1 | //===-- GCNPreRAOptimizations.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 |
10 | /// This pass combines split register tuple initialization into a single pseudo: |
11 | /// |
12 | /// undef %0.sub1:sreg_64 = S_MOV_B32 1 |
13 | /// %0.sub0:sreg_64 = S_MOV_B32 2 |
14 | /// => |
15 | /// %0:sreg_64 = S_MOV_B64_IMM_PSEUDO 0x200000001 |
16 | /// |
17 | /// This is to allow rematerialization of a value instead of spilling. It is |
18 | /// supposed to be done after register coalescer to allow it to do its job and |
19 | /// before actual register allocation to allow rematerialization. |
20 | /// |
21 | /// Right now the pass only handles 64 bit SGPRs with immediate initializers, |
22 | /// although the same shall be possible with other register classes and |
23 | /// instructions if necessary. |
24 | /// |
25 | //===----------------------------------------------------------------------===// |
26 | |
27 | #include "AMDGPU.h" |
28 | #include "GCNSubtarget.h" |
29 | #include "MCTargetDesc/AMDGPUMCTargetDesc.h" |
30 | #include "llvm/CodeGen/LiveIntervals.h" |
31 | #include "llvm/CodeGen/MachineFunctionPass.h" |
32 | #include "llvm/InitializePasses.h" |
33 | |
34 | using namespace llvm; |
35 | |
36 | #define DEBUG_TYPE "amdgpu-pre-ra-optimizations" |
37 | |
38 | namespace { |
39 | |
40 | class GCNPreRAOptimizations : public MachineFunctionPass { |
41 | private: |
42 | const SIInstrInfo *TII; |
43 | const SIRegisterInfo *TRI; |
44 | MachineRegisterInfo *MRI; |
45 | LiveIntervals *LIS; |
46 | |
47 | bool processReg(Register Reg); |
48 | |
49 | public: |
50 | static char ID; |
51 | |
52 | GCNPreRAOptimizations() : MachineFunctionPass(ID) { |
53 | initializeGCNPreRAOptimizationsPass(*PassRegistry::getPassRegistry()); |
54 | } |
55 | |
56 | bool runOnMachineFunction(MachineFunction &MF) override; |
57 | |
58 | StringRef getPassName() const override { |
59 | return "AMDGPU Pre-RA optimizations" ; |
60 | } |
61 | |
62 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
63 | AU.addRequired<LiveIntervalsWrapperPass>(); |
64 | AU.setPreservesAll(); |
65 | MachineFunctionPass::getAnalysisUsage(AU); |
66 | } |
67 | }; |
68 | |
69 | } // End anonymous namespace. |
70 | |
71 | INITIALIZE_PASS_BEGIN(GCNPreRAOptimizations, DEBUG_TYPE, |
72 | "AMDGPU Pre-RA optimizations" , false, false) |
73 | INITIALIZE_PASS_DEPENDENCY(LiveIntervalsWrapperPass) |
74 | INITIALIZE_PASS_END(GCNPreRAOptimizations, DEBUG_TYPE, "Pre-RA optimizations" , |
75 | false, false) |
76 | |
77 | char GCNPreRAOptimizations::ID = 0; |
78 | |
79 | char &llvm::GCNPreRAOptimizationsID = GCNPreRAOptimizations::ID; |
80 | |
81 | FunctionPass *llvm::createGCNPreRAOptimizationsPass() { |
82 | return new GCNPreRAOptimizations(); |
83 | } |
84 | |
85 | bool GCNPreRAOptimizations::processReg(Register Reg) { |
86 | MachineInstr *Def0 = nullptr; |
87 | MachineInstr *Def1 = nullptr; |
88 | uint64_t Init = 0; |
89 | bool Changed = false; |
90 | SmallSet<Register, 32> ModifiedRegs; |
91 | bool IsAGPRDst = TRI->isAGPRClass(RC: MRI->getRegClass(Reg)); |
92 | |
93 | for (MachineInstr &I : MRI->def_instructions(Reg)) { |
94 | switch (I.getOpcode()) { |
95 | default: |
96 | return false; |
97 | case AMDGPU::V_ACCVGPR_WRITE_B32_e64: |
98 | break; |
99 | case AMDGPU::COPY: { |
100 | // Some subtargets cannot do an AGPR to AGPR copy directly, and need an |
101 | // intermdiate temporary VGPR register. Try to find the defining |
102 | // accvgpr_write to avoid temporary registers. |
103 | |
104 | if (!IsAGPRDst) |
105 | return false; |
106 | |
107 | Register SrcReg = I.getOperand(i: 1).getReg(); |
108 | |
109 | if (!SrcReg.isVirtual()) |
110 | break; |
111 | |
112 | // Check if source of copy is from another AGPR. |
113 | bool IsAGPRSrc = TRI->isAGPRClass(RC: MRI->getRegClass(Reg: SrcReg)); |
114 | if (!IsAGPRSrc) |
115 | break; |
116 | |
117 | // def_instructions() does not look at subregs so it may give us a |
118 | // different instruction that defines the same vreg but different subreg |
119 | // so we have to manually check subreg. |
120 | Register SrcSubReg = I.getOperand(i: 1).getSubReg(); |
121 | for (auto &Def : MRI->def_instructions(Reg: SrcReg)) { |
122 | if (SrcSubReg != Def.getOperand(i: 0).getSubReg()) |
123 | continue; |
124 | |
125 | if (Def.getOpcode() == AMDGPU::V_ACCVGPR_WRITE_B32_e64) { |
126 | MachineOperand DefSrcMO = Def.getOperand(i: 1); |
127 | |
128 | // Immediates are not an issue and can be propagated in |
129 | // postrapseudos pass. Only handle cases where defining |
130 | // accvgpr_write source is a vreg. |
131 | if (DefSrcMO.isReg() && DefSrcMO.getReg().isVirtual()) { |
132 | // Propagate source reg of accvgpr write to this copy instruction |
133 | I.getOperand(i: 1).setReg(DefSrcMO.getReg()); |
134 | I.getOperand(i: 1).setSubReg(DefSrcMO.getSubReg()); |
135 | |
136 | // Reg uses were changed, collect unique set of registers to update |
137 | // live intervals at the end. |
138 | ModifiedRegs.insert(V: DefSrcMO.getReg()); |
139 | ModifiedRegs.insert(V: SrcReg); |
140 | |
141 | Changed = true; |
142 | } |
143 | |
144 | // Found the defining accvgpr_write, stop looking any further. |
145 | break; |
146 | } |
147 | } |
148 | break; |
149 | } |
150 | case AMDGPU::S_MOV_B32: |
151 | if (I.getOperand(i: 0).getReg() != Reg || !I.getOperand(i: 1).isImm() || |
152 | I.getNumOperands() != 2) |
153 | return false; |
154 | |
155 | switch (I.getOperand(i: 0).getSubReg()) { |
156 | default: |
157 | return false; |
158 | case AMDGPU::sub0: |
159 | if (Def0) |
160 | return false; |
161 | Def0 = &I; |
162 | Init |= I.getOperand(i: 1).getImm() & 0xffffffff; |
163 | break; |
164 | case AMDGPU::sub1: |
165 | if (Def1) |
166 | return false; |
167 | Def1 = &I; |
168 | Init |= static_cast<uint64_t>(I.getOperand(i: 1).getImm()) << 32; |
169 | break; |
170 | } |
171 | break; |
172 | } |
173 | } |
174 | |
175 | // For AGPR reg, check if live intervals need to be updated. |
176 | if (IsAGPRDst) { |
177 | if (Changed) { |
178 | for (Register RegToUpdate : ModifiedRegs) { |
179 | LIS->removeInterval(Reg: RegToUpdate); |
180 | LIS->createAndComputeVirtRegInterval(Reg: RegToUpdate); |
181 | } |
182 | } |
183 | |
184 | return Changed; |
185 | } |
186 | |
187 | // For SGPR reg, check if we can combine instructions. |
188 | if (!Def0 || !Def1 || Def0->getParent() != Def1->getParent()) |
189 | return Changed; |
190 | |
191 | LLVM_DEBUG(dbgs() << "Combining:\n " << *Def0 << " " << *Def1 |
192 | << " =>\n" ); |
193 | |
194 | if (SlotIndex::isEarlierInstr(A: LIS->getInstructionIndex(Instr: *Def1), |
195 | B: LIS->getInstructionIndex(Instr: *Def0))) |
196 | std::swap(a&: Def0, b&: Def1); |
197 | |
198 | LIS->RemoveMachineInstrFromMaps(MI&: *Def0); |
199 | LIS->RemoveMachineInstrFromMaps(MI&: *Def1); |
200 | auto NewI = BuildMI(BB&: *Def0->getParent(), I&: *Def0, MIMD: Def0->getDebugLoc(), |
201 | MCID: TII->get(Opcode: AMDGPU::S_MOV_B64_IMM_PSEUDO), DestReg: Reg) |
202 | .addImm(Val: Init); |
203 | |
204 | Def0->eraseFromParent(); |
205 | Def1->eraseFromParent(); |
206 | LIS->InsertMachineInstrInMaps(MI&: *NewI); |
207 | LIS->removeInterval(Reg); |
208 | LIS->createAndComputeVirtRegInterval(Reg); |
209 | |
210 | LLVM_DEBUG(dbgs() << " " << *NewI); |
211 | |
212 | return true; |
213 | } |
214 | |
215 | bool GCNPreRAOptimizations::runOnMachineFunction(MachineFunction &MF) { |
216 | if (skipFunction(F: MF.getFunction())) |
217 | return false; |
218 | |
219 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
220 | TII = ST.getInstrInfo(); |
221 | MRI = &MF.getRegInfo(); |
222 | LIS = &getAnalysis<LiveIntervalsWrapperPass>().getLIS(); |
223 | TRI = ST.getRegisterInfo(); |
224 | |
225 | bool Changed = false; |
226 | |
227 | for (unsigned I = 0, E = MRI->getNumVirtRegs(); I != E; ++I) { |
228 | Register Reg = Register::index2VirtReg(Index: I); |
229 | if (!LIS->hasInterval(Reg)) |
230 | continue; |
231 | const TargetRegisterClass *RC = MRI->getRegClass(Reg); |
232 | if ((RC->MC->getSizeInBits() != 64 || !TRI->isSGPRClass(RC)) && |
233 | (ST.hasGFX90AInsts() || !TRI->isAGPRClass(RC))) |
234 | continue; |
235 | |
236 | Changed |= processReg(Reg); |
237 | } |
238 | |
239 | return Changed; |
240 | } |
241 | |