1 | //===-- GCNPreRALongBranchReg.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 | // \file |
9 | // \brief Pass to estimate pre RA branch size and reserve a pair of SGPRs if |
10 | // there is a long branch. Branch size at this point is difficult to track since |
11 | // we have no idea what spills will be inserted later on. We just assume 8 bytes |
12 | // per instruction to compute approximations without computing the actual |
13 | // instruction size to see if we're in the neighborhood of the maximum branch |
14 | // distrance threshold tuning of what is considered "long" is handled through |
15 | // amdgpu-long-branch-factor cl argument which sets LongBranchFactor. |
16 | //===----------------------------------------------------------------------===// |
17 | #include "GCNPreRALongBranchReg.h" |
18 | #include "AMDGPU.h" |
19 | #include "GCNSubtarget.h" |
20 | #include "SIMachineFunctionInfo.h" |
21 | #include "llvm/CodeGen/MachineFunctionPass.h" |
22 | #include "llvm/InitializePasses.h" |
23 | |
24 | using namespace llvm; |
25 | |
26 | #define DEBUG_TYPE "amdgpu-pre-ra-long-branch-reg" |
27 | |
28 | namespace { |
29 | |
30 | static cl::opt<double> LongBranchFactor( |
31 | "amdgpu-long-branch-factor" , cl::init(Val: 1.0), cl::Hidden, |
32 | cl::desc("Factor to apply to what qualifies as a long branch " |
33 | "to reserve a pair of scalar registers. If this value " |
34 | "is 0 the long branch registers are never reserved. As this " |
35 | "value grows the greater chance the branch distance will fall " |
36 | "within the threshold and the registers will be marked to be " |
37 | "reserved. We lean towards always reserving a register for " |
38 | "long jumps" )); |
39 | |
40 | class GCNPreRALongBranchReg { |
41 | |
42 | struct BasicBlockInfo { |
43 | // Offset - Distance from the beginning of the function to the beginning |
44 | // of this basic block. |
45 | uint64_t Offset = 0; |
46 | // Size - Size of the basic block in bytes |
47 | uint64_t Size = 0; |
48 | }; |
49 | void generateBlockInfo(MachineFunction &MF, |
50 | SmallVectorImpl<BasicBlockInfo> &BlockInfo); |
51 | |
52 | public: |
53 | GCNPreRALongBranchReg() = default; |
54 | bool run(MachineFunction &MF); |
55 | }; |
56 | |
57 | class GCNPreRALongBranchRegLegacy : public MachineFunctionPass { |
58 | public: |
59 | static char ID; |
60 | GCNPreRALongBranchRegLegacy() : MachineFunctionPass(ID) { |
61 | initializeGCNPreRALongBranchRegLegacyPass(*PassRegistry::getPassRegistry()); |
62 | } |
63 | |
64 | bool runOnMachineFunction(MachineFunction &MF) override { |
65 | return GCNPreRALongBranchReg().run(MF); |
66 | } |
67 | |
68 | StringRef getPassName() const override { |
69 | return "AMDGPU Pre-RA Long Branch Reg" ; |
70 | } |
71 | |
72 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
73 | AU.setPreservesAll(); |
74 | MachineFunctionPass::getAnalysisUsage(AU); |
75 | } |
76 | }; |
77 | } // End anonymous namespace. |
78 | |
79 | char GCNPreRALongBranchRegLegacy::ID = 0; |
80 | |
81 | INITIALIZE_PASS(GCNPreRALongBranchRegLegacy, DEBUG_TYPE, |
82 | "AMDGPU Pre-RA Long Branch Reg" , false, false) |
83 | |
84 | char &llvm::GCNPreRALongBranchRegID = GCNPreRALongBranchRegLegacy::ID; |
85 | void GCNPreRALongBranchReg::generateBlockInfo( |
86 | MachineFunction &MF, SmallVectorImpl<BasicBlockInfo> &BlockInfo) { |
87 | |
88 | BlockInfo.resize(N: MF.getNumBlockIDs()); |
89 | |
90 | // Approximate the size of all basic blocks by just |
91 | // assuming 8 bytes per instruction |
92 | for (const MachineBasicBlock &MBB : MF) { |
93 | uint64_t NumInstr = 0; |
94 | // Loop through the basic block and add up all non-debug |
95 | // non-meta instructions |
96 | for (const MachineInstr &MI : MBB) { |
97 | // isMetaInstruction is a superset of isDebugIstr |
98 | if (MI.isMetaInstruction()) |
99 | continue; |
100 | NumInstr += 1; |
101 | } |
102 | // Approximate size as just 8 bytes per instruction |
103 | BlockInfo[MBB.getNumber()].Size = 8 * NumInstr; |
104 | } |
105 | uint64_t PrevNum = (&MF)->begin()->getNumber(); |
106 | for (auto &MBB : |
107 | make_range(x: std::next(x: MachineFunction::iterator((&MF)->begin())), |
108 | y: (&MF)->end())) { |
109 | uint64_t Num = MBB.getNumber(); |
110 | // Compute the offset immediately following this block. |
111 | BlockInfo[Num].Offset = BlockInfo[PrevNum].Offset + BlockInfo[PrevNum].Size; |
112 | PrevNum = Num; |
113 | } |
114 | } |
115 | |
116 | bool GCNPreRALongBranchReg::run(MachineFunction &MF) { |
117 | const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>(); |
118 | const SIInstrInfo *TII = STM.getInstrInfo(); |
119 | const SIRegisterInfo *TRI = STM.getRegisterInfo(); |
120 | SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
121 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
122 | |
123 | // For now, reserve highest available SGPR pair. After RA, |
124 | // shift down to a lower unused pair of SGPRs |
125 | // If all registers are used, then findUnusedRegister will return |
126 | // AMDGPU::NoRegister. |
127 | constexpr bool ReserveHighestRegister = true; |
128 | Register LongBranchReservedReg = TRI->findUnusedRegister( |
129 | MRI, RC: &AMDGPU::SGPR_64RegClass, MF, ReserveHighestVGPR: ReserveHighestRegister); |
130 | if (!LongBranchReservedReg) |
131 | return false; |
132 | |
133 | // Approximate code size and offsets of each basic block |
134 | SmallVector<BasicBlockInfo, 16> BlockInfo; |
135 | generateBlockInfo(MF, BlockInfo); |
136 | |
137 | for (const MachineBasicBlock &MBB : MF) { |
138 | MachineBasicBlock::const_iterator Last = MBB.getLastNonDebugInstr(); |
139 | if (Last == MBB.end() || !Last->isUnconditionalBranch()) |
140 | continue; |
141 | MachineBasicBlock *DestBB = TII->getBranchDestBlock(MI: *Last); |
142 | uint64_t BlockDistance = static_cast<uint64_t>( |
143 | LongBranchFactor * BlockInfo[DestBB->getNumber()].Offset); |
144 | // If the distance falls outside the threshold assume it is a long branch |
145 | // and we need to reserve the registers |
146 | if (!TII->isBranchOffsetInRange(BranchOpc: Last->getOpcode(), BrOffset: BlockDistance)) { |
147 | MFI->setLongBranchReservedReg(LongBranchReservedReg); |
148 | return true; |
149 | } |
150 | } |
151 | return false; |
152 | } |
153 | |
154 | PreservedAnalyses |
155 | GCNPreRALongBranchRegPass::run(MachineFunction &MF, |
156 | MachineFunctionAnalysisManager &MFAM) { |
157 | GCNPreRALongBranchReg().run(MF); |
158 | return PreservedAnalyses::all(); |
159 | } |
160 | |