[DFAJumpThreading] Constraint the number of cloned instructions

[XChy]

Duplicating blocks of threaded paths may cause a significant regression in IR size and slow down compile-time in later optimizations. This patch adds a coarse constraint on the number of duplicated instructions.

New compile-time number: https://llvm-compile-time-tracker.com/compare.php?from=6935a3362628e823451baf86d15694d57f9664ed&to=455bd99ac151dd0445737f6e02210c2a4ec2dac5&stat=instructions%3Au

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Hongyu Chen (XChy)

Changes

Duplicating blocks of threaded paths may cause a significant regression in IR size and slow down compile-time in later optimizations. This patch adds a coarse constraint on the number of duplicated instructions.

New compile-time number: https://llvm-compile-time-tracker.com/compare.php?from=6935a3362628e823451baf86d15694d57f9664ed&to=455bd99ac151dd0445737f6e02210c2a4ec2dac5&stat=instructions%3Au

---

Full diff: https://github.com/llvm/llvm-project/pull/161632.diff

1 Files Affected:

• (modified) llvm/lib/Transforms/Scalar/DFAJumpThreading.cpp (+25-1)

diff --git a/llvm/lib/Transforms/Scalar/DFAJumpThreading.cpp b/llvm/lib/Transforms/Scalar/DFAJumpThreading.cpp
index e9a3e983bc1e2..3bd7f8e2fcf93 100644
--- a/llvm/lib/Transforms/Scalar/DFAJumpThreading.cpp
+++ b/llvm/lib/Transforms/Scalar/DFAJumpThreading.cpp
@@ -120,6 +120,12 @@ static cl::opt<unsigned>
                   cl::desc("Maximum cost accepted for the transformation"),
                   cl::Hidden, cl::init(50));
 
+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));
+
 namespace {
 
 class SelectInstToUnfold {
@@ -828,6 +834,7 @@ struct TransformDFA {
   /// also returns false if it is illegal to clone some required block.
   bool isLegalAndProfitableToTransform() {
     CodeMetrics Metrics;
+    uint64_t NumClonedInst = 0;
     SwitchInst *Switch = SwitchPaths->getSwitchInst();
 
     // Don't thread switch without multiple successors.
@@ -837,7 +844,6 @@ struct TransformDFA {
     // Note that DuplicateBlockMap is not being used as intended here. It is
     // just being used to ensure (BB, State) pairs are only counted once.
     DuplicateBlockMap DuplicateMap;
-
     for (ThreadingPath &TPath : SwitchPaths->getThreadingPaths()) {
       PathType PathBBs = TPath.getPath();
       APInt NextState = TPath.getExitValue();
@@ -848,6 +854,7 @@ struct TransformDFA {
       BasicBlock *VisitedBB = getClonedBB(BB, NextState, DuplicateMap);
       if (!VisitedBB) {
         Metrics.analyzeBasicBlock(BB, *TTI, EphValues);
+        NumClonedInst += range_size(*BB);
         DuplicateMap[BB].push_back({BB, NextState});
       }
 
@@ -865,6 +872,7 @@ struct TransformDFA {
         if (VisitedBB)
           continue;
         Metrics.analyzeBasicBlock(BB, *TTI, EphValues);
+        NumClonedInst += range_size(*BB);
         DuplicateMap[BB].push_back({BB, NextState});
       }
 
@@ -901,6 +909,22 @@ struct TransformDFA {
       }
     }
 
+    // Too much cloned instructions slow down later optimizations, especially
+    // SLPVectorizer.
+    // TODO: Thread the switch partially before reaching the threshold.
+    uint64_t NumOrigInst = 0;
+    for (auto &[BB, _] : DuplicateMap)
+      NumOrigInst += range_size(*BB);
+    if (double(NumClonedInst) / double(NumOrigInst) > MaxClonedRate) {
+      LLVM_DEBUG(dbgs() << "DFA Jump Threading: Not jump threading, too much "
+                           "instructions wll be cloned\n");
+      ORE->emit([&]() {
+        return OptimizationRemarkMissed(DEBUG_TYPE, "NotProfitable", Switch)
+               << "Too much instructions will be cloned.";
+      });
+      return false;
+    }
+
     InstructionCost DuplicationCost = 0;
 
     unsigned JumpTableSize = 0;

[UsmanNadeem]

How many transformations are restricted by this (would it be possible to check stats on the test suite)? I fear that restricting possibly beneficial transformations to avoid second-order effects might not be the best way forward.

[XChy]

How many transformations are restricted by this (would it be possible to check stats on the test suite)?

I have checked the statistics on the test suite. Only reduce one transformation in the outlier libclamav_nsis_LZMADecode.c.

I fear that restricting possibly beneficial transformations to avoid second-order effects might not be the best way forward.

Passes like inlining and loop unrolling would prevent massive expansion; otherwise, the compile time will blow up in later optimizations, although there are actually benefits after expansion. That's an acceptable trade-off for me.

The outlier libclamav_nsis_LZMADecode.c grows from about 1K lines to over 4K lines after DFAJumpThreading, that's unacceptable to me.

[UsmanNadeem]

LGTM.