[x86] fix assert with horizontal math + broadcast of vector (PR43402)
[llvm-core.git] / lib / Target / X86 / X86RegisterBankInfo.cpp
blobdaddf4231897dc0b7e4e8eaccdc0507a508459ba
1 //===- X86RegisterBankInfo.cpp -----------------------------------*- C++ -*-==//
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 /// \file
9 /// This file implements the targeting of the RegisterBankInfo class for X86.
10 /// \todo This should be generated by TableGen.
11 //===----------------------------------------------------------------------===//
13 #include "X86RegisterBankInfo.h"
14 #include "X86InstrInfo.h"
15 #include "llvm/CodeGen/GlobalISel/RegisterBank.h"
16 #include "llvm/CodeGen/GlobalISel/RegisterBankInfo.h"
17 #include "llvm/CodeGen/MachineRegisterInfo.h"
18 #include "llvm/CodeGen/TargetRegisterInfo.h"
20 #define GET_TARGET_REGBANK_IMPL
21 #include "X86GenRegisterBank.inc"
23 using namespace llvm;
24 // This file will be TableGen'ed at some point.
25 #define GET_TARGET_REGBANK_INFO_IMPL
26 #include "X86GenRegisterBankInfo.def"
28 X86RegisterBankInfo::X86RegisterBankInfo(const TargetRegisterInfo &TRI)
29 : X86GenRegisterBankInfo() {
31 // validate RegBank initialization.
32 const RegisterBank &RBGPR = getRegBank(X86::GPRRegBankID);
33 (void)RBGPR;
34 assert(&X86::GPRRegBank == &RBGPR && "Incorrect RegBanks inizalization.");
36 // The GPR register bank is fully defined by all the registers in
37 // GR64 + its subclasses.
38 assert(RBGPR.covers(*TRI.getRegClass(X86::GR64RegClassID)) &&
39 "Subclass not added?");
40 assert(RBGPR.getSize() == 64 && "GPRs should hold up to 64-bit");
43 const RegisterBank &X86RegisterBankInfo::getRegBankFromRegClass(
44 const TargetRegisterClass &RC) const {
46 if (X86::GR8RegClass.hasSubClassEq(&RC) ||
47 X86::GR16RegClass.hasSubClassEq(&RC) ||
48 X86::GR32RegClass.hasSubClassEq(&RC) ||
49 X86::GR64RegClass.hasSubClassEq(&RC) ||
50 X86::LOW32_ADDR_ACCESSRegClass.hasSubClassEq(&RC) ||
51 X86::LOW32_ADDR_ACCESS_RBPRegClass.hasSubClassEq(&RC))
52 return getRegBank(X86::GPRRegBankID);
54 if (X86::FR32XRegClass.hasSubClassEq(&RC) ||
55 X86::FR64XRegClass.hasSubClassEq(&RC) ||
56 X86::VR128XRegClass.hasSubClassEq(&RC) ||
57 X86::VR256XRegClass.hasSubClassEq(&RC) ||
58 X86::VR512RegClass.hasSubClassEq(&RC))
59 return getRegBank(X86::VECRRegBankID);
61 llvm_unreachable("Unsupported register kind yet.");
64 X86GenRegisterBankInfo::PartialMappingIdx
65 X86GenRegisterBankInfo::getPartialMappingIdx(const LLT &Ty, bool isFP) {
66 if ((Ty.isScalar() && !isFP) || Ty.isPointer()) {
67 switch (Ty.getSizeInBits()) {
68 case 1:
69 case 8:
70 return PMI_GPR8;
71 case 16:
72 return PMI_GPR16;
73 case 32:
74 return PMI_GPR32;
75 case 64:
76 return PMI_GPR64;
77 case 128:
78 return PMI_VEC128;
79 break;
80 default:
81 llvm_unreachable("Unsupported register size.");
83 } else if (Ty.isScalar()) {
84 switch (Ty.getSizeInBits()) {
85 case 32:
86 return PMI_FP32;
87 case 64:
88 return PMI_FP64;
89 case 128:
90 return PMI_VEC128;
91 default:
92 llvm_unreachable("Unsupported register size.");
94 } else {
95 switch (Ty.getSizeInBits()) {
96 case 128:
97 return PMI_VEC128;
98 case 256:
99 return PMI_VEC256;
100 case 512:
101 return PMI_VEC512;
102 default:
103 llvm_unreachable("Unsupported register size.");
107 return PMI_None;
110 void X86RegisterBankInfo::getInstrPartialMappingIdxs(
111 const MachineInstr &MI, const MachineRegisterInfo &MRI, const bool isFP,
112 SmallVectorImpl<PartialMappingIdx> &OpRegBankIdx) {
114 unsigned NumOperands = MI.getNumOperands();
115 for (unsigned Idx = 0; Idx < NumOperands; ++Idx) {
116 auto &MO = MI.getOperand(Idx);
117 if (!MO.isReg())
118 OpRegBankIdx[Idx] = PMI_None;
119 else
120 OpRegBankIdx[Idx] = getPartialMappingIdx(MRI.getType(MO.getReg()), isFP);
124 bool X86RegisterBankInfo::getInstrValueMapping(
125 const MachineInstr &MI,
126 const SmallVectorImpl<PartialMappingIdx> &OpRegBankIdx,
127 SmallVectorImpl<const ValueMapping *> &OpdsMapping) {
129 unsigned NumOperands = MI.getNumOperands();
130 for (unsigned Idx = 0; Idx < NumOperands; ++Idx) {
131 if (!MI.getOperand(Idx).isReg())
132 continue;
134 auto Mapping = getValueMapping(OpRegBankIdx[Idx], 1);
135 if (!Mapping->isValid())
136 return false;
138 OpdsMapping[Idx] = Mapping;
140 return true;
143 const RegisterBankInfo::InstructionMapping &
144 X86RegisterBankInfo::getSameOperandsMapping(const MachineInstr &MI,
145 bool isFP) const {
146 const MachineFunction &MF = *MI.getParent()->getParent();
147 const MachineRegisterInfo &MRI = MF.getRegInfo();
149 unsigned NumOperands = MI.getNumOperands();
150 LLT Ty = MRI.getType(MI.getOperand(0).getReg());
152 if (NumOperands != 3 || (Ty != MRI.getType(MI.getOperand(1).getReg())) ||
153 (Ty != MRI.getType(MI.getOperand(2).getReg())))
154 llvm_unreachable("Unsupported operand mapping yet.");
156 auto Mapping = getValueMapping(getPartialMappingIdx(Ty, isFP), 3);
157 return getInstructionMapping(DefaultMappingID, 1, Mapping, NumOperands);
160 const RegisterBankInfo::InstructionMapping &
161 X86RegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
162 const MachineFunction &MF = *MI.getParent()->getParent();
163 const MachineRegisterInfo &MRI = MF.getRegInfo();
164 unsigned Opc = MI.getOpcode();
166 // Try the default logic for non-generic instructions that are either copies
167 // or already have some operands assigned to banks.
168 if (!isPreISelGenericOpcode(Opc) || Opc == TargetOpcode::G_PHI) {
169 const InstructionMapping &Mapping = getInstrMappingImpl(MI);
170 if (Mapping.isValid())
171 return Mapping;
174 switch (Opc) {
175 case TargetOpcode::G_ADD:
176 case TargetOpcode::G_SUB:
177 case TargetOpcode::G_MUL:
178 return getSameOperandsMapping(MI, false);
179 case TargetOpcode::G_FADD:
180 case TargetOpcode::G_FSUB:
181 case TargetOpcode::G_FMUL:
182 case TargetOpcode::G_FDIV:
183 return getSameOperandsMapping(MI, true);
184 case TargetOpcode::G_SHL:
185 case TargetOpcode::G_LSHR:
186 case TargetOpcode::G_ASHR: {
187 unsigned NumOperands = MI.getNumOperands();
188 LLT Ty = MRI.getType(MI.getOperand(0).getReg());
190 auto Mapping = getValueMapping(getPartialMappingIdx(Ty, false), 3);
191 return getInstructionMapping(DefaultMappingID, 1, Mapping, NumOperands);
194 default:
195 break;
198 unsigned NumOperands = MI.getNumOperands();
199 SmallVector<PartialMappingIdx, 4> OpRegBankIdx(NumOperands);
201 switch (Opc) {
202 case TargetOpcode::G_FPEXT:
203 case TargetOpcode::G_FPTRUNC:
204 case TargetOpcode::G_FCONSTANT:
205 // Instruction having only floating-point operands (all scalars in VECRReg)
206 getInstrPartialMappingIdxs(MI, MRI, /* isFP */ true, OpRegBankIdx);
207 break;
208 case TargetOpcode::G_SITOFP:
209 case TargetOpcode::G_FPTOSI: {
210 // Some of the floating-point instructions have mixed GPR and FP operands:
211 // fine-tune the computed mapping.
212 auto &Op0 = MI.getOperand(0);
213 auto &Op1 = MI.getOperand(1);
214 const LLT Ty0 = MRI.getType(Op0.getReg());
215 const LLT Ty1 = MRI.getType(Op1.getReg());
217 bool FirstArgIsFP = Opc == TargetOpcode::G_SITOFP;
218 bool SecondArgIsFP = Opc == TargetOpcode::G_FPTOSI;
219 OpRegBankIdx[0] = getPartialMappingIdx(Ty0, /* isFP */ FirstArgIsFP);
220 OpRegBankIdx[1] = getPartialMappingIdx(Ty1, /* isFP */ SecondArgIsFP);
221 break;
223 case TargetOpcode::G_FCMP: {
224 LLT Ty1 = MRI.getType(MI.getOperand(2).getReg());
225 LLT Ty2 = MRI.getType(MI.getOperand(3).getReg());
226 (void)Ty2;
227 assert(Ty1.getSizeInBits() == Ty2.getSizeInBits() &&
228 "Mismatched operand sizes for G_FCMP");
230 unsigned Size = Ty1.getSizeInBits();
231 (void)Size;
232 assert((Size == 32 || Size == 64) && "Unsupported size for G_FCMP");
234 auto FpRegBank = getPartialMappingIdx(Ty1, /* isFP */ true);
235 OpRegBankIdx = {PMI_GPR8,
236 /* Predicate */ PMI_None, FpRegBank, FpRegBank};
237 break;
239 case TargetOpcode::G_TRUNC:
240 case TargetOpcode::G_ANYEXT: {
241 auto &Op0 = MI.getOperand(0);
242 auto &Op1 = MI.getOperand(1);
243 const LLT Ty0 = MRI.getType(Op0.getReg());
244 const LLT Ty1 = MRI.getType(Op1.getReg());
246 bool isFPTrunc = (Ty0.getSizeInBits() == 32 || Ty0.getSizeInBits() == 64) &&
247 Ty1.getSizeInBits() == 128 && Opc == TargetOpcode::G_TRUNC;
248 bool isFPAnyExt =
249 Ty0.getSizeInBits() == 128 &&
250 (Ty1.getSizeInBits() == 32 || Ty1.getSizeInBits() == 64) &&
251 Opc == TargetOpcode::G_ANYEXT;
253 getInstrPartialMappingIdxs(MI, MRI, /* isFP */ isFPTrunc || isFPAnyExt,
254 OpRegBankIdx);
255 } break;
256 default:
257 // Track the bank of each register, use NotFP mapping (all scalars in GPRs)
258 getInstrPartialMappingIdxs(MI, MRI, /* isFP */ false, OpRegBankIdx);
259 break;
262 // Finally construct the computed mapping.
263 SmallVector<const ValueMapping *, 8> OpdsMapping(NumOperands);
264 if (!getInstrValueMapping(MI, OpRegBankIdx, OpdsMapping))
265 return getInvalidInstructionMapping();
267 return getInstructionMapping(DefaultMappingID, /* Cost */ 1,
268 getOperandsMapping(OpdsMapping), NumOperands);
271 void X86RegisterBankInfo::applyMappingImpl(
272 const OperandsMapper &OpdMapper) const {
273 return applyDefaultMapping(OpdMapper);
276 RegisterBankInfo::InstructionMappings
277 X86RegisterBankInfo::getInstrAlternativeMappings(const MachineInstr &MI) const {
279 const MachineFunction &MF = *MI.getParent()->getParent();
280 const TargetSubtargetInfo &STI = MF.getSubtarget();
281 const TargetRegisterInfo &TRI = *STI.getRegisterInfo();
282 const MachineRegisterInfo &MRI = MF.getRegInfo();
284 switch (MI.getOpcode()) {
285 case TargetOpcode::G_LOAD:
286 case TargetOpcode::G_STORE:
287 case TargetOpcode::G_IMPLICIT_DEF: {
288 // we going to try to map 32/64 bit to PMI_FP32/PMI_FP64
289 unsigned Size = getSizeInBits(MI.getOperand(0).getReg(), MRI, TRI);
290 if (Size != 32 && Size != 64)
291 break;
293 unsigned NumOperands = MI.getNumOperands();
295 // Track the bank of each register, use FP mapping (all scalars in VEC)
296 SmallVector<PartialMappingIdx, 4> OpRegBankIdx(NumOperands);
297 getInstrPartialMappingIdxs(MI, MRI, /* isFP */ true, OpRegBankIdx);
299 // Finally construct the computed mapping.
300 SmallVector<const ValueMapping *, 8> OpdsMapping(NumOperands);
301 if (!getInstrValueMapping(MI, OpRegBankIdx, OpdsMapping))
302 break;
304 const RegisterBankInfo::InstructionMapping &Mapping = getInstructionMapping(
305 /*ID*/ 1, /*Cost*/ 1, getOperandsMapping(OpdsMapping), NumOperands);
306 InstructionMappings AltMappings;
307 AltMappings.push_back(&Mapping);
308 return AltMappings;
310 default:
311 break;
313 return RegisterBankInfo::getInstrAlternativeMappings(MI);