Skip to content

[MLInliner] Handle CGSCC changes from #94815 #96274

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 6 commits into from
Jul 3, 2024
Merged

Conversation

aeubanks
Copy link
Contributor

With #94815, the nodes belonging to dead functions are no longer invalidated, but kept around to batch delete at the end of the call graph walk.

The ML inliner needs to be updated to handle this. This fixes some asserts getting hit, e.g. https://crbug.com/348376263.

With llvm#94815, the nodes belonging to dead functions are no longer invalidated, but kept around to batch delete at the end of the call graph walk.

The ML inliner needs to be updated to handle this. This fixes some asserts getting hit, e.g. https://crbug.com/348376263.
@aeubanks aeubanks requested a review from mtrofin June 21, 2024 04:27
@llvmbot llvmbot added mlgo llvm:analysis Includes value tracking, cost tables and constant folding llvm:transforms labels Jun 21, 2024
@aeubanks
Copy link
Contributor Author

Still needs tests, installing tflite seems annoying

@llvmbot
Copy link
Member

llvmbot commented Jun 21, 2024

@llvm/pr-subscribers-mlgo

@llvm/pr-subscribers-llvm-transforms

Author: Arthur Eubanks (aeubanks)

Changes

With #94815, the nodes belonging to dead functions are no longer invalidated, but kept around to batch delete at the end of the call graph walk.

The ML inliner needs to be updated to handle this. This fixes some asserts getting hit, e.g. https://crbug.com/348376263.


Full diff: https://github.com/llvm/llvm-project/pull/96274.diff

2 Files Affected:

  • (modified) llvm/lib/Analysis/MLInlineAdvisor.cpp (+21-22)
  • (modified) llvm/lib/Transforms/IPO/Inliner.cpp (+2-2)
diff --git a/llvm/lib/Analysis/MLInlineAdvisor.cpp b/llvm/lib/Analysis/MLInlineAdvisor.cpp
index 21946572339b9..57c7801053555 100644
--- a/llvm/lib/Analysis/MLInlineAdvisor.cpp
+++ b/llvm/lib/Analysis/MLInlineAdvisor.cpp
@@ -187,8 +187,8 @@ unsigned MLInlineAdvisor::getInitialFunctionLevel(const Function &F) const {
   return CG.lookup(F) ? FunctionLevels.at(CG.lookup(F)) : 0;
 }
 
-void MLInlineAdvisor::onPassEntry(LazyCallGraph::SCC *LastSCC) {
-  if (!LastSCC || ForceStop)
+void MLInlineAdvisor::onPassEntry(LazyCallGraph::SCC *CurSCC) {
+  if (!CurSCC || ForceStop)
     return;
   FPICache.clear();
   // Function passes executed between InlinerPass runs may have changed the
@@ -206,16 +206,10 @@ void MLInlineAdvisor::onPassEntry(LazyCallGraph::SCC *LastSCC) {
   // care about the nature of the Edge (call or ref). `FunctionLevels`-wise, we
   // record them at the same level as the original node (this is a choice, may
   // need revisiting).
-  NodeCount -= static_cast<int64_t>(NodesInLastSCC.size());
   while (!NodesInLastSCC.empty()) {
     const auto *N = *NodesInLastSCC.begin();
+    assert(!N->isDead());
     NodesInLastSCC.erase(N);
-    // The Function wrapped by N could have been deleted since we last saw it.
-    if (N->isDead()) {
-      assert(!N->getFunction().isDeclaration());
-      continue;
-    }
-    ++NodeCount;
     EdgeCount += getLocalCalls(N->getFunction());
     const auto NLevel = FunctionLevels.at(N);
     for (const auto &E : *(*N)) {
@@ -223,6 +217,7 @@ void MLInlineAdvisor::onPassEntry(LazyCallGraph::SCC *LastSCC) {
       assert(!AdjNode->isDead() && !AdjNode->getFunction().isDeclaration());
       auto I = AllNodes.insert(AdjNode);
       if (I.second) {
+        ++NodeCount;
         NodesInLastSCC.insert(AdjNode);
         FunctionLevels[AdjNode] = NLevel;
       }
@@ -235,15 +230,15 @@ void MLInlineAdvisor::onPassEntry(LazyCallGraph::SCC *LastSCC) {
   // (Re)use NodesInLastSCC to remember the nodes in the SCC right now,
   // in case the SCC is split before onPassExit and some nodes are split out
   assert(NodesInLastSCC.empty());
-  for (const auto &N : *LastSCC)
+  for (const auto &N : *CurSCC)
     NodesInLastSCC.insert(&N);
 }
 
-void MLInlineAdvisor::onPassExit(LazyCallGraph::SCC *LastSCC) {
+void MLInlineAdvisor::onPassExit(LazyCallGraph::SCC *CurSCC) {
   // No need to keep this around - function passes will invalidate it.
   if (!KeepFPICache)
     FPICache.clear();
-  if (!LastSCC || ForceStop)
+  if (!CurSCC || ForceStop)
     return;
   // Keep track of the nodes and edges we last saw. Then, in onPassEntry,
   // we update the node count and edge count from the subset of these nodes that
@@ -251,15 +246,13 @@ void MLInlineAdvisor::onPassExit(LazyCallGraph::SCC *LastSCC) {
   EdgesOfLastSeenNodes = 0;
 
   // Check on nodes that were in SCC onPassEntry
-  for (auto I = NodesInLastSCC.begin(); I != NodesInLastSCC.end();) {
-    if ((*I)->isDead())
-      NodesInLastSCC.erase(*I++);
-    else
-      EdgesOfLastSeenNodes += getLocalCalls((*I++)->getFunction());
+  for (const LazyCallGraph::Node *N : NodesInLastSCC) {
+    assert(!N->isDead());
+    EdgesOfLastSeenNodes += getLocalCalls(N->getFunction());
   }
 
   // Check on nodes that may have got added to SCC
-  for (const auto &N : *LastSCC) {
+  for (const auto &N : *CurSCC) {
     assert(!N.isDead());
     auto I = NodesInLastSCC.insert(&N);
     if (I.second)
@@ -306,11 +299,18 @@ void MLInlineAdvisor::onSuccessfulInlining(const MLInlineAdvice &Advice,
   int64_t NewCallerAndCalleeEdges =
       getCachedFPI(*Caller).DirectCallsToDefinedFunctions;
 
-  if (CalleeWasDeleted)
+  // A dead function's node is not actually removed from the call graph until
+  // the end of the call graph walk, but the node no longer belongs to any valid
+  // SCC.
+  if (CalleeWasDeleted) {
     --NodeCount;
-  else
+    LazyCallGraph::Node *CalleeN = CG.lookup(*Callee);
+    NodesInLastSCC.erase(CalleeN);
+    AllNodes.erase(CalleeN);
+  } else {
     NewCallerAndCalleeEdges +=
         getCachedFPI(*Callee).DirectCallsToDefinedFunctions;
+  }
   EdgeCount += (NewCallerAndCalleeEdges - Advice.CallerAndCalleeEdges);
   assert(CurrentIRSize >= 0 && EdgeCount >= 0 && NodeCount >= 0);
 }
@@ -484,8 +484,7 @@ void MLInlineAdvisor::print(raw_ostream &OS) const {
   OS << "\n";
   OS << "[MLInlineAdvisor] FuncLevels:\n";
   for (auto I : FunctionLevels)
-    OS << (I.first->isDead() ? "<deleted>" : I.first->getFunction().getName())
-       << " : " << I.second << "\n";
+    OS << I.first->getFunction().getName() << " : " << I.second << "\n";
 
   OS << "\n";
 }
diff --git a/llvm/lib/Transforms/IPO/Inliner.cpp b/llvm/lib/Transforms/IPO/Inliner.cpp
index 1a7b9bc8e3e77..6f0146142d494 100644
--- a/llvm/lib/Transforms/IPO/Inliner.cpp
+++ b/llvm/lib/Transforms/IPO/Inliner.cpp
@@ -223,8 +223,6 @@ PreservedAnalyses InlinerPass::run(LazyCallGraph::SCC &InitialC,
   InlineAdvisor &Advisor = getAdvisor(MAMProxy, FAM, M);
   Advisor.onPassEntry(&InitialC);
 
-  auto AdvisorOnExit = make_scope_exit([&] { Advisor.onPassExit(&InitialC); });
-
   // We use a single common worklist for calls across the entire SCC. We
   // process these in-order and append new calls introduced during inlining to
   // the end. The PriorityInlineOrder is optional here, in which the smaller
@@ -564,6 +562,8 @@ PreservedAnalyses InlinerPass::run(LazyCallGraph::SCC &InitialC,
     ++NumDeleted;
   }
 
+  Advisor.onPassExit(C);
+
   if (!Changed)
     return PreservedAnalyses::all();
 

@dtcxzyw dtcxzyw requested review from nikic and alinas June 21, 2024 11:44
while (!NodesInLastSCC.empty()) {
const auto *N = *NodesInLastSCC.begin();
assert(!N->isDead());
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

could you add a note in the comment blurb above the while, to the "The cgscc pass manager rules", saying "Nodes are never deleted"? (IIUC that's now the case, right?)

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

added comment

@aeubanks aeubanks changed the title [Inliner] Handle CGSCC changes from #94815 [MLInliner] Handle CGSCC changes from #94815 Jun 24, 2024
@aeubanks
Copy link
Contributor Author

aeubanks commented Jul 2, 2024

I've split out a bunch of the changes into other PRs that have been merge, PTAL again

@aeubanks aeubanks merged commit 94471e6 into llvm:main Jul 3, 2024
7 checks passed
@aeubanks aeubanks deleted the mlinliner branch July 3, 2024 17:14
kbluck pushed a commit to kbluck/llvm-project that referenced this pull request Jul 6, 2024
With llvm#94815, the nodes belonging to dead functions are no longer
invalidated, but kept around to batch delete at the end of the call
graph walk.

The ML inliner needs to be updated to handle this. This fixes some
asserts getting hit, e.g. https://crbug.com/348376263.
anton-bannykh pushed a commit to Kotlin/llvm-project that referenced this pull request Aug 23, 2024
With llvm#94815, the nodes belonging to dead functions are no longer
invalidated, but kept around to batch delete at the end of the call
graph walk.

The ML inliner needs to be updated to handle this. This fixes some
asserts getting hit, e.g. https://crbug.com/348376263.

(cherry picked from commit 94471e6)
anton-bannykh pushed a commit to Kotlin/llvm-project that referenced this pull request Aug 23, 2024
With llvm#94815, the nodes belonging to dead functions are no longer
invalidated, but kept around to batch delete at the end of the call
graph walk.

The ML inliner needs to be updated to handle this. This fixes some
asserts getting hit, e.g. https://crbug.com/348376263.

(cherry picked from commit 94471e6)
anton-bannykh pushed a commit to Kotlin/llvm-project that referenced this pull request Aug 26, 2024
With llvm#94815, the nodes belonging to dead functions are no longer
invalidated, but kept around to batch delete at the end of the call
graph walk.

The ML inliner needs to be updated to handle this. This fixes some
asserts getting hit, e.g. https://crbug.com/348376263.

(cherry picked from commit 94471e6)
anton-bannykh pushed a commit to Kotlin/llvm-project that referenced this pull request Aug 26, 2024
With llvm#94815, the nodes belonging to dead functions are no longer
invalidated, but kept around to batch delete at the end of the call
graph walk.

The ML inliner needs to be updated to handle this. This fixes some
asserts getting hit, e.g. https://crbug.com/348376263.

(cherry picked from commit 94471e6)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
llvm:analysis Includes value tracking, cost tables and constant folding llvm:transforms mlgo
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants