xref: /freebsd/contrib/llvm-project/llvm/include/llvm/CodeGen/GlobalISel/LegalizerHelper.h (revision 5956d97f4b3204318ceb6aa9c77bd0bc6ea87a41)
1 //== llvm/CodeGen/GlobalISel/LegalizerHelper.h ---------------- -*- 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 A pass to convert the target-illegal operations created by IR -> MIR
10 /// translation into ones the target expects to be able to select. This may
11 /// occur in multiple phases, for example G_ADD <2 x i8> -> G_ADD <2 x i16> ->
12 /// G_ADD <4 x i16>.
13 ///
14 /// The LegalizerHelper class is where most of the work happens, and is
15 /// designed to be callable from other passes that find themselves with an
16 /// illegal instruction.
17 //
18 //===----------------------------------------------------------------------===//
19 
20 #ifndef LLVM_CODEGEN_GLOBALISEL_LEGALIZERHELPER_H
21 #define LLVM_CODEGEN_GLOBALISEL_LEGALIZERHELPER_H
22 
23 #include "llvm/CodeGen/GlobalISel/CallLowering.h"
24 #include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h"
25 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
26 #include "llvm/CodeGen/LowLevelType.h"
27 #include "llvm/CodeGen/MachineFunctionPass.h"
28 #include "llvm/CodeGen/RuntimeLibcalls.h"
29 
30 namespace llvm {
31 // Forward declarations.
32 class LegalizerInfo;
33 class MachineRegisterInfo;
34 class GISelChangeObserver;
35 class LostDebugLocObserver;
36 class TargetLowering;
37 
38 class LegalizerHelper {
39 public:
40   /// Expose MIRBuilder so clients can set their own RecordInsertInstruction
41   /// functions
42   MachineIRBuilder &MIRBuilder;
43 
44   /// To keep track of changes made by the LegalizerHelper.
45   GISelChangeObserver &Observer;
46 
47 private:
48   MachineRegisterInfo &MRI;
49   const LegalizerInfo &LI;
50   const TargetLowering &TLI;
51 
52 public:
53   enum LegalizeResult {
54     /// Instruction was already legal and no change was made to the
55     /// MachineFunction.
56     AlreadyLegal,
57 
58     /// Instruction has been legalized and the MachineFunction changed.
59     Legalized,
60 
61     /// Some kind of error has occurred and we could not legalize this
62     /// instruction.
63     UnableToLegalize,
64   };
65 
66   /// Expose LegalizerInfo so the clients can re-use.
67   const LegalizerInfo &getLegalizerInfo() const { return LI; }
68   const TargetLowering &getTargetLowering() const { return TLI; }
69 
70   LegalizerHelper(MachineFunction &MF, GISelChangeObserver &Observer,
71                   MachineIRBuilder &B);
72   LegalizerHelper(MachineFunction &MF, const LegalizerInfo &LI,
73                   GISelChangeObserver &Observer, MachineIRBuilder &B);
74 
75   /// Replace \p MI by a sequence of legal instructions that can implement the
76   /// same operation. Note that this means \p MI may be deleted, so any iterator
77   /// steps should be performed before calling this function. \p Helper should
78   /// be initialized to the MachineFunction containing \p MI.
79   ///
80   /// Considered as an opaque blob, the legal code will use and define the same
81   /// registers as \p MI.
82   LegalizeResult legalizeInstrStep(MachineInstr &MI,
83                                    LostDebugLocObserver &LocObserver);
84 
85   /// Legalize an instruction by emiting a runtime library call instead.
86   LegalizeResult libcall(MachineInstr &MI, LostDebugLocObserver &LocObserver);
87 
88   /// Legalize an instruction by reducing the width of the underlying scalar
89   /// type.
90   LegalizeResult narrowScalar(MachineInstr &MI, unsigned TypeIdx, LLT NarrowTy);
91 
92   /// Legalize an instruction by performing the operation on a wider scalar type
93   /// (for example a 16-bit addition can be safely performed at 32-bits
94   /// precision, ignoring the unused bits).
95   LegalizeResult widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy);
96 
97   /// Legalize an instruction by replacing the value type
98   LegalizeResult bitcast(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
99 
100   /// Legalize an instruction by splitting it into simpler parts, hopefully
101   /// understood by the target.
102   LegalizeResult lower(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
103 
104   /// Legalize a vector instruction by splitting into multiple components, each
105   /// acting on the same scalar type as the original but with fewer elements.
106   LegalizeResult fewerElementsVector(MachineInstr &MI, unsigned TypeIdx,
107                                      LLT NarrowTy);
108 
109   /// Legalize a vector instruction by increasing the number of vector elements
110   /// involved and ignoring the added elements later.
111   LegalizeResult moreElementsVector(MachineInstr &MI, unsigned TypeIdx,
112                                     LLT MoreTy);
113 
114   /// Cast the given value to an LLT::scalar with an equivalent size. Returns
115   /// the register to use if an instruction was inserted. Returns the original
116   /// register if no coercion was necessary.
117   //
118   // This may also fail and return Register() if there is no legal way to cast.
119   Register coerceToScalar(Register Val);
120 
121   /// Legalize a single operand \p OpIdx of the machine instruction \p MI as a
122   /// Use by extending the operand's type to \p WideTy using the specified \p
123   /// ExtOpcode for the extension instruction, and replacing the vreg of the
124   /// operand in place.
125   void widenScalarSrc(MachineInstr &MI, LLT WideTy, unsigned OpIdx,
126                       unsigned ExtOpcode);
127 
128   /// Legalize a single operand \p OpIdx of the machine instruction \p MI as a
129   /// Use by truncating the operand's type to \p NarrowTy using G_TRUNC, and
130   /// replacing the vreg of the operand in place.
131   void narrowScalarSrc(MachineInstr &MI, LLT NarrowTy, unsigned OpIdx);
132 
133   /// Legalize a single operand \p OpIdx of the machine instruction \p MI as a
134   /// Def by extending the operand's type to \p WideTy and truncating it back
135   /// with the \p TruncOpcode, and replacing the vreg of the operand in place.
136   void widenScalarDst(MachineInstr &MI, LLT WideTy, unsigned OpIdx = 0,
137                       unsigned TruncOpcode = TargetOpcode::G_TRUNC);
138 
139   // Legalize a single operand \p OpIdx of the machine instruction \p MI as a
140   // Def by truncating the operand's type to \p NarrowTy, replacing in place and
141   // extending back with \p ExtOpcode.
142   void narrowScalarDst(MachineInstr &MI, LLT NarrowTy, unsigned OpIdx,
143                        unsigned ExtOpcode);
144   /// Legalize a single operand \p OpIdx of the machine instruction \p MI as a
145   /// Def by performing it with additional vector elements and extracting the
146   /// result elements, and replacing the vreg of the operand in place.
147   void moreElementsVectorDst(MachineInstr &MI, LLT MoreTy, unsigned OpIdx);
148 
149   /// Legalize a single operand \p OpIdx of the machine instruction \p MI as a
150   /// Use by producing a vector with undefined high elements, extracting the
151   /// original vector type, and replacing the vreg of the operand in place.
152   void moreElementsVectorSrc(MachineInstr &MI, LLT MoreTy, unsigned OpIdx);
153 
154   /// Legalize a single operand \p OpIdx of the machine instruction \p MI as a
155   /// use by inserting a G_BITCAST to \p CastTy
156   void bitcastSrc(MachineInstr &MI, LLT CastTy, unsigned OpIdx);
157 
158   /// Legalize a single operand \p OpIdx of the machine instruction \p MI as a
159   /// def by inserting a G_BITCAST from \p CastTy
160   void bitcastDst(MachineInstr &MI, LLT CastTy, unsigned OpIdx);
161 
162   /// Widen \p OrigReg to \p WideTy by merging to a wider type, padding with
163   /// G_IMPLICIT_DEF, and producing dead results.
164   Register widenWithUnmerge(LLT WideTy, Register OrigReg);
165 
166 private:
167   LegalizeResult
168   widenScalarMergeValues(MachineInstr &MI, unsigned TypeIdx, LLT WideTy);
169   LegalizeResult
170   widenScalarUnmergeValues(MachineInstr &MI, unsigned TypeIdx, LLT WideTy);
171   LegalizeResult
172   widenScalarExtract(MachineInstr &MI, unsigned TypeIdx, LLT WideTy);
173   LegalizeResult
174   widenScalarInsert(MachineInstr &MI, unsigned TypeIdx, LLT WideTy);
175   LegalizeResult widenScalarAddSubOverflow(MachineInstr &MI, unsigned TypeIdx,
176                                            LLT WideTy);
177   LegalizeResult widenScalarAddSubShlSat(MachineInstr &MI, unsigned TypeIdx,
178                                          LLT WideTy);
179   LegalizeResult widenScalarMulo(MachineInstr &MI, unsigned TypeIdx,
180                                  LLT WideTy);
181 
182   /// Helper function to split a wide generic register into bitwise blocks with
183   /// the given Type (which implies the number of blocks needed). The generic
184   /// registers created are appended to Ops, starting at bit 0 of Reg.
185   void extractParts(Register Reg, LLT Ty, int NumParts,
186                     SmallVectorImpl<Register> &VRegs);
187 
188   /// Version which handles irregular splits.
189   bool extractParts(Register Reg, LLT RegTy, LLT MainTy,
190                     LLT &LeftoverTy,
191                     SmallVectorImpl<Register> &VRegs,
192                     SmallVectorImpl<Register> &LeftoverVRegs);
193 
194   /// Version which handles irregular sub-vector splits.
195   void extractVectorParts(Register Reg, unsigned NumElst,
196                           SmallVectorImpl<Register> &VRegs);
197 
198   /// Helper function to build a wide generic register \p DstReg of type \p
199   /// RegTy from smaller parts. This will produce a G_MERGE_VALUES,
200   /// G_BUILD_VECTOR, G_CONCAT_VECTORS, or sequence of G_INSERT as appropriate
201   /// for the types.
202   ///
203   /// \p PartRegs must be registers of type \p PartTy.
204   ///
205   /// If \p ResultTy does not evenly break into \p PartTy sized pieces, the
206   /// remainder must be specified with \p LeftoverRegs of type \p LeftoverTy.
207   void insertParts(Register DstReg, LLT ResultTy,
208                    LLT PartTy, ArrayRef<Register> PartRegs,
209                    LLT LeftoverTy = LLT(), ArrayRef<Register> LeftoverRegs = {});
210 
211   /// Merge \p PartRegs with different types into \p DstReg.
212   void mergeMixedSubvectors(Register DstReg, ArrayRef<Register> PartRegs);
213 
214   void appendVectorElts(SmallVectorImpl<Register> &Elts, Register Reg);
215 
216   /// Unmerge \p SrcReg into smaller sized values, and append them to \p
217   /// Parts. The elements of \p Parts will be the greatest common divisor type
218   /// of \p DstTy, \p NarrowTy and the type of \p SrcReg. This will compute and
219   /// return the GCD type.
220   LLT extractGCDType(SmallVectorImpl<Register> &Parts, LLT DstTy,
221                      LLT NarrowTy, Register SrcReg);
222 
223   /// Unmerge \p SrcReg into \p GCDTy typed registers. This will append all of
224   /// the unpacked registers to \p Parts. This version is if the common unmerge
225   /// type is already known.
226   void extractGCDType(SmallVectorImpl<Register> &Parts, LLT GCDTy,
227                       Register SrcReg);
228 
229   /// Produce a merge of values in \p VRegs to define \p DstReg. Perform a merge
230   /// from the least common multiple type, and convert as appropriate to \p
231   /// DstReg.
232   ///
233   /// \p VRegs should each have type \p GCDTy. This type should be greatest
234   /// common divisor type of \p DstReg, \p NarrowTy, and an undetermined source
235   /// type.
236   ///
237   /// \p NarrowTy is the desired result merge source type. If the source value
238   /// needs to be widened to evenly cover \p DstReg, inserts high bits
239   /// corresponding to the extension opcode \p PadStrategy.
240   ///
241   /// \p VRegs will be cleared, and the the result \p NarrowTy register pieces
242   /// will replace it. Returns The complete LCMTy that \p VRegs will cover when
243   /// merged.
244   LLT buildLCMMergePieces(LLT DstTy, LLT NarrowTy, LLT GCDTy,
245                           SmallVectorImpl<Register> &VRegs,
246                           unsigned PadStrategy = TargetOpcode::G_ANYEXT);
247 
248   /// Merge the values in \p RemergeRegs to an \p LCMTy typed value. Extract the
249   /// low bits into \p DstReg. This is intended to use the outputs from
250   /// buildLCMMergePieces after processing.
251   void buildWidenedRemergeToDst(Register DstReg, LLT LCMTy,
252                                 ArrayRef<Register> RemergeRegs);
253 
254   /// Perform generic multiplication of values held in multiple registers.
255   /// Generated instructions use only types NarrowTy and i1.
256   /// Destination can be same or two times size of the source.
257   void multiplyRegisters(SmallVectorImpl<Register> &DstRegs,
258                          ArrayRef<Register> Src1Regs,
259                          ArrayRef<Register> Src2Regs, LLT NarrowTy);
260 
261   void changeOpcode(MachineInstr &MI, unsigned NewOpcode);
262 
263   LegalizeResult tryNarrowPow2Reduction(MachineInstr &MI, Register SrcReg,
264                                         LLT SrcTy, LLT NarrowTy,
265                                         unsigned ScalarOpc);
266 
267   // Memcpy family legalization helpers.
268   LegalizeResult lowerMemset(MachineInstr &MI, Register Dst, Register Val,
269                              uint64_t KnownLen, Align Alignment,
270                              bool IsVolatile);
271   LegalizeResult lowerMemcpyInline(MachineInstr &MI, Register Dst, Register Src,
272                                    uint64_t KnownLen, Align DstAlign,
273                                    Align SrcAlign, bool IsVolatile);
274   LegalizeResult lowerMemcpy(MachineInstr &MI, Register Dst, Register Src,
275                              uint64_t KnownLen, uint64_t Limit, Align DstAlign,
276                              Align SrcAlign, bool IsVolatile);
277   LegalizeResult lowerMemmove(MachineInstr &MI, Register Dst, Register Src,
278                               uint64_t KnownLen, Align DstAlign, Align SrcAlign,
279                               bool IsVolatile);
280 
281 public:
282   /// Return the alignment to use for a stack temporary object with the given
283   /// type.
284   Align getStackTemporaryAlignment(LLT Type, Align MinAlign = Align()) const;
285 
286   /// Create a stack temporary based on the size in bytes and the alignment
287   MachineInstrBuilder createStackTemporary(TypeSize Bytes, Align Alignment,
288                                            MachinePointerInfo &PtrInfo);
289 
290   /// Get a pointer to vector element \p Index located in memory for a vector of
291   /// type \p VecTy starting at a base address of \p VecPtr. If \p Index is out
292   /// of bounds the returned pointer is unspecified, but will be within the
293   /// vector bounds.
294   Register getVectorElementPointer(Register VecPtr, LLT VecTy, Register Index);
295 
296   /// Handles most opcodes. Split \p MI into same instruction on sub-vectors or
297   /// scalars with \p NumElts elements (1 for scalar). Supports uneven splits:
298   /// there can be leftover sub-vector with fewer then \p NumElts or a leftover
299   /// scalar. To avoid this use moreElements first and set MI number of elements
300   /// to multiple of \p NumElts. Non-vector operands that should be used on all
301   /// sub-instructions without split are listed in \p NonVecOpIndices.
302   LegalizeResult fewerElementsVectorMultiEltType(
303       GenericMachineInstr &MI, unsigned NumElts,
304       std::initializer_list<unsigned> NonVecOpIndices = {});
305 
306   LegalizeResult fewerElementsVectorPhi(GenericMachineInstr &MI,
307                                         unsigned NumElts);
308 
309   LegalizeResult moreElementsVectorPhi(MachineInstr &MI, unsigned TypeIdx,
310                                        LLT MoreTy);
311   LegalizeResult moreElementsVectorShuffle(MachineInstr &MI, unsigned TypeIdx,
312                                            LLT MoreTy);
313 
314   LegalizeResult fewerElementsVectorUnmergeValues(MachineInstr &MI,
315                                                   unsigned TypeIdx,
316                                                   LLT NarrowTy);
317   LegalizeResult fewerElementsVectorMerge(MachineInstr &MI, unsigned TypeIdx,
318                                           LLT NarrowTy);
319   LegalizeResult fewerElementsVectorExtractInsertVectorElt(MachineInstr &MI,
320                                                            unsigned TypeIdx,
321                                                            LLT NarrowTy);
322 
323   LegalizeResult reduceLoadStoreWidth(GLoadStore &MI, unsigned TypeIdx,
324                                       LLT NarrowTy);
325 
326   LegalizeResult fewerElementsVectorSextInReg(MachineInstr &MI, unsigned TypeIdx,
327                                               LLT NarrowTy);
328 
329   LegalizeResult narrowScalarShiftByConstant(MachineInstr &MI, const APInt &Amt,
330                                              LLT HalfTy, LLT ShiftAmtTy);
331 
332   LegalizeResult fewerElementsVectorReductions(MachineInstr &MI,
333                                                unsigned TypeIdx, LLT NarrowTy);
334 
335   LegalizeResult fewerElementsVectorShuffle(MachineInstr &MI, unsigned TypeIdx,
336                                             LLT NarrowTy);
337 
338   LegalizeResult narrowScalarShift(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
339   LegalizeResult narrowScalarAddSub(MachineInstr &MI, unsigned TypeIdx,
340                                     LLT NarrowTy);
341   LegalizeResult narrowScalarMul(MachineInstr &MI, LLT Ty);
342   LegalizeResult narrowScalarFPTOI(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
343   LegalizeResult narrowScalarExtract(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
344   LegalizeResult narrowScalarInsert(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
345 
346   LegalizeResult narrowScalarBasic(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
347   LegalizeResult narrowScalarExt(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
348   LegalizeResult narrowScalarSelect(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
349   LegalizeResult narrowScalarCTLZ(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
350   LegalizeResult narrowScalarCTTZ(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
351   LegalizeResult narrowScalarCTPOP(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
352 
353   /// Perform Bitcast legalize action on G_EXTRACT_VECTOR_ELT.
354   LegalizeResult bitcastExtractVectorElt(MachineInstr &MI, unsigned TypeIdx,
355                                          LLT CastTy);
356 
357   /// Perform Bitcast legalize action on G_INSERT_VECTOR_ELT.
358   LegalizeResult bitcastInsertVectorElt(MachineInstr &MI, unsigned TypeIdx,
359                                         LLT CastTy);
360 
361   LegalizeResult lowerBitcast(MachineInstr &MI);
362   LegalizeResult lowerLoad(GAnyLoad &MI);
363   LegalizeResult lowerStore(GStore &MI);
364   LegalizeResult lowerBitCount(MachineInstr &MI);
365   LegalizeResult lowerFunnelShiftWithInverse(MachineInstr &MI);
366   LegalizeResult lowerFunnelShiftAsShifts(MachineInstr &MI);
367   LegalizeResult lowerFunnelShift(MachineInstr &MI);
368   LegalizeResult lowerRotateWithReverseRotate(MachineInstr &MI);
369   LegalizeResult lowerRotate(MachineInstr &MI);
370 
371   LegalizeResult lowerU64ToF32BitOps(MachineInstr &MI);
372   LegalizeResult lowerUITOFP(MachineInstr &MI);
373   LegalizeResult lowerSITOFP(MachineInstr &MI);
374   LegalizeResult lowerFPTOUI(MachineInstr &MI);
375   LegalizeResult lowerFPTOSI(MachineInstr &MI);
376 
377   LegalizeResult lowerFPTRUNC_F64_TO_F16(MachineInstr &MI);
378   LegalizeResult lowerFPTRUNC(MachineInstr &MI);
379   LegalizeResult lowerFPOWI(MachineInstr &MI);
380 
381   LegalizeResult lowerMinMax(MachineInstr &MI);
382   LegalizeResult lowerFCopySign(MachineInstr &MI);
383   LegalizeResult lowerFMinNumMaxNum(MachineInstr &MI);
384   LegalizeResult lowerFMad(MachineInstr &MI);
385   LegalizeResult lowerIntrinsicRound(MachineInstr &MI);
386   LegalizeResult lowerFFloor(MachineInstr &MI);
387   LegalizeResult lowerMergeValues(MachineInstr &MI);
388   LegalizeResult lowerUnmergeValues(MachineInstr &MI);
389   LegalizeResult lowerExtractInsertVectorElt(MachineInstr &MI);
390   LegalizeResult lowerShuffleVector(MachineInstr &MI);
391   LegalizeResult lowerDynStackAlloc(MachineInstr &MI);
392   LegalizeResult lowerExtract(MachineInstr &MI);
393   LegalizeResult lowerInsert(MachineInstr &MI);
394   LegalizeResult lowerSADDO_SSUBO(MachineInstr &MI);
395   LegalizeResult lowerAddSubSatToMinMax(MachineInstr &MI);
396   LegalizeResult lowerAddSubSatToAddoSubo(MachineInstr &MI);
397   LegalizeResult lowerShlSat(MachineInstr &MI);
398   LegalizeResult lowerBswap(MachineInstr &MI);
399   LegalizeResult lowerBitreverse(MachineInstr &MI);
400   LegalizeResult lowerReadWriteRegister(MachineInstr &MI);
401   LegalizeResult lowerSMULH_UMULH(MachineInstr &MI);
402   LegalizeResult lowerSelect(MachineInstr &MI);
403   LegalizeResult lowerDIVREM(MachineInstr &MI);
404   LegalizeResult lowerAbsToAddXor(MachineInstr &MI);
405   LegalizeResult lowerAbsToMaxNeg(MachineInstr &MI);
406   LegalizeResult lowerVectorReduction(MachineInstr &MI);
407   LegalizeResult lowerMemcpyInline(MachineInstr &MI);
408   LegalizeResult lowerMemCpyFamily(MachineInstr &MI, unsigned MaxLen = 0);
409 };
410 
411 /// Helper function that creates a libcall to the given \p Name using the given
412 /// calling convention \p CC.
413 LegalizerHelper::LegalizeResult
414 createLibcall(MachineIRBuilder &MIRBuilder, const char *Name,
415               const CallLowering::ArgInfo &Result,
416               ArrayRef<CallLowering::ArgInfo> Args, CallingConv::ID CC);
417 
418 /// Helper function that creates the given libcall.
419 LegalizerHelper::LegalizeResult
420 createLibcall(MachineIRBuilder &MIRBuilder, RTLIB::Libcall Libcall,
421               const CallLowering::ArgInfo &Result,
422               ArrayRef<CallLowering::ArgInfo> Args);
423 
424 /// Create a libcall to memcpy et al.
425 LegalizerHelper::LegalizeResult
426 createMemLibcall(MachineIRBuilder &MIRBuilder, MachineRegisterInfo &MRI,
427                  MachineInstr &MI, LostDebugLocObserver &LocObserver);
428 
429 } // End namespace llvm.
430 
431 #endif
432