2 * Copyright 2011 Leiden University. All rights reserved.
3 * Copyright 2012 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
, *res
;
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_set
))
893 id
= isl_multi_pw_aff_get_tuple_id(index
, isl_dim_set
);
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
)
1475 struct pet_expr
*expr
;
1479 access
= isl_map_from_range(isl_set_copy(array
->extent
));
1480 expr
= pet_expr_kill_from_access(access
);
1481 return extract(stmt
, expr
);
1484 /* Construct a pet_scop for a (single) variable declaration.
1486 * The scop contains the variable being declared (as an array)
1487 * and a statement killing the array.
1489 * If the variable is initialized in the AST, then the scop
1490 * also contains an assignment to the variable.
1492 struct pet_scop
*PetScan::extract(DeclStmt
*stmt
)
1496 struct pet_expr
*lhs
, *rhs
, *pe
;
1497 struct pet_scop
*scop_decl
, *scop
;
1498 struct pet_array
*array
;
1500 if (!stmt
->isSingleDecl()) {
1505 decl
= stmt
->getSingleDecl();
1506 vd
= cast
<VarDecl
>(decl
);
1508 array
= extract_array(ctx
, vd
);
1510 array
->declared
= 1;
1511 scop_decl
= kill(stmt
, array
);
1512 scop_decl
= pet_scop_add_array(scop_decl
, array
);
1517 lhs
= extract_access_expr(vd
);
1518 rhs
= extract_expr(vd
->getInit());
1521 assign(lhs
, vd
->getInit());
1523 pe
= pet_expr_new_binary(ctx
, pet_op_assign
, lhs
, rhs
);
1524 scop
= extract(stmt
, pe
);
1526 scop_decl
= pet_scop_prefix(scop_decl
, 0);
1527 scop
= pet_scop_prefix(scop
, 1);
1529 scop
= pet_scop_add_seq(ctx
, scop_decl
, scop
);
1534 /* Construct a pet_expr representing a conditional operation.
1536 * We first try to extract the condition as an affine expression.
1537 * If that fails, we construct a pet_expr tree representing the condition.
1539 struct pet_expr
*PetScan::extract_expr(ConditionalOperator
*expr
)
1541 struct pet_expr
*cond
, *lhs
, *rhs
;
1544 pa
= try_extract_affine(expr
->getCond());
1546 isl_multi_pw_aff
*test
= isl_multi_pw_aff_from_pw_aff(pa
);
1547 test
= isl_multi_pw_aff_from_range(test
);
1548 cond
= pet_expr_from_index(test
);
1550 cond
= extract_expr(expr
->getCond());
1551 lhs
= extract_expr(expr
->getTrueExpr());
1552 rhs
= extract_expr(expr
->getFalseExpr());
1554 return pet_expr_new_ternary(ctx
, cond
, lhs
, rhs
);
1557 struct pet_expr
*PetScan::extract_expr(ImplicitCastExpr
*expr
)
1559 return extract_expr(expr
->getSubExpr());
1562 /* Construct a pet_expr representing a floating point value.
1564 * If the floating point literal does not appear in a macro,
1565 * then we use the original representation in the source code
1566 * as the string representation. Otherwise, we use the pretty
1567 * printer to produce a string representation.
1569 struct pet_expr
*PetScan::extract_expr(FloatingLiteral
*expr
)
1573 const LangOptions
&LO
= PP
.getLangOpts();
1574 SourceLocation loc
= expr
->getLocation();
1576 if (!loc
.isMacroID()) {
1577 SourceManager
&SM
= PP
.getSourceManager();
1578 unsigned len
= Lexer::MeasureTokenLength(loc
, SM
, LO
);
1579 s
= string(SM
.getCharacterData(loc
), len
);
1581 llvm::raw_string_ostream
S(s
);
1582 expr
->printPretty(S
, 0, PrintingPolicy(LO
));
1585 d
= expr
->getValueAsApproximateDouble();
1586 return pet_expr_new_double(ctx
, d
, s
.c_str());
1589 /* Extract an index expression from "expr" and then convert it into
1590 * an access pet_expr.
1592 struct pet_expr
*PetScan::extract_access_expr(Expr
*expr
)
1594 isl_multi_pw_aff
*index
;
1595 struct pet_expr
*pe
;
1598 index
= extract_index(expr
);
1599 depth
= extract_depth(index
);
1601 pe
= pet_expr_from_index_and_depth(index
, depth
);
1606 /* Extract an index expression from "decl" and then convert it into
1607 * an access pet_expr.
1609 struct pet_expr
*PetScan::extract_access_expr(ValueDecl
*decl
)
1611 isl_multi_pw_aff
*index
;
1612 struct pet_expr
*pe
;
1615 index
= extract_index(decl
);
1616 depth
= extract_depth(index
);
1618 pe
= pet_expr_from_index_and_depth(index
, depth
);
1623 struct pet_expr
*PetScan::extract_expr(ParenExpr
*expr
)
1625 return extract_expr(expr
->getSubExpr());
1628 /* Construct a pet_expr representing a function call.
1630 * If we are passing along a pointer to an array element
1631 * or an entire row or even higher dimensional slice of an array,
1632 * then the function being called may write into the array.
1634 * We assume here that if the function is declared to take a pointer
1635 * to a const type, then the function will perform a read
1636 * and that otherwise, it will perform a write.
1638 struct pet_expr
*PetScan::extract_expr(CallExpr
*expr
)
1640 struct pet_expr
*res
= NULL
;
1644 fd
= expr
->getDirectCallee();
1650 name
= fd
->getDeclName().getAsString();
1651 res
= pet_expr_new_call(ctx
, name
.c_str(), expr
->getNumArgs());
1655 for (int i
= 0; i
< expr
->getNumArgs(); ++i
) {
1656 Expr
*arg
= expr
->getArg(i
);
1660 if (arg
->getStmtClass() == Stmt::ImplicitCastExprClass
) {
1661 ImplicitCastExpr
*ice
= cast
<ImplicitCastExpr
>(arg
);
1662 arg
= ice
->getSubExpr();
1664 if (arg
->getStmtClass() == Stmt::UnaryOperatorClass
) {
1665 UnaryOperator
*op
= cast
<UnaryOperator
>(arg
);
1666 if (op
->getOpcode() == UO_AddrOf
) {
1668 arg
= op
->getSubExpr();
1671 res
->args
[i
] = PetScan::extract_expr(arg
);
1672 main_arg
= res
->args
[i
];
1674 res
->args
[i
] = pet_expr_new_unary(ctx
,
1675 pet_op_address_of
, res
->args
[i
]);
1678 if (arg
->getStmtClass() == Stmt::ArraySubscriptExprClass
&&
1679 array_depth(arg
->getType().getTypePtr()) > 0)
1681 if (is_addr
&& main_arg
->type
== pet_expr_access
) {
1683 if (!fd
->hasPrototype()) {
1684 unsupported(expr
, "prototype required");
1687 parm
= fd
->getParamDecl(i
);
1688 if (!const_base(parm
->getType()))
1689 mark_write(main_arg
);
1699 /* Construct a pet_expr representing a (C style) cast.
1701 struct pet_expr
*PetScan::extract_expr(CStyleCastExpr
*expr
)
1703 struct pet_expr
*arg
;
1706 arg
= extract_expr(expr
->getSubExpr());
1710 type
= expr
->getTypeAsWritten();
1711 return pet_expr_new_cast(ctx
, type
.getAsString().c_str(), arg
);
1714 /* Try and onstruct a pet_expr representing "expr".
1716 struct pet_expr
*PetScan::extract_expr(Expr
*expr
)
1718 switch (expr
->getStmtClass()) {
1719 case Stmt::UnaryOperatorClass
:
1720 return extract_expr(cast
<UnaryOperator
>(expr
));
1721 case Stmt::CompoundAssignOperatorClass
:
1722 case Stmt::BinaryOperatorClass
:
1723 return extract_expr(cast
<BinaryOperator
>(expr
));
1724 case Stmt::ImplicitCastExprClass
:
1725 return extract_expr(cast
<ImplicitCastExpr
>(expr
));
1726 case Stmt::ArraySubscriptExprClass
:
1727 case Stmt::DeclRefExprClass
:
1728 case Stmt::IntegerLiteralClass
:
1729 return extract_access_expr(expr
);
1730 case Stmt::FloatingLiteralClass
:
1731 return extract_expr(cast
<FloatingLiteral
>(expr
));
1732 case Stmt::ParenExprClass
:
1733 return extract_expr(cast
<ParenExpr
>(expr
));
1734 case Stmt::ConditionalOperatorClass
:
1735 return extract_expr(cast
<ConditionalOperator
>(expr
));
1736 case Stmt::CallExprClass
:
1737 return extract_expr(cast
<CallExpr
>(expr
));
1738 case Stmt::CStyleCastExprClass
:
1739 return extract_expr(cast
<CStyleCastExpr
>(expr
));
1746 /* Check if the given initialization statement is an assignment.
1747 * If so, return that assignment. Otherwise return NULL.
1749 BinaryOperator
*PetScan::initialization_assignment(Stmt
*init
)
1751 BinaryOperator
*ass
;
1753 if (init
->getStmtClass() != Stmt::BinaryOperatorClass
)
1756 ass
= cast
<BinaryOperator
>(init
);
1757 if (ass
->getOpcode() != BO_Assign
)
1763 /* Check if the given initialization statement is a declaration
1764 * of a single variable.
1765 * If so, return that declaration. Otherwise return NULL.
1767 Decl
*PetScan::initialization_declaration(Stmt
*init
)
1771 if (init
->getStmtClass() != Stmt::DeclStmtClass
)
1774 decl
= cast
<DeclStmt
>(init
);
1776 if (!decl
->isSingleDecl())
1779 return decl
->getSingleDecl();
1782 /* Given the assignment operator in the initialization of a for loop,
1783 * extract the induction variable, i.e., the (integer)variable being
1786 ValueDecl
*PetScan::extract_induction_variable(BinaryOperator
*init
)
1793 lhs
= init
->getLHS();
1794 if (lhs
->getStmtClass() != Stmt::DeclRefExprClass
) {
1799 ref
= cast
<DeclRefExpr
>(lhs
);
1800 decl
= ref
->getDecl();
1801 type
= decl
->getType().getTypePtr();
1803 if (!type
->isIntegerType()) {
1811 /* Given the initialization statement of a for loop and the single
1812 * declaration in this initialization statement,
1813 * extract the induction variable, i.e., the (integer) variable being
1816 VarDecl
*PetScan::extract_induction_variable(Stmt
*init
, Decl
*decl
)
1820 vd
= cast
<VarDecl
>(decl
);
1822 const QualType type
= vd
->getType();
1823 if (!type
->isIntegerType()) {
1828 if (!vd
->getInit()) {
1836 /* Check that op is of the form iv++ or iv--.
1837 * Return an affine expression "1" or "-1" accordingly.
1839 __isl_give isl_pw_aff
*PetScan::extract_unary_increment(
1840 clang::UnaryOperator
*op
, clang::ValueDecl
*iv
)
1847 if (!op
->isIncrementDecrementOp()) {
1852 sub
= op
->getSubExpr();
1853 if (sub
->getStmtClass() != Stmt::DeclRefExprClass
) {
1858 ref
= cast
<DeclRefExpr
>(sub
);
1859 if (ref
->getDecl() != iv
) {
1864 space
= isl_space_params_alloc(ctx
, 0);
1865 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
1867 if (op
->isIncrementOp())
1868 aff
= isl_aff_add_constant_si(aff
, 1);
1870 aff
= isl_aff_add_constant_si(aff
, -1);
1872 return isl_pw_aff_from_aff(aff
);
1875 /* If the isl_pw_aff on which isl_pw_aff_foreach_piece is called
1876 * has a single constant expression, then put this constant in *user.
1877 * The caller is assumed to have checked that this function will
1878 * be called exactly once.
1880 static int extract_cst(__isl_take isl_set
*set
, __isl_take isl_aff
*aff
,
1883 isl_val
**inc
= (isl_val
**)user
;
1886 if (isl_aff_is_cst(aff
))
1887 *inc
= isl_aff_get_constant_val(aff
);
1897 /* Check if op is of the form
1901 * and return inc as an affine expression.
1903 * We extract an affine expression from the RHS, subtract iv and return
1906 __isl_give isl_pw_aff
*PetScan::extract_binary_increment(BinaryOperator
*op
,
1907 clang::ValueDecl
*iv
)
1916 if (op
->getOpcode() != BO_Assign
) {
1922 if (lhs
->getStmtClass() != Stmt::DeclRefExprClass
) {
1927 ref
= cast
<DeclRefExpr
>(lhs
);
1928 if (ref
->getDecl() != iv
) {
1933 val
= extract_affine(op
->getRHS());
1935 id
= isl_id_alloc(ctx
, iv
->getName().str().c_str(), iv
);
1937 dim
= isl_space_params_alloc(ctx
, 1);
1938 dim
= isl_space_set_dim_id(dim
, isl_dim_param
, 0, id
);
1939 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
1940 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_param
, 0, 1);
1942 val
= isl_pw_aff_sub(val
, isl_pw_aff_from_aff(aff
));
1947 /* Check that op is of the form iv += cst or iv -= cst
1948 * and return an affine expression corresponding oto cst or -cst accordingly.
1950 __isl_give isl_pw_aff
*PetScan::extract_compound_increment(
1951 CompoundAssignOperator
*op
, clang::ValueDecl
*iv
)
1957 BinaryOperatorKind opcode
;
1959 opcode
= op
->getOpcode();
1960 if (opcode
!= BO_AddAssign
&& opcode
!= BO_SubAssign
) {
1964 if (opcode
== BO_SubAssign
)
1968 if (lhs
->getStmtClass() != Stmt::DeclRefExprClass
) {
1973 ref
= cast
<DeclRefExpr
>(lhs
);
1974 if (ref
->getDecl() != iv
) {
1979 val
= extract_affine(op
->getRHS());
1981 val
= isl_pw_aff_neg(val
);
1986 /* Check that the increment of the given for loop increments
1987 * (or decrements) the induction variable "iv" and return
1988 * the increment as an affine expression if successful.
1990 __isl_give isl_pw_aff
*PetScan::extract_increment(clang::ForStmt
*stmt
,
1993 Stmt
*inc
= stmt
->getInc();
2000 if (inc
->getStmtClass() == Stmt::UnaryOperatorClass
)
2001 return extract_unary_increment(cast
<UnaryOperator
>(inc
), iv
);
2002 if (inc
->getStmtClass() == Stmt::CompoundAssignOperatorClass
)
2003 return extract_compound_increment(
2004 cast
<CompoundAssignOperator
>(inc
), iv
);
2005 if (inc
->getStmtClass() == Stmt::BinaryOperatorClass
)
2006 return extract_binary_increment(cast
<BinaryOperator
>(inc
), iv
);
2012 /* Embed the given iteration domain in an extra outer loop
2013 * with induction variable "var".
2014 * If this variable appeared as a parameter in the constraints,
2015 * it is replaced by the new outermost dimension.
2017 static __isl_give isl_set
*embed(__isl_take isl_set
*set
,
2018 __isl_take isl_id
*var
)
2022 set
= isl_set_insert_dims(set
, isl_dim_set
, 0, 1);
2023 pos
= isl_set_find_dim_by_id(set
, isl_dim_param
, var
);
2025 set
= isl_set_equate(set
, isl_dim_param
, pos
, isl_dim_set
, 0);
2026 set
= isl_set_project_out(set
, isl_dim_param
, pos
, 1);
2033 /* Return those elements in the space of "cond" that come after
2034 * (based on "sign") an element in "cond".
2036 static __isl_give isl_set
*after(__isl_take isl_set
*cond
, int sign
)
2038 isl_map
*previous_to_this
;
2041 previous_to_this
= isl_map_lex_lt(isl_set_get_space(cond
));
2043 previous_to_this
= isl_map_lex_gt(isl_set_get_space(cond
));
2045 cond
= isl_set_apply(cond
, previous_to_this
);
2050 /* Create the infinite iteration domain
2052 * { [id] : id >= 0 }
2054 * If "scop" has an affine skip of type pet_skip_later,
2055 * then remove those iterations i that have an earlier iteration
2056 * where the skip condition is satisfied, meaning that iteration i
2058 * Since we are dealing with a loop without loop iterator,
2059 * the skip condition cannot refer to the current loop iterator and
2060 * so effectively, the returned set is of the form
2062 * { [0]; [id] : id >= 1 and not skip }
2064 static __isl_give isl_set
*infinite_domain(__isl_take isl_id
*id
,
2065 struct pet_scop
*scop
)
2067 isl_ctx
*ctx
= isl_id_get_ctx(id
);
2071 domain
= isl_set_nat_universe(isl_space_set_alloc(ctx
, 0, 1));
2072 domain
= isl_set_set_dim_id(domain
, isl_dim_set
, 0, id
);
2074 if (!pet_scop_has_affine_skip(scop
, pet_skip_later
))
2077 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_later
);
2078 skip
= embed(skip
, isl_id_copy(id
));
2079 skip
= isl_set_intersect(skip
, isl_set_copy(domain
));
2080 domain
= isl_set_subtract(domain
, after(skip
, 1));
2085 /* Create an identity affine expression on the space containing "domain",
2086 * which is assumed to be one-dimensional.
2088 static __isl_give isl_aff
*identity_aff(__isl_keep isl_set
*domain
)
2090 isl_local_space
*ls
;
2092 ls
= isl_local_space_from_space(isl_set_get_space(domain
));
2093 return isl_aff_var_on_domain(ls
, isl_dim_set
, 0);
2096 /* Create a map that maps elements of a single-dimensional array "id_test"
2097 * to the previous element (according to "inc"), provided this element
2098 * belongs to "domain". That is, create the map
2100 * { id[x] -> id[x - inc] : x - inc in domain }
2102 static __isl_give isl_map
*map_to_previous(__isl_take isl_id
*id_test
,
2103 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
2106 isl_local_space
*ls
;
2110 space
= isl_set_get_space(domain
);
2111 ls
= isl_local_space_from_space(space
);
2112 aff
= isl_aff_var_on_domain(ls
, isl_dim_set
, 0);
2113 aff
= isl_aff_add_constant_val(aff
, isl_val_neg(inc
));
2114 prev
= isl_map_from_aff(aff
);
2115 prev
= isl_map_intersect_range(prev
, domain
);
2116 prev
= isl_map_set_tuple_id(prev
, isl_dim_out
, id_test
);
2121 /* Add an implication to "scop" expressing that if an element of
2122 * virtual array "id_test" has value "satisfied" then all previous elements
2123 * of this array also have that value. The set of previous elements
2124 * is bounded by "domain". If "sign" is negative then iterator
2125 * is decreasing and we express that all subsequent array elements
2126 * (but still defined previously) have the same value.
2128 static struct pet_scop
*add_implication(struct pet_scop
*scop
,
2129 __isl_take isl_id
*id_test
, __isl_take isl_set
*domain
, int sign
,
2135 domain
= isl_set_set_tuple_id(domain
, id_test
);
2136 space
= isl_set_get_space(domain
);
2138 map
= isl_map_lex_ge(space
);
2140 map
= isl_map_lex_le(space
);
2141 map
= isl_map_intersect_range(map
, domain
);
2142 scop
= pet_scop_add_implication(scop
, map
, satisfied
);
2147 /* Add a filter to "scop" that imposes that it is only executed
2148 * when the variable identified by "id_test" has a zero value
2149 * for all previous iterations of "domain".
2151 * In particular, add a filter that imposes that the array
2152 * has a zero value at the previous iteration of domain and
2153 * add an implication that implies that it then has that
2154 * value for all previous iterations.
2156 static struct pet_scop
*scop_add_break(struct pet_scop
*scop
,
2157 __isl_take isl_id
*id_test
, __isl_take isl_set
*domain
,
2158 __isl_take isl_val
*inc
)
2161 int sign
= isl_val_sgn(inc
);
2163 prev
= map_to_previous(isl_id_copy(id_test
), isl_set_copy(domain
), inc
);
2164 scop
= add_implication(scop
, id_test
, domain
, sign
, 0);
2165 scop
= pet_scop_filter(scop
, prev
, 0);
2170 /* Construct a pet_scop for an infinite loop around the given body.
2172 * We extract a pet_scop for the body and then embed it in a loop with
2181 * If the body contains any break, then it is taken into
2182 * account in infinite_domain (if the skip condition is affine)
2183 * or in scop_add_break (if the skip condition is not affine).
2185 struct pet_scop
*PetScan::extract_infinite_loop(Stmt
*body
)
2187 isl_id
*id
, *id_test
;
2190 struct pet_scop
*scop
;
2193 scop
= extract(body
);
2197 id
= isl_id_alloc(ctx
, "t", NULL
);
2198 domain
= infinite_domain(isl_id_copy(id
), scop
);
2199 ident
= identity_aff(domain
);
2201 has_var_break
= pet_scop_has_var_skip(scop
, pet_skip_later
);
2203 id_test
= pet_scop_get_skip_id(scop
, pet_skip_later
);
2205 scop
= pet_scop_embed(scop
, isl_set_copy(domain
),
2206 isl_map_from_aff(isl_aff_copy(ident
)), ident
, id
);
2208 scop
= scop_add_break(scop
, id_test
, domain
, isl_val_one(ctx
));
2210 isl_set_free(domain
);
2215 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
2221 struct pet_scop
*PetScan::extract_infinite_for(ForStmt
*stmt
)
2223 return extract_infinite_loop(stmt
->getBody());
2226 /* Create an index expression for an access to a virtual array
2227 * representing the result of a condition.
2228 * Unlike other accessed data, the id of the array is NULL as
2229 * there is no ValueDecl in the program corresponding to the virtual
2231 * The array starts out as a scalar, but grows along with the
2232 * statement writing to the array in pet_scop_embed.
2234 static __isl_give isl_multi_pw_aff
*create_test_index(isl_ctx
*ctx
, int test_nr
)
2236 isl_space
*dim
= isl_space_alloc(ctx
, 0, 0, 0);
2240 snprintf(name
, sizeof(name
), "__pet_test_%d", test_nr
);
2241 id
= isl_id_alloc(ctx
, name
, NULL
);
2242 dim
= isl_space_set_tuple_id(dim
, isl_dim_out
, id
);
2243 return isl_multi_pw_aff_zero(dim
);
2246 /* Create an access to a virtual array representing the result
2249 static __isl_give isl_map
*create_test_access(isl_ctx
*ctx
, int test_nr
)
2251 return isl_map_from_multi_pw_aff(create_test_index(ctx
, test_nr
));
2254 /* Add an array with the given extent ("access") to the list
2255 * of arrays in "scop" and return the extended pet_scop.
2256 * The array is marked as attaining values 0 and 1 only and
2257 * as each element being assigned at most once.
2259 static struct pet_scop
*scop_add_array(struct pet_scop
*scop
,
2260 __isl_keep isl_map
*access
, clang::ASTContext
&ast_ctx
)
2262 isl_ctx
*ctx
= isl_map_get_ctx(access
);
2264 struct pet_array
*array
;
2271 array
= isl_calloc_type(ctx
, struct pet_array
);
2275 array
->extent
= isl_map_range(isl_map_copy(access
));
2276 dim
= isl_space_params_alloc(ctx
, 0);
2277 array
->context
= isl_set_universe(dim
);
2278 dim
= isl_space_set_alloc(ctx
, 0, 1);
2279 array
->value_bounds
= isl_set_universe(dim
);
2280 array
->value_bounds
= isl_set_lower_bound_si(array
->value_bounds
,
2282 array
->value_bounds
= isl_set_upper_bound_si(array
->value_bounds
,
2284 array
->element_type
= strdup("int");
2285 array
->element_size
= ast_ctx
.getTypeInfo(ast_ctx
.IntTy
).first
/ 8;
2286 array
->uniquely_defined
= 1;
2288 if (!array
->extent
|| !array
->context
)
2289 array
= pet_array_free(array
);
2291 scop
= pet_scop_add_array(scop
, array
);
2295 pet_scop_free(scop
);
2299 /* Construct a pet_scop for a while loop of the form
2304 * In particular, construct a scop for an infinite loop around body and
2305 * intersect the domain with the affine expression.
2306 * Note that this intersection may result in an empty loop.
2308 struct pet_scop
*PetScan::extract_affine_while(__isl_take isl_pw_aff
*pa
,
2311 struct pet_scop
*scop
;
2315 valid
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
2316 dom
= isl_pw_aff_non_zero_set(pa
);
2317 scop
= extract_infinite_loop(body
);
2318 scop
= pet_scop_restrict(scop
, dom
);
2319 scop
= pet_scop_restrict_context(scop
, valid
);
2324 /* Construct a scop for a while, given the scops for the condition
2325 * and the body, the filter identifier and the iteration domain of
2328 * In particular, the scop for the condition is filtered to depend
2329 * on "id_test" evaluating to true for all previous iterations
2330 * of the loop, while the scop for the body is filtered to depend
2331 * on "id_test" evaluating to true for all iterations up to the
2332 * current iteration.
2333 * The actual filter only imposes that this virtual array has
2334 * value one on the previous or the current iteration.
2335 * The fact that this condition also applies to the previous
2336 * iterations is enforced by an implication.
2338 * These filtered scops are then combined into a single scop.
2340 * "sign" is positive if the iterator increases and negative
2343 static struct pet_scop
*scop_add_while(struct pet_scop
*scop_cond
,
2344 struct pet_scop
*scop_body
, __isl_take isl_id
*id_test
,
2345 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
2347 isl_ctx
*ctx
= isl_set_get_ctx(domain
);
2349 isl_map
*test_access
;
2351 int sign
= isl_val_sgn(inc
);
2352 struct pet_scop
*scop
;
2354 prev
= map_to_previous(isl_id_copy(id_test
), isl_set_copy(domain
), inc
);
2355 scop_cond
= pet_scop_filter(scop_cond
, prev
, 1);
2357 space
= isl_space_map_from_set(isl_set_get_space(domain
));
2358 test_access
= isl_map_identity(space
);
2359 test_access
= isl_map_set_tuple_id(test_access
, isl_dim_out
,
2360 isl_id_copy(id_test
));
2361 scop_body
= pet_scop_filter(scop_body
, test_access
, 1);
2363 scop
= pet_scop_add_seq(ctx
, scop_cond
, scop_body
);
2364 scop
= add_implication(scop
, id_test
, domain
, sign
, 1);
2369 /* Check if the while loop is of the form
2371 * while (affine expression)
2374 * If so, call extract_affine_while to construct a scop.
2376 * Otherwise, construct a generic while scop, with iteration domain
2377 * { [t] : t >= 0 }. The scop consists of two parts, one for
2378 * evaluating the condition and one for the body.
2379 * The schedule is adjusted to reflect that the condition is evaluated
2380 * before the body is executed and the body is filtered to depend
2381 * on the result of the condition evaluating to true on all iterations
2382 * up to the current iteration, while the evaluation the condition itself
2383 * is filtered to depend on the result of the condition evaluating to true
2384 * on all previous iterations.
2385 * The context of the scop representing the body is dropped
2386 * because we don't know how many times the body will be executed,
2389 * If the body contains any break, then it is taken into
2390 * account in infinite_domain (if the skip condition is affine)
2391 * or in scop_add_break (if the skip condition is not affine).
2393 struct pet_scop
*PetScan::extract(WhileStmt
*stmt
)
2396 isl_id
*id
, *id_test
, *id_break_test
;
2397 isl_map
*test_access
;
2401 struct pet_scop
*scop
, *scop_body
;
2404 cond
= stmt
->getCond();
2410 clear_assignments
clear(assigned_value
);
2411 clear
.TraverseStmt(stmt
->getBody());
2413 pa
= try_extract_affine_condition(cond
);
2415 return extract_affine_while(pa
, stmt
->getBody());
2417 if (!allow_nested
) {
2422 test_access
= create_test_access(ctx
, n_test
++);
2423 scop
= extract_non_affine_condition(cond
, isl_map_copy(test_access
));
2424 scop
= scop_add_array(scop
, test_access
, ast_context
);
2425 id_test
= isl_map_get_tuple_id(test_access
, isl_dim_out
);
2426 isl_map_free(test_access
);
2427 scop_body
= extract(stmt
->getBody());
2429 id
= isl_id_alloc(ctx
, "t", NULL
);
2430 domain
= infinite_domain(isl_id_copy(id
), scop_body
);
2431 ident
= identity_aff(domain
);
2433 has_var_break
= pet_scop_has_var_skip(scop_body
, pet_skip_later
);
2435 id_break_test
= pet_scop_get_skip_id(scop_body
, pet_skip_later
);
2437 scop
= pet_scop_prefix(scop
, 0);
2438 scop
= pet_scop_embed(scop
, isl_set_copy(domain
),
2439 isl_map_from_aff(isl_aff_copy(ident
)),
2440 isl_aff_copy(ident
), isl_id_copy(id
));
2441 scop_body
= pet_scop_reset_context(scop_body
);
2442 scop_body
= pet_scop_prefix(scop_body
, 1);
2443 scop_body
= pet_scop_embed(scop_body
, isl_set_copy(domain
),
2444 isl_map_from_aff(isl_aff_copy(ident
)), ident
, id
);
2446 if (has_var_break
) {
2447 scop
= scop_add_break(scop
, isl_id_copy(id_break_test
),
2448 isl_set_copy(domain
), isl_val_one(ctx
));
2449 scop_body
= scop_add_break(scop_body
, id_break_test
,
2450 isl_set_copy(domain
), isl_val_one(ctx
));
2452 scop
= scop_add_while(scop
, scop_body
, id_test
, domain
,
2458 /* Check whether "cond" expresses a simple loop bound
2459 * on the only set dimension.
2460 * In particular, if "up" is set then "cond" should contain only
2461 * upper bounds on the set dimension.
2462 * Otherwise, it should contain only lower bounds.
2464 static bool is_simple_bound(__isl_keep isl_set
*cond
, __isl_keep isl_val
*inc
)
2466 if (isl_val_is_pos(inc
))
2467 return !isl_set_dim_has_any_lower_bound(cond
, isl_dim_set
, 0);
2469 return !isl_set_dim_has_any_upper_bound(cond
, isl_dim_set
, 0);
2472 /* Extend a condition on a given iteration of a loop to one that
2473 * imposes the same condition on all previous iterations.
2474 * "domain" expresses the lower [upper] bound on the iterations
2475 * when inc is positive [negative].
2477 * In particular, we construct the condition (when inc is positive)
2479 * forall i' : (domain(i') and i' <= i) => cond(i')
2481 * which is equivalent to
2483 * not exists i' : domain(i') and i' <= i and not cond(i')
2485 * We construct this set by negating cond, applying a map
2487 * { [i'] -> [i] : domain(i') and i' <= i }
2489 * and then negating the result again.
2491 static __isl_give isl_set
*valid_for_each_iteration(__isl_take isl_set
*cond
,
2492 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
2494 isl_map
*previous_to_this
;
2496 if (isl_val_is_pos(inc
))
2497 previous_to_this
= isl_map_lex_le(isl_set_get_space(domain
));
2499 previous_to_this
= isl_map_lex_ge(isl_set_get_space(domain
));
2501 previous_to_this
= isl_map_intersect_domain(previous_to_this
, domain
);
2503 cond
= isl_set_complement(cond
);
2504 cond
= isl_set_apply(cond
, previous_to_this
);
2505 cond
= isl_set_complement(cond
);
2512 /* Construct a domain of the form
2514 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
2516 static __isl_give isl_set
*strided_domain(__isl_take isl_id
*id
,
2517 __isl_take isl_pw_aff
*init
, __isl_take isl_val
*inc
)
2523 init
= isl_pw_aff_insert_dims(init
, isl_dim_in
, 0, 1);
2524 dim
= isl_pw_aff_get_domain_space(init
);
2525 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
2526 aff
= isl_aff_add_coefficient_val(aff
, isl_dim_in
, 0, inc
);
2527 init
= isl_pw_aff_add(init
, isl_pw_aff_from_aff(aff
));
2529 dim
= isl_space_set_alloc(isl_pw_aff_get_ctx(init
), 1, 1);
2530 dim
= isl_space_set_dim_id(dim
, isl_dim_param
, 0, id
);
2531 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
2532 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_param
, 0, 1);
2534 set
= isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff
), init
);
2536 set
= isl_set_lower_bound_si(set
, isl_dim_set
, 0, 0);
2538 return isl_set_params(set
);
2541 /* Assuming "cond" represents a bound on a loop where the loop
2542 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2545 * Under the given assumptions, wrapping is only possible if "cond" allows
2546 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2547 * increasing iterator and 0 in case of a decreasing iterator.
2549 static bool can_wrap(__isl_keep isl_set
*cond
, ValueDecl
*iv
,
2550 __isl_keep isl_val
*inc
)
2557 test
= isl_set_copy(cond
);
2559 ctx
= isl_set_get_ctx(test
);
2560 if (isl_val_is_neg(inc
))
2561 limit
= isl_val_zero(ctx
);
2563 limit
= isl_val_int_from_ui(ctx
, get_type_size(iv
));
2564 limit
= isl_val_2exp(limit
);
2565 limit
= isl_val_sub_ui(limit
, 1);
2568 test
= isl_set_fix_val(cond
, isl_dim_set
, 0, limit
);
2569 cw
= !isl_set_is_empty(test
);
2575 /* Given a one-dimensional space, construct the following affine expression
2578 * { [v] -> [v mod 2^width] }
2580 * where width is the number of bits used to represent the values
2581 * of the unsigned variable "iv".
2583 static __isl_give isl_aff
*compute_wrapping(__isl_take isl_space
*dim
,
2591 ctx
= isl_space_get_ctx(dim
);
2592 mod
= isl_val_int_from_ui(ctx
, get_type_size(iv
));
2593 mod
= isl_val_2exp(mod
);
2595 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
2596 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, 0, 1);
2597 aff
= isl_aff_mod_val(aff
, mod
);
2602 /* Project out the parameter "id" from "set".
2604 static __isl_give isl_set
*set_project_out_by_id(__isl_take isl_set
*set
,
2605 __isl_keep isl_id
*id
)
2609 pos
= isl_set_find_dim_by_id(set
, isl_dim_param
, id
);
2611 set
= isl_set_project_out(set
, isl_dim_param
, pos
, 1);
2616 /* Compute the set of parameters for which "set1" is a subset of "set2".
2618 * set1 is a subset of set2 if
2620 * forall i in set1 : i in set2
2624 * not exists i in set1 and i not in set2
2628 * not exists i in set1 \ set2
2630 static __isl_give isl_set
*enforce_subset(__isl_take isl_set
*set1
,
2631 __isl_take isl_set
*set2
)
2633 return isl_set_complement(isl_set_params(isl_set_subtract(set1
, set2
)));
2636 /* Compute the set of parameter values for which "cond" holds
2637 * on the next iteration for each element of "dom".
2639 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2640 * and then compute the set of parameters for which the result is a subset
2643 static __isl_give isl_set
*valid_on_next(__isl_take isl_set
*cond
,
2644 __isl_take isl_set
*dom
, __isl_take isl_val
*inc
)
2650 space
= isl_set_get_space(dom
);
2651 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
2652 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, 0, 1);
2653 aff
= isl_aff_add_constant_val(aff
, inc
);
2654 next
= isl_map_from_basic_map(isl_basic_map_from_aff(aff
));
2656 dom
= isl_set_apply(dom
, next
);
2658 return enforce_subset(dom
, cond
);
2661 /* Does "id" refer to a nested access?
2663 static bool is_nested_parameter(__isl_keep isl_id
*id
)
2665 return id
&& isl_id_get_user(id
) && !isl_id_get_name(id
);
2668 /* Does parameter "pos" of "space" refer to a nested access?
2670 static bool is_nested_parameter(__isl_keep isl_space
*space
, int pos
)
2675 id
= isl_space_get_dim_id(space
, isl_dim_param
, pos
);
2676 nested
= is_nested_parameter(id
);
2682 /* Does "space" involve any parameters that refer to nested
2683 * accesses, i.e., parameters with no name?
2685 static bool has_nested(__isl_keep isl_space
*space
)
2689 nparam
= isl_space_dim(space
, isl_dim_param
);
2690 for (int i
= 0; i
< nparam
; ++i
)
2691 if (is_nested_parameter(space
, i
))
2697 /* Does "pa" involve any parameters that refer to nested
2698 * accesses, i.e., parameters with no name?
2700 static bool has_nested(__isl_keep isl_pw_aff
*pa
)
2705 space
= isl_pw_aff_get_space(pa
);
2706 nested
= has_nested(space
);
2707 isl_space_free(space
);
2712 /* Construct a pet_scop for a for statement.
2713 * The for loop is required to be of the form
2715 * for (i = init; condition; ++i)
2719 * for (i = init; condition; --i)
2721 * The initialization of the for loop should either be an assignment
2722 * to an integer variable, or a declaration of such a variable with
2725 * The condition is allowed to contain nested accesses, provided
2726 * they are not being written to inside the body of the loop.
2727 * Otherwise, or if the condition is otherwise non-affine, the for loop is
2728 * essentially treated as a while loop, with iteration domain
2729 * { [i] : i >= init }.
2731 * We extract a pet_scop for the body and then embed it in a loop with
2732 * iteration domain and schedule
2734 * { [i] : i >= init and condition' }
2739 * { [i] : i <= init and condition' }
2742 * Where condition' is equal to condition if the latter is
2743 * a simple upper [lower] bound and a condition that is extended
2744 * to apply to all previous iterations otherwise.
2746 * If the condition is non-affine, then we drop the condition from the
2747 * iteration domain and instead create a separate statement
2748 * for evaluating the condition. The body is then filtered to depend
2749 * on the result of the condition evaluating to true on all iterations
2750 * up to the current iteration, while the evaluation the condition itself
2751 * is filtered to depend on the result of the condition evaluating to true
2752 * on all previous iterations.
2753 * The context of the scop representing the body is dropped
2754 * because we don't know how many times the body will be executed,
2757 * If the stride of the loop is not 1, then "i >= init" is replaced by
2759 * (exists a: i = init + stride * a and a >= 0)
2761 * If the loop iterator i is unsigned, then wrapping may occur.
2762 * During the computation, we work with a virtual iterator that
2763 * does not wrap. However, the condition in the code applies
2764 * to the wrapped value, so we need to change condition(i)
2765 * into condition([i % 2^width]).
2766 * After computing the virtual domain and schedule, we apply
2767 * the function { [v] -> [v % 2^width] } to the domain and the domain
2768 * of the schedule. In order not to lose any information, we also
2769 * need to intersect the domain of the schedule with the virtual domain
2770 * first, since some iterations in the wrapped domain may be scheduled
2771 * several times, typically an infinite number of times.
2772 * Note that there may be no need to perform this final wrapping
2773 * if the loop condition (after wrapping) satisfies certain conditions.
2774 * However, the is_simple_bound condition is not enough since it doesn't
2775 * check if there even is an upper bound.
2777 * If the loop condition is non-affine, then we keep the virtual
2778 * iterator in the iteration domain and instead replace all accesses
2779 * to the original iterator by the wrapping of the virtual iterator.
2781 * Wrapping on unsigned iterators can be avoided entirely if
2782 * loop condition is simple, the loop iterator is incremented
2783 * [decremented] by one and the last value before wrapping cannot
2784 * possibly satisfy the loop condition.
2786 * Before extracting a pet_scop from the body we remove all
2787 * assignments in assigned_value to variables that are assigned
2788 * somewhere in the body of the loop.
2790 * Valid parameters for a for loop are those for which the initial
2791 * value itself, the increment on each domain iteration and
2792 * the condition on both the initial value and
2793 * the result of incrementing the iterator for each iteration of the domain
2795 * If the loop condition is non-affine, then we only consider validity
2796 * of the initial value.
2798 * If the body contains any break, then we keep track of it in "skip"
2799 * (if the skip condition is affine) or it is handled in scop_add_break
2800 * (if the skip condition is not affine).
2801 * Note that the affine break condition needs to be considered with
2802 * respect to previous iterations in the virtual domain (if any)
2803 * and that the domain needs to be kept virtual if there is a non-affine
2806 struct pet_scop
*PetScan::extract_for(ForStmt
*stmt
)
2808 BinaryOperator
*ass
;
2816 isl_set
*cond
= NULL
;
2817 isl_set
*skip
= NULL
;
2818 isl_id
*id
, *id_test
= NULL
, *id_break_test
;
2819 struct pet_scop
*scop
, *scop_cond
= NULL
;
2820 assigned_value_cache
cache(assigned_value
);
2826 bool keep_virtual
= false;
2827 bool has_affine_break
;
2829 isl_aff
*wrap
= NULL
;
2830 isl_pw_aff
*pa
, *pa_inc
, *init_val
;
2831 isl_set
*valid_init
;
2832 isl_set
*valid_cond
;
2833 isl_set
*valid_cond_init
;
2834 isl_set
*valid_cond_next
;
2838 if (!stmt
->getInit() && !stmt
->getCond() && !stmt
->getInc())
2839 return extract_infinite_for(stmt
);
2841 init
= stmt
->getInit();
2846 if ((ass
= initialization_assignment(init
)) != NULL
) {
2847 iv
= extract_induction_variable(ass
);
2850 lhs
= ass
->getLHS();
2851 rhs
= ass
->getRHS();
2852 } else if ((decl
= initialization_declaration(init
)) != NULL
) {
2853 VarDecl
*var
= extract_induction_variable(init
, decl
);
2857 rhs
= var
->getInit();
2858 lhs
= create_DeclRefExpr(var
);
2860 unsupported(stmt
->getInit());
2864 pa_inc
= extract_increment(stmt
, iv
);
2869 if (isl_pw_aff_n_piece(pa_inc
) != 1 ||
2870 isl_pw_aff_foreach_piece(pa_inc
, &extract_cst
, &inc
) < 0) {
2871 isl_pw_aff_free(pa_inc
);
2872 unsupported(stmt
->getInc());
2876 valid_inc
= isl_pw_aff_domain(pa_inc
);
2878 is_unsigned
= iv
->getType()->isUnsignedIntegerType();
2880 assigned_value
.erase(iv
);
2881 clear_assignments
clear(assigned_value
);
2882 clear
.TraverseStmt(stmt
->getBody());
2884 id
= isl_id_alloc(ctx
, iv
->getName().str().c_str(), iv
);
2886 pa
= try_extract_nested_condition(stmt
->getCond());
2887 if (allow_nested
&& (!pa
|| has_nested(pa
)))
2890 scop
= extract(stmt
->getBody());
2892 has_affine_break
= scop
&&
2893 pet_scop_has_affine_skip(scop
, pet_skip_later
);
2894 if (has_affine_break
)
2895 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_later
);
2896 has_var_break
= scop
&& pet_scop_has_var_skip(scop
, pet_skip_later
);
2897 if (has_var_break
) {
2898 id_break_test
= pet_scop_get_skip_id(scop
, pet_skip_later
);
2899 keep_virtual
= true;
2902 if (pa
&& !is_nested_allowed(pa
, scop
)) {
2903 isl_pw_aff_free(pa
);
2907 if (!allow_nested
&& !pa
)
2908 pa
= try_extract_affine_condition(stmt
->getCond());
2909 valid_cond
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
2910 cond
= isl_pw_aff_non_zero_set(pa
);
2911 if (allow_nested
&& !cond
) {
2912 isl_map
*test_access
;
2913 int save_n_stmt
= n_stmt
;
2914 test_access
= create_test_access(ctx
, n_test
++);
2916 scop_cond
= extract_non_affine_condition(stmt
->getCond(),
2917 isl_map_copy(test_access
));
2918 n_stmt
= save_n_stmt
;
2919 scop_cond
= scop_add_array(scop_cond
, test_access
, ast_context
);
2920 id_test
= isl_map_get_tuple_id(test_access
, isl_dim_out
);
2921 isl_map_free(test_access
);
2922 scop_cond
= pet_scop_prefix(scop_cond
, 0);
2923 scop
= pet_scop_reset_context(scop
);
2924 scop
= pet_scop_prefix(scop
, 1);
2925 keep_virtual
= true;
2926 cond
= isl_set_universe(isl_space_set_alloc(ctx
, 0, 0));
2929 cond
= embed(cond
, isl_id_copy(id
));
2930 skip
= embed(skip
, isl_id_copy(id
));
2931 valid_cond
= isl_set_coalesce(valid_cond
);
2932 valid_cond
= embed(valid_cond
, isl_id_copy(id
));
2933 valid_inc
= embed(valid_inc
, isl_id_copy(id
));
2934 is_one
= isl_val_is_one(inc
) || isl_val_is_negone(inc
);
2935 is_virtual
= is_unsigned
&& (!is_one
|| can_wrap(cond
, iv
, inc
));
2937 init_val
= extract_affine(rhs
);
2938 valid_cond_init
= enforce_subset(
2939 isl_set_from_pw_aff(isl_pw_aff_copy(init_val
)),
2940 isl_set_copy(valid_cond
));
2941 if (is_one
&& !is_virtual
) {
2942 isl_pw_aff_free(init_val
);
2943 pa
= extract_comparison(isl_val_is_pos(inc
) ? BO_GE
: BO_LE
,
2945 valid_init
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
2946 valid_init
= set_project_out_by_id(valid_init
, id
);
2947 domain
= isl_pw_aff_non_zero_set(pa
);
2949 valid_init
= isl_pw_aff_domain(isl_pw_aff_copy(init_val
));
2950 domain
= strided_domain(isl_id_copy(id
), init_val
,
2954 domain
= embed(domain
, isl_id_copy(id
));
2957 wrap
= compute_wrapping(isl_set_get_space(cond
), iv
);
2958 rev_wrap
= isl_map_from_aff(isl_aff_copy(wrap
));
2959 rev_wrap
= isl_map_reverse(rev_wrap
);
2960 cond
= isl_set_apply(cond
, isl_map_copy(rev_wrap
));
2961 skip
= isl_set_apply(skip
, isl_map_copy(rev_wrap
));
2962 valid_cond
= isl_set_apply(valid_cond
, isl_map_copy(rev_wrap
));
2963 valid_inc
= isl_set_apply(valid_inc
, rev_wrap
);
2965 is_simple
= is_simple_bound(cond
, inc
);
2967 cond
= isl_set_gist(cond
, isl_set_copy(domain
));
2968 is_simple
= is_simple_bound(cond
, inc
);
2971 cond
= valid_for_each_iteration(cond
,
2972 isl_set_copy(domain
), isl_val_copy(inc
));
2973 domain
= isl_set_intersect(domain
, cond
);
2974 if (has_affine_break
) {
2975 skip
= isl_set_intersect(skip
, isl_set_copy(domain
));
2976 skip
= after(skip
, isl_val_sgn(inc
));
2977 domain
= isl_set_subtract(domain
, skip
);
2979 domain
= isl_set_set_dim_id(domain
, isl_dim_set
, 0, isl_id_copy(id
));
2980 space
= isl_space_from_domain(isl_set_get_space(domain
));
2981 space
= isl_space_add_dims(space
, isl_dim_out
, 1);
2982 sched
= isl_map_universe(space
);
2983 if (isl_val_is_pos(inc
))
2984 sched
= isl_map_equate(sched
, isl_dim_in
, 0, isl_dim_out
, 0);
2986 sched
= isl_map_oppose(sched
, isl_dim_in
, 0, isl_dim_out
, 0);
2988 valid_cond_next
= valid_on_next(valid_cond
, isl_set_copy(domain
),
2990 valid_inc
= enforce_subset(isl_set_copy(domain
), valid_inc
);
2992 if (is_virtual
&& !keep_virtual
) {
2993 isl_map
*wrap_map
= isl_map_from_aff(wrap
);
2994 wrap_map
= isl_map_set_dim_id(wrap_map
,
2995 isl_dim_out
, 0, isl_id_copy(id
));
2996 sched
= isl_map_intersect_domain(sched
, isl_set_copy(domain
));
2997 domain
= isl_set_apply(domain
, isl_map_copy(wrap_map
));
2998 sched
= isl_map_apply_domain(sched
, wrap_map
);
3000 if (!(is_virtual
&& keep_virtual
))
3001 wrap
= identity_aff(domain
);
3003 scop_cond
= pet_scop_embed(scop_cond
, isl_set_copy(domain
),
3004 isl_map_copy(sched
), isl_aff_copy(wrap
), isl_id_copy(id
));
3005 scop
= pet_scop_embed(scop
, isl_set_copy(domain
), sched
, wrap
, id
);
3006 scop
= resolve_nested(scop
);
3008 scop
= scop_add_break(scop
, id_break_test
, isl_set_copy(domain
),
3011 scop
= scop_add_while(scop_cond
, scop
, id_test
, domain
,
3013 isl_set_free(valid_inc
);
3015 scop
= pet_scop_restrict_context(scop
, valid_inc
);
3016 scop
= pet_scop_restrict_context(scop
, valid_cond_next
);
3017 scop
= pet_scop_restrict_context(scop
, valid_cond_init
);
3018 isl_set_free(domain
);
3020 clear_assignment(assigned_value
, iv
);
3024 scop
= pet_scop_restrict_context(scop
, valid_init
);
3029 struct pet_scop
*PetScan::extract(CompoundStmt
*stmt
, bool skip_declarations
)
3031 return extract(stmt
->children(), true, skip_declarations
);
3034 /* Does parameter "pos" of "map" refer to a nested access?
3036 static bool is_nested_parameter(__isl_keep isl_map
*map
, int pos
)
3041 id
= isl_map_get_dim_id(map
, isl_dim_param
, pos
);
3042 nested
= is_nested_parameter(id
);
3048 /* How many parameters of "space" refer to nested accesses, i.e., have no name?
3050 static int n_nested_parameter(__isl_keep isl_space
*space
)
3055 nparam
= isl_space_dim(space
, isl_dim_param
);
3056 for (int i
= 0; i
< nparam
; ++i
)
3057 if (is_nested_parameter(space
, i
))
3063 /* How many parameters of "map" refer to nested accesses, i.e., have no name?
3065 static int n_nested_parameter(__isl_keep isl_map
*map
)
3070 space
= isl_map_get_space(map
);
3071 n
= n_nested_parameter(space
);
3072 isl_space_free(space
);
3077 /* For each nested access parameter in "space",
3078 * construct a corresponding pet_expr, place it in args and
3079 * record its position in "param2pos".
3080 * "n_arg" is the number of elements that are already in args.
3081 * The position recorded in "param2pos" takes this number into account.
3082 * If the pet_expr corresponding to a parameter is identical to
3083 * the pet_expr corresponding to an earlier parameter, then these two
3084 * parameters are made to refer to the same element in args.
3086 * Return the final number of elements in args or -1 if an error has occurred.
3088 int PetScan::extract_nested(__isl_keep isl_space
*space
,
3089 int n_arg
, struct pet_expr
**args
, std::map
<int,int> ¶m2pos
)
3093 nparam
= isl_space_dim(space
, isl_dim_param
);
3094 for (int i
= 0; i
< nparam
; ++i
) {
3096 isl_id
*id
= isl_space_get_dim_id(space
, isl_dim_param
, i
);
3099 if (!is_nested_parameter(id
)) {
3104 nested
= (Expr
*) isl_id_get_user(id
);
3105 args
[n_arg
] = extract_expr(nested
);
3109 for (j
= 0; j
< n_arg
; ++j
)
3110 if (pet_expr_is_equal(args
[j
], args
[n_arg
]))
3114 pet_expr_free(args
[n_arg
]);
3118 param2pos
[i
] = n_arg
++;
3126 /* For each nested access parameter in the access relations in "expr",
3127 * construct a corresponding pet_expr, place it in expr->args and
3128 * record its position in "param2pos".
3129 * n is the number of nested access parameters.
3131 struct pet_expr
*PetScan::extract_nested(struct pet_expr
*expr
, int n
,
3132 std::map
<int,int> ¶m2pos
)
3136 expr
->args
= isl_calloc_array(ctx
, struct pet_expr
*, n
);
3141 space
= isl_map_get_space(expr
->acc
.access
);
3142 n
= extract_nested(space
, 0, expr
->args
, param2pos
);
3143 isl_space_free(space
);
3151 pet_expr_free(expr
);
3155 /* Look for parameters in any access relation in "expr" that
3156 * refer to nested accesses. In particular, these are
3157 * parameters with no name.
3159 * If there are any such parameters, then the domain of the access
3160 * relation, which is still [] at this point, is replaced by
3161 * [[] -> [t_1,...,t_n]], with n the number of these parameters
3162 * (after identifying identical nested accesses).
3163 * The parameters are then equated to the corresponding t dimensions
3164 * and subsequently projected out.
3165 * param2pos maps the position of the parameter to the position
3166 * of the corresponding t dimension.
3168 struct pet_expr
*PetScan::resolve_nested(struct pet_expr
*expr
)
3175 std::map
<int,int> param2pos
;
3180 for (int i
= 0; i
< expr
->n_arg
; ++i
) {
3181 expr
->args
[i
] = resolve_nested(expr
->args
[i
]);
3182 if (!expr
->args
[i
]) {
3183 pet_expr_free(expr
);
3188 if (expr
->type
!= pet_expr_access
)
3191 n
= n_nested_parameter(expr
->acc
.access
);
3195 expr
= extract_nested(expr
, n
, param2pos
);
3200 nparam
= isl_map_dim(expr
->acc
.access
, isl_dim_param
);
3201 n_in
= isl_map_dim(expr
->acc
.access
, isl_dim_in
);
3202 dim
= isl_map_get_space(expr
->acc
.access
);
3203 dim
= isl_space_domain(dim
);
3204 dim
= isl_space_from_domain(dim
);
3205 dim
= isl_space_add_dims(dim
, isl_dim_out
, n
);
3206 map
= isl_map_universe(dim
);
3207 map
= isl_map_domain_map(map
);
3208 map
= isl_map_reverse(map
);
3209 expr
->acc
.access
= isl_map_apply_domain(expr
->acc
.access
, map
);
3211 for (int i
= nparam
- 1; i
>= 0; --i
) {
3212 isl_id
*id
= isl_map_get_dim_id(expr
->acc
.access
,
3214 if (!is_nested_parameter(id
)) {
3219 expr
->acc
.access
= isl_map_equate(expr
->acc
.access
,
3220 isl_dim_param
, i
, isl_dim_in
,
3221 n_in
+ param2pos
[i
]);
3222 expr
->acc
.access
= isl_map_project_out(expr
->acc
.access
,
3223 isl_dim_param
, i
, 1);
3230 pet_expr_free(expr
);
3234 /* Return the file offset of the expansion location of "Loc".
3236 static unsigned getExpansionOffset(SourceManager
&SM
, SourceLocation Loc
)
3238 return SM
.getFileOffset(SM
.getExpansionLoc(Loc
));
3241 #ifdef HAVE_FINDLOCATIONAFTERTOKEN
3243 /* Return a SourceLocation for the location after the first semicolon
3244 * after "loc". If Lexer::findLocationAfterToken is available, we simply
3245 * call it and also skip trailing spaces and newline.
3247 static SourceLocation
location_after_semi(SourceLocation loc
, SourceManager
&SM
,
3248 const LangOptions
&LO
)
3250 return Lexer::findLocationAfterToken(loc
, tok::semi
, SM
, LO
, true);
3255 /* Return a SourceLocation for the location after the first semicolon
3256 * after "loc". If Lexer::findLocationAfterToken is not available,
3257 * we look in the underlying character data for the first semicolon.
3259 static SourceLocation
location_after_semi(SourceLocation loc
, SourceManager
&SM
,
3260 const LangOptions
&LO
)
3263 const char *s
= SM
.getCharacterData(loc
);
3265 semi
= strchr(s
, ';');
3267 return SourceLocation();
3268 return loc
.getFileLocWithOffset(semi
+ 1 - s
);
3273 /* If the token at "loc" is the first token on the line, then return
3274 * a location referring to the start of the line.
3275 * Otherwise, return "loc".
3277 * This function is used to extend a scop to the start of the line
3278 * if the first token of the scop is also the first token on the line.
3280 * We look for the first token on the line. If its location is equal to "loc",
3281 * then the latter is the location of the first token on the line.
3283 static SourceLocation
move_to_start_of_line_if_first_token(SourceLocation loc
,
3284 SourceManager
&SM
, const LangOptions
&LO
)
3286 std::pair
<FileID
, unsigned> file_offset_pair
;
3287 llvm::StringRef file
;
3290 SourceLocation token_loc
, line_loc
;
3293 loc
= SM
.getExpansionLoc(loc
);
3294 col
= SM
.getExpansionColumnNumber(loc
);
3295 line_loc
= loc
.getLocWithOffset(1 - col
);
3296 file_offset_pair
= SM
.getDecomposedLoc(line_loc
);
3297 file
= SM
.getBufferData(file_offset_pair
.first
, NULL
);
3298 pos
= file
.data() + file_offset_pair
.second
;
3300 Lexer
lexer(SM
.getLocForStartOfFile(file_offset_pair
.first
), LO
,
3301 file
.begin(), pos
, file
.end());
3302 lexer
.LexFromRawLexer(tok
);
3303 token_loc
= tok
.getLocation();
3305 if (token_loc
== loc
)
3311 /* Convert a top-level pet_expr to a pet_scop with one statement.
3312 * This mainly involves resolving nested expression parameters
3313 * and setting the name of the iteration space.
3314 * The name is given by "label" if it is non-NULL. Otherwise,
3315 * it is of the form S_<n_stmt>.
3316 * start and end of the pet_scop are derived from those of "stmt".
3318 struct pet_scop
*PetScan::extract(Stmt
*stmt
, struct pet_expr
*expr
,
3319 __isl_take isl_id
*label
)
3321 struct pet_stmt
*ps
;
3322 struct pet_scop
*scop
;
3323 SourceLocation loc
= stmt
->getLocStart();
3324 SourceManager
&SM
= PP
.getSourceManager();
3325 const LangOptions
&LO
= PP
.getLangOpts();
3326 int line
= PP
.getSourceManager().getExpansionLineNumber(loc
);
3327 unsigned start
, end
;
3329 expr
= resolve_nested(expr
);
3330 ps
= pet_stmt_from_pet_expr(ctx
, line
, label
, n_stmt
++, expr
);
3331 scop
= pet_scop_from_pet_stmt(ctx
, ps
);
3333 loc
= move_to_start_of_line_if_first_token(loc
, SM
, LO
);
3334 start
= getExpansionOffset(SM
, loc
);
3335 loc
= stmt
->getLocEnd();
3336 loc
= location_after_semi(loc
, SM
, LO
);
3337 end
= getExpansionOffset(SM
, loc
);
3339 scop
= pet_scop_update_start_end(scop
, start
, end
);
3343 /* Check if we can extract an affine expression from "expr".
3344 * Return the expressions as an isl_pw_aff if we can and NULL otherwise.
3345 * We turn on autodetection so that we won't generate any warnings
3346 * and turn off nesting, so that we won't accept any non-affine constructs.
3348 __isl_give isl_pw_aff
*PetScan::try_extract_affine(Expr
*expr
)
3351 int save_autodetect
= options
->autodetect
;
3352 bool save_nesting
= nesting_enabled
;
3354 options
->autodetect
= 1;
3355 nesting_enabled
= false;
3357 pwaff
= extract_affine(expr
);
3359 options
->autodetect
= save_autodetect
;
3360 nesting_enabled
= save_nesting
;
3365 /* Check whether "expr" is an affine expression.
3367 bool PetScan::is_affine(Expr
*expr
)
3371 pwaff
= try_extract_affine(expr
);
3372 isl_pw_aff_free(pwaff
);
3374 return pwaff
!= NULL
;
3377 /* Check if we can extract an affine constraint from "expr".
3378 * Return the constraint as an isl_set if we can and NULL otherwise.
3379 * We turn on autodetection so that we won't generate any warnings
3380 * and turn off nesting, so that we won't accept any non-affine constructs.
3382 __isl_give isl_pw_aff
*PetScan::try_extract_affine_condition(Expr
*expr
)
3385 int save_autodetect
= options
->autodetect
;
3386 bool save_nesting
= nesting_enabled
;
3388 options
->autodetect
= 1;
3389 nesting_enabled
= false;
3391 cond
= extract_condition(expr
);
3393 options
->autodetect
= save_autodetect
;
3394 nesting_enabled
= save_nesting
;
3399 /* Check whether "expr" is an affine constraint.
3401 bool PetScan::is_affine_condition(Expr
*expr
)
3405 cond
= try_extract_affine_condition(expr
);
3406 isl_pw_aff_free(cond
);
3408 return cond
!= NULL
;
3411 /* Check if we can extract a condition from "expr".
3412 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
3413 * If allow_nested is set, then the condition may involve parameters
3414 * corresponding to nested accesses.
3415 * We turn on autodetection so that we won't generate any warnings.
3417 __isl_give isl_pw_aff
*PetScan::try_extract_nested_condition(Expr
*expr
)
3420 int save_autodetect
= options
->autodetect
;
3421 bool save_nesting
= nesting_enabled
;
3423 options
->autodetect
= 1;
3424 nesting_enabled
= allow_nested
;
3425 cond
= extract_condition(expr
);
3427 options
->autodetect
= save_autodetect
;
3428 nesting_enabled
= save_nesting
;
3433 /* If the top-level expression of "stmt" is an assignment, then
3434 * return that assignment as a BinaryOperator.
3435 * Otherwise return NULL.
3437 static BinaryOperator
*top_assignment_or_null(Stmt
*stmt
)
3439 BinaryOperator
*ass
;
3443 if (stmt
->getStmtClass() != Stmt::BinaryOperatorClass
)
3446 ass
= cast
<BinaryOperator
>(stmt
);
3447 if(ass
->getOpcode() != BO_Assign
)
3453 /* Check if the given if statement is a conditional assignement
3454 * with a non-affine condition. If so, construct a pet_scop
3455 * corresponding to this conditional assignment. Otherwise return NULL.
3457 * In particular we check if "stmt" is of the form
3464 * where a is some array or scalar access.
3465 * The constructed pet_scop then corresponds to the expression
3467 * a = condition ? f(...) : g(...)
3469 * All access relations in f(...) are intersected with condition
3470 * while all access relation in g(...) are intersected with the complement.
3472 struct pet_scop
*PetScan::extract_conditional_assignment(IfStmt
*stmt
)
3474 BinaryOperator
*ass_then
, *ass_else
;
3475 isl_multi_pw_aff
*write_then
, *write_else
;
3476 isl_set
*cond
, *comp
;
3477 isl_multi_pw_aff
*index
;
3480 struct pet_expr
*pe_cond
, *pe_then
, *pe_else
, *pe
, *pe_write
;
3481 bool save_nesting
= nesting_enabled
;
3483 if (!options
->detect_conditional_assignment
)
3486 ass_then
= top_assignment_or_null(stmt
->getThen());
3487 ass_else
= top_assignment_or_null(stmt
->getElse());
3489 if (!ass_then
|| !ass_else
)
3492 if (is_affine_condition(stmt
->getCond()))
3495 write_then
= extract_index(ass_then
->getLHS());
3496 write_else
= extract_index(ass_else
->getLHS());
3498 equal
= isl_multi_pw_aff_plain_is_equal(write_then
, write_else
);
3499 isl_multi_pw_aff_free(write_else
);
3500 if (equal
< 0 || !equal
) {
3501 isl_multi_pw_aff_free(write_then
);
3505 nesting_enabled
= allow_nested
;
3506 pa
= extract_condition(stmt
->getCond());
3507 nesting_enabled
= save_nesting
;
3508 cond
= isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa
));
3509 comp
= isl_pw_aff_zero_set(isl_pw_aff_copy(pa
));
3510 index
= isl_multi_pw_aff_from_range(isl_multi_pw_aff_from_pw_aff(pa
));
3512 pe_cond
= pet_expr_from_index(index
);
3514 pe_then
= extract_expr(ass_then
->getRHS());
3515 pe_then
= pet_expr_restrict(pe_then
, cond
);
3516 pe_else
= extract_expr(ass_else
->getRHS());
3517 pe_else
= pet_expr_restrict(pe_else
, comp
);
3519 pe
= pet_expr_new_ternary(ctx
, pe_cond
, pe_then
, pe_else
);
3520 pe_write
= pet_expr_from_index_and_depth(write_then
,
3521 extract_depth(write_then
));
3523 pe_write
->acc
.write
= 1;
3524 pe_write
->acc
.read
= 0;
3526 pe
= pet_expr_new_binary(ctx
, pet_op_assign
, pe_write
, pe
);
3527 return extract(stmt
, pe
);
3530 /* Create a pet_scop with a single statement evaluating "cond"
3531 * and writing the result to a virtual scalar, as expressed by
3534 struct pet_scop
*PetScan::extract_non_affine_condition(Expr
*cond
,
3535 __isl_take isl_map
*access
)
3537 struct pet_expr
*expr
, *write
;
3538 struct pet_stmt
*ps
;
3539 struct pet_scop
*scop
;
3540 SourceLocation loc
= cond
->getLocStart();
3541 int line
= PP
.getSourceManager().getExpansionLineNumber(loc
);
3543 write
= pet_expr_from_access(access
);
3545 write
->acc
.write
= 1;
3546 write
->acc
.read
= 0;
3548 expr
= extract_expr(cond
);
3549 expr
= resolve_nested(expr
);
3550 expr
= pet_expr_new_binary(ctx
, pet_op_assign
, write
, expr
);
3551 ps
= pet_stmt_from_pet_expr(ctx
, line
, NULL
, n_stmt
++, expr
);
3552 scop
= pet_scop_from_pet_stmt(ctx
, ps
);
3553 scop
= resolve_nested(scop
);
3559 static struct pet_expr
*embed_access(struct pet_expr
*expr
, void *user
);
3562 /* Apply the map pointed to by "user" to the domain of the access
3563 * relation associated to "expr", thereby embedding it in the range of the map.
3564 * The domain of both relations is the zero-dimensional domain.
3566 static struct pet_expr
*embed_access(struct pet_expr
*expr
, void *user
)
3568 isl_map
*map
= (isl_map
*) user
;
3570 expr
->acc
.access
= isl_map_apply_domain(expr
->acc
.access
,
3572 if (!expr
->acc
.access
)
3577 pet_expr_free(expr
);
3581 /* Apply "map" to all access relations in "expr".
3583 static struct pet_expr
*embed(struct pet_expr
*expr
, __isl_keep isl_map
*map
)
3585 return pet_expr_map_access(expr
, &embed_access
, map
);
3588 /* How many parameters of "set" refer to nested accesses, i.e., have no name?
3590 static int n_nested_parameter(__isl_keep isl_set
*set
)
3595 space
= isl_set_get_space(set
);
3596 n
= n_nested_parameter(space
);
3597 isl_space_free(space
);
3602 /* Remove all parameters from "map" that refer to nested accesses.
3604 static __isl_give isl_map
*remove_nested_parameters(__isl_take isl_map
*map
)
3609 space
= isl_map_get_space(map
);
3610 nparam
= isl_space_dim(space
, isl_dim_param
);
3611 for (int i
= nparam
- 1; i
>= 0; --i
)
3612 if (is_nested_parameter(space
, i
))
3613 map
= isl_map_project_out(map
, isl_dim_param
, i
, 1);
3614 isl_space_free(space
);
3619 /* Remove all parameters from the access relation of "expr"
3620 * that refer to nested accesses.
3622 static struct pet_expr
*remove_nested_parameters(struct pet_expr
*expr
)
3624 expr
->acc
.access
= remove_nested_parameters(expr
->acc
.access
);
3625 if (!expr
->acc
.access
)
3630 pet_expr_free(expr
);
3635 static struct pet_expr
*expr_remove_nested_parameters(
3636 struct pet_expr
*expr
, void *user
);
3639 static struct pet_expr
*expr_remove_nested_parameters(
3640 struct pet_expr
*expr
, void *user
)
3642 return remove_nested_parameters(expr
);
3645 /* Remove all nested access parameters from the schedule and all
3646 * accesses of "stmt".
3647 * There is no need to remove them from the domain as these parameters
3648 * have already been removed from the domain when this function is called.
3650 static struct pet_stmt
*remove_nested_parameters(struct pet_stmt
*stmt
)
3654 stmt
->schedule
= remove_nested_parameters(stmt
->schedule
);
3655 stmt
->body
= pet_expr_map_access(stmt
->body
,
3656 &expr_remove_nested_parameters
, NULL
);
3657 if (!stmt
->schedule
|| !stmt
->body
)
3659 for (int i
= 0; i
< stmt
->n_arg
; ++i
) {
3660 stmt
->args
[i
] = pet_expr_map_access(stmt
->args
[i
],
3661 &expr_remove_nested_parameters
, NULL
);
3668 pet_stmt_free(stmt
);
3672 /* For each nested access parameter in the domain of "stmt",
3673 * construct a corresponding pet_expr, place it before the original
3674 * elements in stmt->args and record its position in "param2pos".
3675 * n is the number of nested access parameters.
3677 struct pet_stmt
*PetScan::extract_nested(struct pet_stmt
*stmt
, int n
,
3678 std::map
<int,int> ¶m2pos
)
3683 struct pet_expr
**args
;
3685 n_arg
= stmt
->n_arg
;
3686 args
= isl_calloc_array(ctx
, struct pet_expr
*, n
+ n_arg
);
3690 space
= isl_set_get_space(stmt
->domain
);
3691 n_arg
= extract_nested(space
, 0, args
, param2pos
);
3692 isl_space_free(space
);
3697 for (i
= 0; i
< stmt
->n_arg
; ++i
)
3698 args
[n_arg
+ i
] = stmt
->args
[i
];
3701 stmt
->n_arg
+= n_arg
;
3706 for (i
= 0; i
< n
; ++i
)
3707 pet_expr_free(args
[i
]);
3710 pet_stmt_free(stmt
);
3714 /* Check whether any of the arguments i of "stmt" starting at position "n"
3715 * is equal to one of the first "n" arguments j.
3716 * If so, combine the constraints on arguments i and j and remove
3719 static struct pet_stmt
*remove_duplicate_arguments(struct pet_stmt
*stmt
, int n
)
3728 if (n
== stmt
->n_arg
)
3731 map
= isl_set_unwrap(stmt
->domain
);
3733 for (i
= stmt
->n_arg
- 1; i
>= n
; --i
) {
3734 for (j
= 0; j
< n
; ++j
)
3735 if (pet_expr_is_equal(stmt
->args
[i
], stmt
->args
[j
]))
3740 map
= isl_map_equate(map
, isl_dim_out
, i
, isl_dim_out
, j
);
3741 map
= isl_map_project_out(map
, isl_dim_out
, i
, 1);
3743 pet_expr_free(stmt
->args
[i
]);
3744 for (j
= i
; j
+ 1 < stmt
->n_arg
; ++j
)
3745 stmt
->args
[j
] = stmt
->args
[j
+ 1];
3749 stmt
->domain
= isl_map_wrap(map
);
3754 pet_stmt_free(stmt
);
3758 /* Look for parameters in the iteration domain of "stmt" that
3759 * refer to nested accesses. In particular, these are
3760 * parameters with no name.
3762 * If there are any such parameters, then as many extra variables
3763 * (after identifying identical nested accesses) are inserted in the
3764 * range of the map wrapped inside the domain, before the original variables.
3765 * If the original domain is not a wrapped map, then a new wrapped
3766 * map is created with zero output dimensions.
3767 * The parameters are then equated to the corresponding output dimensions
3768 * and subsequently projected out, from the iteration domain,
3769 * the schedule and the access relations.
3770 * For each of the output dimensions, a corresponding argument
3771 * expression is inserted. Initially they are created with
3772 * a zero-dimensional domain, so they have to be embedded
3773 * in the current iteration domain.
3774 * param2pos maps the position of the parameter to the position
3775 * of the corresponding output dimension in the wrapped map.
3777 struct pet_stmt
*PetScan::resolve_nested(struct pet_stmt
*stmt
)
3783 std::map
<int,int> param2pos
;
3788 n
= n_nested_parameter(stmt
->domain
);
3792 n_arg
= stmt
->n_arg
;
3793 stmt
= extract_nested(stmt
, n
, param2pos
);
3797 n
= stmt
->n_arg
- n_arg
;
3798 nparam
= isl_set_dim(stmt
->domain
, isl_dim_param
);
3799 if (isl_set_is_wrapping(stmt
->domain
))
3800 map
= isl_set_unwrap(stmt
->domain
);
3802 map
= isl_map_from_domain(stmt
->domain
);
3803 map
= isl_map_insert_dims(map
, isl_dim_out
, 0, n
);
3805 for (int i
= nparam
- 1; i
>= 0; --i
) {
3808 if (!is_nested_parameter(map
, i
))
3811 id
= pet_expr_access_get_id(stmt
->args
[param2pos
[i
]]);
3812 map
= isl_map_set_dim_id(map
, isl_dim_out
, param2pos
[i
], id
);
3813 map
= isl_map_equate(map
, isl_dim_param
, i
, isl_dim_out
,
3815 map
= isl_map_project_out(map
, isl_dim_param
, i
, 1);
3818 stmt
->domain
= isl_map_wrap(map
);
3820 map
= isl_set_unwrap(isl_set_copy(stmt
->domain
));
3821 map
= isl_map_from_range(isl_map_domain(map
));
3822 for (int pos
= 0; pos
< n
; ++pos
)
3823 stmt
->args
[pos
] = embed(stmt
->args
[pos
], map
);
3826 stmt
= remove_nested_parameters(stmt
);
3827 stmt
= remove_duplicate_arguments(stmt
, n
);
3831 pet_stmt_free(stmt
);
3835 /* For each statement in "scop", move the parameters that correspond
3836 * to nested access into the ranges of the domains and create
3837 * corresponding argument expressions.
3839 struct pet_scop
*PetScan::resolve_nested(struct pet_scop
*scop
)
3844 for (int i
= 0; i
< scop
->n_stmt
; ++i
) {
3845 scop
->stmts
[i
] = resolve_nested(scop
->stmts
[i
]);
3846 if (!scop
->stmts
[i
])
3852 pet_scop_free(scop
);
3856 /* Given an access expression "expr", is the variable accessed by
3857 * "expr" assigned anywhere inside "scop"?
3859 static bool is_assigned(pet_expr
*expr
, pet_scop
*scop
)
3861 bool assigned
= false;
3864 id
= pet_expr_access_get_id(expr
);
3865 assigned
= pet_scop_writes(scop
, id
);
3871 /* Are all nested access parameters in "pa" allowed given "scop".
3872 * In particular, is none of them written by anywhere inside "scop".
3874 * If "scop" has any skip conditions, then no nested access parameters
3875 * are allowed. In particular, if there is any nested access in a guard
3876 * for a piece of code containing a "continue", then we want to introduce
3877 * a separate statement for evaluating this guard so that we can express
3878 * that the result is false for all previous iterations.
3880 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff
*pa
, pet_scop
*scop
)
3887 nparam
= isl_pw_aff_dim(pa
, isl_dim_param
);
3888 for (int i
= 0; i
< nparam
; ++i
) {
3890 isl_id
*id
= isl_pw_aff_get_dim_id(pa
, isl_dim_param
, i
);
3894 if (!is_nested_parameter(id
)) {
3899 if (pet_scop_has_skip(scop
, pet_skip_now
)) {
3904 nested
= (Expr
*) isl_id_get_user(id
);
3905 expr
= extract_expr(nested
);
3906 allowed
= expr
&& expr
->type
== pet_expr_access
&&
3907 !is_assigned(expr
, scop
);
3909 pet_expr_free(expr
);
3919 /* Do we need to construct a skip condition of the given type
3920 * on an if statement, given that the if condition is non-affine?
3922 * pet_scop_filter_skip can only handle the case where the if condition
3923 * holds (the then branch) and the skip condition is universal.
3924 * In any other case, we need to construct a new skip condition.
3926 static bool need_skip(struct pet_scop
*scop_then
, struct pet_scop
*scop_else
,
3927 bool have_else
, enum pet_skip type
)
3929 if (have_else
&& scop_else
&& pet_scop_has_skip(scop_else
, type
))
3931 if (scop_then
&& pet_scop_has_skip(scop_then
, type
) &&
3932 !pet_scop_has_universal_skip(scop_then
, type
))
3937 /* Do we need to construct a skip condition of the given type
3938 * on an if statement, given that the if condition is affine?
3940 * There is no need to construct a new skip condition if all
3941 * the skip conditions are affine.
3943 static bool need_skip_aff(struct pet_scop
*scop_then
,
3944 struct pet_scop
*scop_else
, bool have_else
, enum pet_skip type
)
3946 if (scop_then
&& pet_scop_has_var_skip(scop_then
, type
))
3948 if (have_else
&& scop_else
&& pet_scop_has_var_skip(scop_else
, type
))
3953 /* Do we need to construct a skip condition of the given type
3954 * on an if statement?
3956 static bool need_skip(struct pet_scop
*scop_then
, struct pet_scop
*scop_else
,
3957 bool have_else
, enum pet_skip type
, bool affine
)
3960 return need_skip_aff(scop_then
, scop_else
, have_else
, type
);
3962 return need_skip(scop_then
, scop_else
, have_else
, type
);
3965 /* Construct an affine expression pet_expr that evaluates
3966 * to the constant "val".
3968 static struct pet_expr
*universally(isl_ctx
*ctx
, int val
)
3970 isl_local_space
*ls
;
3973 isl_multi_pw_aff
*mpa
;
3975 ls
= isl_local_space_from_space(isl_space_set_alloc(ctx
, 0, 0));
3976 aff
= isl_aff_val_on_domain(ls
, isl_val_int_from_si(ctx
, val
));
3977 mpa
= isl_multi_pw_aff_from_pw_aff(isl_pw_aff_from_aff(aff
));
3979 return pet_expr_from_index(mpa
);
3982 /* Construct an affine expression pet_expr that evaluates
3983 * to the constant 1.
3985 static struct pet_expr
*universally_true(isl_ctx
*ctx
)
3987 return universally(ctx
, 1);
3990 /* Construct an affine expression pet_expr that evaluates
3991 * to the constant 0.
3993 static struct pet_expr
*universally_false(isl_ctx
*ctx
)
3995 return universally(ctx
, 0);
3998 /* Given an access relation "test_access" for the if condition,
3999 * an access relation "skip_access" for the skip condition and
4000 * scops for the then and else branches, construct a scop for
4001 * computing "skip_access".
4003 * The computed scop contains a single statement that essentially does
4005 * skip_cond = test_cond ? skip_cond_then : skip_cond_else
4007 * If the skip conditions of the then and/or else branch are not affine,
4008 * then they need to be filtered by test_access.
4009 * If they are missing, then this means the skip condition is false.
4011 * Since we are constructing a skip condition for the if statement,
4012 * the skip conditions on the then and else branches are removed.
4014 static struct pet_scop
*extract_skip(PetScan
*scan
,
4015 __isl_take isl_map
*test_access
, __isl_take isl_map
*skip_access
,
4016 struct pet_scop
*scop_then
, struct pet_scop
*scop_else
, bool have_else
,
4019 struct pet_expr
*expr_then
, *expr_else
, *expr
, *expr_skip
;
4020 struct pet_stmt
*stmt
;
4021 struct pet_scop
*scop
;
4022 isl_ctx
*ctx
= scan
->ctx
;
4026 if (have_else
&& !scop_else
)
4029 if (pet_scop_has_skip(scop_then
, type
)) {
4030 expr_then
= pet_scop_get_skip_expr(scop_then
, type
);
4031 pet_scop_reset_skip(scop_then
, type
);
4032 if (!pet_expr_is_affine(expr_then
))
4033 expr_then
= pet_expr_filter(expr_then
,
4034 isl_map_copy(test_access
), 1);
4036 expr_then
= universally_false(ctx
);
4038 if (have_else
&& pet_scop_has_skip(scop_else
, type
)) {
4039 expr_else
= pet_scop_get_skip_expr(scop_else
, type
);
4040 pet_scop_reset_skip(scop_else
, type
);
4041 if (!pet_expr_is_affine(expr_else
))
4042 expr_else
= pet_expr_filter(expr_else
,
4043 isl_map_copy(test_access
), 0);
4045 expr_else
= universally_false(ctx
);
4047 expr
= pet_expr_from_access(test_access
);
4048 expr
= pet_expr_new_ternary(ctx
, expr
, expr_then
, expr_else
);
4049 expr_skip
= pet_expr_from_access(isl_map_copy(skip_access
));
4051 expr_skip
->acc
.write
= 1;
4052 expr_skip
->acc
.read
= 0;
4054 expr
= pet_expr_new_binary(ctx
, pet_op_assign
, expr_skip
, expr
);
4055 stmt
= pet_stmt_from_pet_expr(ctx
, -1, NULL
, scan
->n_stmt
++, expr
);
4057 scop
= pet_scop_from_pet_stmt(ctx
, stmt
);
4058 scop
= scop_add_array(scop
, skip_access
, scan
->ast_context
);
4059 isl_map_free(skip_access
);
4063 isl_map_free(test_access
);
4064 isl_map_free(skip_access
);
4068 /* Is scop's skip_now condition equal to its skip_later condition?
4069 * In particular, this means that it either has no skip_now condition
4070 * or both a skip_now and a skip_later condition (that are equal to each other).
4072 static bool skip_equals_skip_later(struct pet_scop
*scop
)
4074 int has_skip_now
, has_skip_later
;
4076 isl_set
*skip_now
, *skip_later
;
4080 has_skip_now
= pet_scop_has_skip(scop
, pet_skip_now
);
4081 has_skip_later
= pet_scop_has_skip(scop
, pet_skip_later
);
4082 if (has_skip_now
!= has_skip_later
)
4087 skip_now
= pet_scop_get_skip(scop
, pet_skip_now
);
4088 skip_later
= pet_scop_get_skip(scop
, pet_skip_later
);
4089 equal
= isl_set_is_equal(skip_now
, skip_later
);
4090 isl_set_free(skip_now
);
4091 isl_set_free(skip_later
);
4096 /* Drop the skip conditions of type pet_skip_later from scop1 and scop2.
4098 static void drop_skip_later(struct pet_scop
*scop1
, struct pet_scop
*scop2
)
4100 pet_scop_reset_skip(scop1
, pet_skip_later
);
4101 pet_scop_reset_skip(scop2
, pet_skip_later
);
4104 /* Structure that handles the construction of skip conditions.
4106 * scop_then and scop_else represent the then and else branches
4107 * of the if statement
4109 * skip[type] is true if we need to construct a skip condition of that type
4110 * equal is set if the skip conditions of types pet_skip_now and pet_skip_later
4111 * are equal to each other
4112 * access[type] is the virtual array representing the skip condition
4113 * scop[type] is a scop for computing the skip condition
4115 struct pet_skip_info
{
4121 struct pet_scop
*scop
[2];
4123 pet_skip_info(isl_ctx
*ctx
) : ctx(ctx
) {}
4125 operator bool() { return skip
[pet_skip_now
] || skip
[pet_skip_later
]; }
4128 /* Structure that handles the construction of skip conditions on if statements.
4130 * scop_then and scop_else represent the then and else branches
4131 * of the if statement
4133 struct pet_skip_info_if
: public pet_skip_info
{
4134 struct pet_scop
*scop_then
, *scop_else
;
4137 pet_skip_info_if(isl_ctx
*ctx
, struct pet_scop
*scop_then
,
4138 struct pet_scop
*scop_else
, bool have_else
, bool affine
);
4139 void extract(PetScan
*scan
, __isl_keep isl_multi_pw_aff
*index
,
4140 enum pet_skip type
);
4141 void extract(PetScan
*scan
, __isl_keep isl_multi_pw_aff
*index
);
4142 void extract(PetScan
*scan
, __isl_keep isl_pw_aff
*cond
);
4143 struct pet_scop
*add(struct pet_scop
*scop
, enum pet_skip type
,
4145 struct pet_scop
*add(struct pet_scop
*scop
, int offset
);
4148 /* Initialize a pet_skip_info_if structure based on the then and else branches
4149 * and based on whether the if condition is affine or not.
4151 pet_skip_info_if::pet_skip_info_if(isl_ctx
*ctx
, struct pet_scop
*scop_then
,
4152 struct pet_scop
*scop_else
, bool have_else
, bool affine
) :
4153 pet_skip_info(ctx
), scop_then(scop_then
), scop_else(scop_else
),
4154 have_else(have_else
)
4156 skip
[pet_skip_now
] =
4157 need_skip(scop_then
, scop_else
, have_else
, pet_skip_now
, affine
);
4158 equal
= skip
[pet_skip_now
] && skip_equals_skip_later(scop_then
) &&
4159 (!have_else
|| skip_equals_skip_later(scop_else
));
4160 skip
[pet_skip_later
] = skip
[pet_skip_now
] && !equal
&&
4161 need_skip(scop_then
, scop_else
, have_else
, pet_skip_later
, affine
);
4164 /* If we need to construct a skip condition of the given type,
4167 * "index" represents the if condition.
4169 void pet_skip_info_if::extract(PetScan
*scan
,
4170 __isl_keep isl_multi_pw_aff
*index
, enum pet_skip type
)
4177 map
= isl_map_from_multi_pw_aff(isl_multi_pw_aff_copy(index
));
4178 access
[type
] = create_test_access(isl_map_get_ctx(map
), scan
->n_test
++);
4179 scop
[type
] = extract_skip(scan
, map
, isl_map_copy(access
[type
]),
4180 scop_then
, scop_else
, have_else
, type
);
4183 /* Construct the required skip conditions, given the if condition "index".
4185 void pet_skip_info_if::extract(PetScan
*scan
,
4186 __isl_keep isl_multi_pw_aff
*index
)
4188 extract(scan
, index
, pet_skip_now
);
4189 extract(scan
, index
, pet_skip_later
);
4191 drop_skip_later(scop_then
, scop_else
);
4194 /* Construct the required skip conditions, given the if condition "cond".
4196 void pet_skip_info_if::extract(PetScan
*scan
, __isl_keep isl_pw_aff
*cond
)
4198 isl_multi_pw_aff
*test
;
4200 if (!skip
[pet_skip_now
] && !skip
[pet_skip_later
])
4203 test
= isl_multi_pw_aff_from_pw_aff(isl_pw_aff_copy(cond
));
4204 test
= isl_multi_pw_aff_from_range(test
);
4205 extract(scan
, test
);
4206 isl_multi_pw_aff_free(test
);
4209 /* Add the computed skip condition of the give type to "main" and
4210 * add the scop for computing the condition at the given offset.
4212 * If equal is set, then we only computed a skip condition for pet_skip_now,
4213 * but we also need to set it as main's pet_skip_later.
4215 struct pet_scop
*pet_skip_info_if::add(struct pet_scop
*main
,
4216 enum pet_skip type
, int offset
)
4223 skip_set
= isl_map_range(access
[type
]);
4224 access
[type
] = NULL
;
4225 scop
[type
] = pet_scop_prefix(scop
[type
], offset
);
4226 main
= pet_scop_add_par(ctx
, main
, scop
[type
]);
4230 main
= pet_scop_set_skip(main
, pet_skip_later
,
4231 isl_set_copy(skip_set
));
4233 main
= pet_scop_set_skip(main
, type
, skip_set
);
4238 /* Add the computed skip conditions to "main" and
4239 * add the scops for computing the conditions at the given offset.
4241 struct pet_scop
*pet_skip_info_if::add(struct pet_scop
*scop
, int offset
)
4243 scop
= add(scop
, pet_skip_now
, offset
);
4244 scop
= add(scop
, pet_skip_later
, offset
);
4249 /* Construct a pet_scop for a non-affine if statement.
4251 * We create a separate statement that writes the result
4252 * of the non-affine condition to a virtual scalar.
4253 * A constraint requiring the value of this virtual scalar to be one
4254 * is added to the iteration domains of the then branch.
4255 * Similarly, a constraint requiring the value of this virtual scalar
4256 * to be zero is added to the iteration domains of the else branch, if any.
4257 * We adjust the schedules to ensure that the virtual scalar is written
4258 * before it is read.
4260 * If there are any breaks or continues in the then and/or else
4261 * branches, then we may have to compute a new skip condition.
4262 * This is handled using a pet_skip_info_if object.
4263 * On initialization, the object checks if skip conditions need
4264 * to be computed. If so, it does so in "extract" and adds them in "add".
4266 struct pet_scop
*PetScan::extract_non_affine_if(Expr
*cond
,
4267 struct pet_scop
*scop_then
, struct pet_scop
*scop_else
,
4268 bool have_else
, int stmt_id
)
4270 struct pet_scop
*scop
;
4271 isl_multi_pw_aff
*test_index
;
4272 isl_map
*test_access
;
4273 int save_n_stmt
= n_stmt
;
4275 test_index
= create_test_index(ctx
, n_test
++);
4276 test_access
= isl_map_from_multi_pw_aff(
4277 isl_multi_pw_aff_copy(test_index
));
4279 scop
= extract_non_affine_condition(cond
, isl_map_copy(test_access
));
4280 n_stmt
= save_n_stmt
;
4281 scop
= scop_add_array(scop
, test_access
, ast_context
);
4283 pet_skip_info_if
skip(ctx
, scop_then
, scop_else
, have_else
, false);
4284 skip
.extract(this, test_index
);
4285 isl_multi_pw_aff_free(test_index
);
4287 scop
= pet_scop_prefix(scop
, 0);
4288 scop_then
= pet_scop_prefix(scop_then
, 1);
4289 scop_then
= pet_scop_filter(scop_then
, isl_map_copy(test_access
), 1);
4291 scop_else
= pet_scop_prefix(scop_else
, 1);
4292 scop_else
= pet_scop_filter(scop_else
, test_access
, 0);
4293 scop_then
= pet_scop_add_par(ctx
, scop_then
, scop_else
);
4295 isl_map_free(test_access
);
4297 scop
= pet_scop_add_seq(ctx
, scop
, scop_then
);
4299 scop
= skip
.add(scop
, 2);
4304 /* Construct a pet_scop for an if statement.
4306 * If the condition fits the pattern of a conditional assignment,
4307 * then it is handled by extract_conditional_assignment.
4308 * Otherwise, we do the following.
4310 * If the condition is affine, then the condition is added
4311 * to the iteration domains of the then branch, while the
4312 * opposite of the condition in added to the iteration domains
4313 * of the else branch, if any.
4314 * We allow the condition to be dynamic, i.e., to refer to
4315 * scalars or array elements that may be written to outside
4316 * of the given if statement. These nested accesses are then represented
4317 * as output dimensions in the wrapping iteration domain.
4318 * If it also written _inside_ the then or else branch, then
4319 * we treat the condition as non-affine.
4320 * As explained in extract_non_affine_if, this will introduce
4321 * an extra statement.
4322 * For aesthetic reasons, we want this statement to have a statement
4323 * number that is lower than those of the then and else branches.
4324 * In order to evaluate if will need such a statement, however, we
4325 * first construct scops for the then and else branches.
4326 * We therefore reserve a statement number if we might have to
4327 * introduce such an extra statement.
4329 * If the condition is not affine, then the scop is created in
4330 * extract_non_affine_if.
4332 * If there are any breaks or continues in the then and/or else
4333 * branches, then we may have to compute a new skip condition.
4334 * This is handled using a pet_skip_info_if object.
4335 * On initialization, the object checks if skip conditions need
4336 * to be computed. If so, it does so in "extract" and adds them in "add".
4338 struct pet_scop
*PetScan::extract(IfStmt
*stmt
)
4340 struct pet_scop
*scop_then
, *scop_else
= NULL
, *scop
;
4346 scop
= extract_conditional_assignment(stmt
);
4350 cond
= try_extract_nested_condition(stmt
->getCond());
4351 if (allow_nested
&& (!cond
|| has_nested(cond
)))
4355 assigned_value_cache
cache(assigned_value
);
4356 scop_then
= extract(stmt
->getThen());
4359 if (stmt
->getElse()) {
4360 assigned_value_cache
cache(assigned_value
);
4361 scop_else
= extract(stmt
->getElse());
4362 if (options
->autodetect
) {
4363 if (scop_then
&& !scop_else
) {
4365 isl_pw_aff_free(cond
);
4368 if (!scop_then
&& scop_else
) {
4370 isl_pw_aff_free(cond
);
4377 (!is_nested_allowed(cond
, scop_then
) ||
4378 (stmt
->getElse() && !is_nested_allowed(cond
, scop_else
)))) {
4379 isl_pw_aff_free(cond
);
4382 if (allow_nested
&& !cond
)
4383 return extract_non_affine_if(stmt
->getCond(), scop_then
,
4384 scop_else
, stmt
->getElse(), stmt_id
);
4387 cond
= extract_condition(stmt
->getCond());
4389 pet_skip_info_if
skip(ctx
, scop_then
, scop_else
, stmt
->getElse(), true);
4390 skip
.extract(this, cond
);
4392 valid
= isl_pw_aff_domain(isl_pw_aff_copy(cond
));
4393 set
= isl_pw_aff_non_zero_set(cond
);
4394 scop
= pet_scop_restrict(scop_then
, isl_set_copy(set
));
4396 if (stmt
->getElse()) {
4397 set
= isl_set_subtract(isl_set_copy(valid
), set
);
4398 scop_else
= pet_scop_restrict(scop_else
, set
);
4399 scop
= pet_scop_add_par(ctx
, scop
, scop_else
);
4402 scop
= resolve_nested(scop
);
4403 scop
= pet_scop_restrict_context(scop
, valid
);
4406 scop
= pet_scop_prefix(scop
, 0);
4407 scop
= skip
.add(scop
, 1);
4412 /* Try and construct a pet_scop for a label statement.
4413 * We currently only allow labels on expression statements.
4415 struct pet_scop
*PetScan::extract(LabelStmt
*stmt
)
4420 sub
= stmt
->getSubStmt();
4421 if (!isa
<Expr
>(sub
)) {
4426 label
= isl_id_alloc(ctx
, stmt
->getName(), NULL
);
4428 return extract(sub
, extract_expr(cast
<Expr
>(sub
)), label
);
4431 /* Construct a pet_scop for a continue statement.
4433 * We simply create an empty scop with a universal pet_skip_now
4434 * skip condition. This skip condition will then be taken into
4435 * account by the enclosing loop construct, possibly after
4436 * being incorporated into outer skip conditions.
4438 struct pet_scop
*PetScan::extract(ContinueStmt
*stmt
)
4444 scop
= pet_scop_empty(ctx
);
4448 space
= isl_space_set_alloc(ctx
, 0, 1);
4449 set
= isl_set_universe(space
);
4450 set
= isl_set_fix_si(set
, isl_dim_set
, 0, 1);
4451 scop
= pet_scop_set_skip(scop
, pet_skip_now
, set
);
4456 /* Construct a pet_scop for a break statement.
4458 * We simply create an empty scop with both a universal pet_skip_now
4459 * skip condition and a universal pet_skip_later skip condition.
4460 * These skip conditions will then be taken into
4461 * account by the enclosing loop construct, possibly after
4462 * being incorporated into outer skip conditions.
4464 struct pet_scop
*PetScan::extract(BreakStmt
*stmt
)
4470 scop
= pet_scop_empty(ctx
);
4474 space
= isl_space_set_alloc(ctx
, 0, 1);
4475 set
= isl_set_universe(space
);
4476 set
= isl_set_fix_si(set
, isl_dim_set
, 0, 1);
4477 scop
= pet_scop_set_skip(scop
, pet_skip_now
, isl_set_copy(set
));
4478 scop
= pet_scop_set_skip(scop
, pet_skip_later
, set
);
4483 /* Try and construct a pet_scop corresponding to "stmt".
4485 * If "stmt" is a compound statement, then "skip_declarations"
4486 * indicates whether we should skip initial declarations in the
4487 * compound statement.
4489 * If the constructed pet_scop is not a (possibly) partial representation
4490 * of "stmt", we update start and end of the pet_scop to those of "stmt".
4491 * In particular, if skip_declarations, then we may have skipped declarations
4492 * inside "stmt" and so the pet_scop may not represent the entire "stmt".
4493 * Note that this function may be called with "stmt" referring to the entire
4494 * body of the function, including the outer braces. In such cases,
4495 * skip_declarations will be set and the braces will not be taken into
4496 * account in scop->start and scop->end.
4498 struct pet_scop
*PetScan::extract(Stmt
*stmt
, bool skip_declarations
)
4500 struct pet_scop
*scop
;
4501 unsigned start
, end
;
4503 SourceManager
&SM
= PP
.getSourceManager();
4504 const LangOptions
&LO
= PP
.getLangOpts();
4506 if (isa
<Expr
>(stmt
))
4507 return extract(stmt
, extract_expr(cast
<Expr
>(stmt
)));
4509 switch (stmt
->getStmtClass()) {
4510 case Stmt::WhileStmtClass
:
4511 scop
= extract(cast
<WhileStmt
>(stmt
));
4513 case Stmt::ForStmtClass
:
4514 scop
= extract_for(cast
<ForStmt
>(stmt
));
4516 case Stmt::IfStmtClass
:
4517 scop
= extract(cast
<IfStmt
>(stmt
));
4519 case Stmt::CompoundStmtClass
:
4520 scop
= extract(cast
<CompoundStmt
>(stmt
), skip_declarations
);
4522 case Stmt::LabelStmtClass
:
4523 scop
= extract(cast
<LabelStmt
>(stmt
));
4525 case Stmt::ContinueStmtClass
:
4526 scop
= extract(cast
<ContinueStmt
>(stmt
));
4528 case Stmt::BreakStmtClass
:
4529 scop
= extract(cast
<BreakStmt
>(stmt
));
4531 case Stmt::DeclStmtClass
:
4532 scop
= extract(cast
<DeclStmt
>(stmt
));
4539 if (partial
|| skip_declarations
)
4542 loc
= stmt
->getLocStart();
4543 loc
= move_to_start_of_line_if_first_token(loc
, SM
, LO
);
4544 start
= getExpansionOffset(SM
, loc
);
4545 loc
= PP
.getLocForEndOfToken(stmt
->getLocEnd());
4546 end
= getExpansionOffset(SM
, loc
);
4547 scop
= pet_scop_update_start_end(scop
, start
, end
);
4552 /* Do we need to construct a skip condition of the given type
4553 * on a sequence of statements?
4555 * There is no need to construct a new skip condition if only
4556 * only of the two statements has a skip condition or if both
4557 * of their skip conditions are affine.
4559 * In principle we also don't need a new continuation variable if
4560 * the continuation of scop2 is affine, but then we would need
4561 * to allow more complicated forms of continuations.
4563 static bool need_skip_seq(struct pet_scop
*scop1
, struct pet_scop
*scop2
,
4566 if (!scop1
|| !pet_scop_has_skip(scop1
, type
))
4568 if (!scop2
|| !pet_scop_has_skip(scop2
, type
))
4570 if (pet_scop_has_affine_skip(scop1
, type
) &&
4571 pet_scop_has_affine_skip(scop2
, type
))
4576 /* Construct a scop for computing the skip condition of the given type and
4577 * with access relation "skip_access" for a sequence of two scops "scop1"
4580 * The computed scop contains a single statement that essentially does
4582 * skip_cond = skip_cond_1 ? 1 : skip_cond_2
4584 * or, in other words, skip_cond1 || skip_cond2.
4585 * In this expression, skip_cond_2 is filtered to reflect that it is
4586 * only evaluated when skip_cond_1 is false.
4588 * The skip condition on scop1 is not removed because it still needs
4589 * to be applied to scop2 when these two scops are combined.
4591 static struct pet_scop
*extract_skip_seq(PetScan
*ps
,
4592 __isl_take isl_map
*skip_access
,
4593 struct pet_scop
*scop1
, struct pet_scop
*scop2
, enum pet_skip type
)
4596 struct pet_expr
*expr1
, *expr2
, *expr
, *expr_skip
;
4597 struct pet_stmt
*stmt
;
4598 struct pet_scop
*scop
;
4599 isl_ctx
*ctx
= ps
->ctx
;
4601 if (!scop1
|| !scop2
)
4604 expr1
= pet_scop_get_skip_expr(scop1
, type
);
4605 expr2
= pet_scop_get_skip_expr(scop2
, type
);
4606 pet_scop_reset_skip(scop2
, type
);
4608 expr2
= pet_expr_filter(expr2
, isl_map_copy(expr1
->acc
.access
), 0);
4610 expr
= universally_true(ctx
);
4611 expr
= pet_expr_new_ternary(ctx
, expr1
, expr
, expr2
);
4612 expr_skip
= pet_expr_from_access(isl_map_copy(skip_access
));
4614 expr_skip
->acc
.write
= 1;
4615 expr_skip
->acc
.read
= 0;
4617 expr
= pet_expr_new_binary(ctx
, pet_op_assign
, expr_skip
, expr
);
4618 stmt
= pet_stmt_from_pet_expr(ctx
, -1, NULL
, ps
->n_stmt
++, expr
);
4620 scop
= pet_scop_from_pet_stmt(ctx
, stmt
);
4621 scop
= scop_add_array(scop
, skip_access
, ps
->ast_context
);
4622 isl_map_free(skip_access
);
4626 isl_map_free(skip_access
);
4630 /* Structure that handles the construction of skip conditions
4631 * on sequences of statements.
4633 * scop1 and scop2 represent the two statements that are combined
4635 struct pet_skip_info_seq
: public pet_skip_info
{
4636 struct pet_scop
*scop1
, *scop2
;
4638 pet_skip_info_seq(isl_ctx
*ctx
, struct pet_scop
*scop1
,
4639 struct pet_scop
*scop2
);
4640 void extract(PetScan
*scan
, enum pet_skip type
);
4641 void extract(PetScan
*scan
);
4642 struct pet_scop
*add(struct pet_scop
*scop
, enum pet_skip type
,
4644 struct pet_scop
*add(struct pet_scop
*scop
, int offset
);
4647 /* Initialize a pet_skip_info_seq structure based on
4648 * on the two statements that are going to be combined.
4650 pet_skip_info_seq::pet_skip_info_seq(isl_ctx
*ctx
, struct pet_scop
*scop1
,
4651 struct pet_scop
*scop2
) : pet_skip_info(ctx
), scop1(scop1
), scop2(scop2
)
4653 skip
[pet_skip_now
] = need_skip_seq(scop1
, scop2
, pet_skip_now
);
4654 equal
= skip
[pet_skip_now
] && skip_equals_skip_later(scop1
) &&
4655 skip_equals_skip_later(scop2
);
4656 skip
[pet_skip_later
] = skip
[pet_skip_now
] && !equal
&&
4657 need_skip_seq(scop1
, scop2
, pet_skip_later
);
4660 /* If we need to construct a skip condition of the given type,
4663 void pet_skip_info_seq::extract(PetScan
*scan
, enum pet_skip type
)
4668 access
[type
] = create_test_access(ctx
, scan
->n_test
++);
4669 scop
[type
] = extract_skip_seq(scan
, isl_map_copy(access
[type
]),
4670 scop1
, scop2
, type
);
4673 /* Construct the required skip conditions.
4675 void pet_skip_info_seq::extract(PetScan
*scan
)
4677 extract(scan
, pet_skip_now
);
4678 extract(scan
, pet_skip_later
);
4680 drop_skip_later(scop1
, scop2
);
4683 /* Add the computed skip condition of the given type to "main" and
4684 * add the scop for computing the condition at the given offset (the statement
4685 * number). Within this offset, the condition is computed at position 1
4686 * to ensure that it is computed after the corresponding statement.
4688 * If equal is set, then we only computed a skip condition for pet_skip_now,
4689 * but we also need to set it as main's pet_skip_later.
4691 struct pet_scop
*pet_skip_info_seq::add(struct pet_scop
*main
,
4692 enum pet_skip type
, int offset
)
4699 skip_set
= isl_map_range(access
[type
]);
4700 access
[type
] = NULL
;
4701 scop
[type
] = pet_scop_prefix(scop
[type
], 1);
4702 scop
[type
] = pet_scop_prefix(scop
[type
], offset
);
4703 main
= pet_scop_add_par(ctx
, main
, scop
[type
]);
4707 main
= pet_scop_set_skip(main
, pet_skip_later
,
4708 isl_set_copy(skip_set
));
4710 main
= pet_scop_set_skip(main
, type
, skip_set
);
4715 /* Add the computed skip conditions to "main" and
4716 * add the scops for computing the conditions at the given offset.
4718 struct pet_scop
*pet_skip_info_seq::add(struct pet_scop
*scop
, int offset
)
4720 scop
= add(scop
, pet_skip_now
, offset
);
4721 scop
= add(scop
, pet_skip_later
, offset
);
4726 /* Extract a clone of the kill statement in "scop".
4727 * "scop" is expected to have been created from a DeclStmt
4728 * and should have the kill as its first statement.
4730 struct pet_stmt
*PetScan::extract_kill(struct pet_scop
*scop
)
4732 struct pet_expr
*kill
;
4733 struct pet_stmt
*stmt
;
4738 if (scop
->n_stmt
< 1)
4739 isl_die(ctx
, isl_error_internal
,
4740 "expecting at least one statement", return NULL
);
4741 stmt
= scop
->stmts
[0];
4742 if (stmt
->body
->type
!= pet_expr_unary
||
4743 stmt
->body
->op
!= pet_op_kill
)
4744 isl_die(ctx
, isl_error_internal
,
4745 "expecting kill statement", return NULL
);
4747 access
= isl_map_copy(stmt
->body
->args
[0]->acc
.access
);
4748 access
= isl_map_reset_tuple_id(access
, isl_dim_in
);
4749 kill
= pet_expr_kill_from_access(access
);
4750 return pet_stmt_from_pet_expr(ctx
, stmt
->line
, NULL
, n_stmt
++, kill
);
4753 /* Mark all arrays in "scop" as being exposed.
4755 static struct pet_scop
*mark_exposed(struct pet_scop
*scop
)
4759 for (int i
= 0; i
< scop
->n_array
; ++i
)
4760 scop
->arrays
[i
]->exposed
= 1;
4764 /* Try and construct a pet_scop corresponding to (part of)
4765 * a sequence of statements.
4767 * "block" is set if the sequence respresents the children of
4768 * a compound statement.
4769 * "skip_declarations" is set if we should skip initial declarations
4770 * in the sequence of statements.
4772 * If there are any breaks or continues in the individual statements,
4773 * then we may have to compute a new skip condition.
4774 * This is handled using a pet_skip_info_seq object.
4775 * On initialization, the object checks if skip conditions need
4776 * to be computed. If so, it does so in "extract" and adds them in "add".
4778 * If "block" is set, then we need to insert kill statements at
4779 * the end of the block for any array that has been declared by
4780 * one of the statements in the sequence. Each of these declarations
4781 * results in the construction of a kill statement at the place
4782 * of the declaration, so we simply collect duplicates of
4783 * those kill statements and append these duplicates to the constructed scop.
4785 * If "block" is not set, then any array declared by one of the statements
4786 * in the sequence is marked as being exposed.
4788 struct pet_scop
*PetScan::extract(StmtRange stmt_range
, bool block
,
4789 bool skip_declarations
)
4794 bool partial_range
= false;
4795 set
<struct pet_stmt
*> kills
;
4796 set
<struct pet_stmt
*>::iterator it
;
4798 scop
= pet_scop_empty(ctx
);
4799 for (i
= stmt_range
.first
, j
= 0; i
!= stmt_range
.second
; ++i
, ++j
) {
4801 struct pet_scop
*scop_i
;
4803 if (skip_declarations
&&
4804 child
->getStmtClass() == Stmt::DeclStmtClass
)
4807 scop_i
= extract(child
);
4808 if (scop
&& partial
) {
4809 pet_scop_free(scop_i
);
4812 pet_skip_info_seq
skip(ctx
, scop
, scop_i
);
4815 scop_i
= pet_scop_prefix(scop_i
, 0);
4816 if (scop_i
&& child
->getStmtClass() == Stmt::DeclStmtClass
) {
4818 kills
.insert(extract_kill(scop_i
));
4820 scop_i
= mark_exposed(scop_i
);
4822 scop_i
= pet_scop_prefix(scop_i
, j
);
4823 if (options
->autodetect
) {
4825 scop
= pet_scop_add_seq(ctx
, scop
, scop_i
);
4827 partial_range
= true;
4828 if (scop
->n_stmt
!= 0 && !scop_i
)
4831 scop
= pet_scop_add_seq(ctx
, scop
, scop_i
);
4834 scop
= skip
.add(scop
, j
);
4840 for (it
= kills
.begin(); it
!= kills
.end(); ++it
) {
4842 scop_j
= pet_scop_from_pet_stmt(ctx
, *it
);
4843 scop_j
= pet_scop_prefix(scop_j
, j
);
4844 scop
= pet_scop_add_seq(ctx
, scop
, scop_j
);
4847 if (scop
&& partial_range
) {
4848 if (scop
->n_stmt
== 0) {
4849 pet_scop_free(scop
);
4858 /* Check if the scop marked by the user is exactly this Stmt
4859 * or part of this Stmt.
4860 * If so, return a pet_scop corresponding to the marked region.
4861 * Otherwise, return NULL.
4863 struct pet_scop
*PetScan::scan(Stmt
*stmt
)
4865 SourceManager
&SM
= PP
.getSourceManager();
4866 unsigned start_off
, end_off
;
4868 start_off
= getExpansionOffset(SM
, stmt
->getLocStart());
4869 end_off
= getExpansionOffset(SM
, stmt
->getLocEnd());
4871 if (start_off
> loc
.end
)
4873 if (end_off
< loc
.start
)
4875 if (start_off
>= loc
.start
&& end_off
<= loc
.end
) {
4876 return extract(stmt
);
4880 for (start
= stmt
->child_begin(); start
!= stmt
->child_end(); ++start
) {
4881 Stmt
*child
= *start
;
4884 start_off
= getExpansionOffset(SM
, child
->getLocStart());
4885 end_off
= getExpansionOffset(SM
, child
->getLocEnd());
4886 if (start_off
< loc
.start
&& end_off
>= loc
.end
)
4888 if (start_off
>= loc
.start
)
4893 for (end
= start
; end
!= stmt
->child_end(); ++end
) {
4895 start_off
= SM
.getFileOffset(child
->getLocStart());
4896 if (start_off
>= loc
.end
)
4900 return extract(StmtRange(start
, end
), false, false);
4903 /* Set the size of index "pos" of "array" to "size".
4904 * In particular, add a constraint of the form
4908 * to array->extent and a constraint of the form
4912 * to array->context.
4914 static struct pet_array
*update_size(struct pet_array
*array
, int pos
,
4915 __isl_take isl_pw_aff
*size
)
4925 valid
= isl_pw_aff_nonneg_set(isl_pw_aff_copy(size
));
4926 array
->context
= isl_set_intersect(array
->context
, valid
);
4928 dim
= isl_set_get_space(array
->extent
);
4929 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
4930 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, pos
, 1);
4931 univ
= isl_set_universe(isl_aff_get_domain_space(aff
));
4932 index
= isl_pw_aff_alloc(univ
, aff
);
4934 size
= isl_pw_aff_add_dims(size
, isl_dim_in
,
4935 isl_set_dim(array
->extent
, isl_dim_set
));
4936 id
= isl_set_get_tuple_id(array
->extent
);
4937 size
= isl_pw_aff_set_tuple_id(size
, isl_dim_in
, id
);
4938 bound
= isl_pw_aff_lt_set(index
, size
);
4940 array
->extent
= isl_set_intersect(array
->extent
, bound
);
4942 if (!array
->context
|| !array
->extent
)
4947 pet_array_free(array
);
4951 /* Figure out the size of the array at position "pos" and all
4952 * subsequent positions from "type" and update "array" accordingly.
4954 struct pet_array
*PetScan::set_upper_bounds(struct pet_array
*array
,
4955 const Type
*type
, int pos
)
4957 const ArrayType
*atype
;
4963 if (type
->isPointerType()) {
4964 type
= type
->getPointeeType().getTypePtr();
4965 return set_upper_bounds(array
, type
, pos
+ 1);
4967 if (!type
->isArrayType())
4970 type
= type
->getCanonicalTypeInternal().getTypePtr();
4971 atype
= cast
<ArrayType
>(type
);
4973 if (type
->isConstantArrayType()) {
4974 const ConstantArrayType
*ca
= cast
<ConstantArrayType
>(atype
);
4975 size
= extract_affine(ca
->getSize());
4976 array
= update_size(array
, pos
, size
);
4977 } else if (type
->isVariableArrayType()) {
4978 const VariableArrayType
*vla
= cast
<VariableArrayType
>(atype
);
4979 size
= extract_affine(vla
->getSizeExpr());
4980 array
= update_size(array
, pos
, size
);
4983 type
= atype
->getElementType().getTypePtr();
4985 return set_upper_bounds(array
, type
, pos
+ 1);
4988 /* Is "T" the type of a variable length array with static size?
4990 static bool is_vla_with_static_size(QualType T
)
4992 const VariableArrayType
*vlatype
;
4994 if (!T
->isVariableArrayType())
4996 vlatype
= cast
<VariableArrayType
>(T
);
4997 return vlatype
->getSizeModifier() == VariableArrayType::Static
;
5000 /* Return the type of "decl" as an array.
5002 * In particular, if "decl" is a parameter declaration that
5003 * is a variable length array with a static size, then
5004 * return the original type (i.e., the variable length array).
5005 * Otherwise, return the type of decl.
5007 static QualType
get_array_type(ValueDecl
*decl
)
5012 parm
= dyn_cast
<ParmVarDecl
>(decl
);
5014 return decl
->getType();
5016 T
= parm
->getOriginalType();
5017 if (!is_vla_with_static_size(T
))
5018 return decl
->getType();
5022 /* Construct and return a pet_array corresponding to the variable "decl".
5023 * In particular, initialize array->extent to
5025 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
5027 * and then call set_upper_bounds to set the upper bounds on the indices
5028 * based on the type of the variable.
5030 struct pet_array
*PetScan::extract_array(isl_ctx
*ctx
, ValueDecl
*decl
)
5032 struct pet_array
*array
;
5033 QualType qt
= get_array_type(decl
);
5034 const Type
*type
= qt
.getTypePtr();
5035 int depth
= array_depth(type
);
5036 QualType base
= base_type(qt
);
5041 array
= isl_calloc_type(ctx
, struct pet_array
);
5045 id
= isl_id_alloc(ctx
, decl
->getName().str().c_str(), decl
);
5046 dim
= isl_space_set_alloc(ctx
, 0, depth
);
5047 dim
= isl_space_set_tuple_id(dim
, isl_dim_set
, id
);
5049 array
->extent
= isl_set_nat_universe(dim
);
5051 dim
= isl_space_params_alloc(ctx
, 0);
5052 array
->context
= isl_set_universe(dim
);
5054 array
= set_upper_bounds(array
, type
, 0);
5058 name
= base
.getAsString();
5059 array
->element_type
= strdup(name
.c_str());
5060 array
->element_size
= decl
->getASTContext().getTypeInfo(base
).first
/ 8;
5065 /* Construct a list of pet_arrays, one for each array (or scalar)
5066 * accessed inside "scop", add this list to "scop" and return the result.
5068 * The context of "scop" is updated with the intersection of
5069 * the contexts of all arrays, i.e., constraints on the parameters
5070 * that ensure that the arrays have a valid (non-negative) size.
5072 struct pet_scop
*PetScan::scan_arrays(struct pet_scop
*scop
)
5075 set
<ValueDecl
*> arrays
;
5076 set
<ValueDecl
*>::iterator it
;
5078 struct pet_array
**scop_arrays
;
5083 pet_scop_collect_arrays(scop
, arrays
);
5084 if (arrays
.size() == 0)
5087 n_array
= scop
->n_array
;
5089 scop_arrays
= isl_realloc_array(ctx
, scop
->arrays
, struct pet_array
*,
5090 n_array
+ arrays
.size());
5093 scop
->arrays
= scop_arrays
;
5095 for (it
= arrays
.begin(), i
= 0; it
!= arrays
.end(); ++it
, ++i
) {
5096 struct pet_array
*array
;
5097 scop
->arrays
[n_array
+ i
] = array
= extract_array(ctx
, *it
);
5098 if (!scop
->arrays
[n_array
+ i
])
5101 scop
->context
= isl_set_intersect(scop
->context
,
5102 isl_set_copy(array
->context
));
5109 pet_scop_free(scop
);
5113 /* Bound all parameters in scop->context to the possible values
5114 * of the corresponding C variable.
5116 static struct pet_scop
*add_parameter_bounds(struct pet_scop
*scop
)
5123 n
= isl_set_dim(scop
->context
, isl_dim_param
);
5124 for (int i
= 0; i
< n
; ++i
) {
5128 id
= isl_set_get_dim_id(scop
->context
, isl_dim_param
, i
);
5129 if (is_nested_parameter(id
)) {
5131 isl_die(isl_set_get_ctx(scop
->context
),
5133 "unresolved nested parameter", goto error
);
5135 decl
= (ValueDecl
*) isl_id_get_user(id
);
5138 scop
->context
= set_parameter_bounds(scop
->context
, i
, decl
);
5146 pet_scop_free(scop
);
5150 /* Construct a pet_scop from the given function.
5152 * If the scop was delimited by scop and endscop pragmas, then we override
5153 * the file offsets by those derived from the pragmas.
5155 struct pet_scop
*PetScan::scan(FunctionDecl
*fd
)
5160 stmt
= fd
->getBody();
5162 if (options
->autodetect
)
5163 scop
= extract(stmt
, true);
5166 scop
= pet_scop_update_start_end(scop
, loc
.start
, loc
.end
);
5168 scop
= pet_scop_detect_parameter_accesses(scop
);
5169 scop
= scan_arrays(scop
);
5170 scop
= add_parameter_bounds(scop
);
5171 scop
= pet_scop_gist(scop
, value_bounds
);