1 //===-- AMDGPUAlwaysInlinePass.cpp - Promote Allocas ----------------------===//
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 pass marks all internal functions as always_inline and creates
11 /// duplicates of all other functions and marks the duplicates as always_inline.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "AMDGPU.h"
16 #include "AMDGPUTargetMachine.h"
17 #include "Utils/AMDGPUBaseInfo.h"
18 #include "llvm/IR/Module.h"
19 #include "llvm/Pass.h"
20 #include "llvm/Support/CommandLine.h"
21 
22 using namespace llvm;
23 
24 namespace {
25 
26 static cl::opt<bool> StressCalls(
27   "amdgpu-stress-function-calls",
28   cl::Hidden,
29   cl::desc("Force all functions to be noinline"),
30   cl::init(false));
31 
32 class AMDGPUAlwaysInline : public ModulePass {
33   bool GlobalOpt;
34 
35 public:
36   static char ID;
37 
38   AMDGPUAlwaysInline(bool GlobalOpt = false) :
39     ModulePass(ID), GlobalOpt(GlobalOpt) { }
40   bool runOnModule(Module &M) override;
41 
42   void getAnalysisUsage(AnalysisUsage &AU) const override {
43     AU.setPreservesAll();
44  }
45 };
46 
47 } // End anonymous namespace
48 
49 INITIALIZE_PASS(AMDGPUAlwaysInline, "amdgpu-always-inline",
50                 "AMDGPU Inline All Functions", false, false)
51 
52 char AMDGPUAlwaysInline::ID = 0;
53 
54 static void
55 recursivelyVisitUsers(GlobalValue &GV,
56                       SmallPtrSetImpl<Function *> &FuncsToAlwaysInline) {
57   SmallVector<User *, 16> Stack(GV.users());
58 
59   SmallPtrSet<const Value *, 8> Visited;
60 
61   while (!Stack.empty()) {
62     User *U = Stack.pop_back_val();
63     if (!Visited.insert(U).second)
64       continue;
65 
66     if (Instruction *I = dyn_cast<Instruction>(U)) {
67       Function *F = I->getParent()->getParent();
68       if (!AMDGPU::isEntryFunctionCC(F->getCallingConv())) {
69         // FIXME: This is a horrible hack. We should always respect noinline,
70         // and just let us hit the error when we can't handle this.
71         //
72         // Unfortunately, clang adds noinline to all functions at -O0. We have
73         // to override this here. until that's fixed.
74         F->removeFnAttr(Attribute::NoInline);
75 
76         FuncsToAlwaysInline.insert(F);
77         Stack.push_back(F);
78       }
79 
80       // No need to look at further users, but we do need to inline any callers.
81       continue;
82     }
83 
84     for (User *UU : U->users())
85       Stack.push_back(UU);
86   }
87 }
88 
89 static bool alwaysInlineImpl(Module &M, bool GlobalOpt) {
90   std::vector<GlobalAlias*> AliasesToRemove;
91 
92   SmallPtrSet<Function *, 8> FuncsToAlwaysInline;
93   SmallPtrSet<Function *, 8> FuncsToNoInline;
94 
95   for (GlobalAlias &A : M.aliases()) {
96     if (Function* F = dyn_cast<Function>(A.getAliasee())) {
97       A.replaceAllUsesWith(F);
98       AliasesToRemove.push_back(&A);
99     }
100 
101     // FIXME: If the aliasee isn't a function, it's some kind of constant expr
102     // cast that won't be inlined through.
103   }
104 
105   if (GlobalOpt) {
106     for (GlobalAlias* A : AliasesToRemove) {
107       A->eraseFromParent();
108     }
109   }
110 
111   // Always force inlining of any function that uses an LDS global address. This
112   // is something of a workaround because we don't have a way of supporting LDS
113   // objects defined in functions. LDS is always allocated by a kernel, and it
114   // is difficult to manage LDS usage if a function may be used by multiple
115   // kernels.
116   //
117   // OpenCL doesn't allow declaring LDS in non-kernels, so in practice this
118   // should only appear when IPO passes manages to move LDs defined in a kernel
119   // into a single user function.
120 
121   for (GlobalVariable &GV : M.globals()) {
122     // TODO: Region address
123     unsigned AS = GV.getAddressSpace();
124     if (AS != AMDGPUAS::LOCAL_ADDRESS && AS != AMDGPUAS::REGION_ADDRESS)
125       continue;
126 
127     recursivelyVisitUsers(GV, FuncsToAlwaysInline);
128   }
129 
130   if (!AMDGPUTargetMachine::EnableFunctionCalls || StressCalls) {
131     auto IncompatAttr
132       = StressCalls ? Attribute::AlwaysInline : Attribute::NoInline;
133 
134     for (Function &F : M) {
135       if (!F.isDeclaration() && !F.use_empty() &&
136           !F.hasFnAttribute(IncompatAttr)) {
137         if (StressCalls) {
138           if (!FuncsToAlwaysInline.count(&F))
139             FuncsToNoInline.insert(&F);
140         } else
141           FuncsToAlwaysInline.insert(&F);
142       }
143     }
144   }
145 
146   for (Function *F : FuncsToAlwaysInline)
147     F->addFnAttr(Attribute::AlwaysInline);
148 
149   for (Function *F : FuncsToNoInline)
150     F->addFnAttr(Attribute::NoInline);
151 
152   return !FuncsToAlwaysInline.empty() || !FuncsToNoInline.empty();
153 }
154 
155 bool AMDGPUAlwaysInline::runOnModule(Module &M) {
156   return alwaysInlineImpl(M, GlobalOpt);
157 }
158 
159 ModulePass *llvm::createAMDGPUAlwaysInlinePass(bool GlobalOpt) {
160   return new AMDGPUAlwaysInline(GlobalOpt);
161 }
162 
163 PreservedAnalyses AMDGPUAlwaysInlinePass::run(Module &M,
164                                               ModuleAnalysisManager &AM) {
165   alwaysInlineImpl(M, GlobalOpt);
166   return PreservedAnalyses::all();
167 }
168