IRLinker: avoid quadratic behavior #157045
Merged
Conversation
This file contains hidden or 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
Created using spr 1.3.6
|
@llvm/pr-subscribers-lto Author: Vitaly Buka (vitalybuka) Changesis_contained does linear search trough On large binaries it can take up to 30% Full diff: https://github.com/llvm/llvm-project/pull/157045.diff 1 Files Affected:
diff --git a/llvm/lib/Linker/IRMover.cpp b/llvm/lib/Linker/IRMover.cpp
index d6c15de4c4cdc..8903f6e9775d8 100644
--- a/llvm/lib/Linker/IRMover.cpp
+++ b/llvm/lib/Linker/IRMover.cpp
@@ -8,6 +8,7 @@
#include "llvm/Linker/IRMover.h"
#include "LinkDiagnosticInfo.h"
+#include "llvm/ADT/DenseSet.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallString.h"
#include "llvm/IR/AutoUpgrade.h"
@@ -290,6 +291,9 @@ class IRLinker {
Module &DstM;
std::unique_ptr<Module> SrcM;
+ // Lookup table to optimize IRMover::linkNamedMDNodes().
+ DenseMap<StringRef, DenseSet<MDNode *>> NamedMDNodes;
+
/// See IRMover::move().
IRMover::LazyCallback AddLazyFor;
@@ -1132,12 +1136,20 @@ void IRLinker::linkNamedMDNodes() {
continue;
NamedMDNode *DestNMD = DstM.getOrInsertNamedMetadata(NMD.getName());
+
+ auto &Inserted = NamedMDNodes[DestNMD->getName()];
+ if (Inserted.size() < DestNMD->getNumOperands()) {
+ // Must be the first module, copy everything from DestNMD.
+ Inserted.insert(DestNMD->operands().begin(), DestNMD->operands().end());
+ }
+
// Add Src elements into Dest node.
for (const MDNode *Op : NMD.operands()) {
MDNode *MD = Mapper.mapMDNode(*Op);
- if (!is_contained(DestNMD->operands(), MD))
+ if (Inserted.insert(MD).second)
DestNMD->addOperand(MD);
}
+ assert(Inserted.size() == DestNMD->getNumOperands());
}
}
|
There was a problem hiding this comment.
Pull Request Overview
This PR optimizes the IRLinker to avoid quadratic behavior during ThinLTO linking by replacing a linear search with a hash-based lookup. The change addresses performance issues where is_contained performs linear searches through previously inserted operands, causing up to 30% overhead on large binaries with CFI metadata.
Key changes:
- Introduces a
DenseMap<StringRef, DenseSet<MDNode *>>lookup table to track inserted operands - Replaces
is_containedlinear search with O(1) hash-based duplicate detection - Adds initialization logic to populate the lookup table from existing destination operands
arsenm
changed the title
Created using spr 1.3.6
arsenm
approved these changes
Sep 5, 2025
vitalybuka
deleted the
users/vitalybuka/spr/irlinker-avoid-quardatic-behavior
branch
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/181/builds/27321 Here is the relevant piece of the build log for the reference |
vitalybuka
added a commit
that referenced
this pull request
Sep 6, 2025
In #157045 I didn't realize that IRMover object is one for entire LTO, but IRLinker is created per module. We need one chache for combined module. I timed "IRLinker::linkNamedMDNodes" code on one of our internal binaries, not very large, but above average. Before #146020: 0.4859s After #146020: 624.4686s After #157045: 322.3493s After this patch: 0.5574s
|
We're hitting the assert when there are duplicated I'll go ahead and remove the assert for now and add a test. |
zmodem
added a commit
that referenced
this pull request
Sep 8, 2025
It was added in #157045, but will fail if there are duplicated nodes in DstM already, as illustrated by the test. See comments on the PR.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
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.
After #146020
is_containeddoes linear search troughall previously inserted operands.
On large binaries it can take up to 30% for ThinLTO linking with CFI, which
uses long lists of
cfi.functionsmetadata.