2 * Copyright 2011 Leiden University. All rights reserved.
3 * Copyright 2012-2014 Ecole Normale Superieure. All rights reserved.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
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
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>
46 #include <isl/space.h>
60 #include "scop_plus.h"
66 using namespace clang
;
68 static enum pet_op_type
UnaryOperatorKind2pet_op_type(UnaryOperatorKind kind
)
78 return pet_op_post_inc
;
80 return pet_op_post_dec
;
82 return pet_op_pre_inc
;
84 return pet_op_pre_dec
;
90 static enum pet_op_type
BinaryOperatorKind2pet_op_type(BinaryOperatorKind kind
)
94 return pet_op_add_assign
;
96 return pet_op_sub_assign
;
98 return pet_op_mul_assign
;
100 return pet_op_div_assign
;
102 return pet_op_assign
;
144 #if defined(DECLREFEXPR_CREATE_REQUIRES_BOOL)
145 static DeclRefExpr
*create_DeclRefExpr(VarDecl
*var
)
147 return DeclRefExpr::Create(var
->getASTContext(), var
->getQualifierLoc(),
148 SourceLocation(), var
, false, var
->getInnerLocStart(),
149 var
->getType(), VK_LValue
);
151 #elif defined(DECLREFEXPR_CREATE_REQUIRES_SOURCELOCATION)
152 static DeclRefExpr
*create_DeclRefExpr(VarDecl
*var
)
154 return DeclRefExpr::Create(var
->getASTContext(), var
->getQualifierLoc(),
155 SourceLocation(), var
, var
->getInnerLocStart(), var
->getType(),
159 static DeclRefExpr
*create_DeclRefExpr(VarDecl
*var
)
161 return DeclRefExpr::Create(var
->getASTContext(), var
->getQualifierLoc(),
162 var
, var
->getInnerLocStart(), var
->getType(), VK_LValue
);
166 /* Check if the element type corresponding to the given array type
167 * has a const qualifier.
169 static bool const_base(QualType qt
)
171 const Type
*type
= qt
.getTypePtr();
173 if (type
->isPointerType())
174 return const_base(type
->getPointeeType());
175 if (type
->isArrayType()) {
176 const ArrayType
*atype
;
177 type
= type
->getCanonicalTypeInternal().getTypePtr();
178 atype
= cast
<ArrayType
>(type
);
179 return const_base(atype
->getElementType());
182 return qt
.isConstQualified();
185 /* Create an isl_id that refers to the named declarator "decl".
187 static __isl_give isl_id
*create_decl_id(isl_ctx
*ctx
, NamedDecl
*decl
)
189 return isl_id_alloc(ctx
, decl
->getName().str().c_str(), decl
);
192 /* Look for any assignments to scalar variables in part of the parse
193 * tree and mark them as having an unknown value in "pc".
194 * If the address of a scalar variable is being taken, then mark
195 * it as having an unknown value as well. As an exception,
196 * if the address is passed to a function
197 * that is declared to receive a const pointer, then "pc" is not updated.
199 * This ensures that we won't use any previously stored value
200 * in the current subtree and its parents.
202 struct clear_assignments
: RecursiveASTVisitor
<clear_assignments
> {
205 set
<UnaryOperator
*> skip
;
207 clear_assignments(pet_context
*&pc
) : pc(pc
),
208 ctx(pet_context_get_ctx(pc
)) {}
210 /* Check for "address of" operators whose value is passed
211 * to a const pointer argument and add them to "skip", so that
212 * we can skip them in VisitUnaryOperator.
214 bool VisitCallExpr(CallExpr
*expr
) {
216 fd
= expr
->getDirectCallee();
219 for (int i
= 0; i
< expr
->getNumArgs(); ++i
) {
220 Expr
*arg
= expr
->getArg(i
);
222 if (arg
->getStmtClass() == Stmt::ImplicitCastExprClass
) {
223 ImplicitCastExpr
*ice
;
224 ice
= cast
<ImplicitCastExpr
>(arg
);
225 arg
= ice
->getSubExpr();
227 if (arg
->getStmtClass() != Stmt::UnaryOperatorClass
)
229 op
= cast
<UnaryOperator
>(arg
);
230 if (op
->getOpcode() != UO_AddrOf
)
232 if (const_base(fd
->getParamDecl(i
)->getType()))
238 bool VisitUnaryOperator(UnaryOperator
*expr
) {
243 switch (expr
->getOpcode()) {
253 if (skip
.find(expr
) != skip
.end())
256 arg
= expr
->getSubExpr();
257 if (arg
->getStmtClass() != Stmt::DeclRefExprClass
)
259 ref
= cast
<DeclRefExpr
>(arg
);
260 decl
= ref
->getDecl();
261 pc
= pet_context_mark_assigned(pc
, create_decl_id(ctx
, decl
));
265 bool VisitBinaryOperator(BinaryOperator
*expr
) {
270 if (!expr
->isAssignmentOp())
272 lhs
= expr
->getLHS();
273 if (lhs
->getStmtClass() != Stmt::DeclRefExprClass
)
275 ref
= cast
<DeclRefExpr
>(lhs
);
276 decl
= ref
->getDecl();
277 pc
= pet_context_mark_assigned(pc
, create_decl_id(ctx
, decl
));
284 isl_union_map_free(value_bounds
);
287 /* Report a diagnostic, unless autodetect is set.
289 void PetScan::report(Stmt
*stmt
, unsigned id
)
291 if (options
->autodetect
)
294 SourceLocation loc
= stmt
->getLocStart();
295 DiagnosticsEngine
&diag
= PP
.getDiagnostics();
296 DiagnosticBuilder B
= diag
.Report(loc
, id
) << stmt
->getSourceRange();
299 /* Called if we found something we (currently) cannot handle.
300 * We'll provide more informative warnings later.
302 * We only actually complain if autodetect is false.
304 void PetScan::unsupported(Stmt
*stmt
)
306 DiagnosticsEngine
&diag
= PP
.getDiagnostics();
307 unsigned id
= diag
.getCustomDiagID(DiagnosticsEngine::Warning
,
312 /* Report a missing prototype, unless autodetect is set.
314 void PetScan::report_prototype_required(Stmt
*stmt
)
316 DiagnosticsEngine
&diag
= PP
.getDiagnostics();
317 unsigned id
= diag
.getCustomDiagID(DiagnosticsEngine::Warning
,
318 "prototype required");
322 /* Report a missing increment, unless autodetect is set.
324 void PetScan::report_missing_increment(Stmt
*stmt
)
326 DiagnosticsEngine
&diag
= PP
.getDiagnostics();
327 unsigned id
= diag
.getCustomDiagID(DiagnosticsEngine::Warning
,
328 "missing increment");
332 /* Extract an integer from "expr".
334 __isl_give isl_val
*PetScan::extract_int(isl_ctx
*ctx
, IntegerLiteral
*expr
)
336 const Type
*type
= expr
->getType().getTypePtr();
337 int is_signed
= type
->hasSignedIntegerRepresentation();
338 llvm::APInt val
= expr
->getValue();
339 int is_negative
= is_signed
&& val
.isNegative();
345 v
= extract_unsigned(ctx
, val
);
352 /* Extract an integer from "val", which is assumed to be non-negative.
354 __isl_give isl_val
*PetScan::extract_unsigned(isl_ctx
*ctx
,
355 const llvm::APInt
&val
)
358 const uint64_t *data
;
360 data
= val
.getRawData();
361 n
= val
.getNumWords();
362 return isl_val_int_from_chunks(ctx
, n
, sizeof(uint64_t), data
);
365 /* Extract an integer from "expr".
366 * Return NULL if "expr" does not (obviously) represent an integer.
368 __isl_give isl_val
*PetScan::extract_int(clang::ParenExpr
*expr
)
370 return extract_int(expr
->getSubExpr());
373 /* Extract an integer from "expr".
374 * Return NULL if "expr" does not (obviously) represent an integer.
376 __isl_give isl_val
*PetScan::extract_int(clang::Expr
*expr
)
378 if (expr
->getStmtClass() == Stmt::IntegerLiteralClass
)
379 return extract_int(ctx
, cast
<IntegerLiteral
>(expr
));
380 if (expr
->getStmtClass() == Stmt::ParenExprClass
)
381 return extract_int(cast
<ParenExpr
>(expr
));
387 /* Extract a pet_expr from the APInt "val", which is assumed
388 * to be non-negative.
390 __isl_give pet_expr
*PetScan::extract_expr(const llvm::APInt
&val
)
392 return pet_expr_new_int(extract_unsigned(ctx
, val
));
395 /* Return the number of bits needed to represent the type "qt",
396 * if it is an integer type. Otherwise return 0.
397 * If qt is signed then return the opposite of the number of bits.
399 static int get_type_size(QualType qt
, ASTContext
&ast_context
)
403 if (!qt
->isIntegerType())
406 size
= ast_context
.getIntWidth(qt
);
407 if (!qt
->isUnsignedIntegerType())
413 /* Return the number of bits needed to represent the type of "decl",
414 * if it is an integer type. Otherwise return 0.
415 * If qt is signed then return the opposite of the number of bits.
417 static int get_type_size(ValueDecl
*decl
)
419 return get_type_size(decl
->getType(), decl
->getASTContext());
422 /* Bound parameter "pos" of "set" to the possible values of "decl".
424 static __isl_give isl_set
*set_parameter_bounds(__isl_take isl_set
*set
,
425 unsigned pos
, ValueDecl
*decl
)
431 ctx
= isl_set_get_ctx(set
);
432 type_size
= get_type_size(decl
);
434 isl_die(ctx
, isl_error_invalid
, "not an integer type",
435 return isl_set_free(set
));
437 set
= isl_set_lower_bound_si(set
, isl_dim_param
, pos
, 0);
438 bound
= isl_val_int_from_ui(ctx
, type_size
);
439 bound
= isl_val_2exp(bound
);
440 bound
= isl_val_sub_ui(bound
, 1);
441 set
= isl_set_upper_bound_val(set
, isl_dim_param
, pos
, bound
);
443 bound
= isl_val_int_from_ui(ctx
, -type_size
- 1);
444 bound
= isl_val_2exp(bound
);
445 bound
= isl_val_sub_ui(bound
, 1);
446 set
= isl_set_upper_bound_val(set
, isl_dim_param
, pos
,
447 isl_val_copy(bound
));
448 bound
= isl_val_neg(bound
);
449 bound
= isl_val_sub_ui(bound
, 1);
450 set
= isl_set_lower_bound_val(set
, isl_dim_param
, pos
, bound
);
456 /* Return the piecewise affine expression "set ? 1 : 0" defined on "dom".
458 static __isl_give isl_pw_aff
*indicator_function(__isl_take isl_set
*set
,
459 __isl_take isl_set
*dom
)
462 pa
= isl_set_indicator_function(set
);
463 pa
= isl_pw_aff_intersect_domain(pa
, isl_set_coalesce(dom
));
467 /* Extract an affine expression, if possible, from "expr"
468 * within the context "pc", except that nesting is enabled.
469 * Otherwise return NULL.
471 __isl_give isl_pw_aff
*PetScan::extract_affine(Expr
*expr
,
472 __isl_keep pet_context
*pc
)
477 pe
= extract_expr(expr
);
480 pe
= pet_expr_plug_in_args(pe
, pc
);
481 pa
= pet_expr_extract_affine(pe
, pc
);
482 if (isl_pw_aff_involves_nan(pa
)) {
484 pa
= isl_pw_aff_free(pa
);
491 __isl_give pet_expr
*PetScan::extract_index_expr(ImplicitCastExpr
*expr
)
493 return extract_index_expr(expr
->getSubExpr());
496 /* Return the depth of an array of the given type.
498 static int array_depth(const Type
*type
)
500 if (type
->isPointerType())
501 return 1 + array_depth(type
->getPointeeType().getTypePtr());
502 if (type
->isArrayType()) {
503 const ArrayType
*atype
;
504 type
= type
->getCanonicalTypeInternal().getTypePtr();
505 atype
= cast
<ArrayType
>(type
);
506 return 1 + array_depth(atype
->getElementType().getTypePtr());
511 /* Return the depth of the array accessed by the index expression "index".
512 * If "index" is an affine expression, i.e., if it does not access
513 * any array, then return 1.
514 * If "index" represent a member access, i.e., if its range is a wrapped
515 * relation, then return the sum of the depth of the array of structures
516 * and that of the member inside the structure.
518 static int extract_depth(__isl_keep isl_multi_pw_aff
*index
)
526 if (isl_multi_pw_aff_range_is_wrapping(index
)) {
527 int domain_depth
, range_depth
;
528 isl_multi_pw_aff
*domain
, *range
;
530 domain
= isl_multi_pw_aff_copy(index
);
531 domain
= isl_multi_pw_aff_range_factor_domain(domain
);
532 domain_depth
= extract_depth(domain
);
533 isl_multi_pw_aff_free(domain
);
534 range
= isl_multi_pw_aff_copy(index
);
535 range
= isl_multi_pw_aff_range_factor_range(range
);
536 range_depth
= extract_depth(range
);
537 isl_multi_pw_aff_free(range
);
539 return domain_depth
+ range_depth
;
542 if (!isl_multi_pw_aff_has_tuple_id(index
, isl_dim_out
))
545 id
= isl_multi_pw_aff_get_tuple_id(index
, isl_dim_out
);
548 decl
= (ValueDecl
*) isl_id_get_user(id
);
551 return array_depth(decl
->getType().getTypePtr());
554 /* Return the depth of the array accessed by the access expression "expr".
556 static int extract_depth(__isl_keep pet_expr
*expr
)
558 isl_multi_pw_aff
*index
;
561 index
= pet_expr_access_get_index(expr
);
562 depth
= extract_depth(index
);
563 isl_multi_pw_aff_free(index
);
568 /* Construct a pet_expr representing an index expression for an access
569 * to the variable referenced by "expr".
571 __isl_give pet_expr
*PetScan::extract_index_expr(DeclRefExpr
*expr
)
573 return extract_index_expr(expr
->getDecl());
576 /* Construct a pet_expr representing an index expression for an access
577 * to the variable "decl".
579 __isl_give pet_expr
*PetScan::extract_index_expr(ValueDecl
*decl
)
581 isl_id
*id
= create_decl_id(ctx
, decl
);
582 isl_space
*space
= isl_space_alloc(ctx
, 0, 0, 0);
584 space
= isl_space_set_tuple_id(space
, isl_dim_out
, id
);
586 return pet_expr_from_index(isl_multi_pw_aff_zero(space
));
589 /* Construct a pet_expr representing the index expression "expr"
590 * Return NULL on error.
592 __isl_give pet_expr
*PetScan::extract_index_expr(Expr
*expr
)
594 switch (expr
->getStmtClass()) {
595 case Stmt::ImplicitCastExprClass
:
596 return extract_index_expr(cast
<ImplicitCastExpr
>(expr
));
597 case Stmt::DeclRefExprClass
:
598 return extract_index_expr(cast
<DeclRefExpr
>(expr
));
599 case Stmt::ArraySubscriptExprClass
:
600 return extract_index_expr(cast
<ArraySubscriptExpr
>(expr
));
601 case Stmt::IntegerLiteralClass
:
602 return extract_expr(cast
<IntegerLiteral
>(expr
));
603 case Stmt::MemberExprClass
:
604 return extract_index_expr(cast
<MemberExpr
>(expr
));
611 /* Extract an index expression from the given array subscript expression.
613 * We first extract an index expression from the base.
614 * This will result in an index expression with a range that corresponds
615 * to the earlier indices.
616 * We then extract the current index and let
617 * pet_expr_access_subscript combine the two.
619 __isl_give pet_expr
*PetScan::extract_index_expr(ArraySubscriptExpr
*expr
)
621 Expr
*base
= expr
->getBase();
622 Expr
*idx
= expr
->getIdx();
626 base_expr
= extract_index_expr(base
);
627 index
= extract_expr(idx
);
629 base_expr
= pet_expr_access_subscript(base_expr
, index
);
634 /* Extract an index expression from a member expression.
636 * If the base access (to the structure containing the member)
641 * and the member is called "f", then the member access is of
646 * If the member access is to an anonymous struct, then simply return
650 * If the member access in the source code is of the form
654 * then it is treated as
658 __isl_give pet_expr
*PetScan::extract_index_expr(MemberExpr
*expr
)
660 Expr
*base
= expr
->getBase();
661 FieldDecl
*field
= cast
<FieldDecl
>(expr
->getMemberDecl());
662 pet_expr
*base_index
;
665 base_index
= extract_index_expr(base
);
667 if (expr
->isArrow()) {
668 pet_expr
*index
= pet_expr_new_int(isl_val_zero(ctx
));
669 base_index
= pet_expr_access_subscript(base_index
, index
);
672 if (field
->isAnonymousStructOrUnion())
675 id
= create_decl_id(ctx
, field
);
677 return pet_expr_access_member(base_index
, id
);
680 /* Check if "expr" calls function "minmax" with two arguments and if so
681 * make lhs and rhs refer to these two arguments.
683 static bool is_minmax(Expr
*expr
, const char *minmax
, Expr
*&lhs
, Expr
*&rhs
)
689 if (expr
->getStmtClass() != Stmt::CallExprClass
)
692 call
= cast
<CallExpr
>(expr
);
693 fd
= call
->getDirectCallee();
697 if (call
->getNumArgs() != 2)
700 name
= fd
->getDeclName().getAsString();
704 lhs
= call
->getArg(0);
705 rhs
= call
->getArg(1);
710 /* Check if "expr" is of the form min(lhs, rhs) and if so make
711 * lhs and rhs refer to the two arguments.
713 static bool is_min(Expr
*expr
, Expr
*&lhs
, Expr
*&rhs
)
715 return is_minmax(expr
, "min", lhs
, rhs
);
718 /* Check if "expr" is of the form max(lhs, rhs) and if so make
719 * lhs and rhs refer to the two arguments.
721 static bool is_max(Expr
*expr
, Expr
*&lhs
, Expr
*&rhs
)
723 return is_minmax(expr
, "max", lhs
, rhs
);
726 /* Extract an affine expressions representing the comparison "LHS op RHS"
727 * within the context "pc".
728 * "comp" is the original statement that "LHS op RHS" is derived from
729 * and is used for diagnostics.
731 * If the comparison is of the form
735 * then the expression is constructed as the conjunction of
740 * A similar optimization is performed for max(a,b) <= c.
741 * We do this because that will lead to simpler representations
743 * If isl is ever enhanced to explicitly deal with min and max expressions,
744 * this optimization can be removed.
746 __isl_give isl_pw_aff
*PetScan::extract_comparison(BinaryOperatorKind op
,
747 Expr
*LHS
, Expr
*RHS
, Stmt
*comp
, __isl_keep pet_context
*pc
)
754 enum pet_op_type type
;
757 return extract_comparison(BO_LT
, RHS
, LHS
, comp
, pc
);
759 return extract_comparison(BO_LE
, RHS
, LHS
, comp
, pc
);
761 if (op
== BO_LT
|| op
== BO_LE
) {
763 if (is_min(RHS
, expr1
, expr2
)) {
764 lhs
= extract_comparison(op
, LHS
, expr1
, comp
, pc
);
765 rhs
= extract_comparison(op
, LHS
, expr2
, comp
, pc
);
766 return pet_and(lhs
, rhs
);
768 if (is_max(LHS
, expr1
, expr2
)) {
769 lhs
= extract_comparison(op
, expr1
, RHS
, comp
, pc
);
770 rhs
= extract_comparison(op
, expr2
, RHS
, comp
, pc
);
771 return pet_and(lhs
, rhs
);
775 lhs
= extract_affine(LHS
, pc
);
776 rhs
= extract_affine(RHS
, pc
);
778 type
= BinaryOperatorKind2pet_op_type(op
);
779 return pet_comparison(type
, lhs
, rhs
);
782 __isl_give isl_pw_aff
*PetScan::extract_comparison(BinaryOperator
*comp
,
783 __isl_keep pet_context
*pc
)
785 return extract_comparison(comp
->getOpcode(), comp
->getLHS(),
786 comp
->getRHS(), comp
, pc
);
789 /* Extract an affine expression from a boolean expression
790 * within the context "pc".
791 * In particular, return the expression "expr ? 1 : 0".
792 * Return NULL if we are unable to extract an affine expression.
794 * We first convert the clang::Expr to a pet_expr and
795 * then extract an affine expression from that pet_expr.
797 __isl_give isl_pw_aff
*PetScan::extract_condition(Expr
*expr
,
798 __isl_keep pet_context
*pc
)
804 isl_set
*u
= isl_set_universe(isl_space_set_alloc(ctx
, 0, 0));
805 return indicator_function(u
, isl_set_copy(u
));
808 pe
= extract_expr(expr
);
809 pe
= pet_expr_plug_in_args(pe
, pc
);
810 pc
= pet_context_copy(pc
);
811 pc
= pet_context_set_allow_nested(pc
, nesting_enabled
);
812 cond
= pet_expr_extract_affine_condition(pe
, pc
);
813 if (isl_pw_aff_involves_nan(cond
))
814 cond
= isl_pw_aff_free(cond
);
815 pet_context_free(pc
);
820 /* Mark the given access pet_expr as a write.
822 static __isl_give pet_expr
*mark_write(__isl_take pet_expr
*access
)
824 access
= pet_expr_access_set_write(access
, 1);
825 access
= pet_expr_access_set_read(access
, 0);
830 /* Construct a pet_expr representing a unary operator expression.
832 __isl_give pet_expr
*PetScan::extract_expr(UnaryOperator
*expr
)
837 op
= UnaryOperatorKind2pet_op_type(expr
->getOpcode());
838 if (op
== pet_op_last
) {
843 arg
= extract_expr(expr
->getSubExpr());
845 if (expr
->isIncrementDecrementOp() &&
846 pet_expr_get_type(arg
) == pet_expr_access
) {
847 arg
= mark_write(arg
);
848 arg
= pet_expr_access_set_read(arg
, 1);
851 return pet_expr_new_unary(op
, arg
);
854 /* If the access expression "expr" writes to a (non-virtual) scalar,
855 * then mark the scalar as having an unknown value in "pc".
857 static int clear_write(__isl_keep pet_expr
*expr
, void *user
)
860 pet_context
**pc
= (pet_context
**) user
;
862 if (!pet_expr_access_is_write(expr
))
864 if (!pet_expr_is_scalar_access(expr
))
867 id
= pet_expr_access_get_id(expr
);
868 if (isl_id_get_user(id
))
869 *pc
= pet_context_mark_assigned(*pc
, id
);
876 /* Update "pc" by taking into account the writes in "stmt".
877 * That is, first mark all scalar variables that are written by "stmt"
878 * as having an unknown value. Afterwards,
879 * if "stmt" is a top-level (i.e., unconditional) assignment
880 * to a scalar variable, then update "pc" accordingly.
882 * In particular, if the lhs of the assignment is a scalar variable, then mark
883 * the variable as having been assigned. If, furthermore, the rhs
884 * is an affine expression, then keep track of this value in "pc"
885 * so that we can plug it in when we later come across the same variable.
887 * We skip assignments to virtual arrays (those with NULL user pointer).
889 __isl_give pet_context
*PetScan::handle_writes(struct pet_stmt
*stmt
,
890 __isl_take pet_context
*pc
)
892 pet_expr
*body
= stmt
->body
;
897 if (pet_expr_foreach_access_expr(body
, &clear_write
, &pc
) < 0)
898 return pet_context_free(pc
);
900 if (!pet_stmt_is_assign(stmt
))
902 if (!isl_set_plain_is_universe(stmt
->domain
))
904 arg
= pet_expr_get_arg(body
, 0);
905 if (!pet_expr_is_scalar_access(arg
)) {
910 id
= pet_expr_access_get_id(arg
);
913 if (!isl_id_get_user(id
)) {
918 arg
= pet_expr_get_arg(body
, 1);
919 pa
= pet_expr_extract_affine(arg
, pc
);
920 pc
= pet_context_mark_assigned(pc
, isl_id_copy(id
));
923 if (isl_pw_aff_involves_nan(pa
))
924 pa
= isl_pw_aff_free(pa
);
930 pc
= pet_context_set_value(pc
, id
, pa
);
935 /* Update "pc" based on the write accesses (and, in particular,
936 * assignments) in "scop".
938 __isl_give pet_context
*PetScan::handle_writes(struct pet_scop
*scop
,
939 __isl_take pet_context
*pc
)
942 return pet_context_free(pc
);
943 for (int i
= 0; i
< scop
->n_stmt
; ++i
)
944 pc
= handle_writes(scop
->stmts
[i
], pc
);
949 /* Construct a pet_expr representing a binary operator expression.
951 * If the top level operator is an assignment and the LHS is an access,
952 * then we mark that access as a write. If the operator is a compound
953 * assignment, the access is marked as both a read and a write.
955 __isl_give pet_expr
*PetScan::extract_expr(BinaryOperator
*expr
)
961 op
= BinaryOperatorKind2pet_op_type(expr
->getOpcode());
962 if (op
== pet_op_last
) {
967 lhs
= extract_expr(expr
->getLHS());
968 rhs
= extract_expr(expr
->getRHS());
970 if (expr
->isAssignmentOp() &&
971 pet_expr_get_type(lhs
) == pet_expr_access
) {
972 lhs
= mark_write(lhs
);
973 if (expr
->isCompoundAssignmentOp())
974 lhs
= pet_expr_access_set_read(lhs
, 1);
977 type_size
= get_type_size(expr
->getType(), ast_context
);
978 return pet_expr_new_binary(type_size
, op
, lhs
, rhs
);
981 /* Construct a pet_scop with a single statement killing the entire
982 * array "array" within the context "pc".
984 struct pet_scop
*PetScan::kill(Stmt
*stmt
, struct pet_array
*array
,
985 __isl_keep pet_context
*pc
)
989 isl_multi_pw_aff
*index
;
995 access
= isl_map_from_range(isl_set_copy(array
->extent
));
996 id
= isl_set_get_tuple_id(array
->extent
);
997 space
= isl_space_alloc(ctx
, 0, 0, 0);
998 space
= isl_space_set_tuple_id(space
, isl_dim_out
, id
);
999 index
= isl_multi_pw_aff_zero(space
);
1000 expr
= pet_expr_kill_from_access_and_index(access
, index
);
1001 return extract(expr
, stmt
->getSourceRange(), false, pc
);
1004 /* Construct a pet_scop for a (single) variable declaration
1005 * within the context "pc".
1007 * The scop contains the variable being declared (as an array)
1008 * and a statement killing the array.
1010 * If the variable is initialized in the AST, then the scop
1011 * also contains an assignment to the variable.
1013 struct pet_scop
*PetScan::extract(DeclStmt
*stmt
, __isl_keep pet_context
*pc
)
1018 pet_expr
*lhs
, *rhs
, *pe
;
1019 struct pet_scop
*scop_decl
, *scop
;
1020 struct pet_array
*array
;
1022 if (!stmt
->isSingleDecl()) {
1027 decl
= stmt
->getSingleDecl();
1028 vd
= cast
<VarDecl
>(decl
);
1030 array
= extract_array(ctx
, vd
, NULL
, pc
);
1032 array
->declared
= 1;
1033 scop_decl
= kill(stmt
, array
, pc
);
1034 scop_decl
= pet_scop_add_array(scop_decl
, array
);
1039 lhs
= extract_access_expr(vd
);
1040 rhs
= extract_expr(vd
->getInit());
1042 lhs
= mark_write(lhs
);
1044 type_size
= get_type_size(vd
->getType(), ast_context
);
1045 pe
= pet_expr_new_binary(type_size
, pet_op_assign
, lhs
, rhs
);
1046 scop
= extract(pe
, stmt
->getSourceRange(), false, pc
);
1048 scop_decl
= pet_scop_prefix(scop_decl
, 0);
1049 scop
= pet_scop_prefix(scop
, 1);
1051 scop
= pet_scop_add_seq(ctx
, scop_decl
, scop
);
1056 /* Construct a pet_expr representing a conditional operation.
1058 __isl_give pet_expr
*PetScan::extract_expr(ConditionalOperator
*expr
)
1060 pet_expr
*cond
, *lhs
, *rhs
;
1063 cond
= extract_expr(expr
->getCond());
1064 lhs
= extract_expr(expr
->getTrueExpr());
1065 rhs
= extract_expr(expr
->getFalseExpr());
1067 return pet_expr_new_ternary(cond
, lhs
, rhs
);
1070 __isl_give pet_expr
*PetScan::extract_expr(ImplicitCastExpr
*expr
)
1072 return extract_expr(expr
->getSubExpr());
1075 /* Construct a pet_expr representing a floating point value.
1077 * If the floating point literal does not appear in a macro,
1078 * then we use the original representation in the source code
1079 * as the string representation. Otherwise, we use the pretty
1080 * printer to produce a string representation.
1082 __isl_give pet_expr
*PetScan::extract_expr(FloatingLiteral
*expr
)
1086 const LangOptions
&LO
= PP
.getLangOpts();
1087 SourceLocation loc
= expr
->getLocation();
1089 if (!loc
.isMacroID()) {
1090 SourceManager
&SM
= PP
.getSourceManager();
1091 unsigned len
= Lexer::MeasureTokenLength(loc
, SM
, LO
);
1092 s
= string(SM
.getCharacterData(loc
), len
);
1094 llvm::raw_string_ostream
S(s
);
1095 expr
->printPretty(S
, 0, PrintingPolicy(LO
));
1098 d
= expr
->getValueAsApproximateDouble();
1099 return pet_expr_new_double(ctx
, d
, s
.c_str());
1102 /* Convert the index expression "index" into an access pet_expr of type "qt".
1104 __isl_give pet_expr
*PetScan::extract_access_expr(QualType qt
,
1105 __isl_take pet_expr
*index
)
1110 depth
= extract_depth(index
);
1111 type_size
= get_type_size(qt
, ast_context
);
1113 index
= pet_expr_set_type_size(index
, type_size
);
1114 index
= pet_expr_access_set_depth(index
, depth
);
1119 /* Extract an index expression from "expr" and then convert it into
1120 * an access pet_expr.
1122 __isl_give pet_expr
*PetScan::extract_access_expr(Expr
*expr
)
1124 return extract_access_expr(expr
->getType(), extract_index_expr(expr
));
1127 /* Extract an index expression from "decl" and then convert it into
1128 * an access pet_expr.
1130 __isl_give pet_expr
*PetScan::extract_access_expr(ValueDecl
*decl
)
1132 return extract_access_expr(decl
->getType(), extract_index_expr(decl
));
1135 __isl_give pet_expr
*PetScan::extract_expr(ParenExpr
*expr
)
1137 return extract_expr(expr
->getSubExpr());
1140 /* Extract an assume statement from the argument "expr"
1141 * of a __pencil_assume statement.
1143 __isl_give pet_expr
*PetScan::extract_assume(Expr
*expr
)
1145 return pet_expr_new_unary(pet_op_assume
, extract_expr(expr
));
1148 /* Construct a pet_expr corresponding to the function call argument "expr".
1149 * The argument appears in position "pos" of a call to function "fd".
1151 * If we are passing along a pointer to an array element
1152 * or an entire row or even higher dimensional slice of an array,
1153 * then the function being called may write into the array.
1155 * We assume here that if the function is declared to take a pointer
1156 * to a const type, then the function will perform a read
1157 * and that otherwise, it will perform a write.
1159 __isl_give pet_expr
*PetScan::extract_argument(FunctionDecl
*fd
, int pos
,
1163 int is_addr
= 0, is_partial
= 0;
1166 if (expr
->getStmtClass() == Stmt::ImplicitCastExprClass
) {
1167 ImplicitCastExpr
*ice
= cast
<ImplicitCastExpr
>(expr
);
1168 expr
= ice
->getSubExpr();
1170 if (expr
->getStmtClass() == Stmt::UnaryOperatorClass
) {
1171 UnaryOperator
*op
= cast
<UnaryOperator
>(expr
);
1172 if (op
->getOpcode() == UO_AddrOf
) {
1174 expr
= op
->getSubExpr();
1177 res
= extract_expr(expr
);
1180 sc
= expr
->getStmtClass();
1181 if ((sc
== Stmt::ArraySubscriptExprClass
||
1182 sc
== Stmt::MemberExprClass
) &&
1183 array_depth(expr
->getType().getTypePtr()) > 0)
1185 if ((is_addr
|| is_partial
) &&
1186 pet_expr_get_type(res
) == pet_expr_access
) {
1188 if (!fd
->hasPrototype()) {
1189 report_prototype_required(expr
);
1190 return pet_expr_free(res
);
1192 parm
= fd
->getParamDecl(pos
);
1193 if (!const_base(parm
->getType()))
1194 res
= mark_write(res
);
1198 res
= pet_expr_new_unary(pet_op_address_of
, res
);
1202 /* Construct a pet_expr representing a function call.
1204 * In the special case of a "call" to __pencil_assume,
1205 * construct an assume expression instead.
1207 __isl_give pet_expr
*PetScan::extract_expr(CallExpr
*expr
)
1209 pet_expr
*res
= NULL
;
1214 fd
= expr
->getDirectCallee();
1220 name
= fd
->getDeclName().getAsString();
1221 n_arg
= expr
->getNumArgs();
1223 if (n_arg
== 1 && name
== "__pencil_assume")
1224 return extract_assume(expr
->getArg(0));
1226 res
= pet_expr_new_call(ctx
, name
.c_str(), n_arg
);
1230 for (int i
= 0; i
< n_arg
; ++i
) {
1231 Expr
*arg
= expr
->getArg(i
);
1232 res
= pet_expr_set_arg(res
, i
,
1233 PetScan::extract_argument(fd
, i
, arg
));
1239 /* Construct a pet_expr representing a (C style) cast.
1241 __isl_give pet_expr
*PetScan::extract_expr(CStyleCastExpr
*expr
)
1246 arg
= extract_expr(expr
->getSubExpr());
1250 type
= expr
->getTypeAsWritten();
1251 return pet_expr_new_cast(type
.getAsString().c_str(), arg
);
1254 /* Construct a pet_expr representing an integer.
1256 __isl_give pet_expr
*PetScan::extract_expr(IntegerLiteral
*expr
)
1258 return pet_expr_new_int(extract_int(expr
));
1261 /* Try and construct a pet_expr representing "expr".
1263 __isl_give pet_expr
*PetScan::extract_expr(Expr
*expr
)
1265 switch (expr
->getStmtClass()) {
1266 case Stmt::UnaryOperatorClass
:
1267 return extract_expr(cast
<UnaryOperator
>(expr
));
1268 case Stmt::CompoundAssignOperatorClass
:
1269 case Stmt::BinaryOperatorClass
:
1270 return extract_expr(cast
<BinaryOperator
>(expr
));
1271 case Stmt::ImplicitCastExprClass
:
1272 return extract_expr(cast
<ImplicitCastExpr
>(expr
));
1273 case Stmt::ArraySubscriptExprClass
:
1274 case Stmt::DeclRefExprClass
:
1275 case Stmt::MemberExprClass
:
1276 return extract_access_expr(expr
);
1277 case Stmt::IntegerLiteralClass
:
1278 return extract_expr(cast
<IntegerLiteral
>(expr
));
1279 case Stmt::FloatingLiteralClass
:
1280 return extract_expr(cast
<FloatingLiteral
>(expr
));
1281 case Stmt::ParenExprClass
:
1282 return extract_expr(cast
<ParenExpr
>(expr
));
1283 case Stmt::ConditionalOperatorClass
:
1284 return extract_expr(cast
<ConditionalOperator
>(expr
));
1285 case Stmt::CallExprClass
:
1286 return extract_expr(cast
<CallExpr
>(expr
));
1287 case Stmt::CStyleCastExprClass
:
1288 return extract_expr(cast
<CStyleCastExpr
>(expr
));
1295 /* Check if the given initialization statement is an assignment.
1296 * If so, return that assignment. Otherwise return NULL.
1298 BinaryOperator
*PetScan::initialization_assignment(Stmt
*init
)
1300 BinaryOperator
*ass
;
1302 if (init
->getStmtClass() != Stmt::BinaryOperatorClass
)
1305 ass
= cast
<BinaryOperator
>(init
);
1306 if (ass
->getOpcode() != BO_Assign
)
1312 /* Check if the given initialization statement is a declaration
1313 * of a single variable.
1314 * If so, return that declaration. Otherwise return NULL.
1316 Decl
*PetScan::initialization_declaration(Stmt
*init
)
1320 if (init
->getStmtClass() != Stmt::DeclStmtClass
)
1323 decl
= cast
<DeclStmt
>(init
);
1325 if (!decl
->isSingleDecl())
1328 return decl
->getSingleDecl();
1331 /* Given the assignment operator in the initialization of a for loop,
1332 * extract the induction variable, i.e., the (integer)variable being
1335 ValueDecl
*PetScan::extract_induction_variable(BinaryOperator
*init
)
1342 lhs
= init
->getLHS();
1343 if (lhs
->getStmtClass() != Stmt::DeclRefExprClass
) {
1348 ref
= cast
<DeclRefExpr
>(lhs
);
1349 decl
= ref
->getDecl();
1350 type
= decl
->getType().getTypePtr();
1352 if (!type
->isIntegerType()) {
1360 /* Given the initialization statement of a for loop and the single
1361 * declaration in this initialization statement,
1362 * extract the induction variable, i.e., the (integer) variable being
1365 VarDecl
*PetScan::extract_induction_variable(Stmt
*init
, Decl
*decl
)
1369 vd
= cast
<VarDecl
>(decl
);
1371 const QualType type
= vd
->getType();
1372 if (!type
->isIntegerType()) {
1377 if (!vd
->getInit()) {
1385 /* Check that op is of the form iv++ or iv--.
1386 * Return a pet_expr representing "1" or "-1" accordingly.
1388 __isl_give pet_expr
*PetScan::extract_unary_increment(
1389 clang::UnaryOperator
*op
, clang::ValueDecl
*iv
)
1395 if (!op
->isIncrementDecrementOp()) {
1400 sub
= op
->getSubExpr();
1401 if (sub
->getStmtClass() != Stmt::DeclRefExprClass
) {
1406 ref
= cast
<DeclRefExpr
>(sub
);
1407 if (ref
->getDecl() != iv
) {
1412 if (op
->isIncrementOp())
1413 v
= isl_val_one(ctx
);
1415 v
= isl_val_negone(ctx
);
1417 return pet_expr_new_int(v
);
1420 /* Check if op is of the form
1424 * and return the increment "expr - iv" as a pet_expr.
1426 __isl_give pet_expr
*PetScan::extract_binary_increment(BinaryOperator
*op
,
1427 clang::ValueDecl
*iv
)
1432 pet_expr
*expr
, *expr_iv
;
1434 if (op
->getOpcode() != BO_Assign
) {
1440 if (lhs
->getStmtClass() != Stmt::DeclRefExprClass
) {
1445 ref
= cast
<DeclRefExpr
>(lhs
);
1446 if (ref
->getDecl() != iv
) {
1451 expr
= extract_expr(op
->getRHS());
1452 expr_iv
= extract_expr(lhs
);
1454 type_size
= get_type_size(iv
->getType(), ast_context
);
1455 return pet_expr_new_binary(type_size
, pet_op_sub
, expr
, expr_iv
);
1458 /* Check that op is of the form iv += cst or iv -= cst
1459 * and return a pet_expr corresponding to cst or -cst accordingly.
1461 __isl_give pet_expr
*PetScan::extract_compound_increment(
1462 CompoundAssignOperator
*op
, clang::ValueDecl
*iv
)
1468 BinaryOperatorKind opcode
;
1470 opcode
= op
->getOpcode();
1471 if (opcode
!= BO_AddAssign
&& opcode
!= BO_SubAssign
) {
1475 if (opcode
== BO_SubAssign
)
1479 if (lhs
->getStmtClass() != Stmt::DeclRefExprClass
) {
1484 ref
= cast
<DeclRefExpr
>(lhs
);
1485 if (ref
->getDecl() != iv
) {
1490 expr
= extract_expr(op
->getRHS());
1492 expr
= pet_expr_new_unary(pet_op_minus
, expr
);
1497 /* Check that the increment of the given for loop increments
1498 * (or decrements) the induction variable "iv" and return
1499 * the increment as a pet_expr if successful.
1501 __isl_give pet_expr
*PetScan::extract_increment(clang::ForStmt
*stmt
,
1504 Stmt
*inc
= stmt
->getInc();
1507 report_missing_increment(stmt
);
1511 if (inc
->getStmtClass() == Stmt::UnaryOperatorClass
)
1512 return extract_unary_increment(cast
<UnaryOperator
>(inc
), iv
);
1513 if (inc
->getStmtClass() == Stmt::CompoundAssignOperatorClass
)
1514 return extract_compound_increment(
1515 cast
<CompoundAssignOperator
>(inc
), iv
);
1516 if (inc
->getStmtClass() == Stmt::BinaryOperatorClass
)
1517 return extract_binary_increment(cast
<BinaryOperator
>(inc
), iv
);
1523 /* Embed the given iteration domain in an extra outer loop
1524 * with induction variable "var".
1525 * If this variable appeared as a parameter in the constraints,
1526 * it is replaced by the new outermost dimension.
1528 static __isl_give isl_set
*embed(__isl_take isl_set
*set
,
1529 __isl_take isl_id
*var
)
1533 set
= isl_set_insert_dims(set
, isl_dim_set
, 0, 1);
1534 pos
= isl_set_find_dim_by_id(set
, isl_dim_param
, var
);
1536 set
= isl_set_equate(set
, isl_dim_param
, pos
, isl_dim_set
, 0);
1537 set
= isl_set_project_out(set
, isl_dim_param
, pos
, 1);
1544 /* Return those elements in the space of "cond" that come after
1545 * (based on "sign") an element in "cond".
1547 static __isl_give isl_set
*after(__isl_take isl_set
*cond
, int sign
)
1549 isl_map
*previous_to_this
;
1552 previous_to_this
= isl_map_lex_lt(isl_set_get_space(cond
));
1554 previous_to_this
= isl_map_lex_gt(isl_set_get_space(cond
));
1556 cond
= isl_set_apply(cond
, previous_to_this
);
1561 /* Create the infinite iteration domain
1563 * { [id] : id >= 0 }
1565 * If "scop" has an affine skip of type pet_skip_later,
1566 * then remove those iterations i that have an earlier iteration
1567 * where the skip condition is satisfied, meaning that iteration i
1569 * Since we are dealing with a loop without loop iterator,
1570 * the skip condition cannot refer to the current loop iterator and
1571 * so effectively, the returned set is of the form
1573 * { [0]; [id] : id >= 1 and not skip }
1575 static __isl_give isl_set
*infinite_domain(__isl_take isl_id
*id
,
1576 struct pet_scop
*scop
)
1578 isl_ctx
*ctx
= isl_id_get_ctx(id
);
1582 domain
= isl_set_nat_universe(isl_space_set_alloc(ctx
, 0, 1));
1583 domain
= isl_set_set_dim_id(domain
, isl_dim_set
, 0, id
);
1585 if (!pet_scop_has_affine_skip(scop
, pet_skip_later
))
1588 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_later
);
1589 skip
= embed(skip
, isl_id_copy(id
));
1590 skip
= isl_set_intersect(skip
, isl_set_copy(domain
));
1591 domain
= isl_set_subtract(domain
, after(skip
, 1));
1596 /* Create an identity affine expression on the space containing "domain",
1597 * which is assumed to be one-dimensional.
1599 static __isl_give isl_aff
*identity_aff(__isl_keep isl_set
*domain
)
1601 isl_local_space
*ls
;
1603 ls
= isl_local_space_from_space(isl_set_get_space(domain
));
1604 return isl_aff_var_on_domain(ls
, isl_dim_set
, 0);
1607 /* Create an affine expression that maps elements
1608 * of a single-dimensional array "id_test" to the previous element
1609 * (according to "inc"), provided this element belongs to "domain".
1610 * That is, create the affine expression
1612 * { id[x] -> id[x - inc] : x - inc in domain }
1614 static __isl_give isl_multi_pw_aff
*map_to_previous(__isl_take isl_id
*id_test
,
1615 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
1618 isl_local_space
*ls
;
1620 isl_multi_pw_aff
*prev
;
1622 space
= isl_set_get_space(domain
);
1623 ls
= isl_local_space_from_space(space
);
1624 aff
= isl_aff_var_on_domain(ls
, isl_dim_set
, 0);
1625 aff
= isl_aff_add_constant_val(aff
, isl_val_neg(inc
));
1626 prev
= isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff
));
1627 domain
= isl_set_preimage_multi_pw_aff(domain
,
1628 isl_multi_pw_aff_copy(prev
));
1629 prev
= isl_multi_pw_aff_intersect_domain(prev
, domain
);
1630 prev
= isl_multi_pw_aff_set_tuple_id(prev
, isl_dim_out
, id_test
);
1635 /* Add an implication to "scop" expressing that if an element of
1636 * virtual array "id_test" has value "satisfied" then all previous elements
1637 * of this array also have that value. The set of previous elements
1638 * is bounded by "domain". If "sign" is negative then the iterator
1639 * is decreasing and we express that all subsequent array elements
1640 * (but still defined previously) have the same value.
1642 static struct pet_scop
*add_implication(struct pet_scop
*scop
,
1643 __isl_take isl_id
*id_test
, __isl_take isl_set
*domain
, int sign
,
1649 domain
= isl_set_set_tuple_id(domain
, id_test
);
1650 space
= isl_set_get_space(domain
);
1652 map
= isl_map_lex_ge(space
);
1654 map
= isl_map_lex_le(space
);
1655 map
= isl_map_intersect_range(map
, domain
);
1656 scop
= pet_scop_add_implication(scop
, map
, satisfied
);
1661 /* Add a filter to "scop" that imposes that it is only executed
1662 * when the variable identified by "id_test" has a zero value
1663 * for all previous iterations of "domain".
1665 * In particular, add a filter that imposes that the array
1666 * has a zero value at the previous iteration of domain and
1667 * add an implication that implies that it then has that
1668 * value for all previous iterations.
1670 static struct pet_scop
*scop_add_break(struct pet_scop
*scop
,
1671 __isl_take isl_id
*id_test
, __isl_take isl_set
*domain
,
1672 __isl_take isl_val
*inc
)
1674 isl_multi_pw_aff
*prev
;
1675 int sign
= isl_val_sgn(inc
);
1677 prev
= map_to_previous(isl_id_copy(id_test
), isl_set_copy(domain
), inc
);
1678 scop
= add_implication(scop
, id_test
, domain
, sign
, 0);
1679 scop
= pet_scop_filter(scop
, prev
, 0);
1684 /* Construct a pet_scop for an infinite loop around the given body
1685 * within the context "pc".
1687 * We extract a pet_scop for the body and then embed it in a loop with
1696 * If the body contains any break, then it is taken into
1697 * account in infinite_domain (if the skip condition is affine)
1698 * or in scop_add_break (if the skip condition is not affine).
1700 * If we were only able to extract part of the body, then simply
1703 struct pet_scop
*PetScan::extract_infinite_loop(Stmt
*body
,
1704 __isl_keep pet_context
*pc
)
1706 isl_id
*id
, *id_test
;
1709 struct pet_scop
*scop
;
1712 scop
= extract(body
, pc
);
1718 id
= isl_id_alloc(ctx
, "t", NULL
);
1719 domain
= infinite_domain(isl_id_copy(id
), scop
);
1720 ident
= identity_aff(domain
);
1722 has_var_break
= pet_scop_has_var_skip(scop
, pet_skip_later
);
1724 id_test
= pet_scop_get_skip_id(scop
, pet_skip_later
);
1726 scop
= pet_scop_embed(scop
, isl_set_copy(domain
),
1727 isl_aff_copy(ident
), ident
, id
);
1729 scop
= scop_add_break(scop
, id_test
, domain
, isl_val_one(ctx
));
1731 isl_set_free(domain
);
1736 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
1741 * within the context "pc".
1743 struct pet_scop
*PetScan::extract_infinite_for(ForStmt
*stmt
,
1744 __isl_keep pet_context
*pc
)
1746 struct pet_scop
*scop
;
1748 pc
= pet_context_copy(pc
);
1749 clear_assignments
clear(pc
);
1750 clear
.TraverseStmt(stmt
->getBody());
1752 scop
= extract_infinite_loop(stmt
->getBody(), pc
);
1753 pet_context_free(pc
);
1757 /* Add an array with the given extent (range of "index") to the list
1758 * of arrays in "scop" and return the extended pet_scop.
1759 * The array is marked as attaining values 0 and 1 only and
1760 * as each element being assigned at most once.
1762 static struct pet_scop
*scop_add_array(struct pet_scop
*scop
,
1763 __isl_keep isl_multi_pw_aff
*index
, clang::ASTContext
&ast_ctx
)
1765 int int_size
= ast_ctx
.getTypeInfo(ast_ctx
.IntTy
).first
/ 8;
1767 return pet_scop_add_boolean_array(scop
, isl_multi_pw_aff_copy(index
),
1771 /* Construct a pet_scop for a while loop of the form
1776 * within the context "pc".
1777 * In particular, construct a scop for an infinite loop around body and
1778 * intersect the domain with the affine expression.
1779 * Note that this intersection may result in an empty loop.
1781 struct pet_scop
*PetScan::extract_affine_while(__isl_take isl_pw_aff
*pa
,
1782 Stmt
*body
, __isl_take pet_context
*pc
)
1784 struct pet_scop
*scop
;
1788 valid
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
1789 dom
= isl_pw_aff_non_zero_set(pa
);
1790 scop
= extract_infinite_loop(body
, pc
);
1791 scop
= pet_scop_restrict(scop
, isl_set_params(dom
));
1792 scop
= pet_scop_restrict_context(scop
, isl_set_params(valid
));
1794 pet_context_free(pc
);
1798 /* Construct a scop for a while, given the scops for the condition
1799 * and the body, the filter identifier and the iteration domain of
1802 * In particular, the scop for the condition is filtered to depend
1803 * on "id_test" evaluating to true for all previous iterations
1804 * of the loop, while the scop for the body is filtered to depend
1805 * on "id_test" evaluating to true for all iterations up to the
1806 * current iteration.
1807 * The actual filter only imposes that this virtual array has
1808 * value one on the previous or the current iteration.
1809 * The fact that this condition also applies to the previous
1810 * iterations is enforced by an implication.
1812 * These filtered scops are then combined into a single scop.
1814 * "sign" is positive if the iterator increases and negative
1817 static struct pet_scop
*scop_add_while(struct pet_scop
*scop_cond
,
1818 struct pet_scop
*scop_body
, __isl_take isl_id
*id_test
,
1819 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
1821 isl_ctx
*ctx
= isl_set_get_ctx(domain
);
1823 isl_multi_pw_aff
*test_index
;
1824 isl_multi_pw_aff
*prev
;
1825 int sign
= isl_val_sgn(inc
);
1826 struct pet_scop
*scop
;
1828 prev
= map_to_previous(isl_id_copy(id_test
), isl_set_copy(domain
), inc
);
1829 scop_cond
= pet_scop_filter(scop_cond
, prev
, 1);
1831 space
= isl_space_map_from_set(isl_set_get_space(domain
));
1832 test_index
= isl_multi_pw_aff_identity(space
);
1833 test_index
= isl_multi_pw_aff_set_tuple_id(test_index
, isl_dim_out
,
1834 isl_id_copy(id_test
));
1835 scop_body
= pet_scop_filter(scop_body
, test_index
, 1);
1837 scop
= pet_scop_add_seq(ctx
, scop_cond
, scop_body
);
1838 scop
= add_implication(scop
, id_test
, domain
, sign
, 1);
1843 /* Check if the while loop is of the form
1845 * while (affine expression)
1848 * If so, call extract_affine_while to construct a scop.
1850 * Otherwise, extract the body and pass control to extract_while
1851 * to extend the iteration domain with an infinite loop.
1852 * If we were only able to extract part of the body, then simply
1855 * "pc" is the context in which the affine expressions in the scop are created.
1857 struct pet_scop
*PetScan::extract(WhileStmt
*stmt
, __isl_keep pet_context
*pc
)
1860 int test_nr
, stmt_nr
;
1862 struct pet_scop
*scop
, *scop_body
;
1864 cond
= stmt
->getCond();
1870 pc
= pet_context_copy(pc
);
1871 clear_assignments
clear(pc
);
1872 clear
.TraverseStmt(stmt
->getBody());
1874 pa
= extract_condition(cond
, pc
);
1876 return extract_affine_while(pa
, stmt
->getBody(), pc
);
1878 if (!allow_nested
) {
1880 pet_context_free(pc
);
1886 scop_body
= extract(stmt
->getBody(), pc
);
1888 pet_context_free(pc
);
1892 return extract_while(cond
, test_nr
, stmt_nr
, scop_body
, NULL
, pc
);
1895 /* Construct a generic while scop, with iteration domain
1896 * { [t] : t >= 0 } around "scop_body" within the context "pc".
1897 * The scop consists of two parts,
1898 * one for evaluating the condition "cond" and one for the body.
1899 * "test_nr" is the sequence number of the virtual test variable that contains
1900 * the result of the condition and "stmt_nr" is the sequence number
1901 * of the statement that evaluates the condition.
1902 * If "scop_inc" is not NULL, then it is added at the end of the body,
1903 * after replacing any skip conditions resulting from continue statements
1904 * by the skip conditions resulting from break statements (if any).
1906 * The schedule is adjusted to reflect that the condition is evaluated
1907 * before the body is executed and the body is filtered to depend
1908 * on the result of the condition evaluating to true on all iterations
1909 * up to the current iteration, while the evaluation of the condition itself
1910 * is filtered to depend on the result of the condition evaluating to true
1911 * on all previous iterations.
1912 * The context of the scop representing the body is dropped
1913 * because we don't know how many times the body will be executed,
1916 * If the body contains any break, then it is taken into
1917 * account in infinite_domain (if the skip condition is affine)
1918 * or in scop_add_break (if the skip condition is not affine).
1920 struct pet_scop
*PetScan::extract_while(Expr
*cond
, int test_nr
, int stmt_nr
,
1921 struct pet_scop
*scop_body
, struct pet_scop
*scop_inc
,
1922 __isl_take pet_context
*pc
)
1924 isl_id
*id
, *id_test
, *id_break_test
;
1927 isl_multi_pw_aff
*test_index
;
1928 struct pet_scop
*scop
;
1931 test_index
= pet_create_test_index(ctx
, test_nr
);
1932 scop
= extract_non_affine_condition(cond
, stmt_nr
,
1933 isl_multi_pw_aff_copy(test_index
), pc
);
1934 scop
= scop_add_array(scop
, test_index
, ast_context
);
1935 id_test
= isl_multi_pw_aff_get_tuple_id(test_index
, isl_dim_out
);
1936 isl_multi_pw_aff_free(test_index
);
1938 id
= isl_id_alloc(ctx
, "t", NULL
);
1939 domain
= infinite_domain(isl_id_copy(id
), scop_body
);
1940 ident
= identity_aff(domain
);
1942 has_var_break
= pet_scop_has_var_skip(scop_body
, pet_skip_later
);
1944 id_break_test
= pet_scop_get_skip_id(scop_body
, pet_skip_later
);
1946 scop
= pet_scop_prefix(scop
, 0);
1947 scop
= pet_scop_embed(scop
, isl_set_copy(domain
), isl_aff_copy(ident
),
1948 isl_aff_copy(ident
), isl_id_copy(id
));
1949 scop_body
= pet_scop_reset_context(scop_body
);
1950 scop_body
= pet_scop_prefix(scop_body
, 1);
1952 scop_inc
= pet_scop_prefix(scop_inc
, 2);
1953 if (pet_scop_has_skip(scop_body
, pet_skip_later
)) {
1954 isl_multi_pw_aff
*skip
;
1955 skip
= pet_scop_get_skip(scop_body
, pet_skip_later
);
1956 scop_body
= pet_scop_set_skip(scop_body
,
1957 pet_skip_now
, skip
);
1959 pet_scop_reset_skip(scop_body
, pet_skip_now
);
1960 scop_body
= pet_scop_add_seq(ctx
, scop_body
, scop_inc
);
1962 scop_body
= pet_scop_embed(scop_body
, isl_set_copy(domain
),
1963 isl_aff_copy(ident
), ident
, id
);
1965 if (has_var_break
) {
1966 scop
= scop_add_break(scop
, isl_id_copy(id_break_test
),
1967 isl_set_copy(domain
), isl_val_one(ctx
));
1968 scop_body
= scop_add_break(scop_body
, id_break_test
,
1969 isl_set_copy(domain
), isl_val_one(ctx
));
1971 scop
= scop_add_while(scop
, scop_body
, id_test
, domain
,
1974 pet_context_free(pc
);
1978 /* Check whether "cond" expresses a simple loop bound
1979 * on the only set dimension.
1980 * In particular, if "up" is set then "cond" should contain only
1981 * upper bounds on the set dimension.
1982 * Otherwise, it should contain only lower bounds.
1984 static bool is_simple_bound(__isl_keep isl_set
*cond
, __isl_keep isl_val
*inc
)
1986 if (isl_val_is_pos(inc
))
1987 return !isl_set_dim_has_any_lower_bound(cond
, isl_dim_set
, 0);
1989 return !isl_set_dim_has_any_upper_bound(cond
, isl_dim_set
, 0);
1992 /* Extend a condition on a given iteration of a loop to one that
1993 * imposes the same condition on all previous iterations.
1994 * "domain" expresses the lower [upper] bound on the iterations
1995 * when inc is positive [negative].
1997 * In particular, we construct the condition (when inc is positive)
1999 * forall i' : (domain(i') and i' <= i) => cond(i')
2001 * which is equivalent to
2003 * not exists i' : domain(i') and i' <= i and not cond(i')
2005 * We construct this set by negating cond, applying a map
2007 * { [i'] -> [i] : domain(i') and i' <= i }
2009 * and then negating the result again.
2011 static __isl_give isl_set
*valid_for_each_iteration(__isl_take isl_set
*cond
,
2012 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
2014 isl_map
*previous_to_this
;
2016 if (isl_val_is_pos(inc
))
2017 previous_to_this
= isl_map_lex_le(isl_set_get_space(domain
));
2019 previous_to_this
= isl_map_lex_ge(isl_set_get_space(domain
));
2021 previous_to_this
= isl_map_intersect_domain(previous_to_this
, domain
);
2023 cond
= isl_set_complement(cond
);
2024 cond
= isl_set_apply(cond
, previous_to_this
);
2025 cond
= isl_set_complement(cond
);
2032 /* Construct a domain of the form
2034 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
2036 static __isl_give isl_set
*strided_domain(__isl_take isl_id
*id
,
2037 __isl_take isl_pw_aff
*init
, __isl_take isl_val
*inc
)
2043 init
= isl_pw_aff_insert_dims(init
, isl_dim_in
, 0, 1);
2044 dim
= isl_pw_aff_get_domain_space(init
);
2045 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
2046 aff
= isl_aff_add_coefficient_val(aff
, isl_dim_in
, 0, inc
);
2047 init
= isl_pw_aff_add(init
, isl_pw_aff_from_aff(aff
));
2049 dim
= isl_space_set_alloc(isl_pw_aff_get_ctx(init
), 1, 1);
2050 dim
= isl_space_set_dim_id(dim
, isl_dim_param
, 0, id
);
2051 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
2052 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_param
, 0, 1);
2054 set
= isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff
), init
);
2056 set
= isl_set_lower_bound_si(set
, isl_dim_set
, 0, 0);
2058 return isl_set_params(set
);
2061 /* Assuming "cond" represents a bound on a loop where the loop
2062 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2065 * Under the given assumptions, wrapping is only possible if "cond" allows
2066 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2067 * increasing iterator and 0 in case of a decreasing iterator.
2069 static bool can_wrap(__isl_keep isl_set
*cond
, ValueDecl
*iv
,
2070 __isl_keep isl_val
*inc
)
2077 test
= isl_set_copy(cond
);
2079 ctx
= isl_set_get_ctx(test
);
2080 if (isl_val_is_neg(inc
))
2081 limit
= isl_val_zero(ctx
);
2083 limit
= isl_val_int_from_ui(ctx
, get_type_size(iv
));
2084 limit
= isl_val_2exp(limit
);
2085 limit
= isl_val_sub_ui(limit
, 1);
2088 test
= isl_set_fix_val(cond
, isl_dim_set
, 0, limit
);
2089 cw
= !isl_set_is_empty(test
);
2095 /* Given a one-dimensional space, construct the following affine expression
2098 * { [v] -> [v mod 2^width] }
2100 * where width is the number of bits used to represent the values
2101 * of the unsigned variable "iv".
2103 static __isl_give isl_aff
*compute_wrapping(__isl_take isl_space
*dim
,
2110 ctx
= isl_space_get_ctx(dim
);
2111 mod
= isl_val_int_from_ui(ctx
, get_type_size(iv
));
2112 mod
= isl_val_2exp(mod
);
2114 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
2115 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, 0, 1);
2116 aff
= isl_aff_mod_val(aff
, mod
);
2121 /* Project out the parameter "id" from "set".
2123 static __isl_give isl_set
*set_project_out_by_id(__isl_take isl_set
*set
,
2124 __isl_keep isl_id
*id
)
2128 pos
= isl_set_find_dim_by_id(set
, isl_dim_param
, id
);
2130 set
= isl_set_project_out(set
, isl_dim_param
, pos
, 1);
2135 /* Compute the set of parameters for which "set1" is a subset of "set2".
2137 * set1 is a subset of set2 if
2139 * forall i in set1 : i in set2
2143 * not exists i in set1 and i not in set2
2147 * not exists i in set1 \ set2
2149 static __isl_give isl_set
*enforce_subset(__isl_take isl_set
*set1
,
2150 __isl_take isl_set
*set2
)
2152 return isl_set_complement(isl_set_params(isl_set_subtract(set1
, set2
)));
2155 /* Compute the set of parameter values for which "cond" holds
2156 * on the next iteration for each element of "dom".
2158 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2159 * and then compute the set of parameters for which the result is a subset
2162 static __isl_give isl_set
*valid_on_next(__isl_take isl_set
*cond
,
2163 __isl_take isl_set
*dom
, __isl_take isl_val
*inc
)
2169 space
= isl_set_get_space(dom
);
2170 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
2171 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, 0, 1);
2172 aff
= isl_aff_add_constant_val(aff
, inc
);
2173 next
= isl_map_from_basic_map(isl_basic_map_from_aff(aff
));
2175 dom
= isl_set_apply(dom
, next
);
2177 return enforce_subset(dom
, cond
);
2180 /* Extract the for loop "stmt" as a while loop within the context "pc".
2181 * "iv" is the loop iterator. "init" is the initialization.
2182 * "inc" is the increment.
2184 * That is, the for loop has the form
2186 * for (iv = init; cond; iv += inc)
2197 * except that the skips resulting from any continue statements
2198 * in body do not apply to the increment, but are replaced by the skips
2199 * resulting from break statements.
2201 * If "iv" is declared in the for loop, then it is killed before
2202 * and after the loop.
2204 struct pet_scop
*PetScan::extract_non_affine_for(ForStmt
*stmt
, ValueDecl
*iv
,
2205 __isl_take pet_expr
*init
, __isl_take pet_expr
*inc
,
2206 __isl_take pet_context
*pc
)
2209 int test_nr
, stmt_nr
;
2211 struct pet_scop
*scop_init
, *scop_inc
, *scop
, *scop_body
;
2213 struct pet_array
*array
;
2214 struct pet_scop
*scop_kill
;
2216 if (!allow_nested
) {
2218 pet_context_free(pc
);
2222 pc
= pet_context_mark_assigned(pc
, create_decl_id(ctx
, iv
));
2224 declared
= !initialization_assignment(stmt
->getInit());
2226 expr_iv
= extract_access_expr(iv
);
2227 expr_iv
= mark_write(expr_iv
);
2228 type_size
= pet_expr_get_type_size(expr_iv
);
2229 init
= pet_expr_new_binary(type_size
, pet_op_assign
, expr_iv
, init
);
2230 scop_init
= extract(init
, stmt
->getInit()->getSourceRange(), false, pc
);
2231 scop_init
= pet_scop_prefix(scop_init
, declared
);
2235 scop_body
= extract(stmt
->getBody(), pc
);
2237 pet_scop_free(scop_init
);
2238 pet_context_free(pc
);
2242 expr_iv
= extract_access_expr(iv
);
2243 expr_iv
= mark_write(expr_iv
);
2244 type_size
= pet_expr_get_type_size(expr_iv
);
2245 inc
= pet_expr_new_binary(type_size
, pet_op_add_assign
, expr_iv
, inc
);
2246 scop_inc
= extract(inc
, stmt
->getInc()->getSourceRange(), false, pc
);
2248 pet_scop_free(scop_init
);
2249 pet_scop_free(scop_body
);
2250 pet_context_free(pc
);
2254 scop
= extract_while(stmt
->getCond(), test_nr
, stmt_nr
, scop_body
,
2255 scop_inc
, pet_context_copy(pc
));
2257 scop
= pet_scop_prefix(scop
, declared
+ 1);
2258 scop
= pet_scop_add_seq(ctx
, scop_init
, scop
);
2261 pet_context_free(pc
);
2265 array
= extract_array(ctx
, iv
, NULL
, pc
);
2267 array
->declared
= 1;
2268 scop_kill
= kill(stmt
, array
, pc
);
2269 scop_kill
= pet_scop_prefix(scop_kill
, 0);
2270 scop
= pet_scop_add_seq(ctx
, scop_kill
, scop
);
2271 scop_kill
= kill(stmt
, array
, pc
);
2272 scop_kill
= pet_scop_add_array(scop_kill
, array
);
2273 scop_kill
= pet_scop_prefix(scop_kill
, 3);
2274 scop
= pet_scop_add_seq(ctx
, scop
, scop_kill
);
2276 pet_context_free(pc
);
2280 /* Construct a pet_scop for a for statement within the context "pc".
2281 * The for loop is required to be of one of the following forms
2283 * for (i = init; condition; ++i)
2284 * for (i = init; condition; --i)
2285 * for (i = init; condition; i += constant)
2286 * for (i = init; condition; i -= constant)
2288 * The initialization of the for loop should either be an assignment
2289 * of a static affine value to an integer variable, or a declaration
2290 * of such a variable with initialization.
2292 * If the initialization or the increment do not satisfy the above
2293 * conditions, i.e., if the initialization is not static affine
2294 * or the increment is not constant, then the for loop is extracted
2295 * as a while loop instead.
2297 * The condition is allowed to contain nested accesses, provided
2298 * they are not being written to inside the body of the loop.
2299 * Otherwise, or if the condition is otherwise non-affine, the for loop is
2300 * essentially treated as a while loop, with iteration domain
2301 * { [i] : i >= init }.
2303 * We extract a pet_scop for the body and then embed it in a loop with
2304 * iteration domain and schedule
2306 * { [i] : i >= init and condition' }
2311 * { [i] : i <= init and condition' }
2314 * Where condition' is equal to condition if the latter is
2315 * a simple upper [lower] bound and a condition that is extended
2316 * to apply to all previous iterations otherwise.
2318 * If the condition is non-affine, then we drop the condition from the
2319 * iteration domain and instead create a separate statement
2320 * for evaluating the condition. The body is then filtered to depend
2321 * on the result of the condition evaluating to true on all iterations
2322 * up to the current iteration, while the evaluation the condition itself
2323 * is filtered to depend on the result of the condition evaluating to true
2324 * on all previous iterations.
2325 * The context of the scop representing the body is dropped
2326 * because we don't know how many times the body will be executed,
2329 * If the stride of the loop is not 1, then "i >= init" is replaced by
2331 * (exists a: i = init + stride * a and a >= 0)
2333 * If the loop iterator i is unsigned, then wrapping may occur.
2334 * We therefore use a virtual iterator instead that does not wrap.
2335 * However, the condition in the code applies
2336 * to the wrapped value, so we need to change condition(i)
2337 * into condition([i % 2^width]). Similarly, we replace all accesses
2338 * to the original iterator by the wrapping of the virtual iterator.
2339 * Note that there may be no need to perform this final wrapping
2340 * if the loop condition (after wrapping) satisfies certain conditions.
2341 * However, the is_simple_bound condition is not enough since it doesn't
2342 * check if there even is an upper bound.
2344 * Wrapping on unsigned iterators can be avoided entirely if
2345 * loop condition is simple, the loop iterator is incremented
2346 * [decremented] by one and the last value before wrapping cannot
2347 * possibly satisfy the loop condition.
2349 * Before extracting a pet_scop from the body we remove all
2350 * assignments in "pc" to variables that are assigned
2351 * somewhere in the body of the loop.
2353 * Valid parameters for a for loop are those for which the initial
2354 * value itself, the increment on each domain iteration and
2355 * the condition on both the initial value and
2356 * the result of incrementing the iterator for each iteration of the domain
2358 * If the loop condition is non-affine, then we only consider validity
2359 * of the initial value.
2361 * If the body contains any break, then we keep track of it in "skip"
2362 * (if the skip condition is affine) or it is handled in scop_add_break
2363 * (if the skip condition is not affine).
2364 * Note that the affine break condition needs to be considered with
2365 * respect to previous iterations in the virtual domain (if any).
2367 * If we were only able to extract part of the body, then simply
2370 struct pet_scop
*PetScan::extract_for(ForStmt
*stmt
, __isl_keep pet_context
*pc
)
2372 BinaryOperator
*ass
;
2377 isl_local_space
*ls
;
2380 isl_set
*cond
= NULL
;
2381 isl_set
*skip
= NULL
;
2382 isl_id
*id
, *id_test
= NULL
, *id_break_test
;
2383 struct pet_scop
*scop
, *scop_cond
= NULL
;
2389 bool has_affine_break
;
2391 isl_aff
*wrap
= NULL
;
2392 isl_pw_aff
*pa
, *pa_inc
, *init_val
;
2393 isl_set
*valid_init
;
2394 isl_set
*valid_cond
;
2395 isl_set
*valid_cond_init
;
2396 isl_set
*valid_cond_next
;
2399 pet_expr
*pe_init
, *pe_inc
;
2400 pet_context
*pc_init_val
;
2402 if (!stmt
->getInit() && !stmt
->getCond() && !stmt
->getInc())
2403 return extract_infinite_for(stmt
, pc
);
2405 init
= stmt
->getInit();
2410 if ((ass
= initialization_assignment(init
)) != NULL
) {
2411 iv
= extract_induction_variable(ass
);
2414 lhs
= ass
->getLHS();
2415 rhs
= ass
->getRHS();
2416 } else if ((decl
= initialization_declaration(init
)) != NULL
) {
2417 VarDecl
*var
= extract_induction_variable(init
, decl
);
2421 rhs
= var
->getInit();
2422 lhs
= create_DeclRefExpr(var
);
2424 unsupported(stmt
->getInit());
2428 id
= create_decl_id(ctx
, iv
);
2430 pc
= pet_context_copy(pc
);
2431 pc
= pet_context_clear_value(pc
, isl_id_copy(id
));
2432 clear_assignments
clear(pc
);
2433 clear
.TraverseStmt(stmt
->getBody());
2435 pe_init
= extract_expr(rhs
);
2436 pe_inc
= extract_increment(stmt
, iv
);
2437 pc_init_val
= pet_context_copy(pc
);
2438 pc_init_val
= pet_context_mark_unknown(pc_init_val
, isl_id_copy(id
));
2439 init_val
= pet_expr_extract_affine(pe_init
, pc_init_val
);
2440 pet_context_free(pc_init_val
);
2441 pa_inc
= pet_expr_extract_affine(pe_inc
, pc
);
2442 inc
= pet_extract_cst(pa_inc
);
2443 if (!pe_init
|| !pe_inc
|| !inc
|| isl_val_is_nan(inc
) ||
2444 isl_pw_aff_involves_nan(pa_inc
) ||
2445 isl_pw_aff_involves_nan(init_val
)) {
2448 isl_pw_aff_free(pa_inc
);
2449 isl_pw_aff_free(init_val
);
2450 if (pe_init
&& pe_inc
&& !(pa_inc
&& !inc
))
2451 return extract_non_affine_for(stmt
, iv
,
2452 pe_init
, pe_inc
, pc
);
2453 pet_expr_free(pe_init
);
2454 pet_expr_free(pe_inc
);
2455 pet_context_free(pc
);
2458 pet_expr_free(pe_init
);
2459 pet_expr_free(pe_inc
);
2461 pa
= try_extract_nested_condition(stmt
->getCond(), pc
);
2462 if (allow_nested
&& (!pa
|| pet_nested_any_in_pw_aff(pa
)))
2465 scop
= extract(stmt
->getBody(), pc
);
2468 isl_pw_aff_free(init_val
);
2469 isl_pw_aff_free(pa_inc
);
2470 isl_pw_aff_free(pa
);
2472 pet_context_free(pc
);
2476 valid_inc
= isl_pw_aff_domain(pa_inc
);
2478 is_unsigned
= iv
->getType()->isUnsignedIntegerType();
2480 has_affine_break
= scop
&&
2481 pet_scop_has_affine_skip(scop
, pet_skip_later
);
2482 if (has_affine_break
)
2483 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_later
);
2484 has_var_break
= scop
&& pet_scop_has_var_skip(scop
, pet_skip_later
);
2486 id_break_test
= pet_scop_get_skip_id(scop
, pet_skip_later
);
2488 if (pa
&& !is_nested_allowed(pa
, scop
)) {
2489 isl_pw_aff_free(pa
);
2493 if (!allow_nested
&& !pa
)
2494 pa
= extract_condition(stmt
->getCond(), pc
);
2495 valid_cond
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
2496 cond
= isl_pw_aff_non_zero_set(pa
);
2497 if (allow_nested
&& !cond
) {
2498 isl_multi_pw_aff
*test_index
;
2499 int save_n_stmt
= n_stmt
;
2500 test_index
= pet_create_test_index(ctx
, n_test
++);
2502 scop_cond
= extract_non_affine_condition(stmt
->getCond(),
2503 n_stmt
++, isl_multi_pw_aff_copy(test_index
), pc
);
2504 n_stmt
= save_n_stmt
;
2505 scop_cond
= scop_add_array(scop_cond
, test_index
, ast_context
);
2506 id_test
= isl_multi_pw_aff_get_tuple_id(test_index
,
2508 isl_multi_pw_aff_free(test_index
);
2509 scop_cond
= pet_scop_prefix(scop_cond
, 0);
2510 scop
= pet_scop_reset_context(scop
);
2511 scop
= pet_scop_prefix(scop
, 1);
2512 cond
= isl_set_universe(isl_space_set_alloc(ctx
, 0, 0));
2515 cond
= embed(cond
, isl_id_copy(id
));
2516 skip
= embed(skip
, isl_id_copy(id
));
2517 valid_cond
= isl_set_coalesce(valid_cond
);
2518 valid_cond
= embed(valid_cond
, isl_id_copy(id
));
2519 valid_inc
= embed(valid_inc
, isl_id_copy(id
));
2520 is_one
= isl_val_is_one(inc
) || isl_val_is_negone(inc
);
2521 is_virtual
= is_unsigned
&& (!is_one
|| can_wrap(cond
, iv
, inc
));
2523 valid_cond_init
= enforce_subset(
2524 isl_map_range(isl_map_from_pw_aff(isl_pw_aff_copy(init_val
))),
2525 isl_set_copy(valid_cond
));
2526 if (is_one
&& !is_virtual
) {
2527 isl_pw_aff_free(init_val
);
2528 pa
= extract_comparison(isl_val_is_pos(inc
) ? BO_GE
: BO_LE
,
2529 lhs
, rhs
, init
, pc
);
2530 valid_init
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
2531 valid_init
= set_project_out_by_id(valid_init
, id
);
2532 domain
= isl_pw_aff_non_zero_set(pa
);
2534 valid_init
= isl_pw_aff_domain(isl_pw_aff_copy(init_val
));
2535 domain
= strided_domain(isl_id_copy(id
), init_val
,
2539 domain
= embed(domain
, isl_id_copy(id
));
2542 wrap
= compute_wrapping(isl_set_get_space(cond
), iv
);
2543 rev_wrap
= isl_map_from_aff(isl_aff_copy(wrap
));
2544 rev_wrap
= isl_map_reverse(rev_wrap
);
2545 cond
= isl_set_apply(cond
, isl_map_copy(rev_wrap
));
2546 skip
= isl_set_apply(skip
, isl_map_copy(rev_wrap
));
2547 valid_cond
= isl_set_apply(valid_cond
, isl_map_copy(rev_wrap
));
2548 valid_inc
= isl_set_apply(valid_inc
, rev_wrap
);
2550 is_simple
= is_simple_bound(cond
, inc
);
2552 cond
= isl_set_gist(cond
, isl_set_copy(domain
));
2553 is_simple
= is_simple_bound(cond
, inc
);
2556 cond
= valid_for_each_iteration(cond
,
2557 isl_set_copy(domain
), isl_val_copy(inc
));
2558 domain
= isl_set_intersect(domain
, cond
);
2559 if (has_affine_break
) {
2560 skip
= isl_set_intersect(skip
, isl_set_copy(domain
));
2561 skip
= after(skip
, isl_val_sgn(inc
));
2562 domain
= isl_set_subtract(domain
, skip
);
2564 domain
= isl_set_set_dim_id(domain
, isl_dim_set
, 0, isl_id_copy(id
));
2565 ls
= isl_local_space_from_space(isl_set_get_space(domain
));
2566 sched
= isl_aff_var_on_domain(ls
, isl_dim_set
, 0);
2567 if (isl_val_is_neg(inc
))
2568 sched
= isl_aff_neg(sched
);
2570 valid_cond_next
= valid_on_next(valid_cond
, isl_set_copy(domain
),
2572 valid_inc
= enforce_subset(isl_set_copy(domain
), valid_inc
);
2575 wrap
= identity_aff(domain
);
2577 scop_cond
= pet_scop_embed(scop_cond
, isl_set_copy(domain
),
2578 isl_aff_copy(sched
), isl_aff_copy(wrap
), isl_id_copy(id
));
2579 scop
= pet_scop_embed(scop
, isl_set_copy(domain
), sched
, wrap
, id
);
2580 scop
= resolve_nested(scop
);
2582 scop
= scop_add_break(scop
, id_break_test
, isl_set_copy(domain
),
2585 scop
= scop_add_while(scop_cond
, scop
, id_test
, domain
,
2587 isl_set_free(valid_inc
);
2589 scop
= pet_scop_restrict_context(scop
, valid_inc
);
2590 scop
= pet_scop_restrict_context(scop
, valid_cond_next
);
2591 scop
= pet_scop_restrict_context(scop
, valid_cond_init
);
2592 isl_set_free(domain
);
2597 scop
= pet_scop_restrict_context(scop
, isl_set_params(valid_init
));
2599 pet_context_free(pc
);
2603 /* Try and construct a pet_scop corresponding to a compound statement
2604 * within the context "pc".
2606 * "skip_declarations" is set if we should skip initial declarations
2607 * in the children of the compound statements. This then implies
2608 * that this sequence of children should not be treated as a block
2609 * since the initial statements may be skipped.
2611 struct pet_scop
*PetScan::extract(CompoundStmt
*stmt
,
2612 __isl_keep pet_context
*pc
, bool skip_declarations
)
2614 return extract(stmt
->children(),
2615 !skip_declarations
, skip_declarations
, pc
);
2618 /* For each nested access parameter in "space",
2619 * construct a corresponding pet_expr, place it in args and
2620 * record its position in "param2pos".
2621 * "n_arg" is the number of elements that are already in args.
2622 * The position recorded in "param2pos" takes this number into account.
2623 * If the pet_expr corresponding to a parameter is identical to
2624 * the pet_expr corresponding to an earlier parameter, then these two
2625 * parameters are made to refer to the same element in args.
2627 * Return the final number of elements in args or -1 if an error has occurred.
2629 int PetScan::extract_nested(__isl_keep isl_space
*space
,
2630 int n_arg
, pet_expr
**args
, std::map
<int,int> ¶m2pos
)
2634 nparam
= isl_space_dim(space
, isl_dim_param
);
2635 for (int i
= 0; i
< nparam
; ++i
) {
2637 isl_id
*id
= isl_space_get_dim_id(space
, isl_dim_param
, i
);
2639 if (!pet_nested_in_id(id
)) {
2644 args
[n_arg
] = pet_nested_extract_expr(id
);
2649 for (j
= 0; j
< n_arg
; ++j
)
2650 if (pet_expr_is_equal(args
[j
], args
[n_arg
]))
2654 pet_expr_free(args
[n_arg
]);
2658 param2pos
[i
] = n_arg
++;
2664 /* For each nested access parameter in the access relations in "expr",
2665 * construct a corresponding pet_expr, append it to the arguments of "expr"
2666 * and record its position in "param2pos" (relative to the initial
2667 * number of arguments).
2668 * n is the number of nested access parameters.
2670 __isl_give pet_expr
*PetScan::extract_nested(__isl_take pet_expr
*expr
, int n
,
2671 std::map
<int,int> ¶m2pos
)
2677 args
= isl_calloc_array(ctx
, pet_expr
*, n
);
2679 return pet_expr_free(expr
);
2681 n_arg
= pet_expr_get_n_arg(expr
);
2682 space
= pet_expr_access_get_parameter_space(expr
);
2683 n
= extract_nested(space
, 0, args
, param2pos
);
2684 isl_space_free(space
);
2687 expr
= pet_expr_free(expr
);
2689 expr
= pet_expr_set_n_arg(expr
, n_arg
+ n
);
2691 for (i
= 0; i
< n
; ++i
)
2692 expr
= pet_expr_set_arg(expr
, n_arg
+ i
, args
[i
]);
2698 /* Are "expr1" and "expr2" both array accesses such that
2699 * the access relation of "expr1" is a subset of that of "expr2"?
2700 * Only take into account the first "n_arg" arguments.
2702 static int is_sub_access(__isl_keep pet_expr
*expr1
, __isl_keep pet_expr
*expr2
,
2706 isl_map
*access1
, *access2
;
2710 if (!expr1
|| !expr2
)
2712 if (pet_expr_get_type(expr1
) != pet_expr_access
)
2714 if (pet_expr_get_type(expr2
) != pet_expr_access
)
2716 if (pet_expr_is_affine(expr1
))
2718 if (pet_expr_is_affine(expr2
))
2720 n1
= pet_expr_get_n_arg(expr1
);
2723 n2
= pet_expr_get_n_arg(expr2
);
2728 for (i
= 0; i
< n1
; ++i
) {
2729 pet_expr
*arg1
, *arg2
;
2731 arg1
= pet_expr_get_arg(expr1
, i
);
2732 arg2
= pet_expr_get_arg(expr2
, i
);
2733 equal
= pet_expr_is_equal(arg1
, arg2
);
2734 pet_expr_free(arg1
);
2735 pet_expr_free(arg2
);
2736 if (equal
< 0 || !equal
)
2739 id1
= pet_expr_access_get_id(expr1
);
2740 id2
= pet_expr_access_get_id(expr2
);
2748 access1
= pet_expr_access_get_access(expr1
);
2749 access2
= pet_expr_access_get_access(expr2
);
2750 is_subset
= isl_map_is_subset(access1
, access2
);
2751 isl_map_free(access1
);
2752 isl_map_free(access2
);
2757 /* Mark self dependences among the arguments of "expr" starting at "first".
2758 * These arguments have already been added to the list of arguments
2759 * but are not yet referenced directly from the index expression.
2760 * Instead, they are still referenced through parameters encoding
2763 * In particular, if "expr" is a read access, then check the arguments
2764 * starting at "first" to see if "expr" accesses a subset of
2765 * the elements accessed by the argument, or under more restrictive conditions.
2766 * If so, then this nested access can be removed from the constraints
2767 * governing the outer access. There is no point in restricting
2768 * accesses to an array if in order to evaluate the restriction,
2769 * we have to access the same elements (or more).
2771 * Rather than removing the argument at this point (which would
2772 * complicate the resolution of the other nested accesses), we simply
2773 * mark it here by replacing it by a NaN pet_expr.
2774 * These NaNs are then later removed in remove_marked_self_dependences.
2776 static __isl_give pet_expr
*mark_self_dependences(__isl_take pet_expr
*expr
,
2781 if (pet_expr_access_is_write(expr
))
2784 n
= pet_expr_get_n_arg(expr
);
2785 for (int i
= first
; i
< n
; ++i
) {
2789 arg
= pet_expr_get_arg(expr
, i
);
2790 mark
= is_sub_access(expr
, arg
, first
);
2793 return pet_expr_free(expr
);
2797 arg
= pet_expr_new_int(isl_val_nan(pet_expr_get_ctx(expr
)));
2798 expr
= pet_expr_set_arg(expr
, i
, arg
);
2804 /* Is "expr" a NaN integer expression?
2806 static int expr_is_nan(__isl_keep pet_expr
*expr
)
2811 if (pet_expr_get_type(expr
) != pet_expr_int
)
2814 v
= pet_expr_int_get_val(expr
);
2815 is_nan
= isl_val_is_nan(v
);
2821 /* Check if we have marked any self dependences (as NaNs)
2822 * in mark_self_dependences and remove them here.
2823 * It is safe to project them out since these arguments
2824 * can at most be referenced from the condition of the access relation,
2825 * but do not appear in the index expression.
2826 * "dim" is the dimension of the iteration domain.
2828 static __isl_give pet_expr
*remove_marked_self_dependences(
2829 __isl_take pet_expr
*expr
, int dim
, int first
)
2833 n
= pet_expr_get_n_arg(expr
);
2834 for (int i
= n
- 1; i
>= first
; --i
) {
2838 arg
= pet_expr_get_arg(expr
, i
);
2839 is_nan
= expr_is_nan(arg
);
2843 expr
= pet_expr_access_project_out_arg(expr
, dim
, i
);
2849 /* Look for parameters in any access relation in "expr" that
2850 * refer to nested accesses. In particular, these are
2851 * parameters with name "__pet_expr".
2853 * If there are any such parameters, then the domain of the index
2854 * expression and the access relation, which is either [] or
2855 * [[] -> [a_1,...,a_m]] at this point, is replaced by [[] -> [t_1,...,t_n]] or
2856 * [[] -> [a_1,...,a_m,t_1,...,t_n]], with m the original number of arguments
2857 * (n_arg) and n the number of these parameters
2858 * (after identifying identical nested accesses).
2860 * This transformation is performed in several steps.
2861 * We first extract the arguments in extract_nested.
2862 * param2pos maps the original parameter position to the position
2863 * of the argument beyond the initial (n_arg) number of arguments.
2864 * Then we move these parameters to input dimensions.
2865 * t2pos maps the positions of these temporary input dimensions
2866 * to the positions of the corresponding arguments.
2867 * Finally, we express these temporary dimensions in terms of the domain
2868 * [[] -> [a_1,...,a_m,t_1,...,t_n]] and precompose index expression and access
2869 * relations with this function.
2871 __isl_give pet_expr
*PetScan::resolve_nested(__isl_take pet_expr
*expr
)
2876 isl_local_space
*ls
;
2879 std::map
<int,int> param2pos
;
2880 std::map
<int,int> t2pos
;
2885 n_arg
= pet_expr_get_n_arg(expr
);
2886 for (int i
= 0; i
< n_arg
; ++i
) {
2888 arg
= pet_expr_get_arg(expr
, i
);
2889 arg
= resolve_nested(arg
);
2890 expr
= pet_expr_set_arg(expr
, i
, arg
);
2893 if (pet_expr_get_type(expr
) != pet_expr_access
)
2896 space
= pet_expr_access_get_parameter_space(expr
);
2897 n
= pet_nested_n_in_space(space
);
2898 isl_space_free(space
);
2902 expr
= extract_nested(expr
, n
, param2pos
);
2906 expr
= pet_expr_access_align_params(expr
);
2907 expr
= mark_self_dependences(expr
, n_arg
);
2912 space
= pet_expr_access_get_parameter_space(expr
);
2913 nparam
= isl_space_dim(space
, isl_dim_param
);
2914 for (int i
= nparam
- 1; i
>= 0; --i
) {
2915 isl_id
*id
= isl_space_get_dim_id(space
, isl_dim_param
, i
);
2916 if (!pet_nested_in_id(id
)) {
2921 expr
= pet_expr_access_move_dims(expr
,
2922 isl_dim_in
, n_arg
+ n
, isl_dim_param
, i
, 1);
2923 t2pos
[n
] = n_arg
+ param2pos
[i
];
2928 isl_space_free(space
);
2930 space
= pet_expr_access_get_parameter_space(expr
);
2931 space
= isl_space_set_from_params(space
);
2932 space
= isl_space_add_dims(space
, isl_dim_set
,
2933 pet_expr_get_n_arg(expr
));
2934 space
= isl_space_wrap(isl_space_from_range(space
));
2935 ls
= isl_local_space_from_space(isl_space_copy(space
));
2936 space
= isl_space_from_domain(space
);
2937 space
= isl_space_add_dims(space
, isl_dim_out
, n_arg
+ n
);
2938 ma
= isl_multi_aff_zero(space
);
2940 for (int i
= 0; i
< n_arg
; ++i
) {
2941 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
2943 ma
= isl_multi_aff_set_aff(ma
, i
, aff
);
2945 for (int i
= 0; i
< n
; ++i
) {
2946 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
2947 isl_dim_set
, t2pos
[i
]);
2948 ma
= isl_multi_aff_set_aff(ma
, n_arg
+ i
, aff
);
2950 isl_local_space_free(ls
);
2952 expr
= pet_expr_access_pullback_multi_aff(expr
, ma
);
2954 expr
= remove_marked_self_dependences(expr
, 0, n_arg
);
2959 /* Return the file offset of the expansion location of "Loc".
2961 static unsigned getExpansionOffset(SourceManager
&SM
, SourceLocation Loc
)
2963 return SM
.getFileOffset(SM
.getExpansionLoc(Loc
));
2966 #ifdef HAVE_FINDLOCATIONAFTERTOKEN
2968 /* Return a SourceLocation for the location after the first semicolon
2969 * after "loc". If Lexer::findLocationAfterToken is available, we simply
2970 * call it and also skip trailing spaces and newline.
2972 static SourceLocation
location_after_semi(SourceLocation loc
, SourceManager
&SM
,
2973 const LangOptions
&LO
)
2975 return Lexer::findLocationAfterToken(loc
, tok::semi
, SM
, LO
, true);
2980 /* Return a SourceLocation for the location after the first semicolon
2981 * after "loc". If Lexer::findLocationAfterToken is not available,
2982 * we look in the underlying character data for the first semicolon.
2984 static SourceLocation
location_after_semi(SourceLocation loc
, SourceManager
&SM
,
2985 const LangOptions
&LO
)
2988 const char *s
= SM
.getCharacterData(loc
);
2990 semi
= strchr(s
, ';');
2992 return SourceLocation();
2993 return loc
.getFileLocWithOffset(semi
+ 1 - s
);
2998 /* If the token at "loc" is the first token on the line, then return
2999 * a location referring to the start of the line.
3000 * Otherwise, return "loc".
3002 * This function is used to extend a scop to the start of the line
3003 * if the first token of the scop is also the first token on the line.
3005 * We look for the first token on the line. If its location is equal to "loc",
3006 * then the latter is the location of the first token on the line.
3008 static SourceLocation
move_to_start_of_line_if_first_token(SourceLocation loc
,
3009 SourceManager
&SM
, const LangOptions
&LO
)
3011 std::pair
<FileID
, unsigned> file_offset_pair
;
3012 llvm::StringRef file
;
3015 SourceLocation token_loc
, line_loc
;
3018 loc
= SM
.getExpansionLoc(loc
);
3019 col
= SM
.getExpansionColumnNumber(loc
);
3020 line_loc
= loc
.getLocWithOffset(1 - col
);
3021 file_offset_pair
= SM
.getDecomposedLoc(line_loc
);
3022 file
= SM
.getBufferData(file_offset_pair
.first
, NULL
);
3023 pos
= file
.data() + file_offset_pair
.second
;
3025 Lexer
lexer(SM
.getLocForStartOfFile(file_offset_pair
.first
), LO
,
3026 file
.begin(), pos
, file
.end());
3027 lexer
.LexFromRawLexer(tok
);
3028 token_loc
= tok
.getLocation();
3030 if (token_loc
== loc
)
3036 /* Update start and end of "scop" to include the region covered by "range".
3037 * If "skip_semi" is set, then we assume "range" is followed by
3038 * a semicolon and also include this semicolon.
3040 struct pet_scop
*PetScan::update_scop_start_end(struct pet_scop
*scop
,
3041 SourceRange range
, bool skip_semi
)
3043 SourceLocation loc
= range
.getBegin();
3044 SourceManager
&SM
= PP
.getSourceManager();
3045 const LangOptions
&LO
= PP
.getLangOpts();
3046 unsigned start
, end
;
3048 loc
= move_to_start_of_line_if_first_token(loc
, SM
, LO
);
3049 start
= getExpansionOffset(SM
, loc
);
3050 loc
= range
.getEnd();
3052 loc
= location_after_semi(loc
, SM
, LO
);
3054 loc
= PP
.getLocForEndOfToken(loc
);
3055 end
= getExpansionOffset(SM
, loc
);
3057 scop
= pet_scop_update_start_end(scop
, start
, end
);
3061 /* Convert a top-level pet_expr to a pet_scop with one statement
3062 * within the context "pc".
3063 * This mainly involves resolving nested expression parameters
3064 * and setting the name of the iteration space.
3065 * The name is given by "label" if it is non-NULL. Otherwise,
3066 * it is of the form S_<n_stmt>.
3067 * start and end of the pet_scop are derived from "range" and "skip_semi".
3068 * In particular, if "skip_semi" is set then the semicolon following "range"
3071 struct pet_scop
*PetScan::extract(__isl_take pet_expr
*expr
, SourceRange range
,
3072 bool skip_semi
, __isl_keep pet_context
*pc
, __isl_take isl_id
*label
)
3074 struct pet_stmt
*ps
;
3075 struct pet_scop
*scop
;
3076 SourceLocation loc
= range
.getBegin();
3077 int line
= PP
.getSourceManager().getExpansionLineNumber(loc
);
3079 expr
= pet_expr_plug_in_args(expr
, pc
);
3080 expr
= resolve_nested(expr
);
3081 expr
= pet_expr_resolve_assume(expr
, pc
);
3082 ps
= pet_stmt_from_pet_expr(line
, label
, n_stmt
++, expr
);
3083 scop
= pet_scop_from_pet_stmt(ctx
, ps
);
3085 scop
= update_scop_start_end(scop
, range
, skip_semi
);
3089 /* Check whether "expr" is an affine constraint within the context "pc".
3091 bool PetScan::is_affine_condition(Expr
*expr
, __isl_keep pet_context
*pc
)
3095 cond
= extract_condition(expr
, pc
);
3096 isl_pw_aff_free(cond
);
3098 return cond
!= NULL
;
3101 /* Check if we can extract a condition from "expr" within the context "pc".
3102 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
3103 * If allow_nested is set, then the condition may involve parameters
3104 * corresponding to nested accesses.
3105 * We turn on autodetection so that we won't generate any warnings.
3107 __isl_give isl_pw_aff
*PetScan::try_extract_nested_condition(Expr
*expr
,
3108 __isl_keep pet_context
*pc
)
3111 int save_autodetect
= options
->autodetect
;
3112 bool save_nesting
= nesting_enabled
;
3114 options
->autodetect
= 1;
3115 nesting_enabled
= allow_nested
;
3116 cond
= extract_condition(expr
, pc
);
3118 options
->autodetect
= save_autodetect
;
3119 nesting_enabled
= save_nesting
;
3124 /* If the top-level expression of "stmt" is an assignment, then
3125 * return that assignment as a BinaryOperator.
3126 * Otherwise return NULL.
3128 static BinaryOperator
*top_assignment_or_null(Stmt
*stmt
)
3130 BinaryOperator
*ass
;
3134 if (stmt
->getStmtClass() != Stmt::BinaryOperatorClass
)
3137 ass
= cast
<BinaryOperator
>(stmt
);
3138 if(ass
->getOpcode() != BO_Assign
)
3144 /* Check if the given if statement is a conditional assignement
3145 * with a non-affine condition within the context "pc".
3146 * If so, construct a pet_scop corresponding to this conditional assignment.
3147 * Otherwise return NULL.
3149 * In particular we check if "stmt" is of the form
3156 * where a is some array or scalar access.
3157 * The constructed pet_scop then corresponds to the expression
3159 * a = condition ? f(...) : g(...)
3161 * All access relations in f(...) are intersected with condition
3162 * while all access relation in g(...) are intersected with the complement.
3164 struct pet_scop
*PetScan::extract_conditional_assignment(IfStmt
*stmt
,
3165 __isl_keep pet_context
*pc
)
3167 BinaryOperator
*ass_then
, *ass_else
;
3168 pet_expr
*write_then
, *write_else
;
3169 isl_set
*cond
, *comp
;
3170 isl_multi_pw_aff
*index
;
3174 pet_expr
*pe_cond
, *pe_then
, *pe_else
, *pe
;
3175 bool save_nesting
= nesting_enabled
;
3177 if (!options
->detect_conditional_assignment
)
3180 ass_then
= top_assignment_or_null(stmt
->getThen());
3181 ass_else
= top_assignment_or_null(stmt
->getElse());
3183 if (!ass_then
|| !ass_else
)
3186 if (is_affine_condition(stmt
->getCond(), pc
))
3189 write_then
= extract_access_expr(ass_then
->getLHS());
3190 write_else
= extract_access_expr(ass_else
->getLHS());
3192 equal
= pet_expr_is_equal(write_then
, write_else
);
3193 pet_expr_free(write_else
);
3194 if (equal
< 0 || !equal
) {
3195 pet_expr_free(write_then
);
3199 nesting_enabled
= allow_nested
;
3200 pa
= extract_condition(stmt
->getCond(), pc
);
3201 nesting_enabled
= save_nesting
;
3202 cond
= isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa
));
3203 comp
= isl_pw_aff_zero_set(isl_pw_aff_copy(pa
));
3204 index
= isl_multi_pw_aff_from_pw_aff(pa
);
3206 pe_cond
= pet_expr_from_index(index
);
3208 pe_then
= extract_expr(ass_then
->getRHS());
3209 pe_then
= pet_expr_restrict(pe_then
, cond
);
3210 pe_else
= extract_expr(ass_else
->getRHS());
3211 pe_else
= pet_expr_restrict(pe_else
, comp
);
3213 pe
= pet_expr_new_ternary(pe_cond
, pe_then
, pe_else
);
3214 write_then
= pet_expr_access_set_write(write_then
, 1);
3215 write_then
= pet_expr_access_set_read(write_then
, 0);
3216 type_size
= get_type_size(ass_then
->getType(), ast_context
);
3217 pe
= pet_expr_new_binary(type_size
, pet_op_assign
, write_then
, pe
);
3218 return extract(pe
, stmt
->getSourceRange(), false, pc
);
3221 /* Create a pet_scop with a single statement with name S_<stmt_nr>,
3222 * evaluating "cond" and writing the result to a virtual scalar,
3223 * as expressed by "index".
3224 * Do so within the context "pc".
3226 struct pet_scop
*PetScan::extract_non_affine_condition(Expr
*cond
, int stmt_nr
,
3227 __isl_take isl_multi_pw_aff
*index
, __isl_keep pet_context
*pc
)
3229 pet_expr
*expr
, *write
;
3230 struct pet_stmt
*ps
;
3231 SourceLocation loc
= cond
->getLocStart();
3232 int line
= PP
.getSourceManager().getExpansionLineNumber(loc
);
3234 write
= pet_expr_from_index(index
);
3235 write
= pet_expr_access_set_write(write
, 1);
3236 write
= pet_expr_access_set_read(write
, 0);
3237 expr
= extract_expr(cond
);
3238 expr
= pet_expr_plug_in_args(expr
, pc
);
3239 expr
= resolve_nested(expr
);
3240 expr
= pet_expr_new_binary(1, pet_op_assign
, write
, expr
);
3241 ps
= pet_stmt_from_pet_expr(line
, NULL
, stmt_nr
, expr
);
3242 return pet_scop_from_pet_stmt(ctx
, ps
);
3246 static __isl_give pet_expr
*embed_access(__isl_take pet_expr
*expr
,
3250 /* Precompose the access relation and the index expression associated
3251 * to "expr" with the function pointed to by "user",
3252 * thereby embedding the access relation in the domain of this function.
3253 * The initial domain of the access relation and the index expression
3254 * is the zero-dimensional domain.
3256 static __isl_give pet_expr
*embed_access(__isl_take pet_expr
*expr
, void *user
)
3258 isl_multi_aff
*ma
= (isl_multi_aff
*) user
;
3260 return pet_expr_access_pullback_multi_aff(expr
, isl_multi_aff_copy(ma
));
3263 /* Precompose all access relations in "expr" with "ma", thereby
3264 * embedding them in the domain of "ma".
3266 static __isl_give pet_expr
*embed(__isl_take pet_expr
*expr
,
3267 __isl_keep isl_multi_aff
*ma
)
3269 return pet_expr_map_access(expr
, &embed_access
, ma
);
3272 /* For each nested access parameter in the domain of "stmt",
3273 * construct a corresponding pet_expr, place it before the original
3274 * elements in stmt->args and record its position in "param2pos".
3275 * n is the number of nested access parameters.
3277 struct pet_stmt
*PetScan::extract_nested(struct pet_stmt
*stmt
, int n
,
3278 std::map
<int,int> ¶m2pos
)
3285 n_arg
= stmt
->n_arg
;
3286 args
= isl_calloc_array(ctx
, pet_expr
*, n
+ n_arg
);
3290 space
= isl_set_get_space(stmt
->domain
);
3291 n_arg
= extract_nested(space
, 0, args
, param2pos
);
3292 isl_space_free(space
);
3297 for (i
= 0; i
< stmt
->n_arg
; ++i
)
3298 args
[n_arg
+ i
] = stmt
->args
[i
];
3301 stmt
->n_arg
+= n_arg
;
3306 for (i
= 0; i
< n
; ++i
)
3307 pet_expr_free(args
[i
]);
3310 pet_stmt_free(stmt
);
3314 /* Check whether any of the arguments i of "stmt" starting at position "n"
3315 * is equal to one of the first "n" arguments j.
3316 * If so, combine the constraints on arguments i and j and remove
3319 static struct pet_stmt
*remove_duplicate_arguments(struct pet_stmt
*stmt
, int n
)
3328 if (n
== stmt
->n_arg
)
3331 map
= isl_set_unwrap(stmt
->domain
);
3333 for (i
= stmt
->n_arg
- 1; i
>= n
; --i
) {
3334 for (j
= 0; j
< n
; ++j
)
3335 if (pet_expr_is_equal(stmt
->args
[i
], stmt
->args
[j
]))
3340 map
= isl_map_equate(map
, isl_dim_out
, i
, isl_dim_out
, j
);
3341 map
= isl_map_project_out(map
, isl_dim_out
, i
, 1);
3343 pet_expr_free(stmt
->args
[i
]);
3344 for (j
= i
; j
+ 1 < stmt
->n_arg
; ++j
)
3345 stmt
->args
[j
] = stmt
->args
[j
+ 1];
3349 stmt
->domain
= isl_map_wrap(map
);
3354 pet_stmt_free(stmt
);
3358 /* Look for parameters in the iteration domain of "stmt" that
3359 * refer to nested accesses. In particular, these are
3360 * parameters with name "__pet_expr".
3362 * If there are any such parameters, then as many extra variables
3363 * (after identifying identical nested accesses) are inserted in the
3364 * range of the map wrapped inside the domain, before the original variables.
3365 * If the original domain is not a wrapped map, then a new wrapped
3366 * map is created with zero output dimensions.
3367 * The parameters are then equated to the corresponding output dimensions
3368 * and subsequently projected out, from the iteration domain,
3369 * the schedule and the access relations.
3370 * For each of the output dimensions, a corresponding argument
3371 * expression is inserted. Initially they are created with
3372 * a zero-dimensional domain, so they have to be embedded
3373 * in the current iteration domain.
3374 * param2pos maps the position of the parameter to the position
3375 * of the corresponding output dimension in the wrapped map.
3377 struct pet_stmt
*PetScan::resolve_nested(struct pet_stmt
*stmt
)
3385 std::map
<int,int> param2pos
;
3390 n
= pet_nested_n_in_set(stmt
->domain
);
3394 n_arg
= stmt
->n_arg
;
3395 stmt
= extract_nested(stmt
, n
, param2pos
);
3399 n
= stmt
->n_arg
- n_arg
;
3400 nparam
= isl_set_dim(stmt
->domain
, isl_dim_param
);
3401 if (isl_set_is_wrapping(stmt
->domain
))
3402 map
= isl_set_unwrap(stmt
->domain
);
3404 map
= isl_map_from_domain(stmt
->domain
);
3405 map
= isl_map_insert_dims(map
, isl_dim_out
, 0, n
);
3407 for (int i
= nparam
- 1; i
>= 0; --i
) {
3410 if (!pet_nested_in_map(map
, i
))
3413 id
= pet_expr_access_get_id(stmt
->args
[param2pos
[i
]]);
3414 map
= isl_map_set_dim_id(map
, isl_dim_out
, param2pos
[i
], id
);
3415 map
= isl_map_equate(map
, isl_dim_param
, i
, isl_dim_out
,
3417 map
= isl_map_project_out(map
, isl_dim_param
, i
, 1);
3420 stmt
->domain
= isl_map_wrap(map
);
3422 space
= isl_space_unwrap(isl_set_get_space(stmt
->domain
));
3423 space
= isl_space_from_domain(isl_space_domain(space
));
3424 ma
= isl_multi_aff_zero(space
);
3425 for (int pos
= 0; pos
< n
; ++pos
)
3426 stmt
->args
[pos
] = embed(stmt
->args
[pos
], ma
);
3427 isl_multi_aff_free(ma
);
3429 stmt
= pet_stmt_remove_nested_parameters(stmt
);
3430 stmt
= remove_duplicate_arguments(stmt
, n
);
3435 /* For each statement in "scop", move the parameters that correspond
3436 * to nested access into the ranges of the domains and create
3437 * corresponding argument expressions.
3439 struct pet_scop
*PetScan::resolve_nested(struct pet_scop
*scop
)
3444 for (int i
= 0; i
< scop
->n_stmt
; ++i
) {
3445 scop
->stmts
[i
] = resolve_nested(scop
->stmts
[i
]);
3446 if (!scop
->stmts
[i
])
3452 pet_scop_free(scop
);
3456 /* Given an access expression "expr", is the variable accessed by
3457 * "expr" assigned anywhere inside "scop"?
3459 static bool is_assigned(__isl_keep pet_expr
*expr
, pet_scop
*scop
)
3461 bool assigned
= false;
3464 id
= pet_expr_access_get_id(expr
);
3465 assigned
= pet_scop_writes(scop
, id
);
3471 /* Are all nested access parameters in "pa" allowed given "scop".
3472 * In particular, is none of them written by anywhere inside "scop".
3474 * If "scop" has any skip conditions, then no nested access parameters
3475 * are allowed. In particular, if there is any nested access in a guard
3476 * for a piece of code containing a "continue", then we want to introduce
3477 * a separate statement for evaluating this guard so that we can express
3478 * that the result is false for all previous iterations.
3480 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff
*pa
, pet_scop
*scop
)
3487 if (!pet_nested_any_in_pw_aff(pa
))
3490 if (pet_scop_has_skip(scop
, pet_skip_now
))
3493 nparam
= isl_pw_aff_dim(pa
, isl_dim_param
);
3494 for (int i
= 0; i
< nparam
; ++i
) {
3495 isl_id
*id
= isl_pw_aff_get_dim_id(pa
, isl_dim_param
, i
);
3499 if (!pet_nested_in_id(id
)) {
3504 expr
= pet_nested_extract_expr(id
);
3505 allowed
= pet_expr_get_type(expr
) == pet_expr_access
&&
3506 !is_assigned(expr
, scop
);
3508 pet_expr_free(expr
);
3518 /* Construct a pet_scop for a non-affine if statement within the context "pc".
3520 * We create a separate statement that writes the result
3521 * of the non-affine condition to a virtual scalar.
3522 * A constraint requiring the value of this virtual scalar to be one
3523 * is added to the iteration domains of the then branch.
3524 * Similarly, a constraint requiring the value of this virtual scalar
3525 * to be zero is added to the iteration domains of the else branch, if any.
3526 * We adjust the schedules to ensure that the virtual scalar is written
3527 * before it is read.
3529 * If there are any breaks or continues in the then and/or else
3530 * branches, then we may have to compute a new skip condition.
3531 * This is handled using a pet_skip_info object.
3532 * On initialization, the object checks if skip conditions need
3533 * to be computed. If so, it does so in pet_skip_info_if_extract_index and
3534 * adds them in pet_skip_info_if_add.
3536 struct pet_scop
*PetScan::extract_non_affine_if(Expr
*cond
,
3537 struct pet_scop
*scop_then
, struct pet_scop
*scop_else
,
3538 bool have_else
, int stmt_id
, __isl_take pet_context
*pc
)
3540 struct pet_scop
*scop
;
3541 isl_multi_pw_aff
*test_index
;
3543 int save_n_stmt
= n_stmt
;
3545 test_index
= pet_create_test_index(ctx
, n_test
++);
3547 scop
= extract_non_affine_condition(cond
, n_stmt
++,
3548 isl_multi_pw_aff_copy(test_index
), pc
);
3549 n_stmt
= save_n_stmt
;
3550 scop
= scop_add_array(scop
, test_index
, ast_context
);
3553 pet_skip_info_if_init(&skip
, ctx
, scop_then
, scop_else
, have_else
, 0);
3554 int_size
= ast_context
.getTypeInfo(ast_context
.IntTy
).first
/ 8;
3555 pet_skip_info_if_extract_index(&skip
, test_index
, int_size
,
3558 scop
= pet_scop_prefix(scop
, 0);
3559 scop_then
= pet_scop_prefix(scop_then
, 1);
3560 scop_then
= pet_scop_filter(scop_then
,
3561 isl_multi_pw_aff_copy(test_index
), 1);
3563 scop_else
= pet_scop_prefix(scop_else
, 1);
3564 scop_else
= pet_scop_filter(scop_else
, test_index
, 0);
3565 scop_then
= pet_scop_add_par(ctx
, scop_then
, scop_else
);
3567 isl_multi_pw_aff_free(test_index
);
3569 scop
= pet_scop_add_seq(ctx
, scop
, scop_then
);
3571 scop
= pet_skip_info_if_add(&skip
, scop
, 2);
3573 pet_context_free(pc
);
3577 /* Construct a pet_scop for an if statement within the context "pc".
3579 * If the condition fits the pattern of a conditional assignment,
3580 * then it is handled by extract_conditional_assignment.
3581 * Otherwise, we do the following.
3583 * If the condition is affine, then the condition is added
3584 * to the iteration domains of the then branch, while the
3585 * opposite of the condition in added to the iteration domains
3586 * of the else branch, if any.
3587 * We allow the condition to be dynamic, i.e., to refer to
3588 * scalars or array elements that may be written to outside
3589 * of the given if statement. These nested accesses are then represented
3590 * as output dimensions in the wrapping iteration domain.
3591 * If it is also written _inside_ the then or else branch, then
3592 * we treat the condition as non-affine.
3593 * As explained in extract_non_affine_if, this will introduce
3594 * an extra statement.
3595 * For aesthetic reasons, we want this statement to have a statement
3596 * number that is lower than those of the then and else branches.
3597 * In order to evaluate if we will need such a statement, however, we
3598 * first construct scops for the then and else branches.
3599 * We therefore reserve a statement number if we might have to
3600 * introduce such an extra statement.
3602 * If the condition is not affine, then the scop is created in
3603 * extract_non_affine_if.
3605 * If there are any breaks or continues in the then and/or else
3606 * branches, then we may have to compute a new skip condition.
3607 * This is handled using a pet_skip_info object.
3608 * On initialization, the object checks if skip conditions need
3609 * to be computed. If so, it does so in pet_skip_info_if_extract_cond and
3610 * adds them in pet_skip_info_if_add.
3612 struct pet_scop
*PetScan::extract(IfStmt
*stmt
, __isl_keep pet_context
*pc
)
3614 struct pet_scop
*scop_then
, *scop_else
= NULL
, *scop
;
3621 pc
= pet_context_copy(pc
);
3622 clear_assignments
clear(pc
);
3623 clear
.TraverseStmt(stmt
->getThen());
3624 if (stmt
->getElse())
3625 clear
.TraverseStmt(stmt
->getElse());
3627 scop
= extract_conditional_assignment(stmt
, pc
);
3629 pet_context_free(pc
);
3633 cond
= try_extract_nested_condition(stmt
->getCond(), pc
);
3634 if (allow_nested
&& (!cond
|| pet_nested_any_in_pw_aff(cond
)))
3637 scop_then
= extract(stmt
->getThen(), pc
);
3639 if (stmt
->getElse()) {
3640 scop_else
= extract(stmt
->getElse(), pc
);
3641 if (options
->autodetect
) {
3642 if (scop_then
&& !scop_else
) {
3644 isl_pw_aff_free(cond
);
3645 pet_context_free(pc
);
3648 if (!scop_then
&& scop_else
) {
3650 isl_pw_aff_free(cond
);
3651 pet_context_free(pc
);
3658 (!is_nested_allowed(cond
, scop_then
) ||
3659 (stmt
->getElse() && !is_nested_allowed(cond
, scop_else
)))) {
3660 isl_pw_aff_free(cond
);
3663 if (allow_nested
&& !cond
)
3664 return extract_non_affine_if(stmt
->getCond(), scop_then
,
3665 scop_else
, stmt
->getElse(), stmt_id
, pc
);
3668 cond
= extract_condition(stmt
->getCond(), pc
);
3671 pet_skip_info_if_init(&skip
, ctx
, scop_then
, scop_else
,
3672 stmt
->getElse() != NULL
, 1);
3673 pet_skip_info_if_extract_cond(&skip
, cond
, int_size
, &n_stmt
, &n_test
);
3675 valid
= isl_pw_aff_domain(isl_pw_aff_copy(cond
));
3676 set
= isl_pw_aff_non_zero_set(cond
);
3677 scop
= pet_scop_restrict(scop_then
, isl_set_params(isl_set_copy(set
)));
3679 if (stmt
->getElse()) {
3680 set
= isl_set_subtract(isl_set_copy(valid
), set
);
3681 scop_else
= pet_scop_restrict(scop_else
, isl_set_params(set
));
3682 scop
= pet_scop_add_par(ctx
, scop
, scop_else
);
3685 scop
= resolve_nested(scop
);
3686 scop
= pet_scop_restrict_context(scop
, isl_set_params(valid
));
3688 if (pet_skip_info_has_skip(&skip
))
3689 scop
= pet_scop_prefix(scop
, 0);
3690 scop
= pet_skip_info_if_add(&skip
, scop
, 1);
3692 pet_context_free(pc
);
3696 /* Try and construct a pet_scop for a label statement within the context "pc".
3697 * We currently only allow labels on expression statements.
3699 struct pet_scop
*PetScan::extract(LabelStmt
*stmt
, __isl_keep pet_context
*pc
)
3704 sub
= stmt
->getSubStmt();
3705 if (!isa
<Expr
>(sub
)) {
3710 label
= isl_id_alloc(ctx
, stmt
->getName(), NULL
);
3712 return extract(extract_expr(cast
<Expr
>(sub
)), stmt
->getSourceRange(),
3716 /* Return a one-dimensional multi piecewise affine expression that is equal
3717 * to the constant 1 and is defined over a zero-dimensional domain.
3719 static __isl_give isl_multi_pw_aff
*one_mpa(isl_ctx
*ctx
)
3722 isl_local_space
*ls
;
3725 space
= isl_space_set_alloc(ctx
, 0, 0);
3726 ls
= isl_local_space_from_space(space
);
3727 aff
= isl_aff_zero_on_domain(ls
);
3728 aff
= isl_aff_set_constant_si(aff
, 1);
3730 return isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff
));
3733 /* Construct a pet_scop for a continue statement.
3735 * We simply create an empty scop with a universal pet_skip_now
3736 * skip condition. This skip condition will then be taken into
3737 * account by the enclosing loop construct, possibly after
3738 * being incorporated into outer skip conditions.
3740 struct pet_scop
*PetScan::extract(ContinueStmt
*stmt
)
3744 scop
= pet_scop_empty(ctx
);
3748 scop
= pet_scop_set_skip(scop
, pet_skip_now
, one_mpa(ctx
));
3753 /* Construct a pet_scop for a break statement.
3755 * We simply create an empty scop with both a universal pet_skip_now
3756 * skip condition and a universal pet_skip_later skip condition.
3757 * These skip conditions will then be taken into
3758 * account by the enclosing loop construct, possibly after
3759 * being incorporated into outer skip conditions.
3761 struct pet_scop
*PetScan::extract(BreakStmt
*stmt
)
3764 isl_multi_pw_aff
*skip
;
3766 scop
= pet_scop_empty(ctx
);
3770 skip
= one_mpa(ctx
);
3771 scop
= pet_scop_set_skip(scop
, pet_skip_now
,
3772 isl_multi_pw_aff_copy(skip
));
3773 scop
= pet_scop_set_skip(scop
, pet_skip_later
, skip
);
3778 /* Try and construct a pet_scop corresponding to "stmt"
3779 * within the context "pc".
3781 * If "stmt" is a compound statement, then "skip_declarations"
3782 * indicates whether we should skip initial declarations in the
3783 * compound statement.
3785 * If the constructed pet_scop is not a (possibly) partial representation
3786 * of "stmt", we update start and end of the pet_scop to those of "stmt".
3787 * In particular, if skip_declarations is set, then we may have skipped
3788 * declarations inside "stmt" and so the pet_scop may not represent
3789 * the entire "stmt".
3790 * Note that this function may be called with "stmt" referring to the entire
3791 * body of the function, including the outer braces. In such cases,
3792 * skip_declarations will be set and the braces will not be taken into
3793 * account in scop->start and scop->end.
3795 struct pet_scop
*PetScan::extract(Stmt
*stmt
, __isl_keep pet_context
*pc
,
3796 bool skip_declarations
)
3798 struct pet_scop
*scop
;
3800 if (isa
<Expr
>(stmt
))
3801 return extract(extract_expr(cast
<Expr
>(stmt
)),
3802 stmt
->getSourceRange(), true, pc
);
3804 switch (stmt
->getStmtClass()) {
3805 case Stmt::WhileStmtClass
:
3806 scop
= extract(cast
<WhileStmt
>(stmt
), pc
);
3808 case Stmt::ForStmtClass
:
3809 scop
= extract_for(cast
<ForStmt
>(stmt
), pc
);
3811 case Stmt::IfStmtClass
:
3812 scop
= extract(cast
<IfStmt
>(stmt
), pc
);
3814 case Stmt::CompoundStmtClass
:
3815 scop
= extract(cast
<CompoundStmt
>(stmt
), pc
, skip_declarations
);
3817 case Stmt::LabelStmtClass
:
3818 scop
= extract(cast
<LabelStmt
>(stmt
), pc
);
3820 case Stmt::ContinueStmtClass
:
3821 scop
= extract(cast
<ContinueStmt
>(stmt
));
3823 case Stmt::BreakStmtClass
:
3824 scop
= extract(cast
<BreakStmt
>(stmt
));
3826 case Stmt::DeclStmtClass
:
3827 scop
= extract(cast
<DeclStmt
>(stmt
), pc
);
3834 if (partial
|| skip_declarations
)
3837 scop
= update_scop_start_end(scop
, stmt
->getSourceRange(), false);
3842 /* Extract a clone of the kill statement in "scop".
3843 * "scop" is expected to have been created from a DeclStmt
3844 * and should have the kill as its first statement.
3846 struct pet_stmt
*PetScan::extract_kill(struct pet_scop
*scop
)
3849 struct pet_stmt
*stmt
;
3850 isl_multi_pw_aff
*index
;
3856 if (scop
->n_stmt
< 1)
3857 isl_die(ctx
, isl_error_internal
,
3858 "expecting at least one statement", return NULL
);
3859 stmt
= scop
->stmts
[0];
3860 if (!pet_stmt_is_kill(stmt
))
3861 isl_die(ctx
, isl_error_internal
,
3862 "expecting kill statement", return NULL
);
3864 arg
= pet_expr_get_arg(stmt
->body
, 0);
3865 index
= pet_expr_access_get_index(arg
);
3866 access
= pet_expr_access_get_access(arg
);
3868 index
= isl_multi_pw_aff_reset_tuple_id(index
, isl_dim_in
);
3869 access
= isl_map_reset_tuple_id(access
, isl_dim_in
);
3870 kill
= pet_expr_kill_from_access_and_index(access
, index
);
3871 return pet_stmt_from_pet_expr(stmt
->line
, NULL
, n_stmt
++, kill
);
3874 /* Mark all arrays in "scop" as being exposed.
3876 static struct pet_scop
*mark_exposed(struct pet_scop
*scop
)
3880 for (int i
= 0; i
< scop
->n_array
; ++i
)
3881 scop
->arrays
[i
]->exposed
= 1;
3885 /* Try and construct a pet_scop corresponding to (part of)
3886 * a sequence of statements within the context "pc".
3888 * "block" is set if the sequence respresents the children of
3889 * a compound statement.
3890 * "skip_declarations" is set if we should skip initial declarations
3891 * in the sequence of statements.
3893 * After extracting a statement, we update "pc"
3894 * based on the top-level assignments in the statement
3895 * so that we can exploit them in subsequent statements in the same block.
3897 * If there are any breaks or continues in the individual statements,
3898 * then we may have to compute a new skip condition.
3899 * This is handled using a pet_skip_info object.
3900 * On initialization, the object checks if skip conditions need
3901 * to be computed. If so, it does so in pet_skip_info_seq_extract and
3902 * adds them in pet_skip_info_seq_add.
3904 * If "block" is set, then we need to insert kill statements at
3905 * the end of the block for any array that has been declared by
3906 * one of the statements in the sequence. Each of these declarations
3907 * results in the construction of a kill statement at the place
3908 * of the declaration, so we simply collect duplicates of
3909 * those kill statements and append these duplicates to the constructed scop.
3911 * If "block" is not set, then any array declared by one of the statements
3912 * in the sequence is marked as being exposed.
3914 * If autodetect is set, then we allow the extraction of only a subrange
3915 * of the sequence of statements. However, if there is at least one statement
3916 * for which we could not construct a scop and the final range contains
3917 * either no statements or at least one kill, then we discard the entire
3920 struct pet_scop
*PetScan::extract(StmtRange stmt_range
, bool block
,
3921 bool skip_declarations
, __isl_keep pet_context
*pc
)
3927 bool partial_range
= false;
3928 set
<struct pet_stmt
*> kills
;
3929 set
<struct pet_stmt
*>::iterator it
;
3931 int_size
= ast_context
.getTypeInfo(ast_context
.IntTy
).first
/ 8;
3933 pc
= pet_context_copy(pc
);
3934 scop
= pet_scop_empty(ctx
);
3935 for (i
= stmt_range
.first
, j
= 0; i
!= stmt_range
.second
; ++i
, ++j
) {
3937 struct pet_scop
*scop_i
;
3939 if (scop
->n_stmt
== 0 && skip_declarations
&&
3940 child
->getStmtClass() == Stmt::DeclStmtClass
)
3943 scop_i
= extract(child
, pc
);
3944 if (scop
->n_stmt
!= 0 && partial
) {
3945 pet_scop_free(scop_i
);
3948 pc
= handle_writes(scop_i
, pc
);
3950 pet_skip_info_seq_init(&skip
, ctx
, scop
, scop_i
);
3951 pet_skip_info_seq_extract(&skip
, int_size
, &n_stmt
, &n_test
);
3952 if (pet_skip_info_has_skip(&skip
))
3953 scop_i
= pet_scop_prefix(scop_i
, 0);
3954 if (scop_i
&& child
->getStmtClass() == Stmt::DeclStmtClass
) {
3956 kills
.insert(extract_kill(scop_i
));
3958 scop_i
= mark_exposed(scop_i
);
3960 scop_i
= pet_scop_prefix(scop_i
, j
);
3961 if (options
->autodetect
) {
3963 scop
= pet_scop_add_seq(ctx
, scop
, scop_i
);
3965 partial_range
= true;
3966 if (scop
->n_stmt
!= 0 && !scop_i
)
3969 scop
= pet_scop_add_seq(ctx
, scop
, scop_i
);
3972 scop
= pet_skip_info_seq_add(&skip
, scop
, j
);
3974 if (partial
|| !scop
)
3978 for (it
= kills
.begin(); it
!= kills
.end(); ++it
) {
3980 scop_j
= pet_scop_from_pet_stmt(ctx
, *it
);
3981 scop_j
= pet_scop_prefix(scop_j
, j
);
3982 scop
= pet_scop_add_seq(ctx
, scop
, scop_j
);
3985 pet_context_free(pc
);
3987 if (scop
&& partial_range
) {
3988 if (scop
->n_stmt
== 0 || kills
.size() != 0) {
3989 pet_scop_free(scop
);
3998 /* Check if the scop marked by the user is exactly this Stmt
3999 * or part of this Stmt.
4000 * If so, return a pet_scop corresponding to the marked region.
4001 * The pet_scop is created within the context "pc".
4002 * Otherwise, return NULL.
4004 struct pet_scop
*PetScan::scan(Stmt
*stmt
, __isl_keep pet_context
*pc
)
4006 SourceManager
&SM
= PP
.getSourceManager();
4007 unsigned start_off
, end_off
;
4009 start_off
= getExpansionOffset(SM
, stmt
->getLocStart());
4010 end_off
= getExpansionOffset(SM
, stmt
->getLocEnd());
4012 if (start_off
> loc
.end
)
4014 if (end_off
< loc
.start
)
4016 if (start_off
>= loc
.start
&& end_off
<= loc
.end
) {
4017 return extract(stmt
, pc
);
4021 for (start
= stmt
->child_begin(); start
!= stmt
->child_end(); ++start
) {
4022 Stmt
*child
= *start
;
4025 start_off
= getExpansionOffset(SM
, child
->getLocStart());
4026 end_off
= getExpansionOffset(SM
, child
->getLocEnd());
4027 if (start_off
< loc
.start
&& end_off
>= loc
.end
)
4028 return scan(child
, pc
);
4029 if (start_off
>= loc
.start
)
4034 for (end
= start
; end
!= stmt
->child_end(); ++end
) {
4036 start_off
= SM
.getFileOffset(child
->getLocStart());
4037 if (start_off
>= loc
.end
)
4041 return extract(StmtRange(start
, end
), false, false, pc
);
4044 /* Set the size of index "pos" of "array" to "size".
4045 * In particular, add a constraint of the form
4049 * to array->extent and a constraint of the form
4053 * to array->context.
4055 static struct pet_array
*update_size(struct pet_array
*array
, int pos
,
4056 __isl_take isl_pw_aff
*size
)
4069 valid
= isl_set_params(isl_pw_aff_nonneg_set(isl_pw_aff_copy(size
)));
4070 array
->context
= isl_set_intersect(array
->context
, valid
);
4072 dim
= isl_set_get_space(array
->extent
);
4073 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
4074 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, pos
, 1);
4075 univ
= isl_set_universe(isl_aff_get_domain_space(aff
));
4076 index
= isl_pw_aff_alloc(univ
, aff
);
4078 size
= isl_pw_aff_add_dims(size
, isl_dim_in
,
4079 isl_set_dim(array
->extent
, isl_dim_set
));
4080 id
= isl_set_get_tuple_id(array
->extent
);
4081 size
= isl_pw_aff_set_tuple_id(size
, isl_dim_in
, id
);
4082 bound
= isl_pw_aff_lt_set(index
, size
);
4084 array
->extent
= isl_set_intersect(array
->extent
, bound
);
4086 if (!array
->context
|| !array
->extent
)
4087 return pet_array_free(array
);
4091 isl_pw_aff_free(size
);
4095 /* Figure out the size of the array at position "pos" and all
4096 * subsequent positions from "type" and update the corresponding
4097 * argument of "expr" accordingly.
4099 __isl_give pet_expr
*PetScan::set_upper_bounds(__isl_take pet_expr
*expr
,
4100 const Type
*type
, int pos
)
4102 const ArrayType
*atype
;
4108 if (type
->isPointerType()) {
4109 type
= type
->getPointeeType().getTypePtr();
4110 return set_upper_bounds(expr
, type
, pos
+ 1);
4112 if (!type
->isArrayType())
4115 type
= type
->getCanonicalTypeInternal().getTypePtr();
4116 atype
= cast
<ArrayType
>(type
);
4118 if (type
->isConstantArrayType()) {
4119 const ConstantArrayType
*ca
= cast
<ConstantArrayType
>(atype
);
4120 size
= extract_expr(ca
->getSize());
4121 expr
= pet_expr_set_arg(expr
, pos
, size
);
4122 } else if (type
->isVariableArrayType()) {
4123 const VariableArrayType
*vla
= cast
<VariableArrayType
>(atype
);
4124 size
= extract_expr(vla
->getSizeExpr());
4125 expr
= pet_expr_set_arg(expr
, pos
, size
);
4128 type
= atype
->getElementType().getTypePtr();
4130 return set_upper_bounds(expr
, type
, pos
+ 1);
4133 /* Does "expr" represent the "integer" infinity?
4135 static int is_infty(__isl_keep pet_expr
*expr
)
4140 if (pet_expr_get_type(expr
) != pet_expr_int
)
4142 v
= pet_expr_int_get_val(expr
);
4143 res
= isl_val_is_infty(v
);
4149 /* Figure out the dimensions of an array "array" based on its type
4150 * "type" and update "array" accordingly.
4152 * We first construct a pet_expr that holds the sizes of the array
4153 * in each dimension. The expression is initialized to infinity
4154 * and updated from the type.
4156 * The arguments of the size expression that have been updated
4157 * are then converted to an affine expression within the context "pc" and
4158 * incorporated into the size of "array". If we are unable to convert
4159 * a size expression to an affine expression, then we leave
4160 * the corresponding size of "array" untouched.
4162 struct pet_array
*PetScan::set_upper_bounds(struct pet_array
*array
,
4163 const Type
*type
, __isl_keep pet_context
*pc
)
4165 int depth
= array_depth(type
);
4166 pet_expr
*expr
, *inf
;
4171 inf
= pet_expr_new_int(isl_val_infty(ctx
));
4172 expr
= pet_expr_new_call(ctx
, "bounds", depth
);
4173 for (int i
= 0; i
< depth
; ++i
)
4174 expr
= pet_expr_set_arg(expr
, i
, pet_expr_copy(inf
));
4177 expr
= set_upper_bounds(expr
, type
, 0);
4179 for (int i
= 0; i
< depth
; ++i
) {
4183 arg
= pet_expr_get_arg(expr
, i
);
4184 if (!is_infty(arg
)) {
4185 size
= pet_expr_extract_affine(arg
, pc
);
4187 array
= pet_array_free(array
);
4188 else if (isl_pw_aff_involves_nan(size
))
4189 isl_pw_aff_free(size
);
4191 array
= update_size(array
, i
, size
);
4195 pet_expr_free(expr
);
4200 /* Is "T" the type of a variable length array with static size?
4202 static bool is_vla_with_static_size(QualType T
)
4204 const VariableArrayType
*vlatype
;
4206 if (!T
->isVariableArrayType())
4208 vlatype
= cast
<VariableArrayType
>(T
);
4209 return vlatype
->getSizeModifier() == VariableArrayType::Static
;
4212 /* Return the type of "decl" as an array.
4214 * In particular, if "decl" is a parameter declaration that
4215 * is a variable length array with a static size, then
4216 * return the original type (i.e., the variable length array).
4217 * Otherwise, return the type of decl.
4219 static QualType
get_array_type(ValueDecl
*decl
)
4224 parm
= dyn_cast
<ParmVarDecl
>(decl
);
4226 return decl
->getType();
4228 T
= parm
->getOriginalType();
4229 if (!is_vla_with_static_size(T
))
4230 return decl
->getType();
4234 /* Does "decl" have definition that we can keep track of in a pet_type?
4236 static bool has_printable_definition(RecordDecl
*decl
)
4238 if (!decl
->getDeclName())
4240 return decl
->getLexicalDeclContext() == decl
->getDeclContext();
4243 /* Construct and return a pet_array corresponding to the variable "decl".
4244 * In particular, initialize array->extent to
4246 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
4248 * and then call set_upper_bounds to set the upper bounds on the indices
4249 * based on the type of the variable. The upper bounds are converted
4250 * to affine expressions within the context "pc".
4252 * If the base type is that of a record with a top-level definition and
4253 * if "types" is not null, then the RecordDecl corresponding to the type
4254 * is added to "types".
4256 * If the base type is that of a record with no top-level definition,
4257 * then we replace it by "<subfield>".
4259 struct pet_array
*PetScan::extract_array(isl_ctx
*ctx
, ValueDecl
*decl
,
4260 lex_recorddecl_set
*types
, __isl_keep pet_context
*pc
)
4262 struct pet_array
*array
;
4263 QualType qt
= get_array_type(decl
);
4264 const Type
*type
= qt
.getTypePtr();
4265 int depth
= array_depth(type
);
4266 QualType base
= pet_clang_base_type(qt
);
4271 array
= isl_calloc_type(ctx
, struct pet_array
);
4275 id
= create_decl_id(ctx
, decl
);
4276 dim
= isl_space_set_alloc(ctx
, 0, depth
);
4277 dim
= isl_space_set_tuple_id(dim
, isl_dim_set
, id
);
4279 array
->extent
= isl_set_nat_universe(dim
);
4281 dim
= isl_space_params_alloc(ctx
, 0);
4282 array
->context
= isl_set_universe(dim
);
4284 array
= set_upper_bounds(array
, type
, pc
);
4288 name
= base
.getAsString();
4290 if (types
&& base
->isRecordType()) {
4291 RecordDecl
*decl
= pet_clang_record_decl(base
);
4292 if (has_printable_definition(decl
))
4293 types
->insert(decl
);
4295 name
= "<subfield>";
4298 array
->element_type
= strdup(name
.c_str());
4299 array
->element_is_record
= base
->isRecordType();
4300 array
->element_size
= decl
->getASTContext().getTypeInfo(base
).first
/ 8;
4305 /* Construct and return a pet_array corresponding to the sequence
4306 * of declarations "decls".
4307 * The upper bounds of the array are converted to affine expressions
4308 * within the context "pc".
4309 * If the sequence contains a single declaration, then it corresponds
4310 * to a simple array access. Otherwise, it corresponds to a member access,
4311 * with the declaration for the substructure following that of the containing
4312 * structure in the sequence of declarations.
4313 * We start with the outermost substructure and then combine it with
4314 * information from the inner structures.
4316 * Additionally, keep track of all required types in "types".
4318 struct pet_array
*PetScan::extract_array(isl_ctx
*ctx
,
4319 vector
<ValueDecl
*> decls
, lex_recorddecl_set
*types
,
4320 __isl_keep pet_context
*pc
)
4322 struct pet_array
*array
;
4323 vector
<ValueDecl
*>::iterator it
;
4327 array
= extract_array(ctx
, *it
, types
, pc
);
4329 for (++it
; it
!= decls
.end(); ++it
) {
4330 struct pet_array
*parent
;
4331 const char *base_name
, *field_name
;
4335 array
= extract_array(ctx
, *it
, types
, pc
);
4337 return pet_array_free(parent
);
4339 base_name
= isl_set_get_tuple_name(parent
->extent
);
4340 field_name
= isl_set_get_tuple_name(array
->extent
);
4341 product_name
= pet_array_member_access_name(ctx
,
4342 base_name
, field_name
);
4344 array
->extent
= isl_set_product(isl_set_copy(parent
->extent
),
4347 array
->extent
= isl_set_set_tuple_name(array
->extent
,
4349 array
->context
= isl_set_intersect(array
->context
,
4350 isl_set_copy(parent
->context
));
4352 pet_array_free(parent
);
4355 if (!array
->extent
|| !array
->context
|| !product_name
)
4356 return pet_array_free(array
);
4362 /* Add a pet_type corresponding to "decl" to "scop, provided
4363 * it is a member of "types" and it has not been added before
4364 * (i.e., it is not a member of "types_done".
4366 * Since we want the user to be able to print the types
4367 * in the order in which they appear in the scop, we need to
4368 * make sure that types of fields in a structure appear before
4369 * that structure. We therefore call ourselves recursively
4370 * on the types of all record subfields.
4372 static struct pet_scop
*add_type(isl_ctx
*ctx
, struct pet_scop
*scop
,
4373 RecordDecl
*decl
, Preprocessor
&PP
, lex_recorddecl_set
&types
,
4374 lex_recorddecl_set
&types_done
)
4377 llvm::raw_string_ostream
S(s
);
4378 RecordDecl::field_iterator it
;
4380 if (types
.find(decl
) == types
.end())
4382 if (types_done
.find(decl
) != types_done
.end())
4385 for (it
= decl
->field_begin(); it
!= decl
->field_end(); ++it
) {
4387 QualType type
= it
->getType();
4389 if (!type
->isRecordType())
4391 record
= pet_clang_record_decl(type
);
4392 scop
= add_type(ctx
, scop
, record
, PP
, types
, types_done
);
4395 if (strlen(decl
->getName().str().c_str()) == 0)
4398 decl
->print(S
, PrintingPolicy(PP
.getLangOpts()));
4401 scop
->types
[scop
->n_type
] = pet_type_alloc(ctx
,
4402 decl
->getName().str().c_str(), s
.c_str());
4403 if (!scop
->types
[scop
->n_type
])
4404 return pet_scop_free(scop
);
4406 types_done
.insert(decl
);
4413 /* Construct a list of pet_arrays, one for each array (or scalar)
4414 * accessed inside "scop", add this list to "scop" and return the result.
4415 * The upper bounds of the arrays are converted to affine expressions
4416 * within the context "pc".
4418 * The context of "scop" is updated with the intersection of
4419 * the contexts of all arrays, i.e., constraints on the parameters
4420 * that ensure that the arrays have a valid (non-negative) size.
4422 * If the any of the extracted arrays refers to a member access,
4423 * then also add the required types to "scop".
4425 struct pet_scop
*PetScan::scan_arrays(struct pet_scop
*scop
,
4426 __isl_keep pet_context
*pc
)
4429 array_desc_set arrays
;
4430 array_desc_set::iterator it
;
4431 lex_recorddecl_set types
;
4432 lex_recorddecl_set types_done
;
4433 lex_recorddecl_set::iterator types_it
;
4435 struct pet_array
**scop_arrays
;
4440 pet_scop_collect_arrays(scop
, arrays
);
4441 if (arrays
.size() == 0)
4444 n_array
= scop
->n_array
;
4446 scop_arrays
= isl_realloc_array(ctx
, scop
->arrays
, struct pet_array
*,
4447 n_array
+ arrays
.size());
4450 scop
->arrays
= scop_arrays
;
4452 for (it
= arrays
.begin(), i
= 0; it
!= arrays
.end(); ++it
, ++i
) {
4453 struct pet_array
*array
;
4454 array
= extract_array(ctx
, *it
, &types
, pc
);
4455 scop
->arrays
[n_array
+ i
] = array
;
4456 if (!scop
->arrays
[n_array
+ i
])
4459 scop
->context
= isl_set_intersect(scop
->context
,
4460 isl_set_copy(array
->context
));
4465 if (types
.size() == 0)
4468 scop
->types
= isl_alloc_array(ctx
, struct pet_type
*, types
.size());
4472 for (types_it
= types
.begin(); types_it
!= types
.end(); ++types_it
)
4473 scop
= add_type(ctx
, scop
, *types_it
, PP
, types
, types_done
);
4477 pet_scop_free(scop
);
4481 /* Bound all parameters in scop->context to the possible values
4482 * of the corresponding C variable.
4484 static struct pet_scop
*add_parameter_bounds(struct pet_scop
*scop
)
4491 n
= isl_set_dim(scop
->context
, isl_dim_param
);
4492 for (int i
= 0; i
< n
; ++i
) {
4496 id
= isl_set_get_dim_id(scop
->context
, isl_dim_param
, i
);
4497 if (pet_nested_in_id(id
)) {
4499 isl_die(isl_set_get_ctx(scop
->context
),
4501 "unresolved nested parameter", goto error
);
4503 decl
= (ValueDecl
*) isl_id_get_user(id
);
4506 scop
->context
= set_parameter_bounds(scop
->context
, i
, decl
);
4514 pet_scop_free(scop
);
4518 /* Construct a pet_scop from the given function.
4520 * If the scop was delimited by scop and endscop pragmas, then we override
4521 * the file offsets by those derived from the pragmas.
4523 struct pet_scop
*PetScan::scan(FunctionDecl
*fd
)
4529 stmt
= fd
->getBody();
4531 pc
= pet_context_alloc(isl_space_set_alloc(ctx
, 0, 0));
4532 if (options
->autodetect
) {
4533 scop
= extract(stmt
, pc
, true);
4535 scop
= scan(stmt
, pc
);
4536 scop
= pet_scop_update_start_end(scop
, loc
.start
, loc
.end
);
4538 scop
= pet_scop_detect_parameter_accesses(scop
);
4539 scop
= scan_arrays(scop
, pc
);
4540 pet_context_free(pc
);
4541 scop
= add_parameter_bounds(scop
);
4542 scop
= pet_scop_gist(scop
, value_bounds
);