remove PetScan::try_extract_affine_condition
[pet.git] / scan.cc
blob93e35a62dd2168b6a032549a346de382e58736ca
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 "aff.h"
51 #include "array.h"
52 #include "clang.h"
53 #include "context.h"
54 #include "expr.h"
55 #include "expr_arg.h"
56 #include "nest.h"
57 #include "options.h"
58 #include "scan.h"
59 #include "scop.h"
60 #include "scop_plus.h"
61 #include "skip.h"
63 #include "config.h"
65 using namespace std;
66 using namespace clang;
68 static enum pet_op_type UnaryOperatorKind2pet_op_type(UnaryOperatorKind kind)
70 switch (kind) {
71 case UO_Minus:
72 return pet_op_minus;
73 case UO_Not:
74 return pet_op_not;
75 case UO_LNot:
76 return pet_op_lnot;
77 case UO_PostInc:
78 return pet_op_post_inc;
79 case UO_PostDec:
80 return pet_op_post_dec;
81 case UO_PreInc:
82 return pet_op_pre_inc;
83 case UO_PreDec:
84 return pet_op_pre_dec;
85 default:
86 return pet_op_last;
90 static enum pet_op_type BinaryOperatorKind2pet_op_type(BinaryOperatorKind kind)
92 switch (kind) {
93 case BO_AddAssign:
94 return pet_op_add_assign;
95 case BO_SubAssign:
96 return pet_op_sub_assign;
97 case BO_MulAssign:
98 return pet_op_mul_assign;
99 case BO_DivAssign:
100 return pet_op_div_assign;
101 case BO_Assign:
102 return pet_op_assign;
103 case BO_Add:
104 return pet_op_add;
105 case BO_Sub:
106 return pet_op_sub;
107 case BO_Mul:
108 return pet_op_mul;
109 case BO_Div:
110 return pet_op_div;
111 case BO_Rem:
112 return pet_op_mod;
113 case BO_Shl:
114 return pet_op_shl;
115 case BO_Shr:
116 return pet_op_shr;
117 case BO_EQ:
118 return pet_op_eq;
119 case BO_NE:
120 return pet_op_ne;
121 case BO_LE:
122 return pet_op_le;
123 case BO_GE:
124 return pet_op_ge;
125 case BO_LT:
126 return pet_op_lt;
127 case BO_GT:
128 return pet_op_gt;
129 case BO_And:
130 return pet_op_and;
131 case BO_Xor:
132 return pet_op_xor;
133 case BO_Or:
134 return pet_op_or;
135 case BO_LAnd:
136 return pet_op_land;
137 case BO_LOr:
138 return pet_op_lor;
139 default:
140 return pet_op_last;
144 #if defined(DECLREFEXPR_CREATE_REQUIRES_BOOL)
145 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
147 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
148 SourceLocation(), var, false, var->getInnerLocStart(),
149 var->getType(), VK_LValue);
151 #elif defined(DECLREFEXPR_CREATE_REQUIRES_SOURCELOCATION)
152 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
154 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
155 SourceLocation(), var, var->getInnerLocStart(), var->getType(),
156 VK_LValue);
158 #else
159 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
161 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
162 var, var->getInnerLocStart(), var->getType(), VK_LValue);
164 #endif
166 /* Check if the element type corresponding to the given array type
167 * has a const qualifier.
169 static bool const_base(QualType qt)
171 const Type *type = qt.getTypePtr();
173 if (type->isPointerType())
174 return const_base(type->getPointeeType());
175 if (type->isArrayType()) {
176 const ArrayType *atype;
177 type = type->getCanonicalTypeInternal().getTypePtr();
178 atype = cast<ArrayType>(type);
179 return const_base(atype->getElementType());
182 return qt.isConstQualified();
185 /* Create an isl_id that refers to the named declarator "decl".
187 static __isl_give isl_id *create_decl_id(isl_ctx *ctx, NamedDecl *decl)
189 return isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
192 /* Mark "decl" as having an unknown value in "assigned_value".
194 * If no (known or unknown) value was assigned to "decl" before,
195 * then it may have been treated as a parameter before and may
196 * therefore appear in a value assigned to another variable.
197 * If so, this assignment needs to be turned into an unknown value too.
199 static void clear_assignment(map<ValueDecl *, isl_pw_aff *> &assigned_value,
200 ValueDecl *decl)
202 map<ValueDecl *, isl_pw_aff *>::iterator it;
204 it = assigned_value.find(decl);
206 assigned_value[decl] = NULL;
208 if (it != assigned_value.end())
209 return;
211 for (it = assigned_value.begin(); it != assigned_value.end(); ++it) {
212 isl_pw_aff *pa = it->second;
213 int nparam = isl_pw_aff_dim(pa, isl_dim_param);
215 for (int i = 0; i < nparam; ++i) {
216 isl_id *id;
218 if (!isl_pw_aff_has_dim_id(pa, isl_dim_param, i))
219 continue;
220 id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
221 if (isl_id_get_user(id) == decl)
222 it->second = NULL;
223 isl_id_free(id);
228 /* Look for any assignments to scalar variables in part of the parse
229 * tree and set assigned_value to NULL for each of them.
230 * Also reset assigned_value if the address of a scalar variable
231 * is being taken. As an exception, if the address is passed to a function
232 * that is declared to receive a const pointer, then assigned_value is
233 * not reset.
235 * This ensures that we won't use any previously stored value
236 * in the current subtree and its parents.
238 struct clear_assignments : RecursiveASTVisitor<clear_assignments> {
239 map<ValueDecl *, isl_pw_aff *> &assigned_value;
240 set<UnaryOperator *> skip;
242 clear_assignments(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
243 assigned_value(assigned_value) {}
245 /* Check for "address of" operators whose value is passed
246 * to a const pointer argument and add them to "skip", so that
247 * we can skip them in VisitUnaryOperator.
249 bool VisitCallExpr(CallExpr *expr) {
250 FunctionDecl *fd;
251 fd = expr->getDirectCallee();
252 if (!fd)
253 return true;
254 for (int i = 0; i < expr->getNumArgs(); ++i) {
255 Expr *arg = expr->getArg(i);
256 UnaryOperator *op;
257 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
258 ImplicitCastExpr *ice;
259 ice = cast<ImplicitCastExpr>(arg);
260 arg = ice->getSubExpr();
262 if (arg->getStmtClass() != Stmt::UnaryOperatorClass)
263 continue;
264 op = cast<UnaryOperator>(arg);
265 if (op->getOpcode() != UO_AddrOf)
266 continue;
267 if (const_base(fd->getParamDecl(i)->getType()))
268 skip.insert(op);
270 return true;
273 bool VisitUnaryOperator(UnaryOperator *expr) {
274 Expr *arg;
275 DeclRefExpr *ref;
276 ValueDecl *decl;
278 switch (expr->getOpcode()) {
279 case UO_AddrOf:
280 case UO_PostInc:
281 case UO_PostDec:
282 case UO_PreInc:
283 case UO_PreDec:
284 break;
285 default:
286 return true;
288 if (skip.find(expr) != skip.end())
289 return true;
291 arg = expr->getSubExpr();
292 if (arg->getStmtClass() != Stmt::DeclRefExprClass)
293 return true;
294 ref = cast<DeclRefExpr>(arg);
295 decl = ref->getDecl();
296 clear_assignment(assigned_value, decl);
297 return true;
300 bool VisitBinaryOperator(BinaryOperator *expr) {
301 Expr *lhs;
302 DeclRefExpr *ref;
303 ValueDecl *decl;
305 if (!expr->isAssignmentOp())
306 return true;
307 lhs = expr->getLHS();
308 if (lhs->getStmtClass() != Stmt::DeclRefExprClass)
309 return true;
310 ref = cast<DeclRefExpr>(lhs);
311 decl = ref->getDecl();
312 clear_assignment(assigned_value, decl);
313 return true;
317 /* Keep a copy of the currently assigned values.
319 * Any variable that is assigned a value inside the current scope
320 * is removed again when we leave the scope (either because it wasn't
321 * stored in the cache or because it has a different value in the cache).
323 struct assigned_value_cache {
324 map<ValueDecl *, isl_pw_aff *> &assigned_value;
325 map<ValueDecl *, isl_pw_aff *> cache;
327 assigned_value_cache(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
328 assigned_value(assigned_value), cache(assigned_value) {}
329 ~assigned_value_cache() {
330 map<ValueDecl *, isl_pw_aff *>::iterator it = cache.begin();
331 for (it = assigned_value.begin(); it != assigned_value.end();
332 ++it) {
333 if (!it->second ||
334 (cache.find(it->first) != cache.end() &&
335 cache[it->first] != it->second))
336 cache[it->first] = NULL;
338 assigned_value = cache;
342 /* Convert the mapping from identifiers to values in "assigned_value"
343 * to a pet_context to be used by pet_expr_extract_*.
344 * In particular, the clang identifiers are wrapped in an isl_id and
345 * a NULL value (representing an unknown value) is replaced by a NaN.
347 static __isl_give pet_context *convert_assignments(isl_ctx *ctx,
348 map<ValueDecl *, isl_pw_aff *> &assigned_value)
350 pet_context *pc;
351 map<ValueDecl *, isl_pw_aff *>::iterator it;
353 pc = pet_context_alloc(isl_space_set_alloc(ctx, 0, 0));
355 for (it = assigned_value.begin(); it != assigned_value.end(); ++it) {
356 ValueDecl *decl = it->first;
357 isl_pw_aff *pa = it->second;
358 isl_id *id;
360 id = create_decl_id(ctx, decl);
361 if (pa)
362 pc = pet_context_set_value(pc, id, isl_pw_aff_copy(pa));
363 else
364 pc = pet_context_mark_unknown(pc, id);
367 return pc;
370 /* Insert an expression into the collection of expressions,
371 * provided it is not already in there.
372 * The isl_pw_affs are freed in the destructor.
374 void PetScan::insert_expression(__isl_take isl_pw_aff *expr)
376 std::set<isl_pw_aff *>::iterator it;
378 if (expressions.find(expr) == expressions.end())
379 expressions.insert(expr);
380 else
381 isl_pw_aff_free(expr);
384 PetScan::~PetScan()
386 std::set<isl_pw_aff *>::iterator it;
388 for (it = expressions.begin(); it != expressions.end(); ++it)
389 isl_pw_aff_free(*it);
391 isl_union_map_free(value_bounds);
394 /* Report a diagnostic, unless autodetect is set.
396 void PetScan::report(Stmt *stmt, unsigned id)
398 if (options->autodetect)
399 return;
401 SourceLocation loc = stmt->getLocStart();
402 DiagnosticsEngine &diag = PP.getDiagnostics();
403 DiagnosticBuilder B = diag.Report(loc, id) << stmt->getSourceRange();
406 /* Called if we found something we (currently) cannot handle.
407 * We'll provide more informative warnings later.
409 * We only actually complain if autodetect is false.
411 void PetScan::unsupported(Stmt *stmt)
413 DiagnosticsEngine &diag = PP.getDiagnostics();
414 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
415 "unsupported");
416 report(stmt, id);
419 /* Report a missing prototype, unless autodetect is set.
421 void PetScan::report_prototype_required(Stmt *stmt)
423 DiagnosticsEngine &diag = PP.getDiagnostics();
424 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
425 "prototype required");
426 report(stmt, id);
429 /* Report a missing increment, unless autodetect is set.
431 void PetScan::report_missing_increment(Stmt *stmt)
433 DiagnosticsEngine &diag = PP.getDiagnostics();
434 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
435 "missing increment");
436 report(stmt, id);
439 /* Extract an integer from "expr".
441 __isl_give isl_val *PetScan::extract_int(isl_ctx *ctx, IntegerLiteral *expr)
443 const Type *type = expr->getType().getTypePtr();
444 int is_signed = type->hasSignedIntegerRepresentation();
445 llvm::APInt val = expr->getValue();
446 int is_negative = is_signed && val.isNegative();
447 isl_val *v;
449 if (is_negative)
450 val = -val;
452 v = extract_unsigned(ctx, val);
454 if (is_negative)
455 v = isl_val_neg(v);
456 return v;
459 /* Extract an integer from "val", which is assumed to be non-negative.
461 __isl_give isl_val *PetScan::extract_unsigned(isl_ctx *ctx,
462 const llvm::APInt &val)
464 unsigned n;
465 const uint64_t *data;
467 data = val.getRawData();
468 n = val.getNumWords();
469 return isl_val_int_from_chunks(ctx, n, sizeof(uint64_t), data);
472 /* Extract an integer from "expr".
473 * Return NULL if "expr" does not (obviously) represent an integer.
475 __isl_give isl_val *PetScan::extract_int(clang::ParenExpr *expr)
477 return extract_int(expr->getSubExpr());
480 /* Extract an integer from "expr".
481 * Return NULL if "expr" does not (obviously) represent an integer.
483 __isl_give isl_val *PetScan::extract_int(clang::Expr *expr)
485 if (expr->getStmtClass() == Stmt::IntegerLiteralClass)
486 return extract_int(ctx, cast<IntegerLiteral>(expr));
487 if (expr->getStmtClass() == Stmt::ParenExprClass)
488 return extract_int(cast<ParenExpr>(expr));
490 unsupported(expr);
491 return NULL;
494 /* Extract a pet_expr from the APInt "val", which is assumed
495 * to be non-negative.
497 __isl_give pet_expr *PetScan::extract_expr(const llvm::APInt &val)
499 return pet_expr_new_int(extract_unsigned(ctx, val));
502 /* Return the number of bits needed to represent the type "qt",
503 * if it is an integer type. Otherwise return 0.
504 * If qt is signed then return the opposite of the number of bits.
506 static int get_type_size(QualType qt, ASTContext &ast_context)
508 int size;
510 if (!qt->isIntegerType())
511 return 0;
513 size = ast_context.getIntWidth(qt);
514 if (!qt->isUnsignedIntegerType())
515 size = -size;
517 return size;
520 /* Return the number of bits needed to represent the type of "decl",
521 * if it is an integer type. Otherwise return 0.
522 * If qt is signed then return the opposite of the number of bits.
524 static int get_type_size(ValueDecl *decl)
526 return get_type_size(decl->getType(), decl->getASTContext());
529 /* Bound parameter "pos" of "set" to the possible values of "decl".
531 static __isl_give isl_set *set_parameter_bounds(__isl_take isl_set *set,
532 unsigned pos, ValueDecl *decl)
534 int type_size;
535 isl_ctx *ctx;
536 isl_val *bound;
538 ctx = isl_set_get_ctx(set);
539 type_size = get_type_size(decl);
540 if (type_size == 0)
541 isl_die(ctx, isl_error_invalid, "not an integer type",
542 return isl_set_free(set));
543 if (type_size > 0) {
544 set = isl_set_lower_bound_si(set, isl_dim_param, pos, 0);
545 bound = isl_val_int_from_ui(ctx, type_size);
546 bound = isl_val_2exp(bound);
547 bound = isl_val_sub_ui(bound, 1);
548 set = isl_set_upper_bound_val(set, isl_dim_param, pos, bound);
549 } else {
550 bound = isl_val_int_from_ui(ctx, -type_size - 1);
551 bound = isl_val_2exp(bound);
552 bound = isl_val_sub_ui(bound, 1);
553 set = isl_set_upper_bound_val(set, isl_dim_param, pos,
554 isl_val_copy(bound));
555 bound = isl_val_neg(bound);
556 bound = isl_val_sub_ui(bound, 1);
557 set = isl_set_lower_bound_val(set, isl_dim_param, pos, bound);
560 return set;
563 /* Return the piecewise affine expression "set ? 1 : 0" defined on "dom".
565 static __isl_give isl_pw_aff *indicator_function(__isl_take isl_set *set,
566 __isl_take isl_set *dom)
568 isl_pw_aff *pa;
569 pa = isl_set_indicator_function(set);
570 pa = isl_pw_aff_intersect_domain(pa, isl_set_coalesce(dom));
571 return pa;
574 /* Extract an affine expression, if possible, from "expr".
575 * Otherwise return NULL.
577 __isl_give isl_pw_aff *PetScan::extract_affine(Expr *expr)
579 pet_expr *pe;
580 pet_context *pc;
581 isl_pw_aff *pa;
583 pe = extract_expr(expr);
584 if (!pe)
585 return NULL;
586 pc = convert_assignments(ctx, assigned_value);
587 pe = pet_expr_plug_in_args(pe, pc);
588 pa = pet_expr_extract_affine(pe, pc);
589 if (isl_pw_aff_involves_nan(pa)) {
590 unsupported(expr);
591 pa = isl_pw_aff_free(pa);
593 pet_context_free(pc);
594 pet_expr_free(pe);
596 return pa;
599 __isl_give pet_expr *PetScan::extract_index_expr(ImplicitCastExpr *expr)
601 return extract_index_expr(expr->getSubExpr());
604 /* Return the depth of an array of the given type.
606 static int array_depth(const Type *type)
608 if (type->isPointerType())
609 return 1 + array_depth(type->getPointeeType().getTypePtr());
610 if (type->isArrayType()) {
611 const ArrayType *atype;
612 type = type->getCanonicalTypeInternal().getTypePtr();
613 atype = cast<ArrayType>(type);
614 return 1 + array_depth(atype->getElementType().getTypePtr());
616 return 0;
619 /* Return the depth of the array accessed by the index expression "index".
620 * If "index" is an affine expression, i.e., if it does not access
621 * any array, then return 1.
622 * If "index" represent a member access, i.e., if its range is a wrapped
623 * relation, then return the sum of the depth of the array of structures
624 * and that of the member inside the structure.
626 static int extract_depth(__isl_keep isl_multi_pw_aff *index)
628 isl_id *id;
629 ValueDecl *decl;
631 if (!index)
632 return -1;
634 if (isl_multi_pw_aff_range_is_wrapping(index)) {
635 int domain_depth, range_depth;
636 isl_multi_pw_aff *domain, *range;
638 domain = isl_multi_pw_aff_copy(index);
639 domain = isl_multi_pw_aff_range_factor_domain(domain);
640 domain_depth = extract_depth(domain);
641 isl_multi_pw_aff_free(domain);
642 range = isl_multi_pw_aff_copy(index);
643 range = isl_multi_pw_aff_range_factor_range(range);
644 range_depth = extract_depth(range);
645 isl_multi_pw_aff_free(range);
647 return domain_depth + range_depth;
650 if (!isl_multi_pw_aff_has_tuple_id(index, isl_dim_out))
651 return 1;
653 id = isl_multi_pw_aff_get_tuple_id(index, isl_dim_out);
654 if (!id)
655 return -1;
656 decl = (ValueDecl *) isl_id_get_user(id);
657 isl_id_free(id);
659 return array_depth(decl->getType().getTypePtr());
662 /* Return the depth of the array accessed by the access expression "expr".
664 static int extract_depth(__isl_keep pet_expr *expr)
666 isl_multi_pw_aff *index;
667 int depth;
669 index = pet_expr_access_get_index(expr);
670 depth = extract_depth(index);
671 isl_multi_pw_aff_free(index);
673 return depth;
676 /* Construct a pet_expr representing an index expression for an access
677 * to the variable referenced by "expr".
679 __isl_give pet_expr *PetScan::extract_index_expr(DeclRefExpr *expr)
681 return extract_index_expr(expr->getDecl());
684 /* Construct a pet_expr representing an index expression for an access
685 * to the variable "decl".
687 __isl_give pet_expr *PetScan::extract_index_expr(ValueDecl *decl)
689 isl_id *id = create_decl_id(ctx, decl);
690 isl_space *space = isl_space_alloc(ctx, 0, 0, 0);
692 space = isl_space_set_tuple_id(space, isl_dim_out, id);
694 return pet_expr_from_index(isl_multi_pw_aff_zero(space));
697 /* Construct a pet_expr representing the index expression "expr"
698 * Return NULL on error.
700 __isl_give pet_expr *PetScan::extract_index_expr(Expr *expr)
702 switch (expr->getStmtClass()) {
703 case Stmt::ImplicitCastExprClass:
704 return extract_index_expr(cast<ImplicitCastExpr>(expr));
705 case Stmt::DeclRefExprClass:
706 return extract_index_expr(cast<DeclRefExpr>(expr));
707 case Stmt::ArraySubscriptExprClass:
708 return extract_index_expr(cast<ArraySubscriptExpr>(expr));
709 case Stmt::IntegerLiteralClass:
710 return extract_expr(cast<IntegerLiteral>(expr));
711 case Stmt::MemberExprClass:
712 return extract_index_expr(cast<MemberExpr>(expr));
713 default:
714 unsupported(expr);
716 return NULL;
719 /* Extract an index expression from the given array subscript expression.
721 * We first extract an index expression from the base.
722 * This will result in an index expression with a range that corresponds
723 * to the earlier indices.
724 * We then extract the current index and let
725 * pet_expr_access_subscript combine the two.
727 __isl_give pet_expr *PetScan::extract_index_expr(ArraySubscriptExpr *expr)
729 Expr *base = expr->getBase();
730 Expr *idx = expr->getIdx();
731 pet_expr *index;
732 pet_expr *base_expr;
734 base_expr = extract_index_expr(base);
735 index = extract_expr(idx);
737 base_expr = pet_expr_access_subscript(base_expr, index);
739 return base_expr;
742 /* Extract an index expression from a member expression.
744 * If the base access (to the structure containing the member)
745 * is of the form
747 * A[..]
749 * and the member is called "f", then the member access is of
750 * the form
752 * A_f[A[..] -> f[]]
754 * If the member access is to an anonymous struct, then simply return
756 * A[..]
758 * If the member access in the source code is of the form
760 * A->f
762 * then it is treated as
764 * A[0].f
766 __isl_give pet_expr *PetScan::extract_index_expr(MemberExpr *expr)
768 Expr *base = expr->getBase();
769 FieldDecl *field = cast<FieldDecl>(expr->getMemberDecl());
770 pet_expr *base_index;
771 isl_id *id;
773 base_index = extract_index_expr(base);
775 if (expr->isArrow()) {
776 pet_expr *index = pet_expr_new_int(isl_val_zero(ctx));
777 base_index = pet_expr_access_subscript(base_index, index);
780 if (field->isAnonymousStructOrUnion())
781 return base_index;
783 id = create_decl_id(ctx, field);
785 return pet_expr_access_member(base_index, id);
788 /* Check if "expr" calls function "minmax" with two arguments and if so
789 * make lhs and rhs refer to these two arguments.
791 static bool is_minmax(Expr *expr, const char *minmax, Expr *&lhs, Expr *&rhs)
793 CallExpr *call;
794 FunctionDecl *fd;
795 string name;
797 if (expr->getStmtClass() != Stmt::CallExprClass)
798 return false;
800 call = cast<CallExpr>(expr);
801 fd = call->getDirectCallee();
802 if (!fd)
803 return false;
805 if (call->getNumArgs() != 2)
806 return false;
808 name = fd->getDeclName().getAsString();
809 if (name != minmax)
810 return false;
812 lhs = call->getArg(0);
813 rhs = call->getArg(1);
815 return true;
818 /* Check if "expr" is of the form min(lhs, rhs) and if so make
819 * lhs and rhs refer to the two arguments.
821 static bool is_min(Expr *expr, Expr *&lhs, Expr *&rhs)
823 return is_minmax(expr, "min", lhs, rhs);
826 /* Check if "expr" is of the form max(lhs, rhs) and if so make
827 * lhs and rhs refer to the two arguments.
829 static bool is_max(Expr *expr, Expr *&lhs, Expr *&rhs)
831 return is_minmax(expr, "max", lhs, rhs);
834 /* Extract an affine expressions representing the comparison "LHS op RHS"
835 * "comp" is the original statement that "LHS op RHS" is derived from
836 * and is used for diagnostics.
838 * If the comparison is of the form
840 * a <= min(b,c)
842 * then the expression is constructed as the conjunction of
843 * the comparisons
845 * a <= b and a <= c
847 * A similar optimization is performed for max(a,b) <= c.
848 * We do this because that will lead to simpler representations
849 * of the expression.
850 * If isl is ever enhanced to explicitly deal with min and max expressions,
851 * this optimization can be removed.
853 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperatorKind op,
854 Expr *LHS, Expr *RHS, Stmt *comp)
856 isl_pw_aff *lhs;
857 isl_pw_aff *rhs;
858 isl_pw_aff *res;
859 isl_set *cond;
860 isl_set *dom;
861 enum pet_op_type type;
863 if (op == BO_GT)
864 return extract_comparison(BO_LT, RHS, LHS, comp);
865 if (op == BO_GE)
866 return extract_comparison(BO_LE, RHS, LHS, comp);
868 if (op == BO_LT || op == BO_LE) {
869 Expr *expr1, *expr2;
870 if (is_min(RHS, expr1, expr2)) {
871 lhs = extract_comparison(op, LHS, expr1, comp);
872 rhs = extract_comparison(op, LHS, expr2, comp);
873 return pet_and(lhs, rhs);
875 if (is_max(LHS, expr1, expr2)) {
876 lhs = extract_comparison(op, expr1, RHS, comp);
877 rhs = extract_comparison(op, expr2, RHS, comp);
878 return pet_and(lhs, rhs);
882 lhs = extract_affine(LHS);
883 rhs = extract_affine(RHS);
885 type = BinaryOperatorKind2pet_op_type(op);
886 return pet_comparison(type, lhs, rhs);
889 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperator *comp)
891 return extract_comparison(comp->getOpcode(), comp->getLHS(),
892 comp->getRHS(), comp);
895 /* Extract an affine expression from a boolean expression.
896 * In particular, return the expression "expr ? 1 : 0".
897 * Return NULL if we are unable to extract an affine expression.
899 * We first convert the clang::Expr to a pet_expr and
900 * then extract an affine expression from that pet_expr.
902 __isl_give isl_pw_aff *PetScan::extract_condition(Expr *expr)
904 isl_pw_aff *cond;
905 pet_expr *pe;
906 pet_context *pc;
908 if (!expr) {
909 isl_set *u = isl_set_universe(isl_space_set_alloc(ctx, 0, 0));
910 return indicator_function(u, isl_set_copy(u));
913 pe = extract_expr(expr);
914 pc = convert_assignments(ctx, assigned_value);
915 pe = pet_expr_plug_in_args(pe, pc);
916 pc = pet_context_set_allow_nested(pc, nesting_enabled);
917 cond = pet_expr_extract_affine_condition(pe, pc);
918 if (isl_pw_aff_involves_nan(cond))
919 cond = isl_pw_aff_free(cond);
920 pet_context_free(pc);
921 pet_expr_free(pe);
922 return cond;
925 /* Mark the given access pet_expr as a write.
927 static __isl_give pet_expr *mark_write(__isl_take pet_expr *access)
929 access = pet_expr_access_set_write(access, 1);
930 access = pet_expr_access_set_read(access, 0);
932 return access;
935 /* Construct a pet_expr representing a unary operator expression.
937 __isl_give pet_expr *PetScan::extract_expr(UnaryOperator *expr)
939 pet_expr *arg;
940 enum pet_op_type op;
942 op = UnaryOperatorKind2pet_op_type(expr->getOpcode());
943 if (op == pet_op_last) {
944 unsupported(expr);
945 return NULL;
948 arg = extract_expr(expr->getSubExpr());
950 if (expr->isIncrementDecrementOp() &&
951 pet_expr_get_type(arg) == pet_expr_access) {
952 arg = mark_write(arg);
953 arg = pet_expr_access_set_read(arg, 1);
956 return pet_expr_new_unary(op, arg);
959 /* If the access expression "expr" writes to a (non-virtual) scalar,
960 * then mark the scalar as having an unknown value in "assigned_value".
962 static int clear_write(__isl_keep pet_expr *expr, void *user)
964 isl_id *id;
965 ValueDecl *decl;
966 PetScan *ps = (PetScan *) user;
968 if (!pet_expr_access_is_write(expr))
969 return 0;
970 if (!pet_expr_is_scalar_access(expr))
971 return 0;
973 id = pet_expr_access_get_id(expr);
974 decl = (ValueDecl *) isl_id_get_user(id);
975 isl_id_free(id);
977 if (decl)
978 clear_assignment(ps->assigned_value, decl);
980 return 0;
983 /* Take into account the writes in "stmt".
984 * That is, first mark all scalar variables that are written by "stmt"
985 * as having an unknown value. Afterwards,
986 * if "stmt" is a top-level (i.e., unconditional) assignment
987 * to a scalar variable, then update "assigned_value" accordingly.
989 * In particular, if the lhs of the assignment is a scalar variable, then mark
990 * the variable as having been assigned. If, furthermore, the rhs
991 * is an affine expression, then keep track of this value in assigned_value
992 * so that we can plug it in when we later come across the same variable.
994 * We skip assignments to virtual arrays (those with NULL user pointer).
996 void PetScan::handle_writes(struct pet_stmt *stmt)
998 pet_expr *body = stmt->body;
999 pet_expr *arg;
1000 isl_id *id;
1001 ValueDecl *decl;
1002 pet_context *pc;
1003 isl_pw_aff *pa;
1005 pet_expr_foreach_access_expr(body, &clear_write, this);
1007 if (!pet_stmt_is_assign(stmt))
1008 return;
1009 if (!isl_set_plain_is_universe(stmt->domain))
1010 return;
1011 arg = pet_expr_get_arg(body, 0);
1012 if (!pet_expr_is_scalar_access(arg)) {
1013 pet_expr_free(arg);
1014 return;
1017 id = pet_expr_access_get_id(arg);
1018 decl = (ValueDecl *) isl_id_get_user(id);
1019 isl_id_free(id);
1020 pet_expr_free(arg);
1022 if (!decl)
1023 return;
1025 arg = pet_expr_get_arg(body, 1);
1026 pc = convert_assignments(ctx, assigned_value);
1027 pa = pet_expr_extract_affine(arg, pc);
1028 pet_context_free(pc);
1029 clear_assignment(assigned_value, decl);
1030 pet_expr_free(arg);
1032 if (isl_pw_aff_involves_nan(pa))
1033 pa = isl_pw_aff_free(pa);
1034 if (!pa)
1035 return;
1036 assigned_value[decl] = pa;
1037 insert_expression(pa);
1040 /* Update "assigned_value" based on the write accesses (and, in particular,
1041 * assignments) in "scop".
1043 void PetScan::handle_writes(struct pet_scop *scop)
1045 if (!scop)
1046 return;
1047 for (int i = 0; i < scop->n_stmt; ++i)
1048 handle_writes(scop->stmts[i]);
1051 /* Construct a pet_expr representing a binary operator expression.
1053 * If the top level operator is an assignment and the LHS is an access,
1054 * then we mark that access as a write. If the operator is a compound
1055 * assignment, the access is marked as both a read and a write.
1057 __isl_give pet_expr *PetScan::extract_expr(BinaryOperator *expr)
1059 int type_size;
1060 pet_expr *lhs, *rhs;
1061 enum pet_op_type op;
1063 op = BinaryOperatorKind2pet_op_type(expr->getOpcode());
1064 if (op == pet_op_last) {
1065 unsupported(expr);
1066 return NULL;
1069 lhs = extract_expr(expr->getLHS());
1070 rhs = extract_expr(expr->getRHS());
1072 if (expr->isAssignmentOp() &&
1073 pet_expr_get_type(lhs) == pet_expr_access) {
1074 lhs = mark_write(lhs);
1075 if (expr->isCompoundAssignmentOp())
1076 lhs = pet_expr_access_set_read(lhs, 1);
1079 type_size = get_type_size(expr->getType(), ast_context);
1080 return pet_expr_new_binary(type_size, op, lhs, rhs);
1083 /* Construct a pet_scop with a single statement killing the entire
1084 * array "array".
1086 struct pet_scop *PetScan::kill(Stmt *stmt, struct pet_array *array)
1088 isl_id *id;
1089 isl_space *space;
1090 isl_multi_pw_aff *index;
1091 isl_map *access;
1092 pet_expr *expr;
1094 if (!array)
1095 return NULL;
1096 access = isl_map_from_range(isl_set_copy(array->extent));
1097 id = isl_set_get_tuple_id(array->extent);
1098 space = isl_space_alloc(ctx, 0, 0, 0);
1099 space = isl_space_set_tuple_id(space, isl_dim_out, id);
1100 index = isl_multi_pw_aff_zero(space);
1101 expr = pet_expr_kill_from_access_and_index(access, index);
1102 return extract(expr, stmt->getSourceRange(), false);
1105 /* Construct a pet_scop for a (single) variable declaration.
1107 * The scop contains the variable being declared (as an array)
1108 * and a statement killing the array.
1110 * If the variable is initialized in the AST, then the scop
1111 * also contains an assignment to the variable.
1113 struct pet_scop *PetScan::extract(DeclStmt *stmt)
1115 int type_size;
1116 Decl *decl;
1117 VarDecl *vd;
1118 pet_expr *lhs, *rhs, *pe;
1119 struct pet_scop *scop_decl, *scop;
1120 struct pet_array *array;
1122 if (!stmt->isSingleDecl()) {
1123 unsupported(stmt);
1124 return NULL;
1127 decl = stmt->getSingleDecl();
1128 vd = cast<VarDecl>(decl);
1130 array = extract_array(ctx, vd, NULL);
1131 if (array)
1132 array->declared = 1;
1133 scop_decl = kill(stmt, array);
1134 scop_decl = pet_scop_add_array(scop_decl, array);
1136 if (!vd->getInit())
1137 return scop_decl;
1139 lhs = extract_access_expr(vd);
1140 rhs = extract_expr(vd->getInit());
1142 lhs = mark_write(lhs);
1144 type_size = get_type_size(vd->getType(), ast_context);
1145 pe = pet_expr_new_binary(type_size, pet_op_assign, lhs, rhs);
1146 scop = extract(pe, stmt->getSourceRange(), false);
1148 scop_decl = pet_scop_prefix(scop_decl, 0);
1149 scop = pet_scop_prefix(scop, 1);
1151 scop = pet_scop_add_seq(ctx, scop_decl, scop);
1153 return scop;
1156 /* Construct a pet_expr representing a conditional operation.
1158 __isl_give pet_expr *PetScan::extract_expr(ConditionalOperator *expr)
1160 pet_expr *cond, *lhs, *rhs;
1161 isl_pw_aff *pa;
1163 cond = extract_expr(expr->getCond());
1164 lhs = extract_expr(expr->getTrueExpr());
1165 rhs = extract_expr(expr->getFalseExpr());
1167 return pet_expr_new_ternary(cond, lhs, rhs);
1170 __isl_give pet_expr *PetScan::extract_expr(ImplicitCastExpr *expr)
1172 return extract_expr(expr->getSubExpr());
1175 /* Construct a pet_expr representing a floating point value.
1177 * If the floating point literal does not appear in a macro,
1178 * then we use the original representation in the source code
1179 * as the string representation. Otherwise, we use the pretty
1180 * printer to produce a string representation.
1182 __isl_give pet_expr *PetScan::extract_expr(FloatingLiteral *expr)
1184 double d;
1185 string s;
1186 const LangOptions &LO = PP.getLangOpts();
1187 SourceLocation loc = expr->getLocation();
1189 if (!loc.isMacroID()) {
1190 SourceManager &SM = PP.getSourceManager();
1191 unsigned len = Lexer::MeasureTokenLength(loc, SM, LO);
1192 s = string(SM.getCharacterData(loc), len);
1193 } else {
1194 llvm::raw_string_ostream S(s);
1195 expr->printPretty(S, 0, PrintingPolicy(LO));
1196 S.str();
1198 d = expr->getValueAsApproximateDouble();
1199 return pet_expr_new_double(ctx, d, s.c_str());
1202 /* Convert the index expression "index" into an access pet_expr of type "qt".
1204 __isl_give pet_expr *PetScan::extract_access_expr(QualType qt,
1205 __isl_take pet_expr *index)
1207 int depth;
1208 int type_size;
1210 depth = extract_depth(index);
1211 type_size = get_type_size(qt, ast_context);
1213 index = pet_expr_set_type_size(index, type_size);
1214 index = pet_expr_access_set_depth(index, depth);
1216 return index;
1219 /* Extract an index expression from "expr" and then convert it into
1220 * an access pet_expr.
1222 __isl_give pet_expr *PetScan::extract_access_expr(Expr *expr)
1224 return extract_access_expr(expr->getType(), extract_index_expr(expr));
1227 /* Extract an index expression from "decl" and then convert it into
1228 * an access pet_expr.
1230 __isl_give pet_expr *PetScan::extract_access_expr(ValueDecl *decl)
1232 return extract_access_expr(decl->getType(), extract_index_expr(decl));
1235 __isl_give pet_expr *PetScan::extract_expr(ParenExpr *expr)
1237 return extract_expr(expr->getSubExpr());
1240 /* Extract an assume statement from the argument "expr"
1241 * of a __pencil_assume statement.
1243 __isl_give pet_expr *PetScan::extract_assume(Expr *expr)
1245 return pet_expr_new_unary(pet_op_assume, extract_expr(expr));
1248 /* Construct a pet_expr corresponding to the function call argument "expr".
1249 * The argument appears in position "pos" of a call to function "fd".
1251 * If we are passing along a pointer to an array element
1252 * or an entire row or even higher dimensional slice of an array,
1253 * then the function being called may write into the array.
1255 * We assume here that if the function is declared to take a pointer
1256 * to a const type, then the function will perform a read
1257 * and that otherwise, it will perform a write.
1259 __isl_give pet_expr *PetScan::extract_argument(FunctionDecl *fd, int pos,
1260 Expr *expr)
1262 pet_expr *res;
1263 int is_addr = 0, is_partial = 0;
1264 Stmt::StmtClass sc;
1266 if (expr->getStmtClass() == Stmt::ImplicitCastExprClass) {
1267 ImplicitCastExpr *ice = cast<ImplicitCastExpr>(expr);
1268 expr = ice->getSubExpr();
1270 if (expr->getStmtClass() == Stmt::UnaryOperatorClass) {
1271 UnaryOperator *op = cast<UnaryOperator>(expr);
1272 if (op->getOpcode() == UO_AddrOf) {
1273 is_addr = 1;
1274 expr = op->getSubExpr();
1277 res = extract_expr(expr);
1278 if (!res)
1279 return NULL;
1280 sc = expr->getStmtClass();
1281 if ((sc == Stmt::ArraySubscriptExprClass ||
1282 sc == Stmt::MemberExprClass) &&
1283 array_depth(expr->getType().getTypePtr()) > 0)
1284 is_partial = 1;
1285 if ((is_addr || is_partial) &&
1286 pet_expr_get_type(res) == pet_expr_access) {
1287 ParmVarDecl *parm;
1288 if (!fd->hasPrototype()) {
1289 report_prototype_required(expr);
1290 return pet_expr_free(res);
1292 parm = fd->getParamDecl(pos);
1293 if (!const_base(parm->getType()))
1294 res = mark_write(res);
1297 if (is_addr)
1298 res = pet_expr_new_unary(pet_op_address_of, res);
1299 return res;
1302 /* Construct a pet_expr representing a function call.
1304 * In the special case of a "call" to __pencil_assume,
1305 * construct an assume expression instead.
1307 __isl_give pet_expr *PetScan::extract_expr(CallExpr *expr)
1309 pet_expr *res = NULL;
1310 FunctionDecl *fd;
1311 string name;
1312 unsigned n_arg;
1314 fd = expr->getDirectCallee();
1315 if (!fd) {
1316 unsupported(expr);
1317 return NULL;
1320 name = fd->getDeclName().getAsString();
1321 n_arg = expr->getNumArgs();
1323 if (n_arg == 1 && name == "__pencil_assume")
1324 return extract_assume(expr->getArg(0));
1326 res = pet_expr_new_call(ctx, name.c_str(), n_arg);
1327 if (!res)
1328 return NULL;
1330 for (int i = 0; i < n_arg; ++i) {
1331 Expr *arg = expr->getArg(i);
1332 res = pet_expr_set_arg(res, i,
1333 PetScan::extract_argument(fd, i, arg));
1336 return res;
1339 /* Construct a pet_expr representing a (C style) cast.
1341 __isl_give pet_expr *PetScan::extract_expr(CStyleCastExpr *expr)
1343 pet_expr *arg;
1344 QualType type;
1346 arg = extract_expr(expr->getSubExpr());
1347 if (!arg)
1348 return NULL;
1350 type = expr->getTypeAsWritten();
1351 return pet_expr_new_cast(type.getAsString().c_str(), arg);
1354 /* Construct a pet_expr representing an integer.
1356 __isl_give pet_expr *PetScan::extract_expr(IntegerLiteral *expr)
1358 return pet_expr_new_int(extract_int(expr));
1361 /* Try and construct a pet_expr representing "expr".
1363 __isl_give pet_expr *PetScan::extract_expr(Expr *expr)
1365 switch (expr->getStmtClass()) {
1366 case Stmt::UnaryOperatorClass:
1367 return extract_expr(cast<UnaryOperator>(expr));
1368 case Stmt::CompoundAssignOperatorClass:
1369 case Stmt::BinaryOperatorClass:
1370 return extract_expr(cast<BinaryOperator>(expr));
1371 case Stmt::ImplicitCastExprClass:
1372 return extract_expr(cast<ImplicitCastExpr>(expr));
1373 case Stmt::ArraySubscriptExprClass:
1374 case Stmt::DeclRefExprClass:
1375 case Stmt::MemberExprClass:
1376 return extract_access_expr(expr);
1377 case Stmt::IntegerLiteralClass:
1378 return extract_expr(cast<IntegerLiteral>(expr));
1379 case Stmt::FloatingLiteralClass:
1380 return extract_expr(cast<FloatingLiteral>(expr));
1381 case Stmt::ParenExprClass:
1382 return extract_expr(cast<ParenExpr>(expr));
1383 case Stmt::ConditionalOperatorClass:
1384 return extract_expr(cast<ConditionalOperator>(expr));
1385 case Stmt::CallExprClass:
1386 return extract_expr(cast<CallExpr>(expr));
1387 case Stmt::CStyleCastExprClass:
1388 return extract_expr(cast<CStyleCastExpr>(expr));
1389 default:
1390 unsupported(expr);
1392 return NULL;
1395 /* Check if the given initialization statement is an assignment.
1396 * If so, return that assignment. Otherwise return NULL.
1398 BinaryOperator *PetScan::initialization_assignment(Stmt *init)
1400 BinaryOperator *ass;
1402 if (init->getStmtClass() != Stmt::BinaryOperatorClass)
1403 return NULL;
1405 ass = cast<BinaryOperator>(init);
1406 if (ass->getOpcode() != BO_Assign)
1407 return NULL;
1409 return ass;
1412 /* Check if the given initialization statement is a declaration
1413 * of a single variable.
1414 * If so, return that declaration. Otherwise return NULL.
1416 Decl *PetScan::initialization_declaration(Stmt *init)
1418 DeclStmt *decl;
1420 if (init->getStmtClass() != Stmt::DeclStmtClass)
1421 return NULL;
1423 decl = cast<DeclStmt>(init);
1425 if (!decl->isSingleDecl())
1426 return NULL;
1428 return decl->getSingleDecl();
1431 /* Given the assignment operator in the initialization of a for loop,
1432 * extract the induction variable, i.e., the (integer)variable being
1433 * assigned.
1435 ValueDecl *PetScan::extract_induction_variable(BinaryOperator *init)
1437 Expr *lhs;
1438 DeclRefExpr *ref;
1439 ValueDecl *decl;
1440 const Type *type;
1442 lhs = init->getLHS();
1443 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1444 unsupported(init);
1445 return NULL;
1448 ref = cast<DeclRefExpr>(lhs);
1449 decl = ref->getDecl();
1450 type = decl->getType().getTypePtr();
1452 if (!type->isIntegerType()) {
1453 unsupported(lhs);
1454 return NULL;
1457 return decl;
1460 /* Given the initialization statement of a for loop and the single
1461 * declaration in this initialization statement,
1462 * extract the induction variable, i.e., the (integer) variable being
1463 * declared.
1465 VarDecl *PetScan::extract_induction_variable(Stmt *init, Decl *decl)
1467 VarDecl *vd;
1469 vd = cast<VarDecl>(decl);
1471 const QualType type = vd->getType();
1472 if (!type->isIntegerType()) {
1473 unsupported(init);
1474 return NULL;
1477 if (!vd->getInit()) {
1478 unsupported(init);
1479 return NULL;
1482 return vd;
1485 /* Check that op is of the form iv++ or iv--.
1486 * Return a pet_expr representing "1" or "-1" accordingly.
1488 __isl_give pet_expr *PetScan::extract_unary_increment(
1489 clang::UnaryOperator *op, clang::ValueDecl *iv)
1491 Expr *sub;
1492 DeclRefExpr *ref;
1493 isl_val *v;
1495 if (!op->isIncrementDecrementOp()) {
1496 unsupported(op);
1497 return NULL;
1500 sub = op->getSubExpr();
1501 if (sub->getStmtClass() != Stmt::DeclRefExprClass) {
1502 unsupported(op);
1503 return NULL;
1506 ref = cast<DeclRefExpr>(sub);
1507 if (ref->getDecl() != iv) {
1508 unsupported(op);
1509 return NULL;
1512 if (op->isIncrementOp())
1513 v = isl_val_one(ctx);
1514 else
1515 v = isl_val_negone(ctx);
1517 return pet_expr_new_int(v);
1520 /* Check if op is of the form
1522 * iv = expr
1524 * and return the increment "expr - iv" as a pet_expr.
1526 __isl_give pet_expr *PetScan::extract_binary_increment(BinaryOperator *op,
1527 clang::ValueDecl *iv)
1529 int type_size;
1530 Expr *lhs;
1531 DeclRefExpr *ref;
1532 pet_expr *expr, *expr_iv;
1534 if (op->getOpcode() != BO_Assign) {
1535 unsupported(op);
1536 return NULL;
1539 lhs = op->getLHS();
1540 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1541 unsupported(op);
1542 return NULL;
1545 ref = cast<DeclRefExpr>(lhs);
1546 if (ref->getDecl() != iv) {
1547 unsupported(op);
1548 return NULL;
1551 expr = extract_expr(op->getRHS());
1552 expr_iv = extract_expr(lhs);
1554 type_size = get_type_size(iv->getType(), ast_context);
1555 return pet_expr_new_binary(type_size, pet_op_sub, expr, expr_iv);
1558 /* Check that op is of the form iv += cst or iv -= cst
1559 * and return a pet_expr corresponding to cst or -cst accordingly.
1561 __isl_give pet_expr *PetScan::extract_compound_increment(
1562 CompoundAssignOperator *op, clang::ValueDecl *iv)
1564 Expr *lhs;
1565 DeclRefExpr *ref;
1566 bool neg = false;
1567 pet_expr *expr;
1568 BinaryOperatorKind opcode;
1570 opcode = op->getOpcode();
1571 if (opcode != BO_AddAssign && opcode != BO_SubAssign) {
1572 unsupported(op);
1573 return NULL;
1575 if (opcode == BO_SubAssign)
1576 neg = true;
1578 lhs = op->getLHS();
1579 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1580 unsupported(op);
1581 return NULL;
1584 ref = cast<DeclRefExpr>(lhs);
1585 if (ref->getDecl() != iv) {
1586 unsupported(op);
1587 return NULL;
1590 expr = extract_expr(op->getRHS());
1591 if (neg)
1592 expr = pet_expr_new_unary(pet_op_minus, expr);
1594 return expr;
1597 /* Check that the increment of the given for loop increments
1598 * (or decrements) the induction variable "iv" and return
1599 * the increment as a pet_expr if successful.
1601 __isl_give pet_expr *PetScan::extract_increment(clang::ForStmt *stmt,
1602 ValueDecl *iv)
1604 Stmt *inc = stmt->getInc();
1606 if (!inc) {
1607 report_missing_increment(stmt);
1608 return NULL;
1611 if (inc->getStmtClass() == Stmt::UnaryOperatorClass)
1612 return extract_unary_increment(cast<UnaryOperator>(inc), iv);
1613 if (inc->getStmtClass() == Stmt::CompoundAssignOperatorClass)
1614 return extract_compound_increment(
1615 cast<CompoundAssignOperator>(inc), iv);
1616 if (inc->getStmtClass() == Stmt::BinaryOperatorClass)
1617 return extract_binary_increment(cast<BinaryOperator>(inc), iv);
1619 unsupported(inc);
1620 return NULL;
1623 /* Embed the given iteration domain in an extra outer loop
1624 * with induction variable "var".
1625 * If this variable appeared as a parameter in the constraints,
1626 * it is replaced by the new outermost dimension.
1628 static __isl_give isl_set *embed(__isl_take isl_set *set,
1629 __isl_take isl_id *var)
1631 int pos;
1633 set = isl_set_insert_dims(set, isl_dim_set, 0, 1);
1634 pos = isl_set_find_dim_by_id(set, isl_dim_param, var);
1635 if (pos >= 0) {
1636 set = isl_set_equate(set, isl_dim_param, pos, isl_dim_set, 0);
1637 set = isl_set_project_out(set, isl_dim_param, pos, 1);
1640 isl_id_free(var);
1641 return set;
1644 /* Return those elements in the space of "cond" that come after
1645 * (based on "sign") an element in "cond".
1647 static __isl_give isl_set *after(__isl_take isl_set *cond, int sign)
1649 isl_map *previous_to_this;
1651 if (sign > 0)
1652 previous_to_this = isl_map_lex_lt(isl_set_get_space(cond));
1653 else
1654 previous_to_this = isl_map_lex_gt(isl_set_get_space(cond));
1656 cond = isl_set_apply(cond, previous_to_this);
1658 return cond;
1661 /* Create the infinite iteration domain
1663 * { [id] : id >= 0 }
1665 * If "scop" has an affine skip of type pet_skip_later,
1666 * then remove those iterations i that have an earlier iteration
1667 * where the skip condition is satisfied, meaning that iteration i
1668 * is not executed.
1669 * Since we are dealing with a loop without loop iterator,
1670 * the skip condition cannot refer to the current loop iterator and
1671 * so effectively, the returned set is of the form
1673 * { [0]; [id] : id >= 1 and not skip }
1675 static __isl_give isl_set *infinite_domain(__isl_take isl_id *id,
1676 struct pet_scop *scop)
1678 isl_ctx *ctx = isl_id_get_ctx(id);
1679 isl_set *domain;
1680 isl_set *skip;
1682 domain = isl_set_nat_universe(isl_space_set_alloc(ctx, 0, 1));
1683 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, id);
1685 if (!pet_scop_has_affine_skip(scop, pet_skip_later))
1686 return domain;
1688 skip = pet_scop_get_affine_skip_domain(scop, pet_skip_later);
1689 skip = embed(skip, isl_id_copy(id));
1690 skip = isl_set_intersect(skip , isl_set_copy(domain));
1691 domain = isl_set_subtract(domain, after(skip, 1));
1693 return domain;
1696 /* Create an identity affine expression on the space containing "domain",
1697 * which is assumed to be one-dimensional.
1699 static __isl_give isl_aff *identity_aff(__isl_keep isl_set *domain)
1701 isl_local_space *ls;
1703 ls = isl_local_space_from_space(isl_set_get_space(domain));
1704 return isl_aff_var_on_domain(ls, isl_dim_set, 0);
1707 /* Create an affine expression that maps elements
1708 * of a single-dimensional array "id_test" to the previous element
1709 * (according to "inc"), provided this element belongs to "domain".
1710 * That is, create the affine expression
1712 * { id[x] -> id[x - inc] : x - inc in domain }
1714 static __isl_give isl_multi_pw_aff *map_to_previous(__isl_take isl_id *id_test,
1715 __isl_take isl_set *domain, __isl_take isl_val *inc)
1717 isl_space *space;
1718 isl_local_space *ls;
1719 isl_aff *aff;
1720 isl_multi_pw_aff *prev;
1722 space = isl_set_get_space(domain);
1723 ls = isl_local_space_from_space(space);
1724 aff = isl_aff_var_on_domain(ls, isl_dim_set, 0);
1725 aff = isl_aff_add_constant_val(aff, isl_val_neg(inc));
1726 prev = isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
1727 domain = isl_set_preimage_multi_pw_aff(domain,
1728 isl_multi_pw_aff_copy(prev));
1729 prev = isl_multi_pw_aff_intersect_domain(prev, domain);
1730 prev = isl_multi_pw_aff_set_tuple_id(prev, isl_dim_out, id_test);
1732 return prev;
1735 /* Add an implication to "scop" expressing that if an element of
1736 * virtual array "id_test" has value "satisfied" then all previous elements
1737 * of this array also have that value. The set of previous elements
1738 * is bounded by "domain". If "sign" is negative then the iterator
1739 * is decreasing and we express that all subsequent array elements
1740 * (but still defined previously) have the same value.
1742 static struct pet_scop *add_implication(struct pet_scop *scop,
1743 __isl_take isl_id *id_test, __isl_take isl_set *domain, int sign,
1744 int satisfied)
1746 isl_space *space;
1747 isl_map *map;
1749 domain = isl_set_set_tuple_id(domain, id_test);
1750 space = isl_set_get_space(domain);
1751 if (sign > 0)
1752 map = isl_map_lex_ge(space);
1753 else
1754 map = isl_map_lex_le(space);
1755 map = isl_map_intersect_range(map, domain);
1756 scop = pet_scop_add_implication(scop, map, satisfied);
1758 return scop;
1761 /* Add a filter to "scop" that imposes that it is only executed
1762 * when the variable identified by "id_test" has a zero value
1763 * for all previous iterations of "domain".
1765 * In particular, add a filter that imposes that the array
1766 * has a zero value at the previous iteration of domain and
1767 * add an implication that implies that it then has that
1768 * value for all previous iterations.
1770 static struct pet_scop *scop_add_break(struct pet_scop *scop,
1771 __isl_take isl_id *id_test, __isl_take isl_set *domain,
1772 __isl_take isl_val *inc)
1774 isl_multi_pw_aff *prev;
1775 int sign = isl_val_sgn(inc);
1777 prev = map_to_previous(isl_id_copy(id_test), isl_set_copy(domain), inc);
1778 scop = add_implication(scop, id_test, domain, sign, 0);
1779 scop = pet_scop_filter(scop, prev, 0);
1781 return scop;
1784 /* Construct a pet_scop for an infinite loop around the given body.
1786 * We extract a pet_scop for the body and then embed it in a loop with
1787 * iteration domain
1789 * { [t] : t >= 0 }
1791 * and schedule
1793 * { [t] -> [t] }
1795 * If the body contains any break, then it is taken into
1796 * account in infinite_domain (if the skip condition is affine)
1797 * or in scop_add_break (if the skip condition is not affine).
1799 * If we were only able to extract part of the body, then simply
1800 * return that part.
1802 struct pet_scop *PetScan::extract_infinite_loop(Stmt *body)
1804 isl_id *id, *id_test;
1805 isl_set *domain;
1806 isl_aff *ident;
1807 struct pet_scop *scop;
1808 bool has_var_break;
1810 scop = extract(body);
1811 if (!scop)
1812 return NULL;
1813 if (partial)
1814 return scop;
1816 id = isl_id_alloc(ctx, "t", NULL);
1817 domain = infinite_domain(isl_id_copy(id), scop);
1818 ident = identity_aff(domain);
1820 has_var_break = pet_scop_has_var_skip(scop, pet_skip_later);
1821 if (has_var_break)
1822 id_test = pet_scop_get_skip_id(scop, pet_skip_later);
1824 scop = pet_scop_embed(scop, isl_set_copy(domain),
1825 isl_aff_copy(ident), ident, id);
1826 if (has_var_break)
1827 scop = scop_add_break(scop, id_test, domain, isl_val_one(ctx));
1828 else
1829 isl_set_free(domain);
1831 return scop;
1834 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
1836 * for (;;)
1837 * body
1840 struct pet_scop *PetScan::extract_infinite_for(ForStmt *stmt)
1842 clear_assignments clear(assigned_value);
1843 clear.TraverseStmt(stmt->getBody());
1845 return extract_infinite_loop(stmt->getBody());
1848 /* Add an array with the given extent (range of "index") to the list
1849 * of arrays in "scop" and return the extended pet_scop.
1850 * The array is marked as attaining values 0 and 1 only and
1851 * as each element being assigned at most once.
1853 static struct pet_scop *scop_add_array(struct pet_scop *scop,
1854 __isl_keep isl_multi_pw_aff *index, clang::ASTContext &ast_ctx)
1856 int int_size = ast_ctx.getTypeInfo(ast_ctx.IntTy).first / 8;
1858 return pet_scop_add_boolean_array(scop, isl_multi_pw_aff_copy(index),
1859 int_size);
1862 /* Construct a pet_scop for a while loop of the form
1864 * while (pa)
1865 * body
1867 * In particular, construct a scop for an infinite loop around body and
1868 * intersect the domain with the affine expression.
1869 * Note that this intersection may result in an empty loop.
1871 struct pet_scop *PetScan::extract_affine_while(__isl_take isl_pw_aff *pa,
1872 Stmt *body)
1874 struct pet_scop *scop;
1875 isl_set *dom;
1876 isl_set *valid;
1878 valid = isl_pw_aff_domain(isl_pw_aff_copy(pa));
1879 dom = isl_pw_aff_non_zero_set(pa);
1880 scop = extract_infinite_loop(body);
1881 scop = pet_scop_restrict(scop, isl_set_params(dom));
1882 scop = pet_scop_restrict_context(scop, isl_set_params(valid));
1884 return scop;
1887 /* Construct a scop for a while, given the scops for the condition
1888 * and the body, the filter identifier and the iteration domain of
1889 * the while loop.
1891 * In particular, the scop for the condition is filtered to depend
1892 * on "id_test" evaluating to true for all previous iterations
1893 * of the loop, while the scop for the body is filtered to depend
1894 * on "id_test" evaluating to true for all iterations up to the
1895 * current iteration.
1896 * The actual filter only imposes that this virtual array has
1897 * value one on the previous or the current iteration.
1898 * The fact that this condition also applies to the previous
1899 * iterations is enforced by an implication.
1901 * These filtered scops are then combined into a single scop.
1903 * "sign" is positive if the iterator increases and negative
1904 * if it decreases.
1906 static struct pet_scop *scop_add_while(struct pet_scop *scop_cond,
1907 struct pet_scop *scop_body, __isl_take isl_id *id_test,
1908 __isl_take isl_set *domain, __isl_take isl_val *inc)
1910 isl_ctx *ctx = isl_set_get_ctx(domain);
1911 isl_space *space;
1912 isl_multi_pw_aff *test_index;
1913 isl_multi_pw_aff *prev;
1914 int sign = isl_val_sgn(inc);
1915 struct pet_scop *scop;
1917 prev = map_to_previous(isl_id_copy(id_test), isl_set_copy(domain), inc);
1918 scop_cond = pet_scop_filter(scop_cond, prev, 1);
1920 space = isl_space_map_from_set(isl_set_get_space(domain));
1921 test_index = isl_multi_pw_aff_identity(space);
1922 test_index = isl_multi_pw_aff_set_tuple_id(test_index, isl_dim_out,
1923 isl_id_copy(id_test));
1924 scop_body = pet_scop_filter(scop_body, test_index, 1);
1926 scop = pet_scop_add_seq(ctx, scop_cond, scop_body);
1927 scop = add_implication(scop, id_test, domain, sign, 1);
1929 return scop;
1932 /* Check if the while loop is of the form
1934 * while (affine expression)
1935 * body
1937 * If so, call extract_affine_while to construct a scop.
1939 * Otherwise, extract the body and pass control to extract_while
1940 * to extend the iteration domain with an infinite loop.
1941 * If we were only able to extract part of the body, then simply
1942 * return that part.
1944 struct pet_scop *PetScan::extract(WhileStmt *stmt)
1946 Expr *cond;
1947 int test_nr, stmt_nr;
1948 isl_pw_aff *pa;
1949 struct pet_scop *scop_body;
1951 cond = stmt->getCond();
1952 if (!cond) {
1953 unsupported(stmt);
1954 return NULL;
1957 clear_assignments clear(assigned_value);
1958 clear.TraverseStmt(stmt->getBody());
1960 pa = extract_condition(cond);
1961 if (pa)
1962 return extract_affine_while(pa, stmt->getBody());
1964 if (!allow_nested) {
1965 unsupported(stmt);
1966 return NULL;
1969 test_nr = n_test++;
1970 stmt_nr = n_stmt++;
1971 scop_body = extract(stmt->getBody());
1972 if (partial)
1973 return scop_body;
1975 return extract_while(cond, test_nr, stmt_nr, scop_body, NULL);
1978 /* Construct a generic while scop, with iteration domain
1979 * { [t] : t >= 0 } around "scop_body". The scop consists of two parts,
1980 * one for evaluating the condition "cond" and one for the body.
1981 * "test_nr" is the sequence number of the virtual test variable that contains
1982 * the result of the condition and "stmt_nr" is the sequence number
1983 * of the statement that evaluates the condition.
1984 * If "scop_inc" is not NULL, then it is added at the end of the body,
1985 * after replacing any skip conditions resulting from continue statements
1986 * by the skip conditions resulting from break statements (if any).
1988 * The schedule is adjusted to reflect that the condition is evaluated
1989 * before the body is executed and the body is filtered to depend
1990 * on the result of the condition evaluating to true on all iterations
1991 * up to the current iteration, while the evaluation of the condition itself
1992 * is filtered to depend on the result of the condition evaluating to true
1993 * on all previous iterations.
1994 * The context of the scop representing the body is dropped
1995 * because we don't know how many times the body will be executed,
1996 * if at all.
1998 * If the body contains any break, then it is taken into
1999 * account in infinite_domain (if the skip condition is affine)
2000 * or in scop_add_break (if the skip condition is not affine).
2002 struct pet_scop *PetScan::extract_while(Expr *cond, int test_nr, int stmt_nr,
2003 struct pet_scop *scop_body, struct pet_scop *scop_inc)
2005 isl_id *id, *id_test, *id_break_test;
2006 isl_set *domain;
2007 isl_aff *ident;
2008 isl_multi_pw_aff *test_index;
2009 struct pet_scop *scop;
2010 bool has_var_break;
2012 test_index = pet_create_test_index(ctx, test_nr);
2013 scop = extract_non_affine_condition(cond, stmt_nr,
2014 isl_multi_pw_aff_copy(test_index));
2015 scop = scop_add_array(scop, test_index, ast_context);
2016 id_test = isl_multi_pw_aff_get_tuple_id(test_index, isl_dim_out);
2017 isl_multi_pw_aff_free(test_index);
2019 id = isl_id_alloc(ctx, "t", NULL);
2020 domain = infinite_domain(isl_id_copy(id), scop_body);
2021 ident = identity_aff(domain);
2023 has_var_break = pet_scop_has_var_skip(scop_body, pet_skip_later);
2024 if (has_var_break)
2025 id_break_test = pet_scop_get_skip_id(scop_body, pet_skip_later);
2027 scop = pet_scop_prefix(scop, 0);
2028 scop = pet_scop_embed(scop, isl_set_copy(domain), isl_aff_copy(ident),
2029 isl_aff_copy(ident), isl_id_copy(id));
2030 scop_body = pet_scop_reset_context(scop_body);
2031 scop_body = pet_scop_prefix(scop_body, 1);
2032 if (scop_inc) {
2033 scop_inc = pet_scop_prefix(scop_inc, 2);
2034 if (pet_scop_has_skip(scop_body, pet_skip_later)) {
2035 isl_multi_pw_aff *skip;
2036 skip = pet_scop_get_skip(scop_body, pet_skip_later);
2037 scop_body = pet_scop_set_skip(scop_body,
2038 pet_skip_now, skip);
2039 } else
2040 pet_scop_reset_skip(scop_body, pet_skip_now);
2041 scop_body = pet_scop_add_seq(ctx, scop_body, scop_inc);
2043 scop_body = pet_scop_embed(scop_body, isl_set_copy(domain),
2044 isl_aff_copy(ident), ident, id);
2046 if (has_var_break) {
2047 scop = scop_add_break(scop, isl_id_copy(id_break_test),
2048 isl_set_copy(domain), isl_val_one(ctx));
2049 scop_body = scop_add_break(scop_body, id_break_test,
2050 isl_set_copy(domain), isl_val_one(ctx));
2052 scop = scop_add_while(scop, scop_body, id_test, domain,
2053 isl_val_one(ctx));
2055 return scop;
2058 /* Check whether "cond" expresses a simple loop bound
2059 * on the only set dimension.
2060 * In particular, if "up" is set then "cond" should contain only
2061 * upper bounds on the set dimension.
2062 * Otherwise, it should contain only lower bounds.
2064 static bool is_simple_bound(__isl_keep isl_set *cond, __isl_keep isl_val *inc)
2066 if (isl_val_is_pos(inc))
2067 return !isl_set_dim_has_any_lower_bound(cond, isl_dim_set, 0);
2068 else
2069 return !isl_set_dim_has_any_upper_bound(cond, isl_dim_set, 0);
2072 /* Extend a condition on a given iteration of a loop to one that
2073 * imposes the same condition on all previous iterations.
2074 * "domain" expresses the lower [upper] bound on the iterations
2075 * when inc is positive [negative].
2077 * In particular, we construct the condition (when inc is positive)
2079 * forall i' : (domain(i') and i' <= i) => cond(i')
2081 * which is equivalent to
2083 * not exists i' : domain(i') and i' <= i and not cond(i')
2085 * We construct this set by negating cond, applying a map
2087 * { [i'] -> [i] : domain(i') and i' <= i }
2089 * and then negating the result again.
2091 static __isl_give isl_set *valid_for_each_iteration(__isl_take isl_set *cond,
2092 __isl_take isl_set *domain, __isl_take isl_val *inc)
2094 isl_map *previous_to_this;
2096 if (isl_val_is_pos(inc))
2097 previous_to_this = isl_map_lex_le(isl_set_get_space(domain));
2098 else
2099 previous_to_this = isl_map_lex_ge(isl_set_get_space(domain));
2101 previous_to_this = isl_map_intersect_domain(previous_to_this, domain);
2103 cond = isl_set_complement(cond);
2104 cond = isl_set_apply(cond, previous_to_this);
2105 cond = isl_set_complement(cond);
2107 isl_val_free(inc);
2109 return cond;
2112 /* Construct a domain of the form
2114 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
2116 static __isl_give isl_set *strided_domain(__isl_take isl_id *id,
2117 __isl_take isl_pw_aff *init, __isl_take isl_val *inc)
2119 isl_aff *aff;
2120 isl_space *dim;
2121 isl_set *set;
2123 init = isl_pw_aff_insert_dims(init, isl_dim_in, 0, 1);
2124 dim = isl_pw_aff_get_domain_space(init);
2125 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2126 aff = isl_aff_add_coefficient_val(aff, isl_dim_in, 0, inc);
2127 init = isl_pw_aff_add(init, isl_pw_aff_from_aff(aff));
2129 dim = isl_space_set_alloc(isl_pw_aff_get_ctx(init), 1, 1);
2130 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
2131 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2132 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2134 set = isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff), init);
2136 set = isl_set_lower_bound_si(set, isl_dim_set, 0, 0);
2138 return isl_set_params(set);
2141 /* Assuming "cond" represents a bound on a loop where the loop
2142 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2143 * is possible.
2145 * Under the given assumptions, wrapping is only possible if "cond" allows
2146 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2147 * increasing iterator and 0 in case of a decreasing iterator.
2149 static bool can_wrap(__isl_keep isl_set *cond, ValueDecl *iv,
2150 __isl_keep isl_val *inc)
2152 bool cw;
2153 isl_ctx *ctx;
2154 isl_val *limit;
2155 isl_set *test;
2157 test = isl_set_copy(cond);
2159 ctx = isl_set_get_ctx(test);
2160 if (isl_val_is_neg(inc))
2161 limit = isl_val_zero(ctx);
2162 else {
2163 limit = isl_val_int_from_ui(ctx, get_type_size(iv));
2164 limit = isl_val_2exp(limit);
2165 limit = isl_val_sub_ui(limit, 1);
2168 test = isl_set_fix_val(cond, isl_dim_set, 0, limit);
2169 cw = !isl_set_is_empty(test);
2170 isl_set_free(test);
2172 return cw;
2175 /* Given a one-dimensional space, construct the following affine expression
2176 * on this space
2178 * { [v] -> [v mod 2^width] }
2180 * where width is the number of bits used to represent the values
2181 * of the unsigned variable "iv".
2183 static __isl_give isl_aff *compute_wrapping(__isl_take isl_space *dim,
2184 ValueDecl *iv)
2186 isl_ctx *ctx;
2187 isl_val *mod;
2188 isl_aff *aff;
2190 ctx = isl_space_get_ctx(dim);
2191 mod = isl_val_int_from_ui(ctx, get_type_size(iv));
2192 mod = isl_val_2exp(mod);
2194 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2195 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2196 aff = isl_aff_mod_val(aff, mod);
2198 return aff;
2201 /* Project out the parameter "id" from "set".
2203 static __isl_give isl_set *set_project_out_by_id(__isl_take isl_set *set,
2204 __isl_keep isl_id *id)
2206 int pos;
2208 pos = isl_set_find_dim_by_id(set, isl_dim_param, id);
2209 if (pos >= 0)
2210 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2212 return set;
2215 /* Compute the set of parameters for which "set1" is a subset of "set2".
2217 * set1 is a subset of set2 if
2219 * forall i in set1 : i in set2
2221 * or
2223 * not exists i in set1 and i not in set2
2225 * i.e.,
2227 * not exists i in set1 \ set2
2229 static __isl_give isl_set *enforce_subset(__isl_take isl_set *set1,
2230 __isl_take isl_set *set2)
2232 return isl_set_complement(isl_set_params(isl_set_subtract(set1, set2)));
2235 /* Compute the set of parameter values for which "cond" holds
2236 * on the next iteration for each element of "dom".
2238 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2239 * and then compute the set of parameters for which the result is a subset
2240 * of "cond".
2242 static __isl_give isl_set *valid_on_next(__isl_take isl_set *cond,
2243 __isl_take isl_set *dom, __isl_take isl_val *inc)
2245 isl_space *space;
2246 isl_aff *aff;
2247 isl_map *next;
2249 space = isl_set_get_space(dom);
2250 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2251 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2252 aff = isl_aff_add_constant_val(aff, inc);
2253 next = isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2255 dom = isl_set_apply(dom, next);
2257 return enforce_subset(dom, cond);
2260 /* Extract the for loop "stmt" as a while loop.
2261 * "iv" is the loop iterator. "init" is the initialization.
2262 * "inc" is the increment.
2264 * That is, the for loop has the form
2266 * for (iv = init; cond; iv += inc)
2267 * body;
2269 * and is treated as
2271 * iv = init;
2272 * while (cond) {
2273 * body;
2274 * iv += inc;
2277 * except that the skips resulting from any continue statements
2278 * in body do not apply to the increment, but are replaced by the skips
2279 * resulting from break statements.
2281 * If "iv" is declared in the for loop, then it is killed before
2282 * and after the loop.
2284 struct pet_scop *PetScan::extract_non_affine_for(ForStmt *stmt, ValueDecl *iv,
2285 __isl_take pet_expr *init, __isl_take pet_expr *inc)
2287 int declared;
2288 int test_nr, stmt_nr;
2289 pet_expr *expr_iv;
2290 struct pet_scop *scop_init, *scop_inc, *scop, *scop_body;
2291 int type_size;
2292 struct pet_array *array;
2293 struct pet_scop *scop_kill;
2295 if (!allow_nested) {
2296 unsupported(stmt);
2297 return NULL;
2300 clear_assignment(assigned_value, iv);
2302 declared = !initialization_assignment(stmt->getInit());
2304 expr_iv = extract_access_expr(iv);
2305 expr_iv = mark_write(expr_iv);
2306 type_size = pet_expr_get_type_size(expr_iv);
2307 init = pet_expr_new_binary(type_size, pet_op_assign, expr_iv, init);
2308 scop_init = extract(init, stmt->getInit()->getSourceRange(), false);
2309 scop_init = pet_scop_prefix(scop_init, declared);
2311 test_nr = n_test++;
2312 stmt_nr = n_stmt++;
2313 scop_body = extract(stmt->getBody());
2314 if (partial) {
2315 pet_scop_free(scop_init);
2316 return scop_body;
2319 expr_iv = extract_access_expr(iv);
2320 expr_iv = mark_write(expr_iv);
2321 type_size = pet_expr_get_type_size(expr_iv);
2322 inc = pet_expr_new_binary(type_size, pet_op_add_assign, expr_iv, inc);
2323 scop_inc = extract(inc, stmt->getInc()->getSourceRange(), false);
2324 if (!scop_inc) {
2325 pet_scop_free(scop_init);
2326 pet_scop_free(scop_body);
2327 return NULL;
2330 scop = extract_while(stmt->getCond(), test_nr, stmt_nr, scop_body,
2331 scop_inc);
2333 scop = pet_scop_prefix(scop, declared + 1);
2334 scop = pet_scop_add_seq(ctx, scop_init, scop);
2336 if (!declared)
2337 return scop;
2339 array = extract_array(ctx, iv, NULL);
2340 if (array)
2341 array->declared = 1;
2342 scop_kill = kill(stmt, array);
2343 scop_kill = pet_scop_prefix(scop_kill, 0);
2344 scop = pet_scop_add_seq(ctx, scop_kill, scop);
2345 scop_kill = kill(stmt, array);
2346 scop_kill = pet_scop_add_array(scop_kill, array);
2347 scop_kill = pet_scop_prefix(scop_kill, 3);
2348 scop = pet_scop_add_seq(ctx, scop, scop_kill);
2350 return scop;
2353 /* Construct a pet_scop for a for statement.
2354 * The for loop is required to be of one of the following forms
2356 * for (i = init; condition; ++i)
2357 * for (i = init; condition; --i)
2358 * for (i = init; condition; i += constant)
2359 * for (i = init; condition; i -= constant)
2361 * The initialization of the for loop should either be an assignment
2362 * of a static affine value to an integer variable, or a declaration
2363 * of such a variable with initialization.
2365 * If the initialization or the increment do not satisfy the above
2366 * conditions, i.e., if the initialization is not static affine
2367 * or the increment is not constant, then the for loop is extracted
2368 * as a while loop instead.
2370 * The condition is allowed to contain nested accesses, provided
2371 * they are not being written to inside the body of the loop.
2372 * Otherwise, or if the condition is otherwise non-affine, the for loop is
2373 * essentially treated as a while loop, with iteration domain
2374 * { [i] : i >= init }.
2376 * We extract a pet_scop for the body and then embed it in a loop with
2377 * iteration domain and schedule
2379 * { [i] : i >= init and condition' }
2380 * { [i] -> [i] }
2382 * or
2384 * { [i] : i <= init and condition' }
2385 * { [i] -> [-i] }
2387 * Where condition' is equal to condition if the latter is
2388 * a simple upper [lower] bound and a condition that is extended
2389 * to apply to all previous iterations otherwise.
2391 * If the condition is non-affine, then we drop the condition from the
2392 * iteration domain and instead create a separate statement
2393 * for evaluating the condition. The body is then filtered to depend
2394 * on the result of the condition evaluating to true on all iterations
2395 * up to the current iteration, while the evaluation the condition itself
2396 * is filtered to depend on the result of the condition evaluating to true
2397 * on all previous iterations.
2398 * The context of the scop representing the body is dropped
2399 * because we don't know how many times the body will be executed,
2400 * if at all.
2402 * If the stride of the loop is not 1, then "i >= init" is replaced by
2404 * (exists a: i = init + stride * a and a >= 0)
2406 * If the loop iterator i is unsigned, then wrapping may occur.
2407 * We therefore use a virtual iterator instead that does not wrap.
2408 * However, the condition in the code applies
2409 * to the wrapped value, so we need to change condition(i)
2410 * into condition([i % 2^width]). Similarly, we replace all accesses
2411 * to the original iterator by the wrapping of the virtual iterator.
2412 * Note that there may be no need to perform this final wrapping
2413 * if the loop condition (after wrapping) satisfies certain conditions.
2414 * However, the is_simple_bound condition is not enough since it doesn't
2415 * check if there even is an upper bound.
2417 * Wrapping on unsigned iterators can be avoided entirely if
2418 * loop condition is simple, the loop iterator is incremented
2419 * [decremented] by one and the last value before wrapping cannot
2420 * possibly satisfy the loop condition.
2422 * Before extracting a pet_scop from the body we remove all
2423 * assignments in assigned_value to variables that are assigned
2424 * somewhere in the body of the loop.
2426 * Valid parameters for a for loop are those for which the initial
2427 * value itself, the increment on each domain iteration and
2428 * the condition on both the initial value and
2429 * the result of incrementing the iterator for each iteration of the domain
2430 * can be evaluated.
2431 * If the loop condition is non-affine, then we only consider validity
2432 * of the initial value.
2434 * If the body contains any break, then we keep track of it in "skip"
2435 * (if the skip condition is affine) or it is handled in scop_add_break
2436 * (if the skip condition is not affine).
2437 * Note that the affine break condition needs to be considered with
2438 * respect to previous iterations in the virtual domain (if any).
2440 * If we were only able to extract part of the body, then simply
2441 * return that part.
2443 struct pet_scop *PetScan::extract_for(ForStmt *stmt)
2445 BinaryOperator *ass;
2446 Decl *decl;
2447 Stmt *init;
2448 Expr *lhs, *rhs;
2449 ValueDecl *iv;
2450 isl_local_space *ls;
2451 isl_set *domain;
2452 isl_aff *sched;
2453 isl_set *cond = NULL;
2454 isl_set *skip = NULL;
2455 isl_id *id, *id_test = NULL, *id_break_test;
2456 struct pet_scop *scop, *scop_cond = NULL;
2457 assigned_value_cache cache(assigned_value);
2458 isl_val *inc;
2459 bool is_one;
2460 bool is_unsigned;
2461 bool is_simple;
2462 bool is_virtual;
2463 bool has_affine_break;
2464 bool has_var_break;
2465 isl_aff *wrap = NULL;
2466 isl_pw_aff *pa, *pa_inc, *init_val;
2467 isl_set *valid_init;
2468 isl_set *valid_cond;
2469 isl_set *valid_cond_init;
2470 isl_set *valid_cond_next;
2471 isl_set *valid_inc;
2472 int stmt_id;
2473 pet_expr *pe_init, *pe_inc;
2474 pet_context *pc, *pc_init_val;
2476 if (!stmt->getInit() && !stmt->getCond() && !stmt->getInc())
2477 return extract_infinite_for(stmt);
2479 init = stmt->getInit();
2480 if (!init) {
2481 unsupported(stmt);
2482 return NULL;
2484 if ((ass = initialization_assignment(init)) != NULL) {
2485 iv = extract_induction_variable(ass);
2486 if (!iv)
2487 return NULL;
2488 lhs = ass->getLHS();
2489 rhs = ass->getRHS();
2490 } else if ((decl = initialization_declaration(init)) != NULL) {
2491 VarDecl *var = extract_induction_variable(init, decl);
2492 if (!var)
2493 return NULL;
2494 iv = var;
2495 rhs = var->getInit();
2496 lhs = create_DeclRefExpr(var);
2497 } else {
2498 unsupported(stmt->getInit());
2499 return NULL;
2502 id = create_decl_id(ctx, iv);
2504 assigned_value.erase(iv);
2505 clear_assignments clear(assigned_value);
2506 clear.TraverseStmt(stmt->getBody());
2508 pe_init = extract_expr(rhs);
2509 pe_inc = extract_increment(stmt, iv);
2510 pc = convert_assignments(ctx, assigned_value);
2511 pc_init_val = pet_context_copy(pc);
2512 pc_init_val = pet_context_mark_unknown(pc_init_val, isl_id_copy(id));
2513 init_val = pet_expr_extract_affine(pe_init, pc_init_val);
2514 pet_context_free(pc_init_val);
2515 pa_inc = pet_expr_extract_affine(pe_inc, pc);
2516 pet_context_free(pc);
2517 inc = pet_extract_cst(pa_inc);
2518 if (!pe_init || !pe_inc || !inc || isl_val_is_nan(inc) ||
2519 isl_pw_aff_involves_nan(pa_inc) ||
2520 isl_pw_aff_involves_nan(init_val)) {
2521 isl_id_free(id);
2522 isl_val_free(inc);
2523 isl_pw_aff_free(pa_inc);
2524 isl_pw_aff_free(init_val);
2525 if (pe_init && pe_inc && !(pa_inc && !inc))
2526 return extract_non_affine_for(stmt, iv,
2527 pe_init, pe_inc);
2528 pet_expr_free(pe_init);
2529 pet_expr_free(pe_inc);
2530 return NULL;
2532 pet_expr_free(pe_init);
2533 pet_expr_free(pe_inc);
2535 pa = try_extract_nested_condition(stmt->getCond());
2536 if (allow_nested && (!pa || pet_nested_any_in_pw_aff(pa)))
2537 stmt_id = n_stmt++;
2539 scop = extract(stmt->getBody());
2540 if (partial) {
2541 isl_id_free(id);
2542 isl_pw_aff_free(init_val);
2543 isl_pw_aff_free(pa_inc);
2544 isl_pw_aff_free(pa);
2545 isl_val_free(inc);
2546 return scop;
2549 valid_inc = isl_pw_aff_domain(pa_inc);
2551 is_unsigned = iv->getType()->isUnsignedIntegerType();
2553 has_affine_break = scop &&
2554 pet_scop_has_affine_skip(scop, pet_skip_later);
2555 if (has_affine_break)
2556 skip = pet_scop_get_affine_skip_domain(scop, pet_skip_later);
2557 has_var_break = scop && pet_scop_has_var_skip(scop, pet_skip_later);
2558 if (has_var_break)
2559 id_break_test = pet_scop_get_skip_id(scop, pet_skip_later);
2561 if (pa && !is_nested_allowed(pa, scop)) {
2562 isl_pw_aff_free(pa);
2563 pa = NULL;
2566 if (!allow_nested && !pa)
2567 pa = extract_condition(stmt->getCond());
2568 valid_cond = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2569 cond = isl_pw_aff_non_zero_set(pa);
2570 if (allow_nested && !cond) {
2571 isl_multi_pw_aff *test_index;
2572 int save_n_stmt = n_stmt;
2573 test_index = pet_create_test_index(ctx, n_test++);
2574 n_stmt = stmt_id;
2575 scop_cond = extract_non_affine_condition(stmt->getCond(),
2576 n_stmt++, isl_multi_pw_aff_copy(test_index));
2577 n_stmt = save_n_stmt;
2578 scop_cond = scop_add_array(scop_cond, test_index, ast_context);
2579 id_test = isl_multi_pw_aff_get_tuple_id(test_index,
2580 isl_dim_out);
2581 isl_multi_pw_aff_free(test_index);
2582 scop_cond = pet_scop_prefix(scop_cond, 0);
2583 scop = pet_scop_reset_context(scop);
2584 scop = pet_scop_prefix(scop, 1);
2585 cond = isl_set_universe(isl_space_set_alloc(ctx, 0, 0));
2588 cond = embed(cond, isl_id_copy(id));
2589 skip = embed(skip, isl_id_copy(id));
2590 valid_cond = isl_set_coalesce(valid_cond);
2591 valid_cond = embed(valid_cond, isl_id_copy(id));
2592 valid_inc = embed(valid_inc, isl_id_copy(id));
2593 is_one = isl_val_is_one(inc) || isl_val_is_negone(inc);
2594 is_virtual = is_unsigned && (!is_one || can_wrap(cond, iv, inc));
2596 valid_cond_init = enforce_subset(
2597 isl_map_range(isl_map_from_pw_aff(isl_pw_aff_copy(init_val))),
2598 isl_set_copy(valid_cond));
2599 if (is_one && !is_virtual) {
2600 isl_pw_aff_free(init_val);
2601 pa = extract_comparison(isl_val_is_pos(inc) ? BO_GE : BO_LE,
2602 lhs, rhs, init);
2603 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2604 valid_init = set_project_out_by_id(valid_init, id);
2605 domain = isl_pw_aff_non_zero_set(pa);
2606 } else {
2607 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(init_val));
2608 domain = strided_domain(isl_id_copy(id), init_val,
2609 isl_val_copy(inc));
2612 domain = embed(domain, isl_id_copy(id));
2613 if (is_virtual) {
2614 isl_map *rev_wrap;
2615 wrap = compute_wrapping(isl_set_get_space(cond), iv);
2616 rev_wrap = isl_map_from_aff(isl_aff_copy(wrap));
2617 rev_wrap = isl_map_reverse(rev_wrap);
2618 cond = isl_set_apply(cond, isl_map_copy(rev_wrap));
2619 skip = isl_set_apply(skip, isl_map_copy(rev_wrap));
2620 valid_cond = isl_set_apply(valid_cond, isl_map_copy(rev_wrap));
2621 valid_inc = isl_set_apply(valid_inc, rev_wrap);
2623 is_simple = is_simple_bound(cond, inc);
2624 if (!is_simple) {
2625 cond = isl_set_gist(cond, isl_set_copy(domain));
2626 is_simple = is_simple_bound(cond, inc);
2628 if (!is_simple)
2629 cond = valid_for_each_iteration(cond,
2630 isl_set_copy(domain), isl_val_copy(inc));
2631 domain = isl_set_intersect(domain, cond);
2632 if (has_affine_break) {
2633 skip = isl_set_intersect(skip , isl_set_copy(domain));
2634 skip = after(skip, isl_val_sgn(inc));
2635 domain = isl_set_subtract(domain, skip);
2637 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, isl_id_copy(id));
2638 ls = isl_local_space_from_space(isl_set_get_space(domain));
2639 sched = isl_aff_var_on_domain(ls, isl_dim_set, 0);
2640 if (isl_val_is_neg(inc))
2641 sched = isl_aff_neg(sched);
2643 valid_cond_next = valid_on_next(valid_cond, isl_set_copy(domain),
2644 isl_val_copy(inc));
2645 valid_inc = enforce_subset(isl_set_copy(domain), valid_inc);
2647 if (!is_virtual)
2648 wrap = identity_aff(domain);
2650 scop_cond = pet_scop_embed(scop_cond, isl_set_copy(domain),
2651 isl_aff_copy(sched), isl_aff_copy(wrap), isl_id_copy(id));
2652 scop = pet_scop_embed(scop, isl_set_copy(domain), sched, wrap, id);
2653 scop = resolve_nested(scop);
2654 if (has_var_break)
2655 scop = scop_add_break(scop, id_break_test, isl_set_copy(domain),
2656 isl_val_copy(inc));
2657 if (id_test) {
2658 scop = scop_add_while(scop_cond, scop, id_test, domain,
2659 isl_val_copy(inc));
2660 isl_set_free(valid_inc);
2661 } else {
2662 scop = pet_scop_restrict_context(scop, valid_inc);
2663 scop = pet_scop_restrict_context(scop, valid_cond_next);
2664 scop = pet_scop_restrict_context(scop, valid_cond_init);
2665 isl_set_free(domain);
2667 clear_assignment(assigned_value, iv);
2669 isl_val_free(inc);
2671 scop = pet_scop_restrict_context(scop, isl_set_params(valid_init));
2673 return scop;
2676 /* Try and construct a pet_scop corresponding to a compound statement.
2678 * "skip_declarations" is set if we should skip initial declarations
2679 * in the children of the compound statements. This then implies
2680 * that this sequence of children should not be treated as a block
2681 * since the initial statements may be skipped.
2683 struct pet_scop *PetScan::extract(CompoundStmt *stmt, bool skip_declarations)
2685 return extract(stmt->children(), !skip_declarations, skip_declarations);
2688 /* For each nested access parameter in "space",
2689 * construct a corresponding pet_expr, place it in args and
2690 * record its position in "param2pos".
2691 * "n_arg" is the number of elements that are already in args.
2692 * The position recorded in "param2pos" takes this number into account.
2693 * If the pet_expr corresponding to a parameter is identical to
2694 * the pet_expr corresponding to an earlier parameter, then these two
2695 * parameters are made to refer to the same element in args.
2697 * Return the final number of elements in args or -1 if an error has occurred.
2699 int PetScan::extract_nested(__isl_keep isl_space *space,
2700 int n_arg, pet_expr **args, std::map<int,int> &param2pos)
2702 int nparam;
2704 nparam = isl_space_dim(space, isl_dim_param);
2705 for (int i = 0; i < nparam; ++i) {
2706 int j;
2707 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
2709 if (!pet_nested_in_id(id)) {
2710 isl_id_free(id);
2711 continue;
2714 args[n_arg] = pet_nested_extract_expr(id);
2715 isl_id_free(id);
2716 if (!args[n_arg])
2717 return -1;
2719 for (j = 0; j < n_arg; ++j)
2720 if (pet_expr_is_equal(args[j], args[n_arg]))
2721 break;
2723 if (j < n_arg) {
2724 pet_expr_free(args[n_arg]);
2725 args[n_arg] = NULL;
2726 param2pos[i] = j;
2727 } else
2728 param2pos[i] = n_arg++;
2731 return n_arg;
2734 /* For each nested access parameter in the access relations in "expr",
2735 * construct a corresponding pet_expr, append it to the arguments of "expr"
2736 * and record its position in "param2pos" (relative to the initial
2737 * number of arguments).
2738 * n is the number of nested access parameters.
2740 __isl_give pet_expr *PetScan::extract_nested(__isl_take pet_expr *expr, int n,
2741 std::map<int,int> &param2pos)
2743 isl_space *space;
2744 int i, n_arg;
2745 pet_expr **args;
2747 args = isl_calloc_array(ctx, pet_expr *, n);
2748 if (!args)
2749 return pet_expr_free(expr);
2751 n_arg = pet_expr_get_n_arg(expr);
2752 space = pet_expr_access_get_parameter_space(expr);
2753 n = extract_nested(space, 0, args, param2pos);
2754 isl_space_free(space);
2756 if (n < 0)
2757 expr = pet_expr_free(expr);
2758 else
2759 expr = pet_expr_set_n_arg(expr, n_arg + n);
2761 for (i = 0; i < n; ++i)
2762 expr = pet_expr_set_arg(expr, n_arg + i, args[i]);
2763 free(args);
2765 return expr;
2768 /* Are "expr1" and "expr2" both array accesses such that
2769 * the access relation of "expr1" is a subset of that of "expr2"?
2770 * Only take into account the first "n_arg" arguments.
2772 static int is_sub_access(__isl_keep pet_expr *expr1, __isl_keep pet_expr *expr2,
2773 int n_arg)
2775 isl_id *id1, *id2;
2776 isl_map *access1, *access2;
2777 int is_subset;
2778 int i, n1, n2;
2780 if (!expr1 || !expr2)
2781 return 0;
2782 if (pet_expr_get_type(expr1) != pet_expr_access)
2783 return 0;
2784 if (pet_expr_get_type(expr2) != pet_expr_access)
2785 return 0;
2786 if (pet_expr_is_affine(expr1))
2787 return 0;
2788 if (pet_expr_is_affine(expr2))
2789 return 0;
2790 n1 = pet_expr_get_n_arg(expr1);
2791 if (n1 > n_arg)
2792 n1 = n_arg;
2793 n2 = pet_expr_get_n_arg(expr2);
2794 if (n2 > n_arg)
2795 n2 = n_arg;
2796 if (n1 != n2)
2797 return 0;
2798 for (i = 0; i < n1; ++i) {
2799 pet_expr *arg1, *arg2;
2800 int equal;
2801 arg1 = pet_expr_get_arg(expr1, i);
2802 arg2 = pet_expr_get_arg(expr2, i);
2803 equal = pet_expr_is_equal(arg1, arg2);
2804 pet_expr_free(arg1);
2805 pet_expr_free(arg2);
2806 if (equal < 0 || !equal)
2807 return equal;
2809 id1 = pet_expr_access_get_id(expr1);
2810 id2 = pet_expr_access_get_id(expr2);
2811 isl_id_free(id1);
2812 isl_id_free(id2);
2813 if (!id1 || !id2)
2814 return 0;
2815 if (id1 != id2)
2816 return 0;
2818 access1 = pet_expr_access_get_access(expr1);
2819 access2 = pet_expr_access_get_access(expr2);
2820 is_subset = isl_map_is_subset(access1, access2);
2821 isl_map_free(access1);
2822 isl_map_free(access2);
2824 return is_subset;
2827 /* Mark self dependences among the arguments of "expr" starting at "first".
2828 * These arguments have already been added to the list of arguments
2829 * but are not yet referenced directly from the index expression.
2830 * Instead, they are still referenced through parameters encoding
2831 * nested accesses.
2833 * In particular, if "expr" is a read access, then check the arguments
2834 * starting at "first" to see if "expr" accesses a subset of
2835 * the elements accessed by the argument, or under more restrictive conditions.
2836 * If so, then this nested access can be removed from the constraints
2837 * governing the outer access. There is no point in restricting
2838 * accesses to an array if in order to evaluate the restriction,
2839 * we have to access the same elements (or more).
2841 * Rather than removing the argument at this point (which would
2842 * complicate the resolution of the other nested accesses), we simply
2843 * mark it here by replacing it by a NaN pet_expr.
2844 * These NaNs are then later removed in remove_marked_self_dependences.
2846 static __isl_give pet_expr *mark_self_dependences(__isl_take pet_expr *expr,
2847 int first)
2849 int n;
2851 if (pet_expr_access_is_write(expr))
2852 return expr;
2854 n = pet_expr_get_n_arg(expr);
2855 for (int i = first; i < n; ++i) {
2856 int mark;
2857 pet_expr *arg;
2859 arg = pet_expr_get_arg(expr, i);
2860 mark = is_sub_access(expr, arg, first);
2861 pet_expr_free(arg);
2862 if (mark < 0)
2863 return pet_expr_free(expr);
2864 if (!mark)
2865 continue;
2867 arg = pet_expr_new_int(isl_val_nan(pet_expr_get_ctx(expr)));
2868 expr = pet_expr_set_arg(expr, i, arg);
2871 return expr;
2874 /* Is "expr" a NaN integer expression?
2876 static int expr_is_nan(__isl_keep pet_expr *expr)
2878 isl_val *v;
2879 int is_nan;
2881 if (pet_expr_get_type(expr) != pet_expr_int)
2882 return 0;
2884 v = pet_expr_int_get_val(expr);
2885 is_nan = isl_val_is_nan(v);
2886 isl_val_free(v);
2888 return is_nan;
2891 /* Check if we have marked any self dependences (as NaNs)
2892 * in mark_self_dependences and remove them here.
2893 * It is safe to project them out since these arguments
2894 * can at most be referenced from the condition of the access relation,
2895 * but do not appear in the index expression.
2896 * "dim" is the dimension of the iteration domain.
2898 static __isl_give pet_expr *remove_marked_self_dependences(
2899 __isl_take pet_expr *expr, int dim, int first)
2901 int n;
2903 n = pet_expr_get_n_arg(expr);
2904 for (int i = n - 1; i >= first; --i) {
2905 int is_nan;
2906 pet_expr *arg;
2908 arg = pet_expr_get_arg(expr, i);
2909 is_nan = expr_is_nan(arg);
2910 pet_expr_free(arg);
2911 if (!is_nan)
2912 continue;
2913 expr = pet_expr_access_project_out_arg(expr, dim, i);
2916 return expr;
2919 /* Look for parameters in any access relation in "expr" that
2920 * refer to nested accesses. In particular, these are
2921 * parameters with name "__pet_expr".
2923 * If there are any such parameters, then the domain of the index
2924 * expression and the access relation, which is either [] or
2925 * [[] -> [a_1,...,a_m]] at this point, is replaced by [[] -> [t_1,...,t_n]] or
2926 * [[] -> [a_1,...,a_m,t_1,...,t_n]], with m the original number of arguments
2927 * (n_arg) and n the number of these parameters
2928 * (after identifying identical nested accesses).
2930 * This transformation is performed in several steps.
2931 * We first extract the arguments in extract_nested.
2932 * param2pos maps the original parameter position to the position
2933 * of the argument beyond the initial (n_arg) number of arguments.
2934 * Then we move these parameters to input dimensions.
2935 * t2pos maps the positions of these temporary input dimensions
2936 * to the positions of the corresponding arguments.
2937 * Finally, we express these temporary dimensions in terms of the domain
2938 * [[] -> [a_1,...,a_m,t_1,...,t_n]] and precompose index expression and access
2939 * relations with this function.
2941 __isl_give pet_expr *PetScan::resolve_nested(__isl_take pet_expr *expr)
2943 int n, n_arg;
2944 int nparam;
2945 isl_space *space;
2946 isl_local_space *ls;
2947 isl_aff *aff;
2948 isl_multi_aff *ma;
2949 std::map<int,int> param2pos;
2950 std::map<int,int> t2pos;
2952 if (!expr)
2953 return expr;
2955 n_arg = pet_expr_get_n_arg(expr);
2956 for (int i = 0; i < n_arg; ++i) {
2957 pet_expr *arg;
2958 arg = pet_expr_get_arg(expr, i);
2959 arg = resolve_nested(arg);
2960 expr = pet_expr_set_arg(expr, i, arg);
2963 if (pet_expr_get_type(expr) != pet_expr_access)
2964 return expr;
2966 space = pet_expr_access_get_parameter_space(expr);
2967 n = pet_nested_n_in_space(space);
2968 isl_space_free(space);
2969 if (n == 0)
2970 return expr;
2972 expr = extract_nested(expr, n, param2pos);
2973 if (!expr)
2974 return NULL;
2976 expr = pet_expr_access_align_params(expr);
2977 expr = mark_self_dependences(expr, n_arg);
2978 if (!expr)
2979 return NULL;
2981 n = 0;
2982 space = pet_expr_access_get_parameter_space(expr);
2983 nparam = isl_space_dim(space, isl_dim_param);
2984 for (int i = nparam - 1; i >= 0; --i) {
2985 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
2986 if (!pet_nested_in_id(id)) {
2987 isl_id_free(id);
2988 continue;
2991 expr = pet_expr_access_move_dims(expr,
2992 isl_dim_in, n_arg + n, isl_dim_param, i, 1);
2993 t2pos[n] = n_arg + param2pos[i];
2994 n++;
2996 isl_id_free(id);
2998 isl_space_free(space);
3000 space = pet_expr_access_get_parameter_space(expr);
3001 space = isl_space_set_from_params(space);
3002 space = isl_space_add_dims(space, isl_dim_set,
3003 pet_expr_get_n_arg(expr));
3004 space = isl_space_wrap(isl_space_from_range(space));
3005 ls = isl_local_space_from_space(isl_space_copy(space));
3006 space = isl_space_from_domain(space);
3007 space = isl_space_add_dims(space, isl_dim_out, n_arg + n);
3008 ma = isl_multi_aff_zero(space);
3010 for (int i = 0; i < n_arg; ++i) {
3011 aff = isl_aff_var_on_domain(isl_local_space_copy(ls),
3012 isl_dim_set, i);
3013 ma = isl_multi_aff_set_aff(ma, i, aff);
3015 for (int i = 0; i < n; ++i) {
3016 aff = isl_aff_var_on_domain(isl_local_space_copy(ls),
3017 isl_dim_set, t2pos[i]);
3018 ma = isl_multi_aff_set_aff(ma, n_arg + i, aff);
3020 isl_local_space_free(ls);
3022 expr = pet_expr_access_pullback_multi_aff(expr, ma);
3024 expr = remove_marked_self_dependences(expr, 0, n_arg);
3026 return expr;
3029 /* Return the file offset of the expansion location of "Loc".
3031 static unsigned getExpansionOffset(SourceManager &SM, SourceLocation Loc)
3033 return SM.getFileOffset(SM.getExpansionLoc(Loc));
3036 #ifdef HAVE_FINDLOCATIONAFTERTOKEN
3038 /* Return a SourceLocation for the location after the first semicolon
3039 * after "loc". If Lexer::findLocationAfterToken is available, we simply
3040 * call it and also skip trailing spaces and newline.
3042 static SourceLocation location_after_semi(SourceLocation loc, SourceManager &SM,
3043 const LangOptions &LO)
3045 return Lexer::findLocationAfterToken(loc, tok::semi, SM, LO, true);
3048 #else
3050 /* Return a SourceLocation for the location after the first semicolon
3051 * after "loc". If Lexer::findLocationAfterToken is not available,
3052 * we look in the underlying character data for the first semicolon.
3054 static SourceLocation location_after_semi(SourceLocation loc, SourceManager &SM,
3055 const LangOptions &LO)
3057 const char *semi;
3058 const char *s = SM.getCharacterData(loc);
3060 semi = strchr(s, ';');
3061 if (!semi)
3062 return SourceLocation();
3063 return loc.getFileLocWithOffset(semi + 1 - s);
3066 #endif
3068 /* If the token at "loc" is the first token on the line, then return
3069 * a location referring to the start of the line.
3070 * Otherwise, return "loc".
3072 * This function is used to extend a scop to the start of the line
3073 * if the first token of the scop is also the first token on the line.
3075 * We look for the first token on the line. If its location is equal to "loc",
3076 * then the latter is the location of the first token on the line.
3078 static SourceLocation move_to_start_of_line_if_first_token(SourceLocation loc,
3079 SourceManager &SM, const LangOptions &LO)
3081 std::pair<FileID, unsigned> file_offset_pair;
3082 llvm::StringRef file;
3083 const char *pos;
3084 Token tok;
3085 SourceLocation token_loc, line_loc;
3086 int col;
3088 loc = SM.getExpansionLoc(loc);
3089 col = SM.getExpansionColumnNumber(loc);
3090 line_loc = loc.getLocWithOffset(1 - col);
3091 file_offset_pair = SM.getDecomposedLoc(line_loc);
3092 file = SM.getBufferData(file_offset_pair.first, NULL);
3093 pos = file.data() + file_offset_pair.second;
3095 Lexer lexer(SM.getLocForStartOfFile(file_offset_pair.first), LO,
3096 file.begin(), pos, file.end());
3097 lexer.LexFromRawLexer(tok);
3098 token_loc = tok.getLocation();
3100 if (token_loc == loc)
3101 return line_loc;
3102 else
3103 return loc;
3106 /* If "expr" is an assume expression, then try and convert
3107 * its single argument to an affine expression.
3109 __isl_give pet_expr *PetScan::resolve_assume(__isl_take pet_expr *expr)
3111 pet_context *pc;
3113 if (!expr)
3114 return NULL;
3115 if (!pet_expr_is_assume(expr))
3116 return expr;
3118 pc = convert_assignments(ctx, assigned_value);
3119 expr = pet_expr_resolve_assume(expr, pc);
3120 pet_context_free(pc);
3122 return expr;
3125 /* Update start and end of "scop" to include the region covered by "range".
3126 * If "skip_semi" is set, then we assume "range" is followed by
3127 * a semicolon and also include this semicolon.
3129 struct pet_scop *PetScan::update_scop_start_end(struct pet_scop *scop,
3130 SourceRange range, bool skip_semi)
3132 SourceLocation loc = range.getBegin();
3133 SourceManager &SM = PP.getSourceManager();
3134 const LangOptions &LO = PP.getLangOpts();
3135 unsigned start, end;
3137 loc = move_to_start_of_line_if_first_token(loc, SM, LO);
3138 start = getExpansionOffset(SM, loc);
3139 loc = range.getEnd();
3140 if (skip_semi)
3141 loc = location_after_semi(loc, SM, LO);
3142 else
3143 loc = PP.getLocForEndOfToken(loc);
3144 end = getExpansionOffset(SM, loc);
3146 scop = pet_scop_update_start_end(scop, start, end);
3147 return scop;
3150 /* Convert a top-level pet_expr to a pet_scop with one statement.
3151 * This mainly involves resolving nested expression parameters
3152 * and setting the name of the iteration space.
3153 * The name is given by "label" if it is non-NULL. Otherwise,
3154 * it is of the form S_<n_stmt>.
3155 * start and end of the pet_scop are derived from "range" and "skip_semi".
3156 * In particular, if "skip_semi" is set then the semicolon following "range"
3157 * is also included.
3159 struct pet_scop *PetScan::extract(__isl_take pet_expr *expr, SourceRange range,
3160 bool skip_semi, __isl_take isl_id *label)
3162 struct pet_stmt *ps;
3163 struct pet_scop *scop;
3164 SourceLocation loc = range.getBegin();
3165 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3166 pet_context *pc;
3168 pc = convert_assignments(ctx, assigned_value);
3169 expr = pet_expr_plug_in_args(expr, pc);
3170 pet_context_free(pc);
3172 expr = resolve_nested(expr);
3173 expr = resolve_assume(expr);
3174 ps = pet_stmt_from_pet_expr(line, label, n_stmt++, expr);
3175 scop = pet_scop_from_pet_stmt(ctx, ps);
3177 scop = update_scop_start_end(scop, range, skip_semi);
3178 return scop;
3181 /* Check whether "expr" is an affine constraint.
3183 bool PetScan::is_affine_condition(Expr *expr)
3185 isl_pw_aff *cond;
3187 cond = extract_condition(expr);
3188 isl_pw_aff_free(cond);
3190 return cond != NULL;
3193 /* Check if we can extract a condition from "expr".
3194 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
3195 * If allow_nested is set, then the condition may involve parameters
3196 * corresponding to nested accesses.
3197 * We turn on autodetection so that we won't generate any warnings.
3199 __isl_give isl_pw_aff *PetScan::try_extract_nested_condition(Expr *expr)
3201 isl_pw_aff *cond;
3202 int save_autodetect = options->autodetect;
3203 bool save_nesting = nesting_enabled;
3205 options->autodetect = 1;
3206 nesting_enabled = allow_nested;
3207 cond = extract_condition(expr);
3209 options->autodetect = save_autodetect;
3210 nesting_enabled = save_nesting;
3212 return cond;
3215 /* If the top-level expression of "stmt" is an assignment, then
3216 * return that assignment as a BinaryOperator.
3217 * Otherwise return NULL.
3219 static BinaryOperator *top_assignment_or_null(Stmt *stmt)
3221 BinaryOperator *ass;
3223 if (!stmt)
3224 return NULL;
3225 if (stmt->getStmtClass() != Stmt::BinaryOperatorClass)
3226 return NULL;
3228 ass = cast<BinaryOperator>(stmt);
3229 if(ass->getOpcode() != BO_Assign)
3230 return NULL;
3232 return ass;
3235 /* Check if the given if statement is a conditional assignement
3236 * with a non-affine condition. If so, construct a pet_scop
3237 * corresponding to this conditional assignment. Otherwise return NULL.
3239 * In particular we check if "stmt" is of the form
3241 * if (condition)
3242 * a = f(...);
3243 * else
3244 * a = g(...);
3246 * where a is some array or scalar access.
3247 * The constructed pet_scop then corresponds to the expression
3249 * a = condition ? f(...) : g(...)
3251 * All access relations in f(...) are intersected with condition
3252 * while all access relation in g(...) are intersected with the complement.
3254 struct pet_scop *PetScan::extract_conditional_assignment(IfStmt *stmt)
3256 BinaryOperator *ass_then, *ass_else;
3257 pet_expr *write_then, *write_else;
3258 isl_set *cond, *comp;
3259 isl_multi_pw_aff *index;
3260 isl_pw_aff *pa;
3261 int equal;
3262 int type_size;
3263 pet_expr *pe_cond, *pe_then, *pe_else, *pe;
3264 bool save_nesting = nesting_enabled;
3266 if (!options->detect_conditional_assignment)
3267 return NULL;
3269 ass_then = top_assignment_or_null(stmt->getThen());
3270 ass_else = top_assignment_or_null(stmt->getElse());
3272 if (!ass_then || !ass_else)
3273 return NULL;
3275 if (is_affine_condition(stmt->getCond()))
3276 return NULL;
3278 write_then = extract_access_expr(ass_then->getLHS());
3279 write_else = extract_access_expr(ass_else->getLHS());
3281 equal = pet_expr_is_equal(write_then, write_else);
3282 pet_expr_free(write_else);
3283 if (equal < 0 || !equal) {
3284 pet_expr_free(write_then);
3285 return NULL;
3288 nesting_enabled = allow_nested;
3289 pa = extract_condition(stmt->getCond());
3290 nesting_enabled = save_nesting;
3291 cond = isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa));
3292 comp = isl_pw_aff_zero_set(isl_pw_aff_copy(pa));
3293 index = isl_multi_pw_aff_from_pw_aff(pa);
3295 pe_cond = pet_expr_from_index(index);
3297 pe_then = extract_expr(ass_then->getRHS());
3298 pe_then = pet_expr_restrict(pe_then, cond);
3299 pe_else = extract_expr(ass_else->getRHS());
3300 pe_else = pet_expr_restrict(pe_else, comp);
3302 pe = pet_expr_new_ternary(pe_cond, pe_then, pe_else);
3303 write_then = pet_expr_access_set_write(write_then, 1);
3304 write_then = pet_expr_access_set_read(write_then, 0);
3305 type_size = get_type_size(ass_then->getType(), ast_context);
3306 pe = pet_expr_new_binary(type_size, pet_op_assign, write_then, pe);
3307 return extract(pe, stmt->getSourceRange(), false);
3310 /* Create a pet_scop with a single statement with name S_<stmt_nr>,
3311 * evaluating "cond" and writing the result to a virtual scalar,
3312 * as expressed by "index".
3314 struct pet_scop *PetScan::extract_non_affine_condition(Expr *cond, int stmt_nr,
3315 __isl_take isl_multi_pw_aff *index)
3317 pet_expr *expr, *write;
3318 struct pet_stmt *ps;
3319 SourceLocation loc = cond->getLocStart();
3320 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3321 pet_context *pc;
3323 write = pet_expr_from_index(index);
3324 write = pet_expr_access_set_write(write, 1);
3325 write = pet_expr_access_set_read(write, 0);
3326 expr = extract_expr(cond);
3328 pc = convert_assignments(ctx, assigned_value);
3329 expr = pet_expr_plug_in_args(expr, pc);
3330 pet_context_free(pc);
3332 expr = resolve_nested(expr);
3333 expr = pet_expr_new_binary(1, pet_op_assign, write, expr);
3334 ps = pet_stmt_from_pet_expr(line, NULL, stmt_nr, expr);
3335 return pet_scop_from_pet_stmt(ctx, ps);
3338 extern "C" {
3339 static __isl_give pet_expr *embed_access(__isl_take pet_expr *expr,
3340 void *user);
3343 /* Precompose the access relation and the index expression associated
3344 * to "expr" with the function pointed to by "user",
3345 * thereby embedding the access relation in the domain of this function.
3346 * The initial domain of the access relation and the index expression
3347 * is the zero-dimensional domain.
3349 static __isl_give pet_expr *embed_access(__isl_take pet_expr *expr, void *user)
3351 isl_multi_aff *ma = (isl_multi_aff *) user;
3353 return pet_expr_access_pullback_multi_aff(expr, isl_multi_aff_copy(ma));
3356 /* Precompose all access relations in "expr" with "ma", thereby
3357 * embedding them in the domain of "ma".
3359 static __isl_give pet_expr *embed(__isl_take pet_expr *expr,
3360 __isl_keep isl_multi_aff *ma)
3362 return pet_expr_map_access(expr, &embed_access, ma);
3365 /* For each nested access parameter in the domain of "stmt",
3366 * construct a corresponding pet_expr, place it before the original
3367 * elements in stmt->args and record its position in "param2pos".
3368 * n is the number of nested access parameters.
3370 struct pet_stmt *PetScan::extract_nested(struct pet_stmt *stmt, int n,
3371 std::map<int,int> &param2pos)
3373 int i;
3374 isl_space *space;
3375 int n_arg;
3376 pet_expr **args;
3378 n_arg = stmt->n_arg;
3379 args = isl_calloc_array(ctx, pet_expr *, n + n_arg);
3380 if (!args)
3381 goto error;
3383 space = isl_set_get_space(stmt->domain);
3384 n_arg = extract_nested(space, 0, args, param2pos);
3385 isl_space_free(space);
3387 if (n_arg < 0)
3388 goto error;
3390 for (i = 0; i < stmt->n_arg; ++i)
3391 args[n_arg + i] = stmt->args[i];
3392 free(stmt->args);
3393 stmt->args = args;
3394 stmt->n_arg += n_arg;
3396 return stmt;
3397 error:
3398 if (args) {
3399 for (i = 0; i < n; ++i)
3400 pet_expr_free(args[i]);
3401 free(args);
3403 pet_stmt_free(stmt);
3404 return NULL;
3407 /* Check whether any of the arguments i of "stmt" starting at position "n"
3408 * is equal to one of the first "n" arguments j.
3409 * If so, combine the constraints on arguments i and j and remove
3410 * argument i.
3412 static struct pet_stmt *remove_duplicate_arguments(struct pet_stmt *stmt, int n)
3414 int i, j;
3415 isl_map *map;
3417 if (!stmt)
3418 return NULL;
3419 if (n == 0)
3420 return stmt;
3421 if (n == stmt->n_arg)
3422 return stmt;
3424 map = isl_set_unwrap(stmt->domain);
3426 for (i = stmt->n_arg - 1; i >= n; --i) {
3427 for (j = 0; j < n; ++j)
3428 if (pet_expr_is_equal(stmt->args[i], stmt->args[j]))
3429 break;
3430 if (j >= n)
3431 continue;
3433 map = isl_map_equate(map, isl_dim_out, i, isl_dim_out, j);
3434 map = isl_map_project_out(map, isl_dim_out, i, 1);
3436 pet_expr_free(stmt->args[i]);
3437 for (j = i; j + 1 < stmt->n_arg; ++j)
3438 stmt->args[j] = stmt->args[j + 1];
3439 stmt->n_arg--;
3442 stmt->domain = isl_map_wrap(map);
3443 if (!stmt->domain)
3444 goto error;
3445 return stmt;
3446 error:
3447 pet_stmt_free(stmt);
3448 return NULL;
3451 /* Look for parameters in the iteration domain of "stmt" that
3452 * refer to nested accesses. In particular, these are
3453 * parameters with name "__pet_expr".
3455 * If there are any such parameters, then as many extra variables
3456 * (after identifying identical nested accesses) are inserted in the
3457 * range of the map wrapped inside the domain, before the original variables.
3458 * If the original domain is not a wrapped map, then a new wrapped
3459 * map is created with zero output dimensions.
3460 * The parameters are then equated to the corresponding output dimensions
3461 * and subsequently projected out, from the iteration domain,
3462 * the schedule and the access relations.
3463 * For each of the output dimensions, a corresponding argument
3464 * expression is inserted. Initially they are created with
3465 * a zero-dimensional domain, so they have to be embedded
3466 * in the current iteration domain.
3467 * param2pos maps the position of the parameter to the position
3468 * of the corresponding output dimension in the wrapped map.
3470 struct pet_stmt *PetScan::resolve_nested(struct pet_stmt *stmt)
3472 int n;
3473 int nparam;
3474 unsigned n_arg;
3475 isl_map *map;
3476 isl_space *space;
3477 isl_multi_aff *ma;
3478 std::map<int,int> param2pos;
3480 if (!stmt)
3481 return NULL;
3483 n = pet_nested_n_in_set(stmt->domain);
3484 if (n == 0)
3485 return stmt;
3487 n_arg = stmt->n_arg;
3488 stmt = extract_nested(stmt, n, param2pos);
3489 if (!stmt)
3490 return NULL;
3492 n = stmt->n_arg - n_arg;
3493 nparam = isl_set_dim(stmt->domain, isl_dim_param);
3494 if (isl_set_is_wrapping(stmt->domain))
3495 map = isl_set_unwrap(stmt->domain);
3496 else
3497 map = isl_map_from_domain(stmt->domain);
3498 map = isl_map_insert_dims(map, isl_dim_out, 0, n);
3500 for (int i = nparam - 1; i >= 0; --i) {
3501 isl_id *id;
3503 if (!pet_nested_in_map(map, i))
3504 continue;
3506 id = pet_expr_access_get_id(stmt->args[param2pos[i]]);
3507 map = isl_map_set_dim_id(map, isl_dim_out, param2pos[i], id);
3508 map = isl_map_equate(map, isl_dim_param, i, isl_dim_out,
3509 param2pos[i]);
3510 map = isl_map_project_out(map, isl_dim_param, i, 1);
3513 stmt->domain = isl_map_wrap(map);
3515 space = isl_space_unwrap(isl_set_get_space(stmt->domain));
3516 space = isl_space_from_domain(isl_space_domain(space));
3517 ma = isl_multi_aff_zero(space);
3518 for (int pos = 0; pos < n; ++pos)
3519 stmt->args[pos] = embed(stmt->args[pos], ma);
3520 isl_multi_aff_free(ma);
3522 stmt = pet_stmt_remove_nested_parameters(stmt);
3523 stmt = remove_duplicate_arguments(stmt, n);
3525 return stmt;
3528 /* For each statement in "scop", move the parameters that correspond
3529 * to nested access into the ranges of the domains and create
3530 * corresponding argument expressions.
3532 struct pet_scop *PetScan::resolve_nested(struct pet_scop *scop)
3534 if (!scop)
3535 return NULL;
3537 for (int i = 0; i < scop->n_stmt; ++i) {
3538 scop->stmts[i] = resolve_nested(scop->stmts[i]);
3539 if (!scop->stmts[i])
3540 goto error;
3543 return scop;
3544 error:
3545 pet_scop_free(scop);
3546 return NULL;
3549 /* Given an access expression "expr", is the variable accessed by
3550 * "expr" assigned anywhere inside "scop"?
3552 static bool is_assigned(__isl_keep pet_expr *expr, pet_scop *scop)
3554 bool assigned = false;
3555 isl_id *id;
3557 id = pet_expr_access_get_id(expr);
3558 assigned = pet_scop_writes(scop, id);
3559 isl_id_free(id);
3561 return assigned;
3564 /* Are all nested access parameters in "pa" allowed given "scop".
3565 * In particular, is none of them written by anywhere inside "scop".
3567 * If "scop" has any skip conditions, then no nested access parameters
3568 * are allowed. In particular, if there is any nested access in a guard
3569 * for a piece of code containing a "continue", then we want to introduce
3570 * a separate statement for evaluating this guard so that we can express
3571 * that the result is false for all previous iterations.
3573 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff *pa, pet_scop *scop)
3575 int nparam;
3577 if (!scop)
3578 return true;
3580 if (!pet_nested_any_in_pw_aff(pa))
3581 return true;
3583 if (pet_scop_has_skip(scop, pet_skip_now))
3584 return false;
3586 nparam = isl_pw_aff_dim(pa, isl_dim_param);
3587 for (int i = 0; i < nparam; ++i) {
3588 isl_id *id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
3589 pet_expr *expr;
3590 bool allowed;
3592 if (!pet_nested_in_id(id)) {
3593 isl_id_free(id);
3594 continue;
3597 expr = pet_nested_extract_expr(id);
3598 allowed = pet_expr_get_type(expr) == pet_expr_access &&
3599 !is_assigned(expr, scop);
3601 pet_expr_free(expr);
3602 isl_id_free(id);
3604 if (!allowed)
3605 return false;
3608 return true;
3611 /* Construct a pet_scop for a non-affine if statement.
3613 * We create a separate statement that writes the result
3614 * of the non-affine condition to a virtual scalar.
3615 * A constraint requiring the value of this virtual scalar to be one
3616 * is added to the iteration domains of the then branch.
3617 * Similarly, a constraint requiring the value of this virtual scalar
3618 * to be zero is added to the iteration domains of the else branch, if any.
3619 * We adjust the schedules to ensure that the virtual scalar is written
3620 * before it is read.
3622 * If there are any breaks or continues in the then and/or else
3623 * branches, then we may have to compute a new skip condition.
3624 * This is handled using a pet_skip_info object.
3625 * On initialization, the object checks if skip conditions need
3626 * to be computed. If so, it does so in pet_skip_info_if_extract_index and
3627 * adds them in pet_skip_info_if_add.
3629 struct pet_scop *PetScan::extract_non_affine_if(Expr *cond,
3630 struct pet_scop *scop_then, struct pet_scop *scop_else,
3631 bool have_else, int stmt_id)
3633 struct pet_scop *scop;
3634 isl_multi_pw_aff *test_index;
3635 int int_size;
3636 int save_n_stmt = n_stmt;
3638 test_index = pet_create_test_index(ctx, n_test++);
3639 n_stmt = stmt_id;
3640 scop = extract_non_affine_condition(cond, n_stmt++,
3641 isl_multi_pw_aff_copy(test_index));
3642 n_stmt = save_n_stmt;
3643 scop = scop_add_array(scop, test_index, ast_context);
3645 pet_skip_info skip;
3646 pet_skip_info_if_init(&skip, ctx, scop_then, scop_else, have_else, 0);
3647 int_size = ast_context.getTypeInfo(ast_context.IntTy).first / 8;
3648 pet_skip_info_if_extract_index(&skip, test_index, int_size,
3649 &n_stmt, &n_test);
3651 scop = pet_scop_prefix(scop, 0);
3652 scop_then = pet_scop_prefix(scop_then, 1);
3653 scop_then = pet_scop_filter(scop_then,
3654 isl_multi_pw_aff_copy(test_index), 1);
3655 if (have_else) {
3656 scop_else = pet_scop_prefix(scop_else, 1);
3657 scop_else = pet_scop_filter(scop_else, test_index, 0);
3658 scop_then = pet_scop_add_par(ctx, scop_then, scop_else);
3659 } else
3660 isl_multi_pw_aff_free(test_index);
3662 scop = pet_scop_add_seq(ctx, scop, scop_then);
3664 scop = pet_skip_info_if_add(&skip, scop, 2);
3666 return scop;
3669 /* Construct a pet_scop for an if statement.
3671 * If the condition fits the pattern of a conditional assignment,
3672 * then it is handled by extract_conditional_assignment.
3673 * Otherwise, we do the following.
3675 * If the condition is affine, then the condition is added
3676 * to the iteration domains of the then branch, while the
3677 * opposite of the condition in added to the iteration domains
3678 * of the else branch, if any.
3679 * We allow the condition to be dynamic, i.e., to refer to
3680 * scalars or array elements that may be written to outside
3681 * of the given if statement. These nested accesses are then represented
3682 * as output dimensions in the wrapping iteration domain.
3683 * If it is also written _inside_ the then or else branch, then
3684 * we treat the condition as non-affine.
3685 * As explained in extract_non_affine_if, this will introduce
3686 * an extra statement.
3687 * For aesthetic reasons, we want this statement to have a statement
3688 * number that is lower than those of the then and else branches.
3689 * In order to evaluate if we will need such a statement, however, we
3690 * first construct scops for the then and else branches.
3691 * We therefore reserve a statement number if we might have to
3692 * introduce such an extra statement.
3694 * If the condition is not affine, then the scop is created in
3695 * extract_non_affine_if.
3697 * If there are any breaks or continues in the then and/or else
3698 * branches, then we may have to compute a new skip condition.
3699 * This is handled using a pet_skip_info object.
3700 * On initialization, the object checks if skip conditions need
3701 * to be computed. If so, it does so in pet_skip_info_if_extract_cond and
3702 * adds them in pet_skip_info_if_add.
3704 struct pet_scop *PetScan::extract(IfStmt *stmt)
3706 struct pet_scop *scop_then, *scop_else = NULL, *scop;
3707 isl_pw_aff *cond;
3708 int stmt_id;
3709 int int_size;
3710 isl_set *set;
3711 isl_set *valid;
3713 clear_assignments clear(assigned_value);
3714 clear.TraverseStmt(stmt->getThen());
3715 if (stmt->getElse())
3716 clear.TraverseStmt(stmt->getElse());
3718 scop = extract_conditional_assignment(stmt);
3719 if (scop)
3720 return scop;
3722 cond = try_extract_nested_condition(stmt->getCond());
3723 if (allow_nested && (!cond || pet_nested_any_in_pw_aff(cond)))
3724 stmt_id = n_stmt++;
3727 assigned_value_cache cache(assigned_value);
3728 scop_then = extract(stmt->getThen());
3731 if (stmt->getElse()) {
3732 assigned_value_cache cache(assigned_value);
3733 scop_else = extract(stmt->getElse());
3734 if (options->autodetect) {
3735 if (scop_then && !scop_else) {
3736 partial = true;
3737 isl_pw_aff_free(cond);
3738 return scop_then;
3740 if (!scop_then && scop_else) {
3741 partial = true;
3742 isl_pw_aff_free(cond);
3743 return scop_else;
3748 if (cond &&
3749 (!is_nested_allowed(cond, scop_then) ||
3750 (stmt->getElse() && !is_nested_allowed(cond, scop_else)))) {
3751 isl_pw_aff_free(cond);
3752 cond = NULL;
3754 if (allow_nested && !cond)
3755 return extract_non_affine_if(stmt->getCond(), scop_then,
3756 scop_else, stmt->getElse(), stmt_id);
3758 if (!cond)
3759 cond = extract_condition(stmt->getCond());
3761 pet_skip_info skip;
3762 pet_skip_info_if_init(&skip, ctx, scop_then, scop_else,
3763 stmt->getElse() != NULL, 1);
3764 pet_skip_info_if_extract_cond(&skip, cond, int_size, &n_stmt, &n_test);
3766 valid = isl_pw_aff_domain(isl_pw_aff_copy(cond));
3767 set = isl_pw_aff_non_zero_set(cond);
3768 scop = pet_scop_restrict(scop_then, isl_set_params(isl_set_copy(set)));
3770 if (stmt->getElse()) {
3771 set = isl_set_subtract(isl_set_copy(valid), set);
3772 scop_else = pet_scop_restrict(scop_else, isl_set_params(set));
3773 scop = pet_scop_add_par(ctx, scop, scop_else);
3774 } else
3775 isl_set_free(set);
3776 scop = resolve_nested(scop);
3777 scop = pet_scop_restrict_context(scop, isl_set_params(valid));
3779 if (pet_skip_info_has_skip(&skip))
3780 scop = pet_scop_prefix(scop, 0);
3781 scop = pet_skip_info_if_add(&skip, scop, 1);
3783 return scop;
3786 /* Try and construct a pet_scop for a label statement.
3787 * We currently only allow labels on expression statements.
3789 struct pet_scop *PetScan::extract(LabelStmt *stmt)
3791 isl_id *label;
3792 Stmt *sub;
3794 sub = stmt->getSubStmt();
3795 if (!isa<Expr>(sub)) {
3796 unsupported(stmt);
3797 return NULL;
3800 label = isl_id_alloc(ctx, stmt->getName(), NULL);
3802 return extract(extract_expr(cast<Expr>(sub)), stmt->getSourceRange(),
3803 true, label);
3806 /* Return a one-dimensional multi piecewise affine expression that is equal
3807 * to the constant 1 and is defined over a zero-dimensional domain.
3809 static __isl_give isl_multi_pw_aff *one_mpa(isl_ctx *ctx)
3811 isl_space *space;
3812 isl_local_space *ls;
3813 isl_aff *aff;
3815 space = isl_space_set_alloc(ctx, 0, 0);
3816 ls = isl_local_space_from_space(space);
3817 aff = isl_aff_zero_on_domain(ls);
3818 aff = isl_aff_set_constant_si(aff, 1);
3820 return isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
3823 /* Construct a pet_scop for a continue statement.
3825 * We simply create an empty scop with a universal pet_skip_now
3826 * skip condition. This skip condition will then be taken into
3827 * account by the enclosing loop construct, possibly after
3828 * being incorporated into outer skip conditions.
3830 struct pet_scop *PetScan::extract(ContinueStmt *stmt)
3832 pet_scop *scop;
3834 scop = pet_scop_empty(ctx);
3835 if (!scop)
3836 return NULL;
3838 scop = pet_scop_set_skip(scop, pet_skip_now, one_mpa(ctx));
3840 return scop;
3843 /* Construct a pet_scop for a break statement.
3845 * We simply create an empty scop with both a universal pet_skip_now
3846 * skip condition and a universal pet_skip_later skip condition.
3847 * These skip conditions will then be taken into
3848 * account by the enclosing loop construct, possibly after
3849 * being incorporated into outer skip conditions.
3851 struct pet_scop *PetScan::extract(BreakStmt *stmt)
3853 pet_scop *scop;
3854 isl_multi_pw_aff *skip;
3856 scop = pet_scop_empty(ctx);
3857 if (!scop)
3858 return NULL;
3860 skip = one_mpa(ctx);
3861 scop = pet_scop_set_skip(scop, pet_skip_now,
3862 isl_multi_pw_aff_copy(skip));
3863 scop = pet_scop_set_skip(scop, pet_skip_later, skip);
3865 return scop;
3868 /* Try and construct a pet_scop corresponding to "stmt".
3870 * If "stmt" is a compound statement, then "skip_declarations"
3871 * indicates whether we should skip initial declarations in the
3872 * compound statement.
3874 * If the constructed pet_scop is not a (possibly) partial representation
3875 * of "stmt", we update start and end of the pet_scop to those of "stmt".
3876 * In particular, if skip_declarations is set, then we may have skipped
3877 * declarations inside "stmt" and so the pet_scop may not represent
3878 * the entire "stmt".
3879 * Note that this function may be called with "stmt" referring to the entire
3880 * body of the function, including the outer braces. In such cases,
3881 * skip_declarations will be set and the braces will not be taken into
3882 * account in scop->start and scop->end.
3884 struct pet_scop *PetScan::extract(Stmt *stmt, bool skip_declarations)
3886 struct pet_scop *scop;
3888 if (isa<Expr>(stmt))
3889 return extract(extract_expr(cast<Expr>(stmt)),
3890 stmt->getSourceRange(), true);
3892 switch (stmt->getStmtClass()) {
3893 case Stmt::WhileStmtClass:
3894 scop = extract(cast<WhileStmt>(stmt));
3895 break;
3896 case Stmt::ForStmtClass:
3897 scop = extract_for(cast<ForStmt>(stmt));
3898 break;
3899 case Stmt::IfStmtClass:
3900 scop = extract(cast<IfStmt>(stmt));
3901 break;
3902 case Stmt::CompoundStmtClass:
3903 scop = extract(cast<CompoundStmt>(stmt), skip_declarations);
3904 break;
3905 case Stmt::LabelStmtClass:
3906 scop = extract(cast<LabelStmt>(stmt));
3907 break;
3908 case Stmt::ContinueStmtClass:
3909 scop = extract(cast<ContinueStmt>(stmt));
3910 break;
3911 case Stmt::BreakStmtClass:
3912 scop = extract(cast<BreakStmt>(stmt));
3913 break;
3914 case Stmt::DeclStmtClass:
3915 scop = extract(cast<DeclStmt>(stmt));
3916 break;
3917 default:
3918 unsupported(stmt);
3919 return NULL;
3922 if (partial || skip_declarations)
3923 return scop;
3925 scop = update_scop_start_end(scop, stmt->getSourceRange(), false);
3927 return scop;
3930 /* Extract a clone of the kill statement in "scop".
3931 * "scop" is expected to have been created from a DeclStmt
3932 * and should have the kill as its first statement.
3934 struct pet_stmt *PetScan::extract_kill(struct pet_scop *scop)
3936 pet_expr *kill;
3937 struct pet_stmt *stmt;
3938 isl_multi_pw_aff *index;
3939 isl_map *access;
3940 pet_expr *arg;
3942 if (!scop)
3943 return NULL;
3944 if (scop->n_stmt < 1)
3945 isl_die(ctx, isl_error_internal,
3946 "expecting at least one statement", return NULL);
3947 stmt = scop->stmts[0];
3948 if (!pet_stmt_is_kill(stmt))
3949 isl_die(ctx, isl_error_internal,
3950 "expecting kill statement", return NULL);
3952 arg = pet_expr_get_arg(stmt->body, 0);
3953 index = pet_expr_access_get_index(arg);
3954 access = pet_expr_access_get_access(arg);
3955 pet_expr_free(arg);
3956 index = isl_multi_pw_aff_reset_tuple_id(index, isl_dim_in);
3957 access = isl_map_reset_tuple_id(access, isl_dim_in);
3958 kill = pet_expr_kill_from_access_and_index(access, index);
3959 return pet_stmt_from_pet_expr(stmt->line, NULL, n_stmt++, kill);
3962 /* Mark all arrays in "scop" as being exposed.
3964 static struct pet_scop *mark_exposed(struct pet_scop *scop)
3966 if (!scop)
3967 return NULL;
3968 for (int i = 0; i < scop->n_array; ++i)
3969 scop->arrays[i]->exposed = 1;
3970 return scop;
3973 /* Try and construct a pet_scop corresponding to (part of)
3974 * a sequence of statements.
3976 * "block" is set if the sequence respresents the children of
3977 * a compound statement.
3978 * "skip_declarations" is set if we should skip initial declarations
3979 * in the sequence of statements.
3981 * After extracting a statement, we update "assigned_value"
3982 * based on the top-level assignments in the statement
3983 * so that we can exploit them in subsequent statements in the same block.
3985 * If there are any breaks or continues in the individual statements,
3986 * then we may have to compute a new skip condition.
3987 * This is handled using a pet_skip_info object.
3988 * On initialization, the object checks if skip conditions need
3989 * to be computed. If so, it does so in pet_skip_info_seq_extract and
3990 * adds them in pet_skip_info_seq_add.
3992 * If "block" is set, then we need to insert kill statements at
3993 * the end of the block for any array that has been declared by
3994 * one of the statements in the sequence. Each of these declarations
3995 * results in the construction of a kill statement at the place
3996 * of the declaration, so we simply collect duplicates of
3997 * those kill statements and append these duplicates to the constructed scop.
3999 * If "block" is not set, then any array declared by one of the statements
4000 * in the sequence is marked as being exposed.
4002 * If autodetect is set, then we allow the extraction of only a subrange
4003 * of the sequence of statements. However, if there is at least one statement
4004 * for which we could not construct a scop and the final range contains
4005 * either no statements or at least one kill, then we discard the entire
4006 * range.
4008 struct pet_scop *PetScan::extract(StmtRange stmt_range, bool block,
4009 bool skip_declarations)
4011 pet_scop *scop;
4012 StmtIterator i;
4013 int int_size;
4014 int j;
4015 bool partial_range = false;
4016 set<struct pet_stmt *> kills;
4017 set<struct pet_stmt *>::iterator it;
4019 int_size = ast_context.getTypeInfo(ast_context.IntTy).first / 8;
4021 scop = pet_scop_empty(ctx);
4022 for (i = stmt_range.first, j = 0; i != stmt_range.second; ++i, ++j) {
4023 Stmt *child = *i;
4024 struct pet_scop *scop_i;
4026 if (scop->n_stmt == 0 && skip_declarations &&
4027 child->getStmtClass() == Stmt::DeclStmtClass)
4028 continue;
4030 scop_i = extract(child);
4031 if (scop->n_stmt != 0 && partial) {
4032 pet_scop_free(scop_i);
4033 break;
4035 handle_writes(scop_i);
4036 pet_skip_info skip;
4037 pet_skip_info_seq_init(&skip, ctx, scop, scop_i);
4038 pet_skip_info_seq_extract(&skip, int_size, &n_stmt, &n_test);
4039 if (pet_skip_info_has_skip(&skip))
4040 scop_i = pet_scop_prefix(scop_i, 0);
4041 if (scop_i && child->getStmtClass() == Stmt::DeclStmtClass) {
4042 if (block)
4043 kills.insert(extract_kill(scop_i));
4044 else
4045 scop_i = mark_exposed(scop_i);
4047 scop_i = pet_scop_prefix(scop_i, j);
4048 if (options->autodetect) {
4049 if (scop_i)
4050 scop = pet_scop_add_seq(ctx, scop, scop_i);
4051 else
4052 partial_range = true;
4053 if (scop->n_stmt != 0 && !scop_i)
4054 partial = true;
4055 } else {
4056 scop = pet_scop_add_seq(ctx, scop, scop_i);
4059 scop = pet_skip_info_seq_add(&skip, scop, j);
4061 if (partial || !scop)
4062 break;
4065 for (it = kills.begin(); it != kills.end(); ++it) {
4066 pet_scop *scop_j;
4067 scop_j = pet_scop_from_pet_stmt(ctx, *it);
4068 scop_j = pet_scop_prefix(scop_j, j);
4069 scop = pet_scop_add_seq(ctx, scop, scop_j);
4072 if (scop && partial_range) {
4073 if (scop->n_stmt == 0 || kills.size() != 0) {
4074 pet_scop_free(scop);
4075 return NULL;
4077 partial = true;
4080 return scop;
4083 /* Check if the scop marked by the user is exactly this Stmt
4084 * or part of this Stmt.
4085 * If so, return a pet_scop corresponding to the marked region.
4086 * Otherwise, return NULL.
4088 struct pet_scop *PetScan::scan(Stmt *stmt)
4090 SourceManager &SM = PP.getSourceManager();
4091 unsigned start_off, end_off;
4093 start_off = getExpansionOffset(SM, stmt->getLocStart());
4094 end_off = getExpansionOffset(SM, stmt->getLocEnd());
4096 if (start_off > loc.end)
4097 return NULL;
4098 if (end_off < loc.start)
4099 return NULL;
4100 if (start_off >= loc.start && end_off <= loc.end) {
4101 return extract(stmt);
4104 StmtIterator start;
4105 for (start = stmt->child_begin(); start != stmt->child_end(); ++start) {
4106 Stmt *child = *start;
4107 if (!child)
4108 continue;
4109 start_off = getExpansionOffset(SM, child->getLocStart());
4110 end_off = getExpansionOffset(SM, child->getLocEnd());
4111 if (start_off < loc.start && end_off >= loc.end)
4112 return scan(child);
4113 if (start_off >= loc.start)
4114 break;
4117 StmtIterator end;
4118 for (end = start; end != stmt->child_end(); ++end) {
4119 Stmt *child = *end;
4120 start_off = SM.getFileOffset(child->getLocStart());
4121 if (start_off >= loc.end)
4122 break;
4125 return extract(StmtRange(start, end), false, false);
4128 /* Set the size of index "pos" of "array" to "size".
4129 * In particular, add a constraint of the form
4131 * i_pos < size
4133 * to array->extent and a constraint of the form
4135 * size >= 0
4137 * to array->context.
4139 static struct pet_array *update_size(struct pet_array *array, int pos,
4140 __isl_take isl_pw_aff *size)
4142 isl_set *valid;
4143 isl_set *univ;
4144 isl_set *bound;
4145 isl_space *dim;
4146 isl_aff *aff;
4147 isl_pw_aff *index;
4148 isl_id *id;
4150 if (!array)
4151 goto error;
4153 valid = isl_set_params(isl_pw_aff_nonneg_set(isl_pw_aff_copy(size)));
4154 array->context = isl_set_intersect(array->context, valid);
4156 dim = isl_set_get_space(array->extent);
4157 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
4158 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, pos, 1);
4159 univ = isl_set_universe(isl_aff_get_domain_space(aff));
4160 index = isl_pw_aff_alloc(univ, aff);
4162 size = isl_pw_aff_add_dims(size, isl_dim_in,
4163 isl_set_dim(array->extent, isl_dim_set));
4164 id = isl_set_get_tuple_id(array->extent);
4165 size = isl_pw_aff_set_tuple_id(size, isl_dim_in, id);
4166 bound = isl_pw_aff_lt_set(index, size);
4168 array->extent = isl_set_intersect(array->extent, bound);
4170 if (!array->context || !array->extent)
4171 return pet_array_free(array);
4173 return array;
4174 error:
4175 isl_pw_aff_free(size);
4176 return NULL;
4179 /* Figure out the size of the array at position "pos" and all
4180 * subsequent positions from "type" and update the corresponding
4181 * argument of "expr" accordingly.
4183 __isl_give pet_expr *PetScan::set_upper_bounds(__isl_take pet_expr *expr,
4184 const Type *type, int pos)
4186 const ArrayType *atype;
4187 pet_expr *size;
4189 if (!expr)
4190 return NULL;
4192 if (type->isPointerType()) {
4193 type = type->getPointeeType().getTypePtr();
4194 return set_upper_bounds(expr, type, pos + 1);
4196 if (!type->isArrayType())
4197 return expr;
4199 type = type->getCanonicalTypeInternal().getTypePtr();
4200 atype = cast<ArrayType>(type);
4202 if (type->isConstantArrayType()) {
4203 const ConstantArrayType *ca = cast<ConstantArrayType>(atype);
4204 size = extract_expr(ca->getSize());
4205 expr = pet_expr_set_arg(expr, pos, size);
4206 } else if (type->isVariableArrayType()) {
4207 const VariableArrayType *vla = cast<VariableArrayType>(atype);
4208 size = extract_expr(vla->getSizeExpr());
4209 expr = pet_expr_set_arg(expr, pos, size);
4212 type = atype->getElementType().getTypePtr();
4214 return set_upper_bounds(expr, type, pos + 1);
4217 /* Does "expr" represent the "integer" infinity?
4219 static int is_infty(__isl_keep pet_expr *expr)
4221 isl_val *v;
4222 int res;
4224 if (pet_expr_get_type(expr) != pet_expr_int)
4225 return 0;
4226 v = pet_expr_int_get_val(expr);
4227 res = isl_val_is_infty(v);
4228 isl_val_free(v);
4230 return res;
4233 /* Figure out the dimensions of an array "array" based on its type
4234 * "type" and update "array" accordingly.
4236 * We first construct a pet_expr that holds the sizes of the array
4237 * in each dimension. The expression is initialized to infinity
4238 * and updated from the type.
4240 * The arguments of the size expression that have been updated
4241 * are then converted to an affine expression and incorporated
4242 * into the size of "array". If we are unable to convert
4243 * a size expression to an affine expression, then we leave
4244 * the corresponding size of "array" untouched.
4246 struct pet_array *PetScan::set_upper_bounds(struct pet_array *array,
4247 const Type *type)
4249 int depth = array_depth(type);
4250 pet_expr *expr, *inf;
4251 pet_context *pc;
4253 if (!array)
4254 return NULL;
4256 inf = pet_expr_new_int(isl_val_infty(ctx));
4257 expr = pet_expr_new_call(ctx, "bounds", depth);
4258 for (int i = 0; i < depth; ++i)
4259 expr = pet_expr_set_arg(expr, i, pet_expr_copy(inf));
4260 pet_expr_free(inf);
4262 expr = set_upper_bounds(expr, type, 0);
4264 pc = convert_assignments(ctx, assigned_value);
4265 for (int i = 0; i < depth; ++i) {
4266 pet_expr *arg;
4267 isl_pw_aff *size;
4269 arg = pet_expr_get_arg(expr, i);
4270 if (!is_infty(arg)) {
4271 size = pet_expr_extract_affine(arg, pc);
4272 if (!size)
4273 array = pet_array_free(array);
4274 else if (isl_pw_aff_involves_nan(size))
4275 isl_pw_aff_free(size);
4276 else
4277 array = update_size(array, i, size);
4279 pet_expr_free(arg);
4281 pet_expr_free(expr);
4282 pet_context_free(pc);
4284 return array;
4287 /* Is "T" the type of a variable length array with static size?
4289 static bool is_vla_with_static_size(QualType T)
4291 const VariableArrayType *vlatype;
4293 if (!T->isVariableArrayType())
4294 return false;
4295 vlatype = cast<VariableArrayType>(T);
4296 return vlatype->getSizeModifier() == VariableArrayType::Static;
4299 /* Return the type of "decl" as an array.
4301 * In particular, if "decl" is a parameter declaration that
4302 * is a variable length array with a static size, then
4303 * return the original type (i.e., the variable length array).
4304 * Otherwise, return the type of decl.
4306 static QualType get_array_type(ValueDecl *decl)
4308 ParmVarDecl *parm;
4309 QualType T;
4311 parm = dyn_cast<ParmVarDecl>(decl);
4312 if (!parm)
4313 return decl->getType();
4315 T = parm->getOriginalType();
4316 if (!is_vla_with_static_size(T))
4317 return decl->getType();
4318 return T;
4321 /* Does "decl" have definition that we can keep track of in a pet_type?
4323 static bool has_printable_definition(RecordDecl *decl)
4325 if (!decl->getDeclName())
4326 return false;
4327 return decl->getLexicalDeclContext() == decl->getDeclContext();
4330 /* Construct and return a pet_array corresponding to the variable "decl".
4331 * In particular, initialize array->extent to
4333 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
4335 * and then call set_upper_bounds to set the upper bounds on the indices
4336 * based on the type of the variable.
4338 * If the base type is that of a record with a top-level definition and
4339 * if "types" is not null, then the RecordDecl corresponding to the type
4340 * is added to "types".
4342 * If the base type is that of a record with no top-level definition,
4343 * then we replace it by "<subfield>".
4345 struct pet_array *PetScan::extract_array(isl_ctx *ctx, ValueDecl *decl,
4346 lex_recorddecl_set *types)
4348 struct pet_array *array;
4349 QualType qt = get_array_type(decl);
4350 const Type *type = qt.getTypePtr();
4351 int depth = array_depth(type);
4352 QualType base = pet_clang_base_type(qt);
4353 string name;
4354 isl_id *id;
4355 isl_space *dim;
4357 array = isl_calloc_type(ctx, struct pet_array);
4358 if (!array)
4359 return NULL;
4361 id = create_decl_id(ctx, decl);
4362 dim = isl_space_set_alloc(ctx, 0, depth);
4363 dim = isl_space_set_tuple_id(dim, isl_dim_set, id);
4365 array->extent = isl_set_nat_universe(dim);
4367 dim = isl_space_params_alloc(ctx, 0);
4368 array->context = isl_set_universe(dim);
4370 array = set_upper_bounds(array, type);
4371 if (!array)
4372 return NULL;
4374 name = base.getAsString();
4376 if (types && base->isRecordType()) {
4377 RecordDecl *decl = pet_clang_record_decl(base);
4378 if (has_printable_definition(decl))
4379 types->insert(decl);
4380 else
4381 name = "<subfield>";
4384 array->element_type = strdup(name.c_str());
4385 array->element_is_record = base->isRecordType();
4386 array->element_size = decl->getASTContext().getTypeInfo(base).first / 8;
4388 return array;
4391 /* Construct and return a pet_array corresponding to the sequence
4392 * of declarations "decls".
4393 * If the sequence contains a single declaration, then it corresponds
4394 * to a simple array access. Otherwise, it corresponds to a member access,
4395 * with the declaration for the substructure following that of the containing
4396 * structure in the sequence of declarations.
4397 * We start with the outermost substructure and then combine it with
4398 * information from the inner structures.
4400 * Additionally, keep track of all required types in "types".
4402 struct pet_array *PetScan::extract_array(isl_ctx *ctx,
4403 vector<ValueDecl *> decls, lex_recorddecl_set *types)
4405 struct pet_array *array;
4406 vector<ValueDecl *>::iterator it;
4408 it = decls.begin();
4410 array = extract_array(ctx, *it, types);
4412 for (++it; it != decls.end(); ++it) {
4413 struct pet_array *parent;
4414 const char *base_name, *field_name;
4415 char *product_name;
4417 parent = array;
4418 array = extract_array(ctx, *it, types);
4419 if (!array)
4420 return pet_array_free(parent);
4422 base_name = isl_set_get_tuple_name(parent->extent);
4423 field_name = isl_set_get_tuple_name(array->extent);
4424 product_name = pet_array_member_access_name(ctx,
4425 base_name, field_name);
4427 array->extent = isl_set_product(isl_set_copy(parent->extent),
4428 array->extent);
4429 if (product_name)
4430 array->extent = isl_set_set_tuple_name(array->extent,
4431 product_name);
4432 array->context = isl_set_intersect(array->context,
4433 isl_set_copy(parent->context));
4435 pet_array_free(parent);
4436 free(product_name);
4438 if (!array->extent || !array->context || !product_name)
4439 return pet_array_free(array);
4442 return array;
4445 /* Add a pet_type corresponding to "decl" to "scop, provided
4446 * it is a member of "types" and it has not been added before
4447 * (i.e., it is not a member of "types_done".
4449 * Since we want the user to be able to print the types
4450 * in the order in which they appear in the scop, we need to
4451 * make sure that types of fields in a structure appear before
4452 * that structure. We therefore call ourselves recursively
4453 * on the types of all record subfields.
4455 static struct pet_scop *add_type(isl_ctx *ctx, struct pet_scop *scop,
4456 RecordDecl *decl, Preprocessor &PP, lex_recorddecl_set &types,
4457 lex_recorddecl_set &types_done)
4459 string s;
4460 llvm::raw_string_ostream S(s);
4461 RecordDecl::field_iterator it;
4463 if (types.find(decl) == types.end())
4464 return scop;
4465 if (types_done.find(decl) != types_done.end())
4466 return scop;
4468 for (it = decl->field_begin(); it != decl->field_end(); ++it) {
4469 RecordDecl *record;
4470 QualType type = it->getType();
4472 if (!type->isRecordType())
4473 continue;
4474 record = pet_clang_record_decl(type);
4475 scop = add_type(ctx, scop, record, PP, types, types_done);
4478 if (strlen(decl->getName().str().c_str()) == 0)
4479 return scop;
4481 decl->print(S, PrintingPolicy(PP.getLangOpts()));
4482 S.str();
4484 scop->types[scop->n_type] = pet_type_alloc(ctx,
4485 decl->getName().str().c_str(), s.c_str());
4486 if (!scop->types[scop->n_type])
4487 return pet_scop_free(scop);
4489 types_done.insert(decl);
4491 scop->n_type++;
4493 return scop;
4496 /* Construct a list of pet_arrays, one for each array (or scalar)
4497 * accessed inside "scop", add this list to "scop" and return the result.
4499 * The context of "scop" is updated with the intersection of
4500 * the contexts of all arrays, i.e., constraints on the parameters
4501 * that ensure that the arrays have a valid (non-negative) size.
4503 * If the any of the extracted arrays refers to a member access,
4504 * then also add the required types to "scop".
4506 struct pet_scop *PetScan::scan_arrays(struct pet_scop *scop)
4508 int i;
4509 array_desc_set arrays;
4510 array_desc_set::iterator it;
4511 lex_recorddecl_set types;
4512 lex_recorddecl_set types_done;
4513 lex_recorddecl_set::iterator types_it;
4514 int n_array;
4515 struct pet_array **scop_arrays;
4517 if (!scop)
4518 return NULL;
4520 pet_scop_collect_arrays(scop, arrays);
4521 if (arrays.size() == 0)
4522 return scop;
4524 n_array = scop->n_array;
4526 scop_arrays = isl_realloc_array(ctx, scop->arrays, struct pet_array *,
4527 n_array + arrays.size());
4528 if (!scop_arrays)
4529 goto error;
4530 scop->arrays = scop_arrays;
4532 for (it = arrays.begin(), i = 0; it != arrays.end(); ++it, ++i) {
4533 struct pet_array *array;
4534 array = extract_array(ctx, *it, &types);
4535 scop->arrays[n_array + i] = array;
4536 if (!scop->arrays[n_array + i])
4537 goto error;
4538 scop->n_array++;
4539 scop->context = isl_set_intersect(scop->context,
4540 isl_set_copy(array->context));
4541 if (!scop->context)
4542 goto error;
4545 if (types.size() == 0)
4546 return scop;
4548 scop->types = isl_alloc_array(ctx, struct pet_type *, types.size());
4549 if (!scop->types)
4550 goto error;
4552 for (types_it = types.begin(); types_it != types.end(); ++types_it)
4553 scop = add_type(ctx, scop, *types_it, PP, types, types_done);
4555 return scop;
4556 error:
4557 pet_scop_free(scop);
4558 return NULL;
4561 /* Bound all parameters in scop->context to the possible values
4562 * of the corresponding C variable.
4564 static struct pet_scop *add_parameter_bounds(struct pet_scop *scop)
4566 int n;
4568 if (!scop)
4569 return NULL;
4571 n = isl_set_dim(scop->context, isl_dim_param);
4572 for (int i = 0; i < n; ++i) {
4573 isl_id *id;
4574 ValueDecl *decl;
4576 id = isl_set_get_dim_id(scop->context, isl_dim_param, i);
4577 if (pet_nested_in_id(id)) {
4578 isl_id_free(id);
4579 isl_die(isl_set_get_ctx(scop->context),
4580 isl_error_internal,
4581 "unresolved nested parameter", goto error);
4583 decl = (ValueDecl *) isl_id_get_user(id);
4584 isl_id_free(id);
4586 scop->context = set_parameter_bounds(scop->context, i, decl);
4588 if (!scop->context)
4589 goto error;
4592 return scop;
4593 error:
4594 pet_scop_free(scop);
4595 return NULL;
4598 /* Construct a pet_scop from the given function.
4600 * If the scop was delimited by scop and endscop pragmas, then we override
4601 * the file offsets by those derived from the pragmas.
4603 struct pet_scop *PetScan::scan(FunctionDecl *fd)
4605 pet_scop *scop;
4606 Stmt *stmt;
4608 stmt = fd->getBody();
4610 if (options->autodetect)
4611 scop = extract(stmt, true);
4612 else {
4613 scop = scan(stmt);
4614 scop = pet_scop_update_start_end(scop, loc.start, loc.end);
4616 scop = pet_scop_detect_parameter_accesses(scop);
4617 scop = scan_arrays(scop);
4618 scop = add_parameter_bounds(scop);
4619 scop = pet_scop_gist(scop, value_bounds);
4621 return scop;