treat fors with non-affine initialization or non-constant increment as whiles
[pet.git] / scan.cc
blobe4f50ae16c0fd33920e4fe19c0e81896a6e8f1de
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 "context.h"
53 #include "expr.h"
54 #include "nest.h"
55 #include "options.h"
56 #include "scan.h"
57 #include "scop.h"
58 #include "scop_plus.h"
59 #include "skip.h"
61 #include "config.h"
63 using namespace std;
64 using namespace clang;
66 static enum pet_op_type UnaryOperatorKind2pet_op_type(UnaryOperatorKind kind)
68 switch (kind) {
69 case UO_Minus:
70 return pet_op_minus;
71 case UO_Not:
72 return pet_op_not;
73 case UO_LNot:
74 return pet_op_lnot;
75 case UO_PostInc:
76 return pet_op_post_inc;
77 case UO_PostDec:
78 return pet_op_post_dec;
79 case UO_PreInc:
80 return pet_op_pre_inc;
81 case UO_PreDec:
82 return pet_op_pre_dec;
83 default:
84 return pet_op_last;
88 static enum pet_op_type BinaryOperatorKind2pet_op_type(BinaryOperatorKind kind)
90 switch (kind) {
91 case BO_AddAssign:
92 return pet_op_add_assign;
93 case BO_SubAssign:
94 return pet_op_sub_assign;
95 case BO_MulAssign:
96 return pet_op_mul_assign;
97 case BO_DivAssign:
98 return pet_op_div_assign;
99 case BO_Assign:
100 return pet_op_assign;
101 case BO_Add:
102 return pet_op_add;
103 case BO_Sub:
104 return pet_op_sub;
105 case BO_Mul:
106 return pet_op_mul;
107 case BO_Div:
108 return pet_op_div;
109 case BO_Rem:
110 return pet_op_mod;
111 case BO_Shl:
112 return pet_op_shl;
113 case BO_Shr:
114 return pet_op_shr;
115 case BO_EQ:
116 return pet_op_eq;
117 case BO_NE:
118 return pet_op_ne;
119 case BO_LE:
120 return pet_op_le;
121 case BO_GE:
122 return pet_op_ge;
123 case BO_LT:
124 return pet_op_lt;
125 case BO_GT:
126 return pet_op_gt;
127 case BO_And:
128 return pet_op_and;
129 case BO_Xor:
130 return pet_op_xor;
131 case BO_Or:
132 return pet_op_or;
133 case BO_LAnd:
134 return pet_op_land;
135 case BO_LOr:
136 return pet_op_lor;
137 default:
138 return pet_op_last;
142 #if defined(DECLREFEXPR_CREATE_REQUIRES_BOOL)
143 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
145 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
146 SourceLocation(), var, false, var->getInnerLocStart(),
147 var->getType(), VK_LValue);
149 #elif defined(DECLREFEXPR_CREATE_REQUIRES_SOURCELOCATION)
150 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
152 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
153 SourceLocation(), var, var->getInnerLocStart(), var->getType(),
154 VK_LValue);
156 #else
157 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
159 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
160 var, var->getInnerLocStart(), var->getType(), VK_LValue);
162 #endif
164 /* Check if the element type corresponding to the given array type
165 * has a const qualifier.
167 static bool const_base(QualType qt)
169 const Type *type = qt.getTypePtr();
171 if (type->isPointerType())
172 return const_base(type->getPointeeType());
173 if (type->isArrayType()) {
174 const ArrayType *atype;
175 type = type->getCanonicalTypeInternal().getTypePtr();
176 atype = cast<ArrayType>(type);
177 return const_base(atype->getElementType());
180 return qt.isConstQualified();
183 /* Create an isl_id that refers to the named declarator "decl".
185 static __isl_give isl_id *create_decl_id(isl_ctx *ctx, NamedDecl *decl)
187 return isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
190 /* Mark "decl" as having an unknown value in "assigned_value".
192 * If no (known or unknown) value was assigned to "decl" before,
193 * then it may have been treated as a parameter before and may
194 * therefore appear in a value assigned to another variable.
195 * If so, this assignment needs to be turned into an unknown value too.
197 static void clear_assignment(map<ValueDecl *, isl_pw_aff *> &assigned_value,
198 ValueDecl *decl)
200 map<ValueDecl *, isl_pw_aff *>::iterator it;
202 it = assigned_value.find(decl);
204 assigned_value[decl] = NULL;
206 if (it != assigned_value.end())
207 return;
209 for (it = assigned_value.begin(); it != assigned_value.end(); ++it) {
210 isl_pw_aff *pa = it->second;
211 int nparam = isl_pw_aff_dim(pa, isl_dim_param);
213 for (int i = 0; i < nparam; ++i) {
214 isl_id *id;
216 if (!isl_pw_aff_has_dim_id(pa, isl_dim_param, i))
217 continue;
218 id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
219 if (isl_id_get_user(id) == decl)
220 it->second = NULL;
221 isl_id_free(id);
226 /* Look for any assignments to scalar variables in part of the parse
227 * tree and set assigned_value to NULL for each of them.
228 * Also reset assigned_value if the address of a scalar variable
229 * is being taken. As an exception, if the address is passed to a function
230 * that is declared to receive a const pointer, then assigned_value is
231 * not reset.
233 * This ensures that we won't use any previously stored value
234 * in the current subtree and its parents.
236 struct clear_assignments : RecursiveASTVisitor<clear_assignments> {
237 map<ValueDecl *, isl_pw_aff *> &assigned_value;
238 set<UnaryOperator *> skip;
240 clear_assignments(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
241 assigned_value(assigned_value) {}
243 /* Check for "address of" operators whose value is passed
244 * to a const pointer argument and add them to "skip", so that
245 * we can skip them in VisitUnaryOperator.
247 bool VisitCallExpr(CallExpr *expr) {
248 FunctionDecl *fd;
249 fd = expr->getDirectCallee();
250 if (!fd)
251 return true;
252 for (int i = 0; i < expr->getNumArgs(); ++i) {
253 Expr *arg = expr->getArg(i);
254 UnaryOperator *op;
255 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
256 ImplicitCastExpr *ice;
257 ice = cast<ImplicitCastExpr>(arg);
258 arg = ice->getSubExpr();
260 if (arg->getStmtClass() != Stmt::UnaryOperatorClass)
261 continue;
262 op = cast<UnaryOperator>(arg);
263 if (op->getOpcode() != UO_AddrOf)
264 continue;
265 if (const_base(fd->getParamDecl(i)->getType()))
266 skip.insert(op);
268 return true;
271 bool VisitUnaryOperator(UnaryOperator *expr) {
272 Expr *arg;
273 DeclRefExpr *ref;
274 ValueDecl *decl;
276 switch (expr->getOpcode()) {
277 case UO_AddrOf:
278 case UO_PostInc:
279 case UO_PostDec:
280 case UO_PreInc:
281 case UO_PreDec:
282 break;
283 default:
284 return true;
286 if (skip.find(expr) != skip.end())
287 return true;
289 arg = expr->getSubExpr();
290 if (arg->getStmtClass() != Stmt::DeclRefExprClass)
291 return true;
292 ref = cast<DeclRefExpr>(arg);
293 decl = ref->getDecl();
294 clear_assignment(assigned_value, decl);
295 return true;
298 bool VisitBinaryOperator(BinaryOperator *expr) {
299 Expr *lhs;
300 DeclRefExpr *ref;
301 ValueDecl *decl;
303 if (!expr->isAssignmentOp())
304 return true;
305 lhs = expr->getLHS();
306 if (lhs->getStmtClass() != Stmt::DeclRefExprClass)
307 return true;
308 ref = cast<DeclRefExpr>(lhs);
309 decl = ref->getDecl();
310 clear_assignment(assigned_value, decl);
311 return true;
315 /* Keep a copy of the currently assigned values.
317 * Any variable that is assigned a value inside the current scope
318 * is removed again when we leave the scope (either because it wasn't
319 * stored in the cache or because it has a different value in the cache).
321 struct assigned_value_cache {
322 map<ValueDecl *, isl_pw_aff *> &assigned_value;
323 map<ValueDecl *, isl_pw_aff *> cache;
325 assigned_value_cache(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
326 assigned_value(assigned_value), cache(assigned_value) {}
327 ~assigned_value_cache() {
328 map<ValueDecl *, isl_pw_aff *>::iterator it = cache.begin();
329 for (it = assigned_value.begin(); it != assigned_value.end();
330 ++it) {
331 if (!it->second ||
332 (cache.find(it->first) != cache.end() &&
333 cache[it->first] != it->second))
334 cache[it->first] = NULL;
336 assigned_value = cache;
340 /* Convert the mapping from identifiers to values in "assigned_value"
341 * to a pet_context to be used by pet_expr_extract_*.
342 * In particular, the clang identifiers are wrapped in an isl_id and
343 * a NULL value (representing an unknown value) is replaced by a NaN.
345 static __isl_give pet_context *convert_assignments(isl_ctx *ctx,
346 map<ValueDecl *, isl_pw_aff *> &assigned_value)
348 pet_context *pc;
349 map<ValueDecl *, isl_pw_aff *>::iterator it;
351 pc = pet_context_alloc(isl_space_set_alloc(ctx, 0, 0));
353 for (it = assigned_value.begin(); it != assigned_value.end(); ++it) {
354 ValueDecl *decl = it->first;
355 isl_pw_aff *pa = it->second;
356 isl_id *id;
358 id = create_decl_id(ctx, decl);
359 if (pa)
360 pc = pet_context_set_value(pc, id, isl_pw_aff_copy(pa));
361 else
362 pc = pet_context_mark_unknown(pc, id);
365 return pc;
368 /* Insert an expression into the collection of expressions,
369 * provided it is not already in there.
370 * The isl_pw_affs are freed in the destructor.
372 void PetScan::insert_expression(__isl_take isl_pw_aff *expr)
374 std::set<isl_pw_aff *>::iterator it;
376 if (expressions.find(expr) == expressions.end())
377 expressions.insert(expr);
378 else
379 isl_pw_aff_free(expr);
382 PetScan::~PetScan()
384 std::set<isl_pw_aff *>::iterator it;
386 for (it = expressions.begin(); it != expressions.end(); ++it)
387 isl_pw_aff_free(*it);
389 isl_union_map_free(value_bounds);
392 /* Report a diagnostic, unless autodetect is set.
394 void PetScan::report(Stmt *stmt, unsigned id)
396 if (options->autodetect)
397 return;
399 SourceLocation loc = stmt->getLocStart();
400 DiagnosticsEngine &diag = PP.getDiagnostics();
401 DiagnosticBuilder B = diag.Report(loc, id) << stmt->getSourceRange();
404 /* Called if we found something we (currently) cannot handle.
405 * We'll provide more informative warnings later.
407 * We only actually complain if autodetect is false.
409 void PetScan::unsupported(Stmt *stmt)
411 DiagnosticsEngine &diag = PP.getDiagnostics();
412 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
413 "unsupported");
414 report(stmt, id);
417 /* Report a missing prototype, unless autodetect is set.
419 void PetScan::report_prototype_required(Stmt *stmt)
421 DiagnosticsEngine &diag = PP.getDiagnostics();
422 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
423 "prototype required");
424 report(stmt, id);
427 /* Report a missing increment, unless autodetect is set.
429 void PetScan::report_missing_increment(Stmt *stmt)
431 DiagnosticsEngine &diag = PP.getDiagnostics();
432 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
433 "missing increment");
434 report(stmt, id);
437 /* Extract an integer from "expr".
439 __isl_give isl_val *PetScan::extract_int(isl_ctx *ctx, IntegerLiteral *expr)
441 const Type *type = expr->getType().getTypePtr();
442 int is_signed = type->hasSignedIntegerRepresentation();
443 llvm::APInt val = expr->getValue();
444 int is_negative = is_signed && val.isNegative();
445 isl_val *v;
447 if (is_negative)
448 val = -val;
450 v = extract_unsigned(ctx, val);
452 if (is_negative)
453 v = isl_val_neg(v);
454 return v;
457 /* Extract an integer from "val", which is assumed to be non-negative.
459 __isl_give isl_val *PetScan::extract_unsigned(isl_ctx *ctx,
460 const llvm::APInt &val)
462 unsigned n;
463 const uint64_t *data;
465 data = val.getRawData();
466 n = val.getNumWords();
467 return isl_val_int_from_chunks(ctx, n, sizeof(uint64_t), data);
470 /* Extract an integer from "expr".
471 * Return NULL if "expr" does not (obviously) represent an integer.
473 __isl_give isl_val *PetScan::extract_int(clang::ParenExpr *expr)
475 return extract_int(expr->getSubExpr());
478 /* Extract an integer from "expr".
479 * Return NULL if "expr" does not (obviously) represent an integer.
481 __isl_give isl_val *PetScan::extract_int(clang::Expr *expr)
483 if (expr->getStmtClass() == Stmt::IntegerLiteralClass)
484 return extract_int(ctx, cast<IntegerLiteral>(expr));
485 if (expr->getStmtClass() == Stmt::ParenExprClass)
486 return extract_int(cast<ParenExpr>(expr));
488 unsupported(expr);
489 return NULL;
492 /* Extract an affine expression from the IntegerLiteral "expr".
493 * If the value of "expr" is "v", then the returned expression
494 * is
496 * { [] -> [v] }
498 __isl_give isl_pw_aff *PetScan::extract_affine(IntegerLiteral *expr)
500 isl_space *space = isl_space_set_alloc(ctx, 0, 0);
501 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(space));
502 isl_aff *aff = isl_aff_zero_on_domain(ls);
503 isl_set *dom = isl_set_universe(space);
504 isl_val *v;
506 v = extract_int(expr);
507 aff = isl_aff_add_constant_val(aff, v);
509 return isl_pw_aff_alloc(dom, aff);
512 /* Extract an affine expression from the APInt "val", which is assumed
513 * to be non-negative.
514 * If the value of "val" is "v", then the returned expression
515 * is
517 * { [] -> [v] }
519 __isl_give isl_pw_aff *PetScan::extract_affine(const llvm::APInt &val)
521 isl_space *space = isl_space_set_alloc(ctx, 0, 0);
522 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(space));
523 isl_aff *aff = isl_aff_zero_on_domain(ls);
524 isl_set *dom = isl_set_universe(space);
525 isl_val *v;
527 v = extract_unsigned(ctx, val);
528 aff = isl_aff_add_constant_val(aff, v);
530 return isl_pw_aff_alloc(dom, aff);
533 __isl_give isl_pw_aff *PetScan::extract_affine(ImplicitCastExpr *expr)
535 return extract_affine(expr->getSubExpr());
538 /* Return the number of bits needed to represent the type "qt",
539 * if it is an integer type. Otherwise return 0.
540 * If qt is signed then return the opposite of the number of bits.
542 static int get_type_size(QualType qt, ASTContext &ast_context)
544 int size;
546 if (!qt->isIntegerType())
547 return 0;
549 size = ast_context.getIntWidth(qt);
550 if (!qt->isUnsignedIntegerType())
551 size = -size;
553 return size;
556 /* Return the number of bits needed to represent the type of "decl",
557 * if it is an integer type. Otherwise return 0.
558 * If qt is signed then return the opposite of the number of bits.
560 static int get_type_size(ValueDecl *decl)
562 return get_type_size(decl->getType(), decl->getASTContext());
565 /* Bound parameter "pos" of "set" to the possible values of "decl".
567 static __isl_give isl_set *set_parameter_bounds(__isl_take isl_set *set,
568 unsigned pos, ValueDecl *decl)
570 int type_size;
571 isl_ctx *ctx;
572 isl_val *bound;
574 ctx = isl_set_get_ctx(set);
575 type_size = get_type_size(decl);
576 if (type_size == 0)
577 isl_die(ctx, isl_error_invalid, "not an integer type",
578 return isl_set_free(set));
579 if (type_size > 0) {
580 set = isl_set_lower_bound_si(set, isl_dim_param, pos, 0);
581 bound = isl_val_int_from_ui(ctx, type_size);
582 bound = isl_val_2exp(bound);
583 bound = isl_val_sub_ui(bound, 1);
584 set = isl_set_upper_bound_val(set, isl_dim_param, pos, bound);
585 } else {
586 bound = isl_val_int_from_ui(ctx, -type_size - 1);
587 bound = isl_val_2exp(bound);
588 bound = isl_val_sub_ui(bound, 1);
589 set = isl_set_upper_bound_val(set, isl_dim_param, pos,
590 isl_val_copy(bound));
591 bound = isl_val_neg(bound);
592 bound = isl_val_sub_ui(bound, 1);
593 set = isl_set_lower_bound_val(set, isl_dim_param, pos, bound);
596 return set;
599 /* Extract an affine expression from the DeclRefExpr "expr".
601 * If the variable has been assigned a value, then we check whether
602 * we know what (affine) value was assigned.
603 * If so, we return this value. Otherwise we convert "expr"
604 * to an extra parameter (provided nesting_enabled is set).
606 * Otherwise, we simply return an expression that is equal
607 * to a parameter corresponding to the referenced variable.
609 __isl_give isl_pw_aff *PetScan::extract_affine(DeclRefExpr *expr)
611 ValueDecl *decl = expr->getDecl();
612 const Type *type = decl->getType().getTypePtr();
613 isl_id *id;
614 isl_space *space;
615 isl_aff *aff;
616 isl_set *dom;
618 if (!type->isIntegerType()) {
619 unsupported(expr);
620 return NULL;
623 if (assigned_value.find(decl) != assigned_value.end()) {
624 if (assigned_value[decl])
625 return isl_pw_aff_copy(assigned_value[decl]);
626 else
627 return nested_access(expr);
630 id = create_decl_id(ctx, decl);
631 space = isl_space_set_alloc(ctx, 1, 0);
633 space = isl_space_set_dim_id(space, isl_dim_param, 0, id);
635 dom = isl_set_universe(isl_space_copy(space));
636 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
637 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
639 return isl_pw_aff_alloc(dom, aff);
642 /* Extract an affine expression from an integer division operation.
643 * In particular, if "expr" is lhs/rhs, then return
645 * lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs)
647 * The second argument (rhs) is required to be a (positive) integer constant.
649 __isl_give isl_pw_aff *PetScan::extract_affine_div(BinaryOperator *expr)
651 int is_cst;
652 isl_pw_aff *rhs, *lhs;
654 rhs = extract_affine(expr->getRHS());
655 is_cst = isl_pw_aff_is_cst(rhs);
656 if (is_cst < 0 || !is_cst) {
657 isl_pw_aff_free(rhs);
658 if (!is_cst)
659 unsupported(expr);
660 return NULL;
663 lhs = extract_affine(expr->getLHS());
665 return isl_pw_aff_tdiv_q(lhs, rhs);
668 /* Extract an affine expression from a modulo operation.
669 * In particular, if "expr" is lhs/rhs, then return
671 * lhs - rhs * (lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs))
673 * The second argument (rhs) is required to be a (positive) integer constant.
675 __isl_give isl_pw_aff *PetScan::extract_affine_mod(BinaryOperator *expr)
677 int is_cst;
678 isl_pw_aff *rhs, *lhs;
680 rhs = extract_affine(expr->getRHS());
681 is_cst = isl_pw_aff_is_cst(rhs);
682 if (is_cst < 0 || !is_cst) {
683 isl_pw_aff_free(rhs);
684 if (!is_cst)
685 unsupported(expr);
686 return NULL;
689 lhs = extract_affine(expr->getLHS());
691 return isl_pw_aff_tdiv_r(lhs, rhs);
694 /* Extract an affine expression from a multiplication operation.
695 * This is only allowed if at least one of the two arguments
696 * is a (piecewise) constant.
698 __isl_give isl_pw_aff *PetScan::extract_affine_mul(BinaryOperator *expr)
700 isl_pw_aff *lhs;
701 isl_pw_aff *rhs;
703 lhs = extract_affine(expr->getLHS());
704 rhs = extract_affine(expr->getRHS());
706 if (!isl_pw_aff_is_cst(lhs) && !isl_pw_aff_is_cst(rhs)) {
707 isl_pw_aff_free(lhs);
708 isl_pw_aff_free(rhs);
709 unsupported(expr);
710 return NULL;
713 return isl_pw_aff_mul(lhs, rhs);
716 /* Extract an affine expression from an addition or subtraction operation.
718 __isl_give isl_pw_aff *PetScan::extract_affine_add(BinaryOperator *expr)
720 isl_pw_aff *lhs;
721 isl_pw_aff *rhs;
723 lhs = extract_affine(expr->getLHS());
724 rhs = extract_affine(expr->getRHS());
726 switch (expr->getOpcode()) {
727 case BO_Add:
728 return isl_pw_aff_add(lhs, rhs);
729 case BO_Sub:
730 return isl_pw_aff_sub(lhs, rhs);
731 default:
732 isl_pw_aff_free(lhs);
733 isl_pw_aff_free(rhs);
734 return NULL;
739 /* Compute
741 * pwaff mod 2^width
743 static __isl_give isl_pw_aff *wrap(__isl_take isl_pw_aff *pwaff,
744 unsigned width)
746 isl_ctx *ctx;
747 isl_val *mod;
749 ctx = isl_pw_aff_get_ctx(pwaff);
750 mod = isl_val_int_from_ui(ctx, width);
751 mod = isl_val_2exp(mod);
753 pwaff = isl_pw_aff_mod_val(pwaff, mod);
755 return pwaff;
758 /* Limit the domain of "pwaff" to those elements where the function
759 * value satisfies
761 * 2^{width-1} <= pwaff < 2^{width-1}
763 static __isl_give isl_pw_aff *avoid_overflow(__isl_take isl_pw_aff *pwaff,
764 unsigned width)
766 isl_ctx *ctx;
767 isl_val *v;
768 isl_space *space = isl_pw_aff_get_domain_space(pwaff);
769 isl_local_space *ls = isl_local_space_from_space(space);
770 isl_aff *bound;
771 isl_set *dom;
772 isl_pw_aff *b;
774 ctx = isl_pw_aff_get_ctx(pwaff);
775 v = isl_val_int_from_ui(ctx, width - 1);
776 v = isl_val_2exp(v);
778 bound = isl_aff_zero_on_domain(ls);
779 bound = isl_aff_add_constant_val(bound, v);
780 b = isl_pw_aff_from_aff(bound);
782 dom = isl_pw_aff_lt_set(isl_pw_aff_copy(pwaff), isl_pw_aff_copy(b));
783 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
785 b = isl_pw_aff_neg(b);
786 dom = isl_pw_aff_ge_set(isl_pw_aff_copy(pwaff), b);
787 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
789 return pwaff;
792 /* Handle potential overflows on signed computations.
794 * If options->signed_overflow is set to PET_OVERFLOW_AVOID,
795 * the we adjust the domain of "pa" to avoid overflows.
797 __isl_give isl_pw_aff *PetScan::signed_overflow(__isl_take isl_pw_aff *pa,
798 unsigned width)
800 if (options->signed_overflow == PET_OVERFLOW_AVOID)
801 pa = avoid_overflow(pa, width);
803 return pa;
806 /* Return the piecewise affine expression "set ? 1 : 0" defined on "dom".
808 static __isl_give isl_pw_aff *indicator_function(__isl_take isl_set *set,
809 __isl_take isl_set *dom)
811 isl_pw_aff *pa;
812 pa = isl_set_indicator_function(set);
813 pa = isl_pw_aff_intersect_domain(pa, isl_set_coalesce(dom));
814 return pa;
817 /* Extract an affine expression from some binary operations.
818 * If the result of the expression is unsigned, then we wrap it
819 * based on the size of the type. Otherwise, we ensure that
820 * no overflow occurs.
822 __isl_give isl_pw_aff *PetScan::extract_affine(BinaryOperator *expr)
824 isl_pw_aff *res;
825 unsigned width;
827 switch (expr->getOpcode()) {
828 case BO_Add:
829 case BO_Sub:
830 res = extract_affine_add(expr);
831 break;
832 case BO_Div:
833 res = extract_affine_div(expr);
834 break;
835 case BO_Rem:
836 res = extract_affine_mod(expr);
837 break;
838 case BO_Mul:
839 res = extract_affine_mul(expr);
840 break;
841 case BO_LT:
842 case BO_LE:
843 case BO_GT:
844 case BO_GE:
845 case BO_EQ:
846 case BO_NE:
847 case BO_LAnd:
848 case BO_LOr:
849 return extract_condition(expr);
850 default:
851 unsupported(expr);
852 return NULL;
855 width = ast_context.getIntWidth(expr->getType());
856 if (expr->getType()->isUnsignedIntegerType())
857 res = wrap(res, width);
858 else
859 res = signed_overflow(res, width);
861 return res;
864 /* Extract an affine expression from a negation operation.
866 __isl_give isl_pw_aff *PetScan::extract_affine(UnaryOperator *expr)
868 if (expr->getOpcode() == UO_Minus)
869 return isl_pw_aff_neg(extract_affine(expr->getSubExpr()));
870 if (expr->getOpcode() == UO_LNot)
871 return extract_condition(expr);
873 unsupported(expr);
874 return NULL;
877 __isl_give isl_pw_aff *PetScan::extract_affine(ParenExpr *expr)
879 return extract_affine(expr->getSubExpr());
882 /* Extract an affine expression from some special function calls.
883 * In particular, we handle "min", "max", "ceild", "floord",
884 * "intMod", "intFloor" and "intCeil".
885 * In case of the latter five, the second argument needs to be
886 * a (positive) integer constant.
888 __isl_give isl_pw_aff *PetScan::extract_affine(CallExpr *expr)
890 FunctionDecl *fd;
891 string name;
892 isl_pw_aff *aff1, *aff2;
894 fd = expr->getDirectCallee();
895 if (!fd) {
896 unsupported(expr);
897 return NULL;
900 name = fd->getDeclName().getAsString();
901 if (!(expr->getNumArgs() == 2 && name == "min") &&
902 !(expr->getNumArgs() == 2 && name == "max") &&
903 !(expr->getNumArgs() == 2 && name == "intMod") &&
904 !(expr->getNumArgs() == 2 && name == "intFloor") &&
905 !(expr->getNumArgs() == 2 && name == "intCeil") &&
906 !(expr->getNumArgs() == 2 && name == "floord") &&
907 !(expr->getNumArgs() == 2 && name == "ceild")) {
908 unsupported(expr);
909 return NULL;
912 if (name == "min" || name == "max") {
913 aff1 = extract_affine(expr->getArg(0));
914 aff2 = extract_affine(expr->getArg(1));
916 if (name == "min")
917 aff1 = isl_pw_aff_min(aff1, aff2);
918 else
919 aff1 = isl_pw_aff_max(aff1, aff2);
920 } else if (name == "intMod") {
921 isl_val *v;
922 Expr *arg2 = expr->getArg(1);
924 if (arg2->getStmtClass() != Stmt::IntegerLiteralClass) {
925 unsupported(expr);
926 return NULL;
928 aff1 = extract_affine(expr->getArg(0));
929 v = extract_int(cast<IntegerLiteral>(arg2));
930 aff1 = isl_pw_aff_mod_val(aff1, v);
931 } else if (name == "floord" || name == "ceild" ||
932 name == "intFloor" || name == "intCeil") {
933 isl_val *v;
934 Expr *arg2 = expr->getArg(1);
936 if (arg2->getStmtClass() != Stmt::IntegerLiteralClass) {
937 unsupported(expr);
938 return NULL;
940 aff1 = extract_affine(expr->getArg(0));
941 v = extract_int(cast<IntegerLiteral>(arg2));
942 aff1 = isl_pw_aff_scale_down_val(aff1, v);
943 if (name == "floord" || name == "intFloor")
944 aff1 = isl_pw_aff_floor(aff1);
945 else
946 aff1 = isl_pw_aff_ceil(aff1);
947 } else {
948 unsupported(expr);
949 return NULL;
952 return aff1;
955 /* This method is called when we come across an access that is
956 * nested in what is supposed to be an affine expression.
957 * If nesting is allowed, we return a new parameter that corresponds
958 * to this nested access. Otherwise, we simply complain.
960 * Note that we currently don't allow nested accesses themselves
961 * to contain any nested accesses, so we check if we can extract
962 * the access without any nesting and complain if we can't.
964 * The new parameter is resolved in resolve_nested.
966 isl_pw_aff *PetScan::nested_access(Expr *expr)
968 isl_id *id;
969 isl_space *space;
970 isl_aff *aff;
971 isl_set *dom;
972 isl_multi_pw_aff *index;
974 if (!nesting_enabled) {
975 unsupported(expr);
976 return NULL;
979 allow_nested = false;
980 index = extract_index(expr);
981 allow_nested = true;
982 if (!index) {
983 unsupported(expr);
984 return NULL;
986 isl_multi_pw_aff_free(index);
988 id = pet_nested_clang_expr(ctx, expr);
989 space = isl_space_set_alloc(ctx, 1, 0);
991 space = isl_space_set_dim_id(space, isl_dim_param, 0, id);
993 dom = isl_set_universe(isl_space_copy(space));
994 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
995 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
997 return isl_pw_aff_alloc(dom, aff);
1000 /* Affine expressions are not supposed to contain array accesses,
1001 * but if nesting is allowed, we return a parameter corresponding
1002 * to the array access.
1004 __isl_give isl_pw_aff *PetScan::extract_affine(ArraySubscriptExpr *expr)
1006 return nested_access(expr);
1009 /* Affine expressions are not supposed to contain member accesses,
1010 * but if nesting is allowed, we return a parameter corresponding
1011 * to the member access.
1013 __isl_give isl_pw_aff *PetScan::extract_affine(MemberExpr *expr)
1015 return nested_access(expr);
1018 /* Extract an affine expression from a conditional operation.
1020 __isl_give isl_pw_aff *PetScan::extract_affine(ConditionalOperator *expr)
1022 isl_pw_aff *cond, *lhs, *rhs;
1024 cond = extract_condition(expr->getCond());
1025 lhs = extract_affine(expr->getTrueExpr());
1026 rhs = extract_affine(expr->getFalseExpr());
1028 return isl_pw_aff_cond(cond, lhs, rhs);
1031 /* Extract an affine expression, if possible, from "expr".
1032 * Otherwise return NULL.
1034 * The result has an anonymous zero-dimensional domain.
1036 __isl_give isl_pw_aff *PetScan::extract_affine(Expr *expr)
1038 switch (expr->getStmtClass()) {
1039 case Stmt::ImplicitCastExprClass:
1040 return extract_affine(cast<ImplicitCastExpr>(expr));
1041 case Stmt::IntegerLiteralClass:
1042 return extract_affine(cast<IntegerLiteral>(expr));
1043 case Stmt::DeclRefExprClass:
1044 return extract_affine(cast<DeclRefExpr>(expr));
1045 case Stmt::BinaryOperatorClass:
1046 return extract_affine(cast<BinaryOperator>(expr));
1047 case Stmt::UnaryOperatorClass:
1048 return extract_affine(cast<UnaryOperator>(expr));
1049 case Stmt::ParenExprClass:
1050 return extract_affine(cast<ParenExpr>(expr));
1051 case Stmt::CallExprClass:
1052 return extract_affine(cast<CallExpr>(expr));
1053 case Stmt::ArraySubscriptExprClass:
1054 return extract_affine(cast<ArraySubscriptExpr>(expr));
1055 case Stmt::MemberExprClass:
1056 return extract_affine(cast<MemberExpr>(expr));
1057 case Stmt::ConditionalOperatorClass:
1058 return extract_affine(cast<ConditionalOperator>(expr));
1059 default:
1060 unsupported(expr);
1062 return NULL;
1065 __isl_give isl_multi_pw_aff *PetScan::extract_index(ImplicitCastExpr *expr)
1067 return extract_index(expr->getSubExpr());
1070 /* Return the depth of an array of the given type.
1072 static int array_depth(const Type *type)
1074 if (type->isPointerType())
1075 return 1 + array_depth(type->getPointeeType().getTypePtr());
1076 if (type->isArrayType()) {
1077 const ArrayType *atype;
1078 type = type->getCanonicalTypeInternal().getTypePtr();
1079 atype = cast<ArrayType>(type);
1080 return 1 + array_depth(atype->getElementType().getTypePtr());
1082 return 0;
1085 /* Return the depth of the array accessed by the index expression "index".
1086 * If "index" is an affine expression, i.e., if it does not access
1087 * any array, then return 1.
1088 * If "index" represent a member access, i.e., if its range is a wrapped
1089 * relation, then return the sum of the depth of the array of structures
1090 * and that of the member inside the structure.
1092 static int extract_depth(__isl_keep isl_multi_pw_aff *index)
1094 isl_id *id;
1095 ValueDecl *decl;
1097 if (!index)
1098 return -1;
1100 if (isl_multi_pw_aff_range_is_wrapping(index)) {
1101 int domain_depth, range_depth;
1102 isl_multi_pw_aff *domain, *range;
1104 domain = isl_multi_pw_aff_copy(index);
1105 domain = isl_multi_pw_aff_range_factor_domain(domain);
1106 domain_depth = extract_depth(domain);
1107 isl_multi_pw_aff_free(domain);
1108 range = isl_multi_pw_aff_copy(index);
1109 range = isl_multi_pw_aff_range_factor_range(range);
1110 range_depth = extract_depth(range);
1111 isl_multi_pw_aff_free(range);
1113 return domain_depth + range_depth;
1116 if (!isl_multi_pw_aff_has_tuple_id(index, isl_dim_out))
1117 return 1;
1119 id = isl_multi_pw_aff_get_tuple_id(index, isl_dim_out);
1120 if (!id)
1121 return -1;
1122 decl = (ValueDecl *) isl_id_get_user(id);
1123 isl_id_free(id);
1125 return array_depth(decl->getType().getTypePtr());
1128 /* Extract an index expression from a reference to a variable.
1129 * If the variable has name "A", then the returned index expression
1130 * is of the form
1132 * { [] -> A[] }
1134 __isl_give isl_multi_pw_aff *PetScan::extract_index(DeclRefExpr *expr)
1136 return extract_index(expr->getDecl());
1139 /* Extract an index expression from a variable.
1140 * If the variable has name "A", then the returned index expression
1141 * is of the form
1143 * { [] -> A[] }
1145 __isl_give isl_multi_pw_aff *PetScan::extract_index(ValueDecl *decl)
1147 isl_id *id = create_decl_id(ctx, decl);
1148 isl_space *space = isl_space_alloc(ctx, 0, 0, 0);
1150 space = isl_space_set_tuple_id(space, isl_dim_out, id);
1152 return isl_multi_pw_aff_zero(space);
1155 /* Extract an index expression from an integer contant.
1156 * If the value of the constant is "v", then the returned access relation
1157 * is
1159 * { [] -> [v] }
1161 __isl_give isl_multi_pw_aff *PetScan::extract_index(IntegerLiteral *expr)
1163 isl_multi_pw_aff *mpa;
1165 mpa = isl_multi_pw_aff_from_pw_aff(extract_affine(expr));
1166 return mpa;
1169 /* Try and extract an index expression from the given Expr.
1170 * Return NULL if it doesn't work out.
1172 __isl_give isl_multi_pw_aff *PetScan::extract_index(Expr *expr)
1174 switch (expr->getStmtClass()) {
1175 case Stmt::ImplicitCastExprClass:
1176 return extract_index(cast<ImplicitCastExpr>(expr));
1177 case Stmt::DeclRefExprClass:
1178 return extract_index(cast<DeclRefExpr>(expr));
1179 case Stmt::ArraySubscriptExprClass:
1180 return extract_index(cast<ArraySubscriptExpr>(expr));
1181 case Stmt::IntegerLiteralClass:
1182 return extract_index(cast<IntegerLiteral>(expr));
1183 case Stmt::MemberExprClass:
1184 return extract_index(cast<MemberExpr>(expr));
1185 default:
1186 unsupported(expr);
1188 return NULL;
1191 /* Given a partial index expression "base" and an extra index "index",
1192 * append the extra index to "base" and return the result.
1193 * Additionally, add the constraints that the extra index is non-negative.
1194 * If "index" represent a member access, i.e., if its range is a wrapped
1195 * relation, then we recursively extend the range of this nested relation.
1197 * The inputs "base" and "index", as well as the result, all have
1198 * an anonymous zero-dimensional domain.
1200 static __isl_give isl_multi_pw_aff *subscript(__isl_take isl_multi_pw_aff *base,
1201 __isl_take isl_pw_aff *index)
1203 isl_id *id;
1204 isl_set *domain;
1205 isl_multi_pw_aff *access;
1206 int member_access;
1208 member_access = isl_multi_pw_aff_range_is_wrapping(base);
1209 if (member_access < 0)
1210 goto error;
1211 if (member_access) {
1212 isl_multi_pw_aff *domain, *range;
1213 isl_id *id;
1215 id = isl_multi_pw_aff_get_tuple_id(base, isl_dim_out);
1216 domain = isl_multi_pw_aff_copy(base);
1217 domain = isl_multi_pw_aff_range_factor_domain(domain);
1218 range = isl_multi_pw_aff_range_factor_range(base);
1219 range = subscript(range, index);
1220 access = isl_multi_pw_aff_range_product(domain, range);
1221 access = isl_multi_pw_aff_set_tuple_id(access, isl_dim_out, id);
1222 return access;
1225 id = isl_multi_pw_aff_get_tuple_id(base, isl_dim_set);
1226 domain = isl_pw_aff_nonneg_set(isl_pw_aff_copy(index));
1227 index = isl_pw_aff_intersect_domain(index, domain);
1228 access = isl_multi_pw_aff_from_pw_aff(index);
1229 access = isl_multi_pw_aff_flat_range_product(base, access);
1230 access = isl_multi_pw_aff_set_tuple_id(access, isl_dim_set, id);
1232 return access;
1233 error:
1234 isl_multi_pw_aff_free(base);
1235 isl_pw_aff_free(index);
1236 return NULL;
1239 /* Extract an index expression from the given array subscript expression.
1240 * If nesting is allowed in general, then we turn it on while
1241 * examining the index expression.
1243 * We first extract an index expression from the base.
1244 * This will result in an index expression with a range that corresponds
1245 * to the earlier indices.
1246 * We then extract the current index, restrict its domain
1247 * to those values that result in a non-negative index and
1248 * append the index to the base index expression.
1250 __isl_give isl_multi_pw_aff *PetScan::extract_index(ArraySubscriptExpr *expr)
1252 Expr *base = expr->getBase();
1253 Expr *idx = expr->getIdx();
1254 isl_pw_aff *index;
1255 isl_multi_pw_aff *base_access;
1256 isl_multi_pw_aff *access;
1257 bool save_nesting = nesting_enabled;
1259 nesting_enabled = allow_nested;
1261 base_access = extract_index(base);
1262 index = extract_affine(idx);
1264 nesting_enabled = save_nesting;
1266 access = subscript(base_access, index);
1268 return access;
1271 /* Construct a name for a member access by concatenating the name
1272 * of the array of structures and the member, separated by an underscore.
1274 * The caller is responsible for freeing the result.
1276 static char *member_access_name(isl_ctx *ctx, const char *base,
1277 const char *field)
1279 int len;
1280 char *name;
1282 len = strlen(base) + 1 + strlen(field);
1283 name = isl_alloc_array(ctx, char, len + 1);
1284 if (!name)
1285 return NULL;
1286 snprintf(name, len + 1, "%s_%s", base, field);
1288 return name;
1291 /* Given an index expression "base" for an element of an array of structures
1292 * and an expression "field" for the field member being accessed, construct
1293 * an index expression for an access to that member of the given structure.
1294 * In particular, take the range product of "base" and "field" and
1295 * attach a name to the result.
1297 static __isl_give isl_multi_pw_aff *member(__isl_take isl_multi_pw_aff *base,
1298 __isl_take isl_multi_pw_aff *field)
1300 isl_ctx *ctx;
1301 isl_multi_pw_aff *access;
1302 const char *base_name, *field_name;
1303 char *name;
1305 ctx = isl_multi_pw_aff_get_ctx(base);
1307 base_name = isl_multi_pw_aff_get_tuple_name(base, isl_dim_out);
1308 field_name = isl_multi_pw_aff_get_tuple_name(field, isl_dim_out);
1309 name = member_access_name(ctx, base_name, field_name);
1311 access = isl_multi_pw_aff_range_product(base, field);
1313 access = isl_multi_pw_aff_set_tuple_name(access, isl_dim_out, name);
1314 free(name);
1316 return access;
1319 /* Extract an index expression from a member expression.
1321 * If the base access (to the structure containing the member)
1322 * is of the form
1324 * [] -> A[..]
1326 * and the member is called "f", then the member access is of
1327 * the form
1329 * [] -> A_f[A[..] -> f[]]
1331 * If the member access is to an anonymous struct, then simply return
1333 * [] -> A[..]
1335 * If the member access in the source code is of the form
1337 * A->f
1339 * then it is treated as
1341 * A[0].f
1343 __isl_give isl_multi_pw_aff *PetScan::extract_index(MemberExpr *expr)
1345 Expr *base = expr->getBase();
1346 FieldDecl *field = cast<FieldDecl>(expr->getMemberDecl());
1347 isl_multi_pw_aff *base_access, *field_access;
1348 isl_id *id;
1349 isl_space *space;
1351 base_access = extract_index(base);
1353 if (expr->isArrow()) {
1354 isl_space *space = isl_space_set_alloc(ctx, 0, 0);
1355 isl_local_space *ls = isl_local_space_from_space(space);
1356 isl_aff *aff = isl_aff_zero_on_domain(ls);
1357 isl_pw_aff *index = isl_pw_aff_from_aff(aff);
1358 base_access = subscript(base_access, index);
1361 if (field->isAnonymousStructOrUnion())
1362 return base_access;
1364 id = create_decl_id(ctx, field);
1365 space = isl_multi_pw_aff_get_domain_space(base_access);
1366 space = isl_space_from_domain(space);
1367 space = isl_space_set_tuple_id(space, isl_dim_out, id);
1368 field_access = isl_multi_pw_aff_zero(space);
1370 return member(base_access, field_access);
1373 /* Check if "expr" calls function "minmax" with two arguments and if so
1374 * make lhs and rhs refer to these two arguments.
1376 static bool is_minmax(Expr *expr, const char *minmax, Expr *&lhs, Expr *&rhs)
1378 CallExpr *call;
1379 FunctionDecl *fd;
1380 string name;
1382 if (expr->getStmtClass() != Stmt::CallExprClass)
1383 return false;
1385 call = cast<CallExpr>(expr);
1386 fd = call->getDirectCallee();
1387 if (!fd)
1388 return false;
1390 if (call->getNumArgs() != 2)
1391 return false;
1393 name = fd->getDeclName().getAsString();
1394 if (name != minmax)
1395 return false;
1397 lhs = call->getArg(0);
1398 rhs = call->getArg(1);
1400 return true;
1403 /* Check if "expr" is of the form min(lhs, rhs) and if so make
1404 * lhs and rhs refer to the two arguments.
1406 static bool is_min(Expr *expr, Expr *&lhs, Expr *&rhs)
1408 return is_minmax(expr, "min", lhs, rhs);
1411 /* Check if "expr" is of the form max(lhs, rhs) and if so make
1412 * lhs and rhs refer to the two arguments.
1414 static bool is_max(Expr *expr, Expr *&lhs, Expr *&rhs)
1416 return is_minmax(expr, "max", lhs, rhs);
1419 /* Extract an affine expressions representing the comparison "LHS op RHS"
1420 * "comp" is the original statement that "LHS op RHS" is derived from
1421 * and is used for diagnostics.
1423 * If the comparison is of the form
1425 * a <= min(b,c)
1427 * then the expression is constructed as the conjunction of
1428 * the comparisons
1430 * a <= b and a <= c
1432 * A similar optimization is performed for max(a,b) <= c.
1433 * We do this because that will lead to simpler representations
1434 * of the expression.
1435 * If isl is ever enhanced to explicitly deal with min and max expressions,
1436 * this optimization can be removed.
1438 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperatorKind op,
1439 Expr *LHS, Expr *RHS, Stmt *comp)
1441 isl_pw_aff *lhs;
1442 isl_pw_aff *rhs;
1443 isl_pw_aff *res;
1444 isl_set *cond;
1445 isl_set *dom;
1446 enum pet_op_type type;
1448 if (op == BO_GT)
1449 return extract_comparison(BO_LT, RHS, LHS, comp);
1450 if (op == BO_GE)
1451 return extract_comparison(BO_LE, RHS, LHS, comp);
1453 if (op == BO_LT || op == BO_LE) {
1454 Expr *expr1, *expr2;
1455 if (is_min(RHS, expr1, expr2)) {
1456 lhs = extract_comparison(op, LHS, expr1, comp);
1457 rhs = extract_comparison(op, LHS, expr2, comp);
1458 return pet_and(lhs, rhs);
1460 if (is_max(LHS, expr1, expr2)) {
1461 lhs = extract_comparison(op, expr1, RHS, comp);
1462 rhs = extract_comparison(op, expr2, RHS, comp);
1463 return pet_and(lhs, rhs);
1467 lhs = extract_affine(LHS);
1468 rhs = extract_affine(RHS);
1470 type = BinaryOperatorKind2pet_op_type(op);
1471 return pet_comparison(type, lhs, rhs);
1474 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperator *comp)
1476 return extract_comparison(comp->getOpcode(), comp->getLHS(),
1477 comp->getRHS(), comp);
1480 /* Extract an affine expression from a boolean expression.
1481 * In particular, return the expression "expr ? 1 : 0".
1482 * Return NULL if we are unable to extract an affine expression.
1484 * We first convert the clang::Expr to a pet_expr and
1485 * then extract an affine expression from that pet_expr.
1487 __isl_give isl_pw_aff *PetScan::extract_condition(Expr *expr)
1489 isl_pw_aff *cond;
1490 pet_expr *pe;
1491 pet_context *pc;
1493 if (!expr) {
1494 isl_set *u = isl_set_universe(isl_space_set_alloc(ctx, 0, 0));
1495 return indicator_function(u, isl_set_copy(u));
1498 pe = extract_expr(expr);
1499 pc = convert_assignments(ctx, assigned_value);
1500 pc = pet_context_set_allow_nested(pc, nesting_enabled);
1501 cond = pet_expr_extract_affine_condition(pe, pc);
1502 if (isl_pw_aff_involves_nan(cond))
1503 cond = isl_pw_aff_free(cond);
1504 pet_context_free(pc);
1505 pet_expr_free(pe);
1506 return cond;
1509 /* Construct a pet_expr representing a unary operator expression.
1511 __isl_give pet_expr *PetScan::extract_expr(UnaryOperator *expr)
1513 pet_expr *arg;
1514 enum pet_op_type op;
1516 op = UnaryOperatorKind2pet_op_type(expr->getOpcode());
1517 if (op == pet_op_last) {
1518 unsupported(expr);
1519 return NULL;
1522 arg = extract_expr(expr->getSubExpr());
1524 if (expr->isIncrementDecrementOp() &&
1525 pet_expr_get_type(arg) == pet_expr_access) {
1526 arg = mark_write(arg);
1527 arg = pet_expr_access_set_read(arg, 1);
1530 return pet_expr_new_unary(op, arg);
1533 /* Mark the given access pet_expr as a write.
1534 * If a scalar is being accessed, then mark its value
1535 * as unknown in assigned_value.
1537 __isl_give pet_expr *PetScan::mark_write(__isl_take pet_expr *access)
1539 isl_id *id;
1540 ValueDecl *decl;
1542 access = pet_expr_access_set_write(access, 1);
1543 access = pet_expr_access_set_read(access, 0);
1545 if (!access || !pet_expr_is_scalar_access(access))
1546 return access;
1548 id = pet_expr_access_get_id(access);
1549 decl = (ValueDecl *) isl_id_get_user(id);
1550 clear_assignment(assigned_value, decl);
1551 isl_id_free(id);
1553 return access;
1556 /* Assign "rhs" to "lhs".
1558 * In particular, if "lhs" is a scalar variable, then mark
1559 * the variable as having been assigned. If, furthermore, "rhs"
1560 * is an affine expression, then keep track of this value in assigned_value
1561 * so that we can plug it in when we later come across the same variable.
1563 void PetScan::assign(__isl_keep pet_expr *lhs, Expr *rhs)
1565 isl_id *id;
1566 ValueDecl *decl;
1567 isl_pw_aff *pa;
1569 if (!lhs)
1570 return;
1571 if (!pet_expr_is_scalar_access(lhs))
1572 return;
1574 id = pet_expr_access_get_id(lhs);
1575 decl = (ValueDecl *) isl_id_get_user(id);
1576 isl_id_free(id);
1578 pa = try_extract_affine(rhs);
1579 clear_assignment(assigned_value, decl);
1580 if (!pa)
1581 return;
1582 assigned_value[decl] = pa;
1583 insert_expression(pa);
1586 /* Construct a pet_expr representing a binary operator expression.
1588 * If the top level operator is an assignment and the LHS is an access,
1589 * then we mark that access as a write. If the operator is a compound
1590 * assignment, the access is marked as both a read and a write.
1592 * If "expr" assigns something to a scalar variable, then we mark
1593 * the variable as having been assigned. If, furthermore, the expression
1594 * is affine, then keep track of this value in assigned_value
1595 * so that we can plug it in when we later come across the same variable.
1597 __isl_give pet_expr *PetScan::extract_expr(BinaryOperator *expr)
1599 int type_size;
1600 pet_expr *lhs, *rhs;
1601 enum pet_op_type op;
1603 op = BinaryOperatorKind2pet_op_type(expr->getOpcode());
1604 if (op == pet_op_last) {
1605 unsupported(expr);
1606 return NULL;
1609 lhs = extract_expr(expr->getLHS());
1610 rhs = extract_expr(expr->getRHS());
1612 if (expr->isAssignmentOp() &&
1613 pet_expr_get_type(lhs) == pet_expr_access) {
1614 lhs = mark_write(lhs);
1615 if (expr->isCompoundAssignmentOp())
1616 lhs = pet_expr_access_set_read(lhs, 1);
1619 if (expr->getOpcode() == BO_Assign)
1620 assign(lhs, expr->getRHS());
1622 type_size = get_type_size(expr->getType(), ast_context);
1623 return pet_expr_new_binary(type_size, op, lhs, rhs);
1626 /* Construct a pet_scop with a single statement killing the entire
1627 * array "array".
1629 struct pet_scop *PetScan::kill(Stmt *stmt, struct pet_array *array)
1631 isl_id *id;
1632 isl_space *space;
1633 isl_multi_pw_aff *index;
1634 isl_map *access;
1635 pet_expr *expr;
1637 if (!array)
1638 return NULL;
1639 access = isl_map_from_range(isl_set_copy(array->extent));
1640 id = isl_set_get_tuple_id(array->extent);
1641 space = isl_space_alloc(ctx, 0, 0, 0);
1642 space = isl_space_set_tuple_id(space, isl_dim_out, id);
1643 index = isl_multi_pw_aff_zero(space);
1644 expr = pet_expr_kill_from_access_and_index(access, index);
1645 return extract(expr, stmt->getSourceRange(), false);
1648 /* Construct a pet_scop for a (single) variable declaration.
1650 * The scop contains the variable being declared (as an array)
1651 * and a statement killing the array.
1653 * If the variable is initialized in the AST, then the scop
1654 * also contains an assignment to the variable.
1656 struct pet_scop *PetScan::extract(DeclStmt *stmt)
1658 int type_size;
1659 Decl *decl;
1660 VarDecl *vd;
1661 pet_expr *lhs, *rhs, *pe;
1662 struct pet_scop *scop_decl, *scop;
1663 struct pet_array *array;
1665 if (!stmt->isSingleDecl()) {
1666 unsupported(stmt);
1667 return NULL;
1670 decl = stmt->getSingleDecl();
1671 vd = cast<VarDecl>(decl);
1673 array = extract_array(ctx, vd, NULL);
1674 if (array)
1675 array->declared = 1;
1676 scop_decl = kill(stmt, array);
1677 scop_decl = pet_scop_add_array(scop_decl, array);
1679 if (!vd->getInit())
1680 return scop_decl;
1682 lhs = extract_access_expr(vd);
1683 rhs = extract_expr(vd->getInit());
1685 lhs = mark_write(lhs);
1686 assign(lhs, vd->getInit());
1688 type_size = get_type_size(vd->getType(), ast_context);
1689 pe = pet_expr_new_binary(type_size, pet_op_assign, lhs, rhs);
1690 scop = extract(pe, stmt->getSourceRange(), false);
1692 scop_decl = pet_scop_prefix(scop_decl, 0);
1693 scop = pet_scop_prefix(scop, 1);
1695 scop = pet_scop_add_seq(ctx, scop_decl, scop);
1697 return scop;
1700 /* Construct a pet_expr representing a conditional operation.
1702 * We first try to extract the condition as an affine expression.
1703 * If that fails, we construct a pet_expr tree representing the condition.
1705 __isl_give pet_expr *PetScan::extract_expr(ConditionalOperator *expr)
1707 pet_expr *cond, *lhs, *rhs;
1708 isl_pw_aff *pa;
1710 pa = try_extract_affine(expr->getCond());
1711 if (pa) {
1712 isl_multi_pw_aff *test = isl_multi_pw_aff_from_pw_aff(pa);
1713 cond = pet_expr_from_index(test);
1714 } else
1715 cond = extract_expr(expr->getCond());
1716 lhs = extract_expr(expr->getTrueExpr());
1717 rhs = extract_expr(expr->getFalseExpr());
1719 return pet_expr_new_ternary(cond, lhs, rhs);
1722 __isl_give pet_expr *PetScan::extract_expr(ImplicitCastExpr *expr)
1724 return extract_expr(expr->getSubExpr());
1727 /* Construct a pet_expr representing a floating point value.
1729 * If the floating point literal does not appear in a macro,
1730 * then we use the original representation in the source code
1731 * as the string representation. Otherwise, we use the pretty
1732 * printer to produce a string representation.
1734 __isl_give pet_expr *PetScan::extract_expr(FloatingLiteral *expr)
1736 double d;
1737 string s;
1738 const LangOptions &LO = PP.getLangOpts();
1739 SourceLocation loc = expr->getLocation();
1741 if (!loc.isMacroID()) {
1742 SourceManager &SM = PP.getSourceManager();
1743 unsigned len = Lexer::MeasureTokenLength(loc, SM, LO);
1744 s = string(SM.getCharacterData(loc), len);
1745 } else {
1746 llvm::raw_string_ostream S(s);
1747 expr->printPretty(S, 0, PrintingPolicy(LO));
1748 S.str();
1750 d = expr->getValueAsApproximateDouble();
1751 return pet_expr_new_double(ctx, d, s.c_str());
1754 /* Convert the index expression "index" into an access pet_expr of type "qt".
1756 __isl_give pet_expr *PetScan::extract_access_expr(QualType qt,
1757 __isl_take isl_multi_pw_aff *index)
1759 pet_expr *pe;
1760 int depth;
1761 int type_size;
1763 depth = extract_depth(index);
1764 type_size = get_type_size(qt, ast_context);
1766 pe = pet_expr_from_index_and_depth(type_size, index, depth);
1768 return pe;
1771 /* Extract an index expression from "expr" and then convert it into
1772 * an access pet_expr.
1774 __isl_give pet_expr *PetScan::extract_access_expr(Expr *expr)
1776 return extract_access_expr(expr->getType(), extract_index(expr));
1779 /* Extract an index expression from "decl" and then convert it into
1780 * an access pet_expr.
1782 __isl_give pet_expr *PetScan::extract_access_expr(ValueDecl *decl)
1784 return extract_access_expr(decl->getType(), extract_index(decl));
1787 __isl_give pet_expr *PetScan::extract_expr(ParenExpr *expr)
1789 return extract_expr(expr->getSubExpr());
1792 /* Extract an assume statement from the argument "expr"
1793 * of a __pencil_assume statement.
1795 __isl_give pet_expr *PetScan::extract_assume(Expr *expr)
1797 isl_pw_aff *cond;
1798 pet_expr *res;
1800 cond = try_extract_affine_condition(expr);
1801 if (!cond) {
1802 res = extract_expr(expr);
1803 } else {
1804 isl_multi_pw_aff *index;
1805 index = isl_multi_pw_aff_from_pw_aff(cond);
1806 res = pet_expr_from_index(index);
1808 return pet_expr_new_unary(pet_op_assume, res);
1811 /* Construct a pet_expr corresponding to the function call argument "expr".
1812 * The argument appears in position "pos" of a call to function "fd".
1814 * If we are passing along a pointer to an array element
1815 * or an entire row or even higher dimensional slice of an array,
1816 * then the function being called may write into the array.
1818 * We assume here that if the function is declared to take a pointer
1819 * to a const type, then the function will perform a read
1820 * and that otherwise, it will perform a write.
1822 __isl_give pet_expr *PetScan::extract_argument(FunctionDecl *fd, int pos,
1823 Expr *expr)
1825 pet_expr *res;
1826 int is_addr = 0, is_partial = 0;
1827 Stmt::StmtClass sc;
1829 if (expr->getStmtClass() == Stmt::ImplicitCastExprClass) {
1830 ImplicitCastExpr *ice = cast<ImplicitCastExpr>(expr);
1831 expr = ice->getSubExpr();
1833 if (expr->getStmtClass() == Stmt::UnaryOperatorClass) {
1834 UnaryOperator *op = cast<UnaryOperator>(expr);
1835 if (op->getOpcode() == UO_AddrOf) {
1836 is_addr = 1;
1837 expr = op->getSubExpr();
1840 res = extract_expr(expr);
1841 if (!res)
1842 return NULL;
1843 sc = expr->getStmtClass();
1844 if ((sc == Stmt::ArraySubscriptExprClass ||
1845 sc == Stmt::MemberExprClass) &&
1846 array_depth(expr->getType().getTypePtr()) > 0)
1847 is_partial = 1;
1848 if ((is_addr || is_partial) &&
1849 pet_expr_get_type(res) == pet_expr_access) {
1850 ParmVarDecl *parm;
1851 if (!fd->hasPrototype()) {
1852 report_prototype_required(expr);
1853 return pet_expr_free(res);
1855 parm = fd->getParamDecl(pos);
1856 if (!const_base(parm->getType()))
1857 res = mark_write(res);
1860 if (is_addr)
1861 res = pet_expr_new_unary(pet_op_address_of, res);
1862 return res;
1865 /* Construct a pet_expr representing a function call.
1867 * In the special case of a "call" to __pencil_assume,
1868 * construct an assume expression instead.
1870 __isl_give pet_expr *PetScan::extract_expr(CallExpr *expr)
1872 pet_expr *res = NULL;
1873 FunctionDecl *fd;
1874 string name;
1875 unsigned n_arg;
1877 fd = expr->getDirectCallee();
1878 if (!fd) {
1879 unsupported(expr);
1880 return NULL;
1883 name = fd->getDeclName().getAsString();
1884 n_arg = expr->getNumArgs();
1886 if (n_arg == 1 && name == "__pencil_assume")
1887 return extract_assume(expr->getArg(0));
1889 res = pet_expr_new_call(ctx, name.c_str(), n_arg);
1890 if (!res)
1891 return NULL;
1893 for (int i = 0; i < n_arg; ++i) {
1894 Expr *arg = expr->getArg(i);
1895 res = pet_expr_set_arg(res, i,
1896 PetScan::extract_argument(fd, i, arg));
1899 return res;
1902 /* Construct a pet_expr representing a (C style) cast.
1904 __isl_give pet_expr *PetScan::extract_expr(CStyleCastExpr *expr)
1906 pet_expr *arg;
1907 QualType type;
1909 arg = extract_expr(expr->getSubExpr());
1910 if (!arg)
1911 return NULL;
1913 type = expr->getTypeAsWritten();
1914 return pet_expr_new_cast(type.getAsString().c_str(), arg);
1917 /* Construct a pet_expr representing an integer.
1919 __isl_give pet_expr *PetScan::extract_expr(IntegerLiteral *expr)
1921 return pet_expr_new_int(extract_int(expr));
1924 /* Try and construct a pet_expr representing "expr".
1926 __isl_give pet_expr *PetScan::extract_expr(Expr *expr)
1928 switch (expr->getStmtClass()) {
1929 case Stmt::UnaryOperatorClass:
1930 return extract_expr(cast<UnaryOperator>(expr));
1931 case Stmt::CompoundAssignOperatorClass:
1932 case Stmt::BinaryOperatorClass:
1933 return extract_expr(cast<BinaryOperator>(expr));
1934 case Stmt::ImplicitCastExprClass:
1935 return extract_expr(cast<ImplicitCastExpr>(expr));
1936 case Stmt::ArraySubscriptExprClass:
1937 case Stmt::DeclRefExprClass:
1938 case Stmt::MemberExprClass:
1939 return extract_access_expr(expr);
1940 case Stmt::IntegerLiteralClass:
1941 return extract_expr(cast<IntegerLiteral>(expr));
1942 case Stmt::FloatingLiteralClass:
1943 return extract_expr(cast<FloatingLiteral>(expr));
1944 case Stmt::ParenExprClass:
1945 return extract_expr(cast<ParenExpr>(expr));
1946 case Stmt::ConditionalOperatorClass:
1947 return extract_expr(cast<ConditionalOperator>(expr));
1948 case Stmt::CallExprClass:
1949 return extract_expr(cast<CallExpr>(expr));
1950 case Stmt::CStyleCastExprClass:
1951 return extract_expr(cast<CStyleCastExpr>(expr));
1952 default:
1953 unsupported(expr);
1955 return NULL;
1958 /* Check if the given initialization statement is an assignment.
1959 * If so, return that assignment. Otherwise return NULL.
1961 BinaryOperator *PetScan::initialization_assignment(Stmt *init)
1963 BinaryOperator *ass;
1965 if (init->getStmtClass() != Stmt::BinaryOperatorClass)
1966 return NULL;
1968 ass = cast<BinaryOperator>(init);
1969 if (ass->getOpcode() != BO_Assign)
1970 return NULL;
1972 return ass;
1975 /* Check if the given initialization statement is a declaration
1976 * of a single variable.
1977 * If so, return that declaration. Otherwise return NULL.
1979 Decl *PetScan::initialization_declaration(Stmt *init)
1981 DeclStmt *decl;
1983 if (init->getStmtClass() != Stmt::DeclStmtClass)
1984 return NULL;
1986 decl = cast<DeclStmt>(init);
1988 if (!decl->isSingleDecl())
1989 return NULL;
1991 return decl->getSingleDecl();
1994 /* Given the assignment operator in the initialization of a for loop,
1995 * extract the induction variable, i.e., the (integer)variable being
1996 * assigned.
1998 ValueDecl *PetScan::extract_induction_variable(BinaryOperator *init)
2000 Expr *lhs;
2001 DeclRefExpr *ref;
2002 ValueDecl *decl;
2003 const Type *type;
2005 lhs = init->getLHS();
2006 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
2007 unsupported(init);
2008 return NULL;
2011 ref = cast<DeclRefExpr>(lhs);
2012 decl = ref->getDecl();
2013 type = decl->getType().getTypePtr();
2015 if (!type->isIntegerType()) {
2016 unsupported(lhs);
2017 return NULL;
2020 return decl;
2023 /* Given the initialization statement of a for loop and the single
2024 * declaration in this initialization statement,
2025 * extract the induction variable, i.e., the (integer) variable being
2026 * declared.
2028 VarDecl *PetScan::extract_induction_variable(Stmt *init, Decl *decl)
2030 VarDecl *vd;
2032 vd = cast<VarDecl>(decl);
2034 const QualType type = vd->getType();
2035 if (!type->isIntegerType()) {
2036 unsupported(init);
2037 return NULL;
2040 if (!vd->getInit()) {
2041 unsupported(init);
2042 return NULL;
2045 return vd;
2048 /* Check that op is of the form iv++ or iv--.
2049 * Return a pet_expr representing "1" or "-1" accordingly.
2051 __isl_give pet_expr *PetScan::extract_unary_increment(
2052 clang::UnaryOperator *op, clang::ValueDecl *iv)
2054 Expr *sub;
2055 DeclRefExpr *ref;
2056 isl_val *v;
2058 if (!op->isIncrementDecrementOp()) {
2059 unsupported(op);
2060 return NULL;
2063 sub = op->getSubExpr();
2064 if (sub->getStmtClass() != Stmt::DeclRefExprClass) {
2065 unsupported(op);
2066 return NULL;
2069 ref = cast<DeclRefExpr>(sub);
2070 if (ref->getDecl() != iv) {
2071 unsupported(op);
2072 return NULL;
2075 if (op->isIncrementOp())
2076 v = isl_val_one(ctx);
2077 else
2078 v = isl_val_negone(ctx);
2080 return pet_expr_new_int(v);
2083 /* Check if op is of the form
2085 * iv = expr
2087 * and return the increment "expr - iv" as a pet_expr.
2089 __isl_give pet_expr *PetScan::extract_binary_increment(BinaryOperator *op,
2090 clang::ValueDecl *iv)
2092 int type_size;
2093 Expr *lhs;
2094 DeclRefExpr *ref;
2095 pet_expr *expr, *expr_iv;
2097 if (op->getOpcode() != BO_Assign) {
2098 unsupported(op);
2099 return NULL;
2102 lhs = op->getLHS();
2103 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
2104 unsupported(op);
2105 return NULL;
2108 ref = cast<DeclRefExpr>(lhs);
2109 if (ref->getDecl() != iv) {
2110 unsupported(op);
2111 return NULL;
2114 expr = extract_expr(op->getRHS());
2115 expr_iv = extract_expr(lhs);
2117 type_size = get_type_size(iv->getType(), ast_context);
2118 return pet_expr_new_binary(type_size, pet_op_sub, expr, expr_iv);
2121 /* Check that op is of the form iv += cst or iv -= cst
2122 * and return a pet_expr corresponding to cst or -cst accordingly.
2124 __isl_give pet_expr *PetScan::extract_compound_increment(
2125 CompoundAssignOperator *op, clang::ValueDecl *iv)
2127 Expr *lhs;
2128 DeclRefExpr *ref;
2129 bool neg = false;
2130 pet_expr *expr;
2131 BinaryOperatorKind opcode;
2133 opcode = op->getOpcode();
2134 if (opcode != BO_AddAssign && opcode != BO_SubAssign) {
2135 unsupported(op);
2136 return NULL;
2138 if (opcode == BO_SubAssign)
2139 neg = true;
2141 lhs = op->getLHS();
2142 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
2143 unsupported(op);
2144 return NULL;
2147 ref = cast<DeclRefExpr>(lhs);
2148 if (ref->getDecl() != iv) {
2149 unsupported(op);
2150 return NULL;
2153 expr = extract_expr(op->getRHS());
2154 if (neg)
2155 expr = pet_expr_new_unary(pet_op_minus, expr);
2157 return expr;
2160 /* Check that the increment of the given for loop increments
2161 * (or decrements) the induction variable "iv" and return
2162 * the increment as a pet_expr if successful.
2164 __isl_give pet_expr *PetScan::extract_increment(clang::ForStmt *stmt,
2165 ValueDecl *iv)
2167 Stmt *inc = stmt->getInc();
2169 if (!inc) {
2170 report_missing_increment(stmt);
2171 return NULL;
2174 if (inc->getStmtClass() == Stmt::UnaryOperatorClass)
2175 return extract_unary_increment(cast<UnaryOperator>(inc), iv);
2176 if (inc->getStmtClass() == Stmt::CompoundAssignOperatorClass)
2177 return extract_compound_increment(
2178 cast<CompoundAssignOperator>(inc), iv);
2179 if (inc->getStmtClass() == Stmt::BinaryOperatorClass)
2180 return extract_binary_increment(cast<BinaryOperator>(inc), iv);
2182 unsupported(inc);
2183 return NULL;
2186 /* Embed the given iteration domain in an extra outer loop
2187 * with induction variable "var".
2188 * If this variable appeared as a parameter in the constraints,
2189 * it is replaced by the new outermost dimension.
2191 static __isl_give isl_set *embed(__isl_take isl_set *set,
2192 __isl_take isl_id *var)
2194 int pos;
2196 set = isl_set_insert_dims(set, isl_dim_set, 0, 1);
2197 pos = isl_set_find_dim_by_id(set, isl_dim_param, var);
2198 if (pos >= 0) {
2199 set = isl_set_equate(set, isl_dim_param, pos, isl_dim_set, 0);
2200 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2203 isl_id_free(var);
2204 return set;
2207 /* Return those elements in the space of "cond" that come after
2208 * (based on "sign") an element in "cond".
2210 static __isl_give isl_set *after(__isl_take isl_set *cond, int sign)
2212 isl_map *previous_to_this;
2214 if (sign > 0)
2215 previous_to_this = isl_map_lex_lt(isl_set_get_space(cond));
2216 else
2217 previous_to_this = isl_map_lex_gt(isl_set_get_space(cond));
2219 cond = isl_set_apply(cond, previous_to_this);
2221 return cond;
2224 /* Create the infinite iteration domain
2226 * { [id] : id >= 0 }
2228 * If "scop" has an affine skip of type pet_skip_later,
2229 * then remove those iterations i that have an earlier iteration
2230 * where the skip condition is satisfied, meaning that iteration i
2231 * is not executed.
2232 * Since we are dealing with a loop without loop iterator,
2233 * the skip condition cannot refer to the current loop iterator and
2234 * so effectively, the returned set is of the form
2236 * { [0]; [id] : id >= 1 and not skip }
2238 static __isl_give isl_set *infinite_domain(__isl_take isl_id *id,
2239 struct pet_scop *scop)
2241 isl_ctx *ctx = isl_id_get_ctx(id);
2242 isl_set *domain;
2243 isl_set *skip;
2245 domain = isl_set_nat_universe(isl_space_set_alloc(ctx, 0, 1));
2246 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, id);
2248 if (!pet_scop_has_affine_skip(scop, pet_skip_later))
2249 return domain;
2251 skip = pet_scop_get_affine_skip_domain(scop, pet_skip_later);
2252 skip = embed(skip, isl_id_copy(id));
2253 skip = isl_set_intersect(skip , isl_set_copy(domain));
2254 domain = isl_set_subtract(domain, after(skip, 1));
2256 return domain;
2259 /* Create an identity affine expression on the space containing "domain",
2260 * which is assumed to be one-dimensional.
2262 static __isl_give isl_aff *identity_aff(__isl_keep isl_set *domain)
2264 isl_local_space *ls;
2266 ls = isl_local_space_from_space(isl_set_get_space(domain));
2267 return isl_aff_var_on_domain(ls, isl_dim_set, 0);
2270 /* Create an affine expression that maps elements
2271 * of a single-dimensional array "id_test" to the previous element
2272 * (according to "inc"), provided this element belongs to "domain".
2273 * That is, create the affine expression
2275 * { id[x] -> id[x - inc] : x - inc in domain }
2277 static __isl_give isl_multi_pw_aff *map_to_previous(__isl_take isl_id *id_test,
2278 __isl_take isl_set *domain, __isl_take isl_val *inc)
2280 isl_space *space;
2281 isl_local_space *ls;
2282 isl_aff *aff;
2283 isl_multi_pw_aff *prev;
2285 space = isl_set_get_space(domain);
2286 ls = isl_local_space_from_space(space);
2287 aff = isl_aff_var_on_domain(ls, isl_dim_set, 0);
2288 aff = isl_aff_add_constant_val(aff, isl_val_neg(inc));
2289 prev = isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
2290 domain = isl_set_preimage_multi_pw_aff(domain,
2291 isl_multi_pw_aff_copy(prev));
2292 prev = isl_multi_pw_aff_intersect_domain(prev, domain);
2293 prev = isl_multi_pw_aff_set_tuple_id(prev, isl_dim_out, id_test);
2295 return prev;
2298 /* Add an implication to "scop" expressing that if an element of
2299 * virtual array "id_test" has value "satisfied" then all previous elements
2300 * of this array also have that value. The set of previous elements
2301 * is bounded by "domain". If "sign" is negative then the iterator
2302 * is decreasing and we express that all subsequent array elements
2303 * (but still defined previously) have the same value.
2305 static struct pet_scop *add_implication(struct pet_scop *scop,
2306 __isl_take isl_id *id_test, __isl_take isl_set *domain, int sign,
2307 int satisfied)
2309 isl_space *space;
2310 isl_map *map;
2312 domain = isl_set_set_tuple_id(domain, id_test);
2313 space = isl_set_get_space(domain);
2314 if (sign > 0)
2315 map = isl_map_lex_ge(space);
2316 else
2317 map = isl_map_lex_le(space);
2318 map = isl_map_intersect_range(map, domain);
2319 scop = pet_scop_add_implication(scop, map, satisfied);
2321 return scop;
2324 /* Add a filter to "scop" that imposes that it is only executed
2325 * when the variable identified by "id_test" has a zero value
2326 * for all previous iterations of "domain".
2328 * In particular, add a filter that imposes that the array
2329 * has a zero value at the previous iteration of domain and
2330 * add an implication that implies that it then has that
2331 * value for all previous iterations.
2333 static struct pet_scop *scop_add_break(struct pet_scop *scop,
2334 __isl_take isl_id *id_test, __isl_take isl_set *domain,
2335 __isl_take isl_val *inc)
2337 isl_multi_pw_aff *prev;
2338 int sign = isl_val_sgn(inc);
2340 prev = map_to_previous(isl_id_copy(id_test), isl_set_copy(domain), inc);
2341 scop = add_implication(scop, id_test, domain, sign, 0);
2342 scop = pet_scop_filter(scop, prev, 0);
2344 return scop;
2347 /* Construct a pet_scop for an infinite loop around the given body.
2349 * We extract a pet_scop for the body and then embed it in a loop with
2350 * iteration domain
2352 * { [t] : t >= 0 }
2354 * and schedule
2356 * { [t] -> [t] }
2358 * If the body contains any break, then it is taken into
2359 * account in infinite_domain (if the skip condition is affine)
2360 * or in scop_add_break (if the skip condition is not affine).
2362 * If we were only able to extract part of the body, then simply
2363 * return that part.
2365 struct pet_scop *PetScan::extract_infinite_loop(Stmt *body)
2367 isl_id *id, *id_test;
2368 isl_set *domain;
2369 isl_aff *ident;
2370 struct pet_scop *scop;
2371 bool has_var_break;
2373 scop = extract(body);
2374 if (!scop)
2375 return NULL;
2376 if (partial)
2377 return scop;
2379 id = isl_id_alloc(ctx, "t", NULL);
2380 domain = infinite_domain(isl_id_copy(id), scop);
2381 ident = identity_aff(domain);
2383 has_var_break = pet_scop_has_var_skip(scop, pet_skip_later);
2384 if (has_var_break)
2385 id_test = pet_scop_get_skip_id(scop, pet_skip_later);
2387 scop = pet_scop_embed(scop, isl_set_copy(domain),
2388 isl_aff_copy(ident), ident, id);
2389 if (has_var_break)
2390 scop = scop_add_break(scop, id_test, domain, isl_val_one(ctx));
2391 else
2392 isl_set_free(domain);
2394 return scop;
2397 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
2399 * for (;;)
2400 * body
2403 struct pet_scop *PetScan::extract_infinite_for(ForStmt *stmt)
2405 clear_assignments clear(assigned_value);
2406 clear.TraverseStmt(stmt->getBody());
2408 return extract_infinite_loop(stmt->getBody());
2411 /* Add an array with the given extent (range of "index") to the list
2412 * of arrays in "scop" and return the extended pet_scop.
2413 * The array is marked as attaining values 0 and 1 only and
2414 * as each element being assigned at most once.
2416 static struct pet_scop *scop_add_array(struct pet_scop *scop,
2417 __isl_keep isl_multi_pw_aff *index, clang::ASTContext &ast_ctx)
2419 int int_size = ast_ctx.getTypeInfo(ast_ctx.IntTy).first / 8;
2421 return pet_scop_add_boolean_array(scop, isl_multi_pw_aff_copy(index),
2422 int_size);
2425 /* Construct a pet_scop for a while loop of the form
2427 * while (pa)
2428 * body
2430 * In particular, construct a scop for an infinite loop around body and
2431 * intersect the domain with the affine expression.
2432 * Note that this intersection may result in an empty loop.
2434 struct pet_scop *PetScan::extract_affine_while(__isl_take isl_pw_aff *pa,
2435 Stmt *body)
2437 struct pet_scop *scop;
2438 isl_set *dom;
2439 isl_set *valid;
2441 valid = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2442 dom = isl_pw_aff_non_zero_set(pa);
2443 scop = extract_infinite_loop(body);
2444 scop = pet_scop_restrict(scop, isl_set_params(dom));
2445 scop = pet_scop_restrict_context(scop, isl_set_params(valid));
2447 return scop;
2450 /* Construct a scop for a while, given the scops for the condition
2451 * and the body, the filter identifier and the iteration domain of
2452 * the while loop.
2454 * In particular, the scop for the condition is filtered to depend
2455 * on "id_test" evaluating to true for all previous iterations
2456 * of the loop, while the scop for the body is filtered to depend
2457 * on "id_test" evaluating to true for all iterations up to the
2458 * current iteration.
2459 * The actual filter only imposes that this virtual array has
2460 * value one on the previous or the current iteration.
2461 * The fact that this condition also applies to the previous
2462 * iterations is enforced by an implication.
2464 * These filtered scops are then combined into a single scop.
2466 * "sign" is positive if the iterator increases and negative
2467 * if it decreases.
2469 static struct pet_scop *scop_add_while(struct pet_scop *scop_cond,
2470 struct pet_scop *scop_body, __isl_take isl_id *id_test,
2471 __isl_take isl_set *domain, __isl_take isl_val *inc)
2473 isl_ctx *ctx = isl_set_get_ctx(domain);
2474 isl_space *space;
2475 isl_multi_pw_aff *test_index;
2476 isl_multi_pw_aff *prev;
2477 int sign = isl_val_sgn(inc);
2478 struct pet_scop *scop;
2480 prev = map_to_previous(isl_id_copy(id_test), isl_set_copy(domain), inc);
2481 scop_cond = pet_scop_filter(scop_cond, prev, 1);
2483 space = isl_space_map_from_set(isl_set_get_space(domain));
2484 test_index = isl_multi_pw_aff_identity(space);
2485 test_index = isl_multi_pw_aff_set_tuple_id(test_index, isl_dim_out,
2486 isl_id_copy(id_test));
2487 scop_body = pet_scop_filter(scop_body, test_index, 1);
2489 scop = pet_scop_add_seq(ctx, scop_cond, scop_body);
2490 scop = add_implication(scop, id_test, domain, sign, 1);
2492 return scop;
2495 /* Check if the while loop is of the form
2497 * while (affine expression)
2498 * body
2500 * If so, call extract_affine_while to construct a scop.
2502 * Otherwise, extract the body and pass control to extract_while
2503 * to extend the iteration domain with an infinite loop.
2504 * If we were only able to extract part of the body, then simply
2505 * return that part.
2507 struct pet_scop *PetScan::extract(WhileStmt *stmt)
2509 Expr *cond;
2510 int test_nr, stmt_nr;
2511 isl_pw_aff *pa;
2512 struct pet_scop *scop_body;
2514 cond = stmt->getCond();
2515 if (!cond) {
2516 unsupported(stmt);
2517 return NULL;
2520 clear_assignments clear(assigned_value);
2521 clear.TraverseStmt(stmt->getBody());
2523 pa = try_extract_affine_condition(cond);
2524 if (pa)
2525 return extract_affine_while(pa, stmt->getBody());
2527 if (!allow_nested) {
2528 unsupported(stmt);
2529 return NULL;
2532 test_nr = n_test++;
2533 stmt_nr = n_stmt++;
2534 scop_body = extract(stmt->getBody());
2535 if (partial)
2536 return scop_body;
2538 return extract_while(cond, test_nr, stmt_nr, scop_body, NULL);
2541 /* Construct a generic while scop, with iteration domain
2542 * { [t] : t >= 0 } around "scop_body". The scop consists of two parts,
2543 * one for evaluating the condition "cond" and one for the body.
2544 * "test_nr" is the sequence number of the virtual test variable that contains
2545 * the result of the condition and "stmt_nr" is the sequence number
2546 * of the statement that evaluates the condition.
2547 * If "scop_inc" is not NULL, then it is added at the end of the body,
2548 * after replacing any skip conditions resulting from continue statements
2549 * by the skip conditions resulting from break statements (if any).
2551 * The schedule is adjusted to reflect that the condition is evaluated
2552 * before the body is executed and the body is filtered to depend
2553 * on the result of the condition evaluating to true on all iterations
2554 * up to the current iteration, while the evaluation of the condition itself
2555 * is filtered to depend on the result of the condition evaluating to true
2556 * on all previous iterations.
2557 * The context of the scop representing the body is dropped
2558 * because we don't know how many times the body will be executed,
2559 * if at all.
2561 * If the body contains any break, then it is taken into
2562 * account in infinite_domain (if the skip condition is affine)
2563 * or in scop_add_break (if the skip condition is not affine).
2565 struct pet_scop *PetScan::extract_while(Expr *cond, int test_nr, int stmt_nr,
2566 struct pet_scop *scop_body, struct pet_scop *scop_inc)
2568 isl_id *id, *id_test, *id_break_test;
2569 isl_set *domain;
2570 isl_aff *ident;
2571 isl_multi_pw_aff *test_index;
2572 struct pet_scop *scop;
2573 bool has_var_break;
2575 test_index = pet_create_test_index(ctx, test_nr);
2576 scop = extract_non_affine_condition(cond, stmt_nr,
2577 isl_multi_pw_aff_copy(test_index));
2578 scop = scop_add_array(scop, test_index, ast_context);
2579 id_test = isl_multi_pw_aff_get_tuple_id(test_index, isl_dim_out);
2580 isl_multi_pw_aff_free(test_index);
2582 id = isl_id_alloc(ctx, "t", NULL);
2583 domain = infinite_domain(isl_id_copy(id), scop_body);
2584 ident = identity_aff(domain);
2586 has_var_break = pet_scop_has_var_skip(scop_body, pet_skip_later);
2587 if (has_var_break)
2588 id_break_test = pet_scop_get_skip_id(scop_body, pet_skip_later);
2590 scop = pet_scop_prefix(scop, 0);
2591 scop = pet_scop_embed(scop, isl_set_copy(domain), isl_aff_copy(ident),
2592 isl_aff_copy(ident), isl_id_copy(id));
2593 scop_body = pet_scop_reset_context(scop_body);
2594 scop_body = pet_scop_prefix(scop_body, 1);
2595 if (scop_inc) {
2596 scop_inc = pet_scop_prefix(scop_inc, 2);
2597 if (pet_scop_has_skip(scop_body, pet_skip_later)) {
2598 isl_multi_pw_aff *skip;
2599 skip = pet_scop_get_skip(scop_body, pet_skip_later);
2600 scop_body = pet_scop_set_skip(scop_body,
2601 pet_skip_now, skip);
2602 } else
2603 pet_scop_reset_skip(scop_body, pet_skip_now);
2604 scop_body = pet_scop_add_seq(ctx, scop_body, scop_inc);
2606 scop_body = pet_scop_embed(scop_body, isl_set_copy(domain),
2607 isl_aff_copy(ident), ident, id);
2609 if (has_var_break) {
2610 scop = scop_add_break(scop, isl_id_copy(id_break_test),
2611 isl_set_copy(domain), isl_val_one(ctx));
2612 scop_body = scop_add_break(scop_body, id_break_test,
2613 isl_set_copy(domain), isl_val_one(ctx));
2615 scop = scop_add_while(scop, scop_body, id_test, domain,
2616 isl_val_one(ctx));
2618 return scop;
2621 /* Check whether "cond" expresses a simple loop bound
2622 * on the only set dimension.
2623 * In particular, if "up" is set then "cond" should contain only
2624 * upper bounds on the set dimension.
2625 * Otherwise, it should contain only lower bounds.
2627 static bool is_simple_bound(__isl_keep isl_set *cond, __isl_keep isl_val *inc)
2629 if (isl_val_is_pos(inc))
2630 return !isl_set_dim_has_any_lower_bound(cond, isl_dim_set, 0);
2631 else
2632 return !isl_set_dim_has_any_upper_bound(cond, isl_dim_set, 0);
2635 /* Extend a condition on a given iteration of a loop to one that
2636 * imposes the same condition on all previous iterations.
2637 * "domain" expresses the lower [upper] bound on the iterations
2638 * when inc is positive [negative].
2640 * In particular, we construct the condition (when inc is positive)
2642 * forall i' : (domain(i') and i' <= i) => cond(i')
2644 * which is equivalent to
2646 * not exists i' : domain(i') and i' <= i and not cond(i')
2648 * We construct this set by negating cond, applying a map
2650 * { [i'] -> [i] : domain(i') and i' <= i }
2652 * and then negating the result again.
2654 static __isl_give isl_set *valid_for_each_iteration(__isl_take isl_set *cond,
2655 __isl_take isl_set *domain, __isl_take isl_val *inc)
2657 isl_map *previous_to_this;
2659 if (isl_val_is_pos(inc))
2660 previous_to_this = isl_map_lex_le(isl_set_get_space(domain));
2661 else
2662 previous_to_this = isl_map_lex_ge(isl_set_get_space(domain));
2664 previous_to_this = isl_map_intersect_domain(previous_to_this, domain);
2666 cond = isl_set_complement(cond);
2667 cond = isl_set_apply(cond, previous_to_this);
2668 cond = isl_set_complement(cond);
2670 isl_val_free(inc);
2672 return cond;
2675 /* Construct a domain of the form
2677 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
2679 static __isl_give isl_set *strided_domain(__isl_take isl_id *id,
2680 __isl_take isl_pw_aff *init, __isl_take isl_val *inc)
2682 isl_aff *aff;
2683 isl_space *dim;
2684 isl_set *set;
2686 init = isl_pw_aff_insert_dims(init, isl_dim_in, 0, 1);
2687 dim = isl_pw_aff_get_domain_space(init);
2688 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2689 aff = isl_aff_add_coefficient_val(aff, isl_dim_in, 0, inc);
2690 init = isl_pw_aff_add(init, isl_pw_aff_from_aff(aff));
2692 dim = isl_space_set_alloc(isl_pw_aff_get_ctx(init), 1, 1);
2693 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
2694 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2695 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2697 set = isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff), init);
2699 set = isl_set_lower_bound_si(set, isl_dim_set, 0, 0);
2701 return isl_set_params(set);
2704 /* Assuming "cond" represents a bound on a loop where the loop
2705 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2706 * is possible.
2708 * Under the given assumptions, wrapping is only possible if "cond" allows
2709 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2710 * increasing iterator and 0 in case of a decreasing iterator.
2712 static bool can_wrap(__isl_keep isl_set *cond, ValueDecl *iv,
2713 __isl_keep isl_val *inc)
2715 bool cw;
2716 isl_ctx *ctx;
2717 isl_val *limit;
2718 isl_set *test;
2720 test = isl_set_copy(cond);
2722 ctx = isl_set_get_ctx(test);
2723 if (isl_val_is_neg(inc))
2724 limit = isl_val_zero(ctx);
2725 else {
2726 limit = isl_val_int_from_ui(ctx, get_type_size(iv));
2727 limit = isl_val_2exp(limit);
2728 limit = isl_val_sub_ui(limit, 1);
2731 test = isl_set_fix_val(cond, isl_dim_set, 0, limit);
2732 cw = !isl_set_is_empty(test);
2733 isl_set_free(test);
2735 return cw;
2738 /* Given a one-dimensional space, construct the following affine expression
2739 * on this space
2741 * { [v] -> [v mod 2^width] }
2743 * where width is the number of bits used to represent the values
2744 * of the unsigned variable "iv".
2746 static __isl_give isl_aff *compute_wrapping(__isl_take isl_space *dim,
2747 ValueDecl *iv)
2749 isl_ctx *ctx;
2750 isl_val *mod;
2751 isl_aff *aff;
2753 ctx = isl_space_get_ctx(dim);
2754 mod = isl_val_int_from_ui(ctx, get_type_size(iv));
2755 mod = isl_val_2exp(mod);
2757 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2758 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2759 aff = isl_aff_mod_val(aff, mod);
2761 return aff;
2764 /* Project out the parameter "id" from "set".
2766 static __isl_give isl_set *set_project_out_by_id(__isl_take isl_set *set,
2767 __isl_keep isl_id *id)
2769 int pos;
2771 pos = isl_set_find_dim_by_id(set, isl_dim_param, id);
2772 if (pos >= 0)
2773 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2775 return set;
2778 /* Compute the set of parameters for which "set1" is a subset of "set2".
2780 * set1 is a subset of set2 if
2782 * forall i in set1 : i in set2
2784 * or
2786 * not exists i in set1 and i not in set2
2788 * i.e.,
2790 * not exists i in set1 \ set2
2792 static __isl_give isl_set *enforce_subset(__isl_take isl_set *set1,
2793 __isl_take isl_set *set2)
2795 return isl_set_complement(isl_set_params(isl_set_subtract(set1, set2)));
2798 /* Compute the set of parameter values for which "cond" holds
2799 * on the next iteration for each element of "dom".
2801 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2802 * and then compute the set of parameters for which the result is a subset
2803 * of "cond".
2805 static __isl_give isl_set *valid_on_next(__isl_take isl_set *cond,
2806 __isl_take isl_set *dom, __isl_take isl_val *inc)
2808 isl_space *space;
2809 isl_aff *aff;
2810 isl_map *next;
2812 space = isl_set_get_space(dom);
2813 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2814 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2815 aff = isl_aff_add_constant_val(aff, inc);
2816 next = isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2818 dom = isl_set_apply(dom, next);
2820 return enforce_subset(dom, cond);
2823 /* Extract the for loop "stmt" as a while loop.
2824 * "iv" is the loop iterator. "init" is the initialization.
2825 * "inc" is the increment.
2827 * That is, the for loop has the form
2829 * for (iv = init; cond; iv += inc)
2830 * body;
2832 * and is treated as
2834 * iv = init;
2835 * while (cond) {
2836 * body;
2837 * iv += inc;
2840 * except that the skips resulting from any continue statements
2841 * in body do not apply to the increment, but are replaced by the skips
2842 * resulting from break statements.
2844 * If "iv" is declared in the for loop, then it is killed before
2845 * and after the loop.
2847 struct pet_scop *PetScan::extract_non_affine_for(ForStmt *stmt, ValueDecl *iv,
2848 __isl_take pet_expr *init, __isl_take pet_expr *inc)
2850 int declared;
2851 int test_nr, stmt_nr;
2852 pet_expr *expr_iv;
2853 struct pet_scop *scop_init, *scop_inc, *scop, *scop_body;
2854 int type_size;
2855 struct pet_array *array;
2856 struct pet_scop *scop_kill;
2858 if (!allow_nested) {
2859 unsupported(stmt);
2860 return NULL;
2863 clear_assignment(assigned_value, iv);
2865 declared = !initialization_assignment(stmt->getInit());
2867 expr_iv = extract_access_expr(iv);
2868 expr_iv = mark_write(expr_iv);
2869 type_size = pet_expr_get_type_size(expr_iv);
2870 init = pet_expr_new_binary(type_size, pet_op_assign, expr_iv, init);
2871 scop_init = extract(init, stmt->getInit()->getSourceRange(), false);
2872 scop_init = pet_scop_prefix(scop_init, declared);
2874 test_nr = n_test++;
2875 stmt_nr = n_stmt++;
2876 scop_body = extract(stmt->getBody());
2877 if (partial) {
2878 pet_scop_free(scop_init);
2879 return scop_body;
2882 expr_iv = extract_access_expr(iv);
2883 expr_iv = mark_write(expr_iv);
2884 type_size = pet_expr_get_type_size(expr_iv);
2885 inc = pet_expr_new_binary(type_size, pet_op_add_assign, expr_iv, inc);
2886 scop_inc = extract(inc, stmt->getInc()->getSourceRange(), false);
2887 if (!scop_inc) {
2888 pet_scop_free(scop_init);
2889 pet_scop_free(scop_body);
2890 return NULL;
2893 scop = extract_while(stmt->getCond(), test_nr, stmt_nr, scop_body,
2894 scop_inc);
2896 scop = pet_scop_prefix(scop, declared + 1);
2897 scop = pet_scop_add_seq(ctx, scop_init, scop);
2899 if (!declared)
2900 return scop;
2902 array = extract_array(ctx, iv, NULL);
2903 if (array)
2904 array->declared = 1;
2905 scop_kill = kill(stmt, array);
2906 scop_kill = pet_scop_prefix(scop_kill, 0);
2907 scop = pet_scop_add_seq(ctx, scop_kill, scop);
2908 scop_kill = kill(stmt, array);
2909 scop_kill = pet_scop_add_array(scop_kill, array);
2910 scop_kill = pet_scop_prefix(scop_kill, 3);
2911 scop = pet_scop_add_seq(ctx, scop, scop_kill);
2913 return scop;
2916 /* Construct a pet_scop for a for statement.
2917 * The for loop is required to be of one of the following forms
2919 * for (i = init; condition; ++i)
2920 * for (i = init; condition; --i)
2921 * for (i = init; condition; i += constant)
2922 * for (i = init; condition; i -= constant)
2924 * The initialization of the for loop should either be an assignment
2925 * of a static affine value to an integer variable, or a declaration
2926 * of such a variable with initialization.
2928 * If the initialization or the increment do not satisfy the above
2929 * conditions, i.e., if the initialization is not static affine
2930 * or the increment is not constant, then the for loop is extracted
2931 * as a while loop instead.
2933 * The condition is allowed to contain nested accesses, provided
2934 * they are not being written to inside the body of the loop.
2935 * Otherwise, or if the condition is otherwise non-affine, the for loop is
2936 * essentially treated as a while loop, with iteration domain
2937 * { [i] : i >= init }.
2939 * We extract a pet_scop for the body and then embed it in a loop with
2940 * iteration domain and schedule
2942 * { [i] : i >= init and condition' }
2943 * { [i] -> [i] }
2945 * or
2947 * { [i] : i <= init and condition' }
2948 * { [i] -> [-i] }
2950 * Where condition' is equal to condition if the latter is
2951 * a simple upper [lower] bound and a condition that is extended
2952 * to apply to all previous iterations otherwise.
2954 * If the condition is non-affine, then we drop the condition from the
2955 * iteration domain and instead create a separate statement
2956 * for evaluating the condition. The body is then filtered to depend
2957 * on the result of the condition evaluating to true on all iterations
2958 * up to the current iteration, while the evaluation the condition itself
2959 * is filtered to depend on the result of the condition evaluating to true
2960 * on all previous iterations.
2961 * The context of the scop representing the body is dropped
2962 * because we don't know how many times the body will be executed,
2963 * if at all.
2965 * If the stride of the loop is not 1, then "i >= init" is replaced by
2967 * (exists a: i = init + stride * a and a >= 0)
2969 * If the loop iterator i is unsigned, then wrapping may occur.
2970 * We therefore use a virtual iterator instead that does not wrap.
2971 * However, the condition in the code applies
2972 * to the wrapped value, so we need to change condition(i)
2973 * into condition([i % 2^width]). Similarly, we replace all accesses
2974 * to the original iterator by the wrapping of the virtual iterator.
2975 * Note that there may be no need to perform this final wrapping
2976 * if the loop condition (after wrapping) satisfies certain conditions.
2977 * However, the is_simple_bound condition is not enough since it doesn't
2978 * check if there even is an upper bound.
2980 * Wrapping on unsigned iterators can be avoided entirely if
2981 * loop condition is simple, the loop iterator is incremented
2982 * [decremented] by one and the last value before wrapping cannot
2983 * possibly satisfy the loop condition.
2985 * Before extracting a pet_scop from the body we remove all
2986 * assignments in assigned_value to variables that are assigned
2987 * somewhere in the body of the loop.
2989 * Valid parameters for a for loop are those for which the initial
2990 * value itself, the increment on each domain iteration and
2991 * the condition on both the initial value and
2992 * the result of incrementing the iterator for each iteration of the domain
2993 * can be evaluated.
2994 * If the loop condition is non-affine, then we only consider validity
2995 * of the initial value.
2997 * If the body contains any break, then we keep track of it in "skip"
2998 * (if the skip condition is affine) or it is handled in scop_add_break
2999 * (if the skip condition is not affine).
3000 * Note that the affine break condition needs to be considered with
3001 * respect to previous iterations in the virtual domain (if any).
3003 * If we were only able to extract part of the body, then simply
3004 * return that part.
3006 struct pet_scop *PetScan::extract_for(ForStmt *stmt)
3008 BinaryOperator *ass;
3009 Decl *decl;
3010 Stmt *init;
3011 Expr *lhs, *rhs;
3012 ValueDecl *iv;
3013 isl_local_space *ls;
3014 isl_set *domain;
3015 isl_aff *sched;
3016 isl_set *cond = NULL;
3017 isl_set *skip = NULL;
3018 isl_id *id, *id_test = NULL, *id_break_test;
3019 struct pet_scop *scop, *scop_cond = NULL;
3020 assigned_value_cache cache(assigned_value);
3021 isl_val *inc;
3022 bool is_one;
3023 bool is_unsigned;
3024 bool is_simple;
3025 bool is_virtual;
3026 bool has_affine_break;
3027 bool has_var_break;
3028 isl_aff *wrap = NULL;
3029 isl_pw_aff *pa, *pa_inc, *init_val;
3030 isl_set *valid_init;
3031 isl_set *valid_cond;
3032 isl_set *valid_cond_init;
3033 isl_set *valid_cond_next;
3034 isl_set *valid_inc;
3035 int stmt_id;
3036 pet_expr *pe_init, *pe_inc;
3037 pet_context *pc, *pc_init_val;
3039 if (!stmt->getInit() && !stmt->getCond() && !stmt->getInc())
3040 return extract_infinite_for(stmt);
3042 init = stmt->getInit();
3043 if (!init) {
3044 unsupported(stmt);
3045 return NULL;
3047 if ((ass = initialization_assignment(init)) != NULL) {
3048 iv = extract_induction_variable(ass);
3049 if (!iv)
3050 return NULL;
3051 lhs = ass->getLHS();
3052 rhs = ass->getRHS();
3053 } else if ((decl = initialization_declaration(init)) != NULL) {
3054 VarDecl *var = extract_induction_variable(init, decl);
3055 if (!var)
3056 return NULL;
3057 iv = var;
3058 rhs = var->getInit();
3059 lhs = create_DeclRefExpr(var);
3060 } else {
3061 unsupported(stmt->getInit());
3062 return NULL;
3065 id = create_decl_id(ctx, iv);
3067 assigned_value.erase(iv);
3068 clear_assignments clear(assigned_value);
3069 clear.TraverseStmt(stmt->getBody());
3071 pe_init = extract_expr(rhs);
3072 pe_inc = extract_increment(stmt, iv);
3073 pc = convert_assignments(ctx, assigned_value);
3074 pc_init_val = pet_context_copy(pc);
3075 pc_init_val = pet_context_mark_unknown(pc_init_val, isl_id_copy(id));
3076 init_val = pet_expr_extract_affine(pe_init, pc_init_val);
3077 pet_context_free(pc_init_val);
3078 pa_inc = pet_expr_extract_affine(pe_inc, pc);
3079 pet_context_free(pc);
3080 inc = pet_extract_cst(pa_inc);
3081 if (!pe_init || !pe_inc || !inc || isl_val_is_nan(inc) ||
3082 isl_pw_aff_involves_nan(pa_inc) ||
3083 isl_pw_aff_involves_nan(init_val)) {
3084 isl_id_free(id);
3085 isl_val_free(inc);
3086 isl_pw_aff_free(pa_inc);
3087 isl_pw_aff_free(init_val);
3088 if (pe_init && pe_inc && !(pa_inc && !inc))
3089 return extract_non_affine_for(stmt, iv,
3090 pe_init, pe_inc);
3091 pet_expr_free(pe_init);
3092 pet_expr_free(pe_inc);
3093 return NULL;
3095 pet_expr_free(pe_init);
3096 pet_expr_free(pe_inc);
3098 pa = try_extract_nested_condition(stmt->getCond());
3099 if (allow_nested && (!pa || pet_nested_any_in_pw_aff(pa)))
3100 stmt_id = n_stmt++;
3102 scop = extract(stmt->getBody());
3103 if (partial) {
3104 isl_id_free(id);
3105 isl_pw_aff_free(init_val);
3106 isl_pw_aff_free(pa_inc);
3107 isl_pw_aff_free(pa);
3108 isl_val_free(inc);
3109 return scop;
3112 valid_inc = isl_pw_aff_domain(pa_inc);
3114 is_unsigned = iv->getType()->isUnsignedIntegerType();
3116 has_affine_break = scop &&
3117 pet_scop_has_affine_skip(scop, pet_skip_later);
3118 if (has_affine_break)
3119 skip = pet_scop_get_affine_skip_domain(scop, pet_skip_later);
3120 has_var_break = scop && pet_scop_has_var_skip(scop, pet_skip_later);
3121 if (has_var_break)
3122 id_break_test = pet_scop_get_skip_id(scop, pet_skip_later);
3124 if (pa && !is_nested_allowed(pa, scop)) {
3125 isl_pw_aff_free(pa);
3126 pa = NULL;
3129 if (!allow_nested && !pa)
3130 pa = try_extract_affine_condition(stmt->getCond());
3131 valid_cond = isl_pw_aff_domain(isl_pw_aff_copy(pa));
3132 cond = isl_pw_aff_non_zero_set(pa);
3133 if (allow_nested && !cond) {
3134 isl_multi_pw_aff *test_index;
3135 int save_n_stmt = n_stmt;
3136 test_index = pet_create_test_index(ctx, n_test++);
3137 n_stmt = stmt_id;
3138 scop_cond = extract_non_affine_condition(stmt->getCond(),
3139 n_stmt++, isl_multi_pw_aff_copy(test_index));
3140 n_stmt = save_n_stmt;
3141 scop_cond = scop_add_array(scop_cond, test_index, ast_context);
3142 id_test = isl_multi_pw_aff_get_tuple_id(test_index,
3143 isl_dim_out);
3144 isl_multi_pw_aff_free(test_index);
3145 scop_cond = pet_scop_prefix(scop_cond, 0);
3146 scop = pet_scop_reset_context(scop);
3147 scop = pet_scop_prefix(scop, 1);
3148 cond = isl_set_universe(isl_space_set_alloc(ctx, 0, 0));
3151 cond = embed(cond, isl_id_copy(id));
3152 skip = embed(skip, isl_id_copy(id));
3153 valid_cond = isl_set_coalesce(valid_cond);
3154 valid_cond = embed(valid_cond, isl_id_copy(id));
3155 valid_inc = embed(valid_inc, isl_id_copy(id));
3156 is_one = isl_val_is_one(inc) || isl_val_is_negone(inc);
3157 is_virtual = is_unsigned && (!is_one || can_wrap(cond, iv, inc));
3159 valid_cond_init = enforce_subset(
3160 isl_map_range(isl_map_from_pw_aff(isl_pw_aff_copy(init_val))),
3161 isl_set_copy(valid_cond));
3162 if (is_one && !is_virtual) {
3163 isl_pw_aff_free(init_val);
3164 pa = extract_comparison(isl_val_is_pos(inc) ? BO_GE : BO_LE,
3165 lhs, rhs, init);
3166 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(pa));
3167 valid_init = set_project_out_by_id(valid_init, id);
3168 domain = isl_pw_aff_non_zero_set(pa);
3169 } else {
3170 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(init_val));
3171 domain = strided_domain(isl_id_copy(id), init_val,
3172 isl_val_copy(inc));
3175 domain = embed(domain, isl_id_copy(id));
3176 if (is_virtual) {
3177 isl_map *rev_wrap;
3178 wrap = compute_wrapping(isl_set_get_space(cond), iv);
3179 rev_wrap = isl_map_from_aff(isl_aff_copy(wrap));
3180 rev_wrap = isl_map_reverse(rev_wrap);
3181 cond = isl_set_apply(cond, isl_map_copy(rev_wrap));
3182 skip = isl_set_apply(skip, isl_map_copy(rev_wrap));
3183 valid_cond = isl_set_apply(valid_cond, isl_map_copy(rev_wrap));
3184 valid_inc = isl_set_apply(valid_inc, rev_wrap);
3186 is_simple = is_simple_bound(cond, inc);
3187 if (!is_simple) {
3188 cond = isl_set_gist(cond, isl_set_copy(domain));
3189 is_simple = is_simple_bound(cond, inc);
3191 if (!is_simple)
3192 cond = valid_for_each_iteration(cond,
3193 isl_set_copy(domain), isl_val_copy(inc));
3194 domain = isl_set_intersect(domain, cond);
3195 if (has_affine_break) {
3196 skip = isl_set_intersect(skip , isl_set_copy(domain));
3197 skip = after(skip, isl_val_sgn(inc));
3198 domain = isl_set_subtract(domain, skip);
3200 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, isl_id_copy(id));
3201 ls = isl_local_space_from_space(isl_set_get_space(domain));
3202 sched = isl_aff_var_on_domain(ls, isl_dim_set, 0);
3203 if (isl_val_is_neg(inc))
3204 sched = isl_aff_neg(sched);
3206 valid_cond_next = valid_on_next(valid_cond, isl_set_copy(domain),
3207 isl_val_copy(inc));
3208 valid_inc = enforce_subset(isl_set_copy(domain), valid_inc);
3210 if (!is_virtual)
3211 wrap = identity_aff(domain);
3213 scop_cond = pet_scop_embed(scop_cond, isl_set_copy(domain),
3214 isl_aff_copy(sched), isl_aff_copy(wrap), isl_id_copy(id));
3215 scop = pet_scop_embed(scop, isl_set_copy(domain), sched, wrap, id);
3216 scop = resolve_nested(scop);
3217 if (has_var_break)
3218 scop = scop_add_break(scop, id_break_test, isl_set_copy(domain),
3219 isl_val_copy(inc));
3220 if (id_test) {
3221 scop = scop_add_while(scop_cond, scop, id_test, domain,
3222 isl_val_copy(inc));
3223 isl_set_free(valid_inc);
3224 } else {
3225 scop = pet_scop_restrict_context(scop, valid_inc);
3226 scop = pet_scop_restrict_context(scop, valid_cond_next);
3227 scop = pet_scop_restrict_context(scop, valid_cond_init);
3228 isl_set_free(domain);
3230 clear_assignment(assigned_value, iv);
3232 isl_val_free(inc);
3234 scop = pet_scop_restrict_context(scop, isl_set_params(valid_init));
3236 return scop;
3239 /* Try and construct a pet_scop corresponding to a compound statement.
3241 * "skip_declarations" is set if we should skip initial declarations
3242 * in the children of the compound statements. This then implies
3243 * that this sequence of children should not be treated as a block
3244 * since the initial statements may be skipped.
3246 struct pet_scop *PetScan::extract(CompoundStmt *stmt, bool skip_declarations)
3248 return extract(stmt->children(), !skip_declarations, skip_declarations);
3251 /* Extract a pet_expr from an isl_id created by either pet_nested_clang_expr or
3252 * pet_nested_pet_expr.
3253 * In the first case, the isl_id has no name and
3254 * the user pointer points to a clang::Expr object.
3255 * In the second case, the isl_id has name "__pet_expr" and
3256 * the user pointer points to a pet_expr object.
3258 __isl_give pet_expr *PetScan::extract_expr(__isl_keep isl_id *id)
3260 if (!isl_id_get_name(id))
3261 return extract_expr((Expr *) isl_id_get_user(id));
3262 else
3263 return pet_expr_copy((pet_expr *) isl_id_get_user(id));
3266 /* For each nested access parameter in "space",
3267 * construct a corresponding pet_expr, place it in args and
3268 * record its position in "param2pos".
3269 * "n_arg" is the number of elements that are already in args.
3270 * The position recorded in "param2pos" takes this number into account.
3271 * If the pet_expr corresponding to a parameter is identical to
3272 * the pet_expr corresponding to an earlier parameter, then these two
3273 * parameters are made to refer to the same element in args.
3275 * Return the final number of elements in args or -1 if an error has occurred.
3277 int PetScan::extract_nested(__isl_keep isl_space *space,
3278 int n_arg, pet_expr **args, std::map<int,int> &param2pos)
3280 int nparam;
3282 nparam = isl_space_dim(space, isl_dim_param);
3283 for (int i = 0; i < nparam; ++i) {
3284 int j;
3285 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
3287 if (!pet_nested_in_id(id)) {
3288 isl_id_free(id);
3289 continue;
3292 args[n_arg] = extract_expr(id);
3293 isl_id_free(id);
3294 if (!args[n_arg])
3295 return -1;
3297 for (j = 0; j < n_arg; ++j)
3298 if (pet_expr_is_equal(args[j], args[n_arg]))
3299 break;
3301 if (j < n_arg) {
3302 pet_expr_free(args[n_arg]);
3303 args[n_arg] = NULL;
3304 param2pos[i] = j;
3305 } else
3306 param2pos[i] = n_arg++;
3309 return n_arg;
3312 /* For each nested access parameter in the access relations in "expr",
3313 * construct a corresponding pet_expr, place it in the arguments of "expr"
3314 * and record its position in "param2pos".
3315 * n is the number of nested access parameters.
3317 __isl_give pet_expr *PetScan::extract_nested(__isl_take pet_expr *expr, int n,
3318 std::map<int,int> &param2pos)
3320 isl_space *space;
3321 int i;
3322 pet_expr **args;
3324 args = isl_calloc_array(ctx, pet_expr *, n);
3325 if (!args)
3326 return pet_expr_free(expr);
3328 space = pet_expr_access_get_parameter_space(expr);
3329 n = extract_nested(space, 0, args, param2pos);
3330 isl_space_free(space);
3332 if (n < 0)
3333 expr = pet_expr_free(expr);
3334 else
3335 expr = pet_expr_set_n_arg(expr, n);
3337 for (i = 0; i < n; ++i)
3338 expr = pet_expr_set_arg(expr, i, args[i]);
3339 free(args);
3341 return expr;
3344 /* Look for parameters in any access relation in "expr" that
3345 * refer to nested accesses. In particular, these are
3346 * parameters with either no name or with name "__pet_expr".
3348 * If there are any such parameters, then the domain of the index
3349 * expression and the access relation, which is still [] at this point,
3350 * is replaced by [[] -> [t_1,...,t_n]], with n the number of these parameters
3351 * (after identifying identical nested accesses).
3353 * This transformation is performed in several steps.
3354 * We first extract the arguments in extract_nested.
3355 * param2pos maps the original parameter position to the position
3356 * of the argument.
3357 * Then we move these parameters to input dimensions.
3358 * t2pos maps the positions of these temporary input dimensions
3359 * to the positions of the corresponding arguments.
3360 * Finally, we express these temporary dimensions in terms of the domain
3361 * [[] -> [t_1,...,t_n]] and precompose index expression and access
3362 * relations with this function.
3364 __isl_give pet_expr *PetScan::resolve_nested(__isl_take pet_expr *expr)
3366 int n;
3367 int nparam;
3368 isl_space *space;
3369 isl_local_space *ls;
3370 isl_aff *aff;
3371 isl_multi_aff *ma;
3372 std::map<int,int> param2pos;
3373 std::map<int,int> t2pos;
3375 if (!expr)
3376 return expr;
3378 n = pet_expr_get_n_arg(expr);
3379 for (int i = 0; i < n; ++i) {
3380 pet_expr *arg;
3381 arg = pet_expr_get_arg(expr, i);
3382 arg = resolve_nested(arg);
3383 expr = pet_expr_set_arg(expr, i, arg);
3386 if (pet_expr_get_type(expr) != pet_expr_access)
3387 return expr;
3389 space = pet_expr_access_get_parameter_space(expr);
3390 n = pet_nested_n_in_space(space);
3391 isl_space_free(space);
3392 if (n == 0)
3393 return expr;
3395 expr = extract_nested(expr, n, param2pos);
3396 if (!expr)
3397 return NULL;
3399 expr = pet_expr_access_align_params(expr);
3400 if (!expr)
3401 return NULL;
3403 n = 0;
3404 space = pet_expr_access_get_parameter_space(expr);
3405 nparam = isl_space_dim(space, isl_dim_param);
3406 for (int i = nparam - 1; i >= 0; --i) {
3407 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
3408 if (!pet_nested_in_id(id)) {
3409 isl_id_free(id);
3410 continue;
3413 expr = pet_expr_access_move_dims(expr,
3414 isl_dim_in, n, isl_dim_param, i, 1);
3415 t2pos[n] = param2pos[i];
3416 n++;
3418 isl_id_free(id);
3420 isl_space_free(space);
3422 space = pet_expr_access_get_parameter_space(expr);
3423 space = isl_space_set_from_params(space);
3424 space = isl_space_add_dims(space, isl_dim_set,
3425 pet_expr_get_n_arg(expr));
3426 space = isl_space_wrap(isl_space_from_range(space));
3427 ls = isl_local_space_from_space(isl_space_copy(space));
3428 space = isl_space_from_domain(space);
3429 space = isl_space_add_dims(space, isl_dim_out, n);
3430 ma = isl_multi_aff_zero(space);
3432 for (int i = 0; i < n; ++i) {
3433 aff = isl_aff_var_on_domain(isl_local_space_copy(ls),
3434 isl_dim_set, t2pos[i]);
3435 ma = isl_multi_aff_set_aff(ma, i, aff);
3437 isl_local_space_free(ls);
3439 expr = pet_expr_access_pullback_multi_aff(expr, ma);
3441 return expr;
3444 /* Return the file offset of the expansion location of "Loc".
3446 static unsigned getExpansionOffset(SourceManager &SM, SourceLocation Loc)
3448 return SM.getFileOffset(SM.getExpansionLoc(Loc));
3451 #ifdef HAVE_FINDLOCATIONAFTERTOKEN
3453 /* Return a SourceLocation for the location after the first semicolon
3454 * after "loc". If Lexer::findLocationAfterToken is available, we simply
3455 * call it and also skip trailing spaces and newline.
3457 static SourceLocation location_after_semi(SourceLocation loc, SourceManager &SM,
3458 const LangOptions &LO)
3460 return Lexer::findLocationAfterToken(loc, tok::semi, SM, LO, true);
3463 #else
3465 /* Return a SourceLocation for the location after the first semicolon
3466 * after "loc". If Lexer::findLocationAfterToken is not available,
3467 * we look in the underlying character data for the first semicolon.
3469 static SourceLocation location_after_semi(SourceLocation loc, SourceManager &SM,
3470 const LangOptions &LO)
3472 const char *semi;
3473 const char *s = SM.getCharacterData(loc);
3475 semi = strchr(s, ';');
3476 if (!semi)
3477 return SourceLocation();
3478 return loc.getFileLocWithOffset(semi + 1 - s);
3481 #endif
3483 /* If the token at "loc" is the first token on the line, then return
3484 * a location referring to the start of the line.
3485 * Otherwise, return "loc".
3487 * This function is used to extend a scop to the start of the line
3488 * if the first token of the scop is also the first token on the line.
3490 * We look for the first token on the line. If its location is equal to "loc",
3491 * then the latter is the location of the first token on the line.
3493 static SourceLocation move_to_start_of_line_if_first_token(SourceLocation loc,
3494 SourceManager &SM, const LangOptions &LO)
3496 std::pair<FileID, unsigned> file_offset_pair;
3497 llvm::StringRef file;
3498 const char *pos;
3499 Token tok;
3500 SourceLocation token_loc, line_loc;
3501 int col;
3503 loc = SM.getExpansionLoc(loc);
3504 col = SM.getExpansionColumnNumber(loc);
3505 line_loc = loc.getLocWithOffset(1 - col);
3506 file_offset_pair = SM.getDecomposedLoc(line_loc);
3507 file = SM.getBufferData(file_offset_pair.first, NULL);
3508 pos = file.data() + file_offset_pair.second;
3510 Lexer lexer(SM.getLocForStartOfFile(file_offset_pair.first), LO,
3511 file.begin(), pos, file.end());
3512 lexer.LexFromRawLexer(tok);
3513 token_loc = tok.getLocation();
3515 if (token_loc == loc)
3516 return line_loc;
3517 else
3518 return loc;
3521 /* Update start and end of "scop" to include the region covered by "range".
3522 * If "skip_semi" is set, then we assume "range" is followed by
3523 * a semicolon and also include this semicolon.
3525 struct pet_scop *PetScan::update_scop_start_end(struct pet_scop *scop,
3526 SourceRange range, bool skip_semi)
3528 SourceLocation loc = range.getBegin();
3529 SourceManager &SM = PP.getSourceManager();
3530 const LangOptions &LO = PP.getLangOpts();
3531 unsigned start, end;
3533 loc = move_to_start_of_line_if_first_token(loc, SM, LO);
3534 start = getExpansionOffset(SM, loc);
3535 loc = range.getEnd();
3536 if (skip_semi)
3537 loc = location_after_semi(loc, SM, LO);
3538 else
3539 loc = PP.getLocForEndOfToken(loc);
3540 end = getExpansionOffset(SM, loc);
3542 scop = pet_scop_update_start_end(scop, start, end);
3543 return scop;
3546 /* Convert a top-level pet_expr to a pet_scop with one statement.
3547 * This mainly involves resolving nested expression parameters
3548 * and setting the name of the iteration space.
3549 * The name is given by "label" if it is non-NULL. Otherwise,
3550 * it is of the form S_<n_stmt>.
3551 * start and end of the pet_scop are derived from "range" and "skip_semi".
3552 * In particular, if "skip_semi" is set then the semicolon following "range"
3553 * is also included.
3555 struct pet_scop *PetScan::extract(__isl_take pet_expr *expr, SourceRange range,
3556 bool skip_semi, __isl_take isl_id *label)
3558 struct pet_stmt *ps;
3559 struct pet_scop *scop;
3560 SourceLocation loc = range.getBegin();
3561 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3563 expr = resolve_nested(expr);
3564 ps = pet_stmt_from_pet_expr(line, label, n_stmt++, expr);
3565 scop = pet_scop_from_pet_stmt(ctx, ps);
3567 scop = update_scop_start_end(scop, range, skip_semi);
3568 return scop;
3571 /* Check if we can extract an affine expression from "expr".
3572 * Return the expressions as an isl_pw_aff if we can and NULL otherwise.
3573 * We turn on autodetection so that we won't generate any warnings
3574 * and turn off nesting, so that we won't accept any non-affine constructs.
3576 __isl_give isl_pw_aff *PetScan::try_extract_affine(Expr *expr)
3578 isl_pw_aff *pwaff;
3579 int save_autodetect = options->autodetect;
3580 bool save_nesting = nesting_enabled;
3582 options->autodetect = 1;
3583 nesting_enabled = false;
3585 pwaff = extract_affine(expr);
3587 options->autodetect = save_autodetect;
3588 nesting_enabled = save_nesting;
3590 return pwaff;
3593 /* Check if we can extract an affine constraint from "expr".
3594 * Return the constraint as an isl_set if we can and NULL otherwise.
3595 * We turn on autodetection so that we won't generate any warnings
3596 * and turn off nesting, so that we won't accept any non-affine constructs.
3598 __isl_give isl_pw_aff *PetScan::try_extract_affine_condition(Expr *expr)
3600 isl_pw_aff *cond;
3601 int save_autodetect = options->autodetect;
3602 bool save_nesting = nesting_enabled;
3604 options->autodetect = 1;
3605 nesting_enabled = false;
3607 cond = extract_condition(expr);
3609 options->autodetect = save_autodetect;
3610 nesting_enabled = save_nesting;
3612 return cond;
3615 /* Check whether "expr" is an affine constraint.
3617 bool PetScan::is_affine_condition(Expr *expr)
3619 isl_pw_aff *cond;
3621 cond = try_extract_affine_condition(expr);
3622 isl_pw_aff_free(cond);
3624 return cond != NULL;
3627 /* Check if we can extract a condition from "expr".
3628 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
3629 * If allow_nested is set, then the condition may involve parameters
3630 * corresponding to nested accesses.
3631 * We turn on autodetection so that we won't generate any warnings.
3633 __isl_give isl_pw_aff *PetScan::try_extract_nested_condition(Expr *expr)
3635 isl_pw_aff *cond;
3636 int save_autodetect = options->autodetect;
3637 bool save_nesting = nesting_enabled;
3639 options->autodetect = 1;
3640 nesting_enabled = allow_nested;
3641 cond = extract_condition(expr);
3643 options->autodetect = save_autodetect;
3644 nesting_enabled = save_nesting;
3646 return cond;
3649 /* If the top-level expression of "stmt" is an assignment, then
3650 * return that assignment as a BinaryOperator.
3651 * Otherwise return NULL.
3653 static BinaryOperator *top_assignment_or_null(Stmt *stmt)
3655 BinaryOperator *ass;
3657 if (!stmt)
3658 return NULL;
3659 if (stmt->getStmtClass() != Stmt::BinaryOperatorClass)
3660 return NULL;
3662 ass = cast<BinaryOperator>(stmt);
3663 if(ass->getOpcode() != BO_Assign)
3664 return NULL;
3666 return ass;
3669 /* Check if the given if statement is a conditional assignement
3670 * with a non-affine condition. If so, construct a pet_scop
3671 * corresponding to this conditional assignment. Otherwise return NULL.
3673 * In particular we check if "stmt" is of the form
3675 * if (condition)
3676 * a = f(...);
3677 * else
3678 * a = g(...);
3680 * where a is some array or scalar access.
3681 * The constructed pet_scop then corresponds to the expression
3683 * a = condition ? f(...) : g(...)
3685 * All access relations in f(...) are intersected with condition
3686 * while all access relation in g(...) are intersected with the complement.
3688 struct pet_scop *PetScan::extract_conditional_assignment(IfStmt *stmt)
3690 BinaryOperator *ass_then, *ass_else;
3691 isl_multi_pw_aff *write_then, *write_else;
3692 isl_set *cond, *comp;
3693 isl_multi_pw_aff *index;
3694 isl_pw_aff *pa;
3695 int equal;
3696 int type_size;
3697 pet_expr *pe_cond, *pe_then, *pe_else, *pe, *pe_write;
3698 bool save_nesting = nesting_enabled;
3700 if (!options->detect_conditional_assignment)
3701 return NULL;
3703 ass_then = top_assignment_or_null(stmt->getThen());
3704 ass_else = top_assignment_or_null(stmt->getElse());
3706 if (!ass_then || !ass_else)
3707 return NULL;
3709 if (is_affine_condition(stmt->getCond()))
3710 return NULL;
3712 write_then = extract_index(ass_then->getLHS());
3713 write_else = extract_index(ass_else->getLHS());
3715 equal = isl_multi_pw_aff_plain_is_equal(write_then, write_else);
3716 isl_multi_pw_aff_free(write_else);
3717 if (equal < 0 || !equal) {
3718 isl_multi_pw_aff_free(write_then);
3719 return NULL;
3722 nesting_enabled = allow_nested;
3723 pa = extract_condition(stmt->getCond());
3724 nesting_enabled = save_nesting;
3725 cond = isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa));
3726 comp = isl_pw_aff_zero_set(isl_pw_aff_copy(pa));
3727 index = isl_multi_pw_aff_from_pw_aff(pa);
3729 pe_cond = pet_expr_from_index(index);
3731 pe_then = extract_expr(ass_then->getRHS());
3732 pe_then = pet_expr_restrict(pe_then, cond);
3733 pe_else = extract_expr(ass_else->getRHS());
3734 pe_else = pet_expr_restrict(pe_else, comp);
3736 pe = pet_expr_new_ternary(pe_cond, pe_then, pe_else);
3737 type_size = get_type_size(ass_then->getType(), ast_context);
3738 pe_write = pet_expr_from_index_and_depth(type_size, write_then,
3739 extract_depth(write_then));
3740 pe_write = pet_expr_access_set_write(pe_write, 1);
3741 pe_write = pet_expr_access_set_read(pe_write, 0);
3742 pe = pet_expr_new_binary(type_size, pet_op_assign, pe_write, pe);
3743 return extract(pe, stmt->getSourceRange(), false);
3746 /* Create a pet_scop with a single statement with name S_<stmt_nr>,
3747 * evaluating "cond" and writing the result to a virtual scalar,
3748 * as expressed by "index".
3750 struct pet_scop *PetScan::extract_non_affine_condition(Expr *cond, int stmt_nr,
3751 __isl_take isl_multi_pw_aff *index)
3753 pet_expr *expr, *write;
3754 struct pet_stmt *ps;
3755 SourceLocation loc = cond->getLocStart();
3756 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3758 write = pet_expr_from_index(index);
3759 write = pet_expr_access_set_write(write, 1);
3760 write = pet_expr_access_set_read(write, 0);
3761 expr = extract_expr(cond);
3762 expr = resolve_nested(expr);
3763 expr = pet_expr_new_binary(1, pet_op_assign, write, expr);
3764 ps = pet_stmt_from_pet_expr(line, NULL, stmt_nr, expr);
3765 return pet_scop_from_pet_stmt(ctx, ps);
3768 extern "C" {
3769 static __isl_give pet_expr *embed_access(__isl_take pet_expr *expr,
3770 void *user);
3773 /* Precompose the access relation and the index expression associated
3774 * to "expr" with the function pointed to by "user",
3775 * thereby embedding the access relation in the domain of this function.
3776 * The initial domain of the access relation and the index expression
3777 * is the zero-dimensional domain.
3779 static __isl_give pet_expr *embed_access(__isl_take pet_expr *expr, void *user)
3781 isl_multi_aff *ma = (isl_multi_aff *) user;
3783 return pet_expr_access_pullback_multi_aff(expr, isl_multi_aff_copy(ma));
3786 /* Precompose all access relations in "expr" with "ma", thereby
3787 * embedding them in the domain of "ma".
3789 static __isl_give pet_expr *embed(__isl_take pet_expr *expr,
3790 __isl_keep isl_multi_aff *ma)
3792 return pet_expr_map_access(expr, &embed_access, ma);
3795 /* For each nested access parameter in the domain of "stmt",
3796 * construct a corresponding pet_expr, place it before the original
3797 * elements in stmt->args and record its position in "param2pos".
3798 * n is the number of nested access parameters.
3800 struct pet_stmt *PetScan::extract_nested(struct pet_stmt *stmt, int n,
3801 std::map<int,int> &param2pos)
3803 int i;
3804 isl_space *space;
3805 int n_arg;
3806 pet_expr **args;
3808 n_arg = stmt->n_arg;
3809 args = isl_calloc_array(ctx, pet_expr *, n + n_arg);
3810 if (!args)
3811 goto error;
3813 space = isl_set_get_space(stmt->domain);
3814 n_arg = extract_nested(space, 0, args, param2pos);
3815 isl_space_free(space);
3817 if (n_arg < 0)
3818 goto error;
3820 for (i = 0; i < stmt->n_arg; ++i)
3821 args[n_arg + i] = stmt->args[i];
3822 free(stmt->args);
3823 stmt->args = args;
3824 stmt->n_arg += n_arg;
3826 return stmt;
3827 error:
3828 if (args) {
3829 for (i = 0; i < n; ++i)
3830 pet_expr_free(args[i]);
3831 free(args);
3833 pet_stmt_free(stmt);
3834 return NULL;
3837 /* Check whether any of the arguments i of "stmt" starting at position "n"
3838 * is equal to one of the first "n" arguments j.
3839 * If so, combine the constraints on arguments i and j and remove
3840 * argument i.
3842 static struct pet_stmt *remove_duplicate_arguments(struct pet_stmt *stmt, int n)
3844 int i, j;
3845 isl_map *map;
3847 if (!stmt)
3848 return NULL;
3849 if (n == 0)
3850 return stmt;
3851 if (n == stmt->n_arg)
3852 return stmt;
3854 map = isl_set_unwrap(stmt->domain);
3856 for (i = stmt->n_arg - 1; i >= n; --i) {
3857 for (j = 0; j < n; ++j)
3858 if (pet_expr_is_equal(stmt->args[i], stmt->args[j]))
3859 break;
3860 if (j >= n)
3861 continue;
3863 map = isl_map_equate(map, isl_dim_out, i, isl_dim_out, j);
3864 map = isl_map_project_out(map, isl_dim_out, i, 1);
3866 pet_expr_free(stmt->args[i]);
3867 for (j = i; j + 1 < stmt->n_arg; ++j)
3868 stmt->args[j] = stmt->args[j + 1];
3869 stmt->n_arg--;
3872 stmt->domain = isl_map_wrap(map);
3873 if (!stmt->domain)
3874 goto error;
3875 return stmt;
3876 error:
3877 pet_stmt_free(stmt);
3878 return NULL;
3881 /* Look for parameters in the iteration domain of "stmt" that
3882 * refer to nested accesses. In particular, these are
3883 * parameters with either no name or with name "__pet_expr".
3885 * If there are any such parameters, then as many extra variables
3886 * (after identifying identical nested accesses) are inserted in the
3887 * range of the map wrapped inside the domain, before the original variables.
3888 * If the original domain is not a wrapped map, then a new wrapped
3889 * map is created with zero output dimensions.
3890 * The parameters are then equated to the corresponding output dimensions
3891 * and subsequently projected out, from the iteration domain,
3892 * the schedule and the access relations.
3893 * For each of the output dimensions, a corresponding argument
3894 * expression is inserted. Initially they are created with
3895 * a zero-dimensional domain, so they have to be embedded
3896 * in the current iteration domain.
3897 * param2pos maps the position of the parameter to the position
3898 * of the corresponding output dimension in the wrapped map.
3900 struct pet_stmt *PetScan::resolve_nested(struct pet_stmt *stmt)
3902 int n;
3903 int nparam;
3904 unsigned n_arg;
3905 isl_map *map;
3906 isl_space *space;
3907 isl_multi_aff *ma;
3908 std::map<int,int> param2pos;
3910 if (!stmt)
3911 return NULL;
3913 n = pet_nested_n_in_set(stmt->domain);
3914 if (n == 0)
3915 return stmt;
3917 n_arg = stmt->n_arg;
3918 stmt = extract_nested(stmt, n, param2pos);
3919 if (!stmt)
3920 return NULL;
3922 n = stmt->n_arg - n_arg;
3923 nparam = isl_set_dim(stmt->domain, isl_dim_param);
3924 if (isl_set_is_wrapping(stmt->domain))
3925 map = isl_set_unwrap(stmt->domain);
3926 else
3927 map = isl_map_from_domain(stmt->domain);
3928 map = isl_map_insert_dims(map, isl_dim_out, 0, n);
3930 for (int i = nparam - 1; i >= 0; --i) {
3931 isl_id *id;
3933 if (!pet_nested_in_map(map, i))
3934 continue;
3936 id = pet_expr_access_get_id(stmt->args[param2pos[i]]);
3937 map = isl_map_set_dim_id(map, isl_dim_out, param2pos[i], id);
3938 map = isl_map_equate(map, isl_dim_param, i, isl_dim_out,
3939 param2pos[i]);
3940 map = isl_map_project_out(map, isl_dim_param, i, 1);
3943 stmt->domain = isl_map_wrap(map);
3945 space = isl_space_unwrap(isl_set_get_space(stmt->domain));
3946 space = isl_space_from_domain(isl_space_domain(space));
3947 ma = isl_multi_aff_zero(space);
3948 for (int pos = 0; pos < n; ++pos)
3949 stmt->args[pos] = embed(stmt->args[pos], ma);
3950 isl_multi_aff_free(ma);
3952 stmt = pet_stmt_remove_nested_parameters(stmt);
3953 stmt = remove_duplicate_arguments(stmt, n);
3955 return stmt;
3958 /* For each statement in "scop", move the parameters that correspond
3959 * to nested access into the ranges of the domains and create
3960 * corresponding argument expressions.
3962 struct pet_scop *PetScan::resolve_nested(struct pet_scop *scop)
3964 if (!scop)
3965 return NULL;
3967 for (int i = 0; i < scop->n_stmt; ++i) {
3968 scop->stmts[i] = resolve_nested(scop->stmts[i]);
3969 if (!scop->stmts[i])
3970 goto error;
3973 return scop;
3974 error:
3975 pet_scop_free(scop);
3976 return NULL;
3979 /* Given an access expression "expr", is the variable accessed by
3980 * "expr" assigned anywhere inside "scop"?
3982 static bool is_assigned(__isl_keep pet_expr *expr, pet_scop *scop)
3984 bool assigned = false;
3985 isl_id *id;
3987 id = pet_expr_access_get_id(expr);
3988 assigned = pet_scop_writes(scop, id);
3989 isl_id_free(id);
3991 return assigned;
3994 /* Are all nested access parameters in "pa" allowed given "scop".
3995 * In particular, is none of them written by anywhere inside "scop".
3997 * If "scop" has any skip conditions, then no nested access parameters
3998 * are allowed. In particular, if there is any nested access in a guard
3999 * for a piece of code containing a "continue", then we want to introduce
4000 * a separate statement for evaluating this guard so that we can express
4001 * that the result is false for all previous iterations.
4003 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff *pa, pet_scop *scop)
4005 int nparam;
4007 if (!scop)
4008 return true;
4010 if (!pet_nested_any_in_pw_aff(pa))
4011 return true;
4013 if (pet_scop_has_skip(scop, pet_skip_now))
4014 return false;
4016 nparam = isl_pw_aff_dim(pa, isl_dim_param);
4017 for (int i = 0; i < nparam; ++i) {
4018 isl_id *id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
4019 pet_expr *expr;
4020 bool allowed;
4022 if (!pet_nested_in_id(id)) {
4023 isl_id_free(id);
4024 continue;
4027 expr = extract_expr(id);
4028 allowed = pet_expr_get_type(expr) == pet_expr_access &&
4029 !is_assigned(expr, scop);
4031 pet_expr_free(expr);
4032 isl_id_free(id);
4034 if (!allowed)
4035 return false;
4038 return true;
4041 /* Construct a pet_scop for a non-affine if statement.
4043 * We create a separate statement that writes the result
4044 * of the non-affine condition to a virtual scalar.
4045 * A constraint requiring the value of this virtual scalar to be one
4046 * is added to the iteration domains of the then branch.
4047 * Similarly, a constraint requiring the value of this virtual scalar
4048 * to be zero is added to the iteration domains of the else branch, if any.
4049 * We adjust the schedules to ensure that the virtual scalar is written
4050 * before it is read.
4052 * If there are any breaks or continues in the then and/or else
4053 * branches, then we may have to compute a new skip condition.
4054 * This is handled using a pet_skip_info object.
4055 * On initialization, the object checks if skip conditions need
4056 * to be computed. If so, it does so in pet_skip_info_if_extract_index and
4057 * adds them in pet_skip_info_if_add.
4059 struct pet_scop *PetScan::extract_non_affine_if(Expr *cond,
4060 struct pet_scop *scop_then, struct pet_scop *scop_else,
4061 bool have_else, int stmt_id)
4063 struct pet_scop *scop;
4064 isl_multi_pw_aff *test_index;
4065 int int_size;
4066 int save_n_stmt = n_stmt;
4068 test_index = pet_create_test_index(ctx, n_test++);
4069 n_stmt = stmt_id;
4070 scop = extract_non_affine_condition(cond, n_stmt++,
4071 isl_multi_pw_aff_copy(test_index));
4072 n_stmt = save_n_stmt;
4073 scop = scop_add_array(scop, test_index, ast_context);
4075 pet_skip_info skip;
4076 pet_skip_info_if_init(&skip, ctx, scop_then, scop_else, have_else, 0);
4077 int_size = ast_context.getTypeInfo(ast_context.IntTy).first / 8;
4078 pet_skip_info_if_extract_index(&skip, test_index, int_size,
4079 &n_stmt, &n_test);
4081 scop = pet_scop_prefix(scop, 0);
4082 scop_then = pet_scop_prefix(scop_then, 1);
4083 scop_then = pet_scop_filter(scop_then,
4084 isl_multi_pw_aff_copy(test_index), 1);
4085 if (have_else) {
4086 scop_else = pet_scop_prefix(scop_else, 1);
4087 scop_else = pet_scop_filter(scop_else, test_index, 0);
4088 scop_then = pet_scop_add_par(ctx, scop_then, scop_else);
4089 } else
4090 isl_multi_pw_aff_free(test_index);
4092 scop = pet_scop_add_seq(ctx, scop, scop_then);
4094 scop = pet_skip_info_if_add(&skip, scop, 2);
4096 return scop;
4099 /* Construct a pet_scop for an if statement.
4101 * If the condition fits the pattern of a conditional assignment,
4102 * then it is handled by extract_conditional_assignment.
4103 * Otherwise, we do the following.
4105 * If the condition is affine, then the condition is added
4106 * to the iteration domains of the then branch, while the
4107 * opposite of the condition in added to the iteration domains
4108 * of the else branch, if any.
4109 * We allow the condition to be dynamic, i.e., to refer to
4110 * scalars or array elements that may be written to outside
4111 * of the given if statement. These nested accesses are then represented
4112 * as output dimensions in the wrapping iteration domain.
4113 * If it is also written _inside_ the then or else branch, then
4114 * we treat the condition as non-affine.
4115 * As explained in extract_non_affine_if, this will introduce
4116 * an extra statement.
4117 * For aesthetic reasons, we want this statement to have a statement
4118 * number that is lower than those of the then and else branches.
4119 * In order to evaluate if we will need such a statement, however, we
4120 * first construct scops for the then and else branches.
4121 * We therefore reserve a statement number if we might have to
4122 * introduce such an extra statement.
4124 * If the condition is not affine, then the scop is created in
4125 * extract_non_affine_if.
4127 * If there are any breaks or continues in the then and/or else
4128 * branches, then we may have to compute a new skip condition.
4129 * This is handled using a pet_skip_info object.
4130 * On initialization, the object checks if skip conditions need
4131 * to be computed. If so, it does so in pet_skip_info_if_extract_cond and
4132 * adds them in pet_skip_info_if_add.
4134 struct pet_scop *PetScan::extract(IfStmt *stmt)
4136 struct pet_scop *scop_then, *scop_else = NULL, *scop;
4137 isl_pw_aff *cond;
4138 int stmt_id;
4139 int int_size;
4140 isl_set *set;
4141 isl_set *valid;
4143 clear_assignments clear(assigned_value);
4144 clear.TraverseStmt(stmt->getThen());
4145 if (stmt->getElse())
4146 clear.TraverseStmt(stmt->getElse());
4148 scop = extract_conditional_assignment(stmt);
4149 if (scop)
4150 return scop;
4152 cond = try_extract_nested_condition(stmt->getCond());
4153 if (allow_nested && (!cond || pet_nested_any_in_pw_aff(cond)))
4154 stmt_id = n_stmt++;
4157 assigned_value_cache cache(assigned_value);
4158 scop_then = extract(stmt->getThen());
4161 if (stmt->getElse()) {
4162 assigned_value_cache cache(assigned_value);
4163 scop_else = extract(stmt->getElse());
4164 if (options->autodetect) {
4165 if (scop_then && !scop_else) {
4166 partial = true;
4167 isl_pw_aff_free(cond);
4168 return scop_then;
4170 if (!scop_then && scop_else) {
4171 partial = true;
4172 isl_pw_aff_free(cond);
4173 return scop_else;
4178 if (cond &&
4179 (!is_nested_allowed(cond, scop_then) ||
4180 (stmt->getElse() && !is_nested_allowed(cond, scop_else)))) {
4181 isl_pw_aff_free(cond);
4182 cond = NULL;
4184 if (allow_nested && !cond)
4185 return extract_non_affine_if(stmt->getCond(), scop_then,
4186 scop_else, stmt->getElse(), stmt_id);
4188 if (!cond)
4189 cond = extract_condition(stmt->getCond());
4191 pet_skip_info skip;
4192 pet_skip_info_if_init(&skip, ctx, scop_then, scop_else,
4193 stmt->getElse() != NULL, 1);
4194 pet_skip_info_if_extract_cond(&skip, cond, int_size, &n_stmt, &n_test);
4196 valid = isl_pw_aff_domain(isl_pw_aff_copy(cond));
4197 set = isl_pw_aff_non_zero_set(cond);
4198 scop = pet_scop_restrict(scop_then, isl_set_params(isl_set_copy(set)));
4200 if (stmt->getElse()) {
4201 set = isl_set_subtract(isl_set_copy(valid), set);
4202 scop_else = pet_scop_restrict(scop_else, isl_set_params(set));
4203 scop = pet_scop_add_par(ctx, scop, scop_else);
4204 } else
4205 isl_set_free(set);
4206 scop = resolve_nested(scop);
4207 scop = pet_scop_restrict_context(scop, isl_set_params(valid));
4209 if (pet_skip_info_has_skip(&skip))
4210 scop = pet_scop_prefix(scop, 0);
4211 scop = pet_skip_info_if_add(&skip, scop, 1);
4213 return scop;
4216 /* Try and construct a pet_scop for a label statement.
4217 * We currently only allow labels on expression statements.
4219 struct pet_scop *PetScan::extract(LabelStmt *stmt)
4221 isl_id *label;
4222 Stmt *sub;
4224 sub = stmt->getSubStmt();
4225 if (!isa<Expr>(sub)) {
4226 unsupported(stmt);
4227 return NULL;
4230 label = isl_id_alloc(ctx, stmt->getName(), NULL);
4232 return extract(extract_expr(cast<Expr>(sub)), stmt->getSourceRange(),
4233 true, label);
4236 /* Return a one-dimensional multi piecewise affine expression that is equal
4237 * to the constant 1 and is defined over a zero-dimensional domain.
4239 static __isl_give isl_multi_pw_aff *one_mpa(isl_ctx *ctx)
4241 isl_space *space;
4242 isl_local_space *ls;
4243 isl_aff *aff;
4245 space = isl_space_set_alloc(ctx, 0, 0);
4246 ls = isl_local_space_from_space(space);
4247 aff = isl_aff_zero_on_domain(ls);
4248 aff = isl_aff_set_constant_si(aff, 1);
4250 return isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
4253 /* Construct a pet_scop for a continue statement.
4255 * We simply create an empty scop with a universal pet_skip_now
4256 * skip condition. This skip condition will then be taken into
4257 * account by the enclosing loop construct, possibly after
4258 * being incorporated into outer skip conditions.
4260 struct pet_scop *PetScan::extract(ContinueStmt *stmt)
4262 pet_scop *scop;
4264 scop = pet_scop_empty(ctx);
4265 if (!scop)
4266 return NULL;
4268 scop = pet_scop_set_skip(scop, pet_skip_now, one_mpa(ctx));
4270 return scop;
4273 /* Construct a pet_scop for a break statement.
4275 * We simply create an empty scop with both a universal pet_skip_now
4276 * skip condition and a universal pet_skip_later skip condition.
4277 * These skip conditions will then be taken into
4278 * account by the enclosing loop construct, possibly after
4279 * being incorporated into outer skip conditions.
4281 struct pet_scop *PetScan::extract(BreakStmt *stmt)
4283 pet_scop *scop;
4284 isl_multi_pw_aff *skip;
4286 scop = pet_scop_empty(ctx);
4287 if (!scop)
4288 return NULL;
4290 skip = one_mpa(ctx);
4291 scop = pet_scop_set_skip(scop, pet_skip_now,
4292 isl_multi_pw_aff_copy(skip));
4293 scop = pet_scop_set_skip(scop, pet_skip_later, skip);
4295 return scop;
4298 /* Try and construct a pet_scop corresponding to "stmt".
4300 * If "stmt" is a compound statement, then "skip_declarations"
4301 * indicates whether we should skip initial declarations in the
4302 * compound statement.
4304 * If the constructed pet_scop is not a (possibly) partial representation
4305 * of "stmt", we update start and end of the pet_scop to those of "stmt".
4306 * In particular, if skip_declarations is set, then we may have skipped
4307 * declarations inside "stmt" and so the pet_scop may not represent
4308 * the entire "stmt".
4309 * Note that this function may be called with "stmt" referring to the entire
4310 * body of the function, including the outer braces. In such cases,
4311 * skip_declarations will be set and the braces will not be taken into
4312 * account in scop->start and scop->end.
4314 struct pet_scop *PetScan::extract(Stmt *stmt, bool skip_declarations)
4316 struct pet_scop *scop;
4318 if (isa<Expr>(stmt))
4319 return extract(extract_expr(cast<Expr>(stmt)),
4320 stmt->getSourceRange(), true);
4322 switch (stmt->getStmtClass()) {
4323 case Stmt::WhileStmtClass:
4324 scop = extract(cast<WhileStmt>(stmt));
4325 break;
4326 case Stmt::ForStmtClass:
4327 scop = extract_for(cast<ForStmt>(stmt));
4328 break;
4329 case Stmt::IfStmtClass:
4330 scop = extract(cast<IfStmt>(stmt));
4331 break;
4332 case Stmt::CompoundStmtClass:
4333 scop = extract(cast<CompoundStmt>(stmt), skip_declarations);
4334 break;
4335 case Stmt::LabelStmtClass:
4336 scop = extract(cast<LabelStmt>(stmt));
4337 break;
4338 case Stmt::ContinueStmtClass:
4339 scop = extract(cast<ContinueStmt>(stmt));
4340 break;
4341 case Stmt::BreakStmtClass:
4342 scop = extract(cast<BreakStmt>(stmt));
4343 break;
4344 case Stmt::DeclStmtClass:
4345 scop = extract(cast<DeclStmt>(stmt));
4346 break;
4347 default:
4348 unsupported(stmt);
4349 return NULL;
4352 if (partial || skip_declarations)
4353 return scop;
4355 scop = update_scop_start_end(scop, stmt->getSourceRange(), false);
4357 return scop;
4360 /* Extract a clone of the kill statement in "scop".
4361 * "scop" is expected to have been created from a DeclStmt
4362 * and should have the kill as its first statement.
4364 struct pet_stmt *PetScan::extract_kill(struct pet_scop *scop)
4366 pet_expr *kill;
4367 struct pet_stmt *stmt;
4368 isl_multi_pw_aff *index;
4369 isl_map *access;
4370 pet_expr *arg;
4372 if (!scop)
4373 return NULL;
4374 if (scop->n_stmt < 1)
4375 isl_die(ctx, isl_error_internal,
4376 "expecting at least one statement", return NULL);
4377 stmt = scop->stmts[0];
4378 if (!pet_stmt_is_kill(stmt))
4379 isl_die(ctx, isl_error_internal,
4380 "expecting kill statement", return NULL);
4382 arg = pet_expr_get_arg(stmt->body, 0);
4383 index = pet_expr_access_get_index(arg);
4384 access = pet_expr_access_get_access(arg);
4385 pet_expr_free(arg);
4386 index = isl_multi_pw_aff_reset_tuple_id(index, isl_dim_in);
4387 access = isl_map_reset_tuple_id(access, isl_dim_in);
4388 kill = pet_expr_kill_from_access_and_index(access, index);
4389 return pet_stmt_from_pet_expr(stmt->line, NULL, n_stmt++, kill);
4392 /* Mark all arrays in "scop" as being exposed.
4394 static struct pet_scop *mark_exposed(struct pet_scop *scop)
4396 if (!scop)
4397 return NULL;
4398 for (int i = 0; i < scop->n_array; ++i)
4399 scop->arrays[i]->exposed = 1;
4400 return scop;
4403 /* Try and construct a pet_scop corresponding to (part of)
4404 * a sequence of statements.
4406 * "block" is set if the sequence respresents the children of
4407 * a compound statement.
4408 * "skip_declarations" is set if we should skip initial declarations
4409 * in the sequence of statements.
4411 * If there are any breaks or continues in the individual statements,
4412 * then we may have to compute a new skip condition.
4413 * This is handled using a pet_skip_info object.
4414 * On initialization, the object checks if skip conditions need
4415 * to be computed. If so, it does so in pet_skip_info_seq_extract and
4416 * adds them in pet_skip_info_seq_add.
4418 * If "block" is set, then we need to insert kill statements at
4419 * the end of the block for any array that has been declared by
4420 * one of the statements in the sequence. Each of these declarations
4421 * results in the construction of a kill statement at the place
4422 * of the declaration, so we simply collect duplicates of
4423 * those kill statements and append these duplicates to the constructed scop.
4425 * If "block" is not set, then any array declared by one of the statements
4426 * in the sequence is marked as being exposed.
4428 * If autodetect is set, then we allow the extraction of only a subrange
4429 * of the sequence of statements. However, if there is at least one statement
4430 * for which we could not construct a scop and the final range contains
4431 * either no statements or at least one kill, then we discard the entire
4432 * range.
4434 struct pet_scop *PetScan::extract(StmtRange stmt_range, bool block,
4435 bool skip_declarations)
4437 pet_scop *scop;
4438 StmtIterator i;
4439 int int_size;
4440 int j;
4441 bool partial_range = false;
4442 set<struct pet_stmt *> kills;
4443 set<struct pet_stmt *>::iterator it;
4445 int_size = ast_context.getTypeInfo(ast_context.IntTy).first / 8;
4447 scop = pet_scop_empty(ctx);
4448 for (i = stmt_range.first, j = 0; i != stmt_range.second; ++i, ++j) {
4449 Stmt *child = *i;
4450 struct pet_scop *scop_i;
4452 if (scop->n_stmt == 0 && skip_declarations &&
4453 child->getStmtClass() == Stmt::DeclStmtClass)
4454 continue;
4456 scop_i = extract(child);
4457 if (scop->n_stmt != 0 && partial) {
4458 pet_scop_free(scop_i);
4459 break;
4461 pet_skip_info skip;
4462 pet_skip_info_seq_init(&skip, ctx, scop, scop_i);
4463 pet_skip_info_seq_extract(&skip, int_size, &n_stmt, &n_test);
4464 if (pet_skip_info_has_skip(&skip))
4465 scop_i = pet_scop_prefix(scop_i, 0);
4466 if (scop_i && child->getStmtClass() == Stmt::DeclStmtClass) {
4467 if (block)
4468 kills.insert(extract_kill(scop_i));
4469 else
4470 scop_i = mark_exposed(scop_i);
4472 scop_i = pet_scop_prefix(scop_i, j);
4473 if (options->autodetect) {
4474 if (scop_i)
4475 scop = pet_scop_add_seq(ctx, scop, scop_i);
4476 else
4477 partial_range = true;
4478 if (scop->n_stmt != 0 && !scop_i)
4479 partial = true;
4480 } else {
4481 scop = pet_scop_add_seq(ctx, scop, scop_i);
4484 scop = pet_skip_info_seq_add(&skip, scop, j);
4486 if (partial || !scop)
4487 break;
4490 for (it = kills.begin(); it != kills.end(); ++it) {
4491 pet_scop *scop_j;
4492 scop_j = pet_scop_from_pet_stmt(ctx, *it);
4493 scop_j = pet_scop_prefix(scop_j, j);
4494 scop = pet_scop_add_seq(ctx, scop, scop_j);
4497 if (scop && partial_range) {
4498 if (scop->n_stmt == 0 || kills.size() != 0) {
4499 pet_scop_free(scop);
4500 return NULL;
4502 partial = true;
4505 return scop;
4508 /* Check if the scop marked by the user is exactly this Stmt
4509 * or part of this Stmt.
4510 * If so, return a pet_scop corresponding to the marked region.
4511 * Otherwise, return NULL.
4513 struct pet_scop *PetScan::scan(Stmt *stmt)
4515 SourceManager &SM = PP.getSourceManager();
4516 unsigned start_off, end_off;
4518 start_off = getExpansionOffset(SM, stmt->getLocStart());
4519 end_off = getExpansionOffset(SM, stmt->getLocEnd());
4521 if (start_off > loc.end)
4522 return NULL;
4523 if (end_off < loc.start)
4524 return NULL;
4525 if (start_off >= loc.start && end_off <= loc.end) {
4526 return extract(stmt);
4529 StmtIterator start;
4530 for (start = stmt->child_begin(); start != stmt->child_end(); ++start) {
4531 Stmt *child = *start;
4532 if (!child)
4533 continue;
4534 start_off = getExpansionOffset(SM, child->getLocStart());
4535 end_off = getExpansionOffset(SM, child->getLocEnd());
4536 if (start_off < loc.start && end_off >= loc.end)
4537 return scan(child);
4538 if (start_off >= loc.start)
4539 break;
4542 StmtIterator end;
4543 for (end = start; end != stmt->child_end(); ++end) {
4544 Stmt *child = *end;
4545 start_off = SM.getFileOffset(child->getLocStart());
4546 if (start_off >= loc.end)
4547 break;
4550 return extract(StmtRange(start, end), false, false);
4553 /* Set the size of index "pos" of "array" to "size".
4554 * In particular, add a constraint of the form
4556 * i_pos < size
4558 * to array->extent and a constraint of the form
4560 * size >= 0
4562 * to array->context.
4564 static struct pet_array *update_size(struct pet_array *array, int pos,
4565 __isl_take isl_pw_aff *size)
4567 isl_set *valid;
4568 isl_set *univ;
4569 isl_set *bound;
4570 isl_space *dim;
4571 isl_aff *aff;
4572 isl_pw_aff *index;
4573 isl_id *id;
4575 valid = isl_set_params(isl_pw_aff_nonneg_set(isl_pw_aff_copy(size)));
4576 array->context = isl_set_intersect(array->context, valid);
4578 dim = isl_set_get_space(array->extent);
4579 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
4580 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, pos, 1);
4581 univ = isl_set_universe(isl_aff_get_domain_space(aff));
4582 index = isl_pw_aff_alloc(univ, aff);
4584 size = isl_pw_aff_add_dims(size, isl_dim_in,
4585 isl_set_dim(array->extent, isl_dim_set));
4586 id = isl_set_get_tuple_id(array->extent);
4587 size = isl_pw_aff_set_tuple_id(size, isl_dim_in, id);
4588 bound = isl_pw_aff_lt_set(index, size);
4590 array->extent = isl_set_intersect(array->extent, bound);
4592 if (!array->context || !array->extent)
4593 goto error;
4595 return array;
4596 error:
4597 pet_array_free(array);
4598 return NULL;
4601 /* Figure out the size of the array at position "pos" and all
4602 * subsequent positions from "type" and update "array" accordingly.
4604 struct pet_array *PetScan::set_upper_bounds(struct pet_array *array,
4605 const Type *type, int pos)
4607 const ArrayType *atype;
4608 isl_pw_aff *size;
4610 if (!array)
4611 return NULL;
4613 if (type->isPointerType()) {
4614 type = type->getPointeeType().getTypePtr();
4615 return set_upper_bounds(array, type, pos + 1);
4617 if (!type->isArrayType())
4618 return array;
4620 type = type->getCanonicalTypeInternal().getTypePtr();
4621 atype = cast<ArrayType>(type);
4623 if (type->isConstantArrayType()) {
4624 const ConstantArrayType *ca = cast<ConstantArrayType>(atype);
4625 size = extract_affine(ca->getSize());
4626 array = update_size(array, pos, size);
4627 } else if (type->isVariableArrayType()) {
4628 const VariableArrayType *vla = cast<VariableArrayType>(atype);
4629 size = extract_affine(vla->getSizeExpr());
4630 array = update_size(array, pos, size);
4633 type = atype->getElementType().getTypePtr();
4635 return set_upper_bounds(array, type, pos + 1);
4638 /* Is "T" the type of a variable length array with static size?
4640 static bool is_vla_with_static_size(QualType T)
4642 const VariableArrayType *vlatype;
4644 if (!T->isVariableArrayType())
4645 return false;
4646 vlatype = cast<VariableArrayType>(T);
4647 return vlatype->getSizeModifier() == VariableArrayType::Static;
4650 /* Return the type of "decl" as an array.
4652 * In particular, if "decl" is a parameter declaration that
4653 * is a variable length array with a static size, then
4654 * return the original type (i.e., the variable length array).
4655 * Otherwise, return the type of decl.
4657 static QualType get_array_type(ValueDecl *decl)
4659 ParmVarDecl *parm;
4660 QualType T;
4662 parm = dyn_cast<ParmVarDecl>(decl);
4663 if (!parm)
4664 return decl->getType();
4666 T = parm->getOriginalType();
4667 if (!is_vla_with_static_size(T))
4668 return decl->getType();
4669 return T;
4672 /* Does "decl" have definition that we can keep track of in a pet_type?
4674 static bool has_printable_definition(RecordDecl *decl)
4676 if (!decl->getDeclName())
4677 return false;
4678 return decl->getLexicalDeclContext() == decl->getDeclContext();
4681 /* Construct and return a pet_array corresponding to the variable "decl".
4682 * In particular, initialize array->extent to
4684 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
4686 * and then call set_upper_bounds to set the upper bounds on the indices
4687 * based on the type of the variable.
4689 * If the base type is that of a record with a top-level definition and
4690 * if "types" is not null, then the RecordDecl corresponding to the type
4691 * is added to "types".
4693 * If the base type is that of a record with no top-level definition,
4694 * then we replace it by "<subfield>".
4696 struct pet_array *PetScan::extract_array(isl_ctx *ctx, ValueDecl *decl,
4697 lex_recorddecl_set *types)
4699 struct pet_array *array;
4700 QualType qt = get_array_type(decl);
4701 const Type *type = qt.getTypePtr();
4702 int depth = array_depth(type);
4703 QualType base = pet_clang_base_type(qt);
4704 string name;
4705 isl_id *id;
4706 isl_space *dim;
4708 array = isl_calloc_type(ctx, struct pet_array);
4709 if (!array)
4710 return NULL;
4712 id = create_decl_id(ctx, decl);
4713 dim = isl_space_set_alloc(ctx, 0, depth);
4714 dim = isl_space_set_tuple_id(dim, isl_dim_set, id);
4716 array->extent = isl_set_nat_universe(dim);
4718 dim = isl_space_params_alloc(ctx, 0);
4719 array->context = isl_set_universe(dim);
4721 array = set_upper_bounds(array, type, 0);
4722 if (!array)
4723 return NULL;
4725 name = base.getAsString();
4727 if (types && base->isRecordType()) {
4728 RecordDecl *decl = pet_clang_record_decl(base);
4729 if (has_printable_definition(decl))
4730 types->insert(decl);
4731 else
4732 name = "<subfield>";
4735 array->element_type = strdup(name.c_str());
4736 array->element_is_record = base->isRecordType();
4737 array->element_size = decl->getASTContext().getTypeInfo(base).first / 8;
4739 return array;
4742 /* Construct and return a pet_array corresponding to the sequence
4743 * of declarations "decls".
4744 * If the sequence contains a single declaration, then it corresponds
4745 * to a simple array access. Otherwise, it corresponds to a member access,
4746 * with the declaration for the substructure following that of the containing
4747 * structure in the sequence of declarations.
4748 * We start with the outermost substructure and then combine it with
4749 * information from the inner structures.
4751 * Additionally, keep track of all required types in "types".
4753 struct pet_array *PetScan::extract_array(isl_ctx *ctx,
4754 vector<ValueDecl *> decls, lex_recorddecl_set *types)
4756 struct pet_array *array;
4757 vector<ValueDecl *>::iterator it;
4759 it = decls.begin();
4761 array = extract_array(ctx, *it, types);
4763 for (++it; it != decls.end(); ++it) {
4764 struct pet_array *parent;
4765 const char *base_name, *field_name;
4766 char *product_name;
4768 parent = array;
4769 array = extract_array(ctx, *it, types);
4770 if (!array)
4771 return pet_array_free(parent);
4773 base_name = isl_set_get_tuple_name(parent->extent);
4774 field_name = isl_set_get_tuple_name(array->extent);
4775 product_name = member_access_name(ctx, base_name, field_name);
4777 array->extent = isl_set_product(isl_set_copy(parent->extent),
4778 array->extent);
4779 if (product_name)
4780 array->extent = isl_set_set_tuple_name(array->extent,
4781 product_name);
4782 array->context = isl_set_intersect(array->context,
4783 isl_set_copy(parent->context));
4785 pet_array_free(parent);
4786 free(product_name);
4788 if (!array->extent || !array->context || !product_name)
4789 return pet_array_free(array);
4792 return array;
4795 /* Add a pet_type corresponding to "decl" to "scop, provided
4796 * it is a member of "types" and it has not been added before
4797 * (i.e., it is not a member of "types_done".
4799 * Since we want the user to be able to print the types
4800 * in the order in which they appear in the scop, we need to
4801 * make sure that types of fields in a structure appear before
4802 * that structure. We therefore call ourselves recursively
4803 * on the types of all record subfields.
4805 static struct pet_scop *add_type(isl_ctx *ctx, struct pet_scop *scop,
4806 RecordDecl *decl, Preprocessor &PP, lex_recorddecl_set &types,
4807 lex_recorddecl_set &types_done)
4809 string s;
4810 llvm::raw_string_ostream S(s);
4811 RecordDecl::field_iterator it;
4813 if (types.find(decl) == types.end())
4814 return scop;
4815 if (types_done.find(decl) != types_done.end())
4816 return scop;
4818 for (it = decl->field_begin(); it != decl->field_end(); ++it) {
4819 RecordDecl *record;
4820 QualType type = it->getType();
4822 if (!type->isRecordType())
4823 continue;
4824 record = pet_clang_record_decl(type);
4825 scop = add_type(ctx, scop, record, PP, types, types_done);
4828 if (strlen(decl->getName().str().c_str()) == 0)
4829 return scop;
4831 decl->print(S, PrintingPolicy(PP.getLangOpts()));
4832 S.str();
4834 scop->types[scop->n_type] = pet_type_alloc(ctx,
4835 decl->getName().str().c_str(), s.c_str());
4836 if (!scop->types[scop->n_type])
4837 return pet_scop_free(scop);
4839 types_done.insert(decl);
4841 scop->n_type++;
4843 return scop;
4846 /* Construct a list of pet_arrays, one for each array (or scalar)
4847 * accessed inside "scop", add this list to "scop" and return the result.
4849 * The context of "scop" is updated with the intersection of
4850 * the contexts of all arrays, i.e., constraints on the parameters
4851 * that ensure that the arrays have a valid (non-negative) size.
4853 * If the any of the extracted arrays refers to a member access,
4854 * then also add the required types to "scop".
4856 struct pet_scop *PetScan::scan_arrays(struct pet_scop *scop)
4858 int i;
4859 array_desc_set arrays;
4860 array_desc_set::iterator it;
4861 lex_recorddecl_set types;
4862 lex_recorddecl_set types_done;
4863 lex_recorddecl_set::iterator types_it;
4864 int n_array;
4865 struct pet_array **scop_arrays;
4867 if (!scop)
4868 return NULL;
4870 pet_scop_collect_arrays(scop, arrays);
4871 if (arrays.size() == 0)
4872 return scop;
4874 n_array = scop->n_array;
4876 scop_arrays = isl_realloc_array(ctx, scop->arrays, struct pet_array *,
4877 n_array + arrays.size());
4878 if (!scop_arrays)
4879 goto error;
4880 scop->arrays = scop_arrays;
4882 for (it = arrays.begin(), i = 0; it != arrays.end(); ++it, ++i) {
4883 struct pet_array *array;
4884 array = extract_array(ctx, *it, &types);
4885 scop->arrays[n_array + i] = array;
4886 if (!scop->arrays[n_array + i])
4887 goto error;
4888 scop->n_array++;
4889 scop->context = isl_set_intersect(scop->context,
4890 isl_set_copy(array->context));
4891 if (!scop->context)
4892 goto error;
4895 if (types.size() == 0)
4896 return scop;
4898 scop->types = isl_alloc_array(ctx, struct pet_type *, types.size());
4899 if (!scop->types)
4900 goto error;
4902 for (types_it = types.begin(); types_it != types.end(); ++types_it)
4903 scop = add_type(ctx, scop, *types_it, PP, types, types_done);
4905 return scop;
4906 error:
4907 pet_scop_free(scop);
4908 return NULL;
4911 /* Bound all parameters in scop->context to the possible values
4912 * of the corresponding C variable.
4914 static struct pet_scop *add_parameter_bounds(struct pet_scop *scop)
4916 int n;
4918 if (!scop)
4919 return NULL;
4921 n = isl_set_dim(scop->context, isl_dim_param);
4922 for (int i = 0; i < n; ++i) {
4923 isl_id *id;
4924 ValueDecl *decl;
4926 id = isl_set_get_dim_id(scop->context, isl_dim_param, i);
4927 if (pet_nested_in_id(id)) {
4928 isl_id_free(id);
4929 isl_die(isl_set_get_ctx(scop->context),
4930 isl_error_internal,
4931 "unresolved nested parameter", goto error);
4933 decl = (ValueDecl *) isl_id_get_user(id);
4934 isl_id_free(id);
4936 scop->context = set_parameter_bounds(scop->context, i, decl);
4938 if (!scop->context)
4939 goto error;
4942 return scop;
4943 error:
4944 pet_scop_free(scop);
4945 return NULL;
4948 /* Construct a pet_scop from the given function.
4950 * If the scop was delimited by scop and endscop pragmas, then we override
4951 * the file offsets by those derived from the pragmas.
4953 struct pet_scop *PetScan::scan(FunctionDecl *fd)
4955 pet_scop *scop;
4956 Stmt *stmt;
4958 stmt = fd->getBody();
4960 if (options->autodetect)
4961 scop = extract(stmt, true);
4962 else {
4963 scop = scan(stmt);
4964 scop = pet_scop_update_start_end(scop, loc.start, loc.end);
4966 scop = pet_scop_detect_parameter_accesses(scop);
4967 scop = scan_arrays(scop);
4968 scop = add_parameter_bounds(scop);
4969 scop = pet_scop_gist(scop, value_bounds);
4971 return scop;