lib/Target/AMDGPU/AMDGPUInline.cpp

   1 //===- AMDGPUInline.cpp - Code to perform simple function inlining --------===//
   2 //
   3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   4 // See https://llvm.org/LICENSE.txt for license information.
   5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8 //
   9 /// \file
  10 /// This is AMDGPU specific replacement of the standard inliner.
  11 /// The main purpose is to account for the fact that calls not only expensive
  12 /// on the AMDGPU, but much more expensive if a private memory pointer is
  13 /// passed to a function as an argument. In this situation, we are unable to
  14 /// eliminate private memory in the caller unless inlined and end up with slow
  15 /// and expensive scratch access. Thus, we boost the inline threshold for such
  16 /// functions here.
  17 ///
  18 //===----------------------------------------------------------------------===//
  19
  20
  21 #include "AMDGPU.h"
  22 #include "llvm/Transforms/IPO.h"
  23 #include "llvm/Analysis/AssumptionCache.h"
  24 #include "llvm/Analysis/CallGraph.h"
  25 #include "llvm/Analysis/InlineCost.h"
  26 #include "llvm/Analysis/ValueTracking.h"
  27 #include "llvm/Analysis/TargetTransformInfo.h"
  28 #include "llvm/IR/CallSite.h"
  29 #include "llvm/IR/DataLayout.h"
  30 #include "llvm/IR/Instructions.h"
  31 #include "llvm/IR/Module.h"
  32 #include "llvm/IR/Type.h"
  33 #include "llvm/Support/CommandLine.h"
  34 #include "llvm/Support/Debug.h"
  35 #include "llvm/Transforms/IPO/Inliner.h"
  36
  37 using namespace llvm;
  38
  39 #define DEBUG_TYPE "inline"
  40
  41 static cl::opt<int>
  42 ArgAllocaCost("amdgpu-inline-arg-alloca-cost", cl::Hidden, cl::init(1500),
  43               cl::desc("Cost of alloca argument"));
  44
  45 // If the amount of scratch memory to eliminate exceeds our ability to allocate
  46 // it into registers we gain nothing by aggressively inlining functions for that
  47 // heuristic.
  48 static cl::opt<unsigned>
  49 ArgAllocaCutoff("amdgpu-inline-arg-alloca-cutoff", cl::Hidden, cl::init(256),
  50                 cl::desc("Maximum alloca size to use for inline cost"));
  51
  52 // Inliner constraint to achieve reasonable compilation time
  53 static cl::opt<size_t>
  54 MaxBB("amdgpu-inline-max-bb", cl::Hidden, cl::init(300),
  55       cl::desc("Maximum BB number allowed in a function after inlining"
  56                " (compile time constraint)"));
  57
  58 namespace {
  59
  60 class AMDGPUInliner : public LegacyInlinerBase {
  61
  62 public:
  63   AMDGPUInliner() : LegacyInlinerBase(ID) {
  64     initializeAMDGPUInlinerPass(*PassRegistry::getPassRegistry());
  65     Params = getInlineParams();
  66   }
  67
  68   static char ID; // Pass identification, replacement for typeid
  69
  70   unsigned getInlineThreshold(CallSite CS) const;
  71
  72   InlineCost getInlineCost(CallSite CS) override;
  73
  74   bool runOnSCC(CallGraphSCC &SCC) override;
  75
  76   void getAnalysisUsage(AnalysisUsage &AU) const override;
  77
  78 private:
  79   TargetTransformInfoWrapperPass *TTIWP;
  80
  81   InlineParams Params;
  82 };
  83
  84 } // end anonymous namespace
  85
  86 char AMDGPUInliner::ID = 0;
  87 INITIALIZE_PASS_BEGIN(AMDGPUInliner, "amdgpu-inline",
  88                 "AMDGPU Function Integration/Inlining", false, false)
  89 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
  90 INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)
  91 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
  92 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
  93 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
  94 INITIALIZE_PASS_END(AMDGPUInliner, "amdgpu-inline",
  95                 "AMDGPU Function Integration/Inlining", false, false)
  96
  97 Pass *llvm::createAMDGPUFunctionInliningPass() { return new AMDGPUInliner(); }
  98
  99 bool AMDGPUInliner::runOnSCC(CallGraphSCC &SCC) {
 100   TTIWP = &getAnalysis<TargetTransformInfoWrapperPass>();
 101   return LegacyInlinerBase::runOnSCC(SCC);
 102 }
 103
 104 void AMDGPUInliner::getAnalysisUsage(AnalysisUsage &AU) const {
 105   AU.addRequired<TargetTransformInfoWrapperPass>();
 106   LegacyInlinerBase::getAnalysisUsage(AU);
 107 }
 108
 109 unsigned AMDGPUInliner::getInlineThreshold(CallSite CS) const {
 110   int Thres = Params.DefaultThreshold;
 111
 112   Function *Caller = CS.getCaller();
 113   // Listen to the inlinehint attribute when it would increase the threshold
 114   // and the caller does not need to minimize its size.
 115   Function *Callee = CS.getCalledFunction();
 116   bool InlineHint = Callee && !Callee->isDeclaration() &&
 117     Callee->hasFnAttribute(Attribute::InlineHint);
 118   if (InlineHint && Params.HintThreshold && Params.HintThreshold > Thres
 119       && !Caller->hasFnAttribute(Attribute::MinSize))
 120     Thres = Params.HintThreshold.getValue() *
 121             TTIWP->getTTI(*Callee).getInliningThresholdMultiplier();
 122
 123   const DataLayout &DL = Caller->getParent()->getDataLayout();
 124   if (!Callee)
 125     return (unsigned)Thres;
 126
 127   // If we have a pointer to private array passed into a function
 128   // it will not be optimized out, leaving scratch usage.
 129   // Increase the inline threshold to allow inliniting in this case.
 130   uint64_t AllocaSize = 0;
 131   SmallPtrSet<const AllocaInst *, 8> AIVisited;
 132   for (Value *PtrArg : CS.args()) {
 133     PointerType *Ty = dyn_cast<PointerType>(PtrArg->getType());
 134     if (!Ty || (Ty->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS &&
 135                 Ty->getAddressSpace() != AMDGPUAS::FLAT_ADDRESS))
 136       continue;
 137
 138     PtrArg = GetUnderlyingObject(PtrArg, DL);
 139     if (const AllocaInst *AI = dyn_cast<AllocaInst>(PtrArg)) {
 140       if (!AI->isStaticAlloca() || !AIVisited.insert(AI).second)
 141         continue;
 142       AllocaSize += DL.getTypeAllocSize(AI->getAllocatedType());
 143       // If the amount of stack memory is excessive we will not be able
 144       // to get rid of the scratch anyway, bail out.
 145       if (AllocaSize > ArgAllocaCutoff) {
 146         AllocaSize = 0;
 147         break;
 148       }
 149     }
 150   }
 151   if (AllocaSize)
 152     Thres += ArgAllocaCost;
 153
 154   return (unsigned)Thres;
 155 }
 156
 157 // Check if call is just a wrapper around another call.
 158 // In this case we only have call and ret instructions.
 159 static bool isWrapperOnlyCall(CallSite CS) {
 160   Function *Callee = CS.getCalledFunction();
 161   if (!Callee || Callee->size() != 1)
 162     return false;
 163   const BasicBlock &BB = Callee->getEntryBlock();
 164   if (const Instruction *I = BB.getFirstNonPHI()) {
 165     if (!isa<CallInst>(I)) {
 166       return false;
 167     }
 168     if (isa<ReturnInst>(*std::next(I->getIterator()))) {
 169       LLVM_DEBUG(dbgs() << "    Wrapper only call detected: "
 170                         << Callee->getName() << '\n');
 171       return true;
 172     }
 173   }
 174   return false;
 175 }
 176
 177 InlineCost AMDGPUInliner::getInlineCost(CallSite CS) {
 178   Function *Callee = CS.getCalledFunction();
 179   Function *Caller = CS.getCaller();
 180
 181   if (!Callee || Callee->isDeclaration())
 182     return llvm::InlineCost::getNever("undefined callee");
 183
 184   if (CS.isNoInline())
 185     return llvm::InlineCost::getNever("noinline");
 186
 187   TargetTransformInfo &TTI = TTIWP->getTTI(*Callee);
 188   if (!TTI.areInlineCompatible(Caller, Callee))
 189     return llvm::InlineCost::getNever("incompatible");
 190
 191   if (CS.hasFnAttr(Attribute::AlwaysInline)) {
 192     auto IsViable = isInlineViable(*Callee);
 193     if (IsViable)
 194       return llvm::InlineCost::getAlways("alwaysinline viable");
 195     return llvm::InlineCost::getNever(IsViable.message);
 196   }
 197
 198   if (isWrapperOnlyCall(CS))
 199     return llvm::InlineCost::getAlways("wrapper-only call");
 200
 201   InlineParams LocalParams = Params;
 202   LocalParams.DefaultThreshold = (int)getInlineThreshold(CS);
 203   bool RemarksEnabled = false;
 204   const auto &BBs = Caller->getBasicBlockList();
 205   if (!BBs.empty()) {
 206     auto DI = OptimizationRemark(DEBUG_TYPE, "", DebugLoc(), &BBs.front());
 207     if (DI.isEnabled())
 208       RemarksEnabled = true;
 209   }
 210
 211   OptimizationRemarkEmitter ORE(Caller);
 212   std::function<AssumptionCache &(Function &)> GetAssumptionCache =
 213       [this](Function &F) -> AssumptionCache & {
 214     return ACT->getAssumptionCache(F);
 215   };
 216
 217   auto IC = llvm::getInlineCost(cast<CallBase>(*CS.getInstruction()), Callee,
 218                              LocalParams, TTI, GetAssumptionCache, None, PSI,
 219                              RemarksEnabled ? &ORE : nullptr);
 220
 221   if (IC && !IC.isAlways() && !Callee->hasFnAttribute(Attribute::InlineHint)) {
 222     // Single BB does not increase total BB amount, thus subtract 1
 223     size_t Size = Caller->size() + Callee->size() - 1;
 224     if (MaxBB && Size > MaxBB)
 225       return llvm::InlineCost::getNever("max number of bb exceeded");
 226   }
 227   return IC;
 228 }