1 //===- AMDGPUInstructionSelector.cpp ----------------------------*- C++ -*-==//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This file implements the targeting of the InstructionSelector class for
12 /// \todo This should be generated by TableGen.
13 //===----------------------------------------------------------------------===//
15 #include "AMDGPUInstructionSelector.h"
16 #include "AMDGPUInstrInfo.h"
17 #include "AMDGPURegisterBankInfo.h"
18 #include "AMDGPURegisterInfo.h"
19 #include "AMDGPUSubtarget.h"
20 #include "llvm/CodeGen/MachineBasicBlock.h"
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineInstr.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/IR/Type.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/raw_ostream.h"
29 #define DEBUG_TYPE "amdgpu-isel"
33 AMDGPUInstructionSelector::AMDGPUInstructionSelector(
34 const SISubtarget &STI, const AMDGPURegisterBankInfo &RBI)
35 : InstructionSelector(), TII(*STI.getInstrInfo()),
36 TRI(*STI.getRegisterInfo()), RBI(RBI), AMDGPUASI(STI.getAMDGPUAS()) {}
39 AMDGPUInstructionSelector::getSubOperand64(MachineOperand &MO,
40 unsigned SubIdx) const {
42 MachineInstr *MI = MO.getParent();
43 MachineBasicBlock *BB = MO.getParent()->getParent();
44 MachineFunction *MF = BB->getParent();
45 MachineRegisterInfo &MRI = MF->getRegInfo();
46 unsigned DstReg = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
49 unsigned ComposedSubIdx = TRI.composeSubRegIndices(MO.getSubReg(), SubIdx);
50 unsigned Reg = MO.getReg();
51 BuildMI(*BB, MI, MI->getDebugLoc(), TII.get(AMDGPU::COPY), DstReg)
52 .addReg(Reg, 0, ComposedSubIdx);
54 return MachineOperand::CreateReg(DstReg, MO.isDef(), MO.isImplicit(),
55 MO.isKill(), MO.isDead(), MO.isUndef(),
56 MO.isEarlyClobber(), 0, MO.isDebug(),
62 APInt Imm(64, MO.getImm());
66 llvm_unreachable("do not know to split immediate with this sub index.");
68 return MachineOperand::CreateImm(Imm.getLoBits(32).getSExtValue());
70 return MachineOperand::CreateImm(Imm.getHiBits(32).getSExtValue());
74 bool AMDGPUInstructionSelector::selectG_ADD(MachineInstr &I) const {
75 MachineBasicBlock *BB = I.getParent();
76 MachineFunction *MF = BB->getParent();
77 MachineRegisterInfo &MRI = MF->getRegInfo();
78 unsigned Size = RBI.getSizeInBits(I.getOperand(0).getReg(), MRI, TRI);
79 unsigned DstLo = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
80 unsigned DstHi = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
85 DebugLoc DL = I.getDebugLoc();
87 MachineOperand Lo1(getSubOperand64(I.getOperand(1), AMDGPU::sub0));
88 MachineOperand Lo2(getSubOperand64(I.getOperand(2), AMDGPU::sub0));
90 BuildMI(*BB, &I, DL, TII.get(AMDGPU::S_ADD_U32), DstLo)
94 MachineOperand Hi1(getSubOperand64(I.getOperand(1), AMDGPU::sub1));
95 MachineOperand Hi2(getSubOperand64(I.getOperand(2), AMDGPU::sub1));
97 BuildMI(*BB, &I, DL, TII.get(AMDGPU::S_ADDC_U32), DstHi)
101 BuildMI(*BB, &I, DL, TII.get(AMDGPU::REG_SEQUENCE), I.getOperand(0).getReg())
103 .addImm(AMDGPU::sub0)
105 .addImm(AMDGPU::sub1);
107 for (MachineOperand &MO : I.explicit_operands()) {
108 if (!MO.isReg() || TargetRegisterInfo::isPhysicalRegister(MO.getReg()))
110 RBI.constrainGenericRegister(MO.getReg(), AMDGPU::SReg_64RegClass, MRI);
117 bool AMDGPUInstructionSelector::selectG_GEP(MachineInstr &I) const {
118 return selectG_ADD(I);
121 bool AMDGPUInstructionSelector::selectG_STORE(MachineInstr &I) const {
122 MachineBasicBlock *BB = I.getParent();
123 DebugLoc DL = I.getDebugLoc();
125 // FIXME: Select store instruction based on address space
126 MachineInstr *Flat = BuildMI(*BB, &I, DL, TII.get(AMDGPU::FLAT_STORE_DWORD))
127 .add(I.getOperand(1))
128 .add(I.getOperand(0))
134 // Now that we selected an opcode, we need to constrain the register
135 // operands to use appropriate classes.
136 bool Ret = constrainSelectedInstRegOperands(*Flat, TII, TRI, RBI);
142 bool AMDGPUInstructionSelector::selectG_CONSTANT(MachineInstr &I) const {
143 MachineBasicBlock *BB = I.getParent();
144 MachineFunction *MF = BB->getParent();
145 MachineRegisterInfo &MRI = MF->getRegInfo();
146 unsigned DstReg = I.getOperand(0).getReg();
147 unsigned Size = RBI.getSizeInBits(DstReg, MRI, TRI);
150 I.setDesc(TII.get(AMDGPU::S_MOV_B32));
151 return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
156 DebugLoc DL = I.getDebugLoc();
157 unsigned LoReg = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
158 unsigned HiReg = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
159 const APInt &Imm = I.getOperand(1).getCImm()->getValue();
161 BuildMI(*BB, &I, DL, TII.get(AMDGPU::S_MOV_B32), LoReg)
162 .addImm(Imm.trunc(32).getZExtValue());
164 BuildMI(*BB, &I, DL, TII.get(AMDGPU::S_MOV_B32), HiReg)
165 .addImm(Imm.ashr(32).getZExtValue());
167 BuildMI(*BB, &I, DL, TII.get(AMDGPU::REG_SEQUENCE), DstReg)
169 .addImm(AMDGPU::sub0)
171 .addImm(AMDGPU::sub1);
172 // We can't call constrainSelectedInstRegOperands here, because it doesn't
173 // work for target independent opcodes
175 return RBI.constrainGenericRegister(DstReg, AMDGPU::SReg_64RegClass, MRI);
178 static bool isConstant(const MachineInstr &MI) {
179 return MI.getOpcode() == TargetOpcode::G_CONSTANT;
182 void AMDGPUInstructionSelector::getAddrModeInfo(const MachineInstr &Load,
183 const MachineRegisterInfo &MRI, SmallVectorImpl<GEPInfo> &AddrInfo) const {
185 const MachineInstr *PtrMI = MRI.getUniqueVRegDef(Load.getOperand(1).getReg());
189 if (PtrMI->getOpcode() != TargetOpcode::G_GEP)
192 GEPInfo GEPInfo(*PtrMI);
194 for (unsigned i = 1, e = 3; i < e; ++i) {
195 const MachineOperand &GEPOp = PtrMI->getOperand(i);
196 const MachineInstr *OpDef = MRI.getUniqueVRegDef(GEPOp.getReg());
198 if (isConstant(*OpDef)) {
199 // FIXME: Is it possible to have multiple Imm parts? Maybe if we
200 // are lacking other optimizations.
201 assert(GEPInfo.Imm == 0);
202 GEPInfo.Imm = OpDef->getOperand(1).getCImm()->getSExtValue();
205 const RegisterBank *OpBank = RBI.getRegBank(GEPOp.getReg(), MRI, TRI);
206 if (OpBank->getID() == AMDGPU::SGPRRegBankID)
207 GEPInfo.SgprParts.push_back(GEPOp.getReg());
209 GEPInfo.VgprParts.push_back(GEPOp.getReg());
212 AddrInfo.push_back(GEPInfo);
213 getAddrModeInfo(*PtrMI, MRI, AddrInfo);
216 static bool isInstrUniform(const MachineInstr &MI) {
217 if (!MI.hasOneMemOperand())
220 const MachineMemOperand *MMO = *MI.memoperands_begin();
221 const Value *Ptr = MMO->getValue();
223 // UndefValue means this is a load of a kernel input. These are uniform.
224 // Sometimes LDS instructions have constant pointers.
225 // If Ptr is null, then that means this mem operand contains a
226 // PseudoSourceValue like GOT.
227 if (!Ptr || isa<UndefValue>(Ptr) || isa<Argument>(Ptr) ||
228 isa<Constant>(Ptr) || isa<GlobalValue>(Ptr))
231 const Instruction *I = dyn_cast<Instruction>(Ptr);
232 return I && I->getMetadata("amdgpu.uniform");
235 static unsigned getSmrdOpcode(unsigned BaseOpcode, unsigned LoadSize) {
240 switch (BaseOpcode) {
241 case AMDGPU::S_LOAD_DWORD_IMM:
244 return AMDGPU::S_LOAD_DWORDX2_IMM;
246 return AMDGPU::S_LOAD_DWORDX4_IMM;
248 return AMDGPU::S_LOAD_DWORDX8_IMM;
250 return AMDGPU::S_LOAD_DWORDX16_IMM;
253 case AMDGPU::S_LOAD_DWORD_IMM_ci:
256 return AMDGPU::S_LOAD_DWORDX2_IMM_ci;
258 return AMDGPU::S_LOAD_DWORDX4_IMM_ci;
260 return AMDGPU::S_LOAD_DWORDX8_IMM_ci;
262 return AMDGPU::S_LOAD_DWORDX16_IMM_ci;
265 case AMDGPU::S_LOAD_DWORD_SGPR:
268 return AMDGPU::S_LOAD_DWORDX2_SGPR;
270 return AMDGPU::S_LOAD_DWORDX4_SGPR;
272 return AMDGPU::S_LOAD_DWORDX8_SGPR;
274 return AMDGPU::S_LOAD_DWORDX16_SGPR;
278 llvm_unreachable("Invalid base smrd opcode or size");
281 bool AMDGPUInstructionSelector::hasVgprParts(ArrayRef<GEPInfo> AddrInfo) const {
282 for (const GEPInfo &GEPInfo : AddrInfo) {
283 if (!GEPInfo.VgprParts.empty())
289 bool AMDGPUInstructionSelector::selectSMRD(MachineInstr &I,
290 ArrayRef<GEPInfo> AddrInfo) const {
292 if (!I.hasOneMemOperand())
295 if ((*I.memoperands_begin())->getAddrSpace() != AMDGPUASI.CONSTANT_ADDRESS)
298 if (!isInstrUniform(I))
301 if (hasVgprParts(AddrInfo))
304 MachineBasicBlock *BB = I.getParent();
305 MachineFunction *MF = BB->getParent();
306 const SISubtarget &Subtarget = MF->getSubtarget<SISubtarget>();
307 MachineRegisterInfo &MRI = MF->getRegInfo();
308 unsigned DstReg = I.getOperand(0).getReg();
309 const DebugLoc &DL = I.getDebugLoc();
311 unsigned LoadSize = RBI.getSizeInBits(DstReg, MRI, TRI);
313 if (!AddrInfo.empty() && AddrInfo[0].SgprParts.size() == 1) {
315 const GEPInfo &GEPInfo = AddrInfo[0];
317 unsigned PtrReg = GEPInfo.SgprParts[0];
318 int64_t EncodedImm = AMDGPU::getSMRDEncodedOffset(Subtarget, GEPInfo.Imm);
319 if (AMDGPU::isLegalSMRDImmOffset(Subtarget, GEPInfo.Imm)) {
320 Opcode = getSmrdOpcode(AMDGPU::S_LOAD_DWORD_IMM, LoadSize);
322 MachineInstr *SMRD = BuildMI(*BB, &I, DL, TII.get(Opcode), DstReg)
326 return constrainSelectedInstRegOperands(*SMRD, TII, TRI, RBI);
329 if (Subtarget.getGeneration() == AMDGPUSubtarget::SEA_ISLANDS &&
330 isUInt<32>(EncodedImm)) {
331 Opcode = getSmrdOpcode(AMDGPU::S_LOAD_DWORD_IMM_ci, LoadSize);
332 MachineInstr *SMRD = BuildMI(*BB, &I, DL, TII.get(Opcode), DstReg)
336 return constrainSelectedInstRegOperands(*SMRD, TII, TRI, RBI);
339 if (isUInt<32>(GEPInfo.Imm)) {
340 Opcode = getSmrdOpcode(AMDGPU::S_LOAD_DWORD_SGPR, LoadSize);
341 unsigned OffsetReg = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
342 BuildMI(*BB, &I, DL, TII.get(AMDGPU::S_MOV_B32), OffsetReg)
343 .addImm(GEPInfo.Imm);
345 MachineInstr *SMRD = BuildMI(*BB, &I, DL, TII.get(Opcode), DstReg)
349 return constrainSelectedInstRegOperands(*SMRD, TII, TRI, RBI);
353 unsigned PtrReg = I.getOperand(1).getReg();
354 Opcode = getSmrdOpcode(AMDGPU::S_LOAD_DWORD_IMM, LoadSize);
355 MachineInstr *SMRD = BuildMI(*BB, &I, DL, TII.get(Opcode), DstReg)
359 return constrainSelectedInstRegOperands(*SMRD, TII, TRI, RBI);
363 bool AMDGPUInstructionSelector::selectG_LOAD(MachineInstr &I) const {
364 MachineBasicBlock *BB = I.getParent();
365 MachineFunction *MF = BB->getParent();
366 MachineRegisterInfo &MRI = MF->getRegInfo();
367 DebugLoc DL = I.getDebugLoc();
368 unsigned DstReg = I.getOperand(0).getReg();
369 unsigned PtrReg = I.getOperand(1).getReg();
370 unsigned LoadSize = RBI.getSizeInBits(DstReg, MRI, TRI);
373 SmallVector<GEPInfo, 4> AddrInfo;
375 getAddrModeInfo(I, MRI, AddrInfo);
377 if (selectSMRD(I, AddrInfo)) {
384 llvm_unreachable("Load size not supported\n");
386 Opcode = AMDGPU::FLAT_LOAD_DWORD;
389 Opcode = AMDGPU::FLAT_LOAD_DWORDX2;
393 MachineInstr *Flat = BuildMI(*BB, &I, DL, TII.get(Opcode))
394 .add(I.getOperand(0))
400 bool Ret = constrainSelectedInstRegOperands(*Flat, TII, TRI, RBI);
405 bool AMDGPUInstructionSelector::select(MachineInstr &I) const {
407 if (!isPreISelGenericOpcode(I.getOpcode()))
410 switch (I.getOpcode()) {
413 case TargetOpcode::G_ADD:
414 return selectG_ADD(I);
415 case TargetOpcode::G_CONSTANT:
416 return selectG_CONSTANT(I);
417 case TargetOpcode::G_GEP:
418 return selectG_GEP(I);
419 case TargetOpcode::G_LOAD:
420 return selectG_LOAD(I);
421 case TargetOpcode::G_STORE:
422 return selectG_STORE(I);