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 "AMDGPUTargetMachine.h"
20 #include "MCTargetDesc/AMDGPUInstPrinter.h"
21 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
22 #include "llvm/CodeGen/MachineBasicBlock.h"
23 #include "llvm/CodeGen/MachineInstr.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/Function.h"
26 #include "llvm/IR/GlobalVariable.h"
27 #include "llvm/MC/MCCodeEmitter.h"
28 #include "llvm/MC/MCContext.h"
29 #include "llvm/MC/MCExpr.h"
30 #include "llvm/MC/MCInst.h"
31 #include "llvm/MC/MCObjectStreamer.h"
32 #include "llvm/MC/MCStreamer.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/Format.h"
35 #include <algorithm>
36
37 using namespace llvm;
38
39 #include "AMDGPUGenMCPseudoLowering.inc"
40
AMDGPUMCInstLower(MCContext & ctx,const TargetSubtargetInfo & st,const AsmPrinter & ap)41 AMDGPUMCInstLower::AMDGPUMCInstLower(MCContext &ctx,
42 const TargetSubtargetInfo &st,
43 const AsmPrinter &ap):
44 Ctx(ctx), ST(st), AP(ap) { }
45
getVariantKind(unsigned MOFlags)46 static MCSymbolRefExpr::VariantKind getVariantKind(unsigned MOFlags) {
47 switch (MOFlags) {
48 default:
49 return MCSymbolRefExpr::VK_None;
50 case SIInstrInfo::MO_GOTPCREL:
51 return MCSymbolRefExpr::VK_GOTPCREL;
52 case SIInstrInfo::MO_GOTPCREL32_LO:
53 return MCSymbolRefExpr::VK_AMDGPU_GOTPCREL32_LO;
54 case SIInstrInfo::MO_GOTPCREL32_HI:
55 return MCSymbolRefExpr::VK_AMDGPU_GOTPCREL32_HI;
56 case SIInstrInfo::MO_REL32_LO:
57 return MCSymbolRefExpr::VK_AMDGPU_REL32_LO;
58 case SIInstrInfo::MO_REL32_HI:
59 return MCSymbolRefExpr::VK_AMDGPU_REL32_HI;
60 case SIInstrInfo::MO_ABS32_LO:
61 return MCSymbolRefExpr::VK_AMDGPU_ABS32_LO;
62 case SIInstrInfo::MO_ABS32_HI:
63 return MCSymbolRefExpr::VK_AMDGPU_ABS32_HI;
64 }
65 }
66
lowerOperand(const MachineOperand & MO,MCOperand & MCOp) const67 bool AMDGPUMCInstLower::lowerOperand(const MachineOperand &MO,
68 MCOperand &MCOp) const {
69 switch (MO.getType()) {
70 default:
71 break;
72 case MachineOperand::MO_Immediate:
73 MCOp = MCOperand::createImm(MO.getImm());
74 return true;
75 case MachineOperand::MO_Register:
76 MCOp = MCOperand::createReg(AMDGPU::getMCReg(MO.getReg(), ST));
77 return true;
78 case MachineOperand::MO_MachineBasicBlock:
79 MCOp = MCOperand::createExpr(
80 MCSymbolRefExpr::create(MO.getMBB()->getSymbol(), Ctx));
81 return true;
82 case MachineOperand::MO_GlobalAddress: {
83 const GlobalValue *GV = MO.getGlobal();
84 SmallString<128> SymbolName;
85 AP.getNameWithPrefix(SymbolName, GV);
86 MCSymbol *Sym = Ctx.getOrCreateSymbol(SymbolName);
87 const MCExpr *Expr =
88 MCSymbolRefExpr::create(Sym, getVariantKind(MO.getTargetFlags()),Ctx);
89 int64_t Offset = MO.getOffset();
90 if (Offset != 0) {
91 Expr = MCBinaryExpr::createAdd(Expr,
92 MCConstantExpr::create(Offset, Ctx), Ctx);
93 }
94 MCOp = MCOperand::createExpr(Expr);
95 return true;
96 }
97 case MachineOperand::MO_ExternalSymbol: {
98 MCSymbol *Sym = Ctx.getOrCreateSymbol(StringRef(MO.getSymbolName()));
99 const MCSymbolRefExpr *Expr = MCSymbolRefExpr::create(Sym, Ctx);
100 MCOp = MCOperand::createExpr(Expr);
101 return true;
102 }
103 case MachineOperand::MO_RegisterMask:
104 // Regmasks are like implicit defs.
105 return false;
106 case MachineOperand::MO_MCSymbol:
107 if (MO.getTargetFlags() == SIInstrInfo::MO_FAR_BRANCH_OFFSET) {
108 MCSymbol *Sym = MO.getMCSymbol();
109 MCOp = MCOperand::createExpr(Sym->getVariableValue());
110 return true;
111 }
112 break;
113 }
114 llvm_unreachable("unknown operand type");
115 }
116
lower(const MachineInstr * MI,MCInst & OutMI) const117 void AMDGPUMCInstLower::lower(const MachineInstr *MI, MCInst &OutMI) const {
118 unsigned Opcode = MI->getOpcode();
119 const auto *TII = static_cast<const SIInstrInfo*>(ST.getInstrInfo());
120
121 // FIXME: Should be able to handle this with emitPseudoExpansionLowering. We
122 // need to select it to the subtarget specific version, and there's no way to
123 // do that with a single pseudo source operation.
124 if (Opcode == AMDGPU::S_SETPC_B64_return)
125 Opcode = AMDGPU::S_SETPC_B64;
126 else if (Opcode == AMDGPU::SI_CALL) {
127 // SI_CALL is just S_SWAPPC_B64 with an additional operand to track the
128 // called function (which we need to remove here).
129 OutMI.setOpcode(TII->pseudoToMCOpcode(AMDGPU::S_SWAPPC_B64));
130 MCOperand Dest, Src;
131 lowerOperand(MI->getOperand(0), Dest);
132 lowerOperand(MI->getOperand(1), Src);
133 OutMI.addOperand(Dest);
134 OutMI.addOperand(Src);
135 return;
136 } else if (Opcode == AMDGPU::SI_TCRETURN ||
137 Opcode == AMDGPU::SI_TCRETURN_GFX) {
138 // TODO: How to use branch immediate and avoid register+add?
139 Opcode = AMDGPU::S_SETPC_B64;
140 }
141
142 int MCOpcode = TII->pseudoToMCOpcode(Opcode);
143 if (MCOpcode == -1) {
144 LLVMContext &C = MI->getParent()->getParent()->getFunction().getContext();
145 C.emitError("AMDGPUMCInstLower::lower - Pseudo instruction doesn't have "
146 "a target-specific version: " + Twine(MI->getOpcode()));
147 }
148
149 OutMI.setOpcode(MCOpcode);
150
151 for (const MachineOperand &MO : MI->explicit_operands()) {
152 MCOperand MCOp;
153 lowerOperand(MO, MCOp);
154 OutMI.addOperand(MCOp);
155 }
156
157 int FIIdx = AMDGPU::getNamedOperandIdx(MCOpcode, AMDGPU::OpName::fi);
158 if (FIIdx >= (int)OutMI.getNumOperands())
159 OutMI.addOperand(MCOperand::createImm(0));
160 }
161
lowerOperand(const MachineOperand & MO,MCOperand & MCOp) const162 bool AMDGPUAsmPrinter::lowerOperand(const MachineOperand &MO,
163 MCOperand &MCOp) const {
164 const GCNSubtarget &STI = MF->getSubtarget<GCNSubtarget>();
165 AMDGPUMCInstLower MCInstLowering(OutContext, STI, *this);
166 return MCInstLowering.lowerOperand(MO, MCOp);
167 }
168
lowerConstant(const Constant * CV)169 const MCExpr *AMDGPUAsmPrinter::lowerConstant(const Constant *CV) {
170
171 // Intercept LDS variables with known addresses
172 if (const GlobalVariable *GV = dyn_cast<const GlobalVariable>(CV)) {
173 if (std::optional<uint32_t> Address =
174 AMDGPUMachineFunction::getLDSAbsoluteAddress(*GV)) {
175 auto *IntTy = Type::getInt32Ty(CV->getContext());
176 return AsmPrinter::lowerConstant(ConstantInt::get(IntTy, *Address));
177 }
178 }
179
180 if (const MCExpr *E = lowerAddrSpaceCast(TM, CV, OutContext))
181 return E;
182 return AsmPrinter::lowerConstant(CV);
183 }
184
emitInstruction(const MachineInstr * MI)185 void AMDGPUAsmPrinter::emitInstruction(const MachineInstr *MI) {
186 // FIXME: Enable feature predicate checks once all the test pass.
187 // AMDGPU_MC::verifyInstructionPredicates(MI->getOpcode(),
188 // getSubtargetInfo().getFeatureBits());
189
190 if (emitPseudoExpansionLowering(*OutStreamer, MI))
191 return;
192
193 const GCNSubtarget &STI = MF->getSubtarget<GCNSubtarget>();
194 AMDGPUMCInstLower MCInstLowering(OutContext, STI, *this);
195
196 StringRef Err;
197 if (!STI.getInstrInfo()->verifyInstruction(*MI, Err)) {
198 LLVMContext &C = MI->getParent()->getParent()->getFunction().getContext();
199 C.emitError("Illegal instruction detected: " + Err);
200 MI->print(errs());
201 }
202
203 if (MI->isBundle()) {
204 const MachineBasicBlock *MBB = MI->getParent();
205 MachineBasicBlock::const_instr_iterator I = ++MI->getIterator();
206 while (I != MBB->instr_end() && I->isInsideBundle()) {
207 emitInstruction(&*I);
208 ++I;
209 }
210 } else {
211 // We don't want these pseudo instructions encoded. They are
212 // placeholder terminator instructions and should only be printed as
213 // comments.
214 if (MI->getOpcode() == AMDGPU::SI_RETURN_TO_EPILOG) {
215 if (isVerbose())
216 OutStreamer->emitRawComment(" return to shader part epilog");
217 return;
218 }
219
220 if (MI->getOpcode() == AMDGPU::WAVE_BARRIER) {
221 if (isVerbose())
222 OutStreamer->emitRawComment(" wave barrier");
223 return;
224 }
225
226 if (MI->getOpcode() == AMDGPU::SCHED_BARRIER) {
227 if (isVerbose()) {
228 std::string HexString;
229 raw_string_ostream HexStream(HexString);
230 HexStream << format_hex(MI->getOperand(0).getImm(), 10, true);
231 OutStreamer->emitRawComment(" sched_barrier mask(" + HexString + ")");
232 }
233 return;
234 }
235
236 if (MI->getOpcode() == AMDGPU::SCHED_GROUP_BARRIER) {
237 if (isVerbose()) {
238 std::string HexString;
239 raw_string_ostream HexStream(HexString);
240 HexStream << format_hex(MI->getOperand(0).getImm(), 10, true);
241 OutStreamer->emitRawComment(
242 " sched_group_barrier mask(" + HexString + ") size(" +
243 Twine(MI->getOperand(1).getImm()) + ") SyncID(" +
244 Twine(MI->getOperand(2).getImm()) + ")");
245 }
246 return;
247 }
248
249 if (MI->getOpcode() == AMDGPU::IGLP_OPT) {
250 if (isVerbose()) {
251 std::string HexString;
252 raw_string_ostream HexStream(HexString);
253 HexStream << format_hex(MI->getOperand(0).getImm(), 10, true);
254 OutStreamer->emitRawComment(" iglp_opt mask(" + HexString + ")");
255 }
256 return;
257 }
258
259 if (MI->getOpcode() == AMDGPU::SI_MASKED_UNREACHABLE) {
260 if (isVerbose())
261 OutStreamer->emitRawComment(" divergent unreachable");
262 return;
263 }
264
265 if (MI->isMetaInstruction()) {
266 if (isVerbose())
267 OutStreamer->emitRawComment(" meta instruction");
268 return;
269 }
270
271 MCInst TmpInst;
272 MCInstLowering.lower(MI, TmpInst);
273 EmitToStreamer(*OutStreamer, TmpInst);
274
275 #ifdef EXPENSIVE_CHECKS
276 // Check getInstSizeInBytes on explicitly specified CPUs (it cannot
277 // work correctly for the generic CPU).
278 //
279 // The isPseudo check really shouldn't be here, but unfortunately there are
280 // some negative lit tests that depend on being able to continue through
281 // here even when pseudo instructions haven't been lowered.
282 //
283 // We also overestimate branch sizes with the offset bug.
284 if (!MI->isPseudo() && STI.isCPUStringValid(STI.getCPU()) &&
285 (!STI.hasOffset3fBug() || !MI->isBranch())) {
286 SmallVector<MCFixup, 4> Fixups;
287 SmallVector<char, 16> CodeBytes;
288
289 std::unique_ptr<MCCodeEmitter> InstEmitter(createAMDGPUMCCodeEmitter(
290 *STI.getInstrInfo(), OutContext));
291 InstEmitter->encodeInstruction(TmpInst, CodeBytes, Fixups, STI);
292
293 assert(CodeBytes.size() == STI.getInstrInfo()->getInstSizeInBytes(*MI));
294 }
295 #endif
296
297 if (DumpCodeInstEmitter) {
298 // Disassemble instruction/operands to text
299 DisasmLines.resize(DisasmLines.size() + 1);
300 std::string &DisasmLine = DisasmLines.back();
301 raw_string_ostream DisasmStream(DisasmLine);
302
303 AMDGPUInstPrinter InstPrinter(*TM.getMCAsmInfo(), *STI.getInstrInfo(),
304 *STI.getRegisterInfo());
305 InstPrinter.printInst(&TmpInst, 0, StringRef(), STI, DisasmStream);
306
307 // Disassemble instruction/operands to hex representation.
308 SmallVector<MCFixup, 4> Fixups;
309 SmallVector<char, 16> CodeBytes;
310
311 DumpCodeInstEmitter->encodeInstruction(
312 TmpInst, CodeBytes, Fixups, MF->getSubtarget<MCSubtargetInfo>());
313 HexLines.resize(HexLines.size() + 1);
314 std::string &HexLine = HexLines.back();
315 raw_string_ostream HexStream(HexLine);
316
317 for (size_t i = 0; i < CodeBytes.size(); i += 4) {
318 unsigned int CodeDWord = *(unsigned int *)&CodeBytes[i];
319 HexStream << format("%s%08X", (i > 0 ? " " : ""), CodeDWord);
320 }
321
322 DisasmStream.flush();
323 DisasmLineMaxLen = std::max(DisasmLineMaxLen, DisasmLine.size());
324 }
325 }
326 }
327