AMDGPUISelLowering.cpp revision 286684 
1//===-- AMDGPUISelLowering.cpp - AMDGPU Common DAG lowering functions -----===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10/// \file
11/// \brief This is the parent TargetLowering class for hardware code gen
12/// targets.
13//
14//===----------------------------------------------------------------------===//
15
16#include "AMDGPUISelLowering.h"
17#include "AMDGPU.h"
18#include "AMDGPUFrameLowering.h"
19#include "AMDGPUIntrinsicInfo.h"
20#include "AMDGPURegisterInfo.h"
21#include "AMDGPUSubtarget.h"
22#include "R600MachineFunctionInfo.h"
23#include "SIMachineFunctionInfo.h"
24#include "llvm/CodeGen/CallingConvLower.h"
25#include "llvm/CodeGen/MachineFunction.h"
26#include "llvm/CodeGen/MachineRegisterInfo.h"
27#include "llvm/CodeGen/SelectionDAG.h"
28#include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
29#include "llvm/IR/DataLayout.h"
30#include "llvm/IR/DiagnosticInfo.h"
31#include "llvm/IR/DiagnosticPrinter.h"
32
33using namespace llvm;
34
35namespace {
36
37/// Diagnostic information for unimplemented or unsupported feature reporting.
38class DiagnosticInfoUnsupported : public DiagnosticInfo {
39private:
40  const Twine &Description;
41  const Function &Fn;
42
43  static int KindID;
44
45  static int getKindID() {
46    if (KindID == 0)
47      KindID = llvm::getNextAvailablePluginDiagnosticKind();
48    return KindID;
49  }
50
51public:
52  DiagnosticInfoUnsupported(const Function &Fn, const Twine &Desc,
53                          DiagnosticSeverity Severity = DS_Error)
54    : DiagnosticInfo(getKindID(), Severity),
55      Description(Desc),
56      Fn(Fn) { }
57
58  const Function &getFunction() const { return Fn; }
59  const Twine &getDescription() const { return Description; }
60
61  void print(DiagnosticPrinter &DP) const override {
62    DP << "unsupported " << getDescription() << " in " << Fn.getName();
63  }
64
65  static bool classof(const DiagnosticInfo *DI) {
66    return DI->getKind() == getKindID();
67  }
68};
69
70int DiagnosticInfoUnsupported::KindID = 0;
71}
72
73
74static bool allocateStack(unsigned ValNo, MVT ValVT, MVT LocVT,
75                      CCValAssign::LocInfo LocInfo,
76                      ISD::ArgFlagsTy ArgFlags, CCState &State) {
77  unsigned Offset = State.AllocateStack(ValVT.getStoreSize(),
78                                        ArgFlags.getOrigAlign());
79  State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
80
81  return true;
82}
83
84#include "AMDGPUGenCallingConv.inc"
85
86// Find a larger type to do a load / store of a vector with.
87EVT AMDGPUTargetLowering::getEquivalentMemType(LLVMContext &Ctx, EVT VT) {
88  unsigned StoreSize = VT.getStoreSizeInBits();
89  if (StoreSize <= 32)
90    return EVT::getIntegerVT(Ctx, StoreSize);
91
92  assert(StoreSize % 32 == 0 && "Store size not a multiple of 32");
93  return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
94}
95
96// Type for a vector that will be loaded to.
97EVT AMDGPUTargetLowering::getEquivalentLoadRegType(LLVMContext &Ctx, EVT VT) {
98  unsigned StoreSize = VT.getStoreSizeInBits();
99  if (StoreSize <= 32)
100    return EVT::getIntegerVT(Ctx, 32);
101
102  return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
103}
104
105AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM,
106                                           const AMDGPUSubtarget &STI)
107    : TargetLowering(TM), Subtarget(&STI) {
108  setOperationAction(ISD::Constant, MVT::i32, Legal);
109  setOperationAction(ISD::Constant, MVT::i64, Legal);
110  setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
111  setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
112
113  setOperationAction(ISD::BR_JT, MVT::Other, Expand);
114  setOperationAction(ISD::BRIND, MVT::Other, Expand);
115
116  // We need to custom lower some of the intrinsics
117  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
118
119  // Library functions.  These default to Expand, but we have instructions
120  // for them.
121  setOperationAction(ISD::FCEIL,  MVT::f32, Legal);
122  setOperationAction(ISD::FEXP2,  MVT::f32, Legal);
123  setOperationAction(ISD::FPOW,   MVT::f32, Legal);
124  setOperationAction(ISD::FLOG2,  MVT::f32, Legal);
125  setOperationAction(ISD::FABS,   MVT::f32, Legal);
126  setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
127  setOperationAction(ISD::FRINT,  MVT::f32, Legal);
128  setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
129  setOperationAction(ISD::FMINNUM, MVT::f32, Legal);
130  setOperationAction(ISD::FMAXNUM, MVT::f32, Legal);
131
132  setOperationAction(ISD::FROUND, MVT::f32, Custom);
133  setOperationAction(ISD::FROUND, MVT::f64, Custom);
134
135  setOperationAction(ISD::FREM, MVT::f32, Custom);
136  setOperationAction(ISD::FREM, MVT::f64, Custom);
137
138  // v_mad_f32 does not support denormals according to some sources.
139  if (!Subtarget->hasFP32Denormals())
140    setOperationAction(ISD::FMAD, MVT::f32, Legal);
141
142  // Expand to fneg + fadd.
143  setOperationAction(ISD::FSUB, MVT::f64, Expand);
144
145  // Lower floating point store/load to integer store/load to reduce the number
146  // of patterns in tablegen.
147  setOperationAction(ISD::STORE, MVT::f32, Promote);
148  AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
149
150  setOperationAction(ISD::STORE, MVT::v2f32, Promote);
151  AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
152
153  setOperationAction(ISD::STORE, MVT::v4f32, Promote);
154  AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
155
156  setOperationAction(ISD::STORE, MVT::v8f32, Promote);
157  AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
158
159  setOperationAction(ISD::STORE, MVT::v16f32, Promote);
160  AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
161
162  setOperationAction(ISD::STORE, MVT::f64, Promote);
163  AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
164
165  setOperationAction(ISD::STORE, MVT::v2f64, Promote);
166  AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v2i64);
167
168  // Custom lowering of vector stores is required for local address space
169  // stores.
170  setOperationAction(ISD::STORE, MVT::v4i32, Custom);
171
172  setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
173  setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
174  setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
175
176  // XXX: This can be change to Custom, once ExpandVectorStores can
177  // handle 64-bit stores.
178  setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
179
180  setTruncStoreAction(MVT::i64, MVT::i16, Expand);
181  setTruncStoreAction(MVT::i64, MVT::i8, Expand);
182  setTruncStoreAction(MVT::i64, MVT::i1, Expand);
183  setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
184  setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
185
186
187  setOperationAction(ISD::LOAD, MVT::f32, Promote);
188  AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
189
190  setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
191  AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
192
193  setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
194  AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
195
196  setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
197  AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
198
199  setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
200  AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
201
202  setOperationAction(ISD::LOAD, MVT::f64, Promote);
203  AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
204
205  setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
206  AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v2i64);
207
208  setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
209  setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
210  setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
211  setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
212  setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
213  setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
214  setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
215  setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
216  setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
217  setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
218
219  // There are no 64-bit extloads. These should be done as a 32-bit extload and
220  // an extension to 64-bit.
221  for (MVT VT : MVT::integer_valuetypes()) {
222    setLoadExtAction(ISD::EXTLOAD, MVT::i64, VT, Expand);
223    setLoadExtAction(ISD::SEXTLOAD, MVT::i64, VT, Expand);
224    setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, VT, Expand);
225  }
226
227  for (MVT VT : MVT::integer_vector_valuetypes()) {
228    setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Expand);
229    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Expand);
230    setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i8, Expand);
231    setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Expand);
232    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Expand);
233    setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i8, Expand);
234    setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Expand);
235    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Expand);
236    setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i16, Expand);
237    setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Expand);
238    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Expand);
239    setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i16, Expand);
240  }
241
242  setOperationAction(ISD::BR_CC, MVT::i1, Expand);
243
244  if (Subtarget->getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
245    setOperationAction(ISD::FCEIL, MVT::f64, Custom);
246    setOperationAction(ISD::FTRUNC, MVT::f64, Custom);
247    setOperationAction(ISD::FRINT, MVT::f64, Custom);
248    setOperationAction(ISD::FFLOOR, MVT::f64, Custom);
249  }
250
251  if (!Subtarget->hasBFI()) {
252    // fcopysign can be done in a single instruction with BFI.
253    setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
254    setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
255  }
256
257  setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
258
259  setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
260  setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, MVT::v2f16, Expand);
261  setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, MVT::v4f16, Expand);
262  setLoadExtAction(ISD::EXTLOAD, MVT::v8f32, MVT::v8f16, Expand);
263
264  setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
265  setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f16, Expand);
266  setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f16, Expand);
267  setLoadExtAction(ISD::EXTLOAD, MVT::v8f64, MVT::v8f16, Expand);
268
269  setTruncStoreAction(MVT::f32, MVT::f16, Expand);
270  setTruncStoreAction(MVT::v2f32, MVT::v2f16, Expand);
271  setTruncStoreAction(MVT::v4f32, MVT::v4f16, Expand);
272  setTruncStoreAction(MVT::v8f32, MVT::v8f16, Expand);
273
274  setTruncStoreAction(MVT::f64, MVT::f16, Expand);
275  setTruncStoreAction(MVT::f64, MVT::f32, Expand);
276
277  const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
278  for (MVT VT : ScalarIntVTs) {
279    setOperationAction(ISD::SREM, VT, Expand);
280    setOperationAction(ISD::SDIV, VT, Expand);
281
282    // GPU does not have divrem function for signed or unsigned.
283    setOperationAction(ISD::SDIVREM, VT, Custom);
284    setOperationAction(ISD::UDIVREM, VT, Custom);
285
286    // GPU does not have [S|U]MUL_LOHI functions as a single instruction.
287    setOperationAction(ISD::SMUL_LOHI, VT, Expand);
288    setOperationAction(ISD::UMUL_LOHI, VT, Expand);
289
290    setOperationAction(ISD::BSWAP, VT, Expand);
291    setOperationAction(ISD::CTTZ, VT, Expand);
292    setOperationAction(ISD::CTLZ, VT, Expand);
293  }
294
295  if (!Subtarget->hasBCNT(32))
296    setOperationAction(ISD::CTPOP, MVT::i32, Expand);
297
298  if (!Subtarget->hasBCNT(64))
299    setOperationAction(ISD::CTPOP, MVT::i64, Expand);
300
301  // The hardware supports 32-bit ROTR, but not ROTL.
302  setOperationAction(ISD::ROTL, MVT::i32, Expand);
303  setOperationAction(ISD::ROTL, MVT::i64, Expand);
304  setOperationAction(ISD::ROTR, MVT::i64, Expand);
305
306  setOperationAction(ISD::MUL, MVT::i64, Expand);
307  setOperationAction(ISD::MULHU, MVT::i64, Expand);
308  setOperationAction(ISD::MULHS, MVT::i64, Expand);
309  setOperationAction(ISD::UDIV, MVT::i32, Expand);
310  setOperationAction(ISD::UREM, MVT::i32, Expand);
311  setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
312  setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
313  setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
314  setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
315  setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
316
317  setOperationAction(ISD::SMIN, MVT::i32, Legal);
318  setOperationAction(ISD::UMIN, MVT::i32, Legal);
319  setOperationAction(ISD::SMAX, MVT::i32, Legal);
320  setOperationAction(ISD::UMAX, MVT::i32, Legal);
321
322  if (!Subtarget->hasFFBH())
323    setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
324
325  if (!Subtarget->hasFFBL())
326    setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
327
328  static const MVT::SimpleValueType VectorIntTypes[] = {
329    MVT::v2i32, MVT::v4i32
330  };
331
332  for (MVT VT : VectorIntTypes) {
333    // Expand the following operations for the current type by default.
334    setOperationAction(ISD::ADD,  VT, Expand);
335    setOperationAction(ISD::AND,  VT, Expand);
336    setOperationAction(ISD::FP_TO_SINT, VT, Expand);
337    setOperationAction(ISD::FP_TO_UINT, VT, Expand);
338    setOperationAction(ISD::MUL,  VT, Expand);
339    setOperationAction(ISD::OR,   VT, Expand);
340    setOperationAction(ISD::SHL,  VT, Expand);
341    setOperationAction(ISD::SRA,  VT, Expand);
342    setOperationAction(ISD::SRL,  VT, Expand);
343    setOperationAction(ISD::ROTL, VT, Expand);
344    setOperationAction(ISD::ROTR, VT, Expand);
345    setOperationAction(ISD::SUB,  VT, Expand);
346    setOperationAction(ISD::SINT_TO_FP, VT, Expand);
347    setOperationAction(ISD::UINT_TO_FP, VT, Expand);
348    setOperationAction(ISD::SDIV, VT, Expand);
349    setOperationAction(ISD::UDIV, VT, Expand);
350    setOperationAction(ISD::SREM, VT, Expand);
351    setOperationAction(ISD::UREM, VT, Expand);
352    setOperationAction(ISD::SMUL_LOHI, VT, Expand);
353    setOperationAction(ISD::UMUL_LOHI, VT, Expand);
354    setOperationAction(ISD::SDIVREM, VT, Custom);
355    setOperationAction(ISD::UDIVREM, VT, Custom);
356    setOperationAction(ISD::ADDC, VT, Expand);
357    setOperationAction(ISD::SUBC, VT, Expand);
358    setOperationAction(ISD::ADDE, VT, Expand);
359    setOperationAction(ISD::SUBE, VT, Expand);
360    setOperationAction(ISD::SELECT, VT, Expand);
361    setOperationAction(ISD::VSELECT, VT, Expand);
362    setOperationAction(ISD::SELECT_CC, VT, Expand);
363    setOperationAction(ISD::XOR,  VT, Expand);
364    setOperationAction(ISD::BSWAP, VT, Expand);
365    setOperationAction(ISD::CTPOP, VT, Expand);
366    setOperationAction(ISD::CTTZ, VT, Expand);
367    setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
368    setOperationAction(ISD::CTLZ, VT, Expand);
369    setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
370    setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
371  }
372
373  static const MVT::SimpleValueType FloatVectorTypes[] = {
374    MVT::v2f32, MVT::v4f32
375  };
376
377  for (MVT VT : FloatVectorTypes) {
378    setOperationAction(ISD::FABS, VT, Expand);
379    setOperationAction(ISD::FMINNUM, VT, Expand);
380    setOperationAction(ISD::FMAXNUM, VT, Expand);
381    setOperationAction(ISD::FADD, VT, Expand);
382    setOperationAction(ISD::FCEIL, VT, Expand);
383    setOperationAction(ISD::FCOS, VT, Expand);
384    setOperationAction(ISD::FDIV, VT, Expand);
385    setOperationAction(ISD::FEXP2, VT, Expand);
386    setOperationAction(ISD::FLOG2, VT, Expand);
387    setOperationAction(ISD::FREM, VT, Expand);
388    setOperationAction(ISD::FPOW, VT, Expand);
389    setOperationAction(ISD::FFLOOR, VT, Expand);
390    setOperationAction(ISD::FTRUNC, VT, Expand);
391    setOperationAction(ISD::FMUL, VT, Expand);
392    setOperationAction(ISD::FMA, VT, Expand);
393    setOperationAction(ISD::FRINT, VT, Expand);
394    setOperationAction(ISD::FNEARBYINT, VT, Expand);
395    setOperationAction(ISD::FSQRT, VT, Expand);
396    setOperationAction(ISD::FSIN, VT, Expand);
397    setOperationAction(ISD::FSUB, VT, Expand);
398    setOperationAction(ISD::FNEG, VT, Expand);
399    setOperationAction(ISD::SELECT, VT, Expand);
400    setOperationAction(ISD::VSELECT, VT, Expand);
401    setOperationAction(ISD::SELECT_CC, VT, Expand);
402    setOperationAction(ISD::FCOPYSIGN, VT, Expand);
403    setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
404  }
405
406  setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom);
407  setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
408
409  setTargetDAGCombine(ISD::SHL);
410  setTargetDAGCombine(ISD::MUL);
411  setTargetDAGCombine(ISD::SELECT);
412  setTargetDAGCombine(ISD::SELECT_CC);
413  setTargetDAGCombine(ISD::STORE);
414
415  setTargetDAGCombine(ISD::FADD);
416  setTargetDAGCombine(ISD::FSUB);
417
418  setBooleanContents(ZeroOrNegativeOneBooleanContent);
419  setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
420
421  setSchedulingPreference(Sched::RegPressure);
422  setJumpIsExpensive(true);
423
424  // SI at least has hardware support for floating point exceptions, but no way
425  // of using or handling them is implemented. They are also optional in OpenCL
426  // (Section 7.3)
427  setHasFloatingPointExceptions(false);
428
429  setSelectIsExpensive(false);
430  PredictableSelectIsExpensive = false;
431
432  // There are no integer divide instructions, and these expand to a pretty
433  // large sequence of instructions.
434  setIntDivIsCheap(false);
435  setPow2SDivIsCheap(false);
436  setFsqrtIsCheap(true);
437
438  // FIXME: Need to really handle these.
439  MaxStoresPerMemcpy  = 4096;
440  MaxStoresPerMemmove = 4096;
441  MaxStoresPerMemset  = 4096;
442}
443
444//===----------------------------------------------------------------------===//
445// Target Information
446//===----------------------------------------------------------------------===//
447
448MVT AMDGPUTargetLowering::getVectorIdxTy(const DataLayout &) const {
449  return MVT::i32;
450}
451
452bool AMDGPUTargetLowering::isSelectSupported(SelectSupportKind SelType) const {
453  return true;
454}
455
456// The backend supports 32 and 64 bit floating point immediates.
457// FIXME: Why are we reporting vectors of FP immediates as legal?
458bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
459  EVT ScalarVT = VT.getScalarType();
460  return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64);
461}
462
463// We don't want to shrink f64 / f32 constants.
464bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
465  EVT ScalarVT = VT.getScalarType();
466  return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
467}
468
469bool AMDGPUTargetLowering::shouldReduceLoadWidth(SDNode *N,
470                                                 ISD::LoadExtType,
471                                                 EVT NewVT) const {
472
473  unsigned NewSize = NewVT.getStoreSizeInBits();
474
475  // If we are reducing to a 32-bit load, this is always better.
476  if (NewSize == 32)
477    return true;
478
479  EVT OldVT = N->getValueType(0);
480  unsigned OldSize = OldVT.getStoreSizeInBits();
481
482  // Don't produce extloads from sub 32-bit types. SI doesn't have scalar
483  // extloads, so doing one requires using a buffer_load. In cases where we
484  // still couldn't use a scalar load, using the wider load shouldn't really
485  // hurt anything.
486
487  // If the old size already had to be an extload, there's no harm in continuing
488  // to reduce the width.
489  return (OldSize < 32);
490}
491
492bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
493                                                   EVT CastTy) const {
494  if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
495    return true;
496
497  unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
498  unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
499
500  return ((LScalarSize <= CastScalarSize) ||
501          (CastScalarSize >= 32) ||
502          (LScalarSize < 32));
503}
504
505// SI+ has instructions for cttz / ctlz for 32-bit values. This is probably also
506// profitable with the expansion for 64-bit since it's generally good to
507// speculate things.
508// FIXME: These should really have the size as a parameter.
509bool AMDGPUTargetLowering::isCheapToSpeculateCttz() const {
510  return true;
511}
512
513bool AMDGPUTargetLowering::isCheapToSpeculateCtlz() const {
514  return true;
515}
516
517//===---------------------------------------------------------------------===//
518// Target Properties
519//===---------------------------------------------------------------------===//
520
521bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
522  assert(VT.isFloatingPoint());
523  return VT == MVT::f32 || VT == MVT::f64;
524}
525
526bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
527  assert(VT.isFloatingPoint());
528  return VT == MVT::f32 || VT == MVT::f64;
529}
530
531bool AMDGPUTargetLowering:: storeOfVectorConstantIsCheap(EVT MemVT,
532                                                         unsigned NumElem,
533                                                         unsigned AS) const {
534  return true;
535}
536
537bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
538  // Truncate is just accessing a subregister.
539  return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
540}
541
542bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
543  // Truncate is just accessing a subregister.
544  return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
545         (Dest->getPrimitiveSizeInBits() % 32 == 0);
546}
547
548bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
549  unsigned SrcSize = Src->getScalarSizeInBits();
550  unsigned DestSize = Dest->getScalarSizeInBits();
551
552  return SrcSize == 32 && DestSize == 64;
553}
554
555bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
556  // Any register load of a 64-bit value really requires 2 32-bit moves. For all
557  // practical purposes, the extra mov 0 to load a 64-bit is free.  As used,
558  // this will enable reducing 64-bit operations the 32-bit, which is always
559  // good.
560  return Src == MVT::i32 && Dest == MVT::i64;
561}
562
563bool AMDGPUTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
564  return isZExtFree(Val.getValueType(), VT2);
565}
566
567bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
568  // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
569  // limited number of native 64-bit operations. Shrinking an operation to fit
570  // in a single 32-bit register should always be helpful. As currently used,
571  // this is much less general than the name suggests, and is only used in
572  // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
573  // not profitable, and may actually be harmful.
574  return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
575}
576
577//===---------------------------------------------------------------------===//
578// TargetLowering Callbacks
579//===---------------------------------------------------------------------===//
580
581void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
582                             const SmallVectorImpl<ISD::InputArg> &Ins) const {
583
584  State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
585}
586
587SDValue AMDGPUTargetLowering::LowerReturn(
588                                     SDValue Chain,
589                                     CallingConv::ID CallConv,
590                                     bool isVarArg,
591                                     const SmallVectorImpl<ISD::OutputArg> &Outs,
592                                     const SmallVectorImpl<SDValue> &OutVals,
593                                     SDLoc DL, SelectionDAG &DAG) const {
594  return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
595}
596
597//===---------------------------------------------------------------------===//
598// Target specific lowering
599//===---------------------------------------------------------------------===//
600
601SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
602                                        SmallVectorImpl<SDValue> &InVals) const {
603  SDValue Callee = CLI.Callee;
604  SelectionDAG &DAG = CLI.DAG;
605
606  const Function &Fn = *DAG.getMachineFunction().getFunction();
607
608  StringRef FuncName("<unknown>");
609
610  if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
611    FuncName = G->getSymbol();
612  else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
613    FuncName = G->getGlobal()->getName();
614
615  DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
616  DAG.getContext()->diagnose(NoCalls);
617  return SDValue();
618}
619
620SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
621                                             SelectionDAG &DAG) const {
622  switch (Op.getOpcode()) {
623  default:
624    Op.getNode()->dump();
625    llvm_unreachable("Custom lowering code for this"
626                     "instruction is not implemented yet!");
627    break;
628  case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
629  case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
630  case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
631  case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
632  case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
633  case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
634  case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
635  case ISD::FREM: return LowerFREM(Op, DAG);
636  case ISD::FCEIL: return LowerFCEIL(Op, DAG);
637  case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
638  case ISD::FRINT: return LowerFRINT(Op, DAG);
639  case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
640  case ISD::FROUND: return LowerFROUND(Op, DAG);
641  case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
642  case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
643  case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
644  case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
645  case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
646  }
647  return Op;
648}
649
650void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
651                                              SmallVectorImpl<SDValue> &Results,
652                                              SelectionDAG &DAG) const {
653  switch (N->getOpcode()) {
654  case ISD::SIGN_EXTEND_INREG:
655    // Different parts of legalization seem to interpret which type of
656    // sign_extend_inreg is the one to check for custom lowering. The extended
657    // from type is what really matters, but some places check for custom
658    // lowering of the result type. This results in trying to use
659    // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
660    // nothing here and let the illegal result integer be handled normally.
661    return;
662  case ISD::LOAD: {
663    SDNode *Node = LowerLOAD(SDValue(N, 0), DAG).getNode();
664    if (!Node)
665      return;
666
667    Results.push_back(SDValue(Node, 0));
668    Results.push_back(SDValue(Node, 1));
669    // XXX: LLVM seems not to replace Chain Value inside CustomWidenLowerNode
670    // function
671    DAG.ReplaceAllUsesOfValueWith(SDValue(N,1), SDValue(Node, 1));
672    return;
673  }
674  case ISD::STORE: {
675    SDValue Lowered = LowerSTORE(SDValue(N, 0), DAG);
676    if (Lowered.getNode())
677      Results.push_back(Lowered);
678    return;
679  }
680  default:
681    return;
682  }
683}
684
685// FIXME: This implements accesses to initialized globals in the constant
686// address space by copying them to private and accessing that. It does not
687// properly handle illegal types or vectors. The private vector loads are not
688// scalarized, and the illegal scalars hit an assertion. This technique will not
689// work well with large initializers, and this should eventually be
690// removed. Initialized globals should be placed into a data section that the
691// runtime will load into a buffer before the kernel is executed. Uses of the
692// global need to be replaced with a pointer loaded from an implicit kernel
693// argument into this buffer holding the copy of the data, which will remove the
694// need for any of this.
695SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
696                                                       const GlobalValue *GV,
697                                                       const SDValue &InitPtr,
698                                                       SDValue Chain,
699                                                       SelectionDAG &DAG) const {
700  const DataLayout &TD = DAG.getDataLayout();
701  SDLoc DL(InitPtr);
702  Type *InitTy = Init->getType();
703
704  if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
705    EVT VT = EVT::getEVT(InitTy);
706    PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
707    return DAG.getStore(Chain, DL, DAG.getConstant(*CI, DL, VT), InitPtr,
708                        MachinePointerInfo(UndefValue::get(PtrTy)), false,
709                        false, TD.getPrefTypeAlignment(InitTy));
710  }
711
712  if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
713    EVT VT = EVT::getEVT(CFP->getType());
714    PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
715    return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, DL, VT), InitPtr,
716                        MachinePointerInfo(UndefValue::get(PtrTy)), false,
717                        false, TD.getPrefTypeAlignment(CFP->getType()));
718  }
719
720  if (StructType *ST = dyn_cast<StructType>(InitTy)) {
721    const StructLayout *SL = TD.getStructLayout(ST);
722
723    EVT PtrVT = InitPtr.getValueType();
724    SmallVector<SDValue, 8> Chains;
725
726    for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
727      SDValue Offset = DAG.getConstant(SL->getElementOffset(I), DL, PtrVT);
728      SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
729
730      Constant *Elt = Init->getAggregateElement(I);
731      Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
732    }
733
734    return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
735  }
736
737  if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
738    EVT PtrVT = InitPtr.getValueType();
739
740    unsigned NumElements;
741    if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
742      NumElements = AT->getNumElements();
743    else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
744      NumElements = VT->getNumElements();
745    else
746      llvm_unreachable("Unexpected type");
747
748    unsigned EltSize = TD.getTypeAllocSize(SeqTy->getElementType());
749    SmallVector<SDValue, 8> Chains;
750    for (unsigned i = 0; i < NumElements; ++i) {
751      SDValue Offset = DAG.getConstant(i * EltSize, DL, PtrVT);
752      SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
753
754      Constant *Elt = Init->getAggregateElement(i);
755      Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
756    }
757
758    return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
759  }
760
761  if (isa<UndefValue>(Init)) {
762    EVT VT = EVT::getEVT(InitTy);
763    PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
764    return DAG.getStore(Chain, DL, DAG.getUNDEF(VT), InitPtr,
765                        MachinePointerInfo(UndefValue::get(PtrTy)), false,
766                        false, TD.getPrefTypeAlignment(InitTy));
767  }
768
769  Init->dump();
770  llvm_unreachable("Unhandled constant initializer");
771}
772
773static bool hasDefinedInitializer(const GlobalValue *GV) {
774  const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
775  if (!GVar || !GVar->hasInitializer())
776    return false;
777
778  if (isa<UndefValue>(GVar->getInitializer()))
779    return false;
780
781  return true;
782}
783
784SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
785                                                 SDValue Op,
786                                                 SelectionDAG &DAG) const {
787
788  const DataLayout &DL = DAG.getDataLayout();
789  GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
790  const GlobalValue *GV = G->getGlobal();
791
792  switch (G->getAddressSpace()) {
793  case AMDGPUAS::LOCAL_ADDRESS: {
794    // XXX: What does the value of G->getOffset() mean?
795    assert(G->getOffset() == 0 &&
796         "Do not know what to do with an non-zero offset");
797
798    // TODO: We could emit code to handle the initialization somewhere.
799    if (hasDefinedInitializer(GV))
800      break;
801
802    unsigned Offset;
803    if (MFI->LocalMemoryObjects.count(GV) == 0) {
804      uint64_t Size = DL.getTypeAllocSize(GV->getType()->getElementType());
805      Offset = MFI->LDSSize;
806      MFI->LocalMemoryObjects[GV] = Offset;
807      // XXX: Account for alignment?
808      MFI->LDSSize += Size;
809    } else {
810      Offset = MFI->LocalMemoryObjects[GV];
811    }
812
813    return DAG.getConstant(Offset, SDLoc(Op),
814                           getPointerTy(DL, AMDGPUAS::LOCAL_ADDRESS));
815  }
816  case AMDGPUAS::CONSTANT_ADDRESS: {
817    MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
818    Type *EltType = GV->getType()->getElementType();
819    unsigned Size = DL.getTypeAllocSize(EltType);
820    unsigned Alignment = DL.getPrefTypeAlignment(EltType);
821
822    MVT PrivPtrVT = getPointerTy(DL, AMDGPUAS::PRIVATE_ADDRESS);
823    MVT ConstPtrVT = getPointerTy(DL, AMDGPUAS::CONSTANT_ADDRESS);
824
825    int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
826    SDValue InitPtr = DAG.getFrameIndex(FI, PrivPtrVT);
827
828    const GlobalVariable *Var = cast<GlobalVariable>(GV);
829    if (!Var->hasInitializer()) {
830      // This has no use, but bugpoint will hit it.
831      return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
832    }
833
834    const Constant *Init = Var->getInitializer();
835    SmallVector<SDNode*, 8> WorkList;
836
837    for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
838                              E = DAG.getEntryNode()->use_end(); I != E; ++I) {
839      if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
840        continue;
841      WorkList.push_back(*I);
842    }
843    SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
844    for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
845                                           E = WorkList.end(); I != E; ++I) {
846      SmallVector<SDValue, 8> Ops;
847      Ops.push_back(Chain);
848      for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
849        Ops.push_back((*I)->getOperand(i));
850      }
851      DAG.UpdateNodeOperands(*I, Ops);
852    }
853    return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
854  }
855  }
856
857  const Function &Fn = *DAG.getMachineFunction().getFunction();
858  DiagnosticInfoUnsupported BadInit(Fn,
859                                    "initializer for address space");
860  DAG.getContext()->diagnose(BadInit);
861  return SDValue();
862}
863
864SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
865                                                  SelectionDAG &DAG) const {
866  SmallVector<SDValue, 8> Args;
867
868  for (const SDUse &U : Op->ops())
869    DAG.ExtractVectorElements(U.get(), Args);
870
871  return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
872}
873
874SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
875                                                     SelectionDAG &DAG) const {
876
877  SmallVector<SDValue, 8> Args;
878  unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
879  EVT VT = Op.getValueType();
880  DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
881                            VT.getVectorNumElements());
882
883  return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
884}
885
886SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
887                                              SelectionDAG &DAG) const {
888
889  MachineFunction &MF = DAG.getMachineFunction();
890  const AMDGPUFrameLowering *TFL = Subtarget->getFrameLowering();
891
892  FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
893
894  unsigned FrameIndex = FIN->getIndex();
895  unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
896  return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF), SDLoc(Op),
897                         Op.getValueType());
898}
899
900SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
901    SelectionDAG &DAG) const {
902  unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
903  SDLoc DL(Op);
904  EVT VT = Op.getValueType();
905
906  switch (IntrinsicID) {
907    default: return Op;
908    case AMDGPUIntrinsic::AMDGPU_abs:
909    case AMDGPUIntrinsic::AMDIL_abs: // Legacy name.
910      return LowerIntrinsicIABS(Op, DAG);
911    case AMDGPUIntrinsic::AMDGPU_lrp:
912      return LowerIntrinsicLRP(Op, DAG);
913
914    case AMDGPUIntrinsic::AMDGPU_clamp:
915    case AMDGPUIntrinsic::AMDIL_clamp: // Legacy name.
916      return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
917                         Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
918
919    case Intrinsic::AMDGPU_div_scale: {
920      // 3rd parameter required to be a constant.
921      const ConstantSDNode *Param = dyn_cast<ConstantSDNode>(Op.getOperand(3));
922      if (!Param)
923        return DAG.getUNDEF(VT);
924
925      // Translate to the operands expected by the machine instruction. The
926      // first parameter must be the same as the first instruction.
927      SDValue Numerator = Op.getOperand(1);
928      SDValue Denominator = Op.getOperand(2);
929
930      // Note this order is opposite of the machine instruction's operations,
931      // which is s0.f = Quotient, s1.f = Denominator, s2.f = Numerator. The
932      // intrinsic has the numerator as the first operand to match a normal
933      // division operation.
934
935      SDValue Src0 = Param->isAllOnesValue() ? Numerator : Denominator;
936
937      return DAG.getNode(AMDGPUISD::DIV_SCALE, DL, Op->getVTList(), Src0,
938                         Denominator, Numerator);
939    }
940
941    case Intrinsic::AMDGPU_div_fmas:
942      return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT,
943                         Op.getOperand(1), Op.getOperand(2), Op.getOperand(3),
944                         Op.getOperand(4));
945
946    case Intrinsic::AMDGPU_div_fixup:
947      return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT,
948                         Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
949
950    case Intrinsic::AMDGPU_trig_preop:
951      return DAG.getNode(AMDGPUISD::TRIG_PREOP, DL, VT,
952                         Op.getOperand(1), Op.getOperand(2));
953
954    case Intrinsic::AMDGPU_rcp:
955      return DAG.getNode(AMDGPUISD::RCP, DL, VT, Op.getOperand(1));
956
957    case Intrinsic::AMDGPU_rsq:
958      return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
959
960    case AMDGPUIntrinsic::AMDGPU_legacy_rsq:
961      return DAG.getNode(AMDGPUISD::RSQ_LEGACY, DL, VT, Op.getOperand(1));
962
963    case Intrinsic::AMDGPU_rsq_clamped:
964      if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
965        Type *Type = VT.getTypeForEVT(*DAG.getContext());
966        APFloat Max = APFloat::getLargest(Type->getFltSemantics());
967        APFloat Min = APFloat::getLargest(Type->getFltSemantics(), true);
968
969        SDValue Rsq = DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
970        SDValue Tmp = DAG.getNode(ISD::FMINNUM, DL, VT, Rsq,
971                                  DAG.getConstantFP(Max, DL, VT));
972        return DAG.getNode(ISD::FMAXNUM, DL, VT, Tmp,
973                           DAG.getConstantFP(Min, DL, VT));
974      } else {
975        return DAG.getNode(AMDGPUISD::RSQ_CLAMPED, DL, VT, Op.getOperand(1));
976      }
977
978    case Intrinsic::AMDGPU_ldexp:
979      return DAG.getNode(AMDGPUISD::LDEXP, DL, VT, Op.getOperand(1),
980                                                   Op.getOperand(2));
981
982    case AMDGPUIntrinsic::AMDGPU_imax:
983      return DAG.getNode(ISD::SMAX, DL, VT, Op.getOperand(1),
984                                            Op.getOperand(2));
985    case AMDGPUIntrinsic::AMDGPU_umax:
986      return DAG.getNode(ISD::UMAX, DL, VT, Op.getOperand(1),
987                                            Op.getOperand(2));
988    case AMDGPUIntrinsic::AMDGPU_imin:
989      return DAG.getNode(ISD::SMIN, DL, VT, Op.getOperand(1),
990                                            Op.getOperand(2));
991    case AMDGPUIntrinsic::AMDGPU_umin:
992      return DAG.getNode(ISD::UMIN, DL, VT, Op.getOperand(1),
993                                            Op.getOperand(2));
994
995    case AMDGPUIntrinsic::AMDGPU_umul24:
996      return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
997                         Op.getOperand(1), Op.getOperand(2));
998
999    case AMDGPUIntrinsic::AMDGPU_imul24:
1000      return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
1001                         Op.getOperand(1), Op.getOperand(2));
1002
1003    case AMDGPUIntrinsic::AMDGPU_umad24:
1004      return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
1005                         Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
1006
1007    case AMDGPUIntrinsic::AMDGPU_imad24:
1008      return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
1009                         Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
1010
1011    case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
1012      return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
1013
1014    case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
1015      return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
1016
1017    case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
1018      return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
1019
1020    case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
1021      return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
1022
1023    case AMDGPUIntrinsic::AMDGPU_bfe_i32:
1024      return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
1025                         Op.getOperand(1),
1026                         Op.getOperand(2),
1027                         Op.getOperand(3));
1028
1029    case AMDGPUIntrinsic::AMDGPU_bfe_u32:
1030      return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
1031                         Op.getOperand(1),
1032                         Op.getOperand(2),
1033                         Op.getOperand(3));
1034
1035    case AMDGPUIntrinsic::AMDGPU_bfi:
1036      return DAG.getNode(AMDGPUISD::BFI, DL, VT,
1037                         Op.getOperand(1),
1038                         Op.getOperand(2),
1039                         Op.getOperand(3));
1040
1041    case AMDGPUIntrinsic::AMDGPU_bfm:
1042      return DAG.getNode(AMDGPUISD::BFM, DL, VT,
1043                         Op.getOperand(1),
1044                         Op.getOperand(2));
1045
1046    case AMDGPUIntrinsic::AMDGPU_brev:
1047      return DAG.getNode(AMDGPUISD::BREV, DL, VT, Op.getOperand(1));
1048
1049  case Intrinsic::AMDGPU_class:
1050    return DAG.getNode(AMDGPUISD::FP_CLASS, DL, VT,
1051                       Op.getOperand(1), Op.getOperand(2));
1052
1053    case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
1054      return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
1055
1056    case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
1057      return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
1058    case AMDGPUIntrinsic::AMDGPU_trunc: // Legacy name.
1059      return DAG.getNode(ISD::FTRUNC, DL, VT, Op.getOperand(1));
1060  }
1061}
1062
1063///IABS(a) = SMAX(sub(0, a), a)
1064SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
1065                                                 SelectionDAG &DAG) const {
1066  SDLoc DL(Op);
1067  EVT VT = Op.getValueType();
1068  SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
1069                            Op.getOperand(1));
1070
1071  return DAG.getNode(ISD::SMAX, DL, VT, Neg, Op.getOperand(1));
1072}
1073
1074/// Linear Interpolation
1075/// LRP(a, b, c) = muladd(a,  b, (1 - a) * c)
1076SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
1077                                                SelectionDAG &DAG) const {
1078  SDLoc DL(Op);
1079  EVT VT = Op.getValueType();
1080  SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
1081                                DAG.getConstantFP(1.0f, DL, MVT::f32),
1082                                Op.getOperand(1));
1083  SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
1084                                                    Op.getOperand(3));
1085  return DAG.getNode(ISD::FADD, DL, VT,
1086      DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
1087      OneSubAC);
1088}
1089
1090/// \brief Generate Min/Max node
1091SDValue AMDGPUTargetLowering::CombineFMinMaxLegacy(SDLoc DL,
1092                                                   EVT VT,
1093                                                   SDValue LHS,
1094                                                   SDValue RHS,
1095                                                   SDValue True,
1096                                                   SDValue False,
1097                                                   SDValue CC,
1098                                                   DAGCombinerInfo &DCI) const {
1099  if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
1100    return SDValue();
1101
1102  if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
1103    return SDValue();
1104
1105  SelectionDAG &DAG = DCI.DAG;
1106  ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1107  switch (CCOpcode) {
1108  case ISD::SETOEQ:
1109  case ISD::SETONE:
1110  case ISD::SETUNE:
1111  case ISD::SETNE:
1112  case ISD::SETUEQ:
1113  case ISD::SETEQ:
1114  case ISD::SETFALSE:
1115  case ISD::SETFALSE2:
1116  case ISD::SETTRUE:
1117  case ISD::SETTRUE2:
1118  case ISD::SETUO:
1119  case ISD::SETO:
1120    break;
1121  case ISD::SETULE:
1122  case ISD::SETULT: {
1123    if (LHS == True)
1124      return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1125    return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1126  }
1127  case ISD::SETOLE:
1128  case ISD::SETOLT:
1129  case ISD::SETLE:
1130  case ISD::SETLT: {
1131    // Ordered. Assume ordered for undefined.
1132
1133    // Only do this after legalization to avoid interfering with other combines
1134    // which might occur.
1135    if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
1136        !DCI.isCalledByLegalizer())
1137      return SDValue();
1138
1139    // We need to permute the operands to get the correct NaN behavior. The
1140    // selected operand is the second one based on the failing compare with NaN,
1141    // so permute it based on the compare type the hardware uses.
1142    if (LHS == True)
1143      return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1144    return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1145  }
1146  case ISD::SETUGE:
1147  case ISD::SETUGT: {
1148    if (LHS == True)
1149      return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1150    return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1151  }
1152  case ISD::SETGT:
1153  case ISD::SETGE:
1154  case ISD::SETOGE:
1155  case ISD::SETOGT: {
1156    if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
1157        !DCI.isCalledByLegalizer())
1158      return SDValue();
1159
1160    if (LHS == True)
1161      return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1162    return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1163  }
1164  case ISD::SETCC_INVALID:
1165    llvm_unreachable("Invalid setcc condcode!");
1166  }
1167  return SDValue();
1168}
1169
1170// FIXME: Remove this when combines added to DAGCombiner.
1171SDValue AMDGPUTargetLowering::CombineIMinMax(SDLoc DL,
1172                                             EVT VT,
1173                                             SDValue LHS,
1174                                             SDValue RHS,
1175                                             SDValue True,
1176                                             SDValue False,
1177                                             SDValue CC,
1178                                             SelectionDAG &DAG) const {
1179  if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
1180    return SDValue();
1181
1182  ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1183  switch (CCOpcode) {
1184  case ISD::SETULE:
1185  case ISD::SETULT: {
1186    unsigned Opc = (LHS == True) ? ISD::UMIN : ISD::UMAX;
1187    return DAG.getNode(Opc, DL, VT, LHS, RHS);
1188  }
1189  case ISD::SETLE:
1190  case ISD::SETLT: {
1191    unsigned Opc = (LHS == True) ? ISD::SMIN : ISD::SMAX;
1192    return DAG.getNode(Opc, DL, VT, LHS, RHS);
1193  }
1194  case ISD::SETGT:
1195  case ISD::SETGE: {
1196    unsigned Opc = (LHS == True) ? ISD::SMAX : ISD::SMIN;
1197    return DAG.getNode(Opc, DL, VT, LHS, RHS);
1198  }
1199  case ISD::SETUGE:
1200  case ISD::SETUGT: {
1201    unsigned Opc = (LHS == True) ? ISD::UMAX : ISD::UMIN;
1202    return DAG.getNode(Opc, DL, VT, LHS, RHS);
1203  }
1204  default:
1205    return SDValue();
1206  }
1207}
1208
1209SDValue AMDGPUTargetLowering::ScalarizeVectorLoad(const SDValue Op,
1210                                                  SelectionDAG &DAG) const {
1211  LoadSDNode *Load = cast<LoadSDNode>(Op);
1212  EVT MemVT = Load->getMemoryVT();
1213  EVT MemEltVT = MemVT.getVectorElementType();
1214
1215  EVT LoadVT = Op.getValueType();
1216  EVT EltVT = LoadVT.getVectorElementType();
1217  EVT PtrVT = Load->getBasePtr().getValueType();
1218
1219  unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
1220  SmallVector<SDValue, 8> Loads;
1221  SmallVector<SDValue, 8> Chains;
1222
1223  SDLoc SL(Op);
1224  unsigned MemEltSize = MemEltVT.getStoreSize();
1225  MachinePointerInfo SrcValue(Load->getMemOperand()->getValue());
1226
1227  for (unsigned i = 0; i < NumElts; ++i) {
1228    SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
1229                              DAG.getConstant(i * MemEltSize, SL, PtrVT));
1230
1231    SDValue NewLoad
1232      = DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
1233                       Load->getChain(), Ptr,
1234                       SrcValue.getWithOffset(i * MemEltSize),
1235                       MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
1236                       Load->isInvariant(), Load->getAlignment());
1237    Loads.push_back(NewLoad.getValue(0));
1238    Chains.push_back(NewLoad.getValue(1));
1239  }
1240
1241  SDValue Ops[] = {
1242    DAG.getNode(ISD::BUILD_VECTOR, SL, LoadVT, Loads),
1243    DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains)
1244  };
1245
1246  return DAG.getMergeValues(Ops, SL);
1247}
1248
1249SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue Op,
1250                                              SelectionDAG &DAG) const {
1251  EVT VT = Op.getValueType();
1252
1253  // If this is a 2 element vector, we really want to scalarize and not create
1254  // weird 1 element vectors.
1255  if (VT.getVectorNumElements() == 2)
1256    return ScalarizeVectorLoad(Op, DAG);
1257
1258  LoadSDNode *Load = cast<LoadSDNode>(Op);
1259  SDValue BasePtr = Load->getBasePtr();
1260  EVT PtrVT = BasePtr.getValueType();
1261  EVT MemVT = Load->getMemoryVT();
1262  SDLoc SL(Op);
1263  MachinePointerInfo SrcValue(Load->getMemOperand()->getValue());
1264
1265  EVT LoVT, HiVT;
1266  EVT LoMemVT, HiMemVT;
1267  SDValue Lo, Hi;
1268
1269  std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1270  std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1271  std::tie(Lo, Hi) = DAG.SplitVector(Op, SL, LoVT, HiVT);
1272  SDValue LoLoad
1273    = DAG.getExtLoad(Load->getExtensionType(), SL, LoVT,
1274                     Load->getChain(), BasePtr,
1275                     SrcValue,
1276                     LoMemVT, Load->isVolatile(), Load->isNonTemporal(),
1277                     Load->isInvariant(), Load->getAlignment());
1278
1279  SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1280                              DAG.getConstant(LoMemVT.getStoreSize(), SL,
1281                                              PtrVT));
1282
1283  SDValue HiLoad
1284    = DAG.getExtLoad(Load->getExtensionType(), SL, HiVT,
1285                     Load->getChain(), HiPtr,
1286                     SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1287                     HiMemVT, Load->isVolatile(), Load->isNonTemporal(),
1288                     Load->isInvariant(), Load->getAlignment());
1289
1290  SDValue Ops[] = {
1291    DAG.getNode(ISD::CONCAT_VECTORS, SL, VT, LoLoad, HiLoad),
1292    DAG.getNode(ISD::TokenFactor, SL, MVT::Other,
1293                LoLoad.getValue(1), HiLoad.getValue(1))
1294  };
1295
1296  return DAG.getMergeValues(Ops, SL);
1297}
1298
1299SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
1300                                               SelectionDAG &DAG) const {
1301  StoreSDNode *Store = cast<StoreSDNode>(Op);
1302  EVT MemVT = Store->getMemoryVT();
1303  unsigned MemBits = MemVT.getSizeInBits();
1304
1305  // Byte stores are really expensive, so if possible, try to pack 32-bit vector
1306  // truncating store into an i32 store.
1307  // XXX: We could also handle optimize other vector bitwidths.
1308  if (!MemVT.isVector() || MemBits > 32) {
1309    return SDValue();
1310  }
1311
1312  SDLoc DL(Op);
1313  SDValue Value = Store->getValue();
1314  EVT VT = Value.getValueType();
1315  EVT ElemVT = VT.getVectorElementType();
1316  SDValue Ptr = Store->getBasePtr();
1317  EVT MemEltVT = MemVT.getVectorElementType();
1318  unsigned MemEltBits = MemEltVT.getSizeInBits();
1319  unsigned MemNumElements = MemVT.getVectorNumElements();
1320  unsigned PackedSize = MemVT.getStoreSizeInBits();
1321  SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, DL, MVT::i32);
1322
1323  assert(Value.getValueType().getScalarSizeInBits() >= 32);
1324
1325  SDValue PackedValue;
1326  for (unsigned i = 0; i < MemNumElements; ++i) {
1327    SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
1328                              DAG.getConstant(i, DL, MVT::i32));
1329    Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
1330    Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
1331
1332    SDValue Shift = DAG.getConstant(MemEltBits * i, DL, MVT::i32);
1333    Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
1334
1335    if (i == 0) {
1336      PackedValue = Elt;
1337    } else {
1338      PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
1339    }
1340  }
1341
1342  if (PackedSize < 32) {
1343    EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
1344    return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
1345                             Store->getMemOperand()->getPointerInfo(),
1346                             PackedVT,
1347                             Store->isNonTemporal(), Store->isVolatile(),
1348                             Store->getAlignment());
1349  }
1350
1351  return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
1352                      Store->getMemOperand()->getPointerInfo(),
1353                      Store->isVolatile(),  Store->isNonTemporal(),
1354                      Store->getAlignment());
1355}
1356
1357SDValue AMDGPUTargetLowering::ScalarizeVectorStore(SDValue Op,
1358                                                   SelectionDAG &DAG) const {
1359  StoreSDNode *Store = cast<StoreSDNode>(Op);
1360  EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
1361  EVT EltVT = Store->getValue().getValueType().getVectorElementType();
1362  EVT PtrVT = Store->getBasePtr().getValueType();
1363  unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1364  SDLoc SL(Op);
1365
1366  SmallVector<SDValue, 8> Chains;
1367
1368  unsigned EltSize = MemEltVT.getStoreSize();
1369  MachinePointerInfo SrcValue(Store->getMemOperand()->getValue());
1370
1371  for (unsigned i = 0, e = NumElts; i != e; ++i) {
1372    SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1373                              Store->getValue(),
1374                              DAG.getConstant(i, SL, MVT::i32));
1375
1376    SDValue Offset = DAG.getConstant(i * MemEltVT.getStoreSize(), SL, PtrVT);
1377    SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Store->getBasePtr(), Offset);
1378    SDValue NewStore =
1379      DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1380                        SrcValue.getWithOffset(i * EltSize),
1381                        MemEltVT, Store->isNonTemporal(), Store->isVolatile(),
1382                        Store->getAlignment());
1383    Chains.push_back(NewStore);
1384  }
1385
1386  return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1387}
1388
1389SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1390                                               SelectionDAG &DAG) const {
1391  StoreSDNode *Store = cast<StoreSDNode>(Op);
1392  SDValue Val = Store->getValue();
1393  EVT VT = Val.getValueType();
1394
1395  // If this is a 2 element vector, we really want to scalarize and not create
1396  // weird 1 element vectors.
1397  if (VT.getVectorNumElements() == 2)
1398    return ScalarizeVectorStore(Op, DAG);
1399
1400  EVT MemVT = Store->getMemoryVT();
1401  SDValue Chain = Store->getChain();
1402  SDValue BasePtr = Store->getBasePtr();
1403  SDLoc SL(Op);
1404
1405  EVT LoVT, HiVT;
1406  EVT LoMemVT, HiMemVT;
1407  SDValue Lo, Hi;
1408
1409  std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1410  std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1411  std::tie(Lo, Hi) = DAG.SplitVector(Val, SL, LoVT, HiVT);
1412
1413  EVT PtrVT = BasePtr.getValueType();
1414  SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1415                              DAG.getConstant(LoMemVT.getStoreSize(), SL,
1416                                              PtrVT));
1417
1418  MachinePointerInfo SrcValue(Store->getMemOperand()->getValue());
1419  SDValue LoStore
1420    = DAG.getTruncStore(Chain, SL, Lo,
1421                        BasePtr,
1422                        SrcValue,
1423                        LoMemVT,
1424                        Store->isNonTemporal(),
1425                        Store->isVolatile(),
1426                        Store->getAlignment());
1427  SDValue HiStore
1428    = DAG.getTruncStore(Chain, SL, Hi,
1429                        HiPtr,
1430                        SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1431                        HiMemVT,
1432                        Store->isNonTemporal(),
1433                        Store->isVolatile(),
1434                        Store->getAlignment());
1435
1436  return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoStore, HiStore);
1437}
1438
1439
1440SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1441  SDLoc DL(Op);
1442  LoadSDNode *Load = cast<LoadSDNode>(Op);
1443  ISD::LoadExtType ExtType = Load->getExtensionType();
1444  EVT VT = Op.getValueType();
1445  EVT MemVT = Load->getMemoryVT();
1446
1447  if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1448    assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1449    // FIXME: Copied from PPC
1450    // First, load into 32 bits, then truncate to 1 bit.
1451
1452    SDValue Chain = Load->getChain();
1453    SDValue BasePtr = Load->getBasePtr();
1454    MachineMemOperand *MMO = Load->getMemOperand();
1455
1456    SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1457                                   BasePtr, MVT::i8, MMO);
1458
1459    SDValue Ops[] = {
1460      DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD),
1461      NewLD.getValue(1)
1462    };
1463
1464    return DAG.getMergeValues(Ops, DL);
1465  }
1466
1467  if (Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS ||
1468      Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1469      ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1470    return SDValue();
1471
1472  // <SI && AS=PRIVATE && EXTLOAD && size < 32bit,
1473  // register (2-)byte extract.
1474
1475  // Get Register holding the target.
1476  SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1477                            DAG.getConstant(2, DL, MVT::i32));
1478  // Load the Register.
1479  SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1480                            Load->getChain(), Ptr,
1481                            DAG.getTargetConstant(0, DL, MVT::i32),
1482                            Op.getOperand(2));
1483
1484  // Get offset within the register.
1485  SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1486                                Load->getBasePtr(),
1487                                DAG.getConstant(0x3, DL, MVT::i32));
1488
1489  // Bit offset of target byte (byteIdx * 8).
1490  SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1491                                 DAG.getConstant(3, DL, MVT::i32));
1492
1493  // Shift to the right.
1494  Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1495
1496  // Eliminate the upper bits by setting them to ...
1497  EVT MemEltVT = MemVT.getScalarType();
1498
1499  // ... ones.
1500  if (ExtType == ISD::SEXTLOAD) {
1501    SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1502
1503    SDValue Ops[] = {
1504      DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode),
1505      Load->getChain()
1506    };
1507
1508    return DAG.getMergeValues(Ops, DL);
1509  }
1510
1511  // ... or zeros.
1512  SDValue Ops[] = {
1513    DAG.getZeroExtendInReg(Ret, DL, MemEltVT),
1514    Load->getChain()
1515  };
1516
1517  return DAG.getMergeValues(Ops, DL);
1518}
1519
1520SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1521  SDLoc DL(Op);
1522  SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1523  if (Result.getNode()) {
1524    return Result;
1525  }
1526
1527  StoreSDNode *Store = cast<StoreSDNode>(Op);
1528  SDValue Chain = Store->getChain();
1529  if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1530       Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1531      Store->getValue().getValueType().isVector()) {
1532    return ScalarizeVectorStore(Op, DAG);
1533  }
1534
1535  EVT MemVT = Store->getMemoryVT();
1536  if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1537      MemVT.bitsLT(MVT::i32)) {
1538    unsigned Mask = 0;
1539    if (Store->getMemoryVT() == MVT::i8) {
1540      Mask = 0xff;
1541    } else if (Store->getMemoryVT() == MVT::i16) {
1542      Mask = 0xffff;
1543    }
1544    SDValue BasePtr = Store->getBasePtr();
1545    SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1546                              DAG.getConstant(2, DL, MVT::i32));
1547    SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1548                              Chain, Ptr,
1549                              DAG.getTargetConstant(0, DL, MVT::i32));
1550
1551    SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1552                                  DAG.getConstant(0x3, DL, MVT::i32));
1553
1554    SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1555                                   DAG.getConstant(3, DL, MVT::i32));
1556
1557    SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1558                                    Store->getValue());
1559
1560    SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1561
1562    SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1563                                       MaskedValue, ShiftAmt);
1564
1565    SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32,
1566                                  DAG.getConstant(Mask, DL, MVT::i32),
1567                                  ShiftAmt);
1568    DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1569                          DAG.getConstant(0xffffffff, DL, MVT::i32));
1570    Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1571
1572    SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1573    return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1574                       Chain, Value, Ptr,
1575                       DAG.getTargetConstant(0, DL, MVT::i32));
1576  }
1577  return SDValue();
1578}
1579
1580// This is a shortcut for integer division because we have fast i32<->f32
1581// conversions, and fast f32 reciprocal instructions. The fractional part of a
1582// float is enough to accurately represent up to a 24-bit integer.
1583SDValue AMDGPUTargetLowering::LowerDIVREM24(SDValue Op, SelectionDAG &DAG, bool sign) const {
1584  SDLoc DL(Op);
1585  EVT VT = Op.getValueType();
1586  SDValue LHS = Op.getOperand(0);
1587  SDValue RHS = Op.getOperand(1);
1588  MVT IntVT = MVT::i32;
1589  MVT FltVT = MVT::f32;
1590
1591  ISD::NodeType ToFp  = sign ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
1592  ISD::NodeType ToInt = sign ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;
1593
1594  if (VT.isVector()) {
1595    unsigned NElts = VT.getVectorNumElements();
1596    IntVT = MVT::getVectorVT(MVT::i32, NElts);
1597    FltVT = MVT::getVectorVT(MVT::f32, NElts);
1598  }
1599
1600  unsigned BitSize = VT.getScalarType().getSizeInBits();
1601
1602  SDValue jq = DAG.getConstant(1, DL, IntVT);
1603
1604  if (sign) {
1605    // char|short jq = ia ^ ib;
1606    jq = DAG.getNode(ISD::XOR, DL, VT, LHS, RHS);
1607
1608    // jq = jq >> (bitsize - 2)
1609    jq = DAG.getNode(ISD::SRA, DL, VT, jq,
1610                     DAG.getConstant(BitSize - 2, DL, VT));
1611
1612    // jq = jq | 0x1
1613    jq = DAG.getNode(ISD::OR, DL, VT, jq, DAG.getConstant(1, DL, VT));
1614
1615    // jq = (int)jq
1616    jq = DAG.getSExtOrTrunc(jq, DL, IntVT);
1617  }
1618
1619  // int ia = (int)LHS;
1620  SDValue ia = sign ?
1621    DAG.getSExtOrTrunc(LHS, DL, IntVT) : DAG.getZExtOrTrunc(LHS, DL, IntVT);
1622
1623  // int ib, (int)RHS;
1624  SDValue ib = sign ?
1625    DAG.getSExtOrTrunc(RHS, DL, IntVT) : DAG.getZExtOrTrunc(RHS, DL, IntVT);
1626
1627  // float fa = (float)ia;
1628  SDValue fa = DAG.getNode(ToFp, DL, FltVT, ia);
1629
1630  // float fb = (float)ib;
1631  SDValue fb = DAG.getNode(ToFp, DL, FltVT, ib);
1632
1633  // float fq = native_divide(fa, fb);
1634  SDValue fq = DAG.getNode(ISD::FMUL, DL, FltVT,
1635                           fa, DAG.getNode(AMDGPUISD::RCP, DL, FltVT, fb));
1636
1637  // fq = trunc(fq);
1638  fq = DAG.getNode(ISD::FTRUNC, DL, FltVT, fq);
1639
1640  // float fqneg = -fq;
1641  SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FltVT, fq);
1642
1643  // float fr = mad(fqneg, fb, fa);
1644  SDValue fr = DAG.getNode(ISD::FADD, DL, FltVT,
1645                           DAG.getNode(ISD::FMUL, DL, FltVT, fqneg, fb), fa);
1646
1647  // int iq = (int)fq;
1648  SDValue iq = DAG.getNode(ToInt, DL, IntVT, fq);
1649
1650  // fr = fabs(fr);
1651  fr = DAG.getNode(ISD::FABS, DL, FltVT, fr);
1652
1653  // fb = fabs(fb);
1654  fb = DAG.getNode(ISD::FABS, DL, FltVT, fb);
1655
1656  EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
1657
1658  // int cv = fr >= fb;
1659  SDValue cv = DAG.getSetCC(DL, SetCCVT, fr, fb, ISD::SETOGE);
1660
1661  // jq = (cv ? jq : 0);
1662  jq = DAG.getNode(ISD::SELECT, DL, VT, cv, jq, DAG.getConstant(0, DL, VT));
1663
1664  // dst = trunc/extend to legal type
1665  iq = sign ? DAG.getSExtOrTrunc(iq, DL, VT) : DAG.getZExtOrTrunc(iq, DL, VT);
1666
1667  // dst = iq + jq;
1668  SDValue Div = DAG.getNode(ISD::ADD, DL, VT, iq, jq);
1669
1670  // Rem needs compensation, it's easier to recompute it
1671  SDValue Rem = DAG.getNode(ISD::MUL, DL, VT, Div, RHS);
1672  Rem = DAG.getNode(ISD::SUB, DL, VT, LHS, Rem);
1673
1674  SDValue Res[2] = {
1675    Div,
1676    Rem
1677  };
1678  return DAG.getMergeValues(Res, DL);
1679}
1680
1681void AMDGPUTargetLowering::LowerUDIVREM64(SDValue Op,
1682                                      SelectionDAG &DAG,
1683                                      SmallVectorImpl<SDValue> &Results) const {
1684  assert(Op.getValueType() == MVT::i64);
1685
1686  SDLoc DL(Op);
1687  EVT VT = Op.getValueType();
1688  EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
1689
1690  SDValue one = DAG.getConstant(1, DL, HalfVT);
1691  SDValue zero = DAG.getConstant(0, DL, HalfVT);
1692
1693  //HiLo split
1694  SDValue LHS = Op.getOperand(0);
1695  SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
1696  SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
1697
1698  SDValue RHS = Op.getOperand(1);
1699  SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
1700  SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
1701
1702  if (VT == MVT::i64 &&
1703    DAG.MaskedValueIsZero(RHS, APInt::getHighBitsSet(64, 32)) &&
1704    DAG.MaskedValueIsZero(LHS, APInt::getHighBitsSet(64, 32))) {
1705
1706    SDValue Res = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
1707                              LHS_Lo, RHS_Lo);
1708
1709    SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, Res.getValue(0), zero);
1710    SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, Res.getValue(1), zero);
1711    Results.push_back(DIV);
1712    Results.push_back(REM);
1713    return;
1714  }
1715
1716  // Get Speculative values
1717  SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
1718  SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
1719
1720  SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
1721  SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, zero);
1722
1723  SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
1724  SDValue DIV_Lo = zero;
1725
1726  const unsigned halfBitWidth = HalfVT.getSizeInBits();
1727
1728  for (unsigned i = 0; i < halfBitWidth; ++i) {
1729    const unsigned bitPos = halfBitWidth - i - 1;
1730    SDValue POS = DAG.getConstant(bitPos, DL, HalfVT);
1731    // Get value of high bit
1732    SDValue HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
1733    HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
1734    HBit = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, HBit);
1735
1736    // Shift
1737    REM = DAG.getNode(ISD::SHL, DL, VT, REM, DAG.getConstant(1, DL, VT));
1738    // Add LHS high bit
1739    REM = DAG.getNode(ISD::OR, DL, VT, REM, HBit);
1740
1741    SDValue BIT = DAG.getConstant(1 << bitPos, DL, HalfVT);
1742    SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETUGE);
1743
1744    DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
1745
1746    // Update REM
1747    SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
1748    REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETUGE);
1749  }
1750
1751  SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
1752  Results.push_back(DIV);
1753  Results.push_back(REM);
1754}
1755
1756SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1757                                           SelectionDAG &DAG) const {
1758  SDLoc DL(Op);
1759  EVT VT = Op.getValueType();
1760
1761  if (VT == MVT::i64) {
1762    SmallVector<SDValue, 2> Results;
1763    LowerUDIVREM64(Op, DAG, Results);
1764    return DAG.getMergeValues(Results, DL);
1765  }
1766
1767  SDValue Num = Op.getOperand(0);
1768  SDValue Den = Op.getOperand(1);
1769
1770  if (VT == MVT::i32) {
1771    if (DAG.MaskedValueIsZero(Num, APInt::getHighBitsSet(32, 8)) &&
1772        DAG.MaskedValueIsZero(Den, APInt::getHighBitsSet(32, 8))) {
1773      // TODO: We technically could do this for i64, but shouldn't that just be
1774      // handled by something generally reducing 64-bit division on 32-bit
1775      // values to 32-bit?
1776      return LowerDIVREM24(Op, DAG, false);
1777    }
1778  }
1779
1780  // RCP =  URECIP(Den) = 2^32 / Den + e
1781  // e is rounding error.
1782  SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1783
1784  // RCP_LO = mul(RCP, Den) */
1785  SDValue RCP_LO = DAG.getNode(ISD::MUL, DL, VT, RCP, Den);
1786
1787  // RCP_HI = mulhu (RCP, Den) */
1788  SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1789
1790  // NEG_RCP_LO = -RCP_LO
1791  SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
1792                                                     RCP_LO);
1793
1794  // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1795  SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, DL, VT),
1796                                           NEG_RCP_LO, RCP_LO,
1797                                           ISD::SETEQ);
1798  // Calculate the rounding error from the URECIP instruction
1799  // E = mulhu(ABS_RCP_LO, RCP)
1800  SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1801
1802  // RCP_A_E = RCP + E
1803  SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1804
1805  // RCP_S_E = RCP - E
1806  SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1807
1808  // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1809  SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, DL, VT),
1810                                     RCP_A_E, RCP_S_E,
1811                                     ISD::SETEQ);
1812  // Quotient = mulhu(Tmp0, Num)
1813  SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1814
1815  // Num_S_Remainder = Quotient * Den
1816  SDValue Num_S_Remainder = DAG.getNode(ISD::MUL, DL, VT, Quotient, Den);
1817
1818  // Remainder = Num - Num_S_Remainder
1819  SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1820
1821  // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1822  SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1823                                                 DAG.getConstant(-1, DL, VT),
1824                                                 DAG.getConstant(0, DL, VT),
1825                                                 ISD::SETUGE);
1826  // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1827  SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1828                                                  Num_S_Remainder,
1829                                                  DAG.getConstant(-1, DL, VT),
1830                                                  DAG.getConstant(0, DL, VT),
1831                                                  ISD::SETUGE);
1832  // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1833  SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1834                                               Remainder_GE_Zero);
1835
1836  // Calculate Division result:
1837
1838  // Quotient_A_One = Quotient + 1
1839  SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1840                                       DAG.getConstant(1, DL, VT));
1841
1842  // Quotient_S_One = Quotient - 1
1843  SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1844                                       DAG.getConstant(1, DL, VT));
1845
1846  // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1847  SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, DL, VT),
1848                                     Quotient, Quotient_A_One, ISD::SETEQ);
1849
1850  // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1851  Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, DL, VT),
1852                            Quotient_S_One, Div, ISD::SETEQ);
1853
1854  // Calculate Rem result:
1855
1856  // Remainder_S_Den = Remainder - Den
1857  SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1858
1859  // Remainder_A_Den = Remainder + Den
1860  SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1861
1862  // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1863  SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, DL, VT),
1864                                    Remainder, Remainder_S_Den, ISD::SETEQ);
1865
1866  // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1867  Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, DL, VT),
1868                            Remainder_A_Den, Rem, ISD::SETEQ);
1869  SDValue Ops[2] = {
1870    Div,
1871    Rem
1872  };
1873  return DAG.getMergeValues(Ops, DL);
1874}
1875
1876SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
1877                                           SelectionDAG &DAG) const {
1878  SDLoc DL(Op);
1879  EVT VT = Op.getValueType();
1880
1881  SDValue LHS = Op.getOperand(0);
1882  SDValue RHS = Op.getOperand(1);
1883
1884  SDValue Zero = DAG.getConstant(0, DL, VT);
1885  SDValue NegOne = DAG.getConstant(-1, DL, VT);
1886
1887  if (VT == MVT::i32 &&
1888      DAG.ComputeNumSignBits(LHS) > 8 &&
1889      DAG.ComputeNumSignBits(RHS) > 8) {
1890    return LowerDIVREM24(Op, DAG, true);
1891  }
1892  if (VT == MVT::i64 &&
1893      DAG.ComputeNumSignBits(LHS) > 32 &&
1894      DAG.ComputeNumSignBits(RHS) > 32) {
1895    EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
1896
1897    //HiLo split
1898    SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero);
1899    SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero);
1900    SDValue DIVREM = DAG.getNode(ISD::SDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
1901                                 LHS_Lo, RHS_Lo);
1902    SDValue Res[2] = {
1903      DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(0)),
1904      DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(1))
1905    };
1906    return DAG.getMergeValues(Res, DL);
1907  }
1908
1909  SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
1910  SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
1911  SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
1912  SDValue RSign = LHSign; // Remainder sign is the same as LHS
1913
1914  LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
1915  RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);
1916
1917  LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
1918  RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);
1919
1920  SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
1921  SDValue Rem = Div.getValue(1);
1922
1923  Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
1924  Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);
1925
1926  Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
1927  Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);
1928
1929  SDValue Res[2] = {
1930    Div,
1931    Rem
1932  };
1933  return DAG.getMergeValues(Res, DL);
1934}
1935
1936// (frem x, y) -> (fsub x, (fmul (ftrunc (fdiv x, y)), y))
1937SDValue AMDGPUTargetLowering::LowerFREM(SDValue Op, SelectionDAG &DAG) const {
1938  SDLoc SL(Op);
1939  EVT VT = Op.getValueType();
1940  SDValue X = Op.getOperand(0);
1941  SDValue Y = Op.getOperand(1);
1942
1943  SDValue Div = DAG.getNode(ISD::FDIV, SL, VT, X, Y);
1944  SDValue Floor = DAG.getNode(ISD::FTRUNC, SL, VT, Div);
1945  SDValue Mul = DAG.getNode(ISD::FMUL, SL, VT, Floor, Y);
1946
1947  return DAG.getNode(ISD::FSUB, SL, VT, X, Mul);
1948}
1949
1950SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
1951  SDLoc SL(Op);
1952  SDValue Src = Op.getOperand(0);
1953
1954  // result = trunc(src)
1955  // if (src > 0.0 && src != result)
1956  //   result += 1.0
1957
1958  SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1959
1960  const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64);
1961  const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f64);
1962
1963  EVT SetCCVT =
1964      getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
1965
1966  SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
1967  SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1968  SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1969
1970  SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
1971  return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1972}
1973
1974static SDValue extractF64Exponent(SDValue Hi, SDLoc SL, SelectionDAG &DAG) {
1975  const unsigned FractBits = 52;
1976  const unsigned ExpBits = 11;
1977
1978  SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_U32, SL, MVT::i32,
1979                                Hi,
1980                                DAG.getConstant(FractBits - 32, SL, MVT::i32),
1981                                DAG.getConstant(ExpBits, SL, MVT::i32));
1982  SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
1983                            DAG.getConstant(1023, SL, MVT::i32));
1984
1985  return Exp;
1986}
1987
1988SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
1989  SDLoc SL(Op);
1990  SDValue Src = Op.getOperand(0);
1991
1992  assert(Op.getValueType() == MVT::f64);
1993
1994  const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
1995  const SDValue One = DAG.getConstant(1, SL, MVT::i32);
1996
1997  SDValue VecSrc = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1998
1999  // Extract the upper half, since this is where we will find the sign and
2000  // exponent.
2001  SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
2002
2003  SDValue Exp = extractF64Exponent(Hi, SL, DAG);
2004
2005  const unsigned FractBits = 52;
2006
2007  // Extract the sign bit.
2008  const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, SL, MVT::i32);
2009  SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);
2010
2011  // Extend back to to 64-bits.
2012  SDValue SignBit64 = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
2013                                  Zero, SignBit);
2014  SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);
2015
2016  SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
2017  const SDValue FractMask
2018    = DAG.getConstant((UINT64_C(1) << FractBits) - 1, SL, MVT::i64);
2019
2020  SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
2021  SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
2022  SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);
2023
2024  EVT SetCCVT =
2025      getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i32);
2026
2027  const SDValue FiftyOne = DAG.getConstant(FractBits - 1, SL, MVT::i32);
2028
2029  SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
2030  SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
2031
2032  SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
2033  SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);
2034
2035  return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
2036}
2037
2038SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
2039  SDLoc SL(Op);
2040  SDValue Src = Op.getOperand(0);
2041
2042  assert(Op.getValueType() == MVT::f64);
2043
2044  APFloat C1Val(APFloat::IEEEdouble, "0x1.0p+52");
2045  SDValue C1 = DAG.getConstantFP(C1Val, SL, MVT::f64);
2046  SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
2047
2048  SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
2049  SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
2050
2051  SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
2052
2053  APFloat C2Val(APFloat::IEEEdouble, "0x1.fffffffffffffp+51");
2054  SDValue C2 = DAG.getConstantFP(C2Val, SL, MVT::f64);
2055
2056  EVT SetCCVT =
2057      getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
2058  SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
2059
2060  return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
2061}
2062
2063SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const {
2064  // FNEARBYINT and FRINT are the same, except in their handling of FP
2065  // exceptions. Those aren't really meaningful for us, and OpenCL only has
2066  // rint, so just treat them as equivalent.
2067  return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
2068}
2069
2070// XXX - May require not supporting f32 denormals?
2071SDValue AMDGPUTargetLowering::LowerFROUND32(SDValue Op, SelectionDAG &DAG) const {
2072  SDLoc SL(Op);
2073  SDValue X = Op.getOperand(0);
2074
2075  SDValue T = DAG.getNode(ISD::FTRUNC, SL, MVT::f32, X);
2076
2077  SDValue Diff = DAG.getNode(ISD::FSUB, SL, MVT::f32, X, T);
2078
2079  SDValue AbsDiff = DAG.getNode(ISD::FABS, SL, MVT::f32, Diff);
2080
2081  const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f32);
2082  const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f32);
2083  const SDValue Half = DAG.getConstantFP(0.5, SL, MVT::f32);
2084
2085  SDValue SignOne = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f32, One, X);
2086
2087  EVT SetCCVT =
2088      getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f32);
2089
2090  SDValue Cmp = DAG.getSetCC(SL, SetCCVT, AbsDiff, Half, ISD::SETOGE);
2091
2092  SDValue Sel = DAG.getNode(ISD::SELECT, SL, MVT::f32, Cmp, SignOne, Zero);
2093
2094  return DAG.getNode(ISD::FADD, SL, MVT::f32, T, Sel);
2095}
2096
2097SDValue AMDGPUTargetLowering::LowerFROUND64(SDValue Op, SelectionDAG &DAG) const {
2098  SDLoc SL(Op);
2099  SDValue X = Op.getOperand(0);
2100
2101  SDValue L = DAG.getNode(ISD::BITCAST, SL, MVT::i64, X);
2102
2103  const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
2104  const SDValue One = DAG.getConstant(1, SL, MVT::i32);
2105  const SDValue NegOne = DAG.getConstant(-1, SL, MVT::i32);
2106  const SDValue FiftyOne = DAG.getConstant(51, SL, MVT::i32);
2107  EVT SetCCVT =
2108      getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i32);
2109
2110  SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, X);
2111
2112  SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC, One);
2113
2114  SDValue Exp = extractF64Exponent(Hi, SL, DAG);
2115
2116  const SDValue Mask = DAG.getConstant(INT64_C(0x000fffffffffffff), SL,
2117                                       MVT::i64);
2118
2119  SDValue M = DAG.getNode(ISD::SRA, SL, MVT::i64, Mask, Exp);
2120  SDValue D = DAG.getNode(ISD::SRA, SL, MVT::i64,
2121                          DAG.getConstant(INT64_C(0x0008000000000000), SL,
2122                                          MVT::i64),
2123                          Exp);
2124
2125  SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, L, M);
2126  SDValue Tmp1 = DAG.getSetCC(SL, SetCCVT,
2127                              DAG.getConstant(0, SL, MVT::i64), Tmp0,
2128                              ISD::SETNE);
2129
2130  SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, Tmp1,
2131                             D, DAG.getConstant(0, SL, MVT::i64));
2132  SDValue K = DAG.getNode(ISD::ADD, SL, MVT::i64, L, Tmp2);
2133
2134  K = DAG.getNode(ISD::AND, SL, MVT::i64, K, DAG.getNOT(SL, M, MVT::i64));
2135  K = DAG.getNode(ISD::BITCAST, SL, MVT::f64, K);
2136
2137  SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
2138  SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
2139  SDValue ExpEqNegOne = DAG.getSetCC(SL, SetCCVT, NegOne, Exp, ISD::SETEQ);
2140
2141  SDValue Mag = DAG.getNode(ISD::SELECT, SL, MVT::f64,
2142                            ExpEqNegOne,
2143                            DAG.getConstantFP(1.0, SL, MVT::f64),
2144                            DAG.getConstantFP(0.0, SL, MVT::f64));
2145
2146  SDValue S = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, Mag, X);
2147
2148  K = DAG.getNode(ISD::SELECT, SL, MVT::f64, ExpLt0, S, K);
2149  K = DAG.getNode(ISD::SELECT, SL, MVT::f64, ExpGt51, X, K);
2150
2151  return K;
2152}
2153
2154SDValue AMDGPUTargetLowering::LowerFROUND(SDValue Op, SelectionDAG &DAG) const {
2155  EVT VT = Op.getValueType();
2156
2157  if (VT == MVT::f32)
2158    return LowerFROUND32(Op, DAG);
2159
2160  if (VT == MVT::f64)
2161    return LowerFROUND64(Op, DAG);
2162
2163  llvm_unreachable("unhandled type");
2164}
2165
2166SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
2167  SDLoc SL(Op);
2168  SDValue Src = Op.getOperand(0);
2169
2170  // result = trunc(src);
2171  // if (src < 0.0 && src != result)
2172  //   result += -1.0.
2173
2174  SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2175
2176  const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64);
2177  const SDValue NegOne = DAG.getConstantFP(-1.0, SL, MVT::f64);
2178
2179  EVT SetCCVT =
2180      getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
2181
2182  SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
2183  SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
2184  SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
2185
2186  SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
2187  return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
2188}
2189
2190SDValue AMDGPUTargetLowering::LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG,
2191                                               bool Signed) const {
2192  SDLoc SL(Op);
2193  SDValue Src = Op.getOperand(0);
2194
2195  SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
2196
2197  SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
2198                           DAG.getConstant(0, SL, MVT::i32));
2199  SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
2200                           DAG.getConstant(1, SL, MVT::i32));
2201
2202  SDValue CvtHi = DAG.getNode(Signed ? ISD::SINT_TO_FP : ISD::UINT_TO_FP,
2203                              SL, MVT::f64, Hi);
2204
2205  SDValue CvtLo = DAG.getNode(ISD::UINT_TO_FP, SL, MVT::f64, Lo);
2206
2207  SDValue LdExp = DAG.getNode(AMDGPUISD::LDEXP, SL, MVT::f64, CvtHi,
2208                              DAG.getConstant(32, SL, MVT::i32));
2209
2210  return DAG.getNode(ISD::FADD, SL, MVT::f64, LdExp, CvtLo);
2211}
2212
2213SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
2214                                               SelectionDAG &DAG) const {
2215  SDValue S0 = Op.getOperand(0);
2216  if (S0.getValueType() != MVT::i64)
2217    return SDValue();
2218
2219  EVT DestVT = Op.getValueType();
2220  if (DestVT == MVT::f64)
2221    return LowerINT_TO_FP64(Op, DAG, false);
2222
2223  assert(DestVT == MVT::f32);
2224
2225  SDLoc DL(Op);
2226
2227  // f32 uint_to_fp i64
2228  SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
2229                           DAG.getConstant(0, DL, MVT::i32));
2230  SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
2231  SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
2232                           DAG.getConstant(1, DL, MVT::i32));
2233  SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
2234  FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
2235                        DAG.getConstantFP(4294967296.0f, DL, MVT::f32)); // 2^32
2236  return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
2237}
2238
2239SDValue AMDGPUTargetLowering::LowerSINT_TO_FP(SDValue Op,
2240                                              SelectionDAG &DAG) const {
2241  SDValue Src = Op.getOperand(0);
2242  if (Src.getValueType() == MVT::i64 && Op.getValueType() == MVT::f64)
2243    return LowerINT_TO_FP64(Op, DAG, true);
2244
2245  return SDValue();
2246}
2247
2248SDValue AMDGPUTargetLowering::LowerFP64_TO_INT(SDValue Op, SelectionDAG &DAG,
2249                                               bool Signed) const {
2250  SDLoc SL(Op);
2251
2252  SDValue Src = Op.getOperand(0);
2253
2254  SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2255
2256  SDValue K0 = DAG.getConstantFP(BitsToDouble(UINT64_C(0x3df0000000000000)), SL,
2257                                 MVT::f64);
2258  SDValue K1 = DAG.getConstantFP(BitsToDouble(UINT64_C(0xc1f0000000000000)), SL,
2259                                 MVT::f64);
2260
2261  SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f64, Trunc, K0);
2262
2263  SDValue FloorMul = DAG.getNode(ISD::FFLOOR, SL, MVT::f64, Mul);
2264
2265
2266  SDValue Fma = DAG.getNode(ISD::FMA, SL, MVT::f64, FloorMul, K1, Trunc);
2267
2268  SDValue Hi = DAG.getNode(Signed ? ISD::FP_TO_SINT : ISD::FP_TO_UINT, SL,
2269                           MVT::i32, FloorMul);
2270  SDValue Lo = DAG.getNode(ISD::FP_TO_UINT, SL, MVT::i32, Fma);
2271
2272  SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32, Lo, Hi);
2273
2274  return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Result);
2275}
2276
2277SDValue AMDGPUTargetLowering::LowerFP_TO_SINT(SDValue Op,
2278                                              SelectionDAG &DAG) const {
2279  SDValue Src = Op.getOperand(0);
2280
2281  if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2282    return LowerFP64_TO_INT(Op, DAG, true);
2283
2284  return SDValue();
2285}
2286
2287SDValue AMDGPUTargetLowering::LowerFP_TO_UINT(SDValue Op,
2288                                              SelectionDAG &DAG) const {
2289  SDValue Src = Op.getOperand(0);
2290
2291  if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2292    return LowerFP64_TO_INT(Op, DAG, false);
2293
2294  return SDValue();
2295}
2296
2297SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
2298                                                     SelectionDAG &DAG) const {
2299  EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
2300  MVT VT = Op.getSimpleValueType();
2301  MVT ScalarVT = VT.getScalarType();
2302
2303  if (!VT.isVector())
2304    return SDValue();
2305
2306  SDValue Src = Op.getOperand(0);
2307  SDLoc DL(Op);
2308
2309  // TODO: Don't scalarize on Evergreen?
2310  unsigned NElts = VT.getVectorNumElements();
2311  SmallVector<SDValue, 8> Args;
2312  DAG.ExtractVectorElements(Src, Args, 0, NElts);
2313
2314  SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
2315  for (unsigned I = 0; I < NElts; ++I)
2316    Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
2317
2318  return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
2319}
2320
2321//===----------------------------------------------------------------------===//
2322// Custom DAG optimizations
2323//===----------------------------------------------------------------------===//
2324
2325static bool isU24(SDValue Op, SelectionDAG &DAG) {
2326  APInt KnownZero, KnownOne;
2327  EVT VT = Op.getValueType();
2328  DAG.computeKnownBits(Op, KnownZero, KnownOne);
2329
2330  return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
2331}
2332
2333static bool isI24(SDValue Op, SelectionDAG &DAG) {
2334  EVT VT = Op.getValueType();
2335
2336  // In order for this to be a signed 24-bit value, bit 23, must
2337  // be a sign bit.
2338  return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
2339                                     // as unsigned 24-bit values.
2340         (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
2341}
2342
2343static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
2344
2345  SelectionDAG &DAG = DCI.DAG;
2346  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2347  EVT VT = Op.getValueType();
2348
2349  APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
2350  APInt KnownZero, KnownOne;
2351  TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
2352  if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
2353    DCI.CommitTargetLoweringOpt(TLO);
2354}
2355
2356template <typename IntTy>
2357static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
2358                               uint32_t Offset, uint32_t Width, SDLoc DL) {
2359  if (Width + Offset < 32) {
2360    uint32_t Shl = static_cast<uint32_t>(Src0) << (32 - Offset - Width);
2361    IntTy Result = static_cast<IntTy>(Shl) >> (32 - Width);
2362    return DAG.getConstant(Result, DL, MVT::i32);
2363  }
2364
2365  return DAG.getConstant(Src0 >> Offset, DL, MVT::i32);
2366}
2367
2368static bool usesAllNormalStores(SDNode *LoadVal) {
2369  for (SDNode::use_iterator I = LoadVal->use_begin(); !I.atEnd(); ++I) {
2370    if (!ISD::isNormalStore(*I))
2371      return false;
2372  }
2373
2374  return true;
2375}
2376
2377// If we have a copy of an illegal type, replace it with a load / store of an
2378// equivalently sized legal type. This avoids intermediate bit pack / unpack
2379// instructions emitted when handling extloads and truncstores. Ideally we could
2380// recognize the pack / unpack pattern to eliminate it.
2381SDValue AMDGPUTargetLowering::performStoreCombine(SDNode *N,
2382                                                  DAGCombinerInfo &DCI) const {
2383  if (!DCI.isBeforeLegalize())
2384    return SDValue();
2385
2386  StoreSDNode *SN = cast<StoreSDNode>(N);
2387  SDValue Value = SN->getValue();
2388  EVT VT = Value.getValueType();
2389
2390  if (isTypeLegal(VT) || SN->isVolatile() ||
2391      !ISD::isNormalLoad(Value.getNode()) || VT.getSizeInBits() < 8)
2392    return SDValue();
2393
2394  LoadSDNode *LoadVal = cast<LoadSDNode>(Value);
2395  if (LoadVal->isVolatile() || !usesAllNormalStores(LoadVal))
2396    return SDValue();
2397
2398  EVT MemVT = LoadVal->getMemoryVT();
2399
2400  SDLoc SL(N);
2401  SelectionDAG &DAG = DCI.DAG;
2402  EVT LoadVT = getEquivalentMemType(*DAG.getContext(), MemVT);
2403
2404  SDValue NewLoad = DAG.getLoad(ISD::UNINDEXED, ISD::NON_EXTLOAD,
2405                                LoadVT, SL,
2406                                LoadVal->getChain(),
2407                                LoadVal->getBasePtr(),
2408                                LoadVal->getOffset(),
2409                                LoadVT,
2410                                LoadVal->getMemOperand());
2411
2412  SDValue CastLoad = DAG.getNode(ISD::BITCAST, SL, VT, NewLoad.getValue(0));
2413  DCI.CombineTo(LoadVal, CastLoad, NewLoad.getValue(1), false);
2414
2415  return DAG.getStore(SN->getChain(), SL, NewLoad,
2416                      SN->getBasePtr(), SN->getMemOperand());
2417}
2418
2419SDValue AMDGPUTargetLowering::performShlCombine(SDNode *N,
2420                                                DAGCombinerInfo &DCI) const {
2421  if (N->getValueType(0) != MVT::i64)
2422    return SDValue();
2423
2424  // i64 (shl x, 32) -> (build_pair 0, x)
2425
2426  // Doing this with moves theoretically helps MI optimizations that understand
2427  // copies. 2 v_mov_b32_e32 will have the same code size / cycle count as
2428  // v_lshl_b64. In the SALU case, I think this is slightly worse since it
2429  // doubles the code size and I'm unsure about cycle count.
2430  const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
2431  if (!RHS || RHS->getZExtValue() != 32)
2432    return SDValue();
2433
2434  SDValue LHS = N->getOperand(0);
2435
2436  SDLoc SL(N);
2437  SelectionDAG &DAG = DCI.DAG;
2438
2439  // Extract low 32-bits.
2440  SDValue Lo = DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, LHS);
2441
2442  const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
2443  return DAG.getNode(ISD::BUILD_PAIR, SL, MVT::i64, Zero, Lo);
2444}
2445
2446SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N,
2447                                                DAGCombinerInfo &DCI) const {
2448  EVT VT = N->getValueType(0);
2449
2450  if (VT.isVector() || VT.getSizeInBits() > 32)
2451    return SDValue();
2452
2453  SelectionDAG &DAG = DCI.DAG;
2454  SDLoc DL(N);
2455
2456  SDValue N0 = N->getOperand(0);
2457  SDValue N1 = N->getOperand(1);
2458  SDValue Mul;
2459
2460  if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
2461    N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
2462    N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
2463    Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
2464  } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
2465    N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
2466    N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
2467    Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
2468  } else {
2469    return SDValue();
2470  }
2471
2472  // We need to use sext even for MUL_U24, because MUL_U24 is used
2473  // for signed multiply of 8 and 16-bit types.
2474  return DAG.getSExtOrTrunc(Mul, DL, VT);
2475}
2476
2477SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
2478                                                DAGCombinerInfo &DCI) const {
2479  SelectionDAG &DAG = DCI.DAG;
2480  SDLoc DL(N);
2481
2482  switch(N->getOpcode()) {
2483  default:
2484    break;
2485  case ISD::SHL: {
2486    if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
2487      break;
2488
2489    return performShlCombine(N, DCI);
2490  }
2491  case ISD::MUL:
2492    return performMulCombine(N, DCI);
2493  case AMDGPUISD::MUL_I24:
2494  case AMDGPUISD::MUL_U24: {
2495    SDValue N0 = N->getOperand(0);
2496    SDValue N1 = N->getOperand(1);
2497    simplifyI24(N0, DCI);
2498    simplifyI24(N1, DCI);
2499    return SDValue();
2500  }
2501  case ISD::SELECT: {
2502    SDValue Cond = N->getOperand(0);
2503    if (Cond.getOpcode() == ISD::SETCC && Cond.hasOneUse()) {
2504      EVT VT = N->getValueType(0);
2505      SDValue LHS = Cond.getOperand(0);
2506      SDValue RHS = Cond.getOperand(1);
2507      SDValue CC = Cond.getOperand(2);
2508
2509      SDValue True = N->getOperand(1);
2510      SDValue False = N->getOperand(2);
2511
2512      if (VT == MVT::f32)
2513        return CombineFMinMaxLegacy(DL, VT, LHS, RHS, True, False, CC, DCI);
2514
2515      // TODO: Implement min / max Evergreen instructions.
2516      if (VT == MVT::i32 &&
2517          Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
2518        return CombineIMinMax(DL, VT, LHS, RHS, True, False, CC, DAG);
2519      }
2520    }
2521
2522    break;
2523  }
2524  case AMDGPUISD::BFE_I32:
2525  case AMDGPUISD::BFE_U32: {
2526    assert(!N->getValueType(0).isVector() &&
2527           "Vector handling of BFE not implemented");
2528    ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
2529    if (!Width)
2530      break;
2531
2532    uint32_t WidthVal = Width->getZExtValue() & 0x1f;
2533    if (WidthVal == 0)
2534      return DAG.getConstant(0, DL, MVT::i32);
2535
2536    ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
2537    if (!Offset)
2538      break;
2539
2540    SDValue BitsFrom = N->getOperand(0);
2541    uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
2542
2543    bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
2544
2545    if (OffsetVal == 0) {
2546      // This is already sign / zero extended, so try to fold away extra BFEs.
2547      unsigned SignBits =  Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
2548
2549      unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
2550      if (OpSignBits >= SignBits)
2551        return BitsFrom;
2552
2553      EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
2554      if (Signed) {
2555        // This is a sign_extend_inreg. Replace it to take advantage of existing
2556        // DAG Combines. If not eliminated, we will match back to BFE during
2557        // selection.
2558
2559        // TODO: The sext_inreg of extended types ends, although we can could
2560        // handle them in a single BFE.
2561        return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
2562                           DAG.getValueType(SmallVT));
2563      }
2564
2565      return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
2566    }
2567
2568    if (ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(BitsFrom)) {
2569      if (Signed) {
2570        return constantFoldBFE<int32_t>(DAG,
2571                                        CVal->getSExtValue(),
2572                                        OffsetVal,
2573                                        WidthVal,
2574                                        DL);
2575      }
2576
2577      return constantFoldBFE<uint32_t>(DAG,
2578                                       CVal->getZExtValue(),
2579                                       OffsetVal,
2580                                       WidthVal,
2581                                       DL);
2582    }
2583
2584    if ((OffsetVal + WidthVal) >= 32) {
2585      SDValue ShiftVal = DAG.getConstant(OffsetVal, DL, MVT::i32);
2586      return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
2587                         BitsFrom, ShiftVal);
2588    }
2589
2590    if (BitsFrom.hasOneUse()) {
2591      APInt Demanded = APInt::getBitsSet(32,
2592                                         OffsetVal,
2593                                         OffsetVal + WidthVal);
2594
2595      APInt KnownZero, KnownOne;
2596      TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
2597                                            !DCI.isBeforeLegalizeOps());
2598      const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2599      if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
2600          TLI.SimplifyDemandedBits(BitsFrom, Demanded,
2601                                   KnownZero, KnownOne, TLO)) {
2602        DCI.CommitTargetLoweringOpt(TLO);
2603      }
2604    }
2605
2606    break;
2607  }
2608
2609  case ISD::STORE:
2610    return performStoreCombine(N, DCI);
2611  }
2612  return SDValue();
2613}
2614
2615//===----------------------------------------------------------------------===//
2616// Helper functions
2617//===----------------------------------------------------------------------===//
2618
2619void AMDGPUTargetLowering::getOriginalFunctionArgs(
2620                               SelectionDAG &DAG,
2621                               const Function *F,
2622                               const SmallVectorImpl<ISD::InputArg> &Ins,
2623                               SmallVectorImpl<ISD::InputArg> &OrigIns) const {
2624
2625  for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
2626    if (Ins[i].ArgVT == Ins[i].VT) {
2627      OrigIns.push_back(Ins[i]);
2628      continue;
2629    }
2630
2631    EVT VT;
2632    if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
2633      // Vector has been split into scalars.
2634      VT = Ins[i].ArgVT.getVectorElementType();
2635    } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
2636               Ins[i].ArgVT.getVectorElementType() !=
2637               Ins[i].VT.getVectorElementType()) {
2638      // Vector elements have been promoted
2639      VT = Ins[i].ArgVT;
2640    } else {
2641      // Vector has been spilt into smaller vectors.
2642      VT = Ins[i].VT;
2643    }
2644
2645    ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
2646                      Ins[i].OrigArgIndex, Ins[i].PartOffset);
2647    OrigIns.push_back(Arg);
2648  }
2649}
2650
2651bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
2652  if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2653    return CFP->isExactlyValue(1.0);
2654  }
2655  if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2656    return C->isAllOnesValue();
2657  }
2658  return false;
2659}
2660
2661bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
2662  if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2663    return CFP->getValueAPF().isZero();
2664  }
2665  if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2666    return C->isNullValue();
2667  }
2668  return false;
2669}
2670
2671SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
2672                                                  const TargetRegisterClass *RC,
2673                                                   unsigned Reg, EVT VT) const {
2674  MachineFunction &MF = DAG.getMachineFunction();
2675  MachineRegisterInfo &MRI = MF.getRegInfo();
2676  unsigned VirtualRegister;
2677  if (!MRI.isLiveIn(Reg)) {
2678    VirtualRegister = MRI.createVirtualRegister(RC);
2679    MRI.addLiveIn(Reg, VirtualRegister);
2680  } else {
2681    VirtualRegister = MRI.getLiveInVirtReg(Reg);
2682  }
2683  return DAG.getRegister(VirtualRegister, VT);
2684}
2685
2686uint32_t AMDGPUTargetLowering::getImplicitParameterOffset(
2687    const AMDGPUMachineFunction *MFI, const ImplicitParameter Param) const {
2688  uint64_t ArgOffset = MFI->ABIArgOffset;
2689  switch (Param) {
2690  case GRID_DIM:
2691    return ArgOffset;
2692  case GRID_OFFSET:
2693    return ArgOffset + 4;
2694  }
2695  llvm_unreachable("unexpected implicit parameter type");
2696}
2697
2698#define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
2699
2700const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
2701  switch ((AMDGPUISD::NodeType)Opcode) {
2702  case AMDGPUISD::FIRST_NUMBER: break;
2703  // AMDIL DAG nodes
2704  NODE_NAME_CASE(CALL);
2705  NODE_NAME_CASE(UMUL);
2706  NODE_NAME_CASE(RET_FLAG);
2707  NODE_NAME_CASE(BRANCH_COND);
2708
2709  // AMDGPU DAG nodes
2710  NODE_NAME_CASE(DWORDADDR)
2711  NODE_NAME_CASE(FRACT)
2712  NODE_NAME_CASE(CLAMP)
2713  NODE_NAME_CASE(COS_HW)
2714  NODE_NAME_CASE(SIN_HW)
2715  NODE_NAME_CASE(FMAX_LEGACY)
2716  NODE_NAME_CASE(FMIN_LEGACY)
2717  NODE_NAME_CASE(FMAX3)
2718  NODE_NAME_CASE(SMAX3)
2719  NODE_NAME_CASE(UMAX3)
2720  NODE_NAME_CASE(FMIN3)
2721  NODE_NAME_CASE(SMIN3)
2722  NODE_NAME_CASE(UMIN3)
2723  NODE_NAME_CASE(URECIP)
2724  NODE_NAME_CASE(DIV_SCALE)
2725  NODE_NAME_CASE(DIV_FMAS)
2726  NODE_NAME_CASE(DIV_FIXUP)
2727  NODE_NAME_CASE(TRIG_PREOP)
2728  NODE_NAME_CASE(RCP)
2729  NODE_NAME_CASE(RSQ)
2730  NODE_NAME_CASE(RSQ_LEGACY)
2731  NODE_NAME_CASE(RSQ_CLAMPED)
2732  NODE_NAME_CASE(LDEXP)
2733  NODE_NAME_CASE(FP_CLASS)
2734  NODE_NAME_CASE(DOT4)
2735  NODE_NAME_CASE(CARRY)
2736  NODE_NAME_CASE(BORROW)
2737  NODE_NAME_CASE(BFE_U32)
2738  NODE_NAME_CASE(BFE_I32)
2739  NODE_NAME_CASE(BFI)
2740  NODE_NAME_CASE(BFM)
2741  NODE_NAME_CASE(BREV)
2742  NODE_NAME_CASE(MUL_U24)
2743  NODE_NAME_CASE(MUL_I24)
2744  NODE_NAME_CASE(MAD_U24)
2745  NODE_NAME_CASE(MAD_I24)
2746  NODE_NAME_CASE(TEXTURE_FETCH)
2747  NODE_NAME_CASE(EXPORT)
2748  NODE_NAME_CASE(CONST_ADDRESS)
2749  NODE_NAME_CASE(REGISTER_LOAD)
2750  NODE_NAME_CASE(REGISTER_STORE)
2751  NODE_NAME_CASE(LOAD_CONSTANT)
2752  NODE_NAME_CASE(LOAD_INPUT)
2753  NODE_NAME_CASE(SAMPLE)
2754  NODE_NAME_CASE(SAMPLEB)
2755  NODE_NAME_CASE(SAMPLED)
2756  NODE_NAME_CASE(SAMPLEL)
2757  NODE_NAME_CASE(CVT_F32_UBYTE0)
2758  NODE_NAME_CASE(CVT_F32_UBYTE1)
2759  NODE_NAME_CASE(CVT_F32_UBYTE2)
2760  NODE_NAME_CASE(CVT_F32_UBYTE3)
2761  NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
2762  NODE_NAME_CASE(CONST_DATA_PTR)
2763  case AMDGPUISD::FIRST_MEM_OPCODE_NUMBER: break;
2764  NODE_NAME_CASE(SENDMSG)
2765  NODE_NAME_CASE(INTERP_MOV)
2766  NODE_NAME_CASE(INTERP_P1)
2767  NODE_NAME_CASE(INTERP_P2)
2768  NODE_NAME_CASE(STORE_MSKOR)
2769  NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
2770  case AMDGPUISD::LAST_AMDGPU_ISD_NUMBER: break;
2771  }
2772  return nullptr;
2773}
2774
2775SDValue AMDGPUTargetLowering::getRsqrtEstimate(SDValue Operand,
2776                                               DAGCombinerInfo &DCI,
2777                                               unsigned &RefinementSteps,
2778                                               bool &UseOneConstNR) const {
2779  SelectionDAG &DAG = DCI.DAG;
2780  EVT VT = Operand.getValueType();
2781
2782  if (VT == MVT::f32) {
2783    RefinementSteps = 0;
2784    return DAG.getNode(AMDGPUISD::RSQ, SDLoc(Operand), VT, Operand);
2785  }
2786
2787  // TODO: There is also f64 rsq instruction, but the documentation is less
2788  // clear on its precision.
2789
2790  return SDValue();
2791}
2792
2793SDValue AMDGPUTargetLowering::getRecipEstimate(SDValue Operand,
2794                                               DAGCombinerInfo &DCI,
2795                                               unsigned &RefinementSteps) const {
2796  SelectionDAG &DAG = DCI.DAG;
2797  EVT VT = Operand.getValueType();
2798
2799  if (VT == MVT::f32) {
2800    // Reciprocal, < 1 ulp error.
2801    //
2802    // This reciprocal approximation converges to < 0.5 ulp error with one
2803    // newton rhapson performed with two fused multiple adds (FMAs).
2804
2805    RefinementSteps = 0;
2806    return DAG.getNode(AMDGPUISD::RCP, SDLoc(Operand), VT, Operand);
2807  }
2808
2809  // TODO: There is also f64 rcp instruction, but the documentation is less
2810  // clear on its precision.
2811
2812  return SDValue();
2813}
2814
2815static void computeKnownBitsForMinMax(const SDValue Op0,
2816                                      const SDValue Op1,
2817                                      APInt &KnownZero,
2818                                      APInt &KnownOne,
2819                                      const SelectionDAG &DAG,
2820                                      unsigned Depth) {
2821  APInt Op0Zero, Op0One;
2822  APInt Op1Zero, Op1One;
2823  DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
2824  DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
2825
2826  KnownZero = Op0Zero & Op1Zero;
2827  KnownOne = Op0One & Op1One;
2828}
2829
2830void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
2831  const SDValue Op,
2832  APInt &KnownZero,
2833  APInt &KnownOne,
2834  const SelectionDAG &DAG,
2835  unsigned Depth) const {
2836
2837  KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
2838
2839  APInt KnownZero2;
2840  APInt KnownOne2;
2841  unsigned Opc = Op.getOpcode();
2842
2843  switch (Opc) {
2844  default:
2845    break;
2846  case ISD::INTRINSIC_WO_CHAIN: {
2847    // FIXME: The intrinsic should just use the node.
2848    switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
2849    case AMDGPUIntrinsic::AMDGPU_imax:
2850    case AMDGPUIntrinsic::AMDGPU_umax:
2851    case AMDGPUIntrinsic::AMDGPU_imin:
2852    case AMDGPUIntrinsic::AMDGPU_umin:
2853      computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
2854                                KnownZero, KnownOne, DAG, Depth);
2855      break;
2856    default:
2857      break;
2858    }
2859
2860    break;
2861  }
2862  case AMDGPUISD::CARRY:
2863  case AMDGPUISD::BORROW: {
2864    KnownZero = APInt::getHighBitsSet(32, 31);
2865    break;
2866  }
2867
2868  case AMDGPUISD::BFE_I32:
2869  case AMDGPUISD::BFE_U32: {
2870    ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2871    if (!CWidth)
2872      return;
2873
2874    unsigned BitWidth = 32;
2875    uint32_t Width = CWidth->getZExtValue() & 0x1f;
2876
2877    if (Opc == AMDGPUISD::BFE_U32)
2878      KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2879
2880    break;
2881  }
2882  }
2883}
2884
2885unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
2886  SDValue Op,
2887  const SelectionDAG &DAG,
2888  unsigned Depth) const {
2889  switch (Op.getOpcode()) {
2890  case AMDGPUISD::BFE_I32: {
2891    ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2892    if (!Width)
2893      return 1;
2894
2895    unsigned SignBits = 32 - Width->getZExtValue() + 1;
2896    ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2897    if (!Offset || !Offset->isNullValue())
2898      return SignBits;
2899
2900    // TODO: Could probably figure something out with non-0 offsets.
2901    unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
2902    return std::max(SignBits, Op0SignBits);
2903  }
2904
2905  case AMDGPUISD::BFE_U32: {
2906    ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2907    return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;
2908  }
2909
2910  case AMDGPUISD::CARRY:
2911  case AMDGPUISD::BORROW:
2912    return 31;
2913
2914  default:
2915    return 1;
2916  }
2917}
2918