[decl]);
427 else
428 return nested_access(expr);
431 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
432 dim = isl_space_params_alloc(ctx, 1);
434 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
436 dom = isl_set_universe(isl_space_copy(dim));
437 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
438 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
440 return isl_pw_aff_alloc(dom, aff);
443 /* Extract an affine expression from an integer division operation.
444 * In particular, if "expr" is lhs/rhs, then return
446 * lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs)
448 * The second argument (rhs) is required to be a (positive) integer constant.
450 __isl_give isl_pw_aff *PetScan::extract_affine_div(BinaryOperator *expr)
452 int is_cst;
453 isl_pw_aff *rhs, *lhs;
455 rhs = extract_affine(expr->getRHS());
456 is_cst = isl_pw_aff_is_cst(rhs);
457 if (is_cst < 0 || !is_cst) {
458 isl_pw_aff_free(rhs);
459 if (!is_cst)
460 unsupported(expr);
461 return NULL;
464 lhs = extract_affine(expr->getLHS());
466 return isl_pw_aff_tdiv_q(lhs, rhs);
469 /* Extract an affine expression from a modulo operation.
470 * In particular, if "expr" is lhs/rhs, then return
472 * lhs - rhs * (lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs))
474 * The second argument (rhs) is required to be a (positive) integer constant.
476 __isl_give isl_pw_aff *PetScan::extract_affine_mod(BinaryOperator *expr)
478 int is_cst;
479 isl_pw_aff *rhs, *lhs;
481 rhs = extract_affine(expr->getRHS());
482 is_cst = isl_pw_aff_is_cst(rhs);
483 if (is_cst < 0 || !is_cst) {
484 isl_pw_aff_free(rhs);
485 if (!is_cst)
486 unsupported(expr);
487 return NULL;
490 lhs = extract_affine(expr->getLHS());
492 return isl_pw_aff_tdiv_r(lhs, rhs);
495 /* Extract an affine expression from a multiplication operation.
496 * This is only allowed if at least one of the two arguments
497 * is a (piecewise) constant.
499 __isl_give isl_pw_aff *PetScan::extract_affine_mul(BinaryOperator *expr)
501 isl_pw_aff *lhs;
502 isl_pw_aff *rhs;
504 lhs = extract_affine(expr->getLHS());
505 rhs = extract_affine(expr->getRHS());
507 if (!isl_pw_aff_is_cst(lhs) && !isl_pw_aff_is_cst(rhs)) {
508 isl_pw_aff_free(lhs);
509 isl_pw_aff_free(rhs);
510 unsupported(expr);
511 return NULL;
514 return isl_pw_aff_mul(lhs, rhs);
517 /* Extract an affine expression from an addition or subtraction operation.
519 __isl_give isl_pw_aff *PetScan::extract_affine_add(BinaryOperator *expr)
521 isl_pw_aff *lhs;
522 isl_pw_aff *rhs;
524 lhs = extract_affine(expr->getLHS());
525 rhs = extract_affine(expr->getRHS());
527 switch (expr->getOpcode()) {
528 case BO_Add:
529 return isl_pw_aff_add(lhs, rhs);
530 case BO_Sub:
531 return isl_pw_aff_sub(lhs, rhs);
532 default:
533 isl_pw_aff_free(lhs);
534 isl_pw_aff_free(rhs);
535 return NULL;
540 /* Compute
542 * pwaff mod 2^width
544 static __isl_give isl_pw_aff *wrap(__isl_take isl_pw_aff *pwaff,
545 unsigned width)
547 isl_int mod;
549 isl_int_init(mod);
550 isl_int_set_si(mod, 1);
551 isl_int_mul_2exp(mod, mod, width);
553 pwaff = isl_pw_aff_mod(pwaff, mod);
555 isl_int_clear(mod);
557 return pwaff;
560 /* Limit the domain of "pwaff" to those elements where the function
561 * value satisfies
563 * 2^{width-1} <= pwaff < 2^{width-1}
565 static __isl_give isl_pw_aff *avoid_overflow(__isl_take isl_pw_aff *pwaff,
566 unsigned width)
568 isl_int v;
569 isl_space *space = isl_pw_aff_get_domain_space(pwaff);
570 isl_local_space *ls = isl_local_space_from_space(space);
571 isl_aff *bound;
572 isl_set *dom;
573 isl_pw_aff *b;
575 isl_int_init(v);
576 isl_int_set_si(v, 1);
577 isl_int_mul_2exp(v, v, width - 1);
579 bound = isl_aff_zero_on_domain(ls);
580 bound = isl_aff_add_constant(bound, v);
581 b = isl_pw_aff_from_aff(bound);
583 dom = isl_pw_aff_lt_set(isl_pw_aff_copy(pwaff), isl_pw_aff_copy(b));
584 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
586 b = isl_pw_aff_neg(b);
587 dom = isl_pw_aff_ge_set(isl_pw_aff_copy(pwaff), b);
588 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
590 isl_int_clear(v);
592 return pwaff;
595 /* Handle potential overflows on signed computations.
597 * If options->signed_overflow is set to PET_OVERFLOW_AVOID,
598 * the we adjust the domain of "pa" to avoid overflows.
600 __isl_give isl_pw_aff *PetScan::signed_overflow(__isl_take isl_pw_aff *pa,
601 unsigned width)
603 if (options->signed_overflow == PET_OVERFLOW_AVOID)
604 pa = avoid_overflow(pa, width);
606 return pa;
609 /* Return the piecewise affine expression "set ? 1 : 0" defined on "dom".
611 static __isl_give isl_pw_aff *indicator_function(__isl_take isl_set *set,
612 __isl_take isl_set *dom)
614 isl_pw_aff *pa;
615 pa = isl_set_indicator_function(set);
616 pa = isl_pw_aff_intersect_domain(pa, dom);
617 return pa;
620 /* Extract an affine expression from some binary operations.
621 * If the result of the expression is unsigned, then we wrap it
622 * based on the size of the type. Otherwise, we ensure that
623 * no overflow occurs.
625 __isl_give isl_pw_aff *PetScan::extract_affine(BinaryOperator *expr)
627 isl_pw_aff *res;
628 unsigned width;
630 switch (expr->getOpcode()) {
631 case BO_Add:
632 case BO_Sub:
633 res = extract_affine_add(expr);
634 break;
635 case BO_Div:
636 res = extract_affine_div(expr);
637 break;
638 case BO_Rem:
639 res = extract_affine_mod(expr);
640 break;
641 case BO_Mul:
642 res = extract_affine_mul(expr);
643 break;
644 case BO_LT:
645 case BO_LE:
646 case BO_GT:
647 case BO_GE:
648 case BO_EQ:
649 case BO_NE:
650 case BO_LAnd:
651 case BO_LOr:
652 return extract_condition(expr);
653 default:
654 unsupported(expr);
655 return NULL;
658 width = ast_context.getIntWidth(expr->getType());
659 if (expr->getType()->isUnsignedIntegerType())
660 res = wrap(res, width);
661 else
662 res = signed_overflow(res, width);
664 return res;
667 /* Extract an affine expression from a negation operation.
669 __isl_give isl_pw_aff *PetScan::extract_affine(UnaryOperator *expr)
671 if (expr->getOpcode() == UO_Minus)
672 return isl_pw_aff_neg(extract_affine(expr->getSubExpr()));
673 if (expr->getOpcode() == UO_LNot)
674 return extract_condition(expr);
676 unsupported(expr);
677 return NULL;
680 __isl_give isl_pw_aff *PetScan::extract_affine(ParenExpr *expr)
682 return extract_affine(expr->getSubExpr());
685 /* Extract an affine expression from some special function calls.
686 * In particular, we handle "min", "max", "ceild" and "floord".
687 * In case of the latter two, the second argument needs to be
688 * a (positive) integer constant.
690 __isl_give isl_pw_aff *PetScan::extract_affine(CallExpr *expr)
692 FunctionDecl *fd;
693 string name;
694 isl_pw_aff *aff1, *aff2;
696 fd = expr->getDirectCallee();
697 if (!fd) {
698 unsupported(expr);
699 return NULL;
702 name = fd->getDeclName().getAsString();
703 if (!(expr->getNumArgs() == 2 && name == "min") &&
704 !(expr->getNumArgs() == 2 && name == "max") &&
705 !(expr->getNumArgs() == 2 && name == "floord") &&
706 !(expr->getNumArgs() == 2 && name == "ceild")) {
707 unsupported(expr);
708 return NULL;
711 if (name == "min" || name == "max") {
712 aff1 = extract_affine(expr->getArg(0));
713 aff2 = extract_affine(expr->getArg(1));
715 if (name == "min")
716 aff1 = isl_pw_aff_min(aff1, aff2);
717 else
718 aff1 = isl_pw_aff_max(aff1, aff2);
719 } else if (name == "floord" || name == "ceild") {
720 isl_int v;
721 Expr *arg2 = expr->getArg(1);
723 if (arg2->getStmtClass() != Stmt::IntegerLiteralClass) {
724 unsupported(expr);
725 return NULL;
727 aff1 = extract_affine(expr->getArg(0));
728 isl_int_init(v);
729 extract_int(cast<IntegerLiteral>(arg2), &v);
730 aff1 = isl_pw_aff_scale_down(aff1, v);
731 isl_int_clear(v);
732 if (name == "floord")
733 aff1 = isl_pw_aff_floor(aff1);
734 else
735 aff1 = isl_pw_aff_ceil(aff1);
736 } else {
737 unsupported(expr);
738 return NULL;
741 return aff1;
744 /* This method is called when we come across an access that is
745 * nested in what is supposed to be an affine expression.
746 * If nesting is allowed, we return a new parameter that corresponds
747 * to this nested access. Otherwise, we simply complain.
749 * Note that we currently don't allow nested accesses themselves
750 * to contain any nested accesses, so we check if we can extract
751 * the access without any nesting and complain if we can't.
753 * The new parameter is resolved in resolve_nested.
755 isl_pw_aff *PetScan::nested_access(Expr *expr)
757 isl_id *id;
758 isl_space *dim;
759 isl_aff *aff;
760 isl_set *dom;
761 isl_map *access;
763 if (!nesting_enabled) {
764 unsupported(expr);
765 return NULL;
768 allow_nested = false;
769 access = extract_access(expr);
770 allow_nested = true;
771 if (!access) {
772 unsupported(expr);
773 return NULL;
775 isl_map_free(access);
777 id = isl_id_alloc(ctx, NULL, expr);
778 dim = isl_space_params_alloc(ctx, 1);
780 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
782 dom = isl_set_universe(isl_space_copy(dim));
783 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
784 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
786 return isl_pw_aff_alloc(dom, aff);
789 /* Affine expressions are not supposed to contain array accesses,
790 * but if nesting is allowed, we return a parameter corresponding
791 * to the array access.
793 __isl_give isl_pw_aff *PetScan::extract_affine(ArraySubscriptExpr *expr)
795 return nested_access(expr);
798 /* Extract an affine expression from a conditional operation.
800 __isl_give isl_pw_aff *PetScan::extract_affine(ConditionalOperator *expr)
802 isl_pw_aff *cond, *lhs, *rhs, *res;
804 cond = extract_condition(expr->getCond());
805 lhs = extract_affine(expr->getTrueExpr());
806 rhs = extract_affine(expr->getFalseExpr());
808 return isl_pw_aff_cond(cond, lhs, rhs);
811 /* Extract an affine expression, if possible, from "expr".
812 * Otherwise return NULL.
814 __isl_give isl_pw_aff *PetScan::extract_affine(Expr *expr)
816 switch (expr->getStmtClass()) {
817 case Stmt::ImplicitCastExprClass:
818 return extract_affine(cast<ImplicitCastExpr>(expr));
819 case Stmt::IntegerLiteralClass:
820 return extract_affine(cast<IntegerLiteral>(expr));
821 case Stmt::DeclRefExprClass:
822 return extract_affine(cast<DeclRefExpr>(expr));
823 case Stmt::BinaryOperatorClass:
824 return extract_affine(cast<BinaryOperator>(expr));
825 case Stmt::UnaryOperatorClass:
826 return extract_affine(cast<UnaryOperator>(expr));
827 case Stmt::ParenExprClass:
828 return extract_affine(cast<ParenExpr>(expr));
829 case Stmt::CallExprClass:
830 return extract_affine(cast<CallExpr>(expr));
831 case Stmt::ArraySubscriptExprClass:
832 return extract_affine(cast<ArraySubscriptExpr>(expr));
833 case Stmt::ConditionalOperatorClass:
834 return extract_affine(cast<ConditionalOperator>(expr));
835 default:
836 unsupported(expr);
838 return NULL;
841 __isl_give isl_map *PetScan::extract_access(ImplicitCastExpr *expr)
843 return extract_access(expr->getSubExpr());
846 /* Return the depth of an array of the given type.
848 static int array_depth(const Type *type)
850 if (type->isPointerType())
851 return 1 + array_depth(type->getPointeeType().getTypePtr());
852 if (type->isArrayType()) {
853 const ArrayType *atype;
854 type = type->getCanonicalTypeInternal().getTypePtr();
855 atype = cast<ArrayType>(type);
856 return 1 + array_depth(atype->getElementType().getTypePtr());
858 return 0;
861 /* Return the element type of the given array type.
863 static QualType base_type(QualType qt)
865 const Type *type = qt.getTypePtr();
867 if (type->isPointerType())
868 return base_type(type->getPointeeType());
869 if (type->isArrayType()) {
870 const ArrayType *atype;
871 type = type->getCanonicalTypeInternal().getTypePtr();
872 atype = cast<ArrayType>(type);
873 return base_type(atype->getElementType());
875 return qt;
878 /* Extract an access relation from a reference to a variable.
879 * If the variable has name "A" and its type corresponds to an
880 * array of depth d, then the returned access relation is of the
881 * form
883 * { [] -> A[i_1,...,i_d] }
885 __isl_give isl_map *PetScan::extract_access(DeclRefExpr *expr)
887 ValueDecl *decl = expr->getDecl();
888 int depth = array_depth(decl->getType().getTypePtr());
889 isl_id *id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
890 isl_space *dim = isl_space_alloc(ctx, 0, 0, depth);
891 isl_map *access_rel;
893 dim = isl_space_set_tuple_id(dim, isl_dim_out, id);
895 access_rel = isl_map_universe(dim);
897 return access_rel;
900 /* Extract an access relation from an integer contant.
901 * If the value of the constant is "v", then the returned access relation
902 * is
904 * { [] -> [v] }
906 __isl_give isl_map *PetScan::extract_access(IntegerLiteral *expr)
908 return isl_map_from_range(isl_set_from_pw_aff(extract_affine(expr)));
911 /* Try and extract an access relation from the given Expr.
912 * Return NULL if it doesn't work out.
914 __isl_give isl_map *PetScan::extract_access(Expr *expr)
916 switch (expr->getStmtClass()) {
917 case Stmt::ImplicitCastExprClass:
918 return extract_access(cast<ImplicitCastExpr>(expr));
919 case Stmt::DeclRefExprClass:
920 return extract_access(cast<DeclRefExpr>(expr));
921 case Stmt::ArraySubscriptExprClass:
922 return extract_access(cast<ArraySubscriptExpr>(expr));
923 case Stmt::IntegerLiteralClass:
924 return extract_access(cast<IntegerLiteral>(expr));
925 default:
926 unsupported(expr);
928 return NULL;
931 /* Assign the affine expression "index" to the output dimension "pos" of "map",
932 * restrict the domain to those values that result in a non-negative index
933 * and return the result.
935 __isl_give isl_map *set_index(__isl_take isl_map *map, int pos,
936 __isl_take isl_pw_aff *index)
938 isl_map *index_map;
939 int len = isl_map_dim(map, isl_dim_out);
940 isl_id *id;
941 isl_set *domain;
943 domain = isl_pw_aff_nonneg_set(isl_pw_aff_copy(index));
944 index = isl_pw_aff_intersect_domain(index, domain);
945 index_map = isl_map_from_range(isl_set_from_pw_aff(index));
946 index_map = isl_map_insert_dims(index_map, isl_dim_out, 0, pos);
947 index_map = isl_map_add_dims(index_map, isl_dim_out, len - pos - 1);
948 id = isl_map_get_tuple_id(map, isl_dim_out);
949 index_map = isl_map_set_tuple_id(index_map, isl_dim_out, id);
951 map = isl_map_intersect(map, index_map);
953 return map;
956 /* Extract an access relation from the given array subscript expression.
957 * If nesting is allowed in general, then we turn it on while
958 * examining the index expression.
960 * We first extract an access relation from the base.
961 * This will result in an access relation with a range that corresponds
962 * to the array being accessed and with earlier indices filled in already.
963 * We then extract the current index and fill that in as well.
964 * The position of the current index is based on the type of base.
965 * If base is the actual array variable, then the depth of this type
966 * will be the same as the depth of the array and we will fill in
967 * the first array index.
968 * Otherwise, the depth of the base type will be smaller and we will fill
969 * in a later index.
971 __isl_give isl_map *PetScan::extract_access(ArraySubscriptExpr *expr)
973 Expr *base = expr->getBase();
974 Expr *idx = expr->getIdx();
975 isl_pw_aff *index;
976 isl_map *base_access;
977 isl_map *access;
978 int depth = array_depth(base->getType().getTypePtr());
979 int pos;
980 bool save_nesting = nesting_enabled;
982 nesting_enabled = allow_nested;
984 base_access = extract_access(base);
985 index = extract_affine(idx);
987 nesting_enabled = save_nesting;
989 pos = isl_map_dim(base_access, isl_dim_out) - depth;
990 access = set_index(base_access, pos, index);
992 return access;
995 /* Check if "expr" calls function "minmax" with two arguments and if so
996 * make lhs and rhs refer to these two arguments.
998 static bool is_minmax(Expr *expr, const char *minmax, Expr *&lhs, Expr *&rhs)
1000 CallExpr *call;
1001 FunctionDecl *fd;
1002 string name;
1004 if (expr->getStmtClass() != Stmt::CallExprClass)
1005 return false;
1007 call = cast<CallExpr>(expr);
1008 fd = call->getDirectCallee();
1009 if (!fd)
1010 return false;
1012 if (call->getNumArgs() != 2)
1013 return false;
1015 name = fd->getDeclName().getAsString();
1016 if (name != minmax)
1017 return false;
1019 lhs = call->getArg(0);
1020 rhs = call->getArg(1);
1022 return true;
1025 /* Check if "expr" is of the form min(lhs, rhs) and if so make
1026 * lhs and rhs refer to the two arguments.
1028 static bool is_min(Expr *expr, Expr *&lhs, Expr *&rhs)
1030 return is_minmax(expr, "min", lhs, rhs);
1033 /* Check if "expr" is of the form max(lhs, rhs) and if so make
1034 * lhs and rhs refer to the two arguments.
1036 static bool is_max(Expr *expr, Expr *&lhs, Expr *&rhs)
1038 return is_minmax(expr, "max", lhs, rhs);
1041 /* Return "lhs && rhs", defined on the shared definition domain.
1043 static __isl_give isl_pw_aff *pw_aff_and(__isl_take isl_pw_aff *lhs,
1044 __isl_take isl_pw_aff *rhs)
1046 isl_set *cond;
1047 isl_set *dom;
1049 dom = isl_set_intersect(isl_pw_aff_domain(isl_pw_aff_copy(lhs)),
1050 isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1051 cond = isl_set_intersect(isl_pw_aff_non_zero_set(lhs),
1052 isl_pw_aff_non_zero_set(rhs));
1053 return indicator_function(cond, dom);
1056 /* Return "lhs && rhs", with shortcut semantics.
1057 * That is, if lhs is false, then the result is defined even if rhs is not.
1058 * In practice, we compute lhs ? rhs : lhs.
1060 static __isl_give isl_pw_aff *pw_aff_and_then(__isl_take isl_pw_aff *lhs,
1061 __isl_take isl_pw_aff *rhs)
1063 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), rhs, lhs);
1066 /* Return "lhs || rhs", with shortcut semantics.
1067 * That is, if lhs is true, then the result is defined even if rhs is not.
1068 * In practice, we compute lhs ? lhs : rhs.
1070 static __isl_give isl_pw_aff *pw_aff_or_else(__isl_take isl_pw_aff *lhs,
1071 __isl_take isl_pw_aff *rhs)
1073 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), lhs, rhs);
1076 /* Extract an affine expressions representing the comparison "LHS op RHS"
1077 * "comp" is the original statement that "LHS op RHS" is derived from
1078 * and is used for diagnostics.
1080 * If the comparison is of the form
1082 * a <= min(b,c)
1084 * then the expression is constructed as the conjunction of
1085 * the comparisons
1087 * a <= b and a <= c
1089 * A similar optimization is performed for max(a,b) <= c.
1090 * We do this because that will lead to simpler representations
1091 * of the expression.
1092 * If isl is ever enhanced to explicitly deal with min and max expressions,
1093 * this optimization can be removed.
1095 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperatorKind op,
1096 Expr *LHS, Expr *RHS, Stmt *comp)
1098 isl_pw_aff *lhs;
1099 isl_pw_aff *rhs;
1100 isl_pw_aff *res;
1101 isl_set *cond;
1102 isl_set *dom;
1104 if (op == BO_GT)
1105 return extract_comparison(BO_LT, RHS, LHS, comp);
1106 if (op == BO_GE)
1107 return extract_comparison(BO_LE, RHS, LHS, comp);
1109 if (op == BO_LT || op == BO_LE) {
1110 Expr *expr1, *expr2;
1111 if (is_min(RHS, expr1, expr2)) {
1112 lhs = extract_comparison(op, LHS, expr1, comp);
1113 rhs = extract_comparison(op, LHS, expr2, comp);
1114 return pw_aff_and(lhs, rhs);
1116 if (is_max(LHS, expr1, expr2)) {
1117 lhs = extract_comparison(op, expr1, RHS, comp);
1118 rhs = extract_comparison(op, expr2, RHS, comp);
1119 return pw_aff_and(lhs, rhs);
1123 lhs = extract_affine(LHS);
1124 rhs = extract_affine(RHS);
1126 dom = isl_pw_aff_domain(isl_pw_aff_copy(lhs));
1127 dom = isl_set_intersect(dom, isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1129 switch (op) {
1130 case BO_LT:
1131 cond = isl_pw_aff_lt_set(lhs, rhs);
1132 break;
1133 case BO_LE:
1134 cond = isl_pw_aff_le_set(lhs, rhs);
1135 break;
1136 case BO_EQ:
1137 cond = isl_pw_aff_eq_set(lhs, rhs);
1138 break;
1139 case BO_NE:
1140 cond = isl_pw_aff_ne_set(lhs, rhs);
1141 break;
1142 default:
1143 isl_pw_aff_free(lhs);
1144 isl_pw_aff_free(rhs);
1145 isl_set_free(dom);
1146 unsupported(comp);
1147 return NULL;
1150 cond = isl_set_coalesce(cond);
1151 res = indicator_function(cond, dom);
1153 return res;
1156 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperator *comp)
1158 return extract_comparison(comp->getOpcode(), comp->getLHS(),
1159 comp->getRHS(), comp);
1162 /* Extract an affine expression representing the negation (logical not)
1163 * of a subexpression.
1165 __isl_give isl_pw_aff *PetScan::extract_boolean(UnaryOperator *op)
1167 isl_set *set_cond, *dom;
1168 isl_pw_aff *cond, *res;
1170 cond = extract_condition(op->getSubExpr());
1172 dom = isl_pw_aff_domain(isl_pw_aff_copy(cond));
1174 set_cond = isl_pw_aff_zero_set(cond);
1176 res = indicator_function(set_cond, dom);
1178 return res;
1181 /* Extract an affine expression representing the disjunction (logical or)
1182 * or conjunction (logical and) of two subexpressions.
1184 __isl_give isl_pw_aff *PetScan::extract_boolean(BinaryOperator *comp)
1186 isl_pw_aff *lhs, *rhs;
1188 lhs = extract_condition(comp->getLHS());
1189 rhs = extract_condition(comp->getRHS());
1191 switch (comp->getOpcode()) {
1192 case BO_LAnd:
1193 return pw_aff_and_then(lhs, rhs);
1194 case BO_LOr:
1195 return pw_aff_or_else(lhs, rhs);
1196 default:
1197 isl_pw_aff_free(lhs);
1198 isl_pw_aff_free(rhs);
1201 unsupported(comp);
1202 return NULL;
1205 __isl_give isl_pw_aff *PetScan::extract_condition(UnaryOperator *expr)
1207 switch (expr->getOpcode()) {
1208 case UO_LNot:
1209 return extract_boolean(expr);
1210 default:
1211 unsupported(expr);
1212 return NULL;
1216 /* Extract the affine expression "expr != 0 ? 1 : 0".
1218 __isl_give isl_pw_aff *PetScan::extract_implicit_condition(Expr *expr)
1220 isl_pw_aff *res;
1221 isl_set *set, *dom;
1223 res = extract_affine(expr);
1225 dom = isl_pw_aff_domain(isl_pw_aff_copy(res));
1226 set = isl_pw_aff_non_zero_set(res);
1228 res = indicator_function(set, dom);
1230 return res;
1233 /* Extract an affine expression from a boolean expression.
1234 * In particular, return the expression "expr ? 1 : 0".
1236 * If the expression doesn't look like a condition, we assume it
1237 * is an affine expression and return the condition "expr != 0 ? 1 : 0".
1239 __isl_give isl_pw_aff *PetScan::extract_condition(Expr *expr)
1241 BinaryOperator *comp;
1243 if (!expr) {
1244 isl_set *u = isl_set_universe(isl_space_params_alloc(ctx, 0));
1245 return indicator_function(u, isl_set_copy(u));
1248 if (expr->getStmtClass() == Stmt::ParenExprClass)
1249 return extract_condition(cast<ParenExpr>(expr)->getSubExpr());
1251 if (expr->getStmtClass() == Stmt::UnaryOperatorClass)
1252 return extract_condition(cast<UnaryOperator>(expr));
1254 if (expr->getStmtClass() != Stmt::BinaryOperatorClass)
1255 return extract_implicit_condition(expr);
1257 comp = cast<BinaryOperator>(expr);
1258 switch (comp->getOpcode()) {
1259 case BO_LT:
1260 case BO_LE:
1261 case BO_GT:
1262 case BO_GE:
1263 case BO_EQ:
1264 case BO_NE:
1265 return extract_comparison(comp);
1266 case BO_LAnd:
1267 case BO_LOr:
1268 return extract_boolean(comp);
1269 default:
1270 return extract_implicit_condition(expr);
1274 static enum pet_op_type UnaryOperatorKind2pet_op_type(UnaryOperatorKind kind)
1276 switch (kind) {
1277 case UO_Minus:
1278 return pet_op_minus;
1279 case UO_PostInc:
1280 return pet_op_post_inc;
1281 case UO_PostDec:
1282 return pet_op_post_dec;
1283 case UO_PreInc:
1284 return pet_op_pre_inc;
1285 case UO_PreDec:
1286 return pet_op_pre_dec;
1287 default:
1288 return pet_op_last;
1292 static enum pet_op_type BinaryOperatorKind2pet_op_type(BinaryOperatorKind kind)
1294 switch (kind) {
1295 case BO_AddAssign:
1296 return pet_op_add_assign;
1297 case BO_SubAssign:
1298 return pet_op_sub_assign;
1299 case BO_MulAssign:
1300 return pet_op_mul_assign;
1301 case BO_DivAssign:
1302 return pet_op_div_assign;
1303 case BO_Assign:
1304 return pet_op_assign;
1305 case BO_Add:
1306 return pet_op_add;
1307 case BO_Sub:
1308 return pet_op_sub;
1309 case BO_Mul:
1310 return pet_op_mul;
1311 case BO_Div:
1312 return pet_op_div;
1313 case BO_Rem:
1314 return pet_op_mod;
1315 case BO_EQ:
1316 return pet_op_eq;
1317 case BO_LE:
1318 return pet_op_le;
1319 case BO_LT:
1320 return pet_op_lt;
1321 case BO_GT:
1322 return pet_op_gt;
1323 default:
1324 return pet_op_last;
1328 /* Construct a pet_expr representing a unary operator expression.
1330 struct pet_expr *PetScan::extract_expr(UnaryOperator *expr)
1332 struct pet_expr *arg;
1333 enum pet_op_type op;
1335 op = UnaryOperatorKind2pet_op_type(expr->getOpcode());
1336 if (op == pet_op_last) {
1337 unsupported(expr);
1338 return NULL;
1341 arg = extract_expr(expr->getSubExpr());
1343 if (expr->isIncrementDecrementOp() &&
1344 arg && arg->type == pet_expr_access) {
1345 mark_write(arg);
1346 arg->acc.read = 1;
1349 return pet_expr_new_unary(ctx, op, arg);
1352 /* Mark the given access pet_expr as a write.
1353 * If a scalar is being accessed, then mark its value
1354 * as unknown in assigned_value.
1356 void PetScan::mark_write(struct pet_expr *access)
1358 isl_id *id;
1359 ValueDecl *decl;
1361 access->acc.write = 1;
1362 access->acc.read = 0;
1364 if (isl_map_dim(access->acc.access, isl_dim_out) != 0)
1365 return;
1367 id = isl_map_get_tuple_id(access->acc.access, isl_dim_out);
1368 decl = (ValueDecl *) isl_id_get_user(id);
1369 clear_assignment(assigned_value, decl);
1370 isl_id_free(id);
1373 /* Construct a pet_expr representing a binary operator expression.
1375 * If the top level operator is an assignment and the LHS is an access,
1376 * then we mark that access as a write. If the operator is a compound
1377 * assignment, the access is marked as both a read and a write.
1379 * If "expr" assigns something to a scalar variable, then we mark
1380 * the variable as having been assigned. If, furthermore, the expression
1381 * is affine, then keep track of this value in assigned_value
1382 * so that we can plug it in when we later come across the same variable.
1384 struct pet_expr *PetScan::extract_expr(BinaryOperator *expr)
1386 struct pet_expr *lhs, *rhs;
1387 enum pet_op_type op;
1389 op = BinaryOperatorKind2pet_op_type(expr->getOpcode());
1390 if (op == pet_op_last) {
1391 unsupported(expr);
1392 return NULL;
1395 lhs = extract_expr(expr->getLHS());
1396 rhs = extract_expr(expr->getRHS());
1398 if (expr->isAssignmentOp() && lhs && lhs->type == pet_expr_access) {
1399 mark_write(lhs);
1400 if (expr->isCompoundAssignmentOp())
1401 lhs->acc.read = 1;
1404 if (expr->getOpcode() == BO_Assign &&
1405 lhs && lhs->type == pet_expr_access &&
1406 isl_map_dim(lhs->acc.access, isl_dim_out) == 0) {
1407 isl_id *id = isl_map_get_tuple_id(lhs->acc.access, isl_dim_out);
1408 ValueDecl *decl = (ValueDecl *) isl_id_get_user(id);
1409 Expr *rhs = expr->getRHS();
1410 isl_pw_aff *pa = try_extract_affine(rhs);
1411 clear_assignment(assigned_value, decl);
1412 if (pa) {
1413 assigned_value[decl] = pa;
1414 insert_expression(pa);
1416 isl_id_free(id);
1419 return pet_expr_new_binary(ctx, op, lhs, rhs);
1422 /* Construct a pet_expr representing a conditional operation.
1424 * We first try to extract the condition as an affine expression.
1425 * If that fails, we construct a pet_expr tree representing the condition.
1427 struct pet_expr *PetScan::extract_expr(ConditionalOperator *expr)
1429 struct pet_expr *cond, *lhs, *rhs;
1430 isl_pw_aff *pa;
1432 pa = try_extract_affine(expr->getCond());
1433 if (pa) {
1434 isl_set *test = isl_set_from_pw_aff(pa);
1435 cond = pet_expr_from_access(isl_map_from_range(test));
1436 } else
1437 cond = extract_expr(expr->getCond());
1438 lhs = extract_expr(expr->getTrueExpr());
1439 rhs = extract_expr(expr->getFalseExpr());
1441 return pet_expr_new_ternary(ctx, cond, lhs, rhs);
1444 struct pet_expr *PetScan::extract_expr(ImplicitCastExpr *expr)
1446 return extract_expr(expr->getSubExpr());
1449 /* Construct a pet_expr representing a floating point value.
1451 struct pet_expr *PetScan::extract_expr(FloatingLiteral *expr)
1453 return pet_expr_new_double(ctx, expr->getValueAsApproximateDouble());
1456 /* Extract an access relation from "expr" and then convert it into
1457 * a pet_expr.
1459 struct pet_expr *PetScan::extract_access_expr(Expr *expr)
1461 isl_map *access;
1462 struct pet_expr *pe;
1464 access = extract_access(expr);
1466 pe = pet_expr_from_access(access);
1468 return pe;
1471 struct pet_expr *PetScan::extract_expr(ParenExpr *expr)
1473 return extract_expr(expr->getSubExpr());
1476 /* Construct a pet_expr representing a function call.
1478 * If we are passing along a pointer to an array element
1479 * or an entire row or even higher dimensional slice of an array,
1480 * then the function being called may write into the array.
1482 * We assume here that if the function is declared to take a pointer
1483 * to a const type, then the function will perform a read
1484 * and that otherwise, it will perform a write.
1486 struct pet_expr *PetScan::extract_expr(CallExpr *expr)
1488 struct pet_expr *res = NULL;
1489 FunctionDecl *fd;
1490 string name;
1492 fd = expr->getDirectCallee();
1493 if (!fd) {
1494 unsupported(expr);
1495 return NULL;
1498 name = fd->getDeclName().getAsString();
1499 res = pet_expr_new_call(ctx, name.c_str(), expr->getNumArgs());
1500 if (!res)
1501 return NULL;
1503 for (int i = 0; i < expr->getNumArgs(); ++i) {
1504 Expr *arg = expr->getArg(i);
1505 int is_addr = 0;
1506 pet_expr *main_arg;
1508 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
1509 ImplicitCastExpr *ice = cast<ImplicitCastExpr>(arg);
1510 arg = ice->getSubExpr();
1512 if (arg->getStmtClass() == Stmt::UnaryOperatorClass) {
1513 UnaryOperator *op = cast<UnaryOperator>(arg);
1514 if (op->getOpcode() == UO_AddrOf) {
1515 is_addr = 1;
1516 arg = op->getSubExpr();
1519 res->args[i] = PetScan::extract_expr(arg);
1520 main_arg = res->args[i];
1521 if (is_addr)
1522 res->args[i] = pet_expr_new_unary(ctx,
1523 pet_op_address_of, res->args[i]);
1524 if (!res->args[i])
1525 goto error;
1526 if (arg->getStmtClass() == Stmt::ArraySubscriptExprClass &&
1527 array_depth(arg->getType().getTypePtr()) > 0)
1528 is_addr = 1;
1529 if (is_addr && main_arg->type == pet_expr_access) {
1530 ParmVarDecl *parm;
1531 if (!fd->hasPrototype()) {
1532 unsupported(expr, "prototype required");
1533 goto error;
1535 parm = fd->getParamDecl(i);
1536 if (!const_base(parm->getType()))
1537 mark_write(main_arg);
1541 return res;
1542 error:
1543 pet_expr_free(res);
1544 return NULL;
1547 /* Try and onstruct a pet_expr representing "expr".
1549 struct pet_expr *PetScan::extract_expr(Expr *expr)
1551 switch (expr->getStmtClass()) {
1552 case Stmt::UnaryOperatorClass:
1553 return extract_expr(cast<UnaryOperator>(expr));
1554 case Stmt::CompoundAssignOperatorClass:
1555 case Stmt::BinaryOperatorClass:
1556 return extract_expr(cast<BinaryOperator>(expr));
1557 case Stmt::ImplicitCastExprClass:
1558 return extract_expr(cast<ImplicitCastExpr>(expr));
1559 case Stmt::ArraySubscriptExprClass:
1560 case Stmt::DeclRefExprClass:
1561 case Stmt::IntegerLiteralClass:
1562 return extract_access_expr(expr);
1563 case Stmt::FloatingLiteralClass:
1564 return extract_expr(cast<FloatingLiteral>(expr));
1565 case Stmt::ParenExprClass:
1566 return extract_expr(cast<ParenExpr>(expr));
1567 case Stmt::ConditionalOperatorClass:
1568 return extract_expr(cast<ConditionalOperator>(expr));
1569 case Stmt::CallExprClass:
1570 return extract_expr(cast<CallExpr>(expr));
1571 default:
1572 unsupported(expr);
1574 return NULL;
1577 /* Check if the given initialization statement is an assignment.
1578 * If so, return that assignment. Otherwise return NULL.
1580 BinaryOperator *PetScan::initialization_assignment(Stmt *init)
1582 BinaryOperator *ass;
1584 if (init->getStmtClass() != Stmt::BinaryOperatorClass)
1585 return NULL;
1587 ass = cast<BinaryOperator>(init);
1588 if (ass->getOpcode() != BO_Assign)
1589 return NULL;
1591 return ass;
1594 /* Check if the given initialization statement is a declaration
1595 * of a single variable.
1596 * If so, return that declaration. Otherwise return NULL.
1598 Decl *PetScan::initialization_declaration(Stmt *init)
1600 DeclStmt *decl;
1602 if (init->getStmtClass() != Stmt::DeclStmtClass)
1603 return NULL;
1605 decl = cast<DeclStmt>(init);
1607 if (!decl->isSingleDecl())
1608 return NULL;
1610 return decl->getSingleDecl();
1613 /* Given the assignment operator in the initialization of a for loop,
1614 * extract the induction variable, i.e., the (integer)variable being
1615 * assigned.
1617 ValueDecl *PetScan::extract_induction_variable(BinaryOperator *init)
1619 Expr *lhs;
1620 DeclRefExpr *ref;
1621 ValueDecl *decl;
1622 const Type *type;
1624 lhs = init->getLHS();
1625 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1626 unsupported(init);
1627 return NULL;
1630 ref = cast<DeclRefExpr>(lhs);
1631 decl = ref->getDecl();
1632 type = decl->getType().getTypePtr();
1634 if (!type->isIntegerType()) {
1635 unsupported(lhs);
1636 return NULL;
1639 return decl;
1642 /* Given the initialization statement of a for loop and the single
1643 * declaration in this initialization statement,
1644 * extract the induction variable, i.e., the (integer) variable being
1645 * declared.
1647 VarDecl *PetScan::extract_induction_variable(Stmt *init, Decl *decl)
1649 VarDecl *vd;
1651 vd = cast<VarDecl>(decl);
1653 const QualType type = vd->getType();
1654 if (!type->isIntegerType()) {
1655 unsupported(init);
1656 return NULL;
1659 if (!vd->getInit()) {
1660 unsupported(init);
1661 return NULL;
1664 return vd;
1667 /* Check that op is of the form iv++ or iv--.
1668 * Return an affine expression "1" or "-1" accordingly.
1670 __isl_give isl_pw_aff *PetScan::extract_unary_increment(
1671 clang::UnaryOperator *op, clang::ValueDecl *iv)
1673 Expr *sub;
1674 DeclRefExpr *ref;
1675 isl_space *space;
1676 isl_aff *aff;
1678 if (!op->isIncrementDecrementOp()) {
1679 unsupported(op);
1680 return NULL;
1683 sub = op->getSubExpr();
1684 if (sub->getStmtClass() != Stmt::DeclRefExprClass) {
1685 unsupported(op);
1686 return NULL;
1689 ref = cast<DeclRefExpr>(sub);
1690 if (ref->getDecl() != iv) {
1691 unsupported(op);
1692 return NULL;
1695 space = isl_space_params_alloc(ctx, 0);
1696 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
1698 if (op->isIncrementOp())
1699 aff = isl_aff_add_constant_si(aff, 1);
1700 else
1701 aff = isl_aff_add_constant_si(aff, -1);
1703 return isl_pw_aff_from_aff(aff);
1706 /* If the isl_pw_aff on which isl_pw_aff_foreach_piece is called
1707 * has a single constant expression, then put this constant in *user.
1708 * The caller is assumed to have checked that this function will
1709 * be called exactly once.
1711 static int extract_cst(__isl_take isl_set *set, __isl_take isl_aff *aff,
1712 void *user)
1714 isl_int *inc = (isl_int *)user;
1715 int res = 0;
1717 if (isl_aff_is_cst(aff))
1718 isl_aff_get_constant(aff, inc);
1719 else
1720 res = -1;
1722 isl_set_free(set);
1723 isl_aff_free(aff);
1725 return res;
1728 /* Check if op is of the form
1730 * iv = iv + inc
1732 * and return inc as an affine expression.
1734 * We extract an affine expression from the RHS, subtract iv and return
1735 * the result.
1737 __isl_give isl_pw_aff *PetScan::extract_binary_increment(BinaryOperator *op,
1738 clang::ValueDecl *iv)
1740 Expr *lhs;
1741 DeclRefExpr *ref;
1742 isl_id *id;
1743 isl_space *dim;
1744 isl_aff *aff;
1745 isl_pw_aff *val;
1747 if (op->getOpcode() != BO_Assign) {
1748 unsupported(op);
1749 return NULL;
1752 lhs = op->getLHS();
1753 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1754 unsupported(op);
1755 return NULL;
1758 ref = cast<DeclRefExpr>(lhs);
1759 if (ref->getDecl() != iv) {
1760 unsupported(op);
1761 return NULL;
1764 val = extract_affine(op->getRHS());
1766 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
1768 dim = isl_space_params_alloc(ctx, 1);
1769 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
1770 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
1771 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
1773 val = isl_pw_aff_sub(val, isl_pw_aff_from_aff(aff));
1775 return val;
1778 /* Check that op is of the form iv += cst or iv -= cst
1779 * and return an affine expression corresponding oto cst or -cst accordingly.
1781 __isl_give isl_pw_aff *PetScan::extract_compound_increment(
1782 CompoundAssignOperator *op, clang::ValueDecl *iv)
1784 Expr *lhs;
1785 DeclRefExpr *ref;
1786 bool neg = false;
1787 isl_pw_aff *val;
1788 BinaryOperatorKind opcode;
1790 opcode = op->getOpcode();
1791 if (opcode != BO_AddAssign && opcode != BO_SubAssign) {
1792 unsupported(op);
1793 return NULL;
1795 if (opcode == BO_SubAssign)
1796 neg = true;
1798 lhs = op->getLHS();
1799 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1800 unsupported(op);
1801 return NULL;
1804 ref = cast<DeclRefExpr>(lhs);
1805 if (ref->getDecl() != iv) {
1806 unsupported(op);
1807 return NULL;
1810 val = extract_affine(op->getRHS());
1811 if (neg)
1812 val = isl_pw_aff_neg(val);
1814 return val;
1817 /* Check that the increment of the given for loop increments
1818 * (or decrements) the induction variable "iv" and return
1819 * the increment as an affine expression if successful.
1821 __isl_give isl_pw_aff *PetScan::extract_increment(clang::ForStmt *stmt,
1822 ValueDecl *iv)
1824 Stmt *inc = stmt->getInc();
1826 if (!inc) {
1827 unsupported(stmt);
1828 return NULL;
1831 if (inc->getStmtClass() == Stmt::UnaryOperatorClass)
1832 return extract_unary_increment(cast<UnaryOperator>(inc), iv);
1833 if (inc->getStmtClass() == Stmt::CompoundAssignOperatorClass)
1834 return extract_compound_increment(
1835 cast<CompoundAssignOperator>(inc), iv);
1836 if (inc->getStmtClass() == Stmt::BinaryOperatorClass)
1837 return extract_binary_increment(cast<BinaryOperator>(inc), iv);
1839 unsupported(inc);
1840 return NULL;
1843 /* Embed the given iteration domain in an extra outer loop
1844 * with induction variable "var".
1845 * If this variable appeared as a parameter in the constraints,
1846 * it is replaced by the new outermost dimension.
1848 static __isl_give isl_set *embed(__isl_take isl_set *set,
1849 __isl_take isl_id *var)
1851 int pos;
1853 set = isl_set_insert_dims(set, isl_dim_set, 0, 1);
1854 pos = isl_set_find_dim_by_id(set, isl_dim_param, var);
1855 if (pos >= 0) {
1856 set = isl_set_equate(set, isl_dim_param, pos, isl_dim_set, 0);
1857 set = isl_set_project_out(set, isl_dim_param, pos, 1);
1860 isl_id_free(var);
1861 return set;
1864 /* Return those elements in the space of "cond" that come after
1865 * (based on "sign") an element in "cond".
1867 static __isl_give isl_set *after(__isl_take isl_set *cond, int sign)
1869 isl_map *previous_to_this;
1871 if (sign > 0)
1872 previous_to_this = isl_map_lex_lt(isl_set_get_space(cond));
1873 else
1874 previous_to_this = isl_map_lex_gt(isl_set_get_space(cond));
1876 cond = isl_set_apply(cond, previous_to_this);
1878 return cond;
1881 /* Create the infinite iteration domain
1883 * { [id] : id >= 0 }
1885 * If "scop" has an affine skip of type pet_skip_later,
1886 * then remove those iterations i that have an earlier iteration
1887 * where the skip condition is satisfied, meaning that iteration i
1888 * is not executed.
1889 * Since we are dealing with a loop without loop iterator,
1890 * the skip condition cannot refer to the current loop iterator and
1891 * so effectively, the returned set is of the form
1893 * { [0]; [id] : id >= 1 and not skip }
1895 static __isl_give isl_set *infinite_domain(__isl_take isl_id *id,
1896 struct pet_scop *scop)
1898 isl_ctx *ctx = isl_id_get_ctx(id);
1899 isl_set *domain;
1900 isl_set *skip;
1902 domain = isl_set_nat_universe(isl_space_set_alloc(ctx, 0, 1));
1903 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, id);
1905 if (!pet_scop_has_affine_skip(scop, pet_skip_later))
1906 return domain;
1908 skip = pet_scop_get_skip(scop, pet_skip_later);
1909 skip = isl_set_fix_si(skip, isl_dim_set, 0, 1);
1910 skip = isl_set_params(skip);
1911 skip = embed(skip, isl_id_copy(id));
1912 skip = isl_set_intersect(skip , isl_set_copy(domain));
1913 domain = isl_set_subtract(domain, after(skip, 1));
1915 return domain;
1918 /* Create an identity mapping on the space containing "domain".
1920 static __isl_give isl_map *identity_map(__isl_keep isl_set *domain)
1922 isl_space *space;
1923 isl_map *id;
1925 space = isl_space_map_from_set(isl_set_get_space(domain));
1926 id = isl_map_identity(space);
1928 return id;
1931 /* Add a filter to "scop" that imposes that it is only executed
1932 * when "break_access" has a zero value for all previous iterations
1933 * of "domain".
1935 * The input "break_access" has a zero-dimensional domain and range.
1937 static struct pet_scop *scop_add_break(struct pet_scop *scop,
1938 __isl_take isl_map *break_access, __isl_take isl_set *domain, int sign)
1940 isl_ctx *ctx = isl_set_get_ctx(domain);
1941 isl_id *id_test;
1942 isl_map *prev;
1944 id_test = isl_map_get_tuple_id(break_access, isl_dim_out);
1945 break_access = isl_map_add_dims(break_access, isl_dim_in, 1);
1946 break_access = isl_map_add_dims(break_access, isl_dim_out, 1);
1947 break_access = isl_map_intersect_range(break_access, domain);
1948 break_access = isl_map_set_tuple_id(break_access, isl_dim_out, id_test);
1949 if (sign > 0)
1950 prev = isl_map_lex_gt_first(isl_map_get_space(break_access), 1);
1951 else
1952 prev = isl_map_lex_lt_first(isl_map_get_space(break_access), 1);
1953 break_access = isl_map_intersect(break_access, prev);
1954 scop = pet_scop_filter(scop, break_access, 0);
1955 scop = pet_scop_merge_filters(scop);
1957 return scop;
1960 /* Construct a pet_scop for an infinite loop around the given body.
1962 * We extract a pet_scop for the body and then embed it in a loop with
1963 * iteration domain
1965 * { [t] : t >= 0 }
1967 * and schedule
1969 * { [t] -> [t] }
1971 * If the body contains any break, then it is taken into
1972 * account in infinite_domain (if the skip condition is affine)
1973 * or in scop_add_break (if the skip condition is not affine).
1975 struct pet_scop *PetScan::extract_infinite_loop(Stmt *body)
1977 isl_id *id;
1978 isl_set *domain;
1979 isl_map *ident;
1980 isl_map *access;
1981 struct pet_scop *scop;
1982 bool has_var_break;
1984 scop = extract(body);
1985 if (!scop)
1986 return NULL;
1988 id = isl_id_alloc(ctx, "t", NULL);
1989 domain = infinite_domain(isl_id_copy(id), scop);
1990 ident = identity_map(domain);
1992 has_var_break = pet_scop_has_var_skip(scop, pet_skip_later);
1993 if (has_var_break)
1994 access = pet_scop_get_skip_map(scop, pet_skip_later);
1996 scop = pet_scop_embed(scop, isl_set_copy(domain),
1997 isl_map_copy(ident), ident, id);
1998 if (has_var_break)
1999 scop = scop_add_break(scop, access, domain, 1);
2000 else
2001 isl_set_free(domain);
2003 return scop;
2006 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
2008 * for (;;)
2009 * body
2012 struct pet_scop *PetScan::extract_infinite_for(ForStmt *stmt)
2014 return extract_infinite_loop(stmt->getBody());
2017 /* Create an access to a virtual array representing the result
2018 * of a condition.
2019 * Unlike other accessed data, the id of the array is NULL as
2020 * there is no ValueDecl in the program corresponding to the virtual
2021 * array.
2022 * The array starts out as a scalar, but grows along with the
2023 * statement writing to the array in pet_scop_embed.
2025 static __isl_give isl_map *create_test_access(isl_ctx *ctx, int test_nr)
2027 isl_space *dim = isl_space_alloc(ctx, 0, 0, 0);
2028 isl_id *id;
2029 char name[50];
2031 snprintf(name, sizeof(name), "__pet_test_%d", test_nr);
2032 id = isl_id_alloc(ctx, name, NULL);
2033 dim = isl_space_set_tuple_id(dim, isl_dim_out, id);
2034 return isl_map_universe(dim);
2037 /* Add an array with the given extent ("access") to the list
2038 * of arrays in "scop" and return the extended pet_scop.
2039 * The array is marked as attaining values 0 and 1 only and
2040 * as each element being assigned at most once.
2042 static struct pet_scop *scop_add_array(struct pet_scop *scop,
2043 __isl_keep isl_map *access, clang::ASTContext &ast_ctx)
2045 isl_ctx *ctx = isl_map_get_ctx(access);
2046 isl_space *dim;
2047 struct pet_array **arrays;
2048 struct pet_array *array;
2050 if (!scop)
2051 return NULL;
2052 if (!ctx)
2053 goto error;
2055 arrays = isl_realloc_array(ctx, scop->arrays, struct pet_array *,
2056 scop->n_array + 1);
2057 if (!arrays)
2058 goto error;
2059 scop->arrays = arrays;
2061 array = isl_calloc_type(ctx, struct pet_array);
2062 if (!array)
2063 goto error;
2065 array->extent = isl_map_range(isl_map_copy(access));
2066 dim = isl_space_params_alloc(ctx, 0);
2067 array->context = isl_set_universe(dim);
2068 dim = isl_space_set_alloc(ctx, 0, 1);
2069 array->value_bounds = isl_set_universe(dim);
2070 array->value_bounds = isl_set_lower_bound_si(array->value_bounds,
2071 isl_dim_set, 0, 0);
2072 array->value_bounds = isl_set_upper_bound_si(array->value_bounds,
2073 isl_dim_set, 0, 1);
2074 array->element_type = strdup("int");
2075 array->element_size = ast_ctx.getTypeInfo(ast_ctx.IntTy).first / 8;
2076 array->uniquely_defined = 1;
2078 scop->arrays[scop->n_array] = array;
2079 scop->n_array++;
2081 if (!array->extent || !array->context)
2082 goto error;
2084 return scop;
2085 error:
2086 pet_scop_free(scop);
2087 return NULL;
2090 /* Construct a pet_scop for a while loop of the form
2092 * while (pa)
2093 * body
2095 * In particular, construct a scop for an infinite loop around body and
2096 * intersect the domain with the affine expression.
2097 * Note that this intersection may result in an empty loop.
2099 struct pet_scop *PetScan::extract_affine_while(__isl_take isl_pw_aff *pa,
2100 Stmt *body)
2102 struct pet_scop *scop;
2103 isl_set *dom;
2104 isl_set *valid;
2106 valid = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2107 dom = isl_pw_aff_non_zero_set(pa);
2108 scop = extract_infinite_loop(body);
2109 scop = pet_scop_restrict(scop, dom);
2110 scop = pet_scop_restrict_context(scop, valid);
2112 return scop;
2115 /* Construct a scop for a while, given the scops for the condition
2116 * and the body, the filter access and the iteration domain of
2117 * the while loop.
2119 * In particular, the scop for the condition is filtered to depend
2120 * on "test_access" evaluating to true for all previous iterations
2121 * of the loop, while the scop for the body is filtered to depend
2122 * on "test_access" evaluating to true for all iterations up to the
2123 * current iteration.
2125 * These filtered scops are then combined into a single scop.
2127 * "sign" is positive if the iterator increases and negative
2128 * if it decreases.
2130 static struct pet_scop *scop_add_while(struct pet_scop *scop_cond,
2131 struct pet_scop *scop_body, __isl_take isl_map *test_access,
2132 __isl_take isl_set *domain, int sign)
2134 isl_ctx *ctx = isl_set_get_ctx(domain);
2135 isl_id *id_test;
2136 isl_map *prev;
2138 id_test = isl_map_get_tuple_id(test_access, isl_dim_out);
2139 test_access = isl_map_add_dims(test_access, isl_dim_in, 1);
2140 test_access = isl_map_add_dims(test_access, isl_dim_out, 1);
2141 test_access = isl_map_intersect_range(test_access, domain);
2142 test_access = isl_map_set_tuple_id(test_access, isl_dim_out, id_test);
2143 if (sign > 0)
2144 prev = isl_map_lex_ge_first(isl_map_get_space(test_access), 1);
2145 else
2146 prev = isl_map_lex_le_first(isl_map_get_space(test_access), 1);
2147 test_access = isl_map_intersect(test_access, prev);
2148 scop_body = pet_scop_filter(scop_body, isl_map_copy(test_access), 1);
2149 if (sign > 0)
2150 prev = isl_map_lex_gt_first(isl_map_get_space(test_access), 1);
2151 else
2152 prev = isl_map_lex_lt_first(isl_map_get_space(test_access), 1);
2153 test_access = isl_map_intersect(test_access, prev);
2154 scop_cond = pet_scop_filter(scop_cond, test_access, 1);
2156 return pet_scop_add_seq(ctx, scop_cond, scop_body);
2159 /* Check if the while loop is of the form
2161 * while (affine expression)
2162 * body
2164 * If so, call extract_affine_while to construct a scop.
2166 * Otherwise, construct a generic while scop, with iteration domain
2167 * { [t] : t >= 0 }. The scop consists of two parts, one for
2168 * evaluating the condition and one for the body.
2169 * The schedule is adjusted to reflect that the condition is evaluated
2170 * before the body is executed and the body is filtered to depend
2171 * on the result of the condition evaluating to true on all iterations
2172 * up to the current iteration, while the evaluation the condition itself
2173 * is filtered to depend on the result of the condition evaluating to true
2174 * on all previous iterations.
2175 * The context of the scop representing the body is dropped
2176 * because we don't know how many times the body will be executed,
2177 * if at all.
2179 * If the body contains any break, then it is taken into
2180 * account in infinite_domain (if the skip condition is affine)
2181 * or in scop_add_break (if the skip condition is not affine).
2183 struct pet_scop *PetScan::extract(WhileStmt *stmt)
2185 Expr *cond;
2186 isl_id *id;
2187 isl_map *test_access;
2188 isl_set *domain;
2189 isl_map *ident;
2190 isl_pw_aff *pa;
2191 struct pet_scop *scop, *scop_body;
2192 bool has_var_break;
2193 isl_map *break_access;
2195 cond = stmt->getCond();
2196 if (!cond) {
2197 unsupported(stmt);
2198 return NULL;
2201 pa = try_extract_affine_condition(cond);
2202 if (pa)
2203 return extract_affine_while(pa, stmt->getBody());
2205 if (!allow_nested) {
2206 unsupported(stmt);
2207 return NULL;
2210 test_access = create_test_access(ctx, n_test++);
2211 scop = extract_non_affine_condition(cond, isl_map_copy(test_access));
2212 scop = scop_add_array(scop, test_access, ast_context);
2213 scop_body = extract(stmt->getBody());
2215 id = isl_id_alloc(ctx, "t", NULL);
2216 domain = infinite_domain(isl_id_copy(id), scop_body);
2217 ident = identity_map(domain);
2219 has_var_break = pet_scop_has_var_skip(scop_body, pet_skip_later);
2220 if (has_var_break)
2221 break_access = pet_scop_get_skip_map(scop_body, pet_skip_later);
2223 scop = pet_scop_prefix(scop, 0);
2224 scop = pet_scop_embed(scop, isl_set_copy(domain), isl_map_copy(ident),
2225 isl_map_copy(ident), isl_id_copy(id));
2226 scop_body = pet_scop_reset_context(scop_body);
2227 scop_body = pet_scop_prefix(scop_body, 1);
2228 scop_body = pet_scop_embed(scop_body, isl_set_copy(domain),
2229 isl_map_copy(ident), ident, id);
2231 if (has_var_break) {
2232 scop = scop_add_break(scop, isl_map_copy(break_access),
2233 isl_set_copy(domain), 1);
2234 scop_body = scop_add_break(scop_body, break_access,
2235 isl_set_copy(domain), 1);
2237 scop = scop_add_while(scop, scop_body, test_access, domain, 1);
2239 return scop;
2242 /* Check whether "cond" expresses a simple loop bound
2243 * on the only set dimension.
2244 * In particular, if "up" is set then "cond" should contain only
2245 * upper bounds on the set dimension.
2246 * Otherwise, it should contain only lower bounds.
2248 static bool is_simple_bound(__isl_keep isl_set *cond, isl_int inc)
2250 if (isl_int_is_pos(inc))
2251 return !isl_set_dim_has_any_lower_bound(cond, isl_dim_set, 0);
2252 else
2253 return !isl_set_dim_has_any_upper_bound(cond, isl_dim_set, 0);
2256 /* Extend a condition on a given iteration of a loop to one that
2257 * imposes the same condition on all previous iterations.
2258 * "domain" expresses the lower [upper] bound on the iterations
2259 * when inc is positive [negative].
2261 * In particular, we construct the condition (when inc is positive)
2263 * forall i' : (domain(i') and i' <= i) => cond(i')
2265 * which is equivalent to
2267 * not exists i' : domain(i') and i' <= i and not cond(i')
2269 * We construct this set by negating cond, applying a map
2271 * { [i'] -> [i] : domain(i') and i' <= i }
2273 * and then negating the result again.
2275 static __isl_give isl_set *valid_for_each_iteration(__isl_take isl_set *cond,
2276 __isl_take isl_set *domain, isl_int inc)
2278 isl_map *previous_to_this;
2280 if (isl_int_is_pos(inc))
2281 previous_to_this = isl_map_lex_le(isl_set_get_space(domain));
2282 else
2283 previous_to_this = isl_map_lex_ge(isl_set_get_space(domain));
2285 previous_to_this = isl_map_intersect_domain(previous_to_this, domain);
2287 cond = isl_set_complement(cond);
2288 cond = isl_set_apply(cond, previous_to_this);
2289 cond = isl_set_complement(cond);
2291 return cond;
2294 /* Construct a domain of the form
2296 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
2298 static __isl_give isl_set *strided_domain(__isl_take isl_id *id,
2299 __isl_take isl_pw_aff *init, isl_int inc)
2301 isl_aff *aff;
2302 isl_space *dim;
2303 isl_set *set;
2305 init = isl_pw_aff_insert_dims(init, isl_dim_in, 0, 1);
2306 dim = isl_pw_aff_get_domain_space(init);
2307 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2308 aff = isl_aff_add_coefficient(aff, isl_dim_in, 0, inc);
2309 init = isl_pw_aff_add(init, isl_pw_aff_from_aff(aff));
2311 dim = isl_space_set_alloc(isl_pw_aff_get_ctx(init), 1, 1);
2312 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
2313 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2314 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2316 set = isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff), init);
2318 set = isl_set_lower_bound_si(set, isl_dim_set, 0, 0);
2320 return isl_set_params(set);
2323 /* Assuming "cond" represents a bound on a loop where the loop
2324 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2325 * is possible.
2327 * Under the given assumptions, wrapping is only possible if "cond" allows
2328 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2329 * increasing iterator and 0 in case of a decreasing iterator.
2331 static bool can_wrap(__isl_keep isl_set *cond, ValueDecl *iv, isl_int inc)
2333 bool cw;
2334 isl_int limit;
2335 isl_set *test;
2337 test = isl_set_copy(cond);
2339 isl_int_init(limit);
2340 if (isl_int_is_neg(inc))
2341 isl_int_set_si(limit, 0);
2342 else {
2343 isl_int_set_si(limit, 1);
2344 isl_int_mul_2exp(limit, limit, get_type_size(iv));
2345 isl_int_sub_ui(limit, limit, 1);
2348 test = isl_set_fix(cond, isl_dim_set, 0, limit);
2349 cw = !isl_set_is_empty(test);
2350 isl_set_free(test);
2352 isl_int_clear(limit);
2354 return cw;
2357 /* Given a one-dimensional space, construct the following mapping on this
2358 * space
2360 * { [v] -> [v mod 2^width] }
2362 * where width is the number of bits used to represent the values
2363 * of the unsigned variable "iv".
2365 static __isl_give isl_map *compute_wrapping(__isl_take isl_space *dim,
2366 ValueDecl *iv)
2368 isl_int mod;
2369 isl_aff *aff;
2370 isl_map *map;
2372 isl_int_init(mod);
2373 isl_int_set_si(mod, 1);
2374 isl_int_mul_2exp(mod, mod, get_type_size(iv));
2376 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2377 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2378 aff = isl_aff_mod(aff, mod);
2380 isl_int_clear(mod);
2382 return isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2383 map = isl_map_reverse(map);
2386 /* Project out the parameter "id" from "set".
2388 static __isl_give isl_set *set_project_out_by_id(__isl_take isl_set *set,
2389 __isl_keep isl_id *id)
2391 int pos;
2393 pos = isl_set_find_dim_by_id(set, isl_dim_param, id);
2394 if (pos >= 0)
2395 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2397 return set;
2400 /* Compute the set of parameters for which "set1" is a subset of "set2".
2402 * set1 is a subset of set2 if
2404 * forall i in set1 : i in set2
2406 * or
2408 * not exists i in set1 and i not in set2
2410 * i.e.,
2412 * not exists i in set1 \ set2
2414 static __isl_give isl_set *enforce_subset(__isl_take isl_set *set1,
2415 __isl_take isl_set *set2)
2417 return isl_set_complement(isl_set_params(isl_set_subtract(set1, set2)));
2420 /* Compute the set of parameter values for which "cond" holds
2421 * on the next iteration for each element of "dom".
2423 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2424 * and then compute the set of parameters for which the result is a subset
2425 * of "cond".
2427 static __isl_give isl_set *valid_on_next(__isl_take isl_set *cond,
2428 __isl_take isl_set *dom, isl_int inc)
2430 isl_space *space;
2431 isl_aff *aff;
2432 isl_map *next;
2434 space = isl_set_get_space(dom);
2435 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2436 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2437 aff = isl_aff_add_constant(aff, inc);
2438 next = isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2440 dom = isl_set_apply(dom, next);
2442 return enforce_subset(dom, cond);
2445 /* Does "id" refer to a nested access?
2447 static bool is_nested_parameter(__isl_keep isl_id *id)
2449 return id && isl_id_get_user(id) && !isl_id_get_name(id);
2452 /* Does parameter "pos" of "space" refer to a nested access?
2454 static bool is_nested_parameter(__isl_keep isl_space *space, int pos)
2456 bool nested;
2457 isl_id *id;
2459 id = isl_space_get_dim_id(space, isl_dim_param, pos);
2460 nested = is_nested_parameter(id);
2461 isl_id_free(id);
2463 return nested;
2466 /* Does "space" involve any parameters that refer to nested
2467 * accesses, i.e., parameters with no name?
2469 static bool has_nested(__isl_keep isl_space *space)
2471 int nparam;
2473 nparam = isl_space_dim(space, isl_dim_param);
2474 for (int i = 0; i < nparam; ++i)
2475 if (is_nested_parameter(space, i))
2476 return true;
2478 return false;
2481 /* Does "pa" involve any parameters that refer to nested
2482 * accesses, i.e., parameters with no name?
2484 static bool has_nested(__isl_keep isl_pw_aff *pa)
2486 isl_space *space;
2487 bool nested;
2489 space = isl_pw_aff_get_space(pa);
2490 nested = has_nested(space);
2491 isl_space_free(space);
2493 return nested;
2496 /* Construct a pet_scop for a for statement.
2497 * The for loop is required to be of the form
2499 * for (i = init; condition; ++i)
2501 * or
2503 * for (i = init; condition; --i)
2505 * The initialization of the for loop should either be an assignment
2506 * to an integer variable, or a declaration of such a variable with
2507 * initialization.
2509 * The condition is allowed to contain nested accesses, provided
2510 * they are not being written to inside the body of the loop.
2511 * Otherwise, or if the condition is otherwise non-affine, the for loop is
2512 * essentially treated as a while loop, with iteration domain
2513 * { [i] : i >= init }.
2515 * We extract a pet_scop for the body and then embed it in a loop with
2516 * iteration domain and schedule
2518 * { [i] : i >= init and condition' }
2519 * { [i] -> [i] }
2521 * or
2523 * { [i] : i <= init and condition' }
2524 * { [i] -> [-i] }
2526 * Where condition' is equal to condition if the latter is
2527 * a simple upper [lower] bound and a condition that is extended
2528 * to apply to all previous iterations otherwise.
2530 * If the condition is non-affine, then we drop the condition from the
2531 * iteration domain and instead create a separate statement
2532 * for evaluating the condition. The body is then filtered to depend
2533 * on the result of the condition evaluating to true on all iterations
2534 * up to the current iteration, while the evaluation the condition itself
2535 * is filtered to depend on the result of the condition evaluating to true
2536 * on all previous iterations.
2537 * The context of the scop representing the body is dropped
2538 * because we don't know how many times the body will be executed,
2539 * if at all.
2541 * If the stride of the loop is not 1, then "i >= init" is replaced by
2543 * (exists a: i = init + stride * a and a >= 0)
2545 * If the loop iterator i is unsigned, then wrapping may occur.
2546 * During the computation, we work with a virtual iterator that
2547 * does not wrap. However, the condition in the code applies
2548 * to the wrapped value, so we need to change condition(i)
2549 * into condition([i % 2^width]).
2550 * After computing the virtual domain and schedule, we apply
2551 * the function { [v] -> [v % 2^width] } to the domain and the domain
2552 * of the schedule. In order not to lose any information, we also
2553 * need to intersect the domain of the schedule with the virtual domain
2554 * first, since some iterations in the wrapped domain may be scheduled
2555 * several times, typically an infinite number of times.
2556 * Note that there may be no need to perform this final wrapping
2557 * if the loop condition (after wrapping) satisfies certain conditions.
2558 * However, the is_simple_bound condition is not enough since it doesn't
2559 * check if there even is an upper bound.
2561 * If the loop condition is non-affine, then we keep the virtual
2562 * iterator in the iteration domain and instead replace all accesses
2563 * to the original iterator by the wrapping of the virtual iterator.
2565 * Wrapping on unsigned iterators can be avoided entirely if
2566 * loop condition is simple, the loop iterator is incremented
2567 * [decremented] by one and the last value before wrapping cannot
2568 * possibly satisfy the loop condition.
2570 * Before extracting a pet_scop from the body we remove all
2571 * assignments in assigned_value to variables that are assigned
2572 * somewhere in the body of the loop.
2574 * Valid parameters for a for loop are those for which the initial
2575 * value itself, the increment on each domain iteration and
2576 * the condition on both the initial value and
2577 * the result of incrementing the iterator for each iteration of the domain
2578 * can be evaluated.
2579 * If the loop condition is non-affine, then we only consider validity
2580 * of the initial value.
2582 * If the body contains any break, then we keep track of it in "skip"
2583 * (if the skip condition is affine) or it is handled in scop_add_break
2584 * (if the skip condition is not affine).
2585 * Note that the affine break condition needs to be considered with
2586 * respect to previous iterations in the virtual domain (if any)
2587 * and that the domain needs to be kept virtual if there is a non-affine
2588 * break condition.
2590 struct pet_scop *PetScan::extract_for(ForStmt *stmt)
2592 BinaryOperator *ass;
2593 Decl *decl;
2594 Stmt *init;
2595 Expr *lhs, *rhs;
2596 ValueDecl *iv;
2597 isl_space *space;
2598 isl_set *domain;
2599 isl_map *sched;
2600 isl_set *cond = NULL;
2601 isl_set *skip = NULL;
2602 isl_id *id;
2603 struct pet_scop *scop, *scop_cond = NULL;
2604 assigned_value_cache cache(assigned_value);
2605 isl_int inc;
2606 bool is_one;
2607 bool is_unsigned;
2608 bool is_simple;
2609 bool is_virtual;
2610 bool keep_virtual = false;
2611 bool has_affine_break;
2612 bool has_var_break;
2613 isl_map *wrap = NULL;
2614 isl_pw_aff *pa, *pa_inc, *init_val;
2615 isl_set *valid_init;
2616 isl_set *valid_cond;
2617 isl_set *valid_cond_init;
2618 isl_set *valid_cond_next;
2619 isl_set *valid_inc;
2620 isl_map *test_access = NULL, *break_access = NULL;
2621 int stmt_id;
2623 if (!stmt->getInit() && !stmt->getCond() && !stmt->getInc())
2624 return extract_infinite_for(stmt);
2626 init = stmt->getInit();
2627 if (!init) {
2628 unsupported(stmt);
2629 return NULL;
2631 if ((ass = initialization_assignment(init)) != NULL) {
2632 iv = extract_induction_variable(ass);
2633 if (!iv)
2634 return NULL;
2635 lhs = ass->getLHS();
2636 rhs = ass->getRHS();
2637 } else if ((decl = initialization_declaration(init)) != NULL) {
2638 VarDecl *var = extract_induction_variable(init, decl);
2639 if (!var)
2640 return NULL;
2641 iv = var;
2642 rhs = var->getInit();
2643 lhs = create_DeclRefExpr(var);
2644 } else {
2645 unsupported(stmt->getInit());
2646 return NULL;
2649 pa_inc = extract_increment(stmt, iv);
2650 if (!pa_inc)
2651 return NULL;
2653 isl_int_init(inc);
2654 if (isl_pw_aff_n_piece(pa_inc) != 1 ||
2655 isl_pw_aff_foreach_piece(pa_inc, &extract_cst, &inc) < 0) {
2656 isl_pw_aff_free(pa_inc);
2657 unsupported(stmt->getInc());
2658 isl_int_clear(inc);
2659 return NULL;
2661 valid_inc = isl_pw_aff_domain(pa_inc);
2663 is_unsigned = iv->getType()->isUnsignedIntegerType();
2665 assigned_value.erase(iv);
2666 clear_assignments clear(assigned_value);
2667 clear.TraverseStmt(stmt->getBody());
2669 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
2671 pa = try_extract_nested_condition(stmt->getCond());
2672 if (allow_nested && (!pa || has_nested(pa)))
2673 stmt_id = n_stmt++;
2675 scop = extract(stmt->getBody());
2677 has_affine_break = scop &&
2678 pet_scop_has_affine_skip(scop, pet_skip_later);
2679 if (has_affine_break) {
2680 skip = pet_scop_get_skip(scop, pet_skip_later);
2681 skip = isl_set_fix_si(skip, isl_dim_set, 0, 1);
2682 skip = isl_set_params(skip);
2684 has_var_break = scop && pet_scop_has_var_skip(scop, pet_skip_later);
2685 if (has_var_break) {
2686 break_access = pet_scop_get_skip_map(scop, pet_skip_later);
2687 keep_virtual = true;
2690 if (pa && !is_nested_allowed(pa, scop)) {
2691 isl_pw_aff_free(pa);
2692 pa = NULL;
2695 if (!allow_nested && !pa)
2696 pa = try_extract_affine_condition(stmt->getCond());
2697 valid_cond = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2698 cond = isl_pw_aff_non_zero_set(pa);
2699 if (allow_nested && !cond) {
2700 int save_n_stmt = n_stmt;
2701 test_access = create_test_access(ctx, n_test++);
2702 n_stmt = stmt_id;
2703 scop_cond = extract_non_affine_condition(stmt->getCond(),
2704 isl_map_copy(test_access));
2705 n_stmt = save_n_stmt;
2706 scop_cond = scop_add_array(scop_cond, test_access, ast_context);
2707 scop_cond = pet_scop_prefix(scop_cond, 0);
2708 scop = pet_scop_reset_context(scop);
2709 scop = pet_scop_prefix(scop, 1);
2710 keep_virtual = true;
2711 cond = isl_set_universe(isl_space_set_alloc(ctx, 0, 0));
2714 cond = embed(cond, isl_id_copy(id));
2715 skip = embed(skip, isl_id_copy(id));
2716 valid_cond = isl_set_coalesce(valid_cond);
2717 valid_cond = embed(valid_cond, isl_id_copy(id));
2718 valid_inc = embed(valid_inc, isl_id_copy(id));
2719 is_one = isl_int_is_one(inc) || isl_int_is_negone(inc);
2720 is_virtual = is_unsigned && (!is_one || can_wrap(cond, iv, inc));
2722 init_val = extract_affine(rhs);
2723 valid_cond_init = enforce_subset(
2724 isl_set_from_pw_aff(isl_pw_aff_copy(init_val)),
2725 isl_set_copy(valid_cond));
2726 if (is_one && !is_virtual) {
2727 isl_pw_aff_free(init_val);
2728 pa = extract_comparison(isl_int_is_pos(inc) ? BO_GE : BO_LE,
2729 lhs, rhs, init);
2730 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2731 valid_init = set_project_out_by_id(valid_init, id);
2732 domain = isl_pw_aff_non_zero_set(pa);
2733 } else {
2734 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(init_val));
2735 domain = strided_domain(isl_id_copy(id), init_val, inc);
2738 domain = embed(domain, isl_id_copy(id));
2739 if (is_virtual) {
2740 isl_map *rev_wrap;
2741 wrap = compute_wrapping(isl_set_get_space(cond), iv);
2742 rev_wrap = isl_map_reverse(isl_map_copy(wrap));
2743 cond = isl_set_apply(cond, isl_map_copy(rev_wrap));
2744 skip = isl_set_apply(skip, isl_map_copy(rev_wrap));
2745 valid_cond = isl_set_apply(valid_cond, isl_map_copy(rev_wrap));
2746 valid_inc = isl_set_apply(valid_inc, rev_wrap);
2748 is_simple = is_simple_bound(cond, inc);
2749 if (!is_simple) {
2750 cond = isl_set_gist(cond, isl_set_copy(domain));
2751 is_simple = is_simple_bound(cond, inc);
2753 if (!is_simple)
2754 cond = valid_for_each_iteration(cond,
2755 isl_set_copy(domain), inc);
2756 domain = isl_set_intersect(domain, cond);
2757 if (has_affine_break) {
2758 skip = isl_set_intersect(skip , isl_set_copy(domain));
2759 skip = after(skip, isl_int_sgn(inc));
2760 domain = isl_set_subtract(domain, skip);
2762 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, isl_id_copy(id));
2763 space = isl_space_from_domain(isl_set_get_space(domain));
2764 space = isl_space_add_dims(space, isl_dim_out, 1);
2765 sched = isl_map_universe(space);
2766 if (isl_int_is_pos(inc))
2767 sched = isl_map_equate(sched, isl_dim_in, 0, isl_dim_out, 0);
2768 else
2769 sched = isl_map_oppose(sched, isl_dim_in, 0, isl_dim_out, 0);
2771 valid_cond_next = valid_on_next(valid_cond, isl_set_copy(domain), inc);
2772 valid_inc = enforce_subset(isl_set_copy(domain), valid_inc);
2774 if (is_virtual && !keep_virtual) {
2775 wrap = isl_map_set_dim_id(wrap,
2776 isl_dim_out, 0, isl_id_copy(id));
2777 sched = isl_map_intersect_domain(sched, isl_set_copy(domain));
2778 domain = isl_set_apply(domain, isl_map_copy(wrap));
2779 sched = isl_map_apply_domain(sched, wrap);
2781 if (!(is_virtual && keep_virtual)) {
2782 space = isl_set_get_space(domain);
2783 wrap = isl_map_identity(isl_space_map_from_set(space));
2786 scop_cond = pet_scop_embed(scop_cond, isl_set_copy(domain),
2787 isl_map_copy(sched), isl_map_copy(wrap), isl_id_copy(id));
2788 scop = pet_scop_embed(scop, isl_set_copy(domain), sched, wrap, id);
2789 scop = resolve_nested(scop);
2790 if (has_var_break)
2791 scop = scop_add_break(scop, break_access, isl_set_copy(domain),
2792 isl_int_sgn(inc));
2793 if (test_access) {
2794 scop = scop_add_while(scop_cond, scop, test_access, domain,
2795 isl_int_sgn(inc));
2796 isl_set_free(valid_inc);
2797 } else {
2798 scop = pet_scop_restrict_context(scop, valid_inc);
2799 scop = pet_scop_restrict_context(scop, valid_cond_next);
2800 scop = pet_scop_restrict_context(scop, valid_cond_init);
2801 isl_set_free(domain);
2803 clear_assignment(assigned_value, iv);
2805 isl_int_clear(inc);
2807 scop = pet_scop_restrict_context(scop, valid_init);
2809 return scop;
2812 struct pet_scop *PetScan::extract(CompoundStmt *stmt)
2814 return extract(stmt->children());
2817 /* Does parameter "pos" of "map" refer to a nested access?
2819 static bool is_nested_parameter(__isl_keep isl_map *map, int pos)
2821 bool nested;
2822 isl_id *id;
2824 id = isl_map_get_dim_id(map, isl_dim_param, pos);
2825 nested = is_nested_parameter(id);
2826 isl_id_free(id);
2828 return nested;
2831 /* How many parameters of "space" refer to nested accesses, i.e., have no name?
2833 static int n_nested_parameter(__isl_keep isl_space *space)
2835 int n = 0;
2836 int nparam;
2838 nparam = isl_space_dim(space, isl_dim_param);
2839 for (int i = 0; i < nparam; ++i)
2840 if (is_nested_parameter(space, i))
2841 ++n;
2843 return n;
2846 /* How many parameters of "map" refer to nested accesses, i.e., have no name?
2848 static int n_nested_parameter(__isl_keep isl_map *map)
2850 isl_space *space;
2851 int n;
2853 space = isl_map_get_space(map);
2854 n = n_nested_parameter(space);
2855 isl_space_free(space);
2857 return n;
2860 /* For each nested access parameter in "space",
2861 * construct a corresponding pet_expr, place it in args and
2862 * record its position in "param2pos".
2863 * "n_arg" is the number of elements that are already in args.
2864 * The position recorded in "param2pos" takes this number into account.
2865 * If the pet_expr corresponding to a parameter is identical to
2866 * the pet_expr corresponding to an earlier parameter, then these two
2867 * parameters are made to refer to the same element in args.
2869 * Return the final number of elements in args or -1 if an error has occurred.
2871 int PetScan::extract_nested(__isl_keep isl_space *space,
2872 int n_arg, struct pet_expr **args, std::map<int,int> &param2pos)
2874 int nparam;
2876 nparam = isl_space_dim(space, isl_dim_param);
2877 for (int i = 0; i < nparam; ++i) {
2878 int j;
2879 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
2880 Expr *nested;
2882 if (!is_nested_parameter(id)) {
2883 isl_id_free(id);
2884 continue;
2887 nested = (Expr *) isl_id_get_user(id);
2888 args[n_arg] = extract_expr(nested);
2889 if (!args[n_arg])
2890 return -1;
2892 for (j = 0; j < n_arg; ++j)
2893 if (pet_expr_is_equal(args[j], args[n_arg]))
2894 break;
2896 if (j < n_arg) {
2897 pet_expr_free(args[n_arg]);
2898 args[n_arg] = NULL;
2899 param2pos[i] = j;
2900 } else
2901 param2pos[i] = n_arg++;
2903 isl_id_free(id);
2906 return n_arg;
2909 /* For each nested access parameter in the access relations in "expr",
2910 * construct a corresponding pet_expr, place it in expr->args and
2911 * record its position in "param2pos".
2912 * n is the number of nested access parameters.
2914 struct pet_expr *PetScan::extract_nested(struct pet_expr *expr, int n,
2915 std::map<int,int> &param2pos)
2917 isl_space *space;
2919 expr->args = isl_calloc_array(ctx, struct pet_expr *, n);
2920 expr->n_arg = n;
2921 if (!expr->args)
2922 goto error;
2924 space = isl_map_get_space(expr->acc.access);
2925 n = extract_nested(space, 0, expr->args, param2pos);
2926 isl_space_free(space);
2928 if (n < 0)
2929 goto error;
2931 expr->n_arg = n;
2932 return expr;
2933 error:
2934 pet_expr_free(expr);
2935 return NULL;
2938 /* Look for parameters in any access relation in "expr" that
2939 * refer to nested accesses. In particular, these are
2940 * parameters with no name.
2942 * If there are any such parameters, then the domain of the access
2943 * relation, which is still [] at this point, is replaced by
2944 * [[] -> [t_1,...,t_n]], with n the number of these parameters
2945 * (after identifying identical nested accesses).
2946 * The parameters are then equated to the corresponding t dimensions
2947 * and subsequently projected out.
2948 * param2pos maps the position of the parameter to the position
2949 * of the corresponding t dimension.
2951 struct pet_expr *PetScan::resolve_nested(struct pet_expr *expr)
2953 int n;
2954 int nparam;
2955 int n_in;
2956 isl_space *dim;
2957 isl_map *map;
2958 std::map<int,int> param2pos;
2960 if (!expr)
2961 return expr;
2963 for (int i = 0; i < expr->n_arg; ++i) {
2964 expr->args[i] = resolve_nested(expr->args[i]);
2965 if (!expr->args[i]) {
2966 pet_expr_free(expr);
2967 return NULL;
2971 if (expr->type != pet_expr_access)
2972 return expr;
2974 n = n_nested_parameter(expr->acc.access);
2975 if (n == 0)
2976 return expr;
2978 expr = extract_nested(expr, n, param2pos);
2979 if (!expr)
2980 return NULL;
2982 n = expr->n_arg;
2983 nparam = isl_map_dim(expr->acc.access, isl_dim_param);
2984 n_in = isl_map_dim(expr->acc.access, isl_dim_in);
2985 dim = isl_map_get_space(expr->acc.access);
2986 dim = isl_space_domain(dim);
2987 dim = isl_space_from_domain(dim);
2988 dim = isl_space_add_dims(dim, isl_dim_out, n);
2989 map = isl_map_universe(dim);
2990 map = isl_map_domain_map(map);
2991 map = isl_map_reverse(map);
2992 expr->acc.access = isl_map_apply_domain(expr->acc.access, map);
2994 for (int i = nparam - 1; i >= 0; --i) {
2995 isl_id *id = isl_map_get_dim_id(expr->acc.access,
2996 isl_dim_param, i);
2997 if (!is_nested_parameter(id)) {
2998 isl_id_free(id);
2999 continue;
3002 expr->acc.access = isl_map_equate(expr->acc.access,
3003 isl_dim_param, i, isl_dim_in,
3004 n_in + param2pos[i]);
3005 expr->acc.access = isl_map_project_out(expr->acc.access,
3006 isl_dim_param, i, 1);
3008 isl_id_free(id);
3011 return expr;
3012 error:
3013 pet_expr_free(expr);
3014 return NULL;
3017 /* Convert a top-level pet_expr to a pet_scop with one statement.
3018 * This mainly involves resolving nested expression parameters
3019 * and setting the name of the iteration space.
3020 * The name is given by "label" if it is non-NULL. Otherwise,
3021 * it is of the form S_<n_stmt>.
3023 struct pet_scop *PetScan::extract(Stmt *stmt, struct pet_expr *expr,
3024 __isl_take isl_id *label)
3026 struct pet_stmt *ps;
3027 SourceLocation loc = stmt->getLocStart();
3028 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3030 expr = resolve_nested(expr);
3031 ps = pet_stmt_from_pet_expr(ctx, line, label, n_stmt++, expr);
3032 return pet_scop_from_pet_stmt(ctx, ps);
3035 /* Check if we can extract an affine expression from "expr".
3036 * Return the expressions as an isl_pw_aff if we can and NULL otherwise.
3037 * We turn on autodetection so that we won't generate any warnings
3038 * and turn off nesting, so that we won't accept any non-affine constructs.
3040 __isl_give isl_pw_aff *PetScan::try_extract_affine(Expr *expr)
3042 isl_pw_aff *pwaff;
3043 int save_autodetect = options->autodetect;
3044 bool save_nesting = nesting_enabled;
3046 options->autodetect = 1;
3047 nesting_enabled = false;
3049 pwaff = extract_affine(expr);
3051 options->autodetect = save_autodetect;
3052 nesting_enabled = save_nesting;
3054 return pwaff;
3057 /* Check whether "expr" is an affine expression.
3059 bool PetScan::is_affine(Expr *expr)
3061 isl_pw_aff *pwaff;
3063 pwaff = try_extract_affine(expr);
3064 isl_pw_aff_free(pwaff);
3066 return pwaff != NULL;
3069 /* Check if we can extract an affine constraint from "expr".
3070 * Return the constraint as an isl_set if we can and NULL otherwise.
3071 * We turn on autodetection so that we won't generate any warnings
3072 * and turn off nesting, so that we won't accept any non-affine constructs.
3074 __isl_give isl_pw_aff *PetScan::try_extract_affine_condition(Expr *expr)
3076 isl_pw_aff *cond;
3077 int save_autodetect = options->autodetect;
3078 bool save_nesting = nesting_enabled;
3080 options->autodetect = 1;
3081 nesting_enabled = false;
3083 cond = extract_condition(expr);
3085 options->autodetect = save_autodetect;
3086 nesting_enabled = save_nesting;
3088 return cond;
3091 /* Check whether "expr" is an affine constraint.
3093 bool PetScan::is_affine_condition(Expr *expr)
3095 isl_pw_aff *cond;
3097 cond = try_extract_affine_condition(expr);
3098 isl_pw_aff_free(cond);
3100 return cond != NULL;
3103 /* Check if we can extract a condition from "expr".
3104 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
3105 * If allow_nested is set, then the condition may involve parameters
3106 * corresponding to nested accesses.
3107 * We turn on autodetection so that we won't generate any warnings.
3109 __isl_give isl_pw_aff *PetScan::try_extract_nested_condition(Expr *expr)
3111 isl_pw_aff *cond;
3112 int save_autodetect = options->autodetect;
3113 bool save_nesting = nesting_enabled;
3115 options->autodetect = 1;
3116 nesting_enabled = allow_nested;
3117 cond = extract_condition(expr);
3119 options->autodetect = save_autodetect;
3120 nesting_enabled = save_nesting;
3122 return cond;
3125 /* If the top-level expression of "stmt" is an assignment, then
3126 * return that assignment as a BinaryOperator.
3127 * Otherwise return NULL.
3129 static BinaryOperator *top_assignment_or_null(Stmt *stmt)
3131 BinaryOperator *ass;
3133 if (!stmt)
3134 return NULL;
3135 if (stmt->getStmtClass() != Stmt::BinaryOperatorClass)
3136 return NULL;
3138 ass = cast<BinaryOperator>(stmt);
3139 if(ass->getOpcode() != BO_Assign)
3140 return NULL;
3142 return ass;
3145 /* Check if the given if statement is a conditional assignement
3146 * with a non-affine condition. If so, construct a pet_scop
3147 * corresponding to this conditional assignment. Otherwise return NULL.
3149 * In particular we check if "stmt" is of the form
3151 * if (condition)
3152 * a = f(...);
3153 * else
3154 * a = g(...);
3156 * where a is some array or scalar access.
3157 * The constructed pet_scop then corresponds to the expression
3159 * a = condition ? f(...) : g(...)
3161 * All access relations in f(...) are intersected with condition
3162 * while all access relation in g(...) are intersected with the complement.
3164 struct pet_scop *PetScan::extract_conditional_assignment(IfStmt *stmt)
3166 BinaryOperator *ass_then, *ass_else;
3167 isl_map *write_then, *write_else;
3168 isl_set *cond, *comp;
3169 isl_map *map;
3170 isl_pw_aff *pa;
3171 int equal;
3172 struct pet_expr *pe_cond, *pe_then, *pe_else, *pe, *pe_write;
3173 bool save_nesting = nesting_enabled;
3175 if (!options->detect_conditional_assignment)
3176 return NULL;
3178 ass_then = top_assignment_or_null(stmt->getThen());
3179 ass_else = top_assignment_or_null(stmt->getElse());
3181 if (!ass_then || !ass_else)
3182 return NULL;
3184 if (is_affine_condition(stmt->getCond()))
3185 return NULL;
3187 write_then = extract_access(ass_then->getLHS());
3188 write_else = extract_access(ass_else->getLHS());
3190 equal = isl_map_is_equal(write_then, write_else);
3191 isl_map_free(write_else);
3192 if (equal < 0 || !equal) {
3193 isl_map_free(write_then);
3194 return NULL;
3197 nesting_enabled = allow_nested;
3198 pa = extract_condition(stmt->getCond());
3199 nesting_enabled = save_nesting;
3200 cond = isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa));
3201 comp = isl_pw_aff_zero_set(isl_pw_aff_copy(pa));
3202 map = isl_map_from_range(isl_set_from_pw_aff(pa));
3204 pe_cond = pet_expr_from_access(map);
3206 pe_then = extract_expr(ass_then->getRHS());
3207 pe_then = pet_expr_restrict(pe_then, cond);
3208 pe_else = extract_expr(ass_else->getRHS());
3209 pe_else = pet_expr_restrict(pe_else, comp);
3211 pe = pet_expr_new_ternary(ctx, pe_cond, pe_then, pe_else);
3212 pe_write = pet_expr_from_access(write_then);
3213 if (pe_write) {
3214 pe_write->acc.write = 1;
3215 pe_write->acc.read = 0;
3217 pe = pet_expr_new_binary(ctx, pet_op_assign, pe_write, pe);
3218 return extract(stmt, pe);
3221 /* Create a pet_scop with a single statement evaluating "cond"
3222 * and writing the result to a virtual scalar, as expressed by
3223 * "access".
3225 struct pet_scop *PetScan::extract_non_affine_condition(Expr *cond,
3226 __isl_take isl_map *access)
3228 struct pet_expr *expr, *write;
3229 struct pet_stmt *ps;
3230 struct pet_scop *scop;
3231 SourceLocation loc = cond->getLocStart();
3232 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3234 write = pet_expr_from_access(access);
3235 if (write) {
3236 write->acc.write = 1;
3237 write->acc.read = 0;
3239 expr = extract_expr(cond);
3240 expr = resolve_nested(expr);
3241 expr = pet_expr_new_binary(ctx, pet_op_assign, write, expr);
3242 ps = pet_stmt_from_pet_expr(ctx, line, NULL, n_stmt++, expr);
3243 scop = pet_scop_from_pet_stmt(ctx, ps);
3244 scop = resolve_nested(scop);
3246 return scop;
3249 extern "C" {
3250 static __isl_give isl_map *embed_access(__isl_take isl_map *access,
3251 void *user);
3254 /* Apply the map pointed to by "user" to the domain of the access
3255 * relation, thereby embedding it in the range of the map.
3256 * The domain of both relations is the zero-dimensional domain.
3258 static __isl_give isl_map *embed_access(__isl_take isl_map *access, void *user)
3260 isl_map *map = (isl_map *) user;
3262 return isl_map_apply_domain(access, isl_map_copy(map));
3265 /* Apply "map" to all access relations in "expr".
3267 static struct pet_expr *embed(struct pet_expr *expr, __isl_keep isl_map *map)
3269 return pet_expr_foreach_access(expr, &embed_access, map);
3272 /* How many parameters of "set" refer to nested accesses, i.e., have no name?
3274 static int n_nested_parameter(__isl_keep isl_set *set)
3276 isl_space *space;
3277 int n;
3279 space = isl_set_get_space(set);
3280 n = n_nested_parameter(space);
3281 isl_space_free(space);
3283 return n;
3286 /* Remove all parameters from "map" that refer to nested accesses.
3288 static __isl_give isl_map *remove_nested_parameters(__isl_take isl_map *map)
3290 int nparam;
3291 isl_space *space;
3293 space = isl_map_get_space(map);
3294 nparam = isl_space_dim(space, isl_dim_param);
3295 for (int i = nparam - 1; i >= 0; --i)
3296 if (is_nested_parameter(space, i))
3297 map = isl_map_project_out(map, isl_dim_param, i, 1);
3298 isl_space_free(space);
3300 return map;
3303 extern "C" {
3304 static __isl_give isl_map *access_remove_nested_parameters(
3305 __isl_take isl_map *access, void *user);
3308 static __isl_give isl_map *access_remove_nested_parameters(
3309 __isl_take isl_map *access, void *user)
3311 return remove_nested_parameters(access);
3314 /* Remove all nested access parameters from the schedule and all
3315 * accesses of "stmt".
3316 * There is no need to remove them from the domain as these parameters
3317 * have already been removed from the domain when this function is called.
3319 static struct pet_stmt *remove_nested_parameters(struct pet_stmt *stmt)
3321 if (!stmt)
3322 return NULL;
3323 stmt->schedule = remove_nested_parameters(stmt->schedule);
3324 stmt->body = pet_expr_foreach_access(stmt->body,
3325 &access_remove_nested_parameters, NULL);
3326 if (!stmt->schedule || !stmt->body)
3327 goto error;
3328 for (int i = 0; i < stmt->n_arg; ++i) {
3329 stmt->args[i] = pet_expr_foreach_access(stmt->args[i],
3330 &access_remove_nested_parameters, NULL);
3331 if (!stmt->args[i])
3332 goto error;
3335 return stmt;
3336 error:
3337 pet_stmt_free(stmt);
3338 return NULL;
3341 /* For each nested access parameter in the domain of "stmt",
3342 * construct a corresponding pet_expr, place it before the original
3343 * elements in stmt->args and record its position in "param2pos".
3344 * n is the number of nested access parameters.
3346 struct pet_stmt *PetScan::extract_nested(struct pet_stmt *stmt, int n,
3347 std::map<int,int> &param2pos)
3349 int i;
3350 isl_space *space;
3351 int n_arg;
3352 struct pet_expr **args;
3354 n_arg = stmt->n_arg;
3355 args = isl_calloc_array(ctx, struct pet_expr *, n + n_arg);
3356 if (!args)
3357 goto error;
3359 space = isl_set_get_space(stmt->domain);
3360 n_arg = extract_nested(space, 0, args, param2pos);
3361 isl_space_free(space);
3363 if (n_arg < 0)
3364 goto error;
3366 for (i = 0; i < stmt->n_arg; ++i)
3367 args[n_arg + i] = stmt->args[i];
3368 free(stmt->args);
3369 stmt->args = args;
3370 stmt->n_arg += n_arg;
3372 return stmt;
3373 error:
3374 if (args) {
3375 for (i = 0; i < n; ++i)
3376 pet_expr_free(args[i]);
3377 free(args);
3379 pet_stmt_free(stmt);
3380 return NULL;
3383 /* Check whether any of the arguments i of "stmt" starting at position "n"
3384 * is equal to one of the first "n" arguments j.
3385 * If so, combine the constraints on arguments i and j and remove
3386 * argument i.
3388 static struct pet_stmt *remove_duplicate_arguments(struct pet_stmt *stmt, int n)
3390 int i, j;
3391 isl_map *map;
3393 if (!stmt)
3394 return NULL;
3395 if (n == 0)
3396 return stmt;
3397 if (n == stmt->n_arg)
3398 return stmt;
3400 map = isl_set_unwrap(stmt->domain);
3402 for (i = stmt->n_arg - 1; i >= n; --i) {
3403 for (j = 0; j < n; ++j)
3404 if (pet_expr_is_equal(stmt->args[i], stmt->args[j]))
3405 break;
3406 if (j >= n)
3407 continue;
3409 map = isl_map_equate(map, isl_dim_out, i, isl_dim_out, j);
3410 map = isl_map_project_out(map, isl_dim_out, i, 1);
3412 pet_expr_free(stmt->args[i]);
3413 for (j = i; j + 1 < stmt->n_arg; ++j)
3414 stmt->args[j] = stmt->args[j + 1];
3415 stmt->n_arg--;
3418 stmt->domain = isl_map_wrap(map);
3419 if (!stmt->domain)
3420 goto error;
3421 return stmt;
3422 error:
3423 pet_stmt_free(stmt);
3424 return NULL;
3427 /* Look for parameters in the iteration domain of "stmt" that
3428 * refer to nested accesses. In particular, these are
3429 * parameters with no name.
3431 * If there are any such parameters, then as many extra variables
3432 * (after identifying identical nested accesses) are inserted in the
3433 * range of the map wrapped inside the domain, before the original variables.
3434 * If the original domain is not a wrapped map, then a new wrapped
3435 * map is created with zero output dimensions.
3436 * The parameters are then equated to the corresponding output dimensions
3437 * and subsequently projected out, from the iteration domain,
3438 * the schedule and the access relations.
3439 * For each of the output dimensions, a corresponding argument
3440 * expression is inserted. Initially they are created with
3441 * a zero-dimensional domain, so they have to be embedded
3442 * in the current iteration domain.
3443 * param2pos maps the position of the parameter to the position
3444 * of the corresponding output dimension in the wrapped map.
3446 struct pet_stmt *PetScan::resolve_nested(struct pet_stmt *stmt)
3448 int n;
3449 int nparam;
3450 unsigned n_arg;
3451 isl_map *map;
3452 std::map<int,int> param2pos;
3454 if (!stmt)
3455 return NULL;
3457 n = n_nested_parameter(stmt->domain);
3458 if (n == 0)
3459 return stmt;
3461 n_arg = stmt->n_arg;
3462 stmt = extract_nested(stmt, n, param2pos);
3463 if (!stmt)
3464 return NULL;
3466 n = stmt->n_arg - n_arg;
3467 nparam = isl_set_dim(stmt->domain, isl_dim_param);
3468 if (isl_set_is_wrapping(stmt->domain))
3469 map = isl_set_unwrap(stmt->domain);
3470 else
3471 map = isl_map_from_domain(stmt->domain);
3472 map = isl_map_insert_dims(map, isl_dim_out, 0, n);
3474 for (int i = nparam - 1; i >= 0; --i) {
3475 isl_id *id;
3477 if (!is_nested_parameter(map, i))
3478 continue;
3480 id = isl_map_get_tuple_id(stmt->args[param2pos[i]]->acc.access,
3481 isl_dim_out);
3482 map = isl_map_set_dim_id(map, isl_dim_out, param2pos[i], id);
3483 map = isl_map_equate(map, isl_dim_param, i, isl_dim_out,
3484 param2pos[i]);
3485 map = isl_map_project_out(map, isl_dim_param, i, 1);
3488 stmt->domain = isl_map_wrap(map);
3490 map = isl_set_unwrap(isl_set_copy(stmt->domain));
3491 map = isl_map_from_range(isl_map_domain(map));
3492 for (int pos = 0; pos < n; ++pos)
3493 stmt->args[pos] = embed(stmt->args[pos], map);
3494 isl_map_free(map);
3496 stmt = remove_nested_parameters(stmt);
3497 stmt = remove_duplicate_arguments(stmt, n);
3499 return stmt;
3500 error:
3501 pet_stmt_free(stmt);
3502 return NULL;
3505 /* For each statement in "scop", move the parameters that correspond
3506 * to nested access into the ranges of the domains and create
3507 * corresponding argument expressions.
3509 struct pet_scop *PetScan::resolve_nested(struct pet_scop *scop)
3511 if (!scop)
3512 return NULL;
3514 for (int i = 0; i < scop->n_stmt; ++i) {
3515 scop->stmts[i] = resolve_nested(scop->stmts[i]);
3516 if (!scop->stmts[i])
3517 goto error;
3520 return scop;
3521 error:
3522 pet_scop_free(scop);
3523 return NULL;
3526 /* Given an access expression "expr", is the variable accessed by
3527 * "expr" assigned anywhere inside "scop"?
3529 static bool is_assigned(pet_expr *expr, pet_scop *scop)
3531 bool assigned = false;
3532 isl_id *id;
3534 id = isl_map_get_tuple_id(expr->acc.access, isl_dim_out);
3535 assigned = pet_scop_writes(scop, id);
3536 isl_id_free(id);
3538 return assigned;
3541 /* Are all nested access parameters in "pa" allowed given "scop".
3542 * In particular, is none of them written by anywhere inside "scop".
3544 * If "scop" has any skip conditions, then no nested access parameters
3545 * are allowed. In particular, if there is any nested access in a guard
3546 * for a piece of code containing a "continue", then we want to introduce
3547 * a separate statement for evaluating this guard so that we can express
3548 * that the result is false for all previous iterations.
3550 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff *pa, pet_scop *scop)
3552 int nparam;
3554 if (!scop)
3555 return true;
3557 nparam = isl_pw_aff_dim(pa, isl_dim_param);
3558 for (int i = 0; i < nparam; ++i) {
3559 Expr *nested;
3560 isl_id *id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
3561 pet_expr *expr;
3562 bool allowed;
3564 if (!is_nested_parameter(id)) {
3565 isl_id_free(id);
3566 continue;
3569 if (pet_scop_has_skip(scop, pet_skip_now)) {
3570 isl_id_free(id);
3571 return false;
3574 nested = (Expr *) isl_id_get_user(id);
3575 expr = extract_expr(nested);
3576 allowed = expr && expr->type == pet_expr_access &&
3577 !is_assigned(expr, scop);
3579 pet_expr_free(expr);
3580 isl_id_free(id);
3582 if (!allowed)
3583 return false;
3586 return true;
3589 /* Do we need to construct a skip condition of the given type
3590 * on an if statement, given that the if condition is non-affine?
3592 * pet_scop_filter_skip can only handle the case where the if condition
3593 * holds (the then branch) and the skip condition is universal.
3594 * In any other case, we need to construct a new skip condition.
3596 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
3597 bool have_else, enum pet_skip type)
3599 if (have_else && scop_else && pet_scop_has_skip(scop_else, type))
3600 return true;
3601 if (scop_then && pet_scop_has_skip(scop_then, type) &&
3602 !pet_scop_has_universal_skip(scop_then, type))
3603 return true;
3604 return false;
3607 /* Do we need to construct a skip condition of the given type
3608 * on an if statement, given that the if condition is affine?
3610 * There is no need to construct a new skip condition if all
3611 * the skip conditions are affine.
3613 static bool need_skip_aff(struct pet_scop *scop_then,
3614 struct pet_scop *scop_else, bool have_else, enum pet_skip type)
3616 if (scop_then && pet_scop_has_var_skip(scop_then, type))
3617 return true;
3618 if (have_else && scop_else && pet_scop_has_var_skip(scop_else, type))
3619 return true;
3620 return false;
3623 /* Do we need to construct a skip condition of the given type
3624 * on an if statement?
3626 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
3627 bool have_else, enum pet_skip type, bool affine)
3629 if (affine)
3630 return need_skip_aff(scop_then, scop_else, have_else, type);
3631 else
3632 return need_skip(scop_then, scop_else, have_else, type);
3635 /* Construct an affine expression pet_expr that is evaluates
3636 * to the constant "val".
3638 static struct pet_expr *universally(isl_ctx *ctx, int val)
3640 isl_space *space;
3641 isl_map *map;
3643 space = isl_space_alloc(ctx, 0, 0, 1);
3644 map = isl_map_universe(space);
3645 map = isl_map_fix_si(map, isl_dim_out, 0, val);
3647 return pet_expr_from_access(map);
3650 /* Construct an affine expression pet_expr that is evaluates
3651 * to the constant 1.
3653 static struct pet_expr *universally_true(isl_ctx *ctx)
3655 return universally(ctx, 1);
3658 /* Construct an affine expression pet_expr that is evaluates
3659 * to the constant 0.
3661 static struct pet_expr *universally_false(isl_ctx *ctx)
3663 return universally(ctx, 0);
3666 /* Given an access relation "test_access" for the if condition,
3667 * an access relation "skip_access" for the skip condition and
3668 * scops for the then and else branches, construct a scop for
3669 * computing "skip_access".
3671 * The computed scop contains a single statement that essentially does
3673 * skip_cond = test_cond ? skip_cond_then : skip_cond_else
3675 * If the skip conditions of the then and/or else branch are not affine,
3676 * then they need to be filtered by test_access.
3677 * If they are missing, then this means the skip condition is false.
3679 * Since we are constructing a skip condition for the if statement,
3680 * the skip conditions on the then and else branches are removed.
3682 static struct pet_scop *extract_skip(PetScan *scan,
3683 __isl_take isl_map *test_access, __isl_take isl_map *skip_access,
3684 struct pet_scop *scop_then, struct pet_scop *scop_else, bool have_else,
3685 enum pet_skip type)
3687 struct pet_expr *expr_then, *expr_else, *expr, *expr_skip;
3688 struct pet_stmt *stmt;
3689 struct pet_scop *scop;
3690 isl_ctx *ctx = scan->ctx;
3692 if (!scop_then)
3693 goto error;
3694 if (have_else && !scop_else)
3695 goto error;
3697 if (pet_scop_has_skip(scop_then, type)) {
3698 expr_then = pet_scop_get_skip_expr(scop_then, type);
3699 pet_scop_reset_skip(scop_then, type);
3700 if (!pet_expr_is_affine(expr_then))
3701 expr_then = pet_expr_filter(expr_then,
3702 isl_map_copy(test_access), 1);
3703 } else
3704 expr_then = universally_false(ctx);
3706 if (have_else && pet_scop_has_skip(scop_else, type)) {
3707 expr_else = pet_scop_get_skip_expr(scop_else, type);
3708 pet_scop_reset_skip(scop_else, type);
3709 if (!pet_expr_is_affine(expr_else))
3710 expr_else = pet_expr_filter(expr_else,
3711 isl_map_copy(test_access), 0);
3712 } else
3713 expr_else = universally_false(ctx);
3715 expr = pet_expr_from_access(test_access);
3716 expr = pet_expr_new_ternary(ctx, expr, expr_then, expr_else);
3717 expr_skip = pet_expr_from_access(isl_map_copy(skip_access));
3718 if (expr_skip) {
3719 expr_skip->acc.write = 1;
3720 expr_skip->acc.read = 0;
3722 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
3723 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, scan->n_stmt++, expr);
3725 scop = pet_scop_from_pet_stmt(ctx, stmt);
3726 scop = scop_add_array(scop, skip_access, scan->ast_context);
3727 isl_map_free(skip_access);
3729 return scop;
3730 error:
3731 isl_map_free(test_access);
3732 isl_map_free(skip_access);
3733 return NULL;
3736 /* Is scop's skip_now condition equal to its skip_later condition?
3737 * In particular, this means that it either has no skip_now condition
3738 * or both a skip_now and a skip_later condition (that are equal to each other).
3740 static bool skip_equals_skip_later(struct pet_scop *scop)
3742 int has_skip_now, has_skip_later;
3743 int equal;
3744 isl_set *skip_now, *skip_later;
3746 if (!scop)
3747 return false;
3748 has_skip_now = pet_scop_has_skip(scop, pet_skip_now);
3749 has_skip_later = pet_scop_has_skip(scop, pet_skip_later);
3750 if (has_skip_now != has_skip_later)
3751 return false;
3752 if (!has_skip_now)
3753 return true;
3755 skip_now = pet_scop_get_skip(scop, pet_skip_now);
3756 skip_later = pet_scop_get_skip(scop, pet_skip_later);
3757 equal = isl_set_is_equal(skip_now, skip_later);
3758 isl_set_free(skip_now);
3759 isl_set_free(skip_later);
3761 return equal;
3764 /* Drop the skip conditions of type pet_skip_later from scop1 and scop2.
3766 static void drop_skip_later(struct pet_scop *scop1, struct pet_scop *scop2)
3768 pet_scop_reset_skip(scop1, pet_skip_later);
3769 pet_scop_reset_skip(scop2, pet_skip_later);
3772 /* Structure that handles the construction of skip conditions.
3774 * scop_then and scop_else represent the then and else branches
3775 * of the if statement
3777 * skip[type] is true if we need to construct a skip condition of that type
3778 * equal is set if the skip conditions of types pet_skip_now and pet_skip_later
3779 * are equal to each other
3780 * access[type] is the virtual array representing the skip condition
3781 * scop[type] is a scop for computing the skip condition
3783 struct pet_skip_info {
3784 isl_ctx *ctx;
3786 bool skip[2];
3787 bool equal;
3788 isl_map *access[2];
3789 struct pet_scop *scop[2];
3791 pet_skip_info(isl_ctx *ctx) : ctx(ctx) {}
3793 operator bool() { return skip[pet_skip_now] || skip[pet_skip_later]; }
3796 /* Structure that handles the construction of skip conditions on if statements.
3798 * scop_then and scop_else represent the then and else branches
3799 * of the if statement
3801 struct pet_skip_info_if : public pet_skip_info {
3802 struct pet_scop *scop_then, *scop_else;
3803 bool have_else;
3805 pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
3806 struct pet_scop *scop_else, bool have_else, bool affine);
3807 void extract(PetScan *scan, __isl_keep isl_map *access,
3808 enum pet_skip type);
3809 void extract(PetScan *scan, __isl_keep isl_map *access);
3810 void extract(PetScan *scan, __isl_keep isl_pw_aff *cond);
3811 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
3812 int offset);
3813 struct pet_scop *add(struct pet_scop *scop, int offset);
3816 /* Initialize a pet_skip_info_if structure based on the then and else branches
3817 * and based on whether the if condition is affine or not.
3819 pet_skip_info_if::pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
3820 struct pet_scop *scop_else, bool have_else, bool affine) :
3821 pet_skip_info(ctx), scop_then(scop_then), scop_else(scop_else),
3822 have_else(have_else)
3824 skip[pet_skip_now] =
3825 need_skip(scop_then, scop_else, have_else, pet_skip_now, affine);
3826 equal = skip[pet_skip_now] && skip_equals_skip_later(scop_then) &&
3827 (!have_else || skip_equals_skip_later(scop_else));
3828 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
3829 need_skip(scop_then, scop_else, have_else, pet_skip_later, affine);
3832 /* If we need to construct a skip condition of the given type,
3833 * then do so now.
3835 * "map" represents the if condition.
3837 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_map *map,
3838 enum pet_skip type)
3840 if (!skip[type])
3841 return;
3843 access[type] = create_test_access(isl_map_get_ctx(map), scan->n_test++);
3844 scop[type] = extract_skip(scan, isl_map_copy(map),
3845 isl_map_copy(access[type]),
3846 scop_then, scop_else, have_else, type);
3849 /* Construct the required skip conditions, given the if condition "map".
3851 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_map *map)
3853 extract(scan, map, pet_skip_now);
3854 extract(scan, map, pet_skip_later);
3855 if (equal)
3856 drop_skip_later(scop_then, scop_else);
3859 /* Construct the required skip conditions, given the if condition "cond".
3861 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_pw_aff *cond)
3863 isl_set *test_set;
3864 isl_map *test;
3866 if (!skip[pet_skip_now] && !skip[pet_skip_later])
3867 return;
3869 test_set = isl_set_from_pw_aff(isl_pw_aff_copy(cond));
3870 test = isl_map_from_range(test_set);
3871 extract(scan, test);
3872 isl_map_free(test);
3875 /* Add the computed skip condition of the give type to "main" and
3876 * add the scop for computing the condition at the given offset.
3878 * If equal is set, then we only computed a skip condition for pet_skip_now,
3879 * but we also need to set it as main's pet_skip_later.
3881 struct pet_scop *pet_skip_info_if::add(struct pet_scop *main,
3882 enum pet_skip type, int offset)
3884 isl_set *skip_set;
3886 if (!skip[type])
3887 return main;
3889 skip_set = isl_map_range(access[type]);
3890 access[type] = NULL;
3891 scop[type] = pet_scop_prefix(scop[type], offset);
3892 main = pet_scop_add_par(ctx, main, scop[type]);
3893 scop[type] = NULL;
3895 if (equal)
3896 main = pet_scop_set_skip(main, pet_skip_later,
3897 isl_set_copy(skip_set));
3899 main = pet_scop_set_skip(main, type, skip_set);
3901 return main;
3904 /* Add the computed skip conditions to "main" and
3905 * add the scops for computing the conditions at the given offset.
3907 struct pet_scop *pet_skip_info_if::add(struct pet_scop *scop, int offset)
3909 scop = add(scop, pet_skip_now, offset);
3910 scop = add(scop, pet_skip_later, offset);
3912 return scop;
3915 /* Construct a pet_scop for a non-affine if statement.
3917 * We create a separate statement that writes the result
3918 * of the non-affine condition to a virtual scalar.
3919 * A constraint requiring the value of this virtual scalar to be one
3920 * is added to the iteration domains of the then branch.
3921 * Similarly, a constraint requiring the value of this virtual scalar
3922 * to be zero is added to the iteration domains of the else branch, if any.
3923 * We adjust the schedules to ensure that the virtual scalar is written
3924 * before it is read.
3926 * If there are any breaks or continues in the then and/or else
3927 * branches, then we may have to compute a new skip condition.
3928 * This is handled using a pet_skip_info_if object.
3929 * On initialization, the object checks if skip conditions need
3930 * to be computed. If so, it does so in "extract" and adds them in "add".
3932 struct pet_scop *PetScan::extract_non_affine_if(Expr *cond,
3933 struct pet_scop *scop_then, struct pet_scop *scop_else,
3934 bool have_else, int stmt_id)
3936 struct pet_scop *scop;
3937 isl_map *test_access;
3938 int save_n_stmt = n_stmt;
3940 test_access = create_test_access(ctx, n_test++);
3941 n_stmt = stmt_id;
3942 scop = extract_non_affine_condition(cond, isl_map_copy(test_access));
3943 n_stmt = save_n_stmt;
3944 scop = scop_add_array(scop, test_access, ast_context);
3946 pet_skip_info_if skip(ctx, scop_then, scop_else, have_else, false);
3947 skip.extract(this, test_access);
3949 scop = pet_scop_prefix(scop, 0);
3950 scop_then = pet_scop_prefix(scop_then, 1);
3951 scop_then = pet_scop_filter(scop_then, isl_map_copy(test_access), 1);
3952 if (have_else) {
3953 scop_else = pet_scop_prefix(scop_else, 1);
3954 scop_else = pet_scop_filter(scop_else, test_access, 0);
3955 scop_then = pet_scop_add_par(ctx, scop_then, scop_else);
3956 } else
3957 isl_map_free(test_access);
3959 scop = pet_scop_add_seq(ctx, scop, scop_then);
3961 scop = skip.add(scop, 2);
3963 return scop;
3966 /* Construct a pet_scop for an if statement.
3968 * If the condition fits the pattern of a conditional assignment,
3969 * then it is handled by extract_conditional_assignment.
3970 * Otherwise, we do the following.
3972 * If the condition is affine, then the condition is added
3973 * to the iteration domains of the then branch, while the
3974 * opposite of the condition in added to the iteration domains
3975 * of the else branch, if any.
3976 * We allow the condition to be dynamic, i.e., to refer to
3977 * scalars or array elements that may be written to outside
3978 * of the given if statement. These nested accesses are then represented
3979 * as output dimensions in the wrapping iteration domain.
3980 * If it also written _inside_ the then or else branch, then
3981 * we treat the condition as non-affine.
3982 * As explained in extract_non_affine_if, this will introduce
3983 * an extra statement.
3984 * For aesthetic reasons, we want this statement to have a statement
3985 * number that is lower than those of the then and else branches.
3986 * In order to evaluate if will need such a statement, however, we
3987 * first construct scops for the then and else branches.
3988 * We therefore reserve a statement number if we might have to
3989 * introduce such an extra statement.
3991 * If the condition is not affine, then the scop is created in
3992 * extract_non_affine_if.
3994 * If there are any breaks or continues in the then and/or else
3995 * branches, then we may have to compute a new skip condition.
3996 * This is handled using a pet_skip_info_if object.
3997 * On initialization, the object checks if skip conditions need
3998 * to be computed. If so, it does so in "extract" and adds them in "add".
4000 struct pet_scop *PetScan::extract(IfStmt *stmt)
4002 struct pet_scop *scop_then, *scop_else = NULL, *scop;
4003 isl_pw_aff *cond;
4004 int stmt_id;
4005 isl_set *set;
4006 isl_set *valid;
4008 scop = extract_conditional_assignment(stmt);
4009 if (scop)
4010 return scop;
4012 cond = try_extract_nested_condition(stmt->getCond());
4013 if (allow_nested && (!cond || has_nested(cond)))
4014 stmt_id = n_stmt++;
4017 assigned_value_cache cache(assigned_value);
4018 scop_then = extract(stmt->getThen());
4021 if (stmt->getElse()) {
4022 assigned_value_cache cache(assigned_value);
4023 scop_else = extract(stmt->getElse());
4024 if (options->autodetect) {
4025 if (scop_then && !scop_else) {
4026 partial = true;
4027 isl_pw_aff_free(cond);
4028 return scop_then;
4030 if (!scop_then && scop_else) {
4031 partial = true;
4032 isl_pw_aff_free(cond);
4033 return scop_else;
4038 if (cond &&
4039 (!is_nested_allowed(cond, scop_then) ||
4040 (stmt->getElse() && !is_nested_allowed(cond, scop_else)))) {
4041 isl_pw_aff_free(cond);
4042 cond = NULL;
4044 if (allow_nested && !cond)
4045 return extract_non_affine_if(stmt->getCond(), scop_then,
4046 scop_else, stmt->getElse(), stmt_id);
4048 if (!cond)
4049 cond = extract_condition(stmt->getCond());
4051 pet_skip_info_if skip(ctx, scop_then, scop_else, stmt->getElse(), true);
4052 skip.extract(this, cond);
4054 valid = isl_pw_aff_domain(isl_pw_aff_copy(cond));
4055 set = isl_pw_aff_non_zero_set(cond);
4056 scop = pet_scop_restrict(scop_then, isl_set_copy(set));
4058 if (stmt->getElse()) {
4059 set = isl_set_subtract(isl_set_copy(valid), set);
4060 scop_else = pet_scop_restrict(scop_else, set);
4061 scop = pet_scop_add_par(ctx, scop, scop_else);
4062 } else
4063 isl_set_free(set);
4064 scop = resolve_nested(scop);
4065 scop = pet_scop_restrict_context(scop, valid);
4067 if (skip)
4068 scop = pet_scop_prefix(scop, 0);
4069 scop = skip.add(scop, 1);
4071 return scop;
4074 /* Try and construct a pet_scop for a label statement.
4075 * We currently only allow labels on expression statements.
4077 struct pet_scop *PetScan::extract(LabelStmt *stmt)
4079 isl_id *label;
4080 Stmt *sub;
4082 sub = stmt->getSubStmt();
4083 if (!isa<Expr>(sub)) {
4084 unsupported(stmt);
4085 return NULL;
4088 label = isl_id_alloc(ctx, stmt->getName(), NULL);
4090 return extract(sub, extract_expr(cast<Expr>(sub)), label);
4093 /* Construct a pet_scop for a continue statement.
4095 * We simply create an empty scop with a universal pet_skip_now
4096 * skip condition. This skip condition will then be taken into
4097 * account by the enclosing loop construct, possibly after
4098 * being incorporated into outer skip conditions.
4100 struct pet_scop *PetScan::extract(ContinueStmt *stmt)
4102 pet_scop *scop;
4103 isl_space *space;
4104 isl_set *set;
4106 scop = pet_scop_empty(ctx);
4107 if (!scop)
4108 return NULL;
4110 space = isl_space_set_alloc(ctx, 0, 1);
4111 set = isl_set_universe(space);
4112 set = isl_set_fix_si(set, isl_dim_set, 0, 1);
4113 scop = pet_scop_set_skip(scop, pet_skip_now, set);
4115 return scop;
4118 /* Construct a pet_scop for a break statement.
4120 * We simply create an empty scop with both a universal pet_skip_now
4121 * skip condition and a universal pet_skip_later skip condition.
4122 * These skip conditions will then be taken into
4123 * account by the enclosing loop construct, possibly after
4124 * being incorporated into outer skip conditions.
4126 struct pet_scop *PetScan::extract(BreakStmt *stmt)
4128 pet_scop *scop;
4129 isl_space *space;
4130 isl_set *set;
4132 scop = pet_scop_empty(ctx);
4133 if (!scop)
4134 return NULL;
4136 space = isl_space_set_alloc(ctx, 0, 1);
4137 set = isl_set_universe(space);
4138 set = isl_set_fix_si(set, isl_dim_set, 0, 1);
4139 scop = pet_scop_set_skip(scop, pet_skip_now, isl_set_copy(set));
4140 scop = pet_scop_set_skip(scop, pet_skip_later, set);
4142 return scop;
4145 /* Try and construct a pet_scop corresponding to "stmt".
4147 struct pet_scop *PetScan::extract(Stmt *stmt)
4149 if (isa<Expr>(stmt))
4150 return extract(stmt, extract_expr(cast<Expr>(stmt)));
4152 switch (stmt->getStmtClass()) {
4153 case Stmt::WhileStmtClass:
4154 return extract(cast<WhileStmt>(stmt));
4155 case Stmt::ForStmtClass:
4156 return extract_for(cast<ForStmt>(stmt));
4157 case Stmt::IfStmtClass:
4158 return extract(cast<IfStmt>(stmt));
4159 case Stmt::CompoundStmtClass:
4160 return extract(cast<CompoundStmt>(stmt));
4161 case Stmt::LabelStmtClass:
4162 return extract(cast<LabelStmt>(stmt));
4163 case Stmt::ContinueStmtClass:
4164 return extract(cast<ContinueStmt>(stmt));
4165 case Stmt::BreakStmtClass:
4166 return extract(cast<BreakStmt>(stmt));
4167 default:
4168 unsupported(stmt);
4171 return NULL;
4174 /* Do we need to construct a skip condition of the given type
4175 * on a sequence of statements?
4177 * There is no need to construct a new skip condition if only
4178 * only of the two statements has a skip condition or if both
4179 * of their skip conditions are affine.
4181 * In principle we also don't need a new continuation variable if
4182 * the continuation of scop2 is affine, but then we would need
4183 * to allow more complicated forms of continuations.
4185 static bool need_skip_seq(struct pet_scop *scop1, struct pet_scop *scop2,
4186 enum pet_skip type)
4188 if (!scop1 || !pet_scop_has_skip(scop1, type))
4189 return false;
4190 if (!scop2 || !pet_scop_has_skip(scop2, type))
4191 return false;
4192 if (pet_scop_has_affine_skip(scop1, type) &&
4193 pet_scop_has_affine_skip(scop2, type))
4194 return false;
4195 return true;
4198 /* Construct a scop for computing the skip condition of the given type and
4199 * with access relation "skip_access" for a sequence of two scops "scop1"
4200 * and "scop2".
4202 * The computed scop contains a single statement that essentially does
4204 * skip_cond = skip_cond_1 ? 1 : skip_cond_2
4206 * or, in other words, skip_cond1 || skip_cond2.
4207 * In this expression, skip_cond_2 is filtered to reflect that it is
4208 * only evaluated when skip_cond_1 is false.
4210 * The skip condition on scop1 is not removed because it still needs
4211 * to be applied to scop2 when these two scops are combined.
4213 static struct pet_scop *extract_skip_seq(PetScan *ps,
4214 __isl_take isl_map *skip_access,
4215 struct pet_scop *scop1, struct pet_scop *scop2, enum pet_skip type)
4217 isl_map *access;
4218 struct pet_expr *expr1, *expr2, *expr, *expr_skip;
4219 struct pet_stmt *stmt;
4220 struct pet_scop *scop;
4221 isl_ctx *ctx = ps->ctx;
4223 if (!scop1 || !scop2)
4224 goto error;
4226 expr1 = pet_scop_get_skip_expr(scop1, type);
4227 expr2 = pet_scop_get_skip_expr(scop2, type);
4228 pet_scop_reset_skip(scop2, type);
4230 expr2 = pet_expr_filter(expr2, isl_map_copy(expr1->acc.access), 0);
4232 expr = universally_true(ctx);
4233 expr = pet_expr_new_ternary(ctx, expr1, expr, expr2);
4234 expr_skip = pet_expr_from_access(isl_map_copy(skip_access));
4235 if (expr_skip) {
4236 expr_skip->acc.write = 1;
4237 expr_skip->acc.read = 0;
4239 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
4240 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, ps->n_stmt++, expr);
4242 scop = pet_scop_from_pet_stmt(ctx, stmt);
4243 scop = scop_add_array(scop, skip_access, ps->ast_context);
4244 isl_map_free(skip_access);
4246 return scop;
4247 error:
4248 isl_map_free(skip_access);
4249 return NULL;
4252 /* Structure that handles the construction of skip conditions
4253 * on sequences of statements.
4255 * scop1 and scop2 represent the two statements that are combined
4257 struct pet_skip_info_seq : public pet_skip_info {
4258 struct pet_scop *scop1, *scop2;
4260 pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4261 struct pet_scop *scop2);
4262 void extract(PetScan *scan, enum pet_skip type);
4263 void extract(PetScan *scan);
4264 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
4265 int offset);
4266 struct pet_scop *add(struct pet_scop *scop, int offset);
4269 /* Initialize a pet_skip_info_seq structure based on
4270 * on the two statements that are going to be combined.
4272 pet_skip_info_seq::pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4273 struct pet_scop *scop2) : pet_skip_info(ctx), scop1(scop1), scop2(scop2)
4275 skip[pet_skip_now] = need_skip_seq(scop1, scop2, pet_skip_now);
4276 equal = skip[pet_skip_now] && skip_equals_skip_later(scop1) &&
4277 skip_equals_skip_later(scop2);
4278 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
4279 need_skip_seq(scop1, scop2, pet_skip_later);
4282 /* If we need to construct a skip condition of the given type,
4283 * then do so now.
4285 void pet_skip_info_seq::extract(PetScan *scan, enum pet_skip type)
4287 if (!skip[type])
4288 return;
4290 access[type] = create_test_access(ctx, scan->n_test++);
4291 scop[type] = extract_skip_seq(scan, isl_map_copy(access[type]),
4292 scop1, scop2, type);
4295 /* Construct the required skip conditions.
4297 void pet_skip_info_seq::extract(PetScan *scan)
4299 extract(scan, pet_skip_now);
4300 extract(scan, pet_skip_later);
4301 if (equal)
4302 drop_skip_later(scop1, scop2);
4305 /* Add the computed skip condition of the give type to "main" and
4306 * add the scop for computing the condition at the given offset (the statement
4307 * number). Within this offset, the condition is computed at position 1
4308 * to ensure that it is computed after the corresponding statement.
4310 * If equal is set, then we only computed a skip condition for pet_skip_now,
4311 * but we also need to set it as main's pet_skip_later.
4313 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *main,
4314 enum pet_skip type, int offset)
4316 isl_set *skip_set;
4318 if (!skip[type])
4319 return main;
4321 skip_set = isl_map_range(access[type]);
4322 access[type] = NULL;
4323 scop[type] = pet_scop_prefix(scop[type], 1);
4324 scop[type] = pet_scop_prefix(scop[type], offset);
4325 main = pet_scop_add_par(ctx, main, scop[type]);
4326 scop[type] = NULL;
4328 if (equal)
4329 main = pet_scop_set_skip(main, pet_skip_later,
4330 isl_set_copy(skip_set));
4332 main = pet_scop_set_skip(main, type, skip_set);
4334 return main;
4337 /* Add the computed skip conditions to "main" and
4338 * add the scops for computing the conditions at the given offset.
4340 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *scop, int offset)
4342 scop = add(scop, pet_skip_now, offset);
4343 scop = add(scop, pet_skip_later, offset);
4345 return scop;
4348 /* Try and construct a pet_scop corresponding to (part of)
4349 * a sequence of statements.
4351 * If there are any breaks or continues in the individual statements,
4352 * then we may have to compute a new skip condition.
4353 * This is handled using a pet_skip_info_seq object.
4354 * On initialization, the object checks if skip conditions need
4355 * to be computed. If so, it does so in "extract" and adds them in "add".
4357 struct pet_scop *PetScan::extract(StmtRange stmt_range)
4359 pet_scop *scop;
4360 StmtIterator i;
4361 int j;
4362 bool partial_range = false;
4364 scop = pet_scop_empty(ctx);
4365 for (i = stmt_range.first, j = 0; i != stmt_range.second; ++i, ++j) {
4366 Stmt *child = *i;
4367 struct pet_scop *scop_i;
4369 scop_i = extract(child);
4370 if (scop && partial) {
4371 pet_scop_free(scop_i);
4372 break;
4374 pet_skip_info_seq skip(ctx, scop, scop_i);
4375 skip.extract(this);
4376 if (skip)
4377 scop_i = pet_scop_prefix(scop_i, 0);
4378 scop_i = pet_scop_prefix(scop_i, j);
4379 if (options->autodetect) {
4380 if (scop_i)
4381 scop = pet_scop_add_seq(ctx, scop, scop_i);
4382 else
4383 partial_range = true;
4384 if (scop->n_stmt != 0 && !scop_i)
4385 partial = true;
4386 } else {
4387 scop = pet_scop_add_seq(ctx, scop, scop_i);
4390 scop = skip.add(scop, j);
4392 if (partial)
4393 break;
4396 if (scop && partial_range)
4397 partial = true;
4399 return scop;
4402 /* Return the file offset of the expansion location of "Loc".
4404 static unsigned getExpansionOffset(SourceManager &SM, SourceLocation Loc)
4406 return SM.getFileOffset(SM.getExpansionLoc(Loc));
4409 /* Check if the scop marked by the user is exactly this Stmt
4410 * or part of this Stmt.
4411 * If so, return a pet_scop corresponding to the marked region.
4412 * Otherwise, return NULL.
4414 struct pet_scop *PetScan::scan(Stmt *stmt)
4416 SourceManager &SM = PP.getSourceManager();
4417 unsigned start_off, end_off;
4419 start_off = getExpansionOffset(SM, stmt->getLocStart());
4420 end_off = getExpansionOffset(SM, stmt->getLocEnd());
4422 if (start_off > loc.end)
4423 return NULL;
4424 if (end_off < loc.start)
4425 return NULL;
4426 if (start_off >= loc.start && end_off <= loc.end) {
4427 return extract(stmt);
4430 StmtIterator start;
4431 for (start = stmt->child_begin(); start != stmt->child_end(); ++start) {
4432 Stmt *child = *start;
4433 if (!child)
4434 continue;
4435 start_off = getExpansionOffset(SM, child->getLocStart());
4436 end_off = getExpansionOffset(SM, child->getLocEnd());
4437 if (start_off < loc.start && end_off > loc.end)
4438 return scan(child);
4439 if (start_off >= loc.start)
4440 break;
4443 StmtIterator end;
4444 for (end = start; end != stmt->child_end(); ++end) {
4445 Stmt *child = *end;
4446 start_off = SM.getFileOffset(child->getLocStart());
4447 if (start_off >= loc.end)
4448 break;
4451 return extract(StmtRange(start, end));
4454 /* Set the size of index "pos" of "array" to "size".
4455 * In particular, add a constraint of the form
4457 * i_pos < size
4459 * to array->extent and a constraint of the form
4461 * size >= 0
4463 * to array->context.
4465 static struct pet_array *update_size(struct pet_array *array, int pos,
4466 __isl_take isl_pw_aff *size)
4468 isl_set *valid;
4469 isl_set *univ;
4470 isl_set *bound;
4471 isl_space *dim;
4472 isl_aff *aff;
4473 isl_pw_aff *index;
4474 isl_id *id;
4476 valid = isl_pw_aff_nonneg_set(isl_pw_aff_copy(size));
4477 array->context = isl_set_intersect(array->context, valid);
4479 dim = isl_set_get_space(array->extent);
4480 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
4481 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, pos, 1);
4482 univ = isl_set_universe(isl_aff_get_domain_space(aff));
4483 index = isl_pw_aff_alloc(univ, aff);
4485 size = isl_pw_aff_add_dims(size, isl_dim_in,
4486 isl_set_dim(array->extent, isl_dim_set));
4487 id = isl_set_get_tuple_id(array->extent);
4488 size = isl_pw_aff_set_tuple_id(size, isl_dim_in, id);
4489 bound = isl_pw_aff_lt_set(index, size);
4491 array->extent = isl_set_intersect(array->extent, bound);
4493 if (!array->context || !array->extent)
4494 goto error;
4496 return array;
4497 error:
4498 pet_array_free(array);
4499 return NULL;
4502 /* Figure out the size of the array at position "pos" and all
4503 * subsequent positions from "type" and update "array" accordingly.
4505 struct pet_array *PetScan::set_upper_bounds(struct pet_array *array,
4506 const Type *type, int pos)
4508 const ArrayType *atype;
4509 isl_pw_aff *size;
4511 if (!array)
4512 return NULL;
4514 if (type->isPointerType()) {
4515 type = type->getPointeeType().getTypePtr();
4516 return set_upper_bounds(array, type, pos + 1);
4518 if (!type->isArrayType())
4519 return array;
4521 type = type->getCanonicalTypeInternal().getTypePtr();
4522 atype = cast<ArrayType>(type);
4524 if (type->isConstantArrayType()) {
4525 const ConstantArrayType *ca = cast<ConstantArrayType>(atype);
4526 size = extract_affine(ca->getSize());
4527 array = update_size(array, pos, size);
4528 } else if (type->isVariableArrayType()) {
4529 const VariableArrayType *vla = cast<VariableArrayType>(atype);
4530 size = extract_affine(vla->getSizeExpr());
4531 array = update_size(array, pos, size);
4534 type = atype->getElementType().getTypePtr();
4536 return set_upper_bounds(array, type, pos + 1);
4539 /* Construct and return a pet_array corresponding to the variable "decl".
4540 * In particular, initialize array->extent to
4542 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
4544 * and then call set_upper_bounds to set the upper bounds on the indices
4545 * based on the type of the variable.
4547 struct pet_array *PetScan::extract_array(isl_ctx *ctx, ValueDecl *decl)
4549 struct pet_array *array;
4550 QualType qt = decl->getType();
4551 const Type *type = qt.getTypePtr();
4552 int depth = array_depth(type);
4553 QualType base = base_type(qt);
4554 string name;
4555 isl_id *id;
4556 isl_space *dim;
4558 array = isl_calloc_type(ctx, struct pet_array);
4559 if (!array)
4560 return NULL;
4562 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
4563 dim = isl_space_set_alloc(ctx, 0, depth);
4564 dim = isl_space_set_tuple_id(dim, isl_dim_set, id);
4566 array->extent = isl_set_nat_universe(dim);
4568 dim = isl_space_params_alloc(ctx, 0);
4569 array->context = isl_set_universe(dim);
4571 array = set_upper_bounds(array, type, 0);
4572 if (!array)
4573 return NULL;
4575 name = base.getAsString();
4576 array->element_type = strdup(name.c_str());
4577 array->element_size = decl->getASTContext().getTypeInfo(base).first / 8;
4579 return array;
4582 /* Construct a list of pet_arrays, one for each array (or scalar)
4583 * accessed inside "scop", add this list to "scop" and return the result.
4585 * The context of "scop" is updated with the intersection of
4586 * the contexts of all arrays, i.e., constraints on the parameters
4587 * that ensure that the arrays have a valid (non-negative) size.
4589 struct pet_scop *PetScan::scan_arrays(struct pet_scop *scop)
4591 int i;
4592 set<ValueDecl *> arrays;
4593 set<ValueDecl *>::iterator it;
4594 int n_array;
4595 struct pet_array **scop_arrays;
4597 if (!scop)
4598 return NULL;
4600 pet_scop_collect_arrays(scop, arrays);
4601 if (arrays.size() == 0)
4602 return scop;
4604 n_array = scop->n_array;
4606 scop_arrays = isl_realloc_array(ctx, scop->arrays, struct pet_array *,
4607 n_array + arrays.size());
4608 if (!scop_arrays)
4609 goto error;
4610 scop->arrays = scop_arrays;
4612 for (it = arrays.begin(), i = 0; it != arrays.end(); ++it, ++i) {
4613 struct pet_array *array;
4614 scop->arrays[n_array + i] = array = extract_array(ctx, *it);
4615 if (!scop->arrays[n_array + i])
4616 goto error;
4617 scop->n_array++;
4618 scop->context = isl_set_intersect(scop->context,
4619 isl_set_copy(array->context));
4620 if (!scop->context)
4621 goto error;
4624 return scop;
4625 error:
4626 pet_scop_free(scop);
4627 return NULL;
4630 /* Bound all parameters in scop->context to the possible values
4631 * of the corresponding C variable.
4633 static struct pet_scop *add_parameter_bounds(struct pet_scop *scop)
4635 int n;
4637 if (!scop)
4638 return NULL;
4640 n = isl_set_dim(scop->context, isl_dim_param);
4641 for (int i = 0; i < n; ++i) {
4642 isl_id *id;
4643 ValueDecl *decl;
4645 id = isl_set_get_dim_id(scop->context, isl_dim_param, i);
4646 if (is_nested_parameter(id)) {
4647 isl_id_free(id);
4648 isl_die(isl_set_get_ctx(scop->context),
4649 isl_error_internal,
4650 "unresolved nested parameter", goto error);
4652 decl = (ValueDecl *) isl_id_get_user(id);
4653 isl_id_free(id);
4655 scop->context = set_parameter_bounds(scop->context, i, decl);
4657 if (!scop->context)
4658 goto error;
4661 return scop;
4662 error:
4663 pet_scop_free(scop);
4664 return NULL;
4667 /* Construct a pet_scop from the given function.
4669 struct pet_scop *PetScan::scan(FunctionDecl *fd)
4671 pet_scop *scop;
4672 Stmt *stmt;
4674 stmt = fd->getBody();
4676 if (options->autodetect)
4677 scop = extract(stmt);
4678 else
4679 scop = scan(stmt);
4680 scop = pet_scop_detect_parameter_accesses(scop);
4681 scop = scan_arrays(scop);
4682 scop = add_parameter_bounds(scop);
4683 scop = pet_scop_gist(scop, value_bounds);
4685 return scop;