[BOLT] Drop high discrepancy profiles in matching

[shawbyoung]

Summary: Functions with high discrepancy
(measured by matched function blocks)
can be ignored with an added command line
argument for better performance.

Test Plan: Added
stale-matching-min-matched-block.test

[llvmbot]

@llvm/pr-subscribers-bolt
@llvm/pr-subscribers-pgo

@llvm/pr-subscribers-llvm-transforms

Author: shaw young (shawbyoung)

Changes

Test Plan: tbd

---

Full diff: https://github.com/llvm/llvm-project/pull/95156.diff

2 Files Affected:

• (modified) bolt/lib/Profile/StaleProfileMatching.cpp (+14-2)
• (modified) llvm/include/llvm/Transforms/Utils/SampleProfileInference.h (+2-1)

diff --git a/bolt/lib/Profile/StaleProfileMatching.cpp b/bolt/lib/Profile/StaleProfileMatching.cpp
index 5969f4b9bcbc1..41afa6b4bbb19 100644
--- a/bolt/lib/Profile/StaleProfileMatching.cpp
+++ b/bolt/lib/Profile/StaleProfileMatching.cpp
@@ -51,6 +51,12 @@ cl::opt<bool>
                       cl::desc("Infer counts from stale profile data."),
                       cl::init(false), cl::Hidden, cl::cat(BoltOptCategory));
 
+cl::opt<unsigned> MatchedProfileThreshold(
+    "matched-profile-threshold",
+    cl::desc("Percentage threshold of matched execution counts at which stale "
+             "profile inference is executed."),
+    cl::init(5), cl::Hidden, cl::cat(BoltOptCategory));
+
 cl::opt<unsigned> StaleMatchingMaxFuncSize(
     "stale-matching-max-func-size",
     cl::desc("The maximum size of a function to consider for inference."),
@@ -451,6 +457,7 @@ void matchWeightsByHashes(BinaryContext &BC,
       if (Matcher.isHighConfidenceMatch(BinHash, YamlHash)) {
         ++BC.Stats.NumMatchedBlocks;
         BC.Stats.MatchedSampleCount += YamlBB.ExecCount;
+        Func.MatchedExecCount += YamlBB.ExecCount;
         LLVM_DEBUG(dbgs() << "  exact match\n");
       } else {
         LLVM_DEBUG(dbgs() << "  loose match\n");
@@ -592,10 +599,15 @@ void preprocessUnreachableBlocks(FlowFunction &Func) {
 /// Decide if stale profile matching can be applied for a given function.
 /// Currently we skip inference for (very) large instances and for instances
 /// having "unexpected" control flow (e.g., having no sink basic blocks).
-bool canApplyInference(const FlowFunction &Func) {
+bool canApplyInference(const FlowFunction &Func,
+                       const yaml::bolt::BinaryFunctionProfile &YamlBF) {
   if (Func.Blocks.size() > opts::StaleMatchingMaxFuncSize)
     return false;
 
+  if (Func.MatchedExecCount / YamlBF.ExecCount >=
+      opts::MatchedProfileThreshold / 100)
+    return false;
+
   bool HasExitBlocks = llvm::any_of(
       Func.Blocks, [&](const FlowBlock &Block) { return Block.isExit(); });
   if (!HasExitBlocks)
@@ -756,7 +768,7 @@ bool YAMLProfileReader::inferStaleProfile(
   preprocessUnreachableBlocks(Func);
 
   // Check if profile inference can be applied for the instance.
-  if (!canApplyInference(Func))
+  if (!canApplyInference(Func, YamlBF))
     return false;
 
   // Apply the profile inference algorithm.
diff --git a/llvm/include/llvm/Transforms/Utils/SampleProfileInference.h b/llvm/include/llvm/Transforms/Utils/SampleProfileInference.h
index 70a80b91405d0..c654715c0ae9f 100644
--- a/llvm/include/llvm/Transforms/Utils/SampleProfileInference.h
+++ b/llvm/include/llvm/Transforms/Utils/SampleProfileInference.h
@@ -58,7 +58,8 @@ struct FlowFunction {
   std::vector<FlowJump> Jumps;
   /// The index of the entry block.
   uint64_t Entry{0};
-  uint64_t Sink{UINT64_MAX};
+  // Matched execution count for the function.
+  uint64_t MatchedExecCount{0};
 };
 
 /// Various thresholds and options controlling the behavior of the profile

[WenleiHe]

cc @wlei-llvm

[WenleiHe]

lgtm with a nit, thanks.

[aaupov]

LGTM with one nit