[llvm] [BOLT] Flow conservation scores (PR #127954)

[Amir Ayupov via llvm-commits]

================
@@ -0,0 +1,579 @@
+//===- bolt/Passes/ProfileQualityStats.cpp ----------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements the profile quality stats calculation pass.
+//
+//===----------------------------------------------------------------------===//
+
+#include "bolt/Passes/ProfileQualityStats.h"
+#include "bolt/Core/BinaryBasicBlock.h"
+#include "bolt/Core/BinaryFunction.h"
+#include "bolt/Utils/CommandLineOpts.h"
+#include "llvm/Support/CommandLine.h"
+#include <queue>
+#include <unordered_map>
+#include <unordered_set>
+
+#define DEBUG_TYPE "bolt-opts"
+
+using namespace llvm;
+using namespace bolt;
+
+namespace opts {
+extern cl::opt<unsigned> Verbosity;
+cl::opt<unsigned> NumFunctionsForProfileQualityCheck(
+    "num-functions-for-profile-quality-check",
+    cl::desc("number of hottest functions to print aggregated "
+             "profile quality stats of."),
+    cl::init(1000), cl::ZeroOrMore, cl::Hidden, cl::cat(BoltOptCategory));
+cl::opt<unsigned> PercentileForProfileQualityCheck(
+    "percentile-for-profile-quality-check",
+    cl::desc("Percentile of profile quality distributions over hottest "
+             "functions to display."),
+    cl::init(95), cl::ZeroOrMore, cl::Hidden, cl::cat(BoltOptCategory));
+} // namespace opts
+
+namespace {
+using FunctionListType = std::vector<const BinaryFunction *>;
+using function_iterator = FunctionListType::iterator;
+
+// BB index -> flow count
+using FlowMapTy = std::unordered_map<unsigned, uint64_t>;
+// Function number -> FlowMapTy
+using TotalFlowMapTy = std::unordered_map<uint64_t, FlowMapTy>;
+// Function number -> flow count
+using FunctionFlowMapTy = std::unordered_map<uint64_t, uint64_t>;
+struct FlowInfo {
+  TotalFlowMapTy TotalIncomingMaps;
+  TotalFlowMapTy TotalOutgoingMaps;
+  TotalFlowMapTy TotalMaxCountMaps;
+  TotalFlowMapTy TotalMinCountMaps;
+  FunctionFlowMapTy CallGraphIncomingMap;
+};
+
+template <typename T>
+void printDistribution(raw_ostream &OS, std::vector<T> &values,
+                       bool Fraction = false) {
+  // Assume values are sorted.
+  if (values.empty())
+    return;
+
+  OS << "  Length     : " << values.size() << "\n";
+
+  auto printLine = [&](std::string Text, double Percent) {
+    int Rank = int(values.size() * (100 - Percent) / 100);
+    if (Percent == 0)
+      Rank = values.size() - 1;
+    if (Fraction)
+      OS << "  " << Text << std::string(11 - Text.length(), ' ') << ": "
+         << format("%.2lf%%", values[Rank] * 100) << "\n";
+    else
+      OS << "  " << Text << std::string(11 - Text.length(), ' ') << ": "
+         << values[Rank] << "\n";
+  };
+
+  printLine("MAX", 0);
+  const int percentages[] = {1, 5, 10, 20, 50, 80};
+  for (size_t i = 0; i < sizeof(percentages) / sizeof(percentages[0]); ++i) {
+    printLine("TOP " + std::to_string(percentages[i]) + "%", percentages[i]);
+  }
+  printLine("MIN", 100);
+}
+
+void printCFGContinuityStats(raw_ostream &OS,
+                             iterator_range<function_iterator> &Functions) {
+  // Given a perfect profile, every positive-execution-count BB should be
+  // connected to an entry of the function through a positive-execution-count
+  // directed path in the control flow graph.
+  std::vector<size_t> NumUnreachables;
+  std::vector<size_t> SumECUnreachables;
+  std::vector<double> FractionECUnreachables;
+
+  for (auto it = Functions.begin(); it != Functions.end(); ++it) {
+    const BinaryFunction *Function = *it;
+    if (Function->size() <= 1)
+      continue;
+
+    // Compute the sum of all BB execution counts (ECs).
+    size_t NumPosECBBs = 0;
+    size_t SumAllBBEC = 0;
+    for (const BinaryBasicBlock &BB : *Function) {
+      const size_t BBEC = BB.getKnownExecutionCount();
+      NumPosECBBs += BBEC > 0 ? 1 : 0;
+      SumAllBBEC += BBEC;
+    }
+
+    // Perform BFS on subgraph of CFG induced by positive weight edges.
+    // Compute the number of BBs reachable from the entry(s) of the function and
+    // the sum of their execution counts (ECs).
+    std::unordered_map<unsigned, const BinaryBasicBlock *> IndexToBB;
+    std::unordered_set<unsigned> Visited;
+    std::queue<unsigned> Queue;
+    for (const BinaryBasicBlock &BB : *Function) {
+      // Make sure BB.getIndex() is not already in IndexToBB.
+      assert(IndexToBB.find(BB.getIndex()) == IndexToBB.end());
+      IndexToBB[BB.getIndex()] = &BB;
+      if (BB.isEntryPoint() && BB.getKnownExecutionCount() > 0) {
+        Queue.push(BB.getIndex());
+        Visited.insert(BB.getIndex());
+      }
+    }
+    while (!Queue.empty()) {
+      const unsigned BBIndex = Queue.front();
+      const BinaryBasicBlock *BB = IndexToBB[BBIndex];
+      Queue.pop();
+      auto SuccBIIter = BB->branch_info_begin();
+      for (const BinaryBasicBlock *Succ : BB->successors()) {
+        const uint64_t Count = SuccBIIter->Count;
+        if (Count == BinaryBasicBlock::COUNT_NO_PROFILE || Count == 0) {
+          ++SuccBIIter;
+          continue;
+        }
+        if (!Visited.insert(Succ->getIndex()).second) {
+          ++SuccBIIter;
+          continue;
+        }
+        Queue.push(Succ->getIndex());
+        ++SuccBIIter;
+      }
+    }
+
+    const size_t NumReachableBBs = Visited.size();
+
+    // Loop through Visited, and sum the corresponding BBs' execution counts
+    // (ECs).
+    size_t SumReachableBBEC = 0;
----------------
aaupov wrote:

Can compute during traversal

https://github.com/llvm/llvm-project/pull/127954