ConstraintSystem: replace data structure with Matrix #98895
Conversation
artagnon
requested review from
MaskRay,
antoniofrighetto,
arsenm,
fhahn,
jayfoad and
nikic
llvmbot
added
llvm:analysis
|
@llvm/pr-subscribers-llvm-transforms @llvm/pr-subscribers-llvm-analysis Author: Ramkumar Ramachandra (artagnon) ChangesConstraintSystem currently stores constraints in a vector-of-vectors, performing inefficient memory operations on it, as part of its Since Matrix requires knowing an upper bounds on the number of columns ahead-of-time, add a cl::opt to ConstraintElimination upper-bounding this, and change addVariableRowFill to not add constraints whose length exceeds this upper bound. -- 8< -- Patch is 44.36 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/98895.diff 7 Files Affected:
diff --git a/llvm/include/llvm/ADT/ArrayRef.h b/llvm/include/llvm/ADT/ArrayRef.h
index 1c6799f1c56ed..6fc7acff7ba8c 100644
--- a/llvm/include/llvm/ADT/ArrayRef.h
+++ b/llvm/include/llvm/ADT/ArrayRef.h
@@ -25,6 +25,7 @@
namespace llvm {
template<typename T> class [[nodiscard]] MutableArrayRef;
+ template <typename T> struct [[nodiscard]] MatrixRowView;
/// ArrayRef - Represent a constant reference to an array (0 or more elements
/// consecutively in memory), i.e. a start pointer and a length. It allows
@@ -59,19 +60,21 @@ namespace llvm {
/// The number of elements.
size_type Length = 0;
+ friend MatrixRowView<T>;
+
public:
/// @name Constructors
/// @{
/// Construct an empty ArrayRef.
- /*implicit*/ ArrayRef() = default;
+ /*implicit*/ constexpr ArrayRef() = default;
/// Construct an empty ArrayRef from std::nullopt.
- /*implicit*/ ArrayRef(std::nullopt_t) {}
+ /*implicit*/ constexpr ArrayRef(std::nullopt_t) {}
/// Construct an ArrayRef from a single element.
- /*implicit*/ ArrayRef(const T &OneElt)
- : Data(&OneElt), Length(1) {}
+ /*implicit*/ constexpr ArrayRef(const T &OneElt)
+ : Data(&OneElt), Length(1) {}
/// Construct an ArrayRef from a pointer and length.
constexpr /*implicit*/ ArrayRef(const T *data, size_t length)
@@ -123,9 +126,10 @@ namespace llvm {
/// Construct an ArrayRef<const T*> from ArrayRef<T*>. This uses SFINAE to
/// ensure that only ArrayRefs of pointers can be converted.
template <typename U>
- ArrayRef(const ArrayRef<U *> &A,
- std::enable_if_t<std::is_convertible<U *const *, T const *>::value>
- * = nullptr)
+ constexpr ArrayRef(
+ const ArrayRef<U *> &A,
+ std::enable_if_t<std::is_convertible<U *const *, T const *>::value> * =
+ nullptr)
: Data(A.data()), Length(A.size()) {}
/// Construct an ArrayRef<const T*> from a SmallVector<T*>. This is
@@ -150,28 +154,32 @@ namespace llvm {
/// @name Simple Operations
/// @{
- iterator begin() const { return Data; }
- iterator end() const { return Data + Length; }
+ constexpr iterator begin() const { return Data; }
+ constexpr iterator end() const { return Data + Length; }
- reverse_iterator rbegin() const { return reverse_iterator(end()); }
- reverse_iterator rend() const { return reverse_iterator(begin()); }
+ constexpr reverse_iterator rbegin() const {
+ return reverse_iterator(end());
+ }
+ constexpr reverse_iterator rend() const {
+ return reverse_iterator(begin());
+ }
/// empty - Check if the array is empty.
- bool empty() const { return Length == 0; }
+ constexpr bool empty() const { return Length == 0; }
- const T *data() const { return Data; }
+ constexpr const T *data() const { return Data; }
/// size - Get the array size.
- size_t size() const { return Length; }
+ constexpr size_t size() const { return Length; }
/// front - Get the first element.
- const T &front() const {
+ constexpr const T &front() const {
assert(!empty());
return Data[0];
}
/// back - Get the last element.
- const T &back() const {
+ constexpr const T &back() const {
assert(!empty());
return Data[Length-1];
}
@@ -184,7 +192,7 @@ namespace llvm {
}
/// equals - Check for element-wise equality.
- bool equals(ArrayRef RHS) const {
+ constexpr bool equals(ArrayRef RHS) const {
if (Length != RHS.Length)
return false;
return std::equal(begin(), end(), RHS.begin());
@@ -192,22 +200,22 @@ namespace llvm {
/// slice(n, m) - Chop off the first N elements of the array, and keep M
/// elements in the array.
- ArrayRef<T> slice(size_t N, size_t M) const {
+ constexpr ArrayRef<T> slice(size_t N, size_t M) const {
assert(N+M <= size() && "Invalid specifier");
return ArrayRef<T>(data()+N, M);
}
/// slice(n) - Chop off the first N elements of the array.
- ArrayRef<T> slice(size_t N) const { return slice(N, size() - N); }
+ constexpr ArrayRef<T> slice(size_t N) const { return slice(N, size() - N); }
/// Drop the first \p N elements of the array.
- ArrayRef<T> drop_front(size_t N = 1) const {
+ constexpr ArrayRef<T> drop_front(size_t N = 1) const {
assert(size() >= N && "Dropping more elements than exist");
return slice(N, size() - N);
}
/// Drop the last \p N elements of the array.
- ArrayRef<T> drop_back(size_t N = 1) const {
+ constexpr ArrayRef<T> drop_back(size_t N = 1) const {
assert(size() >= N && "Dropping more elements than exist");
return slice(0, size() - N);
}
@@ -225,14 +233,14 @@ namespace llvm {
}
/// Return a copy of *this with only the first \p N elements.
- ArrayRef<T> take_front(size_t N = 1) const {
+ constexpr ArrayRef<T> take_front(size_t N = 1) const {
if (N >= size())
return *this;
return drop_back(size() - N);
}
/// Return a copy of *this with only the last \p N elements.
- ArrayRef<T> take_back(size_t N = 1) const {
+ constexpr ArrayRef<T> take_back(size_t N = 1) const {
if (N >= size())
return *this;
return drop_front(size() - N);
@@ -253,7 +261,7 @@ namespace llvm {
/// @}
/// @name Operator Overloads
/// @{
- const T &operator[](size_t Index) const {
+ constexpr const T &operator[](size_t Index) const {
assert(Index < Length && "Invalid index!");
return Data[Index];
}
@@ -319,20 +327,20 @@ namespace llvm {
using difference_type = ptrdiff_t;
/// Construct an empty MutableArrayRef.
- /*implicit*/ MutableArrayRef() = default;
+ /*implicit*/ constexpr MutableArrayRef() = default;
/// Construct an empty MutableArrayRef from std::nullopt.
- /*implicit*/ MutableArrayRef(std::nullopt_t) : ArrayRef<T>() {}
+ /*implicit*/ constexpr MutableArrayRef(std::nullopt_t) : ArrayRef<T>() {}
/// Construct a MutableArrayRef from a single element.
- /*implicit*/ MutableArrayRef(T &OneElt) : ArrayRef<T>(OneElt) {}
+ /*implicit*/ constexpr MutableArrayRef(T &OneElt) : ArrayRef<T>(OneElt) {}
/// Construct a MutableArrayRef from a pointer and length.
- /*implicit*/ MutableArrayRef(T *data, size_t length)
- : ArrayRef<T>(data, length) {}
+ /*implicit*/ constexpr MutableArrayRef(T *data, size_t length)
+ : ArrayRef<T>(data, length) {}
/// Construct a MutableArrayRef from a range.
- MutableArrayRef(T *begin, T *end) : ArrayRef<T>(begin, end) {}
+ constexpr MutableArrayRef(T *begin, T *end) : ArrayRef<T>(begin, end) {}
/// Construct a MutableArrayRef from a SmallVector.
/*implicit*/ MutableArrayRef(SmallVectorImpl<T> &Vec)
@@ -351,45 +359,49 @@ namespace llvm {
template <size_t N>
/*implicit*/ constexpr MutableArrayRef(T (&Arr)[N]) : ArrayRef<T>(Arr) {}
- T *data() const { return const_cast<T*>(ArrayRef<T>::data()); }
+ constexpr T *data() const { return const_cast<T *>(ArrayRef<T>::data()); }
- iterator begin() const { return data(); }
- iterator end() const { return data() + this->size(); }
+ constexpr iterator begin() const { return data(); }
+ constexpr iterator end() const { return data() + this->size(); }
- reverse_iterator rbegin() const { return reverse_iterator(end()); }
- reverse_iterator rend() const { return reverse_iterator(begin()); }
+ constexpr reverse_iterator rbegin() const {
+ return reverse_iterator(end());
+ }
+ constexpr reverse_iterator rend() const {
+ return reverse_iterator(begin());
+ }
/// front - Get the first element.
- T &front() const {
+ constexpr T &front() const {
assert(!this->empty());
return data()[0];
}
/// back - Get the last element.
- T &back() const {
+ constexpr T &back() const {
assert(!this->empty());
return data()[this->size()-1];
}
/// slice(n, m) - Chop off the first N elements of the array, and keep M
/// elements in the array.
- MutableArrayRef<T> slice(size_t N, size_t M) const {
+ constexpr MutableArrayRef<T> slice(size_t N, size_t M) const {
assert(N + M <= this->size() && "Invalid specifier");
return MutableArrayRef<T>(this->data() + N, M);
}
/// slice(n) - Chop off the first N elements of the array.
- MutableArrayRef<T> slice(size_t N) const {
+ constexpr MutableArrayRef<T> slice(size_t N) const {
return slice(N, this->size() - N);
}
/// Drop the first \p N elements of the array.
- MutableArrayRef<T> drop_front(size_t N = 1) const {
+ constexpr MutableArrayRef<T> drop_front(size_t N = 1) const {
assert(this->size() >= N && "Dropping more elements than exist");
return slice(N, this->size() - N);
}
- MutableArrayRef<T> drop_back(size_t N = 1) const {
+ constexpr MutableArrayRef<T> drop_back(size_t N = 1) const {
assert(this->size() >= N && "Dropping more elements than exist");
return slice(0, this->size() - N);
}
@@ -409,14 +421,14 @@ namespace llvm {
}
/// Return a copy of *this with only the first \p N elements.
- MutableArrayRef<T> take_front(size_t N = 1) const {
+ constexpr MutableArrayRef<T> take_front(size_t N = 1) const {
if (N >= this->size())
return *this;
return drop_back(this->size() - N);
}
/// Return a copy of *this with only the last \p N elements.
- MutableArrayRef<T> take_back(size_t N = 1) const {
+ constexpr MutableArrayRef<T> take_back(size_t N = 1) const {
if (N >= this->size())
return *this;
return drop_front(this->size() - N);
@@ -439,7 +451,7 @@ namespace llvm {
/// @}
/// @name Operator Overloads
/// @{
- T &operator[](size_t Index) const {
+ constexpr T &operator[](size_t Index) const {
assert(Index < this->size() && "Invalid index!");
return data()[Index];
}
diff --git a/llvm/include/llvm/ADT/Matrix.h b/llvm/include/llvm/ADT/Matrix.h
new file mode 100644
index 0000000000000..eeb9702772738
--- /dev/null
+++ b/llvm/include/llvm/ADT/Matrix.h
@@ -0,0 +1,339 @@
+//===- Matrix.h - Two-dimensional Container with View -----------*- 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
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_ADT_MATRIX_H
+#define LLVM_ADT_MATRIX_H
+
+#include "llvm/ADT/ArrayRef.h"
+#include "llvm/ADT/SmallVector.h"
+
+namespace llvm {
+template <typename T, size_t M, size_t NStorageInline> class MatrixView;
+
+/// Due to the SmallVector infrastructure using SmallVectorAlignmentAndOffset
+/// that depends on the exact data layout, no derived classes can have extra
+/// members.
+template <typename T, size_t N>
+struct MatrixStorageBase : public SmallVectorImpl<T>, SmallVectorStorage<T, N> {
+ MatrixStorageBase() : SmallVectorImpl<T>(N) {}
+ MatrixStorageBase(size_t Size) : SmallVectorImpl<T>(N) { resize(Size); }
+ ~MatrixStorageBase() { destroy_range(this->begin(), this->end()); }
+ MatrixStorageBase(const MatrixStorageBase &RHS) : SmallVectorImpl<T>(N) {
+ if (!RHS.empty())
+ SmallVectorImpl<T>::operator=(RHS);
+ }
+ MatrixStorageBase(MatrixStorageBase &&RHS) : SmallVectorImpl<T>(N) {
+ if (!RHS.empty())
+ SmallVectorImpl<T>::operator=(::std::move(RHS));
+ }
+ using SmallVectorImpl<T>::size;
+ using SmallVectorImpl<T>::resize;
+ using SmallVectorImpl<T>::append;
+ using SmallVectorImpl<T>::erase;
+ using SmallVectorImpl<T>::destroy_range;
+
+ T *begin() const { return const_cast<T *>(SmallVectorImpl<T>::begin()); }
+ T *end() const { return const_cast<T *>(SmallVectorImpl<T>::end()); }
+};
+
+/// A two-dimensional container storage, whose upper bound on the number of
+/// columns should be known ahead of time. Not menat to be used directly: the
+/// primary usage API is MatrixView.
+template <typename T,
+ size_t N = CalculateSmallVectorDefaultInlinedElements<T>::value>
+class MatrixStorage {
+public:
+ MatrixStorage() = delete;
+ MatrixStorage(size_t NRows, size_t NCols)
+ : Base(NRows * NCols), NCols(NCols) {}
+ MatrixStorage(size_t NCols) : Base(), NCols(NCols) {}
+
+ size_t size() const { return Base.size(); }
+ bool empty() const { return !size(); }
+ size_t getNumRows() const { return size() / NCols; }
+ size_t getNumCols() const { return NCols; }
+ void setNumCols(size_t NCols) {
+ assert(empty() && "Column-resizing a non-empty MatrixStorage");
+ this->NCols = NCols;
+ }
+ void resize(size_t NRows) { Base.resize(NCols * NRows); }
+
+protected:
+ template <typename U, size_t M, size_t NStorageInline>
+ friend class MatrixView;
+
+ T *begin() const { return Base.begin(); }
+ T *rowFromIdx(size_t RowIdx, size_t Offset = 0) const {
+ return begin() + RowIdx * NCols + Offset;
+ }
+ std::pair<size_t, size_t> idxFromRow(T *Ptr) const {
+ assert(Ptr >= begin() && "Internal error");
+ size_t Offset = (Ptr - begin()) % NCols;
+ return {(Ptr - begin()) / NCols, Offset};
+ }
+
+ // If Arg.size() < NCols, the number of columns won't be changed, and the
+ // difference is default-constructed.
+ void addRow(const SmallVectorImpl<T> &Arg) {
+ assert(Arg.size() <= NCols &&
+ "MatrixStorage has insufficient number of columns");
+ size_t Diff = NCols - Arg.size();
+ Base.append(Arg.begin(), Arg.end());
+ Base.append(Diff, T());
+ }
+
+private:
+ MatrixStorageBase<T, N> Base;
+ size_t NCols;
+};
+
+/// MutableArrayRef with a copy-assign, and extra APIs.
+template <typename T>
+struct [[nodiscard]] MatrixRowView : public MutableArrayRef<T> {
+ using pointer = typename MutableArrayRef<T>::pointer;
+ using iterator = typename MutableArrayRef<T>::iterator;
+ using const_iterator = typename MutableArrayRef<T>::const_iterator;
+
+ constexpr MatrixRowView() = delete;
+ constexpr MatrixRowView(pointer Data, size_t Length)
+ : MutableArrayRef<T>(Data, Length) {}
+ constexpr MatrixRowView(iterator Begin, iterator End)
+ : MutableArrayRef<T>(Begin, End) {}
+ constexpr MatrixRowView(const_iterator Begin, const_iterator End)
+ : MutableArrayRef<T>(Begin, End) {}
+ constexpr MatrixRowView(MutableArrayRef<T> Other)
+ : MutableArrayRef<T>(Other.data(), Other.size()) {}
+ MatrixRowView(const SmallVectorImpl<T> &Vec) : MutableArrayRef<T>(Vec) {}
+
+ using MutableArrayRef<T>::size;
+ using MutableArrayRef<T>::data;
+
+ constexpr T &back() const { return MutableArrayRef<T>::back(); }
+ constexpr T &front() const { return MutableArrayRef<T>::front(); }
+ constexpr MatrixRowView<T> drop_back(size_t N = 1) const { // NOLINT
+ return MutableArrayRef<T>::drop_back(N);
+ }
+ constexpr MatrixRowView<T> drop_front(size_t N = 1) const { // NOLINT
+ return MutableArrayRef<T>::drop_front(N);
+ }
+ // This slice is different from the MutableArrayRef slice, and specifies a
+ // Begin and End index, instead of a Begin and Length.
+ constexpr MatrixRowView<T> slice(size_t Begin, size_t End) {
+ return MutableArrayRef<T>::slice(Begin, End - Begin);
+ }
+ constexpr void pop_back(size_t N = 1) { // NOLINT
+ this->Length -= N;
+ }
+ constexpr void pop_front(size_t N = 1) { // NOLINT
+ this->Data += N;
+ this->Length -= N;
+ }
+
+ MatrixRowView &operator=(const SmallVectorImpl<T> &Vec) {
+ copy_assign(Vec.begin(), Vec.end());
+ return *this;
+ }
+ MatrixRowView &operator=(std::initializer_list<T> IL) {
+ copy_assign(IL.begin(), IL.end());
+ return *this;
+ }
+
+ void swap(MatrixRowView<T> &Other) {
+ std::swap(this->Data, Other.Data);
+ std::swap(this->Length, Other.Length);
+ }
+
+protected:
+ void copy_assign(iterator Begin, iterator End) { // NOLINT
+ std::uninitialized_copy(Begin, End, data());
+ this->Length = End - Begin;
+ }
+ void copy_assign(const_iterator Begin, const_iterator End) { // NOLINT
+ std::uninitialized_copy(Begin, End, data());
+ this->Length = End - Begin;
+ }
+};
+
+/// The primary usage API of MatrixStorage. Abstracts out indexing-arithmetic,
+/// eliminating memory operations on the underlying data. Supports
+/// variable-length columns.
+template <typename T,
+ size_t N = CalculateSmallVectorDefaultInlinedElements<T>::value,
+ size_t NStorageInline =
+ CalculateSmallVectorDefaultInlinedElements<T>::value>
+class [[nodiscard]] MatrixView {
+public:
+ using row_type = MatrixRowView<T>;
+ using container_type = SmallVector<row_type, N>;
+ using iterator = typename container_type::iterator;
+ using const_iterator = typename container_type::const_iterator;
+
+ constexpr MatrixView(MatrixStorage<T, NStorageInline> &Mat, size_t RowSpan,
+ size_t ColSpan)
+ : Mat(Mat) {
+ RowView.reserve(RowSpan);
+ for (size_t RowIdx = 0; RowIdx < RowSpan; ++RowIdx) {
+ auto RangeBegin = Mat.begin() + RowIdx * ColSpan;
+ RowView.emplace_back(RangeBegin, RangeBegin + ColSpan);
+ }
+ }
+
+ // Constructor with a full View of the underlying MatrixStorage, if
+ // MatrixStorage has a non-zero number of Columns. Otherwise, creates an empty
+ // view.
+ constexpr MatrixView(MatrixStorage<T, NStorageInline> &Mat)
+ : MatrixView(Mat, Mat.getNumRows(), Mat.getNumCols()) {}
+
+ // Obvious copy-construator is deleted, since the underlying storage could
+ // have changed.
+ constexpr MatrixView(const MatrixView &) = delete;
+
+ // Copy-assignment operator should not be used when the underlying storage
+ // changes.
+ constexpr MatrixView &operator=(const MatrixView &Other) {
+ assert(Mat.begin() == Other.Mat.begin() &&
+ "Underlying storage has changed: use custom copy-constructor");
+ RowView = Other.RowView;
+ return *this;
+ }
+
+ // The actual copy-constructor: to be used when the underlying storage is
+ // copy-constructed.
+ MatrixView(const MatrixView &OldView,
+ MatrixStorage<T, NStorageInline> &NewMat)
+ : Mat(NewMat) {
+ assert(OldView.Mat.size() == Mat.size() &&
+ "Custom copy-constructor called on non-copied storage");
+
+ // The underlying storage will change. Construct a new RowView by performing
+ // pointer-arithmetic on the underlying storage of OldView, using pointers
+ // from OldVie.
+ for (const auto &R : OldView.RowView) {
+ auto [StorageIdx, StartOffset] = OldView.Mat.idxFromRow(R.data());
+ RowView.emplace_back(Mat.rowFromIdx(StorageIdx, StartOffset), R.size());
+ }
+ }
+
+ void addRow(const SmallVectorImpl<T> &Row) {
+ // The underlying storage may be resized, performing reallocations. The
+ // pointers in RowView will no longer be valid, so save and restore the
+ // data. Construct RestoreData by performing pointer-arithmetic on the
+ // underlying storgge.
+ SmallVector<std::tuple<size_t, size_t, size_t>> RestoreData;
+ RestoreData.reserve(RowView.size());
+ for (const auto &R : RowView) {
+ auto [StorageIdx, StartOffset] = Mat.idxFromRow(R.data());
+ RestoreData.emplace_back(StorageIdx, StartOffset, R.size());
+ }
+
+ Mat.addRow(Row);
+
+ // Restore the RowView by performing pointer-arithmetic on the
+ // possibly-reallocated storage, using information from RestoreData.
+ RowView.clear();
+ for (const auto &[StorageIdx, StartOffset, Len] : RestoreData)
+ RowView.emplace_back(Mat.rowFromIdx(StorageIdx, StartOffset), Len);
+
+ // Finally, add the new row to the VRowView.
+ RowView.emplace_back(Mat.rowFromIdx(Mat.getNumRows() - 1), Row.size());
+ }
+
+ // To support addRow(View[Idx]).
+ void addRow(const row_type &Row) { addRow(SmallVector<T>{Row}); }
+
+ void addRow(std::initializer_list<T> Row) { addRow(SmallVector<T>{Row}); }
+
+ constexpr row_type &operator[](size_t RowIdx) {
+ assert(RowIdx < RowView.size() && "Indexing out of bounds");
+ return RowView[RowIdx];
+ }
+
+ constexpr T *data() const {
+ assert(!empty() && "Non-empty view expected");
+ return RowView.front().data();
+ }
+ size_t size() const { return getRowSpan() * getMaxColSpan(); }
+...
[truncated]
|
|
The LLVM Compile Time Tracker results are in, and they look good: https://llvm-compile-time-tracker.com/compare.php?from=c5ee3c05ca61f3fae11337c5853aee7b450a9dc6&to=f78bc6008c7dea8acce3a4171dd3ece15103419f&stat=instructions:u. |
artagnon
force-pushed
the
perf/constraintsystem-matrix
branch
from
b1c870b to
f78bc60
Compare
|
Thanks for working on improving the representation used!
I think the current cutoff is probably too low, I think in practice we often have systems with more than 16 columns/variables (there are a number of binary changes in CTMark due to this I think https://llvm-compile-time-tracker.com/compare.php?from=ef1851a4daa5a9b48cdff9bac12a116c6518d39b&to=f78bc6008c7dea8acce3a4171dd3ece15103419f&stat=size-text). To compare apples to apples compile-time wise, it would be good to pick the default so we handle all cases handled without this change (might pessimize this change due to large up-front allocations though) or possibly resize the matrix on demand (which would need to repack the existing data) |
I see. Should I attempt to change it to 32 or 64? What's a good upper-bound on the number of columns, if the number of rows are upper-bounded at 500?
I think one large upfront allocation wouldn't hurt as much as repeatedly calling malloc/free to resize the underlying storage. I think things can be improved by doing a dry-run, upper-bounding the number of rows and columns, and performing the entire allocation upfront. |
Another thing I noticed was a call to a copy-constructor that copies the underlying storage. Is it really necessary to copy the underlying storage? |
artagnon
force-pushed
the
perf/constraintsystem-matrix
branch
4 times, most recently
from
4f19a2b to
94bdfca
Compare
|
Looks like Clang crashes with the dry-run version on CTmark https://llvm-compile-time-tracker.com/show_error.php?commit=94bdfca2db58d6e6ff507d073e407acba1046701 |
artagnon
force-pushed
the
perf/constraintsystem-matrix
branch
from
94bdfca to
82ddf54
Compare
Should be fixed now. However, LLVM Compile Time Tracker times out now, and I have no idea how to go about debugging that: https://llvm-compile-time-tracker.com/show_error.php?commit=82ddf545ab4688a550881a7a7404d40fc585143e. |
artagnon
force-pushed
the
perf/constraintsystem-matrix
branch
from
82ddf54 to
bab070c
Compare
This looks like a hang while building clang in stage 2 (while building LLVM/Clang with a clang that contains 82ddf545ab4688a550881a7a7404d40fc585143e ). Doing a stage2 build should hopefully reproduce the issue locally |
artagnon
force-pushed
the
perf/constraintsystem-matrix
branch
from
9d14061 to
9683576
Compare
llvm/include/llvm/ADT/Matrix.h
Outdated
There was a problem hiding this comment.
Would it be possible to first have this code live somewhere local to ConstraintSystem and once it matures / gains more usage, promote it to ADT?
There was a problem hiding this comment.
Possible, since it's header-only. I would like some other opinions as well.
There was a problem hiding this comment.
I'd strongly prefer this option then. Also, please do not click 'resolve' on conversations that haven't been resolved.
artagnon
force-pushed
the
perf/constraintsystem-matrix
branch
3 times, most recently
from
9e59d0e to
28aa6d4
Compare
Leverage the excellent SmallVector infrastructure to write a two-dimensional container, along with a View that abstracts out indexing-arithmetic, eliminating memory operations on the underlying storage. The immediate applicability of Matrix is to replace uses of the vector-of-vectors idiom, with one caveat: an upper bound on the number of columns should be known ahead of time.
Add a dry-run routine that computes a conservative estimate of the number of rows and columns that the transform will require, and fail early if the estimates exceed the upper bounds. This patch has a small overhead, but improves compile-time on one benchmark significantly. The overhead will be compensated for in a follow-up patch, where ConstraintSystem is ported to use a Matrix data structure, performing the full allocation ahead-of-time using these estimates.
ConstraintSystem currently stores constraints in a vector-of-vectors, performing inefficient memory operations on it, as part of its operation. Replace this data structure with the newly-minted Matrix, using MatrixView to eliminate row-swaps and truncation of the underlying storage. Since Matrix requires knowing an upper bounds on the number of columns ahead-of-time, add a cl::opt to ConstraintElimination upper-bounding this, and change addVariableRowFill to not add constraints whose length exceeds this upper bound.
artagnon
force-pushed
the
perf/constraintsystem-matrix
branch
from
28aa6d4 to
15dd685
Compare
|
@artagnon let me know once there's a stable point for the patches, then I can run them on a larger set of programs to check the wider impact |
Will do, thanks! Sorry about the heavy load of notifications from pushes, while this is still in development: I didn't know to check |
|
Okay, the final results from LLVM Compile Time Tracker are back, and it looks like this change has a significant impact on one benchmark, and negligible regressions in ReleaseLTO: http://llvm-compile-time-tracker.com/compare.php?from=c5ee3c05ca61f3fae11337c5853aee7b450a9dc6&to=15dd68506d479cc50349de1cf1e70bd896b9c82e&stat=instructions%3Au. @fhahn Could you check it on some more programs, when you have the time? |
|
It looks like, due to the small Matrix size, the gains were all observed in the dry-run cutoff. Subtracting the impact of the dry-run patch, switching the data structure to Matrix has no impact according to LLVM Compile Time Tracker: http://llvm-compile-time-tracker.com/compare.php?from=47c8510fd2749c1f44cc33db75f7eac142919e71&to=15dd68506d479cc50349de1cf1e70bd896b9c82e&stat=instructions:u. |
6 participants
ConstraintSystem currently stores constraints in a vector-of-vectors, performing inefficient memory operations on it, as part of its operation. Replace this data structure with the newly-minted Matrix, using JaggedArrayView to eliminate row-swaps and truncation of the underlying storage.
Since Matrix requires knowing an upper bounds on the number of columns ahead-of-time, use the newly-added dry-run routine to do the entire allocation upfront.
-- 8< --
Based on #98893 and #99670.