[MachineOutliner] Leaf Descendants #90275
Merged
+462
−36
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
xuanzhang816
force-pushed
the
leaf-descendants
branch
from
75c515f to
887ef54
Compare
|
@llvm/pr-subscribers-backend-aarch64 @llvm/pr-subscribers-llvm-support Author: Xuan Zhang (xuanzh-meta) ChangesThis PR depends on #90264 In the current implementation, only leaf children of each internal node in the suffix tree are included as candidates for outlining. But all leaf descendants are outlining candidates, which we include in the new implementation. This is enabled on a flag The reason for enabling this on a flag is because machine outliner is not the only pass that uses suffix tree. The reason for having this default to be true is because including all leaf descendants show consistent size win.
This is part of an enhanced version of machine outliner -- see RFC. Patch is 32.77 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/90275.diff 15 Files Affected:
diff --git a/llvm/include/llvm/Support/SuffixTree.h b/llvm/include/llvm/Support/SuffixTree.h
index 4940fbbf308d8..37b7366640430 100644
--- a/llvm/include/llvm/Support/SuffixTree.h
+++ b/llvm/include/llvm/Support/SuffixTree.h
@@ -42,6 +42,9 @@ class SuffixTree {
/// Each element is an integer representing an instruction in the module.
ArrayRef<unsigned> Str;
+ /// Whether to consider leaf descendants or only leaf children.
+ bool OutlinerLeafDescendants;
+
/// A repeated substring in the tree.
struct RepeatedSubstring {
/// The length of the string.
@@ -130,11 +133,27 @@ class SuffixTree {
/// this step.
unsigned extend(unsigned EndIdx, unsigned SuffixesToAdd);
+ /// This vector contains all leaf nodes of this suffix tree. These leaf nodes
+ /// are identified using post-order depth-first traversal, so that the order
+ /// of these leaf nodes in the vector matches the order of the leaves in the
+ /// tree from left to right if one were to draw the tree on paper.
+ std::vector<SuffixTreeLeafNode *> LeafNodes;
+
+ /// Perform a post-order depth-first traversal of the tree and perform two
+ /// tasks during the traversal. The first is to populate LeafNodes, adding
+ /// nodes in order of the traversal. The second is to keep track of the leaf
+ /// descendants of every internal node by assigning values to LeftLeafIndex
+ /// and RightLefIndex fields of SuffixTreeNode for all internal nodes.
+ void setLeafNodes();
+
public:
/// Construct a suffix tree from a sequence of unsigned integers.
///
/// \param Str The string to construct the suffix tree for.
- SuffixTree(const ArrayRef<unsigned> &Str);
+ /// \param OutlinerLeafDescendants Whether to consider leaf descendants or
+ /// only leaf children (used by Machine Outliner).
+ SuffixTree(const ArrayRef<unsigned> &Str,
+ bool OutlinerLeafDescendants = false);
/// Iterator for finding all repeated substrings in the suffix tree.
struct RepeatedSubstringIterator {
@@ -154,6 +173,12 @@ class SuffixTree {
/// instruction lengths.
const unsigned MinLength = 2;
+ /// Vector of leaf nodes of the suffix tree.
+ const std::vector<SuffixTreeLeafNode *> &LeafNodes;
+
+ /// Whether to consider leaf descendants or only leaf children.
+ bool OutlinerLeafDescendants = !LeafNodes.empty();
+
/// Move the iterator to the next repeated substring.
void advance();
@@ -179,7 +204,10 @@ class SuffixTree {
return !(*this == Other);
}
- RepeatedSubstringIterator(SuffixTreeInternalNode *N) : N(N) {
+ RepeatedSubstringIterator(
+ SuffixTreeInternalNode *N,
+ const std::vector<SuffixTreeLeafNode *> &LeafNodes = {})
+ : N(N), LeafNodes(LeafNodes) {
// Do we have a non-null node?
if (!N)
return;
@@ -191,7 +219,7 @@ class SuffixTree {
};
typedef RepeatedSubstringIterator iterator;
- iterator begin() { return iterator(Root); }
+ iterator begin() { return iterator(Root, LeafNodes); }
iterator end() { return iterator(nullptr); }
};
diff --git a/llvm/include/llvm/Support/SuffixTreeNode.h b/llvm/include/llvm/Support/SuffixTreeNode.h
index 7d0d1cf0c58b9..84b590f2deb0c 100644
--- a/llvm/include/llvm/Support/SuffixTreeNode.h
+++ b/llvm/include/llvm/Support/SuffixTreeNode.h
@@ -46,6 +46,17 @@ struct SuffixTreeNode {
/// the root to this node.
unsigned ConcatLen = 0;
+ /// These two indices give a range of indices for its leaf descendants.
+ /// Imagine drawing a tree on paper and assigning a unique index to each leaf
+ /// node in monotonically increasing order from left to right. This way of
+ /// numbering the leaf nodes allows us to associate a continuous range of
+ /// indices with each internal node. For example, if a node has leaf
+ /// descendants with indices i, i+1, ..., j, then its LeftLeafIdx is i and
+ /// its RightLeafIdx is j. These indices are for LeafNodes in the SuffixTree
+ /// class, which is constructed using post-order depth-first traversal.
+ unsigned LeftLeafIdx = EmptyIdx;
+ unsigned RightLeafIdx = EmptyIdx;
+
public:
// LLVM RTTI boilerplate.
NodeKind getKind() const { return Kind; }
@@ -56,6 +67,18 @@ struct SuffixTreeNode {
/// \returns the end index of this node.
virtual unsigned getEndIdx() const = 0;
+ /// \return the index of this node's left most leaf node.
+ unsigned getLeftLeafIdx() const;
+
+ /// \return the index of this node's right most leaf node.
+ unsigned getRightLeafIdx() const;
+
+ /// Set the index of the left most leaf node of this node to \p Idx.
+ void setLeftLeafIdx(unsigned Idx);
+
+ /// Set the index of the right most leaf node of this node to \p Idx.
+ void setRightLeafIdx(unsigned Idx);
+
/// Advance this node's StartIdx by \p Inc.
void incrementStartIdx(unsigned Inc);
@@ -168,4 +191,4 @@ struct SuffixTreeLeafNode : SuffixTreeNode {
virtual ~SuffixTreeLeafNode() = default;
};
} // namespace llvm
-#endif // LLVM_SUPPORT_SUFFIXTREE_NODE_H
\ No newline at end of file
+#endif // LLVM_SUPPORT_SUFFIXTREE_NODE_H
diff --git a/llvm/lib/CodeGen/MachineOutliner.cpp b/llvm/lib/CodeGen/MachineOutliner.cpp
index 9e2e6316dc6d1..6836f6cbc5409 100644
--- a/llvm/lib/CodeGen/MachineOutliner.cpp
+++ b/llvm/lib/CodeGen/MachineOutliner.cpp
@@ -121,6 +121,12 @@ static cl::opt<unsigned> OutlinerBenefitThreshold(
cl::desc(
"The minimum size in bytes before an outlining candidate is accepted"));
+static cl::opt<bool> OutlinerLeafDescendants(
+ "outliner-leaf-descendants", cl::init(true), cl::Hidden,
+ cl::desc("Consider all leaf descendants of internal nodes of the suffix "
+ "tree as candidates for outlining (if false, only leaf children "
+ "are considered)"));
+
namespace {
/// Maps \p MachineInstrs to unsigned integers and stores the mappings.
@@ -576,7 +582,7 @@ void MachineOutliner::emitOutlinedFunctionRemark(OutlinedFunction &OF) {
void MachineOutliner::findCandidates(
InstructionMapper &Mapper, std::vector<OutlinedFunction> &FunctionList) {
FunctionList.clear();
- SuffixTree ST(Mapper.UnsignedVec);
+ SuffixTree ST(Mapper.UnsignedVec, OutlinerLeafDescendants);
// First, find all of the repeated substrings in the tree of minimum length
// 2.
diff --git a/llvm/lib/Support/SuffixTree.cpp b/llvm/lib/Support/SuffixTree.cpp
index c00c7989d1a64..e868dfe513860 100644
--- a/llvm/lib/Support/SuffixTree.cpp
+++ b/llvm/lib/Support/SuffixTree.cpp
@@ -11,9 +11,11 @@
//===----------------------------------------------------------------------===//
#include "llvm/Support/SuffixTree.h"
+#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/Support/Allocator.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/SuffixTreeNode.h"
+#include <stack>
using namespace llvm;
@@ -26,7 +28,9 @@ static size_t numElementsInSubstring(const SuffixTreeNode *N) {
return N->getEndIdx() - N->getStartIdx() + 1;
}
-SuffixTree::SuffixTree(const ArrayRef<unsigned> &Str) : Str(Str) {
+SuffixTree::SuffixTree(const ArrayRef<unsigned> &Str,
+ bool OutlinerLeafDescendants)
+ : Str(Str), OutlinerLeafDescendants(OutlinerLeafDescendants) {
Root = insertRoot();
Active.Node = Root;
@@ -46,6 +50,11 @@ SuffixTree::SuffixTree(const ArrayRef<unsigned> &Str) : Str(Str) {
// Set the suffix indices of each leaf.
assert(Root && "Root node can't be nullptr!");
setSuffixIndices();
+
+ // Collect all leaf nodes of the suffix tree. And for each internal node,
+ // record the range of leaf nodes that are descendants of it.
+ if (OutlinerLeafDescendants)
+ setLeafNodes();
}
SuffixTreeNode *SuffixTree::insertLeaf(SuffixTreeInternalNode &Parent,
@@ -105,6 +114,68 @@ void SuffixTree::setSuffixIndices() {
}
}
+void SuffixTree::setLeafNodes() {
+ // A stack that keeps track of nodes to visit for post-order DFS traversal.
+ std::stack<SuffixTreeNode *> ToVisit;
+ ToVisit.push(Root);
+
+ // This keeps track of the index of the next leaf node to be added to
+ // the LeafNodes vector of the suffix tree.
+ unsigned LeafCounter = 0;
+
+ // This keeps track of nodes whose children have been added to the stack
+ // during the post-order depth-first traversal of the tree.
+ llvm::SmallPtrSet<SuffixTreeInternalNode *, 32> ChildrenAddedToStack;
+
+ // Traverse the tree in post-order.
+ while (!ToVisit.empty()) {
+ SuffixTreeNode *CurrNode = ToVisit.top();
+ ToVisit.pop();
+ if (auto *CurrInternalNode = dyn_cast<SuffixTreeInternalNode>(CurrNode)) {
+ // The current node is an internal node.
+ if (ChildrenAddedToStack.find(CurrInternalNode) !=
+ ChildrenAddedToStack.end()) {
+ // If the children of the current node has been added to the stack,
+ // then this is the second time we visit this node and at this point,
+ // all of its children have already been processed. Now, we can
+ // set its LeftLeafIdx and RightLeafIdx;
+ auto it = CurrInternalNode->Children.begin();
+ if (it != CurrInternalNode->Children.end()) {
+ // Get the first child to use its RightLeafIdx. The RightLeafIdx is
+ // used as the first child is the initial one added to the stack, so
+ // it's the last one to be processed. This implies that the leaf
+ // descendants of the first child are assigned the largest index
+ // numbers.
+ CurrNode->setRightLeafIdx(it->second->getRightLeafIdx());
+ // get the last child to use its LeftLeafIdx.
+ while (std::next(it) != CurrInternalNode->Children.end())
+ it = std::next(it);
+ CurrNode->setLeftLeafIdx(it->second->getLeftLeafIdx());
+ assert(CurrNode->getLeftLeafIdx() <= CurrNode->getRightLeafIdx() &&
+ "LeftLeafIdx should not be larger than RightLeafIdx");
+ }
+ } else {
+ // This is the first time we visit this node. This means that its
+ // children have not been added to the stack yet. Hence, we will add
+ // the current node back to the stack and add its children to the
+ // stack for processing.
+ ToVisit.push(CurrNode);
+ for (auto &ChildPair : CurrInternalNode->Children)
+ ToVisit.push(ChildPair.second);
+ ChildrenAddedToStack.insert(CurrInternalNode);
+ }
+ } else {
+ // The current node is a leaf node.
+ // We can simplyset its LeftLeafIdx and RightLeafIdx.
+ CurrNode->setLeftLeafIdx(LeafCounter);
+ CurrNode->setRightLeafIdx(LeafCounter);
+ LeafCounter++;
+ auto *CurrLeafNode = cast<SuffixTreeLeafNode>(CurrNode);
+ LeafNodes.push_back(CurrLeafNode);
+ }
+ }
+}
+
unsigned SuffixTree::extend(unsigned EndIdx, unsigned SuffixesToAdd) {
SuffixTreeInternalNode *NeedsLink = nullptr;
@@ -230,6 +301,7 @@ void SuffixTree::RepeatedSubstringIterator::advance() {
// Each leaf node represents a repeat of a string.
SmallVector<unsigned> RepeatedSubstringStarts;
+ SmallVector<SuffixTreeLeafNode *> LeafDescendants;
// Continue visiting nodes until we find one which repeats more than once.
while (!InternalNodesToVisit.empty()) {
@@ -241,30 +313,35 @@ void SuffixTree::RepeatedSubstringIterator::advance() {
// it's too short, we'll quit.
unsigned Length = Curr->getConcatLen();
- // Iterate over each child, saving internal nodes for visiting, and
- // leaf nodes' SuffixIdx in RepeatedSubstringStarts. Internal nodes
- // represent individual strings, which may repeat.
- for (auto &ChildPair : Curr->Children) {
+ // Iterate over each child, saving internal nodes for visiting.
+ // Internal nodes represent individual strings, which may repeat.
+ for (auto &ChildPair : Curr->Children)
// Save all of this node's children for processing.
if (auto *InternalChild =
- dyn_cast<SuffixTreeInternalNode>(ChildPair.second)) {
+ dyn_cast<SuffixTreeInternalNode>(ChildPair.second))
InternalNodesToVisit.push_back(InternalChild);
- continue;
- }
-
- if (Length < MinLength)
- continue;
-
- // Have an occurrence of a potentially repeated string. Save it.
- auto *Leaf = cast<SuffixTreeLeafNode>(ChildPair.second);
- RepeatedSubstringStarts.push_back(Leaf->getSuffixIdx());
- }
+
+ // If length of repeated substring is below threshold, then skip it.
+ if (Length < MinLength)
+ continue;
// The root never represents a repeated substring. If we're looking at
// that, then skip it.
if (Curr->isRoot())
continue;
+ // Collect leaf children or leaf descendants by OutlinerLeafDescendants.
+ if (!OutlinerLeafDescendants) {
+ for (auto &ChildPair : Curr->Children)
+ if (auto *Leaf = dyn_cast<SuffixTreeLeafNode>(ChildPair.second))
+ RepeatedSubstringStarts.push_back(Leaf->getSuffixIdx());
+ } else {
+ LeafDescendants.assign(LeafNodes.begin() + Curr->getLeftLeafIdx(),
+ LeafNodes.begin() + Curr->getRightLeafIdx() + 1);
+ for (SuffixTreeLeafNode *Leaf : LeafDescendants)
+ RepeatedSubstringStarts.push_back(Leaf->getSuffixIdx());
+ }
+
// Do we have any repeated substrings?
if (RepeatedSubstringStarts.size() < 2)
continue;
diff --git a/llvm/lib/Support/SuffixTreeNode.cpp b/llvm/lib/Support/SuffixTreeNode.cpp
index 113b990fd352f..9f1f94a39895e 100644
--- a/llvm/lib/Support/SuffixTreeNode.cpp
+++ b/llvm/lib/Support/SuffixTreeNode.cpp
@@ -38,3 +38,8 @@ unsigned SuffixTreeLeafNode::getEndIdx() const {
unsigned SuffixTreeLeafNode::getSuffixIdx() const { return SuffixIdx; }
void SuffixTreeLeafNode::setSuffixIdx(unsigned Idx) { SuffixIdx = Idx; }
+
+unsigned SuffixTreeNode::getLeftLeafIdx() const { return LeftLeafIdx; }
+unsigned SuffixTreeNode::getRightLeafIdx() const { return RightLeafIdx; }
+void SuffixTreeNode::setLeftLeafIdx(unsigned Idx) { LeftLeafIdx = Idx; }
+void SuffixTreeNode::setRightLeafIdx(unsigned Idx) { RightLeafIdx = Idx; }
diff --git a/llvm/test/CodeGen/AArch64/machine-outliner-cfi-tail-some.mir b/llvm/test/CodeGen/AArch64/machine-outliner-cfi-tail-some.mir
index 67d411962ce4f..3afa1d5559a58 100644
--- a/llvm/test/CodeGen/AArch64/machine-outliner-cfi-tail-some.mir
+++ b/llvm/test/CodeGen/AArch64/machine-outliner-cfi-tail-some.mir
@@ -1,5 +1,5 @@
# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -mtriple=aarch64-apple-unknown -run-pass=machine-outliner -verify-machineinstrs %s -o - | FileCheck %s
+# RUN: llc -mtriple=aarch64-apple-unknown -run-pass=machine-outliner -verify-machineinstrs -outliner-leaf-descendants=false %s -o - | FileCheck %s
# Outlining CFI instructions is unsafe if we cannot outline all of the CFI
# instructions from a function. This shows that we choose not to outline the
diff --git a/llvm/test/CodeGen/AArch64/machine-outliner-leaf-descendants.ll b/llvm/test/CodeGen/AArch64/machine-outliner-leaf-descendants.ll
new file mode 100644
index 0000000000000..29076204e89f6
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/machine-outliner-leaf-descendants.ll
@@ -0,0 +1,124 @@
+; This test is mainly for the -outliner-leaf-descendants flag for MachineOutliner.
+;
+; ===================== -outliner-leaf-descendants=false =====================
+; MachineOutliner finds THREE key `OutlinedFunction` and outlines them. They are:
+; ```
+; mov w0, #1
+; mov w1, #2
+; mov w2, #3
+; mov w3, #4
+; mov w4, #5
+; mov w5, #6 or #7 or #8
+; b
+; ```
+; Each has:
+; - `SequenceSize=28` and `OccurrenceCount=2`
+; - each Candidate has `CallOverhead=4` and `FrameOverhead=0`
+; - `NotOutlinedCost=28*2=56` and `OutliningCost=4*2+28+0=36`
+; - `Benefit=56-36=20` and `Priority=56/36=1.56`
+;
+; ===================== -outliner-leaf-descendants=true =====================
+; MachineOutliner finds a FOURTH key `OutlinedFunction`, which is:
+; ```
+; mov w0, #1
+; mov w1, #2
+; mov w2, #3
+; mov w3, #4
+; mov w4, #5
+; ```
+; This corresponds to an internal node that has ZERO leaf children, but SIX leaf descendants.
+; It has:
+; - `SequenceSize=20` and `OccurrenceCount=6`
+; - each Candidate has `CallOverhead=12` and `FrameOverhead=4`
+; - `NotOutlinedCost=20*6=120` and `OutliningCost=12*6+20+4=96`
+; - `Benefit=120-96=24` and `Priority=120/96=1.25`
+;
+; The FOURTH `OutlinedFunction` has lower _priority_ compared to the first THREE `OutlinedFunction`
+; Hence, if we additionally include the `-sort-per-priority` flag, the first THREE `OutlinedFunction` are outlined.
+
+; RUN: llc %s -enable-machine-outliner=always -outliner-leaf-descendants=false -filetype=obj -o %t
+; RUN: llvm-objdump -d %t | FileCheck %s --check-prefix=CHECK-BASELINE
+
+; RUN: llc %s -enable-machine-outliner=always -outliner-leaf-descendants=false -outliner-benefit-threshold=22 -filetype=obj -o %t
+; RUN: llvm-objdump -d %t | FileCheck %s --check-prefix=CHECK-NO-CANDIDATE
+
+; RUN: llc %s -enable-machine-outliner=always -outliner-leaf-descendants=true -filetype=obj -o %t
+; RUN: llvm-objdump -d %t | FileCheck %s --check-prefix=CHECK-BASELINE
+
+; RUN: llc %s -enable-machine-outliner=always -outliner-leaf-descendants=true -outliner-benefit-threshold=22 -filetype=obj -o %t
+; RUN: llvm-objdump -d %t | FileCheck %s --check-prefix=CHECK-LEAF-DESCENDANTS
+
+
+target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"
+target triple = "arm64-apple-macosx14.0.0"
+
+declare i32 @_Z3fooiiii(i32 noundef, i32 noundef, i32 noundef, i32 noundef, i32 noundef, i32 noundef)
+
+define i32 @_Z2f1v() minsize {
+ %1 = tail call i32 @_Z3fooiiii(i32 noundef 1, i32 noundef 2, i32 noundef 3, i32 noundef 4, i32 noundef 5, i32 noundef 6)
+ ret i32 %1
+}
+
+define i32 @_Z2f2v() minsize {
+ %1 = tail call i32 @_Z3fooiiii(i32 noundef 1, i32 noundef 2, i32 noundef 3, i32 noundef 4, i32 noundef 5, i32 noundef 6)
+ ret i32 %1
+}
+
+define i32 @_Z2f3v() minsize {
+ %1 = tail call i32 @_Z3fooiiii(i32 noundef 1, i32 noundef 2, i32 noundef 3, i32 noundef 4, i32 noundef 5, i32 noundef 7)
+ ret i32 %1
+}
+
+define i32 @_Z2f4v() minsize {
+ %1 = tail call i32 @_Z3fooiiii(i32 noundef 1, i32 noundef 2, i32 noundef 3, i32 noundef 4, i32 noundef 5, i32 noundef 7)
+ ret i32 %1
+}
+
+define i32 @_Z2f5v() minsize {
+ %1 = tail call i32 @_Z3fooiiii(i32 noundef 1, i32 noundef 2, i32 noundef 3, i32 noundef 4, i32 noundef 5, i32 noundef 8)
+ ret i32 %1
+}
+
+define i32 @_Z2f6v() minsize {
+ %1 = tail call i32 @_Z3fooiiii(i32 noundef 1, i32 noundef 2, i32 noundef 3, i32 noundef 4, i32 noundef 5, i32 noundef 8)
+ ret i32 %1
+}
+
+; CHECK-BASELINE: <_OUTLINED_FUNCTION_0>:
+; CHECK-BASELINE-NEXT: mov w0, #0x1
+; CHECK-BASELINE-NEXT: mov w1, #0x2
+; CHECK-BASELINE-NEXT: mov w2, #0x3
+; CHECK-BASELINE-NEXT: mov w3, #0x4
+; CHECK-BASELINE-NEXT: mov w4, #0x5
+; CHECK-BASELINE-NEXT: mov w5, #0x6
+; CHECK-BASELINE-NEXT: b
+
+; CHECK-BASELINE: <_OUTLINED_FUNCTION_1>:
+; CHECK-BASELINE-NEXT: mov w0, #0x1
+; CHECK-BASELINE-NEXT: mov w1, #0x2
+; CHECK-BASELINE-NEXT: mov w2, #0x3
+; CHECK-BASELINE-NEXT: mov w3, #0x4
+; CHECK-BASELINE-NEXT: mov w4, #0x5
+; CHECK-BASELINE-NEXT: mov w5, #0x8
+; CHECK-BASELINE-NEXT: b
+
+; CHECK-BASELINE: <_OUTLINED_FUNCTION_2>:
+; CHECK-BASELINE-NEXT: mov w0, #0x1
+; CHECK-BASELINE-NEXT: mov w1, #0x2
+; CHECK-BASELINE-NEXT: mov w2, #0x3
+; CHECK-BASELINE-NEXT: mov w3, #0x4
+; CHECK-BASELINE-NEXT: mov w4, #0x5
+; CHECK-BASELINE-NEXT: mov w5, #0x7
+; CHECK-BASELINE-NEXT: b
+
+; CHECK-LEAF-DESCENDANTS: <_OUTLINED_FUNCTION_0>:
+; CHECK-LEAF-DESCENDANTS-NEXT: mov w0, #0x1
+; CHECK-LEAF-DESCENDANTS-NEXT: mov w1, #0x2
+; CHECK-LEAF-DESCENDANTS-NEXT: mov w2, #0x3
+; CHECK-LEAF-DESCENDANTS-NEXT: mov w3, #0x4
+; CHECK-LEAF-DESCENDANTS-NEXT: mov w4, #0x5
+; CHECK-LEAF-DESCENDANTS-NEXT: ret
+
+; CHECK-LEAF-DESCENDANTS-NOT: <_OUTLINED_FUNCTION_1>:
+
+; CHECK-NO-CANDIDATE-NOT: <_OUTLINED_FUNCTION_0>:
diff --git a/llvm/test/CodeGen/AArch64/machine-outliner-overlap.mir b/llvm/test/CodeGen/AArch64/machine-outliner-overlap.mir
index c6bd4c1d04d87..8b0e03dbee8d8 100644
--- a/llvm/test/CodeGen/AArch64/machine-outliner-overlap.mir
+++ b/llvm/test/CodeGen/AArch64/machine-outliner-overlap.mir
@@ -1,5 +1,6 @@
# NOTE: Assertions have been autogen...
[truncated]
|
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
kyulee-com
requested review from
francisvm,
ornata,
ellishg,
plotfi,
kyulee-com and
alx32
ellishg
reviewed
Jun 7, 2024
kyulee-com
reviewed
Jun 7, 2024
kyulee-com
reviewed
Jun 8, 2024
| @@ -121,6 +121,12 @@ static cl::opt<unsigned> OutlinerBenefitThreshold( | |||
| cl::desc( | |||
| "The minimum size in bytes before an outlining candidate is accepted")); | |||
|
|
|||
| static cl::opt<bool> OutlinerLeafDescendants( | |||
| "outliner-leaf-descendants", cl::init(true), cl::Hidden, | |||
There was a problem hiding this comment.
I don't have a strong preference, but wonder if we want a disable flag instead, like -disable-outliner-leaf-descendants and we can drop cl::init(false), or completely delete a flag while updating all tests. I'm all open for either option.
kyulee-com
reviewed
Jun 14, 2024
ellishg
reviewed
Jun 15, 2024
kyulee-com
reviewed
Jun 15, 2024
kyulee-com
reviewed
Jun 15, 2024
AlexisPerry
pushed a commit
to llvm-project-tlp/llvm-project
that referenced
this pull request
Jul 9, 2024
This PR depends on llvm#90264 In the current implementation, only leaf children of each internal node in the suffix tree are included as candidates for outlining. But all leaf descendants are outlining candidates, which we include in the new implementation. This is enabled on a flag `outliner-leaf-descendants` which is default to be true. The reason for _enabling this on a flag_ is because machine outliner is not the only pass that uses suffix tree. The reason for _having this default to be true_ is because including all leaf descendants show consistent size win. * For Clang/LLD, it shows around 3% reduction in text segment size when compared to the baseline `-Oz` linker binary. * For selected benchmark tests in LLVM test suite | run (CTMark/) | only leaf children | all leaf descendants | reduction % | |------------------|--------------------|----------------------|-------------| | lencod | 349624 | 348564 | -0.2004% | | SPASS | 219672 | 218440 | -0.4738% | | kc | 271956 | 250068 | -0.4506% | | sqlite3 | 223920 | 222484 | -0.5471% | | 7zip-benchmark | 405364 | 401244 | -0.3428% | | bullet | 139820 | 138340 | -0.8315% | | consumer-typeset | 295684 | 286628 | -1.2295% | | pairlocalalign | 72236 | 71936 | -0.2164% | | tramp3d-v4 | 189572 | 183676 | -2.9668% | This is part of an enhanced version of machine outliner -- see [RFC](https://discourse.llvm.org/t/rfc-enhanced-machine-outliner-part-1-fulllto-part-2-thinlto-nolto-to-come/78732).
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR depends on #90264
In the current implementation, only leaf children of each internal node in the suffix tree are included as candidates for outlining. But all leaf descendants are outlining candidates, which we include in the new implementation. This is enabled on a flag
outliner-leaf-descendantswhich is default to be true.The reason for enabling this on a flag is because machine outliner is not the only pass that uses suffix tree.
The reason for having this default to be true is because including all leaf descendants show consistent size win.
-Ozlinker binary.This is part of an enhanced version of machine outliner -- see RFC.