add support for structs
[pet.git] / scan.cc
blob8322d3e3c3bec0fc57b7be0c91a7b75045c07ae4
1 /*
2 * Copyright 2011 Leiden University. All rights reserved.
3 * Copyright 2012-2013 Ecole Normale Superieure. All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 *
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above
13 * copyright notice, this list of conditions and the following
14 * disclaimer in the documentation and/or other materials provided
15 * with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY LEIDEN UNIVERSITY ''AS IS'' AND ANY
18 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL LEIDEN UNIVERSITY OR
21 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
22 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
23 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA,
24 * OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
27 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 * The views and conclusions contained in the software and documentation
30 * are those of the authors and should not be interpreted as
31 * representing official policies, either expressed or implied, of
32 * Leiden University.
33 */
35 #include <string.h>
36 #include <set>
37 #include <map>
38 #include <iostream>
39 #include <llvm/Support/raw_ostream.h>
40 #include <clang/AST/ASTContext.h>
41 #include <clang/AST/ASTDiagnostic.h>
42 #include <clang/AST/Expr.h>
43 #include <clang/AST/RecursiveASTVisitor.h>
45 #include <isl/id.h>
46 #include <isl/space.h>
47 #include <isl/aff.h>
48 #include <isl/set.h>
50 #include "clang.h"
51 #include "options.h"
52 #include "scan.h"
53 #include "scop.h"
54 #include "scop_plus.h"
56 #include "config.h"
58 using namespace std;
59 using namespace clang;
61 #if defined(DECLREFEXPR_CREATE_REQUIRES_BOOL)
62 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
64 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
65 SourceLocation(), var, false, var->getInnerLocStart(),
66 var->getType(), VK_LValue);
68 #elif defined(DECLREFEXPR_CREATE_REQUIRES_SOURCELOCATION)
69 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
71 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
72 SourceLocation(), var, var->getInnerLocStart(), var->getType(),
73 VK_LValue);
75 #else
76 static DeclRefExpr *create_DeclRefExpr(VarDecl *var)
78 return DeclRefExpr::Create(var->getASTContext(), var->getQualifierLoc(),
79 var, var->getInnerLocStart(), var->getType(), VK_LValue);
81 #endif
83 /* Check if the element type corresponding to the given array type
84 * has a const qualifier.
86 static bool const_base(QualType qt)
88 const Type *type = qt.getTypePtr();
90 if (type->isPointerType())
91 return const_base(type->getPointeeType());
92 if (type->isArrayType()) {
93 const ArrayType *atype;
94 type = type->getCanonicalTypeInternal().getTypePtr();
95 atype = cast<ArrayType>(type);
96 return const_base(atype->getElementType());
99 return qt.isConstQualified();
102 /* Mark "decl" as having an unknown value in "assigned_value".
104 * If no (known or unknown) value was assigned to "decl" before,
105 * then it may have been treated as a parameter before and may
106 * therefore appear in a value assigned to another variable.
107 * If so, this assignment needs to be turned into an unknown value too.
109 static void clear_assignment(map<ValueDecl *, isl_pw_aff *> &assigned_value,
110 ValueDecl *decl)
112 map<ValueDecl *, isl_pw_aff *>::iterator it;
114 it = assigned_value.find(decl);
116 assigned_value[decl] = NULL;
118 if (it == assigned_value.end())
119 return;
121 for (it = assigned_value.begin(); it != assigned_value.end(); ++it) {
122 isl_pw_aff *pa = it->second;
123 int nparam = isl_pw_aff_dim(pa, isl_dim_param);
125 for (int i = 0; i < nparam; ++i) {
126 isl_id *id;
128 if (!isl_pw_aff_has_dim_id(pa, isl_dim_param, i))
129 continue;
130 id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
131 if (isl_id_get_user(id) == decl)
132 it->second = NULL;
133 isl_id_free(id);
138 /* Look for any assignments to scalar variables in part of the parse
139 * tree and set assigned_value to NULL for each of them.
140 * Also reset assigned_value if the address of a scalar variable
141 * is being taken. As an exception, if the address is passed to a function
142 * that is declared to receive a const pointer, then assigned_value is
143 * not reset.
145 * This ensures that we won't use any previously stored value
146 * in the current subtree and its parents.
148 struct clear_assignments : RecursiveASTVisitor<clear_assignments> {
149 map<ValueDecl *, isl_pw_aff *> &assigned_value;
150 set<UnaryOperator *> skip;
152 clear_assignments(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
153 assigned_value(assigned_value) {}
155 /* Check for "address of" operators whose value is passed
156 * to a const pointer argument and add them to "skip", so that
157 * we can skip them in VisitUnaryOperator.
159 bool VisitCallExpr(CallExpr *expr) {
160 FunctionDecl *fd;
161 fd = expr->getDirectCallee();
162 if (!fd)
163 return true;
164 for (int i = 0; i < expr->getNumArgs(); ++i) {
165 Expr *arg = expr->getArg(i);
166 UnaryOperator *op;
167 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
168 ImplicitCastExpr *ice;
169 ice = cast<ImplicitCastExpr>(arg);
170 arg = ice->getSubExpr();
172 if (arg->getStmtClass() != Stmt::UnaryOperatorClass)
173 continue;
174 op = cast<UnaryOperator>(arg);
175 if (op->getOpcode() != UO_AddrOf)
176 continue;
177 if (const_base(fd->getParamDecl(i)->getType()))
178 skip.insert(op);
180 return true;
183 bool VisitUnaryOperator(UnaryOperator *expr) {
184 Expr *arg;
185 DeclRefExpr *ref;
186 ValueDecl *decl;
188 switch (expr->getOpcode()) {
189 case UO_AddrOf:
190 case UO_PostInc:
191 case UO_PostDec:
192 case UO_PreInc:
193 case UO_PreDec:
194 break;
195 default:
196 return true;
198 if (skip.find(expr) != skip.end())
199 return true;
201 arg = expr->getSubExpr();
202 if (arg->getStmtClass() != Stmt::DeclRefExprClass)
203 return true;
204 ref = cast<DeclRefExpr>(arg);
205 decl = ref->getDecl();
206 clear_assignment(assigned_value, decl);
207 return true;
210 bool VisitBinaryOperator(BinaryOperator *expr) {
211 Expr *lhs;
212 DeclRefExpr *ref;
213 ValueDecl *decl;
215 if (!expr->isAssignmentOp())
216 return true;
217 lhs = expr->getLHS();
218 if (lhs->getStmtClass() != Stmt::DeclRefExprClass)
219 return true;
220 ref = cast<DeclRefExpr>(lhs);
221 decl = ref->getDecl();
222 clear_assignment(assigned_value, decl);
223 return true;
227 /* Keep a copy of the currently assigned values.
229 * Any variable that is assigned a value inside the current scope
230 * is removed again when we leave the scope (either because it wasn't
231 * stored in the cache or because it has a different value in the cache).
233 struct assigned_value_cache {
234 map<ValueDecl *, isl_pw_aff *> &assigned_value;
235 map<ValueDecl *, isl_pw_aff *> cache;
237 assigned_value_cache(map<ValueDecl *, isl_pw_aff *> &assigned_value) :
238 assigned_value(assigned_value), cache(assigned_value) {}
239 ~assigned_value_cache() {
240 map<ValueDecl *, isl_pw_aff *>::iterator it = cache.begin();
241 for (it = assigned_value.begin(); it != assigned_value.end();
242 ++it) {
243 if (!it->second ||
244 (cache.find(it->first) != cache.end() &&
245 cache[it->first] != it->second))
246 cache[it->first] = NULL;
248 assigned_value = cache;
252 /* Insert an expression into the collection of expressions,
253 * provided it is not already in there.
254 * The isl_pw_affs are freed in the destructor.
256 void PetScan::insert_expression(__isl_take isl_pw_aff *expr)
258 std::set<isl_pw_aff *>::iterator it;
260 if (expressions.find(expr) == expressions.end())
261 expressions.insert(expr);
262 else
263 isl_pw_aff_free(expr);
266 PetScan::~PetScan()
268 std::set<isl_pw_aff *>::iterator it;
270 for (it = expressions.begin(); it != expressions.end(); ++it)
271 isl_pw_aff_free(*it);
273 isl_union_map_free(value_bounds);
276 /* Called if we found something we (currently) cannot handle.
277 * We'll provide more informative warnings later.
279 * We only actually complain if autodetect is false.
281 void PetScan::unsupported(Stmt *stmt, const char *msg)
283 if (options->autodetect)
284 return;
286 SourceLocation loc = stmt->getLocStart();
287 DiagnosticsEngine &diag = PP.getDiagnostics();
288 unsigned id = diag.getCustomDiagID(DiagnosticsEngine::Warning,
289 msg ? msg : "unsupported");
290 DiagnosticBuilder B = diag.Report(loc, id) << stmt->getSourceRange();
293 /* Extract an integer from "expr".
295 __isl_give isl_val *PetScan::extract_int(isl_ctx *ctx, IntegerLiteral *expr)
297 const Type *type = expr->getType().getTypePtr();
298 int is_signed = type->hasSignedIntegerRepresentation();
299 llvm::APInt val = expr->getValue();
300 int is_negative = is_signed && val.isNegative();
301 isl_val *v;
303 if (is_negative)
304 val = -val;
306 v = extract_unsigned(ctx, val);
308 if (is_negative)
309 v = isl_val_neg(v);
310 return v;
313 /* Extract an integer from "val", which assumed to be non-negative.
315 __isl_give isl_val *PetScan::extract_unsigned(isl_ctx *ctx,
316 const llvm::APInt &val)
318 unsigned n;
319 const uint64_t *data;
321 data = val.getRawData();
322 n = val.getNumWords();
323 return isl_val_int_from_chunks(ctx, n, sizeof(uint64_t), data);
326 /* Extract an integer from "expr".
327 * Return NULL if "expr" does not (obviously) represent an integer.
329 __isl_give isl_val *PetScan::extract_int(clang::ParenExpr *expr)
331 return extract_int(expr->getSubExpr());
334 /* Extract an integer from "expr".
335 * Return NULL if "expr" does not (obviously) represent an integer.
337 __isl_give isl_val *PetScan::extract_int(clang::Expr *expr)
339 if (expr->getStmtClass() == Stmt::IntegerLiteralClass)
340 return extract_int(ctx, cast<IntegerLiteral>(expr));
341 if (expr->getStmtClass() == Stmt::ParenExprClass)
342 return extract_int(cast<ParenExpr>(expr));
344 unsupported(expr);
345 return NULL;
348 /* Extract an affine expression from the IntegerLiteral "expr".
350 __isl_give isl_pw_aff *PetScan::extract_affine(IntegerLiteral *expr)
352 isl_space *dim = isl_space_params_alloc(ctx, 0);
353 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
354 isl_aff *aff = isl_aff_zero_on_domain(ls);
355 isl_set *dom = isl_set_universe(dim);
356 isl_val *v;
358 v = extract_int(expr);
359 aff = isl_aff_add_constant_val(aff, v);
361 return isl_pw_aff_alloc(dom, aff);
364 /* Extract an affine expression from the APInt "val", which is assumed
365 * to be non-negative.
367 __isl_give isl_pw_aff *PetScan::extract_affine(const llvm::APInt &val)
369 isl_space *dim = isl_space_params_alloc(ctx, 0);
370 isl_local_space *ls = isl_local_space_from_space(isl_space_copy(dim));
371 isl_aff *aff = isl_aff_zero_on_domain(ls);
372 isl_set *dom = isl_set_universe(dim);
373 isl_val *v;
375 v = extract_unsigned(ctx, val);
376 aff = isl_aff_add_constant_val(aff, v);
378 return isl_pw_aff_alloc(dom, aff);
381 __isl_give isl_pw_aff *PetScan::extract_affine(ImplicitCastExpr *expr)
383 return extract_affine(expr->getSubExpr());
386 static unsigned get_type_size(ValueDecl *decl)
388 return decl->getASTContext().getIntWidth(decl->getType());
391 /* Bound parameter "pos" of "set" to the possible values of "decl".
393 static __isl_give isl_set *set_parameter_bounds(__isl_take isl_set *set,
394 unsigned pos, ValueDecl *decl)
396 unsigned width;
397 isl_ctx *ctx;
398 isl_val *bound;
400 ctx = isl_set_get_ctx(set);
401 width = get_type_size(decl);
402 if (decl->getType()->isUnsignedIntegerType()) {
403 set = isl_set_lower_bound_si(set, isl_dim_param, pos, 0);
404 bound = isl_val_int_from_ui(ctx, width);
405 bound = isl_val_2exp(bound);
406 bound = isl_val_sub_ui(bound, 1);
407 set = isl_set_upper_bound_val(set, isl_dim_param, pos, bound);
408 } else {
409 bound = isl_val_int_from_ui(ctx, width - 1);
410 bound = isl_val_2exp(bound);
411 bound = isl_val_sub_ui(bound, 1);
412 set = isl_set_upper_bound_val(set, isl_dim_param, pos,
413 isl_val_copy(bound));
414 bound = isl_val_neg(bound);
415 bound = isl_val_sub_ui(bound, 1);
416 set = isl_set_lower_bound_val(set, isl_dim_param, pos, bound);
419 return set;
422 /* Extract an affine expression from the DeclRefExpr "expr".
424 * If the variable has been assigned a value, then we check whether
425 * we know what (affine) value was assigned.
426 * If so, we return this value. Otherwise we convert "expr"
427 * to an extra parameter (provided nesting_enabled is set).
429 * Otherwise, we simply return an expression that is equal
430 * to a parameter corresponding to the referenced variable.
432 __isl_give isl_pw_aff *PetScan::extract_affine(DeclRefExpr *expr)
434 ValueDecl *decl = expr->getDecl();
435 const Type *type = decl->getType().getTypePtr();
436 isl_id *id;
437 isl_space *dim;
438 isl_aff *aff;
439 isl_set *dom;
441 if (!type->isIntegerType()) {
442 unsupported(expr);
443 return NULL;
446 if (assigned_value.find(decl) != assigned_value.end()) {
447 if (assigned_value[decl])
448 return isl_pw_aff_copy(assigned_value[decl]);
449 else
450 return nested_access(expr);
453 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
454 dim = isl_space_params_alloc(ctx, 1);
456 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
458 dom = isl_set_universe(isl_space_copy(dim));
459 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
460 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
462 return isl_pw_aff_alloc(dom, aff);
465 /* Extract an affine expression from an integer division operation.
466 * In particular, if "expr" is lhs/rhs, then return
468 * lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs)
470 * The second argument (rhs) is required to be a (positive) integer constant.
472 __isl_give isl_pw_aff *PetScan::extract_affine_div(BinaryOperator *expr)
474 int is_cst;
475 isl_pw_aff *rhs, *lhs;
477 rhs = extract_affine(expr->getRHS());
478 is_cst = isl_pw_aff_is_cst(rhs);
479 if (is_cst < 0 || !is_cst) {
480 isl_pw_aff_free(rhs);
481 if (!is_cst)
482 unsupported(expr);
483 return NULL;
486 lhs = extract_affine(expr->getLHS());
488 return isl_pw_aff_tdiv_q(lhs, rhs);
491 /* Extract an affine expression from a modulo operation.
492 * In particular, if "expr" is lhs/rhs, then return
494 * lhs - rhs * (lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs))
496 * The second argument (rhs) is required to be a (positive) integer constant.
498 __isl_give isl_pw_aff *PetScan::extract_affine_mod(BinaryOperator *expr)
500 int is_cst;
501 isl_pw_aff *rhs, *lhs;
503 rhs = extract_affine(expr->getRHS());
504 is_cst = isl_pw_aff_is_cst(rhs);
505 if (is_cst < 0 || !is_cst) {
506 isl_pw_aff_free(rhs);
507 if (!is_cst)
508 unsupported(expr);
509 return NULL;
512 lhs = extract_affine(expr->getLHS());
514 return isl_pw_aff_tdiv_r(lhs, rhs);
517 /* Extract an affine expression from a multiplication operation.
518 * This is only allowed if at least one of the two arguments
519 * is a (piecewise) constant.
521 __isl_give isl_pw_aff *PetScan::extract_affine_mul(BinaryOperator *expr)
523 isl_pw_aff *lhs;
524 isl_pw_aff *rhs;
526 lhs = extract_affine(expr->getLHS());
527 rhs = extract_affine(expr->getRHS());
529 if (!isl_pw_aff_is_cst(lhs) && !isl_pw_aff_is_cst(rhs)) {
530 isl_pw_aff_free(lhs);
531 isl_pw_aff_free(rhs);
532 unsupported(expr);
533 return NULL;
536 return isl_pw_aff_mul(lhs, rhs);
539 /* Extract an affine expression from an addition or subtraction operation.
541 __isl_give isl_pw_aff *PetScan::extract_affine_add(BinaryOperator *expr)
543 isl_pw_aff *lhs;
544 isl_pw_aff *rhs;
546 lhs = extract_affine(expr->getLHS());
547 rhs = extract_affine(expr->getRHS());
549 switch (expr->getOpcode()) {
550 case BO_Add:
551 return isl_pw_aff_add(lhs, rhs);
552 case BO_Sub:
553 return isl_pw_aff_sub(lhs, rhs);
554 default:
555 isl_pw_aff_free(lhs);
556 isl_pw_aff_free(rhs);
557 return NULL;
562 /* Compute
564 * pwaff mod 2^width
566 static __isl_give isl_pw_aff *wrap(__isl_take isl_pw_aff *pwaff,
567 unsigned width)
569 isl_ctx *ctx;
570 isl_val *mod;
572 ctx = isl_pw_aff_get_ctx(pwaff);
573 mod = isl_val_int_from_ui(ctx, width);
574 mod = isl_val_2exp(mod);
576 pwaff = isl_pw_aff_mod_val(pwaff, mod);
578 return pwaff;
581 /* Limit the domain of "pwaff" to those elements where the function
582 * value satisfies
584 * 2^{width-1} <= pwaff < 2^{width-1}
586 static __isl_give isl_pw_aff *avoid_overflow(__isl_take isl_pw_aff *pwaff,
587 unsigned width)
589 isl_ctx *ctx;
590 isl_val *v;
591 isl_space *space = isl_pw_aff_get_domain_space(pwaff);
592 isl_local_space *ls = isl_local_space_from_space(space);
593 isl_aff *bound;
594 isl_set *dom;
595 isl_pw_aff *b;
597 ctx = isl_pw_aff_get_ctx(pwaff);
598 v = isl_val_int_from_ui(ctx, width - 1);
599 v = isl_val_2exp(v);
601 bound = isl_aff_zero_on_domain(ls);
602 bound = isl_aff_add_constant_val(bound, v);
603 b = isl_pw_aff_from_aff(bound);
605 dom = isl_pw_aff_lt_set(isl_pw_aff_copy(pwaff), isl_pw_aff_copy(b));
606 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
608 b = isl_pw_aff_neg(b);
609 dom = isl_pw_aff_ge_set(isl_pw_aff_copy(pwaff), b);
610 pwaff = isl_pw_aff_intersect_domain(pwaff, dom);
612 return pwaff;
615 /* Handle potential overflows on signed computations.
617 * If options->signed_overflow is set to PET_OVERFLOW_AVOID,
618 * the we adjust the domain of "pa" to avoid overflows.
620 __isl_give isl_pw_aff *PetScan::signed_overflow(__isl_take isl_pw_aff *pa,
621 unsigned width)
623 if (options->signed_overflow == PET_OVERFLOW_AVOID)
624 pa = avoid_overflow(pa, width);
626 return pa;
629 /* Return the piecewise affine expression "set ? 1 : 0" defined on "dom".
631 static __isl_give isl_pw_aff *indicator_function(__isl_take isl_set *set,
632 __isl_take isl_set *dom)
634 isl_pw_aff *pa;
635 pa = isl_set_indicator_function(set);
636 pa = isl_pw_aff_intersect_domain(pa, dom);
637 return pa;
640 /* Extract an affine expression from some binary operations.
641 * If the result of the expression is unsigned, then we wrap it
642 * based on the size of the type. Otherwise, we ensure that
643 * no overflow occurs.
645 __isl_give isl_pw_aff *PetScan::extract_affine(BinaryOperator *expr)
647 isl_pw_aff *res;
648 unsigned width;
650 switch (expr->getOpcode()) {
651 case BO_Add:
652 case BO_Sub:
653 res = extract_affine_add(expr);
654 break;
655 case BO_Div:
656 res = extract_affine_div(expr);
657 break;
658 case BO_Rem:
659 res = extract_affine_mod(expr);
660 break;
661 case BO_Mul:
662 res = extract_affine_mul(expr);
663 break;
664 case BO_LT:
665 case BO_LE:
666 case BO_GT:
667 case BO_GE:
668 case BO_EQ:
669 case BO_NE:
670 case BO_LAnd:
671 case BO_LOr:
672 return extract_condition(expr);
673 default:
674 unsupported(expr);
675 return NULL;
678 width = ast_context.getIntWidth(expr->getType());
679 if (expr->getType()->isUnsignedIntegerType())
680 res = wrap(res, width);
681 else
682 res = signed_overflow(res, width);
684 return res;
687 /* Extract an affine expression from a negation operation.
689 __isl_give isl_pw_aff *PetScan::extract_affine(UnaryOperator *expr)
691 if (expr->getOpcode() == UO_Minus)
692 return isl_pw_aff_neg(extract_affine(expr->getSubExpr()));
693 if (expr->getOpcode() == UO_LNot)
694 return extract_condition(expr);
696 unsupported(expr);
697 return NULL;
700 __isl_give isl_pw_aff *PetScan::extract_affine(ParenExpr *expr)
702 return extract_affine(expr->getSubExpr());
705 /* Extract an affine expression from some special function calls.
706 * In particular, we handle "min", "max", "ceild" and "floord".
707 * In case of the latter two, the second argument needs to be
708 * a (positive) integer constant.
710 __isl_give isl_pw_aff *PetScan::extract_affine(CallExpr *expr)
712 FunctionDecl *fd;
713 string name;
714 isl_pw_aff *aff1, *aff2;
716 fd = expr->getDirectCallee();
717 if (!fd) {
718 unsupported(expr);
719 return NULL;
722 name = fd->getDeclName().getAsString();
723 if (!(expr->getNumArgs() == 2 && name == "min") &&
724 !(expr->getNumArgs() == 2 && name == "max") &&
725 !(expr->getNumArgs() == 2 && name == "floord") &&
726 !(expr->getNumArgs() == 2 && name == "ceild")) {
727 unsupported(expr);
728 return NULL;
731 if (name == "min" || name == "max") {
732 aff1 = extract_affine(expr->getArg(0));
733 aff2 = extract_affine(expr->getArg(1));
735 if (name == "min")
736 aff1 = isl_pw_aff_min(aff1, aff2);
737 else
738 aff1 = isl_pw_aff_max(aff1, aff2);
739 } else if (name == "floord" || name == "ceild") {
740 isl_val *v;
741 Expr *arg2 = expr->getArg(1);
743 if (arg2->getStmtClass() != Stmt::IntegerLiteralClass) {
744 unsupported(expr);
745 return NULL;
747 aff1 = extract_affine(expr->getArg(0));
748 v = extract_int(cast<IntegerLiteral>(arg2));
749 aff1 = isl_pw_aff_scale_down_val(aff1, v);
750 if (name == "floord")
751 aff1 = isl_pw_aff_floor(aff1);
752 else
753 aff1 = isl_pw_aff_ceil(aff1);
754 } else {
755 unsupported(expr);
756 return NULL;
759 return aff1;
762 /* This method is called when we come across an access that is
763 * nested in what is supposed to be an affine expression.
764 * If nesting is allowed, we return a new parameter that corresponds
765 * to this nested access. Otherwise, we simply complain.
767 * Note that we currently don't allow nested accesses themselves
768 * to contain any nested accesses, so we check if we can extract
769 * the access without any nesting and complain if we can't.
771 * The new parameter is resolved in resolve_nested.
773 isl_pw_aff *PetScan::nested_access(Expr *expr)
775 isl_id *id;
776 isl_space *dim;
777 isl_aff *aff;
778 isl_set *dom;
779 isl_multi_pw_aff *index;
781 if (!nesting_enabled) {
782 unsupported(expr);
783 return NULL;
786 allow_nested = false;
787 index = extract_index(expr);
788 allow_nested = true;
789 if (!index) {
790 unsupported(expr);
791 return NULL;
793 isl_multi_pw_aff_free(index);
795 id = isl_id_alloc(ctx, NULL, expr);
796 dim = isl_space_params_alloc(ctx, 1);
798 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
800 dom = isl_set_universe(isl_space_copy(dim));
801 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
802 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
804 return isl_pw_aff_alloc(dom, aff);
807 /* Affine expressions are not supposed to contain array accesses,
808 * but if nesting is allowed, we return a parameter corresponding
809 * to the array access.
811 __isl_give isl_pw_aff *PetScan::extract_affine(ArraySubscriptExpr *expr)
813 return nested_access(expr);
816 /* Affine expressions are not supposed to contain member accesses,
817 * but if nesting is allowed, we return a parameter corresponding
818 * to the member access.
820 __isl_give isl_pw_aff *PetScan::extract_affine(MemberExpr *expr)
822 return nested_access(expr);
825 /* Extract an affine expression from a conditional operation.
827 __isl_give isl_pw_aff *PetScan::extract_affine(ConditionalOperator *expr)
829 isl_pw_aff *cond, *lhs, *rhs;
831 cond = extract_condition(expr->getCond());
832 lhs = extract_affine(expr->getTrueExpr());
833 rhs = extract_affine(expr->getFalseExpr());
835 return isl_pw_aff_cond(cond, lhs, rhs);
838 /* Extract an affine expression, if possible, from "expr".
839 * Otherwise return NULL.
841 __isl_give isl_pw_aff *PetScan::extract_affine(Expr *expr)
843 switch (expr->getStmtClass()) {
844 case Stmt::ImplicitCastExprClass:
845 return extract_affine(cast<ImplicitCastExpr>(expr));
846 case Stmt::IntegerLiteralClass:
847 return extract_affine(cast<IntegerLiteral>(expr));
848 case Stmt::DeclRefExprClass:
849 return extract_affine(cast<DeclRefExpr>(expr));
850 case Stmt::BinaryOperatorClass:
851 return extract_affine(cast<BinaryOperator>(expr));
852 case Stmt::UnaryOperatorClass:
853 return extract_affine(cast<UnaryOperator>(expr));
854 case Stmt::ParenExprClass:
855 return extract_affine(cast<ParenExpr>(expr));
856 case Stmt::CallExprClass:
857 return extract_affine(cast<CallExpr>(expr));
858 case Stmt::ArraySubscriptExprClass:
859 return extract_affine(cast<ArraySubscriptExpr>(expr));
860 case Stmt::MemberExprClass:
861 return extract_affine(cast<MemberExpr>(expr));
862 case Stmt::ConditionalOperatorClass:
863 return extract_affine(cast<ConditionalOperator>(expr));
864 default:
865 unsupported(expr);
867 return NULL;
870 __isl_give isl_multi_pw_aff *PetScan::extract_index(ImplicitCastExpr *expr)
872 return extract_index(expr->getSubExpr());
875 /* Return the depth of an array of the given type.
877 static int array_depth(const Type *type)
879 if (type->isPointerType())
880 return 1 + array_depth(type->getPointeeType().getTypePtr());
881 if (type->isArrayType()) {
882 const ArrayType *atype;
883 type = type->getCanonicalTypeInternal().getTypePtr();
884 atype = cast<ArrayType>(type);
885 return 1 + array_depth(atype->getElementType().getTypePtr());
887 return 0;
890 /* Return the depth of the array accessed by the index expression "index".
891 * If "index" is an affine expression, i.e., if it does not access
892 * any array, then return 1.
893 * If "index" represent a member access, i.e., if its range is a wrapped
894 * relation, then return the sum of the depth of the array of structures
895 * and that of the member inside the structure.
897 static int extract_depth(__isl_keep isl_multi_pw_aff *index)
899 isl_id *id;
900 ValueDecl *decl;
902 if (!index)
903 return -1;
905 if (isl_multi_pw_aff_range_is_wrapping(index)) {
906 int domain_depth, range_depth;
907 isl_multi_pw_aff *domain, *range;
909 domain = isl_multi_pw_aff_copy(index);
910 domain = isl_multi_pw_aff_range_factor_domain(domain);
911 domain_depth = extract_depth(domain);
912 isl_multi_pw_aff_free(domain);
913 range = isl_multi_pw_aff_copy(index);
914 range = isl_multi_pw_aff_range_factor_range(range);
915 range_depth = extract_depth(range);
916 isl_multi_pw_aff_free(range);
918 return domain_depth + range_depth;
921 if (!isl_multi_pw_aff_has_tuple_id(index, isl_dim_out))
922 return 1;
924 id = isl_multi_pw_aff_get_tuple_id(index, isl_dim_out);
925 if (!id)
926 return -1;
927 decl = (ValueDecl *) isl_id_get_user(id);
928 isl_id_free(id);
930 return array_depth(decl->getType().getTypePtr());
933 /* Extract an index expression from a reference to a variable.
934 * If the variable has name "A", then the returned index expression
935 * is of the form
937 * { [] -> A[] }
939 __isl_give isl_multi_pw_aff *PetScan::extract_index(DeclRefExpr *expr)
941 return extract_index(expr->getDecl());
944 /* Extract an index expression from a variable.
945 * If the variable has name "A", then the returned index expression
946 * is of the form
948 * { [] -> A[] }
950 __isl_give isl_multi_pw_aff *PetScan::extract_index(ValueDecl *decl)
952 isl_id *id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
953 isl_space *space = isl_space_alloc(ctx, 0, 0, 0);
955 space = isl_space_set_tuple_id(space, isl_dim_out, id);
957 return isl_multi_pw_aff_zero(space);
960 /* Extract an index expression from an integer contant.
961 * If the value of the constant is "v", then the returned access relation
962 * is
964 * { [] -> [v] }
966 __isl_give isl_multi_pw_aff *PetScan::extract_index(IntegerLiteral *expr)
968 isl_multi_pw_aff *mpa;
970 mpa = isl_multi_pw_aff_from_pw_aff(extract_affine(expr));
971 mpa = isl_multi_pw_aff_from_range(mpa);
972 return mpa;
975 /* Try and extract an index expression from the given Expr.
976 * Return NULL if it doesn't work out.
978 __isl_give isl_multi_pw_aff *PetScan::extract_index(Expr *expr)
980 switch (expr->getStmtClass()) {
981 case Stmt::ImplicitCastExprClass:
982 return extract_index(cast<ImplicitCastExpr>(expr));
983 case Stmt::DeclRefExprClass:
984 return extract_index(cast<DeclRefExpr>(expr));
985 case Stmt::ArraySubscriptExprClass:
986 return extract_index(cast<ArraySubscriptExpr>(expr));
987 case Stmt::IntegerLiteralClass:
988 return extract_index(cast<IntegerLiteral>(expr));
989 case Stmt::MemberExprClass:
990 return extract_index(cast<MemberExpr>(expr));
991 default:
992 unsupported(expr);
994 return NULL;
997 /* Given a partial index expression "base" and an extra index "index",
998 * append the extra index to "base" and return the result.
999 * Additionally, add the constraints that the extra index is non-negative.
1000 * If "index" represent a member access, i.e., if its range is a wrapped
1001 * relation, then we recursively extend the range of this nested relation.
1003 static __isl_give isl_multi_pw_aff *subscript(__isl_take isl_multi_pw_aff *base,
1004 __isl_take isl_pw_aff *index)
1006 isl_id *id;
1007 isl_set *domain;
1008 isl_multi_pw_aff *access;
1009 int member_access;
1011 member_access = isl_multi_pw_aff_range_is_wrapping(base);
1012 if (member_access < 0)
1013 goto error;
1014 if (member_access) {
1015 isl_multi_pw_aff *domain, *range;
1016 isl_id *id;
1018 id = isl_multi_pw_aff_get_tuple_id(base, isl_dim_out);
1019 domain = isl_multi_pw_aff_copy(base);
1020 domain = isl_multi_pw_aff_range_factor_domain(domain);
1021 range = isl_multi_pw_aff_range_factor_range(base);
1022 range = subscript(range, index);
1023 access = isl_multi_pw_aff_range_product(domain, range);
1024 access = isl_multi_pw_aff_set_tuple_id(access, isl_dim_out, id);
1025 return access;
1028 id = isl_multi_pw_aff_get_tuple_id(base, isl_dim_set);
1029 index = isl_pw_aff_from_range(index);
1030 domain = isl_pw_aff_nonneg_set(isl_pw_aff_copy(index));
1031 index = isl_pw_aff_intersect_domain(index, domain);
1032 access = isl_multi_pw_aff_from_pw_aff(index);
1033 access = isl_multi_pw_aff_flat_range_product(base, access);
1034 access = isl_multi_pw_aff_set_tuple_id(access, isl_dim_set, id);
1036 return access;
1037 error:
1038 isl_multi_pw_aff_free(base);
1039 isl_pw_aff_free(index);
1040 return NULL;
1043 /* Extract an index expression from the given array subscript expression.
1044 * If nesting is allowed in general, then we turn it on while
1045 * examining the index expression.
1047 * We first extract an index expression from the base.
1048 * This will result in an index expression with a range that corresponds
1049 * to the earlier indices.
1050 * We then extract the current index, restrict its domain
1051 * to those values that result in a non-negative index and
1052 * append the index to the base index expression.
1054 __isl_give isl_multi_pw_aff *PetScan::extract_index(ArraySubscriptExpr *expr)
1056 Expr *base = expr->getBase();
1057 Expr *idx = expr->getIdx();
1058 isl_pw_aff *index;
1059 isl_multi_pw_aff *base_access;
1060 isl_multi_pw_aff *access;
1061 bool save_nesting = nesting_enabled;
1063 nesting_enabled = allow_nested;
1065 base_access = extract_index(base);
1066 index = extract_affine(idx);
1068 nesting_enabled = save_nesting;
1070 access = subscript(base_access, index);
1072 return access;
1075 /* Construct a name for a member access by concatenating the name
1076 * of the array of structures and the member, separated by an underscore.
1078 * The caller is responsible for freeing the result.
1080 static char *member_access_name(isl_ctx *ctx, const char *base,
1081 const char *field)
1083 int len;
1084 char *name;
1086 len = strlen(base) + 1 + strlen(field);
1087 name = isl_alloc_array(ctx, char, len + 1);
1088 if (!name)
1089 return NULL;
1090 snprintf(name, len + 1, "%s_%s", base, field);
1092 return name;
1095 /* Given an index expression "base" for an element of an array of structures
1096 * and an expression "field" for the field member being accessed, construct
1097 * an index expression for an access to that member of the given structure.
1098 * In particular, take the range product of "base" and "field" and
1099 * attach a name to the result.
1101 static __isl_give isl_multi_pw_aff *member(__isl_take isl_multi_pw_aff *base,
1102 __isl_take isl_multi_pw_aff *field)
1104 isl_ctx *ctx;
1105 isl_multi_pw_aff *access;
1106 const char *base_name, *field_name;
1107 char *name;
1109 ctx = isl_multi_pw_aff_get_ctx(base);
1111 base_name = isl_multi_pw_aff_get_tuple_name(base, isl_dim_out);
1112 field_name = isl_multi_pw_aff_get_tuple_name(field, isl_dim_out);
1113 name = member_access_name(ctx, base_name, field_name);
1115 access = isl_multi_pw_aff_range_product(base, field);
1117 access = isl_multi_pw_aff_set_tuple_name(access, isl_dim_out, name);
1118 free(name);
1120 return access;
1123 /* Extract an index expression from a member expression.
1125 * If the base access (to the structure containing the member)
1126 * is of the form
1128 * [] -> A[..]
1130 * and the member is called "f", then the member access is of
1131 * the form
1133 * [] -> A_f[A[..] -> f[]]
1135 * If the member access is to an anonymous struct, then simply return
1137 * [] -> A[..]
1139 * If the member access in the source code is of the form
1141 * A->f
1143 * then it is treated as
1145 * A[0].f
1147 __isl_give isl_multi_pw_aff *PetScan::extract_index(MemberExpr *expr)
1149 Expr *base = expr->getBase();
1150 FieldDecl *field = cast<FieldDecl>(expr->getMemberDecl());
1151 isl_multi_pw_aff *base_access, *field_access;
1152 isl_id *id;
1153 isl_space *space;
1155 base_access = extract_index(base);
1157 if (expr->isArrow()) {
1158 isl_space *space = isl_space_params_alloc(ctx, 0);
1159 isl_local_space *ls = isl_local_space_from_space(space);
1160 isl_aff *aff = isl_aff_zero_on_domain(ls);
1161 isl_pw_aff *index = isl_pw_aff_from_aff(aff);
1162 base_access = subscript(base_access, index);
1165 if (field->isAnonymousStructOrUnion())
1166 return base_access;
1168 id = isl_id_alloc(ctx, field->getName().str().c_str(), field);
1169 space = isl_multi_pw_aff_get_domain_space(base_access);
1170 space = isl_space_from_domain(space);
1171 space = isl_space_set_tuple_id(space, isl_dim_out, id);
1172 field_access = isl_multi_pw_aff_zero(space);
1174 return member(base_access, field_access);
1177 /* Check if "expr" calls function "minmax" with two arguments and if so
1178 * make lhs and rhs refer to these two arguments.
1180 static bool is_minmax(Expr *expr, const char *minmax, Expr *&lhs, Expr *&rhs)
1182 CallExpr *call;
1183 FunctionDecl *fd;
1184 string name;
1186 if (expr->getStmtClass() != Stmt::CallExprClass)
1187 return false;
1189 call = cast<CallExpr>(expr);
1190 fd = call->getDirectCallee();
1191 if (!fd)
1192 return false;
1194 if (call->getNumArgs() != 2)
1195 return false;
1197 name = fd->getDeclName().getAsString();
1198 if (name != minmax)
1199 return false;
1201 lhs = call->getArg(0);
1202 rhs = call->getArg(1);
1204 return true;
1207 /* Check if "expr" is of the form min(lhs, rhs) and if so make
1208 * lhs and rhs refer to the two arguments.
1210 static bool is_min(Expr *expr, Expr *&lhs, Expr *&rhs)
1212 return is_minmax(expr, "min", lhs, rhs);
1215 /* Check if "expr" is of the form max(lhs, rhs) and if so make
1216 * lhs and rhs refer to the two arguments.
1218 static bool is_max(Expr *expr, Expr *&lhs, Expr *&rhs)
1220 return is_minmax(expr, "max", lhs, rhs);
1223 /* Return "lhs && rhs", defined on the shared definition domain.
1225 static __isl_give isl_pw_aff *pw_aff_and(__isl_take isl_pw_aff *lhs,
1226 __isl_take isl_pw_aff *rhs)
1228 isl_set *cond;
1229 isl_set *dom;
1231 dom = isl_set_intersect(isl_pw_aff_domain(isl_pw_aff_copy(lhs)),
1232 isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1233 cond = isl_set_intersect(isl_pw_aff_non_zero_set(lhs),
1234 isl_pw_aff_non_zero_set(rhs));
1235 return indicator_function(cond, dom);
1238 /* Return "lhs && rhs", with shortcut semantics.
1239 * That is, if lhs is false, then the result is defined even if rhs is not.
1240 * In practice, we compute lhs ? rhs : lhs.
1242 static __isl_give isl_pw_aff *pw_aff_and_then(__isl_take isl_pw_aff *lhs,
1243 __isl_take isl_pw_aff *rhs)
1245 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), rhs, lhs);
1248 /* Return "lhs || rhs", with shortcut semantics.
1249 * That is, if lhs is true, then the result is defined even if rhs is not.
1250 * In practice, we compute lhs ? lhs : rhs.
1252 static __isl_give isl_pw_aff *pw_aff_or_else(__isl_take isl_pw_aff *lhs,
1253 __isl_take isl_pw_aff *rhs)
1255 return isl_pw_aff_cond(isl_pw_aff_copy(lhs), lhs, rhs);
1258 /* Extract an affine expressions representing the comparison "LHS op RHS"
1259 * "comp" is the original statement that "LHS op RHS" is derived from
1260 * and is used for diagnostics.
1262 * If the comparison is of the form
1264 * a <= min(b,c)
1266 * then the expression is constructed as the conjunction of
1267 * the comparisons
1269 * a <= b and a <= c
1271 * A similar optimization is performed for max(a,b) <= c.
1272 * We do this because that will lead to simpler representations
1273 * of the expression.
1274 * If isl is ever enhanced to explicitly deal with min and max expressions,
1275 * this optimization can be removed.
1277 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperatorKind op,
1278 Expr *LHS, Expr *RHS, Stmt *comp)
1280 isl_pw_aff *lhs;
1281 isl_pw_aff *rhs;
1282 isl_pw_aff *res;
1283 isl_set *cond;
1284 isl_set *dom;
1286 if (op == BO_GT)
1287 return extract_comparison(BO_LT, RHS, LHS, comp);
1288 if (op == BO_GE)
1289 return extract_comparison(BO_LE, RHS, LHS, comp);
1291 if (op == BO_LT || op == BO_LE) {
1292 Expr *expr1, *expr2;
1293 if (is_min(RHS, expr1, expr2)) {
1294 lhs = extract_comparison(op, LHS, expr1, comp);
1295 rhs = extract_comparison(op, LHS, expr2, comp);
1296 return pw_aff_and(lhs, rhs);
1298 if (is_max(LHS, expr1, expr2)) {
1299 lhs = extract_comparison(op, expr1, RHS, comp);
1300 rhs = extract_comparison(op, expr2, RHS, comp);
1301 return pw_aff_and(lhs, rhs);
1305 lhs = extract_affine(LHS);
1306 rhs = extract_affine(RHS);
1308 dom = isl_pw_aff_domain(isl_pw_aff_copy(lhs));
1309 dom = isl_set_intersect(dom, isl_pw_aff_domain(isl_pw_aff_copy(rhs)));
1311 switch (op) {
1312 case BO_LT:
1313 cond = isl_pw_aff_lt_set(lhs, rhs);
1314 break;
1315 case BO_LE:
1316 cond = isl_pw_aff_le_set(lhs, rhs);
1317 break;
1318 case BO_EQ:
1319 cond = isl_pw_aff_eq_set(lhs, rhs);
1320 break;
1321 case BO_NE:
1322 cond = isl_pw_aff_ne_set(lhs, rhs);
1323 break;
1324 default:
1325 isl_pw_aff_free(lhs);
1326 isl_pw_aff_free(rhs);
1327 isl_set_free(dom);
1328 unsupported(comp);
1329 return NULL;
1332 cond = isl_set_coalesce(cond);
1333 res = indicator_function(cond, dom);
1335 return res;
1338 __isl_give isl_pw_aff *PetScan::extract_comparison(BinaryOperator *comp)
1340 return extract_comparison(comp->getOpcode(), comp->getLHS(),
1341 comp->getRHS(), comp);
1344 /* Extract an affine expression representing the negation (logical not)
1345 * of a subexpression.
1347 __isl_give isl_pw_aff *PetScan::extract_boolean(UnaryOperator *op)
1349 isl_set *set_cond, *dom;
1350 isl_pw_aff *cond, *res;
1352 cond = extract_condition(op->getSubExpr());
1354 dom = isl_pw_aff_domain(isl_pw_aff_copy(cond));
1356 set_cond = isl_pw_aff_zero_set(cond);
1358 res = indicator_function(set_cond, dom);
1360 return res;
1363 /* Extract an affine expression representing the disjunction (logical or)
1364 * or conjunction (logical and) of two subexpressions.
1366 __isl_give isl_pw_aff *PetScan::extract_boolean(BinaryOperator *comp)
1368 isl_pw_aff *lhs, *rhs;
1370 lhs = extract_condition(comp->getLHS());
1371 rhs = extract_condition(comp->getRHS());
1373 switch (comp->getOpcode()) {
1374 case BO_LAnd:
1375 return pw_aff_and_then(lhs, rhs);
1376 case BO_LOr:
1377 return pw_aff_or_else(lhs, rhs);
1378 default:
1379 isl_pw_aff_free(lhs);
1380 isl_pw_aff_free(rhs);
1383 unsupported(comp);
1384 return NULL;
1387 __isl_give isl_pw_aff *PetScan::extract_condition(UnaryOperator *expr)
1389 switch (expr->getOpcode()) {
1390 case UO_LNot:
1391 return extract_boolean(expr);
1392 default:
1393 unsupported(expr);
1394 return NULL;
1398 /* Extract the affine expression "expr != 0 ? 1 : 0".
1400 __isl_give isl_pw_aff *PetScan::extract_implicit_condition(Expr *expr)
1402 isl_pw_aff *res;
1403 isl_set *set, *dom;
1405 res = extract_affine(expr);
1407 dom = isl_pw_aff_domain(isl_pw_aff_copy(res));
1408 set = isl_pw_aff_non_zero_set(res);
1410 res = indicator_function(set, dom);
1412 return res;
1415 /* Extract an affine expression from a boolean expression.
1416 * In particular, return the expression "expr ? 1 : 0".
1418 * If the expression doesn't look like a condition, we assume it
1419 * is an affine expression and return the condition "expr != 0 ? 1 : 0".
1421 __isl_give isl_pw_aff *PetScan::extract_condition(Expr *expr)
1423 BinaryOperator *comp;
1425 if (!expr) {
1426 isl_set *u = isl_set_universe(isl_space_params_alloc(ctx, 0));
1427 return indicator_function(u, isl_set_copy(u));
1430 if (expr->getStmtClass() == Stmt::ParenExprClass)
1431 return extract_condition(cast<ParenExpr>(expr)->getSubExpr());
1433 if (expr->getStmtClass() == Stmt::UnaryOperatorClass)
1434 return extract_condition(cast<UnaryOperator>(expr));
1436 if (expr->getStmtClass() != Stmt::BinaryOperatorClass)
1437 return extract_implicit_condition(expr);
1439 comp = cast<BinaryOperator>(expr);
1440 switch (comp->getOpcode()) {
1441 case BO_LT:
1442 case BO_LE:
1443 case BO_GT:
1444 case BO_GE:
1445 case BO_EQ:
1446 case BO_NE:
1447 return extract_comparison(comp);
1448 case BO_LAnd:
1449 case BO_LOr:
1450 return extract_boolean(comp);
1451 default:
1452 return extract_implicit_condition(expr);
1456 static enum pet_op_type UnaryOperatorKind2pet_op_type(UnaryOperatorKind kind)
1458 switch (kind) {
1459 case UO_Minus:
1460 return pet_op_minus;
1461 case UO_PostInc:
1462 return pet_op_post_inc;
1463 case UO_PostDec:
1464 return pet_op_post_dec;
1465 case UO_PreInc:
1466 return pet_op_pre_inc;
1467 case UO_PreDec:
1468 return pet_op_pre_dec;
1469 default:
1470 return pet_op_last;
1474 static enum pet_op_type BinaryOperatorKind2pet_op_type(BinaryOperatorKind kind)
1476 switch (kind) {
1477 case BO_AddAssign:
1478 return pet_op_add_assign;
1479 case BO_SubAssign:
1480 return pet_op_sub_assign;
1481 case BO_MulAssign:
1482 return pet_op_mul_assign;
1483 case BO_DivAssign:
1484 return pet_op_div_assign;
1485 case BO_Assign:
1486 return pet_op_assign;
1487 case BO_Add:
1488 return pet_op_add;
1489 case BO_Sub:
1490 return pet_op_sub;
1491 case BO_Mul:
1492 return pet_op_mul;
1493 case BO_Div:
1494 return pet_op_div;
1495 case BO_Rem:
1496 return pet_op_mod;
1497 case BO_EQ:
1498 return pet_op_eq;
1499 case BO_LE:
1500 return pet_op_le;
1501 case BO_LT:
1502 return pet_op_lt;
1503 case BO_GT:
1504 return pet_op_gt;
1505 default:
1506 return pet_op_last;
1510 /* Construct a pet_expr representing a unary operator expression.
1512 struct pet_expr *PetScan::extract_expr(UnaryOperator *expr)
1514 struct pet_expr *arg;
1515 enum pet_op_type op;
1517 op = UnaryOperatorKind2pet_op_type(expr->getOpcode());
1518 if (op == pet_op_last) {
1519 unsupported(expr);
1520 return NULL;
1523 arg = extract_expr(expr->getSubExpr());
1525 if (expr->isIncrementDecrementOp() &&
1526 arg && arg->type == pet_expr_access) {
1527 mark_write(arg);
1528 arg->acc.read = 1;
1531 return pet_expr_new_unary(ctx, op, arg);
1534 /* Mark the given access pet_expr as a write.
1535 * If a scalar is being accessed, then mark its value
1536 * as unknown in assigned_value.
1538 void PetScan::mark_write(struct pet_expr *access)
1540 isl_id *id;
1541 ValueDecl *decl;
1543 if (!access)
1544 return;
1546 access->acc.write = 1;
1547 access->acc.read = 0;
1549 if (!pet_expr_is_scalar_access(access))
1550 return;
1552 id = pet_expr_access_get_id(access);
1553 decl = (ValueDecl *) isl_id_get_user(id);
1554 clear_assignment(assigned_value, decl);
1555 isl_id_free(id);
1558 /* Assign "rhs" to "lhs".
1560 * In particular, if "lhs" is a scalar variable, then mark
1561 * the variable as having been assigned. If, furthermore, "rhs"
1562 * is an affine expression, then keep track of this value in assigned_value
1563 * so that we can plug it in when we later come across the same variable.
1565 void PetScan::assign(struct pet_expr *lhs, Expr *rhs)
1567 isl_id *id;
1568 ValueDecl *decl;
1569 isl_pw_aff *pa;
1571 if (!lhs)
1572 return;
1573 if (!pet_expr_is_scalar_access(lhs))
1574 return;
1576 id = pet_expr_access_get_id(lhs);
1577 decl = (ValueDecl *) isl_id_get_user(id);
1578 isl_id_free(id);
1580 pa = try_extract_affine(rhs);
1581 clear_assignment(assigned_value, decl);
1582 if (!pa)
1583 return;
1584 assigned_value[decl] = pa;
1585 insert_expression(pa);
1588 /* Construct a pet_expr representing a binary operator expression.
1590 * If the top level operator is an assignment and the LHS is an access,
1591 * then we mark that access as a write. If the operator is a compound
1592 * assignment, the access is marked as both a read and a write.
1594 * If "expr" assigns something to a scalar variable, then we mark
1595 * the variable as having been assigned. If, furthermore, the expression
1596 * is affine, then keep track of this value in assigned_value
1597 * so that we can plug it in when we later come across the same variable.
1599 struct pet_expr *PetScan::extract_expr(BinaryOperator *expr)
1601 struct pet_expr *lhs, *rhs;
1602 enum pet_op_type op;
1604 op = BinaryOperatorKind2pet_op_type(expr->getOpcode());
1605 if (op == pet_op_last) {
1606 unsupported(expr);
1607 return NULL;
1610 lhs = extract_expr(expr->getLHS());
1611 rhs = extract_expr(expr->getRHS());
1613 if (expr->isAssignmentOp() && lhs && lhs->type == pet_expr_access) {
1614 mark_write(lhs);
1615 if (expr->isCompoundAssignmentOp())
1616 lhs->acc.read = 1;
1619 if (expr->getOpcode() == BO_Assign)
1620 assign(lhs, expr->getRHS());
1622 return pet_expr_new_binary(ctx, op, lhs, rhs);
1625 /* Construct a pet_scop with a single statement killing the entire
1626 * array "array".
1628 struct pet_scop *PetScan::kill(Stmt *stmt, struct pet_array *array)
1630 isl_id *id;
1631 isl_space *space;
1632 isl_multi_pw_aff *index;
1633 isl_map *access;
1634 struct pet_expr *expr;
1636 if (!array)
1637 return NULL;
1638 access = isl_map_from_range(isl_set_copy(array->extent));
1639 id = isl_set_get_tuple_id(array->extent);
1640 space = isl_space_alloc(ctx, 0, 0, 0);
1641 space = isl_space_set_tuple_id(space, isl_dim_out, id);
1642 index = isl_multi_pw_aff_zero(space);
1643 expr = pet_expr_kill_from_access_and_index(access, index);
1644 return extract(stmt, expr);
1647 /* Construct a pet_scop for a (single) variable declaration.
1649 * The scop contains the variable being declared (as an array)
1650 * and a statement killing the array.
1652 * If the variable is initialized in the AST, then the scop
1653 * also contains an assignment to the variable.
1655 struct pet_scop *PetScan::extract(DeclStmt *stmt)
1657 Decl *decl;
1658 VarDecl *vd;
1659 struct pet_expr *lhs, *rhs, *pe;
1660 struct pet_scop *scop_decl, *scop;
1661 struct pet_array *array;
1663 if (!stmt->isSingleDecl()) {
1664 unsupported(stmt);
1665 return NULL;
1668 decl = stmt->getSingleDecl();
1669 vd = cast<VarDecl>(decl);
1671 array = extract_array(ctx, vd, NULL);
1672 if (array)
1673 array->declared = 1;
1674 scop_decl = kill(stmt, array);
1675 scop_decl = pet_scop_add_array(scop_decl, array);
1677 if (!vd->getInit())
1678 return scop_decl;
1680 lhs = extract_access_expr(vd);
1681 rhs = extract_expr(vd->getInit());
1683 mark_write(lhs);
1684 assign(lhs, vd->getInit());
1686 pe = pet_expr_new_binary(ctx, pet_op_assign, lhs, rhs);
1687 scop = extract(stmt, pe);
1689 scop_decl = pet_scop_prefix(scop_decl, 0);
1690 scop = pet_scop_prefix(scop, 1);
1692 scop = pet_scop_add_seq(ctx, scop_decl, scop);
1694 return scop;
1697 /* Construct a pet_expr representing a conditional operation.
1699 * We first try to extract the condition as an affine expression.
1700 * If that fails, we construct a pet_expr tree representing the condition.
1702 struct pet_expr *PetScan::extract_expr(ConditionalOperator *expr)
1704 struct pet_expr *cond, *lhs, *rhs;
1705 isl_pw_aff *pa;
1707 pa = try_extract_affine(expr->getCond());
1708 if (pa) {
1709 isl_multi_pw_aff *test = isl_multi_pw_aff_from_pw_aff(pa);
1710 test = isl_multi_pw_aff_from_range(test);
1711 cond = pet_expr_from_index(test);
1712 } else
1713 cond = extract_expr(expr->getCond());
1714 lhs = extract_expr(expr->getTrueExpr());
1715 rhs = extract_expr(expr->getFalseExpr());
1717 return pet_expr_new_ternary(ctx, cond, lhs, rhs);
1720 struct pet_expr *PetScan::extract_expr(ImplicitCastExpr *expr)
1722 return extract_expr(expr->getSubExpr());
1725 /* Construct a pet_expr representing a floating point value.
1727 * If the floating point literal does not appear in a macro,
1728 * then we use the original representation in the source code
1729 * as the string representation. Otherwise, we use the pretty
1730 * printer to produce a string representation.
1732 struct pet_expr *PetScan::extract_expr(FloatingLiteral *expr)
1734 double d;
1735 string s;
1736 const LangOptions &LO = PP.getLangOpts();
1737 SourceLocation loc = expr->getLocation();
1739 if (!loc.isMacroID()) {
1740 SourceManager &SM = PP.getSourceManager();
1741 unsigned len = Lexer::MeasureTokenLength(loc, SM, LO);
1742 s = string(SM.getCharacterData(loc), len);
1743 } else {
1744 llvm::raw_string_ostream S(s);
1745 expr->printPretty(S, 0, PrintingPolicy(LO));
1746 S.str();
1748 d = expr->getValueAsApproximateDouble();
1749 return pet_expr_new_double(ctx, d, s.c_str());
1752 /* Extract an index expression from "expr" and then convert it into
1753 * an access pet_expr.
1755 struct pet_expr *PetScan::extract_access_expr(Expr *expr)
1757 isl_multi_pw_aff *index;
1758 struct pet_expr *pe;
1759 int depth;
1761 index = extract_index(expr);
1762 depth = extract_depth(index);
1764 pe = pet_expr_from_index_and_depth(index, depth);
1766 return pe;
1769 /* Extract an index expression from "decl" and then convert it into
1770 * an access pet_expr.
1772 struct pet_expr *PetScan::extract_access_expr(ValueDecl *decl)
1774 isl_multi_pw_aff *index;
1775 struct pet_expr *pe;
1776 int depth;
1778 index = extract_index(decl);
1779 depth = extract_depth(index);
1781 pe = pet_expr_from_index_and_depth(index, depth);
1783 return pe;
1786 struct pet_expr *PetScan::extract_expr(ParenExpr *expr)
1788 return extract_expr(expr->getSubExpr());
1791 /* Construct a pet_expr representing a function call.
1793 * If we are passing along a pointer to an array element
1794 * or an entire row or even higher dimensional slice of an array,
1795 * then the function being called may write into the array.
1797 * We assume here that if the function is declared to take a pointer
1798 * to a const type, then the function will perform a read
1799 * and that otherwise, it will perform a write.
1801 struct pet_expr *PetScan::extract_expr(CallExpr *expr)
1803 struct pet_expr *res = NULL;
1804 FunctionDecl *fd;
1805 string name;
1807 fd = expr->getDirectCallee();
1808 if (!fd) {
1809 unsupported(expr);
1810 return NULL;
1813 name = fd->getDeclName().getAsString();
1814 res = pet_expr_new_call(ctx, name.c_str(), expr->getNumArgs());
1815 if (!res)
1816 return NULL;
1818 for (int i = 0; i < expr->getNumArgs(); ++i) {
1819 Expr *arg = expr->getArg(i);
1820 int is_addr = 0;
1821 pet_expr *main_arg;
1822 Stmt::StmtClass sc;
1824 if (arg->getStmtClass() == Stmt::ImplicitCastExprClass) {
1825 ImplicitCastExpr *ice = cast<ImplicitCastExpr>(arg);
1826 arg = ice->getSubExpr();
1828 if (arg->getStmtClass() == Stmt::UnaryOperatorClass) {
1829 UnaryOperator *op = cast<UnaryOperator>(arg);
1830 if (op->getOpcode() == UO_AddrOf) {
1831 is_addr = 1;
1832 arg = op->getSubExpr();
1835 res->args[i] = PetScan::extract_expr(arg);
1836 main_arg = res->args[i];
1837 if (is_addr)
1838 res->args[i] = pet_expr_new_unary(ctx,
1839 pet_op_address_of, res->args[i]);
1840 if (!res->args[i])
1841 goto error;
1842 sc = arg->getStmtClass();
1843 if ((sc == Stmt::ArraySubscriptExprClass ||
1844 sc == Stmt::MemberExprClass) &&
1845 array_depth(arg->getType().getTypePtr()) > 0)
1846 is_addr = 1;
1847 if (is_addr && main_arg->type == pet_expr_access) {
1848 ParmVarDecl *parm;
1849 if (!fd->hasPrototype()) {
1850 unsupported(expr, "prototype required");
1851 goto error;
1853 parm = fd->getParamDecl(i);
1854 if (!const_base(parm->getType()))
1855 mark_write(main_arg);
1859 return res;
1860 error:
1861 pet_expr_free(res);
1862 return NULL;
1865 /* Construct a pet_expr representing a (C style) cast.
1867 struct pet_expr *PetScan::extract_expr(CStyleCastExpr *expr)
1869 struct pet_expr *arg;
1870 QualType type;
1872 arg = extract_expr(expr->getSubExpr());
1873 if (!arg)
1874 return NULL;
1876 type = expr->getTypeAsWritten();
1877 return pet_expr_new_cast(ctx, type.getAsString().c_str(), arg);
1880 /* Try and onstruct a pet_expr representing "expr".
1882 struct pet_expr *PetScan::extract_expr(Expr *expr)
1884 switch (expr->getStmtClass()) {
1885 case Stmt::UnaryOperatorClass:
1886 return extract_expr(cast<UnaryOperator>(expr));
1887 case Stmt::CompoundAssignOperatorClass:
1888 case Stmt::BinaryOperatorClass:
1889 return extract_expr(cast<BinaryOperator>(expr));
1890 case Stmt::ImplicitCastExprClass:
1891 return extract_expr(cast<ImplicitCastExpr>(expr));
1892 case Stmt::ArraySubscriptExprClass:
1893 case Stmt::DeclRefExprClass:
1894 case Stmt::IntegerLiteralClass:
1895 case Stmt::MemberExprClass:
1896 return extract_access_expr(expr);
1897 case Stmt::FloatingLiteralClass:
1898 return extract_expr(cast<FloatingLiteral>(expr));
1899 case Stmt::ParenExprClass:
1900 return extract_expr(cast<ParenExpr>(expr));
1901 case Stmt::ConditionalOperatorClass:
1902 return extract_expr(cast<ConditionalOperator>(expr));
1903 case Stmt::CallExprClass:
1904 return extract_expr(cast<CallExpr>(expr));
1905 case Stmt::CStyleCastExprClass:
1906 return extract_expr(cast<CStyleCastExpr>(expr));
1907 default:
1908 unsupported(expr);
1910 return NULL;
1913 /* Check if the given initialization statement is an assignment.
1914 * If so, return that assignment. Otherwise return NULL.
1916 BinaryOperator *PetScan::initialization_assignment(Stmt *init)
1918 BinaryOperator *ass;
1920 if (init->getStmtClass() != Stmt::BinaryOperatorClass)
1921 return NULL;
1923 ass = cast<BinaryOperator>(init);
1924 if (ass->getOpcode() != BO_Assign)
1925 return NULL;
1927 return ass;
1930 /* Check if the given initialization statement is a declaration
1931 * of a single variable.
1932 * If so, return that declaration. Otherwise return NULL.
1934 Decl *PetScan::initialization_declaration(Stmt *init)
1936 DeclStmt *decl;
1938 if (init->getStmtClass() != Stmt::DeclStmtClass)
1939 return NULL;
1941 decl = cast<DeclStmt>(init);
1943 if (!decl->isSingleDecl())
1944 return NULL;
1946 return decl->getSingleDecl();
1949 /* Given the assignment operator in the initialization of a for loop,
1950 * extract the induction variable, i.e., the (integer)variable being
1951 * assigned.
1953 ValueDecl *PetScan::extract_induction_variable(BinaryOperator *init)
1955 Expr *lhs;
1956 DeclRefExpr *ref;
1957 ValueDecl *decl;
1958 const Type *type;
1960 lhs = init->getLHS();
1961 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
1962 unsupported(init);
1963 return NULL;
1966 ref = cast<DeclRefExpr>(lhs);
1967 decl = ref->getDecl();
1968 type = decl->getType().getTypePtr();
1970 if (!type->isIntegerType()) {
1971 unsupported(lhs);
1972 return NULL;
1975 return decl;
1978 /* Given the initialization statement of a for loop and the single
1979 * declaration in this initialization statement,
1980 * extract the induction variable, i.e., the (integer) variable being
1981 * declared.
1983 VarDecl *PetScan::extract_induction_variable(Stmt *init, Decl *decl)
1985 VarDecl *vd;
1987 vd = cast<VarDecl>(decl);
1989 const QualType type = vd->getType();
1990 if (!type->isIntegerType()) {
1991 unsupported(init);
1992 return NULL;
1995 if (!vd->getInit()) {
1996 unsupported(init);
1997 return NULL;
2000 return vd;
2003 /* Check that op is of the form iv++ or iv--.
2004 * Return an affine expression "1" or "-1" accordingly.
2006 __isl_give isl_pw_aff *PetScan::extract_unary_increment(
2007 clang::UnaryOperator *op, clang::ValueDecl *iv)
2009 Expr *sub;
2010 DeclRefExpr *ref;
2011 isl_space *space;
2012 isl_aff *aff;
2014 if (!op->isIncrementDecrementOp()) {
2015 unsupported(op);
2016 return NULL;
2019 sub = op->getSubExpr();
2020 if (sub->getStmtClass() != Stmt::DeclRefExprClass) {
2021 unsupported(op);
2022 return NULL;
2025 ref = cast<DeclRefExpr>(sub);
2026 if (ref->getDecl() != iv) {
2027 unsupported(op);
2028 return NULL;
2031 space = isl_space_params_alloc(ctx, 0);
2032 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2034 if (op->isIncrementOp())
2035 aff = isl_aff_add_constant_si(aff, 1);
2036 else
2037 aff = isl_aff_add_constant_si(aff, -1);
2039 return isl_pw_aff_from_aff(aff);
2042 /* If the isl_pw_aff on which isl_pw_aff_foreach_piece is called
2043 * has a single constant expression, then put this constant in *user.
2044 * The caller is assumed to have checked that this function will
2045 * be called exactly once.
2047 static int extract_cst(__isl_take isl_set *set, __isl_take isl_aff *aff,
2048 void *user)
2050 isl_val **inc = (isl_val **)user;
2051 int res = 0;
2053 if (isl_aff_is_cst(aff))
2054 *inc = isl_aff_get_constant_val(aff);
2055 else
2056 res = -1;
2058 isl_set_free(set);
2059 isl_aff_free(aff);
2061 return res;
2064 /* Check if op is of the form
2066 * iv = iv + inc
2068 * and return inc as an affine expression.
2070 * We extract an affine expression from the RHS, subtract iv and return
2071 * the result.
2073 __isl_give isl_pw_aff *PetScan::extract_binary_increment(BinaryOperator *op,
2074 clang::ValueDecl *iv)
2076 Expr *lhs;
2077 DeclRefExpr *ref;
2078 isl_id *id;
2079 isl_space *dim;
2080 isl_aff *aff;
2081 isl_pw_aff *val;
2083 if (op->getOpcode() != BO_Assign) {
2084 unsupported(op);
2085 return NULL;
2088 lhs = op->getLHS();
2089 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
2090 unsupported(op);
2091 return NULL;
2094 ref = cast<DeclRefExpr>(lhs);
2095 if (ref->getDecl() != iv) {
2096 unsupported(op);
2097 return NULL;
2100 val = extract_affine(op->getRHS());
2102 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
2104 dim = isl_space_params_alloc(ctx, 1);
2105 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
2106 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2107 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2109 val = isl_pw_aff_sub(val, isl_pw_aff_from_aff(aff));
2111 return val;
2114 /* Check that op is of the form iv += cst or iv -= cst
2115 * and return an affine expression corresponding oto cst or -cst accordingly.
2117 __isl_give isl_pw_aff *PetScan::extract_compound_increment(
2118 CompoundAssignOperator *op, clang::ValueDecl *iv)
2120 Expr *lhs;
2121 DeclRefExpr *ref;
2122 bool neg = false;
2123 isl_pw_aff *val;
2124 BinaryOperatorKind opcode;
2126 opcode = op->getOpcode();
2127 if (opcode != BO_AddAssign && opcode != BO_SubAssign) {
2128 unsupported(op);
2129 return NULL;
2131 if (opcode == BO_SubAssign)
2132 neg = true;
2134 lhs = op->getLHS();
2135 if (lhs->getStmtClass() != Stmt::DeclRefExprClass) {
2136 unsupported(op);
2137 return NULL;
2140 ref = cast<DeclRefExpr>(lhs);
2141 if (ref->getDecl() != iv) {
2142 unsupported(op);
2143 return NULL;
2146 val = extract_affine(op->getRHS());
2147 if (neg)
2148 val = isl_pw_aff_neg(val);
2150 return val;
2153 /* Check that the increment of the given for loop increments
2154 * (or decrements) the induction variable "iv" and return
2155 * the increment as an affine expression if successful.
2157 __isl_give isl_pw_aff *PetScan::extract_increment(clang::ForStmt *stmt,
2158 ValueDecl *iv)
2160 Stmt *inc = stmt->getInc();
2162 if (!inc) {
2163 unsupported(stmt);
2164 return NULL;
2167 if (inc->getStmtClass() == Stmt::UnaryOperatorClass)
2168 return extract_unary_increment(cast<UnaryOperator>(inc), iv);
2169 if (inc->getStmtClass() == Stmt::CompoundAssignOperatorClass)
2170 return extract_compound_increment(
2171 cast<CompoundAssignOperator>(inc), iv);
2172 if (inc->getStmtClass() == Stmt::BinaryOperatorClass)
2173 return extract_binary_increment(cast<BinaryOperator>(inc), iv);
2175 unsupported(inc);
2176 return NULL;
2179 /* Embed the given iteration domain in an extra outer loop
2180 * with induction variable "var".
2181 * If this variable appeared as a parameter in the constraints,
2182 * it is replaced by the new outermost dimension.
2184 static __isl_give isl_set *embed(__isl_take isl_set *set,
2185 __isl_take isl_id *var)
2187 int pos;
2189 set = isl_set_insert_dims(set, isl_dim_set, 0, 1);
2190 pos = isl_set_find_dim_by_id(set, isl_dim_param, var);
2191 if (pos >= 0) {
2192 set = isl_set_equate(set, isl_dim_param, pos, isl_dim_set, 0);
2193 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2196 isl_id_free(var);
2197 return set;
2200 /* Return those elements in the space of "cond" that come after
2201 * (based on "sign") an element in "cond".
2203 static __isl_give isl_set *after(__isl_take isl_set *cond, int sign)
2205 isl_map *previous_to_this;
2207 if (sign > 0)
2208 previous_to_this = isl_map_lex_lt(isl_set_get_space(cond));
2209 else
2210 previous_to_this = isl_map_lex_gt(isl_set_get_space(cond));
2212 cond = isl_set_apply(cond, previous_to_this);
2214 return cond;
2217 /* Create the infinite iteration domain
2219 * { [id] : id >= 0 }
2221 * If "scop" has an affine skip of type pet_skip_later,
2222 * then remove those iterations i that have an earlier iteration
2223 * where the skip condition is satisfied, meaning that iteration i
2224 * is not executed.
2225 * Since we are dealing with a loop without loop iterator,
2226 * the skip condition cannot refer to the current loop iterator and
2227 * so effectively, the returned set is of the form
2229 * { [0]; [id] : id >= 1 and not skip }
2231 static __isl_give isl_set *infinite_domain(__isl_take isl_id *id,
2232 struct pet_scop *scop)
2234 isl_ctx *ctx = isl_id_get_ctx(id);
2235 isl_set *domain;
2236 isl_set *skip;
2238 domain = isl_set_nat_universe(isl_space_set_alloc(ctx, 0, 1));
2239 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, id);
2241 if (!pet_scop_has_affine_skip(scop, pet_skip_later))
2242 return domain;
2244 skip = pet_scop_get_affine_skip_domain(scop, pet_skip_later);
2245 skip = embed(skip, isl_id_copy(id));
2246 skip = isl_set_intersect(skip , isl_set_copy(domain));
2247 domain = isl_set_subtract(domain, after(skip, 1));
2249 return domain;
2252 /* Create an identity affine expression on the space containing "domain",
2253 * which is assumed to be one-dimensional.
2255 static __isl_give isl_aff *identity_aff(__isl_keep isl_set *domain)
2257 isl_local_space *ls;
2259 ls = isl_local_space_from_space(isl_set_get_space(domain));
2260 return isl_aff_var_on_domain(ls, isl_dim_set, 0);
2263 /* Create an affine expression that maps elements
2264 * of a single-dimensional array "id_test" to the previous element
2265 * (according to "inc"), provided this element belongs to "domain".
2266 * That is, create the affine expression
2268 * { id[x] -> id[x - inc] : x - inc in domain }
2270 static __isl_give isl_multi_pw_aff *map_to_previous(__isl_take isl_id *id_test,
2271 __isl_take isl_set *domain, __isl_take isl_val *inc)
2273 isl_space *space;
2274 isl_local_space *ls;
2275 isl_aff *aff;
2276 isl_multi_pw_aff *prev;
2278 space = isl_set_get_space(domain);
2279 ls = isl_local_space_from_space(space);
2280 aff = isl_aff_var_on_domain(ls, isl_dim_set, 0);
2281 aff = isl_aff_add_constant_val(aff, isl_val_neg(inc));
2282 prev = isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
2283 domain = isl_set_preimage_multi_pw_aff(domain,
2284 isl_multi_pw_aff_copy(prev));
2285 prev = isl_multi_pw_aff_intersect_domain(prev, domain);
2286 prev = isl_multi_pw_aff_set_tuple_id(prev, isl_dim_out, id_test);
2288 return prev;
2291 /* Add an implication to "scop" expressing that if an element of
2292 * virtual array "id_test" has value "satisfied" then all previous elements
2293 * of this array also have that value. The set of previous elements
2294 * is bounded by "domain". If "sign" is negative then iterator
2295 * is decreasing and we express that all subsequent array elements
2296 * (but still defined previously) have the same value.
2298 static struct pet_scop *add_implication(struct pet_scop *scop,
2299 __isl_take isl_id *id_test, __isl_take isl_set *domain, int sign,
2300 int satisfied)
2302 isl_space *space;
2303 isl_map *map;
2305 domain = isl_set_set_tuple_id(domain, id_test);
2306 space = isl_set_get_space(domain);
2307 if (sign > 0)
2308 map = isl_map_lex_ge(space);
2309 else
2310 map = isl_map_lex_le(space);
2311 map = isl_map_intersect_range(map, domain);
2312 scop = pet_scop_add_implication(scop, map, satisfied);
2314 return scop;
2317 /* Add a filter to "scop" that imposes that it is only executed
2318 * when the variable identified by "id_test" has a zero value
2319 * for all previous iterations of "domain".
2321 * In particular, add a filter that imposes that the array
2322 * has a zero value at the previous iteration of domain and
2323 * add an implication that implies that it then has that
2324 * value for all previous iterations.
2326 static struct pet_scop *scop_add_break(struct pet_scop *scop,
2327 __isl_take isl_id *id_test, __isl_take isl_set *domain,
2328 __isl_take isl_val *inc)
2330 isl_multi_pw_aff *prev;
2331 int sign = isl_val_sgn(inc);
2333 prev = map_to_previous(isl_id_copy(id_test), isl_set_copy(domain), inc);
2334 scop = add_implication(scop, id_test, domain, sign, 0);
2335 scop = pet_scop_filter(scop, prev, 0);
2337 return scop;
2340 /* Construct a pet_scop for an infinite loop around the given body.
2342 * We extract a pet_scop for the body and then embed it in a loop with
2343 * iteration domain
2345 * { [t] : t >= 0 }
2347 * and schedule
2349 * { [t] -> [t] }
2351 * If the body contains any break, then it is taken into
2352 * account in infinite_domain (if the skip condition is affine)
2353 * or in scop_add_break (if the skip condition is not affine).
2355 struct pet_scop *PetScan::extract_infinite_loop(Stmt *body)
2357 isl_id *id, *id_test;
2358 isl_set *domain;
2359 isl_aff *ident;
2360 struct pet_scop *scop;
2361 bool has_var_break;
2363 scop = extract(body);
2364 if (!scop)
2365 return NULL;
2367 id = isl_id_alloc(ctx, "t", NULL);
2368 domain = infinite_domain(isl_id_copy(id), scop);
2369 ident = identity_aff(domain);
2371 has_var_break = pet_scop_has_var_skip(scop, pet_skip_later);
2372 if (has_var_break)
2373 id_test = pet_scop_get_skip_id(scop, pet_skip_later);
2375 scop = pet_scop_embed(scop, isl_set_copy(domain),
2376 isl_map_from_aff(isl_aff_copy(ident)), ident, id);
2377 if (has_var_break)
2378 scop = scop_add_break(scop, id_test, domain, isl_val_one(ctx));
2379 else
2380 isl_set_free(domain);
2382 return scop;
2385 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
2387 * for (;;)
2388 * body
2391 struct pet_scop *PetScan::extract_infinite_for(ForStmt *stmt)
2393 return extract_infinite_loop(stmt->getBody());
2396 /* Create an index expression for an access to a virtual array
2397 * representing the result of a condition.
2398 * Unlike other accessed data, the id of the array is NULL as
2399 * there is no ValueDecl in the program corresponding to the virtual
2400 * array.
2401 * The array starts out as a scalar, but grows along with the
2402 * statement writing to the array in pet_scop_embed.
2404 static __isl_give isl_multi_pw_aff *create_test_index(isl_ctx *ctx, int test_nr)
2406 isl_space *dim = isl_space_alloc(ctx, 0, 0, 0);
2407 isl_id *id;
2408 char name[50];
2410 snprintf(name, sizeof(name), "__pet_test_%d", test_nr);
2411 id = isl_id_alloc(ctx, name, NULL);
2412 dim = isl_space_set_tuple_id(dim, isl_dim_out, id);
2413 return isl_multi_pw_aff_zero(dim);
2416 /* Add an array with the given extent (range of "index") to the list
2417 * of arrays in "scop" and return the extended pet_scop.
2418 * The array is marked as attaining values 0 and 1 only and
2419 * as each element being assigned at most once.
2421 static struct pet_scop *scop_add_array(struct pet_scop *scop,
2422 __isl_keep isl_multi_pw_aff *index, clang::ASTContext &ast_ctx)
2424 isl_ctx *ctx = isl_multi_pw_aff_get_ctx(index);
2425 isl_space *dim;
2426 struct pet_array *array;
2427 isl_map *access;
2429 if (!scop)
2430 return NULL;
2431 if (!ctx)
2432 goto error;
2434 array = isl_calloc_type(ctx, struct pet_array);
2435 if (!array)
2436 goto error;
2438 access = isl_map_from_multi_pw_aff(isl_multi_pw_aff_copy(index));
2439 array->extent = isl_map_range(access);
2440 dim = isl_space_params_alloc(ctx, 0);
2441 array->context = isl_set_universe(dim);
2442 dim = isl_space_set_alloc(ctx, 0, 1);
2443 array->value_bounds = isl_set_universe(dim);
2444 array->value_bounds = isl_set_lower_bound_si(array->value_bounds,
2445 isl_dim_set, 0, 0);
2446 array->value_bounds = isl_set_upper_bound_si(array->value_bounds,
2447 isl_dim_set, 0, 1);
2448 array->element_type = strdup("int");
2449 array->element_size = ast_ctx.getTypeInfo(ast_ctx.IntTy).first / 8;
2450 array->uniquely_defined = 1;
2452 if (!array->extent || !array->context)
2453 array = pet_array_free(array);
2455 scop = pet_scop_add_array(scop, array);
2457 return scop;
2458 error:
2459 pet_scop_free(scop);
2460 return NULL;
2463 /* Construct a pet_scop for a while loop of the form
2465 * while (pa)
2466 * body
2468 * In particular, construct a scop for an infinite loop around body and
2469 * intersect the domain with the affine expression.
2470 * Note that this intersection may result in an empty loop.
2472 struct pet_scop *PetScan::extract_affine_while(__isl_take isl_pw_aff *pa,
2473 Stmt *body)
2475 struct pet_scop *scop;
2476 isl_set *dom;
2477 isl_set *valid;
2479 valid = isl_pw_aff_domain(isl_pw_aff_copy(pa));
2480 dom = isl_pw_aff_non_zero_set(pa);
2481 scop = extract_infinite_loop(body);
2482 scop = pet_scop_restrict(scop, dom);
2483 scop = pet_scop_restrict_context(scop, valid);
2485 return scop;
2488 /* Construct a scop for a while, given the scops for the condition
2489 * and the body, the filter identifier and the iteration domain of
2490 * the while loop.
2492 * In particular, the scop for the condition is filtered to depend
2493 * on "id_test" evaluating to true for all previous iterations
2494 * of the loop, while the scop for the body is filtered to depend
2495 * on "id_test" evaluating to true for all iterations up to the
2496 * current iteration.
2497 * The actual filter only imposes that this virtual array has
2498 * value one on the previous or the current iteration.
2499 * The fact that this condition also applies to the previous
2500 * iterations is enforced by an implication.
2502 * These filtered scops are then combined into a single scop.
2504 * "sign" is positive if the iterator increases and negative
2505 * if it decreases.
2507 static struct pet_scop *scop_add_while(struct pet_scop *scop_cond,
2508 struct pet_scop *scop_body, __isl_take isl_id *id_test,
2509 __isl_take isl_set *domain, __isl_take isl_val *inc)
2511 isl_ctx *ctx = isl_set_get_ctx(domain);
2512 isl_space *space;
2513 isl_multi_pw_aff *test_index;
2514 isl_multi_pw_aff *prev;
2515 int sign = isl_val_sgn(inc);
2516 struct pet_scop *scop;
2518 prev = map_to_previous(isl_id_copy(id_test), isl_set_copy(domain), inc);
2519 scop_cond = pet_scop_filter(scop_cond, prev, 1);
2521 space = isl_space_map_from_set(isl_set_get_space(domain));
2522 test_index = isl_multi_pw_aff_identity(space);
2523 test_index = isl_multi_pw_aff_set_tuple_id(test_index, isl_dim_out,
2524 isl_id_copy(id_test));
2525 scop_body = pet_scop_filter(scop_body, test_index, 1);
2527 scop = pet_scop_add_seq(ctx, scop_cond, scop_body);
2528 scop = add_implication(scop, id_test, domain, sign, 1);
2530 return scop;
2533 /* Check if the while loop is of the form
2535 * while (affine expression)
2536 * body
2538 * If so, call extract_affine_while to construct a scop.
2540 * Otherwise, construct a generic while scop, with iteration domain
2541 * { [t] : t >= 0 }. The scop consists of two parts, one for
2542 * evaluating the condition and one for the body.
2543 * The schedule is adjusted to reflect that the condition is evaluated
2544 * before the body is executed and the body is filtered to depend
2545 * on the result of the condition evaluating to true on all iterations
2546 * up to the current iteration, while the evaluation the condition itself
2547 * is filtered to depend on the result of the condition evaluating to true
2548 * on all previous iterations.
2549 * The context of the scop representing the body is dropped
2550 * because we don't know how many times the body will be executed,
2551 * if at all.
2553 * If the body contains any break, then it is taken into
2554 * account in infinite_domain (if the skip condition is affine)
2555 * or in scop_add_break (if the skip condition is not affine).
2557 struct pet_scop *PetScan::extract(WhileStmt *stmt)
2559 Expr *cond;
2560 isl_id *id, *id_test, *id_break_test;
2561 isl_multi_pw_aff *test_index;
2562 isl_set *domain;
2563 isl_aff *ident;
2564 isl_pw_aff *pa;
2565 struct pet_scop *scop, *scop_body;
2566 bool has_var_break;
2568 cond = stmt->getCond();
2569 if (!cond) {
2570 unsupported(stmt);
2571 return NULL;
2574 clear_assignments clear(assigned_value);
2575 clear.TraverseStmt(stmt->getBody());
2577 pa = try_extract_affine_condition(cond);
2578 if (pa)
2579 return extract_affine_while(pa, stmt->getBody());
2581 if (!allow_nested) {
2582 unsupported(stmt);
2583 return NULL;
2586 test_index = create_test_index(ctx, n_test++);
2587 scop = extract_non_affine_condition(cond,
2588 isl_multi_pw_aff_copy(test_index));
2589 scop = scop_add_array(scop, test_index, ast_context);
2590 id_test = isl_multi_pw_aff_get_tuple_id(test_index, isl_dim_out);
2591 isl_multi_pw_aff_free(test_index);
2592 scop_body = extract(stmt->getBody());
2594 id = isl_id_alloc(ctx, "t", NULL);
2595 domain = infinite_domain(isl_id_copy(id), scop_body);
2596 ident = identity_aff(domain);
2598 has_var_break = pet_scop_has_var_skip(scop_body, pet_skip_later);
2599 if (has_var_break)
2600 id_break_test = pet_scop_get_skip_id(scop_body, pet_skip_later);
2602 scop = pet_scop_prefix(scop, 0);
2603 scop = pet_scop_embed(scop, isl_set_copy(domain),
2604 isl_map_from_aff(isl_aff_copy(ident)),
2605 isl_aff_copy(ident), isl_id_copy(id));
2606 scop_body = pet_scop_reset_context(scop_body);
2607 scop_body = pet_scop_prefix(scop_body, 1);
2608 scop_body = pet_scop_embed(scop_body, isl_set_copy(domain),
2609 isl_map_from_aff(isl_aff_copy(ident)), ident, id);
2611 if (has_var_break) {
2612 scop = scop_add_break(scop, isl_id_copy(id_break_test),
2613 isl_set_copy(domain), isl_val_one(ctx));
2614 scop_body = scop_add_break(scop_body, id_break_test,
2615 isl_set_copy(domain), isl_val_one(ctx));
2617 scop = scop_add_while(scop, scop_body, id_test, domain,
2618 isl_val_one(ctx));
2620 return scop;
2623 /* Check whether "cond" expresses a simple loop bound
2624 * on the only set dimension.
2625 * In particular, if "up" is set then "cond" should contain only
2626 * upper bounds on the set dimension.
2627 * Otherwise, it should contain only lower bounds.
2629 static bool is_simple_bound(__isl_keep isl_set *cond, __isl_keep isl_val *inc)
2631 if (isl_val_is_pos(inc))
2632 return !isl_set_dim_has_any_lower_bound(cond, isl_dim_set, 0);
2633 else
2634 return !isl_set_dim_has_any_upper_bound(cond, isl_dim_set, 0);
2637 /* Extend a condition on a given iteration of a loop to one that
2638 * imposes the same condition on all previous iterations.
2639 * "domain" expresses the lower [upper] bound on the iterations
2640 * when inc is positive [negative].
2642 * In particular, we construct the condition (when inc is positive)
2644 * forall i' : (domain(i') and i' <= i) => cond(i')
2646 * which is equivalent to
2648 * not exists i' : domain(i') and i' <= i and not cond(i')
2650 * We construct this set by negating cond, applying a map
2652 * { [i'] -> [i] : domain(i') and i' <= i }
2654 * and then negating the result again.
2656 static __isl_give isl_set *valid_for_each_iteration(__isl_take isl_set *cond,
2657 __isl_take isl_set *domain, __isl_take isl_val *inc)
2659 isl_map *previous_to_this;
2661 if (isl_val_is_pos(inc))
2662 previous_to_this = isl_map_lex_le(isl_set_get_space(domain));
2663 else
2664 previous_to_this = isl_map_lex_ge(isl_set_get_space(domain));
2666 previous_to_this = isl_map_intersect_domain(previous_to_this, domain);
2668 cond = isl_set_complement(cond);
2669 cond = isl_set_apply(cond, previous_to_this);
2670 cond = isl_set_complement(cond);
2672 isl_val_free(inc);
2674 return cond;
2677 /* Construct a domain of the form
2679 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
2681 static __isl_give isl_set *strided_domain(__isl_take isl_id *id,
2682 __isl_take isl_pw_aff *init, __isl_take isl_val *inc)
2684 isl_aff *aff;
2685 isl_space *dim;
2686 isl_set *set;
2688 init = isl_pw_aff_insert_dims(init, isl_dim_in, 0, 1);
2689 dim = isl_pw_aff_get_domain_space(init);
2690 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2691 aff = isl_aff_add_coefficient_val(aff, isl_dim_in, 0, inc);
2692 init = isl_pw_aff_add(init, isl_pw_aff_from_aff(aff));
2694 dim = isl_space_set_alloc(isl_pw_aff_get_ctx(init), 1, 1);
2695 dim = isl_space_set_dim_id(dim, isl_dim_param, 0, id);
2696 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2697 aff = isl_aff_add_coefficient_si(aff, isl_dim_param, 0, 1);
2699 set = isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff), init);
2701 set = isl_set_lower_bound_si(set, isl_dim_set, 0, 0);
2703 return isl_set_params(set);
2706 /* Assuming "cond" represents a bound on a loop where the loop
2707 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2708 * is possible.
2710 * Under the given assumptions, wrapping is only possible if "cond" allows
2711 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2712 * increasing iterator and 0 in case of a decreasing iterator.
2714 static bool can_wrap(__isl_keep isl_set *cond, ValueDecl *iv,
2715 __isl_keep isl_val *inc)
2717 bool cw;
2718 isl_ctx *ctx;
2719 isl_val *limit;
2720 isl_set *test;
2722 test = isl_set_copy(cond);
2724 ctx = isl_set_get_ctx(test);
2725 if (isl_val_is_neg(inc))
2726 limit = isl_val_zero(ctx);
2727 else {
2728 limit = isl_val_int_from_ui(ctx, get_type_size(iv));
2729 limit = isl_val_2exp(limit);
2730 limit = isl_val_sub_ui(limit, 1);
2733 test = isl_set_fix_val(cond, isl_dim_set, 0, limit);
2734 cw = !isl_set_is_empty(test);
2735 isl_set_free(test);
2737 return cw;
2740 /* Given a one-dimensional space, construct the following affine expression
2741 * on this space
2743 * { [v] -> [v mod 2^width] }
2745 * where width is the number of bits used to represent the values
2746 * of the unsigned variable "iv".
2748 static __isl_give isl_aff *compute_wrapping(__isl_take isl_space *dim,
2749 ValueDecl *iv)
2751 isl_ctx *ctx;
2752 isl_val *mod;
2753 isl_aff *aff;
2755 ctx = isl_space_get_ctx(dim);
2756 mod = isl_val_int_from_ui(ctx, get_type_size(iv));
2757 mod = isl_val_2exp(mod);
2759 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
2760 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2761 aff = isl_aff_mod_val(aff, mod);
2763 return aff;
2766 /* Project out the parameter "id" from "set".
2768 static __isl_give isl_set *set_project_out_by_id(__isl_take isl_set *set,
2769 __isl_keep isl_id *id)
2771 int pos;
2773 pos = isl_set_find_dim_by_id(set, isl_dim_param, id);
2774 if (pos >= 0)
2775 set = isl_set_project_out(set, isl_dim_param, pos, 1);
2777 return set;
2780 /* Compute the set of parameters for which "set1" is a subset of "set2".
2782 * set1 is a subset of set2 if
2784 * forall i in set1 : i in set2
2786 * or
2788 * not exists i in set1 and i not in set2
2790 * i.e.,
2792 * not exists i in set1 \ set2
2794 static __isl_give isl_set *enforce_subset(__isl_take isl_set *set1,
2795 __isl_take isl_set *set2)
2797 return isl_set_complement(isl_set_params(isl_set_subtract(set1, set2)));
2800 /* Compute the set of parameter values for which "cond" holds
2801 * on the next iteration for each element of "dom".
2803 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2804 * and then compute the set of parameters for which the result is a subset
2805 * of "cond".
2807 static __isl_give isl_set *valid_on_next(__isl_take isl_set *cond,
2808 __isl_take isl_set *dom, __isl_take isl_val *inc)
2810 isl_space *space;
2811 isl_aff *aff;
2812 isl_map *next;
2814 space = isl_set_get_space(dom);
2815 aff = isl_aff_zero_on_domain(isl_local_space_from_space(space));
2816 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, 0, 1);
2817 aff = isl_aff_add_constant_val(aff, inc);
2818 next = isl_map_from_basic_map(isl_basic_map_from_aff(aff));
2820 dom = isl_set_apply(dom, next);
2822 return enforce_subset(dom, cond);
2825 /* Does "id" refer to a nested access?
2827 static bool is_nested_parameter(__isl_keep isl_id *id)
2829 return id && isl_id_get_user(id) && !isl_id_get_name(id);
2832 /* Does parameter "pos" of "space" refer to a nested access?
2834 static bool is_nested_parameter(__isl_keep isl_space *space, int pos)
2836 bool nested;
2837 isl_id *id;
2839 id = isl_space_get_dim_id(space, isl_dim_param, pos);
2840 nested = is_nested_parameter(id);
2841 isl_id_free(id);
2843 return nested;
2846 /* Does "space" involve any parameters that refer to nested
2847 * accesses, i.e., parameters with no name?
2849 static bool has_nested(__isl_keep isl_space *space)
2851 int nparam;
2853 nparam = isl_space_dim(space, isl_dim_param);
2854 for (int i = 0; i < nparam; ++i)
2855 if (is_nested_parameter(space, i))
2856 return true;
2858 return false;
2861 /* Does "pa" involve any parameters that refer to nested
2862 * accesses, i.e., parameters with no name?
2864 static bool has_nested(__isl_keep isl_pw_aff *pa)
2866 isl_space *space;
2867 bool nested;
2869 space = isl_pw_aff_get_space(pa);
2870 nested = has_nested(space);
2871 isl_space_free(space);
2873 return nested;
2876 /* Construct a pet_scop for a for statement.
2877 * The for loop is required to be of the form
2879 * for (i = init; condition; ++i)
2881 * or
2883 * for (i = init; condition; --i)
2885 * The initialization of the for loop should either be an assignment
2886 * to an integer variable, or a declaration of such a variable with
2887 * initialization.
2889 * The condition is allowed to contain nested accesses, provided
2890 * they are not being written to inside the body of the loop.
2891 * Otherwise, or if the condition is otherwise non-affine, the for loop is
2892 * essentially treated as a while loop, with iteration domain
2893 * { [i] : i >= init }.
2895 * We extract a pet_scop for the body and then embed it in a loop with
2896 * iteration domain and schedule
2898 * { [i] : i >= init and condition' }
2899 * { [i] -> [i] }
2901 * or
2903 * { [i] : i <= init and condition' }
2904 * { [i] -> [-i] }
2906 * Where condition' is equal to condition if the latter is
2907 * a simple upper [lower] bound and a condition that is extended
2908 * to apply to all previous iterations otherwise.
2910 * If the condition is non-affine, then we drop the condition from the
2911 * iteration domain and instead create a separate statement
2912 * for evaluating the condition. The body is then filtered to depend
2913 * on the result of the condition evaluating to true on all iterations
2914 * up to the current iteration, while the evaluation the condition itself
2915 * is filtered to depend on the result of the condition evaluating to true
2916 * on all previous iterations.
2917 * The context of the scop representing the body is dropped
2918 * because we don't know how many times the body will be executed,
2919 * if at all.
2921 * If the stride of the loop is not 1, then "i >= init" is replaced by
2923 * (exists a: i = init + stride * a and a >= 0)
2925 * If the loop iterator i is unsigned, then wrapping may occur.
2926 * During the computation, we work with a virtual iterator that
2927 * does not wrap. However, the condition in the code applies
2928 * to the wrapped value, so we need to change condition(i)
2929 * into condition([i % 2^width]).
2930 * After computing the virtual domain and schedule, we apply
2931 * the function { [v] -> [v % 2^width] } to the domain and the domain
2932 * of the schedule. In order not to lose any information, we also
2933 * need to intersect the domain of the schedule with the virtual domain
2934 * first, since some iterations in the wrapped domain may be scheduled
2935 * several times, typically an infinite number of times.
2936 * Note that there may be no need to perform this final wrapping
2937 * if the loop condition (after wrapping) satisfies certain conditions.
2938 * However, the is_simple_bound condition is not enough since it doesn't
2939 * check if there even is an upper bound.
2941 * If the loop condition is non-affine, then we keep the virtual
2942 * iterator in the iteration domain and instead replace all accesses
2943 * to the original iterator by the wrapping of the virtual iterator.
2945 * Wrapping on unsigned iterators can be avoided entirely if
2946 * loop condition is simple, the loop iterator is incremented
2947 * [decremented] by one and the last value before wrapping cannot
2948 * possibly satisfy the loop condition.
2950 * Before extracting a pet_scop from the body we remove all
2951 * assignments in assigned_value to variables that are assigned
2952 * somewhere in the body of the loop.
2954 * Valid parameters for a for loop are those for which the initial
2955 * value itself, the increment on each domain iteration and
2956 * the condition on both the initial value and
2957 * the result of incrementing the iterator for each iteration of the domain
2958 * can be evaluated.
2959 * If the loop condition is non-affine, then we only consider validity
2960 * of the initial value.
2962 * If the body contains any break, then we keep track of it in "skip"
2963 * (if the skip condition is affine) or it is handled in scop_add_break
2964 * (if the skip condition is not affine).
2965 * Note that the affine break condition needs to be considered with
2966 * respect to previous iterations in the virtual domain (if any)
2967 * and that the domain needs to be kept virtual if there is a non-affine
2968 * break condition.
2970 struct pet_scop *PetScan::extract_for(ForStmt *stmt)
2972 BinaryOperator *ass;
2973 Decl *decl;
2974 Stmt *init;
2975 Expr *lhs, *rhs;
2976 ValueDecl *iv;
2977 isl_space *space;
2978 isl_set *domain;
2979 isl_map *sched;
2980 isl_set *cond = NULL;
2981 isl_set *skip = NULL;
2982 isl_id *id, *id_test = NULL, *id_break_test;
2983 struct pet_scop *scop, *scop_cond = NULL;
2984 assigned_value_cache cache(assigned_value);
2985 isl_val *inc;
2986 bool is_one;
2987 bool is_unsigned;
2988 bool is_simple;
2989 bool is_virtual;
2990 bool keep_virtual = false;
2991 bool has_affine_break;
2992 bool has_var_break;
2993 isl_aff *wrap = NULL;
2994 isl_pw_aff *pa, *pa_inc, *init_val;
2995 isl_set *valid_init;
2996 isl_set *valid_cond;
2997 isl_set *valid_cond_init;
2998 isl_set *valid_cond_next;
2999 isl_set *valid_inc;
3000 int stmt_id;
3002 if (!stmt->getInit() && !stmt->getCond() && !stmt->getInc())
3003 return extract_infinite_for(stmt);
3005 init = stmt->getInit();
3006 if (!init) {
3007 unsupported(stmt);
3008 return NULL;
3010 if ((ass = initialization_assignment(init)) != NULL) {
3011 iv = extract_induction_variable(ass);
3012 if (!iv)
3013 return NULL;
3014 lhs = ass->getLHS();
3015 rhs = ass->getRHS();
3016 } else if ((decl = initialization_declaration(init)) != NULL) {
3017 VarDecl *var = extract_induction_variable(init, decl);
3018 if (!var)
3019 return NULL;
3020 iv = var;
3021 rhs = var->getInit();
3022 lhs = create_DeclRefExpr(var);
3023 } else {
3024 unsupported(stmt->getInit());
3025 return NULL;
3028 pa_inc = extract_increment(stmt, iv);
3029 if (!pa_inc)
3030 return NULL;
3032 inc = NULL;
3033 if (isl_pw_aff_n_piece(pa_inc) != 1 ||
3034 isl_pw_aff_foreach_piece(pa_inc, &extract_cst, &inc) < 0) {
3035 isl_pw_aff_free(pa_inc);
3036 unsupported(stmt->getInc());
3037 isl_val_free(inc);
3038 return NULL;
3040 valid_inc = isl_pw_aff_domain(pa_inc);
3042 is_unsigned = iv->getType()->isUnsignedIntegerType();
3044 assigned_value.erase(iv);
3045 clear_assignments clear(assigned_value);
3046 clear.TraverseStmt(stmt->getBody());
3048 id = isl_id_alloc(ctx, iv->getName().str().c_str(), iv);
3050 pa = try_extract_nested_condition(stmt->getCond());
3051 if (allow_nested && (!pa || has_nested(pa)))
3052 stmt_id = n_stmt++;
3054 scop = extract(stmt->getBody());
3056 has_affine_break = scop &&
3057 pet_scop_has_affine_skip(scop, pet_skip_later);
3058 if (has_affine_break)
3059 skip = pet_scop_get_affine_skip_domain(scop, pet_skip_later);
3060 has_var_break = scop && pet_scop_has_var_skip(scop, pet_skip_later);
3061 if (has_var_break) {
3062 id_break_test = pet_scop_get_skip_id(scop, pet_skip_later);
3063 keep_virtual = true;
3066 if (pa && !is_nested_allowed(pa, scop)) {
3067 isl_pw_aff_free(pa);
3068 pa = NULL;
3071 if (!allow_nested && !pa)
3072 pa = try_extract_affine_condition(stmt->getCond());
3073 valid_cond = isl_pw_aff_domain(isl_pw_aff_copy(pa));
3074 cond = isl_pw_aff_non_zero_set(pa);
3075 if (allow_nested && !cond) {
3076 isl_multi_pw_aff *test_index;
3077 int save_n_stmt = n_stmt;
3078 test_index = create_test_index(ctx, n_test++);
3079 n_stmt = stmt_id;
3080 scop_cond = extract_non_affine_condition(stmt->getCond(),
3081 isl_multi_pw_aff_copy(test_index));
3082 n_stmt = save_n_stmt;
3083 scop_cond = scop_add_array(scop_cond, test_index, ast_context);
3084 id_test = isl_multi_pw_aff_get_tuple_id(test_index,
3085 isl_dim_out);
3086 isl_multi_pw_aff_free(test_index);
3087 scop_cond = pet_scop_prefix(scop_cond, 0);
3088 scop = pet_scop_reset_context(scop);
3089 scop = pet_scop_prefix(scop, 1);
3090 keep_virtual = true;
3091 cond = isl_set_universe(isl_space_set_alloc(ctx, 0, 0));
3094 cond = embed(cond, isl_id_copy(id));
3095 skip = embed(skip, isl_id_copy(id));
3096 valid_cond = isl_set_coalesce(valid_cond);
3097 valid_cond = embed(valid_cond, isl_id_copy(id));
3098 valid_inc = embed(valid_inc, isl_id_copy(id));
3099 is_one = isl_val_is_one(inc) || isl_val_is_negone(inc);
3100 is_virtual = is_unsigned && (!is_one || can_wrap(cond, iv, inc));
3102 init_val = extract_affine(rhs);
3103 valid_cond_init = enforce_subset(
3104 isl_set_from_pw_aff(isl_pw_aff_copy(init_val)),
3105 isl_set_copy(valid_cond));
3106 if (is_one && !is_virtual) {
3107 isl_pw_aff_free(init_val);
3108 pa = extract_comparison(isl_val_is_pos(inc) ? BO_GE : BO_LE,
3109 lhs, rhs, init);
3110 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(pa));
3111 valid_init = set_project_out_by_id(valid_init, id);
3112 domain = isl_pw_aff_non_zero_set(pa);
3113 } else {
3114 valid_init = isl_pw_aff_domain(isl_pw_aff_copy(init_val));
3115 domain = strided_domain(isl_id_copy(id), init_val,
3116 isl_val_copy(inc));
3119 domain = embed(domain, isl_id_copy(id));
3120 if (is_virtual) {
3121 isl_map *rev_wrap;
3122 wrap = compute_wrapping(isl_set_get_space(cond), iv);
3123 rev_wrap = isl_map_from_aff(isl_aff_copy(wrap));
3124 rev_wrap = isl_map_reverse(rev_wrap);
3125 cond = isl_set_apply(cond, isl_map_copy(rev_wrap));
3126 skip = isl_set_apply(skip, isl_map_copy(rev_wrap));
3127 valid_cond = isl_set_apply(valid_cond, isl_map_copy(rev_wrap));
3128 valid_inc = isl_set_apply(valid_inc, rev_wrap);
3130 is_simple = is_simple_bound(cond, inc);
3131 if (!is_simple) {
3132 cond = isl_set_gist(cond, isl_set_copy(domain));
3133 is_simple = is_simple_bound(cond, inc);
3135 if (!is_simple)
3136 cond = valid_for_each_iteration(cond,
3137 isl_set_copy(domain), isl_val_copy(inc));
3138 domain = isl_set_intersect(domain, cond);
3139 if (has_affine_break) {
3140 skip = isl_set_intersect(skip , isl_set_copy(domain));
3141 skip = after(skip, isl_val_sgn(inc));
3142 domain = isl_set_subtract(domain, skip);
3144 domain = isl_set_set_dim_id(domain, isl_dim_set, 0, isl_id_copy(id));
3145 space = isl_space_from_domain(isl_set_get_space(domain));
3146 space = isl_space_add_dims(space, isl_dim_out, 1);
3147 sched = isl_map_universe(space);
3148 if (isl_val_is_pos(inc))
3149 sched = isl_map_equate(sched, isl_dim_in, 0, isl_dim_out, 0);
3150 else
3151 sched = isl_map_oppose(sched, isl_dim_in, 0, isl_dim_out, 0);
3153 valid_cond_next = valid_on_next(valid_cond, isl_set_copy(domain),
3154 isl_val_copy(inc));
3155 valid_inc = enforce_subset(isl_set_copy(domain), valid_inc);
3157 if (is_virtual && !keep_virtual) {
3158 isl_map *wrap_map = isl_map_from_aff(wrap);
3159 wrap_map = isl_map_set_dim_id(wrap_map,
3160 isl_dim_out, 0, isl_id_copy(id));
3161 sched = isl_map_intersect_domain(sched, isl_set_copy(domain));
3162 domain = isl_set_apply(domain, isl_map_copy(wrap_map));
3163 sched = isl_map_apply_domain(sched, wrap_map);
3165 if (!(is_virtual && keep_virtual))
3166 wrap = identity_aff(domain);
3168 scop_cond = pet_scop_embed(scop_cond, isl_set_copy(domain),
3169 isl_map_copy(sched), isl_aff_copy(wrap), isl_id_copy(id));
3170 scop = pet_scop_embed(scop, isl_set_copy(domain), sched, wrap, id);
3171 scop = resolve_nested(scop);
3172 if (has_var_break)
3173 scop = scop_add_break(scop, id_break_test, isl_set_copy(domain),
3174 isl_val_copy(inc));
3175 if (id_test) {
3176 scop = scop_add_while(scop_cond, scop, id_test, domain,
3177 isl_val_copy(inc));
3178 isl_set_free(valid_inc);
3179 } else {
3180 scop = pet_scop_restrict_context(scop, valid_inc);
3181 scop = pet_scop_restrict_context(scop, valid_cond_next);
3182 scop = pet_scop_restrict_context(scop, valid_cond_init);
3183 isl_set_free(domain);
3185 clear_assignment(assigned_value, iv);
3187 isl_val_free(inc);
3189 scop = pet_scop_restrict_context(scop, valid_init);
3191 return scop;
3194 struct pet_scop *PetScan::extract(CompoundStmt *stmt, bool skip_declarations)
3196 return extract(stmt->children(), true, skip_declarations);
3199 /* Does parameter "pos" of "map" refer to a nested access?
3201 static bool is_nested_parameter(__isl_keep isl_map *map, int pos)
3203 bool nested;
3204 isl_id *id;
3206 id = isl_map_get_dim_id(map, isl_dim_param, pos);
3207 nested = is_nested_parameter(id);
3208 isl_id_free(id);
3210 return nested;
3213 /* How many parameters of "space" refer to nested accesses, i.e., have no name?
3215 static int n_nested_parameter(__isl_keep isl_space *space)
3217 int n = 0;
3218 int nparam;
3220 nparam = isl_space_dim(space, isl_dim_param);
3221 for (int i = 0; i < nparam; ++i)
3222 if (is_nested_parameter(space, i))
3223 ++n;
3225 return n;
3228 /* How many parameters of "map" refer to nested accesses, i.e., have no name?
3230 static int n_nested_parameter(__isl_keep isl_map *map)
3232 isl_space *space;
3233 int n;
3235 space = isl_map_get_space(map);
3236 n = n_nested_parameter(space);
3237 isl_space_free(space);
3239 return n;
3242 /* For each nested access parameter in "space",
3243 * construct a corresponding pet_expr, place it in args and
3244 * record its position in "param2pos".
3245 * "n_arg" is the number of elements that are already in args.
3246 * The position recorded in "param2pos" takes this number into account.
3247 * If the pet_expr corresponding to a parameter is identical to
3248 * the pet_expr corresponding to an earlier parameter, then these two
3249 * parameters are made to refer to the same element in args.
3251 * Return the final number of elements in args or -1 if an error has occurred.
3253 int PetScan::extract_nested(__isl_keep isl_space *space,
3254 int n_arg, struct pet_expr **args, std::map<int,int> &param2pos)
3256 int nparam;
3258 nparam = isl_space_dim(space, isl_dim_param);
3259 for (int i = 0; i < nparam; ++i) {
3260 int j;
3261 isl_id *id = isl_space_get_dim_id(space, isl_dim_param, i);
3262 Expr *nested;
3264 if (!is_nested_parameter(id)) {
3265 isl_id_free(id);
3266 continue;
3269 nested = (Expr *) isl_id_get_user(id);
3270 args[n_arg] = extract_expr(nested);
3271 if (!args[n_arg])
3272 return -1;
3274 for (j = 0; j < n_arg; ++j)
3275 if (pet_expr_is_equal(args[j], args[n_arg]))
3276 break;
3278 if (j < n_arg) {
3279 pet_expr_free(args[n_arg]);
3280 args[n_arg] = NULL;
3281 param2pos[i] = j;
3282 } else
3283 param2pos[i] = n_arg++;
3285 isl_id_free(id);
3288 return n_arg;
3291 /* For each nested access parameter in the access relations in "expr",
3292 * construct a corresponding pet_expr, place it in expr->args and
3293 * record its position in "param2pos".
3294 * n is the number of nested access parameters.
3296 struct pet_expr *PetScan::extract_nested(struct pet_expr *expr, int n,
3297 std::map<int,int> &param2pos)
3299 isl_space *space;
3301 expr->args = isl_calloc_array(ctx, struct pet_expr *, n);
3302 expr->n_arg = n;
3303 if (!expr->args)
3304 goto error;
3306 space = isl_map_get_space(expr->acc.access);
3307 n = extract_nested(space, 0, expr->args, param2pos);
3308 isl_space_free(space);
3310 if (n < 0)
3311 goto error;
3313 expr->n_arg = n;
3314 return expr;
3315 error:
3316 pet_expr_free(expr);
3317 return NULL;
3320 /* Look for parameters in any access relation in "expr" that
3321 * refer to nested accesses. In particular, these are
3322 * parameters with no name.
3324 * If there are any such parameters, then the domain of the index
3325 * expression and the access relation, which is still [] at this point,
3326 * is replaced by [[] -> [t_1,...,t_n]], with n the number of these parameters
3327 * (after identifying identical nested accesses).
3329 * This transformation is performed in several steps.
3330 * We first extract the arguments in extract_nested.
3331 * param2pos maps the original parameter position to the position
3332 * of the argument.
3333 * Then we move these parameters to input dimension.
3334 * t2pos maps the positions of these temporary input dimensions
3335 * to the positions of the corresponding arguments.
3336 * Finally, we express there temporary dimensions in term of the domain
3337 * [[] -> [t_1,...,t_n]] and precompose index expression and access
3338 * relations with this function.
3340 struct pet_expr *PetScan::resolve_nested(struct pet_expr *expr)
3342 int n;
3343 int nparam;
3344 isl_space *space;
3345 isl_local_space *ls;
3346 isl_aff *aff;
3347 isl_multi_aff *ma;
3348 std::map<int,int> param2pos;
3349 std::map<int,int> t2pos;
3351 if (!expr)
3352 return expr;
3354 for (int i = 0; i < expr->n_arg; ++i) {
3355 expr->args[i] = resolve_nested(expr->args[i]);
3356 if (!expr->args[i]) {
3357 pet_expr_free(expr);
3358 return NULL;
3362 if (expr->type != pet_expr_access)
3363 return expr;
3365 n = n_nested_parameter(expr->acc.access);
3366 if (n == 0)
3367 return expr;
3369 expr = extract_nested(expr, n, param2pos);
3370 if (!expr)
3371 return NULL;
3373 expr = pet_expr_access_align_params(expr);
3374 if (!expr)
3375 return NULL;
3376 nparam = isl_map_dim(expr->acc.access, isl_dim_param);
3378 n = 0;
3379 for (int i = nparam - 1; i >= 0; --i) {
3380 isl_id *id = isl_map_get_dim_id(expr->acc.access,
3381 isl_dim_param, i);
3382 if (!is_nested_parameter(id)) {
3383 isl_id_free(id);
3384 continue;
3387 expr->acc.access = isl_map_move_dims(expr->acc.access,
3388 isl_dim_in, n, isl_dim_param, i, 1);
3389 expr->acc.index = isl_multi_pw_aff_move_dims(expr->acc.index,
3390 isl_dim_in, n, isl_dim_param, i, 1);
3391 t2pos[n] = param2pos[i];
3392 n++;
3394 isl_id_free(id);
3397 space = isl_multi_pw_aff_get_space(expr->acc.index);
3398 space = isl_space_set_from_params(isl_space_params(space));
3399 space = isl_space_add_dims(space, isl_dim_set, expr->n_arg);
3400 space = isl_space_wrap(isl_space_from_range(space));
3401 ls = isl_local_space_from_space(isl_space_copy(space));
3402 space = isl_space_from_domain(space);
3403 space = isl_space_add_dims(space, isl_dim_out, n);
3404 ma = isl_multi_aff_zero(space);
3406 for (int i = 0; i < n; ++i) {
3407 aff = isl_aff_var_on_domain(isl_local_space_copy(ls),
3408 isl_dim_set, t2pos[i]);
3409 ma = isl_multi_aff_set_aff(ma, i, aff);
3411 isl_local_space_free(ls);
3413 expr->acc.access = isl_map_preimage_domain_multi_aff(expr->acc.access,
3414 isl_multi_aff_copy(ma));
3415 expr->acc.index = isl_multi_pw_aff_pullback_multi_aff(expr->acc.index,
3416 ma);
3418 return expr;
3421 /* Return the file offset of the expansion location of "Loc".
3423 static unsigned getExpansionOffset(SourceManager &SM, SourceLocation Loc)
3425 return SM.getFileOffset(SM.getExpansionLoc(Loc));
3428 #ifdef HAVE_FINDLOCATIONAFTERTOKEN
3430 /* Return a SourceLocation for the location after the first semicolon
3431 * after "loc". If Lexer::findLocationAfterToken is available, we simply
3432 * call it and also skip trailing spaces and newline.
3434 static SourceLocation location_after_semi(SourceLocation loc, SourceManager &SM,
3435 const LangOptions &LO)
3437 return Lexer::findLocationAfterToken(loc, tok::semi, SM, LO, true);
3440 #else
3442 /* Return a SourceLocation for the location after the first semicolon
3443 * after "loc". If Lexer::findLocationAfterToken is not available,
3444 * we look in the underlying character data for the first semicolon.
3446 static SourceLocation location_after_semi(SourceLocation loc, SourceManager &SM,
3447 const LangOptions &LO)
3449 const char *semi;
3450 const char *s = SM.getCharacterData(loc);
3452 semi = strchr(s, ';');
3453 if (!semi)
3454 return SourceLocation();
3455 return loc.getFileLocWithOffset(semi + 1 - s);
3458 #endif
3460 /* If the token at "loc" is the first token on the line, then return
3461 * a location referring to the start of the line.
3462 * Otherwise, return "loc".
3464 * This function is used to extend a scop to the start of the line
3465 * if the first token of the scop is also the first token on the line.
3467 * We look for the first token on the line. If its location is equal to "loc",
3468 * then the latter is the location of the first token on the line.
3470 static SourceLocation move_to_start_of_line_if_first_token(SourceLocation loc,
3471 SourceManager &SM, const LangOptions &LO)
3473 std::pair<FileID, unsigned> file_offset_pair;
3474 llvm::StringRef file;
3475 const char *pos;
3476 Token tok;
3477 SourceLocation token_loc, line_loc;
3478 int col;
3480 loc = SM.getExpansionLoc(loc);
3481 col = SM.getExpansionColumnNumber(loc);
3482 line_loc = loc.getLocWithOffset(1 - col);
3483 file_offset_pair = SM.getDecomposedLoc(line_loc);
3484 file = SM.getBufferData(file_offset_pair.first, NULL);
3485 pos = file.data() + file_offset_pair.second;
3487 Lexer lexer(SM.getLocForStartOfFile(file_offset_pair.first), LO,
3488 file.begin(), pos, file.end());
3489 lexer.LexFromRawLexer(tok);
3490 token_loc = tok.getLocation();
3492 if (token_loc == loc)
3493 return line_loc;
3494 else
3495 return loc;
3498 /* Convert a top-level pet_expr to a pet_scop with one statement.
3499 * This mainly involves resolving nested expression parameters
3500 * and setting the name of the iteration space.
3501 * The name is given by "label" if it is non-NULL. Otherwise,
3502 * it is of the form S_<n_stmt>.
3503 * start and end of the pet_scop are derived from those of "stmt".
3505 struct pet_scop *PetScan::extract(Stmt *stmt, struct pet_expr *expr,
3506 __isl_take isl_id *label)
3508 struct pet_stmt *ps;
3509 struct pet_scop *scop;
3510 SourceLocation loc = stmt->getLocStart();
3511 SourceManager &SM = PP.getSourceManager();
3512 const LangOptions &LO = PP.getLangOpts();
3513 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3514 unsigned start, end;
3516 expr = resolve_nested(expr);
3517 ps = pet_stmt_from_pet_expr(ctx, line, label, n_stmt++, expr);
3518 scop = pet_scop_from_pet_stmt(ctx, ps);
3520 loc = move_to_start_of_line_if_first_token(loc, SM, LO);
3521 start = getExpansionOffset(SM, loc);
3522 loc = stmt->getLocEnd();
3523 loc = location_after_semi(loc, SM, LO);
3524 end = getExpansionOffset(SM, loc);
3526 scop = pet_scop_update_start_end(scop, start, end);
3527 return scop;
3530 /* Check if we can extract an affine expression from "expr".
3531 * Return the expressions as an isl_pw_aff if we can and NULL otherwise.
3532 * We turn on autodetection so that we won't generate any warnings
3533 * and turn off nesting, so that we won't accept any non-affine constructs.
3535 __isl_give isl_pw_aff *PetScan::try_extract_affine(Expr *expr)
3537 isl_pw_aff *pwaff;
3538 int save_autodetect = options->autodetect;
3539 bool save_nesting = nesting_enabled;
3541 options->autodetect = 1;
3542 nesting_enabled = false;
3544 pwaff = extract_affine(expr);
3546 options->autodetect = save_autodetect;
3547 nesting_enabled = save_nesting;
3549 return pwaff;
3552 /* Check whether "expr" is an affine expression.
3554 bool PetScan::is_affine(Expr *expr)
3556 isl_pw_aff *pwaff;
3558 pwaff = try_extract_affine(expr);
3559 isl_pw_aff_free(pwaff);
3561 return pwaff != NULL;
3564 /* Check if we can extract an affine constraint from "expr".
3565 * Return the constraint as an isl_set if we can and NULL otherwise.
3566 * We turn on autodetection so that we won't generate any warnings
3567 * and turn off nesting, so that we won't accept any non-affine constructs.
3569 __isl_give isl_pw_aff *PetScan::try_extract_affine_condition(Expr *expr)
3571 isl_pw_aff *cond;
3572 int save_autodetect = options->autodetect;
3573 bool save_nesting = nesting_enabled;
3575 options->autodetect = 1;
3576 nesting_enabled = false;
3578 cond = extract_condition(expr);
3580 options->autodetect = save_autodetect;
3581 nesting_enabled = save_nesting;
3583 return cond;
3586 /* Check whether "expr" is an affine constraint.
3588 bool PetScan::is_affine_condition(Expr *expr)
3590 isl_pw_aff *cond;
3592 cond = try_extract_affine_condition(expr);
3593 isl_pw_aff_free(cond);
3595 return cond != NULL;
3598 /* Check if we can extract a condition from "expr".
3599 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
3600 * If allow_nested is set, then the condition may involve parameters
3601 * corresponding to nested accesses.
3602 * We turn on autodetection so that we won't generate any warnings.
3604 __isl_give isl_pw_aff *PetScan::try_extract_nested_condition(Expr *expr)
3606 isl_pw_aff *cond;
3607 int save_autodetect = options->autodetect;
3608 bool save_nesting = nesting_enabled;
3610 options->autodetect = 1;
3611 nesting_enabled = allow_nested;
3612 cond = extract_condition(expr);
3614 options->autodetect = save_autodetect;
3615 nesting_enabled = save_nesting;
3617 return cond;
3620 /* If the top-level expression of "stmt" is an assignment, then
3621 * return that assignment as a BinaryOperator.
3622 * Otherwise return NULL.
3624 static BinaryOperator *top_assignment_or_null(Stmt *stmt)
3626 BinaryOperator *ass;
3628 if (!stmt)
3629 return NULL;
3630 if (stmt->getStmtClass() != Stmt::BinaryOperatorClass)
3631 return NULL;
3633 ass = cast<BinaryOperator>(stmt);
3634 if(ass->getOpcode() != BO_Assign)
3635 return NULL;
3637 return ass;
3640 /* Check if the given if statement is a conditional assignement
3641 * with a non-affine condition. If so, construct a pet_scop
3642 * corresponding to this conditional assignment. Otherwise return NULL.
3644 * In particular we check if "stmt" is of the form
3646 * if (condition)
3647 * a = f(...);
3648 * else
3649 * a = g(...);
3651 * where a is some array or scalar access.
3652 * The constructed pet_scop then corresponds to the expression
3654 * a = condition ? f(...) : g(...)
3656 * All access relations in f(...) are intersected with condition
3657 * while all access relation in g(...) are intersected with the complement.
3659 struct pet_scop *PetScan::extract_conditional_assignment(IfStmt *stmt)
3661 BinaryOperator *ass_then, *ass_else;
3662 isl_multi_pw_aff *write_then, *write_else;
3663 isl_set *cond, *comp;
3664 isl_multi_pw_aff *index;
3665 isl_pw_aff *pa;
3666 int equal;
3667 struct pet_expr *pe_cond, *pe_then, *pe_else, *pe, *pe_write;
3668 bool save_nesting = nesting_enabled;
3670 if (!options->detect_conditional_assignment)
3671 return NULL;
3673 ass_then = top_assignment_or_null(stmt->getThen());
3674 ass_else = top_assignment_or_null(stmt->getElse());
3676 if (!ass_then || !ass_else)
3677 return NULL;
3679 if (is_affine_condition(stmt->getCond()))
3680 return NULL;
3682 write_then = extract_index(ass_then->getLHS());
3683 write_else = extract_index(ass_else->getLHS());
3685 equal = isl_multi_pw_aff_plain_is_equal(write_then, write_else);
3686 isl_multi_pw_aff_free(write_else);
3687 if (equal < 0 || !equal) {
3688 isl_multi_pw_aff_free(write_then);
3689 return NULL;
3692 nesting_enabled = allow_nested;
3693 pa = extract_condition(stmt->getCond());
3694 nesting_enabled = save_nesting;
3695 cond = isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa));
3696 comp = isl_pw_aff_zero_set(isl_pw_aff_copy(pa));
3697 index = isl_multi_pw_aff_from_range(isl_multi_pw_aff_from_pw_aff(pa));
3699 pe_cond = pet_expr_from_index(index);
3701 pe_then = extract_expr(ass_then->getRHS());
3702 pe_then = pet_expr_restrict(pe_then, cond);
3703 pe_else = extract_expr(ass_else->getRHS());
3704 pe_else = pet_expr_restrict(pe_else, comp);
3706 pe = pet_expr_new_ternary(ctx, pe_cond, pe_then, pe_else);
3707 pe_write = pet_expr_from_index_and_depth(write_then,
3708 extract_depth(write_then));
3709 if (pe_write) {
3710 pe_write->acc.write = 1;
3711 pe_write->acc.read = 0;
3713 pe = pet_expr_new_binary(ctx, pet_op_assign, pe_write, pe);
3714 return extract(stmt, pe);
3717 /* Create a pet_scop with a single statement evaluating "cond"
3718 * and writing the result to a virtual scalar, as expressed by
3719 * "index".
3721 struct pet_scop *PetScan::extract_non_affine_condition(Expr *cond,
3722 __isl_take isl_multi_pw_aff *index)
3724 struct pet_expr *expr, *write;
3725 struct pet_stmt *ps;
3726 struct pet_scop *scop;
3727 SourceLocation loc = cond->getLocStart();
3728 int line = PP.getSourceManager().getExpansionLineNumber(loc);
3730 write = pet_expr_from_index(index);
3731 if (write) {
3732 write->acc.write = 1;
3733 write->acc.read = 0;
3735 expr = extract_expr(cond);
3736 expr = resolve_nested(expr);
3737 expr = pet_expr_new_binary(ctx, pet_op_assign, write, expr);
3738 ps = pet_stmt_from_pet_expr(ctx, line, NULL, n_stmt++, expr);
3739 scop = pet_scop_from_pet_stmt(ctx, ps);
3740 scop = resolve_nested(scop);
3742 return scop;
3745 extern "C" {
3746 static struct pet_expr *embed_access(struct pet_expr *expr, void *user);
3749 /* Precompose the access relation and the index expression associated
3750 * to "expr" with the function pointed to by "user",
3751 * thereby embedding the access relation in the domain of this function.
3752 * The initial domain of the access relation and the index expression
3753 * is the zero-dimensional domain.
3755 static struct pet_expr *embed_access(struct pet_expr *expr, void *user)
3757 isl_multi_aff *ma = (isl_multi_aff *) user;
3759 expr->acc.access = isl_map_preimage_domain_multi_aff(expr->acc.access,
3760 isl_multi_aff_copy(ma));
3761 expr->acc.index = isl_multi_pw_aff_pullback_multi_aff(expr->acc.index,
3762 isl_multi_aff_copy(ma));
3763 if (!expr->acc.access || !expr->acc.index)
3764 goto error;
3766 return expr;
3767 error:
3768 pet_expr_free(expr);
3769 return NULL;
3772 /* Precompose all access relations in "expr" with "ma", thereby
3773 * embedding them in the domain of "ma".
3775 static struct pet_expr *embed(struct pet_expr *expr,
3776 __isl_keep isl_multi_aff *ma)
3778 return pet_expr_map_access(expr, &embed_access, ma);
3781 /* How many parameters of "set" refer to nested accesses, i.e., have no name?
3783 static int n_nested_parameter(__isl_keep isl_set *set)
3785 isl_space *space;
3786 int n;
3788 space = isl_set_get_space(set);
3789 n = n_nested_parameter(space);
3790 isl_space_free(space);
3792 return n;
3795 /* Remove all parameters from "map" that refer to nested accesses.
3797 static __isl_give isl_map *remove_nested_parameters(__isl_take isl_map *map)
3799 int nparam;
3800 isl_space *space;
3802 space = isl_map_get_space(map);
3803 nparam = isl_space_dim(space, isl_dim_param);
3804 for (int i = nparam - 1; i >= 0; --i)
3805 if (is_nested_parameter(space, i))
3806 map = isl_map_project_out(map, isl_dim_param, i, 1);
3807 isl_space_free(space);
3809 return map;
3812 /* Remove all parameters from "mpa" that refer to nested accesses.
3814 static __isl_give isl_multi_pw_aff *remove_nested_parameters(
3815 __isl_take isl_multi_pw_aff *mpa)
3817 int nparam;
3818 isl_space *space;
3820 space = isl_multi_pw_aff_get_space(mpa);
3821 nparam = isl_space_dim(space, isl_dim_param);
3822 for (int i = nparam - 1; i >= 0; --i) {
3823 if (!is_nested_parameter(space, i))
3824 continue;
3825 mpa = isl_multi_pw_aff_drop_dims(mpa, isl_dim_param, i, 1);
3827 isl_space_free(space);
3829 return mpa;
3832 /* Remove all parameters from the index expression and access relation of "expr"
3833 * that refer to nested accesses.
3835 static struct pet_expr *remove_nested_parameters(struct pet_expr *expr)
3837 expr->acc.access = remove_nested_parameters(expr->acc.access);
3838 expr->acc.index = remove_nested_parameters(expr->acc.index);
3839 if (!expr->acc.access || !expr->acc.index)
3840 goto error;
3842 return expr;
3843 error:
3844 pet_expr_free(expr);
3845 return NULL;
3848 extern "C" {
3849 static struct pet_expr *expr_remove_nested_parameters(
3850 struct pet_expr *expr, void *user);
3853 static struct pet_expr *expr_remove_nested_parameters(
3854 struct pet_expr *expr, void *user)
3856 return remove_nested_parameters(expr);
3859 /* Remove all nested access parameters from the schedule and all
3860 * accesses of "stmt".
3861 * There is no need to remove them from the domain as these parameters
3862 * have already been removed from the domain when this function is called.
3864 static struct pet_stmt *remove_nested_parameters(struct pet_stmt *stmt)
3866 if (!stmt)
3867 return NULL;
3868 stmt->schedule = remove_nested_parameters(stmt->schedule);
3869 stmt->body = pet_expr_map_access(stmt->body,
3870 &expr_remove_nested_parameters, NULL);
3871 if (!stmt->schedule || !stmt->body)
3872 goto error;
3873 for (int i = 0; i < stmt->n_arg; ++i) {
3874 stmt->args[i] = pet_expr_map_access(stmt->args[i],
3875 &expr_remove_nested_parameters, NULL);
3876 if (!stmt->args[i])
3877 goto error;
3880 return stmt;
3881 error:
3882 pet_stmt_free(stmt);
3883 return NULL;
3886 /* For each nested access parameter in the domain of "stmt",
3887 * construct a corresponding pet_expr, place it before the original
3888 * elements in stmt->args and record its position in "param2pos".
3889 * n is the number of nested access parameters.
3891 struct pet_stmt *PetScan::extract_nested(struct pet_stmt *stmt, int n,
3892 std::map<int,int> &param2pos)
3894 int i;
3895 isl_space *space;
3896 int n_arg;
3897 struct pet_expr **args;
3899 n_arg = stmt->n_arg;
3900 args = isl_calloc_array(ctx, struct pet_expr *, n + n_arg);
3901 if (!args)
3902 goto error;
3904 space = isl_set_get_space(stmt->domain);
3905 n_arg = extract_nested(space, 0, args, param2pos);
3906 isl_space_free(space);
3908 if (n_arg < 0)
3909 goto error;
3911 for (i = 0; i < stmt->n_arg; ++i)
3912 args[n_arg + i] = stmt->args[i];
3913 free(stmt->args);
3914 stmt->args = args;
3915 stmt->n_arg += n_arg;
3917 return stmt;
3918 error:
3919 if (args) {
3920 for (i = 0; i < n; ++i)
3921 pet_expr_free(args[i]);
3922 free(args);
3924 pet_stmt_free(stmt);
3925 return NULL;
3928 /* Check whether any of the arguments i of "stmt" starting at position "n"
3929 * is equal to one of the first "n" arguments j.
3930 * If so, combine the constraints on arguments i and j and remove
3931 * argument i.
3933 static struct pet_stmt *remove_duplicate_arguments(struct pet_stmt *stmt, int n)
3935 int i, j;
3936 isl_map *map;
3938 if (!stmt)
3939 return NULL;
3940 if (n == 0)
3941 return stmt;
3942 if (n == stmt->n_arg)
3943 return stmt;
3945 map = isl_set_unwrap(stmt->domain);
3947 for (i = stmt->n_arg - 1; i >= n; --i) {
3948 for (j = 0; j < n; ++j)
3949 if (pet_expr_is_equal(stmt->args[i], stmt->args[j]))
3950 break;
3951 if (j >= n)
3952 continue;
3954 map = isl_map_equate(map, isl_dim_out, i, isl_dim_out, j);
3955 map = isl_map_project_out(map, isl_dim_out, i, 1);
3957 pet_expr_free(stmt->args[i]);
3958 for (j = i; j + 1 < stmt->n_arg; ++j)
3959 stmt->args[j] = stmt->args[j + 1];
3960 stmt->n_arg--;
3963 stmt->domain = isl_map_wrap(map);
3964 if (!stmt->domain)
3965 goto error;
3966 return stmt;
3967 error:
3968 pet_stmt_free(stmt);
3969 return NULL;
3972 /* Look for parameters in the iteration domain of "stmt" that
3973 * refer to nested accesses. In particular, these are
3974 * parameters with no name.
3976 * If there are any such parameters, then as many extra variables
3977 * (after identifying identical nested accesses) are inserted in the
3978 * range of the map wrapped inside the domain, before the original variables.
3979 * If the original domain is not a wrapped map, then a new wrapped
3980 * map is created with zero output dimensions.
3981 * The parameters are then equated to the corresponding output dimensions
3982 * and subsequently projected out, from the iteration domain,
3983 * the schedule and the access relations.
3984 * For each of the output dimensions, a corresponding argument
3985 * expression is inserted. Initially they are created with
3986 * a zero-dimensional domain, so they have to be embedded
3987 * in the current iteration domain.
3988 * param2pos maps the position of the parameter to the position
3989 * of the corresponding output dimension in the wrapped map.
3991 struct pet_stmt *PetScan::resolve_nested(struct pet_stmt *stmt)
3993 int n;
3994 int nparam;
3995 unsigned n_arg;
3996 isl_map *map;
3997 isl_space *space;
3998 isl_multi_aff *ma;
3999 std::map<int,int> param2pos;
4001 if (!stmt)
4002 return NULL;
4004 n = n_nested_parameter(stmt->domain);
4005 if (n == 0)
4006 return stmt;
4008 n_arg = stmt->n_arg;
4009 stmt = extract_nested(stmt, n, param2pos);
4010 if (!stmt)
4011 return NULL;
4013 n = stmt->n_arg - n_arg;
4014 nparam = isl_set_dim(stmt->domain, isl_dim_param);
4015 if (isl_set_is_wrapping(stmt->domain))
4016 map = isl_set_unwrap(stmt->domain);
4017 else
4018 map = isl_map_from_domain(stmt->domain);
4019 map = isl_map_insert_dims(map, isl_dim_out, 0, n);
4021 for (int i = nparam - 1; i >= 0; --i) {
4022 isl_id *id;
4024 if (!is_nested_parameter(map, i))
4025 continue;
4027 id = pet_expr_access_get_id(stmt->args[param2pos[i]]);
4028 map = isl_map_set_dim_id(map, isl_dim_out, param2pos[i], id);
4029 map = isl_map_equate(map, isl_dim_param, i, isl_dim_out,
4030 param2pos[i]);
4031 map = isl_map_project_out(map, isl_dim_param, i, 1);
4034 stmt->domain = isl_map_wrap(map);
4036 space = isl_space_unwrap(isl_set_get_space(stmt->domain));
4037 space = isl_space_from_domain(isl_space_domain(space));
4038 ma = isl_multi_aff_zero(space);
4039 for (int pos = 0; pos < n; ++pos)
4040 stmt->args[pos] = embed(stmt->args[pos], ma);
4041 isl_multi_aff_free(ma);
4043 stmt = remove_nested_parameters(stmt);
4044 stmt = remove_duplicate_arguments(stmt, n);
4046 return stmt;
4049 /* For each statement in "scop", move the parameters that correspond
4050 * to nested access into the ranges of the domains and create
4051 * corresponding argument expressions.
4053 struct pet_scop *PetScan::resolve_nested(struct pet_scop *scop)
4055 if (!scop)
4056 return NULL;
4058 for (int i = 0; i < scop->n_stmt; ++i) {
4059 scop->stmts[i] = resolve_nested(scop->stmts[i]);
4060 if (!scop->stmts[i])
4061 goto error;
4064 return scop;
4065 error:
4066 pet_scop_free(scop);
4067 return NULL;
4070 /* Given an access expression "expr", is the variable accessed by
4071 * "expr" assigned anywhere inside "scop"?
4073 static bool is_assigned(pet_expr *expr, pet_scop *scop)
4075 bool assigned = false;
4076 isl_id *id;
4078 id = pet_expr_access_get_id(expr);
4079 assigned = pet_scop_writes(scop, id);
4080 isl_id_free(id);
4082 return assigned;
4085 /* Are all nested access parameters in "pa" allowed given "scop".
4086 * In particular, is none of them written by anywhere inside "scop".
4088 * If "scop" has any skip conditions, then no nested access parameters
4089 * are allowed. In particular, if there is any nested access in a guard
4090 * for a piece of code containing a "continue", then we want to introduce
4091 * a separate statement for evaluating this guard so that we can express
4092 * that the result is false for all previous iterations.
4094 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff *pa, pet_scop *scop)
4096 int nparam;
4098 if (!scop)
4099 return true;
4101 nparam = isl_pw_aff_dim(pa, isl_dim_param);
4102 for (int i = 0; i < nparam; ++i) {
4103 Expr *nested;
4104 isl_id *id = isl_pw_aff_get_dim_id(pa, isl_dim_param, i);
4105 pet_expr *expr;
4106 bool allowed;
4108 if (!is_nested_parameter(id)) {
4109 isl_id_free(id);
4110 continue;
4113 if (pet_scop_has_skip(scop, pet_skip_now)) {
4114 isl_id_free(id);
4115 return false;
4118 nested = (Expr *) isl_id_get_user(id);
4119 expr = extract_expr(nested);
4120 allowed = expr && expr->type == pet_expr_access &&
4121 !is_assigned(expr, scop);
4123 pet_expr_free(expr);
4124 isl_id_free(id);
4126 if (!allowed)
4127 return false;
4130 return true;
4133 /* Do we need to construct a skip condition of the given type
4134 * on an if statement, given that the if condition is non-affine?
4136 * pet_scop_filter_skip can only handle the case where the if condition
4137 * holds (the then branch) and the skip condition is universal.
4138 * In any other case, we need to construct a new skip condition.
4140 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
4141 bool have_else, enum pet_skip type)
4143 if (have_else && scop_else && pet_scop_has_skip(scop_else, type))
4144 return true;
4145 if (scop_then && pet_scop_has_skip(scop_then, type) &&
4146 !pet_scop_has_universal_skip(scop_then, type))
4147 return true;
4148 return false;
4151 /* Do we need to construct a skip condition of the given type
4152 * on an if statement, given that the if condition is affine?
4154 * There is no need to construct a new skip condition if all
4155 * the skip conditions are affine.
4157 static bool need_skip_aff(struct pet_scop *scop_then,
4158 struct pet_scop *scop_else, bool have_else, enum pet_skip type)
4160 if (scop_then && pet_scop_has_var_skip(scop_then, type))
4161 return true;
4162 if (have_else && scop_else && pet_scop_has_var_skip(scop_else, type))
4163 return true;
4164 return false;
4167 /* Do we need to construct a skip condition of the given type
4168 * on an if statement?
4170 static bool need_skip(struct pet_scop *scop_then, struct pet_scop *scop_else,
4171 bool have_else, enum pet_skip type, bool affine)
4173 if (affine)
4174 return need_skip_aff(scop_then, scop_else, have_else, type);
4175 else
4176 return need_skip(scop_then, scop_else, have_else, type);
4179 /* Construct an affine expression pet_expr that evaluates
4180 * to the constant "val".
4182 static struct pet_expr *universally(isl_ctx *ctx, int val)
4184 isl_local_space *ls;
4185 isl_aff *aff;
4186 isl_multi_pw_aff *mpa;
4188 ls = isl_local_space_from_space(isl_space_set_alloc(ctx, 0, 0));
4189 aff = isl_aff_val_on_domain(ls, isl_val_int_from_si(ctx, val));
4190 mpa = isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
4192 return pet_expr_from_index(mpa);
4195 /* Construct an affine expression pet_expr that evaluates
4196 * to the constant 1.
4198 static struct pet_expr *universally_true(isl_ctx *ctx)
4200 return universally(ctx, 1);
4203 /* Construct an affine expression pet_expr that evaluates
4204 * to the constant 0.
4206 static struct pet_expr *universally_false(isl_ctx *ctx)
4208 return universally(ctx, 0);
4211 /* Given an index expression "test_index" for the if condition,
4212 * an index expression "skip_index" for the skip condition and
4213 * scops for the then and else branches, construct a scop for
4214 * computing "skip_index".
4216 * The computed scop contains a single statement that essentially does
4218 * skip_index = test_cond ? skip_cond_then : skip_cond_else
4220 * If the skip conditions of the then and/or else branch are not affine,
4221 * then they need to be filtered by test_index.
4222 * If they are missing, then this means the skip condition is false.
4224 * Since we are constructing a skip condition for the if statement,
4225 * the skip conditions on the then and else branches are removed.
4227 static struct pet_scop *extract_skip(PetScan *scan,
4228 __isl_take isl_multi_pw_aff *test_index,
4229 __isl_take isl_multi_pw_aff *skip_index,
4230 struct pet_scop *scop_then, struct pet_scop *scop_else, bool have_else,
4231 enum pet_skip type)
4233 struct pet_expr *expr_then, *expr_else, *expr, *expr_skip;
4234 struct pet_stmt *stmt;
4235 struct pet_scop *scop;
4236 isl_ctx *ctx = scan->ctx;
4238 if (!scop_then)
4239 goto error;
4240 if (have_else && !scop_else)
4241 goto error;
4243 if (pet_scop_has_skip(scop_then, type)) {
4244 expr_then = pet_scop_get_skip_expr(scop_then, type);
4245 pet_scop_reset_skip(scop_then, type);
4246 if (!pet_expr_is_affine(expr_then))
4247 expr_then = pet_expr_filter(expr_then,
4248 isl_multi_pw_aff_copy(test_index), 1);
4249 } else
4250 expr_then = universally_false(ctx);
4252 if (have_else && pet_scop_has_skip(scop_else, type)) {
4253 expr_else = pet_scop_get_skip_expr(scop_else, type);
4254 pet_scop_reset_skip(scop_else, type);
4255 if (!pet_expr_is_affine(expr_else))
4256 expr_else = pet_expr_filter(expr_else,
4257 isl_multi_pw_aff_copy(test_index), 0);
4258 } else
4259 expr_else = universally_false(ctx);
4261 expr = pet_expr_from_index(test_index);
4262 expr = pet_expr_new_ternary(ctx, expr, expr_then, expr_else);
4263 expr_skip = pet_expr_from_index(isl_multi_pw_aff_copy(skip_index));
4264 if (expr_skip) {
4265 expr_skip->acc.write = 1;
4266 expr_skip->acc.read = 0;
4268 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
4269 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, scan->n_stmt++, expr);
4271 scop = pet_scop_from_pet_stmt(ctx, stmt);
4272 scop = scop_add_array(scop, skip_index, scan->ast_context);
4273 isl_multi_pw_aff_free(skip_index);
4275 return scop;
4276 error:
4277 isl_multi_pw_aff_free(test_index);
4278 isl_multi_pw_aff_free(skip_index);
4279 return NULL;
4282 /* Is scop's skip_now condition equal to its skip_later condition?
4283 * In particular, this means that it either has no skip_now condition
4284 * or both a skip_now and a skip_later condition (that are equal to each other).
4286 static bool skip_equals_skip_later(struct pet_scop *scop)
4288 int has_skip_now, has_skip_later;
4289 int equal;
4290 isl_multi_pw_aff *skip_now, *skip_later;
4292 if (!scop)
4293 return false;
4294 has_skip_now = pet_scop_has_skip(scop, pet_skip_now);
4295 has_skip_later = pet_scop_has_skip(scop, pet_skip_later);
4296 if (has_skip_now != has_skip_later)
4297 return false;
4298 if (!has_skip_now)
4299 return true;
4301 skip_now = pet_scop_get_skip(scop, pet_skip_now);
4302 skip_later = pet_scop_get_skip(scop, pet_skip_later);
4303 equal = isl_multi_pw_aff_is_equal(skip_now, skip_later);
4304 isl_multi_pw_aff_free(skip_now);
4305 isl_multi_pw_aff_free(skip_later);
4307 return equal;
4310 /* Drop the skip conditions of type pet_skip_later from scop1 and scop2.
4312 static void drop_skip_later(struct pet_scop *scop1, struct pet_scop *scop2)
4314 pet_scop_reset_skip(scop1, pet_skip_later);
4315 pet_scop_reset_skip(scop2, pet_skip_later);
4318 /* Structure that handles the construction of skip conditions.
4320 * scop_then and scop_else represent the then and else branches
4321 * of the if statement
4323 * skip[type] is true if we need to construct a skip condition of that type
4324 * equal is set if the skip conditions of types pet_skip_now and pet_skip_later
4325 * are equal to each other
4326 * index[type] is an index expression from a zero-dimension domain
4327 * to the virtual array representing the skip condition
4328 * scop[type] is a scop for computing the skip condition
4330 struct pet_skip_info {
4331 isl_ctx *ctx;
4333 bool skip[2];
4334 bool equal;
4335 isl_multi_pw_aff *index[2];
4336 struct pet_scop *scop[2];
4338 pet_skip_info(isl_ctx *ctx) : ctx(ctx) {}
4340 operator bool() { return skip[pet_skip_now] || skip[pet_skip_later]; }
4343 /* Structure that handles the construction of skip conditions on if statements.
4345 * scop_then and scop_else represent the then and else branches
4346 * of the if statement
4348 struct pet_skip_info_if : public pet_skip_info {
4349 struct pet_scop *scop_then, *scop_else;
4350 bool have_else;
4352 pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
4353 struct pet_scop *scop_else, bool have_else, bool affine);
4354 void extract(PetScan *scan, __isl_keep isl_multi_pw_aff *index,
4355 enum pet_skip type);
4356 void extract(PetScan *scan, __isl_keep isl_multi_pw_aff *index);
4357 void extract(PetScan *scan, __isl_keep isl_pw_aff *cond);
4358 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
4359 int offset);
4360 struct pet_scop *add(struct pet_scop *scop, int offset);
4363 /* Initialize a pet_skip_info_if structure based on the then and else branches
4364 * and based on whether the if condition is affine or not.
4366 pet_skip_info_if::pet_skip_info_if(isl_ctx *ctx, struct pet_scop *scop_then,
4367 struct pet_scop *scop_else, bool have_else, bool affine) :
4368 pet_skip_info(ctx), scop_then(scop_then), scop_else(scop_else),
4369 have_else(have_else)
4371 skip[pet_skip_now] =
4372 need_skip(scop_then, scop_else, have_else, pet_skip_now, affine);
4373 equal = skip[pet_skip_now] && skip_equals_skip_later(scop_then) &&
4374 (!have_else || skip_equals_skip_later(scop_else));
4375 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
4376 need_skip(scop_then, scop_else, have_else, pet_skip_later, affine);
4379 /* If we need to construct a skip condition of the given type,
4380 * then do so now.
4382 * "mpa" represents the if condition.
4384 void pet_skip_info_if::extract(PetScan *scan,
4385 __isl_keep isl_multi_pw_aff *mpa, enum pet_skip type)
4387 isl_ctx *ctx;
4389 if (!skip[type])
4390 return;
4392 ctx = isl_multi_pw_aff_get_ctx(mpa);
4393 index[type] = create_test_index(ctx, scan->n_test++);
4394 scop[type] = extract_skip(scan, isl_multi_pw_aff_copy(mpa),
4395 isl_multi_pw_aff_copy(index[type]),
4396 scop_then, scop_else, have_else, type);
4399 /* Construct the required skip conditions, given the if condition "index".
4401 void pet_skip_info_if::extract(PetScan *scan,
4402 __isl_keep isl_multi_pw_aff *index)
4404 extract(scan, index, pet_skip_now);
4405 extract(scan, index, pet_skip_later);
4406 if (equal)
4407 drop_skip_later(scop_then, scop_else);
4410 /* Construct the required skip conditions, given the if condition "cond".
4412 void pet_skip_info_if::extract(PetScan *scan, __isl_keep isl_pw_aff *cond)
4414 isl_multi_pw_aff *test;
4416 if (!skip[pet_skip_now] && !skip[pet_skip_later])
4417 return;
4419 test = isl_multi_pw_aff_from_pw_aff(isl_pw_aff_copy(cond));
4420 test = isl_multi_pw_aff_from_range(test);
4421 extract(scan, test);
4422 isl_multi_pw_aff_free(test);
4425 /* Add the computed skip condition of the give type to "main" and
4426 * add the scop for computing the condition at the given offset.
4428 * If equal is set, then we only computed a skip condition for pet_skip_now,
4429 * but we also need to set it as main's pet_skip_later.
4431 struct pet_scop *pet_skip_info_if::add(struct pet_scop *main,
4432 enum pet_skip type, int offset)
4434 if (!skip[type])
4435 return main;
4437 scop[type] = pet_scop_prefix(scop[type], offset);
4438 main = pet_scop_add_par(ctx, main, scop[type]);
4439 scop[type] = NULL;
4441 if (equal)
4442 main = pet_scop_set_skip(main, pet_skip_later,
4443 isl_multi_pw_aff_copy(index[type]));
4445 main = pet_scop_set_skip(main, type, index[type]);
4446 index[type] = NULL;
4448 return main;
4451 /* Add the computed skip conditions to "main" and
4452 * add the scops for computing the conditions at the given offset.
4454 struct pet_scop *pet_skip_info_if::add(struct pet_scop *scop, int offset)
4456 scop = add(scop, pet_skip_now, offset);
4457 scop = add(scop, pet_skip_later, offset);
4459 return scop;
4462 /* Construct a pet_scop for a non-affine if statement.
4464 * We create a separate statement that writes the result
4465 * of the non-affine condition to a virtual scalar.
4466 * A constraint requiring the value of this virtual scalar to be one
4467 * is added to the iteration domains of the then branch.
4468 * Similarly, a constraint requiring the value of this virtual scalar
4469 * to be zero is added to the iteration domains of the else branch, if any.
4470 * We adjust the schedules to ensure that the virtual scalar is written
4471 * before it is read.
4473 * If there are any breaks or continues in the then and/or else
4474 * branches, then we may have to compute a new skip condition.
4475 * This is handled using a pet_skip_info_if object.
4476 * On initialization, the object checks if skip conditions need
4477 * to be computed. If so, it does so in "extract" and adds them in "add".
4479 struct pet_scop *PetScan::extract_non_affine_if(Expr *cond,
4480 struct pet_scop *scop_then, struct pet_scop *scop_else,
4481 bool have_else, int stmt_id)
4483 struct pet_scop *scop;
4484 isl_multi_pw_aff *test_index;
4485 int save_n_stmt = n_stmt;
4487 test_index = create_test_index(ctx, n_test++);
4488 n_stmt = stmt_id;
4489 scop = extract_non_affine_condition(cond,
4490 isl_multi_pw_aff_copy(test_index));
4491 n_stmt = save_n_stmt;
4492 scop = scop_add_array(scop, test_index, ast_context);
4494 pet_skip_info_if skip(ctx, scop_then, scop_else, have_else, false);
4495 skip.extract(this, test_index);
4497 scop = pet_scop_prefix(scop, 0);
4498 scop_then = pet_scop_prefix(scop_then, 1);
4499 scop_then = pet_scop_filter(scop_then,
4500 isl_multi_pw_aff_copy(test_index), 1);
4501 if (have_else) {
4502 scop_else = pet_scop_prefix(scop_else, 1);
4503 scop_else = pet_scop_filter(scop_else, test_index, 0);
4504 scop_then = pet_scop_add_par(ctx, scop_then, scop_else);
4505 } else
4506 isl_multi_pw_aff_free(test_index);
4508 scop = pet_scop_add_seq(ctx, scop, scop_then);
4510 scop = skip.add(scop, 2);
4512 return scop;
4515 /* Construct a pet_scop for an if statement.
4517 * If the condition fits the pattern of a conditional assignment,
4518 * then it is handled by extract_conditional_assignment.
4519 * Otherwise, we do the following.
4521 * If the condition is affine, then the condition is added
4522 * to the iteration domains of the then branch, while the
4523 * opposite of the condition in added to the iteration domains
4524 * of the else branch, if any.
4525 * We allow the condition to be dynamic, i.e., to refer to
4526 * scalars or array elements that may be written to outside
4527 * of the given if statement. These nested accesses are then represented
4528 * as output dimensions in the wrapping iteration domain.
4529 * If it also written _inside_ the then or else branch, then
4530 * we treat the condition as non-affine.
4531 * As explained in extract_non_affine_if, this will introduce
4532 * an extra statement.
4533 * For aesthetic reasons, we want this statement to have a statement
4534 * number that is lower than those of the then and else branches.
4535 * In order to evaluate if will need such a statement, however, we
4536 * first construct scops for the then and else branches.
4537 * We therefore reserve a statement number if we might have to
4538 * introduce such an extra statement.
4540 * If the condition is not affine, then the scop is created in
4541 * extract_non_affine_if.
4543 * If there are any breaks or continues in the then and/or else
4544 * branches, then we may have to compute a new skip condition.
4545 * This is handled using a pet_skip_info_if object.
4546 * On initialization, the object checks if skip conditions need
4547 * to be computed. If so, it does so in "extract" and adds them in "add".
4549 struct pet_scop *PetScan::extract(IfStmt *stmt)
4551 struct pet_scop *scop_then, *scop_else = NULL, *scop;
4552 isl_pw_aff *cond;
4553 int stmt_id;
4554 isl_set *set;
4555 isl_set *valid;
4557 scop = extract_conditional_assignment(stmt);
4558 if (scop)
4559 return scop;
4561 cond = try_extract_nested_condition(stmt->getCond());
4562 if (allow_nested && (!cond || has_nested(cond)))
4563 stmt_id = n_stmt++;
4566 assigned_value_cache cache(assigned_value);
4567 scop_then = extract(stmt->getThen());
4570 if (stmt->getElse()) {
4571 assigned_value_cache cache(assigned_value);
4572 scop_else = extract(stmt->getElse());
4573 if (options->autodetect) {
4574 if (scop_then && !scop_else) {
4575 partial = true;
4576 isl_pw_aff_free(cond);
4577 return scop_then;
4579 if (!scop_then && scop_else) {
4580 partial = true;
4581 isl_pw_aff_free(cond);
4582 return scop_else;
4587 if (cond &&
4588 (!is_nested_allowed(cond, scop_then) ||
4589 (stmt->getElse() && !is_nested_allowed(cond, scop_else)))) {
4590 isl_pw_aff_free(cond);
4591 cond = NULL;
4593 if (allow_nested && !cond)
4594 return extract_non_affine_if(stmt->getCond(), scop_then,
4595 scop_else, stmt->getElse(), stmt_id);
4597 if (!cond)
4598 cond = extract_condition(stmt->getCond());
4600 pet_skip_info_if skip(ctx, scop_then, scop_else, stmt->getElse(), true);
4601 skip.extract(this, cond);
4603 valid = isl_pw_aff_domain(isl_pw_aff_copy(cond));
4604 set = isl_pw_aff_non_zero_set(cond);
4605 scop = pet_scop_restrict(scop_then, isl_set_copy(set));
4607 if (stmt->getElse()) {
4608 set = isl_set_subtract(isl_set_copy(valid), set);
4609 scop_else = pet_scop_restrict(scop_else, set);
4610 scop = pet_scop_add_par(ctx, scop, scop_else);
4611 } else
4612 isl_set_free(set);
4613 scop = resolve_nested(scop);
4614 scop = pet_scop_restrict_context(scop, valid);
4616 if (skip)
4617 scop = pet_scop_prefix(scop, 0);
4618 scop = skip.add(scop, 1);
4620 return scop;
4623 /* Try and construct a pet_scop for a label statement.
4624 * We currently only allow labels on expression statements.
4626 struct pet_scop *PetScan::extract(LabelStmt *stmt)
4628 isl_id *label;
4629 Stmt *sub;
4631 sub = stmt->getSubStmt();
4632 if (!isa<Expr>(sub)) {
4633 unsupported(stmt);
4634 return NULL;
4637 label = isl_id_alloc(ctx, stmt->getName(), NULL);
4639 return extract(sub, extract_expr(cast<Expr>(sub)), label);
4642 /* Return a one-dimensional multi piecewise affine expression that is equal
4643 * to the constant 1 and is defined over a zero-dimensional domain.
4645 static __isl_give isl_multi_pw_aff *one_mpa(isl_ctx *ctx)
4647 isl_space *space;
4648 isl_local_space *ls;
4649 isl_aff *aff;
4651 space = isl_space_set_alloc(ctx, 0, 0);
4652 ls = isl_local_space_from_space(space);
4653 aff = isl_aff_zero_on_domain(ls);
4654 aff = isl_aff_set_constant_si(aff, 1);
4656 return isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff));
4659 /* Construct a pet_scop for a continue statement.
4661 * We simply create an empty scop with a universal pet_skip_now
4662 * skip condition. This skip condition will then be taken into
4663 * account by the enclosing loop construct, possibly after
4664 * being incorporated into outer skip conditions.
4666 struct pet_scop *PetScan::extract(ContinueStmt *stmt)
4668 pet_scop *scop;
4670 scop = pet_scop_empty(ctx);
4671 if (!scop)
4672 return NULL;
4674 scop = pet_scop_set_skip(scop, pet_skip_now, one_mpa(ctx));
4676 return scop;
4679 /* Construct a pet_scop for a break statement.
4681 * We simply create an empty scop with both a universal pet_skip_now
4682 * skip condition and a universal pet_skip_later skip condition.
4683 * These skip conditions will then be taken into
4684 * account by the enclosing loop construct, possibly after
4685 * being incorporated into outer skip conditions.
4687 struct pet_scop *PetScan::extract(BreakStmt *stmt)
4689 pet_scop *scop;
4690 isl_multi_pw_aff *skip;
4692 scop = pet_scop_empty(ctx);
4693 if (!scop)
4694 return NULL;
4696 skip = one_mpa(ctx);
4697 scop = pet_scop_set_skip(scop, pet_skip_now,
4698 isl_multi_pw_aff_copy(skip));
4699 scop = pet_scop_set_skip(scop, pet_skip_later, skip);
4701 return scop;
4704 /* Try and construct a pet_scop corresponding to "stmt".
4706 * If "stmt" is a compound statement, then "skip_declarations"
4707 * indicates whether we should skip initial declarations in the
4708 * compound statement.
4710 * If the constructed pet_scop is not a (possibly) partial representation
4711 * of "stmt", we update start and end of the pet_scop to those of "stmt".
4712 * In particular, if skip_declarations, then we may have skipped declarations
4713 * inside "stmt" and so the pet_scop may not represent the entire "stmt".
4714 * Note that this function may be called with "stmt" referring to the entire
4715 * body of the function, including the outer braces. In such cases,
4716 * skip_declarations will be set and the braces will not be taken into
4717 * account in scop->start and scop->end.
4719 struct pet_scop *PetScan::extract(Stmt *stmt, bool skip_declarations)
4721 struct pet_scop *scop;
4722 unsigned start, end;
4723 SourceLocation loc;
4724 SourceManager &SM = PP.getSourceManager();
4725 const LangOptions &LO = PP.getLangOpts();
4727 if (isa<Expr>(stmt))
4728 return extract(stmt, extract_expr(cast<Expr>(stmt)));
4730 switch (stmt->getStmtClass()) {
4731 case Stmt::WhileStmtClass:
4732 scop = extract(cast<WhileStmt>(stmt));
4733 break;
4734 case Stmt::ForStmtClass:
4735 scop = extract_for(cast<ForStmt>(stmt));
4736 break;
4737 case Stmt::IfStmtClass:
4738 scop = extract(cast<IfStmt>(stmt));
4739 break;
4740 case Stmt::CompoundStmtClass:
4741 scop = extract(cast<CompoundStmt>(stmt), skip_declarations);
4742 break;
4743 case Stmt::LabelStmtClass:
4744 scop = extract(cast<LabelStmt>(stmt));
4745 break;
4746 case Stmt::ContinueStmtClass:
4747 scop = extract(cast<ContinueStmt>(stmt));
4748 break;
4749 case Stmt::BreakStmtClass:
4750 scop = extract(cast<BreakStmt>(stmt));
4751 break;
4752 case Stmt::DeclStmtClass:
4753 scop = extract(cast<DeclStmt>(stmt));
4754 break;
4755 default:
4756 unsupported(stmt);
4757 return NULL;
4760 if (partial || skip_declarations)
4761 return scop;
4763 loc = stmt->getLocStart();
4764 loc = move_to_start_of_line_if_first_token(loc, SM, LO);
4765 start = getExpansionOffset(SM, loc);
4766 loc = PP.getLocForEndOfToken(stmt->getLocEnd());
4767 end = getExpansionOffset(SM, loc);
4768 scop = pet_scop_update_start_end(scop, start, end);
4770 return scop;
4773 /* Do we need to construct a skip condition of the given type
4774 * on a sequence of statements?
4776 * There is no need to construct a new skip condition if only
4777 * only of the two statements has a skip condition or if both
4778 * of their skip conditions are affine.
4780 * In principle we also don't need a new continuation variable if
4781 * the continuation of scop2 is affine, but then we would need
4782 * to allow more complicated forms of continuations.
4784 static bool need_skip_seq(struct pet_scop *scop1, struct pet_scop *scop2,
4785 enum pet_skip type)
4787 if (!scop1 || !pet_scop_has_skip(scop1, type))
4788 return false;
4789 if (!scop2 || !pet_scop_has_skip(scop2, type))
4790 return false;
4791 if (pet_scop_has_affine_skip(scop1, type) &&
4792 pet_scop_has_affine_skip(scop2, type))
4793 return false;
4794 return true;
4797 /* Construct a scop for computing the skip condition of the given type and
4798 * with index expression "skip_index" for a sequence of two scops "scop1"
4799 * and "scop2".
4801 * The computed scop contains a single statement that essentially does
4803 * skip_index = skip_cond_1 ? 1 : skip_cond_2
4805 * or, in other words, skip_cond1 || skip_cond2.
4806 * In this expression, skip_cond_2 is filtered to reflect that it is
4807 * only evaluated when skip_cond_1 is false.
4809 * The skip condition on scop1 is not removed because it still needs
4810 * to be applied to scop2 when these two scops are combined.
4812 static struct pet_scop *extract_skip_seq(PetScan *ps,
4813 __isl_take isl_multi_pw_aff *skip_index,
4814 struct pet_scop *scop1, struct pet_scop *scop2, enum pet_skip type)
4816 struct pet_expr *expr1, *expr2, *expr, *expr_skip;
4817 struct pet_stmt *stmt;
4818 struct pet_scop *scop;
4819 isl_ctx *ctx = ps->ctx;
4821 if (!scop1 || !scop2)
4822 goto error;
4824 expr1 = pet_scop_get_skip_expr(scop1, type);
4825 expr2 = pet_scop_get_skip_expr(scop2, type);
4826 pet_scop_reset_skip(scop2, type);
4828 expr2 = pet_expr_filter(expr2,
4829 isl_multi_pw_aff_copy(expr1->acc.index), 0);
4831 expr = universally_true(ctx);
4832 expr = pet_expr_new_ternary(ctx, expr1, expr, expr2);
4833 expr_skip = pet_expr_from_index(isl_multi_pw_aff_copy(skip_index));
4834 if (expr_skip) {
4835 expr_skip->acc.write = 1;
4836 expr_skip->acc.read = 0;
4838 expr = pet_expr_new_binary(ctx, pet_op_assign, expr_skip, expr);
4839 stmt = pet_stmt_from_pet_expr(ctx, -1, NULL, ps->n_stmt++, expr);
4841 scop = pet_scop_from_pet_stmt(ctx, stmt);
4842 scop = scop_add_array(scop, skip_index, ps->ast_context);
4843 isl_multi_pw_aff_free(skip_index);
4845 return scop;
4846 error:
4847 isl_multi_pw_aff_free(skip_index);
4848 return NULL;
4851 /* Structure that handles the construction of skip conditions
4852 * on sequences of statements.
4854 * scop1 and scop2 represent the two statements that are combined
4856 struct pet_skip_info_seq : public pet_skip_info {
4857 struct pet_scop *scop1, *scop2;
4859 pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4860 struct pet_scop *scop2);
4861 void extract(PetScan *scan, enum pet_skip type);
4862 void extract(PetScan *scan);
4863 struct pet_scop *add(struct pet_scop *scop, enum pet_skip type,
4864 int offset);
4865 struct pet_scop *add(struct pet_scop *scop, int offset);
4868 /* Initialize a pet_skip_info_seq structure based on
4869 * on the two statements that are going to be combined.
4871 pet_skip_info_seq::pet_skip_info_seq(isl_ctx *ctx, struct pet_scop *scop1,
4872 struct pet_scop *scop2) : pet_skip_info(ctx), scop1(scop1), scop2(scop2)
4874 skip[pet_skip_now] = need_skip_seq(scop1, scop2, pet_skip_now);
4875 equal = skip[pet_skip_now] && skip_equals_skip_later(scop1) &&
4876 skip_equals_skip_later(scop2);
4877 skip[pet_skip_later] = skip[pet_skip_now] && !equal &&
4878 need_skip_seq(scop1, scop2, pet_skip_later);
4881 /* If we need to construct a skip condition of the given type,
4882 * then do so now.
4884 void pet_skip_info_seq::extract(PetScan *scan, enum pet_skip type)
4886 if (!skip[type])
4887 return;
4889 index[type] = create_test_index(ctx, scan->n_test++);
4890 scop[type] = extract_skip_seq(scan, isl_multi_pw_aff_copy(index[type]),
4891 scop1, scop2, type);
4894 /* Construct the required skip conditions.
4896 void pet_skip_info_seq::extract(PetScan *scan)
4898 extract(scan, pet_skip_now);
4899 extract(scan, pet_skip_later);
4900 if (equal)
4901 drop_skip_later(scop1, scop2);
4904 /* Add the computed skip condition of the given type to "main" and
4905 * add the scop for computing the condition at the given offset (the statement
4906 * number). Within this offset, the condition is computed at position 1
4907 * to ensure that it is computed after the corresponding statement.
4909 * If equal is set, then we only computed a skip condition for pet_skip_now,
4910 * but we also need to set it as main's pet_skip_later.
4912 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *main,
4913 enum pet_skip type, int offset)
4915 if (!skip[type])
4916 return main;
4918 scop[type] = pet_scop_prefix(scop[type], 1);
4919 scop[type] = pet_scop_prefix(scop[type], offset);
4920 main = pet_scop_add_par(ctx, main, scop[type]);
4921 scop[type] = NULL;
4923 if (equal)
4924 main = pet_scop_set_skip(main, pet_skip_later,
4925 isl_multi_pw_aff_copy(index[type]));
4927 main = pet_scop_set_skip(main, type, index[type]);
4928 index[type] = NULL;
4930 return main;
4933 /* Add the computed skip conditions to "main" and
4934 * add the scops for computing the conditions at the given offset.
4936 struct pet_scop *pet_skip_info_seq::add(struct pet_scop *scop, int offset)
4938 scop = add(scop, pet_skip_now, offset);
4939 scop = add(scop, pet_skip_later, offset);
4941 return scop;
4944 /* Extract a clone of the kill statement in "scop".
4945 * "scop" is expected to have been created from a DeclStmt
4946 * and should have the kill as its first statement.
4948 struct pet_stmt *PetScan::extract_kill(struct pet_scop *scop)
4950 struct pet_expr *kill;
4951 struct pet_stmt *stmt;
4952 isl_multi_pw_aff *index;
4953 isl_map *access;
4955 if (!scop)
4956 return NULL;
4957 if (scop->n_stmt < 1)
4958 isl_die(ctx, isl_error_internal,
4959 "expecting at least one statement", return NULL);
4960 stmt = scop->stmts[0];
4961 if (stmt->body->type != pet_expr_unary ||
4962 stmt->body->op != pet_op_kill)
4963 isl_die(ctx, isl_error_internal,
4964 "expecting kill statement", return NULL);
4966 index = isl_multi_pw_aff_copy(stmt->body->args[0]->acc.index);
4967 access = isl_map_copy(stmt->body->args[0]->acc.access);
4968 index = isl_multi_pw_aff_reset_tuple_id(index, isl_dim_in);
4969 access = isl_map_reset_tuple_id(access, isl_dim_in);
4970 kill = pet_expr_kill_from_access_and_index(access, index);
4971 return pet_stmt_from_pet_expr(ctx, stmt->line, NULL, n_stmt++, kill);
4974 /* Mark all arrays in "scop" as being exposed.
4976 static struct pet_scop *mark_exposed(struct pet_scop *scop)
4978 if (!scop)
4979 return NULL;
4980 for (int i = 0; i < scop->n_array; ++i)
4981 scop->arrays[i]->exposed = 1;
4982 return scop;
4985 /* Try and construct a pet_scop corresponding to (part of)
4986 * a sequence of statements.
4988 * "block" is set if the sequence respresents the children of
4989 * a compound statement.
4990 * "skip_declarations" is set if we should skip initial declarations
4991 * in the sequence of statements.
4993 * If there are any breaks or continues in the individual statements,
4994 * then we may have to compute a new skip condition.
4995 * This is handled using a pet_skip_info_seq object.
4996 * On initialization, the object checks if skip conditions need
4997 * to be computed. If so, it does so in "extract" and adds them in "add".
4999 * If "block" is set, then we need to insert kill statements at
5000 * the end of the block for any array that has been declared by
5001 * one of the statements in the sequence. Each of these declarations
5002 * results in the construction of a kill statement at the place
5003 * of the declaration, so we simply collect duplicates of
5004 * those kill statements and append these duplicates to the constructed scop.
5006 * If "block" is not set, then any array declared by one of the statements
5007 * in the sequence is marked as being exposed.
5009 * If autodetect is set, then we allow the extraction of only a subrange
5010 * of the sequence of statements. However, if there is at least one statement
5011 * for which we could not construct a scop and the final range contains
5012 * either no statements or at least one kill, then we discard the entire
5013 * range.
5015 struct pet_scop *PetScan::extract(StmtRange stmt_range, bool block,
5016 bool skip_declarations)
5018 pet_scop *scop;
5019 StmtIterator i;
5020 int j;
5021 bool partial_range = false;
5022 set<struct pet_stmt *> kills;
5023 set<struct pet_stmt *>::iterator it;
5025 scop = pet_scop_empty(ctx);
5026 for (i = stmt_range.first, j = 0; i != stmt_range.second; ++i, ++j) {
5027 Stmt *child = *i;
5028 struct pet_scop *scop_i;
5030 if (skip_declarations &&
5031 child->getStmtClass() == Stmt::DeclStmtClass)
5032 continue;
5034 scop_i = extract(child);
5035 if (scop->n_stmt != 0 && partial) {
5036 pet_scop_free(scop_i);
5037 break;
5039 pet_skip_info_seq skip(ctx, scop, scop_i);
5040 skip.extract(this);
5041 if (skip)
5042 scop_i = pet_scop_prefix(scop_i, 0);
5043 if (scop_i && child->getStmtClass() == Stmt::DeclStmtClass) {
5044 if (block)
5045 kills.insert(extract_kill(scop_i));
5046 else
5047 scop_i = mark_exposed(scop_i);
5049 scop_i = pet_scop_prefix(scop_i, j);
5050 if (options->autodetect) {
5051 if (scop_i)
5052 scop = pet_scop_add_seq(ctx, scop, scop_i);
5053 else
5054 partial_range = true;
5055 if (scop->n_stmt != 0 && !scop_i)
5056 partial = true;
5057 } else {
5058 scop = pet_scop_add_seq(ctx, scop, scop_i);
5061 scop = skip.add(scop, j);
5063 if (partial || !scop)
5064 break;
5067 for (it = kills.begin(); it != kills.end(); ++it) {
5068 pet_scop *scop_j;
5069 scop_j = pet_scop_from_pet_stmt(ctx, *it);
5070 scop_j = pet_scop_prefix(scop_j, j);
5071 scop = pet_scop_add_seq(ctx, scop, scop_j);
5074 if (scop && partial_range) {
5075 if (scop->n_stmt == 0 || kills.size() != 0) {
5076 pet_scop_free(scop);
5077 return NULL;
5079 partial = true;
5082 return scop;
5085 /* Check if the scop marked by the user is exactly this Stmt
5086 * or part of this Stmt.
5087 * If so, return a pet_scop corresponding to the marked region.
5088 * Otherwise, return NULL.
5090 struct pet_scop *PetScan::scan(Stmt *stmt)
5092 SourceManager &SM = PP.getSourceManager();
5093 unsigned start_off, end_off;
5095 start_off = getExpansionOffset(SM, stmt->getLocStart());
5096 end_off = getExpansionOffset(SM, stmt->getLocEnd());
5098 if (start_off > loc.end)
5099 return NULL;
5100 if (end_off < loc.start)
5101 return NULL;
5102 if (start_off >= loc.start && end_off <= loc.end) {
5103 return extract(stmt);
5106 StmtIterator start;
5107 for (start = stmt->child_begin(); start != stmt->child_end(); ++start) {
5108 Stmt *child = *start;
5109 if (!child)
5110 continue;
5111 start_off = getExpansionOffset(SM, child->getLocStart());
5112 end_off = getExpansionOffset(SM, child->getLocEnd());
5113 if (start_off < loc.start && end_off >= loc.end)
5114 return scan(child);
5115 if (start_off >= loc.start)
5116 break;
5119 StmtIterator end;
5120 for (end = start; end != stmt->child_end(); ++end) {
5121 Stmt *child = *end;
5122 start_off = SM.getFileOffset(child->getLocStart());
5123 if (start_off >= loc.end)
5124 break;
5127 return extract(StmtRange(start, end), false, false);
5130 /* Set the size of index "pos" of "array" to "size".
5131 * In particular, add a constraint of the form
5133 * i_pos < size
5135 * to array->extent and a constraint of the form
5137 * size >= 0
5139 * to array->context.
5141 static struct pet_array *update_size(struct pet_array *array, int pos,
5142 __isl_take isl_pw_aff *size)
5144 isl_set *valid;
5145 isl_set *univ;
5146 isl_set *bound;
5147 isl_space *dim;
5148 isl_aff *aff;
5149 isl_pw_aff *index;
5150 isl_id *id;
5152 valid = isl_pw_aff_nonneg_set(isl_pw_aff_copy(size));
5153 array->context = isl_set_intersect(array->context, valid);
5155 dim = isl_set_get_space(array->extent);
5156 aff = isl_aff_zero_on_domain(isl_local_space_from_space(dim));
5157 aff = isl_aff_add_coefficient_si(aff, isl_dim_in, pos, 1);
5158 univ = isl_set_universe(isl_aff_get_domain_space(aff));
5159 index = isl_pw_aff_alloc(univ, aff);
5161 size = isl_pw_aff_add_dims(size, isl_dim_in,
5162 isl_set_dim(array->extent, isl_dim_set));
5163 id = isl_set_get_tuple_id(array->extent);
5164 size = isl_pw_aff_set_tuple_id(size, isl_dim_in, id);
5165 bound = isl_pw_aff_lt_set(index, size);
5167 array->extent = isl_set_intersect(array->extent, bound);
5169 if (!array->context || !array->extent)
5170 goto error;
5172 return array;
5173 error:
5174 pet_array_free(array);
5175 return NULL;
5178 /* Figure out the size of the array at position "pos" and all
5179 * subsequent positions from "type" and update "array" accordingly.
5181 struct pet_array *PetScan::set_upper_bounds(struct pet_array *array,
5182 const Type *type, int pos)
5184 const ArrayType *atype;
5185 isl_pw_aff *size;
5187 if (!array)
5188 return NULL;
5190 if (type->isPointerType()) {
5191 type = type->getPointeeType().getTypePtr();
5192 return set_upper_bounds(array, type, pos + 1);
5194 if (!type->isArrayType())
5195 return array;
5197 type = type->getCanonicalTypeInternal().getTypePtr();
5198 atype = cast<ArrayType>(type);
5200 if (type->isConstantArrayType()) {
5201 const ConstantArrayType *ca = cast<ConstantArrayType>(atype);
5202 size = extract_affine(ca->getSize());
5203 array = update_size(array, pos, size);
5204 } else if (type->isVariableArrayType()) {
5205 const VariableArrayType *vla = cast<VariableArrayType>(atype);
5206 size = extract_affine(vla->getSizeExpr());
5207 array = update_size(array, pos, size);
5210 type = atype->getElementType().getTypePtr();
5212 return set_upper_bounds(array, type, pos + 1);
5215 /* Is "T" the type of a variable length array with static size?
5217 static bool is_vla_with_static_size(QualType T)
5219 const VariableArrayType *vlatype;
5221 if (!T->isVariableArrayType())
5222 return false;
5223 vlatype = cast<VariableArrayType>(T);
5224 return vlatype->getSizeModifier() == VariableArrayType::Static;
5227 /* Return the type of "decl" as an array.
5229 * In particular, if "decl" is a parameter declaration that
5230 * is a variable length array with a static size, then
5231 * return the original type (i.e., the variable length array).
5232 * Otherwise, return the type of decl.
5234 static QualType get_array_type(ValueDecl *decl)
5236 ParmVarDecl *parm;
5237 QualType T;
5239 parm = dyn_cast<ParmVarDecl>(decl);
5240 if (!parm)
5241 return decl->getType();
5243 T = parm->getOriginalType();
5244 if (!is_vla_with_static_size(T))
5245 return decl->getType();
5246 return T;
5249 /* Does "decl" have definition that we can keep track of in a pet_type?
5251 static bool has_printable_definition(RecordDecl *decl)
5253 if (!decl->getDeclName())
5254 return false;
5255 return decl->getLexicalDeclContext() == decl->getDeclContext();
5258 /* Construct and return a pet_array corresponding to the variable "decl".
5259 * In particular, initialize array->extent to
5261 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
5263 * and then call set_upper_bounds to set the upper bounds on the indices
5264 * based on the type of the variable.
5266 * If the base type is that of a record with a top-level definition and
5267 * if "types" is not null, then the RecordDecl corresponding to the type
5268 * is added to "types".
5270 * If the base type is that of a record with no top-level definition,
5271 * then we replace it by "<subfield>".
5273 struct pet_array *PetScan::extract_array(isl_ctx *ctx, ValueDecl *decl,
5274 lex_recorddecl_set *types)
5276 struct pet_array *array;
5277 QualType qt = get_array_type(decl);
5278 const Type *type = qt.getTypePtr();
5279 int depth = array_depth(type);
5280 QualType base = pet_clang_base_type(qt);
5281 string name;
5282 isl_id *id;
5283 isl_space *dim;
5285 array = isl_calloc_type(ctx, struct pet_array);
5286 if (!array)
5287 return NULL;
5289 id = isl_id_alloc(ctx, decl->getName().str().c_str(), decl);
5290 dim = isl_space_set_alloc(ctx, 0, depth);
5291 dim = isl_space_set_tuple_id(dim, isl_dim_set, id);
5293 array->extent = isl_set_nat_universe(dim);
5295 dim = isl_space_params_alloc(ctx, 0);
5296 array->context = isl_set_universe(dim);
5298 array = set_upper_bounds(array, type, 0);
5299 if (!array)
5300 return NULL;
5302 name = base.getAsString();
5304 if (types && base->isRecordType()) {
5305 RecordDecl *decl = pet_clang_record_decl(base);
5306 if (has_printable_definition(decl))
5307 types->insert(decl);
5308 else
5309 name = "<subfield>";
5312 array->element_type = strdup(name.c_str());
5313 array->element_is_record = base->isRecordType();
5314 array->element_size = decl->getASTContext().getTypeInfo(base).first / 8;
5316 return array;
5319 /* Construct and return a pet_array corresponding to the sequence
5320 * of declarations "decls".
5321 * If the sequence contains a single declaration, then it corresponds
5322 * to a simple array access. Otherwise, it corresponds to a member access,
5323 * with the declaration for the substructure following that of the containing
5324 * structure in the sequence of declarations.
5325 * We start with the outermost substructure and then combine it with
5326 * information from the inner structures.
5328 * Additionally, keep track of all required types in "types".
5330 struct pet_array *PetScan::extract_array(isl_ctx *ctx,
5331 vector<ValueDecl *> decls, lex_recorddecl_set *types)
5333 struct pet_array *array;
5334 vector<ValueDecl *>::iterator it;
5336 it = decls.begin();
5338 array = extract_array(ctx, *it, types);
5340 for (++it; it != decls.end(); ++it) {
5341 struct pet_array *parent;
5342 const char *base_name, *field_name;
5343 char *product_name;
5345 parent = array;
5346 array = extract_array(ctx, *it, types);
5347 if (!array)
5348 return pet_array_free(parent);
5350 base_name = isl_set_get_tuple_name(parent->extent);
5351 field_name = isl_set_get_tuple_name(array->extent);
5352 product_name = member_access_name(ctx, base_name, field_name);
5354 array->extent = isl_set_product(isl_set_copy(parent->extent),
5355 array->extent);
5356 if (product_name)
5357 array->extent = isl_set_set_tuple_name(array->extent,
5358 product_name);
5359 array->context = isl_set_intersect(array->context,
5360 isl_set_copy(parent->context));
5362 pet_array_free(parent);
5363 free(product_name);
5365 if (!array->extent || !array->context || !product_name)
5366 return pet_array_free(array);
5369 return array;
5372 /* Add a pet_type corresponding to "decl" to "scop, provided
5373 * it is a member of "types" and it has not been added before
5374 * (i.e., it is not a member of "types_done".
5376 * Since we want the user to be able to print the types
5377 * in the order in which they appear in the scop, we need to
5378 * make sure that types of fields in a structure appear before
5379 * that structure. We therefore call ourselves recursively
5380 * on the types of all record subfields.
5382 static struct pet_scop *add_type(isl_ctx *ctx, struct pet_scop *scop,
5383 RecordDecl *decl, Preprocessor &PP, lex_recorddecl_set &types,
5384 lex_recorddecl_set &types_done)
5386 string s;
5387 llvm::raw_string_ostream S(s);
5388 RecordDecl::field_iterator it;
5390 if (types.find(decl) == types.end())
5391 return scop;
5392 if (types_done.find(decl) != types_done.end())
5393 return scop;
5395 for (it = decl->field_begin(); it != decl->field_end(); ++it) {
5396 RecordDecl *record;
5397 QualType type = it->getType();
5399 if (!type->isRecordType())
5400 continue;
5401 record = pet_clang_record_decl(type);
5402 scop = add_type(ctx, scop, record, PP, types, types_done);
5405 if (strlen(decl->getName().str().c_str()) == 0)
5406 return scop;
5408 decl->print(S, PrintingPolicy(PP.getLangOpts()));
5409 S.str();
5411 scop->types[scop->n_type] = pet_type_alloc(ctx,
5412 decl->getName().str().c_str(), s.c_str());
5413 if (!scop->types[scop->n_type])
5414 return pet_scop_free(scop);
5416 types_done.insert(decl);
5418 scop->n_type++;
5420 return scop;
5423 /* Construct a list of pet_arrays, one for each array (or scalar)
5424 * accessed inside "scop", add this list to "scop" and return the result.
5426 * The context of "scop" is updated with the intersection of
5427 * the contexts of all arrays, i.e., constraints on the parameters
5428 * that ensure that the arrays have a valid (non-negative) size.
5430 * If the any of the extracted arrays refers to a member access,
5431 * then also add the required types to "scop".
5433 struct pet_scop *PetScan::scan_arrays(struct pet_scop *scop)
5435 int i;
5436 set<vector<ValueDecl *> > arrays;
5437 set<vector<ValueDecl *> >::iterator it;
5438 lex_recorddecl_set types;
5439 lex_recorddecl_set types_done;
5440 lex_recorddecl_set::iterator types_it;
5441 int n_array;
5442 struct pet_array **scop_arrays;
5444 if (!scop)
5445 return NULL;
5447 pet_scop_collect_arrays(scop, arrays);
5448 if (arrays.size() == 0)
5449 return scop;
5451 n_array = scop->n_array;
5453 scop_arrays = isl_realloc_array(ctx, scop->arrays, struct pet_array *,
5454 n_array + arrays.size());
5455 if (!scop_arrays)
5456 goto error;
5457 scop->arrays = scop_arrays;
5459 for (it = arrays.begin(), i = 0; it != arrays.end(); ++it, ++i) {
5460 struct pet_array *array;
5461 array = extract_array(ctx, *it, &types);
5462 scop->arrays[n_array + i] = array;
5463 if (!scop->arrays[n_array + i])
5464 goto error;
5465 scop->n_array++;
5466 scop->context = isl_set_intersect(scop->context,
5467 isl_set_copy(array->context));
5468 if (!scop->context)
5469 goto error;
5472 if (types.size() == 0)
5473 return scop;
5475 scop->types = isl_alloc_array(ctx, struct pet_type *, types.size());
5476 if (!scop->types)
5477 goto error;
5479 for (types_it = types.begin(); types_it != types.end(); ++types_it)
5480 scop = add_type(ctx, scop, *types_it, PP, types, types_done);
5482 return scop;
5483 error:
5484 pet_scop_free(scop);
5485 return NULL;
5488 /* Bound all parameters in scop->context to the possible values
5489 * of the corresponding C variable.
5491 static struct pet_scop *add_parameter_bounds(struct pet_scop *scop)
5493 int n;
5495 if (!scop)
5496 return NULL;
5498 n = isl_set_dim(scop->context, isl_dim_param);
5499 for (int i = 0; i < n; ++i) {
5500 isl_id *id;
5501 ValueDecl *decl;
5503 id = isl_set_get_dim_id(scop->context, isl_dim_param, i);
5504 if (is_nested_parameter(id)) {
5505 isl_id_free(id);
5506 isl_die(isl_set_get_ctx(scop->context),
5507 isl_error_internal,
5508 "unresolved nested parameter", goto error);
5510 decl = (ValueDecl *) isl_id_get_user(id);
5511 isl_id_free(id);
5513 scop->context = set_parameter_bounds(scop->context, i, decl);
5515 if (!scop->context)
5516 goto error;
5519 return scop;
5520 error:
5521 pet_scop_free(scop);
5522 return NULL;
5525 /* Construct a pet_scop from the given function.
5527 * If the scop was delimited by scop and endscop pragmas, then we override
5528 * the file offsets by those derived from the pragmas.
5530 struct pet_scop *PetScan::scan(FunctionDecl *fd)
5532 pet_scop *scop;
5533 Stmt *stmt;
5535 stmt = fd->getBody();
5537 if (options->autodetect)
5538 scop = extract(stmt, true);
5539 else {
5540 scop = scan(stmt);
5541 scop = pet_scop_update_start_end(scop, loc.start, loc.end);
5543 scop = pet_scop_detect_parameter_accesses(scop);
5544 scop = scan_arrays(scop);
5545 scop = add_parameter_bounds(scop);
5546 scop = pet_scop_gist(scop, value_bounds);
5548 return scop;