1 //===- AMDGPUInstructionSelector --------------------------------*- 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 declares the targeting of the InstructionSelector class for
10 /// AMDGPU.
11 //===----------------------------------------------------------------------===//
12 
13 #ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUINSTRUCTIONSELECTOR_H
14 #define LLVM_LIB_TARGET_AMDGPU_AMDGPUINSTRUCTIONSELECTOR_H
15 
16 #include "llvm/CodeGen/GlobalISel/InstructionSelector.h"
17 #include "llvm/IR/InstrTypes.h"
18 
19 namespace {
20 #define GET_GLOBALISEL_PREDICATE_BITSET
21 #define AMDGPUSubtarget GCNSubtarget
22 #include "AMDGPUGenGlobalISel.inc"
23 #undef GET_GLOBALISEL_PREDICATE_BITSET
24 #undef AMDGPUSubtarget
25 }
26 
27 namespace llvm {
28 
29 namespace AMDGPU {
30 struct ImageDimIntrinsicInfo;
31 }
32 
33 class AMDGPURegisterBankInfo;
34 class AMDGPUTargetMachine;
35 class BlockFrequencyInfo;
36 class ProfileSummaryInfo;
37 class GCNSubtarget;
38 class MachineInstr;
39 class MachineIRBuilder;
40 class MachineOperand;
41 class MachineRegisterInfo;
42 class RegisterBank;
43 class SIInstrInfo;
44 class SIRegisterInfo;
45 class TargetRegisterClass;
46 
47 class AMDGPUInstructionSelector final : public InstructionSelector {
48 private:
49   MachineRegisterInfo *MRI;
50   const GCNSubtarget *Subtarget;
51 
52 public:
53   AMDGPUInstructionSelector(const GCNSubtarget &STI,
54                             const AMDGPURegisterBankInfo &RBI,
55                             const AMDGPUTargetMachine &TM);
56 
57   bool select(MachineInstr &I) override;
58   static const char *getName();
59 
60   void setupMF(MachineFunction &MF, GISelKnownBits *KB,
61                CodeGenCoverage &CoverageInfo, ProfileSummaryInfo *PSI,
62                BlockFrequencyInfo *BFI) override;
63 
64 private:
65   struct GEPInfo {
66     const MachineInstr &GEP;
67     SmallVector<unsigned, 2> SgprParts;
68     SmallVector<unsigned, 2> VgprParts;
69     int64_t Imm;
70     GEPInfo(const MachineInstr &GEP) : GEP(GEP), Imm(0) { }
71   };
72 
73   bool isSGPR(Register Reg) const;
74 
75   bool isInstrUniform(const MachineInstr &MI) const;
76   bool isVCC(Register Reg, const MachineRegisterInfo &MRI) const;
77 
78   const RegisterBank *getArtifactRegBank(
79     Register Reg, const MachineRegisterInfo &MRI,
80     const TargetRegisterInfo &TRI) const;
81 
82   /// tblgen-erated 'select' implementation.
83   bool selectImpl(MachineInstr &I, CodeGenCoverage &CoverageInfo) const;
84 
85   MachineOperand getSubOperand64(MachineOperand &MO,
86                                  const TargetRegisterClass &SubRC,
87                                  unsigned SubIdx) const;
88 
89   bool constrainCopyLikeIntrin(MachineInstr &MI, unsigned NewOpc) const;
90   bool selectCOPY(MachineInstr &I) const;
91   bool selectPHI(MachineInstr &I) const;
92   bool selectG_TRUNC(MachineInstr &I) const;
93   bool selectG_SZA_EXT(MachineInstr &I) const;
94   bool selectG_CONSTANT(MachineInstr &I) const;
95   bool selectG_FNEG(MachineInstr &I) const;
96   bool selectG_FABS(MachineInstr &I) const;
97   bool selectG_AND_OR_XOR(MachineInstr &I) const;
98   bool selectG_ADD_SUB(MachineInstr &I) const;
99   bool selectG_UADDO_USUBO_UADDE_USUBE(MachineInstr &I) const;
100   bool selectG_EXTRACT(MachineInstr &I) const;
101   bool selectG_MERGE_VALUES(MachineInstr &I) const;
102   bool selectG_UNMERGE_VALUES(MachineInstr &I) const;
103   bool selectG_BUILD_VECTOR_TRUNC(MachineInstr &I) const;
104   bool selectG_PTR_ADD(MachineInstr &I) const;
105   bool selectG_IMPLICIT_DEF(MachineInstr &I) const;
106   bool selectG_INSERT(MachineInstr &I) const;
107   bool selectG_SBFX_UBFX(MachineInstr &I) const;
108 
109   bool selectInterpP1F16(MachineInstr &MI) const;
110   bool selectWritelane(MachineInstr &MI) const;
111   bool selectDivScale(MachineInstr &MI) const;
112   bool selectIntrinsicIcmp(MachineInstr &MI) const;
113   bool selectBallot(MachineInstr &I) const;
114   bool selectRelocConstant(MachineInstr &I) const;
115   bool selectGroupStaticSize(MachineInstr &I) const;
116   bool selectReturnAddress(MachineInstr &I) const;
117   bool selectG_INTRINSIC(MachineInstr &I) const;
118 
119   bool selectEndCfIntrinsic(MachineInstr &MI) const;
120   bool selectDSOrderedIntrinsic(MachineInstr &MI, Intrinsic::ID IID) const;
121   bool selectDSGWSIntrinsic(MachineInstr &MI, Intrinsic::ID IID) const;
122   bool selectDSAppendConsume(MachineInstr &MI, bool IsAppend) const;
123   bool selectSBarrier(MachineInstr &MI) const;
124 
125   bool selectImageIntrinsic(MachineInstr &MI,
126                             const AMDGPU::ImageDimIntrinsicInfo *Intr) const;
127   bool selectG_INTRINSIC_W_SIDE_EFFECTS(MachineInstr &I) const;
128   int getS_CMPOpcode(CmpInst::Predicate P, unsigned Size) const;
129   bool selectG_ICMP(MachineInstr &I) const;
130   bool hasVgprParts(ArrayRef<GEPInfo> AddrInfo) const;
131   void getAddrModeInfo(const MachineInstr &Load, const MachineRegisterInfo &MRI,
132                        SmallVectorImpl<GEPInfo> &AddrInfo) const;
133 
134   void initM0(MachineInstr &I) const;
135   bool selectG_LOAD_STORE_ATOMICRMW(MachineInstr &I) const;
136   bool selectG_AMDGPU_ATOMIC_CMPXCHG(MachineInstr &I) const;
137   bool selectG_SELECT(MachineInstr &I) const;
138   bool selectG_BRCOND(MachineInstr &I) const;
139   bool selectG_GLOBAL_VALUE(MachineInstr &I) const;
140   bool selectG_PTRMASK(MachineInstr &I) const;
141   bool selectG_EXTRACT_VECTOR_ELT(MachineInstr &I) const;
142   bool selectG_INSERT_VECTOR_ELT(MachineInstr &I) const;
143   bool selectG_SHUFFLE_VECTOR(MachineInstr &I) const;
144   bool selectAMDGPU_BUFFER_ATOMIC_FADD(MachineInstr &I) const;
145   bool selectGlobalAtomicFadd(MachineInstr &I, MachineOperand &AddrOp,
146                               MachineOperand &DataOp) const;
147   bool selectBVHIntrinsic(MachineInstr &I) const;
148   bool selectWaveAddress(MachineInstr &I) const;
149 
150   std::pair<Register, unsigned> selectVOP3ModsImpl(MachineOperand &Root,
151                                                    bool AllowAbs = true) const;
152 
153   InstructionSelector::ComplexRendererFns
154   selectVCSRC(MachineOperand &Root) const;
155 
156   InstructionSelector::ComplexRendererFns
157   selectVSRC0(MachineOperand &Root) const;
158 
159   InstructionSelector::ComplexRendererFns
160   selectVOP3Mods0(MachineOperand &Root) const;
161   InstructionSelector::ComplexRendererFns
162   selectVOP3BMods0(MachineOperand &Root) const;
163   InstructionSelector::ComplexRendererFns
164   selectVOP3OMods(MachineOperand &Root) const;
165   InstructionSelector::ComplexRendererFns
166   selectVOP3Mods(MachineOperand &Root) const;
167   InstructionSelector::ComplexRendererFns
168   selectVOP3BMods(MachineOperand &Root) const;
169 
170   ComplexRendererFns selectVOP3NoMods(MachineOperand &Root) const;
171 
172   InstructionSelector::ComplexRendererFns
173   selectVOP3Mods_nnan(MachineOperand &Root) const;
174 
175   std::pair<Register, unsigned>
176   selectVOP3PModsImpl(Register Src, const MachineRegisterInfo &MRI,
177                       bool IsDOT = false) const;
178 
179   InstructionSelector::ComplexRendererFns
180   selectVOP3PMods(MachineOperand &Root) const;
181 
182   InstructionSelector::ComplexRendererFns
183   selectVOP3PModsDOT(MachineOperand &Root) const;
184 
185   InstructionSelector::ComplexRendererFns
186   selectVOP3OpSelMods(MachineOperand &Root) const;
187 
188   InstructionSelector::ComplexRendererFns
189   selectSmrdImm(MachineOperand &Root) const;
190   InstructionSelector::ComplexRendererFns
191   selectSmrdImm32(MachineOperand &Root) const;
192   InstructionSelector::ComplexRendererFns
193   selectSmrdSgpr(MachineOperand &Root) const;
194 
195   std::pair<Register, int> selectFlatOffsetImpl(MachineOperand &Root,
196                                                 uint64_t FlatVariant) const;
197 
198   InstructionSelector::ComplexRendererFns
199   selectFlatOffset(MachineOperand &Root) const;
200   InstructionSelector::ComplexRendererFns
201   selectGlobalOffset(MachineOperand &Root) const;
202   InstructionSelector::ComplexRendererFns
203   selectScratchOffset(MachineOperand &Root) const;
204 
205   InstructionSelector::ComplexRendererFns
206   selectGlobalSAddr(MachineOperand &Root) const;
207 
208   InstructionSelector::ComplexRendererFns
209   selectScratchSAddr(MachineOperand &Root) const;
210   InstructionSelector::ComplexRendererFns
211   selectScratchSVAddr(MachineOperand &Root) const;
212 
213   InstructionSelector::ComplexRendererFns
214   selectMUBUFScratchOffen(MachineOperand &Root) const;
215   InstructionSelector::ComplexRendererFns
216   selectMUBUFScratchOffset(MachineOperand &Root) const;
217 
218   bool isDSOffsetLegal(Register Base, int64_t Offset) const;
219   bool isDSOffset2Legal(Register Base, int64_t Offset0, int64_t Offset1,
220                         unsigned Size) const;
221 
222   std::pair<Register, unsigned>
223   selectDS1Addr1OffsetImpl(MachineOperand &Root) const;
224   InstructionSelector::ComplexRendererFns
225   selectDS1Addr1Offset(MachineOperand &Root) const;
226 
227   InstructionSelector::ComplexRendererFns
228   selectDS64Bit4ByteAligned(MachineOperand &Root) const;
229 
230   InstructionSelector::ComplexRendererFns
231   selectDS128Bit8ByteAligned(MachineOperand &Root) const;
232 
233   std::pair<Register, unsigned> selectDSReadWrite2Impl(MachineOperand &Root,
234                                                        unsigned size) const;
235   InstructionSelector::ComplexRendererFns
236   selectDSReadWrite2(MachineOperand &Root, unsigned size) const;
237 
238   std::pair<Register, int64_t>
239   getPtrBaseWithConstantOffset(Register Root,
240                                const MachineRegisterInfo &MRI) const;
241 
242   // Parse out a chain of up to two g_ptr_add instructions.
243   // g_ptr_add (n0, _)
244   // g_ptr_add (n0, (n1 = g_ptr_add n2, n3))
245   struct MUBUFAddressData {
246     Register N0, N2, N3;
247     int64_t Offset = 0;
248   };
249 
250   bool shouldUseAddr64(MUBUFAddressData AddrData) const;
251 
252   void splitIllegalMUBUFOffset(MachineIRBuilder &B,
253                                Register &SOffset, int64_t &ImmOffset) const;
254 
255   MUBUFAddressData parseMUBUFAddress(Register Src) const;
256 
257   bool selectMUBUFAddr64Impl(MachineOperand &Root, Register &VAddr,
258                              Register &RSrcReg, Register &SOffset,
259                              int64_t &Offset) const;
260 
261   bool selectMUBUFOffsetImpl(MachineOperand &Root, Register &RSrcReg,
262                              Register &SOffset, int64_t &Offset) const;
263 
264   InstructionSelector::ComplexRendererFns
265   selectMUBUFAddr64(MachineOperand &Root) const;
266 
267   InstructionSelector::ComplexRendererFns
268   selectMUBUFOffset(MachineOperand &Root) const;
269 
270   InstructionSelector::ComplexRendererFns
271   selectMUBUFOffsetAtomic(MachineOperand &Root) const;
272 
273   InstructionSelector::ComplexRendererFns
274   selectMUBUFAddr64Atomic(MachineOperand &Root) const;
275 
276   ComplexRendererFns selectSMRDBufferImm(MachineOperand &Root) const;
277   ComplexRendererFns selectSMRDBufferImm32(MachineOperand &Root) const;
278 
279   void renderTruncImm32(MachineInstrBuilder &MIB, const MachineInstr &MI,
280                         int OpIdx = -1) const;
281 
282   void renderTruncTImm(MachineInstrBuilder &MIB, const MachineInstr &MI,
283                        int OpIdx) const;
284 
285   void renderNegateImm(MachineInstrBuilder &MIB, const MachineInstr &MI,
286                        int OpIdx) const;
287 
288   void renderBitcastImm(MachineInstrBuilder &MIB, const MachineInstr &MI,
289                         int OpIdx) const;
290 
291   void renderPopcntImm(MachineInstrBuilder &MIB, const MachineInstr &MI,
292                        int OpIdx) const;
293   void renderExtractCPol(MachineInstrBuilder &MIB, const MachineInstr &MI,
294                          int OpIdx) const;
295   void renderExtractSWZ(MachineInstrBuilder &MIB, const MachineInstr &MI,
296                         int OpIdx) const;
297   void renderSetGLC(MachineInstrBuilder &MIB, const MachineInstr &MI,
298                     int OpIdx) const;
299 
300   void renderFrameIndex(MachineInstrBuilder &MIB, const MachineInstr &MI,
301                         int OpIdx) const;
302 
303   bool isInlineImmediate16(int64_t Imm) const;
304   bool isInlineImmediate32(int64_t Imm) const;
305   bool isInlineImmediate64(int64_t Imm) const;
306   bool isInlineImmediate(const APFloat &Imm) const;
307 
308   // Returns true if TargetOpcode::G_AND MachineInstr `MI`'s masking of the
309   // shift amount operand's `ShAmtBits` bits is unneeded.
310   bool isUnneededShiftMask(const MachineInstr &MI, unsigned ShAmtBits) const;
311 
312   const SIInstrInfo &TII;
313   const SIRegisterInfo &TRI;
314   const AMDGPURegisterBankInfo &RBI;
315   const AMDGPUTargetMachine &TM;
316   const GCNSubtarget &STI;
317   bool EnableLateStructurizeCFG;
318 #define GET_GLOBALISEL_PREDICATES_DECL
319 #define AMDGPUSubtarget GCNSubtarget
320 #include "AMDGPUGenGlobalISel.inc"
321 #undef GET_GLOBALISEL_PREDICATES_DECL
322 #undef AMDGPUSubtarget
323 
324 #define GET_GLOBALISEL_TEMPORARIES_DECL
325 #include "AMDGPUGenGlobalISel.inc"
326 #undef GET_GLOBALISEL_TEMPORARIES_DECL
327 };
328 
329 } // End llvm namespace.
330 #endif
331