scop.c: extract out expr_collect_access
[pet.git] / scan.cc
blob82c2a6a4f43158213682932d326ba55dbdadce1e
1 /*
2 * Copyright 2011 Leiden University. All rights reserved.
3 * Copyright 2012-2013 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 /* Called if we found something we (currently) cannot handle.
277 * We'll provide more informative warnings later.
279 * We only actually complain if autodetect is false.
281 void PetScan::unsupported(Stmt *stmt, const char *msg)
283 if (options->autodetect)
284 return;
286 SourceLocation loc = stmt->getLocStart();
287 DiagnosticsEngine &diag = PP.getDiagnostics();
288 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
289 msg ? msg : "unsupported");
290 DiagnosticBuilder B = diag.Report(loc, id) << stmt->getSourceRange();
293 /* Extract an integer from "expr".
295 __isl_give isl_val *PetScan::extract_int(isl_ctx *ctx, IntegerLiteral *expr)
297 const Type *type = expr->getType().getTypePtr();
298 int is_signed = type->hasSignedIntegerRepresentation();
299 llvm::APInt val = expr->getValue();
300 int is_negative = is_signed && val.isNegative();
301 isl_val *v;
303 if (is_negative)
304 val = -val;
306 v = extract_unsigned(ctx, val);
308 if (is_negative)
309 v = isl_val_neg(v);
310 return v;
313 /* Extract an integer from "val", which assumed to be non-negative.
315 __isl_give isl_val *PetScan::extract_unsigned(isl_ctx *ctx,
316 const llvm::APInt &val)
318 unsigned n;
319 const uint64_t *data;
321 data = val.getRawData();
322 n = val.getNumWords();
323 return isl_val_int_from_chunks(ctx, n, sizeof(uint64_t), data);
326 /* Extract an integer from "expr".
327 * Return NULL if "expr" does not (obviously) represent an integer.
329 __isl_give isl_val *PetScan::extract_int(clang::ParenExpr *expr)
331 return extract_int(expr->getSubExpr());
334 /* Extract an integer from "expr".
335 * Return NULL if "expr" does not (obviously) represent an integer.
337 __isl_give isl_val *PetScan::extract_int(clang::Expr *expr)
339 if (expr->getStmtClass() == Stmt::IntegerLiteralClass)
340 return extract_int(ctx, cast<IntegerLiteral>(expr));
341 if (expr->getStmtClass() == Stmt::ParenExprClass)
342 return extract_int(cast<ParenExpr>(expr));
344 unsupported(expr);
345 return NULL;
348 /* Extract an affine expression from the IntegerLiteral "expr".
350 __isl_give isl_pw_aff *PetScan::extract_affine(IntegerLiteral *expr)
352 isl_space *dim = isl_space_params_alloc(ctx, 0);
353 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
354 isl_aff *aff = isl_aff_zero_on_domain(ls);
355 isl_set *dom = isl_set_universe(dim);
356 isl_val *v;
358 v = extract_int(expr);
359 aff = isl_aff_add_constant_val(aff, v);
361 return isl_pw_aff_alloc(dom, aff);
364 /* Extract an affine expression from the APInt "val", which is assumed
365 * to be non-negative.
367 __isl_give isl_pw_aff *PetScan::extract_affine(const llvm::APInt &val)
369 isl_space *dim = isl_space_params_alloc(ctx, 0);
370 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
371 isl_aff *aff = isl_aff_zero_on_domain(ls);
372 isl_set *dom = isl_set_universe(dim);
373 isl_val *v;
375 v = extract_unsigned(ctx, val);
376 aff = isl_aff_add_constant_val(aff, v);
378 return isl_pw_aff_alloc(dom, aff);
381 __isl_give isl_pw_aff *PetScan::extract_affine(ImplicitCastExpr *expr)
383 return extract_affine(expr->getSubExpr());
386 static unsigned get_type_size(ValueDecl *decl)
388 return decl->getASTContext().getIntWidth(decl->getType());
391 /* Bound parameter "pos" of "set" to the possible values of "decl".
393 static __isl_give isl_set *set_parameter_bounds(__isl_take isl_set *set,
394 unsigned pos, ValueDecl *decl)
396 unsigned width;
397 isl_ctx *ctx;
398 isl_val *bound;
400 ctx = isl_set_get_ctx(set);
401 width = get_type_size(decl);
402 if (decl->getType()->isUnsignedIntegerType()) {
403 set = isl_set_lower_bound_si(set, isl_dim_param, pos, 0);
404 bound = isl_val_int_from_ui(ctx, width);
405 bound = isl_val_2exp(bound);
406 bound = isl_val_sub_ui(bound, 1);
407 set = isl_set_upper_bound_val(set, isl_dim_param, pos, bound);
408 } else {
409 bound = isl_val_int_from_ui(ctx, width - 1);
410 bound = isl_val_2exp(bound);
411 bound = isl_val_sub_ui(bound, 1);
412 set = isl_set_upper_bound_val(set, isl_dim_param, pos,
413 isl_val_copy(bound));
414 bound = isl_val_neg(bound);
415 bound = isl_val_sub_ui(bound, 1);
416 set = isl_set_lower_bound_val(set, isl_dim_param, pos, bound);
419 return set;
422 /* Extract an affine expression from the DeclRefExpr "expr".
424 * If the variable has been assigned a value, then we check whether
425 * we know what (affine) value was assigned.
426 * If so, we return this value. Otherwise we convert "expr"
427 * to an extra parameter (provided nesting_enabled is set).
429 * Otherwise, we simply return an expression that is equal
430 * to a parameter corresponding to the referenced variable.
432 __isl_give isl_pw_aff *PetScan::extract_affine(DeclRefExpr *expr)
434 ValueDecl *decl = expr->getDecl();
435 const Type *type = decl->getType().getTypePtr();
436 isl_id *id;
437 isl_space *dim;
438 isl_aff *aff;
439 isl_set *dom;
441 if (!type->isIntegerType()) {
442 unsupported(expr);
443 return NULL;
446 if (assigned_value.find(decl) != assigned_value.end()) {
447 if (assigned_value[decl])
448 return isl_pw_aff_copy(assigned_value[decl]);
449 else
450 return nested_access(expr);
453 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
454 dim = isl_space_params_alloc(ctx, 1);
456 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
458 dom = isl_set_universe(isl_space_copy(dim));
459 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
460 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
462 return isl_pw_aff_alloc(dom, aff);
465 /* Extract an affine expression from an integer division operation.
466 * In particular, if "expr" is lhs/rhs, then return
468 * lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs)
470 * The second argument (rhs) is required to be a (positive) integer constant.
472 __isl_give isl_pw_aff *PetScan::extract_affine_div(BinaryOperator *expr)
474 int is_cst;
475 isl_pw_aff *rhs, *lhs;
477 rhs = extract_affine(expr->getRHS());
478 is_cst = isl_pw_aff_is_cst(rhs);
479 if (is_cst < 0 || !is_cst) {
480 isl_pw_aff_free(rhs);
481 if (!is_cst)
482 unsupported(expr);
483 return NULL;
486 lhs = extract_affine(expr->getLHS());
488 return isl_pw_aff_tdiv_q(lhs, rhs);
491 /* Extract an affine expression from a modulo operation.
492 * In particular, if "expr" is lhs/rhs, then return
494 * lhs - rhs * (lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs))
496 * The second argument (rhs) is required to be a (positive) integer constant.
498 __isl_give isl_pw_aff *PetScan::extract_affine_mod(BinaryOperator *expr)
500 int is_cst;
501 isl_pw_aff *rhs, *lhs;
503 rhs = extract_affine(expr->getRHS());
504 is_cst = isl_pw_aff_is_cst(rhs);
505 if (is_cst < 0 || !is_cst) {
506 isl_pw_aff_free(rhs);
507 if (!is_cst)
508 unsupported(expr);
509 return NULL;
512 lhs = extract_affine(expr->getLHS());
514 return isl_pw_aff_tdiv_r(lhs, rhs);
517 /* Extract an affine expression from a multiplication operation.
518 * This is only allowed if at least one of the two arguments
519 * is a (piecewise) constant.
521 __isl_give isl_pw_aff *PetScan::extract_affine_mul(BinaryOperator *expr)
523 isl_pw_aff *lhs;
524 isl_pw_aff *rhs;
526 lhs = extract_affine(expr->getLHS());
527 rhs = extract_affine(expr->getRHS());
529 if (!isl_pw_aff_is_cst(lhs) && !isl_pw_aff_is_cst(rhs)) {
530 isl_pw_aff_free(lhs);
531 isl_pw_aff_free(rhs);
532 unsupported(expr);
533 return NULL;
536 return isl_pw_aff_mul(lhs, rhs);
539 /* Extract an affine expression from an addition or subtraction operation.
541 __isl_give isl_pw_aff *PetScan::extract_affine_add(BinaryOperator *expr)
543 isl_pw_aff *lhs;
544 isl_pw_aff *rhs;
546 lhs = extract_affine(expr->getLHS());
547 rhs = extract_affine(expr->getRHS());
549 switch (expr->getOpcode()) {
550 case BO_Add:
551 return isl_pw_aff_add(lhs, rhs);
552 case BO_Sub:
553 return isl_pw_aff_sub(lhs, rhs);
554 default:
555 isl_pw_aff_free(lhs);
556 isl_pw_aff_free(rhs);
557 return NULL;
562 /* Compute
564 * pwaff mod 2^width
566 static __isl_give isl_pw_aff *wrap(__isl_take isl_pw_aff *pwaff,
567 unsigned width)
569 isl_ctx *ctx;
570 isl_val *mod;
572 ctx = isl_pw_aff_get_ctx(pwaff);
573 mod = isl_val_int_from_ui(ctx, width);
574 mod = isl_val_2exp(mod);
576 pwaff = isl_pw_aff_mod_val(pwaff, mod);
578 return pwaff;
581 /* Limit the domain of "pwaff" to those elements where the function
582 * value satisfies
584 * 2^{width-1} <= pwaff < 2^{width-1}
586 static __isl_give isl_pw_aff *avoid_overflow(__isl_take isl_pw_aff *pwaff,
587 unsigned width)
589 isl_ctx *ctx;
590 isl_val *v;
591 isl_space *space = isl_pw_aff_get_domain_space(pwaff);
592 isl_local_space *ls = isl_local_space_from_space(space);
593 isl_aff *bound;
594 isl_set *dom;
595 isl_pw_aff *b;
597 ctx = isl_pw_aff_get_ctx(pwaff);
598 v = isl_val_int_from_ui(ctx, width - 1);
599 v = isl_val_2exp(v);
601 bound = isl_aff_zero_on_domain(ls);
602 bound = isl_aff_add_constant_val(bound, v);
603 b = isl_pw_aff_from_aff(bound);
605 dom = isl_pw_aff_lt_set(isl_pw_aff_copy(pwaff), isl_pw_aff_copy(b));
606 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
608 b = isl_pw_aff_neg(b);
609 dom = isl_pw_aff_ge_set(isl_pw_aff_copy(pwaff), b);
610 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
612 return pwaff;
615 /* Handle potential overflows on signed computations.
617 * If options->signed_overflow is set to PET_OVERFLOW_AVOID,
618 * the we adjust the domain of "pa" to avoid overflows.
620 __isl_give isl_pw_aff *PetScan::signed_overflow(__isl_take isl_pw_aff *pa,
621 unsigned width)
623 if (options->signed_overflow == PET_OVERFLOW_AVOID)
624 pa = avoid_overflow(pa, width);
626 return pa;
629 /* Return the piecewise affine expression "set ? 1 : 0" defined on "dom".
631 static __isl_give isl_pw_aff *indicator_function(__isl_take isl_set *set,
632 __isl_take isl_set *dom)
634 isl_pw_aff *pa;
635 pa = isl_set_indicator_function(set);
636 pa = isl_pw_aff_intersect_domain(pa, dom);
637 return pa;
640 /* Extract an affine expression from some binary operations.
641 * If the result of the expression is unsigned, then we wrap it
642 * based on the size of the type. Otherwise, we ensure that
643 * no overflow occurs.
645 __isl_give isl_pw_aff *PetScan::extract_affine(BinaryOperator *expr)
647 isl_pw_aff *res;
648 unsigned width;
650 switch (expr->getOpcode()) {
651 case BO_Add:
652 case BO_Sub:
653 res = extract_affine_add(expr);
654 break;
655 case BO_Div:
656 res = extract_affine_div(expr);
657 break;
658 case BO_Rem:
659 res = extract_affine_mod(expr);
660 break;
661 case BO_Mul:
662 res = extract_affine_mul(expr);
663 break;
664 case BO_LT:
665 case BO_LE:
666 case BO_GT:
667 case BO_GE:
668 case BO_EQ:
669 case BO_NE:
670 case BO_LAnd:
671 case BO_LOr:
672 return extract_condition(expr);
673 default:
674 unsupported(expr);
675 return NULL;
678 width = ast_context.getIntWidth(expr->getType());
679 if (expr->getType()->isUnsignedIntegerType())
680 res = wrap(res, width);
681 else
682 res = signed_overflow(res, width);
684 return res;
687 /* Extract an affine expression from a negation operation.
689 __isl_give isl_pw_aff *PetScan::extract_affine(UnaryOperator *expr)
691 if (expr->getOpcode() == UO_Minus)
692 return isl_pw_aff_neg(extract_affine(expr->getSubExpr()));
693 if (expr->getOpcode() == UO_LNot)
694 return extract_condition(expr);
696 unsupported(expr);
697 return NULL;
700 __isl_give isl_pw_aff *PetScan::extract_affine(ParenExpr *expr)
702 return extract_affine(expr->getSubExpr());
705 /* Extract an affine expression from some special function calls.
706 * In particular, we handle "min", "max", "ceild" and "floord".
707 * In case of the latter two, the second argument needs to be
708 * a (positive) integer constant.
710 __isl_give isl_pw_aff *PetScan::extract_affine(CallExpr *expr)
712 FunctionDecl *fd;
713 string name;
714 isl_pw_aff *aff1, *aff2;
716 fd = expr->getDirectCallee();
717 if (!fd) {
718 unsupported(expr);
719 return NULL;
722 name = fd->getDeclName().getAsString();
723 if (!(expr->getNumArgs() == 2 && name == "min") &&
724 !(expr->getNumArgs() == 2 && name == "max") &&
725 !(expr->getNumArgs() == 2 && name == "floord") &&
726 !(expr->getNumArgs() == 2 && name == "ceild")) {
727 unsupported(expr);
728 return NULL;
731 if (name == "min" || name == "max") {
732 aff1 = extract_affine(expr->getArg(0));
733 aff2 = extract_affine(expr->getArg(1));
735 if (name == "min")
736 aff1 = isl_pw_aff_min(aff1, aff2);
737 else
738 aff1 = isl_pw_aff_max(aff1, aff2);
739 } else if (name == "floord" || name == "ceild") {
740 isl_val *v;
741 Expr *arg2 = expr->getArg(1);
743 if (arg2->getStmtClass() != Stmt::IntegerLiteralClass) {
744 unsupported(expr);
745 return NULL;
747 aff1 = extract_affine(expr->getArg(0));
748 v = extract_int(cast<IntegerLiteral>(arg2));
749 aff1 = isl_pw_aff_scale_down_val(aff1, v);
750 if (name == "floord")
751 aff1 = isl_pw_aff_floor(aff1);
752 else
753 aff1 = isl_pw_aff_ceil(aff1);
754 } else {
755 unsupported(expr);
756 return NULL;
759 return aff1;
762 /* This method is called when we come across an access that is
763 * nested in what is supposed to be an affine expression.
764 * If nesting is allowed, we return a new parameter that corresponds
765 * to this nested access. Otherwise, we simply complain.
767 * Note that we currently don't allow nested accesses themselves
768 * to contain any nested accesses, so we check if we can extract
769 * the access without any nesting and complain if we can't.
771 * The new parameter is resolved in resolve_nested.
773 isl_pw_aff *PetScan::nested_access(Expr *expr)
775 isl_id *id;
776 isl_space *dim;
777 isl_aff *aff;
778 isl_set *dom;
779 isl_multi_pw_aff *index;
781 if (!nesting_enabled) {
782 unsupported(expr);
783 return NULL;
786 allow_nested = false;
787 index = extract_index(expr);
788 allow_nested = true;
789 if (!index) {
790 unsupported(expr);
791 return NULL;
793 isl_multi_pw_aff_free(index);
795 id = isl_id_alloc(ctx, NULL, expr);
796 dim = isl_space_params_alloc(ctx, 1);
798 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
800 dom = isl_set_universe(isl_space_copy(dim));
801 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
802 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
804 return isl_pw_aff_alloc(dom, aff);
807 /* Affine expressions are not supposed to contain array accesses,
808 * but if nesting is allowed, we return a parameter corresponding
809 * to the array access.
811 __isl_give isl_pw_aff *PetScan::extract_affine(ArraySubscriptExpr *expr)
813 return nested_access(expr);
816 /* Extract an affine expression from a conditional operation.
818 __isl_give isl_pw_aff *PetScan::extract_affine(ConditionalOperator *expr)
820 isl_pw_aff *cond, *lhs, *rhs;
822 cond = extract_condition(expr->getCond());
823 lhs = extract_affine(expr->getTrueExpr());
824 rhs = extract_affine(expr->getFalseExpr());
826 return isl_pw_aff_cond(cond, lhs, rhs);
829 /* Extract an affine expression, if possible, from "expr".
830 * Otherwise return NULL.
832 __isl_give isl_pw_aff *PetScan::extract_affine(Expr *expr)
834 switch (expr->getStmtClass()) {
835 case Stmt::ImplicitCastExprClass:
836 return extract_affine(cast<ImplicitCastExpr>(expr));
837 case Stmt::IntegerLiteralClass:
838 return extract_affine(cast<IntegerLiteral>(expr));
839 case Stmt::DeclRefExprClass:
840 return extract_affine(cast<DeclRefExpr>(expr));
841 case Stmt::BinaryOperatorClass:
842 return extract_affine(cast<BinaryOperator>(expr));
843 case Stmt::UnaryOperatorClass:
844 return extract_affine(cast<UnaryOperator>(expr));
845 case Stmt::ParenExprClass:
846 return extract_affine(cast<ParenExpr>(expr));
847 case Stmt::CallExprClass:
848 return extract_affine(cast<CallExpr>(expr));
849 case Stmt::ArraySubscriptExprClass:
850 return extract_affine(cast<ArraySubscriptExpr>(expr));
851 case Stmt::ConditionalOperatorClass:
852 return extract_affine(cast<ConditionalOperator>(expr));
853 default:
854 unsupported(expr);
856 return NULL;
859 __isl_give isl_multi_pw_aff *PetScan::extract_index(ImplicitCastExpr *expr)
861 return extract_index(expr->getSubExpr());
864 /* Return the depth of an array of the given type.
866 static int array_depth(const Type *type)
868 if (type->isPointerType())
869 return 1 + array_depth(type->getPointeeType().getTypePtr());
870 if (type->isArrayType()) {
871 const ArrayType *atype;
872 type = type->getCanonicalTypeInternal().getTypePtr();
873 atype = cast<ArrayType>(type);
874 return 1 + array_depth(atype->getElementType().getTypePtr());
876 return 0;
879 /* Return the depth of the array accessed by the index expression "index".
880 * If "index" is an affine expression, i.e., if it does not access
881 * any array, then return 1.
883 static int extract_depth(__isl_keep isl_multi_pw_aff *index)
885 isl_id *id;
886 ValueDecl *decl;
888 if (!index)
889 return -1;
891 if (!isl_multi_pw_aff_has_tuple_id(index, isl_dim_out))
892 return 1;
894 id = isl_multi_pw_aff_get_tuple_id(index, isl_dim_out);
895 if (!id)
896 return -1;
897 decl = (ValueDecl *) isl_id_get_user(id);
898 isl_id_free(id);
900 return array_depth(decl->getType().getTypePtr());
903 /* Extract an index expression from a reference to a variable.
904 * If the variable has name "A", then the returned index expression
905 * is of the form
907 * { [] -> A[] }
909 __isl_give isl_multi_pw_aff *PetScan::extract_index(DeclRefExpr *expr)
911 return extract_index(expr->getDecl());
914 /* Extract an index expression from a variable.
915 * If the variable has name "A", then the returned index expression
916 * is of the form
918 * { [] -> A[] }
920 __isl_give isl_multi_pw_aff *PetScan::extract_index(ValueDecl *decl)
922 isl_id *id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
923 isl_space *space = isl_space_alloc(ctx, 0, 0, 0);
925 space = isl_space_set_tuple_id(space, isl_dim_out, id);
927 return isl_multi_pw_aff_zero(space);
930 /* Extract an index expression from an integer contant.
931 * If the value of the constant is "v", then the returned access relation
932 * is
934 * { [] -> [v] }
936 __isl_give isl_multi_pw_aff *PetScan::extract_index(IntegerLiteral *expr)
938 isl_multi_pw_aff *mpa;
940 mpa = isl_multi_pw_aff_from_pw_aff(extract_affine(expr));
941 mpa = isl_multi_pw_aff_from_range(mpa);
942 return mpa;
945 /* Try and extract an index expression from the given Expr.
946 * Return NULL if it doesn't work out.
948 __isl_give isl_multi_pw_aff *PetScan::extract_index(Expr *expr)
950 switch (expr->getStmtClass()) {
951 case Stmt::ImplicitCastExprClass:
952 return extract_index(cast<ImplicitCastExpr>(expr));
953 case Stmt::DeclRefExprClass:
954 return extract_index(cast<DeclRefExpr>(expr));
955 case Stmt::ArraySubscriptExprClass:
956 return extract_index(cast<ArraySubscriptExpr>(expr));
957 case Stmt::IntegerLiteralClass:
958 return extract_index(cast<IntegerLiteral>(expr));
959 default:
960 unsupported(expr);
962 return NULL;
965 /* Given a partial index expression "base" and an extra index "index",
966 * append the extra index to "base" and return the result.
967 * Additionally, add the constraints that the extra index is non-negative.
969 static __isl_give isl_multi_pw_aff *subscript(__isl_take isl_multi_pw_aff *base,
970 __isl_take isl_pw_aff *index)
972 isl_id *id;
973 isl_set *domain;
974 isl_multi_pw_aff *access;
976 id = isl_multi_pw_aff_get_tuple_id(base, isl_dim_set);
977 index = isl_pw_aff_from_range(index);
978 domain = isl_pw_aff_nonneg_set(isl_pw_aff_copy(index));
979 index = isl_pw_aff_intersect_domain(index, domain);
980 access = isl_multi_pw_aff_from_pw_aff(index);
981 access = isl_multi_pw_aff_flat_range_product(base, access);
982 access = isl_multi_pw_aff_set_tuple_id(access, isl_dim_set, id);
984 return access;
987 /* Extract an index expression from the given array subscript expression.
988 * If nesting is allowed in general, then we turn it on while
989 * examining the index expression.
991 * We first extract an index expression from the base.
992 * This will result in an index expression with a range that corresponds
993 * to the earlier indices.
994 * We then extract the current index, restrict its domain
995 * to those values that result in a non-negative index and
996 * append the index to the base index expression.
998 __isl_give isl_multi_pw_aff *PetScan::extract_index(ArraySubscriptExpr *expr)
1000 Expr *base = expr->getBase();
1001 Expr *idx = expr->getIdx();
1002 isl_pw_aff *index;
1003 isl_multi_pw_aff *base_access;
1004 isl_multi_pw_aff *access;
1005 bool save_nesting = nesting_enabled;
1007 nesting_enabled = allow_nested;
1009 base_access = extract_index(base);
1010 index = extract_affine(idx);
1012 nesting_enabled = save_nesting;
1014 access = subscript(base_access, index);
1016 return access;
1019 /* Check if "expr" calls function "minmax" with two arguments and if so
1020 * make lhs and rhs refer to these two arguments.
1022 static bool is_minmax(Expr *expr, const char *minmax, Expr *&lhs, Expr *&rhs)
1024 CallExpr *call;
1025 FunctionDecl *fd;
1026 string name;
1028 if (expr->getStmtClass() != Stmt::CallExprClass)
1029 return false;
1031 call = cast<CallExpr>(expr);
1032 fd = call->getDirectCallee();
1033 if (!fd)
1034 return false;
1036 if (call->getNumArgs() != 2)
1037 return false;
1039 name = fd->getDeclName().getAsString();
1040 if (name != minmax)
1041 return false;
1043 lhs = call->getArg(0);
1044 rhs = call->getArg(1);
1046 return true;
1049 /* Check if "expr" is of the form min(lhs, rhs) and if so make
1050 * lhs and rhs refer to the two arguments.
1052 static bool is_min(Expr *expr, Expr *&lhs, Expr *&rhs)
1054 return is_minmax(expr, "min", lhs, rhs);
1057 /* Check if "expr" is of the form max(lhs, rhs) and if so make
1058 * lhs and rhs refer to the two arguments.
1060 static bool is_max(Expr *expr, Expr *&lhs, Expr *&rhs)
1062 return is_minmax(expr, "max", lhs, rhs);
1065 /* Return "lhs && rhs", defined on the shared definition domain.
1067 static __isl_give isl_pw_aff *pw_aff_and(__isl_take isl_pw_aff *lhs,
1068 __isl_take isl_pw_aff *rhs)
1070 isl_set *cond;
1071 isl_set *dom;
1073 dom = isl_set_intersect(isl_pw_aff_domain(isl_pw_aff_copy(lhs)),
1074 isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1075 cond = isl_set_intersect(isl_pw_aff_non_zero_set(lhs),
1076 isl_pw_aff_non_zero_set(rhs));
1077 return indicator_function(cond, dom);
1080 /* Return "lhs && rhs", with shortcut semantics.
1081 * That is, if lhs is false, then the result is defined even if rhs is not.
1082 * In practice, we compute lhs ? rhs : lhs.
1084 static __isl_give isl_pw_aff *pw_aff_and_then(__isl_take isl_pw_aff *lhs,
1085 __isl_take isl_pw_aff *rhs)
1087 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), rhs, lhs);
1090 /* Return "lhs || rhs", with shortcut semantics.
1091 * That is, if lhs is true, then the result is defined even if rhs is not.
1092 * In practice, we compute lhs ? lhs : rhs.
1094 static __isl_give isl_pw_aff *pw_aff_or_else(__isl_take isl_pw_aff *lhs,
1095 __isl_take isl_pw_aff *rhs)
1097 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), lhs, rhs);
1100 /* Extract an affine expressions representing the comparison "LHS op RHS"
1101 * "comp" is the original statement that "LHS op RHS" is derived from
1102 * and is used for diagnostics.
1104 * If the comparison is of the form
1106 * a <= min(b,c)
1108 * then the expression is constructed as the conjunction of
1109 * the comparisons
1111 * a <= b and a <= c
1113 * A similar optimization is performed for max(a,b) <= c.
1114 * We do this because that will lead to simpler representations
1115 * of the expression.
1116 * If isl is ever enhanced to explicitly deal with min and max expressions,
1117 * this optimization can be removed.
1119 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperatorKind op,
1120 Expr *LHS, Expr *RHS, Stmt *comp)
1122 isl_pw_aff *lhs;
1123 isl_pw_aff *rhs;
1124 isl_pw_aff *res;
1125 isl_set *cond;
1126 isl_set *dom;
1128 if (op == BO_GT)
1129 return extract_comparison(BO_LT, RHS, LHS, comp);
1130 if (op == BO_GE)
1131 return extract_comparison(BO_LE, RHS, LHS, comp);
1133 if (op == BO_LT || op == BO_LE) {
1134 Expr *expr1, *expr2;
1135 if (is_min(RHS, expr1, expr2)) {
1136 lhs = extract_comparison(op, LHS, expr1, comp);
1137 rhs = extract_comparison(op, LHS, expr2, comp);
1138 return pw_aff_and(lhs, rhs);
1140 if (is_max(LHS, expr1, expr2)) {
1141 lhs = extract_comparison(op, expr1, RHS, comp);
1142 rhs = extract_comparison(op, expr2, RHS, comp);
1143 return pw_aff_and(lhs, rhs);
1147 lhs = extract_affine(LHS);
1148 rhs = extract_affine(RHS);
1150 dom = isl_pw_aff_domain(isl_pw_aff_copy(lhs));
1151 dom = isl_set_intersect(dom, isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1153 switch (op) {
1154 case BO_LT:
1155 cond = isl_pw_aff_lt_set(lhs, rhs);
1156 break;
1157 case BO_LE:
1158 cond = isl_pw_aff_le_set(lhs, rhs);
1159 break;
1160 case BO_EQ:
1161 cond = isl_pw_aff_eq_set(lhs, rhs);
1162 break;
1163 case BO_NE:
1164 cond = isl_pw_aff_ne_set(lhs, rhs);
1165 break;
1166 default:
1167 isl_pw_aff_free(lhs);
1168 isl_pw_aff_free(rhs);
1169 isl_set_free(dom);
1170 unsupported(comp);
1171 return NULL;
1174 cond = isl_set_coalesce(cond);
1175 res = indicator_function(cond, dom);
1177 return res;
1180 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperator *comp)
1182 return extract_comparison(comp->getOpcode(), comp->getLHS(),
1183 comp->getRHS(), comp);
1186 /* Extract an affine expression representing the negation (logical not)
1187 * of a subexpression.
1189 __isl_give isl_pw_aff *PetScan::extract_boolean(UnaryOperator *op)
1191 isl_set *set_cond, *dom;
1192 isl_pw_aff *cond, *res;
1194 cond = extract_condition(op->getSubExpr());
1196 dom = isl_pw_aff_domain(isl_pw_aff_copy(cond));
1198 set_cond = isl_pw_aff_zero_set(cond);
1200 res = indicator_function(set_cond, dom);
1202 return res;
1205 /* Extract an affine expression representing the disjunction (logical or)
1206 * or conjunction (logical and) of two subexpressions.
1208 __isl_give isl_pw_aff *PetScan::extract_boolean(BinaryOperator *comp)
1210 isl_pw_aff *lhs, *rhs;
1212 lhs = extract_condition(comp->getLHS());
1213 rhs = extract_condition(comp->getRHS());
1215 switch (comp->getOpcode()) {
1216 case BO_LAnd:
1217 return pw_aff_and_then(lhs, rhs);
1218 case BO_LOr:
1219 return pw_aff_or_else(lhs, rhs);
1220 default:
1221 isl_pw_aff_free(lhs);
1222 isl_pw_aff_free(rhs);
1225 unsupported(comp);
1226 return NULL;
1229 __isl_give isl_pw_aff *PetScan::extract_condition(UnaryOperator *expr)
1231 switch (expr->getOpcode()) {
1232 case UO_LNot:
1233 return extract_boolean(expr);
1234 default:
1235 unsupported(expr);
1236 return NULL;
1240 /* Extract the affine expression "expr != 0 ? 1 : 0".
1242 __isl_give isl_pw_aff *PetScan::extract_implicit_condition(Expr *expr)
1244 isl_pw_aff *res;
1245 isl_set *set, *dom;
1247 res = extract_affine(expr);
1249 dom = isl_pw_aff_domain(isl_pw_aff_copy(res));
1250 set = isl_pw_aff_non_zero_set(res);
1252 res = indicator_function(set, dom);
1254 return res;
1257 /* Extract an affine expression from a boolean expression.
1258 * In particular, return the expression "expr ? 1 : 0".
1260 * If the expression doesn't look like a condition, we assume it
1261 * is an affine expression and return the condition "expr != 0 ? 1 : 0".
1263 __isl_give isl_pw_aff *PetScan::extract_condition(Expr *expr)
1265 BinaryOperator *comp;
1267 if (!expr) {
1268 isl_set *u = isl_set_universe(isl_space_params_alloc(ctx, 0));
1269 return indicator_function(u, isl_set_copy(u));
1272 if (expr->getStmtClass() == Stmt::ParenExprClass)
1273 return extract_condition(cast<ParenExpr>(expr)->getSubExpr());
1275 if (expr->getStmtClass() == Stmt::UnaryOperatorClass)
1276 return extract_condition(cast<UnaryOperator>(expr));
1278 if (expr->getStmtClass() != Stmt::BinaryOperatorClass)
1279 return extract_implicit_condition(expr);
1281 comp = cast<BinaryOperator>(expr);
1282 switch (comp->getOpcode()) {
1283 case BO_LT:
1284 case BO_LE:
1285 case BO_GT:
1286 case BO_GE:
1287 case BO_EQ:
1288 case BO_NE:
1289 return extract_comparison(comp);
1290 case BO_LAnd:
1291 case BO_LOr:
1292 return extract_boolean(comp);
1293 default:
1294 return extract_implicit_condition(expr);
1298 static enum pet_op_type UnaryOperatorKind2pet_op_type(UnaryOperatorKind kind)
1300 switch (kind) {
1301 case UO_Minus:
1302 return pet_op_minus;
1303 case UO_PostInc:
1304 return pet_op_post_inc;
1305 case UO_PostDec:
1306 return pet_op_post_dec;
1307 case UO_PreInc:
1308 return pet_op_pre_inc;
1309 case UO_PreDec:
1310 return pet_op_pre_dec;
1311 default:
1312 return pet_op_last;
1316 static enum pet_op_type BinaryOperatorKind2pet_op_type(BinaryOperatorKind kind)
1318 switch (kind) {
1319 case BO_AddAssign:
1320 return pet_op_add_assign;
1321 case BO_SubAssign:
1322 return pet_op_sub_assign;
1323 case BO_MulAssign:
1324 return pet_op_mul_assign;
1325 case BO_DivAssign:
1326 return pet_op_div_assign;
1327 case BO_Assign:
1328 return pet_op_assign;
1329 case BO_Add:
1330 return pet_op_add;
1331 case BO_Sub:
1332 return pet_op_sub;
1333 case BO_Mul:
1334 return pet_op_mul;
1335 case BO_Div:
1336 return pet_op_div;
1337 case BO_Rem:
1338 return pet_op_mod;
1339 case BO_EQ:
1340 return pet_op_eq;
1341 case BO_LE:
1342 return pet_op_le;
1343 case BO_LT:
1344 return pet_op_lt;
1345 case BO_GT:
1346 return pet_op_gt;
1347 default:
1348 return pet_op_last;
1352 /* Construct a pet_expr representing a unary operator expression.
1354 struct pet_expr *PetScan::extract_expr(UnaryOperator *expr)
1356 struct pet_expr *arg;
1357 enum pet_op_type op;
1359 op = UnaryOperatorKind2pet_op_type(expr->getOpcode());
1360 if (op == pet_op_last) {
1361 unsupported(expr);
1362 return NULL;
1365 arg = extract_expr(expr->getSubExpr());
1367 if (expr->isIncrementDecrementOp() &&
1368 arg && arg->type == pet_expr_access) {
1369 mark_write(arg);
1370 arg->acc.read = 1;
1373 return pet_expr_new_unary(ctx, op, arg);
1376 /* Mark the given access pet_expr as a write.
1377 * If a scalar is being accessed, then mark its value
1378 * as unknown in assigned_value.
1380 void PetScan::mark_write(struct pet_expr *access)
1382 isl_id *id;
1383 ValueDecl *decl;
1385 if (!access)
1386 return;
1388 access->acc.write = 1;
1389 access->acc.read = 0;
1391 if (!pet_expr_is_scalar_access(access))
1392 return;
1394 id = pet_expr_access_get_id(access);
1395 decl = (ValueDecl *) isl_id_get_user(id);
1396 clear_assignment(assigned_value, decl);
1397 isl_id_free(id);
1400 /* Assign "rhs" to "lhs".
1402 * In particular, if "lhs" is a scalar variable, then mark
1403 * the variable as having been assigned. If, furthermore, "rhs"
1404 * is an affine expression, then keep track of this value in assigned_value
1405 * so that we can plug it in when we later come across the same variable.
1407 void PetScan::assign(struct pet_expr *lhs, Expr *rhs)
1409 isl_id *id;
1410 ValueDecl *decl;
1411 isl_pw_aff *pa;
1413 if (!lhs)
1414 return;
1415 if (!pet_expr_is_scalar_access(lhs))
1416 return;
1418 id = pet_expr_access_get_id(lhs);
1419 decl = (ValueDecl *) isl_id_get_user(id);
1420 isl_id_free(id);
1422 pa = try_extract_affine(rhs);
1423 clear_assignment(assigned_value, decl);
1424 if (!pa)
1425 return;
1426 assigned_value[decl] = pa;
1427 insert_expression(pa);
1430 /* Construct a pet_expr representing a binary operator expression.
1432 * If the top level operator is an assignment and the LHS is an access,
1433 * then we mark that access as a write. If the operator is a compound
1434 * assignment, the access is marked as both a read and a write.
1436 * If "expr" assigns something to a scalar variable, then we mark
1437 * the variable as having been assigned. If, furthermore, the expression
1438 * is affine, then keep track of this value in assigned_value
1439 * so that we can plug it in when we later come across the same variable.
1441 struct pet_expr *PetScan::extract_expr(BinaryOperator *expr)
1443 struct pet_expr *lhs, *rhs;
1444 enum pet_op_type op;
1446 op = BinaryOperatorKind2pet_op_type(expr->getOpcode());
1447 if (op == pet_op_last) {
1448 unsupported(expr);
1449 return NULL;
1452 lhs = extract_expr(expr->getLHS());
1453 rhs = extract_expr(expr->getRHS());
1455 if (expr->isAssignmentOp() && lhs && lhs->type == pet_expr_access) {
1456 mark_write(lhs);
1457 if (expr->isCompoundAssignmentOp())
1458 lhs->acc.read = 1;
1461 if (expr->getOpcode() == BO_Assign)
1462 assign(lhs, expr->getRHS());
1464 return pet_expr_new_binary(ctx, op, lhs, rhs);
1467 /* Construct a pet_scop with a single statement killing the entire
1468 * array "array".
1470 struct pet_scop *PetScan::kill(Stmt *stmt, struct pet_array *array)
1472 isl_id *id;
1473 isl_space *space;
1474 isl_multi_pw_aff *index;
1475 isl_map *access;
1476 struct pet_expr *expr;
1478 if (!array)
1479 return NULL;
1480 access = isl_map_from_range(isl_set_copy(array->extent));
1481 id = isl_set_get_tuple_id(array->extent);
1482 space = isl_space_alloc(ctx, 0, 0, 0);
1483 space = isl_space_set_tuple_id(space, isl_dim_out, id);
1484 index = isl_multi_pw_aff_zero(space);
1485 expr = pet_expr_kill_from_access_and_index(access, index);
1486 return extract(stmt, expr);
1489 /* Construct a pet_scop for a (single) variable declaration.
1491 * The scop contains the variable being declared (as an array)
1492 * and a statement killing the array.
1494 * If the variable is initialized in the AST, then the scop
1495 * also contains an assignment to the variable.
1497 struct pet_scop *PetScan::extract(DeclStmt *stmt)
1499 Decl *decl;
1500 VarDecl *vd;
1501 struct pet_expr *lhs, *rhs, *pe;
1502 struct pet_scop *scop_decl, *scop;
1503 struct pet_array *array;
1505 if (!stmt->isSingleDecl()) {
1506 unsupported(stmt);
1507 return NULL;
1510 decl = stmt->getSingleDecl();
1511 vd = cast<VarDecl>(decl);
1513 array = extract_array(ctx, vd);
1514 if (array)
1515 array->declared = 1;
1516 scop_decl = kill(stmt, array);
1517 scop_decl = pet_scop_add_array(scop_decl, array);
1519 if (!vd->getInit())
1520 return scop_decl;
1522 lhs = extract_access_expr(vd);
1523 rhs = extract_expr(vd->getInit());
1525 mark_write(lhs);
1526 assign(lhs, vd->getInit());
1528 pe = pet_expr_new_binary(ctx, pet_op_assign, lhs, rhs);
1529 scop = extract(stmt, pe);
1531 scop_decl = pet_scop_prefix(scop_decl, 0);
1532 scop = pet_scop_prefix(scop, 1);
1534 scop = pet_scop_add_seq(ctx, scop_decl, scop);
1536 return scop;
1539 /* Construct a pet_expr representing a conditional operation.
1541 * We first try to extract the condition as an affine expression.
1542 * If that fails, we construct a pet_expr tree representing the condition.
1544 struct pet_expr *PetScan::extract_expr(ConditionalOperator *expr)
1546 struct pet_expr *cond, *lhs, *rhs;
1547 isl_pw_aff *pa;
1549 pa = try_extract_affine(expr->getCond());
1550 if (pa) {
1551 isl_multi_pw_aff *test = isl_multi_pw_aff_from_pw_aff(pa);
1552 test = isl_multi_pw_aff_from_range(test);
1553 cond = pet_expr_from_index(test);
1554 } else
1555 cond = extract_expr(expr->getCond());
1556 lhs = extract_expr(expr->getTrueExpr());
1557 rhs = extract_expr(expr->getFalseExpr());
1559 return pet_expr_new_ternary(ctx, cond, lhs, rhs);
1562 struct pet_expr *PetScan::extract_expr(ImplicitCastExpr *expr)
1564 return extract_expr(expr->getSubExpr());
1567 /* Construct a pet_expr representing a floating point value.
1569 * If the floating point literal does not appear in a macro,
1570 * then we use the original representation in the source code
1571 * as the string representation. Otherwise, we use the pretty
1572 * printer to produce a string representation.
1574 struct pet_expr *PetScan::extract_expr(FloatingLiteral *expr)
1576 double d;
1577 string s;
1578 const LangOptions &LO = PP.getLangOpts();
1579 SourceLocation loc = expr->getLocation();
1581 if (!loc.isMacroID()) {
1582 SourceManager &SM = PP.getSourceManager();
1583 unsigned len = Lexer::MeasureTokenLength(loc, SM, LO);
1584 s = string(SM.getCharacterData(loc), len);
1585 } else {
1586 llvm::raw_string_ostream S(s);
1587 expr->printPretty(S, 0, PrintingPolicy(LO));
1588 S.str();
1590 d = expr->getValueAsApproximateDouble();
1591 return pet_expr_new_double(ctx, d, s.c_str());
1594 /* Extract an index expression from "expr" and then convert it into
1595 * an access pet_expr.
1597 struct pet_expr *PetScan::extract_access_expr(Expr *expr)
1599 isl_multi_pw_aff *index;
1600 struct pet_expr *pe;
1601 int depth;
1603 index = extract_index(expr);
1604 depth = extract_depth(index);
1606 pe = pet_expr_from_index_and_depth(index, depth);
1608 return pe;
1611 /* Extract an index expression from "decl" and then convert it into
1612 * an access pet_expr.
1614 struct pet_expr *PetScan::extract_access_expr(ValueDecl *decl)
1616 isl_multi_pw_aff *index;
1617 struct pet_expr *pe;
1618 int depth;
1620 index = extract_index(decl);
1621 depth = extract_depth(index);
1623 pe = pet_expr_from_index_and_depth(index, depth);
1625 return pe;
1628 struct pet_expr *PetScan::extract_expr(ParenExpr *expr)
1630 return extract_expr(expr->getSubExpr());
1633 /* Construct a pet_expr representing a function call.
1635 * If we are passing along a pointer to an array element
1636 * or an entire row or even higher dimensional slice of an array,
1637 * then the function being called may write into the array.
1639 * We assume here that if the function is declared to take a pointer
1640 * to a const type, then the function will perform a read
1641 * and that otherwise, it will perform a write.
1643 struct pet_expr *PetScan::extract_expr(CallExpr *expr)
1645 struct pet_expr *res = NULL;
1646 FunctionDecl *fd;
1647 string name;
1649 fd = expr->getDirectCallee();
1650 if (!fd) {
1651 unsupported(expr);
1652 return NULL;
1655 name = fd->getDeclName().getAsString();
1656 res = pet_expr_new_call(ctx, name.c_str(), expr->getNumArgs());
1657 if (!res)
1658 return NULL;
1660 for (int i = 0; i < expr->getNumArgs(); ++i) {
1661 Expr *arg = expr->getArg(i);
1662 int is_addr = 0;
1663 pet_expr *main_arg;
1665 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
1666 ImplicitCastExpr *ice = cast<ImplicitCastExpr>(arg);
1667 arg = ice->getSubExpr();
1669 if (arg->getStmtClass() == Stmt::UnaryOperatorClass) {
1670 UnaryOperator *op = cast<UnaryOperator>(arg);
1671 if (op->getOpcode() == UO_AddrOf) {
1672 is_addr = 1;
1673 arg = op->getSubExpr();
1676 res->args[i] = PetScan::extract_expr(arg);
1677 main_arg = res->args[i];
1678 if (is_addr)
1679 res->args[i] = pet_expr_new_unary(ctx,
1680 pet_op_address_of, res->args[i]);
1681 if (!res->args[i])
1682 goto error;
1683 if (arg->getStmtClass() == Stmt::ArraySubscriptExprClass &&
1684 array_depth(arg->getType().getTypePtr()) > 0)
1685 is_addr = 1;
1686 if (is_addr && main_arg->type == pet_expr_access) {
1687 ParmVarDecl *parm;
1688 if (!fd->hasPrototype()) {
1689 unsupported(expr, "prototype required");
1690 goto error;
1692 parm = fd->getParamDecl(i);
1693 if (!const_base(parm->getType()))
1694 mark_write(main_arg);
1698 return res;
1699 error:
1700 pet_expr_free(res);
1701 return NULL;
1704 /* Construct a pet_expr representing a (C style) cast.
1706 struct pet_expr *PetScan::extract_expr(CStyleCastExpr *expr)
1708 struct pet_expr *arg;
1709 QualType type;
1711 arg = extract_expr(expr->getSubExpr());
1712 if (!arg)
1713 return NULL;
1715 type = expr->getTypeAsWritten();
1716 return pet_expr_new_cast(ctx, type.getAsString().c_str(), arg);
1719 /* Try and onstruct a pet_expr representing "expr".
1721 struct pet_expr *PetScan::extract_expr(Expr *expr)
1723 switch (expr->getStmtClass()) {
1724 case Stmt::UnaryOperatorClass:
1725 return extract_expr(cast<UnaryOperator>(expr));
1726 case Stmt::CompoundAssignOperatorClass:
1727 case Stmt::BinaryOperatorClass:
1728 return extract_expr(cast<BinaryOperator>(expr));
1729 case Stmt::ImplicitCastExprClass:
1730 return extract_expr(cast<ImplicitCastExpr>(expr));
1731 case Stmt::ArraySubscriptExprClass:
1732 case Stmt::DeclRefExprClass:
1733 case Stmt::IntegerLiteralClass:
1734 return extract_access_expr(expr);
1735 case Stmt::FloatingLiteralClass:
1736 return extract_expr(cast<FloatingLiteral>(expr));
1737 case Stmt::ParenExprClass:
1738 return extract_expr(cast<ParenExpr>(expr));
1739 case Stmt::ConditionalOperatorClass:
1740 return extract_expr(cast<ConditionalOperator>(expr));
1741 case Stmt::CallExprClass:
1742 return extract_expr(cast<CallExpr>(expr));
1743 case Stmt::CStyleCastExprClass:
1744 return extract_expr(cast<CStyleCastExpr>(expr));
1745 default:
1746 unsupported(expr);
1748 return NULL;
1751 /* Check if the given initialization statement is an assignment.
1752 * If so, return that assignment. Otherwise return NULL.
1754 BinaryOperator *PetScan::initialization_assignment(Stmt *init)
1756 BinaryOperator *ass;
1758 if (init->getStmtClass() != Stmt::BinaryOperatorClass)
1759 return NULL;
1761 ass = cast<BinaryOperator>(init);
1762 if (ass->getOpcode() != BO_Assign)
1763 return NULL;
1765 return ass;
1768 /* Check if the given initialization statement is a declaration
1769 * of a single variable.
1770 * If so, return that declaration. Otherwise return NULL.
1772 Decl *PetScan::initialization_declaration(Stmt *init)
1774 DeclStmt *decl;
1776 if (init->getStmtClass() != Stmt::DeclStmtClass)
1777 return NULL;
1779 decl = cast<DeclStmt>(init);
1781 if (!decl->isSingleDecl())
1782 return NULL;
1784 return decl->getSingleDecl();
1787 /* Given the assignment operator in the initialization of a for loop,
1788 * extract the induction variable, i.e., the (integer)variable being
1789 * assigned.
1791 ValueDecl *PetScan::extract_induction_variable(BinaryOperator *init)
1793 Expr *lhs;
1794 DeclRefExpr *ref;
1795 ValueDecl *decl;
1796 const Type *type;
1798 lhs = init->getLHS();
1799 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1800 unsupported(init);
1801 return NULL;
1804 ref = cast<DeclRefExpr>(lhs);
1805 decl = ref->getDecl();
1806 type = decl->getType().getTypePtr();
1808 if (!type->isIntegerType()) {
1809 unsupported(lhs);
1810 return NULL;
1813 return decl;
1816 /* Given the initialization statement of a for loop and the single
1817 * declaration in this initialization statement,
1818 * extract the induction variable, i.e., the (integer) variable being
1819 * declared.
1821 VarDecl *PetScan::extract_induction_variable(Stmt *init, Decl *decl)
1823 VarDecl *vd;
1825 vd = cast<VarDecl>(decl);
1827 const QualType type = vd->getType();
1828 if (!type->isIntegerType()) {
1829 unsupported(init);
1830 return NULL;
1833 if (!vd->getInit()) {
1834 unsupported(init);
1835 return NULL;
1838 return vd;
1841 /* Check that op is of the form iv++ or iv--.
1842 * Return an affine expression "1" or "-1" accordingly.
1844 __isl_give isl_pw_aff *PetScan::extract_unary_increment(
1845 clang::UnaryOperator *op, clang::ValueDecl *iv)
1847 Expr *sub;
1848 DeclRefExpr *ref;
1849 isl_space *space;
1850 isl_aff *aff;
1852 if (!op->isIncrementDecrementOp()) {
1853 unsupported(op);
1854 return NULL;
1857 sub = op->getSubExpr();
1858 if (sub->getStmtClass() != Stmt::DeclRefExprClass) {
1859 unsupported(op);
1860 return NULL;
1863 ref = cast<DeclRefExpr>(sub);
1864 if (ref->getDecl() != iv) {
1865 unsupported(op);
1866 return NULL;
1869 space = isl_space_params_alloc(ctx, 0);
1870 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
1872 if (op->isIncrementOp())
1873 aff = isl_aff_add_constant_si(aff, 1);
1874 else
1875 aff = isl_aff_add_constant_si(aff, -1);
1877 return isl_pw_aff_from_aff(aff);
1880 /* If the isl_pw_aff on which isl_pw_aff_foreach_piece is called
1881 * has a single constant expression, then put this constant in *user.
1882 * The caller is assumed to have checked that this function will
1883 * be called exactly once.
1885 static int extract_cst(__isl_take isl_set *set, __isl_take isl_aff *aff,
1886 void *user)
1888 isl_val **inc = (isl_val **)user;
1889 int res = 0;
1891 if (isl_aff_is_cst(aff))
1892 *inc = isl_aff_get_constant_val(aff);
1893 else
1894 res = -1;
1896 isl_set_free(set);
1897 isl_aff_free(aff);
1899 return res;
1902 /* Check if op is of the form
1904 * iv = iv + inc
1906 * and return inc as an affine expression.
1908 * We extract an affine expression from the RHS, subtract iv and return
1909 * the result.
1911 __isl_give isl_pw_aff *PetScan::extract_binary_increment(BinaryOperator *op,
1912 clang::ValueDecl *iv)
1914 Expr *lhs;
1915 DeclRefExpr *ref;
1916 isl_id *id;
1917 isl_space *dim;
1918 isl_aff *aff;
1919 isl_pw_aff *val;
1921 if (op->getOpcode() != BO_Assign) {
1922 unsupported(op);
1923 return NULL;
1926 lhs = op->getLHS();
1927 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1928 unsupported(op);
1929 return NULL;
1932 ref = cast<DeclRefExpr>(lhs);
1933 if (ref->getDecl() != iv) {
1934 unsupported(op);
1935 return NULL;
1938 val = extract_affine(op->getRHS());
1940 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
1942 dim = isl_space_params_alloc(ctx, 1);
1943 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
1944 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
1945 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
1947 val = isl_pw_aff_sub(val, isl_pw_aff_from_aff(aff));
1949 return val;
1952 /* Check that op is of the form iv += cst or iv -= cst
1953 * and return an affine expression corresponding oto cst or -cst accordingly.
1955 __isl_give isl_pw_aff *PetScan::extract_compound_increment(
1956 CompoundAssignOperator *op, clang::ValueDecl *iv)
1958 Expr *lhs;
1959 DeclRefExpr *ref;
1960 bool neg = false;
1961 isl_pw_aff *val;
1962 BinaryOperatorKind opcode;
1964 opcode = op->getOpcode();
1965 if (opcode != BO_AddAssign && opcode != BO_SubAssign) {
1966 unsupported(op);
1967 return NULL;
1969 if (opcode == BO_SubAssign)
1970 neg = true;
1972 lhs = op->getLHS();
1973 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1974 unsupported(op);
1975 return NULL;
1978 ref = cast<DeclRefExpr>(lhs);
1979 if (ref->getDecl() != iv) {
1980 unsupported(op);
1981 return NULL;
1984 val = extract_affine(op->getRHS());
1985 if (neg)
1986 val = isl_pw_aff_neg(val);
1988 return val;
1991 /* Check that the increment of the given for loop increments
1992 * (or decrements) the induction variable "iv" and return
1993 * the increment as an affine expression if successful.
1995 __isl_give isl_pw_aff *PetScan::extract_increment(clang::ForStmt *stmt,
1996 ValueDecl *iv)
1998 Stmt *inc = stmt->getInc();
2000 if (!inc) {
2001 unsupported(stmt);
2002 return NULL;
2005 if (inc->getStmtClass() == Stmt::UnaryOperatorClass)
2006 return extract_unary_increment(cast<UnaryOperator>(inc), iv);
2007 if (inc->getStmtClass() == Stmt::CompoundAssignOperatorClass)
2008 return extract_compound_increment(
2009 cast<CompoundAssignOperator>(inc), iv);
2010 if (inc->getStmtClass() == Stmt::BinaryOperatorClass)
2011 return extract_binary_increment(cast<BinaryOperator>(inc), iv);
2013 unsupported(inc);
2014 return NULL;
2017 /* Embed the given iteration domain in an extra outer loop
2018 * with induction variable "var".
2019 * If this variable appeared as a parameter in the constraints,
2020 * it is replaced by the new outermost dimension.
2022 static __isl_give isl_set *embed(__isl_take isl_set *set,
2023 __isl_take isl_id *var)
2025 int pos;
2027 set = isl_set_insert_dims(set, isl_dim_set, 0, 1);
2028 pos = isl_set_find_dim_by_id(set, isl_dim_param, var);
2029 if (pos >= 0) {
2030 set = isl_set_equate(set, isl_dim_param, pos, isl_dim_set, 0);
2031 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2034 isl_id_free(var);
2035 return set;
2038 /* Return those elements in the space of "cond" that come after
2039 * (based on "sign") an element in "cond".
2041 static __isl_give isl_set *after(__isl_take isl_set *cond, int sign)
2043 isl_map *previous_to_this;
2045 if (sign > 0)
2046 previous_to_this = isl_map_lex_lt(isl_set_get_space(cond));
2047 else
2048 previous_to_this = isl_map_lex_gt(isl_set_get_space(cond));
2050 cond = isl_set_apply(cond, previous_to_this);
2052 return cond;
2055 /* Create the infinite iteration domain
2057 * { [id] : id >= 0 }
2059 * If "scop" has an affine skip of type pet_skip_later,
2060 * then remove those iterations i that have an earlier iteration
2061 * where the skip condition is satisfied, meaning that iteration i
2062 * is not executed.
2063 * Since we are dealing with a loop without loop iterator,
2064 * the skip condition cannot refer to the current loop iterator and
2065 * so effectively, the returned set is of the form
2067 * { [0]; [id] : id >= 1 and not skip }
2069 static __isl_give isl_set *infinite_domain(__isl_take isl_id *id,
2070 struct pet_scop *scop)
2072 isl_ctx *ctx = isl_id_get_ctx(id);
2073 isl_set *domain;
2074 isl_set *skip;
2076 domain = isl_set_nat_universe(isl_space_set_alloc(ctx, 0, 1));
2077 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, id);
2079 if (!pet_scop_has_affine_skip(scop, pet_skip_later))
2080 return domain;
2082 skip = pet_scop_get_affine_skip_domain(scop, pet_skip_later);
2083 skip = embed(skip, isl_id_copy(id));
2084 skip = isl_set_intersect(skip , isl_set_copy(domain));
2085 domain = isl_set_subtract(domain, after(skip, 1));
2087 return domain;
2090 /* Create an identity affine expression on the space containing "domain",
2091 * which is assumed to be one-dimensional.
2093 static __isl_give isl_aff *identity_aff(__isl_keep isl_set *domain)
2095 isl_local_space *ls;
2097 ls = isl_local_space_from_space(isl_set_get_space(domain));
2098 return isl_aff_var_on_domain(ls, isl_dim_set, 0);
2101 /* Create an affine expression that maps elements
2102 * of a single-dimensional array "id_test" to the previous element
2103 * (according to "inc"), provided this element belongs to "domain".
2104 * That is, create the affine expression
2106 * { id[x] -> id[x - inc] : x - inc in domain }
2108 static __isl_give isl_multi_pw_aff *map_to_previous(__isl_take isl_id *id_test,
2109 __isl_take isl_set *domain, __isl_take isl_val *inc)
2111 isl_space *space;
2112 isl_local_space *ls;
2113 isl_aff *aff;
2114 isl_multi_pw_aff *prev;
2116 space = isl_set_get_space(domain);
2117 ls = isl_local_space_from_space(space);
2118 aff = isl_aff_var_on_domain(ls, isl_dim_set, 0);
2119 aff = isl_aff_add_constant_val(aff, isl_val_neg(inc));
2120 prev = isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
2121 domain = isl_set_preimage_multi_pw_aff(domain,
2122 isl_multi_pw_aff_copy(prev));
2123 prev = isl_multi_pw_aff_intersect_domain(prev, domain);
2124 prev = isl_multi_pw_aff_set_tuple_id(prev, isl_dim_out, id_test);
2126 return prev;
2129 /* Add an implication to "scop" expressing that if an element of
2130 * virtual array "id_test" has value "satisfied" then all previous elements
2131 * of this array also have that value. The set of previous elements
2132 * is bounded by "domain". If "sign" is negative then iterator
2133 * is decreasing and we express that all subsequent array elements
2134 * (but still defined previously) have the same value.
2136 static struct pet_scop *add_implication(struct pet_scop *scop,
2137 __isl_take isl_id *id_test, __isl_take isl_set *domain, int sign,
2138 int satisfied)
2140 isl_space *space;
2141 isl_map *map;
2143 domain = isl_set_set_tuple_id(domain, id_test);
2144 space = isl_set_get_space(domain);
2145 if (sign > 0)
2146 map = isl_map_lex_ge(space);
2147 else
2148 map = isl_map_lex_le(space);
2149 map = isl_map_intersect_range(map, domain);
2150 scop = pet_scop_add_implication(scop, map, satisfied);
2152 return scop;
2155 /* Add a filter to "scop" that imposes that it is only executed
2156 * when the variable identified by "id_test" has a zero value
2157 * for all previous iterations of "domain".
2159 * In particular, add a filter that imposes that the array
2160 * has a zero value at the previous iteration of domain and
2161 * add an implication that implies that it then has that
2162 * value for all previous iterations.
2164 static struct pet_scop *scop_add_break(struct pet_scop *scop,
2165 __isl_take isl_id *id_test, __isl_take isl_set *domain,
2166 __isl_take isl_val *inc)
2168 isl_multi_pw_aff *prev;
2169 int sign = isl_val_sgn(inc);
2171 prev = map_to_previous(isl_id_copy(id_test), isl_set_copy(domain), inc);
2172 scop = add_implication(scop, id_test, domain, sign, 0);
2173 scop = pet_scop_filter(scop, prev, 0);
2175 return scop;
2178 /* Construct a pet_scop for an infinite loop around the given body.
2180 * We extract a pet_scop for the body and then embed it in a loop with
2181 * iteration domain
2183 * { [t] : t >= 0 }
2185 * and schedule
2187 * { [t] -> [t] }
2189 * If the body contains any break, then it is taken into
2190 * account in infinite_domain (if the skip condition is affine)
2191 * or in scop_add_break (if the skip condition is not affine).
2193 struct pet_scop *PetScan::extract_infinite_loop(Stmt *body)
2195 isl_id *id, *id_test;
2196 isl_set *domain;
2197 isl_aff *ident;
2198 struct pet_scop *scop;
2199 bool has_var_break;
2201 scop = extract(body);
2202 if (!scop)
2203 return NULL;
2205 id = isl_id_alloc(ctx, "t", NULL);
2206 domain = infinite_domain(isl_id_copy(id), scop);
2207 ident = identity_aff(domain);
2209 has_var_break = pet_scop_has_var_skip(scop, pet_skip_later);
2210 if (has_var_break)
2211 id_test = pet_scop_get_skip_id(scop, pet_skip_later);
2213 scop = pet_scop_embed(scop, isl_set_copy(domain),
2214 isl_map_from_aff(isl_aff_copy(ident)), ident, id);
2215 if (has_var_break)
2216 scop = scop_add_break(scop, id_test, domain, isl_val_one(ctx));
2217 else
2218 isl_set_free(domain);
2220 return scop;
2223 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
2225 * for (;;)
2226 * body
2229 struct pet_scop *PetScan::extract_infinite_for(ForStmt *stmt)
2231 return extract_infinite_loop(stmt->getBody());
2234 /* Create an index expression for an access to a virtual array
2235 * representing the result of a condition.
2236 * Unlike other accessed data, the id of the array is NULL as
2237 * there is no ValueDecl in the program corresponding to the virtual
2238 * array.
2239 * The array starts out as a scalar, but grows along with the
2240 * statement writing to the array in pet_scop_embed.
2242 static __isl_give isl_multi_pw_aff *create_test_index(isl_ctx *ctx, int test_nr)
2244 isl_space *dim = isl_space_alloc(ctx, 0, 0, 0);
2245 isl_id *id;
2246 char name[50];
2248 snprintf(name, sizeof(name), "__pet_test_%d", test_nr);
2249 id = isl_id_alloc(ctx, name, NULL);
2250 dim = isl_space_set_tuple_id(dim, isl_dim_out, id);
2251 return isl_multi_pw_aff_zero(dim);
2254 /* Add an array with the given extent (range of "index") to the list
2255 * of arrays in "scop" and return the extended pet_scop.
2256 * The array is marked as attaining values 0 and 1 only and
2257 * as each element being assigned at most once.
2259 static struct pet_scop *scop_add_array(struct pet_scop *scop,
2260 __isl_keep isl_multi_pw_aff *index, clang::ASTContext &ast_ctx)
2262 isl_ctx *ctx = isl_multi_pw_aff_get_ctx(index);
2263 isl_space *dim;
2264 struct pet_array *array;
2265 isl_map *access;
2267 if (!scop)
2268 return NULL;
2269 if (!ctx)
2270 goto error;
2272 array = isl_calloc_type(ctx, struct pet_array);
2273 if (!array)
2274 goto error;
2276 access = isl_map_from_multi_pw_aff(isl_multi_pw_aff_copy(index));
2277 array->extent = isl_map_range(access);
2278 dim = isl_space_params_alloc(ctx, 0);
2279 array->context = isl_set_universe(dim);
2280 dim = isl_space_set_alloc(ctx, 0, 1);
2281 array->value_bounds = isl_set_universe(dim);
2282 array->value_bounds = isl_set_lower_bound_si(array->value_bounds,
2283 isl_dim_set, 0, 0);
2284 array->value_bounds = isl_set_upper_bound_si(array->value_bounds,
2285 isl_dim_set, 0, 1);
2286 array->element_type = strdup("int");
2287 array->element_size = ast_ctx.getTypeInfo(ast_ctx.IntTy).first / 8;
2288 array->uniquely_defined = 1;
2290 if (!array->extent || !array->context)
2291 array = pet_array_free(array);
2293 scop = pet_scop_add_array(scop, array);
2295 return scop;
2296 error:
2297 pet_scop_free(scop);
2298 return NULL;
2301 /* Construct a pet_scop for a while loop of the form
2303 * while (pa)
2304 * body
2306 * In particular, construct a scop for an infinite loop around body and
2307 * intersect the domain with the affine expression.
2308 * Note that this intersection may result in an empty loop.
2310 struct pet_scop *PetScan::extract_affine_while(__isl_take isl_pw_aff *pa,
2311 Stmt *body)
2313 struct pet_scop *scop;
2314 isl_set *dom;
2315 isl_set *valid;
2317 valid = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2318 dom = isl_pw_aff_non_zero_set(pa);
2319 scop = extract_infinite_loop(body);
2320 scop = pet_scop_restrict(scop, dom);
2321 scop = pet_scop_restrict_context(scop, valid);
2323 return scop;
2326 /* Construct a scop for a while, given the scops for the condition
2327 * and the body, the filter identifier and the iteration domain of
2328 * the while loop.
2330 * In particular, the scop for the condition is filtered to depend
2331 * on "id_test" evaluating to true for all previous iterations
2332 * of the loop, while the scop for the body is filtered to depend
2333 * on "id_test" evaluating to true for all iterations up to the
2334 * current iteration.
2335 * The actual filter only imposes that this virtual array has
2336 * value one on the previous or the current iteration.
2337 * The fact that this condition also applies to the previous
2338 * iterations is enforced by an implication.
2340 * These filtered scops are then combined into a single scop.
2342 * "sign" is positive if the iterator increases and negative
2343 * if it decreases.
2345 static struct pet_scop *scop_add_while(struct pet_scop *scop_cond,
2346 struct pet_scop *scop_body, __isl_take isl_id *id_test,
2347 __isl_take isl_set *domain, __isl_take isl_val *inc)
2349 isl_ctx *ctx = isl_set_get_ctx(domain);
2350 isl_space *space;
2351 isl_multi_pw_aff *test_index;
2352 isl_multi_pw_aff *prev;
2353 int sign = isl_val_sgn(inc);
2354 struct pet_scop *scop;
2356 prev = map_to_previous(isl_id_copy(id_test), isl_set_copy(domain), inc);
2357 scop_cond = pet_scop_filter(scop_cond, prev, 1);
2359 space = isl_space_map_from_set(isl_set_get_space(domain));
2360 test_index = isl_multi_pw_aff_identity(space);
2361 test_index = isl_multi_pw_aff_set_tuple_id(test_index, isl_dim_out,
2362 isl_id_copy(id_test));
2363 scop_body = pet_scop_filter(scop_body, test_index, 1);
2365 scop = pet_scop_add_seq(ctx, scop_cond, scop_body);
2366 scop = add_implication(scop, id_test, domain, sign, 1);
2368 return scop;
2371 /* Check if the while loop is of the form
2373 * while (affine expression)
2374 * body
2376 * If so, call extract_affine_while to construct a scop.
2378 * Otherwise, construct a generic while scop, with iteration domain
2379 * { [t] : t >= 0 }. The scop consists of two parts, one for
2380 * evaluating the condition and one for the body.
2381 * The schedule is adjusted to reflect that the condition is evaluated
2382 * before the body is executed and the body is filtered to depend
2383 * on the result of the condition evaluating to true on all iterations
2384 * up to the current iteration, while the evaluation the condition itself
2385 * is filtered to depend on the result of the condition evaluating to true
2386 * on all previous iterations.
2387 * The context of the scop representing the body is dropped
2388 * because we don't know how many times the body will be executed,
2389 * if at all.
2391 * If the body contains any break, then it is taken into
2392 * account in infinite_domain (if the skip condition is affine)
2393 * or in scop_add_break (if the skip condition is not affine).
2395 struct pet_scop *PetScan::extract(WhileStmt *stmt)
2397 Expr *cond;
2398 isl_id *id, *id_test, *id_break_test;
2399 isl_multi_pw_aff *test_index;
2400 isl_set *domain;
2401 isl_aff *ident;
2402 isl_pw_aff *pa;
2403 struct pet_scop *scop, *scop_body;
2404 bool has_var_break;
2406 cond = stmt->getCond();
2407 if (!cond) {
2408 unsupported(stmt);
2409 return NULL;
2412 clear_assignments clear(assigned_value);
2413 clear.TraverseStmt(stmt->getBody());
2415 pa = try_extract_affine_condition(cond);
2416 if (pa)
2417 return extract_affine_while(pa, stmt->getBody());
2419 if (!allow_nested) {
2420 unsupported(stmt);
2421 return NULL;
2424 test_index = create_test_index(ctx, n_test++);
2425 scop = extract_non_affine_condition(cond,
2426 isl_multi_pw_aff_copy(test_index));
2427 scop = scop_add_array(scop, test_index, ast_context);
2428 id_test = isl_multi_pw_aff_get_tuple_id(test_index, isl_dim_out);
2429 isl_multi_pw_aff_free(test_index);
2430 scop_body = extract(stmt->getBody());
2432 id = isl_id_alloc(ctx, "t", NULL);
2433 domain = infinite_domain(isl_id_copy(id), scop_body);
2434 ident = identity_aff(domain);
2436 has_var_break = pet_scop_has_var_skip(scop_body, pet_skip_later);
2437 if (has_var_break)
2438 id_break_test = pet_scop_get_skip_id(scop_body, pet_skip_later);
2440 scop = pet_scop_prefix(scop, 0);
2441 scop = pet_scop_embed(scop, isl_set_copy(domain),
2442 isl_map_from_aff(isl_aff_copy(ident)),
2443 isl_aff_copy(ident), isl_id_copy(id));
2444 scop_body = pet_scop_reset_context(scop_body);
2445 scop_body = pet_scop_prefix(scop_body, 1);
2446 scop_body = pet_scop_embed(scop_body, isl_set_copy(domain),
2447 isl_map_from_aff(isl_aff_copy(ident)), ident, id);
2449 if (has_var_break) {
2450 scop = scop_add_break(scop, isl_id_copy(id_break_test),
2451 isl_set_copy(domain), isl_val_one(ctx));
2452 scop_body = scop_add_break(scop_body, id_break_test,
2453 isl_set_copy(domain), isl_val_one(ctx));
2455 scop = scop_add_while(scop, scop_body, id_test, domain,
2456 isl_val_one(ctx));
2458 return scop;
2461 /* Check whether "cond" expresses a simple loop bound
2462 * on the only set dimension.
2463 * In particular, if "up" is set then "cond" should contain only
2464 * upper bounds on the set dimension.
2465 * Otherwise, it should contain only lower bounds.
2467 static bool is_simple_bound(__isl_keep isl_set *cond, __isl_keep isl_val *inc)
2469 if (isl_val_is_pos(inc))
2470 return !isl_set_dim_has_any_lower_bound(cond, isl_dim_set, 0);
2471 else
2472 return !isl_set_dim_has_any_upper_bound(cond, isl_dim_set, 0);
2475 /* Extend a condition on a given iteration of a loop to one that
2476 * imposes the same condition on all previous iterations.
2477 * "domain" expresses the lower [upper] bound on the iterations
2478 * when inc is positive [negative].
2480 * In particular, we construct the condition (when inc is positive)
2482 * forall i' : (domain(i') and i' <= i) => cond(i')
2484 * which is equivalent to
2486 * not exists i' : domain(i') and i' <= i and not cond(i')
2488 * We construct this set by negating cond, applying a map
2490 * { [i'] -> [i] : domain(i') and i' <= i }
2492 * and then negating the result again.
2494 static __isl_give isl_set *valid_for_each_iteration(__isl_take isl_set *cond,
2495 __isl_take isl_set *domain, __isl_take isl_val *inc)
2497 isl_map *previous_to_this;
2499 if (isl_val_is_pos(inc))
2500 previous_to_this = isl_map_lex_le(isl_set_get_space(domain));
2501 else
2502 previous_to_this = isl_map_lex_ge(isl_set_get_space(domain));
2504 previous_to_this = isl_map_intersect_domain(previous_to_this, domain);
2506 cond = isl_set_complement(cond);
2507 cond = isl_set_apply(cond, previous_to_this);
2508 cond = isl_set_complement(cond);
2510 isl_val_free(inc);
2512 return cond;
2515 /* Construct a domain of the form
2517 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
2519 static __isl_give isl_set *strided_domain(__isl_take isl_id *id,
2520 __isl_take isl_pw_aff *init, __isl_take isl_val *inc)
2522 isl_aff *aff;
2523 isl_space *dim;
2524 isl_set *set;
2526 init = isl_pw_aff_insert_dims(init, isl_dim_in, 0, 1);
2527 dim = isl_pw_aff_get_domain_space(init);
2528 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2529 aff = isl_aff_add_coefficient_val(aff, isl_dim_in, 0, inc);
2530 init = isl_pw_aff_add(init, isl_pw_aff_from_aff(aff));
2532 dim = isl_space_set_alloc(isl_pw_aff_get_ctx(init), 1, 1);
2533 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
2534 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2535 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2537 set = isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff), init);
2539 set = isl_set_lower_bound_si(set, isl_dim_set, 0, 0);
2541 return isl_set_params(set);
2544 /* Assuming "cond" represents a bound on a loop where the loop
2545 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2546 * is possible.
2548 * Under the given assumptions, wrapping is only possible if "cond" allows
2549 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2550 * increasing iterator and 0 in case of a decreasing iterator.
2552 static bool can_wrap(__isl_keep isl_set *cond, ValueDecl *iv,
2553 __isl_keep isl_val *inc)
2555 bool cw;
2556 isl_ctx *ctx;
2557 isl_val *limit;
2558 isl_set *test;
2560 test = isl_set_copy(cond);
2562 ctx = isl_set_get_ctx(test);
2563 if (isl_val_is_neg(inc))
2564 limit = isl_val_zero(ctx);
2565 else {
2566 limit = isl_val_int_from_ui(ctx, get_type_size(iv));
2567 limit = isl_val_2exp(limit);
2568 limit = isl_val_sub_ui(limit, 1);
2571 test = isl_set_fix_val(cond, isl_dim_set, 0, limit);
2572 cw = !isl_set_is_empty(test);
2573 isl_set_free(test);
2575 return cw;
2578 /* Given a one-dimensional space, construct the following affine expression
2579 * on this space
2581 * { [v] -> [v mod 2^width] }
2583 * where width is the number of bits used to represent the values
2584 * of the unsigned variable "iv".
2586 static __isl_give isl_aff *compute_wrapping(__isl_take isl_space *dim,
2587 ValueDecl *iv)
2589 isl_ctx *ctx;
2590 isl_val *mod;
2591 isl_aff *aff;
2593 ctx = isl_space_get_ctx(dim);
2594 mod = isl_val_int_from_ui(ctx, get_type_size(iv));
2595 mod = isl_val_2exp(mod);
2597 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2598 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2599 aff = isl_aff_mod_val(aff, mod);
2601 return aff;
2604 /* Project out the parameter "id" from "set".
2606 static __isl_give isl_set *set_project_out_by_id(__isl_take isl_set *set,
2607 __isl_keep isl_id *id)
2609 int pos;
2611 pos = isl_set_find_dim_by_id(set, isl_dim_param, id);
2612 if (pos >= 0)
2613 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2615 return set;
2618 /* Compute the set of parameters for which "set1" is a subset of "set2".
2620 * set1 is a subset of set2 if
2622 * forall i in set1 : i in set2
2624 * or
2626 * not exists i in set1 and i not in set2
2628 * i.e.,
2630 * not exists i in set1 \ set2
2632 static __isl_give isl_set *enforce_subset(__isl_take isl_set *set1,
2633 __isl_take isl_set *set2)
2635 return isl_set_complement(isl_set_params(isl_set_subtract(set1, set2)));
2638 /* Compute the set of parameter values for which "cond" holds
2639 * on the next iteration for each element of "dom".
2641 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2642 * and then compute the set of parameters for which the result is a subset
2643 * of "cond".
2645 static __isl_give isl_set *valid_on_next(__isl_take isl_set *cond,
2646 __isl_take isl_set *dom, __isl_take isl_val *inc)
2648 isl_space *space;
2649 isl_aff *aff;
2650 isl_map *next;
2652 space = isl_set_get_space(dom);
2653 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2654 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2655 aff = isl_aff_add_constant_val(aff, inc);
2656 next = isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2658 dom = isl_set_apply(dom, next);
2660 return enforce_subset(dom, cond);
2663 /* Does "id" refer to a nested access?
2665 static bool is_nested_parameter(__isl_keep isl_id *id)
2667 return id && isl_id_get_user(id) && !isl_id_get_name(id);
2670 /* Does parameter "pos" of "space" refer to a nested access?
2672 static bool is_nested_parameter(__isl_keep isl_space *space, int pos)
2674 bool nested;
2675 isl_id *id;
2677 id = isl_space_get_dim_id(space, isl_dim_param, pos);
2678 nested = is_nested_parameter(id);
2679 isl_id_free(id);
2681 return nested;
2684 /* Does "space" involve any parameters that refer to nested
2685 * accesses, i.e., parameters with no name?
2687 static bool has_nested(__isl_keep isl_space *space)
2689 int nparam;
2691 nparam = isl_space_dim(space, isl_dim_param);
2692 for (int i = 0; i < nparam; ++i)
2693 if (is_nested_parameter(space, i))
2694 return true;
2696 return false;
2699 /* Does "pa" involve any parameters that refer to nested
2700 * accesses, i.e., parameters with no name?
2702 static bool has_nested(__isl_keep isl_pw_aff *pa)
2704 isl_space *space;
2705 bool nested;
2707 space = isl_pw_aff_get_space(pa);
2708 nested = has_nested(space);
2709 isl_space_free(space);
2711 return nested;
2714 /* Construct a pet_scop for a for statement.
2715 * The for loop is required to be of the form
2717 * for (i = init; condition; ++i)
2719 * or
2721 * for (i = init; condition; --i)
2723 * The initialization of the for loop should either be an assignment
2724 * to an integer variable, or a declaration of such a variable with
2725 * initialization.
2727 * The condition is allowed to contain nested accesses, provided
2728 * they are not being written to inside the body of the loop.
2729 * Otherwise, or if the condition is otherwise non-affine, the for loop is
2730 * essentially treated as a while loop, with iteration domain
2731 * { [i] : i >= init }.
2733 * We extract a pet_scop for the body and then embed it in a loop with
2734 * iteration domain and schedule
2736 * { [i] : i >= init and condition' }
2737 * { [i] -> [i] }
2739 * or
2741 * { [i] : i <= init and condition' }
2742 * { [i] -> [-i] }
2744 * Where condition' is equal to condition if the latter is
2745 * a simple upper [lower] bound and a condition that is extended
2746 * to apply to all previous iterations otherwise.
2748 * If the condition is non-affine, then we drop the condition from the
2749 * iteration domain and instead create a separate statement
2750 * for evaluating the condition. The body is then filtered to depend
2751 * on the result of the condition evaluating to true on all iterations
2752 * up to the current iteration, while the evaluation the condition itself
2753 * is filtered to depend on the result of the condition evaluating to true
2754 * on all previous iterations.
2755 * The context of the scop representing the body is dropped
2756 * because we don't know how many times the body will be executed,
2757 * if at all.
2759 * If the stride of the loop is not 1, then "i >= init" is replaced by
2761 * (exists a: i = init + stride * a and a >= 0)
2763 * If the loop iterator i is unsigned, then wrapping may occur.
2764 * During the computation, we work with a virtual iterator that
2765 * does not wrap. However, the condition in the code applies
2766 * to the wrapped value, so we need to change condition(i)
2767 * into condition([i % 2^width]).
2768 * After computing the virtual domain and schedule, we apply
2769 * the function { [v] -> [v % 2^width] } to the domain and the domain
2770 * of the schedule. In order not to lose any information, we also
2771 * need to intersect the domain of the schedule with the virtual domain
2772 * first, since some iterations in the wrapped domain may be scheduled
2773 * several times, typically an infinite number of times.
2774 * Note that there may be no need to perform this final wrapping
2775 * if the loop condition (after wrapping) satisfies certain conditions.
2776 * However, the is_simple_bound condition is not enough since it doesn't
2777 * check if there even is an upper bound.
2779 * If the loop condition is non-affine, then we keep the virtual
2780 * iterator in the iteration domain and instead replace all accesses
2781 * to the original iterator by the wrapping of the virtual iterator.
2783 * Wrapping on unsigned iterators can be avoided entirely if
2784 * loop condition is simple, the loop iterator is incremented
2785 * [decremented] by one and the last value before wrapping cannot
2786 * possibly satisfy the loop condition.
2788 * Before extracting a pet_scop from the body we remove all
2789 * assignments in assigned_value to variables that are assigned
2790 * somewhere in the body of the loop.
2792 * Valid parameters for a for loop are those for which the initial
2793 * value itself, the increment on each domain iteration and
2794 * the condition on both the initial value and
2795 * the result of incrementing the iterator for each iteration of the domain
2796 * can be evaluated.
2797 * If the loop condition is non-affine, then we only consider validity
2798 * of the initial value.
2800 * If the body contains any break, then we keep track of it in "skip"
2801 * (if the skip condition is affine) or it is handled in scop_add_break
2802 * (if the skip condition is not affine).
2803 * Note that the affine break condition needs to be considered with
2804 * respect to previous iterations in the virtual domain (if any)
2805 * and that the domain needs to be kept virtual if there is a non-affine
2806 * break condition.
2808 struct pet_scop *PetScan::extract_for(ForStmt *stmt)
2810 BinaryOperator *ass;
2811 Decl *decl;
2812 Stmt *init;
2813 Expr *lhs, *rhs;
2814 ValueDecl *iv;
2815 isl_space *space;
2816 isl_set *domain;
2817 isl_map *sched;
2818 isl_set *cond = NULL;
2819 isl_set *skip = NULL;
2820 isl_id *id, *id_test = NULL, *id_break_test;
2821 struct pet_scop *scop, *scop_cond = NULL;
2822 assigned_value_cache cache(assigned_value);
2823 isl_val *inc;
2824 bool is_one;
2825 bool is_unsigned;
2826 bool is_simple;
2827 bool is_virtual;
2828 bool keep_virtual = false;
2829 bool has_affine_break;
2830 bool has_var_break;
2831 isl_aff *wrap = NULL;
2832 isl_pw_aff *pa, *pa_inc, *init_val;
2833 isl_set *valid_init;
2834 isl_set *valid_cond;
2835 isl_set *valid_cond_init;
2836 isl_set *valid_cond_next;
2837 isl_set *valid_inc;
2838 int stmt_id;
2840 if (!stmt->getInit() && !stmt->getCond() && !stmt->getInc())
2841 return extract_infinite_for(stmt);
2843 init = stmt->getInit();
2844 if (!init) {
2845 unsupported(stmt);
2846 return NULL;
2848 if ((ass = initialization_assignment(init)) != NULL) {
2849 iv = extract_induction_variable(ass);
2850 if (!iv)
2851 return NULL;
2852 lhs = ass->getLHS();
2853 rhs = ass->getRHS();
2854 } else if ((decl = initialization_declaration(init)) != NULL) {
2855 VarDecl *var = extract_induction_variable(init, decl);
2856 if (!var)
2857 return NULL;
2858 iv = var;
2859 rhs = var->getInit();
2860 lhs = create_DeclRefExpr(var);
2861 } else {
2862 unsupported(stmt->getInit());
2863 return NULL;
2866 pa_inc = extract_increment(stmt, iv);
2867 if (!pa_inc)
2868 return NULL;
2870 inc = NULL;
2871 if (isl_pw_aff_n_piece(pa_inc) != 1 ||
2872 isl_pw_aff_foreach_piece(pa_inc, &extract_cst, &inc) < 0) {
2873 isl_pw_aff_free(pa_inc);
2874 unsupported(stmt->getInc());
2875 isl_val_free(inc);
2876 return NULL;
2878 valid_inc = isl_pw_aff_domain(pa_inc);
2880 is_unsigned = iv->getType()->isUnsignedIntegerType();
2882 assigned_value.erase(iv);
2883 clear_assignments clear(assigned_value);
2884 clear.TraverseStmt(stmt->getBody());
2886 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
2888 pa = try_extract_nested_condition(stmt->getCond());
2889 if (allow_nested && (!pa || has_nested(pa)))
2890 stmt_id = n_stmt++;
2892 scop = extract(stmt->getBody());
2894 has_affine_break = scop &&
2895 pet_scop_has_affine_skip(scop, pet_skip_later);
2896 if (has_affine_break)
2897 skip = pet_scop_get_affine_skip_domain(scop, pet_skip_later);
2898 has_var_break = scop && pet_scop_has_var_skip(scop, pet_skip_later);
2899 if (has_var_break) {
2900 id_break_test = pet_scop_get_skip_id(scop, pet_skip_later);
2901 keep_virtual = true;
2904 if (pa && !is_nested_allowed(pa, scop)) {
2905 isl_pw_aff_free(pa);
2906 pa = NULL;
2909 if (!allow_nested && !pa)
2910 pa = try_extract_affine_condition(stmt->getCond());
2911 valid_cond = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2912 cond = isl_pw_aff_non_zero_set(pa);
2913 if (allow_nested && !cond) {
2914 isl_multi_pw_aff *test_index;
2915 int save_n_stmt = n_stmt;
2916 test_index = create_test_index(ctx, n_test++);
2917 n_stmt = stmt_id;
2918 scop_cond = extract_non_affine_condition(stmt->getCond(),
2919 isl_multi_pw_aff_copy(test_index));
2920 n_stmt = save_n_stmt;
2921 scop_cond = scop_add_array(scop_cond, test_index, ast_context);
2922 id_test = isl_multi_pw_aff_get_tuple_id(test_index,
2923 isl_dim_out);
2924 isl_multi_pw_aff_free(test_index);
2925 scop_cond = pet_scop_prefix(scop_cond, 0);
2926 scop = pet_scop_reset_context(scop);
2927 scop = pet_scop_prefix(scop, 1);
2928 keep_virtual = true;
2929 cond = isl_set_universe(isl_space_set_alloc(ctx, 0, 0));
2932 cond = embed(cond, isl_id_copy(id));
2933 skip = embed(skip, isl_id_copy(id));
2934 valid_cond = isl_set_coalesce(valid_cond);
2935 valid_cond = embed(valid_cond, isl_id_copy(id));
2936 valid_inc = embed(valid_inc, isl_id_copy(id));
2937 is_one = isl_val_is_one(inc) || isl_val_is_negone(inc);
2938 is_virtual = is_unsigned && (!is_one || can_wrap(cond, iv, inc));
2940 init_val = extract_affine(rhs);
2941 valid_cond_init = enforce_subset(
2942 isl_set_from_pw_aff(isl_pw_aff_copy(init_val)),
2943 isl_set_copy(valid_cond));
2944 if (is_one && !is_virtual) {
2945 isl_pw_aff_free(init_val);
2946 pa = extract_comparison(isl_val_is_pos(inc) ? BO_GE : BO_LE,
2947 lhs, rhs, init);
2948 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2949 valid_init = set_project_out_by_id(valid_init, id);
2950 domain = isl_pw_aff_non_zero_set(pa);
2951 } else {
2952 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(init_val));
2953 domain = strided_domain(isl_id_copy(id), init_val,
2954 isl_val_copy(inc));
2957 domain = embed(domain, isl_id_copy(id));
2958 if (is_virtual) {
2959 isl_map *rev_wrap;
2960 wrap = compute_wrapping(isl_set_get_space(cond), iv);
2961 rev_wrap = isl_map_from_aff(isl_aff_copy(wrap));
2962 rev_wrap = isl_map_reverse(rev_wrap);
2963 cond = isl_set_apply(cond, isl_map_copy(rev_wrap));
2964 skip = isl_set_apply(skip, isl_map_copy(rev_wrap));
2965 valid_cond = isl_set_apply(valid_cond, isl_map_copy(rev_wrap));
2966 valid_inc = isl_set_apply(valid_inc, rev_wrap);
2968 is_simple = is_simple_bound(cond, inc);
2969 if (!is_simple) {
2970 cond = isl_set_gist(cond, isl_set_copy(domain));
2971 is_simple = is_simple_bound(cond, inc);
2973 if (!is_simple)
2974 cond = valid_for_each_iteration(cond,
2975 isl_set_copy(domain), isl_val_copy(inc));
2976 domain = isl_set_intersect(domain, cond);
2977 if (has_affine_break) {
2978 skip = isl_set_intersect(skip , isl_set_copy(domain));
2979 skip = after(skip, isl_val_sgn(inc));
2980 domain = isl_set_subtract(domain, skip);
2982 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, isl_id_copy(id));
2983 space = isl_space_from_domain(isl_set_get_space(domain));
2984 space = isl_space_add_dims(space, isl_dim_out, 1);
2985 sched = isl_map_universe(space);
2986 if (isl_val_is_pos(inc))
2987 sched = isl_map_equate(sched, isl_dim_in, 0, isl_dim_out, 0);
2988 else
2989 sched = isl_map_oppose(sched, isl_dim_in, 0, isl_dim_out, 0);
2991 valid_cond_next = valid_on_next(valid_cond, isl_set_copy(domain),
2992 isl_val_copy(inc));
2993 valid_inc = enforce_subset(isl_set_copy(domain), valid_inc);
2995 if (is_virtual && !keep_virtual) {
2996 isl_map *wrap_map = isl_map_from_aff(wrap);
2997 wrap_map = isl_map_set_dim_id(wrap_map,
2998 isl_dim_out, 0, isl_id_copy(id));
2999 sched = isl_map_intersect_domain(sched, isl_set_copy(domain));
3000 domain = isl_set_apply(domain, isl_map_copy(wrap_map));
3001 sched = isl_map_apply_domain(sched, wrap_map);
3003 if (!(is_virtual && keep_virtual))
3004 wrap = identity_aff(domain);
3006 scop_cond = pet_scop_embed(scop_cond, isl_set_copy(domain),
3007 isl_map_copy(sched), isl_aff_copy(wrap), isl_id_copy(id));
3008 scop = pet_scop_embed(scop, isl_set_copy(domain), sched, wrap, id);
3009 scop = resolve_nested(scop);
3010 if (has_var_break)
3011 scop = scop_add_break(scop, id_break_test, isl_set_copy(domain),
3012 isl_val_copy(inc));
3013 if (id_test) {
3014 scop = scop_add_while(scop_cond, scop, id_test, domain,
3015 isl_val_copy(inc));
3016 isl_set_free(valid_inc);
3017 } else {
3018 scop = pet_scop_restrict_context(scop, valid_inc);
3019 scop = pet_scop_restrict_context(scop, valid_cond_next);
3020 scop = pet_scop_restrict_context(scop, valid_cond_init);
3021 isl_set_free(domain);
3023 clear_assignment(assigned_value, iv);
3025 isl_val_free(inc);
3027 scop = pet_scop_restrict_context(scop, valid_init);
3029 return scop;
3032 struct pet_scop *PetScan::extract(CompoundStmt *stmt, bool skip_declarations)
3034 return extract(stmt->children(), true, skip_declarations);
3037 /* Does parameter "pos" of "map" refer to a nested access?
3039 static bool is_nested_parameter(__isl_keep isl_map *map, int pos)
3041 bool nested;
3042 isl_id *id;
3044 id = isl_map_get_dim_id(map, isl_dim_param, pos);
3045 nested = is_nested_parameter(id);
3046 isl_id_free(id);
3048 return nested;
3051 /* How many parameters of "space" refer to nested accesses, i.e., have no name?
3053 static int n_nested_parameter(__isl_keep isl_space *space)
3055 int n = 0;
3056 int nparam;
3058 nparam = isl_space_dim(space, isl_dim_param);
3059 for (int i = 0; i < nparam; ++i)
3060 if (is_nested_parameter(space, i))
3061 ++n;
3063 return n;
3066 /* How many parameters of "map" refer to nested accesses, i.e., have no name?
3068 static int n_nested_parameter(__isl_keep isl_map *map)
3070 isl_space *space;
3071 int n;
3073 space = isl_map_get_space(map);
3074 n = n_nested_parameter(space);
3075 isl_space_free(space);
3077 return n;
3080 /* For each nested access parameter in "space",
3081 * construct a corresponding pet_expr, place it in args and
3082 * record its position in "param2pos".
3083 * "n_arg" is the number of elements that are already in args.
3084 * The position recorded in "param2pos" takes this number into account.
3085 * If the pet_expr corresponding to a parameter is identical to
3086 * the pet_expr corresponding to an earlier parameter, then these two
3087 * parameters are made to refer to the same element in args.
3089 * Return the final number of elements in args or -1 if an error has occurred.
3091 int PetScan::extract_nested(__isl_keep isl_space *space,
3092 int n_arg, struct pet_expr **args, std::map<int,int> &param2pos)
3094 int nparam;
3096 nparam = isl_space_dim(space, isl_dim_param);
3097 for (int i = 0; i < nparam; ++i) {
3098 int j;
3099 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
3100 Expr *nested;
3102 if (!is_nested_parameter(id)) {
3103 isl_id_free(id);
3104 continue;
3107 nested = (Expr *) isl_id_get_user(id);
3108 args[n_arg] = extract_expr(nested);
3109 if (!args[n_arg])
3110 return -1;
3112 for (j = 0; j < n_arg; ++j)
3113 if (pet_expr_is_equal(args[j], args[n_arg]))
3114 break;
3116 if (j < n_arg) {
3117 pet_expr_free(args[n_arg]);
3118 args[n_arg] = NULL;
3119 param2pos[i] = j;
3120 } else
3121 param2pos[i] = n_arg++;
3123 isl_id_free(id);
3126 return n_arg;
3129 /* For each nested access parameter in the access relations in "expr",
3130 * construct a corresponding pet_expr, place it in expr->args and
3131 * record its position in "param2pos".
3132 * n is the number of nested access parameters.
3134 struct pet_expr *PetScan::extract_nested(struct pet_expr *expr, int n,
3135 std::map<int,int> &param2pos)
3137 isl_space *space;
3139 expr->args = isl_calloc_array(ctx, struct pet_expr *, n);
3140 expr->n_arg = n;
3141 if (!expr->args)
3142 goto error;
3144 space = isl_map_get_space(expr->acc.access);
3145 n = extract_nested(space, 0, expr->args, param2pos);
3146 isl_space_free(space);
3148 if (n < 0)
3149 goto error;
3151 expr->n_arg = n;
3152 return expr;
3153 error:
3154 pet_expr_free(expr);
3155 return NULL;
3158 /* Look for parameters in any access relation in "expr" that
3159 * refer to nested accesses. In particular, these are
3160 * parameters with no name.
3162 * If there are any such parameters, then the domain of the index
3163 * expression and the access relation, which is still [] at this point,
3164 * is replaced by [[] -> [t_1,...,t_n]], with n the number of these parameters
3165 * (after identifying identical nested accesses).
3167 * This transformation is performed in several steps.
3168 * We first extract the arguments in extract_nested.
3169 * param2pos maps the original parameter position to the position
3170 * of the argument.
3171 * Then we move these parameters to input dimension.
3172 * t2pos maps the positions of these temporary input dimensions
3173 * to the positions of the corresponding arguments.
3174 * Finally, we express there temporary dimensions in term of the domain
3175 * [[] -> [t_1,...,t_n]] and precompose index expression and access
3176 * relations with this function.
3178 struct pet_expr *PetScan::resolve_nested(struct pet_expr *expr)
3180 int n;
3181 int nparam;
3182 isl_space *space;
3183 isl_local_space *ls;
3184 isl_aff *aff;
3185 isl_multi_aff *ma;
3186 std::map<int,int> param2pos;
3187 std::map<int,int> t2pos;
3189 if (!expr)
3190 return expr;
3192 for (int i = 0; i < expr->n_arg; ++i) {
3193 expr->args[i] = resolve_nested(expr->args[i]);
3194 if (!expr->args[i]) {
3195 pet_expr_free(expr);
3196 return NULL;
3200 if (expr->type != pet_expr_access)
3201 return expr;
3203 n = n_nested_parameter(expr->acc.access);
3204 if (n == 0)
3205 return expr;
3207 expr = extract_nested(expr, n, param2pos);
3208 if (!expr)
3209 return NULL;
3211 expr = pet_expr_access_align_params(expr);
3212 if (!expr)
3213 return NULL;
3214 nparam = isl_map_dim(expr->acc.access, isl_dim_param);
3216 n = 0;
3217 for (int i = nparam - 1; i >= 0; --i) {
3218 isl_id *id = isl_map_get_dim_id(expr->acc.access,
3219 isl_dim_param, i);
3220 if (!is_nested_parameter(id)) {
3221 isl_id_free(id);
3222 continue;
3225 expr->acc.access = isl_map_move_dims(expr->acc.access,
3226 isl_dim_in, n, isl_dim_param, i, 1);
3227 expr->acc.index = isl_multi_pw_aff_move_dims(expr->acc.index,
3228 isl_dim_in, n, isl_dim_param, i, 1);
3229 t2pos[n] = param2pos[i];
3230 n++;
3232 isl_id_free(id);
3235 space = isl_multi_pw_aff_get_space(expr->acc.index);
3236 space = isl_space_set_from_params(isl_space_params(space));
3237 space = isl_space_add_dims(space, isl_dim_set, expr->n_arg);
3238 space = isl_space_wrap(isl_space_from_range(space));
3239 ls = isl_local_space_from_space(isl_space_copy(space));
3240 space = isl_space_from_domain(space);
3241 space = isl_space_add_dims(space, isl_dim_out, n);
3242 ma = isl_multi_aff_zero(space);
3244 for (int i = 0; i < n; ++i) {
3245 aff = isl_aff_var_on_domain(isl_local_space_copy(ls),
3246 isl_dim_set, t2pos[i]);
3247 ma = isl_multi_aff_set_aff(ma, i, aff);
3249 isl_local_space_free(ls);
3251 expr->acc.access = isl_map_preimage_domain_multi_aff(expr->acc.access,
3252 isl_multi_aff_copy(ma));
3253 expr->acc.index = isl_multi_pw_aff_pullback_multi_aff(expr->acc.index,
3254 ma);
3256 return expr;
3259 /* Return the file offset of the expansion location of "Loc".
3261 static unsigned getExpansionOffset(SourceManager &SM, SourceLocation Loc)
3263 return SM.getFileOffset(SM.getExpansionLoc(Loc));
3266 #ifdef HAVE_FINDLOCATIONAFTERTOKEN
3268 /* Return a SourceLocation for the location after the first semicolon
3269 * after "loc". If Lexer::findLocationAfterToken is available, we simply
3270 * call it and also skip trailing spaces and newline.
3272 static SourceLocation location_after_semi(SourceLocation loc, SourceManager &SM,
3273 const LangOptions &LO)
3275 return Lexer::findLocationAfterToken(loc, tok::semi, SM, LO, true);
3278 #else
3280 /* Return a SourceLocation for the location after the first semicolon
3281 * after "loc". If Lexer::findLocationAfterToken is not available,
3282 * we look in the underlying character data for the first semicolon.
3284 static SourceLocation location_after_semi(SourceLocation loc, SourceManager &SM,
3285 const LangOptions &LO)
3287 const char *semi;
3288 const char *s = SM.getCharacterData(loc);
3290 semi = strchr(s, ';');
3291 if (!semi)
3292 return SourceLocation();
3293 return loc.getFileLocWithOffset(semi + 1 - s);
3296 #endif
3298 /* If the token at "loc" is the first token on the line, then return
3299 * a location referring to the start of the line.
3300 * Otherwise, return "loc".
3302 * This function is used to extend a scop to the start of the line
3303 * if the first token of the scop is also the first token on the line.
3305 * We look for the first token on the line. If its location is equal to "loc",
3306 * then the latter is the location of the first token on the line.
3308 static SourceLocation move_to_start_of_line_if_first_token(SourceLocation loc,
3309 SourceManager &SM, const LangOptions &LO)
3311 std::pair<FileID, unsigned> file_offset_pair;
3312 llvm::StringRef file;
3313 const char *pos;
3314 Token tok;
3315 SourceLocation token_loc, line_loc;
3316 int col;
3318 loc = SM.getExpansionLoc(loc);
3319 col = SM.getExpansionColumnNumber(loc);
3320 line_loc = loc.getLocWithOffset(1 - col);
3321 file_offset_pair = SM.getDecomposedLoc(line_loc);
3322 file = SM.getBufferData(file_offset_pair.first, NULL);
3323 pos = file.data() + file_offset_pair.second;
3325 Lexer lexer(SM.getLocForStartOfFile(file_offset_pair.first), LO,
3326 file.begin(), pos, file.end());
3327 lexer.LexFromRawLexer(tok);
3328 token_loc = tok.getLocation();
3330 if (token_loc == loc)
3331 return line_loc;
3332 else
3333 return loc;
3336 /* Convert a top-level pet_expr to a pet_scop with one statement.
3337 * This mainly involves resolving nested expression parameters
3338 * and setting the name of the iteration space.
3339 * The name is given by "label" if it is non-NULL. Otherwise,
3340 * it is of the form S_<n_stmt>.
3341 * start and end of the pet_scop are derived from those of "stmt".
3343 struct pet_scop *PetScan::extract(Stmt *stmt, struct pet_expr *expr,
3344 __isl_take isl_id *label)
3346 struct pet_stmt *ps;
3347 struct pet_scop *scop;
3348 SourceLocation loc = stmt->getLocStart();
3349 SourceManager &SM = PP.getSourceManager();
3350 const LangOptions &LO = PP.getLangOpts();
3351 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3352 unsigned start, end;
3354 expr = resolve_nested(expr);
3355 ps = pet_stmt_from_pet_expr(ctx, line, label, n_stmt++, expr);
3356 scop = pet_scop_from_pet_stmt(ctx, ps);
3358 loc = move_to_start_of_line_if_first_token(loc, SM, LO);
3359 start = getExpansionOffset(SM, loc);
3360 loc = stmt->getLocEnd();
3361 loc = location_after_semi(loc, SM, LO);
3362 end = getExpansionOffset(SM, loc);
3364 scop = pet_scop_update_start_end(scop, start, end);
3365 return scop;
3368 /* Check if we can extract an affine expression from "expr".
3369 * Return the expressions as an isl_pw_aff if we can and NULL otherwise.
3370 * We turn on autodetection so that we won't generate any warnings
3371 * and turn off nesting, so that we won't accept any non-affine constructs.
3373 __isl_give isl_pw_aff *PetScan::try_extract_affine(Expr *expr)
3375 isl_pw_aff *pwaff;
3376 int save_autodetect = options->autodetect;
3377 bool save_nesting = nesting_enabled;
3379 options->autodetect = 1;
3380 nesting_enabled = false;
3382 pwaff = extract_affine(expr);
3384 options->autodetect = save_autodetect;
3385 nesting_enabled = save_nesting;
3387 return pwaff;
3390 /* Check whether "expr" is an affine expression.
3392 bool PetScan::is_affine(Expr *expr)
3394 isl_pw_aff *pwaff;
3396 pwaff = try_extract_affine(expr);
3397 isl_pw_aff_free(pwaff);
3399 return pwaff != NULL;
3402 /* Check if we can extract an affine constraint from "expr".
3403 * Return the constraint as an isl_set if we can and NULL otherwise.
3404 * We turn on autodetection so that we won't generate any warnings
3405 * and turn off nesting, so that we won't accept any non-affine constructs.
3407 __isl_give isl_pw_aff *PetScan::try_extract_affine_condition(Expr *expr)
3409 isl_pw_aff *cond;
3410 int save_autodetect = options->autodetect;
3411 bool save_nesting = nesting_enabled;
3413 options->autodetect = 1;
3414 nesting_enabled = false;
3416 cond = extract_condition(expr);
3418 options->autodetect = save_autodetect;
3419 nesting_enabled = save_nesting;
3421 return cond;
3424 /* Check whether "expr" is an affine constraint.
3426 bool PetScan::is_affine_condition(Expr *expr)
3428 isl_pw_aff *cond;
3430 cond = try_extract_affine_condition(expr);
3431 isl_pw_aff_free(cond);
3433 return cond != NULL;
3436 /* Check if we can extract a condition from "expr".
3437 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
3438 * If allow_nested is set, then the condition may involve parameters
3439 * corresponding to nested accesses.
3440 * We turn on autodetection so that we won't generate any warnings.
3442 __isl_give isl_pw_aff *PetScan::try_extract_nested_condition(Expr *expr)
3444 isl_pw_aff *cond;
3445 int save_autodetect = options->autodetect;
3446 bool save_nesting = nesting_enabled;
3448 options->autodetect = 1;
3449 nesting_enabled = allow_nested;
3450 cond = extract_condition(expr);
3452 options->autodetect = save_autodetect;
3453 nesting_enabled = save_nesting;
3455 return cond;
3458 /* If the top-level expression of "stmt" is an assignment, then
3459 * return that assignment as a BinaryOperator.
3460 * Otherwise return NULL.
3462 static BinaryOperator *top_assignment_or_null(Stmt *stmt)
3464 BinaryOperator *ass;
3466 if (!stmt)
3467 return NULL;
3468 if (stmt->getStmtClass() != Stmt::BinaryOperatorClass)
3469 return NULL;
3471 ass = cast<BinaryOperator>(stmt);
3472 if(ass->getOpcode() != BO_Assign)
3473 return NULL;
3475 return ass;
3478 /* Check if the given if statement is a conditional assignement
3479 * with a non-affine condition. If so, construct a pet_scop
3480 * corresponding to this conditional assignment. Otherwise return NULL.
3482 * In particular we check if "stmt" is of the form
3484 * if (condition)
3485 * a = f(...);
3486 * else
3487 * a = g(...);
3489 * where a is some array or scalar access.
3490 * The constructed pet_scop then corresponds to the expression
3492 * a = condition ? f(...) : g(...)
3494 * All access relations in f(...) are intersected with condition
3495 * while all access relation in g(...) are intersected with the complement.
3497 struct pet_scop *PetScan::extract_conditional_assignment(IfStmt *stmt)
3499 BinaryOperator *ass_then, *ass_else;
3500 isl_multi_pw_aff *write_then, *write_else;
3501 isl_set *cond, *comp;
3502 isl_multi_pw_aff *index;
3503 isl_pw_aff *pa;
3504 int equal;
3505 struct pet_expr *pe_cond, *pe_then, *pe_else, *pe, *pe_write;
3506 bool save_nesting = nesting_enabled;
3508 if (!options->detect_conditional_assignment)
3509 return NULL;
3511 ass_then = top_assignment_or_null(stmt->getThen());
3512 ass_else = top_assignment_or_null(stmt->getElse());
3514 if (!ass_then || !ass_else)
3515 return NULL;
3517 if (is_affine_condition(stmt->getCond()))
3518 return NULL;
3520 write_then = extract_index(ass_then->getLHS());
3521 write_else = extract_index(ass_else->getLHS());
3523 equal = isl_multi_pw_aff_plain_is_equal(write_then, write_else);
3524 isl_multi_pw_aff_free(write_else);
3525 if (equal < 0 || !equal) {
3526 isl_multi_pw_aff_free(write_then);
3527 return NULL;
3530 nesting_enabled = allow_nested;
3531 pa = extract_condition(stmt->getCond());
3532 nesting_enabled = save_nesting;
3533 cond = isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa));
3534 comp = isl_pw_aff_zero_set(isl_pw_aff_copy(pa));
3535 index = isl_multi_pw_aff_from_range(isl_multi_pw_aff_from_pw_aff(pa));
3537 pe_cond = pet_expr_from_index(index);
3539 pe_then = extract_expr(ass_then->getRHS());
3540 pe_then = pet_expr_restrict(pe_then, cond);
3541 pe_else = extract_expr(ass_else->getRHS());
3542 pe_else = pet_expr_restrict(pe_else, comp);
3544 pe = pet_expr_new_ternary(ctx, pe_cond, pe_then, pe_else);
3545 pe_write = pet_expr_from_index_and_depth(write_then,
3546 extract_depth(write_then));
3547 if (pe_write) {
3548 pe_write->acc.write = 1;
3549 pe_write->acc.read = 0;
3551 pe = pet_expr_new_binary(ctx, pet_op_assign, pe_write, pe);
3552 return extract(stmt, pe);
3555 /* Create a pet_scop with a single statement evaluating "cond"
3556 * and writing the result to a virtual scalar, as expressed by
3557 * "index".
3559 struct pet_scop *PetScan::extract_non_affine_condition(Expr *cond,
3560 __isl_take isl_multi_pw_aff *index)
3562 struct pet_expr *expr, *write;
3563 struct pet_stmt *ps;
3564 struct pet_scop *scop;
3565 SourceLocation loc = cond->getLocStart();
3566 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3568 write = pet_expr_from_index(index);
3569 if (write) {
3570 write->acc.write = 1;
3571 write->acc.read = 0;
3573 expr = extract_expr(cond);
3574 expr = resolve_nested(expr);
3575 expr = pet_expr_new_binary(ctx, pet_op_assign, write, expr);
3576 ps = pet_stmt_from_pet_expr(ctx, line, NULL, n_stmt++, expr);
3577 scop = pet_scop_from_pet_stmt(ctx, ps);
3578 scop = resolve_nested(scop);
3580 return scop;
3583 extern "C" {
3584 static struct pet_expr *embed_access(struct pet_expr *expr, void *user);
3587 /* Precompose the access relation and the index expression associated
3588 * to "expr" with the function pointed to by "user",
3589 * thereby embedding the access relation in the domain of this function.
3590 * The initial domain of the access relation and the index expression
3591 * is the zero-dimensional domain.
3593 static struct pet_expr *embed_access(struct pet_expr *expr, void *user)
3595 isl_multi_aff *ma = (isl_multi_aff *) user;
3597 expr->acc.access = isl_map_preimage_domain_multi_aff(expr->acc.access,
3598 isl_multi_aff_copy(ma));
3599 expr->acc.index = isl_multi_pw_aff_pullback_multi_aff(expr->acc.index,
3600 isl_multi_aff_copy(ma));
3601 if (!expr->acc.access || !expr->acc.index)
3602 goto error;
3604 return expr;
3605 error:
3606 pet_expr_free(expr);
3607 return NULL;
3610 /* Precompose all access relations in "expr" with "ma", thereby
3611 * embedding them in the domain of "ma".
3613 static struct pet_expr *embed(struct pet_expr *expr,
3614 __isl_keep isl_multi_aff *ma)
3616 return pet_expr_map_access(expr, &embed_access, ma);
3619 /* How many parameters of "set" refer to nested accesses, i.e., have no name?
3621 static int n_nested_parameter(__isl_keep isl_set *set)
3623 isl_space *space;
3624 int n;
3626 space = isl_set_get_space(set);
3627 n = n_nested_parameter(space);
3628 isl_space_free(space);
3630 return n;
3633 /* Remove all parameters from "map" that refer to nested accesses.
3635 static __isl_give isl_map *remove_nested_parameters(__isl_take isl_map *map)
3637 int nparam;
3638 isl_space *space;
3640 space = isl_map_get_space(map);
3641 nparam = isl_space_dim(space, isl_dim_param);
3642 for (int i = nparam - 1; i >= 0; --i)
3643 if (is_nested_parameter(space, i))
3644 map = isl_map_project_out(map, isl_dim_param, i, 1);
3645 isl_space_free(space);
3647 return map;
3650 /* Remove all parameters from "mpa" that refer to nested accesses.
3652 static __isl_give isl_multi_pw_aff *remove_nested_parameters(
3653 __isl_take isl_multi_pw_aff *mpa)
3655 int nparam;
3656 isl_space *space;
3658 space = isl_multi_pw_aff_get_space(mpa);
3659 nparam = isl_space_dim(space, isl_dim_param);
3660 for (int i = nparam - 1; i >= 0; --i) {
3661 if (!is_nested_parameter(space, i))
3662 continue;
3663 mpa = isl_multi_pw_aff_drop_dims(mpa, isl_dim_param, i, 1);
3665 isl_space_free(space);
3667 return mpa;
3670 /* Remove all parameters from the index expression and access relation of "expr"
3671 * that refer to nested accesses.
3673 static struct pet_expr *remove_nested_parameters(struct pet_expr *expr)
3675 expr->acc.access = remove_nested_parameters(expr->acc.access);
3676 expr->acc.index = remove_nested_parameters(expr->acc.index);
3677 if (!expr->acc.access || !expr->acc.index)
3678 goto error;
3680 return expr;
3681 error:
3682 pet_expr_free(expr);
3683 return NULL;
3686 extern "C" {
3687 static struct pet_expr *expr_remove_nested_parameters(
3688 struct pet_expr *expr, void *user);
3691 static struct pet_expr *expr_remove_nested_parameters(
3692 struct pet_expr *expr, void *user)
3694 return remove_nested_parameters(expr);
3697 /* Remove all nested access parameters from the schedule and all
3698 * accesses of "stmt".
3699 * There is no need to remove them from the domain as these parameters
3700 * have already been removed from the domain when this function is called.
3702 static struct pet_stmt *remove_nested_parameters(struct pet_stmt *stmt)
3704 if (!stmt)
3705 return NULL;
3706 stmt->schedule = remove_nested_parameters(stmt->schedule);
3707 stmt->body = pet_expr_map_access(stmt->body,
3708 &expr_remove_nested_parameters, NULL);
3709 if (!stmt->schedule || !stmt->body)
3710 goto error;
3711 for (int i = 0; i < stmt->n_arg; ++i) {
3712 stmt->args[i] = pet_expr_map_access(stmt->args[i],
3713 &expr_remove_nested_parameters, NULL);
3714 if (!stmt->args[i])
3715 goto error;
3718 return stmt;
3719 error:
3720 pet_stmt_free(stmt);
3721 return NULL;
3724 /* For each nested access parameter in the domain of "stmt",
3725 * construct a corresponding pet_expr, place it before the original
3726 * elements in stmt->args and record its position in "param2pos".
3727 * n is the number of nested access parameters.
3729 struct pet_stmt *PetScan::extract_nested(struct pet_stmt *stmt, int n,
3730 std::map<int,int> &param2pos)
3732 int i;
3733 isl_space *space;
3734 int n_arg;
3735 struct pet_expr **args;
3737 n_arg = stmt->n_arg;
3738 args = isl_calloc_array(ctx, struct pet_expr *, n + n_arg);
3739 if (!args)
3740 goto error;
3742 space = isl_set_get_space(stmt->domain);
3743 n_arg = extract_nested(space, 0, args, param2pos);
3744 isl_space_free(space);
3746 if (n_arg < 0)
3747 goto error;
3749 for (i = 0; i < stmt->n_arg; ++i)
3750 args[n_arg + i] = stmt->args[i];
3751 free(stmt->args);
3752 stmt->args = args;
3753 stmt->n_arg += n_arg;
3755 return stmt;
3756 error:
3757 if (args) {
3758 for (i = 0; i < n; ++i)
3759 pet_expr_free(args[i]);
3760 free(args);
3762 pet_stmt_free(stmt);
3763 return NULL;
3766 /* Check whether any of the arguments i of "stmt" starting at position "n"
3767 * is equal to one of the first "n" arguments j.
3768 * If so, combine the constraints on arguments i and j and remove
3769 * argument i.
3771 static struct pet_stmt *remove_duplicate_arguments(struct pet_stmt *stmt, int n)
3773 int i, j;
3774 isl_map *map;
3776 if (!stmt)
3777 return NULL;
3778 if (n == 0)
3779 return stmt;
3780 if (n == stmt->n_arg)
3781 return stmt;
3783 map = isl_set_unwrap(stmt->domain);
3785 for (i = stmt->n_arg - 1; i >= n; --i) {
3786 for (j = 0; j < n; ++j)
3787 if (pet_expr_is_equal(stmt->args[i], stmt->args[j]))
3788 break;
3789 if (j >= n)
3790 continue;
3792 map = isl_map_equate(map, isl_dim_out, i, isl_dim_out, j);
3793 map = isl_map_project_out(map, isl_dim_out, i, 1);
3795 pet_expr_free(stmt->args[i]);
3796 for (j = i; j + 1 < stmt->n_arg; ++j)
3797 stmt->args[j] = stmt->args[j + 1];
3798 stmt->n_arg--;
3801 stmt->domain = isl_map_wrap(map);
3802 if (!stmt->domain)
3803 goto error;
3804 return stmt;
3805 error:
3806 pet_stmt_free(stmt);
3807 return NULL;
3810 /* Look for parameters in the iteration domain of "stmt" that
3811 * refer to nested accesses. In particular, these are
3812 * parameters with no name.
3814 * If there are any such parameters, then as many extra variables
3815 * (after identifying identical nested accesses) are inserted in the
3816 * range of the map wrapped inside the domain, before the original variables.
3817 * If the original domain is not a wrapped map, then a new wrapped
3818 * map is created with zero output dimensions.
3819 * The parameters are then equated to the corresponding output dimensions
3820 * and subsequently projected out, from the iteration domain,
3821 * the schedule and the access relations.
3822 * For each of the output dimensions, a corresponding argument
3823 * expression is inserted. Initially they are created with
3824 * a zero-dimensional domain, so they have to be embedded
3825 * in the current iteration domain.
3826 * param2pos maps the position of the parameter to the position
3827 * of the corresponding output dimension in the wrapped map.
3829 struct pet_stmt *PetScan::resolve_nested(struct pet_stmt *stmt)
3831 int n;
3832 int nparam;
3833 unsigned n_arg;
3834 isl_map *map;
3835 isl_space *space;
3836 isl_multi_aff *ma;
3837 std::map<int,int> param2pos;
3839 if (!stmt)
3840 return NULL;
3842 n = n_nested_parameter(stmt->domain);
3843 if (n == 0)
3844 return stmt;
3846 n_arg = stmt->n_arg;
3847 stmt = extract_nested(stmt, n, param2pos);
3848 if (!stmt)
3849 return NULL;
3851 n = stmt->n_arg - n_arg;
3852 nparam = isl_set_dim(stmt->domain, isl_dim_param);
3853 if (isl_set_is_wrapping(stmt->domain))
3854 map = isl_set_unwrap(stmt->domain);
3855 else
3856 map = isl_map_from_domain(stmt->domain);
3857 map = isl_map_insert_dims(map, isl_dim_out, 0, n);
3859 for (int i = nparam - 1; i >= 0; --i) {
3860 isl_id *id;
3862 if (!is_nested_parameter(map, i))
3863 continue;
3865 id = pet_expr_access_get_id(stmt->args[param2pos[i]]);
3866 map = isl_map_set_dim_id(map, isl_dim_out, param2pos[i], id);
3867 map = isl_map_equate(map, isl_dim_param, i, isl_dim_out,
3868 param2pos[i]);
3869 map = isl_map_project_out(map, isl_dim_param, i, 1);
3872 stmt->domain = isl_map_wrap(map);
3874 space = isl_space_unwrap(isl_set_get_space(stmt->domain));
3875 space = isl_space_from_domain(isl_space_domain(space));
3876 ma = isl_multi_aff_zero(space);
3877 for (int pos = 0; pos < n; ++pos)
3878 stmt->args[pos] = embed(stmt->args[pos], ma);
3879 isl_multi_aff_free(ma);
3881 stmt = remove_nested_parameters(stmt);
3882 stmt = remove_duplicate_arguments(stmt, n);
3884 return stmt;
3887 /* For each statement in "scop", move the parameters that correspond
3888 * to nested access into the ranges of the domains and create
3889 * corresponding argument expressions.
3891 struct pet_scop *PetScan::resolve_nested(struct pet_scop *scop)
3893 if (!scop)
3894 return NULL;
3896 for (int i = 0; i < scop->n_stmt; ++i) {
3897 scop->stmts[i] = resolve_nested(scop->stmts[i]);
3898 if (!scop->stmts[i])
3899 goto error;
3902 return scop;
3903 error:
3904 pet_scop_free(scop);
3905 return NULL;
3908 /* Given an access expression "expr", is the variable accessed by
3909 * "expr" assigned anywhere inside "scop"?
3911 static bool is_assigned(pet_expr *expr, pet_scop *scop)
3913 bool assigned = false;
3914 isl_id *id;
3916 id = pet_expr_access_get_id(expr);
3917 assigned = pet_scop_writes(scop, id);
3918 isl_id_free(id);
3920 return assigned;
3923 /* Are all nested access parameters in "pa" allowed given "scop".
3924 * In particular, is none of them written by anywhere inside "scop".
3926 * If "scop" has any skip conditions, then no nested access parameters
3927 * are allowed. In particular, if there is any nested access in a guard
3928 * for a piece of code containing a "continue", then we want to introduce
3929 * a separate statement for evaluating this guard so that we can express
3930 * that the result is false for all previous iterations.
3932 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff *pa, pet_scop *scop)
3934 int nparam;
3936 if (!scop)
3937 return true;
3939 nparam = isl_pw_aff_dim(pa, isl_dim_param);
3940 for (int i = 0; i < nparam; ++i) {
3941 Expr *nested;
3942 isl_id *id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
3943 pet_expr *expr;
3944 bool allowed;
3946 if (!is_nested_parameter(id)) {
3947 isl_id_free(id);
3948 continue;
3951 if (pet_scop_has_skip(scop, pet_skip_now)) {
3952 isl_id_free(id);
3953 return false;
3956 nested = (Expr *) isl_id_get_user(id);
3957 expr = extract_expr(nested);
3958 allowed = expr && expr->type == pet_expr_access &&
3959 !is_assigned(expr, scop);
3961 pet_expr_free(expr);
3962 isl_id_free(id);
3964 if (!allowed)
3965 return false;
3968 return true;
3971 /* Do we need to construct a skip condition of the given type
3972 * on an if statement, given that the if condition is non-affine?
3974 * pet_scop_filter_skip can only handle the case where the if condition
3975 * holds (the then branch) and the skip condition is universal.
3976 * In any other case, we need to construct a new skip condition.
3978 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
3979 bool have_else, enum pet_skip type)
3981 if (have_else && scop_else && pet_scop_has_skip(scop_else, type))
3982 return true;
3983 if (scop_then && pet_scop_has_skip(scop_then, type) &&
3984 !pet_scop_has_universal_skip(scop_then, type))
3985 return true;
3986 return false;
3989 /* Do we need to construct a skip condition of the given type
3990 * on an if statement, given that the if condition is affine?
3992 * There is no need to construct a new skip condition if all
3993 * the skip conditions are affine.
3995 static bool need_skip_aff(struct pet_scop *scop_then,
3996 struct pet_scop *scop_else, bool have_else, enum pet_skip type)
3998 if (scop_then && pet_scop_has_var_skip(scop_then, type))
3999 return true;
4000 if (have_else && scop_else && pet_scop_has_var_skip(scop_else, type))
4001 return true;
4002 return false;
4005 /* Do we need to construct a skip condition of the given type
4006 * on an if statement?
4008 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
4009 bool have_else, enum pet_skip type, bool affine)
4011 if (affine)
4012 return need_skip_aff(scop_then, scop_else, have_else, type);
4013 else
4014 return need_skip(scop_then, scop_else, have_else, type);
4017 /* Construct an affine expression pet_expr that evaluates
4018 * to the constant "val".
4020 static struct pet_expr *universally(isl_ctx *ctx, int val)
4022 isl_local_space *ls;
4023 isl_aff *aff;
4024 isl_multi_pw_aff *mpa;
4026 ls = isl_local_space_from_space(isl_space_set_alloc(ctx, 0, 0));
4027 aff = isl_aff_val_on_domain(ls, isl_val_int_from_si(ctx, val));
4028 mpa = isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
4030 return pet_expr_from_index(mpa);
4033 /* Construct an affine expression pet_expr that evaluates
4034 * to the constant 1.
4036 static struct pet_expr *universally_true(isl_ctx *ctx)
4038 return universally(ctx, 1);
4041 /* Construct an affine expression pet_expr that evaluates
4042 * to the constant 0.
4044 static struct pet_expr *universally_false(isl_ctx *ctx)
4046 return universally(ctx, 0);
4049 /* Given an index expression "test_index" for the if condition,
4050 * an index expression "skip_index" for the skip condition and
4051 * scops for the then and else branches, construct a scop for
4052 * computing "skip_index".
4054 * The computed scop contains a single statement that essentially does
4056 * skip_index = test_cond ? skip_cond_then : skip_cond_else
4058 * If the skip conditions of the then and/or else branch are not affine,
4059 * then they need to be filtered by test_index.
4060 * If they are missing, then this means the skip condition is false.
4062 * Since we are constructing a skip condition for the if statement,
4063 * the skip conditions on the then and else branches are removed.
4065 static struct pet_scop *extract_skip(PetScan *scan,
4066 __isl_take isl_multi_pw_aff *test_index,
4067 __isl_take isl_multi_pw_aff *skip_index,
4068 struct pet_scop *scop_then, struct pet_scop *scop_else, bool have_else,
4069 enum pet_skip type)
4071 struct pet_expr *expr_then, *expr_else, *expr, *expr_skip;
4072 struct pet_stmt *stmt;
4073 struct pet_scop *scop;
4074 isl_ctx *ctx = scan->ctx;
4076 if (!scop_then)
4077 goto error;
4078 if (have_else && !scop_else)
4079 goto error;
4081 if (pet_scop_has_skip(scop_then, type)) {
4082 expr_then = pet_scop_get_skip_expr(scop_then, type);
4083 pet_scop_reset_skip(scop_then, type);
4084 if (!pet_expr_is_affine(expr_then))
4085 expr_then = pet_expr_filter(expr_then,
4086 isl_multi_pw_aff_copy(test_index), 1);
4087 } else
4088 expr_then = universally_false(ctx);
4090 if (have_else && pet_scop_has_skip(scop_else, type)) {
4091 expr_else = pet_scop_get_skip_expr(scop_else, type);
4092 pet_scop_reset_skip(scop_else, type);
4093 if (!pet_expr_is_affine(expr_else))
4094 expr_else = pet_expr_filter(expr_else,
4095 isl_multi_pw_aff_copy(test_index), 0);
4096 } else
4097 expr_else = universally_false(ctx);
4099 expr = pet_expr_from_index(test_index);
4100 expr = pet_expr_new_ternary(ctx, expr, expr_then, expr_else);
4101 expr_skip = pet_expr_from_index(isl_multi_pw_aff_copy(skip_index));
4102 if (expr_skip) {
4103 expr_skip->acc.write = 1;
4104 expr_skip->acc.read = 0;
4106 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
4107 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, scan->n_stmt++, expr);
4109 scop = pet_scop_from_pet_stmt(ctx, stmt);
4110 scop = scop_add_array(scop, skip_index, scan->ast_context);
4111 isl_multi_pw_aff_free(skip_index);
4113 return scop;
4114 error:
4115 isl_multi_pw_aff_free(test_index);
4116 isl_multi_pw_aff_free(skip_index);
4117 return NULL;
4120 /* Is scop's skip_now condition equal to its skip_later condition?
4121 * In particular, this means that it either has no skip_now condition
4122 * or both a skip_now and a skip_later condition (that are equal to each other).
4124 static bool skip_equals_skip_later(struct pet_scop *scop)
4126 int has_skip_now, has_skip_later;
4127 int equal;
4128 isl_multi_pw_aff *skip_now, *skip_later;
4130 if (!scop)
4131 return false;
4132 has_skip_now = pet_scop_has_skip(scop, pet_skip_now);
4133 has_skip_later = pet_scop_has_skip(scop, pet_skip_later);
4134 if (has_skip_now != has_skip_later)
4135 return false;
4136 if (!has_skip_now)
4137 return true;
4139 skip_now = pet_scop_get_skip(scop, pet_skip_now);
4140 skip_later = pet_scop_get_skip(scop, pet_skip_later);
4141 equal = isl_multi_pw_aff_is_equal(skip_now, skip_later);
4142 isl_multi_pw_aff_free(skip_now);
4143 isl_multi_pw_aff_free(skip_later);
4145 return equal;
4148 /* Drop the skip conditions of type pet_skip_later from scop1 and scop2.
4150 static void drop_skip_later(struct pet_scop *scop1, struct pet_scop *scop2)
4152 pet_scop_reset_skip(scop1, pet_skip_later);
4153 pet_scop_reset_skip(scop2, pet_skip_later);
4156 /* Structure that handles the construction of skip conditions.
4158 * scop_then and scop_else represent the then and else branches
4159 * of the if statement
4161 * skip[type] is true if we need to construct a skip condition of that type
4162 * equal is set if the skip conditions of types pet_skip_now and pet_skip_later
4163 * are equal to each other
4164 * index[type] is an index expression from a zero-dimension domain
4165 * to the virtual array representing the skip condition
4166 * scop[type] is a scop for computing the skip condition
4168 struct pet_skip_info {
4169 isl_ctx *ctx;
4171 bool skip[2];
4172 bool equal;
4173 isl_multi_pw_aff *index[2];
4174 struct pet_scop *scop[2];
4176 pet_skip_info(isl_ctx *ctx) : ctx(ctx) {}
4178 operator bool() { return skip[pet_skip_now] || skip[pet_skip_later]; }
4181 /* Structure that handles the construction of skip conditions on if statements.
4183 * scop_then and scop_else represent the then and else branches
4184 * of the if statement
4186 struct pet_skip_info_if : public pet_skip_info {
4187 struct pet_scop *scop_then, *scop_else;
4188 bool have_else;
4190 pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
4191 struct pet_scop *scop_else, bool have_else, bool affine);
4192 void extract(PetScan *scan, __isl_keep isl_multi_pw_aff *index,
4193 enum pet_skip type);
4194 void extract(PetScan *scan, __isl_keep isl_multi_pw_aff *index);
4195 void extract(PetScan *scan, __isl_keep isl_pw_aff *cond);
4196 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
4197 int offset);
4198 struct pet_scop *add(struct pet_scop *scop, int offset);
4201 /* Initialize a pet_skip_info_if structure based on the then and else branches
4202 * and based on whether the if condition is affine or not.
4204 pet_skip_info_if::pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
4205 struct pet_scop *scop_else, bool have_else, bool affine) :
4206 pet_skip_info(ctx), scop_then(scop_then), scop_else(scop_else),
4207 have_else(have_else)
4209 skip[pet_skip_now] =
4210 need_skip(scop_then, scop_else, have_else, pet_skip_now, affine);
4211 equal = skip[pet_skip_now] && skip_equals_skip_later(scop_then) &&
4212 (!have_else || skip_equals_skip_later(scop_else));
4213 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
4214 need_skip(scop_then, scop_else, have_else, pet_skip_later, affine);
4217 /* If we need to construct a skip condition of the given type,
4218 * then do so now.
4220 * "mpa" represents the if condition.
4222 void pet_skip_info_if::extract(PetScan *scan,
4223 __isl_keep isl_multi_pw_aff *mpa, enum pet_skip type)
4225 isl_ctx *ctx;
4227 if (!skip[type])
4228 return;
4230 ctx = isl_multi_pw_aff_get_ctx(mpa);
4231 index[type] = create_test_index(ctx, scan->n_test++);
4232 scop[type] = extract_skip(scan, isl_multi_pw_aff_copy(mpa),
4233 isl_multi_pw_aff_copy(index[type]),
4234 scop_then, scop_else, have_else, type);
4237 /* Construct the required skip conditions, given the if condition "index".
4239 void pet_skip_info_if::extract(PetScan *scan,
4240 __isl_keep isl_multi_pw_aff *index)
4242 extract(scan, index, pet_skip_now);
4243 extract(scan, index, pet_skip_later);
4244 if (equal)
4245 drop_skip_later(scop_then, scop_else);
4248 /* Construct the required skip conditions, given the if condition "cond".
4250 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_pw_aff *cond)
4252 isl_multi_pw_aff *test;
4254 if (!skip[pet_skip_now] && !skip[pet_skip_later])
4255 return;
4257 test = isl_multi_pw_aff_from_pw_aff(isl_pw_aff_copy(cond));
4258 test = isl_multi_pw_aff_from_range(test);
4259 extract(scan, test);
4260 isl_multi_pw_aff_free(test);
4263 /* Add the computed skip condition of the give type to "main" and
4264 * add the scop for computing the condition at the given offset.
4266 * If equal is set, then we only computed a skip condition for pet_skip_now,
4267 * but we also need to set it as main's pet_skip_later.
4269 struct pet_scop *pet_skip_info_if::add(struct pet_scop *main,
4270 enum pet_skip type, int offset)
4272 if (!skip[type])
4273 return main;
4275 scop[type] = pet_scop_prefix(scop[type], offset);
4276 main = pet_scop_add_par(ctx, main, scop[type]);
4277 scop[type] = NULL;
4279 if (equal)
4280 main = pet_scop_set_skip(main, pet_skip_later,
4281 isl_multi_pw_aff_copy(index[type]));
4283 main = pet_scop_set_skip(main, type, index[type]);
4284 index[type] = NULL;
4286 return main;
4289 /* Add the computed skip conditions to "main" and
4290 * add the scops for computing the conditions at the given offset.
4292 struct pet_scop *pet_skip_info_if::add(struct pet_scop *scop, int offset)
4294 scop = add(scop, pet_skip_now, offset);
4295 scop = add(scop, pet_skip_later, offset);
4297 return scop;
4300 /* Construct a pet_scop for a non-affine if statement.
4302 * We create a separate statement that writes the result
4303 * of the non-affine condition to a virtual scalar.
4304 * A constraint requiring the value of this virtual scalar to be one
4305 * is added to the iteration domains of the then branch.
4306 * Similarly, a constraint requiring the value of this virtual scalar
4307 * to be zero is added to the iteration domains of the else branch, if any.
4308 * We adjust the schedules to ensure that the virtual scalar is written
4309 * before it is read.
4311 * If there are any breaks or continues in the then and/or else
4312 * branches, then we may have to compute a new skip condition.
4313 * This is handled using a pet_skip_info_if object.
4314 * On initialization, the object checks if skip conditions need
4315 * to be computed. If so, it does so in "extract" and adds them in "add".
4317 struct pet_scop *PetScan::extract_non_affine_if(Expr *cond,
4318 struct pet_scop *scop_then, struct pet_scop *scop_else,
4319 bool have_else, int stmt_id)
4321 struct pet_scop *scop;
4322 isl_multi_pw_aff *test_index;
4323 int save_n_stmt = n_stmt;
4325 test_index = create_test_index(ctx, n_test++);
4326 n_stmt = stmt_id;
4327 scop = extract_non_affine_condition(cond,
4328 isl_multi_pw_aff_copy(test_index));
4329 n_stmt = save_n_stmt;
4330 scop = scop_add_array(scop, test_index, ast_context);
4332 pet_skip_info_if skip(ctx, scop_then, scop_else, have_else, false);
4333 skip.extract(this, test_index);
4335 scop = pet_scop_prefix(scop, 0);
4336 scop_then = pet_scop_prefix(scop_then, 1);
4337 scop_then = pet_scop_filter(scop_then,
4338 isl_multi_pw_aff_copy(test_index), 1);
4339 if (have_else) {
4340 scop_else = pet_scop_prefix(scop_else, 1);
4341 scop_else = pet_scop_filter(scop_else, test_index, 0);
4342 scop_then = pet_scop_add_par(ctx, scop_then, scop_else);
4343 } else
4344 isl_multi_pw_aff_free(test_index);
4346 scop = pet_scop_add_seq(ctx, scop, scop_then);
4348 scop = skip.add(scop, 2);
4350 return scop;
4353 /* Construct a pet_scop for an if statement.
4355 * If the condition fits the pattern of a conditional assignment,
4356 * then it is handled by extract_conditional_assignment.
4357 * Otherwise, we do the following.
4359 * If the condition is affine, then the condition is added
4360 * to the iteration domains of the then branch, while the
4361 * opposite of the condition in added to the iteration domains
4362 * of the else branch, if any.
4363 * We allow the condition to be dynamic, i.e., to refer to
4364 * scalars or array elements that may be written to outside
4365 * of the given if statement. These nested accesses are then represented
4366 * as output dimensions in the wrapping iteration domain.
4367 * If it also written _inside_ the then or else branch, then
4368 * we treat the condition as non-affine.
4369 * As explained in extract_non_affine_if, this will introduce
4370 * an extra statement.
4371 * For aesthetic reasons, we want this statement to have a statement
4372 * number that is lower than those of the then and else branches.
4373 * In order to evaluate if will need such a statement, however, we
4374 * first construct scops for the then and else branches.
4375 * We therefore reserve a statement number if we might have to
4376 * introduce such an extra statement.
4378 * If the condition is not affine, then the scop is created in
4379 * extract_non_affine_if.
4381 * If there are any breaks or continues in the then and/or else
4382 * branches, then we may have to compute a new skip condition.
4383 * This is handled using a pet_skip_info_if object.
4384 * On initialization, the object checks if skip conditions need
4385 * to be computed. If so, it does so in "extract" and adds them in "add".
4387 struct pet_scop *PetScan::extract(IfStmt *stmt)
4389 struct pet_scop *scop_then, *scop_else = NULL, *scop;
4390 isl_pw_aff *cond;
4391 int stmt_id;
4392 isl_set *set;
4393 isl_set *valid;
4395 scop = extract_conditional_assignment(stmt);
4396 if (scop)
4397 return scop;
4399 cond = try_extract_nested_condition(stmt->getCond());
4400 if (allow_nested && (!cond || has_nested(cond)))
4401 stmt_id = n_stmt++;
4404 assigned_value_cache cache(assigned_value);
4405 scop_then = extract(stmt->getThen());
4408 if (stmt->getElse()) {
4409 assigned_value_cache cache(assigned_value);
4410 scop_else = extract(stmt->getElse());
4411 if (options->autodetect) {
4412 if (scop_then && !scop_else) {
4413 partial = true;
4414 isl_pw_aff_free(cond);
4415 return scop_then;
4417 if (!scop_then && scop_else) {
4418 partial = true;
4419 isl_pw_aff_free(cond);
4420 return scop_else;
4425 if (cond &&
4426 (!is_nested_allowed(cond, scop_then) ||
4427 (stmt->getElse() && !is_nested_allowed(cond, scop_else)))) {
4428 isl_pw_aff_free(cond);
4429 cond = NULL;
4431 if (allow_nested && !cond)
4432 return extract_non_affine_if(stmt->getCond(), scop_then,
4433 scop_else, stmt->getElse(), stmt_id);
4435 if (!cond)
4436 cond = extract_condition(stmt->getCond());
4438 pet_skip_info_if skip(ctx, scop_then, scop_else, stmt->getElse(), true);
4439 skip.extract(this, cond);
4441 valid = isl_pw_aff_domain(isl_pw_aff_copy(cond));
4442 set = isl_pw_aff_non_zero_set(cond);
4443 scop = pet_scop_restrict(scop_then, isl_set_copy(set));
4445 if (stmt->getElse()) {
4446 set = isl_set_subtract(isl_set_copy(valid), set);
4447 scop_else = pet_scop_restrict(scop_else, set);
4448 scop = pet_scop_add_par(ctx, scop, scop_else);
4449 } else
4450 isl_set_free(set);
4451 scop = resolve_nested(scop);
4452 scop = pet_scop_restrict_context(scop, valid);
4454 if (skip)
4455 scop = pet_scop_prefix(scop, 0);
4456 scop = skip.add(scop, 1);
4458 return scop;
4461 /* Try and construct a pet_scop for a label statement.
4462 * We currently only allow labels on expression statements.
4464 struct pet_scop *PetScan::extract(LabelStmt *stmt)
4466 isl_id *label;
4467 Stmt *sub;
4469 sub = stmt->getSubStmt();
4470 if (!isa<Expr>(sub)) {
4471 unsupported(stmt);
4472 return NULL;
4475 label = isl_id_alloc(ctx, stmt->getName(), NULL);
4477 return extract(sub, extract_expr(cast<Expr>(sub)), label);
4480 /* Return a one-dimensional multi piecewise affine expression that is equal
4481 * to the constant 1 and is defined over a zero-dimensional domain.
4483 static __isl_give isl_multi_pw_aff *one_mpa(isl_ctx *ctx)
4485 isl_space *space;
4486 isl_local_space *ls;
4487 isl_aff *aff;
4489 space = isl_space_set_alloc(ctx, 0, 0);
4490 ls = isl_local_space_from_space(space);
4491 aff = isl_aff_zero_on_domain(ls);
4492 aff = isl_aff_set_constant_si(aff, 1);
4494 return isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
4497 /* Construct a pet_scop for a continue statement.
4499 * We simply create an empty scop with a universal pet_skip_now
4500 * skip condition. This skip condition will then be taken into
4501 * account by the enclosing loop construct, possibly after
4502 * being incorporated into outer skip conditions.
4504 struct pet_scop *PetScan::extract(ContinueStmt *stmt)
4506 pet_scop *scop;
4508 scop = pet_scop_empty(ctx);
4509 if (!scop)
4510 return NULL;
4512 scop = pet_scop_set_skip(scop, pet_skip_now, one_mpa(ctx));
4514 return scop;
4517 /* Construct a pet_scop for a break statement.
4519 * We simply create an empty scop with both a universal pet_skip_now
4520 * skip condition and a universal pet_skip_later skip condition.
4521 * These skip conditions will then be taken into
4522 * account by the enclosing loop construct, possibly after
4523 * being incorporated into outer skip conditions.
4525 struct pet_scop *PetScan::extract(BreakStmt *stmt)
4527 pet_scop *scop;
4528 isl_multi_pw_aff *skip;
4530 scop = pet_scop_empty(ctx);
4531 if (!scop)
4532 return NULL;
4534 skip = one_mpa(ctx);
4535 scop = pet_scop_set_skip(scop, pet_skip_now,
4536 isl_multi_pw_aff_copy(skip));
4537 scop = pet_scop_set_skip(scop, pet_skip_later, skip);
4539 return scop;
4542 /* Try and construct a pet_scop corresponding to "stmt".
4544 * If "stmt" is a compound statement, then "skip_declarations"
4545 * indicates whether we should skip initial declarations in the
4546 * compound statement.
4548 * If the constructed pet_scop is not a (possibly) partial representation
4549 * of "stmt", we update start and end of the pet_scop to those of "stmt".
4550 * In particular, if skip_declarations, then we may have skipped declarations
4551 * inside "stmt" and so the pet_scop may not represent the entire "stmt".
4552 * Note that this function may be called with "stmt" referring to the entire
4553 * body of the function, including the outer braces. In such cases,
4554 * skip_declarations will be set and the braces will not be taken into
4555 * account in scop->start and scop->end.
4557 struct pet_scop *PetScan::extract(Stmt *stmt, bool skip_declarations)
4559 struct pet_scop *scop;
4560 unsigned start, end;
4561 SourceLocation loc;
4562 SourceManager &SM = PP.getSourceManager();
4563 const LangOptions &LO = PP.getLangOpts();
4565 if (isa<Expr>(stmt))
4566 return extract(stmt, extract_expr(cast<Expr>(stmt)));
4568 switch (stmt->getStmtClass()) {
4569 case Stmt::WhileStmtClass:
4570 scop = extract(cast<WhileStmt>(stmt));
4571 break;
4572 case Stmt::ForStmtClass:
4573 scop = extract_for(cast<ForStmt>(stmt));
4574 break;
4575 case Stmt::IfStmtClass:
4576 scop = extract(cast<IfStmt>(stmt));
4577 break;
4578 case Stmt::CompoundStmtClass:
4579 scop = extract(cast<CompoundStmt>(stmt), skip_declarations);
4580 break;
4581 case Stmt::LabelStmtClass:
4582 scop = extract(cast<LabelStmt>(stmt));
4583 break;
4584 case Stmt::ContinueStmtClass:
4585 scop = extract(cast<ContinueStmt>(stmt));
4586 break;
4587 case Stmt::BreakStmtClass:
4588 scop = extract(cast<BreakStmt>(stmt));
4589 break;
4590 case Stmt::DeclStmtClass:
4591 scop = extract(cast<DeclStmt>(stmt));
4592 break;
4593 default:
4594 unsupported(stmt);
4595 return NULL;
4598 if (partial || skip_declarations)
4599 return scop;
4601 loc = stmt->getLocStart();
4602 loc = move_to_start_of_line_if_first_token(loc, SM, LO);
4603 start = getExpansionOffset(SM, loc);
4604 loc = PP.getLocForEndOfToken(stmt->getLocEnd());
4605 end = getExpansionOffset(SM, loc);
4606 scop = pet_scop_update_start_end(scop, start, end);
4608 return scop;
4611 /* Do we need to construct a skip condition of the given type
4612 * on a sequence of statements?
4614 * There is no need to construct a new skip condition if only
4615 * only of the two statements has a skip condition or if both
4616 * of their skip conditions are affine.
4618 * In principle we also don't need a new continuation variable if
4619 * the continuation of scop2 is affine, but then we would need
4620 * to allow more complicated forms of continuations.
4622 static bool need_skip_seq(struct pet_scop *scop1, struct pet_scop *scop2,
4623 enum pet_skip type)
4625 if (!scop1 || !pet_scop_has_skip(scop1, type))
4626 return false;
4627 if (!scop2 || !pet_scop_has_skip(scop2, type))
4628 return false;
4629 if (pet_scop_has_affine_skip(scop1, type) &&
4630 pet_scop_has_affine_skip(scop2, type))
4631 return false;
4632 return true;
4635 /* Construct a scop for computing the skip condition of the given type and
4636 * with index expression "skip_index" for a sequence of two scops "scop1"
4637 * and "scop2".
4639 * The computed scop contains a single statement that essentially does
4641 * skip_index = skip_cond_1 ? 1 : skip_cond_2
4643 * or, in other words, skip_cond1 || skip_cond2.
4644 * In this expression, skip_cond_2 is filtered to reflect that it is
4645 * only evaluated when skip_cond_1 is false.
4647 * The skip condition on scop1 is not removed because it still needs
4648 * to be applied to scop2 when these two scops are combined.
4650 static struct pet_scop *extract_skip_seq(PetScan *ps,
4651 __isl_take isl_multi_pw_aff *skip_index,
4652 struct pet_scop *scop1, struct pet_scop *scop2, enum pet_skip type)
4654 struct pet_expr *expr1, *expr2, *expr, *expr_skip;
4655 struct pet_stmt *stmt;
4656 struct pet_scop *scop;
4657 isl_ctx *ctx = ps->ctx;
4659 if (!scop1 || !scop2)
4660 goto error;
4662 expr1 = pet_scop_get_skip_expr(scop1, type);
4663 expr2 = pet_scop_get_skip_expr(scop2, type);
4664 pet_scop_reset_skip(scop2, type);
4666 expr2 = pet_expr_filter(expr2,
4667 isl_multi_pw_aff_copy(expr1->acc.index), 0);
4669 expr = universally_true(ctx);
4670 expr = pet_expr_new_ternary(ctx, expr1, expr, expr2);
4671 expr_skip = pet_expr_from_index(isl_multi_pw_aff_copy(skip_index));
4672 if (expr_skip) {
4673 expr_skip->acc.write = 1;
4674 expr_skip->acc.read = 0;
4676 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
4677 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, ps->n_stmt++, expr);
4679 scop = pet_scop_from_pet_stmt(ctx, stmt);
4680 scop = scop_add_array(scop, skip_index, ps->ast_context);
4681 isl_multi_pw_aff_free(skip_index);
4683 return scop;
4684 error:
4685 isl_multi_pw_aff_free(skip_index);
4686 return NULL;
4689 /* Structure that handles the construction of skip conditions
4690 * on sequences of statements.
4692 * scop1 and scop2 represent the two statements that are combined
4694 struct pet_skip_info_seq : public pet_skip_info {
4695 struct pet_scop *scop1, *scop2;
4697 pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4698 struct pet_scop *scop2);
4699 void extract(PetScan *scan, enum pet_skip type);
4700 void extract(PetScan *scan);
4701 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
4702 int offset);
4703 struct pet_scop *add(struct pet_scop *scop, int offset);
4706 /* Initialize a pet_skip_info_seq structure based on
4707 * on the two statements that are going to be combined.
4709 pet_skip_info_seq::pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4710 struct pet_scop *scop2) : pet_skip_info(ctx), scop1(scop1), scop2(scop2)
4712 skip[pet_skip_now] = need_skip_seq(scop1, scop2, pet_skip_now);
4713 equal = skip[pet_skip_now] && skip_equals_skip_later(scop1) &&
4714 skip_equals_skip_later(scop2);
4715 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
4716 need_skip_seq(scop1, scop2, pet_skip_later);
4719 /* If we need to construct a skip condition of the given type,
4720 * then do so now.
4722 void pet_skip_info_seq::extract(PetScan *scan, enum pet_skip type)
4724 if (!skip[type])
4725 return;
4727 index[type] = create_test_index(ctx, scan->n_test++);
4728 scop[type] = extract_skip_seq(scan, isl_multi_pw_aff_copy(index[type]),
4729 scop1, scop2, type);
4732 /* Construct the required skip conditions.
4734 void pet_skip_info_seq::extract(PetScan *scan)
4736 extract(scan, pet_skip_now);
4737 extract(scan, pet_skip_later);
4738 if (equal)
4739 drop_skip_later(scop1, scop2);
4742 /* Add the computed skip condition of the given type to "main" and
4743 * add the scop for computing the condition at the given offset (the statement
4744 * number). Within this offset, the condition is computed at position 1
4745 * to ensure that it is computed after the corresponding statement.
4747 * If equal is set, then we only computed a skip condition for pet_skip_now,
4748 * but we also need to set it as main's pet_skip_later.
4750 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *main,
4751 enum pet_skip type, int offset)
4753 if (!skip[type])
4754 return main;
4756 scop[type] = pet_scop_prefix(scop[type], 1);
4757 scop[type] = pet_scop_prefix(scop[type], offset);
4758 main = pet_scop_add_par(ctx, main, scop[type]);
4759 scop[type] = NULL;
4761 if (equal)
4762 main = pet_scop_set_skip(main, pet_skip_later,
4763 isl_multi_pw_aff_copy(index[type]));
4765 main = pet_scop_set_skip(main, type, index[type]);
4766 index[type] = NULL;
4768 return main;
4771 /* Add the computed skip conditions to "main" and
4772 * add the scops for computing the conditions at the given offset.
4774 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *scop, int offset)
4776 scop = add(scop, pet_skip_now, offset);
4777 scop = add(scop, pet_skip_later, offset);
4779 return scop;
4782 /* Extract a clone of the kill statement in "scop".
4783 * "scop" is expected to have been created from a DeclStmt
4784 * and should have the kill as its first statement.
4786 struct pet_stmt *PetScan::extract_kill(struct pet_scop *scop)
4788 struct pet_expr *kill;
4789 struct pet_stmt *stmt;
4790 isl_multi_pw_aff *index;
4791 isl_map *access;
4793 if (!scop)
4794 return NULL;
4795 if (scop->n_stmt < 1)
4796 isl_die(ctx, isl_error_internal,
4797 "expecting at least one statement", return NULL);
4798 stmt = scop->stmts[0];
4799 if (stmt->body->type != pet_expr_unary ||
4800 stmt->body->op != pet_op_kill)
4801 isl_die(ctx, isl_error_internal,
4802 "expecting kill statement", return NULL);
4804 index = isl_multi_pw_aff_copy(stmt->body->args[0]->acc.index);
4805 access = isl_map_copy(stmt->body->args[0]->acc.access);
4806 index = isl_multi_pw_aff_reset_tuple_id(index, isl_dim_in);
4807 access = isl_map_reset_tuple_id(access, isl_dim_in);
4808 kill = pet_expr_kill_from_access_and_index(access, index);
4809 return pet_stmt_from_pet_expr(ctx, stmt->line, NULL, n_stmt++, kill);
4812 /* Mark all arrays in "scop" as being exposed.
4814 static struct pet_scop *mark_exposed(struct pet_scop *scop)
4816 if (!scop)
4817 return NULL;
4818 for (int i = 0; i < scop->n_array; ++i)
4819 scop->arrays[i]->exposed = 1;
4820 return scop;
4823 /* Try and construct a pet_scop corresponding to (part of)
4824 * a sequence of statements.
4826 * "block" is set if the sequence respresents the children of
4827 * a compound statement.
4828 * "skip_declarations" is set if we should skip initial declarations
4829 * in the sequence of statements.
4831 * If there are any breaks or continues in the individual statements,
4832 * then we may have to compute a new skip condition.
4833 * This is handled using a pet_skip_info_seq object.
4834 * On initialization, the object checks if skip conditions need
4835 * to be computed. If so, it does so in "extract" and adds them in "add".
4837 * If "block" is set, then we need to insert kill statements at
4838 * the end of the block for any array that has been declared by
4839 * one of the statements in the sequence. Each of these declarations
4840 * results in the construction of a kill statement at the place
4841 * of the declaration, so we simply collect duplicates of
4842 * those kill statements and append these duplicates to the constructed scop.
4844 * If "block" is not set, then any array declared by one of the statements
4845 * in the sequence is marked as being exposed.
4847 * If autodetect is set, then we allow the extraction of only a subrange
4848 * of the sequence of statements. However, if there is at least one statement
4849 * for which we could not construct a scop and the final range contains
4850 * either no statements or at least one kill, then we discard the entire
4851 * range.
4853 struct pet_scop *PetScan::extract(StmtRange stmt_range, bool block,
4854 bool skip_declarations)
4856 pet_scop *scop;
4857 StmtIterator i;
4858 int j;
4859 bool partial_range = false;
4860 set<struct pet_stmt *> kills;
4861 set<struct pet_stmt *>::iterator it;
4863 scop = pet_scop_empty(ctx);
4864 for (i = stmt_range.first, j = 0; i != stmt_range.second; ++i, ++j) {
4865 Stmt *child = *i;
4866 struct pet_scop *scop_i;
4868 if (skip_declarations &&
4869 child->getStmtClass() == Stmt::DeclStmtClass)
4870 continue;
4872 scop_i = extract(child);
4873 if (scop->n_stmt != 0 && partial) {
4874 pet_scop_free(scop_i);
4875 break;
4877 pet_skip_info_seq skip(ctx, scop, scop_i);
4878 skip.extract(this);
4879 if (skip)
4880 scop_i = pet_scop_prefix(scop_i, 0);
4881 if (scop_i && child->getStmtClass() == Stmt::DeclStmtClass) {
4882 if (block)
4883 kills.insert(extract_kill(scop_i));
4884 else
4885 scop_i = mark_exposed(scop_i);
4887 scop_i = pet_scop_prefix(scop_i, j);
4888 if (options->autodetect) {
4889 if (scop_i)
4890 scop = pet_scop_add_seq(ctx, scop, scop_i);
4891 else
4892 partial_range = true;
4893 if (scop->n_stmt != 0 && !scop_i)
4894 partial = true;
4895 } else {
4896 scop = pet_scop_add_seq(ctx, scop, scop_i);
4899 scop = skip.add(scop, j);
4901 if (partial || !scop)
4902 break;
4905 for (it = kills.begin(); it != kills.end(); ++it) {
4906 pet_scop *scop_j;
4907 scop_j = pet_scop_from_pet_stmt(ctx, *it);
4908 scop_j = pet_scop_prefix(scop_j, j);
4909 scop = pet_scop_add_seq(ctx, scop, scop_j);
4912 if (scop && partial_range) {
4913 if (scop->n_stmt == 0 || kills.size() != 0) {
4914 pet_scop_free(scop);
4915 return NULL;
4917 partial = true;
4920 return scop;
4923 /* Check if the scop marked by the user is exactly this Stmt
4924 * or part of this Stmt.
4925 * If so, return a pet_scop corresponding to the marked region.
4926 * Otherwise, return NULL.
4928 struct pet_scop *PetScan::scan(Stmt *stmt)
4930 SourceManager &SM = PP.getSourceManager();
4931 unsigned start_off, end_off;
4933 start_off = getExpansionOffset(SM, stmt->getLocStart());
4934 end_off = getExpansionOffset(SM, stmt->getLocEnd());
4936 if (start_off > loc.end)
4937 return NULL;
4938 if (end_off < loc.start)
4939 return NULL;
4940 if (start_off >= loc.start && end_off <= loc.end) {
4941 return extract(stmt);
4944 StmtIterator start;
4945 for (start = stmt->child_begin(); start != stmt->child_end(); ++start) {
4946 Stmt *child = *start;
4947 if (!child)
4948 continue;
4949 start_off = getExpansionOffset(SM, child->getLocStart());
4950 end_off = getExpansionOffset(SM, child->getLocEnd());
4951 if (start_off < loc.start && end_off >= loc.end)
4952 return scan(child);
4953 if (start_off >= loc.start)
4954 break;
4957 StmtIterator end;
4958 for (end = start; end != stmt->child_end(); ++end) {
4959 Stmt *child = *end;
4960 start_off = SM.getFileOffset(child->getLocStart());
4961 if (start_off >= loc.end)
4962 break;
4965 return extract(StmtRange(start, end), false, false);
4968 /* Set the size of index "pos" of "array" to "size".
4969 * In particular, add a constraint of the form
4971 * i_pos < size
4973 * to array->extent and a constraint of the form
4975 * size >= 0
4977 * to array->context.
4979 static struct pet_array *update_size(struct pet_array *array, int pos,
4980 __isl_take isl_pw_aff *size)
4982 isl_set *valid;
4983 isl_set *univ;
4984 isl_set *bound;
4985 isl_space *dim;
4986 isl_aff *aff;
4987 isl_pw_aff *index;
4988 isl_id *id;
4990 valid = isl_pw_aff_nonneg_set(isl_pw_aff_copy(size));
4991 array->context = isl_set_intersect(array->context, valid);
4993 dim = isl_set_get_space(array->extent);
4994 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
4995 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, pos, 1);
4996 univ = isl_set_universe(isl_aff_get_domain_space(aff));
4997 index = isl_pw_aff_alloc(univ, aff);
4999 size = isl_pw_aff_add_dims(size, isl_dim_in,
5000 isl_set_dim(array->extent, isl_dim_set));
5001 id = isl_set_get_tuple_id(array->extent);
5002 size = isl_pw_aff_set_tuple_id(size, isl_dim_in, id);
5003 bound = isl_pw_aff_lt_set(index, size);
5005 array->extent = isl_set_intersect(array->extent, bound);
5007 if (!array->context || !array->extent)
5008 goto error;
5010 return array;
5011 error:
5012 pet_array_free(array);
5013 return NULL;
5016 /* Figure out the size of the array at position "pos" and all
5017 * subsequent positions from "type" and update "array" accordingly.
5019 struct pet_array *PetScan::set_upper_bounds(struct pet_array *array,
5020 const Type *type, int pos)
5022 const ArrayType *atype;
5023 isl_pw_aff *size;
5025 if (!array)
5026 return NULL;
5028 if (type->isPointerType()) {
5029 type = type->getPointeeType().getTypePtr();
5030 return set_upper_bounds(array, type, pos + 1);
5032 if (!type->isArrayType())
5033 return array;
5035 type = type->getCanonicalTypeInternal().getTypePtr();
5036 atype = cast<ArrayType>(type);
5038 if (type->isConstantArrayType()) {
5039 const ConstantArrayType *ca = cast<ConstantArrayType>(atype);
5040 size = extract_affine(ca->getSize());
5041 array = update_size(array, pos, size);
5042 } else if (type->isVariableArrayType()) {
5043 const VariableArrayType *vla = cast<VariableArrayType>(atype);
5044 size = extract_affine(vla->getSizeExpr());
5045 array = update_size(array, pos, size);
5048 type = atype->getElementType().getTypePtr();
5050 return set_upper_bounds(array, type, pos + 1);
5053 /* Is "T" the type of a variable length array with static size?
5055 static bool is_vla_with_static_size(QualType T)
5057 const VariableArrayType *vlatype;
5059 if (!T->isVariableArrayType())
5060 return false;
5061 vlatype = cast<VariableArrayType>(T);
5062 return vlatype->getSizeModifier() == VariableArrayType::Static;
5065 /* Return the type of "decl" as an array.
5067 * In particular, if "decl" is a parameter declaration that
5068 * is a variable length array with a static size, then
5069 * return the original type (i.e., the variable length array).
5070 * Otherwise, return the type of decl.
5072 static QualType get_array_type(ValueDecl *decl)
5074 ParmVarDecl *parm;
5075 QualType T;
5077 parm = dyn_cast<ParmVarDecl>(decl);
5078 if (!parm)
5079 return decl->getType();
5081 T = parm->getOriginalType();
5082 if (!is_vla_with_static_size(T))
5083 return decl->getType();
5084 return T;
5087 /* Construct and return a pet_array corresponding to the variable "decl".
5088 * In particular, initialize array->extent to
5090 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
5092 * and then call set_upper_bounds to set the upper bounds on the indices
5093 * based on the type of the variable.
5095 struct pet_array *PetScan::extract_array(isl_ctx *ctx, ValueDecl *decl)
5097 struct pet_array *array;
5098 QualType qt = get_array_type(decl);
5099 const Type *type = qt.getTypePtr();
5100 int depth = array_depth(type);
5101 QualType base = pet_clang_base_type(qt);
5102 string name;
5103 isl_id *id;
5104 isl_space *dim;
5106 array = isl_calloc_type(ctx, struct pet_array);
5107 if (!array)
5108 return NULL;
5110 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
5111 dim = isl_space_set_alloc(ctx, 0, depth);
5112 dim = isl_space_set_tuple_id(dim, isl_dim_set, id);
5114 array->extent = isl_set_nat_universe(dim);
5116 dim = isl_space_params_alloc(ctx, 0);
5117 array->context = isl_set_universe(dim);
5119 array = set_upper_bounds(array, type, 0);
5120 if (!array)
5121 return NULL;
5123 name = base.getAsString();
5124 array->element_type = strdup(name.c_str());
5125 array->element_size = decl->getASTContext().getTypeInfo(base).first / 8;
5127 return array;
5130 /* Construct a list of pet_arrays, one for each array (or scalar)
5131 * accessed inside "scop", add this list to "scop" and return the result.
5133 * The context of "scop" is updated with the intersection of
5134 * the contexts of all arrays, i.e., constraints on the parameters
5135 * that ensure that the arrays have a valid (non-negative) size.
5137 struct pet_scop *PetScan::scan_arrays(struct pet_scop *scop)
5139 int i;
5140 set<ValueDecl *> arrays;
5141 set<ValueDecl *>::iterator it;
5142 int n_array;
5143 struct pet_array **scop_arrays;
5145 if (!scop)
5146 return NULL;
5148 pet_scop_collect_arrays(scop, arrays);
5149 if (arrays.size() == 0)
5150 return scop;
5152 n_array = scop->n_array;
5154 scop_arrays = isl_realloc_array(ctx, scop->arrays, struct pet_array *,
5155 n_array + arrays.size());
5156 if (!scop_arrays)
5157 goto error;
5158 scop->arrays = scop_arrays;
5160 for (it = arrays.begin(), i = 0; it != arrays.end(); ++it, ++i) {
5161 struct pet_array *array;
5162 scop->arrays[n_array + i] = array = extract_array(ctx, *it);
5163 if (!scop->arrays[n_array + i])
5164 goto error;
5165 scop->n_array++;
5166 scop->context = isl_set_intersect(scop->context,
5167 isl_set_copy(array->context));
5168 if (!scop->context)
5169 goto error;
5172 return scop;
5173 error:
5174 pet_scop_free(scop);
5175 return NULL;
5178 /* Bound all parameters in scop->context to the possible values
5179 * of the corresponding C variable.
5181 static struct pet_scop *add_parameter_bounds(struct pet_scop *scop)
5183 int n;
5185 if (!scop)
5186 return NULL;
5188 n = isl_set_dim(scop->context, isl_dim_param);
5189 for (int i = 0; i < n; ++i) {
5190 isl_id *id;
5191 ValueDecl *decl;
5193 id = isl_set_get_dim_id(scop->context, isl_dim_param, i);
5194 if (is_nested_parameter(id)) {
5195 isl_id_free(id);
5196 isl_die(isl_set_get_ctx(scop->context),
5197 isl_error_internal,
5198 "unresolved nested parameter", goto error);
5200 decl = (ValueDecl *) isl_id_get_user(id);
5201 isl_id_free(id);
5203 scop->context = set_parameter_bounds(scop->context, i, decl);
5205 if (!scop->context)
5206 goto error;
5209 return scop;
5210 error:
5211 pet_scop_free(scop);
5212 return NULL;
5215 /* Construct a pet_scop from the given function.
5217 * If the scop was delimited by scop and endscop pragmas, then we override
5218 * the file offsets by those derived from the pragmas.
5220 struct pet_scop *PetScan::scan(FunctionDecl *fd)
5222 pet_scop *scop;
5223 Stmt *stmt;
5225 stmt = fd->getBody();
5227 if (options->autodetect)
5228 scop = extract(stmt, true);
5229 else {
5230 scop = scan(stmt);
5231 scop = pet_scop_update_start_end(scop, loc.start, loc.end);
5233 scop = pet_scop_detect_parameter_accesses(scop);
5234 scop = scan_arrays(scop);
5235 scop = add_parameter_bounds(scop);
5236 scop = pet_scop_gist(scop, value_bounds);
5238 return scop;