1 //===- AMDGPUMCInstLower.cpp - Lower AMDGPU MachineInstr to an MCInst -----===// 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 /// Code to lower AMDGPU MachineInstrs to their corresponding MCInst. 11 // 12 //===----------------------------------------------------------------------===// 13 // 14 15 #include "AMDGPUMCInstLower.h" 16 #include "AMDGPU.h" 17 #include "AMDGPUAsmPrinter.h" 18 #include "AMDGPUMachineFunction.h" 19 #include "MCTargetDesc/AMDGPUInstPrinter.h" 20 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 21 #include "llvm/CodeGen/MachineBasicBlock.h" 22 #include "llvm/CodeGen/MachineInstr.h" 23 #include "llvm/IR/Constants.h" 24 #include "llvm/IR/Function.h" 25 #include "llvm/IR/GlobalVariable.h" 26 #include "llvm/MC/MCCodeEmitter.h" 27 #include "llvm/MC/MCContext.h" 28 #include "llvm/MC/MCExpr.h" 29 #include "llvm/MC/MCInst.h" 30 #include "llvm/MC/MCObjectStreamer.h" 31 #include "llvm/MC/MCStreamer.h" 32 #include "llvm/Support/ErrorHandling.h" 33 #include "llvm/Support/Format.h" 34 #include <algorithm> 35 36 using namespace llvm; 37 38 #include "AMDGPUGenMCPseudoLowering.inc" 39 40 AMDGPUMCInstLower::AMDGPUMCInstLower(MCContext &ctx, 41 const TargetSubtargetInfo &st, 42 const AsmPrinter &ap): 43 Ctx(ctx), ST(st), AP(ap) { } 44 45 static MCSymbolRefExpr::VariantKind getVariantKind(unsigned MOFlags) { 46 switch (MOFlags) { 47 default: 48 return MCSymbolRefExpr::VK_None; 49 case SIInstrInfo::MO_GOTPCREL: 50 return MCSymbolRefExpr::VK_GOTPCREL; 51 case SIInstrInfo::MO_GOTPCREL32_LO: 52 return MCSymbolRefExpr::VK_AMDGPU_GOTPCREL32_LO; 53 case SIInstrInfo::MO_GOTPCREL32_HI: 54 return MCSymbolRefExpr::VK_AMDGPU_GOTPCREL32_HI; 55 case SIInstrInfo::MO_REL32_LO: 56 return MCSymbolRefExpr::VK_AMDGPU_REL32_LO; 57 case SIInstrInfo::MO_REL32_HI: 58 return MCSymbolRefExpr::VK_AMDGPU_REL32_HI; 59 case SIInstrInfo::MO_ABS32_LO: 60 return MCSymbolRefExpr::VK_AMDGPU_ABS32_LO; 61 case SIInstrInfo::MO_ABS32_HI: 62 return MCSymbolRefExpr::VK_AMDGPU_ABS32_HI; 63 } 64 } 65 66 bool AMDGPUMCInstLower::lowerOperand(const MachineOperand &MO, 67 MCOperand &MCOp) const { 68 switch (MO.getType()) { 69 default: 70 break; 71 case MachineOperand::MO_Immediate: 72 MCOp = MCOperand::createImm(MO.getImm()); 73 return true; 74 case MachineOperand::MO_Register: 75 MCOp = MCOperand::createReg(AMDGPU::getMCReg(MO.getReg(), ST)); 76 return true; 77 case MachineOperand::MO_MachineBasicBlock: 78 MCOp = MCOperand::createExpr( 79 MCSymbolRefExpr::create(MO.getMBB()->getSymbol(), Ctx)); 80 return true; 81 case MachineOperand::MO_GlobalAddress: { 82 const GlobalValue *GV = MO.getGlobal(); 83 SmallString<128> SymbolName; 84 AP.getNameWithPrefix(SymbolName, GV); 85 MCSymbol *Sym = Ctx.getOrCreateSymbol(SymbolName); 86 const MCExpr *Expr = 87 MCSymbolRefExpr::create(Sym, getVariantKind(MO.getTargetFlags()),Ctx); 88 int64_t Offset = MO.getOffset(); 89 if (Offset != 0) { 90 Expr = MCBinaryExpr::createAdd(Expr, 91 MCConstantExpr::create(Offset, Ctx), Ctx); 92 } 93 MCOp = MCOperand::createExpr(Expr); 94 return true; 95 } 96 case MachineOperand::MO_ExternalSymbol: { 97 MCSymbol *Sym = Ctx.getOrCreateSymbol(StringRef(MO.getSymbolName())); 98 const MCSymbolRefExpr *Expr = MCSymbolRefExpr::create(Sym, Ctx); 99 MCOp = MCOperand::createExpr(Expr); 100 return true; 101 } 102 case MachineOperand::MO_RegisterMask: 103 // Regmasks are like implicit defs. 104 return false; 105 case MachineOperand::MO_MCSymbol: 106 if (MO.getTargetFlags() == SIInstrInfo::MO_FAR_BRANCH_OFFSET) { 107 MCSymbol *Sym = MO.getMCSymbol(); 108 MCOp = MCOperand::createExpr(Sym->getVariableValue()); 109 return true; 110 } 111 break; 112 } 113 llvm_unreachable("unknown operand type"); 114 } 115 116 void AMDGPUMCInstLower::lower(const MachineInstr *MI, MCInst &OutMI) const { 117 unsigned Opcode = MI->getOpcode(); 118 const auto *TII = static_cast<const SIInstrInfo*>(ST.getInstrInfo()); 119 120 // FIXME: Should be able to handle this with lowerPseudoInstExpansion. We 121 // need to select it to the subtarget specific version, and there's no way to 122 // do that with a single pseudo source operation. 123 if (Opcode == AMDGPU::S_SETPC_B64_return) 124 Opcode = AMDGPU::S_SETPC_B64; 125 else if (Opcode == AMDGPU::SI_CALL) { 126 // SI_CALL is just S_SWAPPC_B64 with an additional operand to track the 127 // called function (which we need to remove here). 128 OutMI.setOpcode(TII->pseudoToMCOpcode(AMDGPU::S_SWAPPC_B64)); 129 MCOperand Dest, Src; 130 lowerOperand(MI->getOperand(0), Dest); 131 lowerOperand(MI->getOperand(1), Src); 132 OutMI.addOperand(Dest); 133 OutMI.addOperand(Src); 134 return; 135 } else if (Opcode == AMDGPU::SI_TCRETURN || 136 Opcode == AMDGPU::SI_TCRETURN_GFX) { 137 // TODO: How to use branch immediate and avoid register+add? 138 Opcode = AMDGPU::S_SETPC_B64; 139 } 140 141 int MCOpcode = TII->pseudoToMCOpcode(Opcode); 142 if (MCOpcode == -1) { 143 LLVMContext &C = MI->getParent()->getParent()->getFunction().getContext(); 144 C.emitError("AMDGPUMCInstLower::lower - Pseudo instruction doesn't have " 145 "a target-specific version: " + Twine(MI->getOpcode())); 146 } 147 148 OutMI.setOpcode(MCOpcode); 149 150 for (const MachineOperand &MO : MI->explicit_operands()) { 151 MCOperand MCOp; 152 lowerOperand(MO, MCOp); 153 OutMI.addOperand(MCOp); 154 } 155 156 int FIIdx = AMDGPU::getNamedOperandIdx(MCOpcode, AMDGPU::OpName::fi); 157 if (FIIdx >= (int)OutMI.getNumOperands()) 158 OutMI.addOperand(MCOperand::createImm(0)); 159 } 160 161 bool AMDGPUAsmPrinter::lowerOperand(const MachineOperand &MO, 162 MCOperand &MCOp) const { 163 const GCNSubtarget &STI = MF->getSubtarget<GCNSubtarget>(); 164 AMDGPUMCInstLower MCInstLowering(OutContext, STI, *this); 165 return MCInstLowering.lowerOperand(MO, MCOp); 166 } 167 168 const MCExpr *AMDGPUAsmPrinter::lowerConstant(const Constant *CV) { 169 170 // Intercept LDS variables with known addresses 171 if (const GlobalVariable *GV = dyn_cast<const GlobalVariable>(CV)) { 172 if (std::optional<uint32_t> Address = 173 AMDGPUMachineFunction::getLDSAbsoluteAddress(*GV)) { 174 auto *IntTy = Type::getInt32Ty(CV->getContext()); 175 return AsmPrinter::lowerConstant(ConstantInt::get(IntTy, *Address)); 176 } 177 } 178 179 if (const MCExpr *E = lowerAddrSpaceCast(TM, CV, OutContext)) 180 return E; 181 return AsmPrinter::lowerConstant(CV); 182 } 183 184 void AMDGPUAsmPrinter::emitInstruction(const MachineInstr *MI) { 185 // FIXME: Enable feature predicate checks once all the test pass. 186 // AMDGPU_MC::verifyInstructionPredicates(MI->getOpcode(), 187 // getSubtargetInfo().getFeatureBits()); 188 189 if (MCInst OutInst; lowerPseudoInstExpansion(MI, OutInst)) { 190 EmitToStreamer(*OutStreamer, OutInst); 191 return; 192 } 193 194 const GCNSubtarget &STI = MF->getSubtarget<GCNSubtarget>(); 195 AMDGPUMCInstLower MCInstLowering(OutContext, STI, *this); 196 197 StringRef Err; 198 if (!STI.getInstrInfo()->verifyInstruction(*MI, Err)) { 199 LLVMContext &C = MI->getParent()->getParent()->getFunction().getContext(); 200 C.emitError("Illegal instruction detected: " + Err); 201 MI->print(errs()); 202 } 203 204 if (MI->isBundle()) { 205 const MachineBasicBlock *MBB = MI->getParent(); 206 MachineBasicBlock::const_instr_iterator I = ++MI->getIterator(); 207 while (I != MBB->instr_end() && I->isInsideBundle()) { 208 emitInstruction(&*I); 209 ++I; 210 } 211 } else { 212 // We don't want these pseudo instructions encoded. They are 213 // placeholder terminator instructions and should only be printed as 214 // comments. 215 if (MI->getOpcode() == AMDGPU::SI_RETURN_TO_EPILOG) { 216 if (isVerbose()) 217 OutStreamer->emitRawComment(" return to shader part epilog"); 218 return; 219 } 220 221 if (MI->getOpcode() == AMDGPU::WAVE_BARRIER) { 222 if (isVerbose()) 223 OutStreamer->emitRawComment(" wave barrier"); 224 return; 225 } 226 227 if (MI->getOpcode() == AMDGPU::SCHED_BARRIER) { 228 if (isVerbose()) { 229 std::string HexString; 230 raw_string_ostream HexStream(HexString); 231 HexStream << format_hex(MI->getOperand(0).getImm(), 10, true); 232 OutStreamer->emitRawComment(" sched_barrier mask(" + HexString + ")"); 233 } 234 return; 235 } 236 237 if (MI->getOpcode() == AMDGPU::SCHED_GROUP_BARRIER) { 238 if (isVerbose()) { 239 std::string HexString; 240 raw_string_ostream HexStream(HexString); 241 HexStream << format_hex(MI->getOperand(0).getImm(), 10, true); 242 OutStreamer->emitRawComment( 243 " sched_group_barrier mask(" + HexString + ") size(" + 244 Twine(MI->getOperand(1).getImm()) + ") SyncID(" + 245 Twine(MI->getOperand(2).getImm()) + ")"); 246 } 247 return; 248 } 249 250 if (MI->getOpcode() == AMDGPU::IGLP_OPT) { 251 if (isVerbose()) { 252 std::string HexString; 253 raw_string_ostream HexStream(HexString); 254 HexStream << format_hex(MI->getOperand(0).getImm(), 10, true); 255 OutStreamer->emitRawComment(" iglp_opt mask(" + HexString + ")"); 256 } 257 return; 258 } 259 260 if (MI->getOpcode() == AMDGPU::SI_MASKED_UNREACHABLE) { 261 if (isVerbose()) 262 OutStreamer->emitRawComment(" divergent unreachable"); 263 return; 264 } 265 266 if (MI->isMetaInstruction()) { 267 if (isVerbose()) 268 OutStreamer->emitRawComment(" meta instruction"); 269 return; 270 } 271 272 MCInst TmpInst; 273 MCInstLowering.lower(MI, TmpInst); 274 EmitToStreamer(*OutStreamer, TmpInst); 275 276 #ifdef EXPENSIVE_CHECKS 277 // Check getInstSizeInBytes on explicitly specified CPUs (it cannot 278 // work correctly for the generic CPU). 279 // 280 // The isPseudo check really shouldn't be here, but unfortunately there are 281 // some negative lit tests that depend on being able to continue through 282 // here even when pseudo instructions haven't been lowered. 283 // 284 // We also overestimate branch sizes with the offset bug. 285 if (!MI->isPseudo() && STI.isCPUStringValid(STI.getCPU()) && 286 (!STI.hasOffset3fBug() || !MI->isBranch())) { 287 SmallVector<MCFixup, 4> Fixups; 288 SmallVector<char, 16> CodeBytes; 289 290 std::unique_ptr<MCCodeEmitter> InstEmitter(createAMDGPUMCCodeEmitter( 291 *STI.getInstrInfo(), OutContext)); 292 InstEmitter->encodeInstruction(TmpInst, CodeBytes, Fixups, STI); 293 294 assert(CodeBytes.size() == STI.getInstrInfo()->getInstSizeInBytes(*MI)); 295 } 296 #endif 297 298 if (DumpCodeInstEmitter) { 299 // Disassemble instruction/operands to text 300 DisasmLines.resize(DisasmLines.size() + 1); 301 std::string &DisasmLine = DisasmLines.back(); 302 raw_string_ostream DisasmStream(DisasmLine); 303 304 AMDGPUInstPrinter InstPrinter(*TM.getMCAsmInfo(), *STI.getInstrInfo(), 305 *STI.getRegisterInfo()); 306 InstPrinter.printInst(&TmpInst, 0, StringRef(), STI, DisasmStream); 307 308 // Disassemble instruction/operands to hex representation. 309 SmallVector<MCFixup, 4> Fixups; 310 SmallVector<char, 16> CodeBytes; 311 312 DumpCodeInstEmitter->encodeInstruction( 313 TmpInst, CodeBytes, Fixups, MF->getSubtarget<MCSubtargetInfo>()); 314 HexLines.resize(HexLines.size() + 1); 315 std::string &HexLine = HexLines.back(); 316 raw_string_ostream HexStream(HexLine); 317 318 for (size_t i = 0; i < CodeBytes.size(); i += 4) { 319 unsigned int CodeDWord = *(unsigned int *)&CodeBytes[i]; 320 HexStream << format("%s%08X", (i > 0 ? " " : ""), CodeDWord); 321 } 322 323 DisasmLineMaxLen = std::max(DisasmLineMaxLen, DisasmLine.size()); 324 } 325 } 326 } 327