xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/R600ISelLowering.cpp (revision 04eeddc0aa8e0a417a16eaf9d7d095207f4a8623)
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);
251   setTargetDAGCombine(ISD::FP_TO_SINT);
252   setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
253   setTargetDAGCombine(ISD::SELECT_CC);
254   setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
255   setTargetDAGCombine(ISD::LOAD);
256 }
257 
258 static inline bool isEOP(MachineBasicBlock::iterator I) {
259   if (std::next(I) == I->getParent()->end())
260     return false;
261   return std::next(I)->getOpcode() == R600::RETURN;
262 }
263 
264 MachineBasicBlock *
265 R600TargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
266                                                 MachineBasicBlock *BB) const {
267   MachineFunction *MF = BB->getParent();
268   MachineRegisterInfo &MRI = MF->getRegInfo();
269   MachineBasicBlock::iterator I = MI;
270   const R600InstrInfo *TII = Subtarget->getInstrInfo();
271 
272   switch (MI.getOpcode()) {
273   default:
274     // Replace LDS_*_RET instruction that don't have any uses with the
275     // equivalent LDS_*_NORET instruction.
276     if (TII->isLDSRetInstr(MI.getOpcode())) {
277       int DstIdx = TII->getOperandIdx(MI.getOpcode(), R600::OpName::dst);
278       assert(DstIdx != -1);
279       MachineInstrBuilder NewMI;
280       // FIXME: getLDSNoRetOp method only handles LDS_1A1D LDS ops. Add
281       //        LDS_1A2D support and remove this special case.
282       if (!MRI.use_empty(MI.getOperand(DstIdx).getReg()) ||
283           MI.getOpcode() == R600::LDS_CMPST_RET)
284         return BB;
285 
286       NewMI = BuildMI(*BB, I, BB->findDebugLoc(I),
287                       TII->get(R600::getLDSNoRetOp(MI.getOpcode())));
288       for (const MachineOperand &MO : llvm::drop_begin(MI.operands()))
289         NewMI.add(MO);
290     } else {
291       return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
292     }
293     break;
294 
295   case R600::FABS_R600: {
296     MachineInstr *NewMI = TII->buildDefaultInstruction(
297         *BB, I, R600::MOV, MI.getOperand(0).getReg(),
298         MI.getOperand(1).getReg());
299     TII->addFlag(*NewMI, 0, MO_FLAG_ABS);
300     break;
301   }
302 
303   case R600::FNEG_R600: {
304     MachineInstr *NewMI = TII->buildDefaultInstruction(
305         *BB, I, R600::MOV, MI.getOperand(0).getReg(),
306         MI.getOperand(1).getReg());
307     TII->addFlag(*NewMI, 0, MO_FLAG_NEG);
308     break;
309   }
310 
311   case R600::MASK_WRITE: {
312     Register maskedRegister = MI.getOperand(0).getReg();
313     assert(maskedRegister.isVirtual());
314     MachineInstr * defInstr = MRI.getVRegDef(maskedRegister);
315     TII->addFlag(*defInstr, 0, MO_FLAG_MASK);
316     break;
317   }
318 
319   case R600::MOV_IMM_F32:
320     TII->buildMovImm(*BB, I, MI.getOperand(0).getReg(), MI.getOperand(1)
321                                                             .getFPImm()
322                                                             ->getValueAPF()
323                                                             .bitcastToAPInt()
324                                                             .getZExtValue());
325     break;
326 
327   case R600::MOV_IMM_I32:
328     TII->buildMovImm(*BB, I, MI.getOperand(0).getReg(),
329                      MI.getOperand(1).getImm());
330     break;
331 
332   case R600::MOV_IMM_GLOBAL_ADDR: {
333     //TODO: Perhaps combine this instruction with the next if possible
334     auto MIB = TII->buildDefaultInstruction(
335         *BB, MI, R600::MOV, MI.getOperand(0).getReg(), R600::ALU_LITERAL_X);
336     int Idx = TII->getOperandIdx(*MIB, R600::OpName::literal);
337     //TODO: Ugh this is rather ugly
338     const MachineOperand &MO = MI.getOperand(1);
339     MIB->getOperand(Idx).ChangeToGA(MO.getGlobal(), MO.getOffset(),
340                                     MO.getTargetFlags());
341     break;
342   }
343 
344   case R600::CONST_COPY: {
345     MachineInstr *NewMI = TII->buildDefaultInstruction(
346         *BB, MI, R600::MOV, MI.getOperand(0).getReg(), R600::ALU_CONST);
347     TII->setImmOperand(*NewMI, R600::OpName::src0_sel,
348                        MI.getOperand(1).getImm());
349     break;
350   }
351 
352   case R600::RAT_WRITE_CACHELESS_32_eg:
353   case R600::RAT_WRITE_CACHELESS_64_eg:
354   case R600::RAT_WRITE_CACHELESS_128_eg:
355     BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI.getOpcode()))
356         .add(MI.getOperand(0))
357         .add(MI.getOperand(1))
358         .addImm(isEOP(I)); // Set End of program bit
359     break;
360 
361   case R600::RAT_STORE_TYPED_eg:
362     BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI.getOpcode()))
363         .add(MI.getOperand(0))
364         .add(MI.getOperand(1))
365         .add(MI.getOperand(2))
366         .addImm(isEOP(I)); // Set End of program bit
367     break;
368 
369   case R600::BRANCH:
370     BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(R600::JUMP))
371         .add(MI.getOperand(0));
372     break;
373 
374   case R600::BRANCH_COND_f32: {
375     MachineInstr *NewMI =
376         BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(R600::PRED_X),
377                 R600::PREDICATE_BIT)
378             .add(MI.getOperand(1))
379             .addImm(R600::PRED_SETNE)
380             .addImm(0); // Flags
381     TII->addFlag(*NewMI, 0, MO_FLAG_PUSH);
382     BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(R600::JUMP_COND))
383         .add(MI.getOperand(0))
384         .addReg(R600::PREDICATE_BIT, RegState::Kill);
385     break;
386   }
387 
388   case R600::BRANCH_COND_i32: {
389     MachineInstr *NewMI =
390         BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(R600::PRED_X),
391                 R600::PREDICATE_BIT)
392             .add(MI.getOperand(1))
393             .addImm(R600::PRED_SETNE_INT)
394             .addImm(0); // Flags
395     TII->addFlag(*NewMI, 0, MO_FLAG_PUSH);
396     BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(R600::JUMP_COND))
397         .add(MI.getOperand(0))
398         .addReg(R600::PREDICATE_BIT, RegState::Kill);
399     break;
400   }
401 
402   case R600::EG_ExportSwz:
403   case R600::R600_ExportSwz: {
404     // Instruction is left unmodified if its not the last one of its type
405     bool isLastInstructionOfItsType = true;
406     unsigned InstExportType = MI.getOperand(1).getImm();
407     for (MachineBasicBlock::iterator NextExportInst = std::next(I),
408          EndBlock = BB->end(); NextExportInst != EndBlock;
409          NextExportInst = std::next(NextExportInst)) {
410       if (NextExportInst->getOpcode() == R600::EG_ExportSwz ||
411           NextExportInst->getOpcode() == R600::R600_ExportSwz) {
412         unsigned CurrentInstExportType = NextExportInst->getOperand(1)
413             .getImm();
414         if (CurrentInstExportType == InstExportType) {
415           isLastInstructionOfItsType = false;
416           break;
417         }
418       }
419     }
420     bool EOP = isEOP(I);
421     if (!EOP && !isLastInstructionOfItsType)
422       return BB;
423     unsigned CfInst = (MI.getOpcode() == R600::EG_ExportSwz) ? 84 : 40;
424     BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI.getOpcode()))
425         .add(MI.getOperand(0))
426         .add(MI.getOperand(1))
427         .add(MI.getOperand(2))
428         .add(MI.getOperand(3))
429         .add(MI.getOperand(4))
430         .add(MI.getOperand(5))
431         .add(MI.getOperand(6))
432         .addImm(CfInst)
433         .addImm(EOP);
434     break;
435   }
436   case R600::RETURN: {
437     return BB;
438   }
439   }
440 
441   MI.eraseFromParent();
442   return BB;
443 }
444 
445 //===----------------------------------------------------------------------===//
446 // Custom DAG Lowering Operations
447 //===----------------------------------------------------------------------===//
448 
449 SDValue R600TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
450   MachineFunction &MF = DAG.getMachineFunction();
451   R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
452   switch (Op.getOpcode()) {
453   default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
454   case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
455   case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
456   case ISD::SHL_PARTS:
457   case ISD::SRA_PARTS:
458   case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
459   case ISD::UADDO: return LowerUADDSUBO(Op, DAG, ISD::ADD, AMDGPUISD::CARRY);
460   case ISD::USUBO: return LowerUADDSUBO(Op, DAG, ISD::SUB, AMDGPUISD::BORROW);
461   case ISD::FCOS:
462   case ISD::FSIN: return LowerTrig(Op, DAG);
463   case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
464   case ISD::STORE: return LowerSTORE(Op, DAG);
465   case ISD::LOAD: {
466     SDValue Result = LowerLOAD(Op, DAG);
467     assert((!Result.getNode() ||
468             Result.getNode()->getNumValues() == 2) &&
469            "Load should return a value and a chain");
470     return Result;
471   }
472 
473   case ISD::BRCOND: return LowerBRCOND(Op, DAG);
474   case ISD::GlobalAddress: return LowerGlobalAddress(MFI, Op, DAG);
475   case ISD::FrameIndex: return lowerFrameIndex(Op, DAG);
476   case ISD::INTRINSIC_VOID: {
477     SDValue Chain = Op.getOperand(0);
478     unsigned IntrinsicID =
479                          cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
480     switch (IntrinsicID) {
481     case Intrinsic::r600_store_swizzle: {
482       SDLoc DL(Op);
483       const SDValue Args[8] = {
484         Chain,
485         Op.getOperand(2), // Export Value
486         Op.getOperand(3), // ArrayBase
487         Op.getOperand(4), // Type
488         DAG.getConstant(0, DL, MVT::i32), // SWZ_X
489         DAG.getConstant(1, DL, MVT::i32), // SWZ_Y
490         DAG.getConstant(2, DL, MVT::i32), // SWZ_Z
491         DAG.getConstant(3, DL, MVT::i32) // SWZ_W
492       };
493       return DAG.getNode(AMDGPUISD::R600_EXPORT, DL, Op.getValueType(), Args);
494     }
495 
496     // default for switch(IntrinsicID)
497     default: break;
498     }
499     // break out of case ISD::INTRINSIC_VOID in switch(Op.getOpcode())
500     break;
501   }
502   case ISD::INTRINSIC_WO_CHAIN: {
503     unsigned IntrinsicID =
504                          cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
505     EVT VT = Op.getValueType();
506     SDLoc DL(Op);
507     switch (IntrinsicID) {
508     case Intrinsic::r600_tex:
509     case Intrinsic::r600_texc: {
510       unsigned TextureOp;
511       switch (IntrinsicID) {
512       case Intrinsic::r600_tex:
513         TextureOp = 0;
514         break;
515       case Intrinsic::r600_texc:
516         TextureOp = 1;
517         break;
518       default:
519         llvm_unreachable("unhandled texture operation");
520       }
521 
522       SDValue TexArgs[19] = {
523         DAG.getConstant(TextureOp, DL, MVT::i32),
524         Op.getOperand(1),
525         DAG.getConstant(0, DL, MVT::i32),
526         DAG.getConstant(1, DL, MVT::i32),
527         DAG.getConstant(2, DL, MVT::i32),
528         DAG.getConstant(3, DL, MVT::i32),
529         Op.getOperand(2),
530         Op.getOperand(3),
531         Op.getOperand(4),
532         DAG.getConstant(0, DL, MVT::i32),
533         DAG.getConstant(1, DL, MVT::i32),
534         DAG.getConstant(2, DL, MVT::i32),
535         DAG.getConstant(3, DL, MVT::i32),
536         Op.getOperand(5),
537         Op.getOperand(6),
538         Op.getOperand(7),
539         Op.getOperand(8),
540         Op.getOperand(9),
541         Op.getOperand(10)
542       };
543       return DAG.getNode(AMDGPUISD::TEXTURE_FETCH, DL, MVT::v4f32, TexArgs);
544     }
545     case Intrinsic::r600_dot4: {
546       SDValue Args[8] = {
547       DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
548           DAG.getConstant(0, DL, MVT::i32)),
549       DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
550           DAG.getConstant(0, DL, MVT::i32)),
551       DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
552           DAG.getConstant(1, DL, MVT::i32)),
553       DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
554           DAG.getConstant(1, DL, MVT::i32)),
555       DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
556           DAG.getConstant(2, DL, MVT::i32)),
557       DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
558           DAG.getConstant(2, DL, MVT::i32)),
559       DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
560           DAG.getConstant(3, DL, MVT::i32)),
561       DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
562           DAG.getConstant(3, DL, MVT::i32))
563       };
564       return DAG.getNode(AMDGPUISD::DOT4, DL, MVT::f32, Args);
565     }
566 
567     case Intrinsic::r600_implicitarg_ptr: {
568       MVT PtrVT = getPointerTy(DAG.getDataLayout(), AMDGPUAS::PARAM_I_ADDRESS);
569       uint32_t ByteOffset = getImplicitParameterOffset(MF, FIRST_IMPLICIT);
570       return DAG.getConstant(ByteOffset, DL, PtrVT);
571     }
572     case Intrinsic::r600_read_ngroups_x:
573       return LowerImplicitParameter(DAG, VT, DL, 0);
574     case Intrinsic::r600_read_ngroups_y:
575       return LowerImplicitParameter(DAG, VT, DL, 1);
576     case Intrinsic::r600_read_ngroups_z:
577       return LowerImplicitParameter(DAG, VT, DL, 2);
578     case Intrinsic::r600_read_global_size_x:
579       return LowerImplicitParameter(DAG, VT, DL, 3);
580     case Intrinsic::r600_read_global_size_y:
581       return LowerImplicitParameter(DAG, VT, DL, 4);
582     case Intrinsic::r600_read_global_size_z:
583       return LowerImplicitParameter(DAG, VT, DL, 5);
584     case Intrinsic::r600_read_local_size_x:
585       return LowerImplicitParameter(DAG, VT, DL, 6);
586     case Intrinsic::r600_read_local_size_y:
587       return LowerImplicitParameter(DAG, VT, DL, 7);
588     case Intrinsic::r600_read_local_size_z:
589       return LowerImplicitParameter(DAG, VT, DL, 8);
590 
591     case Intrinsic::r600_read_tgid_x:
592     case Intrinsic::amdgcn_workgroup_id_x:
593       return CreateLiveInRegisterRaw(DAG, &R600::R600_TReg32RegClass,
594                                      R600::T1_X, VT);
595     case Intrinsic::r600_read_tgid_y:
596     case Intrinsic::amdgcn_workgroup_id_y:
597       return CreateLiveInRegisterRaw(DAG, &R600::R600_TReg32RegClass,
598                                      R600::T1_Y, VT);
599     case Intrinsic::r600_read_tgid_z:
600     case Intrinsic::amdgcn_workgroup_id_z:
601       return CreateLiveInRegisterRaw(DAG, &R600::R600_TReg32RegClass,
602                                      R600::T1_Z, VT);
603     case Intrinsic::r600_read_tidig_x:
604     case Intrinsic::amdgcn_workitem_id_x:
605       return CreateLiveInRegisterRaw(DAG, &R600::R600_TReg32RegClass,
606                                      R600::T0_X, VT);
607     case Intrinsic::r600_read_tidig_y:
608     case Intrinsic::amdgcn_workitem_id_y:
609       return CreateLiveInRegisterRaw(DAG, &R600::R600_TReg32RegClass,
610                                      R600::T0_Y, VT);
611     case Intrinsic::r600_read_tidig_z:
612     case Intrinsic::amdgcn_workitem_id_z:
613       return CreateLiveInRegisterRaw(DAG, &R600::R600_TReg32RegClass,
614                                      R600::T0_Z, VT);
615 
616     case Intrinsic::r600_recipsqrt_ieee:
617       return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
618 
619     case Intrinsic::r600_recipsqrt_clamped:
620       return DAG.getNode(AMDGPUISD::RSQ_CLAMP, DL, VT, Op.getOperand(1));
621     default:
622       return Op;
623     }
624 
625     // break out of case ISD::INTRINSIC_WO_CHAIN in switch(Op.getOpcode())
626     break;
627   }
628   } // end switch(Op.getOpcode())
629   return SDValue();
630 }
631 
632 void R600TargetLowering::ReplaceNodeResults(SDNode *N,
633                                             SmallVectorImpl<SDValue> &Results,
634                                             SelectionDAG &DAG) const {
635   switch (N->getOpcode()) {
636   default:
637     AMDGPUTargetLowering::ReplaceNodeResults(N, Results, DAG);
638     return;
639   case ISD::FP_TO_UINT:
640     if (N->getValueType(0) == MVT::i1) {
641       Results.push_back(lowerFP_TO_UINT(N->getOperand(0), DAG));
642       return;
643     }
644     // Since we don't care about out of bounds values we can use FP_TO_SINT for
645     // uints too. The DAGLegalizer code for uint considers some extra cases
646     // which are not necessary here.
647     LLVM_FALLTHROUGH;
648   case ISD::FP_TO_SINT: {
649     if (N->getValueType(0) == MVT::i1) {
650       Results.push_back(lowerFP_TO_SINT(N->getOperand(0), DAG));
651       return;
652     }
653 
654     SDValue Result;
655     if (expandFP_TO_SINT(N, Result, DAG))
656       Results.push_back(Result);
657     return;
658   }
659   case ISD::SDIVREM: {
660     SDValue Op = SDValue(N, 1);
661     SDValue RES = LowerSDIVREM(Op, DAG);
662     Results.push_back(RES);
663     Results.push_back(RES.getValue(1));
664     break;
665   }
666   case ISD::UDIVREM: {
667     SDValue Op = SDValue(N, 0);
668     LowerUDIVREM64(Op, DAG, Results);
669     break;
670   }
671   }
672 }
673 
674 SDValue R600TargetLowering::vectorToVerticalVector(SelectionDAG &DAG,
675                                                    SDValue Vector) const {
676   SDLoc DL(Vector);
677   EVT VecVT = Vector.getValueType();
678   EVT EltVT = VecVT.getVectorElementType();
679   SmallVector<SDValue, 8> Args;
680 
681   for (unsigned i = 0, e = VecVT.getVectorNumElements(); i != e; ++i) {
682     Args.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, Vector,
683                                DAG.getVectorIdxConstant(i, DL)));
684   }
685 
686   return DAG.getNode(AMDGPUISD::BUILD_VERTICAL_VECTOR, DL, VecVT, Args);
687 }
688 
689 SDValue R600TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
690                                                     SelectionDAG &DAG) const {
691   SDLoc DL(Op);
692   SDValue Vector = Op.getOperand(0);
693   SDValue Index = Op.getOperand(1);
694 
695   if (isa<ConstantSDNode>(Index) ||
696       Vector.getOpcode() == AMDGPUISD::BUILD_VERTICAL_VECTOR)
697     return Op;
698 
699   Vector = vectorToVerticalVector(DAG, Vector);
700   return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, Op.getValueType(),
701                      Vector, Index);
702 }
703 
704 SDValue R600TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
705                                                    SelectionDAG &DAG) const {
706   SDLoc DL(Op);
707   SDValue Vector = Op.getOperand(0);
708   SDValue Value = Op.getOperand(1);
709   SDValue Index = Op.getOperand(2);
710 
711   if (isa<ConstantSDNode>(Index) ||
712       Vector.getOpcode() == AMDGPUISD::BUILD_VERTICAL_VECTOR)
713     return Op;
714 
715   Vector = vectorToVerticalVector(DAG, Vector);
716   SDValue Insert = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, Op.getValueType(),
717                                Vector, Value, Index);
718   return vectorToVerticalVector(DAG, Insert);
719 }
720 
721 SDValue R600TargetLowering::LowerGlobalAddress(AMDGPUMachineFunction *MFI,
722                                                SDValue Op,
723                                                SelectionDAG &DAG) const {
724   GlobalAddressSDNode *GSD = cast<GlobalAddressSDNode>(Op);
725   if (GSD->getAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS)
726     return AMDGPUTargetLowering::LowerGlobalAddress(MFI, Op, DAG);
727 
728   const DataLayout &DL = DAG.getDataLayout();
729   const GlobalValue *GV = GSD->getGlobal();
730   MVT ConstPtrVT = getPointerTy(DL, AMDGPUAS::CONSTANT_ADDRESS);
731 
732   SDValue GA = DAG.getTargetGlobalAddress(GV, SDLoc(GSD), ConstPtrVT);
733   return DAG.getNode(AMDGPUISD::CONST_DATA_PTR, SDLoc(GSD), ConstPtrVT, GA);
734 }
735 
736 SDValue R600TargetLowering::LowerTrig(SDValue Op, SelectionDAG &DAG) const {
737   // On hw >= R700, COS/SIN input must be between -1. and 1.
738   // Thus we lower them to TRIG ( FRACT ( x / 2Pi + 0.5) - 0.5)
739   EVT VT = Op.getValueType();
740   SDValue Arg = Op.getOperand(0);
741   SDLoc DL(Op);
742 
743   // TODO: Should this propagate fast-math-flags?
744   SDValue FractPart = DAG.getNode(AMDGPUISD::FRACT, DL, VT,
745       DAG.getNode(ISD::FADD, DL, VT,
746         DAG.getNode(ISD::FMUL, DL, VT, Arg,
747           DAG.getConstantFP(0.15915494309, DL, MVT::f32)),
748         DAG.getConstantFP(0.5, DL, MVT::f32)));
749   unsigned TrigNode;
750   switch (Op.getOpcode()) {
751   case ISD::FCOS:
752     TrigNode = AMDGPUISD::COS_HW;
753     break;
754   case ISD::FSIN:
755     TrigNode = AMDGPUISD::SIN_HW;
756     break;
757   default:
758     llvm_unreachable("Wrong trig opcode");
759   }
760   SDValue TrigVal = DAG.getNode(TrigNode, DL, VT,
761       DAG.getNode(ISD::FADD, DL, VT, FractPart,
762         DAG.getConstantFP(-0.5, DL, MVT::f32)));
763   if (Gen >= AMDGPUSubtarget::R700)
764     return TrigVal;
765   // On R600 hw, COS/SIN input must be between -Pi and Pi.
766   return DAG.getNode(ISD::FMUL, DL, VT, TrigVal,
767       DAG.getConstantFP(numbers::pif, DL, MVT::f32));
768 }
769 
770 SDValue R600TargetLowering::LowerShiftParts(SDValue Op,
771                                             SelectionDAG &DAG) const {
772   SDValue Lo, Hi;
773   expandShiftParts(Op.getNode(), Lo, Hi, DAG);
774   return DAG.getMergeValues({Lo, Hi}, SDLoc(Op));
775 }
776 
777 SDValue R600TargetLowering::LowerUADDSUBO(SDValue Op, SelectionDAG &DAG,
778                                           unsigned mainop, unsigned ovf) const {
779   SDLoc DL(Op);
780   EVT VT = Op.getValueType();
781 
782   SDValue Lo = Op.getOperand(0);
783   SDValue Hi = Op.getOperand(1);
784 
785   SDValue OVF = DAG.getNode(ovf, DL, VT, Lo, Hi);
786   // Extend sign.
787   OVF = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, OVF,
788                     DAG.getValueType(MVT::i1));
789 
790   SDValue Res = DAG.getNode(mainop, DL, VT, Lo, Hi);
791 
792   return DAG.getNode(ISD::MERGE_VALUES, DL, DAG.getVTList(VT, VT), Res, OVF);
793 }
794 
795 SDValue R600TargetLowering::lowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const {
796   SDLoc DL(Op);
797   return DAG.getNode(
798       ISD::SETCC,
799       DL,
800       MVT::i1,
801       Op, DAG.getConstantFP(1.0f, DL, MVT::f32),
802       DAG.getCondCode(ISD::SETEQ));
803 }
804 
805 SDValue R600TargetLowering::lowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const {
806   SDLoc DL(Op);
807   return DAG.getNode(
808       ISD::SETCC,
809       DL,
810       MVT::i1,
811       Op, DAG.getConstantFP(-1.0f, DL, MVT::f32),
812       DAG.getCondCode(ISD::SETEQ));
813 }
814 
815 SDValue R600TargetLowering::LowerImplicitParameter(SelectionDAG &DAG, EVT VT,
816                                                    const SDLoc &DL,
817                                                    unsigned DwordOffset) const {
818   unsigned ByteOffset = DwordOffset * 4;
819   PointerType * PtrType = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
820                                       AMDGPUAS::PARAM_I_ADDRESS);
821 
822   // We shouldn't be using an offset wider than 16-bits for implicit parameters.
823   assert(isInt<16>(ByteOffset));
824 
825   return DAG.getLoad(VT, DL, DAG.getEntryNode(),
826                      DAG.getConstant(ByteOffset, DL, MVT::i32), // PTR
827                      MachinePointerInfo(ConstantPointerNull::get(PtrType)));
828 }
829 
830 bool R600TargetLowering::isZero(SDValue Op) const {
831   if(ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op)) {
832     return Cst->isZero();
833   } else if(ConstantFPSDNode *CstFP = dyn_cast<ConstantFPSDNode>(Op)){
834     return CstFP->isZero();
835   } else {
836     return false;
837   }
838 }
839 
840 bool R600TargetLowering::isHWTrueValue(SDValue Op) const {
841   if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
842     return CFP->isExactlyValue(1.0);
843   }
844   return isAllOnesConstant(Op);
845 }
846 
847 bool R600TargetLowering::isHWFalseValue(SDValue Op) const {
848   if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
849     return CFP->getValueAPF().isZero();
850   }
851   return isNullConstant(Op);
852 }
853 
854 SDValue R600TargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
855   SDLoc DL(Op);
856   EVT VT = Op.getValueType();
857 
858   SDValue LHS = Op.getOperand(0);
859   SDValue RHS = Op.getOperand(1);
860   SDValue True = Op.getOperand(2);
861   SDValue False = Op.getOperand(3);
862   SDValue CC = Op.getOperand(4);
863   SDValue Temp;
864 
865   if (VT == MVT::f32) {
866     DAGCombinerInfo DCI(DAG, AfterLegalizeVectorOps, true, nullptr);
867     SDValue MinMax = combineFMinMaxLegacy(DL, VT, LHS, RHS, True, False, CC, DCI);
868     if (MinMax)
869       return MinMax;
870   }
871 
872   // LHS and RHS are guaranteed to be the same value type
873   EVT CompareVT = LHS.getValueType();
874 
875   // Check if we can lower this to a native operation.
876 
877   // Try to lower to a SET* instruction:
878   //
879   // SET* can match the following patterns:
880   //
881   // select_cc f32, f32, -1,  0, cc_supported
882   // select_cc f32, f32, 1.0f, 0.0f, cc_supported
883   // select_cc i32, i32, -1,  0, cc_supported
884   //
885 
886   // Move hardware True/False values to the correct operand.
887   if (isHWTrueValue(False) && isHWFalseValue(True)) {
888     ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
889     ISD::CondCode InverseCC = ISD::getSetCCInverse(CCOpcode, CompareVT);
890     if (isCondCodeLegal(InverseCC, CompareVT.getSimpleVT())) {
891       std::swap(False, True);
892       CC = DAG.getCondCode(InverseCC);
893     } else {
894       ISD::CondCode SwapInvCC = ISD::getSetCCSwappedOperands(InverseCC);
895       if (isCondCodeLegal(SwapInvCC, CompareVT.getSimpleVT())) {
896         std::swap(False, True);
897         std::swap(LHS, RHS);
898         CC = DAG.getCondCode(SwapInvCC);
899       }
900     }
901   }
902 
903   if (isHWTrueValue(True) && isHWFalseValue(False) &&
904       (CompareVT == VT || VT == MVT::i32)) {
905     // This can be matched by a SET* instruction.
906     return DAG.getNode(ISD::SELECT_CC, DL, VT, LHS, RHS, True, False, CC);
907   }
908 
909   // Try to lower to a CND* instruction:
910   //
911   // CND* can match the following patterns:
912   //
913   // select_cc f32, 0.0, f32, f32, cc_supported
914   // select_cc f32, 0.0, i32, i32, cc_supported
915   // select_cc i32, 0,   f32, f32, cc_supported
916   // select_cc i32, 0,   i32, i32, cc_supported
917   //
918 
919   // Try to move the zero value to the RHS
920   if (isZero(LHS)) {
921     ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
922     // Try swapping the operands
923     ISD::CondCode CCSwapped = ISD::getSetCCSwappedOperands(CCOpcode);
924     if (isCondCodeLegal(CCSwapped, CompareVT.getSimpleVT())) {
925       std::swap(LHS, RHS);
926       CC = DAG.getCondCode(CCSwapped);
927     } else {
928       // Try inverting the condition and then swapping the operands
929       ISD::CondCode CCInv = ISD::getSetCCInverse(CCOpcode, CompareVT);
930       CCSwapped = ISD::getSetCCSwappedOperands(CCInv);
931       if (isCondCodeLegal(CCSwapped, CompareVT.getSimpleVT())) {
932         std::swap(True, False);
933         std::swap(LHS, RHS);
934         CC = DAG.getCondCode(CCSwapped);
935       }
936     }
937   }
938   if (isZero(RHS)) {
939     SDValue Cond = LHS;
940     SDValue Zero = RHS;
941     ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
942     if (CompareVT != VT) {
943       // Bitcast True / False to the correct types.  This will end up being
944       // a nop, but it allows us to define only a single pattern in the
945       // .TD files for each CND* instruction rather than having to have
946       // one pattern for integer True/False and one for fp True/False
947       True = DAG.getNode(ISD::BITCAST, DL, CompareVT, True);
948       False = DAG.getNode(ISD::BITCAST, DL, CompareVT, False);
949     }
950 
951     switch (CCOpcode) {
952     case ISD::SETONE:
953     case ISD::SETUNE:
954     case ISD::SETNE:
955       CCOpcode = ISD::getSetCCInverse(CCOpcode, CompareVT);
956       Temp = True;
957       True = False;
958       False = Temp;
959       break;
960     default:
961       break;
962     }
963     SDValue SelectNode = DAG.getNode(ISD::SELECT_CC, DL, CompareVT,
964         Cond, Zero,
965         True, False,
966         DAG.getCondCode(CCOpcode));
967     return DAG.getNode(ISD::BITCAST, DL, VT, SelectNode);
968   }
969 
970   // If we make it this for it means we have no native instructions to handle
971   // this SELECT_CC, so we must lower it.
972   SDValue HWTrue, HWFalse;
973 
974   if (CompareVT == MVT::f32) {
975     HWTrue = DAG.getConstantFP(1.0f, DL, CompareVT);
976     HWFalse = DAG.getConstantFP(0.0f, DL, CompareVT);
977   } else if (CompareVT == MVT::i32) {
978     HWTrue = DAG.getConstant(-1, DL, CompareVT);
979     HWFalse = DAG.getConstant(0, DL, CompareVT);
980   }
981   else {
982     llvm_unreachable("Unhandled value type in LowerSELECT_CC");
983   }
984 
985   // Lower this unsupported SELECT_CC into a combination of two supported
986   // SELECT_CC operations.
987   SDValue Cond = DAG.getNode(ISD::SELECT_CC, DL, CompareVT, LHS, RHS, HWTrue, HWFalse, CC);
988 
989   return DAG.getNode(ISD::SELECT_CC, DL, VT,
990       Cond, HWFalse,
991       True, False,
992       DAG.getCondCode(ISD::SETNE));
993 }
994 
995 /// LLVM generates byte-addressed pointers.  For indirect addressing, we need to
996 /// convert these pointers to a register index.  Each register holds
997 /// 16 bytes, (4 x 32bit sub-register), but we need to take into account the
998 /// \p StackWidth, which tells us how many of the 4 sub-registrers will be used
999 /// for indirect addressing.
1000 SDValue R600TargetLowering::stackPtrToRegIndex(SDValue Ptr,
1001                                                unsigned StackWidth,
1002                                                SelectionDAG &DAG) const {
1003   unsigned SRLPad;
1004   switch(StackWidth) {
1005   case 1:
1006     SRLPad = 2;
1007     break;
1008   case 2:
1009     SRLPad = 3;
1010     break;
1011   case 4:
1012     SRLPad = 4;
1013     break;
1014   default: llvm_unreachable("Invalid stack width");
1015   }
1016 
1017   SDLoc DL(Ptr);
1018   return DAG.getNode(ISD::SRL, DL, Ptr.getValueType(), Ptr,
1019                      DAG.getConstant(SRLPad, DL, MVT::i32));
1020 }
1021 
1022 void R600TargetLowering::getStackAddress(unsigned StackWidth,
1023                                          unsigned ElemIdx,
1024                                          unsigned &Channel,
1025                                          unsigned &PtrIncr) const {
1026   switch (StackWidth) {
1027   default:
1028   case 1:
1029     Channel = 0;
1030     if (ElemIdx > 0) {
1031       PtrIncr = 1;
1032     } else {
1033       PtrIncr = 0;
1034     }
1035     break;
1036   case 2:
1037     Channel = ElemIdx % 2;
1038     if (ElemIdx == 2) {
1039       PtrIncr = 1;
1040     } else {
1041       PtrIncr = 0;
1042     }
1043     break;
1044   case 4:
1045     Channel = ElemIdx;
1046     PtrIncr = 0;
1047     break;
1048   }
1049 }
1050 
1051 SDValue R600TargetLowering::lowerPrivateTruncStore(StoreSDNode *Store,
1052                                                    SelectionDAG &DAG) const {
1053   SDLoc DL(Store);
1054   //TODO: Who creates the i8 stores?
1055   assert(Store->isTruncatingStore()
1056          || Store->getValue().getValueType() == MVT::i8);
1057   assert(Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS);
1058 
1059   SDValue Mask;
1060   if (Store->getMemoryVT() == MVT::i8) {
1061     assert(Store->getAlignment() >= 1);
1062     Mask = DAG.getConstant(0xff, DL, MVT::i32);
1063   } else if (Store->getMemoryVT() == MVT::i16) {
1064     assert(Store->getAlignment() >= 2);
1065     Mask = DAG.getConstant(0xffff, DL, MVT::i32);
1066   } else {
1067     llvm_unreachable("Unsupported private trunc store");
1068   }
1069 
1070   SDValue OldChain = Store->getChain();
1071   bool VectorTrunc = (OldChain.getOpcode() == AMDGPUISD::DUMMY_CHAIN);
1072   // Skip dummy
1073   SDValue Chain = VectorTrunc ? OldChain->getOperand(0) : OldChain;
1074   SDValue BasePtr = Store->getBasePtr();
1075   SDValue Offset = Store->getOffset();
1076   EVT MemVT = Store->getMemoryVT();
1077 
1078   SDValue LoadPtr = BasePtr;
1079   if (!Offset.isUndef()) {
1080     LoadPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr, Offset);
1081   }
1082 
1083   // Get dword location
1084   // TODO: this should be eliminated by the future SHR ptr, 2
1085   SDValue Ptr = DAG.getNode(ISD::AND, DL, MVT::i32, LoadPtr,
1086                             DAG.getConstant(0xfffffffc, DL, MVT::i32));
1087 
1088   // Load dword
1089   // TODO: can we be smarter about machine pointer info?
1090   MachinePointerInfo PtrInfo(AMDGPUAS::PRIVATE_ADDRESS);
1091   SDValue Dst = DAG.getLoad(MVT::i32, DL, Chain, Ptr, PtrInfo);
1092 
1093   Chain = Dst.getValue(1);
1094 
1095   // Get offset in dword
1096   SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, LoadPtr,
1097                                 DAG.getConstant(0x3, DL, MVT::i32));
1098 
1099   // Convert byte offset to bit shift
1100   SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1101                                  DAG.getConstant(3, DL, MVT::i32));
1102 
1103   // TODO: Contrary to the name of the functiom,
1104   // it also handles sub i32 non-truncating stores (like i1)
1105   SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1106                                   Store->getValue());
1107 
1108   // Mask the value to the right type
1109   SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1110 
1111   // Shift the value in place
1112   SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1113                                      MaskedValue, ShiftAmt);
1114 
1115   // Shift the mask in place
1116   SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, Mask, ShiftAmt);
1117 
1118   // Invert the mask. NOTE: if we had native ROL instructions we could
1119   // use inverted mask
1120   DstMask = DAG.getNOT(DL, DstMask, MVT::i32);
1121 
1122   // Cleanup the target bits
1123   Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1124 
1125   // Add the new bits
1126   SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1127 
1128   // Store dword
1129   // TODO: Can we be smarter about MachinePointerInfo?
1130   SDValue NewStore = DAG.getStore(Chain, DL, Value, Ptr, PtrInfo);
1131 
1132   // If we are part of expanded vector, make our neighbors depend on this store
1133   if (VectorTrunc) {
1134     // Make all other vector elements depend on this store
1135     Chain = DAG.getNode(AMDGPUISD::DUMMY_CHAIN, DL, MVT::Other, NewStore);
1136     DAG.ReplaceAllUsesOfValueWith(OldChain, Chain);
1137   }
1138   return NewStore;
1139 }
1140 
1141 SDValue R600TargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1142   StoreSDNode *StoreNode = cast<StoreSDNode>(Op);
1143   unsigned AS = StoreNode->getAddressSpace();
1144 
1145   SDValue Chain = StoreNode->getChain();
1146   SDValue Ptr = StoreNode->getBasePtr();
1147   SDValue Value = StoreNode->getValue();
1148 
1149   EVT VT = Value.getValueType();
1150   EVT MemVT = StoreNode->getMemoryVT();
1151   EVT PtrVT = Ptr.getValueType();
1152 
1153   SDLoc DL(Op);
1154 
1155   const bool TruncatingStore = StoreNode->isTruncatingStore();
1156 
1157   // Neither LOCAL nor PRIVATE can do vectors at the moment
1158   if ((AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::PRIVATE_ADDRESS ||
1159        TruncatingStore) &&
1160       VT.isVector()) {
1161     if ((AS == AMDGPUAS::PRIVATE_ADDRESS) && TruncatingStore) {
1162       // Add an extra level of chain to isolate this vector
1163       SDValue NewChain = DAG.getNode(AMDGPUISD::DUMMY_CHAIN, DL, MVT::Other, Chain);
1164       // TODO: can the chain be replaced without creating a new store?
1165       SDValue NewStore = DAG.getTruncStore(
1166           NewChain, DL, Value, Ptr, StoreNode->getPointerInfo(),
1167           MemVT, StoreNode->getAlignment(),
1168           StoreNode->getMemOperand()->getFlags(), StoreNode->getAAInfo());
1169       StoreNode = cast<StoreSDNode>(NewStore);
1170     }
1171 
1172     return scalarizeVectorStore(StoreNode, DAG);
1173   }
1174 
1175   Align Alignment = StoreNode->getAlign();
1176   if (Alignment < MemVT.getStoreSize() &&
1177       !allowsMisalignedMemoryAccesses(MemVT, AS, Alignment,
1178                                       StoreNode->getMemOperand()->getFlags(),
1179                                       nullptr)) {
1180     return expandUnalignedStore(StoreNode, DAG);
1181   }
1182 
1183   SDValue DWordAddr = DAG.getNode(ISD::SRL, DL, PtrVT, Ptr,
1184                                   DAG.getConstant(2, DL, PtrVT));
1185 
1186   if (AS == AMDGPUAS::GLOBAL_ADDRESS) {
1187     // It is beneficial to create MSKOR here instead of combiner to avoid
1188     // artificial dependencies introduced by RMW
1189     if (TruncatingStore) {
1190       assert(VT.bitsLE(MVT::i32));
1191       SDValue MaskConstant;
1192       if (MemVT == MVT::i8) {
1193         MaskConstant = DAG.getConstant(0xFF, DL, MVT::i32);
1194       } else {
1195         assert(MemVT == MVT::i16);
1196         assert(StoreNode->getAlignment() >= 2);
1197         MaskConstant = DAG.getConstant(0xFFFF, DL, MVT::i32);
1198       }
1199 
1200       SDValue ByteIndex = DAG.getNode(ISD::AND, DL, PtrVT, Ptr,
1201                                       DAG.getConstant(0x00000003, DL, PtrVT));
1202       SDValue BitShift = DAG.getNode(ISD::SHL, DL, VT, ByteIndex,
1203                                      DAG.getConstant(3, DL, VT));
1204 
1205       // Put the mask in correct place
1206       SDValue Mask = DAG.getNode(ISD::SHL, DL, VT, MaskConstant, BitShift);
1207 
1208       // Put the value bits in correct place
1209       SDValue TruncValue = DAG.getNode(ISD::AND, DL, VT, Value, MaskConstant);
1210       SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, VT, TruncValue, BitShift);
1211 
1212       // XXX: If we add a 64-bit ZW register class, then we could use a 2 x i32
1213       // vector instead.
1214       SDValue Src[4] = {
1215         ShiftedValue,
1216         DAG.getConstant(0, DL, MVT::i32),
1217         DAG.getConstant(0, DL, MVT::i32),
1218         Mask
1219       };
1220       SDValue Input = DAG.getBuildVector(MVT::v4i32, DL, Src);
1221       SDValue Args[3] = { Chain, Input, DWordAddr };
1222       return DAG.getMemIntrinsicNode(AMDGPUISD::STORE_MSKOR, DL,
1223                                      Op->getVTList(), Args, MemVT,
1224                                      StoreNode->getMemOperand());
1225     } else if (Ptr->getOpcode() != AMDGPUISD::DWORDADDR && VT.bitsGE(MVT::i32)) {
1226       // Convert pointer from byte address to dword address.
1227       Ptr = DAG.getNode(AMDGPUISD::DWORDADDR, DL, PtrVT, DWordAddr);
1228 
1229       if (StoreNode->isIndexed()) {
1230         llvm_unreachable("Indexed stores not supported yet");
1231       } else {
1232         Chain = DAG.getStore(Chain, DL, Value, Ptr, StoreNode->getMemOperand());
1233       }
1234       return Chain;
1235     }
1236   }
1237 
1238   // GLOBAL_ADDRESS has been handled above, LOCAL_ADDRESS allows all sizes
1239   if (AS != AMDGPUAS::PRIVATE_ADDRESS)
1240     return SDValue();
1241 
1242   if (MemVT.bitsLT(MVT::i32))
1243     return lowerPrivateTruncStore(StoreNode, DAG);
1244 
1245   // Standard i32+ store, tag it with DWORDADDR to note that the address
1246   // has been shifted
1247   if (Ptr.getOpcode() != AMDGPUISD::DWORDADDR) {
1248     Ptr = DAG.getNode(AMDGPUISD::DWORDADDR, DL, PtrVT, DWordAddr);
1249     return DAG.getStore(Chain, DL, Value, Ptr, StoreNode->getMemOperand());
1250   }
1251 
1252   // Tagged i32+ stores will be matched by patterns
1253   return SDValue();
1254 }
1255 
1256 // return (512 + (kc_bank << 12)
1257 static int
1258 ConstantAddressBlock(unsigned AddressSpace) {
1259   switch (AddressSpace) {
1260   case AMDGPUAS::CONSTANT_BUFFER_0:
1261     return 512;
1262   case AMDGPUAS::CONSTANT_BUFFER_1:
1263     return 512 + 4096;
1264   case AMDGPUAS::CONSTANT_BUFFER_2:
1265     return 512 + 4096 * 2;
1266   case AMDGPUAS::CONSTANT_BUFFER_3:
1267     return 512 + 4096 * 3;
1268   case AMDGPUAS::CONSTANT_BUFFER_4:
1269     return 512 + 4096 * 4;
1270   case AMDGPUAS::CONSTANT_BUFFER_5:
1271     return 512 + 4096 * 5;
1272   case AMDGPUAS::CONSTANT_BUFFER_6:
1273     return 512 + 4096 * 6;
1274   case AMDGPUAS::CONSTANT_BUFFER_7:
1275     return 512 + 4096 * 7;
1276   case AMDGPUAS::CONSTANT_BUFFER_8:
1277     return 512 + 4096 * 8;
1278   case AMDGPUAS::CONSTANT_BUFFER_9:
1279     return 512 + 4096 * 9;
1280   case AMDGPUAS::CONSTANT_BUFFER_10:
1281     return 512 + 4096 * 10;
1282   case AMDGPUAS::CONSTANT_BUFFER_11:
1283     return 512 + 4096 * 11;
1284   case AMDGPUAS::CONSTANT_BUFFER_12:
1285     return 512 + 4096 * 12;
1286   case AMDGPUAS::CONSTANT_BUFFER_13:
1287     return 512 + 4096 * 13;
1288   case AMDGPUAS::CONSTANT_BUFFER_14:
1289     return 512 + 4096 * 14;
1290   case AMDGPUAS::CONSTANT_BUFFER_15:
1291     return 512 + 4096 * 15;
1292   default:
1293     return -1;
1294   }
1295 }
1296 
1297 SDValue R600TargetLowering::lowerPrivateExtLoad(SDValue Op,
1298                                                 SelectionDAG &DAG) const {
1299   SDLoc DL(Op);
1300   LoadSDNode *Load = cast<LoadSDNode>(Op);
1301   ISD::LoadExtType ExtType = Load->getExtensionType();
1302   EVT MemVT = Load->getMemoryVT();
1303   assert(Load->getAlignment() >= MemVT.getStoreSize());
1304 
1305   SDValue BasePtr = Load->getBasePtr();
1306   SDValue Chain = Load->getChain();
1307   SDValue Offset = Load->getOffset();
1308 
1309   SDValue LoadPtr = BasePtr;
1310   if (!Offset.isUndef()) {
1311     LoadPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr, Offset);
1312   }
1313 
1314   // Get dword location
1315   // NOTE: this should be eliminated by the future SHR ptr, 2
1316   SDValue Ptr = DAG.getNode(ISD::AND, DL, MVT::i32, LoadPtr,
1317                             DAG.getConstant(0xfffffffc, DL, MVT::i32));
1318 
1319   // Load dword
1320   // TODO: can we be smarter about machine pointer info?
1321   MachinePointerInfo PtrInfo(AMDGPUAS::PRIVATE_ADDRESS);
1322   SDValue Read = DAG.getLoad(MVT::i32, DL, Chain, Ptr, PtrInfo);
1323 
1324   // Get offset within the register.
1325   SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1326                                 LoadPtr, DAG.getConstant(0x3, DL, MVT::i32));
1327 
1328   // Bit offset of target byte (byteIdx * 8).
1329   SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1330                                  DAG.getConstant(3, DL, MVT::i32));
1331 
1332   // Shift to the right.
1333   SDValue Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Read, ShiftAmt);
1334 
1335   // Eliminate the upper bits by setting them to ...
1336   EVT MemEltVT = MemVT.getScalarType();
1337 
1338   if (ExtType == ISD::SEXTLOAD) { // ... ones.
1339     SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1340     Ret = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
1341   } else { // ... or zeros.
1342     Ret = DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
1343   }
1344 
1345   SDValue Ops[] = {
1346     Ret,
1347     Read.getValue(1) // This should be our output chain
1348   };
1349 
1350   return DAG.getMergeValues(Ops, DL);
1351 }
1352 
1353 SDValue R600TargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1354   LoadSDNode *LoadNode = cast<LoadSDNode>(Op);
1355   unsigned AS = LoadNode->getAddressSpace();
1356   EVT MemVT = LoadNode->getMemoryVT();
1357   ISD::LoadExtType ExtType = LoadNode->getExtensionType();
1358 
1359   if (AS == AMDGPUAS::PRIVATE_ADDRESS &&
1360       ExtType != ISD::NON_EXTLOAD && MemVT.bitsLT(MVT::i32)) {
1361     return lowerPrivateExtLoad(Op, DAG);
1362   }
1363 
1364   SDLoc DL(Op);
1365   EVT VT = Op.getValueType();
1366   SDValue Chain = LoadNode->getChain();
1367   SDValue Ptr = LoadNode->getBasePtr();
1368 
1369   if ((LoadNode->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1370       LoadNode->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1371       VT.isVector()) {
1372     SDValue Ops[2];
1373     std::tie(Ops[0], Ops[1]) = scalarizeVectorLoad(LoadNode, DAG);
1374     return DAG.getMergeValues(Ops, DL);
1375   }
1376 
1377   // This is still used for explicit load from addrspace(8)
1378   int ConstantBlock = ConstantAddressBlock(LoadNode->getAddressSpace());
1379   if (ConstantBlock > -1 &&
1380       ((LoadNode->getExtensionType() == ISD::NON_EXTLOAD) ||
1381        (LoadNode->getExtensionType() == ISD::ZEXTLOAD))) {
1382     SDValue Result;
1383     if (isa<Constant>(LoadNode->getMemOperand()->getValue()) ||
1384         isa<ConstantSDNode>(Ptr)) {
1385       return constBufferLoad(LoadNode, LoadNode->getAddressSpace(), DAG);
1386     } else {
1387       //TODO: Does this even work?
1388       // non-constant ptr can't be folded, keeps it as a v4f32 load
1389       Result = DAG.getNode(AMDGPUISD::CONST_ADDRESS, DL, MVT::v4i32,
1390           DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
1391                       DAG.getConstant(4, DL, MVT::i32)),
1392                       DAG.getConstant(LoadNode->getAddressSpace() -
1393                                       AMDGPUAS::CONSTANT_BUFFER_0, DL, MVT::i32)
1394           );
1395     }
1396 
1397     if (!VT.isVector()) {
1398       Result = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, Result,
1399                            DAG.getConstant(0, DL, MVT::i32));
1400     }
1401 
1402     SDValue MergedValues[2] = {
1403       Result,
1404       Chain
1405     };
1406     return DAG.getMergeValues(MergedValues, DL);
1407   }
1408 
1409   // For most operations returning SDValue() will result in the node being
1410   // expanded by the DAG Legalizer. This is not the case for ISD::LOAD, so we
1411   // need to manually expand loads that may be legal in some address spaces and
1412   // illegal in others. SEXT loads from CONSTANT_BUFFER_0 are supported for
1413   // compute shaders, since the data is sign extended when it is uploaded to the
1414   // buffer. However SEXT loads from other address spaces are not supported, so
1415   // we need to expand them here.
1416   if (LoadNode->getExtensionType() == ISD::SEXTLOAD) {
1417     assert(!MemVT.isVector() && (MemVT == MVT::i16 || MemVT == MVT::i8));
1418     SDValue NewLoad = DAG.getExtLoad(
1419         ISD::EXTLOAD, DL, VT, Chain, Ptr, LoadNode->getPointerInfo(), MemVT,
1420         LoadNode->getAlignment(), LoadNode->getMemOperand()->getFlags());
1421     SDValue Res = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, NewLoad,
1422                               DAG.getValueType(MemVT));
1423 
1424     SDValue MergedValues[2] = { Res, Chain };
1425     return DAG.getMergeValues(MergedValues, DL);
1426   }
1427 
1428   if (LoadNode->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
1429     return SDValue();
1430   }
1431 
1432   // DWORDADDR ISD marks already shifted address
1433   if (Ptr.getOpcode() != AMDGPUISD::DWORDADDR) {
1434     assert(VT == MVT::i32);
1435     Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr, DAG.getConstant(2, DL, MVT::i32));
1436     Ptr = DAG.getNode(AMDGPUISD::DWORDADDR, DL, MVT::i32, Ptr);
1437     return DAG.getLoad(MVT::i32, DL, Chain, Ptr, LoadNode->getMemOperand());
1438   }
1439   return SDValue();
1440 }
1441 
1442 SDValue R600TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
1443   SDValue Chain = Op.getOperand(0);
1444   SDValue Cond  = Op.getOperand(1);
1445   SDValue Jump  = Op.getOperand(2);
1446 
1447   return DAG.getNode(AMDGPUISD::BRANCH_COND, SDLoc(Op), Op.getValueType(),
1448                      Chain, Jump, Cond);
1449 }
1450 
1451 SDValue R600TargetLowering::lowerFrameIndex(SDValue Op,
1452                                             SelectionDAG &DAG) const {
1453   MachineFunction &MF = DAG.getMachineFunction();
1454   const R600FrameLowering *TFL = Subtarget->getFrameLowering();
1455 
1456   FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
1457 
1458   unsigned FrameIndex = FIN->getIndex();
1459   Register IgnoredFrameReg;
1460   StackOffset Offset =
1461       TFL->getFrameIndexReference(MF, FrameIndex, IgnoredFrameReg);
1462   return DAG.getConstant(Offset.getFixed() * 4 * TFL->getStackWidth(MF),
1463                          SDLoc(Op), Op.getValueType());
1464 }
1465 
1466 CCAssignFn *R600TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
1467                                                   bool IsVarArg) const {
1468   switch (CC) {
1469   case CallingConv::AMDGPU_KERNEL:
1470   case CallingConv::SPIR_KERNEL:
1471   case CallingConv::C:
1472   case CallingConv::Fast:
1473   case CallingConv::Cold:
1474     llvm_unreachable("kernels should not be handled here");
1475   case CallingConv::AMDGPU_VS:
1476   case CallingConv::AMDGPU_GS:
1477   case CallingConv::AMDGPU_PS:
1478   case CallingConv::AMDGPU_CS:
1479   case CallingConv::AMDGPU_HS:
1480   case CallingConv::AMDGPU_ES:
1481   case CallingConv::AMDGPU_LS:
1482     return CC_R600;
1483   default:
1484     report_fatal_error("Unsupported calling convention.");
1485   }
1486 }
1487 
1488 /// XXX Only kernel functions are supported, so we can assume for now that
1489 /// every function is a kernel function, but in the future we should use
1490 /// separate calling conventions for kernel and non-kernel functions.
1491 SDValue R600TargetLowering::LowerFormalArguments(
1492     SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1493     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
1494     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
1495   SmallVector<CCValAssign, 16> ArgLocs;
1496   CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1497                  *DAG.getContext());
1498   MachineFunction &MF = DAG.getMachineFunction();
1499   SmallVector<ISD::InputArg, 8> LocalIns;
1500 
1501   if (AMDGPU::isShader(CallConv)) {
1502     CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForCall(CallConv, isVarArg));
1503   } else {
1504     analyzeFormalArgumentsCompute(CCInfo, Ins);
1505   }
1506 
1507   for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1508     CCValAssign &VA = ArgLocs[i];
1509     const ISD::InputArg &In = Ins[i];
1510     EVT VT = In.VT;
1511     EVT MemVT = VA.getLocVT();
1512     if (!VT.isVector() && MemVT.isVector()) {
1513       // Get load source type if scalarized.
1514       MemVT = MemVT.getVectorElementType();
1515     }
1516 
1517     if (AMDGPU::isShader(CallConv)) {
1518       Register Reg = MF.addLiveIn(VA.getLocReg(), &R600::R600_Reg128RegClass);
1519       SDValue Register = DAG.getCopyFromReg(Chain, DL, Reg, VT);
1520       InVals.push_back(Register);
1521       continue;
1522     }
1523 
1524     // i64 isn't a legal type, so the register type used ends up as i32, which
1525     // isn't expected here. It attempts to create this sextload, but it ends up
1526     // being invalid. Somehow this seems to work with i64 arguments, but breaks
1527     // for <1 x i64>.
1528 
1529     // The first 36 bytes of the input buffer contains information about
1530     // thread group and global sizes.
1531     ISD::LoadExtType Ext = ISD::NON_EXTLOAD;
1532     if (MemVT.getScalarSizeInBits() != VT.getScalarSizeInBits()) {
1533       // FIXME: This should really check the extload type, but the handling of
1534       // extload vector parameters seems to be broken.
1535 
1536       // Ext = In.Flags.isSExt() ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
1537       Ext = ISD::SEXTLOAD;
1538     }
1539 
1540     // Compute the offset from the value.
1541     // XXX - I think PartOffset should give you this, but it seems to give the
1542     // size of the register which isn't useful.
1543 
1544     unsigned PartOffset = VA.getLocMemOffset();
1545     unsigned Alignment = MinAlign(VT.getStoreSize(), PartOffset);
1546 
1547     MachinePointerInfo PtrInfo(AMDGPUAS::PARAM_I_ADDRESS);
1548     SDValue Arg = DAG.getLoad(
1549         ISD::UNINDEXED, Ext, VT, DL, Chain,
1550         DAG.getConstant(PartOffset, DL, MVT::i32), DAG.getUNDEF(MVT::i32),
1551         PtrInfo,
1552         MemVT, Alignment, MachineMemOperand::MONonTemporal |
1553                                         MachineMemOperand::MODereferenceable |
1554                                         MachineMemOperand::MOInvariant);
1555 
1556     InVals.push_back(Arg);
1557   }
1558   return Chain;
1559 }
1560 
1561 EVT R600TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1562                                            EVT VT) const {
1563    if (!VT.isVector())
1564      return MVT::i32;
1565    return VT.changeVectorElementTypeToInteger();
1566 }
1567 
1568 bool R600TargetLowering::canMergeStoresTo(unsigned AS, EVT MemVT,
1569                                           const MachineFunction &MF) const {
1570   // Local and Private addresses do not handle vectors. Limit to i32
1571   if ((AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::PRIVATE_ADDRESS)) {
1572     return (MemVT.getSizeInBits() <= 32);
1573   }
1574   return true;
1575 }
1576 
1577 bool R600TargetLowering::allowsMisalignedMemoryAccesses(
1578     EVT VT, unsigned AddrSpace, Align Alignment, MachineMemOperand::Flags Flags,
1579     bool *IsFast) const {
1580   if (IsFast)
1581     *IsFast = false;
1582 
1583   if (!VT.isSimple() || VT == MVT::Other)
1584     return false;
1585 
1586   if (VT.bitsLT(MVT::i32))
1587     return false;
1588 
1589   // TODO: This is a rough estimate.
1590   if (IsFast)
1591     *IsFast = true;
1592 
1593   return VT.bitsGT(MVT::i32) && Alignment >= Align(4);
1594 }
1595 
1596 static SDValue CompactSwizzlableVector(
1597   SelectionDAG &DAG, SDValue VectorEntry,
1598   DenseMap<unsigned, unsigned> &RemapSwizzle) {
1599   assert(RemapSwizzle.empty());
1600 
1601   SDLoc DL(VectorEntry);
1602   EVT EltTy = VectorEntry.getValueType().getVectorElementType();
1603 
1604   SDValue NewBldVec[4];
1605   for (unsigned i = 0; i < 4; i++)
1606     NewBldVec[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltTy, VectorEntry,
1607                                DAG.getIntPtrConstant(i, DL));
1608 
1609   for (unsigned i = 0; i < 4; i++) {
1610     if (NewBldVec[i].isUndef())
1611       // We mask write here to teach later passes that the ith element of this
1612       // vector is undef. Thus we can use it to reduce 128 bits reg usage,
1613       // break false dependencies and additionnaly make assembly easier to read.
1614       RemapSwizzle[i] = 7; // SEL_MASK_WRITE
1615     if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(NewBldVec[i])) {
1616       if (C->isZero()) {
1617         RemapSwizzle[i] = 4; // SEL_0
1618         NewBldVec[i] = DAG.getUNDEF(MVT::f32);
1619       } else if (C->isExactlyValue(1.0)) {
1620         RemapSwizzle[i] = 5; // SEL_1
1621         NewBldVec[i] = DAG.getUNDEF(MVT::f32);
1622       }
1623     }
1624 
1625     if (NewBldVec[i].isUndef())
1626       continue;
1627 
1628     for (unsigned j = 0; j < i; j++) {
1629       if (NewBldVec[i] == NewBldVec[j]) {
1630         NewBldVec[i] = DAG.getUNDEF(NewBldVec[i].getValueType());
1631         RemapSwizzle[i] = j;
1632         break;
1633       }
1634     }
1635   }
1636 
1637   return DAG.getBuildVector(VectorEntry.getValueType(), SDLoc(VectorEntry),
1638                             NewBldVec);
1639 }
1640 
1641 static SDValue ReorganizeVector(SelectionDAG &DAG, SDValue VectorEntry,
1642                                 DenseMap<unsigned, unsigned> &RemapSwizzle) {
1643   assert(RemapSwizzle.empty());
1644 
1645   SDLoc DL(VectorEntry);
1646   EVT EltTy = VectorEntry.getValueType().getVectorElementType();
1647 
1648   SDValue NewBldVec[4];
1649   bool isUnmovable[4] = {false, false, false, false};
1650   for (unsigned i = 0; i < 4; i++)
1651     NewBldVec[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltTy, VectorEntry,
1652                                DAG.getIntPtrConstant(i, DL));
1653 
1654   for (unsigned i = 0; i < 4; i++) {
1655     RemapSwizzle[i] = i;
1656     if (NewBldVec[i].getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
1657       unsigned Idx = cast<ConstantSDNode>(NewBldVec[i].getOperand(1))
1658           ->getZExtValue();
1659       if (i == Idx)
1660         isUnmovable[Idx] = true;
1661     }
1662   }
1663 
1664   for (unsigned i = 0; i < 4; i++) {
1665     if (NewBldVec[i].getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
1666       unsigned Idx = cast<ConstantSDNode>(NewBldVec[i].getOperand(1))
1667           ->getZExtValue();
1668       if (isUnmovable[Idx])
1669         continue;
1670       // Swap i and Idx
1671       std::swap(NewBldVec[Idx], NewBldVec[i]);
1672       std::swap(RemapSwizzle[i], RemapSwizzle[Idx]);
1673       break;
1674     }
1675   }
1676 
1677   return DAG.getBuildVector(VectorEntry.getValueType(), SDLoc(VectorEntry),
1678                             NewBldVec);
1679 }
1680 
1681 SDValue R600TargetLowering::OptimizeSwizzle(SDValue BuildVector, SDValue Swz[4],
1682                                             SelectionDAG &DAG,
1683                                             const SDLoc &DL) const {
1684   // Old -> New swizzle values
1685   DenseMap<unsigned, unsigned> SwizzleRemap;
1686 
1687   BuildVector = CompactSwizzlableVector(DAG, BuildVector, SwizzleRemap);
1688   for (unsigned i = 0; i < 4; i++) {
1689     unsigned Idx = cast<ConstantSDNode>(Swz[i])->getZExtValue();
1690     if (SwizzleRemap.find(Idx) != SwizzleRemap.end())
1691       Swz[i] = DAG.getConstant(SwizzleRemap[Idx], DL, MVT::i32);
1692   }
1693 
1694   SwizzleRemap.clear();
1695   BuildVector = ReorganizeVector(DAG, BuildVector, SwizzleRemap);
1696   for (unsigned i = 0; i < 4; i++) {
1697     unsigned Idx = cast<ConstantSDNode>(Swz[i])->getZExtValue();
1698     if (SwizzleRemap.find(Idx) != SwizzleRemap.end())
1699       Swz[i] = DAG.getConstant(SwizzleRemap[Idx], DL, MVT::i32);
1700   }
1701 
1702   return BuildVector;
1703 }
1704 
1705 SDValue R600TargetLowering::constBufferLoad(LoadSDNode *LoadNode, int Block,
1706                                             SelectionDAG &DAG) const {
1707   SDLoc DL(LoadNode);
1708   EVT VT = LoadNode->getValueType(0);
1709   SDValue Chain = LoadNode->getChain();
1710   SDValue Ptr = LoadNode->getBasePtr();
1711   assert (isa<ConstantSDNode>(Ptr));
1712 
1713   //TODO: Support smaller loads
1714   if (LoadNode->getMemoryVT().getScalarType() != MVT::i32 || !ISD::isNON_EXTLoad(LoadNode))
1715     return SDValue();
1716 
1717   if (LoadNode->getAlignment() < 4)
1718     return SDValue();
1719 
1720   int ConstantBlock = ConstantAddressBlock(Block);
1721 
1722   SDValue Slots[4];
1723   for (unsigned i = 0; i < 4; i++) {
1724     // We want Const position encoded with the following formula :
1725     // (((512 + (kc_bank << 12) + const_index) << 2) + chan)
1726     // const_index is Ptr computed by llvm using an alignment of 16.
1727     // Thus we add (((512 + (kc_bank << 12)) + chan ) * 4 here and
1728     // then div by 4 at the ISel step
1729     SDValue NewPtr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
1730         DAG.getConstant(4 * i + ConstantBlock * 16, DL, MVT::i32));
1731     Slots[i] = DAG.getNode(AMDGPUISD::CONST_ADDRESS, DL, MVT::i32, NewPtr);
1732   }
1733   EVT NewVT = MVT::v4i32;
1734   unsigned NumElements = 4;
1735   if (VT.isVector()) {
1736     NewVT = VT;
1737     NumElements = VT.getVectorNumElements();
1738   }
1739   SDValue Result = DAG.getBuildVector(NewVT, DL, makeArrayRef(Slots, NumElements));
1740   if (!VT.isVector()) {
1741     Result = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, Result,
1742                          DAG.getConstant(0, DL, MVT::i32));
1743   }
1744   SDValue MergedValues[2] = {
1745     Result,
1746     Chain
1747   };
1748   return DAG.getMergeValues(MergedValues, DL);
1749 }
1750 
1751 //===----------------------------------------------------------------------===//
1752 // Custom DAG Optimizations
1753 //===----------------------------------------------------------------------===//
1754 
1755 SDValue R600TargetLowering::PerformDAGCombine(SDNode *N,
1756                                               DAGCombinerInfo &DCI) const {
1757   SelectionDAG &DAG = DCI.DAG;
1758   SDLoc DL(N);
1759 
1760   switch (N->getOpcode()) {
1761   // (f32 fp_round (f64 uint_to_fp a)) -> (f32 uint_to_fp a)
1762   case ISD::FP_ROUND: {
1763       SDValue Arg = N->getOperand(0);
1764       if (Arg.getOpcode() == ISD::UINT_TO_FP && Arg.getValueType() == MVT::f64) {
1765         return DAG.getNode(ISD::UINT_TO_FP, DL, N->getValueType(0),
1766                            Arg.getOperand(0));
1767       }
1768       break;
1769     }
1770 
1771   // (i32 fp_to_sint (fneg (select_cc f32, f32, 1.0, 0.0 cc))) ->
1772   // (i32 select_cc f32, f32, -1, 0 cc)
1773   //
1774   // Mesa's GLSL frontend generates the above pattern a lot and we can lower
1775   // this to one of the SET*_DX10 instructions.
1776   case ISD::FP_TO_SINT: {
1777     SDValue FNeg = N->getOperand(0);
1778     if (FNeg.getOpcode() != ISD::FNEG) {
1779       return SDValue();
1780     }
1781     SDValue SelectCC = FNeg.getOperand(0);
1782     if (SelectCC.getOpcode() != ISD::SELECT_CC ||
1783         SelectCC.getOperand(0).getValueType() != MVT::f32 || // LHS
1784         SelectCC.getOperand(2).getValueType() != MVT::f32 || // True
1785         !isHWTrueValue(SelectCC.getOperand(2)) ||
1786         !isHWFalseValue(SelectCC.getOperand(3))) {
1787       return SDValue();
1788     }
1789 
1790     return DAG.getNode(ISD::SELECT_CC, DL, N->getValueType(0),
1791                            SelectCC.getOperand(0), // LHS
1792                            SelectCC.getOperand(1), // RHS
1793                            DAG.getConstant(-1, DL, MVT::i32), // True
1794                            DAG.getConstant(0, DL, MVT::i32),  // False
1795                            SelectCC.getOperand(4)); // CC
1796   }
1797 
1798   // insert_vector_elt (build_vector elt0, ... , eltN), NewEltIdx, idx
1799   // => build_vector elt0, ... , NewEltIdx, ... , eltN
1800   case ISD::INSERT_VECTOR_ELT: {
1801     SDValue InVec = N->getOperand(0);
1802     SDValue InVal = N->getOperand(1);
1803     SDValue EltNo = N->getOperand(2);
1804 
1805     // If the inserted element is an UNDEF, just use the input vector.
1806     if (InVal.isUndef())
1807       return InVec;
1808 
1809     EVT VT = InVec.getValueType();
1810 
1811     // If we can't generate a legal BUILD_VECTOR, exit
1812     if (!isOperationLegal(ISD::BUILD_VECTOR, VT))
1813       return SDValue();
1814 
1815     // Check that we know which element is being inserted
1816     if (!isa<ConstantSDNode>(EltNo))
1817       return SDValue();
1818     unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
1819 
1820     // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
1821     // be converted to a BUILD_VECTOR).  Fill in the Ops vector with the
1822     // vector elements.
1823     SmallVector<SDValue, 8> Ops;
1824     if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
1825       Ops.append(InVec.getNode()->op_begin(),
1826                  InVec.getNode()->op_end());
1827     } else if (InVec.isUndef()) {
1828       unsigned NElts = VT.getVectorNumElements();
1829       Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
1830     } else {
1831       return SDValue();
1832     }
1833 
1834     // Insert the element
1835     if (Elt < Ops.size()) {
1836       // All the operands of BUILD_VECTOR must have the same type;
1837       // we enforce that here.
1838       EVT OpVT = Ops[0].getValueType();
1839       if (InVal.getValueType() != OpVT)
1840         InVal = OpVT.bitsGT(InVal.getValueType()) ?
1841           DAG.getNode(ISD::ANY_EXTEND, DL, OpVT, InVal) :
1842           DAG.getNode(ISD::TRUNCATE, DL, OpVT, InVal);
1843       Ops[Elt] = InVal;
1844     }
1845 
1846     // Return the new vector
1847     return DAG.getBuildVector(VT, DL, Ops);
1848   }
1849 
1850   // Extract_vec (Build_vector) generated by custom lowering
1851   // also needs to be customly combined
1852   case ISD::EXTRACT_VECTOR_ELT: {
1853     SDValue Arg = N->getOperand(0);
1854     if (Arg.getOpcode() == ISD::BUILD_VECTOR) {
1855       if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1856         unsigned Element = Const->getZExtValue();
1857         return Arg->getOperand(Element);
1858       }
1859     }
1860     if (Arg.getOpcode() == ISD::BITCAST &&
1861         Arg.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
1862         (Arg.getOperand(0).getValueType().getVectorNumElements() ==
1863          Arg.getValueType().getVectorNumElements())) {
1864       if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1865         unsigned Element = Const->getZExtValue();
1866         return DAG.getNode(ISD::BITCAST, DL, N->getVTList(),
1867                            Arg->getOperand(0).getOperand(Element));
1868       }
1869     }
1870     break;
1871   }
1872 
1873   case ISD::SELECT_CC: {
1874     // Try common optimizations
1875     if (SDValue Ret = AMDGPUTargetLowering::PerformDAGCombine(N, DCI))
1876       return Ret;
1877 
1878     // fold selectcc (selectcc x, y, a, b, cc), b, a, b, seteq ->
1879     //      selectcc x, y, a, b, inv(cc)
1880     //
1881     // fold selectcc (selectcc x, y, a, b, cc), b, a, b, setne ->
1882     //      selectcc x, y, a, b, cc
1883     SDValue LHS = N->getOperand(0);
1884     if (LHS.getOpcode() != ISD::SELECT_CC) {
1885       return SDValue();
1886     }
1887 
1888     SDValue RHS = N->getOperand(1);
1889     SDValue True = N->getOperand(2);
1890     SDValue False = N->getOperand(3);
1891     ISD::CondCode NCC = cast<CondCodeSDNode>(N->getOperand(4))->get();
1892 
1893     if (LHS.getOperand(2).getNode() != True.getNode() ||
1894         LHS.getOperand(3).getNode() != False.getNode() ||
1895         RHS.getNode() != False.getNode()) {
1896       return SDValue();
1897     }
1898 
1899     switch (NCC) {
1900     default: return SDValue();
1901     case ISD::SETNE: return LHS;
1902     case ISD::SETEQ: {
1903       ISD::CondCode LHSCC = cast<CondCodeSDNode>(LHS.getOperand(4))->get();
1904       LHSCC = ISD::getSetCCInverse(LHSCC, LHS.getOperand(0).getValueType());
1905       if (DCI.isBeforeLegalizeOps() ||
1906           isCondCodeLegal(LHSCC, LHS.getOperand(0).getSimpleValueType()))
1907         return DAG.getSelectCC(DL,
1908                                LHS.getOperand(0),
1909                                LHS.getOperand(1),
1910                                LHS.getOperand(2),
1911                                LHS.getOperand(3),
1912                                LHSCC);
1913       break;
1914     }
1915     }
1916     return SDValue();
1917   }
1918 
1919   case AMDGPUISD::R600_EXPORT: {
1920     SDValue Arg = N->getOperand(1);
1921     if (Arg.getOpcode() != ISD::BUILD_VECTOR)
1922       break;
1923 
1924     SDValue NewArgs[8] = {
1925       N->getOperand(0), // Chain
1926       SDValue(),
1927       N->getOperand(2), // ArrayBase
1928       N->getOperand(3), // Type
1929       N->getOperand(4), // SWZ_X
1930       N->getOperand(5), // SWZ_Y
1931       N->getOperand(6), // SWZ_Z
1932       N->getOperand(7) // SWZ_W
1933     };
1934     NewArgs[1] = OptimizeSwizzle(N->getOperand(1), &NewArgs[4], DAG, DL);
1935     return DAG.getNode(AMDGPUISD::R600_EXPORT, DL, N->getVTList(), NewArgs);
1936   }
1937   case AMDGPUISD::TEXTURE_FETCH: {
1938     SDValue Arg = N->getOperand(1);
1939     if (Arg.getOpcode() != ISD::BUILD_VECTOR)
1940       break;
1941 
1942     SDValue NewArgs[19] = {
1943       N->getOperand(0),
1944       N->getOperand(1),
1945       N->getOperand(2),
1946       N->getOperand(3),
1947       N->getOperand(4),
1948       N->getOperand(5),
1949       N->getOperand(6),
1950       N->getOperand(7),
1951       N->getOperand(8),
1952       N->getOperand(9),
1953       N->getOperand(10),
1954       N->getOperand(11),
1955       N->getOperand(12),
1956       N->getOperand(13),
1957       N->getOperand(14),
1958       N->getOperand(15),
1959       N->getOperand(16),
1960       N->getOperand(17),
1961       N->getOperand(18),
1962     };
1963     NewArgs[1] = OptimizeSwizzle(N->getOperand(1), &NewArgs[2], DAG, DL);
1964     return DAG.getNode(AMDGPUISD::TEXTURE_FETCH, DL, N->getVTList(), NewArgs);
1965   }
1966 
1967   case ISD::LOAD: {
1968     LoadSDNode *LoadNode = cast<LoadSDNode>(N);
1969     SDValue Ptr = LoadNode->getBasePtr();
1970     if (LoadNode->getAddressSpace() == AMDGPUAS::PARAM_I_ADDRESS &&
1971          isa<ConstantSDNode>(Ptr))
1972       return constBufferLoad(LoadNode, AMDGPUAS::CONSTANT_BUFFER_0, DAG);
1973     break;
1974   }
1975 
1976   default: break;
1977   }
1978 
1979   return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
1980 }
1981 
1982 bool R600TargetLowering::FoldOperand(SDNode *ParentNode, unsigned SrcIdx,
1983                                      SDValue &Src, SDValue &Neg, SDValue &Abs,
1984                                      SDValue &Sel, SDValue &Imm,
1985                                      SelectionDAG &DAG) const {
1986   const R600InstrInfo *TII = Subtarget->getInstrInfo();
1987   if (!Src.isMachineOpcode())
1988     return false;
1989 
1990   switch (Src.getMachineOpcode()) {
1991   case R600::FNEG_R600:
1992     if (!Neg.getNode())
1993       return false;
1994     Src = Src.getOperand(0);
1995     Neg = DAG.getTargetConstant(1, SDLoc(ParentNode), MVT::i32);
1996     return true;
1997   case R600::FABS_R600:
1998     if (!Abs.getNode())
1999       return false;
2000     Src = Src.getOperand(0);
2001     Abs = DAG.getTargetConstant(1, SDLoc(ParentNode), MVT::i32);
2002     return true;
2003   case R600::CONST_COPY: {
2004     unsigned Opcode = ParentNode->getMachineOpcode();
2005     bool HasDst = TII->getOperandIdx(Opcode, R600::OpName::dst) > -1;
2006 
2007     if (!Sel.getNode())
2008       return false;
2009 
2010     SDValue CstOffset = Src.getOperand(0);
2011     if (ParentNode->getValueType(0).isVector())
2012       return false;
2013 
2014     // Gather constants values
2015     int SrcIndices[] = {
2016       TII->getOperandIdx(Opcode, R600::OpName::src0),
2017       TII->getOperandIdx(Opcode, R600::OpName::src1),
2018       TII->getOperandIdx(Opcode, R600::OpName::src2),
2019       TII->getOperandIdx(Opcode, R600::OpName::src0_X),
2020       TII->getOperandIdx(Opcode, R600::OpName::src0_Y),
2021       TII->getOperandIdx(Opcode, R600::OpName::src0_Z),
2022       TII->getOperandIdx(Opcode, R600::OpName::src0_W),
2023       TII->getOperandIdx(Opcode, R600::OpName::src1_X),
2024       TII->getOperandIdx(Opcode, R600::OpName::src1_Y),
2025       TII->getOperandIdx(Opcode, R600::OpName::src1_Z),
2026       TII->getOperandIdx(Opcode, R600::OpName::src1_W)
2027     };
2028     std::vector<unsigned> Consts;
2029     for (int OtherSrcIdx : SrcIndices) {
2030       int OtherSelIdx = TII->getSelIdx(Opcode, OtherSrcIdx);
2031       if (OtherSrcIdx < 0 || OtherSelIdx < 0)
2032         continue;
2033       if (HasDst) {
2034         OtherSrcIdx--;
2035         OtherSelIdx--;
2036       }
2037       if (RegisterSDNode *Reg =
2038           dyn_cast<RegisterSDNode>(ParentNode->getOperand(OtherSrcIdx))) {
2039         if (Reg->getReg() == R600::ALU_CONST) {
2040           ConstantSDNode *Cst
2041             = cast<ConstantSDNode>(ParentNode->getOperand(OtherSelIdx));
2042           Consts.push_back(Cst->getZExtValue());
2043         }
2044       }
2045     }
2046 
2047     ConstantSDNode *Cst = cast<ConstantSDNode>(CstOffset);
2048     Consts.push_back(Cst->getZExtValue());
2049     if (!TII->fitsConstReadLimitations(Consts)) {
2050       return false;
2051     }
2052 
2053     Sel = CstOffset;
2054     Src = DAG.getRegister(R600::ALU_CONST, MVT::f32);
2055     return true;
2056   }
2057   case R600::MOV_IMM_GLOBAL_ADDR:
2058     // Check if the Imm slot is used. Taken from below.
2059     if (cast<ConstantSDNode>(Imm)->getZExtValue())
2060       return false;
2061     Imm = Src.getOperand(0);
2062     Src = DAG.getRegister(R600::ALU_LITERAL_X, MVT::i32);
2063     return true;
2064   case R600::MOV_IMM_I32:
2065   case R600::MOV_IMM_F32: {
2066     unsigned ImmReg = R600::ALU_LITERAL_X;
2067     uint64_t ImmValue = 0;
2068 
2069     if (Src.getMachineOpcode() == R600::MOV_IMM_F32) {
2070       ConstantFPSDNode *FPC = cast<ConstantFPSDNode>(Src.getOperand(0));
2071       float FloatValue = FPC->getValueAPF().convertToFloat();
2072       if (FloatValue == 0.0) {
2073         ImmReg = R600::ZERO;
2074       } else if (FloatValue == 0.5) {
2075         ImmReg = R600::HALF;
2076       } else if (FloatValue == 1.0) {
2077         ImmReg = R600::ONE;
2078       } else {
2079         ImmValue = FPC->getValueAPF().bitcastToAPInt().getZExtValue();
2080       }
2081     } else {
2082       ConstantSDNode *C = cast<ConstantSDNode>(Src.getOperand(0));
2083       uint64_t Value = C->getZExtValue();
2084       if (Value == 0) {
2085         ImmReg = R600::ZERO;
2086       } else if (Value == 1) {
2087         ImmReg = R600::ONE_INT;
2088       } else {
2089         ImmValue = Value;
2090       }
2091     }
2092 
2093     // Check that we aren't already using an immediate.
2094     // XXX: It's possible for an instruction to have more than one
2095     // immediate operand, but this is not supported yet.
2096     if (ImmReg == R600::ALU_LITERAL_X) {
2097       if (!Imm.getNode())
2098         return false;
2099       ConstantSDNode *C = cast<ConstantSDNode>(Imm);
2100       if (C->getZExtValue())
2101         return false;
2102       Imm = DAG.getTargetConstant(ImmValue, SDLoc(ParentNode), MVT::i32);
2103     }
2104     Src = DAG.getRegister(ImmReg, MVT::i32);
2105     return true;
2106   }
2107   default:
2108     return false;
2109   }
2110 }
2111 
2112 /// Fold the instructions after selecting them
2113 SDNode *R600TargetLowering::PostISelFolding(MachineSDNode *Node,
2114                                             SelectionDAG &DAG) const {
2115   const R600InstrInfo *TII = Subtarget->getInstrInfo();
2116   if (!Node->isMachineOpcode())
2117     return Node;
2118 
2119   unsigned Opcode = Node->getMachineOpcode();
2120   SDValue FakeOp;
2121 
2122   std::vector<SDValue> Ops(Node->op_begin(), Node->op_end());
2123 
2124   if (Opcode == R600::DOT_4) {
2125     int OperandIdx[] = {
2126       TII->getOperandIdx(Opcode, R600::OpName::src0_X),
2127       TII->getOperandIdx(Opcode, R600::OpName::src0_Y),
2128       TII->getOperandIdx(Opcode, R600::OpName::src0_Z),
2129       TII->getOperandIdx(Opcode, R600::OpName::src0_W),
2130       TII->getOperandIdx(Opcode, R600::OpName::src1_X),
2131       TII->getOperandIdx(Opcode, R600::OpName::src1_Y),
2132       TII->getOperandIdx(Opcode, R600::OpName::src1_Z),
2133       TII->getOperandIdx(Opcode, R600::OpName::src1_W)
2134         };
2135     int NegIdx[] = {
2136       TII->getOperandIdx(Opcode, R600::OpName::src0_neg_X),
2137       TII->getOperandIdx(Opcode, R600::OpName::src0_neg_Y),
2138       TII->getOperandIdx(Opcode, R600::OpName::src0_neg_Z),
2139       TII->getOperandIdx(Opcode, R600::OpName::src0_neg_W),
2140       TII->getOperandIdx(Opcode, R600::OpName::src1_neg_X),
2141       TII->getOperandIdx(Opcode, R600::OpName::src1_neg_Y),
2142       TII->getOperandIdx(Opcode, R600::OpName::src1_neg_Z),
2143       TII->getOperandIdx(Opcode, R600::OpName::src1_neg_W)
2144     };
2145     int AbsIdx[] = {
2146       TII->getOperandIdx(Opcode, R600::OpName::src0_abs_X),
2147       TII->getOperandIdx(Opcode, R600::OpName::src0_abs_Y),
2148       TII->getOperandIdx(Opcode, R600::OpName::src0_abs_Z),
2149       TII->getOperandIdx(Opcode, R600::OpName::src0_abs_W),
2150       TII->getOperandIdx(Opcode, R600::OpName::src1_abs_X),
2151       TII->getOperandIdx(Opcode, R600::OpName::src1_abs_Y),
2152       TII->getOperandIdx(Opcode, R600::OpName::src1_abs_Z),
2153       TII->getOperandIdx(Opcode, R600::OpName::src1_abs_W)
2154     };
2155     for (unsigned i = 0; i < 8; i++) {
2156       if (OperandIdx[i] < 0)
2157         return Node;
2158       SDValue &Src = Ops[OperandIdx[i] - 1];
2159       SDValue &Neg = Ops[NegIdx[i] - 1];
2160       SDValue &Abs = Ops[AbsIdx[i] - 1];
2161       bool HasDst = TII->getOperandIdx(Opcode, R600::OpName::dst) > -1;
2162       int SelIdx = TII->getSelIdx(Opcode, OperandIdx[i]);
2163       if (HasDst)
2164         SelIdx--;
2165       SDValue &Sel = (SelIdx > -1) ? Ops[SelIdx] : FakeOp;
2166       if (FoldOperand(Node, i, Src, Neg, Abs, Sel, FakeOp, DAG))
2167         return DAG.getMachineNode(Opcode, SDLoc(Node), Node->getVTList(), Ops);
2168     }
2169   } else if (Opcode == R600::REG_SEQUENCE) {
2170     for (unsigned i = 1, e = Node->getNumOperands(); i < e; i += 2) {
2171       SDValue &Src = Ops[i];
2172       if (FoldOperand(Node, i, Src, FakeOp, FakeOp, FakeOp, FakeOp, DAG))
2173         return DAG.getMachineNode(Opcode, SDLoc(Node), Node->getVTList(), Ops);
2174     }
2175   } else {
2176     if (!TII->hasInstrModifiers(Opcode))
2177       return Node;
2178     int OperandIdx[] = {
2179       TII->getOperandIdx(Opcode, R600::OpName::src0),
2180       TII->getOperandIdx(Opcode, R600::OpName::src1),
2181       TII->getOperandIdx(Opcode, R600::OpName::src2)
2182     };
2183     int NegIdx[] = {
2184       TII->getOperandIdx(Opcode, R600::OpName::src0_neg),
2185       TII->getOperandIdx(Opcode, R600::OpName::src1_neg),
2186       TII->getOperandIdx(Opcode, R600::OpName::src2_neg)
2187     };
2188     int AbsIdx[] = {
2189       TII->getOperandIdx(Opcode, R600::OpName::src0_abs),
2190       TII->getOperandIdx(Opcode, R600::OpName::src1_abs),
2191       -1
2192     };
2193     for (unsigned i = 0; i < 3; i++) {
2194       if (OperandIdx[i] < 0)
2195         return Node;
2196       SDValue &Src = Ops[OperandIdx[i] - 1];
2197       SDValue &Neg = Ops[NegIdx[i] - 1];
2198       SDValue FakeAbs;
2199       SDValue &Abs = (AbsIdx[i] > -1) ? Ops[AbsIdx[i] - 1] : FakeAbs;
2200       bool HasDst = TII->getOperandIdx(Opcode, R600::OpName::dst) > -1;
2201       int SelIdx = TII->getSelIdx(Opcode, OperandIdx[i]);
2202       int ImmIdx = TII->getOperandIdx(Opcode, R600::OpName::literal);
2203       if (HasDst) {
2204         SelIdx--;
2205         ImmIdx--;
2206       }
2207       SDValue &Sel = (SelIdx > -1) ? Ops[SelIdx] : FakeOp;
2208       SDValue &Imm = Ops[ImmIdx];
2209       if (FoldOperand(Node, i, Src, Neg, Abs, Sel, Imm, DAG))
2210         return DAG.getMachineNode(Opcode, SDLoc(Node), Node->getVTList(), Ops);
2211     }
2212   }
2213 
2214   return Node;
2215 }
2216