1 | //===-- AMDGPUISelDAGToDAG.h - A dag to dag inst selector for AMDGPU ----===// |
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 | /// Defines an instruction selector for the AMDGPU target. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUISELDAGTODAG_H |
15 | #define LLVM_LIB_TARGET_AMDGPU_AMDGPUISELDAGTODAG_H |
16 | |
17 | #include "GCNSubtarget.h" |
18 | #include "SIMachineFunctionInfo.h" |
19 | #include "SIModeRegisterDefaults.h" |
20 | #include "llvm/Analysis/ValueTracking.h" |
21 | #include "llvm/CodeGen/SelectionDAGISel.h" |
22 | #include "llvm/Target/TargetMachine.h" |
23 | |
24 | namespace llvm { |
25 | |
26 | static inline bool getConstantValue(SDValue N, uint32_t &Out) { |
27 | // This is only used for packed vectors, where using 0 for undef should |
28 | // always be good. |
29 | if (N.isUndef()) { |
30 | Out = 0; |
31 | return true; |
32 | } |
33 | |
34 | if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: N)) { |
35 | Out = C->getAPIntValue().getSExtValue(); |
36 | return true; |
37 | } |
38 | |
39 | if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val&: N)) { |
40 | Out = C->getValueAPF().bitcastToAPInt().getSExtValue(); |
41 | return true; |
42 | } |
43 | |
44 | return false; |
45 | } |
46 | |
47 | // TODO: Handle undef as zero |
48 | static inline SDNode *packConstantV2I16(const SDNode *N, SelectionDAG &DAG) { |
49 | assert(N->getOpcode() == ISD::BUILD_VECTOR && N->getNumOperands() == 2); |
50 | uint32_t LHSVal, RHSVal; |
51 | if (getConstantValue(N: N->getOperand(Num: 0), Out&: LHSVal) && |
52 | getConstantValue(N: N->getOperand(Num: 1), Out&: RHSVal)) { |
53 | SDLoc SL(N); |
54 | uint32_t K = (LHSVal & 0xffff) | (RHSVal << 16); |
55 | return DAG.getMachineNode(Opcode: AMDGPU::S_MOV_B32, dl: SL, VT: N->getValueType(ResNo: 0), |
56 | Op1: DAG.getTargetConstant(Val: K, DL: SL, VT: MVT::i32)); |
57 | } |
58 | |
59 | return nullptr; |
60 | } |
61 | |
62 | /// AMDGPU specific code to select AMDGPU machine instructions for |
63 | /// SelectionDAG operations. |
64 | class AMDGPUDAGToDAGISel : public SelectionDAGISel { |
65 | // Subtarget - Keep a pointer to the AMDGPU Subtarget around so that we can |
66 | // make the right decision when generating code for different targets. |
67 | const GCNSubtarget *Subtarget; |
68 | |
69 | // Default FP mode for the current function. |
70 | SIModeRegisterDefaults Mode; |
71 | |
72 | // Instructions that will be lowered with a final instruction that zeros the |
73 | // high result bits. |
74 | bool fp16SrcZerosHighBits(unsigned Opc) const; |
75 | |
76 | public: |
77 | AMDGPUDAGToDAGISel() = delete; |
78 | |
79 | explicit AMDGPUDAGToDAGISel(TargetMachine &TM, CodeGenOptLevel OptLevel); |
80 | |
81 | bool runOnMachineFunction(MachineFunction &MF) override; |
82 | bool matchLoadD16FromBuildVector(SDNode *N) const; |
83 | void PreprocessISelDAG() override; |
84 | void Select(SDNode *N) override; |
85 | void PostprocessISelDAG() override; |
86 | |
87 | protected: |
88 | void SelectBuildVector(SDNode *N, unsigned RegClassID); |
89 | void SelectVectorShuffle(SDNode *N); |
90 | |
91 | private: |
92 | std::pair<SDValue, SDValue> foldFrameIndex(SDValue N) const; |
93 | |
94 | bool isInlineImmediate(const SDNode *N) const; |
95 | |
96 | bool isInlineImmediate(const APInt &Imm) const { |
97 | return Subtarget->getInstrInfo()->isInlineConstant(Imm); |
98 | } |
99 | |
100 | bool isInlineImmediate(const APFloat &Imm) const { |
101 | return Subtarget->getInstrInfo()->isInlineConstant(Imm); |
102 | } |
103 | |
104 | bool isVGPRImm(const SDNode *N) const; |
105 | bool isUniformLoad(const SDNode *N) const; |
106 | bool isUniformBr(const SDNode *N) const; |
107 | |
108 | // Returns true if ISD::AND SDNode `N`'s masking of the shift amount operand's |
109 | // `ShAmtBits` bits is unneeded. |
110 | bool isUnneededShiftMask(const SDNode *N, unsigned ShAmtBits) const; |
111 | |
112 | bool isBaseWithConstantOffset64(SDValue Addr, SDValue &LHS, |
113 | SDValue &RHS) const; |
114 | |
115 | MachineSDNode *buildSMovImm64(SDLoc &DL, uint64_t Val, EVT VT) const; |
116 | |
117 | SDNode *glueCopyToOp(SDNode *N, SDValue NewChain, SDValue Glue) const; |
118 | SDNode *glueCopyToM0(SDNode *N, SDValue Val) const; |
119 | SDNode *glueCopyToM0LDSInit(SDNode *N) const; |
120 | |
121 | const TargetRegisterClass *getOperandRegClass(SDNode *N, unsigned OpNo) const; |
122 | virtual bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base, SDValue &Offset); |
123 | virtual bool SelectADDRIndirect(SDValue Addr, SDValue &Base, SDValue &Offset); |
124 | bool isDSOffsetLegal(SDValue Base, unsigned Offset) const; |
125 | bool isDSOffset2Legal(SDValue Base, unsigned Offset0, unsigned Offset1, |
126 | unsigned Size) const; |
127 | |
128 | bool isFlatScratchBaseLegal(SDValue Addr) const; |
129 | bool isFlatScratchBaseLegalSV(SDValue Addr) const; |
130 | bool isFlatScratchBaseLegalSVImm(SDValue Addr) const; |
131 | bool isSOffsetLegalWithImmOffset(SDValue *SOffset, bool Imm32Only, |
132 | bool IsBuffer, int64_t ImmOffset = 0) const; |
133 | |
134 | bool SelectDS1Addr1Offset(SDValue Ptr, SDValue &Base, SDValue &Offset) const; |
135 | bool SelectDS64Bit4ByteAligned(SDValue Ptr, SDValue &Base, SDValue &Offset0, |
136 | SDValue &Offset1) const; |
137 | bool SelectDS128Bit8ByteAligned(SDValue Ptr, SDValue &Base, SDValue &Offset0, |
138 | SDValue &Offset1) const; |
139 | bool SelectDSReadWrite2(SDValue Ptr, SDValue &Base, SDValue &Offset0, |
140 | SDValue &Offset1, unsigned Size) const; |
141 | bool SelectMUBUF(SDValue Addr, SDValue &SRsrc, SDValue &VAddr, |
142 | SDValue &SOffset, SDValue &Offset, SDValue &Offen, |
143 | SDValue &Idxen, SDValue &Addr64) const; |
144 | bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, SDValue &VAddr, |
145 | SDValue &SOffset, SDValue &Offset) const; |
146 | bool SelectMUBUFScratchOffen(SDNode *Parent, SDValue Addr, SDValue &RSrc, |
147 | SDValue &VAddr, SDValue &SOffset, |
148 | SDValue &ImmOffset) const; |
149 | bool SelectMUBUFScratchOffset(SDNode *Parent, SDValue Addr, SDValue &SRsrc, |
150 | SDValue &Soffset, SDValue &Offset) const; |
151 | |
152 | bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset, |
153 | SDValue &Offset) const; |
154 | bool SelectBUFSOffset(SDValue Addr, SDValue &SOffset) const; |
155 | |
156 | bool SelectFlatOffsetImpl(SDNode *N, SDValue Addr, SDValue &VAddr, |
157 | SDValue &Offset, uint64_t FlatVariant) const; |
158 | bool SelectFlatOffset(SDNode *N, SDValue Addr, SDValue &VAddr, |
159 | SDValue &Offset) const; |
160 | bool SelectGlobalOffset(SDNode *N, SDValue Addr, SDValue &VAddr, |
161 | SDValue &Offset) const; |
162 | bool SelectScratchOffset(SDNode *N, SDValue Addr, SDValue &VAddr, |
163 | SDValue &Offset) const; |
164 | bool SelectGlobalSAddr(SDNode *N, SDValue Addr, SDValue &SAddr, |
165 | SDValue &VOffset, SDValue &Offset) const; |
166 | bool SelectScratchSAddr(SDNode *N, SDValue Addr, SDValue &SAddr, |
167 | SDValue &Offset) const; |
168 | bool checkFlatScratchSVSSwizzleBug(SDValue VAddr, SDValue SAddr, |
169 | uint64_t ImmOffset) const; |
170 | bool SelectScratchSVAddr(SDNode *N, SDValue Addr, SDValue &VAddr, |
171 | SDValue &SAddr, SDValue &Offset) const; |
172 | |
173 | bool SelectSMRDOffset(SDValue ByteOffsetNode, SDValue *SOffset, |
174 | SDValue *Offset, bool Imm32Only = false, |
175 | bool IsBuffer = false, bool HasSOffset = false, |
176 | int64_t ImmOffset = 0) const; |
177 | SDValue Expand32BitAddress(SDValue Addr) const; |
178 | bool SelectSMRDBaseOffset(SDValue Addr, SDValue &SBase, SDValue *SOffset, |
179 | SDValue *Offset, bool Imm32Only = false, |
180 | bool IsBuffer = false, bool HasSOffset = false, |
181 | int64_t ImmOffset = 0) const; |
182 | bool SelectSMRD(SDValue Addr, SDValue &SBase, SDValue *SOffset, |
183 | SDValue *Offset, bool Imm32Only = false) const; |
184 | bool SelectSMRDImm(SDValue Addr, SDValue &SBase, SDValue &Offset) const; |
185 | bool SelectSMRDImm32(SDValue Addr, SDValue &SBase, SDValue &Offset) const; |
186 | bool SelectSMRDSgpr(SDValue Addr, SDValue &SBase, SDValue &SOffset) const; |
187 | bool SelectSMRDSgprImm(SDValue Addr, SDValue &SBase, SDValue &SOffset, |
188 | SDValue &Offset) const; |
189 | bool SelectSMRDBufferImm(SDValue N, SDValue &Offset) const; |
190 | bool SelectSMRDBufferImm32(SDValue N, SDValue &Offset) const; |
191 | bool SelectSMRDBufferSgprImm(SDValue N, SDValue &SOffset, |
192 | SDValue &Offset) const; |
193 | bool SelectSMRDPrefetchImm(SDValue Addr, SDValue &SBase, |
194 | SDValue &Offset) const; |
195 | bool SelectMOVRELOffset(SDValue Index, SDValue &Base, SDValue &Offset) const; |
196 | |
197 | bool SelectVOP3ModsImpl(SDValue In, SDValue &Src, unsigned &SrcMods, |
198 | bool IsCanonicalizing = true, |
199 | bool AllowAbs = true) const; |
200 | bool SelectVOP3Mods(SDValue In, SDValue &Src, SDValue &SrcMods) const; |
201 | bool SelectVOP3ModsNonCanonicalizing(SDValue In, SDValue &Src, |
202 | SDValue &SrcMods) const; |
203 | bool SelectVOP3BMods(SDValue In, SDValue &Src, SDValue &SrcMods) const; |
204 | bool SelectVOP3NoMods(SDValue In, SDValue &Src) const; |
205 | bool SelectVOP3Mods0(SDValue In, SDValue &Src, SDValue &SrcMods, |
206 | SDValue &Clamp, SDValue &Omod) const; |
207 | bool SelectVOP3BMods0(SDValue In, SDValue &Src, SDValue &SrcMods, |
208 | SDValue &Clamp, SDValue &Omod) const; |
209 | bool SelectVOP3NoMods0(SDValue In, SDValue &Src, SDValue &SrcMods, |
210 | SDValue &Clamp, SDValue &Omod) const; |
211 | |
212 | bool SelectVINTERPModsImpl(SDValue In, SDValue &Src, SDValue &SrcMods, |
213 | bool OpSel) const; |
214 | bool SelectVINTERPMods(SDValue In, SDValue &Src, SDValue &SrcMods) const; |
215 | bool SelectVINTERPModsHi(SDValue In, SDValue &Src, SDValue &SrcMods) const; |
216 | |
217 | bool SelectVOP3OMods(SDValue In, SDValue &Src, SDValue &Clamp, |
218 | SDValue &Omod) const; |
219 | |
220 | bool SelectVOP3PMods(SDValue In, SDValue &Src, SDValue &SrcMods, |
221 | bool IsDOT = false) const; |
222 | bool SelectVOP3PModsDOT(SDValue In, SDValue &Src, SDValue &SrcMods) const; |
223 | |
224 | bool SelectVOP3PModsNeg(SDValue In, SDValue &Src) const; |
225 | bool SelectWMMAOpSelVOP3PMods(SDValue In, SDValue &Src) const; |
226 | |
227 | bool SelectWMMAModsF32NegAbs(SDValue In, SDValue &Src, |
228 | SDValue &SrcMods) const; |
229 | bool SelectWMMAModsF16Neg(SDValue In, SDValue &Src, SDValue &SrcMods) const; |
230 | bool SelectWMMAModsF16NegAbs(SDValue In, SDValue &Src, |
231 | SDValue &SrcMods) const; |
232 | bool SelectWMMAVISrc(SDValue In, SDValue &Src) const; |
233 | |
234 | bool SelectSWMMACIndex8(SDValue In, SDValue &Src, SDValue &IndexKey) const; |
235 | bool SelectSWMMACIndex16(SDValue In, SDValue &Src, SDValue &IndexKey) const; |
236 | |
237 | bool SelectVOP3OpSel(SDValue In, SDValue &Src, SDValue &SrcMods) const; |
238 | |
239 | bool SelectVOP3OpSelMods(SDValue In, SDValue &Src, SDValue &SrcMods) const; |
240 | bool SelectVOP3PMadMixModsImpl(SDValue In, SDValue &Src, |
241 | unsigned &Mods) const; |
242 | bool SelectVOP3PMadMixModsExt(SDValue In, SDValue &Src, |
243 | SDValue &SrcMods) const; |
244 | bool SelectVOP3PMadMixMods(SDValue In, SDValue &Src, SDValue &SrcMods) const; |
245 | |
246 | bool SelectBITOP3(SDValue In, SDValue &Src0, SDValue &Src1, SDValue &Src2, |
247 | SDValue &Tbl) const; |
248 | |
249 | SDValue getHi16Elt(SDValue In) const; |
250 | |
251 | SDValue getMaterializedScalarImm32(int64_t Val, const SDLoc &DL) const; |
252 | |
253 | void SelectADD_SUB_I64(SDNode *N); |
254 | void SelectAddcSubb(SDNode *N); |
255 | void SelectUADDO_USUBO(SDNode *N); |
256 | void SelectDIV_SCALE(SDNode *N); |
257 | void SelectMAD_64_32(SDNode *N); |
258 | void SelectMUL_LOHI(SDNode *N); |
259 | void SelectFMA_W_CHAIN(SDNode *N); |
260 | void SelectFMUL_W_CHAIN(SDNode *N); |
261 | SDNode *getBFE32(bool IsSigned, const SDLoc &DL, SDValue Val, uint32_t Offset, |
262 | uint32_t Width); |
263 | void SelectS_BFEFromShifts(SDNode *N); |
264 | void SelectS_BFE(SDNode *N); |
265 | bool isCBranchSCC(const SDNode *N) const; |
266 | void SelectBRCOND(SDNode *N); |
267 | void SelectFMAD_FMA(SDNode *N); |
268 | void SelectFP_EXTEND(SDNode *N); |
269 | void SelectDSAppendConsume(SDNode *N, unsigned IntrID); |
270 | void SelectDSBvhStackIntrinsic(SDNode *N, unsigned IntrID); |
271 | void SelectDS_GWS(SDNode *N, unsigned IntrID); |
272 | void SelectInterpP1F16(SDNode *N); |
273 | void SelectINTRINSIC_W_CHAIN(SDNode *N); |
274 | void SelectINTRINSIC_WO_CHAIN(SDNode *N); |
275 | void SelectINTRINSIC_VOID(SDNode *N); |
276 | void SelectWAVE_ADDRESS(SDNode *N); |
277 | void SelectSTACKRESTORE(SDNode *N); |
278 | |
279 | protected: |
280 | // Include the pieces autogenerated from the target description. |
281 | #include "AMDGPUGenDAGISel.inc" |
282 | }; |
283 | |
284 | class AMDGPUISelDAGToDAGPass : public SelectionDAGISelPass { |
285 | public: |
286 | AMDGPUISelDAGToDAGPass(TargetMachine &TM); |
287 | |
288 | PreservedAnalyses run(MachineFunction &MF, |
289 | MachineFunctionAnalysisManager &MFAM); |
290 | }; |
291 | |
292 | class AMDGPUDAGToDAGISelLegacy : public SelectionDAGISelLegacy { |
293 | public: |
294 | static char ID; |
295 | |
296 | AMDGPUDAGToDAGISelLegacy(TargetMachine &TM, CodeGenOptLevel OptLevel); |
297 | |
298 | bool runOnMachineFunction(MachineFunction &MF) override; |
299 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
300 | StringRef getPassName() const override; |
301 | }; |
302 | |
303 | } // namespace llvm |
304 | |
305 | #endif // LLVM_LIB_TARGET_AMDGPU_AMDGPUISELDAGTODAG_H |
306 | |