xref: /llvm-project/llvm/lib/Target/AMDGPU/R600ISelLowering.cpp (revision 662b9fa02ca217f91e683f456d70fccc7746d31c)
1 //===-- R600ISelLowering.cpp - R600 DAG Lowering Implementation -----------===//
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 /// Custom DAG lowering for R600
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "R600ISelLowering.h"
15 #include "AMDGPU.h"
16 #include "MCTargetDesc/R600MCTargetDesc.h"
17 #include "R600Defines.h"
18 #include "R600InstrInfo.h"
19 #include "R600MachineFunctionInfo.h"
20 #include "R600Subtarget.h"
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/IR/IntrinsicsAMDGPU.h"
23 #include "llvm/IR/IntrinsicsR600.h"
24 
25 using namespace llvm;
26 
27 #include "R600GenCallingConv.inc"
28 
29 R600TargetLowering::R600TargetLowering(const TargetMachine &TM,
30                                        const R600Subtarget &STI)
31     : AMDGPUTargetLowering(TM, STI), Subtarget(&STI), Gen(STI.getGeneration()) {
32   addRegisterClass(MVT::f32, &R600::R600_Reg32RegClass);
33   addRegisterClass(MVT::i32, &R600::R600_Reg32RegClass);
34   addRegisterClass(MVT::v2f32, &R600::R600_Reg64RegClass);
35   addRegisterClass(MVT::v2i32, &R600::R600_Reg64RegClass);
36   addRegisterClass(MVT::v4f32, &R600::R600_Reg128RegClass);
37   addRegisterClass(MVT::v4i32, &R600::R600_Reg128RegClass);
38 
39   setBooleanContents(ZeroOrNegativeOneBooleanContent);
40   setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
41 
42   computeRegisterProperties(Subtarget->getRegisterInfo());
43 
44   // Legalize loads and stores to the private address space.
45   setOperationAction(ISD::LOAD, MVT::i32, Custom);
46   setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
47   setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
48 
49   // EXTLOAD should be the same as ZEXTLOAD. It is legal for some address
50   // spaces, so it is custom lowered to handle those where it isn't.
51   for (MVT VT : MVT::integer_valuetypes()) {
52     setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
53     setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Custom);
54     setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i16, Custom);
55 
56     setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
57     setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i8, Custom);
58     setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i16, Custom);
59 
60     setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
61     setLoadExtAction(ISD::EXTLOAD, VT, MVT::i8, Custom);
62     setLoadExtAction(ISD::EXTLOAD, VT, MVT::i16, Custom);
63   }
64 
65   // Workaround for LegalizeDAG asserting on expansion of i1 vector loads.
66   setLoadExtAction(ISD::EXTLOAD, MVT::v2i32, MVT::v2i1, Expand);
67   setLoadExtAction(ISD::SEXTLOAD, MVT::v2i32, MVT::v2i1, Expand);
68   setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i32, MVT::v2i1, Expand);
69 
70   setLoadExtAction(ISD::EXTLOAD, MVT::v4i32, MVT::v4i1, Expand);
71   setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i1, Expand);
72   setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i1, Expand);
73 
74   setOperationAction(ISD::STORE, MVT::i8, Custom);
75   setOperationAction(ISD::STORE, MVT::i32, Custom);
76   setOperationAction(ISD::STORE, MVT::v2i32, Custom);
77   setOperationAction(ISD::STORE, MVT::v4i32, Custom);
78 
79   setTruncStoreAction(MVT::i32, MVT::i8, Custom);
80   setTruncStoreAction(MVT::i32, MVT::i16, Custom);
81   // We need to include these since trunc STORES to PRIVATE need
82   // special handling to accommodate RMW
83   setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
84   setTruncStoreAction(MVT::v4i32, MVT::v4i16, Custom);
85   setTruncStoreAction(MVT::v8i32, MVT::v8i16, Custom);
86   setTruncStoreAction(MVT::v16i32, MVT::v16i16, Custom);
87   setTruncStoreAction(MVT::v32i32, MVT::v32i16, Custom);
88   setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
89   setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
90   setTruncStoreAction(MVT::v8i32, MVT::v8i8, Custom);
91   setTruncStoreAction(MVT::v16i32, MVT::v16i8, Custom);
92   setTruncStoreAction(MVT::v32i32, MVT::v32i8, Custom);
93 
94   // Workaround for LegalizeDAG asserting on expansion of i1 vector stores.
95   setTruncStoreAction(MVT::v2i32, MVT::v2i1, Expand);
96   setTruncStoreAction(MVT::v4i32, MVT::v4i1, Expand);
97 
98   // Set condition code actions
99   setCondCodeAction(ISD::SETO,   MVT::f32, Expand);
100   setCondCodeAction(ISD::SETUO,  MVT::f32, Expand);
101   setCondCodeAction(ISD::SETLT,  MVT::f32, Expand);
102   setCondCodeAction(ISD::SETLE,  MVT::f32, Expand);
103   setCondCodeAction(ISD::SETOLT, MVT::f32, Expand);
104   setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
105   setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
106   setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
107   setCondCodeAction(ISD::SETUGE, MVT::f32, Expand);
108   setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
109   setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
110   setCondCodeAction(ISD::SETULE, MVT::f32, Expand);
111 
112   setCondCodeAction(ISD::SETLE, MVT::i32, Expand);
113   setCondCodeAction(ISD::SETLT, MVT::i32, Expand);
114   setCondCodeAction(ISD::SETULE, MVT::i32, Expand);
115   setCondCodeAction(ISD::SETULT, MVT::i32, Expand);
116 
117   setOperationAction(ISD::FCOS, MVT::f32, Custom);
118   setOperationAction(ISD::FSIN, MVT::f32, Custom);
119 
120   setOperationAction(ISD::SETCC, MVT::v4i32, Expand);
121   setOperationAction(ISD::SETCC, MVT::v2i32, Expand);
122 
123   setOperationAction(ISD::BR_CC, MVT::i32, Expand);
124   setOperationAction(ISD::BR_CC, MVT::f32, Expand);
125   setOperationAction(ISD::BRCOND, MVT::Other, Custom);
126 
127   setOperationAction(ISD::FSUB, MVT::f32, Expand);
128 
129   setOperationAction(ISD::FCEIL, MVT::f64, Custom);
130   setOperationAction(ISD::FTRUNC, MVT::f64, Custom);
131   setOperationAction(ISD::FRINT, MVT::f64, Custom);
132   setOperationAction(ISD::FFLOOR, MVT::f64, Custom);
133 
134   setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
135   setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
136 
137   setOperationAction(ISD::SETCC, MVT::i32, Expand);
138   setOperationAction(ISD::SETCC, MVT::f32, Expand);
139   setOperationAction(ISD::FP_TO_UINT, MVT::i1, Custom);
140   setOperationAction(ISD::FP_TO_SINT, MVT::i1, Custom);
141   setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
142   setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
143 
144   setOperationAction(ISD::SELECT, MVT::i32, Expand);
145   setOperationAction(ISD::SELECT, MVT::f32, Expand);
146   setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
147   setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
148 
149   // ADD, SUB overflow.
150   // TODO: turn these into Legal?
151   if (Subtarget->hasCARRY())
152     setOperationAction(ISD::UADDO, MVT::i32, Custom);
153 
154   if (Subtarget->hasBORROW())
155     setOperationAction(ISD::USUBO, MVT::i32, Custom);
156 
157   // Expand sign extension of vectors
158   if (!Subtarget->hasBFE())
159     setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
160 
161   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Expand);
162   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Expand);
163 
164   if (!Subtarget->hasBFE())
165     setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
166   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Expand);
167   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Expand);
168 
169   if (!Subtarget->hasBFE())
170     setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
171   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Expand);
172   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Expand);
173 
174   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
175   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i32, Expand);
176   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i32, Expand);
177 
178   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Expand);
179 
180   setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
181 
182   setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i32, Custom);
183   setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f32, Custom);
184   setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
185   setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
186 
187   setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i32, Custom);
188   setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f32, Custom);
189   setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
190   setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
191 
192   // We don't have 64-bit shifts. Thus we need either SHX i64 or SHX_PARTS i32
193   //  to be Legal/Custom in order to avoid library calls.
194   setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
195   setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
196   setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
197 
198   if (!Subtarget->hasFMA()) {
199     setOperationAction(ISD::FMA, MVT::f32, Expand);
200     setOperationAction(ISD::FMA, MVT::f64, Expand);
201   }
202 
203   // FIXME: May need no denormals check
204   setOperationAction(ISD::FMAD, MVT::f32, Legal);
205 
206   if (!Subtarget->hasBFI()) {
207     // fcopysign can be done in a single instruction with BFI.
208     setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
209     setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
210   }
211 
212   if (!Subtarget->hasBCNT(32))
213     setOperationAction(ISD::CTPOP, MVT::i32, Expand);
214 
215   if (!Subtarget->hasBCNT(64))
216     setOperationAction(ISD::CTPOP, MVT::i64, Expand);
217 
218   if (Subtarget->hasFFBH())
219     setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Custom);
220 
221   if (Subtarget->hasFFBL())
222     setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Custom);
223 
224   // FIXME: This was moved from AMDGPUTargetLowering, I'm not sure if we
225   // need it for R600.
226   if (Subtarget->hasBFE())
227     setHasExtractBitsInsn(true);
228 
229   setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
230 
231   const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
232   for (MVT VT : ScalarIntVTs) {
233     setOperationAction(ISD::ADDC, VT, Expand);
234     setOperationAction(ISD::SUBC, VT, Expand);
235     setOperationAction(ISD::ADDE, VT, Expand);
236     setOperationAction(ISD::SUBE, VT, Expand);
237   }
238 
239   // LLVM will expand these to atomic_cmp_swap(0)
240   // and atomic_swap, respectively.
241   setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
242   setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
243 
244   // We need to custom lower some of the intrinsics
245   setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
246   setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
247 
248   setSchedulingPreference(Sched::Source);
249 
250   setTargetDAGCombine({ISD::FP_ROUND, ISD::FP_TO_SINT, ISD::EXTRACT_VECTOR_ELT,
251                        ISD::SELECT_CC, ISD::INSERT_VECTOR_ELT, ISD::LOAD});
252 }
253 
254 static inline bool isEOP(MachineBasicBlock::iterator I) {
255   if (std::next(I) == I->getParent()->end())
256     return false;
257   return std::next(I)->getOpcode() == R600::RETURN;
258 }
259 
260 MachineBasicBlock *
261 R600TargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
262                                                 MachineBasicBlock *BB) const {
263   MachineFunction *MF = BB->getParent();
264   MachineRegisterInfo &MRI = MF->getRegInfo();
265   MachineBasicBlock::iterator I = MI;
266   const R600InstrInfo *TII = Subtarget->getInstrInfo();
267 
268   switch (MI.getOpcode()) {
269   default:
270     // Replace LDS_*_RET instruction that don't have any uses with the
271     // equivalent LDS_*_NORET instruction.
272     if (TII->isLDSRetInstr(MI.getOpcode())) {
273       int DstIdx = TII->getOperandIdx(MI.getOpcode(), R600::OpName::dst);
274       assert(DstIdx != -1);
275       MachineInstrBuilder NewMI;
276       // FIXME: getLDSNoRetOp method only handles LDS_1A1D LDS ops. Add
277       //        LDS_1A2D support and remove this special case.
278       if (!MRI.use_empty(MI.getOperand(DstIdx).getReg()) ||
279           MI.getOpcode() == R600::LDS_CMPST_RET)
280         return BB;
281 
282       NewMI = BuildMI(*BB, I, BB->findDebugLoc(I),
283                       TII->get(R600::getLDSNoRetOp(MI.getOpcode())));
284       for (const MachineOperand &MO : llvm::drop_begin(MI.operands()))
285         NewMI.add(MO);
286     } else {
287       return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
288     }
289     break;
290 
291   case R600::FABS_R600: {
292     MachineInstr *NewMI = TII->buildDefaultInstruction(
293         *BB, I, R600::MOV, MI.getOperand(0).getReg(),
294         MI.getOperand(1).getReg());
295     TII->addFlag(*NewMI, 0, MO_FLAG_ABS);
296     break;
297   }
298 
299   case R600::FNEG_R600: {
300     MachineInstr *NewMI = TII->buildDefaultInstruction(
301         *BB, I, R600::MOV, MI.getOperand(0).getReg(),
302         MI.getOperand(1).getReg());
303     TII->addFlag(*NewMI, 0, MO_FLAG_NEG);
304     break;
305   }
306 
307   case R600::MASK_WRITE: {
308     Register maskedRegister = MI.getOperand(0).getReg();
309     assert(maskedRegister.isVirtual());
310     MachineInstr * defInstr = MRI.getVRegDef(maskedRegister);
311     TII->addFlag(*defInstr, 0, MO_FLAG_MASK);
312     break;
313   }
314 
315   case R600::MOV_IMM_F32:
316     TII->buildMovImm(*BB, I, MI.getOperand(0).getReg(), MI.getOperand(1)
317                                                             .getFPImm()
318                                                             ->getValueAPF()
319                                                             .bitcastToAPInt()
320                                                             .getZExtValue());
321     break;
322 
323   case R600::MOV_IMM_I32:
324     TII->buildMovImm(*BB, I, MI.getOperand(0).getReg(),
325                      MI.getOperand(1).getImm());
326     break;
327 
328   case R600::MOV_IMM_GLOBAL_ADDR: {
329     //TODO: Perhaps combine this instruction with the next if possible
330     auto MIB = TII->buildDefaultInstruction(
331         *BB, MI, R600::MOV, MI.getOperand(0).getReg(), R600::ALU_LITERAL_X);
332     int Idx = TII->getOperandIdx(*MIB, R600::OpName::literal);
333     //TODO: Ugh this is rather ugly
334     const MachineOperand &MO = MI.getOperand(1);
335     MIB->getOperand(Idx).ChangeToGA(MO.getGlobal(), MO.getOffset(),
336                                     MO.getTargetFlags());
337     break;
338   }
339 
340   case R600::CONST_COPY: {
341     MachineInstr *NewMI = TII->buildDefaultInstruction(
342         *BB, MI, R600::MOV, MI.getOperand(0).getReg(), R600::ALU_CONST);
343     TII->setImmOperand(*NewMI, R600::OpName::src0_sel,
344                        MI.getOperand(1).getImm());
345     break;
346   }
347 
348   case R600::RAT_WRITE_CACHELESS_32_eg:
349   case R600::RAT_WRITE_CACHELESS_64_eg:
350   case R600::RAT_WRITE_CACHELESS_128_eg:
351     BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI.getOpcode()))
352         .add(MI.getOperand(0))
353         .add(MI.getOperand(1))
354         .addImm(isEOP(I)); // Set End of program bit
355     break;
356 
357   case R600::RAT_STORE_TYPED_eg:
358     BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI.getOpcode()))
359         .add(MI.getOperand(0))
360         .add(MI.getOperand(1))
361         .add(MI.getOperand(2))
362         .addImm(isEOP(I)); // Set End of program bit
363     break;
364 
365   case R600::BRANCH:
366     BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(R600::JUMP))
367         .add(MI.getOperand(0));
368     break;
369 
370   case R600::BRANCH_COND_f32: {
371     MachineInstr *NewMI =
372         BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(R600::PRED_X),
373                 R600::PREDICATE_BIT)
374             .add(MI.getOperand(1))
375             .addImm(R600::PRED_SETNE)
376             .addImm(0); // Flags
377     TII->addFlag(*NewMI, 0, MO_FLAG_PUSH);
378     BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(R600::JUMP_COND))
379         .add(MI.getOperand(0))
380         .addReg(R600::PREDICATE_BIT, RegState::Kill);
381     break;
382   }
383 
384   case R600::BRANCH_COND_i32: {
385     MachineInstr *NewMI =
386         BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(R600::PRED_X),
387                 R600::PREDICATE_BIT)
388             .add(MI.getOperand(1))
389             .addImm(R600::PRED_SETNE_INT)
390             .addImm(0); // Flags
391     TII->addFlag(*NewMI, 0, MO_FLAG_PUSH);
392     BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(R600::JUMP_COND))
393         .add(MI.getOperand(0))
394         .addReg(R600::PREDICATE_BIT, RegState::Kill);
395     break;
396   }
397 
398   case R600::EG_ExportSwz:
399   case R600::R600_ExportSwz: {
400     // Instruction is left unmodified if its not the last one of its type
401     bool isLastInstructionOfItsType = true;
402     unsigned InstExportType = MI.getOperand(1).getImm();
403     for (MachineBasicBlock::iterator NextExportInst = std::next(I),
404          EndBlock = BB->end(); NextExportInst != EndBlock;
405          NextExportInst = std::next(NextExportInst)) {
406       if (NextExportInst->getOpcode() == R600::EG_ExportSwz ||
407           NextExportInst->getOpcode() == R600::R600_ExportSwz) {
408         unsigned CurrentInstExportType = NextExportInst->getOperand(1)
409             .getImm();
410         if (CurrentInstExportType == InstExportType) {
411           isLastInstructionOfItsType = false;
412           break;
413         }
414       }
415     }
416     bool EOP = isEOP(I);
417     if (!EOP && !isLastInstructionOfItsType)
418       return BB;
419     unsigned CfInst = (MI.getOpcode() == R600::EG_ExportSwz) ? 84 : 40;
420     BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI.getOpcode()))
421         .add(MI.getOperand(0))
422         .add(MI.getOperand(1))
423         .add(MI.getOperand(2))
424         .add(MI.getOperand(3))
425         .add(MI.getOperand(4))
426         .add(MI.getOperand(5))
427         .add(MI.getOperand(6))
428         .addImm(CfInst)
429         .addImm(EOP);
430     break;
431   }
432   case R600::RETURN: {
433     return BB;
434   }
435   }
436 
437   MI.eraseFromParent();
438   return BB;
439 }
440 
441 //===----------------------------------------------------------------------===//
442 // Custom DAG Lowering Operations
443 //===----------------------------------------------------------------------===//
444 
445 SDValue R600TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
446   MachineFunction &MF = DAG.getMachineFunction();
447   R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
448   switch (Op.getOpcode()) {
449   default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
450   case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
451   case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
452   case ISD::SHL_PARTS:
453   case ISD::SRA_PARTS:
454   case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
455   case ISD::UADDO: return LowerUADDSUBO(Op, DAG, ISD::ADD, AMDGPUISD::CARRY);
456   case ISD::USUBO: return LowerUADDSUBO(Op, DAG, ISD::SUB, AMDGPUISD::BORROW);
457   case ISD::FCOS:
458   case ISD::FSIN: return LowerTrig(Op, DAG);
459   case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
460   case ISD::STORE: return LowerSTORE(Op, DAG);
461   case ISD::LOAD: {
462     SDValue Result = LowerLOAD(Op, DAG);
463     assert((!Result.getNode() ||
464             Result.getNode()->getNumValues() == 2) &&
465            "Load should return a value and a chain");
466     return Result;
467   }
468 
469   case ISD::BRCOND: return LowerBRCOND(Op, DAG);
470   case ISD::GlobalAddress: return LowerGlobalAddress(MFI, Op, DAG);
471   case ISD::FrameIndex: return lowerFrameIndex(Op, DAG);
472   case ISD::INTRINSIC_VOID: {
473     SDValue Chain = Op.getOperand(0);
474     unsigned IntrinsicID =
475                          cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
476     switch (IntrinsicID) {
477     case Intrinsic::r600_store_swizzle: {
478       SDLoc DL(Op);
479       const SDValue Args[8] = {
480         Chain,
481         Op.getOperand(2), // Export Value
482         Op.getOperand(3), // ArrayBase
483         Op.getOperand(4), // Type
484         DAG.getConstant(0, DL, MVT::i32), // SWZ_X
485         DAG.getConstant(1, DL, MVT::i32), // SWZ_Y
486         DAG.getConstant(2, DL, MVT::i32), // SWZ_Z
487         DAG.getConstant(3, DL, MVT::i32) // SWZ_W
488       };
489       return DAG.getNode(AMDGPUISD::R600_EXPORT, DL, Op.getValueType(), Args);
490     }
491 
492     // default for switch(IntrinsicID)
493     default: break;
494     }
495     // break out of case ISD::INTRINSIC_VOID in switch(Op.getOpcode())
496     break;
497   }
498   case ISD::INTRINSIC_WO_CHAIN: {
499     unsigned IntrinsicID =
500                          cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
501     EVT VT = Op.getValueType();
502     SDLoc DL(Op);
503     switch (IntrinsicID) {
504     case Intrinsic::r600_tex:
505     case Intrinsic::r600_texc: {
506       unsigned TextureOp;
507       switch (IntrinsicID) {
508       case Intrinsic::r600_tex:
509         TextureOp = 0;
510         break;
511       case Intrinsic::r600_texc:
512         TextureOp = 1;
513         break;
514       default:
515         llvm_unreachable("unhandled texture operation");
516       }
517 
518       SDValue TexArgs[19] = {
519         DAG.getConstant(TextureOp, DL, MVT::i32),
520         Op.getOperand(1),
521         DAG.getConstant(0, DL, MVT::i32),
522         DAG.getConstant(1, DL, MVT::i32),
523         DAG.getConstant(2, DL, MVT::i32),
524         DAG.getConstant(3, DL, MVT::i32),
525         Op.getOperand(2),
526         Op.getOperand(3),
527         Op.getOperand(4),
528         DAG.getConstant(0, DL, MVT::i32),
529         DAG.getConstant(1, DL, MVT::i32),
530         DAG.getConstant(2, DL, MVT::i32),
531         DAG.getConstant(3, DL, MVT::i32),
532         Op.getOperand(5),
533         Op.getOperand(6),
534         Op.getOperand(7),
535         Op.getOperand(8),
536         Op.getOperand(9),
537         Op.getOperand(10)
538       };
539       return DAG.getNode(AMDGPUISD::TEXTURE_FETCH, DL, MVT::v4f32, TexArgs);
540     }
541     case Intrinsic::r600_dot4: {
542       SDValue Args[8] = {
543       DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
544           DAG.getConstant(0, DL, MVT::i32)),
545       DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
546           DAG.getConstant(0, DL, MVT::i32)),
547       DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
548           DAG.getConstant(1, DL, MVT::i32)),
549       DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
550           DAG.getConstant(1, DL, MVT::i32)),
551       DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
552           DAG.getConstant(2, DL, MVT::i32)),
553       DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
554           DAG.getConstant(2, DL, MVT::i32)),
555       DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
556           DAG.getConstant(3, DL, MVT::i32)),
557       DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
558           DAG.getConstant(3, DL, MVT::i32))
559       };
560       return DAG.getNode(AMDGPUISD::DOT4, DL, MVT::f32, Args);
561     }
562 
563     case Intrinsic::r600_implicitarg_ptr: {
564       MVT PtrVT = getPointerTy(DAG.getDataLayout(), AMDGPUAS::PARAM_I_ADDRESS);
565       uint32_t ByteOffset = getImplicitParameterOffset(MF, FIRST_IMPLICIT);
566       return DAG.getConstant(ByteOffset, DL, PtrVT);
567     }
568     case Intrinsic::r600_read_ngroups_x:
569       return LowerImplicitParameter(DAG, VT, DL, 0);
570     case Intrinsic::r600_read_ngroups_y:
571       return LowerImplicitParameter(DAG, VT, DL, 1);
572     case Intrinsic::r600_read_ngroups_z:
573       return LowerImplicitParameter(DAG, VT, DL, 2);
574     case Intrinsic::r600_read_global_size_x:
575       return LowerImplicitParameter(DAG, VT, DL, 3);
576     case Intrinsic::r600_read_global_size_y:
577       return LowerImplicitParameter(DAG, VT, DL, 4);
578     case Intrinsic::r600_read_global_size_z:
579       return LowerImplicitParameter(DAG, VT, DL, 5);
580     case Intrinsic::r600_read_local_size_x:
581       return LowerImplicitParameter(DAG, VT, DL, 6);
582     case Intrinsic::r600_read_local_size_y:
583       return LowerImplicitParameter(DAG, VT, DL, 7);
584     case Intrinsic::r600_read_local_size_z:
585       return LowerImplicitParameter(DAG, VT, DL, 8);
586 
587     case Intrinsic::r600_read_tgid_x:
588     case Intrinsic::amdgcn_workgroup_id_x:
589       return CreateLiveInRegisterRaw(DAG, &R600::R600_TReg32RegClass,
590                                      R600::T1_X, VT);
591     case Intrinsic::r600_read_tgid_y:
592     case Intrinsic::amdgcn_workgroup_id_y:
593       return CreateLiveInRegisterRaw(DAG, &R600::R600_TReg32RegClass,
594                                      R600::T1_Y, VT);
595     case Intrinsic::r600_read_tgid_z:
596     case Intrinsic::amdgcn_workgroup_id_z:
597       return CreateLiveInRegisterRaw(DAG, &R600::R600_TReg32RegClass,
598                                      R600::T1_Z, VT);
599     case Intrinsic::r600_read_tidig_x:
600     case Intrinsic::amdgcn_workitem_id_x:
601       return CreateLiveInRegisterRaw(DAG, &R600::R600_TReg32RegClass,
602                                      R600::T0_X, VT);
603     case Intrinsic::r600_read_tidig_y:
604     case Intrinsic::amdgcn_workitem_id_y:
605       return CreateLiveInRegisterRaw(DAG, &R600::R600_TReg32RegClass,
606                                      R600::T0_Y, VT);
607     case Intrinsic::r600_read_tidig_z:
608     case Intrinsic::amdgcn_workitem_id_z:
609       return CreateLiveInRegisterRaw(DAG, &R600::R600_TReg32RegClass,
610                                      R600::T0_Z, VT);
611 
612     case Intrinsic::r600_recipsqrt_ieee:
613       return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
614 
615     case Intrinsic::r600_recipsqrt_clamped:
616       return DAG.getNode(AMDGPUISD::RSQ_CLAMP, DL, VT, Op.getOperand(1));
617     default:
618       return Op;
619     }
620 
621     // break out of case ISD::INTRINSIC_WO_CHAIN in switch(Op.getOpcode())
622     break;
623   }
624   } // end switch(Op.getOpcode())
625   return SDValue();
626 }
627 
628 void R600TargetLowering::ReplaceNodeResults(SDNode *N,
629                                             SmallVectorImpl<SDValue> &Results,
630                                             SelectionDAG &DAG) const {
631   switch (N->getOpcode()) {
632   default:
633     AMDGPUTargetLowering::ReplaceNodeResults(N, Results, DAG);
634     return;
635   case ISD::FP_TO_UINT:
636     if (N->getValueType(0) == MVT::i1) {
637       Results.push_back(lowerFP_TO_UINT(N->getOperand(0), DAG));
638       return;
639     }
640     // Since we don't care about out of bounds values we can use FP_TO_SINT for
641     // uints too. The DAGLegalizer code for uint considers some extra cases
642     // which are not necessary here.
643     LLVM_FALLTHROUGH;
644   case ISD::FP_TO_SINT: {
645     if (N->getValueType(0) == MVT::i1) {
646       Results.push_back(lowerFP_TO_SINT(N->getOperand(0), DAG));
647       return;
648     }
649 
650     SDValue Result;
651     if (expandFP_TO_SINT(N, Result, DAG))
652       Results.push_back(Result);
653     return;
654   }
655   case ISD::SDIVREM: {
656     SDValue Op = SDValue(N, 1);
657     SDValue RES = LowerSDIVREM(Op, DAG);
658     Results.push_back(RES);
659     Results.push_back(RES.getValue(1));
660     break;
661   }
662   case ISD::UDIVREM: {
663     SDValue Op = SDValue(N, 0);
664     LowerUDIVREM64(Op, DAG, Results);
665     break;
666   }
667   }
668 }
669 
670 SDValue R600TargetLowering::vectorToVerticalVector(SelectionDAG &DAG,
671                                                    SDValue Vector) const {
672   SDLoc DL(Vector);
673   EVT VecVT = Vector.getValueType();
674   EVT EltVT = VecVT.getVectorElementType();
675   SmallVector<SDValue, 8> Args;
676 
677   for (unsigned i = 0, e = VecVT.getVectorNumElements(); i != e; ++i) {
678     Args.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, Vector,
679                                DAG.getVectorIdxConstant(i, DL)));
680   }
681 
682   return DAG.getNode(AMDGPUISD::BUILD_VERTICAL_VECTOR, DL, VecVT, Args);
683 }
684 
685 SDValue R600TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
686                                                     SelectionDAG &DAG) const {
687   SDLoc DL(Op);
688   SDValue Vector = Op.getOperand(0);
689   SDValue Index = Op.getOperand(1);
690 
691   if (isa<ConstantSDNode>(Index) ||
692       Vector.getOpcode() == AMDGPUISD::BUILD_VERTICAL_VECTOR)
693     return Op;
694 
695   Vector = vectorToVerticalVector(DAG, Vector);
696   return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, Op.getValueType(),
697                      Vector, Index);
698 }
699 
700 SDValue R600TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
701                                                    SelectionDAG &DAG) const {
702   SDLoc DL(Op);
703   SDValue Vector = Op.getOperand(0);
704   SDValue Value = Op.getOperand(1);
705   SDValue Index = Op.getOperand(2);
706 
707   if (isa<ConstantSDNode>(Index) ||
708       Vector.getOpcode() == AMDGPUISD::BUILD_VERTICAL_VECTOR)
709     return Op;
710 
711   Vector = vectorToVerticalVector(DAG, Vector);
712   SDValue Insert = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, Op.getValueType(),
713                                Vector, Value, Index);
714   return vectorToVerticalVector(DAG, Insert);
715 }
716 
717 SDValue R600TargetLowering::LowerGlobalAddress(AMDGPUMachineFunction *MFI,
718                                                SDValue Op,
719                                                SelectionDAG &DAG) const {
720   GlobalAddressSDNode *GSD = cast<GlobalAddressSDNode>(Op);
721   if (GSD->getAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS)
722     return AMDGPUTargetLowering::LowerGlobalAddress(MFI, Op, DAG);
723 
724   const DataLayout &DL = DAG.getDataLayout();
725   const GlobalValue *GV = GSD->getGlobal();
726   MVT ConstPtrVT = getPointerTy(DL, AMDGPUAS::CONSTANT_ADDRESS);
727 
728   SDValue GA = DAG.getTargetGlobalAddress(GV, SDLoc(GSD), ConstPtrVT);
729   return DAG.getNode(AMDGPUISD::CONST_DATA_PTR, SDLoc(GSD), ConstPtrVT, GA);
730 }
731 
732 SDValue R600TargetLowering::LowerTrig(SDValue Op, SelectionDAG &DAG) const {
733   // On hw >= R700, COS/SIN input must be between -1. and 1.
734   // Thus we lower them to TRIG ( FRACT ( x / 2Pi + 0.5) - 0.5)
735   EVT VT = Op.getValueType();
736   SDValue Arg = Op.getOperand(0);
737   SDLoc DL(Op);
738 
739   // TODO: Should this propagate fast-math-flags?
740   SDValue FractPart = DAG.getNode(AMDGPUISD::FRACT, DL, VT,
741       DAG.getNode(ISD::FADD, DL, VT,
742         DAG.getNode(ISD::FMUL, DL, VT, Arg,
743           DAG.getConstantFP(0.15915494309, DL, MVT::f32)),
744         DAG.getConstantFP(0.5, DL, MVT::f32)));
745   unsigned TrigNode;
746   switch (Op.getOpcode()) {
747   case ISD::FCOS:
748     TrigNode = AMDGPUISD::COS_HW;
749     break;
750   case ISD::FSIN:
751     TrigNode = AMDGPUISD::SIN_HW;
752     break;
753   default:
754     llvm_unreachable("Wrong trig opcode");
755   }
756   SDValue TrigVal = DAG.getNode(TrigNode, DL, VT,
757       DAG.getNode(ISD::FADD, DL, VT, FractPart,
758         DAG.getConstantFP(-0.5, DL, MVT::f32)));
759   if (Gen >= AMDGPUSubtarget::R700)
760     return TrigVal;
761   // On R600 hw, COS/SIN input must be between -Pi and Pi.
762   return DAG.getNode(ISD::FMUL, DL, VT, TrigVal,
763       DAG.getConstantFP(numbers::pif, DL, MVT::f32));
764 }
765 
766 SDValue R600TargetLowering::LowerShiftParts(SDValue Op,
767                                             SelectionDAG &DAG) const {
768   SDValue Lo, Hi;
769   expandShiftParts(Op.getNode(), Lo, Hi, DAG);
770   return DAG.getMergeValues({Lo, Hi}, SDLoc(Op));
771 }
772 
773 SDValue R600TargetLowering::LowerUADDSUBO(SDValue Op, SelectionDAG &DAG,
774                                           unsigned mainop, unsigned ovf) const {
775   SDLoc DL(Op);
776   EVT VT = Op.getValueType();
777 
778   SDValue Lo = Op.getOperand(0);
779   SDValue Hi = Op.getOperand(1);
780 
781   SDValue OVF = DAG.getNode(ovf, DL, VT, Lo, Hi);
782   // Extend sign.
783   OVF = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, OVF,
784                     DAG.getValueType(MVT::i1));
785 
786   SDValue Res = DAG.getNode(mainop, DL, VT, Lo, Hi);
787 
788   return DAG.getNode(ISD::MERGE_VALUES, DL, DAG.getVTList(VT, VT), Res, OVF);
789 }
790 
791 SDValue R600TargetLowering::lowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const {
792   SDLoc DL(Op);
793   return DAG.getNode(
794       ISD::SETCC,
795       DL,
796       MVT::i1,
797       Op, DAG.getConstantFP(1.0f, DL, MVT::f32),
798       DAG.getCondCode(ISD::SETEQ));
799 }
800 
801 SDValue R600TargetLowering::lowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const {
802   SDLoc DL(Op);
803   return DAG.getNode(
804       ISD::SETCC,
805       DL,
806       MVT::i1,
807       Op, DAG.getConstantFP(-1.0f, DL, MVT::f32),
808       DAG.getCondCode(ISD::SETEQ));
809 }
810 
811 SDValue R600TargetLowering::LowerImplicitParameter(SelectionDAG &DAG, EVT VT,
812                                                    const SDLoc &DL,
813                                                    unsigned DwordOffset) const {
814   unsigned ByteOffset = DwordOffset * 4;
815   PointerType * PtrType = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
816                                       AMDGPUAS::PARAM_I_ADDRESS);
817 
818   // We shouldn't be using an offset wider than 16-bits for implicit parameters.
819   assert(isInt<16>(ByteOffset));
820 
821   return DAG.getLoad(VT, DL, DAG.getEntryNode(),
822                      DAG.getConstant(ByteOffset, DL, MVT::i32), // PTR
823                      MachinePointerInfo(ConstantPointerNull::get(PtrType)));
824 }
825 
826 bool R600TargetLowering::isZero(SDValue Op) const {
827   if(ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op)) {
828     return Cst->isZero();
829   } else if(ConstantFPSDNode *CstFP = dyn_cast<ConstantFPSDNode>(Op)){
830     return CstFP->isZero();
831   } else {
832     return false;
833   }
834 }
835 
836 bool R600TargetLowering::isHWTrueValue(SDValue Op) const {
837   if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
838     return CFP->isExactlyValue(1.0);
839   }
840   return isAllOnesConstant(Op);
841 }
842 
843 bool R600TargetLowering::isHWFalseValue(SDValue Op) const {
844   if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
845     return CFP->getValueAPF().isZero();
846   }
847   return isNullConstant(Op);
848 }
849 
850 SDValue R600TargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
851   SDLoc DL(Op);
852   EVT VT = Op.getValueType();
853 
854   SDValue LHS = Op.getOperand(0);
855   SDValue RHS = Op.getOperand(1);
856   SDValue True = Op.getOperand(2);
857   SDValue False = Op.getOperand(3);
858   SDValue CC = Op.getOperand(4);
859   SDValue Temp;
860 
861   if (VT == MVT::f32) {
862     DAGCombinerInfo DCI(DAG, AfterLegalizeVectorOps, true, nullptr);
863     SDValue MinMax = combineFMinMaxLegacy(DL, VT, LHS, RHS, True, False, CC, DCI);
864     if (MinMax)
865       return MinMax;
866   }
867 
868   // LHS and RHS are guaranteed to be the same value type
869   EVT CompareVT = LHS.getValueType();
870 
871   // Check if we can lower this to a native operation.
872 
873   // Try to lower to a SET* instruction:
874   //
875   // SET* can match the following patterns:
876   //
877   // select_cc f32, f32, -1,  0, cc_supported
878   // select_cc f32, f32, 1.0f, 0.0f, cc_supported
879   // select_cc i32, i32, -1,  0, cc_supported
880   //
881 
882   // Move hardware True/False values to the correct operand.
883   if (isHWTrueValue(False) && isHWFalseValue(True)) {
884     ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
885     ISD::CondCode InverseCC = ISD::getSetCCInverse(CCOpcode, CompareVT);
886     if (isCondCodeLegal(InverseCC, CompareVT.getSimpleVT())) {
887       std::swap(False, True);
888       CC = DAG.getCondCode(InverseCC);
889     } else {
890       ISD::CondCode SwapInvCC = ISD::getSetCCSwappedOperands(InverseCC);
891       if (isCondCodeLegal(SwapInvCC, CompareVT.getSimpleVT())) {
892         std::swap(False, True);
893         std::swap(LHS, RHS);
894         CC = DAG.getCondCode(SwapInvCC);
895       }
896     }
897   }
898 
899   if (isHWTrueValue(True) && isHWFalseValue(False) &&
900       (CompareVT == VT || VT == MVT::i32)) {
901     // This can be matched by a SET* instruction.
902     return DAG.getNode(ISD::SELECT_CC, DL, VT, LHS, RHS, True, False, CC);
903   }
904 
905   // Try to lower to a CND* instruction:
906   //
907   // CND* can match the following patterns:
908   //
909   // select_cc f32, 0.0, f32, f32, cc_supported
910   // select_cc f32, 0.0, i32, i32, cc_supported
911   // select_cc i32, 0,   f32, f32, cc_supported
912   // select_cc i32, 0,   i32, i32, cc_supported
913   //
914 
915   // Try to move the zero value to the RHS
916   if (isZero(LHS)) {
917     ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
918     // Try swapping the operands
919     ISD::CondCode CCSwapped = ISD::getSetCCSwappedOperands(CCOpcode);
920     if (isCondCodeLegal(CCSwapped, CompareVT.getSimpleVT())) {
921       std::swap(LHS, RHS);
922       CC = DAG.getCondCode(CCSwapped);
923     } else {
924       // Try inverting the condition and then swapping the operands
925       ISD::CondCode CCInv = ISD::getSetCCInverse(CCOpcode, CompareVT);
926       CCSwapped = ISD::getSetCCSwappedOperands(CCInv);
927       if (isCondCodeLegal(CCSwapped, CompareVT.getSimpleVT())) {
928         std::swap(True, False);
929         std::swap(LHS, RHS);
930         CC = DAG.getCondCode(CCSwapped);
931       }
932     }
933   }
934   if (isZero(RHS)) {
935     SDValue Cond = LHS;
936     SDValue Zero = RHS;
937     ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
938     if (CompareVT != VT) {
939       // Bitcast True / False to the correct types.  This will end up being
940       // a nop, but it allows us to define only a single pattern in the
941       // .TD files for each CND* instruction rather than having to have
942       // one pattern for integer True/False and one for fp True/False
943       True = DAG.getNode(ISD::BITCAST, DL, CompareVT, True);
944       False = DAG.getNode(ISD::BITCAST, DL, CompareVT, False);
945     }
946 
947     switch (CCOpcode) {
948     case ISD::SETONE:
949     case ISD::SETUNE:
950     case ISD::SETNE:
951       CCOpcode = ISD::getSetCCInverse(CCOpcode, CompareVT);
952       Temp = True;
953       True = False;
954       False = Temp;
955       break;
956     default:
957       break;
958     }
959     SDValue SelectNode = DAG.getNode(ISD::SELECT_CC, DL, CompareVT,
960         Cond, Zero,
961         True, False,
962         DAG.getCondCode(CCOpcode));
963     return DAG.getNode(ISD::BITCAST, DL, VT, SelectNode);
964   }
965 
966   // If we make it this for it means we have no native instructions to handle
967   // this SELECT_CC, so we must lower it.
968   SDValue HWTrue, HWFalse;
969 
970   if (CompareVT == MVT::f32) {
971     HWTrue = DAG.getConstantFP(1.0f, DL, CompareVT);
972     HWFalse = DAG.getConstantFP(0.0f, DL, CompareVT);
973   } else if (CompareVT == MVT::i32) {
974     HWTrue = DAG.getConstant(-1, DL, CompareVT);
975     HWFalse = DAG.getConstant(0, DL, CompareVT);
976   }
977   else {
978     llvm_unreachable("Unhandled value type in LowerSELECT_CC");
979   }
980 
981   // Lower this unsupported SELECT_CC into a combination of two supported
982   // SELECT_CC operations.
983   SDValue Cond = DAG.getNode(ISD::SELECT_CC, DL, CompareVT, LHS, RHS, HWTrue, HWFalse, CC);
984 
985   return DAG.getNode(ISD::SELECT_CC, DL, VT,
986       Cond, HWFalse,
987       True, False,
988       DAG.getCondCode(ISD::SETNE));
989 }
990 
991 /// LLVM generates byte-addressed pointers.  For indirect addressing, we need to
992 /// convert these pointers to a register index.  Each register holds
993 /// 16 bytes, (4 x 32bit sub-register), but we need to take into account the
994 /// \p StackWidth, which tells us how many of the 4 sub-registers will be used
995 /// for indirect addressing.
996 SDValue R600TargetLowering::stackPtrToRegIndex(SDValue Ptr,
997                                                unsigned StackWidth,
998                                                SelectionDAG &DAG) const {
999   unsigned SRLPad;
1000   switch(StackWidth) {
1001   case 1:
1002     SRLPad = 2;
1003     break;
1004   case 2:
1005     SRLPad = 3;
1006     break;
1007   case 4:
1008     SRLPad = 4;
1009     break;
1010   default: llvm_unreachable("Invalid stack width");
1011   }
1012 
1013   SDLoc DL(Ptr);
1014   return DAG.getNode(ISD::SRL, DL, Ptr.getValueType(), Ptr,
1015                      DAG.getConstant(SRLPad, DL, MVT::i32));
1016 }
1017 
1018 void R600TargetLowering::getStackAddress(unsigned StackWidth,
1019                                          unsigned ElemIdx,
1020                                          unsigned &Channel,
1021                                          unsigned &PtrIncr) const {
1022   switch (StackWidth) {
1023   default:
1024   case 1:
1025     Channel = 0;
1026     if (ElemIdx > 0) {
1027       PtrIncr = 1;
1028     } else {
1029       PtrIncr = 0;
1030     }
1031     break;
1032   case 2:
1033     Channel = ElemIdx % 2;
1034     if (ElemIdx == 2) {
1035       PtrIncr = 1;
1036     } else {
1037       PtrIncr = 0;
1038     }
1039     break;
1040   case 4:
1041     Channel = ElemIdx;
1042     PtrIncr = 0;
1043     break;
1044   }
1045 }
1046 
1047 SDValue R600TargetLowering::lowerPrivateTruncStore(StoreSDNode *Store,
1048                                                    SelectionDAG &DAG) const {
1049   SDLoc DL(Store);
1050   //TODO: Who creates the i8 stores?
1051   assert(Store->isTruncatingStore()
1052          || Store->getValue().getValueType() == MVT::i8);
1053   assert(Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS);
1054 
1055   SDValue Mask;
1056   if (Store->getMemoryVT() == MVT::i8) {
1057     assert(Store->getAlignment() >= 1);
1058     Mask = DAG.getConstant(0xff, DL, MVT::i32);
1059   } else if (Store->getMemoryVT() == MVT::i16) {
1060     assert(Store->getAlignment() >= 2);
1061     Mask = DAG.getConstant(0xffff, DL, MVT::i32);
1062   } else {
1063     llvm_unreachable("Unsupported private trunc store");
1064   }
1065 
1066   SDValue OldChain = Store->getChain();
1067   bool VectorTrunc = (OldChain.getOpcode() == AMDGPUISD::DUMMY_CHAIN);
1068   // Skip dummy
1069   SDValue Chain = VectorTrunc ? OldChain->getOperand(0) : OldChain;
1070   SDValue BasePtr = Store->getBasePtr();
1071   SDValue Offset = Store->getOffset();
1072   EVT MemVT = Store->getMemoryVT();
1073 
1074   SDValue LoadPtr = BasePtr;
1075   if (!Offset.isUndef()) {
1076     LoadPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr, Offset);
1077   }
1078 
1079   // Get dword location
1080   // TODO: this should be eliminated by the future SHR ptr, 2
1081   SDValue Ptr = DAG.getNode(ISD::AND, DL, MVT::i32, LoadPtr,
1082                             DAG.getConstant(0xfffffffc, DL, MVT::i32));
1083 
1084   // Load dword
1085   // TODO: can we be smarter about machine pointer info?
1086   MachinePointerInfo PtrInfo(AMDGPUAS::PRIVATE_ADDRESS);
1087   SDValue Dst = DAG.getLoad(MVT::i32, DL, Chain, Ptr, PtrInfo);
1088 
1089   Chain = Dst.getValue(1);
1090 
1091   // Get offset in dword
1092   SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, LoadPtr,
1093                                 DAG.getConstant(0x3, DL, MVT::i32));
1094 
1095   // Convert byte offset to bit shift
1096   SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1097                                  DAG.getConstant(3, DL, MVT::i32));
1098 
1099   // TODO: Contrary to the name of the function,
1100   // it also handles sub i32 non-truncating stores (like i1)
1101   SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1102                                   Store->getValue());
1103 
1104   // Mask the value to the right type
1105   SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1106 
1107   // Shift the value in place
1108   SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1109                                      MaskedValue, ShiftAmt);
1110 
1111   // Shift the mask in place
1112   SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, Mask, ShiftAmt);
1113 
1114   // Invert the mask. NOTE: if we had native ROL instructions we could
1115   // use inverted mask
1116   DstMask = DAG.getNOT(DL, DstMask, MVT::i32);
1117 
1118   // Cleanup the target bits
1119   Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1120 
1121   // Add the new bits
1122   SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1123 
1124   // Store dword
1125   // TODO: Can we be smarter about MachinePointerInfo?
1126   SDValue NewStore = DAG.getStore(Chain, DL, Value, Ptr, PtrInfo);
1127 
1128   // If we are part of expanded vector, make our neighbors depend on this store
1129   if (VectorTrunc) {
1130     // Make all other vector elements depend on this store
1131     Chain = DAG.getNode(AMDGPUISD::DUMMY_CHAIN, DL, MVT::Other, NewStore);
1132     DAG.ReplaceAllUsesOfValueWith(OldChain, Chain);
1133   }
1134   return NewStore;
1135 }
1136 
1137 SDValue R600TargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1138   StoreSDNode *StoreNode = cast<StoreSDNode>(Op);
1139   unsigned AS = StoreNode->getAddressSpace();
1140 
1141   SDValue Chain = StoreNode->getChain();
1142   SDValue Ptr = StoreNode->getBasePtr();
1143   SDValue Value = StoreNode->getValue();
1144 
1145   EVT VT = Value.getValueType();
1146   EVT MemVT = StoreNode->getMemoryVT();
1147   EVT PtrVT = Ptr.getValueType();
1148 
1149   SDLoc DL(Op);
1150 
1151   const bool TruncatingStore = StoreNode->isTruncatingStore();
1152 
1153   // Neither LOCAL nor PRIVATE can do vectors at the moment
1154   if ((AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::PRIVATE_ADDRESS ||
1155        TruncatingStore) &&
1156       VT.isVector()) {
1157     if ((AS == AMDGPUAS::PRIVATE_ADDRESS) && TruncatingStore) {
1158       // Add an extra level of chain to isolate this vector
1159       SDValue NewChain = DAG.getNode(AMDGPUISD::DUMMY_CHAIN, DL, MVT::Other, Chain);
1160       // TODO: can the chain be replaced without creating a new store?
1161       SDValue NewStore = DAG.getTruncStore(
1162           NewChain, DL, Value, Ptr, StoreNode->getPointerInfo(),
1163           MemVT, StoreNode->getAlignment(),
1164           StoreNode->getMemOperand()->getFlags(), StoreNode->getAAInfo());
1165       StoreNode = cast<StoreSDNode>(NewStore);
1166     }
1167 
1168     return scalarizeVectorStore(StoreNode, DAG);
1169   }
1170 
1171   Align Alignment = StoreNode->getAlign();
1172   if (Alignment < MemVT.getStoreSize() &&
1173       !allowsMisalignedMemoryAccesses(MemVT, AS, Alignment,
1174                                       StoreNode->getMemOperand()->getFlags(),
1175                                       nullptr)) {
1176     return expandUnalignedStore(StoreNode, DAG);
1177   }
1178 
1179   SDValue DWordAddr = DAG.getNode(ISD::SRL, DL, PtrVT, Ptr,
1180                                   DAG.getConstant(2, DL, PtrVT));
1181 
1182   if (AS == AMDGPUAS::GLOBAL_ADDRESS) {
1183     // It is beneficial to create MSKOR here instead of combiner to avoid
1184     // artificial dependencies introduced by RMW
1185     if (TruncatingStore) {
1186       assert(VT.bitsLE(MVT::i32));
1187       SDValue MaskConstant;
1188       if (MemVT == MVT::i8) {
1189         MaskConstant = DAG.getConstant(0xFF, DL, MVT::i32);
1190       } else {
1191         assert(MemVT == MVT::i16);
1192         assert(StoreNode->getAlignment() >= 2);
1193         MaskConstant = DAG.getConstant(0xFFFF, DL, MVT::i32);
1194       }
1195 
1196       SDValue ByteIndex = DAG.getNode(ISD::AND, DL, PtrVT, Ptr,
1197                                       DAG.getConstant(0x00000003, DL, PtrVT));
1198       SDValue BitShift = DAG.getNode(ISD::SHL, DL, VT, ByteIndex,
1199                                      DAG.getConstant(3, DL, VT));
1200 
1201       // Put the mask in correct place
1202       SDValue Mask = DAG.getNode(ISD::SHL, DL, VT, MaskConstant, BitShift);
1203 
1204       // Put the value bits in correct place
1205       SDValue TruncValue = DAG.getNode(ISD::AND, DL, VT, Value, MaskConstant);
1206       SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, VT, TruncValue, BitShift);
1207 
1208       // XXX: If we add a 64-bit ZW register class, then we could use a 2 x i32
1209       // vector instead.
1210       SDValue Src[4] = {
1211         ShiftedValue,
1212         DAG.getConstant(0, DL, MVT::i32),
1213         DAG.getConstant(0, DL, MVT::i32),
1214         Mask
1215       };
1216       SDValue Input = DAG.getBuildVector(MVT::v4i32, DL, Src);
1217       SDValue Args[3] = { Chain, Input, DWordAddr };
1218       return DAG.getMemIntrinsicNode(AMDGPUISD::STORE_MSKOR, DL,
1219                                      Op->getVTList(), Args, MemVT,
1220                                      StoreNode->getMemOperand());
1221     } else if (Ptr->getOpcode() != AMDGPUISD::DWORDADDR && VT.bitsGE(MVT::i32)) {
1222       // Convert pointer from byte address to dword address.
1223       Ptr = DAG.getNode(AMDGPUISD::DWORDADDR, DL, PtrVT, DWordAddr);
1224 
1225       if (StoreNode->isIndexed()) {
1226         llvm_unreachable("Indexed stores not supported yet");
1227       } else {
1228         Chain = DAG.getStore(Chain, DL, Value, Ptr, StoreNode->getMemOperand());
1229       }
1230       return Chain;
1231     }
1232   }
1233 
1234   // GLOBAL_ADDRESS has been handled above, LOCAL_ADDRESS allows all sizes
1235   if (AS != AMDGPUAS::PRIVATE_ADDRESS)
1236     return SDValue();
1237 
1238   if (MemVT.bitsLT(MVT::i32))
1239     return lowerPrivateTruncStore(StoreNode, DAG);
1240 
1241   // Standard i32+ store, tag it with DWORDADDR to note that the address
1242   // has been shifted
1243   if (Ptr.getOpcode() != AMDGPUISD::DWORDADDR) {
1244     Ptr = DAG.getNode(AMDGPUISD::DWORDADDR, DL, PtrVT, DWordAddr);
1245     return DAG.getStore(Chain, DL, Value, Ptr, StoreNode->getMemOperand());
1246   }
1247 
1248   // Tagged i32+ stores will be matched by patterns
1249   return SDValue();
1250 }
1251 
1252 // return (512 + (kc_bank << 12)
1253 static int
1254 ConstantAddressBlock(unsigned AddressSpace) {
1255   switch (AddressSpace) {
1256   case AMDGPUAS::CONSTANT_BUFFER_0:
1257     return 512;
1258   case AMDGPUAS::CONSTANT_BUFFER_1:
1259     return 512 + 4096;
1260   case AMDGPUAS::CONSTANT_BUFFER_2:
1261     return 512 + 4096 * 2;
1262   case AMDGPUAS::CONSTANT_BUFFER_3:
1263     return 512 + 4096 * 3;
1264   case AMDGPUAS::CONSTANT_BUFFER_4:
1265     return 512 + 4096 * 4;
1266   case AMDGPUAS::CONSTANT_BUFFER_5:
1267     return 512 + 4096 * 5;
1268   case AMDGPUAS::CONSTANT_BUFFER_6:
1269     return 512 + 4096 * 6;
1270   case AMDGPUAS::CONSTANT_BUFFER_7:
1271     return 512 + 4096 * 7;
1272   case AMDGPUAS::CONSTANT_BUFFER_8:
1273     return 512 + 4096 * 8;
1274   case AMDGPUAS::CONSTANT_BUFFER_9:
1275     return 512 + 4096 * 9;
1276   case AMDGPUAS::CONSTANT_BUFFER_10:
1277     return 512 + 4096 * 10;
1278   case AMDGPUAS::CONSTANT_BUFFER_11:
1279     return 512 + 4096 * 11;
1280   case AMDGPUAS::CONSTANT_BUFFER_12:
1281     return 512 + 4096 * 12;
1282   case AMDGPUAS::CONSTANT_BUFFER_13:
1283     return 512 + 4096 * 13;
1284   case AMDGPUAS::CONSTANT_BUFFER_14:
1285     return 512 + 4096 * 14;
1286   case AMDGPUAS::CONSTANT_BUFFER_15:
1287     return 512 + 4096 * 15;
1288   default:
1289     return -1;
1290   }
1291 }
1292 
1293 SDValue R600TargetLowering::lowerPrivateExtLoad(SDValue Op,
1294                                                 SelectionDAG &DAG) const {
1295   SDLoc DL(Op);
1296   LoadSDNode *Load = cast<LoadSDNode>(Op);
1297   ISD::LoadExtType ExtType = Load->getExtensionType();
1298   EVT MemVT = Load->getMemoryVT();
1299   assert(Load->getAlignment() >= MemVT.getStoreSize());
1300 
1301   SDValue BasePtr = Load->getBasePtr();
1302   SDValue Chain = Load->getChain();
1303   SDValue Offset = Load->getOffset();
1304 
1305   SDValue LoadPtr = BasePtr;
1306   if (!Offset.isUndef()) {
1307     LoadPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr, Offset);
1308   }
1309 
1310   // Get dword location
1311   // NOTE: this should be eliminated by the future SHR ptr, 2
1312   SDValue Ptr = DAG.getNode(ISD::AND, DL, MVT::i32, LoadPtr,
1313                             DAG.getConstant(0xfffffffc, DL, MVT::i32));
1314 
1315   // Load dword
1316   // TODO: can we be smarter about machine pointer info?
1317   MachinePointerInfo PtrInfo(AMDGPUAS::PRIVATE_ADDRESS);
1318   SDValue Read = DAG.getLoad(MVT::i32, DL, Chain, Ptr, PtrInfo);
1319 
1320   // Get offset within the register.
1321   SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1322                                 LoadPtr, DAG.getConstant(0x3, DL, MVT::i32));
1323 
1324   // Bit offset of target byte (byteIdx * 8).
1325   SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1326                                  DAG.getConstant(3, DL, MVT::i32));
1327 
1328   // Shift to the right.
1329   SDValue Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Read, ShiftAmt);
1330 
1331   // Eliminate the upper bits by setting them to ...
1332   EVT MemEltVT = MemVT.getScalarType();
1333 
1334   if (ExtType == ISD::SEXTLOAD) { // ... ones.
1335     SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1336     Ret = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
1337   } else { // ... or zeros.
1338     Ret = DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
1339   }
1340 
1341   SDValue Ops[] = {
1342     Ret,
1343     Read.getValue(1) // This should be our output chain
1344   };
1345 
1346   return DAG.getMergeValues(Ops, DL);
1347 }
1348 
1349 SDValue R600TargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1350   LoadSDNode *LoadNode = cast<LoadSDNode>(Op);
1351   unsigned AS = LoadNode->getAddressSpace();
1352   EVT MemVT = LoadNode->getMemoryVT();
1353   ISD::LoadExtType ExtType = LoadNode->getExtensionType();
1354 
1355   if (AS == AMDGPUAS::PRIVATE_ADDRESS &&
1356       ExtType != ISD::NON_EXTLOAD && MemVT.bitsLT(MVT::i32)) {
1357     return lowerPrivateExtLoad(Op, DAG);
1358   }
1359 
1360   SDLoc DL(Op);
1361   EVT VT = Op.getValueType();
1362   SDValue Chain = LoadNode->getChain();
1363   SDValue Ptr = LoadNode->getBasePtr();
1364 
1365   if ((LoadNode->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1366       LoadNode->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1367       VT.isVector()) {
1368     SDValue Ops[2];
1369     std::tie(Ops[0], Ops[1]) = scalarizeVectorLoad(LoadNode, DAG);
1370     return DAG.getMergeValues(Ops, DL);
1371   }
1372 
1373   // This is still used for explicit load from addrspace(8)
1374   int ConstantBlock = ConstantAddressBlock(LoadNode->getAddressSpace());
1375   if (ConstantBlock > -1 &&
1376       ((LoadNode->getExtensionType() == ISD::NON_EXTLOAD) ||
1377        (LoadNode->getExtensionType() == ISD::ZEXTLOAD))) {
1378     SDValue Result;
1379     if (isa<Constant>(LoadNode->getMemOperand()->getValue()) ||
1380         isa<ConstantSDNode>(Ptr)) {
1381       return constBufferLoad(LoadNode, LoadNode->getAddressSpace(), DAG);
1382     } else {
1383       //TODO: Does this even work?
1384       // non-constant ptr can't be folded, keeps it as a v4f32 load
1385       Result = DAG.getNode(AMDGPUISD::CONST_ADDRESS, DL, MVT::v4i32,
1386           DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
1387                       DAG.getConstant(4, DL, MVT::i32)),
1388                       DAG.getConstant(LoadNode->getAddressSpace() -
1389                                       AMDGPUAS::CONSTANT_BUFFER_0, DL, MVT::i32)
1390           );
1391     }
1392 
1393     if (!VT.isVector()) {
1394       Result = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, Result,
1395                            DAG.getConstant(0, DL, MVT::i32));
1396     }
1397 
1398     SDValue MergedValues[2] = {
1399       Result,
1400       Chain
1401     };
1402     return DAG.getMergeValues(MergedValues, DL);
1403   }
1404 
1405   // For most operations returning SDValue() will result in the node being
1406   // expanded by the DAG Legalizer. This is not the case for ISD::LOAD, so we
1407   // need to manually expand loads that may be legal in some address spaces and
1408   // illegal in others. SEXT loads from CONSTANT_BUFFER_0 are supported for
1409   // compute shaders, since the data is sign extended when it is uploaded to the
1410   // buffer. However SEXT loads from other address spaces are not supported, so
1411   // we need to expand them here.
1412   if (LoadNode->getExtensionType() == ISD::SEXTLOAD) {
1413     assert(!MemVT.isVector() && (MemVT == MVT::i16 || MemVT == MVT::i8));
1414     SDValue NewLoad = DAG.getExtLoad(
1415         ISD::EXTLOAD, DL, VT, Chain, Ptr, LoadNode->getPointerInfo(), MemVT,
1416         LoadNode->getAlignment(), LoadNode->getMemOperand()->getFlags());
1417     SDValue Res = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, NewLoad,
1418                               DAG.getValueType(MemVT));
1419 
1420     SDValue MergedValues[2] = { Res, Chain };
1421     return DAG.getMergeValues(MergedValues, DL);
1422   }
1423 
1424   if (LoadNode->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
1425     return SDValue();
1426   }
1427 
1428   // DWORDADDR ISD marks already shifted address
1429   if (Ptr.getOpcode() != AMDGPUISD::DWORDADDR) {
1430     assert(VT == MVT::i32);
1431     Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr, DAG.getConstant(2, DL, MVT::i32));
1432     Ptr = DAG.getNode(AMDGPUISD::DWORDADDR, DL, MVT::i32, Ptr);
1433     return DAG.getLoad(MVT::i32, DL, Chain, Ptr, LoadNode->getMemOperand());
1434   }
1435   return SDValue();
1436 }
1437 
1438 SDValue R600TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
1439   SDValue Chain = Op.getOperand(0);
1440   SDValue Cond  = Op.getOperand(1);
1441   SDValue Jump  = Op.getOperand(2);
1442 
1443   return DAG.getNode(AMDGPUISD::BRANCH_COND, SDLoc(Op), Op.getValueType(),
1444                      Chain, Jump, Cond);
1445 }
1446 
1447 SDValue R600TargetLowering::lowerFrameIndex(SDValue Op,
1448                                             SelectionDAG &DAG) const {
1449   MachineFunction &MF = DAG.getMachineFunction();
1450   const R600FrameLowering *TFL = Subtarget->getFrameLowering();
1451 
1452   FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
1453 
1454   unsigned FrameIndex = FIN->getIndex();
1455   Register IgnoredFrameReg;
1456   StackOffset Offset =
1457       TFL->getFrameIndexReference(MF, FrameIndex, IgnoredFrameReg);
1458   return DAG.getConstant(Offset.getFixed() * 4 * TFL->getStackWidth(MF),
1459                          SDLoc(Op), Op.getValueType());
1460 }
1461 
1462 CCAssignFn *R600TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
1463                                                   bool IsVarArg) const {
1464   switch (CC) {
1465   case CallingConv::AMDGPU_KERNEL:
1466   case CallingConv::SPIR_KERNEL:
1467   case CallingConv::C:
1468   case CallingConv::Fast:
1469   case CallingConv::Cold:
1470     llvm_unreachable("kernels should not be handled here");
1471   case CallingConv::AMDGPU_VS:
1472   case CallingConv::AMDGPU_GS:
1473   case CallingConv::AMDGPU_PS:
1474   case CallingConv::AMDGPU_CS:
1475   case CallingConv::AMDGPU_HS:
1476   case CallingConv::AMDGPU_ES:
1477   case CallingConv::AMDGPU_LS:
1478     return CC_R600;
1479   default:
1480     report_fatal_error("Unsupported calling convention.");
1481   }
1482 }
1483 
1484 /// XXX Only kernel functions are supported, so we can assume for now that
1485 /// every function is a kernel function, but in the future we should use
1486 /// separate calling conventions for kernel and non-kernel functions.
1487 SDValue R600TargetLowering::LowerFormalArguments(
1488     SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1489     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
1490     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
1491   SmallVector<CCValAssign, 16> ArgLocs;
1492   CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1493                  *DAG.getContext());
1494   MachineFunction &MF = DAG.getMachineFunction();
1495   SmallVector<ISD::InputArg, 8> LocalIns;
1496 
1497   if (AMDGPU::isShader(CallConv)) {
1498     CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForCall(CallConv, isVarArg));
1499   } else {
1500     analyzeFormalArgumentsCompute(CCInfo, Ins);
1501   }
1502 
1503   for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1504     CCValAssign &VA = ArgLocs[i];
1505     const ISD::InputArg &In = Ins[i];
1506     EVT VT = In.VT;
1507     EVT MemVT = VA.getLocVT();
1508     if (!VT.isVector() && MemVT.isVector()) {
1509       // Get load source type if scalarized.
1510       MemVT = MemVT.getVectorElementType();
1511     }
1512 
1513     if (AMDGPU::isShader(CallConv)) {
1514       Register Reg = MF.addLiveIn(VA.getLocReg(), &R600::R600_Reg128RegClass);
1515       SDValue Register = DAG.getCopyFromReg(Chain, DL, Reg, VT);
1516       InVals.push_back(Register);
1517       continue;
1518     }
1519 
1520     // i64 isn't a legal type, so the register type used ends up as i32, which
1521     // isn't expected here. It attempts to create this sextload, but it ends up
1522     // being invalid. Somehow this seems to work with i64 arguments, but breaks
1523     // for <1 x i64>.
1524 
1525     // The first 36 bytes of the input buffer contains information about
1526     // thread group and global sizes.
1527     ISD::LoadExtType Ext = ISD::NON_EXTLOAD;
1528     if (MemVT.getScalarSizeInBits() != VT.getScalarSizeInBits()) {
1529       // FIXME: This should really check the extload type, but the handling of
1530       // extload vector parameters seems to be broken.
1531 
1532       // Ext = In.Flags.isSExt() ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
1533       Ext = ISD::SEXTLOAD;
1534     }
1535 
1536     // Compute the offset from the value.
1537     // XXX - I think PartOffset should give you this, but it seems to give the
1538     // size of the register which isn't useful.
1539 
1540     unsigned PartOffset = VA.getLocMemOffset();
1541     unsigned Alignment = MinAlign(VT.getStoreSize(), PartOffset);
1542 
1543     MachinePointerInfo PtrInfo(AMDGPUAS::PARAM_I_ADDRESS);
1544     SDValue Arg = DAG.getLoad(
1545         ISD::UNINDEXED, Ext, VT, DL, Chain,
1546         DAG.getConstant(PartOffset, DL, MVT::i32), DAG.getUNDEF(MVT::i32),
1547         PtrInfo,
1548         MemVT, Alignment, MachineMemOperand::MONonTemporal |
1549                                         MachineMemOperand::MODereferenceable |
1550                                         MachineMemOperand::MOInvariant);
1551 
1552     InVals.push_back(Arg);
1553   }
1554   return Chain;
1555 }
1556 
1557 EVT R600TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1558                                            EVT VT) const {
1559    if (!VT.isVector())
1560      return MVT::i32;
1561    return VT.changeVectorElementTypeToInteger();
1562 }
1563 
1564 bool R600TargetLowering::canMergeStoresTo(unsigned AS, EVT MemVT,
1565                                           const MachineFunction &MF) const {
1566   // Local and Private addresses do not handle vectors. Limit to i32
1567   if ((AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::PRIVATE_ADDRESS)) {
1568     return (MemVT.getSizeInBits() <= 32);
1569   }
1570   return true;
1571 }
1572 
1573 bool R600TargetLowering::allowsMisalignedMemoryAccesses(
1574     EVT VT, unsigned AddrSpace, Align Alignment, MachineMemOperand::Flags Flags,
1575     bool *IsFast) const {
1576   if (IsFast)
1577     *IsFast = false;
1578 
1579   if (!VT.isSimple() || VT == MVT::Other)
1580     return false;
1581 
1582   if (VT.bitsLT(MVT::i32))
1583     return false;
1584 
1585   // TODO: This is a rough estimate.
1586   if (IsFast)
1587     *IsFast = true;
1588 
1589   return VT.bitsGT(MVT::i32) && Alignment >= Align(4);
1590 }
1591 
1592 static SDValue CompactSwizzlableVector(
1593   SelectionDAG &DAG, SDValue VectorEntry,
1594   DenseMap<unsigned, unsigned> &RemapSwizzle) {
1595   assert(RemapSwizzle.empty());
1596 
1597   SDLoc DL(VectorEntry);
1598   EVT EltTy = VectorEntry.getValueType().getVectorElementType();
1599 
1600   SDValue NewBldVec[4];
1601   for (unsigned i = 0; i < 4; i++)
1602     NewBldVec[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltTy, VectorEntry,
1603                                DAG.getIntPtrConstant(i, DL));
1604 
1605   for (unsigned i = 0; i < 4; i++) {
1606     if (NewBldVec[i].isUndef())
1607       // We mask write here to teach later passes that the ith element of this
1608       // vector is undef. Thus we can use it to reduce 128 bits reg usage,
1609       // break false dependencies and additionally make assembly easier to read.
1610       RemapSwizzle[i] = 7; // SEL_MASK_WRITE
1611     if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(NewBldVec[i])) {
1612       if (C->isZero()) {
1613         RemapSwizzle[i] = 4; // SEL_0
1614         NewBldVec[i] = DAG.getUNDEF(MVT::f32);
1615       } else if (C->isExactlyValue(1.0)) {
1616         RemapSwizzle[i] = 5; // SEL_1
1617         NewBldVec[i] = DAG.getUNDEF(MVT::f32);
1618       }
1619     }
1620 
1621     if (NewBldVec[i].isUndef())
1622       continue;
1623 
1624     for (unsigned j = 0; j < i; j++) {
1625       if (NewBldVec[i] == NewBldVec[j]) {
1626         NewBldVec[i] = DAG.getUNDEF(NewBldVec[i].getValueType());
1627         RemapSwizzle[i] = j;
1628         break;
1629       }
1630     }
1631   }
1632 
1633   return DAG.getBuildVector(VectorEntry.getValueType(), SDLoc(VectorEntry),
1634                             NewBldVec);
1635 }
1636 
1637 static SDValue ReorganizeVector(SelectionDAG &DAG, SDValue VectorEntry,
1638                                 DenseMap<unsigned, unsigned> &RemapSwizzle) {
1639   assert(RemapSwizzle.empty());
1640 
1641   SDLoc DL(VectorEntry);
1642   EVT EltTy = VectorEntry.getValueType().getVectorElementType();
1643 
1644   SDValue NewBldVec[4];
1645   bool isUnmovable[4] = {false, false, false, false};
1646   for (unsigned i = 0; i < 4; i++)
1647     NewBldVec[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltTy, VectorEntry,
1648                                DAG.getIntPtrConstant(i, DL));
1649 
1650   for (unsigned i = 0; i < 4; i++) {
1651     RemapSwizzle[i] = i;
1652     if (NewBldVec[i].getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
1653       unsigned Idx = cast<ConstantSDNode>(NewBldVec[i].getOperand(1))
1654           ->getZExtValue();
1655       if (i == Idx)
1656         isUnmovable[Idx] = true;
1657     }
1658   }
1659 
1660   for (unsigned i = 0; i < 4; i++) {
1661     if (NewBldVec[i].getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
1662       unsigned Idx = cast<ConstantSDNode>(NewBldVec[i].getOperand(1))
1663           ->getZExtValue();
1664       if (isUnmovable[Idx])
1665         continue;
1666       // Swap i and Idx
1667       std::swap(NewBldVec[Idx], NewBldVec[i]);
1668       std::swap(RemapSwizzle[i], RemapSwizzle[Idx]);
1669       break;
1670     }
1671   }
1672 
1673   return DAG.getBuildVector(VectorEntry.getValueType(), SDLoc(VectorEntry),
1674                             NewBldVec);
1675 }
1676 
1677 SDValue R600TargetLowering::OptimizeSwizzle(SDValue BuildVector, SDValue Swz[4],
1678                                             SelectionDAG &DAG,
1679                                             const SDLoc &DL) const {
1680   // Old -> New swizzle values
1681   DenseMap<unsigned, unsigned> SwizzleRemap;
1682 
1683   BuildVector = CompactSwizzlableVector(DAG, BuildVector, SwizzleRemap);
1684   for (unsigned i = 0; i < 4; i++) {
1685     unsigned Idx = cast<ConstantSDNode>(Swz[i])->getZExtValue();
1686     if (SwizzleRemap.find(Idx) != SwizzleRemap.end())
1687       Swz[i] = DAG.getConstant(SwizzleRemap[Idx], DL, MVT::i32);
1688   }
1689 
1690   SwizzleRemap.clear();
1691   BuildVector = ReorganizeVector(DAG, BuildVector, SwizzleRemap);
1692   for (unsigned i = 0; i < 4; i++) {
1693     unsigned Idx = cast<ConstantSDNode>(Swz[i])->getZExtValue();
1694     if (SwizzleRemap.find(Idx) != SwizzleRemap.end())
1695       Swz[i] = DAG.getConstant(SwizzleRemap[Idx], DL, MVT::i32);
1696   }
1697 
1698   return BuildVector;
1699 }
1700 
1701 SDValue R600TargetLowering::constBufferLoad(LoadSDNode *LoadNode, int Block,
1702                                             SelectionDAG &DAG) const {
1703   SDLoc DL(LoadNode);
1704   EVT VT = LoadNode->getValueType(0);
1705   SDValue Chain = LoadNode->getChain();
1706   SDValue Ptr = LoadNode->getBasePtr();
1707   assert (isa<ConstantSDNode>(Ptr));
1708 
1709   //TODO: Support smaller loads
1710   if (LoadNode->getMemoryVT().getScalarType() != MVT::i32 || !ISD::isNON_EXTLoad(LoadNode))
1711     return SDValue();
1712 
1713   if (LoadNode->getAlignment() < 4)
1714     return SDValue();
1715 
1716   int ConstantBlock = ConstantAddressBlock(Block);
1717 
1718   SDValue Slots[4];
1719   for (unsigned i = 0; i < 4; i++) {
1720     // We want Const position encoded with the following formula :
1721     // (((512 + (kc_bank << 12) + const_index) << 2) + chan)
1722     // const_index is Ptr computed by llvm using an alignment of 16.
1723     // Thus we add (((512 + (kc_bank << 12)) + chan ) * 4 here and
1724     // then div by 4 at the ISel step
1725     SDValue NewPtr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
1726         DAG.getConstant(4 * i + ConstantBlock * 16, DL, MVT::i32));
1727     Slots[i] = DAG.getNode(AMDGPUISD::CONST_ADDRESS, DL, MVT::i32, NewPtr);
1728   }
1729   EVT NewVT = MVT::v4i32;
1730   unsigned NumElements = 4;
1731   if (VT.isVector()) {
1732     NewVT = VT;
1733     NumElements = VT.getVectorNumElements();
1734   }
1735   SDValue Result = DAG.getBuildVector(NewVT, DL, makeArrayRef(Slots, NumElements));
1736   if (!VT.isVector()) {
1737     Result = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, Result,
1738                          DAG.getConstant(0, DL, MVT::i32));
1739   }
1740   SDValue MergedValues[2] = {
1741     Result,
1742     Chain
1743   };
1744   return DAG.getMergeValues(MergedValues, DL);
1745 }
1746 
1747 //===----------------------------------------------------------------------===//
1748 // Custom DAG Optimizations
1749 //===----------------------------------------------------------------------===//
1750 
1751 SDValue R600TargetLowering::PerformDAGCombine(SDNode *N,
1752                                               DAGCombinerInfo &DCI) const {
1753   SelectionDAG &DAG = DCI.DAG;
1754   SDLoc DL(N);
1755 
1756   switch (N->getOpcode()) {
1757   // (f32 fp_round (f64 uint_to_fp a)) -> (f32 uint_to_fp a)
1758   case ISD::FP_ROUND: {
1759       SDValue Arg = N->getOperand(0);
1760       if (Arg.getOpcode() == ISD::UINT_TO_FP && Arg.getValueType() == MVT::f64) {
1761         return DAG.getNode(ISD::UINT_TO_FP, DL, N->getValueType(0),
1762                            Arg.getOperand(0));
1763       }
1764       break;
1765     }
1766 
1767   // (i32 fp_to_sint (fneg (select_cc f32, f32, 1.0, 0.0 cc))) ->
1768   // (i32 select_cc f32, f32, -1, 0 cc)
1769   //
1770   // Mesa's GLSL frontend generates the above pattern a lot and we can lower
1771   // this to one of the SET*_DX10 instructions.
1772   case ISD::FP_TO_SINT: {
1773     SDValue FNeg = N->getOperand(0);
1774     if (FNeg.getOpcode() != ISD::FNEG) {
1775       return SDValue();
1776     }
1777     SDValue SelectCC = FNeg.getOperand(0);
1778     if (SelectCC.getOpcode() != ISD::SELECT_CC ||
1779         SelectCC.getOperand(0).getValueType() != MVT::f32 || // LHS
1780         SelectCC.getOperand(2).getValueType() != MVT::f32 || // True
1781         !isHWTrueValue(SelectCC.getOperand(2)) ||
1782         !isHWFalseValue(SelectCC.getOperand(3))) {
1783       return SDValue();
1784     }
1785 
1786     return DAG.getNode(ISD::SELECT_CC, DL, N->getValueType(0),
1787                            SelectCC.getOperand(0), // LHS
1788                            SelectCC.getOperand(1), // RHS
1789                            DAG.getConstant(-1, DL, MVT::i32), // True
1790                            DAG.getConstant(0, DL, MVT::i32),  // False
1791                            SelectCC.getOperand(4)); // CC
1792   }
1793 
1794   // insert_vector_elt (build_vector elt0, ... , eltN), NewEltIdx, idx
1795   // => build_vector elt0, ... , NewEltIdx, ... , eltN
1796   case ISD::INSERT_VECTOR_ELT: {
1797     SDValue InVec = N->getOperand(0);
1798     SDValue InVal = N->getOperand(1);
1799     SDValue EltNo = N->getOperand(2);
1800 
1801     // If the inserted element is an UNDEF, just use the input vector.
1802     if (InVal.isUndef())
1803       return InVec;
1804 
1805     EVT VT = InVec.getValueType();
1806 
1807     // If we can't generate a legal BUILD_VECTOR, exit
1808     if (!isOperationLegal(ISD::BUILD_VECTOR, VT))
1809       return SDValue();
1810 
1811     // Check that we know which element is being inserted
1812     if (!isa<ConstantSDNode>(EltNo))
1813       return SDValue();
1814     unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
1815 
1816     // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
1817     // be converted to a BUILD_VECTOR).  Fill in the Ops vector with the
1818     // vector elements.
1819     SmallVector<SDValue, 8> Ops;
1820     if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
1821       Ops.append(InVec.getNode()->op_begin(),
1822                  InVec.getNode()->op_end());
1823     } else if (InVec.isUndef()) {
1824       unsigned NElts = VT.getVectorNumElements();
1825       Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
1826     } else {
1827       return SDValue();
1828     }
1829 
1830     // Insert the element
1831     if (Elt < Ops.size()) {
1832       // All the operands of BUILD_VECTOR must have the same type;
1833       // we enforce that here.
1834       EVT OpVT = Ops[0].getValueType();
1835       if (InVal.getValueType() != OpVT)
1836         InVal = OpVT.bitsGT(InVal.getValueType()) ?
1837           DAG.getNode(ISD::ANY_EXTEND, DL, OpVT, InVal) :
1838           DAG.getNode(ISD::TRUNCATE, DL, OpVT, InVal);
1839       Ops[Elt] = InVal;
1840     }
1841 
1842     // Return the new vector
1843     return DAG.getBuildVector(VT, DL, Ops);
1844   }
1845 
1846   // Extract_vec (Build_vector) generated by custom lowering
1847   // also needs to be customly combined
1848   case ISD::EXTRACT_VECTOR_ELT: {
1849     SDValue Arg = N->getOperand(0);
1850     if (Arg.getOpcode() == ISD::BUILD_VECTOR) {
1851       if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1852         unsigned Element = Const->getZExtValue();
1853         return Arg->getOperand(Element);
1854       }
1855     }
1856     if (Arg.getOpcode() == ISD::BITCAST &&
1857         Arg.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
1858         (Arg.getOperand(0).getValueType().getVectorNumElements() ==
1859          Arg.getValueType().getVectorNumElements())) {
1860       if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1861         unsigned Element = Const->getZExtValue();
1862         return DAG.getNode(ISD::BITCAST, DL, N->getVTList(),
1863                            Arg->getOperand(0).getOperand(Element));
1864       }
1865     }
1866     break;
1867   }
1868 
1869   case ISD::SELECT_CC: {
1870     // Try common optimizations
1871     if (SDValue Ret = AMDGPUTargetLowering::PerformDAGCombine(N, DCI))
1872       return Ret;
1873 
1874     // fold selectcc (selectcc x, y, a, b, cc), b, a, b, seteq ->
1875     //      selectcc x, y, a, b, inv(cc)
1876     //
1877     // fold selectcc (selectcc x, y, a, b, cc), b, a, b, setne ->
1878     //      selectcc x, y, a, b, cc
1879     SDValue LHS = N->getOperand(0);
1880     if (LHS.getOpcode() != ISD::SELECT_CC) {
1881       return SDValue();
1882     }
1883 
1884     SDValue RHS = N->getOperand(1);
1885     SDValue True = N->getOperand(2);
1886     SDValue False = N->getOperand(3);
1887     ISD::CondCode NCC = cast<CondCodeSDNode>(N->getOperand(4))->get();
1888 
1889     if (LHS.getOperand(2).getNode() != True.getNode() ||
1890         LHS.getOperand(3).getNode() != False.getNode() ||
1891         RHS.getNode() != False.getNode()) {
1892       return SDValue();
1893     }
1894 
1895     switch (NCC) {
1896     default: return SDValue();
1897     case ISD::SETNE: return LHS;
1898     case ISD::SETEQ: {
1899       ISD::CondCode LHSCC = cast<CondCodeSDNode>(LHS.getOperand(4))->get();
1900       LHSCC = ISD::getSetCCInverse(LHSCC, LHS.getOperand(0).getValueType());
1901       if (DCI.isBeforeLegalizeOps() ||
1902           isCondCodeLegal(LHSCC, LHS.getOperand(0).getSimpleValueType()))
1903         return DAG.getSelectCC(DL,
1904                                LHS.getOperand(0),
1905                                LHS.getOperand(1),
1906                                LHS.getOperand(2),
1907                                LHS.getOperand(3),
1908                                LHSCC);
1909       break;
1910     }
1911     }
1912     return SDValue();
1913   }
1914 
1915   case AMDGPUISD::R600_EXPORT: {
1916     SDValue Arg = N->getOperand(1);
1917     if (Arg.getOpcode() != ISD::BUILD_VECTOR)
1918       break;
1919 
1920     SDValue NewArgs[8] = {
1921       N->getOperand(0), // Chain
1922       SDValue(),
1923       N->getOperand(2), // ArrayBase
1924       N->getOperand(3), // Type
1925       N->getOperand(4), // SWZ_X
1926       N->getOperand(5), // SWZ_Y
1927       N->getOperand(6), // SWZ_Z
1928       N->getOperand(7) // SWZ_W
1929     };
1930     NewArgs[1] = OptimizeSwizzle(N->getOperand(1), &NewArgs[4], DAG, DL);
1931     return DAG.getNode(AMDGPUISD::R600_EXPORT, DL, N->getVTList(), NewArgs);
1932   }
1933   case AMDGPUISD::TEXTURE_FETCH: {
1934     SDValue Arg = N->getOperand(1);
1935     if (Arg.getOpcode() != ISD::BUILD_VECTOR)
1936       break;
1937 
1938     SDValue NewArgs[19] = {
1939       N->getOperand(0),
1940       N->getOperand(1),
1941       N->getOperand(2),
1942       N->getOperand(3),
1943       N->getOperand(4),
1944       N->getOperand(5),
1945       N->getOperand(6),
1946       N->getOperand(7),
1947       N->getOperand(8),
1948       N->getOperand(9),
1949       N->getOperand(10),
1950       N->getOperand(11),
1951       N->getOperand(12),
1952       N->getOperand(13),
1953       N->getOperand(14),
1954       N->getOperand(15),
1955       N->getOperand(16),
1956       N->getOperand(17),
1957       N->getOperand(18),
1958     };
1959     NewArgs[1] = OptimizeSwizzle(N->getOperand(1), &NewArgs[2], DAG, DL);
1960     return DAG.getNode(AMDGPUISD::TEXTURE_FETCH, DL, N->getVTList(), NewArgs);
1961   }
1962 
1963   case ISD::LOAD: {
1964     LoadSDNode *LoadNode = cast<LoadSDNode>(N);
1965     SDValue Ptr = LoadNode->getBasePtr();
1966     if (LoadNode->getAddressSpace() == AMDGPUAS::PARAM_I_ADDRESS &&
1967          isa<ConstantSDNode>(Ptr))
1968       return constBufferLoad(LoadNode, AMDGPUAS::CONSTANT_BUFFER_0, DAG);
1969     break;
1970   }
1971 
1972   default: break;
1973   }
1974 
1975   return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1976 }
1977 
1978 bool R600TargetLowering::FoldOperand(SDNode *ParentNode, unsigned SrcIdx,
1979                                      SDValue &Src, SDValue &Neg, SDValue &Abs,
1980                                      SDValue &Sel, SDValue &Imm,
1981                                      SelectionDAG &DAG) const {
1982   const R600InstrInfo *TII = Subtarget->getInstrInfo();
1983   if (!Src.isMachineOpcode())
1984     return false;
1985 
1986   switch (Src.getMachineOpcode()) {
1987   case R600::FNEG_R600:
1988     if (!Neg.getNode())
1989       return false;
1990     Src = Src.getOperand(0);
1991     Neg = DAG.getTargetConstant(1, SDLoc(ParentNode), MVT::i32);
1992     return true;
1993   case R600::FABS_R600:
1994     if (!Abs.getNode())
1995       return false;
1996     Src = Src.getOperand(0);
1997     Abs = DAG.getTargetConstant(1, SDLoc(ParentNode), MVT::i32);
1998     return true;
1999   case R600::CONST_COPY: {
2000     unsigned Opcode = ParentNode->getMachineOpcode();
2001     bool HasDst = TII->getOperandIdx(Opcode, R600::OpName::dst) > -1;
2002 
2003     if (!Sel.getNode())
2004       return false;
2005 
2006     SDValue CstOffset = Src.getOperand(0);
2007     if (ParentNode->getValueType(0).isVector())
2008       return false;
2009 
2010     // Gather constants values
2011     int SrcIndices[] = {
2012       TII->getOperandIdx(Opcode, R600::OpName::src0),
2013       TII->getOperandIdx(Opcode, R600::OpName::src1),
2014       TII->getOperandIdx(Opcode, R600::OpName::src2),
2015       TII->getOperandIdx(Opcode, R600::OpName::src0_X),
2016       TII->getOperandIdx(Opcode, R600::OpName::src0_Y),
2017       TII->getOperandIdx(Opcode, R600::OpName::src0_Z),
2018       TII->getOperandIdx(Opcode, R600::OpName::src0_W),
2019       TII->getOperandIdx(Opcode, R600::OpName::src1_X),
2020       TII->getOperandIdx(Opcode, R600::OpName::src1_Y),
2021       TII->getOperandIdx(Opcode, R600::OpName::src1_Z),
2022       TII->getOperandIdx(Opcode, R600::OpName::src1_W)
2023     };
2024     std::vector<unsigned> Consts;
2025     for (int OtherSrcIdx : SrcIndices) {
2026       int OtherSelIdx = TII->getSelIdx(Opcode, OtherSrcIdx);
2027       if (OtherSrcIdx < 0 || OtherSelIdx < 0)
2028         continue;
2029       if (HasDst) {
2030         OtherSrcIdx--;
2031         OtherSelIdx--;
2032       }
2033       if (RegisterSDNode *Reg =
2034           dyn_cast<RegisterSDNode>(ParentNode->getOperand(OtherSrcIdx))) {
2035         if (Reg->getReg() == R600::ALU_CONST) {
2036           ConstantSDNode *Cst
2037             = cast<ConstantSDNode>(ParentNode->getOperand(OtherSelIdx));
2038           Consts.push_back(Cst->getZExtValue());
2039         }
2040       }
2041     }
2042 
2043     ConstantSDNode *Cst = cast<ConstantSDNode>(CstOffset);
2044     Consts.push_back(Cst->getZExtValue());
2045     if (!TII->fitsConstReadLimitations(Consts)) {
2046       return false;
2047     }
2048 
2049     Sel = CstOffset;
2050     Src = DAG.getRegister(R600::ALU_CONST, MVT::f32);
2051     return true;
2052   }
2053   case R600::MOV_IMM_GLOBAL_ADDR:
2054     // Check if the Imm slot is used. Taken from below.
2055     if (cast<ConstantSDNode>(Imm)->getZExtValue())
2056       return false;
2057     Imm = Src.getOperand(0);
2058     Src = DAG.getRegister(R600::ALU_LITERAL_X, MVT::i32);
2059     return true;
2060   case R600::MOV_IMM_I32:
2061   case R600::MOV_IMM_F32: {
2062     unsigned ImmReg = R600::ALU_LITERAL_X;
2063     uint64_t ImmValue = 0;
2064 
2065     if (Src.getMachineOpcode() == R600::MOV_IMM_F32) {
2066       ConstantFPSDNode *FPC = cast<ConstantFPSDNode>(Src.getOperand(0));
2067       float FloatValue = FPC->getValueAPF().convertToFloat();
2068       if (FloatValue == 0.0) {
2069         ImmReg = R600::ZERO;
2070       } else if (FloatValue == 0.5) {
2071         ImmReg = R600::HALF;
2072       } else if (FloatValue == 1.0) {
2073         ImmReg = R600::ONE;
2074       } else {
2075         ImmValue = FPC->getValueAPF().bitcastToAPInt().getZExtValue();
2076       }
2077     } else {
2078       ConstantSDNode *C = cast<ConstantSDNode>(Src.getOperand(0));
2079       uint64_t Value = C->getZExtValue();
2080       if (Value == 0) {
2081         ImmReg = R600::ZERO;
2082       } else if (Value == 1) {
2083         ImmReg = R600::ONE_INT;
2084       } else {
2085         ImmValue = Value;
2086       }
2087     }
2088 
2089     // Check that we aren't already using an immediate.
2090     // XXX: It's possible for an instruction to have more than one
2091     // immediate operand, but this is not supported yet.
2092     if (ImmReg == R600::ALU_LITERAL_X) {
2093       if (!Imm.getNode())
2094         return false;
2095       ConstantSDNode *C = cast<ConstantSDNode>(Imm);
2096       if (C->getZExtValue())
2097         return false;
2098       Imm = DAG.getTargetConstant(ImmValue, SDLoc(ParentNode), MVT::i32);
2099     }
2100     Src = DAG.getRegister(ImmReg, MVT::i32);
2101     return true;
2102   }
2103   default:
2104     return false;
2105   }
2106 }
2107 
2108 /// Fold the instructions after selecting them
2109 SDNode *R600TargetLowering::PostISelFolding(MachineSDNode *Node,
2110                                             SelectionDAG &DAG) const {
2111   const R600InstrInfo *TII = Subtarget->getInstrInfo();
2112   if (!Node->isMachineOpcode())
2113     return Node;
2114 
2115   unsigned Opcode = Node->getMachineOpcode();
2116   SDValue FakeOp;
2117 
2118   std::vector<SDValue> Ops(Node->op_begin(), Node->op_end());
2119 
2120   if (Opcode == R600::DOT_4) {
2121     int OperandIdx[] = {
2122       TII->getOperandIdx(Opcode, R600::OpName::src0_X),
2123       TII->getOperandIdx(Opcode, R600::OpName::src0_Y),
2124       TII->getOperandIdx(Opcode, R600::OpName::src0_Z),
2125       TII->getOperandIdx(Opcode, R600::OpName::src0_W),
2126       TII->getOperandIdx(Opcode, R600::OpName::src1_X),
2127       TII->getOperandIdx(Opcode, R600::OpName::src1_Y),
2128       TII->getOperandIdx(Opcode, R600::OpName::src1_Z),
2129       TII->getOperandIdx(Opcode, R600::OpName::src1_W)
2130         };
2131     int NegIdx[] = {
2132       TII->getOperandIdx(Opcode, R600::OpName::src0_neg_X),
2133       TII->getOperandIdx(Opcode, R600::OpName::src0_neg_Y),
2134       TII->getOperandIdx(Opcode, R600::OpName::src0_neg_Z),
2135       TII->getOperandIdx(Opcode, R600::OpName::src0_neg_W),
2136       TII->getOperandIdx(Opcode, R600::OpName::src1_neg_X),
2137       TII->getOperandIdx(Opcode, R600::OpName::src1_neg_Y),
2138       TII->getOperandIdx(Opcode, R600::OpName::src1_neg_Z),
2139       TII->getOperandIdx(Opcode, R600::OpName::src1_neg_W)
2140     };
2141     int AbsIdx[] = {
2142       TII->getOperandIdx(Opcode, R600::OpName::src0_abs_X),
2143       TII->getOperandIdx(Opcode, R600::OpName::src0_abs_Y),
2144       TII->getOperandIdx(Opcode, R600::OpName::src0_abs_Z),
2145       TII->getOperandIdx(Opcode, R600::OpName::src0_abs_W),
2146       TII->getOperandIdx(Opcode, R600::OpName::src1_abs_X),
2147       TII->getOperandIdx(Opcode, R600::OpName::src1_abs_Y),
2148       TII->getOperandIdx(Opcode, R600::OpName::src1_abs_Z),
2149       TII->getOperandIdx(Opcode, R600::OpName::src1_abs_W)
2150     };
2151     for (unsigned i = 0; i < 8; i++) {
2152       if (OperandIdx[i] < 0)
2153         return Node;
2154       SDValue &Src = Ops[OperandIdx[i] - 1];
2155       SDValue &Neg = Ops[NegIdx[i] - 1];
2156       SDValue &Abs = Ops[AbsIdx[i] - 1];
2157       bool HasDst = TII->getOperandIdx(Opcode, R600::OpName::dst) > -1;
2158       int SelIdx = TII->getSelIdx(Opcode, OperandIdx[i]);
2159       if (HasDst)
2160         SelIdx--;
2161       SDValue &Sel = (SelIdx > -1) ? Ops[SelIdx] : FakeOp;
2162       if (FoldOperand(Node, i, Src, Neg, Abs, Sel, FakeOp, DAG))
2163         return DAG.getMachineNode(Opcode, SDLoc(Node), Node->getVTList(), Ops);
2164     }
2165   } else if (Opcode == R600::REG_SEQUENCE) {
2166     for (unsigned i = 1, e = Node->getNumOperands(); i < e; i += 2) {
2167       SDValue &Src = Ops[i];
2168       if (FoldOperand(Node, i, Src, FakeOp, FakeOp, FakeOp, FakeOp, DAG))
2169         return DAG.getMachineNode(Opcode, SDLoc(Node), Node->getVTList(), Ops);
2170     }
2171   } else {
2172     if (!TII->hasInstrModifiers(Opcode))
2173       return Node;
2174     int OperandIdx[] = {
2175       TII->getOperandIdx(Opcode, R600::OpName::src0),
2176       TII->getOperandIdx(Opcode, R600::OpName::src1),
2177       TII->getOperandIdx(Opcode, R600::OpName::src2)
2178     };
2179     int NegIdx[] = {
2180       TII->getOperandIdx(Opcode, R600::OpName::src0_neg),
2181       TII->getOperandIdx(Opcode, R600::OpName::src1_neg),
2182       TII->getOperandIdx(Opcode, R600::OpName::src2_neg)
2183     };
2184     int AbsIdx[] = {
2185       TII->getOperandIdx(Opcode, R600::OpName::src0_abs),
2186       TII->getOperandIdx(Opcode, R600::OpName::src1_abs),
2187       -1
2188     };
2189     for (unsigned i = 0; i < 3; i++) {
2190       if (OperandIdx[i] < 0)
2191         return Node;
2192       SDValue &Src = Ops[OperandIdx[i] - 1];
2193       SDValue &Neg = Ops[NegIdx[i] - 1];
2194       SDValue FakeAbs;
2195       SDValue &Abs = (AbsIdx[i] > -1) ? Ops[AbsIdx[i] - 1] : FakeAbs;
2196       bool HasDst = TII->getOperandIdx(Opcode, R600::OpName::dst) > -1;
2197       int SelIdx = TII->getSelIdx(Opcode, OperandIdx[i]);
2198       int ImmIdx = TII->getOperandIdx(Opcode, R600::OpName::literal);
2199       if (HasDst) {
2200         SelIdx--;
2201         ImmIdx--;
2202       }
2203       SDValue &Sel = (SelIdx > -1) ? Ops[SelIdx] : FakeOp;
2204       SDValue &Imm = Ops[ImmIdx];
2205       if (FoldOperand(Node, i, Src, Neg, Abs, Sel, Imm, DAG))
2206         return DAG.getMachineNode(Opcode, SDLoc(Node), Node->getVTList(), Ops);
2207     }
2208   }
2209 
2210   return Node;
2211 }
2212