extract out pet_extract_cst
[pet.git] / scan.cc
blob7e5bb53c4a3ab788ef23a1880557a8b0cb02d06c
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 "clang.h"
52 #include "expr.h"
53 #include "nest.h"
54 #include "options.h"
55 #include "scan.h"
56 #include "scop.h"
57 #include "scop_plus.h"
58 #include "skip.h"
60 #include "config.h"
62 using namespace std;
63 using namespace clang;
65 static enum pet_op_type UnaryOperatorKind2pet_op_type(UnaryOperatorKind kind)
67 switch (kind) {
68 case UO_Minus:
69 return pet_op_minus;
70 case UO_Not:
71 return pet_op_not;
72 case UO_LNot:
73 return pet_op_lnot;
74 case UO_PostInc:
75 return pet_op_post_inc;
76 case UO_PostDec:
77 return pet_op_post_dec;
78 case UO_PreInc:
79 return pet_op_pre_inc;
80 case UO_PreDec:
81 return pet_op_pre_dec;
82 default:
83 return pet_op_last;
87 static enum pet_op_type BinaryOperatorKind2pet_op_type(BinaryOperatorKind kind)
89 switch (kind) {
90 case BO_AddAssign:
91 return pet_op_add_assign;
92 case BO_SubAssign:
93 return pet_op_sub_assign;
94 case BO_MulAssign:
95 return pet_op_mul_assign;
96 case BO_DivAssign:
97 return pet_op_div_assign;
98 case BO_Assign:
99 return pet_op_assign;
100 case BO_Add:
101 return pet_op_add;
102 case BO_Sub:
103 return pet_op_sub;
104 case BO_Mul:
105 return pet_op_mul;
106 case BO_Div:
107 return pet_op_div;
108 case BO_Rem:
109 return pet_op_mod;
110 case BO_Shl:
111 return pet_op_shl;
112 case BO_Shr:
113 return pet_op_shr;
114 case BO_EQ:
115 return pet_op_eq;
116 case BO_NE:
117 return pet_op_ne;
118 case BO_LE:
119 return pet_op_le;
120 case BO_GE:
121 return pet_op_ge;
122 case BO_LT:
123 return pet_op_lt;
124 case BO_GT:
125 return pet_op_gt;
126 case BO_And:
127 return pet_op_and;
128 case BO_Xor:
129 return pet_op_xor;
130 case BO_Or:
131 return pet_op_or;
132 case BO_LAnd:
133 return pet_op_land;
134 case BO_LOr:
135 return pet_op_lor;
136 default:
137 return pet_op_last;
141 #if defined(DECLREFEXPR_CREATE_REQUIRES_BOOL)
142 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
144 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
145 SourceLocation(), var, false, var->getInnerLocStart(),
146 var->getType(), VK_LValue);
148 #elif defined(DECLREFEXPR_CREATE_REQUIRES_SOURCELOCATION)
149 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
151 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
152 SourceLocation(), var, var->getInnerLocStart(), var->getType(),
153 VK_LValue);
155 #else
156 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
158 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
159 var, var->getInnerLocStart(), var->getType(), VK_LValue);
161 #endif
163 /* Check if the element type corresponding to the given array type
164 * has a const qualifier.
166 static bool const_base(QualType qt)
168 const Type *type = qt.getTypePtr();
170 if (type->isPointerType())
171 return const_base(type->getPointeeType());
172 if (type->isArrayType()) {
173 const ArrayType *atype;
174 type = type->getCanonicalTypeInternal().getTypePtr();
175 atype = cast<ArrayType>(type);
176 return const_base(atype->getElementType());
179 return qt.isConstQualified();
182 /* Mark "decl" as having an unknown value in "assigned_value".
184 * If no (known or unknown) value was assigned to "decl" before,
185 * then it may have been treated as a parameter before and may
186 * therefore appear in a value assigned to another variable.
187 * If so, this assignment needs to be turned into an unknown value too.
189 static void clear_assignment(map<ValueDecl *, isl_pw_aff *> &assigned_value,
190 ValueDecl *decl)
192 map<ValueDecl *, isl_pw_aff *>::iterator it;
194 it = assigned_value.find(decl);
196 assigned_value[decl] = NULL;
198 if (it != assigned_value.end())
199 return;
201 for (it = assigned_value.begin(); it != assigned_value.end(); ++it) {
202 isl_pw_aff *pa = it->second;
203 int nparam = isl_pw_aff_dim(pa, isl_dim_param);
205 for (int i = 0; i < nparam; ++i) {
206 isl_id *id;
208 if (!isl_pw_aff_has_dim_id(pa, isl_dim_param, i))
209 continue;
210 id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
211 if (isl_id_get_user(id) == decl)
212 it->second = NULL;
213 isl_id_free(id);
218 /* Look for any assignments to scalar variables in part of the parse
219 * tree and set assigned_value to NULL for each of them.
220 * Also reset assigned_value if the address of a scalar variable
221 * is being taken. As an exception, if the address is passed to a function
222 * that is declared to receive a const pointer, then assigned_value is
223 * not reset.
225 * This ensures that we won't use any previously stored value
226 * in the current subtree and its parents.
228 struct clear_assignments : RecursiveASTVisitor<clear_assignments> {
229 map<ValueDecl *, isl_pw_aff *> &assigned_value;
230 set<UnaryOperator *> skip;
232 clear_assignments(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
233 assigned_value(assigned_value) {}
235 /* Check for "address of" operators whose value is passed
236 * to a const pointer argument and add them to "skip", so that
237 * we can skip them in VisitUnaryOperator.
239 bool VisitCallExpr(CallExpr *expr) {
240 FunctionDecl *fd;
241 fd = expr->getDirectCallee();
242 if (!fd)
243 return true;
244 for (int i = 0; i < expr->getNumArgs(); ++i) {
245 Expr *arg = expr->getArg(i);
246 UnaryOperator *op;
247 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
248 ImplicitCastExpr *ice;
249 ice = cast<ImplicitCastExpr>(arg);
250 arg = ice->getSubExpr();
252 if (arg->getStmtClass() != Stmt::UnaryOperatorClass)
253 continue;
254 op = cast<UnaryOperator>(arg);
255 if (op->getOpcode() != UO_AddrOf)
256 continue;
257 if (const_base(fd->getParamDecl(i)->getType()))
258 skip.insert(op);
260 return true;
263 bool VisitUnaryOperator(UnaryOperator *expr) {
264 Expr *arg;
265 DeclRefExpr *ref;
266 ValueDecl *decl;
268 switch (expr->getOpcode()) {
269 case UO_AddrOf:
270 case UO_PostInc:
271 case UO_PostDec:
272 case UO_PreInc:
273 case UO_PreDec:
274 break;
275 default:
276 return true;
278 if (skip.find(expr) != skip.end())
279 return true;
281 arg = expr->getSubExpr();
282 if (arg->getStmtClass() != Stmt::DeclRefExprClass)
283 return true;
284 ref = cast<DeclRefExpr>(arg);
285 decl = ref->getDecl();
286 clear_assignment(assigned_value, decl);
287 return true;
290 bool VisitBinaryOperator(BinaryOperator *expr) {
291 Expr *lhs;
292 DeclRefExpr *ref;
293 ValueDecl *decl;
295 if (!expr->isAssignmentOp())
296 return true;
297 lhs = expr->getLHS();
298 if (lhs->getStmtClass() != Stmt::DeclRefExprClass)
299 return true;
300 ref = cast<DeclRefExpr>(lhs);
301 decl = ref->getDecl();
302 clear_assignment(assigned_value, decl);
303 return true;
307 /* Keep a copy of the currently assigned values.
309 * Any variable that is assigned a value inside the current scope
310 * is removed again when we leave the scope (either because it wasn't
311 * stored in the cache or because it has a different value in the cache).
313 struct assigned_value_cache {
314 map<ValueDecl *, isl_pw_aff *> &assigned_value;
315 map<ValueDecl *, isl_pw_aff *> cache;
317 assigned_value_cache(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
318 assigned_value(assigned_value), cache(assigned_value) {}
319 ~assigned_value_cache() {
320 map<ValueDecl *, isl_pw_aff *>::iterator it = cache.begin();
321 for (it = assigned_value.begin(); it != assigned_value.end();
322 ++it) {
323 if (!it->second ||
324 (cache.find(it->first) != cache.end() &&
325 cache[it->first] != it->second))
326 cache[it->first] = NULL;
328 assigned_value = cache;
332 /* Insert an expression into the collection of expressions,
333 * provided it is not already in there.
334 * The isl_pw_affs are freed in the destructor.
336 void PetScan::insert_expression(__isl_take isl_pw_aff *expr)
338 std::set<isl_pw_aff *>::iterator it;
340 if (expressions.find(expr) == expressions.end())
341 expressions.insert(expr);
342 else
343 isl_pw_aff_free(expr);
346 PetScan::~PetScan()
348 std::set<isl_pw_aff *>::iterator it;
350 for (it = expressions.begin(); it != expressions.end(); ++it)
351 isl_pw_aff_free(*it);
353 isl_union_map_free(value_bounds);
356 /* Report a diagnostic, unless autodetect is set.
358 void PetScan::report(Stmt *stmt, unsigned id)
360 if (options->autodetect)
361 return;
363 SourceLocation loc = stmt->getLocStart();
364 DiagnosticsEngine &diag = PP.getDiagnostics();
365 DiagnosticBuilder B = diag.Report(loc, id) << stmt->getSourceRange();
368 /* Called if we found something we (currently) cannot handle.
369 * We'll provide more informative warnings later.
371 * We only actually complain if autodetect is false.
373 void PetScan::unsupported(Stmt *stmt)
375 DiagnosticsEngine &diag = PP.getDiagnostics();
376 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
377 "unsupported");
378 report(stmt, id);
381 /* Report a missing prototype, unless autodetect is set.
383 void PetScan::report_prototype_required(Stmt *stmt)
385 DiagnosticsEngine &diag = PP.getDiagnostics();
386 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
387 "prototype required");
388 report(stmt, id);
391 /* Report a missing increment, unless autodetect is set.
393 void PetScan::report_missing_increment(Stmt *stmt)
395 DiagnosticsEngine &diag = PP.getDiagnostics();
396 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
397 "missing increment");
398 report(stmt, id);
401 /* Extract an integer from "expr".
403 __isl_give isl_val *PetScan::extract_int(isl_ctx *ctx, IntegerLiteral *expr)
405 const Type *type = expr->getType().getTypePtr();
406 int is_signed = type->hasSignedIntegerRepresentation();
407 llvm::APInt val = expr->getValue();
408 int is_negative = is_signed && val.isNegative();
409 isl_val *v;
411 if (is_negative)
412 val = -val;
414 v = extract_unsigned(ctx, val);
416 if (is_negative)
417 v = isl_val_neg(v);
418 return v;
421 /* Extract an integer from "val", which is assumed to be non-negative.
423 __isl_give isl_val *PetScan::extract_unsigned(isl_ctx *ctx,
424 const llvm::APInt &val)
426 unsigned n;
427 const uint64_t *data;
429 data = val.getRawData();
430 n = val.getNumWords();
431 return isl_val_int_from_chunks(ctx, n, sizeof(uint64_t), data);
434 /* Extract an integer from "expr".
435 * Return NULL if "expr" does not (obviously) represent an integer.
437 __isl_give isl_val *PetScan::extract_int(clang::ParenExpr *expr)
439 return extract_int(expr->getSubExpr());
442 /* Extract an integer from "expr".
443 * Return NULL if "expr" does not (obviously) represent an integer.
445 __isl_give isl_val *PetScan::extract_int(clang::Expr *expr)
447 if (expr->getStmtClass() == Stmt::IntegerLiteralClass)
448 return extract_int(ctx, cast<IntegerLiteral>(expr));
449 if (expr->getStmtClass() == Stmt::ParenExprClass)
450 return extract_int(cast<ParenExpr>(expr));
452 unsupported(expr);
453 return NULL;
456 /* Extract an affine expression from the IntegerLiteral "expr".
458 __isl_give isl_pw_aff *PetScan::extract_affine(IntegerLiteral *expr)
460 isl_space *dim = isl_space_params_alloc(ctx, 0);
461 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
462 isl_aff *aff = isl_aff_zero_on_domain(ls);
463 isl_set *dom = isl_set_universe(dim);
464 isl_val *v;
466 v = extract_int(expr);
467 aff = isl_aff_add_constant_val(aff, v);
469 return isl_pw_aff_alloc(dom, aff);
472 /* Extract an affine expression from the APInt "val", which is assumed
473 * to be non-negative.
475 __isl_give isl_pw_aff *PetScan::extract_affine(const llvm::APInt &val)
477 isl_space *dim = isl_space_params_alloc(ctx, 0);
478 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
479 isl_aff *aff = isl_aff_zero_on_domain(ls);
480 isl_set *dom = isl_set_universe(dim);
481 isl_val *v;
483 v = extract_unsigned(ctx, val);
484 aff = isl_aff_add_constant_val(aff, v);
486 return isl_pw_aff_alloc(dom, aff);
489 __isl_give isl_pw_aff *PetScan::extract_affine(ImplicitCastExpr *expr)
491 return extract_affine(expr->getSubExpr());
494 /* Return the number of bits needed to represent the type "qt",
495 * if it is an integer type. Otherwise return 0.
496 * If qt is signed then return the opposite of the number of bits.
498 static int get_type_size(QualType qt, ASTContext &ast_context)
500 int size;
502 if (!qt->isIntegerType())
503 return 0;
505 size = ast_context.getIntWidth(qt);
506 if (!qt->isUnsignedIntegerType())
507 size = -size;
509 return size;
512 /* Return the number of bits needed to represent the type of "decl",
513 * if it is an integer type. Otherwise return 0.
514 * If qt is signed then return the opposite of the number of bits.
516 static int get_type_size(ValueDecl *decl)
518 return get_type_size(decl->getType(), decl->getASTContext());
521 /* Bound parameter "pos" of "set" to the possible values of "decl".
523 static __isl_give isl_set *set_parameter_bounds(__isl_take isl_set *set,
524 unsigned pos, ValueDecl *decl)
526 int type_size;
527 isl_ctx *ctx;
528 isl_val *bound;
530 ctx = isl_set_get_ctx(set);
531 type_size = get_type_size(decl);
532 if (type_size == 0)
533 isl_die(ctx, isl_error_invalid, "not an integer type",
534 return isl_set_free(set));
535 if (type_size > 0) {
536 set = isl_set_lower_bound_si(set, isl_dim_param, pos, 0);
537 bound = isl_val_int_from_ui(ctx, type_size);
538 bound = isl_val_2exp(bound);
539 bound = isl_val_sub_ui(bound, 1);
540 set = isl_set_upper_bound_val(set, isl_dim_param, pos, bound);
541 } else {
542 bound = isl_val_int_from_ui(ctx, -type_size - 1);
543 bound = isl_val_2exp(bound);
544 bound = isl_val_sub_ui(bound, 1);
545 set = isl_set_upper_bound_val(set, isl_dim_param, pos,
546 isl_val_copy(bound));
547 bound = isl_val_neg(bound);
548 bound = isl_val_sub_ui(bound, 1);
549 set = isl_set_lower_bound_val(set, isl_dim_param, pos, bound);
552 return set;
555 /* Extract an affine expression from the DeclRefExpr "expr".
557 * If the variable has been assigned a value, then we check whether
558 * we know what (affine) value was assigned.
559 * If so, we return this value. Otherwise we convert "expr"
560 * to an extra parameter (provided nesting_enabled is set).
562 * Otherwise, we simply return an expression that is equal
563 * to a parameter corresponding to the referenced variable.
565 __isl_give isl_pw_aff *PetScan::extract_affine(DeclRefExpr *expr)
567 ValueDecl *decl = expr->getDecl();
568 const Type *type = decl->getType().getTypePtr();
569 isl_id *id;
570 isl_space *dim;
571 isl_aff *aff;
572 isl_set *dom;
574 if (!type->isIntegerType()) {
575 unsupported(expr);
576 return NULL;
579 if (assigned_value.find(decl) != assigned_value.end()) {
580 if (assigned_value[decl])
581 return isl_pw_aff_copy(assigned_value[decl]);
582 else
583 return nested_access(expr);
586 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
587 dim = isl_space_params_alloc(ctx, 1);
589 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
591 dom = isl_set_universe(isl_space_copy(dim));
592 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
593 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
595 return isl_pw_aff_alloc(dom, aff);
598 /* Extract an affine expression from an integer division operation.
599 * In particular, if "expr" is lhs/rhs, then return
601 * lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs)
603 * The second argument (rhs) is required to be a (positive) integer constant.
605 __isl_give isl_pw_aff *PetScan::extract_affine_div(BinaryOperator *expr)
607 int is_cst;
608 isl_pw_aff *rhs, *lhs;
610 rhs = extract_affine(expr->getRHS());
611 is_cst = isl_pw_aff_is_cst(rhs);
612 if (is_cst < 0 || !is_cst) {
613 isl_pw_aff_free(rhs);
614 if (!is_cst)
615 unsupported(expr);
616 return NULL;
619 lhs = extract_affine(expr->getLHS());
621 return isl_pw_aff_tdiv_q(lhs, rhs);
624 /* Extract an affine expression from a modulo operation.
625 * In particular, if "expr" is lhs/rhs, then return
627 * lhs - rhs * (lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs))
629 * The second argument (rhs) is required to be a (positive) integer constant.
631 __isl_give isl_pw_aff *PetScan::extract_affine_mod(BinaryOperator *expr)
633 int is_cst;
634 isl_pw_aff *rhs, *lhs;
636 rhs = extract_affine(expr->getRHS());
637 is_cst = isl_pw_aff_is_cst(rhs);
638 if (is_cst < 0 || !is_cst) {
639 isl_pw_aff_free(rhs);
640 if (!is_cst)
641 unsupported(expr);
642 return NULL;
645 lhs = extract_affine(expr->getLHS());
647 return isl_pw_aff_tdiv_r(lhs, rhs);
650 /* Extract an affine expression from a multiplication operation.
651 * This is only allowed if at least one of the two arguments
652 * is a (piecewise) constant.
654 __isl_give isl_pw_aff *PetScan::extract_affine_mul(BinaryOperator *expr)
656 isl_pw_aff *lhs;
657 isl_pw_aff *rhs;
659 lhs = extract_affine(expr->getLHS());
660 rhs = extract_affine(expr->getRHS());
662 if (!isl_pw_aff_is_cst(lhs) && !isl_pw_aff_is_cst(rhs)) {
663 isl_pw_aff_free(lhs);
664 isl_pw_aff_free(rhs);
665 unsupported(expr);
666 return NULL;
669 return isl_pw_aff_mul(lhs, rhs);
672 /* Extract an affine expression from an addition or subtraction operation.
674 __isl_give isl_pw_aff *PetScan::extract_affine_add(BinaryOperator *expr)
676 isl_pw_aff *lhs;
677 isl_pw_aff *rhs;
679 lhs = extract_affine(expr->getLHS());
680 rhs = extract_affine(expr->getRHS());
682 switch (expr->getOpcode()) {
683 case BO_Add:
684 return isl_pw_aff_add(lhs, rhs);
685 case BO_Sub:
686 return isl_pw_aff_sub(lhs, rhs);
687 default:
688 isl_pw_aff_free(lhs);
689 isl_pw_aff_free(rhs);
690 return NULL;
695 /* Compute
697 * pwaff mod 2^width
699 static __isl_give isl_pw_aff *wrap(__isl_take isl_pw_aff *pwaff,
700 unsigned width)
702 isl_ctx *ctx;
703 isl_val *mod;
705 ctx = isl_pw_aff_get_ctx(pwaff);
706 mod = isl_val_int_from_ui(ctx, width);
707 mod = isl_val_2exp(mod);
709 pwaff = isl_pw_aff_mod_val(pwaff, mod);
711 return pwaff;
714 /* Limit the domain of "pwaff" to those elements where the function
715 * value satisfies
717 * 2^{width-1} <= pwaff < 2^{width-1}
719 static __isl_give isl_pw_aff *avoid_overflow(__isl_take isl_pw_aff *pwaff,
720 unsigned width)
722 isl_ctx *ctx;
723 isl_val *v;
724 isl_space *space = isl_pw_aff_get_domain_space(pwaff);
725 isl_local_space *ls = isl_local_space_from_space(space);
726 isl_aff *bound;
727 isl_set *dom;
728 isl_pw_aff *b;
730 ctx = isl_pw_aff_get_ctx(pwaff);
731 v = isl_val_int_from_ui(ctx, width - 1);
732 v = isl_val_2exp(v);
734 bound = isl_aff_zero_on_domain(ls);
735 bound = isl_aff_add_constant_val(bound, v);
736 b = isl_pw_aff_from_aff(bound);
738 dom = isl_pw_aff_lt_set(isl_pw_aff_copy(pwaff), isl_pw_aff_copy(b));
739 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
741 b = isl_pw_aff_neg(b);
742 dom = isl_pw_aff_ge_set(isl_pw_aff_copy(pwaff), b);
743 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
745 return pwaff;
748 /* Handle potential overflows on signed computations.
750 * If options->signed_overflow is set to PET_OVERFLOW_AVOID,
751 * the we adjust the domain of "pa" to avoid overflows.
753 __isl_give isl_pw_aff *PetScan::signed_overflow(__isl_take isl_pw_aff *pa,
754 unsigned width)
756 if (options->signed_overflow == PET_OVERFLOW_AVOID)
757 pa = avoid_overflow(pa, width);
759 return pa;
762 /* Return the piecewise affine expression "set ? 1 : 0" defined on "dom".
764 static __isl_give isl_pw_aff *indicator_function(__isl_take isl_set *set,
765 __isl_take isl_set *dom)
767 isl_pw_aff *pa;
768 pa = isl_set_indicator_function(set);
769 pa = isl_pw_aff_intersect_domain(pa, isl_set_coalesce(dom));
770 return pa;
773 /* Extract an affine expression from some binary operations.
774 * If the result of the expression is unsigned, then we wrap it
775 * based on the size of the type. Otherwise, we ensure that
776 * no overflow occurs.
778 __isl_give isl_pw_aff *PetScan::extract_affine(BinaryOperator *expr)
780 isl_pw_aff *res;
781 unsigned width;
783 switch (expr->getOpcode()) {
784 case BO_Add:
785 case BO_Sub:
786 res = extract_affine_add(expr);
787 break;
788 case BO_Div:
789 res = extract_affine_div(expr);
790 break;
791 case BO_Rem:
792 res = extract_affine_mod(expr);
793 break;
794 case BO_Mul:
795 res = extract_affine_mul(expr);
796 break;
797 case BO_LT:
798 case BO_LE:
799 case BO_GT:
800 case BO_GE:
801 case BO_EQ:
802 case BO_NE:
803 case BO_LAnd:
804 case BO_LOr:
805 return extract_condition(expr);
806 default:
807 unsupported(expr);
808 return NULL;
811 width = ast_context.getIntWidth(expr->getType());
812 if (expr->getType()->isUnsignedIntegerType())
813 res = wrap(res, width);
814 else
815 res = signed_overflow(res, width);
817 return res;
820 /* Extract an affine expression from a negation operation.
822 __isl_give isl_pw_aff *PetScan::extract_affine(UnaryOperator *expr)
824 if (expr->getOpcode() == UO_Minus)
825 return isl_pw_aff_neg(extract_affine(expr->getSubExpr()));
826 if (expr->getOpcode() == UO_LNot)
827 return extract_condition(expr);
829 unsupported(expr);
830 return NULL;
833 __isl_give isl_pw_aff *PetScan::extract_affine(ParenExpr *expr)
835 return extract_affine(expr->getSubExpr());
838 /* Extract an affine expression from some special function calls.
839 * In particular, we handle "min", "max", "ceild", "floord",
840 * "intMod", "intFloor" and "intCeil".
841 * In case of the latter five, the second argument needs to be
842 * a (positive) integer constant.
844 __isl_give isl_pw_aff *PetScan::extract_affine(CallExpr *expr)
846 FunctionDecl *fd;
847 string name;
848 isl_pw_aff *aff1, *aff2;
850 fd = expr->getDirectCallee();
851 if (!fd) {
852 unsupported(expr);
853 return NULL;
856 name = fd->getDeclName().getAsString();
857 if (!(expr->getNumArgs() == 2 && name == "min") &&
858 !(expr->getNumArgs() == 2 && name == "max") &&
859 !(expr->getNumArgs() == 2 && name == "intMod") &&
860 !(expr->getNumArgs() == 2 && name == "intFloor") &&
861 !(expr->getNumArgs() == 2 && name == "intCeil") &&
862 !(expr->getNumArgs() == 2 && name == "floord") &&
863 !(expr->getNumArgs() == 2 && name == "ceild")) {
864 unsupported(expr);
865 return NULL;
868 if (name == "min" || name == "max") {
869 aff1 = extract_affine(expr->getArg(0));
870 aff2 = extract_affine(expr->getArg(1));
872 if (name == "min")
873 aff1 = isl_pw_aff_min(aff1, aff2);
874 else
875 aff1 = isl_pw_aff_max(aff1, aff2);
876 } else if (name == "intMod") {
877 isl_val *v;
878 Expr *arg2 = expr->getArg(1);
880 if (arg2->getStmtClass() != Stmt::IntegerLiteralClass) {
881 unsupported(expr);
882 return NULL;
884 aff1 = extract_affine(expr->getArg(0));
885 v = extract_int(cast<IntegerLiteral>(arg2));
886 aff1 = isl_pw_aff_mod_val(aff1, v);
887 } else if (name == "floord" || name == "ceild" ||
888 name == "intFloor" || name == "intCeil") {
889 isl_val *v;
890 Expr *arg2 = expr->getArg(1);
892 if (arg2->getStmtClass() != Stmt::IntegerLiteralClass) {
893 unsupported(expr);
894 return NULL;
896 aff1 = extract_affine(expr->getArg(0));
897 v = extract_int(cast<IntegerLiteral>(arg2));
898 aff1 = isl_pw_aff_scale_down_val(aff1, v);
899 if (name == "floord" || name == "intFloor")
900 aff1 = isl_pw_aff_floor(aff1);
901 else
902 aff1 = isl_pw_aff_ceil(aff1);
903 } else {
904 unsupported(expr);
905 return NULL;
908 return aff1;
911 /* This method is called when we come across an access that is
912 * nested in what is supposed to be an affine expression.
913 * If nesting is allowed, we return a new parameter that corresponds
914 * to this nested access. Otherwise, we simply complain.
916 * Note that we currently don't allow nested accesses themselves
917 * to contain any nested accesses, so we check if we can extract
918 * the access without any nesting and complain if we can't.
920 * The new parameter is resolved in resolve_nested.
922 isl_pw_aff *PetScan::nested_access(Expr *expr)
924 isl_id *id;
925 isl_space *dim;
926 isl_aff *aff;
927 isl_set *dom;
928 isl_multi_pw_aff *index;
930 if (!nesting_enabled) {
931 unsupported(expr);
932 return NULL;
935 allow_nested = false;
936 index = extract_index(expr);
937 allow_nested = true;
938 if (!index) {
939 unsupported(expr);
940 return NULL;
942 isl_multi_pw_aff_free(index);
944 id = pet_nested_clang_expr(ctx, expr);
945 dim = isl_space_params_alloc(ctx, 1);
947 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
949 dom = isl_set_universe(isl_space_copy(dim));
950 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
951 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
953 return isl_pw_aff_alloc(dom, aff);
956 /* Affine expressions are not supposed to contain array accesses,
957 * but if nesting is allowed, we return a parameter corresponding
958 * to the array access.
960 __isl_give isl_pw_aff *PetScan::extract_affine(ArraySubscriptExpr *expr)
962 return nested_access(expr);
965 /* Affine expressions are not supposed to contain member accesses,
966 * but if nesting is allowed, we return a parameter corresponding
967 * to the member access.
969 __isl_give isl_pw_aff *PetScan::extract_affine(MemberExpr *expr)
971 return nested_access(expr);
974 /* Extract an affine expression from a conditional operation.
976 __isl_give isl_pw_aff *PetScan::extract_affine(ConditionalOperator *expr)
978 isl_pw_aff *cond, *lhs, *rhs;
980 cond = extract_condition(expr->getCond());
981 lhs = extract_affine(expr->getTrueExpr());
982 rhs = extract_affine(expr->getFalseExpr());
984 return isl_pw_aff_cond(cond, lhs, rhs);
987 /* Extract an affine expression, if possible, from "expr".
988 * Otherwise return NULL.
990 __isl_give isl_pw_aff *PetScan::extract_affine(Expr *expr)
992 switch (expr->getStmtClass()) {
993 case Stmt::ImplicitCastExprClass:
994 return extract_affine(cast<ImplicitCastExpr>(expr));
995 case Stmt::IntegerLiteralClass:
996 return extract_affine(cast<IntegerLiteral>(expr));
997 case Stmt::DeclRefExprClass:
998 return extract_affine(cast<DeclRefExpr>(expr));
999 case Stmt::BinaryOperatorClass:
1000 return extract_affine(cast<BinaryOperator>(expr));
1001 case Stmt::UnaryOperatorClass:
1002 return extract_affine(cast<UnaryOperator>(expr));
1003 case Stmt::ParenExprClass:
1004 return extract_affine(cast<ParenExpr>(expr));
1005 case Stmt::CallExprClass:
1006 return extract_affine(cast<CallExpr>(expr));
1007 case Stmt::ArraySubscriptExprClass:
1008 return extract_affine(cast<ArraySubscriptExpr>(expr));
1009 case Stmt::MemberExprClass:
1010 return extract_affine(cast<MemberExpr>(expr));
1011 case Stmt::ConditionalOperatorClass:
1012 return extract_affine(cast<ConditionalOperator>(expr));
1013 default:
1014 unsupported(expr);
1016 return NULL;
1019 __isl_give isl_multi_pw_aff *PetScan::extract_index(ImplicitCastExpr *expr)
1021 return extract_index(expr->getSubExpr());
1024 /* Return the depth of an array of the given type.
1026 static int array_depth(const Type *type)
1028 if (type->isPointerType())
1029 return 1 + array_depth(type->getPointeeType().getTypePtr());
1030 if (type->isArrayType()) {
1031 const ArrayType *atype;
1032 type = type->getCanonicalTypeInternal().getTypePtr();
1033 atype = cast<ArrayType>(type);
1034 return 1 + array_depth(atype->getElementType().getTypePtr());
1036 return 0;
1039 /* Return the depth of the array accessed by the index expression "index".
1040 * If "index" is an affine expression, i.e., if it does not access
1041 * any array, then return 1.
1042 * If "index" represent a member access, i.e., if its range is a wrapped
1043 * relation, then return the sum of the depth of the array of structures
1044 * and that of the member inside the structure.
1046 static int extract_depth(__isl_keep isl_multi_pw_aff *index)
1048 isl_id *id;
1049 ValueDecl *decl;
1051 if (!index)
1052 return -1;
1054 if (isl_multi_pw_aff_range_is_wrapping(index)) {
1055 int domain_depth, range_depth;
1056 isl_multi_pw_aff *domain, *range;
1058 domain = isl_multi_pw_aff_copy(index);
1059 domain = isl_multi_pw_aff_range_factor_domain(domain);
1060 domain_depth = extract_depth(domain);
1061 isl_multi_pw_aff_free(domain);
1062 range = isl_multi_pw_aff_copy(index);
1063 range = isl_multi_pw_aff_range_factor_range(range);
1064 range_depth = extract_depth(range);
1065 isl_multi_pw_aff_free(range);
1067 return domain_depth + range_depth;
1070 if (!isl_multi_pw_aff_has_tuple_id(index, isl_dim_out))
1071 return 1;
1073 id = isl_multi_pw_aff_get_tuple_id(index, isl_dim_out);
1074 if (!id)
1075 return -1;
1076 decl = (ValueDecl *) isl_id_get_user(id);
1077 isl_id_free(id);
1079 return array_depth(decl->getType().getTypePtr());
1082 /* Extract an index expression from a reference to a variable.
1083 * If the variable has name "A", then the returned index expression
1084 * is of the form
1086 * { [] -> A[] }
1088 __isl_give isl_multi_pw_aff *PetScan::extract_index(DeclRefExpr *expr)
1090 return extract_index(expr->getDecl());
1093 /* Extract an index expression from a variable.
1094 * If the variable has name "A", then the returned index expression
1095 * is of the form
1097 * { [] -> A[] }
1099 __isl_give isl_multi_pw_aff *PetScan::extract_index(ValueDecl *decl)
1101 isl_id *id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
1102 isl_space *space = isl_space_alloc(ctx, 0, 0, 0);
1104 space = isl_space_set_tuple_id(space, isl_dim_out, id);
1106 return isl_multi_pw_aff_zero(space);
1109 /* Extract an index expression from an integer contant.
1110 * If the value of the constant is "v", then the returned access relation
1111 * is
1113 * { [] -> [v] }
1115 __isl_give isl_multi_pw_aff *PetScan::extract_index(IntegerLiteral *expr)
1117 isl_multi_pw_aff *mpa;
1119 mpa = isl_multi_pw_aff_from_pw_aff(extract_affine(expr));
1120 mpa = isl_multi_pw_aff_from_range(mpa);
1121 return mpa;
1124 /* Try and extract an index expression from the given Expr.
1125 * Return NULL if it doesn't work out.
1127 __isl_give isl_multi_pw_aff *PetScan::extract_index(Expr *expr)
1129 switch (expr->getStmtClass()) {
1130 case Stmt::ImplicitCastExprClass:
1131 return extract_index(cast<ImplicitCastExpr>(expr));
1132 case Stmt::DeclRefExprClass:
1133 return extract_index(cast<DeclRefExpr>(expr));
1134 case Stmt::ArraySubscriptExprClass:
1135 return extract_index(cast<ArraySubscriptExpr>(expr));
1136 case Stmt::IntegerLiteralClass:
1137 return extract_index(cast<IntegerLiteral>(expr));
1138 case Stmt::MemberExprClass:
1139 return extract_index(cast<MemberExpr>(expr));
1140 default:
1141 unsupported(expr);
1143 return NULL;
1146 /* Given a partial index expression "base" and an extra index "index",
1147 * append the extra index to "base" and return the result.
1148 * Additionally, add the constraints that the extra index is non-negative.
1149 * If "index" represent a member access, i.e., if its range is a wrapped
1150 * relation, then we recursively extend the range of this nested relation.
1152 static __isl_give isl_multi_pw_aff *subscript(__isl_take isl_multi_pw_aff *base,
1153 __isl_take isl_pw_aff *index)
1155 isl_id *id;
1156 isl_set *domain;
1157 isl_multi_pw_aff *access;
1158 int member_access;
1160 member_access = isl_multi_pw_aff_range_is_wrapping(base);
1161 if (member_access < 0)
1162 goto error;
1163 if (member_access) {
1164 isl_multi_pw_aff *domain, *range;
1165 isl_id *id;
1167 id = isl_multi_pw_aff_get_tuple_id(base, isl_dim_out);
1168 domain = isl_multi_pw_aff_copy(base);
1169 domain = isl_multi_pw_aff_range_factor_domain(domain);
1170 range = isl_multi_pw_aff_range_factor_range(base);
1171 range = subscript(range, index);
1172 access = isl_multi_pw_aff_range_product(domain, range);
1173 access = isl_multi_pw_aff_set_tuple_id(access, isl_dim_out, id);
1174 return access;
1177 id = isl_multi_pw_aff_get_tuple_id(base, isl_dim_set);
1178 index = isl_pw_aff_from_range(index);
1179 domain = isl_pw_aff_nonneg_set(isl_pw_aff_copy(index));
1180 index = isl_pw_aff_intersect_domain(index, domain);
1181 access = isl_multi_pw_aff_from_pw_aff(index);
1182 access = isl_multi_pw_aff_flat_range_product(base, access);
1183 access = isl_multi_pw_aff_set_tuple_id(access, isl_dim_set, id);
1185 return access;
1186 error:
1187 isl_multi_pw_aff_free(base);
1188 isl_pw_aff_free(index);
1189 return NULL;
1192 /* Extract an index expression from the given array subscript expression.
1193 * If nesting is allowed in general, then we turn it on while
1194 * examining the index expression.
1196 * We first extract an index expression from the base.
1197 * This will result in an index expression with a range that corresponds
1198 * to the earlier indices.
1199 * We then extract the current index, restrict its domain
1200 * to those values that result in a non-negative index and
1201 * append the index to the base index expression.
1203 __isl_give isl_multi_pw_aff *PetScan::extract_index(ArraySubscriptExpr *expr)
1205 Expr *base = expr->getBase();
1206 Expr *idx = expr->getIdx();
1207 isl_pw_aff *index;
1208 isl_multi_pw_aff *base_access;
1209 isl_multi_pw_aff *access;
1210 bool save_nesting = nesting_enabled;
1212 nesting_enabled = allow_nested;
1214 base_access = extract_index(base);
1215 index = extract_affine(idx);
1217 nesting_enabled = save_nesting;
1219 access = subscript(base_access, index);
1221 return access;
1224 /* Construct a name for a member access by concatenating the name
1225 * of the array of structures and the member, separated by an underscore.
1227 * The caller is responsible for freeing the result.
1229 static char *member_access_name(isl_ctx *ctx, const char *base,
1230 const char *field)
1232 int len;
1233 char *name;
1235 len = strlen(base) + 1 + strlen(field);
1236 name = isl_alloc_array(ctx, char, len + 1);
1237 if (!name)
1238 return NULL;
1239 snprintf(name, len + 1, "%s_%s", base, field);
1241 return name;
1244 /* Given an index expression "base" for an element of an array of structures
1245 * and an expression "field" for the field member being accessed, construct
1246 * an index expression for an access to that member of the given structure.
1247 * In particular, take the range product of "base" and "field" and
1248 * attach a name to the result.
1250 static __isl_give isl_multi_pw_aff *member(__isl_take isl_multi_pw_aff *base,
1251 __isl_take isl_multi_pw_aff *field)
1253 isl_ctx *ctx;
1254 isl_multi_pw_aff *access;
1255 const char *base_name, *field_name;
1256 char *name;
1258 ctx = isl_multi_pw_aff_get_ctx(base);
1260 base_name = isl_multi_pw_aff_get_tuple_name(base, isl_dim_out);
1261 field_name = isl_multi_pw_aff_get_tuple_name(field, isl_dim_out);
1262 name = member_access_name(ctx, base_name, field_name);
1264 access = isl_multi_pw_aff_range_product(base, field);
1266 access = isl_multi_pw_aff_set_tuple_name(access, isl_dim_out, name);
1267 free(name);
1269 return access;
1272 /* Extract an index expression from a member expression.
1274 * If the base access (to the structure containing the member)
1275 * is of the form
1277 * [] -> A[..]
1279 * and the member is called "f", then the member access is of
1280 * the form
1282 * [] -> A_f[A[..] -> f[]]
1284 * If the member access is to an anonymous struct, then simply return
1286 * [] -> A[..]
1288 * If the member access in the source code is of the form
1290 * A->f
1292 * then it is treated as
1294 * A[0].f
1296 __isl_give isl_multi_pw_aff *PetScan::extract_index(MemberExpr *expr)
1298 Expr *base = expr->getBase();
1299 FieldDecl *field = cast<FieldDecl>(expr->getMemberDecl());
1300 isl_multi_pw_aff *base_access, *field_access;
1301 isl_id *id;
1302 isl_space *space;
1304 base_access = extract_index(base);
1306 if (expr->isArrow()) {
1307 isl_space *space = isl_space_params_alloc(ctx, 0);
1308 isl_local_space *ls = isl_local_space_from_space(space);
1309 isl_aff *aff = isl_aff_zero_on_domain(ls);
1310 isl_pw_aff *index = isl_pw_aff_from_aff(aff);
1311 base_access = subscript(base_access, index);
1314 if (field->isAnonymousStructOrUnion())
1315 return base_access;
1317 id = isl_id_alloc(ctx, field->getName().str().c_str(), field);
1318 space = isl_multi_pw_aff_get_domain_space(base_access);
1319 space = isl_space_from_domain(space);
1320 space = isl_space_set_tuple_id(space, isl_dim_out, id);
1321 field_access = isl_multi_pw_aff_zero(space);
1323 return member(base_access, field_access);
1326 /* Check if "expr" calls function "minmax" with two arguments and if so
1327 * make lhs and rhs refer to these two arguments.
1329 static bool is_minmax(Expr *expr, const char *minmax, Expr *&lhs, Expr *&rhs)
1331 CallExpr *call;
1332 FunctionDecl *fd;
1333 string name;
1335 if (expr->getStmtClass() != Stmt::CallExprClass)
1336 return false;
1338 call = cast<CallExpr>(expr);
1339 fd = call->getDirectCallee();
1340 if (!fd)
1341 return false;
1343 if (call->getNumArgs() != 2)
1344 return false;
1346 name = fd->getDeclName().getAsString();
1347 if (name != minmax)
1348 return false;
1350 lhs = call->getArg(0);
1351 rhs = call->getArg(1);
1353 return true;
1356 /* Check if "expr" is of the form min(lhs, rhs) and if so make
1357 * lhs and rhs refer to the two arguments.
1359 static bool is_min(Expr *expr, Expr *&lhs, Expr *&rhs)
1361 return is_minmax(expr, "min", lhs, rhs);
1364 /* Check if "expr" is of the form max(lhs, rhs) and if so make
1365 * lhs and rhs refer to the two arguments.
1367 static bool is_max(Expr *expr, Expr *&lhs, Expr *&rhs)
1369 return is_minmax(expr, "max", lhs, rhs);
1372 /* Extract an affine expressions representing the comparison "LHS op RHS"
1373 * "comp" is the original statement that "LHS op RHS" is derived from
1374 * and is used for diagnostics.
1376 * If the comparison is of the form
1378 * a <= min(b,c)
1380 * then the expression is constructed as the conjunction of
1381 * the comparisons
1383 * a <= b and a <= c
1385 * A similar optimization is performed for max(a,b) <= c.
1386 * We do this because that will lead to simpler representations
1387 * of the expression.
1388 * If isl is ever enhanced to explicitly deal with min and max expressions,
1389 * this optimization can be removed.
1391 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperatorKind op,
1392 Expr *LHS, Expr *RHS, Stmt *comp)
1394 isl_pw_aff *lhs;
1395 isl_pw_aff *rhs;
1396 isl_pw_aff *res;
1397 isl_set *cond;
1398 isl_set *dom;
1399 enum pet_op_type type;
1401 if (op == BO_GT)
1402 return extract_comparison(BO_LT, RHS, LHS, comp);
1403 if (op == BO_GE)
1404 return extract_comparison(BO_LE, RHS, LHS, comp);
1406 if (op == BO_LT || op == BO_LE) {
1407 Expr *expr1, *expr2;
1408 if (is_min(RHS, expr1, expr2)) {
1409 lhs = extract_comparison(op, LHS, expr1, comp);
1410 rhs = extract_comparison(op, LHS, expr2, comp);
1411 return pet_and(lhs, rhs);
1413 if (is_max(LHS, expr1, expr2)) {
1414 lhs = extract_comparison(op, expr1, RHS, comp);
1415 rhs = extract_comparison(op, expr2, RHS, comp);
1416 return pet_and(lhs, rhs);
1420 lhs = extract_affine(LHS);
1421 rhs = extract_affine(RHS);
1423 type = BinaryOperatorKind2pet_op_type(op);
1424 return pet_comparison(type, lhs, rhs);
1427 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperator *comp)
1429 return extract_comparison(comp->getOpcode(), comp->getLHS(),
1430 comp->getRHS(), comp);
1433 /* Extract an affine expression representing the negation (logical not)
1434 * of a subexpression.
1436 __isl_give isl_pw_aff *PetScan::extract_boolean(UnaryOperator *op)
1438 isl_pw_aff *cond;
1440 cond = extract_condition(op->getSubExpr());
1441 return pet_not(cond);
1444 /* Extract an affine expression representing the disjunction (logical or)
1445 * or conjunction (logical and) of two subexpressions.
1447 __isl_give isl_pw_aff *PetScan::extract_boolean(BinaryOperator *comp)
1449 isl_pw_aff *lhs, *rhs;
1451 lhs = extract_condition(comp->getLHS());
1452 rhs = extract_condition(comp->getRHS());
1454 switch (comp->getOpcode()) {
1455 case BO_LAnd:
1456 return pet_boolean(pet_op_land, lhs, rhs);
1457 case BO_LOr:
1458 return pet_boolean(pet_op_lor, lhs, rhs);
1459 default:
1460 isl_pw_aff_free(lhs);
1461 isl_pw_aff_free(rhs);
1464 unsupported(comp);
1465 return NULL;
1468 __isl_give isl_pw_aff *PetScan::extract_condition(UnaryOperator *expr)
1470 switch (expr->getOpcode()) {
1471 case UO_LNot:
1472 return extract_boolean(expr);
1473 default:
1474 unsupported(expr);
1475 return NULL;
1479 /* Extract the affine expression "expr != 0 ? 1 : 0".
1481 __isl_give isl_pw_aff *PetScan::extract_implicit_condition(Expr *expr)
1483 isl_pw_aff *res;
1485 res = extract_affine(expr);
1486 return pet_to_bool(res);
1489 /* Extract an affine expression from a boolean expression.
1490 * In particular, return the expression "expr ? 1 : 0".
1492 * If the expression doesn't look like a condition, we assume it
1493 * is an affine expression and return the condition "expr != 0 ? 1 : 0".
1495 __isl_give isl_pw_aff *PetScan::extract_condition(Expr *expr)
1497 BinaryOperator *comp;
1499 if (!expr) {
1500 isl_set *u = isl_set_universe(isl_space_params_alloc(ctx, 0));
1501 return indicator_function(u, isl_set_copy(u));
1504 if (expr->getStmtClass() == Stmt::ParenExprClass)
1505 return extract_condition(cast<ParenExpr>(expr)->getSubExpr());
1507 if (expr->getStmtClass() == Stmt::UnaryOperatorClass)
1508 return extract_condition(cast<UnaryOperator>(expr));
1510 if (expr->getStmtClass() != Stmt::BinaryOperatorClass)
1511 return extract_implicit_condition(expr);
1513 comp = cast<BinaryOperator>(expr);
1514 switch (comp->getOpcode()) {
1515 case BO_LT:
1516 case BO_LE:
1517 case BO_GT:
1518 case BO_GE:
1519 case BO_EQ:
1520 case BO_NE:
1521 return extract_comparison(comp);
1522 case BO_LAnd:
1523 case BO_LOr:
1524 return extract_boolean(comp);
1525 default:
1526 return extract_implicit_condition(expr);
1530 /* Construct a pet_expr representing a unary operator expression.
1532 __isl_give pet_expr *PetScan::extract_expr(UnaryOperator *expr)
1534 pet_expr *arg;
1535 enum pet_op_type op;
1537 op = UnaryOperatorKind2pet_op_type(expr->getOpcode());
1538 if (op == pet_op_last) {
1539 unsupported(expr);
1540 return NULL;
1543 arg = extract_expr(expr->getSubExpr());
1545 if (expr->isIncrementDecrementOp() &&
1546 pet_expr_get_type(arg) == pet_expr_access) {
1547 arg = mark_write(arg);
1548 arg = pet_expr_access_set_read(arg, 1);
1551 return pet_expr_new_unary(op, arg);
1554 /* Mark the given access pet_expr as a write.
1555 * If a scalar is being accessed, then mark its value
1556 * as unknown in assigned_value.
1558 __isl_give pet_expr *PetScan::mark_write(__isl_take pet_expr *access)
1560 isl_id *id;
1561 ValueDecl *decl;
1563 access = pet_expr_access_set_write(access, 1);
1564 access = pet_expr_access_set_read(access, 0);
1566 if (!access || !pet_expr_is_scalar_access(access))
1567 return access;
1569 id = pet_expr_access_get_id(access);
1570 decl = (ValueDecl *) isl_id_get_user(id);
1571 clear_assignment(assigned_value, decl);
1572 isl_id_free(id);
1574 return access;
1577 /* Assign "rhs" to "lhs".
1579 * In particular, if "lhs" is a scalar variable, then mark
1580 * the variable as having been assigned. If, furthermore, "rhs"
1581 * is an affine expression, then keep track of this value in assigned_value
1582 * so that we can plug it in when we later come across the same variable.
1584 void PetScan::assign(__isl_keep pet_expr *lhs, Expr *rhs)
1586 isl_id *id;
1587 ValueDecl *decl;
1588 isl_pw_aff *pa;
1590 if (!lhs)
1591 return;
1592 if (!pet_expr_is_scalar_access(lhs))
1593 return;
1595 id = pet_expr_access_get_id(lhs);
1596 decl = (ValueDecl *) isl_id_get_user(id);
1597 isl_id_free(id);
1599 pa = try_extract_affine(rhs);
1600 clear_assignment(assigned_value, decl);
1601 if (!pa)
1602 return;
1603 assigned_value[decl] = pa;
1604 insert_expression(pa);
1607 /* Construct a pet_expr representing a binary operator expression.
1609 * If the top level operator is an assignment and the LHS is an access,
1610 * then we mark that access as a write. If the operator is a compound
1611 * assignment, the access is marked as both a read and a write.
1613 * If "expr" assigns something to a scalar variable, then we mark
1614 * the variable as having been assigned. If, furthermore, the expression
1615 * is affine, then keep track of this value in assigned_value
1616 * so that we can plug it in when we later come across the same variable.
1618 __isl_give pet_expr *PetScan::extract_expr(BinaryOperator *expr)
1620 int type_size;
1621 pet_expr *lhs, *rhs;
1622 enum pet_op_type op;
1624 op = BinaryOperatorKind2pet_op_type(expr->getOpcode());
1625 if (op == pet_op_last) {
1626 unsupported(expr);
1627 return NULL;
1630 lhs = extract_expr(expr->getLHS());
1631 rhs = extract_expr(expr->getRHS());
1633 if (expr->isAssignmentOp() &&
1634 pet_expr_get_type(lhs) == pet_expr_access) {
1635 lhs = mark_write(lhs);
1636 if (expr->isCompoundAssignmentOp())
1637 lhs = pet_expr_access_set_read(lhs, 1);
1640 if (expr->getOpcode() == BO_Assign)
1641 assign(lhs, expr->getRHS());
1643 type_size = get_type_size(expr->getType(), ast_context);
1644 return pet_expr_new_binary(type_size, op, lhs, rhs);
1647 /* Construct a pet_scop with a single statement killing the entire
1648 * array "array".
1650 struct pet_scop *PetScan::kill(Stmt *stmt, struct pet_array *array)
1652 isl_id *id;
1653 isl_space *space;
1654 isl_multi_pw_aff *index;
1655 isl_map *access;
1656 pet_expr *expr;
1658 if (!array)
1659 return NULL;
1660 access = isl_map_from_range(isl_set_copy(array->extent));
1661 id = isl_set_get_tuple_id(array->extent);
1662 space = isl_space_alloc(ctx, 0, 0, 0);
1663 space = isl_space_set_tuple_id(space, isl_dim_out, id);
1664 index = isl_multi_pw_aff_zero(space);
1665 expr = pet_expr_kill_from_access_and_index(access, index);
1666 return extract(stmt, expr);
1669 /* Construct a pet_scop for a (single) variable declaration.
1671 * The scop contains the variable being declared (as an array)
1672 * and a statement killing the array.
1674 * If the variable is initialized in the AST, then the scop
1675 * also contains an assignment to the variable.
1677 struct pet_scop *PetScan::extract(DeclStmt *stmt)
1679 int type_size;
1680 Decl *decl;
1681 VarDecl *vd;
1682 pet_expr *lhs, *rhs, *pe;
1683 struct pet_scop *scop_decl, *scop;
1684 struct pet_array *array;
1686 if (!stmt->isSingleDecl()) {
1687 unsupported(stmt);
1688 return NULL;
1691 decl = stmt->getSingleDecl();
1692 vd = cast<VarDecl>(decl);
1694 array = extract_array(ctx, vd, NULL);
1695 if (array)
1696 array->declared = 1;
1697 scop_decl = kill(stmt, array);
1698 scop_decl = pet_scop_add_array(scop_decl, array);
1700 if (!vd->getInit())
1701 return scop_decl;
1703 lhs = extract_access_expr(vd);
1704 rhs = extract_expr(vd->getInit());
1706 lhs = mark_write(lhs);
1707 assign(lhs, vd->getInit());
1709 type_size = get_type_size(vd->getType(), ast_context);
1710 pe = pet_expr_new_binary(type_size, pet_op_assign, lhs, rhs);
1711 scop = extract(stmt, pe);
1713 scop_decl = pet_scop_prefix(scop_decl, 0);
1714 scop = pet_scop_prefix(scop, 1);
1716 scop = pet_scop_add_seq(ctx, scop_decl, scop);
1718 return scop;
1721 /* Construct a pet_expr representing a conditional operation.
1723 * We first try to extract the condition as an affine expression.
1724 * If that fails, we construct a pet_expr tree representing the condition.
1726 __isl_give pet_expr *PetScan::extract_expr(ConditionalOperator *expr)
1728 pet_expr *cond, *lhs, *rhs;
1729 isl_pw_aff *pa;
1731 pa = try_extract_affine(expr->getCond());
1732 if (pa) {
1733 isl_multi_pw_aff *test = isl_multi_pw_aff_from_pw_aff(pa);
1734 test = isl_multi_pw_aff_from_range(test);
1735 cond = pet_expr_from_index(test);
1736 } else
1737 cond = extract_expr(expr->getCond());
1738 lhs = extract_expr(expr->getTrueExpr());
1739 rhs = extract_expr(expr->getFalseExpr());
1741 return pet_expr_new_ternary(cond, lhs, rhs);
1744 __isl_give pet_expr *PetScan::extract_expr(ImplicitCastExpr *expr)
1746 return extract_expr(expr->getSubExpr());
1749 /* Construct a pet_expr representing a floating point value.
1751 * If the floating point literal does not appear in a macro,
1752 * then we use the original representation in the source code
1753 * as the string representation. Otherwise, we use the pretty
1754 * printer to produce a string representation.
1756 __isl_give pet_expr *PetScan::extract_expr(FloatingLiteral *expr)
1758 double d;
1759 string s;
1760 const LangOptions &LO = PP.getLangOpts();
1761 SourceLocation loc = expr->getLocation();
1763 if (!loc.isMacroID()) {
1764 SourceManager &SM = PP.getSourceManager();
1765 unsigned len = Lexer::MeasureTokenLength(loc, SM, LO);
1766 s = string(SM.getCharacterData(loc), len);
1767 } else {
1768 llvm::raw_string_ostream S(s);
1769 expr->printPretty(S, 0, PrintingPolicy(LO));
1770 S.str();
1772 d = expr->getValueAsApproximateDouble();
1773 return pet_expr_new_double(ctx, d, s.c_str());
1776 /* Convert the index expression "index" into an access pet_expr of type "qt".
1778 __isl_give pet_expr *PetScan::extract_access_expr(QualType qt,
1779 __isl_take isl_multi_pw_aff *index)
1781 pet_expr *pe;
1782 int depth;
1783 int type_size;
1785 depth = extract_depth(index);
1786 type_size = get_type_size(qt, ast_context);
1788 pe = pet_expr_from_index_and_depth(type_size, index, depth);
1790 return pe;
1793 /* Extract an index expression from "expr" and then convert it into
1794 * an access pet_expr.
1796 __isl_give pet_expr *PetScan::extract_access_expr(Expr *expr)
1798 return extract_access_expr(expr->getType(), extract_index(expr));
1801 /* Extract an index expression from "decl" and then convert it into
1802 * an access pet_expr.
1804 __isl_give pet_expr *PetScan::extract_access_expr(ValueDecl *decl)
1806 return extract_access_expr(decl->getType(), extract_index(decl));
1809 __isl_give pet_expr *PetScan::extract_expr(ParenExpr *expr)
1811 return extract_expr(expr->getSubExpr());
1814 /* Extract an assume statement from the argument "expr"
1815 * of a __pencil_assume statement.
1817 __isl_give pet_expr *PetScan::extract_assume(Expr *expr)
1819 isl_pw_aff *cond;
1820 pet_expr *res;
1822 cond = try_extract_affine_condition(expr);
1823 if (!cond) {
1824 res = extract_expr(expr);
1825 } else {
1826 isl_multi_pw_aff *index;
1827 index = isl_multi_pw_aff_from_pw_aff(cond);
1828 index = isl_multi_pw_aff_from_range(index);
1829 res = pet_expr_from_index(index);
1831 return pet_expr_new_unary(pet_op_assume, res);
1834 /* Construct a pet_expr corresponding to the function call argument "expr".
1835 * The argument appears in position "pos" of a call to function "fd".
1837 * If we are passing along a pointer to an array element
1838 * or an entire row or even higher dimensional slice of an array,
1839 * then the function being called may write into the array.
1841 * We assume here that if the function is declared to take a pointer
1842 * to a const type, then the function will perform a read
1843 * and that otherwise, it will perform a write.
1845 __isl_give pet_expr *PetScan::extract_argument(FunctionDecl *fd, int pos,
1846 Expr *expr)
1848 pet_expr *res;
1849 int is_addr = 0, is_partial = 0;
1850 Stmt::StmtClass sc;
1852 if (expr->getStmtClass() == Stmt::ImplicitCastExprClass) {
1853 ImplicitCastExpr *ice = cast<ImplicitCastExpr>(expr);
1854 expr = ice->getSubExpr();
1856 if (expr->getStmtClass() == Stmt::UnaryOperatorClass) {
1857 UnaryOperator *op = cast<UnaryOperator>(expr);
1858 if (op->getOpcode() == UO_AddrOf) {
1859 is_addr = 1;
1860 expr = op->getSubExpr();
1863 res = extract_expr(expr);
1864 if (!res)
1865 return NULL;
1866 sc = expr->getStmtClass();
1867 if ((sc == Stmt::ArraySubscriptExprClass ||
1868 sc == Stmt::MemberExprClass) &&
1869 array_depth(expr->getType().getTypePtr()) > 0)
1870 is_partial = 1;
1871 if ((is_addr || is_partial) &&
1872 pet_expr_get_type(res) == pet_expr_access) {
1873 ParmVarDecl *parm;
1874 if (!fd->hasPrototype()) {
1875 report_prototype_required(expr);
1876 return pet_expr_free(res);
1878 parm = fd->getParamDecl(pos);
1879 if (!const_base(parm->getType()))
1880 res = mark_write(res);
1883 if (is_addr)
1884 res = pet_expr_new_unary(pet_op_address_of, res);
1885 return res;
1888 /* Construct a pet_expr representing a function call.
1890 * In the special case of a "call" to __pencil_assume,
1891 * construct an assume expression instead.
1893 __isl_give pet_expr *PetScan::extract_expr(CallExpr *expr)
1895 pet_expr *res = NULL;
1896 FunctionDecl *fd;
1897 string name;
1898 unsigned n_arg;
1900 fd = expr->getDirectCallee();
1901 if (!fd) {
1902 unsupported(expr);
1903 return NULL;
1906 name = fd->getDeclName().getAsString();
1907 n_arg = expr->getNumArgs();
1909 if (n_arg == 1 && name == "__pencil_assume")
1910 return extract_assume(expr->getArg(0));
1912 res = pet_expr_new_call(ctx, name.c_str(), n_arg);
1913 if (!res)
1914 return NULL;
1916 for (int i = 0; i < n_arg; ++i) {
1917 Expr *arg = expr->getArg(i);
1918 res = pet_expr_set_arg(res, i,
1919 PetScan::extract_argument(fd, i, arg));
1922 return res;
1925 /* Construct a pet_expr representing a (C style) cast.
1927 __isl_give pet_expr *PetScan::extract_expr(CStyleCastExpr *expr)
1929 pet_expr *arg;
1930 QualType type;
1932 arg = extract_expr(expr->getSubExpr());
1933 if (!arg)
1934 return NULL;
1936 type = expr->getTypeAsWritten();
1937 return pet_expr_new_cast(type.getAsString().c_str(), arg);
1940 /* Construct a pet_expr representing an integer.
1942 __isl_give pet_expr *PetScan::extract_expr(IntegerLiteral *expr)
1944 return pet_expr_new_int(extract_int(expr));
1947 /* Try and construct a pet_expr representing "expr".
1949 __isl_give pet_expr *PetScan::extract_expr(Expr *expr)
1951 switch (expr->getStmtClass()) {
1952 case Stmt::UnaryOperatorClass:
1953 return extract_expr(cast<UnaryOperator>(expr));
1954 case Stmt::CompoundAssignOperatorClass:
1955 case Stmt::BinaryOperatorClass:
1956 return extract_expr(cast<BinaryOperator>(expr));
1957 case Stmt::ImplicitCastExprClass:
1958 return extract_expr(cast<ImplicitCastExpr>(expr));
1959 case Stmt::ArraySubscriptExprClass:
1960 case Stmt::DeclRefExprClass:
1961 case Stmt::MemberExprClass:
1962 return extract_access_expr(expr);
1963 case Stmt::IntegerLiteralClass:
1964 return extract_expr(cast<IntegerLiteral>(expr));
1965 case Stmt::FloatingLiteralClass:
1966 return extract_expr(cast<FloatingLiteral>(expr));
1967 case Stmt::ParenExprClass:
1968 return extract_expr(cast<ParenExpr>(expr));
1969 case Stmt::ConditionalOperatorClass:
1970 return extract_expr(cast<ConditionalOperator>(expr));
1971 case Stmt::CallExprClass:
1972 return extract_expr(cast<CallExpr>(expr));
1973 case Stmt::CStyleCastExprClass:
1974 return extract_expr(cast<CStyleCastExpr>(expr));
1975 default:
1976 unsupported(expr);
1978 return NULL;
1981 /* Check if the given initialization statement is an assignment.
1982 * If so, return that assignment. Otherwise return NULL.
1984 BinaryOperator *PetScan::initialization_assignment(Stmt *init)
1986 BinaryOperator *ass;
1988 if (init->getStmtClass() != Stmt::BinaryOperatorClass)
1989 return NULL;
1991 ass = cast<BinaryOperator>(init);
1992 if (ass->getOpcode() != BO_Assign)
1993 return NULL;
1995 return ass;
1998 /* Check if the given initialization statement is a declaration
1999 * of a single variable.
2000 * If so, return that declaration. Otherwise return NULL.
2002 Decl *PetScan::initialization_declaration(Stmt *init)
2004 DeclStmt *decl;
2006 if (init->getStmtClass() != Stmt::DeclStmtClass)
2007 return NULL;
2009 decl = cast<DeclStmt>(init);
2011 if (!decl->isSingleDecl())
2012 return NULL;
2014 return decl->getSingleDecl();
2017 /* Given the assignment operator in the initialization of a for loop,
2018 * extract the induction variable, i.e., the (integer)variable being
2019 * assigned.
2021 ValueDecl *PetScan::extract_induction_variable(BinaryOperator *init)
2023 Expr *lhs;
2024 DeclRefExpr *ref;
2025 ValueDecl *decl;
2026 const Type *type;
2028 lhs = init->getLHS();
2029 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
2030 unsupported(init);
2031 return NULL;
2034 ref = cast<DeclRefExpr>(lhs);
2035 decl = ref->getDecl();
2036 type = decl->getType().getTypePtr();
2038 if (!type->isIntegerType()) {
2039 unsupported(lhs);
2040 return NULL;
2043 return decl;
2046 /* Given the initialization statement of a for loop and the single
2047 * declaration in this initialization statement,
2048 * extract the induction variable, i.e., the (integer) variable being
2049 * declared.
2051 VarDecl *PetScan::extract_induction_variable(Stmt *init, Decl *decl)
2053 VarDecl *vd;
2055 vd = cast<VarDecl>(decl);
2057 const QualType type = vd->getType();
2058 if (!type->isIntegerType()) {
2059 unsupported(init);
2060 return NULL;
2063 if (!vd->getInit()) {
2064 unsupported(init);
2065 return NULL;
2068 return vd;
2071 /* Check that op is of the form iv++ or iv--.
2072 * Return an affine expression "1" or "-1" accordingly.
2074 __isl_give isl_pw_aff *PetScan::extract_unary_increment(
2075 clang::UnaryOperator *op, clang::ValueDecl *iv)
2077 Expr *sub;
2078 DeclRefExpr *ref;
2079 isl_space *space;
2080 isl_aff *aff;
2082 if (!op->isIncrementDecrementOp()) {
2083 unsupported(op);
2084 return NULL;
2087 sub = op->getSubExpr();
2088 if (sub->getStmtClass() != Stmt::DeclRefExprClass) {
2089 unsupported(op);
2090 return NULL;
2093 ref = cast<DeclRefExpr>(sub);
2094 if (ref->getDecl() != iv) {
2095 unsupported(op);
2096 return NULL;
2099 space = isl_space_params_alloc(ctx, 0);
2100 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2102 if (op->isIncrementOp())
2103 aff = isl_aff_add_constant_si(aff, 1);
2104 else
2105 aff = isl_aff_add_constant_si(aff, -1);
2107 return isl_pw_aff_from_aff(aff);
2110 /* Check if op is of the form
2112 * iv = iv + inc
2114 * and return inc as an affine expression.
2116 * We extract an affine expression from the RHS, subtract iv and return
2117 * the result.
2119 __isl_give isl_pw_aff *PetScan::extract_binary_increment(BinaryOperator *op,
2120 clang::ValueDecl *iv)
2122 Expr *lhs;
2123 DeclRefExpr *ref;
2124 isl_id *id;
2125 isl_space *dim;
2126 isl_aff *aff;
2127 isl_pw_aff *val;
2129 if (op->getOpcode() != BO_Assign) {
2130 unsupported(op);
2131 return NULL;
2134 lhs = op->getLHS();
2135 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
2136 unsupported(op);
2137 return NULL;
2140 ref = cast<DeclRefExpr>(lhs);
2141 if (ref->getDecl() != iv) {
2142 unsupported(op);
2143 return NULL;
2146 val = extract_affine(op->getRHS());
2148 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
2150 dim = isl_space_params_alloc(ctx, 1);
2151 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
2152 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2153 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2155 val = isl_pw_aff_sub(val, isl_pw_aff_from_aff(aff));
2157 return val;
2160 /* Check that op is of the form iv += cst or iv -= cst
2161 * and return an affine expression corresponding oto cst or -cst accordingly.
2163 __isl_give isl_pw_aff *PetScan::extract_compound_increment(
2164 CompoundAssignOperator *op, clang::ValueDecl *iv)
2166 Expr *lhs;
2167 DeclRefExpr *ref;
2168 bool neg = false;
2169 isl_pw_aff *val;
2170 BinaryOperatorKind opcode;
2172 opcode = op->getOpcode();
2173 if (opcode != BO_AddAssign && opcode != BO_SubAssign) {
2174 unsupported(op);
2175 return NULL;
2177 if (opcode == BO_SubAssign)
2178 neg = true;
2180 lhs = op->getLHS();
2181 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
2182 unsupported(op);
2183 return NULL;
2186 ref = cast<DeclRefExpr>(lhs);
2187 if (ref->getDecl() != iv) {
2188 unsupported(op);
2189 return NULL;
2192 val = extract_affine(op->getRHS());
2193 if (neg)
2194 val = isl_pw_aff_neg(val);
2196 return val;
2199 /* Check that the increment of the given for loop increments
2200 * (or decrements) the induction variable "iv" and return
2201 * the increment as an affine expression if successful.
2203 __isl_give isl_pw_aff *PetScan::extract_increment(clang::ForStmt *stmt,
2204 ValueDecl *iv)
2206 Stmt *inc = stmt->getInc();
2208 if (!inc) {
2209 report_missing_increment(stmt);
2210 return NULL;
2213 if (inc->getStmtClass() == Stmt::UnaryOperatorClass)
2214 return extract_unary_increment(cast<UnaryOperator>(inc), iv);
2215 if (inc->getStmtClass() == Stmt::CompoundAssignOperatorClass)
2216 return extract_compound_increment(
2217 cast<CompoundAssignOperator>(inc), iv);
2218 if (inc->getStmtClass() == Stmt::BinaryOperatorClass)
2219 return extract_binary_increment(cast<BinaryOperator>(inc), iv);
2221 unsupported(inc);
2222 return NULL;
2225 /* Embed the given iteration domain in an extra outer loop
2226 * with induction variable "var".
2227 * If this variable appeared as a parameter in the constraints,
2228 * it is replaced by the new outermost dimension.
2230 static __isl_give isl_set *embed(__isl_take isl_set *set,
2231 __isl_take isl_id *var)
2233 int pos;
2235 set = isl_set_insert_dims(set, isl_dim_set, 0, 1);
2236 pos = isl_set_find_dim_by_id(set, isl_dim_param, var);
2237 if (pos >= 0) {
2238 set = isl_set_equate(set, isl_dim_param, pos, isl_dim_set, 0);
2239 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2242 isl_id_free(var);
2243 return set;
2246 /* Return those elements in the space of "cond" that come after
2247 * (based on "sign") an element in "cond".
2249 static __isl_give isl_set *after(__isl_take isl_set *cond, int sign)
2251 isl_map *previous_to_this;
2253 if (sign > 0)
2254 previous_to_this = isl_map_lex_lt(isl_set_get_space(cond));
2255 else
2256 previous_to_this = isl_map_lex_gt(isl_set_get_space(cond));
2258 cond = isl_set_apply(cond, previous_to_this);
2260 return cond;
2263 /* Create the infinite iteration domain
2265 * { [id] : id >= 0 }
2267 * If "scop" has an affine skip of type pet_skip_later,
2268 * then remove those iterations i that have an earlier iteration
2269 * where the skip condition is satisfied, meaning that iteration i
2270 * is not executed.
2271 * Since we are dealing with a loop without loop iterator,
2272 * the skip condition cannot refer to the current loop iterator and
2273 * so effectively, the returned set is of the form
2275 * { [0]; [id] : id >= 1 and not skip }
2277 static __isl_give isl_set *infinite_domain(__isl_take isl_id *id,
2278 struct pet_scop *scop)
2280 isl_ctx *ctx = isl_id_get_ctx(id);
2281 isl_set *domain;
2282 isl_set *skip;
2284 domain = isl_set_nat_universe(isl_space_set_alloc(ctx, 0, 1));
2285 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, id);
2287 if (!pet_scop_has_affine_skip(scop, pet_skip_later))
2288 return domain;
2290 skip = pet_scop_get_affine_skip_domain(scop, pet_skip_later);
2291 skip = embed(skip, isl_id_copy(id));
2292 skip = isl_set_intersect(skip , isl_set_copy(domain));
2293 domain = isl_set_subtract(domain, after(skip, 1));
2295 return domain;
2298 /* Create an identity affine expression on the space containing "domain",
2299 * which is assumed to be one-dimensional.
2301 static __isl_give isl_aff *identity_aff(__isl_keep isl_set *domain)
2303 isl_local_space *ls;
2305 ls = isl_local_space_from_space(isl_set_get_space(domain));
2306 return isl_aff_var_on_domain(ls, isl_dim_set, 0);
2309 /* Create an affine expression that maps elements
2310 * of a single-dimensional array "id_test" to the previous element
2311 * (according to "inc"), provided this element belongs to "domain".
2312 * That is, create the affine expression
2314 * { id[x] -> id[x - inc] : x - inc in domain }
2316 static __isl_give isl_multi_pw_aff *map_to_previous(__isl_take isl_id *id_test,
2317 __isl_take isl_set *domain, __isl_take isl_val *inc)
2319 isl_space *space;
2320 isl_local_space *ls;
2321 isl_aff *aff;
2322 isl_multi_pw_aff *prev;
2324 space = isl_set_get_space(domain);
2325 ls = isl_local_space_from_space(space);
2326 aff = isl_aff_var_on_domain(ls, isl_dim_set, 0);
2327 aff = isl_aff_add_constant_val(aff, isl_val_neg(inc));
2328 prev = isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
2329 domain = isl_set_preimage_multi_pw_aff(domain,
2330 isl_multi_pw_aff_copy(prev));
2331 prev = isl_multi_pw_aff_intersect_domain(prev, domain);
2332 prev = isl_multi_pw_aff_set_tuple_id(prev, isl_dim_out, id_test);
2334 return prev;
2337 /* Add an implication to "scop" expressing that if an element of
2338 * virtual array "id_test" has value "satisfied" then all previous elements
2339 * of this array also have that value. The set of previous elements
2340 * is bounded by "domain". If "sign" is negative then the iterator
2341 * is decreasing and we express that all subsequent array elements
2342 * (but still defined previously) have the same value.
2344 static struct pet_scop *add_implication(struct pet_scop *scop,
2345 __isl_take isl_id *id_test, __isl_take isl_set *domain, int sign,
2346 int satisfied)
2348 isl_space *space;
2349 isl_map *map;
2351 domain = isl_set_set_tuple_id(domain, id_test);
2352 space = isl_set_get_space(domain);
2353 if (sign > 0)
2354 map = isl_map_lex_ge(space);
2355 else
2356 map = isl_map_lex_le(space);
2357 map = isl_map_intersect_range(map, domain);
2358 scop = pet_scop_add_implication(scop, map, satisfied);
2360 return scop;
2363 /* Add a filter to "scop" that imposes that it is only executed
2364 * when the variable identified by "id_test" has a zero value
2365 * for all previous iterations of "domain".
2367 * In particular, add a filter that imposes that the array
2368 * has a zero value at the previous iteration of domain and
2369 * add an implication that implies that it then has that
2370 * value for all previous iterations.
2372 static struct pet_scop *scop_add_break(struct pet_scop *scop,
2373 __isl_take isl_id *id_test, __isl_take isl_set *domain,
2374 __isl_take isl_val *inc)
2376 isl_multi_pw_aff *prev;
2377 int sign = isl_val_sgn(inc);
2379 prev = map_to_previous(isl_id_copy(id_test), isl_set_copy(domain), inc);
2380 scop = add_implication(scop, id_test, domain, sign, 0);
2381 scop = pet_scop_filter(scop, prev, 0);
2383 return scop;
2386 /* Construct a pet_scop for an infinite loop around the given body.
2388 * We extract a pet_scop for the body and then embed it in a loop with
2389 * iteration domain
2391 * { [t] : t >= 0 }
2393 * and schedule
2395 * { [t] -> [t] }
2397 * If the body contains any break, then it is taken into
2398 * account in infinite_domain (if the skip condition is affine)
2399 * or in scop_add_break (if the skip condition is not affine).
2401 * If we were only able to extract part of the body, then simply
2402 * return that part.
2404 struct pet_scop *PetScan::extract_infinite_loop(Stmt *body)
2406 isl_id *id, *id_test;
2407 isl_set *domain;
2408 isl_aff *ident;
2409 struct pet_scop *scop;
2410 bool has_var_break;
2412 scop = extract(body);
2413 if (!scop)
2414 return NULL;
2415 if (partial)
2416 return scop;
2418 id = isl_id_alloc(ctx, "t", NULL);
2419 domain = infinite_domain(isl_id_copy(id), scop);
2420 ident = identity_aff(domain);
2422 has_var_break = pet_scop_has_var_skip(scop, pet_skip_later);
2423 if (has_var_break)
2424 id_test = pet_scop_get_skip_id(scop, pet_skip_later);
2426 scop = pet_scop_embed(scop, isl_set_copy(domain),
2427 isl_aff_copy(ident), ident, id);
2428 if (has_var_break)
2429 scop = scop_add_break(scop, id_test, domain, isl_val_one(ctx));
2430 else
2431 isl_set_free(domain);
2433 return scop;
2436 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
2438 * for (;;)
2439 * body
2442 struct pet_scop *PetScan::extract_infinite_for(ForStmt *stmt)
2444 clear_assignments clear(assigned_value);
2445 clear.TraverseStmt(stmt->getBody());
2447 return extract_infinite_loop(stmt->getBody());
2450 /* Add an array with the given extent (range of "index") to the list
2451 * of arrays in "scop" and return the extended pet_scop.
2452 * The array is marked as attaining values 0 and 1 only and
2453 * as each element being assigned at most once.
2455 static struct pet_scop *scop_add_array(struct pet_scop *scop,
2456 __isl_keep isl_multi_pw_aff *index, clang::ASTContext &ast_ctx)
2458 int int_size = ast_ctx.getTypeInfo(ast_ctx.IntTy).first / 8;
2460 return pet_scop_add_boolean_array(scop, isl_multi_pw_aff_copy(index),
2461 int_size);
2464 /* Construct a pet_scop for a while loop of the form
2466 * while (pa)
2467 * body
2469 * In particular, construct a scop for an infinite loop around body and
2470 * intersect the domain with the affine expression.
2471 * Note that this intersection may result in an empty loop.
2473 struct pet_scop *PetScan::extract_affine_while(__isl_take isl_pw_aff *pa,
2474 Stmt *body)
2476 struct pet_scop *scop;
2477 isl_set *dom;
2478 isl_set *valid;
2480 valid = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2481 dom = isl_pw_aff_non_zero_set(pa);
2482 scop = extract_infinite_loop(body);
2483 scop = pet_scop_restrict(scop, dom);
2484 scop = pet_scop_restrict_context(scop, valid);
2486 return scop;
2489 /* Construct a scop for a while, given the scops for the condition
2490 * and the body, the filter identifier and the iteration domain of
2491 * the while loop.
2493 * In particular, the scop for the condition is filtered to depend
2494 * on "id_test" evaluating to true for all previous iterations
2495 * of the loop, while the scop for the body is filtered to depend
2496 * on "id_test" evaluating to true for all iterations up to the
2497 * current iteration.
2498 * The actual filter only imposes that this virtual array has
2499 * value one on the previous or the current iteration.
2500 * The fact that this condition also applies to the previous
2501 * iterations is enforced by an implication.
2503 * These filtered scops are then combined into a single scop.
2505 * "sign" is positive if the iterator increases and negative
2506 * if it decreases.
2508 static struct pet_scop *scop_add_while(struct pet_scop *scop_cond,
2509 struct pet_scop *scop_body, __isl_take isl_id *id_test,
2510 __isl_take isl_set *domain, __isl_take isl_val *inc)
2512 isl_ctx *ctx = isl_set_get_ctx(domain);
2513 isl_space *space;
2514 isl_multi_pw_aff *test_index;
2515 isl_multi_pw_aff *prev;
2516 int sign = isl_val_sgn(inc);
2517 struct pet_scop *scop;
2519 prev = map_to_previous(isl_id_copy(id_test), isl_set_copy(domain), inc);
2520 scop_cond = pet_scop_filter(scop_cond, prev, 1);
2522 space = isl_space_map_from_set(isl_set_get_space(domain));
2523 test_index = isl_multi_pw_aff_identity(space);
2524 test_index = isl_multi_pw_aff_set_tuple_id(test_index, isl_dim_out,
2525 isl_id_copy(id_test));
2526 scop_body = pet_scop_filter(scop_body, test_index, 1);
2528 scop = pet_scop_add_seq(ctx, scop_cond, scop_body);
2529 scop = add_implication(scop, id_test, domain, sign, 1);
2531 return scop;
2534 /* Check if the while loop is of the form
2536 * while (affine expression)
2537 * body
2539 * If so, call extract_affine_while to construct a scop.
2541 * Otherwise, construct a generic while scop, with iteration domain
2542 * { [t] : t >= 0 }. The scop consists of two parts, one for
2543 * evaluating the condition and one for the body.
2544 * The schedule is adjusted to reflect that the condition is evaluated
2545 * before the body is executed and the body is filtered to depend
2546 * on the result of the condition evaluating to true on all iterations
2547 * up to the current iteration, while the evaluation of the condition itself
2548 * is filtered to depend on the result of the condition evaluating to true
2549 * on all previous iterations.
2550 * The context of the scop representing the body is dropped
2551 * because we don't know how many times the body will be executed,
2552 * if at all.
2554 * If the body contains any break, then it is taken into
2555 * account in infinite_domain (if the skip condition is affine)
2556 * or in scop_add_break (if the skip condition is not affine).
2558 * If we were only able to extract part of the body, then simply
2559 * return that part.
2561 struct pet_scop *PetScan::extract(WhileStmt *stmt)
2563 Expr *cond;
2564 int test_nr, stmt_nr;
2565 isl_id *id, *id_test, *id_break_test;
2566 isl_multi_pw_aff *test_index;
2567 isl_set *domain;
2568 isl_aff *ident;
2569 isl_pw_aff *pa;
2570 struct pet_scop *scop, *scop_body;
2571 bool has_var_break;
2573 cond = stmt->getCond();
2574 if (!cond) {
2575 unsupported(stmt);
2576 return NULL;
2579 clear_assignments clear(assigned_value);
2580 clear.TraverseStmt(stmt->getBody());
2582 pa = try_extract_affine_condition(cond);
2583 if (pa)
2584 return extract_affine_while(pa, stmt->getBody());
2586 if (!allow_nested) {
2587 unsupported(stmt);
2588 return NULL;
2591 test_nr = n_test++;
2592 stmt_nr = n_stmt++;
2593 scop_body = extract(stmt->getBody());
2594 if (partial)
2595 return scop_body;
2597 test_index = pet_create_test_index(ctx, test_nr);
2598 scop = extract_non_affine_condition(cond, stmt_nr,
2599 isl_multi_pw_aff_copy(test_index));
2600 scop = scop_add_array(scop, test_index, ast_context);
2601 id_test = isl_multi_pw_aff_get_tuple_id(test_index, isl_dim_out);
2602 isl_multi_pw_aff_free(test_index);
2604 id = isl_id_alloc(ctx, "t", NULL);
2605 domain = infinite_domain(isl_id_copy(id), scop_body);
2606 ident = identity_aff(domain);
2608 has_var_break = pet_scop_has_var_skip(scop_body, pet_skip_later);
2609 if (has_var_break)
2610 id_break_test = pet_scop_get_skip_id(scop_body, pet_skip_later);
2612 scop = pet_scop_prefix(scop, 0);
2613 scop = pet_scop_embed(scop, isl_set_copy(domain), isl_aff_copy(ident),
2614 isl_aff_copy(ident), isl_id_copy(id));
2615 scop_body = pet_scop_reset_context(scop_body);
2616 scop_body = pet_scop_prefix(scop_body, 1);
2617 scop_body = pet_scop_embed(scop_body, isl_set_copy(domain),
2618 isl_aff_copy(ident), ident, id);
2620 if (has_var_break) {
2621 scop = scop_add_break(scop, isl_id_copy(id_break_test),
2622 isl_set_copy(domain), isl_val_one(ctx));
2623 scop_body = scop_add_break(scop_body, id_break_test,
2624 isl_set_copy(domain), isl_val_one(ctx));
2626 scop = scop_add_while(scop, scop_body, id_test, domain,
2627 isl_val_one(ctx));
2629 return scop;
2632 /* Check whether "cond" expresses a simple loop bound
2633 * on the only set dimension.
2634 * In particular, if "up" is set then "cond" should contain only
2635 * upper bounds on the set dimension.
2636 * Otherwise, it should contain only lower bounds.
2638 static bool is_simple_bound(__isl_keep isl_set *cond, __isl_keep isl_val *inc)
2640 if (isl_val_is_pos(inc))
2641 return !isl_set_dim_has_any_lower_bound(cond, isl_dim_set, 0);
2642 else
2643 return !isl_set_dim_has_any_upper_bound(cond, isl_dim_set, 0);
2646 /* Extend a condition on a given iteration of a loop to one that
2647 * imposes the same condition on all previous iterations.
2648 * "domain" expresses the lower [upper] bound on the iterations
2649 * when inc is positive [negative].
2651 * In particular, we construct the condition (when inc is positive)
2653 * forall i' : (domain(i') and i' <= i) => cond(i')
2655 * which is equivalent to
2657 * not exists i' : domain(i') and i' <= i and not cond(i')
2659 * We construct this set by negating cond, applying a map
2661 * { [i'] -> [i] : domain(i') and i' <= i }
2663 * and then negating the result again.
2665 static __isl_give isl_set *valid_for_each_iteration(__isl_take isl_set *cond,
2666 __isl_take isl_set *domain, __isl_take isl_val *inc)
2668 isl_map *previous_to_this;
2670 if (isl_val_is_pos(inc))
2671 previous_to_this = isl_map_lex_le(isl_set_get_space(domain));
2672 else
2673 previous_to_this = isl_map_lex_ge(isl_set_get_space(domain));
2675 previous_to_this = isl_map_intersect_domain(previous_to_this, domain);
2677 cond = isl_set_complement(cond);
2678 cond = isl_set_apply(cond, previous_to_this);
2679 cond = isl_set_complement(cond);
2681 isl_val_free(inc);
2683 return cond;
2686 /* Construct a domain of the form
2688 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
2690 static __isl_give isl_set *strided_domain(__isl_take isl_id *id,
2691 __isl_take isl_pw_aff *init, __isl_take isl_val *inc)
2693 isl_aff *aff;
2694 isl_space *dim;
2695 isl_set *set;
2697 init = isl_pw_aff_insert_dims(init, isl_dim_in, 0, 1);
2698 dim = isl_pw_aff_get_domain_space(init);
2699 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2700 aff = isl_aff_add_coefficient_val(aff, isl_dim_in, 0, inc);
2701 init = isl_pw_aff_add(init, isl_pw_aff_from_aff(aff));
2703 dim = isl_space_set_alloc(isl_pw_aff_get_ctx(init), 1, 1);
2704 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
2705 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2706 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2708 set = isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff), init);
2710 set = isl_set_lower_bound_si(set, isl_dim_set, 0, 0);
2712 return isl_set_params(set);
2715 /* Assuming "cond" represents a bound on a loop where the loop
2716 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2717 * is possible.
2719 * Under the given assumptions, wrapping is only possible if "cond" allows
2720 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2721 * increasing iterator and 0 in case of a decreasing iterator.
2723 static bool can_wrap(__isl_keep isl_set *cond, ValueDecl *iv,
2724 __isl_keep isl_val *inc)
2726 bool cw;
2727 isl_ctx *ctx;
2728 isl_val *limit;
2729 isl_set *test;
2731 test = isl_set_copy(cond);
2733 ctx = isl_set_get_ctx(test);
2734 if (isl_val_is_neg(inc))
2735 limit = isl_val_zero(ctx);
2736 else {
2737 limit = isl_val_int_from_ui(ctx, get_type_size(iv));
2738 limit = isl_val_2exp(limit);
2739 limit = isl_val_sub_ui(limit, 1);
2742 test = isl_set_fix_val(cond, isl_dim_set, 0, limit);
2743 cw = !isl_set_is_empty(test);
2744 isl_set_free(test);
2746 return cw;
2749 /* Given a one-dimensional space, construct the following affine expression
2750 * on this space
2752 * { [v] -> [v mod 2^width] }
2754 * where width is the number of bits used to represent the values
2755 * of the unsigned variable "iv".
2757 static __isl_give isl_aff *compute_wrapping(__isl_take isl_space *dim,
2758 ValueDecl *iv)
2760 isl_ctx *ctx;
2761 isl_val *mod;
2762 isl_aff *aff;
2764 ctx = isl_space_get_ctx(dim);
2765 mod = isl_val_int_from_ui(ctx, get_type_size(iv));
2766 mod = isl_val_2exp(mod);
2768 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2769 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2770 aff = isl_aff_mod_val(aff, mod);
2772 return aff;
2775 /* Project out the parameter "id" from "set".
2777 static __isl_give isl_set *set_project_out_by_id(__isl_take isl_set *set,
2778 __isl_keep isl_id *id)
2780 int pos;
2782 pos = isl_set_find_dim_by_id(set, isl_dim_param, id);
2783 if (pos >= 0)
2784 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2786 return set;
2789 /* Compute the set of parameters for which "set1" is a subset of "set2".
2791 * set1 is a subset of set2 if
2793 * forall i in set1 : i in set2
2795 * or
2797 * not exists i in set1 and i not in set2
2799 * i.e.,
2801 * not exists i in set1 \ set2
2803 static __isl_give isl_set *enforce_subset(__isl_take isl_set *set1,
2804 __isl_take isl_set *set2)
2806 return isl_set_complement(isl_set_params(isl_set_subtract(set1, set2)));
2809 /* Compute the set of parameter values for which "cond" holds
2810 * on the next iteration for each element of "dom".
2812 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2813 * and then compute the set of parameters for which the result is a subset
2814 * of "cond".
2816 static __isl_give isl_set *valid_on_next(__isl_take isl_set *cond,
2817 __isl_take isl_set *dom, __isl_take isl_val *inc)
2819 isl_space *space;
2820 isl_aff *aff;
2821 isl_map *next;
2823 space = isl_set_get_space(dom);
2824 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2825 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2826 aff = isl_aff_add_constant_val(aff, inc);
2827 next = isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2829 dom = isl_set_apply(dom, next);
2831 return enforce_subset(dom, cond);
2834 /* Construct a pet_scop for a for statement.
2835 * The for loop is required to be of the form
2837 * for (i = init; condition; ++i)
2839 * or
2841 * for (i = init; condition; --i)
2843 * The initialization of the for loop should either be an assignment
2844 * to an integer variable, or a declaration of such a variable with
2845 * initialization.
2847 * The condition is allowed to contain nested accesses, provided
2848 * they are not being written to inside the body of the loop.
2849 * Otherwise, or if the condition is otherwise non-affine, the for loop is
2850 * essentially treated as a while loop, with iteration domain
2851 * { [i] : i >= init }.
2853 * We extract a pet_scop for the body and then embed it in a loop with
2854 * iteration domain and schedule
2856 * { [i] : i >= init and condition' }
2857 * { [i] -> [i] }
2859 * or
2861 * { [i] : i <= init and condition' }
2862 * { [i] -> [-i] }
2864 * Where condition' is equal to condition if the latter is
2865 * a simple upper [lower] bound and a condition that is extended
2866 * to apply to all previous iterations otherwise.
2868 * If the condition is non-affine, then we drop the condition from the
2869 * iteration domain and instead create a separate statement
2870 * for evaluating the condition. The body is then filtered to depend
2871 * on the result of the condition evaluating to true on all iterations
2872 * up to the current iteration, while the evaluation the condition itself
2873 * is filtered to depend on the result of the condition evaluating to true
2874 * on all previous iterations.
2875 * The context of the scop representing the body is dropped
2876 * because we don't know how many times the body will be executed,
2877 * if at all.
2879 * If the stride of the loop is not 1, then "i >= init" is replaced by
2881 * (exists a: i = init + stride * a and a >= 0)
2883 * If the loop iterator i is unsigned, then wrapping may occur.
2884 * We therefore use a virtual iterator instead that does not wrap.
2885 * However, the condition in the code applies
2886 * to the wrapped value, so we need to change condition(i)
2887 * into condition([i % 2^width]). Similarly, we replace all accesses
2888 * to the original iterator by the wrapping of the virtual iterator.
2889 * Note that there may be no need to perform this final wrapping
2890 * if the loop condition (after wrapping) satisfies certain conditions.
2891 * However, the is_simple_bound condition is not enough since it doesn't
2892 * check if there even is an upper bound.
2894 * Wrapping on unsigned iterators can be avoided entirely if
2895 * loop condition is simple, the loop iterator is incremented
2896 * [decremented] by one and the last value before wrapping cannot
2897 * possibly satisfy the loop condition.
2899 * Before extracting a pet_scop from the body we remove all
2900 * assignments in assigned_value to variables that are assigned
2901 * somewhere in the body of the loop.
2903 * Valid parameters for a for loop are those for which the initial
2904 * value itself, the increment on each domain iteration and
2905 * the condition on both the initial value and
2906 * the result of incrementing the iterator for each iteration of the domain
2907 * can be evaluated.
2908 * If the loop condition is non-affine, then we only consider validity
2909 * of the initial value.
2911 * If the body contains any break, then we keep track of it in "skip"
2912 * (if the skip condition is affine) or it is handled in scop_add_break
2913 * (if the skip condition is not affine).
2914 * Note that the affine break condition needs to be considered with
2915 * respect to previous iterations in the virtual domain (if any).
2917 * If we were only able to extract part of the body, then simply
2918 * return that part.
2920 struct pet_scop *PetScan::extract_for(ForStmt *stmt)
2922 BinaryOperator *ass;
2923 Decl *decl;
2924 Stmt *init;
2925 Expr *lhs, *rhs;
2926 ValueDecl *iv;
2927 isl_local_space *ls;
2928 isl_set *domain;
2929 isl_aff *sched;
2930 isl_set *cond = NULL;
2931 isl_set *skip = NULL;
2932 isl_id *id, *id_test = NULL, *id_break_test;
2933 struct pet_scop *scop, *scop_cond = NULL;
2934 assigned_value_cache cache(assigned_value);
2935 isl_val *inc;
2936 bool was_assigned;
2937 bool is_one;
2938 bool is_unsigned;
2939 bool is_simple;
2940 bool is_virtual;
2941 bool has_affine_break;
2942 bool has_var_break;
2943 isl_aff *wrap = NULL;
2944 isl_pw_aff *pa, *pa_inc, *init_val;
2945 isl_set *valid_init;
2946 isl_set *valid_cond;
2947 isl_set *valid_cond_init;
2948 isl_set *valid_cond_next;
2949 isl_set *valid_inc;
2950 int stmt_id;
2952 if (!stmt->getInit() && !stmt->getCond() && !stmt->getInc())
2953 return extract_infinite_for(stmt);
2955 init = stmt->getInit();
2956 if (!init) {
2957 unsupported(stmt);
2958 return NULL;
2960 if ((ass = initialization_assignment(init)) != NULL) {
2961 iv = extract_induction_variable(ass);
2962 if (!iv)
2963 return NULL;
2964 lhs = ass->getLHS();
2965 rhs = ass->getRHS();
2966 } else if ((decl = initialization_declaration(init)) != NULL) {
2967 VarDecl *var = extract_induction_variable(init, decl);
2968 if (!var)
2969 return NULL;
2970 iv = var;
2971 rhs = var->getInit();
2972 lhs = create_DeclRefExpr(var);
2973 } else {
2974 unsupported(stmt->getInit());
2975 return NULL;
2978 assigned_value.erase(iv);
2979 clear_assignments clear(assigned_value);
2980 clear.TraverseStmt(stmt->getBody());
2982 was_assigned = assigned_value.find(iv) != assigned_value.end();
2983 clear_assignment(assigned_value, iv);
2984 init_val = extract_affine(rhs);
2985 if (!was_assigned)
2986 assigned_value.erase(iv);
2987 if (!init_val)
2988 return NULL;
2990 pa_inc = extract_increment(stmt, iv);
2991 if (!pa_inc) {
2992 isl_pw_aff_free(init_val);
2993 return NULL;
2996 inc = pet_extract_cst(pa_inc);
2997 if (!inc || isl_val_is_nan(inc)) {
2998 isl_pw_aff_free(init_val);
2999 isl_pw_aff_free(pa_inc);
3000 unsupported(stmt->getInc());
3001 isl_val_free(inc);
3002 return NULL;
3005 pa = try_extract_nested_condition(stmt->getCond());
3006 if (allow_nested && (!pa || pet_nested_any_in_pw_aff(pa)))
3007 stmt_id = n_stmt++;
3009 scop = extract(stmt->getBody());
3010 if (partial) {
3011 isl_pw_aff_free(init_val);
3012 isl_pw_aff_free(pa_inc);
3013 isl_pw_aff_free(pa);
3014 isl_val_free(inc);
3015 return scop;
3018 valid_inc = isl_pw_aff_domain(pa_inc);
3020 is_unsigned = iv->getType()->isUnsignedIntegerType();
3022 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
3024 has_affine_break = scop &&
3025 pet_scop_has_affine_skip(scop, pet_skip_later);
3026 if (has_affine_break)
3027 skip = pet_scop_get_affine_skip_domain(scop, pet_skip_later);
3028 has_var_break = scop && pet_scop_has_var_skip(scop, pet_skip_later);
3029 if (has_var_break)
3030 id_break_test = pet_scop_get_skip_id(scop, pet_skip_later);
3032 if (pa && !is_nested_allowed(pa, scop)) {
3033 isl_pw_aff_free(pa);
3034 pa = NULL;
3037 if (!allow_nested && !pa)
3038 pa = try_extract_affine_condition(stmt->getCond());
3039 valid_cond = isl_pw_aff_domain(isl_pw_aff_copy(pa));
3040 cond = isl_pw_aff_non_zero_set(pa);
3041 if (allow_nested && !cond) {
3042 isl_multi_pw_aff *test_index;
3043 int save_n_stmt = n_stmt;
3044 test_index = pet_create_test_index(ctx, n_test++);
3045 n_stmt = stmt_id;
3046 scop_cond = extract_non_affine_condition(stmt->getCond(),
3047 n_stmt++, isl_multi_pw_aff_copy(test_index));
3048 n_stmt = save_n_stmt;
3049 scop_cond = scop_add_array(scop_cond, test_index, ast_context);
3050 id_test = isl_multi_pw_aff_get_tuple_id(test_index,
3051 isl_dim_out);
3052 isl_multi_pw_aff_free(test_index);
3053 scop_cond = pet_scop_prefix(scop_cond, 0);
3054 scop = pet_scop_reset_context(scop);
3055 scop = pet_scop_prefix(scop, 1);
3056 cond = isl_set_universe(isl_space_set_alloc(ctx, 0, 0));
3059 cond = embed(cond, isl_id_copy(id));
3060 skip = embed(skip, isl_id_copy(id));
3061 valid_cond = isl_set_coalesce(valid_cond);
3062 valid_cond = embed(valid_cond, isl_id_copy(id));
3063 valid_inc = embed(valid_inc, isl_id_copy(id));
3064 is_one = isl_val_is_one(inc) || isl_val_is_negone(inc);
3065 is_virtual = is_unsigned && (!is_one || can_wrap(cond, iv, inc));
3067 valid_cond_init = enforce_subset(
3068 isl_set_from_pw_aff(isl_pw_aff_copy(init_val)),
3069 isl_set_copy(valid_cond));
3070 if (is_one && !is_virtual) {
3071 isl_pw_aff_free(init_val);
3072 pa = extract_comparison(isl_val_is_pos(inc) ? BO_GE : BO_LE,
3073 lhs, rhs, init);
3074 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(pa));
3075 valid_init = set_project_out_by_id(valid_init, id);
3076 domain = isl_pw_aff_non_zero_set(pa);
3077 } else {
3078 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(init_val));
3079 domain = strided_domain(isl_id_copy(id), init_val,
3080 isl_val_copy(inc));
3083 domain = embed(domain, isl_id_copy(id));
3084 if (is_virtual) {
3085 isl_map *rev_wrap;
3086 wrap = compute_wrapping(isl_set_get_space(cond), iv);
3087 rev_wrap = isl_map_from_aff(isl_aff_copy(wrap));
3088 rev_wrap = isl_map_reverse(rev_wrap);
3089 cond = isl_set_apply(cond, isl_map_copy(rev_wrap));
3090 skip = isl_set_apply(skip, isl_map_copy(rev_wrap));
3091 valid_cond = isl_set_apply(valid_cond, isl_map_copy(rev_wrap));
3092 valid_inc = isl_set_apply(valid_inc, rev_wrap);
3094 is_simple = is_simple_bound(cond, inc);
3095 if (!is_simple) {
3096 cond = isl_set_gist(cond, isl_set_copy(domain));
3097 is_simple = is_simple_bound(cond, inc);
3099 if (!is_simple)
3100 cond = valid_for_each_iteration(cond,
3101 isl_set_copy(domain), isl_val_copy(inc));
3102 domain = isl_set_intersect(domain, cond);
3103 if (has_affine_break) {
3104 skip = isl_set_intersect(skip , isl_set_copy(domain));
3105 skip = after(skip, isl_val_sgn(inc));
3106 domain = isl_set_subtract(domain, skip);
3108 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, isl_id_copy(id));
3109 ls = isl_local_space_from_space(isl_set_get_space(domain));
3110 sched = isl_aff_var_on_domain(ls, isl_dim_set, 0);
3111 if (isl_val_is_neg(inc))
3112 sched = isl_aff_neg(sched);
3114 valid_cond_next = valid_on_next(valid_cond, isl_set_copy(domain),
3115 isl_val_copy(inc));
3116 valid_inc = enforce_subset(isl_set_copy(domain), valid_inc);
3118 if (!is_virtual)
3119 wrap = identity_aff(domain);
3121 scop_cond = pet_scop_embed(scop_cond, isl_set_copy(domain),
3122 isl_aff_copy(sched), isl_aff_copy(wrap), isl_id_copy(id));
3123 scop = pet_scop_embed(scop, isl_set_copy(domain), sched, wrap, id);
3124 scop = resolve_nested(scop);
3125 if (has_var_break)
3126 scop = scop_add_break(scop, id_break_test, isl_set_copy(domain),
3127 isl_val_copy(inc));
3128 if (id_test) {
3129 scop = scop_add_while(scop_cond, scop, id_test, domain,
3130 isl_val_copy(inc));
3131 isl_set_free(valid_inc);
3132 } else {
3133 scop = pet_scop_restrict_context(scop, valid_inc);
3134 scop = pet_scop_restrict_context(scop, valid_cond_next);
3135 scop = pet_scop_restrict_context(scop, valid_cond_init);
3136 isl_set_free(domain);
3138 clear_assignment(assigned_value, iv);
3140 isl_val_free(inc);
3142 scop = pet_scop_restrict_context(scop, valid_init);
3144 return scop;
3147 /* Try and construct a pet_scop corresponding to a compound statement.
3149 * "skip_declarations" is set if we should skip initial declarations
3150 * in the children of the compound statements. This then implies
3151 * that this sequence of children should not be treated as a block
3152 * since the initial statements may be skipped.
3154 struct pet_scop *PetScan::extract(CompoundStmt *stmt, bool skip_declarations)
3156 return extract(stmt->children(), !skip_declarations, skip_declarations);
3159 /* For each nested access parameter in "space",
3160 * construct a corresponding pet_expr, place it in args and
3161 * record its position in "param2pos".
3162 * "n_arg" is the number of elements that are already in args.
3163 * The position recorded in "param2pos" takes this number into account.
3164 * If the pet_expr corresponding to a parameter is identical to
3165 * the pet_expr corresponding to an earlier parameter, then these two
3166 * parameters are made to refer to the same element in args.
3168 * Return the final number of elements in args or -1 if an error has occurred.
3170 int PetScan::extract_nested(__isl_keep isl_space *space,
3171 int n_arg, pet_expr **args, std::map<int,int> &param2pos)
3173 int nparam;
3175 nparam = isl_space_dim(space, isl_dim_param);
3176 for (int i = 0; i < nparam; ++i) {
3177 int j;
3178 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
3179 Expr *nested;
3181 if (!pet_nested_in_id(id)) {
3182 isl_id_free(id);
3183 continue;
3186 nested = (Expr *) isl_id_get_user(id);
3187 args[n_arg] = extract_expr(nested);
3188 isl_id_free(id);
3189 if (!args[n_arg])
3190 return -1;
3192 for (j = 0; j < n_arg; ++j)
3193 if (pet_expr_is_equal(args[j], args[n_arg]))
3194 break;
3196 if (j < n_arg) {
3197 pet_expr_free(args[n_arg]);
3198 args[n_arg] = NULL;
3199 param2pos[i] = j;
3200 } else
3201 param2pos[i] = n_arg++;
3204 return n_arg;
3207 /* For each nested access parameter in the access relations in "expr",
3208 * construct a corresponding pet_expr, place it in the arguments of "expr"
3209 * and record its position in "param2pos".
3210 * n is the number of nested access parameters.
3212 __isl_give pet_expr *PetScan::extract_nested(__isl_take pet_expr *expr, int n,
3213 std::map<int,int> &param2pos)
3215 isl_space *space;
3216 int i;
3217 pet_expr **args;
3219 args = isl_calloc_array(ctx, pet_expr *, n);
3220 if (!args)
3221 return pet_expr_free(expr);
3223 space = pet_expr_access_get_parameter_space(expr);
3224 n = extract_nested(space, 0, args, param2pos);
3225 isl_space_free(space);
3227 if (n < 0)
3228 expr = pet_expr_free(expr);
3229 else
3230 expr = pet_expr_set_n_arg(expr, n);
3232 for (i = 0; i < n; ++i)
3233 expr = pet_expr_set_arg(expr, i, args[i]);
3234 free(args);
3236 return expr;
3239 /* Look for parameters in any access relation in "expr" that
3240 * refer to nested accesses. In particular, these are
3241 * parameters with no name.
3243 * If there are any such parameters, then the domain of the index
3244 * expression and the access relation, which is still [] at this point,
3245 * is replaced by [[] -> [t_1,...,t_n]], with n the number of these parameters
3246 * (after identifying identical nested accesses).
3248 * This transformation is performed in several steps.
3249 * We first extract the arguments in extract_nested.
3250 * param2pos maps the original parameter position to the position
3251 * of the argument.
3252 * Then we move these parameters to input dimensions.
3253 * t2pos maps the positions of these temporary input dimensions
3254 * to the positions of the corresponding arguments.
3255 * Finally, we express these temporary dimensions in terms of the domain
3256 * [[] -> [t_1,...,t_n]] and precompose index expression and access
3257 * relations with this function.
3259 __isl_give pet_expr *PetScan::resolve_nested(__isl_take pet_expr *expr)
3261 int n;
3262 int nparam;
3263 isl_space *space;
3264 isl_local_space *ls;
3265 isl_aff *aff;
3266 isl_multi_aff *ma;
3267 std::map<int,int> param2pos;
3268 std::map<int,int> t2pos;
3270 if (!expr)
3271 return expr;
3273 n = pet_expr_get_n_arg(expr);
3274 for (int i = 0; i < n; ++i) {
3275 pet_expr *arg;
3276 arg = pet_expr_get_arg(expr, i);
3277 arg = resolve_nested(arg);
3278 expr = pet_expr_set_arg(expr, i, arg);
3281 if (pet_expr_get_type(expr) != pet_expr_access)
3282 return expr;
3284 space = pet_expr_access_get_parameter_space(expr);
3285 n = pet_nested_n_in_space(space);
3286 isl_space_free(space);
3287 if (n == 0)
3288 return expr;
3290 expr = extract_nested(expr, n, param2pos);
3291 if (!expr)
3292 return NULL;
3294 expr = pet_expr_access_align_params(expr);
3295 if (!expr)
3296 return NULL;
3298 n = 0;
3299 space = pet_expr_access_get_parameter_space(expr);
3300 nparam = isl_space_dim(space, isl_dim_param);
3301 for (int i = nparam - 1; i >= 0; --i) {
3302 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
3303 if (!pet_nested_in_id(id)) {
3304 isl_id_free(id);
3305 continue;
3308 expr = pet_expr_access_move_dims(expr,
3309 isl_dim_in, n, isl_dim_param, i, 1);
3310 t2pos[n] = param2pos[i];
3311 n++;
3313 isl_id_free(id);
3315 isl_space_free(space);
3317 space = pet_expr_access_get_parameter_space(expr);
3318 space = isl_space_set_from_params(space);
3319 space = isl_space_add_dims(space, isl_dim_set,
3320 pet_expr_get_n_arg(expr));
3321 space = isl_space_wrap(isl_space_from_range(space));
3322 ls = isl_local_space_from_space(isl_space_copy(space));
3323 space = isl_space_from_domain(space);
3324 space = isl_space_add_dims(space, isl_dim_out, n);
3325 ma = isl_multi_aff_zero(space);
3327 for (int i = 0; i < n; ++i) {
3328 aff = isl_aff_var_on_domain(isl_local_space_copy(ls),
3329 isl_dim_set, t2pos[i]);
3330 ma = isl_multi_aff_set_aff(ma, i, aff);
3332 isl_local_space_free(ls);
3334 expr = pet_expr_access_pullback_multi_aff(expr, ma);
3336 return expr;
3339 /* Return the file offset of the expansion location of "Loc".
3341 static unsigned getExpansionOffset(SourceManager &SM, SourceLocation Loc)
3343 return SM.getFileOffset(SM.getExpansionLoc(Loc));
3346 #ifdef HAVE_FINDLOCATIONAFTERTOKEN
3348 /* Return a SourceLocation for the location after the first semicolon
3349 * after "loc". If Lexer::findLocationAfterToken is available, we simply
3350 * call it and also skip trailing spaces and newline.
3352 static SourceLocation location_after_semi(SourceLocation loc, SourceManager &SM,
3353 const LangOptions &LO)
3355 return Lexer::findLocationAfterToken(loc, tok::semi, SM, LO, true);
3358 #else
3360 /* Return a SourceLocation for the location after the first semicolon
3361 * after "loc". If Lexer::findLocationAfterToken is not available,
3362 * we look in the underlying character data for the first semicolon.
3364 static SourceLocation location_after_semi(SourceLocation loc, SourceManager &SM,
3365 const LangOptions &LO)
3367 const char *semi;
3368 const char *s = SM.getCharacterData(loc);
3370 semi = strchr(s, ';');
3371 if (!semi)
3372 return SourceLocation();
3373 return loc.getFileLocWithOffset(semi + 1 - s);
3376 #endif
3378 /* If the token at "loc" is the first token on the line, then return
3379 * a location referring to the start of the line.
3380 * Otherwise, return "loc".
3382 * This function is used to extend a scop to the start of the line
3383 * if the first token of the scop is also the first token on the line.
3385 * We look for the first token on the line. If its location is equal to "loc",
3386 * then the latter is the location of the first token on the line.
3388 static SourceLocation move_to_start_of_line_if_first_token(SourceLocation loc,
3389 SourceManager &SM, const LangOptions &LO)
3391 std::pair<FileID, unsigned> file_offset_pair;
3392 llvm::StringRef file;
3393 const char *pos;
3394 Token tok;
3395 SourceLocation token_loc, line_loc;
3396 int col;
3398 loc = SM.getExpansionLoc(loc);
3399 col = SM.getExpansionColumnNumber(loc);
3400 line_loc = loc.getLocWithOffset(1 - col);
3401 file_offset_pair = SM.getDecomposedLoc(line_loc);
3402 file = SM.getBufferData(file_offset_pair.first, NULL);
3403 pos = file.data() + file_offset_pair.second;
3405 Lexer lexer(SM.getLocForStartOfFile(file_offset_pair.first), LO,
3406 file.begin(), pos, file.end());
3407 lexer.LexFromRawLexer(tok);
3408 token_loc = tok.getLocation();
3410 if (token_loc == loc)
3411 return line_loc;
3412 else
3413 return loc;
3416 /* Update start and end of "scop" to include the region covered by "range".
3417 * If "skip_semi" is set, then we assume "range" is followed by
3418 * a semicolon and also include this semicolon.
3420 struct pet_scop *PetScan::update_scop_start_end(struct pet_scop *scop,
3421 SourceRange range, bool skip_semi)
3423 SourceLocation loc = range.getBegin();
3424 SourceManager &SM = PP.getSourceManager();
3425 const LangOptions &LO = PP.getLangOpts();
3426 unsigned start, end;
3428 loc = move_to_start_of_line_if_first_token(loc, SM, LO);
3429 start = getExpansionOffset(SM, loc);
3430 loc = range.getEnd();
3431 if (skip_semi)
3432 loc = location_after_semi(loc, SM, LO);
3433 else
3434 loc = PP.getLocForEndOfToken(loc);
3435 end = getExpansionOffset(SM, loc);
3437 scop = pet_scop_update_start_end(scop, start, end);
3438 return scop;
3441 /* Convert a top-level pet_expr to a pet_scop with one statement.
3442 * This mainly involves resolving nested expression parameters
3443 * and setting the name of the iteration space.
3444 * The name is given by "label" if it is non-NULL. Otherwise,
3445 * it is of the form S_<n_stmt>.
3446 * start and end of the pet_scop are derived from those of "stmt".
3447 * If "stmt" is an expression statement, then its range does not
3448 * include the semicolon, while it should be included in the pet_scop.
3450 struct pet_scop *PetScan::extract(Stmt *stmt, __isl_take pet_expr *expr,
3451 __isl_take isl_id *label)
3453 struct pet_stmt *ps;
3454 struct pet_scop *scop;
3455 SourceLocation loc = stmt->getLocStart();
3456 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3457 bool skip_semi;
3459 expr = resolve_nested(expr);
3460 ps = pet_stmt_from_pet_expr(line, label, n_stmt++, expr);
3461 scop = pet_scop_from_pet_stmt(ctx, ps);
3463 skip_semi = isa<Expr>(stmt);
3464 scop = update_scop_start_end(scop, stmt->getSourceRange(), skip_semi);
3465 return scop;
3468 /* Check if we can extract an affine expression from "expr".
3469 * Return the expressions as an isl_pw_aff if we can and NULL otherwise.
3470 * We turn on autodetection so that we won't generate any warnings
3471 * and turn off nesting, so that we won't accept any non-affine constructs.
3473 __isl_give isl_pw_aff *PetScan::try_extract_affine(Expr *expr)
3475 isl_pw_aff *pwaff;
3476 int save_autodetect = options->autodetect;
3477 bool save_nesting = nesting_enabled;
3479 options->autodetect = 1;
3480 nesting_enabled = false;
3482 pwaff = extract_affine(expr);
3484 options->autodetect = save_autodetect;
3485 nesting_enabled = save_nesting;
3487 return pwaff;
3490 /* Check if we can extract an affine constraint from "expr".
3491 * Return the constraint as an isl_set if we can and NULL otherwise.
3492 * We turn on autodetection so that we won't generate any warnings
3493 * and turn off nesting, so that we won't accept any non-affine constructs.
3495 __isl_give isl_pw_aff *PetScan::try_extract_affine_condition(Expr *expr)
3497 isl_pw_aff *cond;
3498 int save_autodetect = options->autodetect;
3499 bool save_nesting = nesting_enabled;
3501 options->autodetect = 1;
3502 nesting_enabled = false;
3504 cond = extract_condition(expr);
3506 options->autodetect = save_autodetect;
3507 nesting_enabled = save_nesting;
3509 return cond;
3512 /* Check whether "expr" is an affine constraint.
3514 bool PetScan::is_affine_condition(Expr *expr)
3516 isl_pw_aff *cond;
3518 cond = try_extract_affine_condition(expr);
3519 isl_pw_aff_free(cond);
3521 return cond != NULL;
3524 /* Check if we can extract a condition from "expr".
3525 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
3526 * If allow_nested is set, then the condition may involve parameters
3527 * corresponding to nested accesses.
3528 * We turn on autodetection so that we won't generate any warnings.
3530 __isl_give isl_pw_aff *PetScan::try_extract_nested_condition(Expr *expr)
3532 isl_pw_aff *cond;
3533 int save_autodetect = options->autodetect;
3534 bool save_nesting = nesting_enabled;
3536 options->autodetect = 1;
3537 nesting_enabled = allow_nested;
3538 cond = extract_condition(expr);
3540 options->autodetect = save_autodetect;
3541 nesting_enabled = save_nesting;
3543 return cond;
3546 /* If the top-level expression of "stmt" is an assignment, then
3547 * return that assignment as a BinaryOperator.
3548 * Otherwise return NULL.
3550 static BinaryOperator *top_assignment_or_null(Stmt *stmt)
3552 BinaryOperator *ass;
3554 if (!stmt)
3555 return NULL;
3556 if (stmt->getStmtClass() != Stmt::BinaryOperatorClass)
3557 return NULL;
3559 ass = cast<BinaryOperator>(stmt);
3560 if(ass->getOpcode() != BO_Assign)
3561 return NULL;
3563 return ass;
3566 /* Check if the given if statement is a conditional assignement
3567 * with a non-affine condition. If so, construct a pet_scop
3568 * corresponding to this conditional assignment. Otherwise return NULL.
3570 * In particular we check if "stmt" is of the form
3572 * if (condition)
3573 * a = f(...);
3574 * else
3575 * a = g(...);
3577 * where a is some array or scalar access.
3578 * The constructed pet_scop then corresponds to the expression
3580 * a = condition ? f(...) : g(...)
3582 * All access relations in f(...) are intersected with condition
3583 * while all access relation in g(...) are intersected with the complement.
3585 struct pet_scop *PetScan::extract_conditional_assignment(IfStmt *stmt)
3587 BinaryOperator *ass_then, *ass_else;
3588 isl_multi_pw_aff *write_then, *write_else;
3589 isl_set *cond, *comp;
3590 isl_multi_pw_aff *index;
3591 isl_pw_aff *pa;
3592 int equal;
3593 int type_size;
3594 pet_expr *pe_cond, *pe_then, *pe_else, *pe, *pe_write;
3595 bool save_nesting = nesting_enabled;
3597 if (!options->detect_conditional_assignment)
3598 return NULL;
3600 ass_then = top_assignment_or_null(stmt->getThen());
3601 ass_else = top_assignment_or_null(stmt->getElse());
3603 if (!ass_then || !ass_else)
3604 return NULL;
3606 if (is_affine_condition(stmt->getCond()))
3607 return NULL;
3609 write_then = extract_index(ass_then->getLHS());
3610 write_else = extract_index(ass_else->getLHS());
3612 equal = isl_multi_pw_aff_plain_is_equal(write_then, write_else);
3613 isl_multi_pw_aff_free(write_else);
3614 if (equal < 0 || !equal) {
3615 isl_multi_pw_aff_free(write_then);
3616 return NULL;
3619 nesting_enabled = allow_nested;
3620 pa = extract_condition(stmt->getCond());
3621 nesting_enabled = save_nesting;
3622 cond = isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa));
3623 comp = isl_pw_aff_zero_set(isl_pw_aff_copy(pa));
3624 index = isl_multi_pw_aff_from_range(isl_multi_pw_aff_from_pw_aff(pa));
3626 pe_cond = pet_expr_from_index(index);
3628 pe_then = extract_expr(ass_then->getRHS());
3629 pe_then = pet_expr_restrict(pe_then, cond);
3630 pe_else = extract_expr(ass_else->getRHS());
3631 pe_else = pet_expr_restrict(pe_else, comp);
3633 pe = pet_expr_new_ternary(pe_cond, pe_then, pe_else);
3634 type_size = get_type_size(ass_then->getType(), ast_context);
3635 pe_write = pet_expr_from_index_and_depth(type_size, write_then,
3636 extract_depth(write_then));
3637 pe_write = pet_expr_access_set_write(pe_write, 1);
3638 pe_write = pet_expr_access_set_read(pe_write, 0);
3639 pe = pet_expr_new_binary(type_size, pet_op_assign, pe_write, pe);
3640 return extract(stmt, pe);
3643 /* Create a pet_scop with a single statement with name S_<stmt_nr>,
3644 * evaluating "cond" and writing the result to a virtual scalar,
3645 * as expressed by "index".
3647 struct pet_scop *PetScan::extract_non_affine_condition(Expr *cond, int stmt_nr,
3648 __isl_take isl_multi_pw_aff *index)
3650 pet_expr *expr, *write;
3651 struct pet_stmt *ps;
3652 SourceLocation loc = cond->getLocStart();
3653 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3655 write = pet_expr_from_index(index);
3656 write = pet_expr_access_set_write(write, 1);
3657 write = pet_expr_access_set_read(write, 0);
3658 expr = extract_expr(cond);
3659 expr = resolve_nested(expr);
3660 expr = pet_expr_new_binary(1, pet_op_assign, write, expr);
3661 ps = pet_stmt_from_pet_expr(line, NULL, stmt_nr, expr);
3662 return pet_scop_from_pet_stmt(ctx, ps);
3665 extern "C" {
3666 static __isl_give pet_expr *embed_access(__isl_take pet_expr *expr,
3667 void *user);
3670 /* Precompose the access relation and the index expression associated
3671 * to "expr" with the function pointed to by "user",
3672 * thereby embedding the access relation in the domain of this function.
3673 * The initial domain of the access relation and the index expression
3674 * is the zero-dimensional domain.
3676 static __isl_give pet_expr *embed_access(__isl_take pet_expr *expr, void *user)
3678 isl_multi_aff *ma = (isl_multi_aff *) user;
3680 return pet_expr_access_pullback_multi_aff(expr, isl_multi_aff_copy(ma));
3683 /* Precompose all access relations in "expr" with "ma", thereby
3684 * embedding them in the domain of "ma".
3686 static __isl_give pet_expr *embed(__isl_take pet_expr *expr,
3687 __isl_keep isl_multi_aff *ma)
3689 return pet_expr_map_access(expr, &embed_access, ma);
3692 /* For each nested access parameter in the domain of "stmt",
3693 * construct a corresponding pet_expr, place it before the original
3694 * elements in stmt->args and record its position in "param2pos".
3695 * n is the number of nested access parameters.
3697 struct pet_stmt *PetScan::extract_nested(struct pet_stmt *stmt, int n,
3698 std::map<int,int> &param2pos)
3700 int i;
3701 isl_space *space;
3702 int n_arg;
3703 pet_expr **args;
3705 n_arg = stmt->n_arg;
3706 args = isl_calloc_array(ctx, pet_expr *, n + n_arg);
3707 if (!args)
3708 goto error;
3710 space = isl_set_get_space(stmt->domain);
3711 n_arg = extract_nested(space, 0, args, param2pos);
3712 isl_space_free(space);
3714 if (n_arg < 0)
3715 goto error;
3717 for (i = 0; i < stmt->n_arg; ++i)
3718 args[n_arg + i] = stmt->args[i];
3719 free(stmt->args);
3720 stmt->args = args;
3721 stmt->n_arg += n_arg;
3723 return stmt;
3724 error:
3725 if (args) {
3726 for (i = 0; i < n; ++i)
3727 pet_expr_free(args[i]);
3728 free(args);
3730 pet_stmt_free(stmt);
3731 return NULL;
3734 /* Check whether any of the arguments i of "stmt" starting at position "n"
3735 * is equal to one of the first "n" arguments j.
3736 * If so, combine the constraints on arguments i and j and remove
3737 * argument i.
3739 static struct pet_stmt *remove_duplicate_arguments(struct pet_stmt *stmt, int n)
3741 int i, j;
3742 isl_map *map;
3744 if (!stmt)
3745 return NULL;
3746 if (n == 0)
3747 return stmt;
3748 if (n == stmt->n_arg)
3749 return stmt;
3751 map = isl_set_unwrap(stmt->domain);
3753 for (i = stmt->n_arg - 1; i >= n; --i) {
3754 for (j = 0; j < n; ++j)
3755 if (pet_expr_is_equal(stmt->args[i], stmt->args[j]))
3756 break;
3757 if (j >= n)
3758 continue;
3760 map = isl_map_equate(map, isl_dim_out, i, isl_dim_out, j);
3761 map = isl_map_project_out(map, isl_dim_out, i, 1);
3763 pet_expr_free(stmt->args[i]);
3764 for (j = i; j + 1 < stmt->n_arg; ++j)
3765 stmt->args[j] = stmt->args[j + 1];
3766 stmt->n_arg--;
3769 stmt->domain = isl_map_wrap(map);
3770 if (!stmt->domain)
3771 goto error;
3772 return stmt;
3773 error:
3774 pet_stmt_free(stmt);
3775 return NULL;
3778 /* Look for parameters in the iteration domain of "stmt" that
3779 * refer to nested accesses. In particular, these are
3780 * parameters with no name.
3782 * If there are any such parameters, then as many extra variables
3783 * (after identifying identical nested accesses) are inserted in the
3784 * range of the map wrapped inside the domain, before the original variables.
3785 * If the original domain is not a wrapped map, then a new wrapped
3786 * map is created with zero output dimensions.
3787 * The parameters are then equated to the corresponding output dimensions
3788 * and subsequently projected out, from the iteration domain,
3789 * the schedule and the access relations.
3790 * For each of the output dimensions, a corresponding argument
3791 * expression is inserted. Initially they are created with
3792 * a zero-dimensional domain, so they have to be embedded
3793 * in the current iteration domain.
3794 * param2pos maps the position of the parameter to the position
3795 * of the corresponding output dimension in the wrapped map.
3797 struct pet_stmt *PetScan::resolve_nested(struct pet_stmt *stmt)
3799 int n;
3800 int nparam;
3801 unsigned n_arg;
3802 isl_map *map;
3803 isl_space *space;
3804 isl_multi_aff *ma;
3805 std::map<int,int> param2pos;
3807 if (!stmt)
3808 return NULL;
3810 n = pet_nested_n_in_set(stmt->domain);
3811 if (n == 0)
3812 return stmt;
3814 n_arg = stmt->n_arg;
3815 stmt = extract_nested(stmt, n, param2pos);
3816 if (!stmt)
3817 return NULL;
3819 n = stmt->n_arg - n_arg;
3820 nparam = isl_set_dim(stmt->domain, isl_dim_param);
3821 if (isl_set_is_wrapping(stmt->domain))
3822 map = isl_set_unwrap(stmt->domain);
3823 else
3824 map = isl_map_from_domain(stmt->domain);
3825 map = isl_map_insert_dims(map, isl_dim_out, 0, n);
3827 for (int i = nparam - 1; i >= 0; --i) {
3828 isl_id *id;
3830 if (!pet_nested_in_map(map, i))
3831 continue;
3833 id = pet_expr_access_get_id(stmt->args[param2pos[i]]);
3834 map = isl_map_set_dim_id(map, isl_dim_out, param2pos[i], id);
3835 map = isl_map_equate(map, isl_dim_param, i, isl_dim_out,
3836 param2pos[i]);
3837 map = isl_map_project_out(map, isl_dim_param, i, 1);
3840 stmt->domain = isl_map_wrap(map);
3842 space = isl_space_unwrap(isl_set_get_space(stmt->domain));
3843 space = isl_space_from_domain(isl_space_domain(space));
3844 ma = isl_multi_aff_zero(space);
3845 for (int pos = 0; pos < n; ++pos)
3846 stmt->args[pos] = embed(stmt->args[pos], ma);
3847 isl_multi_aff_free(ma);
3849 stmt = pet_stmt_remove_nested_parameters(stmt);
3850 stmt = remove_duplicate_arguments(stmt, n);
3852 return stmt;
3855 /* For each statement in "scop", move the parameters that correspond
3856 * to nested access into the ranges of the domains and create
3857 * corresponding argument expressions.
3859 struct pet_scop *PetScan::resolve_nested(struct pet_scop *scop)
3861 if (!scop)
3862 return NULL;
3864 for (int i = 0; i < scop->n_stmt; ++i) {
3865 scop->stmts[i] = resolve_nested(scop->stmts[i]);
3866 if (!scop->stmts[i])
3867 goto error;
3870 return scop;
3871 error:
3872 pet_scop_free(scop);
3873 return NULL;
3876 /* Given an access expression "expr", is the variable accessed by
3877 * "expr" assigned anywhere inside "scop"?
3879 static bool is_assigned(__isl_keep pet_expr *expr, pet_scop *scop)
3881 bool assigned = false;
3882 isl_id *id;
3884 id = pet_expr_access_get_id(expr);
3885 assigned = pet_scop_writes(scop, id);
3886 isl_id_free(id);
3888 return assigned;
3891 /* Are all nested access parameters in "pa" allowed given "scop".
3892 * In particular, is none of them written by anywhere inside "scop".
3894 * If "scop" has any skip conditions, then no nested access parameters
3895 * are allowed. In particular, if there is any nested access in a guard
3896 * for a piece of code containing a "continue", then we want to introduce
3897 * a separate statement for evaluating this guard so that we can express
3898 * that the result is false for all previous iterations.
3900 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff *pa, pet_scop *scop)
3902 int nparam;
3904 if (!scop)
3905 return true;
3907 if (!pet_nested_any_in_pw_aff(pa))
3908 return true;
3910 if (pet_scop_has_skip(scop, pet_skip_now))
3911 return false;
3913 nparam = isl_pw_aff_dim(pa, isl_dim_param);
3914 for (int i = 0; i < nparam; ++i) {
3915 Expr *nested;
3916 isl_id *id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
3917 pet_expr *expr;
3918 bool allowed;
3920 if (!pet_nested_in_id(id)) {
3921 isl_id_free(id);
3922 continue;
3925 nested = (Expr *) isl_id_get_user(id);
3926 expr = extract_expr(nested);
3927 allowed = pet_expr_get_type(expr) == pet_expr_access &&
3928 !is_assigned(expr, scop);
3930 pet_expr_free(expr);
3931 isl_id_free(id);
3933 if (!allowed)
3934 return false;
3937 return true;
3940 /* Construct a pet_scop for a non-affine if statement.
3942 * We create a separate statement that writes the result
3943 * of the non-affine condition to a virtual scalar.
3944 * A constraint requiring the value of this virtual scalar to be one
3945 * is added to the iteration domains of the then branch.
3946 * Similarly, a constraint requiring the value of this virtual scalar
3947 * to be zero is added to the iteration domains of the else branch, if any.
3948 * We adjust the schedules to ensure that the virtual scalar is written
3949 * before it is read.
3951 * If there are any breaks or continues in the then and/or else
3952 * branches, then we may have to compute a new skip condition.
3953 * This is handled using a pet_skip_info object.
3954 * On initialization, the object checks if skip conditions need
3955 * to be computed. If so, it does so in pet_skip_info_if_extract_index and
3956 * adds them in pet_skip_info_if_add.
3958 struct pet_scop *PetScan::extract_non_affine_if(Expr *cond,
3959 struct pet_scop *scop_then, struct pet_scop *scop_else,
3960 bool have_else, int stmt_id)
3962 struct pet_scop *scop;
3963 isl_multi_pw_aff *test_index;
3964 int int_size;
3965 int save_n_stmt = n_stmt;
3967 test_index = pet_create_test_index(ctx, n_test++);
3968 n_stmt = stmt_id;
3969 scop = extract_non_affine_condition(cond, n_stmt++,
3970 isl_multi_pw_aff_copy(test_index));
3971 n_stmt = save_n_stmt;
3972 scop = scop_add_array(scop, test_index, ast_context);
3974 pet_skip_info skip;
3975 pet_skip_info_if_init(&skip, ctx, scop_then, scop_else, have_else, 0);
3976 int_size = ast_context.getTypeInfo(ast_context.IntTy).first / 8;
3977 pet_skip_info_if_extract_index(&skip, test_index, int_size,
3978 &n_stmt, &n_test);
3980 scop = pet_scop_prefix(scop, 0);
3981 scop_then = pet_scop_prefix(scop_then, 1);
3982 scop_then = pet_scop_filter(scop_then,
3983 isl_multi_pw_aff_copy(test_index), 1);
3984 if (have_else) {
3985 scop_else = pet_scop_prefix(scop_else, 1);
3986 scop_else = pet_scop_filter(scop_else, test_index, 0);
3987 scop_then = pet_scop_add_par(ctx, scop_then, scop_else);
3988 } else
3989 isl_multi_pw_aff_free(test_index);
3991 scop = pet_scop_add_seq(ctx, scop, scop_then);
3993 scop = pet_skip_info_if_add(&skip, scop, 2);
3995 return scop;
3998 /* Construct a pet_scop for an if statement.
4000 * If the condition fits the pattern of a conditional assignment,
4001 * then it is handled by extract_conditional_assignment.
4002 * Otherwise, we do the following.
4004 * If the condition is affine, then the condition is added
4005 * to the iteration domains of the then branch, while the
4006 * opposite of the condition in added to the iteration domains
4007 * of the else branch, if any.
4008 * We allow the condition to be dynamic, i.e., to refer to
4009 * scalars or array elements that may be written to outside
4010 * of the given if statement. These nested accesses are then represented
4011 * as output dimensions in the wrapping iteration domain.
4012 * If it is also written _inside_ the then or else branch, then
4013 * we treat the condition as non-affine.
4014 * As explained in extract_non_affine_if, this will introduce
4015 * an extra statement.
4016 * For aesthetic reasons, we want this statement to have a statement
4017 * number that is lower than those of the then and else branches.
4018 * In order to evaluate if we will need such a statement, however, we
4019 * first construct scops for the then and else branches.
4020 * We therefore reserve a statement number if we might have to
4021 * introduce such an extra statement.
4023 * If the condition is not affine, then the scop is created in
4024 * extract_non_affine_if.
4026 * If there are any breaks or continues in the then and/or else
4027 * branches, then we may have to compute a new skip condition.
4028 * This is handled using a pet_skip_info object.
4029 * On initialization, the object checks if skip conditions need
4030 * to be computed. If so, it does so in pet_skip_info_if_extract_cond and
4031 * adds them in pet_skip_info_if_add.
4033 struct pet_scop *PetScan::extract(IfStmt *stmt)
4035 struct pet_scop *scop_then, *scop_else = NULL, *scop;
4036 isl_pw_aff *cond;
4037 int stmt_id;
4038 int int_size;
4039 isl_set *set;
4040 isl_set *valid;
4042 clear_assignments clear(assigned_value);
4043 clear.TraverseStmt(stmt->getThen());
4044 if (stmt->getElse())
4045 clear.TraverseStmt(stmt->getElse());
4047 scop = extract_conditional_assignment(stmt);
4048 if (scop)
4049 return scop;
4051 cond = try_extract_nested_condition(stmt->getCond());
4052 if (allow_nested && (!cond || pet_nested_any_in_pw_aff(cond)))
4053 stmt_id = n_stmt++;
4056 assigned_value_cache cache(assigned_value);
4057 scop_then = extract(stmt->getThen());
4060 if (stmt->getElse()) {
4061 assigned_value_cache cache(assigned_value);
4062 scop_else = extract(stmt->getElse());
4063 if (options->autodetect) {
4064 if (scop_then && !scop_else) {
4065 partial = true;
4066 isl_pw_aff_free(cond);
4067 return scop_then;
4069 if (!scop_then && scop_else) {
4070 partial = true;
4071 isl_pw_aff_free(cond);
4072 return scop_else;
4077 if (cond &&
4078 (!is_nested_allowed(cond, scop_then) ||
4079 (stmt->getElse() && !is_nested_allowed(cond, scop_else)))) {
4080 isl_pw_aff_free(cond);
4081 cond = NULL;
4083 if (allow_nested && !cond)
4084 return extract_non_affine_if(stmt->getCond(), scop_then,
4085 scop_else, stmt->getElse(), stmt_id);
4087 if (!cond)
4088 cond = extract_condition(stmt->getCond());
4090 pet_skip_info skip;
4091 pet_skip_info_if_init(&skip, ctx, scop_then, scop_else,
4092 stmt->getElse() != NULL, 1);
4093 pet_skip_info_if_extract_cond(&skip, cond, int_size, &n_stmt, &n_test);
4095 valid = isl_pw_aff_domain(isl_pw_aff_copy(cond));
4096 set = isl_pw_aff_non_zero_set(cond);
4097 scop = pet_scop_restrict(scop_then, isl_set_copy(set));
4099 if (stmt->getElse()) {
4100 set = isl_set_subtract(isl_set_copy(valid), set);
4101 scop_else = pet_scop_restrict(scop_else, set);
4102 scop = pet_scop_add_par(ctx, scop, scop_else);
4103 } else
4104 isl_set_free(set);
4105 scop = resolve_nested(scop);
4106 scop = pet_scop_restrict_context(scop, valid);
4108 if (pet_skip_info_has_skip(&skip))
4109 scop = pet_scop_prefix(scop, 0);
4110 scop = pet_skip_info_if_add(&skip, scop, 1);
4112 return scop;
4115 /* Try and construct a pet_scop for a label statement.
4116 * We currently only allow labels on expression statements.
4118 struct pet_scop *PetScan::extract(LabelStmt *stmt)
4120 isl_id *label;
4121 Stmt *sub;
4123 sub = stmt->getSubStmt();
4124 if (!isa<Expr>(sub)) {
4125 unsupported(stmt);
4126 return NULL;
4129 label = isl_id_alloc(ctx, stmt->getName(), NULL);
4131 return extract(sub, extract_expr(cast<Expr>(sub)), label);
4134 /* Return a one-dimensional multi piecewise affine expression that is equal
4135 * to the constant 1 and is defined over a zero-dimensional domain.
4137 static __isl_give isl_multi_pw_aff *one_mpa(isl_ctx *ctx)
4139 isl_space *space;
4140 isl_local_space *ls;
4141 isl_aff *aff;
4143 space = isl_space_set_alloc(ctx, 0, 0);
4144 ls = isl_local_space_from_space(space);
4145 aff = isl_aff_zero_on_domain(ls);
4146 aff = isl_aff_set_constant_si(aff, 1);
4148 return isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
4151 /* Construct a pet_scop for a continue statement.
4153 * We simply create an empty scop with a universal pet_skip_now
4154 * skip condition. This skip condition will then be taken into
4155 * account by the enclosing loop construct, possibly after
4156 * being incorporated into outer skip conditions.
4158 struct pet_scop *PetScan::extract(ContinueStmt *stmt)
4160 pet_scop *scop;
4162 scop = pet_scop_empty(ctx);
4163 if (!scop)
4164 return NULL;
4166 scop = pet_scop_set_skip(scop, pet_skip_now, one_mpa(ctx));
4168 return scop;
4171 /* Construct a pet_scop for a break statement.
4173 * We simply create an empty scop with both a universal pet_skip_now
4174 * skip condition and a universal pet_skip_later skip condition.
4175 * These skip conditions will then be taken into
4176 * account by the enclosing loop construct, possibly after
4177 * being incorporated into outer skip conditions.
4179 struct pet_scop *PetScan::extract(BreakStmt *stmt)
4181 pet_scop *scop;
4182 isl_multi_pw_aff *skip;
4184 scop = pet_scop_empty(ctx);
4185 if (!scop)
4186 return NULL;
4188 skip = one_mpa(ctx);
4189 scop = pet_scop_set_skip(scop, pet_skip_now,
4190 isl_multi_pw_aff_copy(skip));
4191 scop = pet_scop_set_skip(scop, pet_skip_later, skip);
4193 return scop;
4196 /* Try and construct a pet_scop corresponding to "stmt".
4198 * If "stmt" is a compound statement, then "skip_declarations"
4199 * indicates whether we should skip initial declarations in the
4200 * compound statement.
4202 * If the constructed pet_scop is not a (possibly) partial representation
4203 * of "stmt", we update start and end of the pet_scop to those of "stmt".
4204 * In particular, if skip_declarations is set, then we may have skipped
4205 * declarations inside "stmt" and so the pet_scop may not represent
4206 * the entire "stmt".
4207 * Note that this function may be called with "stmt" referring to the entire
4208 * body of the function, including the outer braces. In such cases,
4209 * skip_declarations will be set and the braces will not be taken into
4210 * account in scop->start and scop->end.
4212 struct pet_scop *PetScan::extract(Stmt *stmt, bool skip_declarations)
4214 struct pet_scop *scop;
4216 if (isa<Expr>(stmt))
4217 return extract(stmt, extract_expr(cast<Expr>(stmt)));
4219 switch (stmt->getStmtClass()) {
4220 case Stmt::WhileStmtClass:
4221 scop = extract(cast<WhileStmt>(stmt));
4222 break;
4223 case Stmt::ForStmtClass:
4224 scop = extract_for(cast<ForStmt>(stmt));
4225 break;
4226 case Stmt::IfStmtClass:
4227 scop = extract(cast<IfStmt>(stmt));
4228 break;
4229 case Stmt::CompoundStmtClass:
4230 scop = extract(cast<CompoundStmt>(stmt), skip_declarations);
4231 break;
4232 case Stmt::LabelStmtClass:
4233 scop = extract(cast<LabelStmt>(stmt));
4234 break;
4235 case Stmt::ContinueStmtClass:
4236 scop = extract(cast<ContinueStmt>(stmt));
4237 break;
4238 case Stmt::BreakStmtClass:
4239 scop = extract(cast<BreakStmt>(stmt));
4240 break;
4241 case Stmt::DeclStmtClass:
4242 scop = extract(cast<DeclStmt>(stmt));
4243 break;
4244 default:
4245 unsupported(stmt);
4246 return NULL;
4249 if (partial || skip_declarations)
4250 return scop;
4252 scop = update_scop_start_end(scop, stmt->getSourceRange(), false);
4254 return scop;
4257 /* Extract a clone of the kill statement in "scop".
4258 * "scop" is expected to have been created from a DeclStmt
4259 * and should have the kill as its first statement.
4261 struct pet_stmt *PetScan::extract_kill(struct pet_scop *scop)
4263 pet_expr *kill;
4264 struct pet_stmt *stmt;
4265 isl_multi_pw_aff *index;
4266 isl_map *access;
4267 pet_expr *arg;
4269 if (!scop)
4270 return NULL;
4271 if (scop->n_stmt < 1)
4272 isl_die(ctx, isl_error_internal,
4273 "expecting at least one statement", return NULL);
4274 stmt = scop->stmts[0];
4275 if (!pet_stmt_is_kill(stmt))
4276 isl_die(ctx, isl_error_internal,
4277 "expecting kill statement", return NULL);
4279 arg = pet_expr_get_arg(stmt->body, 0);
4280 index = pet_expr_access_get_index(arg);
4281 access = pet_expr_access_get_access(arg);
4282 pet_expr_free(arg);
4283 index = isl_multi_pw_aff_reset_tuple_id(index, isl_dim_in);
4284 access = isl_map_reset_tuple_id(access, isl_dim_in);
4285 kill = pet_expr_kill_from_access_and_index(access, index);
4286 return pet_stmt_from_pet_expr(stmt->line, NULL, n_stmt++, kill);
4289 /* Mark all arrays in "scop" as being exposed.
4291 static struct pet_scop *mark_exposed(struct pet_scop *scop)
4293 if (!scop)
4294 return NULL;
4295 for (int i = 0; i < scop->n_array; ++i)
4296 scop->arrays[i]->exposed = 1;
4297 return scop;
4300 /* Try and construct a pet_scop corresponding to (part of)
4301 * a sequence of statements.
4303 * "block" is set if the sequence respresents the children of
4304 * a compound statement.
4305 * "skip_declarations" is set if we should skip initial declarations
4306 * in the sequence of statements.
4308 * If there are any breaks or continues in the individual statements,
4309 * then we may have to compute a new skip condition.
4310 * This is handled using a pet_skip_info object.
4311 * On initialization, the object checks if skip conditions need
4312 * to be computed. If so, it does so in pet_skip_info_seq_extract and
4313 * adds them in pet_skip_info_seq_add.
4315 * If "block" is set, then we need to insert kill statements at
4316 * the end of the block for any array that has been declared by
4317 * one of the statements in the sequence. Each of these declarations
4318 * results in the construction of a kill statement at the place
4319 * of the declaration, so we simply collect duplicates of
4320 * those kill statements and append these duplicates to the constructed scop.
4322 * If "block" is not set, then any array declared by one of the statements
4323 * in the sequence is marked as being exposed.
4325 * If autodetect is set, then we allow the extraction of only a subrange
4326 * of the sequence of statements. However, if there is at least one statement
4327 * for which we could not construct a scop and the final range contains
4328 * either no statements or at least one kill, then we discard the entire
4329 * range.
4331 struct pet_scop *PetScan::extract(StmtRange stmt_range, bool block,
4332 bool skip_declarations)
4334 pet_scop *scop;
4335 StmtIterator i;
4336 int int_size;
4337 int j;
4338 bool partial_range = false;
4339 set<struct pet_stmt *> kills;
4340 set<struct pet_stmt *>::iterator it;
4342 int_size = ast_context.getTypeInfo(ast_context.IntTy).first / 8;
4344 scop = pet_scop_empty(ctx);
4345 for (i = stmt_range.first, j = 0; i != stmt_range.second; ++i, ++j) {
4346 Stmt *child = *i;
4347 struct pet_scop *scop_i;
4349 if (scop->n_stmt == 0 && skip_declarations &&
4350 child->getStmtClass() == Stmt::DeclStmtClass)
4351 continue;
4353 scop_i = extract(child);
4354 if (scop->n_stmt != 0 && partial) {
4355 pet_scop_free(scop_i);
4356 break;
4358 pet_skip_info skip;
4359 pet_skip_info_seq_init(&skip, ctx, scop, scop_i);
4360 pet_skip_info_seq_extract(&skip, int_size, &n_stmt, &n_test);
4361 if (pet_skip_info_has_skip(&skip))
4362 scop_i = pet_scop_prefix(scop_i, 0);
4363 if (scop_i && child->getStmtClass() == Stmt::DeclStmtClass) {
4364 if (block)
4365 kills.insert(extract_kill(scop_i));
4366 else
4367 scop_i = mark_exposed(scop_i);
4369 scop_i = pet_scop_prefix(scop_i, j);
4370 if (options->autodetect) {
4371 if (scop_i)
4372 scop = pet_scop_add_seq(ctx, scop, scop_i);
4373 else
4374 partial_range = true;
4375 if (scop->n_stmt != 0 && !scop_i)
4376 partial = true;
4377 } else {
4378 scop = pet_scop_add_seq(ctx, scop, scop_i);
4381 scop = pet_skip_info_seq_add(&skip, scop, j);
4383 if (partial || !scop)
4384 break;
4387 for (it = kills.begin(); it != kills.end(); ++it) {
4388 pet_scop *scop_j;
4389 scop_j = pet_scop_from_pet_stmt(ctx, *it);
4390 scop_j = pet_scop_prefix(scop_j, j);
4391 scop = pet_scop_add_seq(ctx, scop, scop_j);
4394 if (scop && partial_range) {
4395 if (scop->n_stmt == 0 || kills.size() != 0) {
4396 pet_scop_free(scop);
4397 return NULL;
4399 partial = true;
4402 return scop;
4405 /* Check if the scop marked by the user is exactly this Stmt
4406 * or part of this Stmt.
4407 * If so, return a pet_scop corresponding to the marked region.
4408 * Otherwise, return NULL.
4410 struct pet_scop *PetScan::scan(Stmt *stmt)
4412 SourceManager &SM = PP.getSourceManager();
4413 unsigned start_off, end_off;
4415 start_off = getExpansionOffset(SM, stmt->getLocStart());
4416 end_off = getExpansionOffset(SM, stmt->getLocEnd());
4418 if (start_off > loc.end)
4419 return NULL;
4420 if (end_off < loc.start)
4421 return NULL;
4422 if (start_off >= loc.start && end_off <= loc.end) {
4423 return extract(stmt);
4426 StmtIterator start;
4427 for (start = stmt->child_begin(); start != stmt->child_end(); ++start) {
4428 Stmt *child = *start;
4429 if (!child)
4430 continue;
4431 start_off = getExpansionOffset(SM, child->getLocStart());
4432 end_off = getExpansionOffset(SM, child->getLocEnd());
4433 if (start_off < loc.start && end_off >= loc.end)
4434 return scan(child);
4435 if (start_off >= loc.start)
4436 break;
4439 StmtIterator end;
4440 for (end = start; end != stmt->child_end(); ++end) {
4441 Stmt *child = *end;
4442 start_off = SM.getFileOffset(child->getLocStart());
4443 if (start_off >= loc.end)
4444 break;
4447 return extract(StmtRange(start, end), false, false);
4450 /* Set the size of index "pos" of "array" to "size".
4451 * In particular, add a constraint of the form
4453 * i_pos < size
4455 * to array->extent and a constraint of the form
4457 * size >= 0
4459 * to array->context.
4461 static struct pet_array *update_size(struct pet_array *array, int pos,
4462 __isl_take isl_pw_aff *size)
4464 isl_set *valid;
4465 isl_set *univ;
4466 isl_set *bound;
4467 isl_space *dim;
4468 isl_aff *aff;
4469 isl_pw_aff *index;
4470 isl_id *id;
4472 valid = isl_pw_aff_nonneg_set(isl_pw_aff_copy(size));
4473 array->context = isl_set_intersect(array->context, valid);
4475 dim = isl_set_get_space(array->extent);
4476 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
4477 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, pos, 1);
4478 univ = isl_set_universe(isl_aff_get_domain_space(aff));
4479 index = isl_pw_aff_alloc(univ, aff);
4481 size = isl_pw_aff_add_dims(size, isl_dim_in,
4482 isl_set_dim(array->extent, isl_dim_set));
4483 id = isl_set_get_tuple_id(array->extent);
4484 size = isl_pw_aff_set_tuple_id(size, isl_dim_in, id);
4485 bound = isl_pw_aff_lt_set(index, size);
4487 array->extent = isl_set_intersect(array->extent, bound);
4489 if (!array->context || !array->extent)
4490 goto error;
4492 return array;
4493 error:
4494 pet_array_free(array);
4495 return NULL;
4498 /* Figure out the size of the array at position "pos" and all
4499 * subsequent positions from "type" and update "array" accordingly.
4501 struct pet_array *PetScan::set_upper_bounds(struct pet_array *array,
4502 const Type *type, int pos)
4504 const ArrayType *atype;
4505 isl_pw_aff *size;
4507 if (!array)
4508 return NULL;
4510 if (type->isPointerType()) {
4511 type = type->getPointeeType().getTypePtr();
4512 return set_upper_bounds(array, type, pos + 1);
4514 if (!type->isArrayType())
4515 return array;
4517 type = type->getCanonicalTypeInternal().getTypePtr();
4518 atype = cast<ArrayType>(type);
4520 if (type->isConstantArrayType()) {
4521 const ConstantArrayType *ca = cast<ConstantArrayType>(atype);
4522 size = extract_affine(ca->getSize());
4523 array = update_size(array, pos, size);
4524 } else if (type->isVariableArrayType()) {
4525 const VariableArrayType *vla = cast<VariableArrayType>(atype);
4526 size = extract_affine(vla->getSizeExpr());
4527 array = update_size(array, pos, size);
4530 type = atype->getElementType().getTypePtr();
4532 return set_upper_bounds(array, type, pos + 1);
4535 /* Is "T" the type of a variable length array with static size?
4537 static bool is_vla_with_static_size(QualType T)
4539 const VariableArrayType *vlatype;
4541 if (!T->isVariableArrayType())
4542 return false;
4543 vlatype = cast<VariableArrayType>(T);
4544 return vlatype->getSizeModifier() == VariableArrayType::Static;
4547 /* Return the type of "decl" as an array.
4549 * In particular, if "decl" is a parameter declaration that
4550 * is a variable length array with a static size, then
4551 * return the original type (i.e., the variable length array).
4552 * Otherwise, return the type of decl.
4554 static QualType get_array_type(ValueDecl *decl)
4556 ParmVarDecl *parm;
4557 QualType T;
4559 parm = dyn_cast<ParmVarDecl>(decl);
4560 if (!parm)
4561 return decl->getType();
4563 T = parm->getOriginalType();
4564 if (!is_vla_with_static_size(T))
4565 return decl->getType();
4566 return T;
4569 /* Does "decl" have definition that we can keep track of in a pet_type?
4571 static bool has_printable_definition(RecordDecl *decl)
4573 if (!decl->getDeclName())
4574 return false;
4575 return decl->getLexicalDeclContext() == decl->getDeclContext();
4578 /* Construct and return a pet_array corresponding to the variable "decl".
4579 * In particular, initialize array->extent to
4581 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
4583 * and then call set_upper_bounds to set the upper bounds on the indices
4584 * based on the type of the variable.
4586 * If the base type is that of a record with a top-level definition and
4587 * if "types" is not null, then the RecordDecl corresponding to the type
4588 * is added to "types".
4590 * If the base type is that of a record with no top-level definition,
4591 * then we replace it by "<subfield>".
4593 struct pet_array *PetScan::extract_array(isl_ctx *ctx, ValueDecl *decl,
4594 lex_recorddecl_set *types)
4596 struct pet_array *array;
4597 QualType qt = get_array_type(decl);
4598 const Type *type = qt.getTypePtr();
4599 int depth = array_depth(type);
4600 QualType base = pet_clang_base_type(qt);
4601 string name;
4602 isl_id *id;
4603 isl_space *dim;
4605 array = isl_calloc_type(ctx, struct pet_array);
4606 if (!array)
4607 return NULL;
4609 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
4610 dim = isl_space_set_alloc(ctx, 0, depth);
4611 dim = isl_space_set_tuple_id(dim, isl_dim_set, id);
4613 array->extent = isl_set_nat_universe(dim);
4615 dim = isl_space_params_alloc(ctx, 0);
4616 array->context = isl_set_universe(dim);
4618 array = set_upper_bounds(array, type, 0);
4619 if (!array)
4620 return NULL;
4622 name = base.getAsString();
4624 if (types && base->isRecordType()) {
4625 RecordDecl *decl = pet_clang_record_decl(base);
4626 if (has_printable_definition(decl))
4627 types->insert(decl);
4628 else
4629 name = "<subfield>";
4632 array->element_type = strdup(name.c_str());
4633 array->element_is_record = base->isRecordType();
4634 array->element_size = decl->getASTContext().getTypeInfo(base).first / 8;
4636 return array;
4639 /* Construct and return a pet_array corresponding to the sequence
4640 * of declarations "decls".
4641 * If the sequence contains a single declaration, then it corresponds
4642 * to a simple array access. Otherwise, it corresponds to a member access,
4643 * with the declaration for the substructure following that of the containing
4644 * structure in the sequence of declarations.
4645 * We start with the outermost substructure and then combine it with
4646 * information from the inner structures.
4648 * Additionally, keep track of all required types in "types".
4650 struct pet_array *PetScan::extract_array(isl_ctx *ctx,
4651 vector<ValueDecl *> decls, lex_recorddecl_set *types)
4653 struct pet_array *array;
4654 vector<ValueDecl *>::iterator it;
4656 it = decls.begin();
4658 array = extract_array(ctx, *it, types);
4660 for (++it; it != decls.end(); ++it) {
4661 struct pet_array *parent;
4662 const char *base_name, *field_name;
4663 char *product_name;
4665 parent = array;
4666 array = extract_array(ctx, *it, types);
4667 if (!array)
4668 return pet_array_free(parent);
4670 base_name = isl_set_get_tuple_name(parent->extent);
4671 field_name = isl_set_get_tuple_name(array->extent);
4672 product_name = member_access_name(ctx, base_name, field_name);
4674 array->extent = isl_set_product(isl_set_copy(parent->extent),
4675 array->extent);
4676 if (product_name)
4677 array->extent = isl_set_set_tuple_name(array->extent,
4678 product_name);
4679 array->context = isl_set_intersect(array->context,
4680 isl_set_copy(parent->context));
4682 pet_array_free(parent);
4683 free(product_name);
4685 if (!array->extent || !array->context || !product_name)
4686 return pet_array_free(array);
4689 return array;
4692 /* Add a pet_type corresponding to "decl" to "scop, provided
4693 * it is a member of "types" and it has not been added before
4694 * (i.e., it is not a member of "types_done".
4696 * Since we want the user to be able to print the types
4697 * in the order in which they appear in the scop, we need to
4698 * make sure that types of fields in a structure appear before
4699 * that structure. We therefore call ourselves recursively
4700 * on the types of all record subfields.
4702 static struct pet_scop *add_type(isl_ctx *ctx, struct pet_scop *scop,
4703 RecordDecl *decl, Preprocessor &PP, lex_recorddecl_set &types,
4704 lex_recorddecl_set &types_done)
4706 string s;
4707 llvm::raw_string_ostream S(s);
4708 RecordDecl::field_iterator it;
4710 if (types.find(decl) == types.end())
4711 return scop;
4712 if (types_done.find(decl) != types_done.end())
4713 return scop;
4715 for (it = decl->field_begin(); it != decl->field_end(); ++it) {
4716 RecordDecl *record;
4717 QualType type = it->getType();
4719 if (!type->isRecordType())
4720 continue;
4721 record = pet_clang_record_decl(type);
4722 scop = add_type(ctx, scop, record, PP, types, types_done);
4725 if (strlen(decl->getName().str().c_str()) == 0)
4726 return scop;
4728 decl->print(S, PrintingPolicy(PP.getLangOpts()));
4729 S.str();
4731 scop->types[scop->n_type] = pet_type_alloc(ctx,
4732 decl->getName().str().c_str(), s.c_str());
4733 if (!scop->types[scop->n_type])
4734 return pet_scop_free(scop);
4736 types_done.insert(decl);
4738 scop->n_type++;
4740 return scop;
4743 /* Construct a list of pet_arrays, one for each array (or scalar)
4744 * accessed inside "scop", add this list to "scop" and return the result.
4746 * The context of "scop" is updated with the intersection of
4747 * the contexts of all arrays, i.e., constraints on the parameters
4748 * that ensure that the arrays have a valid (non-negative) size.
4750 * If the any of the extracted arrays refers to a member access,
4751 * then also add the required types to "scop".
4753 struct pet_scop *PetScan::scan_arrays(struct pet_scop *scop)
4755 int i;
4756 array_desc_set arrays;
4757 array_desc_set::iterator it;
4758 lex_recorddecl_set types;
4759 lex_recorddecl_set types_done;
4760 lex_recorddecl_set::iterator types_it;
4761 int n_array;
4762 struct pet_array **scop_arrays;
4764 if (!scop)
4765 return NULL;
4767 pet_scop_collect_arrays(scop, arrays);
4768 if (arrays.size() == 0)
4769 return scop;
4771 n_array = scop->n_array;
4773 scop_arrays = isl_realloc_array(ctx, scop->arrays, struct pet_array *,
4774 n_array + arrays.size());
4775 if (!scop_arrays)
4776 goto error;
4777 scop->arrays = scop_arrays;
4779 for (it = arrays.begin(), i = 0; it != arrays.end(); ++it, ++i) {
4780 struct pet_array *array;
4781 array = extract_array(ctx, *it, &types);
4782 scop->arrays[n_array + i] = array;
4783 if (!scop->arrays[n_array + i])
4784 goto error;
4785 scop->n_array++;
4786 scop->context = isl_set_intersect(scop->context,
4787 isl_set_copy(array->context));
4788 if (!scop->context)
4789 goto error;
4792 if (types.size() == 0)
4793 return scop;
4795 scop->types = isl_alloc_array(ctx, struct pet_type *, types.size());
4796 if (!scop->types)
4797 goto error;
4799 for (types_it = types.begin(); types_it != types.end(); ++types_it)
4800 scop = add_type(ctx, scop, *types_it, PP, types, types_done);
4802 return scop;
4803 error:
4804 pet_scop_free(scop);
4805 return NULL;
4808 /* Bound all parameters in scop->context to the possible values
4809 * of the corresponding C variable.
4811 static struct pet_scop *add_parameter_bounds(struct pet_scop *scop)
4813 int n;
4815 if (!scop)
4816 return NULL;
4818 n = isl_set_dim(scop->context, isl_dim_param);
4819 for (int i = 0; i < n; ++i) {
4820 isl_id *id;
4821 ValueDecl *decl;
4823 id = isl_set_get_dim_id(scop->context, isl_dim_param, i);
4824 if (pet_nested_in_id(id)) {
4825 isl_id_free(id);
4826 isl_die(isl_set_get_ctx(scop->context),
4827 isl_error_internal,
4828 "unresolved nested parameter", goto error);
4830 decl = (ValueDecl *) isl_id_get_user(id);
4831 isl_id_free(id);
4833 scop->context = set_parameter_bounds(scop->context, i, decl);
4835 if (!scop->context)
4836 goto error;
4839 return scop;
4840 error:
4841 pet_scop_free(scop);
4842 return NULL;
4845 /* Construct a pet_scop from the given function.
4847 * If the scop was delimited by scop and endscop pragmas, then we override
4848 * the file offsets by those derived from the pragmas.
4850 struct pet_scop *PetScan::scan(FunctionDecl *fd)
4852 pet_scop *scop;
4853 Stmt *stmt;
4855 stmt = fd->getBody();
4857 if (options->autodetect)
4858 scop = extract(stmt, true);
4859 else {
4860 scop = scan(stmt);
4861 scop = pet_scop_update_start_end(scop, loc.start, loc.end);
4863 scop = pet_scop_detect_parameter_accesses(scop);
4864 scop = scan_arrays(scop);
4865 scop = add_parameter_bounds(scop);
4866 scop = pet_scop_gist(scop, value_bounds);
4868 return scop;