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 "AMDGPUSelectionDAGInfo.h"
18#include "GCNSubtarget.h"
19#include "SIMachineFunctionInfo.h"
20#include "SIModeRegisterDefaults.h"
21#include "llvm/Analysis/ValueTracking.h"
22#include "llvm/CodeGen/SelectionDAGISel.h"
23#include "llvm/Support/AMDGPUAddrSpace.h"
24#include "llvm/Target/TargetMachine.h"
25
26namespace llvm {
27
28static inline bool getConstantValue(SDValue N, uint32_t &Out) {
29 // This is only used for packed vectors, where using 0 for undef should
30 // always be good.
31 if (N.isUndef()) {
32 Out = 0;
33 return true;
34 }
35
36 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: N)) {
37 Out = C->getAPIntValue().getSExtValue();
38 return true;
39 }
40
41 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val&: N)) {
42 Out = C->getValueAPF().bitcastToAPInt().getSExtValue();
43 return true;
44 }
45
46 return false;
47}
48
49/// AMDGPU specific code to select AMDGPU machine instructions for
50/// SelectionDAG operations.
51class AMDGPUDAGToDAGISel : public SelectionDAGISel {
52 // Subtarget - Keep a pointer to the AMDGPU Subtarget around so that we can
53 // make the right decision when generating code for different targets.
54 const GCNSubtarget *Subtarget;
55
56 // Default FP mode for the current function.
57 SIModeRegisterDefaults Mode;
58
59 // Instructions that will be lowered with a final instruction that zeros the
60 // high result bits.
61 bool fp16SrcZerosHighBits(unsigned Opc) const;
62
63public:
64 AMDGPUDAGToDAGISel() = delete;
65
66 explicit AMDGPUDAGToDAGISel(TargetMachine &TM, CodeGenOptLevel OptLevel);
67
68 bool runOnMachineFunction(MachineFunction &MF) override;
69 bool matchLoadD16FromBuildVector(SDNode *N) const;
70 void PreprocessISelDAG() override;
71 void Select(SDNode *N) override;
72 void PostprocessISelDAG() override;
73
74protected:
75 void SelectBuildVector(SDNode *N, unsigned RegClassID);
76 void SelectVectorShuffle(SDNode *N);
77
78private:
79 std::pair<SDValue, SDValue> foldFrameIndex(SDValue N) const;
80
81 bool isInlineImmediate(const SDNode *N) const;
82
83 bool isInlineImmediate(const APInt &Imm) const {
84 return Subtarget->getInstrInfo()->isInlineConstant(Imm);
85 }
86
87 bool isInlineImmediate(const APFloat &Imm) const {
88 return Subtarget->getInstrInfo()->isInlineConstant(Imm);
89 }
90
91 bool isVGPRImm(const SDNode *N) const;
92 bool isUniformLoad(const SDNode *N) const;
93 bool isUniformBr(const SDNode *N) const;
94
95 // Returns true if ISD::AND SDNode `N`'s masking of the shift amount operand's
96 // `ShAmtBits` bits is unneeded.
97 bool isUnneededShiftMask(const SDNode *N, unsigned ShAmtBits) const;
98
99 bool isBaseWithConstantOffset64(SDValue Addr, SDValue &LHS,
100 SDValue &RHS) const;
101
102 MachineSDNode *buildSMovImm64(SDLoc &DL, uint64_t Val, EVT VT) const;
103
104 SDNode *packConstantV2I16(const SDNode *N, SelectionDAG &DAG) const;
105
106 SDNode *glueCopyToOp(SDNode *N, SDValue NewChain, SDValue Glue) const;
107 SDNode *glueCopyToM0(SDNode *N, SDValue Val) const;
108 SDNode *glueCopyToM0LDSInit(SDNode *N) const;
109
110 const TargetRegisterClass *getOperandRegClass(SDNode *N, unsigned OpNo) const;
111 virtual bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base, SDValue &Offset);
112 virtual bool SelectADDRIndirect(SDValue Addr, SDValue &Base, SDValue &Offset);
113 bool isDSOffsetLegal(SDValue Base, unsigned Offset) const;
114 bool isDSOffset2Legal(SDValue Base, unsigned Offset0, unsigned Offset1,
115 unsigned Size) const;
116
117 bool isFlatScratchBaseLegal(SDValue Addr) const;
118 bool isFlatScratchBaseLegalSV(SDValue Addr) const;
119 bool isFlatScratchBaseLegalSVImm(SDValue Addr) const;
120 bool isSOffsetLegalWithImmOffset(SDValue *SOffset, bool Imm32Only,
121 bool IsBuffer, int64_t ImmOffset = 0) const;
122
123 bool SelectDS1Addr1Offset(SDValue Ptr, SDValue &Base, SDValue &Offset) const;
124 bool SelectDS64Bit4ByteAligned(SDValue Ptr, SDValue &Base, SDValue &Offset0,
125 SDValue &Offset1) const;
126 bool SelectDS128Bit8ByteAligned(SDValue Ptr, SDValue &Base, SDValue &Offset0,
127 SDValue &Offset1) const;
128 bool SelectDSReadWrite2(SDValue Ptr, SDValue &Base, SDValue &Offset0,
129 SDValue &Offset1, unsigned Size) const;
130 bool SelectMUBUF(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
131 SDValue &SOffset, SDValue &Offset, SDValue &Offen,
132 SDValue &Idxen, SDValue &Addr64) const;
133 bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
134 SDValue &SOffset, SDValue &Offset) const;
135 bool SelectMUBUFScratchOffen(SDNode *Parent, SDValue Addr, SDValue &RSrc,
136 SDValue &VAddr, SDValue &SOffset,
137 SDValue &ImmOffset) const;
138 bool SelectMUBUFScratchOffset(SDNode *Parent, SDValue Addr, SDValue &SRsrc,
139 SDValue &Soffset, SDValue &Offset) const;
140
141 bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset,
142 SDValue &Offset) const;
143 bool SelectBUFSOffset(SDValue Addr, SDValue &SOffset) const;
144
145 bool SelectFlatOffsetImpl(SDNode *N, SDValue Addr, SDValue &VAddr,
146 SDValue &Offset, uint64_t FlatVariant) const;
147 bool SelectFlatOffset(SDNode *N, SDValue Addr, SDValue &VAddr,
148 SDValue &Offset) const;
149 bool SelectGlobalOffset(SDNode *N, SDValue Addr, SDValue &VAddr,
150 SDValue &Offset) const;
151 bool SelectScratchOffset(SDNode *N, SDValue Addr, SDValue &VAddr,
152 SDValue &Offset) const;
153 bool SelectGlobalSAddr(SDNode *N, SDValue Addr, SDValue &SAddr,
154 SDValue &VOffset, SDValue &Offset, bool &ScaleOffset,
155 bool NeedIOffset = true) const;
156 bool SelectGlobalSAddr(SDNode *N, SDValue Addr, SDValue &SAddr,
157 SDValue &VOffset, SDValue &Offset,
158 SDValue &CPol) const;
159 bool SelectGlobalSAddrCPol(SDNode *N, SDValue Addr, SDValue &SAddr,
160 SDValue &VOffset, SDValue &Offset,
161 SDValue &CPol) const;
162 bool SelectGlobalSAddrCPolM0(SDNode *N, SDValue Addr, SDValue &SAddr,
163 SDValue &VOffset, SDValue &Offset,
164 SDValue &CPol) const;
165 bool SelectGlobalSAddrGLC(SDNode *N, SDValue Addr, SDValue &SAddr,
166 SDValue &VOffset, SDValue &Offset,
167 SDValue &CPol) const;
168 bool SelectGlobalSAddrNoIOffset(SDNode *N, SDValue Addr, SDValue &SAddr,
169 SDValue &VOffset, SDValue &CPol) const;
170 bool SelectGlobalSAddrNoIOffsetM0(SDNode *N, SDValue Addr, SDValue &SAddr,
171 SDValue &VOffset, SDValue &CPol) const;
172 bool SelectScratchSAddr(SDNode *N, SDValue Addr, SDValue &SAddr,
173 SDValue &Offset) const;
174 bool checkFlatScratchSVSSwizzleBug(SDValue VAddr, SDValue SAddr,
175 uint64_t ImmOffset) const;
176 bool SelectScratchSVAddr(SDNode *N, SDValue Addr, SDValue &VAddr,
177 SDValue &SAddr, SDValue &Offset,
178 SDValue &CPol) const;
179
180 bool SelectSMRDOffset(SDNode *N, SDValue ByteOffsetNode, SDValue *SOffset,
181 SDValue *Offset, bool Imm32Only = false,
182 bool IsBuffer = false, bool HasSOffset = false,
183 int64_t ImmOffset = 0,
184 bool *ScaleOffset = nullptr) const;
185 SDValue Expand32BitAddress(SDValue Addr) const;
186 bool SelectSMRDBaseOffset(SDNode *N, SDValue Addr, SDValue &SBase,
187 SDValue *SOffset, SDValue *Offset,
188 bool Imm32Only = false, bool IsBuffer = false,
189 bool HasSOffset = false, int64_t ImmOffset = 0,
190 bool *ScaleOffset = nullptr) const;
191 bool SelectSMRD(SDNode *N, SDValue Addr, SDValue &SBase, SDValue *SOffset,
192 SDValue *Offset, bool Imm32Only = false,
193 bool *ScaleOffset = nullptr) const;
194 bool SelectSMRDImm(SDValue Addr, SDValue &SBase, SDValue &Offset) const;
195 bool SelectSMRDImm32(SDValue Addr, SDValue &SBase, SDValue &Offset) const;
196 bool SelectScaleOffset(SDNode *N, SDValue &Offset, bool IsSigned) const;
197 bool SelectSMRDSgpr(SDNode *N, SDValue Addr, SDValue &SBase, SDValue &SOffset,
198 SDValue &CPol) const;
199 bool SelectSMRDSgprImm(SDNode *N, SDValue Addr, SDValue &SBase,
200 SDValue &SOffset, SDValue &Offset,
201 SDValue &CPol) const;
202 bool SelectSMRDBufferImm(SDValue N, SDValue &Offset) const;
203 bool SelectSMRDBufferImm32(SDValue N, SDValue &Offset) const;
204 bool SelectSMRDBufferSgprImm(SDValue N, SDValue &SOffset,
205 SDValue &Offset) const;
206 bool SelectSMRDPrefetchImm(SDValue Addr, SDValue &SBase,
207 SDValue &Offset) const;
208 bool SelectMOVRELOffset(SDValue Index, SDValue &Base, SDValue &Offset) const;
209
210 bool SelectVOP3ModsImpl(SDValue In, SDValue &Src, unsigned &SrcMods,
211 bool IsCanonicalizing = true,
212 bool AllowAbs = true) const;
213 bool SelectVOP3Mods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
214 bool SelectVOP3ModsNonCanonicalizing(SDValue In, SDValue &Src,
215 SDValue &SrcMods) const;
216 bool SelectVOP3BMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
217 bool SelectVOP3NoMods(SDValue In, SDValue &Src) const;
218 bool SelectVOP3Mods0(SDValue In, SDValue &Src, SDValue &SrcMods,
219 SDValue &Clamp, SDValue &Omod) const;
220 bool SelectVOP3BMods0(SDValue In, SDValue &Src, SDValue &SrcMods,
221 SDValue &Clamp, SDValue &Omod) const;
222 bool SelectVOP3NoMods0(SDValue In, SDValue &Src, SDValue &SrcMods,
223 SDValue &Clamp, SDValue &Omod) const;
224
225 bool SelectVINTERPModsImpl(SDValue In, SDValue &Src, SDValue &SrcMods,
226 bool OpSel) const;
227 bool SelectVINTERPMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
228 bool SelectVINTERPModsHi(SDValue In, SDValue &Src, SDValue &SrcMods) const;
229
230 bool SelectVOP3OMods(SDValue In, SDValue &Src, SDValue &Clamp,
231 SDValue &Omod) const;
232
233 bool SelectVOP3PMods(SDValue In, SDValue &Src, SDValue &SrcMods,
234 bool IsDOT = false) const;
235 bool SelectVOP3PModsDOT(SDValue In, SDValue &Src, SDValue &SrcMods) const;
236
237 bool SelectWMMAOpSelVOP3PMods(SDValue In, SDValue &Src) const;
238
239 bool SelectWMMAModsF32NegAbs(SDValue In, SDValue &Src,
240 SDValue &SrcMods) const;
241 bool SelectWMMAModsF16Neg(SDValue In, SDValue &Src, SDValue &SrcMods) const;
242 bool SelectWMMAModsF16NegAbs(SDValue In, SDValue &Src,
243 SDValue &SrcMods) const;
244 bool SelectWMMAVISrc(SDValue In, SDValue &Src) const;
245
246 bool SelectSWMMACIndex8(SDValue In, SDValue &Src, SDValue &IndexKey) const;
247 bool SelectSWMMACIndex16(SDValue In, SDValue &Src, SDValue &IndexKey) const;
248 bool SelectSWMMACIndex32(SDValue In, SDValue &Src, SDValue &IndexKey) const;
249
250 bool SelectVOP3OpSel(SDValue In, SDValue &Src, SDValue &SrcMods) const;
251
252 bool SelectVOP3OpSelMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
253 bool SelectVOP3PMadMixModsImpl(SDValue In, SDValue &Src, unsigned &Mods,
254 MVT VT) const;
255 bool SelectVOP3PMadMixModsExt(SDValue In, SDValue &Src,
256 SDValue &SrcMods) const;
257 bool SelectVOP3PMadMixMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
258 bool SelectVOP3PMadMixBF16ModsExt(SDValue In, SDValue &Src,
259 SDValue &SrcMods) const;
260 bool SelectVOP3PMadMixBF16Mods(SDValue In, SDValue &Src,
261 SDValue &SrcMods) const;
262
263 bool SelectBITOP3(SDValue In, SDValue &Src0, SDValue &Src1, SDValue &Src2,
264 SDValue &Tbl) const;
265
266 SDValue getHi16Elt(SDValue In) const;
267
268 SDValue getMaterializedScalarImm32(int64_t Val, const SDLoc &DL) const;
269
270 void SelectADD_SUB_I64(SDNode *N);
271 void SelectAddcSubb(SDNode *N);
272 void SelectUADDO_USUBO(SDNode *N);
273 void SelectDIV_SCALE(SDNode *N);
274 void SelectMAD_64_32(SDNode *N);
275 void SelectMUL_LOHI(SDNode *N);
276 void SelectFMA_W_CHAIN(SDNode *N);
277 void SelectFMUL_W_CHAIN(SDNode *N);
278 SDNode *getBFE32(bool IsSigned, const SDLoc &DL, SDValue Val, uint32_t Offset,
279 uint32_t Width);
280 void SelectS_BFEFromShifts(SDNode *N);
281 void SelectS_BFE(SDNode *N);
282 bool isCBranchSCC(const SDNode *N) const;
283 void SelectBRCOND(SDNode *N);
284 void SelectFMAD_FMA(SDNode *N);
285 void SelectFP_EXTEND(SDNode *N);
286 void SelectDSAppendConsume(SDNode *N, unsigned IntrID);
287 void SelectDSBvhStackIntrinsic(SDNode *N, unsigned IntrID);
288 void SelectDS_GWS(SDNode *N, unsigned IntrID);
289 void SelectInterpP1F16(SDNode *N);
290 void SelectINTRINSIC_W_CHAIN(SDNode *N);
291 void SelectINTRINSIC_WO_CHAIN(SDNode *N);
292 void SelectINTRINSIC_VOID(SDNode *N);
293 void SelectWAVE_ADDRESS(SDNode *N);
294 void SelectSTACKRESTORE(SDNode *N);
295
296protected:
297 // Include the pieces autogenerated from the target description.
298#include "AMDGPUGenDAGISel.inc"
299};
300
301class AMDGPUISelDAGToDAGPass : public SelectionDAGISelPass {
302public:
303 AMDGPUISelDAGToDAGPass(TargetMachine &TM);
304
305 PreservedAnalyses run(MachineFunction &MF,
306 MachineFunctionAnalysisManager &MFAM);
307};
308
309class AMDGPUDAGToDAGISelLegacy : public SelectionDAGISelLegacy {
310public:
311 static char ID;
312
313 AMDGPUDAGToDAGISelLegacy(TargetMachine &TM, CodeGenOptLevel OptLevel);
314
315 bool runOnMachineFunction(MachineFunction &MF) override;
316 void getAnalysisUsage(AnalysisUsage &AU) const override;
317 StringRef getPassName() const override;
318};
319
320} // namespace llvm
321
322#endif // LLVM_LIB_TARGET_AMDGPU_AMDGPUISELDAGTODAG_H
323