scan.cc: add missing include
[pet.git] / scan.cc
blob21d6ffe3998dfeddae1d4a2c3242f242700ebca2
1 /*
2 * Copyright 2011 Leiden University. All rights reserved.
3 * Copyright 2012 Ecole Normale Superieure. All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 *
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above
13 * copyright notice, this list of conditions and the following
14 * disclaimer in the documentation and/or other materials provided
15 * with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY LEIDEN UNIVERSITY ''AS IS'' AND ANY
18 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL LEIDEN UNIVERSITY OR
21 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
22 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
23 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA,
24 * OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
27 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 * The views and conclusions contained in the software and documentation
30 * are those of the authors and should not be interpreted as
31 * representing official policies, either expressed or implied, of
32 * Leiden University.
33 */
35 #include <set>
36 #include <map>
37 #include <iostream>
38 #include <clang/AST/ASTContext.h>
39 #include <clang/AST/ASTDiagnostic.h>
40 #include <clang/AST/Expr.h>
41 #include <clang/AST/RecursiveASTVisitor.h>
43 #include <isl/id.h>
44 #include <isl/space.h>
45 #include <isl/aff.h>
46 #include <isl/set.h>
48 #include "options.h"
49 #include "scan.h"
50 #include "scop.h"
51 #include "scop_plus.h"
53 #include "config.h"
55 using namespace std;
56 using namespace clang;
58 #if defined(DECLREFEXPR_CREATE_REQUIRES_BOOL)
59 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
61 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
62 SourceLocation(), var, false, var->getInnerLocStart(),
63 var->getType(), VK_LValue);
65 #elif defined(DECLREFEXPR_CREATE_REQUIRES_SOURCELOCATION)
66 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
68 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
69 SourceLocation(), var, var->getInnerLocStart(), var->getType(),
70 VK_LValue);
72 #else
73 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
75 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
76 var, var->getInnerLocStart(), var->getType(), VK_LValue);
78 #endif
80 /* Check if the element type corresponding to the given array type
81 * has a const qualifier.
83 static bool const_base(QualType qt)
85 const Type *type = qt.getTypePtr();
87 if (type->isPointerType())
88 return const_base(type->getPointeeType());
89 if (type->isArrayType()) {
90 const ArrayType *atype;
91 type = type->getCanonicalTypeInternal().getTypePtr();
92 atype = cast<ArrayType>(type);
93 return const_base(atype->getElementType());
96 return qt.isConstQualified();
99 /* Mark "decl" as having an unknown value in "assigned_value".
101 * If no (known or unknown) value was assigned to "decl" before,
102 * then it may have been treated as a parameter before and may
103 * therefore appear in a value assigned to another variable.
104 * If so, this assignment needs to be turned into an unknown value too.
106 static void clear_assignment(map<ValueDecl *, isl_pw_aff *> &assigned_value,
107 ValueDecl *decl)
109 map<ValueDecl *, isl_pw_aff *>::iterator it;
111 it = assigned_value.find(decl);
113 assigned_value[decl] = NULL;
115 if (it == assigned_value.end())
116 return;
118 for (it = assigned_value.begin(); it != assigned_value.end(); ++it) {
119 isl_pw_aff *pa = it->second;
120 int nparam = isl_pw_aff_dim(pa, isl_dim_param);
122 for (int i = 0; i < nparam; ++i) {
123 isl_id *id;
125 if (!isl_pw_aff_has_dim_id(pa, isl_dim_param, i))
126 continue;
127 id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
128 if (isl_id_get_user(id) == decl)
129 it->second = NULL;
130 isl_id_free(id);
135 /* Look for any assignments to scalar variables in part of the parse
136 * tree and set assigned_value to NULL for each of them.
137 * Also reset assigned_value if the address of a scalar variable
138 * is being taken. As an exception, if the address is passed to a function
139 * that is declared to receive a const pointer, then assigned_value is
140 * not reset.
142 * This ensures that we won't use any previously stored value
143 * in the current subtree and its parents.
145 struct clear_assignments : RecursiveASTVisitor<clear_assignments> {
146 map<ValueDecl *, isl_pw_aff *> &assigned_value;
147 set<UnaryOperator *> skip;
149 clear_assignments(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
150 assigned_value(assigned_value) {}
152 /* Check for "address of" operators whose value is passed
153 * to a const pointer argument and add them to "skip", so that
154 * we can skip them in VisitUnaryOperator.
156 bool VisitCallExpr(CallExpr *expr) {
157 FunctionDecl *fd;
158 fd = expr->getDirectCallee();
159 if (!fd)
160 return true;
161 for (int i = 0; i < expr->getNumArgs(); ++i) {
162 Expr *arg = expr->getArg(i);
163 UnaryOperator *op;
164 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
165 ImplicitCastExpr *ice;
166 ice = cast<ImplicitCastExpr>(arg);
167 arg = ice->getSubExpr();
169 if (arg->getStmtClass() != Stmt::UnaryOperatorClass)
170 continue;
171 op = cast<UnaryOperator>(arg);
172 if (op->getOpcode() != UO_AddrOf)
173 continue;
174 if (const_base(fd->getParamDecl(i)->getType()))
175 skip.insert(op);
177 return true;
180 bool VisitUnaryOperator(UnaryOperator *expr) {
181 Expr *arg;
182 DeclRefExpr *ref;
183 ValueDecl *decl;
185 if (expr->getOpcode() != UO_AddrOf)
186 return true;
187 if (skip.find(expr) != skip.end())
188 return true;
190 arg = expr->getSubExpr();
191 if (arg->getStmtClass() != Stmt::DeclRefExprClass)
192 return true;
193 ref = cast<DeclRefExpr>(arg);
194 decl = ref->getDecl();
195 clear_assignment(assigned_value, decl);
196 return true;
199 bool VisitBinaryOperator(BinaryOperator *expr) {
200 Expr *lhs;
201 DeclRefExpr *ref;
202 ValueDecl *decl;
204 if (!expr->isAssignmentOp())
205 return true;
206 lhs = expr->getLHS();
207 if (lhs->getStmtClass() != Stmt::DeclRefExprClass)
208 return true;
209 ref = cast<DeclRefExpr>(lhs);
210 decl = ref->getDecl();
211 clear_assignment(assigned_value, decl);
212 return true;
216 /* Keep a copy of the currently assigned values.
218 * Any variable that is assigned a value inside the current scope
219 * is removed again when we leave the scope (either because it wasn't
220 * stored in the cache or because it has a different value in the cache).
222 struct assigned_value_cache {
223 map<ValueDecl *, isl_pw_aff *> &assigned_value;
224 map<ValueDecl *, isl_pw_aff *> cache;
226 assigned_value_cache(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
227 assigned_value(assigned_value), cache(assigned_value) {}
228 ~assigned_value_cache() {
229 map<ValueDecl *, isl_pw_aff *>::iterator it = cache.begin();
230 for (it = assigned_value.begin(); it != assigned_value.end();
231 ++it) {
232 if (!it->second ||
233 (cache.find(it->first) != cache.end() &&
234 cache[it->first] != it->second))
235 cache[it->first] = NULL;
237 assigned_value = cache;
241 /* Insert an expression into the collection of expressions,
242 * provided it is not already in there.
243 * The isl_pw_affs are freed in the destructor.
245 void PetScan::insert_expression(__isl_take isl_pw_aff *expr)
247 std::set<isl_pw_aff *>::iterator it;
249 if (expressions.find(expr) == expressions.end())
250 expressions.insert(expr);
251 else
252 isl_pw_aff_free(expr);
255 PetScan::~PetScan()
257 std::set<isl_pw_aff *>::iterator it;
259 for (it = expressions.begin(); it != expressions.end(); ++it)
260 isl_pw_aff_free(*it);
262 isl_union_map_free(value_bounds);
265 /* Called if we found something we (currently) cannot handle.
266 * We'll provide more informative warnings later.
268 * We only actually complain if autodetect is false.
270 void PetScan::unsupported(Stmt *stmt, const char *msg)
272 if (options->autodetect)
273 return;
275 SourceLocation loc = stmt->getLocStart();
276 DiagnosticsEngine &diag = PP.getDiagnostics();
277 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
278 msg ? msg : "unsupported");
279 DiagnosticBuilder B = diag.Report(loc, id) << stmt->getSourceRange();
282 /* Extract an integer from "expr" and store it in "v".
284 int PetScan::extract_int(IntegerLiteral *expr, isl_int *v)
286 const Type *type = expr->getType().getTypePtr();
287 int is_signed = type->hasSignedIntegerRepresentation();
289 if (is_signed) {
290 int64_t i = expr->getValue().getSExtValue();
291 isl_int_set_si(*v, i);
292 } else {
293 uint64_t i = expr->getValue().getZExtValue();
294 isl_int_set_ui(*v, i);
297 return 0;
300 /* Extract an integer from "expr" and store it in "v".
301 * Return -1 if "expr" does not (obviously) represent an integer.
303 int PetScan::extract_int(clang::ParenExpr *expr, isl_int *v)
305 return extract_int(expr->getSubExpr(), v);
308 /* Extract an integer from "expr" and store it in "v".
309 * Return -1 if "expr" does not (obviously) represent an integer.
311 int PetScan::extract_int(clang::Expr *expr, isl_int *v)
313 if (expr->getStmtClass() == Stmt::IntegerLiteralClass)
314 return extract_int(cast<IntegerLiteral>(expr), v);
315 if (expr->getStmtClass() == Stmt::ParenExprClass)
316 return extract_int(cast<ParenExpr>(expr), v);
318 unsupported(expr);
319 return -1;
322 /* Extract an affine expression from the IntegerLiteral "expr".
324 __isl_give isl_pw_aff *PetScan::extract_affine(IntegerLiteral *expr)
326 isl_space *dim = isl_space_params_alloc(ctx, 0);
327 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
328 isl_aff *aff = isl_aff_zero_on_domain(ls);
329 isl_set *dom = isl_set_universe(dim);
330 isl_int v;
332 isl_int_init(v);
333 extract_int(expr, &v);
334 aff = isl_aff_add_constant(aff, v);
335 isl_int_clear(v);
337 return isl_pw_aff_alloc(dom, aff);
340 /* Extract an affine expression from the APInt "val".
342 __isl_give isl_pw_aff *PetScan::extract_affine(const llvm::APInt &val)
344 isl_space *dim = isl_space_params_alloc(ctx, 0);
345 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
346 isl_aff *aff = isl_aff_zero_on_domain(ls);
347 isl_set *dom = isl_set_universe(dim);
348 isl_int v;
350 isl_int_init(v);
351 isl_int_set_ui(v, val.getZExtValue());
352 aff = isl_aff_add_constant(aff, v);
353 isl_int_clear(v);
355 return isl_pw_aff_alloc(dom, aff);
358 __isl_give isl_pw_aff *PetScan::extract_affine(ImplicitCastExpr *expr)
360 return extract_affine(expr->getSubExpr());
363 static unsigned get_type_size(ValueDecl *decl)
365 return decl->getASTContext().getIntWidth(decl->getType());
368 /* Bound parameter "pos" of "set" to the possible values of "decl".
370 static __isl_give isl_set *set_parameter_bounds(__isl_take isl_set *set,
371 unsigned pos, ValueDecl *decl)
373 unsigned width;
374 isl_int v;
376 isl_int_init(v);
378 width = get_type_size(decl);
379 if (decl->getType()->isUnsignedIntegerType()) {
380 set = isl_set_lower_bound_si(set, isl_dim_param, pos, 0);
381 isl_int_set_si(v, 1);
382 isl_int_mul_2exp(v, v, width);
383 isl_int_sub_ui(v, v, 1);
384 set = isl_set_upper_bound(set, isl_dim_param, pos, v);
385 } else {
386 isl_int_set_si(v, 1);
387 isl_int_mul_2exp(v, v, width - 1);
388 isl_int_sub_ui(v, v, 1);
389 set = isl_set_upper_bound(set, isl_dim_param, pos, v);
390 isl_int_neg(v, v);
391 isl_int_sub_ui(v, v, 1);
392 set = isl_set_lower_bound(set, isl_dim_param, pos, v);
395 isl_int_clear(v);
397 return set;
400 /* Extract an affine expression from the DeclRefExpr "expr".
402 * If the variable has been assigned a value, then we check whether
403 * we know what (affine) value was assigned.
404 * If so, we return this value. Otherwise we convert "expr"
405 * to an extra parameter (provided nesting_enabled is set).
407 * Otherwise, we simply return an expression that is equal
408 * to a parameter corresponding to the referenced variable.
410 __isl_give isl_pw_aff *PetScan::extract_affine(DeclRefExpr *expr)
412 ValueDecl *decl = expr->getDecl();
413 const Type *type = decl->getType().getTypePtr();
414 isl_id *id;
415 isl_space *dim;
416 isl_aff *aff;
417 isl_set *dom;
419 if (!type->isIntegerType()) {
420 unsupported(expr);
421 return NULL;
424 if (assigned_value.find(decl) != assigned_value.end()) {
425 if (assigned_value[decl])
426 return isl_pw_aff_copy(assigned_value[decl]);
427 else
428 return nested_access(expr);
431 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
432 dim = isl_space_params_alloc(ctx, 1);
434 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
436 dom = isl_set_universe(isl_space_copy(dim));
437 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
438 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
440 return isl_pw_aff_alloc(dom, aff);
443 /* Extract an affine expression from an integer division operation.
444 * In particular, if "expr" is lhs/rhs, then return
446 * lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs)
448 * The second argument (rhs) is required to be a (positive) integer constant.
450 __isl_give isl_pw_aff *PetScan::extract_affine_div(BinaryOperator *expr)
452 Expr *rhs_expr;
453 isl_pw_aff *lhs, *lhs_f, *lhs_c;
454 isl_pw_aff *res;
455 isl_int v;
456 isl_set *cond;
458 rhs_expr = expr->getRHS();
459 isl_int_init(v);
460 if (extract_int(rhs_expr, &v) < 0) {
461 isl_int_clear(v);
462 return NULL;
465 lhs = extract_affine(expr->getLHS());
466 cond = isl_pw_aff_nonneg_set(isl_pw_aff_copy(lhs));
468 lhs = isl_pw_aff_scale_down(lhs, v);
469 isl_int_clear(v);
471 lhs_f = isl_pw_aff_floor(isl_pw_aff_copy(lhs));
472 lhs_c = isl_pw_aff_ceil(lhs);
473 res = isl_pw_aff_cond(isl_set_indicator_function(cond), lhs_f, lhs_c);
475 return res;
478 /* Extract an affine expression from a modulo operation.
479 * In particular, if "expr" is lhs/rhs, then return
481 * lhs - rhs * (lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs))
483 * The second argument (rhs) is required to be a (positive) integer constant.
485 __isl_give isl_pw_aff *PetScan::extract_affine_mod(BinaryOperator *expr)
487 Expr *rhs_expr;
488 isl_pw_aff *lhs, *lhs_f, *lhs_c;
489 isl_pw_aff *res;
490 isl_int v;
491 isl_set *cond;
493 rhs_expr = expr->getRHS();
494 if (rhs_expr->getStmtClass() != Stmt::IntegerLiteralClass) {
495 unsupported(expr);
496 return NULL;
499 lhs = extract_affine(expr->getLHS());
500 cond = isl_pw_aff_nonneg_set(isl_pw_aff_copy(lhs));
502 isl_int_init(v);
503 extract_int(cast<IntegerLiteral>(rhs_expr), &v);
504 res = isl_pw_aff_scale_down(isl_pw_aff_copy(lhs), v);
506 lhs_f = isl_pw_aff_floor(isl_pw_aff_copy(res));
507 lhs_c = isl_pw_aff_ceil(res);
508 res = isl_pw_aff_cond(isl_set_indicator_function(cond), lhs_f, lhs_c);
510 res = isl_pw_aff_scale(res, v);
511 isl_int_clear(v);
513 res = isl_pw_aff_sub(lhs, res);
515 return res;
518 /* Extract an affine expression from a multiplication operation.
519 * This is only allowed if at least one of the two arguments
520 * is a (piecewise) constant.
522 __isl_give isl_pw_aff *PetScan::extract_affine_mul(BinaryOperator *expr)
524 isl_pw_aff *lhs;
525 isl_pw_aff *rhs;
527 lhs = extract_affine(expr->getLHS());
528 rhs = extract_affine(expr->getRHS());
530 if (!isl_pw_aff_is_cst(lhs) && !isl_pw_aff_is_cst(rhs)) {
531 isl_pw_aff_free(lhs);
532 isl_pw_aff_free(rhs);
533 unsupported(expr);
534 return NULL;
537 return isl_pw_aff_mul(lhs, rhs);
540 /* Extract an affine expression from an addition or subtraction operation.
542 __isl_give isl_pw_aff *PetScan::extract_affine_add(BinaryOperator *expr)
544 isl_pw_aff *lhs;
545 isl_pw_aff *rhs;
547 lhs = extract_affine(expr->getLHS());
548 rhs = extract_affine(expr->getRHS());
550 switch (expr->getOpcode()) {
551 case BO_Add:
552 return isl_pw_aff_add(lhs, rhs);
553 case BO_Sub:
554 return isl_pw_aff_sub(lhs, rhs);
555 default:
556 isl_pw_aff_free(lhs);
557 isl_pw_aff_free(rhs);
558 return NULL;
563 /* Compute
565 * pwaff mod 2^width
567 static __isl_give isl_pw_aff *wrap(__isl_take isl_pw_aff *pwaff,
568 unsigned width)
570 isl_int mod;
572 isl_int_init(mod);
573 isl_int_set_si(mod, 1);
574 isl_int_mul_2exp(mod, mod, width);
576 pwaff = isl_pw_aff_mod(pwaff, mod);
578 isl_int_clear(mod);
580 return pwaff;
583 /* Limit the domain of "pwaff" to those elements where the function
584 * value satisfies
586 * 2^{width-1} <= pwaff < 2^{width-1}
588 static __isl_give isl_pw_aff *avoid_overflow(__isl_take isl_pw_aff *pwaff,
589 unsigned width)
591 isl_int v;
592 isl_space *space = isl_pw_aff_get_domain_space(pwaff);
593 isl_local_space *ls = isl_local_space_from_space(space);
594 isl_aff *bound;
595 isl_set *dom;
596 isl_pw_aff *b;
598 isl_int_init(v);
599 isl_int_set_si(v, 1);
600 isl_int_mul_2exp(v, v, width - 1);
602 bound = isl_aff_zero_on_domain(ls);
603 bound = isl_aff_add_constant(bound, v);
604 b = isl_pw_aff_from_aff(bound);
606 dom = isl_pw_aff_lt_set(isl_pw_aff_copy(pwaff), isl_pw_aff_copy(b));
607 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
609 b = isl_pw_aff_neg(b);
610 dom = isl_pw_aff_ge_set(isl_pw_aff_copy(pwaff), b);
611 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
613 isl_int_clear(v);
615 return pwaff;
618 /* Return the piecewise affine expression "set ? 1 : 0" defined on "dom".
620 static __isl_give isl_pw_aff *indicator_function(__isl_take isl_set *set,
621 __isl_take isl_set *dom)
623 isl_pw_aff *pa;
624 pa = isl_set_indicator_function(set);
625 pa = isl_pw_aff_intersect_domain(pa, dom);
626 return pa;
629 /* Extract an affine expression from some binary operations.
630 * If the result of the expression is unsigned, then we wrap it
631 * based on the size of the type. Otherwise, we ensure that
632 * no overflow occurs.
634 __isl_give isl_pw_aff *PetScan::extract_affine(BinaryOperator *expr)
636 isl_pw_aff *res;
637 unsigned width;
639 switch (expr->getOpcode()) {
640 case BO_Add:
641 case BO_Sub:
642 res = extract_affine_add(expr);
643 break;
644 case BO_Div:
645 res = extract_affine_div(expr);
646 break;
647 case BO_Rem:
648 res = extract_affine_mod(expr);
649 break;
650 case BO_Mul:
651 res = extract_affine_mul(expr);
652 break;
653 case BO_LT:
654 case BO_LE:
655 case BO_GT:
656 case BO_GE:
657 case BO_EQ:
658 case BO_NE:
659 case BO_LAnd:
660 case BO_LOr:
661 return extract_condition(expr);
662 default:
663 unsupported(expr);
664 return NULL;
667 width = ast_context.getIntWidth(expr->getType());
668 if (expr->getType()->isUnsignedIntegerType())
669 res = wrap(res, width);
670 else
671 res = avoid_overflow(res, width);
673 return res;
676 /* Extract an affine expression from a negation operation.
678 __isl_give isl_pw_aff *PetScan::extract_affine(UnaryOperator *expr)
680 if (expr->getOpcode() == UO_Minus)
681 return isl_pw_aff_neg(extract_affine(expr->getSubExpr()));
682 if (expr->getOpcode() == UO_LNot)
683 return extract_condition(expr);
685 unsupported(expr);
686 return NULL;
689 __isl_give isl_pw_aff *PetScan::extract_affine(ParenExpr *expr)
691 return extract_affine(expr->getSubExpr());
694 /* Extract an affine expression from some special function calls.
695 * In particular, we handle "min", "max", "ceild" and "floord".
696 * In case of the latter two, the second argument needs to be
697 * a (positive) integer constant.
699 __isl_give isl_pw_aff *PetScan::extract_affine(CallExpr *expr)
701 FunctionDecl *fd;
702 string name;
703 isl_pw_aff *aff1, *aff2;
705 fd = expr->getDirectCallee();
706 if (!fd) {
707 unsupported(expr);
708 return NULL;
711 name = fd->getDeclName().getAsString();
712 if (!(expr->getNumArgs() == 2 && name == "min") &&
713 !(expr->getNumArgs() == 2 && name == "max") &&
714 !(expr->getNumArgs() == 2 && name == "floord") &&
715 !(expr->getNumArgs() == 2 && name == "ceild")) {
716 unsupported(expr);
717 return NULL;
720 if (name == "min" || name == "max") {
721 aff1 = extract_affine(expr->getArg(0));
722 aff2 = extract_affine(expr->getArg(1));
724 if (name == "min")
725 aff1 = isl_pw_aff_min(aff1, aff2);
726 else
727 aff1 = isl_pw_aff_max(aff1, aff2);
728 } else if (name == "floord" || name == "ceild") {
729 isl_int v;
730 Expr *arg2 = expr->getArg(1);
732 if (arg2->getStmtClass() != Stmt::IntegerLiteralClass) {
733 unsupported(expr);
734 return NULL;
736 aff1 = extract_affine(expr->getArg(0));
737 isl_int_init(v);
738 extract_int(cast<IntegerLiteral>(arg2), &v);
739 aff1 = isl_pw_aff_scale_down(aff1, v);
740 isl_int_clear(v);
741 if (name == "floord")
742 aff1 = isl_pw_aff_floor(aff1);
743 else
744 aff1 = isl_pw_aff_ceil(aff1);
745 } else {
746 unsupported(expr);
747 return NULL;
750 return aff1;
753 /* This method is called when we come across an access that is
754 * nested in what is supposed to be an affine expression.
755 * If nesting is allowed, we return a new parameter that corresponds
756 * to this nested access. Otherwise, we simply complain.
758 * Note that we currently don't allow nested accesses themselves
759 * to contain any nested accesses, so we check if we can extract
760 * the access without any nesting and complain if we can't.
762 * The new parameter is resolved in resolve_nested.
764 isl_pw_aff *PetScan::nested_access(Expr *expr)
766 isl_id *id;
767 isl_space *dim;
768 isl_aff *aff;
769 isl_set *dom;
770 isl_map *access;
772 if (!nesting_enabled) {
773 unsupported(expr);
774 return NULL;
777 allow_nested = false;
778 access = extract_access(expr);
779 allow_nested = true;
780 if (!access) {
781 unsupported(expr);
782 return NULL;
784 isl_map_free(access);
786 id = isl_id_alloc(ctx, NULL, expr);
787 dim = isl_space_params_alloc(ctx, 1);
789 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
791 dom = isl_set_universe(isl_space_copy(dim));
792 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
793 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
795 return isl_pw_aff_alloc(dom, aff);
798 /* Affine expressions are not supposed to contain array accesses,
799 * but if nesting is allowed, we return a parameter corresponding
800 * to the array access.
802 __isl_give isl_pw_aff *PetScan::extract_affine(ArraySubscriptExpr *expr)
804 return nested_access(expr);
807 /* Extract an affine expression from a conditional operation.
809 __isl_give isl_pw_aff *PetScan::extract_affine(ConditionalOperator *expr)
811 isl_pw_aff *cond, *lhs, *rhs, *res;
813 cond = extract_condition(expr->getCond());
814 lhs = extract_affine(expr->getTrueExpr());
815 rhs = extract_affine(expr->getFalseExpr());
817 return isl_pw_aff_cond(cond, lhs, rhs);
820 /* Extract an affine expression, if possible, from "expr".
821 * Otherwise return NULL.
823 __isl_give isl_pw_aff *PetScan::extract_affine(Expr *expr)
825 switch (expr->getStmtClass()) {
826 case Stmt::ImplicitCastExprClass:
827 return extract_affine(cast<ImplicitCastExpr>(expr));
828 case Stmt::IntegerLiteralClass:
829 return extract_affine(cast<IntegerLiteral>(expr));
830 case Stmt::DeclRefExprClass:
831 return extract_affine(cast<DeclRefExpr>(expr));
832 case Stmt::BinaryOperatorClass:
833 return extract_affine(cast<BinaryOperator>(expr));
834 case Stmt::UnaryOperatorClass:
835 return extract_affine(cast<UnaryOperator>(expr));
836 case Stmt::ParenExprClass:
837 return extract_affine(cast<ParenExpr>(expr));
838 case Stmt::CallExprClass:
839 return extract_affine(cast<CallExpr>(expr));
840 case Stmt::ArraySubscriptExprClass:
841 return extract_affine(cast<ArraySubscriptExpr>(expr));
842 case Stmt::ConditionalOperatorClass:
843 return extract_affine(cast<ConditionalOperator>(expr));
844 default:
845 unsupported(expr);
847 return NULL;
850 __isl_give isl_map *PetScan::extract_access(ImplicitCastExpr *expr)
852 return extract_access(expr->getSubExpr());
855 /* Return the depth of an array of the given type.
857 static int array_depth(const Type *type)
859 if (type->isPointerType())
860 return 1 + array_depth(type->getPointeeType().getTypePtr());
861 if (type->isArrayType()) {
862 const ArrayType *atype;
863 type = type->getCanonicalTypeInternal().getTypePtr();
864 atype = cast<ArrayType>(type);
865 return 1 + array_depth(atype->getElementType().getTypePtr());
867 return 0;
870 /* Return the element type of the given array type.
872 static QualType base_type(QualType qt)
874 const Type *type = qt.getTypePtr();
876 if (type->isPointerType())
877 return base_type(type->getPointeeType());
878 if (type->isArrayType()) {
879 const ArrayType *atype;
880 type = type->getCanonicalTypeInternal().getTypePtr();
881 atype = cast<ArrayType>(type);
882 return base_type(atype->getElementType());
884 return qt;
887 /* Extract an access relation from a reference to a variable.
888 * If the variable has name "A" and its type corresponds to an
889 * array of depth d, then the returned access relation is of the
890 * form
892 * { [] -> A[i_1,...,i_d] }
894 __isl_give isl_map *PetScan::extract_access(DeclRefExpr *expr)
896 ValueDecl *decl = expr->getDecl();
897 int depth = array_depth(decl->getType().getTypePtr());
898 isl_id *id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
899 isl_space *dim = isl_space_alloc(ctx, 0, 0, depth);
900 isl_map *access_rel;
902 dim = isl_space_set_tuple_id(dim, isl_dim_out, id);
904 access_rel = isl_map_universe(dim);
906 return access_rel;
909 /* Extract an access relation from an integer contant.
910 * If the value of the constant is "v", then the returned access relation
911 * is
913 * { [] -> [v] }
915 __isl_give isl_map *PetScan::extract_access(IntegerLiteral *expr)
917 return isl_map_from_range(isl_set_from_pw_aff(extract_affine(expr)));
920 /* Try and extract an access relation from the given Expr.
921 * Return NULL if it doesn't work out.
923 __isl_give isl_map *PetScan::extract_access(Expr *expr)
925 switch (expr->getStmtClass()) {
926 case Stmt::ImplicitCastExprClass:
927 return extract_access(cast<ImplicitCastExpr>(expr));
928 case Stmt::DeclRefExprClass:
929 return extract_access(cast<DeclRefExpr>(expr));
930 case Stmt::ArraySubscriptExprClass:
931 return extract_access(cast<ArraySubscriptExpr>(expr));
932 case Stmt::IntegerLiteralClass:
933 return extract_access(cast<IntegerLiteral>(expr));
934 default:
935 unsupported(expr);
937 return NULL;
940 /* Assign the affine expression "index" to the output dimension "pos" of "map",
941 * restrict the domain to those values that result in a non-negative index
942 * and return the result.
944 __isl_give isl_map *set_index(__isl_take isl_map *map, int pos,
945 __isl_take isl_pw_aff *index)
947 isl_map *index_map;
948 int len = isl_map_dim(map, isl_dim_out);
949 isl_id *id;
950 isl_set *domain;
952 domain = isl_pw_aff_nonneg_set(isl_pw_aff_copy(index));
953 index = isl_pw_aff_intersect_domain(index, domain);
954 index_map = isl_map_from_range(isl_set_from_pw_aff(index));
955 index_map = isl_map_insert_dims(index_map, isl_dim_out, 0, pos);
956 index_map = isl_map_add_dims(index_map, isl_dim_out, len - pos - 1);
957 id = isl_map_get_tuple_id(map, isl_dim_out);
958 index_map = isl_map_set_tuple_id(index_map, isl_dim_out, id);
960 map = isl_map_intersect(map, index_map);
962 return map;
965 /* Extract an access relation from the given array subscript expression.
966 * If nesting is allowed in general, then we turn it on while
967 * examining the index expression.
969 * We first extract an access relation from the base.
970 * This will result in an access relation with a range that corresponds
971 * to the array being accessed and with earlier indices filled in already.
972 * We then extract the current index and fill that in as well.
973 * The position of the current index is based on the type of base.
974 * If base is the actual array variable, then the depth of this type
975 * will be the same as the depth of the array and we will fill in
976 * the first array index.
977 * Otherwise, the depth of the base type will be smaller and we will fill
978 * in a later index.
980 __isl_give isl_map *PetScan::extract_access(ArraySubscriptExpr *expr)
982 Expr *base = expr->getBase();
983 Expr *idx = expr->getIdx();
984 isl_pw_aff *index;
985 isl_map *base_access;
986 isl_map *access;
987 int depth = array_depth(base->getType().getTypePtr());
988 int pos;
989 bool save_nesting = nesting_enabled;
991 nesting_enabled = allow_nested;
993 base_access = extract_access(base);
994 index = extract_affine(idx);
996 nesting_enabled = save_nesting;
998 pos = isl_map_dim(base_access, isl_dim_out) - depth;
999 access = set_index(base_access, pos, index);
1001 return access;
1004 /* Check if "expr" calls function "minmax" with two arguments and if so
1005 * make lhs and rhs refer to these two arguments.
1007 static bool is_minmax(Expr *expr, const char *minmax, Expr *&lhs, Expr *&rhs)
1009 CallExpr *call;
1010 FunctionDecl *fd;
1011 string name;
1013 if (expr->getStmtClass() != Stmt::CallExprClass)
1014 return false;
1016 call = cast<CallExpr>(expr);
1017 fd = call->getDirectCallee();
1018 if (!fd)
1019 return false;
1021 if (call->getNumArgs() != 2)
1022 return false;
1024 name = fd->getDeclName().getAsString();
1025 if (name != minmax)
1026 return false;
1028 lhs = call->getArg(0);
1029 rhs = call->getArg(1);
1031 return true;
1034 /* Check if "expr" is of the form min(lhs, rhs) and if so make
1035 * lhs and rhs refer to the two arguments.
1037 static bool is_min(Expr *expr, Expr *&lhs, Expr *&rhs)
1039 return is_minmax(expr, "min", lhs, rhs);
1042 /* Check if "expr" is of the form max(lhs, rhs) and if so make
1043 * lhs and rhs refer to the two arguments.
1045 static bool is_max(Expr *expr, Expr *&lhs, Expr *&rhs)
1047 return is_minmax(expr, "max", lhs, rhs);
1050 /* Return "lhs && rhs", defined on the shared definition domain.
1052 static __isl_give isl_pw_aff *pw_aff_and(__isl_take isl_pw_aff *lhs,
1053 __isl_take isl_pw_aff *rhs)
1055 isl_set *cond;
1056 isl_set *dom;
1058 dom = isl_set_intersect(isl_pw_aff_domain(isl_pw_aff_copy(lhs)),
1059 isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1060 cond = isl_set_intersect(isl_pw_aff_non_zero_set(lhs),
1061 isl_pw_aff_non_zero_set(rhs));
1062 return indicator_function(cond, dom);
1065 /* Return "lhs && rhs", with shortcut semantics.
1066 * That is, if lhs is false, then the result is defined even if rhs is not.
1067 * In practice, we compute lhs ? rhs : lhs.
1069 static __isl_give isl_pw_aff *pw_aff_and_then(__isl_take isl_pw_aff *lhs,
1070 __isl_take isl_pw_aff *rhs)
1072 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), rhs, lhs);
1075 /* Return "lhs || rhs", with shortcut semantics.
1076 * That is, if lhs is true, then the result is defined even if rhs is not.
1077 * In practice, we compute lhs ? lhs : rhs.
1079 static __isl_give isl_pw_aff *pw_aff_or_else(__isl_take isl_pw_aff *lhs,
1080 __isl_take isl_pw_aff *rhs)
1082 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), lhs, rhs);
1085 /* Extract an affine expressions representing the comparison "LHS op RHS"
1086 * "comp" is the original statement that "LHS op RHS" is derived from
1087 * and is used for diagnostics.
1089 * If the comparison is of the form
1091 * a <= min(b,c)
1093 * then the expression is constructed as the conjunction of
1094 * the comparisons
1096 * a <= b and a <= c
1098 * A similar optimization is performed for max(a,b) <= c.
1099 * We do this because that will lead to simpler representations
1100 * of the expression.
1101 * If isl is ever enhanced to explicitly deal with min and max expressions,
1102 * this optimization can be removed.
1104 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperatorKind op,
1105 Expr *LHS, Expr *RHS, Stmt *comp)
1107 isl_pw_aff *lhs;
1108 isl_pw_aff *rhs;
1109 isl_pw_aff *res;
1110 isl_set *cond;
1111 isl_set *dom;
1113 if (op == BO_GT)
1114 return extract_comparison(BO_LT, RHS, LHS, comp);
1115 if (op == BO_GE)
1116 return extract_comparison(BO_LE, RHS, LHS, comp);
1118 if (op == BO_LT || op == BO_LE) {
1119 Expr *expr1, *expr2;
1120 if (is_min(RHS, expr1, expr2)) {
1121 lhs = extract_comparison(op, LHS, expr1, comp);
1122 rhs = extract_comparison(op, LHS, expr2, comp);
1123 return pw_aff_and(lhs, rhs);
1125 if (is_max(LHS, expr1, expr2)) {
1126 lhs = extract_comparison(op, expr1, RHS, comp);
1127 rhs = extract_comparison(op, expr2, RHS, comp);
1128 return pw_aff_and(lhs, rhs);
1132 lhs = extract_affine(LHS);
1133 rhs = extract_affine(RHS);
1135 dom = isl_pw_aff_domain(isl_pw_aff_copy(lhs));
1136 dom = isl_set_intersect(dom, isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1138 switch (op) {
1139 case BO_LT:
1140 cond = isl_pw_aff_lt_set(lhs, rhs);
1141 break;
1142 case BO_LE:
1143 cond = isl_pw_aff_le_set(lhs, rhs);
1144 break;
1145 case BO_EQ:
1146 cond = isl_pw_aff_eq_set(lhs, rhs);
1147 break;
1148 case BO_NE:
1149 cond = isl_pw_aff_ne_set(lhs, rhs);
1150 break;
1151 default:
1152 isl_pw_aff_free(lhs);
1153 isl_pw_aff_free(rhs);
1154 isl_set_free(dom);
1155 unsupported(comp);
1156 return NULL;
1159 cond = isl_set_coalesce(cond);
1160 res = indicator_function(cond, dom);
1162 return res;
1165 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperator *comp)
1167 return extract_comparison(comp->getOpcode(), comp->getLHS(),
1168 comp->getRHS(), comp);
1171 /* Extract an affine expression representing the negation (logical not)
1172 * of a subexpression.
1174 __isl_give isl_pw_aff *PetScan::extract_boolean(UnaryOperator *op)
1176 isl_set *set_cond, *dom;
1177 isl_pw_aff *cond, *res;
1179 cond = extract_condition(op->getSubExpr());
1181 dom = isl_pw_aff_domain(isl_pw_aff_copy(cond));
1183 set_cond = isl_pw_aff_zero_set(cond);
1185 res = indicator_function(set_cond, dom);
1187 return res;
1190 /* Extract an affine expression representing the disjunction (logical or)
1191 * or conjunction (logical and) of two subexpressions.
1193 __isl_give isl_pw_aff *PetScan::extract_boolean(BinaryOperator *comp)
1195 isl_pw_aff *lhs, *rhs;
1197 lhs = extract_condition(comp->getLHS());
1198 rhs = extract_condition(comp->getRHS());
1200 switch (comp->getOpcode()) {
1201 case BO_LAnd:
1202 return pw_aff_and_then(lhs, rhs);
1203 case BO_LOr:
1204 return pw_aff_or_else(lhs, rhs);
1205 default:
1206 isl_pw_aff_free(lhs);
1207 isl_pw_aff_free(rhs);
1210 unsupported(comp);
1211 return NULL;
1214 __isl_give isl_pw_aff *PetScan::extract_condition(UnaryOperator *expr)
1216 switch (expr->getOpcode()) {
1217 case UO_LNot:
1218 return extract_boolean(expr);
1219 default:
1220 unsupported(expr);
1221 return NULL;
1225 /* Extract the affine expression "expr != 0 ? 1 : 0".
1227 __isl_give isl_pw_aff *PetScan::extract_implicit_condition(Expr *expr)
1229 isl_pw_aff *res;
1230 isl_set *set, *dom;
1232 res = extract_affine(expr);
1234 dom = isl_pw_aff_domain(isl_pw_aff_copy(res));
1235 set = isl_pw_aff_non_zero_set(res);
1237 res = indicator_function(set, dom);
1239 return res;
1242 /* Extract an affine expression from a boolean expression.
1243 * In particular, return the expression "expr ? 1 : 0".
1245 * If the expression doesn't look like a condition, we assume it
1246 * is an affine expression and return the condition "expr != 0 ? 1 : 0".
1248 __isl_give isl_pw_aff *PetScan::extract_condition(Expr *expr)
1250 BinaryOperator *comp;
1252 if (!expr) {
1253 isl_set *u = isl_set_universe(isl_space_params_alloc(ctx, 0));
1254 return indicator_function(u, isl_set_copy(u));
1257 if (expr->getStmtClass() == Stmt::ParenExprClass)
1258 return extract_condition(cast<ParenExpr>(expr)->getSubExpr());
1260 if (expr->getStmtClass() == Stmt::UnaryOperatorClass)
1261 return extract_condition(cast<UnaryOperator>(expr));
1263 if (expr->getStmtClass() != Stmt::BinaryOperatorClass)
1264 return extract_implicit_condition(expr);
1266 comp = cast<BinaryOperator>(expr);
1267 switch (comp->getOpcode()) {
1268 case BO_LT:
1269 case BO_LE:
1270 case BO_GT:
1271 case BO_GE:
1272 case BO_EQ:
1273 case BO_NE:
1274 return extract_comparison(comp);
1275 case BO_LAnd:
1276 case BO_LOr:
1277 return extract_boolean(comp);
1278 default:
1279 return extract_implicit_condition(expr);
1283 static enum pet_op_type UnaryOperatorKind2pet_op_type(UnaryOperatorKind kind)
1285 switch (kind) {
1286 case UO_Minus:
1287 return pet_op_minus;
1288 case UO_PostInc:
1289 return pet_op_post_inc;
1290 case UO_PostDec:
1291 return pet_op_post_dec;
1292 case UO_PreInc:
1293 return pet_op_pre_inc;
1294 case UO_PreDec:
1295 return pet_op_pre_dec;
1296 default:
1297 return pet_op_last;
1301 static enum pet_op_type BinaryOperatorKind2pet_op_type(BinaryOperatorKind kind)
1303 switch (kind) {
1304 case BO_AddAssign:
1305 return pet_op_add_assign;
1306 case BO_SubAssign:
1307 return pet_op_sub_assign;
1308 case BO_MulAssign:
1309 return pet_op_mul_assign;
1310 case BO_DivAssign:
1311 return pet_op_div_assign;
1312 case BO_Assign:
1313 return pet_op_assign;
1314 case BO_Add:
1315 return pet_op_add;
1316 case BO_Sub:
1317 return pet_op_sub;
1318 case BO_Mul:
1319 return pet_op_mul;
1320 case BO_Div:
1321 return pet_op_div;
1322 case BO_EQ:
1323 return pet_op_eq;
1324 case BO_LE:
1325 return pet_op_le;
1326 case BO_LT:
1327 return pet_op_lt;
1328 case BO_GT:
1329 return pet_op_gt;
1330 default:
1331 return pet_op_last;
1335 /* Construct a pet_expr representing a unary operator expression.
1337 struct pet_expr *PetScan::extract_expr(UnaryOperator *expr)
1339 struct pet_expr *arg;
1340 enum pet_op_type op;
1342 op = UnaryOperatorKind2pet_op_type(expr->getOpcode());
1343 if (op == pet_op_last) {
1344 unsupported(expr);
1345 return NULL;
1348 arg = extract_expr(expr->getSubExpr());
1350 if (expr->isIncrementDecrementOp() &&
1351 arg && arg->type == pet_expr_access) {
1352 mark_write(arg);
1353 arg->acc.read = 1;
1356 return pet_expr_new_unary(ctx, op, arg);
1359 /* Mark the given access pet_expr as a write.
1360 * If a scalar is being accessed, then mark its value
1361 * as unknown in assigned_value.
1363 void PetScan::mark_write(struct pet_expr *access)
1365 isl_id *id;
1366 ValueDecl *decl;
1368 access->acc.write = 1;
1369 access->acc.read = 0;
1371 if (isl_map_dim(access->acc.access, isl_dim_out) != 0)
1372 return;
1374 id = isl_map_get_tuple_id(access->acc.access, isl_dim_out);
1375 decl = (ValueDecl *) isl_id_get_user(id);
1376 clear_assignment(assigned_value, decl);
1377 isl_id_free(id);
1380 /* Construct a pet_expr representing a binary operator expression.
1382 * If the top level operator is an assignment and the LHS is an access,
1383 * then we mark that access as a write. If the operator is a compound
1384 * assignment, the access is marked as both a read and a write.
1386 * If "expr" assigns something to a scalar variable, then we mark
1387 * the variable as having been assigned. If, furthermore, the expression
1388 * is affine, then keep track of this value in assigned_value
1389 * so that we can plug it in when we later come across the same variable.
1391 struct pet_expr *PetScan::extract_expr(BinaryOperator *expr)
1393 struct pet_expr *lhs, *rhs;
1394 enum pet_op_type op;
1396 op = BinaryOperatorKind2pet_op_type(expr->getOpcode());
1397 if (op == pet_op_last) {
1398 unsupported(expr);
1399 return NULL;
1402 lhs = extract_expr(expr->getLHS());
1403 rhs = extract_expr(expr->getRHS());
1405 if (expr->isAssignmentOp() && lhs && lhs->type == pet_expr_access) {
1406 mark_write(lhs);
1407 if (expr->isCompoundAssignmentOp())
1408 lhs->acc.read = 1;
1411 if (expr->getOpcode() == BO_Assign &&
1412 lhs && lhs->type == pet_expr_access &&
1413 isl_map_dim(lhs->acc.access, isl_dim_out) == 0) {
1414 isl_id *id = isl_map_get_tuple_id(lhs->acc.access, isl_dim_out);
1415 ValueDecl *decl = (ValueDecl *) isl_id_get_user(id);
1416 Expr *rhs = expr->getRHS();
1417 isl_pw_aff *pa = try_extract_affine(rhs);
1418 clear_assignment(assigned_value, decl);
1419 if (pa) {
1420 assigned_value[decl] = pa;
1421 insert_expression(pa);
1423 isl_id_free(id);
1426 return pet_expr_new_binary(ctx, op, lhs, rhs);
1429 /* Construct a pet_expr representing a conditional operation.
1431 * We first try to extract the condition as an affine expression.
1432 * If that fails, we construct a pet_expr tree representing the condition.
1434 struct pet_expr *PetScan::extract_expr(ConditionalOperator *expr)
1436 struct pet_expr *cond, *lhs, *rhs;
1437 isl_pw_aff *pa;
1439 pa = try_extract_affine(expr->getCond());
1440 if (pa) {
1441 isl_set *test = isl_set_from_pw_aff(pa);
1442 cond = pet_expr_from_access(isl_map_from_range(test));
1443 } else
1444 cond = extract_expr(expr->getCond());
1445 lhs = extract_expr(expr->getTrueExpr());
1446 rhs = extract_expr(expr->getFalseExpr());
1448 return pet_expr_new_ternary(ctx, cond, lhs, rhs);
1451 struct pet_expr *PetScan::extract_expr(ImplicitCastExpr *expr)
1453 return extract_expr(expr->getSubExpr());
1456 /* Construct a pet_expr representing a floating point value.
1458 struct pet_expr *PetScan::extract_expr(FloatingLiteral *expr)
1460 return pet_expr_new_double(ctx, expr->getValueAsApproximateDouble());
1463 /* Extract an access relation from "expr" and then convert it into
1464 * a pet_expr.
1466 struct pet_expr *PetScan::extract_access_expr(Expr *expr)
1468 isl_map *access;
1469 struct pet_expr *pe;
1471 access = extract_access(expr);
1473 pe = pet_expr_from_access(access);
1475 return pe;
1478 struct pet_expr *PetScan::extract_expr(ParenExpr *expr)
1480 return extract_expr(expr->getSubExpr());
1483 /* Construct a pet_expr representing a function call.
1485 * If we are passing along a pointer to an array element
1486 * or an entire row or even higher dimensional slice of an array,
1487 * then the function being called may write into the array.
1489 * We assume here that if the function is declared to take a pointer
1490 * to a const type, then the function will perform a read
1491 * and that otherwise, it will perform a write.
1493 struct pet_expr *PetScan::extract_expr(CallExpr *expr)
1495 struct pet_expr *res = NULL;
1496 FunctionDecl *fd;
1497 string name;
1499 fd = expr->getDirectCallee();
1500 if (!fd) {
1501 unsupported(expr);
1502 return NULL;
1505 name = fd->getDeclName().getAsString();
1506 res = pet_expr_new_call(ctx, name.c_str(), expr->getNumArgs());
1507 if (!res)
1508 return NULL;
1510 for (int i = 0; i < expr->getNumArgs(); ++i) {
1511 Expr *arg = expr->getArg(i);
1512 int is_addr = 0;
1513 pet_expr *main_arg;
1515 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
1516 ImplicitCastExpr *ice = cast<ImplicitCastExpr>(arg);
1517 arg = ice->getSubExpr();
1519 if (arg->getStmtClass() == Stmt::UnaryOperatorClass) {
1520 UnaryOperator *op = cast<UnaryOperator>(arg);
1521 if (op->getOpcode() == UO_AddrOf) {
1522 is_addr = 1;
1523 arg = op->getSubExpr();
1526 res->args[i] = PetScan::extract_expr(arg);
1527 main_arg = res->args[i];
1528 if (is_addr)
1529 res->args[i] = pet_expr_new_unary(ctx,
1530 pet_op_address_of, res->args[i]);
1531 if (!res->args[i])
1532 goto error;
1533 if (arg->getStmtClass() == Stmt::ArraySubscriptExprClass &&
1534 array_depth(arg->getType().getTypePtr()) > 0)
1535 is_addr = 1;
1536 if (is_addr && main_arg->type == pet_expr_access) {
1537 ParmVarDecl *parm;
1538 if (!fd->hasPrototype()) {
1539 unsupported(expr, "prototype required");
1540 goto error;
1542 parm = fd->getParamDecl(i);
1543 if (!const_base(parm->getType()))
1544 mark_write(main_arg);
1548 return res;
1549 error:
1550 pet_expr_free(res);
1551 return NULL;
1554 /* Try and onstruct a pet_expr representing "expr".
1556 struct pet_expr *PetScan::extract_expr(Expr *expr)
1558 switch (expr->getStmtClass()) {
1559 case Stmt::UnaryOperatorClass:
1560 return extract_expr(cast<UnaryOperator>(expr));
1561 case Stmt::CompoundAssignOperatorClass:
1562 case Stmt::BinaryOperatorClass:
1563 return extract_expr(cast<BinaryOperator>(expr));
1564 case Stmt::ImplicitCastExprClass:
1565 return extract_expr(cast<ImplicitCastExpr>(expr));
1566 case Stmt::ArraySubscriptExprClass:
1567 case Stmt::DeclRefExprClass:
1568 case Stmt::IntegerLiteralClass:
1569 return extract_access_expr(expr);
1570 case Stmt::FloatingLiteralClass:
1571 return extract_expr(cast<FloatingLiteral>(expr));
1572 case Stmt::ParenExprClass:
1573 return extract_expr(cast<ParenExpr>(expr));
1574 case Stmt::ConditionalOperatorClass:
1575 return extract_expr(cast<ConditionalOperator>(expr));
1576 case Stmt::CallExprClass:
1577 return extract_expr(cast<CallExpr>(expr));
1578 default:
1579 unsupported(expr);
1581 return NULL;
1584 /* Check if the given initialization statement is an assignment.
1585 * If so, return that assignment. Otherwise return NULL.
1587 BinaryOperator *PetScan::initialization_assignment(Stmt *init)
1589 BinaryOperator *ass;
1591 if (init->getStmtClass() != Stmt::BinaryOperatorClass)
1592 return NULL;
1594 ass = cast<BinaryOperator>(init);
1595 if (ass->getOpcode() != BO_Assign)
1596 return NULL;
1598 return ass;
1601 /* Check if the given initialization statement is a declaration
1602 * of a single variable.
1603 * If so, return that declaration. Otherwise return NULL.
1605 Decl *PetScan::initialization_declaration(Stmt *init)
1607 DeclStmt *decl;
1609 if (init->getStmtClass() != Stmt::DeclStmtClass)
1610 return NULL;
1612 decl = cast<DeclStmt>(init);
1614 if (!decl->isSingleDecl())
1615 return NULL;
1617 return decl->getSingleDecl();
1620 /* Given the assignment operator in the initialization of a for loop,
1621 * extract the induction variable, i.e., the (integer)variable being
1622 * assigned.
1624 ValueDecl *PetScan::extract_induction_variable(BinaryOperator *init)
1626 Expr *lhs;
1627 DeclRefExpr *ref;
1628 ValueDecl *decl;
1629 const Type *type;
1631 lhs = init->getLHS();
1632 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1633 unsupported(init);
1634 return NULL;
1637 ref = cast<DeclRefExpr>(lhs);
1638 decl = ref->getDecl();
1639 type = decl->getType().getTypePtr();
1641 if (!type->isIntegerType()) {
1642 unsupported(lhs);
1643 return NULL;
1646 return decl;
1649 /* Given the initialization statement of a for loop and the single
1650 * declaration in this initialization statement,
1651 * extract the induction variable, i.e., the (integer) variable being
1652 * declared.
1654 VarDecl *PetScan::extract_induction_variable(Stmt *init, Decl *decl)
1656 VarDecl *vd;
1658 vd = cast<VarDecl>(decl);
1660 const QualType type = vd->getType();
1661 if (!type->isIntegerType()) {
1662 unsupported(init);
1663 return NULL;
1666 if (!vd->getInit()) {
1667 unsupported(init);
1668 return NULL;
1671 return vd;
1674 /* Check that op is of the form iv++ or iv--.
1675 * Return an affine expression "1" or "-1" accordingly.
1677 __isl_give isl_pw_aff *PetScan::extract_unary_increment(
1678 clang::UnaryOperator *op, clang::ValueDecl *iv)
1680 Expr *sub;
1681 DeclRefExpr *ref;
1682 isl_space *space;
1683 isl_aff *aff;
1685 if (!op->isIncrementDecrementOp()) {
1686 unsupported(op);
1687 return NULL;
1690 sub = op->getSubExpr();
1691 if (sub->getStmtClass() != Stmt::DeclRefExprClass) {
1692 unsupported(op);
1693 return NULL;
1696 ref = cast<DeclRefExpr>(sub);
1697 if (ref->getDecl() != iv) {
1698 unsupported(op);
1699 return NULL;
1702 space = isl_space_params_alloc(ctx, 0);
1703 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
1705 if (op->isIncrementOp())
1706 aff = isl_aff_add_constant_si(aff, 1);
1707 else
1708 aff = isl_aff_add_constant_si(aff, -1);
1710 return isl_pw_aff_from_aff(aff);
1713 /* If the isl_pw_aff on which isl_pw_aff_foreach_piece is called
1714 * has a single constant expression, then put this constant in *user.
1715 * The caller is assumed to have checked that this function will
1716 * be called exactly once.
1718 static int extract_cst(__isl_take isl_set *set, __isl_take isl_aff *aff,
1719 void *user)
1721 isl_int *inc = (isl_int *)user;
1722 int res = 0;
1724 if (isl_aff_is_cst(aff))
1725 isl_aff_get_constant(aff, inc);
1726 else
1727 res = -1;
1729 isl_set_free(set);
1730 isl_aff_free(aff);
1732 return res;
1735 /* Check if op is of the form
1737 * iv = iv + inc
1739 * and return inc as an affine expression.
1741 * We extract an affine expression from the RHS, subtract iv and return
1742 * the result.
1744 __isl_give isl_pw_aff *PetScan::extract_binary_increment(BinaryOperator *op,
1745 clang::ValueDecl *iv)
1747 Expr *lhs;
1748 DeclRefExpr *ref;
1749 isl_id *id;
1750 isl_space *dim;
1751 isl_aff *aff;
1752 isl_pw_aff *val;
1754 if (op->getOpcode() != BO_Assign) {
1755 unsupported(op);
1756 return NULL;
1759 lhs = op->getLHS();
1760 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1761 unsupported(op);
1762 return NULL;
1765 ref = cast<DeclRefExpr>(lhs);
1766 if (ref->getDecl() != iv) {
1767 unsupported(op);
1768 return NULL;
1771 val = extract_affine(op->getRHS());
1773 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
1775 dim = isl_space_params_alloc(ctx, 1);
1776 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
1777 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
1778 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
1780 val = isl_pw_aff_sub(val, isl_pw_aff_from_aff(aff));
1782 return val;
1785 /* Check that op is of the form iv += cst or iv -= cst
1786 * and return an affine expression corresponding oto cst or -cst accordingly.
1788 __isl_give isl_pw_aff *PetScan::extract_compound_increment(
1789 CompoundAssignOperator *op, clang::ValueDecl *iv)
1791 Expr *lhs;
1792 DeclRefExpr *ref;
1793 bool neg = false;
1794 isl_pw_aff *val;
1795 BinaryOperatorKind opcode;
1797 opcode = op->getOpcode();
1798 if (opcode != BO_AddAssign && opcode != BO_SubAssign) {
1799 unsupported(op);
1800 return NULL;
1802 if (opcode == BO_SubAssign)
1803 neg = true;
1805 lhs = op->getLHS();
1806 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1807 unsupported(op);
1808 return NULL;
1811 ref = cast<DeclRefExpr>(lhs);
1812 if (ref->getDecl() != iv) {
1813 unsupported(op);
1814 return NULL;
1817 val = extract_affine(op->getRHS());
1818 if (neg)
1819 val = isl_pw_aff_neg(val);
1821 return val;
1824 /* Check that the increment of the given for loop increments
1825 * (or decrements) the induction variable "iv" and return
1826 * the increment as an affine expression if successful.
1828 __isl_give isl_pw_aff *PetScan::extract_increment(clang::ForStmt *stmt,
1829 ValueDecl *iv)
1831 Stmt *inc = stmt->getInc();
1833 if (!inc) {
1834 unsupported(stmt);
1835 return NULL;
1838 if (inc->getStmtClass() == Stmt::UnaryOperatorClass)
1839 return extract_unary_increment(cast<UnaryOperator>(inc), iv);
1840 if (inc->getStmtClass() == Stmt::CompoundAssignOperatorClass)
1841 return extract_compound_increment(
1842 cast<CompoundAssignOperator>(inc), iv);
1843 if (inc->getStmtClass() == Stmt::BinaryOperatorClass)
1844 return extract_binary_increment(cast<BinaryOperator>(inc), iv);
1846 unsupported(inc);
1847 return NULL;
1850 /* Embed the given iteration domain in an extra outer loop
1851 * with induction variable "var".
1852 * If this variable appeared as a parameter in the constraints,
1853 * it is replaced by the new outermost dimension.
1855 static __isl_give isl_set *embed(__isl_take isl_set *set,
1856 __isl_take isl_id *var)
1858 int pos;
1860 set = isl_set_insert_dims(set, isl_dim_set, 0, 1);
1861 pos = isl_set_find_dim_by_id(set, isl_dim_param, var);
1862 if (pos >= 0) {
1863 set = isl_set_equate(set, isl_dim_param, pos, isl_dim_set, 0);
1864 set = isl_set_project_out(set, isl_dim_param, pos, 1);
1867 isl_id_free(var);
1868 return set;
1871 /* Return those elements in the space of "cond" that come after
1872 * (based on "sign") an element in "cond".
1874 static __isl_give isl_set *after(__isl_take isl_set *cond, int sign)
1876 isl_map *previous_to_this;
1878 if (sign > 0)
1879 previous_to_this = isl_map_lex_lt(isl_set_get_space(cond));
1880 else
1881 previous_to_this = isl_map_lex_gt(isl_set_get_space(cond));
1883 cond = isl_set_apply(cond, previous_to_this);
1885 return cond;
1888 /* Create the infinite iteration domain
1890 * { [id] : id >= 0 }
1892 * If "scop" has an affine skip of type pet_skip_later,
1893 * then remove those iterations i that have an earlier iteration
1894 * where the skip condition is satisfied, meaning that iteration i
1895 * is not executed.
1896 * Since we are dealing with a loop without loop iterator,
1897 * the skip condition cannot refer to the current loop iterator and
1898 * so effectively, the returned set is of the form
1900 * { [0]; [id] : id >= 1 and not skip }
1902 static __isl_give isl_set *infinite_domain(__isl_take isl_id *id,
1903 struct pet_scop *scop)
1905 isl_ctx *ctx = isl_id_get_ctx(id);
1906 isl_set *domain;
1907 isl_set *skip;
1909 domain = isl_set_nat_universe(isl_space_set_alloc(ctx, 0, 1));
1910 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, id);
1912 if (!pet_scop_has_affine_skip(scop, pet_skip_later))
1913 return domain;
1915 skip = pet_scop_get_skip(scop, pet_skip_later);
1916 skip = isl_set_fix_si(skip, isl_dim_set, 0, 1);
1917 skip = isl_set_params(skip);
1918 skip = embed(skip, isl_id_copy(id));
1919 skip = isl_set_intersect(skip , isl_set_copy(domain));
1920 domain = isl_set_subtract(domain, after(skip, 1));
1922 return domain;
1925 /* Create an identity mapping on the space containing "domain".
1927 static __isl_give isl_map *identity_map(__isl_keep isl_set *domain)
1929 isl_space *space;
1930 isl_map *id;
1932 space = isl_space_map_from_set(isl_set_get_space(domain));
1933 id = isl_map_identity(space);
1935 return id;
1938 /* Add a filter to "scop" that imposes that it is only executed
1939 * when "break_access" has a zero value for all previous iterations
1940 * of "domain".
1942 * The input "break_access" has a zero-dimensional domain and range.
1944 static struct pet_scop *scop_add_break(struct pet_scop *scop,
1945 __isl_take isl_map *break_access, __isl_take isl_set *domain, int sign)
1947 isl_ctx *ctx = isl_set_get_ctx(domain);
1948 isl_id *id_test;
1949 isl_map *prev;
1951 id_test = isl_map_get_tuple_id(break_access, isl_dim_out);
1952 break_access = isl_map_add_dims(break_access, isl_dim_in, 1);
1953 break_access = isl_map_add_dims(break_access, isl_dim_out, 1);
1954 break_access = isl_map_intersect_range(break_access, domain);
1955 break_access = isl_map_set_tuple_id(break_access, isl_dim_out, id_test);
1956 if (sign > 0)
1957 prev = isl_map_lex_gt_first(isl_map_get_space(break_access), 1);
1958 else
1959 prev = isl_map_lex_lt_first(isl_map_get_space(break_access), 1);
1960 break_access = isl_map_intersect(break_access, prev);
1961 scop = pet_scop_filter(scop, break_access, 0);
1962 scop = pet_scop_merge_filters(scop);
1964 return scop;
1967 /* Construct a pet_scop for an infinite loop around the given body.
1969 * We extract a pet_scop for the body and then embed it in a loop with
1970 * iteration domain
1972 * { [t] : t >= 0 }
1974 * and schedule
1976 * { [t] -> [t] }
1978 * If the body contains any break, then it is taken into
1979 * account in infinite_domain (if the skip condition is affine)
1980 * or in scop_add_break (if the skip condition is not affine).
1982 struct pet_scop *PetScan::extract_infinite_loop(Stmt *body)
1984 isl_id *id;
1985 isl_set *domain;
1986 isl_map *ident;
1987 isl_map *access;
1988 struct pet_scop *scop;
1989 bool has_var_break;
1991 scop = extract(body);
1992 if (!scop)
1993 return NULL;
1995 id = isl_id_alloc(ctx, "t", NULL);
1996 domain = infinite_domain(isl_id_copy(id), scop);
1997 ident = identity_map(domain);
1999 has_var_break = pet_scop_has_var_skip(scop, pet_skip_later);
2000 if (has_var_break)
2001 access = pet_scop_get_skip_map(scop, pet_skip_later);
2003 scop = pet_scop_embed(scop, isl_set_copy(domain),
2004 isl_map_copy(ident), ident, id);
2005 if (has_var_break)
2006 scop = scop_add_break(scop, access, domain, 1);
2007 else
2008 isl_set_free(domain);
2010 return scop;
2013 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
2015 * for (;;)
2016 * body
2019 struct pet_scop *PetScan::extract_infinite_for(ForStmt *stmt)
2021 return extract_infinite_loop(stmt->getBody());
2024 /* Create an access to a virtual array representing the result
2025 * of a condition.
2026 * Unlike other accessed data, the id of the array is NULL as
2027 * there is no ValueDecl in the program corresponding to the virtual
2028 * array.
2029 * The array starts out as a scalar, but grows along with the
2030 * statement writing to the array in pet_scop_embed.
2032 static __isl_give isl_map *create_test_access(isl_ctx *ctx, int test_nr)
2034 isl_space *dim = isl_space_alloc(ctx, 0, 0, 0);
2035 isl_id *id;
2036 char name[50];
2038 snprintf(name, sizeof(name), "__pet_test_%d", test_nr);
2039 id = isl_id_alloc(ctx, name, NULL);
2040 dim = isl_space_set_tuple_id(dim, isl_dim_out, id);
2041 return isl_map_universe(dim);
2044 /* Add an array with the given extent ("access") to the list
2045 * of arrays in "scop" and return the extended pet_scop.
2046 * The array is marked as attaining values 0 and 1 only and
2047 * as each element being assigned at most once.
2049 static struct pet_scop *scop_add_array(struct pet_scop *scop,
2050 __isl_keep isl_map *access, clang::ASTContext &ast_ctx)
2052 isl_ctx *ctx = isl_map_get_ctx(access);
2053 isl_space *dim;
2054 struct pet_array **arrays;
2055 struct pet_array *array;
2057 if (!scop)
2058 return NULL;
2059 if (!ctx)
2060 goto error;
2062 arrays = isl_realloc_array(ctx, scop->arrays, struct pet_array *,
2063 scop->n_array + 1);
2064 if (!arrays)
2065 goto error;
2066 scop->arrays = arrays;
2068 array = isl_calloc_type(ctx, struct pet_array);
2069 if (!array)
2070 goto error;
2072 array->extent = isl_map_range(isl_map_copy(access));
2073 dim = isl_space_params_alloc(ctx, 0);
2074 array->context = isl_set_universe(dim);
2075 dim = isl_space_set_alloc(ctx, 0, 1);
2076 array->value_bounds = isl_set_universe(dim);
2077 array->value_bounds = isl_set_lower_bound_si(array->value_bounds,
2078 isl_dim_set, 0, 0);
2079 array->value_bounds = isl_set_upper_bound_si(array->value_bounds,
2080 isl_dim_set, 0, 1);
2081 array->element_type = strdup("int");
2082 array->element_size = ast_ctx.getTypeInfo(ast_ctx.IntTy).first / 8;
2083 array->uniquely_defined = 1;
2085 scop->arrays[scop->n_array] = array;
2086 scop->n_array++;
2088 if (!array->extent || !array->context)
2089 goto error;
2091 return scop;
2092 error:
2093 pet_scop_free(scop);
2094 return NULL;
2097 /* Construct a pet_scop for a while loop of the form
2099 * while (pa)
2100 * body
2102 * In particular, construct a scop for an infinite loop around body and
2103 * intersect the domain with the affine expression.
2104 * Note that this intersection may result in an empty loop.
2106 struct pet_scop *PetScan::extract_affine_while(__isl_take isl_pw_aff *pa,
2107 Stmt *body)
2109 struct pet_scop *scop;
2110 isl_set *dom;
2111 isl_set *valid;
2113 valid = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2114 dom = isl_pw_aff_non_zero_set(pa);
2115 scop = extract_infinite_loop(body);
2116 scop = pet_scop_restrict(scop, dom);
2117 scop = pet_scop_restrict_context(scop, valid);
2119 return scop;
2122 /* Construct a scop for a while, given the scops for the condition
2123 * and the body, the filter access and the iteration domain of
2124 * the while loop.
2126 * In particular, the scop for the condition is filtered to depend
2127 * on "test_access" evaluating to true for all previous iterations
2128 * of the loop, while the scop for the body is filtered to depend
2129 * on "test_access" evaluating to true for all iterations up to the
2130 * current iteration.
2132 * These filtered scops are then combined into a single scop.
2134 * "sign" is positive if the iterator increases and negative
2135 * if it decreases.
2137 static struct pet_scop *scop_add_while(struct pet_scop *scop_cond,
2138 struct pet_scop *scop_body, __isl_take isl_map *test_access,
2139 __isl_take isl_set *domain, int sign)
2141 isl_ctx *ctx = isl_set_get_ctx(domain);
2142 isl_id *id_test;
2143 isl_map *prev;
2145 id_test = isl_map_get_tuple_id(test_access, isl_dim_out);
2146 test_access = isl_map_add_dims(test_access, isl_dim_in, 1);
2147 test_access = isl_map_add_dims(test_access, isl_dim_out, 1);
2148 test_access = isl_map_intersect_range(test_access, domain);
2149 test_access = isl_map_set_tuple_id(test_access, isl_dim_out, id_test);
2150 if (sign > 0)
2151 prev = isl_map_lex_ge_first(isl_map_get_space(test_access), 1);
2152 else
2153 prev = isl_map_lex_le_first(isl_map_get_space(test_access), 1);
2154 test_access = isl_map_intersect(test_access, prev);
2155 scop_body = pet_scop_filter(scop_body, isl_map_copy(test_access), 1);
2156 if (sign > 0)
2157 prev = isl_map_lex_gt_first(isl_map_get_space(test_access), 1);
2158 else
2159 prev = isl_map_lex_lt_first(isl_map_get_space(test_access), 1);
2160 test_access = isl_map_intersect(test_access, prev);
2161 scop_cond = pet_scop_filter(scop_cond, test_access, 1);
2163 return pet_scop_add_seq(ctx, scop_cond, scop_body);
2166 /* Check if the while loop is of the form
2168 * while (affine expression)
2169 * body
2171 * If so, call extract_affine_while to construct a scop.
2173 * Otherwise, construct a generic while scop, with iteration domain
2174 * { [t] : t >= 0 }. The scop consists of two parts, one for
2175 * evaluating the condition and one for the body.
2176 * The schedule is adjusted to reflect that the condition is evaluated
2177 * before the body is executed and the body is filtered to depend
2178 * on the result of the condition evaluating to true on all iterations
2179 * up to the current iteration, while the evaluation the condition itself
2180 * is filtered to depend on the result of the condition evaluating to true
2181 * on all previous iterations.
2182 * The context of the scop representing the body is dropped
2183 * because we don't know how many times the body will be executed,
2184 * if at all.
2186 * If the body contains any break, then it is taken into
2187 * account in infinite_domain (if the skip condition is affine)
2188 * or in scop_add_break (if the skip condition is not affine).
2190 struct pet_scop *PetScan::extract(WhileStmt *stmt)
2192 Expr *cond;
2193 isl_id *id;
2194 isl_map *test_access;
2195 isl_set *domain;
2196 isl_map *ident;
2197 isl_pw_aff *pa;
2198 struct pet_scop *scop, *scop_body;
2199 bool has_var_break;
2200 isl_map *break_access;
2202 cond = stmt->getCond();
2203 if (!cond) {
2204 unsupported(stmt);
2205 return NULL;
2208 pa = try_extract_affine_condition(cond);
2209 if (pa)
2210 return extract_affine_while(pa, stmt->getBody());
2212 if (!allow_nested) {
2213 unsupported(stmt);
2214 return NULL;
2217 test_access = create_test_access(ctx, n_test++);
2218 scop = extract_non_affine_condition(cond, isl_map_copy(test_access));
2219 scop = scop_add_array(scop, test_access, ast_context);
2220 scop_body = extract(stmt->getBody());
2222 id = isl_id_alloc(ctx, "t", NULL);
2223 domain = infinite_domain(isl_id_copy(id), scop_body);
2224 ident = identity_map(domain);
2226 has_var_break = pet_scop_has_var_skip(scop_body, pet_skip_later);
2227 if (has_var_break)
2228 break_access = pet_scop_get_skip_map(scop_body, pet_skip_later);
2230 scop = pet_scop_prefix(scop, 0);
2231 scop = pet_scop_embed(scop, isl_set_copy(domain), isl_map_copy(ident),
2232 isl_map_copy(ident), isl_id_copy(id));
2233 scop_body = pet_scop_reset_context(scop_body);
2234 scop_body = pet_scop_prefix(scop_body, 1);
2235 scop_body = pet_scop_embed(scop_body, isl_set_copy(domain),
2236 isl_map_copy(ident), ident, id);
2238 if (has_var_break) {
2239 scop = scop_add_break(scop, isl_map_copy(break_access),
2240 isl_set_copy(domain), 1);
2241 scop_body = scop_add_break(scop_body, break_access,
2242 isl_set_copy(domain), 1);
2244 scop = scop_add_while(scop, scop_body, test_access, domain, 1);
2246 return scop;
2249 /* Check whether "cond" expresses a simple loop bound
2250 * on the only set dimension.
2251 * In particular, if "up" is set then "cond" should contain only
2252 * upper bounds on the set dimension.
2253 * Otherwise, it should contain only lower bounds.
2255 static bool is_simple_bound(__isl_keep isl_set *cond, isl_int inc)
2257 if (isl_int_is_pos(inc))
2258 return !isl_set_dim_has_lower_bound(cond, isl_dim_set, 0);
2259 else
2260 return !isl_set_dim_has_upper_bound(cond, isl_dim_set, 0);
2263 /* Extend a condition on a given iteration of a loop to one that
2264 * imposes the same condition on all previous iterations.
2265 * "domain" expresses the lower [upper] bound on the iterations
2266 * when inc is positive [negative].
2268 * In particular, we construct the condition (when inc is positive)
2270 * forall i' : (domain(i') and i' <= i) => cond(i')
2272 * which is equivalent to
2274 * not exists i' : domain(i') and i' <= i and not cond(i')
2276 * We construct this set by negating cond, applying a map
2278 * { [i'] -> [i] : domain(i') and i' <= i }
2280 * and then negating the result again.
2282 static __isl_give isl_set *valid_for_each_iteration(__isl_take isl_set *cond,
2283 __isl_take isl_set *domain, isl_int inc)
2285 isl_map *previous_to_this;
2287 if (isl_int_is_pos(inc))
2288 previous_to_this = isl_map_lex_le(isl_set_get_space(domain));
2289 else
2290 previous_to_this = isl_map_lex_ge(isl_set_get_space(domain));
2292 previous_to_this = isl_map_intersect_domain(previous_to_this, domain);
2294 cond = isl_set_complement(cond);
2295 cond = isl_set_apply(cond, previous_to_this);
2296 cond = isl_set_complement(cond);
2298 return cond;
2301 /* Construct a domain of the form
2303 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
2305 static __isl_give isl_set *strided_domain(__isl_take isl_id *id,
2306 __isl_take isl_pw_aff *init, isl_int inc)
2308 isl_aff *aff;
2309 isl_space *dim;
2310 isl_set *set;
2312 init = isl_pw_aff_insert_dims(init, isl_dim_in, 0, 1);
2313 dim = isl_pw_aff_get_domain_space(init);
2314 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2315 aff = isl_aff_add_coefficient(aff, isl_dim_in, 0, inc);
2316 init = isl_pw_aff_add(init, isl_pw_aff_from_aff(aff));
2318 dim = isl_space_set_alloc(isl_pw_aff_get_ctx(init), 1, 1);
2319 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
2320 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2321 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2323 set = isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff), init);
2325 set = isl_set_lower_bound_si(set, isl_dim_set, 0, 0);
2327 return isl_set_params(set);
2330 /* Assuming "cond" represents a bound on a loop where the loop
2331 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2332 * is possible.
2334 * Under the given assumptions, wrapping is only possible if "cond" allows
2335 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2336 * increasing iterator and 0 in case of a decreasing iterator.
2338 static bool can_wrap(__isl_keep isl_set *cond, ValueDecl *iv, isl_int inc)
2340 bool cw;
2341 isl_int limit;
2342 isl_set *test;
2344 test = isl_set_copy(cond);
2346 isl_int_init(limit);
2347 if (isl_int_is_neg(inc))
2348 isl_int_set_si(limit, 0);
2349 else {
2350 isl_int_set_si(limit, 1);
2351 isl_int_mul_2exp(limit, limit, get_type_size(iv));
2352 isl_int_sub_ui(limit, limit, 1);
2355 test = isl_set_fix(cond, isl_dim_set, 0, limit);
2356 cw = !isl_set_is_empty(test);
2357 isl_set_free(test);
2359 isl_int_clear(limit);
2361 return cw;
2364 /* Given a one-dimensional space, construct the following mapping on this
2365 * space
2367 * { [v] -> [v mod 2^width] }
2369 * where width is the number of bits used to represent the values
2370 * of the unsigned variable "iv".
2372 static __isl_give isl_map *compute_wrapping(__isl_take isl_space *dim,
2373 ValueDecl *iv)
2375 isl_int mod;
2376 isl_aff *aff;
2377 isl_map *map;
2379 isl_int_init(mod);
2380 isl_int_set_si(mod, 1);
2381 isl_int_mul_2exp(mod, mod, get_type_size(iv));
2383 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2384 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2385 aff = isl_aff_mod(aff, mod);
2387 isl_int_clear(mod);
2389 return isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2390 map = isl_map_reverse(map);
2393 /* Project out the parameter "id" from "set".
2395 static __isl_give isl_set *set_project_out_by_id(__isl_take isl_set *set,
2396 __isl_keep isl_id *id)
2398 int pos;
2400 pos = isl_set_find_dim_by_id(set, isl_dim_param, id);
2401 if (pos >= 0)
2402 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2404 return set;
2407 /* Compute the set of parameters for which "set1" is a subset of "set2".
2409 * set1 is a subset of set2 if
2411 * forall i in set1 : i in set2
2413 * or
2415 * not exists i in set1 and i not in set2
2417 * i.e.,
2419 * not exists i in set1 \ set2
2421 static __isl_give isl_set *enforce_subset(__isl_take isl_set *set1,
2422 __isl_take isl_set *set2)
2424 return isl_set_complement(isl_set_params(isl_set_subtract(set1, set2)));
2427 /* Compute the set of parameter values for which "cond" holds
2428 * on the next iteration for each element of "dom".
2430 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2431 * and then compute the set of parameters for which the result is a subset
2432 * of "cond".
2434 static __isl_give isl_set *valid_on_next(__isl_take isl_set *cond,
2435 __isl_take isl_set *dom, isl_int inc)
2437 isl_space *space;
2438 isl_aff *aff;
2439 isl_map *next;
2441 space = isl_set_get_space(dom);
2442 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2443 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2444 aff = isl_aff_add_constant(aff, inc);
2445 next = isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2447 dom = isl_set_apply(dom, next);
2449 return enforce_subset(dom, cond);
2452 /* Does "id" refer to a nested access?
2454 static bool is_nested_parameter(__isl_keep isl_id *id)
2456 return id && isl_id_get_user(id) && !isl_id_get_name(id);
2459 /* Does parameter "pos" of "space" refer to a nested access?
2461 static bool is_nested_parameter(__isl_keep isl_space *space, int pos)
2463 bool nested;
2464 isl_id *id;
2466 id = isl_space_get_dim_id(space, isl_dim_param, pos);
2467 nested = is_nested_parameter(id);
2468 isl_id_free(id);
2470 return nested;
2473 /* Does "space" involve any parameters that refer to nested
2474 * accesses, i.e., parameters with no name?
2476 static bool has_nested(__isl_keep isl_space *space)
2478 int nparam;
2480 nparam = isl_space_dim(space, isl_dim_param);
2481 for (int i = 0; i < nparam; ++i)
2482 if (is_nested_parameter(space, i))
2483 return true;
2485 return false;
2488 /* Does "pa" involve any parameters that refer to nested
2489 * accesses, i.e., parameters with no name?
2491 static bool has_nested(__isl_keep isl_pw_aff *pa)
2493 isl_space *space;
2494 bool nested;
2496 space = isl_pw_aff_get_space(pa);
2497 nested = has_nested(space);
2498 isl_space_free(space);
2500 return nested;
2503 /* Construct a pet_scop for a for statement.
2504 * The for loop is required to be of the form
2506 * for (i = init; condition; ++i)
2508 * or
2510 * for (i = init; condition; --i)
2512 * The initialization of the for loop should either be an assignment
2513 * to an integer variable, or a declaration of such a variable with
2514 * initialization.
2516 * The condition is allowed to contain nested accesses, provided
2517 * they are not being written to inside the body of the loop.
2518 * Otherwise, or if the condition is otherwise non-affine, the for loop is
2519 * essentially treated as a while loop, with iteration domain
2520 * { [i] : i >= init }.
2522 * We extract a pet_scop for the body and then embed it in a loop with
2523 * iteration domain and schedule
2525 * { [i] : i >= init and condition' }
2526 * { [i] -> [i] }
2528 * or
2530 * { [i] : i <= init and condition' }
2531 * { [i] -> [-i] }
2533 * Where condition' is equal to condition if the latter is
2534 * a simple upper [lower] bound and a condition that is extended
2535 * to apply to all previous iterations otherwise.
2537 * If the condition is non-affine, then we drop the condition from the
2538 * iteration domain and instead create a separate statement
2539 * for evaluating the condition. The body is then filtered to depend
2540 * on the result of the condition evaluating to true on all iterations
2541 * up to the current iteration, while the evaluation the condition itself
2542 * is filtered to depend on the result of the condition evaluating to true
2543 * on all previous iterations.
2544 * The context of the scop representing the body is dropped
2545 * because we don't know how many times the body will be executed,
2546 * if at all.
2548 * If the stride of the loop is not 1, then "i >= init" is replaced by
2550 * (exists a: i = init + stride * a and a >= 0)
2552 * If the loop iterator i is unsigned, then wrapping may occur.
2553 * During the computation, we work with a virtual iterator that
2554 * does not wrap. However, the condition in the code applies
2555 * to the wrapped value, so we need to change condition(i)
2556 * into condition([i % 2^width]).
2557 * After computing the virtual domain and schedule, we apply
2558 * the function { [v] -> [v % 2^width] } to the domain and the domain
2559 * of the schedule. In order not to lose any information, we also
2560 * need to intersect the domain of the schedule with the virtual domain
2561 * first, since some iterations in the wrapped domain may be scheduled
2562 * several times, typically an infinite number of times.
2563 * Note that there may be no need to perform this final wrapping
2564 * if the loop condition (after wrapping) satisfies certain conditions.
2565 * However, the is_simple_bound condition is not enough since it doesn't
2566 * check if there even is an upper bound.
2568 * If the loop condition is non-affine, then we keep the virtual
2569 * iterator in the iteration domain and instead replace all accesses
2570 * to the original iterator by the wrapping of the virtual iterator.
2572 * Wrapping on unsigned iterators can be avoided entirely if
2573 * loop condition is simple, the loop iterator is incremented
2574 * [decremented] by one and the last value before wrapping cannot
2575 * possibly satisfy the loop condition.
2577 * Before extracting a pet_scop from the body we remove all
2578 * assignments in assigned_value to variables that are assigned
2579 * somewhere in the body of the loop.
2581 * Valid parameters for a for loop are those for which the initial
2582 * value itself, the increment on each domain iteration and
2583 * the condition on both the initial value and
2584 * the result of incrementing the iterator for each iteration of the domain
2585 * can be evaluated.
2586 * If the loop condition is non-affine, then we only consider validity
2587 * of the initial value.
2589 * If the body contains any break, then we keep track of it in "skip"
2590 * (if the skip condition is affine) or it is handled in scop_add_break
2591 * (if the skip condition is not affine).
2592 * Note that the affine break condition needs to be considered with
2593 * respect to previous iterations in the virtual domain (if any)
2594 * and that the domain needs to be kept virtual if there is a non-affine
2595 * break condition.
2597 struct pet_scop *PetScan::extract_for(ForStmt *stmt)
2599 BinaryOperator *ass;
2600 Decl *decl;
2601 Stmt *init;
2602 Expr *lhs, *rhs;
2603 ValueDecl *iv;
2604 isl_space *space;
2605 isl_set *domain;
2606 isl_map *sched;
2607 isl_set *cond = NULL;
2608 isl_set *skip = NULL;
2609 isl_id *id;
2610 struct pet_scop *scop, *scop_cond = NULL;
2611 assigned_value_cache cache(assigned_value);
2612 isl_int inc;
2613 bool is_one;
2614 bool is_unsigned;
2615 bool is_simple;
2616 bool is_virtual;
2617 bool keep_virtual = false;
2618 bool has_affine_break;
2619 bool has_var_break;
2620 isl_map *wrap = NULL;
2621 isl_pw_aff *pa, *pa_inc, *init_val;
2622 isl_set *valid_init;
2623 isl_set *valid_cond;
2624 isl_set *valid_cond_init;
2625 isl_set *valid_cond_next;
2626 isl_set *valid_inc;
2627 isl_map *test_access = NULL, *break_access = NULL;
2628 int stmt_id;
2630 if (!stmt->getInit() && !stmt->getCond() && !stmt->getInc())
2631 return extract_infinite_for(stmt);
2633 init = stmt->getInit();
2634 if (!init) {
2635 unsupported(stmt);
2636 return NULL;
2638 if ((ass = initialization_assignment(init)) != NULL) {
2639 iv = extract_induction_variable(ass);
2640 if (!iv)
2641 return NULL;
2642 lhs = ass->getLHS();
2643 rhs = ass->getRHS();
2644 } else if ((decl = initialization_declaration(init)) != NULL) {
2645 VarDecl *var = extract_induction_variable(init, decl);
2646 if (!var)
2647 return NULL;
2648 iv = var;
2649 rhs = var->getInit();
2650 lhs = create_DeclRefExpr(var);
2651 } else {
2652 unsupported(stmt->getInit());
2653 return NULL;
2656 pa_inc = extract_increment(stmt, iv);
2657 if (!pa_inc)
2658 return NULL;
2660 isl_int_init(inc);
2661 if (isl_pw_aff_n_piece(pa_inc) != 1 ||
2662 isl_pw_aff_foreach_piece(pa_inc, &extract_cst, &inc) < 0) {
2663 isl_pw_aff_free(pa_inc);
2664 unsupported(stmt->getInc());
2665 isl_int_clear(inc);
2666 return NULL;
2668 valid_inc = isl_pw_aff_domain(pa_inc);
2670 is_unsigned = iv->getType()->isUnsignedIntegerType();
2672 assigned_value.erase(iv);
2673 clear_assignments clear(assigned_value);
2674 clear.TraverseStmt(stmt->getBody());
2676 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
2678 pa = try_extract_nested_condition(stmt->getCond());
2679 if (allow_nested && (!pa || has_nested(pa)))
2680 stmt_id = n_stmt++;
2682 scop = extract(stmt->getBody());
2684 has_affine_break = scop &&
2685 pet_scop_has_affine_skip(scop, pet_skip_later);
2686 if (has_affine_break) {
2687 skip = pet_scop_get_skip(scop, pet_skip_later);
2688 skip = isl_set_fix_si(skip, isl_dim_set, 0, 1);
2689 skip = isl_set_params(skip);
2691 has_var_break = scop && pet_scop_has_var_skip(scop, pet_skip_later);
2692 if (has_var_break) {
2693 break_access = pet_scop_get_skip_map(scop, pet_skip_later);
2694 keep_virtual = true;
2697 if (pa && !is_nested_allowed(pa, scop)) {
2698 isl_pw_aff_free(pa);
2699 pa = NULL;
2702 if (!allow_nested && !pa)
2703 pa = try_extract_affine_condition(stmt->getCond());
2704 valid_cond = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2705 cond = isl_pw_aff_non_zero_set(pa);
2706 if (allow_nested && !cond) {
2707 int save_n_stmt = n_stmt;
2708 test_access = create_test_access(ctx, n_test++);
2709 n_stmt = stmt_id;
2710 scop_cond = extract_non_affine_condition(stmt->getCond(),
2711 isl_map_copy(test_access));
2712 n_stmt = save_n_stmt;
2713 scop_cond = scop_add_array(scop_cond, test_access, ast_context);
2714 scop_cond = pet_scop_prefix(scop_cond, 0);
2715 scop = pet_scop_reset_context(scop);
2716 scop = pet_scop_prefix(scop, 1);
2717 keep_virtual = true;
2718 cond = isl_set_universe(isl_space_set_alloc(ctx, 0, 0));
2721 cond = embed(cond, isl_id_copy(id));
2722 skip = embed(skip, isl_id_copy(id));
2723 valid_cond = isl_set_coalesce(valid_cond);
2724 valid_cond = embed(valid_cond, isl_id_copy(id));
2725 valid_inc = embed(valid_inc, isl_id_copy(id));
2726 is_one = isl_int_is_one(inc) || isl_int_is_negone(inc);
2727 is_virtual = is_unsigned && (!is_one || can_wrap(cond, iv, inc));
2729 init_val = extract_affine(rhs);
2730 valid_cond_init = enforce_subset(
2731 isl_set_from_pw_aff(isl_pw_aff_copy(init_val)),
2732 isl_set_copy(valid_cond));
2733 if (is_one && !is_virtual) {
2734 isl_pw_aff_free(init_val);
2735 pa = extract_comparison(isl_int_is_pos(inc) ? BO_GE : BO_LE,
2736 lhs, rhs, init);
2737 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2738 valid_init = set_project_out_by_id(valid_init, id);
2739 domain = isl_pw_aff_non_zero_set(pa);
2740 } else {
2741 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(init_val));
2742 domain = strided_domain(isl_id_copy(id), init_val, inc);
2745 domain = embed(domain, isl_id_copy(id));
2746 if (is_virtual) {
2747 isl_map *rev_wrap;
2748 wrap = compute_wrapping(isl_set_get_space(cond), iv);
2749 rev_wrap = isl_map_reverse(isl_map_copy(wrap));
2750 cond = isl_set_apply(cond, isl_map_copy(rev_wrap));
2751 skip = isl_set_apply(skip, isl_map_copy(rev_wrap));
2752 valid_cond = isl_set_apply(valid_cond, isl_map_copy(rev_wrap));
2753 valid_inc = isl_set_apply(valid_inc, rev_wrap);
2755 cond = isl_set_gist(cond, isl_set_copy(domain));
2756 is_simple = is_simple_bound(cond, inc);
2757 if (!is_simple)
2758 cond = valid_for_each_iteration(cond,
2759 isl_set_copy(domain), inc);
2760 domain = isl_set_intersect(domain, cond);
2761 if (has_affine_break) {
2762 skip = isl_set_intersect(skip , isl_set_copy(domain));
2763 skip = after(skip, isl_int_sgn(inc));
2764 domain = isl_set_subtract(domain, skip);
2766 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, isl_id_copy(id));
2767 space = isl_space_from_domain(isl_set_get_space(domain));
2768 space = isl_space_add_dims(space, isl_dim_out, 1);
2769 sched = isl_map_universe(space);
2770 if (isl_int_is_pos(inc))
2771 sched = isl_map_equate(sched, isl_dim_in, 0, isl_dim_out, 0);
2772 else
2773 sched = isl_map_oppose(sched, isl_dim_in, 0, isl_dim_out, 0);
2775 valid_cond_next = valid_on_next(valid_cond, isl_set_copy(domain), inc);
2776 valid_inc = enforce_subset(isl_set_copy(domain), valid_inc);
2778 if (is_virtual && !keep_virtual) {
2779 wrap = isl_map_set_dim_id(wrap,
2780 isl_dim_out, 0, isl_id_copy(id));
2781 sched = isl_map_intersect_domain(sched, isl_set_copy(domain));
2782 domain = isl_set_apply(domain, isl_map_copy(wrap));
2783 sched = isl_map_apply_domain(sched, wrap);
2785 if (!(is_virtual && keep_virtual)) {
2786 space = isl_set_get_space(domain);
2787 wrap = isl_map_identity(isl_space_map_from_set(space));
2790 scop_cond = pet_scop_embed(scop_cond, isl_set_copy(domain),
2791 isl_map_copy(sched), isl_map_copy(wrap), isl_id_copy(id));
2792 scop = pet_scop_embed(scop, isl_set_copy(domain), sched, wrap, id);
2793 scop = resolve_nested(scop);
2794 if (has_var_break)
2795 scop = scop_add_break(scop, break_access, isl_set_copy(domain),
2796 isl_int_sgn(inc));
2797 if (test_access) {
2798 scop = scop_add_while(scop_cond, scop, test_access, domain,
2799 isl_int_sgn(inc));
2800 isl_set_free(valid_inc);
2801 } else {
2802 scop = pet_scop_restrict_context(scop, valid_inc);
2803 scop = pet_scop_restrict_context(scop, valid_cond_next);
2804 scop = pet_scop_restrict_context(scop, valid_cond_init);
2805 isl_set_free(domain);
2807 clear_assignment(assigned_value, iv);
2809 isl_int_clear(inc);
2811 scop = pet_scop_restrict_context(scop, valid_init);
2813 return scop;
2816 struct pet_scop *PetScan::extract(CompoundStmt *stmt)
2818 return extract(stmt->children());
2821 /* Does parameter "pos" of "map" refer to a nested access?
2823 static bool is_nested_parameter(__isl_keep isl_map *map, int pos)
2825 bool nested;
2826 isl_id *id;
2828 id = isl_map_get_dim_id(map, isl_dim_param, pos);
2829 nested = is_nested_parameter(id);
2830 isl_id_free(id);
2832 return nested;
2835 /* How many parameters of "space" refer to nested accesses, i.e., have no name?
2837 static int n_nested_parameter(__isl_keep isl_space *space)
2839 int n = 0;
2840 int nparam;
2842 nparam = isl_space_dim(space, isl_dim_param);
2843 for (int i = 0; i < nparam; ++i)
2844 if (is_nested_parameter(space, i))
2845 ++n;
2847 return n;
2850 /* How many parameters of "map" refer to nested accesses, i.e., have no name?
2852 static int n_nested_parameter(__isl_keep isl_map *map)
2854 isl_space *space;
2855 int n;
2857 space = isl_map_get_space(map);
2858 n = n_nested_parameter(space);
2859 isl_space_free(space);
2861 return n;
2864 /* For each nested access parameter in "space",
2865 * construct a corresponding pet_expr, place it in args and
2866 * record its position in "param2pos".
2867 * "n_arg" is the number of elements that are already in args.
2868 * The position recorded in "param2pos" takes this number into account.
2869 * If the pet_expr corresponding to a parameter is identical to
2870 * the pet_expr corresponding to an earlier parameter, then these two
2871 * parameters are made to refer to the same element in args.
2873 * Return the final number of elements in args or -1 if an error has occurred.
2875 int PetScan::extract_nested(__isl_keep isl_space *space,
2876 int n_arg, struct pet_expr **args, std::map<int,int> &param2pos)
2878 int nparam;
2880 nparam = isl_space_dim(space, isl_dim_param);
2881 for (int i = 0; i < nparam; ++i) {
2882 int j;
2883 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
2884 Expr *nested;
2886 if (!is_nested_parameter(id)) {
2887 isl_id_free(id);
2888 continue;
2891 nested = (Expr *) isl_id_get_user(id);
2892 args[n_arg] = extract_expr(nested);
2893 if (!args[n_arg])
2894 return -1;
2896 for (j = 0; j < n_arg; ++j)
2897 if (pet_expr_is_equal(args[j], args[n_arg]))
2898 break;
2900 if (j < n_arg) {
2901 pet_expr_free(args[n_arg]);
2902 args[n_arg] = NULL;
2903 param2pos[i] = j;
2904 } else
2905 param2pos[i] = n_arg++;
2907 isl_id_free(id);
2910 return n_arg;
2913 /* For each nested access parameter in the access relations in "expr",
2914 * construct a corresponding pet_expr, place it in expr->args and
2915 * record its position in "param2pos".
2916 * n is the number of nested access parameters.
2918 struct pet_expr *PetScan::extract_nested(struct pet_expr *expr, int n,
2919 std::map<int,int> &param2pos)
2921 isl_space *space;
2923 expr->args = isl_calloc_array(ctx, struct pet_expr *, n);
2924 expr->n_arg = n;
2925 if (!expr->args)
2926 goto error;
2928 space = isl_map_get_space(expr->acc.access);
2929 n = extract_nested(space, 0, expr->args, param2pos);
2930 isl_space_free(space);
2932 if (n < 0)
2933 goto error;
2935 expr->n_arg = n;
2936 return expr;
2937 error:
2938 pet_expr_free(expr);
2939 return NULL;
2942 /* Look for parameters in any access relation in "expr" that
2943 * refer to nested accesses. In particular, these are
2944 * parameters with no name.
2946 * If there are any such parameters, then the domain of the access
2947 * relation, which is still [] at this point, is replaced by
2948 * [[] -> [t_1,...,t_n]], with n the number of these parameters
2949 * (after identifying identical nested accesses).
2950 * The parameters are then equated to the corresponding t dimensions
2951 * and subsequently projected out.
2952 * param2pos maps the position of the parameter to the position
2953 * of the corresponding t dimension.
2955 struct pet_expr *PetScan::resolve_nested(struct pet_expr *expr)
2957 int n;
2958 int nparam;
2959 int n_in;
2960 isl_space *dim;
2961 isl_map *map;
2962 std::map<int,int> param2pos;
2964 if (!expr)
2965 return expr;
2967 for (int i = 0; i < expr->n_arg; ++i) {
2968 expr->args[i] = resolve_nested(expr->args[i]);
2969 if (!expr->args[i]) {
2970 pet_expr_free(expr);
2971 return NULL;
2975 if (expr->type != pet_expr_access)
2976 return expr;
2978 n = n_nested_parameter(expr->acc.access);
2979 if (n == 0)
2980 return expr;
2982 expr = extract_nested(expr, n, param2pos);
2983 if (!expr)
2984 return NULL;
2986 n = expr->n_arg;
2987 nparam = isl_map_dim(expr->acc.access, isl_dim_param);
2988 n_in = isl_map_dim(expr->acc.access, isl_dim_in);
2989 dim = isl_map_get_space(expr->acc.access);
2990 dim = isl_space_domain(dim);
2991 dim = isl_space_from_domain(dim);
2992 dim = isl_space_add_dims(dim, isl_dim_out, n);
2993 map = isl_map_universe(dim);
2994 map = isl_map_domain_map(map);
2995 map = isl_map_reverse(map);
2996 expr->acc.access = isl_map_apply_domain(expr->acc.access, map);
2998 for (int i = nparam - 1; i >= 0; --i) {
2999 isl_id *id = isl_map_get_dim_id(expr->acc.access,
3000 isl_dim_param, i);
3001 if (!is_nested_parameter(id)) {
3002 isl_id_free(id);
3003 continue;
3006 expr->acc.access = isl_map_equate(expr->acc.access,
3007 isl_dim_param, i, isl_dim_in,
3008 n_in + param2pos[i]);
3009 expr->acc.access = isl_map_project_out(expr->acc.access,
3010 isl_dim_param, i, 1);
3012 isl_id_free(id);
3015 return expr;
3016 error:
3017 pet_expr_free(expr);
3018 return NULL;
3021 /* Convert a top-level pet_expr to a pet_scop with one statement.
3022 * This mainly involves resolving nested expression parameters
3023 * and setting the name of the iteration space.
3024 * The name is given by "label" if it is non-NULL. Otherwise,
3025 * it is of the form S_<n_stmt>.
3027 struct pet_scop *PetScan::extract(Stmt *stmt, struct pet_expr *expr,
3028 __isl_take isl_id *label)
3030 struct pet_stmt *ps;
3031 SourceLocation loc = stmt->getLocStart();
3032 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3034 expr = resolve_nested(expr);
3035 ps = pet_stmt_from_pet_expr(ctx, line, label, n_stmt++, expr);
3036 return pet_scop_from_pet_stmt(ctx, ps);
3039 /* Check if we can extract an affine expression from "expr".
3040 * Return the expressions as an isl_pw_aff if we can and NULL otherwise.
3041 * We turn on autodetection so that we won't generate any warnings
3042 * and turn off nesting, so that we won't accept any non-affine constructs.
3044 __isl_give isl_pw_aff *PetScan::try_extract_affine(Expr *expr)
3046 isl_pw_aff *pwaff;
3047 int save_autodetect = options->autodetect;
3048 bool save_nesting = nesting_enabled;
3050 options->autodetect = 1;
3051 nesting_enabled = false;
3053 pwaff = extract_affine(expr);
3055 options->autodetect = save_autodetect;
3056 nesting_enabled = save_nesting;
3058 return pwaff;
3061 /* Check whether "expr" is an affine expression.
3063 bool PetScan::is_affine(Expr *expr)
3065 isl_pw_aff *pwaff;
3067 pwaff = try_extract_affine(expr);
3068 isl_pw_aff_free(pwaff);
3070 return pwaff != NULL;
3073 /* Check if we can extract an affine constraint from "expr".
3074 * Return the constraint as an isl_set if we can and NULL otherwise.
3075 * We turn on autodetection so that we won't generate any warnings
3076 * and turn off nesting, so that we won't accept any non-affine constructs.
3078 __isl_give isl_pw_aff *PetScan::try_extract_affine_condition(Expr *expr)
3080 isl_pw_aff *cond;
3081 int save_autodetect = options->autodetect;
3082 bool save_nesting = nesting_enabled;
3084 options->autodetect = 1;
3085 nesting_enabled = false;
3087 cond = extract_condition(expr);
3089 options->autodetect = save_autodetect;
3090 nesting_enabled = save_nesting;
3092 return cond;
3095 /* Check whether "expr" is an affine constraint.
3097 bool PetScan::is_affine_condition(Expr *expr)
3099 isl_pw_aff *cond;
3101 cond = try_extract_affine_condition(expr);
3102 isl_pw_aff_free(cond);
3104 return cond != NULL;
3107 /* Check if we can extract a condition from "expr".
3108 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
3109 * If allow_nested is set, then the condition may involve parameters
3110 * corresponding to nested accesses.
3111 * We turn on autodetection so that we won't generate any warnings.
3113 __isl_give isl_pw_aff *PetScan::try_extract_nested_condition(Expr *expr)
3115 isl_pw_aff *cond;
3116 int save_autodetect = options->autodetect;
3117 bool save_nesting = nesting_enabled;
3119 options->autodetect = 1;
3120 nesting_enabled = allow_nested;
3121 cond = extract_condition(expr);
3123 options->autodetect = save_autodetect;
3124 nesting_enabled = save_nesting;
3126 return cond;
3129 /* If the top-level expression of "stmt" is an assignment, then
3130 * return that assignment as a BinaryOperator.
3131 * Otherwise return NULL.
3133 static BinaryOperator *top_assignment_or_null(Stmt *stmt)
3135 BinaryOperator *ass;
3137 if (!stmt)
3138 return NULL;
3139 if (stmt->getStmtClass() != Stmt::BinaryOperatorClass)
3140 return NULL;
3142 ass = cast<BinaryOperator>(stmt);
3143 if(ass->getOpcode() != BO_Assign)
3144 return NULL;
3146 return ass;
3149 /* Check if the given if statement is a conditional assignement
3150 * with a non-affine condition. If so, construct a pet_scop
3151 * corresponding to this conditional assignment. Otherwise return NULL.
3153 * In particular we check if "stmt" is of the form
3155 * if (condition)
3156 * a = f(...);
3157 * else
3158 * a = g(...);
3160 * where a is some array or scalar access.
3161 * The constructed pet_scop then corresponds to the expression
3163 * a = condition ? f(...) : g(...)
3165 * All access relations in f(...) are intersected with condition
3166 * while all access relation in g(...) are intersected with the complement.
3168 struct pet_scop *PetScan::extract_conditional_assignment(IfStmt *stmt)
3170 BinaryOperator *ass_then, *ass_else;
3171 isl_map *write_then, *write_else;
3172 isl_set *cond, *comp;
3173 isl_map *map;
3174 isl_pw_aff *pa;
3175 int equal;
3176 struct pet_expr *pe_cond, *pe_then, *pe_else, *pe, *pe_write;
3177 bool save_nesting = nesting_enabled;
3179 if (!options->detect_conditional_assignment)
3180 return NULL;
3182 ass_then = top_assignment_or_null(stmt->getThen());
3183 ass_else = top_assignment_or_null(stmt->getElse());
3185 if (!ass_then || !ass_else)
3186 return NULL;
3188 if (is_affine_condition(stmt->getCond()))
3189 return NULL;
3191 write_then = extract_access(ass_then->getLHS());
3192 write_else = extract_access(ass_else->getLHS());
3194 equal = isl_map_is_equal(write_then, write_else);
3195 isl_map_free(write_else);
3196 if (equal < 0 || !equal) {
3197 isl_map_free(write_then);
3198 return NULL;
3201 nesting_enabled = allow_nested;
3202 pa = extract_condition(stmt->getCond());
3203 nesting_enabled = save_nesting;
3204 cond = isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa));
3205 comp = isl_pw_aff_zero_set(isl_pw_aff_copy(pa));
3206 map = isl_map_from_range(isl_set_from_pw_aff(pa));
3208 pe_cond = pet_expr_from_access(map);
3210 pe_then = extract_expr(ass_then->getRHS());
3211 pe_then = pet_expr_restrict(pe_then, cond);
3212 pe_else = extract_expr(ass_else->getRHS());
3213 pe_else = pet_expr_restrict(pe_else, comp);
3215 pe = pet_expr_new_ternary(ctx, pe_cond, pe_then, pe_else);
3216 pe_write = pet_expr_from_access(write_then);
3217 if (pe_write) {
3218 pe_write->acc.write = 1;
3219 pe_write->acc.read = 0;
3221 pe = pet_expr_new_binary(ctx, pet_op_assign, pe_write, pe);
3222 return extract(stmt, pe);
3225 /* Create a pet_scop with a single statement evaluating "cond"
3226 * and writing the result to a virtual scalar, as expressed by
3227 * "access".
3229 struct pet_scop *PetScan::extract_non_affine_condition(Expr *cond,
3230 __isl_take isl_map *access)
3232 struct pet_expr *expr, *write;
3233 struct pet_stmt *ps;
3234 struct pet_scop *scop;
3235 SourceLocation loc = cond->getLocStart();
3236 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3238 write = pet_expr_from_access(access);
3239 if (write) {
3240 write->acc.write = 1;
3241 write->acc.read = 0;
3243 expr = extract_expr(cond);
3244 expr = resolve_nested(expr);
3245 expr = pet_expr_new_binary(ctx, pet_op_assign, write, expr);
3246 ps = pet_stmt_from_pet_expr(ctx, line, NULL, n_stmt++, expr);
3247 scop = pet_scop_from_pet_stmt(ctx, ps);
3248 scop = resolve_nested(scop);
3250 return scop;
3253 extern "C" {
3254 static __isl_give isl_map *embed_access(__isl_take isl_map *access,
3255 void *user);
3258 /* Apply the map pointed to by "user" to the domain of the access
3259 * relation, thereby embedding it in the range of the map.
3260 * The domain of both relations is the zero-dimensional domain.
3262 static __isl_give isl_map *embed_access(__isl_take isl_map *access, void *user)
3264 isl_map *map = (isl_map *) user;
3266 return isl_map_apply_domain(access, isl_map_copy(map));
3269 /* Apply "map" to all access relations in "expr".
3271 static struct pet_expr *embed(struct pet_expr *expr, __isl_keep isl_map *map)
3273 return pet_expr_foreach_access(expr, &embed_access, map);
3276 /* How many parameters of "set" refer to nested accesses, i.e., have no name?
3278 static int n_nested_parameter(__isl_keep isl_set *set)
3280 isl_space *space;
3281 int n;
3283 space = isl_set_get_space(set);
3284 n = n_nested_parameter(space);
3285 isl_space_free(space);
3287 return n;
3290 /* Remove all parameters from "map" that refer to nested accesses.
3292 static __isl_give isl_map *remove_nested_parameters(__isl_take isl_map *map)
3294 int nparam;
3295 isl_space *space;
3297 space = isl_map_get_space(map);
3298 nparam = isl_space_dim(space, isl_dim_param);
3299 for (int i = nparam - 1; i >= 0; --i)
3300 if (is_nested_parameter(space, i))
3301 map = isl_map_project_out(map, isl_dim_param, i, 1);
3302 isl_space_free(space);
3304 return map;
3307 extern "C" {
3308 static __isl_give isl_map *access_remove_nested_parameters(
3309 __isl_take isl_map *access, void *user);
3312 static __isl_give isl_map *access_remove_nested_parameters(
3313 __isl_take isl_map *access, void *user)
3315 return remove_nested_parameters(access);
3318 /* Remove all nested access parameters from the schedule and all
3319 * accesses of "stmt".
3320 * There is no need to remove them from the domain as these parameters
3321 * have already been removed from the domain when this function is called.
3323 static struct pet_stmt *remove_nested_parameters(struct pet_stmt *stmt)
3325 if (!stmt)
3326 return NULL;
3327 stmt->schedule = remove_nested_parameters(stmt->schedule);
3328 stmt->body = pet_expr_foreach_access(stmt->body,
3329 &access_remove_nested_parameters, NULL);
3330 if (!stmt->schedule || !stmt->body)
3331 goto error;
3332 for (int i = 0; i < stmt->n_arg; ++i) {
3333 stmt->args[i] = pet_expr_foreach_access(stmt->args[i],
3334 &access_remove_nested_parameters, NULL);
3335 if (!stmt->args[i])
3336 goto error;
3339 return stmt;
3340 error:
3341 pet_stmt_free(stmt);
3342 return NULL;
3345 /* For each nested access parameter in the domain of "stmt",
3346 * construct a corresponding pet_expr, place it before the original
3347 * elements in stmt->args and record its position in "param2pos".
3348 * n is the number of nested access parameters.
3350 struct pet_stmt *PetScan::extract_nested(struct pet_stmt *stmt, int n,
3351 std::map<int,int> &param2pos)
3353 int i;
3354 isl_space *space;
3355 int n_arg;
3356 struct pet_expr **args;
3358 n_arg = stmt->n_arg;
3359 args = isl_calloc_array(ctx, struct pet_expr *, n + n_arg);
3360 if (!args)
3361 goto error;
3363 space = isl_set_get_space(stmt->domain);
3364 n_arg = extract_nested(space, 0, args, param2pos);
3365 isl_space_free(space);
3367 if (n_arg < 0)
3368 goto error;
3370 for (i = 0; i < stmt->n_arg; ++i)
3371 args[n_arg + i] = stmt->args[i];
3372 free(stmt->args);
3373 stmt->args = args;
3374 stmt->n_arg += n_arg;
3376 return stmt;
3377 error:
3378 if (args) {
3379 for (i = 0; i < n; ++i)
3380 pet_expr_free(args[i]);
3381 free(args);
3383 pet_stmt_free(stmt);
3384 return NULL;
3387 /* Check whether any of the arguments i of "stmt" starting at position "n"
3388 * is equal to one of the first "n" arguments j.
3389 * If so, combine the constraints on arguments i and j and remove
3390 * argument i.
3392 static struct pet_stmt *remove_duplicate_arguments(struct pet_stmt *stmt, int n)
3394 int i, j;
3395 isl_map *map;
3397 if (!stmt)
3398 return NULL;
3399 if (n == 0)
3400 return stmt;
3401 if (n == stmt->n_arg)
3402 return stmt;
3404 map = isl_set_unwrap(stmt->domain);
3406 for (i = stmt->n_arg - 1; i >= n; --i) {
3407 for (j = 0; j < n; ++j)
3408 if (pet_expr_is_equal(stmt->args[i], stmt->args[j]))
3409 break;
3410 if (j >= n)
3411 continue;
3413 map = isl_map_equate(map, isl_dim_out, i, isl_dim_out, j);
3414 map = isl_map_project_out(map, isl_dim_out, i, 1);
3416 pet_expr_free(stmt->args[i]);
3417 for (j = i; j + 1 < stmt->n_arg; ++j)
3418 stmt->args[j] = stmt->args[j + 1];
3419 stmt->n_arg--;
3422 stmt->domain = isl_map_wrap(map);
3423 if (!stmt->domain)
3424 goto error;
3425 return stmt;
3426 error:
3427 pet_stmt_free(stmt);
3428 return NULL;
3431 /* Look for parameters in the iteration domain of "stmt" that
3432 * refer to nested accesses. In particular, these are
3433 * parameters with no name.
3435 * If there are any such parameters, then as many extra variables
3436 * (after identifying identical nested accesses) are inserted in the
3437 * range of the map wrapped inside the domain, before the original variables.
3438 * If the original domain is not a wrapped map, then a new wrapped
3439 * map is created with zero output dimensions.
3440 * The parameters are then equated to the corresponding output dimensions
3441 * and subsequently projected out, from the iteration domain,
3442 * the schedule and the access relations.
3443 * For each of the output dimensions, a corresponding argument
3444 * expression is inserted. Initially they are created with
3445 * a zero-dimensional domain, so they have to be embedded
3446 * in the current iteration domain.
3447 * param2pos maps the position of the parameter to the position
3448 * of the corresponding output dimension in the wrapped map.
3450 struct pet_stmt *PetScan::resolve_nested(struct pet_stmt *stmt)
3452 int n;
3453 int nparam;
3454 unsigned n_arg;
3455 isl_map *map;
3456 std::map<int,int> param2pos;
3458 if (!stmt)
3459 return NULL;
3461 n = n_nested_parameter(stmt->domain);
3462 if (n == 0)
3463 return stmt;
3465 n_arg = stmt->n_arg;
3466 stmt = extract_nested(stmt, n, param2pos);
3467 if (!stmt)
3468 return NULL;
3470 n = stmt->n_arg - n_arg;
3471 nparam = isl_set_dim(stmt->domain, isl_dim_param);
3472 if (isl_set_is_wrapping(stmt->domain))
3473 map = isl_set_unwrap(stmt->domain);
3474 else
3475 map = isl_map_from_domain(stmt->domain);
3476 map = isl_map_insert_dims(map, isl_dim_out, 0, n);
3478 for (int i = nparam - 1; i >= 0; --i) {
3479 isl_id *id;
3481 if (!is_nested_parameter(map, i))
3482 continue;
3484 id = isl_map_get_tuple_id(stmt->args[param2pos[i]]->acc.access,
3485 isl_dim_out);
3486 map = isl_map_set_dim_id(map, isl_dim_out, param2pos[i], id);
3487 map = isl_map_equate(map, isl_dim_param, i, isl_dim_out,
3488 param2pos[i]);
3489 map = isl_map_project_out(map, isl_dim_param, i, 1);
3492 stmt->domain = isl_map_wrap(map);
3494 map = isl_set_unwrap(isl_set_copy(stmt->domain));
3495 map = isl_map_from_range(isl_map_domain(map));
3496 for (int pos = 0; pos < n; ++pos)
3497 stmt->args[pos] = embed(stmt->args[pos], map);
3498 isl_map_free(map);
3500 stmt = remove_nested_parameters(stmt);
3501 stmt = remove_duplicate_arguments(stmt, n);
3503 return stmt;
3504 error:
3505 pet_stmt_free(stmt);
3506 return NULL;
3509 /* For each statement in "scop", move the parameters that correspond
3510 * to nested access into the ranges of the domains and create
3511 * corresponding argument expressions.
3513 struct pet_scop *PetScan::resolve_nested(struct pet_scop *scop)
3515 if (!scop)
3516 return NULL;
3518 for (int i = 0; i < scop->n_stmt; ++i) {
3519 scop->stmts[i] = resolve_nested(scop->stmts[i]);
3520 if (!scop->stmts[i])
3521 goto error;
3524 return scop;
3525 error:
3526 pet_scop_free(scop);
3527 return NULL;
3530 /* Given an access expression "expr", is the variable accessed by
3531 * "expr" assigned anywhere inside "scop"?
3533 static bool is_assigned(pet_expr *expr, pet_scop *scop)
3535 bool assigned = false;
3536 isl_id *id;
3538 id = isl_map_get_tuple_id(expr->acc.access, isl_dim_out);
3539 assigned = pet_scop_writes(scop, id);
3540 isl_id_free(id);
3542 return assigned;
3545 /* Are all nested access parameters in "pa" allowed given "scop".
3546 * In particular, is none of them written by anywhere inside "scop".
3548 * If "scop" has any skip conditions, then no nested access parameters
3549 * are allowed. In particular, if there is any nested access in a guard
3550 * for a piece of code containing a "continue", then we want to introduce
3551 * a separate statement for evaluating this guard so that we can express
3552 * that the result is false for all previous iterations.
3554 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff *pa, pet_scop *scop)
3556 int nparam;
3558 if (!scop)
3559 return true;
3561 nparam = isl_pw_aff_dim(pa, isl_dim_param);
3562 for (int i = 0; i < nparam; ++i) {
3563 Expr *nested;
3564 isl_id *id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
3565 pet_expr *expr;
3566 bool allowed;
3568 if (!is_nested_parameter(id)) {
3569 isl_id_free(id);
3570 continue;
3573 if (pet_scop_has_skip(scop, pet_skip_now)) {
3574 isl_id_free(id);
3575 return false;
3578 nested = (Expr *) isl_id_get_user(id);
3579 expr = extract_expr(nested);
3580 allowed = expr && expr->type == pet_expr_access &&
3581 !is_assigned(expr, scop);
3583 pet_expr_free(expr);
3584 isl_id_free(id);
3586 if (!allowed)
3587 return false;
3590 return true;
3593 /* Do we need to construct a skip condition of the given type
3594 * on an if statement, given that the if condition is non-affine?
3596 * pet_scop_filter_skip can only handle the case where the if condition
3597 * holds (the then branch) and the skip condition is universal.
3598 * In any other case, we need to construct a new skip condition.
3600 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
3601 bool have_else, enum pet_skip type)
3603 if (have_else && scop_else && pet_scop_has_skip(scop_else, type))
3604 return true;
3605 if (scop_then && pet_scop_has_skip(scop_then, type) &&
3606 !pet_scop_has_universal_skip(scop_then, type))
3607 return true;
3608 return false;
3611 /* Do we need to construct a skip condition of the given type
3612 * on an if statement, given that the if condition is affine?
3614 * There is no need to construct a new skip condition if all
3615 * the skip conditions are affine.
3617 static bool need_skip_aff(struct pet_scop *scop_then,
3618 struct pet_scop *scop_else, bool have_else, enum pet_skip type)
3620 if (scop_then && pet_scop_has_var_skip(scop_then, type))
3621 return true;
3622 if (have_else && scop_else && pet_scop_has_var_skip(scop_else, type))
3623 return true;
3624 return false;
3627 /* Do we need to construct a skip condition of the given type
3628 * on an if statement?
3630 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
3631 bool have_else, enum pet_skip type, bool affine)
3633 if (affine)
3634 return need_skip_aff(scop_then, scop_else, have_else, type);
3635 else
3636 return need_skip(scop_then, scop_else, have_else, type);
3639 /* Construct an affine expression pet_expr that is evaluates
3640 * to the constant "val".
3642 static struct pet_expr *universally(isl_ctx *ctx, int val)
3644 isl_space *space;
3645 isl_map *map;
3647 space = isl_space_alloc(ctx, 0, 0, 1);
3648 map = isl_map_universe(space);
3649 map = isl_map_fix_si(map, isl_dim_out, 0, val);
3651 return pet_expr_from_access(map);
3654 /* Construct an affine expression pet_expr that is evaluates
3655 * to the constant 1.
3657 static struct pet_expr *universally_true(isl_ctx *ctx)
3659 return universally(ctx, 1);
3662 /* Construct an affine expression pet_expr that is evaluates
3663 * to the constant 0.
3665 static struct pet_expr *universally_false(isl_ctx *ctx)
3667 return universally(ctx, 0);
3670 /* Given an access relation "test_access" for the if condition,
3671 * an access relation "skip_access" for the skip condition and
3672 * scops for the then and else branches, construct a scop for
3673 * computing "skip_access".
3675 * The computed scop contains a single statement that essentially does
3677 * skip_cond = test_cond ? skip_cond_then : skip_cond_else
3679 * If the skip conditions of the then and/or else branch are not affine,
3680 * then they need to be filtered by test_access.
3681 * If they are missing, then this means the skip condition is false.
3683 * Since we are constructing a skip condition for the if statement,
3684 * the skip conditions on the then and else branches are removed.
3686 static struct pet_scop *extract_skip(PetScan *scan,
3687 __isl_take isl_map *test_access, __isl_take isl_map *skip_access,
3688 struct pet_scop *scop_then, struct pet_scop *scop_else, bool have_else,
3689 enum pet_skip type)
3691 struct pet_expr *expr_then, *expr_else, *expr, *expr_skip;
3692 struct pet_stmt *stmt;
3693 struct pet_scop *scop;
3694 isl_ctx *ctx = scan->ctx;
3696 if (!scop_then)
3697 goto error;
3698 if (have_else && !scop_else)
3699 goto error;
3701 if (pet_scop_has_skip(scop_then, type)) {
3702 expr_then = pet_scop_get_skip_expr(scop_then, type);
3703 pet_scop_reset_skip(scop_then, type);
3704 if (!pet_expr_is_affine(expr_then))
3705 expr_then = pet_expr_filter(expr_then,
3706 isl_map_copy(test_access), 1);
3707 } else
3708 expr_then = universally_false(ctx);
3710 if (have_else && pet_scop_has_skip(scop_else, type)) {
3711 expr_else = pet_scop_get_skip_expr(scop_else, type);
3712 pet_scop_reset_skip(scop_else, type);
3713 if (!pet_expr_is_affine(expr_else))
3714 expr_else = pet_expr_filter(expr_else,
3715 isl_map_copy(test_access), 0);
3716 } else
3717 expr_else = universally_false(ctx);
3719 expr = pet_expr_from_access(test_access);
3720 expr = pet_expr_new_ternary(ctx, expr, expr_then, expr_else);
3721 expr_skip = pet_expr_from_access(isl_map_copy(skip_access));
3722 if (expr_skip) {
3723 expr_skip->acc.write = 1;
3724 expr_skip->acc.read = 0;
3726 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
3727 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, scan->n_stmt++, expr);
3729 scop = pet_scop_from_pet_stmt(ctx, stmt);
3730 scop = scop_add_array(scop, skip_access, scan->ast_context);
3731 isl_map_free(skip_access);
3733 return scop;
3734 error:
3735 isl_map_free(test_access);
3736 isl_map_free(skip_access);
3737 return NULL;
3740 /* Is scop's skip_now condition equal to its skip_later condition?
3741 * In particular, this means that it either has no skip_now condition
3742 * or both a skip_now and a skip_later condition (that are equal to each other).
3744 static bool skip_equals_skip_later(struct pet_scop *scop)
3746 int has_skip_now, has_skip_later;
3747 int equal;
3748 isl_set *skip_now, *skip_later;
3750 if (!scop)
3751 return false;
3752 has_skip_now = pet_scop_has_skip(scop, pet_skip_now);
3753 has_skip_later = pet_scop_has_skip(scop, pet_skip_later);
3754 if (has_skip_now != has_skip_later)
3755 return false;
3756 if (!has_skip_now)
3757 return true;
3759 skip_now = pet_scop_get_skip(scop, pet_skip_now);
3760 skip_later = pet_scop_get_skip(scop, pet_skip_later);
3761 equal = isl_set_is_equal(skip_now, skip_later);
3762 isl_set_free(skip_now);
3763 isl_set_free(skip_later);
3765 return equal;
3768 /* Drop the skip conditions of type pet_skip_later from scop1 and scop2.
3770 static void drop_skip_later(struct pet_scop *scop1, struct pet_scop *scop2)
3772 pet_scop_reset_skip(scop1, pet_skip_later);
3773 pet_scop_reset_skip(scop2, pet_skip_later);
3776 /* Structure that handles the construction of skip conditions.
3778 * scop_then and scop_else represent the then and else branches
3779 * of the if statement
3781 * skip[type] is true if we need to construct a skip condition of that type
3782 * equal is set if the skip conditions of types pet_skip_now and pet_skip_later
3783 * are equal to each other
3784 * access[type] is the virtual array representing the skip condition
3785 * scop[type] is a scop for computing the skip condition
3787 struct pet_skip_info {
3788 isl_ctx *ctx;
3790 bool skip[2];
3791 bool equal;
3792 isl_map *access[2];
3793 struct pet_scop *scop[2];
3795 pet_skip_info(isl_ctx *ctx) : ctx(ctx) {}
3797 operator bool() { return skip[pet_skip_now] || skip[pet_skip_later]; }
3800 /* Structure that handles the construction of skip conditions on if statements.
3802 * scop_then and scop_else represent the then and else branches
3803 * of the if statement
3805 struct pet_skip_info_if : public pet_skip_info {
3806 struct pet_scop *scop_then, *scop_else;
3807 bool have_else;
3809 pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
3810 struct pet_scop *scop_else, bool have_else, bool affine);
3811 void extract(PetScan *scan, __isl_keep isl_map *access,
3812 enum pet_skip type);
3813 void extract(PetScan *scan, __isl_keep isl_map *access);
3814 void extract(PetScan *scan, __isl_keep isl_pw_aff *cond);
3815 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
3816 int offset);
3817 struct pet_scop *add(struct pet_scop *scop, int offset);
3820 /* Initialize a pet_skip_info_if structure based on the then and else branches
3821 * and based on whether the if condition is affine or not.
3823 pet_skip_info_if::pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
3824 struct pet_scop *scop_else, bool have_else, bool affine) :
3825 pet_skip_info(ctx), scop_then(scop_then), scop_else(scop_else),
3826 have_else(have_else)
3828 skip[pet_skip_now] =
3829 need_skip(scop_then, scop_else, have_else, pet_skip_now, affine);
3830 equal = skip[pet_skip_now] && skip_equals_skip_later(scop_then) &&
3831 (!have_else || skip_equals_skip_later(scop_else));
3832 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
3833 need_skip(scop_then, scop_else, have_else, pet_skip_later, affine);
3836 /* If we need to construct a skip condition of the given type,
3837 * then do so now.
3839 * "map" represents the if condition.
3841 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_map *map,
3842 enum pet_skip type)
3844 if (!skip[type])
3845 return;
3847 access[type] = create_test_access(isl_map_get_ctx(map), scan->n_test++);
3848 scop[type] = extract_skip(scan, isl_map_copy(map),
3849 isl_map_copy(access[type]),
3850 scop_then, scop_else, have_else, type);
3853 /* Construct the required skip conditions, given the if condition "map".
3855 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_map *map)
3857 extract(scan, map, pet_skip_now);
3858 extract(scan, map, pet_skip_later);
3859 if (equal)
3860 drop_skip_later(scop_then, scop_else);
3863 /* Construct the required skip conditions, given the if condition "cond".
3865 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_pw_aff *cond)
3867 isl_set *test_set;
3868 isl_map *test;
3870 if (!skip[pet_skip_now] && !skip[pet_skip_later])
3871 return;
3873 test_set = isl_set_from_pw_aff(isl_pw_aff_copy(cond));
3874 test = isl_map_from_range(test_set);
3875 extract(scan, test);
3876 isl_map_free(test);
3879 /* Add the computed skip condition of the give type to "main" and
3880 * add the scop for computing the condition at the given offset.
3882 * If equal is set, then we only computed a skip condition for pet_skip_now,
3883 * but we also need to set it as main's pet_skip_later.
3885 struct pet_scop *pet_skip_info_if::add(struct pet_scop *main,
3886 enum pet_skip type, int offset)
3888 isl_set *skip_set;
3890 if (!skip[type])
3891 return main;
3893 skip_set = isl_map_range(access[type]);
3894 access[type] = NULL;
3895 scop[type] = pet_scop_prefix(scop[type], offset);
3896 main = pet_scop_add_par(ctx, main, scop[type]);
3897 scop[type] = NULL;
3899 if (equal)
3900 main = pet_scop_set_skip(main, pet_skip_later,
3901 isl_set_copy(skip_set));
3903 main = pet_scop_set_skip(main, type, skip_set);
3905 return main;
3908 /* Add the computed skip conditions to "main" and
3909 * add the scops for computing the conditions at the given offset.
3911 struct pet_scop *pet_skip_info_if::add(struct pet_scop *scop, int offset)
3913 scop = add(scop, pet_skip_now, offset);
3914 scop = add(scop, pet_skip_later, offset);
3916 return scop;
3919 /* Construct a pet_scop for a non-affine if statement.
3921 * We create a separate statement that writes the result
3922 * of the non-affine condition to a virtual scalar.
3923 * A constraint requiring the value of this virtual scalar to be one
3924 * is added to the iteration domains of the then branch.
3925 * Similarly, a constraint requiring the value of this virtual scalar
3926 * to be zero is added to the iteration domains of the else branch, if any.
3927 * We adjust the schedules to ensure that the virtual scalar is written
3928 * before it is read.
3930 * If there are any breaks or continues in the then and/or else
3931 * branches, then we may have to compute a new skip condition.
3932 * This is handled using a pet_skip_info_if object.
3933 * On initialization, the object checks if skip conditions need
3934 * to be computed. If so, it does so in "extract" and adds them in "add".
3936 struct pet_scop *PetScan::extract_non_affine_if(Expr *cond,
3937 struct pet_scop *scop_then, struct pet_scop *scop_else,
3938 bool have_else, int stmt_id)
3940 struct pet_scop *scop;
3941 isl_map *test_access;
3942 int save_n_stmt = n_stmt;
3944 test_access = create_test_access(ctx, n_test++);
3945 n_stmt = stmt_id;
3946 scop = extract_non_affine_condition(cond, isl_map_copy(test_access));
3947 n_stmt = save_n_stmt;
3948 scop = scop_add_array(scop, test_access, ast_context);
3950 pet_skip_info_if skip(ctx, scop_then, scop_else, have_else, false);
3951 skip.extract(this, test_access);
3953 scop = pet_scop_prefix(scop, 0);
3954 scop_then = pet_scop_prefix(scop_then, 1);
3955 scop_then = pet_scop_filter(scop_then, isl_map_copy(test_access), 1);
3956 if (have_else) {
3957 scop_else = pet_scop_prefix(scop_else, 1);
3958 scop_else = pet_scop_filter(scop_else, test_access, 0);
3959 scop_then = pet_scop_add_par(ctx, scop_then, scop_else);
3960 } else
3961 isl_map_free(test_access);
3963 scop = pet_scop_add_seq(ctx, scop, scop_then);
3965 scop = skip.add(scop, 2);
3967 return scop;
3970 /* Construct a pet_scop for an if statement.
3972 * If the condition fits the pattern of a conditional assignment,
3973 * then it is handled by extract_conditional_assignment.
3974 * Otherwise, we do the following.
3976 * If the condition is affine, then the condition is added
3977 * to the iteration domains of the then branch, while the
3978 * opposite of the condition in added to the iteration domains
3979 * of the else branch, if any.
3980 * We allow the condition to be dynamic, i.e., to refer to
3981 * scalars or array elements that may be written to outside
3982 * of the given if statement. These nested accesses are then represented
3983 * as output dimensions in the wrapping iteration domain.
3984 * If it also written _inside_ the then or else branch, then
3985 * we treat the condition as non-affine.
3986 * As explained in extract_non_affine_if, this will introduce
3987 * an extra statement.
3988 * For aesthetic reasons, we want this statement to have a statement
3989 * number that is lower than those of the then and else branches.
3990 * In order to evaluate if will need such a statement, however, we
3991 * first construct scops for the then and else branches.
3992 * We therefore reserve a statement number if we might have to
3993 * introduce such an extra statement.
3995 * If the condition is not affine, then the scop is created in
3996 * extract_non_affine_if.
3998 * If there are any breaks or continues in the then and/or else
3999 * branches, then we may have to compute a new skip condition.
4000 * This is handled using a pet_skip_info_if object.
4001 * On initialization, the object checks if skip conditions need
4002 * to be computed. If so, it does so in "extract" and adds them in "add".
4004 struct pet_scop *PetScan::extract(IfStmt *stmt)
4006 struct pet_scop *scop_then, *scop_else = NULL, *scop;
4007 isl_pw_aff *cond;
4008 int stmt_id;
4009 isl_set *set;
4010 isl_set *valid;
4012 scop = extract_conditional_assignment(stmt);
4013 if (scop)
4014 return scop;
4016 cond = try_extract_nested_condition(stmt->getCond());
4017 if (allow_nested && (!cond || has_nested(cond)))
4018 stmt_id = n_stmt++;
4021 assigned_value_cache cache(assigned_value);
4022 scop_then = extract(stmt->getThen());
4025 if (stmt->getElse()) {
4026 assigned_value_cache cache(assigned_value);
4027 scop_else = extract(stmt->getElse());
4028 if (options->autodetect) {
4029 if (scop_then && !scop_else) {
4030 partial = true;
4031 isl_pw_aff_free(cond);
4032 return scop_then;
4034 if (!scop_then && scop_else) {
4035 partial = true;
4036 isl_pw_aff_free(cond);
4037 return scop_else;
4042 if (cond &&
4043 (!is_nested_allowed(cond, scop_then) ||
4044 (stmt->getElse() && !is_nested_allowed(cond, scop_else)))) {
4045 isl_pw_aff_free(cond);
4046 cond = NULL;
4048 if (allow_nested && !cond)
4049 return extract_non_affine_if(stmt->getCond(), scop_then,
4050 scop_else, stmt->getElse(), stmt_id);
4052 if (!cond)
4053 cond = extract_condition(stmt->getCond());
4055 pet_skip_info_if skip(ctx, scop_then, scop_else, stmt->getElse(), true);
4056 skip.extract(this, cond);
4058 valid = isl_pw_aff_domain(isl_pw_aff_copy(cond));
4059 set = isl_pw_aff_non_zero_set(cond);
4060 scop = pet_scop_restrict(scop_then, isl_set_copy(set));
4062 if (stmt->getElse()) {
4063 set = isl_set_subtract(isl_set_copy(valid), set);
4064 scop_else = pet_scop_restrict(scop_else, set);
4065 scop = pet_scop_add_par(ctx, scop, scop_else);
4066 } else
4067 isl_set_free(set);
4068 scop = resolve_nested(scop);
4069 scop = pet_scop_restrict_context(scop, valid);
4071 if (skip)
4072 scop = pet_scop_prefix(scop, 0);
4073 scop = skip.add(scop, 1);
4075 return scop;
4078 /* Try and construct a pet_scop for a label statement.
4079 * We currently only allow labels on expression statements.
4081 struct pet_scop *PetScan::extract(LabelStmt *stmt)
4083 isl_id *label;
4084 Stmt *sub;
4086 sub = stmt->getSubStmt();
4087 if (!isa<Expr>(sub)) {
4088 unsupported(stmt);
4089 return NULL;
4092 label = isl_id_alloc(ctx, stmt->getName(), NULL);
4094 return extract(sub, extract_expr(cast<Expr>(sub)), label);
4097 /* Construct a pet_scop for a continue statement.
4099 * We simply create an empty scop with a universal pet_skip_now
4100 * skip condition. This skip condition will then be taken into
4101 * account by the enclosing loop construct, possibly after
4102 * being incorporated into outer skip conditions.
4104 struct pet_scop *PetScan::extract(ContinueStmt *stmt)
4106 pet_scop *scop;
4107 isl_space *space;
4108 isl_set *set;
4110 scop = pet_scop_empty(ctx);
4111 if (!scop)
4112 return NULL;
4114 space = isl_space_set_alloc(ctx, 0, 1);
4115 set = isl_set_universe(space);
4116 set = isl_set_fix_si(set, isl_dim_set, 0, 1);
4117 scop = pet_scop_set_skip(scop, pet_skip_now, set);
4119 return scop;
4122 /* Construct a pet_scop for a break statement.
4124 * We simply create an empty scop with both a universal pet_skip_now
4125 * skip condition and a universal pet_skip_later skip condition.
4126 * These skip conditions will then be taken into
4127 * account by the enclosing loop construct, possibly after
4128 * being incorporated into outer skip conditions.
4130 struct pet_scop *PetScan::extract(BreakStmt *stmt)
4132 pet_scop *scop;
4133 isl_space *space;
4134 isl_set *set;
4136 scop = pet_scop_empty(ctx);
4137 if (!scop)
4138 return NULL;
4140 space = isl_space_set_alloc(ctx, 0, 1);
4141 set = isl_set_universe(space);
4142 set = isl_set_fix_si(set, isl_dim_set, 0, 1);
4143 scop = pet_scop_set_skip(scop, pet_skip_now, isl_set_copy(set));
4144 scop = pet_scop_set_skip(scop, pet_skip_later, set);
4146 return scop;
4149 /* Try and construct a pet_scop corresponding to "stmt".
4151 struct pet_scop *PetScan::extract(Stmt *stmt)
4153 if (isa<Expr>(stmt))
4154 return extract(stmt, extract_expr(cast<Expr>(stmt)));
4156 switch (stmt->getStmtClass()) {
4157 case Stmt::WhileStmtClass:
4158 return extract(cast<WhileStmt>(stmt));
4159 case Stmt::ForStmtClass:
4160 return extract_for(cast<ForStmt>(stmt));
4161 case Stmt::IfStmtClass:
4162 return extract(cast<IfStmt>(stmt));
4163 case Stmt::CompoundStmtClass:
4164 return extract(cast<CompoundStmt>(stmt));
4165 case Stmt::LabelStmtClass:
4166 return extract(cast<LabelStmt>(stmt));
4167 case Stmt::ContinueStmtClass:
4168 return extract(cast<ContinueStmt>(stmt));
4169 case Stmt::BreakStmtClass:
4170 return extract(cast<BreakStmt>(stmt));
4171 default:
4172 unsupported(stmt);
4175 return NULL;
4178 /* Do we need to construct a skip condition of the given type
4179 * on a sequence of statements?
4181 * There is no need to construct a new skip condition if only
4182 * only of the two statements has a skip condition or if both
4183 * of their skip conditions are affine.
4185 * In principle we also don't need a new continuation variable if
4186 * the continuation of scop2 is affine, but then we would need
4187 * to allow more complicated forms of continuations.
4189 static bool need_skip_seq(struct pet_scop *scop1, struct pet_scop *scop2,
4190 enum pet_skip type)
4192 if (!scop1 || !pet_scop_has_skip(scop1, type))
4193 return false;
4194 if (!scop2 || !pet_scop_has_skip(scop2, type))
4195 return false;
4196 if (pet_scop_has_affine_skip(scop1, type) &&
4197 pet_scop_has_affine_skip(scop2, type))
4198 return false;
4199 return true;
4202 /* Construct a scop for computing the skip condition of the given type and
4203 * with access relation "skip_access" for a sequence of two scops "scop1"
4204 * and "scop2".
4206 * The computed scop contains a single statement that essentially does
4208 * skip_cond = skip_cond_1 ? 1 : skip_cond_2
4210 * or, in other words, skip_cond1 || skip_cond2.
4211 * In this expression, skip_cond_2 is filtered to reflect that it is
4212 * only evaluated when skip_cond_1 is false.
4214 * The skip condition on scop1 is not removed because it still needs
4215 * to be applied to scop2 when these two scops are combined.
4217 static struct pet_scop *extract_skip_seq(PetScan *ps,
4218 __isl_take isl_map *skip_access,
4219 struct pet_scop *scop1, struct pet_scop *scop2, enum pet_skip type)
4221 isl_map *access;
4222 struct pet_expr *expr1, *expr2, *expr, *expr_skip;
4223 struct pet_stmt *stmt;
4224 struct pet_scop *scop;
4225 isl_ctx *ctx = ps->ctx;
4227 if (!scop1 || !scop2)
4228 goto error;
4230 expr1 = pet_scop_get_skip_expr(scop1, type);
4231 expr2 = pet_scop_get_skip_expr(scop2, type);
4232 pet_scop_reset_skip(scop2, type);
4234 expr2 = pet_expr_filter(expr2, isl_map_copy(expr1->acc.access), 0);
4236 expr = universally_true(ctx);
4237 expr = pet_expr_new_ternary(ctx, expr1, expr, expr2);
4238 expr_skip = pet_expr_from_access(isl_map_copy(skip_access));
4239 if (expr_skip) {
4240 expr_skip->acc.write = 1;
4241 expr_skip->acc.read = 0;
4243 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
4244 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, ps->n_stmt++, expr);
4246 scop = pet_scop_from_pet_stmt(ctx, stmt);
4247 scop = scop_add_array(scop, skip_access, ps->ast_context);
4248 isl_map_free(skip_access);
4250 return scop;
4251 error:
4252 isl_map_free(skip_access);
4253 return NULL;
4256 /* Structure that handles the construction of skip conditions
4257 * on sequences of statements.
4259 * scop1 and scop2 represent the two statements that are combined
4261 struct pet_skip_info_seq : public pet_skip_info {
4262 struct pet_scop *scop1, *scop2;
4264 pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4265 struct pet_scop *scop2);
4266 void extract(PetScan *scan, enum pet_skip type);
4267 void extract(PetScan *scan);
4268 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
4269 int offset);
4270 struct pet_scop *add(struct pet_scop *scop, int offset);
4273 /* Initialize a pet_skip_info_seq structure based on
4274 * on the two statements that are going to be combined.
4276 pet_skip_info_seq::pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4277 struct pet_scop *scop2) : pet_skip_info(ctx), scop1(scop1), scop2(scop2)
4279 skip[pet_skip_now] = need_skip_seq(scop1, scop2, pet_skip_now);
4280 equal = skip[pet_skip_now] && skip_equals_skip_later(scop1) &&
4281 skip_equals_skip_later(scop2);
4282 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
4283 need_skip_seq(scop1, scop2, pet_skip_later);
4286 /* If we need to construct a skip condition of the given type,
4287 * then do so now.
4289 void pet_skip_info_seq::extract(PetScan *scan, enum pet_skip type)
4291 if (!skip[type])
4292 return;
4294 access[type] = create_test_access(ctx, scan->n_test++);
4295 scop[type] = extract_skip_seq(scan, isl_map_copy(access[type]),
4296 scop1, scop2, type);
4299 /* Construct the required skip conditions.
4301 void pet_skip_info_seq::extract(PetScan *scan)
4303 extract(scan, pet_skip_now);
4304 extract(scan, pet_skip_later);
4305 if (equal)
4306 drop_skip_later(scop1, scop2);
4309 /* Add the computed skip condition of the give type to "main" and
4310 * add the scop for computing the condition at the given offset (the statement
4311 * number). Within this offset, the condition is computed at position 1
4312 * to ensure that it is computed after the corresponding statement.
4314 * If equal is set, then we only computed a skip condition for pet_skip_now,
4315 * but we also need to set it as main's pet_skip_later.
4317 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *main,
4318 enum pet_skip type, int offset)
4320 isl_set *skip_set;
4322 if (!skip[type])
4323 return main;
4325 skip_set = isl_map_range(access[type]);
4326 access[type] = NULL;
4327 scop[type] = pet_scop_prefix(scop[type], 1);
4328 scop[type] = pet_scop_prefix(scop[type], offset);
4329 main = pet_scop_add_par(ctx, main, scop[type]);
4330 scop[type] = NULL;
4332 if (equal)
4333 main = pet_scop_set_skip(main, pet_skip_later,
4334 isl_set_copy(skip_set));
4336 main = pet_scop_set_skip(main, type, skip_set);
4338 return main;
4341 /* Add the computed skip conditions to "main" and
4342 * add the scops for computing the conditions at the given offset.
4344 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *scop, int offset)
4346 scop = add(scop, pet_skip_now, offset);
4347 scop = add(scop, pet_skip_later, offset);
4349 return scop;
4352 /* Try and construct a pet_scop corresponding to (part of)
4353 * a sequence of statements.
4355 * If there are any breaks or continues in the individual statements,
4356 * then we may have to compute a new skip condition.
4357 * This is handled using a pet_skip_info_seq object.
4358 * On initialization, the object checks if skip conditions need
4359 * to be computed. If so, it does so in "extract" and adds them in "add".
4361 struct pet_scop *PetScan::extract(StmtRange stmt_range)
4363 pet_scop *scop;
4364 StmtIterator i;
4365 int j;
4366 bool partial_range = false;
4368 scop = pet_scop_empty(ctx);
4369 for (i = stmt_range.first, j = 0; i != stmt_range.second; ++i, ++j) {
4370 Stmt *child = *i;
4371 struct pet_scop *scop_i;
4373 scop_i = extract(child);
4374 if (scop && partial) {
4375 pet_scop_free(scop_i);
4376 break;
4378 pet_skip_info_seq skip(ctx, scop, scop_i);
4379 skip.extract(this);
4380 if (skip)
4381 scop_i = pet_scop_prefix(scop_i, 0);
4382 scop_i = pet_scop_prefix(scop_i, j);
4383 if (options->autodetect) {
4384 if (scop_i)
4385 scop = pet_scop_add_seq(ctx, scop, scop_i);
4386 else
4387 partial_range = true;
4388 if (scop->n_stmt != 0 && !scop_i)
4389 partial = true;
4390 } else {
4391 scop = pet_scop_add_seq(ctx, scop, scop_i);
4394 scop = skip.add(scop, j);
4396 if (partial)
4397 break;
4400 if (scop && partial_range)
4401 partial = true;
4403 return scop;
4406 /* Check if the scop marked by the user is exactly this Stmt
4407 * or part of this Stmt.
4408 * If so, return a pet_scop corresponding to the marked region.
4409 * Otherwise, return NULL.
4411 struct pet_scop *PetScan::scan(Stmt *stmt)
4413 SourceManager &SM = PP.getSourceManager();
4414 unsigned start_off, end_off;
4416 start_off = SM.getFileOffset(stmt->getLocStart());
4417 end_off = SM.getFileOffset(stmt->getLocEnd());
4419 if (start_off > loc.end)
4420 return NULL;
4421 if (end_off < loc.start)
4422 return NULL;
4423 if (start_off >= loc.start && end_off <= loc.end) {
4424 return extract(stmt);
4427 StmtIterator start;
4428 for (start = stmt->child_begin(); start != stmt->child_end(); ++start) {
4429 Stmt *child = *start;
4430 if (!child)
4431 continue;
4432 start_off = SM.getFileOffset(child->getLocStart());
4433 end_off = SM.getFileOffset(child->getLocEnd());
4434 if (start_off < loc.start && end_off > loc.end)
4435 return scan(child);
4436 if (start_off >= loc.start)
4437 break;
4440 StmtIterator end;
4441 for (end = start; end != stmt->child_end(); ++end) {
4442 Stmt *child = *end;
4443 start_off = SM.getFileOffset(child->getLocStart());
4444 if (start_off >= loc.end)
4445 break;
4448 return extract(StmtRange(start, end));
4451 /* Set the size of index "pos" of "array" to "size".
4452 * In particular, add a constraint of the form
4454 * i_pos < size
4456 * to array->extent and a constraint of the form
4458 * size >= 0
4460 * to array->context.
4462 static struct pet_array *update_size(struct pet_array *array, int pos,
4463 __isl_take isl_pw_aff *size)
4465 isl_set *valid;
4466 isl_set *univ;
4467 isl_set *bound;
4468 isl_space *dim;
4469 isl_aff *aff;
4470 isl_pw_aff *index;
4471 isl_id *id;
4473 valid = isl_pw_aff_nonneg_set(isl_pw_aff_copy(size));
4474 array->context = isl_set_intersect(array->context, valid);
4476 dim = isl_set_get_space(array->extent);
4477 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
4478 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, pos, 1);
4479 univ = isl_set_universe(isl_aff_get_domain_space(aff));
4480 index = isl_pw_aff_alloc(univ, aff);
4482 size = isl_pw_aff_add_dims(size, isl_dim_in,
4483 isl_set_dim(array->extent, isl_dim_set));
4484 id = isl_set_get_tuple_id(array->extent);
4485 size = isl_pw_aff_set_tuple_id(size, isl_dim_in, id);
4486 bound = isl_pw_aff_lt_set(index, size);
4488 array->extent = isl_set_intersect(array->extent, bound);
4490 if (!array->context || !array->extent)
4491 goto error;
4493 return array;
4494 error:
4495 pet_array_free(array);
4496 return NULL;
4499 /* Figure out the size of the array at position "pos" and all
4500 * subsequent positions from "type" and update "array" accordingly.
4502 struct pet_array *PetScan::set_upper_bounds(struct pet_array *array,
4503 const Type *type, int pos)
4505 const ArrayType *atype;
4506 isl_pw_aff *size;
4508 if (!array)
4509 return NULL;
4511 if (type->isPointerType()) {
4512 type = type->getPointeeType().getTypePtr();
4513 return set_upper_bounds(array, type, pos + 1);
4515 if (!type->isArrayType())
4516 return array;
4518 type = type->getCanonicalTypeInternal().getTypePtr();
4519 atype = cast<ArrayType>(type);
4521 if (type->isConstantArrayType()) {
4522 const ConstantArrayType *ca = cast<ConstantArrayType>(atype);
4523 size = extract_affine(ca->getSize());
4524 array = update_size(array, pos, size);
4525 } else if (type->isVariableArrayType()) {
4526 const VariableArrayType *vla = cast<VariableArrayType>(atype);
4527 size = extract_affine(vla->getSizeExpr());
4528 array = update_size(array, pos, size);
4531 type = atype->getElementType().getTypePtr();
4533 return set_upper_bounds(array, type, pos + 1);
4536 /* Construct and return a pet_array corresponding to the variable "decl".
4537 * In particular, initialize array->extent to
4539 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
4541 * and then call set_upper_bounds to set the upper bounds on the indices
4542 * based on the type of the variable.
4544 struct pet_array *PetScan::extract_array(isl_ctx *ctx, ValueDecl *decl)
4546 struct pet_array *array;
4547 QualType qt = decl->getType();
4548 const Type *type = qt.getTypePtr();
4549 int depth = array_depth(type);
4550 QualType base = base_type(qt);
4551 string name;
4552 isl_id *id;
4553 isl_space *dim;
4555 array = isl_calloc_type(ctx, struct pet_array);
4556 if (!array)
4557 return NULL;
4559 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
4560 dim = isl_space_set_alloc(ctx, 0, depth);
4561 dim = isl_space_set_tuple_id(dim, isl_dim_set, id);
4563 array->extent = isl_set_nat_universe(dim);
4565 dim = isl_space_params_alloc(ctx, 0);
4566 array->context = isl_set_universe(dim);
4568 array = set_upper_bounds(array, type, 0);
4569 if (!array)
4570 return NULL;
4572 name = base.getAsString();
4573 array->element_type = strdup(name.c_str());
4574 array->element_size = decl->getASTContext().getTypeInfo(base).first / 8;
4576 return array;
4579 /* Construct a list of pet_arrays, one for each array (or scalar)
4580 * accessed inside "scop", add this list to "scop" and return the result.
4582 * The context of "scop" is updated with the intersection of
4583 * the contexts of all arrays, i.e., constraints on the parameters
4584 * that ensure that the arrays have a valid (non-negative) size.
4586 struct pet_scop *PetScan::scan_arrays(struct pet_scop *scop)
4588 int i;
4589 set<ValueDecl *> arrays;
4590 set<ValueDecl *>::iterator it;
4591 int n_array;
4592 struct pet_array **scop_arrays;
4594 if (!scop)
4595 return NULL;
4597 pet_scop_collect_arrays(scop, arrays);
4598 if (arrays.size() == 0)
4599 return scop;
4601 n_array = scop->n_array;
4603 scop_arrays = isl_realloc_array(ctx, scop->arrays, struct pet_array *,
4604 n_array + arrays.size());
4605 if (!scop_arrays)
4606 goto error;
4607 scop->arrays = scop_arrays;
4609 for (it = arrays.begin(), i = 0; it != arrays.end(); ++it, ++i) {
4610 struct pet_array *array;
4611 scop->arrays[n_array + i] = array = extract_array(ctx, *it);
4612 if (!scop->arrays[n_array + i])
4613 goto error;
4614 scop->n_array++;
4615 scop->context = isl_set_intersect(scop->context,
4616 isl_set_copy(array->context));
4617 if (!scop->context)
4618 goto error;
4621 return scop;
4622 error:
4623 pet_scop_free(scop);
4624 return NULL;
4627 /* Bound all parameters in scop->context to the possible values
4628 * of the corresponding C variable.
4630 static struct pet_scop *add_parameter_bounds(struct pet_scop *scop)
4632 int n;
4634 if (!scop)
4635 return NULL;
4637 n = isl_set_dim(scop->context, isl_dim_param);
4638 for (int i = 0; i < n; ++i) {
4639 isl_id *id;
4640 ValueDecl *decl;
4642 id = isl_set_get_dim_id(scop->context, isl_dim_param, i);
4643 if (is_nested_parameter(id)) {
4644 isl_id_free(id);
4645 isl_die(isl_set_get_ctx(scop->context),
4646 isl_error_internal,
4647 "unresolved nested parameter", goto error);
4649 decl = (ValueDecl *) isl_id_get_user(id);
4650 isl_id_free(id);
4652 scop->context = set_parameter_bounds(scop->context, i, decl);
4654 if (!scop->context)
4655 goto error;
4658 return scop;
4659 error:
4660 pet_scop_free(scop);
4661 return NULL;
4664 /* Construct a pet_scop from the given function.
4666 struct pet_scop *PetScan::scan(FunctionDecl *fd)
4668 pet_scop *scop;
4669 Stmt *stmt;
4671 stmt = fd->getBody();
4673 if (options->autodetect)
4674 scop = extract(stmt);
4675 else
4676 scop = scan(stmt);
4677 scop = pet_scop_detect_parameter_accesses(scop);
4678 scop = scan_arrays(scop);
4679 scop = add_parameter_bounds(scop);
4680 scop = pet_scop_gist(scop, value_bounds);
4682 return scop;