pet_expr: merge unary, binary and ternary types into operation type
[pet.git] / scan.cc
blobcb95428677e10f1984bebf93131b9fbc39f72c9e
1 /*
2 * Copyright 2011 Leiden University. All rights reserved.
3 * Copyright 2012-2014 Ecole Normale Superieure. All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 *
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above
13 * copyright notice, this list of conditions and the following
14 * disclaimer in the documentation and/or other materials provided
15 * with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY LEIDEN UNIVERSITY ''AS IS'' AND ANY
18 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL LEIDEN UNIVERSITY OR
21 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
22 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
23 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA,
24 * OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
27 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 * The views and conclusions contained in the software and documentation
30 * are those of the authors and should not be interpreted as
31 * representing official policies, either expressed or implied, of
32 * Leiden University.
33 */
35 #include <string.h>
36 #include <set>
37 #include <map>
38 #include <iostream>
39 #include <llvm/Support/raw_ostream.h>
40 #include <clang/AST/ASTContext.h>
41 #include <clang/AST/ASTDiagnostic.h>
42 #include <clang/AST/Expr.h>
43 #include <clang/AST/RecursiveASTVisitor.h>
45 #include <isl/id.h>
46 #include <isl/space.h>
47 #include <isl/aff.h>
48 #include <isl/set.h>
50 #include "clang.h"
51 #include "options.h"
52 #include "scan.h"
53 #include "scop.h"
54 #include "scop_plus.h"
56 #include "config.h"
58 using namespace std;
59 using namespace clang;
61 #if defined(DECLREFEXPR_CREATE_REQUIRES_BOOL)
62 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
64 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
65 SourceLocation(), var, false, var->getInnerLocStart(),
66 var->getType(), VK_LValue);
68 #elif defined(DECLREFEXPR_CREATE_REQUIRES_SOURCELOCATION)
69 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
71 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
72 SourceLocation(), var, var->getInnerLocStart(), var->getType(),
73 VK_LValue);
75 #else
76 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
78 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
79 var, var->getInnerLocStart(), var->getType(), VK_LValue);
81 #endif
83 /* Check if the element type corresponding to the given array type
84 * has a const qualifier.
86 static bool const_base(QualType qt)
88 const Type *type = qt.getTypePtr();
90 if (type->isPointerType())
91 return const_base(type->getPointeeType());
92 if (type->isArrayType()) {
93 const ArrayType *atype;
94 type = type->getCanonicalTypeInternal().getTypePtr();
95 atype = cast<ArrayType>(type);
96 return const_base(atype->getElementType());
99 return qt.isConstQualified();
102 /* Mark "decl" as having an unknown value in "assigned_value".
104 * If no (known or unknown) value was assigned to "decl" before,
105 * then it may have been treated as a parameter before and may
106 * therefore appear in a value assigned to another variable.
107 * If so, this assignment needs to be turned into an unknown value too.
109 static void clear_assignment(map<ValueDecl *, isl_pw_aff *> &assigned_value,
110 ValueDecl *decl)
112 map<ValueDecl *, isl_pw_aff *>::iterator it;
114 it = assigned_value.find(decl);
116 assigned_value[decl] = NULL;
118 if (it != assigned_value.end())
119 return;
121 for (it = assigned_value.begin(); it != assigned_value.end(); ++it) {
122 isl_pw_aff *pa = it->second;
123 int nparam = isl_pw_aff_dim(pa, isl_dim_param);
125 for (int i = 0; i < nparam; ++i) {
126 isl_id *id;
128 if (!isl_pw_aff_has_dim_id(pa, isl_dim_param, i))
129 continue;
130 id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
131 if (isl_id_get_user(id) == decl)
132 it->second = NULL;
133 isl_id_free(id);
138 /* Look for any assignments to scalar variables in part of the parse
139 * tree and set assigned_value to NULL for each of them.
140 * Also reset assigned_value if the address of a scalar variable
141 * is being taken. As an exception, if the address is passed to a function
142 * that is declared to receive a const pointer, then assigned_value is
143 * not reset.
145 * This ensures that we won't use any previously stored value
146 * in the current subtree and its parents.
148 struct clear_assignments : RecursiveASTVisitor<clear_assignments> {
149 map<ValueDecl *, isl_pw_aff *> &assigned_value;
150 set<UnaryOperator *> skip;
152 clear_assignments(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
153 assigned_value(assigned_value) {}
155 /* Check for "address of" operators whose value is passed
156 * to a const pointer argument and add them to "skip", so that
157 * we can skip them in VisitUnaryOperator.
159 bool VisitCallExpr(CallExpr *expr) {
160 FunctionDecl *fd;
161 fd = expr->getDirectCallee();
162 if (!fd)
163 return true;
164 for (int i = 0; i < expr->getNumArgs(); ++i) {
165 Expr *arg = expr->getArg(i);
166 UnaryOperator *op;
167 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
168 ImplicitCastExpr *ice;
169 ice = cast<ImplicitCastExpr>(arg);
170 arg = ice->getSubExpr();
172 if (arg->getStmtClass() != Stmt::UnaryOperatorClass)
173 continue;
174 op = cast<UnaryOperator>(arg);
175 if (op->getOpcode() != UO_AddrOf)
176 continue;
177 if (const_base(fd->getParamDecl(i)->getType()))
178 skip.insert(op);
180 return true;
183 bool VisitUnaryOperator(UnaryOperator *expr) {
184 Expr *arg;
185 DeclRefExpr *ref;
186 ValueDecl *decl;
188 switch (expr->getOpcode()) {
189 case UO_AddrOf:
190 case UO_PostInc:
191 case UO_PostDec:
192 case UO_PreInc:
193 case UO_PreDec:
194 break;
195 default:
196 return true;
198 if (skip.find(expr) != skip.end())
199 return true;
201 arg = expr->getSubExpr();
202 if (arg->getStmtClass() != Stmt::DeclRefExprClass)
203 return true;
204 ref = cast<DeclRefExpr>(arg);
205 decl = ref->getDecl();
206 clear_assignment(assigned_value, decl);
207 return true;
210 bool VisitBinaryOperator(BinaryOperator *expr) {
211 Expr *lhs;
212 DeclRefExpr *ref;
213 ValueDecl *decl;
215 if (!expr->isAssignmentOp())
216 return true;
217 lhs = expr->getLHS();
218 if (lhs->getStmtClass() != Stmt::DeclRefExprClass)
219 return true;
220 ref = cast<DeclRefExpr>(lhs);
221 decl = ref->getDecl();
222 clear_assignment(assigned_value, decl);
223 return true;
227 /* Keep a copy of the currently assigned values.
229 * Any variable that is assigned a value inside the current scope
230 * is removed again when we leave the scope (either because it wasn't
231 * stored in the cache or because it has a different value in the cache).
233 struct assigned_value_cache {
234 map<ValueDecl *, isl_pw_aff *> &assigned_value;
235 map<ValueDecl *, isl_pw_aff *> cache;
237 assigned_value_cache(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
238 assigned_value(assigned_value), cache(assigned_value) {}
239 ~assigned_value_cache() {
240 map<ValueDecl *, isl_pw_aff *>::iterator it = cache.begin();
241 for (it = assigned_value.begin(); it != assigned_value.end();
242 ++it) {
243 if (!it->second ||
244 (cache.find(it->first) != cache.end() &&
245 cache[it->first] != it->second))
246 cache[it->first] = NULL;
248 assigned_value = cache;
252 /* Insert an expression into the collection of expressions,
253 * provided it is not already in there.
254 * The isl_pw_affs are freed in the destructor.
256 void PetScan::insert_expression(__isl_take isl_pw_aff *expr)
258 std::set<isl_pw_aff *>::iterator it;
260 if (expressions.find(expr) == expressions.end())
261 expressions.insert(expr);
262 else
263 isl_pw_aff_free(expr);
266 PetScan::~PetScan()
268 std::set<isl_pw_aff *>::iterator it;
270 for (it = expressions.begin(); it != expressions.end(); ++it)
271 isl_pw_aff_free(*it);
273 isl_union_map_free(value_bounds);
276 /* Report a diagnostic, unless autodetect is set.
278 void PetScan::report(Stmt *stmt, unsigned id)
280 if (options->autodetect)
281 return;
283 SourceLocation loc = stmt->getLocStart();
284 DiagnosticsEngine &diag = PP.getDiagnostics();
285 DiagnosticBuilder B = diag.Report(loc, id) << stmt->getSourceRange();
288 /* Called if we found something we (currently) cannot handle.
289 * We'll provide more informative warnings later.
291 * We only actually complain if autodetect is false.
293 void PetScan::unsupported(Stmt *stmt)
295 DiagnosticsEngine &diag = PP.getDiagnostics();
296 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
297 "unsupported");
298 report(stmt, id);
301 /* Report a missing prototype, unless autodetect is set.
303 void PetScan::report_prototype_required(Stmt *stmt)
305 DiagnosticsEngine &diag = PP.getDiagnostics();
306 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
307 "prototype required");
308 report(stmt, id);
311 /* Report a missing increment, unless autodetect is set.
313 void PetScan::report_missing_increment(Stmt *stmt)
315 DiagnosticsEngine &diag = PP.getDiagnostics();
316 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
317 "missing increment");
318 report(stmt, id);
321 /* Extract an integer from "expr".
323 __isl_give isl_val *PetScan::extract_int(isl_ctx *ctx, IntegerLiteral *expr)
325 const Type *type = expr->getType().getTypePtr();
326 int is_signed = type->hasSignedIntegerRepresentation();
327 llvm::APInt val = expr->getValue();
328 int is_negative = is_signed && val.isNegative();
329 isl_val *v;
331 if (is_negative)
332 val = -val;
334 v = extract_unsigned(ctx, val);
336 if (is_negative)
337 v = isl_val_neg(v);
338 return v;
341 /* Extract an integer from "val", which is assumed to be non-negative.
343 __isl_give isl_val *PetScan::extract_unsigned(isl_ctx *ctx,
344 const llvm::APInt &val)
346 unsigned n;
347 const uint64_t *data;
349 data = val.getRawData();
350 n = val.getNumWords();
351 return isl_val_int_from_chunks(ctx, n, sizeof(uint64_t), data);
354 /* Extract an integer from "expr".
355 * Return NULL if "expr" does not (obviously) represent an integer.
357 __isl_give isl_val *PetScan::extract_int(clang::ParenExpr *expr)
359 return extract_int(expr->getSubExpr());
362 /* Extract an integer from "expr".
363 * Return NULL if "expr" does not (obviously) represent an integer.
365 __isl_give isl_val *PetScan::extract_int(clang::Expr *expr)
367 if (expr->getStmtClass() == Stmt::IntegerLiteralClass)
368 return extract_int(ctx, cast<IntegerLiteral>(expr));
369 if (expr->getStmtClass() == Stmt::ParenExprClass)
370 return extract_int(cast<ParenExpr>(expr));
372 unsupported(expr);
373 return NULL;
376 /* Extract an affine expression from the IntegerLiteral "expr".
378 __isl_give isl_pw_aff *PetScan::extract_affine(IntegerLiteral *expr)
380 isl_space *dim = isl_space_params_alloc(ctx, 0);
381 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
382 isl_aff *aff = isl_aff_zero_on_domain(ls);
383 isl_set *dom = isl_set_universe(dim);
384 isl_val *v;
386 v = extract_int(expr);
387 aff = isl_aff_add_constant_val(aff, v);
389 return isl_pw_aff_alloc(dom, aff);
392 /* Extract an affine expression from the APInt "val", which is assumed
393 * to be non-negative.
395 __isl_give isl_pw_aff *PetScan::extract_affine(const llvm::APInt &val)
397 isl_space *dim = isl_space_params_alloc(ctx, 0);
398 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
399 isl_aff *aff = isl_aff_zero_on_domain(ls);
400 isl_set *dom = isl_set_universe(dim);
401 isl_val *v;
403 v = extract_unsigned(ctx, val);
404 aff = isl_aff_add_constant_val(aff, v);
406 return isl_pw_aff_alloc(dom, aff);
409 __isl_give isl_pw_aff *PetScan::extract_affine(ImplicitCastExpr *expr)
411 return extract_affine(expr->getSubExpr());
414 static unsigned get_type_size(ValueDecl *decl)
416 return decl->getASTContext().getIntWidth(decl->getType());
419 /* Bound parameter "pos" of "set" to the possible values of "decl".
421 static __isl_give isl_set *set_parameter_bounds(__isl_take isl_set *set,
422 unsigned pos, ValueDecl *decl)
424 unsigned width;
425 isl_ctx *ctx;
426 isl_val *bound;
428 ctx = isl_set_get_ctx(set);
429 width = get_type_size(decl);
430 if (decl->getType()->isUnsignedIntegerType()) {
431 set = isl_set_lower_bound_si(set, isl_dim_param, pos, 0);
432 bound = isl_val_int_from_ui(ctx, width);
433 bound = isl_val_2exp(bound);
434 bound = isl_val_sub_ui(bound, 1);
435 set = isl_set_upper_bound_val(set, isl_dim_param, pos, bound);
436 } else {
437 bound = isl_val_int_from_ui(ctx, width - 1);
438 bound = isl_val_2exp(bound);
439 bound = isl_val_sub_ui(bound, 1);
440 set = isl_set_upper_bound_val(set, isl_dim_param, pos,
441 isl_val_copy(bound));
442 bound = isl_val_neg(bound);
443 bound = isl_val_sub_ui(bound, 1);
444 set = isl_set_lower_bound_val(set, isl_dim_param, pos, bound);
447 return set;
450 /* Extract an affine expression from the DeclRefExpr "expr".
452 * If the variable has been assigned a value, then we check whether
453 * we know what (affine) value was assigned.
454 * If so, we return this value. Otherwise we convert "expr"
455 * to an extra parameter (provided nesting_enabled is set).
457 * Otherwise, we simply return an expression that is equal
458 * to a parameter corresponding to the referenced variable.
460 __isl_give isl_pw_aff *PetScan::extract_affine(DeclRefExpr *expr)
462 ValueDecl *decl = expr->getDecl();
463 const Type *type = decl->getType().getTypePtr();
464 isl_id *id;
465 isl_space *dim;
466 isl_aff *aff;
467 isl_set *dom;
469 if (!type->isIntegerType()) {
470 unsupported(expr);
471 return NULL;
474 if (assigned_value.find(decl) != assigned_value.end()) {
475 if (assigned_value[decl])
476 return isl_pw_aff_copy(assigned_value[decl]);
477 else
478 return nested_access(expr);
481 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
482 dim = isl_space_params_alloc(ctx, 1);
484 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
486 dom = isl_set_universe(isl_space_copy(dim));
487 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
488 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
490 return isl_pw_aff_alloc(dom, aff);
493 /* Extract an affine expression from an integer division operation.
494 * In particular, if "expr" is lhs/rhs, then return
496 * lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs)
498 * The second argument (rhs) is required to be a (positive) integer constant.
500 __isl_give isl_pw_aff *PetScan::extract_affine_div(BinaryOperator *expr)
502 int is_cst;
503 isl_pw_aff *rhs, *lhs;
505 rhs = extract_affine(expr->getRHS());
506 is_cst = isl_pw_aff_is_cst(rhs);
507 if (is_cst < 0 || !is_cst) {
508 isl_pw_aff_free(rhs);
509 if (!is_cst)
510 unsupported(expr);
511 return NULL;
514 lhs = extract_affine(expr->getLHS());
516 return isl_pw_aff_tdiv_q(lhs, rhs);
519 /* Extract an affine expression from a modulo operation.
520 * In particular, if "expr" is lhs/rhs, then return
522 * lhs - rhs * (lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs))
524 * The second argument (rhs) is required to be a (positive) integer constant.
526 __isl_give isl_pw_aff *PetScan::extract_affine_mod(BinaryOperator *expr)
528 int is_cst;
529 isl_pw_aff *rhs, *lhs;
531 rhs = extract_affine(expr->getRHS());
532 is_cst = isl_pw_aff_is_cst(rhs);
533 if (is_cst < 0 || !is_cst) {
534 isl_pw_aff_free(rhs);
535 if (!is_cst)
536 unsupported(expr);
537 return NULL;
540 lhs = extract_affine(expr->getLHS());
542 return isl_pw_aff_tdiv_r(lhs, rhs);
545 /* Extract an affine expression from a multiplication operation.
546 * This is only allowed if at least one of the two arguments
547 * is a (piecewise) constant.
549 __isl_give isl_pw_aff *PetScan::extract_affine_mul(BinaryOperator *expr)
551 isl_pw_aff *lhs;
552 isl_pw_aff *rhs;
554 lhs = extract_affine(expr->getLHS());
555 rhs = extract_affine(expr->getRHS());
557 if (!isl_pw_aff_is_cst(lhs) && !isl_pw_aff_is_cst(rhs)) {
558 isl_pw_aff_free(lhs);
559 isl_pw_aff_free(rhs);
560 unsupported(expr);
561 return NULL;
564 return isl_pw_aff_mul(lhs, rhs);
567 /* Extract an affine expression from an addition or subtraction operation.
569 __isl_give isl_pw_aff *PetScan::extract_affine_add(BinaryOperator *expr)
571 isl_pw_aff *lhs;
572 isl_pw_aff *rhs;
574 lhs = extract_affine(expr->getLHS());
575 rhs = extract_affine(expr->getRHS());
577 switch (expr->getOpcode()) {
578 case BO_Add:
579 return isl_pw_aff_add(lhs, rhs);
580 case BO_Sub:
581 return isl_pw_aff_sub(lhs, rhs);
582 default:
583 isl_pw_aff_free(lhs);
584 isl_pw_aff_free(rhs);
585 return NULL;
590 /* Compute
592 * pwaff mod 2^width
594 static __isl_give isl_pw_aff *wrap(__isl_take isl_pw_aff *pwaff,
595 unsigned width)
597 isl_ctx *ctx;
598 isl_val *mod;
600 ctx = isl_pw_aff_get_ctx(pwaff);
601 mod = isl_val_int_from_ui(ctx, width);
602 mod = isl_val_2exp(mod);
604 pwaff = isl_pw_aff_mod_val(pwaff, mod);
606 return pwaff;
609 /* Limit the domain of "pwaff" to those elements where the function
610 * value satisfies
612 * 2^{width-1} <= pwaff < 2^{width-1}
614 static __isl_give isl_pw_aff *avoid_overflow(__isl_take isl_pw_aff *pwaff,
615 unsigned width)
617 isl_ctx *ctx;
618 isl_val *v;
619 isl_space *space = isl_pw_aff_get_domain_space(pwaff);
620 isl_local_space *ls = isl_local_space_from_space(space);
621 isl_aff *bound;
622 isl_set *dom;
623 isl_pw_aff *b;
625 ctx = isl_pw_aff_get_ctx(pwaff);
626 v = isl_val_int_from_ui(ctx, width - 1);
627 v = isl_val_2exp(v);
629 bound = isl_aff_zero_on_domain(ls);
630 bound = isl_aff_add_constant_val(bound, v);
631 b = isl_pw_aff_from_aff(bound);
633 dom = isl_pw_aff_lt_set(isl_pw_aff_copy(pwaff), isl_pw_aff_copy(b));
634 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
636 b = isl_pw_aff_neg(b);
637 dom = isl_pw_aff_ge_set(isl_pw_aff_copy(pwaff), b);
638 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
640 return pwaff;
643 /* Handle potential overflows on signed computations.
645 * If options->signed_overflow is set to PET_OVERFLOW_AVOID,
646 * the we adjust the domain of "pa" to avoid overflows.
648 __isl_give isl_pw_aff *PetScan::signed_overflow(__isl_take isl_pw_aff *pa,
649 unsigned width)
651 if (options->signed_overflow == PET_OVERFLOW_AVOID)
652 pa = avoid_overflow(pa, width);
654 return pa;
657 /* Return the piecewise affine expression "set ? 1 : 0" defined on "dom".
659 static __isl_give isl_pw_aff *indicator_function(__isl_take isl_set *set,
660 __isl_take isl_set *dom)
662 isl_pw_aff *pa;
663 pa = isl_set_indicator_function(set);
664 pa = isl_pw_aff_intersect_domain(pa, isl_set_coalesce(dom));
665 return pa;
668 /* Extract an affine expression from some binary operations.
669 * If the result of the expression is unsigned, then we wrap it
670 * based on the size of the type. Otherwise, we ensure that
671 * no overflow occurs.
673 __isl_give isl_pw_aff *PetScan::extract_affine(BinaryOperator *expr)
675 isl_pw_aff *res;
676 unsigned width;
678 switch (expr->getOpcode()) {
679 case BO_Add:
680 case BO_Sub:
681 res = extract_affine_add(expr);
682 break;
683 case BO_Div:
684 res = extract_affine_div(expr);
685 break;
686 case BO_Rem:
687 res = extract_affine_mod(expr);
688 break;
689 case BO_Mul:
690 res = extract_affine_mul(expr);
691 break;
692 case BO_LT:
693 case BO_LE:
694 case BO_GT:
695 case BO_GE:
696 case BO_EQ:
697 case BO_NE:
698 case BO_LAnd:
699 case BO_LOr:
700 return extract_condition(expr);
701 default:
702 unsupported(expr);
703 return NULL;
706 width = ast_context.getIntWidth(expr->getType());
707 if (expr->getType()->isUnsignedIntegerType())
708 res = wrap(res, width);
709 else
710 res = signed_overflow(res, width);
712 return res;
715 /* Extract an affine expression from a negation operation.
717 __isl_give isl_pw_aff *PetScan::extract_affine(UnaryOperator *expr)
719 if (expr->getOpcode() == UO_Minus)
720 return isl_pw_aff_neg(extract_affine(expr->getSubExpr()));
721 if (expr->getOpcode() == UO_LNot)
722 return extract_condition(expr);
724 unsupported(expr);
725 return NULL;
728 __isl_give isl_pw_aff *PetScan::extract_affine(ParenExpr *expr)
730 return extract_affine(expr->getSubExpr());
733 /* Extract an affine expression from some special function calls.
734 * In particular, we handle "min", "max", "ceild", "floord",
735 * "intMod", "intFloor" and "intCeil".
736 * In case of the latter five, the second argument needs to be
737 * a (positive) integer constant.
739 __isl_give isl_pw_aff *PetScan::extract_affine(CallExpr *expr)
741 FunctionDecl *fd;
742 string name;
743 isl_pw_aff *aff1, *aff2;
745 fd = expr->getDirectCallee();
746 if (!fd) {
747 unsupported(expr);
748 return NULL;
751 name = fd->getDeclName().getAsString();
752 if (!(expr->getNumArgs() == 2 && name == "min") &&
753 !(expr->getNumArgs() == 2 && name == "max") &&
754 !(expr->getNumArgs() == 2 && name == "intMod") &&
755 !(expr->getNumArgs() == 2 && name == "intFloor") &&
756 !(expr->getNumArgs() == 2 && name == "intCeil") &&
757 !(expr->getNumArgs() == 2 && name == "floord") &&
758 !(expr->getNumArgs() == 2 && name == "ceild")) {
759 unsupported(expr);
760 return NULL;
763 if (name == "min" || name == "max") {
764 aff1 = extract_affine(expr->getArg(0));
765 aff2 = extract_affine(expr->getArg(1));
767 if (name == "min")
768 aff1 = isl_pw_aff_min(aff1, aff2);
769 else
770 aff1 = isl_pw_aff_max(aff1, aff2);
771 } else if (name == "intMod") {
772 isl_val *v;
773 Expr *arg2 = expr->getArg(1);
775 if (arg2->getStmtClass() != Stmt::IntegerLiteralClass) {
776 unsupported(expr);
777 return NULL;
779 aff1 = extract_affine(expr->getArg(0));
780 v = extract_int(cast<IntegerLiteral>(arg2));
781 aff1 = isl_pw_aff_mod_val(aff1, v);
782 } else if (name == "floord" || name == "ceild" ||
783 name == "intFloor" || name == "intCeil") {
784 isl_val *v;
785 Expr *arg2 = expr->getArg(1);
787 if (arg2->getStmtClass() != Stmt::IntegerLiteralClass) {
788 unsupported(expr);
789 return NULL;
791 aff1 = extract_affine(expr->getArg(0));
792 v = extract_int(cast<IntegerLiteral>(arg2));
793 aff1 = isl_pw_aff_scale_down_val(aff1, v);
794 if (name == "floord" || name == "intFloor")
795 aff1 = isl_pw_aff_floor(aff1);
796 else
797 aff1 = isl_pw_aff_ceil(aff1);
798 } else {
799 unsupported(expr);
800 return NULL;
803 return aff1;
806 /* This method is called when we come across an access that is
807 * nested in what is supposed to be an affine expression.
808 * If nesting is allowed, we return a new parameter that corresponds
809 * to this nested access. Otherwise, we simply complain.
811 * Note that we currently don't allow nested accesses themselves
812 * to contain any nested accesses, so we check if we can extract
813 * the access without any nesting and complain if we can't.
815 * The new parameter is resolved in resolve_nested.
817 isl_pw_aff *PetScan::nested_access(Expr *expr)
819 isl_id *id;
820 isl_space *dim;
821 isl_aff *aff;
822 isl_set *dom;
823 isl_multi_pw_aff *index;
825 if (!nesting_enabled) {
826 unsupported(expr);
827 return NULL;
830 allow_nested = false;
831 index = extract_index(expr);
832 allow_nested = true;
833 if (!index) {
834 unsupported(expr);
835 return NULL;
837 isl_multi_pw_aff_free(index);
839 id = isl_id_alloc(ctx, NULL, expr);
840 dim = isl_space_params_alloc(ctx, 1);
842 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
844 dom = isl_set_universe(isl_space_copy(dim));
845 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
846 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
848 return isl_pw_aff_alloc(dom, aff);
851 /* Affine expressions are not supposed to contain array accesses,
852 * but if nesting is allowed, we return a parameter corresponding
853 * to the array access.
855 __isl_give isl_pw_aff *PetScan::extract_affine(ArraySubscriptExpr *expr)
857 return nested_access(expr);
860 /* Affine expressions are not supposed to contain member accesses,
861 * but if nesting is allowed, we return a parameter corresponding
862 * to the member access.
864 __isl_give isl_pw_aff *PetScan::extract_affine(MemberExpr *expr)
866 return nested_access(expr);
869 /* Extract an affine expression from a conditional operation.
871 __isl_give isl_pw_aff *PetScan::extract_affine(ConditionalOperator *expr)
873 isl_pw_aff *cond, *lhs, *rhs;
875 cond = extract_condition(expr->getCond());
876 lhs = extract_affine(expr->getTrueExpr());
877 rhs = extract_affine(expr->getFalseExpr());
879 return isl_pw_aff_cond(cond, lhs, rhs);
882 /* Extract an affine expression, if possible, from "expr".
883 * Otherwise return NULL.
885 __isl_give isl_pw_aff *PetScan::extract_affine(Expr *expr)
887 switch (expr->getStmtClass()) {
888 case Stmt::ImplicitCastExprClass:
889 return extract_affine(cast<ImplicitCastExpr>(expr));
890 case Stmt::IntegerLiteralClass:
891 return extract_affine(cast<IntegerLiteral>(expr));
892 case Stmt::DeclRefExprClass:
893 return extract_affine(cast<DeclRefExpr>(expr));
894 case Stmt::BinaryOperatorClass:
895 return extract_affine(cast<BinaryOperator>(expr));
896 case Stmt::UnaryOperatorClass:
897 return extract_affine(cast<UnaryOperator>(expr));
898 case Stmt::ParenExprClass:
899 return extract_affine(cast<ParenExpr>(expr));
900 case Stmt::CallExprClass:
901 return extract_affine(cast<CallExpr>(expr));
902 case Stmt::ArraySubscriptExprClass:
903 return extract_affine(cast<ArraySubscriptExpr>(expr));
904 case Stmt::MemberExprClass:
905 return extract_affine(cast<MemberExpr>(expr));
906 case Stmt::ConditionalOperatorClass:
907 return extract_affine(cast<ConditionalOperator>(expr));
908 default:
909 unsupported(expr);
911 return NULL;
914 __isl_give isl_multi_pw_aff *PetScan::extract_index(ImplicitCastExpr *expr)
916 return extract_index(expr->getSubExpr());
919 /* Return the depth of an array of the given type.
921 static int array_depth(const Type *type)
923 if (type->isPointerType())
924 return 1 + array_depth(type->getPointeeType().getTypePtr());
925 if (type->isArrayType()) {
926 const ArrayType *atype;
927 type = type->getCanonicalTypeInternal().getTypePtr();
928 atype = cast<ArrayType>(type);
929 return 1 + array_depth(atype->getElementType().getTypePtr());
931 return 0;
934 /* Return the depth of the array accessed by the index expression "index".
935 * If "index" is an affine expression, i.e., if it does not access
936 * any array, then return 1.
937 * If "index" represent a member access, i.e., if its range is a wrapped
938 * relation, then return the sum of the depth of the array of structures
939 * and that of the member inside the structure.
941 static int extract_depth(__isl_keep isl_multi_pw_aff *index)
943 isl_id *id;
944 ValueDecl *decl;
946 if (!index)
947 return -1;
949 if (isl_multi_pw_aff_range_is_wrapping(index)) {
950 int domain_depth, range_depth;
951 isl_multi_pw_aff *domain, *range;
953 domain = isl_multi_pw_aff_copy(index);
954 domain = isl_multi_pw_aff_range_factor_domain(domain);
955 domain_depth = extract_depth(domain);
956 isl_multi_pw_aff_free(domain);
957 range = isl_multi_pw_aff_copy(index);
958 range = isl_multi_pw_aff_range_factor_range(range);
959 range_depth = extract_depth(range);
960 isl_multi_pw_aff_free(range);
962 return domain_depth + range_depth;
965 if (!isl_multi_pw_aff_has_tuple_id(index, isl_dim_out))
966 return 1;
968 id = isl_multi_pw_aff_get_tuple_id(index, isl_dim_out);
969 if (!id)
970 return -1;
971 decl = (ValueDecl *) isl_id_get_user(id);
972 isl_id_free(id);
974 return array_depth(decl->getType().getTypePtr());
977 /* Extract an index expression from a reference to a variable.
978 * If the variable has name "A", then the returned index expression
979 * is of the form
981 * { [] -> A[] }
983 __isl_give isl_multi_pw_aff *PetScan::extract_index(DeclRefExpr *expr)
985 return extract_index(expr->getDecl());
988 /* Extract an index expression from a variable.
989 * If the variable has name "A", then the returned index expression
990 * is of the form
992 * { [] -> A[] }
994 __isl_give isl_multi_pw_aff *PetScan::extract_index(ValueDecl *decl)
996 isl_id *id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
997 isl_space *space = isl_space_alloc(ctx, 0, 0, 0);
999 space = isl_space_set_tuple_id(space, isl_dim_out, id);
1001 return isl_multi_pw_aff_zero(space);
1004 /* Extract an index expression from an integer contant.
1005 * If the value of the constant is "v", then the returned access relation
1006 * is
1008 * { [] -> [v] }
1010 __isl_give isl_multi_pw_aff *PetScan::extract_index(IntegerLiteral *expr)
1012 isl_multi_pw_aff *mpa;
1014 mpa = isl_multi_pw_aff_from_pw_aff(extract_affine(expr));
1015 mpa = isl_multi_pw_aff_from_range(mpa);
1016 return mpa;
1019 /* Try and extract an index expression from the given Expr.
1020 * Return NULL if it doesn't work out.
1022 __isl_give isl_multi_pw_aff *PetScan::extract_index(Expr *expr)
1024 switch (expr->getStmtClass()) {
1025 case Stmt::ImplicitCastExprClass:
1026 return extract_index(cast<ImplicitCastExpr>(expr));
1027 case Stmt::DeclRefExprClass:
1028 return extract_index(cast<DeclRefExpr>(expr));
1029 case Stmt::ArraySubscriptExprClass:
1030 return extract_index(cast<ArraySubscriptExpr>(expr));
1031 case Stmt::IntegerLiteralClass:
1032 return extract_index(cast<IntegerLiteral>(expr));
1033 case Stmt::MemberExprClass:
1034 return extract_index(cast<MemberExpr>(expr));
1035 default:
1036 unsupported(expr);
1038 return NULL;
1041 /* Given a partial index expression "base" and an extra index "index",
1042 * append the extra index to "base" and return the result.
1043 * Additionally, add the constraints that the extra index is non-negative.
1044 * If "index" represent a member access, i.e., if its range is a wrapped
1045 * relation, then we recursively extend the range of this nested relation.
1047 static __isl_give isl_multi_pw_aff *subscript(__isl_take isl_multi_pw_aff *base,
1048 __isl_take isl_pw_aff *index)
1050 isl_id *id;
1051 isl_set *domain;
1052 isl_multi_pw_aff *access;
1053 int member_access;
1055 member_access = isl_multi_pw_aff_range_is_wrapping(base);
1056 if (member_access < 0)
1057 goto error;
1058 if (member_access) {
1059 isl_multi_pw_aff *domain, *range;
1060 isl_id *id;
1062 id = isl_multi_pw_aff_get_tuple_id(base, isl_dim_out);
1063 domain = isl_multi_pw_aff_copy(base);
1064 domain = isl_multi_pw_aff_range_factor_domain(domain);
1065 range = isl_multi_pw_aff_range_factor_range(base);
1066 range = subscript(range, index);
1067 access = isl_multi_pw_aff_range_product(domain, range);
1068 access = isl_multi_pw_aff_set_tuple_id(access, isl_dim_out, id);
1069 return access;
1072 id = isl_multi_pw_aff_get_tuple_id(base, isl_dim_set);
1073 index = isl_pw_aff_from_range(index);
1074 domain = isl_pw_aff_nonneg_set(isl_pw_aff_copy(index));
1075 index = isl_pw_aff_intersect_domain(index, domain);
1076 access = isl_multi_pw_aff_from_pw_aff(index);
1077 access = isl_multi_pw_aff_flat_range_product(base, access);
1078 access = isl_multi_pw_aff_set_tuple_id(access, isl_dim_set, id);
1080 return access;
1081 error:
1082 isl_multi_pw_aff_free(base);
1083 isl_pw_aff_free(index);
1084 return NULL;
1087 /* Extract an index expression from the given array subscript expression.
1088 * If nesting is allowed in general, then we turn it on while
1089 * examining the index expression.
1091 * We first extract an index expression from the base.
1092 * This will result in an index expression with a range that corresponds
1093 * to the earlier indices.
1094 * We then extract the current index, restrict its domain
1095 * to those values that result in a non-negative index and
1096 * append the index to the base index expression.
1098 __isl_give isl_multi_pw_aff *PetScan::extract_index(ArraySubscriptExpr *expr)
1100 Expr *base = expr->getBase();
1101 Expr *idx = expr->getIdx();
1102 isl_pw_aff *index;
1103 isl_multi_pw_aff *base_access;
1104 isl_multi_pw_aff *access;
1105 bool save_nesting = nesting_enabled;
1107 nesting_enabled = allow_nested;
1109 base_access = extract_index(base);
1110 index = extract_affine(idx);
1112 nesting_enabled = save_nesting;
1114 access = subscript(base_access, index);
1116 return access;
1119 /* Construct a name for a member access by concatenating the name
1120 * of the array of structures and the member, separated by an underscore.
1122 * The caller is responsible for freeing the result.
1124 static char *member_access_name(isl_ctx *ctx, const char *base,
1125 const char *field)
1127 int len;
1128 char *name;
1130 len = strlen(base) + 1 + strlen(field);
1131 name = isl_alloc_array(ctx, char, len + 1);
1132 if (!name)
1133 return NULL;
1134 snprintf(name, len + 1, "%s_%s", base, field);
1136 return name;
1139 /* Given an index expression "base" for an element of an array of structures
1140 * and an expression "field" for the field member being accessed, construct
1141 * an index expression for an access to that member of the given structure.
1142 * In particular, take the range product of "base" and "field" and
1143 * attach a name to the result.
1145 static __isl_give isl_multi_pw_aff *member(__isl_take isl_multi_pw_aff *base,
1146 __isl_take isl_multi_pw_aff *field)
1148 isl_ctx *ctx;
1149 isl_multi_pw_aff *access;
1150 const char *base_name, *field_name;
1151 char *name;
1153 ctx = isl_multi_pw_aff_get_ctx(base);
1155 base_name = isl_multi_pw_aff_get_tuple_name(base, isl_dim_out);
1156 field_name = isl_multi_pw_aff_get_tuple_name(field, isl_dim_out);
1157 name = member_access_name(ctx, base_name, field_name);
1159 access = isl_multi_pw_aff_range_product(base, field);
1161 access = isl_multi_pw_aff_set_tuple_name(access, isl_dim_out, name);
1162 free(name);
1164 return access;
1167 /* Extract an index expression from a member expression.
1169 * If the base access (to the structure containing the member)
1170 * is of the form
1172 * [] -> A[..]
1174 * and the member is called "f", then the member access is of
1175 * the form
1177 * [] -> A_f[A[..] -> f[]]
1179 * If the member access is to an anonymous struct, then simply return
1181 * [] -> A[..]
1183 * If the member access in the source code is of the form
1185 * A->f
1187 * then it is treated as
1189 * A[0].f
1191 __isl_give isl_multi_pw_aff *PetScan::extract_index(MemberExpr *expr)
1193 Expr *base = expr->getBase();
1194 FieldDecl *field = cast<FieldDecl>(expr->getMemberDecl());
1195 isl_multi_pw_aff *base_access, *field_access;
1196 isl_id *id;
1197 isl_space *space;
1199 base_access = extract_index(base);
1201 if (expr->isArrow()) {
1202 isl_space *space = isl_space_params_alloc(ctx, 0);
1203 isl_local_space *ls = isl_local_space_from_space(space);
1204 isl_aff *aff = isl_aff_zero_on_domain(ls);
1205 isl_pw_aff *index = isl_pw_aff_from_aff(aff);
1206 base_access = subscript(base_access, index);
1209 if (field->isAnonymousStructOrUnion())
1210 return base_access;
1212 id = isl_id_alloc(ctx, field->getName().str().c_str(), field);
1213 space = isl_multi_pw_aff_get_domain_space(base_access);
1214 space = isl_space_from_domain(space);
1215 space = isl_space_set_tuple_id(space, isl_dim_out, id);
1216 field_access = isl_multi_pw_aff_zero(space);
1218 return member(base_access, field_access);
1221 /* Check if "expr" calls function "minmax" with two arguments and if so
1222 * make lhs and rhs refer to these two arguments.
1224 static bool is_minmax(Expr *expr, const char *minmax, Expr *&lhs, Expr *&rhs)
1226 CallExpr *call;
1227 FunctionDecl *fd;
1228 string name;
1230 if (expr->getStmtClass() != Stmt::CallExprClass)
1231 return false;
1233 call = cast<CallExpr>(expr);
1234 fd = call->getDirectCallee();
1235 if (!fd)
1236 return false;
1238 if (call->getNumArgs() != 2)
1239 return false;
1241 name = fd->getDeclName().getAsString();
1242 if (name != minmax)
1243 return false;
1245 lhs = call->getArg(0);
1246 rhs = call->getArg(1);
1248 return true;
1251 /* Check if "expr" is of the form min(lhs, rhs) and if so make
1252 * lhs and rhs refer to the two arguments.
1254 static bool is_min(Expr *expr, Expr *&lhs, Expr *&rhs)
1256 return is_minmax(expr, "min", lhs, rhs);
1259 /* Check if "expr" is of the form max(lhs, rhs) and if so make
1260 * lhs and rhs refer to the two arguments.
1262 static bool is_max(Expr *expr, Expr *&lhs, Expr *&rhs)
1264 return is_minmax(expr, "max", lhs, rhs);
1267 /* Return "lhs && rhs", defined on the shared definition domain.
1269 static __isl_give isl_pw_aff *pw_aff_and(__isl_take isl_pw_aff *lhs,
1270 __isl_take isl_pw_aff *rhs)
1272 isl_set *cond;
1273 isl_set *dom;
1275 dom = isl_set_intersect(isl_pw_aff_domain(isl_pw_aff_copy(lhs)),
1276 isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1277 cond = isl_set_intersect(isl_pw_aff_non_zero_set(lhs),
1278 isl_pw_aff_non_zero_set(rhs));
1279 return indicator_function(cond, dom);
1282 /* Return "lhs && rhs", with shortcut semantics.
1283 * That is, if lhs is false, then the result is defined even if rhs is not.
1284 * In practice, we compute lhs ? rhs : lhs.
1286 static __isl_give isl_pw_aff *pw_aff_and_then(__isl_take isl_pw_aff *lhs,
1287 __isl_take isl_pw_aff *rhs)
1289 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), rhs, lhs);
1292 /* Return "lhs || rhs", with shortcut semantics.
1293 * That is, if lhs is true, then the result is defined even if rhs is not.
1294 * In practice, we compute lhs ? lhs : rhs.
1296 static __isl_give isl_pw_aff *pw_aff_or_else(__isl_take isl_pw_aff *lhs,
1297 __isl_take isl_pw_aff *rhs)
1299 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), lhs, rhs);
1302 /* Extract an affine expressions representing the comparison "LHS op RHS"
1303 * "comp" is the original statement that "LHS op RHS" is derived from
1304 * and is used for diagnostics.
1306 * If the comparison is of the form
1308 * a <= min(b,c)
1310 * then the expression is constructed as the conjunction of
1311 * the comparisons
1313 * a <= b and a <= c
1315 * A similar optimization is performed for max(a,b) <= c.
1316 * We do this because that will lead to simpler representations
1317 * of the expression.
1318 * If isl is ever enhanced to explicitly deal with min and max expressions,
1319 * this optimization can be removed.
1321 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperatorKind op,
1322 Expr *LHS, Expr *RHS, Stmt *comp)
1324 isl_pw_aff *lhs;
1325 isl_pw_aff *rhs;
1326 isl_pw_aff *res;
1327 isl_set *cond;
1328 isl_set *dom;
1330 if (op == BO_GT)
1331 return extract_comparison(BO_LT, RHS, LHS, comp);
1332 if (op == BO_GE)
1333 return extract_comparison(BO_LE, RHS, LHS, comp);
1335 if (op == BO_LT || op == BO_LE) {
1336 Expr *expr1, *expr2;
1337 if (is_min(RHS, expr1, expr2)) {
1338 lhs = extract_comparison(op, LHS, expr1, comp);
1339 rhs = extract_comparison(op, LHS, expr2, comp);
1340 return pw_aff_and(lhs, rhs);
1342 if (is_max(LHS, expr1, expr2)) {
1343 lhs = extract_comparison(op, expr1, RHS, comp);
1344 rhs = extract_comparison(op, expr2, RHS, comp);
1345 return pw_aff_and(lhs, rhs);
1349 lhs = extract_affine(LHS);
1350 rhs = extract_affine(RHS);
1352 dom = isl_pw_aff_domain(isl_pw_aff_copy(lhs));
1353 dom = isl_set_intersect(dom, isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1355 switch (op) {
1356 case BO_LT:
1357 cond = isl_pw_aff_lt_set(lhs, rhs);
1358 break;
1359 case BO_LE:
1360 cond = isl_pw_aff_le_set(lhs, rhs);
1361 break;
1362 case BO_EQ:
1363 cond = isl_pw_aff_eq_set(lhs, rhs);
1364 break;
1365 case BO_NE:
1366 cond = isl_pw_aff_ne_set(lhs, rhs);
1367 break;
1368 default:
1369 isl_pw_aff_free(lhs);
1370 isl_pw_aff_free(rhs);
1371 isl_set_free(dom);
1372 unsupported(comp);
1373 return NULL;
1376 cond = isl_set_coalesce(cond);
1377 res = indicator_function(cond, dom);
1379 return res;
1382 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperator *comp)
1384 return extract_comparison(comp->getOpcode(), comp->getLHS(),
1385 comp->getRHS(), comp);
1388 /* Extract an affine expression representing the negation (logical not)
1389 * of a subexpression.
1391 __isl_give isl_pw_aff *PetScan::extract_boolean(UnaryOperator *op)
1393 isl_set *set_cond, *dom;
1394 isl_pw_aff *cond, *res;
1396 cond = extract_condition(op->getSubExpr());
1398 dom = isl_pw_aff_domain(isl_pw_aff_copy(cond));
1400 set_cond = isl_pw_aff_zero_set(cond);
1402 res = indicator_function(set_cond, dom);
1404 return res;
1407 /* Extract an affine expression representing the disjunction (logical or)
1408 * or conjunction (logical and) of two subexpressions.
1410 __isl_give isl_pw_aff *PetScan::extract_boolean(BinaryOperator *comp)
1412 isl_pw_aff *lhs, *rhs;
1414 lhs = extract_condition(comp->getLHS());
1415 rhs = extract_condition(comp->getRHS());
1417 switch (comp->getOpcode()) {
1418 case BO_LAnd:
1419 return pw_aff_and_then(lhs, rhs);
1420 case BO_LOr:
1421 return pw_aff_or_else(lhs, rhs);
1422 default:
1423 isl_pw_aff_free(lhs);
1424 isl_pw_aff_free(rhs);
1427 unsupported(comp);
1428 return NULL;
1431 __isl_give isl_pw_aff *PetScan::extract_condition(UnaryOperator *expr)
1433 switch (expr->getOpcode()) {
1434 case UO_LNot:
1435 return extract_boolean(expr);
1436 default:
1437 unsupported(expr);
1438 return NULL;
1442 /* Extract the affine expression "expr != 0 ? 1 : 0".
1444 __isl_give isl_pw_aff *PetScan::extract_implicit_condition(Expr *expr)
1446 isl_pw_aff *res;
1447 isl_set *set, *dom;
1449 res = extract_affine(expr);
1451 dom = isl_pw_aff_domain(isl_pw_aff_copy(res));
1452 set = isl_pw_aff_non_zero_set(res);
1454 res = indicator_function(set, dom);
1456 return res;
1459 /* Extract an affine expression from a boolean expression.
1460 * In particular, return the expression "expr ? 1 : 0".
1462 * If the expression doesn't look like a condition, we assume it
1463 * is an affine expression and return the condition "expr != 0 ? 1 : 0".
1465 __isl_give isl_pw_aff *PetScan::extract_condition(Expr *expr)
1467 BinaryOperator *comp;
1469 if (!expr) {
1470 isl_set *u = isl_set_universe(isl_space_params_alloc(ctx, 0));
1471 return indicator_function(u, isl_set_copy(u));
1474 if (expr->getStmtClass() == Stmt::ParenExprClass)
1475 return extract_condition(cast<ParenExpr>(expr)->getSubExpr());
1477 if (expr->getStmtClass() == Stmt::UnaryOperatorClass)
1478 return extract_condition(cast<UnaryOperator>(expr));
1480 if (expr->getStmtClass() != Stmt::BinaryOperatorClass)
1481 return extract_implicit_condition(expr);
1483 comp = cast<BinaryOperator>(expr);
1484 switch (comp->getOpcode()) {
1485 case BO_LT:
1486 case BO_LE:
1487 case BO_GT:
1488 case BO_GE:
1489 case BO_EQ:
1490 case BO_NE:
1491 return extract_comparison(comp);
1492 case BO_LAnd:
1493 case BO_LOr:
1494 return extract_boolean(comp);
1495 default:
1496 return extract_implicit_condition(expr);
1500 static enum pet_op_type UnaryOperatorKind2pet_op_type(UnaryOperatorKind kind)
1502 switch (kind) {
1503 case UO_Minus:
1504 return pet_op_minus;
1505 case UO_Not:
1506 return pet_op_not;
1507 case UO_LNot:
1508 return pet_op_lnot;
1509 case UO_PostInc:
1510 return pet_op_post_inc;
1511 case UO_PostDec:
1512 return pet_op_post_dec;
1513 case UO_PreInc:
1514 return pet_op_pre_inc;
1515 case UO_PreDec:
1516 return pet_op_pre_dec;
1517 default:
1518 return pet_op_last;
1522 static enum pet_op_type BinaryOperatorKind2pet_op_type(BinaryOperatorKind kind)
1524 switch (kind) {
1525 case BO_AddAssign:
1526 return pet_op_add_assign;
1527 case BO_SubAssign:
1528 return pet_op_sub_assign;
1529 case BO_MulAssign:
1530 return pet_op_mul_assign;
1531 case BO_DivAssign:
1532 return pet_op_div_assign;
1533 case BO_Assign:
1534 return pet_op_assign;
1535 case BO_Add:
1536 return pet_op_add;
1537 case BO_Sub:
1538 return pet_op_sub;
1539 case BO_Mul:
1540 return pet_op_mul;
1541 case BO_Div:
1542 return pet_op_div;
1543 case BO_Rem:
1544 return pet_op_mod;
1545 case BO_Shl:
1546 return pet_op_shl;
1547 case BO_Shr:
1548 return pet_op_shr;
1549 case BO_EQ:
1550 return pet_op_eq;
1551 case BO_NE:
1552 return pet_op_ne;
1553 case BO_LE:
1554 return pet_op_le;
1555 case BO_GE:
1556 return pet_op_ge;
1557 case BO_LT:
1558 return pet_op_lt;
1559 case BO_GT:
1560 return pet_op_gt;
1561 case BO_And:
1562 return pet_op_and;
1563 case BO_Xor:
1564 return pet_op_xor;
1565 case BO_Or:
1566 return pet_op_or;
1567 case BO_LAnd:
1568 return pet_op_land;
1569 case BO_LOr:
1570 return pet_op_lor;
1571 default:
1572 return pet_op_last;
1576 /* Construct a pet_expr representing a unary operator expression.
1578 struct pet_expr *PetScan::extract_expr(UnaryOperator *expr)
1580 struct pet_expr *arg;
1581 enum pet_op_type op;
1583 op = UnaryOperatorKind2pet_op_type(expr->getOpcode());
1584 if (op == pet_op_last) {
1585 unsupported(expr);
1586 return NULL;
1589 arg = extract_expr(expr->getSubExpr());
1591 if (expr->isIncrementDecrementOp() &&
1592 arg && arg->type == pet_expr_access) {
1593 mark_write(arg);
1594 arg->acc.read = 1;
1597 return pet_expr_new_unary(ctx, op, arg);
1600 /* Mark the given access pet_expr as a write.
1601 * If a scalar is being accessed, then mark its value
1602 * as unknown in assigned_value.
1604 void PetScan::mark_write(struct pet_expr *access)
1606 isl_id *id;
1607 ValueDecl *decl;
1609 if (!access)
1610 return;
1612 access->acc.write = 1;
1613 access->acc.read = 0;
1615 if (!pet_expr_is_scalar_access(access))
1616 return;
1618 id = pet_expr_access_get_id(access);
1619 decl = (ValueDecl *) isl_id_get_user(id);
1620 clear_assignment(assigned_value, decl);
1621 isl_id_free(id);
1624 /* Assign "rhs" to "lhs".
1626 * In particular, if "lhs" is a scalar variable, then mark
1627 * the variable as having been assigned. If, furthermore, "rhs"
1628 * is an affine expression, then keep track of this value in assigned_value
1629 * so that we can plug it in when we later come across the same variable.
1631 void PetScan::assign(struct pet_expr *lhs, Expr *rhs)
1633 isl_id *id;
1634 ValueDecl *decl;
1635 isl_pw_aff *pa;
1637 if (!lhs)
1638 return;
1639 if (!pet_expr_is_scalar_access(lhs))
1640 return;
1642 id = pet_expr_access_get_id(lhs);
1643 decl = (ValueDecl *) isl_id_get_user(id);
1644 isl_id_free(id);
1646 pa = try_extract_affine(rhs);
1647 clear_assignment(assigned_value, decl);
1648 if (!pa)
1649 return;
1650 assigned_value[decl] = pa;
1651 insert_expression(pa);
1654 /* Construct a pet_expr representing a binary operator expression.
1656 * If the top level operator is an assignment and the LHS is an access,
1657 * then we mark that access as a write. If the operator is a compound
1658 * assignment, the access is marked as both a read and a write.
1660 * If "expr" assigns something to a scalar variable, then we mark
1661 * the variable as having been assigned. If, furthermore, the expression
1662 * is affine, then keep track of this value in assigned_value
1663 * so that we can plug it in when we later come across the same variable.
1665 struct pet_expr *PetScan::extract_expr(BinaryOperator *expr)
1667 struct pet_expr *lhs, *rhs;
1668 enum pet_op_type op;
1670 op = BinaryOperatorKind2pet_op_type(expr->getOpcode());
1671 if (op == pet_op_last) {
1672 unsupported(expr);
1673 return NULL;
1676 lhs = extract_expr(expr->getLHS());
1677 rhs = extract_expr(expr->getRHS());
1679 if (expr->isAssignmentOp() && lhs && lhs->type == pet_expr_access) {
1680 mark_write(lhs);
1681 if (expr->isCompoundAssignmentOp())
1682 lhs->acc.read = 1;
1685 if (expr->getOpcode() == BO_Assign)
1686 assign(lhs, expr->getRHS());
1688 return pet_expr_new_binary(ctx, op, lhs, rhs);
1691 /* Construct a pet_scop with a single statement killing the entire
1692 * array "array".
1694 struct pet_scop *PetScan::kill(Stmt *stmt, struct pet_array *array)
1696 isl_id *id;
1697 isl_space *space;
1698 isl_multi_pw_aff *index;
1699 isl_map *access;
1700 struct pet_expr *expr;
1702 if (!array)
1703 return NULL;
1704 access = isl_map_from_range(isl_set_copy(array->extent));
1705 id = isl_set_get_tuple_id(array->extent);
1706 space = isl_space_alloc(ctx, 0, 0, 0);
1707 space = isl_space_set_tuple_id(space, isl_dim_out, id);
1708 index = isl_multi_pw_aff_zero(space);
1709 expr = pet_expr_kill_from_access_and_index(access, index);
1710 return extract(stmt, expr);
1713 /* Construct a pet_scop for a (single) variable declaration.
1715 * The scop contains the variable being declared (as an array)
1716 * and a statement killing the array.
1718 * If the variable is initialized in the AST, then the scop
1719 * also contains an assignment to the variable.
1721 struct pet_scop *PetScan::extract(DeclStmt *stmt)
1723 Decl *decl;
1724 VarDecl *vd;
1725 struct pet_expr *lhs, *rhs, *pe;
1726 struct pet_scop *scop_decl, *scop;
1727 struct pet_array *array;
1729 if (!stmt->isSingleDecl()) {
1730 unsupported(stmt);
1731 return NULL;
1734 decl = stmt->getSingleDecl();
1735 vd = cast<VarDecl>(decl);
1737 array = extract_array(ctx, vd, NULL);
1738 if (array)
1739 array->declared = 1;
1740 scop_decl = kill(stmt, array);
1741 scop_decl = pet_scop_add_array(scop_decl, array);
1743 if (!vd->getInit())
1744 return scop_decl;
1746 lhs = extract_access_expr(vd);
1747 rhs = extract_expr(vd->getInit());
1749 mark_write(lhs);
1750 assign(lhs, vd->getInit());
1752 pe = pet_expr_new_binary(ctx, pet_op_assign, lhs, rhs);
1753 scop = extract(stmt, pe);
1755 scop_decl = pet_scop_prefix(scop_decl, 0);
1756 scop = pet_scop_prefix(scop, 1);
1758 scop = pet_scop_add_seq(ctx, scop_decl, scop);
1760 return scop;
1763 /* Construct a pet_expr representing a conditional operation.
1765 * We first try to extract the condition as an affine expression.
1766 * If that fails, we construct a pet_expr tree representing the condition.
1768 struct pet_expr *PetScan::extract_expr(ConditionalOperator *expr)
1770 struct pet_expr *cond, *lhs, *rhs;
1771 isl_pw_aff *pa;
1773 pa = try_extract_affine(expr->getCond());
1774 if (pa) {
1775 isl_multi_pw_aff *test = isl_multi_pw_aff_from_pw_aff(pa);
1776 test = isl_multi_pw_aff_from_range(test);
1777 cond = pet_expr_from_index(test);
1778 } else
1779 cond = extract_expr(expr->getCond());
1780 lhs = extract_expr(expr->getTrueExpr());
1781 rhs = extract_expr(expr->getFalseExpr());
1783 return pet_expr_new_ternary(ctx, cond, lhs, rhs);
1786 struct pet_expr *PetScan::extract_expr(ImplicitCastExpr *expr)
1788 return extract_expr(expr->getSubExpr());
1791 /* Construct a pet_expr representing a floating point value.
1793 * If the floating point literal does not appear in a macro,
1794 * then we use the original representation in the source code
1795 * as the string representation. Otherwise, we use the pretty
1796 * printer to produce a string representation.
1798 struct pet_expr *PetScan::extract_expr(FloatingLiteral *expr)
1800 double d;
1801 string s;
1802 const LangOptions &LO = PP.getLangOpts();
1803 SourceLocation loc = expr->getLocation();
1805 if (!loc.isMacroID()) {
1806 SourceManager &SM = PP.getSourceManager();
1807 unsigned len = Lexer::MeasureTokenLength(loc, SM, LO);
1808 s = string(SM.getCharacterData(loc), len);
1809 } else {
1810 llvm::raw_string_ostream S(s);
1811 expr->printPretty(S, 0, PrintingPolicy(LO));
1812 S.str();
1814 d = expr->getValueAsApproximateDouble();
1815 return pet_expr_new_double(ctx, d, s.c_str());
1818 /* Extract an index expression from "expr" and then convert it into
1819 * an access pet_expr.
1821 struct pet_expr *PetScan::extract_access_expr(Expr *expr)
1823 isl_multi_pw_aff *index;
1824 struct pet_expr *pe;
1825 int depth;
1827 index = extract_index(expr);
1828 depth = extract_depth(index);
1830 pe = pet_expr_from_index_and_depth(index, depth);
1832 return pe;
1835 /* Extract an index expression from "decl" and then convert it into
1836 * an access pet_expr.
1838 struct pet_expr *PetScan::extract_access_expr(ValueDecl *decl)
1840 isl_multi_pw_aff *index;
1841 struct pet_expr *pe;
1842 int depth;
1844 index = extract_index(decl);
1845 depth = extract_depth(index);
1847 pe = pet_expr_from_index_and_depth(index, depth);
1849 return pe;
1852 struct pet_expr *PetScan::extract_expr(ParenExpr *expr)
1854 return extract_expr(expr->getSubExpr());
1857 /* Extract an assume statement from the argument "expr"
1858 * of a __pencil_assume statement.
1860 struct pet_expr *PetScan::extract_assume(Expr *expr)
1862 isl_pw_aff *cond;
1863 struct pet_expr *res;
1865 cond = try_extract_affine_condition(expr);
1866 if (!cond) {
1867 res = extract_expr(expr);
1868 } else {
1869 isl_multi_pw_aff *index;
1870 index = isl_multi_pw_aff_from_pw_aff(cond);
1871 index = isl_multi_pw_aff_from_range(index);
1872 res = pet_expr_from_index(index);
1874 return pet_expr_new_unary(ctx, pet_op_assume, res);
1877 /* Construct a pet_expr corresponding to the function call argument "expr".
1878 * The argument appears in position "pos" of a call to function "fd".
1880 * If we are passing along a pointer to an array element
1881 * or an entire row or even higher dimensional slice of an array,
1882 * then the function being called may write into the array.
1884 * We assume here that if the function is declared to take a pointer
1885 * to a const type, then the function will perform a read
1886 * and that otherwise, it will perform a write.
1888 struct pet_expr *PetScan::extract_argument(FunctionDecl *fd, int pos,
1889 Expr *expr)
1891 struct pet_expr *res;
1892 int is_addr = 0;
1893 pet_expr *main_arg;
1894 Stmt::StmtClass sc;
1896 if (expr->getStmtClass() == Stmt::ImplicitCastExprClass) {
1897 ImplicitCastExpr *ice = cast<ImplicitCastExpr>(expr);
1898 expr = ice->getSubExpr();
1900 if (expr->getStmtClass() == Stmt::UnaryOperatorClass) {
1901 UnaryOperator *op = cast<UnaryOperator>(expr);
1902 if (op->getOpcode() == UO_AddrOf) {
1903 is_addr = 1;
1904 expr = op->getSubExpr();
1907 res = extract_expr(expr);
1908 main_arg = res;
1909 if (is_addr)
1910 res = pet_expr_new_unary(ctx, pet_op_address_of, res);
1911 if (!res)
1912 return NULL;
1913 sc = expr->getStmtClass();
1914 if ((sc == Stmt::ArraySubscriptExprClass ||
1915 sc == Stmt::MemberExprClass) &&
1916 array_depth(expr->getType().getTypePtr()) > 0)
1917 is_addr = 1;
1918 if (is_addr && main_arg->type == pet_expr_access) {
1919 ParmVarDecl *parm;
1920 if (!fd->hasPrototype()) {
1921 report_prototype_required(expr);
1922 return pet_expr_free(res);
1924 parm = fd->getParamDecl(pos);
1925 if (!const_base(parm->getType()))
1926 mark_write(main_arg);
1929 return res;
1932 /* Construct a pet_expr representing a function call.
1934 * In the special case of a "call" to __pencil_assume,
1935 * construct an assume expression instead.
1937 struct pet_expr *PetScan::extract_expr(CallExpr *expr)
1939 struct pet_expr *res = NULL;
1940 FunctionDecl *fd;
1941 string name;
1942 unsigned n_arg;
1944 fd = expr->getDirectCallee();
1945 if (!fd) {
1946 unsupported(expr);
1947 return NULL;
1950 name = fd->getDeclName().getAsString();
1951 n_arg = expr->getNumArgs();
1953 if (n_arg == 1 && name == "__pencil_assume")
1954 return extract_assume(expr->getArg(0));
1956 res = pet_expr_new_call(ctx, name.c_str(), n_arg);
1957 if (!res)
1958 return NULL;
1960 for (int i = 0; i < n_arg; ++i) {
1961 Expr *arg = expr->getArg(i);
1962 res->args[i] = PetScan::extract_argument(fd, i, arg);
1963 if (!res->args[i])
1964 goto error;
1967 return res;
1968 error:
1969 pet_expr_free(res);
1970 return NULL;
1973 /* Construct a pet_expr representing a (C style) cast.
1975 struct pet_expr *PetScan::extract_expr(CStyleCastExpr *expr)
1977 struct pet_expr *arg;
1978 QualType type;
1980 arg = extract_expr(expr->getSubExpr());
1981 if (!arg)
1982 return NULL;
1984 type = expr->getTypeAsWritten();
1985 return pet_expr_new_cast(ctx, type.getAsString().c_str(), arg);
1988 /* Construct a pet_expr representing an integer.
1990 struct pet_expr *PetScan::extract_expr(IntegerLiteral *expr)
1992 return pet_expr_new_int(extract_int(expr));
1995 /* Try and construct a pet_expr representing "expr".
1997 struct pet_expr *PetScan::extract_expr(Expr *expr)
1999 switch (expr->getStmtClass()) {
2000 case Stmt::UnaryOperatorClass:
2001 return extract_expr(cast<UnaryOperator>(expr));
2002 case Stmt::CompoundAssignOperatorClass:
2003 case Stmt::BinaryOperatorClass:
2004 return extract_expr(cast<BinaryOperator>(expr));
2005 case Stmt::ImplicitCastExprClass:
2006 return extract_expr(cast<ImplicitCastExpr>(expr));
2007 case Stmt::ArraySubscriptExprClass:
2008 case Stmt::DeclRefExprClass:
2009 case Stmt::MemberExprClass:
2010 return extract_access_expr(expr);
2011 case Stmt::IntegerLiteralClass:
2012 return extract_expr(cast<IntegerLiteral>(expr));
2013 case Stmt::FloatingLiteralClass:
2014 return extract_expr(cast<FloatingLiteral>(expr));
2015 case Stmt::ParenExprClass:
2016 return extract_expr(cast<ParenExpr>(expr));
2017 case Stmt::ConditionalOperatorClass:
2018 return extract_expr(cast<ConditionalOperator>(expr));
2019 case Stmt::CallExprClass:
2020 return extract_expr(cast<CallExpr>(expr));
2021 case Stmt::CStyleCastExprClass:
2022 return extract_expr(cast<CStyleCastExpr>(expr));
2023 default:
2024 unsupported(expr);
2026 return NULL;
2029 /* Check if the given initialization statement is an assignment.
2030 * If so, return that assignment. Otherwise return NULL.
2032 BinaryOperator *PetScan::initialization_assignment(Stmt *init)
2034 BinaryOperator *ass;
2036 if (init->getStmtClass() != Stmt::BinaryOperatorClass)
2037 return NULL;
2039 ass = cast<BinaryOperator>(init);
2040 if (ass->getOpcode() != BO_Assign)
2041 return NULL;
2043 return ass;
2046 /* Check if the given initialization statement is a declaration
2047 * of a single variable.
2048 * If so, return that declaration. Otherwise return NULL.
2050 Decl *PetScan::initialization_declaration(Stmt *init)
2052 DeclStmt *decl;
2054 if (init->getStmtClass() != Stmt::DeclStmtClass)
2055 return NULL;
2057 decl = cast<DeclStmt>(init);
2059 if (!decl->isSingleDecl())
2060 return NULL;
2062 return decl->getSingleDecl();
2065 /* Given the assignment operator in the initialization of a for loop,
2066 * extract the induction variable, i.e., the (integer)variable being
2067 * assigned.
2069 ValueDecl *PetScan::extract_induction_variable(BinaryOperator *init)
2071 Expr *lhs;
2072 DeclRefExpr *ref;
2073 ValueDecl *decl;
2074 const Type *type;
2076 lhs = init->getLHS();
2077 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
2078 unsupported(init);
2079 return NULL;
2082 ref = cast<DeclRefExpr>(lhs);
2083 decl = ref->getDecl();
2084 type = decl->getType().getTypePtr();
2086 if (!type->isIntegerType()) {
2087 unsupported(lhs);
2088 return NULL;
2091 return decl;
2094 /* Given the initialization statement of a for loop and the single
2095 * declaration in this initialization statement,
2096 * extract the induction variable, i.e., the (integer) variable being
2097 * declared.
2099 VarDecl *PetScan::extract_induction_variable(Stmt *init, Decl *decl)
2101 VarDecl *vd;
2103 vd = cast<VarDecl>(decl);
2105 const QualType type = vd->getType();
2106 if (!type->isIntegerType()) {
2107 unsupported(init);
2108 return NULL;
2111 if (!vd->getInit()) {
2112 unsupported(init);
2113 return NULL;
2116 return vd;
2119 /* Check that op is of the form iv++ or iv--.
2120 * Return an affine expression "1" or "-1" accordingly.
2122 __isl_give isl_pw_aff *PetScan::extract_unary_increment(
2123 clang::UnaryOperator *op, clang::ValueDecl *iv)
2125 Expr *sub;
2126 DeclRefExpr *ref;
2127 isl_space *space;
2128 isl_aff *aff;
2130 if (!op->isIncrementDecrementOp()) {
2131 unsupported(op);
2132 return NULL;
2135 sub = op->getSubExpr();
2136 if (sub->getStmtClass() != Stmt::DeclRefExprClass) {
2137 unsupported(op);
2138 return NULL;
2141 ref = cast<DeclRefExpr>(sub);
2142 if (ref->getDecl() != iv) {
2143 unsupported(op);
2144 return NULL;
2147 space = isl_space_params_alloc(ctx, 0);
2148 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2150 if (op->isIncrementOp())
2151 aff = isl_aff_add_constant_si(aff, 1);
2152 else
2153 aff = isl_aff_add_constant_si(aff, -1);
2155 return isl_pw_aff_from_aff(aff);
2158 /* If the isl_pw_aff on which isl_pw_aff_foreach_piece is called
2159 * has a single constant expression, then put this constant in *user.
2160 * The caller is assumed to have checked that this function will
2161 * be called exactly once.
2163 static int extract_cst(__isl_take isl_set *set, __isl_take isl_aff *aff,
2164 void *user)
2166 isl_val **inc = (isl_val **)user;
2167 int res = 0;
2169 if (isl_aff_is_cst(aff))
2170 *inc = isl_aff_get_constant_val(aff);
2171 else
2172 res = -1;
2174 isl_set_free(set);
2175 isl_aff_free(aff);
2177 return res;
2180 /* Check if op is of the form
2182 * iv = iv + inc
2184 * and return inc as an affine expression.
2186 * We extract an affine expression from the RHS, subtract iv and return
2187 * the result.
2189 __isl_give isl_pw_aff *PetScan::extract_binary_increment(BinaryOperator *op,
2190 clang::ValueDecl *iv)
2192 Expr *lhs;
2193 DeclRefExpr *ref;
2194 isl_id *id;
2195 isl_space *dim;
2196 isl_aff *aff;
2197 isl_pw_aff *val;
2199 if (op->getOpcode() != BO_Assign) {
2200 unsupported(op);
2201 return NULL;
2204 lhs = op->getLHS();
2205 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
2206 unsupported(op);
2207 return NULL;
2210 ref = cast<DeclRefExpr>(lhs);
2211 if (ref->getDecl() != iv) {
2212 unsupported(op);
2213 return NULL;
2216 val = extract_affine(op->getRHS());
2218 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
2220 dim = isl_space_params_alloc(ctx, 1);
2221 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
2222 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2223 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2225 val = isl_pw_aff_sub(val, isl_pw_aff_from_aff(aff));
2227 return val;
2230 /* Check that op is of the form iv += cst or iv -= cst
2231 * and return an affine expression corresponding oto cst or -cst accordingly.
2233 __isl_give isl_pw_aff *PetScan::extract_compound_increment(
2234 CompoundAssignOperator *op, clang::ValueDecl *iv)
2236 Expr *lhs;
2237 DeclRefExpr *ref;
2238 bool neg = false;
2239 isl_pw_aff *val;
2240 BinaryOperatorKind opcode;
2242 opcode = op->getOpcode();
2243 if (opcode != BO_AddAssign && opcode != BO_SubAssign) {
2244 unsupported(op);
2245 return NULL;
2247 if (opcode == BO_SubAssign)
2248 neg = true;
2250 lhs = op->getLHS();
2251 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
2252 unsupported(op);
2253 return NULL;
2256 ref = cast<DeclRefExpr>(lhs);
2257 if (ref->getDecl() != iv) {
2258 unsupported(op);
2259 return NULL;
2262 val = extract_affine(op->getRHS());
2263 if (neg)
2264 val = isl_pw_aff_neg(val);
2266 return val;
2269 /* Check that the increment of the given for loop increments
2270 * (or decrements) the induction variable "iv" and return
2271 * the increment as an affine expression if successful.
2273 __isl_give isl_pw_aff *PetScan::extract_increment(clang::ForStmt *stmt,
2274 ValueDecl *iv)
2276 Stmt *inc = stmt->getInc();
2278 if (!inc) {
2279 report_missing_increment(stmt);
2280 return NULL;
2283 if (inc->getStmtClass() == Stmt::UnaryOperatorClass)
2284 return extract_unary_increment(cast<UnaryOperator>(inc), iv);
2285 if (inc->getStmtClass() == Stmt::CompoundAssignOperatorClass)
2286 return extract_compound_increment(
2287 cast<CompoundAssignOperator>(inc), iv);
2288 if (inc->getStmtClass() == Stmt::BinaryOperatorClass)
2289 return extract_binary_increment(cast<BinaryOperator>(inc), iv);
2291 unsupported(inc);
2292 return NULL;
2295 /* Embed the given iteration domain in an extra outer loop
2296 * with induction variable "var".
2297 * If this variable appeared as a parameter in the constraints,
2298 * it is replaced by the new outermost dimension.
2300 static __isl_give isl_set *embed(__isl_take isl_set *set,
2301 __isl_take isl_id *var)
2303 int pos;
2305 set = isl_set_insert_dims(set, isl_dim_set, 0, 1);
2306 pos = isl_set_find_dim_by_id(set, isl_dim_param, var);
2307 if (pos >= 0) {
2308 set = isl_set_equate(set, isl_dim_param, pos, isl_dim_set, 0);
2309 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2312 isl_id_free(var);
2313 return set;
2316 /* Return those elements in the space of "cond" that come after
2317 * (based on "sign") an element in "cond".
2319 static __isl_give isl_set *after(__isl_take isl_set *cond, int sign)
2321 isl_map *previous_to_this;
2323 if (sign > 0)
2324 previous_to_this = isl_map_lex_lt(isl_set_get_space(cond));
2325 else
2326 previous_to_this = isl_map_lex_gt(isl_set_get_space(cond));
2328 cond = isl_set_apply(cond, previous_to_this);
2330 return cond;
2333 /* Create the infinite iteration domain
2335 * { [id] : id >= 0 }
2337 * If "scop" has an affine skip of type pet_skip_later,
2338 * then remove those iterations i that have an earlier iteration
2339 * where the skip condition is satisfied, meaning that iteration i
2340 * is not executed.
2341 * Since we are dealing with a loop without loop iterator,
2342 * the skip condition cannot refer to the current loop iterator and
2343 * so effectively, the returned set is of the form
2345 * { [0]; [id] : id >= 1 and not skip }
2347 static __isl_give isl_set *infinite_domain(__isl_take isl_id *id,
2348 struct pet_scop *scop)
2350 isl_ctx *ctx = isl_id_get_ctx(id);
2351 isl_set *domain;
2352 isl_set *skip;
2354 domain = isl_set_nat_universe(isl_space_set_alloc(ctx, 0, 1));
2355 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, id);
2357 if (!pet_scop_has_affine_skip(scop, pet_skip_later))
2358 return domain;
2360 skip = pet_scop_get_affine_skip_domain(scop, pet_skip_later);
2361 skip = embed(skip, isl_id_copy(id));
2362 skip = isl_set_intersect(skip , isl_set_copy(domain));
2363 domain = isl_set_subtract(domain, after(skip, 1));
2365 return domain;
2368 /* Create an identity affine expression on the space containing "domain",
2369 * which is assumed to be one-dimensional.
2371 static __isl_give isl_aff *identity_aff(__isl_keep isl_set *domain)
2373 isl_local_space *ls;
2375 ls = isl_local_space_from_space(isl_set_get_space(domain));
2376 return isl_aff_var_on_domain(ls, isl_dim_set, 0);
2379 /* Create an affine expression that maps elements
2380 * of a single-dimensional array "id_test" to the previous element
2381 * (according to "inc"), provided this element belongs to "domain".
2382 * That is, create the affine expression
2384 * { id[x] -> id[x - inc] : x - inc in domain }
2386 static __isl_give isl_multi_pw_aff *map_to_previous(__isl_take isl_id *id_test,
2387 __isl_take isl_set *domain, __isl_take isl_val *inc)
2389 isl_space *space;
2390 isl_local_space *ls;
2391 isl_aff *aff;
2392 isl_multi_pw_aff *prev;
2394 space = isl_set_get_space(domain);
2395 ls = isl_local_space_from_space(space);
2396 aff = isl_aff_var_on_domain(ls, isl_dim_set, 0);
2397 aff = isl_aff_add_constant_val(aff, isl_val_neg(inc));
2398 prev = isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
2399 domain = isl_set_preimage_multi_pw_aff(domain,
2400 isl_multi_pw_aff_copy(prev));
2401 prev = isl_multi_pw_aff_intersect_domain(prev, domain);
2402 prev = isl_multi_pw_aff_set_tuple_id(prev, isl_dim_out, id_test);
2404 return prev;
2407 /* Add an implication to "scop" expressing that if an element of
2408 * virtual array "id_test" has value "satisfied" then all previous elements
2409 * of this array also have that value. The set of previous elements
2410 * is bounded by "domain". If "sign" is negative then the iterator
2411 * is decreasing and we express that all subsequent array elements
2412 * (but still defined previously) have the same value.
2414 static struct pet_scop *add_implication(struct pet_scop *scop,
2415 __isl_take isl_id *id_test, __isl_take isl_set *domain, int sign,
2416 int satisfied)
2418 isl_space *space;
2419 isl_map *map;
2421 domain = isl_set_set_tuple_id(domain, id_test);
2422 space = isl_set_get_space(domain);
2423 if (sign > 0)
2424 map = isl_map_lex_ge(space);
2425 else
2426 map = isl_map_lex_le(space);
2427 map = isl_map_intersect_range(map, domain);
2428 scop = pet_scop_add_implication(scop, map, satisfied);
2430 return scop;
2433 /* Add a filter to "scop" that imposes that it is only executed
2434 * when the variable identified by "id_test" has a zero value
2435 * for all previous iterations of "domain".
2437 * In particular, add a filter that imposes that the array
2438 * has a zero value at the previous iteration of domain and
2439 * add an implication that implies that it then has that
2440 * value for all previous iterations.
2442 static struct pet_scop *scop_add_break(struct pet_scop *scop,
2443 __isl_take isl_id *id_test, __isl_take isl_set *domain,
2444 __isl_take isl_val *inc)
2446 isl_multi_pw_aff *prev;
2447 int sign = isl_val_sgn(inc);
2449 prev = map_to_previous(isl_id_copy(id_test), isl_set_copy(domain), inc);
2450 scop = add_implication(scop, id_test, domain, sign, 0);
2451 scop = pet_scop_filter(scop, prev, 0);
2453 return scop;
2456 /* Construct a pet_scop for an infinite loop around the given body.
2458 * We extract a pet_scop for the body and then embed it in a loop with
2459 * iteration domain
2461 * { [t] : t >= 0 }
2463 * and schedule
2465 * { [t] -> [t] }
2467 * If the body contains any break, then it is taken into
2468 * account in infinite_domain (if the skip condition is affine)
2469 * or in scop_add_break (if the skip condition is not affine).
2471 * If we were only able to extract part of the body, then simply
2472 * return that part.
2474 struct pet_scop *PetScan::extract_infinite_loop(Stmt *body)
2476 isl_id *id, *id_test;
2477 isl_set *domain;
2478 isl_aff *ident;
2479 struct pet_scop *scop;
2480 bool has_var_break;
2482 scop = extract(body);
2483 if (!scop)
2484 return NULL;
2485 if (partial)
2486 return scop;
2488 id = isl_id_alloc(ctx, "t", NULL);
2489 domain = infinite_domain(isl_id_copy(id), scop);
2490 ident = identity_aff(domain);
2492 has_var_break = pet_scop_has_var_skip(scop, pet_skip_later);
2493 if (has_var_break)
2494 id_test = pet_scop_get_skip_id(scop, pet_skip_later);
2496 scop = pet_scop_embed(scop, isl_set_copy(domain),
2497 isl_aff_copy(ident), ident, id);
2498 if (has_var_break)
2499 scop = scop_add_break(scop, id_test, domain, isl_val_one(ctx));
2500 else
2501 isl_set_free(domain);
2503 return scop;
2506 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
2508 * for (;;)
2509 * body
2512 struct pet_scop *PetScan::extract_infinite_for(ForStmt *stmt)
2514 clear_assignments clear(assigned_value);
2515 clear.TraverseStmt(stmt->getBody());
2517 return extract_infinite_loop(stmt->getBody());
2520 /* Create an index expression for an access to a virtual array
2521 * representing the result of a condition.
2522 * Unlike other accessed data, the id of the array is NULL as
2523 * there is no ValueDecl in the program corresponding to the virtual
2524 * array.
2525 * The array starts out as a scalar, but grows along with the
2526 * statement writing to the array in pet_scop_embed.
2528 static __isl_give isl_multi_pw_aff *create_test_index(isl_ctx *ctx, int test_nr)
2530 isl_space *dim = isl_space_alloc(ctx, 0, 0, 0);
2531 isl_id *id;
2532 char name[50];
2534 snprintf(name, sizeof(name), "__pet_test_%d", test_nr);
2535 id = isl_id_alloc(ctx, name, NULL);
2536 dim = isl_space_set_tuple_id(dim, isl_dim_out, id);
2537 return isl_multi_pw_aff_zero(dim);
2540 /* Add an array with the given extent (range of "index") to the list
2541 * of arrays in "scop" and return the extended pet_scop.
2542 * The array is marked as attaining values 0 and 1 only and
2543 * as each element being assigned at most once.
2545 static struct pet_scop *scop_add_array(struct pet_scop *scop,
2546 __isl_keep isl_multi_pw_aff *index, clang::ASTContext &ast_ctx)
2548 isl_ctx *ctx = isl_multi_pw_aff_get_ctx(index);
2549 isl_space *dim;
2550 struct pet_array *array;
2551 isl_map *access;
2553 if (!scop)
2554 return NULL;
2555 if (!ctx)
2556 goto error;
2558 array = isl_calloc_type(ctx, struct pet_array);
2559 if (!array)
2560 goto error;
2562 access = isl_map_from_multi_pw_aff(isl_multi_pw_aff_copy(index));
2563 array->extent = isl_map_range(access);
2564 dim = isl_space_params_alloc(ctx, 0);
2565 array->context = isl_set_universe(dim);
2566 dim = isl_space_set_alloc(ctx, 0, 1);
2567 array->value_bounds = isl_set_universe(dim);
2568 array->value_bounds = isl_set_lower_bound_si(array->value_bounds,
2569 isl_dim_set, 0, 0);
2570 array->value_bounds = isl_set_upper_bound_si(array->value_bounds,
2571 isl_dim_set, 0, 1);
2572 array->element_type = strdup("int");
2573 array->element_size = ast_ctx.getTypeInfo(ast_ctx.IntTy).first / 8;
2574 array->uniquely_defined = 1;
2576 if (!array->extent || !array->context)
2577 array = pet_array_free(array);
2579 scop = pet_scop_add_array(scop, array);
2581 return scop;
2582 error:
2583 pet_scop_free(scop);
2584 return NULL;
2587 /* Construct a pet_scop for a while loop of the form
2589 * while (pa)
2590 * body
2592 * In particular, construct a scop for an infinite loop around body and
2593 * intersect the domain with the affine expression.
2594 * Note that this intersection may result in an empty loop.
2596 struct pet_scop *PetScan::extract_affine_while(__isl_take isl_pw_aff *pa,
2597 Stmt *body)
2599 struct pet_scop *scop;
2600 isl_set *dom;
2601 isl_set *valid;
2603 valid = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2604 dom = isl_pw_aff_non_zero_set(pa);
2605 scop = extract_infinite_loop(body);
2606 scop = pet_scop_restrict(scop, dom);
2607 scop = pet_scop_restrict_context(scop, valid);
2609 return scop;
2612 /* Construct a scop for a while, given the scops for the condition
2613 * and the body, the filter identifier and the iteration domain of
2614 * the while loop.
2616 * In particular, the scop for the condition is filtered to depend
2617 * on "id_test" evaluating to true for all previous iterations
2618 * of the loop, while the scop for the body is filtered to depend
2619 * on "id_test" evaluating to true for all iterations up to the
2620 * current iteration.
2621 * The actual filter only imposes that this virtual array has
2622 * value one on the previous or the current iteration.
2623 * The fact that this condition also applies to the previous
2624 * iterations is enforced by an implication.
2626 * These filtered scops are then combined into a single scop.
2628 * "sign" is positive if the iterator increases and negative
2629 * if it decreases.
2631 static struct pet_scop *scop_add_while(struct pet_scop *scop_cond,
2632 struct pet_scop *scop_body, __isl_take isl_id *id_test,
2633 __isl_take isl_set *domain, __isl_take isl_val *inc)
2635 isl_ctx *ctx = isl_set_get_ctx(domain);
2636 isl_space *space;
2637 isl_multi_pw_aff *test_index;
2638 isl_multi_pw_aff *prev;
2639 int sign = isl_val_sgn(inc);
2640 struct pet_scop *scop;
2642 prev = map_to_previous(isl_id_copy(id_test), isl_set_copy(domain), inc);
2643 scop_cond = pet_scop_filter(scop_cond, prev, 1);
2645 space = isl_space_map_from_set(isl_set_get_space(domain));
2646 test_index = isl_multi_pw_aff_identity(space);
2647 test_index = isl_multi_pw_aff_set_tuple_id(test_index, isl_dim_out,
2648 isl_id_copy(id_test));
2649 scop_body = pet_scop_filter(scop_body, test_index, 1);
2651 scop = pet_scop_add_seq(ctx, scop_cond, scop_body);
2652 scop = add_implication(scop, id_test, domain, sign, 1);
2654 return scop;
2657 /* Check if the while loop is of the form
2659 * while (affine expression)
2660 * body
2662 * If so, call extract_affine_while to construct a scop.
2664 * Otherwise, construct a generic while scop, with iteration domain
2665 * { [t] : t >= 0 }. The scop consists of two parts, one for
2666 * evaluating the condition and one for the body.
2667 * The schedule is adjusted to reflect that the condition is evaluated
2668 * before the body is executed and the body is filtered to depend
2669 * on the result of the condition evaluating to true on all iterations
2670 * up to the current iteration, while the evaluation of the condition itself
2671 * is filtered to depend on the result of the condition evaluating to true
2672 * on all previous iterations.
2673 * The context of the scop representing the body is dropped
2674 * because we don't know how many times the body will be executed,
2675 * if at all.
2677 * If the body contains any break, then it is taken into
2678 * account in infinite_domain (if the skip condition is affine)
2679 * or in scop_add_break (if the skip condition is not affine).
2681 * If we were only able to extract part of the body, then simply
2682 * return that part.
2684 struct pet_scop *PetScan::extract(WhileStmt *stmt)
2686 Expr *cond;
2687 int test_nr, stmt_nr;
2688 isl_id *id, *id_test, *id_break_test;
2689 isl_multi_pw_aff *test_index;
2690 isl_set *domain;
2691 isl_aff *ident;
2692 isl_pw_aff *pa;
2693 struct pet_scop *scop, *scop_body;
2694 bool has_var_break;
2696 cond = stmt->getCond();
2697 if (!cond) {
2698 unsupported(stmt);
2699 return NULL;
2702 clear_assignments clear(assigned_value);
2703 clear.TraverseStmt(stmt->getBody());
2705 pa = try_extract_affine_condition(cond);
2706 if (pa)
2707 return extract_affine_while(pa, stmt->getBody());
2709 if (!allow_nested) {
2710 unsupported(stmt);
2711 return NULL;
2714 test_nr = n_test++;
2715 stmt_nr = n_stmt++;
2716 scop_body = extract(stmt->getBody());
2717 if (partial)
2718 return scop_body;
2720 test_index = create_test_index(ctx, test_nr);
2721 scop = extract_non_affine_condition(cond, stmt_nr,
2722 isl_multi_pw_aff_copy(test_index));
2723 scop = scop_add_array(scop, test_index, ast_context);
2724 id_test = isl_multi_pw_aff_get_tuple_id(test_index, isl_dim_out);
2725 isl_multi_pw_aff_free(test_index);
2727 id = isl_id_alloc(ctx, "t", NULL);
2728 domain = infinite_domain(isl_id_copy(id), scop_body);
2729 ident = identity_aff(domain);
2731 has_var_break = pet_scop_has_var_skip(scop_body, pet_skip_later);
2732 if (has_var_break)
2733 id_break_test = pet_scop_get_skip_id(scop_body, pet_skip_later);
2735 scop = pet_scop_prefix(scop, 0);
2736 scop = pet_scop_embed(scop, isl_set_copy(domain), isl_aff_copy(ident),
2737 isl_aff_copy(ident), isl_id_copy(id));
2738 scop_body = pet_scop_reset_context(scop_body);
2739 scop_body = pet_scop_prefix(scop_body, 1);
2740 scop_body = pet_scop_embed(scop_body, isl_set_copy(domain),
2741 isl_aff_copy(ident), ident, id);
2743 if (has_var_break) {
2744 scop = scop_add_break(scop, isl_id_copy(id_break_test),
2745 isl_set_copy(domain), isl_val_one(ctx));
2746 scop_body = scop_add_break(scop_body, id_break_test,
2747 isl_set_copy(domain), isl_val_one(ctx));
2749 scop = scop_add_while(scop, scop_body, id_test, domain,
2750 isl_val_one(ctx));
2752 return scop;
2755 /* Check whether "cond" expresses a simple loop bound
2756 * on the only set dimension.
2757 * In particular, if "up" is set then "cond" should contain only
2758 * upper bounds on the set dimension.
2759 * Otherwise, it should contain only lower bounds.
2761 static bool is_simple_bound(__isl_keep isl_set *cond, __isl_keep isl_val *inc)
2763 if (isl_val_is_pos(inc))
2764 return !isl_set_dim_has_any_lower_bound(cond, isl_dim_set, 0);
2765 else
2766 return !isl_set_dim_has_any_upper_bound(cond, isl_dim_set, 0);
2769 /* Extend a condition on a given iteration of a loop to one that
2770 * imposes the same condition on all previous iterations.
2771 * "domain" expresses the lower [upper] bound on the iterations
2772 * when inc is positive [negative].
2774 * In particular, we construct the condition (when inc is positive)
2776 * forall i' : (domain(i') and i' <= i) => cond(i')
2778 * which is equivalent to
2780 * not exists i' : domain(i') and i' <= i and not cond(i')
2782 * We construct this set by negating cond, applying a map
2784 * { [i'] -> [i] : domain(i') and i' <= i }
2786 * and then negating the result again.
2788 static __isl_give isl_set *valid_for_each_iteration(__isl_take isl_set *cond,
2789 __isl_take isl_set *domain, __isl_take isl_val *inc)
2791 isl_map *previous_to_this;
2793 if (isl_val_is_pos(inc))
2794 previous_to_this = isl_map_lex_le(isl_set_get_space(domain));
2795 else
2796 previous_to_this = isl_map_lex_ge(isl_set_get_space(domain));
2798 previous_to_this = isl_map_intersect_domain(previous_to_this, domain);
2800 cond = isl_set_complement(cond);
2801 cond = isl_set_apply(cond, previous_to_this);
2802 cond = isl_set_complement(cond);
2804 isl_val_free(inc);
2806 return cond;
2809 /* Construct a domain of the form
2811 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
2813 static __isl_give isl_set *strided_domain(__isl_take isl_id *id,
2814 __isl_take isl_pw_aff *init, __isl_take isl_val *inc)
2816 isl_aff *aff;
2817 isl_space *dim;
2818 isl_set *set;
2820 init = isl_pw_aff_insert_dims(init, isl_dim_in, 0, 1);
2821 dim = isl_pw_aff_get_domain_space(init);
2822 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2823 aff = isl_aff_add_coefficient_val(aff, isl_dim_in, 0, inc);
2824 init = isl_pw_aff_add(init, isl_pw_aff_from_aff(aff));
2826 dim = isl_space_set_alloc(isl_pw_aff_get_ctx(init), 1, 1);
2827 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
2828 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2829 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2831 set = isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff), init);
2833 set = isl_set_lower_bound_si(set, isl_dim_set, 0, 0);
2835 return isl_set_params(set);
2838 /* Assuming "cond" represents a bound on a loop where the loop
2839 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2840 * is possible.
2842 * Under the given assumptions, wrapping is only possible if "cond" allows
2843 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2844 * increasing iterator and 0 in case of a decreasing iterator.
2846 static bool can_wrap(__isl_keep isl_set *cond, ValueDecl *iv,
2847 __isl_keep isl_val *inc)
2849 bool cw;
2850 isl_ctx *ctx;
2851 isl_val *limit;
2852 isl_set *test;
2854 test = isl_set_copy(cond);
2856 ctx = isl_set_get_ctx(test);
2857 if (isl_val_is_neg(inc))
2858 limit = isl_val_zero(ctx);
2859 else {
2860 limit = isl_val_int_from_ui(ctx, get_type_size(iv));
2861 limit = isl_val_2exp(limit);
2862 limit = isl_val_sub_ui(limit, 1);
2865 test = isl_set_fix_val(cond, isl_dim_set, 0, limit);
2866 cw = !isl_set_is_empty(test);
2867 isl_set_free(test);
2869 return cw;
2872 /* Given a one-dimensional space, construct the following affine expression
2873 * on this space
2875 * { [v] -> [v mod 2^width] }
2877 * where width is the number of bits used to represent the values
2878 * of the unsigned variable "iv".
2880 static __isl_give isl_aff *compute_wrapping(__isl_take isl_space *dim,
2881 ValueDecl *iv)
2883 isl_ctx *ctx;
2884 isl_val *mod;
2885 isl_aff *aff;
2887 ctx = isl_space_get_ctx(dim);
2888 mod = isl_val_int_from_ui(ctx, get_type_size(iv));
2889 mod = isl_val_2exp(mod);
2891 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2892 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2893 aff = isl_aff_mod_val(aff, mod);
2895 return aff;
2898 /* Project out the parameter "id" from "set".
2900 static __isl_give isl_set *set_project_out_by_id(__isl_take isl_set *set,
2901 __isl_keep isl_id *id)
2903 int pos;
2905 pos = isl_set_find_dim_by_id(set, isl_dim_param, id);
2906 if (pos >= 0)
2907 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2909 return set;
2912 /* Compute the set of parameters for which "set1" is a subset of "set2".
2914 * set1 is a subset of set2 if
2916 * forall i in set1 : i in set2
2918 * or
2920 * not exists i in set1 and i not in set2
2922 * i.e.,
2924 * not exists i in set1 \ set2
2926 static __isl_give isl_set *enforce_subset(__isl_take isl_set *set1,
2927 __isl_take isl_set *set2)
2929 return isl_set_complement(isl_set_params(isl_set_subtract(set1, set2)));
2932 /* Compute the set of parameter values for which "cond" holds
2933 * on the next iteration for each element of "dom".
2935 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2936 * and then compute the set of parameters for which the result is a subset
2937 * of "cond".
2939 static __isl_give isl_set *valid_on_next(__isl_take isl_set *cond,
2940 __isl_take isl_set *dom, __isl_take isl_val *inc)
2942 isl_space *space;
2943 isl_aff *aff;
2944 isl_map *next;
2946 space = isl_set_get_space(dom);
2947 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2948 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2949 aff = isl_aff_add_constant_val(aff, inc);
2950 next = isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2952 dom = isl_set_apply(dom, next);
2954 return enforce_subset(dom, cond);
2957 /* Does "id" refer to a nested access?
2959 static bool is_nested_parameter(__isl_keep isl_id *id)
2961 return id && isl_id_get_user(id) && !isl_id_get_name(id);
2964 /* Does parameter "pos" of "space" refer to a nested access?
2966 static bool is_nested_parameter(__isl_keep isl_space *space, int pos)
2968 bool nested;
2969 isl_id *id;
2971 id = isl_space_get_dim_id(space, isl_dim_param, pos);
2972 nested = is_nested_parameter(id);
2973 isl_id_free(id);
2975 return nested;
2978 /* Does "space" involve any parameters that refer to nested
2979 * accesses, i.e., parameters with no name?
2981 static bool has_nested(__isl_keep isl_space *space)
2983 int nparam;
2985 nparam = isl_space_dim(space, isl_dim_param);
2986 for (int i = 0; i < nparam; ++i)
2987 if (is_nested_parameter(space, i))
2988 return true;
2990 return false;
2993 /* Does "pa" involve any parameters that refer to nested
2994 * accesses, i.e., parameters with no name?
2996 static bool has_nested(__isl_keep isl_pw_aff *pa)
2998 isl_space *space;
2999 bool nested;
3001 space = isl_pw_aff_get_space(pa);
3002 nested = has_nested(space);
3003 isl_space_free(space);
3005 return nested;
3008 /* Construct a pet_scop for a for statement.
3009 * The for loop is required to be of the form
3011 * for (i = init; condition; ++i)
3013 * or
3015 * for (i = init; condition; --i)
3017 * The initialization of the for loop should either be an assignment
3018 * to an integer variable, or a declaration of such a variable with
3019 * initialization.
3021 * The condition is allowed to contain nested accesses, provided
3022 * they are not being written to inside the body of the loop.
3023 * Otherwise, or if the condition is otherwise non-affine, the for loop is
3024 * essentially treated as a while loop, with iteration domain
3025 * { [i] : i >= init }.
3027 * We extract a pet_scop for the body and then embed it in a loop with
3028 * iteration domain and schedule
3030 * { [i] : i >= init and condition' }
3031 * { [i] -> [i] }
3033 * or
3035 * { [i] : i <= init and condition' }
3036 * { [i] -> [-i] }
3038 * Where condition' is equal to condition if the latter is
3039 * a simple upper [lower] bound and a condition that is extended
3040 * to apply to all previous iterations otherwise.
3042 * If the condition is non-affine, then we drop the condition from the
3043 * iteration domain and instead create a separate statement
3044 * for evaluating the condition. The body is then filtered to depend
3045 * on the result of the condition evaluating to true on all iterations
3046 * up to the current iteration, while the evaluation the condition itself
3047 * is filtered to depend on the result of the condition evaluating to true
3048 * on all previous iterations.
3049 * The context of the scop representing the body is dropped
3050 * because we don't know how many times the body will be executed,
3051 * if at all.
3053 * If the stride of the loop is not 1, then "i >= init" is replaced by
3055 * (exists a: i = init + stride * a and a >= 0)
3057 * If the loop iterator i is unsigned, then wrapping may occur.
3058 * We therefore use a virtual iterator instead that does not wrap.
3059 * However, the condition in the code applies
3060 * to the wrapped value, so we need to change condition(i)
3061 * into condition([i % 2^width]). Similarly, we replace all accesses
3062 * to the original iterator by the wrapping of the virtual iterator.
3063 * Note that there may be no need to perform this final wrapping
3064 * if the loop condition (after wrapping) satisfies certain conditions.
3065 * However, the is_simple_bound condition is not enough since it doesn't
3066 * check if there even is an upper bound.
3068 * Wrapping on unsigned iterators can be avoided entirely if
3069 * loop condition is simple, the loop iterator is incremented
3070 * [decremented] by one and the last value before wrapping cannot
3071 * possibly satisfy the loop condition.
3073 * Before extracting a pet_scop from the body we remove all
3074 * assignments in assigned_value to variables that are assigned
3075 * somewhere in the body of the loop.
3077 * Valid parameters for a for loop are those for which the initial
3078 * value itself, the increment on each domain iteration and
3079 * the condition on both the initial value and
3080 * the result of incrementing the iterator for each iteration of the domain
3081 * can be evaluated.
3082 * If the loop condition is non-affine, then we only consider validity
3083 * of the initial value.
3085 * If the body contains any break, then we keep track of it in "skip"
3086 * (if the skip condition is affine) or it is handled in scop_add_break
3087 * (if the skip condition is not affine).
3088 * Note that the affine break condition needs to be considered with
3089 * respect to previous iterations in the virtual domain (if any).
3091 * If we were only able to extract part of the body, then simply
3092 * return that part.
3094 struct pet_scop *PetScan::extract_for(ForStmt *stmt)
3096 BinaryOperator *ass;
3097 Decl *decl;
3098 Stmt *init;
3099 Expr *lhs, *rhs;
3100 ValueDecl *iv;
3101 isl_local_space *ls;
3102 isl_set *domain;
3103 isl_aff *sched;
3104 isl_set *cond = NULL;
3105 isl_set *skip = NULL;
3106 isl_id *id, *id_test = NULL, *id_break_test;
3107 struct pet_scop *scop, *scop_cond = NULL;
3108 assigned_value_cache cache(assigned_value);
3109 isl_val *inc;
3110 bool was_assigned;
3111 bool is_one;
3112 bool is_unsigned;
3113 bool is_simple;
3114 bool is_virtual;
3115 bool has_affine_break;
3116 bool has_var_break;
3117 isl_aff *wrap = NULL;
3118 isl_pw_aff *pa, *pa_inc, *init_val;
3119 isl_set *valid_init;
3120 isl_set *valid_cond;
3121 isl_set *valid_cond_init;
3122 isl_set *valid_cond_next;
3123 isl_set *valid_inc;
3124 int stmt_id;
3126 if (!stmt->getInit() && !stmt->getCond() && !stmt->getInc())
3127 return extract_infinite_for(stmt);
3129 init = stmt->getInit();
3130 if (!init) {
3131 unsupported(stmt);
3132 return NULL;
3134 if ((ass = initialization_assignment(init)) != NULL) {
3135 iv = extract_induction_variable(ass);
3136 if (!iv)
3137 return NULL;
3138 lhs = ass->getLHS();
3139 rhs = ass->getRHS();
3140 } else if ((decl = initialization_declaration(init)) != NULL) {
3141 VarDecl *var = extract_induction_variable(init, decl);
3142 if (!var)
3143 return NULL;
3144 iv = var;
3145 rhs = var->getInit();
3146 lhs = create_DeclRefExpr(var);
3147 } else {
3148 unsupported(stmt->getInit());
3149 return NULL;
3152 assigned_value.erase(iv);
3153 clear_assignments clear(assigned_value);
3154 clear.TraverseStmt(stmt->getBody());
3156 was_assigned = assigned_value.find(iv) != assigned_value.end();
3157 clear_assignment(assigned_value, iv);
3158 init_val = extract_affine(rhs);
3159 if (!was_assigned)
3160 assigned_value.erase(iv);
3161 if (!init_val)
3162 return NULL;
3164 pa_inc = extract_increment(stmt, iv);
3165 if (!pa_inc) {
3166 isl_pw_aff_free(init_val);
3167 return NULL;
3170 inc = NULL;
3171 if (isl_pw_aff_n_piece(pa_inc) != 1 ||
3172 isl_pw_aff_foreach_piece(pa_inc, &extract_cst, &inc) < 0) {
3173 isl_pw_aff_free(init_val);
3174 isl_pw_aff_free(pa_inc);
3175 unsupported(stmt->getInc());
3176 isl_val_free(inc);
3177 return NULL;
3180 pa = try_extract_nested_condition(stmt->getCond());
3181 if (allow_nested && (!pa || has_nested(pa)))
3182 stmt_id = n_stmt++;
3184 scop = extract(stmt->getBody());
3185 if (partial) {
3186 isl_pw_aff_free(init_val);
3187 isl_pw_aff_free(pa_inc);
3188 isl_pw_aff_free(pa);
3189 isl_val_free(inc);
3190 return scop;
3193 valid_inc = isl_pw_aff_domain(pa_inc);
3195 is_unsigned = iv->getType()->isUnsignedIntegerType();
3197 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
3199 has_affine_break = scop &&
3200 pet_scop_has_affine_skip(scop, pet_skip_later);
3201 if (has_affine_break)
3202 skip = pet_scop_get_affine_skip_domain(scop, pet_skip_later);
3203 has_var_break = scop && pet_scop_has_var_skip(scop, pet_skip_later);
3204 if (has_var_break)
3205 id_break_test = pet_scop_get_skip_id(scop, pet_skip_later);
3207 if (pa && !is_nested_allowed(pa, scop)) {
3208 isl_pw_aff_free(pa);
3209 pa = NULL;
3212 if (!allow_nested && !pa)
3213 pa = try_extract_affine_condition(stmt->getCond());
3214 valid_cond = isl_pw_aff_domain(isl_pw_aff_copy(pa));
3215 cond = isl_pw_aff_non_zero_set(pa);
3216 if (allow_nested && !cond) {
3217 isl_multi_pw_aff *test_index;
3218 int save_n_stmt = n_stmt;
3219 test_index = create_test_index(ctx, n_test++);
3220 n_stmt = stmt_id;
3221 scop_cond = extract_non_affine_condition(stmt->getCond(),
3222 n_stmt++, isl_multi_pw_aff_copy(test_index));
3223 n_stmt = save_n_stmt;
3224 scop_cond = scop_add_array(scop_cond, test_index, ast_context);
3225 id_test = isl_multi_pw_aff_get_tuple_id(test_index,
3226 isl_dim_out);
3227 isl_multi_pw_aff_free(test_index);
3228 scop_cond = pet_scop_prefix(scop_cond, 0);
3229 scop = pet_scop_reset_context(scop);
3230 scop = pet_scop_prefix(scop, 1);
3231 cond = isl_set_universe(isl_space_set_alloc(ctx, 0, 0));
3234 cond = embed(cond, isl_id_copy(id));
3235 skip = embed(skip, isl_id_copy(id));
3236 valid_cond = isl_set_coalesce(valid_cond);
3237 valid_cond = embed(valid_cond, isl_id_copy(id));
3238 valid_inc = embed(valid_inc, isl_id_copy(id));
3239 is_one = isl_val_is_one(inc) || isl_val_is_negone(inc);
3240 is_virtual = is_unsigned && (!is_one || can_wrap(cond, iv, inc));
3242 valid_cond_init = enforce_subset(
3243 isl_set_from_pw_aff(isl_pw_aff_copy(init_val)),
3244 isl_set_copy(valid_cond));
3245 if (is_one && !is_virtual) {
3246 isl_pw_aff_free(init_val);
3247 pa = extract_comparison(isl_val_is_pos(inc) ? BO_GE : BO_LE,
3248 lhs, rhs, init);
3249 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(pa));
3250 valid_init = set_project_out_by_id(valid_init, id);
3251 domain = isl_pw_aff_non_zero_set(pa);
3252 } else {
3253 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(init_val));
3254 domain = strided_domain(isl_id_copy(id), init_val,
3255 isl_val_copy(inc));
3258 domain = embed(domain, isl_id_copy(id));
3259 if (is_virtual) {
3260 isl_map *rev_wrap;
3261 wrap = compute_wrapping(isl_set_get_space(cond), iv);
3262 rev_wrap = isl_map_from_aff(isl_aff_copy(wrap));
3263 rev_wrap = isl_map_reverse(rev_wrap);
3264 cond = isl_set_apply(cond, isl_map_copy(rev_wrap));
3265 skip = isl_set_apply(skip, isl_map_copy(rev_wrap));
3266 valid_cond = isl_set_apply(valid_cond, isl_map_copy(rev_wrap));
3267 valid_inc = isl_set_apply(valid_inc, rev_wrap);
3269 is_simple = is_simple_bound(cond, inc);
3270 if (!is_simple) {
3271 cond = isl_set_gist(cond, isl_set_copy(domain));
3272 is_simple = is_simple_bound(cond, inc);
3274 if (!is_simple)
3275 cond = valid_for_each_iteration(cond,
3276 isl_set_copy(domain), isl_val_copy(inc));
3277 domain = isl_set_intersect(domain, cond);
3278 if (has_affine_break) {
3279 skip = isl_set_intersect(skip , isl_set_copy(domain));
3280 skip = after(skip, isl_val_sgn(inc));
3281 domain = isl_set_subtract(domain, skip);
3283 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, isl_id_copy(id));
3284 ls = isl_local_space_from_space(isl_set_get_space(domain));
3285 sched = isl_aff_var_on_domain(ls, isl_dim_set, 0);
3286 if (isl_val_is_neg(inc))
3287 sched = isl_aff_neg(sched);
3289 valid_cond_next = valid_on_next(valid_cond, isl_set_copy(domain),
3290 isl_val_copy(inc));
3291 valid_inc = enforce_subset(isl_set_copy(domain), valid_inc);
3293 if (!is_virtual)
3294 wrap = identity_aff(domain);
3296 scop_cond = pet_scop_embed(scop_cond, isl_set_copy(domain),
3297 isl_aff_copy(sched), isl_aff_copy(wrap), isl_id_copy(id));
3298 scop = pet_scop_embed(scop, isl_set_copy(domain), sched, wrap, id);
3299 scop = resolve_nested(scop);
3300 if (has_var_break)
3301 scop = scop_add_break(scop, id_break_test, isl_set_copy(domain),
3302 isl_val_copy(inc));
3303 if (id_test) {
3304 scop = scop_add_while(scop_cond, scop, id_test, domain,
3305 isl_val_copy(inc));
3306 isl_set_free(valid_inc);
3307 } else {
3308 scop = pet_scop_restrict_context(scop, valid_inc);
3309 scop = pet_scop_restrict_context(scop, valid_cond_next);
3310 scop = pet_scop_restrict_context(scop, valid_cond_init);
3311 isl_set_free(domain);
3313 clear_assignment(assigned_value, iv);
3315 isl_val_free(inc);
3317 scop = pet_scop_restrict_context(scop, valid_init);
3319 return scop;
3322 /* Try and construct a pet_scop corresponding to a compound statement.
3324 * "skip_declarations" is set if we should skip initial declarations
3325 * in the children of the compound statements. This then implies
3326 * that this sequence of children should not be treated as a block
3327 * since the initial statements may be skipped.
3329 struct pet_scop *PetScan::extract(CompoundStmt *stmt, bool skip_declarations)
3331 return extract(stmt->children(), !skip_declarations, skip_declarations);
3334 /* Does parameter "pos" of "map" refer to a nested access?
3336 static bool is_nested_parameter(__isl_keep isl_map *map, int pos)
3338 bool nested;
3339 isl_id *id;
3341 id = isl_map_get_dim_id(map, isl_dim_param, pos);
3342 nested = is_nested_parameter(id);
3343 isl_id_free(id);
3345 return nested;
3348 /* How many parameters of "space" refer to nested accesses, i.e., have no name?
3350 static int n_nested_parameter(__isl_keep isl_space *space)
3352 int n = 0;
3353 int nparam;
3355 nparam = isl_space_dim(space, isl_dim_param);
3356 for (int i = 0; i < nparam; ++i)
3357 if (is_nested_parameter(space, i))
3358 ++n;
3360 return n;
3363 /* How many parameters of "map" refer to nested accesses, i.e., have no name?
3365 static int n_nested_parameter(__isl_keep isl_map *map)
3367 isl_space *space;
3368 int n;
3370 space = isl_map_get_space(map);
3371 n = n_nested_parameter(space);
3372 isl_space_free(space);
3374 return n;
3377 /* For each nested access parameter in "space",
3378 * construct a corresponding pet_expr, place it in args and
3379 * record its position in "param2pos".
3380 * "n_arg" is the number of elements that are already in args.
3381 * The position recorded in "param2pos" takes this number into account.
3382 * If the pet_expr corresponding to a parameter is identical to
3383 * the pet_expr corresponding to an earlier parameter, then these two
3384 * parameters are made to refer to the same element in args.
3386 * Return the final number of elements in args or -1 if an error has occurred.
3388 int PetScan::extract_nested(__isl_keep isl_space *space,
3389 int n_arg, struct pet_expr **args, std::map<int,int> &param2pos)
3391 int nparam;
3393 nparam = isl_space_dim(space, isl_dim_param);
3394 for (int i = 0; i < nparam; ++i) {
3395 int j;
3396 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
3397 Expr *nested;
3399 if (!is_nested_parameter(id)) {
3400 isl_id_free(id);
3401 continue;
3404 nested = (Expr *) isl_id_get_user(id);
3405 args[n_arg] = extract_expr(nested);
3406 isl_id_free(id);
3407 if (!args[n_arg])
3408 return -1;
3410 for (j = 0; j < n_arg; ++j)
3411 if (pet_expr_is_equal(args[j], args[n_arg]))
3412 break;
3414 if (j < n_arg) {
3415 pet_expr_free(args[n_arg]);
3416 args[n_arg] = NULL;
3417 param2pos[i] = j;
3418 } else
3419 param2pos[i] = n_arg++;
3422 return n_arg;
3425 /* For each nested access parameter in the access relations in "expr",
3426 * construct a corresponding pet_expr, place it in expr->args and
3427 * record its position in "param2pos".
3428 * n is the number of nested access parameters.
3430 struct pet_expr *PetScan::extract_nested(struct pet_expr *expr, int n,
3431 std::map<int,int> &param2pos)
3433 isl_space *space;
3435 expr->args = isl_calloc_array(ctx, struct pet_expr *, n);
3436 expr->n_arg = n;
3437 if (!expr->args)
3438 goto error;
3440 space = isl_map_get_space(expr->acc.access);
3441 n = extract_nested(space, 0, expr->args, param2pos);
3442 isl_space_free(space);
3444 if (n < 0)
3445 goto error;
3447 expr->n_arg = n;
3448 return expr;
3449 error:
3450 pet_expr_free(expr);
3451 return NULL;
3454 /* Look for parameters in any access relation in "expr" that
3455 * refer to nested accesses. In particular, these are
3456 * parameters with no name.
3458 * If there are any such parameters, then the domain of the index
3459 * expression and the access relation, which is still [] at this point,
3460 * is replaced by [[] -> [t_1,...,t_n]], with n the number of these parameters
3461 * (after identifying identical nested accesses).
3463 * This transformation is performed in several steps.
3464 * We first extract the arguments in extract_nested.
3465 * param2pos maps the original parameter position to the position
3466 * of the argument.
3467 * Then we move these parameters to input dimensions.
3468 * t2pos maps the positions of these temporary input dimensions
3469 * to the positions of the corresponding arguments.
3470 * Finally, we express these temporary dimensions in terms of the domain
3471 * [[] -> [t_1,...,t_n]] and precompose index expression and access
3472 * relations with this function.
3474 struct pet_expr *PetScan::resolve_nested(struct pet_expr *expr)
3476 int n;
3477 int nparam;
3478 isl_space *space;
3479 isl_local_space *ls;
3480 isl_aff *aff;
3481 isl_multi_aff *ma;
3482 std::map<int,int> param2pos;
3483 std::map<int,int> t2pos;
3485 if (!expr)
3486 return expr;
3488 for (int i = 0; i < expr->n_arg; ++i) {
3489 expr->args[i] = resolve_nested(expr->args[i]);
3490 if (!expr->args[i]) {
3491 pet_expr_free(expr);
3492 return NULL;
3496 if (expr->type != pet_expr_access)
3497 return expr;
3499 n = n_nested_parameter(expr->acc.access);
3500 if (n == 0)
3501 return expr;
3503 expr = extract_nested(expr, n, param2pos);
3504 if (!expr)
3505 return NULL;
3507 expr = pet_expr_access_align_params(expr);
3508 if (!expr)
3509 return NULL;
3510 nparam = isl_map_dim(expr->acc.access, isl_dim_param);
3512 n = 0;
3513 for (int i = nparam - 1; i >= 0; --i) {
3514 isl_id *id = isl_map_get_dim_id(expr->acc.access,
3515 isl_dim_param, i);
3516 if (!is_nested_parameter(id)) {
3517 isl_id_free(id);
3518 continue;
3521 expr->acc.access = isl_map_move_dims(expr->acc.access,
3522 isl_dim_in, n, isl_dim_param, i, 1);
3523 expr->acc.index = isl_multi_pw_aff_move_dims(expr->acc.index,
3524 isl_dim_in, n, isl_dim_param, i, 1);
3525 t2pos[n] = param2pos[i];
3526 n++;
3528 isl_id_free(id);
3531 space = isl_multi_pw_aff_get_space(expr->acc.index);
3532 space = isl_space_set_from_params(isl_space_params(space));
3533 space = isl_space_add_dims(space, isl_dim_set, expr->n_arg);
3534 space = isl_space_wrap(isl_space_from_range(space));
3535 ls = isl_local_space_from_space(isl_space_copy(space));
3536 space = isl_space_from_domain(space);
3537 space = isl_space_add_dims(space, isl_dim_out, n);
3538 ma = isl_multi_aff_zero(space);
3540 for (int i = 0; i < n; ++i) {
3541 aff = isl_aff_var_on_domain(isl_local_space_copy(ls),
3542 isl_dim_set, t2pos[i]);
3543 ma = isl_multi_aff_set_aff(ma, i, aff);
3545 isl_local_space_free(ls);
3547 expr->acc.access = isl_map_preimage_domain_multi_aff(expr->acc.access,
3548 isl_multi_aff_copy(ma));
3549 expr->acc.index = isl_multi_pw_aff_pullback_multi_aff(expr->acc.index,
3550 ma);
3552 return expr;
3555 /* Return the file offset of the expansion location of "Loc".
3557 static unsigned getExpansionOffset(SourceManager &SM, SourceLocation Loc)
3559 return SM.getFileOffset(SM.getExpansionLoc(Loc));
3562 #ifdef HAVE_FINDLOCATIONAFTERTOKEN
3564 /* Return a SourceLocation for the location after the first semicolon
3565 * after "loc". If Lexer::findLocationAfterToken is available, we simply
3566 * call it and also skip trailing spaces and newline.
3568 static SourceLocation location_after_semi(SourceLocation loc, SourceManager &SM,
3569 const LangOptions &LO)
3571 return Lexer::findLocationAfterToken(loc, tok::semi, SM, LO, true);
3574 #else
3576 /* Return a SourceLocation for the location after the first semicolon
3577 * after "loc". If Lexer::findLocationAfterToken is not available,
3578 * we look in the underlying character data for the first semicolon.
3580 static SourceLocation location_after_semi(SourceLocation loc, SourceManager &SM,
3581 const LangOptions &LO)
3583 const char *semi;
3584 const char *s = SM.getCharacterData(loc);
3586 semi = strchr(s, ';');
3587 if (!semi)
3588 return SourceLocation();
3589 return loc.getFileLocWithOffset(semi + 1 - s);
3592 #endif
3594 /* If the token at "loc" is the first token on the line, then return
3595 * a location referring to the start of the line.
3596 * Otherwise, return "loc".
3598 * This function is used to extend a scop to the start of the line
3599 * if the first token of the scop is also the first token on the line.
3601 * We look for the first token on the line. If its location is equal to "loc",
3602 * then the latter is the location of the first token on the line.
3604 static SourceLocation move_to_start_of_line_if_first_token(SourceLocation loc,
3605 SourceManager &SM, const LangOptions &LO)
3607 std::pair<FileID, unsigned> file_offset_pair;
3608 llvm::StringRef file;
3609 const char *pos;
3610 Token tok;
3611 SourceLocation token_loc, line_loc;
3612 int col;
3614 loc = SM.getExpansionLoc(loc);
3615 col = SM.getExpansionColumnNumber(loc);
3616 line_loc = loc.getLocWithOffset(1 - col);
3617 file_offset_pair = SM.getDecomposedLoc(line_loc);
3618 file = SM.getBufferData(file_offset_pair.first, NULL);
3619 pos = file.data() + file_offset_pair.second;
3621 Lexer lexer(SM.getLocForStartOfFile(file_offset_pair.first), LO,
3622 file.begin(), pos, file.end());
3623 lexer.LexFromRawLexer(tok);
3624 token_loc = tok.getLocation();
3626 if (token_loc == loc)
3627 return line_loc;
3628 else
3629 return loc;
3632 /* Update start and end of "scop" to include the region covered by "range".
3633 * If "skip_semi" is set, then we assume "range" is followed by
3634 * a semicolon and also include this semicolon.
3636 struct pet_scop *PetScan::update_scop_start_end(struct pet_scop *scop,
3637 SourceRange range, bool skip_semi)
3639 SourceLocation loc = range.getBegin();
3640 SourceManager &SM = PP.getSourceManager();
3641 const LangOptions &LO = PP.getLangOpts();
3642 unsigned start, end;
3644 loc = move_to_start_of_line_if_first_token(loc, SM, LO);
3645 start = getExpansionOffset(SM, loc);
3646 loc = range.getEnd();
3647 if (skip_semi)
3648 loc = location_after_semi(loc, SM, LO);
3649 else
3650 loc = PP.getLocForEndOfToken(loc);
3651 end = getExpansionOffset(SM, loc);
3653 scop = pet_scop_update_start_end(scop, start, end);
3654 return scop;
3657 /* Convert a top-level pet_expr to a pet_scop with one statement.
3658 * This mainly involves resolving nested expression parameters
3659 * and setting the name of the iteration space.
3660 * The name is given by "label" if it is non-NULL. Otherwise,
3661 * it is of the form S_<n_stmt>.
3662 * start and end of the pet_scop are derived from those of "stmt".
3663 * If "stmt" is an expression statement, then its range does not
3664 * include the semicolon, while it should be included in the pet_scop.
3666 struct pet_scop *PetScan::extract(Stmt *stmt, struct pet_expr *expr,
3667 __isl_take isl_id *label)
3669 struct pet_stmt *ps;
3670 struct pet_scop *scop;
3671 SourceLocation loc = stmt->getLocStart();
3672 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3673 bool skip_semi;
3675 expr = resolve_nested(expr);
3676 ps = pet_stmt_from_pet_expr(ctx, line, label, n_stmt++, expr);
3677 scop = pet_scop_from_pet_stmt(ctx, ps);
3679 skip_semi = isa<Expr>(stmt);
3680 scop = update_scop_start_end(scop, stmt->getSourceRange(), skip_semi);
3681 return scop;
3684 /* Check if we can extract an affine expression from "expr".
3685 * Return the expressions as an isl_pw_aff if we can and NULL otherwise.
3686 * We turn on autodetection so that we won't generate any warnings
3687 * and turn off nesting, so that we won't accept any non-affine constructs.
3689 __isl_give isl_pw_aff *PetScan::try_extract_affine(Expr *expr)
3691 isl_pw_aff *pwaff;
3692 int save_autodetect = options->autodetect;
3693 bool save_nesting = nesting_enabled;
3695 options->autodetect = 1;
3696 nesting_enabled = false;
3698 pwaff = extract_affine(expr);
3700 options->autodetect = save_autodetect;
3701 nesting_enabled = save_nesting;
3703 return pwaff;
3706 /* Check if we can extract an affine constraint from "expr".
3707 * Return the constraint as an isl_set if we can and NULL otherwise.
3708 * We turn on autodetection so that we won't generate any warnings
3709 * and turn off nesting, so that we won't accept any non-affine constructs.
3711 __isl_give isl_pw_aff *PetScan::try_extract_affine_condition(Expr *expr)
3713 isl_pw_aff *cond;
3714 int save_autodetect = options->autodetect;
3715 bool save_nesting = nesting_enabled;
3717 options->autodetect = 1;
3718 nesting_enabled = false;
3720 cond = extract_condition(expr);
3722 options->autodetect = save_autodetect;
3723 nesting_enabled = save_nesting;
3725 return cond;
3728 /* Check whether "expr" is an affine constraint.
3730 bool PetScan::is_affine_condition(Expr *expr)
3732 isl_pw_aff *cond;
3734 cond = try_extract_affine_condition(expr);
3735 isl_pw_aff_free(cond);
3737 return cond != NULL;
3740 /* Check if we can extract a condition from "expr".
3741 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
3742 * If allow_nested is set, then the condition may involve parameters
3743 * corresponding to nested accesses.
3744 * We turn on autodetection so that we won't generate any warnings.
3746 __isl_give isl_pw_aff *PetScan::try_extract_nested_condition(Expr *expr)
3748 isl_pw_aff *cond;
3749 int save_autodetect = options->autodetect;
3750 bool save_nesting = nesting_enabled;
3752 options->autodetect = 1;
3753 nesting_enabled = allow_nested;
3754 cond = extract_condition(expr);
3756 options->autodetect = save_autodetect;
3757 nesting_enabled = save_nesting;
3759 return cond;
3762 /* If the top-level expression of "stmt" is an assignment, then
3763 * return that assignment as a BinaryOperator.
3764 * Otherwise return NULL.
3766 static BinaryOperator *top_assignment_or_null(Stmt *stmt)
3768 BinaryOperator *ass;
3770 if (!stmt)
3771 return NULL;
3772 if (stmt->getStmtClass() != Stmt::BinaryOperatorClass)
3773 return NULL;
3775 ass = cast<BinaryOperator>(stmt);
3776 if(ass->getOpcode() != BO_Assign)
3777 return NULL;
3779 return ass;
3782 /* Check if the given if statement is a conditional assignement
3783 * with a non-affine condition. If so, construct a pet_scop
3784 * corresponding to this conditional assignment. Otherwise return NULL.
3786 * In particular we check if "stmt" is of the form
3788 * if (condition)
3789 * a = f(...);
3790 * else
3791 * a = g(...);
3793 * where a is some array or scalar access.
3794 * The constructed pet_scop then corresponds to the expression
3796 * a = condition ? f(...) : g(...)
3798 * All access relations in f(...) are intersected with condition
3799 * while all access relation in g(...) are intersected with the complement.
3801 struct pet_scop *PetScan::extract_conditional_assignment(IfStmt *stmt)
3803 BinaryOperator *ass_then, *ass_else;
3804 isl_multi_pw_aff *write_then, *write_else;
3805 isl_set *cond, *comp;
3806 isl_multi_pw_aff *index;
3807 isl_pw_aff *pa;
3808 int equal;
3809 struct pet_expr *pe_cond, *pe_then, *pe_else, *pe, *pe_write;
3810 bool save_nesting = nesting_enabled;
3812 if (!options->detect_conditional_assignment)
3813 return NULL;
3815 ass_then = top_assignment_or_null(stmt->getThen());
3816 ass_else = top_assignment_or_null(stmt->getElse());
3818 if (!ass_then || !ass_else)
3819 return NULL;
3821 if (is_affine_condition(stmt->getCond()))
3822 return NULL;
3824 write_then = extract_index(ass_then->getLHS());
3825 write_else = extract_index(ass_else->getLHS());
3827 equal = isl_multi_pw_aff_plain_is_equal(write_then, write_else);
3828 isl_multi_pw_aff_free(write_else);
3829 if (equal < 0 || !equal) {
3830 isl_multi_pw_aff_free(write_then);
3831 return NULL;
3834 nesting_enabled = allow_nested;
3835 pa = extract_condition(stmt->getCond());
3836 nesting_enabled = save_nesting;
3837 cond = isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa));
3838 comp = isl_pw_aff_zero_set(isl_pw_aff_copy(pa));
3839 index = isl_multi_pw_aff_from_range(isl_multi_pw_aff_from_pw_aff(pa));
3841 pe_cond = pet_expr_from_index(index);
3843 pe_then = extract_expr(ass_then->getRHS());
3844 pe_then = pet_expr_restrict(pe_then, cond);
3845 pe_else = extract_expr(ass_else->getRHS());
3846 pe_else = pet_expr_restrict(pe_else, comp);
3848 pe = pet_expr_new_ternary(ctx, pe_cond, pe_then, pe_else);
3849 pe_write = pet_expr_from_index_and_depth(write_then,
3850 extract_depth(write_then));
3851 if (pe_write) {
3852 pe_write->acc.write = 1;
3853 pe_write->acc.read = 0;
3855 pe = pet_expr_new_binary(ctx, pet_op_assign, pe_write, pe);
3856 return extract(stmt, pe);
3859 /* Create a pet_scop with a single statement with name S_<stmt_nr>,
3860 * evaluating "cond" and writing the result to a virtual scalar,
3861 * as expressed by "index".
3863 struct pet_scop *PetScan::extract_non_affine_condition(Expr *cond, int stmt_nr,
3864 __isl_take isl_multi_pw_aff *index)
3866 struct pet_expr *expr, *write;
3867 struct pet_stmt *ps;
3868 SourceLocation loc = cond->getLocStart();
3869 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3871 write = pet_expr_from_index(index);
3872 if (write) {
3873 write->acc.write = 1;
3874 write->acc.read = 0;
3876 expr = extract_expr(cond);
3877 expr = resolve_nested(expr);
3878 expr = pet_expr_new_binary(ctx, pet_op_assign, write, expr);
3879 ps = pet_stmt_from_pet_expr(ctx, line, NULL, stmt_nr, expr);
3880 return pet_scop_from_pet_stmt(ctx, ps);
3883 extern "C" {
3884 static struct pet_expr *embed_access(struct pet_expr *expr, void *user);
3887 /* Precompose the access relation and the index expression associated
3888 * to "expr" with the function pointed to by "user",
3889 * thereby embedding the access relation in the domain of this function.
3890 * The initial domain of the access relation and the index expression
3891 * is the zero-dimensional domain.
3893 static struct pet_expr *embed_access(struct pet_expr *expr, void *user)
3895 isl_multi_aff *ma = (isl_multi_aff *) user;
3897 expr->acc.access = isl_map_preimage_domain_multi_aff(expr->acc.access,
3898 isl_multi_aff_copy(ma));
3899 expr->acc.index = isl_multi_pw_aff_pullback_multi_aff(expr->acc.index,
3900 isl_multi_aff_copy(ma));
3901 if (!expr->acc.access || !expr->acc.index)
3902 goto error;
3904 return expr;
3905 error:
3906 pet_expr_free(expr);
3907 return NULL;
3910 /* Precompose all access relations in "expr" with "ma", thereby
3911 * embedding them in the domain of "ma".
3913 static struct pet_expr *embed(struct pet_expr *expr,
3914 __isl_keep isl_multi_aff *ma)
3916 return pet_expr_map_access(expr, &embed_access, ma);
3919 /* How many parameters of "set" refer to nested accesses, i.e., have no name?
3921 static int n_nested_parameter(__isl_keep isl_set *set)
3923 isl_space *space;
3924 int n;
3926 space = isl_set_get_space(set);
3927 n = n_nested_parameter(space);
3928 isl_space_free(space);
3930 return n;
3933 /* Remove all parameters from "map" that refer to nested accesses.
3935 static __isl_give isl_map *remove_nested_parameters(__isl_take isl_map *map)
3937 int nparam;
3938 isl_space *space;
3940 space = isl_map_get_space(map);
3941 nparam = isl_space_dim(space, isl_dim_param);
3942 for (int i = nparam - 1; i >= 0; --i)
3943 if (is_nested_parameter(space, i))
3944 map = isl_map_project_out(map, isl_dim_param, i, 1);
3945 isl_space_free(space);
3947 return map;
3950 /* Remove all parameters from "mpa" that refer to nested accesses.
3952 static __isl_give isl_multi_pw_aff *remove_nested_parameters(
3953 __isl_take isl_multi_pw_aff *mpa)
3955 int nparam;
3956 isl_space *space;
3958 space = isl_multi_pw_aff_get_space(mpa);
3959 nparam = isl_space_dim(space, isl_dim_param);
3960 for (int i = nparam - 1; i >= 0; --i) {
3961 if (!is_nested_parameter(space, i))
3962 continue;
3963 mpa = isl_multi_pw_aff_drop_dims(mpa, isl_dim_param, i, 1);
3965 isl_space_free(space);
3967 return mpa;
3970 /* Remove all parameters from the index expression and access relation of "expr"
3971 * that refer to nested accesses.
3973 static struct pet_expr *remove_nested_parameters(struct pet_expr *expr)
3975 expr->acc.access = remove_nested_parameters(expr->acc.access);
3976 expr->acc.index = remove_nested_parameters(expr->acc.index);
3977 if (!expr->acc.access || !expr->acc.index)
3978 goto error;
3980 return expr;
3981 error:
3982 pet_expr_free(expr);
3983 return NULL;
3986 extern "C" {
3987 static struct pet_expr *expr_remove_nested_parameters(
3988 struct pet_expr *expr, void *user);
3991 static struct pet_expr *expr_remove_nested_parameters(
3992 struct pet_expr *expr, void *user)
3994 return remove_nested_parameters(expr);
3997 /* Remove all nested access parameters from the schedule and all
3998 * accesses of "stmt".
3999 * There is no need to remove them from the domain as these parameters
4000 * have already been removed from the domain when this function is called.
4002 static struct pet_stmt *remove_nested_parameters(struct pet_stmt *stmt)
4004 if (!stmt)
4005 return NULL;
4006 stmt->schedule = remove_nested_parameters(stmt->schedule);
4007 stmt->body = pet_expr_map_access(stmt->body,
4008 &expr_remove_nested_parameters, NULL);
4009 if (!stmt->schedule || !stmt->body)
4010 goto error;
4011 for (int i = 0; i < stmt->n_arg; ++i) {
4012 stmt->args[i] = pet_expr_map_access(stmt->args[i],
4013 &expr_remove_nested_parameters, NULL);
4014 if (!stmt->args[i])
4015 goto error;
4018 return stmt;
4019 error:
4020 pet_stmt_free(stmt);
4021 return NULL;
4024 /* For each nested access parameter in the domain of "stmt",
4025 * construct a corresponding pet_expr, place it before the original
4026 * elements in stmt->args and record its position in "param2pos".
4027 * n is the number of nested access parameters.
4029 struct pet_stmt *PetScan::extract_nested(struct pet_stmt *stmt, int n,
4030 std::map<int,int> &param2pos)
4032 int i;
4033 isl_space *space;
4034 int n_arg;
4035 struct pet_expr **args;
4037 n_arg = stmt->n_arg;
4038 args = isl_calloc_array(ctx, struct pet_expr *, n + n_arg);
4039 if (!args)
4040 goto error;
4042 space = isl_set_get_space(stmt->domain);
4043 n_arg = extract_nested(space, 0, args, param2pos);
4044 isl_space_free(space);
4046 if (n_arg < 0)
4047 goto error;
4049 for (i = 0; i < stmt->n_arg; ++i)
4050 args[n_arg + i] = stmt->args[i];
4051 free(stmt->args);
4052 stmt->args = args;
4053 stmt->n_arg += n_arg;
4055 return stmt;
4056 error:
4057 if (args) {
4058 for (i = 0; i < n; ++i)
4059 pet_expr_free(args[i]);
4060 free(args);
4062 pet_stmt_free(stmt);
4063 return NULL;
4066 /* Check whether any of the arguments i of "stmt" starting at position "n"
4067 * is equal to one of the first "n" arguments j.
4068 * If so, combine the constraints on arguments i and j and remove
4069 * argument i.
4071 static struct pet_stmt *remove_duplicate_arguments(struct pet_stmt *stmt, int n)
4073 int i, j;
4074 isl_map *map;
4076 if (!stmt)
4077 return NULL;
4078 if (n == 0)
4079 return stmt;
4080 if (n == stmt->n_arg)
4081 return stmt;
4083 map = isl_set_unwrap(stmt->domain);
4085 for (i = stmt->n_arg - 1; i >= n; --i) {
4086 for (j = 0; j < n; ++j)
4087 if (pet_expr_is_equal(stmt->args[i], stmt->args[j]))
4088 break;
4089 if (j >= n)
4090 continue;
4092 map = isl_map_equate(map, isl_dim_out, i, isl_dim_out, j);
4093 map = isl_map_project_out(map, isl_dim_out, i, 1);
4095 pet_expr_free(stmt->args[i]);
4096 for (j = i; j + 1 < stmt->n_arg; ++j)
4097 stmt->args[j] = stmt->args[j + 1];
4098 stmt->n_arg--;
4101 stmt->domain = isl_map_wrap(map);
4102 if (!stmt->domain)
4103 goto error;
4104 return stmt;
4105 error:
4106 pet_stmt_free(stmt);
4107 return NULL;
4110 /* Look for parameters in the iteration domain of "stmt" that
4111 * refer to nested accesses. In particular, these are
4112 * parameters with no name.
4114 * If there are any such parameters, then as many extra variables
4115 * (after identifying identical nested accesses) are inserted in the
4116 * range of the map wrapped inside the domain, before the original variables.
4117 * If the original domain is not a wrapped map, then a new wrapped
4118 * map is created with zero output dimensions.
4119 * The parameters are then equated to the corresponding output dimensions
4120 * and subsequently projected out, from the iteration domain,
4121 * the schedule and the access relations.
4122 * For each of the output dimensions, a corresponding argument
4123 * expression is inserted. Initially they are created with
4124 * a zero-dimensional domain, so they have to be embedded
4125 * in the current iteration domain.
4126 * param2pos maps the position of the parameter to the position
4127 * of the corresponding output dimension in the wrapped map.
4129 struct pet_stmt *PetScan::resolve_nested(struct pet_stmt *stmt)
4131 int n;
4132 int nparam;
4133 unsigned n_arg;
4134 isl_map *map;
4135 isl_space *space;
4136 isl_multi_aff *ma;
4137 std::map<int,int> param2pos;
4139 if (!stmt)
4140 return NULL;
4142 n = n_nested_parameter(stmt->domain);
4143 if (n == 0)
4144 return stmt;
4146 n_arg = stmt->n_arg;
4147 stmt = extract_nested(stmt, n, param2pos);
4148 if (!stmt)
4149 return NULL;
4151 n = stmt->n_arg - n_arg;
4152 nparam = isl_set_dim(stmt->domain, isl_dim_param);
4153 if (isl_set_is_wrapping(stmt->domain))
4154 map = isl_set_unwrap(stmt->domain);
4155 else
4156 map = isl_map_from_domain(stmt->domain);
4157 map = isl_map_insert_dims(map, isl_dim_out, 0, n);
4159 for (int i = nparam - 1; i >= 0; --i) {
4160 isl_id *id;
4162 if (!is_nested_parameter(map, i))
4163 continue;
4165 id = pet_expr_access_get_id(stmt->args[param2pos[i]]);
4166 map = isl_map_set_dim_id(map, isl_dim_out, param2pos[i], id);
4167 map = isl_map_equate(map, isl_dim_param, i, isl_dim_out,
4168 param2pos[i]);
4169 map = isl_map_project_out(map, isl_dim_param, i, 1);
4172 stmt->domain = isl_map_wrap(map);
4174 space = isl_space_unwrap(isl_set_get_space(stmt->domain));
4175 space = isl_space_from_domain(isl_space_domain(space));
4176 ma = isl_multi_aff_zero(space);
4177 for (int pos = 0; pos < n; ++pos)
4178 stmt->args[pos] = embed(stmt->args[pos], ma);
4179 isl_multi_aff_free(ma);
4181 stmt = remove_nested_parameters(stmt);
4182 stmt = remove_duplicate_arguments(stmt, n);
4184 return stmt;
4187 /* For each statement in "scop", move the parameters that correspond
4188 * to nested access into the ranges of the domains and create
4189 * corresponding argument expressions.
4191 struct pet_scop *PetScan::resolve_nested(struct pet_scop *scop)
4193 if (!scop)
4194 return NULL;
4196 for (int i = 0; i < scop->n_stmt; ++i) {
4197 scop->stmts[i] = resolve_nested(scop->stmts[i]);
4198 if (!scop->stmts[i])
4199 goto error;
4202 return scop;
4203 error:
4204 pet_scop_free(scop);
4205 return NULL;
4208 /* Given an access expression "expr", is the variable accessed by
4209 * "expr" assigned anywhere inside "scop"?
4211 static bool is_assigned(pet_expr *expr, pet_scop *scop)
4213 bool assigned = false;
4214 isl_id *id;
4216 id = pet_expr_access_get_id(expr);
4217 assigned = pet_scop_writes(scop, id);
4218 isl_id_free(id);
4220 return assigned;
4223 /* Are all nested access parameters in "pa" allowed given "scop".
4224 * In particular, is none of them written by anywhere inside "scop".
4226 * If "scop" has any skip conditions, then no nested access parameters
4227 * are allowed. In particular, if there is any nested access in a guard
4228 * for a piece of code containing a "continue", then we want to introduce
4229 * a separate statement for evaluating this guard so that we can express
4230 * that the result is false for all previous iterations.
4232 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff *pa, pet_scop *scop)
4234 int nparam;
4236 if (!scop)
4237 return true;
4239 if (!has_nested(pa))
4240 return true;
4242 if (pet_scop_has_skip(scop, pet_skip_now))
4243 return false;
4245 nparam = isl_pw_aff_dim(pa, isl_dim_param);
4246 for (int i = 0; i < nparam; ++i) {
4247 Expr *nested;
4248 isl_id *id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
4249 pet_expr *expr;
4250 bool allowed;
4252 if (!is_nested_parameter(id)) {
4253 isl_id_free(id);
4254 continue;
4257 nested = (Expr *) isl_id_get_user(id);
4258 expr = extract_expr(nested);
4259 allowed = expr && expr->type == pet_expr_access &&
4260 !is_assigned(expr, scop);
4262 pet_expr_free(expr);
4263 isl_id_free(id);
4265 if (!allowed)
4266 return false;
4269 return true;
4272 /* Do we need to construct a skip condition of the given type
4273 * on an if statement, given that the if condition is non-affine?
4275 * pet_scop_filter_skip can only handle the case where the if condition
4276 * holds (the then branch) and the skip condition is universal.
4277 * In any other case, we need to construct a new skip condition.
4279 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
4280 bool have_else, enum pet_skip type)
4282 if (have_else && scop_else && pet_scop_has_skip(scop_else, type))
4283 return true;
4284 if (scop_then && pet_scop_has_skip(scop_then, type) &&
4285 !pet_scop_has_universal_skip(scop_then, type))
4286 return true;
4287 return false;
4290 /* Do we need to construct a skip condition of the given type
4291 * on an if statement, given that the if condition is affine?
4293 * There is no need to construct a new skip condition if all
4294 * the skip conditions are affine.
4296 static bool need_skip_aff(struct pet_scop *scop_then,
4297 struct pet_scop *scop_else, bool have_else, enum pet_skip type)
4299 if (scop_then && pet_scop_has_var_skip(scop_then, type))
4300 return true;
4301 if (have_else && scop_else && pet_scop_has_var_skip(scop_else, type))
4302 return true;
4303 return false;
4306 /* Do we need to construct a skip condition of the given type
4307 * on an if statement?
4309 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
4310 bool have_else, enum pet_skip type, bool affine)
4312 if (affine)
4313 return need_skip_aff(scop_then, scop_else, have_else, type);
4314 else
4315 return need_skip(scop_then, scop_else, have_else, type);
4318 /* Construct an affine expression pet_expr that evaluates
4319 * to the constant "val".
4321 static struct pet_expr *universally(isl_ctx *ctx, int val)
4323 isl_local_space *ls;
4324 isl_aff *aff;
4325 isl_multi_pw_aff *mpa;
4327 ls = isl_local_space_from_space(isl_space_set_alloc(ctx, 0, 0));
4328 aff = isl_aff_val_on_domain(ls, isl_val_int_from_si(ctx, val));
4329 mpa = isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
4331 return pet_expr_from_index(mpa);
4334 /* Construct an affine expression pet_expr that evaluates
4335 * to the constant 1.
4337 static struct pet_expr *universally_true(isl_ctx *ctx)
4339 return universally(ctx, 1);
4342 /* Construct an affine expression pet_expr that evaluates
4343 * to the constant 0.
4345 static struct pet_expr *universally_false(isl_ctx *ctx)
4347 return universally(ctx, 0);
4350 /* Given an index expression "test_index" for the if condition,
4351 * an index expression "skip_index" for the skip condition and
4352 * scops for the then and else branches, construct a scop for
4353 * computing "skip_index".
4355 * The computed scop contains a single statement that essentially does
4357 * skip_index = test_cond ? skip_cond_then : skip_cond_else
4359 * If the skip conditions of the then and/or else branch are not affine,
4360 * then they need to be filtered by test_index.
4361 * If they are missing, then this means the skip condition is false.
4363 * Since we are constructing a skip condition for the if statement,
4364 * the skip conditions on the then and else branches are removed.
4366 static struct pet_scop *extract_skip(PetScan *scan,
4367 __isl_take isl_multi_pw_aff *test_index,
4368 __isl_take isl_multi_pw_aff *skip_index,
4369 struct pet_scop *scop_then, struct pet_scop *scop_else, bool have_else,
4370 enum pet_skip type)
4372 struct pet_expr *expr_then, *expr_else, *expr, *expr_skip;
4373 struct pet_stmt *stmt;
4374 struct pet_scop *scop;
4375 isl_ctx *ctx = scan->ctx;
4377 if (!scop_then)
4378 goto error;
4379 if (have_else && !scop_else)
4380 goto error;
4382 if (pet_scop_has_skip(scop_then, type)) {
4383 expr_then = pet_scop_get_skip_expr(scop_then, type);
4384 pet_scop_reset_skip(scop_then, type);
4385 if (!pet_expr_is_affine(expr_then))
4386 expr_then = pet_expr_filter(expr_then,
4387 isl_multi_pw_aff_copy(test_index), 1);
4388 } else
4389 expr_then = universally_false(ctx);
4391 if (have_else && pet_scop_has_skip(scop_else, type)) {
4392 expr_else = pet_scop_get_skip_expr(scop_else, type);
4393 pet_scop_reset_skip(scop_else, type);
4394 if (!pet_expr_is_affine(expr_else))
4395 expr_else = pet_expr_filter(expr_else,
4396 isl_multi_pw_aff_copy(test_index), 0);
4397 } else
4398 expr_else = universally_false(ctx);
4400 expr = pet_expr_from_index(test_index);
4401 expr = pet_expr_new_ternary(ctx, expr, expr_then, expr_else);
4402 expr_skip = pet_expr_from_index(isl_multi_pw_aff_copy(skip_index));
4403 if (expr_skip) {
4404 expr_skip->acc.write = 1;
4405 expr_skip->acc.read = 0;
4407 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
4408 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, scan->n_stmt++, expr);
4410 scop = pet_scop_from_pet_stmt(ctx, stmt);
4411 scop = scop_add_array(scop, skip_index, scan->ast_context);
4412 isl_multi_pw_aff_free(skip_index);
4414 return scop;
4415 error:
4416 isl_multi_pw_aff_free(test_index);
4417 isl_multi_pw_aff_free(skip_index);
4418 return NULL;
4421 /* Is scop's skip_now condition equal to its skip_later condition?
4422 * In particular, this means that it either has no skip_now condition
4423 * or both a skip_now and a skip_later condition (that are equal to each other).
4425 static bool skip_equals_skip_later(struct pet_scop *scop)
4427 int has_skip_now, has_skip_later;
4428 int equal;
4429 isl_multi_pw_aff *skip_now, *skip_later;
4431 if (!scop)
4432 return false;
4433 has_skip_now = pet_scop_has_skip(scop, pet_skip_now);
4434 has_skip_later = pet_scop_has_skip(scop, pet_skip_later);
4435 if (has_skip_now != has_skip_later)
4436 return false;
4437 if (!has_skip_now)
4438 return true;
4440 skip_now = pet_scop_get_skip(scop, pet_skip_now);
4441 skip_later = pet_scop_get_skip(scop, pet_skip_later);
4442 equal = isl_multi_pw_aff_is_equal(skip_now, skip_later);
4443 isl_multi_pw_aff_free(skip_now);
4444 isl_multi_pw_aff_free(skip_later);
4446 return equal;
4449 /* Drop the skip conditions of type pet_skip_later from scop1 and scop2.
4451 static void drop_skip_later(struct pet_scop *scop1, struct pet_scop *scop2)
4453 pet_scop_reset_skip(scop1, pet_skip_later);
4454 pet_scop_reset_skip(scop2, pet_skip_later);
4457 /* Structure that handles the construction of skip conditions.
4459 * scop_then and scop_else represent the then and else branches
4460 * of the if statement
4462 * skip[type] is true if we need to construct a skip condition of that type
4463 * equal is set if the skip conditions of types pet_skip_now and pet_skip_later
4464 * are equal to each other
4465 * index[type] is an index expression from a zero-dimension domain
4466 * to the virtual array representing the skip condition
4467 * scop[type] is a scop for computing the skip condition
4469 struct pet_skip_info {
4470 isl_ctx *ctx;
4472 bool skip[2];
4473 bool equal;
4474 isl_multi_pw_aff *index[2];
4475 struct pet_scop *scop[2];
4477 pet_skip_info(isl_ctx *ctx) : ctx(ctx) {}
4479 operator bool() { return skip[pet_skip_now] || skip[pet_skip_later]; }
4482 /* Structure that handles the construction of skip conditions on if statements.
4484 * scop_then and scop_else represent the then and else branches
4485 * of the if statement
4487 struct pet_skip_info_if : public pet_skip_info {
4488 struct pet_scop *scop_then, *scop_else;
4489 bool have_else;
4491 pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
4492 struct pet_scop *scop_else, bool have_else, bool affine);
4493 void extract(PetScan *scan, __isl_keep isl_multi_pw_aff *index,
4494 enum pet_skip type);
4495 void extract(PetScan *scan, __isl_keep isl_multi_pw_aff *index);
4496 void extract(PetScan *scan, __isl_keep isl_pw_aff *cond);
4497 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
4498 int offset);
4499 struct pet_scop *add(struct pet_scop *scop, int offset);
4502 /* Initialize a pet_skip_info_if structure based on the then and else branches
4503 * and based on whether the if condition is affine or not.
4505 pet_skip_info_if::pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
4506 struct pet_scop *scop_else, bool have_else, bool affine) :
4507 pet_skip_info(ctx), scop_then(scop_then), scop_else(scop_else),
4508 have_else(have_else)
4510 skip[pet_skip_now] =
4511 need_skip(scop_then, scop_else, have_else, pet_skip_now, affine);
4512 equal = skip[pet_skip_now] && skip_equals_skip_later(scop_then) &&
4513 (!have_else || skip_equals_skip_later(scop_else));
4514 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
4515 need_skip(scop_then, scop_else, have_else, pet_skip_later, affine);
4518 /* If we need to construct a skip condition of the given type,
4519 * then do so now.
4521 * "mpa" represents the if condition.
4523 void pet_skip_info_if::extract(PetScan *scan,
4524 __isl_keep isl_multi_pw_aff *mpa, enum pet_skip type)
4526 isl_ctx *ctx;
4528 if (!skip[type])
4529 return;
4531 ctx = isl_multi_pw_aff_get_ctx(mpa);
4532 index[type] = create_test_index(ctx, scan->n_test++);
4533 scop[type] = extract_skip(scan, isl_multi_pw_aff_copy(mpa),
4534 isl_multi_pw_aff_copy(index[type]),
4535 scop_then, scop_else, have_else, type);
4538 /* Construct the required skip conditions, given the if condition "index".
4540 void pet_skip_info_if::extract(PetScan *scan,
4541 __isl_keep isl_multi_pw_aff *index)
4543 extract(scan, index, pet_skip_now);
4544 extract(scan, index, pet_skip_later);
4545 if (equal)
4546 drop_skip_later(scop_then, scop_else);
4549 /* Construct the required skip conditions, given the if condition "cond".
4551 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_pw_aff *cond)
4553 isl_multi_pw_aff *test;
4555 if (!skip[pet_skip_now] && !skip[pet_skip_later])
4556 return;
4558 test = isl_multi_pw_aff_from_pw_aff(isl_pw_aff_copy(cond));
4559 test = isl_multi_pw_aff_from_range(test);
4560 extract(scan, test);
4561 isl_multi_pw_aff_free(test);
4564 /* Add the computed skip condition of the give type to "main" and
4565 * add the scop for computing the condition at the given offset.
4567 * If equal is set, then we only computed a skip condition for pet_skip_now,
4568 * but we also need to set it as main's pet_skip_later.
4570 struct pet_scop *pet_skip_info_if::add(struct pet_scop *main,
4571 enum pet_skip type, int offset)
4573 if (!skip[type])
4574 return main;
4576 scop[type] = pet_scop_prefix(scop[type], offset);
4577 main = pet_scop_add_par(ctx, main, scop[type]);
4578 scop[type] = NULL;
4580 if (equal)
4581 main = pet_scop_set_skip(main, pet_skip_later,
4582 isl_multi_pw_aff_copy(index[type]));
4584 main = pet_scop_set_skip(main, type, index[type]);
4585 index[type] = NULL;
4587 return main;
4590 /* Add the computed skip conditions to "main" and
4591 * add the scops for computing the conditions at the given offset.
4593 struct pet_scop *pet_skip_info_if::add(struct pet_scop *scop, int offset)
4595 scop = add(scop, pet_skip_now, offset);
4596 scop = add(scop, pet_skip_later, offset);
4598 return scop;
4601 /* Construct a pet_scop for a non-affine if statement.
4603 * We create a separate statement that writes the result
4604 * of the non-affine condition to a virtual scalar.
4605 * A constraint requiring the value of this virtual scalar to be one
4606 * is added to the iteration domains of the then branch.
4607 * Similarly, a constraint requiring the value of this virtual scalar
4608 * to be zero is added to the iteration domains of the else branch, if any.
4609 * We adjust the schedules to ensure that the virtual scalar is written
4610 * before it is read.
4612 * If there are any breaks or continues in the then and/or else
4613 * branches, then we may have to compute a new skip condition.
4614 * This is handled using a pet_skip_info_if object.
4615 * On initialization, the object checks if skip conditions need
4616 * to be computed. If so, it does so in "extract" and adds them in "add".
4618 struct pet_scop *PetScan::extract_non_affine_if(Expr *cond,
4619 struct pet_scop *scop_then, struct pet_scop *scop_else,
4620 bool have_else, int stmt_id)
4622 struct pet_scop *scop;
4623 isl_multi_pw_aff *test_index;
4624 int save_n_stmt = n_stmt;
4626 test_index = create_test_index(ctx, n_test++);
4627 n_stmt = stmt_id;
4628 scop = extract_non_affine_condition(cond, n_stmt++,
4629 isl_multi_pw_aff_copy(test_index));
4630 n_stmt = save_n_stmt;
4631 scop = scop_add_array(scop, test_index, ast_context);
4633 pet_skip_info_if skip(ctx, scop_then, scop_else, have_else, false);
4634 skip.extract(this, test_index);
4636 scop = pet_scop_prefix(scop, 0);
4637 scop_then = pet_scop_prefix(scop_then, 1);
4638 scop_then = pet_scop_filter(scop_then,
4639 isl_multi_pw_aff_copy(test_index), 1);
4640 if (have_else) {
4641 scop_else = pet_scop_prefix(scop_else, 1);
4642 scop_else = pet_scop_filter(scop_else, test_index, 0);
4643 scop_then = pet_scop_add_par(ctx, scop_then, scop_else);
4644 } else
4645 isl_multi_pw_aff_free(test_index);
4647 scop = pet_scop_add_seq(ctx, scop, scop_then);
4649 scop = skip.add(scop, 2);
4651 return scop;
4654 /* Construct a pet_scop for an if statement.
4656 * If the condition fits the pattern of a conditional assignment,
4657 * then it is handled by extract_conditional_assignment.
4658 * Otherwise, we do the following.
4660 * If the condition is affine, then the condition is added
4661 * to the iteration domains of the then branch, while the
4662 * opposite of the condition in added to the iteration domains
4663 * of the else branch, if any.
4664 * We allow the condition to be dynamic, i.e., to refer to
4665 * scalars or array elements that may be written to outside
4666 * of the given if statement. These nested accesses are then represented
4667 * as output dimensions in the wrapping iteration domain.
4668 * If it is also written _inside_ the then or else branch, then
4669 * we treat the condition as non-affine.
4670 * As explained in extract_non_affine_if, this will introduce
4671 * an extra statement.
4672 * For aesthetic reasons, we want this statement to have a statement
4673 * number that is lower than those of the then and else branches.
4674 * In order to evaluate if we will need such a statement, however, we
4675 * first construct scops for the then and else branches.
4676 * We therefore reserve a statement number if we might have to
4677 * introduce such an extra statement.
4679 * If the condition is not affine, then the scop is created in
4680 * extract_non_affine_if.
4682 * If there are any breaks or continues in the then and/or else
4683 * branches, then we may have to compute a new skip condition.
4684 * This is handled using a pet_skip_info_if object.
4685 * On initialization, the object checks if skip conditions need
4686 * to be computed. If so, it does so in "extract" and adds them in "add".
4688 struct pet_scop *PetScan::extract(IfStmt *stmt)
4690 struct pet_scop *scop_then, *scop_else = NULL, *scop;
4691 isl_pw_aff *cond;
4692 int stmt_id;
4693 isl_set *set;
4694 isl_set *valid;
4696 clear_assignments clear(assigned_value);
4697 clear.TraverseStmt(stmt->getThen());
4698 if (stmt->getElse())
4699 clear.TraverseStmt(stmt->getElse());
4701 scop = extract_conditional_assignment(stmt);
4702 if (scop)
4703 return scop;
4705 cond = try_extract_nested_condition(stmt->getCond());
4706 if (allow_nested && (!cond || has_nested(cond)))
4707 stmt_id = n_stmt++;
4710 assigned_value_cache cache(assigned_value);
4711 scop_then = extract(stmt->getThen());
4714 if (stmt->getElse()) {
4715 assigned_value_cache cache(assigned_value);
4716 scop_else = extract(stmt->getElse());
4717 if (options->autodetect) {
4718 if (scop_then && !scop_else) {
4719 partial = true;
4720 isl_pw_aff_free(cond);
4721 return scop_then;
4723 if (!scop_then && scop_else) {
4724 partial = true;
4725 isl_pw_aff_free(cond);
4726 return scop_else;
4731 if (cond &&
4732 (!is_nested_allowed(cond, scop_then) ||
4733 (stmt->getElse() && !is_nested_allowed(cond, scop_else)))) {
4734 isl_pw_aff_free(cond);
4735 cond = NULL;
4737 if (allow_nested && !cond)
4738 return extract_non_affine_if(stmt->getCond(), scop_then,
4739 scop_else, stmt->getElse(), stmt_id);
4741 if (!cond)
4742 cond = extract_condition(stmt->getCond());
4744 pet_skip_info_if skip(ctx, scop_then, scop_else, stmt->getElse(), true);
4745 skip.extract(this, cond);
4747 valid = isl_pw_aff_domain(isl_pw_aff_copy(cond));
4748 set = isl_pw_aff_non_zero_set(cond);
4749 scop = pet_scop_restrict(scop_then, isl_set_copy(set));
4751 if (stmt->getElse()) {
4752 set = isl_set_subtract(isl_set_copy(valid), set);
4753 scop_else = pet_scop_restrict(scop_else, set);
4754 scop = pet_scop_add_par(ctx, scop, scop_else);
4755 } else
4756 isl_set_free(set);
4757 scop = resolve_nested(scop);
4758 scop = pet_scop_restrict_context(scop, valid);
4760 if (skip)
4761 scop = pet_scop_prefix(scop, 0);
4762 scop = skip.add(scop, 1);
4764 return scop;
4767 /* Try and construct a pet_scop for a label statement.
4768 * We currently only allow labels on expression statements.
4770 struct pet_scop *PetScan::extract(LabelStmt *stmt)
4772 isl_id *label;
4773 Stmt *sub;
4775 sub = stmt->getSubStmt();
4776 if (!isa<Expr>(sub)) {
4777 unsupported(stmt);
4778 return NULL;
4781 label = isl_id_alloc(ctx, stmt->getName(), NULL);
4783 return extract(sub, extract_expr(cast<Expr>(sub)), label);
4786 /* Return a one-dimensional multi piecewise affine expression that is equal
4787 * to the constant 1 and is defined over a zero-dimensional domain.
4789 static __isl_give isl_multi_pw_aff *one_mpa(isl_ctx *ctx)
4791 isl_space *space;
4792 isl_local_space *ls;
4793 isl_aff *aff;
4795 space = isl_space_set_alloc(ctx, 0, 0);
4796 ls = isl_local_space_from_space(space);
4797 aff = isl_aff_zero_on_domain(ls);
4798 aff = isl_aff_set_constant_si(aff, 1);
4800 return isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
4803 /* Construct a pet_scop for a continue statement.
4805 * We simply create an empty scop with a universal pet_skip_now
4806 * skip condition. This skip condition will then be taken into
4807 * account by the enclosing loop construct, possibly after
4808 * being incorporated into outer skip conditions.
4810 struct pet_scop *PetScan::extract(ContinueStmt *stmt)
4812 pet_scop *scop;
4814 scop = pet_scop_empty(ctx);
4815 if (!scop)
4816 return NULL;
4818 scop = pet_scop_set_skip(scop, pet_skip_now, one_mpa(ctx));
4820 return scop;
4823 /* Construct a pet_scop for a break statement.
4825 * We simply create an empty scop with both a universal pet_skip_now
4826 * skip condition and a universal pet_skip_later skip condition.
4827 * These skip conditions will then be taken into
4828 * account by the enclosing loop construct, possibly after
4829 * being incorporated into outer skip conditions.
4831 struct pet_scop *PetScan::extract(BreakStmt *stmt)
4833 pet_scop *scop;
4834 isl_multi_pw_aff *skip;
4836 scop = pet_scop_empty(ctx);
4837 if (!scop)
4838 return NULL;
4840 skip = one_mpa(ctx);
4841 scop = pet_scop_set_skip(scop, pet_skip_now,
4842 isl_multi_pw_aff_copy(skip));
4843 scop = pet_scop_set_skip(scop, pet_skip_later, skip);
4845 return scop;
4848 /* Try and construct a pet_scop corresponding to "stmt".
4850 * If "stmt" is a compound statement, then "skip_declarations"
4851 * indicates whether we should skip initial declarations in the
4852 * compound statement.
4854 * If the constructed pet_scop is not a (possibly) partial representation
4855 * of "stmt", we update start and end of the pet_scop to those of "stmt".
4856 * In particular, if skip_declarations is set, then we may have skipped
4857 * declarations inside "stmt" and so the pet_scop may not represent
4858 * the entire "stmt".
4859 * Note that this function may be called with "stmt" referring to the entire
4860 * body of the function, including the outer braces. In such cases,
4861 * skip_declarations will be set and the braces will not be taken into
4862 * account in scop->start and scop->end.
4864 struct pet_scop *PetScan::extract(Stmt *stmt, bool skip_declarations)
4866 struct pet_scop *scop;
4868 if (isa<Expr>(stmt))
4869 return extract(stmt, extract_expr(cast<Expr>(stmt)));
4871 switch (stmt->getStmtClass()) {
4872 case Stmt::WhileStmtClass:
4873 scop = extract(cast<WhileStmt>(stmt));
4874 break;
4875 case Stmt::ForStmtClass:
4876 scop = extract_for(cast<ForStmt>(stmt));
4877 break;
4878 case Stmt::IfStmtClass:
4879 scop = extract(cast<IfStmt>(stmt));
4880 break;
4881 case Stmt::CompoundStmtClass:
4882 scop = extract(cast<CompoundStmt>(stmt), skip_declarations);
4883 break;
4884 case Stmt::LabelStmtClass:
4885 scop = extract(cast<LabelStmt>(stmt));
4886 break;
4887 case Stmt::ContinueStmtClass:
4888 scop = extract(cast<ContinueStmt>(stmt));
4889 break;
4890 case Stmt::BreakStmtClass:
4891 scop = extract(cast<BreakStmt>(stmt));
4892 break;
4893 case Stmt::DeclStmtClass:
4894 scop = extract(cast<DeclStmt>(stmt));
4895 break;
4896 default:
4897 unsupported(stmt);
4898 return NULL;
4901 if (partial || skip_declarations)
4902 return scop;
4904 scop = update_scop_start_end(scop, stmt->getSourceRange(), false);
4906 return scop;
4909 /* Do we need to construct a skip condition of the given type
4910 * on a sequence of statements?
4912 * There is no need to construct a new skip condition if only
4913 * only of the two statements has a skip condition or if both
4914 * of their skip conditions are affine.
4916 * In principle we also don't need a new continuation variable if
4917 * the continuation of scop2 is affine, but then we would need
4918 * to allow more complicated forms of continuations.
4920 static bool need_skip_seq(struct pet_scop *scop1, struct pet_scop *scop2,
4921 enum pet_skip type)
4923 if (!scop1 || !pet_scop_has_skip(scop1, type))
4924 return false;
4925 if (!scop2 || !pet_scop_has_skip(scop2, type))
4926 return false;
4927 if (pet_scop_has_affine_skip(scop1, type) &&
4928 pet_scop_has_affine_skip(scop2, type))
4929 return false;
4930 return true;
4933 /* Construct a scop for computing the skip condition of the given type and
4934 * with index expression "skip_index" for a sequence of two scops "scop1"
4935 * and "scop2".
4937 * The computed scop contains a single statement that essentially does
4939 * skip_index = skip_cond_1 ? 1 : skip_cond_2
4941 * or, in other words, skip_cond1 || skip_cond2.
4942 * In this expression, skip_cond_2 is filtered to reflect that it is
4943 * only evaluated when skip_cond_1 is false.
4945 * The skip condition on scop1 is not removed because it still needs
4946 * to be applied to scop2 when these two scops are combined.
4948 static struct pet_scop *extract_skip_seq(PetScan *ps,
4949 __isl_take isl_multi_pw_aff *skip_index,
4950 struct pet_scop *scop1, struct pet_scop *scop2, enum pet_skip type)
4952 struct pet_expr *expr1, *expr2, *expr, *expr_skip;
4953 struct pet_stmt *stmt;
4954 struct pet_scop *scop;
4955 isl_ctx *ctx = ps->ctx;
4957 if (!scop1 || !scop2)
4958 goto error;
4960 expr1 = pet_scop_get_skip_expr(scop1, type);
4961 expr2 = pet_scop_get_skip_expr(scop2, type);
4962 pet_scop_reset_skip(scop2, type);
4964 expr2 = pet_expr_filter(expr2,
4965 isl_multi_pw_aff_copy(expr1->acc.index), 0);
4967 expr = universally_true(ctx);
4968 expr = pet_expr_new_ternary(ctx, expr1, expr, expr2);
4969 expr_skip = pet_expr_from_index(isl_multi_pw_aff_copy(skip_index));
4970 if (expr_skip) {
4971 expr_skip->acc.write = 1;
4972 expr_skip->acc.read = 0;
4974 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
4975 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, ps->n_stmt++, expr);
4977 scop = pet_scop_from_pet_stmt(ctx, stmt);
4978 scop = scop_add_array(scop, skip_index, ps->ast_context);
4979 isl_multi_pw_aff_free(skip_index);
4981 return scop;
4982 error:
4983 isl_multi_pw_aff_free(skip_index);
4984 return NULL;
4987 /* Structure that handles the construction of skip conditions
4988 * on sequences of statements.
4990 * scop1 and scop2 represent the two statements that are combined
4992 struct pet_skip_info_seq : public pet_skip_info {
4993 struct pet_scop *scop1, *scop2;
4995 pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4996 struct pet_scop *scop2);
4997 void extract(PetScan *scan, enum pet_skip type);
4998 void extract(PetScan *scan);
4999 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
5000 int offset);
5001 struct pet_scop *add(struct pet_scop *scop, int offset);
5004 /* Initialize a pet_skip_info_seq structure based on
5005 * on the two statements that are going to be combined.
5007 pet_skip_info_seq::pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
5008 struct pet_scop *scop2) : pet_skip_info(ctx), scop1(scop1), scop2(scop2)
5010 skip[pet_skip_now] = need_skip_seq(scop1, scop2, pet_skip_now);
5011 equal = skip[pet_skip_now] && skip_equals_skip_later(scop1) &&
5012 skip_equals_skip_later(scop2);
5013 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
5014 need_skip_seq(scop1, scop2, pet_skip_later);
5017 /* If we need to construct a skip condition of the given type,
5018 * then do so now.
5020 void pet_skip_info_seq::extract(PetScan *scan, enum pet_skip type)
5022 if (!skip[type])
5023 return;
5025 index[type] = create_test_index(ctx, scan->n_test++);
5026 scop[type] = extract_skip_seq(scan, isl_multi_pw_aff_copy(index[type]),
5027 scop1, scop2, type);
5030 /* Construct the required skip conditions.
5032 void pet_skip_info_seq::extract(PetScan *scan)
5034 extract(scan, pet_skip_now);
5035 extract(scan, pet_skip_later);
5036 if (equal)
5037 drop_skip_later(scop1, scop2);
5040 /* Add the computed skip condition of the given type to "main" and
5041 * add the scop for computing the condition at the given offset (the statement
5042 * number). Within this offset, the condition is computed at position 1
5043 * to ensure that it is computed after the corresponding statement.
5045 * If equal is set, then we only computed a skip condition for pet_skip_now,
5046 * but we also need to set it as main's pet_skip_later.
5048 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *main,
5049 enum pet_skip type, int offset)
5051 if (!skip[type])
5052 return main;
5054 scop[type] = pet_scop_prefix(scop[type], 1);
5055 scop[type] = pet_scop_prefix(scop[type], offset);
5056 main = pet_scop_add_par(ctx, main, scop[type]);
5057 scop[type] = NULL;
5059 if (equal)
5060 main = pet_scop_set_skip(main, pet_skip_later,
5061 isl_multi_pw_aff_copy(index[type]));
5063 main = pet_scop_set_skip(main, type, index[type]);
5064 index[type] = NULL;
5066 return main;
5069 /* Add the computed skip conditions to "main" and
5070 * add the scops for computing the conditions at the given offset.
5072 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *scop, int offset)
5074 scop = add(scop, pet_skip_now, offset);
5075 scop = add(scop, pet_skip_later, offset);
5077 return scop;
5080 /* Extract a clone of the kill statement in "scop".
5081 * "scop" is expected to have been created from a DeclStmt
5082 * and should have the kill as its first statement.
5084 struct pet_stmt *PetScan::extract_kill(struct pet_scop *scop)
5086 struct pet_expr *kill;
5087 struct pet_stmt *stmt;
5088 isl_multi_pw_aff *index;
5089 isl_map *access;
5091 if (!scop)
5092 return NULL;
5093 if (scop->n_stmt < 1)
5094 isl_die(ctx, isl_error_internal,
5095 "expecting at least one statement", return NULL);
5096 stmt = scop->stmts[0];
5097 if (!pet_stmt_is_kill(stmt))
5098 isl_die(ctx, isl_error_internal,
5099 "expecting kill statement", return NULL);
5101 index = isl_multi_pw_aff_copy(stmt->body->args[0]->acc.index);
5102 access = isl_map_copy(stmt->body->args[0]->acc.access);
5103 index = isl_multi_pw_aff_reset_tuple_id(index, isl_dim_in);
5104 access = isl_map_reset_tuple_id(access, isl_dim_in);
5105 kill = pet_expr_kill_from_access_and_index(access, index);
5106 return pet_stmt_from_pet_expr(ctx, stmt->line, NULL, n_stmt++, kill);
5109 /* Mark all arrays in "scop" as being exposed.
5111 static struct pet_scop *mark_exposed(struct pet_scop *scop)
5113 if (!scop)
5114 return NULL;
5115 for (int i = 0; i < scop->n_array; ++i)
5116 scop->arrays[i]->exposed = 1;
5117 return scop;
5120 /* Try and construct a pet_scop corresponding to (part of)
5121 * a sequence of statements.
5123 * "block" is set if the sequence respresents the children of
5124 * a compound statement.
5125 * "skip_declarations" is set if we should skip initial declarations
5126 * in the sequence of statements.
5128 * If there are any breaks or continues in the individual statements,
5129 * then we may have to compute a new skip condition.
5130 * This is handled using a pet_skip_info_seq object.
5131 * On initialization, the object checks if skip conditions need
5132 * to be computed. If so, it does so in "extract" and adds them in "add".
5134 * If "block" is set, then we need to insert kill statements at
5135 * the end of the block for any array that has been declared by
5136 * one of the statements in the sequence. Each of these declarations
5137 * results in the construction of a kill statement at the place
5138 * of the declaration, so we simply collect duplicates of
5139 * those kill statements and append these duplicates to the constructed scop.
5141 * If "block" is not set, then any array declared by one of the statements
5142 * in the sequence is marked as being exposed.
5144 * If autodetect is set, then we allow the extraction of only a subrange
5145 * of the sequence of statements. However, if there is at least one statement
5146 * for which we could not construct a scop and the final range contains
5147 * either no statements or at least one kill, then we discard the entire
5148 * range.
5150 struct pet_scop *PetScan::extract(StmtRange stmt_range, bool block,
5151 bool skip_declarations)
5153 pet_scop *scop;
5154 StmtIterator i;
5155 int j;
5156 bool partial_range = false;
5157 set<struct pet_stmt *> kills;
5158 set<struct pet_stmt *>::iterator it;
5160 scop = pet_scop_empty(ctx);
5161 for (i = stmt_range.first, j = 0; i != stmt_range.second; ++i, ++j) {
5162 Stmt *child = *i;
5163 struct pet_scop *scop_i;
5165 if (scop->n_stmt == 0 && skip_declarations &&
5166 child->getStmtClass() == Stmt::DeclStmtClass)
5167 continue;
5169 scop_i = extract(child);
5170 if (scop->n_stmt != 0 && partial) {
5171 pet_scop_free(scop_i);
5172 break;
5174 pet_skip_info_seq skip(ctx, scop, scop_i);
5175 skip.extract(this);
5176 if (skip)
5177 scop_i = pet_scop_prefix(scop_i, 0);
5178 if (scop_i && child->getStmtClass() == Stmt::DeclStmtClass) {
5179 if (block)
5180 kills.insert(extract_kill(scop_i));
5181 else
5182 scop_i = mark_exposed(scop_i);
5184 scop_i = pet_scop_prefix(scop_i, j);
5185 if (options->autodetect) {
5186 if (scop_i)
5187 scop = pet_scop_add_seq(ctx, scop, scop_i);
5188 else
5189 partial_range = true;
5190 if (scop->n_stmt != 0 && !scop_i)
5191 partial = true;
5192 } else {
5193 scop = pet_scop_add_seq(ctx, scop, scop_i);
5196 scop = skip.add(scop, j);
5198 if (partial || !scop)
5199 break;
5202 for (it = kills.begin(); it != kills.end(); ++it) {
5203 pet_scop *scop_j;
5204 scop_j = pet_scop_from_pet_stmt(ctx, *it);
5205 scop_j = pet_scop_prefix(scop_j, j);
5206 scop = pet_scop_add_seq(ctx, scop, scop_j);
5209 if (scop && partial_range) {
5210 if (scop->n_stmt == 0 || kills.size() != 0) {
5211 pet_scop_free(scop);
5212 return NULL;
5214 partial = true;
5217 return scop;
5220 /* Check if the scop marked by the user is exactly this Stmt
5221 * or part of this Stmt.
5222 * If so, return a pet_scop corresponding to the marked region.
5223 * Otherwise, return NULL.
5225 struct pet_scop *PetScan::scan(Stmt *stmt)
5227 SourceManager &SM = PP.getSourceManager();
5228 unsigned start_off, end_off;
5230 start_off = getExpansionOffset(SM, stmt->getLocStart());
5231 end_off = getExpansionOffset(SM, stmt->getLocEnd());
5233 if (start_off > loc.end)
5234 return NULL;
5235 if (end_off < loc.start)
5236 return NULL;
5237 if (start_off >= loc.start && end_off <= loc.end) {
5238 return extract(stmt);
5241 StmtIterator start;
5242 for (start = stmt->child_begin(); start != stmt->child_end(); ++start) {
5243 Stmt *child = *start;
5244 if (!child)
5245 continue;
5246 start_off = getExpansionOffset(SM, child->getLocStart());
5247 end_off = getExpansionOffset(SM, child->getLocEnd());
5248 if (start_off < loc.start && end_off >= loc.end)
5249 return scan(child);
5250 if (start_off >= loc.start)
5251 break;
5254 StmtIterator end;
5255 for (end = start; end != stmt->child_end(); ++end) {
5256 Stmt *child = *end;
5257 start_off = SM.getFileOffset(child->getLocStart());
5258 if (start_off >= loc.end)
5259 break;
5262 return extract(StmtRange(start, end), false, false);
5265 /* Set the size of index "pos" of "array" to "size".
5266 * In particular, add a constraint of the form
5268 * i_pos < size
5270 * to array->extent and a constraint of the form
5272 * size >= 0
5274 * to array->context.
5276 static struct pet_array *update_size(struct pet_array *array, int pos,
5277 __isl_take isl_pw_aff *size)
5279 isl_set *valid;
5280 isl_set *univ;
5281 isl_set *bound;
5282 isl_space *dim;
5283 isl_aff *aff;
5284 isl_pw_aff *index;
5285 isl_id *id;
5287 valid = isl_pw_aff_nonneg_set(isl_pw_aff_copy(size));
5288 array->context = isl_set_intersect(array->context, valid);
5290 dim = isl_set_get_space(array->extent);
5291 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
5292 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, pos, 1);
5293 univ = isl_set_universe(isl_aff_get_domain_space(aff));
5294 index = isl_pw_aff_alloc(univ, aff);
5296 size = isl_pw_aff_add_dims(size, isl_dim_in,
5297 isl_set_dim(array->extent, isl_dim_set));
5298 id = isl_set_get_tuple_id(array->extent);
5299 size = isl_pw_aff_set_tuple_id(size, isl_dim_in, id);
5300 bound = isl_pw_aff_lt_set(index, size);
5302 array->extent = isl_set_intersect(array->extent, bound);
5304 if (!array->context || !array->extent)
5305 goto error;
5307 return array;
5308 error:
5309 pet_array_free(array);
5310 return NULL;
5313 /* Figure out the size of the array at position "pos" and all
5314 * subsequent positions from "type" and update "array" accordingly.
5316 struct pet_array *PetScan::set_upper_bounds(struct pet_array *array,
5317 const Type *type, int pos)
5319 const ArrayType *atype;
5320 isl_pw_aff *size;
5322 if (!array)
5323 return NULL;
5325 if (type->isPointerType()) {
5326 type = type->getPointeeType().getTypePtr();
5327 return set_upper_bounds(array, type, pos + 1);
5329 if (!type->isArrayType())
5330 return array;
5332 type = type->getCanonicalTypeInternal().getTypePtr();
5333 atype = cast<ArrayType>(type);
5335 if (type->isConstantArrayType()) {
5336 const ConstantArrayType *ca = cast<ConstantArrayType>(atype);
5337 size = extract_affine(ca->getSize());
5338 array = update_size(array, pos, size);
5339 } else if (type->isVariableArrayType()) {
5340 const VariableArrayType *vla = cast<VariableArrayType>(atype);
5341 size = extract_affine(vla->getSizeExpr());
5342 array = update_size(array, pos, size);
5345 type = atype->getElementType().getTypePtr();
5347 return set_upper_bounds(array, type, pos + 1);
5350 /* Is "T" the type of a variable length array with static size?
5352 static bool is_vla_with_static_size(QualType T)
5354 const VariableArrayType *vlatype;
5356 if (!T->isVariableArrayType())
5357 return false;
5358 vlatype = cast<VariableArrayType>(T);
5359 return vlatype->getSizeModifier() == VariableArrayType::Static;
5362 /* Return the type of "decl" as an array.
5364 * In particular, if "decl" is a parameter declaration that
5365 * is a variable length array with a static size, then
5366 * return the original type (i.e., the variable length array).
5367 * Otherwise, return the type of decl.
5369 static QualType get_array_type(ValueDecl *decl)
5371 ParmVarDecl *parm;
5372 QualType T;
5374 parm = dyn_cast<ParmVarDecl>(decl);
5375 if (!parm)
5376 return decl->getType();
5378 T = parm->getOriginalType();
5379 if (!is_vla_with_static_size(T))
5380 return decl->getType();
5381 return T;
5384 /* Does "decl" have definition that we can keep track of in a pet_type?
5386 static bool has_printable_definition(RecordDecl *decl)
5388 if (!decl->getDeclName())
5389 return false;
5390 return decl->getLexicalDeclContext() == decl->getDeclContext();
5393 /* Construct and return a pet_array corresponding to the variable "decl".
5394 * In particular, initialize array->extent to
5396 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
5398 * and then call set_upper_bounds to set the upper bounds on the indices
5399 * based on the type of the variable.
5401 * If the base type is that of a record with a top-level definition and
5402 * if "types" is not null, then the RecordDecl corresponding to the type
5403 * is added to "types".
5405 * If the base type is that of a record with no top-level definition,
5406 * then we replace it by "<subfield>".
5408 struct pet_array *PetScan::extract_array(isl_ctx *ctx, ValueDecl *decl,
5409 lex_recorddecl_set *types)
5411 struct pet_array *array;
5412 QualType qt = get_array_type(decl);
5413 const Type *type = qt.getTypePtr();
5414 int depth = array_depth(type);
5415 QualType base = pet_clang_base_type(qt);
5416 string name;
5417 isl_id *id;
5418 isl_space *dim;
5420 array = isl_calloc_type(ctx, struct pet_array);
5421 if (!array)
5422 return NULL;
5424 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
5425 dim = isl_space_set_alloc(ctx, 0, depth);
5426 dim = isl_space_set_tuple_id(dim, isl_dim_set, id);
5428 array->extent = isl_set_nat_universe(dim);
5430 dim = isl_space_params_alloc(ctx, 0);
5431 array->context = isl_set_universe(dim);
5433 array = set_upper_bounds(array, type, 0);
5434 if (!array)
5435 return NULL;
5437 name = base.getAsString();
5439 if (types && base->isRecordType()) {
5440 RecordDecl *decl = pet_clang_record_decl(base);
5441 if (has_printable_definition(decl))
5442 types->insert(decl);
5443 else
5444 name = "<subfield>";
5447 array->element_type = strdup(name.c_str());
5448 array->element_is_record = base->isRecordType();
5449 array->element_size = decl->getASTContext().getTypeInfo(base).first / 8;
5451 return array;
5454 /* Construct and return a pet_array corresponding to the sequence
5455 * of declarations "decls".
5456 * If the sequence contains a single declaration, then it corresponds
5457 * to a simple array access. Otherwise, it corresponds to a member access,
5458 * with the declaration for the substructure following that of the containing
5459 * structure in the sequence of declarations.
5460 * We start with the outermost substructure and then combine it with
5461 * information from the inner structures.
5463 * Additionally, keep track of all required types in "types".
5465 struct pet_array *PetScan::extract_array(isl_ctx *ctx,
5466 vector<ValueDecl *> decls, lex_recorddecl_set *types)
5468 struct pet_array *array;
5469 vector<ValueDecl *>::iterator it;
5471 it = decls.begin();
5473 array = extract_array(ctx, *it, types);
5475 for (++it; it != decls.end(); ++it) {
5476 struct pet_array *parent;
5477 const char *base_name, *field_name;
5478 char *product_name;
5480 parent = array;
5481 array = extract_array(ctx, *it, types);
5482 if (!array)
5483 return pet_array_free(parent);
5485 base_name = isl_set_get_tuple_name(parent->extent);
5486 field_name = isl_set_get_tuple_name(array->extent);
5487 product_name = member_access_name(ctx, base_name, field_name);
5489 array->extent = isl_set_product(isl_set_copy(parent->extent),
5490 array->extent);
5491 if (product_name)
5492 array->extent = isl_set_set_tuple_name(array->extent,
5493 product_name);
5494 array->context = isl_set_intersect(array->context,
5495 isl_set_copy(parent->context));
5497 pet_array_free(parent);
5498 free(product_name);
5500 if (!array->extent || !array->context || !product_name)
5501 return pet_array_free(array);
5504 return array;
5507 /* Add a pet_type corresponding to "decl" to "scop, provided
5508 * it is a member of "types" and it has not been added before
5509 * (i.e., it is not a member of "types_done".
5511 * Since we want the user to be able to print the types
5512 * in the order in which they appear in the scop, we need to
5513 * make sure that types of fields in a structure appear before
5514 * that structure. We therefore call ourselves recursively
5515 * on the types of all record subfields.
5517 static struct pet_scop *add_type(isl_ctx *ctx, struct pet_scop *scop,
5518 RecordDecl *decl, Preprocessor &PP, lex_recorddecl_set &types,
5519 lex_recorddecl_set &types_done)
5521 string s;
5522 llvm::raw_string_ostream S(s);
5523 RecordDecl::field_iterator it;
5525 if (types.find(decl) == types.end())
5526 return scop;
5527 if (types_done.find(decl) != types_done.end())
5528 return scop;
5530 for (it = decl->field_begin(); it != decl->field_end(); ++it) {
5531 RecordDecl *record;
5532 QualType type = it->getType();
5534 if (!type->isRecordType())
5535 continue;
5536 record = pet_clang_record_decl(type);
5537 scop = add_type(ctx, scop, record, PP, types, types_done);
5540 if (strlen(decl->getName().str().c_str()) == 0)
5541 return scop;
5543 decl->print(S, PrintingPolicy(PP.getLangOpts()));
5544 S.str();
5546 scop->types[scop->n_type] = pet_type_alloc(ctx,
5547 decl->getName().str().c_str(), s.c_str());
5548 if (!scop->types[scop->n_type])
5549 return pet_scop_free(scop);
5551 types_done.insert(decl);
5553 scop->n_type++;
5555 return scop;
5558 /* Construct a list of pet_arrays, one for each array (or scalar)
5559 * accessed inside "scop", add this list to "scop" and return the result.
5561 * The context of "scop" is updated with the intersection of
5562 * the contexts of all arrays, i.e., constraints on the parameters
5563 * that ensure that the arrays have a valid (non-negative) size.
5565 * If the any of the extracted arrays refers to a member access,
5566 * then also add the required types to "scop".
5568 struct pet_scop *PetScan::scan_arrays(struct pet_scop *scop)
5570 int i;
5571 set<vector<ValueDecl *> > arrays;
5572 set<vector<ValueDecl *> >::iterator it;
5573 lex_recorddecl_set types;
5574 lex_recorddecl_set types_done;
5575 lex_recorddecl_set::iterator types_it;
5576 int n_array;
5577 struct pet_array **scop_arrays;
5579 if (!scop)
5580 return NULL;
5582 pet_scop_collect_arrays(scop, arrays);
5583 if (arrays.size() == 0)
5584 return scop;
5586 n_array = scop->n_array;
5588 scop_arrays = isl_realloc_array(ctx, scop->arrays, struct pet_array *,
5589 n_array + arrays.size());
5590 if (!scop_arrays)
5591 goto error;
5592 scop->arrays = scop_arrays;
5594 for (it = arrays.begin(), i = 0; it != arrays.end(); ++it, ++i) {
5595 struct pet_array *array;
5596 array = extract_array(ctx, *it, &types);
5597 scop->arrays[n_array + i] = array;
5598 if (!scop->arrays[n_array + i])
5599 goto error;
5600 scop->n_array++;
5601 scop->context = isl_set_intersect(scop->context,
5602 isl_set_copy(array->context));
5603 if (!scop->context)
5604 goto error;
5607 if (types.size() == 0)
5608 return scop;
5610 scop->types = isl_alloc_array(ctx, struct pet_type *, types.size());
5611 if (!scop->types)
5612 goto error;
5614 for (types_it = types.begin(); types_it != types.end(); ++types_it)
5615 scop = add_type(ctx, scop, *types_it, PP, types, types_done);
5617 return scop;
5618 error:
5619 pet_scop_free(scop);
5620 return NULL;
5623 /* Bound all parameters in scop->context to the possible values
5624 * of the corresponding C variable.
5626 static struct pet_scop *add_parameter_bounds(struct pet_scop *scop)
5628 int n;
5630 if (!scop)
5631 return NULL;
5633 n = isl_set_dim(scop->context, isl_dim_param);
5634 for (int i = 0; i < n; ++i) {
5635 isl_id *id;
5636 ValueDecl *decl;
5638 id = isl_set_get_dim_id(scop->context, isl_dim_param, i);
5639 if (is_nested_parameter(id)) {
5640 isl_id_free(id);
5641 isl_die(isl_set_get_ctx(scop->context),
5642 isl_error_internal,
5643 "unresolved nested parameter", goto error);
5645 decl = (ValueDecl *) isl_id_get_user(id);
5646 isl_id_free(id);
5648 scop->context = set_parameter_bounds(scop->context, i, decl);
5650 if (!scop->context)
5651 goto error;
5654 return scop;
5655 error:
5656 pet_scop_free(scop);
5657 return NULL;
5660 /* Construct a pet_scop from the given function.
5662 * If the scop was delimited by scop and endscop pragmas, then we override
5663 * the file offsets by those derived from the pragmas.
5665 struct pet_scop *PetScan::scan(FunctionDecl *fd)
5667 pet_scop *scop;
5668 Stmt *stmt;
5670 stmt = fd->getBody();
5672 if (options->autodetect)
5673 scop = extract(stmt, true);
5674 else {
5675 scop = scan(stmt);
5676 scop = pet_scop_update_start_end(scop, loc.start, loc.end);
5678 scop = pet_scop_detect_parameter_accesses(scop);
5679 scop = scan_arrays(scop);
5680 scop = add_parameter_bounds(scop);
5681 scop = pet_scop_gist(scop, value_bounds);
5683 return scop;