[llvm] [CodeGen] Introduce Static Data Splitter pass (PR #122183)

[Mingming Liu via llvm-commits]

================
@@ -0,0 +1,153 @@
+//===- StaticDataSplitter.cpp ---------------------------------------------===//
+//
+// 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 pass uses profile information to partition static data sections into
+// hot and cold ones. It begins to split jump tables based on profile, and
+// subsequent patches will handle constant pools and other module internal data.
+//
+// For the original RFC of this pass please see
+// https://discourse.llvm.org/t/rfc-profile-guided-static-data-partitioning/83744.
+
+#include "llvm/ADT/Statistic.h"
+#include "llvm/Analysis/ProfileSummaryInfo.h"
+#include "llvm/CodeGen/MBFIWrapper.h"
+#include "llvm/CodeGen/MachineBasicBlock.h"
+#include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
+#include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
+#include "llvm/CodeGen/MachineConstantPool.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineJumpTableInfo.h"
+#include "llvm/CodeGen/Passes.h"
+#include "llvm/InitializePasses.h"
+#include "llvm/Pass.h"
+#include "llvm/Support/CommandLine.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "static-data-splitter"
+
+STATISTIC(NumHotJumpTables, "Number of hot jump tables seen");
+STATISTIC(NumColdJumpTables, "Number of cold jump tables seen");
+STATISTIC(NumUnknownJumpTables,
+          "Number of jump tables with unknown hotness. Such jump tables will "
+          "be placed in the hot-suffixed section by default.");
+
+class StaticDataSplitter : public MachineFunctionPass {
+  const MachineBranchProbabilityInfo *MBPI = nullptr;
+  const MachineBlockFrequencyInfo *MBFI = nullptr;
+  const ProfileSummaryInfo *PSI = nullptr;
+
+  // Returns true iff any jump table is hot-cold categorized.
+  bool splitJumpTables(MachineFunction &MF);
+
+  // Same as above but works on functions with profile information.
+  bool splitJumpTablesWithProfiles(MachineFunction &MF,
+                                   MachineJumpTableInfo &MJTI);
+
+public:
+  static char ID;
+
+  StaticDataSplitter() : MachineFunctionPass(ID) {
+    initializeStaticDataSplitterPass(*PassRegistry::getPassRegistry());
+  }
+
+  StringRef getPassName() const override { return "Static Data Splitter"; }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    MachineFunctionPass::getAnalysisUsage(AU);
+    AU.addRequired<MachineBranchProbabilityInfoWrapperPass>();
+    AU.addRequired<MachineBlockFrequencyInfoWrapperPass>();
+    AU.addRequired<ProfileSummaryInfoWrapperPass>();
+  }
+
+  bool runOnMachineFunction(MachineFunction &MF) override;
+};
+
+bool StaticDataSplitter::runOnMachineFunction(MachineFunction &MF) {
+  MBPI = &getAnalysis<MachineBranchProbabilityInfoWrapperPass>().getMBPI();
+  MBFI = &getAnalysis<MachineBlockFrequencyInfoWrapperPass>().getMBFI();
+  PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
+
+  // Split jump tables based on profile information. Subsequent patches will
+  // handle other data types like constant pools, module-internal data, etc.
+  return splitJumpTables(MF);
+}
+
+bool StaticDataSplitter::splitJumpTablesWithProfiles(
+    MachineFunction &MF, MachineJumpTableInfo &MJTI) {
+  int NumChangedJumpTables = 0;
+  // Regard a jump table as hot by default. If the source and all of destination
+  // blocks are cold, regard the jump table as cold.
+  DataHotness Hotness = DataHotness::Hot;
+  for (const auto &MBB : MF) {
+    // IMPORTANT, `getJumpTableIndex` is a thin wrapper around per-target
+    // interface `TargetInstrInfo::getjumpTableIndex`, and only X86 implements
+    // it so far.
+    const int JTI = MBB.getJumpTableIndex();
+    // This is not a source block of jump table.
+    if (JTI == -1)
+      continue;
+
+    bool AllBlocksCold = true;
+
+    if (!PSI->isColdBlock(&MBB, MBFI))
+      AllBlocksCold = false;
+
+    for (const MachineBasicBlock *MBB : MJTI.getJumpTables()[JTI].MBBs)
----------------
mingmingl-llvm wrote:

> I will do a sanity check (with and without destination block hotness) using a PGO binary, and update the data early next week.

The TL,DR is that counting destination blocks' hotness on top of counting source block hotness on hot / cold jump table size (as shown by the first table below), but it doesn't change the cold/hot jump table size ratio of a binary too much (as shown by the second table)

Moreover, I printed the function names in which destination block changed jump table hotness for two iFDO-optimized binaries, and manually checked a couple of such function's cycle percentage out of the whole binary. They are mostly cold functions themselves. So it's fine to not account for destination block hotness.

---
Data size 

**Binary**|**hot jump table size (in bytes) before**|**hot jump table size (in bytes) after**|**cold jump table size (in bytes) before**|**cold jump table size (in bytes) after**
:-----:|:-----:|:-----:|:-----:|:-----:
binary1|298736|326496|1510720|1482960
binary2|73912|81448|1298560|1291024

**binary**|**cold / hot jump table size ratio before**|**cold / hot jump table size ratio after**
:-----:|:-----:|:-----:
binary 1|5.05704033|4.542046457
binary 2|17.56900097|15.85089873










https://github.com/llvm/llvm-project/pull/122183