| 1 | //===- AMDGPUResourceUsageAnalysis.h ---- analysis of resources -----------===// |
| 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 | /// \brief Analyzes how many registers and other resources are used by |
| 11 | /// functions. |
| 12 | /// |
| 13 | /// The results of this analysis are used to fill the register usage, flat |
| 14 | /// usage, etc. into hardware registers. |
| 15 | /// |
| 16 | //===----------------------------------------------------------------------===// |
| 17 | |
| 18 | #include "AMDGPUResourceUsageAnalysis.h" |
| 19 | #include "AMDGPU.h" |
| 20 | #include "GCNSubtarget.h" |
| 21 | #include "SIMachineFunctionInfo.h" |
| 22 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 23 | #include "llvm/CodeGen/MachineModuleInfo.h" |
| 24 | #include "llvm/CodeGen/TargetPassConfig.h" |
| 25 | #include "llvm/IR/GlobalValue.h" |
| 26 | #include "llvm/Target/TargetMachine.h" |
| 27 | |
| 28 | using namespace llvm; |
| 29 | using namespace llvm::AMDGPU; |
| 30 | |
| 31 | #define DEBUG_TYPE "amdgpu-resource-usage" |
| 32 | |
| 33 | char llvm::AMDGPUResourceUsageAnalysisWrapperPass::ID = 0; |
| 34 | char &llvm::AMDGPUResourceUsageAnalysisID = |
| 35 | AMDGPUResourceUsageAnalysisWrapperPass::ID; |
| 36 | |
| 37 | // In code object v4 and older, we need to tell the runtime some amount ahead of |
| 38 | // time if we don't know the true stack size. Assume a smaller number if this is |
| 39 | // only due to dynamic / non-entry block allocas. |
| 40 | static cl::opt<uint32_t> clAssumedStackSizeForExternalCall( |
| 41 | "amdgpu-assume-external-call-stack-size" , |
| 42 | cl::desc("Assumed stack use of any external call (in bytes)" ), cl::Hidden, |
| 43 | cl::init(Val: 16384)); |
| 44 | |
| 45 | static cl::opt<uint32_t> clAssumedStackSizeForDynamicSizeObjects( |
| 46 | "amdgpu-assume-dynamic-stack-object-size" , |
| 47 | cl::desc("Assumed extra stack use if there are any " |
| 48 | "variable sized objects (in bytes)" ), |
| 49 | cl::Hidden, cl::init(Val: 4096)); |
| 50 | |
| 51 | INITIALIZE_PASS(AMDGPUResourceUsageAnalysisWrapperPass, DEBUG_TYPE, |
| 52 | "Function register usage analysis" , true, true) |
| 53 | |
| 54 | static const Function *getCalleeFunction(const MachineOperand &Op) { |
| 55 | if (Op.isImm()) { |
| 56 | assert(Op.getImm() == 0); |
| 57 | return nullptr; |
| 58 | } |
| 59 | return cast<Function>(Val: Op.getGlobal()->stripPointerCastsAndAliases()); |
| 60 | } |
| 61 | |
| 62 | static bool hasAnyNonFlatUseOfReg(const MachineRegisterInfo &MRI, |
| 63 | const SIInstrInfo &TII, unsigned Reg) { |
| 64 | for (const MachineOperand &UseOp : MRI.reg_operands(Reg)) { |
| 65 | if (!UseOp.isImplicit() || !TII.isFLAT(MI: *UseOp.getParent())) |
| 66 | return true; |
| 67 | } |
| 68 | |
| 69 | return false; |
| 70 | } |
| 71 | |
| 72 | bool AMDGPUResourceUsageAnalysisWrapperPass::runOnMachineFunction( |
| 73 | MachineFunction &MF) { |
| 74 | auto *TPC = getAnalysisIfAvailable<TargetPassConfig>(); |
| 75 | if (!TPC) |
| 76 | return false; |
| 77 | |
| 78 | const TargetMachine &TM = TPC->getTM<TargetMachine>(); |
| 79 | const MCSubtargetInfo &STI = *TM.getMCSubtargetInfo(); |
| 80 | |
| 81 | // By default, for code object v5 and later, track only the minimum scratch |
| 82 | // size |
| 83 | uint32_t AssumedStackSizeForDynamicSizeObjects = |
| 84 | clAssumedStackSizeForDynamicSizeObjects; |
| 85 | uint32_t AssumedStackSizeForExternalCall = clAssumedStackSizeForExternalCall; |
| 86 | if (AMDGPU::getAMDHSACodeObjectVersion(M: *MF.getFunction().getParent()) >= |
| 87 | AMDGPU::AMDHSA_COV5 || |
| 88 | STI.getTargetTriple().getOS() == Triple::AMDPAL) { |
| 89 | if (!clAssumedStackSizeForDynamicSizeObjects.getNumOccurrences()) |
| 90 | AssumedStackSizeForDynamicSizeObjects = 0; |
| 91 | if (!clAssumedStackSizeForExternalCall.getNumOccurrences()) |
| 92 | AssumedStackSizeForExternalCall = 0; |
| 93 | } |
| 94 | |
| 95 | ResourceInfo = AMDGPUResourceUsageAnalysisImpl().analyzeResourceUsage( |
| 96 | MF, AssumedStackSizeForDynamicSizeObjects, |
| 97 | AssumedStackSizeForExternalCall); |
| 98 | |
| 99 | return false; |
| 100 | } |
| 101 | |
| 102 | AnalysisKey AMDGPUResourceUsageAnalysis::Key; |
| 103 | AMDGPUResourceUsageAnalysis::Result |
| 104 | AMDGPUResourceUsageAnalysis::run(MachineFunction &MF, |
| 105 | MachineFunctionAnalysisManager &MFAM) { |
| 106 | const MCSubtargetInfo &STI = *TM.getMCSubtargetInfo(); |
| 107 | |
| 108 | // By default, for code object v5 and later, track only the minimum scratch |
| 109 | // size |
| 110 | uint32_t AssumedStackSizeForDynamicSizeObjects = |
| 111 | clAssumedStackSizeForDynamicSizeObjects; |
| 112 | uint32_t AssumedStackSizeForExternalCall = clAssumedStackSizeForExternalCall; |
| 113 | if (AMDGPU::getAMDHSACodeObjectVersion(M: *MF.getFunction().getParent()) >= |
| 114 | AMDGPU::AMDHSA_COV5 || |
| 115 | STI.getTargetTriple().getOS() == Triple::AMDPAL) { |
| 116 | if (!clAssumedStackSizeForDynamicSizeObjects.getNumOccurrences()) |
| 117 | AssumedStackSizeForDynamicSizeObjects = 0; |
| 118 | if (!clAssumedStackSizeForExternalCall.getNumOccurrences()) |
| 119 | AssumedStackSizeForExternalCall = 0; |
| 120 | } |
| 121 | |
| 122 | return AMDGPUResourceUsageAnalysisImpl().analyzeResourceUsage( |
| 123 | MF, AssumedStackSizeForDynamicSizeObjects, |
| 124 | AssumedStackSizeForExternalCall); |
| 125 | } |
| 126 | |
| 127 | AMDGPUResourceUsageAnalysisImpl::SIFunctionResourceInfo |
| 128 | AMDGPUResourceUsageAnalysisImpl::analyzeResourceUsage( |
| 129 | const MachineFunction &MF, uint32_t AssumedStackSizeForDynamicSizeObjects, |
| 130 | uint32_t AssumedStackSizeForExternalCall) const { |
| 131 | SIFunctionResourceInfo Info; |
| 132 | |
| 133 | const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
| 134 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 135 | const MachineFrameInfo &FrameInfo = MF.getFrameInfo(); |
| 136 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 137 | const SIInstrInfo *TII = ST.getInstrInfo(); |
| 138 | const SIRegisterInfo &TRI = TII->getRegisterInfo(); |
| 139 | |
| 140 | Info.UsesFlatScratch = MRI.isPhysRegUsed(PhysReg: AMDGPU::FLAT_SCR_LO) || |
| 141 | MRI.isPhysRegUsed(PhysReg: AMDGPU::FLAT_SCR_HI) || |
| 142 | MRI.isLiveIn(Reg: MFI->getPreloadedReg( |
| 143 | Value: AMDGPUFunctionArgInfo::FLAT_SCRATCH_INIT)); |
| 144 | |
| 145 | Info.NumNamedBarrier = MFI->getNumNamedBarriers(); |
| 146 | |
| 147 | // Even if FLAT_SCRATCH is implicitly used, it has no effect if flat |
| 148 | // instructions aren't used to access the scratch buffer. Inline assembly may |
| 149 | // need it though. |
| 150 | // |
| 151 | // If we only have implicit uses of flat_scr on flat instructions, it is not |
| 152 | // really needed. |
| 153 | if (Info.UsesFlatScratch && !MFI->getUserSGPRInfo().hasFlatScratchInit() && |
| 154 | (!hasAnyNonFlatUseOfReg(MRI, TII: *TII, Reg: AMDGPU::FLAT_SCR) && |
| 155 | !hasAnyNonFlatUseOfReg(MRI, TII: *TII, Reg: AMDGPU::FLAT_SCR_LO) && |
| 156 | !hasAnyNonFlatUseOfReg(MRI, TII: *TII, Reg: AMDGPU::FLAT_SCR_HI))) { |
| 157 | Info.UsesFlatScratch = false; |
| 158 | } |
| 159 | |
| 160 | Info.PrivateSegmentSize = FrameInfo.getStackSize(); |
| 161 | |
| 162 | // Assume a big number if there are any unknown sized objects. |
| 163 | Info.HasDynamicallySizedStack = FrameInfo.hasVarSizedObjects(); |
| 164 | if (Info.HasDynamicallySizedStack) |
| 165 | Info.PrivateSegmentSize += AssumedStackSizeForDynamicSizeObjects; |
| 166 | |
| 167 | if (MFI->isStackRealigned()) |
| 168 | Info.PrivateSegmentSize += FrameInfo.getMaxAlign().value(); |
| 169 | |
| 170 | Info.UsesVCC = |
| 171 | MRI.isPhysRegUsed(PhysReg: AMDGPU::VCC_LO) || MRI.isPhysRegUsed(PhysReg: AMDGPU::VCC_HI); |
| 172 | Info.NumExplicitSGPR = TRI.getNumUsedPhysRegs(MRI, RC: AMDGPU::SGPR_32RegClass, |
| 173 | /*IncludeCalls=*/false); |
| 174 | if (ST.hasMAIInsts()) |
| 175 | Info.NumAGPR = TRI.getNumUsedPhysRegs(MRI, RC: AMDGPU::AGPR_32RegClass, |
| 176 | /*IncludeCalls=*/false); |
| 177 | |
| 178 | // If there are no calls, MachineRegisterInfo can tell us the used register |
| 179 | // count easily. |
| 180 | // A tail call isn't considered a call for MachineFrameInfo's purposes. |
| 181 | if (!FrameInfo.hasCalls() && !FrameInfo.hasTailCall()) { |
| 182 | Info.NumVGPR = TRI.getNumUsedPhysRegs(MRI, RC: AMDGPU::VGPR_32RegClass, |
| 183 | /*IncludeCalls=*/false); |
| 184 | return Info; |
| 185 | } |
| 186 | |
| 187 | int32_t MaxVGPR = -1; |
| 188 | Info.CalleeSegmentSize = 0; |
| 189 | |
| 190 | for (const MachineBasicBlock &MBB : MF) { |
| 191 | for (const MachineInstr &MI : MBB) { |
| 192 | for (unsigned I = 0; I < MI.getNumOperands(); ++I) { |
| 193 | const MachineOperand &MO = MI.getOperand(i: I); |
| 194 | |
| 195 | if (!MO.isReg()) |
| 196 | continue; |
| 197 | |
| 198 | Register Reg = MO.getReg(); |
| 199 | switch (Reg) { |
| 200 | case AMDGPU::NoRegister: |
| 201 | assert(MI.isDebugInstr() && |
| 202 | "Instruction uses invalid noreg register" ); |
| 203 | continue; |
| 204 | |
| 205 | case AMDGPU::XNACK_MASK: |
| 206 | case AMDGPU::XNACK_MASK_LO: |
| 207 | case AMDGPU::XNACK_MASK_HI: |
| 208 | llvm_unreachable("xnack_mask registers should not be used" ); |
| 209 | |
| 210 | case AMDGPU::LDS_DIRECT: |
| 211 | llvm_unreachable("lds_direct register should not be used" ); |
| 212 | |
| 213 | case AMDGPU::TBA: |
| 214 | case AMDGPU::TBA_LO: |
| 215 | case AMDGPU::TBA_HI: |
| 216 | case AMDGPU::TMA: |
| 217 | case AMDGPU::TMA_LO: |
| 218 | case AMDGPU::TMA_HI: |
| 219 | llvm_unreachable("trap handler registers should not be used" ); |
| 220 | |
| 221 | case AMDGPU::SRC_VCCZ: |
| 222 | llvm_unreachable("src_vccz register should not be used" ); |
| 223 | |
| 224 | case AMDGPU::SRC_EXECZ: |
| 225 | llvm_unreachable("src_execz register should not be used" ); |
| 226 | |
| 227 | case AMDGPU::SRC_SCC: |
| 228 | llvm_unreachable("src_scc register should not be used" ); |
| 229 | |
| 230 | default: |
| 231 | break; |
| 232 | } |
| 233 | |
| 234 | const TargetRegisterClass *RC = TRI.getPhysRegBaseClass(Reg); |
| 235 | assert((!RC || TRI.isVGPRClass(RC) || TRI.isSGPRClass(RC) || |
| 236 | TRI.isAGPRClass(RC) || AMDGPU::TTMP_32RegClass.contains(Reg) || |
| 237 | AMDGPU::TTMP_64RegClass.contains(Reg) || |
| 238 | AMDGPU::TTMP_128RegClass.contains(Reg) || |
| 239 | AMDGPU::TTMP_256RegClass.contains(Reg) || |
| 240 | AMDGPU::TTMP_512RegClass.contains(Reg)) && |
| 241 | "Unknown register class" ); |
| 242 | |
| 243 | if (!RC || !TRI.isVGPRClass(RC)) |
| 244 | continue; |
| 245 | |
| 246 | if (MI.isCall() || MI.isMetaInstruction()) |
| 247 | continue; |
| 248 | |
| 249 | unsigned Width = divideCeil(Numerator: TRI.getRegSizeInBits(RC: *RC), Denominator: 32); |
| 250 | unsigned HWReg = TRI.getHWRegIndex(Reg); |
| 251 | int MaxUsed = HWReg + Width - 1; |
| 252 | MaxVGPR = std::max(a: MaxUsed, b: MaxVGPR); |
| 253 | } |
| 254 | |
| 255 | if (MI.isCall()) { |
| 256 | // Pseudo used just to encode the underlying global. Is there a better |
| 257 | // way to track this? |
| 258 | |
| 259 | // TODO: Some of the generic call-like pseudos do not encode the callee, |
| 260 | // so we overly conservatively treat this as an indirect call. |
| 261 | const MachineOperand *CalleeOp = |
| 262 | TII->getNamedOperand(MI, OperandName: AMDGPU::OpName::callee); |
| 263 | |
| 264 | const Function *Callee = |
| 265 | CalleeOp ? getCalleeFunction(Op: *CalleeOp) : nullptr; |
| 266 | |
| 267 | auto isSameFunction = [](const MachineFunction &MF, const Function *F) { |
| 268 | return F == &MF.getFunction(); |
| 269 | }; |
| 270 | |
| 271 | if (Callee && !isSameFunction(MF, Callee)) |
| 272 | Info.Callees.push_back(Elt: Callee); |
| 273 | |
| 274 | bool IsIndirect = !Callee || Callee->isDeclaration(); |
| 275 | |
| 276 | // FIXME: Call site could have norecurse on it |
| 277 | if (!Callee || !Callee->doesNotRecurse()) { |
| 278 | Info.HasRecursion = true; |
| 279 | |
| 280 | // TODO: If we happen to know there is no stack usage in the |
| 281 | // callgraph, we don't need to assume an infinitely growing stack. |
| 282 | if (!MI.isReturn()) { |
| 283 | // We don't need to assume an unknown stack size for tail calls. |
| 284 | |
| 285 | // FIXME: This only benefits in the case where the kernel does not |
| 286 | // directly call the tail called function. If a kernel directly |
| 287 | // calls a tail recursive function, we'll assume maximum stack size |
| 288 | // based on the regular call instruction. |
| 289 | Info.CalleeSegmentSize = std::max( |
| 290 | a: Info.CalleeSegmentSize, |
| 291 | b: static_cast<uint64_t>(AssumedStackSizeForExternalCall)); |
| 292 | } |
| 293 | } |
| 294 | |
| 295 | if (IsIndirect) { |
| 296 | Info.CalleeSegmentSize = |
| 297 | std::max(a: Info.CalleeSegmentSize, |
| 298 | b: static_cast<uint64_t>(AssumedStackSizeForExternalCall)); |
| 299 | |
| 300 | // Register usage of indirect calls gets handled later |
| 301 | Info.UsesVCC = true; |
| 302 | Info.UsesFlatScratch = ST.hasFlatAddressSpace(); |
| 303 | Info.HasDynamicallySizedStack = true; |
| 304 | Info.HasIndirectCall = true; |
| 305 | } |
| 306 | } |
| 307 | } |
| 308 | } |
| 309 | |
| 310 | Info.NumVGPR = MaxVGPR + 1; |
| 311 | |
| 312 | return Info; |
| 313 | } |
| 314 | |