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