1 //===-- AMDGPUISelLowering.cpp - AMDGPU Common DAG lowering functions -----===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 /// \file
10 /// This is the parent TargetLowering class for hardware code gen
11 /// targets.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "AMDGPUISelLowering.h"
16 #include "AMDGPU.h"
17 #include "AMDGPUInstrInfo.h"
18 #include "AMDGPUMachineFunction.h"
19 #include "GCNSubtarget.h"
20 #include "SIMachineFunctionInfo.h"
21 #include "llvm/CodeGen/Analysis.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/IR/DiagnosticInfo.h"
24 #include "llvm/IR/IntrinsicsAMDGPU.h"
25 #include "llvm/Support/CommandLine.h"
26 #include "llvm/Support/KnownBits.h"
27 #include "llvm/Target/TargetMachine.h"
28 
29 using namespace llvm;
30 
31 #include "AMDGPUGenCallingConv.inc"
32 
33 static cl::opt<bool> AMDGPUBypassSlowDiv(
34   "amdgpu-bypass-slow-div",
35   cl::desc("Skip 64-bit divide for dynamic 32-bit values"),
36   cl::init(true));
37 
38 // Find a larger type to do a load / store of a vector with.
39 EVT AMDGPUTargetLowering::getEquivalentMemType(LLVMContext &Ctx, EVT VT) {
40   unsigned StoreSize = VT.getStoreSizeInBits();
41   if (StoreSize <= 32)
42     return EVT::getIntegerVT(Ctx, StoreSize);
43 
44   assert(StoreSize % 32 == 0 && "Store size not a multiple of 32");
45   return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
46 }
47 
48 unsigned AMDGPUTargetLowering::numBitsUnsigned(SDValue Op, SelectionDAG &DAG) {
49   return DAG.computeKnownBits(Op).countMaxActiveBits();
50 }
51 
52 unsigned AMDGPUTargetLowering::numBitsSigned(SDValue Op, SelectionDAG &DAG) {
53   // In order for this to be a signed 24-bit value, bit 23, must
54   // be a sign bit.
55   return DAG.ComputeMaxSignificantBits(Op);
56 }
57 
58 AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
59                                            const AMDGPUSubtarget &STI)
60     : TargetLowering(TM), Subtarget(&STI) {
61   // Lower floating point store/load to integer store/load to reduce the number
62   // of patterns in tablegen.
63   setOperationAction(ISD::LOAD, MVT::f32, Promote);
64   AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
65 
66   setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
67   AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
68 
69   setOperationAction(ISD::LOAD, MVT::v3f32, Promote);
70   AddPromotedToType(ISD::LOAD, MVT::v3f32, MVT::v3i32);
71 
72   setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
73   AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
74 
75   setOperationAction(ISD::LOAD, MVT::v5f32, Promote);
76   AddPromotedToType(ISD::LOAD, MVT::v5f32, MVT::v5i32);
77 
78   setOperationAction(ISD::LOAD, MVT::v6f32, Promote);
79   AddPromotedToType(ISD::LOAD, MVT::v6f32, MVT::v6i32);
80 
81   setOperationAction(ISD::LOAD, MVT::v7f32, Promote);
82   AddPromotedToType(ISD::LOAD, MVT::v7f32, MVT::v7i32);
83 
84   setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
85   AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
86 
87   setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
88   AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
89 
90   setOperationAction(ISD::LOAD, MVT::v32f32, Promote);
91   AddPromotedToType(ISD::LOAD, MVT::v32f32, MVT::v32i32);
92 
93   setOperationAction(ISD::LOAD, MVT::i64, Promote);
94   AddPromotedToType(ISD::LOAD, MVT::i64, MVT::v2i32);
95 
96   setOperationAction(ISD::LOAD, MVT::v2i64, Promote);
97   AddPromotedToType(ISD::LOAD, MVT::v2i64, MVT::v4i32);
98 
99   setOperationAction(ISD::LOAD, MVT::f64, Promote);
100   AddPromotedToType(ISD::LOAD, MVT::f64, MVT::v2i32);
101 
102   setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
103   AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v4i32);
104 
105   setOperationAction(ISD::LOAD, MVT::v3i64, Promote);
106   AddPromotedToType(ISD::LOAD, MVT::v3i64, MVT::v6i32);
107 
108   setOperationAction(ISD::LOAD, MVT::v4i64, Promote);
109   AddPromotedToType(ISD::LOAD, MVT::v4i64, MVT::v8i32);
110 
111   setOperationAction(ISD::LOAD, MVT::v3f64, Promote);
112   AddPromotedToType(ISD::LOAD, MVT::v3f64, MVT::v6i32);
113 
114   setOperationAction(ISD::LOAD, MVT::v4f64, Promote);
115   AddPromotedToType(ISD::LOAD, MVT::v4f64, MVT::v8i32);
116 
117   setOperationAction(ISD::LOAD, MVT::v8i64, Promote);
118   AddPromotedToType(ISD::LOAD, MVT::v8i64, MVT::v16i32);
119 
120   setOperationAction(ISD::LOAD, MVT::v8f64, Promote);
121   AddPromotedToType(ISD::LOAD, MVT::v8f64, MVT::v16i32);
122 
123   setOperationAction(ISD::LOAD, MVT::v16i64, Promote);
124   AddPromotedToType(ISD::LOAD, MVT::v16i64, MVT::v32i32);
125 
126   setOperationAction(ISD::LOAD, MVT::v16f64, Promote);
127   AddPromotedToType(ISD::LOAD, MVT::v16f64, MVT::v32i32);
128 
129   // There are no 64-bit extloads. These should be done as a 32-bit extload and
130   // an extension to 64-bit.
131   for (MVT VT : MVT::integer_valuetypes())
132     setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, MVT::i64, VT,
133                      Expand);
134 
135   for (MVT VT : MVT::integer_valuetypes()) {
136     if (VT == MVT::i64)
137       continue;
138 
139     for (auto Op : {ISD::SEXTLOAD, ISD::ZEXTLOAD, ISD::EXTLOAD}) {
140       setLoadExtAction(Op, VT, MVT::i1, Promote);
141       setLoadExtAction(Op, VT, MVT::i8, Legal);
142       setLoadExtAction(Op, VT, MVT::i16, Legal);
143       setLoadExtAction(Op, VT, MVT::i32, Expand);
144     }
145   }
146 
147   for (MVT VT : MVT::integer_fixedlen_vector_valuetypes())
148     for (auto MemVT :
149          {MVT::v2i8, MVT::v4i8, MVT::v2i16, MVT::v3i16, MVT::v4i16})
150       setLoadExtAction({ISD::SEXTLOAD, ISD::ZEXTLOAD, ISD::EXTLOAD}, VT, MemVT,
151                        Expand);
152 
153   setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
154   setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, MVT::v2f16, Expand);
155   setLoadExtAction(ISD::EXTLOAD, MVT::v3f32, MVT::v3f16, Expand);
156   setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, MVT::v4f16, Expand);
157   setLoadExtAction(ISD::EXTLOAD, MVT::v8f32, MVT::v8f16, Expand);
158   setLoadExtAction(ISD::EXTLOAD, MVT::v16f32, MVT::v16f16, Expand);
159   setLoadExtAction(ISD::EXTLOAD, MVT::v32f32, MVT::v32f16, Expand);
160 
161   setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
162   setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f32, Expand);
163   setLoadExtAction(ISD::EXTLOAD, MVT::v3f64, MVT::v3f32, Expand);
164   setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f32, Expand);
165   setLoadExtAction(ISD::EXTLOAD, MVT::v8f64, MVT::v8f32, Expand);
166   setLoadExtAction(ISD::EXTLOAD, MVT::v16f64, MVT::v16f32, Expand);
167 
168   setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
169   setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f16, Expand);
170   setLoadExtAction(ISD::EXTLOAD, MVT::v3f64, MVT::v3f16, Expand);
171   setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f16, Expand);
172   setLoadExtAction(ISD::EXTLOAD, MVT::v8f64, MVT::v8f16, Expand);
173   setLoadExtAction(ISD::EXTLOAD, MVT::v16f64, MVT::v16f16, Expand);
174 
175   setOperationAction(ISD::STORE, MVT::f32, Promote);
176   AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
177 
178   setOperationAction(ISD::STORE, MVT::v2f32, Promote);
179   AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
180 
181   setOperationAction(ISD::STORE, MVT::v3f32, Promote);
182   AddPromotedToType(ISD::STORE, MVT::v3f32, MVT::v3i32);
183 
184   setOperationAction(ISD::STORE, MVT::v4f32, Promote);
185   AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
186 
187   setOperationAction(ISD::STORE, MVT::v5f32, Promote);
188   AddPromotedToType(ISD::STORE, MVT::v5f32, MVT::v5i32);
189 
190   setOperationAction(ISD::STORE, MVT::v6f32, Promote);
191   AddPromotedToType(ISD::STORE, MVT::v6f32, MVT::v6i32);
192 
193   setOperationAction(ISD::STORE, MVT::v7f32, Promote);
194   AddPromotedToType(ISD::STORE, MVT::v7f32, MVT::v7i32);
195 
196   setOperationAction(ISD::STORE, MVT::v8f32, Promote);
197   AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
198 
199   setOperationAction(ISD::STORE, MVT::v16f32, Promote);
200   AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
201 
202   setOperationAction(ISD::STORE, MVT::v32f32, Promote);
203   AddPromotedToType(ISD::STORE, MVT::v32f32, MVT::v32i32);
204 
205   setOperationAction(ISD::STORE, MVT::i64, Promote);
206   AddPromotedToType(ISD::STORE, MVT::i64, MVT::v2i32);
207 
208   setOperationAction(ISD::STORE, MVT::v2i64, Promote);
209   AddPromotedToType(ISD::STORE, MVT::v2i64, MVT::v4i32);
210 
211   setOperationAction(ISD::STORE, MVT::f64, Promote);
212   AddPromotedToType(ISD::STORE, MVT::f64, MVT::v2i32);
213 
214   setOperationAction(ISD::STORE, MVT::v2f64, Promote);
215   AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v4i32);
216 
217   setOperationAction(ISD::STORE, MVT::v3i64, Promote);
218   AddPromotedToType(ISD::STORE, MVT::v3i64, MVT::v6i32);
219 
220   setOperationAction(ISD::STORE, MVT::v3f64, Promote);
221   AddPromotedToType(ISD::STORE, MVT::v3f64, MVT::v6i32);
222 
223   setOperationAction(ISD::STORE, MVT::v4i64, Promote);
224   AddPromotedToType(ISD::STORE, MVT::v4i64, MVT::v8i32);
225 
226   setOperationAction(ISD::STORE, MVT::v4f64, Promote);
227   AddPromotedToType(ISD::STORE, MVT::v4f64, MVT::v8i32);
228 
229   setOperationAction(ISD::STORE, MVT::v8i64, Promote);
230   AddPromotedToType(ISD::STORE, MVT::v8i64, MVT::v16i32);
231 
232   setOperationAction(ISD::STORE, MVT::v8f64, Promote);
233   AddPromotedToType(ISD::STORE, MVT::v8f64, MVT::v16i32);
234 
235   setOperationAction(ISD::STORE, MVT::v16i64, Promote);
236   AddPromotedToType(ISD::STORE, MVT::v16i64, MVT::v32i32);
237 
238   setOperationAction(ISD::STORE, MVT::v16f64, Promote);
239   AddPromotedToType(ISD::STORE, MVT::v16f64, MVT::v32i32);
240 
241   setTruncStoreAction(MVT::i64, MVT::i1, Expand);
242   setTruncStoreAction(MVT::i64, MVT::i8, Expand);
243   setTruncStoreAction(MVT::i64, MVT::i16, Expand);
244   setTruncStoreAction(MVT::i64, MVT::i32, Expand);
245 
246   setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
247   setTruncStoreAction(MVT::v2i64, MVT::v2i8, Expand);
248   setTruncStoreAction(MVT::v2i64, MVT::v2i16, Expand);
249   setTruncStoreAction(MVT::v2i64, MVT::v2i32, Expand);
250 
251   setTruncStoreAction(MVT::f32, MVT::f16, Expand);
252   setTruncStoreAction(MVT::v2f32, MVT::v2f16, Expand);
253   setTruncStoreAction(MVT::v3f32, MVT::v3f16, Expand);
254   setTruncStoreAction(MVT::v4f32, MVT::v4f16, Expand);
255   setTruncStoreAction(MVT::v8f32, MVT::v8f16, Expand);
256   setTruncStoreAction(MVT::v16f32, MVT::v16f16, Expand);
257   setTruncStoreAction(MVT::v32f32, MVT::v32f16, Expand);
258 
259   setTruncStoreAction(MVT::f64, MVT::f16, Expand);
260   setTruncStoreAction(MVT::f64, MVT::f32, Expand);
261 
262   setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
263   setTruncStoreAction(MVT::v2f64, MVT::v2f16, Expand);
264 
265   setTruncStoreAction(MVT::v3i64, MVT::v3i32, Expand);
266   setTruncStoreAction(MVT::v3i64, MVT::v3i16, Expand);
267   setTruncStoreAction(MVT::v3f64, MVT::v3f32, Expand);
268   setTruncStoreAction(MVT::v3f64, MVT::v3f16, Expand);
269 
270   setTruncStoreAction(MVT::v4i64, MVT::v4i32, Expand);
271   setTruncStoreAction(MVT::v4i64, MVT::v4i16, Expand);
272   setTruncStoreAction(MVT::v4f64, MVT::v4f32, Expand);
273   setTruncStoreAction(MVT::v4f64, MVT::v4f16, Expand);
274 
275   setTruncStoreAction(MVT::v8f64, MVT::v8f32, Expand);
276   setTruncStoreAction(MVT::v8f64, MVT::v8f16, Expand);
277 
278   setTruncStoreAction(MVT::v16f64, MVT::v16f32, Expand);
279   setTruncStoreAction(MVT::v16f64, MVT::v16f16, Expand);
280   setTruncStoreAction(MVT::v16i64, MVT::v16i16, Expand);
281   setTruncStoreAction(MVT::v16i64, MVT::v16i16, Expand);
282   setTruncStoreAction(MVT::v16i64, MVT::v16i8, Expand);
283   setTruncStoreAction(MVT::v16i64, MVT::v16i8, Expand);
284   setTruncStoreAction(MVT::v16i64, MVT::v16i1, Expand);
285 
286   setOperationAction(ISD::Constant, {MVT::i32, MVT::i64}, Legal);
287   setOperationAction(ISD::ConstantFP, {MVT::f32, MVT::f64}, Legal);
288 
289   setOperationAction({ISD::BR_JT, ISD::BRIND}, MVT::Other, Expand);
290 
291   // This is totally unsupported, just custom lower to produce an error.
292   setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
293 
294   // Library functions.  These default to Expand, but we have instructions
295   // for them.
296   setOperationAction({ISD::FCEIL, ISD::FEXP2, ISD::FPOW, ISD::FLOG2, ISD::FABS,
297                       ISD::FFLOOR, ISD::FRINT, ISD::FTRUNC, ISD::FMINNUM,
298                       ISD::FMAXNUM},
299                      MVT::f32, Legal);
300 
301   setOperationAction(ISD::FROUND, {MVT::f32, MVT::f64}, Custom);
302 
303   setOperationAction({ISD::FLOG, ISD::FLOG10, ISD::FEXP}, MVT::f32, Custom);
304 
305   setOperationAction(ISD::FNEARBYINT, {MVT::f32, MVT::f64}, Custom);
306 
307   setOperationAction(ISD::FREM, {MVT::f16, MVT::f32, MVT::f64}, Custom);
308 
309   // Expand to fneg + fadd.
310   setOperationAction(ISD::FSUB, MVT::f64, Expand);
311 
312   setOperationAction(ISD::CONCAT_VECTORS,
313                      {MVT::v3i32, MVT::v3f32, MVT::v4i32, MVT::v4f32,
314                       MVT::v5i32, MVT::v5f32, MVT::v6i32, MVT::v6f32,
315                       MVT::v7i32, MVT::v7f32, MVT::v8i32, MVT::v8f32},
316                      Custom);
317   setOperationAction(ISD::EXTRACT_SUBVECTOR,
318                      {MVT::v2f16,  MVT::v2i16,  MVT::v4f16,  MVT::v4i16,
319                       MVT::v2f32,  MVT::v2i32,  MVT::v3f32,  MVT::v3i32,
320                       MVT::v4f32,  MVT::v4i32,  MVT::v5f32,  MVT::v5i32,
321                       MVT::v6f32,  MVT::v6i32,  MVT::v7f32,  MVT::v7i32,
322                       MVT::v8f32,  MVT::v8i32,  MVT::v16f32, MVT::v16i32,
323                       MVT::v32f32, MVT::v32i32, MVT::v2f64,  MVT::v2i64,
324                       MVT::v3f64,  MVT::v3i64,  MVT::v4f64,  MVT::v4i64,
325                       MVT::v8f64,  MVT::v8i64,  MVT::v16f64, MVT::v16i64},
326                      Custom);
327 
328   setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
329   setOperationAction(ISD::FP_TO_FP16, {MVT::f64, MVT::f32}, Custom);
330 
331   const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
332   for (MVT VT : ScalarIntVTs) {
333     // These should use [SU]DIVREM, so set them to expand
334     setOperationAction({ISD::SDIV, ISD::UDIV, ISD::SREM, ISD::UREM}, VT,
335                        Expand);
336 
337     // GPU does not have divrem function for signed or unsigned.
338     setOperationAction({ISD::SDIVREM, ISD::UDIVREM}, VT, Custom);
339 
340     // GPU does not have [S|U]MUL_LOHI functions as a single instruction.
341     setOperationAction({ISD::SMUL_LOHI, ISD::UMUL_LOHI}, VT, Expand);
342 
343     setOperationAction({ISD::BSWAP, ISD::CTTZ, ISD::CTLZ}, VT, Expand);
344 
345     // AMDGPU uses ADDC/SUBC/ADDE/SUBE
346     setOperationAction({ISD::ADDC, ISD::SUBC, ISD::ADDE, ISD::SUBE}, VT, Legal);
347   }
348 
349   // The hardware supports 32-bit FSHR, but not FSHL.
350   setOperationAction(ISD::FSHR, MVT::i32, Legal);
351 
352   // The hardware supports 32-bit ROTR, but not ROTL.
353   setOperationAction(ISD::ROTL, {MVT::i32, MVT::i64}, Expand);
354   setOperationAction(ISD::ROTR, MVT::i64, Expand);
355 
356   setOperationAction({ISD::MULHU, ISD::MULHS}, MVT::i16, Expand);
357 
358   setOperationAction({ISD::MUL, ISD::MULHU, ISD::MULHS}, MVT::i64, Expand);
359   setOperationAction(
360       {ISD::UINT_TO_FP, ISD::SINT_TO_FP, ISD::FP_TO_SINT, ISD::FP_TO_UINT},
361       MVT::i64, Custom);
362   setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
363 
364   setOperationAction({ISD::SMIN, ISD::UMIN, ISD::SMAX, ISD::UMAX}, MVT::i32,
365                      Legal);
366 
367   setOperationAction(
368       {ISD::CTTZ, ISD::CTTZ_ZERO_UNDEF, ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF},
369       MVT::i64, Custom);
370 
371   static const MVT::SimpleValueType VectorIntTypes[] = {
372       MVT::v2i32, MVT::v3i32, MVT::v4i32, MVT::v5i32, MVT::v6i32, MVT::v7i32};
373 
374   for (MVT VT : VectorIntTypes) {
375     // Expand the following operations for the current type by default.
376     setOperationAction({ISD::ADD,        ISD::AND,     ISD::FP_TO_SINT,
377                         ISD::FP_TO_UINT, ISD::MUL,     ISD::MULHU,
378                         ISD::MULHS,      ISD::OR,      ISD::SHL,
379                         ISD::SRA,        ISD::SRL,     ISD::ROTL,
380                         ISD::ROTR,       ISD::SUB,     ISD::SINT_TO_FP,
381                         ISD::UINT_TO_FP, ISD::SDIV,    ISD::UDIV,
382                         ISD::SREM,       ISD::UREM,    ISD::SMUL_LOHI,
383                         ISD::UMUL_LOHI,  ISD::SDIVREM, ISD::UDIVREM,
384                         ISD::SELECT,     ISD::VSELECT, ISD::SELECT_CC,
385                         ISD::XOR,        ISD::BSWAP,   ISD::CTPOP,
386                         ISD::CTTZ,       ISD::CTLZ,    ISD::VECTOR_SHUFFLE,
387                         ISD::SETCC},
388                        VT, Expand);
389   }
390 
391   static const MVT::SimpleValueType FloatVectorTypes[] = {
392       MVT::v2f32, MVT::v3f32, MVT::v4f32, MVT::v5f32, MVT::v6f32, MVT::v7f32};
393 
394   for (MVT VT : FloatVectorTypes) {
395     setOperationAction(
396         {ISD::FABS,    ISD::FMINNUM,      ISD::FMAXNUM,   ISD::FADD,
397          ISD::FCEIL,   ISD::FCOS,         ISD::FDIV,      ISD::FEXP2,
398          ISD::FEXP,    ISD::FLOG2,        ISD::FREM,      ISD::FLOG,
399          ISD::FLOG10,  ISD::FPOW,         ISD::FFLOOR,    ISD::FTRUNC,
400          ISD::FMUL,    ISD::FMA,          ISD::FRINT,     ISD::FNEARBYINT,
401          ISD::FSQRT,   ISD::FSIN,         ISD::FSUB,      ISD::FNEG,
402          ISD::VSELECT, ISD::SELECT_CC,    ISD::FCOPYSIGN, ISD::VECTOR_SHUFFLE,
403          ISD::SETCC,   ISD::FCANONICALIZE},
404         VT, Expand);
405   }
406 
407   // This causes using an unrolled select operation rather than expansion with
408   // bit operations. This is in general better, but the alternative using BFI
409   // instructions may be better if the select sources are SGPRs.
410   setOperationAction(ISD::SELECT, MVT::v2f32, Promote);
411   AddPromotedToType(ISD::SELECT, MVT::v2f32, MVT::v2i32);
412 
413   setOperationAction(ISD::SELECT, MVT::v3f32, Promote);
414   AddPromotedToType(ISD::SELECT, MVT::v3f32, MVT::v3i32);
415 
416   setOperationAction(ISD::SELECT, MVT::v4f32, Promote);
417   AddPromotedToType(ISD::SELECT, MVT::v4f32, MVT::v4i32);
418 
419   setOperationAction(ISD::SELECT, MVT::v5f32, Promote);
420   AddPromotedToType(ISD::SELECT, MVT::v5f32, MVT::v5i32);
421 
422   setOperationAction(ISD::SELECT, MVT::v6f32, Promote);
423   AddPromotedToType(ISD::SELECT, MVT::v6f32, MVT::v6i32);
424 
425   setOperationAction(ISD::SELECT, MVT::v7f32, Promote);
426   AddPromotedToType(ISD::SELECT, MVT::v7f32, MVT::v7i32);
427 
428   // There are no libcalls of any kind.
429   for (int I = 0; I < RTLIB::UNKNOWN_LIBCALL; ++I)
430     setLibcallName(static_cast<RTLIB::Libcall>(I), nullptr);
431 
432   setSchedulingPreference(Sched::RegPressure);
433   setJumpIsExpensive(true);
434 
435   // FIXME: This is only partially true. If we have to do vector compares, any
436   // SGPR pair can be a condition register. If we have a uniform condition, we
437   // are better off doing SALU operations, where there is only one SCC. For now,
438   // we don't have a way of knowing during instruction selection if a condition
439   // will be uniform and we always use vector compares. Assume we are using
440   // vector compares until that is fixed.
441   setHasMultipleConditionRegisters(true);
442 
443   setMinCmpXchgSizeInBits(32);
444   setSupportsUnalignedAtomics(false);
445 
446   PredictableSelectIsExpensive = false;
447 
448   // We want to find all load dependencies for long chains of stores to enable
449   // merging into very wide vectors. The problem is with vectors with > 4
450   // elements. MergeConsecutiveStores will attempt to merge these because x8/x16
451   // vectors are a legal type, even though we have to split the loads
452   // usually. When we can more precisely specify load legality per address
453   // space, we should be able to make FindBetterChain/MergeConsecutiveStores
454   // smarter so that they can figure out what to do in 2 iterations without all
455   // N > 4 stores on the same chain.
456   GatherAllAliasesMaxDepth = 16;
457 
458   // memcpy/memmove/memset are expanded in the IR, so we shouldn't need to worry
459   // about these during lowering.
460   MaxStoresPerMemcpy  = 0xffffffff;
461   MaxStoresPerMemmove = 0xffffffff;
462   MaxStoresPerMemset  = 0xffffffff;
463 
464   // The expansion for 64-bit division is enormous.
465   if (AMDGPUBypassSlowDiv)
466     addBypassSlowDiv(64, 32);
467 
468   setTargetDAGCombine({ISD::BITCAST,    ISD::SHL,
469                        ISD::SRA,        ISD::SRL,
470                        ISD::TRUNCATE,   ISD::MUL,
471                        ISD::SMUL_LOHI,  ISD::UMUL_LOHI,
472                        ISD::MULHU,      ISD::MULHS,
473                        ISD::SELECT,     ISD::SELECT_CC,
474                        ISD::STORE,      ISD::FADD,
475                        ISD::FSUB,       ISD::FNEG,
476                        ISD::FABS,       ISD::AssertZext,
477                        ISD::AssertSext, ISD::INTRINSIC_WO_CHAIN});
478 }
479 
480 bool AMDGPUTargetLowering::mayIgnoreSignedZero(SDValue Op) const {
481   if (getTargetMachine().Options.NoSignedZerosFPMath)
482     return true;
483 
484   const auto Flags = Op.getNode()->getFlags();
485   if (Flags.hasNoSignedZeros())
486     return true;
487 
488   return false;
489 }
490 
491 //===----------------------------------------------------------------------===//
492 // Target Information
493 //===----------------------------------------------------------------------===//
494 
495 LLVM_READNONE
496 static bool fnegFoldsIntoOp(unsigned Opc) {
497   switch (Opc) {
498   case ISD::FADD:
499   case ISD::FSUB:
500   case ISD::FMUL:
501   case ISD::FMA:
502   case ISD::FMAD:
503   case ISD::FMINNUM:
504   case ISD::FMAXNUM:
505   case ISD::FMINNUM_IEEE:
506   case ISD::FMAXNUM_IEEE:
507   case ISD::FSIN:
508   case ISD::FTRUNC:
509   case ISD::FRINT:
510   case ISD::FNEARBYINT:
511   case ISD::FCANONICALIZE:
512   case AMDGPUISD::RCP:
513   case AMDGPUISD::RCP_LEGACY:
514   case AMDGPUISD::RCP_IFLAG:
515   case AMDGPUISD::SIN_HW:
516   case AMDGPUISD::FMUL_LEGACY:
517   case AMDGPUISD::FMIN_LEGACY:
518   case AMDGPUISD::FMAX_LEGACY:
519   case AMDGPUISD::FMED3:
520     // TODO: handle llvm.amdgcn.fma.legacy
521     return true;
522   default:
523     return false;
524   }
525 }
526 
527 /// \p returns true if the operation will definitely need to use a 64-bit
528 /// encoding, and thus will use a VOP3 encoding regardless of the source
529 /// modifiers.
530 LLVM_READONLY
531 static bool opMustUseVOP3Encoding(const SDNode *N, MVT VT) {
532   return N->getNumOperands() > 2 || VT == MVT::f64;
533 }
534 
535 // Most FP instructions support source modifiers, but this could be refined
536 // slightly.
537 LLVM_READONLY
538 static bool hasSourceMods(const SDNode *N) {
539   if (isa<MemSDNode>(N))
540     return false;
541 
542   switch (N->getOpcode()) {
543   case ISD::CopyToReg:
544   case ISD::SELECT:
545   case ISD::FDIV:
546   case ISD::FREM:
547   case ISD::INLINEASM:
548   case ISD::INLINEASM_BR:
549   case AMDGPUISD::DIV_SCALE:
550   case ISD::INTRINSIC_W_CHAIN:
551 
552   // TODO: Should really be looking at the users of the bitcast. These are
553   // problematic because bitcasts are used to legalize all stores to integer
554   // types.
555   case ISD::BITCAST:
556     return false;
557   case ISD::INTRINSIC_WO_CHAIN: {
558     switch (cast<ConstantSDNode>(N->getOperand(0))->getZExtValue()) {
559     case Intrinsic::amdgcn_interp_p1:
560     case Intrinsic::amdgcn_interp_p2:
561     case Intrinsic::amdgcn_interp_mov:
562     case Intrinsic::amdgcn_interp_p1_f16:
563     case Intrinsic::amdgcn_interp_p2_f16:
564       return false;
565     default:
566       return true;
567     }
568   }
569   default:
570     return true;
571   }
572 }
573 
574 bool AMDGPUTargetLowering::allUsesHaveSourceMods(const SDNode *N,
575                                                  unsigned CostThreshold) {
576   // Some users (such as 3-operand FMA/MAD) must use a VOP3 encoding, and thus
577   // it is truly free to use a source modifier in all cases. If there are
578   // multiple users but for each one will necessitate using VOP3, there will be
579   // a code size increase. Try to avoid increasing code size unless we know it
580   // will save on the instruction count.
581   unsigned NumMayIncreaseSize = 0;
582   MVT VT = N->getValueType(0).getScalarType().getSimpleVT();
583 
584   // XXX - Should this limit number of uses to check?
585   for (const SDNode *U : N->uses()) {
586     if (!hasSourceMods(U))
587       return false;
588 
589     if (!opMustUseVOP3Encoding(U, VT)) {
590       if (++NumMayIncreaseSize > CostThreshold)
591         return false;
592     }
593   }
594 
595   return true;
596 }
597 
598 EVT AMDGPUTargetLowering::getTypeForExtReturn(LLVMContext &Context, EVT VT,
599                                               ISD::NodeType ExtendKind) const {
600   assert(!VT.isVector() && "only scalar expected");
601 
602   // Round to the next multiple of 32-bits.
603   unsigned Size = VT.getSizeInBits();
604   if (Size <= 32)
605     return MVT::i32;
606   return EVT::getIntegerVT(Context, 32 * ((Size + 31) / 32));
607 }
608 
609 MVT AMDGPUTargetLowering::getVectorIdxTy(const DataLayout &) const {
610   return MVT::i32;
611 }
612 
613 bool AMDGPUTargetLowering::isSelectSupported(SelectSupportKind SelType) const {
614   return true;
615 }
616 
617 // The backend supports 32 and 64 bit floating point immediates.
618 // FIXME: Why are we reporting vectors of FP immediates as legal?
619 bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT,
620                                         bool ForCodeSize) const {
621   EVT ScalarVT = VT.getScalarType();
622   return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64 ||
623          (ScalarVT == MVT::f16 && Subtarget->has16BitInsts()));
624 }
625 
626 // We don't want to shrink f64 / f32 constants.
627 bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
628   EVT ScalarVT = VT.getScalarType();
629   return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
630 }
631 
632 bool AMDGPUTargetLowering::shouldReduceLoadWidth(SDNode *N,
633                                                  ISD::LoadExtType ExtTy,
634                                                  EVT NewVT) const {
635   // TODO: This may be worth removing. Check regression tests for diffs.
636   if (!TargetLoweringBase::shouldReduceLoadWidth(N, ExtTy, NewVT))
637     return false;
638 
639   unsigned NewSize = NewVT.getStoreSizeInBits();
640 
641   // If we are reducing to a 32-bit load or a smaller multi-dword load,
642   // this is always better.
643   if (NewSize >= 32)
644     return true;
645 
646   EVT OldVT = N->getValueType(0);
647   unsigned OldSize = OldVT.getStoreSizeInBits();
648 
649   MemSDNode *MN = cast<MemSDNode>(N);
650   unsigned AS = MN->getAddressSpace();
651   // Do not shrink an aligned scalar load to sub-dword.
652   // Scalar engine cannot do sub-dword loads.
653   if (OldSize >= 32 && NewSize < 32 && MN->getAlign() >= Align(4) &&
654       (AS == AMDGPUAS::CONSTANT_ADDRESS ||
655        AS == AMDGPUAS::CONSTANT_ADDRESS_32BIT ||
656        (isa<LoadSDNode>(N) && AS == AMDGPUAS::GLOBAL_ADDRESS &&
657         MN->isInvariant())) &&
658       AMDGPUInstrInfo::isUniformMMO(MN->getMemOperand()))
659     return false;
660 
661   // Don't produce extloads from sub 32-bit types. SI doesn't have scalar
662   // extloads, so doing one requires using a buffer_load. In cases where we
663   // still couldn't use a scalar load, using the wider load shouldn't really
664   // hurt anything.
665 
666   // If the old size already had to be an extload, there's no harm in continuing
667   // to reduce the width.
668   return (OldSize < 32);
669 }
670 
671 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy, EVT CastTy,
672                                                    const SelectionDAG &DAG,
673                                                    const MachineMemOperand &MMO) const {
674 
675   assert(LoadTy.getSizeInBits() == CastTy.getSizeInBits());
676 
677   if (LoadTy.getScalarType() == MVT::i32)
678     return false;
679 
680   unsigned LScalarSize = LoadTy.getScalarSizeInBits();
681   unsigned CastScalarSize = CastTy.getScalarSizeInBits();
682 
683   if ((LScalarSize >= CastScalarSize) && (CastScalarSize < 32))
684     return false;
685 
686   bool Fast = false;
687   return allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
688                                         CastTy, MMO, &Fast) &&
689          Fast;
690 }
691 
692 // SI+ has instructions for cttz / ctlz for 32-bit values. This is probably also
693 // profitable with the expansion for 64-bit since it's generally good to
694 // speculate things.
695 // FIXME: These should really have the size as a parameter.
696 bool AMDGPUTargetLowering::isCheapToSpeculateCttz() const {
697   return true;
698 }
699 
700 bool AMDGPUTargetLowering::isCheapToSpeculateCtlz() const {
701   return true;
702 }
703 
704 bool AMDGPUTargetLowering::isSDNodeAlwaysUniform(const SDNode *N) const {
705   switch (N->getOpcode()) {
706   case ISD::EntryToken:
707   case ISD::TokenFactor:
708     return true;
709   case ISD::INTRINSIC_WO_CHAIN: {
710     unsigned IntrID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
711     switch (IntrID) {
712     case Intrinsic::amdgcn_readfirstlane:
713     case Intrinsic::amdgcn_readlane:
714       return true;
715     }
716     return false;
717   }
718   case ISD::LOAD:
719     if (cast<LoadSDNode>(N)->getMemOperand()->getAddrSpace() ==
720         AMDGPUAS::CONSTANT_ADDRESS_32BIT)
721       return true;
722     return false;
723   case AMDGPUISD::SETCC: // ballot-style instruction
724     return true;
725   }
726   return false;
727 }
728 
729 SDValue AMDGPUTargetLowering::getNegatedExpression(
730     SDValue Op, SelectionDAG &DAG, bool LegalOperations, bool ForCodeSize,
731     NegatibleCost &Cost, unsigned Depth) const {
732 
733   switch (Op.getOpcode()) {
734   case ISD::FMA:
735   case ISD::FMAD: {
736     // Negating a fma is not free if it has users without source mods.
737     if (!allUsesHaveSourceMods(Op.getNode()))
738       return SDValue();
739     break;
740   }
741   default:
742     break;
743   }
744 
745   return TargetLowering::getNegatedExpression(Op, DAG, LegalOperations,
746                                               ForCodeSize, Cost, Depth);
747 }
748 
749 //===---------------------------------------------------------------------===//
750 // Target Properties
751 //===---------------------------------------------------------------------===//
752 
753 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
754   assert(VT.isFloatingPoint());
755 
756   // Packed operations do not have a fabs modifier.
757   return VT == MVT::f32 || VT == MVT::f64 ||
758          (Subtarget->has16BitInsts() && VT == MVT::f16);
759 }
760 
761 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
762   assert(VT.isFloatingPoint());
763   // Report this based on the end legalized type.
764   VT = VT.getScalarType();
765   return VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f16;
766 }
767 
768 bool AMDGPUTargetLowering:: storeOfVectorConstantIsCheap(EVT MemVT,
769                                                          unsigned NumElem,
770                                                          unsigned AS) const {
771   return true;
772 }
773 
774 bool AMDGPUTargetLowering::aggressivelyPreferBuildVectorSources(EVT VecVT) const {
775   // There are few operations which truly have vector input operands. Any vector
776   // operation is going to involve operations on each component, and a
777   // build_vector will be a copy per element, so it always makes sense to use a
778   // build_vector input in place of the extracted element to avoid a copy into a
779   // super register.
780   //
781   // We should probably only do this if all users are extracts only, but this
782   // should be the common case.
783   return true;
784 }
785 
786 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
787   // Truncate is just accessing a subregister.
788 
789   unsigned SrcSize = Source.getSizeInBits();
790   unsigned DestSize = Dest.getSizeInBits();
791 
792   return DestSize < SrcSize && DestSize % 32 == 0 ;
793 }
794 
795 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
796   // Truncate is just accessing a subregister.
797 
798   unsigned SrcSize = Source->getScalarSizeInBits();
799   unsigned DestSize = Dest->getScalarSizeInBits();
800 
801   if (DestSize== 16 && Subtarget->has16BitInsts())
802     return SrcSize >= 32;
803 
804   return DestSize < SrcSize && DestSize % 32 == 0;
805 }
806 
807 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
808   unsigned SrcSize = Src->getScalarSizeInBits();
809   unsigned DestSize = Dest->getScalarSizeInBits();
810 
811   if (SrcSize == 16 && Subtarget->has16BitInsts())
812     return DestSize >= 32;
813 
814   return SrcSize == 32 && DestSize == 64;
815 }
816 
817 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
818   // Any register load of a 64-bit value really requires 2 32-bit moves. For all
819   // practical purposes, the extra mov 0 to load a 64-bit is free.  As used,
820   // this will enable reducing 64-bit operations the 32-bit, which is always
821   // good.
822 
823   if (Src == MVT::i16)
824     return Dest == MVT::i32 ||Dest == MVT::i64 ;
825 
826   return Src == MVT::i32 && Dest == MVT::i64;
827 }
828 
829 bool AMDGPUTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
830   return isZExtFree(Val.getValueType(), VT2);
831 }
832 
833 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
834   // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
835   // limited number of native 64-bit operations. Shrinking an operation to fit
836   // in a single 32-bit register should always be helpful. As currently used,
837   // this is much less general than the name suggests, and is only used in
838   // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
839   // not profitable, and may actually be harmful.
840   return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
841 }
842 
843 //===---------------------------------------------------------------------===//
844 // TargetLowering Callbacks
845 //===---------------------------------------------------------------------===//
846 
847 CCAssignFn *AMDGPUCallLowering::CCAssignFnForCall(CallingConv::ID CC,
848                                                   bool IsVarArg) {
849   switch (CC) {
850   case CallingConv::AMDGPU_VS:
851   case CallingConv::AMDGPU_GS:
852   case CallingConv::AMDGPU_PS:
853   case CallingConv::AMDGPU_CS:
854   case CallingConv::AMDGPU_HS:
855   case CallingConv::AMDGPU_ES:
856   case CallingConv::AMDGPU_LS:
857     return CC_AMDGPU;
858   case CallingConv::C:
859   case CallingConv::Fast:
860   case CallingConv::Cold:
861     return CC_AMDGPU_Func;
862   case CallingConv::AMDGPU_Gfx:
863     return CC_SI_Gfx;
864   case CallingConv::AMDGPU_KERNEL:
865   case CallingConv::SPIR_KERNEL:
866   default:
867     report_fatal_error("Unsupported calling convention for call");
868   }
869 }
870 
871 CCAssignFn *AMDGPUCallLowering::CCAssignFnForReturn(CallingConv::ID CC,
872                                                     bool IsVarArg) {
873   switch (CC) {
874   case CallingConv::AMDGPU_KERNEL:
875   case CallingConv::SPIR_KERNEL:
876     llvm_unreachable("kernels should not be handled here");
877   case CallingConv::AMDGPU_VS:
878   case CallingConv::AMDGPU_GS:
879   case CallingConv::AMDGPU_PS:
880   case CallingConv::AMDGPU_CS:
881   case CallingConv::AMDGPU_HS:
882   case CallingConv::AMDGPU_ES:
883   case CallingConv::AMDGPU_LS:
884     return RetCC_SI_Shader;
885   case CallingConv::AMDGPU_Gfx:
886     return RetCC_SI_Gfx;
887   case CallingConv::C:
888   case CallingConv::Fast:
889   case CallingConv::Cold:
890     return RetCC_AMDGPU_Func;
891   default:
892     report_fatal_error("Unsupported calling convention.");
893   }
894 }
895 
896 /// The SelectionDAGBuilder will automatically promote function arguments
897 /// with illegal types.  However, this does not work for the AMDGPU targets
898 /// since the function arguments are stored in memory as these illegal types.
899 /// In order to handle this properly we need to get the original types sizes
900 /// from the LLVM IR Function and fixup the ISD:InputArg values before
901 /// passing them to AnalyzeFormalArguments()
902 
903 /// When the SelectionDAGBuilder computes the Ins, it takes care of splitting
904 /// input values across multiple registers.  Each item in the Ins array
905 /// represents a single value that will be stored in registers.  Ins[x].VT is
906 /// the value type of the value that will be stored in the register, so
907 /// whatever SDNode we lower the argument to needs to be this type.
908 ///
909 /// In order to correctly lower the arguments we need to know the size of each
910 /// argument.  Since Ins[x].VT gives us the size of the register that will
911 /// hold the value, we need to look at Ins[x].ArgVT to see the 'real' type
912 /// for the original function argument so that we can deduce the correct memory
913 /// type to use for Ins[x].  In most cases the correct memory type will be
914 /// Ins[x].ArgVT.  However, this will not always be the case.  If, for example,
915 /// we have a kernel argument of type v8i8, this argument will be split into
916 /// 8 parts and each part will be represented by its own item in the Ins array.
917 /// For each part the Ins[x].ArgVT will be the v8i8, which is the full type of
918 /// the argument before it was split.  From this, we deduce that the memory type
919 /// for each individual part is i8.  We pass the memory type as LocVT to the
920 /// calling convention analysis function and the register type (Ins[x].VT) as
921 /// the ValVT.
922 void AMDGPUTargetLowering::analyzeFormalArgumentsCompute(
923   CCState &State,
924   const SmallVectorImpl<ISD::InputArg> &Ins) const {
925   const MachineFunction &MF = State.getMachineFunction();
926   const Function &Fn = MF.getFunction();
927   LLVMContext &Ctx = Fn.getParent()->getContext();
928   const AMDGPUSubtarget &ST = AMDGPUSubtarget::get(MF);
929   const unsigned ExplicitOffset = ST.getExplicitKernelArgOffset(Fn);
930   CallingConv::ID CC = Fn.getCallingConv();
931 
932   Align MaxAlign = Align(1);
933   uint64_t ExplicitArgOffset = 0;
934   const DataLayout &DL = Fn.getParent()->getDataLayout();
935 
936   unsigned InIndex = 0;
937 
938   for (const Argument &Arg : Fn.args()) {
939     const bool IsByRef = Arg.hasByRefAttr();
940     Type *BaseArgTy = Arg.getType();
941     Type *MemArgTy = IsByRef ? Arg.getParamByRefType() : BaseArgTy;
942     Align Alignment = DL.getValueOrABITypeAlignment(
943         IsByRef ? Arg.getParamAlign() : None, MemArgTy);
944     MaxAlign = std::max(Alignment, MaxAlign);
945     uint64_t AllocSize = DL.getTypeAllocSize(MemArgTy);
946 
947     uint64_t ArgOffset = alignTo(ExplicitArgOffset, Alignment) + ExplicitOffset;
948     ExplicitArgOffset = alignTo(ExplicitArgOffset, Alignment) + AllocSize;
949 
950     // We're basically throwing away everything passed into us and starting over
951     // to get accurate in-memory offsets. The "PartOffset" is completely useless
952     // to us as computed in Ins.
953     //
954     // We also need to figure out what type legalization is trying to do to get
955     // the correct memory offsets.
956 
957     SmallVector<EVT, 16> ValueVTs;
958     SmallVector<uint64_t, 16> Offsets;
959     ComputeValueVTs(*this, DL, BaseArgTy, ValueVTs, &Offsets, ArgOffset);
960 
961     for (unsigned Value = 0, NumValues = ValueVTs.size();
962          Value != NumValues; ++Value) {
963       uint64_t BasePartOffset = Offsets[Value];
964 
965       EVT ArgVT = ValueVTs[Value];
966       EVT MemVT = ArgVT;
967       MVT RegisterVT = getRegisterTypeForCallingConv(Ctx, CC, ArgVT);
968       unsigned NumRegs = getNumRegistersForCallingConv(Ctx, CC, ArgVT);
969 
970       if (NumRegs == 1) {
971         // This argument is not split, so the IR type is the memory type.
972         if (ArgVT.isExtended()) {
973           // We have an extended type, like i24, so we should just use the
974           // register type.
975           MemVT = RegisterVT;
976         } else {
977           MemVT = ArgVT;
978         }
979       } else if (ArgVT.isVector() && RegisterVT.isVector() &&
980                  ArgVT.getScalarType() == RegisterVT.getScalarType()) {
981         assert(ArgVT.getVectorNumElements() > RegisterVT.getVectorNumElements());
982         // We have a vector value which has been split into a vector with
983         // the same scalar type, but fewer elements.  This should handle
984         // all the floating-point vector types.
985         MemVT = RegisterVT;
986       } else if (ArgVT.isVector() &&
987                  ArgVT.getVectorNumElements() == NumRegs) {
988         // This arg has been split so that each element is stored in a separate
989         // register.
990         MemVT = ArgVT.getScalarType();
991       } else if (ArgVT.isExtended()) {
992         // We have an extended type, like i65.
993         MemVT = RegisterVT;
994       } else {
995         unsigned MemoryBits = ArgVT.getStoreSizeInBits() / NumRegs;
996         assert(ArgVT.getStoreSizeInBits() % NumRegs == 0);
997         if (RegisterVT.isInteger()) {
998           MemVT = EVT::getIntegerVT(State.getContext(), MemoryBits);
999         } else if (RegisterVT.isVector()) {
1000           assert(!RegisterVT.getScalarType().isFloatingPoint());
1001           unsigned NumElements = RegisterVT.getVectorNumElements();
1002           assert(MemoryBits % NumElements == 0);
1003           // This vector type has been split into another vector type with
1004           // a different elements size.
1005           EVT ScalarVT = EVT::getIntegerVT(State.getContext(),
1006                                            MemoryBits / NumElements);
1007           MemVT = EVT::getVectorVT(State.getContext(), ScalarVT, NumElements);
1008         } else {
1009           llvm_unreachable("cannot deduce memory type.");
1010         }
1011       }
1012 
1013       // Convert one element vectors to scalar.
1014       if (MemVT.isVector() && MemVT.getVectorNumElements() == 1)
1015         MemVT = MemVT.getScalarType();
1016 
1017       // Round up vec3/vec5 argument.
1018       if (MemVT.isVector() && !MemVT.isPow2VectorType()) {
1019         assert(MemVT.getVectorNumElements() == 3 ||
1020                MemVT.getVectorNumElements() == 5);
1021         MemVT = MemVT.getPow2VectorType(State.getContext());
1022       } else if (!MemVT.isSimple() && !MemVT.isVector()) {
1023         MemVT = MemVT.getRoundIntegerType(State.getContext());
1024       }
1025 
1026       unsigned PartOffset = 0;
1027       for (unsigned i = 0; i != NumRegs; ++i) {
1028         State.addLoc(CCValAssign::getCustomMem(InIndex++, RegisterVT,
1029                                                BasePartOffset + PartOffset,
1030                                                MemVT.getSimpleVT(),
1031                                                CCValAssign::Full));
1032         PartOffset += MemVT.getStoreSize();
1033       }
1034     }
1035   }
1036 }
1037 
1038 SDValue AMDGPUTargetLowering::LowerReturn(
1039   SDValue Chain, CallingConv::ID CallConv,
1040   bool isVarArg,
1041   const SmallVectorImpl<ISD::OutputArg> &Outs,
1042   const SmallVectorImpl<SDValue> &OutVals,
1043   const SDLoc &DL, SelectionDAG &DAG) const {
1044   // FIXME: Fails for r600 tests
1045   //assert(!isVarArg && Outs.empty() && OutVals.empty() &&
1046   // "wave terminate should not have return values");
1047   return DAG.getNode(AMDGPUISD::ENDPGM, DL, MVT::Other, Chain);
1048 }
1049 
1050 //===---------------------------------------------------------------------===//
1051 // Target specific lowering
1052 //===---------------------------------------------------------------------===//
1053 
1054 /// Selects the correct CCAssignFn for a given CallingConvention value.
1055 CCAssignFn *AMDGPUTargetLowering::CCAssignFnForCall(CallingConv::ID CC,
1056                                                     bool IsVarArg) {
1057   return AMDGPUCallLowering::CCAssignFnForCall(CC, IsVarArg);
1058 }
1059 
1060 CCAssignFn *AMDGPUTargetLowering::CCAssignFnForReturn(CallingConv::ID CC,
1061                                                       bool IsVarArg) {
1062   return AMDGPUCallLowering::CCAssignFnForReturn(CC, IsVarArg);
1063 }
1064 
1065 SDValue AMDGPUTargetLowering::addTokenForArgument(SDValue Chain,
1066                                                   SelectionDAG &DAG,
1067                                                   MachineFrameInfo &MFI,
1068                                                   int ClobberedFI) const {
1069   SmallVector<SDValue, 8> ArgChains;
1070   int64_t FirstByte = MFI.getObjectOffset(ClobberedFI);
1071   int64_t LastByte = FirstByte + MFI.getObjectSize(ClobberedFI) - 1;
1072 
1073   // Include the original chain at the beginning of the list. When this is
1074   // used by target LowerCall hooks, this helps legalize find the
1075   // CALLSEQ_BEGIN node.
1076   ArgChains.push_back(Chain);
1077 
1078   // Add a chain value for each stack argument corresponding
1079   for (SDNode *U : DAG.getEntryNode().getNode()->uses()) {
1080     if (LoadSDNode *L = dyn_cast<LoadSDNode>(U)) {
1081       if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr())) {
1082         if (FI->getIndex() < 0) {
1083           int64_t InFirstByte = MFI.getObjectOffset(FI->getIndex());
1084           int64_t InLastByte = InFirstByte;
1085           InLastByte += MFI.getObjectSize(FI->getIndex()) - 1;
1086 
1087           if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
1088               (FirstByte <= InFirstByte && InFirstByte <= LastByte))
1089             ArgChains.push_back(SDValue(L, 1));
1090         }
1091       }
1092     }
1093   }
1094 
1095   // Build a tokenfactor for all the chains.
1096   return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
1097 }
1098 
1099 SDValue AMDGPUTargetLowering::lowerUnhandledCall(CallLoweringInfo &CLI,
1100                                                  SmallVectorImpl<SDValue> &InVals,
1101                                                  StringRef Reason) const {
1102   SDValue Callee = CLI.Callee;
1103   SelectionDAG &DAG = CLI.DAG;
1104 
1105   const Function &Fn = DAG.getMachineFunction().getFunction();
1106 
1107   StringRef FuncName("<unknown>");
1108 
1109   if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
1110     FuncName = G->getSymbol();
1111   else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1112     FuncName = G->getGlobal()->getName();
1113 
1114   DiagnosticInfoUnsupported NoCalls(
1115     Fn, Reason + FuncName, CLI.DL.getDebugLoc());
1116   DAG.getContext()->diagnose(NoCalls);
1117 
1118   if (!CLI.IsTailCall) {
1119     for (unsigned I = 0, E = CLI.Ins.size(); I != E; ++I)
1120       InVals.push_back(DAG.getUNDEF(CLI.Ins[I].VT));
1121   }
1122 
1123   return DAG.getEntryNode();
1124 }
1125 
1126 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
1127                                         SmallVectorImpl<SDValue> &InVals) const {
1128   return lowerUnhandledCall(CLI, InVals, "unsupported call to function ");
1129 }
1130 
1131 SDValue AMDGPUTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1132                                                       SelectionDAG &DAG) const {
1133   const Function &Fn = DAG.getMachineFunction().getFunction();
1134 
1135   DiagnosticInfoUnsupported NoDynamicAlloca(Fn, "unsupported dynamic alloca",
1136                                             SDLoc(Op).getDebugLoc());
1137   DAG.getContext()->diagnose(NoDynamicAlloca);
1138   auto Ops = {DAG.getConstant(0, SDLoc(), Op.getValueType()), Op.getOperand(0)};
1139   return DAG.getMergeValues(Ops, SDLoc());
1140 }
1141 
1142 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
1143                                              SelectionDAG &DAG) const {
1144   switch (Op.getOpcode()) {
1145   default:
1146     Op->print(errs(), &DAG);
1147     llvm_unreachable("Custom lowering code for this "
1148                      "instruction is not implemented yet!");
1149     break;
1150   case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
1151   case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
1152   case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
1153   case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
1154   case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
1155   case ISD::FREM: return LowerFREM(Op, DAG);
1156   case ISD::FCEIL: return LowerFCEIL(Op, DAG);
1157   case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
1158   case ISD::FRINT: return LowerFRINT(Op, DAG);
1159   case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
1160   case ISD::FROUND: return LowerFROUND(Op, DAG);
1161   case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
1162   case ISD::FLOG:
1163     return LowerFLOG(Op, DAG, numbers::ln2f);
1164   case ISD::FLOG10:
1165     return LowerFLOG(Op, DAG, numbers::ln2f / numbers::ln10f);
1166   case ISD::FEXP:
1167     return lowerFEXP(Op, DAG);
1168   case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
1169   case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
1170   case ISD::FP_TO_FP16: return LowerFP_TO_FP16(Op, DAG);
1171   case ISD::FP_TO_SINT:
1172   case ISD::FP_TO_UINT:
1173     return LowerFP_TO_INT(Op, DAG);
1174   case ISD::CTTZ:
1175   case ISD::CTTZ_ZERO_UNDEF:
1176   case ISD::CTLZ:
1177   case ISD::CTLZ_ZERO_UNDEF:
1178     return LowerCTLZ_CTTZ(Op, DAG);
1179   case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
1180   }
1181   return Op;
1182 }
1183 
1184 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
1185                                               SmallVectorImpl<SDValue> &Results,
1186                                               SelectionDAG &DAG) const {
1187   switch (N->getOpcode()) {
1188   case ISD::SIGN_EXTEND_INREG:
1189     // Different parts of legalization seem to interpret which type of
1190     // sign_extend_inreg is the one to check for custom lowering. The extended
1191     // from type is what really matters, but some places check for custom
1192     // lowering of the result type. This results in trying to use
1193     // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
1194     // nothing here and let the illegal result integer be handled normally.
1195     return;
1196   default:
1197     return;
1198   }
1199 }
1200 
1201 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
1202                                                  SDValue Op,
1203                                                  SelectionDAG &DAG) const {
1204 
1205   const DataLayout &DL = DAG.getDataLayout();
1206   GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
1207   const GlobalValue *GV = G->getGlobal();
1208 
1209   if (G->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1210       G->getAddressSpace() == AMDGPUAS::REGION_ADDRESS) {
1211     if (!MFI->isModuleEntryFunction() &&
1212         !GV->getName().equals("llvm.amdgcn.module.lds")) {
1213       SDLoc DL(Op);
1214       const Function &Fn = DAG.getMachineFunction().getFunction();
1215       DiagnosticInfoUnsupported BadLDSDecl(
1216         Fn, "local memory global used by non-kernel function",
1217         DL.getDebugLoc(), DS_Warning);
1218       DAG.getContext()->diagnose(BadLDSDecl);
1219 
1220       // We currently don't have a way to correctly allocate LDS objects that
1221       // aren't directly associated with a kernel. We do force inlining of
1222       // functions that use local objects. However, if these dead functions are
1223       // not eliminated, we don't want a compile time error. Just emit a warning
1224       // and a trap, since there should be no callable path here.
1225       SDValue Trap = DAG.getNode(ISD::TRAP, DL, MVT::Other, DAG.getEntryNode());
1226       SDValue OutputChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
1227                                         Trap, DAG.getRoot());
1228       DAG.setRoot(OutputChain);
1229       return DAG.getUNDEF(Op.getValueType());
1230     }
1231 
1232     // XXX: What does the value of G->getOffset() mean?
1233     assert(G->getOffset() == 0 &&
1234          "Do not know what to do with an non-zero offset");
1235 
1236     // TODO: We could emit code to handle the initialization somewhere.
1237     // We ignore the initializer for now and legalize it to allow selection.
1238     // The initializer will anyway get errored out during assembly emission.
1239     unsigned Offset = MFI->allocateLDSGlobal(DL, *cast<GlobalVariable>(GV));
1240     return DAG.getConstant(Offset, SDLoc(Op), Op.getValueType());
1241   }
1242   return SDValue();
1243 }
1244 
1245 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
1246                                                   SelectionDAG &DAG) const {
1247   SmallVector<SDValue, 8> Args;
1248 
1249   EVT VT = Op.getValueType();
1250   if (VT == MVT::v4i16 || VT == MVT::v4f16) {
1251     SDLoc SL(Op);
1252     SDValue Lo = DAG.getNode(ISD::BITCAST, SL, MVT::i32, Op.getOperand(0));
1253     SDValue Hi = DAG.getNode(ISD::BITCAST, SL, MVT::i32, Op.getOperand(1));
1254 
1255     SDValue BV = DAG.getBuildVector(MVT::v2i32, SL, { Lo, Hi });
1256     return DAG.getNode(ISD::BITCAST, SL, VT, BV);
1257   }
1258 
1259   for (const SDUse &U : Op->ops())
1260     DAG.ExtractVectorElements(U.get(), Args);
1261 
1262   return DAG.getBuildVector(Op.getValueType(), SDLoc(Op), Args);
1263 }
1264 
1265 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
1266                                                      SelectionDAG &DAG) const {
1267 
1268   SmallVector<SDValue, 8> Args;
1269   unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1270   EVT VT = Op.getValueType();
1271   EVT SrcVT = Op.getOperand(0).getValueType();
1272 
1273   // For these types, we have some TableGen patterns except if the index is 1
1274   if (((SrcVT == MVT::v4f16 && VT == MVT::v2f16) ||
1275        (SrcVT == MVT::v4i16 && VT == MVT::v2i16)) &&
1276       Start != 1)
1277     return Op;
1278 
1279   if (((SrcVT == MVT::v8f16 && VT == MVT::v4f16) ||
1280        (SrcVT == MVT::v8i16 && VT == MVT::v4i16)) &&
1281       (Start == 0 || Start == 4))
1282     return Op;
1283 
1284   DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
1285                             VT.getVectorNumElements());
1286 
1287   return DAG.getBuildVector(Op.getValueType(), SDLoc(Op), Args);
1288 }
1289 
1290 /// Generate Min/Max node
1291 SDValue AMDGPUTargetLowering::combineFMinMaxLegacy(const SDLoc &DL, EVT VT,
1292                                                    SDValue LHS, SDValue RHS,
1293                                                    SDValue True, SDValue False,
1294                                                    SDValue CC,
1295                                                    DAGCombinerInfo &DCI) const {
1296   if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
1297     return SDValue();
1298 
1299   SelectionDAG &DAG = DCI.DAG;
1300   ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1301   switch (CCOpcode) {
1302   case ISD::SETOEQ:
1303   case ISD::SETONE:
1304   case ISD::SETUNE:
1305   case ISD::SETNE:
1306   case ISD::SETUEQ:
1307   case ISD::SETEQ:
1308   case ISD::SETFALSE:
1309   case ISD::SETFALSE2:
1310   case ISD::SETTRUE:
1311   case ISD::SETTRUE2:
1312   case ISD::SETUO:
1313   case ISD::SETO:
1314     break;
1315   case ISD::SETULE:
1316   case ISD::SETULT: {
1317     if (LHS == True)
1318       return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1319     return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1320   }
1321   case ISD::SETOLE:
1322   case ISD::SETOLT:
1323   case ISD::SETLE:
1324   case ISD::SETLT: {
1325     // Ordered. Assume ordered for undefined.
1326 
1327     // Only do this after legalization to avoid interfering with other combines
1328     // which might occur.
1329     if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
1330         !DCI.isCalledByLegalizer())
1331       return SDValue();
1332 
1333     // We need to permute the operands to get the correct NaN behavior. The
1334     // selected operand is the second one based on the failing compare with NaN,
1335     // so permute it based on the compare type the hardware uses.
1336     if (LHS == True)
1337       return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1338     return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1339   }
1340   case ISD::SETUGE:
1341   case ISD::SETUGT: {
1342     if (LHS == True)
1343       return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1344     return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1345   }
1346   case ISD::SETGT:
1347   case ISD::SETGE:
1348   case ISD::SETOGE:
1349   case ISD::SETOGT: {
1350     if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
1351         !DCI.isCalledByLegalizer())
1352       return SDValue();
1353 
1354     if (LHS == True)
1355       return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1356     return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1357   }
1358   case ISD::SETCC_INVALID:
1359     llvm_unreachable("Invalid setcc condcode!");
1360   }
1361   return SDValue();
1362 }
1363 
1364 std::pair<SDValue, SDValue>
1365 AMDGPUTargetLowering::split64BitValue(SDValue Op, SelectionDAG &DAG) const {
1366   SDLoc SL(Op);
1367 
1368   SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
1369 
1370   const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
1371   const SDValue One = DAG.getConstant(1, SL, MVT::i32);
1372 
1373   SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero);
1374   SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One);
1375 
1376   return std::make_pair(Lo, Hi);
1377 }
1378 
1379 SDValue AMDGPUTargetLowering::getLoHalf64(SDValue Op, SelectionDAG &DAG) const {
1380   SDLoc SL(Op);
1381 
1382   SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
1383   const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
1384   return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero);
1385 }
1386 
1387 SDValue AMDGPUTargetLowering::getHiHalf64(SDValue Op, SelectionDAG &DAG) const {
1388   SDLoc SL(Op);
1389 
1390   SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
1391   const SDValue One = DAG.getConstant(1, SL, MVT::i32);
1392   return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One);
1393 }
1394 
1395 // Split a vector type into two parts. The first part is a power of two vector.
1396 // The second part is whatever is left over, and is a scalar if it would
1397 // otherwise be a 1-vector.
1398 std::pair<EVT, EVT>
1399 AMDGPUTargetLowering::getSplitDestVTs(const EVT &VT, SelectionDAG &DAG) const {
1400   EVT LoVT, HiVT;
1401   EVT EltVT = VT.getVectorElementType();
1402   unsigned NumElts = VT.getVectorNumElements();
1403   unsigned LoNumElts = PowerOf2Ceil((NumElts + 1) / 2);
1404   LoVT = EVT::getVectorVT(*DAG.getContext(), EltVT, LoNumElts);
1405   HiVT = NumElts - LoNumElts == 1
1406              ? EltVT
1407              : EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts - LoNumElts);
1408   return std::make_pair(LoVT, HiVT);
1409 }
1410 
1411 // Split a vector value into two parts of types LoVT and HiVT. HiVT could be
1412 // scalar.
1413 std::pair<SDValue, SDValue>
1414 AMDGPUTargetLowering::splitVector(const SDValue &N, const SDLoc &DL,
1415                                   const EVT &LoVT, const EVT &HiVT,
1416                                   SelectionDAG &DAG) const {
1417   assert(LoVT.getVectorNumElements() +
1418                  (HiVT.isVector() ? HiVT.getVectorNumElements() : 1) <=
1419              N.getValueType().getVectorNumElements() &&
1420          "More vector elements requested than available!");
1421   SDValue Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, LoVT, N,
1422                            DAG.getVectorIdxConstant(0, DL));
1423   SDValue Hi = DAG.getNode(
1424       HiVT.isVector() ? ISD::EXTRACT_SUBVECTOR : ISD::EXTRACT_VECTOR_ELT, DL,
1425       HiVT, N, DAG.getVectorIdxConstant(LoVT.getVectorNumElements(), DL));
1426   return std::make_pair(Lo, Hi);
1427 }
1428 
1429 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue Op,
1430                                               SelectionDAG &DAG) const {
1431   LoadSDNode *Load = cast<LoadSDNode>(Op);
1432   EVT VT = Op.getValueType();
1433   SDLoc SL(Op);
1434 
1435 
1436   // If this is a 2 element vector, we really want to scalarize and not create
1437   // weird 1 element vectors.
1438   if (VT.getVectorNumElements() == 2) {
1439     SDValue Ops[2];
1440     std::tie(Ops[0], Ops[1]) = scalarizeVectorLoad(Load, DAG);
1441     return DAG.getMergeValues(Ops, SL);
1442   }
1443 
1444   SDValue BasePtr = Load->getBasePtr();
1445   EVT MemVT = Load->getMemoryVT();
1446 
1447   const MachinePointerInfo &SrcValue = Load->getMemOperand()->getPointerInfo();
1448 
1449   EVT LoVT, HiVT;
1450   EVT LoMemVT, HiMemVT;
1451   SDValue Lo, Hi;
1452 
1453   std::tie(LoVT, HiVT) = getSplitDestVTs(VT, DAG);
1454   std::tie(LoMemVT, HiMemVT) = getSplitDestVTs(MemVT, DAG);
1455   std::tie(Lo, Hi) = splitVector(Op, SL, LoVT, HiVT, DAG);
1456 
1457   unsigned Size = LoMemVT.getStoreSize();
1458   Align BaseAlign = Load->getAlign();
1459   Align HiAlign = commonAlignment(BaseAlign, Size);
1460 
1461   SDValue LoLoad = DAG.getExtLoad(Load->getExtensionType(), SL, LoVT,
1462                                   Load->getChain(), BasePtr, SrcValue, LoMemVT,
1463                                   BaseAlign, Load->getMemOperand()->getFlags());
1464   SDValue HiPtr = DAG.getObjectPtrOffset(SL, BasePtr, TypeSize::Fixed(Size));
1465   SDValue HiLoad =
1466       DAG.getExtLoad(Load->getExtensionType(), SL, HiVT, Load->getChain(),
1467                      HiPtr, SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1468                      HiMemVT, HiAlign, Load->getMemOperand()->getFlags());
1469 
1470   SDValue Join;
1471   if (LoVT == HiVT) {
1472     // This is the case that the vector is power of two so was evenly split.
1473     Join = DAG.getNode(ISD::CONCAT_VECTORS, SL, VT, LoLoad, HiLoad);
1474   } else {
1475     Join = DAG.getNode(ISD::INSERT_SUBVECTOR, SL, VT, DAG.getUNDEF(VT), LoLoad,
1476                        DAG.getVectorIdxConstant(0, SL));
1477     Join = DAG.getNode(
1478         HiVT.isVector() ? ISD::INSERT_SUBVECTOR : ISD::INSERT_VECTOR_ELT, SL,
1479         VT, Join, HiLoad,
1480         DAG.getVectorIdxConstant(LoVT.getVectorNumElements(), SL));
1481   }
1482 
1483   SDValue Ops[] = {Join, DAG.getNode(ISD::TokenFactor, SL, MVT::Other,
1484                                      LoLoad.getValue(1), HiLoad.getValue(1))};
1485 
1486   return DAG.getMergeValues(Ops, SL);
1487 }
1488 
1489 SDValue AMDGPUTargetLowering::WidenOrSplitVectorLoad(SDValue Op,
1490                                                      SelectionDAG &DAG) const {
1491   LoadSDNode *Load = cast<LoadSDNode>(Op);
1492   EVT VT = Op.getValueType();
1493   SDValue BasePtr = Load->getBasePtr();
1494   EVT MemVT = Load->getMemoryVT();
1495   SDLoc SL(Op);
1496   const MachinePointerInfo &SrcValue = Load->getMemOperand()->getPointerInfo();
1497   Align BaseAlign = Load->getAlign();
1498   unsigned NumElements = MemVT.getVectorNumElements();
1499 
1500   // Widen from vec3 to vec4 when the load is at least 8-byte aligned
1501   // or 16-byte fully dereferenceable. Otherwise, split the vector load.
1502   if (NumElements != 3 ||
1503       (BaseAlign < Align(8) &&
1504        !SrcValue.isDereferenceable(16, *DAG.getContext(), DAG.getDataLayout())))
1505     return SplitVectorLoad(Op, DAG);
1506 
1507   assert(NumElements == 3);
1508 
1509   EVT WideVT =
1510       EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), 4);
1511   EVT WideMemVT =
1512       EVT::getVectorVT(*DAG.getContext(), MemVT.getVectorElementType(), 4);
1513   SDValue WideLoad = DAG.getExtLoad(
1514       Load->getExtensionType(), SL, WideVT, Load->getChain(), BasePtr, SrcValue,
1515       WideMemVT, BaseAlign, Load->getMemOperand()->getFlags());
1516   return DAG.getMergeValues(
1517       {DAG.getNode(ISD::EXTRACT_SUBVECTOR, SL, VT, WideLoad,
1518                    DAG.getVectorIdxConstant(0, SL)),
1519        WideLoad.getValue(1)},
1520       SL);
1521 }
1522 
1523 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1524                                                SelectionDAG &DAG) const {
1525   StoreSDNode *Store = cast<StoreSDNode>(Op);
1526   SDValue Val = Store->getValue();
1527   EVT VT = Val.getValueType();
1528 
1529   // If this is a 2 element vector, we really want to scalarize and not create
1530   // weird 1 element vectors.
1531   if (VT.getVectorNumElements() == 2)
1532     return scalarizeVectorStore(Store, DAG);
1533 
1534   EVT MemVT = Store->getMemoryVT();
1535   SDValue Chain = Store->getChain();
1536   SDValue BasePtr = Store->getBasePtr();
1537   SDLoc SL(Op);
1538 
1539   EVT LoVT, HiVT;
1540   EVT LoMemVT, HiMemVT;
1541   SDValue Lo, Hi;
1542 
1543   std::tie(LoVT, HiVT) = getSplitDestVTs(VT, DAG);
1544   std::tie(LoMemVT, HiMemVT) = getSplitDestVTs(MemVT, DAG);
1545   std::tie(Lo, Hi) = splitVector(Val, SL, LoVT, HiVT, DAG);
1546 
1547   SDValue HiPtr = DAG.getObjectPtrOffset(SL, BasePtr, LoMemVT.getStoreSize());
1548 
1549   const MachinePointerInfo &SrcValue = Store->getMemOperand()->getPointerInfo();
1550   Align BaseAlign = Store->getAlign();
1551   unsigned Size = LoMemVT.getStoreSize();
1552   Align HiAlign = commonAlignment(BaseAlign, Size);
1553 
1554   SDValue LoStore =
1555       DAG.getTruncStore(Chain, SL, Lo, BasePtr, SrcValue, LoMemVT, BaseAlign,
1556                         Store->getMemOperand()->getFlags());
1557   SDValue HiStore =
1558       DAG.getTruncStore(Chain, SL, Hi, HiPtr, SrcValue.getWithOffset(Size),
1559                         HiMemVT, HiAlign, Store->getMemOperand()->getFlags());
1560 
1561   return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoStore, HiStore);
1562 }
1563 
1564 // This is a shortcut for integer division because we have fast i32<->f32
1565 // conversions, and fast f32 reciprocal instructions. The fractional part of a
1566 // float is enough to accurately represent up to a 24-bit signed integer.
1567 SDValue AMDGPUTargetLowering::LowerDIVREM24(SDValue Op, SelectionDAG &DAG,
1568                                             bool Sign) const {
1569   SDLoc DL(Op);
1570   EVT VT = Op.getValueType();
1571   SDValue LHS = Op.getOperand(0);
1572   SDValue RHS = Op.getOperand(1);
1573   MVT IntVT = MVT::i32;
1574   MVT FltVT = MVT::f32;
1575 
1576   unsigned LHSSignBits = DAG.ComputeNumSignBits(LHS);
1577   if (LHSSignBits < 9)
1578     return SDValue();
1579 
1580   unsigned RHSSignBits = DAG.ComputeNumSignBits(RHS);
1581   if (RHSSignBits < 9)
1582     return SDValue();
1583 
1584   unsigned BitSize = VT.getSizeInBits();
1585   unsigned SignBits = std::min(LHSSignBits, RHSSignBits);
1586   unsigned DivBits = BitSize - SignBits;
1587   if (Sign)
1588     ++DivBits;
1589 
1590   ISD::NodeType ToFp = Sign ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
1591   ISD::NodeType ToInt = Sign ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;
1592 
1593   SDValue jq = DAG.getConstant(1, DL, IntVT);
1594 
1595   if (Sign) {
1596     // char|short jq = ia ^ ib;
1597     jq = DAG.getNode(ISD::XOR, DL, VT, LHS, RHS);
1598 
1599     // jq = jq >> (bitsize - 2)
1600     jq = DAG.getNode(ISD::SRA, DL, VT, jq,
1601                      DAG.getConstant(BitSize - 2, DL, VT));
1602 
1603     // jq = jq | 0x1
1604     jq = DAG.getNode(ISD::OR, DL, VT, jq, DAG.getConstant(1, DL, VT));
1605   }
1606 
1607   // int ia = (int)LHS;
1608   SDValue ia = LHS;
1609 
1610   // int ib, (int)RHS;
1611   SDValue ib = RHS;
1612 
1613   // float fa = (float)ia;
1614   SDValue fa = DAG.getNode(ToFp, DL, FltVT, ia);
1615 
1616   // float fb = (float)ib;
1617   SDValue fb = DAG.getNode(ToFp, DL, FltVT, ib);
1618 
1619   SDValue fq = DAG.getNode(ISD::FMUL, DL, FltVT,
1620                            fa, DAG.getNode(AMDGPUISD::RCP, DL, FltVT, fb));
1621 
1622   // fq = trunc(fq);
1623   fq = DAG.getNode(ISD::FTRUNC, DL, FltVT, fq);
1624 
1625   // float fqneg = -fq;
1626   SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FltVT, fq);
1627 
1628   MachineFunction &MF = DAG.getMachineFunction();
1629   const AMDGPUMachineFunction *MFI = MF.getInfo<AMDGPUMachineFunction>();
1630 
1631   // float fr = mad(fqneg, fb, fa);
1632   unsigned OpCode = !Subtarget->hasMadMacF32Insts() ?
1633                     (unsigned)ISD::FMA :
1634                     !MFI->getMode().allFP32Denormals() ?
1635                     (unsigned)ISD::FMAD :
1636                     (unsigned)AMDGPUISD::FMAD_FTZ;
1637   SDValue fr = DAG.getNode(OpCode, DL, FltVT, fqneg, fb, fa);
1638 
1639   // int iq = (int)fq;
1640   SDValue iq = DAG.getNode(ToInt, DL, IntVT, fq);
1641 
1642   // fr = fabs(fr);
1643   fr = DAG.getNode(ISD::FABS, DL, FltVT, fr);
1644 
1645   // fb = fabs(fb);
1646   fb = DAG.getNode(ISD::FABS, DL, FltVT, fb);
1647 
1648   EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
1649 
1650   // int cv = fr >= fb;
1651   SDValue cv = DAG.getSetCC(DL, SetCCVT, fr, fb, ISD::SETOGE);
1652 
1653   // jq = (cv ? jq : 0);
1654   jq = DAG.getNode(ISD::SELECT, DL, VT, cv, jq, DAG.getConstant(0, DL, VT));
1655 
1656   // dst = iq + jq;
1657   SDValue Div = DAG.getNode(ISD::ADD, DL, VT, iq, jq);
1658 
1659   // Rem needs compensation, it's easier to recompute it
1660   SDValue Rem = DAG.getNode(ISD::MUL, DL, VT, Div, RHS);
1661   Rem = DAG.getNode(ISD::SUB, DL, VT, LHS, Rem);
1662 
1663   // Truncate to number of bits this divide really is.
1664   if (Sign) {
1665     SDValue InRegSize
1666       = DAG.getValueType(EVT::getIntegerVT(*DAG.getContext(), DivBits));
1667     Div = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Div, InRegSize);
1668     Rem = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Rem, InRegSize);
1669   } else {
1670     SDValue TruncMask = DAG.getConstant((UINT64_C(1) << DivBits) - 1, DL, VT);
1671     Div = DAG.getNode(ISD::AND, DL, VT, Div, TruncMask);
1672     Rem = DAG.getNode(ISD::AND, DL, VT, Rem, TruncMask);
1673   }
1674 
1675   return DAG.getMergeValues({ Div, Rem }, DL);
1676 }
1677 
1678 void AMDGPUTargetLowering::LowerUDIVREM64(SDValue Op,
1679                                       SelectionDAG &DAG,
1680                                       SmallVectorImpl<SDValue> &Results) const {
1681   SDLoc DL(Op);
1682   EVT VT = Op.getValueType();
1683 
1684   assert(VT == MVT::i64 && "LowerUDIVREM64 expects an i64");
1685 
1686   EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
1687 
1688   SDValue One = DAG.getConstant(1, DL, HalfVT);
1689   SDValue Zero = DAG.getConstant(0, DL, HalfVT);
1690 
1691   //HiLo split
1692   SDValue LHS = Op.getOperand(0);
1693   SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero);
1694   SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, One);
1695 
1696   SDValue RHS = Op.getOperand(1);
1697   SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero);
1698   SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, One);
1699 
1700   if (DAG.MaskedValueIsZero(RHS, APInt::getHighBitsSet(64, 32)) &&
1701       DAG.MaskedValueIsZero(LHS, APInt::getHighBitsSet(64, 32))) {
1702 
1703     SDValue Res = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
1704                               LHS_Lo, RHS_Lo);
1705 
1706     SDValue DIV = DAG.getBuildVector(MVT::v2i32, DL, {Res.getValue(0), Zero});
1707     SDValue REM = DAG.getBuildVector(MVT::v2i32, DL, {Res.getValue(1), Zero});
1708 
1709     Results.push_back(DAG.getNode(ISD::BITCAST, DL, MVT::i64, DIV));
1710     Results.push_back(DAG.getNode(ISD::BITCAST, DL, MVT::i64, REM));
1711     return;
1712   }
1713 
1714   if (isTypeLegal(MVT::i64)) {
1715     // The algorithm here is based on ideas from "Software Integer Division",
1716     // Tom Rodeheffer, August 2008.
1717 
1718     MachineFunction &MF = DAG.getMachineFunction();
1719     const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1720 
1721     // Compute denominator reciprocal.
1722     unsigned FMAD = !Subtarget->hasMadMacF32Insts() ?
1723                     (unsigned)ISD::FMA :
1724                     !MFI->getMode().allFP32Denormals() ?
1725                     (unsigned)ISD::FMAD :
1726                     (unsigned)AMDGPUISD::FMAD_FTZ;
1727 
1728     SDValue Cvt_Lo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, RHS_Lo);
1729     SDValue Cvt_Hi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, RHS_Hi);
1730     SDValue Mad1 = DAG.getNode(FMAD, DL, MVT::f32, Cvt_Hi,
1731       DAG.getConstantFP(APInt(32, 0x4f800000).bitsToFloat(), DL, MVT::f32),
1732       Cvt_Lo);
1733     SDValue Rcp = DAG.getNode(AMDGPUISD::RCP, DL, MVT::f32, Mad1);
1734     SDValue Mul1 = DAG.getNode(ISD::FMUL, DL, MVT::f32, Rcp,
1735       DAG.getConstantFP(APInt(32, 0x5f7ffffc).bitsToFloat(), DL, MVT::f32));
1736     SDValue Mul2 = DAG.getNode(ISD::FMUL, DL, MVT::f32, Mul1,
1737       DAG.getConstantFP(APInt(32, 0x2f800000).bitsToFloat(), DL, MVT::f32));
1738     SDValue Trunc = DAG.getNode(ISD::FTRUNC, DL, MVT::f32, Mul2);
1739     SDValue Mad2 = DAG.getNode(FMAD, DL, MVT::f32, Trunc,
1740       DAG.getConstantFP(APInt(32, 0xcf800000).bitsToFloat(), DL, MVT::f32),
1741       Mul1);
1742     SDValue Rcp_Lo = DAG.getNode(ISD::FP_TO_UINT, DL, HalfVT, Mad2);
1743     SDValue Rcp_Hi = DAG.getNode(ISD::FP_TO_UINT, DL, HalfVT, Trunc);
1744     SDValue Rcp64 = DAG.getBitcast(VT,
1745                         DAG.getBuildVector(MVT::v2i32, DL, {Rcp_Lo, Rcp_Hi}));
1746 
1747     SDValue Zero64 = DAG.getConstant(0, DL, VT);
1748     SDValue One64  = DAG.getConstant(1, DL, VT);
1749     SDValue Zero1 = DAG.getConstant(0, DL, MVT::i1);
1750     SDVTList HalfCarryVT = DAG.getVTList(HalfVT, MVT::i1);
1751 
1752     // First round of UNR (Unsigned integer Newton-Raphson).
1753     SDValue Neg_RHS = DAG.getNode(ISD::SUB, DL, VT, Zero64, RHS);
1754     SDValue Mullo1 = DAG.getNode(ISD::MUL, DL, VT, Neg_RHS, Rcp64);
1755     SDValue Mulhi1 = DAG.getNode(ISD::MULHU, DL, VT, Rcp64, Mullo1);
1756     SDValue Mulhi1_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi1,
1757                                     Zero);
1758     SDValue Mulhi1_Hi =
1759         DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi1, One);
1760     SDValue Add1_Lo = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Rcp_Lo,
1761                                   Mulhi1_Lo, Zero1);
1762     SDValue Add1_Hi = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Rcp_Hi,
1763                                   Mulhi1_Hi, Add1_Lo.getValue(1));
1764     SDValue Add1 = DAG.getBitcast(VT,
1765                         DAG.getBuildVector(MVT::v2i32, DL, {Add1_Lo, Add1_Hi}));
1766 
1767     // Second round of UNR.
1768     SDValue Mullo2 = DAG.getNode(ISD::MUL, DL, VT, Neg_RHS, Add1);
1769     SDValue Mulhi2 = DAG.getNode(ISD::MULHU, DL, VT, Add1, Mullo2);
1770     SDValue Mulhi2_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi2,
1771                                     Zero);
1772     SDValue Mulhi2_Hi =
1773         DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi2, One);
1774     SDValue Add2_Lo = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Add1_Lo,
1775                                   Mulhi2_Lo, Zero1);
1776     SDValue Add2_Hi = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Add1_Hi,
1777                                   Mulhi2_Hi, Add2_Lo.getValue(1));
1778     SDValue Add2 = DAG.getBitcast(VT,
1779                         DAG.getBuildVector(MVT::v2i32, DL, {Add2_Lo, Add2_Hi}));
1780 
1781     SDValue Mulhi3 = DAG.getNode(ISD::MULHU, DL, VT, LHS, Add2);
1782 
1783     SDValue Mul3 = DAG.getNode(ISD::MUL, DL, VT, RHS, Mulhi3);
1784 
1785     SDValue Mul3_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mul3, Zero);
1786     SDValue Mul3_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mul3, One);
1787     SDValue Sub1_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, LHS_Lo,
1788                                   Mul3_Lo, Zero1);
1789     SDValue Sub1_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, LHS_Hi,
1790                                   Mul3_Hi, Sub1_Lo.getValue(1));
1791     SDValue Sub1_Mi = DAG.getNode(ISD::SUB, DL, HalfVT, LHS_Hi, Mul3_Hi);
1792     SDValue Sub1 = DAG.getBitcast(VT,
1793                         DAG.getBuildVector(MVT::v2i32, DL, {Sub1_Lo, Sub1_Hi}));
1794 
1795     SDValue MinusOne = DAG.getConstant(0xffffffffu, DL, HalfVT);
1796     SDValue C1 = DAG.getSelectCC(DL, Sub1_Hi, RHS_Hi, MinusOne, Zero,
1797                                  ISD::SETUGE);
1798     SDValue C2 = DAG.getSelectCC(DL, Sub1_Lo, RHS_Lo, MinusOne, Zero,
1799                                  ISD::SETUGE);
1800     SDValue C3 = DAG.getSelectCC(DL, Sub1_Hi, RHS_Hi, C2, C1, ISD::SETEQ);
1801 
1802     // TODO: Here and below portions of the code can be enclosed into if/endif.
1803     // Currently control flow is unconditional and we have 4 selects after
1804     // potential endif to substitute PHIs.
1805 
1806     // if C3 != 0 ...
1807     SDValue Sub2_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub1_Lo,
1808                                   RHS_Lo, Zero1);
1809     SDValue Sub2_Mi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub1_Mi,
1810                                   RHS_Hi, Sub1_Lo.getValue(1));
1811     SDValue Sub2_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Mi,
1812                                   Zero, Sub2_Lo.getValue(1));
1813     SDValue Sub2 = DAG.getBitcast(VT,
1814                         DAG.getBuildVector(MVT::v2i32, DL, {Sub2_Lo, Sub2_Hi}));
1815 
1816     SDValue Add3 = DAG.getNode(ISD::ADD, DL, VT, Mulhi3, One64);
1817 
1818     SDValue C4 = DAG.getSelectCC(DL, Sub2_Hi, RHS_Hi, MinusOne, Zero,
1819                                  ISD::SETUGE);
1820     SDValue C5 = DAG.getSelectCC(DL, Sub2_Lo, RHS_Lo, MinusOne, Zero,
1821                                  ISD::SETUGE);
1822     SDValue C6 = DAG.getSelectCC(DL, Sub2_Hi, RHS_Hi, C5, C4, ISD::SETEQ);
1823 
1824     // if (C6 != 0)
1825     SDValue Add4 = DAG.getNode(ISD::ADD, DL, VT, Add3, One64);
1826 
1827     SDValue Sub3_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Lo,
1828                                   RHS_Lo, Zero1);
1829     SDValue Sub3_Mi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Mi,
1830                                   RHS_Hi, Sub2_Lo.getValue(1));
1831     SDValue Sub3_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub3_Mi,
1832                                   Zero, Sub3_Lo.getValue(1));
1833     SDValue Sub3 = DAG.getBitcast(VT,
1834                         DAG.getBuildVector(MVT::v2i32, DL, {Sub3_Lo, Sub3_Hi}));
1835 
1836     // endif C6
1837     // endif C3
1838 
1839     SDValue Sel1 = DAG.getSelectCC(DL, C6, Zero, Add4, Add3, ISD::SETNE);
1840     SDValue Div  = DAG.getSelectCC(DL, C3, Zero, Sel1, Mulhi3, ISD::SETNE);
1841 
1842     SDValue Sel2 = DAG.getSelectCC(DL, C6, Zero, Sub3, Sub2, ISD::SETNE);
1843     SDValue Rem  = DAG.getSelectCC(DL, C3, Zero, Sel2, Sub1, ISD::SETNE);
1844 
1845     Results.push_back(Div);
1846     Results.push_back(Rem);
1847 
1848     return;
1849   }
1850 
1851   // r600 expandion.
1852   // Get Speculative values
1853   SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
1854   SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
1855 
1856   SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, Zero, REM_Part, LHS_Hi, ISD::SETEQ);
1857   SDValue REM = DAG.getBuildVector(MVT::v2i32, DL, {REM_Lo, Zero});
1858   REM = DAG.getNode(ISD::BITCAST, DL, MVT::i64, REM);
1859 
1860   SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, Zero, DIV_Part, Zero, ISD::SETEQ);
1861   SDValue DIV_Lo = Zero;
1862 
1863   const unsigned halfBitWidth = HalfVT.getSizeInBits();
1864 
1865   for (unsigned i = 0; i < halfBitWidth; ++i) {
1866     const unsigned bitPos = halfBitWidth - i - 1;
1867     SDValue POS = DAG.getConstant(bitPos, DL, HalfVT);
1868     // Get value of high bit
1869     SDValue HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
1870     HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, One);
1871     HBit = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, HBit);
1872 
1873     // Shift
1874     REM = DAG.getNode(ISD::SHL, DL, VT, REM, DAG.getConstant(1, DL, VT));
1875     // Add LHS high bit
1876     REM = DAG.getNode(ISD::OR, DL, VT, REM, HBit);
1877 
1878     SDValue BIT = DAG.getConstant(1ULL << bitPos, DL, HalfVT);
1879     SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, Zero, ISD::SETUGE);
1880 
1881     DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
1882 
1883     // Update REM
1884     SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
1885     REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETUGE);
1886   }
1887 
1888   SDValue DIV = DAG.getBuildVector(MVT::v2i32, DL, {DIV_Lo, DIV_Hi});
1889   DIV = DAG.getNode(ISD::BITCAST, DL, MVT::i64, DIV);
1890   Results.push_back(DIV);
1891   Results.push_back(REM);
1892 }
1893 
1894 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1895                                            SelectionDAG &DAG) const {
1896   SDLoc DL(Op);
1897   EVT VT = Op.getValueType();
1898 
1899   if (VT == MVT::i64) {
1900     SmallVector<SDValue, 2> Results;
1901     LowerUDIVREM64(Op, DAG, Results);
1902     return DAG.getMergeValues(Results, DL);
1903   }
1904 
1905   if (VT == MVT::i32) {
1906     if (SDValue Res = LowerDIVREM24(Op, DAG, false))
1907       return Res;
1908   }
1909 
1910   SDValue X = Op.getOperand(0);
1911   SDValue Y = Op.getOperand(1);
1912 
1913   // See AMDGPUCodeGenPrepare::expandDivRem32 for a description of the
1914   // algorithm used here.
1915 
1916   // Initial estimate of inv(y).
1917   SDValue Z = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Y);
1918 
1919   // One round of UNR.
1920   SDValue NegY = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Y);
1921   SDValue NegYZ = DAG.getNode(ISD::MUL, DL, VT, NegY, Z);
1922   Z = DAG.getNode(ISD::ADD, DL, VT, Z,
1923                   DAG.getNode(ISD::MULHU, DL, VT, Z, NegYZ));
1924 
1925   // Quotient/remainder estimate.
1926   SDValue Q = DAG.getNode(ISD::MULHU, DL, VT, X, Z);
1927   SDValue R =
1928       DAG.getNode(ISD::SUB, DL, VT, X, DAG.getNode(ISD::MUL, DL, VT, Q, Y));
1929 
1930   // First quotient/remainder refinement.
1931   EVT CCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
1932   SDValue One = DAG.getConstant(1, DL, VT);
1933   SDValue Cond = DAG.getSetCC(DL, CCVT, R, Y, ISD::SETUGE);
1934   Q = DAG.getNode(ISD::SELECT, DL, VT, Cond,
1935                   DAG.getNode(ISD::ADD, DL, VT, Q, One), Q);
1936   R = DAG.getNode(ISD::SELECT, DL, VT, Cond,
1937                   DAG.getNode(ISD::SUB, DL, VT, R, Y), R);
1938 
1939   // Second quotient/remainder refinement.
1940   Cond = DAG.getSetCC(DL, CCVT, R, Y, ISD::SETUGE);
1941   Q = DAG.getNode(ISD::SELECT, DL, VT, Cond,
1942                   DAG.getNode(ISD::ADD, DL, VT, Q, One), Q);
1943   R = DAG.getNode(ISD::SELECT, DL, VT, Cond,
1944                   DAG.getNode(ISD::SUB, DL, VT, R, Y), R);
1945 
1946   return DAG.getMergeValues({Q, R}, DL);
1947 }
1948 
1949 SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
1950                                            SelectionDAG &DAG) const {
1951   SDLoc DL(Op);
1952   EVT VT = Op.getValueType();
1953 
1954   SDValue LHS = Op.getOperand(0);
1955   SDValue RHS = Op.getOperand(1);
1956 
1957   SDValue Zero = DAG.getConstant(0, DL, VT);
1958   SDValue NegOne = DAG.getConstant(-1, DL, VT);
1959 
1960   if (VT == MVT::i32) {
1961     if (SDValue Res = LowerDIVREM24(Op, DAG, true))
1962       return Res;
1963   }
1964 
1965   if (VT == MVT::i64 &&
1966       DAG.ComputeNumSignBits(LHS) > 32 &&
1967       DAG.ComputeNumSignBits(RHS) > 32) {
1968     EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
1969 
1970     //HiLo split
1971     SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero);
1972     SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero);
1973     SDValue DIVREM = DAG.getNode(ISD::SDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
1974                                  LHS_Lo, RHS_Lo);
1975     SDValue Res[2] = {
1976       DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(0)),
1977       DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(1))
1978     };
1979     return DAG.getMergeValues(Res, DL);
1980   }
1981 
1982   SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
1983   SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
1984   SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
1985   SDValue RSign = LHSign; // Remainder sign is the same as LHS
1986 
1987   LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
1988   RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);
1989 
1990   LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
1991   RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);
1992 
1993   SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
1994   SDValue Rem = Div.getValue(1);
1995 
1996   Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
1997   Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);
1998 
1999   Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
2000   Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);
2001 
2002   SDValue Res[2] = {
2003     Div,
2004     Rem
2005   };
2006   return DAG.getMergeValues(Res, DL);
2007 }
2008 
2009 // (frem x, y) -> (fma (fneg (ftrunc (fdiv x, y))), y, x)
2010 SDValue AMDGPUTargetLowering::LowerFREM(SDValue Op, SelectionDAG &DAG) const {
2011   SDLoc SL(Op);
2012   EVT VT = Op.getValueType();
2013   auto Flags = Op->getFlags();
2014   SDValue X = Op.getOperand(0);
2015   SDValue Y = Op.getOperand(1);
2016 
2017   SDValue Div = DAG.getNode(ISD::FDIV, SL, VT, X, Y, Flags);
2018   SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, VT, Div, Flags);
2019   SDValue Neg = DAG.getNode(ISD::FNEG, SL, VT, Trunc, Flags);
2020   // TODO: For f32 use FMAD instead if !hasFastFMA32?
2021   return DAG.getNode(ISD::FMA, SL, VT, Neg, Y, X, Flags);
2022 }
2023 
2024 SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
2025   SDLoc SL(Op);
2026   SDValue Src = Op.getOperand(0);
2027 
2028   // result = trunc(src)
2029   // if (src > 0.0 && src != result)
2030   //   result += 1.0
2031 
2032   SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2033 
2034   const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64);
2035   const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f64);
2036 
2037   EVT SetCCVT =
2038       getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
2039 
2040   SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
2041   SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
2042   SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
2043 
2044   SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
2045   // TODO: Should this propagate fast-math-flags?
2046   return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
2047 }
2048 
2049 static SDValue extractF64Exponent(SDValue Hi, const SDLoc &SL,
2050                                   SelectionDAG &DAG) {
2051   const unsigned FractBits = 52;
2052   const unsigned ExpBits = 11;
2053 
2054   SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_U32, SL, MVT::i32,
2055                                 Hi,
2056                                 DAG.getConstant(FractBits - 32, SL, MVT::i32),
2057                                 DAG.getConstant(ExpBits, SL, MVT::i32));
2058   SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
2059                             DAG.getConstant(1023, SL, MVT::i32));
2060 
2061   return Exp;
2062 }
2063 
2064 SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
2065   SDLoc SL(Op);
2066   SDValue Src = Op.getOperand(0);
2067 
2068   assert(Op.getValueType() == MVT::f64);
2069 
2070   const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
2071 
2072   // Extract the upper half, since this is where we will find the sign and
2073   // exponent.
2074   SDValue Hi = getHiHalf64(Src, DAG);
2075 
2076   SDValue Exp = extractF64Exponent(Hi, SL, DAG);
2077 
2078   const unsigned FractBits = 52;
2079 
2080   // Extract the sign bit.
2081   const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, SL, MVT::i32);
2082   SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);
2083 
2084   // Extend back to 64-bits.
2085   SDValue SignBit64 = DAG.getBuildVector(MVT::v2i32, SL, {Zero, SignBit});
2086   SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);
2087 
2088   SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
2089   const SDValue FractMask
2090     = DAG.getConstant((UINT64_C(1) << FractBits) - 1, SL, MVT::i64);
2091 
2092   SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
2093   SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
2094   SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);
2095 
2096   EVT SetCCVT =
2097       getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i32);
2098 
2099   const SDValue FiftyOne = DAG.getConstant(FractBits - 1, SL, MVT::i32);
2100 
2101   SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
2102   SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
2103 
2104   SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
2105   SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);
2106 
2107   return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
2108 }
2109 
2110 SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
2111   SDLoc SL(Op);
2112   SDValue Src = Op.getOperand(0);
2113 
2114   assert(Op.getValueType() == MVT::f64);
2115 
2116   APFloat C1Val(APFloat::IEEEdouble(), "0x1.0p+52");
2117   SDValue C1 = DAG.getConstantFP(C1Val, SL, MVT::f64);
2118   SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
2119 
2120   // TODO: Should this propagate fast-math-flags?
2121 
2122   SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
2123   SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
2124 
2125   SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
2126 
2127   APFloat C2Val(APFloat::IEEEdouble(), "0x1.fffffffffffffp+51");
2128   SDValue C2 = DAG.getConstantFP(C2Val, SL, MVT::f64);
2129 
2130   EVT SetCCVT =
2131       getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
2132   SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
2133 
2134   return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
2135 }
2136 
2137 SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const {
2138   // FNEARBYINT and FRINT are the same, except in their handling of FP
2139   // exceptions. Those aren't really meaningful for us, and OpenCL only has
2140   // rint, so just treat them as equivalent.
2141   return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
2142 }
2143 
2144 // XXX - May require not supporting f32 denormals?
2145 
2146 // Don't handle v2f16. The extra instructions to scalarize and repack around the
2147 // compare and vselect end up producing worse code than scalarizing the whole
2148 // operation.
2149 SDValue AMDGPUTargetLowering::LowerFROUND(SDValue Op, SelectionDAG &DAG) const {
2150   SDLoc SL(Op);
2151   SDValue X = Op.getOperand(0);
2152   EVT VT = Op.getValueType();
2153 
2154   SDValue T = DAG.getNode(ISD::FTRUNC, SL, VT, X);
2155 
2156   // TODO: Should this propagate fast-math-flags?
2157 
2158   SDValue Diff = DAG.getNode(ISD::FSUB, SL, VT, X, T);
2159 
2160   SDValue AbsDiff = DAG.getNode(ISD::FABS, SL, VT, Diff);
2161 
2162   const SDValue Zero = DAG.getConstantFP(0.0, SL, VT);
2163   const SDValue One = DAG.getConstantFP(1.0, SL, VT);
2164   const SDValue Half = DAG.getConstantFP(0.5, SL, VT);
2165 
2166   SDValue SignOne = DAG.getNode(ISD::FCOPYSIGN, SL, VT, One, X);
2167 
2168   EVT SetCCVT =
2169       getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
2170 
2171   SDValue Cmp = DAG.getSetCC(SL, SetCCVT, AbsDiff, Half, ISD::SETOGE);
2172 
2173   SDValue Sel = DAG.getNode(ISD::SELECT, SL, VT, Cmp, SignOne, Zero);
2174 
2175   return DAG.getNode(ISD::FADD, SL, VT, T, Sel);
2176 }
2177 
2178 SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
2179   SDLoc SL(Op);
2180   SDValue Src = Op.getOperand(0);
2181 
2182   // result = trunc(src);
2183   // if (src < 0.0 && src != result)
2184   //   result += -1.0.
2185 
2186   SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2187 
2188   const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64);
2189   const SDValue NegOne = DAG.getConstantFP(-1.0, SL, MVT::f64);
2190 
2191   EVT SetCCVT =
2192       getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
2193 
2194   SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
2195   SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
2196   SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
2197 
2198   SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
2199   // TODO: Should this propagate fast-math-flags?
2200   return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
2201 }
2202 
2203 SDValue AMDGPUTargetLowering::LowerFLOG(SDValue Op, SelectionDAG &DAG,
2204                                         double Log2BaseInverted) const {
2205   EVT VT = Op.getValueType();
2206 
2207   SDLoc SL(Op);
2208   SDValue Operand = Op.getOperand(0);
2209   SDValue Log2Operand = DAG.getNode(ISD::FLOG2, SL, VT, Operand);
2210   SDValue Log2BaseInvertedOperand = DAG.getConstantFP(Log2BaseInverted, SL, VT);
2211 
2212   return DAG.getNode(ISD::FMUL, SL, VT, Log2Operand, Log2BaseInvertedOperand);
2213 }
2214 
2215 // exp2(M_LOG2E_F * f);
2216 SDValue AMDGPUTargetLowering::lowerFEXP(SDValue Op, SelectionDAG &DAG) const {
2217   EVT VT = Op.getValueType();
2218   SDLoc SL(Op);
2219   SDValue Src = Op.getOperand(0);
2220 
2221   const SDValue K = DAG.getConstantFP(numbers::log2e, SL, VT);
2222   SDValue Mul = DAG.getNode(ISD::FMUL, SL, VT, Src, K, Op->getFlags());
2223   return DAG.getNode(ISD::FEXP2, SL, VT, Mul, Op->getFlags());
2224 }
2225 
2226 static bool isCtlzOpc(unsigned Opc) {
2227   return Opc == ISD::CTLZ || Opc == ISD::CTLZ_ZERO_UNDEF;
2228 }
2229 
2230 static bool isCttzOpc(unsigned Opc) {
2231   return Opc == ISD::CTTZ || Opc == ISD::CTTZ_ZERO_UNDEF;
2232 }
2233 
2234 SDValue AMDGPUTargetLowering::LowerCTLZ_CTTZ(SDValue Op, SelectionDAG &DAG) const {
2235   SDLoc SL(Op);
2236   SDValue Src = Op.getOperand(0);
2237 
2238   assert(isCtlzOpc(Op.getOpcode()) || isCttzOpc(Op.getOpcode()));
2239   bool Ctlz = isCtlzOpc(Op.getOpcode());
2240   unsigned NewOpc = Ctlz ? AMDGPUISD::FFBH_U32 : AMDGPUISD::FFBL_B32;
2241 
2242   bool ZeroUndef = Op.getOpcode() == ISD::CTLZ_ZERO_UNDEF ||
2243                    Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF;
2244 
2245   if (Src.getValueType() == MVT::i32) {
2246     // (ctlz hi:lo) -> (umin (ffbh src), 32)
2247     // (cttz hi:lo) -> (umin (ffbl src), 32)
2248     // (ctlz_zero_undef src) -> (ffbh src)
2249     // (cttz_zero_undef src) -> (ffbl src)
2250     SDValue NewOpr = DAG.getNode(NewOpc, SL, MVT::i32, Src);
2251     if (!ZeroUndef) {
2252       const SDValue Const32 = DAG.getConstant(32, SL, MVT::i32);
2253       NewOpr = DAG.getNode(ISD::UMIN, SL, MVT::i32, NewOpr, Const32);
2254     }
2255     return NewOpr;
2256   }
2257 
2258   SDValue Lo, Hi;
2259   std::tie(Lo, Hi) = split64BitValue(Src, DAG);
2260 
2261   SDValue OprLo = DAG.getNode(NewOpc, SL, MVT::i32, Lo);
2262   SDValue OprHi = DAG.getNode(NewOpc, SL, MVT::i32, Hi);
2263 
2264   // (ctlz hi:lo) -> (umin3 (ffbh hi), (uaddsat (ffbh lo), 32), 64)
2265   // (cttz hi:lo) -> (umin3 (uaddsat (ffbl hi), 32), (ffbl lo), 64)
2266   // (ctlz_zero_undef hi:lo) -> (umin (ffbh hi), (add (ffbh lo), 32))
2267   // (cttz_zero_undef hi:lo) -> (umin (add (ffbl hi), 32), (ffbl lo))
2268 
2269   unsigned AddOpc = ZeroUndef ? ISD::ADD : ISD::UADDSAT;
2270   const SDValue Const32 = DAG.getConstant(32, SL, MVT::i32);
2271   if (Ctlz)
2272     OprLo = DAG.getNode(AddOpc, SL, MVT::i32, OprLo, Const32);
2273   else
2274     OprHi = DAG.getNode(AddOpc, SL, MVT::i32, OprHi, Const32);
2275 
2276   SDValue NewOpr;
2277   NewOpr = DAG.getNode(ISD::UMIN, SL, MVT::i32, OprLo, OprHi);
2278   if (!ZeroUndef) {
2279     const SDValue Const64 = DAG.getConstant(64, SL, MVT::i32);
2280     NewOpr = DAG.getNode(ISD::UMIN, SL, MVT::i32, NewOpr, Const64);
2281   }
2282 
2283   return DAG.getNode(ISD::ZERO_EXTEND, SL, MVT::i64, NewOpr);
2284 }
2285 
2286 SDValue AMDGPUTargetLowering::LowerINT_TO_FP32(SDValue Op, SelectionDAG &DAG,
2287                                                bool Signed) const {
2288   // The regular method converting a 64-bit integer to float roughly consists of
2289   // 2 steps: normalization and rounding. In fact, after normalization, the
2290   // conversion from a 64-bit integer to a float is essentially the same as the
2291   // one from a 32-bit integer. The only difference is that it has more
2292   // trailing bits to be rounded. To leverage the native 32-bit conversion, a
2293   // 64-bit integer could be preprocessed and fit into a 32-bit integer then
2294   // converted into the correct float number. The basic steps for the unsigned
2295   // conversion are illustrated in the following pseudo code:
2296   //
2297   // f32 uitofp(i64 u) {
2298   //   i32 hi, lo = split(u);
2299   //   // Only count the leading zeros in hi as we have native support of the
2300   //   // conversion from i32 to f32. If hi is all 0s, the conversion is
2301   //   // reduced to a 32-bit one automatically.
2302   //   i32 shamt = clz(hi); // Return 32 if hi is all 0s.
2303   //   u <<= shamt;
2304   //   hi, lo = split(u);
2305   //   hi |= (lo != 0) ? 1 : 0; // Adjust rounding bit in hi based on lo.
2306   //   // convert it as a 32-bit integer and scale the result back.
2307   //   return uitofp(hi) * 2^(32 - shamt);
2308   // }
2309   //
2310   // The signed one follows the same principle but uses 'ffbh_i32' to count its
2311   // sign bits instead. If 'ffbh_i32' is not available, its absolute value is
2312   // converted instead followed by negation based its sign bit.
2313 
2314   SDLoc SL(Op);
2315   SDValue Src = Op.getOperand(0);
2316 
2317   SDValue Lo, Hi;
2318   std::tie(Lo, Hi) = split64BitValue(Src, DAG);
2319   SDValue Sign;
2320   SDValue ShAmt;
2321   if (Signed && Subtarget->isGCN()) {
2322     // We also need to consider the sign bit in Lo if Hi has just sign bits,
2323     // i.e. Hi is 0 or -1. However, that only needs to take the MSB into
2324     // account. That is, the maximal shift is
2325     // - 32 if Lo and Hi have opposite signs;
2326     // - 33 if Lo and Hi have the same sign.
2327     //
2328     // Or, MaxShAmt = 33 + OppositeSign, where
2329     //
2330     // OppositeSign is defined as ((Lo ^ Hi) >> 31), which is
2331     // - -1 if Lo and Hi have opposite signs; and
2332     // -  0 otherwise.
2333     //
2334     // All in all, ShAmt is calculated as
2335     //
2336     //  umin(sffbh(Hi), 33 + (Lo^Hi)>>31) - 1.
2337     //
2338     // or
2339     //
2340     //  umin(sffbh(Hi) - 1, 32 + (Lo^Hi)>>31).
2341     //
2342     // to reduce the critical path.
2343     SDValue OppositeSign = DAG.getNode(
2344         ISD::SRA, SL, MVT::i32, DAG.getNode(ISD::XOR, SL, MVT::i32, Lo, Hi),
2345         DAG.getConstant(31, SL, MVT::i32));
2346     SDValue MaxShAmt =
2347         DAG.getNode(ISD::ADD, SL, MVT::i32, DAG.getConstant(32, SL, MVT::i32),
2348                     OppositeSign);
2349     // Count the leading sign bits.
2350     ShAmt = DAG.getNode(AMDGPUISD::FFBH_I32, SL, MVT::i32, Hi);
2351     // Different from unsigned conversion, the shift should be one bit less to
2352     // preserve the sign bit.
2353     ShAmt = DAG.getNode(ISD::SUB, SL, MVT::i32, ShAmt,
2354                         DAG.getConstant(1, SL, MVT::i32));
2355     ShAmt = DAG.getNode(ISD::UMIN, SL, MVT::i32, ShAmt, MaxShAmt);
2356   } else {
2357     if (Signed) {
2358       // Without 'ffbh_i32', only leading zeros could be counted. Take the
2359       // absolute value first.
2360       Sign = DAG.getNode(ISD::SRA, SL, MVT::i64, Src,
2361                          DAG.getConstant(63, SL, MVT::i64));
2362       SDValue Abs =
2363           DAG.getNode(ISD::XOR, SL, MVT::i64,
2364                       DAG.getNode(ISD::ADD, SL, MVT::i64, Src, Sign), Sign);
2365       std::tie(Lo, Hi) = split64BitValue(Abs, DAG);
2366     }
2367     // Count the leading zeros.
2368     ShAmt = DAG.getNode(ISD::CTLZ, SL, MVT::i32, Hi);
2369     // The shift amount for signed integers is [0, 32].
2370   }
2371   // Normalize the given 64-bit integer.
2372   SDValue Norm = DAG.getNode(ISD::SHL, SL, MVT::i64, Src, ShAmt);
2373   // Split it again.
2374   std::tie(Lo, Hi) = split64BitValue(Norm, DAG);
2375   // Calculate the adjust bit for rounding.
2376   // (lo != 0) ? 1 : 0 => (lo >= 1) ? 1 : 0 => umin(1, lo)
2377   SDValue Adjust = DAG.getNode(ISD::UMIN, SL, MVT::i32,
2378                                DAG.getConstant(1, SL, MVT::i32), Lo);
2379   // Get the 32-bit normalized integer.
2380   Norm = DAG.getNode(ISD::OR, SL, MVT::i32, Hi, Adjust);
2381   // Convert the normalized 32-bit integer into f32.
2382   unsigned Opc =
2383       (Signed && Subtarget->isGCN()) ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
2384   SDValue FVal = DAG.getNode(Opc, SL, MVT::f32, Norm);
2385 
2386   // Finally, need to scale back the converted floating number as the original
2387   // 64-bit integer is converted as a 32-bit one.
2388   ShAmt = DAG.getNode(ISD::SUB, SL, MVT::i32, DAG.getConstant(32, SL, MVT::i32),
2389                       ShAmt);
2390   // On GCN, use LDEXP directly.
2391   if (Subtarget->isGCN())
2392     return DAG.getNode(AMDGPUISD::LDEXP, SL, MVT::f32, FVal, ShAmt);
2393 
2394   // Otherwise, align 'ShAmt' to the exponent part and add it into the exponent
2395   // part directly to emulate the multiplication of 2^ShAmt. That 8-bit
2396   // exponent is enough to avoid overflowing into the sign bit.
2397   SDValue Exp = DAG.getNode(ISD::SHL, SL, MVT::i32, ShAmt,
2398                             DAG.getConstant(23, SL, MVT::i32));
2399   SDValue IVal =
2400       DAG.getNode(ISD::ADD, SL, MVT::i32,
2401                   DAG.getNode(ISD::BITCAST, SL, MVT::i32, FVal), Exp);
2402   if (Signed) {
2403     // Set the sign bit.
2404     Sign = DAG.getNode(ISD::SHL, SL, MVT::i32,
2405                        DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, Sign),
2406                        DAG.getConstant(31, SL, MVT::i32));
2407     IVal = DAG.getNode(ISD::OR, SL, MVT::i32, IVal, Sign);
2408   }
2409   return DAG.getNode(ISD::BITCAST, SL, MVT::f32, IVal);
2410 }
2411 
2412 SDValue AMDGPUTargetLowering::LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG,
2413                                                bool Signed) const {
2414   SDLoc SL(Op);
2415   SDValue Src = Op.getOperand(0);
2416 
2417   SDValue Lo, Hi;
2418   std::tie(Lo, Hi) = split64BitValue(Src, DAG);
2419 
2420   SDValue CvtHi = DAG.getNode(Signed ? ISD::SINT_TO_FP : ISD::UINT_TO_FP,
2421                               SL, MVT::f64, Hi);
2422 
2423   SDValue CvtLo = DAG.getNode(ISD::UINT_TO_FP, SL, MVT::f64, Lo);
2424 
2425   SDValue LdExp = DAG.getNode(AMDGPUISD::LDEXP, SL, MVT::f64, CvtHi,
2426                               DAG.getConstant(32, SL, MVT::i32));
2427   // TODO: Should this propagate fast-math-flags?
2428   return DAG.getNode(ISD::FADD, SL, MVT::f64, LdExp, CvtLo);
2429 }
2430 
2431 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
2432                                                SelectionDAG &DAG) const {
2433   // TODO: Factor out code common with LowerSINT_TO_FP.
2434   EVT DestVT = Op.getValueType();
2435   SDValue Src = Op.getOperand(0);
2436   EVT SrcVT = Src.getValueType();
2437 
2438   if (SrcVT == MVT::i16) {
2439     if (DestVT == MVT::f16)
2440       return Op;
2441     SDLoc DL(Op);
2442 
2443     // Promote src to i32
2444     SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, Src);
2445     return DAG.getNode(ISD::UINT_TO_FP, DL, DestVT, Ext);
2446   }
2447 
2448   assert(SrcVT == MVT::i64 && "operation should be legal");
2449 
2450   if (Subtarget->has16BitInsts() && DestVT == MVT::f16) {
2451     SDLoc DL(Op);
2452 
2453     SDValue IntToFp32 = DAG.getNode(Op.getOpcode(), DL, MVT::f32, Src);
2454     SDValue FPRoundFlag = DAG.getIntPtrConstant(0, SDLoc(Op));
2455     SDValue FPRound =
2456         DAG.getNode(ISD::FP_ROUND, DL, MVT::f16, IntToFp32, FPRoundFlag);
2457 
2458     return FPRound;
2459   }
2460 
2461   if (DestVT == MVT::f32)
2462     return LowerINT_TO_FP32(Op, DAG, false);
2463 
2464   assert(DestVT == MVT::f64);
2465   return LowerINT_TO_FP64(Op, DAG, false);
2466 }
2467 
2468 SDValue AMDGPUTargetLowering::LowerSINT_TO_FP(SDValue Op,
2469                                               SelectionDAG &DAG) const {
2470   EVT DestVT = Op.getValueType();
2471 
2472   SDValue Src = Op.getOperand(0);
2473   EVT SrcVT = Src.getValueType();
2474 
2475   if (SrcVT == MVT::i16) {
2476     if (DestVT == MVT::f16)
2477       return Op;
2478 
2479     SDLoc DL(Op);
2480     // Promote src to i32
2481     SDValue Ext = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32, Src);
2482     return DAG.getNode(ISD::SINT_TO_FP, DL, DestVT, Ext);
2483   }
2484 
2485   assert(SrcVT == MVT::i64 && "operation should be legal");
2486 
2487   // TODO: Factor out code common with LowerUINT_TO_FP.
2488 
2489   if (Subtarget->has16BitInsts() && DestVT == MVT::f16) {
2490     SDLoc DL(Op);
2491     SDValue Src = Op.getOperand(0);
2492 
2493     SDValue IntToFp32 = DAG.getNode(Op.getOpcode(), DL, MVT::f32, Src);
2494     SDValue FPRoundFlag = DAG.getIntPtrConstant(0, SDLoc(Op));
2495     SDValue FPRound =
2496         DAG.getNode(ISD::FP_ROUND, DL, MVT::f16, IntToFp32, FPRoundFlag);
2497 
2498     return FPRound;
2499   }
2500 
2501   if (DestVT == MVT::f32)
2502     return LowerINT_TO_FP32(Op, DAG, true);
2503 
2504   assert(DestVT == MVT::f64);
2505   return LowerINT_TO_FP64(Op, DAG, true);
2506 }
2507 
2508 SDValue AMDGPUTargetLowering::LowerFP_TO_INT64(SDValue Op, SelectionDAG &DAG,
2509                                                bool Signed) const {
2510   SDLoc SL(Op);
2511 
2512   SDValue Src = Op.getOperand(0);
2513   EVT SrcVT = Src.getValueType();
2514 
2515   assert(SrcVT == MVT::f32 || SrcVT == MVT::f64);
2516 
2517   // The basic idea of converting a floating point number into a pair of 32-bit
2518   // integers is illustrated as follows:
2519   //
2520   //     tf := trunc(val);
2521   //    hif := floor(tf * 2^-32);
2522   //    lof := tf - hif * 2^32; // lof is always positive due to floor.
2523   //     hi := fptoi(hif);
2524   //     lo := fptoi(lof);
2525   //
2526   SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, SrcVT, Src);
2527   SDValue Sign;
2528   if (Signed && SrcVT == MVT::f32) {
2529     // However, a 32-bit floating point number has only 23 bits mantissa and
2530     // it's not enough to hold all the significant bits of `lof` if val is
2531     // negative. To avoid the loss of precision, We need to take the absolute
2532     // value after truncating and flip the result back based on the original
2533     // signedness.
2534     Sign = DAG.getNode(ISD::SRA, SL, MVT::i32,
2535                        DAG.getNode(ISD::BITCAST, SL, MVT::i32, Trunc),
2536                        DAG.getConstant(31, SL, MVT::i32));
2537     Trunc = DAG.getNode(ISD::FABS, SL, SrcVT, Trunc);
2538   }
2539 
2540   SDValue K0, K1;
2541   if (SrcVT == MVT::f64) {
2542     K0 = DAG.getConstantFP(BitsToDouble(UINT64_C(/*2^-32*/ 0x3df0000000000000)),
2543                            SL, SrcVT);
2544     K1 = DAG.getConstantFP(BitsToDouble(UINT64_C(/*-2^32*/ 0xc1f0000000000000)),
2545                            SL, SrcVT);
2546   } else {
2547     K0 = DAG.getConstantFP(BitsToFloat(UINT32_C(/*2^-32*/ 0x2f800000)), SL,
2548                            SrcVT);
2549     K1 = DAG.getConstantFP(BitsToFloat(UINT32_C(/*-2^32*/ 0xcf800000)), SL,
2550                            SrcVT);
2551   }
2552   // TODO: Should this propagate fast-math-flags?
2553   SDValue Mul = DAG.getNode(ISD::FMUL, SL, SrcVT, Trunc, K0);
2554 
2555   SDValue FloorMul = DAG.getNode(ISD::FFLOOR, SL, SrcVT, Mul);
2556 
2557   SDValue Fma = DAG.getNode(ISD::FMA, SL, SrcVT, FloorMul, K1, Trunc);
2558 
2559   SDValue Hi = DAG.getNode((Signed && SrcVT == MVT::f64) ? ISD::FP_TO_SINT
2560                                                          : ISD::FP_TO_UINT,
2561                            SL, MVT::i32, FloorMul);
2562   SDValue Lo = DAG.getNode(ISD::FP_TO_UINT, SL, MVT::i32, Fma);
2563 
2564   SDValue Result = DAG.getNode(ISD::BITCAST, SL, MVT::i64,
2565                                DAG.getBuildVector(MVT::v2i32, SL, {Lo, Hi}));
2566 
2567   if (Signed && SrcVT == MVT::f32) {
2568     assert(Sign);
2569     // Flip the result based on the signedness, which is either all 0s or 1s.
2570     Sign = DAG.getNode(ISD::BITCAST, SL, MVT::i64,
2571                        DAG.getBuildVector(MVT::v2i32, SL, {Sign, Sign}));
2572     // r := xor(r, sign) - sign;
2573     Result =
2574         DAG.getNode(ISD::SUB, SL, MVT::i64,
2575                     DAG.getNode(ISD::XOR, SL, MVT::i64, Result, Sign), Sign);
2576   }
2577 
2578   return Result;
2579 }
2580 
2581 SDValue AMDGPUTargetLowering::LowerFP_TO_FP16(SDValue Op, SelectionDAG &DAG) const {
2582   SDLoc DL(Op);
2583   SDValue N0 = Op.getOperand(0);
2584 
2585   // Convert to target node to get known bits
2586   if (N0.getValueType() == MVT::f32)
2587     return DAG.getNode(AMDGPUISD::FP_TO_FP16, DL, Op.getValueType(), N0);
2588 
2589   if (getTargetMachine().Options.UnsafeFPMath) {
2590     // There is a generic expand for FP_TO_FP16 with unsafe fast math.
2591     return SDValue();
2592   }
2593 
2594   assert(N0.getSimpleValueType() == MVT::f64);
2595 
2596   // f64 -> f16 conversion using round-to-nearest-even rounding mode.
2597   const unsigned ExpMask = 0x7ff;
2598   const unsigned ExpBiasf64 = 1023;
2599   const unsigned ExpBiasf16 = 15;
2600   SDValue Zero = DAG.getConstant(0, DL, MVT::i32);
2601   SDValue One = DAG.getConstant(1, DL, MVT::i32);
2602   SDValue U = DAG.getNode(ISD::BITCAST, DL, MVT::i64, N0);
2603   SDValue UH = DAG.getNode(ISD::SRL, DL, MVT::i64, U,
2604                            DAG.getConstant(32, DL, MVT::i64));
2605   UH = DAG.getZExtOrTrunc(UH, DL, MVT::i32);
2606   U = DAG.getZExtOrTrunc(U, DL, MVT::i32);
2607   SDValue E = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
2608                           DAG.getConstant(20, DL, MVT::i64));
2609   E = DAG.getNode(ISD::AND, DL, MVT::i32, E,
2610                   DAG.getConstant(ExpMask, DL, MVT::i32));
2611   // Subtract the fp64 exponent bias (1023) to get the real exponent and
2612   // add the f16 bias (15) to get the biased exponent for the f16 format.
2613   E = DAG.getNode(ISD::ADD, DL, MVT::i32, E,
2614                   DAG.getConstant(-ExpBiasf64 + ExpBiasf16, DL, MVT::i32));
2615 
2616   SDValue M = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
2617                           DAG.getConstant(8, DL, MVT::i32));
2618   M = DAG.getNode(ISD::AND, DL, MVT::i32, M,
2619                   DAG.getConstant(0xffe, DL, MVT::i32));
2620 
2621   SDValue MaskedSig = DAG.getNode(ISD::AND, DL, MVT::i32, UH,
2622                                   DAG.getConstant(0x1ff, DL, MVT::i32));
2623   MaskedSig = DAG.getNode(ISD::OR, DL, MVT::i32, MaskedSig, U);
2624 
2625   SDValue Lo40Set = DAG.getSelectCC(DL, MaskedSig, Zero, Zero, One, ISD::SETEQ);
2626   M = DAG.getNode(ISD::OR, DL, MVT::i32, M, Lo40Set);
2627 
2628   // (M != 0 ? 0x0200 : 0) | 0x7c00;
2629   SDValue I = DAG.getNode(ISD::OR, DL, MVT::i32,
2630       DAG.getSelectCC(DL, M, Zero, DAG.getConstant(0x0200, DL, MVT::i32),
2631                       Zero, ISD::SETNE), DAG.getConstant(0x7c00, DL, MVT::i32));
2632 
2633   // N = M | (E << 12);
2634   SDValue N = DAG.getNode(ISD::OR, DL, MVT::i32, M,
2635       DAG.getNode(ISD::SHL, DL, MVT::i32, E,
2636                   DAG.getConstant(12, DL, MVT::i32)));
2637 
2638   // B = clamp(1-E, 0, 13);
2639   SDValue OneSubExp = DAG.getNode(ISD::SUB, DL, MVT::i32,
2640                                   One, E);
2641   SDValue B = DAG.getNode(ISD::SMAX, DL, MVT::i32, OneSubExp, Zero);
2642   B = DAG.getNode(ISD::SMIN, DL, MVT::i32, B,
2643                   DAG.getConstant(13, DL, MVT::i32));
2644 
2645   SDValue SigSetHigh = DAG.getNode(ISD::OR, DL, MVT::i32, M,
2646                                    DAG.getConstant(0x1000, DL, MVT::i32));
2647 
2648   SDValue D = DAG.getNode(ISD::SRL, DL, MVT::i32, SigSetHigh, B);
2649   SDValue D0 = DAG.getNode(ISD::SHL, DL, MVT::i32, D, B);
2650   SDValue D1 = DAG.getSelectCC(DL, D0, SigSetHigh, One, Zero, ISD::SETNE);
2651   D = DAG.getNode(ISD::OR, DL, MVT::i32, D, D1);
2652 
2653   SDValue V = DAG.getSelectCC(DL, E, One, D, N, ISD::SETLT);
2654   SDValue VLow3 = DAG.getNode(ISD::AND, DL, MVT::i32, V,
2655                               DAG.getConstant(0x7, DL, MVT::i32));
2656   V = DAG.getNode(ISD::SRL, DL, MVT::i32, V,
2657                   DAG.getConstant(2, DL, MVT::i32));
2658   SDValue V0 = DAG.getSelectCC(DL, VLow3, DAG.getConstant(3, DL, MVT::i32),
2659                                One, Zero, ISD::SETEQ);
2660   SDValue V1 = DAG.getSelectCC(DL, VLow3, DAG.getConstant(5, DL, MVT::i32),
2661                                One, Zero, ISD::SETGT);
2662   V1 = DAG.getNode(ISD::OR, DL, MVT::i32, V0, V1);
2663   V = DAG.getNode(ISD::ADD, DL, MVT::i32, V, V1);
2664 
2665   V = DAG.getSelectCC(DL, E, DAG.getConstant(30, DL, MVT::i32),
2666                       DAG.getConstant(0x7c00, DL, MVT::i32), V, ISD::SETGT);
2667   V = DAG.getSelectCC(DL, E, DAG.getConstant(1039, DL, MVT::i32),
2668                       I, V, ISD::SETEQ);
2669 
2670   // Extract the sign bit.
2671   SDValue Sign = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
2672                             DAG.getConstant(16, DL, MVT::i32));
2673   Sign = DAG.getNode(ISD::AND, DL, MVT::i32, Sign,
2674                      DAG.getConstant(0x8000, DL, MVT::i32));
2675 
2676   V = DAG.getNode(ISD::OR, DL, MVT::i32, Sign, V);
2677   return DAG.getZExtOrTrunc(V, DL, Op.getValueType());
2678 }
2679 
2680 SDValue AMDGPUTargetLowering::LowerFP_TO_INT(SDValue Op,
2681                                              SelectionDAG &DAG) const {
2682   SDValue Src = Op.getOperand(0);
2683   unsigned OpOpcode = Op.getOpcode();
2684   EVT SrcVT = Src.getValueType();
2685   EVT DestVT = Op.getValueType();
2686 
2687   // Will be selected natively
2688   if (SrcVT == MVT::f16 && DestVT == MVT::i16)
2689     return Op;
2690 
2691   // Promote i16 to i32
2692   if (DestVT == MVT::i16 && (SrcVT == MVT::f32 || SrcVT == MVT::f64)) {
2693     SDLoc DL(Op);
2694 
2695     SDValue FpToInt32 = DAG.getNode(OpOpcode, DL, MVT::i32, Src);
2696     return DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, FpToInt32);
2697   }
2698 
2699   if (SrcVT == MVT::f16 ||
2700       (SrcVT == MVT::f32 && Src.getOpcode() == ISD::FP16_TO_FP)) {
2701     SDLoc DL(Op);
2702 
2703     SDValue FpToInt32 = DAG.getNode(OpOpcode, DL, MVT::i32, Src);
2704     unsigned Ext =
2705         OpOpcode == ISD::FP_TO_SINT ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
2706     return DAG.getNode(Ext, DL, MVT::i64, FpToInt32);
2707   }
2708 
2709   if (DestVT == MVT::i64 && (SrcVT == MVT::f32 || SrcVT == MVT::f64))
2710     return LowerFP_TO_INT64(Op, DAG, OpOpcode == ISD::FP_TO_SINT);
2711 
2712   return SDValue();
2713 }
2714 
2715 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
2716                                                      SelectionDAG &DAG) const {
2717   EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
2718   MVT VT = Op.getSimpleValueType();
2719   MVT ScalarVT = VT.getScalarType();
2720 
2721   assert(VT.isVector());
2722 
2723   SDValue Src = Op.getOperand(0);
2724   SDLoc DL(Op);
2725 
2726   // TODO: Don't scalarize on Evergreen?
2727   unsigned NElts = VT.getVectorNumElements();
2728   SmallVector<SDValue, 8> Args;
2729   DAG.ExtractVectorElements(Src, Args, 0, NElts);
2730 
2731   SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
2732   for (unsigned I = 0; I < NElts; ++I)
2733     Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
2734 
2735   return DAG.getBuildVector(VT, DL, Args);
2736 }
2737 
2738 //===----------------------------------------------------------------------===//
2739 // Custom DAG optimizations
2740 //===----------------------------------------------------------------------===//
2741 
2742 static bool isU24(SDValue Op, SelectionDAG &DAG) {
2743   return AMDGPUTargetLowering::numBitsUnsigned(Op, DAG) <= 24;
2744 }
2745 
2746 static bool isI24(SDValue Op, SelectionDAG &DAG) {
2747   EVT VT = Op.getValueType();
2748   return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
2749                                      // as unsigned 24-bit values.
2750          AMDGPUTargetLowering::numBitsSigned(Op, DAG) <= 24;
2751 }
2752 
2753 static SDValue simplifyMul24(SDNode *Node24,
2754                              TargetLowering::DAGCombinerInfo &DCI) {
2755   SelectionDAG &DAG = DCI.DAG;
2756   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2757   bool IsIntrin = Node24->getOpcode() == ISD::INTRINSIC_WO_CHAIN;
2758 
2759   SDValue LHS = IsIntrin ? Node24->getOperand(1) : Node24->getOperand(0);
2760   SDValue RHS = IsIntrin ? Node24->getOperand(2) : Node24->getOperand(1);
2761   unsigned NewOpcode = Node24->getOpcode();
2762   if (IsIntrin) {
2763     unsigned IID = cast<ConstantSDNode>(Node24->getOperand(0))->getZExtValue();
2764     switch (IID) {
2765     case Intrinsic::amdgcn_mul_i24:
2766       NewOpcode = AMDGPUISD::MUL_I24;
2767       break;
2768     case Intrinsic::amdgcn_mul_u24:
2769       NewOpcode = AMDGPUISD::MUL_U24;
2770       break;
2771     case Intrinsic::amdgcn_mulhi_i24:
2772       NewOpcode = AMDGPUISD::MULHI_I24;
2773       break;
2774     case Intrinsic::amdgcn_mulhi_u24:
2775       NewOpcode = AMDGPUISD::MULHI_U24;
2776       break;
2777     default:
2778       llvm_unreachable("Expected 24-bit mul intrinsic");
2779     }
2780   }
2781 
2782   APInt Demanded = APInt::getLowBitsSet(LHS.getValueSizeInBits(), 24);
2783 
2784   // First try to simplify using SimplifyMultipleUseDemandedBits which allows
2785   // the operands to have other uses, but will only perform simplifications that
2786   // involve bypassing some nodes for this user.
2787   SDValue DemandedLHS = TLI.SimplifyMultipleUseDemandedBits(LHS, Demanded, DAG);
2788   SDValue DemandedRHS = TLI.SimplifyMultipleUseDemandedBits(RHS, Demanded, DAG);
2789   if (DemandedLHS || DemandedRHS)
2790     return DAG.getNode(NewOpcode, SDLoc(Node24), Node24->getVTList(),
2791                        DemandedLHS ? DemandedLHS : LHS,
2792                        DemandedRHS ? DemandedRHS : RHS);
2793 
2794   // Now try SimplifyDemandedBits which can simplify the nodes used by our
2795   // operands if this node is the only user.
2796   if (TLI.SimplifyDemandedBits(LHS, Demanded, DCI))
2797     return SDValue(Node24, 0);
2798   if (TLI.SimplifyDemandedBits(RHS, Demanded, DCI))
2799     return SDValue(Node24, 0);
2800 
2801   return SDValue();
2802 }
2803 
2804 template <typename IntTy>
2805 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0, uint32_t Offset,
2806                                uint32_t Width, const SDLoc &DL) {
2807   if (Width + Offset < 32) {
2808     uint32_t Shl = static_cast<uint32_t>(Src0) << (32 - Offset - Width);
2809     IntTy Result = static_cast<IntTy>(Shl) >> (32 - Width);
2810     return DAG.getConstant(Result, DL, MVT::i32);
2811   }
2812 
2813   return DAG.getConstant(Src0 >> Offset, DL, MVT::i32);
2814 }
2815 
2816 static bool hasVolatileUser(SDNode *Val) {
2817   for (SDNode *U : Val->uses()) {
2818     if (MemSDNode *M = dyn_cast<MemSDNode>(U)) {
2819       if (M->isVolatile())
2820         return true;
2821     }
2822   }
2823 
2824   return false;
2825 }
2826 
2827 bool AMDGPUTargetLowering::shouldCombineMemoryType(EVT VT) const {
2828   // i32 vectors are the canonical memory type.
2829   if (VT.getScalarType() == MVT::i32 || isTypeLegal(VT))
2830     return false;
2831 
2832   if (!VT.isByteSized())
2833     return false;
2834 
2835   unsigned Size = VT.getStoreSize();
2836 
2837   if ((Size == 1 || Size == 2 || Size == 4) && !VT.isVector())
2838     return false;
2839 
2840   if (Size == 3 || (Size > 4 && (Size % 4 != 0)))
2841     return false;
2842 
2843   return true;
2844 }
2845 
2846 // Replace load of an illegal type with a store of a bitcast to a friendlier
2847 // type.
2848 SDValue AMDGPUTargetLowering::performLoadCombine(SDNode *N,
2849                                                  DAGCombinerInfo &DCI) const {
2850   if (!DCI.isBeforeLegalize())
2851     return SDValue();
2852 
2853   LoadSDNode *LN = cast<LoadSDNode>(N);
2854   if (!LN->isSimple() || !ISD::isNormalLoad(LN) || hasVolatileUser(LN))
2855     return SDValue();
2856 
2857   SDLoc SL(N);
2858   SelectionDAG &DAG = DCI.DAG;
2859   EVT VT = LN->getMemoryVT();
2860 
2861   unsigned Size = VT.getStoreSize();
2862   Align Alignment = LN->getAlign();
2863   if (Alignment < Size && isTypeLegal(VT)) {
2864     bool IsFast;
2865     unsigned AS = LN->getAddressSpace();
2866 
2867     // Expand unaligned loads earlier than legalization. Due to visitation order
2868     // problems during legalization, the emitted instructions to pack and unpack
2869     // the bytes again are not eliminated in the case of an unaligned copy.
2870     if (!allowsMisalignedMemoryAccesses(
2871             VT, AS, Alignment, LN->getMemOperand()->getFlags(), &IsFast)) {
2872       if (VT.isVector())
2873         return SplitVectorLoad(SDValue(LN, 0), DAG);
2874 
2875       SDValue Ops[2];
2876       std::tie(Ops[0], Ops[1]) = expandUnalignedLoad(LN, DAG);
2877 
2878       return DAG.getMergeValues(Ops, SDLoc(N));
2879     }
2880 
2881     if (!IsFast)
2882       return SDValue();
2883   }
2884 
2885   if (!shouldCombineMemoryType(VT))
2886     return SDValue();
2887 
2888   EVT NewVT = getEquivalentMemType(*DAG.getContext(), VT);
2889 
2890   SDValue NewLoad
2891     = DAG.getLoad(NewVT, SL, LN->getChain(),
2892                   LN->getBasePtr(), LN->getMemOperand());
2893 
2894   SDValue BC = DAG.getNode(ISD::BITCAST, SL, VT, NewLoad);
2895   DCI.CombineTo(N, BC, NewLoad.getValue(1));
2896   return SDValue(N, 0);
2897 }
2898 
2899 // Replace store of an illegal type with a store of a bitcast to a friendlier
2900 // type.
2901 SDValue AMDGPUTargetLowering::performStoreCombine(SDNode *N,
2902                                                   DAGCombinerInfo &DCI) const {
2903   if (!DCI.isBeforeLegalize())
2904     return SDValue();
2905 
2906   StoreSDNode *SN = cast<StoreSDNode>(N);
2907   if (!SN->isSimple() || !ISD::isNormalStore(SN))
2908     return SDValue();
2909 
2910   EVT VT = SN->getMemoryVT();
2911   unsigned Size = VT.getStoreSize();
2912 
2913   SDLoc SL(N);
2914   SelectionDAG &DAG = DCI.DAG;
2915   Align Alignment = SN->getAlign();
2916   if (Alignment < Size && isTypeLegal(VT)) {
2917     bool IsFast;
2918     unsigned AS = SN->getAddressSpace();
2919 
2920     // Expand unaligned stores earlier than legalization. Due to visitation
2921     // order problems during legalization, the emitted instructions to pack and
2922     // unpack the bytes again are not eliminated in the case of an unaligned
2923     // copy.
2924     if (!allowsMisalignedMemoryAccesses(
2925             VT, AS, Alignment, SN->getMemOperand()->getFlags(), &IsFast)) {
2926       if (VT.isVector())
2927         return SplitVectorStore(SDValue(SN, 0), DAG);
2928 
2929       return expandUnalignedStore(SN, DAG);
2930     }
2931 
2932     if (!IsFast)
2933       return SDValue();
2934   }
2935 
2936   if (!shouldCombineMemoryType(VT))
2937     return SDValue();
2938 
2939   EVT NewVT = getEquivalentMemType(*DAG.getContext(), VT);
2940   SDValue Val = SN->getValue();
2941 
2942   //DCI.AddToWorklist(Val.getNode());
2943 
2944   bool OtherUses = !Val.hasOneUse();
2945   SDValue CastVal = DAG.getNode(ISD::BITCAST, SL, NewVT, Val);
2946   if (OtherUses) {
2947     SDValue CastBack = DAG.getNode(ISD::BITCAST, SL, VT, CastVal);
2948     DAG.ReplaceAllUsesOfValueWith(Val, CastBack);
2949   }
2950 
2951   return DAG.getStore(SN->getChain(), SL, CastVal,
2952                       SN->getBasePtr(), SN->getMemOperand());
2953 }
2954 
2955 // FIXME: This should go in generic DAG combiner with an isTruncateFree check,
2956 // but isTruncateFree is inaccurate for i16 now because of SALU vs. VALU
2957 // issues.
2958 SDValue AMDGPUTargetLowering::performAssertSZExtCombine(SDNode *N,
2959                                                         DAGCombinerInfo &DCI) const {
2960   SelectionDAG &DAG = DCI.DAG;
2961   SDValue N0 = N->getOperand(0);
2962 
2963   // (vt2 (assertzext (truncate vt0:x), vt1)) ->
2964   //     (vt2 (truncate (assertzext vt0:x, vt1)))
2965   if (N0.getOpcode() == ISD::TRUNCATE) {
2966     SDValue N1 = N->getOperand(1);
2967     EVT ExtVT = cast<VTSDNode>(N1)->getVT();
2968     SDLoc SL(N);
2969 
2970     SDValue Src = N0.getOperand(0);
2971     EVT SrcVT = Src.getValueType();
2972     if (SrcVT.bitsGE(ExtVT)) {
2973       SDValue NewInReg = DAG.getNode(N->getOpcode(), SL, SrcVT, Src, N1);
2974       return DAG.getNode(ISD::TRUNCATE, SL, N->getValueType(0), NewInReg);
2975     }
2976   }
2977 
2978   return SDValue();
2979 }
2980 
2981 SDValue AMDGPUTargetLowering::performIntrinsicWOChainCombine(
2982   SDNode *N, DAGCombinerInfo &DCI) const {
2983   unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
2984   switch (IID) {
2985   case Intrinsic::amdgcn_mul_i24:
2986   case Intrinsic::amdgcn_mul_u24:
2987   case Intrinsic::amdgcn_mulhi_i24:
2988   case Intrinsic::amdgcn_mulhi_u24:
2989     return simplifyMul24(N, DCI);
2990   case Intrinsic::amdgcn_fract:
2991   case Intrinsic::amdgcn_rsq:
2992   case Intrinsic::amdgcn_rcp_legacy:
2993   case Intrinsic::amdgcn_rsq_legacy:
2994   case Intrinsic::amdgcn_rsq_clamp:
2995   case Intrinsic::amdgcn_ldexp: {
2996     // FIXME: This is probably wrong. If src is an sNaN, it won't be quieted
2997     SDValue Src = N->getOperand(1);
2998     return Src.isUndef() ? Src : SDValue();
2999   }
3000   default:
3001     return SDValue();
3002   }
3003 }
3004 
3005 /// Split the 64-bit value \p LHS into two 32-bit components, and perform the
3006 /// binary operation \p Opc to it with the corresponding constant operands.
3007 SDValue AMDGPUTargetLowering::splitBinaryBitConstantOpImpl(
3008   DAGCombinerInfo &DCI, const SDLoc &SL,
3009   unsigned Opc, SDValue LHS,
3010   uint32_t ValLo, uint32_t ValHi) const {
3011   SelectionDAG &DAG = DCI.DAG;
3012   SDValue Lo, Hi;
3013   std::tie(Lo, Hi) = split64BitValue(LHS, DAG);
3014 
3015   SDValue LoRHS = DAG.getConstant(ValLo, SL, MVT::i32);
3016   SDValue HiRHS = DAG.getConstant(ValHi, SL, MVT::i32);
3017 
3018   SDValue LoAnd = DAG.getNode(Opc, SL, MVT::i32, Lo, LoRHS);
3019   SDValue HiAnd = DAG.getNode(Opc, SL, MVT::i32, Hi, HiRHS);
3020 
3021   // Re-visit the ands. It's possible we eliminated one of them and it could
3022   // simplify the vector.
3023   DCI.AddToWorklist(Lo.getNode());
3024   DCI.AddToWorklist(Hi.getNode());
3025 
3026   SDValue Vec = DAG.getBuildVector(MVT::v2i32, SL, {LoAnd, HiAnd});
3027   return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Vec);
3028 }
3029 
3030 SDValue AMDGPUTargetLowering::performShlCombine(SDNode *N,
3031                                                 DAGCombinerInfo &DCI) const {
3032   EVT VT = N->getValueType(0);
3033 
3034   ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
3035   if (!RHS)
3036     return SDValue();
3037 
3038   SDValue LHS = N->getOperand(0);
3039   unsigned RHSVal = RHS->getZExtValue();
3040   if (!RHSVal)
3041     return LHS;
3042 
3043   SDLoc SL(N);
3044   SelectionDAG &DAG = DCI.DAG;
3045 
3046   switch (LHS->getOpcode()) {
3047   default:
3048     break;
3049   case ISD::ZERO_EXTEND:
3050   case ISD::SIGN_EXTEND:
3051   case ISD::ANY_EXTEND: {
3052     SDValue X = LHS->getOperand(0);
3053 
3054     if (VT == MVT::i32 && RHSVal == 16 && X.getValueType() == MVT::i16 &&
3055         isOperationLegal(ISD::BUILD_VECTOR, MVT::v2i16)) {
3056       // Prefer build_vector as the canonical form if packed types are legal.
3057       // (shl ([asz]ext i16:x), 16 -> build_vector 0, x
3058       SDValue Vec = DAG.getBuildVector(MVT::v2i16, SL,
3059        { DAG.getConstant(0, SL, MVT::i16), LHS->getOperand(0) });
3060       return DAG.getNode(ISD::BITCAST, SL, MVT::i32, Vec);
3061     }
3062 
3063     // shl (ext x) => zext (shl x), if shift does not overflow int
3064     if (VT != MVT::i64)
3065       break;
3066     KnownBits Known = DAG.computeKnownBits(X);
3067     unsigned LZ = Known.countMinLeadingZeros();
3068     if (LZ < RHSVal)
3069       break;
3070     EVT XVT = X.getValueType();
3071     SDValue Shl = DAG.getNode(ISD::SHL, SL, XVT, X, SDValue(RHS, 0));
3072     return DAG.getZExtOrTrunc(Shl, SL, VT);
3073   }
3074   }
3075 
3076   if (VT != MVT::i64)
3077     return SDValue();
3078 
3079   // i64 (shl x, C) -> (build_pair 0, (shl x, C -32))
3080 
3081   // On some subtargets, 64-bit shift is a quarter rate instruction. In the
3082   // common case, splitting this into a move and a 32-bit shift is faster and
3083   // the same code size.
3084   if (RHSVal < 32)
3085     return SDValue();
3086 
3087   SDValue ShiftAmt = DAG.getConstant(RHSVal - 32, SL, MVT::i32);
3088 
3089   SDValue Lo = DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, LHS);
3090   SDValue NewShift = DAG.getNode(ISD::SHL, SL, MVT::i32, Lo, ShiftAmt);
3091 
3092   const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
3093 
3094   SDValue Vec = DAG.getBuildVector(MVT::v2i32, SL, {Zero, NewShift});
3095   return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Vec);
3096 }
3097 
3098 SDValue AMDGPUTargetLowering::performSraCombine(SDNode *N,
3099                                                 DAGCombinerInfo &DCI) const {
3100   if (N->getValueType(0) != MVT::i64)
3101     return SDValue();
3102 
3103   const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
3104   if (!RHS)
3105     return SDValue();
3106 
3107   SelectionDAG &DAG = DCI.DAG;
3108   SDLoc SL(N);
3109   unsigned RHSVal = RHS->getZExtValue();
3110 
3111   // (sra i64:x, 32) -> build_pair x, (sra hi_32(x), 31)
3112   if (RHSVal == 32) {
3113     SDValue Hi = getHiHalf64(N->getOperand(0), DAG);
3114     SDValue NewShift = DAG.getNode(ISD::SRA, SL, MVT::i32, Hi,
3115                                    DAG.getConstant(31, SL, MVT::i32));
3116 
3117     SDValue BuildVec = DAG.getBuildVector(MVT::v2i32, SL, {Hi, NewShift});
3118     return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildVec);
3119   }
3120 
3121   // (sra i64:x, 63) -> build_pair (sra hi_32(x), 31), (sra hi_32(x), 31)
3122   if (RHSVal == 63) {
3123     SDValue Hi = getHiHalf64(N->getOperand(0), DAG);
3124     SDValue NewShift = DAG.getNode(ISD::SRA, SL, MVT::i32, Hi,
3125                                    DAG.getConstant(31, SL, MVT::i32));
3126     SDValue BuildVec = DAG.getBuildVector(MVT::v2i32, SL, {NewShift, NewShift});
3127     return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildVec);
3128   }
3129 
3130   return SDValue();
3131 }
3132 
3133 SDValue AMDGPUTargetLowering::performSrlCombine(SDNode *N,
3134                                                 DAGCombinerInfo &DCI) const {
3135   auto *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
3136   if (!RHS)
3137     return SDValue();
3138 
3139   EVT VT = N->getValueType(0);
3140   SDValue LHS = N->getOperand(0);
3141   unsigned ShiftAmt = RHS->getZExtValue();
3142   SelectionDAG &DAG = DCI.DAG;
3143   SDLoc SL(N);
3144 
3145   // fold (srl (and x, c1 << c2), c2) -> (and (srl(x, c2), c1)
3146   // this improves the ability to match BFE patterns in isel.
3147   if (LHS.getOpcode() == ISD::AND) {
3148     if (auto *Mask = dyn_cast<ConstantSDNode>(LHS.getOperand(1))) {
3149       unsigned MaskIdx, MaskLen;
3150       if (Mask->getAPIntValue().isShiftedMask(MaskIdx, MaskLen) &&
3151           MaskIdx == ShiftAmt) {
3152         return DAG.getNode(
3153             ISD::AND, SL, VT,
3154             DAG.getNode(ISD::SRL, SL, VT, LHS.getOperand(0), N->getOperand(1)),
3155             DAG.getNode(ISD::SRL, SL, VT, LHS.getOperand(1), N->getOperand(1)));
3156       }
3157     }
3158   }
3159 
3160   if (VT != MVT::i64)
3161     return SDValue();
3162 
3163   if (ShiftAmt < 32)
3164     return SDValue();
3165 
3166   // srl i64:x, C for C >= 32
3167   // =>
3168   //   build_pair (srl hi_32(x), C - 32), 0
3169   SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
3170 
3171   SDValue Hi = getHiHalf64(LHS, DAG);
3172 
3173   SDValue NewConst = DAG.getConstant(ShiftAmt - 32, SL, MVT::i32);
3174   SDValue NewShift = DAG.getNode(ISD::SRL, SL, MVT::i32, Hi, NewConst);
3175 
3176   SDValue BuildPair = DAG.getBuildVector(MVT::v2i32, SL, {NewShift, Zero});
3177 
3178   return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildPair);
3179 }
3180 
3181 SDValue AMDGPUTargetLowering::performTruncateCombine(
3182   SDNode *N, DAGCombinerInfo &DCI) const {
3183   SDLoc SL(N);
3184   SelectionDAG &DAG = DCI.DAG;
3185   EVT VT = N->getValueType(0);
3186   SDValue Src = N->getOperand(0);
3187 
3188   // vt1 (truncate (bitcast (build_vector vt0:x, ...))) -> vt1 (bitcast vt0:x)
3189   if (Src.getOpcode() == ISD::BITCAST && !VT.isVector()) {
3190     SDValue Vec = Src.getOperand(0);
3191     if (Vec.getOpcode() == ISD::BUILD_VECTOR) {
3192       SDValue Elt0 = Vec.getOperand(0);
3193       EVT EltVT = Elt0.getValueType();
3194       if (VT.getFixedSizeInBits() <= EltVT.getFixedSizeInBits()) {
3195         if (EltVT.isFloatingPoint()) {
3196           Elt0 = DAG.getNode(ISD::BITCAST, SL,
3197                              EltVT.changeTypeToInteger(), Elt0);
3198         }
3199 
3200         return DAG.getNode(ISD::TRUNCATE, SL, VT, Elt0);
3201       }
3202     }
3203   }
3204 
3205   // Equivalent of above for accessing the high element of a vector as an
3206   // integer operation.
3207   // trunc (srl (bitcast (build_vector x, y))), 16 -> trunc (bitcast y)
3208   if (Src.getOpcode() == ISD::SRL && !VT.isVector()) {
3209     if (auto K = isConstOrConstSplat(Src.getOperand(1))) {
3210       if (2 * K->getZExtValue() == Src.getValueType().getScalarSizeInBits()) {
3211         SDValue BV = stripBitcast(Src.getOperand(0));
3212         if (BV.getOpcode() == ISD::BUILD_VECTOR &&
3213             BV.getValueType().getVectorNumElements() == 2) {
3214           SDValue SrcElt = BV.getOperand(1);
3215           EVT SrcEltVT = SrcElt.getValueType();
3216           if (SrcEltVT.isFloatingPoint()) {
3217             SrcElt = DAG.getNode(ISD::BITCAST, SL,
3218                                  SrcEltVT.changeTypeToInteger(), SrcElt);
3219           }
3220 
3221           return DAG.getNode(ISD::TRUNCATE, SL, VT, SrcElt);
3222         }
3223       }
3224     }
3225   }
3226 
3227   // Partially shrink 64-bit shifts to 32-bit if reduced to 16-bit.
3228   //
3229   // i16 (trunc (srl i64:x, K)), K <= 16 ->
3230   //     i16 (trunc (srl (i32 (trunc x), K)))
3231   if (VT.getScalarSizeInBits() < 32) {
3232     EVT SrcVT = Src.getValueType();
3233     if (SrcVT.getScalarSizeInBits() > 32 &&
3234         (Src.getOpcode() == ISD::SRL ||
3235          Src.getOpcode() == ISD::SRA ||
3236          Src.getOpcode() == ISD::SHL)) {
3237       SDValue Amt = Src.getOperand(1);
3238       KnownBits Known = DAG.computeKnownBits(Amt);
3239       unsigned Size = VT.getScalarSizeInBits();
3240       if ((Known.isConstant() && Known.getConstant().ule(Size)) ||
3241           (Known.countMaxActiveBits() <= Log2_32(Size))) {
3242         EVT MidVT = VT.isVector() ?
3243           EVT::getVectorVT(*DAG.getContext(), MVT::i32,
3244                            VT.getVectorNumElements()) : MVT::i32;
3245 
3246         EVT NewShiftVT = getShiftAmountTy(MidVT, DAG.getDataLayout());
3247         SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SL, MidVT,
3248                                     Src.getOperand(0));
3249         DCI.AddToWorklist(Trunc.getNode());
3250 
3251         if (Amt.getValueType() != NewShiftVT) {
3252           Amt = DAG.getZExtOrTrunc(Amt, SL, NewShiftVT);
3253           DCI.AddToWorklist(Amt.getNode());
3254         }
3255 
3256         SDValue ShrunkShift = DAG.getNode(Src.getOpcode(), SL, MidVT,
3257                                           Trunc, Amt);
3258         return DAG.getNode(ISD::TRUNCATE, SL, VT, ShrunkShift);
3259       }
3260     }
3261   }
3262 
3263   return SDValue();
3264 }
3265 
3266 // We need to specifically handle i64 mul here to avoid unnecessary conversion
3267 // instructions. If we only match on the legalized i64 mul expansion,
3268 // SimplifyDemandedBits will be unable to remove them because there will be
3269 // multiple uses due to the separate mul + mulh[su].
3270 static SDValue getMul24(SelectionDAG &DAG, const SDLoc &SL,
3271                         SDValue N0, SDValue N1, unsigned Size, bool Signed) {
3272   if (Size <= 32) {
3273     unsigned MulOpc = Signed ? AMDGPUISD::MUL_I24 : AMDGPUISD::MUL_U24;
3274     return DAG.getNode(MulOpc, SL, MVT::i32, N0, N1);
3275   }
3276 
3277   unsigned MulLoOpc = Signed ? AMDGPUISD::MUL_I24 : AMDGPUISD::MUL_U24;
3278   unsigned MulHiOpc = Signed ? AMDGPUISD::MULHI_I24 : AMDGPUISD::MULHI_U24;
3279 
3280   SDValue MulLo = DAG.getNode(MulLoOpc, SL, MVT::i32, N0, N1);
3281   SDValue MulHi = DAG.getNode(MulHiOpc, SL, MVT::i32, N0, N1);
3282 
3283   return DAG.getNode(ISD::BUILD_PAIR, SL, MVT::i64, MulLo, MulHi);
3284 }
3285 
3286 SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N,
3287                                                 DAGCombinerInfo &DCI) const {
3288   EVT VT = N->getValueType(0);
3289 
3290   // Don't generate 24-bit multiplies on values that are in SGPRs, since
3291   // we only have a 32-bit scalar multiply (avoid values being moved to VGPRs
3292   // unnecessarily). isDivergent() is used as an approximation of whether the
3293   // value is in an SGPR.
3294   if (!N->isDivergent())
3295     return SDValue();
3296 
3297   unsigned Size = VT.getSizeInBits();
3298   if (VT.isVector() || Size > 64)
3299     return SDValue();
3300 
3301   // There are i16 integer mul/mad.
3302   if (Subtarget->has16BitInsts() && VT.getScalarType().bitsLE(MVT::i16))
3303     return SDValue();
3304 
3305   SelectionDAG &DAG = DCI.DAG;
3306   SDLoc DL(N);
3307 
3308   SDValue N0 = N->getOperand(0);
3309   SDValue N1 = N->getOperand(1);
3310 
3311   // SimplifyDemandedBits has the annoying habit of turning useful zero_extends
3312   // in the source into any_extends if the result of the mul is truncated. Since
3313   // we can assume the high bits are whatever we want, use the underlying value
3314   // to avoid the unknown high bits from interfering.
3315   if (N0.getOpcode() == ISD::ANY_EXTEND)
3316     N0 = N0.getOperand(0);
3317 
3318   if (N1.getOpcode() == ISD::ANY_EXTEND)
3319     N1 = N1.getOperand(0);
3320 
3321   SDValue Mul;
3322 
3323   if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
3324     N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
3325     N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
3326     Mul = getMul24(DAG, DL, N0, N1, Size, false);
3327   } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
3328     N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
3329     N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
3330     Mul = getMul24(DAG, DL, N0, N1, Size, true);
3331   } else {
3332     return SDValue();
3333   }
3334 
3335   // We need to use sext even for MUL_U24, because MUL_U24 is used
3336   // for signed multiply of 8 and 16-bit types.
3337   return DAG.getSExtOrTrunc(Mul, DL, VT);
3338 }
3339 
3340 SDValue
3341 AMDGPUTargetLowering::performMulLoHiCombine(SDNode *N,
3342                                             DAGCombinerInfo &DCI) const {
3343   if (N->getValueType(0) != MVT::i32)
3344     return SDValue();
3345 
3346   SelectionDAG &DAG = DCI.DAG;
3347   SDLoc DL(N);
3348 
3349   SDValue N0 = N->getOperand(0);
3350   SDValue N1 = N->getOperand(1);
3351 
3352   // SimplifyDemandedBits has the annoying habit of turning useful zero_extends
3353   // in the source into any_extends if the result of the mul is truncated. Since
3354   // we can assume the high bits are whatever we want, use the underlying value
3355   // to avoid the unknown high bits from interfering.
3356   if (N0.getOpcode() == ISD::ANY_EXTEND)
3357     N0 = N0.getOperand(0);
3358   if (N1.getOpcode() == ISD::ANY_EXTEND)
3359     N1 = N1.getOperand(0);
3360 
3361   // Try to use two fast 24-bit multiplies (one for each half of the result)
3362   // instead of one slow extending multiply.
3363   unsigned LoOpcode, HiOpcode;
3364   if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
3365     N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
3366     N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
3367     LoOpcode = AMDGPUISD::MUL_U24;
3368     HiOpcode = AMDGPUISD::MULHI_U24;
3369   } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
3370     N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
3371     N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
3372     LoOpcode = AMDGPUISD::MUL_I24;
3373     HiOpcode = AMDGPUISD::MULHI_I24;
3374   } else {
3375     return SDValue();
3376   }
3377 
3378   SDValue Lo = DAG.getNode(LoOpcode, DL, MVT::i32, N0, N1);
3379   SDValue Hi = DAG.getNode(HiOpcode, DL, MVT::i32, N0, N1);
3380   DCI.CombineTo(N, Lo, Hi);
3381   return SDValue(N, 0);
3382 }
3383 
3384 SDValue AMDGPUTargetLowering::performMulhsCombine(SDNode *N,
3385                                                   DAGCombinerInfo &DCI) const {
3386   EVT VT = N->getValueType(0);
3387 
3388   if (!Subtarget->hasMulI24() || VT.isVector())
3389     return SDValue();
3390 
3391   // Don't generate 24-bit multiplies on values that are in SGPRs, since
3392   // we only have a 32-bit scalar multiply (avoid values being moved to VGPRs
3393   // unnecessarily). isDivergent() is used as an approximation of whether the
3394   // value is in an SGPR.
3395   // This doesn't apply if no s_mul_hi is available (since we'll end up with a
3396   // valu op anyway)
3397   if (Subtarget->hasSMulHi() && !N->isDivergent())
3398     return SDValue();
3399 
3400   SelectionDAG &DAG = DCI.DAG;
3401   SDLoc DL(N);
3402 
3403   SDValue N0 = N->getOperand(0);
3404   SDValue N1 = N->getOperand(1);
3405 
3406   if (!isI24(N0, DAG) || !isI24(N1, DAG))
3407     return SDValue();
3408 
3409   N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
3410   N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
3411 
3412   SDValue Mulhi = DAG.getNode(AMDGPUISD::MULHI_I24, DL, MVT::i32, N0, N1);
3413   DCI.AddToWorklist(Mulhi.getNode());
3414   return DAG.getSExtOrTrunc(Mulhi, DL, VT);
3415 }
3416 
3417 SDValue AMDGPUTargetLowering::performMulhuCombine(SDNode *N,
3418                                                   DAGCombinerInfo &DCI) const {
3419   EVT VT = N->getValueType(0);
3420 
3421   if (!Subtarget->hasMulU24() || VT.isVector() || VT.getSizeInBits() > 32)
3422     return SDValue();
3423 
3424   // Don't generate 24-bit multiplies on values that are in SGPRs, since
3425   // we only have a 32-bit scalar multiply (avoid values being moved to VGPRs
3426   // unnecessarily). isDivergent() is used as an approximation of whether the
3427   // value is in an SGPR.
3428   // This doesn't apply if no s_mul_hi is available (since we'll end up with a
3429   // valu op anyway)
3430   if (Subtarget->hasSMulHi() && !N->isDivergent())
3431     return SDValue();
3432 
3433   SelectionDAG &DAG = DCI.DAG;
3434   SDLoc DL(N);
3435 
3436   SDValue N0 = N->getOperand(0);
3437   SDValue N1 = N->getOperand(1);
3438 
3439   if (!isU24(N0, DAG) || !isU24(N1, DAG))
3440     return SDValue();
3441 
3442   N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
3443   N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
3444 
3445   SDValue Mulhi = DAG.getNode(AMDGPUISD::MULHI_U24, DL, MVT::i32, N0, N1);
3446   DCI.AddToWorklist(Mulhi.getNode());
3447   return DAG.getZExtOrTrunc(Mulhi, DL, VT);
3448 }
3449 
3450 static bool isNegativeOne(SDValue Val) {
3451   if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val))
3452     return C->isAllOnes();
3453   return false;
3454 }
3455 
3456 SDValue AMDGPUTargetLowering::getFFBX_U32(SelectionDAG &DAG,
3457                                           SDValue Op,
3458                                           const SDLoc &DL,
3459                                           unsigned Opc) const {
3460   EVT VT = Op.getValueType();
3461   EVT LegalVT = getTypeToTransformTo(*DAG.getContext(), VT);
3462   if (LegalVT != MVT::i32 && (Subtarget->has16BitInsts() &&
3463                               LegalVT != MVT::i16))
3464     return SDValue();
3465 
3466   if (VT != MVT::i32)
3467     Op = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, Op);
3468 
3469   SDValue FFBX = DAG.getNode(Opc, DL, MVT::i32, Op);
3470   if (VT != MVT::i32)
3471     FFBX = DAG.getNode(ISD::TRUNCATE, DL, VT, FFBX);
3472 
3473   return FFBX;
3474 }
3475 
3476 // The native instructions return -1 on 0 input. Optimize out a select that
3477 // produces -1 on 0.
3478 //
3479 // TODO: If zero is not undef, we could also do this if the output is compared
3480 // against the bitwidth.
3481 //
3482 // TODO: Should probably combine against FFBH_U32 instead of ctlz directly.
3483 SDValue AMDGPUTargetLowering::performCtlz_CttzCombine(const SDLoc &SL, SDValue Cond,
3484                                                  SDValue LHS, SDValue RHS,
3485                                                  DAGCombinerInfo &DCI) const {
3486   ConstantSDNode *CmpRhs = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3487   if (!CmpRhs || !CmpRhs->isZero())
3488     return SDValue();
3489 
3490   SelectionDAG &DAG = DCI.DAG;
3491   ISD::CondCode CCOpcode = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
3492   SDValue CmpLHS = Cond.getOperand(0);
3493 
3494   // select (setcc x, 0, eq), -1, (ctlz_zero_undef x) -> ffbh_u32 x
3495   // select (setcc x, 0, eq), -1, (cttz_zero_undef x) -> ffbl_u32 x
3496   if (CCOpcode == ISD::SETEQ &&
3497       (isCtlzOpc(RHS.getOpcode()) || isCttzOpc(RHS.getOpcode())) &&
3498       RHS.getOperand(0) == CmpLHS && isNegativeOne(LHS)) {
3499     unsigned Opc =
3500         isCttzOpc(RHS.getOpcode()) ? AMDGPUISD::FFBL_B32 : AMDGPUISD::FFBH_U32;
3501     return getFFBX_U32(DAG, CmpLHS, SL, Opc);
3502   }
3503 
3504   // select (setcc x, 0, ne), (ctlz_zero_undef x), -1 -> ffbh_u32 x
3505   // select (setcc x, 0, ne), (cttz_zero_undef x), -1 -> ffbl_u32 x
3506   if (CCOpcode == ISD::SETNE &&
3507       (isCtlzOpc(LHS.getOpcode()) || isCttzOpc(LHS.getOpcode())) &&
3508       LHS.getOperand(0) == CmpLHS && isNegativeOne(RHS)) {
3509     unsigned Opc =
3510         isCttzOpc(LHS.getOpcode()) ? AMDGPUISD::FFBL_B32 : AMDGPUISD::FFBH_U32;
3511 
3512     return getFFBX_U32(DAG, CmpLHS, SL, Opc);
3513   }
3514 
3515   return SDValue();
3516 }
3517 
3518 static SDValue distributeOpThroughSelect(TargetLowering::DAGCombinerInfo &DCI,
3519                                          unsigned Op,
3520                                          const SDLoc &SL,
3521                                          SDValue Cond,
3522                                          SDValue N1,
3523                                          SDValue N2) {
3524   SelectionDAG &DAG = DCI.DAG;
3525   EVT VT = N1.getValueType();
3526 
3527   SDValue NewSelect = DAG.getNode(ISD::SELECT, SL, VT, Cond,
3528                                   N1.getOperand(0), N2.getOperand(0));
3529   DCI.AddToWorklist(NewSelect.getNode());
3530   return DAG.getNode(Op, SL, VT, NewSelect);
3531 }
3532 
3533 // Pull a free FP operation out of a select so it may fold into uses.
3534 //
3535 // select c, (fneg x), (fneg y) -> fneg (select c, x, y)
3536 // select c, (fneg x), k -> fneg (select c, x, (fneg k))
3537 //
3538 // select c, (fabs x), (fabs y) -> fabs (select c, x, y)
3539 // select c, (fabs x), +k -> fabs (select c, x, k)
3540 static SDValue foldFreeOpFromSelect(TargetLowering::DAGCombinerInfo &DCI,
3541                                     SDValue N) {
3542   SelectionDAG &DAG = DCI.DAG;
3543   SDValue Cond = N.getOperand(0);
3544   SDValue LHS = N.getOperand(1);
3545   SDValue RHS = N.getOperand(2);
3546 
3547   EVT VT = N.getValueType();
3548   if ((LHS.getOpcode() == ISD::FABS && RHS.getOpcode() == ISD::FABS) ||
3549       (LHS.getOpcode() == ISD::FNEG && RHS.getOpcode() == ISD::FNEG)) {
3550     return distributeOpThroughSelect(DCI, LHS.getOpcode(),
3551                                      SDLoc(N), Cond, LHS, RHS);
3552   }
3553 
3554   bool Inv = false;
3555   if (RHS.getOpcode() == ISD::FABS || RHS.getOpcode() == ISD::FNEG) {
3556     std::swap(LHS, RHS);
3557     Inv = true;
3558   }
3559 
3560   // TODO: Support vector constants.
3561   ConstantFPSDNode *CRHS = dyn_cast<ConstantFPSDNode>(RHS);
3562   if ((LHS.getOpcode() == ISD::FNEG || LHS.getOpcode() == ISD::FABS) && CRHS) {
3563     SDLoc SL(N);
3564     // If one side is an fneg/fabs and the other is a constant, we can push the
3565     // fneg/fabs down. If it's an fabs, the constant needs to be non-negative.
3566     SDValue NewLHS = LHS.getOperand(0);
3567     SDValue NewRHS = RHS;
3568 
3569     // Careful: if the neg can be folded up, don't try to pull it back down.
3570     bool ShouldFoldNeg = true;
3571 
3572     if (NewLHS.hasOneUse()) {
3573       unsigned Opc = NewLHS.getOpcode();
3574       if (LHS.getOpcode() == ISD::FNEG && fnegFoldsIntoOp(Opc))
3575         ShouldFoldNeg = false;
3576       if (LHS.getOpcode() == ISD::FABS && Opc == ISD::FMUL)
3577         ShouldFoldNeg = false;
3578     }
3579 
3580     if (ShouldFoldNeg) {
3581       if (LHS.getOpcode() == ISD::FNEG)
3582         NewRHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
3583       else if (CRHS->isNegative())
3584         return SDValue();
3585 
3586       if (Inv)
3587         std::swap(NewLHS, NewRHS);
3588 
3589       SDValue NewSelect = DAG.getNode(ISD::SELECT, SL, VT,
3590                                       Cond, NewLHS, NewRHS);
3591       DCI.AddToWorklist(NewSelect.getNode());
3592       return DAG.getNode(LHS.getOpcode(), SL, VT, NewSelect);
3593     }
3594   }
3595 
3596   return SDValue();
3597 }
3598 
3599 
3600 SDValue AMDGPUTargetLowering::performSelectCombine(SDNode *N,
3601                                                    DAGCombinerInfo &DCI) const {
3602   if (SDValue Folded = foldFreeOpFromSelect(DCI, SDValue(N, 0)))
3603     return Folded;
3604 
3605   SDValue Cond = N->getOperand(0);
3606   if (Cond.getOpcode() != ISD::SETCC)
3607     return SDValue();
3608 
3609   EVT VT = N->getValueType(0);
3610   SDValue LHS = Cond.getOperand(0);
3611   SDValue RHS = Cond.getOperand(1);
3612   SDValue CC = Cond.getOperand(2);
3613 
3614   SDValue True = N->getOperand(1);
3615   SDValue False = N->getOperand(2);
3616 
3617   if (Cond.hasOneUse()) { // TODO: Look for multiple select uses.
3618     SelectionDAG &DAG = DCI.DAG;
3619     if (DAG.isConstantValueOfAnyType(True) &&
3620         !DAG.isConstantValueOfAnyType(False)) {
3621       // Swap cmp + select pair to move constant to false input.
3622       // This will allow using VOPC cndmasks more often.
3623       // select (setcc x, y), k, x -> select (setccinv x, y), x, k
3624 
3625       SDLoc SL(N);
3626       ISD::CondCode NewCC =
3627           getSetCCInverse(cast<CondCodeSDNode>(CC)->get(), LHS.getValueType());
3628 
3629       SDValue NewCond = DAG.getSetCC(SL, Cond.getValueType(), LHS, RHS, NewCC);
3630       return DAG.getNode(ISD::SELECT, SL, VT, NewCond, False, True);
3631     }
3632 
3633     if (VT == MVT::f32 && Subtarget->hasFminFmaxLegacy()) {
3634       SDValue MinMax
3635         = combineFMinMaxLegacy(SDLoc(N), VT, LHS, RHS, True, False, CC, DCI);
3636       // Revisit this node so we can catch min3/max3/med3 patterns.
3637       //DCI.AddToWorklist(MinMax.getNode());
3638       return MinMax;
3639     }
3640   }
3641 
3642   // There's no reason to not do this if the condition has other uses.
3643   return performCtlz_CttzCombine(SDLoc(N), Cond, True, False, DCI);
3644 }
3645 
3646 static bool isInv2Pi(const APFloat &APF) {
3647   static const APFloat KF16(APFloat::IEEEhalf(), APInt(16, 0x3118));
3648   static const APFloat KF32(APFloat::IEEEsingle(), APInt(32, 0x3e22f983));
3649   static const APFloat KF64(APFloat::IEEEdouble(), APInt(64, 0x3fc45f306dc9c882));
3650 
3651   return APF.bitwiseIsEqual(KF16) ||
3652          APF.bitwiseIsEqual(KF32) ||
3653          APF.bitwiseIsEqual(KF64);
3654 }
3655 
3656 // 0 and 1.0 / (0.5 * pi) do not have inline immmediates, so there is an
3657 // additional cost to negate them.
3658 bool AMDGPUTargetLowering::isConstantCostlierToNegate(SDValue N) const {
3659   if (const ConstantFPSDNode *C = isConstOrConstSplatFP(N)) {
3660     if (C->isZero() && !C->isNegative())
3661       return true;
3662 
3663     if (Subtarget->hasInv2PiInlineImm() && isInv2Pi(C->getValueAPF()))
3664       return true;
3665   }
3666 
3667   return false;
3668 }
3669 
3670 static unsigned inverseMinMax(unsigned Opc) {
3671   switch (Opc) {
3672   case ISD::FMAXNUM:
3673     return ISD::FMINNUM;
3674   case ISD::FMINNUM:
3675     return ISD::FMAXNUM;
3676   case ISD::FMAXNUM_IEEE:
3677     return ISD::FMINNUM_IEEE;
3678   case ISD::FMINNUM_IEEE:
3679     return ISD::FMAXNUM_IEEE;
3680   case AMDGPUISD::FMAX_LEGACY:
3681     return AMDGPUISD::FMIN_LEGACY;
3682   case AMDGPUISD::FMIN_LEGACY:
3683     return  AMDGPUISD::FMAX_LEGACY;
3684   default:
3685     llvm_unreachable("invalid min/max opcode");
3686   }
3687 }
3688 
3689 SDValue AMDGPUTargetLowering::performFNegCombine(SDNode *N,
3690                                                  DAGCombinerInfo &DCI) const {
3691   SelectionDAG &DAG = DCI.DAG;
3692   SDValue N0 = N->getOperand(0);
3693   EVT VT = N->getValueType(0);
3694 
3695   unsigned Opc = N0.getOpcode();
3696 
3697   // If the input has multiple uses and we can either fold the negate down, or
3698   // the other uses cannot, give up. This both prevents unprofitable
3699   // transformations and infinite loops: we won't repeatedly try to fold around
3700   // a negate that has no 'good' form.
3701   if (N0.hasOneUse()) {
3702     // This may be able to fold into the source, but at a code size cost. Don't
3703     // fold if the fold into the user is free.
3704     if (allUsesHaveSourceMods(N, 0))
3705       return SDValue();
3706   } else {
3707     if (fnegFoldsIntoOp(Opc) &&
3708         (allUsesHaveSourceMods(N) || !allUsesHaveSourceMods(N0.getNode())))
3709       return SDValue();
3710   }
3711 
3712   SDLoc SL(N);
3713   switch (Opc) {
3714   case ISD::FADD: {
3715     if (!mayIgnoreSignedZero(N0))
3716       return SDValue();
3717 
3718     // (fneg (fadd x, y)) -> (fadd (fneg x), (fneg y))
3719     SDValue LHS = N0.getOperand(0);
3720     SDValue RHS = N0.getOperand(1);
3721 
3722     if (LHS.getOpcode() != ISD::FNEG)
3723       LHS = DAG.getNode(ISD::FNEG, SL, VT, LHS);
3724     else
3725       LHS = LHS.getOperand(0);
3726 
3727     if (RHS.getOpcode() != ISD::FNEG)
3728       RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
3729     else
3730       RHS = RHS.getOperand(0);
3731 
3732     SDValue Res = DAG.getNode(ISD::FADD, SL, VT, LHS, RHS, N0->getFlags());
3733     if (Res.getOpcode() != ISD::FADD)
3734       return SDValue(); // Op got folded away.
3735     if (!N0.hasOneUse())
3736       DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
3737     return Res;
3738   }
3739   case ISD::FMUL:
3740   case AMDGPUISD::FMUL_LEGACY: {
3741     // (fneg (fmul x, y)) -> (fmul x, (fneg y))
3742     // (fneg (fmul_legacy x, y)) -> (fmul_legacy x, (fneg y))
3743     SDValue LHS = N0.getOperand(0);
3744     SDValue RHS = N0.getOperand(1);
3745 
3746     if (LHS.getOpcode() == ISD::FNEG)
3747       LHS = LHS.getOperand(0);
3748     else if (RHS.getOpcode() == ISD::FNEG)
3749       RHS = RHS.getOperand(0);
3750     else
3751       RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
3752 
3753     SDValue Res = DAG.getNode(Opc, SL, VT, LHS, RHS, N0->getFlags());
3754     if (Res.getOpcode() != Opc)
3755       return SDValue(); // Op got folded away.
3756     if (!N0.hasOneUse())
3757       DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
3758     return Res;
3759   }
3760   case ISD::FMA:
3761   case ISD::FMAD: {
3762     // TODO: handle llvm.amdgcn.fma.legacy
3763     if (!mayIgnoreSignedZero(N0))
3764       return SDValue();
3765 
3766     // (fneg (fma x, y, z)) -> (fma x, (fneg y), (fneg z))
3767     SDValue LHS = N0.getOperand(0);
3768     SDValue MHS = N0.getOperand(1);
3769     SDValue RHS = N0.getOperand(2);
3770 
3771     if (LHS.getOpcode() == ISD::FNEG)
3772       LHS = LHS.getOperand(0);
3773     else if (MHS.getOpcode() == ISD::FNEG)
3774       MHS = MHS.getOperand(0);
3775     else
3776       MHS = DAG.getNode(ISD::FNEG, SL, VT, MHS);
3777 
3778     if (RHS.getOpcode() != ISD::FNEG)
3779       RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
3780     else
3781       RHS = RHS.getOperand(0);
3782 
3783     SDValue Res = DAG.getNode(Opc, SL, VT, LHS, MHS, RHS);
3784     if (Res.getOpcode() != Opc)
3785       return SDValue(); // Op got folded away.
3786     if (!N0.hasOneUse())
3787       DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
3788     return Res;
3789   }
3790   case ISD::FMAXNUM:
3791   case ISD::FMINNUM:
3792   case ISD::FMAXNUM_IEEE:
3793   case ISD::FMINNUM_IEEE:
3794   case AMDGPUISD::FMAX_LEGACY:
3795   case AMDGPUISD::FMIN_LEGACY: {
3796     // fneg (fmaxnum x, y) -> fminnum (fneg x), (fneg y)
3797     // fneg (fminnum x, y) -> fmaxnum (fneg x), (fneg y)
3798     // fneg (fmax_legacy x, y) -> fmin_legacy (fneg x), (fneg y)
3799     // fneg (fmin_legacy x, y) -> fmax_legacy (fneg x), (fneg y)
3800 
3801     SDValue LHS = N0.getOperand(0);
3802     SDValue RHS = N0.getOperand(1);
3803 
3804     // 0 doesn't have a negated inline immediate.
3805     // TODO: This constant check should be generalized to other operations.
3806     if (isConstantCostlierToNegate(RHS))
3807       return SDValue();
3808 
3809     SDValue NegLHS = DAG.getNode(ISD::FNEG, SL, VT, LHS);
3810     SDValue NegRHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
3811     unsigned Opposite = inverseMinMax(Opc);
3812 
3813     SDValue Res = DAG.getNode(Opposite, SL, VT, NegLHS, NegRHS, N0->getFlags());
3814     if (Res.getOpcode() != Opposite)
3815       return SDValue(); // Op got folded away.
3816     if (!N0.hasOneUse())
3817       DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
3818     return Res;
3819   }
3820   case AMDGPUISD::FMED3: {
3821     SDValue Ops[3];
3822     for (unsigned I = 0; I < 3; ++I)
3823       Ops[I] = DAG.getNode(ISD::FNEG, SL, VT, N0->getOperand(I), N0->getFlags());
3824 
3825     SDValue Res = DAG.getNode(AMDGPUISD::FMED3, SL, VT, Ops, N0->getFlags());
3826     if (Res.getOpcode() != AMDGPUISD::FMED3)
3827       return SDValue(); // Op got folded away.
3828 
3829     if (!N0.hasOneUse()) {
3830       SDValue Neg = DAG.getNode(ISD::FNEG, SL, VT, Res);
3831       DAG.ReplaceAllUsesWith(N0, Neg);
3832 
3833       for (SDNode *U : Neg->uses())
3834         DCI.AddToWorklist(U);
3835     }
3836 
3837     return Res;
3838   }
3839   case ISD::FP_EXTEND:
3840   case ISD::FTRUNC:
3841   case ISD::FRINT:
3842   case ISD::FNEARBYINT: // XXX - Should fround be handled?
3843   case ISD::FSIN:
3844   case ISD::FCANONICALIZE:
3845   case AMDGPUISD::RCP:
3846   case AMDGPUISD::RCP_LEGACY:
3847   case AMDGPUISD::RCP_IFLAG:
3848   case AMDGPUISD::SIN_HW: {
3849     SDValue CvtSrc = N0.getOperand(0);
3850     if (CvtSrc.getOpcode() == ISD::FNEG) {
3851       // (fneg (fp_extend (fneg x))) -> (fp_extend x)
3852       // (fneg (rcp (fneg x))) -> (rcp x)
3853       return DAG.getNode(Opc, SL, VT, CvtSrc.getOperand(0));
3854     }
3855 
3856     if (!N0.hasOneUse())
3857       return SDValue();
3858 
3859     // (fneg (fp_extend x)) -> (fp_extend (fneg x))
3860     // (fneg (rcp x)) -> (rcp (fneg x))
3861     SDValue Neg = DAG.getNode(ISD::FNEG, SL, CvtSrc.getValueType(), CvtSrc);
3862     return DAG.getNode(Opc, SL, VT, Neg, N0->getFlags());
3863   }
3864   case ISD::FP_ROUND: {
3865     SDValue CvtSrc = N0.getOperand(0);
3866 
3867     if (CvtSrc.getOpcode() == ISD::FNEG) {
3868       // (fneg (fp_round (fneg x))) -> (fp_round x)
3869       return DAG.getNode(ISD::FP_ROUND, SL, VT,
3870                          CvtSrc.getOperand(0), N0.getOperand(1));
3871     }
3872 
3873     if (!N0.hasOneUse())
3874       return SDValue();
3875 
3876     // (fneg (fp_round x)) -> (fp_round (fneg x))
3877     SDValue Neg = DAG.getNode(ISD::FNEG, SL, CvtSrc.getValueType(), CvtSrc);
3878     return DAG.getNode(ISD::FP_ROUND, SL, VT, Neg, N0.getOperand(1));
3879   }
3880   case ISD::FP16_TO_FP: {
3881     // v_cvt_f32_f16 supports source modifiers on pre-VI targets without legal
3882     // f16, but legalization of f16 fneg ends up pulling it out of the source.
3883     // Put the fneg back as a legal source operation that can be matched later.
3884     SDLoc SL(N);
3885 
3886     SDValue Src = N0.getOperand(0);
3887     EVT SrcVT = Src.getValueType();
3888 
3889     // fneg (fp16_to_fp x) -> fp16_to_fp (xor x, 0x8000)
3890     SDValue IntFNeg = DAG.getNode(ISD::XOR, SL, SrcVT, Src,
3891                                   DAG.getConstant(0x8000, SL, SrcVT));
3892     return DAG.getNode(ISD::FP16_TO_FP, SL, N->getValueType(0), IntFNeg);
3893   }
3894   default:
3895     return SDValue();
3896   }
3897 }
3898 
3899 SDValue AMDGPUTargetLowering::performFAbsCombine(SDNode *N,
3900                                                  DAGCombinerInfo &DCI) const {
3901   SelectionDAG &DAG = DCI.DAG;
3902   SDValue N0 = N->getOperand(0);
3903 
3904   if (!N0.hasOneUse())
3905     return SDValue();
3906 
3907   switch (N0.getOpcode()) {
3908   case ISD::FP16_TO_FP: {
3909     assert(!Subtarget->has16BitInsts() && "should only see if f16 is illegal");
3910     SDLoc SL(N);
3911     SDValue Src = N0.getOperand(0);
3912     EVT SrcVT = Src.getValueType();
3913 
3914     // fabs (fp16_to_fp x) -> fp16_to_fp (and x, 0x7fff)
3915     SDValue IntFAbs = DAG.getNode(ISD::AND, SL, SrcVT, Src,
3916                                   DAG.getConstant(0x7fff, SL, SrcVT));
3917     return DAG.getNode(ISD::FP16_TO_FP, SL, N->getValueType(0), IntFAbs);
3918   }
3919   default:
3920     return SDValue();
3921   }
3922 }
3923 
3924 SDValue AMDGPUTargetLowering::performRcpCombine(SDNode *N,
3925                                                 DAGCombinerInfo &DCI) const {
3926   const auto *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0));
3927   if (!CFP)
3928     return SDValue();
3929 
3930   // XXX - Should this flush denormals?
3931   const APFloat &Val = CFP->getValueAPF();
3932   APFloat One(Val.getSemantics(), "1.0");
3933   return DCI.DAG.getConstantFP(One / Val, SDLoc(N), N->getValueType(0));
3934 }
3935 
3936 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
3937                                                 DAGCombinerInfo &DCI) const {
3938   SelectionDAG &DAG = DCI.DAG;
3939   SDLoc DL(N);
3940 
3941   switch(N->getOpcode()) {
3942   default:
3943     break;
3944   case ISD::BITCAST: {
3945     EVT DestVT = N->getValueType(0);
3946 
3947     // Push casts through vector builds. This helps avoid emitting a large
3948     // number of copies when materializing floating point vector constants.
3949     //
3950     // vNt1 bitcast (vNt0 (build_vector t0:x, t0:y)) =>
3951     //   vnt1 = build_vector (t1 (bitcast t0:x)), (t1 (bitcast t0:y))
3952     if (DestVT.isVector()) {
3953       SDValue Src = N->getOperand(0);
3954       if (Src.getOpcode() == ISD::BUILD_VECTOR) {
3955         EVT SrcVT = Src.getValueType();
3956         unsigned NElts = DestVT.getVectorNumElements();
3957 
3958         if (SrcVT.getVectorNumElements() == NElts) {
3959           EVT DestEltVT = DestVT.getVectorElementType();
3960 
3961           SmallVector<SDValue, 8> CastedElts;
3962           SDLoc SL(N);
3963           for (unsigned I = 0, E = SrcVT.getVectorNumElements(); I != E; ++I) {
3964             SDValue Elt = Src.getOperand(I);
3965             CastedElts.push_back(DAG.getNode(ISD::BITCAST, DL, DestEltVT, Elt));
3966           }
3967 
3968           return DAG.getBuildVector(DestVT, SL, CastedElts);
3969         }
3970       }
3971     }
3972 
3973     if (DestVT.getSizeInBits() != 64 || !DestVT.isVector())
3974       break;
3975 
3976     // Fold bitcasts of constants.
3977     //
3978     // v2i32 (bitcast i64:k) -> build_vector lo_32(k), hi_32(k)
3979     // TODO: Generalize and move to DAGCombiner
3980     SDValue Src = N->getOperand(0);
3981     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Src)) {
3982       SDLoc SL(N);
3983       uint64_t CVal = C->getZExtValue();
3984       SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
3985                                DAG.getConstant(Lo_32(CVal), SL, MVT::i32),
3986                                DAG.getConstant(Hi_32(CVal), SL, MVT::i32));
3987       return DAG.getNode(ISD::BITCAST, SL, DestVT, BV);
3988     }
3989 
3990     if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Src)) {
3991       const APInt &Val = C->getValueAPF().bitcastToAPInt();
3992       SDLoc SL(N);
3993       uint64_t CVal = Val.getZExtValue();
3994       SDValue Vec = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
3995                                 DAG.getConstant(Lo_32(CVal), SL, MVT::i32),
3996                                 DAG.getConstant(Hi_32(CVal), SL, MVT::i32));
3997 
3998       return DAG.getNode(ISD::BITCAST, SL, DestVT, Vec);
3999     }
4000 
4001     break;
4002   }
4003   case ISD::SHL: {
4004     if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
4005       break;
4006 
4007     return performShlCombine(N, DCI);
4008   }
4009   case ISD::SRL: {
4010     if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
4011       break;
4012 
4013     return performSrlCombine(N, DCI);
4014   }
4015   case ISD::SRA: {
4016     if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
4017       break;
4018 
4019     return performSraCombine(N, DCI);
4020   }
4021   case ISD::TRUNCATE:
4022     return performTruncateCombine(N, DCI);
4023   case ISD::MUL:
4024     return performMulCombine(N, DCI);
4025   case ISD::SMUL_LOHI:
4026   case ISD::UMUL_LOHI:
4027     return performMulLoHiCombine(N, DCI);
4028   case ISD::MULHS:
4029     return performMulhsCombine(N, DCI);
4030   case ISD::MULHU:
4031     return performMulhuCombine(N, DCI);
4032   case AMDGPUISD::MUL_I24:
4033   case AMDGPUISD::MUL_U24:
4034   case AMDGPUISD::MULHI_I24:
4035   case AMDGPUISD::MULHI_U24:
4036     return simplifyMul24(N, DCI);
4037   case ISD::SELECT:
4038     return performSelectCombine(N, DCI);
4039   case ISD::FNEG:
4040     return performFNegCombine(N, DCI);
4041   case ISD::FABS:
4042     return performFAbsCombine(N, DCI);
4043   case AMDGPUISD::BFE_I32:
4044   case AMDGPUISD::BFE_U32: {
4045     assert(!N->getValueType(0).isVector() &&
4046            "Vector handling of BFE not implemented");
4047     ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
4048     if (!Width)
4049       break;
4050 
4051     uint32_t WidthVal = Width->getZExtValue() & 0x1f;
4052     if (WidthVal == 0)
4053       return DAG.getConstant(0, DL, MVT::i32);
4054 
4055     ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
4056     if (!Offset)
4057       break;
4058 
4059     SDValue BitsFrom = N->getOperand(0);
4060     uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
4061 
4062     bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
4063 
4064     if (OffsetVal == 0) {
4065       // This is already sign / zero extended, so try to fold away extra BFEs.
4066       unsigned SignBits =  Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
4067 
4068       unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
4069       if (OpSignBits >= SignBits)
4070         return BitsFrom;
4071 
4072       EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
4073       if (Signed) {
4074         // This is a sign_extend_inreg. Replace it to take advantage of existing
4075         // DAG Combines. If not eliminated, we will match back to BFE during
4076         // selection.
4077 
4078         // TODO: The sext_inreg of extended types ends, although we can could
4079         // handle them in a single BFE.
4080         return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
4081                            DAG.getValueType(SmallVT));
4082       }
4083 
4084       return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
4085     }
4086 
4087     if (ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(BitsFrom)) {
4088       if (Signed) {
4089         return constantFoldBFE<int32_t>(DAG,
4090                                         CVal->getSExtValue(),
4091                                         OffsetVal,
4092                                         WidthVal,
4093                                         DL);
4094       }
4095 
4096       return constantFoldBFE<uint32_t>(DAG,
4097                                        CVal->getZExtValue(),
4098                                        OffsetVal,
4099                                        WidthVal,
4100                                        DL);
4101     }
4102 
4103     if ((OffsetVal + WidthVal) >= 32 &&
4104         !(Subtarget->hasSDWA() && OffsetVal == 16 && WidthVal == 16)) {
4105       SDValue ShiftVal = DAG.getConstant(OffsetVal, DL, MVT::i32);
4106       return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
4107                          BitsFrom, ShiftVal);
4108     }
4109 
4110     if (BitsFrom.hasOneUse()) {
4111       APInt Demanded = APInt::getBitsSet(32,
4112                                          OffsetVal,
4113                                          OffsetVal + WidthVal);
4114 
4115       KnownBits Known;
4116       TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
4117                                             !DCI.isBeforeLegalizeOps());
4118       const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4119       if (TLI.ShrinkDemandedConstant(BitsFrom, Demanded, TLO) ||
4120           TLI.SimplifyDemandedBits(BitsFrom, Demanded, Known, TLO)) {
4121         DCI.CommitTargetLoweringOpt(TLO);
4122       }
4123     }
4124 
4125     break;
4126   }
4127   case ISD::LOAD:
4128     return performLoadCombine(N, DCI);
4129   case ISD::STORE:
4130     return performStoreCombine(N, DCI);
4131   case AMDGPUISD::RCP:
4132   case AMDGPUISD::RCP_IFLAG:
4133     return performRcpCombine(N, DCI);
4134   case ISD::AssertZext:
4135   case ISD::AssertSext:
4136     return performAssertSZExtCombine(N, DCI);
4137   case ISD::INTRINSIC_WO_CHAIN:
4138     return performIntrinsicWOChainCombine(N, DCI);
4139   }
4140   return SDValue();
4141 }
4142 
4143 //===----------------------------------------------------------------------===//
4144 // Helper functions
4145 //===----------------------------------------------------------------------===//
4146 
4147 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
4148                                                    const TargetRegisterClass *RC,
4149                                                    Register Reg, EVT VT,
4150                                                    const SDLoc &SL,
4151                                                    bool RawReg) const {
4152   MachineFunction &MF = DAG.getMachineFunction();
4153   MachineRegisterInfo &MRI = MF.getRegInfo();
4154   Register VReg;
4155 
4156   if (!MRI.isLiveIn(Reg)) {
4157     VReg = MRI.createVirtualRegister(RC);
4158     MRI.addLiveIn(Reg, VReg);
4159   } else {
4160     VReg = MRI.getLiveInVirtReg(Reg);
4161   }
4162 
4163   if (RawReg)
4164     return DAG.getRegister(VReg, VT);
4165 
4166   return DAG.getCopyFromReg(DAG.getEntryNode(), SL, VReg, VT);
4167 }
4168 
4169 // This may be called multiple times, and nothing prevents creating multiple
4170 // objects at the same offset. See if we already defined this object.
4171 static int getOrCreateFixedStackObject(MachineFrameInfo &MFI, unsigned Size,
4172                                        int64_t Offset) {
4173   for (int I = MFI.getObjectIndexBegin(); I < 0; ++I) {
4174     if (MFI.getObjectOffset(I) == Offset) {
4175       assert(MFI.getObjectSize(I) == Size);
4176       return I;
4177     }
4178   }
4179 
4180   return MFI.CreateFixedObject(Size, Offset, true);
4181 }
4182 
4183 SDValue AMDGPUTargetLowering::loadStackInputValue(SelectionDAG &DAG,
4184                                                   EVT VT,
4185                                                   const SDLoc &SL,
4186                                                   int64_t Offset) const {
4187   MachineFunction &MF = DAG.getMachineFunction();
4188   MachineFrameInfo &MFI = MF.getFrameInfo();
4189   int FI = getOrCreateFixedStackObject(MFI, VT.getStoreSize(), Offset);
4190 
4191   auto SrcPtrInfo = MachinePointerInfo::getStack(MF, Offset);
4192   SDValue Ptr = DAG.getFrameIndex(FI, MVT::i32);
4193 
4194   return DAG.getLoad(VT, SL, DAG.getEntryNode(), Ptr, SrcPtrInfo, Align(4),
4195                      MachineMemOperand::MODereferenceable |
4196                          MachineMemOperand::MOInvariant);
4197 }
4198 
4199 SDValue AMDGPUTargetLowering::storeStackInputValue(SelectionDAG &DAG,
4200                                                    const SDLoc &SL,
4201                                                    SDValue Chain,
4202                                                    SDValue ArgVal,
4203                                                    int64_t Offset) const {
4204   MachineFunction &MF = DAG.getMachineFunction();
4205   MachinePointerInfo DstInfo = MachinePointerInfo::getStack(MF, Offset);
4206   const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
4207 
4208   SDValue Ptr = DAG.getConstant(Offset, SL, MVT::i32);
4209   // Stores to the argument stack area are relative to the stack pointer.
4210   SDValue SP =
4211       DAG.getCopyFromReg(Chain, SL, Info->getStackPtrOffsetReg(), MVT::i32);
4212   Ptr = DAG.getNode(ISD::ADD, SL, MVT::i32, SP, Ptr);
4213   SDValue Store = DAG.getStore(Chain, SL, ArgVal, Ptr, DstInfo, Align(4),
4214                                MachineMemOperand::MODereferenceable);
4215   return Store;
4216 }
4217 
4218 SDValue AMDGPUTargetLowering::loadInputValue(SelectionDAG &DAG,
4219                                              const TargetRegisterClass *RC,
4220                                              EVT VT, const SDLoc &SL,
4221                                              const ArgDescriptor &Arg) const {
4222   assert(Arg && "Attempting to load missing argument");
4223 
4224   SDValue V = Arg.isRegister() ?
4225     CreateLiveInRegister(DAG, RC, Arg.getRegister(), VT, SL) :
4226     loadStackInputValue(DAG, VT, SL, Arg.getStackOffset());
4227 
4228   if (!Arg.isMasked())
4229     return V;
4230 
4231   unsigned Mask = Arg.getMask();
4232   unsigned Shift = countTrailingZeros<unsigned>(Mask);
4233   V = DAG.getNode(ISD::SRL, SL, VT, V,
4234                   DAG.getShiftAmountConstant(Shift, VT, SL));
4235   return DAG.getNode(ISD::AND, SL, VT, V,
4236                      DAG.getConstant(Mask >> Shift, SL, VT));
4237 }
4238 
4239 uint32_t AMDGPUTargetLowering::getImplicitParameterOffset(
4240     const MachineFunction &MF, const ImplicitParameter Param) const {
4241   const AMDGPUMachineFunction *MFI = MF.getInfo<AMDGPUMachineFunction>();
4242   const AMDGPUSubtarget &ST =
4243       AMDGPUSubtarget::get(getTargetMachine(), MF.getFunction());
4244   unsigned ExplicitArgOffset = ST.getExplicitKernelArgOffset(MF.getFunction());
4245   const Align Alignment = ST.getAlignmentForImplicitArgPtr();
4246   uint64_t ArgOffset = alignTo(MFI->getExplicitKernArgSize(), Alignment) +
4247                        ExplicitArgOffset;
4248   switch (Param) {
4249   case FIRST_IMPLICIT:
4250     return ArgOffset;
4251   case PRIVATE_BASE:
4252     return ArgOffset + AMDGPU::ImplicitArg::PRIVATE_BASE_OFFSET;
4253   case SHARED_BASE:
4254     return ArgOffset + AMDGPU::ImplicitArg::SHARED_BASE_OFFSET;
4255   case QUEUE_PTR:
4256     return ArgOffset + AMDGPU::ImplicitArg::QUEUE_PTR_OFFSET;
4257   }
4258   llvm_unreachable("unexpected implicit parameter type");
4259 }
4260 
4261 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
4262 
4263 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
4264   switch ((AMDGPUISD::NodeType)Opcode) {
4265   case AMDGPUISD::FIRST_NUMBER: break;
4266   // AMDIL DAG nodes
4267   NODE_NAME_CASE(UMUL);
4268   NODE_NAME_CASE(BRANCH_COND);
4269 
4270   // AMDGPU DAG nodes
4271   NODE_NAME_CASE(IF)
4272   NODE_NAME_CASE(ELSE)
4273   NODE_NAME_CASE(LOOP)
4274   NODE_NAME_CASE(CALL)
4275   NODE_NAME_CASE(TC_RETURN)
4276   NODE_NAME_CASE(TRAP)
4277   NODE_NAME_CASE(RET_FLAG)
4278   NODE_NAME_CASE(RETURN_TO_EPILOG)
4279   NODE_NAME_CASE(ENDPGM)
4280   NODE_NAME_CASE(DWORDADDR)
4281   NODE_NAME_CASE(FRACT)
4282   NODE_NAME_CASE(SETCC)
4283   NODE_NAME_CASE(SETREG)
4284   NODE_NAME_CASE(DENORM_MODE)
4285   NODE_NAME_CASE(FMA_W_CHAIN)
4286   NODE_NAME_CASE(FMUL_W_CHAIN)
4287   NODE_NAME_CASE(CLAMP)
4288   NODE_NAME_CASE(COS_HW)
4289   NODE_NAME_CASE(SIN_HW)
4290   NODE_NAME_CASE(FMAX_LEGACY)
4291   NODE_NAME_CASE(FMIN_LEGACY)
4292   NODE_NAME_CASE(FMAX3)
4293   NODE_NAME_CASE(SMAX3)
4294   NODE_NAME_CASE(UMAX3)
4295   NODE_NAME_CASE(FMIN3)
4296   NODE_NAME_CASE(SMIN3)
4297   NODE_NAME_CASE(UMIN3)
4298   NODE_NAME_CASE(FMED3)
4299   NODE_NAME_CASE(SMED3)
4300   NODE_NAME_CASE(UMED3)
4301   NODE_NAME_CASE(FDOT2)
4302   NODE_NAME_CASE(URECIP)
4303   NODE_NAME_CASE(DIV_SCALE)
4304   NODE_NAME_CASE(DIV_FMAS)
4305   NODE_NAME_CASE(DIV_FIXUP)
4306   NODE_NAME_CASE(FMAD_FTZ)
4307   NODE_NAME_CASE(RCP)
4308   NODE_NAME_CASE(RSQ)
4309   NODE_NAME_CASE(RCP_LEGACY)
4310   NODE_NAME_CASE(RCP_IFLAG)
4311   NODE_NAME_CASE(FMUL_LEGACY)
4312   NODE_NAME_CASE(RSQ_CLAMP)
4313   NODE_NAME_CASE(LDEXP)
4314   NODE_NAME_CASE(FP_CLASS)
4315   NODE_NAME_CASE(DOT4)
4316   NODE_NAME_CASE(CARRY)
4317   NODE_NAME_CASE(BORROW)
4318   NODE_NAME_CASE(BFE_U32)
4319   NODE_NAME_CASE(BFE_I32)
4320   NODE_NAME_CASE(BFI)
4321   NODE_NAME_CASE(BFM)
4322   NODE_NAME_CASE(FFBH_U32)
4323   NODE_NAME_CASE(FFBH_I32)
4324   NODE_NAME_CASE(FFBL_B32)
4325   NODE_NAME_CASE(MUL_U24)
4326   NODE_NAME_CASE(MUL_I24)
4327   NODE_NAME_CASE(MULHI_U24)
4328   NODE_NAME_CASE(MULHI_I24)
4329   NODE_NAME_CASE(MAD_U24)
4330   NODE_NAME_CASE(MAD_I24)
4331   NODE_NAME_CASE(MAD_I64_I32)
4332   NODE_NAME_CASE(MAD_U64_U32)
4333   NODE_NAME_CASE(PERM)
4334   NODE_NAME_CASE(TEXTURE_FETCH)
4335   NODE_NAME_CASE(R600_EXPORT)
4336   NODE_NAME_CASE(CONST_ADDRESS)
4337   NODE_NAME_CASE(REGISTER_LOAD)
4338   NODE_NAME_CASE(REGISTER_STORE)
4339   NODE_NAME_CASE(SAMPLE)
4340   NODE_NAME_CASE(SAMPLEB)
4341   NODE_NAME_CASE(SAMPLED)
4342   NODE_NAME_CASE(SAMPLEL)
4343   NODE_NAME_CASE(CVT_F32_UBYTE0)
4344   NODE_NAME_CASE(CVT_F32_UBYTE1)
4345   NODE_NAME_CASE(CVT_F32_UBYTE2)
4346   NODE_NAME_CASE(CVT_F32_UBYTE3)
4347   NODE_NAME_CASE(CVT_PKRTZ_F16_F32)
4348   NODE_NAME_CASE(CVT_PKNORM_I16_F32)
4349   NODE_NAME_CASE(CVT_PKNORM_U16_F32)
4350   NODE_NAME_CASE(CVT_PK_I16_I32)
4351   NODE_NAME_CASE(CVT_PK_U16_U32)
4352   NODE_NAME_CASE(FP_TO_FP16)
4353   NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
4354   NODE_NAME_CASE(CONST_DATA_PTR)
4355   NODE_NAME_CASE(PC_ADD_REL_OFFSET)
4356   NODE_NAME_CASE(LDS)
4357   NODE_NAME_CASE(FPTRUNC_ROUND_UPWARD)
4358   NODE_NAME_CASE(FPTRUNC_ROUND_DOWNWARD)
4359   NODE_NAME_CASE(DUMMY_CHAIN)
4360   case AMDGPUISD::FIRST_MEM_OPCODE_NUMBER: break;
4361   NODE_NAME_CASE(LOAD_D16_HI)
4362   NODE_NAME_CASE(LOAD_D16_LO)
4363   NODE_NAME_CASE(LOAD_D16_HI_I8)
4364   NODE_NAME_CASE(LOAD_D16_HI_U8)
4365   NODE_NAME_CASE(LOAD_D16_LO_I8)
4366   NODE_NAME_CASE(LOAD_D16_LO_U8)
4367   NODE_NAME_CASE(STORE_MSKOR)
4368   NODE_NAME_CASE(LOAD_CONSTANT)
4369   NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
4370   NODE_NAME_CASE(TBUFFER_STORE_FORMAT_D16)
4371   NODE_NAME_CASE(TBUFFER_LOAD_FORMAT)
4372   NODE_NAME_CASE(TBUFFER_LOAD_FORMAT_D16)
4373   NODE_NAME_CASE(DS_ORDERED_COUNT)
4374   NODE_NAME_CASE(ATOMIC_CMP_SWAP)
4375   NODE_NAME_CASE(ATOMIC_INC)
4376   NODE_NAME_CASE(ATOMIC_DEC)
4377   NODE_NAME_CASE(ATOMIC_LOAD_FMIN)
4378   NODE_NAME_CASE(ATOMIC_LOAD_FMAX)
4379   NODE_NAME_CASE(BUFFER_LOAD)
4380   NODE_NAME_CASE(BUFFER_LOAD_UBYTE)
4381   NODE_NAME_CASE(BUFFER_LOAD_USHORT)
4382   NODE_NAME_CASE(BUFFER_LOAD_BYTE)
4383   NODE_NAME_CASE(BUFFER_LOAD_SHORT)
4384   NODE_NAME_CASE(BUFFER_LOAD_FORMAT)
4385   NODE_NAME_CASE(BUFFER_LOAD_FORMAT_D16)
4386   NODE_NAME_CASE(SBUFFER_LOAD)
4387   NODE_NAME_CASE(BUFFER_STORE)
4388   NODE_NAME_CASE(BUFFER_STORE_BYTE)
4389   NODE_NAME_CASE(BUFFER_STORE_SHORT)
4390   NODE_NAME_CASE(BUFFER_STORE_FORMAT)
4391   NODE_NAME_CASE(BUFFER_STORE_FORMAT_D16)
4392   NODE_NAME_CASE(BUFFER_ATOMIC_SWAP)
4393   NODE_NAME_CASE(BUFFER_ATOMIC_ADD)
4394   NODE_NAME_CASE(BUFFER_ATOMIC_SUB)
4395   NODE_NAME_CASE(BUFFER_ATOMIC_SMIN)
4396   NODE_NAME_CASE(BUFFER_ATOMIC_UMIN)
4397   NODE_NAME_CASE(BUFFER_ATOMIC_SMAX)
4398   NODE_NAME_CASE(BUFFER_ATOMIC_UMAX)
4399   NODE_NAME_CASE(BUFFER_ATOMIC_AND)
4400   NODE_NAME_CASE(BUFFER_ATOMIC_OR)
4401   NODE_NAME_CASE(BUFFER_ATOMIC_XOR)
4402   NODE_NAME_CASE(BUFFER_ATOMIC_INC)
4403   NODE_NAME_CASE(BUFFER_ATOMIC_DEC)
4404   NODE_NAME_CASE(BUFFER_ATOMIC_CMPSWAP)
4405   NODE_NAME_CASE(BUFFER_ATOMIC_CSUB)
4406   NODE_NAME_CASE(BUFFER_ATOMIC_FADD)
4407   NODE_NAME_CASE(BUFFER_ATOMIC_FMIN)
4408   NODE_NAME_CASE(BUFFER_ATOMIC_FMAX)
4409 
4410   case AMDGPUISD::LAST_AMDGPU_ISD_NUMBER: break;
4411   }
4412   return nullptr;
4413 }
4414 
4415 SDValue AMDGPUTargetLowering::getSqrtEstimate(SDValue Operand,
4416                                               SelectionDAG &DAG, int Enabled,
4417                                               int &RefinementSteps,
4418                                               bool &UseOneConstNR,
4419                                               bool Reciprocal) const {
4420   EVT VT = Operand.getValueType();
4421 
4422   if (VT == MVT::f32) {
4423     RefinementSteps = 0;
4424     return DAG.getNode(AMDGPUISD::RSQ, SDLoc(Operand), VT, Operand);
4425   }
4426 
4427   // TODO: There is also f64 rsq instruction, but the documentation is less
4428   // clear on its precision.
4429 
4430   return SDValue();
4431 }
4432 
4433 SDValue AMDGPUTargetLowering::getRecipEstimate(SDValue Operand,
4434                                                SelectionDAG &DAG, int Enabled,
4435                                                int &RefinementSteps) const {
4436   EVT VT = Operand.getValueType();
4437 
4438   if (VT == MVT::f32) {
4439     // Reciprocal, < 1 ulp error.
4440     //
4441     // This reciprocal approximation converges to < 0.5 ulp error with one
4442     // newton rhapson performed with two fused multiple adds (FMAs).
4443 
4444     RefinementSteps = 0;
4445     return DAG.getNode(AMDGPUISD::RCP, SDLoc(Operand), VT, Operand);
4446   }
4447 
4448   // TODO: There is also f64 rcp instruction, but the documentation is less
4449   // clear on its precision.
4450 
4451   return SDValue();
4452 }
4453 
4454 static unsigned workitemIntrinsicDim(unsigned ID) {
4455   switch (ID) {
4456   case Intrinsic::amdgcn_workitem_id_x:
4457     return 0;
4458   case Intrinsic::amdgcn_workitem_id_y:
4459     return 1;
4460   case Intrinsic::amdgcn_workitem_id_z:
4461     return 2;
4462   default:
4463     llvm_unreachable("not a workitem intrinsic");
4464   }
4465 }
4466 
4467 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
4468     const SDValue Op, KnownBits &Known,
4469     const APInt &DemandedElts, const SelectionDAG &DAG, unsigned Depth) const {
4470 
4471   Known.resetAll(); // Don't know anything.
4472 
4473   unsigned Opc = Op.getOpcode();
4474 
4475   switch (Opc) {
4476   default:
4477     break;
4478   case AMDGPUISD::CARRY:
4479   case AMDGPUISD::BORROW: {
4480     Known.Zero = APInt::getHighBitsSet(32, 31);
4481     break;
4482   }
4483 
4484   case AMDGPUISD::BFE_I32:
4485   case AMDGPUISD::BFE_U32: {
4486     ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4487     if (!CWidth)
4488       return;
4489 
4490     uint32_t Width = CWidth->getZExtValue() & 0x1f;
4491 
4492     if (Opc == AMDGPUISD::BFE_U32)
4493       Known.Zero = APInt::getHighBitsSet(32, 32 - Width);
4494 
4495     break;
4496   }
4497   case AMDGPUISD::FP_TO_FP16: {
4498     unsigned BitWidth = Known.getBitWidth();
4499 
4500     // High bits are zero.
4501     Known.Zero = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
4502     break;
4503   }
4504   case AMDGPUISD::MUL_U24:
4505   case AMDGPUISD::MUL_I24: {
4506     KnownBits LHSKnown = DAG.computeKnownBits(Op.getOperand(0), Depth + 1);
4507     KnownBits RHSKnown = DAG.computeKnownBits(Op.getOperand(1), Depth + 1);
4508     unsigned TrailZ = LHSKnown.countMinTrailingZeros() +
4509                       RHSKnown.countMinTrailingZeros();
4510     Known.Zero.setLowBits(std::min(TrailZ, 32u));
4511     // Skip extra check if all bits are known zeros.
4512     if (TrailZ >= 32)
4513       break;
4514 
4515     // Truncate to 24 bits.
4516     LHSKnown = LHSKnown.trunc(24);
4517     RHSKnown = RHSKnown.trunc(24);
4518 
4519     if (Opc == AMDGPUISD::MUL_I24) {
4520       unsigned LHSValBits = LHSKnown.countMaxSignificantBits();
4521       unsigned RHSValBits = RHSKnown.countMaxSignificantBits();
4522       unsigned MaxValBits = LHSValBits + RHSValBits;
4523       if (MaxValBits > 32)
4524         break;
4525       unsigned SignBits = 32 - MaxValBits + 1;
4526       bool LHSNegative = LHSKnown.isNegative();
4527       bool LHSNonNegative = LHSKnown.isNonNegative();
4528       bool LHSPositive = LHSKnown.isStrictlyPositive();
4529       bool RHSNegative = RHSKnown.isNegative();
4530       bool RHSNonNegative = RHSKnown.isNonNegative();
4531       bool RHSPositive = RHSKnown.isStrictlyPositive();
4532 
4533       if ((LHSNonNegative && RHSNonNegative) || (LHSNegative && RHSNegative))
4534         Known.Zero.setHighBits(SignBits);
4535       else if ((LHSNegative && RHSPositive) || (LHSPositive && RHSNegative))
4536         Known.One.setHighBits(SignBits);
4537     } else {
4538       unsigned LHSValBits = LHSKnown.countMaxActiveBits();
4539       unsigned RHSValBits = RHSKnown.countMaxActiveBits();
4540       unsigned MaxValBits = LHSValBits + RHSValBits;
4541       if (MaxValBits >= 32)
4542         break;
4543       Known.Zero.setBitsFrom(MaxValBits);
4544     }
4545     break;
4546   }
4547   case AMDGPUISD::PERM: {
4548     ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4549     if (!CMask)
4550       return;
4551 
4552     KnownBits LHSKnown = DAG.computeKnownBits(Op.getOperand(0), Depth + 1);
4553     KnownBits RHSKnown = DAG.computeKnownBits(Op.getOperand(1), Depth + 1);
4554     unsigned Sel = CMask->getZExtValue();
4555 
4556     for (unsigned I = 0; I < 32; I += 8) {
4557       unsigned SelBits = Sel & 0xff;
4558       if (SelBits < 4) {
4559         SelBits *= 8;
4560         Known.One |= ((RHSKnown.One.getZExtValue() >> SelBits) & 0xff) << I;
4561         Known.Zero |= ((RHSKnown.Zero.getZExtValue() >> SelBits) & 0xff) << I;
4562       } else if (SelBits < 7) {
4563         SelBits = (SelBits & 3) * 8;
4564         Known.One |= ((LHSKnown.One.getZExtValue() >> SelBits) & 0xff) << I;
4565         Known.Zero |= ((LHSKnown.Zero.getZExtValue() >> SelBits) & 0xff) << I;
4566       } else if (SelBits == 0x0c) {
4567         Known.Zero |= 0xFFull << I;
4568       } else if (SelBits > 0x0c) {
4569         Known.One |= 0xFFull << I;
4570       }
4571       Sel >>= 8;
4572     }
4573     break;
4574   }
4575   case AMDGPUISD::BUFFER_LOAD_UBYTE:  {
4576     Known.Zero.setHighBits(24);
4577     break;
4578   }
4579   case AMDGPUISD::BUFFER_LOAD_USHORT: {
4580     Known.Zero.setHighBits(16);
4581     break;
4582   }
4583   case AMDGPUISD::LDS: {
4584     auto GA = cast<GlobalAddressSDNode>(Op.getOperand(0).getNode());
4585     Align Alignment = GA->getGlobal()->getPointerAlignment(DAG.getDataLayout());
4586 
4587     Known.Zero.setHighBits(16);
4588     Known.Zero.setLowBits(Log2(Alignment));
4589     break;
4590   }
4591   case ISD::INTRINSIC_WO_CHAIN: {
4592     unsigned IID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4593     switch (IID) {
4594     case Intrinsic::amdgcn_mbcnt_lo:
4595     case Intrinsic::amdgcn_mbcnt_hi: {
4596       const GCNSubtarget &ST =
4597           DAG.getMachineFunction().getSubtarget<GCNSubtarget>();
4598       // These return at most the wavefront size - 1.
4599       unsigned Size = Op.getValueType().getSizeInBits();
4600       Known.Zero.setHighBits(Size - ST.getWavefrontSizeLog2());
4601       break;
4602     }
4603     case Intrinsic::amdgcn_workitem_id_x:
4604     case Intrinsic::amdgcn_workitem_id_y:
4605     case Intrinsic::amdgcn_workitem_id_z: {
4606       unsigned MaxValue = Subtarget->getMaxWorkitemID(
4607           DAG.getMachineFunction().getFunction(), workitemIntrinsicDim(IID));
4608       Known.Zero.setHighBits(countLeadingZeros(MaxValue));
4609       break;
4610     }
4611     default:
4612       break;
4613     }
4614   }
4615   }
4616 }
4617 
4618 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
4619     SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
4620     unsigned Depth) const {
4621   switch (Op.getOpcode()) {
4622   case AMDGPUISD::BFE_I32: {
4623     ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4624     if (!Width)
4625       return 1;
4626 
4627     unsigned SignBits = 32 - Width->getZExtValue() + 1;
4628     if (!isNullConstant(Op.getOperand(1)))
4629       return SignBits;
4630 
4631     // TODO: Could probably figure something out with non-0 offsets.
4632     unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
4633     return std::max(SignBits, Op0SignBits);
4634   }
4635 
4636   case AMDGPUISD::BFE_U32: {
4637     ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4638     return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;
4639   }
4640 
4641   case AMDGPUISD::CARRY:
4642   case AMDGPUISD::BORROW:
4643     return 31;
4644   case AMDGPUISD::BUFFER_LOAD_BYTE:
4645     return 25;
4646   case AMDGPUISD::BUFFER_LOAD_SHORT:
4647     return 17;
4648   case AMDGPUISD::BUFFER_LOAD_UBYTE:
4649     return 24;
4650   case AMDGPUISD::BUFFER_LOAD_USHORT:
4651     return 16;
4652   case AMDGPUISD::FP_TO_FP16:
4653     return 16;
4654   default:
4655     return 1;
4656   }
4657 }
4658 
4659 unsigned AMDGPUTargetLowering::computeNumSignBitsForTargetInstr(
4660   GISelKnownBits &Analysis, Register R,
4661   const APInt &DemandedElts, const MachineRegisterInfo &MRI,
4662   unsigned Depth) const {
4663   const MachineInstr *MI = MRI.getVRegDef(R);
4664   if (!MI)
4665     return 1;
4666 
4667   // TODO: Check range metadata on MMO.
4668   switch (MI->getOpcode()) {
4669   case AMDGPU::G_AMDGPU_BUFFER_LOAD_SBYTE:
4670     return 25;
4671   case AMDGPU::G_AMDGPU_BUFFER_LOAD_SSHORT:
4672     return 17;
4673   case AMDGPU::G_AMDGPU_BUFFER_LOAD_UBYTE:
4674     return 24;
4675   case AMDGPU::G_AMDGPU_BUFFER_LOAD_USHORT:
4676     return 16;
4677   default:
4678     return 1;
4679   }
4680 }
4681 
4682 bool AMDGPUTargetLowering::isKnownNeverNaNForTargetNode(SDValue Op,
4683                                                         const SelectionDAG &DAG,
4684                                                         bool SNaN,
4685                                                         unsigned Depth) const {
4686   unsigned Opcode = Op.getOpcode();
4687   switch (Opcode) {
4688   case AMDGPUISD::FMIN_LEGACY:
4689   case AMDGPUISD::FMAX_LEGACY: {
4690     if (SNaN)
4691       return true;
4692 
4693     // TODO: Can check no nans on one of the operands for each one, but which
4694     // one?
4695     return false;
4696   }
4697   case AMDGPUISD::FMUL_LEGACY:
4698   case AMDGPUISD::CVT_PKRTZ_F16_F32: {
4699     if (SNaN)
4700       return true;
4701     return DAG.isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1) &&
4702            DAG.isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1);
4703   }
4704   case AMDGPUISD::FMED3:
4705   case AMDGPUISD::FMIN3:
4706   case AMDGPUISD::FMAX3:
4707   case AMDGPUISD::FMAD_FTZ: {
4708     if (SNaN)
4709       return true;
4710     return DAG.isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1) &&
4711            DAG.isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1) &&
4712            DAG.isKnownNeverNaN(Op.getOperand(2), SNaN, Depth + 1);
4713   }
4714   case AMDGPUISD::CVT_F32_UBYTE0:
4715   case AMDGPUISD::CVT_F32_UBYTE1:
4716   case AMDGPUISD::CVT_F32_UBYTE2:
4717   case AMDGPUISD::CVT_F32_UBYTE3:
4718     return true;
4719 
4720   case AMDGPUISD::RCP:
4721   case AMDGPUISD::RSQ:
4722   case AMDGPUISD::RCP_LEGACY:
4723   case AMDGPUISD::RSQ_CLAMP: {
4724     if (SNaN)
4725       return true;
4726 
4727     // TODO: Need is known positive check.
4728     return false;
4729   }
4730   case AMDGPUISD::LDEXP:
4731   case AMDGPUISD::FRACT: {
4732     if (SNaN)
4733       return true;
4734     return DAG.isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1);
4735   }
4736   case AMDGPUISD::DIV_SCALE:
4737   case AMDGPUISD::DIV_FMAS:
4738   case AMDGPUISD::DIV_FIXUP:
4739     // TODO: Refine on operands.
4740     return SNaN;
4741   case AMDGPUISD::SIN_HW:
4742   case AMDGPUISD::COS_HW: {
4743     // TODO: Need check for infinity
4744     return SNaN;
4745   }
4746   case ISD::INTRINSIC_WO_CHAIN: {
4747     unsigned IntrinsicID
4748       = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4749     // TODO: Handle more intrinsics
4750     switch (IntrinsicID) {
4751     case Intrinsic::amdgcn_cubeid:
4752       return true;
4753 
4754     case Intrinsic::amdgcn_frexp_mant: {
4755       if (SNaN)
4756         return true;
4757       return DAG.isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1);
4758     }
4759     case Intrinsic::amdgcn_cvt_pkrtz: {
4760       if (SNaN)
4761         return true;
4762       return DAG.isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1) &&
4763              DAG.isKnownNeverNaN(Op.getOperand(2), SNaN, Depth + 1);
4764     }
4765     case Intrinsic::amdgcn_rcp:
4766     case Intrinsic::amdgcn_rsq:
4767     case Intrinsic::amdgcn_rcp_legacy:
4768     case Intrinsic::amdgcn_rsq_legacy:
4769     case Intrinsic::amdgcn_rsq_clamp: {
4770       if (SNaN)
4771         return true;
4772 
4773       // TODO: Need is known positive check.
4774       return false;
4775     }
4776     case Intrinsic::amdgcn_trig_preop:
4777     case Intrinsic::amdgcn_fdot2:
4778       // TODO: Refine on operand
4779       return SNaN;
4780     case Intrinsic::amdgcn_fma_legacy:
4781       if (SNaN)
4782         return true;
4783       return DAG.isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1) &&
4784              DAG.isKnownNeverNaN(Op.getOperand(2), SNaN, Depth + 1) &&
4785              DAG.isKnownNeverNaN(Op.getOperand(3), SNaN, Depth + 1);
4786     default:
4787       return false;
4788     }
4789   }
4790   default:
4791     return false;
4792   }
4793 }
4794 
4795 TargetLowering::AtomicExpansionKind
4796 AMDGPUTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const {
4797   switch (RMW->getOperation()) {
4798   case AtomicRMWInst::Nand:
4799   case AtomicRMWInst::FAdd:
4800   case AtomicRMWInst::FSub:
4801     return AtomicExpansionKind::CmpXChg;
4802   default:
4803     return AtomicExpansionKind::None;
4804   }
4805 }
4806 
4807 bool AMDGPUTargetLowering::isConstantUnsignedBitfieldExtractLegal(
4808     unsigned Opc, LLT Ty1, LLT Ty2) const {
4809   return (Ty1 == LLT::scalar(32) || Ty1 == LLT::scalar(64)) &&
4810          Ty2 == LLT::scalar(32);
4811 }
4812