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(2200),
  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 namespace {
  53
  54 class AMDGPUInliner : public LegacyInlinerBase {
  55
  56 public:
  57   AMDGPUInliner() : LegacyInlinerBase(ID) {
  58     initializeAMDGPUInlinerPass(*PassRegistry::getPassRegistry());
  59     Params = getInlineParams();
  60   }
  61
  62   static char ID; // Pass identification, replacement for typeid
  63
  64   unsigned getInlineThreshold(CallSite CS) const;
  65
  66   InlineCost getInlineCost(CallSite CS) override;
  67
  68   bool runOnSCC(CallGraphSCC &SCC) override;
  69
  70   void getAnalysisUsage(AnalysisUsage &AU) const override;
  71
  72 private:
  73   TargetTransformInfoWrapperPass *TTIWP;
  74
  75   InlineParams Params;
  76 };
  77
  78 } // end anonymous namespace
  79
  80 char AMDGPUInliner::ID = 0;
  81 INITIALIZE_PASS_BEGIN(AMDGPUInliner, "amdgpu-inline",
  82                 "AMDGPU Function Integration/Inlining", false, false)
  83 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
  84 INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)
  85 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
  86 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
  87 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
  88 INITIALIZE_PASS_END(AMDGPUInliner, "amdgpu-inline",
  89                 "AMDGPU Function Integration/Inlining", false, false)
  90
  91 Pass *llvm::createAMDGPUFunctionInliningPass() { return new AMDGPUInliner(); }
  92
  93 bool AMDGPUInliner::runOnSCC(CallGraphSCC &SCC) {
  94   TTIWP = &getAnalysis<TargetTransformInfoWrapperPass>();
  95   return LegacyInlinerBase::runOnSCC(SCC);
  96 }
  97
  98 void AMDGPUInliner::getAnalysisUsage(AnalysisUsage &AU) const {
  99   AU.addRequired<TargetTransformInfoWrapperPass>();
 100   LegacyInlinerBase::getAnalysisUsage(AU);
 101 }
 102
 103 unsigned AMDGPUInliner::getInlineThreshold(CallSite CS) const {
 104   int Thres = Params.DefaultThreshold;
 105
 106   Function *Caller = CS.getCaller();
 107   // Listen to the inlinehint attribute when it would increase the threshold
 108   // and the caller does not need to minimize its size.
 109   Function *Callee = CS.getCalledFunction();
 110   bool InlineHint = Callee && !Callee->isDeclaration() &&
 111     Callee->hasFnAttribute(Attribute::InlineHint);
 112   if (InlineHint && Params.HintThreshold && Params.HintThreshold > Thres
 113       && !Caller->hasFnAttribute(Attribute::MinSize))
 114     Thres = Params.HintThreshold.getValue();
 115
 116   const DataLayout &DL = Caller->getParent()->getDataLayout();
 117   if (!Callee)
 118     return (unsigned)Thres;
 119
 120   // If we have a pointer to private array passed into a function
 121   // it will not be optimized out, leaving scratch usage.
 122   // Increase the inline threshold to allow inliniting in this case.
 123   uint64_t AllocaSize = 0;
 124   SmallPtrSet<const AllocaInst *, 8> AIVisited;
 125   for (Value *PtrArg : CS.args()) {
 126     Type *Ty = PtrArg->getType();
 127     if (!Ty->isPointerTy() ||
 128         Ty->getPointerAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS)
 129       continue;
 130     PtrArg = GetUnderlyingObject(PtrArg, DL);
 131     if (const AllocaInst *AI = dyn_cast<AllocaInst>(PtrArg)) {
 132       if (!AI->isStaticAlloca() || !AIVisited.insert(AI).second)
 133         continue;
 134       AllocaSize += DL.getTypeAllocSize(AI->getAllocatedType());
 135       // If the amount of stack memory is excessive we will not be able
 136       // to get rid of the scratch anyway, bail out.
 137       if (AllocaSize > ArgAllocaCutoff) {
 138         AllocaSize = 0;
 139         break;
 140       }
 141     }
 142   }
 143   if (AllocaSize)
 144     Thres += ArgAllocaCost;
 145
 146   return (unsigned)Thres;
 147 }
 148
 149 // Check if call is just a wrapper around another call.
 150 // In this case we only have call and ret instructions.
 151 static bool isWrapperOnlyCall(CallSite CS) {
 152   Function *Callee = CS.getCalledFunction();
 153   if (!Callee || Callee->size() != 1)
 154     return false;
 155   const BasicBlock &BB = Callee->getEntryBlock();
 156   if (const Instruction *I = BB.getFirstNonPHI()) {
 157     if (!isa<CallInst>(I)) {
 158       return false;
 159     }
 160     if (isa<ReturnInst>(*std::next(I->getIterator()))) {
 161       LLVM_DEBUG(dbgs() << "    Wrapper only call detected: "
 162                         << Callee->getName() << '\n');
 163       return true;
 164     }
 165   }
 166   return false;
 167 }
 168
 169 InlineCost AMDGPUInliner::getInlineCost(CallSite CS) {
 170   Function *Callee = CS.getCalledFunction();
 171   Function *Caller = CS.getCaller();
 172   TargetTransformInfo &TTI = TTIWP->getTTI(*Callee);
 173
 174   if (!Callee || Callee->isDeclaration())
 175     return llvm::InlineCost::getNever("undefined callee");
 176
 177   if (CS.isNoInline())
 178     return llvm::InlineCost::getNever("noinline");
 179
 180   if (!TTI.areInlineCompatible(Caller, Callee))
 181     return llvm::InlineCost::getNever("incompatible");
 182
 183   if (CS.hasFnAttr(Attribute::AlwaysInline)) {
 184     auto IsViable = isInlineViable(*Callee);
 185     if (IsViable)
 186       return llvm::InlineCost::getAlways("alwaysinline viable");
 187     return llvm::InlineCost::getNever(IsViable.message);
 188   }
 189
 190   if (isWrapperOnlyCall(CS))
 191     return llvm::InlineCost::getAlways("wrapper-only call");
 192
 193   InlineParams LocalParams = Params;
 194   LocalParams.DefaultThreshold = (int)getInlineThreshold(CS);
 195   bool RemarksEnabled = false;
 196   const auto &BBs = Caller->getBasicBlockList();
 197   if (!BBs.empty()) {
 198     auto DI = OptimizationRemark(DEBUG_TYPE, "", DebugLoc(), &BBs.front());
 199     if (DI.isEnabled())
 200       RemarksEnabled = true;
 201   }
 202
 203   OptimizationRemarkEmitter ORE(Caller);
 204   std::function<AssumptionCache &(Function &)> GetAssumptionCache =
 205       [this](Function &F) -> AssumptionCache & {
 206     return ACT->getAssumptionCache(F);
 207   };
 208
 209   return llvm::getInlineCost(CS, Callee, LocalParams, TTI, GetAssumptionCache,
 210                              None, PSI, RemarksEnabled ? &ORE : nullptr);
 211 }