xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AArch64/GISel/AArch64PostLegalizerLowering.cpp (revision 77013d11e6483b970af25e13c9b892075742f7e5)
1 //=== AArch64PostLegalizerLowering.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 ///
9 /// \file
10 /// Post-legalization lowering for instructions.
11 ///
12 /// This is used to offload pattern matching from the selector.
13 ///
14 /// For example, this combiner will notice that a G_SHUFFLE_VECTOR is actually
15 /// a G_ZIP, G_UZP, etc.
16 ///
17 /// General optimization combines should be handled by either the
18 /// AArch64PostLegalizerCombiner or the AArch64PreLegalizerCombiner.
19 ///
20 //===----------------------------------------------------------------------===//
21 
22 #include "AArch64TargetMachine.h"
23 #include "AArch64GlobalISelUtils.h"
24 #include "MCTargetDesc/AArch64MCTargetDesc.h"
25 #include "llvm/CodeGen/GlobalISel/Combiner.h"
26 #include "llvm/CodeGen/GlobalISel/CombinerHelper.h"
27 #include "llvm/CodeGen/GlobalISel/CombinerInfo.h"
28 #include "llvm/CodeGen/GlobalISel/MIPatternMatch.h"
29 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
30 #include "llvm/CodeGen/GlobalISel/Utils.h"
31 #include "llvm/CodeGen/MachineFunctionPass.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/CodeGen/TargetOpcodes.h"
35 #include "llvm/CodeGen/TargetPassConfig.h"
36 #include "llvm/InitializePasses.h"
37 #include "llvm/Support/Debug.h"
38 
39 #define DEBUG_TYPE "aarch64-postlegalizer-lowering"
40 
41 using namespace llvm;
42 using namespace MIPatternMatch;
43 using namespace AArch64GISelUtils;
44 
45 /// Represents a pseudo instruction which replaces a G_SHUFFLE_VECTOR.
46 ///
47 /// Used for matching target-supported shuffles before codegen.
48 struct ShuffleVectorPseudo {
49   unsigned Opc; ///< Opcode for the instruction. (E.g. G_ZIP1)
50   Register Dst; ///< Destination register.
51   SmallVector<SrcOp, 2> SrcOps; ///< Source registers.
52   ShuffleVectorPseudo(unsigned Opc, Register Dst,
53                       std::initializer_list<SrcOp> SrcOps)
54       : Opc(Opc), Dst(Dst), SrcOps(SrcOps){};
55   ShuffleVectorPseudo() {}
56 };
57 
58 /// Check if a vector shuffle corresponds to a REV instruction with the
59 /// specified blocksize.
60 static bool isREVMask(ArrayRef<int> M, unsigned EltSize, unsigned NumElts,
61                       unsigned BlockSize) {
62   assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
63          "Only possible block sizes for REV are: 16, 32, 64");
64   assert(EltSize != 64 && "EltSize cannot be 64 for REV mask.");
65 
66   unsigned BlockElts = M[0] + 1;
67 
68   // If the first shuffle index is UNDEF, be optimistic.
69   if (M[0] < 0)
70     BlockElts = BlockSize / EltSize;
71 
72   if (BlockSize <= EltSize || BlockSize != BlockElts * EltSize)
73     return false;
74 
75   for (unsigned i = 0; i < NumElts; ++i) {
76     // Ignore undef indices.
77     if (M[i] < 0)
78       continue;
79     if (static_cast<unsigned>(M[i]) !=
80         (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
81       return false;
82   }
83 
84   return true;
85 }
86 
87 /// Determines if \p M is a shuffle vector mask for a TRN of \p NumElts.
88 /// Whether or not G_TRN1 or G_TRN2 should be used is stored in \p WhichResult.
89 static bool isTRNMask(ArrayRef<int> M, unsigned NumElts,
90                       unsigned &WhichResult) {
91   if (NumElts % 2 != 0)
92     return false;
93   WhichResult = (M[0] == 0 ? 0 : 1);
94   for (unsigned i = 0; i < NumElts; i += 2) {
95     if ((M[i] >= 0 && static_cast<unsigned>(M[i]) != i + WhichResult) ||
96         (M[i + 1] >= 0 &&
97          static_cast<unsigned>(M[i + 1]) != i + NumElts + WhichResult))
98       return false;
99   }
100   return true;
101 }
102 
103 /// Check if a G_EXT instruction can handle a shuffle mask \p M when the vector
104 /// sources of the shuffle are different.
105 static Optional<std::pair<bool, uint64_t>> getExtMask(ArrayRef<int> M,
106                                                       unsigned NumElts) {
107   // Look for the first non-undef element.
108   auto FirstRealElt = find_if(M, [](int Elt) { return Elt >= 0; });
109   if (FirstRealElt == M.end())
110     return None;
111 
112   // Use APInt to handle overflow when calculating expected element.
113   unsigned MaskBits = APInt(32, NumElts * 2).logBase2();
114   APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1);
115 
116   // The following shuffle indices must be the successive elements after the
117   // first real element.
118   if (any_of(
119           make_range(std::next(FirstRealElt), M.end()),
120           [&ExpectedElt](int Elt) { return Elt != ExpectedElt++ && Elt >= 0; }))
121     return None;
122 
123   // The index of an EXT is the first element if it is not UNDEF.
124   // Watch out for the beginning UNDEFs. The EXT index should be the expected
125   // value of the first element.  E.g.
126   // <-1, -1, 3, ...> is treated as <1, 2, 3, ...>.
127   // <-1, -1, 0, 1, ...> is treated as <2*NumElts-2, 2*NumElts-1, 0, 1, ...>.
128   // ExpectedElt is the last mask index plus 1.
129   uint64_t Imm = ExpectedElt.getZExtValue();
130   bool ReverseExt = false;
131 
132   // There are two difference cases requiring to reverse input vectors.
133   // For example, for vector <4 x i32> we have the following cases,
134   // Case 1: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, -1, 0>)
135   // Case 2: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, 7, 0>)
136   // For both cases, we finally use mask <5, 6, 7, 0>, which requires
137   // to reverse two input vectors.
138   if (Imm < NumElts)
139     ReverseExt = true;
140   else
141     Imm -= NumElts;
142   return std::make_pair(ReverseExt, Imm);
143 }
144 
145 /// Determines if \p M is a shuffle vector mask for a UZP of \p NumElts.
146 /// Whether or not G_UZP1 or G_UZP2 should be used is stored in \p WhichResult.
147 static bool isUZPMask(ArrayRef<int> M, unsigned NumElts,
148                       unsigned &WhichResult) {
149   WhichResult = (M[0] == 0 ? 0 : 1);
150   for (unsigned i = 0; i != NumElts; ++i) {
151     // Skip undef indices.
152     if (M[i] < 0)
153       continue;
154     if (static_cast<unsigned>(M[i]) != 2 * i + WhichResult)
155       return false;
156   }
157   return true;
158 }
159 
160 /// \return true if \p M is a zip mask for a shuffle vector of \p NumElts.
161 /// Whether or not G_ZIP1 or G_ZIP2 should be used is stored in \p WhichResult.
162 static bool isZipMask(ArrayRef<int> M, unsigned NumElts,
163                       unsigned &WhichResult) {
164   if (NumElts % 2 != 0)
165     return false;
166 
167   // 0 means use ZIP1, 1 means use ZIP2.
168   WhichResult = (M[0] == 0 ? 0 : 1);
169   unsigned Idx = WhichResult * NumElts / 2;
170   for (unsigned i = 0; i != NumElts; i += 2) {
171       if ((M[i] >= 0 && static_cast<unsigned>(M[i]) != Idx) ||
172           (M[i + 1] >= 0 && static_cast<unsigned>(M[i + 1]) != Idx + NumElts))
173         return false;
174     Idx += 1;
175   }
176   return true;
177 }
178 
179 /// \return true if a G_SHUFFLE_VECTOR instruction \p MI can be replaced with a
180 /// G_REV instruction. Returns the appropriate G_REV opcode in \p Opc.
181 static bool matchREV(MachineInstr &MI, MachineRegisterInfo &MRI,
182                      ShuffleVectorPseudo &MatchInfo) {
183   assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR);
184   ArrayRef<int> ShuffleMask = MI.getOperand(3).getShuffleMask();
185   Register Dst = MI.getOperand(0).getReg();
186   Register Src = MI.getOperand(1).getReg();
187   LLT Ty = MRI.getType(Dst);
188   unsigned EltSize = Ty.getScalarSizeInBits();
189 
190   // Element size for a rev cannot be 64.
191   if (EltSize == 64)
192     return false;
193 
194   unsigned NumElts = Ty.getNumElements();
195 
196   // Try to produce G_REV64
197   if (isREVMask(ShuffleMask, EltSize, NumElts, 64)) {
198     MatchInfo = ShuffleVectorPseudo(AArch64::G_REV64, Dst, {Src});
199     return true;
200   }
201 
202   // TODO: Produce G_REV32 and G_REV16 once we have proper legalization support.
203   // This should be identical to above, but with a constant 32 and constant
204   // 16.
205   return false;
206 }
207 
208 /// \return true if a G_SHUFFLE_VECTOR instruction \p MI can be replaced with
209 /// a G_TRN1 or G_TRN2 instruction.
210 static bool matchTRN(MachineInstr &MI, MachineRegisterInfo &MRI,
211                      ShuffleVectorPseudo &MatchInfo) {
212   assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR);
213   unsigned WhichResult;
214   ArrayRef<int> ShuffleMask = MI.getOperand(3).getShuffleMask();
215   Register Dst = MI.getOperand(0).getReg();
216   unsigned NumElts = MRI.getType(Dst).getNumElements();
217   if (!isTRNMask(ShuffleMask, NumElts, WhichResult))
218     return false;
219   unsigned Opc = (WhichResult == 0) ? AArch64::G_TRN1 : AArch64::G_TRN2;
220   Register V1 = MI.getOperand(1).getReg();
221   Register V2 = MI.getOperand(2).getReg();
222   MatchInfo = ShuffleVectorPseudo(Opc, Dst, {V1, V2});
223   return true;
224 }
225 
226 /// \return true if a G_SHUFFLE_VECTOR instruction \p MI can be replaced with
227 /// a G_UZP1 or G_UZP2 instruction.
228 ///
229 /// \param [in] MI - The shuffle vector instruction.
230 /// \param [out] MatchInfo - Either G_UZP1 or G_UZP2 on success.
231 static bool matchUZP(MachineInstr &MI, MachineRegisterInfo &MRI,
232                      ShuffleVectorPseudo &MatchInfo) {
233   assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR);
234   unsigned WhichResult;
235   ArrayRef<int> ShuffleMask = MI.getOperand(3).getShuffleMask();
236   Register Dst = MI.getOperand(0).getReg();
237   unsigned NumElts = MRI.getType(Dst).getNumElements();
238   if (!isUZPMask(ShuffleMask, NumElts, WhichResult))
239     return false;
240   unsigned Opc = (WhichResult == 0) ? AArch64::G_UZP1 : AArch64::G_UZP2;
241   Register V1 = MI.getOperand(1).getReg();
242   Register V2 = MI.getOperand(2).getReg();
243   MatchInfo = ShuffleVectorPseudo(Opc, Dst, {V1, V2});
244   return true;
245 }
246 
247 static bool matchZip(MachineInstr &MI, MachineRegisterInfo &MRI,
248                      ShuffleVectorPseudo &MatchInfo) {
249   assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR);
250   unsigned WhichResult;
251   ArrayRef<int> ShuffleMask = MI.getOperand(3).getShuffleMask();
252   Register Dst = MI.getOperand(0).getReg();
253   unsigned NumElts = MRI.getType(Dst).getNumElements();
254   if (!isZipMask(ShuffleMask, NumElts, WhichResult))
255     return false;
256   unsigned Opc = (WhichResult == 0) ? AArch64::G_ZIP1 : AArch64::G_ZIP2;
257   Register V1 = MI.getOperand(1).getReg();
258   Register V2 = MI.getOperand(2).getReg();
259   MatchInfo = ShuffleVectorPseudo(Opc, Dst, {V1, V2});
260   return true;
261 }
262 
263 /// Helper function for matchDup.
264 static bool matchDupFromInsertVectorElt(int Lane, MachineInstr &MI,
265                                         MachineRegisterInfo &MRI,
266                                         ShuffleVectorPseudo &MatchInfo) {
267   if (Lane != 0)
268     return false;
269 
270   // Try to match a vector splat operation into a dup instruction.
271   // We're looking for this pattern:
272   //
273   // %scalar:gpr(s64) = COPY $x0
274   // %undef:fpr(<2 x s64>) = G_IMPLICIT_DEF
275   // %cst0:gpr(s32) = G_CONSTANT i32 0
276   // %zerovec:fpr(<2 x s32>) = G_BUILD_VECTOR %cst0(s32), %cst0(s32)
277   // %ins:fpr(<2 x s64>) = G_INSERT_VECTOR_ELT %undef, %scalar(s64), %cst0(s32)
278   // %splat:fpr(<2 x s64>) = G_SHUFFLE_VECTOR %ins(<2 x s64>), %undef, %zerovec(<2 x s32>)
279   //
280   // ...into:
281   // %splat = G_DUP %scalar
282 
283   // Begin matching the insert.
284   auto *InsMI = getOpcodeDef(TargetOpcode::G_INSERT_VECTOR_ELT,
285                              MI.getOperand(1).getReg(), MRI);
286   if (!InsMI)
287     return false;
288   // Match the undef vector operand.
289   if (!getOpcodeDef(TargetOpcode::G_IMPLICIT_DEF, InsMI->getOperand(1).getReg(),
290                     MRI))
291     return false;
292 
293   // Match the index constant 0.
294   if (!mi_match(InsMI->getOperand(3).getReg(), MRI, m_ZeroInt()))
295     return false;
296 
297   MatchInfo = ShuffleVectorPseudo(AArch64::G_DUP, MI.getOperand(0).getReg(),
298                                   {InsMI->getOperand(2).getReg()});
299   return true;
300 }
301 
302 /// Helper function for matchDup.
303 static bool matchDupFromBuildVector(int Lane, MachineInstr &MI,
304                                     MachineRegisterInfo &MRI,
305                                     ShuffleVectorPseudo &MatchInfo) {
306   assert(Lane >= 0 && "Expected positive lane?");
307   // Test if the LHS is a BUILD_VECTOR. If it is, then we can just reference the
308   // lane's definition directly.
309   auto *BuildVecMI = getOpcodeDef(TargetOpcode::G_BUILD_VECTOR,
310                                   MI.getOperand(1).getReg(), MRI);
311   if (!BuildVecMI)
312     return false;
313   Register Reg = BuildVecMI->getOperand(Lane + 1).getReg();
314   MatchInfo =
315       ShuffleVectorPseudo(AArch64::G_DUP, MI.getOperand(0).getReg(), {Reg});
316   return true;
317 }
318 
319 static bool matchDup(MachineInstr &MI, MachineRegisterInfo &MRI,
320                      ShuffleVectorPseudo &MatchInfo) {
321   assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR);
322   auto MaybeLane = getSplatIndex(MI);
323   if (!MaybeLane)
324     return false;
325   int Lane = *MaybeLane;
326   // If this is undef splat, generate it via "just" vdup, if possible.
327   if (Lane < 0)
328     Lane = 0;
329   if (matchDupFromInsertVectorElt(Lane, MI, MRI, MatchInfo))
330     return true;
331   if (matchDupFromBuildVector(Lane, MI, MRI, MatchInfo))
332     return true;
333   return false;
334 }
335 
336 static bool matchEXT(MachineInstr &MI, MachineRegisterInfo &MRI,
337                      ShuffleVectorPseudo &MatchInfo) {
338   assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR);
339   Register Dst = MI.getOperand(0).getReg();
340   auto ExtInfo = getExtMask(MI.getOperand(3).getShuffleMask(),
341                             MRI.getType(Dst).getNumElements());
342   if (!ExtInfo)
343     return false;
344   bool ReverseExt;
345   uint64_t Imm;
346   std::tie(ReverseExt, Imm) = *ExtInfo;
347   Register V1 = MI.getOperand(1).getReg();
348   Register V2 = MI.getOperand(2).getReg();
349   if (ReverseExt)
350     std::swap(V1, V2);
351   uint64_t ExtFactor = MRI.getType(V1).getScalarSizeInBits() / 8;
352   Imm *= ExtFactor;
353   MatchInfo = ShuffleVectorPseudo(AArch64::G_EXT, Dst, {V1, V2, Imm});
354   return true;
355 }
356 
357 /// Replace a G_SHUFFLE_VECTOR instruction with a pseudo.
358 /// \p Opc is the opcode to use. \p MI is the G_SHUFFLE_VECTOR.
359 static bool applyShuffleVectorPseudo(MachineInstr &MI,
360                                      ShuffleVectorPseudo &MatchInfo) {
361   MachineIRBuilder MIRBuilder(MI);
362   MIRBuilder.buildInstr(MatchInfo.Opc, {MatchInfo.Dst}, MatchInfo.SrcOps);
363   MI.eraseFromParent();
364   return true;
365 }
366 
367 /// Replace a G_SHUFFLE_VECTOR instruction with G_EXT.
368 /// Special-cased because the constant operand must be emitted as a G_CONSTANT
369 /// for the imported tablegen patterns to work.
370 static bool applyEXT(MachineInstr &MI, ShuffleVectorPseudo &MatchInfo) {
371   MachineIRBuilder MIRBuilder(MI);
372   // Tablegen patterns expect an i32 G_CONSTANT as the final op.
373   auto Cst =
374       MIRBuilder.buildConstant(LLT::scalar(32), MatchInfo.SrcOps[2].getImm());
375   MIRBuilder.buildInstr(MatchInfo.Opc, {MatchInfo.Dst},
376                         {MatchInfo.SrcOps[0], MatchInfo.SrcOps[1], Cst});
377   MI.eraseFromParent();
378   return true;
379 }
380 
381 /// isVShiftRImm - Check if this is a valid vector for the immediate
382 /// operand of a vector shift right operation. The value must be in the range:
383 ///   1 <= Value <= ElementBits for a right shift.
384 static bool isVShiftRImm(Register Reg, MachineRegisterInfo &MRI, LLT Ty,
385                          int64_t &Cnt) {
386   assert(Ty.isVector() && "vector shift count is not a vector type");
387   MachineInstr *MI = MRI.getVRegDef(Reg);
388   auto Cst = getBuildVectorConstantSplat(*MI, MRI);
389   if (!Cst)
390     return false;
391   Cnt = *Cst;
392   int64_t ElementBits = Ty.getScalarSizeInBits();
393   return Cnt >= 1 && Cnt <= ElementBits;
394 }
395 
396 /// Match a vector G_ASHR or G_LSHR with a valid immediate shift.
397 static bool matchVAshrLshrImm(MachineInstr &MI, MachineRegisterInfo &MRI,
398                               int64_t &Imm) {
399   assert(MI.getOpcode() == TargetOpcode::G_ASHR ||
400          MI.getOpcode() == TargetOpcode::G_LSHR);
401   LLT Ty = MRI.getType(MI.getOperand(1).getReg());
402   if (!Ty.isVector())
403     return false;
404   return isVShiftRImm(MI.getOperand(2).getReg(), MRI, Ty, Imm);
405 }
406 
407 static bool applyVAshrLshrImm(MachineInstr &MI, MachineRegisterInfo &MRI,
408                               int64_t &Imm) {
409   unsigned Opc = MI.getOpcode();
410   assert(Opc == TargetOpcode::G_ASHR || Opc == TargetOpcode::G_LSHR);
411   unsigned NewOpc =
412       Opc == TargetOpcode::G_ASHR ? AArch64::G_VASHR : AArch64::G_VLSHR;
413   MachineIRBuilder MIB(MI);
414   auto ImmDef = MIB.buildConstant(LLT::scalar(32), Imm);
415   MIB.buildInstr(NewOpc, {MI.getOperand(0)}, {MI.getOperand(1), ImmDef});
416   MI.eraseFromParent();
417   return true;
418 }
419 
420 /// Determine if it is possible to modify the \p RHS and predicate \p P of a
421 /// G_ICMP instruction such that the right-hand side is an arithmetic immediate.
422 ///
423 /// \returns A pair containing the updated immediate and predicate which may
424 /// be used to optimize the instruction.
425 ///
426 /// \note This assumes that the comparison has been legalized.
427 Optional<std::pair<uint64_t, CmpInst::Predicate>>
428 tryAdjustICmpImmAndPred(Register RHS, CmpInst::Predicate P,
429                           const MachineRegisterInfo &MRI) {
430   const auto &Ty = MRI.getType(RHS);
431   if (Ty.isVector())
432     return None;
433   unsigned Size = Ty.getSizeInBits();
434   assert((Size == 32 || Size == 64) && "Expected 32 or 64 bit compare only?");
435 
436   // If the RHS is not a constant, or the RHS is already a valid arithmetic
437   // immediate, then there is nothing to change.
438   auto ValAndVReg = getConstantVRegValWithLookThrough(RHS, MRI);
439   if (!ValAndVReg)
440     return None;
441   uint64_t C = ValAndVReg->Value.getZExtValue();
442   if (isLegalArithImmed(C))
443     return None;
444 
445   // We have a non-arithmetic immediate. Check if adjusting the immediate and
446   // adjusting the predicate will result in a legal arithmetic immediate.
447   switch (P) {
448   default:
449     return None;
450   case CmpInst::ICMP_SLT:
451   case CmpInst::ICMP_SGE:
452     // Check for
453     //
454     // x slt c => x sle c - 1
455     // x sge c => x sgt c - 1
456     //
457     // When c is not the smallest possible negative number.
458     if ((Size == 64 && static_cast<int64_t>(C) == INT64_MIN) ||
459         (Size == 32 && static_cast<int32_t>(C) == INT32_MIN))
460       return None;
461     P = (P == CmpInst::ICMP_SLT) ? CmpInst::ICMP_SLE : CmpInst::ICMP_SGT;
462     C -= 1;
463     break;
464   case CmpInst::ICMP_ULT:
465   case CmpInst::ICMP_UGE:
466     // Check for
467     //
468     // x ult c => x ule c - 1
469     // x uge c => x ugt c - 1
470     //
471     // When c is not zero.
472     if (C == 0)
473       return None;
474     P = (P == CmpInst::ICMP_ULT) ? CmpInst::ICMP_ULE : CmpInst::ICMP_UGT;
475     C -= 1;
476     break;
477   case CmpInst::ICMP_SLE:
478   case CmpInst::ICMP_SGT:
479     // Check for
480     //
481     // x sle c => x slt c + 1
482     // x sgt c => s sge c + 1
483     //
484     // When c is not the largest possible signed integer.
485     if ((Size == 32 && static_cast<int32_t>(C) == INT32_MAX) ||
486         (Size == 64 && static_cast<int64_t>(C) == INT64_MAX))
487       return None;
488     P = (P == CmpInst::ICMP_SLE) ? CmpInst::ICMP_SLT : CmpInst::ICMP_SGE;
489     C += 1;
490     break;
491   case CmpInst::ICMP_ULE:
492   case CmpInst::ICMP_UGT:
493     // Check for
494     //
495     // x ule c => x ult c + 1
496     // x ugt c => s uge c + 1
497     //
498     // When c is not the largest possible unsigned integer.
499     if ((Size == 32 && static_cast<uint32_t>(C) == UINT32_MAX) ||
500         (Size == 64 && C == UINT64_MAX))
501       return None;
502     P = (P == CmpInst::ICMP_ULE) ? CmpInst::ICMP_ULT : CmpInst::ICMP_UGE;
503     C += 1;
504     break;
505   }
506 
507   // Check if the new constant is valid, and return the updated constant and
508   // predicate if it is.
509   if (Size == 32)
510     C = static_cast<uint32_t>(C);
511   if (!isLegalArithImmed(C))
512     return None;
513   return {{C, P}};
514 }
515 
516 /// Determine whether or not it is possible to update the RHS and predicate of
517 /// a G_ICMP instruction such that the RHS will be selected as an arithmetic
518 /// immediate.
519 ///
520 /// \p MI - The G_ICMP instruction
521 /// \p MatchInfo - The new RHS immediate and predicate on success
522 ///
523 /// See tryAdjustICmpImmAndPred for valid transformations.
524 bool matchAdjustICmpImmAndPred(
525     MachineInstr &MI, const MachineRegisterInfo &MRI,
526     std::pair<uint64_t, CmpInst::Predicate> &MatchInfo) {
527   assert(MI.getOpcode() == TargetOpcode::G_ICMP);
528   Register RHS = MI.getOperand(3).getReg();
529   auto Pred = static_cast<CmpInst::Predicate>(MI.getOperand(1).getPredicate());
530   if (auto MaybeNewImmAndPred = tryAdjustICmpImmAndPred(RHS, Pred, MRI)) {
531     MatchInfo = *MaybeNewImmAndPred;
532     return true;
533   }
534   return false;
535 }
536 
537 bool applyAdjustICmpImmAndPred(
538     MachineInstr &MI, std::pair<uint64_t, CmpInst::Predicate> &MatchInfo,
539     MachineIRBuilder &MIB, GISelChangeObserver &Observer) {
540   MIB.setInstrAndDebugLoc(MI);
541   MachineOperand &RHS = MI.getOperand(3);
542   MachineRegisterInfo &MRI = *MIB.getMRI();
543   auto Cst = MIB.buildConstant(MRI.cloneVirtualRegister(RHS.getReg()),
544                                MatchInfo.first);
545   Observer.changingInstr(MI);
546   RHS.setReg(Cst->getOperand(0).getReg());
547   MI.getOperand(1).setPredicate(MatchInfo.second);
548   Observer.changedInstr(MI);
549   return true;
550 }
551 
552 bool matchDupLane(MachineInstr &MI, MachineRegisterInfo &MRI,
553                   std::pair<unsigned, int> &MatchInfo) {
554   assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR);
555   Register Src1Reg = MI.getOperand(1).getReg();
556   const LLT SrcTy = MRI.getType(Src1Reg);
557   const LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
558 
559   auto LaneIdx = getSplatIndex(MI);
560   if (!LaneIdx)
561     return false;
562 
563   // The lane idx should be within the first source vector.
564   if (*LaneIdx >= SrcTy.getNumElements())
565     return false;
566 
567   if (DstTy != SrcTy)
568     return false;
569 
570   LLT ScalarTy = SrcTy.getElementType();
571   unsigned ScalarSize = ScalarTy.getSizeInBits();
572 
573   unsigned Opc = 0;
574   switch (SrcTy.getNumElements()) {
575   case 2:
576     if (ScalarSize == 64)
577       Opc = AArch64::G_DUPLANE64;
578     break;
579   case 4:
580     if (ScalarSize == 32)
581       Opc = AArch64::G_DUPLANE32;
582     break;
583   case 8:
584     if (ScalarSize == 16)
585       Opc = AArch64::G_DUPLANE16;
586     break;
587   case 16:
588     if (ScalarSize == 8)
589       Opc = AArch64::G_DUPLANE8;
590     break;
591   default:
592     break;
593   }
594   if (!Opc)
595     return false;
596 
597   MatchInfo.first = Opc;
598   MatchInfo.second = *LaneIdx;
599   return true;
600 }
601 
602 bool applyDupLane(MachineInstr &MI, MachineRegisterInfo &MRI,
603                   MachineIRBuilder &B, std::pair<unsigned, int> &MatchInfo) {
604   assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR);
605   B.setInstrAndDebugLoc(MI);
606   auto Lane = B.buildConstant(LLT::scalar(64), MatchInfo.second);
607   B.buildInstr(MatchInfo.first, {MI.getOperand(0).getReg()},
608                {MI.getOperand(1).getReg(), Lane});
609   MI.eraseFromParent();
610   return true;
611 }
612 
613 #define AARCH64POSTLEGALIZERLOWERINGHELPER_GENCOMBINERHELPER_DEPS
614 #include "AArch64GenPostLegalizeGILowering.inc"
615 #undef AARCH64POSTLEGALIZERLOWERINGHELPER_GENCOMBINERHELPER_DEPS
616 
617 namespace {
618 #define AARCH64POSTLEGALIZERLOWERINGHELPER_GENCOMBINERHELPER_H
619 #include "AArch64GenPostLegalizeGILowering.inc"
620 #undef AARCH64POSTLEGALIZERLOWERINGHELPER_GENCOMBINERHELPER_H
621 
622 class AArch64PostLegalizerLoweringInfo : public CombinerInfo {
623 public:
624   AArch64GenPostLegalizerLoweringHelperRuleConfig GeneratedRuleCfg;
625 
626   AArch64PostLegalizerLoweringInfo(bool OptSize, bool MinSize)
627       : CombinerInfo(/*AllowIllegalOps*/ true, /*ShouldLegalizeIllegal*/ false,
628                      /*LegalizerInfo*/ nullptr, /*OptEnabled = */ true, OptSize,
629                      MinSize) {
630     if (!GeneratedRuleCfg.parseCommandLineOption())
631       report_fatal_error("Invalid rule identifier");
632   }
633 
634   virtual bool combine(GISelChangeObserver &Observer, MachineInstr &MI,
635                        MachineIRBuilder &B) const override;
636 };
637 
638 bool AArch64PostLegalizerLoweringInfo::combine(GISelChangeObserver &Observer,
639                                                MachineInstr &MI,
640                                                MachineIRBuilder &B) const {
641   CombinerHelper Helper(Observer, B);
642   AArch64GenPostLegalizerLoweringHelper Generated(GeneratedRuleCfg);
643   return Generated.tryCombineAll(Observer, MI, B, Helper);
644 }
645 
646 #define AARCH64POSTLEGALIZERLOWERINGHELPER_GENCOMBINERHELPER_CPP
647 #include "AArch64GenPostLegalizeGILowering.inc"
648 #undef AARCH64POSTLEGALIZERLOWERINGHELPER_GENCOMBINERHELPER_CPP
649 
650 class AArch64PostLegalizerLowering : public MachineFunctionPass {
651 public:
652   static char ID;
653 
654   AArch64PostLegalizerLowering();
655 
656   StringRef getPassName() const override {
657     return "AArch64PostLegalizerLowering";
658   }
659 
660   bool runOnMachineFunction(MachineFunction &MF) override;
661   void getAnalysisUsage(AnalysisUsage &AU) const override;
662 };
663 } // end anonymous namespace
664 
665 void AArch64PostLegalizerLowering::getAnalysisUsage(AnalysisUsage &AU) const {
666   AU.addRequired<TargetPassConfig>();
667   AU.setPreservesCFG();
668   getSelectionDAGFallbackAnalysisUsage(AU);
669   MachineFunctionPass::getAnalysisUsage(AU);
670 }
671 
672 AArch64PostLegalizerLowering::AArch64PostLegalizerLowering()
673     : MachineFunctionPass(ID) {
674   initializeAArch64PostLegalizerLoweringPass(*PassRegistry::getPassRegistry());
675 }
676 
677 bool AArch64PostLegalizerLowering::runOnMachineFunction(MachineFunction &MF) {
678   if (MF.getProperties().hasProperty(
679           MachineFunctionProperties::Property::FailedISel))
680     return false;
681   assert(MF.getProperties().hasProperty(
682              MachineFunctionProperties::Property::Legalized) &&
683          "Expected a legalized function?");
684   auto *TPC = &getAnalysis<TargetPassConfig>();
685   const Function &F = MF.getFunction();
686   AArch64PostLegalizerLoweringInfo PCInfo(F.hasOptSize(), F.hasMinSize());
687   Combiner C(PCInfo, TPC);
688   return C.combineMachineInstrs(MF, /*CSEInfo*/ nullptr);
689 }
690 
691 char AArch64PostLegalizerLowering::ID = 0;
692 INITIALIZE_PASS_BEGIN(AArch64PostLegalizerLowering, DEBUG_TYPE,
693                       "Lower AArch64 MachineInstrs after legalization", false,
694                       false)
695 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
696 INITIALIZE_PASS_END(AArch64PostLegalizerLowering, DEBUG_TYPE,
697                     "Lower AArch64 MachineInstrs after legalization", false,
698                     false)
699 
700 namespace llvm {
701 FunctionPass *createAArch64PostLegalizerLowering() {
702   return new AArch64PostLegalizerLowering();
703 }
704 } // end namespace llvm
705