1 //===- AMDGPUOpenCLEnqueuedBlockLowering.cpp - Lower enqueued block -------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // \file
11 // \brief This post-linking pass replaces the function pointer of enqueued
12 // block kernel with a global variable (runtime handle) and adds
13 // "runtime-handle" attribute to the enqueued block kernel.
14 //
15 // In LLVM CodeGen the runtime-handle metadata will be translated to
16 // RuntimeHandle metadata in code object. Runtime allocates a global buffer
17 // for each kernel with RuntimeHandel metadata and saves the kernel address
18 // required for the AQL packet into the buffer. __enqueue_kernel function
19 // in device library knows that the invoke function pointer in the block
20 // literal is actually runtime handle and loads the kernel address from it
21 // and put it into AQL packet for dispatching.
22 //
23 // This cannot be done in FE since FE cannot create a unique global variable
24 // with external linkage across LLVM modules. The global variable with internal
25 // linkage does not work since optimization passes will try to replace loads
26 // of the global variable with its initialization value.
27 //
28 // It also identifies the kernels directly or indirectly enqueues kernels
29 // and adds "calls-enqueue-kernel" function attribute to them, which will
30 // be used to determine whether to emit runtime metadata for the kernel
31 // enqueue related hidden kernel arguments.
32 //
33 //===----------------------------------------------------------------------===//
34 
35 #include "AMDGPU.h"
36 #include "llvm/ADT/DenseSet.h"
37 #include "llvm/ADT/StringRef.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/Instructions.h"
40 #include "llvm/IR/Module.h"
41 #include "llvm/IR/User.h"
42 #include "llvm/Pass.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/raw_ostream.h"
45 
46 #define DEBUG_TYPE "amdgpu-lower-enqueued-block"
47 
48 using namespace llvm;
49 
50 namespace {
51 
52 /// \brief Lower enqueued blocks.
53 class AMDGPUOpenCLEnqueuedBlockLowering : public ModulePass {
54 public:
55   static char ID;
56 
57   explicit AMDGPUOpenCLEnqueuedBlockLowering() : ModulePass(ID) {}
58 
59 private:
60   bool runOnModule(Module &M) override;
61 };
62 
63 } // end anonymous namespace
64 
65 char AMDGPUOpenCLEnqueuedBlockLowering::ID = 0;
66 
67 char &llvm::AMDGPUOpenCLEnqueuedBlockLoweringID =
68     AMDGPUOpenCLEnqueuedBlockLowering::ID;
69 
70 INITIALIZE_PASS(AMDGPUOpenCLEnqueuedBlockLowering, DEBUG_TYPE,
71                 "Lower OpenCL enqueued blocks", false, false)
72 
73 ModulePass* llvm::createAMDGPUOpenCLEnqueuedBlockLoweringPass() {
74   return new AMDGPUOpenCLEnqueuedBlockLowering();
75 }
76 
77 /// Collect direct or indrect callers of \p F and save them
78 /// to \p Callers.
79 static void collectCallers(Function *F, DenseSet<Function *> &Callers) {
80   for (auto U : F->users()) {
81     if (auto *CI = dyn_cast<CallInst>(&*U)) {
82       auto *Caller = CI->getParent()->getParent();
83       if (Callers.count(Caller))
84         continue;
85       Callers.insert(Caller);
86       collectCallers(Caller, Callers);
87     }
88   }
89 }
90 
91 bool AMDGPUOpenCLEnqueuedBlockLowering::runOnModule(Module &M) {
92   DenseSet<Function *> Callers;
93   auto &C = M.getContext();
94   bool Changed = false;
95   for (auto &F : M.functions()) {
96     if (F.hasFnAttribute("enqueued-block")) {
97       if (!F.hasOneUse() || !F.user_begin()->hasOneUse() ||
98           !isa<ConstantExpr>(*F.user_begin()) ||
99           !isa<ConstantExpr>(*F.user_begin()->user_begin())) {
100         continue;
101       }
102       auto *BitCast = cast<ConstantExpr>(*F.user_begin());
103       auto *AddrCast = cast<ConstantExpr>(*BitCast->user_begin());
104       auto RuntimeHandle = (F.getName() + "_runtime_handle").str();
105       auto *GV = new GlobalVariable(
106           M, Type::getInt8Ty(C)->getPointerTo(AMDGPUAS::GLOBAL_ADDRESS),
107           /*IsConstant=*/true, GlobalValue::ExternalLinkage,
108           /*Initializer=*/nullptr, RuntimeHandle, /*InsertBefore=*/nullptr,
109           GlobalValue::NotThreadLocal, AMDGPUAS::GLOBAL_ADDRESS,
110           /*IsExternallyInitialized=*/true);
111       DEBUG(dbgs() << "runtime handle created: " << *GV << '\n');
112       auto *NewPtr = ConstantExpr::getPointerCast(GV, AddrCast->getType());
113       AddrCast->replaceAllUsesWith(NewPtr);
114       F.addFnAttr("runtime-handle", RuntimeHandle);
115       F.setLinkage(GlobalValue::ExternalLinkage);
116 
117       // Collect direct or indirect callers of enqueue_kernel.
118       for (auto U : NewPtr->users()) {
119         if (auto *I = dyn_cast<Instruction>(&*U)) {
120           auto *F = I->getParent()->getParent();
121           Callers.insert(F);
122           collectCallers(F, Callers);
123         }
124       }
125       Changed = true;
126     }
127   }
128 
129   for (auto F : Callers) {
130     if (F->getCallingConv() != CallingConv::AMDGPU_KERNEL)
131       continue;
132     F->addFnAttr("calls-enqueue-kernel");
133   }
134   return Changed;
135 }
136