xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.h (revision 0ad011ececb978e22a9bff2acf76633b094f1ff6)
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/CodeGen/SelectionDAGISel.h"
21 #include "llvm/Target/TargetMachine.h"
22 
23 using namespace llvm;
24 
25 namespace {
26 
27 static inline bool isNullConstantOrUndef(SDValue V) {
28   return V.isUndef() || isNullConstant(V);
29 }
30 
31 static inline bool getConstantValue(SDValue N, uint32_t &Out) {
32   // This is only used for packed vectors, where using 0 for undef should
33   // always be good.
34   if (N.isUndef()) {
35     Out = 0;
36     return true;
37   }
38 
39   if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
40     Out = C->getAPIntValue().getSExtValue();
41     return true;
42   }
43 
44   if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N)) {
45     Out = C->getValueAPF().bitcastToAPInt().getSExtValue();
46     return true;
47   }
48 
49   return false;
50 }
51 
52 // TODO: Handle undef as zero
53 static inline SDNode *packConstantV2I16(const SDNode *N, SelectionDAG &DAG,
54                                         bool Negate = false) {
55   assert(N->getOpcode() == ISD::BUILD_VECTOR && N->getNumOperands() == 2);
56   uint32_t LHSVal, RHSVal;
57   if (getConstantValue(N->getOperand(0), LHSVal) &&
58       getConstantValue(N->getOperand(1), RHSVal)) {
59     SDLoc SL(N);
60     uint32_t K = Negate ? (-LHSVal & 0xffff) | (-RHSVal << 16)
61                         : (LHSVal & 0xffff) | (RHSVal << 16);
62     return DAG.getMachineNode(AMDGPU::S_MOV_B32, SL, N->getValueType(0),
63                               DAG.getTargetConstant(K, SL, MVT::i32));
64   }
65 
66   return nullptr;
67 }
68 
69 static inline SDNode *packNegConstantV2I16(const SDNode *N, SelectionDAG &DAG) {
70   return packConstantV2I16(N, DAG, true);
71 }
72 } // namespace
73 
74 /// AMDGPU specific code to select AMDGPU machine instructions for
75 /// SelectionDAG operations.
76 class AMDGPUDAGToDAGISel : public SelectionDAGISel {
77   // Subtarget - Keep a pointer to the AMDGPU Subtarget around so that we can
78   // make the right decision when generating code for different targets.
79   const GCNSubtarget *Subtarget;
80 
81   // Default FP mode for the current function.
82   SIModeRegisterDefaults Mode;
83 
84   bool EnableLateStructurizeCFG;
85 
86   // Instructions that will be lowered with a final instruction that zeros the
87   // high result bits.
88   bool fp16SrcZerosHighBits(unsigned Opc) const;
89 
90 public:
91   static char ID;
92 
93   AMDGPUDAGToDAGISel() = delete;
94 
95   explicit AMDGPUDAGToDAGISel(TargetMachine &TM, CodeGenOpt::Level OptLevel);
96   ~AMDGPUDAGToDAGISel() override = default;
97 
98   void getAnalysisUsage(AnalysisUsage &AU) const override;
99 
100   bool matchLoadD16FromBuildVector(SDNode *N) const;
101 
102   bool runOnMachineFunction(MachineFunction &MF) override;
103   void PreprocessISelDAG() override;
104   void Select(SDNode *N) override;
105   StringRef getPassName() const override;
106   void PostprocessISelDAG() override;
107 
108 protected:
109   void SelectBuildVector(SDNode *N, unsigned RegClassID);
110 
111 private:
112   std::pair<SDValue, SDValue> foldFrameIndex(SDValue N) const;
113   bool isInlineImmediate(const SDNode *N, bool Negated = false) const;
114   bool isNegInlineImmediate(const SDNode *N) const {
115     return isInlineImmediate(N, true);
116   }
117 
118   bool isInlineImmediate16(int64_t Imm) const {
119     return AMDGPU::isInlinableLiteral16(Imm, Subtarget->hasInv2PiInlineImm());
120   }
121 
122   bool isInlineImmediate32(int64_t Imm) const {
123     return AMDGPU::isInlinableLiteral32(Imm, Subtarget->hasInv2PiInlineImm());
124   }
125 
126   bool isInlineImmediate64(int64_t Imm) const {
127     return AMDGPU::isInlinableLiteral64(Imm, Subtarget->hasInv2PiInlineImm());
128   }
129 
130   bool isInlineImmediate(const APFloat &Imm) const {
131     return Subtarget->getInstrInfo()->isInlineConstant(Imm);
132   }
133 
134   bool isVGPRImm(const SDNode *N) const;
135   bool isUniformLoad(const SDNode *N) const;
136   bool isUniformBr(const SDNode *N) const;
137 
138   // Returns true if ISD::AND SDNode `N`'s masking of the shift amount operand's
139   // `ShAmtBits` bits is unneeded.
140   bool isUnneededShiftMask(const SDNode *N, unsigned ShAmtBits) const;
141 
142   bool isBaseWithConstantOffset64(SDValue Addr, SDValue &LHS,
143                                   SDValue &RHS) const;
144 
145   MachineSDNode *buildSMovImm64(SDLoc &DL, uint64_t Val, EVT VT) const;
146 
147   SDNode *glueCopyToOp(SDNode *N, SDValue NewChain, SDValue Glue) const;
148   SDNode *glueCopyToM0(SDNode *N, SDValue Val) const;
149   SDNode *glueCopyToM0LDSInit(SDNode *N) const;
150 
151   const TargetRegisterClass *getOperandRegClass(SDNode *N, unsigned OpNo) const;
152   virtual bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base, SDValue &Offset);
153   virtual bool SelectADDRIndirect(SDValue Addr, SDValue &Base, SDValue &Offset);
154   bool isDSOffsetLegal(SDValue Base, unsigned Offset) const;
155   bool isDSOffset2Legal(SDValue Base, unsigned Offset0, unsigned Offset1,
156                         unsigned Size) const;
157   bool isFlatScratchBaseLegal(
158       SDValue Base, uint64_t FlatVariant = SIInstrFlags::FlatScratch) const;
159 
160   bool SelectDS1Addr1Offset(SDValue Ptr, SDValue &Base, SDValue &Offset) const;
161   bool SelectDS64Bit4ByteAligned(SDValue Ptr, SDValue &Base, SDValue &Offset0,
162                                  SDValue &Offset1) const;
163   bool SelectDS128Bit8ByteAligned(SDValue Ptr, SDValue &Base, SDValue &Offset0,
164                                   SDValue &Offset1) const;
165   bool SelectDSReadWrite2(SDValue Ptr, SDValue &Base, SDValue &Offset0,
166                           SDValue &Offset1, unsigned Size) const;
167   bool SelectMUBUF(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
168                    SDValue &SOffset, SDValue &Offset, SDValue &Offen,
169                    SDValue &Idxen, SDValue &Addr64) const;
170   bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
171                          SDValue &SOffset, SDValue &Offset) const;
172   bool SelectMUBUFScratchOffen(SDNode *Parent, SDValue Addr, SDValue &RSrc,
173                                SDValue &VAddr, SDValue &SOffset,
174                                SDValue &ImmOffset) const;
175   bool SelectMUBUFScratchOffset(SDNode *Parent, SDValue Addr, SDValue &SRsrc,
176                                 SDValue &Soffset, SDValue &Offset) const;
177 
178   bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset,
179                          SDValue &Offset) const;
180 
181   bool SelectFlatOffsetImpl(SDNode *N, SDValue Addr, SDValue &VAddr,
182                             SDValue &Offset, uint64_t FlatVariant) const;
183   bool SelectFlatOffset(SDNode *N, SDValue Addr, SDValue &VAddr,
184                         SDValue &Offset) const;
185   bool SelectGlobalOffset(SDNode *N, SDValue Addr, SDValue &VAddr,
186                           SDValue &Offset) const;
187   bool SelectScratchOffset(SDNode *N, SDValue Addr, SDValue &VAddr,
188                            SDValue &Offset) const;
189   bool SelectGlobalSAddr(SDNode *N, SDValue Addr, SDValue &SAddr,
190                          SDValue &VOffset, SDValue &Offset) const;
191   bool SelectScratchSAddr(SDNode *N, SDValue Addr, SDValue &SAddr,
192                           SDValue &Offset) const;
193   bool checkFlatScratchSVSSwizzleBug(SDValue VAddr, SDValue SAddr,
194                                      uint64_t ImmOffset) const;
195   bool SelectScratchSVAddr(SDNode *N, SDValue Addr, SDValue &VAddr,
196                            SDValue &SAddr, SDValue &Offset) const;
197 
198   bool SelectSMRDOffset(SDValue ByteOffsetNode, SDValue *SOffset,
199                         SDValue *Offset, bool Imm32Only = false,
200                         bool IsBuffer = false) const;
201   SDValue Expand32BitAddress(SDValue Addr) const;
202   bool SelectSMRDBaseOffset(SDValue Addr, SDValue &SBase, SDValue *SOffset,
203                             SDValue *Offset, bool Imm32Only = false,
204                             bool IsBuffer = false) const;
205   bool SelectSMRD(SDValue Addr, SDValue &SBase, SDValue *SOffset,
206                   SDValue *Offset, bool Imm32Only = false) const;
207   bool SelectSMRDImm(SDValue Addr, SDValue &SBase, SDValue &Offset) const;
208   bool SelectSMRDImm32(SDValue Addr, SDValue &SBase, SDValue &Offset) const;
209   bool SelectSMRDSgpr(SDValue Addr, SDValue &SBase, SDValue &SOffset) const;
210   bool SelectSMRDSgprImm(SDValue Addr, SDValue &SBase, SDValue &SOffset,
211                          SDValue &Offset) const;
212   bool SelectSMRDBufferImm(SDValue N, SDValue &Offset) const;
213   bool SelectSMRDBufferImm32(SDValue N, SDValue &Offset) const;
214   bool SelectSMRDBufferSgprImm(SDValue N, SDValue &SOffset,
215                                SDValue &Offset) const;
216   bool SelectMOVRELOffset(SDValue Index, SDValue &Base, SDValue &Offset) const;
217 
218   bool SelectVOP3ModsImpl(SDValue In, SDValue &Src, unsigned &SrcMods,
219                           bool IsCanonicalizing = true,
220                           bool AllowAbs = true) const;
221   bool SelectVOP3Mods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
222   bool SelectVOP3ModsNonCanonicalizing(SDValue In, SDValue &Src,
223                                        SDValue &SrcMods) const;
224   bool SelectVOP3BMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
225   bool SelectVOP3NoMods(SDValue In, SDValue &Src) const;
226   bool SelectVOP3Mods0(SDValue In, SDValue &Src, SDValue &SrcMods,
227                        SDValue &Clamp, SDValue &Omod) const;
228   bool SelectVOP3BMods0(SDValue In, SDValue &Src, SDValue &SrcMods,
229                         SDValue &Clamp, SDValue &Omod) const;
230   bool SelectVOP3NoMods0(SDValue In, SDValue &Src, SDValue &SrcMods,
231                          SDValue &Clamp, SDValue &Omod) const;
232 
233   bool SelectVINTERPModsImpl(SDValue In, SDValue &Src, SDValue &SrcMods,
234                              bool OpSel) const;
235   bool SelectVINTERPMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
236   bool SelectVINTERPModsHi(SDValue In, SDValue &Src, SDValue &SrcMods) const;
237 
238   bool SelectVOP3OMods(SDValue In, SDValue &Src, SDValue &Clamp,
239                        SDValue &Omod) const;
240 
241   bool SelectVOP3PMods(SDValue In, SDValue &Src, SDValue &SrcMods,
242                        bool IsDOT = false) const;
243   bool SelectVOP3PModsDOT(SDValue In, SDValue &Src, SDValue &SrcMods) const;
244 
245   bool SelectDotIUVOP3PMods(SDValue In, SDValue &Src) const;
246   bool SelectWMMAOpSelVOP3PMods(SDValue In, SDValue &Src) const;
247 
248   bool SelectVOP3OpSel(SDValue In, SDValue &Src, SDValue &SrcMods) const;
249 
250   bool SelectVOP3OpSelMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
251   bool SelectVOP3PMadMixModsImpl(SDValue In, SDValue &Src,
252                                  unsigned &Mods) const;
253   bool SelectVOP3PMadMixModsExt(SDValue In, SDValue &Src,
254                                 SDValue &SrcMods) const;
255   bool SelectVOP3PMadMixMods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
256 
257   SDValue getHi16Elt(SDValue In) const;
258 
259   SDValue getMaterializedScalarImm32(int64_t Val, const SDLoc &DL) const;
260 
261   void SelectADD_SUB_I64(SDNode *N);
262   void SelectAddcSubb(SDNode *N);
263   void SelectUADDO_USUBO(SDNode *N);
264   void SelectDIV_SCALE(SDNode *N);
265   void SelectMAD_64_32(SDNode *N);
266   void SelectMUL_LOHI(SDNode *N);
267   void SelectFMA_W_CHAIN(SDNode *N);
268   void SelectFMUL_W_CHAIN(SDNode *N);
269   SDNode *getBFE32(bool IsSigned, const SDLoc &DL, SDValue Val, uint32_t Offset,
270                    uint32_t Width);
271   void SelectS_BFEFromShifts(SDNode *N);
272   void SelectS_BFE(SDNode *N);
273   bool isCBranchSCC(const SDNode *N) const;
274   void SelectBRCOND(SDNode *N);
275   void SelectFMAD_FMA(SDNode *N);
276   void SelectDSAppendConsume(SDNode *N, unsigned IntrID);
277   void SelectDSBvhStackIntrinsic(SDNode *N);
278   void SelectDS_GWS(SDNode *N, unsigned IntrID);
279   void SelectInterpP1F16(SDNode *N);
280   void SelectINTRINSIC_W_CHAIN(SDNode *N);
281   void SelectINTRINSIC_WO_CHAIN(SDNode *N);
282   void SelectINTRINSIC_VOID(SDNode *N);
283 
284 protected:
285   // Include the pieces autogenerated from the target description.
286 #include "AMDGPUGenDAGISel.inc"
287 };
288 
289 #endif // LLVM_LIB_TARGET_AMDGPU_AMDGPUISELDAGTODAG_H
290