[x86] fix assert with horizontal math + broadcast of vector (PR43402)
[llvm-core.git] / lib / CodeGen / GlobalISel / Utils.cpp
bloba93e5153089d08cdaca1871846c177ddb4e37a28
1 //===- llvm/CodeGen/GlobalISel/Utils.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 This file implements the utility functions used by the GlobalISel
9 /// pipeline.
10 //===----------------------------------------------------------------------===//
12 #include "llvm/CodeGen/GlobalISel/Utils.h"
13 #include "llvm/ADT/APFloat.h"
14 #include "llvm/ADT/Twine.h"
15 #include "llvm/CodeGen/GlobalISel/RegisterBankInfo.h"
16 #include "llvm/CodeGen/MachineInstr.h"
17 #include "llvm/CodeGen/MachineInstrBuilder.h"
18 #include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h"
19 #include "llvm/CodeGen/MachineRegisterInfo.h"
20 #include "llvm/CodeGen/StackProtector.h"
21 #include "llvm/CodeGen/TargetInstrInfo.h"
22 #include "llvm/CodeGen/TargetPassConfig.h"
23 #include "llvm/CodeGen/TargetRegisterInfo.h"
24 #include "llvm/IR/Constants.h"
26 #define DEBUG_TYPE "globalisel-utils"
28 using namespace llvm;
30 unsigned llvm::constrainRegToClass(MachineRegisterInfo &MRI,
31 const TargetInstrInfo &TII,
32 const RegisterBankInfo &RBI, unsigned Reg,
33 const TargetRegisterClass &RegClass) {
34 if (!RBI.constrainGenericRegister(Reg, RegClass, MRI))
35 return MRI.createVirtualRegister(&RegClass);
37 return Reg;
40 unsigned llvm::constrainOperandRegClass(
41 const MachineFunction &MF, const TargetRegisterInfo &TRI,
42 MachineRegisterInfo &MRI, const TargetInstrInfo &TII,
43 const RegisterBankInfo &RBI, MachineInstr &InsertPt,
44 const TargetRegisterClass &RegClass, const MachineOperand &RegMO,
45 unsigned OpIdx) {
46 Register Reg = RegMO.getReg();
47 // Assume physical registers are properly constrained.
48 assert(Register::isVirtualRegister(Reg) && "PhysReg not implemented");
50 unsigned ConstrainedReg = constrainRegToClass(MRI, TII, RBI, Reg, RegClass);
51 // If we created a new virtual register because the class is not compatible
52 // then create a copy between the new and the old register.
53 if (ConstrainedReg != Reg) {
54 MachineBasicBlock::iterator InsertIt(&InsertPt);
55 MachineBasicBlock &MBB = *InsertPt.getParent();
56 if (RegMO.isUse()) {
57 BuildMI(MBB, InsertIt, InsertPt.getDebugLoc(),
58 TII.get(TargetOpcode::COPY), ConstrainedReg)
59 .addReg(Reg);
60 } else {
61 assert(RegMO.isDef() && "Must be a definition");
62 BuildMI(MBB, std::next(InsertIt), InsertPt.getDebugLoc(),
63 TII.get(TargetOpcode::COPY), Reg)
64 .addReg(ConstrainedReg);
67 return ConstrainedReg;
70 unsigned llvm::constrainOperandRegClass(
71 const MachineFunction &MF, const TargetRegisterInfo &TRI,
72 MachineRegisterInfo &MRI, const TargetInstrInfo &TII,
73 const RegisterBankInfo &RBI, MachineInstr &InsertPt, const MCInstrDesc &II,
74 const MachineOperand &RegMO, unsigned OpIdx) {
75 Register Reg = RegMO.getReg();
76 // Assume physical registers are properly constrained.
77 assert(Register::isVirtualRegister(Reg) && "PhysReg not implemented");
79 const TargetRegisterClass *RegClass = TII.getRegClass(II, OpIdx, &TRI, MF);
80 // Some of the target independent instructions, like COPY, may not impose any
81 // register class constraints on some of their operands: If it's a use, we can
82 // skip constraining as the instruction defining the register would constrain
83 // it.
85 // We can't constrain unallocatable register classes, because we can't create
86 // virtual registers for these classes, so we need to let targets handled this
87 // case.
88 if (RegClass && !RegClass->isAllocatable())
89 RegClass = TRI.getConstrainedRegClassForOperand(RegMO, MRI);
91 if (!RegClass) {
92 assert((!isTargetSpecificOpcode(II.getOpcode()) || RegMO.isUse()) &&
93 "Register class constraint is required unless either the "
94 "instruction is target independent or the operand is a use");
95 // FIXME: Just bailing out like this here could be not enough, unless we
96 // expect the users of this function to do the right thing for PHIs and
97 // COPY:
98 // v1 = COPY v0
99 // v2 = COPY v1
100 // v1 here may end up not being constrained at all. Please notice that to
101 // reproduce the issue we likely need a destination pattern of a selection
102 // rule producing such extra copies, not just an input GMIR with them as
103 // every existing target using selectImpl handles copies before calling it
104 // and they never reach this function.
105 return Reg;
107 return constrainOperandRegClass(MF, TRI, MRI, TII, RBI, InsertPt, *RegClass,
108 RegMO, OpIdx);
111 bool llvm::constrainSelectedInstRegOperands(MachineInstr &I,
112 const TargetInstrInfo &TII,
113 const TargetRegisterInfo &TRI,
114 const RegisterBankInfo &RBI) {
115 assert(!isPreISelGenericOpcode(I.getOpcode()) &&
116 "A selected instruction is expected");
117 MachineBasicBlock &MBB = *I.getParent();
118 MachineFunction &MF = *MBB.getParent();
119 MachineRegisterInfo &MRI = MF.getRegInfo();
121 for (unsigned OpI = 0, OpE = I.getNumExplicitOperands(); OpI != OpE; ++OpI) {
122 MachineOperand &MO = I.getOperand(OpI);
124 // There's nothing to be done on non-register operands.
125 if (!MO.isReg())
126 continue;
128 LLVM_DEBUG(dbgs() << "Converting operand: " << MO << '\n');
129 assert(MO.isReg() && "Unsupported non-reg operand");
131 Register Reg = MO.getReg();
132 // Physical registers don't need to be constrained.
133 if (Register::isPhysicalRegister(Reg))
134 continue;
136 // Register operands with a value of 0 (e.g. predicate operands) don't need
137 // to be constrained.
138 if (Reg == 0)
139 continue;
141 // If the operand is a vreg, we should constrain its regclass, and only
142 // insert COPYs if that's impossible.
143 // constrainOperandRegClass does that for us.
144 MO.setReg(constrainOperandRegClass(MF, TRI, MRI, TII, RBI, I, I.getDesc(),
145 MO, OpI));
147 // Tie uses to defs as indicated in MCInstrDesc if this hasn't already been
148 // done.
149 if (MO.isUse()) {
150 int DefIdx = I.getDesc().getOperandConstraint(OpI, MCOI::TIED_TO);
151 if (DefIdx != -1 && !I.isRegTiedToUseOperand(DefIdx))
152 I.tieOperands(DefIdx, OpI);
155 return true;
158 bool llvm::isTriviallyDead(const MachineInstr &MI,
159 const MachineRegisterInfo &MRI) {
160 // If we can move an instruction, we can remove it. Otherwise, it has
161 // a side-effect of some sort.
162 bool SawStore = false;
163 if (!MI.isSafeToMove(/*AA=*/nullptr, SawStore) && !MI.isPHI())
164 return false;
166 // Instructions without side-effects are dead iff they only define dead vregs.
167 for (auto &MO : MI.operands()) {
168 if (!MO.isReg() || !MO.isDef())
169 continue;
171 Register Reg = MO.getReg();
172 if (Register::isPhysicalRegister(Reg) || !MRI.use_nodbg_empty(Reg))
173 return false;
175 return true;
178 void llvm::reportGISelFailure(MachineFunction &MF, const TargetPassConfig &TPC,
179 MachineOptimizationRemarkEmitter &MORE,
180 MachineOptimizationRemarkMissed &R) {
181 MF.getProperties().set(MachineFunctionProperties::Property::FailedISel);
183 // Print the function name explicitly if we don't have a debug location (which
184 // makes the diagnostic less useful) or if we're going to emit a raw error.
185 if (!R.getLocation().isValid() || TPC.isGlobalISelAbortEnabled())
186 R << (" (in function: " + MF.getName() + ")").str();
188 if (TPC.isGlobalISelAbortEnabled())
189 report_fatal_error(R.getMsg());
190 else
191 MORE.emit(R);
194 void llvm::reportGISelFailure(MachineFunction &MF, const TargetPassConfig &TPC,
195 MachineOptimizationRemarkEmitter &MORE,
196 const char *PassName, StringRef Msg,
197 const MachineInstr &MI) {
198 MachineOptimizationRemarkMissed R(PassName, "GISelFailure: ",
199 MI.getDebugLoc(), MI.getParent());
200 R << Msg;
201 // Printing MI is expensive; only do it if expensive remarks are enabled.
202 if (TPC.isGlobalISelAbortEnabled() || MORE.allowExtraAnalysis(PassName))
203 R << ": " << ore::MNV("Inst", MI);
204 reportGISelFailure(MF, TPC, MORE, R);
207 Optional<int64_t> llvm::getConstantVRegVal(unsigned VReg,
208 const MachineRegisterInfo &MRI) {
209 Optional<ValueAndVReg> ValAndVReg =
210 getConstantVRegValWithLookThrough(VReg, MRI, /*LookThroughInstrs*/ false);
211 assert((!ValAndVReg || ValAndVReg->VReg == VReg) &&
212 "Value found while looking through instrs");
213 if (!ValAndVReg)
214 return None;
215 return ValAndVReg->Value;
218 Optional<ValueAndVReg> llvm::getConstantVRegValWithLookThrough(
219 unsigned VReg, const MachineRegisterInfo &MRI, bool LookThroughInstrs) {
220 SmallVector<std::pair<unsigned, unsigned>, 4> SeenOpcodes;
221 MachineInstr *MI;
222 while ((MI = MRI.getVRegDef(VReg)) &&
223 MI->getOpcode() != TargetOpcode::G_CONSTANT && LookThroughInstrs) {
224 switch (MI->getOpcode()) {
225 case TargetOpcode::G_TRUNC:
226 case TargetOpcode::G_SEXT:
227 case TargetOpcode::G_ZEXT:
228 SeenOpcodes.push_back(std::make_pair(
229 MI->getOpcode(),
230 MRI.getType(MI->getOperand(0).getReg()).getSizeInBits()));
231 VReg = MI->getOperand(1).getReg();
232 break;
233 case TargetOpcode::COPY:
234 VReg = MI->getOperand(1).getReg();
235 if (Register::isPhysicalRegister(VReg))
236 return None;
237 break;
238 case TargetOpcode::G_INTTOPTR:
239 VReg = MI->getOperand(1).getReg();
240 break;
241 default:
242 return None;
245 if (!MI || MI->getOpcode() != TargetOpcode::G_CONSTANT ||
246 (!MI->getOperand(1).isImm() && !MI->getOperand(1).isCImm()))
247 return None;
249 const MachineOperand &CstVal = MI->getOperand(1);
250 unsigned BitWidth = MRI.getType(MI->getOperand(0).getReg()).getSizeInBits();
251 APInt Val = CstVal.isImm() ? APInt(BitWidth, CstVal.getImm())
252 : CstVal.getCImm()->getValue();
253 assert(Val.getBitWidth() == BitWidth &&
254 "Value bitwidth doesn't match definition type");
255 while (!SeenOpcodes.empty()) {
256 std::pair<unsigned, unsigned> OpcodeAndSize = SeenOpcodes.pop_back_val();
257 switch (OpcodeAndSize.first) {
258 case TargetOpcode::G_TRUNC:
259 Val = Val.trunc(OpcodeAndSize.second);
260 break;
261 case TargetOpcode::G_SEXT:
262 Val = Val.sext(OpcodeAndSize.second);
263 break;
264 case TargetOpcode::G_ZEXT:
265 Val = Val.zext(OpcodeAndSize.second);
266 break;
270 if (Val.getBitWidth() > 64)
271 return None;
273 return ValueAndVReg{Val.getSExtValue(), VReg};
276 const llvm::ConstantFP* llvm::getConstantFPVRegVal(unsigned VReg,
277 const MachineRegisterInfo &MRI) {
278 MachineInstr *MI = MRI.getVRegDef(VReg);
279 if (TargetOpcode::G_FCONSTANT != MI->getOpcode())
280 return nullptr;
281 return MI->getOperand(1).getFPImm();
284 llvm::MachineInstr *llvm::getDefIgnoringCopies(Register Reg,
285 const MachineRegisterInfo &MRI) {
286 auto *DefMI = MRI.getVRegDef(Reg);
287 auto DstTy = MRI.getType(DefMI->getOperand(0).getReg());
288 if (!DstTy.isValid())
289 return nullptr;
290 while (DefMI->getOpcode() == TargetOpcode::COPY) {
291 Register SrcReg = DefMI->getOperand(1).getReg();
292 auto SrcTy = MRI.getType(SrcReg);
293 if (!SrcTy.isValid() || SrcTy != DstTy)
294 break;
295 DefMI = MRI.getVRegDef(SrcReg);
297 return DefMI;
300 llvm::MachineInstr *llvm::getOpcodeDef(unsigned Opcode, Register Reg,
301 const MachineRegisterInfo &MRI) {
302 MachineInstr *DefMI = getDefIgnoringCopies(Reg, MRI);
303 return DefMI && DefMI->getOpcode() == Opcode ? DefMI : nullptr;
306 APFloat llvm::getAPFloatFromSize(double Val, unsigned Size) {
307 if (Size == 32)
308 return APFloat(float(Val));
309 if (Size == 64)
310 return APFloat(Val);
311 if (Size != 16)
312 llvm_unreachable("Unsupported FPConstant size");
313 bool Ignored;
314 APFloat APF(Val);
315 APF.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven, &Ignored);
316 return APF;
319 Optional<APInt> llvm::ConstantFoldBinOp(unsigned Opcode, const unsigned Op1,
320 const unsigned Op2,
321 const MachineRegisterInfo &MRI) {
322 auto MaybeOp1Cst = getConstantVRegVal(Op1, MRI);
323 auto MaybeOp2Cst = getConstantVRegVal(Op2, MRI);
324 if (MaybeOp1Cst && MaybeOp2Cst) {
325 LLT Ty = MRI.getType(Op1);
326 APInt C1(Ty.getSizeInBits(), *MaybeOp1Cst, true);
327 APInt C2(Ty.getSizeInBits(), *MaybeOp2Cst, true);
328 switch (Opcode) {
329 default:
330 break;
331 case TargetOpcode::G_ADD:
332 return C1 + C2;
333 case TargetOpcode::G_AND:
334 return C1 & C2;
335 case TargetOpcode::G_ASHR:
336 return C1.ashr(C2);
337 case TargetOpcode::G_LSHR:
338 return C1.lshr(C2);
339 case TargetOpcode::G_MUL:
340 return C1 * C2;
341 case TargetOpcode::G_OR:
342 return C1 | C2;
343 case TargetOpcode::G_SHL:
344 return C1 << C2;
345 case TargetOpcode::G_SUB:
346 return C1 - C2;
347 case TargetOpcode::G_XOR:
348 return C1 ^ C2;
349 case TargetOpcode::G_UDIV:
350 if (!C2.getBoolValue())
351 break;
352 return C1.udiv(C2);
353 case TargetOpcode::G_SDIV:
354 if (!C2.getBoolValue())
355 break;
356 return C1.sdiv(C2);
357 case TargetOpcode::G_UREM:
358 if (!C2.getBoolValue())
359 break;
360 return C1.urem(C2);
361 case TargetOpcode::G_SREM:
362 if (!C2.getBoolValue())
363 break;
364 return C1.srem(C2);
367 return None;
370 bool llvm::isKnownNeverNaN(Register Val, const MachineRegisterInfo &MRI,
371 bool SNaN) {
372 const MachineInstr *DefMI = MRI.getVRegDef(Val);
373 if (!DefMI)
374 return false;
376 if (DefMI->getFlag(MachineInstr::FmNoNans))
377 return true;
379 if (SNaN) {
380 // FP operations quiet. For now, just handle the ones inserted during
381 // legalization.
382 switch (DefMI->getOpcode()) {
383 case TargetOpcode::G_FPEXT:
384 case TargetOpcode::G_FPTRUNC:
385 case TargetOpcode::G_FCANONICALIZE:
386 return true;
387 default:
388 return false;
392 return false;
395 Optional<APInt> llvm::ConstantFoldExtOp(unsigned Opcode, const unsigned Op1,
396 uint64_t Imm,
397 const MachineRegisterInfo &MRI) {
398 auto MaybeOp1Cst = getConstantVRegVal(Op1, MRI);
399 if (MaybeOp1Cst) {
400 LLT Ty = MRI.getType(Op1);
401 APInt C1(Ty.getSizeInBits(), *MaybeOp1Cst, true);
402 switch (Opcode) {
403 default:
404 break;
405 case TargetOpcode::G_SEXT_INREG:
406 return C1.trunc(Imm).sext(C1.getBitWidth());
409 return None;
412 void llvm::getSelectionDAGFallbackAnalysisUsage(AnalysisUsage &AU) {
413 AU.addPreserved<StackProtector>();
416 MVT llvm::getMVTForLLT(LLT Ty) {
417 if (!Ty.isVector())
418 return MVT::getIntegerVT(Ty.getSizeInBits());
420 return MVT::getVectorVT(
421 MVT::getIntegerVT(Ty.getElementType().getSizeInBits()),
422 Ty.getNumElements());
425 LLT llvm::getLLTForMVT(MVT Ty) {
426 if (!Ty.isVector())
427 return LLT::scalar(Ty.getSizeInBits());
429 return LLT::vector(Ty.getVectorNumElements(),
430 Ty.getVectorElementType().getSizeInBits());