xref: /llvm-project/llvm/lib/Target/AMDGPU/AMDGPULowerKernelAttributes.cpp (revision 7ecbe0c7a01848fce88dcf3b6977cec866e9938b)
1 //===-- AMDGPULowerKernelAttributes.cpp ------------------------------------------===//
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 This pass does attempts to make use of reqd_work_group_size metadata
10 /// to eliminate loads from the dispatch packet and to constant fold OpenCL
11 /// get_local_size-like functions.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "AMDGPU.h"
16 #include "AMDGPUTargetMachine.h"
17 #include "llvm/Analysis/ValueTracking.h"
18 #include "llvm/CodeGen/Passes.h"
19 #include "llvm/CodeGen/TargetPassConfig.h"
20 #include "llvm/IR/Constants.h"
21 #include "llvm/IR/Function.h"
22 #include "llvm/IR/InstIterator.h"
23 #include "llvm/IR/Instructions.h"
24 #include "llvm/IR/PassManager.h"
25 #include "llvm/IR/PatternMatch.h"
26 #include "llvm/Pass.h"
27 
28 #define DEBUG_TYPE "amdgpu-lower-kernel-attributes"
29 
30 using namespace llvm;
31 
32 namespace {
33 
34 // Field offsets in hsa_kernel_dispatch_packet_t.
35 enum DispatchPackedOffsets {
36   WORKGROUP_SIZE_X = 4,
37   WORKGROUP_SIZE_Y = 6,
38   WORKGROUP_SIZE_Z = 8,
39 
40   GRID_SIZE_X = 12,
41   GRID_SIZE_Y = 16,
42   GRID_SIZE_Z = 20
43 };
44 
45 class AMDGPULowerKernelAttributes : public ModulePass {
46 public:
47   static char ID;
48 
49   AMDGPULowerKernelAttributes() : ModulePass(ID) {}
50 
51   bool runOnModule(Module &M) override;
52 
53   StringRef getPassName() const override {
54     return "AMDGPU Kernel Attributes";
55   }
56 
57   void getAnalysisUsage(AnalysisUsage &AU) const override {
58     AU.setPreservesAll();
59  }
60 };
61 
62 } // end anonymous namespace
63 
64 static bool processUse(CallInst *CI) {
65   Function *F = CI->getParent()->getParent();
66 
67   auto MD = F->getMetadata("reqd_work_group_size");
68   const bool HasReqdWorkGroupSize = MD && MD->getNumOperands() == 3;
69 
70   const bool HasUniformWorkGroupSize =
71     F->getFnAttribute("uniform-work-group-size").getValueAsString() == "true";
72 
73   if (!HasReqdWorkGroupSize && !HasUniformWorkGroupSize)
74     return false;
75 
76   Value *WorkGroupSizeX = nullptr;
77   Value *WorkGroupSizeY = nullptr;
78   Value *WorkGroupSizeZ = nullptr;
79 
80   Value *GridSizeX = nullptr;
81   Value *GridSizeY = nullptr;
82   Value *GridSizeZ = nullptr;
83 
84   const DataLayout &DL = F->getParent()->getDataLayout();
85 
86   // We expect to see several GEP users, casted to the appropriate type and
87   // loaded.
88   for (User *U : CI->users()) {
89     if (!U->hasOneUse())
90       continue;
91 
92     int64_t Offset = 0;
93     if (GetPointerBaseWithConstantOffset(U, Offset, DL) != CI)
94       continue;
95 
96     auto *BCI = dyn_cast<BitCastInst>(*U->user_begin());
97     if (!BCI || !BCI->hasOneUse())
98       continue;
99 
100     auto *Load = dyn_cast<LoadInst>(*BCI->user_begin());
101     if (!Load || !Load->isSimple())
102       continue;
103 
104     unsigned LoadSize = DL.getTypeStoreSize(Load->getType());
105 
106     // TODO: Handle merged loads.
107     switch (Offset) {
108     case WORKGROUP_SIZE_X:
109       if (LoadSize == 2)
110         WorkGroupSizeX = Load;
111       break;
112     case WORKGROUP_SIZE_Y:
113       if (LoadSize == 2)
114         WorkGroupSizeY = Load;
115       break;
116     case WORKGROUP_SIZE_Z:
117       if (LoadSize == 2)
118         WorkGroupSizeZ = Load;
119       break;
120     case GRID_SIZE_X:
121       if (LoadSize == 4)
122         GridSizeX = Load;
123       break;
124     case GRID_SIZE_Y:
125       if (LoadSize == 4)
126         GridSizeY = Load;
127       break;
128     case GRID_SIZE_Z:
129       if (LoadSize == 4)
130         GridSizeZ = Load;
131       break;
132     default:
133       break;
134     }
135   }
136 
137   // Pattern match the code used to handle partial workgroup dispatches in the
138   // library implementation of get_local_size, so the entire function can be
139   // constant folded with a known group size.
140   //
141   // uint r = grid_size - group_id * group_size;
142   // get_local_size = (r < group_size) ? r : group_size;
143   //
144   // If we have uniform-work-group-size (which is the default in OpenCL 1.2),
145   // the grid_size is required to be a multiple of group_size). In this case:
146   //
147   // grid_size - (group_id * group_size) < group_size
148   // ->
149   // grid_size < group_size + (group_id * group_size)
150   //
151   // (grid_size / group_size) < 1 + group_id
152   //
153   // grid_size / group_size is at least 1, so we can conclude the select
154   // condition is false (except for group_id == 0, where the select result is
155   // the same).
156 
157   bool MadeChange = false;
158   Value *WorkGroupSizes[3] = { WorkGroupSizeX, WorkGroupSizeY, WorkGroupSizeZ };
159   Value *GridSizes[3] = { GridSizeX, GridSizeY, GridSizeZ };
160 
161   for (int I = 0; HasUniformWorkGroupSize && I < 3; ++I) {
162     Value *GroupSize = WorkGroupSizes[I];
163     Value *GridSize = GridSizes[I];
164     if (!GroupSize || !GridSize)
165       continue;
166 
167     for (User *U : GroupSize->users()) {
168       auto *ZextGroupSize = dyn_cast<ZExtInst>(U);
169       if (!ZextGroupSize)
170         continue;
171 
172       for (User *ZextUser : ZextGroupSize->users()) {
173         auto *SI = dyn_cast<SelectInst>(ZextUser);
174         if (!SI)
175           continue;
176 
177         using namespace llvm::PatternMatch;
178         auto GroupIDIntrin = I == 0 ?
179           m_Intrinsic<Intrinsic::amdgcn_workgroup_id_x>() :
180             (I == 1 ? m_Intrinsic<Intrinsic::amdgcn_workgroup_id_y>() :
181                       m_Intrinsic<Intrinsic::amdgcn_workgroup_id_z>());
182 
183         auto SubExpr = m_Sub(m_Specific(GridSize),
184                              m_Mul(GroupIDIntrin, m_Specific(ZextGroupSize)));
185 
186         ICmpInst::Predicate Pred;
187         if (match(SI,
188                   m_Select(m_ICmp(Pred, SubExpr, m_Specific(ZextGroupSize)),
189                            SubExpr,
190                            m_Specific(ZextGroupSize))) &&
191             Pred == ICmpInst::ICMP_ULT) {
192           if (HasReqdWorkGroupSize) {
193             ConstantInt *KnownSize
194               = mdconst::extract<ConstantInt>(MD->getOperand(I));
195             SI->replaceAllUsesWith(ConstantExpr::getIntegerCast(KnownSize,
196                                                                 SI->getType(),
197                                                                 false));
198           } else {
199             SI->replaceAllUsesWith(ZextGroupSize);
200           }
201 
202           MadeChange = true;
203         }
204       }
205     }
206   }
207 
208   if (!HasReqdWorkGroupSize)
209     return MadeChange;
210 
211   // Eliminate any other loads we can from the dispatch packet.
212   for (int I = 0; I < 3; ++I) {
213     Value *GroupSize = WorkGroupSizes[I];
214     if (!GroupSize)
215       continue;
216 
217     ConstantInt *KnownSize = mdconst::extract<ConstantInt>(MD->getOperand(I));
218     GroupSize->replaceAllUsesWith(
219       ConstantExpr::getIntegerCast(KnownSize,
220                                    GroupSize->getType(),
221                                    false));
222     MadeChange = true;
223   }
224 
225   return MadeChange;
226 }
227 
228 // TODO: Move makeLIDRangeMetadata usage into here. Seem to not get
229 // TargetPassConfig for subtarget.
230 bool AMDGPULowerKernelAttributes::runOnModule(Module &M) {
231   StringRef DispatchPtrName
232     = Intrinsic::getName(Intrinsic::amdgcn_dispatch_ptr);
233 
234   Function *DispatchPtr = M.getFunction(DispatchPtrName);
235   if (!DispatchPtr) // Dispatch ptr not used.
236     return false;
237 
238   bool MadeChange = false;
239 
240   SmallPtrSet<Instruction *, 4> HandledUses;
241   for (auto *U : DispatchPtr->users()) {
242     CallInst *CI = cast<CallInst>(U);
243     if (HandledUses.insert(CI).second) {
244       if (processUse(CI))
245         MadeChange = true;
246     }
247   }
248 
249   return MadeChange;
250 }
251 
252 INITIALIZE_PASS_BEGIN(AMDGPULowerKernelAttributes, DEBUG_TYPE,
253                       "AMDGPU IR optimizations", false, false)
254 INITIALIZE_PASS_END(AMDGPULowerKernelAttributes, DEBUG_TYPE, "AMDGPU IR optimizations",
255                     false, false)
256 
257 char AMDGPULowerKernelAttributes::ID = 0;
258 
259 ModulePass *llvm::createAMDGPULowerKernelAttributesPass() {
260   return new AMDGPULowerKernelAttributes();
261 }
262 
263 PreservedAnalyses
264 AMDGPULowerKernelAttributesPass::run(Function &F, FunctionAnalysisManager &AM) {
265   StringRef DispatchPtrName =
266       Intrinsic::getName(Intrinsic::amdgcn_dispatch_ptr);
267 
268   Function *DispatchPtr = F.getParent()->getFunction(DispatchPtrName);
269   if (!DispatchPtr) // Dispatch ptr not used.
270     return PreservedAnalyses::all();
271 
272   for (Instruction &I : instructions(F)) {
273     if (CallInst *CI = dyn_cast<CallInst>(&I)) {
274       if (CI->getCalledFunction() == DispatchPtr)
275         processUse(CI);
276     }
277   }
278 
279   return PreservedAnalyses::all();
280 }
281