1 //===------ polly/ScopInfo.h -----------------------------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Store the polyhedral model representation of a static control flow region,
11 // also called SCoP (Static Control Part).
13 // This representation is shared among several tools in the polyhedral
14 // community, which are e.g. CLooG, Pluto, Loopo, Graphite.
16 //===----------------------------------------------------------------------===//
18 #ifndef POLLY_SCOP_INFO_H
19 #define POLLY_SCOP_INFO_H
21 #include "polly/ScopDetection.h"
22 #include "polly/Support/SCEVAffinator.h"
24 #include "llvm/ADT/MapVector.h"
25 #include "llvm/Analysis/RegionPass.h"
26 #include "llvm/IR/PassManager.h"
31 #include "isl-noexceptions.h"
34 #include <forward_list>
39 class AssumptionCache
;
43 class ScalarEvolution
;
58 struct isl_constraint
;
60 struct isl_pw_multi_aff
;
70 //===---------------------------------------------------------------------===//
72 extern bool UseInstructionNames
;
74 /// Enumeration of assumptions Polly can take.
88 /// Enum to distinguish between assumptions and restrictions.
89 enum AssumptionSign
{ AS_ASSUMPTION
, AS_RESTRICTION
};
91 /// The different memory kinds used in Polly.
93 /// We distinguish between arrays and various scalar memory objects. We use
94 /// the term ``array'' to describe memory objects that consist of a set of
95 /// individual data elements arranged in a multi-dimensional grid. A scalar
96 /// memory object describes an individual data element and is used to model
97 /// the definition and uses of llvm::Values.
99 /// The polyhedral model does traditionally not reason about SSA values. To
100 /// reason about llvm::Values we model them "as if" they were zero-dimensional
101 /// memory objects, even though they were not actually allocated in (main)
102 /// memory. Memory for such objects is only alloca[ed] at CodeGeneration
103 /// time. To relate the memory slots used during code generation with the
104 /// llvm::Values they belong to the new names for these corresponding stack
105 /// slots are derived by appending suffixes (currently ".s2a" and ".phiops")
106 /// to the name of the original llvm::Value. To describe how def/uses are
107 /// modeled exactly we use these suffixes here as well.
109 /// There are currently four different kinds of memory objects:
110 enum class MemoryKind
{
111 /// MemoryKind::Array: Models a one or multi-dimensional array
113 /// A memory object that can be described by a multi-dimensional array.
114 /// Memory objects of this type are used to model actual multi-dimensional
115 /// arrays as they exist in LLVM-IR, but they are also used to describe
117 /// - A single data element allocated on the stack using 'alloca' is
118 /// modeled as a one-dimensional, single-element array.
119 /// - A single data element allocated as a global variable is modeled as
120 /// one-dimensional, single-element array.
121 /// - Certain multi-dimensional arrays with variable size, which in
122 /// LLVM-IR are commonly expressed as a single-dimensional access with a
123 /// complicated access function, are modeled as multi-dimensional
124 /// memory objects (grep for "delinearization").
127 /// MemoryKind::Value: Models an llvm::Value
129 /// Memory objects of type MemoryKind::Value are used to model the data flow
130 /// induced by llvm::Values. For each llvm::Value that is used across
131 /// BasicBocks one ScopArrayInfo object is created. A single memory WRITE
132 /// stores the llvm::Value at its definition into the memory object and at
133 /// each use of the llvm::Value (ignoring trivial intra-block uses) a
134 /// corresponding READ is added. For instance, the use/def chain of a
135 /// llvm::Value %V depicted below
136 /// ______________________
138 /// | %V = float op ... |
139 /// ----------------------
141 /// _________________ _________________
142 /// |UseBB1: | |UseBB2: |
143 /// | use float %V | | use float %V |
144 /// ----------------- -----------------
146 /// is modeled as if the following memory accesses occurred:
148 /// __________________________
150 /// | %V.s2a = alloca float |
151 /// --------------------------
153 /// ___________________________________
155 /// | store %float %V, float* %V.s2a |
156 /// -----------------------------------
158 /// ____________________________________ ___________________________________
159 /// |UseBB1: | |UseBB2: |
160 /// | %V.reload1 = load float* %V.s2a | | %V.reload2 = load float* %V.s2a|
161 /// | use float %V.reload1 | | use float %V.reload2 |
162 /// ------------------------------------ -----------------------------------
166 /// MemoryKind::PHI: Models PHI nodes within the SCoP
168 /// Besides the MemoryKind::Value memory object used to model the normal
169 /// llvm::Value dependences described above, PHI nodes require an additional
170 /// memory object of type MemoryKind::PHI to describe the forwarding of values
174 /// As an example, a PHIInst instructions
176 /// %PHI = phi float [ %Val1, %IncomingBlock1 ], [ %Val2, %IncomingBlock2 ]
178 /// is modeled as if the accesses occurred this way:
180 /// _______________________________
182 /// | %PHI.phiops = alloca float |
183 /// -------------------------------
185 /// __________________________________ __________________________________
186 /// |IncomingBlock1: | |IncomingBlock2: |
188 /// | store float %Val1 %PHI.phiops | | store float %Val2 %PHI.phiops |
189 /// | br label % JoinBlock | | br label %JoinBlock |
190 /// ---------------------------------- ----------------------------------
193 /// _________________________________________
195 /// | %PHI = load float, float* PHI.phiops |
196 /// -----------------------------------------
198 /// Note that there can also be a scalar write access for %PHI if used in a
199 /// different BasicBlock, i.e. there can be a memory object %PHI.phiops as
200 /// well as a memory object %PHI.s2a.
203 /// MemoryKind::ExitPHI: Models PHI nodes in the SCoP's exit block
205 /// For PHI nodes in the Scop's exit block a special memory object kind is
206 /// used. The modeling used is identical to MemoryKind::PHI, with the
208 /// that there are no READs from these memory objects. The PHINode's
209 /// llvm::Value is treated as a value escaping the SCoP. WRITE accesses
210 /// write directly to the escaping value's ".s2a" alloca.
214 /// Maps from a loop to the affine function expressing its backedge taken count.
215 /// The backedge taken count already enough to express iteration domain as we
216 /// only allow loops with canonical induction variable.
217 /// A canonical induction variable is:
218 /// an integer recurrence that starts at 0 and increments by one each time
219 /// through the loop.
220 typedef std::map
<const Loop
*, const SCEV
*> LoopBoundMapType
;
222 typedef std::vector
<std::unique_ptr
<MemoryAccess
>> AccFuncVector
;
224 /// A class to store information about arrays in the SCoP.
226 /// Objects are accessible via the ScoP, MemoryAccess or the id associated with
227 /// the MemoryAccess access function.
229 class ScopArrayInfo
{
231 /// Construct a ScopArrayInfo object.
233 /// @param BasePtr The array base pointer.
234 /// @param ElementType The type of the elements stored in the array.
235 /// @param IslCtx The isl context used to create the base pointer id.
236 /// @param DimensionSizes A vector containing the size of each dimension.
237 /// @param Kind The kind of the array object.
238 /// @param DL The data layout of the module.
239 /// @param S The scop this array object belongs to.
240 /// @param BaseName The optional name of this memory reference.
241 ScopArrayInfo(Value
*BasePtr
, Type
*ElementType
, isl_ctx
*IslCtx
,
242 ArrayRef
<const SCEV
*> DimensionSizes
, MemoryKind Kind
,
243 const DataLayout
&DL
, Scop
*S
, const char *BaseName
= nullptr);
245 /// Update the element type of the ScopArrayInfo object.
247 /// Memory accesses referencing this ScopArrayInfo object may use
248 /// different element sizes. This function ensures the canonical element type
249 /// stored is small enough to model accesses to the current element type as
250 /// well as to @p NewElementType.
252 /// @param NewElementType An element type that is used to access this array.
253 void updateElementType(Type
*NewElementType
);
255 /// Update the sizes of the ScopArrayInfo object.
257 /// A ScopArrayInfo object may be created without all outer dimensions being
258 /// available. This function is called when new memory accesses are added for
259 /// this ScopArrayInfo object. It verifies that sizes are compatible and adds
260 /// additional outer array dimensions, if needed.
262 /// @param Sizes A vector of array sizes where the rightmost array
263 /// sizes need to match the innermost array sizes already
265 /// @param CheckConsistency Update sizes, even if new sizes are inconsistent
267 bool updateSizes(ArrayRef
<const SCEV
*> Sizes
, bool CheckConsistency
= true);
269 /// Make the ScopArrayInfo model a Fortran array.
270 /// It receives the Fortran array descriptor and stores this.
271 /// It also adds a piecewise expression for the outermost dimension
272 /// since this information is available for Fortran arrays at runtime.
273 void applyAndSetFAD(Value
*FAD
);
275 /// Destructor to free the isl id of the base pointer.
278 /// Set the base pointer to @p BP.
279 void setBasePtr(Value
*BP
) { BasePtr
= BP
; }
281 /// Return the base pointer.
282 Value
*getBasePtr() const { return BasePtr
; }
284 // Set IsOnHeap to the value in parameter.
285 void setIsOnHeap(bool value
) { IsOnHeap
= value
; }
287 /// For indirect accesses return the origin SAI of the BP, else null.
288 const ScopArrayInfo
*getBasePtrOriginSAI() const { return BasePtrOriginSAI
; }
290 /// The set of derived indirect SAIs for this origin SAI.
291 const SmallSetVector
<ScopArrayInfo
*, 2> &getDerivedSAIs() const {
295 /// Return the number of dimensions.
296 unsigned getNumberOfDimensions() const {
297 if (Kind
== MemoryKind::PHI
|| Kind
== MemoryKind::ExitPHI
||
298 Kind
== MemoryKind::Value
)
300 return DimensionSizes
.size();
303 /// Return the size of dimension @p dim as SCEV*.
305 // Scalars do not have array dimensions and the first dimension of
306 // a (possibly multi-dimensional) array also does not carry any size
307 // information, in case the array is not newly created.
308 const SCEV
*getDimensionSize(unsigned Dim
) const {
309 assert(Dim
< getNumberOfDimensions() && "Invalid dimension");
310 return DimensionSizes
[Dim
];
313 /// Return the size of dimension @p dim as isl_pw_aff.
315 // Scalars do not have array dimensions and the first dimension of
316 // a (possibly multi-dimensional) array also does not carry any size
317 // information, in case the array is not newly created.
318 __isl_give isl_pw_aff
*getDimensionSizePw(unsigned Dim
) const {
319 assert(Dim
< getNumberOfDimensions() && "Invalid dimension");
320 return isl_pw_aff_copy(DimensionSizesPw
[Dim
]);
323 /// Get the canonical element type of this array.
325 /// @returns The canonical element type of this array.
326 Type
*getElementType() const { return ElementType
; }
328 /// Get element size in bytes.
329 int getElemSizeInBytes() const;
331 /// Get the name of this memory reference.
332 std::string
getName() const;
334 /// Return the isl id for the base pointer.
335 __isl_give isl_id
*getBasePtrId() const;
337 /// Return what kind of memory this represents.
338 MemoryKind
getKind() const { return Kind
; }
340 /// Is this array info modeling an llvm::Value?
341 bool isValueKind() const { return Kind
== MemoryKind::Value
; }
343 /// Is this array info modeling special PHI node memory?
345 /// During code generation of PHI nodes, there is a need for two kinds of
346 /// virtual storage. The normal one as it is used for all scalar dependences,
347 /// where the result of the PHI node is stored and later loaded from as well
348 /// as a second one where the incoming values of the PHI nodes are stored
349 /// into and reloaded when the PHI is executed. As both memories use the
350 /// original PHI node as virtual base pointer, we have this additional
351 /// attribute to distinguish the PHI node specific array modeling from the
352 /// normal scalar array modeling.
353 bool isPHIKind() const { return Kind
== MemoryKind::PHI
; }
355 /// Is this array info modeling an MemoryKind::ExitPHI?
356 bool isExitPHIKind() const { return Kind
== MemoryKind::ExitPHI
; }
358 /// Is this array info modeling an array?
359 bool isArrayKind() const { return Kind
== MemoryKind::Array
; }
361 /// Is this array allocated on heap
363 /// This property is only relevant if the array is allocated by Polly instead
364 /// of pre-existing. If false, it is allocated using alloca instead malloca.
365 bool isOnHeap() const { return IsOnHeap
; }
367 /// Dump a readable representation to stderr.
370 /// Print a readable representation to @p OS.
372 /// @param SizeAsPwAff Print the size as isl_pw_aff
373 void print(raw_ostream
&OS
, bool SizeAsPwAff
= false) const;
375 /// Access the ScopArrayInfo associated with an access function.
376 static const ScopArrayInfo
*
377 getFromAccessFunction(__isl_keep isl_pw_multi_aff
*PMA
);
379 /// Access the ScopArrayInfo associated with an isl Id.
380 static const ScopArrayInfo
*getFromId(__isl_take isl_id
*Id
);
382 /// Get the space of this array access.
383 __isl_give isl_space
*getSpace() const;
385 /// If the array is read only
388 /// Verify that @p Array is compatible to this ScopArrayInfo.
390 /// Two arrays are compatible if their dimensionality, the sizes of their
391 /// dimensions, and their element sizes match.
393 /// @param Array The array to compare against.
395 /// @returns True, if the arrays are compatible, False otherwise.
396 bool isCompatibleWith(const ScopArrayInfo
*Array
) const;
399 void addDerivedSAI(ScopArrayInfo
*DerivedSAI
) {
400 DerivedSAIs
.insert(DerivedSAI
);
403 /// For indirect accesses this is the SAI of the BP origin.
404 const ScopArrayInfo
*BasePtrOriginSAI
;
406 /// For origin SAIs the set of derived indirect SAIs.
407 SmallSetVector
<ScopArrayInfo
*, 2> DerivedSAIs
;
409 /// The base pointer.
410 AssertingVH
<Value
> BasePtr
;
412 /// The canonical element type of this array.
414 /// The canonical element type describes the minimal accessible element in
415 /// this array. Not all elements accessed, need to be of the very same type,
416 /// but the allocation size of the type of the elements loaded/stored from/to
417 /// this array needs to be a multiple of the allocation size of the canonical
421 /// The isl id for the base pointer.
424 /// True if the newly allocated array is on heap.
427 /// The sizes of each dimension as SCEV*.
428 SmallVector
<const SCEV
*, 4> DimensionSizes
;
430 /// The sizes of each dimension as isl_pw_aff.
431 SmallVector
<isl_pw_aff
*, 4> DimensionSizesPw
;
433 /// The type of this scop array info object.
435 /// We distinguish between SCALAR, PHI and ARRAY objects.
438 /// The data layout of the module.
439 const DataLayout
&DL
;
441 /// The scop this SAI object belongs to.
444 /// If this array models a Fortran array, then this points
445 /// to the Fortran array descriptor.
449 /// Represent memory accesses in statements.
452 friend class ScopStmt
;
455 /// The access type of a memory access
457 /// There are three kind of access types:
461 /// A certain set of memory locations are read and may be used for internal
464 /// * A must-write access
466 /// A certain set of memory locations is definitely written. The old value is
467 /// replaced by a newly calculated value. The old value is not read or used at
470 /// * A may-write access
472 /// A certain set of memory locations may be written. The memory location may
473 /// contain a new value if there is actually a write or the old value may
474 /// remain, if no write happens.
481 /// Reduction access type
483 /// Commutative and associative binary operations suitable for reductions
485 RT_NONE
, ///< Indicate no reduction at all
486 RT_ADD
, ///< Addition
487 RT_MUL
, ///< Multiplication
488 RT_BOR
, ///< Bitwise Or
489 RT_BXOR
, ///< Bitwise XOr
490 RT_BAND
, ///< Bitwise And
494 MemoryAccess(const MemoryAccess
&) = delete;
495 const MemoryAccess
&operator=(const MemoryAccess
&) = delete;
497 /// A unique identifier for this memory access.
499 /// The identifier is unique between all memory accesses belonging to the same
503 /// What is modeled by this MemoryAccess.
507 /// Whether it a reading or writing access, and if writing, whether it
508 /// is conditional (MAY_WRITE).
509 enum AccessType AccType
;
511 /// Reduction type for reduction like accesses, RT_NONE otherwise
513 /// An access is reduction like if it is part of a load-store chain in which
514 /// both access the same memory location (use the same LLVM-IR value
515 /// as pointer reference). Furthermore, between the load and the store there
516 /// is exactly one binary operator which is known to be associative and
521 /// We can later lift the constraint that the same LLVM-IR value defines the
522 /// memory location to handle scops such as the following:
526 /// sum[i+j] = sum[i] + 3;
528 /// Here not all iterations access the same memory location, but iterations
529 /// for which j = 0 holds do. After lifting the equality check in ScopBuilder,
530 /// subsequent transformations do not only need check if a statement is
531 /// reduction like, but they also need to verify that that the reduction
532 /// property is only exploited for statement instances that load from and
533 /// store to the same data location. Doing so at dependence analysis time
534 /// could allow us to handle the above example.
535 ReductionType RedType
= RT_NONE
;
537 /// Parent ScopStmt of this access.
540 /// The domain under which this access is not modeled precisely.
542 /// The invalid domain for an access describes all parameter combinations
543 /// under which the statement looks to be executed but is in fact not because
544 /// some assumption/restriction makes the access invalid.
545 isl_set
*InvalidDomain
;
547 // Properties describing the accessed array.
548 // TODO: It might be possible to move them to ScopArrayInfo.
551 /// The base address (e.g., A for A[i+j]).
553 /// The #BaseAddr of a memory access of kind MemoryKind::Array is the base
554 /// pointer of the memory access.
555 /// The #BaseAddr of a memory access of kind MemoryKind::PHI or
556 /// MemoryKind::ExitPHI is the PHI node itself.
557 /// The #BaseAddr of a memory access of kind MemoryKind::Value is the
558 /// instruction defining the value.
559 AssertingVH
<Value
> BaseAddr
;
561 /// Type a single array element wrt. this access.
564 /// Size of each dimension of the accessed array.
565 SmallVector
<const SCEV
*, 4> Sizes
;
568 // Properties describing the accessed element.
571 /// The access instruction of this memory access.
573 /// For memory accesses of kind MemoryKind::Array the access instruction is
574 /// the Load or Store instruction performing the access.
576 /// For memory accesses of kind MemoryKind::PHI or MemoryKind::ExitPHI the
577 /// access instruction of a load access is the PHI instruction. The access
578 /// instruction of a PHI-store is the incoming's block's terminator
581 /// For memory accesses of kind MemoryKind::Value the access instruction of a
582 /// load access is nullptr because generally there can be multiple
583 /// instructions in the statement using the same llvm::Value. The access
584 /// instruction of a write access is the instruction that defines the
586 Instruction
*AccessInstruction
;
588 /// Incoming block and value of a PHINode.
589 SmallVector
<std::pair
<BasicBlock
*, Value
*>, 4> Incoming
;
591 /// The value associated with this memory access.
593 /// - For array memory accesses (MemoryKind::Array) it is the loaded result
594 /// or the stored value. If the access instruction is a memory intrinsic it
595 /// the access value is also the memory intrinsic.
596 /// - For accesses of kind MemoryKind::Value it is the access instruction
598 /// - For accesses of kind MemoryKind::PHI or MemoryKind::ExitPHI it is the
599 /// PHI node itself (for both, READ and WRITE accesses).
601 AssertingVH
<Value
> AccessValue
;
603 /// Are all the subscripts affine expression?
606 /// Subscript expression for each dimension.
607 SmallVector
<const SCEV
*, 4> Subscripts
;
609 /// Relation from statement instances to the accessed array elements.
611 /// In the common case this relation is a function that maps a set of loop
612 /// indices to the memory address from which a value is loaded/stored:
616 /// S: A[i + 3 j] = ...
618 /// => { S[i,j] -> A[i + 3j] }
620 /// In case the exact access function is not known, the access relation may
621 /// also be a one to all mapping { S[i,j] -> A[o] } describing that any
622 /// element accessible through A might be accessed.
624 /// In case of an access to a larger element belonging to an array that also
625 /// contains smaller elements, the access relation models the larger access
626 /// with multiple smaller accesses of the size of the minimal array element
632 /// S: A[i] = *((double*)&A[4 * i]);
634 /// => { S[i] -> A[i]; S[i] -> A[o] : 4i <= o <= 4i + 3 }
635 isl_map
*AccessRelation
;
637 /// Updated access relation read from JSCOP file.
638 isl_map
*NewAccessRelation
;
640 /// Fortran arrays whose sizes are not statically known are stored in terms
641 /// of a descriptor struct. This maintains a raw pointer to the memory,
642 /// along with auxiliary fields with information such as dimensions.
643 /// We hold a reference to the descriptor corresponding to a MemoryAccess
644 /// into a Fortran array. FAD for "Fortran Array Descriptor"
645 AssertingVH
<Value
> FAD
;
648 __isl_give isl_basic_map
*createBasicAccessMap(ScopStmt
*Statement
);
650 void assumeNoOutOfBound();
652 /// Compute bounds on an over approximated access relation.
654 /// @param ElementSize The size of one element accessed.
655 void computeBoundsOnAccessRelation(unsigned ElementSize
);
657 /// Get the original access function as read from IR.
658 __isl_give isl_map
*getOriginalAccessRelation() const;
660 /// Return the space in which the access relation lives in.
661 __isl_give isl_space
*getOriginalAccessRelationSpace() const;
663 /// Get the new access function imported or set by a pass
664 __isl_give isl_map
*getNewAccessRelation() const;
666 /// Fold the memory access to consider parametric offsets
668 /// To recover memory accesses with array size parameters in the subscript
669 /// expression we post-process the delinearization results.
671 /// We would normally recover from an access A[exp0(i) * N + exp1(i)] into an
672 /// array A[][N] the 2D access A[exp0(i)][exp1(i)]. However, another valid
673 /// delinearization is A[exp0(i) - 1][exp1(i) + N] which - depending on the
674 /// range of exp1(i) - may be preferable. Specifically, for cases where we
675 /// know exp1(i) is negative, we want to choose the latter expression.
677 /// As we commonly do not have any information about the range of exp1(i),
678 /// we do not choose one of the two options, but instead create a piecewise
679 /// access function that adds the (-1, N) offsets as soon as exp1(i) becomes
680 /// negative. For a 2D array such an access function is created by applying
681 /// the piecewise map:
683 /// [i,j] -> [i, j] : j >= 0
684 /// [i,j] -> [i-1, j+N] : j < 0
686 /// We can generalize this mapping to arbitrary dimensions by applying this
687 /// piecewise mapping pairwise from the rightmost to the leftmost access
688 /// dimension. It would also be possible to cover a wider range by introducing
689 /// more cases and adding multiple of Ns to these cases. However, this has
690 /// not yet been necessary.
691 /// The introduction of different cases necessarily complicates the memory
692 /// access function, but cases that can be statically proven to not happen
693 /// will be eliminated later on.
694 void foldAccessRelation();
696 /// Create the access relation for the underlying memory intrinsic.
697 void buildMemIntrinsicAccessRelation();
699 /// Assemble the access relation from all available information.
701 /// In particular, used the information passes in the constructor and the
702 /// parent ScopStmt set by setStatment().
704 /// @param SAI Info object for the accessed array.
705 void buildAccessRelation(const ScopArrayInfo
*SAI
);
707 /// Carry index overflows of dimensions with constant size to the next higher
710 /// For dimensions that have constant size, modulo the index by the size and
711 /// add up the carry (floored division) to the next higher dimension. This is
712 /// how overflow is defined in row-major order.
713 /// It happens e.g. when ScalarEvolution computes the offset to the base
714 /// pointer and would algebraically sum up all lower dimensions' indices of
719 /// A[1][6] -> A[2][2]
720 void wrapConstantDimensions();
723 /// Create a new MemoryAccess.
725 /// @param Stmt The parent statement.
726 /// @param AccessInst The instruction doing the access.
727 /// @param BaseAddr The accessed array's address.
728 /// @param ElemType The type of the accessed array elements.
729 /// @param AccType Whether read or write access.
730 /// @param IsAffine Whether the subscripts are affine expressions.
731 /// @param Kind The kind of memory accessed.
732 /// @param Subscripts Subscript expressions
733 /// @param Sizes Dimension lengths of the accessed array.
734 MemoryAccess(ScopStmt
*Stmt
, Instruction
*AccessInst
, AccessType AccType
,
735 Value
*BaseAddress
, Type
*ElemType
, bool Affine
,
736 ArrayRef
<const SCEV
*> Subscripts
, ArrayRef
<const SCEV
*> Sizes
,
737 Value
*AccessValue
, MemoryKind Kind
);
739 /// Create a new MemoryAccess that corresponds to @p AccRel.
741 /// Along with @p Stmt and @p AccType it uses information about dimension
742 /// lengths of the accessed array, the type of the accessed array elements,
743 /// the name of the accessed array that is derived from the object accessible
746 /// @param Stmt The parent statement.
747 /// @param AccType Whether read or write access.
748 /// @param AccRel The access relation that describes the memory access.
749 MemoryAccess(ScopStmt
*Stmt
, AccessType AccType
, __isl_take isl_map
*AccRel
);
753 /// Add a new incoming block/value pairs for this PHI/ExitPHI access.
755 /// @param IncomingBlock The PHI's incoming block.
756 /// @param IncomingValue The value when reaching the PHI from the @p
758 void addIncoming(BasicBlock
*IncomingBlock
, Value
*IncomingValue
) {
760 assert(isAnyPHIKind());
761 Incoming
.emplace_back(std::make_pair(IncomingBlock
, IncomingValue
));
764 /// Return the list of possible PHI/ExitPHI values.
766 /// After code generation moves some PHIs around during region simplification,
767 /// we cannot reliably locate the original PHI node and its incoming values
768 /// anymore. For this reason we remember these explicitly for all PHI-kind
770 ArrayRef
<std::pair
<BasicBlock
*, Value
*>> getIncoming() const {
771 assert(isAnyPHIKind());
775 /// Get the type of a memory access.
776 enum AccessType
getType() { return AccType
; }
778 /// Is this a reduction like access?
779 bool isReductionLike() const { return RedType
!= RT_NONE
; }
781 /// Is this a read memory access?
782 bool isRead() const { return AccType
== MemoryAccess::READ
; }
784 /// Is this a must-write memory access?
785 bool isMustWrite() const { return AccType
== MemoryAccess::MUST_WRITE
; }
787 /// Is this a may-write memory access?
788 bool isMayWrite() const { return AccType
== MemoryAccess::MAY_WRITE
; }
790 /// Is this a write memory access?
791 bool isWrite() const { return isMustWrite() || isMayWrite(); }
793 /// Is this a memory intrinsic access (memcpy, memset, memmove)?
794 bool isMemoryIntrinsic() const {
795 return isa
<MemIntrinsic
>(getAccessInstruction());
798 /// Check if a new access relation was imported or set by a pass.
799 bool hasNewAccessRelation() const { return NewAccessRelation
; }
801 /// Return the newest access relation of this access.
803 /// There are two possibilities:
804 /// 1) The original access relation read from the LLVM-IR.
805 /// 2) A new access relation imported from a json file or set by another
806 /// pass (e.g., for privatization).
808 /// As 2) is by construction "newer" than 1) we return the new access
809 /// relation if present.
811 __isl_give isl_map
*getLatestAccessRelation() const {
812 return hasNewAccessRelation() ? getNewAccessRelation()
813 : getOriginalAccessRelation();
816 /// Old name of getLatestAccessRelation().
817 __isl_give isl_map
*getAccessRelation() const {
818 return getLatestAccessRelation();
821 /// Get an isl map describing the memory address accessed.
823 /// In most cases the memory address accessed is well described by the access
824 /// relation obtained with getAccessRelation. However, in case of arrays
825 /// accessed with types of different size the access relation maps one access
826 /// to multiple smaller address locations. This method returns an isl map that
827 /// relates each dynamic statement instance to the unique memory location
828 /// that is loaded from / stored to.
830 /// For an access relation { S[i] -> A[o] : 4i <= o <= 4i + 3 } this method
831 /// will return the address function { S[i] -> A[4i] }.
833 /// @returns The address function for this memory access.
834 __isl_give isl_map
*getAddressFunction() const;
836 /// Return the access relation after the schedule was applied.
837 __isl_give isl_pw_multi_aff
*
838 applyScheduleToAccessRelation(__isl_take isl_union_map
*Schedule
) const;
840 /// Get an isl string representing the access function read from IR.
841 std::string
getOriginalAccessRelationStr() const;
843 /// Get an isl string representing a new access function, if available.
844 std::string
getNewAccessRelationStr() const;
846 /// Get the original base address of this access (e.g. A for A[i+j]) when
849 /// This adress may differ from the base address referenced by the Original
850 /// ScopArrayInfo to which this array belongs, as this memory access may
851 /// have been unified to a ScopArray which has a different but identically
852 /// valued base pointer in case invariant load hoisting is enabled.
853 Value
*getOriginalBaseAddr() const { return BaseAddr
; }
855 /// Get the detection-time base array isl_id for this access.
856 __isl_give isl_id
*getOriginalArrayId() const;
858 /// Get the base array isl_id for this access, modifiable through
859 /// setNewAccessRelation().
860 __isl_give isl_id
*getLatestArrayId() const;
862 /// Old name of getOriginalArrayId().
863 __isl_give isl_id
*getArrayId() const { return getOriginalArrayId(); }
865 /// Get the detection-time ScopArrayInfo object for the base address.
866 const ScopArrayInfo
*getOriginalScopArrayInfo() const;
868 /// Get the ScopArrayInfo object for the base address, or the one set
869 /// by setNewAccessRelation().
870 const ScopArrayInfo
*getLatestScopArrayInfo() const;
872 /// Legacy name of getOriginalScopArrayInfo().
873 const ScopArrayInfo
*getScopArrayInfo() const {
874 return getOriginalScopArrayInfo();
877 /// Return a string representation of the access's reduction type.
878 const std::string
getReductionOperatorStr() const;
880 /// Return a string representation of the reduction type @p RT.
881 static const std::string
getReductionOperatorStr(ReductionType RT
);
883 /// Return the element type of the accessed array wrt. this access.
884 Type
*getElementType() const { return ElementType
; }
886 /// Return the access value of this memory access.
887 Value
*getAccessValue() const { return AccessValue
; }
889 /// Return the access instruction of this memory access.
890 Instruction
*getAccessInstruction() const { return AccessInstruction
; }
892 /// Return the number of access function subscript.
893 unsigned getNumSubscripts() const { return Subscripts
.size(); }
895 /// Return the access function subscript in the dimension @p Dim.
896 const SCEV
*getSubscript(unsigned Dim
) const { return Subscripts
[Dim
]; }
898 /// Compute the isl representation for the SCEV @p E wrt. this access.
900 /// Note that this function will also adjust the invalid context accordingly.
901 __isl_give isl_pw_aff
*getPwAff(const SCEV
*E
);
903 /// Get the invalid domain for this access.
904 __isl_give isl_set
*getInvalidDomain() const {
905 return isl_set_copy(InvalidDomain
);
908 /// Get the invalid context for this access.
909 __isl_give isl_set
*getInvalidContext() const {
910 return isl_set_params(getInvalidDomain());
913 /// Get the stride of this memory access in the specified Schedule. Schedule
914 /// is a map from the statement to a schedule where the innermost dimension is
915 /// the dimension of the innermost loop containing the statement.
916 __isl_give isl_set
*getStride(__isl_take
const isl_map
*Schedule
) const;
918 /// Get the FortranArrayDescriptor corresponding to this memory access if
919 /// it exists, and nullptr otherwise.
920 Value
*getFortranArrayDescriptor() const { return this->FAD
; };
922 /// Is the stride of the access equal to a certain width? Schedule is a map
923 /// from the statement to a schedule where the innermost dimension is the
924 /// dimension of the innermost loop containing the statement.
925 bool isStrideX(__isl_take
const isl_map
*Schedule
, int StrideWidth
) const;
927 /// Is consecutive memory accessed for a given statement instance set?
928 /// Schedule is a map from the statement to a schedule where the innermost
929 /// dimension is the dimension of the innermost loop containing the
931 bool isStrideOne(__isl_take
const isl_map
*Schedule
) const;
933 /// Is always the same memory accessed for a given statement instance set?
934 /// Schedule is a map from the statement to a schedule where the innermost
935 /// dimension is the dimension of the innermost loop containing the
937 bool isStrideZero(__isl_take
const isl_map
*Schedule
) const;
939 /// Return the kind when this access was first detected.
940 MemoryKind
getOriginalKind() const {
941 assert(!getOriginalScopArrayInfo() /* not yet initialized */ ||
942 getOriginalScopArrayInfo()->getKind() == Kind
);
946 /// Return the kind considering a potential setNewAccessRelation.
947 MemoryKind
getLatestKind() const {
948 return getLatestScopArrayInfo()->getKind();
951 /// Whether this is an access of an explicit load or store in the IR.
952 bool isOriginalArrayKind() const {
953 return getOriginalKind() == MemoryKind::Array
;
956 /// Whether storage memory is either an custom .s2a/.phiops alloca
957 /// (false) or an existing pointer into an array (true).
958 bool isLatestArrayKind() const {
959 return getLatestKind() == MemoryKind::Array
;
962 /// Old name of isOriginalArrayKind.
963 bool isArrayKind() const { return isOriginalArrayKind(); }
965 /// Whether this access is an array to a scalar memory object, without
966 /// considering changes by setNewAccessRelation.
968 /// Scalar accesses are accesses to MemoryKind::Value, MemoryKind::PHI or
969 /// MemoryKind::ExitPHI.
970 bool isOriginalScalarKind() const {
971 return getOriginalKind() != MemoryKind::Array
;
974 /// Whether this access is an array to a scalar memory object, also
975 /// considering changes by setNewAccessRelation.
976 bool isLatestScalarKind() const {
977 return getLatestKind() != MemoryKind::Array
;
980 /// Old name of isOriginalScalarKind.
981 bool isScalarKind() const { return isOriginalScalarKind(); }
983 /// Was this MemoryAccess detected as a scalar dependences?
984 bool isOriginalValueKind() const {
985 return getOriginalKind() == MemoryKind::Value
;
988 /// Is this MemoryAccess currently modeling scalar dependences?
989 bool isLatestValueKind() const {
990 return getLatestKind() == MemoryKind::Value
;
993 /// Old name of isOriginalValueKind().
994 bool isValueKind() const { return isOriginalValueKind(); }
996 /// Was this MemoryAccess detected as a special PHI node access?
997 bool isOriginalPHIKind() const {
998 return getOriginalKind() == MemoryKind::PHI
;
1001 /// Is this MemoryAccess modeling special PHI node accesses, also
1002 /// considering a potential change by setNewAccessRelation?
1003 bool isLatestPHIKind() const { return getLatestKind() == MemoryKind::PHI
; }
1005 /// Old name of isOriginalPHIKind.
1006 bool isPHIKind() const { return isOriginalPHIKind(); }
1008 /// Was this MemoryAccess detected as the accesses of a PHI node in the
1009 /// SCoP's exit block?
1010 bool isOriginalExitPHIKind() const {
1011 return getOriginalKind() == MemoryKind::ExitPHI
;
1014 /// Is this MemoryAccess modeling the accesses of a PHI node in the
1015 /// SCoP's exit block? Can be changed to an array access using
1016 /// setNewAccessRelation().
1017 bool isLatestExitPHIKind() const {
1018 return getLatestKind() == MemoryKind::ExitPHI
;
1021 /// Old name of isOriginalExitPHIKind().
1022 bool isExitPHIKind() const { return isOriginalExitPHIKind(); }
1024 /// Was this access detected as one of the two PHI types?
1025 bool isOriginalAnyPHIKind() const {
1026 return isOriginalPHIKind() || isOriginalExitPHIKind();
1029 /// Does this access originate from one of the two PHI types? Can be
1030 /// changed to an array access using setNewAccessRelation().
1031 bool isLatestAnyPHIKind() const {
1032 return isLatestPHIKind() || isLatestExitPHIKind();
1035 /// Old name of isOriginalAnyPHIKind().
1036 bool isAnyPHIKind() const { return isOriginalAnyPHIKind(); }
1038 /// Get the statement that contains this memory access.
1039 ScopStmt
*getStatement() const { return Statement
; }
1041 /// Get the reduction type of this access
1042 ReductionType
getReductionType() const { return RedType
; }
1044 /// Set the array descriptor corresponding to the Array on which the
1045 /// memory access is performed.
1046 void setFortranArrayDescriptor(Value
*FAD
);
1048 /// Update the original access relation.
1050 /// We need to update the original access relation during scop construction,
1051 /// when unifying the memory accesses that access the same scop array info
1052 /// object. After the scop has been constructed, the original access relation
1053 /// should not be changed any more. Instead setNewAccessRelation should
1055 void setAccessRelation(__isl_take isl_map
*AccessRelation
);
1057 /// Set the updated access relation read from JSCOP file.
1058 void setNewAccessRelation(__isl_take isl_map
*NewAccessRelation
);
1060 /// Return whether the MemoryyAccess is a partial access. That is, the access
1061 /// is not executed in some instances of the parent statement's domain.
1062 bool isLatestPartialAccess() const;
1064 /// Mark this a reduction like access
1065 void markAsReductionLike(ReductionType RT
) { RedType
= RT
; }
1067 /// Align the parameters in the access relation to the scop context
1068 void realignParams();
1070 /// Update the dimensionality of the memory access.
1072 /// During scop construction some memory accesses may not be constructed with
1073 /// their full dimensionality, but outer dimensions may have been omitted if
1074 /// they took the value 'zero'. By updating the dimensionality of the
1075 /// statement we add additional zero-valued dimensions to match the
1076 /// dimensionality of the ScopArrayInfo object that belongs to this memory
1078 void updateDimensionality();
1080 /// Get identifier for the memory access.
1082 /// This identifier is unique for all accesses that belong to the same scop
1084 __isl_give isl_id
*getId() const;
1086 /// Print the MemoryAccess.
1088 /// @param OS The output stream the MemoryAccess is printed to.
1089 void print(raw_ostream
&OS
) const;
1091 /// Print the MemoryAccess to stderr.
1094 /// Is the memory access affine?
1095 bool isAffine() const { return IsAffine
; }
1098 llvm::raw_ostream
&operator<<(llvm::raw_ostream
&OS
,
1099 MemoryAccess::ReductionType RT
);
1101 /// Ordered list type to hold accesses.
1102 using MemoryAccessList
= std::forward_list
<MemoryAccess
*>;
1104 /// Helper structure for invariant memory accesses.
1105 struct InvariantAccess
{
1106 /// The memory access that is (partially) invariant.
1109 /// The context under which the access is not invariant.
1110 isl_set
*NonHoistableCtx
;
1113 /// Ordered container type to hold invariant accesses.
1114 using InvariantAccessesTy
= SmallVector
<InvariantAccess
, 8>;
1116 /// Type for equivalent invariant accesses and their domain context.
1117 struct InvariantEquivClassTy
{
1119 /// The pointer that identifies this equivalence class
1120 const SCEV
*IdentifyingPointer
;
1122 /// Memory accesses now treated invariant
1124 /// These memory accesses access the pointer location that identifies
1125 /// this equivalence class. They are treated as invariant and hoisted during
1126 /// code generation.
1127 MemoryAccessList InvariantAccesses
;
1129 /// The execution context under which the memory location is accessed
1131 /// It is the union of the execution domains of the memory accesses in the
1132 /// InvariantAccesses list.
1133 isl_set
*ExecutionContext
;
1135 /// The type of the invariant access
1137 /// It is used to differentiate between differently typed invariant loads from
1138 /// the same location.
1142 /// Type for invariant accesses equivalence classes.
1143 using InvariantEquivClassesTy
= SmallVector
<InvariantEquivClassTy
, 8>;
1145 /// Statement of the Scop
1147 /// A Scop statement represents an instruction in the Scop.
1149 /// It is further described by its iteration domain, its schedule and its data
1151 /// At the moment every statement represents a single basic block of LLVM-IR.
1154 ScopStmt(const ScopStmt
&) = delete;
1155 const ScopStmt
&operator=(const ScopStmt
&) = delete;
1157 /// Create the ScopStmt from a BasicBlock.
1158 ScopStmt(Scop
&parent
, BasicBlock
&bb
, Loop
*SurroundingLoop
,
1159 std::vector
<Instruction
*> Instructions
);
1161 /// Create an overapproximating ScopStmt for the region @p R.
1162 ScopStmt(Scop
&parent
, Region
&R
, Loop
*SurroundingLoop
);
1164 /// Create a copy statement.
1166 /// @param Stmt The parent statement.
1167 /// @param SourceRel The source location.
1168 /// @param TargetRel The target location.
1169 /// @param Domain The original domain under which the copy statement would
1171 ScopStmt(Scop
&parent
, __isl_take isl_map
*SourceRel
,
1172 __isl_take isl_map
*TargetRel
, __isl_take isl_set
*Domain
);
1174 /// Initialize members after all MemoryAccesses have been added.
1175 void init(LoopInfo
&LI
);
1178 /// Polyhedral description
1181 /// The Scop containing this ScopStmt.
1184 /// The domain under which this statement is not modeled precisely.
1186 /// The invalid domain for a statement describes all parameter combinations
1187 /// under which the statement looks to be executed but is in fact not because
1188 /// some assumption/restriction makes the statement/scop invalid.
1189 isl_set
*InvalidDomain
;
1191 /// The iteration domain describes the set of iterations for which this
1192 /// statement is executed.
1195 /// for (i = 0; i < 100 + b; ++i)
1196 /// for (j = 0; j < i; ++j)
1199 /// 'S' is executed for different values of i and j. A vector of all
1200 /// induction variables around S (i, j) is called iteration vector.
1201 /// The domain describes the set of possible iteration vectors.
1203 /// In this case it is:
1205 /// Domain: 0 <= i <= 100 + b
1208 /// A pair of statement and iteration vector (S, (5,3)) is called statement
1212 /// The memory accesses of this statement.
1214 /// The only side effects of a statement are its memory accesses.
1215 typedef SmallVector
<MemoryAccess
*, 8> MemoryAccessVec
;
1216 MemoryAccessVec MemAccs
;
1218 /// Mapping from instructions to (scalar) memory accesses.
1219 DenseMap
<const Instruction
*, MemoryAccessList
> InstructionToAccess
;
1221 /// The set of values defined elsewhere required in this ScopStmt and
1222 /// their MemoryKind::Value READ MemoryAccesses.
1223 DenseMap
<Value
*, MemoryAccess
*> ValueReads
;
1225 /// The set of values defined in this ScopStmt that are required
1226 /// elsewhere, mapped to their MemoryKind::Value WRITE MemoryAccesses.
1227 DenseMap
<Instruction
*, MemoryAccess
*> ValueWrites
;
1229 /// Map from PHI nodes to its incoming value when coming from this
1232 /// Non-affine subregions can have multiple exiting blocks that are incoming
1233 /// blocks of the PHI nodes. This map ensures that there is only one write
1234 /// operation for the complete subregion. A PHI selecting the relevant value
1235 /// will be inserted.
1236 DenseMap
<PHINode
*, MemoryAccess
*> PHIWrites
;
1240 /// A SCoP statement represents either a basic block (affine/precise case) or
1241 /// a whole region (non-affine case).
1243 /// Only one of the following two members will therefore be set and indicate
1244 /// which kind of statement this is.
1248 /// The BasicBlock represented by this statement (in the affine case).
1251 /// The region represented by this statement (in the non-affine case).
1256 /// The isl AST build for the new generated AST.
1257 isl_ast_build
*Build
;
1259 SmallVector
<Loop
*, 4> NestLoops
;
1261 std::string BaseName
;
1263 /// The closest loop that contains this statement.
1264 Loop
*SurroundingLoop
;
1266 /// Vector for Instructions in a BB.
1267 std::vector
<Instruction
*> Instructions
;
1269 /// Build the statement.
1273 /// Fill NestLoops with loops surrounding this statement.
1274 void collectSurroundingLoops();
1276 /// Build the access relation of all memory accesses.
1277 void buildAccessRelations();
1279 /// Detect and mark reductions in the ScopStmt
1280 void checkForReductions();
1282 /// Collect loads which might form a reduction chain with @p StoreMA
1284 collectCandiateReductionLoads(MemoryAccess
*StoreMA
,
1285 llvm::SmallVectorImpl
<MemoryAccess
*> &Loads
);
1288 /// Remove @p MA from dictionaries pointing to them.
1289 void removeAccessData(MemoryAccess
*MA
);
1294 /// Get an isl_ctx pointer.
1295 isl_ctx
*getIslCtx() const;
1297 /// Get the iteration domain of this ScopStmt.
1299 /// @return The iteration domain of this ScopStmt.
1300 __isl_give isl_set
*getDomain() const;
1302 /// Get the space of the iteration domain
1304 /// @return The space of the iteration domain
1305 __isl_give isl_space
*getDomainSpace() const;
1307 /// Get the id of the iteration domain space
1309 /// @return The id of the iteration domain space
1310 __isl_give isl_id
*getDomainId() const;
1312 /// Get an isl string representing this domain.
1313 std::string
getDomainStr() const;
1315 /// Get the schedule function of this ScopStmt.
1317 /// @return The schedule function of this ScopStmt, if it does not contain
1318 /// extension nodes, and nullptr, otherwise.
1319 __isl_give isl_map
*getSchedule() const;
1321 /// Get an isl string representing this schedule.
1323 /// @return An isl string representing this schedule, if it does not contain
1324 /// extension nodes, and an empty string, otherwise.
1325 std::string
getScheduleStr() const;
1327 /// Get the invalid domain for this statement.
1328 __isl_give isl_set
*getInvalidDomain() const {
1329 return isl_set_copy(InvalidDomain
);
1332 /// Get the invalid context for this statement.
1333 __isl_give isl_set
*getInvalidContext() const {
1334 return isl_set_params(getInvalidDomain());
1337 /// Set the invalid context for this statement to @p ID.
1338 void setInvalidDomain(__isl_take isl_set
*ID
);
1340 /// Get the BasicBlock represented by this ScopStmt (if any).
1342 /// @return The BasicBlock represented by this ScopStmt, or null if the
1343 /// statement represents a region.
1344 BasicBlock
*getBasicBlock() const { return BB
; }
1346 /// Return true if this statement represents a single basic block.
1347 bool isBlockStmt() const { return BB
!= nullptr; }
1349 /// Return true if this is a copy statement.
1350 bool isCopyStmt() const { return BB
== nullptr && R
== nullptr; }
1352 /// Get the region represented by this ScopStmt (if any).
1354 /// @return The region represented by this ScopStmt, or null if the statement
1355 /// represents a basic block.
1356 Region
*getRegion() const { return R
; }
1358 /// Return true if this statement represents a whole region.
1359 bool isRegionStmt() const { return R
!= nullptr; }
1361 /// Return a BasicBlock from this statement.
1363 /// For block statements, it returns the BasicBlock itself. For subregion
1364 /// statements, return its entry block.
1365 BasicBlock
*getEntryBlock() const;
1367 /// Return whether @p L is boxed within this statement.
1368 bool contains(const Loop
*L
) const {
1369 // Block statements never contain loops.
1373 return getRegion()->contains(L
);
1376 /// Return whether this statement contains @p BB.
1377 bool contains(BasicBlock
*BB
) const {
1381 return BB
== getBasicBlock();
1382 return getRegion()->contains(BB
);
1385 /// Return whether this statement contains @p Inst.
1386 bool contains(Instruction
*Inst
) const {
1389 return contains(Inst
->getParent());
1392 /// Return the closest innermost loop that contains this statement, but is not
1393 /// contained in it.
1395 /// For block statement, this is just the loop that contains the block. Region
1396 /// statements can contain boxed loops, so getting the loop of one of the
1397 /// region's BBs might return such an inner loop. For instance, the region's
1398 /// entry could be a header of a loop, but the region might extend to BBs
1399 /// after the loop exit. Similarly, the region might only contain parts of the
1400 /// loop body and still include the loop header.
1402 /// Most of the time the surrounding loop is the top element of #NestLoops,
1403 /// except when it is empty. In that case it return the loop that the whole
1404 /// SCoP is contained in. That can be nullptr if there is no such loop.
1405 Loop
*getSurroundingLoop() const {
1406 assert(!isCopyStmt() &&
1407 "No surrounding loop for artificially created statements");
1408 return SurroundingLoop
;
1411 /// Return true if this statement does not contain any accesses.
1412 bool isEmpty() const { return MemAccs
.empty(); }
1414 /// Return the only array access for @p Inst, if existing.
1416 /// @param Inst The instruction for which to look up the access.
1417 /// @returns The unique array memory access related to Inst or nullptr if
1418 /// no array access exists
1419 MemoryAccess
*getArrayAccessOrNULLFor(const Instruction
*Inst
) const {
1420 auto It
= InstructionToAccess
.find(Inst
);
1421 if (It
== InstructionToAccess
.end())
1424 MemoryAccess
*ArrayAccess
= nullptr;
1426 for (auto Access
: It
->getSecond()) {
1427 if (!Access
->isArrayKind())
1430 assert(!ArrayAccess
&& "More then one array access for instruction");
1432 ArrayAccess
= Access
;
1438 /// Return the only array access for @p Inst.
1440 /// @param Inst The instruction for which to look up the access.
1441 /// @returns The unique array memory access related to Inst.
1442 MemoryAccess
&getArrayAccessFor(const Instruction
*Inst
) const {
1443 MemoryAccess
*ArrayAccess
= getArrayAccessOrNULLFor(Inst
);
1445 assert(ArrayAccess
&& "No array access found for instruction!");
1446 return *ArrayAccess
;
1449 /// Return the MemoryAccess that writes the value of an instruction
1450 /// defined in this statement, or nullptr if not existing, respectively
1452 MemoryAccess
*lookupValueWriteOf(Instruction
*Inst
) const {
1453 assert((isRegionStmt() && R
->contains(Inst
)) ||
1454 (!isRegionStmt() && Inst
->getParent() == BB
));
1455 return ValueWrites
.lookup(Inst
);
1458 /// Return the MemoryAccess that reloads a value, or nullptr if not
1459 /// existing, respectively not yet added.
1460 MemoryAccess
*lookupValueReadOf(Value
*Inst
) const {
1461 return ValueReads
.lookup(Inst
);
1464 /// Return the MemoryAccess that loads a PHINode value, or nullptr if not
1465 /// existing, respectively not yet added.
1466 MemoryAccess
*lookupPHIReadOf(PHINode
*PHI
) const;
1468 /// Return the PHI write MemoryAccess for the incoming values from any
1469 /// basic block in this ScopStmt, or nullptr if not existing,
1470 /// respectively not yet added.
1471 MemoryAccess
*lookupPHIWriteOf(PHINode
*PHI
) const {
1472 assert(isBlockStmt() || R
->getExit() == PHI
->getParent());
1473 return PHIWrites
.lookup(PHI
);
1476 /// Return the input access of the value, or null if no such MemoryAccess
1479 /// The input access is the MemoryAccess that makes an inter-statement value
1480 /// available in this statement by reading it at the start of this statement.
1481 /// This can be a MemoryKind::Value if defined in another statement or a
1482 /// MemoryKind::PHI if the value is a PHINode in this statement.
1483 MemoryAccess
*lookupInputAccessOf(Value
*Val
) const {
1484 if (isa
<PHINode
>(Val
))
1485 if (auto InputMA
= lookupPHIReadOf(cast
<PHINode
>(Val
))) {
1486 assert(!lookupValueReadOf(Val
) && "input accesses must be unique; a "
1487 "statement cannot read a .s2a and "
1488 ".phiops simultaneously");
1492 if (auto *InputMA
= lookupValueReadOf(Val
))
1498 /// Add @p Access to this statement's list of accesses.
1499 void addAccess(MemoryAccess
*Access
);
1501 /// Remove a MemoryAccess from this statement.
1503 /// Note that scalar accesses that are caused by MA will
1504 /// be eliminated too.
1505 void removeMemoryAccess(MemoryAccess
*MA
);
1507 /// Remove @p MA from this statement.
1509 /// In contrast to removeMemoryAccess(), no other access will be eliminated.
1510 void removeSingleMemoryAccess(MemoryAccess
*MA
);
1512 typedef MemoryAccessVec::iterator iterator
;
1513 typedef MemoryAccessVec::const_iterator const_iterator
;
1515 iterator
begin() { return MemAccs
.begin(); }
1516 iterator
end() { return MemAccs
.end(); }
1517 const_iterator
begin() const { return MemAccs
.begin(); }
1518 const_iterator
end() const { return MemAccs
.end(); }
1519 size_t size() const { return MemAccs
.size(); }
1521 unsigned getNumIterators() const;
1523 Scop
*getParent() { return &Parent
; }
1524 const Scop
*getParent() const { return &Parent
; }
1526 const std::vector
<Instruction
*> &getInstructions() const {
1527 return Instructions
;
1530 const char *getBaseName() const;
1532 /// Set the isl AST build.
1533 void setAstBuild(__isl_keep isl_ast_build
*B
) { Build
= B
; }
1535 /// Get the isl AST build.
1536 __isl_keep isl_ast_build
*getAstBuild() const { return Build
; }
1538 /// Restrict the domain of the statement.
1540 /// @param NewDomain The new statement domain.
1541 void restrictDomain(__isl_take isl_set
*NewDomain
);
1543 /// Get the loop for a dimension.
1545 /// @param Dimension The dimension of the induction variable
1546 /// @return The loop at a certain dimension.
1547 Loop
*getLoopForDimension(unsigned Dimension
) const;
1549 /// Align the parameters in the statement to the scop context
1550 void realignParams();
1552 /// Print the ScopStmt.
1554 /// @param OS The output stream the ScopStmt is printed to.
1555 void print(raw_ostream
&OS
) const;
1557 /// Print the instructions in ScopStmt.
1559 void printInstructions(raw_ostream
&OS
) const;
1561 /// Print the ScopStmt to stderr.
1565 /// Print ScopStmt S to raw_ostream O.
1566 static inline raw_ostream
&operator<<(raw_ostream
&O
, const ScopStmt
&S
) {
1571 /// Static Control Part
1573 /// A Scop is the polyhedral representation of a control flow region detected
1574 /// by the Scop detection. It is generated by translating the LLVM-IR and
1575 /// abstracting its effects.
1577 /// A Scop consists of a set of:
1579 /// * A set of statements executed in the Scop.
1581 /// * A set of global parameters
1582 /// Those parameters are scalar integer values, which are constant during
1586 /// This context contains information about the values the parameters
1587 /// can take and relations between different parameters.
1590 /// Type to represent a pair of minimal/maximal access to an array.
1591 using MinMaxAccessTy
= std::pair
<isl_pw_multi_aff
*, isl_pw_multi_aff
*>;
1593 /// Vector of minimal/maximal accesses to different arrays.
1594 using MinMaxVectorTy
= SmallVector
<MinMaxAccessTy
, 4>;
1596 /// Pair of minimal/maximal access vectors representing
1597 /// read write and read only accesses
1598 using MinMaxVectorPairTy
= std::pair
<MinMaxVectorTy
, MinMaxVectorTy
>;
1600 /// Vector of pair of minimal/maximal access vectors representing
1601 /// non read only and read only accesses for each alias group.
1602 using MinMaxVectorPairVectorTy
= SmallVector
<MinMaxVectorPairTy
, 4>;
1605 Scop(const Scop
&) = delete;
1606 const Scop
&operator=(const Scop
&) = delete;
1608 ScalarEvolution
*SE
;
1610 /// The underlying Region.
1613 /// The name of the SCoP (identical to the regions name)
1616 // Access functions of the SCoP.
1618 // This owns all the MemoryAccess objects of the Scop created in this pass.
1619 AccFuncVector AccessFunctions
;
1621 /// Flag to indicate that the scheduler actually optimized the SCoP.
1624 /// True if the underlying region has a single exiting block.
1625 bool HasSingleExitEdge
;
1627 /// Flag to remember if the SCoP contained an error block or not.
1631 unsigned MaxLoopDepth
;
1633 /// Number of copy statements.
1634 unsigned CopyStmtsNum
;
1636 typedef std::list
<ScopStmt
> StmtSet
;
1637 /// The statements in this Scop.
1640 /// Parameters of this Scop
1641 ParameterSetTy Parameters
;
1643 /// Mapping from parameters to their ids.
1644 DenseMap
<const SCEV
*, isl_id
*> ParameterIds
;
1646 /// The context of the SCoP created during SCoP detection.
1647 ScopDetection::DetectionContext
&DC
;
1651 /// We need a shared_ptr with reference counter to delete the context when all
1652 /// isl objects are deleted. We will distribute the shared_ptr to all objects
1653 /// that use the context to create isl objects, and increase the reference
1654 /// counter. By doing this, we guarantee that the context is deleted when we
1655 /// delete the last object that creates isl objects with the context.
1656 std::shared_ptr
<isl_ctx
> IslCtx
;
1658 /// A map from basic blocks to SCoP statements.
1659 DenseMap
<BasicBlock
*, ScopStmt
*> StmtMap
;
1661 /// A map from basic blocks to their domains.
1662 DenseMap
<BasicBlock
*, isl_set
*> DomainMap
;
1664 /// Constraints on parameters.
1667 /// The affinator used to translate SCEVs to isl expressions.
1668 SCEVAffinator Affinator
;
1670 typedef std::map
<std::pair
<AssertingVH
<const Value
>, MemoryKind
>,
1671 std::unique_ptr
<ScopArrayInfo
>>
1674 typedef StringMap
<std::unique_ptr
<ScopArrayInfo
>> ArrayNameMapTy
;
1676 typedef SetVector
<ScopArrayInfo
*> ArrayInfoSetTy
;
1678 /// A map to remember ScopArrayInfo objects for all base pointers.
1680 /// As PHI nodes may have two array info objects associated, we add a flag
1681 /// that distinguishes between the PHI node specific ArrayInfo object
1682 /// and the normal one.
1683 ArrayInfoMapTy ScopArrayInfoMap
;
1685 /// A map to remember ScopArrayInfo objects for all names of memory
1687 ArrayNameMapTy ScopArrayNameMap
;
1689 /// A set to remember ScopArrayInfo objects.
1690 /// @see Scop::ScopArrayInfoMap
1691 ArrayInfoSetTy ScopArrayInfoSet
;
1693 /// The assumptions under which this scop was built.
1695 /// When constructing a scop sometimes the exact representation of a statement
1696 /// or condition would be very complex, but there is a common case which is a
1697 /// lot simpler, but which is only valid under certain assumptions. The
1698 /// assumed context records the assumptions taken during the construction of
1699 /// this scop and that need to be code generated as a run-time test.
1700 isl_set
*AssumedContext
;
1702 /// The restrictions under which this SCoP was built.
1704 /// The invalid context is similar to the assumed context as it contains
1705 /// constraints over the parameters. However, while we need the constraints
1706 /// in the assumed context to be "true" the constraints in the invalid context
1707 /// need to be "false". Otherwise they behave the same.
1708 isl_set
*InvalidContext
;
1710 /// Helper struct to remember assumptions.
1713 /// The kind of the assumption (e.g., WRAPPING).
1714 AssumptionKind Kind
;
1716 /// Flag to distinguish assumptions and restrictions.
1717 AssumptionSign Sign
;
1719 /// The valid/invalid context if this is an assumption/restriction.
1722 /// The location that caused this assumption.
1725 /// An optional block whose domain can simplify the assumption.
1729 /// Collection to hold taken assumptions.
1731 /// There are two reasons why we want to record assumptions first before we
1732 /// add them to the assumed/invalid context:
1733 /// 1) If the SCoP is not profitable or otherwise invalid without the
1734 /// assumed/invalid context we do not have to compute it.
1735 /// 2) Information about the context are gathered rather late in the SCoP
1736 /// construction (basically after we know all parameters), thus the user
1737 /// might see overly complicated assumptions to be taken while they will
1738 /// only be simplified later on.
1739 SmallVector
<Assumption
, 8> RecordedAssumptions
;
1741 /// The schedule of the SCoP
1743 /// The schedule of the SCoP describes the execution order of the statements
1744 /// in the scop by assigning each statement instance a possibly
1745 /// multi-dimensional execution time. The schedule is stored as a tree of
1748 /// The most common nodes in a schedule tree are so-called band nodes. Band
1749 /// nodes map statement instances into a multi dimensional schedule space.
1750 /// This space can be seen as a multi-dimensional clock.
1754 /// <S,(5,4)> may be mapped to (5,4) by this schedule:
1756 /// s0 = i (Year of execution)
1757 /// s1 = j (Day of execution)
1759 /// or to (9, 20) by this schedule:
1761 /// s0 = i + j (Year of execution)
1762 /// s1 = 20 (Day of execution)
1764 /// The order statement instances are executed is defined by the
1765 /// schedule vectors they are mapped to. A statement instance
1766 /// <A, (i, j, ..)> is executed before a statement instance <B, (i', ..)>, if
1767 /// the schedule vector of A is lexicographic smaller than the schedule
1770 /// Besides band nodes, schedule trees contain additional nodes that specify
1771 /// a textual ordering between two subtrees or filter nodes that filter the
1772 /// set of statement instances that will be scheduled in a subtree. There
1773 /// are also several other nodes. A full description of the different nodes
1774 /// in a schedule tree is given in the isl manual.
1775 isl_schedule
*Schedule
;
1777 /// The set of minimal/maximal accesses for each alias group.
1779 /// When building runtime alias checks we look at all memory instructions and
1780 /// build so called alias groups. Each group contains a set of accesses to
1781 /// different base arrays which might alias with each other. However, between
1782 /// alias groups there is no aliasing possible.
1784 /// In a program with int and float pointers annotated with tbaa information
1785 /// we would probably generate two alias groups, one for the int pointers and
1786 /// one for the float pointers.
1788 /// During code generation we will create a runtime alias check for each alias
1789 /// group to ensure the SCoP is executed in an alias free environment.
1790 MinMaxVectorPairVectorTy MinMaxAliasGroups
;
1792 /// Mapping from invariant loads to the representing invariant load of
1793 /// their equivalence class.
1794 ValueToValueMap InvEquivClassVMap
;
1796 /// List of invariant accesses.
1797 InvariantEquivClassesTy InvariantEquivClasses
;
1799 /// The smallest array index not yet assigned.
1802 /// The smallest statement index not yet assigned.
1805 /// Scop constructor; invoked from ScopBuilder::buildScop.
1806 Scop(Region
&R
, ScalarEvolution
&SE
, LoopInfo
&LI
,
1807 ScopDetection::DetectionContext
&DC
);
1811 /// Initialize this ScopBuilder.
1812 void init(AliasAnalysis
&AA
, AssumptionCache
&AC
, DominatorTree
&DT
,
1815 /// Propagate domains that are known due to graph properties.
1817 /// As a CFG is mostly structured we use the graph properties to propagate
1818 /// domains without the need to compute all path conditions. In particular, if
1819 /// a block A dominates a block B and B post-dominates A we know that the
1820 /// domain of B is a superset of the domain of A. As we do not have
1821 /// post-dominator information available here we use the less precise region
1822 /// information. Given a region R, we know that the exit is always executed if
1823 /// the entry was executed, thus the domain of the exit is a superset of the
1824 /// domain of the entry. In case the exit can only be reached from within the
1825 /// region the domains are in fact equal. This function will use this property
1826 /// to avoid the generation of condition constraints that determine when a
1827 /// branch is taken. If @p BB is a region entry block we will propagate its
1828 /// domain to the region exit block. Additionally, we put the region exit
1829 /// block in the @p FinishedExitBlocks set so we can later skip edges from
1830 /// within the region to that block.
1832 /// @param BB The block for which the domain is currently
1834 /// @param BBLoop The innermost affine loop surrounding @p BB.
1835 /// @param FinishedExitBlocks Set of region exits the domain was set for.
1836 /// @param LI The LoopInfo for the current function.
1837 /// @param InvalidDomainMap BB to InvalidDomain map for the BB of current
1839 void propagateDomainConstraintsToRegionExit(
1840 BasicBlock
*BB
, Loop
*BBLoop
,
1841 SmallPtrSetImpl
<BasicBlock
*> &FinishedExitBlocks
, LoopInfo
&LI
,
1842 DenseMap
<BasicBlock
*, __isl_keep isl_set
*> &InvalidDomainMap
);
1844 /// Compute the union of predecessor domains for @p BB.
1846 /// To compute the union of all domains of predecessors of @p BB this
1847 /// function applies similar reasoning on the CFG structure as described for
1848 /// @see propagateDomainConstraintsToRegionExit
1850 /// @param BB The block for which the predecessor domains are collected.
1851 /// @param Domain The domain under which BB is executed.
1852 /// @param DT The DominatorTree for the current function.
1853 /// @param LI The LoopInfo for the current function.
1855 /// @returns The domain under which @p BB is executed.
1856 __isl_give isl_set
*
1857 getPredecessorDomainConstraints(BasicBlock
*BB
, __isl_keep isl_set
*Domain
,
1858 DominatorTree
&DT
, LoopInfo
&LI
);
1860 /// Add loop carried constraints to the header block of the loop @p L.
1862 /// @param L The loop to process.
1863 /// @param LI The LoopInfo for the current function.
1864 /// @param InvalidDomainMap BB to InvalidDomain map for the BB of current
1867 /// @returns True if there was no problem and false otherwise.
1868 bool addLoopBoundsToHeaderDomain(
1869 Loop
*L
, LoopInfo
&LI
,
1870 DenseMap
<BasicBlock
*, __isl_keep isl_set
*> &InvalidDomainMap
);
1872 /// Compute the branching constraints for each basic block in @p R.
1874 /// @param R The region we currently build branching conditions
1876 /// @param DT The DominatorTree for the current function.
1877 /// @param LI The LoopInfo for the current function.
1878 /// @param InvalidDomainMap BB to InvalidDomain map for the BB of current
1881 /// @returns True if there was no problem and false otherwise.
1882 bool buildDomainsWithBranchConstraints(
1883 Region
*R
, DominatorTree
&DT
, LoopInfo
&LI
,
1884 DenseMap
<BasicBlock
*, __isl_keep isl_set
*> &InvalidDomainMap
);
1886 /// Propagate the domain constraints through the region @p R.
1888 /// @param R The region we currently build branching conditions
1890 /// @param DT The DominatorTree for the current function.
1891 /// @param LI The LoopInfo for the current function.
1892 /// @param InvalidDomainMap BB to InvalidDomain map for the BB of current
1895 /// @returns True if there was no problem and false otherwise.
1896 bool propagateDomainConstraints(
1897 Region
*R
, DominatorTree
&DT
, LoopInfo
&LI
,
1898 DenseMap
<BasicBlock
*, __isl_keep isl_set
*> &InvalidDomainMap
);
1900 /// Propagate invalid domains of statements through @p R.
1902 /// This method will propagate invalid statement domains through @p R and at
1903 /// the same time add error block domains to them. Additionally, the domains
1904 /// of error statements and those only reachable via error statements will be
1905 /// replaced by an empty set. Later those will be removed completely.
1907 /// @param R The currently traversed region.
1908 /// @param DT The DominatorTree for the current function.
1909 /// @param LI The LoopInfo for the current function.
1910 /// @param InvalidDomainMap BB to InvalidDomain map for the BB of current
1913 /// @returns True if there was no problem and false otherwise.
1914 bool propagateInvalidStmtDomains(
1915 Region
*R
, DominatorTree
&DT
, LoopInfo
&LI
,
1916 DenseMap
<BasicBlock
*, __isl_keep isl_set
*> &InvalidDomainMap
);
1918 /// Compute the domain for each basic block in @p R.
1920 /// @param R The region we currently traverse.
1921 /// @param DT The DominatorTree for the current function.
1922 /// @param LI The LoopInfo for the current function.
1923 /// @param InvalidDomainMap BB to InvalidDomain map for the BB of current
1926 /// @returns True if there was no problem and false otherwise.
1928 buildDomains(Region
*R
, DominatorTree
&DT
, LoopInfo
&LI
,
1929 DenseMap
<BasicBlock
*, __isl_keep isl_set
*> &InvalidDomainMap
);
1931 /// Add parameter constraints to @p C that imply a non-empty domain.
1932 __isl_give isl_set
*addNonEmptyDomainConstraints(__isl_take isl_set
*C
) const;
1934 /// Return the access for the base ptr of @p MA if any.
1935 MemoryAccess
*lookupBasePtrAccess(MemoryAccess
*MA
);
1937 /// Check if the base ptr of @p MA is in the SCoP but not hoistable.
1938 bool hasNonHoistableBasePtrInScop(MemoryAccess
*MA
, isl::union_map Writes
);
1940 /// Create equivalence classes for required invariant accesses.
1942 /// These classes will consolidate multiple required invariant loads from the
1943 /// same address in order to keep the number of dimensions in the SCoP
1944 /// description small. For each such class equivalence class only one
1945 /// representing element, hence one required invariant load, will be chosen
1946 /// and modeled as parameter. The method
1947 /// Scop::getRepresentingInvariantLoadSCEV() will replace each element from an
1948 /// equivalence class with the representing element that is modeled. As a
1949 /// consequence Scop::getIdForParam() will only return an id for the
1950 /// representing element of each equivalence class, thus for each required
1951 /// invariant location.
1952 void buildInvariantEquivalenceClasses();
1954 /// Return the context under which the access cannot be hoisted.
1956 /// @param Access The access to check.
1957 /// @param Writes The set of all memory writes in the scop.
1959 /// @return Return the context under which the access cannot be hoisted or a
1960 /// nullptr if it cannot be hoisted at all.
1961 isl::set
getNonHoistableCtx(MemoryAccess
*Access
, isl::union_map Writes
);
1963 /// Verify that all required invariant loads have been hoisted.
1965 /// Invariant load hoisting is not guaranteed to hoist all loads that were
1966 /// assumed to be scop invariant during scop detection. This function checks
1967 /// for cases where the hoisting failed, but where it would have been
1968 /// necessary for our scop modeling to be correct. In case of insufficient
1969 /// hoisting the scop is marked as invalid.
1971 /// In the example below Bound[1] is required to be invariant:
1973 /// for (int i = 1; i < Bound[0]; i++)
1974 /// for (int j = 1; j < Bound[1]; j++)
1977 void verifyInvariantLoads();
1979 /// Hoist invariant memory loads and check for required ones.
1981 /// We first identify "common" invariant loads, thus loads that are invariant
1982 /// and can be hoisted. Then we check if all required invariant loads have
1983 /// been identified as (common) invariant. A load is a required invariant load
1984 /// if it was assumed to be invariant during SCoP detection, e.g., to assume
1985 /// loop bounds to be affine or runtime alias checks to be placeable. In case
1986 /// a required invariant load was not identified as (common) invariant we will
1987 /// drop this SCoP. An example for both "common" as well as required invariant
1988 /// loads is given below:
1990 /// for (int i = 1; i < *LB[0]; i++)
1991 /// for (int j = 1; j < *LB[1]; j++)
1992 /// A[i][j] += A[0][0] + (*V);
1994 /// Common inv. loads: V, A[0][0], LB[0], LB[1]
1995 /// Required inv. loads: LB[0], LB[1], (V, if it may alias with A or LB)
1997 void hoistInvariantLoads();
1999 /// Canonicalize arrays with base pointers from the same equivalence class.
2001 /// Some context: in our normal model we assume that each base pointer is
2002 /// related to a single specific memory region, where memory regions
2003 /// associated with different base pointers are disjoint. Consequently we do
2004 /// not need to compute additional data dependences that model possible
2005 /// overlaps of these memory regions. To verify our assumption we compute
2006 /// alias checks that verify that modeled arrays indeed do not overlap. In
2007 /// case an overlap is detected the runtime check fails and we fall back to
2008 /// the original code.
2010 /// In case of arrays where the base pointers are know to be identical,
2011 /// because they are dynamically loaded by accesses that are in the same
2012 /// invariant load equivalence class, such run-time alias check would always
2015 /// This function makes sure that we do not generate consistently failing
2016 /// run-time checks for code that contains distinct arrays with known
2017 /// equivalent base pointers. It identifies for each invariant load
2018 /// equivalence class a single canonical array and canonicalizes all memory
2019 /// accesses that reference arrays that have base pointers that are known to
2020 /// be equal to the base pointer of such a canonical array to this canonical
2023 /// We currently do not canonicalize arrays for which certain memory accesses
2024 /// have been hoisted as loop invariant.
2025 void canonicalizeDynamicBasePtrs();
2027 /// Add invariant loads listed in @p InvMAs with the domain of @p Stmt.
2028 void addInvariantLoads(ScopStmt
&Stmt
, InvariantAccessesTy
&InvMAs
);
2030 /// Create an id for @p Param and store it in the ParameterIds map.
2031 void createParameterId(const SCEV
*Param
);
2033 /// Build the Context of the Scop.
2034 void buildContext();
2036 /// Add user provided parameter constraints to context (source code).
2037 void addUserAssumptions(
2038 AssumptionCache
&AC
, DominatorTree
&DT
, LoopInfo
&LI
,
2039 DenseMap
<BasicBlock
*, __isl_keep isl_set
*> &InvalidDomainMap
);
2041 /// Add user provided parameter constraints to context (command line).
2042 void addUserContext();
2044 /// Add the bounds of the parameters to the context.
2045 void addParameterBounds();
2047 /// Simplify the assumed and invalid context.
2048 void simplifyContexts();
2050 /// Get the representing SCEV for @p S if applicable, otherwise @p S.
2052 /// Invariant loads of the same location are put in an equivalence class and
2053 /// only one of them is chosen as a representing element that will be
2054 /// modeled as a parameter. The others have to be normalized, i.e.,
2055 /// replaced by the representing element of their equivalence class, in order
2056 /// to get the correct parameter value, e.g., in the SCEVAffinator.
2058 /// @param S The SCEV to normalize.
2060 /// @return The representing SCEV for invariant loads or @p S if none.
2061 const SCEV
*getRepresentingInvariantLoadSCEV(const SCEV
*S
);
2063 /// Create a new SCoP statement for @p BB.
2065 /// A new statement for @p BB will be created and added to the statement
2069 /// @param BB The basic block we build the statement for.
2070 /// @param SurroundingLoop The loop the created statement is contained in.
2071 /// @param Instructions The instructions in the basic block.
2072 void addScopStmt(BasicBlock
*BB
, Loop
*SurroundingLoop
,
2073 std::vector
<Instruction
*> Instructions
);
2075 /// Create a new SCoP statement for @p R.
2077 /// A new statement for @p R will be created and added to the statement vector
2080 /// @param R The region we build the statement for.
2081 /// @param SurroundingLoop The loop the created statement is contained in.
2082 void addScopStmt(Region
*R
, Loop
*SurroundingLoop
);
2084 /// Update access dimensionalities.
2086 /// When detecting memory accesses different accesses to the same array may
2087 /// have built with different dimensionality, as outer zero-values dimensions
2088 /// may not have been recognized as separate dimensions. This function goes
2089 /// again over all memory accesses and updates their dimensionality to match
2090 /// the dimensionality of the underlying ScopArrayInfo object.
2091 void updateAccessDimensionality();
2093 /// Fold size constants to the right.
2095 /// In case all memory accesses in a given dimension are multiplied with a
2096 /// common constant, we can remove this constant from the individual access
2097 /// functions and move it to the size of the memory access. We do this as this
2098 /// increases the size of the innermost dimension, consequently widens the
2099 /// valid range the array subscript in this dimension can evaluate to, and
2100 /// as a result increases the likelihood that our delinearization is
2106 /// S[i,j] -> A[2i][2j+1]
2107 /// S[i,j] -> A[2i][2j]
2112 /// S[i,j] -> A[i][2j+1]
2113 /// S[i,j] -> A[i][2j]
2115 /// Constants in outer dimensions can arise when the elements of a parametric
2116 /// multi-dimensional array are not elementary data types, but e.g.,
2118 void foldSizeConstantsToRight();
2120 /// Fold memory accesses to handle parametric offset.
2122 /// As a post-processing step, we 'fold' memory accesses to parametric
2123 /// offsets in the access functions. @see MemoryAccess::foldAccess for
2125 void foldAccessRelations();
2127 /// Assume that all memory accesses are within bounds.
2129 /// After we have built a model of all memory accesses, we need to assume
2130 /// that the model we built matches reality -- aka. all modeled memory
2131 /// accesses always remain within bounds. We do this as last step, after
2132 /// all memory accesses have been modeled and canonicalized.
2133 void assumeNoOutOfBounds();
2135 /// Mark arrays that have memory accesses with FortranArrayDescriptor.
2136 void markFortranArrays();
2138 /// Finalize all access relations.
2140 /// When building up access relations, temporary access relations that
2141 /// correctly represent each individual access are constructed. However, these
2142 /// access relations can be inconsistent or non-optimal when looking at the
2143 /// set of accesses as a whole. This function finalizes the memory accesses
2144 /// and constructs a globally consistent state.
2145 void finalizeAccesses();
2147 /// Construct the schedule of this SCoP.
2149 /// @param LI The LoopInfo for the current function.
2150 void buildSchedule(LoopInfo
&LI
);
2152 /// A loop stack element to keep track of per-loop information during
2153 /// schedule construction.
2154 typedef struct LoopStackElement
{
2155 // The loop for which we keep information.
2158 // The (possibly incomplete) schedule for this loop.
2159 isl_schedule
*Schedule
;
2161 // The number of basic blocks in the current loop, for which a schedule has
2162 // already been constructed.
2163 unsigned NumBlocksProcessed
;
2165 LoopStackElement(Loop
*L
, __isl_give isl_schedule
*S
,
2166 unsigned NumBlocksProcessed
)
2167 : L(L
), Schedule(S
), NumBlocksProcessed(NumBlocksProcessed
) {}
2168 } LoopStackElementTy
;
2170 /// The loop stack used for schedule construction.
2172 /// The loop stack keeps track of schedule information for a set of nested
2173 /// loops as well as an (optional) 'nullptr' loop that models the outermost
2174 /// schedule dimension. The loops in a loop stack always have a parent-child
2175 /// relation where the loop at position n is the parent of the loop at
2177 typedef SmallVector
<LoopStackElementTy
, 4> LoopStackTy
;
2179 /// Construct schedule information for a given Region and add the
2180 /// derived information to @p LoopStack.
2182 /// Given a Region we derive schedule information for all RegionNodes
2183 /// contained in this region ensuring that the assigned execution times
2184 /// correctly model the existing control flow relations.
2186 /// @param R The region which to process.
2187 /// @param LoopStack A stack of loops that are currently under
2189 /// @param LI The LoopInfo for the current function.
2190 void buildSchedule(Region
*R
, LoopStackTy
&LoopStack
, LoopInfo
&LI
);
2192 /// Build Schedule for the region node @p RN and add the derived
2193 /// information to @p LoopStack.
2195 /// In case @p RN is a BasicBlock or a non-affine Region, we construct the
2196 /// schedule for this @p RN and also finalize loop schedules in case the
2197 /// current @p RN completes the loop.
2199 /// In case @p RN is a not-non-affine Region, we delegate the construction to
2200 /// buildSchedule(Region *R, ...).
2202 /// @param RN The RegionNode region traversed.
2203 /// @param LoopStack A stack of loops that are currently under
2205 /// @param LI The LoopInfo for the current function.
2206 void buildSchedule(RegionNode
*RN
, LoopStackTy
&LoopStack
, LoopInfo
&LI
);
2208 /// Collect all memory access relations of a given type.
2210 /// @param Predicate A predicate function that returns true if an access is
2211 /// of a given type.
2213 /// @returns The set of memory accesses in the scop that match the predicate.
2214 __isl_give isl_union_map
*
2215 getAccessesOfType(std::function
<bool(MemoryAccess
&)> Predicate
);
2217 /// @name Helper functions for printing the Scop.
2220 void printContext(raw_ostream
&OS
) const;
2221 void printArrayInfo(raw_ostream
&OS
) const;
2222 void printStatements(raw_ostream
&OS
) const;
2223 void printAliasAssumptions(raw_ostream
&OS
) const;
2226 friend class ScopBuilder
;
2231 /// Get the count of copy statements added to this Scop.
2233 /// @return The count of copy statements added to this Scop.
2234 unsigned getCopyStmtsNum() { return CopyStmtsNum
; }
2236 /// Create a new copy statement.
2238 /// A new statement will be created and added to the statement vector.
2240 /// @param Stmt The parent statement.
2241 /// @param SourceRel The source location.
2242 /// @param TargetRel The target location.
2243 /// @param Domain The original domain under which the copy statement would
2245 ScopStmt
*addScopStmt(__isl_take isl_map
*SourceRel
,
2246 __isl_take isl_map
*TargetRel
,
2247 __isl_take isl_set
*Domain
);
2249 /// Add the access function to all MemoryAccess objects of the Scop
2250 /// created in this pass.
2251 void addAccessFunction(MemoryAccess
*Access
) {
2252 AccessFunctions
.emplace_back(Access
);
2255 ScalarEvolution
*getSE() const;
2257 /// Get the count of parameters used in this Scop.
2259 /// @return The count of parameters used in this Scop.
2260 size_t getNumParams() const { return Parameters
.size(); }
2262 /// Take a list of parameters and add the new ones to the scop.
2263 void addParams(const ParameterSetTy
&NewParameters
);
2265 /// Return an iterator range containing the scop parameters.
2266 iterator_range
<ParameterSetTy::iterator
> parameters() const {
2267 return make_range(Parameters
.begin(), Parameters
.end());
2270 /// Return whether this scop is empty, i.e. contains no statements that
2271 /// could be executed.
2272 bool isEmpty() const { return Stmts
.empty(); }
2274 const StringRef
getName() const { return name
; }
2276 typedef ArrayInfoSetTy::iterator array_iterator
;
2277 typedef ArrayInfoSetTy::const_iterator const_array_iterator
;
2278 typedef iterator_range
<ArrayInfoSetTy::iterator
> array_range
;
2279 typedef iterator_range
<ArrayInfoSetTy::const_iterator
> const_array_range
;
2281 inline array_iterator
array_begin() { return ScopArrayInfoSet
.begin(); }
2283 inline array_iterator
array_end() { return ScopArrayInfoSet
.end(); }
2285 inline const_array_iterator
array_begin() const {
2286 return ScopArrayInfoSet
.begin();
2289 inline const_array_iterator
array_end() const {
2290 return ScopArrayInfoSet
.end();
2293 inline array_range
arrays() {
2294 return array_range(array_begin(), array_end());
2297 inline const_array_range
arrays() const {
2298 return const_array_range(array_begin(), array_end());
2301 /// Return the isl_id that represents a certain parameter.
2303 /// @param Parameter A SCEV that was recognized as a Parameter.
2305 /// @return The corresponding isl_id or NULL otherwise.
2306 __isl_give isl_id
*getIdForParam(const SCEV
*Parameter
);
2308 /// Get the maximum region of this static control part.
2310 /// @return The maximum region of this static control part.
2311 inline const Region
&getRegion() const { return R
; }
2312 inline Region
&getRegion() { return R
; }
2314 /// Return the function this SCoP is in.
2315 Function
&getFunction() const { return *R
.getEntry()->getParent(); }
2317 /// Check if @p L is contained in the SCoP.
2318 bool contains(const Loop
*L
) const { return R
.contains(L
); }
2320 /// Check if @p BB is contained in the SCoP.
2321 bool contains(const BasicBlock
*BB
) const { return R
.contains(BB
); }
2323 /// Check if @p I is contained in the SCoP.
2324 bool contains(const Instruction
*I
) const { return R
.contains(I
); }
2326 /// Return the unique exit block of the SCoP.
2327 BasicBlock
*getExit() const { return R
.getExit(); }
2329 /// Return the unique exiting block of the SCoP if any.
2330 BasicBlock
*getExitingBlock() const { return R
.getExitingBlock(); }
2332 /// Return the unique entry block of the SCoP.
2333 BasicBlock
*getEntry() const { return R
.getEntry(); }
2335 /// Return the unique entering block of the SCoP if any.
2336 BasicBlock
*getEnteringBlock() const { return R
.getEnteringBlock(); }
2338 /// Return true if @p BB is the exit block of the SCoP.
2339 bool isExit(BasicBlock
*BB
) const { return getExit() == BB
; }
2341 /// Return a range of all basic blocks in the SCoP.
2342 Region::block_range
blocks() const { return R
.blocks(); }
2344 /// Return true if and only if @p BB dominates the SCoP.
2345 bool isDominatedBy(const DominatorTree
&DT
, BasicBlock
*BB
) const;
2347 /// Get the maximum depth of the loop.
2349 /// @return The maximum depth of the loop.
2350 inline unsigned getMaxLoopDepth() const { return MaxLoopDepth
; }
2352 /// Return the invariant equivalence class for @p Val if any.
2353 InvariantEquivClassTy
*lookupInvariantEquivClass(Value
*Val
);
2355 /// Return the set of invariant accesses.
2356 InvariantEquivClassesTy
&getInvariantAccesses() {
2357 return InvariantEquivClasses
;
2360 /// Check if the scop has any invariant access.
2361 bool hasInvariantAccesses() { return !InvariantEquivClasses
.empty(); }
2363 /// Mark the SCoP as optimized by the scheduler.
2364 void markAsOptimized() { IsOptimized
= true; }
2366 /// Check if the SCoP has been optimized by the scheduler.
2367 bool isOptimized() const { return IsOptimized
; }
2369 /// Get the name of the entry and exit blocks of this Scop.
2371 /// These along with the function name can uniquely identify a Scop.
2373 /// @return std::pair whose first element is the entry name & second element
2374 /// is the exit name.
2375 std::pair
<std::string
, std::string
> getEntryExitStr() const;
2377 /// Get the name of this Scop.
2378 std::string
getNameStr() const;
2380 /// Get the constraint on parameter of this Scop.
2382 /// @return The constraint on parameter of this Scop.
2383 __isl_give isl_set
*getContext() const;
2384 __isl_give isl_space
*getParamSpace() const;
2386 /// Get the assumed context for this Scop.
2388 /// @return The assumed context of this Scop.
2389 __isl_give isl_set
*getAssumedContext() const;
2391 /// Return true if the optimized SCoP can be executed.
2393 /// In addition to the runtime check context this will also utilize the domain
2394 /// constraints to decide it the optimized version can actually be executed.
2396 /// @returns True if the optimized SCoP can be executed.
2397 bool hasFeasibleRuntimeContext() const;
2399 /// Check if the assumption in @p Set is trivial or not.
2401 /// @param Set The relations between parameters that are assumed to hold.
2402 /// @param Sign Enum to indicate if the assumptions in @p Set are positive
2403 /// (needed/assumptions) or negative (invalid/restrictions).
2405 /// @returns True if the assumption @p Set is not trivial.
2406 bool isEffectiveAssumption(__isl_keep isl_set
*Set
, AssumptionSign Sign
);
2408 /// Track and report an assumption.
2410 /// Use 'clang -Rpass-analysis=polly-scops' or 'opt
2411 /// -pass-remarks-analysis=polly-scops' to output the assumptions.
2413 /// @param Kind The assumption kind describing the underlying cause.
2414 /// @param Set The relations between parameters that are assumed to hold.
2415 /// @param Loc The location in the source that caused this assumption.
2416 /// @param Sign Enum to indicate if the assumptions in @p Set are positive
2417 /// (needed/assumptions) or negative (invalid/restrictions).
2419 /// @returns True if the assumption is not trivial.
2420 bool trackAssumption(AssumptionKind Kind
, __isl_keep isl_set
*Set
,
2421 DebugLoc Loc
, AssumptionSign Sign
);
2423 /// Add assumptions to assumed context.
2425 /// The assumptions added will be assumed to hold during the execution of the
2426 /// scop. However, as they are generally not statically provable, at code
2427 /// generation time run-time checks will be generated that ensure the
2428 /// assumptions hold.
2430 /// WARNING: We currently exploit in simplifyAssumedContext the knowledge
2431 /// that assumptions do not change the set of statement instances
2434 /// @param Kind The assumption kind describing the underlying cause.
2435 /// @param Set The relations between parameters that are assumed to hold.
2436 /// @param Loc The location in the source that caused this assumption.
2437 /// @param Sign Enum to indicate if the assumptions in @p Set are positive
2438 /// (needed/assumptions) or negative (invalid/restrictions).
2439 void addAssumption(AssumptionKind Kind
, __isl_take isl_set
*Set
, DebugLoc Loc
,
2440 AssumptionSign Sign
);
2442 /// Record an assumption for later addition to the assumed context.
2444 /// This function will add the assumption to the RecordedAssumptions. This
2445 /// collection will be added (@see addAssumption) to the assumed context once
2446 /// all paramaters are known and the context is fully build.
2448 /// @param Kind The assumption kind describing the underlying cause.
2449 /// @param Set The relations between parameters that are assumed to hold.
2450 /// @param Loc The location in the source that caused this assumption.
2451 /// @param Sign Enum to indicate if the assumptions in @p Set are positive
2452 /// (needed/assumptions) or negative (invalid/restrictions).
2453 /// @param BB The block in which this assumption was taken. If it is
2454 /// set, the domain of that block will be used to simplify the
2455 /// actual assumption in @p Set once it is added. This is useful
2456 /// if the assumption was created prior to the domain.
2457 void recordAssumption(AssumptionKind Kind
, __isl_take isl_set
*Set
,
2458 DebugLoc Loc
, AssumptionSign Sign
,
2459 BasicBlock
*BB
= nullptr);
2461 /// Add all recorded assumptions to the assumed context.
2462 void addRecordedAssumptions();
2464 /// Mark the scop as invalid.
2466 /// This method adds an assumption to the scop that is always invalid. As a
2467 /// result, the scop will not be optimized later on. This function is commonly
2468 /// called when a condition makes it impossible (or too compile time
2469 /// expensive) to process this scop any further.
2471 /// @param Kind The assumption kind describing the underlying cause.
2472 /// @param Loc The location in the source that triggered .
2473 void invalidate(AssumptionKind Kind
, DebugLoc Loc
);
2475 /// Get the invalid context for this Scop.
2477 /// @return The invalid context of this Scop.
2478 __isl_give isl_set
*getInvalidContext() const;
2480 /// Return true if and only if the InvalidContext is trivial (=empty).
2481 bool hasTrivialInvalidContext() const {
2482 return isl_set_is_empty(InvalidContext
);
2485 /// A vector of memory accesses that belong to an alias group.
2486 typedef SmallVector
<MemoryAccess
*, 4> AliasGroupTy
;
2488 /// A vector of alias groups.
2489 typedef SmallVector
<Scop::AliasGroupTy
, 4> AliasGroupVectorTy
;
2491 /// Build the alias checks for this SCoP.
2492 bool buildAliasChecks(AliasAnalysis
&AA
);
2494 /// Build all alias groups for this SCoP.
2496 /// @returns True if __no__ error occurred, false otherwise.
2497 bool buildAliasGroups(AliasAnalysis
&AA
);
2499 /// Build alias groups for all memory accesses in the Scop.
2501 /// Using the alias analysis and an alias set tracker we build alias sets
2502 /// for all memory accesses inside the Scop. For each alias set we then map
2503 /// the aliasing pointers back to the memory accesses we know, thus obtain
2504 /// groups of memory accesses which might alias. We also collect the set of
2505 /// arrays through which memory is written.
2507 /// @param AA A reference to the alias analysis.
2509 /// @returns A pair consistent of a vector of alias groups and a set of arrays
2510 /// through which memory is written.
2511 std::tuple
<AliasGroupVectorTy
, DenseSet
<const ScopArrayInfo
*>>
2512 buildAliasGroupsForAccesses(AliasAnalysis
&AA
);
2514 /// Split alias groups by iteration domains.
2516 /// We split each group based on the domains of the minimal/maximal accesses.
2517 /// That means two minimal/maximal accesses are only in a group if their
2518 /// access domains intersect. Otherwise, they are in different groups.
2520 /// @param AliasGroups The alias groups to split
2521 void splitAliasGroupsByDomain(AliasGroupVectorTy
&AliasGroups
);
2523 /// Build a given alias group and its access data.
2525 /// @param AliasGroup The alias group to build.
2526 /// @param HasWriteAccess A set of arrays through which memory is not only
2527 /// read, but also written.
2529 /// @returns True if __no__ error occurred, false otherwise.
2530 bool buildAliasGroup(Scop::AliasGroupTy
&AliasGroup
,
2531 DenseSet
<const ScopArrayInfo
*> HasWriteAccess
);
2533 /// Return all alias groups for this SCoP.
2534 const MinMaxVectorPairVectorTy
&getAliasGroups() const {
2535 return MinMaxAliasGroups
;
2538 /// Get an isl string representing the context.
2539 std::string
getContextStr() const;
2541 /// Get an isl string representing the assumed context.
2542 std::string
getAssumedContextStr() const;
2544 /// Get an isl string representing the invalid context.
2545 std::string
getInvalidContextStr() const;
2547 /// Return the ScopStmt for the given @p BB or nullptr if there is
2549 ScopStmt
*getStmtFor(BasicBlock
*BB
) const;
2551 /// Return the last statement representing @p BB.
2553 /// Of the sequence of statements that represent a @p BB, this is the last one
2554 /// to be executed. It is typically used to determine which instruction to add
2555 /// a MemoryKind::PHI WRITE to. For this purpose, it is not strictly required
2556 /// to be executed last, only that the incoming value is available in it.
2557 ScopStmt
*getLastStmtFor(BasicBlock
*BB
) const { return getStmtFor(BB
); }
2559 /// Return the ScopStmt that represents the Region @p R, or nullptr if
2560 /// it is not represented by any statement in this Scop.
2561 ScopStmt
*getStmtFor(Region
*R
) const;
2563 /// Return the ScopStmt that represents @p RN; can return nullptr if
2564 /// the RegionNode is not within the SCoP or has been removed due to
2565 /// simplifications.
2566 ScopStmt
*getStmtFor(RegionNode
*RN
) const;
2568 /// Return the ScopStmt an instruction belongs to, or nullptr if it
2569 /// does not belong to any statement in this Scop.
2570 ScopStmt
*getStmtFor(Instruction
*Inst
) const {
2571 return getStmtFor(Inst
->getParent());
2574 /// Return the number of statements in the SCoP.
2575 size_t getSize() const { return Stmts
.size(); }
2577 /// @name Statements Iterators
2579 /// These iterators iterate over all statements of this Scop.
2581 typedef StmtSet::iterator iterator
;
2582 typedef StmtSet::const_iterator const_iterator
;
2584 iterator
begin() { return Stmts
.begin(); }
2585 iterator
end() { return Stmts
.end(); }
2586 const_iterator
begin() const { return Stmts
.begin(); }
2587 const_iterator
end() const { return Stmts
.end(); }
2589 typedef StmtSet::reverse_iterator reverse_iterator
;
2590 typedef StmtSet::const_reverse_iterator const_reverse_iterator
;
2592 reverse_iterator
rbegin() { return Stmts
.rbegin(); }
2593 reverse_iterator
rend() { return Stmts
.rend(); }
2594 const_reverse_iterator
rbegin() const { return Stmts
.rbegin(); }
2595 const_reverse_iterator
rend() const { return Stmts
.rend(); }
2598 /// Return the set of required invariant loads.
2599 const InvariantLoadsSetTy
&getRequiredInvariantLoads() const {
2600 return DC
.RequiredILS
;
2603 /// Add @p LI to the set of required invariant loads.
2604 void addRequiredInvariantLoad(LoadInst
*LI
) { DC
.RequiredILS
.insert(LI
); }
2606 /// Return true if and only if @p LI is a required invariant load.
2607 bool isRequiredInvariantLoad(LoadInst
*LI
) const {
2608 return getRequiredInvariantLoads().count(LI
);
2611 /// Return the set of boxed (thus overapproximated) loops.
2612 const BoxedLoopsSetTy
&getBoxedLoops() const { return DC
.BoxedLoopsSet
; }
2614 /// Return true if and only if @p R is a non-affine subregion.
2615 bool isNonAffineSubRegion(const Region
*R
) {
2616 return DC
.NonAffineSubRegionSet
.count(R
);
2619 const MapInsnToMemAcc
&getInsnToMemAccMap() const { return DC
.InsnToMemAcc
; }
2621 /// Return the (possibly new) ScopArrayInfo object for @p Access.
2623 /// @param ElementType The type of the elements stored in this array.
2624 /// @param Kind The kind of the array info object.
2625 /// @param BaseName The optional name of this memory reference.
2626 ScopArrayInfo
*getOrCreateScopArrayInfo(Value
*BasePtr
, Type
*ElementType
,
2627 ArrayRef
<const SCEV
*> Sizes
,
2629 const char *BaseName
= nullptr);
2631 /// Create an array and return the corresponding ScopArrayInfo object.
2633 /// @param ElementType The type of the elements stored in this array.
2634 /// @param BaseName The name of this memory reference.
2635 /// @param Sizes The sizes of dimensions.
2636 ScopArrayInfo
*createScopArrayInfo(Type
*ElementType
,
2637 const std::string
&BaseName
,
2638 const std::vector
<unsigned> &Sizes
);
2640 /// Return the cached ScopArrayInfo object for @p BasePtr.
2642 /// @param BasePtr The base pointer the object has been stored for.
2643 /// @param Kind The kind of array info object.
2645 /// @returns The ScopArrayInfo pointer or NULL if no such pointer is
2647 const ScopArrayInfo
*getScopArrayInfoOrNull(Value
*BasePtr
, MemoryKind Kind
);
2649 /// Return the cached ScopArrayInfo object for @p BasePtr.
2651 /// @param BasePtr The base pointer the object has been stored for.
2652 /// @param Kind The kind of array info object.
2654 /// @returns The ScopArrayInfo pointer (may assert if no such pointer is
2656 const ScopArrayInfo
*getScopArrayInfo(Value
*BasePtr
, MemoryKind Kind
);
2658 /// Invalidate ScopArrayInfo object for base address.
2660 /// @param BasePtr The base pointer of the ScopArrayInfo object to invalidate.
2661 /// @param Kind The Kind of the ScopArrayInfo object.
2662 void invalidateScopArrayInfo(Value
*BasePtr
, MemoryKind Kind
) {
2663 auto It
= ScopArrayInfoMap
.find(std::make_pair(BasePtr
, Kind
));
2664 if (It
== ScopArrayInfoMap
.end())
2666 ScopArrayInfoSet
.remove(It
->second
.get());
2667 ScopArrayInfoMap
.erase(It
);
2670 void setContext(__isl_take isl_set
*NewContext
);
2672 /// Align the parameters in the statement to the scop context
2673 void realignParams();
2675 /// Return true if this SCoP can be profitably optimized.
2677 /// @param ScalarsAreUnprofitable Never consider statements with scalar writes
2678 /// as profitably optimizable.
2680 /// @return Whether this SCoP can be profitably optimized.
2681 bool isProfitable(bool ScalarsAreUnprofitable
) const;
2683 /// Return true if the SCoP contained at least one error block.
2684 bool hasErrorBlock() const { return HasErrorBlock
; }
2686 /// Return true if the underlying region has a single exiting block.
2687 bool hasSingleExitEdge() const { return HasSingleExitEdge
; }
2689 /// Print the static control part.
2691 /// @param OS The output stream the static control part is printed to.
2692 void print(raw_ostream
&OS
) const;
2694 /// Print the ScopStmt to stderr.
2697 /// Get the isl context of this static control part.
2699 /// @return The isl context of this static control part.
2700 isl_ctx
*getIslCtx() const;
2702 /// Directly return the shared_ptr of the context.
2703 const std::shared_ptr
<isl_ctx
> &getSharedIslCtx() const { return IslCtx
; }
2705 /// Compute the isl representation for the SCEV @p E
2707 /// @param E The SCEV that should be translated.
2708 /// @param BB An (optional) basic block in which the isl_pw_aff is computed.
2709 /// SCEVs known to not reference any loops in the SCoP can be
2710 /// passed without a @p BB.
2711 /// @param NonNegative Flag to indicate the @p E has to be non-negative.
2713 /// Note that this function will always return a valid isl_pw_aff. However, if
2714 /// the translation of @p E was deemed to complex the SCoP is invalidated and
2715 /// a dummy value of appropriate dimension is returned. This allows to bail
2716 /// for complex cases without "error handling code" needed on the users side.
2717 __isl_give PWACtx
getPwAff(const SCEV
*E
, BasicBlock
*BB
= nullptr,
2718 bool NonNegative
= false);
2720 /// Compute the isl representation for the SCEV @p E
2722 /// This function is like @see Scop::getPwAff() but strips away the invalid
2723 /// domain part associated with the piecewise affine function.
2724 __isl_give isl_pw_aff
*getPwAffOnly(const SCEV
*E
, BasicBlock
*BB
= nullptr);
2726 /// Return the domain of @p Stmt.
2728 /// @param Stmt The statement for which the conditions should be returned.
2729 __isl_give isl_set
*getDomainConditions(const ScopStmt
*Stmt
) const;
2731 /// Return the domain of @p BB.
2733 /// @param BB The block for which the conditions should be returned.
2734 __isl_give isl_set
*getDomainConditions(BasicBlock
*BB
) const;
2736 /// Get a union set containing the iteration domains of all statements.
2737 __isl_give isl_union_set
*getDomains() const;
2739 /// Get a union map of all may-writes performed in the SCoP.
2740 __isl_give isl_union_map
*getMayWrites();
2742 /// Get a union map of all must-writes performed in the SCoP.
2743 __isl_give isl_union_map
*getMustWrites();
2745 /// Get a union map of all writes performed in the SCoP.
2746 __isl_give isl_union_map
*getWrites();
2748 /// Get a union map of all reads performed in the SCoP.
2749 __isl_give isl_union_map
*getReads();
2751 /// Get a union map of all memory accesses performed in the SCoP.
2752 __isl_give isl_union_map
*getAccesses();
2754 /// Get the schedule of all the statements in the SCoP.
2756 /// @return The schedule of all the statements in the SCoP, if the schedule of
2757 /// the Scop does not contain extension nodes, and nullptr, otherwise.
2758 __isl_give isl_union_map
*getSchedule() const;
2760 /// Get a schedule tree describing the schedule of all statements.
2761 __isl_give isl_schedule
*getScheduleTree() const;
2763 /// Update the current schedule
2765 /// NewSchedule The new schedule (given as a flat union-map).
2766 void setSchedule(__isl_take isl_union_map
*NewSchedule
);
2768 /// Update the current schedule
2770 /// NewSchedule The new schedule (given as schedule tree).
2771 void setScheduleTree(__isl_take isl_schedule
*NewSchedule
);
2773 /// Intersects the domains of all statements in the SCoP.
2775 /// @return true if a change was made
2776 bool restrictDomains(__isl_take isl_union_set
*Domain
);
2778 /// Get the depth of a loop relative to the outermost loop in the Scop.
2780 /// This will return
2781 /// 0 if @p L is an outermost loop in the SCoP
2782 /// >0 for other loops in the SCoP
2783 /// -1 if @p L is nullptr or there is no outermost loop in the SCoP
2784 int getRelativeLoopDepth(const Loop
*L
) const;
2786 /// Find the ScopArrayInfo associated with an isl Id
2787 /// that has name @p Name.
2788 ScopArrayInfo
*getArrayInfoByName(const std::string BaseName
);
2790 /// Check whether @p Schedule contains extension nodes.
2792 /// @return true if @p Schedule contains extension nodes.
2793 static bool containsExtensionNode(__isl_keep isl_schedule
*Schedule
);
2795 /// Simplify the SCoP representation.
2797 /// @param AfterHoisting Whether it is called after invariant load hoisting.
2798 /// When true, also removes statements without
2800 void simplifySCoP(bool AfterHoisting
);
2802 /// Get the next free array index.
2804 /// This function returns a unique index which can be used to identify an
2806 long getNextArrayIdx() { return ArrayIdx
++; }
2808 /// Get the next free statement index.
2810 /// This function returns a unique index which can be used to identify a
2812 long getNextStmtIdx() { return StmtIdx
++; }
2815 /// Print Scop scop to raw_ostream O.
2816 static inline raw_ostream
&operator<<(raw_ostream
&O
, const Scop
&scop
) {
2821 /// The legacy pass manager's analysis pass to compute scop information
2823 class ScopInfoRegionPass
: public RegionPass
{
2824 /// The Scop pointer which is used to construct a Scop.
2825 std::unique_ptr
<Scop
> S
;
2828 static char ID
; // Pass identification, replacement for typeid
2830 ScopInfoRegionPass() : RegionPass(ID
) {}
2831 ~ScopInfoRegionPass() {}
2833 /// Build Scop object, the Polly IR of static control
2834 /// part for the current SESE-Region.
2836 /// @return If the current region is a valid for a static control part,
2837 /// return the Polly IR representing this static control part,
2838 /// return null otherwise.
2839 Scop
*getScop() { return S
.get(); }
2840 const Scop
*getScop() const { return S
.get(); }
2842 /// Calculate the polyhedral scop information for a given Region.
2843 bool runOnRegion(Region
*R
, RGPassManager
&RGM
) override
;
2845 void releaseMemory() override
{ S
.reset(); }
2847 void print(raw_ostream
&O
, const Module
*M
= nullptr) const override
;
2849 void getAnalysisUsage(AnalysisUsage
&AU
) const override
;
2854 using RegionToScopMapTy
= DenseMap
<Region
*, std::unique_ptr
<Scop
>>;
2855 using iterator
= RegionToScopMapTy::iterator
;
2856 using const_iterator
= RegionToScopMapTy::const_iterator
;
2859 /// A map of Region to its Scop object containing
2860 /// Polly IR of static control part.
2861 RegionToScopMapTy RegionToScopMap
;
2864 ScopInfo(const DataLayout
&DL
, ScopDetection
&SD
, ScalarEvolution
&SE
,
2865 LoopInfo
&LI
, AliasAnalysis
&AA
, DominatorTree
&DT
,
2866 AssumptionCache
&AC
);
2868 /// Get the Scop object for the given Region.
2870 /// @return If the given region is the maximal region within a scop, return
2871 /// the scop object. If the given region is a subregion, return a
2872 /// nullptr. Top level region containing the entry block of a function
2873 /// is not considered in the scop creation.
2874 Scop
*getScop(Region
*R
) const {
2875 auto MapIt
= RegionToScopMap
.find(R
);
2876 if (MapIt
!= RegionToScopMap
.end())
2877 return MapIt
->second
.get();
2881 iterator
begin() { return RegionToScopMap
.begin(); }
2882 iterator
end() { return RegionToScopMap
.end(); }
2883 const_iterator
begin() const { return RegionToScopMap
.begin(); }
2884 const_iterator
end() const { return RegionToScopMap
.end(); }
2885 bool empty() const { return RegionToScopMap
.empty(); }
2888 struct ScopInfoAnalysis
: public AnalysisInfoMixin
<ScopInfoAnalysis
> {
2889 static AnalysisKey Key
;
2890 using Result
= ScopInfo
;
2891 Result
run(Function
&, FunctionAnalysisManager
&);
2894 struct ScopInfoPrinterPass
: public PassInfoMixin
<ScopInfoPrinterPass
> {
2895 ScopInfoPrinterPass(raw_ostream
&O
) : Stream(O
) {}
2896 PreservedAnalyses
run(Function
&, FunctionAnalysisManager
&);
2897 raw_ostream
&Stream
;
2900 //===----------------------------------------------------------------------===//
2901 /// The legacy pass manager's analysis pass to compute scop information
2902 /// for the whole function.
2904 /// This pass will maintain a map of the maximal region within a scop to its
2905 /// scop object for all the feasible scops present in a function.
2906 /// This pass is an alternative to the ScopInfoRegionPass in order to avoid a
2907 /// region pass manager.
2908 class ScopInfoWrapperPass
: public FunctionPass
{
2909 std::unique_ptr
<ScopInfo
> Result
;
2912 ScopInfoWrapperPass() : FunctionPass(ID
) {}
2913 ~ScopInfoWrapperPass() = default;
2915 static char ID
; // Pass identification, replacement for typeid
2917 ScopInfo
*getSI() { return Result
.get(); }
2918 const ScopInfo
*getSI() const { return Result
.get(); }
2920 /// Calculate all the polyhedral scops for a given function.
2921 bool runOnFunction(Function
&F
) override
;
2923 void releaseMemory() override
{ Result
.reset(); }
2925 void print(raw_ostream
&O
, const Module
*M
= nullptr) const override
;
2927 void getAnalysisUsage(AnalysisUsage
&AU
) const override
;
2930 } // end namespace polly
2934 void initializeScopInfoRegionPassPass(llvm::PassRegistry
&);
2935 void initializeScopInfoWrapperPassPass(llvm::PassRegistry
&);