[llvm] [DFAJumpThreading] Add MaxOuterUseBlocks threshold (PR #163428)

[Hongyu Chen via llvm-commits]

https://github.com/XChy created https://github.com/llvm/llvm-project/pull/163428

For every threadable path `B1 -> B2 -> ... -> Bn`, we need to insert phi nodes into every unduplicated successor of `Bi` if there are outer uses of duplicated definitions in `B_i`. To prevent the booming of phi nodes, this patch adds a threshold for the maximum number of unduplicated successors that may contain outer uses. This threshold makes sense especially when multi-target branches like switch/indirectbr/callbr are duplicated.

Note that the O3 statistics in llvm-test-suite are not influenced.

>From 8704e9a50ad3beb894e3119d1bf0926996bf080c Mon Sep 17 00:00:00 2001
From: XChy <xxs_chy at outlook.com>
Date: Wed, 15 Oct 2025 01:30:40 +0800
Subject: [PATCH] [DFAJumpThreading] Add MaxOuterUseBlocks threshold

---
 .../Transforms/Scalar/DFAJumpThreading.cpp    | 38 ++++++++++++++++---
 1 file changed, 33 insertions(+), 5 deletions(-)

diff --git a/llvm/lib/Transforms/Scalar/DFAJumpThreading.cpp b/llvm/lib/Transforms/Scalar/DFAJumpThreading.cpp
index ff5f390d6fe18..11d65d5a59175 100644
--- a/llvm/lib/Transforms/Scalar/DFAJumpThreading.cpp
+++ b/llvm/lib/Transforms/Scalar/DFAJumpThreading.cpp
@@ -122,16 +122,22 @@ static cl::opt<unsigned>
                   cl::desc("Maximum cost accepted for the transformation"),
                   cl::Hidden, cl::init(50));
 
-extern cl::opt<bool> ProfcheckDisableMetadataFixes;
-
-} // namespace llvm
-
 static cl::opt<double> MaxClonedRate(
     "dfa-max-cloned-rate",
     cl::desc(
         "Maximum cloned instructions rate accepted for the transformation"),
     cl::Hidden, cl::init(7.5));
 
+static cl::opt<unsigned>
+    MaxOuterUseBlocks("dfa-max-out-use-blocks",
+                      cl::desc("Maximum unduplicated blocks with outer uses "
+                               "accepted for the transformation"),
+                      cl::Hidden, cl::init(40));
+
+extern cl::opt<bool> ProfcheckDisableMetadataFixes;
+
+} // namespace llvm
+
 namespace {
 class SelectInstToUnfold {
   SelectInst *SI;
@@ -966,8 +972,16 @@ struct TransformDFA {
     // SLPVectorizer.
     // TODO: Thread the switch partially before reaching the threshold.
     uint64_t NumOrigInst = 0;
-    for (auto *BB : DuplicateMap.keys())
+    uint64_t NumOuterUseBlock = 0;
+    for (auto *BB : DuplicateMap.keys()) {
       NumOrigInst += BB->sizeWithoutDebug();
+      // Only unduplicated blocks with single predecessor require new phi
+      // nodes.
+      for (auto *Succ : successors(BB))
+        if (!DuplicateMap.count(Succ) && Succ->getSinglePredecessor())
+          NumOuterUseBlock++;
+    }
+
     if (double(NumClonedInst) / double(NumOrigInst) > MaxClonedRate) {
       LLVM_DEBUG(dbgs() << "DFA Jump Threading: Not jump threading, too much "
                            "instructions wll be cloned\n");
@@ -978,6 +992,20 @@ struct TransformDFA {
       return false;
     }
 
+    // Too much unduplicated blocks with outer uses may cause too much
+    // insertions of phi nodes for duplicated definitions. TODO: Drop this
+    // threshold if we come up with another way to reduce the number of inserted
+    // phi nodes.
+    if (NumOuterUseBlock > MaxOuterUseBlocks) {
+      LLVM_DEBUG(dbgs() << "DFA Jump Threading: Not jump threading, too much "
+                           "blocks with outer uses\n");
+      ORE->emit([&]() {
+        return OptimizationRemarkMissed(DEBUG_TYPE, "NotProfitable", Switch)
+               << "Too much blocks with outer uses.";
+      });
+      return false;
+    }
+
     InstructionCost DuplicationCost = 0;
 
     unsigned JumpTableSize = 0;