llvm-project/polly/lib/Transform/ScheduleTreeTransform.cpp

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

1273 lines
47 KiB
C++
Raw Normal View History

//===- polly/ScheduleTreeTransform.cpp --------------------------*- C++ -*-===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// Make changes to isl's schedule tree data structure.
//
//===----------------------------------------------------------------------===//
#include "polly/ScheduleTreeTransform.h"
#include "polly/Support/GICHelper.h"
#include "polly/Support/ISLTools.h"
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
#include "polly/Support/ScopHelper.h"
#include "llvm/ADT/ArrayRef.h"
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
#include "llvm/ADT/Sequence.h"
#include "llvm/ADT/SmallVector.h"
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
#include "llvm/IR/Constants.h"
#include "llvm/IR/Metadata.h"
#include "llvm/Transforms/Utils/UnrollLoop.h"
#define DEBUG_TYPE "polly-opt-isl"
using namespace polly;
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
using namespace llvm;
namespace {
/// Copy the band member attributes (coincidence, loop type, isolate ast loop
/// type) from one band to another.
static isl::schedule_node_band
applyBandMemberAttributes(isl::schedule_node_band Target, int TargetIdx,
const isl::schedule_node_band &Source,
int SourceIdx) {
bool Coincident = Source.member_get_coincident(SourceIdx).release();
Target = Target.member_set_coincident(TargetIdx, Coincident);
isl_ast_loop_type LoopType =
isl_schedule_node_band_member_get_ast_loop_type(Source.get(), SourceIdx);
Target = isl::manage(isl_schedule_node_band_member_set_ast_loop_type(
Target.release(), TargetIdx, LoopType))
.as<isl::schedule_node_band>();
isl_ast_loop_type IsolateType =
isl_schedule_node_band_member_get_isolate_ast_loop_type(Source.get(),
SourceIdx);
Target = isl::manage(isl_schedule_node_band_member_set_isolate_ast_loop_type(
Target.release(), TargetIdx, IsolateType))
.as<isl::schedule_node_band>();
return Target;
}
/// Create a new band by copying members from another @p Band. @p IncludeCb
/// decides which band indices are copied to the result.
template <typename CbTy>
static isl::schedule rebuildBand(isl::schedule_node_band OldBand,
isl::schedule Body, CbTy IncludeCb) {
int NumBandDims = OldBand.n_member().release();
bool ExcludeAny = false;
bool IncludeAny = false;
for (auto OldIdx : seq<int>(0, NumBandDims)) {
if (IncludeCb(OldIdx))
IncludeAny = true;
else
ExcludeAny = true;
}
// Instead of creating a zero-member band, don't create a band at all.
if (!IncludeAny)
return Body;
isl::multi_union_pw_aff PartialSched = OldBand.get_partial_schedule();
isl::multi_union_pw_aff NewPartialSched;
if (ExcludeAny) {
// Select the included partial scatter functions.
isl::union_pw_aff_list List = PartialSched.list();
int NewIdx = 0;
for (auto OldIdx : seq<int>(0, NumBandDims)) {
if (IncludeCb(OldIdx))
NewIdx += 1;
else
List = List.drop(NewIdx, 1);
}
isl::space ParamSpace = PartialSched.get_space().params();
isl::space NewScatterSpace = ParamSpace.add_unnamed_tuple(NewIdx);
NewPartialSched = isl::multi_union_pw_aff(NewScatterSpace, List);
} else {
// Just reuse original scatter function of copying all of them.
NewPartialSched = PartialSched;
}
// Create the new band node.
isl::schedule_node_band NewBand =
Body.insert_partial_schedule(NewPartialSched)
.get_root()
.child(0)
.as<isl::schedule_node_band>();
// If OldBand was permutable, so is the new one, even if some dimensions are
// missing.
bool IsPermutable = OldBand.permutable().release();
NewBand = NewBand.set_permutable(IsPermutable);
// Reapply member attributes.
int NewIdx = 0;
for (auto OldIdx : seq<int>(0, NumBandDims)) {
if (!IncludeCb(OldIdx))
continue;
NewBand =
applyBandMemberAttributes(std::move(NewBand), NewIdx, OldBand, OldIdx);
NewIdx += 1;
}
return NewBand.get_schedule();
}
/// Recursively visit all nodes of a schedule tree while allowing changes.
///
/// The visit methods return an isl::schedule_node that is used to continue
/// visiting the tree. Structural changes such as returning a different node
/// will confuse the visitor.
template <typename Derived, typename... Args>
struct ScheduleNodeRewriter
: public RecursiveScheduleTreeVisitor<Derived, isl::schedule_node,
Args...> {
Derived &getDerived() { return *static_cast<Derived *>(this); }
const Derived &getDerived() const {
return *static_cast<const Derived *>(this);
}
isl::schedule_node visitNode(const isl::schedule_node &Node, Args... args) {
if (!Node.has_children())
return Node;
isl::schedule_node It = Node.first_child();
while (true) {
It = getDerived().visit(It, std::forward<Args>(args)...);
if (!It.has_next_sibling())
break;
It = It.next_sibling();
}
return It.parent();
}
};
/// Rewrite a schedule tree by reconstructing it bottom-up.
///
/// By default, the original schedule tree is reconstructed. To build a
/// different tree, redefine visitor methods in a derived class (CRTP).
///
/// Note that AST build options are not applied; Setting the isolate[] option
/// makes the schedule tree 'anchored' and cannot be modified afterwards. Hence,
/// AST build options must be set after the tree has been constructed.
template <typename Derived, typename... Args>
struct ScheduleTreeRewriter
: public RecursiveScheduleTreeVisitor<Derived, isl::schedule, Args...> {
Derived &getDerived() { return *static_cast<Derived *>(this); }
const Derived &getDerived() const {
return *static_cast<const Derived *>(this);
}
isl::schedule visitDomain(isl::schedule_node_domain Node, Args... args) {
// Every schedule_tree already has a domain node, no need to add one.
return getDerived().visit(Node.first_child(), std::forward<Args>(args)...);
}
isl::schedule visitBand(isl::schedule_node_band Band, Args... args) {
isl::schedule NewChild =
getDerived().visit(Band.child(0), std::forward<Args>(args)...);
return rebuildBand(Band, NewChild, [](int) { return true; });
}
isl::schedule visitSequence(isl::schedule_node_sequence Sequence,
Args... args) {
int NumChildren = isl_schedule_node_n_children(Sequence.get());
isl::schedule Result =
getDerived().visit(Sequence.child(0), std::forward<Args>(args)...);
for (int i = 1; i < NumChildren; i += 1)
Result = Result.sequence(
getDerived().visit(Sequence.child(i), std::forward<Args>(args)...));
return Result;
}
isl::schedule visitSet(isl::schedule_node_set Set, Args... args) {
int NumChildren = isl_schedule_node_n_children(Set.get());
isl::schedule Result =
getDerived().visit(Set.child(0), std::forward<Args>(args)...);
for (int i = 1; i < NumChildren; i += 1)
Result = isl::manage(
isl_schedule_set(Result.release(),
getDerived()
.visit(Set.child(i), std::forward<Args>(args)...)
.release()));
return Result;
}
isl::schedule visitLeaf(isl::schedule_node_leaf Leaf, Args... args) {
return isl::schedule::from_domain(Leaf.get_domain());
}
isl::schedule visitMark(const isl::schedule_node &Mark, Args... args) {
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
isl::id TheMark = Mark.as<isl::schedule_node_mark>().get_id();
isl::schedule_node NewChild =
getDerived()
.visit(Mark.first_child(), std::forward<Args>(args)...)
.get_root()
.first_child();
return NewChild.insert_mark(TheMark).get_schedule();
}
isl::schedule visitExtension(isl::schedule_node_extension Extension,
Args... args) {
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
isl::union_map TheExtension =
Extension.as<isl::schedule_node_extension>().get_extension();
isl::schedule_node NewChild = getDerived()
.visit(Extension.child(0), args...)
.get_root()
.first_child();
isl::schedule_node NewExtension =
isl::schedule_node::from_extension(TheExtension);
return NewChild.graft_before(NewExtension).get_schedule();
}
isl::schedule visitFilter(isl::schedule_node_filter Filter, Args... args) {
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
isl::union_set FilterDomain =
Filter.as<isl::schedule_node_filter>().get_filter();
isl::schedule NewSchedule =
getDerived().visit(Filter.child(0), std::forward<Args>(args)...);
return NewSchedule.intersect_domain(FilterDomain);
}
isl::schedule visitNode(isl::schedule_node Node, Args... args) {
llvm_unreachable("Not implemented");
}
};
/// Rewrite the schedule tree without any changes. Useful to copy a subtree into
/// a new schedule, discarding everything but.
struct IdentityRewriter : public ScheduleTreeRewriter<IdentityRewriter> {};
/// Rewrite a schedule tree to an equivalent one without extension nodes.
///
/// Each visit method takes two additional arguments:
///
/// * The new domain the node, which is the inherited domain plus any domains
/// added by extension nodes.
///
/// * A map of extension domains of all children is returned; it is required by
/// band nodes to schedule the additional domains at the same position as the
/// extension node would.
///
struct ExtensionNodeRewriter
: public ScheduleTreeRewriter<ExtensionNodeRewriter, const isl::union_set &,
isl::union_map &> {
using BaseTy = ScheduleTreeRewriter<ExtensionNodeRewriter,
const isl::union_set &, isl::union_map &>;
BaseTy &getBase() { return *this; }
const BaseTy &getBase() const { return *this; }
isl::schedule visitSchedule(isl::schedule Schedule) {
isl::union_map Extensions;
isl::schedule Result =
visit(Schedule.get_root(), Schedule.get_domain(), Extensions);
assert(!Extensions.is_null() && Extensions.is_empty());
return Result;
}
isl::schedule visitSequence(isl::schedule_node_sequence Sequence,
const isl::union_set &Domain,
isl::union_map &Extensions) {
int NumChildren = isl_schedule_node_n_children(Sequence.get());
isl::schedule NewNode = visit(Sequence.first_child(), Domain, Extensions);
for (int i = 1; i < NumChildren; i += 1) {
isl::schedule_node OldChild = Sequence.child(i);
isl::union_map NewChildExtensions;
isl::schedule NewChildNode = visit(OldChild, Domain, NewChildExtensions);
NewNode = NewNode.sequence(NewChildNode);
Extensions = Extensions.unite(NewChildExtensions);
}
return NewNode;
}
isl::schedule visitSet(isl::schedule_node_set Set,
const isl::union_set &Domain,
isl::union_map &Extensions) {
int NumChildren = isl_schedule_node_n_children(Set.get());
isl::schedule NewNode = visit(Set.first_child(), Domain, Extensions);
for (int i = 1; i < NumChildren; i += 1) {
isl::schedule_node OldChild = Set.child(i);
isl::union_map NewChildExtensions;
isl::schedule NewChildNode = visit(OldChild, Domain, NewChildExtensions);
NewNode = isl::manage(
isl_schedule_set(NewNode.release(), NewChildNode.release()));
Extensions = Extensions.unite(NewChildExtensions);
}
return NewNode;
}
isl::schedule visitLeaf(isl::schedule_node_leaf Leaf,
const isl::union_set &Domain,
isl::union_map &Extensions) {
Extensions = isl::union_map::empty(Leaf.ctx());
return isl::schedule::from_domain(Domain);
}
isl::schedule visitBand(isl::schedule_node_band OldNode,
const isl::union_set &Domain,
isl::union_map &OuterExtensions) {
isl::schedule_node OldChild = OldNode.first_child();
isl::multi_union_pw_aff PartialSched =
isl::manage(isl_schedule_node_band_get_partial_schedule(OldNode.get()));
isl::union_map NewChildExtensions;
isl::schedule NewChild = visit(OldChild, Domain, NewChildExtensions);
// Add the extensions to the partial schedule.
OuterExtensions = isl::union_map::empty(NewChildExtensions.ctx());
isl::union_map NewPartialSchedMap = isl::union_map::from(PartialSched);
unsigned BandDims = isl_schedule_node_band_n_member(OldNode.get());
for (isl::map Ext : NewChildExtensions.get_map_list()) {
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
unsigned ExtDims = Ext.domain_tuple_dim().release();
assert(ExtDims >= BandDims);
unsigned OuterDims = ExtDims - BandDims;
isl::map BandSched =
Ext.project_out(isl::dim::in, 0, OuterDims).reverse();
NewPartialSchedMap = NewPartialSchedMap.unite(BandSched);
// There might be more outer bands that have to schedule the extensions.
if (OuterDims > 0) {
isl::map OuterSched =
Ext.project_out(isl::dim::in, OuterDims, BandDims);
OuterExtensions = OuterExtensions.unite(OuterSched);
}
}
isl::multi_union_pw_aff NewPartialSchedAsAsMultiUnionPwAff =
isl::multi_union_pw_aff::from_union_map(NewPartialSchedMap);
isl::schedule_node NewNode =
NewChild.insert_partial_schedule(NewPartialSchedAsAsMultiUnionPwAff)
.get_root()
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
.child(0);
// Reapply permutability and coincidence attributes.
NewNode = isl::manage(isl_schedule_node_band_set_permutable(
NewNode.release(),
isl_schedule_node_band_get_permutable(OldNode.get())));
for (unsigned i = 0; i < BandDims; i += 1)
NewNode = applyBandMemberAttributes(NewNode.as<isl::schedule_node_band>(),
i, OldNode, i);
return NewNode.get_schedule();
}
isl::schedule visitFilter(isl::schedule_node_filter Filter,
const isl::union_set &Domain,
isl::union_map &Extensions) {
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
isl::union_set FilterDomain =
Filter.as<isl::schedule_node_filter>().get_filter();
isl::union_set NewDomain = Domain.intersect(FilterDomain);
// A filter is added implicitly if necessary when joining schedule trees.
return visit(Filter.first_child(), NewDomain, Extensions);
}
isl::schedule visitExtension(isl::schedule_node_extension Extension,
const isl::union_set &Domain,
isl::union_map &Extensions) {
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
isl::union_map ExtDomain =
Extension.as<isl::schedule_node_extension>().get_extension();
isl::union_set NewDomain = Domain.unite(ExtDomain.range());
isl::union_map ChildExtensions;
isl::schedule NewChild =
visit(Extension.first_child(), NewDomain, ChildExtensions);
Extensions = ChildExtensions.unite(ExtDomain);
return NewChild;
}
};
/// Collect all AST build options in any schedule tree band.
///
/// ScheduleTreeRewriter cannot apply the schedule tree options. This class
/// collects these options to apply them later.
struct CollectASTBuildOptions
: public RecursiveScheduleTreeVisitor<CollectASTBuildOptions> {
using BaseTy = RecursiveScheduleTreeVisitor<CollectASTBuildOptions>;
BaseTy &getBase() { return *this; }
const BaseTy &getBase() const { return *this; }
llvm::SmallVector<isl::union_set, 8> ASTBuildOptions;
void visitBand(isl::schedule_node_band Band) {
ASTBuildOptions.push_back(
isl::manage(isl_schedule_node_band_get_ast_build_options(Band.get())));
return getBase().visitBand(Band);
}
};
/// Apply AST build options to the bands in a schedule tree.
///
/// This rewrites a schedule tree with the AST build options applied. We assume
/// that the band nodes are visited in the same order as they were when the
/// build options were collected, typically by CollectASTBuildOptions.
struct ApplyASTBuildOptions
: public ScheduleNodeRewriter<ApplyASTBuildOptions> {
using BaseTy = ScheduleNodeRewriter<ApplyASTBuildOptions>;
BaseTy &getBase() { return *this; }
const BaseTy &getBase() const { return *this; }
size_t Pos;
llvm::ArrayRef<isl::union_set> ASTBuildOptions;
ApplyASTBuildOptions(llvm::ArrayRef<isl::union_set> ASTBuildOptions)
: ASTBuildOptions(ASTBuildOptions) {}
isl::schedule visitSchedule(isl::schedule Schedule) {
Pos = 0;
isl::schedule Result = visit(Schedule).get_schedule();
assert(Pos == ASTBuildOptions.size() &&
"AST build options must match to band nodes");
return Result;
}
isl::schedule_node visitBand(isl::schedule_node_band Band) {
isl::schedule_node_band Result =
Band.set_ast_build_options(ASTBuildOptions[Pos]);
Pos += 1;
return getBase().visitBand(Result);
}
};
/// Return whether the schedule contains an extension node.
static bool containsExtensionNode(isl::schedule Schedule) {
assert(!Schedule.is_null());
auto Callback = [](__isl_keep isl_schedule_node *Node,
void *User) -> isl_bool {
if (isl_schedule_node_get_type(Node) == isl_schedule_node_extension) {
// Stop walking the schedule tree.
return isl_bool_error;
}
// Continue searching the subtree.
return isl_bool_true;
};
isl_stat RetVal = isl_schedule_foreach_schedule_node_top_down(
Schedule.get(), Callback, nullptr);
// We assume that the traversal itself does not fail, i.e. the only reason to
// return isl_stat_error is that an extension node was found.
return RetVal == isl_stat_error;
}
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
/// Find a named MDNode property in a LoopID.
static MDNode *findOptionalNodeOperand(MDNode *LoopMD, StringRef Name) {
return dyn_cast_or_null<MDNode>(
findMetadataOperand(LoopMD, Name).getValueOr(nullptr));
}
/// Is this node of type mark?
static bool isMark(const isl::schedule_node &Node) {
return isl_schedule_node_get_type(Node.get()) == isl_schedule_node_mark;
}
/// Is this node of type band?
static bool isBand(const isl::schedule_node &Node) {
return isl_schedule_node_get_type(Node.get()) == isl_schedule_node_band;
}
#ifndef NDEBUG
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
/// Is this node a band of a single dimension (i.e. could represent a loop)?
static bool isBandWithSingleLoop(const isl::schedule_node &Node) {
return isBand(Node) && isl_schedule_node_band_n_member(Node.get()) == 1;
}
#endif
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
static bool isLeaf(const isl::schedule_node &Node) {
return isl_schedule_node_get_type(Node.get()) == isl_schedule_node_leaf;
}
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
/// Create an isl::id representing the output loop after a transformation.
static isl::id createGeneratedLoopAttr(isl::ctx Ctx, MDNode *FollowupLoopMD) {
// Don't need to id the followup.
// TODO: Append llvm.loop.disable_heustistics metadata unless overridden by
// user followup-MD
if (!FollowupLoopMD)
return {};
BandAttr *Attr = new BandAttr();
Attr->Metadata = FollowupLoopMD;
return getIslLoopAttr(Ctx, Attr);
}
/// A loop consists of a band and an optional marker that wraps it. Return the
/// outermost of the two.
/// That is, either the mark or, if there is not mark, the loop itself. Can
/// start with either the mark or the band.
static isl::schedule_node moveToBandMark(isl::schedule_node BandOrMark) {
if (isBandMark(BandOrMark)) {
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
assert(isBandWithSingleLoop(BandOrMark.child(0)));
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
return BandOrMark;
}
assert(isBandWithSingleLoop(BandOrMark));
isl::schedule_node Mark = BandOrMark.parent();
if (isBandMark(Mark))
return Mark;
// Band has no loop marker.
return BandOrMark;
}
static isl::schedule_node removeMark(isl::schedule_node MarkOrBand,
BandAttr *&Attr) {
MarkOrBand = moveToBandMark(MarkOrBand);
isl::schedule_node Band;
if (isMark(MarkOrBand)) {
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
Attr = getLoopAttr(MarkOrBand.as<isl::schedule_node_mark>().get_id());
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
Band = isl::manage(isl_schedule_node_delete(MarkOrBand.release()));
} else {
Attr = nullptr;
Band = MarkOrBand;
}
assert(isBandWithSingleLoop(Band));
return Band;
}
/// Remove the mark that wraps a loop. Return the band representing the loop.
static isl::schedule_node removeMark(isl::schedule_node MarkOrBand) {
BandAttr *Attr;
return removeMark(MarkOrBand, Attr);
}
static isl::schedule_node insertMark(isl::schedule_node Band, isl::id Mark) {
assert(isBand(Band));
assert(moveToBandMark(Band).is_equal(Band) &&
"Don't add a two marks for a band");
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
return Band.insert_mark(Mark).child(0);
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
}
/// Return the (one-dimensional) set of numbers that are divisible by @p Factor
/// with remainder @p Offset.
///
/// isDivisibleBySet(Ctx, 4, 0) = { [i] : floord(i,4) = 0 }
/// isDivisibleBySet(Ctx, 4, 1) = { [i] : floord(i,4) = 1 }
///
static isl::basic_set isDivisibleBySet(isl::ctx &Ctx, long Factor,
long Offset) {
isl::val ValFactor{Ctx, Factor};
isl::val ValOffset{Ctx, Offset};
isl::space Unispace{Ctx, 0, 1};
isl::local_space LUnispace{Unispace};
isl::aff AffFactor{LUnispace, ValFactor};
isl::aff AffOffset{LUnispace, ValOffset};
isl::aff Id = isl::aff::var_on_domain(LUnispace, isl::dim::out, 0);
isl::aff DivMul = Id.mod(ValFactor);
isl::basic_map Divisible = isl::basic_map::from_aff(DivMul);
isl::basic_map Modulo = Divisible.fix_val(isl::dim::out, 0, ValOffset);
return Modulo.domain();
}
/// Make the last dimension of Set to take values from 0 to VectorWidth - 1.
///
/// @param Set A set, which should be modified.
/// @param VectorWidth A parameter, which determines the constraint.
static isl::set addExtentConstraints(isl::set Set, int VectorWidth) {
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
unsigned Dims = Set.tuple_dim().release();
isl::space Space = Set.get_space();
isl::local_space LocalSpace = isl::local_space(Space);
isl::constraint ExtConstr = isl::constraint::alloc_inequality(LocalSpace);
ExtConstr = ExtConstr.set_constant_si(0);
ExtConstr = ExtConstr.set_coefficient_si(isl::dim::set, Dims - 1, 1);
Set = Set.add_constraint(ExtConstr);
ExtConstr = isl::constraint::alloc_inequality(LocalSpace);
ExtConstr = ExtConstr.set_constant_si(VectorWidth - 1);
ExtConstr = ExtConstr.set_coefficient_si(isl::dim::set, Dims - 1, -1);
return Set.add_constraint(ExtConstr);
}
/// Collapse perfectly nested bands into a single band.
class BandCollapseRewriter : public ScheduleTreeRewriter<BandCollapseRewriter> {
private:
using BaseTy = ScheduleTreeRewriter<BandCollapseRewriter>;
BaseTy &getBase() { return *this; }
const BaseTy &getBase() const { return *this; }
public:
isl::schedule visitBand(isl::schedule_node_band RootBand) {
isl::schedule_node_band Band = RootBand;
isl::ctx Ctx = Band.ctx();
// Do not merge permutable band to avoid loosing the permutability property.
// Cannot collapse even two permutable loops, they might be permutable
// individually, but not necassarily accross.
if (Band.n_member().release() > 1 && Band.permutable())
return getBase().visitBand(Band);
// Find collapsable bands.
SmallVector<isl::schedule_node_band> Nest;
int NumTotalLoops = 0;
isl::schedule_node Body;
while (true) {
Nest.push_back(Band);
NumTotalLoops += Band.n_member().release();
Body = Band.first_child();
if (!Body.isa<isl::schedule_node_band>())
break;
Band = Body.as<isl::schedule_node_band>();
// Do not include next band if it is permutable to not lose its
// permutability property.
if (Band.n_member().release() > 1 && Band.permutable())
break;
}
// Nothing to collapse, preserve permutability.
if (Nest.size() <= 1)
return getBase().visitBand(Band);
LLVM_DEBUG({
dbgs() << "Found loops to collapse between\n";
dumpIslObj(RootBand, dbgs());
dbgs() << "and\n";
dumpIslObj(Body, dbgs());
dbgs() << "\n";
});
isl::schedule NewBody = visit(Body);
// Collect partial schedules from all members.
isl::union_pw_aff_list PartScheds{Ctx, NumTotalLoops};
for (isl::schedule_node_band Band : Nest) {
int NumLoops = Band.n_member().release();
isl::multi_union_pw_aff BandScheds = Band.get_partial_schedule();
for (auto j : seq<int>(0, NumLoops))
PartScheds = PartScheds.add(BandScheds.at(j));
}
isl::space ScatterSpace = isl::space(Ctx, 0, NumTotalLoops);
isl::multi_union_pw_aff PartSchedsMulti{ScatterSpace, PartScheds};
isl::schedule_node_band CollapsedBand =
NewBody.insert_partial_schedule(PartSchedsMulti)
.get_root()
.first_child()
.as<isl::schedule_node_band>();
// Copy over loop attributes form original bands.
int LoopIdx = 0;
for (isl::schedule_node_band Band : Nest) {
int NumLoops = Band.n_member().release();
for (int i : seq<int>(0, NumLoops)) {
CollapsedBand = applyBandMemberAttributes(std::move(CollapsedBand),
LoopIdx, Band, i);
LoopIdx += 1;
}
}
assert(LoopIdx == NumTotalLoops &&
"Expect the same number of loops to add up again");
return CollapsedBand.get_schedule();
}
};
static isl::schedule collapseBands(isl::schedule Sched) {
LLVM_DEBUG(dbgs() << "Collapse bands in schedule\n");
BandCollapseRewriter Rewriter;
return Rewriter.visit(Sched);
}
/// Collect sequentially executed bands (or anything else), even if nested in a
/// mark or other nodes whose child is executed just once. If we can
/// successfully fuse the bands, we allow them to be removed.
static void collectPotentiallyFusableBands(
isl::schedule_node Node,
SmallVectorImpl<std::pair<isl::schedule_node, isl::schedule_node>>
&ScheduleBands,
const isl::schedule_node &DirectChild) {
switch (isl_schedule_node_get_type(Node.get())) {
case isl_schedule_node_sequence:
case isl_schedule_node_set:
case isl_schedule_node_mark:
case isl_schedule_node_domain:
case isl_schedule_node_filter:
if (Node.has_children()) {
isl::schedule_node C = Node.first_child();
while (true) {
collectPotentiallyFusableBands(C, ScheduleBands, DirectChild);
if (!C.has_next_sibling())
break;
C = C.next_sibling();
}
}
break;
default:
// Something that does not execute suquentially (e.g. a band)
ScheduleBands.push_back({Node, DirectChild});
break;
}
}
/// Remove dependencies that are resolved by @p PartSched. That is, remove
/// everything that we already know is executed in-order.
static isl::union_map remainingDepsFromPartialSchedule(isl::union_map PartSched,
isl::union_map Deps) {
int NumDims = getNumScatterDims(PartSched);
auto ParamSpace = PartSched.get_space().params();
// { Scatter[] }
isl::space ScatterSpace =
ParamSpace.set_from_params().add_dims(isl::dim::set, NumDims);
// { Scatter[] -> Domain[] }
isl::union_map PartSchedRev = PartSched.reverse();
// { Scatter[] -> Scatter[] }
isl::map MaybeBefore = isl::map::lex_le(ScatterSpace);
// { Domain[] -> Domain[] }
isl::union_map DomMaybeBefore =
MaybeBefore.apply_domain(PartSchedRev).apply_range(PartSchedRev);
// { Domain[] -> Domain[] }
isl::union_map ChildRemainingDeps = Deps.intersect(DomMaybeBefore);
return ChildRemainingDeps;
}
/// Remove dependencies that are resolved by executing them in the order
/// specified by @p Domains;
static isl::union_map remainigDepsFromSequence(ArrayRef<isl::union_set> Domains,
isl::union_map Deps) {
isl::ctx Ctx = Deps.ctx();
isl::space ParamSpace = Deps.get_space().params();
// Create a partial schedule mapping to constants that reflect the execution
// order.
isl::union_map PartialSchedules = isl::union_map::empty(Ctx);
for (auto P : enumerate(Domains)) {
isl::val ExecTime = isl::val(Ctx, P.index());
isl::union_pw_aff DomSched{P.value(), ExecTime};
PartialSchedules = PartialSchedules.unite(DomSched.as_union_map());
}
return remainingDepsFromPartialSchedule(PartialSchedules, Deps);
}
/// Determine whether the outermost loop of to bands can be fused while
/// respecting validity dependencies.
static bool canFuseOutermost(const isl::schedule_node_band &LHS,
const isl::schedule_node_band &RHS,
const isl::union_map &Deps) {
// { LHSDomain[] -> Scatter[] }
isl::union_map LHSPartSched =
LHS.get_partial_schedule().get_at(0).as_union_map();
// { Domain[] -> Scatter[] }
isl::union_map RHSPartSched =
RHS.get_partial_schedule().get_at(0).as_union_map();
// Dependencies that are already resolved because LHS executes before RHS, but
// will not be anymore after fusion. { DefDomain[] -> UseDomain[] }
isl::union_map OrderedBySequence =
Deps.intersect_domain(LHSPartSched.domain())
.intersect_range(RHSPartSched.domain());
isl::space ParamSpace = OrderedBySequence.get_space().params();
isl::space NewScatterSpace = ParamSpace.add_unnamed_tuple(1);
// { Scatter[] -> Scatter[] }
isl::map After = isl::map::lex_gt(NewScatterSpace);
// After fusion, instances with smaller (or equal, which means they will be
// executed in the same iteration, but the LHS instance is still sequenced
// before RHS) scatter value will still be executed before. This are the
// orderings where this is not necessarily the case.
// { LHSDomain[] -> RHSDomain[] }
isl::union_map MightBeAfterDoms = After.apply_domain(LHSPartSched.reverse())
.apply_range(RHSPartSched.reverse());
// Dependencies that are not resolved by the new execution order.
isl::union_map WithBefore = OrderedBySequence.intersect(MightBeAfterDoms);
return WithBefore.is_empty();
}
/// Fuse @p LHS and @p RHS if possible while preserving validity dependenvies.
static isl::schedule tryGreedyFuse(isl::schedule_node_band LHS,
isl::schedule_node_band RHS,
const isl::union_map &Deps) {
if (!canFuseOutermost(LHS, RHS, Deps))
return {};
LLVM_DEBUG({
dbgs() << "Found loops for greedy fusion:\n";
dumpIslObj(LHS, dbgs());
dbgs() << "and\n";
dumpIslObj(RHS, dbgs());
dbgs() << "\n";
});
// The partial schedule of the bands outermost loop that we need to combine
// for the fusion.
isl::union_pw_aff LHSPartOuterSched = LHS.get_partial_schedule().get_at(0);
isl::union_pw_aff RHSPartOuterSched = RHS.get_partial_schedule().get_at(0);
// Isolate band bodies as roots of their own schedule trees.
IdentityRewriter Rewriter;
isl::schedule LHSBody = Rewriter.visit(LHS.first_child());
isl::schedule RHSBody = Rewriter.visit(RHS.first_child());
// Reconstruct the non-outermost (not going to be fused) loops from both
// bands.
// TODO: Maybe it is possibly to transfer the 'permutability' property from
// LHS+RHS. At minimum we need merge multiple band members at once, otherwise
// permutability has no meaning.
isl::schedule LHSNewBody =
rebuildBand(LHS, LHSBody, [](int i) { return i > 0; });
isl::schedule RHSNewBody =
rebuildBand(RHS, RHSBody, [](int i) { return i > 0; });
// The loop body of the fused loop.
isl::schedule NewCommonBody = LHSNewBody.sequence(RHSNewBody);
// Combine the partial schedules of both loops to a new one. Instances with
// the same scatter value are put together.
isl::union_map NewCommonPartialSched =
LHSPartOuterSched.as_union_map().unite(RHSPartOuterSched.as_union_map());
isl::schedule NewCommonSchedule = NewCommonBody.insert_partial_schedule(
NewCommonPartialSched.as_multi_union_pw_aff());
return NewCommonSchedule;
}
static isl::schedule tryGreedyFuse(isl::schedule_node LHS,
isl::schedule_node RHS,
const isl::union_map &Deps) {
// TODO: Non-bands could be interpreted as a band with just as single
// iteration. However, this is only useful if both ends of a fused loop were
// originally loops themselves.
if (!LHS.isa<isl::schedule_node_band>())
return {};
if (!RHS.isa<isl::schedule_node_band>())
return {};
return tryGreedyFuse(LHS.as<isl::schedule_node_band>(),
RHS.as<isl::schedule_node_band>(), Deps);
}
/// Fuse all fusable loop top-down in a schedule tree.
///
/// The isl::union_map parameters is the set of validity dependencies that have
/// not been resolved/carried by a parent schedule node.
class GreedyFusionRewriter
: public ScheduleTreeRewriter<GreedyFusionRewriter, isl::union_map> {
private:
using BaseTy = ScheduleTreeRewriter<GreedyFusionRewriter, isl::union_map>;
BaseTy &getBase() { return *this; }
const BaseTy &getBase() const { return *this; }
public:
/// Is set to true if anything has been fused.
bool AnyChange = false;
isl::schedule visitBand(isl::schedule_node_band Band, isl::union_map Deps) {
int NumLoops = Band.n_member().release();
// { Domain[] -> Scatter[] }
isl::union_map PartSched =
isl::union_map::from(Band.get_partial_schedule());
assert(getNumScatterDims(PartSched) == NumLoops);
isl::space ParamSpace = PartSched.get_space().params();
// { Scatter[] -> Domain[] }
isl::union_map PartSchedRev = PartSched.reverse();
// Possible within the same iteration. Dependencies with smaller scatter
// value are carried by this loop and therefore have been resolved by the
// in-order execution if the loop iteration. A dependency with small scatter
// value would be a dependency violation that we assume did not happen. {
// Domain[] -> Domain[] }
isl::union_map Unsequenced = PartSchedRev.apply_domain(PartSchedRev);
// Actual dependencies within the same iteration.
// { DefDomain[] -> UseDomain[] }
isl::union_map RemDeps = Deps.intersect(Unsequenced);
return getBase().visitBand(Band, RemDeps);
}
isl::schedule visitSequence(isl::schedule_node_sequence Sequence,
isl::union_map Deps) {
int NumChildren = isl_schedule_node_n_children(Sequence.get());
// List of fusion candidates. The first element is the fusion candidate, the
// second is candidate's ancestor that is the sequence's direct child. It is
// preferable to use the direct child if not if its non-direct children is
// fused to preserve its structure such as mark nodes.
SmallVector<std::pair<isl::schedule_node, isl::schedule_node>> Bands;
for (auto i : seq<int>(0, NumChildren)) {
isl::schedule_node Child = Sequence.child(i);
collectPotentiallyFusableBands(Child, Bands, Child);
}
// Direct children that had at least one of its decendants fused.
SmallDenseSet<isl_schedule_node *, 4> ChangedDirectChildren;
// Fuse neigboring bands until reaching the end of candidates.
int i = 0;
while (i + 1 < (int)Bands.size()) {
isl::schedule Fused =
tryGreedyFuse(Bands[i].first, Bands[i + 1].first, Deps);
if (Fused.is_null()) {
// Cannot merge this node with the next; look at next pair.
i += 1;
continue;
}
// Mark the direct children as (partially) fused.
if (!Bands[i].second.is_null())
ChangedDirectChildren.insert(Bands[i].second.get());
if (!Bands[i + 1].second.is_null())
ChangedDirectChildren.insert(Bands[i + 1].second.get());
// Collapse the neigbros to a single new candidate that could be fused
// with the next candidate.
Bands[i] = {Fused.get_root(), {}};
Bands.erase(Bands.begin() + i + 1);
AnyChange = true;
}
// By construction equal if done with collectPotentiallyFusableBands's
// output.
SmallVector<isl::union_set> SubDomains;
SubDomains.reserve(NumChildren);
for (int i = 0; i < NumChildren; i += 1)
SubDomains.push_back(Sequence.child(i).domain());
auto SubRemainingDeps = remainigDepsFromSequence(SubDomains, Deps);
// We may iterate over direct children multiple times, be sure to add each
// at most once.
SmallDenseSet<isl_schedule_node *, 4> AlreadyAdded;
isl::schedule Result;
for (auto &P : Bands) {
isl::schedule_node MaybeFused = P.first;
isl::schedule_node DirectChild = P.second;
// If not modified, use the direct child.
if (!DirectChild.is_null() &&
!ChangedDirectChildren.count(DirectChild.get())) {
if (AlreadyAdded.count(DirectChild.get()))
continue;
AlreadyAdded.insert(DirectChild.get());
MaybeFused = DirectChild;
} else {
assert(AnyChange &&
"Need changed flag for be consistent with actual change");
}
// Top-down recursion: If the outermost loop has been fused, their nested
// bands might be fusable now as well.
isl::schedule InnerFused = visit(MaybeFused, SubRemainingDeps);
// Reconstruct the sequence, with some of the children fused.
if (Result.is_null())
Result = InnerFused;
else
Result = Result.sequence(InnerFused);
}
return Result;
}
};
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
} // namespace
bool polly::isBandMark(const isl::schedule_node &Node) {
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
return isMark(Node) &&
isLoopAttr(Node.as<isl::schedule_node_mark>().get_id());
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
}
BandAttr *polly::getBandAttr(isl::schedule_node MarkOrBand) {
MarkOrBand = moveToBandMark(MarkOrBand);
if (!isMark(MarkOrBand))
return nullptr;
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
return getLoopAttr(MarkOrBand.as<isl::schedule_node_mark>().get_id());
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
}
isl::schedule polly::hoistExtensionNodes(isl::schedule Sched) {
// If there is no extension node in the first place, return the original
// schedule tree.
if (!containsExtensionNode(Sched))
return Sched;
// Build options can anchor schedule nodes, such that the schedule tree cannot
// be modified anymore. Therefore, apply build options after the tree has been
// created.
CollectASTBuildOptions Collector;
Collector.visit(Sched);
// Rewrite the schedule tree without extension nodes.
ExtensionNodeRewriter Rewriter;
isl::schedule NewSched = Rewriter.visitSchedule(Sched);
// Reapply the AST build options. The rewriter must not change the iteration
// order of bands. Any other node type is ignored.
ApplyASTBuildOptions Applicator(Collector.ASTBuildOptions);
NewSched = Applicator.visitSchedule(NewSched);
return NewSched;
}
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
isl::schedule polly::applyFullUnroll(isl::schedule_node BandToUnroll) {
isl::ctx Ctx = BandToUnroll.ctx();
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
// Remove the loop's mark, the loop will disappear anyway.
BandToUnroll = removeMark(BandToUnroll);
assert(isBandWithSingleLoop(BandToUnroll));
isl::multi_union_pw_aff PartialSched = isl::manage(
isl_schedule_node_band_get_partial_schedule(BandToUnroll.get()));
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
assert(PartialSched.dim(isl::dim::out).release() == 1 &&
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
"Can only unroll a single dimension");
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
isl::union_pw_aff PartialSchedUAff = PartialSched.at(0);
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
isl::union_set Domain = BandToUnroll.get_domain();
PartialSchedUAff = PartialSchedUAff.intersect_domain(Domain);
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
isl::union_map PartialSchedUMap =
isl::union_map::from(isl::union_pw_multi_aff(PartialSchedUAff));
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
// Enumerator only the scatter elements.
isl::union_set ScatterList = PartialSchedUMap.range();
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
// Enumerate all loop iterations.
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
// TODO: Diagnose if not enumerable or depends on a parameter.
SmallVector<isl::point, 16> Elts;
ScatterList.foreach_point([&Elts](isl::point P) -> isl::stat {
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
Elts.push_back(P);
return isl::stat::ok();
});
// Don't assume that foreach_point returns in execution order.
llvm::sort(Elts, [](isl::point P1, isl::point P2) -> bool {
isl::val C1 = P1.get_coordinate_val(isl::dim::set, 0);
isl::val C2 = P2.get_coordinate_val(isl::dim::set, 0);
return C1.lt(C2);
});
// Convert the points to a sequence of filters.
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
isl::union_set_list List = isl::union_set_list(Ctx, Elts.size());
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
for (isl::point P : Elts) {
// Determine the domains that map this scatter element.
isl::union_set DomainFilter = PartialSchedUMap.intersect_range(P).domain();
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
List = List.add(DomainFilter);
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
}
// Replace original band with unrolled sequence.
isl::schedule_node Body =
isl::manage(isl_schedule_node_delete(BandToUnroll.release()));
Body = Body.insert_sequence(List);
return Body.get_schedule();
}
isl::schedule polly::applyPartialUnroll(isl::schedule_node BandToUnroll,
int Factor) {
assert(Factor > 0 && "Positive unroll factor required");
isl::ctx Ctx = BandToUnroll.ctx();
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
// Remove the mark, save the attribute for later use.
BandAttr *Attr;
BandToUnroll = removeMark(BandToUnroll, Attr);
assert(isBandWithSingleLoop(BandToUnroll));
isl::multi_union_pw_aff PartialSched = isl::manage(
isl_schedule_node_band_get_partial_schedule(BandToUnroll.get()));
// { Stmt[] -> [x] }
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
isl::union_pw_aff PartialSchedUAff = PartialSched.at(0);
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
// Here we assume the schedule stride is one and starts with 0, which is not
// necessarily the case.
isl::union_pw_aff StridedPartialSchedUAff =
isl::union_pw_aff::empty(PartialSchedUAff.get_space());
isl::val ValFactor{Ctx, Factor};
PartialSchedUAff.foreach_pw_aff([&StridedPartialSchedUAff,
&ValFactor](isl::pw_aff PwAff) -> isl::stat {
isl::space Space = PwAff.get_space();
isl::set Universe = isl::set::universe(Space.domain());
isl::pw_aff AffFactor{Universe, ValFactor};
isl::pw_aff DivSchedAff = PwAff.div(AffFactor).floor().mul(AffFactor);
StridedPartialSchedUAff = StridedPartialSchedUAff.union_add(DivSchedAff);
return isl::stat::ok();
});
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
isl::union_set_list List = isl::union_set_list(Ctx, Factor);
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
for (auto i : seq<int>(0, Factor)) {
// { Stmt[] -> [x] }
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
isl::union_map UMap =
isl::union_map::from(isl::union_pw_multi_aff(PartialSchedUAff));
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
// { [x] }
isl::basic_set Divisible = isDivisibleBySet(Ctx, Factor, i);
// { Stmt[] }
isl::union_set UnrolledDomain = UMap.intersect_range(Divisible).domain();
List = List.add(UnrolledDomain);
}
isl::schedule_node Body =
isl::manage(isl_schedule_node_delete(BandToUnroll.copy()));
Body = Body.insert_sequence(List);
isl::schedule_node NewLoop =
Body.insert_partial_schedule(StridedPartialSchedUAff);
MDNode *FollowupMD = nullptr;
if (Attr && Attr->Metadata)
FollowupMD =
findOptionalNodeOperand(Attr->Metadata, LLVMLoopUnrollFollowupUnrolled);
isl::id NewBandId = createGeneratedLoopAttr(Ctx, FollowupMD);
if (!NewBandId.is_null())
[Polly][Optimizer] Apply user-directed unrolling. Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule. While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default. This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter. Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088. Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks. Reviewed By: sebastiankreutzer Differential Revision: https://reviews.llvm.org/D97977
2021-03-16 01:13:21 +08:00
NewLoop = insertMark(NewLoop, NewBandId);
return NewLoop.get_schedule();
}
isl::set polly::getPartialTilePrefixes(isl::set ScheduleRange,
int VectorWidth) {
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
isl_size Dims = ScheduleRange.tuple_dim().release();
isl::set LoopPrefixes =
ScheduleRange.drop_constraints_involving_dims(isl::dim::set, Dims - 1, 1);
auto ExtentPrefixes = addExtentConstraints(LoopPrefixes, VectorWidth);
isl::set BadPrefixes = ExtentPrefixes.subtract(ScheduleRange);
BadPrefixes = BadPrefixes.project_out(isl::dim::set, Dims - 1, 1);
LoopPrefixes = LoopPrefixes.project_out(isl::dim::set, Dims - 1, 1);
return LoopPrefixes.subtract(BadPrefixes);
}
isl::union_set polly::getIsolateOptions(isl::set IsolateDomain,
isl_size OutDimsNum) {
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
isl_size Dims = IsolateDomain.tuple_dim().release();
assert(OutDimsNum <= Dims &&
"The isl::set IsolateDomain is used to describe the range of schedule "
"dimensions values, which should be isolated. Consequently, the "
"number of its dimensions should be greater than or equal to the "
"number of the schedule dimensions.");
isl::map IsolateRelation = isl::map::from_domain(IsolateDomain);
IsolateRelation = IsolateRelation.move_dims(isl::dim::out, 0, isl::dim::in,
Dims - OutDimsNum, OutDimsNum);
isl::set IsolateOption = IsolateRelation.wrap();
isl::id Id = isl::id::alloc(IsolateOption.ctx(), "isolate", nullptr);
IsolateOption = IsolateOption.set_tuple_id(Id);
return isl::union_set(IsolateOption);
}
isl::union_set polly::getDimOptions(isl::ctx Ctx, const char *Option) {
isl::space Space(Ctx, 0, 1);
auto DimOption = isl::set::universe(Space);
auto Id = isl::id::alloc(Ctx, Option, nullptr);
DimOption = DimOption.set_tuple_id(Id);
return isl::union_set(DimOption);
}
isl::schedule_node polly::tileNode(isl::schedule_node Node,
const char *Identifier,
ArrayRef<int> TileSizes,
int DefaultTileSize) {
auto Space = isl::manage(isl_schedule_node_band_get_space(Node.get()));
auto Dims = Space.dim(isl::dim::set);
auto Sizes = isl::multi_val::zero(Space);
std::string IdentifierString(Identifier);
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
for (auto i : seq<isl_size>(0, Dims.release())) {
auto tileSize =
i < (isl_size)TileSizes.size() ? TileSizes[i] : DefaultTileSize;
Sizes = Sizes.set_val(i, isl::val(Node.ctx(), tileSize));
}
auto TileLoopMarkerStr = IdentifierString + " - Tiles";
auto TileLoopMarker = isl::id::alloc(Node.ctx(), TileLoopMarkerStr, nullptr);
Node = Node.insert_mark(TileLoopMarker);
Node = Node.child(0);
Node =
isl::manage(isl_schedule_node_band_tile(Node.release(), Sizes.release()));
Node = Node.child(0);
auto PointLoopMarkerStr = IdentifierString + " - Points";
auto PointLoopMarker =
isl::id::alloc(Node.ctx(), PointLoopMarkerStr, nullptr);
Node = Node.insert_mark(PointLoopMarker);
return Node.child(0);
}
isl::schedule_node polly::applyRegisterTiling(isl::schedule_node Node,
ArrayRef<int> TileSizes,
int DefaultTileSize) {
Node = tileNode(Node, "Register tiling", TileSizes, DefaultTileSize);
auto Ctx = Node.ctx();
[Polly][Isl] Move to the new-polly-generator branch version of isl-noexceptions.h. NFCI This is part of an effort to reduce the differences between the custom C++ bindings used right now by polly in `lib/External/isl/include/isl/isl-noxceptions.h` and the official isl C++ interface. With this commit we are moving from the `polly-generator` branch to the `new-polly-generator` branch that is more mantainable and is based on the official C++ interface `cpp-checked.h`. Changes made: - There are now many sublcasses for `isl::ast_node` representing different isl types. Use `isl::ast_node_for`, `isl::ast_node_user`, `isl::ast_node_block` and `isl::ast_node_mark` where needed. - There are now many sublcasses for `isl::schedule_node` representing different isl types. Use `isl::schedule_node_mark`, `isl::schedule_node_extension`, `isl::schedule_node_band` and `isl::schedule_node_filter` where needed. - Replace the `isl::*::dump` with `dumpIslObj` since the isl dump method is not exposed in the C++ interface. - `isl::schedule_node::get_child` has been renamed to `isl::schedule_node::child` - `isl::pw_multi_aff::get_pw_aff` has been renamed to `isl::pw_multi_aff::at` - The constructor `isl::union_map(isl::union_pw_multi_aff)` has been replaced with the static method `isl::union_map::from()` - Replace usages of `isl::val::add_ui` with `isl::val::add` - `isl::union_set_list::alloc` is now a constructor - All the `isl_size` values are now wrapped inside the class `isl::size` use `isl::size::release` to get the internal `isl_size` value where needed. - `isl-noexceptions.h` has been generated by https://github.com/patacca/isl/commit/73f5ed1f4d1f72582f731590ef9e43d9ab1956ad No functional change intended. Reviewed By: Meinersbur Differential Revision: https://reviews.llvm.org/D107225
2021-08-16 21:52:24 +08:00
return Node.as<isl::schedule_node_band>().set_ast_build_options(
isl::union_set(Ctx, "{unroll[x]}"));
}
/// Find statements and sub-loops in (possibly nested) sequences.
static void
collectFussionableStmts(isl::schedule_node Node,
SmallVectorImpl<isl::schedule_node> &ScheduleStmts) {
if (isBand(Node) || isLeaf(Node)) {
ScheduleStmts.push_back(Node);
return;
}
if (Node.has_children()) {
isl::schedule_node C = Node.first_child();
while (true) {
collectFussionableStmts(C, ScheduleStmts);
if (!C.has_next_sibling())
break;
C = C.next_sibling();
}
}
}
isl::schedule polly::applyMaxFission(isl::schedule_node BandToFission) {
isl::ctx Ctx = BandToFission.ctx();
BandToFission = removeMark(BandToFission);
isl::schedule_node BandBody = BandToFission.child(0);
SmallVector<isl::schedule_node> FissionableStmts;
collectFussionableStmts(BandBody, FissionableStmts);
size_t N = FissionableStmts.size();
// Collect the domain for each of the statements that will get their own loop.
isl::union_set_list DomList = isl::union_set_list(Ctx, N);
for (size_t i = 0; i < N; ++i) {
isl::schedule_node BodyPart = FissionableStmts[i];
DomList = DomList.add(BodyPart.get_domain());
}
// Apply the fission by copying the entire loop, but inserting a filter for
// the statement domains for each fissioned loop.
isl::schedule_node Fissioned = BandToFission.insert_sequence(DomList);
return Fissioned.get_schedule();
}
isl::schedule polly::applyGreedyFusion(isl::schedule Sched,
const isl::union_map &Deps) {
LLVM_DEBUG(dbgs() << "Greedy loop fusion\n");
GreedyFusionRewriter Rewriter;
isl::schedule Result = Rewriter.visit(Sched, Deps);
if (!Rewriter.AnyChange) {
LLVM_DEBUG(dbgs() << "Found nothing to fuse\n");
return Sched;
}
// GreedyFusionRewriter due to working loop-by-loop, bands with multiple loops
// may have been split into multiple bands.
return collapseBands(Result);
}