[llvm] [CGData] Global Merge Functions (PR #112671)

Zhaoxuan Jiang via llvm-commits llvm-commits at lists.llvm.org
Sun Nov 17 01:25:40 PST 2024


================
@@ -0,0 +1,672 @@
+//===---- GlobalMergeFunctions.cpp - Global merge functions -------*- 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 pass implements the global merge function pass.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/CodeGen/GlobalMergeFunctions.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/Analysis/ModuleSummaryAnalysis.h"
+#include "llvm/CGData/CodeGenData.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/StructuralHash.h"
+#include "llvm/InitializePasses.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Transforms/Utils/ModuleUtils.h"
+
+#define DEBUG_TYPE "global-merge-func"
+
+using namespace llvm;
+using namespace llvm::support;
+
+static cl::opt<bool> DisableCGDataForMerging(
+    "disable-cgdata-for-merging", cl::Hidden,
+    cl::desc("Disable codegen data for function merging. Local "
+             "merging is still enabled within a module."),
+    cl::init(false));
+
+STATISTIC(NumMismatchedFunctionHash,
+          "Number of mismatched function hash for global merge function");
+STATISTIC(NumMismatchedInstCount,
+          "Number of mismatched instruction count for global merge function");
+STATISTIC(NumMismatchedConstHash,
+          "Number of mismatched const hash for global merge function");
+STATISTIC(NumMismatchedModuleId,
+          "Number of mismatched Module Id for global merge function");
+STATISTIC(NumMergedFunctions,
+          "Number of functions that are actually merged using function hash");
+STATISTIC(NumAnalyzedModues, "Number of modules that are analyzed");
+STATISTIC(NumAnalyzedFunctions, "Number of functions that are analyzed");
+STATISTIC(NumEligibleFunctions, "Number of functions that are eligible");
+
+/// Returns true if the \OpIdx operand of \p CI is the callee operand.
+static bool isCalleeOperand(const CallBase *CI, unsigned OpIdx) {
+  return &CI->getCalledOperandUse() == &CI->getOperandUse(OpIdx);
+}
+
+static bool canParameterizeCallOperand(const CallBase *CI, unsigned OpIdx) {
+  if (CI->isInlineAsm())
+    return false;
+  Function *Callee = CI->getCalledOperand()
+                         ? dyn_cast_or_null<Function>(
+                               CI->getCalledOperand()->stripPointerCasts())
+                         : nullptr;
+  if (Callee) {
+    if (Callee->isIntrinsic())
+      return false;
+    auto Name = Callee->getName();
+    // objc_msgSend stubs must be called, and can't have their address taken.
+    if (Name.starts_with("objc_msgSend$"))
+      return false;
+    // Calls to dtrace probes must generate unique patchpoints.
+    if (Name.starts_with("__dtrace"))
+      return false;
+  }
+  if (isCalleeOperand(CI, OpIdx) &&
+      CI->getOperandBundle(LLVMContext::OB_ptrauth).has_value()) {
+    // The operand is the callee and it has already been signed. Ignore this
+    // because we cannot add another ptrauth bundle to the call instruction.
+    return false;
+  }
+  return true;
+}
+
+/// Returns true if function \p F is eligible for merging.
+bool isEligibleFunction(Function *F) {
+  if (F->isDeclaration())
+    return false;
+
+  if (F->hasFnAttribute(llvm::Attribute::NoMerge) ||
+      F->hasFnAttribute(llvm::Attribute::AlwaysInline))
+    return false;
+
+  if (F->hasAvailableExternallyLinkage())
+    return false;
+
+  if (F->getFunctionType()->isVarArg())
+    return false;
+
+  if (F->getCallingConv() == CallingConv::SwiftTail)
+    return false;
+
+  // If function contains callsites with musttail, if we merge
+  // it, the merged function will have the musttail callsite, but
+  // the number of parameters can change, thus the parameter count
+  // of the callsite will mismatch with the function itself.
+  for (const BasicBlock &BB : *F) {
+    for (const Instruction &I : BB) {
+      const auto *CB = dyn_cast<CallBase>(&I);
+      if (CB && CB->isMustTailCall())
+        return false;
+    }
+  }
+
+  return true;
+}
+
+static bool isEligibleInstrunctionForConstantSharing(const Instruction *I) {
+  switch (I->getOpcode()) {
+  case Instruction::Load:
+  case Instruction::Store:
+  case Instruction::Call:
+  case Instruction::Invoke:
+    return true;
+  default:
+    return false;
+  }
+}
+
+static bool ignoreOp(const Instruction *I, unsigned OpIdx) {
+  assert(OpIdx < I->getNumOperands() && "Invalid operand index");
+
+  if (!isEligibleInstrunctionForConstantSharing(I))
+    return false;
+
+  if (!isa<Constant>(I->getOperand(OpIdx)))
+    return false;
+
+  if (const auto *CI = dyn_cast<CallBase>(I))
+    return canParameterizeCallOperand(CI, OpIdx);
+
+  return true;
+}
+
+static Value *createCast(IRBuilder<> &Builder, Value *V, Type *DestTy) {
+  Type *SrcTy = V->getType();
+  if (SrcTy->isStructTy()) {
+    assert(DestTy->isStructTy());
+    assert(SrcTy->getStructNumElements() == DestTy->getStructNumElements());
+    Value *Result = PoisonValue::get(DestTy);
+    for (unsigned int I = 0, E = SrcTy->getStructNumElements(); I < E; ++I) {
+      Value *Element =
+          createCast(Builder, Builder.CreateExtractValue(V, ArrayRef(I)),
+                     DestTy->getStructElementType(I));
+
+      Result = Builder.CreateInsertValue(Result, Element, ArrayRef(I));
+    }
+    return Result;
+  }
+  assert(!DestTy->isStructTy());
+  if (auto *SrcAT = dyn_cast<ArrayType>(SrcTy)) {
+    auto *DestAT = dyn_cast<ArrayType>(DestTy);
+    assert(DestAT);
+    assert(SrcAT->getNumElements() == DestAT->getNumElements());
+    Value *Result = UndefValue::get(DestTy);
+    for (unsigned int I = 0, E = SrcAT->getNumElements(); I < E; ++I) {
+      Value *Element =
+          createCast(Builder, Builder.CreateExtractValue(V, ArrayRef(I)),
+                     DestAT->getElementType());
+
+      Result = Builder.CreateInsertValue(Result, Element, ArrayRef(I));
+    }
+    return Result;
+  }
+  assert(!DestTy->isArrayTy());
+  if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
+    return Builder.CreateIntToPtr(V, DestTy);
+  if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
+    return Builder.CreatePtrToInt(V, DestTy);
+  return Builder.CreateBitCast(V, DestTy);
+}
+
+void GlobalMergeFunc::analyze(Module &M) {
+  ++NumAnalyzedModues;
+  for (Function &Func : M) {
+    ++NumAnalyzedFunctions;
+    if (isEligibleFunction(&Func)) {
+      ++NumEligibleFunctions;
+
+      auto FI = llvm::StructuralHashWithDifferences(Func, ignoreOp);
+
+      // Convert the operand map to a vector for a serialization-friendly
+      // format.
+      IndexOperandHashVecType IndexOperandHashes;
----------------
nocchijiang wrote:

Should we bail out for functions without parameterizable constants here? 

https://github.com/llvm/llvm-project/pull/112671


More information about the llvm-commits mailing list