| 1 | //===-- AMDGPUAsmBackend.cpp - AMDGPU Assembler Backend -------------------===// |
| 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 | /// \file |
| 8 | //===----------------------------------------------------------------------===// |
| 9 | |
| 10 | #include "MCTargetDesc/AMDGPUFixupKinds.h" |
| 11 | #include "MCTargetDesc/AMDGPUMCTargetDesc.h" |
| 12 | #include "Utils/AMDGPUBaseInfo.h" |
| 13 | #include "llvm/ADT/StringSwitch.h" |
| 14 | #include "llvm/BinaryFormat/ELF.h" |
| 15 | #include "llvm/MC/MCAsmBackend.h" |
| 16 | #include "llvm/MC/MCAsmInfo.h" |
| 17 | #include "llvm/MC/MCAssembler.h" |
| 18 | #include "llvm/MC/MCContext.h" |
| 19 | #include "llvm/MC/MCObjectWriter.h" |
| 20 | #include "llvm/MC/MCSubtargetInfo.h" |
| 21 | #include "llvm/MC/MCValue.h" |
| 22 | #include "llvm/MC/TargetRegistry.h" |
| 23 | #include "llvm/Support/EndianStream.h" |
| 24 | #include "llvm/TargetParser/TargetParser.h" |
| 25 | |
| 26 | using namespace llvm; |
| 27 | using namespace llvm::AMDGPU; |
| 28 | |
| 29 | namespace { |
| 30 | |
| 31 | class AMDGPUAsmBackend : public MCAsmBackend { |
| 32 | public: |
| 33 | AMDGPUAsmBackend(const Target &T) : MCAsmBackend(llvm::endianness::little) {} |
| 34 | |
| 35 | void applyFixup(const MCFragment &, const MCFixup &, const MCValue &Target, |
| 36 | uint8_t *Data, uint64_t Value, bool IsResolved) override; |
| 37 | bool fixupNeedsRelaxationAdvanced(const MCFragment &, const MCFixup &, |
| 38 | const MCValue &, uint64_t, |
| 39 | bool) const override; |
| 40 | |
| 41 | void relaxInstruction(MCInst &Inst, |
| 42 | const MCSubtargetInfo &STI) const override; |
| 43 | |
| 44 | bool mayNeedRelaxation(unsigned Opcode, ArrayRef<MCOperand> Operands, |
| 45 | const MCSubtargetInfo &STI) const override; |
| 46 | |
| 47 | unsigned getMinimumNopSize() const override; |
| 48 | bool writeNopData(raw_ostream &OS, uint64_t Count, |
| 49 | const MCSubtargetInfo *STI) const override; |
| 50 | |
| 51 | std::optional<MCFixupKind> getFixupKind(StringRef Name) const override; |
| 52 | MCFixupKindInfo getFixupKindInfo(MCFixupKind Kind) const override; |
| 53 | }; |
| 54 | |
| 55 | } //End anonymous namespace |
| 56 | |
| 57 | void AMDGPUAsmBackend::relaxInstruction(MCInst &Inst, |
| 58 | const MCSubtargetInfo &STI) const { |
| 59 | MCInst Res; |
| 60 | unsigned RelaxedOpcode = AMDGPU::getSOPPWithRelaxation(Opcode: Inst.getOpcode()); |
| 61 | Res.setOpcode(RelaxedOpcode); |
| 62 | Res.addOperand(Op: Inst.getOperand(i: 0)); |
| 63 | Inst = std::move(Res); |
| 64 | } |
| 65 | |
| 66 | bool AMDGPUAsmBackend::fixupNeedsRelaxationAdvanced(const MCFragment &, |
| 67 | const MCFixup &Fixup, |
| 68 | const MCValue &, |
| 69 | uint64_t Value, |
| 70 | bool Resolved) const { |
| 71 | if (!Resolved) |
| 72 | return true; |
| 73 | // if the branch target has an offset of x3f this needs to be relaxed to |
| 74 | // add a s_nop 0 immediately after branch to effectively increment offset |
| 75 | // for hardware workaround in gfx1010 |
| 76 | return (((int64_t(Value)/4)-1) == 0x3f); |
| 77 | } |
| 78 | |
| 79 | bool AMDGPUAsmBackend::mayNeedRelaxation(unsigned Opcode, |
| 80 | ArrayRef<MCOperand> Operands, |
| 81 | const MCSubtargetInfo &STI) const { |
| 82 | if (!STI.hasFeature(Feature: AMDGPU::FeatureOffset3fBug)) |
| 83 | return false; |
| 84 | |
| 85 | if (AMDGPU::getSOPPWithRelaxation(Opcode) >= 0) |
| 86 | return true; |
| 87 | |
| 88 | return false; |
| 89 | } |
| 90 | |
| 91 | static unsigned getFixupKindNumBytes(unsigned Kind) { |
| 92 | switch (Kind) { |
| 93 | case AMDGPU::fixup_si_sopp_br: |
| 94 | return 2; |
| 95 | case FK_SecRel_1: |
| 96 | case FK_Data_1: |
| 97 | return 1; |
| 98 | case FK_SecRel_2: |
| 99 | case FK_Data_2: |
| 100 | return 2; |
| 101 | case FK_SecRel_4: |
| 102 | case FK_Data_4: |
| 103 | return 4; |
| 104 | case FK_SecRel_8: |
| 105 | case FK_Data_8: |
| 106 | return 8; |
| 107 | default: |
| 108 | llvm_unreachable("Unknown fixup kind!" ); |
| 109 | } |
| 110 | } |
| 111 | |
| 112 | static uint64_t adjustFixupValue(const MCFixup &Fixup, uint64_t Value, |
| 113 | MCContext *Ctx) { |
| 114 | int64_t SignedValue = static_cast<int64_t>(Value); |
| 115 | |
| 116 | switch (Fixup.getKind()) { |
| 117 | case AMDGPU::fixup_si_sopp_br: { |
| 118 | int64_t BrImm = (SignedValue - 4) / 4; |
| 119 | |
| 120 | if (Ctx && !isInt<16>(x: BrImm)) |
| 121 | Ctx->reportError(L: Fixup.getLoc(), Msg: "branch size exceeds simm16" ); |
| 122 | |
| 123 | return BrImm; |
| 124 | } |
| 125 | case FK_Data_1: |
| 126 | case FK_Data_2: |
| 127 | case FK_Data_4: |
| 128 | case FK_Data_8: |
| 129 | case FK_SecRel_4: |
| 130 | return Value; |
| 131 | default: |
| 132 | llvm_unreachable("unhandled fixup kind" ); |
| 133 | } |
| 134 | } |
| 135 | |
| 136 | void AMDGPUAsmBackend::applyFixup(const MCFragment &F, const MCFixup &Fixup, |
| 137 | const MCValue &Target, uint8_t *Data, |
| 138 | uint64_t Value, bool IsResolved) { |
| 139 | if (Target.getSpecifier()) |
| 140 | IsResolved = false; |
| 141 | maybeAddReloc(F, Fixup, Target, Value, IsResolved); |
| 142 | if (mc::isRelocation(FixupKind: Fixup.getKind())) |
| 143 | return; |
| 144 | |
| 145 | Value = adjustFixupValue(Fixup, Value, Ctx: &getContext()); |
| 146 | if (!Value) |
| 147 | return; // Doesn't change encoding. |
| 148 | |
| 149 | MCFixupKindInfo Info = getFixupKindInfo(Kind: Fixup.getKind()); |
| 150 | |
| 151 | // Shift the value into position. |
| 152 | Value <<= Info.TargetOffset; |
| 153 | |
| 154 | unsigned NumBytes = getFixupKindNumBytes(Kind: Fixup.getKind()); |
| 155 | assert(Fixup.getOffset() + NumBytes <= F.getSize() && |
| 156 | "Invalid fixup offset!" ); |
| 157 | |
| 158 | // For each byte of the fragment that the fixup touches, mask in the bits from |
| 159 | // the fixup value. |
| 160 | for (unsigned i = 0; i != NumBytes; ++i) |
| 161 | Data[i] |= static_cast<uint8_t>((Value >> (i * 8)) & 0xff); |
| 162 | } |
| 163 | |
| 164 | std::optional<MCFixupKind> |
| 165 | AMDGPUAsmBackend::getFixupKind(StringRef Name) const { |
| 166 | auto Type = StringSwitch<unsigned>(Name) |
| 167 | #define ELF_RELOC(Name, Value) .Case(#Name, Value) |
| 168 | #include "llvm/BinaryFormat/ELFRelocs/AMDGPU.def" |
| 169 | #undef ELF_RELOC |
| 170 | .Case(S: "BFD_RELOC_NONE" , Value: ELF::R_AMDGPU_NONE) |
| 171 | .Case(S: "BFD_RELOC_32" , Value: ELF::R_AMDGPU_ABS32) |
| 172 | .Case(S: "BFD_RELOC_64" , Value: ELF::R_AMDGPU_ABS64) |
| 173 | .Default(Value: -1u); |
| 174 | if (Type != -1u) |
| 175 | return static_cast<MCFixupKind>(FirstLiteralRelocationKind + Type); |
| 176 | return std::nullopt; |
| 177 | } |
| 178 | |
| 179 | MCFixupKindInfo AMDGPUAsmBackend::getFixupKindInfo(MCFixupKind Kind) const { |
| 180 | const static MCFixupKindInfo Infos[AMDGPU::NumTargetFixupKinds] = { |
| 181 | // name offset bits flags |
| 182 | {.Name: "fixup_si_sopp_br" , .TargetOffset: 0, .TargetSize: 16, .Flags: 0}, |
| 183 | }; |
| 184 | |
| 185 | if (mc::isRelocation(FixupKind: Kind)) |
| 186 | return {}; |
| 187 | |
| 188 | if (Kind < FirstTargetFixupKind) |
| 189 | return MCAsmBackend::getFixupKindInfo(Kind); |
| 190 | |
| 191 | assert(unsigned(Kind - FirstTargetFixupKind) < AMDGPU::NumTargetFixupKinds && |
| 192 | "Invalid kind!" ); |
| 193 | return Infos[Kind - FirstTargetFixupKind]; |
| 194 | } |
| 195 | |
| 196 | unsigned AMDGPUAsmBackend::getMinimumNopSize() const { |
| 197 | return 4; |
| 198 | } |
| 199 | |
| 200 | bool AMDGPUAsmBackend::writeNopData(raw_ostream &OS, uint64_t Count, |
| 201 | const MCSubtargetInfo *STI) const { |
| 202 | // If the count is not aligned to the minimum instruction alignment, we must |
| 203 | // be writing data into the text section (otherwise we have unaligned |
| 204 | // instructions, and thus have far bigger problems), so just write zeros |
| 205 | // instead. |
| 206 | unsigned MinInstAlignment = getContext().getAsmInfo()->getMinInstAlignment(); |
| 207 | OS.write_zeros(NumZeros: Count % MinInstAlignment); |
| 208 | |
| 209 | // We are properly aligned, so write NOPs as requested. |
| 210 | Count /= MinInstAlignment; |
| 211 | |
| 212 | // FIXME: R600 support. |
| 213 | // s_nop 0 |
| 214 | const uint32_t Encoded_S_NOP_0 = 0xbf800000; |
| 215 | |
| 216 | assert(MinInstAlignment == sizeof(Encoded_S_NOP_0)); |
| 217 | for (uint64_t I = 0; I != Count; ++I) |
| 218 | support::endian::write<uint32_t>(os&: OS, value: Encoded_S_NOP_0, endian: Endian); |
| 219 | |
| 220 | return true; |
| 221 | } |
| 222 | |
| 223 | //===----------------------------------------------------------------------===// |
| 224 | // ELFAMDGPUAsmBackend class |
| 225 | //===----------------------------------------------------------------------===// |
| 226 | |
| 227 | namespace { |
| 228 | |
| 229 | class ELFAMDGPUAsmBackend : public AMDGPUAsmBackend { |
| 230 | bool Is64Bit; |
| 231 | bool HasRelocationAddend; |
| 232 | uint8_t OSABI = ELF::ELFOSABI_NONE; |
| 233 | |
| 234 | public: |
| 235 | ELFAMDGPUAsmBackend(const Target &T, const Triple &TT) |
| 236 | : AMDGPUAsmBackend(T), Is64Bit(TT.isAMDGCN()), |
| 237 | HasRelocationAddend(TT.getOS() == Triple::AMDHSA) { |
| 238 | switch (TT.getOS()) { |
| 239 | case Triple::AMDHSA: |
| 240 | OSABI = ELF::ELFOSABI_AMDGPU_HSA; |
| 241 | break; |
| 242 | case Triple::AMDPAL: |
| 243 | OSABI = ELF::ELFOSABI_AMDGPU_PAL; |
| 244 | break; |
| 245 | case Triple::Mesa3D: |
| 246 | OSABI = ELF::ELFOSABI_AMDGPU_MESA3D; |
| 247 | break; |
| 248 | default: |
| 249 | break; |
| 250 | } |
| 251 | } |
| 252 | |
| 253 | std::unique_ptr<MCObjectTargetWriter> |
| 254 | createObjectTargetWriter() const override { |
| 255 | return createAMDGPUELFObjectWriter(Is64Bit, OSABI, HasRelocationAddend); |
| 256 | } |
| 257 | }; |
| 258 | |
| 259 | } // end anonymous namespace |
| 260 | |
| 261 | MCAsmBackend *llvm::createAMDGPUAsmBackend(const Target &T, |
| 262 | const MCSubtargetInfo &STI, |
| 263 | const MCRegisterInfo &MRI, |
| 264 | const MCTargetOptions &Options) { |
| 265 | return new ELFAMDGPUAsmBackend(T, STI.getTargetTriple()); |
| 266 | } |
| 267 | |