1 //===- AMDGPUOpenCLEnqueuedBlockLowering.cpp - Lower enqueued block -------===//
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 post-linking pass replaces the function pointer of enqueued
11 // block kernel with a global variable (runtime handle) and adds
12 // "runtime-handle" attribute to the enqueued block kernel.
13 //
14 // In LLVM CodeGen the runtime-handle metadata will be translated to
15 // RuntimeHandle metadata in code object. Runtime allocates a global buffer
16 // for each kernel with RuntimeHandle metadata and saves the kernel address
17 // required for the AQL packet into the buffer. __enqueue_kernel function
18 // in device library knows that the invoke function pointer in the block
19 // literal is actually runtime handle and loads the kernel address from it
20 // and put it into AQL packet for dispatching.
21 //
22 // This cannot be done in FE since FE cannot create a unique global variable
23 // with external linkage across LLVM modules. The global variable with internal
24 // linkage does not work since optimization passes will try to replace loads
25 // of the global variable with its initialization value.
26 //
27 // It also identifies the kernels directly or indirectly enqueues kernels
28 // and adds "calls-enqueue-kernel" function attribute to them, which will
29 // be used to determine whether to emit runtime metadata for the kernel
30 // enqueue related hidden kernel arguments.
31 //
32 //===----------------------------------------------------------------------===//
33 
34 #include "AMDGPU.h"
35 #include "llvm/ADT/DenseSet.h"
36 #include "llvm/ADT/SmallString.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/Instructions.h"
39 #include "llvm/IR/Mangler.h"
40 #include "llvm/IR/Module.h"
41 #include "llvm/Pass.h"
42 #include "llvm/Support/Debug.h"
43 
44 #define DEBUG_TYPE "amdgpu-lower-enqueued-block"
45 
46 using namespace llvm;
47 
48 namespace {
49 
50 /// Lower enqueued blocks.
51 class AMDGPUOpenCLEnqueuedBlockLowering : public ModulePass {
52 public:
53   static char ID;
54 
55   explicit AMDGPUOpenCLEnqueuedBlockLowering() : ModulePass(ID) {}
56 
57 private:
58   bool runOnModule(Module &M) override;
59 };
60 
61 } // end anonymous namespace
62 
63 char AMDGPUOpenCLEnqueuedBlockLowering::ID = 0;
64 
65 char &llvm::AMDGPUOpenCLEnqueuedBlockLoweringID =
66     AMDGPUOpenCLEnqueuedBlockLowering::ID;
67 
68 INITIALIZE_PASS(AMDGPUOpenCLEnqueuedBlockLowering, DEBUG_TYPE,
69                 "Lower OpenCL enqueued blocks", false, false)
70 
71 ModulePass* llvm::createAMDGPUOpenCLEnqueuedBlockLoweringPass() {
72   return new AMDGPUOpenCLEnqueuedBlockLowering();
73 }
74 
75 /// Collect direct or indirect callers of \p F and save them
76 /// to \p Callers.
77 static void collectCallers(Function *F, DenseSet<Function *> &Callers) {
78   for (auto U : F->users()) {
79     if (auto *CI = dyn_cast<CallInst>(&*U)) {
80       auto *Caller = CI->getParent()->getParent();
81       if (Callers.insert(Caller).second)
82         collectCallers(Caller, Callers);
83     }
84   }
85 }
86 
87 /// If \p U is instruction or constant, collect functions which directly or
88 /// indirectly use it.
89 static void collectFunctionUsers(User *U, DenseSet<Function *> &Funcs) {
90   if (auto *I = dyn_cast<Instruction>(U)) {
91     auto *F = I->getParent()->getParent();
92     if (Funcs.insert(F).second)
93       collectCallers(F, Funcs);
94     return;
95   }
96   if (!isa<Constant>(U))
97     return;
98   for (auto UU : U->users())
99     collectFunctionUsers(&*UU, Funcs);
100 }
101 
102 bool AMDGPUOpenCLEnqueuedBlockLowering::runOnModule(Module &M) {
103   DenseSet<Function *> Callers;
104   auto &C = M.getContext();
105   bool Changed = false;
106   for (auto &F : M.functions()) {
107     if (F.hasFnAttribute("enqueued-block")) {
108       if (!F.hasName()) {
109         SmallString<64> Name;
110         Mangler::getNameWithPrefix(Name, "__amdgpu_enqueued_kernel",
111                                    M.getDataLayout());
112         F.setName(Name);
113       }
114       LLVM_DEBUG(dbgs() << "found enqueued kernel: " << F.getName() << '\n');
115       auto RuntimeHandle = (F.getName() + ".runtime_handle").str();
116       auto T = ArrayType::get(Type::getInt64Ty(C), 2);
117       auto *GV = new GlobalVariable(
118           M, T,
119           /*isConstant=*/false, GlobalValue::ExternalLinkage,
120           /*Initializer=*/Constant::getNullValue(T), RuntimeHandle,
121           /*InsertBefore=*/nullptr, GlobalValue::NotThreadLocal,
122           AMDGPUAS::GLOBAL_ADDRESS,
123           /*isExternallyInitialized=*/false);
124       LLVM_DEBUG(dbgs() << "runtime handle created: " << *GV << '\n');
125 
126       for (auto U : F.users()) {
127         auto *UU = &*U;
128         if (!isa<ConstantExpr>(UU))
129           continue;
130         collectFunctionUsers(UU, Callers);
131         auto *BitCast = cast<ConstantExpr>(UU);
132         auto *NewPtr = ConstantExpr::getPointerCast(GV, BitCast->getType());
133         BitCast->replaceAllUsesWith(NewPtr);
134         F.addFnAttr("runtime-handle", RuntimeHandle);
135         F.setLinkage(GlobalValue::ExternalLinkage);
136         Changed = true;
137       }
138     }
139   }
140 
141   for (auto F : Callers) {
142     if (F->getCallingConv() != CallingConv::AMDGPU_KERNEL)
143       continue;
144     F->addFnAttr("calls-enqueue-kernel");
145     LLVM_DEBUG(dbgs() << "mark enqueue_kernel caller:" << F->getName() << '\n');
146   }
147   return Changed;
148 }
149