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
)
780 if (!nesting_enabled
) {
785 allow_nested
= false;
786 access
= extract_access(expr
);
792 isl_map_free(access
);
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_map
*PetScan::extract_access(ImplicitCastExpr
*expr
)
860 return extract_access(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 element type of the given array type.
880 static QualType
base_type(QualType qt
)
882 const Type
*type
= qt
.getTypePtr();
884 if (type
->isPointerType())
885 return base_type(type
->getPointeeType());
886 if (type
->isArrayType()) {
887 const ArrayType
*atype
;
888 type
= type
->getCanonicalTypeInternal().getTypePtr();
889 atype
= cast
<ArrayType
>(type
);
890 return base_type(atype
->getElementType());
895 /* Extract an access relation from a reference to a variable.
896 * If the variable has name "A" and its type corresponds to an
897 * array of depth d, then the returned access relation is of the
900 * { [] -> A[i_1,...,i_d] }
902 __isl_give isl_map
*PetScan::extract_access(DeclRefExpr
*expr
)
904 return extract_access(expr
->getDecl());
907 /* Extract an access relation from a variable.
908 * If the variable has name "A" and its type corresponds to an
909 * array of depth d, then the returned access relation is of the
912 * { [] -> A[i_1,...,i_d] }
914 __isl_give isl_map
*PetScan::extract_access(ValueDecl
*decl
)
916 int depth
= array_depth(decl
->getType().getTypePtr());
917 isl_id
*id
= isl_id_alloc(ctx
, decl
->getName().str().c_str(), decl
);
918 isl_space
*dim
= isl_space_alloc(ctx
, 0, 0, depth
);
921 dim
= isl_space_set_tuple_id(dim
, isl_dim_out
, id
);
923 access_rel
= isl_map_universe(dim
);
928 /* Extract an access relation from an integer contant.
929 * If the value of the constant is "v", then the returned access relation
934 __isl_give isl_map
*PetScan::extract_access(IntegerLiteral
*expr
)
936 return isl_map_from_range(isl_set_from_pw_aff(extract_affine(expr
)));
939 /* Try and extract an access relation from the given Expr.
940 * Return NULL if it doesn't work out.
942 __isl_give isl_map
*PetScan::extract_access(Expr
*expr
)
944 switch (expr
->getStmtClass()) {
945 case Stmt::ImplicitCastExprClass
:
946 return extract_access(cast
<ImplicitCastExpr
>(expr
));
947 case Stmt::DeclRefExprClass
:
948 return extract_access(cast
<DeclRefExpr
>(expr
));
949 case Stmt::ArraySubscriptExprClass
:
950 return extract_access(cast
<ArraySubscriptExpr
>(expr
));
951 case Stmt::IntegerLiteralClass
:
952 return extract_access(cast
<IntegerLiteral
>(expr
));
959 /* Assign the affine expression "index" to the output dimension "pos" of "map",
960 * restrict the domain to those values that result in a non-negative index
961 * and return the result.
963 __isl_give isl_map
*set_index(__isl_take isl_map
*map
, int pos
,
964 __isl_take isl_pw_aff
*index
)
967 int len
= isl_map_dim(map
, isl_dim_out
);
971 domain
= isl_pw_aff_nonneg_set(isl_pw_aff_copy(index
));
972 index
= isl_pw_aff_intersect_domain(index
, domain
);
973 index_map
= isl_map_from_range(isl_set_from_pw_aff(index
));
974 index_map
= isl_map_insert_dims(index_map
, isl_dim_out
, 0, pos
);
975 index_map
= isl_map_add_dims(index_map
, isl_dim_out
, len
- pos
- 1);
976 id
= isl_map_get_tuple_id(map
, isl_dim_out
);
977 index_map
= isl_map_set_tuple_id(index_map
, isl_dim_out
, id
);
979 map
= isl_map_intersect(map
, index_map
);
984 /* Extract an access relation from the given array subscript expression.
985 * If nesting is allowed in general, then we turn it on while
986 * examining the index expression.
988 * We first extract an access relation from the base.
989 * This will result in an access relation with a range that corresponds
990 * to the array being accessed and with earlier indices filled in already.
991 * We then extract the current index and fill that in as well.
992 * The position of the current index is based on the type of base.
993 * If base is the actual array variable, then the depth of this type
994 * will be the same as the depth of the array and we will fill in
995 * the first array index.
996 * Otherwise, the depth of the base type will be smaller and we will fill
999 __isl_give isl_map
*PetScan::extract_access(ArraySubscriptExpr
*expr
)
1001 Expr
*base
= expr
->getBase();
1002 Expr
*idx
= expr
->getIdx();
1004 isl_map
*base_access
;
1006 int depth
= array_depth(base
->getType().getTypePtr());
1008 bool save_nesting
= nesting_enabled
;
1010 nesting_enabled
= allow_nested
;
1012 base_access
= extract_access(base
);
1013 index
= extract_affine(idx
);
1015 nesting_enabled
= save_nesting
;
1017 pos
= isl_map_dim(base_access
, isl_dim_out
) - depth
;
1018 access
= set_index(base_access
, pos
, index
);
1023 /* Check if "expr" calls function "minmax" with two arguments and if so
1024 * make lhs and rhs refer to these two arguments.
1026 static bool is_minmax(Expr
*expr
, const char *minmax
, Expr
*&lhs
, Expr
*&rhs
)
1032 if (expr
->getStmtClass() != Stmt::CallExprClass
)
1035 call
= cast
<CallExpr
>(expr
);
1036 fd
= call
->getDirectCallee();
1040 if (call
->getNumArgs() != 2)
1043 name
= fd
->getDeclName().getAsString();
1047 lhs
= call
->getArg(0);
1048 rhs
= call
->getArg(1);
1053 /* Check if "expr" is of the form min(lhs, rhs) and if so make
1054 * lhs and rhs refer to the two arguments.
1056 static bool is_min(Expr
*expr
, Expr
*&lhs
, Expr
*&rhs
)
1058 return is_minmax(expr
, "min", lhs
, rhs
);
1061 /* Check if "expr" is of the form max(lhs, rhs) and if so make
1062 * lhs and rhs refer to the two arguments.
1064 static bool is_max(Expr
*expr
, Expr
*&lhs
, Expr
*&rhs
)
1066 return is_minmax(expr
, "max", lhs
, rhs
);
1069 /* Return "lhs && rhs", defined on the shared definition domain.
1071 static __isl_give isl_pw_aff
*pw_aff_and(__isl_take isl_pw_aff
*lhs
,
1072 __isl_take isl_pw_aff
*rhs
)
1077 dom
= isl_set_intersect(isl_pw_aff_domain(isl_pw_aff_copy(lhs
)),
1078 isl_pw_aff_domain(isl_pw_aff_copy(rhs
)));
1079 cond
= isl_set_intersect(isl_pw_aff_non_zero_set(lhs
),
1080 isl_pw_aff_non_zero_set(rhs
));
1081 return indicator_function(cond
, dom
);
1084 /* Return "lhs && rhs", with shortcut semantics.
1085 * That is, if lhs is false, then the result is defined even if rhs is not.
1086 * In practice, we compute lhs ? rhs : lhs.
1088 static __isl_give isl_pw_aff
*pw_aff_and_then(__isl_take isl_pw_aff
*lhs
,
1089 __isl_take isl_pw_aff
*rhs
)
1091 return isl_pw_aff_cond(isl_pw_aff_copy(lhs
), rhs
, lhs
);
1094 /* Return "lhs || rhs", with shortcut semantics.
1095 * That is, if lhs is true, then the result is defined even if rhs is not.
1096 * In practice, we compute lhs ? lhs : rhs.
1098 static __isl_give isl_pw_aff
*pw_aff_or_else(__isl_take isl_pw_aff
*lhs
,
1099 __isl_take isl_pw_aff
*rhs
)
1101 return isl_pw_aff_cond(isl_pw_aff_copy(lhs
), lhs
, rhs
);
1104 /* Extract an affine expressions representing the comparison "LHS op RHS"
1105 * "comp" is the original statement that "LHS op RHS" is derived from
1106 * and is used for diagnostics.
1108 * If the comparison is of the form
1112 * then the expression is constructed as the conjunction of
1117 * A similar optimization is performed for max(a,b) <= c.
1118 * We do this because that will lead to simpler representations
1119 * of the expression.
1120 * If isl is ever enhanced to explicitly deal with min and max expressions,
1121 * this optimization can be removed.
1123 __isl_give isl_pw_aff
*PetScan::extract_comparison(BinaryOperatorKind op
,
1124 Expr
*LHS
, Expr
*RHS
, Stmt
*comp
)
1133 return extract_comparison(BO_LT
, RHS
, LHS
, comp
);
1135 return extract_comparison(BO_LE
, RHS
, LHS
, comp
);
1137 if (op
== BO_LT
|| op
== BO_LE
) {
1138 Expr
*expr1
, *expr2
;
1139 if (is_min(RHS
, expr1
, expr2
)) {
1140 lhs
= extract_comparison(op
, LHS
, expr1
, comp
);
1141 rhs
= extract_comparison(op
, LHS
, expr2
, comp
);
1142 return pw_aff_and(lhs
, rhs
);
1144 if (is_max(LHS
, expr1
, expr2
)) {
1145 lhs
= extract_comparison(op
, expr1
, RHS
, comp
);
1146 rhs
= extract_comparison(op
, expr2
, RHS
, comp
);
1147 return pw_aff_and(lhs
, rhs
);
1151 lhs
= extract_affine(LHS
);
1152 rhs
= extract_affine(RHS
);
1154 dom
= isl_pw_aff_domain(isl_pw_aff_copy(lhs
));
1155 dom
= isl_set_intersect(dom
, isl_pw_aff_domain(isl_pw_aff_copy(rhs
)));
1159 cond
= isl_pw_aff_lt_set(lhs
, rhs
);
1162 cond
= isl_pw_aff_le_set(lhs
, rhs
);
1165 cond
= isl_pw_aff_eq_set(lhs
, rhs
);
1168 cond
= isl_pw_aff_ne_set(lhs
, rhs
);
1171 isl_pw_aff_free(lhs
);
1172 isl_pw_aff_free(rhs
);
1178 cond
= isl_set_coalesce(cond
);
1179 res
= indicator_function(cond
, dom
);
1184 __isl_give isl_pw_aff
*PetScan::extract_comparison(BinaryOperator
*comp
)
1186 return extract_comparison(comp
->getOpcode(), comp
->getLHS(),
1187 comp
->getRHS(), comp
);
1190 /* Extract an affine expression representing the negation (logical not)
1191 * of a subexpression.
1193 __isl_give isl_pw_aff
*PetScan::extract_boolean(UnaryOperator
*op
)
1195 isl_set
*set_cond
, *dom
;
1196 isl_pw_aff
*cond
, *res
;
1198 cond
= extract_condition(op
->getSubExpr());
1200 dom
= isl_pw_aff_domain(isl_pw_aff_copy(cond
));
1202 set_cond
= isl_pw_aff_zero_set(cond
);
1204 res
= indicator_function(set_cond
, dom
);
1209 /* Extract an affine expression representing the disjunction (logical or)
1210 * or conjunction (logical and) of two subexpressions.
1212 __isl_give isl_pw_aff
*PetScan::extract_boolean(BinaryOperator
*comp
)
1214 isl_pw_aff
*lhs
, *rhs
;
1216 lhs
= extract_condition(comp
->getLHS());
1217 rhs
= extract_condition(comp
->getRHS());
1219 switch (comp
->getOpcode()) {
1221 return pw_aff_and_then(lhs
, rhs
);
1223 return pw_aff_or_else(lhs
, rhs
);
1225 isl_pw_aff_free(lhs
);
1226 isl_pw_aff_free(rhs
);
1233 __isl_give isl_pw_aff
*PetScan::extract_condition(UnaryOperator
*expr
)
1235 switch (expr
->getOpcode()) {
1237 return extract_boolean(expr
);
1244 /* Extract the affine expression "expr != 0 ? 1 : 0".
1246 __isl_give isl_pw_aff
*PetScan::extract_implicit_condition(Expr
*expr
)
1251 res
= extract_affine(expr
);
1253 dom
= isl_pw_aff_domain(isl_pw_aff_copy(res
));
1254 set
= isl_pw_aff_non_zero_set(res
);
1256 res
= indicator_function(set
, dom
);
1261 /* Extract an affine expression from a boolean expression.
1262 * In particular, return the expression "expr ? 1 : 0".
1264 * If the expression doesn't look like a condition, we assume it
1265 * is an affine expression and return the condition "expr != 0 ? 1 : 0".
1267 __isl_give isl_pw_aff
*PetScan::extract_condition(Expr
*expr
)
1269 BinaryOperator
*comp
;
1272 isl_set
*u
= isl_set_universe(isl_space_params_alloc(ctx
, 0));
1273 return indicator_function(u
, isl_set_copy(u
));
1276 if (expr
->getStmtClass() == Stmt::ParenExprClass
)
1277 return extract_condition(cast
<ParenExpr
>(expr
)->getSubExpr());
1279 if (expr
->getStmtClass() == Stmt::UnaryOperatorClass
)
1280 return extract_condition(cast
<UnaryOperator
>(expr
));
1282 if (expr
->getStmtClass() != Stmt::BinaryOperatorClass
)
1283 return extract_implicit_condition(expr
);
1285 comp
= cast
<BinaryOperator
>(expr
);
1286 switch (comp
->getOpcode()) {
1293 return extract_comparison(comp
);
1296 return extract_boolean(comp
);
1298 return extract_implicit_condition(expr
);
1302 static enum pet_op_type
UnaryOperatorKind2pet_op_type(UnaryOperatorKind kind
)
1306 return pet_op_minus
;
1308 return pet_op_post_inc
;
1310 return pet_op_post_dec
;
1312 return pet_op_pre_inc
;
1314 return pet_op_pre_dec
;
1320 static enum pet_op_type
BinaryOperatorKind2pet_op_type(BinaryOperatorKind kind
)
1324 return pet_op_add_assign
;
1326 return pet_op_sub_assign
;
1328 return pet_op_mul_assign
;
1330 return pet_op_div_assign
;
1332 return pet_op_assign
;
1356 /* Construct a pet_expr representing a unary operator expression.
1358 struct pet_expr
*PetScan::extract_expr(UnaryOperator
*expr
)
1360 struct pet_expr
*arg
;
1361 enum pet_op_type op
;
1363 op
= UnaryOperatorKind2pet_op_type(expr
->getOpcode());
1364 if (op
== pet_op_last
) {
1369 arg
= extract_expr(expr
->getSubExpr());
1371 if (expr
->isIncrementDecrementOp() &&
1372 arg
&& arg
->type
== pet_expr_access
) {
1377 return pet_expr_new_unary(ctx
, op
, arg
);
1380 /* Mark the given access pet_expr as a write.
1381 * If a scalar is being accessed, then mark its value
1382 * as unknown in assigned_value.
1384 void PetScan::mark_write(struct pet_expr
*access
)
1392 access
->acc
.write
= 1;
1393 access
->acc
.read
= 0;
1395 if (!pet_expr_is_scalar_access(access
))
1398 id
= pet_expr_access_get_id(access
);
1399 decl
= (ValueDecl
*) isl_id_get_user(id
);
1400 clear_assignment(assigned_value
, decl
);
1404 /* Assign "rhs" to "lhs".
1406 * In particular, if "lhs" is a scalar variable, then mark
1407 * the variable as having been assigned. If, furthermore, "rhs"
1408 * is an affine expression, then keep track of this value in assigned_value
1409 * so that we can plug it in when we later come across the same variable.
1411 void PetScan::assign(struct pet_expr
*lhs
, Expr
*rhs
)
1419 if (!pet_expr_is_scalar_access(lhs
))
1422 id
= pet_expr_access_get_id(lhs
);
1423 decl
= (ValueDecl
*) isl_id_get_user(id
);
1426 pa
= try_extract_affine(rhs
);
1427 clear_assignment(assigned_value
, decl
);
1430 assigned_value
[decl
] = pa
;
1431 insert_expression(pa
);
1434 /* Construct a pet_expr representing a binary operator expression.
1436 * If the top level operator is an assignment and the LHS is an access,
1437 * then we mark that access as a write. If the operator is a compound
1438 * assignment, the access is marked as both a read and a write.
1440 * If "expr" assigns something to a scalar variable, then we mark
1441 * the variable as having been assigned. If, furthermore, the expression
1442 * is affine, then keep track of this value in assigned_value
1443 * so that we can plug it in when we later come across the same variable.
1445 struct pet_expr
*PetScan::extract_expr(BinaryOperator
*expr
)
1447 struct pet_expr
*lhs
, *rhs
;
1448 enum pet_op_type op
;
1450 op
= BinaryOperatorKind2pet_op_type(expr
->getOpcode());
1451 if (op
== pet_op_last
) {
1456 lhs
= extract_expr(expr
->getLHS());
1457 rhs
= extract_expr(expr
->getRHS());
1459 if (expr
->isAssignmentOp() && lhs
&& lhs
->type
== pet_expr_access
) {
1461 if (expr
->isCompoundAssignmentOp())
1465 if (expr
->getOpcode() == BO_Assign
)
1466 assign(lhs
, expr
->getRHS());
1468 return pet_expr_new_binary(ctx
, op
, lhs
, rhs
);
1471 /* Construct a pet_scop with a single statement killing the entire
1474 struct pet_scop
*PetScan::kill(Stmt
*stmt
, struct pet_array
*array
)
1477 struct pet_expr
*expr
;
1481 access
= isl_map_from_range(isl_set_copy(array
->extent
));
1482 expr
= pet_expr_kill_from_access(access
);
1483 return extract(stmt
, expr
);
1486 /* Construct a pet_scop for a (single) variable declaration.
1488 * The scop contains the variable being declared (as an array)
1489 * and a statement killing the array.
1491 * If the variable is initialized in the AST, then the scop
1492 * also contains an assignment to the variable.
1494 struct pet_scop
*PetScan::extract(DeclStmt
*stmt
)
1498 struct pet_expr
*lhs
, *rhs
, *pe
;
1499 struct pet_scop
*scop_decl
, *scop
;
1500 struct pet_array
*array
;
1502 if (!stmt
->isSingleDecl()) {
1507 decl
= stmt
->getSingleDecl();
1508 vd
= cast
<VarDecl
>(decl
);
1510 array
= extract_array(ctx
, vd
);
1512 array
->declared
= 1;
1513 scop_decl
= kill(stmt
, array
);
1514 scop_decl
= pet_scop_add_array(scop_decl
, array
);
1519 lhs
= pet_expr_from_access(extract_access(vd
));
1520 rhs
= extract_expr(vd
->getInit());
1523 assign(lhs
, vd
->getInit());
1525 pe
= pet_expr_new_binary(ctx
, pet_op_assign
, lhs
, rhs
);
1526 scop
= extract(stmt
, pe
);
1528 scop_decl
= pet_scop_prefix(scop_decl
, 0);
1529 scop
= pet_scop_prefix(scop
, 1);
1531 scop
= pet_scop_add_seq(ctx
, scop_decl
, scop
);
1536 /* Construct a pet_expr representing a conditional operation.
1538 * We first try to extract the condition as an affine expression.
1539 * If that fails, we construct a pet_expr tree representing the condition.
1541 struct pet_expr
*PetScan::extract_expr(ConditionalOperator
*expr
)
1543 struct pet_expr
*cond
, *lhs
, *rhs
;
1546 pa
= try_extract_affine(expr
->getCond());
1548 isl_set
*test
= isl_set_from_pw_aff(pa
);
1549 cond
= pet_expr_from_access(isl_map_from_range(test
));
1551 cond
= extract_expr(expr
->getCond());
1552 lhs
= extract_expr(expr
->getTrueExpr());
1553 rhs
= extract_expr(expr
->getFalseExpr());
1555 return pet_expr_new_ternary(ctx
, cond
, lhs
, rhs
);
1558 struct pet_expr
*PetScan::extract_expr(ImplicitCastExpr
*expr
)
1560 return extract_expr(expr
->getSubExpr());
1563 /* Construct a pet_expr representing a floating point value.
1565 * If the floating point literal does not appear in a macro,
1566 * then we use the original representation in the source code
1567 * as the string representation. Otherwise, we use the pretty
1568 * printer to produce a string representation.
1570 struct pet_expr
*PetScan::extract_expr(FloatingLiteral
*expr
)
1574 const LangOptions
&LO
= PP
.getLangOpts();
1575 SourceLocation loc
= expr
->getLocation();
1577 if (!loc
.isMacroID()) {
1578 SourceManager
&SM
= PP
.getSourceManager();
1579 unsigned len
= Lexer::MeasureTokenLength(loc
, SM
, LO
);
1580 s
= string(SM
.getCharacterData(loc
), len
);
1582 llvm::raw_string_ostream
S(s
);
1583 expr
->printPretty(S
, 0, PrintingPolicy(LO
));
1586 d
= expr
->getValueAsApproximateDouble();
1587 return pet_expr_new_double(ctx
, d
, s
.c_str());
1590 /* Extract an access relation from "expr" and then convert it into
1593 struct pet_expr
*PetScan::extract_access_expr(Expr
*expr
)
1596 struct pet_expr
*pe
;
1598 access
= extract_access(expr
);
1600 pe
= pet_expr_from_access(access
);
1605 struct pet_expr
*PetScan::extract_expr(ParenExpr
*expr
)
1607 return extract_expr(expr
->getSubExpr());
1610 /* Construct a pet_expr representing a function call.
1612 * If we are passing along a pointer to an array element
1613 * or an entire row or even higher dimensional slice of an array,
1614 * then the function being called may write into the array.
1616 * We assume here that if the function is declared to take a pointer
1617 * to a const type, then the function will perform a read
1618 * and that otherwise, it will perform a write.
1620 struct pet_expr
*PetScan::extract_expr(CallExpr
*expr
)
1622 struct pet_expr
*res
= NULL
;
1626 fd
= expr
->getDirectCallee();
1632 name
= fd
->getDeclName().getAsString();
1633 res
= pet_expr_new_call(ctx
, name
.c_str(), expr
->getNumArgs());
1637 for (int i
= 0; i
< expr
->getNumArgs(); ++i
) {
1638 Expr
*arg
= expr
->getArg(i
);
1642 if (arg
->getStmtClass() == Stmt::ImplicitCastExprClass
) {
1643 ImplicitCastExpr
*ice
= cast
<ImplicitCastExpr
>(arg
);
1644 arg
= ice
->getSubExpr();
1646 if (arg
->getStmtClass() == Stmt::UnaryOperatorClass
) {
1647 UnaryOperator
*op
= cast
<UnaryOperator
>(arg
);
1648 if (op
->getOpcode() == UO_AddrOf
) {
1650 arg
= op
->getSubExpr();
1653 res
->args
[i
] = PetScan::extract_expr(arg
);
1654 main_arg
= res
->args
[i
];
1656 res
->args
[i
] = pet_expr_new_unary(ctx
,
1657 pet_op_address_of
, res
->args
[i
]);
1660 if (arg
->getStmtClass() == Stmt::ArraySubscriptExprClass
&&
1661 array_depth(arg
->getType().getTypePtr()) > 0)
1663 if (is_addr
&& main_arg
->type
== pet_expr_access
) {
1665 if (!fd
->hasPrototype()) {
1666 unsupported(expr
, "prototype required");
1669 parm
= fd
->getParamDecl(i
);
1670 if (!const_base(parm
->getType()))
1671 mark_write(main_arg
);
1681 /* Construct a pet_expr representing a (C style) cast.
1683 struct pet_expr
*PetScan::extract_expr(CStyleCastExpr
*expr
)
1685 struct pet_expr
*arg
;
1688 arg
= extract_expr(expr
->getSubExpr());
1692 type
= expr
->getTypeAsWritten();
1693 return pet_expr_new_cast(ctx
, type
.getAsString().c_str(), arg
);
1696 /* Try and onstruct a pet_expr representing "expr".
1698 struct pet_expr
*PetScan::extract_expr(Expr
*expr
)
1700 switch (expr
->getStmtClass()) {
1701 case Stmt::UnaryOperatorClass
:
1702 return extract_expr(cast
<UnaryOperator
>(expr
));
1703 case Stmt::CompoundAssignOperatorClass
:
1704 case Stmt::BinaryOperatorClass
:
1705 return extract_expr(cast
<BinaryOperator
>(expr
));
1706 case Stmt::ImplicitCastExprClass
:
1707 return extract_expr(cast
<ImplicitCastExpr
>(expr
));
1708 case Stmt::ArraySubscriptExprClass
:
1709 case Stmt::DeclRefExprClass
:
1710 case Stmt::IntegerLiteralClass
:
1711 return extract_access_expr(expr
);
1712 case Stmt::FloatingLiteralClass
:
1713 return extract_expr(cast
<FloatingLiteral
>(expr
));
1714 case Stmt::ParenExprClass
:
1715 return extract_expr(cast
<ParenExpr
>(expr
));
1716 case Stmt::ConditionalOperatorClass
:
1717 return extract_expr(cast
<ConditionalOperator
>(expr
));
1718 case Stmt::CallExprClass
:
1719 return extract_expr(cast
<CallExpr
>(expr
));
1720 case Stmt::CStyleCastExprClass
:
1721 return extract_expr(cast
<CStyleCastExpr
>(expr
));
1728 /* Check if the given initialization statement is an assignment.
1729 * If so, return that assignment. Otherwise return NULL.
1731 BinaryOperator
*PetScan::initialization_assignment(Stmt
*init
)
1733 BinaryOperator
*ass
;
1735 if (init
->getStmtClass() != Stmt::BinaryOperatorClass
)
1738 ass
= cast
<BinaryOperator
>(init
);
1739 if (ass
->getOpcode() != BO_Assign
)
1745 /* Check if the given initialization statement is a declaration
1746 * of a single variable.
1747 * If so, return that declaration. Otherwise return NULL.
1749 Decl
*PetScan::initialization_declaration(Stmt
*init
)
1753 if (init
->getStmtClass() != Stmt::DeclStmtClass
)
1756 decl
= cast
<DeclStmt
>(init
);
1758 if (!decl
->isSingleDecl())
1761 return decl
->getSingleDecl();
1764 /* Given the assignment operator in the initialization of a for loop,
1765 * extract the induction variable, i.e., the (integer)variable being
1768 ValueDecl
*PetScan::extract_induction_variable(BinaryOperator
*init
)
1775 lhs
= init
->getLHS();
1776 if (lhs
->getStmtClass() != Stmt::DeclRefExprClass
) {
1781 ref
= cast
<DeclRefExpr
>(lhs
);
1782 decl
= ref
->getDecl();
1783 type
= decl
->getType().getTypePtr();
1785 if (!type
->isIntegerType()) {
1793 /* Given the initialization statement of a for loop and the single
1794 * declaration in this initialization statement,
1795 * extract the induction variable, i.e., the (integer) variable being
1798 VarDecl
*PetScan::extract_induction_variable(Stmt
*init
, Decl
*decl
)
1802 vd
= cast
<VarDecl
>(decl
);
1804 const QualType type
= vd
->getType();
1805 if (!type
->isIntegerType()) {
1810 if (!vd
->getInit()) {
1818 /* Check that op is of the form iv++ or iv--.
1819 * Return an affine expression "1" or "-1" accordingly.
1821 __isl_give isl_pw_aff
*PetScan::extract_unary_increment(
1822 clang::UnaryOperator
*op
, clang::ValueDecl
*iv
)
1829 if (!op
->isIncrementDecrementOp()) {
1834 sub
= op
->getSubExpr();
1835 if (sub
->getStmtClass() != Stmt::DeclRefExprClass
) {
1840 ref
= cast
<DeclRefExpr
>(sub
);
1841 if (ref
->getDecl() != iv
) {
1846 space
= isl_space_params_alloc(ctx
, 0);
1847 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
1849 if (op
->isIncrementOp())
1850 aff
= isl_aff_add_constant_si(aff
, 1);
1852 aff
= isl_aff_add_constant_si(aff
, -1);
1854 return isl_pw_aff_from_aff(aff
);
1857 /* If the isl_pw_aff on which isl_pw_aff_foreach_piece is called
1858 * has a single constant expression, then put this constant in *user.
1859 * The caller is assumed to have checked that this function will
1860 * be called exactly once.
1862 static int extract_cst(__isl_take isl_set
*set
, __isl_take isl_aff
*aff
,
1865 isl_val
**inc
= (isl_val
**)user
;
1868 if (isl_aff_is_cst(aff
))
1869 *inc
= isl_aff_get_constant_val(aff
);
1879 /* Check if op is of the form
1883 * and return inc as an affine expression.
1885 * We extract an affine expression from the RHS, subtract iv and return
1888 __isl_give isl_pw_aff
*PetScan::extract_binary_increment(BinaryOperator
*op
,
1889 clang::ValueDecl
*iv
)
1898 if (op
->getOpcode() != BO_Assign
) {
1904 if (lhs
->getStmtClass() != Stmt::DeclRefExprClass
) {
1909 ref
= cast
<DeclRefExpr
>(lhs
);
1910 if (ref
->getDecl() != iv
) {
1915 val
= extract_affine(op
->getRHS());
1917 id
= isl_id_alloc(ctx
, iv
->getName().str().c_str(), iv
);
1919 dim
= isl_space_params_alloc(ctx
, 1);
1920 dim
= isl_space_set_dim_id(dim
, isl_dim_param
, 0, id
);
1921 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
1922 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_param
, 0, 1);
1924 val
= isl_pw_aff_sub(val
, isl_pw_aff_from_aff(aff
));
1929 /* Check that op is of the form iv += cst or iv -= cst
1930 * and return an affine expression corresponding oto cst or -cst accordingly.
1932 __isl_give isl_pw_aff
*PetScan::extract_compound_increment(
1933 CompoundAssignOperator
*op
, clang::ValueDecl
*iv
)
1939 BinaryOperatorKind opcode
;
1941 opcode
= op
->getOpcode();
1942 if (opcode
!= BO_AddAssign
&& opcode
!= BO_SubAssign
) {
1946 if (opcode
== BO_SubAssign
)
1950 if (lhs
->getStmtClass() != Stmt::DeclRefExprClass
) {
1955 ref
= cast
<DeclRefExpr
>(lhs
);
1956 if (ref
->getDecl() != iv
) {
1961 val
= extract_affine(op
->getRHS());
1963 val
= isl_pw_aff_neg(val
);
1968 /* Check that the increment of the given for loop increments
1969 * (or decrements) the induction variable "iv" and return
1970 * the increment as an affine expression if successful.
1972 __isl_give isl_pw_aff
*PetScan::extract_increment(clang::ForStmt
*stmt
,
1975 Stmt
*inc
= stmt
->getInc();
1982 if (inc
->getStmtClass() == Stmt::UnaryOperatorClass
)
1983 return extract_unary_increment(cast
<UnaryOperator
>(inc
), iv
);
1984 if (inc
->getStmtClass() == Stmt::CompoundAssignOperatorClass
)
1985 return extract_compound_increment(
1986 cast
<CompoundAssignOperator
>(inc
), iv
);
1987 if (inc
->getStmtClass() == Stmt::BinaryOperatorClass
)
1988 return extract_binary_increment(cast
<BinaryOperator
>(inc
), iv
);
1994 /* Embed the given iteration domain in an extra outer loop
1995 * with induction variable "var".
1996 * If this variable appeared as a parameter in the constraints,
1997 * it is replaced by the new outermost dimension.
1999 static __isl_give isl_set
*embed(__isl_take isl_set
*set
,
2000 __isl_take isl_id
*var
)
2004 set
= isl_set_insert_dims(set
, isl_dim_set
, 0, 1);
2005 pos
= isl_set_find_dim_by_id(set
, isl_dim_param
, var
);
2007 set
= isl_set_equate(set
, isl_dim_param
, pos
, isl_dim_set
, 0);
2008 set
= isl_set_project_out(set
, isl_dim_param
, pos
, 1);
2015 /* Return those elements in the space of "cond" that come after
2016 * (based on "sign") an element in "cond".
2018 static __isl_give isl_set
*after(__isl_take isl_set
*cond
, int sign
)
2020 isl_map
*previous_to_this
;
2023 previous_to_this
= isl_map_lex_lt(isl_set_get_space(cond
));
2025 previous_to_this
= isl_map_lex_gt(isl_set_get_space(cond
));
2027 cond
= isl_set_apply(cond
, previous_to_this
);
2032 /* Create the infinite iteration domain
2034 * { [id] : id >= 0 }
2036 * If "scop" has an affine skip of type pet_skip_later,
2037 * then remove those iterations i that have an earlier iteration
2038 * where the skip condition is satisfied, meaning that iteration i
2040 * Since we are dealing with a loop without loop iterator,
2041 * the skip condition cannot refer to the current loop iterator and
2042 * so effectively, the returned set is of the form
2044 * { [0]; [id] : id >= 1 and not skip }
2046 static __isl_give isl_set
*infinite_domain(__isl_take isl_id
*id
,
2047 struct pet_scop
*scop
)
2049 isl_ctx
*ctx
= isl_id_get_ctx(id
);
2053 domain
= isl_set_nat_universe(isl_space_set_alloc(ctx
, 0, 1));
2054 domain
= isl_set_set_dim_id(domain
, isl_dim_set
, 0, id
);
2056 if (!pet_scop_has_affine_skip(scop
, pet_skip_later
))
2059 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_later
);
2060 skip
= embed(skip
, isl_id_copy(id
));
2061 skip
= isl_set_intersect(skip
, isl_set_copy(domain
));
2062 domain
= isl_set_subtract(domain
, after(skip
, 1));
2067 /* Create an identity mapping on the space containing "domain".
2069 static __isl_give isl_map
*identity_map(__isl_keep isl_set
*domain
)
2074 space
= isl_space_map_from_set(isl_set_get_space(domain
));
2075 id
= isl_map_identity(space
);
2080 /* Create a map that maps elements of a single-dimensional array "id_test"
2081 * to the previous element (according to "inc"), provided this element
2082 * belongs to "domain". That is, create the map
2084 * { id[x] -> id[x - inc] : x - inc in domain }
2086 static __isl_give isl_map
*map_to_previous(__isl_take isl_id
*id_test
,
2087 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
2090 isl_local_space
*ls
;
2094 space
= isl_set_get_space(domain
);
2095 ls
= isl_local_space_from_space(space
);
2096 aff
= isl_aff_var_on_domain(ls
, isl_dim_set
, 0);
2097 aff
= isl_aff_add_constant_val(aff
, isl_val_neg(inc
));
2098 prev
= isl_map_from_aff(aff
);
2099 prev
= isl_map_intersect_range(prev
, domain
);
2100 prev
= isl_map_set_tuple_id(prev
, isl_dim_out
, id_test
);
2105 /* Add an implication to "scop" expressing that if an element of
2106 * virtual array "id_test" has value "satisfied" then all previous elements
2107 * of this array also have that value. The set of previous elements
2108 * is bounded by "domain". If "sign" is negative then iterator
2109 * is decreasing and we express that all subsequent array elements
2110 * (but still defined previously) have the same value.
2112 static struct pet_scop
*add_implication(struct pet_scop
*scop
,
2113 __isl_take isl_id
*id_test
, __isl_take isl_set
*domain
, int sign
,
2119 domain
= isl_set_set_tuple_id(domain
, id_test
);
2120 space
= isl_set_get_space(domain
);
2122 map
= isl_map_lex_ge(space
);
2124 map
= isl_map_lex_le(space
);
2125 map
= isl_map_intersect_range(map
, domain
);
2126 scop
= pet_scop_add_implication(scop
, map
, satisfied
);
2131 /* Add a filter to "scop" that imposes that it is only executed
2132 * when the variable identified by "id_test" has a zero value
2133 * for all previous iterations of "domain".
2135 * In particular, add a filter that imposes that the array
2136 * has a zero value at the previous iteration of domain and
2137 * add an implication that implies that it then has that
2138 * value for all previous iterations.
2140 static struct pet_scop
*scop_add_break(struct pet_scop
*scop
,
2141 __isl_take isl_id
*id_test
, __isl_take isl_set
*domain
,
2142 __isl_take isl_val
*inc
)
2145 int sign
= isl_val_sgn(inc
);
2147 prev
= map_to_previous(isl_id_copy(id_test
), isl_set_copy(domain
), inc
);
2148 scop
= add_implication(scop
, id_test
, domain
, sign
, 0);
2149 scop
= pet_scop_filter(scop
, prev
, 0);
2154 /* Construct a pet_scop for an infinite loop around the given body.
2156 * We extract a pet_scop for the body and then embed it in a loop with
2165 * If the body contains any break, then it is taken into
2166 * account in infinite_domain (if the skip condition is affine)
2167 * or in scop_add_break (if the skip condition is not affine).
2169 struct pet_scop
*PetScan::extract_infinite_loop(Stmt
*body
)
2171 isl_id
*id
, *id_test
;
2174 struct pet_scop
*scop
;
2177 scop
= extract(body
);
2181 id
= isl_id_alloc(ctx
, "t", NULL
);
2182 domain
= infinite_domain(isl_id_copy(id
), scop
);
2183 ident
= identity_map(domain
);
2185 has_var_break
= pet_scop_has_var_skip(scop
, pet_skip_later
);
2187 id_test
= pet_scop_get_skip_id(scop
, pet_skip_later
);
2189 scop
= pet_scop_embed(scop
, isl_set_copy(domain
),
2190 isl_map_copy(ident
), ident
, id
);
2192 scop
= scop_add_break(scop
, id_test
, domain
, isl_val_one(ctx
));
2194 isl_set_free(domain
);
2199 /* Construct a pet_scop for an infinite loop, i.e., a loop of the form
2205 struct pet_scop
*PetScan::extract_infinite_for(ForStmt
*stmt
)
2207 return extract_infinite_loop(stmt
->getBody());
2210 /* Create an access to a virtual array representing the result
2212 * Unlike other accessed data, the id of the array is NULL as
2213 * there is no ValueDecl in the program corresponding to the virtual
2215 * The array starts out as a scalar, but grows along with the
2216 * statement writing to the array in pet_scop_embed.
2218 static __isl_give isl_map
*create_test_access(isl_ctx
*ctx
, int test_nr
)
2220 isl_space
*dim
= isl_space_alloc(ctx
, 0, 0, 0);
2224 snprintf(name
, sizeof(name
), "__pet_test_%d", test_nr
);
2225 id
= isl_id_alloc(ctx
, name
, NULL
);
2226 dim
= isl_space_set_tuple_id(dim
, isl_dim_out
, id
);
2227 return isl_map_universe(dim
);
2230 /* Add an array with the given extent ("access") to the list
2231 * of arrays in "scop" and return the extended pet_scop.
2232 * The array is marked as attaining values 0 and 1 only and
2233 * as each element being assigned at most once.
2235 static struct pet_scop
*scop_add_array(struct pet_scop
*scop
,
2236 __isl_keep isl_map
*access
, clang::ASTContext
&ast_ctx
)
2238 isl_ctx
*ctx
= isl_map_get_ctx(access
);
2240 struct pet_array
*array
;
2247 array
= isl_calloc_type(ctx
, struct pet_array
);
2251 array
->extent
= isl_map_range(isl_map_copy(access
));
2252 dim
= isl_space_params_alloc(ctx
, 0);
2253 array
->context
= isl_set_universe(dim
);
2254 dim
= isl_space_set_alloc(ctx
, 0, 1);
2255 array
->value_bounds
= isl_set_universe(dim
);
2256 array
->value_bounds
= isl_set_lower_bound_si(array
->value_bounds
,
2258 array
->value_bounds
= isl_set_upper_bound_si(array
->value_bounds
,
2260 array
->element_type
= strdup("int");
2261 array
->element_size
= ast_ctx
.getTypeInfo(ast_ctx
.IntTy
).first
/ 8;
2262 array
->uniquely_defined
= 1;
2264 if (!array
->extent
|| !array
->context
)
2265 array
= pet_array_free(array
);
2267 scop
= pet_scop_add_array(scop
, array
);
2271 pet_scop_free(scop
);
2275 /* Construct a pet_scop for a while loop of the form
2280 * In particular, construct a scop for an infinite loop around body and
2281 * intersect the domain with the affine expression.
2282 * Note that this intersection may result in an empty loop.
2284 struct pet_scop
*PetScan::extract_affine_while(__isl_take isl_pw_aff
*pa
,
2287 struct pet_scop
*scop
;
2291 valid
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
2292 dom
= isl_pw_aff_non_zero_set(pa
);
2293 scop
= extract_infinite_loop(body
);
2294 scop
= pet_scop_restrict(scop
, dom
);
2295 scop
= pet_scop_restrict_context(scop
, valid
);
2300 /* Construct a scop for a while, given the scops for the condition
2301 * and the body, the filter identifier and the iteration domain of
2304 * In particular, the scop for the condition is filtered to depend
2305 * on "id_test" evaluating to true for all previous iterations
2306 * of the loop, while the scop for the body is filtered to depend
2307 * on "id_test" evaluating to true for all iterations up to the
2308 * current iteration.
2309 * The actual filter only imposes that this virtual array has
2310 * value one on the previous or the current iteration.
2311 * The fact that this condition also applies to the previous
2312 * iterations is enforced by an implication.
2314 * These filtered scops are then combined into a single scop.
2316 * "sign" is positive if the iterator increases and negative
2319 static struct pet_scop
*scop_add_while(struct pet_scop
*scop_cond
,
2320 struct pet_scop
*scop_body
, __isl_take isl_id
*id_test
,
2321 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
2323 isl_ctx
*ctx
= isl_set_get_ctx(domain
);
2325 isl_map
*test_access
;
2327 int sign
= isl_val_sgn(inc
);
2328 struct pet_scop
*scop
;
2330 prev
= map_to_previous(isl_id_copy(id_test
), isl_set_copy(domain
), inc
);
2331 scop_cond
= pet_scop_filter(scop_cond
, prev
, 1);
2333 space
= isl_space_map_from_set(isl_set_get_space(domain
));
2334 test_access
= isl_map_identity(space
);
2335 test_access
= isl_map_set_tuple_id(test_access
, isl_dim_out
,
2336 isl_id_copy(id_test
));
2337 scop_body
= pet_scop_filter(scop_body
, test_access
, 1);
2339 scop
= pet_scop_add_seq(ctx
, scop_cond
, scop_body
);
2340 scop
= add_implication(scop
, id_test
, domain
, sign
, 1);
2345 /* Check if the while loop is of the form
2347 * while (affine expression)
2350 * If so, call extract_affine_while to construct a scop.
2352 * Otherwise, construct a generic while scop, with iteration domain
2353 * { [t] : t >= 0 }. The scop consists of two parts, one for
2354 * evaluating the condition and one for the body.
2355 * The schedule is adjusted to reflect that the condition is evaluated
2356 * before the body is executed and the body is filtered to depend
2357 * on the result of the condition evaluating to true on all iterations
2358 * up to the current iteration, while the evaluation the condition itself
2359 * is filtered to depend on the result of the condition evaluating to true
2360 * on all previous iterations.
2361 * The context of the scop representing the body is dropped
2362 * because we don't know how many times the body will be executed,
2365 * If the body contains any break, then it is taken into
2366 * account in infinite_domain (if the skip condition is affine)
2367 * or in scop_add_break (if the skip condition is not affine).
2369 struct pet_scop
*PetScan::extract(WhileStmt
*stmt
)
2372 isl_id
*id
, *id_test
, *id_break_test
;
2373 isl_map
*test_access
;
2377 struct pet_scop
*scop
, *scop_body
;
2380 cond
= stmt
->getCond();
2386 clear_assignments
clear(assigned_value
);
2387 clear
.TraverseStmt(stmt
->getBody());
2389 pa
= try_extract_affine_condition(cond
);
2391 return extract_affine_while(pa
, stmt
->getBody());
2393 if (!allow_nested
) {
2398 test_access
= create_test_access(ctx
, n_test
++);
2399 scop
= extract_non_affine_condition(cond
, isl_map_copy(test_access
));
2400 scop
= scop_add_array(scop
, test_access
, ast_context
);
2401 id_test
= isl_map_get_tuple_id(test_access
, isl_dim_out
);
2402 isl_map_free(test_access
);
2403 scop_body
= extract(stmt
->getBody());
2405 id
= isl_id_alloc(ctx
, "t", NULL
);
2406 domain
= infinite_domain(isl_id_copy(id
), scop_body
);
2407 ident
= identity_map(domain
);
2409 has_var_break
= pet_scop_has_var_skip(scop_body
, pet_skip_later
);
2411 id_break_test
= pet_scop_get_skip_id(scop_body
, pet_skip_later
);
2413 scop
= pet_scop_prefix(scop
, 0);
2414 scop
= pet_scop_embed(scop
, isl_set_copy(domain
), isl_map_copy(ident
),
2415 isl_map_copy(ident
), isl_id_copy(id
));
2416 scop_body
= pet_scop_reset_context(scop_body
);
2417 scop_body
= pet_scop_prefix(scop_body
, 1);
2418 scop_body
= pet_scop_embed(scop_body
, isl_set_copy(domain
),
2419 isl_map_copy(ident
), ident
, id
);
2421 if (has_var_break
) {
2422 scop
= scop_add_break(scop
, isl_id_copy(id_break_test
),
2423 isl_set_copy(domain
), isl_val_one(ctx
));
2424 scop_body
= scop_add_break(scop_body
, id_break_test
,
2425 isl_set_copy(domain
), isl_val_one(ctx
));
2427 scop
= scop_add_while(scop
, scop_body
, id_test
, domain
,
2433 /* Check whether "cond" expresses a simple loop bound
2434 * on the only set dimension.
2435 * In particular, if "up" is set then "cond" should contain only
2436 * upper bounds on the set dimension.
2437 * Otherwise, it should contain only lower bounds.
2439 static bool is_simple_bound(__isl_keep isl_set
*cond
, __isl_keep isl_val
*inc
)
2441 if (isl_val_is_pos(inc
))
2442 return !isl_set_dim_has_any_lower_bound(cond
, isl_dim_set
, 0);
2444 return !isl_set_dim_has_any_upper_bound(cond
, isl_dim_set
, 0);
2447 /* Extend a condition on a given iteration of a loop to one that
2448 * imposes the same condition on all previous iterations.
2449 * "domain" expresses the lower [upper] bound on the iterations
2450 * when inc is positive [negative].
2452 * In particular, we construct the condition (when inc is positive)
2454 * forall i' : (domain(i') and i' <= i) => cond(i')
2456 * which is equivalent to
2458 * not exists i' : domain(i') and i' <= i and not cond(i')
2460 * We construct this set by negating cond, applying a map
2462 * { [i'] -> [i] : domain(i') and i' <= i }
2464 * and then negating the result again.
2466 static __isl_give isl_set
*valid_for_each_iteration(__isl_take isl_set
*cond
,
2467 __isl_take isl_set
*domain
, __isl_take isl_val
*inc
)
2469 isl_map
*previous_to_this
;
2471 if (isl_val_is_pos(inc
))
2472 previous_to_this
= isl_map_lex_le(isl_set_get_space(domain
));
2474 previous_to_this
= isl_map_lex_ge(isl_set_get_space(domain
));
2476 previous_to_this
= isl_map_intersect_domain(previous_to_this
, domain
);
2478 cond
= isl_set_complement(cond
);
2479 cond
= isl_set_apply(cond
, previous_to_this
);
2480 cond
= isl_set_complement(cond
);
2487 /* Construct a domain of the form
2489 * [id] -> { : exists a: id = init + a * inc and a >= 0 }
2491 static __isl_give isl_set
*strided_domain(__isl_take isl_id
*id
,
2492 __isl_take isl_pw_aff
*init
, __isl_take isl_val
*inc
)
2498 init
= isl_pw_aff_insert_dims(init
, isl_dim_in
, 0, 1);
2499 dim
= isl_pw_aff_get_domain_space(init
);
2500 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
2501 aff
= isl_aff_add_coefficient_val(aff
, isl_dim_in
, 0, inc
);
2502 init
= isl_pw_aff_add(init
, isl_pw_aff_from_aff(aff
));
2504 dim
= isl_space_set_alloc(isl_pw_aff_get_ctx(init
), 1, 1);
2505 dim
= isl_space_set_dim_id(dim
, isl_dim_param
, 0, id
);
2506 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
2507 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_param
, 0, 1);
2509 set
= isl_pw_aff_eq_set(isl_pw_aff_from_aff(aff
), init
);
2511 set
= isl_set_lower_bound_si(set
, isl_dim_set
, 0, 0);
2513 return isl_set_params(set
);
2516 /* Assuming "cond" represents a bound on a loop where the loop
2517 * iterator "iv" is incremented (or decremented) by one, check if wrapping
2520 * Under the given assumptions, wrapping is only possible if "cond" allows
2521 * for the last value before wrapping, i.e., 2^width - 1 in case of an
2522 * increasing iterator and 0 in case of a decreasing iterator.
2524 static bool can_wrap(__isl_keep isl_set
*cond
, ValueDecl
*iv
,
2525 __isl_keep isl_val
*inc
)
2532 test
= isl_set_copy(cond
);
2534 ctx
= isl_set_get_ctx(test
);
2535 if (isl_val_is_neg(inc
))
2536 limit
= isl_val_zero(ctx
);
2538 limit
= isl_val_int_from_ui(ctx
, get_type_size(iv
));
2539 limit
= isl_val_2exp(limit
);
2540 limit
= isl_val_sub_ui(limit
, 1);
2543 test
= isl_set_fix_val(cond
, isl_dim_set
, 0, limit
);
2544 cw
= !isl_set_is_empty(test
);
2550 /* Given a one-dimensional space, construct the following mapping on this
2553 * { [v] -> [v mod 2^width] }
2555 * where width is the number of bits used to represent the values
2556 * of the unsigned variable "iv".
2558 static __isl_give isl_map
*compute_wrapping(__isl_take isl_space
*dim
,
2566 ctx
= isl_space_get_ctx(dim
);
2567 mod
= isl_val_int_from_ui(ctx
, get_type_size(iv
));
2568 mod
= isl_val_2exp(mod
);
2570 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
2571 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, 0, 1);
2572 aff
= isl_aff_mod_val(aff
, mod
);
2574 return isl_map_from_basic_map(isl_basic_map_from_aff(aff
));
2577 /* Project out the parameter "id" from "set".
2579 static __isl_give isl_set
*set_project_out_by_id(__isl_take isl_set
*set
,
2580 __isl_keep isl_id
*id
)
2584 pos
= isl_set_find_dim_by_id(set
, isl_dim_param
, id
);
2586 set
= isl_set_project_out(set
, isl_dim_param
, pos
, 1);
2591 /* Compute the set of parameters for which "set1" is a subset of "set2".
2593 * set1 is a subset of set2 if
2595 * forall i in set1 : i in set2
2599 * not exists i in set1 and i not in set2
2603 * not exists i in set1 \ set2
2605 static __isl_give isl_set
*enforce_subset(__isl_take isl_set
*set1
,
2606 __isl_take isl_set
*set2
)
2608 return isl_set_complement(isl_set_params(isl_set_subtract(set1
, set2
)));
2611 /* Compute the set of parameter values for which "cond" holds
2612 * on the next iteration for each element of "dom".
2614 * We first construct mapping { [i] -> [i + inc] }, apply that to "dom"
2615 * and then compute the set of parameters for which the result is a subset
2618 static __isl_give isl_set
*valid_on_next(__isl_take isl_set
*cond
,
2619 __isl_take isl_set
*dom
, __isl_take isl_val
*inc
)
2625 space
= isl_set_get_space(dom
);
2626 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
2627 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, 0, 1);
2628 aff
= isl_aff_add_constant_val(aff
, inc
);
2629 next
= isl_map_from_basic_map(isl_basic_map_from_aff(aff
));
2631 dom
= isl_set_apply(dom
, next
);
2633 return enforce_subset(dom
, cond
);
2636 /* Does "id" refer to a nested access?
2638 static bool is_nested_parameter(__isl_keep isl_id
*id
)
2640 return id
&& isl_id_get_user(id
) && !isl_id_get_name(id
);
2643 /* Does parameter "pos" of "space" refer to a nested access?
2645 static bool is_nested_parameter(__isl_keep isl_space
*space
, int pos
)
2650 id
= isl_space_get_dim_id(space
, isl_dim_param
, pos
);
2651 nested
= is_nested_parameter(id
);
2657 /* Does "space" involve any parameters that refer to nested
2658 * accesses, i.e., parameters with no name?
2660 static bool has_nested(__isl_keep isl_space
*space
)
2664 nparam
= isl_space_dim(space
, isl_dim_param
);
2665 for (int i
= 0; i
< nparam
; ++i
)
2666 if (is_nested_parameter(space
, i
))
2672 /* Does "pa" involve any parameters that refer to nested
2673 * accesses, i.e., parameters with no name?
2675 static bool has_nested(__isl_keep isl_pw_aff
*pa
)
2680 space
= isl_pw_aff_get_space(pa
);
2681 nested
= has_nested(space
);
2682 isl_space_free(space
);
2687 /* Construct a pet_scop for a for statement.
2688 * The for loop is required to be of the form
2690 * for (i = init; condition; ++i)
2694 * for (i = init; condition; --i)
2696 * The initialization of the for loop should either be an assignment
2697 * to an integer variable, or a declaration of such a variable with
2700 * The condition is allowed to contain nested accesses, provided
2701 * they are not being written to inside the body of the loop.
2702 * Otherwise, or if the condition is otherwise non-affine, the for loop is
2703 * essentially treated as a while loop, with iteration domain
2704 * { [i] : i >= init }.
2706 * We extract a pet_scop for the body and then embed it in a loop with
2707 * iteration domain and schedule
2709 * { [i] : i >= init and condition' }
2714 * { [i] : i <= init and condition' }
2717 * Where condition' is equal to condition if the latter is
2718 * a simple upper [lower] bound and a condition that is extended
2719 * to apply to all previous iterations otherwise.
2721 * If the condition is non-affine, then we drop the condition from the
2722 * iteration domain and instead create a separate statement
2723 * for evaluating the condition. The body is then filtered to depend
2724 * on the result of the condition evaluating to true on all iterations
2725 * up to the current iteration, while the evaluation the condition itself
2726 * is filtered to depend on the result of the condition evaluating to true
2727 * on all previous iterations.
2728 * The context of the scop representing the body is dropped
2729 * because we don't know how many times the body will be executed,
2732 * If the stride of the loop is not 1, then "i >= init" is replaced by
2734 * (exists a: i = init + stride * a and a >= 0)
2736 * If the loop iterator i is unsigned, then wrapping may occur.
2737 * During the computation, we work with a virtual iterator that
2738 * does not wrap. However, the condition in the code applies
2739 * to the wrapped value, so we need to change condition(i)
2740 * into condition([i % 2^width]).
2741 * After computing the virtual domain and schedule, we apply
2742 * the function { [v] -> [v % 2^width] } to the domain and the domain
2743 * of the schedule. In order not to lose any information, we also
2744 * need to intersect the domain of the schedule with the virtual domain
2745 * first, since some iterations in the wrapped domain may be scheduled
2746 * several times, typically an infinite number of times.
2747 * Note that there may be no need to perform this final wrapping
2748 * if the loop condition (after wrapping) satisfies certain conditions.
2749 * However, the is_simple_bound condition is not enough since it doesn't
2750 * check if there even is an upper bound.
2752 * If the loop condition is non-affine, then we keep the virtual
2753 * iterator in the iteration domain and instead replace all accesses
2754 * to the original iterator by the wrapping of the virtual iterator.
2756 * Wrapping on unsigned iterators can be avoided entirely if
2757 * loop condition is simple, the loop iterator is incremented
2758 * [decremented] by one and the last value before wrapping cannot
2759 * possibly satisfy the loop condition.
2761 * Before extracting a pet_scop from the body we remove all
2762 * assignments in assigned_value to variables that are assigned
2763 * somewhere in the body of the loop.
2765 * Valid parameters for a for loop are those for which the initial
2766 * value itself, the increment on each domain iteration and
2767 * the condition on both the initial value and
2768 * the result of incrementing the iterator for each iteration of the domain
2770 * If the loop condition is non-affine, then we only consider validity
2771 * of the initial value.
2773 * If the body contains any break, then we keep track of it in "skip"
2774 * (if the skip condition is affine) or it is handled in scop_add_break
2775 * (if the skip condition is not affine).
2776 * Note that the affine break condition needs to be considered with
2777 * respect to previous iterations in the virtual domain (if any)
2778 * and that the domain needs to be kept virtual if there is a non-affine
2781 struct pet_scop
*PetScan::extract_for(ForStmt
*stmt
)
2783 BinaryOperator
*ass
;
2791 isl_set
*cond
= NULL
;
2792 isl_set
*skip
= NULL
;
2793 isl_id
*id
, *id_test
= NULL
, *id_break_test
;
2794 struct pet_scop
*scop
, *scop_cond
= NULL
;
2795 assigned_value_cache
cache(assigned_value
);
2801 bool keep_virtual
= false;
2802 bool has_affine_break
;
2804 isl_map
*wrap
= NULL
;
2805 isl_pw_aff
*pa
, *pa_inc
, *init_val
;
2806 isl_set
*valid_init
;
2807 isl_set
*valid_cond
;
2808 isl_set
*valid_cond_init
;
2809 isl_set
*valid_cond_next
;
2813 if (!stmt
->getInit() && !stmt
->getCond() && !stmt
->getInc())
2814 return extract_infinite_for(stmt
);
2816 init
= stmt
->getInit();
2821 if ((ass
= initialization_assignment(init
)) != NULL
) {
2822 iv
= extract_induction_variable(ass
);
2825 lhs
= ass
->getLHS();
2826 rhs
= ass
->getRHS();
2827 } else if ((decl
= initialization_declaration(init
)) != NULL
) {
2828 VarDecl
*var
= extract_induction_variable(init
, decl
);
2832 rhs
= var
->getInit();
2833 lhs
= create_DeclRefExpr(var
);
2835 unsupported(stmt
->getInit());
2839 pa_inc
= extract_increment(stmt
, iv
);
2844 if (isl_pw_aff_n_piece(pa_inc
) != 1 ||
2845 isl_pw_aff_foreach_piece(pa_inc
, &extract_cst
, &inc
) < 0) {
2846 isl_pw_aff_free(pa_inc
);
2847 unsupported(stmt
->getInc());
2851 valid_inc
= isl_pw_aff_domain(pa_inc
);
2853 is_unsigned
= iv
->getType()->isUnsignedIntegerType();
2855 assigned_value
.erase(iv
);
2856 clear_assignments
clear(assigned_value
);
2857 clear
.TraverseStmt(stmt
->getBody());
2859 id
= isl_id_alloc(ctx
, iv
->getName().str().c_str(), iv
);
2861 pa
= try_extract_nested_condition(stmt
->getCond());
2862 if (allow_nested
&& (!pa
|| has_nested(pa
)))
2865 scop
= extract(stmt
->getBody());
2867 has_affine_break
= scop
&&
2868 pet_scop_has_affine_skip(scop
, pet_skip_later
);
2869 if (has_affine_break
)
2870 skip
= pet_scop_get_affine_skip_domain(scop
, pet_skip_later
);
2871 has_var_break
= scop
&& pet_scop_has_var_skip(scop
, pet_skip_later
);
2872 if (has_var_break
) {
2873 id_break_test
= pet_scop_get_skip_id(scop
, pet_skip_later
);
2874 keep_virtual
= true;
2877 if (pa
&& !is_nested_allowed(pa
, scop
)) {
2878 isl_pw_aff_free(pa
);
2882 if (!allow_nested
&& !pa
)
2883 pa
= try_extract_affine_condition(stmt
->getCond());
2884 valid_cond
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
2885 cond
= isl_pw_aff_non_zero_set(pa
);
2886 if (allow_nested
&& !cond
) {
2887 isl_map
*test_access
;
2888 int save_n_stmt
= n_stmt
;
2889 test_access
= create_test_access(ctx
, n_test
++);
2891 scop_cond
= extract_non_affine_condition(stmt
->getCond(),
2892 isl_map_copy(test_access
));
2893 n_stmt
= save_n_stmt
;
2894 scop_cond
= scop_add_array(scop_cond
, test_access
, ast_context
);
2895 id_test
= isl_map_get_tuple_id(test_access
, isl_dim_out
);
2896 isl_map_free(test_access
);
2897 scop_cond
= pet_scop_prefix(scop_cond
, 0);
2898 scop
= pet_scop_reset_context(scop
);
2899 scop
= pet_scop_prefix(scop
, 1);
2900 keep_virtual
= true;
2901 cond
= isl_set_universe(isl_space_set_alloc(ctx
, 0, 0));
2904 cond
= embed(cond
, isl_id_copy(id
));
2905 skip
= embed(skip
, isl_id_copy(id
));
2906 valid_cond
= isl_set_coalesce(valid_cond
);
2907 valid_cond
= embed(valid_cond
, isl_id_copy(id
));
2908 valid_inc
= embed(valid_inc
, isl_id_copy(id
));
2909 is_one
= isl_val_is_one(inc
) || isl_val_is_negone(inc
);
2910 is_virtual
= is_unsigned
&& (!is_one
|| can_wrap(cond
, iv
, inc
));
2912 init_val
= extract_affine(rhs
);
2913 valid_cond_init
= enforce_subset(
2914 isl_set_from_pw_aff(isl_pw_aff_copy(init_val
)),
2915 isl_set_copy(valid_cond
));
2916 if (is_one
&& !is_virtual
) {
2917 isl_pw_aff_free(init_val
);
2918 pa
= extract_comparison(isl_val_is_pos(inc
) ? BO_GE
: BO_LE
,
2920 valid_init
= isl_pw_aff_domain(isl_pw_aff_copy(pa
));
2921 valid_init
= set_project_out_by_id(valid_init
, id
);
2922 domain
= isl_pw_aff_non_zero_set(pa
);
2924 valid_init
= isl_pw_aff_domain(isl_pw_aff_copy(init_val
));
2925 domain
= strided_domain(isl_id_copy(id
), init_val
,
2929 domain
= embed(domain
, isl_id_copy(id
));
2932 wrap
= compute_wrapping(isl_set_get_space(cond
), iv
);
2933 rev_wrap
= isl_map_reverse(isl_map_copy(wrap
));
2934 cond
= isl_set_apply(cond
, isl_map_copy(rev_wrap
));
2935 skip
= isl_set_apply(skip
, isl_map_copy(rev_wrap
));
2936 valid_cond
= isl_set_apply(valid_cond
, isl_map_copy(rev_wrap
));
2937 valid_inc
= isl_set_apply(valid_inc
, rev_wrap
);
2939 is_simple
= is_simple_bound(cond
, inc
);
2941 cond
= isl_set_gist(cond
, isl_set_copy(domain
));
2942 is_simple
= is_simple_bound(cond
, inc
);
2945 cond
= valid_for_each_iteration(cond
,
2946 isl_set_copy(domain
), isl_val_copy(inc
));
2947 domain
= isl_set_intersect(domain
, cond
);
2948 if (has_affine_break
) {
2949 skip
= isl_set_intersect(skip
, isl_set_copy(domain
));
2950 skip
= after(skip
, isl_val_sgn(inc
));
2951 domain
= isl_set_subtract(domain
, skip
);
2953 domain
= isl_set_set_dim_id(domain
, isl_dim_set
, 0, isl_id_copy(id
));
2954 space
= isl_space_from_domain(isl_set_get_space(domain
));
2955 space
= isl_space_add_dims(space
, isl_dim_out
, 1);
2956 sched
= isl_map_universe(space
);
2957 if (isl_val_is_pos(inc
))
2958 sched
= isl_map_equate(sched
, isl_dim_in
, 0, isl_dim_out
, 0);
2960 sched
= isl_map_oppose(sched
, isl_dim_in
, 0, isl_dim_out
, 0);
2962 valid_cond_next
= valid_on_next(valid_cond
, isl_set_copy(domain
),
2964 valid_inc
= enforce_subset(isl_set_copy(domain
), valid_inc
);
2966 if (is_virtual
&& !keep_virtual
) {
2967 wrap
= isl_map_set_dim_id(wrap
,
2968 isl_dim_out
, 0, isl_id_copy(id
));
2969 sched
= isl_map_intersect_domain(sched
, isl_set_copy(domain
));
2970 domain
= isl_set_apply(domain
, isl_map_copy(wrap
));
2971 sched
= isl_map_apply_domain(sched
, wrap
);
2973 if (!(is_virtual
&& keep_virtual
))
2974 wrap
= identity_map(domain
);
2976 scop_cond
= pet_scop_embed(scop_cond
, isl_set_copy(domain
),
2977 isl_map_copy(sched
), isl_map_copy(wrap
), isl_id_copy(id
));
2978 scop
= pet_scop_embed(scop
, isl_set_copy(domain
), sched
, wrap
, id
);
2979 scop
= resolve_nested(scop
);
2981 scop
= scop_add_break(scop
, id_break_test
, isl_set_copy(domain
),
2984 scop
= scop_add_while(scop_cond
, scop
, id_test
, domain
,
2986 isl_set_free(valid_inc
);
2988 scop
= pet_scop_restrict_context(scop
, valid_inc
);
2989 scop
= pet_scop_restrict_context(scop
, valid_cond_next
);
2990 scop
= pet_scop_restrict_context(scop
, valid_cond_init
);
2991 isl_set_free(domain
);
2993 clear_assignment(assigned_value
, iv
);
2997 scop
= pet_scop_restrict_context(scop
, valid_init
);
3002 struct pet_scop
*PetScan::extract(CompoundStmt
*stmt
, bool skip_declarations
)
3004 return extract(stmt
->children(), true, skip_declarations
);
3007 /* Does parameter "pos" of "map" refer to a nested access?
3009 static bool is_nested_parameter(__isl_keep isl_map
*map
, int pos
)
3014 id
= isl_map_get_dim_id(map
, isl_dim_param
, pos
);
3015 nested
= is_nested_parameter(id
);
3021 /* How many parameters of "space" refer to nested accesses, i.e., have no name?
3023 static int n_nested_parameter(__isl_keep isl_space
*space
)
3028 nparam
= isl_space_dim(space
, isl_dim_param
);
3029 for (int i
= 0; i
< nparam
; ++i
)
3030 if (is_nested_parameter(space
, i
))
3036 /* How many parameters of "map" refer to nested accesses, i.e., have no name?
3038 static int n_nested_parameter(__isl_keep isl_map
*map
)
3043 space
= isl_map_get_space(map
);
3044 n
= n_nested_parameter(space
);
3045 isl_space_free(space
);
3050 /* For each nested access parameter in "space",
3051 * construct a corresponding pet_expr, place it in args and
3052 * record its position in "param2pos".
3053 * "n_arg" is the number of elements that are already in args.
3054 * The position recorded in "param2pos" takes this number into account.
3055 * If the pet_expr corresponding to a parameter is identical to
3056 * the pet_expr corresponding to an earlier parameter, then these two
3057 * parameters are made to refer to the same element in args.
3059 * Return the final number of elements in args or -1 if an error has occurred.
3061 int PetScan::extract_nested(__isl_keep isl_space
*space
,
3062 int n_arg
, struct pet_expr
**args
, std::map
<int,int> ¶m2pos
)
3066 nparam
= isl_space_dim(space
, isl_dim_param
);
3067 for (int i
= 0; i
< nparam
; ++i
) {
3069 isl_id
*id
= isl_space_get_dim_id(space
, isl_dim_param
, i
);
3072 if (!is_nested_parameter(id
)) {
3077 nested
= (Expr
*) isl_id_get_user(id
);
3078 args
[n_arg
] = extract_expr(nested
);
3082 for (j
= 0; j
< n_arg
; ++j
)
3083 if (pet_expr_is_equal(args
[j
], args
[n_arg
]))
3087 pet_expr_free(args
[n_arg
]);
3091 param2pos
[i
] = n_arg
++;
3099 /* For each nested access parameter in the access relations in "expr",
3100 * construct a corresponding pet_expr, place it in expr->args and
3101 * record its position in "param2pos".
3102 * n is the number of nested access parameters.
3104 struct pet_expr
*PetScan::extract_nested(struct pet_expr
*expr
, int n
,
3105 std::map
<int,int> ¶m2pos
)
3109 expr
->args
= isl_calloc_array(ctx
, struct pet_expr
*, n
);
3114 space
= isl_map_get_space(expr
->acc
.access
);
3115 n
= extract_nested(space
, 0, expr
->args
, param2pos
);
3116 isl_space_free(space
);
3124 pet_expr_free(expr
);
3128 /* Look for parameters in any access relation in "expr" that
3129 * refer to nested accesses. In particular, these are
3130 * parameters with no name.
3132 * If there are any such parameters, then the domain of the access
3133 * relation, which is still [] at this point, is replaced by
3134 * [[] -> [t_1,...,t_n]], with n the number of these parameters
3135 * (after identifying identical nested accesses).
3136 * The parameters are then equated to the corresponding t dimensions
3137 * and subsequently projected out.
3138 * param2pos maps the position of the parameter to the position
3139 * of the corresponding t dimension.
3141 struct pet_expr
*PetScan::resolve_nested(struct pet_expr
*expr
)
3148 std::map
<int,int> param2pos
;
3153 for (int i
= 0; i
< expr
->n_arg
; ++i
) {
3154 expr
->args
[i
] = resolve_nested(expr
->args
[i
]);
3155 if (!expr
->args
[i
]) {
3156 pet_expr_free(expr
);
3161 if (expr
->type
!= pet_expr_access
)
3164 n
= n_nested_parameter(expr
->acc
.access
);
3168 expr
= extract_nested(expr
, n
, param2pos
);
3173 nparam
= isl_map_dim(expr
->acc
.access
, isl_dim_param
);
3174 n_in
= isl_map_dim(expr
->acc
.access
, isl_dim_in
);
3175 dim
= isl_map_get_space(expr
->acc
.access
);
3176 dim
= isl_space_domain(dim
);
3177 dim
= isl_space_from_domain(dim
);
3178 dim
= isl_space_add_dims(dim
, isl_dim_out
, n
);
3179 map
= isl_map_universe(dim
);
3180 map
= isl_map_domain_map(map
);
3181 map
= isl_map_reverse(map
);
3182 expr
->acc
.access
= isl_map_apply_domain(expr
->acc
.access
, map
);
3184 for (int i
= nparam
- 1; i
>= 0; --i
) {
3185 isl_id
*id
= isl_map_get_dim_id(expr
->acc
.access
,
3187 if (!is_nested_parameter(id
)) {
3192 expr
->acc
.access
= isl_map_equate(expr
->acc
.access
,
3193 isl_dim_param
, i
, isl_dim_in
,
3194 n_in
+ param2pos
[i
]);
3195 expr
->acc
.access
= isl_map_project_out(expr
->acc
.access
,
3196 isl_dim_param
, i
, 1);
3203 pet_expr_free(expr
);
3207 /* Return the file offset of the expansion location of "Loc".
3209 static unsigned getExpansionOffset(SourceManager
&SM
, SourceLocation Loc
)
3211 return SM
.getFileOffset(SM
.getExpansionLoc(Loc
));
3214 #ifdef HAVE_FINDLOCATIONAFTERTOKEN
3216 /* Return a SourceLocation for the location after the first semicolon
3217 * after "loc". If Lexer::findLocationAfterToken is available, we simply
3218 * call it and also skip trailing spaces and newline.
3220 static SourceLocation
location_after_semi(SourceLocation loc
, SourceManager
&SM
,
3221 const LangOptions
&LO
)
3223 return Lexer::findLocationAfterToken(loc
, tok::semi
, SM
, LO
, true);
3228 /* Return a SourceLocation for the location after the first semicolon
3229 * after "loc". If Lexer::findLocationAfterToken is not available,
3230 * we look in the underlying character data for the first semicolon.
3232 static SourceLocation
location_after_semi(SourceLocation loc
, SourceManager
&SM
,
3233 const LangOptions
&LO
)
3236 const char *s
= SM
.getCharacterData(loc
);
3238 semi
= strchr(s
, ';');
3240 return SourceLocation();
3241 return loc
.getFileLocWithOffset(semi
+ 1 - s
);
3246 /* If the token at "loc" is the first token on the line, then return
3247 * a location referring to the start of the line.
3248 * Otherwise, return "loc".
3250 * This function is used to extend a scop to the start of the line
3251 * if the first token of the scop is also the first token on the line.
3253 * We look for the first token on the line. If its location is equal to "loc",
3254 * then the latter is the location of the first token on the line.
3256 static SourceLocation
move_to_start_of_line_if_first_token(SourceLocation loc
,
3257 SourceManager
&SM
, const LangOptions
&LO
)
3259 std::pair
<FileID
, unsigned> file_offset_pair
;
3260 llvm::StringRef file
;
3263 SourceLocation token_loc
, line_loc
;
3266 loc
= SM
.getExpansionLoc(loc
);
3267 col
= SM
.getExpansionColumnNumber(loc
);
3268 line_loc
= loc
.getLocWithOffset(1 - col
);
3269 file_offset_pair
= SM
.getDecomposedLoc(line_loc
);
3270 file
= SM
.getBufferData(file_offset_pair
.first
, NULL
);
3271 pos
= file
.data() + file_offset_pair
.second
;
3273 Lexer
lexer(SM
.getLocForStartOfFile(file_offset_pair
.first
), LO
,
3274 file
.begin(), pos
, file
.end());
3275 lexer
.LexFromRawLexer(tok
);
3276 token_loc
= tok
.getLocation();
3278 if (token_loc
== loc
)
3284 /* Convert a top-level pet_expr to a pet_scop with one statement.
3285 * This mainly involves resolving nested expression parameters
3286 * and setting the name of the iteration space.
3287 * The name is given by "label" if it is non-NULL. Otherwise,
3288 * it is of the form S_<n_stmt>.
3289 * start and end of the pet_scop are derived from those of "stmt".
3291 struct pet_scop
*PetScan::extract(Stmt
*stmt
, struct pet_expr
*expr
,
3292 __isl_take isl_id
*label
)
3294 struct pet_stmt
*ps
;
3295 struct pet_scop
*scop
;
3296 SourceLocation loc
= stmt
->getLocStart();
3297 SourceManager
&SM
= PP
.getSourceManager();
3298 const LangOptions
&LO
= PP
.getLangOpts();
3299 int line
= PP
.getSourceManager().getExpansionLineNumber(loc
);
3300 unsigned start
, end
;
3302 expr
= resolve_nested(expr
);
3303 ps
= pet_stmt_from_pet_expr(ctx
, line
, label
, n_stmt
++, expr
);
3304 scop
= pet_scop_from_pet_stmt(ctx
, ps
);
3306 loc
= move_to_start_of_line_if_first_token(loc
, SM
, LO
);
3307 start
= getExpansionOffset(SM
, loc
);
3308 loc
= stmt
->getLocEnd();
3309 loc
= location_after_semi(loc
, SM
, LO
);
3310 end
= getExpansionOffset(SM
, loc
);
3312 scop
= pet_scop_update_start_end(scop
, start
, end
);
3316 /* Check if we can extract an affine expression from "expr".
3317 * Return the expressions as an isl_pw_aff if we can and NULL otherwise.
3318 * We turn on autodetection so that we won't generate any warnings
3319 * and turn off nesting, so that we won't accept any non-affine constructs.
3321 __isl_give isl_pw_aff
*PetScan::try_extract_affine(Expr
*expr
)
3324 int save_autodetect
= options
->autodetect
;
3325 bool save_nesting
= nesting_enabled
;
3327 options
->autodetect
= 1;
3328 nesting_enabled
= false;
3330 pwaff
= extract_affine(expr
);
3332 options
->autodetect
= save_autodetect
;
3333 nesting_enabled
= save_nesting
;
3338 /* Check whether "expr" is an affine expression.
3340 bool PetScan::is_affine(Expr
*expr
)
3344 pwaff
= try_extract_affine(expr
);
3345 isl_pw_aff_free(pwaff
);
3347 return pwaff
!= NULL
;
3350 /* Check if we can extract an affine constraint from "expr".
3351 * Return the constraint as an isl_set if we can and NULL otherwise.
3352 * We turn on autodetection so that we won't generate any warnings
3353 * and turn off nesting, so that we won't accept any non-affine constructs.
3355 __isl_give isl_pw_aff
*PetScan::try_extract_affine_condition(Expr
*expr
)
3358 int save_autodetect
= options
->autodetect
;
3359 bool save_nesting
= nesting_enabled
;
3361 options
->autodetect
= 1;
3362 nesting_enabled
= false;
3364 cond
= extract_condition(expr
);
3366 options
->autodetect
= save_autodetect
;
3367 nesting_enabled
= save_nesting
;
3372 /* Check whether "expr" is an affine constraint.
3374 bool PetScan::is_affine_condition(Expr
*expr
)
3378 cond
= try_extract_affine_condition(expr
);
3379 isl_pw_aff_free(cond
);
3381 return cond
!= NULL
;
3384 /* Check if we can extract a condition from "expr".
3385 * Return the condition as an isl_pw_aff if we can and NULL otherwise.
3386 * If allow_nested is set, then the condition may involve parameters
3387 * corresponding to nested accesses.
3388 * We turn on autodetection so that we won't generate any warnings.
3390 __isl_give isl_pw_aff
*PetScan::try_extract_nested_condition(Expr
*expr
)
3393 int save_autodetect
= options
->autodetect
;
3394 bool save_nesting
= nesting_enabled
;
3396 options
->autodetect
= 1;
3397 nesting_enabled
= allow_nested
;
3398 cond
= extract_condition(expr
);
3400 options
->autodetect
= save_autodetect
;
3401 nesting_enabled
= save_nesting
;
3406 /* If the top-level expression of "stmt" is an assignment, then
3407 * return that assignment as a BinaryOperator.
3408 * Otherwise return NULL.
3410 static BinaryOperator
*top_assignment_or_null(Stmt
*stmt
)
3412 BinaryOperator
*ass
;
3416 if (stmt
->getStmtClass() != Stmt::BinaryOperatorClass
)
3419 ass
= cast
<BinaryOperator
>(stmt
);
3420 if(ass
->getOpcode() != BO_Assign
)
3426 /* Check if the given if statement is a conditional assignement
3427 * with a non-affine condition. If so, construct a pet_scop
3428 * corresponding to this conditional assignment. Otherwise return NULL.
3430 * In particular we check if "stmt" is of the form
3437 * where a is some array or scalar access.
3438 * The constructed pet_scop then corresponds to the expression
3440 * a = condition ? f(...) : g(...)
3442 * All access relations in f(...) are intersected with condition
3443 * while all access relation in g(...) are intersected with the complement.
3445 struct pet_scop
*PetScan::extract_conditional_assignment(IfStmt
*stmt
)
3447 BinaryOperator
*ass_then
, *ass_else
;
3448 isl_map
*write_then
, *write_else
;
3449 isl_set
*cond
, *comp
;
3453 struct pet_expr
*pe_cond
, *pe_then
, *pe_else
, *pe
, *pe_write
;
3454 bool save_nesting
= nesting_enabled
;
3456 if (!options
->detect_conditional_assignment
)
3459 ass_then
= top_assignment_or_null(stmt
->getThen());
3460 ass_else
= top_assignment_or_null(stmt
->getElse());
3462 if (!ass_then
|| !ass_else
)
3465 if (is_affine_condition(stmt
->getCond()))
3468 write_then
= extract_access(ass_then
->getLHS());
3469 write_else
= extract_access(ass_else
->getLHS());
3471 equal
= isl_map_is_equal(write_then
, write_else
);
3472 isl_map_free(write_else
);
3473 if (equal
< 0 || !equal
) {
3474 isl_map_free(write_then
);
3478 nesting_enabled
= allow_nested
;
3479 pa
= extract_condition(stmt
->getCond());
3480 nesting_enabled
= save_nesting
;
3481 cond
= isl_pw_aff_non_zero_set(isl_pw_aff_copy(pa
));
3482 comp
= isl_pw_aff_zero_set(isl_pw_aff_copy(pa
));
3483 map
= isl_map_from_range(isl_set_from_pw_aff(pa
));
3485 pe_cond
= pet_expr_from_access(map
);
3487 pe_then
= extract_expr(ass_then
->getRHS());
3488 pe_then
= pet_expr_restrict(pe_then
, cond
);
3489 pe_else
= extract_expr(ass_else
->getRHS());
3490 pe_else
= pet_expr_restrict(pe_else
, comp
);
3492 pe
= pet_expr_new_ternary(ctx
, pe_cond
, pe_then
, pe_else
);
3493 pe_write
= pet_expr_from_access(write_then
);
3495 pe_write
->acc
.write
= 1;
3496 pe_write
->acc
.read
= 0;
3498 pe
= pet_expr_new_binary(ctx
, pet_op_assign
, pe_write
, pe
);
3499 return extract(stmt
, pe
);
3502 /* Create a pet_scop with a single statement evaluating "cond"
3503 * and writing the result to a virtual scalar, as expressed by
3506 struct pet_scop
*PetScan::extract_non_affine_condition(Expr
*cond
,
3507 __isl_take isl_map
*access
)
3509 struct pet_expr
*expr
, *write
;
3510 struct pet_stmt
*ps
;
3511 struct pet_scop
*scop
;
3512 SourceLocation loc
= cond
->getLocStart();
3513 int line
= PP
.getSourceManager().getExpansionLineNumber(loc
);
3515 write
= pet_expr_from_access(access
);
3517 write
->acc
.write
= 1;
3518 write
->acc
.read
= 0;
3520 expr
= extract_expr(cond
);
3521 expr
= resolve_nested(expr
);
3522 expr
= pet_expr_new_binary(ctx
, pet_op_assign
, write
, expr
);
3523 ps
= pet_stmt_from_pet_expr(ctx
, line
, NULL
, n_stmt
++, expr
);
3524 scop
= pet_scop_from_pet_stmt(ctx
, ps
);
3525 scop
= resolve_nested(scop
);
3531 static struct pet_expr
*embed_access(struct pet_expr
*expr
, void *user
);
3534 /* Apply the map pointed to by "user" to the domain of the access
3535 * relation associated to "expr", thereby embedding it in the range of the map.
3536 * The domain of both relations is the zero-dimensional domain.
3538 static struct pet_expr
*embed_access(struct pet_expr
*expr
, void *user
)
3540 isl_map
*map
= (isl_map
*) user
;
3542 expr
->acc
.access
= isl_map_apply_domain(expr
->acc
.access
,
3544 if (!expr
->acc
.access
)
3549 pet_expr_free(expr
);
3553 /* Apply "map" to all access relations in "expr".
3555 static struct pet_expr
*embed(struct pet_expr
*expr
, __isl_keep isl_map
*map
)
3557 return pet_expr_map_access(expr
, &embed_access
, map
);
3560 /* How many parameters of "set" refer to nested accesses, i.e., have no name?
3562 static int n_nested_parameter(__isl_keep isl_set
*set
)
3567 space
= isl_set_get_space(set
);
3568 n
= n_nested_parameter(space
);
3569 isl_space_free(space
);
3574 /* Remove all parameters from "map" that refer to nested accesses.
3576 static __isl_give isl_map
*remove_nested_parameters(__isl_take isl_map
*map
)
3581 space
= isl_map_get_space(map
);
3582 nparam
= isl_space_dim(space
, isl_dim_param
);
3583 for (int i
= nparam
- 1; i
>= 0; --i
)
3584 if (is_nested_parameter(space
, i
))
3585 map
= isl_map_project_out(map
, isl_dim_param
, i
, 1);
3586 isl_space_free(space
);
3591 /* Remove all parameters from the access relation of "expr"
3592 * that refer to nested accesses.
3594 static struct pet_expr
*remove_nested_parameters(struct pet_expr
*expr
)
3596 expr
->acc
.access
= remove_nested_parameters(expr
->acc
.access
);
3597 if (!expr
->acc
.access
)
3602 pet_expr_free(expr
);
3607 static struct pet_expr
*expr_remove_nested_parameters(
3608 struct pet_expr
*expr
, void *user
);
3611 static struct pet_expr
*expr_remove_nested_parameters(
3612 struct pet_expr
*expr
, void *user
)
3614 return remove_nested_parameters(expr
);
3617 /* Remove all nested access parameters from the schedule and all
3618 * accesses of "stmt".
3619 * There is no need to remove them from the domain as these parameters
3620 * have already been removed from the domain when this function is called.
3622 static struct pet_stmt
*remove_nested_parameters(struct pet_stmt
*stmt
)
3626 stmt
->schedule
= remove_nested_parameters(stmt
->schedule
);
3627 stmt
->body
= pet_expr_map_access(stmt
->body
,
3628 &expr_remove_nested_parameters
, NULL
);
3629 if (!stmt
->schedule
|| !stmt
->body
)
3631 for (int i
= 0; i
< stmt
->n_arg
; ++i
) {
3632 stmt
->args
[i
] = pet_expr_map_access(stmt
->args
[i
],
3633 &expr_remove_nested_parameters
, NULL
);
3640 pet_stmt_free(stmt
);
3644 /* For each nested access parameter in the domain of "stmt",
3645 * construct a corresponding pet_expr, place it before the original
3646 * elements in stmt->args and record its position in "param2pos".
3647 * n is the number of nested access parameters.
3649 struct pet_stmt
*PetScan::extract_nested(struct pet_stmt
*stmt
, int n
,
3650 std::map
<int,int> ¶m2pos
)
3655 struct pet_expr
**args
;
3657 n_arg
= stmt
->n_arg
;
3658 args
= isl_calloc_array(ctx
, struct pet_expr
*, n
+ n_arg
);
3662 space
= isl_set_get_space(stmt
->domain
);
3663 n_arg
= extract_nested(space
, 0, args
, param2pos
);
3664 isl_space_free(space
);
3669 for (i
= 0; i
< stmt
->n_arg
; ++i
)
3670 args
[n_arg
+ i
] = stmt
->args
[i
];
3673 stmt
->n_arg
+= n_arg
;
3678 for (i
= 0; i
< n
; ++i
)
3679 pet_expr_free(args
[i
]);
3682 pet_stmt_free(stmt
);
3686 /* Check whether any of the arguments i of "stmt" starting at position "n"
3687 * is equal to one of the first "n" arguments j.
3688 * If so, combine the constraints on arguments i and j and remove
3691 static struct pet_stmt
*remove_duplicate_arguments(struct pet_stmt
*stmt
, int n
)
3700 if (n
== stmt
->n_arg
)
3703 map
= isl_set_unwrap(stmt
->domain
);
3705 for (i
= stmt
->n_arg
- 1; i
>= n
; --i
) {
3706 for (j
= 0; j
< n
; ++j
)
3707 if (pet_expr_is_equal(stmt
->args
[i
], stmt
->args
[j
]))
3712 map
= isl_map_equate(map
, isl_dim_out
, i
, isl_dim_out
, j
);
3713 map
= isl_map_project_out(map
, isl_dim_out
, i
, 1);
3715 pet_expr_free(stmt
->args
[i
]);
3716 for (j
= i
; j
+ 1 < stmt
->n_arg
; ++j
)
3717 stmt
->args
[j
] = stmt
->args
[j
+ 1];
3721 stmt
->domain
= isl_map_wrap(map
);
3726 pet_stmt_free(stmt
);
3730 /* Look for parameters in the iteration domain of "stmt" that
3731 * refer to nested accesses. In particular, these are
3732 * parameters with no name.
3734 * If there are any such parameters, then as many extra variables
3735 * (after identifying identical nested accesses) are inserted in the
3736 * range of the map wrapped inside the domain, before the original variables.
3737 * If the original domain is not a wrapped map, then a new wrapped
3738 * map is created with zero output dimensions.
3739 * The parameters are then equated to the corresponding output dimensions
3740 * and subsequently projected out, from the iteration domain,
3741 * the schedule and the access relations.
3742 * For each of the output dimensions, a corresponding argument
3743 * expression is inserted. Initially they are created with
3744 * a zero-dimensional domain, so they have to be embedded
3745 * in the current iteration domain.
3746 * param2pos maps the position of the parameter to the position
3747 * of the corresponding output dimension in the wrapped map.
3749 struct pet_stmt
*PetScan::resolve_nested(struct pet_stmt
*stmt
)
3755 std::map
<int,int> param2pos
;
3760 n
= n_nested_parameter(stmt
->domain
);
3764 n_arg
= stmt
->n_arg
;
3765 stmt
= extract_nested(stmt
, n
, param2pos
);
3769 n
= stmt
->n_arg
- n_arg
;
3770 nparam
= isl_set_dim(stmt
->domain
, isl_dim_param
);
3771 if (isl_set_is_wrapping(stmt
->domain
))
3772 map
= isl_set_unwrap(stmt
->domain
);
3774 map
= isl_map_from_domain(stmt
->domain
);
3775 map
= isl_map_insert_dims(map
, isl_dim_out
, 0, n
);
3777 for (int i
= nparam
- 1; i
>= 0; --i
) {
3780 if (!is_nested_parameter(map
, i
))
3783 id
= pet_expr_access_get_id(stmt
->args
[param2pos
[i
]]);
3784 map
= isl_map_set_dim_id(map
, isl_dim_out
, param2pos
[i
], id
);
3785 map
= isl_map_equate(map
, isl_dim_param
, i
, isl_dim_out
,
3787 map
= isl_map_project_out(map
, isl_dim_param
, i
, 1);
3790 stmt
->domain
= isl_map_wrap(map
);
3792 map
= isl_set_unwrap(isl_set_copy(stmt
->domain
));
3793 map
= isl_map_from_range(isl_map_domain(map
));
3794 for (int pos
= 0; pos
< n
; ++pos
)
3795 stmt
->args
[pos
] = embed(stmt
->args
[pos
], map
);
3798 stmt
= remove_nested_parameters(stmt
);
3799 stmt
= remove_duplicate_arguments(stmt
, n
);
3803 pet_stmt_free(stmt
);
3807 /* For each statement in "scop", move the parameters that correspond
3808 * to nested access into the ranges of the domains and create
3809 * corresponding argument expressions.
3811 struct pet_scop
*PetScan::resolve_nested(struct pet_scop
*scop
)
3816 for (int i
= 0; i
< scop
->n_stmt
; ++i
) {
3817 scop
->stmts
[i
] = resolve_nested(scop
->stmts
[i
]);
3818 if (!scop
->stmts
[i
])
3824 pet_scop_free(scop
);
3828 /* Given an access expression "expr", is the variable accessed by
3829 * "expr" assigned anywhere inside "scop"?
3831 static bool is_assigned(pet_expr
*expr
, pet_scop
*scop
)
3833 bool assigned
= false;
3836 id
= pet_expr_access_get_id(expr
);
3837 assigned
= pet_scop_writes(scop
, id
);
3843 /* Are all nested access parameters in "pa" allowed given "scop".
3844 * In particular, is none of them written by anywhere inside "scop".
3846 * If "scop" has any skip conditions, then no nested access parameters
3847 * are allowed. In particular, if there is any nested access in a guard
3848 * for a piece of code containing a "continue", then we want to introduce
3849 * a separate statement for evaluating this guard so that we can express
3850 * that the result is false for all previous iterations.
3852 bool PetScan::is_nested_allowed(__isl_keep isl_pw_aff
*pa
, pet_scop
*scop
)
3859 nparam
= isl_pw_aff_dim(pa
, isl_dim_param
);
3860 for (int i
= 0; i
< nparam
; ++i
) {
3862 isl_id
*id
= isl_pw_aff_get_dim_id(pa
, isl_dim_param
, i
);
3866 if (!is_nested_parameter(id
)) {
3871 if (pet_scop_has_skip(scop
, pet_skip_now
)) {
3876 nested
= (Expr
*) isl_id_get_user(id
);
3877 expr
= extract_expr(nested
);
3878 allowed
= expr
&& expr
->type
== pet_expr_access
&&
3879 !is_assigned(expr
, scop
);
3881 pet_expr_free(expr
);
3891 /* Do we need to construct a skip condition of the given type
3892 * on an if statement, given that the if condition is non-affine?
3894 * pet_scop_filter_skip can only handle the case where the if condition
3895 * holds (the then branch) and the skip condition is universal.
3896 * In any other case, we need to construct a new skip condition.
3898 static bool need_skip(struct pet_scop
*scop_then
, struct pet_scop
*scop_else
,
3899 bool have_else
, enum pet_skip type
)
3901 if (have_else
&& scop_else
&& pet_scop_has_skip(scop_else
, type
))
3903 if (scop_then
&& pet_scop_has_skip(scop_then
, type
) &&
3904 !pet_scop_has_universal_skip(scop_then
, type
))
3909 /* Do we need to construct a skip condition of the given type
3910 * on an if statement, given that the if condition is affine?
3912 * There is no need to construct a new skip condition if all
3913 * the skip conditions are affine.
3915 static bool need_skip_aff(struct pet_scop
*scop_then
,
3916 struct pet_scop
*scop_else
, bool have_else
, enum pet_skip type
)
3918 if (scop_then
&& pet_scop_has_var_skip(scop_then
, type
))
3920 if (have_else
&& scop_else
&& pet_scop_has_var_skip(scop_else
, type
))
3925 /* Do we need to construct a skip condition of the given type
3926 * on an if statement?
3928 static bool need_skip(struct pet_scop
*scop_then
, struct pet_scop
*scop_else
,
3929 bool have_else
, enum pet_skip type
, bool affine
)
3932 return need_skip_aff(scop_then
, scop_else
, have_else
, type
);
3934 return need_skip(scop_then
, scop_else
, have_else
, type
);
3937 /* Construct an affine expression pet_expr that evaluates
3938 * to the constant "val".
3940 static struct pet_expr
*universally(isl_ctx
*ctx
, int val
)
3945 space
= isl_space_alloc(ctx
, 0, 0, 1);
3946 map
= isl_map_universe(space
);
3947 map
= isl_map_fix_si(map
, isl_dim_out
, 0, val
);
3949 return pet_expr_from_access(map
);
3952 /* Construct an affine expression pet_expr that evaluates
3953 * to the constant 1.
3955 static struct pet_expr
*universally_true(isl_ctx
*ctx
)
3957 return universally(ctx
, 1);
3960 /* Construct an affine expression pet_expr that evaluates
3961 * to the constant 0.
3963 static struct pet_expr
*universally_false(isl_ctx
*ctx
)
3965 return universally(ctx
, 0);
3968 /* Given an access relation "test_access" for the if condition,
3969 * an access relation "skip_access" for the skip condition and
3970 * scops for the then and else branches, construct a scop for
3971 * computing "skip_access".
3973 * The computed scop contains a single statement that essentially does
3975 * skip_cond = test_cond ? skip_cond_then : skip_cond_else
3977 * If the skip conditions of the then and/or else branch are not affine,
3978 * then they need to be filtered by test_access.
3979 * If they are missing, then this means the skip condition is false.
3981 * Since we are constructing a skip condition for the if statement,
3982 * the skip conditions on the then and else branches are removed.
3984 static struct pet_scop
*extract_skip(PetScan
*scan
,
3985 __isl_take isl_map
*test_access
, __isl_take isl_map
*skip_access
,
3986 struct pet_scop
*scop_then
, struct pet_scop
*scop_else
, bool have_else
,
3989 struct pet_expr
*expr_then
, *expr_else
, *expr
, *expr_skip
;
3990 struct pet_stmt
*stmt
;
3991 struct pet_scop
*scop
;
3992 isl_ctx
*ctx
= scan
->ctx
;
3996 if (have_else
&& !scop_else
)
3999 if (pet_scop_has_skip(scop_then
, type
)) {
4000 expr_then
= pet_scop_get_skip_expr(scop_then
, type
);
4001 pet_scop_reset_skip(scop_then
, type
);
4002 if (!pet_expr_is_affine(expr_then
))
4003 expr_then
= pet_expr_filter(expr_then
,
4004 isl_map_copy(test_access
), 1);
4006 expr_then
= universally_false(ctx
);
4008 if (have_else
&& pet_scop_has_skip(scop_else
, type
)) {
4009 expr_else
= pet_scop_get_skip_expr(scop_else
, type
);
4010 pet_scop_reset_skip(scop_else
, type
);
4011 if (!pet_expr_is_affine(expr_else
))
4012 expr_else
= pet_expr_filter(expr_else
,
4013 isl_map_copy(test_access
), 0);
4015 expr_else
= universally_false(ctx
);
4017 expr
= pet_expr_from_access(test_access
);
4018 expr
= pet_expr_new_ternary(ctx
, expr
, expr_then
, expr_else
);
4019 expr_skip
= pet_expr_from_access(isl_map_copy(skip_access
));
4021 expr_skip
->acc
.write
= 1;
4022 expr_skip
->acc
.read
= 0;
4024 expr
= pet_expr_new_binary(ctx
, pet_op_assign
, expr_skip
, expr
);
4025 stmt
= pet_stmt_from_pet_expr(ctx
, -1, NULL
, scan
->n_stmt
++, expr
);
4027 scop
= pet_scop_from_pet_stmt(ctx
, stmt
);
4028 scop
= scop_add_array(scop
, skip_access
, scan
->ast_context
);
4029 isl_map_free(skip_access
);
4033 isl_map_free(test_access
);
4034 isl_map_free(skip_access
);
4038 /* Is scop's skip_now condition equal to its skip_later condition?
4039 * In particular, this means that it either has no skip_now condition
4040 * or both a skip_now and a skip_later condition (that are equal to each other).
4042 static bool skip_equals_skip_later(struct pet_scop
*scop
)
4044 int has_skip_now
, has_skip_later
;
4046 isl_set
*skip_now
, *skip_later
;
4050 has_skip_now
= pet_scop_has_skip(scop
, pet_skip_now
);
4051 has_skip_later
= pet_scop_has_skip(scop
, pet_skip_later
);
4052 if (has_skip_now
!= has_skip_later
)
4057 skip_now
= pet_scop_get_skip(scop
, pet_skip_now
);
4058 skip_later
= pet_scop_get_skip(scop
, pet_skip_later
);
4059 equal
= isl_set_is_equal(skip_now
, skip_later
);
4060 isl_set_free(skip_now
);
4061 isl_set_free(skip_later
);
4066 /* Drop the skip conditions of type pet_skip_later from scop1 and scop2.
4068 static void drop_skip_later(struct pet_scop
*scop1
, struct pet_scop
*scop2
)
4070 pet_scop_reset_skip(scop1
, pet_skip_later
);
4071 pet_scop_reset_skip(scop2
, pet_skip_later
);
4074 /* Structure that handles the construction of skip conditions.
4076 * scop_then and scop_else represent the then and else branches
4077 * of the if statement
4079 * skip[type] is true if we need to construct a skip condition of that type
4080 * equal is set if the skip conditions of types pet_skip_now and pet_skip_later
4081 * are equal to each other
4082 * access[type] is the virtual array representing the skip condition
4083 * scop[type] is a scop for computing the skip condition
4085 struct pet_skip_info
{
4091 struct pet_scop
*scop
[2];
4093 pet_skip_info(isl_ctx
*ctx
) : ctx(ctx
) {}
4095 operator bool() { return skip
[pet_skip_now
] || skip
[pet_skip_later
]; }
4098 /* Structure that handles the construction of skip conditions on if statements.
4100 * scop_then and scop_else represent the then and else branches
4101 * of the if statement
4103 struct pet_skip_info_if
: public pet_skip_info
{
4104 struct pet_scop
*scop_then
, *scop_else
;
4107 pet_skip_info_if(isl_ctx
*ctx
, struct pet_scop
*scop_then
,
4108 struct pet_scop
*scop_else
, bool have_else
, bool affine
);
4109 void extract(PetScan
*scan
, __isl_keep isl_map
*access
,
4110 enum pet_skip type
);
4111 void extract(PetScan
*scan
, __isl_keep isl_map
*access
);
4112 void extract(PetScan
*scan
, __isl_keep isl_pw_aff
*cond
);
4113 struct pet_scop
*add(struct pet_scop
*scop
, enum pet_skip type
,
4115 struct pet_scop
*add(struct pet_scop
*scop
, int offset
);
4118 /* Initialize a pet_skip_info_if structure based on the then and else branches
4119 * and based on whether the if condition is affine or not.
4121 pet_skip_info_if::pet_skip_info_if(isl_ctx
*ctx
, struct pet_scop
*scop_then
,
4122 struct pet_scop
*scop_else
, bool have_else
, bool affine
) :
4123 pet_skip_info(ctx
), scop_then(scop_then
), scop_else(scop_else
),
4124 have_else(have_else
)
4126 skip
[pet_skip_now
] =
4127 need_skip(scop_then
, scop_else
, have_else
, pet_skip_now
, affine
);
4128 equal
= skip
[pet_skip_now
] && skip_equals_skip_later(scop_then
) &&
4129 (!have_else
|| skip_equals_skip_later(scop_else
));
4130 skip
[pet_skip_later
] = skip
[pet_skip_now
] && !equal
&&
4131 need_skip(scop_then
, scop_else
, have_else
, pet_skip_later
, affine
);
4134 /* If we need to construct a skip condition of the given type,
4137 * "map" represents the if condition.
4139 void pet_skip_info_if::extract(PetScan
*scan
, __isl_keep isl_map
*map
,
4145 access
[type
] = create_test_access(isl_map_get_ctx(map
), scan
->n_test
++);
4146 scop
[type
] = extract_skip(scan
, isl_map_copy(map
),
4147 isl_map_copy(access
[type
]),
4148 scop_then
, scop_else
, have_else
, type
);
4151 /* Construct the required skip conditions, given the if condition "map".
4153 void pet_skip_info_if::extract(PetScan
*scan
, __isl_keep isl_map
*map
)
4155 extract(scan
, map
, pet_skip_now
);
4156 extract(scan
, map
, pet_skip_later
);
4158 drop_skip_later(scop_then
, scop_else
);
4161 /* Construct the required skip conditions, given the if condition "cond".
4163 void pet_skip_info_if::extract(PetScan
*scan
, __isl_keep isl_pw_aff
*cond
)
4168 if (!skip
[pet_skip_now
] && !skip
[pet_skip_later
])
4171 test_set
= isl_set_from_pw_aff(isl_pw_aff_copy(cond
));
4172 test
= isl_map_from_range(test_set
);
4173 extract(scan
, test
);
4177 /* Add the computed skip condition of the give type to "main" and
4178 * add the scop for computing the condition at the given offset.
4180 * If equal is set, then we only computed a skip condition for pet_skip_now,
4181 * but we also need to set it as main's pet_skip_later.
4183 struct pet_scop
*pet_skip_info_if::add(struct pet_scop
*main
,
4184 enum pet_skip type
, int offset
)
4191 skip_set
= isl_map_range(access
[type
]);
4192 access
[type
] = NULL
;
4193 scop
[type
] = pet_scop_prefix(scop
[type
], offset
);
4194 main
= pet_scop_add_par(ctx
, main
, scop
[type
]);
4198 main
= pet_scop_set_skip(main
, pet_skip_later
,
4199 isl_set_copy(skip_set
));
4201 main
= pet_scop_set_skip(main
, type
, skip_set
);
4206 /* Add the computed skip conditions to "main" and
4207 * add the scops for computing the conditions at the given offset.
4209 struct pet_scop
*pet_skip_info_if::add(struct pet_scop
*scop
, int offset
)
4211 scop
= add(scop
, pet_skip_now
, offset
);
4212 scop
= add(scop
, pet_skip_later
, offset
);
4217 /* Construct a pet_scop for a non-affine if statement.
4219 * We create a separate statement that writes the result
4220 * of the non-affine condition to a virtual scalar.
4221 * A constraint requiring the value of this virtual scalar to be one
4222 * is added to the iteration domains of the then branch.
4223 * Similarly, a constraint requiring the value of this virtual scalar
4224 * to be zero is added to the iteration domains of the else branch, if any.
4225 * We adjust the schedules to ensure that the virtual scalar is written
4226 * before it is read.
4228 * If there are any breaks or continues in the then and/or else
4229 * branches, then we may have to compute a new skip condition.
4230 * This is handled using a pet_skip_info_if object.
4231 * On initialization, the object checks if skip conditions need
4232 * to be computed. If so, it does so in "extract" and adds them in "add".
4234 struct pet_scop
*PetScan::extract_non_affine_if(Expr
*cond
,
4235 struct pet_scop
*scop_then
, struct pet_scop
*scop_else
,
4236 bool have_else
, int stmt_id
)
4238 struct pet_scop
*scop
;
4239 isl_map
*test_access
;
4240 int save_n_stmt
= n_stmt
;
4242 test_access
= create_test_access(ctx
, n_test
++);
4244 scop
= extract_non_affine_condition(cond
, isl_map_copy(test_access
));
4245 n_stmt
= save_n_stmt
;
4246 scop
= scop_add_array(scop
, test_access
, ast_context
);
4248 pet_skip_info_if
skip(ctx
, scop_then
, scop_else
, have_else
, false);
4249 skip
.extract(this, test_access
);
4251 scop
= pet_scop_prefix(scop
, 0);
4252 scop_then
= pet_scop_prefix(scop_then
, 1);
4253 scop_then
= pet_scop_filter(scop_then
, isl_map_copy(test_access
), 1);
4255 scop_else
= pet_scop_prefix(scop_else
, 1);
4256 scop_else
= pet_scop_filter(scop_else
, test_access
, 0);
4257 scop_then
= pet_scop_add_par(ctx
, scop_then
, scop_else
);
4259 isl_map_free(test_access
);
4261 scop
= pet_scop_add_seq(ctx
, scop
, scop_then
);
4263 scop
= skip
.add(scop
, 2);
4268 /* Construct a pet_scop for an if statement.
4270 * If the condition fits the pattern of a conditional assignment,
4271 * then it is handled by extract_conditional_assignment.
4272 * Otherwise, we do the following.
4274 * If the condition is affine, then the condition is added
4275 * to the iteration domains of the then branch, while the
4276 * opposite of the condition in added to the iteration domains
4277 * of the else branch, if any.
4278 * We allow the condition to be dynamic, i.e., to refer to
4279 * scalars or array elements that may be written to outside
4280 * of the given if statement. These nested accesses are then represented
4281 * as output dimensions in the wrapping iteration domain.
4282 * If it also written _inside_ the then or else branch, then
4283 * we treat the condition as non-affine.
4284 * As explained in extract_non_affine_if, this will introduce
4285 * an extra statement.
4286 * For aesthetic reasons, we want this statement to have a statement
4287 * number that is lower than those of the then and else branches.
4288 * In order to evaluate if will need such a statement, however, we
4289 * first construct scops for the then and else branches.
4290 * We therefore reserve a statement number if we might have to
4291 * introduce such an extra statement.
4293 * If the condition is not affine, then the scop is created in
4294 * extract_non_affine_if.
4296 * If there are any breaks or continues in the then and/or else
4297 * branches, then we may have to compute a new skip condition.
4298 * This is handled using a pet_skip_info_if object.
4299 * On initialization, the object checks if skip conditions need
4300 * to be computed. If so, it does so in "extract" and adds them in "add".
4302 struct pet_scop
*PetScan::extract(IfStmt
*stmt
)
4304 struct pet_scop
*scop_then
, *scop_else
= NULL
, *scop
;
4310 scop
= extract_conditional_assignment(stmt
);
4314 cond
= try_extract_nested_condition(stmt
->getCond());
4315 if (allow_nested
&& (!cond
|| has_nested(cond
)))
4319 assigned_value_cache
cache(assigned_value
);
4320 scop_then
= extract(stmt
->getThen());
4323 if (stmt
->getElse()) {
4324 assigned_value_cache
cache(assigned_value
);
4325 scop_else
= extract(stmt
->getElse());
4326 if (options
->autodetect
) {
4327 if (scop_then
&& !scop_else
) {
4329 isl_pw_aff_free(cond
);
4332 if (!scop_then
&& scop_else
) {
4334 isl_pw_aff_free(cond
);
4341 (!is_nested_allowed(cond
, scop_then
) ||
4342 (stmt
->getElse() && !is_nested_allowed(cond
, scop_else
)))) {
4343 isl_pw_aff_free(cond
);
4346 if (allow_nested
&& !cond
)
4347 return extract_non_affine_if(stmt
->getCond(), scop_then
,
4348 scop_else
, stmt
->getElse(), stmt_id
);
4351 cond
= extract_condition(stmt
->getCond());
4353 pet_skip_info_if
skip(ctx
, scop_then
, scop_else
, stmt
->getElse(), true);
4354 skip
.extract(this, cond
);
4356 valid
= isl_pw_aff_domain(isl_pw_aff_copy(cond
));
4357 set
= isl_pw_aff_non_zero_set(cond
);
4358 scop
= pet_scop_restrict(scop_then
, isl_set_copy(set
));
4360 if (stmt
->getElse()) {
4361 set
= isl_set_subtract(isl_set_copy(valid
), set
);
4362 scop_else
= pet_scop_restrict(scop_else
, set
);
4363 scop
= pet_scop_add_par(ctx
, scop
, scop_else
);
4366 scop
= resolve_nested(scop
);
4367 scop
= pet_scop_restrict_context(scop
, valid
);
4370 scop
= pet_scop_prefix(scop
, 0);
4371 scop
= skip
.add(scop
, 1);
4376 /* Try and construct a pet_scop for a label statement.
4377 * We currently only allow labels on expression statements.
4379 struct pet_scop
*PetScan::extract(LabelStmt
*stmt
)
4384 sub
= stmt
->getSubStmt();
4385 if (!isa
<Expr
>(sub
)) {
4390 label
= isl_id_alloc(ctx
, stmt
->getName(), NULL
);
4392 return extract(sub
, extract_expr(cast
<Expr
>(sub
)), label
);
4395 /* Construct a pet_scop for a continue statement.
4397 * We simply create an empty scop with a universal pet_skip_now
4398 * skip condition. This skip condition will then be taken into
4399 * account by the enclosing loop construct, possibly after
4400 * being incorporated into outer skip conditions.
4402 struct pet_scop
*PetScan::extract(ContinueStmt
*stmt
)
4408 scop
= pet_scop_empty(ctx
);
4412 space
= isl_space_set_alloc(ctx
, 0, 1);
4413 set
= isl_set_universe(space
);
4414 set
= isl_set_fix_si(set
, isl_dim_set
, 0, 1);
4415 scop
= pet_scop_set_skip(scop
, pet_skip_now
, set
);
4420 /* Construct a pet_scop for a break statement.
4422 * We simply create an empty scop with both a universal pet_skip_now
4423 * skip condition and a universal pet_skip_later skip condition.
4424 * These skip conditions will then be taken into
4425 * account by the enclosing loop construct, possibly after
4426 * being incorporated into outer skip conditions.
4428 struct pet_scop
*PetScan::extract(BreakStmt
*stmt
)
4434 scop
= pet_scop_empty(ctx
);
4438 space
= isl_space_set_alloc(ctx
, 0, 1);
4439 set
= isl_set_universe(space
);
4440 set
= isl_set_fix_si(set
, isl_dim_set
, 0, 1);
4441 scop
= pet_scop_set_skip(scop
, pet_skip_now
, isl_set_copy(set
));
4442 scop
= pet_scop_set_skip(scop
, pet_skip_later
, set
);
4447 /* Try and construct a pet_scop corresponding to "stmt".
4449 * If "stmt" is a compound statement, then "skip_declarations"
4450 * indicates whether we should skip initial declarations in the
4451 * compound statement.
4453 * If the constructed pet_scop is not a (possibly) partial representation
4454 * of "stmt", we update start and end of the pet_scop to those of "stmt".
4455 * In particular, if skip_declarations, then we may have skipped declarations
4456 * inside "stmt" and so the pet_scop may not represent the entire "stmt".
4457 * Note that this function may be called with "stmt" referring to the entire
4458 * body of the function, including the outer braces. In such cases,
4459 * skip_declarations will be set and the braces will not be taken into
4460 * account in scop->start and scop->end.
4462 struct pet_scop
*PetScan::extract(Stmt
*stmt
, bool skip_declarations
)
4464 struct pet_scop
*scop
;
4465 unsigned start
, end
;
4467 SourceManager
&SM
= PP
.getSourceManager();
4468 const LangOptions
&LO
= PP
.getLangOpts();
4470 if (isa
<Expr
>(stmt
))
4471 return extract(stmt
, extract_expr(cast
<Expr
>(stmt
)));
4473 switch (stmt
->getStmtClass()) {
4474 case Stmt::WhileStmtClass
:
4475 scop
= extract(cast
<WhileStmt
>(stmt
));
4477 case Stmt::ForStmtClass
:
4478 scop
= extract_for(cast
<ForStmt
>(stmt
));
4480 case Stmt::IfStmtClass
:
4481 scop
= extract(cast
<IfStmt
>(stmt
));
4483 case Stmt::CompoundStmtClass
:
4484 scop
= extract(cast
<CompoundStmt
>(stmt
), skip_declarations
);
4486 case Stmt::LabelStmtClass
:
4487 scop
= extract(cast
<LabelStmt
>(stmt
));
4489 case Stmt::ContinueStmtClass
:
4490 scop
= extract(cast
<ContinueStmt
>(stmt
));
4492 case Stmt::BreakStmtClass
:
4493 scop
= extract(cast
<BreakStmt
>(stmt
));
4495 case Stmt::DeclStmtClass
:
4496 scop
= extract(cast
<DeclStmt
>(stmt
));
4503 if (partial
|| skip_declarations
)
4506 loc
= stmt
->getLocStart();
4507 loc
= move_to_start_of_line_if_first_token(loc
, SM
, LO
);
4508 start
= getExpansionOffset(SM
, loc
);
4509 loc
= PP
.getLocForEndOfToken(stmt
->getLocEnd());
4510 end
= getExpansionOffset(SM
, loc
);
4511 scop
= pet_scop_update_start_end(scop
, start
, end
);
4516 /* Do we need to construct a skip condition of the given type
4517 * on a sequence of statements?
4519 * There is no need to construct a new skip condition if only
4520 * only of the two statements has a skip condition or if both
4521 * of their skip conditions are affine.
4523 * In principle we also don't need a new continuation variable if
4524 * the continuation of scop2 is affine, but then we would need
4525 * to allow more complicated forms of continuations.
4527 static bool need_skip_seq(struct pet_scop
*scop1
, struct pet_scop
*scop2
,
4530 if (!scop1
|| !pet_scop_has_skip(scop1
, type
))
4532 if (!scop2
|| !pet_scop_has_skip(scop2
, type
))
4534 if (pet_scop_has_affine_skip(scop1
, type
) &&
4535 pet_scop_has_affine_skip(scop2
, type
))
4540 /* Construct a scop for computing the skip condition of the given type and
4541 * with access relation "skip_access" for a sequence of two scops "scop1"
4544 * The computed scop contains a single statement that essentially does
4546 * skip_cond = skip_cond_1 ? 1 : skip_cond_2
4548 * or, in other words, skip_cond1 || skip_cond2.
4549 * In this expression, skip_cond_2 is filtered to reflect that it is
4550 * only evaluated when skip_cond_1 is false.
4552 * The skip condition on scop1 is not removed because it still needs
4553 * to be applied to scop2 when these two scops are combined.
4555 static struct pet_scop
*extract_skip_seq(PetScan
*ps
,
4556 __isl_take isl_map
*skip_access
,
4557 struct pet_scop
*scop1
, struct pet_scop
*scop2
, enum pet_skip type
)
4560 struct pet_expr
*expr1
, *expr2
, *expr
, *expr_skip
;
4561 struct pet_stmt
*stmt
;
4562 struct pet_scop
*scop
;
4563 isl_ctx
*ctx
= ps
->ctx
;
4565 if (!scop1
|| !scop2
)
4568 expr1
= pet_scop_get_skip_expr(scop1
, type
);
4569 expr2
= pet_scop_get_skip_expr(scop2
, type
);
4570 pet_scop_reset_skip(scop2
, type
);
4572 expr2
= pet_expr_filter(expr2
, isl_map_copy(expr1
->acc
.access
), 0);
4574 expr
= universally_true(ctx
);
4575 expr
= pet_expr_new_ternary(ctx
, expr1
, expr
, expr2
);
4576 expr_skip
= pet_expr_from_access(isl_map_copy(skip_access
));
4578 expr_skip
->acc
.write
= 1;
4579 expr_skip
->acc
.read
= 0;
4581 expr
= pet_expr_new_binary(ctx
, pet_op_assign
, expr_skip
, expr
);
4582 stmt
= pet_stmt_from_pet_expr(ctx
, -1, NULL
, ps
->n_stmt
++, expr
);
4584 scop
= pet_scop_from_pet_stmt(ctx
, stmt
);
4585 scop
= scop_add_array(scop
, skip_access
, ps
->ast_context
);
4586 isl_map_free(skip_access
);
4590 isl_map_free(skip_access
);
4594 /* Structure that handles the construction of skip conditions
4595 * on sequences of statements.
4597 * scop1 and scop2 represent the two statements that are combined
4599 struct pet_skip_info_seq
: public pet_skip_info
{
4600 struct pet_scop
*scop1
, *scop2
;
4602 pet_skip_info_seq(isl_ctx
*ctx
, struct pet_scop
*scop1
,
4603 struct pet_scop
*scop2
);
4604 void extract(PetScan
*scan
, enum pet_skip type
);
4605 void extract(PetScan
*scan
);
4606 struct pet_scop
*add(struct pet_scop
*scop
, enum pet_skip type
,
4608 struct pet_scop
*add(struct pet_scop
*scop
, int offset
);
4611 /* Initialize a pet_skip_info_seq structure based on
4612 * on the two statements that are going to be combined.
4614 pet_skip_info_seq::pet_skip_info_seq(isl_ctx
*ctx
, struct pet_scop
*scop1
,
4615 struct pet_scop
*scop2
) : pet_skip_info(ctx
), scop1(scop1
), scop2(scop2
)
4617 skip
[pet_skip_now
] = need_skip_seq(scop1
, scop2
, pet_skip_now
);
4618 equal
= skip
[pet_skip_now
] && skip_equals_skip_later(scop1
) &&
4619 skip_equals_skip_later(scop2
);
4620 skip
[pet_skip_later
] = skip
[pet_skip_now
] && !equal
&&
4621 need_skip_seq(scop1
, scop2
, pet_skip_later
);
4624 /* If we need to construct a skip condition of the given type,
4627 void pet_skip_info_seq::extract(PetScan
*scan
, enum pet_skip type
)
4632 access
[type
] = create_test_access(ctx
, scan
->n_test
++);
4633 scop
[type
] = extract_skip_seq(scan
, isl_map_copy(access
[type
]),
4634 scop1
, scop2
, type
);
4637 /* Construct the required skip conditions.
4639 void pet_skip_info_seq::extract(PetScan
*scan
)
4641 extract(scan
, pet_skip_now
);
4642 extract(scan
, pet_skip_later
);
4644 drop_skip_later(scop1
, scop2
);
4647 /* Add the computed skip condition of the given type to "main" and
4648 * add the scop for computing the condition at the given offset (the statement
4649 * number). Within this offset, the condition is computed at position 1
4650 * to ensure that it is computed after the corresponding statement.
4652 * If equal is set, then we only computed a skip condition for pet_skip_now,
4653 * but we also need to set it as main's pet_skip_later.
4655 struct pet_scop
*pet_skip_info_seq::add(struct pet_scop
*main
,
4656 enum pet_skip type
, int offset
)
4663 skip_set
= isl_map_range(access
[type
]);
4664 access
[type
] = NULL
;
4665 scop
[type
] = pet_scop_prefix(scop
[type
], 1);
4666 scop
[type
] = pet_scop_prefix(scop
[type
], offset
);
4667 main
= pet_scop_add_par(ctx
, main
, scop
[type
]);
4671 main
= pet_scop_set_skip(main
, pet_skip_later
,
4672 isl_set_copy(skip_set
));
4674 main
= pet_scop_set_skip(main
, type
, skip_set
);
4679 /* Add the computed skip conditions to "main" and
4680 * add the scops for computing the conditions at the given offset.
4682 struct pet_scop
*pet_skip_info_seq::add(struct pet_scop
*scop
, int offset
)
4684 scop
= add(scop
, pet_skip_now
, offset
);
4685 scop
= add(scop
, pet_skip_later
, offset
);
4690 /* Extract a clone of the kill statement in "scop".
4691 * "scop" is expected to have been created from a DeclStmt
4692 * and should have the kill as its first statement.
4694 struct pet_stmt
*PetScan::extract_kill(struct pet_scop
*scop
)
4696 struct pet_expr
*kill
;
4697 struct pet_stmt
*stmt
;
4702 if (scop
->n_stmt
< 1)
4703 isl_die(ctx
, isl_error_internal
,
4704 "expecting at least one statement", return NULL
);
4705 stmt
= scop
->stmts
[0];
4706 if (stmt
->body
->type
!= pet_expr_unary
||
4707 stmt
->body
->op
!= pet_op_kill
)
4708 isl_die(ctx
, isl_error_internal
,
4709 "expecting kill statement", return NULL
);
4711 access
= isl_map_copy(stmt
->body
->args
[0]->acc
.access
);
4712 access
= isl_map_reset_tuple_id(access
, isl_dim_in
);
4713 kill
= pet_expr_kill_from_access(access
);
4714 return pet_stmt_from_pet_expr(ctx
, stmt
->line
, NULL
, n_stmt
++, kill
);
4717 /* Mark all arrays in "scop" as being exposed.
4719 static struct pet_scop
*mark_exposed(struct pet_scop
*scop
)
4723 for (int i
= 0; i
< scop
->n_array
; ++i
)
4724 scop
->arrays
[i
]->exposed
= 1;
4728 /* Try and construct a pet_scop corresponding to (part of)
4729 * a sequence of statements.
4731 * "block" is set if the sequence respresents the children of
4732 * a compound statement.
4733 * "skip_declarations" is set if we should skip initial declarations
4734 * in the sequence of statements.
4736 * If there are any breaks or continues in the individual statements,
4737 * then we may have to compute a new skip condition.
4738 * This is handled using a pet_skip_info_seq object.
4739 * On initialization, the object checks if skip conditions need
4740 * to be computed. If so, it does so in "extract" and adds them in "add".
4742 * If "block" is set, then we need to insert kill statements at
4743 * the end of the block for any array that has been declared by
4744 * one of the statements in the sequence. Each of these declarations
4745 * results in the construction of a kill statement at the place
4746 * of the declaration, so we simply collect duplicates of
4747 * those kill statements and append these duplicates to the constructed scop.
4749 * If "block" is not set, then any array declared by one of the statements
4750 * in the sequence is marked as being exposed.
4752 struct pet_scop
*PetScan::extract(StmtRange stmt_range
, bool block
,
4753 bool skip_declarations
)
4758 bool partial_range
= false;
4759 set
<struct pet_stmt
*> kills
;
4760 set
<struct pet_stmt
*>::iterator it
;
4762 scop
= pet_scop_empty(ctx
);
4763 for (i
= stmt_range
.first
, j
= 0; i
!= stmt_range
.second
; ++i
, ++j
) {
4765 struct pet_scop
*scop_i
;
4767 if (skip_declarations
&&
4768 child
->getStmtClass() == Stmt::DeclStmtClass
)
4771 scop_i
= extract(child
);
4772 if (scop
&& partial
) {
4773 pet_scop_free(scop_i
);
4776 pet_skip_info_seq
skip(ctx
, scop
, scop_i
);
4779 scop_i
= pet_scop_prefix(scop_i
, 0);
4780 if (scop_i
&& child
->getStmtClass() == Stmt::DeclStmtClass
) {
4782 kills
.insert(extract_kill(scop_i
));
4784 scop_i
= mark_exposed(scop_i
);
4786 scop_i
= pet_scop_prefix(scop_i
, j
);
4787 if (options
->autodetect
) {
4789 scop
= pet_scop_add_seq(ctx
, scop
, scop_i
);
4791 partial_range
= true;
4792 if (scop
->n_stmt
!= 0 && !scop_i
)
4795 scop
= pet_scop_add_seq(ctx
, scop
, scop_i
);
4798 scop
= skip
.add(scop
, j
);
4804 for (it
= kills
.begin(); it
!= kills
.end(); ++it
) {
4806 scop_j
= pet_scop_from_pet_stmt(ctx
, *it
);
4807 scop_j
= pet_scop_prefix(scop_j
, j
);
4808 scop
= pet_scop_add_seq(ctx
, scop
, scop_j
);
4811 if (scop
&& partial_range
) {
4812 if (scop
->n_stmt
== 0) {
4813 pet_scop_free(scop
);
4822 /* Check if the scop marked by the user is exactly this Stmt
4823 * or part of this Stmt.
4824 * If so, return a pet_scop corresponding to the marked region.
4825 * Otherwise, return NULL.
4827 struct pet_scop
*PetScan::scan(Stmt
*stmt
)
4829 SourceManager
&SM
= PP
.getSourceManager();
4830 unsigned start_off
, end_off
;
4832 start_off
= getExpansionOffset(SM
, stmt
->getLocStart());
4833 end_off
= getExpansionOffset(SM
, stmt
->getLocEnd());
4835 if (start_off
> loc
.end
)
4837 if (end_off
< loc
.start
)
4839 if (start_off
>= loc
.start
&& end_off
<= loc
.end
) {
4840 return extract(stmt
);
4844 for (start
= stmt
->child_begin(); start
!= stmt
->child_end(); ++start
) {
4845 Stmt
*child
= *start
;
4848 start_off
= getExpansionOffset(SM
, child
->getLocStart());
4849 end_off
= getExpansionOffset(SM
, child
->getLocEnd());
4850 if (start_off
< loc
.start
&& end_off
>= loc
.end
)
4852 if (start_off
>= loc
.start
)
4857 for (end
= start
; end
!= stmt
->child_end(); ++end
) {
4859 start_off
= SM
.getFileOffset(child
->getLocStart());
4860 if (start_off
>= loc
.end
)
4864 return extract(StmtRange(start
, end
), false, false);
4867 /* Set the size of index "pos" of "array" to "size".
4868 * In particular, add a constraint of the form
4872 * to array->extent and a constraint of the form
4876 * to array->context.
4878 static struct pet_array
*update_size(struct pet_array
*array
, int pos
,
4879 __isl_take isl_pw_aff
*size
)
4889 valid
= isl_pw_aff_nonneg_set(isl_pw_aff_copy(size
));
4890 array
->context
= isl_set_intersect(array
->context
, valid
);
4892 dim
= isl_set_get_space(array
->extent
);
4893 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(dim
));
4894 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, pos
, 1);
4895 univ
= isl_set_universe(isl_aff_get_domain_space(aff
));
4896 index
= isl_pw_aff_alloc(univ
, aff
);
4898 size
= isl_pw_aff_add_dims(size
, isl_dim_in
,
4899 isl_set_dim(array
->extent
, isl_dim_set
));
4900 id
= isl_set_get_tuple_id(array
->extent
);
4901 size
= isl_pw_aff_set_tuple_id(size
, isl_dim_in
, id
);
4902 bound
= isl_pw_aff_lt_set(index
, size
);
4904 array
->extent
= isl_set_intersect(array
->extent
, bound
);
4906 if (!array
->context
|| !array
->extent
)
4911 pet_array_free(array
);
4915 /* Figure out the size of the array at position "pos" and all
4916 * subsequent positions from "type" and update "array" accordingly.
4918 struct pet_array
*PetScan::set_upper_bounds(struct pet_array
*array
,
4919 const Type
*type
, int pos
)
4921 const ArrayType
*atype
;
4927 if (type
->isPointerType()) {
4928 type
= type
->getPointeeType().getTypePtr();
4929 return set_upper_bounds(array
, type
, pos
+ 1);
4931 if (!type
->isArrayType())
4934 type
= type
->getCanonicalTypeInternal().getTypePtr();
4935 atype
= cast
<ArrayType
>(type
);
4937 if (type
->isConstantArrayType()) {
4938 const ConstantArrayType
*ca
= cast
<ConstantArrayType
>(atype
);
4939 size
= extract_affine(ca
->getSize());
4940 array
= update_size(array
, pos
, size
);
4941 } else if (type
->isVariableArrayType()) {
4942 const VariableArrayType
*vla
= cast
<VariableArrayType
>(atype
);
4943 size
= extract_affine(vla
->getSizeExpr());
4944 array
= update_size(array
, pos
, size
);
4947 type
= atype
->getElementType().getTypePtr();
4949 return set_upper_bounds(array
, type
, pos
+ 1);
4952 /* Is "T" the type of a variable length array with static size?
4954 static bool is_vla_with_static_size(QualType T
)
4956 const VariableArrayType
*vlatype
;
4958 if (!T
->isVariableArrayType())
4960 vlatype
= cast
<VariableArrayType
>(T
);
4961 return vlatype
->getSizeModifier() == VariableArrayType::Static
;
4964 /* Return the type of "decl" as an array.
4966 * In particular, if "decl" is a parameter declaration that
4967 * is a variable length array with a static size, then
4968 * return the original type (i.e., the variable length array).
4969 * Otherwise, return the type of decl.
4971 static QualType
get_array_type(ValueDecl
*decl
)
4976 parm
= dyn_cast
<ParmVarDecl
>(decl
);
4978 return decl
->getType();
4980 T
= parm
->getOriginalType();
4981 if (!is_vla_with_static_size(T
))
4982 return decl
->getType();
4986 /* Construct and return a pet_array corresponding to the variable "decl".
4987 * In particular, initialize array->extent to
4989 * { name[i_1,...,i_d] : i_1,...,i_d >= 0 }
4991 * and then call set_upper_bounds to set the upper bounds on the indices
4992 * based on the type of the variable.
4994 struct pet_array
*PetScan::extract_array(isl_ctx
*ctx
, ValueDecl
*decl
)
4996 struct pet_array
*array
;
4997 QualType qt
= get_array_type(decl
);
4998 const Type
*type
= qt
.getTypePtr();
4999 int depth
= array_depth(type
);
5000 QualType base
= base_type(qt
);
5005 array
= isl_calloc_type(ctx
, struct pet_array
);
5009 id
= isl_id_alloc(ctx
, decl
->getName().str().c_str(), decl
);
5010 dim
= isl_space_set_alloc(ctx
, 0, depth
);
5011 dim
= isl_space_set_tuple_id(dim
, isl_dim_set
, id
);
5013 array
->extent
= isl_set_nat_universe(dim
);
5015 dim
= isl_space_params_alloc(ctx
, 0);
5016 array
->context
= isl_set_universe(dim
);
5018 array
= set_upper_bounds(array
, type
, 0);
5022 name
= base
.getAsString();
5023 array
->element_type
= strdup(name
.c_str());
5024 array
->element_size
= decl
->getASTContext().getTypeInfo(base
).first
/ 8;
5029 /* Construct a list of pet_arrays, one for each array (or scalar)
5030 * accessed inside "scop", add this list to "scop" and return the result.
5032 * The context of "scop" is updated with the intersection of
5033 * the contexts of all arrays, i.e., constraints on the parameters
5034 * that ensure that the arrays have a valid (non-negative) size.
5036 struct pet_scop
*PetScan::scan_arrays(struct pet_scop
*scop
)
5039 set
<ValueDecl
*> arrays
;
5040 set
<ValueDecl
*>::iterator it
;
5042 struct pet_array
**scop_arrays
;
5047 pet_scop_collect_arrays(scop
, arrays
);
5048 if (arrays
.size() == 0)
5051 n_array
= scop
->n_array
;
5053 scop_arrays
= isl_realloc_array(ctx
, scop
->arrays
, struct pet_array
*,
5054 n_array
+ arrays
.size());
5057 scop
->arrays
= scop_arrays
;
5059 for (it
= arrays
.begin(), i
= 0; it
!= arrays
.end(); ++it
, ++i
) {
5060 struct pet_array
*array
;
5061 scop
->arrays
[n_array
+ i
] = array
= extract_array(ctx
, *it
);
5062 if (!scop
->arrays
[n_array
+ i
])
5065 scop
->context
= isl_set_intersect(scop
->context
,
5066 isl_set_copy(array
->context
));
5073 pet_scop_free(scop
);
5077 /* Bound all parameters in scop->context to the possible values
5078 * of the corresponding C variable.
5080 static struct pet_scop
*add_parameter_bounds(struct pet_scop
*scop
)
5087 n
= isl_set_dim(scop
->context
, isl_dim_param
);
5088 for (int i
= 0; i
< n
; ++i
) {
5092 id
= isl_set_get_dim_id(scop
->context
, isl_dim_param
, i
);
5093 if (is_nested_parameter(id
)) {
5095 isl_die(isl_set_get_ctx(scop
->context
),
5097 "unresolved nested parameter", goto error
);
5099 decl
= (ValueDecl
*) isl_id_get_user(id
);
5102 scop
->context
= set_parameter_bounds(scop
->context
, i
, decl
);
5110 pet_scop_free(scop
);
5114 /* Construct a pet_scop from the given function.
5116 * If the scop was delimited by scop and endscop pragmas, then we override
5117 * the file offsets by those derived from the pragmas.
5119 struct pet_scop
*PetScan::scan(FunctionDecl
*fd
)
5124 stmt
= fd
->getBody();
5126 if (options
->autodetect
)
5127 scop
= extract(stmt
, true);
5130 scop
= pet_scop_update_start_end(scop
, loc
.start
, loc
.end
);
5132 scop
= pet_scop_detect_parameter_accesses(scop
);
5133 scop
= scan_arrays(scop
);
5134 scop
= add_parameter_bounds(scop
);
5135 scop
= pet_scop_gist(scop
, value_bounds
);