[PATCH] D156916: [CodeGen] Improve speed of ComplexDeinterleaving pass

[Igor Kirillov via Phabricator via llvm-commits]

igor.kirillov created this revision.
Herald added subscribers: mgabka, hiraditya.
Herald added a project: All.
igor.kirillov requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

Cache all results of running `identifyNode`, even those that do not identify
potential complex operations. This patch prevents ComplexDeinterleaving pass
from repeatedly trying to identify Nodes for the same pair of instructions.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D156916

Files:
  llvm/lib/CodeGen/ComplexDeinterleavingPass.cpp


Index: llvm/lib/CodeGen/ComplexDeinterleavingPass.cpp
===================================================================
--- llvm/lib/CodeGen/ComplexDeinterleavingPass.cpp
+++ llvm/lib/CodeGen/ComplexDeinterleavingPass.cpp
@@ -226,6 +226,7 @@
   const TargetLowering *TL = nullptr;
   const TargetLibraryInfo *TLI = nullptr;
   SmallVector<NodePtr> CompositeNodes;
+  DenseMap<std::pair<Value *, Value *>, NodePtr> CachedResult;
 
   SmallPtrSet<Instruction *, 16> FinalInstructions;
 
@@ -292,17 +293,11 @@
 
   NodePtr submitCompositeNode(NodePtr Node) {
     CompositeNodes.push_back(Node);
+    if (Node->Real && Node->Imag)
+      CachedResult[{Node->Real, Node->Imag}] = Node;
     return Node;
   }
 
-  NodePtr getContainingComposite(Value *R, Value *I) {
-    for (const auto &CN : CompositeNodes) {
-      if (CN->Real == R && CN->Imag == I)
-        return CN;
-    }
-    return nullptr;
-  }
-
   /// Identifies a complex partial multiply pattern and its rotation, based on
   /// the following patterns
   ///
@@ -900,9 +895,11 @@
   LLVM_DEBUG(dbgs() << "identifyNode on " << *R << " / " << *I << "\n");
   assert(R->getType() == I->getType() &&
          "Real and imaginary parts should not have different types");
-  if (NodePtr CN = getContainingComposite(R, I)) {
+
+  auto It = CachedResult.find({R, I});
+  if (It != CachedResult.end()) {
     LLVM_DEBUG(dbgs() << " - Folding to existing node\n");
-    return CN;
+    return It->second;
   }
 
   if (NodePtr CN = identifySplat(R, I))
@@ -949,6 +946,7 @@
     return CN;
 
   LLVM_DEBUG(dbgs() << "  - Not recognised as a valid pattern.\n");
+  CachedResult[{R, I}] = nullptr;
   return nullptr;
 }
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D156916.546530.patch
Type: text/x-patch
Size: 1676 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230802/76ac5b58/attachment.bin>