2 * Copyright 2011 Leiden University. All rights reserved.
3 * Copyright 2012-2013 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>
53 #include "scop_plus.h"
58 using namespace clang
;
60 #if defined(DECLREFEXPR_CREATE_REQUIRES_BOOL)
61 static DeclRefExpr
*create_DeclRefExpr(VarDecl
*var
)
63 return DeclRefExpr::Create(var
->getASTContext(), var
->getQualifierLoc(),
64 SourceLocation(), var
, false, var
->getInnerLocStart(),
65 var
->getType(), VK_LValue
);
67 #elif defined(DECLREFEXPR_CREATE_REQUIRES_SOURCELOCATION)
68 static DeclRefExpr
*create_DeclRefExpr(VarDecl
*var
)
70 return DeclRefExpr::Create(var
->getASTContext(), var
->getQualifierLoc(),
71 SourceLocation(), var
, var
->getInnerLocStart(), var
->getType(),
75 static DeclRefExpr
*create_DeclRefExpr(VarDecl
*var
)
77 return DeclRefExpr::Create(var
->getASTContext(), var
->getQualifierLoc(),
78 var
, var
->getInnerLocStart(), var
->getType(), VK_LValue
);
82 /* Check if the element type corresponding to the given array type
83 * has a const qualifier.
85 static bool const_base(QualType qt
)
87 const Type
*type
= qt
.getTypePtr();
89 if (type
->isPointerType())
90 return const_base(type
->getPointeeType());
91 if (type
->isArrayType()) {
92 const ArrayType
*atype
;
93 type
= type
->getCanonicalTypeInternal().getTypePtr();
94 atype
= cast
<ArrayType
>(type
);
95 return const_base(atype
->getElementType());
98 return qt
.isConstQualified();
101 /* Mark "decl" as having an unknown value in "assigned_value".
103 * If no (known or unknown) value was assigned to "decl" before,
104 * then it may have been treated as a parameter before and may
105 * therefore appear in a value assigned to another variable.
106 * If so, this assignment needs to be turned into an unknown value too.
108 static void clear_assignment(map
<ValueDecl
*, isl_pw_aff
*> &assigned_value
,
111 map
<ValueDecl
*, isl_pw_aff
*>::iterator it
;
113 it
= assigned_value
.find(decl
);
115 assigned_value
[decl
] = NULL
;
117 if (it
== assigned_value
.end())
120 for (it
= assigned_value
.begin(); it
!= assigned_value
.end(); ++it
) {
121 isl_pw_aff
*pa
= it
->second
;
122 int nparam
= isl_pw_aff_dim(pa
, isl_dim_param
);
124 for (int i
= 0; i
< nparam
; ++i
) {
127 if (!isl_pw_aff_has_dim_id(pa
, isl_dim_param
, i
))
129 id
= isl_pw_aff_get_dim_id(pa
, isl_dim_param
, i
);
130 if (isl_id_get_user(id
) == decl
)
137 /* Look for any assignments to scalar variables in part of the parse
138 * tree and set assigned_value to NULL for each of them.
139 * Also reset assigned_value if the address of a scalar variable
140 * is being taken. As an exception, if the address is passed to a function
141 * that is declared to receive a const pointer, then assigned_value is
144 * This ensures that we won't use any previously stored value
145 * in the current subtree and its parents.
147 struct clear_assignments
: RecursiveASTVisitor
<clear_assignments
> {
148 map
<ValueDecl
*, isl_pw_aff
*> &assigned_value
;
149 set
<UnaryOperator
*> skip
;
151 clear_assignments(map
<ValueDecl
*, isl_pw_aff
*> &assigned_value
) :
152 assigned_value(assigned_value
) {}
154 /* Check for "address of" operators whose value is passed
155 * to a const pointer argument and add them to "skip", so that
156 * we can skip them in VisitUnaryOperator.
158 bool VisitCallExpr(CallExpr
*expr
) {
160 fd
= expr
->getDirectCallee();
163 for (int i
= 0; i
< expr
->getNumArgs(); ++i
) {
164 Expr
*arg
= expr
->getArg(i
);
166 if (arg
->getStmtClass() == Stmt::ImplicitCastExprClass
) {
167 ImplicitCastExpr
*ice
;
168 ice
= cast
<ImplicitCastExpr
>(arg
);
169 arg
= ice
->getSubExpr();
171 if (arg
->getStmtClass() != Stmt::UnaryOperatorClass
)
173 op
= cast
<UnaryOperator
>(arg
);
174 if (op
->getOpcode() != UO_AddrOf
)
176 if (const_base(fd
->getParamDecl(i
)->getType()))
182 bool VisitUnaryOperator(UnaryOperator
*expr
) {
187 switch (expr
->getOpcode()) {
197 if (skip
.find(expr
) != skip
.end())
200 arg
= expr
->getSubExpr();
201 if (arg
->getStmtClass() != Stmt::DeclRefExprClass
)
203 ref
= cast
<DeclRefExpr
>(arg
);
204 decl
= ref
->getDecl();
205 clear_assignment(assigned_value
, decl
);
209 bool VisitBinaryOperator(BinaryOperator
*expr
) {
214 if (!expr
->isAssignmentOp())
216 lhs
= expr
->getLHS();
217 if (lhs
->getStmtClass() != Stmt::DeclRefExprClass
)
219 ref
= cast
<DeclRefExpr
>(lhs
);
220 decl
= ref
->getDecl();
221 clear_assignment(assigned_value
, decl
);
226 /* Keep a copy of the currently assigned values.
228 * Any variable that is assigned a value inside the current scope
229 * is removed again when we leave the scope (either because it wasn't
230 * stored in the cache or because it has a different value in the cache).
232 struct assigned_value_cache
{
233 map
<ValueDecl
*, isl_pw_aff
*> &assigned_value
;
234 map
<ValueDecl
*, isl_pw_aff
*> cache
;
236 assigned_value_cache(map
<ValueDecl
*, isl_pw_aff
*> &assigned_value
) :
237 assigned_value(assigned_value
), cache(assigned_value
) {}
238 ~assigned_value_cache() {
239 map
<ValueDecl
*, isl_pw_aff
*>::iterator it
= cache
.begin();
240 for (it
= assigned_value
.begin(); it
!= assigned_value
.end();
243 (cache
.find(it
->first
) != cache
.end() &&
244 cache
[it
->first
] != it
->second
))
245 cache
[it
->first
] = NULL
;
247 assigned_value
= cache
;
251 /* Insert an expression into the collection of expressions,
252 * provided it is not already in there.
253 * The isl_pw_affs are freed in the destructor.
255 void PetScan::insert_expression(__isl_take isl_pw_aff
*expr
)
257 std::set
<isl_pw_aff
*>::iterator it
;
259 if (expressions
.find(expr
) == expressions
.end())
260 expressions
.insert(expr
);
262 isl_pw_aff_free(expr
);
267 std::set
<isl_pw_aff
*>::iterator it
;
269 for (it
= expressions
.begin(); it
!= expressions
.end(); ++it
)
270 isl_pw_aff_free(*it
);
272 isl_union_map_free(value_bounds
);
275 /* Called if we found something we (currently) cannot handle.
276 * We'll provide more informative warnings later.
278 * We only actually complain if autodetect is false.
280 void PetScan::unsupported(Stmt
*stmt
, const char *msg
)
282 if (options
->autodetect
)
285 SourceLocation loc
= stmt
->getLocStart();
286 DiagnosticsEngine
&diag
= PP
.getDiagnostics();
287 unsigned id
= diag
.getCustomDiagID(DiagnosticsEngine::Warning
,
288 msg
? msg
: "unsupported");
289 DiagnosticBuilder B
= diag
.Report(loc
, id
) << stmt
->getSourceRange();
292 /* Extract an integer from "expr".
294 __isl_give isl_val
*PetScan::extract_int(isl_ctx
*ctx
, IntegerLiteral
*expr
)
296 const Type
*type
= expr
->getType().getTypePtr();
297 int is_signed
= type
->hasSignedIntegerRepresentation();
298 llvm::APInt val
= expr
->getValue();
299 int is_negative
= is_signed
&& val
.isNegative();
305 v
= extract_unsigned(ctx
, val
);
312 /* Extract an integer from "val", which assumed to be non-negative.
314 __isl_give isl_val
*PetScan::extract_unsigned(isl_ctx
*ctx
,
315 const llvm::APInt
&val
)
318 const uint64_t *data
;
320 data
= val
.getRawData();
321 n
= val
.getNumWords();
322 return isl_val_int_from_chunks(ctx
, n
, sizeof(uint64_t), data
);
325 /* Extract an integer from "expr".
326 * Return NULL if "expr" does not (obviously) represent an integer.
328 __isl_give isl_val
*PetScan::extract_int(clang::ParenExpr
*expr
)
330 return extract_int(expr
->getSubExpr());
333 /* Extract an integer from "expr".
334 * Return NULL if "expr" does not (obviously) represent an integer.
336 __isl_give isl_val
*PetScan::extract_int(clang::Expr
*expr
)
338 if (expr
->getStmtClass() == Stmt::IntegerLiteralClass
)
339 return extract_int(ctx
, cast
<IntegerLiteral
>(expr
));
340 if (expr
->getStmtClass() == Stmt::ParenExprClass
)
341 return extract_int(cast
<ParenExpr
>(expr
));
347 /* Extract an affine expression from the IntegerLiteral "expr".
349 __isl_give isl_pw_aff
*PetScan::extract_affine(IntegerLiteral
*expr
)
351 isl_space
*dim
= isl_space_params_alloc(ctx
, 0);
352 isl_local_space
*ls
= isl_local_space_from_space(isl_space_copy(dim
));
353 isl_aff
*aff
= isl_aff_zero_on_domain(ls
);
354 isl_set
*dom
= isl_set_universe(dim
);
357 v
= extract_int(expr
);
358 aff
= isl_aff_add_constant_val(aff
, v
);
360 return isl_pw_aff_alloc(dom
, aff
);
363 /* Extract an affine expression from the APInt "val", which is assumed
364 * to be non-negative.
366 __isl_give isl_pw_aff
*PetScan::extract_affine(const llvm::APInt
&val
)
368 isl_space
*dim
= isl_space_params_alloc(ctx
, 0);
369 isl_local_space
*ls
= isl_local_space_from_space(isl_space_copy(dim
));
370 isl_aff
*aff
= isl_aff_zero_on_domain(ls
);
371 isl_set
*dom
= isl_set_universe(dim
);
374 v
= extract_unsigned(ctx
, val
);
375 aff
= isl_aff_add_constant_val(aff
, v
);
377 return isl_pw_aff_alloc(dom
, aff
);
380 __isl_give isl_pw_aff
*PetScan::extract_affine(ImplicitCastExpr
*expr
)
382 return extract_affine(expr
->getSubExpr());
385 static unsigned get_type_size(ValueDecl
*decl
)
387 return decl
->getASTContext().getIntWidth(decl
->getType());
390 /* Bound parameter "pos" of "set" to the possible values of "decl".
392 static __isl_give isl_set
*set_parameter_bounds(__isl_take isl_set
*set
,
393 unsigned pos
, ValueDecl
*decl
)
399 ctx
= isl_set_get_ctx(set
);
400 width
= get_type_size(decl
);
401 if (decl
->getType()->isUnsignedIntegerType()) {
402 set
= isl_set_lower_bound_si(set
, isl_dim_param
, pos
, 0);
403 bound
= isl_val_int_from_ui(ctx
, width
);
404 bound
= isl_val_2exp(bound
);
405 bound
= isl_val_sub_ui(bound
, 1);
406 set
= isl_set_upper_bound_val(set
, isl_dim_param
, pos
, bound
);
408 bound
= isl_val_int_from_ui(ctx
, width
- 1);
409 bound
= isl_val_2exp(bound
);
410 bound
= isl_val_sub_ui(bound
, 1);
411 set
= isl_set_upper_bound_val(set
, isl_dim_param
, pos
,
412 isl_val_copy(bound
));
413 bound
= isl_val_neg(bound
);
414 bound
= isl_val_sub_ui(bound
, 1);
415 set
= isl_set_lower_bound_val(set
, isl_dim_param
, pos
, bound
);
421 /* Extract an affine expression from the DeclRefExpr "expr".
423 * If the variable has been assigned a value, then we check whether
424 * we know what (affine) value was assigned.
425 * If so, we return this value. Otherwise we convert "expr"
426 * to an extra parameter (provided nesting_enabled is set).
428 * Otherwise, we simply return an expression that is equal
429 * to a parameter corresponding to the referenced variable.
431 __isl_give isl_pw_aff
*PetScan::extract_affine(DeclRefExpr
*expr
)
433 ValueDecl
*decl
= expr
->getDecl();
434 const Type
*type
= decl
->getType().getTypePtr();
440 if (!type
->isIntegerType()) {
445 if (assigned_value
.find(decl
) != assigned_value
.end()) {
446 if (assigned_value
[decl
])
447 return isl_pw_aff_copy(assigned_value
[decl
]);
449 return nested_access(expr
);
452 id
= isl_id_alloc(ctx
, decl
->getName().str().c_str(), decl
);
453 dim
= isl_space_params_alloc(ctx
, 1);
455 dim
= isl_space_set_dim_id(dim
, isl_dim_param
, 0, id
);
457 dom
= isl_set_universe(isl_space_copy(dim
));
458 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
459 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_param
, 0, 1);
461 return isl_pw_aff_alloc(dom
, aff
);
464 /* Extract an affine expression from an integer division operation.
465 * In particular, if "expr" is lhs/rhs, then return
467 * lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs)
469 * The second argument (rhs) is required to be a (positive) integer constant.
471 __isl_give isl_pw_aff
*PetScan::extract_affine_div(BinaryOperator
*expr
)
474 isl_pw_aff
*rhs
, *lhs
;
476 rhs
= extract_affine(expr
->getRHS());
477 is_cst
= isl_pw_aff_is_cst(rhs
);
478 if (is_cst
< 0 || !is_cst
) {
479 isl_pw_aff_free(rhs
);
485 lhs
= extract_affine(expr
->getLHS());
487 return isl_pw_aff_tdiv_q(lhs
, rhs
);
490 /* Extract an affine expression from a modulo operation.
491 * In particular, if "expr" is lhs/rhs, then return
493 * lhs - rhs * (lhs >= 0 ? floor(lhs/rhs) : ceil(lhs/rhs))
495 * The second argument (rhs) is required to be a (positive) integer constant.
497 __isl_give isl_pw_aff
*PetScan::extract_affine_mod(BinaryOperator
*expr
)
500 isl_pw_aff
*rhs
, *lhs
;
502 rhs
= extract_affine(expr
->getRHS());
503 is_cst
= isl_pw_aff_is_cst(rhs
);
504 if (is_cst
< 0 || !is_cst
) {
505 isl_pw_aff_free(rhs
);
511 lhs
= extract_affine(expr
->getLHS());
513 return isl_pw_aff_tdiv_r(lhs
, rhs
);
516 /* Extract an affine expression from a multiplication operation.
517 * This is only allowed if at least one of the two arguments
518 * is a (piecewise) constant.
520 __isl_give isl_pw_aff
*PetScan::extract_affine_mul(BinaryOperator
*expr
)
525 lhs
= extract_affine(expr
->getLHS());
526 rhs
= extract_affine(expr
->getRHS());
528 if (!isl_pw_aff_is_cst(lhs
) && !isl_pw_aff_is_cst(rhs
)) {
529 isl_pw_aff_free(lhs
);
530 isl_pw_aff_free(rhs
);
535 return isl_pw_aff_mul(lhs
, rhs
);
538 /* Extract an affine expression from an addition or subtraction operation.
540 __isl_give isl_pw_aff
*PetScan::extract_affine_add(BinaryOperator
*expr
)
545 lhs
= extract_affine(expr
->getLHS());
546 rhs
= extract_affine(expr
->getRHS());
548 switch (expr
->getOpcode()) {
550 return isl_pw_aff_add(lhs
, rhs
);
552 return isl_pw_aff_sub(lhs
, rhs
);
554 isl_pw_aff_free(lhs
);
555 isl_pw_aff_free(rhs
);
565 static __isl_give isl_pw_aff
*wrap(__isl_take isl_pw_aff
*pwaff
,
571 ctx
= isl_pw_aff_get_ctx(pwaff
);
572 mod
= isl_val_int_from_ui(ctx
, width
);
573 mod
= isl_val_2exp(mod
);
575 pwaff
= isl_pw_aff_mod_val(pwaff
, mod
);
580 /* Limit the domain of "pwaff" to those elements where the function
583 * 2^{width-1} <= pwaff < 2^{width-1}
585 static __isl_give isl_pw_aff
*avoid_overflow(__isl_take isl_pw_aff
*pwaff
,
590 isl_space
*space
= isl_pw_aff_get_domain_space(pwaff
);
591 isl_local_space
*ls
= isl_local_space_from_space(space
);
596 ctx
= isl_pw_aff_get_ctx(pwaff
);
597 v
= isl_val_int_from_ui(ctx
, width
- 1);
600 bound
= isl_aff_zero_on_domain(ls
);
601 bound
= isl_aff_add_constant_val(bound
, v
);
602 b
= isl_pw_aff_from_aff(bound
);
604 dom
= isl_pw_aff_lt_set(isl_pw_aff_copy(pwaff
), isl_pw_aff_copy(b
));
605 pwaff
= isl_pw_aff_intersect_domain(pwaff
, dom
);
607 b
= isl_pw_aff_neg(b
);
608 dom
= isl_pw_aff_ge_set(isl_pw_aff_copy(pwaff
), b
);
609 pwaff
= isl_pw_aff_intersect_domain(pwaff
, dom
);
614 /* Handle potential overflows on signed computations.
616 * If options->signed_overflow is set to PET_OVERFLOW_AVOID,
617 * the we adjust the domain of "pa" to avoid overflows.
619 __isl_give isl_pw_aff
*PetScan::signed_overflow(__isl_take isl_pw_aff
*pa
,
622 if (options
->signed_overflow
== PET_OVERFLOW_AVOID
)
623 pa
= avoid_overflow(pa
, width
);
628 /* Return the piecewise affine expression "set ? 1 : 0" defined on "dom".
630 static __isl_give isl_pw_aff
*indicator_function(__isl_take isl_set
*set
,
631 __isl_take isl_set
*dom
)
634 pa
= isl_set_indicator_function(set
);
635 pa
= isl_pw_aff_intersect_domain(pa
, dom
);
639 /* Extract an affine expression from some binary operations.
640 * If the result of the expression is unsigned, then we wrap it
641 * based on the size of the type. Otherwise, we ensure that
642 * no overflow occurs.
644 __isl_give isl_pw_aff
*PetScan::extract_affine(BinaryOperator
*expr
)
649 switch (expr
->getOpcode()) {
652 res
= extract_affine_add(expr
);
655 res
= extract_affine_div(expr
);
658 res
= extract_affine_mod(expr
);
661 res
= extract_affine_mul(expr
);
671 return extract_condition(expr
);
677 width
= ast_context
.getIntWidth(expr
->getType());
678 if (expr
->getType()->isUnsignedIntegerType())
679 res
= wrap(res
, width
);
681 res
= signed_overflow(res
, width
);
686 /* Extract an affine expression from a negation operation.
688 __isl_give isl_pw_aff
*PetScan::extract_affine(UnaryOperator
*expr
)
690 if (expr
->getOpcode() == UO_Minus
)
691 return isl_pw_aff_neg(extract_affine(expr
->getSubExpr()));
692 if (expr
->getOpcode() == UO_LNot
)
693 return extract_condition(expr
);
699 __isl_give isl_pw_aff
*PetScan::extract_affine(ParenExpr
*expr
)
701 return extract_affine(expr
->getSubExpr());
704 /* Extract an affine expression from some special function calls.
705 * In particular, we handle "min", "max", "ceild" and "floord".
706 * In case of the latter two, the second argument needs to be
707 * a (positive) integer constant.
709 __isl_give isl_pw_aff
*PetScan::extract_affine(CallExpr
*expr
)
713 isl_pw_aff
*aff1
, *aff2
;
715 fd
= expr
->getDirectCallee();
721 name
= fd
->getDeclName().getAsString();
722 if (!(expr
->getNumArgs() == 2 && name
== "min") &&
723 !(expr
->getNumArgs() == 2 && name
== "max") &&
724 !(expr
->getNumArgs() == 2 && name
== "floord") &&
725 !(expr
->getNumArgs() == 2 && name
== "ceild")) {
730 if (name
== "min" || name
== "max") {
731 aff1
= extract_affine(expr
->getArg(0));
732 aff2
= extract_affine(expr
->getArg(1));
735 aff1
= isl_pw_aff_min(aff1
, aff2
);
737 aff1
= isl_pw_aff_max(aff1
, aff2
);
738 } else if (name
== "floord" || name
== "ceild") {
740 Expr
*arg2
= expr
->getArg(1);
742 if (arg2
->getStmtClass() != Stmt::IntegerLiteralClass
) {
746 aff1
= extract_affine(expr
->getArg(0));
747 v
= extract_int(cast
<IntegerLiteral
>(arg2
));
748 aff1
= isl_pw_aff_scale_down_val(aff1
, v
);
749 if (name
== "floord")
750 aff1
= isl_pw_aff_floor(aff1
);
752 aff1
= isl_pw_aff_ceil(aff1
);
761 /* This method is called when we come across an access that is
762 * nested in what is supposed to be an affine expression.
763 * If nesting is allowed, we return a new parameter that corresponds
764 * to this nested access. Otherwise, we simply complain.
766 * Note that we currently don't allow nested accesses themselves
767 * to contain any nested accesses, so we check if we can extract
768 * the access without any nesting and complain if we can't.
770 * The new parameter is resolved in resolve_nested.
772 isl_pw_aff
*PetScan::nested_access(Expr
*expr
)
778 isl_multi_pw_aff
*index
;
780 if (!nesting_enabled
) {
785 allow_nested
= false;
786 index
= extract_index(expr
);
792 isl_multi_pw_aff_free(index
);
794 id
= isl_id_alloc(ctx
, NULL
, expr
);
795 dim
= isl_space_params_alloc(ctx
, 1);
797 dim
= isl_space_set_dim_id(dim
, isl_dim_param
, 0, id
);
799 dom
= isl_set_universe(isl_space_copy(dim
));
800 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
801 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_param
, 0, 1);
803 return isl_pw_aff_alloc(dom
, aff
);
806 /* Affine expressions are not supposed to contain array accesses,
807 * but if nesting is allowed, we return a parameter corresponding
808 * to the array access.
810 __isl_give isl_pw_aff
*PetScan::extract_affine(ArraySubscriptExpr
*expr
)
812 return nested_access(expr
);
815 /* Extract an affine expression from a conditional operation.
817 __isl_give isl_pw_aff
*PetScan::extract_affine(ConditionalOperator
*expr
)
819 isl_pw_aff
*cond
, *lhs
, *rhs
;
821 cond
= extract_condition(expr
->getCond());
822 lhs
= extract_affine(expr
->getTrueExpr());
823 rhs
= extract_affine(expr
->getFalseExpr());
825 return isl_pw_aff_cond(cond
, lhs
, rhs
);
828 /* Extract an affine expression, if possible, from "expr".
829 * Otherwise return NULL.
831 __isl_give isl_pw_aff
*PetScan::extract_affine(Expr
*expr
)
833 switch (expr
->getStmtClass()) {
834 case Stmt::ImplicitCastExprClass
:
835 return extract_affine(cast
<ImplicitCastExpr
>(expr
));
836 case Stmt::IntegerLiteralClass
:
837 return extract_affine(cast
<IntegerLiteral
>(expr
));
838 case Stmt::DeclRefExprClass
:
839 return extract_affine(cast
<DeclRefExpr
>(expr
));
840 case Stmt::BinaryOperatorClass
:
841 return extract_affine(cast
<BinaryOperator
>(expr
));
842 case Stmt::UnaryOperatorClass
:
843 return extract_affine(cast
<UnaryOperator
>(expr
));
844 case Stmt::ParenExprClass
:
845 return extract_affine(cast
<ParenExpr
>(expr
));
846 case Stmt::CallExprClass
:
847 return extract_affine(cast
<CallExpr
>(expr
));
848 case Stmt::ArraySubscriptExprClass
:
849 return extract_affine(cast
<ArraySubscriptExpr
>(expr
));
850 case Stmt::ConditionalOperatorClass
:
851 return extract_affine(cast
<ConditionalOperator
>(expr
));
858 __isl_give isl_multi_pw_aff
*PetScan::extract_index(ImplicitCastExpr
*expr
)
860 return extract_index(expr
->getSubExpr());
863 /* Return the depth of an array of the given type.
865 static int array_depth(const Type
*type
)
867 if (type
->isPointerType())
868 return 1 + array_depth(type
->getPointeeType().getTypePtr());
869 if (type
->isArrayType()) {
870 const ArrayType
*atype
;
871 type
= type
->getCanonicalTypeInternal().getTypePtr();
872 atype
= cast
<ArrayType
>(type
);
873 return 1 + array_depth(atype
->getElementType().getTypePtr());
878 /* Return the depth of the array accessed by the index expression "index".
879 * If "index" is an affine expression, i.e., if it does not access
880 * any array, then return 1.
882 static int extract_depth(__isl_keep isl_multi_pw_aff
*index
)
890 if (!isl_multi_pw_aff_has_tuple_id(index
, isl_dim_out
))
893 id
= isl_multi_pw_aff_get_tuple_id(index
, isl_dim_out
);
896 decl
= (ValueDecl
*) isl_id_get_user(id
);
899 return array_depth(decl
->getType().getTypePtr());
902 /* Return the element type of the given array type.
904 static QualType
base_type(QualType qt
)
906 const Type
*type
= qt
.getTypePtr();
908 if (type
->isPointerType())
909 return base_type(type
->getPointeeType());
910 if (type
->isArrayType()) {
911 const ArrayType
*atype
;
912 type
= type
->getCanonicalTypeInternal().getTypePtr();
913 atype
= cast
<ArrayType
>(type
);
914 return base_type(atype
->getElementType());
919 /* Extract an index expression from a reference to a variable.
920 * If the variable has name "A", then the returned index expression
925 __isl_give isl_multi_pw_aff
*PetScan::extract_index(DeclRefExpr
*expr
)
927 return extract_index(expr
->getDecl());
930 /* Extract an index expression from a variable.
931 * If the variable has name "A", then the returned index expression
936 __isl_give isl_multi_pw_aff
*PetScan::extract_index(ValueDecl
*decl
)
938 isl_id
*id
= isl_id_alloc(ctx
, decl
->getName().str().c_str(), decl
);
939 isl_space
*space
= isl_space_alloc(ctx
, 0, 0, 0);
941 space
= isl_space_set_tuple_id(space
, isl_dim_out
, id
);
943 return isl_multi_pw_aff_zero(space
);
946 /* Extract an index expression from an integer contant.
947 * If the value of the constant is "v", then the returned access relation
952 __isl_give isl_multi_pw_aff
*PetScan::extract_index(IntegerLiteral
*expr
)
954 isl_multi_pw_aff
*mpa
;
956 mpa
= isl_multi_pw_aff_from_pw_aff(extract_affine(expr
));
957 mpa
= isl_multi_pw_aff_from_range(mpa
);
961 /* Try and extract an index expression from the given Expr.
962 * Return NULL if it doesn't work out.
964 __isl_give isl_multi_pw_aff
*PetScan::extract_index(Expr
*expr
)
966 switch (expr
->getStmtClass()) {
967 case Stmt::ImplicitCastExprClass
:
968 return extract_index(cast
<ImplicitCastExpr
>(expr
));
969 case Stmt::DeclRefExprClass
:
970 return extract_index(cast
<DeclRefExpr
>(expr
));
971 case Stmt::ArraySubscriptExprClass
:
972 return extract_index(cast
<ArraySubscriptExpr
>(expr
));
973 case Stmt::IntegerLiteralClass
:
974 return extract_index(cast
<IntegerLiteral
>(expr
));
981 /* Extract an index expression from the given array subscript expression.
982 * If nesting is allowed in general, then we turn it on while
983 * examining the index expression.
985 * We first extract an index expression from the base.
986 * This will result in an index expression with a range that corresponds
987 * to the earlier indices.
988 * We then extract the current index, restrict its domain
989 * to those values that result in a non-negative index and
990 * append the index to the base index expression.
992 __isl_give isl_multi_pw_aff
*PetScan::extract_index(ArraySubscriptExpr
*expr
)
994 Expr
*base
= expr
->getBase();
995 Expr
*idx
= expr
->getIdx();
998 isl_multi_pw_aff
*base_access
;
999 isl_multi_pw_aff
*access
;
1001 bool save_nesting
= nesting_enabled
;
1003 nesting_enabled
= allow_nested
;
1005 base_access
= extract_index(base
);
1006 index
= extract_affine(idx
);
1008 nesting_enabled
= save_nesting
;
1010 id
= isl_multi_pw_aff_get_tuple_id(base_access
, isl_dim_set
);
1011 index
= isl_pw_aff_from_range(index
);
1012 domain
= isl_pw_aff_nonneg_set(isl_pw_aff_copy(index
));
1013 index
= isl_pw_aff_intersect_domain(index
, domain
);
1014 access
= isl_multi_pw_aff_from_pw_aff(index
);
1015 access
= isl_multi_pw_aff_flat_range_product(base_access
, access
);
1016 access
= isl_multi_pw_aff_set_tuple_id(access
, isl_dim_set
, id
);
1021 /* Check if "expr" calls function "minmax" with two arguments and if so
1022 * make lhs and rhs refer to these two arguments.
1024 static bool is_minmax(Expr
*expr
, const char *minmax
, Expr
*&lhs
, Expr
*&rhs
)
1030 if (expr
->getStmtClass() != Stmt::CallExprClass
)
1033 call
= cast
<CallExpr
>(expr
);
1034 fd
= call
->getDirectCallee();
1038 if (call
->getNumArgs() != 2)
1041 name
= fd
->getDeclName().getAsString();
1045 lhs
= call
->getArg(0);
1046 rhs
= call
->getArg(1);
1051 /* Check if "expr" is of the form min(lhs, rhs) and if so make
1052 * lhs and rhs refer to the two arguments.
1054 static bool is_min(Expr
*expr
, Expr
*&lhs
, Expr
*&rhs
)
1056 return is_minmax(expr
, "min", lhs
, rhs
);
1059 /* Check if "expr" is of the form max(lhs, rhs) and if so make
1060 * lhs and rhs refer to the two arguments.
1062 static bool is_max(Expr
*expr
, Expr
*&lhs
, Expr
*&rhs
)
1064 return is_minmax(expr
, "max", lhs
, rhs
);
1067 /* Return "lhs && rhs", defined on the shared definition domain.
1069 static __isl_give isl_pw_aff
*pw_aff_and(__isl_take isl_pw_aff
*lhs
,
1070 __isl_take isl_pw_aff
*rhs
)
1075 dom
= isl_set_intersect(isl_pw_aff_domain(isl_pw_aff_copy(lhs
)),
1076 isl_pw_aff_domain(isl_pw_aff_copy(rhs
)));
1077 cond
= isl_set_intersect(isl_pw_aff_non_zero_set(lhs
),
1078 isl_pw_aff_non_zero_set(rhs
));
1079 return indicator_function(cond
, dom
);
1082 /* Return "lhs && rhs", with shortcut semantics.
1083 * That is, if lhs is false, then the result is defined even if rhs is not.
1084 * In practice, we compute lhs ? rhs : lhs.
1086 static __isl_give isl_pw_aff
*pw_aff_and_then(__isl_take isl_pw_aff
*lhs
,
1087 __isl_take isl_pw_aff
*rhs
)
1089 return isl_pw_aff_cond(isl_pw_aff_copy(lhs
), rhs
, lhs
);
1092 /* Return "lhs || rhs", with shortcut semantics.
1093 * That is, if lhs is true, then the result is defined even if rhs is not.
1094 * In practice, we compute lhs ? lhs : rhs.
1096 static __isl_give isl_pw_aff
*pw_aff_or_else(__isl_take isl_pw_aff
*lhs
,
1097 __isl_take isl_pw_aff
*rhs
)
1099 return isl_pw_aff_cond(isl_pw_aff_copy(lhs
), lhs
, rhs
);
1102 /* Extract an affine expressions representing the comparison "LHS op RHS"
1103 * "comp" is the original statement that "LHS op RHS" is derived from
1104 * and is used for diagnostics.
1106 * If the comparison is of the form
1110 * then the expression is constructed as the conjunction of
1115 * A similar optimization is performed for max(a,b) <= c.
1116 * We do this because that will lead to simpler representations
1117 * of the expression.
1118 * If isl is ever enhanced to explicitly deal with min and max expressions,
1119 * this optimization can be removed.
1121 __isl_give isl_pw_aff
*PetScan::extract_comparison(BinaryOperatorKind op
,
1122 Expr
*LHS
, Expr
*RHS
, Stmt
*comp
)
1131 return extract_comparison(BO_LT
, RHS
, LHS
, comp
);
1133 return extract_comparison(BO_LE
, RHS
, LHS
, comp
);
1135 if (op
== BO_LT
|| op
== BO_LE
) {
1136 Expr
*expr1
, *expr2
;
1137 if (is_min(RHS
, expr1
, expr2
)) {
1138 lhs
= extract_comparison(op
, LHS
, expr1
, comp
);
1139 rhs
= extract_comparison(op
, LHS
, expr2
, comp
);
1140 return pw_aff_and(lhs
, rhs
);
1142 if (is_max(LHS
, expr1
, expr2
)) {
1143 lhs
= extract_comparison(op
, expr1
, RHS
, comp
);
1144 rhs
= extract_comparison(op
, expr2
, RHS
, comp
);
1145 return pw_aff_and(lhs
, rhs
);
1149 lhs
= extract_affine(LHS
);
1150 rhs
= extract_affine(RHS
);
1152 dom
= isl_pw_aff_domain(isl_pw_aff_copy(lhs
));
1153 dom
= isl_set_intersect(dom
, isl_pw_aff_domain(isl_pw_aff_copy(rhs
)));
1157 cond
= isl_pw_aff_lt_set(lhs
, rhs
);
1160 cond
= isl_pw_aff_le_set(lhs
, rhs
);
1163 cond
= isl_pw_aff_eq_set(lhs
, rhs
);
1166 cond
= isl_pw_aff_ne_set(lhs
, rhs
);
1169 isl_pw_aff_free(lhs
);
1170 isl_pw_aff_free(rhs
);
1176 cond
= isl_set_coalesce(cond
);
1177 res
= indicator_function(cond
, dom
);
1182 __isl_give isl_pw_aff
*PetScan::extract_comparison(BinaryOperator
*comp
)
1184 return extract_comparison(comp
->getOpcode(), comp
->getLHS(),
1185 comp
->getRHS(), comp
);
1188 /* Extract an affine expression representing the negation (logical not)
1189 * of a subexpression.
1191 __isl_give isl_pw_aff
*PetScan::extract_boolean(UnaryOperator
*op
)
1193 isl_set
*set_cond
, *dom
;
1194 isl_pw_aff
*cond
, *res
;
1196 cond
= extract_condition(op
->getSubExpr());
1198 dom
= isl_pw_aff_domain(isl_pw_aff_copy(cond
));
1200 set_cond
= isl_pw_aff_zero_set(cond
);
1202 res
= indicator_function(set_cond
, dom
);
1207 /* Extract an affine expression representing the disjunction (logical or)
1208 * or conjunction (logical and) of two subexpressions.
1210 __isl_give isl_pw_aff
*PetScan::extract_boolean(BinaryOperator
*comp
)
1212 isl_pw_aff
*lhs
, *rhs
;
1214 lhs
= extract_condition(comp
->getLHS());
1215 rhs
= extract_condition(comp
->getRHS());
1217 switch (comp
->getOpcode()) {
1219 return pw_aff_and_then(lhs
, rhs
);
1221 return pw_aff_or_else(lhs
, rhs
);
1223 isl_pw_aff_free(lhs
);
1224 isl_pw_aff_free(rhs
);
1231 __isl_give isl_pw_aff
*PetScan::extract_condition(UnaryOperator
*expr
)
1233 switch (expr
->getOpcode()) {
1235 return extract_boolean(expr
);
1242 /* Extract the affine expression "expr != 0 ? 1 : 0".
1244 __isl_give isl_pw_aff
*PetScan::extract_implicit_condition(Expr
*expr
)
1249 res
= extract_affine(expr
);
1251 dom
= isl_pw_aff_domain(isl_pw_aff_copy(res
));
1252 set
= isl_pw_aff_non_zero_set(res
);
1254 res
= indicator_function(set
, dom
);
1259 /* Extract an affine expression from a boolean expression.
1260 * In particular, return the expression "expr ? 1 : 0".
1262 * If the expression doesn't look like a condition, we assume it
1263 * is an affine expression and return the condition "expr != 0 ? 1 : 0".
1265 __isl_give isl_pw_aff
*PetScan::extract_condition(Expr
*expr
)
1267 BinaryOperator
*comp
;
1270 isl_set
*u
= isl_set_universe(isl_space_params_alloc(ctx
, 0));
1271 return indicator_function(u
, isl_set_copy(u
));
1274 if (expr
->getStmtClass() == Stmt::ParenExprClass
)
1275 return extract_condition(cast
<ParenExpr
>(expr
)->getSubExpr());
1277 if (expr
->getStmtClass() == Stmt::UnaryOperatorClass
)
1278 return extract_condition(cast
<UnaryOperator
>(expr
));
1280 if (expr
->getStmtClass() != Stmt::BinaryOperatorClass
)
1281 return extract_implicit_condition(expr
);
1283 comp
= cast
<BinaryOperator
>(expr
);
1284 switch (comp
->getOpcode()) {
1291 return extract_comparison(comp
);
1294 return extract_boolean(comp
);
1296 return extract_implicit_condition(expr
);
1300 static enum pet_op_type
UnaryOperatorKind2pet_op_type(UnaryOperatorKind kind
)
1304 return pet_op_minus
;
1306 return pet_op_post_inc
;
1308 return pet_op_post_dec
;
1310 return pet_op_pre_inc
;
1312 return pet_op_pre_dec
;
1318 static enum pet_op_type
BinaryOperatorKind2pet_op_type(BinaryOperatorKind kind
)
1322 return pet_op_add_assign
;
1324 return pet_op_sub_assign
;
1326 return pet_op_mul_assign
;
1328 return pet_op_div_assign
;
1330 return pet_op_assign
;
1354 /* Construct a pet_expr representing a unary operator expression.
1356 struct pet_expr
*PetScan::extract_expr(UnaryOperator
*expr
)
1358 struct pet_expr
*arg
;
1359 enum pet_op_type op
;
1361 op
= UnaryOperatorKind2pet_op_type(expr
->getOpcode());
1362 if (op
== pet_op_last
) {
1367 arg
= extract_expr(expr
->getSubExpr());
1369 if (expr
->isIncrementDecrementOp() &&
1370 arg
&& arg
->type
== pet_expr_access
) {
1375 return pet_expr_new_unary(ctx
, op
, arg
);
1378 /* Mark the given access pet_expr as a write.
1379 * If a scalar is being accessed, then mark its value
1380 * as unknown in assigned_value.
1382 void PetScan::mark_write(struct pet_expr
*access
)
1390 access
->acc
.write
= 1;
1391 access
->acc
.read
= 0;
1393 if (!pet_expr_is_scalar_access(access
))
1396 id
= pet_expr_access_get_id(access
);
1397 decl
= (ValueDecl
*) isl_id_get_user(id
);
1398 clear_assignment(assigned_value
, decl
);
1402 /* Assign "rhs" to "lhs".
1404 * In particular, if "lhs" is a scalar variable, then mark
1405 * the variable as having been assigned. If, furthermore, "rhs"
1406 * is an affine expression, then keep track of this value in assigned_value
1407 * so that we can plug it in when we later come across the same variable.
1409 void PetScan::assign(struct pet_expr
*lhs
, Expr
*rhs
)
1417 if (!pet_expr_is_scalar_access(lhs
))
1420 id
= pet_expr_access_get_id(lhs
);
1421 decl
= (ValueDecl
*) isl_id_get_user(id
);
1424 pa
= try_extract_affine(rhs
);
1425 clear_assignment(assigned_value
, decl
);
1428 assigned_value
[decl
] = pa
;
1429 insert_expression(pa
);
1432 /* Construct a pet_expr representing a binary operator expression.
1434 * If the top level operator is an assignment and the LHS is an access,
1435 * then we mark that access as a write. If the operator is a compound
1436 * assignment, the access is marked as both a read and a write.
1438 * If "expr" assigns something to a scalar variable, then we mark
1439 * the variable as having been assigned. If, furthermore, the expression
1440 * is affine, then keep track of this value in assigned_value
1441 * so that we can plug it in when we later come across the same variable.
1443 struct pet_expr
*PetScan::extract_expr(BinaryOperator
*expr
)
1445 struct pet_expr
*lhs
, *rhs
;
1446 enum pet_op_type op
;
1448 op
= BinaryOperatorKind2pet_op_type(expr
->getOpcode());
1449 if (op
== pet_op_last
) {
1454 lhs
= extract_expr(expr
->getLHS());
1455 rhs
= extract_expr(expr
->getRHS());
1457 if (expr
->isAssignmentOp() && lhs
&& lhs
->type
== pet_expr_access
) {
1459 if (expr
->isCompoundAssignmentOp())
1463 if (expr
->getOpcode() == BO_Assign
)
1464 assign(lhs
, expr
->getRHS());
1466 return pet_expr_new_binary(ctx
, op
, lhs
, rhs
);
1469 /* Construct a pet_scop with a single statement killing the entire
1472 struct pet_scop
*PetScan::kill(Stmt
*stmt
, struct pet_array
*array
)
1476 isl_multi_pw_aff
*index
;
1478 struct pet_expr
*expr
;
1482 access
= isl_map_from_range(isl_set_copy(array
->extent
));
1483 id
= isl_set_get_tuple_id(array
->extent
);
1484 space
= isl_space_alloc(ctx
, 0, 0, 0);
1485 space
= isl_space_set_tuple_id(space
, isl_dim_out
, id
);
1486 index
= isl_multi_pw_aff_zero(space
);
1487 expr
= pet_expr_kill_from_access_and_index(access
, index
);
1488 return extract(stmt
, expr
);
1491 /* Construct a pet_scop for a (single) variable declaration.
1493 * The scop contains the variable being declared (as an array)
1494 * and a statement killing the array.
1496 * If the variable is initialized in the AST, then the scop
1497 * also contains an assignment to the variable.
1499 struct pet_scop
*PetScan::extract(DeclStmt
*stmt
)
1503 struct pet_expr
*lhs
, *rhs
, *pe
;
1504 struct pet_scop
*scop_decl
, *scop
;
1505 struct pet_array
*array
;
1507 if (!stmt
->isSingleDecl()) {
1512 decl
= stmt
->getSingleDecl();
1513 vd
= cast
<VarDecl
>(decl
);
1515 array
= extract_array(ctx
, vd
);
1517 array
->declared
= 1;
1518 scop_decl
= kill(stmt
, array
);
1519 scop_decl
= pet_scop_add_array(scop_decl
, array
);
1524 lhs
= extract_access_expr(vd
);
1525 rhs
= extract_expr(vd
->getInit());
1528 assign(lhs
, vd
->getInit());
1530 pe
= pet_expr_new_binary(ctx
, pet_op_assign
, lhs
, rhs
);
1531 scop
= extract(stmt
, pe
);
1533 scop_decl
= pet_scop_prefix(scop_decl
, 0);
1534 scop
= pet_scop_prefix(scop
, 1);
1536 scop
= pet_scop_add_seq(ctx
, scop_decl
, scop
);
1541 /* Construct a pet_expr representing a conditional operation.
1543 * We first try to extract the condition as an affine expression.
1544 * If that fails, we construct a pet_expr tree representing the condition.
1546 struct pet_expr
*PetScan::extract_expr(ConditionalOperator
*expr
)
1548 struct pet_expr
*cond
, *lhs
, *rhs
;
1551 pa
= try_extract_affine(expr
->getCond());
1553 isl_multi_pw_aff
*test
= isl_multi_pw_aff_from_pw_aff(pa
);
1554 test
= isl_multi_pw_aff_from_range(test
);
1555 cond
= pet_expr_from_index(test
);
1557 cond
= extract_expr(expr
->getCond());
1558 lhs
= extract_expr(expr
->getTrueExpr());
1559 rhs
= extract_expr(expr
->getFalseExpr());
1561 return pet_expr_new_ternary(ctx
, cond
, lhs
, rhs
);
1564 struct pet_expr
*PetScan::extract_expr(ImplicitCastExpr
*expr
)
1566 return extract_expr(expr
->getSubExpr());
1569 /* Construct a pet_expr representing a floating point value.
1571 * If the floating point literal does not appear in a macro,
1572 * then we use the original representation in the source code
1573 * as the string representation. Otherwise, we use the pretty
1574 * printer to produce a string representation.
1576 struct pet_expr
*PetScan::extract_expr(FloatingLiteral
*expr
)
1580 const LangOptions
&LO
= PP
.getLangOpts();
1581 SourceLocation loc
= expr
->getLocation();
1583 if (!loc
.isMacroID()) {
1584 SourceManager
&SM
= PP
.getSourceManager();
1585 unsigned len
= Lexer::MeasureTokenLength(loc
, SM
, LO
);
1586 s
= string(SM
.getCharacterData(loc
), len
);
1588 llvm::raw_string_ostream
S(s
);
1589 expr
->printPretty(S
, 0, PrintingPolicy(LO
));
1592 d
= expr
->getValueAsApproximateDouble();
1593 return pet_expr_new_double(ctx
, d
, s
.c_str());
1596 /* Extract an index expression from "expr" and then convert it into
1597 * an access pet_expr.
1599 struct pet_expr
*PetScan::extract_access_expr(Expr
*expr
)
1601 isl_multi_pw_aff
*index
;
1602 struct pet_expr
*pe
;
1605 index
= extract_index(expr
);
1606 depth
= extract_depth(index
);
1608 pe
= pet_expr_from_index_and_depth(index
, depth
);
1613 /* Extract an index expression from "decl" and then convert it into
1614 * an access pet_expr.
1616 struct pet_expr
*PetScan::extract_access_expr(ValueDecl
*decl
)
1618 isl_multi_pw_aff
*index
;
1619 struct pet_expr
*pe
;
1622 index
= extract_index(decl
);
1623 depth
= extract_depth(index
);
1625 pe
= pet_expr_from_index_and_depth(index
, depth
);
1630 struct pet_expr
*PetScan::extract_expr(ParenExpr
*expr
)
1632 return extract_expr(expr
->getSubExpr());
1635 /* Construct a pet_expr representing a function call.
1637 * If we are passing along a pointer to an array element
1638 * or an entire row or even higher dimensional slice of an array,
1639 * then the function being called may write into the array.
1641 * We assume here that if the function is declared to take a pointer
1642 * to a const type, then the function will perform a read
1643 * and that otherwise, it will perform a write.
1645 struct pet_expr
*PetScan::extract_expr(CallExpr
*expr
)
1647 struct pet_expr
*res
= NULL
;
1651 fd
= expr
->getDirectCallee();
1657 name
= fd
->getDeclName().getAsString();
1658 res
= pet_expr_new_call(ctx
, name
.c_str(), expr
->getNumArgs());
1662 for (int i
= 0; i
< expr
->getNumArgs(); ++i
) {
1663 Expr
*arg
= expr
->getArg(i
);
1667 if (arg
->getStmtClass() == Stmt::ImplicitCastExprClass
) {
1668 ImplicitCastExpr
*ice
= cast
<ImplicitCastExpr
>(arg
);
1669 arg
= ice
->getSubExpr();
1671 if (arg
->getStmtClass() == Stmt::UnaryOperatorClass
) {
1672 UnaryOperator
*op
= cast
<UnaryOperator
>(arg
);
1673 if (op
->getOpcode() == UO_AddrOf
) {
1675 arg
= op
->getSubExpr();
1678 res
->args
[i
] = PetScan::extract_expr(arg
);
1679 main_arg
= res
->args
[i
];
1681 res
->args
[i
] = pet_expr_new_unary(ctx
,
1682 pet_op_address_of
, res
->args
[i
]);
1685 if (arg
->getStmtClass() == Stmt::ArraySubscriptExprClass
&&
1686 array_depth(arg
->getType().getTypePtr()) > 0)
1688 if (is_addr
&& main_arg
->type
== pet_expr_access
) {
1690 if (!fd
->hasPrototype()) {
1691 unsupported(expr
, "prototype required");
1694 parm
= fd
->getParamDecl(i
);
1695 if (!const_base(parm
->getType()))
1696 mark_write(main_arg
);
1706 /* Construct a pet_expr representing a (C style) cast.
1708 struct pet_expr
*PetScan::extract_expr(CStyleCastExpr
*expr
)
1710 struct pet_expr
*arg
;
1713 arg
= extract_expr(expr
->getSubExpr());
1717 type
= expr
->getTypeAsWritten();
1718 return pet_expr_new_cast(ctx
, type
.getAsString().c_str(), arg
);
1721 /* Try and onstruct a pet_expr representing "expr".
1723 struct pet_expr
*PetScan::extract_expr(Expr
*expr
)
1725 switch (expr
->getStmtClass()) {
1726 case Stmt::UnaryOperatorClass
:
1727 return extract_expr(cast
<UnaryOperator
>(expr
));
1728 case Stmt::CompoundAssignOperatorClass
:
1729 case Stmt::BinaryOperatorClass
:
1730 return extract_expr(cast
<BinaryOperator
>(expr
));
1731 case Stmt::ImplicitCastExprClass
:
1732 return extract_expr(cast
<ImplicitCastExpr
>(expr
));
1733 case Stmt::ArraySubscriptExprClass
:
1734 case Stmt::DeclRefExprClass
:
1735 case Stmt::IntegerLiteralClass
:
1736 return extract_access_expr(expr
);
1737 case Stmt::FloatingLiteralClass
:
1738 return extract_expr(cast
<FloatingLiteral
>(expr
));
1739 case Stmt::ParenExprClass
:
1740 return extract_expr(cast
<ParenExpr
>(expr
));
1741 case Stmt::ConditionalOperatorClass
:
1742 return extract_expr(cast
<ConditionalOperator
>(expr
));
1743 case Stmt::CallExprClass
:
1744 return extract_expr(cast
<CallExpr
>(expr
));
1745 case Stmt::CStyleCastExprClass
:
1746 return extract_expr(cast
<CStyleCastExpr
>(expr
));
1753 /* Check if the given initialization statement is an assignment.
1754 * If so, return that assignment. Otherwise return NULL.
1756 BinaryOperator
*PetScan::initialization_assignment(Stmt
*init
)
1758 BinaryOperator
*ass
;
1760 if (init
->getStmtClass() != Stmt::BinaryOperatorClass
)
1763 ass
= cast
<BinaryOperator
>(init
);
1764 if (ass
->getOpcode() != BO_Assign
)
1770 /* Check if the given initialization statement is a declaration
1771 * of a single variable.
1772 * If so, return that declaration. Otherwise return NULL.
1774 Decl
*PetScan::initialization_declaration(Stmt
*init
)
1778 if (init
->getStmtClass() != Stmt::DeclStmtClass
)
1781 decl
= cast
<DeclStmt
>(init
);
1783 if (!decl
->isSingleDecl())
1786 return decl
->getSingleDecl();
1789 /* Given the assignment operator in the initialization of a for loop,
1790 * extract the induction variable, i.e., the (integer)variable being
1793 ValueDecl
*PetScan::extract_induction_variable(BinaryOperator
*init
)
1800 lhs
= init
->getLHS();
1801 if (lhs
->getStmtClass() != Stmt::DeclRefExprClass
) {
1806 ref
= cast
<DeclRefExpr
>(lhs
);
1807 decl
= ref
->getDecl();
1808 type
= decl
->getType().getTypePtr();
1810 if (!type
->isIntegerType()) {
1818 /* Given the initialization statement of a for loop and the single
1819 * declaration in this initialization statement,
1820 * extract the induction variable, i.e., the (integer) variable being
1823 VarDecl
*PetScan::extract_induction_variable(Stmt
*init
, Decl
*decl
)
1827 vd
= cast
<VarDecl
>(decl
);
1829 const QualType type
= vd
->getType();
1830 if (!type
->isIntegerType()) {
1835 if (!vd
->getInit()) {
1843 /* Check that op is of the form iv++ or iv--.
1844 * Return an affine expression "1" or "-1" accordingly.
1846 __isl_give isl_pw_aff
*PetScan::extract_unary_increment(
1847 clang::UnaryOperator
*op
, clang::ValueDecl
*iv
)
1854 if (!op
->isIncrementDecrementOp()) {
1859 sub
= op
->getSubExpr();
1860 if (sub
->getStmtClass() != Stmt::DeclRefExprClass
) {
1865 ref
= cast
<DeclRefExpr
>(sub
);
1866 if (ref
->getDecl() != iv
) {
1871 space
= isl_space_params_alloc(ctx
, 0);
1872 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
1874 if (op
->isIncrementOp())
1875 aff
= isl_aff_add_constant_si(aff
, 1);
1877 aff
= isl_aff_add_constant_si(aff
, -1);
1879 return isl_pw_aff_from_aff(aff
);
1882 /* If the isl_pw_aff on which isl_pw_aff_foreach_piece is called
1883 * has a single constant expression, then put this constant in *user.
1884 * The caller is assumed to have checked that this function will
1885 * be called exactly once.
1887 static int extract_cst(__isl_take isl_set
*set
, __isl_take isl_aff
*aff
,
1890 isl_val
**inc
= (isl_val
**)user
;
1893 if (isl_aff_is_cst(aff
))
1894 *inc
= isl_aff_get_constant_val(aff
);
1904 /* Check if op is of the form
1908 * and return inc as an affine expression.
1910 * We extract an affine expression from the RHS, subtract iv and return
1913 __isl_give isl_pw_aff
*PetScan::extract_binary_increment(BinaryOperator
*op
,
1914 clang::ValueDecl
*iv
)
1923 if (op
->getOpcode() != BO_Assign
) {
1929 if (lhs
->getStmtClass() != Stmt::DeclRefExprClass
) {
1934 ref
= cast
<DeclRefExpr
>(lhs
);
1935 if (ref
->getDecl() != iv
) {
1940 val
= extract_affine(op
->getRHS());
1942 id
= isl_id_alloc(ctx
, iv
->getName().str().c_str(), iv
);
1944 dim
= isl_space_params_alloc(ctx
, 1);
1945 dim
= isl_space_set_dim_id(dim
, isl_dim_param
, 0, id
);
1946 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
1947 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_param
, 0, 1);
1949 val
= isl_pw_aff_sub(val
, isl_pw_aff_from_aff(aff
));
1954 /* Check that op is of the form iv += cst or iv -= cst
1955 * and return an affine expression corresponding oto cst or -cst accordingly.
1957 __isl_give isl_pw_aff
*PetScan::extract_compound_increment(
1958 CompoundAssignOperator
*op
, clang::ValueDecl
*iv
)
1964 BinaryOperatorKind opcode
;
1966 opcode
= op
->getOpcode();
1967 if (opcode
!= BO_AddAssign
&& opcode
!= BO_SubAssign
) {
1971 if (opcode
== BO_SubAssign
)
1975 if (lhs
->getStmtClass() != Stmt::DeclRefExprClass
) {
1980 ref
= cast
<DeclRefExpr
>(lhs
);
1981 if (ref
->getDecl() != iv
) {
1986 val
= extract_affine(op
->getRHS());
1988 val
= isl_pw_aff_neg(val
);
1993 /* Check that the increment of the given for loop increments
1994 * (or decrements) the induction variable "iv" and return
1995 * the increment as an affine expression if successful.
1997 __isl_give isl_pw_aff
*PetScan::extract_increment(clang::ForStmt
*stmt
,
2000 Stmt
*inc
= stmt
->getInc();
2007 if (inc
->getStmtClass() == Stmt::UnaryOperatorClass
)
2008 return extract_unary_increment(cast
<UnaryOperator
>(inc
), iv
);
2009 if (inc
->getStmtClass() == Stmt::CompoundAssignOperatorClass
)
2010 return extract_compound_increment(
2011 cast
<CompoundAssignOperator
>(inc
), iv
);
2012 if (inc
->getStmtClass() == Stmt::BinaryOperatorClass
)
2013 return extract_binary_increment(cast
<BinaryOperator
>(inc
), iv
);
2019 /* Embed the given iteration domain in an extra outer loop
2020 * with induction variable "var".
2021 * If this variable appeared as a parameter in the constraints,
2022 * it is replaced by the new outermost dimension.
2024 static __isl_give isl_set
*embed(__isl_take isl_set
*set
,
2025 __isl_take isl_id
*var
)
2029 set
= isl_set_insert_dims(set
, isl_dim_set
, 0, 1);
2030 pos
= isl_set_find_dim_by_id(set
, isl_dim_param
, var
);
2032 set
= isl_set_equate(set
, isl_dim_param
, pos
, isl_dim_set
, 0);
2033 set
= isl_set_project_out(set
, isl_dim_param
, pos
, 1);
2040 /* Return those elements in the space of "cond" that come after
2041 * (based on "sign") an element in "cond".
2043 static __isl_give isl_set
*after(__isl_take isl_set
*cond
, int sign
)
2045 isl_map
*previous_to_this
;
2048 previous_to_this
= isl_map_lex_lt(isl_set_get_space(cond
));
2050 previous_to_this
= isl_map_lex_gt(isl_set_get_space(cond
));
2052 cond
= isl_set_apply(cond
, previous_to_this
);
2057 /* Create the infinite iteration domain
2059 * { [id] : id >= 0 }
2061 * If "scop" has an affine skip of type pet_skip_later,
2062 * then remove those iterations i that have an earlier iteration
2063 * where the skip condition is satisfied, meaning that iteration i
2065 * Since we are dealing with a loop without loop iterator,
2066 * the skip condition cannot refer to the current loop iterator and
2067 * so effectively, the returned set is of the form
2069 * { [0]; [id] : id >= 1 and not skip }
2071 static __isl_give isl_set
*infinite_domain(__isl_take isl_id
*id
,
2072 struct pet_scop
*scop
)
2074 isl_ctx
*ctx
= isl_id_get_ctx(id
);
2078 domain
= isl_set_nat_universe(isl_space_set_alloc(ctx
, 0, 1));
2079 domain
= isl_set_set_dim_id(domain
, isl_dim_set
, 0, id
);
2081 if (!pet_scop_has_affine_skip(scop
, pet_skip_later
))
2084 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_later
);
2085 skip
= embed(skip
, isl_id_copy(id
));
2086 skip
= isl_set_intersect(skip
, isl_set_copy(domain
));
2087 domain
= isl_set_subtract(domain
, after(skip
, 1));
2092 /* Create an identity affine expression on the space containing "domain",
2093 * which is assumed to be one-dimensional.
2095 static __isl_give isl_aff
*identity_aff(__isl_keep isl_set
*domain
)
2097 isl_local_space
*ls
;
2099 ls
= isl_local_space_from_space(isl_set_get_space(domain
));
2100 return isl_aff_var_on_domain(ls
, isl_dim_set
, 0);
2103 /* Create an affine expression that maps elements
2104 * of a single-dimensional array "id_test" to the previous element
2105 * (according to "inc"), provided this element belongs to "domain".
2106 * That is, create the affine expression
2108 * { id[x] -> id[x - inc] : x - inc in domain }
2110 static __isl_give isl_multi_pw_aff
*map_to_previous(__isl_take isl_id
*id_test
,
2111 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
2114 isl_local_space
*ls
;
2116 isl_multi_pw_aff
*prev
;
2118 space
= isl_set_get_space(domain
);
2119 ls
= isl_local_space_from_space(space
);
2120 aff
= isl_aff_var_on_domain(ls
, isl_dim_set
, 0);
2121 aff
= isl_aff_add_constant_val(aff
, isl_val_neg(inc
));
2122 prev
= isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff
));
2123 domain
= isl_set_preimage_multi_pw_aff(domain
,
2124 isl_multi_pw_aff_copy(prev
));
2125 prev
= isl_multi_pw_aff_intersect_domain(prev
, domain
);
2126 prev
= isl_multi_pw_aff_set_tuple_id(prev
, isl_dim_out
, id_test
);
2131 /* Add an implication to "scop" expressing that if an element of
2132 * virtual array "id_test" has value "satisfied" then all previous elements
2133 * of this array also have that value. The set of previous elements
2134 * is bounded by "domain". If "sign" is negative then iterator
2135 * is decreasing and we express that all subsequent array elements
2136 * (but still defined previously) have the same value.
2138 static struct pet_scop
*add_implication(struct pet_scop
*scop
,
2139 __isl_take isl_id
*id_test
, __isl_take isl_set
*domain
, int sign
,
2145 domain
= isl_set_set_tuple_id(domain
, id_test
);
2146 space
= isl_set_get_space(domain
);
2148 map
= isl_map_lex_ge(space
);
2150 map
= isl_map_lex_le(space
);
2151 map
= isl_map_intersect_range(map
, domain
);
2152 scop
= pet_scop_add_implication(scop
, map
, satisfied
);
2157 /* Add a filter to "scop" that imposes that it is only executed
2158 * when the variable identified by "id_test" has a zero value
2159 * for all previous iterations of "domain".
2161 * In particular, add a filter that imposes that the array
2162 * has a zero value at the previous iteration of domain and
2163 * add an implication that implies that it then has that
2164 * value for all previous iterations.
2166 static struct pet_scop
*scop_add_break(struct pet_scop
*scop
,
2167 __isl_take isl_id
*id_test
, __isl_take isl_set
*domain
,
2168 __isl_take isl_val
*inc
)
2170 isl_multi_pw_aff
*prev
;
2171 int sign
= isl_val_sgn(inc
);
2173 prev
= map_to_previous(isl_id_copy(id_test
), isl_set_copy(domain
), inc
);
2174 scop
= add_implication(scop
, id_test
, domain
, sign
, 0);
2175 scop
= pet_scop_filter(scop
, prev
, 0);
2180 /* Construct a pet_scop for an infinite loop around the given body.
2182 * We extract a pet_scop for the body and then embed it in a loop with
2191 * If the body contains any break, then it is taken into
2192 * account in infinite_domain (if the skip condition is affine)
2193 * or in scop_add_break (if the skip condition is not affine).
2195 struct pet_scop
*PetScan::extract_infinite_loop(Stmt
*body
)
2197 isl_id
*id
, *id_test
;
2200 struct pet_scop
*scop
;
2203 scop
= extract(body
);
2207 id
= isl_id_alloc(ctx
, "t", NULL
);
2208 domain
= infinite_domain(isl_id_copy(id
), scop
);
2209 ident
= identity_aff(domain
);
2211 has_var_break
= pet_scop_has_var_skip(scop
, pet_skip_later
);
2213 id_test
= pet_scop_get_skip_id(scop
, pet_skip_later
);
2215 scop
= pet_scop_embed(scop
, isl_set_copy(domain
),
2216 isl_map_from_aff(isl_aff_copy(ident
)), ident
, id
);
2218 scop
= scop_add_break(scop
, id_test
, domain
, isl_val_one(ctx
));
2220 isl_set_free(domain
);
2225 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
2231 struct pet_scop
*PetScan::extract_infinite_for(ForStmt
*stmt
)
2233 return extract_infinite_loop(stmt
->getBody());
2236 /* Create an index expression for an access to a virtual array
2237 * representing the result of a condition.
2238 * Unlike other accessed data, the id of the array is NULL as
2239 * there is no ValueDecl in the program corresponding to the virtual
2241 * The array starts out as a scalar, but grows along with the
2242 * statement writing to the array in pet_scop_embed.
2244 static __isl_give isl_multi_pw_aff
*create_test_index(isl_ctx
*ctx
, int test_nr
)
2246 isl_space
*dim
= isl_space_alloc(ctx
, 0, 0, 0);
2250 snprintf(name
, sizeof(name
), "__pet_test_%d", test_nr
);
2251 id
= isl_id_alloc(ctx
, name
, NULL
);
2252 dim
= isl_space_set_tuple_id(dim
, isl_dim_out
, id
);
2253 return isl_multi_pw_aff_zero(dim
);
2256 /* Add an array with the given extent (range of "index") to the list
2257 * of arrays in "scop" and return the extended pet_scop.
2258 * The array is marked as attaining values 0 and 1 only and
2259 * as each element being assigned at most once.
2261 static struct pet_scop
*scop_add_array(struct pet_scop
*scop
,
2262 __isl_keep isl_multi_pw_aff
*index
, clang::ASTContext
&ast_ctx
)
2264 isl_ctx
*ctx
= isl_multi_pw_aff_get_ctx(index
);
2266 struct pet_array
*array
;
2274 array
= isl_calloc_type(ctx
, struct pet_array
);
2278 access
= isl_map_from_multi_pw_aff(isl_multi_pw_aff_copy(index
));
2279 array
->extent
= isl_map_range(access
);
2280 dim
= isl_space_params_alloc(ctx
, 0);
2281 array
->context
= isl_set_universe(dim
);
2282 dim
= isl_space_set_alloc(ctx
, 0, 1);
2283 array
->value_bounds
= isl_set_universe(dim
);
2284 array
->value_bounds
= isl_set_lower_bound_si(array
->value_bounds
,
2286 array
->value_bounds
= isl_set_upper_bound_si(array
->value_bounds
,
2288 array
->element_type
= strdup("int");
2289 array
->element_size
= ast_ctx
.getTypeInfo(ast_ctx
.IntTy
).first
/ 8;
2290 array
->uniquely_defined
= 1;
2292 if (!array
->extent
|| !array
->context
)
2293 array
= pet_array_free(array
);
2295 scop
= pet_scop_add_array(scop
, array
);
2299 pet_scop_free(scop
);
2303 /* Construct a pet_scop for a while loop of the form
2308 * In particular, construct a scop for an infinite loop around body and
2309 * intersect the domain with the affine expression.
2310 * Note that this intersection may result in an empty loop.
2312 struct pet_scop
*PetScan::extract_affine_while(__isl_take isl_pw_aff
*pa
,
2315 struct pet_scop
*scop
;
2319 valid
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
2320 dom
= isl_pw_aff_non_zero_set(pa
);
2321 scop
= extract_infinite_loop(body
);
2322 scop
= pet_scop_restrict(scop
, dom
);
2323 scop
= pet_scop_restrict_context(scop
, valid
);
2328 /* Construct a scop for a while, given the scops for the condition
2329 * and the body, the filter identifier and the iteration domain of
2332 * In particular, the scop for the condition is filtered to depend
2333 * on "id_test" evaluating to true for all previous iterations
2334 * of the loop, while the scop for the body is filtered to depend
2335 * on "id_test" evaluating to true for all iterations up to the
2336 * current iteration.
2337 * The actual filter only imposes that this virtual array has
2338 * value one on the previous or the current iteration.
2339 * The fact that this condition also applies to the previous
2340 * iterations is enforced by an implication.
2342 * These filtered scops are then combined into a single scop.
2344 * "sign" is positive if the iterator increases and negative
2347 static struct pet_scop
*scop_add_while(struct pet_scop
*scop_cond
,
2348 struct pet_scop
*scop_body
, __isl_take isl_id
*id_test
,
2349 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
2351 isl_ctx
*ctx
= isl_set_get_ctx(domain
);
2353 isl_multi_pw_aff
*test_index
;
2354 isl_multi_pw_aff
*prev
;
2355 int sign
= isl_val_sgn(inc
);
2356 struct pet_scop
*scop
;
2358 prev
= map_to_previous(isl_id_copy(id_test
), isl_set_copy(domain
), inc
);
2359 scop_cond
= pet_scop_filter(scop_cond
, prev
, 1);
2361 space
= isl_space_map_from_set(isl_set_get_space(domain
));
2362 test_index
= isl_multi_pw_aff_identity(space
);
2363 test_index
= isl_multi_pw_aff_set_tuple_id(test_index
, isl_dim_out
,
2364 isl_id_copy(id_test
));
2365 scop_body
= pet_scop_filter(scop_body
, test_index
, 1);
2367 scop
= pet_scop_add_seq(ctx
, scop_cond
, scop_body
);
2368 scop
= add_implication(scop
, id_test
, domain
, sign
, 1);
2373 /* Check if the while loop is of the form
2375 * while (affine expression)
2378 * If so, call extract_affine_while to construct a scop.
2380 * Otherwise, construct a generic while scop, with iteration domain
2381 * { [t] : t >= 0 }. The scop consists of two parts, one for
2382 * evaluating the condition and one for the body.
2383 * The schedule is adjusted to reflect that the condition is evaluated
2384 * before the body is executed and the body is filtered to depend
2385 * on the result of the condition evaluating to true on all iterations
2386 * up to the current iteration, while the evaluation the condition itself
2387 * is filtered to depend on the result of the condition evaluating to true
2388 * on all previous iterations.
2389 * The context of the scop representing the body is dropped
2390 * because we don't know how many times the body will be executed,
2393 * If the body contains any break, then it is taken into
2394 * account in infinite_domain (if the skip condition is affine)
2395 * or in scop_add_break (if the skip condition is not affine).
2397 struct pet_scop
*PetScan::extract(WhileStmt
*stmt
)
2400 isl_id
*id
, *id_test
, *id_break_test
;
2401 isl_multi_pw_aff
*test_index
;
2405 struct pet_scop
*scop
, *scop_body
;
2408 cond
= stmt
->getCond();
2414 clear_assignments
clear(assigned_value
);
2415 clear
.TraverseStmt(stmt
->getBody());
2417 pa
= try_extract_affine_condition(cond
);
2419 return extract_affine_while(pa
, stmt
->getBody());
2421 if (!allow_nested
) {
2426 test_index
= create_test_index(ctx
, n_test
++);
2427 scop
= extract_non_affine_condition(cond
,
2428 isl_multi_pw_aff_copy(test_index
));
2429 scop
= scop_add_array(scop
, test_index
, ast_context
);
2430 id_test
= isl_multi_pw_aff_get_tuple_id(test_index
, isl_dim_out
);
2431 isl_multi_pw_aff_free(test_index
);
2432 scop_body
= extract(stmt
->getBody());
2434 id
= isl_id_alloc(ctx
, "t", NULL
);
2435 domain
= infinite_domain(isl_id_copy(id
), scop_body
);
2436 ident
= identity_aff(domain
);
2438 has_var_break
= pet_scop_has_var_skip(scop_body
, pet_skip_later
);
2440 id_break_test
= pet_scop_get_skip_id(scop_body
, pet_skip_later
);
2442 scop
= pet_scop_prefix(scop
, 0);
2443 scop
= pet_scop_embed(scop
, isl_set_copy(domain
),
2444 isl_map_from_aff(isl_aff_copy(ident
)),
2445 isl_aff_copy(ident
), isl_id_copy(id
));
2446 scop_body
= pet_scop_reset_context(scop_body
);
2447 scop_body
= pet_scop_prefix(scop_body
, 1);
2448 scop_body
= pet_scop_embed(scop_body
, isl_set_copy(domain
),
2449 isl_map_from_aff(isl_aff_copy(ident
)), ident
, id
);
2451 if (has_var_break
) {
2452 scop
= scop_add_break(scop
, isl_id_copy(id_break_test
),
2453 isl_set_copy(domain
), isl_val_one(ctx
));
2454 scop_body
= scop_add_break(scop_body
, id_break_test
,
2455 isl_set_copy(domain
), isl_val_one(ctx
));
2457 scop
= scop_add_while(scop
, scop_body
, id_test
, domain
,
2463 /* Check whether "cond" expresses a simple loop bound
2464 * on the only set dimension.
2465 * In particular, if "up" is set then "cond" should contain only
2466 * upper bounds on the set dimension.
2467 * Otherwise, it should contain only lower bounds.
2469 static bool is_simple_bound(__isl_keep isl_set
*cond
, __isl_keep isl_val
*inc
)
2471 if (isl_val_is_pos(inc
))
2472 return !isl_set_dim_has_any_lower_bound(cond
, isl_dim_set
, 0);
2474 return !isl_set_dim_has_any_upper_bound(cond
, isl_dim_set
, 0);
2477 /* Extend a condition on a given iteration of a loop to one that
2478 * imposes the same condition on all previous iterations.
2479 * "domain" expresses the lower [upper] bound on the iterations
2480 * when inc is positive [negative].
2482 * In particular, we construct the condition (when inc is positive)
2484 * forall i' : (domain(i') and i' <= i) => cond(i')
2486 * which is equivalent to
2488 * not exists i' : domain(i') and i' <= i and not cond(i')
2490 * We construct this set by negating cond, applying a map
2492 * { [i'] -> [i] : domain(i') and i' <= i }
2494 * and then negating the result again.
2496 static __isl_give isl_set
*valid_for_each_iteration(__isl_take isl_set
*cond
,
2497 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
2499 isl_map
*previous_to_this
;
2501 if (isl_val_is_pos(inc
))
2502 previous_to_this
= isl_map_lex_le(isl_set_get_space(domain
));
2504 previous_to_this
= isl_map_lex_ge(isl_set_get_space(domain
));
2506 previous_to_this
= isl_map_intersect_domain(previous_to_this
, domain
);
2508 cond
= isl_set_complement(cond
);
2509 cond
= isl_set_apply(cond
, previous_to_this
);
2510 cond
= isl_set_complement(cond
);
2517 /* Construct a domain of the form
2519 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
2521 static __isl_give isl_set
*strided_domain(__isl_take isl_id
*id
,
2522 __isl_take isl_pw_aff
*init
, __isl_take isl_val
*inc
)
2528 init
= isl_pw_aff_insert_dims(init
, isl_dim_in
, 0, 1);
2529 dim
= isl_pw_aff_get_domain_space(init
);
2530 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
2531 aff
= isl_aff_add_coefficient_val(aff
, isl_dim_in
, 0, inc
);
2532 init
= isl_pw_aff_add(init
, isl_pw_aff_from_aff(aff
));
2534 dim
= isl_space_set_alloc(isl_pw_aff_get_ctx(init
), 1, 1);
2535 dim
= isl_space_set_dim_id(dim
, isl_dim_param
, 0, id
);
2536 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
2537 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_param
, 0, 1);
2539 set
= isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff
), init
);
2541 set
= isl_set_lower_bound_si(set
, isl_dim_set
, 0, 0);
2543 return isl_set_params(set
);
2546 /* Assuming "cond" represents a bound on a loop where the loop
2547 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2550 * Under the given assumptions, wrapping is only possible if "cond" allows
2551 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2552 * increasing iterator and 0 in case of a decreasing iterator.
2554 static bool can_wrap(__isl_keep isl_set
*cond
, ValueDecl
*iv
,
2555 __isl_keep isl_val
*inc
)
2562 test
= isl_set_copy(cond
);
2564 ctx
= isl_set_get_ctx(test
);
2565 if (isl_val_is_neg(inc
))
2566 limit
= isl_val_zero(ctx
);
2568 limit
= isl_val_int_from_ui(ctx
, get_type_size(iv
));
2569 limit
= isl_val_2exp(limit
);
2570 limit
= isl_val_sub_ui(limit
, 1);
2573 test
= isl_set_fix_val(cond
, isl_dim_set
, 0, limit
);
2574 cw
= !isl_set_is_empty(test
);
2580 /* Given a one-dimensional space, construct the following affine expression
2583 * { [v] -> [v mod 2^width] }
2585 * where width is the number of bits used to represent the values
2586 * of the unsigned variable "iv".
2588 static __isl_give isl_aff
*compute_wrapping(__isl_take isl_space
*dim
,
2595 ctx
= isl_space_get_ctx(dim
);
2596 mod
= isl_val_int_from_ui(ctx
, get_type_size(iv
));
2597 mod
= isl_val_2exp(mod
);
2599 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
2600 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, 0, 1);
2601 aff
= isl_aff_mod_val(aff
, mod
);
2606 /* Project out the parameter "id" from "set".
2608 static __isl_give isl_set
*set_project_out_by_id(__isl_take isl_set
*set
,
2609 __isl_keep isl_id
*id
)
2613 pos
= isl_set_find_dim_by_id(set
, isl_dim_param
, id
);
2615 set
= isl_set_project_out(set
, isl_dim_param
, pos
, 1);
2620 /* Compute the set of parameters for which "set1" is a subset of "set2".
2622 * set1 is a subset of set2 if
2624 * forall i in set1 : i in set2
2628 * not exists i in set1 and i not in set2
2632 * not exists i in set1 \ set2
2634 static __isl_give isl_set
*enforce_subset(__isl_take isl_set
*set1
,
2635 __isl_take isl_set
*set2
)
2637 return isl_set_complement(isl_set_params(isl_set_subtract(set1
, set2
)));
2640 /* Compute the set of parameter values for which "cond" holds
2641 * on the next iteration for each element of "dom".
2643 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2644 * and then compute the set of parameters for which the result is a subset
2647 static __isl_give isl_set
*valid_on_next(__isl_take isl_set
*cond
,
2648 __isl_take isl_set
*dom
, __isl_take isl_val
*inc
)
2654 space
= isl_set_get_space(dom
);
2655 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
2656 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, 0, 1);
2657 aff
= isl_aff_add_constant_val(aff
, inc
);
2658 next
= isl_map_from_basic_map(isl_basic_map_from_aff(aff
));
2660 dom
= isl_set_apply(dom
, next
);
2662 return enforce_subset(dom
, cond
);
2665 /* Does "id" refer to a nested access?
2667 static bool is_nested_parameter(__isl_keep isl_id
*id
)
2669 return id
&& isl_id_get_user(id
) && !isl_id_get_name(id
);
2672 /* Does parameter "pos" of "space" refer to a nested access?
2674 static bool is_nested_parameter(__isl_keep isl_space
*space
, int pos
)
2679 id
= isl_space_get_dim_id(space
, isl_dim_param
, pos
);
2680 nested
= is_nested_parameter(id
);
2686 /* Does "space" involve any parameters that refer to nested
2687 * accesses, i.e., parameters with no name?
2689 static bool has_nested(__isl_keep isl_space
*space
)
2693 nparam
= isl_space_dim(space
, isl_dim_param
);
2694 for (int i
= 0; i
< nparam
; ++i
)
2695 if (is_nested_parameter(space
, i
))
2701 /* Does "pa" involve any parameters that refer to nested
2702 * accesses, i.e., parameters with no name?
2704 static bool has_nested(__isl_keep isl_pw_aff
*pa
)
2709 space
= isl_pw_aff_get_space(pa
);
2710 nested
= has_nested(space
);
2711 isl_space_free(space
);
2716 /* Construct a pet_scop for a for statement.
2717 * The for loop is required to be of the form
2719 * for (i = init; condition; ++i)
2723 * for (i = init; condition; --i)
2725 * The initialization of the for loop should either be an assignment
2726 * to an integer variable, or a declaration of such a variable with
2729 * The condition is allowed to contain nested accesses, provided
2730 * they are not being written to inside the body of the loop.
2731 * Otherwise, or if the condition is otherwise non-affine, the for loop is
2732 * essentially treated as a while loop, with iteration domain
2733 * { [i] : i >= init }.
2735 * We extract a pet_scop for the body and then embed it in a loop with
2736 * iteration domain and schedule
2738 * { [i] : i >= init and condition' }
2743 * { [i] : i <= init and condition' }
2746 * Where condition' is equal to condition if the latter is
2747 * a simple upper [lower] bound and a condition that is extended
2748 * to apply to all previous iterations otherwise.
2750 * If the condition is non-affine, then we drop the condition from the
2751 * iteration domain and instead create a separate statement
2752 * for evaluating the condition. The body is then filtered to depend
2753 * on the result of the condition evaluating to true on all iterations
2754 * up to the current iteration, while the evaluation the condition itself
2755 * is filtered to depend on the result of the condition evaluating to true
2756 * on all previous iterations.
2757 * The context of the scop representing the body is dropped
2758 * because we don't know how many times the body will be executed,
2761 * If the stride of the loop is not 1, then "i >= init" is replaced by
2763 * (exists a: i = init + stride * a and a >= 0)
2765 * If the loop iterator i is unsigned, then wrapping may occur.
2766 * During the computation, we work with a virtual iterator that
2767 * does not wrap. However, the condition in the code applies
2768 * to the wrapped value, so we need to change condition(i)
2769 * into condition([i % 2^width]).
2770 * After computing the virtual domain and schedule, we apply
2771 * the function { [v] -> [v % 2^width] } to the domain and the domain
2772 * of the schedule. In order not to lose any information, we also
2773 * need to intersect the domain of the schedule with the virtual domain
2774 * first, since some iterations in the wrapped domain may be scheduled
2775 * several times, typically an infinite number of times.
2776 * Note that there may be no need to perform this final wrapping
2777 * if the loop condition (after wrapping) satisfies certain conditions.
2778 * However, the is_simple_bound condition is not enough since it doesn't
2779 * check if there even is an upper bound.
2781 * If the loop condition is non-affine, then we keep the virtual
2782 * iterator in the iteration domain and instead replace all accesses
2783 * to the original iterator by the wrapping of the virtual iterator.
2785 * Wrapping on unsigned iterators can be avoided entirely if
2786 * loop condition is simple, the loop iterator is incremented
2787 * [decremented] by one and the last value before wrapping cannot
2788 * possibly satisfy the loop condition.
2790 * Before extracting a pet_scop from the body we remove all
2791 * assignments in assigned_value to variables that are assigned
2792 * somewhere in the body of the loop.
2794 * Valid parameters for a for loop are those for which the initial
2795 * value itself, the increment on each domain iteration and
2796 * the condition on both the initial value and
2797 * the result of incrementing the iterator for each iteration of the domain
2799 * If the loop condition is non-affine, then we only consider validity
2800 * of the initial value.
2802 * If the body contains any break, then we keep track of it in "skip"
2803 * (if the skip condition is affine) or it is handled in scop_add_break
2804 * (if the skip condition is not affine).
2805 * Note that the affine break condition needs to be considered with
2806 * respect to previous iterations in the virtual domain (if any)
2807 * and that the domain needs to be kept virtual if there is a non-affine
2810 struct pet_scop
*PetScan::extract_for(ForStmt
*stmt
)
2812 BinaryOperator
*ass
;
2820 isl_set
*cond
= NULL
;
2821 isl_set
*skip
= NULL
;
2822 isl_id
*id
, *id_test
= NULL
, *id_break_test
;
2823 struct pet_scop
*scop
, *scop_cond
= NULL
;
2824 assigned_value_cache
cache(assigned_value
);
2830 bool keep_virtual
= false;
2831 bool has_affine_break
;
2833 isl_aff
*wrap
= NULL
;
2834 isl_pw_aff
*pa
, *pa_inc
, *init_val
;
2835 isl_set
*valid_init
;
2836 isl_set
*valid_cond
;
2837 isl_set
*valid_cond_init
;
2838 isl_set
*valid_cond_next
;
2842 if (!stmt
->getInit() && !stmt
->getCond() && !stmt
->getInc())
2843 return extract_infinite_for(stmt
);
2845 init
= stmt
->getInit();
2850 if ((ass
= initialization_assignment(init
)) != NULL
) {
2851 iv
= extract_induction_variable(ass
);
2854 lhs
= ass
->getLHS();
2855 rhs
= ass
->getRHS();
2856 } else if ((decl
= initialization_declaration(init
)) != NULL
) {
2857 VarDecl
*var
= extract_induction_variable(init
, decl
);
2861 rhs
= var
->getInit();
2862 lhs
= create_DeclRefExpr(var
);
2864 unsupported(stmt
->getInit());
2868 pa_inc
= extract_increment(stmt
, iv
);
2873 if (isl_pw_aff_n_piece(pa_inc
) != 1 ||
2874 isl_pw_aff_foreach_piece(pa_inc
, &extract_cst
, &inc
) < 0) {
2875 isl_pw_aff_free(pa_inc
);
2876 unsupported(stmt
->getInc());
2880 valid_inc
= isl_pw_aff_domain(pa_inc
);
2882 is_unsigned
= iv
->getType()->isUnsignedIntegerType();
2884 assigned_value
.erase(iv
);
2885 clear_assignments
clear(assigned_value
);
2886 clear
.TraverseStmt(stmt
->getBody());
2888 id
= isl_id_alloc(ctx
, iv
->getName().str().c_str(), iv
);
2890 pa
= try_extract_nested_condition(stmt
->getCond());
2891 if (allow_nested
&& (!pa
|| has_nested(pa
)))
2894 scop
= extract(stmt
->getBody());
2896 has_affine_break
= scop
&&
2897 pet_scop_has_affine_skip(scop
, pet_skip_later
);
2898 if (has_affine_break
)
2899 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_later
);
2900 has_var_break
= scop
&& pet_scop_has_var_skip(scop
, pet_skip_later
);
2901 if (has_var_break
) {
2902 id_break_test
= pet_scop_get_skip_id(scop
, pet_skip_later
);
2903 keep_virtual
= true;
2906 if (pa
&& !is_nested_allowed(pa
, scop
)) {
2907 isl_pw_aff_free(pa
);
2911 if (!allow_nested
&& !pa
)
2912 pa
= try_extract_affine_condition(stmt
->getCond());
2913 valid_cond
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
2914 cond
= isl_pw_aff_non_zero_set(pa
);
2915 if (allow_nested
&& !cond
) {
2916 isl_multi_pw_aff
*test_index
;
2917 int save_n_stmt
= n_stmt
;
2918 test_index
= create_test_index(ctx
, n_test
++);
2920 scop_cond
= extract_non_affine_condition(stmt
->getCond(),
2921 isl_multi_pw_aff_copy(test_index
));
2922 n_stmt
= save_n_stmt
;
2923 scop_cond
= scop_add_array(scop_cond
, test_index
, ast_context
);
2924 id_test
= isl_multi_pw_aff_get_tuple_id(test_index
,
2926 isl_multi_pw_aff_free(test_index
);
2927 scop_cond
= pet_scop_prefix(scop_cond
, 0);
2928 scop
= pet_scop_reset_context(scop
);
2929 scop
= pet_scop_prefix(scop
, 1);
2930 keep_virtual
= true;
2931 cond
= isl_set_universe(isl_space_set_alloc(ctx
, 0, 0));
2934 cond
= embed(cond
, isl_id_copy(id
));
2935 skip
= embed(skip
, isl_id_copy(id
));
2936 valid_cond
= isl_set_coalesce(valid_cond
);
2937 valid_cond
= embed(valid_cond
, isl_id_copy(id
));
2938 valid_inc
= embed(valid_inc
, isl_id_copy(id
));
2939 is_one
= isl_val_is_one(inc
) || isl_val_is_negone(inc
);
2940 is_virtual
= is_unsigned
&& (!is_one
|| can_wrap(cond
, iv
, inc
));
2942 init_val
= extract_affine(rhs
);
2943 valid_cond_init
= enforce_subset(
2944 isl_set_from_pw_aff(isl_pw_aff_copy(init_val
)),
2945 isl_set_copy(valid_cond
));
2946 if (is_one
&& !is_virtual
) {
2947 isl_pw_aff_free(init_val
);
2948 pa
= extract_comparison(isl_val_is_pos(inc
) ? BO_GE
: BO_LE
,
2950 valid_init
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
2951 valid_init
= set_project_out_by_id(valid_init
, id
);
2952 domain
= isl_pw_aff_non_zero_set(pa
);
2954 valid_init
= isl_pw_aff_domain(isl_pw_aff_copy(init_val
));
2955 domain
= strided_domain(isl_id_copy(id
), init_val
,
2959 domain
= embed(domain
, isl_id_copy(id
));
2962 wrap
= compute_wrapping(isl_set_get_space(cond
), iv
);
2963 rev_wrap
= isl_map_from_aff(isl_aff_copy(wrap
));
2964 rev_wrap
= isl_map_reverse(rev_wrap
);
2965 cond
= isl_set_apply(cond
, isl_map_copy(rev_wrap
));
2966 skip
= isl_set_apply(skip
, isl_map_copy(rev_wrap
));
2967 valid_cond
= isl_set_apply(valid_cond
, isl_map_copy(rev_wrap
));
2968 valid_inc
= isl_set_apply(valid_inc
, rev_wrap
);
2970 is_simple
= is_simple_bound(cond
, inc
);
2972 cond
= isl_set_gist(cond
, isl_set_copy(domain
));
2973 is_simple
= is_simple_bound(cond
, inc
);
2976 cond
= valid_for_each_iteration(cond
,
2977 isl_set_copy(domain
), isl_val_copy(inc
));
2978 domain
= isl_set_intersect(domain
, cond
);
2979 if (has_affine_break
) {
2980 skip
= isl_set_intersect(skip
, isl_set_copy(domain
));
2981 skip
= after(skip
, isl_val_sgn(inc
));
2982 domain
= isl_set_subtract(domain
, skip
);
2984 domain
= isl_set_set_dim_id(domain
, isl_dim_set
, 0, isl_id_copy(id
));
2985 space
= isl_space_from_domain(isl_set_get_space(domain
));
2986 space
= isl_space_add_dims(space
, isl_dim_out
, 1);
2987 sched
= isl_map_universe(space
);
2988 if (isl_val_is_pos(inc
))
2989 sched
= isl_map_equate(sched
, isl_dim_in
, 0, isl_dim_out
, 0);
2991 sched
= isl_map_oppose(sched
, isl_dim_in
, 0, isl_dim_out
, 0);
2993 valid_cond_next
= valid_on_next(valid_cond
, isl_set_copy(domain
),
2995 valid_inc
= enforce_subset(isl_set_copy(domain
), valid_inc
);
2997 if (is_virtual
&& !keep_virtual
) {
2998 isl_map
*wrap_map
= isl_map_from_aff(wrap
);
2999 wrap_map
= isl_map_set_dim_id(wrap_map
,
3000 isl_dim_out
, 0, isl_id_copy(id
));
3001 sched
= isl_map_intersect_domain(sched
, isl_set_copy(domain
));
3002 domain
= isl_set_apply(domain
, isl_map_copy(wrap_map
));
3003 sched
= isl_map_apply_domain(sched
, wrap_map
);
3005 if (!(is_virtual
&& keep_virtual
))
3006 wrap
= identity_aff(domain
);
3008 scop_cond
= pet_scop_embed(scop_cond
, isl_set_copy(domain
),
3009 isl_map_copy(sched
), isl_aff_copy(wrap
), isl_id_copy(id
));
3010 scop
= pet_scop_embed(scop
, isl_set_copy(domain
), sched
, wrap
, id
);
3011 scop
= resolve_nested(scop
);
3013 scop
= scop_add_break(scop
, id_break_test
, isl_set_copy(domain
),
3016 scop
= scop_add_while(scop_cond
, scop
, id_test
, domain
,
3018 isl_set_free(valid_inc
);
3020 scop
= pet_scop_restrict_context(scop
, valid_inc
);
3021 scop
= pet_scop_restrict_context(scop
, valid_cond_next
);
3022 scop
= pet_scop_restrict_context(scop
, valid_cond_init
);
3023 isl_set_free(domain
);
3025 clear_assignment(assigned_value
, iv
);
3029 scop
= pet_scop_restrict_context(scop
, valid_init
);
3034 struct pet_scop
*PetScan::extract(CompoundStmt
*stmt
, bool skip_declarations
)
3036 return extract(stmt
->children(), true, skip_declarations
);
3039 /* Does parameter "pos" of "map" refer to a nested access?
3041 static bool is_nested_parameter(__isl_keep isl_map
*map
, int pos
)
3046 id
= isl_map_get_dim_id(map
, isl_dim_param
, pos
);
3047 nested
= is_nested_parameter(id
);
3053 /* How many parameters of "space" refer to nested accesses, i.e., have no name?
3055 static int n_nested_parameter(__isl_keep isl_space
*space
)
3060 nparam
= isl_space_dim(space
, isl_dim_param
);
3061 for (int i
= 0; i
< nparam
; ++i
)
3062 if (is_nested_parameter(space
, i
))
3068 /* How many parameters of "map" refer to nested accesses, i.e., have no name?
3070 static int n_nested_parameter(__isl_keep isl_map
*map
)
3075 space
= isl_map_get_space(map
);
3076 n
= n_nested_parameter(space
);
3077 isl_space_free(space
);
3082 /* For each nested access parameter in "space",
3083 * construct a corresponding pet_expr, place it in args and
3084 * record its position in "param2pos".
3085 * "n_arg" is the number of elements that are already in args.
3086 * The position recorded in "param2pos" takes this number into account.
3087 * If the pet_expr corresponding to a parameter is identical to
3088 * the pet_expr corresponding to an earlier parameter, then these two
3089 * parameters are made to refer to the same element in args.
3091 * Return the final number of elements in args or -1 if an error has occurred.
3093 int PetScan::extract_nested(__isl_keep isl_space
*space
,
3094 int n_arg
, struct pet_expr
**args
, std::map
<int,int> ¶m2pos
)
3098 nparam
= isl_space_dim(space
, isl_dim_param
);
3099 for (int i
= 0; i
< nparam
; ++i
) {
3101 isl_id
*id
= isl_space_get_dim_id(space
, isl_dim_param
, i
);
3104 if (!is_nested_parameter(id
)) {
3109 nested
= (Expr
*) isl_id_get_user(id
);
3110 args
[n_arg
] = extract_expr(nested
);
3114 for (j
= 0; j
< n_arg
; ++j
)
3115 if (pet_expr_is_equal(args
[j
], args
[n_arg
]))
3119 pet_expr_free(args
[n_arg
]);
3123 param2pos
[i
] = n_arg
++;
3131 /* For each nested access parameter in the access relations in "expr",
3132 * construct a corresponding pet_expr, place it in expr->args and
3133 * record its position in "param2pos".
3134 * n is the number of nested access parameters.
3136 struct pet_expr
*PetScan::extract_nested(struct pet_expr
*expr
, int n
,
3137 std::map
<int,int> ¶m2pos
)
3141 expr
->args
= isl_calloc_array(ctx
, struct pet_expr
*, n
);
3146 space
= isl_map_get_space(expr
->acc
.access
);
3147 n
= extract_nested(space
, 0, expr
->args
, param2pos
);
3148 isl_space_free(space
);
3156 pet_expr_free(expr
);
3160 /* Look for parameters in any access relation in "expr" that
3161 * refer to nested accesses. In particular, these are
3162 * parameters with no name.
3164 * If there are any such parameters, then the domain of the index
3165 * expression and the access relation, which is still [] at this point,
3166 * is replaced by [[] -> [t_1,...,t_n]], with n the number of these parameters
3167 * (after identifying identical nested accesses).
3169 * This transformation is performed in several steps.
3170 * We first extract the arguments in extract_nested.
3171 * param2pos maps the original parameter position to the position
3173 * Then we move these parameters to input dimension.
3174 * t2pos maps the positions of these temporary input dimensions
3175 * to the positions of the corresponding arguments.
3176 * Finally, we express there temporary dimensions in term of the domain
3177 * [[] -> [t_1,...,t_n]] and precompose index expression and access
3178 * relations with this function.
3180 struct pet_expr
*PetScan::resolve_nested(struct pet_expr
*expr
)
3185 isl_local_space
*ls
;
3188 std::map
<int,int> param2pos
;
3189 std::map
<int,int> t2pos
;
3194 for (int i
= 0; i
< expr
->n_arg
; ++i
) {
3195 expr
->args
[i
] = resolve_nested(expr
->args
[i
]);
3196 if (!expr
->args
[i
]) {
3197 pet_expr_free(expr
);
3202 if (expr
->type
!= pet_expr_access
)
3205 n
= n_nested_parameter(expr
->acc
.access
);
3209 expr
= extract_nested(expr
, n
, param2pos
);
3213 expr
= pet_expr_access_align_params(expr
);
3216 nparam
= isl_map_dim(expr
->acc
.access
, isl_dim_param
);
3219 for (int i
= nparam
- 1; i
>= 0; --i
) {
3220 isl_id
*id
= isl_map_get_dim_id(expr
->acc
.access
,
3222 if (!is_nested_parameter(id
)) {
3227 expr
->acc
.access
= isl_map_move_dims(expr
->acc
.access
,
3228 isl_dim_in
, n
, isl_dim_param
, i
, 1);
3229 expr
->acc
.index
= isl_multi_pw_aff_move_dims(expr
->acc
.index
,
3230 isl_dim_in
, n
, isl_dim_param
, i
, 1);
3231 t2pos
[n
] = param2pos
[i
];
3237 space
= isl_multi_pw_aff_get_space(expr
->acc
.index
);
3238 space
= isl_space_set_from_params(isl_space_params(space
));
3239 space
= isl_space_add_dims(space
, isl_dim_set
, expr
->n_arg
);
3240 space
= isl_space_wrap(isl_space_from_range(space
));
3241 ls
= isl_local_space_from_space(isl_space_copy(space
));
3242 space
= isl_space_from_domain(space
);
3243 space
= isl_space_add_dims(space
, isl_dim_out
, n
);
3244 ma
= isl_multi_aff_zero(space
);
3246 for (int i
= 0; i
< n
; ++i
) {
3247 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
3248 isl_dim_set
, t2pos
[i
]);
3249 ma
= isl_multi_aff_set_aff(ma
, i
, aff
);
3251 isl_local_space_free(ls
);
3253 expr
->acc
.access
= isl_map_preimage_domain_multi_aff(expr
->acc
.access
,
3254 isl_multi_aff_copy(ma
));
3255 expr
->acc
.index
= isl_multi_pw_aff_pullback_multi_aff(expr
->acc
.index
,
3261 /* Return the file offset of the expansion location of "Loc".
3263 static unsigned getExpansionOffset(SourceManager
&SM
, SourceLocation Loc
)
3265 return SM
.getFileOffset(SM
.getExpansionLoc(Loc
));
3268 #ifdef HAVE_FINDLOCATIONAFTERTOKEN
3270 /* Return a SourceLocation for the location after the first semicolon
3271 * after "loc". If Lexer::findLocationAfterToken is available, we simply
3272 * call it and also skip trailing spaces and newline.
3274 static SourceLocation
location_after_semi(SourceLocation loc
, SourceManager
&SM
,
3275 const LangOptions
&LO
)
3277 return Lexer::findLocationAfterToken(loc
, tok::semi
, SM
, LO
, true);
3282 /* Return a SourceLocation for the location after the first semicolon
3283 * after "loc". If Lexer::findLocationAfterToken is not available,
3284 * we look in the underlying character data for the first semicolon.
3286 static SourceLocation
location_after_semi(SourceLocation loc
, SourceManager
&SM
,
3287 const LangOptions
&LO
)
3290 const char *s
= SM
.getCharacterData(loc
);
3292 semi
= strchr(s
, ';');
3294 return SourceLocation();
3295 return loc
.getFileLocWithOffset(semi
+ 1 - s
);
3300 /* If the token at "loc" is the first token on the line, then return
3301 * a location referring to the start of the line.
3302 * Otherwise, return "loc".
3304 * This function is used to extend a scop to the start of the line
3305 * if the first token of the scop is also the first token on the line.
3307 * We look for the first token on the line. If its location is equal to "loc",
3308 * then the latter is the location of the first token on the line.
3310 static SourceLocation
move_to_start_of_line_if_first_token(SourceLocation loc
,
3311 SourceManager
&SM
, const LangOptions
&LO
)
3313 std::pair
<FileID
, unsigned> file_offset_pair
;
3314 llvm::StringRef file
;
3317 SourceLocation token_loc
, line_loc
;
3320 loc
= SM
.getExpansionLoc(loc
);
3321 col
= SM
.getExpansionColumnNumber(loc
);
3322 line_loc
= loc
.getLocWithOffset(1 - col
);
3323 file_offset_pair
= SM
.getDecomposedLoc(line_loc
);
3324 file
= SM
.getBufferData(file_offset_pair
.first
, NULL
);
3325 pos
= file
.data() + file_offset_pair
.second
;
3327 Lexer
lexer(SM
.getLocForStartOfFile(file_offset_pair
.first
), LO
,
3328 file
.begin(), pos
, file
.end());
3329 lexer
.LexFromRawLexer(tok
);
3330 token_loc
= tok
.getLocation();
3332 if (token_loc
== loc
)
3338 /* Convert a top-level pet_expr to a pet_scop with one statement.
3339 * This mainly involves resolving nested expression parameters
3340 * and setting the name of the iteration space.
3341 * The name is given by "label" if it is non-NULL. Otherwise,
3342 * it is of the form S_<n_stmt>.
3343 * start and end of the pet_scop are derived from those of "stmt".
3345 struct pet_scop
*PetScan::extract(Stmt
*stmt
, struct pet_expr
*expr
,
3346 __isl_take isl_id
*label
)
3348 struct pet_stmt
*ps
;
3349 struct pet_scop
*scop
;
3350 SourceLocation loc
= stmt
->getLocStart();
3351 SourceManager
&SM
= PP
.getSourceManager();
3352 const LangOptions
&LO
= PP
.getLangOpts();
3353 int line
= PP
.getSourceManager().getExpansionLineNumber(loc
);
3354 unsigned start
, end
;
3356 expr
= resolve_nested(expr
);
3357 ps
= pet_stmt_from_pet_expr(ctx
, line
, label
, n_stmt
++, expr
);
3358 scop
= pet_scop_from_pet_stmt(ctx
, ps
);
3360 loc
= move_to_start_of_line_if_first_token(loc
, SM
, LO
);
3361 start
= getExpansionOffset(SM
, loc
);
3362 loc
= stmt
->getLocEnd();
3363 loc
= location_after_semi(loc
, SM
, LO
);
3364 end
= getExpansionOffset(SM
, loc
);
3366 scop
= pet_scop_update_start_end(scop
, start
, end
);
3370 /* Check if we can extract an affine expression from "expr".
3371 * Return the expressions as an isl_pw_aff if we can and NULL otherwise.
3372 * We turn on autodetection so that we won't generate any warnings
3373 * and turn off nesting, so that we won't accept any non-affine constructs.
3375 __isl_give isl_pw_aff
*PetScan::try_extract_affine(Expr
*expr
)
3378 int save_autodetect
= options
->autodetect
;
3379 bool save_nesting
= nesting_enabled
;
3381 options
->autodetect
= 1;
3382 nesting_enabled
= false;
3384 pwaff
= extract_affine(expr
);
3386 options
->autodetect
= save_autodetect
;
3387 nesting_enabled
= save_nesting
;
3392 /* Check whether "expr" is an affine expression.
3394 bool PetScan::is_affine(Expr
*expr
)
3398 pwaff
= try_extract_affine(expr
);
3399 isl_pw_aff_free(pwaff
);
3401 return pwaff
!= NULL
;
3404 /* Check if we can extract an affine constraint from "expr".
3405 * Return the constraint as an isl_set if we can and NULL otherwise.
3406 * We turn on autodetection so that we won't generate any warnings
3407 * and turn off nesting, so that we won't accept any non-affine constructs.
3409 __isl_give isl_pw_aff
*PetScan::try_extract_affine_condition(Expr
*expr
)
3412 int save_autodetect
= options
->autodetect
;
3413 bool save_nesting
= nesting_enabled
;
3415 options
->autodetect
= 1;
3416 nesting_enabled
= false;
3418 cond
= extract_condition(expr
);
3420 options
->autodetect
= save_autodetect
;
3421 nesting_enabled
= save_nesting
;
3426 /* Check whether "expr" is an affine constraint.
3428 bool PetScan::is_affine_condition(Expr
*expr
)
3432 cond
= try_extract_affine_condition(expr
);
3433 isl_pw_aff_free(cond
);
3435 return cond
!= NULL
;
3438 /* Check if we can extract a condition from "expr".
3439 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
3440 * If allow_nested is set, then the condition may involve parameters
3441 * corresponding to nested accesses.
3442 * We turn on autodetection so that we won't generate any warnings.
3444 __isl_give isl_pw_aff
*PetScan::try_extract_nested_condition(Expr
*expr
)
3447 int save_autodetect
= options
->autodetect
;
3448 bool save_nesting
= nesting_enabled
;
3450 options
->autodetect
= 1;
3451 nesting_enabled
= allow_nested
;
3452 cond
= extract_condition(expr
);
3454 options
->autodetect
= save_autodetect
;
3455 nesting_enabled
= save_nesting
;
3460 /* If the top-level expression of "stmt" is an assignment, then
3461 * return that assignment as a BinaryOperator.
3462 * Otherwise return NULL.
3464 static BinaryOperator
*top_assignment_or_null(Stmt
*stmt
)
3466 BinaryOperator
*ass
;
3470 if (stmt
->getStmtClass() != Stmt::BinaryOperatorClass
)
3473 ass
= cast
<BinaryOperator
>(stmt
);
3474 if(ass
->getOpcode() != BO_Assign
)
3480 /* Check if the given if statement is a conditional assignement
3481 * with a non-affine condition. If so, construct a pet_scop
3482 * corresponding to this conditional assignment. Otherwise return NULL.
3484 * In particular we check if "stmt" is of the form
3491 * where a is some array or scalar access.
3492 * The constructed pet_scop then corresponds to the expression
3494 * a = condition ? f(...) : g(...)
3496 * All access relations in f(...) are intersected with condition
3497 * while all access relation in g(...) are intersected with the complement.
3499 struct pet_scop
*PetScan::extract_conditional_assignment(IfStmt
*stmt
)
3501 BinaryOperator
*ass_then
, *ass_else
;
3502 isl_multi_pw_aff
*write_then
, *write_else
;
3503 isl_set
*cond
, *comp
;
3504 isl_multi_pw_aff
*index
;
3507 struct pet_expr
*pe_cond
, *pe_then
, *pe_else
, *pe
, *pe_write
;
3508 bool save_nesting
= nesting_enabled
;
3510 if (!options
->detect_conditional_assignment
)
3513 ass_then
= top_assignment_or_null(stmt
->getThen());
3514 ass_else
= top_assignment_or_null(stmt
->getElse());
3516 if (!ass_then
|| !ass_else
)
3519 if (is_affine_condition(stmt
->getCond()))
3522 write_then
= extract_index(ass_then
->getLHS());
3523 write_else
= extract_index(ass_else
->getLHS());
3525 equal
= isl_multi_pw_aff_plain_is_equal(write_then
, write_else
);
3526 isl_multi_pw_aff_free(write_else
);
3527 if (equal
< 0 || !equal
) {
3528 isl_multi_pw_aff_free(write_then
);
3532 nesting_enabled
= allow_nested
;
3533 pa
= extract_condition(stmt
->getCond());
3534 nesting_enabled
= save_nesting
;
3535 cond
= isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa
));
3536 comp
= isl_pw_aff_zero_set(isl_pw_aff_copy(pa
));
3537 index
= isl_multi_pw_aff_from_range(isl_multi_pw_aff_from_pw_aff(pa
));
3539 pe_cond
= pet_expr_from_index(index
);
3541 pe_then
= extract_expr(ass_then
->getRHS());
3542 pe_then
= pet_expr_restrict(pe_then
, cond
);
3543 pe_else
= extract_expr(ass_else
->getRHS());
3544 pe_else
= pet_expr_restrict(pe_else
, comp
);
3546 pe
= pet_expr_new_ternary(ctx
, pe_cond
, pe_then
, pe_else
);
3547 pe_write
= pet_expr_from_index_and_depth(write_then
,
3548 extract_depth(write_then
));
3550 pe_write
->acc
.write
= 1;
3551 pe_write
->acc
.read
= 0;
3553 pe
= pet_expr_new_binary(ctx
, pet_op_assign
, pe_write
, pe
);
3554 return extract(stmt
, pe
);
3557 /* Create a pet_scop with a single statement evaluating "cond"
3558 * and writing the result to a virtual scalar, as expressed by
3561 struct pet_scop
*PetScan::extract_non_affine_condition(Expr
*cond
,
3562 __isl_take isl_multi_pw_aff
*index
)
3564 struct pet_expr
*expr
, *write
;
3565 struct pet_stmt
*ps
;
3566 struct pet_scop
*scop
;
3567 SourceLocation loc
= cond
->getLocStart();
3568 int line
= PP
.getSourceManager().getExpansionLineNumber(loc
);
3570 write
= pet_expr_from_index(index
);
3572 write
->acc
.write
= 1;
3573 write
->acc
.read
= 0;
3575 expr
= extract_expr(cond
);
3576 expr
= resolve_nested(expr
);
3577 expr
= pet_expr_new_binary(ctx
, pet_op_assign
, write
, expr
);
3578 ps
= pet_stmt_from_pet_expr(ctx
, line
, NULL
, n_stmt
++, expr
);
3579 scop
= pet_scop_from_pet_stmt(ctx
, ps
);
3580 scop
= resolve_nested(scop
);
3586 static struct pet_expr
*embed_access(struct pet_expr
*expr
, void *user
);
3589 /* Precompose the access relation and the index expression associated
3590 * to "expr" with the function pointed to by "user",
3591 * thereby embedding the access relation in the domain of this function.
3592 * The initial domain of the access relation and the index expression
3593 * is the zero-dimensional domain.
3595 static struct pet_expr
*embed_access(struct pet_expr
*expr
, void *user
)
3597 isl_multi_aff
*ma
= (isl_multi_aff
*) user
;
3599 expr
->acc
.access
= isl_map_preimage_domain_multi_aff(expr
->acc
.access
,
3600 isl_multi_aff_copy(ma
));
3601 expr
->acc
.index
= isl_multi_pw_aff_pullback_multi_aff(expr
->acc
.index
,
3602 isl_multi_aff_copy(ma
));
3603 if (!expr
->acc
.access
|| !expr
->acc
.index
)
3608 pet_expr_free(expr
);
3612 /* Precompose all access relations in "expr" with "ma", thereby
3613 * embedding them in the domain of "ma".
3615 static struct pet_expr
*embed(struct pet_expr
*expr
,
3616 __isl_keep isl_multi_aff
*ma
)
3618 return pet_expr_map_access(expr
, &embed_access
, ma
);
3621 /* How many parameters of "set" refer to nested accesses, i.e., have no name?
3623 static int n_nested_parameter(__isl_keep isl_set
*set
)
3628 space
= isl_set_get_space(set
);
3629 n
= n_nested_parameter(space
);
3630 isl_space_free(space
);
3635 /* Remove all parameters from "map" that refer to nested accesses.
3637 static __isl_give isl_map
*remove_nested_parameters(__isl_take isl_map
*map
)
3642 space
= isl_map_get_space(map
);
3643 nparam
= isl_space_dim(space
, isl_dim_param
);
3644 for (int i
= nparam
- 1; i
>= 0; --i
)
3645 if (is_nested_parameter(space
, i
))
3646 map
= isl_map_project_out(map
, isl_dim_param
, i
, 1);
3647 isl_space_free(space
);
3652 /* Remove all parameters from "mpa" that refer to nested accesses.
3654 static __isl_give isl_multi_pw_aff
*remove_nested_parameters(
3655 __isl_take isl_multi_pw_aff
*mpa
)
3660 space
= isl_multi_pw_aff_get_space(mpa
);
3661 nparam
= isl_space_dim(space
, isl_dim_param
);
3662 for (int i
= nparam
- 1; i
>= 0; --i
) {
3663 if (!is_nested_parameter(space
, i
))
3665 mpa
= isl_multi_pw_aff_drop_dims(mpa
, isl_dim_param
, i
, 1);
3667 isl_space_free(space
);
3672 /* Remove all parameters from the index expression and access relation of "expr"
3673 * that refer to nested accesses.
3675 static struct pet_expr
*remove_nested_parameters(struct pet_expr
*expr
)
3677 expr
->acc
.access
= remove_nested_parameters(expr
->acc
.access
);
3678 expr
->acc
.index
= remove_nested_parameters(expr
->acc
.index
);
3679 if (!expr
->acc
.access
|| !expr
->acc
.index
)
3684 pet_expr_free(expr
);
3689 static struct pet_expr
*expr_remove_nested_parameters(
3690 struct pet_expr
*expr
, void *user
);
3693 static struct pet_expr
*expr_remove_nested_parameters(
3694 struct pet_expr
*expr
, void *user
)
3696 return remove_nested_parameters(expr
);
3699 /* Remove all nested access parameters from the schedule and all
3700 * accesses of "stmt".
3701 * There is no need to remove them from the domain as these parameters
3702 * have already been removed from the domain when this function is called.
3704 static struct pet_stmt
*remove_nested_parameters(struct pet_stmt
*stmt
)
3708 stmt
->schedule
= remove_nested_parameters(stmt
->schedule
);
3709 stmt
->body
= pet_expr_map_access(stmt
->body
,
3710 &expr_remove_nested_parameters
, NULL
);
3711 if (!stmt
->schedule
|| !stmt
->body
)
3713 for (int i
= 0; i
< stmt
->n_arg
; ++i
) {
3714 stmt
->args
[i
] = pet_expr_map_access(stmt
->args
[i
],
3715 &expr_remove_nested_parameters
, NULL
);
3722 pet_stmt_free(stmt
);
3726 /* For each nested access parameter in the domain of "stmt",
3727 * construct a corresponding pet_expr, place it before the original
3728 * elements in stmt->args and record its position in "param2pos".
3729 * n is the number of nested access parameters.
3731 struct pet_stmt
*PetScan::extract_nested(struct pet_stmt
*stmt
, int n
,
3732 std::map
<int,int> ¶m2pos
)
3737 struct pet_expr
**args
;
3739 n_arg
= stmt
->n_arg
;
3740 args
= isl_calloc_array(ctx
, struct pet_expr
*, n
+ n_arg
);
3744 space
= isl_set_get_space(stmt
->domain
);
3745 n_arg
= extract_nested(space
, 0, args
, param2pos
);
3746 isl_space_free(space
);
3751 for (i
= 0; i
< stmt
->n_arg
; ++i
)
3752 args
[n_arg
+ i
] = stmt
->args
[i
];
3755 stmt
->n_arg
+= n_arg
;
3760 for (i
= 0; i
< n
; ++i
)
3761 pet_expr_free(args
[i
]);
3764 pet_stmt_free(stmt
);
3768 /* Check whether any of the arguments i of "stmt" starting at position "n"
3769 * is equal to one of the first "n" arguments j.
3770 * If so, combine the constraints on arguments i and j and remove
3773 static struct pet_stmt
*remove_duplicate_arguments(struct pet_stmt
*stmt
, int n
)
3782 if (n
== stmt
->n_arg
)
3785 map
= isl_set_unwrap(stmt
->domain
);
3787 for (i
= stmt
->n_arg
- 1; i
>= n
; --i
) {
3788 for (j
= 0; j
< n
; ++j
)
3789 if (pet_expr_is_equal(stmt
->args
[i
], stmt
->args
[j
]))
3794 map
= isl_map_equate(map
, isl_dim_out
, i
, isl_dim_out
, j
);
3795 map
= isl_map_project_out(map
, isl_dim_out
, i
, 1);
3797 pet_expr_free(stmt
->args
[i
]);
3798 for (j
= i
; j
+ 1 < stmt
->n_arg
; ++j
)
3799 stmt
->args
[j
] = stmt
->args
[j
+ 1];
3803 stmt
->domain
= isl_map_wrap(map
);
3808 pet_stmt_free(stmt
);
3812 /* Look for parameters in the iteration domain of "stmt" that
3813 * refer to nested accesses. In particular, these are
3814 * parameters with no name.
3816 * If there are any such parameters, then as many extra variables
3817 * (after identifying identical nested accesses) are inserted in the
3818 * range of the map wrapped inside the domain, before the original variables.
3819 * If the original domain is not a wrapped map, then a new wrapped
3820 * map is created with zero output dimensions.
3821 * The parameters are then equated to the corresponding output dimensions
3822 * and subsequently projected out, from the iteration domain,
3823 * the schedule and the access relations.
3824 * For each of the output dimensions, a corresponding argument
3825 * expression is inserted. Initially they are created with
3826 * a zero-dimensional domain, so they have to be embedded
3827 * in the current iteration domain.
3828 * param2pos maps the position of the parameter to the position
3829 * of the corresponding output dimension in the wrapped map.
3831 struct pet_stmt
*PetScan::resolve_nested(struct pet_stmt
*stmt
)
3839 std::map
<int,int> param2pos
;
3844 n
= n_nested_parameter(stmt
->domain
);
3848 n_arg
= stmt
->n_arg
;
3849 stmt
= extract_nested(stmt
, n
, param2pos
);
3853 n
= stmt
->n_arg
- n_arg
;
3854 nparam
= isl_set_dim(stmt
->domain
, isl_dim_param
);
3855 if (isl_set_is_wrapping(stmt
->domain
))
3856 map
= isl_set_unwrap(stmt
->domain
);
3858 map
= isl_map_from_domain(stmt
->domain
);
3859 map
= isl_map_insert_dims(map
, isl_dim_out
, 0, n
);
3861 for (int i
= nparam
- 1; i
>= 0; --i
) {
3864 if (!is_nested_parameter(map
, i
))
3867 id
= pet_expr_access_get_id(stmt
->args
[param2pos
[i
]]);
3868 map
= isl_map_set_dim_id(map
, isl_dim_out
, param2pos
[i
], id
);
3869 map
= isl_map_equate(map
, isl_dim_param
, i
, isl_dim_out
,
3871 map
= isl_map_project_out(map
, isl_dim_param
, i
, 1);
3874 stmt
->domain
= isl_map_wrap(map
);
3876 space
= isl_space_unwrap(isl_set_get_space(stmt
->domain
));
3877 space
= isl_space_from_domain(isl_space_domain(space
));
3878 ma
= isl_multi_aff_zero(space
);
3879 for (int pos
= 0; pos
< n
; ++pos
)
3880 stmt
->args
[pos
] = embed(stmt
->args
[pos
], ma
);
3881 isl_multi_aff_free(ma
);
3883 stmt
= remove_nested_parameters(stmt
);
3884 stmt
= remove_duplicate_arguments(stmt
, n
);
3888 pet_stmt_free(stmt
);
3892 /* For each statement in "scop", move the parameters that correspond
3893 * to nested access into the ranges of the domains and create
3894 * corresponding argument expressions.
3896 struct pet_scop
*PetScan::resolve_nested(struct pet_scop
*scop
)
3901 for (int i
= 0; i
< scop
->n_stmt
; ++i
) {
3902 scop
->stmts
[i
] = resolve_nested(scop
->stmts
[i
]);
3903 if (!scop
->stmts
[i
])
3909 pet_scop_free(scop
);
3913 /* Given an access expression "expr", is the variable accessed by
3914 * "expr" assigned anywhere inside "scop"?
3916 static bool is_assigned(pet_expr
*expr
, pet_scop
*scop
)
3918 bool assigned
= false;
3921 id
= pet_expr_access_get_id(expr
);
3922 assigned
= pet_scop_writes(scop
, id
);
3928 /* Are all nested access parameters in "pa" allowed given "scop".
3929 * In particular, is none of them written by anywhere inside "scop".
3931 * If "scop" has any skip conditions, then no nested access parameters
3932 * are allowed. In particular, if there is any nested access in a guard
3933 * for a piece of code containing a "continue", then we want to introduce
3934 * a separate statement for evaluating this guard so that we can express
3935 * that the result is false for all previous iterations.
3937 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff
*pa
, pet_scop
*scop
)
3944 nparam
= isl_pw_aff_dim(pa
, isl_dim_param
);
3945 for (int i
= 0; i
< nparam
; ++i
) {
3947 isl_id
*id
= isl_pw_aff_get_dim_id(pa
, isl_dim_param
, i
);
3951 if (!is_nested_parameter(id
)) {
3956 if (pet_scop_has_skip(scop
, pet_skip_now
)) {
3961 nested
= (Expr
*) isl_id_get_user(id
);
3962 expr
= extract_expr(nested
);
3963 allowed
= expr
&& expr
->type
== pet_expr_access
&&
3964 !is_assigned(expr
, scop
);
3966 pet_expr_free(expr
);
3976 /* Do we need to construct a skip condition of the given type
3977 * on an if statement, given that the if condition is non-affine?
3979 * pet_scop_filter_skip can only handle the case where the if condition
3980 * holds (the then branch) and the skip condition is universal.
3981 * In any other case, we need to construct a new skip condition.
3983 static bool need_skip(struct pet_scop
*scop_then
, struct pet_scop
*scop_else
,
3984 bool have_else
, enum pet_skip type
)
3986 if (have_else
&& scop_else
&& pet_scop_has_skip(scop_else
, type
))
3988 if (scop_then
&& pet_scop_has_skip(scop_then
, type
) &&
3989 !pet_scop_has_universal_skip(scop_then
, type
))
3994 /* Do we need to construct a skip condition of the given type
3995 * on an if statement, given that the if condition is affine?
3997 * There is no need to construct a new skip condition if all
3998 * the skip conditions are affine.
4000 static bool need_skip_aff(struct pet_scop
*scop_then
,
4001 struct pet_scop
*scop_else
, bool have_else
, enum pet_skip type
)
4003 if (scop_then
&& pet_scop_has_var_skip(scop_then
, type
))
4005 if (have_else
&& scop_else
&& pet_scop_has_var_skip(scop_else
, type
))
4010 /* Do we need to construct a skip condition of the given type
4011 * on an if statement?
4013 static bool need_skip(struct pet_scop
*scop_then
, struct pet_scop
*scop_else
,
4014 bool have_else
, enum pet_skip type
, bool affine
)
4017 return need_skip_aff(scop_then
, scop_else
, have_else
, type
);
4019 return need_skip(scop_then
, scop_else
, have_else
, type
);
4022 /* Construct an affine expression pet_expr that evaluates
4023 * to the constant "val".
4025 static struct pet_expr
*universally(isl_ctx
*ctx
, int val
)
4027 isl_local_space
*ls
;
4029 isl_multi_pw_aff
*mpa
;
4031 ls
= isl_local_space_from_space(isl_space_set_alloc(ctx
, 0, 0));
4032 aff
= isl_aff_val_on_domain(ls
, isl_val_int_from_si(ctx
, val
));
4033 mpa
= isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff
));
4035 return pet_expr_from_index(mpa
);
4038 /* Construct an affine expression pet_expr that evaluates
4039 * to the constant 1.
4041 static struct pet_expr
*universally_true(isl_ctx
*ctx
)
4043 return universally(ctx
, 1);
4046 /* Construct an affine expression pet_expr that evaluates
4047 * to the constant 0.
4049 static struct pet_expr
*universally_false(isl_ctx
*ctx
)
4051 return universally(ctx
, 0);
4054 /* Given an index expression "test_index" for the if condition,
4055 * an index expression "skip_index" for the skip condition and
4056 * scops for the then and else branches, construct a scop for
4057 * computing "skip_index".
4059 * The computed scop contains a single statement that essentially does
4061 * skip_index = test_cond ? skip_cond_then : skip_cond_else
4063 * If the skip conditions of the then and/or else branch are not affine,
4064 * then they need to be filtered by test_index.
4065 * If they are missing, then this means the skip condition is false.
4067 * Since we are constructing a skip condition for the if statement,
4068 * the skip conditions on the then and else branches are removed.
4070 static struct pet_scop
*extract_skip(PetScan
*scan
,
4071 __isl_take isl_multi_pw_aff
*test_index
,
4072 __isl_take isl_multi_pw_aff
*skip_index
,
4073 struct pet_scop
*scop_then
, struct pet_scop
*scop_else
, bool have_else
,
4076 struct pet_expr
*expr_then
, *expr_else
, *expr
, *expr_skip
;
4077 struct pet_stmt
*stmt
;
4078 struct pet_scop
*scop
;
4079 isl_ctx
*ctx
= scan
->ctx
;
4083 if (have_else
&& !scop_else
)
4086 if (pet_scop_has_skip(scop_then
, type
)) {
4087 expr_then
= pet_scop_get_skip_expr(scop_then
, type
);
4088 pet_scop_reset_skip(scop_then
, type
);
4089 if (!pet_expr_is_affine(expr_then
))
4090 expr_then
= pet_expr_filter(expr_then
,
4091 isl_multi_pw_aff_copy(test_index
), 1);
4093 expr_then
= universally_false(ctx
);
4095 if (have_else
&& pet_scop_has_skip(scop_else
, type
)) {
4096 expr_else
= pet_scop_get_skip_expr(scop_else
, type
);
4097 pet_scop_reset_skip(scop_else
, type
);
4098 if (!pet_expr_is_affine(expr_else
))
4099 expr_else
= pet_expr_filter(expr_else
,
4100 isl_multi_pw_aff_copy(test_index
), 0);
4102 expr_else
= universally_false(ctx
);
4104 expr
= pet_expr_from_index(test_index
);
4105 expr
= pet_expr_new_ternary(ctx
, expr
, expr_then
, expr_else
);
4106 expr_skip
= pet_expr_from_index(isl_multi_pw_aff_copy(skip_index
));
4108 expr_skip
->acc
.write
= 1;
4109 expr_skip
->acc
.read
= 0;
4111 expr
= pet_expr_new_binary(ctx
, pet_op_assign
, expr_skip
, expr
);
4112 stmt
= pet_stmt_from_pet_expr(ctx
, -1, NULL
, scan
->n_stmt
++, expr
);
4114 scop
= pet_scop_from_pet_stmt(ctx
, stmt
);
4115 scop
= scop_add_array(scop
, skip_index
, scan
->ast_context
);
4116 isl_multi_pw_aff_free(skip_index
);
4120 isl_multi_pw_aff_free(test_index
);
4121 isl_multi_pw_aff_free(skip_index
);
4125 /* Is scop's skip_now condition equal to its skip_later condition?
4126 * In particular, this means that it either has no skip_now condition
4127 * or both a skip_now and a skip_later condition (that are equal to each other).
4129 static bool skip_equals_skip_later(struct pet_scop
*scop
)
4131 int has_skip_now
, has_skip_later
;
4133 isl_multi_pw_aff
*skip_now
, *skip_later
;
4137 has_skip_now
= pet_scop_has_skip(scop
, pet_skip_now
);
4138 has_skip_later
= pet_scop_has_skip(scop
, pet_skip_later
);
4139 if (has_skip_now
!= has_skip_later
)
4144 skip_now
= pet_scop_get_skip(scop
, pet_skip_now
);
4145 skip_later
= pet_scop_get_skip(scop
, pet_skip_later
);
4146 equal
= isl_multi_pw_aff_is_equal(skip_now
, skip_later
);
4147 isl_multi_pw_aff_free(skip_now
);
4148 isl_multi_pw_aff_free(skip_later
);
4153 /* Drop the skip conditions of type pet_skip_later from scop1 and scop2.
4155 static void drop_skip_later(struct pet_scop
*scop1
, struct pet_scop
*scop2
)
4157 pet_scop_reset_skip(scop1
, pet_skip_later
);
4158 pet_scop_reset_skip(scop2
, pet_skip_later
);
4161 /* Structure that handles the construction of skip conditions.
4163 * scop_then and scop_else represent the then and else branches
4164 * of the if statement
4166 * skip[type] is true if we need to construct a skip condition of that type
4167 * equal is set if the skip conditions of types pet_skip_now and pet_skip_later
4168 * are equal to each other
4169 * index[type] is an index expression from a zero-dimension domain
4170 * to the virtual array representing the skip condition
4171 * scop[type] is a scop for computing the skip condition
4173 struct pet_skip_info
{
4178 isl_multi_pw_aff
*index
[2];
4179 struct pet_scop
*scop
[2];
4181 pet_skip_info(isl_ctx
*ctx
) : ctx(ctx
) {}
4183 operator bool() { return skip
[pet_skip_now
] || skip
[pet_skip_later
]; }
4186 /* Structure that handles the construction of skip conditions on if statements.
4188 * scop_then and scop_else represent the then and else branches
4189 * of the if statement
4191 struct pet_skip_info_if
: public pet_skip_info
{
4192 struct pet_scop
*scop_then
, *scop_else
;
4195 pet_skip_info_if(isl_ctx
*ctx
, struct pet_scop
*scop_then
,
4196 struct pet_scop
*scop_else
, bool have_else
, bool affine
);
4197 void extract(PetScan
*scan
, __isl_keep isl_multi_pw_aff
*index
,
4198 enum pet_skip type
);
4199 void extract(PetScan
*scan
, __isl_keep isl_multi_pw_aff
*index
);
4200 void extract(PetScan
*scan
, __isl_keep isl_pw_aff
*cond
);
4201 struct pet_scop
*add(struct pet_scop
*scop
, enum pet_skip type
,
4203 struct pet_scop
*add(struct pet_scop
*scop
, int offset
);
4206 /* Initialize a pet_skip_info_if structure based on the then and else branches
4207 * and based on whether the if condition is affine or not.
4209 pet_skip_info_if::pet_skip_info_if(isl_ctx
*ctx
, struct pet_scop
*scop_then
,
4210 struct pet_scop
*scop_else
, bool have_else
, bool affine
) :
4211 pet_skip_info(ctx
), scop_then(scop_then
), scop_else(scop_else
),
4212 have_else(have_else
)
4214 skip
[pet_skip_now
] =
4215 need_skip(scop_then
, scop_else
, have_else
, pet_skip_now
, affine
);
4216 equal
= skip
[pet_skip_now
] && skip_equals_skip_later(scop_then
) &&
4217 (!have_else
|| skip_equals_skip_later(scop_else
));
4218 skip
[pet_skip_later
] = skip
[pet_skip_now
] && !equal
&&
4219 need_skip(scop_then
, scop_else
, have_else
, pet_skip_later
, affine
);
4222 /* If we need to construct a skip condition of the given type,
4225 * "mpa" represents the if condition.
4227 void pet_skip_info_if::extract(PetScan
*scan
,
4228 __isl_keep isl_multi_pw_aff
*mpa
, enum pet_skip type
)
4235 ctx
= isl_multi_pw_aff_get_ctx(mpa
);
4236 index
[type
] = create_test_index(ctx
, scan
->n_test
++);
4237 scop
[type
] = extract_skip(scan
, isl_multi_pw_aff_copy(mpa
),
4238 isl_multi_pw_aff_copy(index
[type
]),
4239 scop_then
, scop_else
, have_else
, type
);
4242 /* Construct the required skip conditions, given the if condition "index".
4244 void pet_skip_info_if::extract(PetScan
*scan
,
4245 __isl_keep isl_multi_pw_aff
*index
)
4247 extract(scan
, index
, pet_skip_now
);
4248 extract(scan
, index
, pet_skip_later
);
4250 drop_skip_later(scop_then
, scop_else
);
4253 /* Construct the required skip conditions, given the if condition "cond".
4255 void pet_skip_info_if::extract(PetScan
*scan
, __isl_keep isl_pw_aff
*cond
)
4257 isl_multi_pw_aff
*test
;
4259 if (!skip
[pet_skip_now
] && !skip
[pet_skip_later
])
4262 test
= isl_multi_pw_aff_from_pw_aff(isl_pw_aff_copy(cond
));
4263 test
= isl_multi_pw_aff_from_range(test
);
4264 extract(scan
, test
);
4265 isl_multi_pw_aff_free(test
);
4268 /* Add the computed skip condition of the give type to "main" and
4269 * add the scop for computing the condition at the given offset.
4271 * If equal is set, then we only computed a skip condition for pet_skip_now,
4272 * but we also need to set it as main's pet_skip_later.
4274 struct pet_scop
*pet_skip_info_if::add(struct pet_scop
*main
,
4275 enum pet_skip type
, int offset
)
4280 scop
[type
] = pet_scop_prefix(scop
[type
], offset
);
4281 main
= pet_scop_add_par(ctx
, main
, scop
[type
]);
4285 main
= pet_scop_set_skip(main
, pet_skip_later
,
4286 isl_multi_pw_aff_copy(index
[type
]));
4288 main
= pet_scop_set_skip(main
, type
, index
[type
]);
4294 /* Add the computed skip conditions to "main" and
4295 * add the scops for computing the conditions at the given offset.
4297 struct pet_scop
*pet_skip_info_if::add(struct pet_scop
*scop
, int offset
)
4299 scop
= add(scop
, pet_skip_now
, offset
);
4300 scop
= add(scop
, pet_skip_later
, offset
);
4305 /* Construct a pet_scop for a non-affine if statement.
4307 * We create a separate statement that writes the result
4308 * of the non-affine condition to a virtual scalar.
4309 * A constraint requiring the value of this virtual scalar to be one
4310 * is added to the iteration domains of the then branch.
4311 * Similarly, a constraint requiring the value of this virtual scalar
4312 * to be zero is added to the iteration domains of the else branch, if any.
4313 * We adjust the schedules to ensure that the virtual scalar is written
4314 * before it is read.
4316 * If there are any breaks or continues in the then and/or else
4317 * branches, then we may have to compute a new skip condition.
4318 * This is handled using a pet_skip_info_if object.
4319 * On initialization, the object checks if skip conditions need
4320 * to be computed. If so, it does so in "extract" and adds them in "add".
4322 struct pet_scop
*PetScan::extract_non_affine_if(Expr
*cond
,
4323 struct pet_scop
*scop_then
, struct pet_scop
*scop_else
,
4324 bool have_else
, int stmt_id
)
4326 struct pet_scop
*scop
;
4327 isl_multi_pw_aff
*test_index
;
4328 int save_n_stmt
= n_stmt
;
4330 test_index
= create_test_index(ctx
, n_test
++);
4332 scop
= extract_non_affine_condition(cond
,
4333 isl_multi_pw_aff_copy(test_index
));
4334 n_stmt
= save_n_stmt
;
4335 scop
= scop_add_array(scop
, test_index
, ast_context
);
4337 pet_skip_info_if
skip(ctx
, scop_then
, scop_else
, have_else
, false);
4338 skip
.extract(this, test_index
);
4340 scop
= pet_scop_prefix(scop
, 0);
4341 scop_then
= pet_scop_prefix(scop_then
, 1);
4342 scop_then
= pet_scop_filter(scop_then
,
4343 isl_multi_pw_aff_copy(test_index
), 1);
4345 scop_else
= pet_scop_prefix(scop_else
, 1);
4346 scop_else
= pet_scop_filter(scop_else
, test_index
, 0);
4347 scop_then
= pet_scop_add_par(ctx
, scop_then
, scop_else
);
4349 isl_multi_pw_aff_free(test_index
);
4351 scop
= pet_scop_add_seq(ctx
, scop
, scop_then
);
4353 scop
= skip
.add(scop
, 2);
4358 /* Construct a pet_scop for an if statement.
4360 * If the condition fits the pattern of a conditional assignment,
4361 * then it is handled by extract_conditional_assignment.
4362 * Otherwise, we do the following.
4364 * If the condition is affine, then the condition is added
4365 * to the iteration domains of the then branch, while the
4366 * opposite of the condition in added to the iteration domains
4367 * of the else branch, if any.
4368 * We allow the condition to be dynamic, i.e., to refer to
4369 * scalars or array elements that may be written to outside
4370 * of the given if statement. These nested accesses are then represented
4371 * as output dimensions in the wrapping iteration domain.
4372 * If it also written _inside_ the then or else branch, then
4373 * we treat the condition as non-affine.
4374 * As explained in extract_non_affine_if, this will introduce
4375 * an extra statement.
4376 * For aesthetic reasons, we want this statement to have a statement
4377 * number that is lower than those of the then and else branches.
4378 * In order to evaluate if will need such a statement, however, we
4379 * first construct scops for the then and else branches.
4380 * We therefore reserve a statement number if we might have to
4381 * introduce such an extra statement.
4383 * If the condition is not affine, then the scop is created in
4384 * extract_non_affine_if.
4386 * If there are any breaks or continues in the then and/or else
4387 * branches, then we may have to compute a new skip condition.
4388 * This is handled using a pet_skip_info_if object.
4389 * On initialization, the object checks if skip conditions need
4390 * to be computed. If so, it does so in "extract" and adds them in "add".
4392 struct pet_scop
*PetScan::extract(IfStmt
*stmt
)
4394 struct pet_scop
*scop_then
, *scop_else
= NULL
, *scop
;
4400 scop
= extract_conditional_assignment(stmt
);
4404 cond
= try_extract_nested_condition(stmt
->getCond());
4405 if (allow_nested
&& (!cond
|| has_nested(cond
)))
4409 assigned_value_cache
cache(assigned_value
);
4410 scop_then
= extract(stmt
->getThen());
4413 if (stmt
->getElse()) {
4414 assigned_value_cache
cache(assigned_value
);
4415 scop_else
= extract(stmt
->getElse());
4416 if (options
->autodetect
) {
4417 if (scop_then
&& !scop_else
) {
4419 isl_pw_aff_free(cond
);
4422 if (!scop_then
&& scop_else
) {
4424 isl_pw_aff_free(cond
);
4431 (!is_nested_allowed(cond
, scop_then
) ||
4432 (stmt
->getElse() && !is_nested_allowed(cond
, scop_else
)))) {
4433 isl_pw_aff_free(cond
);
4436 if (allow_nested
&& !cond
)
4437 return extract_non_affine_if(stmt
->getCond(), scop_then
,
4438 scop_else
, stmt
->getElse(), stmt_id
);
4441 cond
= extract_condition(stmt
->getCond());
4443 pet_skip_info_if
skip(ctx
, scop_then
, scop_else
, stmt
->getElse(), true);
4444 skip
.extract(this, cond
);
4446 valid
= isl_pw_aff_domain(isl_pw_aff_copy(cond
));
4447 set
= isl_pw_aff_non_zero_set(cond
);
4448 scop
= pet_scop_restrict(scop_then
, isl_set_copy(set
));
4450 if (stmt
->getElse()) {
4451 set
= isl_set_subtract(isl_set_copy(valid
), set
);
4452 scop_else
= pet_scop_restrict(scop_else
, set
);
4453 scop
= pet_scop_add_par(ctx
, scop
, scop_else
);
4456 scop
= resolve_nested(scop
);
4457 scop
= pet_scop_restrict_context(scop
, valid
);
4460 scop
= pet_scop_prefix(scop
, 0);
4461 scop
= skip
.add(scop
, 1);
4466 /* Try and construct a pet_scop for a label statement.
4467 * We currently only allow labels on expression statements.
4469 struct pet_scop
*PetScan::extract(LabelStmt
*stmt
)
4474 sub
= stmt
->getSubStmt();
4475 if (!isa
<Expr
>(sub
)) {
4480 label
= isl_id_alloc(ctx
, stmt
->getName(), NULL
);
4482 return extract(sub
, extract_expr(cast
<Expr
>(sub
)), label
);
4485 /* Return a one-dimensional multi piecewise affine expression that is equal
4486 * to the constant 1 and is defined over a zero-dimensional domain.
4488 static __isl_give isl_multi_pw_aff
*one_mpa(isl_ctx
*ctx
)
4491 isl_local_space
*ls
;
4494 space
= isl_space_set_alloc(ctx
, 0, 0);
4495 ls
= isl_local_space_from_space(space
);
4496 aff
= isl_aff_zero_on_domain(ls
);
4497 aff
= isl_aff_set_constant_si(aff
, 1);
4499 return isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff
));
4502 /* Construct a pet_scop for a continue statement.
4504 * We simply create an empty scop with a universal pet_skip_now
4505 * skip condition. This skip condition will then be taken into
4506 * account by the enclosing loop construct, possibly after
4507 * being incorporated into outer skip conditions.
4509 struct pet_scop
*PetScan::extract(ContinueStmt
*stmt
)
4513 scop
= pet_scop_empty(ctx
);
4517 scop
= pet_scop_set_skip(scop
, pet_skip_now
, one_mpa(ctx
));
4522 /* Construct a pet_scop for a break statement.
4524 * We simply create an empty scop with both a universal pet_skip_now
4525 * skip condition and a universal pet_skip_later skip condition.
4526 * These skip conditions will then be taken into
4527 * account by the enclosing loop construct, possibly after
4528 * being incorporated into outer skip conditions.
4530 struct pet_scop
*PetScan::extract(BreakStmt
*stmt
)
4533 isl_multi_pw_aff
*skip
;
4535 scop
= pet_scop_empty(ctx
);
4539 skip
= one_mpa(ctx
);
4540 scop
= pet_scop_set_skip(scop
, pet_skip_now
,
4541 isl_multi_pw_aff_copy(skip
));
4542 scop
= pet_scop_set_skip(scop
, pet_skip_later
, skip
);
4547 /* Try and construct a pet_scop corresponding to "stmt".
4549 * If "stmt" is a compound statement, then "skip_declarations"
4550 * indicates whether we should skip initial declarations in the
4551 * compound statement.
4553 * If the constructed pet_scop is not a (possibly) partial representation
4554 * of "stmt", we update start and end of the pet_scop to those of "stmt".
4555 * In particular, if skip_declarations, then we may have skipped declarations
4556 * inside "stmt" and so the pet_scop may not represent the entire "stmt".
4557 * Note that this function may be called with "stmt" referring to the entire
4558 * body of the function, including the outer braces. In such cases,
4559 * skip_declarations will be set and the braces will not be taken into
4560 * account in scop->start and scop->end.
4562 struct pet_scop
*PetScan::extract(Stmt
*stmt
, bool skip_declarations
)
4564 struct pet_scop
*scop
;
4565 unsigned start
, end
;
4567 SourceManager
&SM
= PP
.getSourceManager();
4568 const LangOptions
&LO
= PP
.getLangOpts();
4570 if (isa
<Expr
>(stmt
))
4571 return extract(stmt
, extract_expr(cast
<Expr
>(stmt
)));
4573 switch (stmt
->getStmtClass()) {
4574 case Stmt::WhileStmtClass
:
4575 scop
= extract(cast
<WhileStmt
>(stmt
));
4577 case Stmt::ForStmtClass
:
4578 scop
= extract_for(cast
<ForStmt
>(stmt
));
4580 case Stmt::IfStmtClass
:
4581 scop
= extract(cast
<IfStmt
>(stmt
));
4583 case Stmt::CompoundStmtClass
:
4584 scop
= extract(cast
<CompoundStmt
>(stmt
), skip_declarations
);
4586 case Stmt::LabelStmtClass
:
4587 scop
= extract(cast
<LabelStmt
>(stmt
));
4589 case Stmt::ContinueStmtClass
:
4590 scop
= extract(cast
<ContinueStmt
>(stmt
));
4592 case Stmt::BreakStmtClass
:
4593 scop
= extract(cast
<BreakStmt
>(stmt
));
4595 case Stmt::DeclStmtClass
:
4596 scop
= extract(cast
<DeclStmt
>(stmt
));
4603 if (partial
|| skip_declarations
)
4606 loc
= stmt
->getLocStart();
4607 loc
= move_to_start_of_line_if_first_token(loc
, SM
, LO
);
4608 start
= getExpansionOffset(SM
, loc
);
4609 loc
= PP
.getLocForEndOfToken(stmt
->getLocEnd());
4610 end
= getExpansionOffset(SM
, loc
);
4611 scop
= pet_scop_update_start_end(scop
, start
, end
);
4616 /* Do we need to construct a skip condition of the given type
4617 * on a sequence of statements?
4619 * There is no need to construct a new skip condition if only
4620 * only of the two statements has a skip condition or if both
4621 * of their skip conditions are affine.
4623 * In principle we also don't need a new continuation variable if
4624 * the continuation of scop2 is affine, but then we would need
4625 * to allow more complicated forms of continuations.
4627 static bool need_skip_seq(struct pet_scop
*scop1
, struct pet_scop
*scop2
,
4630 if (!scop1
|| !pet_scop_has_skip(scop1
, type
))
4632 if (!scop2
|| !pet_scop_has_skip(scop2
, type
))
4634 if (pet_scop_has_affine_skip(scop1
, type
) &&
4635 pet_scop_has_affine_skip(scop2
, type
))
4640 /* Construct a scop for computing the skip condition of the given type and
4641 * with index expression "skip_index" for a sequence of two scops "scop1"
4644 * The computed scop contains a single statement that essentially does
4646 * skip_index = skip_cond_1 ? 1 : skip_cond_2
4648 * or, in other words, skip_cond1 || skip_cond2.
4649 * In this expression, skip_cond_2 is filtered to reflect that it is
4650 * only evaluated when skip_cond_1 is false.
4652 * The skip condition on scop1 is not removed because it still needs
4653 * to be applied to scop2 when these two scops are combined.
4655 static struct pet_scop
*extract_skip_seq(PetScan
*ps
,
4656 __isl_take isl_multi_pw_aff
*skip_index
,
4657 struct pet_scop
*scop1
, struct pet_scop
*scop2
, enum pet_skip type
)
4659 struct pet_expr
*expr1
, *expr2
, *expr
, *expr_skip
;
4660 struct pet_stmt
*stmt
;
4661 struct pet_scop
*scop
;
4662 isl_ctx
*ctx
= ps
->ctx
;
4664 if (!scop1
|| !scop2
)
4667 expr1
= pet_scop_get_skip_expr(scop1
, type
);
4668 expr2
= pet_scop_get_skip_expr(scop2
, type
);
4669 pet_scop_reset_skip(scop2
, type
);
4671 expr2
= pet_expr_filter(expr2
,
4672 isl_multi_pw_aff_copy(expr1
->acc
.index
), 0);
4674 expr
= universally_true(ctx
);
4675 expr
= pet_expr_new_ternary(ctx
, expr1
, expr
, expr2
);
4676 expr_skip
= pet_expr_from_index(isl_multi_pw_aff_copy(skip_index
));
4678 expr_skip
->acc
.write
= 1;
4679 expr_skip
->acc
.read
= 0;
4681 expr
= pet_expr_new_binary(ctx
, pet_op_assign
, expr_skip
, expr
);
4682 stmt
= pet_stmt_from_pet_expr(ctx
, -1, NULL
, ps
->n_stmt
++, expr
);
4684 scop
= pet_scop_from_pet_stmt(ctx
, stmt
);
4685 scop
= scop_add_array(scop
, skip_index
, ps
->ast_context
);
4686 isl_multi_pw_aff_free(skip_index
);
4690 isl_multi_pw_aff_free(skip_index
);
4694 /* Structure that handles the construction of skip conditions
4695 * on sequences of statements.
4697 * scop1 and scop2 represent the two statements that are combined
4699 struct pet_skip_info_seq
: public pet_skip_info
{
4700 struct pet_scop
*scop1
, *scop2
;
4702 pet_skip_info_seq(isl_ctx
*ctx
, struct pet_scop
*scop1
,
4703 struct pet_scop
*scop2
);
4704 void extract(PetScan
*scan
, enum pet_skip type
);
4705 void extract(PetScan
*scan
);
4706 struct pet_scop
*add(struct pet_scop
*scop
, enum pet_skip type
,
4708 struct pet_scop
*add(struct pet_scop
*scop
, int offset
);
4711 /* Initialize a pet_skip_info_seq structure based on
4712 * on the two statements that are going to be combined.
4714 pet_skip_info_seq::pet_skip_info_seq(isl_ctx
*ctx
, struct pet_scop
*scop1
,
4715 struct pet_scop
*scop2
) : pet_skip_info(ctx
), scop1(scop1
), scop2(scop2
)
4717 skip
[pet_skip_now
] = need_skip_seq(scop1
, scop2
, pet_skip_now
);
4718 equal
= skip
[pet_skip_now
] && skip_equals_skip_later(scop1
) &&
4719 skip_equals_skip_later(scop2
);
4720 skip
[pet_skip_later
] = skip
[pet_skip_now
] && !equal
&&
4721 need_skip_seq(scop1
, scop2
, pet_skip_later
);
4724 /* If we need to construct a skip condition of the given type,
4727 void pet_skip_info_seq::extract(PetScan
*scan
, enum pet_skip type
)
4732 index
[type
] = create_test_index(ctx
, scan
->n_test
++);
4733 scop
[type
] = extract_skip_seq(scan
, isl_multi_pw_aff_copy(index
[type
]),
4734 scop1
, scop2
, type
);
4737 /* Construct the required skip conditions.
4739 void pet_skip_info_seq::extract(PetScan
*scan
)
4741 extract(scan
, pet_skip_now
);
4742 extract(scan
, pet_skip_later
);
4744 drop_skip_later(scop1
, scop2
);
4747 /* Add the computed skip condition of the given type to "main" and
4748 * add the scop for computing the condition at the given offset (the statement
4749 * number). Within this offset, the condition is computed at position 1
4750 * to ensure that it is computed after the corresponding statement.
4752 * If equal is set, then we only computed a skip condition for pet_skip_now,
4753 * but we also need to set it as main's pet_skip_later.
4755 struct pet_scop
*pet_skip_info_seq::add(struct pet_scop
*main
,
4756 enum pet_skip type
, int offset
)
4761 scop
[type
] = pet_scop_prefix(scop
[type
], 1);
4762 scop
[type
] = pet_scop_prefix(scop
[type
], offset
);
4763 main
= pet_scop_add_par(ctx
, main
, scop
[type
]);
4767 main
= pet_scop_set_skip(main
, pet_skip_later
,
4768 isl_multi_pw_aff_copy(index
[type
]));
4770 main
= pet_scop_set_skip(main
, type
, index
[type
]);
4776 /* Add the computed skip conditions to "main" and
4777 * add the scops for computing the conditions at the given offset.
4779 struct pet_scop
*pet_skip_info_seq::add(struct pet_scop
*scop
, int offset
)
4781 scop
= add(scop
, pet_skip_now
, offset
);
4782 scop
= add(scop
, pet_skip_later
, offset
);
4787 /* Extract a clone of the kill statement in "scop".
4788 * "scop" is expected to have been created from a DeclStmt
4789 * and should have the kill as its first statement.
4791 struct pet_stmt
*PetScan::extract_kill(struct pet_scop
*scop
)
4793 struct pet_expr
*kill
;
4794 struct pet_stmt
*stmt
;
4795 isl_multi_pw_aff
*index
;
4800 if (scop
->n_stmt
< 1)
4801 isl_die(ctx
, isl_error_internal
,
4802 "expecting at least one statement", return NULL
);
4803 stmt
= scop
->stmts
[0];
4804 if (stmt
->body
->type
!= pet_expr_unary
||
4805 stmt
->body
->op
!= pet_op_kill
)
4806 isl_die(ctx
, isl_error_internal
,
4807 "expecting kill statement", return NULL
);
4809 index
= isl_multi_pw_aff_copy(stmt
->body
->args
[0]->acc
.index
);
4810 access
= isl_map_copy(stmt
->body
->args
[0]->acc
.access
);
4811 index
= isl_multi_pw_aff_reset_tuple_id(index
, isl_dim_in
);
4812 access
= isl_map_reset_tuple_id(access
, isl_dim_in
);
4813 kill
= pet_expr_kill_from_access_and_index(access
, index
);
4814 return pet_stmt_from_pet_expr(ctx
, stmt
->line
, NULL
, n_stmt
++, kill
);
4817 /* Mark all arrays in "scop" as being exposed.
4819 static struct pet_scop
*mark_exposed(struct pet_scop
*scop
)
4823 for (int i
= 0; i
< scop
->n_array
; ++i
)
4824 scop
->arrays
[i
]->exposed
= 1;
4828 /* Try and construct a pet_scop corresponding to (part of)
4829 * a sequence of statements.
4831 * "block" is set if the sequence respresents the children of
4832 * a compound statement.
4833 * "skip_declarations" is set if we should skip initial declarations
4834 * in the sequence of statements.
4836 * If there are any breaks or continues in the individual statements,
4837 * then we may have to compute a new skip condition.
4838 * This is handled using a pet_skip_info_seq object.
4839 * On initialization, the object checks if skip conditions need
4840 * to be computed. If so, it does so in "extract" and adds them in "add".
4842 * If "block" is set, then we need to insert kill statements at
4843 * the end of the block for any array that has been declared by
4844 * one of the statements in the sequence. Each of these declarations
4845 * results in the construction of a kill statement at the place
4846 * of the declaration, so we simply collect duplicates of
4847 * those kill statements and append these duplicates to the constructed scop.
4849 * If "block" is not set, then any array declared by one of the statements
4850 * in the sequence is marked as being exposed.
4852 * If autodetect is set, then we allow the extraction of only a subrange
4853 * of the sequence of statements. However, if there is at least one statement
4854 * for which we could not construct a scop and the final range contains
4855 * either no statements or at least one kill, then we discard the entire
4858 struct pet_scop
*PetScan::extract(StmtRange stmt_range
, bool block
,
4859 bool skip_declarations
)
4864 bool partial_range
= false;
4865 set
<struct pet_stmt
*> kills
;
4866 set
<struct pet_stmt
*>::iterator it
;
4868 scop
= pet_scop_empty(ctx
);
4869 for (i
= stmt_range
.first
, j
= 0; i
!= stmt_range
.second
; ++i
, ++j
) {
4871 struct pet_scop
*scop_i
;
4873 if (skip_declarations
&&
4874 child
->getStmtClass() == Stmt::DeclStmtClass
)
4877 scop_i
= extract(child
);
4878 if (scop
->n_stmt
!= 0 && partial
) {
4879 pet_scop_free(scop_i
);
4882 pet_skip_info_seq
skip(ctx
, scop
, scop_i
);
4885 scop_i
= pet_scop_prefix(scop_i
, 0);
4886 if (scop_i
&& child
->getStmtClass() == Stmt::DeclStmtClass
) {
4888 kills
.insert(extract_kill(scop_i
));
4890 scop_i
= mark_exposed(scop_i
);
4892 scop_i
= pet_scop_prefix(scop_i
, j
);
4893 if (options
->autodetect
) {
4895 scop
= pet_scop_add_seq(ctx
, scop
, scop_i
);
4897 partial_range
= true;
4898 if (scop
->n_stmt
!= 0 && !scop_i
)
4901 scop
= pet_scop_add_seq(ctx
, scop
, scop_i
);
4904 scop
= skip
.add(scop
, j
);
4906 if (partial
|| !scop
)
4910 for (it
= kills
.begin(); it
!= kills
.end(); ++it
) {
4912 scop_j
= pet_scop_from_pet_stmt(ctx
, *it
);
4913 scop_j
= pet_scop_prefix(scop_j
, j
);
4914 scop
= pet_scop_add_seq(ctx
, scop
, scop_j
);
4917 if (scop
&& partial_range
) {
4918 if (scop
->n_stmt
== 0 || kills
.size() != 0) {
4919 pet_scop_free(scop
);
4928 /* Check if the scop marked by the user is exactly this Stmt
4929 * or part of this Stmt.
4930 * If so, return a pet_scop corresponding to the marked region.
4931 * Otherwise, return NULL.
4933 struct pet_scop
*PetScan::scan(Stmt
*stmt
)
4935 SourceManager
&SM
= PP
.getSourceManager();
4936 unsigned start_off
, end_off
;
4938 start_off
= getExpansionOffset(SM
, stmt
->getLocStart());
4939 end_off
= getExpansionOffset(SM
, stmt
->getLocEnd());
4941 if (start_off
> loc
.end
)
4943 if (end_off
< loc
.start
)
4945 if (start_off
>= loc
.start
&& end_off
<= loc
.end
) {
4946 return extract(stmt
);
4950 for (start
= stmt
->child_begin(); start
!= stmt
->child_end(); ++start
) {
4951 Stmt
*child
= *start
;
4954 start_off
= getExpansionOffset(SM
, child
->getLocStart());
4955 end_off
= getExpansionOffset(SM
, child
->getLocEnd());
4956 if (start_off
< loc
.start
&& end_off
>= loc
.end
)
4958 if (start_off
>= loc
.start
)
4963 for (end
= start
; end
!= stmt
->child_end(); ++end
) {
4965 start_off
= SM
.getFileOffset(child
->getLocStart());
4966 if (start_off
>= loc
.end
)
4970 return extract(StmtRange(start
, end
), false, false);
4973 /* Set the size of index "pos" of "array" to "size".
4974 * In particular, add a constraint of the form
4978 * to array->extent and a constraint of the form
4982 * to array->context.
4984 static struct pet_array
*update_size(struct pet_array
*array
, int pos
,
4985 __isl_take isl_pw_aff
*size
)
4995 valid
= isl_pw_aff_nonneg_set(isl_pw_aff_copy(size
));
4996 array
->context
= isl_set_intersect(array
->context
, valid
);
4998 dim
= isl_set_get_space(array
->extent
);
4999 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
5000 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, pos
, 1);
5001 univ
= isl_set_universe(isl_aff_get_domain_space(aff
));
5002 index
= isl_pw_aff_alloc(univ
, aff
);
5004 size
= isl_pw_aff_add_dims(size
, isl_dim_in
,
5005 isl_set_dim(array
->extent
, isl_dim_set
));
5006 id
= isl_set_get_tuple_id(array
->extent
);
5007 size
= isl_pw_aff_set_tuple_id(size
, isl_dim_in
, id
);
5008 bound
= isl_pw_aff_lt_set(index
, size
);
5010 array
->extent
= isl_set_intersect(array
->extent
, bound
);
5012 if (!array
->context
|| !array
->extent
)
5017 pet_array_free(array
);
5021 /* Figure out the size of the array at position "pos" and all
5022 * subsequent positions from "type" and update "array" accordingly.
5024 struct pet_array
*PetScan::set_upper_bounds(struct pet_array
*array
,
5025 const Type
*type
, int pos
)
5027 const ArrayType
*atype
;
5033 if (type
->isPointerType()) {
5034 type
= type
->getPointeeType().getTypePtr();
5035 return set_upper_bounds(array
, type
, pos
+ 1);
5037 if (!type
->isArrayType())
5040 type
= type
->getCanonicalTypeInternal().getTypePtr();
5041 atype
= cast
<ArrayType
>(type
);
5043 if (type
->isConstantArrayType()) {
5044 const ConstantArrayType
*ca
= cast
<ConstantArrayType
>(atype
);
5045 size
= extract_affine(ca
->getSize());
5046 array
= update_size(array
, pos
, size
);
5047 } else if (type
->isVariableArrayType()) {
5048 const VariableArrayType
*vla
= cast
<VariableArrayType
>(atype
);
5049 size
= extract_affine(vla
->getSizeExpr());
5050 array
= update_size(array
, pos
, size
);
5053 type
= atype
->getElementType().getTypePtr();
5055 return set_upper_bounds(array
, type
, pos
+ 1);
5058 /* Is "T" the type of a variable length array with static size?
5060 static bool is_vla_with_static_size(QualType T
)
5062 const VariableArrayType
*vlatype
;
5064 if (!T
->isVariableArrayType())
5066 vlatype
= cast
<VariableArrayType
>(T
);
5067 return vlatype
->getSizeModifier() == VariableArrayType::Static
;
5070 /* Return the type of "decl" as an array.
5072 * In particular, if "decl" is a parameter declaration that
5073 * is a variable length array with a static size, then
5074 * return the original type (i.e., the variable length array).
5075 * Otherwise, return the type of decl.
5077 static QualType
get_array_type(ValueDecl
*decl
)
5082 parm
= dyn_cast
<ParmVarDecl
>(decl
);
5084 return decl
->getType();
5086 T
= parm
->getOriginalType();
5087 if (!is_vla_with_static_size(T
))
5088 return decl
->getType();
5092 /* Construct and return a pet_array corresponding to the variable "decl".
5093 * In particular, initialize array->extent to
5095 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
5097 * and then call set_upper_bounds to set the upper bounds on the indices
5098 * based on the type of the variable.
5100 struct pet_array
*PetScan::extract_array(isl_ctx
*ctx
, ValueDecl
*decl
)
5102 struct pet_array
*array
;
5103 QualType qt
= get_array_type(decl
);
5104 const Type
*type
= qt
.getTypePtr();
5105 int depth
= array_depth(type
);
5106 QualType base
= base_type(qt
);
5111 array
= isl_calloc_type(ctx
, struct pet_array
);
5115 id
= isl_id_alloc(ctx
, decl
->getName().str().c_str(), decl
);
5116 dim
= isl_space_set_alloc(ctx
, 0, depth
);
5117 dim
= isl_space_set_tuple_id(dim
, isl_dim_set
, id
);
5119 array
->extent
= isl_set_nat_universe(dim
);
5121 dim
= isl_space_params_alloc(ctx
, 0);
5122 array
->context
= isl_set_universe(dim
);
5124 array
= set_upper_bounds(array
, type
, 0);
5128 name
= base
.getAsString();
5129 array
->element_type
= strdup(name
.c_str());
5130 array
->element_size
= decl
->getASTContext().getTypeInfo(base
).first
/ 8;
5135 /* Construct a list of pet_arrays, one for each array (or scalar)
5136 * accessed inside "scop", add this list to "scop" and return the result.
5138 * The context of "scop" is updated with the intersection of
5139 * the contexts of all arrays, i.e., constraints on the parameters
5140 * that ensure that the arrays have a valid (non-negative) size.
5142 struct pet_scop
*PetScan::scan_arrays(struct pet_scop
*scop
)
5145 set
<ValueDecl
*> arrays
;
5146 set
<ValueDecl
*>::iterator it
;
5148 struct pet_array
**scop_arrays
;
5153 pet_scop_collect_arrays(scop
, arrays
);
5154 if (arrays
.size() == 0)
5157 n_array
= scop
->n_array
;
5159 scop_arrays
= isl_realloc_array(ctx
, scop
->arrays
, struct pet_array
*,
5160 n_array
+ arrays
.size());
5163 scop
->arrays
= scop_arrays
;
5165 for (it
= arrays
.begin(), i
= 0; it
!= arrays
.end(); ++it
, ++i
) {
5166 struct pet_array
*array
;
5167 scop
->arrays
[n_array
+ i
] = array
= extract_array(ctx
, *it
);
5168 if (!scop
->arrays
[n_array
+ i
])
5171 scop
->context
= isl_set_intersect(scop
->context
,
5172 isl_set_copy(array
->context
));
5179 pet_scop_free(scop
);
5183 /* Bound all parameters in scop->context to the possible values
5184 * of the corresponding C variable.
5186 static struct pet_scop
*add_parameter_bounds(struct pet_scop
*scop
)
5193 n
= isl_set_dim(scop
->context
, isl_dim_param
);
5194 for (int i
= 0; i
< n
; ++i
) {
5198 id
= isl_set_get_dim_id(scop
->context
, isl_dim_param
, i
);
5199 if (is_nested_parameter(id
)) {
5201 isl_die(isl_set_get_ctx(scop
->context
),
5203 "unresolved nested parameter", goto error
);
5205 decl
= (ValueDecl
*) isl_id_get_user(id
);
5208 scop
->context
= set_parameter_bounds(scop
->context
, i
, decl
);
5216 pet_scop_free(scop
);
5220 /* Construct a pet_scop from the given function.
5222 * If the scop was delimited by scop and endscop pragmas, then we override
5223 * the file offsets by those derived from the pragmas.
5225 struct pet_scop
*PetScan::scan(FunctionDecl
*fd
)
5230 stmt
= fd
->getBody();
5232 if (options
->autodetect
)
5233 scop
= extract(stmt
, true);
5236 scop
= pet_scop_update_start_end(scop
, loc
.start
, loc
.end
);
5238 scop
= pet_scop_detect_parameter_accesses(scop
);
5239 scop
= scan_arrays(scop
);
5240 scop
= add_parameter_bounds(scop
);
5241 scop
= pet_scop_gist(scop
, value_bounds
);