update isl for __isl_null memory management annotation
[pet.git] / scan.cc
blob49780e31875ce767d00a5f207c4897b01616d3e6
1 /*
2 * Copyright 2011 Leiden University. All rights reserved.
3 * Copyright 2012-2014 Ecole Normale Superieure. All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 *
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above
13 * copyright notice, this list of conditions and the following
14 * disclaimer in the documentation and/or other materials provided
15 * with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY LEIDEN UNIVERSITY ''AS IS'' AND ANY
18 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL LEIDEN UNIVERSITY OR
21 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
22 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
23 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA,
24 * OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
27 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 * The views and conclusions contained in the software and documentation
30 * are those of the authors and should not be interpreted as
31 * representing official policies, either expressed or implied, of
32 * Leiden University.
33 */
35 #include <string.h>
36 #include <set>
37 #include <map>
38 #include <iostream>
39 #include <llvm/Support/raw_ostream.h>
40 #include <clang/AST/ASTContext.h>
41 #include <clang/AST/ASTDiagnostic.h>
42 #include <clang/AST/Expr.h>
43 #include <clang/AST/RecursiveASTVisitor.h>
45 #include <isl/id.h>
46 #include <isl/space.h>
47 #include <isl/aff.h>
48 #include <isl/set.h>
50 #include "clang.h"
51 #include "expr.h"
52 #include "nest.h"
53 #include "options.h"
54 #include "scan.h"
55 #include "scop.h"
56 #include "scop_plus.h"
58 #include "config.h"
60 using namespace std;
61 using namespace clang;
63 #if defined(DECLREFEXPR_CREATE_REQUIRES_BOOL)
64 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
66 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
67 SourceLocation(), var, false, var->getInnerLocStart(),
68 var->getType(), VK_LValue);
70 #elif defined(DECLREFEXPR_CREATE_REQUIRES_SOURCELOCATION)
71 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
73 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
74 SourceLocation(), var, var->getInnerLocStart(), var->getType(),
75 VK_LValue);
77 #else
78 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
80 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
81 var, var->getInnerLocStart(), var->getType(), VK_LValue);
83 #endif
85 /* Check if the element type corresponding to the given array type
86 * has a const qualifier.
88 static bool const_base(QualType qt)
90 const Type *type = qt.getTypePtr();
92 if (type->isPointerType())
93 return const_base(type->getPointeeType());
94 if (type->isArrayType()) {
95 const ArrayType *atype;
96 type = type->getCanonicalTypeInternal().getTypePtr();
97 atype = cast<ArrayType>(type);
98 return const_base(atype->getElementType());
101 return qt.isConstQualified();
104 /* Mark "decl" as having an unknown value in "assigned_value".
106 * If no (known or unknown) value was assigned to "decl" before,
107 * then it may have been treated as a parameter before and may
108 * therefore appear in a value assigned to another variable.
109 * If so, this assignment needs to be turned into an unknown value too.
111 static void clear_assignment(map<ValueDecl *, isl_pw_aff *> &assigned_value,
112 ValueDecl *decl)
114 map<ValueDecl *, isl_pw_aff *>::iterator it;
116 it = assigned_value.find(decl);
118 assigned_value[decl] = NULL;
120 if (it != assigned_value.end())
121 return;
123 for (it = assigned_value.begin(); it != assigned_value.end(); ++it) {
124 isl_pw_aff *pa = it->second;
125 int nparam = isl_pw_aff_dim(pa, isl_dim_param);
127 for (int i = 0; i < nparam; ++i) {
128 isl_id *id;
130 if (!isl_pw_aff_has_dim_id(pa, isl_dim_param, i))
131 continue;
132 id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
133 if (isl_id_get_user(id) == decl)
134 it->second = NULL;
135 isl_id_free(id);
140 /* Look for any assignments to scalar variables in part of the parse
141 * tree and set assigned_value to NULL for each of them.
142 * Also reset assigned_value if the address of a scalar variable
143 * is being taken. As an exception, if the address is passed to a function
144 * that is declared to receive a const pointer, then assigned_value is
145 * not reset.
147 * This ensures that we won't use any previously stored value
148 * in the current subtree and its parents.
150 struct clear_assignments : RecursiveASTVisitor<clear_assignments> {
151 map<ValueDecl *, isl_pw_aff *> &assigned_value;
152 set<UnaryOperator *> skip;
154 clear_assignments(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
155 assigned_value(assigned_value) {}
157 /* Check for "address of" operators whose value is passed
158 * to a const pointer argument and add them to "skip", so that
159 * we can skip them in VisitUnaryOperator.
161 bool VisitCallExpr(CallExpr *expr) {
162 FunctionDecl *fd;
163 fd = expr->getDirectCallee();
164 if (!fd)
165 return true;
166 for (int i = 0; i < expr->getNumArgs(); ++i) {
167 Expr *arg = expr->getArg(i);
168 UnaryOperator *op;
169 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
170 ImplicitCastExpr *ice;
171 ice = cast<ImplicitCastExpr>(arg);
172 arg = ice->getSubExpr();
174 if (arg->getStmtClass() != Stmt::UnaryOperatorClass)
175 continue;
176 op = cast<UnaryOperator>(arg);
177 if (op->getOpcode() != UO_AddrOf)
178 continue;
179 if (const_base(fd->getParamDecl(i)->getType()))
180 skip.insert(op);
182 return true;
185 bool VisitUnaryOperator(UnaryOperator *expr) {
186 Expr *arg;
187 DeclRefExpr *ref;
188 ValueDecl *decl;
190 switch (expr->getOpcode()) {
191 case UO_AddrOf:
192 case UO_PostInc:
193 case UO_PostDec:
194 case UO_PreInc:
195 case UO_PreDec:
196 break;
197 default:
198 return true;
200 if (skip.find(expr) != skip.end())
201 return true;
203 arg = expr->getSubExpr();
204 if (arg->getStmtClass() != Stmt::DeclRefExprClass)
205 return true;
206 ref = cast<DeclRefExpr>(arg);
207 decl = ref->getDecl();
208 clear_assignment(assigned_value, decl);
209 return true;
212 bool VisitBinaryOperator(BinaryOperator *expr) {
213 Expr *lhs;
214 DeclRefExpr *ref;
215 ValueDecl *decl;
217 if (!expr->isAssignmentOp())
218 return true;
219 lhs = expr->getLHS();
220 if (lhs->getStmtClass() != Stmt::DeclRefExprClass)
221 return true;
222 ref = cast<DeclRefExpr>(lhs);
223 decl = ref->getDecl();
224 clear_assignment(assigned_value, decl);
225 return true;
229 /* Keep a copy of the currently assigned values.
231 * Any variable that is assigned a value inside the current scope
232 * is removed again when we leave the scope (either because it wasn't
233 * stored in the cache or because it has a different value in the cache).
235 struct assigned_value_cache {
236 map<ValueDecl *, isl_pw_aff *> &assigned_value;
237 map<ValueDecl *, isl_pw_aff *> cache;
239 assigned_value_cache(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
240 assigned_value(assigned_value), cache(assigned_value) {}
241 ~assigned_value_cache() {
242 map<ValueDecl *, isl_pw_aff *>::iterator it = cache.begin();
243 for (it = assigned_value.begin(); it != assigned_value.end();
244 ++it) {
245 if (!it->second ||
246 (cache.find(it->first) != cache.end() &&
247 cache[it->first] != it->second))
248 cache[it->first] = NULL;
250 assigned_value = cache;
254 /* Insert an expression into the collection of expressions,
255 * provided it is not already in there.
256 * The isl_pw_affs are freed in the destructor.
258 void PetScan::insert_expression(__isl_take isl_pw_aff *expr)
260 std::set<isl_pw_aff *>::iterator it;
262 if (expressions.find(expr) == expressions.end())
263 expressions.insert(expr);
264 else
265 isl_pw_aff_free(expr);
268 PetScan::~PetScan()
270 std::set<isl_pw_aff *>::iterator it;
272 for (it = expressions.begin(); it != expressions.end(); ++it)
273 isl_pw_aff_free(*it);
275 isl_union_map_free(value_bounds);
278 /* Report a diagnostic, unless autodetect is set.
280 void PetScan::report(Stmt *stmt, unsigned id)
282 if (options->autodetect)
283 return;
285 SourceLocation loc = stmt->getLocStart();
286 DiagnosticsEngine &diag = PP.getDiagnostics();
287 DiagnosticBuilder B = diag.Report(loc, id) << stmt->getSourceRange();
290 /* Called if we found something we (currently) cannot handle.
291 * We'll provide more informative warnings later.
293 * We only actually complain if autodetect is false.
295 void PetScan::unsupported(Stmt *stmt)
297 DiagnosticsEngine &diag = PP.getDiagnostics();
298 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
299 "unsupported");
300 report(stmt, id);
303 /* Report a missing prototype, unless autodetect is set.
305 void PetScan::report_prototype_required(Stmt *stmt)
307 DiagnosticsEngine &diag = PP.getDiagnostics();
308 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
309 "prototype required");
310 report(stmt, id);
313 /* Report a missing increment, unless autodetect is set.
315 void PetScan::report_missing_increment(Stmt *stmt)
317 DiagnosticsEngine &diag = PP.getDiagnostics();
318 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
319 "missing increment");
320 report(stmt, id);
323 /* Extract an integer from "expr".
325 __isl_give isl_val *PetScan::extract_int(isl_ctx *ctx, IntegerLiteral *expr)
327 const Type *type = expr->getType().getTypePtr();
328 int is_signed = type->hasSignedIntegerRepresentation();
329 llvm::APInt val = expr->getValue();
330 int is_negative = is_signed && val.isNegative();
331 isl_val *v;
333 if (is_negative)
334 val = -val;
336 v = extract_unsigned(ctx, val);
338 if (is_negative)
339 v = isl_val_neg(v);
340 return v;
343 /* Extract an integer from "val", which is assumed to be non-negative.
345 __isl_give isl_val *PetScan::extract_unsigned(isl_ctx *ctx,
346 const llvm::APInt &val)
348 unsigned n;
349 const uint64_t *data;
351 data = val.getRawData();
352 n = val.getNumWords();
353 return isl_val_int_from_chunks(ctx, n, sizeof(uint64_t), data);
356 /* Extract an integer from "expr".
357 * Return NULL if "expr" does not (obviously) represent an integer.
359 __isl_give isl_val *PetScan::extract_int(clang::ParenExpr *expr)
361 return extract_int(expr->getSubExpr());
364 /* Extract an integer from "expr".
365 * Return NULL if "expr" does not (obviously) represent an integer.
367 __isl_give isl_val *PetScan::extract_int(clang::Expr *expr)
369 if (expr->getStmtClass() == Stmt::IntegerLiteralClass)
370 return extract_int(ctx, cast<IntegerLiteral>(expr));
371 if (expr->getStmtClass() == Stmt::ParenExprClass)
372 return extract_int(cast<ParenExpr>(expr));
374 unsupported(expr);
375 return NULL;
378 /* Extract an affine expression from the IntegerLiteral "expr".
380 __isl_give isl_pw_aff *PetScan::extract_affine(IntegerLiteral *expr)
382 isl_space *dim = isl_space_params_alloc(ctx, 0);
383 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
384 isl_aff *aff = isl_aff_zero_on_domain(ls);
385 isl_set *dom = isl_set_universe(dim);
386 isl_val *v;
388 v = extract_int(expr);
389 aff = isl_aff_add_constant_val(aff, v);
391 return isl_pw_aff_alloc(dom, aff);
394 /* Extract an affine expression from the APInt "val", which is assumed
395 * to be non-negative.
397 __isl_give isl_pw_aff *PetScan::extract_affine(const llvm::APInt &val)
399 isl_space *dim = isl_space_params_alloc(ctx, 0);
400 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
401 isl_aff *aff = isl_aff_zero_on_domain(ls);
402 isl_set *dom = isl_set_universe(dim);
403 isl_val *v;
405 v = extract_unsigned(ctx, val);
406 aff = isl_aff_add_constant_val(aff, v);
408 return isl_pw_aff_alloc(dom, aff);
411 __isl_give isl_pw_aff *PetScan::extract_affine(ImplicitCastExpr *expr)
413 return extract_affine(expr->getSubExpr());
416 static unsigned get_type_size(ValueDecl *decl)
418 return decl->getASTContext().getIntWidth(decl->getType());
421 /* Bound parameter "pos" of "set" to the possible values of "decl".
423 static __isl_give isl_set *set_parameter_bounds(__isl_take isl_set *set,
424 unsigned pos, ValueDecl *decl)
426 unsigned width;
427 isl_ctx *ctx;
428 isl_val *bound;
430 ctx = isl_set_get_ctx(set);
431 width = get_type_size(decl);
432 if (decl->getType()->isUnsignedIntegerType()) {
433 set = isl_set_lower_bound_si(set, isl_dim_param, pos, 0);
434 bound = isl_val_int_from_ui(ctx, width);
435 bound = isl_val_2exp(bound);
436 bound = isl_val_sub_ui(bound, 1);
437 set = isl_set_upper_bound_val(set, isl_dim_param, pos, bound);
438 } else {
439 bound = isl_val_int_from_ui(ctx, width - 1);
440 bound = isl_val_2exp(bound);
441 bound = isl_val_sub_ui(bound, 1);
442 set = isl_set_upper_bound_val(set, isl_dim_param, pos,
443 isl_val_copy(bound));
444 bound = isl_val_neg(bound);
445 bound = isl_val_sub_ui(bound, 1);
446 set = isl_set_lower_bound_val(set, isl_dim_param, pos, bound);
449 return set;
452 /* Extract an affine expression from the DeclRefExpr "expr".
454 * If the variable has been assigned a value, then we check whether
455 * we know what (affine) value was assigned.
456 * If so, we return this value. Otherwise we convert "expr"
457 * to an extra parameter (provided nesting_enabled is set).
459 * Otherwise, we simply return an expression that is equal
460 * to a parameter corresponding to the referenced variable.
462 __isl_give isl_pw_aff *PetScan::extract_affine(DeclRefExpr *expr)
464 ValueDecl *decl = expr->getDecl();
465 const Type *type = decl->getType().getTypePtr();
466 isl_id *id;
467 isl_space *dim;
468 isl_aff *aff;
469 isl_set *dom;
471 if (!type->isIntegerType()) {
472 unsupported(expr);
473 return NULL;
476 if (assigned_value.find(decl) != assigned_value.end()) {
477 if (assigned_value[decl])
478 return isl_pw_aff_copy(assigned_value[decl]);
479 else
480 return nested_access(expr);
483 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
484 dim = isl_space_params_alloc(ctx, 1);
486 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
488 dom = isl_set_universe(isl_space_copy(dim));
489 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
490 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
492 return isl_pw_aff_alloc(dom, aff);
495 /* Extract an affine expression from an integer division operation.
496 * In particular, if "expr" is lhs/rhs, then return
498 * lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs)
500 * The second argument (rhs) is required to be a (positive) integer constant.
502 __isl_give isl_pw_aff *PetScan::extract_affine_div(BinaryOperator *expr)
504 int is_cst;
505 isl_pw_aff *rhs, *lhs;
507 rhs = extract_affine(expr->getRHS());
508 is_cst = isl_pw_aff_is_cst(rhs);
509 if (is_cst < 0 || !is_cst) {
510 isl_pw_aff_free(rhs);
511 if (!is_cst)
512 unsupported(expr);
513 return NULL;
516 lhs = extract_affine(expr->getLHS());
518 return isl_pw_aff_tdiv_q(lhs, rhs);
521 /* Extract an affine expression from a modulo operation.
522 * In particular, if "expr" is lhs/rhs, then return
524 * lhs - rhs * (lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs))
526 * The second argument (rhs) is required to be a (positive) integer constant.
528 __isl_give isl_pw_aff *PetScan::extract_affine_mod(BinaryOperator *expr)
530 int is_cst;
531 isl_pw_aff *rhs, *lhs;
533 rhs = extract_affine(expr->getRHS());
534 is_cst = isl_pw_aff_is_cst(rhs);
535 if (is_cst < 0 || !is_cst) {
536 isl_pw_aff_free(rhs);
537 if (!is_cst)
538 unsupported(expr);
539 return NULL;
542 lhs = extract_affine(expr->getLHS());
544 return isl_pw_aff_tdiv_r(lhs, rhs);
547 /* Extract an affine expression from a multiplication operation.
548 * This is only allowed if at least one of the two arguments
549 * is a (piecewise) constant.
551 __isl_give isl_pw_aff *PetScan::extract_affine_mul(BinaryOperator *expr)
553 isl_pw_aff *lhs;
554 isl_pw_aff *rhs;
556 lhs = extract_affine(expr->getLHS());
557 rhs = extract_affine(expr->getRHS());
559 if (!isl_pw_aff_is_cst(lhs) && !isl_pw_aff_is_cst(rhs)) {
560 isl_pw_aff_free(lhs);
561 isl_pw_aff_free(rhs);
562 unsupported(expr);
563 return NULL;
566 return isl_pw_aff_mul(lhs, rhs);
569 /* Extract an affine expression from an addition or subtraction operation.
571 __isl_give isl_pw_aff *PetScan::extract_affine_add(BinaryOperator *expr)
573 isl_pw_aff *lhs;
574 isl_pw_aff *rhs;
576 lhs = extract_affine(expr->getLHS());
577 rhs = extract_affine(expr->getRHS());
579 switch (expr->getOpcode()) {
580 case BO_Add:
581 return isl_pw_aff_add(lhs, rhs);
582 case BO_Sub:
583 return isl_pw_aff_sub(lhs, rhs);
584 default:
585 isl_pw_aff_free(lhs);
586 isl_pw_aff_free(rhs);
587 return NULL;
592 /* Compute
594 * pwaff mod 2^width
596 static __isl_give isl_pw_aff *wrap(__isl_take isl_pw_aff *pwaff,
597 unsigned width)
599 isl_ctx *ctx;
600 isl_val *mod;
602 ctx = isl_pw_aff_get_ctx(pwaff);
603 mod = isl_val_int_from_ui(ctx, width);
604 mod = isl_val_2exp(mod);
606 pwaff = isl_pw_aff_mod_val(pwaff, mod);
608 return pwaff;
611 /* Limit the domain of "pwaff" to those elements where the function
612 * value satisfies
614 * 2^{width-1} <= pwaff < 2^{width-1}
616 static __isl_give isl_pw_aff *avoid_overflow(__isl_take isl_pw_aff *pwaff,
617 unsigned width)
619 isl_ctx *ctx;
620 isl_val *v;
621 isl_space *space = isl_pw_aff_get_domain_space(pwaff);
622 isl_local_space *ls = isl_local_space_from_space(space);
623 isl_aff *bound;
624 isl_set *dom;
625 isl_pw_aff *b;
627 ctx = isl_pw_aff_get_ctx(pwaff);
628 v = isl_val_int_from_ui(ctx, width - 1);
629 v = isl_val_2exp(v);
631 bound = isl_aff_zero_on_domain(ls);
632 bound = isl_aff_add_constant_val(bound, v);
633 b = isl_pw_aff_from_aff(bound);
635 dom = isl_pw_aff_lt_set(isl_pw_aff_copy(pwaff), isl_pw_aff_copy(b));
636 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
638 b = isl_pw_aff_neg(b);
639 dom = isl_pw_aff_ge_set(isl_pw_aff_copy(pwaff), b);
640 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
642 return pwaff;
645 /* Handle potential overflows on signed computations.
647 * If options->signed_overflow is set to PET_OVERFLOW_AVOID,
648 * the we adjust the domain of "pa" to avoid overflows.
650 __isl_give isl_pw_aff *PetScan::signed_overflow(__isl_take isl_pw_aff *pa,
651 unsigned width)
653 if (options->signed_overflow == PET_OVERFLOW_AVOID)
654 pa = avoid_overflow(pa, width);
656 return pa;
659 /* Return the piecewise affine expression "set ? 1 : 0" defined on "dom".
661 static __isl_give isl_pw_aff *indicator_function(__isl_take isl_set *set,
662 __isl_take isl_set *dom)
664 isl_pw_aff *pa;
665 pa = isl_set_indicator_function(set);
666 pa = isl_pw_aff_intersect_domain(pa, isl_set_coalesce(dom));
667 return pa;
670 /* Extract an affine expression from some binary operations.
671 * If the result of the expression is unsigned, then we wrap it
672 * based on the size of the type. Otherwise, we ensure that
673 * no overflow occurs.
675 __isl_give isl_pw_aff *PetScan::extract_affine(BinaryOperator *expr)
677 isl_pw_aff *res;
678 unsigned width;
680 switch (expr->getOpcode()) {
681 case BO_Add:
682 case BO_Sub:
683 res = extract_affine_add(expr);
684 break;
685 case BO_Div:
686 res = extract_affine_div(expr);
687 break;
688 case BO_Rem:
689 res = extract_affine_mod(expr);
690 break;
691 case BO_Mul:
692 res = extract_affine_mul(expr);
693 break;
694 case BO_LT:
695 case BO_LE:
696 case BO_GT:
697 case BO_GE:
698 case BO_EQ:
699 case BO_NE:
700 case BO_LAnd:
701 case BO_LOr:
702 return extract_condition(expr);
703 default:
704 unsupported(expr);
705 return NULL;
708 width = ast_context.getIntWidth(expr->getType());
709 if (expr->getType()->isUnsignedIntegerType())
710 res = wrap(res, width);
711 else
712 res = signed_overflow(res, width);
714 return res;
717 /* Extract an affine expression from a negation operation.
719 __isl_give isl_pw_aff *PetScan::extract_affine(UnaryOperator *expr)
721 if (expr->getOpcode() == UO_Minus)
722 return isl_pw_aff_neg(extract_affine(expr->getSubExpr()));
723 if (expr->getOpcode() == UO_LNot)
724 return extract_condition(expr);
726 unsupported(expr);
727 return NULL;
730 __isl_give isl_pw_aff *PetScan::extract_affine(ParenExpr *expr)
732 return extract_affine(expr->getSubExpr());
735 /* Extract an affine expression from some special function calls.
736 * In particular, we handle "min", "max", "ceild", "floord",
737 * "intMod", "intFloor" and "intCeil".
738 * In case of the latter five, the second argument needs to be
739 * a (positive) integer constant.
741 __isl_give isl_pw_aff *PetScan::extract_affine(CallExpr *expr)
743 FunctionDecl *fd;
744 string name;
745 isl_pw_aff *aff1, *aff2;
747 fd = expr->getDirectCallee();
748 if (!fd) {
749 unsupported(expr);
750 return NULL;
753 name = fd->getDeclName().getAsString();
754 if (!(expr->getNumArgs() == 2 && name == "min") &&
755 !(expr->getNumArgs() == 2 && name == "max") &&
756 !(expr->getNumArgs() == 2 && name == "intMod") &&
757 !(expr->getNumArgs() == 2 && name == "intFloor") &&
758 !(expr->getNumArgs() == 2 && name == "intCeil") &&
759 !(expr->getNumArgs() == 2 && name == "floord") &&
760 !(expr->getNumArgs() == 2 && name == "ceild")) {
761 unsupported(expr);
762 return NULL;
765 if (name == "min" || name == "max") {
766 aff1 = extract_affine(expr->getArg(0));
767 aff2 = extract_affine(expr->getArg(1));
769 if (name == "min")
770 aff1 = isl_pw_aff_min(aff1, aff2);
771 else
772 aff1 = isl_pw_aff_max(aff1, aff2);
773 } else if (name == "intMod") {
774 isl_val *v;
775 Expr *arg2 = expr->getArg(1);
777 if (arg2->getStmtClass() != Stmt::IntegerLiteralClass) {
778 unsupported(expr);
779 return NULL;
781 aff1 = extract_affine(expr->getArg(0));
782 v = extract_int(cast<IntegerLiteral>(arg2));
783 aff1 = isl_pw_aff_mod_val(aff1, v);
784 } else if (name == "floord" || name == "ceild" ||
785 name == "intFloor" || name == "intCeil") {
786 isl_val *v;
787 Expr *arg2 = expr->getArg(1);
789 if (arg2->getStmtClass() != Stmt::IntegerLiteralClass) {
790 unsupported(expr);
791 return NULL;
793 aff1 = extract_affine(expr->getArg(0));
794 v = extract_int(cast<IntegerLiteral>(arg2));
795 aff1 = isl_pw_aff_scale_down_val(aff1, v);
796 if (name == "floord" || name == "intFloor")
797 aff1 = isl_pw_aff_floor(aff1);
798 else
799 aff1 = isl_pw_aff_ceil(aff1);
800 } else {
801 unsupported(expr);
802 return NULL;
805 return aff1;
808 /* This method is called when we come across an access that is
809 * nested in what is supposed to be an affine expression.
810 * If nesting is allowed, we return a new parameter that corresponds
811 * to this nested access. Otherwise, we simply complain.
813 * Note that we currently don't allow nested accesses themselves
814 * to contain any nested accesses, so we check if we can extract
815 * the access without any nesting and complain if we can't.
817 * The new parameter is resolved in resolve_nested.
819 isl_pw_aff *PetScan::nested_access(Expr *expr)
821 isl_id *id;
822 isl_space *dim;
823 isl_aff *aff;
824 isl_set *dom;
825 isl_multi_pw_aff *index;
827 if (!nesting_enabled) {
828 unsupported(expr);
829 return NULL;
832 allow_nested = false;
833 index = extract_index(expr);
834 allow_nested = true;
835 if (!index) {
836 unsupported(expr);
837 return NULL;
839 isl_multi_pw_aff_free(index);
841 id = pet_nested_clang_expr(ctx, expr);
842 dim = isl_space_params_alloc(ctx, 1);
844 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
846 dom = isl_set_universe(isl_space_copy(dim));
847 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
848 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
850 return isl_pw_aff_alloc(dom, aff);
853 /* Affine expressions are not supposed to contain array accesses,
854 * but if nesting is allowed, we return a parameter corresponding
855 * to the array access.
857 __isl_give isl_pw_aff *PetScan::extract_affine(ArraySubscriptExpr *expr)
859 return nested_access(expr);
862 /* Affine expressions are not supposed to contain member accesses,
863 * but if nesting is allowed, we return a parameter corresponding
864 * to the member access.
866 __isl_give isl_pw_aff *PetScan::extract_affine(MemberExpr *expr)
868 return nested_access(expr);
871 /* Extract an affine expression from a conditional operation.
873 __isl_give isl_pw_aff *PetScan::extract_affine(ConditionalOperator *expr)
875 isl_pw_aff *cond, *lhs, *rhs;
877 cond = extract_condition(expr->getCond());
878 lhs = extract_affine(expr->getTrueExpr());
879 rhs = extract_affine(expr->getFalseExpr());
881 return isl_pw_aff_cond(cond, lhs, rhs);
884 /* Extract an affine expression, if possible, from "expr".
885 * Otherwise return NULL.
887 __isl_give isl_pw_aff *PetScan::extract_affine(Expr *expr)
889 switch (expr->getStmtClass()) {
890 case Stmt::ImplicitCastExprClass:
891 return extract_affine(cast<ImplicitCastExpr>(expr));
892 case Stmt::IntegerLiteralClass:
893 return extract_affine(cast<IntegerLiteral>(expr));
894 case Stmt::DeclRefExprClass:
895 return extract_affine(cast<DeclRefExpr>(expr));
896 case Stmt::BinaryOperatorClass:
897 return extract_affine(cast<BinaryOperator>(expr));
898 case Stmt::UnaryOperatorClass:
899 return extract_affine(cast<UnaryOperator>(expr));
900 case Stmt::ParenExprClass:
901 return extract_affine(cast<ParenExpr>(expr));
902 case Stmt::CallExprClass:
903 return extract_affine(cast<CallExpr>(expr));
904 case Stmt::ArraySubscriptExprClass:
905 return extract_affine(cast<ArraySubscriptExpr>(expr));
906 case Stmt::MemberExprClass:
907 return extract_affine(cast<MemberExpr>(expr));
908 case Stmt::ConditionalOperatorClass:
909 return extract_affine(cast<ConditionalOperator>(expr));
910 default:
911 unsupported(expr);
913 return NULL;
916 __isl_give isl_multi_pw_aff *PetScan::extract_index(ImplicitCastExpr *expr)
918 return extract_index(expr->getSubExpr());
921 /* Return the depth of an array of the given type.
923 static int array_depth(const Type *type)
925 if (type->isPointerType())
926 return 1 + array_depth(type->getPointeeType().getTypePtr());
927 if (type->isArrayType()) {
928 const ArrayType *atype;
929 type = type->getCanonicalTypeInternal().getTypePtr();
930 atype = cast<ArrayType>(type);
931 return 1 + array_depth(atype->getElementType().getTypePtr());
933 return 0;
936 /* Return the depth of the array accessed by the index expression "index".
937 * If "index" is an affine expression, i.e., if it does not access
938 * any array, then return 1.
939 * If "index" represent a member access, i.e., if its range is a wrapped
940 * relation, then return the sum of the depth of the array of structures
941 * and that of the member inside the structure.
943 static int extract_depth(__isl_keep isl_multi_pw_aff *index)
945 isl_id *id;
946 ValueDecl *decl;
948 if (!index)
949 return -1;
951 if (isl_multi_pw_aff_range_is_wrapping(index)) {
952 int domain_depth, range_depth;
953 isl_multi_pw_aff *domain, *range;
955 domain = isl_multi_pw_aff_copy(index);
956 domain = isl_multi_pw_aff_range_factor_domain(domain);
957 domain_depth = extract_depth(domain);
958 isl_multi_pw_aff_free(domain);
959 range = isl_multi_pw_aff_copy(index);
960 range = isl_multi_pw_aff_range_factor_range(range);
961 range_depth = extract_depth(range);
962 isl_multi_pw_aff_free(range);
964 return domain_depth + range_depth;
967 if (!isl_multi_pw_aff_has_tuple_id(index, isl_dim_out))
968 return 1;
970 id = isl_multi_pw_aff_get_tuple_id(index, isl_dim_out);
971 if (!id)
972 return -1;
973 decl = (ValueDecl *) isl_id_get_user(id);
974 isl_id_free(id);
976 return array_depth(decl->getType().getTypePtr());
979 /* Extract an index expression from a reference to a variable.
980 * If the variable has name "A", then the returned index expression
981 * is of the form
983 * { [] -> A[] }
985 __isl_give isl_multi_pw_aff *PetScan::extract_index(DeclRefExpr *expr)
987 return extract_index(expr->getDecl());
990 /* Extract an index expression from a variable.
991 * If the variable has name "A", then the returned index expression
992 * is of the form
994 * { [] -> A[] }
996 __isl_give isl_multi_pw_aff *PetScan::extract_index(ValueDecl *decl)
998 isl_id *id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
999 isl_space *space = isl_space_alloc(ctx, 0, 0, 0);
1001 space = isl_space_set_tuple_id(space, isl_dim_out, id);
1003 return isl_multi_pw_aff_zero(space);
1006 /* Extract an index expression from an integer contant.
1007 * If the value of the constant is "v", then the returned access relation
1008 * is
1010 * { [] -> [v] }
1012 __isl_give isl_multi_pw_aff *PetScan::extract_index(IntegerLiteral *expr)
1014 isl_multi_pw_aff *mpa;
1016 mpa = isl_multi_pw_aff_from_pw_aff(extract_affine(expr));
1017 mpa = isl_multi_pw_aff_from_range(mpa);
1018 return mpa;
1021 /* Try and extract an index expression from the given Expr.
1022 * Return NULL if it doesn't work out.
1024 __isl_give isl_multi_pw_aff *PetScan::extract_index(Expr *expr)
1026 switch (expr->getStmtClass()) {
1027 case Stmt::ImplicitCastExprClass:
1028 return extract_index(cast<ImplicitCastExpr>(expr));
1029 case Stmt::DeclRefExprClass:
1030 return extract_index(cast<DeclRefExpr>(expr));
1031 case Stmt::ArraySubscriptExprClass:
1032 return extract_index(cast<ArraySubscriptExpr>(expr));
1033 case Stmt::IntegerLiteralClass:
1034 return extract_index(cast<IntegerLiteral>(expr));
1035 case Stmt::MemberExprClass:
1036 return extract_index(cast<MemberExpr>(expr));
1037 default:
1038 unsupported(expr);
1040 return NULL;
1043 /* Given a partial index expression "base" and an extra index "index",
1044 * append the extra index to "base" and return the result.
1045 * Additionally, add the constraints that the extra index is non-negative.
1046 * If "index" represent a member access, i.e., if its range is a wrapped
1047 * relation, then we recursively extend the range of this nested relation.
1049 static __isl_give isl_multi_pw_aff *subscript(__isl_take isl_multi_pw_aff *base,
1050 __isl_take isl_pw_aff *index)
1052 isl_id *id;
1053 isl_set *domain;
1054 isl_multi_pw_aff *access;
1055 int member_access;
1057 member_access = isl_multi_pw_aff_range_is_wrapping(base);
1058 if (member_access < 0)
1059 goto error;
1060 if (member_access) {
1061 isl_multi_pw_aff *domain, *range;
1062 isl_id *id;
1064 id = isl_multi_pw_aff_get_tuple_id(base, isl_dim_out);
1065 domain = isl_multi_pw_aff_copy(base);
1066 domain = isl_multi_pw_aff_range_factor_domain(domain);
1067 range = isl_multi_pw_aff_range_factor_range(base);
1068 range = subscript(range, index);
1069 access = isl_multi_pw_aff_range_product(domain, range);
1070 access = isl_multi_pw_aff_set_tuple_id(access, isl_dim_out, id);
1071 return access;
1074 id = isl_multi_pw_aff_get_tuple_id(base, isl_dim_set);
1075 index = isl_pw_aff_from_range(index);
1076 domain = isl_pw_aff_nonneg_set(isl_pw_aff_copy(index));
1077 index = isl_pw_aff_intersect_domain(index, domain);
1078 access = isl_multi_pw_aff_from_pw_aff(index);
1079 access = isl_multi_pw_aff_flat_range_product(base, access);
1080 access = isl_multi_pw_aff_set_tuple_id(access, isl_dim_set, id);
1082 return access;
1083 error:
1084 isl_multi_pw_aff_free(base);
1085 isl_pw_aff_free(index);
1086 return NULL;
1089 /* Extract an index expression from the given array subscript expression.
1090 * If nesting is allowed in general, then we turn it on while
1091 * examining the index expression.
1093 * We first extract an index expression from the base.
1094 * This will result in an index expression with a range that corresponds
1095 * to the earlier indices.
1096 * We then extract the current index, restrict its domain
1097 * to those values that result in a non-negative index and
1098 * append the index to the base index expression.
1100 __isl_give isl_multi_pw_aff *PetScan::extract_index(ArraySubscriptExpr *expr)
1102 Expr *base = expr->getBase();
1103 Expr *idx = expr->getIdx();
1104 isl_pw_aff *index;
1105 isl_multi_pw_aff *base_access;
1106 isl_multi_pw_aff *access;
1107 bool save_nesting = nesting_enabled;
1109 nesting_enabled = allow_nested;
1111 base_access = extract_index(base);
1112 index = extract_affine(idx);
1114 nesting_enabled = save_nesting;
1116 access = subscript(base_access, index);
1118 return access;
1121 /* Construct a name for a member access by concatenating the name
1122 * of the array of structures and the member, separated by an underscore.
1124 * The caller is responsible for freeing the result.
1126 static char *member_access_name(isl_ctx *ctx, const char *base,
1127 const char *field)
1129 int len;
1130 char *name;
1132 len = strlen(base) + 1 + strlen(field);
1133 name = isl_alloc_array(ctx, char, len + 1);
1134 if (!name)
1135 return NULL;
1136 snprintf(name, len + 1, "%s_%s", base, field);
1138 return name;
1141 /* Given an index expression "base" for an element of an array of structures
1142 * and an expression "field" for the field member being accessed, construct
1143 * an index expression for an access to that member of the given structure.
1144 * In particular, take the range product of "base" and "field" and
1145 * attach a name to the result.
1147 static __isl_give isl_multi_pw_aff *member(__isl_take isl_multi_pw_aff *base,
1148 __isl_take isl_multi_pw_aff *field)
1150 isl_ctx *ctx;
1151 isl_multi_pw_aff *access;
1152 const char *base_name, *field_name;
1153 char *name;
1155 ctx = isl_multi_pw_aff_get_ctx(base);
1157 base_name = isl_multi_pw_aff_get_tuple_name(base, isl_dim_out);
1158 field_name = isl_multi_pw_aff_get_tuple_name(field, isl_dim_out);
1159 name = member_access_name(ctx, base_name, field_name);
1161 access = isl_multi_pw_aff_range_product(base, field);
1163 access = isl_multi_pw_aff_set_tuple_name(access, isl_dim_out, name);
1164 free(name);
1166 return access;
1169 /* Extract an index expression from a member expression.
1171 * If the base access (to the structure containing the member)
1172 * is of the form
1174 * [] -> A[..]
1176 * and the member is called "f", then the member access is of
1177 * the form
1179 * [] -> A_f[A[..] -> f[]]
1181 * If the member access is to an anonymous struct, then simply return
1183 * [] -> A[..]
1185 * If the member access in the source code is of the form
1187 * A->f
1189 * then it is treated as
1191 * A[0].f
1193 __isl_give isl_multi_pw_aff *PetScan::extract_index(MemberExpr *expr)
1195 Expr *base = expr->getBase();
1196 FieldDecl *field = cast<FieldDecl>(expr->getMemberDecl());
1197 isl_multi_pw_aff *base_access, *field_access;
1198 isl_id *id;
1199 isl_space *space;
1201 base_access = extract_index(base);
1203 if (expr->isArrow()) {
1204 isl_space *space = isl_space_params_alloc(ctx, 0);
1205 isl_local_space *ls = isl_local_space_from_space(space);
1206 isl_aff *aff = isl_aff_zero_on_domain(ls);
1207 isl_pw_aff *index = isl_pw_aff_from_aff(aff);
1208 base_access = subscript(base_access, index);
1211 if (field->isAnonymousStructOrUnion())
1212 return base_access;
1214 id = isl_id_alloc(ctx, field->getName().str().c_str(), field);
1215 space = isl_multi_pw_aff_get_domain_space(base_access);
1216 space = isl_space_from_domain(space);
1217 space = isl_space_set_tuple_id(space, isl_dim_out, id);
1218 field_access = isl_multi_pw_aff_zero(space);
1220 return member(base_access, field_access);
1223 /* Check if "expr" calls function "minmax" with two arguments and if so
1224 * make lhs and rhs refer to these two arguments.
1226 static bool is_minmax(Expr *expr, const char *minmax, Expr *&lhs, Expr *&rhs)
1228 CallExpr *call;
1229 FunctionDecl *fd;
1230 string name;
1232 if (expr->getStmtClass() != Stmt::CallExprClass)
1233 return false;
1235 call = cast<CallExpr>(expr);
1236 fd = call->getDirectCallee();
1237 if (!fd)
1238 return false;
1240 if (call->getNumArgs() != 2)
1241 return false;
1243 name = fd->getDeclName().getAsString();
1244 if (name != minmax)
1245 return false;
1247 lhs = call->getArg(0);
1248 rhs = call->getArg(1);
1250 return true;
1253 /* Check if "expr" is of the form min(lhs, rhs) and if so make
1254 * lhs and rhs refer to the two arguments.
1256 static bool is_min(Expr *expr, Expr *&lhs, Expr *&rhs)
1258 return is_minmax(expr, "min", lhs, rhs);
1261 /* Check if "expr" is of the form max(lhs, rhs) and if so make
1262 * lhs and rhs refer to the two arguments.
1264 static bool is_max(Expr *expr, Expr *&lhs, Expr *&rhs)
1266 return is_minmax(expr, "max", lhs, rhs);
1269 /* Return "lhs && rhs", defined on the shared definition domain.
1271 static __isl_give isl_pw_aff *pw_aff_and(__isl_take isl_pw_aff *lhs,
1272 __isl_take isl_pw_aff *rhs)
1274 isl_set *cond;
1275 isl_set *dom;
1277 dom = isl_set_intersect(isl_pw_aff_domain(isl_pw_aff_copy(lhs)),
1278 isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1279 cond = isl_set_intersect(isl_pw_aff_non_zero_set(lhs),
1280 isl_pw_aff_non_zero_set(rhs));
1281 return indicator_function(cond, dom);
1284 /* Return "lhs && rhs", with shortcut semantics.
1285 * That is, if lhs is false, then the result is defined even if rhs is not.
1286 * In practice, we compute lhs ? rhs : lhs.
1288 static __isl_give isl_pw_aff *pw_aff_and_then(__isl_take isl_pw_aff *lhs,
1289 __isl_take isl_pw_aff *rhs)
1291 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), rhs, lhs);
1294 /* Return "lhs || rhs", with shortcut semantics.
1295 * That is, if lhs is true, then the result is defined even if rhs is not.
1296 * In practice, we compute lhs ? lhs : rhs.
1298 static __isl_give isl_pw_aff *pw_aff_or_else(__isl_take isl_pw_aff *lhs,
1299 __isl_take isl_pw_aff *rhs)
1301 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), lhs, rhs);
1304 /* Extract an affine expressions representing the comparison "LHS op RHS"
1305 * "comp" is the original statement that "LHS op RHS" is derived from
1306 * and is used for diagnostics.
1308 * If the comparison is of the form
1310 * a <= min(b,c)
1312 * then the expression is constructed as the conjunction of
1313 * the comparisons
1315 * a <= b and a <= c
1317 * A similar optimization is performed for max(a,b) <= c.
1318 * We do this because that will lead to simpler representations
1319 * of the expression.
1320 * If isl is ever enhanced to explicitly deal with min and max expressions,
1321 * this optimization can be removed.
1323 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperatorKind op,
1324 Expr *LHS, Expr *RHS, Stmt *comp)
1326 isl_pw_aff *lhs;
1327 isl_pw_aff *rhs;
1328 isl_pw_aff *res;
1329 isl_set *cond;
1330 isl_set *dom;
1332 if (op == BO_GT)
1333 return extract_comparison(BO_LT, RHS, LHS, comp);
1334 if (op == BO_GE)
1335 return extract_comparison(BO_LE, RHS, LHS, comp);
1337 if (op == BO_LT || op == BO_LE) {
1338 Expr *expr1, *expr2;
1339 if (is_min(RHS, expr1, expr2)) {
1340 lhs = extract_comparison(op, LHS, expr1, comp);
1341 rhs = extract_comparison(op, LHS, expr2, comp);
1342 return pw_aff_and(lhs, rhs);
1344 if (is_max(LHS, expr1, expr2)) {
1345 lhs = extract_comparison(op, expr1, RHS, comp);
1346 rhs = extract_comparison(op, expr2, RHS, comp);
1347 return pw_aff_and(lhs, rhs);
1351 lhs = extract_affine(LHS);
1352 rhs = extract_affine(RHS);
1354 dom = isl_pw_aff_domain(isl_pw_aff_copy(lhs));
1355 dom = isl_set_intersect(dom, isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1357 switch (op) {
1358 case BO_LT:
1359 cond = isl_pw_aff_lt_set(lhs, rhs);
1360 break;
1361 case BO_LE:
1362 cond = isl_pw_aff_le_set(lhs, rhs);
1363 break;
1364 case BO_EQ:
1365 cond = isl_pw_aff_eq_set(lhs, rhs);
1366 break;
1367 case BO_NE:
1368 cond = isl_pw_aff_ne_set(lhs, rhs);
1369 break;
1370 default:
1371 isl_pw_aff_free(lhs);
1372 isl_pw_aff_free(rhs);
1373 isl_set_free(dom);
1374 unsupported(comp);
1375 return NULL;
1378 cond = isl_set_coalesce(cond);
1379 res = indicator_function(cond, dom);
1381 return res;
1384 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperator *comp)
1386 return extract_comparison(comp->getOpcode(), comp->getLHS(),
1387 comp->getRHS(), comp);
1390 /* Extract an affine expression representing the negation (logical not)
1391 * of a subexpression.
1393 __isl_give isl_pw_aff *PetScan::extract_boolean(UnaryOperator *op)
1395 isl_set *set_cond, *dom;
1396 isl_pw_aff *cond, *res;
1398 cond = extract_condition(op->getSubExpr());
1400 dom = isl_pw_aff_domain(isl_pw_aff_copy(cond));
1402 set_cond = isl_pw_aff_zero_set(cond);
1404 res = indicator_function(set_cond, dom);
1406 return res;
1409 /* Extract an affine expression representing the disjunction (logical or)
1410 * or conjunction (logical and) of two subexpressions.
1412 __isl_give isl_pw_aff *PetScan::extract_boolean(BinaryOperator *comp)
1414 isl_pw_aff *lhs, *rhs;
1416 lhs = extract_condition(comp->getLHS());
1417 rhs = extract_condition(comp->getRHS());
1419 switch (comp->getOpcode()) {
1420 case BO_LAnd:
1421 return pw_aff_and_then(lhs, rhs);
1422 case BO_LOr:
1423 return pw_aff_or_else(lhs, rhs);
1424 default:
1425 isl_pw_aff_free(lhs);
1426 isl_pw_aff_free(rhs);
1429 unsupported(comp);
1430 return NULL;
1433 __isl_give isl_pw_aff *PetScan::extract_condition(UnaryOperator *expr)
1435 switch (expr->getOpcode()) {
1436 case UO_LNot:
1437 return extract_boolean(expr);
1438 default:
1439 unsupported(expr);
1440 return NULL;
1444 /* Extract the affine expression "expr != 0 ? 1 : 0".
1446 __isl_give isl_pw_aff *PetScan::extract_implicit_condition(Expr *expr)
1448 isl_pw_aff *res;
1449 isl_set *set, *dom;
1451 res = extract_affine(expr);
1453 dom = isl_pw_aff_domain(isl_pw_aff_copy(res));
1454 set = isl_pw_aff_non_zero_set(res);
1456 res = indicator_function(set, dom);
1458 return res;
1461 /* Extract an affine expression from a boolean expression.
1462 * In particular, return the expression "expr ? 1 : 0".
1464 * If the expression doesn't look like a condition, we assume it
1465 * is an affine expression and return the condition "expr != 0 ? 1 : 0".
1467 __isl_give isl_pw_aff *PetScan::extract_condition(Expr *expr)
1469 BinaryOperator *comp;
1471 if (!expr) {
1472 isl_set *u = isl_set_universe(isl_space_params_alloc(ctx, 0));
1473 return indicator_function(u, isl_set_copy(u));
1476 if (expr->getStmtClass() == Stmt::ParenExprClass)
1477 return extract_condition(cast<ParenExpr>(expr)->getSubExpr());
1479 if (expr->getStmtClass() == Stmt::UnaryOperatorClass)
1480 return extract_condition(cast<UnaryOperator>(expr));
1482 if (expr->getStmtClass() != Stmt::BinaryOperatorClass)
1483 return extract_implicit_condition(expr);
1485 comp = cast<BinaryOperator>(expr);
1486 switch (comp->getOpcode()) {
1487 case BO_LT:
1488 case BO_LE:
1489 case BO_GT:
1490 case BO_GE:
1491 case BO_EQ:
1492 case BO_NE:
1493 return extract_comparison(comp);
1494 case BO_LAnd:
1495 case BO_LOr:
1496 return extract_boolean(comp);
1497 default:
1498 return extract_implicit_condition(expr);
1502 static enum pet_op_type UnaryOperatorKind2pet_op_type(UnaryOperatorKind kind)
1504 switch (kind) {
1505 case UO_Minus:
1506 return pet_op_minus;
1507 case UO_Not:
1508 return pet_op_not;
1509 case UO_LNot:
1510 return pet_op_lnot;
1511 case UO_PostInc:
1512 return pet_op_post_inc;
1513 case UO_PostDec:
1514 return pet_op_post_dec;
1515 case UO_PreInc:
1516 return pet_op_pre_inc;
1517 case UO_PreDec:
1518 return pet_op_pre_dec;
1519 default:
1520 return pet_op_last;
1524 static enum pet_op_type BinaryOperatorKind2pet_op_type(BinaryOperatorKind kind)
1526 switch (kind) {
1527 case BO_AddAssign:
1528 return pet_op_add_assign;
1529 case BO_SubAssign:
1530 return pet_op_sub_assign;
1531 case BO_MulAssign:
1532 return pet_op_mul_assign;
1533 case BO_DivAssign:
1534 return pet_op_div_assign;
1535 case BO_Assign:
1536 return pet_op_assign;
1537 case BO_Add:
1538 return pet_op_add;
1539 case BO_Sub:
1540 return pet_op_sub;
1541 case BO_Mul:
1542 return pet_op_mul;
1543 case BO_Div:
1544 return pet_op_div;
1545 case BO_Rem:
1546 return pet_op_mod;
1547 case BO_Shl:
1548 return pet_op_shl;
1549 case BO_Shr:
1550 return pet_op_shr;
1551 case BO_EQ:
1552 return pet_op_eq;
1553 case BO_NE:
1554 return pet_op_ne;
1555 case BO_LE:
1556 return pet_op_le;
1557 case BO_GE:
1558 return pet_op_ge;
1559 case BO_LT:
1560 return pet_op_lt;
1561 case BO_GT:
1562 return pet_op_gt;
1563 case BO_And:
1564 return pet_op_and;
1565 case BO_Xor:
1566 return pet_op_xor;
1567 case BO_Or:
1568 return pet_op_or;
1569 case BO_LAnd:
1570 return pet_op_land;
1571 case BO_LOr:
1572 return pet_op_lor;
1573 default:
1574 return pet_op_last;
1578 /* Construct a pet_expr representing a unary operator expression.
1580 struct pet_expr *PetScan::extract_expr(UnaryOperator *expr)
1582 struct pet_expr *arg;
1583 enum pet_op_type op;
1585 op = UnaryOperatorKind2pet_op_type(expr->getOpcode());
1586 if (op == pet_op_last) {
1587 unsupported(expr);
1588 return NULL;
1591 arg = extract_expr(expr->getSubExpr());
1593 if (expr->isIncrementDecrementOp() &&
1594 arg && arg->type == pet_expr_access) {
1595 mark_write(arg);
1596 arg->acc.read = 1;
1599 return pet_expr_new_unary(ctx, op, arg);
1602 /* Mark the given access pet_expr as a write.
1603 * If a scalar is being accessed, then mark its value
1604 * as unknown in assigned_value.
1606 void PetScan::mark_write(struct pet_expr *access)
1608 isl_id *id;
1609 ValueDecl *decl;
1611 if (!access)
1612 return;
1614 access->acc.write = 1;
1615 access->acc.read = 0;
1617 if (!pet_expr_is_scalar_access(access))
1618 return;
1620 id = pet_expr_access_get_id(access);
1621 decl = (ValueDecl *) isl_id_get_user(id);
1622 clear_assignment(assigned_value, decl);
1623 isl_id_free(id);
1626 /* Assign "rhs" to "lhs".
1628 * In particular, if "lhs" is a scalar variable, then mark
1629 * the variable as having been assigned. If, furthermore, "rhs"
1630 * is an affine expression, then keep track of this value in assigned_value
1631 * so that we can plug it in when we later come across the same variable.
1633 void PetScan::assign(struct pet_expr *lhs, Expr *rhs)
1635 isl_id *id;
1636 ValueDecl *decl;
1637 isl_pw_aff *pa;
1639 if (!lhs)
1640 return;
1641 if (!pet_expr_is_scalar_access(lhs))
1642 return;
1644 id = pet_expr_access_get_id(lhs);
1645 decl = (ValueDecl *) isl_id_get_user(id);
1646 isl_id_free(id);
1648 pa = try_extract_affine(rhs);
1649 clear_assignment(assigned_value, decl);
1650 if (!pa)
1651 return;
1652 assigned_value[decl] = pa;
1653 insert_expression(pa);
1656 /* Construct a pet_expr representing a binary operator expression.
1658 * If the top level operator is an assignment and the LHS is an access,
1659 * then we mark that access as a write. If the operator is a compound
1660 * assignment, the access is marked as both a read and a write.
1662 * If "expr" assigns something to a scalar variable, then we mark
1663 * the variable as having been assigned. If, furthermore, the expression
1664 * is affine, then keep track of this value in assigned_value
1665 * so that we can plug it in when we later come across the same variable.
1667 struct pet_expr *PetScan::extract_expr(BinaryOperator *expr)
1669 struct pet_expr *lhs, *rhs;
1670 enum pet_op_type op;
1672 op = BinaryOperatorKind2pet_op_type(expr->getOpcode());
1673 if (op == pet_op_last) {
1674 unsupported(expr);
1675 return NULL;
1678 lhs = extract_expr(expr->getLHS());
1679 rhs = extract_expr(expr->getRHS());
1681 if (expr->isAssignmentOp() && lhs && lhs->type == pet_expr_access) {
1682 mark_write(lhs);
1683 if (expr->isCompoundAssignmentOp())
1684 lhs->acc.read = 1;
1687 if (expr->getOpcode() == BO_Assign)
1688 assign(lhs, expr->getRHS());
1690 return pet_expr_new_binary(ctx, op, lhs, rhs);
1693 /* Construct a pet_scop with a single statement killing the entire
1694 * array "array".
1696 struct pet_scop *PetScan::kill(Stmt *stmt, struct pet_array *array)
1698 isl_id *id;
1699 isl_space *space;
1700 isl_multi_pw_aff *index;
1701 isl_map *access;
1702 struct pet_expr *expr;
1704 if (!array)
1705 return NULL;
1706 access = isl_map_from_range(isl_set_copy(array->extent));
1707 id = isl_set_get_tuple_id(array->extent);
1708 space = isl_space_alloc(ctx, 0, 0, 0);
1709 space = isl_space_set_tuple_id(space, isl_dim_out, id);
1710 index = isl_multi_pw_aff_zero(space);
1711 expr = pet_expr_kill_from_access_and_index(access, index);
1712 return extract(stmt, expr);
1715 /* Construct a pet_scop for a (single) variable declaration.
1717 * The scop contains the variable being declared (as an array)
1718 * and a statement killing the array.
1720 * If the variable is initialized in the AST, then the scop
1721 * also contains an assignment to the variable.
1723 struct pet_scop *PetScan::extract(DeclStmt *stmt)
1725 Decl *decl;
1726 VarDecl *vd;
1727 struct pet_expr *lhs, *rhs, *pe;
1728 struct pet_scop *scop_decl, *scop;
1729 struct pet_array *array;
1731 if (!stmt->isSingleDecl()) {
1732 unsupported(stmt);
1733 return NULL;
1736 decl = stmt->getSingleDecl();
1737 vd = cast<VarDecl>(decl);
1739 array = extract_array(ctx, vd, NULL);
1740 if (array)
1741 array->declared = 1;
1742 scop_decl = kill(stmt, array);
1743 scop_decl = pet_scop_add_array(scop_decl, array);
1745 if (!vd->getInit())
1746 return scop_decl;
1748 lhs = extract_access_expr(vd);
1749 rhs = extract_expr(vd->getInit());
1751 mark_write(lhs);
1752 assign(lhs, vd->getInit());
1754 pe = pet_expr_new_binary(ctx, pet_op_assign, lhs, rhs);
1755 scop = extract(stmt, pe);
1757 scop_decl = pet_scop_prefix(scop_decl, 0);
1758 scop = pet_scop_prefix(scop, 1);
1760 scop = pet_scop_add_seq(ctx, scop_decl, scop);
1762 return scop;
1765 /* Construct a pet_expr representing a conditional operation.
1767 * We first try to extract the condition as an affine expression.
1768 * If that fails, we construct a pet_expr tree representing the condition.
1770 struct pet_expr *PetScan::extract_expr(ConditionalOperator *expr)
1772 struct pet_expr *cond, *lhs, *rhs;
1773 isl_pw_aff *pa;
1775 pa = try_extract_affine(expr->getCond());
1776 if (pa) {
1777 isl_multi_pw_aff *test = isl_multi_pw_aff_from_pw_aff(pa);
1778 test = isl_multi_pw_aff_from_range(test);
1779 cond = pet_expr_from_index(test);
1780 } else
1781 cond = extract_expr(expr->getCond());
1782 lhs = extract_expr(expr->getTrueExpr());
1783 rhs = extract_expr(expr->getFalseExpr());
1785 return pet_expr_new_ternary(ctx, cond, lhs, rhs);
1788 struct pet_expr *PetScan::extract_expr(ImplicitCastExpr *expr)
1790 return extract_expr(expr->getSubExpr());
1793 /* Construct a pet_expr representing a floating point value.
1795 * If the floating point literal does not appear in a macro,
1796 * then we use the original representation in the source code
1797 * as the string representation. Otherwise, we use the pretty
1798 * printer to produce a string representation.
1800 struct pet_expr *PetScan::extract_expr(FloatingLiteral *expr)
1802 double d;
1803 string s;
1804 const LangOptions &LO = PP.getLangOpts();
1805 SourceLocation loc = expr->getLocation();
1807 if (!loc.isMacroID()) {
1808 SourceManager &SM = PP.getSourceManager();
1809 unsigned len = Lexer::MeasureTokenLength(loc, SM, LO);
1810 s = string(SM.getCharacterData(loc), len);
1811 } else {
1812 llvm::raw_string_ostream S(s);
1813 expr->printPretty(S, 0, PrintingPolicy(LO));
1814 S.str();
1816 d = expr->getValueAsApproximateDouble();
1817 return pet_expr_new_double(ctx, d, s.c_str());
1820 /* Extract an index expression from "expr" and then convert it into
1821 * an access pet_expr.
1823 struct pet_expr *PetScan::extract_access_expr(Expr *expr)
1825 isl_multi_pw_aff *index;
1826 struct pet_expr *pe;
1827 int depth;
1829 index = extract_index(expr);
1830 depth = extract_depth(index);
1832 pe = pet_expr_from_index_and_depth(index, depth);
1834 return pe;
1837 /* Extract an index expression from "decl" and then convert it into
1838 * an access pet_expr.
1840 struct pet_expr *PetScan::extract_access_expr(ValueDecl *decl)
1842 isl_multi_pw_aff *index;
1843 struct pet_expr *pe;
1844 int depth;
1846 index = extract_index(decl);
1847 depth = extract_depth(index);
1849 pe = pet_expr_from_index_and_depth(index, depth);
1851 return pe;
1854 struct pet_expr *PetScan::extract_expr(ParenExpr *expr)
1856 return extract_expr(expr->getSubExpr());
1859 /* Extract an assume statement from the argument "expr"
1860 * of a __pencil_assume statement.
1862 struct pet_expr *PetScan::extract_assume(Expr *expr)
1864 isl_pw_aff *cond;
1865 struct pet_expr *res;
1867 cond = try_extract_affine_condition(expr);
1868 if (!cond) {
1869 res = extract_expr(expr);
1870 } else {
1871 isl_multi_pw_aff *index;
1872 index = isl_multi_pw_aff_from_pw_aff(cond);
1873 index = isl_multi_pw_aff_from_range(index);
1874 res = pet_expr_from_index(index);
1876 return pet_expr_new_unary(ctx, pet_op_assume, res);
1879 /* Construct a pet_expr corresponding to the function call argument "expr".
1880 * The argument appears in position "pos" of a call to function "fd".
1882 * If we are passing along a pointer to an array element
1883 * or an entire row or even higher dimensional slice of an array,
1884 * then the function being called may write into the array.
1886 * We assume here that if the function is declared to take a pointer
1887 * to a const type, then the function will perform a read
1888 * and that otherwise, it will perform a write.
1890 struct pet_expr *PetScan::extract_argument(FunctionDecl *fd, int pos,
1891 Expr *expr)
1893 struct pet_expr *res;
1894 int is_addr = 0;
1895 pet_expr *main_arg;
1896 Stmt::StmtClass sc;
1898 if (expr->getStmtClass() == Stmt::ImplicitCastExprClass) {
1899 ImplicitCastExpr *ice = cast<ImplicitCastExpr>(expr);
1900 expr = ice->getSubExpr();
1902 if (expr->getStmtClass() == Stmt::UnaryOperatorClass) {
1903 UnaryOperator *op = cast<UnaryOperator>(expr);
1904 if (op->getOpcode() == UO_AddrOf) {
1905 is_addr = 1;
1906 expr = op->getSubExpr();
1909 res = extract_expr(expr);
1910 main_arg = res;
1911 if (is_addr)
1912 res = pet_expr_new_unary(ctx, pet_op_address_of, res);
1913 if (!res)
1914 return NULL;
1915 sc = expr->getStmtClass();
1916 if ((sc == Stmt::ArraySubscriptExprClass ||
1917 sc == Stmt::MemberExprClass) &&
1918 array_depth(expr->getType().getTypePtr()) > 0)
1919 is_addr = 1;
1920 if (is_addr && main_arg->type == pet_expr_access) {
1921 ParmVarDecl *parm;
1922 if (!fd->hasPrototype()) {
1923 report_prototype_required(expr);
1924 return pet_expr_free(res);
1926 parm = fd->getParamDecl(pos);
1927 if (!const_base(parm->getType()))
1928 mark_write(main_arg);
1931 return res;
1934 /* Construct a pet_expr representing a function call.
1936 * In the special case of a "call" to __pencil_assume,
1937 * construct an assume expression instead.
1939 struct pet_expr *PetScan::extract_expr(CallExpr *expr)
1941 struct pet_expr *res = NULL;
1942 FunctionDecl *fd;
1943 string name;
1944 unsigned n_arg;
1946 fd = expr->getDirectCallee();
1947 if (!fd) {
1948 unsupported(expr);
1949 return NULL;
1952 name = fd->getDeclName().getAsString();
1953 n_arg = expr->getNumArgs();
1955 if (n_arg == 1 && name == "__pencil_assume")
1956 return extract_assume(expr->getArg(0));
1958 res = pet_expr_new_call(ctx, name.c_str(), n_arg);
1959 if (!res)
1960 return NULL;
1962 for (int i = 0; i < n_arg; ++i) {
1963 Expr *arg = expr->getArg(i);
1964 res->args[i] = PetScan::extract_argument(fd, i, arg);
1965 if (!res->args[i])
1966 goto error;
1969 return res;
1970 error:
1971 pet_expr_free(res);
1972 return NULL;
1975 /* Construct a pet_expr representing a (C style) cast.
1977 struct pet_expr *PetScan::extract_expr(CStyleCastExpr *expr)
1979 struct pet_expr *arg;
1980 QualType type;
1982 arg = extract_expr(expr->getSubExpr());
1983 if (!arg)
1984 return NULL;
1986 type = expr->getTypeAsWritten();
1987 return pet_expr_new_cast(ctx, type.getAsString().c_str(), arg);
1990 /* Construct a pet_expr representing an integer.
1992 struct pet_expr *PetScan::extract_expr(IntegerLiteral *expr)
1994 return pet_expr_new_int(extract_int(expr));
1997 /* Try and construct a pet_expr representing "expr".
1999 struct pet_expr *PetScan::extract_expr(Expr *expr)
2001 switch (expr->getStmtClass()) {
2002 case Stmt::UnaryOperatorClass:
2003 return extract_expr(cast<UnaryOperator>(expr));
2004 case Stmt::CompoundAssignOperatorClass:
2005 case Stmt::BinaryOperatorClass:
2006 return extract_expr(cast<BinaryOperator>(expr));
2007 case Stmt::ImplicitCastExprClass:
2008 return extract_expr(cast<ImplicitCastExpr>(expr));
2009 case Stmt::ArraySubscriptExprClass:
2010 case Stmt::DeclRefExprClass:
2011 case Stmt::MemberExprClass:
2012 return extract_access_expr(expr);
2013 case Stmt::IntegerLiteralClass:
2014 return extract_expr(cast<IntegerLiteral>(expr));
2015 case Stmt::FloatingLiteralClass:
2016 return extract_expr(cast<FloatingLiteral>(expr));
2017 case Stmt::ParenExprClass:
2018 return extract_expr(cast<ParenExpr>(expr));
2019 case Stmt::ConditionalOperatorClass:
2020 return extract_expr(cast<ConditionalOperator>(expr));
2021 case Stmt::CallExprClass:
2022 return extract_expr(cast<CallExpr>(expr));
2023 case Stmt::CStyleCastExprClass:
2024 return extract_expr(cast<CStyleCastExpr>(expr));
2025 default:
2026 unsupported(expr);
2028 return NULL;
2031 /* Check if the given initialization statement is an assignment.
2032 * If so, return that assignment. Otherwise return NULL.
2034 BinaryOperator *PetScan::initialization_assignment(Stmt *init)
2036 BinaryOperator *ass;
2038 if (init->getStmtClass() != Stmt::BinaryOperatorClass)
2039 return NULL;
2041 ass = cast<BinaryOperator>(init);
2042 if (ass->getOpcode() != BO_Assign)
2043 return NULL;
2045 return ass;
2048 /* Check if the given initialization statement is a declaration
2049 * of a single variable.
2050 * If so, return that declaration. Otherwise return NULL.
2052 Decl *PetScan::initialization_declaration(Stmt *init)
2054 DeclStmt *decl;
2056 if (init->getStmtClass() != Stmt::DeclStmtClass)
2057 return NULL;
2059 decl = cast<DeclStmt>(init);
2061 if (!decl->isSingleDecl())
2062 return NULL;
2064 return decl->getSingleDecl();
2067 /* Given the assignment operator in the initialization of a for loop,
2068 * extract the induction variable, i.e., the (integer)variable being
2069 * assigned.
2071 ValueDecl *PetScan::extract_induction_variable(BinaryOperator *init)
2073 Expr *lhs;
2074 DeclRefExpr *ref;
2075 ValueDecl *decl;
2076 const Type *type;
2078 lhs = init->getLHS();
2079 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
2080 unsupported(init);
2081 return NULL;
2084 ref = cast<DeclRefExpr>(lhs);
2085 decl = ref->getDecl();
2086 type = decl->getType().getTypePtr();
2088 if (!type->isIntegerType()) {
2089 unsupported(lhs);
2090 return NULL;
2093 return decl;
2096 /* Given the initialization statement of a for loop and the single
2097 * declaration in this initialization statement,
2098 * extract the induction variable, i.e., the (integer) variable being
2099 * declared.
2101 VarDecl *PetScan::extract_induction_variable(Stmt *init, Decl *decl)
2103 VarDecl *vd;
2105 vd = cast<VarDecl>(decl);
2107 const QualType type = vd->getType();
2108 if (!type->isIntegerType()) {
2109 unsupported(init);
2110 return NULL;
2113 if (!vd->getInit()) {
2114 unsupported(init);
2115 return NULL;
2118 return vd;
2121 /* Check that op is of the form iv++ or iv--.
2122 * Return an affine expression "1" or "-1" accordingly.
2124 __isl_give isl_pw_aff *PetScan::extract_unary_increment(
2125 clang::UnaryOperator *op, clang::ValueDecl *iv)
2127 Expr *sub;
2128 DeclRefExpr *ref;
2129 isl_space *space;
2130 isl_aff *aff;
2132 if (!op->isIncrementDecrementOp()) {
2133 unsupported(op);
2134 return NULL;
2137 sub = op->getSubExpr();
2138 if (sub->getStmtClass() != Stmt::DeclRefExprClass) {
2139 unsupported(op);
2140 return NULL;
2143 ref = cast<DeclRefExpr>(sub);
2144 if (ref->getDecl() != iv) {
2145 unsupported(op);
2146 return NULL;
2149 space = isl_space_params_alloc(ctx, 0);
2150 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2152 if (op->isIncrementOp())
2153 aff = isl_aff_add_constant_si(aff, 1);
2154 else
2155 aff = isl_aff_add_constant_si(aff, -1);
2157 return isl_pw_aff_from_aff(aff);
2160 /* If the isl_pw_aff on which isl_pw_aff_foreach_piece is called
2161 * has a single constant expression, then put this constant in *user.
2162 * The caller is assumed to have checked that this function will
2163 * be called exactly once.
2165 static int extract_cst(__isl_take isl_set *set, __isl_take isl_aff *aff,
2166 void *user)
2168 isl_val **inc = (isl_val **)user;
2169 int res = 0;
2171 if (isl_aff_is_cst(aff))
2172 *inc = isl_aff_get_constant_val(aff);
2173 else
2174 res = -1;
2176 isl_set_free(set);
2177 isl_aff_free(aff);
2179 return res;
2182 /* Check if op is of the form
2184 * iv = iv + inc
2186 * and return inc as an affine expression.
2188 * We extract an affine expression from the RHS, subtract iv and return
2189 * the result.
2191 __isl_give isl_pw_aff *PetScan::extract_binary_increment(BinaryOperator *op,
2192 clang::ValueDecl *iv)
2194 Expr *lhs;
2195 DeclRefExpr *ref;
2196 isl_id *id;
2197 isl_space *dim;
2198 isl_aff *aff;
2199 isl_pw_aff *val;
2201 if (op->getOpcode() != BO_Assign) {
2202 unsupported(op);
2203 return NULL;
2206 lhs = op->getLHS();
2207 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
2208 unsupported(op);
2209 return NULL;
2212 ref = cast<DeclRefExpr>(lhs);
2213 if (ref->getDecl() != iv) {
2214 unsupported(op);
2215 return NULL;
2218 val = extract_affine(op->getRHS());
2220 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
2222 dim = isl_space_params_alloc(ctx, 1);
2223 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
2224 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2225 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2227 val = isl_pw_aff_sub(val, isl_pw_aff_from_aff(aff));
2229 return val;
2232 /* Check that op is of the form iv += cst or iv -= cst
2233 * and return an affine expression corresponding oto cst or -cst accordingly.
2235 __isl_give isl_pw_aff *PetScan::extract_compound_increment(
2236 CompoundAssignOperator *op, clang::ValueDecl *iv)
2238 Expr *lhs;
2239 DeclRefExpr *ref;
2240 bool neg = false;
2241 isl_pw_aff *val;
2242 BinaryOperatorKind opcode;
2244 opcode = op->getOpcode();
2245 if (opcode != BO_AddAssign && opcode != BO_SubAssign) {
2246 unsupported(op);
2247 return NULL;
2249 if (opcode == BO_SubAssign)
2250 neg = true;
2252 lhs = op->getLHS();
2253 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
2254 unsupported(op);
2255 return NULL;
2258 ref = cast<DeclRefExpr>(lhs);
2259 if (ref->getDecl() != iv) {
2260 unsupported(op);
2261 return NULL;
2264 val = extract_affine(op->getRHS());
2265 if (neg)
2266 val = isl_pw_aff_neg(val);
2268 return val;
2271 /* Check that the increment of the given for loop increments
2272 * (or decrements) the induction variable "iv" and return
2273 * the increment as an affine expression if successful.
2275 __isl_give isl_pw_aff *PetScan::extract_increment(clang::ForStmt *stmt,
2276 ValueDecl *iv)
2278 Stmt *inc = stmt->getInc();
2280 if (!inc) {
2281 report_missing_increment(stmt);
2282 return NULL;
2285 if (inc->getStmtClass() == Stmt::UnaryOperatorClass)
2286 return extract_unary_increment(cast<UnaryOperator>(inc), iv);
2287 if (inc->getStmtClass() == Stmt::CompoundAssignOperatorClass)
2288 return extract_compound_increment(
2289 cast<CompoundAssignOperator>(inc), iv);
2290 if (inc->getStmtClass() == Stmt::BinaryOperatorClass)
2291 return extract_binary_increment(cast<BinaryOperator>(inc), iv);
2293 unsupported(inc);
2294 return NULL;
2297 /* Embed the given iteration domain in an extra outer loop
2298 * with induction variable "var".
2299 * If this variable appeared as a parameter in the constraints,
2300 * it is replaced by the new outermost dimension.
2302 static __isl_give isl_set *embed(__isl_take isl_set *set,
2303 __isl_take isl_id *var)
2305 int pos;
2307 set = isl_set_insert_dims(set, isl_dim_set, 0, 1);
2308 pos = isl_set_find_dim_by_id(set, isl_dim_param, var);
2309 if (pos >= 0) {
2310 set = isl_set_equate(set, isl_dim_param, pos, isl_dim_set, 0);
2311 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2314 isl_id_free(var);
2315 return set;
2318 /* Return those elements in the space of "cond" that come after
2319 * (based on "sign") an element in "cond".
2321 static __isl_give isl_set *after(__isl_take isl_set *cond, int sign)
2323 isl_map *previous_to_this;
2325 if (sign > 0)
2326 previous_to_this = isl_map_lex_lt(isl_set_get_space(cond));
2327 else
2328 previous_to_this = isl_map_lex_gt(isl_set_get_space(cond));
2330 cond = isl_set_apply(cond, previous_to_this);
2332 return cond;
2335 /* Create the infinite iteration domain
2337 * { [id] : id >= 0 }
2339 * If "scop" has an affine skip of type pet_skip_later,
2340 * then remove those iterations i that have an earlier iteration
2341 * where the skip condition is satisfied, meaning that iteration i
2342 * is not executed.
2343 * Since we are dealing with a loop without loop iterator,
2344 * the skip condition cannot refer to the current loop iterator and
2345 * so effectively, the returned set is of the form
2347 * { [0]; [id] : id >= 1 and not skip }
2349 static __isl_give isl_set *infinite_domain(__isl_take isl_id *id,
2350 struct pet_scop *scop)
2352 isl_ctx *ctx = isl_id_get_ctx(id);
2353 isl_set *domain;
2354 isl_set *skip;
2356 domain = isl_set_nat_universe(isl_space_set_alloc(ctx, 0, 1));
2357 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, id);
2359 if (!pet_scop_has_affine_skip(scop, pet_skip_later))
2360 return domain;
2362 skip = pet_scop_get_affine_skip_domain(scop, pet_skip_later);
2363 skip = embed(skip, isl_id_copy(id));
2364 skip = isl_set_intersect(skip , isl_set_copy(domain));
2365 domain = isl_set_subtract(domain, after(skip, 1));
2367 return domain;
2370 /* Create an identity affine expression on the space containing "domain",
2371 * which is assumed to be one-dimensional.
2373 static __isl_give isl_aff *identity_aff(__isl_keep isl_set *domain)
2375 isl_local_space *ls;
2377 ls = isl_local_space_from_space(isl_set_get_space(domain));
2378 return isl_aff_var_on_domain(ls, isl_dim_set, 0);
2381 /* Create an affine expression that maps elements
2382 * of a single-dimensional array "id_test" to the previous element
2383 * (according to "inc"), provided this element belongs to "domain".
2384 * That is, create the affine expression
2386 * { id[x] -> id[x - inc] : x - inc in domain }
2388 static __isl_give isl_multi_pw_aff *map_to_previous(__isl_take isl_id *id_test,
2389 __isl_take isl_set *domain, __isl_take isl_val *inc)
2391 isl_space *space;
2392 isl_local_space *ls;
2393 isl_aff *aff;
2394 isl_multi_pw_aff *prev;
2396 space = isl_set_get_space(domain);
2397 ls = isl_local_space_from_space(space);
2398 aff = isl_aff_var_on_domain(ls, isl_dim_set, 0);
2399 aff = isl_aff_add_constant_val(aff, isl_val_neg(inc));
2400 prev = isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
2401 domain = isl_set_preimage_multi_pw_aff(domain,
2402 isl_multi_pw_aff_copy(prev));
2403 prev = isl_multi_pw_aff_intersect_domain(prev, domain);
2404 prev = isl_multi_pw_aff_set_tuple_id(prev, isl_dim_out, id_test);
2406 return prev;
2409 /* Add an implication to "scop" expressing that if an element of
2410 * virtual array "id_test" has value "satisfied" then all previous elements
2411 * of this array also have that value. The set of previous elements
2412 * is bounded by "domain". If "sign" is negative then the iterator
2413 * is decreasing and we express that all subsequent array elements
2414 * (but still defined previously) have the same value.
2416 static struct pet_scop *add_implication(struct pet_scop *scop,
2417 __isl_take isl_id *id_test, __isl_take isl_set *domain, int sign,
2418 int satisfied)
2420 isl_space *space;
2421 isl_map *map;
2423 domain = isl_set_set_tuple_id(domain, id_test);
2424 space = isl_set_get_space(domain);
2425 if (sign > 0)
2426 map = isl_map_lex_ge(space);
2427 else
2428 map = isl_map_lex_le(space);
2429 map = isl_map_intersect_range(map, domain);
2430 scop = pet_scop_add_implication(scop, map, satisfied);
2432 return scop;
2435 /* Add a filter to "scop" that imposes that it is only executed
2436 * when the variable identified by "id_test" has a zero value
2437 * for all previous iterations of "domain".
2439 * In particular, add a filter that imposes that the array
2440 * has a zero value at the previous iteration of domain and
2441 * add an implication that implies that it then has that
2442 * value for all previous iterations.
2444 static struct pet_scop *scop_add_break(struct pet_scop *scop,
2445 __isl_take isl_id *id_test, __isl_take isl_set *domain,
2446 __isl_take isl_val *inc)
2448 isl_multi_pw_aff *prev;
2449 int sign = isl_val_sgn(inc);
2451 prev = map_to_previous(isl_id_copy(id_test), isl_set_copy(domain), inc);
2452 scop = add_implication(scop, id_test, domain, sign, 0);
2453 scop = pet_scop_filter(scop, prev, 0);
2455 return scop;
2458 /* Construct a pet_scop for an infinite loop around the given body.
2460 * We extract a pet_scop for the body and then embed it in a loop with
2461 * iteration domain
2463 * { [t] : t >= 0 }
2465 * and schedule
2467 * { [t] -> [t] }
2469 * If the body contains any break, then it is taken into
2470 * account in infinite_domain (if the skip condition is affine)
2471 * or in scop_add_break (if the skip condition is not affine).
2473 * If we were only able to extract part of the body, then simply
2474 * return that part.
2476 struct pet_scop *PetScan::extract_infinite_loop(Stmt *body)
2478 isl_id *id, *id_test;
2479 isl_set *domain;
2480 isl_aff *ident;
2481 struct pet_scop *scop;
2482 bool has_var_break;
2484 scop = extract(body);
2485 if (!scop)
2486 return NULL;
2487 if (partial)
2488 return scop;
2490 id = isl_id_alloc(ctx, "t", NULL);
2491 domain = infinite_domain(isl_id_copy(id), scop);
2492 ident = identity_aff(domain);
2494 has_var_break = pet_scop_has_var_skip(scop, pet_skip_later);
2495 if (has_var_break)
2496 id_test = pet_scop_get_skip_id(scop, pet_skip_later);
2498 scop = pet_scop_embed(scop, isl_set_copy(domain),
2499 isl_aff_copy(ident), ident, id);
2500 if (has_var_break)
2501 scop = scop_add_break(scop, id_test, domain, isl_val_one(ctx));
2502 else
2503 isl_set_free(domain);
2505 return scop;
2508 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
2510 * for (;;)
2511 * body
2514 struct pet_scop *PetScan::extract_infinite_for(ForStmt *stmt)
2516 clear_assignments clear(assigned_value);
2517 clear.TraverseStmt(stmt->getBody());
2519 return extract_infinite_loop(stmt->getBody());
2522 /* Create an index expression for an access to a virtual array
2523 * representing the result of a condition.
2524 * Unlike other accessed data, the id of the array is NULL as
2525 * there is no ValueDecl in the program corresponding to the virtual
2526 * array.
2527 * The array starts out as a scalar, but grows along with the
2528 * statement writing to the array in pet_scop_embed.
2530 static __isl_give isl_multi_pw_aff *create_test_index(isl_ctx *ctx, int test_nr)
2532 isl_space *dim = isl_space_alloc(ctx, 0, 0, 0);
2533 isl_id *id;
2534 char name[50];
2536 snprintf(name, sizeof(name), "__pet_test_%d", test_nr);
2537 id = isl_id_alloc(ctx, name, NULL);
2538 dim = isl_space_set_tuple_id(dim, isl_dim_out, id);
2539 return isl_multi_pw_aff_zero(dim);
2542 /* Add an array with the given extent (range of "index") to the list
2543 * of arrays in "scop" and return the extended pet_scop.
2544 * The array is marked as attaining values 0 and 1 only and
2545 * as each element being assigned at most once.
2547 static struct pet_scop *scop_add_array(struct pet_scop *scop,
2548 __isl_keep isl_multi_pw_aff *index, clang::ASTContext &ast_ctx)
2550 isl_ctx *ctx = isl_multi_pw_aff_get_ctx(index);
2551 isl_space *dim;
2552 struct pet_array *array;
2553 isl_map *access;
2555 if (!scop)
2556 return NULL;
2557 if (!ctx)
2558 goto error;
2560 array = isl_calloc_type(ctx, struct pet_array);
2561 if (!array)
2562 goto error;
2564 access = isl_map_from_multi_pw_aff(isl_multi_pw_aff_copy(index));
2565 array->extent = isl_map_range(access);
2566 dim = isl_space_params_alloc(ctx, 0);
2567 array->context = isl_set_universe(dim);
2568 dim = isl_space_set_alloc(ctx, 0, 1);
2569 array->value_bounds = isl_set_universe(dim);
2570 array->value_bounds = isl_set_lower_bound_si(array->value_bounds,
2571 isl_dim_set, 0, 0);
2572 array->value_bounds = isl_set_upper_bound_si(array->value_bounds,
2573 isl_dim_set, 0, 1);
2574 array->element_type = strdup("int");
2575 array->element_size = ast_ctx.getTypeInfo(ast_ctx.IntTy).first / 8;
2576 array->uniquely_defined = 1;
2578 if (!array->extent || !array->context)
2579 array = pet_array_free(array);
2581 scop = pet_scop_add_array(scop, array);
2583 return scop;
2584 error:
2585 pet_scop_free(scop);
2586 return NULL;
2589 /* Construct a pet_scop for a while loop of the form
2591 * while (pa)
2592 * body
2594 * In particular, construct a scop for an infinite loop around body and
2595 * intersect the domain with the affine expression.
2596 * Note that this intersection may result in an empty loop.
2598 struct pet_scop *PetScan::extract_affine_while(__isl_take isl_pw_aff *pa,
2599 Stmt *body)
2601 struct pet_scop *scop;
2602 isl_set *dom;
2603 isl_set *valid;
2605 valid = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2606 dom = isl_pw_aff_non_zero_set(pa);
2607 scop = extract_infinite_loop(body);
2608 scop = pet_scop_restrict(scop, dom);
2609 scop = pet_scop_restrict_context(scop, valid);
2611 return scop;
2614 /* Construct a scop for a while, given the scops for the condition
2615 * and the body, the filter identifier and the iteration domain of
2616 * the while loop.
2618 * In particular, the scop for the condition is filtered to depend
2619 * on "id_test" evaluating to true for all previous iterations
2620 * of the loop, while the scop for the body is filtered to depend
2621 * on "id_test" evaluating to true for all iterations up to the
2622 * current iteration.
2623 * The actual filter only imposes that this virtual array has
2624 * value one on the previous or the current iteration.
2625 * The fact that this condition also applies to the previous
2626 * iterations is enforced by an implication.
2628 * These filtered scops are then combined into a single scop.
2630 * "sign" is positive if the iterator increases and negative
2631 * if it decreases.
2633 static struct pet_scop *scop_add_while(struct pet_scop *scop_cond,
2634 struct pet_scop *scop_body, __isl_take isl_id *id_test,
2635 __isl_take isl_set *domain, __isl_take isl_val *inc)
2637 isl_ctx *ctx = isl_set_get_ctx(domain);
2638 isl_space *space;
2639 isl_multi_pw_aff *test_index;
2640 isl_multi_pw_aff *prev;
2641 int sign = isl_val_sgn(inc);
2642 struct pet_scop *scop;
2644 prev = map_to_previous(isl_id_copy(id_test), isl_set_copy(domain), inc);
2645 scop_cond = pet_scop_filter(scop_cond, prev, 1);
2647 space = isl_space_map_from_set(isl_set_get_space(domain));
2648 test_index = isl_multi_pw_aff_identity(space);
2649 test_index = isl_multi_pw_aff_set_tuple_id(test_index, isl_dim_out,
2650 isl_id_copy(id_test));
2651 scop_body = pet_scop_filter(scop_body, test_index, 1);
2653 scop = pet_scop_add_seq(ctx, scop_cond, scop_body);
2654 scop = add_implication(scop, id_test, domain, sign, 1);
2656 return scop;
2659 /* Check if the while loop is of the form
2661 * while (affine expression)
2662 * body
2664 * If so, call extract_affine_while to construct a scop.
2666 * Otherwise, construct a generic while scop, with iteration domain
2667 * { [t] : t >= 0 }. The scop consists of two parts, one for
2668 * evaluating the condition and one for the body.
2669 * The schedule is adjusted to reflect that the condition is evaluated
2670 * before the body is executed and the body is filtered to depend
2671 * on the result of the condition evaluating to true on all iterations
2672 * up to the current iteration, while the evaluation of the condition itself
2673 * is filtered to depend on the result of the condition evaluating to true
2674 * on all previous iterations.
2675 * The context of the scop representing the body is dropped
2676 * because we don't know how many times the body will be executed,
2677 * if at all.
2679 * If the body contains any break, then it is taken into
2680 * account in infinite_domain (if the skip condition is affine)
2681 * or in scop_add_break (if the skip condition is not affine).
2683 * If we were only able to extract part of the body, then simply
2684 * return that part.
2686 struct pet_scop *PetScan::extract(WhileStmt *stmt)
2688 Expr *cond;
2689 int test_nr, stmt_nr;
2690 isl_id *id, *id_test, *id_break_test;
2691 isl_multi_pw_aff *test_index;
2692 isl_set *domain;
2693 isl_aff *ident;
2694 isl_pw_aff *pa;
2695 struct pet_scop *scop, *scop_body;
2696 bool has_var_break;
2698 cond = stmt->getCond();
2699 if (!cond) {
2700 unsupported(stmt);
2701 return NULL;
2704 clear_assignments clear(assigned_value);
2705 clear.TraverseStmt(stmt->getBody());
2707 pa = try_extract_affine_condition(cond);
2708 if (pa)
2709 return extract_affine_while(pa, stmt->getBody());
2711 if (!allow_nested) {
2712 unsupported(stmt);
2713 return NULL;
2716 test_nr = n_test++;
2717 stmt_nr = n_stmt++;
2718 scop_body = extract(stmt->getBody());
2719 if (partial)
2720 return scop_body;
2722 test_index = create_test_index(ctx, test_nr);
2723 scop = extract_non_affine_condition(cond, stmt_nr,
2724 isl_multi_pw_aff_copy(test_index));
2725 scop = scop_add_array(scop, test_index, ast_context);
2726 id_test = isl_multi_pw_aff_get_tuple_id(test_index, isl_dim_out);
2727 isl_multi_pw_aff_free(test_index);
2729 id = isl_id_alloc(ctx, "t", NULL);
2730 domain = infinite_domain(isl_id_copy(id), scop_body);
2731 ident = identity_aff(domain);
2733 has_var_break = pet_scop_has_var_skip(scop_body, pet_skip_later);
2734 if (has_var_break)
2735 id_break_test = pet_scop_get_skip_id(scop_body, pet_skip_later);
2737 scop = pet_scop_prefix(scop, 0);
2738 scop = pet_scop_embed(scop, isl_set_copy(domain), isl_aff_copy(ident),
2739 isl_aff_copy(ident), isl_id_copy(id));
2740 scop_body = pet_scop_reset_context(scop_body);
2741 scop_body = pet_scop_prefix(scop_body, 1);
2742 scop_body = pet_scop_embed(scop_body, isl_set_copy(domain),
2743 isl_aff_copy(ident), ident, id);
2745 if (has_var_break) {
2746 scop = scop_add_break(scop, isl_id_copy(id_break_test),
2747 isl_set_copy(domain), isl_val_one(ctx));
2748 scop_body = scop_add_break(scop_body, id_break_test,
2749 isl_set_copy(domain), isl_val_one(ctx));
2751 scop = scop_add_while(scop, scop_body, id_test, domain,
2752 isl_val_one(ctx));
2754 return scop;
2757 /* Check whether "cond" expresses a simple loop bound
2758 * on the only set dimension.
2759 * In particular, if "up" is set then "cond" should contain only
2760 * upper bounds on the set dimension.
2761 * Otherwise, it should contain only lower bounds.
2763 static bool is_simple_bound(__isl_keep isl_set *cond, __isl_keep isl_val *inc)
2765 if (isl_val_is_pos(inc))
2766 return !isl_set_dim_has_any_lower_bound(cond, isl_dim_set, 0);
2767 else
2768 return !isl_set_dim_has_any_upper_bound(cond, isl_dim_set, 0);
2771 /* Extend a condition on a given iteration of a loop to one that
2772 * imposes the same condition on all previous iterations.
2773 * "domain" expresses the lower [upper] bound on the iterations
2774 * when inc is positive [negative].
2776 * In particular, we construct the condition (when inc is positive)
2778 * forall i' : (domain(i') and i' <= i) => cond(i')
2780 * which is equivalent to
2782 * not exists i' : domain(i') and i' <= i and not cond(i')
2784 * We construct this set by negating cond, applying a map
2786 * { [i'] -> [i] : domain(i') and i' <= i }
2788 * and then negating the result again.
2790 static __isl_give isl_set *valid_for_each_iteration(__isl_take isl_set *cond,
2791 __isl_take isl_set *domain, __isl_take isl_val *inc)
2793 isl_map *previous_to_this;
2795 if (isl_val_is_pos(inc))
2796 previous_to_this = isl_map_lex_le(isl_set_get_space(domain));
2797 else
2798 previous_to_this = isl_map_lex_ge(isl_set_get_space(domain));
2800 previous_to_this = isl_map_intersect_domain(previous_to_this, domain);
2802 cond = isl_set_complement(cond);
2803 cond = isl_set_apply(cond, previous_to_this);
2804 cond = isl_set_complement(cond);
2806 isl_val_free(inc);
2808 return cond;
2811 /* Construct a domain of the form
2813 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
2815 static __isl_give isl_set *strided_domain(__isl_take isl_id *id,
2816 __isl_take isl_pw_aff *init, __isl_take isl_val *inc)
2818 isl_aff *aff;
2819 isl_space *dim;
2820 isl_set *set;
2822 init = isl_pw_aff_insert_dims(init, isl_dim_in, 0, 1);
2823 dim = isl_pw_aff_get_domain_space(init);
2824 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2825 aff = isl_aff_add_coefficient_val(aff, isl_dim_in, 0, inc);
2826 init = isl_pw_aff_add(init, isl_pw_aff_from_aff(aff));
2828 dim = isl_space_set_alloc(isl_pw_aff_get_ctx(init), 1, 1);
2829 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
2830 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2831 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2833 set = isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff), init);
2835 set = isl_set_lower_bound_si(set, isl_dim_set, 0, 0);
2837 return isl_set_params(set);
2840 /* Assuming "cond" represents a bound on a loop where the loop
2841 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2842 * is possible.
2844 * Under the given assumptions, wrapping is only possible if "cond" allows
2845 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2846 * increasing iterator and 0 in case of a decreasing iterator.
2848 static bool can_wrap(__isl_keep isl_set *cond, ValueDecl *iv,
2849 __isl_keep isl_val *inc)
2851 bool cw;
2852 isl_ctx *ctx;
2853 isl_val *limit;
2854 isl_set *test;
2856 test = isl_set_copy(cond);
2858 ctx = isl_set_get_ctx(test);
2859 if (isl_val_is_neg(inc))
2860 limit = isl_val_zero(ctx);
2861 else {
2862 limit = isl_val_int_from_ui(ctx, get_type_size(iv));
2863 limit = isl_val_2exp(limit);
2864 limit = isl_val_sub_ui(limit, 1);
2867 test = isl_set_fix_val(cond, isl_dim_set, 0, limit);
2868 cw = !isl_set_is_empty(test);
2869 isl_set_free(test);
2871 return cw;
2874 /* Given a one-dimensional space, construct the following affine expression
2875 * on this space
2877 * { [v] -> [v mod 2^width] }
2879 * where width is the number of bits used to represent the values
2880 * of the unsigned variable "iv".
2882 static __isl_give isl_aff *compute_wrapping(__isl_take isl_space *dim,
2883 ValueDecl *iv)
2885 isl_ctx *ctx;
2886 isl_val *mod;
2887 isl_aff *aff;
2889 ctx = isl_space_get_ctx(dim);
2890 mod = isl_val_int_from_ui(ctx, get_type_size(iv));
2891 mod = isl_val_2exp(mod);
2893 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2894 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2895 aff = isl_aff_mod_val(aff, mod);
2897 return aff;
2900 /* Project out the parameter "id" from "set".
2902 static __isl_give isl_set *set_project_out_by_id(__isl_take isl_set *set,
2903 __isl_keep isl_id *id)
2905 int pos;
2907 pos = isl_set_find_dim_by_id(set, isl_dim_param, id);
2908 if (pos >= 0)
2909 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2911 return set;
2914 /* Compute the set of parameters for which "set1" is a subset of "set2".
2916 * set1 is a subset of set2 if
2918 * forall i in set1 : i in set2
2920 * or
2922 * not exists i in set1 and i not in set2
2924 * i.e.,
2926 * not exists i in set1 \ set2
2928 static __isl_give isl_set *enforce_subset(__isl_take isl_set *set1,
2929 __isl_take isl_set *set2)
2931 return isl_set_complement(isl_set_params(isl_set_subtract(set1, set2)));
2934 /* Compute the set of parameter values for which "cond" holds
2935 * on the next iteration for each element of "dom".
2937 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2938 * and then compute the set of parameters for which the result is a subset
2939 * of "cond".
2941 static __isl_give isl_set *valid_on_next(__isl_take isl_set *cond,
2942 __isl_take isl_set *dom, __isl_take isl_val *inc)
2944 isl_space *space;
2945 isl_aff *aff;
2946 isl_map *next;
2948 space = isl_set_get_space(dom);
2949 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2950 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2951 aff = isl_aff_add_constant_val(aff, inc);
2952 next = isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2954 dom = isl_set_apply(dom, next);
2956 return enforce_subset(dom, cond);
2959 /* Construct a pet_scop for a for statement.
2960 * The for loop is required to be of the form
2962 * for (i = init; condition; ++i)
2964 * or
2966 * for (i = init; condition; --i)
2968 * The initialization of the for loop should either be an assignment
2969 * to an integer variable, or a declaration of such a variable with
2970 * initialization.
2972 * The condition is allowed to contain nested accesses, provided
2973 * they are not being written to inside the body of the loop.
2974 * Otherwise, or if the condition is otherwise non-affine, the for loop is
2975 * essentially treated as a while loop, with iteration domain
2976 * { [i] : i >= init }.
2978 * We extract a pet_scop for the body and then embed it in a loop with
2979 * iteration domain and schedule
2981 * { [i] : i >= init and condition' }
2982 * { [i] -> [i] }
2984 * or
2986 * { [i] : i <= init and condition' }
2987 * { [i] -> [-i] }
2989 * Where condition' is equal to condition if the latter is
2990 * a simple upper [lower] bound and a condition that is extended
2991 * to apply to all previous iterations otherwise.
2993 * If the condition is non-affine, then we drop the condition from the
2994 * iteration domain and instead create a separate statement
2995 * for evaluating the condition. The body is then filtered to depend
2996 * on the result of the condition evaluating to true on all iterations
2997 * up to the current iteration, while the evaluation the condition itself
2998 * is filtered to depend on the result of the condition evaluating to true
2999 * on all previous iterations.
3000 * The context of the scop representing the body is dropped
3001 * because we don't know how many times the body will be executed,
3002 * if at all.
3004 * If the stride of the loop is not 1, then "i >= init" is replaced by
3006 * (exists a: i = init + stride * a and a >= 0)
3008 * If the loop iterator i is unsigned, then wrapping may occur.
3009 * We therefore use a virtual iterator instead that does not wrap.
3010 * However, the condition in the code applies
3011 * to the wrapped value, so we need to change condition(i)
3012 * into condition([i % 2^width]). Similarly, we replace all accesses
3013 * to the original iterator by the wrapping of the virtual iterator.
3014 * Note that there may be no need to perform this final wrapping
3015 * if the loop condition (after wrapping) satisfies certain conditions.
3016 * However, the is_simple_bound condition is not enough since it doesn't
3017 * check if there even is an upper bound.
3019 * Wrapping on unsigned iterators can be avoided entirely if
3020 * loop condition is simple, the loop iterator is incremented
3021 * [decremented] by one and the last value before wrapping cannot
3022 * possibly satisfy the loop condition.
3024 * Before extracting a pet_scop from the body we remove all
3025 * assignments in assigned_value to variables that are assigned
3026 * somewhere in the body of the loop.
3028 * Valid parameters for a for loop are those for which the initial
3029 * value itself, the increment on each domain iteration and
3030 * the condition on both the initial value and
3031 * the result of incrementing the iterator for each iteration of the domain
3032 * can be evaluated.
3033 * If the loop condition is non-affine, then we only consider validity
3034 * of the initial value.
3036 * If the body contains any break, then we keep track of it in "skip"
3037 * (if the skip condition is affine) or it is handled in scop_add_break
3038 * (if the skip condition is not affine).
3039 * Note that the affine break condition needs to be considered with
3040 * respect to previous iterations in the virtual domain (if any).
3042 * If we were only able to extract part of the body, then simply
3043 * return that part.
3045 struct pet_scop *PetScan::extract_for(ForStmt *stmt)
3047 BinaryOperator *ass;
3048 Decl *decl;
3049 Stmt *init;
3050 Expr *lhs, *rhs;
3051 ValueDecl *iv;
3052 isl_local_space *ls;
3053 isl_set *domain;
3054 isl_aff *sched;
3055 isl_set *cond = NULL;
3056 isl_set *skip = NULL;
3057 isl_id *id, *id_test = NULL, *id_break_test;
3058 struct pet_scop *scop, *scop_cond = NULL;
3059 assigned_value_cache cache(assigned_value);
3060 isl_val *inc;
3061 bool was_assigned;
3062 bool is_one;
3063 bool is_unsigned;
3064 bool is_simple;
3065 bool is_virtual;
3066 bool has_affine_break;
3067 bool has_var_break;
3068 isl_aff *wrap = NULL;
3069 isl_pw_aff *pa, *pa_inc, *init_val;
3070 isl_set *valid_init;
3071 isl_set *valid_cond;
3072 isl_set *valid_cond_init;
3073 isl_set *valid_cond_next;
3074 isl_set *valid_inc;
3075 int stmt_id;
3077 if (!stmt->getInit() && !stmt->getCond() && !stmt->getInc())
3078 return extract_infinite_for(stmt);
3080 init = stmt->getInit();
3081 if (!init) {
3082 unsupported(stmt);
3083 return NULL;
3085 if ((ass = initialization_assignment(init)) != NULL) {
3086 iv = extract_induction_variable(ass);
3087 if (!iv)
3088 return NULL;
3089 lhs = ass->getLHS();
3090 rhs = ass->getRHS();
3091 } else if ((decl = initialization_declaration(init)) != NULL) {
3092 VarDecl *var = extract_induction_variable(init, decl);
3093 if (!var)
3094 return NULL;
3095 iv = var;
3096 rhs = var->getInit();
3097 lhs = create_DeclRefExpr(var);
3098 } else {
3099 unsupported(stmt->getInit());
3100 return NULL;
3103 assigned_value.erase(iv);
3104 clear_assignments clear(assigned_value);
3105 clear.TraverseStmt(stmt->getBody());
3107 was_assigned = assigned_value.find(iv) != assigned_value.end();
3108 clear_assignment(assigned_value, iv);
3109 init_val = extract_affine(rhs);
3110 if (!was_assigned)
3111 assigned_value.erase(iv);
3112 if (!init_val)
3113 return NULL;
3115 pa_inc = extract_increment(stmt, iv);
3116 if (!pa_inc) {
3117 isl_pw_aff_free(init_val);
3118 return NULL;
3121 inc = NULL;
3122 if (isl_pw_aff_n_piece(pa_inc) != 1 ||
3123 isl_pw_aff_foreach_piece(pa_inc, &extract_cst, &inc) < 0) {
3124 isl_pw_aff_free(init_val);
3125 isl_pw_aff_free(pa_inc);
3126 unsupported(stmt->getInc());
3127 isl_val_free(inc);
3128 return NULL;
3131 pa = try_extract_nested_condition(stmt->getCond());
3132 if (allow_nested && (!pa || pet_nested_any_in_pw_aff(pa)))
3133 stmt_id = n_stmt++;
3135 scop = extract(stmt->getBody());
3136 if (partial) {
3137 isl_pw_aff_free(init_val);
3138 isl_pw_aff_free(pa_inc);
3139 isl_pw_aff_free(pa);
3140 isl_val_free(inc);
3141 return scop;
3144 valid_inc = isl_pw_aff_domain(pa_inc);
3146 is_unsigned = iv->getType()->isUnsignedIntegerType();
3148 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
3150 has_affine_break = scop &&
3151 pet_scop_has_affine_skip(scop, pet_skip_later);
3152 if (has_affine_break)
3153 skip = pet_scop_get_affine_skip_domain(scop, pet_skip_later);
3154 has_var_break = scop && pet_scop_has_var_skip(scop, pet_skip_later);
3155 if (has_var_break)
3156 id_break_test = pet_scop_get_skip_id(scop, pet_skip_later);
3158 if (pa && !is_nested_allowed(pa, scop)) {
3159 isl_pw_aff_free(pa);
3160 pa = NULL;
3163 if (!allow_nested && !pa)
3164 pa = try_extract_affine_condition(stmt->getCond());
3165 valid_cond = isl_pw_aff_domain(isl_pw_aff_copy(pa));
3166 cond = isl_pw_aff_non_zero_set(pa);
3167 if (allow_nested && !cond) {
3168 isl_multi_pw_aff *test_index;
3169 int save_n_stmt = n_stmt;
3170 test_index = create_test_index(ctx, n_test++);
3171 n_stmt = stmt_id;
3172 scop_cond = extract_non_affine_condition(stmt->getCond(),
3173 n_stmt++, isl_multi_pw_aff_copy(test_index));
3174 n_stmt = save_n_stmt;
3175 scop_cond = scop_add_array(scop_cond, test_index, ast_context);
3176 id_test = isl_multi_pw_aff_get_tuple_id(test_index,
3177 isl_dim_out);
3178 isl_multi_pw_aff_free(test_index);
3179 scop_cond = pet_scop_prefix(scop_cond, 0);
3180 scop = pet_scop_reset_context(scop);
3181 scop = pet_scop_prefix(scop, 1);
3182 cond = isl_set_universe(isl_space_set_alloc(ctx, 0, 0));
3185 cond = embed(cond, isl_id_copy(id));
3186 skip = embed(skip, isl_id_copy(id));
3187 valid_cond = isl_set_coalesce(valid_cond);
3188 valid_cond = embed(valid_cond, isl_id_copy(id));
3189 valid_inc = embed(valid_inc, isl_id_copy(id));
3190 is_one = isl_val_is_one(inc) || isl_val_is_negone(inc);
3191 is_virtual = is_unsigned && (!is_one || can_wrap(cond, iv, inc));
3193 valid_cond_init = enforce_subset(
3194 isl_set_from_pw_aff(isl_pw_aff_copy(init_val)),
3195 isl_set_copy(valid_cond));
3196 if (is_one && !is_virtual) {
3197 isl_pw_aff_free(init_val);
3198 pa = extract_comparison(isl_val_is_pos(inc) ? BO_GE : BO_LE,
3199 lhs, rhs, init);
3200 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(pa));
3201 valid_init = set_project_out_by_id(valid_init, id);
3202 domain = isl_pw_aff_non_zero_set(pa);
3203 } else {
3204 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(init_val));
3205 domain = strided_domain(isl_id_copy(id), init_val,
3206 isl_val_copy(inc));
3209 domain = embed(domain, isl_id_copy(id));
3210 if (is_virtual) {
3211 isl_map *rev_wrap;
3212 wrap = compute_wrapping(isl_set_get_space(cond), iv);
3213 rev_wrap = isl_map_from_aff(isl_aff_copy(wrap));
3214 rev_wrap = isl_map_reverse(rev_wrap);
3215 cond = isl_set_apply(cond, isl_map_copy(rev_wrap));
3216 skip = isl_set_apply(skip, isl_map_copy(rev_wrap));
3217 valid_cond = isl_set_apply(valid_cond, isl_map_copy(rev_wrap));
3218 valid_inc = isl_set_apply(valid_inc, rev_wrap);
3220 is_simple = is_simple_bound(cond, inc);
3221 if (!is_simple) {
3222 cond = isl_set_gist(cond, isl_set_copy(domain));
3223 is_simple = is_simple_bound(cond, inc);
3225 if (!is_simple)
3226 cond = valid_for_each_iteration(cond,
3227 isl_set_copy(domain), isl_val_copy(inc));
3228 domain = isl_set_intersect(domain, cond);
3229 if (has_affine_break) {
3230 skip = isl_set_intersect(skip , isl_set_copy(domain));
3231 skip = after(skip, isl_val_sgn(inc));
3232 domain = isl_set_subtract(domain, skip);
3234 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, isl_id_copy(id));
3235 ls = isl_local_space_from_space(isl_set_get_space(domain));
3236 sched = isl_aff_var_on_domain(ls, isl_dim_set, 0);
3237 if (isl_val_is_neg(inc))
3238 sched = isl_aff_neg(sched);
3240 valid_cond_next = valid_on_next(valid_cond, isl_set_copy(domain),
3241 isl_val_copy(inc));
3242 valid_inc = enforce_subset(isl_set_copy(domain), valid_inc);
3244 if (!is_virtual)
3245 wrap = identity_aff(domain);
3247 scop_cond = pet_scop_embed(scop_cond, isl_set_copy(domain),
3248 isl_aff_copy(sched), isl_aff_copy(wrap), isl_id_copy(id));
3249 scop = pet_scop_embed(scop, isl_set_copy(domain), sched, wrap, id);
3250 scop = resolve_nested(scop);
3251 if (has_var_break)
3252 scop = scop_add_break(scop, id_break_test, isl_set_copy(domain),
3253 isl_val_copy(inc));
3254 if (id_test) {
3255 scop = scop_add_while(scop_cond, scop, id_test, domain,
3256 isl_val_copy(inc));
3257 isl_set_free(valid_inc);
3258 } else {
3259 scop = pet_scop_restrict_context(scop, valid_inc);
3260 scop = pet_scop_restrict_context(scop, valid_cond_next);
3261 scop = pet_scop_restrict_context(scop, valid_cond_init);
3262 isl_set_free(domain);
3264 clear_assignment(assigned_value, iv);
3266 isl_val_free(inc);
3268 scop = pet_scop_restrict_context(scop, valid_init);
3270 return scop;
3273 /* Try and construct a pet_scop corresponding to a compound statement.
3275 * "skip_declarations" is set if we should skip initial declarations
3276 * in the children of the compound statements. This then implies
3277 * that this sequence of children should not be treated as a block
3278 * since the initial statements may be skipped.
3280 struct pet_scop *PetScan::extract(CompoundStmt *stmt, bool skip_declarations)
3282 return extract(stmt->children(), !skip_declarations, skip_declarations);
3285 /* For each nested access parameter in "space",
3286 * construct a corresponding pet_expr, place it in args and
3287 * record its position in "param2pos".
3288 * "n_arg" is the number of elements that are already in args.
3289 * The position recorded in "param2pos" takes this number into account.
3290 * If the pet_expr corresponding to a parameter is identical to
3291 * the pet_expr corresponding to an earlier parameter, then these two
3292 * parameters are made to refer to the same element in args.
3294 * Return the final number of elements in args or -1 if an error has occurred.
3296 int PetScan::extract_nested(__isl_keep isl_space *space,
3297 int n_arg, struct pet_expr **args, std::map<int,int> &param2pos)
3299 int nparam;
3301 nparam = isl_space_dim(space, isl_dim_param);
3302 for (int i = 0; i < nparam; ++i) {
3303 int j;
3304 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
3305 Expr *nested;
3307 if (!pet_nested_in_id(id)) {
3308 isl_id_free(id);
3309 continue;
3312 nested = (Expr *) isl_id_get_user(id);
3313 args[n_arg] = extract_expr(nested);
3314 isl_id_free(id);
3315 if (!args[n_arg])
3316 return -1;
3318 for (j = 0; j < n_arg; ++j)
3319 if (pet_expr_is_equal(args[j], args[n_arg]))
3320 break;
3322 if (j < n_arg) {
3323 pet_expr_free(args[n_arg]);
3324 args[n_arg] = NULL;
3325 param2pos[i] = j;
3326 } else
3327 param2pos[i] = n_arg++;
3330 return n_arg;
3333 /* For each nested access parameter in the access relations in "expr",
3334 * construct a corresponding pet_expr, place it in expr->args and
3335 * record its position in "param2pos".
3336 * n is the number of nested access parameters.
3338 struct pet_expr *PetScan::extract_nested(struct pet_expr *expr, int n,
3339 std::map<int,int> &param2pos)
3341 isl_space *space;
3343 expr->args = isl_calloc_array(ctx, struct pet_expr *, n);
3344 expr->n_arg = n;
3345 if (!expr->args)
3346 goto error;
3348 space = pet_expr_access_get_parameter_space(expr);
3349 n = extract_nested(space, 0, expr->args, param2pos);
3350 isl_space_free(space);
3352 if (n < 0)
3353 goto error;
3355 expr->n_arg = n;
3356 return expr;
3357 error:
3358 pet_expr_free(expr);
3359 return NULL;
3362 /* Look for parameters in any access relation in "expr" that
3363 * refer to nested accesses. In particular, these are
3364 * parameters with no name.
3366 * If there are any such parameters, then the domain of the index
3367 * expression and the access relation, which is still [] at this point,
3368 * is replaced by [[] -> [t_1,...,t_n]], with n the number of these parameters
3369 * (after identifying identical nested accesses).
3371 * This transformation is performed in several steps.
3372 * We first extract the arguments in extract_nested.
3373 * param2pos maps the original parameter position to the position
3374 * of the argument.
3375 * Then we move these parameters to input dimensions.
3376 * t2pos maps the positions of these temporary input dimensions
3377 * to the positions of the corresponding arguments.
3378 * Finally, we express these temporary dimensions in terms of the domain
3379 * [[] -> [t_1,...,t_n]] and precompose index expression and access
3380 * relations with this function.
3382 struct pet_expr *PetScan::resolve_nested(struct pet_expr *expr)
3384 int n;
3385 int nparam;
3386 isl_space *space;
3387 isl_local_space *ls;
3388 isl_aff *aff;
3389 isl_multi_aff *ma;
3390 std::map<int,int> param2pos;
3391 std::map<int,int> t2pos;
3393 if (!expr)
3394 return expr;
3396 for (int i = 0; i < expr->n_arg; ++i) {
3397 expr->args[i] = resolve_nested(expr->args[i]);
3398 if (!expr->args[i]) {
3399 pet_expr_free(expr);
3400 return NULL;
3404 if (expr->type != pet_expr_access)
3405 return expr;
3407 space = pet_expr_access_get_parameter_space(expr);
3408 n = pet_nested_n_in_space(space);
3409 isl_space_free(space);
3410 if (n == 0)
3411 return expr;
3413 expr = extract_nested(expr, n, param2pos);
3414 if (!expr)
3415 return NULL;
3417 expr = pet_expr_access_align_params(expr);
3418 if (!expr)
3419 return NULL;
3421 n = 0;
3422 space = pet_expr_access_get_parameter_space(expr);
3423 nparam = isl_space_dim(space, isl_dim_param);
3424 for (int i = nparam - 1; i >= 0; --i) {
3425 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
3426 if (!pet_nested_in_id(id)) {
3427 isl_id_free(id);
3428 continue;
3431 expr = pet_expr_access_move_dims(expr,
3432 isl_dim_in, n, isl_dim_param, i, 1);
3433 t2pos[n] = param2pos[i];
3434 n++;
3436 isl_id_free(id);
3438 isl_space_free(space);
3440 space = pet_expr_access_get_parameter_space(expr);
3441 space = isl_space_set_from_params(space);
3442 space = isl_space_add_dims(space, isl_dim_set, expr->n_arg);
3443 space = isl_space_wrap(isl_space_from_range(space));
3444 ls = isl_local_space_from_space(isl_space_copy(space));
3445 space = isl_space_from_domain(space);
3446 space = isl_space_add_dims(space, isl_dim_out, n);
3447 ma = isl_multi_aff_zero(space);
3449 for (int i = 0; i < n; ++i) {
3450 aff = isl_aff_var_on_domain(isl_local_space_copy(ls),
3451 isl_dim_set, t2pos[i]);
3452 ma = isl_multi_aff_set_aff(ma, i, aff);
3454 isl_local_space_free(ls);
3456 expr = pet_expr_access_pullback_multi_aff(expr, ma);
3458 return expr;
3461 /* Return the file offset of the expansion location of "Loc".
3463 static unsigned getExpansionOffset(SourceManager &SM, SourceLocation Loc)
3465 return SM.getFileOffset(SM.getExpansionLoc(Loc));
3468 #ifdef HAVE_FINDLOCATIONAFTERTOKEN
3470 /* Return a SourceLocation for the location after the first semicolon
3471 * after "loc". If Lexer::findLocationAfterToken is available, we simply
3472 * call it and also skip trailing spaces and newline.
3474 static SourceLocation location_after_semi(SourceLocation loc, SourceManager &SM,
3475 const LangOptions &LO)
3477 return Lexer::findLocationAfterToken(loc, tok::semi, SM, LO, true);
3480 #else
3482 /* Return a SourceLocation for the location after the first semicolon
3483 * after "loc". If Lexer::findLocationAfterToken is not available,
3484 * we look in the underlying character data for the first semicolon.
3486 static SourceLocation location_after_semi(SourceLocation loc, SourceManager &SM,
3487 const LangOptions &LO)
3489 const char *semi;
3490 const char *s = SM.getCharacterData(loc);
3492 semi = strchr(s, ';');
3493 if (!semi)
3494 return SourceLocation();
3495 return loc.getFileLocWithOffset(semi + 1 - s);
3498 #endif
3500 /* If the token at "loc" is the first token on the line, then return
3501 * a location referring to the start of the line.
3502 * Otherwise, return "loc".
3504 * This function is used to extend a scop to the start of the line
3505 * if the first token of the scop is also the first token on the line.
3507 * We look for the first token on the line. If its location is equal to "loc",
3508 * then the latter is the location of the first token on the line.
3510 static SourceLocation move_to_start_of_line_if_first_token(SourceLocation loc,
3511 SourceManager &SM, const LangOptions &LO)
3513 std::pair<FileID, unsigned> file_offset_pair;
3514 llvm::StringRef file;
3515 const char *pos;
3516 Token tok;
3517 SourceLocation token_loc, line_loc;
3518 int col;
3520 loc = SM.getExpansionLoc(loc);
3521 col = SM.getExpansionColumnNumber(loc);
3522 line_loc = loc.getLocWithOffset(1 - col);
3523 file_offset_pair = SM.getDecomposedLoc(line_loc);
3524 file = SM.getBufferData(file_offset_pair.first, NULL);
3525 pos = file.data() + file_offset_pair.second;
3527 Lexer lexer(SM.getLocForStartOfFile(file_offset_pair.first), LO,
3528 file.begin(), pos, file.end());
3529 lexer.LexFromRawLexer(tok);
3530 token_loc = tok.getLocation();
3532 if (token_loc == loc)
3533 return line_loc;
3534 else
3535 return loc;
3538 /* Update start and end of "scop" to include the region covered by "range".
3539 * If "skip_semi" is set, then we assume "range" is followed by
3540 * a semicolon and also include this semicolon.
3542 struct pet_scop *PetScan::update_scop_start_end(struct pet_scop *scop,
3543 SourceRange range, bool skip_semi)
3545 SourceLocation loc = range.getBegin();
3546 SourceManager &SM = PP.getSourceManager();
3547 const LangOptions &LO = PP.getLangOpts();
3548 unsigned start, end;
3550 loc = move_to_start_of_line_if_first_token(loc, SM, LO);
3551 start = getExpansionOffset(SM, loc);
3552 loc = range.getEnd();
3553 if (skip_semi)
3554 loc = location_after_semi(loc, SM, LO);
3555 else
3556 loc = PP.getLocForEndOfToken(loc);
3557 end = getExpansionOffset(SM, loc);
3559 scop = pet_scop_update_start_end(scop, start, end);
3560 return scop;
3563 /* Convert a top-level pet_expr to a pet_scop with one statement.
3564 * This mainly involves resolving nested expression parameters
3565 * and setting the name of the iteration space.
3566 * The name is given by "label" if it is non-NULL. Otherwise,
3567 * it is of the form S_<n_stmt>.
3568 * start and end of the pet_scop are derived from those of "stmt".
3569 * If "stmt" is an expression statement, then its range does not
3570 * include the semicolon, while it should be included in the pet_scop.
3572 struct pet_scop *PetScan::extract(Stmt *stmt, struct pet_expr *expr,
3573 __isl_take isl_id *label)
3575 struct pet_stmt *ps;
3576 struct pet_scop *scop;
3577 SourceLocation loc = stmt->getLocStart();
3578 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3579 bool skip_semi;
3581 expr = resolve_nested(expr);
3582 ps = pet_stmt_from_pet_expr(ctx, line, label, n_stmt++, expr);
3583 scop = pet_scop_from_pet_stmt(ctx, ps);
3585 skip_semi = isa<Expr>(stmt);
3586 scop = update_scop_start_end(scop, stmt->getSourceRange(), skip_semi);
3587 return scop;
3590 /* Check if we can extract an affine expression from "expr".
3591 * Return the expressions as an isl_pw_aff if we can and NULL otherwise.
3592 * We turn on autodetection so that we won't generate any warnings
3593 * and turn off nesting, so that we won't accept any non-affine constructs.
3595 __isl_give isl_pw_aff *PetScan::try_extract_affine(Expr *expr)
3597 isl_pw_aff *pwaff;
3598 int save_autodetect = options->autodetect;
3599 bool save_nesting = nesting_enabled;
3601 options->autodetect = 1;
3602 nesting_enabled = false;
3604 pwaff = extract_affine(expr);
3606 options->autodetect = save_autodetect;
3607 nesting_enabled = save_nesting;
3609 return pwaff;
3612 /* Check if we can extract an affine constraint from "expr".
3613 * Return the constraint as an isl_set if we can and NULL otherwise.
3614 * We turn on autodetection so that we won't generate any warnings
3615 * and turn off nesting, so that we won't accept any non-affine constructs.
3617 __isl_give isl_pw_aff *PetScan::try_extract_affine_condition(Expr *expr)
3619 isl_pw_aff *cond;
3620 int save_autodetect = options->autodetect;
3621 bool save_nesting = nesting_enabled;
3623 options->autodetect = 1;
3624 nesting_enabled = false;
3626 cond = extract_condition(expr);
3628 options->autodetect = save_autodetect;
3629 nesting_enabled = save_nesting;
3631 return cond;
3634 /* Check whether "expr" is an affine constraint.
3636 bool PetScan::is_affine_condition(Expr *expr)
3638 isl_pw_aff *cond;
3640 cond = try_extract_affine_condition(expr);
3641 isl_pw_aff_free(cond);
3643 return cond != NULL;
3646 /* Check if we can extract a condition from "expr".
3647 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
3648 * If allow_nested is set, then the condition may involve parameters
3649 * corresponding to nested accesses.
3650 * We turn on autodetection so that we won't generate any warnings.
3652 __isl_give isl_pw_aff *PetScan::try_extract_nested_condition(Expr *expr)
3654 isl_pw_aff *cond;
3655 int save_autodetect = options->autodetect;
3656 bool save_nesting = nesting_enabled;
3658 options->autodetect = 1;
3659 nesting_enabled = allow_nested;
3660 cond = extract_condition(expr);
3662 options->autodetect = save_autodetect;
3663 nesting_enabled = save_nesting;
3665 return cond;
3668 /* If the top-level expression of "stmt" is an assignment, then
3669 * return that assignment as a BinaryOperator.
3670 * Otherwise return NULL.
3672 static BinaryOperator *top_assignment_or_null(Stmt *stmt)
3674 BinaryOperator *ass;
3676 if (!stmt)
3677 return NULL;
3678 if (stmt->getStmtClass() != Stmt::BinaryOperatorClass)
3679 return NULL;
3681 ass = cast<BinaryOperator>(stmt);
3682 if(ass->getOpcode() != BO_Assign)
3683 return NULL;
3685 return ass;
3688 /* Check if the given if statement is a conditional assignement
3689 * with a non-affine condition. If so, construct a pet_scop
3690 * corresponding to this conditional assignment. Otherwise return NULL.
3692 * In particular we check if "stmt" is of the form
3694 * if (condition)
3695 * a = f(...);
3696 * else
3697 * a = g(...);
3699 * where a is some array or scalar access.
3700 * The constructed pet_scop then corresponds to the expression
3702 * a = condition ? f(...) : g(...)
3704 * All access relations in f(...) are intersected with condition
3705 * while all access relation in g(...) are intersected with the complement.
3707 struct pet_scop *PetScan::extract_conditional_assignment(IfStmt *stmt)
3709 BinaryOperator *ass_then, *ass_else;
3710 isl_multi_pw_aff *write_then, *write_else;
3711 isl_set *cond, *comp;
3712 isl_multi_pw_aff *index;
3713 isl_pw_aff *pa;
3714 int equal;
3715 struct pet_expr *pe_cond, *pe_then, *pe_else, *pe, *pe_write;
3716 bool save_nesting = nesting_enabled;
3718 if (!options->detect_conditional_assignment)
3719 return NULL;
3721 ass_then = top_assignment_or_null(stmt->getThen());
3722 ass_else = top_assignment_or_null(stmt->getElse());
3724 if (!ass_then || !ass_else)
3725 return NULL;
3727 if (is_affine_condition(stmt->getCond()))
3728 return NULL;
3730 write_then = extract_index(ass_then->getLHS());
3731 write_else = extract_index(ass_else->getLHS());
3733 equal = isl_multi_pw_aff_plain_is_equal(write_then, write_else);
3734 isl_multi_pw_aff_free(write_else);
3735 if (equal < 0 || !equal) {
3736 isl_multi_pw_aff_free(write_then);
3737 return NULL;
3740 nesting_enabled = allow_nested;
3741 pa = extract_condition(stmt->getCond());
3742 nesting_enabled = save_nesting;
3743 cond = isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa));
3744 comp = isl_pw_aff_zero_set(isl_pw_aff_copy(pa));
3745 index = isl_multi_pw_aff_from_range(isl_multi_pw_aff_from_pw_aff(pa));
3747 pe_cond = pet_expr_from_index(index);
3749 pe_then = extract_expr(ass_then->getRHS());
3750 pe_then = pet_expr_restrict(pe_then, cond);
3751 pe_else = extract_expr(ass_else->getRHS());
3752 pe_else = pet_expr_restrict(pe_else, comp);
3754 pe = pet_expr_new_ternary(ctx, pe_cond, pe_then, pe_else);
3755 pe_write = pet_expr_from_index_and_depth(write_then,
3756 extract_depth(write_then));
3757 if (pe_write) {
3758 pe_write->acc.write = 1;
3759 pe_write->acc.read = 0;
3761 pe = pet_expr_new_binary(ctx, pet_op_assign, pe_write, pe);
3762 return extract(stmt, pe);
3765 /* Create a pet_scop with a single statement with name S_<stmt_nr>,
3766 * evaluating "cond" and writing the result to a virtual scalar,
3767 * as expressed by "index".
3769 struct pet_scop *PetScan::extract_non_affine_condition(Expr *cond, int stmt_nr,
3770 __isl_take isl_multi_pw_aff *index)
3772 struct pet_expr *expr, *write;
3773 struct pet_stmt *ps;
3774 SourceLocation loc = cond->getLocStart();
3775 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3777 write = pet_expr_from_index(index);
3778 if (write) {
3779 write->acc.write = 1;
3780 write->acc.read = 0;
3782 expr = extract_expr(cond);
3783 expr = resolve_nested(expr);
3784 expr = pet_expr_new_binary(ctx, pet_op_assign, write, expr);
3785 ps = pet_stmt_from_pet_expr(ctx, line, NULL, stmt_nr, expr);
3786 return pet_scop_from_pet_stmt(ctx, ps);
3789 extern "C" {
3790 static struct pet_expr *embed_access(struct pet_expr *expr, void *user);
3793 /* Precompose the access relation and the index expression associated
3794 * to "expr" with the function pointed to by "user",
3795 * thereby embedding the access relation in the domain of this function.
3796 * The initial domain of the access relation and the index expression
3797 * is the zero-dimensional domain.
3799 static struct pet_expr *embed_access(struct pet_expr *expr, void *user)
3801 isl_multi_aff *ma = (isl_multi_aff *) user;
3803 return pet_expr_access_pullback_multi_aff(expr, isl_multi_aff_copy(ma));
3806 /* Precompose all access relations in "expr" with "ma", thereby
3807 * embedding them in the domain of "ma".
3809 static struct pet_expr *embed(struct pet_expr *expr,
3810 __isl_keep isl_multi_aff *ma)
3812 return pet_expr_map_access(expr, &embed_access, ma);
3815 /* For each nested access parameter in the domain of "stmt",
3816 * construct a corresponding pet_expr, place it before the original
3817 * elements in stmt->args and record its position in "param2pos".
3818 * n is the number of nested access parameters.
3820 struct pet_stmt *PetScan::extract_nested(struct pet_stmt *stmt, int n,
3821 std::map<int,int> &param2pos)
3823 int i;
3824 isl_space *space;
3825 int n_arg;
3826 struct pet_expr **args;
3828 n_arg = stmt->n_arg;
3829 args = isl_calloc_array(ctx, struct pet_expr *, n + n_arg);
3830 if (!args)
3831 goto error;
3833 space = isl_set_get_space(stmt->domain);
3834 n_arg = extract_nested(space, 0, args, param2pos);
3835 isl_space_free(space);
3837 if (n_arg < 0)
3838 goto error;
3840 for (i = 0; i < stmt->n_arg; ++i)
3841 args[n_arg + i] = stmt->args[i];
3842 free(stmt->args);
3843 stmt->args = args;
3844 stmt->n_arg += n_arg;
3846 return stmt;
3847 error:
3848 if (args) {
3849 for (i = 0; i < n; ++i)
3850 pet_expr_free(args[i]);
3851 free(args);
3853 pet_stmt_free(stmt);
3854 return NULL;
3857 /* Check whether any of the arguments i of "stmt" starting at position "n"
3858 * is equal to one of the first "n" arguments j.
3859 * If so, combine the constraints on arguments i and j and remove
3860 * argument i.
3862 static struct pet_stmt *remove_duplicate_arguments(struct pet_stmt *stmt, int n)
3864 int i, j;
3865 isl_map *map;
3867 if (!stmt)
3868 return NULL;
3869 if (n == 0)
3870 return stmt;
3871 if (n == stmt->n_arg)
3872 return stmt;
3874 map = isl_set_unwrap(stmt->domain);
3876 for (i = stmt->n_arg - 1; i >= n; --i) {
3877 for (j = 0; j < n; ++j)
3878 if (pet_expr_is_equal(stmt->args[i], stmt->args[j]))
3879 break;
3880 if (j >= n)
3881 continue;
3883 map = isl_map_equate(map, isl_dim_out, i, isl_dim_out, j);
3884 map = isl_map_project_out(map, isl_dim_out, i, 1);
3886 pet_expr_free(stmt->args[i]);
3887 for (j = i; j + 1 < stmt->n_arg; ++j)
3888 stmt->args[j] = stmt->args[j + 1];
3889 stmt->n_arg--;
3892 stmt->domain = isl_map_wrap(map);
3893 if (!stmt->domain)
3894 goto error;
3895 return stmt;
3896 error:
3897 pet_stmt_free(stmt);
3898 return NULL;
3901 /* Look for parameters in the iteration domain of "stmt" that
3902 * refer to nested accesses. In particular, these are
3903 * parameters with no name.
3905 * If there are any such parameters, then as many extra variables
3906 * (after identifying identical nested accesses) are inserted in the
3907 * range of the map wrapped inside the domain, before the original variables.
3908 * If the original domain is not a wrapped map, then a new wrapped
3909 * map is created with zero output dimensions.
3910 * The parameters are then equated to the corresponding output dimensions
3911 * and subsequently projected out, from the iteration domain,
3912 * the schedule and the access relations.
3913 * For each of the output dimensions, a corresponding argument
3914 * expression is inserted. Initially they are created with
3915 * a zero-dimensional domain, so they have to be embedded
3916 * in the current iteration domain.
3917 * param2pos maps the position of the parameter to the position
3918 * of the corresponding output dimension in the wrapped map.
3920 struct pet_stmt *PetScan::resolve_nested(struct pet_stmt *stmt)
3922 int n;
3923 int nparam;
3924 unsigned n_arg;
3925 isl_map *map;
3926 isl_space *space;
3927 isl_multi_aff *ma;
3928 std::map<int,int> param2pos;
3930 if (!stmt)
3931 return NULL;
3933 n = pet_nested_n_in_set(stmt->domain);
3934 if (n == 0)
3935 return stmt;
3937 n_arg = stmt->n_arg;
3938 stmt = extract_nested(stmt, n, param2pos);
3939 if (!stmt)
3940 return NULL;
3942 n = stmt->n_arg - n_arg;
3943 nparam = isl_set_dim(stmt->domain, isl_dim_param);
3944 if (isl_set_is_wrapping(stmt->domain))
3945 map = isl_set_unwrap(stmt->domain);
3946 else
3947 map = isl_map_from_domain(stmt->domain);
3948 map = isl_map_insert_dims(map, isl_dim_out, 0, n);
3950 for (int i = nparam - 1; i >= 0; --i) {
3951 isl_id *id;
3953 if (!pet_nested_in_map(map, i))
3954 continue;
3956 id = pet_expr_access_get_id(stmt->args[param2pos[i]]);
3957 map = isl_map_set_dim_id(map, isl_dim_out, param2pos[i], id);
3958 map = isl_map_equate(map, isl_dim_param, i, isl_dim_out,
3959 param2pos[i]);
3960 map = isl_map_project_out(map, isl_dim_param, i, 1);
3963 stmt->domain = isl_map_wrap(map);
3965 space = isl_space_unwrap(isl_set_get_space(stmt->domain));
3966 space = isl_space_from_domain(isl_space_domain(space));
3967 ma = isl_multi_aff_zero(space);
3968 for (int pos = 0; pos < n; ++pos)
3969 stmt->args[pos] = embed(stmt->args[pos], ma);
3970 isl_multi_aff_free(ma);
3972 stmt = pet_stmt_remove_nested_parameters(stmt);
3973 stmt = remove_duplicate_arguments(stmt, n);
3975 return stmt;
3978 /* For each statement in "scop", move the parameters that correspond
3979 * to nested access into the ranges of the domains and create
3980 * corresponding argument expressions.
3982 struct pet_scop *PetScan::resolve_nested(struct pet_scop *scop)
3984 if (!scop)
3985 return NULL;
3987 for (int i = 0; i < scop->n_stmt; ++i) {
3988 scop->stmts[i] = resolve_nested(scop->stmts[i]);
3989 if (!scop->stmts[i])
3990 goto error;
3993 return scop;
3994 error:
3995 pet_scop_free(scop);
3996 return NULL;
3999 /* Given an access expression "expr", is the variable accessed by
4000 * "expr" assigned anywhere inside "scop"?
4002 static bool is_assigned(pet_expr *expr, pet_scop *scop)
4004 bool assigned = false;
4005 isl_id *id;
4007 id = pet_expr_access_get_id(expr);
4008 assigned = pet_scop_writes(scop, id);
4009 isl_id_free(id);
4011 return assigned;
4014 /* Are all nested access parameters in "pa" allowed given "scop".
4015 * In particular, is none of them written by anywhere inside "scop".
4017 * If "scop" has any skip conditions, then no nested access parameters
4018 * are allowed. In particular, if there is any nested access in a guard
4019 * for a piece of code containing a "continue", then we want to introduce
4020 * a separate statement for evaluating this guard so that we can express
4021 * that the result is false for all previous iterations.
4023 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff *pa, pet_scop *scop)
4025 int nparam;
4027 if (!scop)
4028 return true;
4030 if (!pet_nested_any_in_pw_aff(pa))
4031 return true;
4033 if (pet_scop_has_skip(scop, pet_skip_now))
4034 return false;
4036 nparam = isl_pw_aff_dim(pa, isl_dim_param);
4037 for (int i = 0; i < nparam; ++i) {
4038 Expr *nested;
4039 isl_id *id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
4040 pet_expr *expr;
4041 bool allowed;
4043 if (!pet_nested_in_id(id)) {
4044 isl_id_free(id);
4045 continue;
4048 nested = (Expr *) isl_id_get_user(id);
4049 expr = extract_expr(nested);
4050 allowed = expr && expr->type == pet_expr_access &&
4051 !is_assigned(expr, scop);
4053 pet_expr_free(expr);
4054 isl_id_free(id);
4056 if (!allowed)
4057 return false;
4060 return true;
4063 /* Do we need to construct a skip condition of the given type
4064 * on an if statement, given that the if condition is non-affine?
4066 * pet_scop_filter_skip can only handle the case where the if condition
4067 * holds (the then branch) and the skip condition is universal.
4068 * In any other case, we need to construct a new skip condition.
4070 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
4071 bool have_else, enum pet_skip type)
4073 if (have_else && scop_else && pet_scop_has_skip(scop_else, type))
4074 return true;
4075 if (scop_then && pet_scop_has_skip(scop_then, type) &&
4076 !pet_scop_has_universal_skip(scop_then, type))
4077 return true;
4078 return false;
4081 /* Do we need to construct a skip condition of the given type
4082 * on an if statement, given that the if condition is affine?
4084 * There is no need to construct a new skip condition if all
4085 * the skip conditions are affine.
4087 static bool need_skip_aff(struct pet_scop *scop_then,
4088 struct pet_scop *scop_else, bool have_else, enum pet_skip type)
4090 if (scop_then && pet_scop_has_var_skip(scop_then, type))
4091 return true;
4092 if (have_else && scop_else && pet_scop_has_var_skip(scop_else, type))
4093 return true;
4094 return false;
4097 /* Do we need to construct a skip condition of the given type
4098 * on an if statement?
4100 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
4101 bool have_else, enum pet_skip type, bool affine)
4103 if (affine)
4104 return need_skip_aff(scop_then, scop_else, have_else, type);
4105 else
4106 return need_skip(scop_then, scop_else, have_else, type);
4109 /* Construct an affine expression pet_expr that evaluates
4110 * to the constant "val".
4112 static struct pet_expr *universally(isl_ctx *ctx, int val)
4114 isl_local_space *ls;
4115 isl_aff *aff;
4116 isl_multi_pw_aff *mpa;
4118 ls = isl_local_space_from_space(isl_space_set_alloc(ctx, 0, 0));
4119 aff = isl_aff_val_on_domain(ls, isl_val_int_from_si(ctx, val));
4120 mpa = isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
4122 return pet_expr_from_index(mpa);
4125 /* Construct an affine expression pet_expr that evaluates
4126 * to the constant 1.
4128 static struct pet_expr *universally_true(isl_ctx *ctx)
4130 return universally(ctx, 1);
4133 /* Construct an affine expression pet_expr that evaluates
4134 * to the constant 0.
4136 static struct pet_expr *universally_false(isl_ctx *ctx)
4138 return universally(ctx, 0);
4141 /* Given an index expression "test_index" for the if condition,
4142 * an index expression "skip_index" for the skip condition and
4143 * scops for the then and else branches, construct a scop for
4144 * computing "skip_index".
4146 * The computed scop contains a single statement that essentially does
4148 * skip_index = test_cond ? skip_cond_then : skip_cond_else
4150 * If the skip conditions of the then and/or else branch are not affine,
4151 * then they need to be filtered by test_index.
4152 * If they are missing, then this means the skip condition is false.
4154 * Since we are constructing a skip condition for the if statement,
4155 * the skip conditions on the then and else branches are removed.
4157 static struct pet_scop *extract_skip(PetScan *scan,
4158 __isl_take isl_multi_pw_aff *test_index,
4159 __isl_take isl_multi_pw_aff *skip_index,
4160 struct pet_scop *scop_then, struct pet_scop *scop_else, bool have_else,
4161 enum pet_skip type)
4163 struct pet_expr *expr_then, *expr_else, *expr, *expr_skip;
4164 struct pet_stmt *stmt;
4165 struct pet_scop *scop;
4166 isl_ctx *ctx = scan->ctx;
4168 if (!scop_then)
4169 goto error;
4170 if (have_else && !scop_else)
4171 goto error;
4173 if (pet_scop_has_skip(scop_then, type)) {
4174 expr_then = pet_scop_get_skip_expr(scop_then, type);
4175 pet_scop_reset_skip(scop_then, type);
4176 if (!pet_expr_is_affine(expr_then))
4177 expr_then = pet_expr_filter(expr_then,
4178 isl_multi_pw_aff_copy(test_index), 1);
4179 } else
4180 expr_then = universally_false(ctx);
4182 if (have_else && pet_scop_has_skip(scop_else, type)) {
4183 expr_else = pet_scop_get_skip_expr(scop_else, type);
4184 pet_scop_reset_skip(scop_else, type);
4185 if (!pet_expr_is_affine(expr_else))
4186 expr_else = pet_expr_filter(expr_else,
4187 isl_multi_pw_aff_copy(test_index), 0);
4188 } else
4189 expr_else = universally_false(ctx);
4191 expr = pet_expr_from_index(test_index);
4192 expr = pet_expr_new_ternary(ctx, expr, expr_then, expr_else);
4193 expr_skip = pet_expr_from_index(isl_multi_pw_aff_copy(skip_index));
4194 if (expr_skip) {
4195 expr_skip->acc.write = 1;
4196 expr_skip->acc.read = 0;
4198 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
4199 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, scan->n_stmt++, expr);
4201 scop = pet_scop_from_pet_stmt(ctx, stmt);
4202 scop = scop_add_array(scop, skip_index, scan->ast_context);
4203 isl_multi_pw_aff_free(skip_index);
4205 return scop;
4206 error:
4207 isl_multi_pw_aff_free(test_index);
4208 isl_multi_pw_aff_free(skip_index);
4209 return NULL;
4212 /* Is scop's skip_now condition equal to its skip_later condition?
4213 * In particular, this means that it either has no skip_now condition
4214 * or both a skip_now and a skip_later condition (that are equal to each other).
4216 static bool skip_equals_skip_later(struct pet_scop *scop)
4218 int has_skip_now, has_skip_later;
4219 int equal;
4220 isl_multi_pw_aff *skip_now, *skip_later;
4222 if (!scop)
4223 return false;
4224 has_skip_now = pet_scop_has_skip(scop, pet_skip_now);
4225 has_skip_later = pet_scop_has_skip(scop, pet_skip_later);
4226 if (has_skip_now != has_skip_later)
4227 return false;
4228 if (!has_skip_now)
4229 return true;
4231 skip_now = pet_scop_get_skip(scop, pet_skip_now);
4232 skip_later = pet_scop_get_skip(scop, pet_skip_later);
4233 equal = isl_multi_pw_aff_is_equal(skip_now, skip_later);
4234 isl_multi_pw_aff_free(skip_now);
4235 isl_multi_pw_aff_free(skip_later);
4237 return equal;
4240 /* Drop the skip conditions of type pet_skip_later from scop1 and scop2.
4242 static void drop_skip_later(struct pet_scop *scop1, struct pet_scop *scop2)
4244 pet_scop_reset_skip(scop1, pet_skip_later);
4245 pet_scop_reset_skip(scop2, pet_skip_later);
4248 /* Structure that handles the construction of skip conditions.
4250 * scop_then and scop_else represent the then and else branches
4251 * of the if statement
4253 * skip[type] is true if we need to construct a skip condition of that type
4254 * equal is set if the skip conditions of types pet_skip_now and pet_skip_later
4255 * are equal to each other
4256 * index[type] is an index expression from a zero-dimension domain
4257 * to the virtual array representing the skip condition
4258 * scop[type] is a scop for computing the skip condition
4260 struct pet_skip_info {
4261 isl_ctx *ctx;
4263 bool skip[2];
4264 bool equal;
4265 isl_multi_pw_aff *index[2];
4266 struct pet_scop *scop[2];
4268 pet_skip_info(isl_ctx *ctx) : ctx(ctx) {}
4270 operator bool() { return skip[pet_skip_now] || skip[pet_skip_later]; }
4273 /* Structure that handles the construction of skip conditions on if statements.
4275 * scop_then and scop_else represent the then and else branches
4276 * of the if statement
4278 struct pet_skip_info_if : public pet_skip_info {
4279 struct pet_scop *scop_then, *scop_else;
4280 bool have_else;
4282 pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
4283 struct pet_scop *scop_else, bool have_else, bool affine);
4284 void extract(PetScan *scan, __isl_keep isl_multi_pw_aff *index,
4285 enum pet_skip type);
4286 void extract(PetScan *scan, __isl_keep isl_multi_pw_aff *index);
4287 void extract(PetScan *scan, __isl_keep isl_pw_aff *cond);
4288 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
4289 int offset);
4290 struct pet_scop *add(struct pet_scop *scop, int offset);
4293 /* Initialize a pet_skip_info_if structure based on the then and else branches
4294 * and based on whether the if condition is affine or not.
4296 pet_skip_info_if::pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
4297 struct pet_scop *scop_else, bool have_else, bool affine) :
4298 pet_skip_info(ctx), scop_then(scop_then), scop_else(scop_else),
4299 have_else(have_else)
4301 skip[pet_skip_now] =
4302 need_skip(scop_then, scop_else, have_else, pet_skip_now, affine);
4303 equal = skip[pet_skip_now] && skip_equals_skip_later(scop_then) &&
4304 (!have_else || skip_equals_skip_later(scop_else));
4305 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
4306 need_skip(scop_then, scop_else, have_else, pet_skip_later, affine);
4309 /* If we need to construct a skip condition of the given type,
4310 * then do so now.
4312 * "mpa" represents the if condition.
4314 void pet_skip_info_if::extract(PetScan *scan,
4315 __isl_keep isl_multi_pw_aff *mpa, enum pet_skip type)
4317 isl_ctx *ctx;
4319 if (!skip[type])
4320 return;
4322 ctx = isl_multi_pw_aff_get_ctx(mpa);
4323 index[type] = create_test_index(ctx, scan->n_test++);
4324 scop[type] = extract_skip(scan, isl_multi_pw_aff_copy(mpa),
4325 isl_multi_pw_aff_copy(index[type]),
4326 scop_then, scop_else, have_else, type);
4329 /* Construct the required skip conditions, given the if condition "index".
4331 void pet_skip_info_if::extract(PetScan *scan,
4332 __isl_keep isl_multi_pw_aff *index)
4334 extract(scan, index, pet_skip_now);
4335 extract(scan, index, pet_skip_later);
4336 if (equal)
4337 drop_skip_later(scop_then, scop_else);
4340 /* Construct the required skip conditions, given the if condition "cond".
4342 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_pw_aff *cond)
4344 isl_multi_pw_aff *test;
4346 if (!skip[pet_skip_now] && !skip[pet_skip_later])
4347 return;
4349 test = isl_multi_pw_aff_from_pw_aff(isl_pw_aff_copy(cond));
4350 test = isl_multi_pw_aff_from_range(test);
4351 extract(scan, test);
4352 isl_multi_pw_aff_free(test);
4355 /* Add the computed skip condition of the give type to "main" and
4356 * add the scop for computing the condition at the given offset.
4358 * If equal is set, then we only computed a skip condition for pet_skip_now,
4359 * but we also need to set it as main's pet_skip_later.
4361 struct pet_scop *pet_skip_info_if::add(struct pet_scop *main,
4362 enum pet_skip type, int offset)
4364 if (!skip[type])
4365 return main;
4367 scop[type] = pet_scop_prefix(scop[type], offset);
4368 main = pet_scop_add_par(ctx, main, scop[type]);
4369 scop[type] = NULL;
4371 if (equal)
4372 main = pet_scop_set_skip(main, pet_skip_later,
4373 isl_multi_pw_aff_copy(index[type]));
4375 main = pet_scop_set_skip(main, type, index[type]);
4376 index[type] = NULL;
4378 return main;
4381 /* Add the computed skip conditions to "main" and
4382 * add the scops for computing the conditions at the given offset.
4384 struct pet_scop *pet_skip_info_if::add(struct pet_scop *scop, int offset)
4386 scop = add(scop, pet_skip_now, offset);
4387 scop = add(scop, pet_skip_later, offset);
4389 return scop;
4392 /* Construct a pet_scop for a non-affine if statement.
4394 * We create a separate statement that writes the result
4395 * of the non-affine condition to a virtual scalar.
4396 * A constraint requiring the value of this virtual scalar to be one
4397 * is added to the iteration domains of the then branch.
4398 * Similarly, a constraint requiring the value of this virtual scalar
4399 * to be zero is added to the iteration domains of the else branch, if any.
4400 * We adjust the schedules to ensure that the virtual scalar is written
4401 * before it is read.
4403 * If there are any breaks or continues in the then and/or else
4404 * branches, then we may have to compute a new skip condition.
4405 * This is handled using a pet_skip_info_if object.
4406 * On initialization, the object checks if skip conditions need
4407 * to be computed. If so, it does so in "extract" and adds them in "add".
4409 struct pet_scop *PetScan::extract_non_affine_if(Expr *cond,
4410 struct pet_scop *scop_then, struct pet_scop *scop_else,
4411 bool have_else, int stmt_id)
4413 struct pet_scop *scop;
4414 isl_multi_pw_aff *test_index;
4415 int save_n_stmt = n_stmt;
4417 test_index = create_test_index(ctx, n_test++);
4418 n_stmt = stmt_id;
4419 scop = extract_non_affine_condition(cond, n_stmt++,
4420 isl_multi_pw_aff_copy(test_index));
4421 n_stmt = save_n_stmt;
4422 scop = scop_add_array(scop, test_index, ast_context);
4424 pet_skip_info_if skip(ctx, scop_then, scop_else, have_else, false);
4425 skip.extract(this, test_index);
4427 scop = pet_scop_prefix(scop, 0);
4428 scop_then = pet_scop_prefix(scop_then, 1);
4429 scop_then = pet_scop_filter(scop_then,
4430 isl_multi_pw_aff_copy(test_index), 1);
4431 if (have_else) {
4432 scop_else = pet_scop_prefix(scop_else, 1);
4433 scop_else = pet_scop_filter(scop_else, test_index, 0);
4434 scop_then = pet_scop_add_par(ctx, scop_then, scop_else);
4435 } else
4436 isl_multi_pw_aff_free(test_index);
4438 scop = pet_scop_add_seq(ctx, scop, scop_then);
4440 scop = skip.add(scop, 2);
4442 return scop;
4445 /* Construct a pet_scop for an if statement.
4447 * If the condition fits the pattern of a conditional assignment,
4448 * then it is handled by extract_conditional_assignment.
4449 * Otherwise, we do the following.
4451 * If the condition is affine, then the condition is added
4452 * to the iteration domains of the then branch, while the
4453 * opposite of the condition in added to the iteration domains
4454 * of the else branch, if any.
4455 * We allow the condition to be dynamic, i.e., to refer to
4456 * scalars or array elements that may be written to outside
4457 * of the given if statement. These nested accesses are then represented
4458 * as output dimensions in the wrapping iteration domain.
4459 * If it is also written _inside_ the then or else branch, then
4460 * we treat the condition as non-affine.
4461 * As explained in extract_non_affine_if, this will introduce
4462 * an extra statement.
4463 * For aesthetic reasons, we want this statement to have a statement
4464 * number that is lower than those of the then and else branches.
4465 * In order to evaluate if we will need such a statement, however, we
4466 * first construct scops for the then and else branches.
4467 * We therefore reserve a statement number if we might have to
4468 * introduce such an extra statement.
4470 * If the condition is not affine, then the scop is created in
4471 * extract_non_affine_if.
4473 * If there are any breaks or continues in the then and/or else
4474 * branches, then we may have to compute a new skip condition.
4475 * This is handled using a pet_skip_info_if object.
4476 * On initialization, the object checks if skip conditions need
4477 * to be computed. If so, it does so in "extract" and adds them in "add".
4479 struct pet_scop *PetScan::extract(IfStmt *stmt)
4481 struct pet_scop *scop_then, *scop_else = NULL, *scop;
4482 isl_pw_aff *cond;
4483 int stmt_id;
4484 isl_set *set;
4485 isl_set *valid;
4487 clear_assignments clear(assigned_value);
4488 clear.TraverseStmt(stmt->getThen());
4489 if (stmt->getElse())
4490 clear.TraverseStmt(stmt->getElse());
4492 scop = extract_conditional_assignment(stmt);
4493 if (scop)
4494 return scop;
4496 cond = try_extract_nested_condition(stmt->getCond());
4497 if (allow_nested && (!cond || pet_nested_any_in_pw_aff(cond)))
4498 stmt_id = n_stmt++;
4501 assigned_value_cache cache(assigned_value);
4502 scop_then = extract(stmt->getThen());
4505 if (stmt->getElse()) {
4506 assigned_value_cache cache(assigned_value);
4507 scop_else = extract(stmt->getElse());
4508 if (options->autodetect) {
4509 if (scop_then && !scop_else) {
4510 partial = true;
4511 isl_pw_aff_free(cond);
4512 return scop_then;
4514 if (!scop_then && scop_else) {
4515 partial = true;
4516 isl_pw_aff_free(cond);
4517 return scop_else;
4522 if (cond &&
4523 (!is_nested_allowed(cond, scop_then) ||
4524 (stmt->getElse() && !is_nested_allowed(cond, scop_else)))) {
4525 isl_pw_aff_free(cond);
4526 cond = NULL;
4528 if (allow_nested && !cond)
4529 return extract_non_affine_if(stmt->getCond(), scop_then,
4530 scop_else, stmt->getElse(), stmt_id);
4532 if (!cond)
4533 cond = extract_condition(stmt->getCond());
4535 pet_skip_info_if skip(ctx, scop_then, scop_else, stmt->getElse(), true);
4536 skip.extract(this, cond);
4538 valid = isl_pw_aff_domain(isl_pw_aff_copy(cond));
4539 set = isl_pw_aff_non_zero_set(cond);
4540 scop = pet_scop_restrict(scop_then, isl_set_copy(set));
4542 if (stmt->getElse()) {
4543 set = isl_set_subtract(isl_set_copy(valid), set);
4544 scop_else = pet_scop_restrict(scop_else, set);
4545 scop = pet_scop_add_par(ctx, scop, scop_else);
4546 } else
4547 isl_set_free(set);
4548 scop = resolve_nested(scop);
4549 scop = pet_scop_restrict_context(scop, valid);
4551 if (skip)
4552 scop = pet_scop_prefix(scop, 0);
4553 scop = skip.add(scop, 1);
4555 return scop;
4558 /* Try and construct a pet_scop for a label statement.
4559 * We currently only allow labels on expression statements.
4561 struct pet_scop *PetScan::extract(LabelStmt *stmt)
4563 isl_id *label;
4564 Stmt *sub;
4566 sub = stmt->getSubStmt();
4567 if (!isa<Expr>(sub)) {
4568 unsupported(stmt);
4569 return NULL;
4572 label = isl_id_alloc(ctx, stmt->getName(), NULL);
4574 return extract(sub, extract_expr(cast<Expr>(sub)), label);
4577 /* Return a one-dimensional multi piecewise affine expression that is equal
4578 * to the constant 1 and is defined over a zero-dimensional domain.
4580 static __isl_give isl_multi_pw_aff *one_mpa(isl_ctx *ctx)
4582 isl_space *space;
4583 isl_local_space *ls;
4584 isl_aff *aff;
4586 space = isl_space_set_alloc(ctx, 0, 0);
4587 ls = isl_local_space_from_space(space);
4588 aff = isl_aff_zero_on_domain(ls);
4589 aff = isl_aff_set_constant_si(aff, 1);
4591 return isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
4594 /* Construct a pet_scop for a continue statement.
4596 * We simply create an empty scop with a universal pet_skip_now
4597 * skip condition. This skip condition will then be taken into
4598 * account by the enclosing loop construct, possibly after
4599 * being incorporated into outer skip conditions.
4601 struct pet_scop *PetScan::extract(ContinueStmt *stmt)
4603 pet_scop *scop;
4605 scop = pet_scop_empty(ctx);
4606 if (!scop)
4607 return NULL;
4609 scop = pet_scop_set_skip(scop, pet_skip_now, one_mpa(ctx));
4611 return scop;
4614 /* Construct a pet_scop for a break statement.
4616 * We simply create an empty scop with both a universal pet_skip_now
4617 * skip condition and a universal pet_skip_later skip condition.
4618 * These skip conditions will then be taken into
4619 * account by the enclosing loop construct, possibly after
4620 * being incorporated into outer skip conditions.
4622 struct pet_scop *PetScan::extract(BreakStmt *stmt)
4624 pet_scop *scop;
4625 isl_multi_pw_aff *skip;
4627 scop = pet_scop_empty(ctx);
4628 if (!scop)
4629 return NULL;
4631 skip = one_mpa(ctx);
4632 scop = pet_scop_set_skip(scop, pet_skip_now,
4633 isl_multi_pw_aff_copy(skip));
4634 scop = pet_scop_set_skip(scop, pet_skip_later, skip);
4636 return scop;
4639 /* Try and construct a pet_scop corresponding to "stmt".
4641 * If "stmt" is a compound statement, then "skip_declarations"
4642 * indicates whether we should skip initial declarations in the
4643 * compound statement.
4645 * If the constructed pet_scop is not a (possibly) partial representation
4646 * of "stmt", we update start and end of the pet_scop to those of "stmt".
4647 * In particular, if skip_declarations is set, then we may have skipped
4648 * declarations inside "stmt" and so the pet_scop may not represent
4649 * the entire "stmt".
4650 * Note that this function may be called with "stmt" referring to the entire
4651 * body of the function, including the outer braces. In such cases,
4652 * skip_declarations will be set and the braces will not be taken into
4653 * account in scop->start and scop->end.
4655 struct pet_scop *PetScan::extract(Stmt *stmt, bool skip_declarations)
4657 struct pet_scop *scop;
4659 if (isa<Expr>(stmt))
4660 return extract(stmt, extract_expr(cast<Expr>(stmt)));
4662 switch (stmt->getStmtClass()) {
4663 case Stmt::WhileStmtClass:
4664 scop = extract(cast<WhileStmt>(stmt));
4665 break;
4666 case Stmt::ForStmtClass:
4667 scop = extract_for(cast<ForStmt>(stmt));
4668 break;
4669 case Stmt::IfStmtClass:
4670 scop = extract(cast<IfStmt>(stmt));
4671 break;
4672 case Stmt::CompoundStmtClass:
4673 scop = extract(cast<CompoundStmt>(stmt), skip_declarations);
4674 break;
4675 case Stmt::LabelStmtClass:
4676 scop = extract(cast<LabelStmt>(stmt));
4677 break;
4678 case Stmt::ContinueStmtClass:
4679 scop = extract(cast<ContinueStmt>(stmt));
4680 break;
4681 case Stmt::BreakStmtClass:
4682 scop = extract(cast<BreakStmt>(stmt));
4683 break;
4684 case Stmt::DeclStmtClass:
4685 scop = extract(cast<DeclStmt>(stmt));
4686 break;
4687 default:
4688 unsupported(stmt);
4689 return NULL;
4692 if (partial || skip_declarations)
4693 return scop;
4695 scop = update_scop_start_end(scop, stmt->getSourceRange(), false);
4697 return scop;
4700 /* Do we need to construct a skip condition of the given type
4701 * on a sequence of statements?
4703 * There is no need to construct a new skip condition if only
4704 * only of the two statements has a skip condition or if both
4705 * of their skip conditions are affine.
4707 * In principle we also don't need a new continuation variable if
4708 * the continuation of scop2 is affine, but then we would need
4709 * to allow more complicated forms of continuations.
4711 static bool need_skip_seq(struct pet_scop *scop1, struct pet_scop *scop2,
4712 enum pet_skip type)
4714 if (!scop1 || !pet_scop_has_skip(scop1, type))
4715 return false;
4716 if (!scop2 || !pet_scop_has_skip(scop2, type))
4717 return false;
4718 if (pet_scop_has_affine_skip(scop1, type) &&
4719 pet_scop_has_affine_skip(scop2, type))
4720 return false;
4721 return true;
4724 /* Construct a scop for computing the skip condition of the given type and
4725 * with index expression "skip_index" for a sequence of two scops "scop1"
4726 * and "scop2".
4728 * The computed scop contains a single statement that essentially does
4730 * skip_index = skip_cond_1 ? 1 : skip_cond_2
4732 * or, in other words, skip_cond1 || skip_cond2.
4733 * In this expression, skip_cond_2 is filtered to reflect that it is
4734 * only evaluated when skip_cond_1 is false.
4736 * The skip condition on scop1 is not removed because it still needs
4737 * to be applied to scop2 when these two scops are combined.
4739 static struct pet_scop *extract_skip_seq(PetScan *ps,
4740 __isl_take isl_multi_pw_aff *skip_index,
4741 struct pet_scop *scop1, struct pet_scop *scop2, enum pet_skip type)
4743 struct pet_expr *expr1, *expr2, *expr, *expr_skip;
4744 struct pet_stmt *stmt;
4745 struct pet_scop *scop;
4746 isl_ctx *ctx = ps->ctx;
4748 if (!scop1 || !scop2)
4749 goto error;
4751 expr1 = pet_scop_get_skip_expr(scop1, type);
4752 expr2 = pet_scop_get_skip_expr(scop2, type);
4753 pet_scop_reset_skip(scop2, type);
4755 expr2 = pet_expr_filter(expr2,
4756 isl_multi_pw_aff_copy(expr1->acc.index), 0);
4758 expr = universally_true(ctx);
4759 expr = pet_expr_new_ternary(ctx, expr1, expr, expr2);
4760 expr_skip = pet_expr_from_index(isl_multi_pw_aff_copy(skip_index));
4761 if (expr_skip) {
4762 expr_skip->acc.write = 1;
4763 expr_skip->acc.read = 0;
4765 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
4766 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, ps->n_stmt++, expr);
4768 scop = pet_scop_from_pet_stmt(ctx, stmt);
4769 scop = scop_add_array(scop, skip_index, ps->ast_context);
4770 isl_multi_pw_aff_free(skip_index);
4772 return scop;
4773 error:
4774 isl_multi_pw_aff_free(skip_index);
4775 return NULL;
4778 /* Structure that handles the construction of skip conditions
4779 * on sequences of statements.
4781 * scop1 and scop2 represent the two statements that are combined
4783 struct pet_skip_info_seq : public pet_skip_info {
4784 struct pet_scop *scop1, *scop2;
4786 pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4787 struct pet_scop *scop2);
4788 void extract(PetScan *scan, enum pet_skip type);
4789 void extract(PetScan *scan);
4790 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
4791 int offset);
4792 struct pet_scop *add(struct pet_scop *scop, int offset);
4795 /* Initialize a pet_skip_info_seq structure based on
4796 * on the two statements that are going to be combined.
4798 pet_skip_info_seq::pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4799 struct pet_scop *scop2) : pet_skip_info(ctx), scop1(scop1), scop2(scop2)
4801 skip[pet_skip_now] = need_skip_seq(scop1, scop2, pet_skip_now);
4802 equal = skip[pet_skip_now] && skip_equals_skip_later(scop1) &&
4803 skip_equals_skip_later(scop2);
4804 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
4805 need_skip_seq(scop1, scop2, pet_skip_later);
4808 /* If we need to construct a skip condition of the given type,
4809 * then do so now.
4811 void pet_skip_info_seq::extract(PetScan *scan, enum pet_skip type)
4813 if (!skip[type])
4814 return;
4816 index[type] = create_test_index(ctx, scan->n_test++);
4817 scop[type] = extract_skip_seq(scan, isl_multi_pw_aff_copy(index[type]),
4818 scop1, scop2, type);
4821 /* Construct the required skip conditions.
4823 void pet_skip_info_seq::extract(PetScan *scan)
4825 extract(scan, pet_skip_now);
4826 extract(scan, pet_skip_later);
4827 if (equal)
4828 drop_skip_later(scop1, scop2);
4831 /* Add the computed skip condition of the given type to "main" and
4832 * add the scop for computing the condition at the given offset (the statement
4833 * number). Within this offset, the condition is computed at position 1
4834 * to ensure that it is computed after the corresponding statement.
4836 * If equal is set, then we only computed a skip condition for pet_skip_now,
4837 * but we also need to set it as main's pet_skip_later.
4839 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *main,
4840 enum pet_skip type, int offset)
4842 if (!skip[type])
4843 return main;
4845 scop[type] = pet_scop_prefix(scop[type], 1);
4846 scop[type] = pet_scop_prefix(scop[type], offset);
4847 main = pet_scop_add_par(ctx, main, scop[type]);
4848 scop[type] = NULL;
4850 if (equal)
4851 main = pet_scop_set_skip(main, pet_skip_later,
4852 isl_multi_pw_aff_copy(index[type]));
4854 main = pet_scop_set_skip(main, type, index[type]);
4855 index[type] = NULL;
4857 return main;
4860 /* Add the computed skip conditions to "main" and
4861 * add the scops for computing the conditions at the given offset.
4863 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *scop, int offset)
4865 scop = add(scop, pet_skip_now, offset);
4866 scop = add(scop, pet_skip_later, offset);
4868 return scop;
4871 /* Extract a clone of the kill statement in "scop".
4872 * "scop" is expected to have been created from a DeclStmt
4873 * and should have the kill as its first statement.
4875 struct pet_stmt *PetScan::extract_kill(struct pet_scop *scop)
4877 struct pet_expr *kill;
4878 struct pet_stmt *stmt;
4879 isl_multi_pw_aff *index;
4880 isl_map *access;
4882 if (!scop)
4883 return NULL;
4884 if (scop->n_stmt < 1)
4885 isl_die(ctx, isl_error_internal,
4886 "expecting at least one statement", return NULL);
4887 stmt = scop->stmts[0];
4888 if (!pet_stmt_is_kill(stmt))
4889 isl_die(ctx, isl_error_internal,
4890 "expecting kill statement", return NULL);
4892 index = isl_multi_pw_aff_copy(stmt->body->args[0]->acc.index);
4893 access = isl_map_copy(stmt->body->args[0]->acc.access);
4894 index = isl_multi_pw_aff_reset_tuple_id(index, isl_dim_in);
4895 access = isl_map_reset_tuple_id(access, isl_dim_in);
4896 kill = pet_expr_kill_from_access_and_index(access, index);
4897 return pet_stmt_from_pet_expr(ctx, stmt->line, NULL, n_stmt++, kill);
4900 /* Mark all arrays in "scop" as being exposed.
4902 static struct pet_scop *mark_exposed(struct pet_scop *scop)
4904 if (!scop)
4905 return NULL;
4906 for (int i = 0; i < scop->n_array; ++i)
4907 scop->arrays[i]->exposed = 1;
4908 return scop;
4911 /* Try and construct a pet_scop corresponding to (part of)
4912 * a sequence of statements.
4914 * "block" is set if the sequence respresents the children of
4915 * a compound statement.
4916 * "skip_declarations" is set if we should skip initial declarations
4917 * in the sequence of statements.
4919 * If there are any breaks or continues in the individual statements,
4920 * then we may have to compute a new skip condition.
4921 * This is handled using a pet_skip_info_seq object.
4922 * On initialization, the object checks if skip conditions need
4923 * to be computed. If so, it does so in "extract" and adds them in "add".
4925 * If "block" is set, then we need to insert kill statements at
4926 * the end of the block for any array that has been declared by
4927 * one of the statements in the sequence. Each of these declarations
4928 * results in the construction of a kill statement at the place
4929 * of the declaration, so we simply collect duplicates of
4930 * those kill statements and append these duplicates to the constructed scop.
4932 * If "block" is not set, then any array declared by one of the statements
4933 * in the sequence is marked as being exposed.
4935 * If autodetect is set, then we allow the extraction of only a subrange
4936 * of the sequence of statements. However, if there is at least one statement
4937 * for which we could not construct a scop and the final range contains
4938 * either no statements or at least one kill, then we discard the entire
4939 * range.
4941 struct pet_scop *PetScan::extract(StmtRange stmt_range, bool block,
4942 bool skip_declarations)
4944 pet_scop *scop;
4945 StmtIterator i;
4946 int j;
4947 bool partial_range = false;
4948 set<struct pet_stmt *> kills;
4949 set<struct pet_stmt *>::iterator it;
4951 scop = pet_scop_empty(ctx);
4952 for (i = stmt_range.first, j = 0; i != stmt_range.second; ++i, ++j) {
4953 Stmt *child = *i;
4954 struct pet_scop *scop_i;
4956 if (scop->n_stmt == 0 && skip_declarations &&
4957 child->getStmtClass() == Stmt::DeclStmtClass)
4958 continue;
4960 scop_i = extract(child);
4961 if (scop->n_stmt != 0 && partial) {
4962 pet_scop_free(scop_i);
4963 break;
4965 pet_skip_info_seq skip(ctx, scop, scop_i);
4966 skip.extract(this);
4967 if (skip)
4968 scop_i = pet_scop_prefix(scop_i, 0);
4969 if (scop_i && child->getStmtClass() == Stmt::DeclStmtClass) {
4970 if (block)
4971 kills.insert(extract_kill(scop_i));
4972 else
4973 scop_i = mark_exposed(scop_i);
4975 scop_i = pet_scop_prefix(scop_i, j);
4976 if (options->autodetect) {
4977 if (scop_i)
4978 scop = pet_scop_add_seq(ctx, scop, scop_i);
4979 else
4980 partial_range = true;
4981 if (scop->n_stmt != 0 && !scop_i)
4982 partial = true;
4983 } else {
4984 scop = pet_scop_add_seq(ctx, scop, scop_i);
4987 scop = skip.add(scop, j);
4989 if (partial || !scop)
4990 break;
4993 for (it = kills.begin(); it != kills.end(); ++it) {
4994 pet_scop *scop_j;
4995 scop_j = pet_scop_from_pet_stmt(ctx, *it);
4996 scop_j = pet_scop_prefix(scop_j, j);
4997 scop = pet_scop_add_seq(ctx, scop, scop_j);
5000 if (scop && partial_range) {
5001 if (scop->n_stmt == 0 || kills.size() != 0) {
5002 pet_scop_free(scop);
5003 return NULL;
5005 partial = true;
5008 return scop;
5011 /* Check if the scop marked by the user is exactly this Stmt
5012 * or part of this Stmt.
5013 * If so, return a pet_scop corresponding to the marked region.
5014 * Otherwise, return NULL.
5016 struct pet_scop *PetScan::scan(Stmt *stmt)
5018 SourceManager &SM = PP.getSourceManager();
5019 unsigned start_off, end_off;
5021 start_off = getExpansionOffset(SM, stmt->getLocStart());
5022 end_off = getExpansionOffset(SM, stmt->getLocEnd());
5024 if (start_off > loc.end)
5025 return NULL;
5026 if (end_off < loc.start)
5027 return NULL;
5028 if (start_off >= loc.start && end_off <= loc.end) {
5029 return extract(stmt);
5032 StmtIterator start;
5033 for (start = stmt->child_begin(); start != stmt->child_end(); ++start) {
5034 Stmt *child = *start;
5035 if (!child)
5036 continue;
5037 start_off = getExpansionOffset(SM, child->getLocStart());
5038 end_off = getExpansionOffset(SM, child->getLocEnd());
5039 if (start_off < loc.start && end_off >= loc.end)
5040 return scan(child);
5041 if (start_off >= loc.start)
5042 break;
5045 StmtIterator end;
5046 for (end = start; end != stmt->child_end(); ++end) {
5047 Stmt *child = *end;
5048 start_off = SM.getFileOffset(child->getLocStart());
5049 if (start_off >= loc.end)
5050 break;
5053 return extract(StmtRange(start, end), false, false);
5056 /* Set the size of index "pos" of "array" to "size".
5057 * In particular, add a constraint of the form
5059 * i_pos < size
5061 * to array->extent and a constraint of the form
5063 * size >= 0
5065 * to array->context.
5067 static struct pet_array *update_size(struct pet_array *array, int pos,
5068 __isl_take isl_pw_aff *size)
5070 isl_set *valid;
5071 isl_set *univ;
5072 isl_set *bound;
5073 isl_space *dim;
5074 isl_aff *aff;
5075 isl_pw_aff *index;
5076 isl_id *id;
5078 valid = isl_pw_aff_nonneg_set(isl_pw_aff_copy(size));
5079 array->context = isl_set_intersect(array->context, valid);
5081 dim = isl_set_get_space(array->extent);
5082 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
5083 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, pos, 1);
5084 univ = isl_set_universe(isl_aff_get_domain_space(aff));
5085 index = isl_pw_aff_alloc(univ, aff);
5087 size = isl_pw_aff_add_dims(size, isl_dim_in,
5088 isl_set_dim(array->extent, isl_dim_set));
5089 id = isl_set_get_tuple_id(array->extent);
5090 size = isl_pw_aff_set_tuple_id(size, isl_dim_in, id);
5091 bound = isl_pw_aff_lt_set(index, size);
5093 array->extent = isl_set_intersect(array->extent, bound);
5095 if (!array->context || !array->extent)
5096 goto error;
5098 return array;
5099 error:
5100 pet_array_free(array);
5101 return NULL;
5104 /* Figure out the size of the array at position "pos" and all
5105 * subsequent positions from "type" and update "array" accordingly.
5107 struct pet_array *PetScan::set_upper_bounds(struct pet_array *array,
5108 const Type *type, int pos)
5110 const ArrayType *atype;
5111 isl_pw_aff *size;
5113 if (!array)
5114 return NULL;
5116 if (type->isPointerType()) {
5117 type = type->getPointeeType().getTypePtr();
5118 return set_upper_bounds(array, type, pos + 1);
5120 if (!type->isArrayType())
5121 return array;
5123 type = type->getCanonicalTypeInternal().getTypePtr();
5124 atype = cast<ArrayType>(type);
5126 if (type->isConstantArrayType()) {
5127 const ConstantArrayType *ca = cast<ConstantArrayType>(atype);
5128 size = extract_affine(ca->getSize());
5129 array = update_size(array, pos, size);
5130 } else if (type->isVariableArrayType()) {
5131 const VariableArrayType *vla = cast<VariableArrayType>(atype);
5132 size = extract_affine(vla->getSizeExpr());
5133 array = update_size(array, pos, size);
5136 type = atype->getElementType().getTypePtr();
5138 return set_upper_bounds(array, type, pos + 1);
5141 /* Is "T" the type of a variable length array with static size?
5143 static bool is_vla_with_static_size(QualType T)
5145 const VariableArrayType *vlatype;
5147 if (!T->isVariableArrayType())
5148 return false;
5149 vlatype = cast<VariableArrayType>(T);
5150 return vlatype->getSizeModifier() == VariableArrayType::Static;
5153 /* Return the type of "decl" as an array.
5155 * In particular, if "decl" is a parameter declaration that
5156 * is a variable length array with a static size, then
5157 * return the original type (i.e., the variable length array).
5158 * Otherwise, return the type of decl.
5160 static QualType get_array_type(ValueDecl *decl)
5162 ParmVarDecl *parm;
5163 QualType T;
5165 parm = dyn_cast<ParmVarDecl>(decl);
5166 if (!parm)
5167 return decl->getType();
5169 T = parm->getOriginalType();
5170 if (!is_vla_with_static_size(T))
5171 return decl->getType();
5172 return T;
5175 /* Does "decl" have definition that we can keep track of in a pet_type?
5177 static bool has_printable_definition(RecordDecl *decl)
5179 if (!decl->getDeclName())
5180 return false;
5181 return decl->getLexicalDeclContext() == decl->getDeclContext();
5184 /* Construct and return a pet_array corresponding to the variable "decl".
5185 * In particular, initialize array->extent to
5187 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
5189 * and then call set_upper_bounds to set the upper bounds on the indices
5190 * based on the type of the variable.
5192 * If the base type is that of a record with a top-level definition and
5193 * if "types" is not null, then the RecordDecl corresponding to the type
5194 * is added to "types".
5196 * If the base type is that of a record with no top-level definition,
5197 * then we replace it by "<subfield>".
5199 struct pet_array *PetScan::extract_array(isl_ctx *ctx, ValueDecl *decl,
5200 lex_recorddecl_set *types)
5202 struct pet_array *array;
5203 QualType qt = get_array_type(decl);
5204 const Type *type = qt.getTypePtr();
5205 int depth = array_depth(type);
5206 QualType base = pet_clang_base_type(qt);
5207 string name;
5208 isl_id *id;
5209 isl_space *dim;
5211 array = isl_calloc_type(ctx, struct pet_array);
5212 if (!array)
5213 return NULL;
5215 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
5216 dim = isl_space_set_alloc(ctx, 0, depth);
5217 dim = isl_space_set_tuple_id(dim, isl_dim_set, id);
5219 array->extent = isl_set_nat_universe(dim);
5221 dim = isl_space_params_alloc(ctx, 0);
5222 array->context = isl_set_universe(dim);
5224 array = set_upper_bounds(array, type, 0);
5225 if (!array)
5226 return NULL;
5228 name = base.getAsString();
5230 if (types && base->isRecordType()) {
5231 RecordDecl *decl = pet_clang_record_decl(base);
5232 if (has_printable_definition(decl))
5233 types->insert(decl);
5234 else
5235 name = "<subfield>";
5238 array->element_type = strdup(name.c_str());
5239 array->element_is_record = base->isRecordType();
5240 array->element_size = decl->getASTContext().getTypeInfo(base).first / 8;
5242 return array;
5245 /* Construct and return a pet_array corresponding to the sequence
5246 * of declarations "decls".
5247 * If the sequence contains a single declaration, then it corresponds
5248 * to a simple array access. Otherwise, it corresponds to a member access,
5249 * with the declaration for the substructure following that of the containing
5250 * structure in the sequence of declarations.
5251 * We start with the outermost substructure and then combine it with
5252 * information from the inner structures.
5254 * Additionally, keep track of all required types in "types".
5256 struct pet_array *PetScan::extract_array(isl_ctx *ctx,
5257 vector<ValueDecl *> decls, lex_recorddecl_set *types)
5259 struct pet_array *array;
5260 vector<ValueDecl *>::iterator it;
5262 it = decls.begin();
5264 array = extract_array(ctx, *it, types);
5266 for (++it; it != decls.end(); ++it) {
5267 struct pet_array *parent;
5268 const char *base_name, *field_name;
5269 char *product_name;
5271 parent = array;
5272 array = extract_array(ctx, *it, types);
5273 if (!array)
5274 return pet_array_free(parent);
5276 base_name = isl_set_get_tuple_name(parent->extent);
5277 field_name = isl_set_get_tuple_name(array->extent);
5278 product_name = member_access_name(ctx, base_name, field_name);
5280 array->extent = isl_set_product(isl_set_copy(parent->extent),
5281 array->extent);
5282 if (product_name)
5283 array->extent = isl_set_set_tuple_name(array->extent,
5284 product_name);
5285 array->context = isl_set_intersect(array->context,
5286 isl_set_copy(parent->context));
5288 pet_array_free(parent);
5289 free(product_name);
5291 if (!array->extent || !array->context || !product_name)
5292 return pet_array_free(array);
5295 return array;
5298 /* Add a pet_type corresponding to "decl" to "scop, provided
5299 * it is a member of "types" and it has not been added before
5300 * (i.e., it is not a member of "types_done".
5302 * Since we want the user to be able to print the types
5303 * in the order in which they appear in the scop, we need to
5304 * make sure that types of fields in a structure appear before
5305 * that structure. We therefore call ourselves recursively
5306 * on the types of all record subfields.
5308 static struct pet_scop *add_type(isl_ctx *ctx, struct pet_scop *scop,
5309 RecordDecl *decl, Preprocessor &PP, lex_recorddecl_set &types,
5310 lex_recorddecl_set &types_done)
5312 string s;
5313 llvm::raw_string_ostream S(s);
5314 RecordDecl::field_iterator it;
5316 if (types.find(decl) == types.end())
5317 return scop;
5318 if (types_done.find(decl) != types_done.end())
5319 return scop;
5321 for (it = decl->field_begin(); it != decl->field_end(); ++it) {
5322 RecordDecl *record;
5323 QualType type = it->getType();
5325 if (!type->isRecordType())
5326 continue;
5327 record = pet_clang_record_decl(type);
5328 scop = add_type(ctx, scop, record, PP, types, types_done);
5331 if (strlen(decl->getName().str().c_str()) == 0)
5332 return scop;
5334 decl->print(S, PrintingPolicy(PP.getLangOpts()));
5335 S.str();
5337 scop->types[scop->n_type] = pet_type_alloc(ctx,
5338 decl->getName().str().c_str(), s.c_str());
5339 if (!scop->types[scop->n_type])
5340 return pet_scop_free(scop);
5342 types_done.insert(decl);
5344 scop->n_type++;
5346 return scop;
5349 /* Construct a list of pet_arrays, one for each array (or scalar)
5350 * accessed inside "scop", add this list to "scop" and return the result.
5352 * The context of "scop" is updated with the intersection of
5353 * the contexts of all arrays, i.e., constraints on the parameters
5354 * that ensure that the arrays have a valid (non-negative) size.
5356 * If the any of the extracted arrays refers to a member access,
5357 * then also add the required types to "scop".
5359 struct pet_scop *PetScan::scan_arrays(struct pet_scop *scop)
5361 int i;
5362 set<vector<ValueDecl *> > arrays;
5363 set<vector<ValueDecl *> >::iterator it;
5364 lex_recorddecl_set types;
5365 lex_recorddecl_set types_done;
5366 lex_recorddecl_set::iterator types_it;
5367 int n_array;
5368 struct pet_array **scop_arrays;
5370 if (!scop)
5371 return NULL;
5373 pet_scop_collect_arrays(scop, arrays);
5374 if (arrays.size() == 0)
5375 return scop;
5377 n_array = scop->n_array;
5379 scop_arrays = isl_realloc_array(ctx, scop->arrays, struct pet_array *,
5380 n_array + arrays.size());
5381 if (!scop_arrays)
5382 goto error;
5383 scop->arrays = scop_arrays;
5385 for (it = arrays.begin(), i = 0; it != arrays.end(); ++it, ++i) {
5386 struct pet_array *array;
5387 array = extract_array(ctx, *it, &types);
5388 scop->arrays[n_array + i] = array;
5389 if (!scop->arrays[n_array + i])
5390 goto error;
5391 scop->n_array++;
5392 scop->context = isl_set_intersect(scop->context,
5393 isl_set_copy(array->context));
5394 if (!scop->context)
5395 goto error;
5398 if (types.size() == 0)
5399 return scop;
5401 scop->types = isl_alloc_array(ctx, struct pet_type *, types.size());
5402 if (!scop->types)
5403 goto error;
5405 for (types_it = types.begin(); types_it != types.end(); ++types_it)
5406 scop = add_type(ctx, scop, *types_it, PP, types, types_done);
5408 return scop;
5409 error:
5410 pet_scop_free(scop);
5411 return NULL;
5414 /* Bound all parameters in scop->context to the possible values
5415 * of the corresponding C variable.
5417 static struct pet_scop *add_parameter_bounds(struct pet_scop *scop)
5419 int n;
5421 if (!scop)
5422 return NULL;
5424 n = isl_set_dim(scop->context, isl_dim_param);
5425 for (int i = 0; i < n; ++i) {
5426 isl_id *id;
5427 ValueDecl *decl;
5429 id = isl_set_get_dim_id(scop->context, isl_dim_param, i);
5430 if (pet_nested_in_id(id)) {
5431 isl_id_free(id);
5432 isl_die(isl_set_get_ctx(scop->context),
5433 isl_error_internal,
5434 "unresolved nested parameter", goto error);
5436 decl = (ValueDecl *) isl_id_get_user(id);
5437 isl_id_free(id);
5439 scop->context = set_parameter_bounds(scop->context, i, decl);
5441 if (!scop->context)
5442 goto error;
5445 return scop;
5446 error:
5447 pet_scop_free(scop);
5448 return NULL;
5451 /* Construct a pet_scop from the given function.
5453 * If the scop was delimited by scop and endscop pragmas, then we override
5454 * the file offsets by those derived from the pragmas.
5456 struct pet_scop *PetScan::scan(FunctionDecl *fd)
5458 pet_scop *scop;
5459 Stmt *stmt;
5461 stmt = fd->getBody();
5463 if (options->autodetect)
5464 scop = extract(stmt, true);
5465 else {
5466 scop = scan(stmt);
5467 scop = pet_scop_update_start_end(scop, loc.start, loc.end);
5469 scop = pet_scop_detect_parameter_accesses(scop);
5470 scop = scan_arrays(scop);
5471 scop = add_parameter_bounds(scop);
5472 scop = pet_scop_gist(scop, value_bounds);
5474 return scop;