?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
[InstCombine] Signed saturation patterns
[llvm-core.git] / lib / Target / AMDGPU / AMDGPUISelLowering.cpp
blob1115d8c2362098f0a5bd2eb6ba8086308c664205
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 "AMDGPUCallLowering.h"
18 #include "AMDGPUFrameLowering.h"
19 #include "AMDGPURegisterInfo.h"
20 #include "AMDGPUSubtarget.h"
21 #include "AMDGPUTargetMachine.h"
22 #include "Utils/AMDGPUBaseInfo.h"
23 #include "R600MachineFunctionInfo.h"
24 #include "SIInstrInfo.h"
25 #include "SIMachineFunctionInfo.h"
26 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
27 #include "llvm/CodeGen/Analysis.h"
28 #include "llvm/CodeGen/CallingConvLower.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/CodeGen/SelectionDAG.h"
32 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
33 #include "llvm/IR/DataLayout.h"
34 #include "llvm/IR/DiagnosticInfo.h"
35 #include "llvm/Support/KnownBits.h"
36 #include "llvm/Support/MathExtras.h"
37 using namespace llvm;
38
39 #include "AMDGPUGenCallingConv.inc"
40
41 // Find a larger type to do a load / store of a vector with.
42 EVT AMDGPUTargetLowering::getEquivalentMemType(LLVMContext &Ctx, EVT VT) {
43 unsigned StoreSize = VT.getStoreSizeInBits();
44 if (StoreSize <= 32)
45 return EVT::getIntegerVT(Ctx, StoreSize);
46
47 assert(StoreSize % 32 == 0 && "Store size not a multiple of 32");
48 return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
49 }
50
51 unsigned AMDGPUTargetLowering::numBitsUnsigned(SDValue Op, SelectionDAG &DAG) {
52 EVT VT = Op.getValueType();
53 KnownBits Known = DAG.computeKnownBits(Op);
54 return VT.getSizeInBits() - Known.countMinLeadingZeros();
55 }
56
57 unsigned AMDGPUTargetLowering::numBitsSigned(SDValue Op, SelectionDAG &DAG) {
58 EVT VT = Op.getValueType();
59
60 // In order for this to be a signed 24-bit value, bit 23, must
61 // be a sign bit.
62 return VT.getSizeInBits() - DAG.ComputeNumSignBits(Op);
63 }
64
65 AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
66 const AMDGPUSubtarget &STI)
67 : TargetLowering(TM), Subtarget(&STI) {
68 // Lower floating point store/load to integer store/load to reduce the number
69 // of patterns in tablegen.
70 setOperationAction(ISD::LOAD, MVT::f32, Promote);
71 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
72
73 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
74 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
75
76 setOperationAction(ISD::LOAD, MVT::v3f32, Promote);
77 AddPromotedToType(ISD::LOAD, MVT::v3f32, MVT::v3i32);
78
79 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
80 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
81
82 setOperationAction(ISD::LOAD, MVT::v5f32, Promote);
83 AddPromotedToType(ISD::LOAD, MVT::v5f32, MVT::v5i32);
84
85 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
86 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
87
88 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
89 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
90
91 setOperationAction(ISD::LOAD, MVT::v32f32, Promote);
92 AddPromotedToType(ISD::LOAD, MVT::v32f32, MVT::v32i32);
93
94 setOperationAction(ISD::LOAD, MVT::i64, Promote);
95 AddPromotedToType(ISD::LOAD, MVT::i64, MVT::v2i32);
96
97 setOperationAction(ISD::LOAD, MVT::v2i64, Promote);
98 AddPromotedToType(ISD::LOAD, MVT::v2i64, MVT::v4i32);
99
100 setOperationAction(ISD::LOAD, MVT::f64, Promote);
101 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::v2i32);
102
103 setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
104 AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v4i32);
105
106 // There are no 64-bit extloads. These should be done as a 32-bit extload and
107 // an extension to 64-bit.
108 for (MVT VT : MVT::integer_valuetypes()) {
109 setLoadExtAction(ISD::EXTLOAD, MVT::i64, VT, Expand);
110 setLoadExtAction(ISD::SEXTLOAD, MVT::i64, VT, Expand);
111 setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, VT, Expand);
112 }
113
114 for (MVT VT : MVT::integer_valuetypes()) {
115 if (VT == MVT::i64)
116 continue;
117
118 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
119 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Legal);
120 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i16, Legal);
121 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i32, Expand);
122
123 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
124 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i8, Legal);
125 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i16, Legal);
126 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i32, Expand);
127
128 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
129 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i8, Legal);
130 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i16, Legal);
131 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i32, Expand);
132 }
133
134 for (MVT VT : MVT::integer_fixedlen_vector_valuetypes()) {
135 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Expand);
136 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Expand);
137 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i8, Expand);
138 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Expand);
139 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Expand);
140 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i8, Expand);
141 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Expand);
142 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Expand);
143 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i16, Expand);
144 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v3i16, Expand);
145 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v3i16, Expand);
146 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v3i16, Expand);
147 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Expand);
148 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Expand);
149 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i16, Expand);
150 }
151
152 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
153 setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, MVT::v2f16, Expand);
154 setLoadExtAction(ISD::EXTLOAD, MVT::v3f32, MVT::v3f16, Expand);
155 setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, MVT::v4f16, Expand);
156 setLoadExtAction(ISD::EXTLOAD, MVT::v8f32, MVT::v8f16, Expand);
157 setLoadExtAction(ISD::EXTLOAD, MVT::v16f32, MVT::v16f16, Expand);
158 setLoadExtAction(ISD::EXTLOAD, MVT::v32f32, MVT::v32f16, Expand);
159
160 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
161 setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f32, Expand);
162 setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f32, Expand);
163 setLoadExtAction(ISD::EXTLOAD, MVT::v8f64, MVT::v8f32, Expand);
164
165 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
166 setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f16, Expand);
167 setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f16, Expand);
168 setLoadExtAction(ISD::EXTLOAD, MVT::v8f64, MVT::v8f16, Expand);
169
170 setOperationAction(ISD::STORE, MVT::f32, Promote);
171 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
172
173 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
174 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
175
176 setOperationAction(ISD::STORE, MVT::v3f32, Promote);
177 AddPromotedToType(ISD::STORE, MVT::v3f32, MVT::v3i32);
178
179 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
180 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
181
182 setOperationAction(ISD::STORE, MVT::v5f32, Promote);
183 AddPromotedToType(ISD::STORE, MVT::v5f32, MVT::v5i32);
184
185 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
186 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
187
188 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
189 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
190
191 setOperationAction(ISD::STORE, MVT::v32f32, Promote);
192 AddPromotedToType(ISD::STORE, MVT::v32f32, MVT::v32i32);
193
194 setOperationAction(ISD::STORE, MVT::i64, Promote);
195 AddPromotedToType(ISD::STORE, MVT::i64, MVT::v2i32);
196
197 setOperationAction(ISD::STORE, MVT::v2i64, Promote);
198 AddPromotedToType(ISD::STORE, MVT::v2i64, MVT::v4i32);
199
200 setOperationAction(ISD::STORE, MVT::f64, Promote);
201 AddPromotedToType(ISD::STORE, MVT::f64, MVT::v2i32);
202
203 setOperationAction(ISD::STORE, MVT::v2f64, Promote);
204 AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v4i32);
205
206 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
207 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
208 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
209 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
210
211 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
212 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Expand);
213 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Expand);
214 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Expand);
215
216 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
217 setTruncStoreAction(MVT::v2f32, MVT::v2f16, Expand);
218 setTruncStoreAction(MVT::v3f32, MVT::v3f16, Expand);
219 setTruncStoreAction(MVT::v4f32, MVT::v4f16, Expand);
220 setTruncStoreAction(MVT::v8f32, MVT::v8f16, Expand);
221 setTruncStoreAction(MVT::v16f32, MVT::v16f16, Expand);
222 setTruncStoreAction(MVT::v32f32, MVT::v32f16, Expand);
223
224 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
225 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
226
227 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
228 setTruncStoreAction(MVT::v2f64, MVT::v2f16, Expand);
229
230 setTruncStoreAction(MVT::v4f64, MVT::v4f32, Expand);
231 setTruncStoreAction(MVT::v4f64, MVT::v4f16, Expand);
232
233 setTruncStoreAction(MVT::v8f64, MVT::v8f32, Expand);
234 setTruncStoreAction(MVT::v8f64, MVT::v8f16, Expand);
235
236
237 setOperationAction(ISD::Constant, MVT::i32, Legal);
238 setOperationAction(ISD::Constant, MVT::i64, Legal);
239 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
240 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
241
242 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
243 setOperationAction(ISD::BRIND, MVT::Other, Expand);
244
245 // This is totally unsupported, just custom lower to produce an error.
246 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
247
248 // Library functions. These default to Expand, but we have instructions
249 // for them.
250 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
251 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
252 setOperationAction(ISD::FPOW, MVT::f32, Legal);
253 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
254 setOperationAction(ISD::FABS, MVT::f32, Legal);
255 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
256 setOperationAction(ISD::FRINT, MVT::f32, Legal);
257 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
258 setOperationAction(ISD::FMINNUM, MVT::f32, Legal);
259 setOperationAction(ISD::FMAXNUM, MVT::f32, Legal);
260
261 setOperationAction(ISD::FROUND, MVT::f32, Custom);
262 setOperationAction(ISD::FROUND, MVT::f64, Custom);
263
264 setOperationAction(ISD::FLOG, MVT::f32, Custom);
265 setOperationAction(ISD::FLOG10, MVT::f32, Custom);
266 setOperationAction(ISD::FEXP, MVT::f32, Custom);
267
268
269 setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom);
270 setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
271
272 setOperationAction(ISD::FREM, MVT::f32, Custom);
273 setOperationAction(ISD::FREM, MVT::f64, Custom);
274
275 // Expand to fneg + fadd.
276 setOperationAction(ISD::FSUB, MVT::f64, Expand);
277
278 setOperationAction(ISD::CONCAT_VECTORS, MVT::v3i32, Custom);
279 setOperationAction(ISD::CONCAT_VECTORS, MVT::v3f32, Custom);
280 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
281 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
282 setOperationAction(ISD::CONCAT_VECTORS, MVT::v5i32, Custom);
283 setOperationAction(ISD::CONCAT_VECTORS, MVT::v5f32, Custom);
284 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
285 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
286 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
287 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
288 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v3f32, Custom);
289 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v3i32, Custom);
290 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
291 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
292 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v5f32, Custom);
293 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v5i32, Custom);
294 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
295 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
296 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v16f32, Custom);
297 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v16i32, Custom);
298 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v32f32, Custom);
299 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v32i32, Custom);
300
301 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
302 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Custom);
303 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Custom);
304
305 const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
306 for (MVT VT : ScalarIntVTs) {
307 // These should use [SU]DIVREM, so set them to expand
308 setOperationAction(ISD::SDIV, VT, Expand);
309 setOperationAction(ISD::UDIV, VT, Expand);
310 setOperationAction(ISD::SREM, VT, Expand);
311 setOperationAction(ISD::UREM, VT, Expand);
312
313 // GPU does not have divrem function for signed or unsigned.
314 setOperationAction(ISD::SDIVREM, VT, Custom);
315 setOperationAction(ISD::UDIVREM, VT, Custom);
316
317 // GPU does not have [S|U]MUL_LOHI functions as a single instruction.
318 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
319 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
320
321 setOperationAction(ISD::BSWAP, VT, Expand);
322 setOperationAction(ISD::CTTZ, VT, Expand);
323 setOperationAction(ISD::CTLZ, VT, Expand);
324
325 // AMDGPU uses ADDC/SUBC/ADDE/SUBE
326 setOperationAction(ISD::ADDC, VT, Legal);
327 setOperationAction(ISD::SUBC, VT, Legal);
328 setOperationAction(ISD::ADDE, VT, Legal);
329 setOperationAction(ISD::SUBE, VT, Legal);
330 }
331
332 // The hardware supports 32-bit ROTR, but not ROTL.
333 setOperationAction(ISD::ROTL, MVT::i32, Expand);
334 setOperationAction(ISD::ROTL, MVT::i64, Expand);
335 setOperationAction(ISD::ROTR, MVT::i64, Expand);
336
337 setOperationAction(ISD::MUL, MVT::i64, Expand);
338 setOperationAction(ISD::MULHU, MVT::i64, Expand);
339 setOperationAction(ISD::MULHS, MVT::i64, Expand);
340 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
341 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
342 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
343 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
344 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
345
346 setOperationAction(ISD::SMIN, MVT::i32, Legal);
347 setOperationAction(ISD::UMIN, MVT::i32, Legal);
348 setOperationAction(ISD::SMAX, MVT::i32, Legal);
349 setOperationAction(ISD::UMAX, MVT::i32, Legal);
350
351 setOperationAction(ISD::CTTZ, MVT::i64, Custom);
352 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Custom);
353 setOperationAction(ISD::CTLZ, MVT::i64, Custom);
354 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
355
356 static const MVT::SimpleValueType VectorIntTypes[] = {
357 MVT::v2i32, MVT::v3i32, MVT::v4i32, MVT::v5i32
358 };
359
360 for (MVT VT : VectorIntTypes) {
361 // Expand the following operations for the current type by default.
362 setOperationAction(ISD::ADD, VT, Expand);
363 setOperationAction(ISD::AND, VT, Expand);
364 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
365 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
366 setOperationAction(ISD::MUL, VT, Expand);
367 setOperationAction(ISD::MULHU, VT, Expand);
368 setOperationAction(ISD::MULHS, VT, Expand);
369 setOperationAction(ISD::OR, VT, Expand);
370 setOperationAction(ISD::SHL, VT, Expand);
371 setOperationAction(ISD::SRA, VT, Expand);
372 setOperationAction(ISD::SRL, VT, Expand);
373 setOperationAction(ISD::ROTL, VT, Expand);
374 setOperationAction(ISD::ROTR, VT, Expand);
375 setOperationAction(ISD::SUB, VT, Expand);
376 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
377 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
378 setOperationAction(ISD::SDIV, VT, Expand);
379 setOperationAction(ISD::UDIV, VT, Expand);
380 setOperationAction(ISD::SREM, VT, Expand);
381 setOperationAction(ISD::UREM, VT, Expand);
382 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
383 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
384 setOperationAction(ISD::SDIVREM, VT, Custom);
385 setOperationAction(ISD::UDIVREM, VT, Expand);
386 setOperationAction(ISD::SELECT, VT, Expand);
387 setOperationAction(ISD::VSELECT, VT, Expand);
388 setOperationAction(ISD::SELECT_CC, VT, Expand);
389 setOperationAction(ISD::XOR, VT, Expand);
390 setOperationAction(ISD::BSWAP, VT, Expand);
391 setOperationAction(ISD::CTPOP, VT, Expand);
392 setOperationAction(ISD::CTTZ, VT, Expand);
393 setOperationAction(ISD::CTLZ, VT, Expand);
394 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
395 setOperationAction(ISD::SETCC, VT, Expand);
396 }
397
398 static const MVT::SimpleValueType FloatVectorTypes[] = {
399 MVT::v2f32, MVT::v3f32, MVT::v4f32, MVT::v5f32
400 };
401
402 for (MVT VT : FloatVectorTypes) {
403 setOperationAction(ISD::FABS, VT, Expand);
404 setOperationAction(ISD::FMINNUM, VT, Expand);
405 setOperationAction(ISD::FMAXNUM, VT, Expand);
406 setOperationAction(ISD::FADD, VT, Expand);
407 setOperationAction(ISD::FCEIL, VT, Expand);
408 setOperationAction(ISD::FCOS, VT, Expand);
409 setOperationAction(ISD::FDIV, VT, Expand);
410 setOperationAction(ISD::FEXP2, VT, Expand);
411 setOperationAction(ISD::FEXP, VT, Expand);
412 setOperationAction(ISD::FLOG2, VT, Expand);
413 setOperationAction(ISD::FREM, VT, Expand);
414 setOperationAction(ISD::FLOG, VT, Expand);
415 setOperationAction(ISD::FLOG10, VT, Expand);
416 setOperationAction(ISD::FPOW, VT, Expand);
417 setOperationAction(ISD::FFLOOR, VT, Expand);
418 setOperationAction(ISD::FTRUNC, VT, Expand);
419 setOperationAction(ISD::FMUL, VT, Expand);
420 setOperationAction(ISD::FMA, VT, Expand);
421 setOperationAction(ISD::FRINT, VT, Expand);
422 setOperationAction(ISD::FNEARBYINT, VT, Expand);
423 setOperationAction(ISD::FSQRT, VT, Expand);
424 setOperationAction(ISD::FSIN, VT, Expand);
425 setOperationAction(ISD::FSUB, VT, Expand);
426 setOperationAction(ISD::FNEG, VT, Expand);
427 setOperationAction(ISD::VSELECT, VT, Expand);
428 setOperationAction(ISD::SELECT_CC, VT, Expand);
429 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
430 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
431 setOperationAction(ISD::SETCC, VT, Expand);
432 setOperationAction(ISD::FCANONICALIZE, VT, Expand);
433 }
434
435 // This causes using an unrolled select operation rather than expansion with
436 // bit operations. This is in general better, but the alternative using BFI
437 // instructions may be better if the select sources are SGPRs.
438 setOperationAction(ISD::SELECT, MVT::v2f32, Promote);
439 AddPromotedToType(ISD::SELECT, MVT::v2f32, MVT::v2i32);
440
441 setOperationAction(ISD::SELECT, MVT::v3f32, Promote);
442 AddPromotedToType(ISD::SELECT, MVT::v3f32, MVT::v3i32);
443
444 setOperationAction(ISD::SELECT, MVT::v4f32, Promote);
445 AddPromotedToType(ISD::SELECT, MVT::v4f32, MVT::v4i32);
446
447 setOperationAction(ISD::SELECT, MVT::v5f32, Promote);
448 AddPromotedToType(ISD::SELECT, MVT::v5f32, MVT::v5i32);
449
450 // There are no libcalls of any kind.
451 for (int I = 0; I < RTLIB::UNKNOWN_LIBCALL; ++I)
452 setLibcallName(static_cast<RTLIB::Libcall>(I), nullptr);
453
454 setBooleanContents(ZeroOrNegativeOneBooleanContent);
455 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
456
457 setSchedulingPreference(Sched::RegPressure);
458 setJumpIsExpensive(true);
459
460 // FIXME: This is only partially true. If we have to do vector compares, any
461 // SGPR pair can be a condition register. If we have a uniform condition, we
462 // are better off doing SALU operations, where there is only one SCC. For now,
463 // we don't have a way of knowing during instruction selection if a condition
464 // will be uniform and we always use vector compares. Assume we are using
465 // vector compares until that is fixed.
466 setHasMultipleConditionRegisters(true);
467
468 setMinCmpXchgSizeInBits(32);
469 setSupportsUnalignedAtomics(false);
470
471 PredictableSelectIsExpensive = false;
472
473 // We want to find all load dependencies for long chains of stores to enable
474 // merging into very wide vectors. The problem is with vectors with > 4
475 // elements. MergeConsecutiveStores will attempt to merge these because x8/x16
476 // vectors are a legal type, even though we have to split the loads
477 // usually. When we can more precisely specify load legality per address
478 // space, we should be able to make FindBetterChain/MergeConsecutiveStores
479 // smarter so that they can figure out what to do in 2 iterations without all
480 // N > 4 stores on the same chain.
481 GatherAllAliasesMaxDepth = 16;
482
483 // memcpy/memmove/memset are expanded in the IR, so we shouldn't need to worry
484 // about these during lowering.
485 MaxStoresPerMemcpy = 0xffffffff;
486 MaxStoresPerMemmove = 0xffffffff;
487 MaxStoresPerMemset = 0xffffffff;
488
489 setTargetDAGCombine(ISD::BITCAST);
490 setTargetDAGCombine(ISD::SHL);
491 setTargetDAGCombine(ISD::SRA);
492 setTargetDAGCombine(ISD::SRL);
493 setTargetDAGCombine(ISD::TRUNCATE);
494 setTargetDAGCombine(ISD::MUL);
495 setTargetDAGCombine(ISD::MULHU);
496 setTargetDAGCombine(ISD::MULHS);
497 setTargetDAGCombine(ISD::SELECT);
498 setTargetDAGCombine(ISD::SELECT_CC);
499 setTargetDAGCombine(ISD::STORE);
500 setTargetDAGCombine(ISD::FADD);
501 setTargetDAGCombine(ISD::FSUB);
502 setTargetDAGCombine(ISD::FNEG);
503 setTargetDAGCombine(ISD::FABS);
504 setTargetDAGCombine(ISD::AssertZext);
505 setTargetDAGCombine(ISD::AssertSext);
506 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
507 }
508
509 //===----------------------------------------------------------------------===//
510 // Target Information
511 //===----------------------------------------------------------------------===//
512
513 LLVM_READNONE
514 static bool fnegFoldsIntoOp(unsigned Opc) {
515 switch (Opc) {
516 case ISD::FADD:
517 case ISD::FSUB:
518 case ISD::FMUL:
519 case ISD::FMA:
520 case ISD::FMAD:
521 case ISD::FMINNUM:
522 case ISD::FMAXNUM:
523 case ISD::FMINNUM_IEEE:
524 case ISD::FMAXNUM_IEEE:
525 case ISD::FSIN:
526 case ISD::FTRUNC:
527 case ISD::FRINT:
528 case ISD::FNEARBYINT:
529 case ISD::FCANONICALIZE:
530 case AMDGPUISD::RCP:
531 case AMDGPUISD::RCP_LEGACY:
532 case AMDGPUISD::RCP_IFLAG:
533 case AMDGPUISD::SIN_HW:
534 case AMDGPUISD::FMUL_LEGACY:
535 case AMDGPUISD::FMIN_LEGACY:
536 case AMDGPUISD::FMAX_LEGACY:
537 case AMDGPUISD::FMED3:
538 return true;
539 default:
540 return false;
541 }
542 }
543
544 /// \p returns true if the operation will definitely need to use a 64-bit
545 /// encoding, and thus will use a VOP3 encoding regardless of the source
546 /// modifiers.
547 LLVM_READONLY
548 static bool opMustUseVOP3Encoding(const SDNode *N, MVT VT) {
549 return N->getNumOperands() > 2 || VT == MVT::f64;
550 }
551
552 // Most FP instructions support source modifiers, but this could be refined
553 // slightly.
554 LLVM_READONLY
555 static bool hasSourceMods(const SDNode *N) {
556 if (isa<MemSDNode>(N))
557 return false;
558
559 switch (N->getOpcode()) {
560 case ISD::CopyToReg:
561 case ISD::SELECT:
562 case ISD::FDIV:
563 case ISD::FREM:
564 case ISD::INLINEASM:
565 case ISD::INLINEASM_BR:
566 case AMDGPUISD::DIV_SCALE:
567 case ISD::INTRINSIC_W_CHAIN:
568
569 // TODO: Should really be looking at the users of the bitcast. These are
570 // problematic because bitcasts are used to legalize all stores to integer
571 // types.
572 case ISD::BITCAST:
573 return false;
574 case ISD::INTRINSIC_WO_CHAIN: {
575 switch (cast<ConstantSDNode>(N->getOperand(0))->getZExtValue()) {
576 case Intrinsic::amdgcn_interp_p1:
577 case Intrinsic::amdgcn_interp_p2:
578 case Intrinsic::amdgcn_interp_mov:
579 case Intrinsic::amdgcn_interp_p1_f16:
580 case Intrinsic::amdgcn_interp_p2_f16:
581 return false;
582 default:
583 return true;
584 }
585 }
586 default:
587 return true;
588 }
589 }
590
591 bool AMDGPUTargetLowering::allUsesHaveSourceMods(const SDNode *N,
592 unsigned CostThreshold) {
593 // Some users (such as 3-operand FMA/MAD) must use a VOP3 encoding, and thus
594 // it is truly free to use a source modifier in all cases. If there are
595 // multiple users but for each one will necessitate using VOP3, there will be
596 // a code size increase. Try to avoid increasing code size unless we know it
597 // will save on the instruction count.
598 unsigned NumMayIncreaseSize = 0;
599 MVT VT = N->getValueType(0).getScalarType().getSimpleVT();
600
601 // XXX - Should this limit number of uses to check?
602 for (const SDNode *U : N->uses()) {
603 if (!hasSourceMods(U))
604 return false;
605
606 if (!opMustUseVOP3Encoding(U, VT)) {
607 if (++NumMayIncreaseSize > CostThreshold)
608 return false;
609 }
610 }
611
612 return true;
613 }
614
615 MVT AMDGPUTargetLowering::getVectorIdxTy(const DataLayout &) const {
616 return MVT::i32;
617 }
618
619 bool AMDGPUTargetLowering::isSelectSupported(SelectSupportKind SelType) const {
620 return true;
621 }
622
623 // The backend supports 32 and 64 bit floating point immediates.
624 // FIXME: Why are we reporting vectors of FP immediates as legal?
625 bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT,
626 bool ForCodeSize) const {
627 EVT ScalarVT = VT.getScalarType();
628 return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64 ||
629 (ScalarVT == MVT::f16 && Subtarget->has16BitInsts()));
630 }
631
632 // We don't want to shrink f64 / f32 constants.
633 bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
634 EVT ScalarVT = VT.getScalarType();
635 return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
636 }
637
638 bool AMDGPUTargetLowering::shouldReduceLoadWidth(SDNode *N,
639 ISD::LoadExtType ExtTy,
640 EVT NewVT) const {
641 // TODO: This may be worth removing. Check regression tests for diffs.
642 if (!TargetLoweringBase::shouldReduceLoadWidth(N, ExtTy, NewVT))
643 return false;
644
645 unsigned NewSize = NewVT.getStoreSizeInBits();
646
647 // If we are reducing to a 32-bit load, this is always better.
648 if (NewSize == 32)
649 return true;
650
651 EVT OldVT = N->getValueType(0);
652 unsigned OldSize = OldVT.getStoreSizeInBits();
653
654 MemSDNode *MN = cast<MemSDNode>(N);
655 unsigned AS = MN->getAddressSpace();
656 // Do not shrink an aligned scalar load to sub-dword.
657 // Scalar engine cannot do sub-dword loads.
658 if (OldSize >= 32 && NewSize < 32 && MN->getAlignment() >= 4 &&
659 (AS == AMDGPUAS::CONSTANT_ADDRESS ||
660 AS == AMDGPUAS::CONSTANT_ADDRESS_32BIT ||
661 (isa<LoadSDNode>(N) &&
662 AS == AMDGPUAS::GLOBAL_ADDRESS && MN->isInvariant())) &&
663 AMDGPUInstrInfo::isUniformMMO(MN->getMemOperand()))
664 return false;
665
666 // Don't produce extloads from sub 32-bit types. SI doesn't have scalar
667 // extloads, so doing one requires using a buffer_load. In cases where we
668 // still couldn't use a scalar load, using the wider load shouldn't really
669 // hurt anything.
670
671 // If the old size already had to be an extload, there's no harm in continuing
672 // to reduce the width.
673 return (OldSize < 32);
674 }
675
676 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy, EVT CastTy,
677 const SelectionDAG &DAG,
678 const MachineMemOperand &MMO) const {
679
680 assert(LoadTy.getSizeInBits() == CastTy.getSizeInBits());
681
682 if (LoadTy.getScalarType() == MVT::i32)
683 return false;
684
685 unsigned LScalarSize = LoadTy.getScalarSizeInBits();
686 unsigned CastScalarSize = CastTy.getScalarSizeInBits();
687
688 if ((LScalarSize >= CastScalarSize) && (CastScalarSize < 32))
689 return false;
690
691 bool Fast = false;
692 return allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
693 CastTy, MMO, &Fast) &&
694 Fast;
695 }
696
697 // SI+ has instructions for cttz / ctlz for 32-bit values. This is probably also
698 // profitable with the expansion for 64-bit since it's generally good to
699 // speculate things.
700 // FIXME: These should really have the size as a parameter.
701 bool AMDGPUTargetLowering::isCheapToSpeculateCttz() const {
702 return true;
703 }
704
705 bool AMDGPUTargetLowering::isCheapToSpeculateCtlz() const {
706 return true;
707 }
708
709 bool AMDGPUTargetLowering::isSDNodeAlwaysUniform(const SDNode * N) const {
710 switch (N->getOpcode()) {
711 default:
712 return false;
713 case ISD::EntryToken:
714 case ISD::TokenFactor:
715 return true;
716 case ISD::INTRINSIC_WO_CHAIN:
717 {
718 unsigned IntrID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
719 switch (IntrID) {
720 default:
721 return false;
722 case Intrinsic::amdgcn_readfirstlane:
723 case Intrinsic::amdgcn_readlane:
724 return true;
725 }
726 }
727 break;
728 case ISD::LOAD:
729 {
730 if (cast<LoadSDNode>(N)->getMemOperand()->getAddrSpace() ==
731 AMDGPUAS::CONSTANT_ADDRESS_32BIT)
732 return true;
733 return false;
734 }
735 break;
736 }
737 }
738
739 //===---------------------------------------------------------------------===//
740 // Target Properties
741 //===---------------------------------------------------------------------===//
742
743 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
744 assert(VT.isFloatingPoint());
745
746 // Packed operations do not have a fabs modifier.
747 return VT == MVT::f32 || VT == MVT::f64 ||
748 (Subtarget->has16BitInsts() && VT == MVT::f16);
749 }
750
751 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
752 assert(VT.isFloatingPoint());
753 return VT == MVT::f32 || VT == MVT::f64 ||
754 (Subtarget->has16BitInsts() && VT == MVT::f16) ||
755 (Subtarget->hasVOP3PInsts() && VT == MVT::v2f16);
756 }
757
758 bool AMDGPUTargetLowering:: storeOfVectorConstantIsCheap(EVT MemVT,
759 unsigned NumElem,
760 unsigned AS) const {
761 return true;
762 }
763
764 bool AMDGPUTargetLowering::aggressivelyPreferBuildVectorSources(EVT VecVT) const {
765 // There are few operations which truly have vector input operands. Any vector
766 // operation is going to involve operations on each component, and a
767 // build_vector will be a copy per element, so it always makes sense to use a
768 // build_vector input in place of the extracted element to avoid a copy into a
769 // super register.
770 //
771 // We should probably only do this if all users are extracts only, but this
772 // should be the common case.
773 return true;
774 }
775
776 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
777 // Truncate is just accessing a subregister.
778
779 unsigned SrcSize = Source.getSizeInBits();
780 unsigned DestSize = Dest.getSizeInBits();
781
782 return DestSize < SrcSize && DestSize % 32 == 0 ;
783 }
784
785 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
786 // Truncate is just accessing a subregister.
787
788 unsigned SrcSize = Source->getScalarSizeInBits();
789 unsigned DestSize = Dest->getScalarSizeInBits();
790
791 if (DestSize== 16 && Subtarget->has16BitInsts())
792 return SrcSize >= 32;
793
794 return DestSize < SrcSize && DestSize % 32 == 0;
795 }
796
797 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
798 unsigned SrcSize = Src->getScalarSizeInBits();
799 unsigned DestSize = Dest->getScalarSizeInBits();
800
801 if (SrcSize == 16 && Subtarget->has16BitInsts())
802 return DestSize >= 32;
803
804 return SrcSize == 32 && DestSize == 64;
805 }
806
807 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
808 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
809 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
810 // this will enable reducing 64-bit operations the 32-bit, which is always
811 // good.
812
813 if (Src == MVT::i16)
814 return Dest == MVT::i32 ||Dest == MVT::i64 ;
815
816 return Src == MVT::i32 && Dest == MVT::i64;
817 }
818
819 bool AMDGPUTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
820 return isZExtFree(Val.getValueType(), VT2);
821 }
822
823 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
824 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
825 // limited number of native 64-bit operations. Shrinking an operation to fit
826 // in a single 32-bit register should always be helpful. As currently used,
827 // this is much less general than the name suggests, and is only used in
828 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
829 // not profitable, and may actually be harmful.
830 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
831 }
832
833 //===---------------------------------------------------------------------===//
834 // TargetLowering Callbacks
835 //===---------------------------------------------------------------------===//
836
837 CCAssignFn *AMDGPUCallLowering::CCAssignFnForCall(CallingConv::ID CC,
838 bool IsVarArg) {
839 switch (CC) {
840 case CallingConv::AMDGPU_VS:
841 case CallingConv::AMDGPU_GS:
842 case CallingConv::AMDGPU_PS:
843 case CallingConv::AMDGPU_CS:
844 case CallingConv::AMDGPU_HS:
845 case CallingConv::AMDGPU_ES:
846 case CallingConv::AMDGPU_LS:
847 return CC_AMDGPU;
848 case CallingConv::C:
849 case CallingConv::Fast:
850 case CallingConv::Cold:
851 return CC_AMDGPU_Func;
852 case CallingConv::AMDGPU_KERNEL:
853 case CallingConv::SPIR_KERNEL:
854 default:
855 report_fatal_error("Unsupported calling convention for call");
856 }
857 }
858
859 CCAssignFn *AMDGPUCallLowering::CCAssignFnForReturn(CallingConv::ID CC,
860 bool IsVarArg) {
861 switch (CC) {
862 case CallingConv::AMDGPU_KERNEL:
863 case CallingConv::SPIR_KERNEL:
864 llvm_unreachable("kernels should not be handled here");
865 case CallingConv::AMDGPU_VS:
866 case CallingConv::AMDGPU_GS:
867 case CallingConv::AMDGPU_PS:
868 case CallingConv::AMDGPU_CS:
869 case CallingConv::AMDGPU_HS:
870 case CallingConv::AMDGPU_ES:
871 case CallingConv::AMDGPU_LS:
872 return RetCC_SI_Shader;
873 case CallingConv::C:
874 case CallingConv::Fast:
875 case CallingConv::Cold:
876 return RetCC_AMDGPU_Func;
877 default:
878 report_fatal_error("Unsupported calling convention.");
879 }
880 }
881
882 /// The SelectionDAGBuilder will automatically promote function arguments
883 /// with illegal types. However, this does not work for the AMDGPU targets
884 /// since the function arguments are stored in memory as these illegal types.
885 /// In order to handle this properly we need to get the original types sizes
886 /// from the LLVM IR Function and fixup the ISD:InputArg values before
887 /// passing them to AnalyzeFormalArguments()
888
889 /// When the SelectionDAGBuilder computes the Ins, it takes care of splitting
890 /// input values across multiple registers. Each item in the Ins array
891 /// represents a single value that will be stored in registers. Ins[x].VT is
892 /// the value type of the value that will be stored in the register, so
893 /// whatever SDNode we lower the argument to needs to be this type.
894 ///
895 /// In order to correctly lower the arguments we need to know the size of each
896 /// argument. Since Ins[x].VT gives us the size of the register that will
897 /// hold the value, we need to look at Ins[x].ArgVT to see the 'real' type
898 /// for the orignal function argument so that we can deduce the correct memory
899 /// type to use for Ins[x]. In most cases the correct memory type will be
900 /// Ins[x].ArgVT. However, this will not always be the case. If, for example,
901 /// we have a kernel argument of type v8i8, this argument will be split into
902 /// 8 parts and each part will be represented by its own item in the Ins array.
903 /// For each part the Ins[x].ArgVT will be the v8i8, which is the full type of
904 /// the argument before it was split. From this, we deduce that the memory type
905 /// for each individual part is i8. We pass the memory type as LocVT to the
906 /// calling convention analysis function and the register type (Ins[x].VT) as
907 /// the ValVT.
908 void AMDGPUTargetLowering::analyzeFormalArgumentsCompute(
909 CCState &State,
910 const SmallVectorImpl<ISD::InputArg> &Ins) const {
911 const MachineFunction &MF = State.getMachineFunction();
912 const Function &Fn = MF.getFunction();
913 LLVMContext &Ctx = Fn.getParent()->getContext();
914 const AMDGPUSubtarget &ST = AMDGPUSubtarget::get(MF);
915 const unsigned ExplicitOffset = ST.getExplicitKernelArgOffset(Fn);
916 CallingConv::ID CC = Fn.getCallingConv();
917
918 unsigned MaxAlign = 1;
919 uint64_t ExplicitArgOffset = 0;
920 const DataLayout &DL = Fn.getParent()->getDataLayout();
921
922 unsigned InIndex = 0;
923
924 for (const Argument &Arg : Fn.args()) {
925 Type *BaseArgTy = Arg.getType();
926 unsigned Align = DL.getABITypeAlignment(BaseArgTy);
927 MaxAlign = std::max(Align, MaxAlign);
928 unsigned AllocSize = DL.getTypeAllocSize(BaseArgTy);
929
930 uint64_t ArgOffset = alignTo(ExplicitArgOffset, Align) + ExplicitOffset;
931 ExplicitArgOffset = alignTo(ExplicitArgOffset, Align) + AllocSize;
932
933 // We're basically throwing away everything passed into us and starting over
934 // to get accurate in-memory offsets. The "PartOffset" is completely useless
935 // to us as computed in Ins.
936 //
937 // We also need to figure out what type legalization is trying to do to get
938 // the correct memory offsets.
939
940 SmallVector<EVT, 16> ValueVTs;
941 SmallVector<uint64_t, 16> Offsets;
942 ComputeValueVTs(*this, DL, BaseArgTy, ValueVTs, &Offsets, ArgOffset);
943
944 for (unsigned Value = 0, NumValues = ValueVTs.size();
945 Value != NumValues; ++Value) {
946 uint64_t BasePartOffset = Offsets[Value];
947
948 EVT ArgVT = ValueVTs[Value];
949 EVT MemVT = ArgVT;
950 MVT RegisterVT = getRegisterTypeForCallingConv(Ctx, CC, ArgVT);
951 unsigned NumRegs = getNumRegistersForCallingConv(Ctx, CC, ArgVT);
952
953 if (NumRegs == 1) {
954 // This argument is not split, so the IR type is the memory type.
955 if (ArgVT.isExtended()) {
956 // We have an extended type, like i24, so we should just use the
957 // register type.
958 MemVT = RegisterVT;
959 } else {
960 MemVT = ArgVT;
961 }
962 } else if (ArgVT.isVector() && RegisterVT.isVector() &&
963 ArgVT.getScalarType() == RegisterVT.getScalarType()) {
964 assert(ArgVT.getVectorNumElements() > RegisterVT.getVectorNumElements());
965 // We have a vector value which has been split into a vector with
966 // the same scalar type, but fewer elements. This should handle
967 // all the floating-point vector types.
968 MemVT = RegisterVT;
969 } else if (ArgVT.isVector() &&
970 ArgVT.getVectorNumElements() == NumRegs) {
971 // This arg has been split so that each element is stored in a separate
972 // register.
973 MemVT = ArgVT.getScalarType();
974 } else if (ArgVT.isExtended()) {
975 // We have an extended type, like i65.
976 MemVT = RegisterVT;
977 } else {
978 unsigned MemoryBits = ArgVT.getStoreSizeInBits() / NumRegs;
979 assert(ArgVT.getStoreSizeInBits() % NumRegs == 0);
980 if (RegisterVT.isInteger()) {
981 MemVT = EVT::getIntegerVT(State.getContext(), MemoryBits);
982 } else if (RegisterVT.isVector()) {
983 assert(!RegisterVT.getScalarType().isFloatingPoint());
984 unsigned NumElements = RegisterVT.getVectorNumElements();
985 assert(MemoryBits % NumElements == 0);
986 // This vector type has been split into another vector type with
987 // a different elements size.
988 EVT ScalarVT = EVT::getIntegerVT(State.getContext(),
989 MemoryBits / NumElements);
990 MemVT = EVT::getVectorVT(State.getContext(), ScalarVT, NumElements);
991 } else {
992 llvm_unreachable("cannot deduce memory type.");
993 }
994 }
995
996 // Convert one element vectors to scalar.
997 if (MemVT.isVector() && MemVT.getVectorNumElements() == 1)
998 MemVT = MemVT.getScalarType();
999
1000 // Round up vec3/vec5 argument.
1001 if (MemVT.isVector() && !MemVT.isPow2VectorType()) {
1002 assert(MemVT.getVectorNumElements() == 3 ||
1003 MemVT.getVectorNumElements() == 5);
1004 MemVT = MemVT.getPow2VectorType(State.getContext());
1005 }
1006
1007 unsigned PartOffset = 0;
1008 for (unsigned i = 0; i != NumRegs; ++i) {
1009 State.addLoc(CCValAssign::getCustomMem(InIndex++, RegisterVT,
1010 BasePartOffset + PartOffset,
1011 MemVT.getSimpleVT(),
1012 CCValAssign::Full));
1013 PartOffset += MemVT.getStoreSize();
1014 }
1015 }
1016 }
1017 }
1018
1019 SDValue AMDGPUTargetLowering::LowerReturn(
1020 SDValue Chain, CallingConv::ID CallConv,
1021 bool isVarArg,
1022 const SmallVectorImpl<ISD::OutputArg> &Outs,
1023 const SmallVectorImpl<SDValue> &OutVals,
1024 const SDLoc &DL, SelectionDAG &DAG) const {
1025 // FIXME: Fails for r600 tests
1026 //assert(!isVarArg && Outs.empty() && OutVals.empty() &&
1027 // "wave terminate should not have return values");
1028 return DAG.getNode(AMDGPUISD::ENDPGM, DL, MVT::Other, Chain);
1029 }
1030
1031 //===---------------------------------------------------------------------===//
1032 // Target specific lowering
1033 //===---------------------------------------------------------------------===//
1034
1035 /// Selects the correct CCAssignFn for a given CallingConvention value.
1036 CCAssignFn *AMDGPUTargetLowering::CCAssignFnForCall(CallingConv::ID CC,
1037 bool IsVarArg) {
1038 return AMDGPUCallLowering::CCAssignFnForCall(CC, IsVarArg);
1039 }
1040
1041 CCAssignFn *AMDGPUTargetLowering::CCAssignFnForReturn(CallingConv::ID CC,
1042 bool IsVarArg) {
1043 return AMDGPUCallLowering::CCAssignFnForReturn(CC, IsVarArg);
1044 }
1045
1046 SDValue AMDGPUTargetLowering::addTokenForArgument(SDValue Chain,
1047 SelectionDAG &DAG,
1048 MachineFrameInfo &MFI,
1049 int ClobberedFI) const {
1050 SmallVector<SDValue, 8> ArgChains;
1051 int64_t FirstByte = MFI.getObjectOffset(ClobberedFI);
1052 int64_t LastByte = FirstByte + MFI.getObjectSize(ClobberedFI) - 1;
1053
1054 // Include the original chain at the beginning of the list. When this is
1055 // used by target LowerCall hooks, this helps legalize find the
1056 // CALLSEQ_BEGIN node.
1057 ArgChains.push_back(Chain);
1058
1059 // Add a chain value for each stack argument corresponding
1060 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
1061 UE = DAG.getEntryNode().getNode()->use_end();
1062 U != UE; ++U) {
1063 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U)) {
1064 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr())) {
1065 if (FI->getIndex() < 0) {
1066 int64_t InFirstByte = MFI.getObjectOffset(FI->getIndex());
1067 int64_t InLastByte = InFirstByte;
1068 InLastByte += MFI.getObjectSize(FI->getIndex()) - 1;
1069
1070 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
1071 (FirstByte <= InFirstByte && InFirstByte <= LastByte))
1072 ArgChains.push_back(SDValue(L, 1));
1073 }
1074 }
1075 }
1076 }
1077
1078 // Build a tokenfactor for all the chains.
1079 return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
1080 }
1081
1082 SDValue AMDGPUTargetLowering::lowerUnhandledCall(CallLoweringInfo &CLI,
1083 SmallVectorImpl<SDValue> &InVals,
1084 StringRef Reason) const {
1085 SDValue Callee = CLI.Callee;
1086 SelectionDAG &DAG = CLI.DAG;
1087
1088 const Function &Fn = DAG.getMachineFunction().getFunction();
1089
1090 StringRef FuncName("<unknown>");
1091
1092 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
1093 FuncName = G->getSymbol();
1094 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1095 FuncName = G->getGlobal()->getName();
1096
1097 DiagnosticInfoUnsupported NoCalls(
1098 Fn, Reason + FuncName, CLI.DL.getDebugLoc());
1099 DAG.getContext()->diagnose(NoCalls);
1100
1101 if (!CLI.IsTailCall) {
1102 for (unsigned I = 0, E = CLI.Ins.size(); I != E; ++I)
1103 InVals.push_back(DAG.getUNDEF(CLI.Ins[I].VT));
1104 }
1105
1106 return DAG.getEntryNode();
1107 }
1108
1109 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
1110 SmallVectorImpl<SDValue> &InVals) const {
1111 return lowerUnhandledCall(CLI, InVals, "unsupported call to function ");
1112 }
1113
1114 SDValue AMDGPUTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1115 SelectionDAG &DAG) const {
1116 const Function &Fn = DAG.getMachineFunction().getFunction();
1117
1118 DiagnosticInfoUnsupported NoDynamicAlloca(Fn, "unsupported dynamic alloca",
1119 SDLoc(Op).getDebugLoc());
1120 DAG.getContext()->diagnose(NoDynamicAlloca);
1121 auto Ops = {DAG.getConstant(0, SDLoc(), Op.getValueType()), Op.getOperand(0)};
1122 return DAG.getMergeValues(Ops, SDLoc());
1123 }
1124
1125 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
1126 SelectionDAG &DAG) const {
1127 switch (Op.getOpcode()) {
1128 default:
1129 Op->print(errs(), &DAG);
1130 llvm_unreachable("Custom lowering code for this"
1131 "instruction is not implemented yet!");
1132 break;
1133 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
1134 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
1135 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
1136 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
1137 case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
1138 case ISD::FREM: return LowerFREM(Op, DAG);
1139 case ISD::FCEIL: return LowerFCEIL(Op, DAG);
1140 case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
1141 case ISD::FRINT: return LowerFRINT(Op, DAG);
1142 case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
1143 case ISD::FROUND: return LowerFROUND(Op, DAG);
1144 case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
1145 case ISD::FLOG:
1146 return LowerFLOG(Op, DAG, 1.0F / numbers::log2ef);
1147 case ISD::FLOG10:
1148 return LowerFLOG(Op, DAG, numbers::ln2f / numbers::ln10f);
1149 case ISD::FEXP:
1150 return lowerFEXP(Op, DAG);
1151 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
1152 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
1153 case ISD::FP_TO_FP16: return LowerFP_TO_FP16(Op, DAG);
1154 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
1155 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
1156 case ISD::CTTZ:
1157 case ISD::CTTZ_ZERO_UNDEF:
1158 case ISD::CTLZ:
1159 case ISD::CTLZ_ZERO_UNDEF:
1160 return LowerCTLZ_CTTZ(Op, DAG);
1161 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
1162 }
1163 return Op;
1164 }
1165
1166 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
1167 SmallVectorImpl<SDValue> &Results,
1168 SelectionDAG &DAG) const {
1169 switch (N->getOpcode()) {
1170 case ISD::SIGN_EXTEND_INREG:
1171 // Different parts of legalization seem to interpret which type of
1172 // sign_extend_inreg is the one to check for custom lowering. The extended
1173 // from type is what really matters, but some places check for custom
1174 // lowering of the result type. This results in trying to use
1175 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
1176 // nothing here and let the illegal result integer be handled normally.
1177 return;
1178 default:
1179 return;
1180 }
1181 }
1182
1183 bool AMDGPUTargetLowering::hasDefinedInitializer(const GlobalValue *GV) {
1184 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
1185 if (!GVar || !GVar->hasInitializer())
1186 return false;
1187
1188 return !isa<UndefValue>(GVar->getInitializer());
1189 }
1190
1191 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
1192 SDValue Op,
1193 SelectionDAG &DAG) const {
1194
1195 const DataLayout &DL = DAG.getDataLayout();
1196 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
1197 const GlobalValue *GV = G->getGlobal();
1198
1199 if (G->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1200 G->getAddressSpace() == AMDGPUAS::REGION_ADDRESS) {
1201 if (!MFI->isEntryFunction()) {
1202 const Function &Fn = DAG.getMachineFunction().getFunction();
1203 DiagnosticInfoUnsupported BadLDSDecl(
1204 Fn, "local memory global used by non-kernel function", SDLoc(Op).getDebugLoc());
1205 DAG.getContext()->diagnose(BadLDSDecl);
1206 }
1207
1208 // XXX: What does the value of G->getOffset() mean?
1209 assert(G->getOffset() == 0 &&
1210 "Do not know what to do with an non-zero offset");
1211
1212 // TODO: We could emit code to handle the initialization somewhere.
1213 if (!hasDefinedInitializer(GV)) {
1214 unsigned Offset = MFI->allocateLDSGlobal(DL, *GV);
1215 return DAG.getConstant(Offset, SDLoc(Op), Op.getValueType());
1216 }
1217 }
1218
1219 const Function &Fn = DAG.getMachineFunction().getFunction();
1220 DiagnosticInfoUnsupported BadInit(
1221 Fn, "unsupported initializer for address space", SDLoc(Op).getDebugLoc());
1222 DAG.getContext()->diagnose(BadInit);
1223 return SDValue();
1224 }
1225
1226 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
1227 SelectionDAG &DAG) const {
1228 SmallVector<SDValue, 8> Args;
1229
1230 EVT VT = Op.getValueType();
1231 if (VT == MVT::v4i16 || VT == MVT::v4f16) {
1232 SDLoc SL(Op);
1233 SDValue Lo = DAG.getNode(ISD::BITCAST, SL, MVT::i32, Op.getOperand(0));
1234 SDValue Hi = DAG.getNode(ISD::BITCAST, SL, MVT::i32, Op.getOperand(1));
1235
1236 SDValue BV = DAG.getBuildVector(MVT::v2i32, SL, { Lo, Hi });
1237 return DAG.getNode(ISD::BITCAST, SL, VT, BV);
1238 }
1239
1240 for (const SDUse &U : Op->ops())
1241 DAG.ExtractVectorElements(U.get(), Args);
1242
1243 return DAG.getBuildVector(Op.getValueType(), SDLoc(Op), Args);
1244 }
1245
1246 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
1247 SelectionDAG &DAG) const {
1248
1249 SmallVector<SDValue, 8> Args;
1250 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1251 EVT VT = Op.getValueType();
1252 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
1253 VT.getVectorNumElements());
1254
1255 return DAG.getBuildVector(Op.getValueType(), SDLoc(Op), Args);
1256 }
1257
1258 /// Generate Min/Max node
1259 SDValue AMDGPUTargetLowering::combineFMinMaxLegacy(const SDLoc &DL, EVT VT,
1260 SDValue LHS, SDValue RHS,
1261 SDValue True, SDValue False,
1262 SDValue CC,
1263 DAGCombinerInfo &DCI) const {
1264 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
1265 return SDValue();
1266
1267 SelectionDAG &DAG = DCI.DAG;
1268 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1269 switch (CCOpcode) {
1270 case ISD::SETOEQ:
1271 case ISD::SETONE:
1272 case ISD::SETUNE:
1273 case ISD::SETNE:
1274 case ISD::SETUEQ:
1275 case ISD::SETEQ:
1276 case ISD::SETFALSE:
1277 case ISD::SETFALSE2:
1278 case ISD::SETTRUE:
1279 case ISD::SETTRUE2:
1280 case ISD::SETUO:
1281 case ISD::SETO:
1282 break;
1283 case ISD::SETULE:
1284 case ISD::SETULT: {
1285 if (LHS == True)
1286 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1287 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1288 }
1289 case ISD::SETOLE:
1290 case ISD::SETOLT:
1291 case ISD::SETLE:
1292 case ISD::SETLT: {
1293 // Ordered. Assume ordered for undefined.
1294
1295 // Only do this after legalization to avoid interfering with other combines
1296 // which might occur.
1297 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
1298 !DCI.isCalledByLegalizer())
1299 return SDValue();
1300
1301 // We need to permute the operands to get the correct NaN behavior. The
1302 // selected operand is the second one based on the failing compare with NaN,
1303 // so permute it based on the compare type the hardware uses.
1304 if (LHS == True)
1305 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1306 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1307 }
1308 case ISD::SETUGE:
1309 case ISD::SETUGT: {
1310 if (LHS == True)
1311 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1312 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1313 }
1314 case ISD::SETGT:
1315 case ISD::SETGE:
1316 case ISD::SETOGE:
1317 case ISD::SETOGT: {
1318 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
1319 !DCI.isCalledByLegalizer())
1320 return SDValue();
1321
1322 if (LHS == True)
1323 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1324 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1325 }
1326 case ISD::SETCC_INVALID:
1327 llvm_unreachable("Invalid setcc condcode!");
1328 }
1329 return SDValue();
1330 }
1331
1332 std::pair<SDValue, SDValue>
1333 AMDGPUTargetLowering::split64BitValue(SDValue Op, SelectionDAG &DAG) const {
1334 SDLoc SL(Op);
1335
1336 SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
1337
1338 const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
1339 const SDValue One = DAG.getConstant(1, SL, MVT::i32);
1340
1341 SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero);
1342 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One);
1343
1344 return std::make_pair(Lo, Hi);
1345 }
1346
1347 SDValue AMDGPUTargetLowering::getLoHalf64(SDValue Op, SelectionDAG &DAG) const {
1348 SDLoc SL(Op);
1349
1350 SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
1351 const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
1352 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero);
1353 }
1354
1355 SDValue AMDGPUTargetLowering::getHiHalf64(SDValue Op, SelectionDAG &DAG) const {
1356 SDLoc SL(Op);
1357
1358 SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
1359 const SDValue One = DAG.getConstant(1, SL, MVT::i32);
1360 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One);
1361 }
1362
1363 // Split a vector type into two parts. The first part is a power of two vector.
1364 // The second part is whatever is left over, and is a scalar if it would
1365 // otherwise be a 1-vector.
1366 std::pair<EVT, EVT>
1367 AMDGPUTargetLowering::getSplitDestVTs(const EVT &VT, SelectionDAG &DAG) const {
1368 EVT LoVT, HiVT;
1369 EVT EltVT = VT.getVectorElementType();
1370 unsigned NumElts = VT.getVectorNumElements();
1371 unsigned LoNumElts = PowerOf2Ceil((NumElts + 1) / 2);
1372 LoVT = EVT::getVectorVT(*DAG.getContext(), EltVT, LoNumElts);
1373 HiVT = NumElts - LoNumElts == 1
1374 ? EltVT
1375 : EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts - LoNumElts);
1376 return std::make_pair(LoVT, HiVT);
1377 }
1378
1379 // Split a vector value into two parts of types LoVT and HiVT. HiVT could be
1380 // scalar.
1381 std::pair<SDValue, SDValue>
1382 AMDGPUTargetLowering::splitVector(const SDValue &N, const SDLoc &DL,
1383 const EVT &LoVT, const EVT &HiVT,
1384 SelectionDAG &DAG) const {
1385 assert(LoVT.getVectorNumElements() +
1386 (HiVT.isVector() ? HiVT.getVectorNumElements() : 1) <=
1387 N.getValueType().getVectorNumElements() &&
1388 "More vector elements requested than available!");
1389 auto IdxTy = getVectorIdxTy(DAG.getDataLayout());
1390 SDValue Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, LoVT, N,
1391 DAG.getConstant(0, DL, IdxTy));
1392 SDValue Hi = DAG.getNode(
1393 HiVT.isVector() ? ISD::EXTRACT_SUBVECTOR : ISD::EXTRACT_VECTOR_ELT, DL,
1394 HiVT, N, DAG.getConstant(LoVT.getVectorNumElements(), DL, IdxTy));
1395 return std::make_pair(Lo, Hi);
1396 }
1397
1398 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue Op,
1399 SelectionDAG &DAG) const {
1400 LoadSDNode *Load = cast<LoadSDNode>(Op);
1401 EVT VT = Op.getValueType();
1402
1403
1404 // If this is a 2 element vector, we really want to scalarize and not create
1405 // weird 1 element vectors.
1406 if (VT.getVectorNumElements() == 2)
1407 return scalarizeVectorLoad(Load, DAG);
1408
1409 SDValue BasePtr = Load->getBasePtr();
1410 EVT MemVT = Load->getMemoryVT();
1411 SDLoc SL(Op);
1412
1413 const MachinePointerInfo &SrcValue = Load->getMemOperand()->getPointerInfo();
1414
1415 EVT LoVT, HiVT;
1416 EVT LoMemVT, HiMemVT;
1417 SDValue Lo, Hi;
1418
1419 std::tie(LoVT, HiVT) = getSplitDestVTs(VT, DAG);
1420 std::tie(LoMemVT, HiMemVT) = getSplitDestVTs(MemVT, DAG);
1421 std::tie(Lo, Hi) = splitVector(Op, SL, LoVT, HiVT, DAG);
1422
1423 unsigned Size = LoMemVT.getStoreSize();
1424 unsigned BaseAlign = Load->getAlignment();
1425 unsigned HiAlign = MinAlign(BaseAlign, Size);
1426
1427 SDValue LoLoad = DAG.getExtLoad(Load->getExtensionType(), SL, LoVT,
1428 Load->getChain(), BasePtr, SrcValue, LoMemVT,
1429 BaseAlign, Load->getMemOperand()->getFlags());
1430 SDValue HiPtr = DAG.getObjectPtrOffset(SL, BasePtr, Size);
1431 SDValue HiLoad =
1432 DAG.getExtLoad(Load->getExtensionType(), SL, HiVT, Load->getChain(),
1433 HiPtr, SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1434 HiMemVT, HiAlign, Load->getMemOperand()->getFlags());
1435
1436 auto IdxTy = getVectorIdxTy(DAG.getDataLayout());
1437 SDValue Join;
1438 if (LoVT == HiVT) {
1439 // This is the case that the vector is power of two so was evenly split.
1440 Join = DAG.getNode(ISD::CONCAT_VECTORS, SL, VT, LoLoad, HiLoad);
1441 } else {
1442 Join = DAG.getNode(ISD::INSERT_SUBVECTOR, SL, VT, DAG.getUNDEF(VT), LoLoad,
1443 DAG.getConstant(0, SL, IdxTy));
1444 Join = DAG.getNode(HiVT.isVector() ? ISD::INSERT_SUBVECTOR
1445 : ISD::INSERT_VECTOR_ELT,
1446 SL, VT, Join, HiLoad,
1447 DAG.getConstant(LoVT.getVectorNumElements(), SL, IdxTy));
1448 }
1449
1450 SDValue Ops[] = {Join, DAG.getNode(ISD::TokenFactor, SL, MVT::Other,
1451 LoLoad.getValue(1), HiLoad.getValue(1))};
1452
1453 return DAG.getMergeValues(Ops, SL);
1454 }
1455
1456 // Widen a vector load from vec3 to vec4.
1457 SDValue AMDGPUTargetLowering::WidenVectorLoad(SDValue Op,
1458 SelectionDAG &DAG) const {
1459 LoadSDNode *Load = cast<LoadSDNode>(Op);
1460 EVT VT = Op.getValueType();
1461 assert(VT.getVectorNumElements() == 3);
1462 SDValue BasePtr = Load->getBasePtr();
1463 EVT MemVT = Load->getMemoryVT();
1464 SDLoc SL(Op);
1465 const MachinePointerInfo &SrcValue = Load->getMemOperand()->getPointerInfo();
1466 unsigned BaseAlign = Load->getAlignment();
1467
1468 EVT WideVT =
1469 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), 4);
1470 EVT WideMemVT =
1471 EVT::getVectorVT(*DAG.getContext(), MemVT.getVectorElementType(), 4);
1472 SDValue WideLoad = DAG.getExtLoad(
1473 Load->getExtensionType(), SL, WideVT, Load->getChain(), BasePtr, SrcValue,
1474 WideMemVT, BaseAlign, Load->getMemOperand()->getFlags());
1475 return DAG.getMergeValues(
1476 {DAG.getNode(ISD::EXTRACT_SUBVECTOR, SL, VT, WideLoad,
1477 DAG.getConstant(0, SL, getVectorIdxTy(DAG.getDataLayout()))),
1478 WideLoad.getValue(1)},
1479 SL);
1480 }
1481
1482 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1483 SelectionDAG &DAG) const {
1484 StoreSDNode *Store = cast<StoreSDNode>(Op);
1485 SDValue Val = Store->getValue();
1486 EVT VT = Val.getValueType();
1487
1488 // If this is a 2 element vector, we really want to scalarize and not create
1489 // weird 1 element vectors.
1490 if (VT.getVectorNumElements() == 2)
1491 return scalarizeVectorStore(Store, DAG);
1492
1493 EVT MemVT = Store->getMemoryVT();
1494 SDValue Chain = Store->getChain();
1495 SDValue BasePtr = Store->getBasePtr();
1496 SDLoc SL(Op);
1497
1498 EVT LoVT, HiVT;
1499 EVT LoMemVT, HiMemVT;
1500 SDValue Lo, Hi;
1501
1502 std::tie(LoVT, HiVT) = getSplitDestVTs(VT, DAG);
1503 std::tie(LoMemVT, HiMemVT) = getSplitDestVTs(MemVT, DAG);
1504 std::tie(Lo, Hi) = splitVector(Val, SL, LoVT, HiVT, DAG);
1505
1506 SDValue HiPtr = DAG.getObjectPtrOffset(SL, BasePtr, LoMemVT.getStoreSize());
1507
1508 const MachinePointerInfo &SrcValue = Store->getMemOperand()->getPointerInfo();
1509 unsigned BaseAlign = Store->getAlignment();
1510 unsigned Size = LoMemVT.getStoreSize();
1511 unsigned HiAlign = MinAlign(BaseAlign, Size);
1512
1513 SDValue LoStore =
1514 DAG.getTruncStore(Chain, SL, Lo, BasePtr, SrcValue, LoMemVT, BaseAlign,
1515 Store->getMemOperand()->getFlags());
1516 SDValue HiStore =
1517 DAG.getTruncStore(Chain, SL, Hi, HiPtr, SrcValue.getWithOffset(Size),
1518 HiMemVT, HiAlign, Store->getMemOperand()->getFlags());
1519
1520 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoStore, HiStore);
1521 }
1522
1523 // This is a shortcut for integer division because we have fast i32<->f32
1524 // conversions, and fast f32 reciprocal instructions. The fractional part of a
1525 // float is enough to accurately represent up to a 24-bit signed integer.
1526 SDValue AMDGPUTargetLowering::LowerDIVREM24(SDValue Op, SelectionDAG &DAG,
1527 bool Sign) const {
1528 SDLoc DL(Op);
1529 EVT VT = Op.getValueType();
1530 SDValue LHS = Op.getOperand(0);
1531 SDValue RHS = Op.getOperand(1);
1532 MVT IntVT = MVT::i32;
1533 MVT FltVT = MVT::f32;
1534
1535 unsigned LHSSignBits = DAG.ComputeNumSignBits(LHS);
1536 if (LHSSignBits < 9)
1537 return SDValue();
1538
1539 unsigned RHSSignBits = DAG.ComputeNumSignBits(RHS);
1540 if (RHSSignBits < 9)
1541 return SDValue();
1542
1543 unsigned BitSize = VT.getSizeInBits();
1544 unsigned SignBits = std::min(LHSSignBits, RHSSignBits);
1545 unsigned DivBits = BitSize - SignBits;
1546 if (Sign)
1547 ++DivBits;
1548
1549 ISD::NodeType ToFp = Sign ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
1550 ISD::NodeType ToInt = Sign ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;
1551
1552 SDValue jq = DAG.getConstant(1, DL, IntVT);
1553
1554 if (Sign) {
1555 // char|short jq = ia ^ ib;
1556 jq = DAG.getNode(ISD::XOR, DL, VT, LHS, RHS);
1557
1558 // jq = jq >> (bitsize - 2)
1559 jq = DAG.getNode(ISD::SRA, DL, VT, jq,
1560 DAG.getConstant(BitSize - 2, DL, VT));
1561
1562 // jq = jq | 0x1
1563 jq = DAG.getNode(ISD::OR, DL, VT, jq, DAG.getConstant(1, DL, VT));
1564 }
1565
1566 // int ia = (int)LHS;
1567 SDValue ia = LHS;
1568
1569 // int ib, (int)RHS;
1570 SDValue ib = RHS;
1571
1572 // float fa = (float)ia;
1573 SDValue fa = DAG.getNode(ToFp, DL, FltVT, ia);
1574
1575 // float fb = (float)ib;
1576 SDValue fb = DAG.getNode(ToFp, DL, FltVT, ib);
1577
1578 SDValue fq = DAG.getNode(ISD::FMUL, DL, FltVT,
1579 fa, DAG.getNode(AMDGPUISD::RCP, DL, FltVT, fb));
1580
1581 // fq = trunc(fq);
1582 fq = DAG.getNode(ISD::FTRUNC, DL, FltVT, fq);
1583
1584 // float fqneg = -fq;
1585 SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FltVT, fq);
1586
1587 // float fr = mad(fqneg, fb, fa);
1588 unsigned OpCode = Subtarget->hasFP32Denormals() ?
1589 (unsigned)AMDGPUISD::FMAD_FTZ :
1590 (unsigned)ISD::FMAD;
1591 SDValue fr = DAG.getNode(OpCode, DL, FltVT, fqneg, fb, fa);
1592
1593 // int iq = (int)fq;
1594 SDValue iq = DAG.getNode(ToInt, DL, IntVT, fq);
1595
1596 // fr = fabs(fr);
1597 fr = DAG.getNode(ISD::FABS, DL, FltVT, fr);
1598
1599 // fb = fabs(fb);
1600 fb = DAG.getNode(ISD::FABS, DL, FltVT, fb);
1601
1602 EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
1603
1604 // int cv = fr >= fb;
1605 SDValue cv = DAG.getSetCC(DL, SetCCVT, fr, fb, ISD::SETOGE);
1606
1607 // jq = (cv ? jq : 0);
1608 jq = DAG.getNode(ISD::SELECT, DL, VT, cv, jq, DAG.getConstant(0, DL, VT));
1609
1610 // dst = iq + jq;
1611 SDValue Div = DAG.getNode(ISD::ADD, DL, VT, iq, jq);
1612
1613 // Rem needs compensation, it's easier to recompute it
1614 SDValue Rem = DAG.getNode(ISD::MUL, DL, VT, Div, RHS);
1615 Rem = DAG.getNode(ISD::SUB, DL, VT, LHS, Rem);
1616
1617 // Truncate to number of bits this divide really is.
1618 if (Sign) {
1619 SDValue InRegSize
1620 = DAG.getValueType(EVT::getIntegerVT(*DAG.getContext(), DivBits));
1621 Div = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Div, InRegSize);
1622 Rem = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Rem, InRegSize);
1623 } else {
1624 SDValue TruncMask = DAG.getConstant((UINT64_C(1) << DivBits) - 1, DL, VT);
1625 Div = DAG.getNode(ISD::AND, DL, VT, Div, TruncMask);
1626 Rem = DAG.getNode(ISD::AND, DL, VT, Rem, TruncMask);
1627 }
1628
1629 return DAG.getMergeValues({ Div, Rem }, DL);
1630 }
1631
1632 void AMDGPUTargetLowering::LowerUDIVREM64(SDValue Op,
1633 SelectionDAG &DAG,
1634 SmallVectorImpl<SDValue> &Results) const {
1635 SDLoc DL(Op);
1636 EVT VT = Op.getValueType();
1637
1638 assert(VT == MVT::i64 && "LowerUDIVREM64 expects an i64");
1639
1640 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
1641
1642 SDValue One = DAG.getConstant(1, DL, HalfVT);
1643 SDValue Zero = DAG.getConstant(0, DL, HalfVT);
1644
1645 //HiLo split
1646 SDValue LHS = Op.getOperand(0);
1647 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero);
1648 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, One);
1649
1650 SDValue RHS = Op.getOperand(1);
1651 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero);
1652 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, One);
1653
1654 if (DAG.MaskedValueIsZero(RHS, APInt::getHighBitsSet(64, 32)) &&
1655 DAG.MaskedValueIsZero(LHS, APInt::getHighBitsSet(64, 32))) {
1656
1657 SDValue Res = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
1658 LHS_Lo, RHS_Lo);
1659
1660 SDValue DIV = DAG.getBuildVector(MVT::v2i32, DL, {Res.getValue(0), Zero});
1661 SDValue REM = DAG.getBuildVector(MVT::v2i32, DL, {Res.getValue(1), Zero});
1662
1663 Results.push_back(DAG.getNode(ISD::BITCAST, DL, MVT::i64, DIV));
1664 Results.push_back(DAG.getNode(ISD::BITCAST, DL, MVT::i64, REM));
1665 return;
1666 }
1667
1668 if (isTypeLegal(MVT::i64)) {
1669 // Compute denominator reciprocal.
1670 unsigned FMAD = Subtarget->hasFP32Denormals() ?
1671 (unsigned)AMDGPUISD::FMAD_FTZ :
1672 (unsigned)ISD::FMAD;
1673
1674 SDValue Cvt_Lo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, RHS_Lo);
1675 SDValue Cvt_Hi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, RHS_Hi);
1676 SDValue Mad1 = DAG.getNode(FMAD, DL, MVT::f32, Cvt_Hi,
1677 DAG.getConstantFP(APInt(32, 0x4f800000).bitsToFloat(), DL, MVT::f32),
1678 Cvt_Lo);
1679 SDValue Rcp = DAG.getNode(AMDGPUISD::RCP, DL, MVT::f32, Mad1);
1680 SDValue Mul1 = DAG.getNode(ISD::FMUL, DL, MVT::f32, Rcp,
1681 DAG.getConstantFP(APInt(32, 0x5f7ffffc).bitsToFloat(), DL, MVT::f32));
1682 SDValue Mul2 = DAG.getNode(ISD::FMUL, DL, MVT::f32, Mul1,
1683 DAG.getConstantFP(APInt(32, 0x2f800000).bitsToFloat(), DL, MVT::f32));
1684 SDValue Trunc = DAG.getNode(ISD::FTRUNC, DL, MVT::f32, Mul2);
1685 SDValue Mad2 = DAG.getNode(FMAD, DL, MVT::f32, Trunc,
1686 DAG.getConstantFP(APInt(32, 0xcf800000).bitsToFloat(), DL, MVT::f32),
1687 Mul1);
1688 SDValue Rcp_Lo = DAG.getNode(ISD::FP_TO_UINT, DL, HalfVT, Mad2);
1689 SDValue Rcp_Hi = DAG.getNode(ISD::FP_TO_UINT, DL, HalfVT, Trunc);
1690 SDValue Rcp64 = DAG.getBitcast(VT,
1691 DAG.getBuildVector(MVT::v2i32, DL, {Rcp_Lo, Rcp_Hi}));
1692
1693 SDValue Zero64 = DAG.getConstant(0, DL, VT);
1694 SDValue One64 = DAG.getConstant(1, DL, VT);
1695 SDValue Zero1 = DAG.getConstant(0, DL, MVT::i1);
1696 SDVTList HalfCarryVT = DAG.getVTList(HalfVT, MVT::i1);
1697
1698 SDValue Neg_RHS = DAG.getNode(ISD::SUB, DL, VT, Zero64, RHS);
1699 SDValue Mullo1 = DAG.getNode(ISD::MUL, DL, VT, Neg_RHS, Rcp64);
1700 SDValue Mulhi1 = DAG.getNode(ISD::MULHU, DL, VT, Rcp64, Mullo1);
1701 SDValue Mulhi1_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi1,
1702 Zero);
1703 SDValue Mulhi1_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi1,
1704 One);
1705
1706 SDValue Add1_Lo = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Rcp_Lo,
1707 Mulhi1_Lo, Zero1);
1708 SDValue Add1_Hi = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Rcp_Hi,
1709 Mulhi1_Hi, Add1_Lo.getValue(1));
1710 SDValue Add1_HiNc = DAG.getNode(ISD::ADD, DL, HalfVT, Rcp_Hi, Mulhi1_Hi);
1711 SDValue Add1 = DAG.getBitcast(VT,
1712 DAG.getBuildVector(MVT::v2i32, DL, {Add1_Lo, Add1_Hi}));
1713
1714 SDValue Mullo2 = DAG.getNode(ISD::MUL, DL, VT, Neg_RHS, Add1);
1715 SDValue Mulhi2 = DAG.getNode(ISD::MULHU, DL, VT, Add1, Mullo2);
1716 SDValue Mulhi2_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi2,
1717 Zero);
1718 SDValue Mulhi2_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi2,
1719 One);
1720
1721 SDValue Add2_Lo = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Add1_Lo,
1722 Mulhi2_Lo, Zero1);
1723 SDValue Add2_HiC = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Add1_HiNc,
1724 Mulhi2_Hi, Add1_Lo.getValue(1));
1725 SDValue Add2_Hi = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Add2_HiC,
1726 Zero, Add2_Lo.getValue(1));
1727 SDValue Add2 = DAG.getBitcast(VT,
1728 DAG.getBuildVector(MVT::v2i32, DL, {Add2_Lo, Add2_Hi}));
1729 SDValue Mulhi3 = DAG.getNode(ISD::MULHU, DL, VT, LHS, Add2);
1730
1731 SDValue Mul3 = DAG.getNode(ISD::MUL, DL, VT, RHS, Mulhi3);
1732
1733 SDValue Mul3_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mul3, Zero);
1734 SDValue Mul3_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mul3, One);
1735 SDValue Sub1_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, LHS_Lo,
1736 Mul3_Lo, Zero1);
1737 SDValue Sub1_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, LHS_Hi,
1738 Mul3_Hi, Sub1_Lo.getValue(1));
1739 SDValue Sub1_Mi = DAG.getNode(ISD::SUB, DL, HalfVT, LHS_Hi, Mul3_Hi);
1740 SDValue Sub1 = DAG.getBitcast(VT,
1741 DAG.getBuildVector(MVT::v2i32, DL, {Sub1_Lo, Sub1_Hi}));
1742
1743 SDValue MinusOne = DAG.getConstant(0xffffffffu, DL, HalfVT);
1744 SDValue C1 = DAG.getSelectCC(DL, Sub1_Hi, RHS_Hi, MinusOne, Zero,
1745 ISD::SETUGE);
1746 SDValue C2 = DAG.getSelectCC(DL, Sub1_Lo, RHS_Lo, MinusOne, Zero,
1747 ISD::SETUGE);
1748 SDValue C3 = DAG.getSelectCC(DL, Sub1_Hi, RHS_Hi, C2, C1, ISD::SETEQ);
1749
1750 // TODO: Here and below portions of the code can be enclosed into if/endif.
1751 // Currently control flow is unconditional and we have 4 selects after
1752 // potential endif to substitute PHIs.
1753
1754 // if C3 != 0 ...
1755 SDValue Sub2_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub1_Lo,
1756 RHS_Lo, Zero1);
1757 SDValue Sub2_Mi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub1_Mi,
1758 RHS_Hi, Sub1_Lo.getValue(1));
1759 SDValue Sub2_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Mi,
1760 Zero, Sub2_Lo.getValue(1));
1761 SDValue Sub2 = DAG.getBitcast(VT,
1762 DAG.getBuildVector(MVT::v2i32, DL, {Sub2_Lo, Sub2_Hi}));
1763
1764 SDValue Add3 = DAG.getNode(ISD::ADD, DL, VT, Mulhi3, One64);
1765
1766 SDValue C4 = DAG.getSelectCC(DL, Sub2_Hi, RHS_Hi, MinusOne, Zero,
1767 ISD::SETUGE);
1768 SDValue C5 = DAG.getSelectCC(DL, Sub2_Lo, RHS_Lo, MinusOne, Zero,
1769 ISD::SETUGE);
1770 SDValue C6 = DAG.getSelectCC(DL, Sub2_Hi, RHS_Hi, C5, C4, ISD::SETEQ);
1771
1772 // if (C6 != 0)
1773 SDValue Add4 = DAG.getNode(ISD::ADD, DL, VT, Add3, One64);
1774
1775 SDValue Sub3_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Lo,
1776 RHS_Lo, Zero1);
1777 SDValue Sub3_Mi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Mi,
1778 RHS_Hi, Sub2_Lo.getValue(1));
1779 SDValue Sub3_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub3_Mi,
1780 Zero, Sub3_Lo.getValue(1));
1781 SDValue Sub3 = DAG.getBitcast(VT,
1782 DAG.getBuildVector(MVT::v2i32, DL, {Sub3_Lo, Sub3_Hi}));
1783
1784 // endif C6
1785 // endif C3
1786
1787 SDValue Sel1 = DAG.getSelectCC(DL, C6, Zero, Add4, Add3, ISD::SETNE);
1788 SDValue Div = DAG.getSelectCC(DL, C3, Zero, Sel1, Mulhi3, ISD::SETNE);
1789
1790 SDValue Sel2 = DAG.getSelectCC(DL, C6, Zero, Sub3, Sub2, ISD::SETNE);
1791 SDValue Rem = DAG.getSelectCC(DL, C3, Zero, Sel2, Sub1, ISD::SETNE);
1792
1793 Results.push_back(Div);
1794 Results.push_back(Rem);
1795
1796 return;
1797 }
1798
1799 // r600 expandion.
1800 // Get Speculative values
1801 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
1802 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
1803
1804 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, Zero, REM_Part, LHS_Hi, ISD::SETEQ);
1805 SDValue REM = DAG.getBuildVector(MVT::v2i32, DL, {REM_Lo, Zero});
1806 REM = DAG.getNode(ISD::BITCAST, DL, MVT::i64, REM);
1807
1808 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, Zero, DIV_Part, Zero, ISD::SETEQ);
1809 SDValue DIV_Lo = Zero;
1810
1811 const unsigned halfBitWidth = HalfVT.getSizeInBits();
1812
1813 for (unsigned i = 0; i < halfBitWidth; ++i) {
1814 const unsigned bitPos = halfBitWidth - i - 1;
1815 SDValue POS = DAG.getConstant(bitPos, DL, HalfVT);
1816 // Get value of high bit
1817 SDValue HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
1818 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, One);
1819 HBit = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, HBit);
1820
1821 // Shift
1822 REM = DAG.getNode(ISD::SHL, DL, VT, REM, DAG.getConstant(1, DL, VT));
1823 // Add LHS high bit
1824 REM = DAG.getNode(ISD::OR, DL, VT, REM, HBit);
1825
1826 SDValue BIT = DAG.getConstant(1ULL << bitPos, DL, HalfVT);
1827 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, Zero, ISD::SETUGE);
1828
1829 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
1830
1831 // Update REM
1832 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
1833 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETUGE);
1834 }
1835
1836 SDValue DIV = DAG.getBuildVector(MVT::v2i32, DL, {DIV_Lo, DIV_Hi});
1837 DIV = DAG.getNode(ISD::BITCAST, DL, MVT::i64, DIV);
1838 Results.push_back(DIV);
1839 Results.push_back(REM);
1840 }
1841
1842 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1843 SelectionDAG &DAG) const {
1844 SDLoc DL(Op);
1845 EVT VT = Op.getValueType();
1846
1847 if (VT == MVT::i64) {
1848 SmallVector<SDValue, 2> Results;
1849 LowerUDIVREM64(Op, DAG, Results);
1850 return DAG.getMergeValues(Results, DL);
1851 }
1852
1853 if (VT == MVT::i32) {
1854 if (SDValue Res = LowerDIVREM24(Op, DAG, false))
1855 return Res;
1856 }
1857
1858 SDValue Num = Op.getOperand(0);
1859 SDValue Den = Op.getOperand(1);
1860
1861 // RCP = URECIP(Den) = 2^32 / Den + e
1862 // e is rounding error.
1863 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1864
1865 // RCP_LO = mul(RCP, Den) */
1866 SDValue RCP_LO = DAG.getNode(ISD::MUL, DL, VT, RCP, Den);
1867
1868 // RCP_HI = mulhu (RCP, Den) */
1869 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1870
1871 // NEG_RCP_LO = -RCP_LO
1872 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
1873 RCP_LO);
1874
1875 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1876 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, DL, VT),
1877 NEG_RCP_LO, RCP_LO,
1878 ISD::SETEQ);
1879 // Calculate the rounding error from the URECIP instruction
1880 // E = mulhu(ABS_RCP_LO, RCP)
1881 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1882
1883 // RCP_A_E = RCP + E
1884 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1885
1886 // RCP_S_E = RCP - E
1887 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1888
1889 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1890 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, DL, VT),
1891 RCP_A_E, RCP_S_E,
1892 ISD::SETEQ);
1893 // Quotient = mulhu(Tmp0, Num)
1894 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1895
1896 // Num_S_Remainder = Quotient * Den
1897 SDValue Num_S_Remainder = DAG.getNode(ISD::MUL, DL, VT, Quotient, Den);
1898
1899 // Remainder = Num - Num_S_Remainder
1900 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1901
1902 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1903 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1904 DAG.getConstant(-1, DL, VT),
1905 DAG.getConstant(0, DL, VT),
1906 ISD::SETUGE);
1907 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1908 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1909 Num_S_Remainder,
1910 DAG.getConstant(-1, DL, VT),
1911 DAG.getConstant(0, DL, VT),
1912 ISD::SETUGE);
1913 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1914 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1915 Remainder_GE_Zero);
1916
1917 // Calculate Division result:
1918
1919 // Quotient_A_One = Quotient + 1
1920 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1921 DAG.getConstant(1, DL, VT));
1922
1923 // Quotient_S_One = Quotient - 1
1924 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1925 DAG.getConstant(1, DL, VT));
1926
1927 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1928 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, DL, VT),
1929 Quotient, Quotient_A_One, ISD::SETEQ);
1930
1931 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1932 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, DL, VT),
1933 Quotient_S_One, Div, ISD::SETEQ);
1934
1935 // Calculate Rem result:
1936
1937 // Remainder_S_Den = Remainder - Den
1938 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1939
1940 // Remainder_A_Den = Remainder + Den
1941 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1942
1943 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1944 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, DL, VT),
1945 Remainder, Remainder_S_Den, ISD::SETEQ);
1946
1947 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1948 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, DL, VT),
1949 Remainder_A_Den, Rem, ISD::SETEQ);
1950 SDValue Ops[2] = {
1951 Div,
1952 Rem
1953 };
1954 return DAG.getMergeValues(Ops, DL);
1955 }
1956
1957 SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
1958 SelectionDAG &DAG) const {
1959 SDLoc DL(Op);
1960 EVT VT = Op.getValueType();
1961
1962 SDValue LHS = Op.getOperand(0);
1963 SDValue RHS = Op.getOperand(1);
1964
1965 SDValue Zero = DAG.getConstant(0, DL, VT);
1966 SDValue NegOne = DAG.getConstant(-1, DL, VT);
1967
1968 if (VT == MVT::i32) {
1969 if (SDValue Res = LowerDIVREM24(Op, DAG, true))
1970 return Res;
1971 }
1972
1973 if (VT == MVT::i64 &&
1974 DAG.ComputeNumSignBits(LHS) > 32 &&
1975 DAG.ComputeNumSignBits(RHS) > 32) {
1976 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
1977
1978 //HiLo split
1979 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero);
1980 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero);
1981 SDValue DIVREM = DAG.getNode(ISD::SDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
1982 LHS_Lo, RHS_Lo);
1983 SDValue Res[2] = {
1984 DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(0)),
1985 DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(1))
1986 };
1987 return DAG.getMergeValues(Res, DL);
1988 }
1989
1990 SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
1991 SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
1992 SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
1993 SDValue RSign = LHSign; // Remainder sign is the same as LHS
1994
1995 LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
1996 RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);
1997
1998 LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
1999 RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);
2000
2001 SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
2002 SDValue Rem = Div.getValue(1);
2003
2004 Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
2005 Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);
2006
2007 Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
2008 Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);
2009
2010 SDValue Res[2] = {
2011 Div,
2012 Rem
2013 };
2014 return DAG.getMergeValues(Res, DL);
2015 }
2016
2017 // (frem x, y) -> (fsub x, (fmul (ftrunc (fdiv x, y)), y))
2018 SDValue AMDGPUTargetLowering::LowerFREM(SDValue Op, SelectionDAG &DAG) const {
2019 SDLoc SL(Op);
2020 EVT VT = Op.getValueType();
2021 SDValue X = Op.getOperand(0);
2022 SDValue Y = Op.getOperand(1);
2023
2024 // TODO: Should this propagate fast-math-flags?
2025
2026 SDValue Div = DAG.getNode(ISD::FDIV, SL, VT, X, Y);
2027 SDValue Floor = DAG.getNode(ISD::FTRUNC, SL, VT, Div);
2028 SDValue Mul = DAG.getNode(ISD::FMUL, SL, VT, Floor, Y);
2029
2030 return DAG.getNode(ISD::FSUB, SL, VT, X, Mul);
2031 }
2032
2033 SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
2034 SDLoc SL(Op);
2035 SDValue Src = Op.getOperand(0);
2036
2037 // result = trunc(src)
2038 // if (src > 0.0 && src != result)
2039 // result += 1.0
2040
2041 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2042
2043 const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64);
2044 const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f64);
2045
2046 EVT SetCCVT =
2047 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
2048
2049 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
2050 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
2051 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
2052
2053 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
2054 // TODO: Should this propagate fast-math-flags?
2055 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
2056 }
2057
2058 static SDValue extractF64Exponent(SDValue Hi, const SDLoc &SL,
2059 SelectionDAG &DAG) {
2060 const unsigned FractBits = 52;
2061 const unsigned ExpBits = 11;
2062
2063 SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_U32, SL, MVT::i32,
2064 Hi,
2065 DAG.getConstant(FractBits - 32, SL, MVT::i32),
2066 DAG.getConstant(ExpBits, SL, MVT::i32));
2067 SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
2068 DAG.getConstant(1023, SL, MVT::i32));
2069
2070 return Exp;
2071 }
2072
2073 SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
2074 SDLoc SL(Op);
2075 SDValue Src = Op.getOperand(0);
2076
2077 assert(Op.getValueType() == MVT::f64);
2078
2079 const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
2080 const SDValue One = DAG.getConstant(1, SL, MVT::i32);
2081
2082 SDValue VecSrc = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
2083
2084 // Extract the upper half, since this is where we will find the sign and
2085 // exponent.
2086 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
2087
2088 SDValue Exp = extractF64Exponent(Hi, SL, DAG);
2089
2090 const unsigned FractBits = 52;
2091
2092 // Extract the sign bit.
2093 const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, SL, MVT::i32);
2094 SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);
2095
2096 // Extend back to 64-bits.
2097 SDValue SignBit64 = DAG.getBuildVector(MVT::v2i32, SL, {Zero, SignBit});
2098 SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);
2099
2100 SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
2101 const SDValue FractMask
2102 = DAG.getConstant((UINT64_C(1) << FractBits) - 1, SL, MVT::i64);
2103
2104 SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
2105 SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
2106 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);
2107
2108 EVT SetCCVT =
2109 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i32);
2110
2111 const SDValue FiftyOne = DAG.getConstant(FractBits - 1, SL, MVT::i32);
2112
2113 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
2114 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
2115
2116 SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
2117 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);
2118
2119 return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
2120 }
2121
2122 SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
2123 SDLoc SL(Op);
2124 SDValue Src = Op.getOperand(0);
2125
2126 assert(Op.getValueType() == MVT::f64);
2127
2128 APFloat C1Val(APFloat::IEEEdouble(), "0x1.0p+52");
2129 SDValue C1 = DAG.getConstantFP(C1Val, SL, MVT::f64);
2130 SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
2131
2132 // TODO: Should this propagate fast-math-flags?
2133
2134 SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
2135 SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
2136
2137 SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
2138
2139 APFloat C2Val(APFloat::IEEEdouble(), "0x1.fffffffffffffp+51");
2140 SDValue C2 = DAG.getConstantFP(C2Val, SL, MVT::f64);
2141
2142 EVT SetCCVT =
2143 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
2144 SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
2145
2146 return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
2147 }
2148
2149 SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const {
2150 // FNEARBYINT and FRINT are the same, except in their handling of FP
2151 // exceptions. Those aren't really meaningful for us, and OpenCL only has
2152 // rint, so just treat them as equivalent.
2153 return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
2154 }
2155
2156 // XXX - May require not supporting f32 denormals?
2157
2158 // Don't handle v2f16. The extra instructions to scalarize and repack around the
2159 // compare and vselect end up producing worse code than scalarizing the whole
2160 // operation.
2161 SDValue AMDGPUTargetLowering::LowerFROUND32_16(SDValue Op, SelectionDAG &DAG) const {
2162 SDLoc SL(Op);
2163 SDValue X = Op.getOperand(0);
2164 EVT VT = Op.getValueType();
2165
2166 SDValue T = DAG.getNode(ISD::FTRUNC, SL, VT, X);
2167
2168 // TODO: Should this propagate fast-math-flags?
2169
2170 SDValue Diff = DAG.getNode(ISD::FSUB, SL, VT, X, T);
2171
2172 SDValue AbsDiff = DAG.getNode(ISD::FABS, SL, VT, Diff);
2173
2174 const SDValue Zero = DAG.getConstantFP(0.0, SL, VT);
2175 const SDValue One = DAG.getConstantFP(1.0, SL, VT);
2176 const SDValue Half = DAG.getConstantFP(0.5, SL, VT);
2177
2178 SDValue SignOne = DAG.getNode(ISD::FCOPYSIGN, SL, VT, One, X);
2179
2180 EVT SetCCVT =
2181 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
2182
2183 SDValue Cmp = DAG.getSetCC(SL, SetCCVT, AbsDiff, Half, ISD::SETOGE);
2184
2185 SDValue Sel = DAG.getNode(ISD::SELECT, SL, VT, Cmp, SignOne, Zero);
2186
2187 return DAG.getNode(ISD::FADD, SL, VT, T, Sel);
2188 }
2189
2190 SDValue AMDGPUTargetLowering::LowerFROUND64(SDValue Op, SelectionDAG &DAG) const {
2191 SDLoc SL(Op);
2192 SDValue X = Op.getOperand(0);
2193
2194 SDValue L = DAG.getNode(ISD::BITCAST, SL, MVT::i64, X);
2195
2196 const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
2197 const SDValue One = DAG.getConstant(1, SL, MVT::i32);
2198 const SDValue NegOne = DAG.getConstant(-1, SL, MVT::i32);
2199 const SDValue FiftyOne = DAG.getConstant(51, SL, MVT::i32);
2200 EVT SetCCVT =
2201 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i32);
2202
2203 SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, X);
2204
2205 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC, One);
2206
2207 SDValue Exp = extractF64Exponent(Hi, SL, DAG);
2208
2209 const SDValue Mask = DAG.getConstant(INT64_C(0x000fffffffffffff), SL,
2210 MVT::i64);
2211
2212 SDValue M = DAG.getNode(ISD::SRA, SL, MVT::i64, Mask, Exp);
2213 SDValue D = DAG.getNode(ISD::SRA, SL, MVT::i64,
2214 DAG.getConstant(INT64_C(0x0008000000000000), SL,
2215 MVT::i64),
2216 Exp);
2217
2218 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, L, M);
2219 SDValue Tmp1 = DAG.getSetCC(SL, SetCCVT,
2220 DAG.getConstant(0, SL, MVT::i64), Tmp0,
2221 ISD::SETNE);
2222
2223 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, Tmp1,
2224 D, DAG.getConstant(0, SL, MVT::i64));
2225 SDValue K = DAG.getNode(ISD::ADD, SL, MVT::i64, L, Tmp2);
2226
2227 K = DAG.getNode(ISD::AND, SL, MVT::i64, K, DAG.getNOT(SL, M, MVT::i64));
2228 K = DAG.getNode(ISD::BITCAST, SL, MVT::f64, K);
2229
2230 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
2231 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
2232 SDValue ExpEqNegOne = DAG.getSetCC(SL, SetCCVT, NegOne, Exp, ISD::SETEQ);
2233
2234 SDValue Mag = DAG.getNode(ISD::SELECT, SL, MVT::f64,
2235 ExpEqNegOne,
2236 DAG.getConstantFP(1.0, SL, MVT::f64),
2237 DAG.getConstantFP(0.0, SL, MVT::f64));
2238
2239 SDValue S = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, Mag, X);
2240
2241 K = DAG.getNode(ISD::SELECT, SL, MVT::f64, ExpLt0, S, K);
2242 K = DAG.getNode(ISD::SELECT, SL, MVT::f64, ExpGt51, X, K);
2243
2244 return K;
2245 }
2246
2247 SDValue AMDGPUTargetLowering::LowerFROUND(SDValue Op, SelectionDAG &DAG) const {
2248 EVT VT = Op.getValueType();
2249
2250 if (VT == MVT::f32 || VT == MVT::f16)
2251 return LowerFROUND32_16(Op, DAG);
2252
2253 if (VT == MVT::f64)
2254 return LowerFROUND64(Op, DAG);
2255
2256 llvm_unreachable("unhandled type");
2257 }
2258
2259 SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
2260 SDLoc SL(Op);
2261 SDValue Src = Op.getOperand(0);
2262
2263 // result = trunc(src);
2264 // if (src < 0.0 && src != result)
2265 // result += -1.0.
2266
2267 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2268
2269 const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64);
2270 const SDValue NegOne = DAG.getConstantFP(-1.0, SL, MVT::f64);
2271
2272 EVT SetCCVT =
2273 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
2274
2275 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
2276 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
2277 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
2278
2279 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
2280 // TODO: Should this propagate fast-math-flags?
2281 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
2282 }
2283
2284 SDValue AMDGPUTargetLowering::LowerFLOG(SDValue Op, SelectionDAG &DAG,
2285 double Log2BaseInverted) const {
2286 EVT VT = Op.getValueType();
2287
2288 SDLoc SL(Op);
2289 SDValue Operand = Op.getOperand(0);
2290 SDValue Log2Operand = DAG.getNode(ISD::FLOG2, SL, VT, Operand);
2291 SDValue Log2BaseInvertedOperand = DAG.getConstantFP(Log2BaseInverted, SL, VT);
2292
2293 return DAG.getNode(ISD::FMUL, SL, VT, Log2Operand, Log2BaseInvertedOperand);
2294 }
2295
2296 // exp2(M_LOG2E_F * f);
2297 SDValue AMDGPUTargetLowering::lowerFEXP(SDValue Op, SelectionDAG &DAG) const {
2298 EVT VT = Op.getValueType();
2299 SDLoc SL(Op);
2300 SDValue Src = Op.getOperand(0);
2301
2302 const SDValue K = DAG.getConstantFP(numbers::log2e, SL, VT);
2303 SDValue Mul = DAG.getNode(ISD::FMUL, SL, VT, Src, K, Op->getFlags());
2304 return DAG.getNode(ISD::FEXP2, SL, VT, Mul, Op->getFlags());
2305 }
2306
2307 static bool isCtlzOpc(unsigned Opc) {
2308 return Opc == ISD::CTLZ || Opc == ISD::CTLZ_ZERO_UNDEF;
2309 }
2310
2311 static bool isCttzOpc(unsigned Opc) {
2312 return Opc == ISD::CTTZ || Opc == ISD::CTTZ_ZERO_UNDEF;
2313 }
2314
2315 SDValue AMDGPUTargetLowering::LowerCTLZ_CTTZ(SDValue Op, SelectionDAG &DAG) const {
2316 SDLoc SL(Op);
2317 SDValue Src = Op.getOperand(0);
2318 bool ZeroUndef = Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF ||
2319 Op.getOpcode() == ISD::CTLZ_ZERO_UNDEF;
2320
2321 unsigned ISDOpc, NewOpc;
2322 if (isCtlzOpc(Op.getOpcode())) {
2323 ISDOpc = ISD::CTLZ_ZERO_UNDEF;
2324 NewOpc = AMDGPUISD::FFBH_U32;
2325 } else if (isCttzOpc(Op.getOpcode())) {
2326 ISDOpc = ISD::CTTZ_ZERO_UNDEF;
2327 NewOpc = AMDGPUISD::FFBL_B32;
2328 } else
2329 llvm_unreachable("Unexpected OPCode!!!");
2330
2331
2332 if (ZeroUndef && Src.getValueType() == MVT::i32)
2333 return DAG.getNode(NewOpc, SL, MVT::i32, Src);
2334
2335 SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
2336
2337 const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
2338 const SDValue One = DAG.getConstant(1, SL, MVT::i32);
2339
2340 SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero);
2341 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One);
2342
2343 EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(),
2344 *DAG.getContext(), MVT::i32);
2345
2346 SDValue HiOrLo = isCtlzOpc(Op.getOpcode()) ? Hi : Lo;
2347 SDValue Hi0orLo0 = DAG.getSetCC(SL, SetCCVT, HiOrLo, Zero, ISD::SETEQ);
2348
2349 SDValue OprLo = DAG.getNode(ISDOpc, SL, MVT::i32, Lo);
2350 SDValue OprHi = DAG.getNode(ISDOpc, SL, MVT::i32, Hi);
2351
2352 const SDValue Bits32 = DAG.getConstant(32, SL, MVT::i32);
2353 SDValue Add, NewOpr;
2354 if (isCtlzOpc(Op.getOpcode())) {
2355 Add = DAG.getNode(ISD::ADD, SL, MVT::i32, OprLo, Bits32);
2356 // ctlz(x) = hi_32(x) == 0 ? ctlz(lo_32(x)) + 32 : ctlz(hi_32(x))
2357 NewOpr = DAG.getNode(ISD::SELECT, SL, MVT::i32, Hi0orLo0, Add, OprHi);
2358 } else {
2359 Add = DAG.getNode(ISD::ADD, SL, MVT::i32, OprHi, Bits32);
2360 // cttz(x) = lo_32(x) == 0 ? cttz(hi_32(x)) + 32 : cttz(lo_32(x))
2361 NewOpr = DAG.getNode(ISD::SELECT, SL, MVT::i32, Hi0orLo0, Add, OprLo);
2362 }
2363
2364 if (!ZeroUndef) {
2365 // Test if the full 64-bit input is zero.
2366
2367 // FIXME: DAG combines turn what should be an s_and_b64 into a v_or_b32,
2368 // which we probably don't want.
2369 SDValue LoOrHi = isCtlzOpc(Op.getOpcode()) ? Lo : Hi;
2370 SDValue Lo0OrHi0 = DAG.getSetCC(SL, SetCCVT, LoOrHi, Zero, ISD::SETEQ);
2371 SDValue SrcIsZero = DAG.getNode(ISD::AND, SL, SetCCVT, Lo0OrHi0, Hi0orLo0);
2372
2373 // TODO: If i64 setcc is half rate, it can result in 1 fewer instruction
2374 // with the same cycles, otherwise it is slower.
2375 // SDValue SrcIsZero = DAG.getSetCC(SL, SetCCVT, Src,
2376 // DAG.getConstant(0, SL, MVT::i64), ISD::SETEQ);
2377
2378 const SDValue Bits32 = DAG.getConstant(64, SL, MVT::i32);
2379
2380 // The instruction returns -1 for 0 input, but the defined intrinsic
2381 // behavior is to return the number of bits.
2382 NewOpr = DAG.getNode(ISD::SELECT, SL, MVT::i32,
2383 SrcIsZero, Bits32, NewOpr);
2384 }
2385
2386 return DAG.getNode(ISD::ZERO_EXTEND, SL, MVT::i64, NewOpr);
2387 }
2388
2389 SDValue AMDGPUTargetLowering::LowerINT_TO_FP32(SDValue Op, SelectionDAG &DAG,
2390 bool Signed) const {
2391 // Unsigned
2392 // cul2f(ulong u)
2393 //{
2394 // uint lz = clz(u);
2395 // uint e = (u != 0) ? 127U + 63U - lz : 0;
2396 // u = (u << lz) & 0x7fffffffffffffffUL;
2397 // ulong t = u & 0xffffffffffUL;
2398 // uint v = (e << 23) | (uint)(u >> 40);
2399 // uint r = t > 0x8000000000UL ? 1U : (t == 0x8000000000UL ? v & 1U : 0U);
2400 // return as_float(v + r);
2401 //}
2402 // Signed
2403 // cl2f(long l)
2404 //{
2405 // long s = l >> 63;
2406 // float r = cul2f((l + s) ^ s);
2407 // return s ? -r : r;
2408 //}
2409
2410 SDLoc SL(Op);
2411 SDValue Src = Op.getOperand(0);
2412 SDValue L = Src;
2413
2414 SDValue S;
2415 if (Signed) {
2416 const SDValue SignBit = DAG.getConstant(63, SL, MVT::i64);
2417 S = DAG.getNode(ISD::SRA, SL, MVT::i64, L, SignBit);
2418
2419 SDValue LPlusS = DAG.getNode(ISD::ADD, SL, MVT::i64, L, S);
2420 L = DAG.getNode(ISD::XOR, SL, MVT::i64, LPlusS, S);
2421 }
2422
2423 EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(),
2424 *DAG.getContext(), MVT::f32);
2425
2426
2427 SDValue ZeroI32 = DAG.getConstant(0, SL, MVT::i32);
2428 SDValue ZeroI64 = DAG.getConstant(0, SL, MVT::i64);
2429 SDValue LZ = DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SL, MVT::i64, L);
2430 LZ = DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, LZ);
2431
2432 SDValue K = DAG.getConstant(127U + 63U, SL, MVT::i32);
2433 SDValue E = DAG.getSelect(SL, MVT::i32,
2434 DAG.getSetCC(SL, SetCCVT, L, ZeroI64, ISD::SETNE),
2435 DAG.getNode(ISD::SUB, SL, MVT::i32, K, LZ),
2436 ZeroI32);
2437
2438 SDValue U = DAG.getNode(ISD::AND, SL, MVT::i64,
2439 DAG.getNode(ISD::SHL, SL, MVT::i64, L, LZ),
2440 DAG.getConstant((-1ULL) >> 1, SL, MVT::i64));
2441
2442 SDValue T = DAG.getNode(ISD::AND, SL, MVT::i64, U,
2443 DAG.getConstant(0xffffffffffULL, SL, MVT::i64));
2444
2445 SDValue UShl = DAG.getNode(ISD::SRL, SL, MVT::i64,
2446 U, DAG.getConstant(40, SL, MVT::i64));
2447
2448 SDValue V = DAG.getNode(ISD::OR, SL, MVT::i32,
2449 DAG.getNode(ISD::SHL, SL, MVT::i32, E, DAG.getConstant(23, SL, MVT::i32)),
2450 DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, UShl));
2451
2452 SDValue C = DAG.getConstant(0x8000000000ULL, SL, MVT::i64);
2453 SDValue RCmp = DAG.getSetCC(SL, SetCCVT, T, C, ISD::SETUGT);
2454 SDValue TCmp = DAG.getSetCC(SL, SetCCVT, T, C, ISD::SETEQ);
2455
2456 SDValue One = DAG.getConstant(1, SL, MVT::i32);
2457
2458 SDValue VTrunc1 = DAG.getNode(ISD::AND, SL, MVT::i32, V, One);
2459
2460 SDValue R = DAG.getSelect(SL, MVT::i32,
2461 RCmp,
2462 One,
2463 DAG.getSelect(SL, MVT::i32, TCmp, VTrunc1, ZeroI32));
2464 R = DAG.getNode(ISD::ADD, SL, MVT::i32, V, R);
2465 R = DAG.getNode(ISD::BITCAST, SL, MVT::f32, R);
2466
2467 if (!Signed)
2468 return R;
2469
2470 SDValue RNeg = DAG.getNode(ISD::FNEG, SL, MVT::f32, R);
2471 return DAG.getSelect(SL, MVT::f32, DAG.getSExtOrTrunc(S, SL, SetCCVT), RNeg, R);
2472 }
2473
2474 SDValue AMDGPUTargetLowering::LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG,
2475 bool Signed) const {
2476 SDLoc SL(Op);
2477 SDValue Src = Op.getOperand(0);
2478
2479 SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
2480
2481 SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
2482 DAG.getConstant(0, SL, MVT::i32));
2483 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
2484 DAG.getConstant(1, SL, MVT::i32));
2485
2486 SDValue CvtHi = DAG.getNode(Signed ? ISD::SINT_TO_FP : ISD::UINT_TO_FP,
2487 SL, MVT::f64, Hi);
2488
2489 SDValue CvtLo = DAG.getNode(ISD::UINT_TO_FP, SL, MVT::f64, Lo);
2490
2491 SDValue LdExp = DAG.getNode(AMDGPUISD::LDEXP, SL, MVT::f64, CvtHi,
2492 DAG.getConstant(32, SL, MVT::i32));
2493 // TODO: Should this propagate fast-math-flags?
2494 return DAG.getNode(ISD::FADD, SL, MVT::f64, LdExp, CvtLo);
2495 }
2496
2497 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
2498 SelectionDAG &DAG) const {
2499 assert(Op.getOperand(0).getValueType() == MVT::i64 &&
2500 "operation should be legal");
2501
2502 // TODO: Factor out code common with LowerSINT_TO_FP.
2503
2504 EVT DestVT = Op.getValueType();
2505 if (Subtarget->has16BitInsts() && DestVT == MVT::f16) {
2506 SDLoc DL(Op);
2507 SDValue Src = Op.getOperand(0);
2508
2509 SDValue IntToFp32 = DAG.getNode(Op.getOpcode(), DL, MVT::f32, Src);
2510 SDValue FPRoundFlag = DAG.getIntPtrConstant(0, SDLoc(Op));
2511 SDValue FPRound =
2512 DAG.getNode(ISD::FP_ROUND, DL, MVT::f16, IntToFp32, FPRoundFlag);
2513
2514 return FPRound;
2515 }
2516
2517 if (DestVT == MVT::f32)
2518 return LowerINT_TO_FP32(Op, DAG, false);
2519
2520 assert(DestVT == MVT::f64);
2521 return LowerINT_TO_FP64(Op, DAG, false);
2522 }
2523
2524 SDValue AMDGPUTargetLowering::LowerSINT_TO_FP(SDValue Op,
2525 SelectionDAG &DAG) const {
2526 assert(Op.getOperand(0).getValueType() == MVT::i64 &&
2527 "operation should be legal");
2528
2529 // TODO: Factor out code common with LowerUINT_TO_FP.
2530
2531 EVT DestVT = Op.getValueType();
2532 if (Subtarget->has16BitInsts() && DestVT == MVT::f16) {
2533 SDLoc DL(Op);
2534 SDValue Src = Op.getOperand(0);
2535
2536 SDValue IntToFp32 = DAG.getNode(Op.getOpcode(), DL, MVT::f32, Src);
2537 SDValue FPRoundFlag = DAG.getIntPtrConstant(0, SDLoc(Op));
2538 SDValue FPRound =
2539 DAG.getNode(ISD::FP_ROUND, DL, MVT::f16, IntToFp32, FPRoundFlag);
2540
2541 return FPRound;
2542 }
2543
2544 if (DestVT == MVT::f32)
2545 return LowerINT_TO_FP32(Op, DAG, true);
2546
2547 assert(DestVT == MVT::f64);
2548 return LowerINT_TO_FP64(Op, DAG, true);
2549 }
2550
2551 SDValue AMDGPUTargetLowering::LowerFP64_TO_INT(SDValue Op, SelectionDAG &DAG,
2552 bool Signed) const {
2553 SDLoc SL(Op);
2554
2555 SDValue Src = Op.getOperand(0);
2556
2557 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2558
2559 SDValue K0 = DAG.getConstantFP(BitsToDouble(UINT64_C(0x3df0000000000000)), SL,
2560 MVT::f64);
2561 SDValue K1 = DAG.getConstantFP(BitsToDouble(UINT64_C(0xc1f0000000000000)), SL,
2562 MVT::f64);
2563 // TODO: Should this propagate fast-math-flags?
2564 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f64, Trunc, K0);
2565
2566 SDValue FloorMul = DAG.getNode(ISD::FFLOOR, SL, MVT::f64, Mul);
2567
2568
2569 SDValue Fma = DAG.getNode(ISD::FMA, SL, MVT::f64, FloorMul, K1, Trunc);
2570
2571 SDValue Hi = DAG.getNode(Signed ? ISD::FP_TO_SINT : ISD::FP_TO_UINT, SL,
2572 MVT::i32, FloorMul);
2573 SDValue Lo = DAG.getNode(ISD::FP_TO_UINT, SL, MVT::i32, Fma);
2574
2575 SDValue Result = DAG.getBuildVector(MVT::v2i32, SL, {Lo, Hi});
2576
2577 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Result);
2578 }
2579
2580 SDValue AMDGPUTargetLowering::LowerFP_TO_FP16(SDValue Op, SelectionDAG &DAG) const {
2581 SDLoc DL(Op);
2582 SDValue N0 = Op.getOperand(0);
2583
2584 // Convert to target node to get known bits
2585 if (N0.getValueType() == MVT::f32)
2586 return DAG.getNode(AMDGPUISD::FP_TO_FP16, DL, Op.getValueType(), N0);
2587
2588 if (getTargetMachine().Options.UnsafeFPMath) {
2589 // There is a generic expand for FP_TO_FP16 with unsafe fast math.
2590 return SDValue();
2591 }
2592
2593 assert(N0.getSimpleValueType() == MVT::f64);
2594
2595 // f64 -> f16 conversion using round-to-nearest-even rounding mode.
2596 const unsigned ExpMask = 0x7ff;
2597 const unsigned ExpBiasf64 = 1023;
2598 const unsigned ExpBiasf16 = 15;
2599 SDValue Zero = DAG.getConstant(0, DL, MVT::i32);
2600 SDValue One = DAG.getConstant(1, DL, MVT::i32);
2601 SDValue U = DAG.getNode(ISD::BITCAST, DL, MVT::i64, N0);
2602 SDValue UH = DAG.getNode(ISD::SRL, DL, MVT::i64, U,
2603 DAG.getConstant(32, DL, MVT::i64));
2604 UH = DAG.getZExtOrTrunc(UH, DL, MVT::i32);
2605 U = DAG.getZExtOrTrunc(U, DL, MVT::i32);
2606 SDValue E = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
2607 DAG.getConstant(20, DL, MVT::i64));
2608 E = DAG.getNode(ISD::AND, DL, MVT::i32, E,
2609 DAG.getConstant(ExpMask, DL, MVT::i32));
2610 // Subtract the fp64 exponent bias (1023) to get the real exponent and
2611 // add the f16 bias (15) to get the biased exponent for the f16 format.
2612 E = DAG.getNode(ISD::ADD, DL, MVT::i32, E,
2613 DAG.getConstant(-ExpBiasf64 + ExpBiasf16, DL, MVT::i32));
2614
2615 SDValue M = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
2616 DAG.getConstant(8, DL, MVT::i32));
2617 M = DAG.getNode(ISD::AND, DL, MVT::i32, M,
2618 DAG.getConstant(0xffe, DL, MVT::i32));
2619
2620 SDValue MaskedSig = DAG.getNode(ISD::AND, DL, MVT::i32, UH,
2621 DAG.getConstant(0x1ff, DL, MVT::i32));
2622 MaskedSig = DAG.getNode(ISD::OR, DL, MVT::i32, MaskedSig, U);
2623
2624 SDValue Lo40Set = DAG.getSelectCC(DL, MaskedSig, Zero, Zero, One, ISD::SETEQ);
2625 M = DAG.getNode(ISD::OR, DL, MVT::i32, M, Lo40Set);
2626
2627 // (M != 0 ? 0x0200 : 0) | 0x7c00;
2628 SDValue I = DAG.getNode(ISD::OR, DL, MVT::i32,
2629 DAG.getSelectCC(DL, M, Zero, DAG.getConstant(0x0200, DL, MVT::i32),
2630 Zero, ISD::SETNE), DAG.getConstant(0x7c00, DL, MVT::i32));
2631
2632 // N = M | (E << 12);
2633 SDValue N = DAG.getNode(ISD::OR, DL, MVT::i32, M,
2634 DAG.getNode(ISD::SHL, DL, MVT::i32, E,
2635 DAG.getConstant(12, DL, MVT::i32)));
2636
2637 // B = clamp(1-E, 0, 13);
2638 SDValue OneSubExp = DAG.getNode(ISD::SUB, DL, MVT::i32,
2639 One, E);
2640 SDValue B = DAG.getNode(ISD::SMAX, DL, MVT::i32, OneSubExp, Zero);
2641 B = DAG.getNode(ISD::SMIN, DL, MVT::i32, B,
2642 DAG.getConstant(13, DL, MVT::i32));
2643
2644 SDValue SigSetHigh = DAG.getNode(ISD::OR, DL, MVT::i32, M,
2645 DAG.getConstant(0x1000, DL, MVT::i32));
2646
2647 SDValue D = DAG.getNode(ISD::SRL, DL, MVT::i32, SigSetHigh, B);
2648 SDValue D0 = DAG.getNode(ISD::SHL, DL, MVT::i32, D, B);
2649 SDValue D1 = DAG.getSelectCC(DL, D0, SigSetHigh, One, Zero, ISD::SETNE);
2650 D = DAG.getNode(ISD::OR, DL, MVT::i32, D, D1);
2651
2652 SDValue V = DAG.getSelectCC(DL, E, One, D, N, ISD::SETLT);
2653 SDValue VLow3 = DAG.getNode(ISD::AND, DL, MVT::i32, V,
2654 DAG.getConstant(0x7, DL, MVT::i32));
2655 V = DAG.getNode(ISD::SRL, DL, MVT::i32, V,
2656 DAG.getConstant(2, DL, MVT::i32));
2657 SDValue V0 = DAG.getSelectCC(DL, VLow3, DAG.getConstant(3, DL, MVT::i32),
2658 One, Zero, ISD::SETEQ);
2659 SDValue V1 = DAG.getSelectCC(DL, VLow3, DAG.getConstant(5, DL, MVT::i32),
2660 One, Zero, ISD::SETGT);
2661 V1 = DAG.getNode(ISD::OR, DL, MVT::i32, V0, V1);
2662 V = DAG.getNode(ISD::ADD, DL, MVT::i32, V, V1);
2663
2664 V = DAG.getSelectCC(DL, E, DAG.getConstant(30, DL, MVT::i32),
2665 DAG.getConstant(0x7c00, DL, MVT::i32), V, ISD::SETGT);
2666 V = DAG.getSelectCC(DL, E, DAG.getConstant(1039, DL, MVT::i32),
2667 I, V, ISD::SETEQ);
2668
2669 // Extract the sign bit.
2670 SDValue Sign = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
2671 DAG.getConstant(16, DL, MVT::i32));
2672 Sign = DAG.getNode(ISD::AND, DL, MVT::i32, Sign,
2673 DAG.getConstant(0x8000, DL, MVT::i32));
2674
2675 V = DAG.getNode(ISD::OR, DL, MVT::i32, Sign, V);
2676 return DAG.getZExtOrTrunc(V, DL, Op.getValueType());
2677 }
2678
2679 SDValue AMDGPUTargetLowering::LowerFP_TO_SINT(SDValue Op,
2680 SelectionDAG &DAG) const {
2681 SDValue Src = Op.getOperand(0);
2682
2683 // TODO: Factor out code common with LowerFP_TO_UINT.
2684
2685 EVT SrcVT = Src.getValueType();
2686 if (Subtarget->has16BitInsts() && SrcVT == MVT::f16) {
2687 SDLoc DL(Op);
2688
2689 SDValue FPExtend = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Src);
2690 SDValue FpToInt32 =
2691 DAG.getNode(Op.getOpcode(), DL, MVT::i64, FPExtend);
2692
2693 return FpToInt32;
2694 }
2695
2696 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2697 return LowerFP64_TO_INT(Op, DAG, true);
2698
2699 return SDValue();
2700 }
2701
2702 SDValue AMDGPUTargetLowering::LowerFP_TO_UINT(SDValue Op,
2703 SelectionDAG &DAG) const {
2704 SDValue Src = Op.getOperand(0);
2705
2706 // TODO: Factor out code common with LowerFP_TO_SINT.
2707
2708 EVT SrcVT = Src.getValueType();
2709 if (Subtarget->has16BitInsts() && SrcVT == MVT::f16) {
2710 SDLoc DL(Op);
2711
2712 SDValue FPExtend = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Src);
2713 SDValue FpToInt32 =
2714 DAG.getNode(Op.getOpcode(), DL, MVT::i64, FPExtend);
2715
2716 return FpToInt32;
2717 }
2718
2719 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2720 return LowerFP64_TO_INT(Op, DAG, false);
2721
2722 return SDValue();
2723 }
2724
2725 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
2726 SelectionDAG &DAG) const {
2727 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
2728 MVT VT = Op.getSimpleValueType();
2729 MVT ScalarVT = VT.getScalarType();
2730
2731 assert(VT.isVector());
2732
2733 SDValue Src = Op.getOperand(0);
2734 SDLoc DL(Op);
2735
2736 // TODO: Don't scalarize on Evergreen?
2737 unsigned NElts = VT.getVectorNumElements();
2738 SmallVector<SDValue, 8> Args;
2739 DAG.ExtractVectorElements(Src, Args, 0, NElts);
2740
2741 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
2742 for (unsigned I = 0; I < NElts; ++I)
2743 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
2744
2745 return DAG.getBuildVector(VT, DL, Args);
2746 }
2747
2748 //===----------------------------------------------------------------------===//
2749 // Custom DAG optimizations
2750 //===----------------------------------------------------------------------===//
2751
2752 static bool isU24(SDValue Op, SelectionDAG &DAG) {
2753 return AMDGPUTargetLowering::numBitsUnsigned(Op, DAG) <= 24;
2754 }
2755
2756 static bool isI24(SDValue Op, SelectionDAG &DAG) {
2757 EVT VT = Op.getValueType();
2758 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
2759 // as unsigned 24-bit values.
2760 AMDGPUTargetLowering::numBitsSigned(Op, DAG) < 24;
2761 }
2762
2763 static SDValue simplifyI24(SDNode *Node24,
2764 TargetLowering::DAGCombinerInfo &DCI) {
2765 SelectionDAG &DAG = DCI.DAG;
2766 bool IsIntrin = Node24->getOpcode() == ISD::INTRINSIC_WO_CHAIN;
2767
2768 SDValue LHS = IsIntrin ? Node24->getOperand(1) : Node24->getOperand(0);
2769 SDValue RHS = IsIntrin ? Node24->getOperand(2) : Node24->getOperand(1);
2770 unsigned NewOpcode = Node24->getOpcode();
2771 if (IsIntrin) {
2772 unsigned IID = cast<ConstantSDNode>(Node24->getOperand(0))->getZExtValue();
2773 NewOpcode = IID == Intrinsic::amdgcn_mul_i24 ?
2774 AMDGPUISD::MUL_I24 : AMDGPUISD::MUL_U24;
2775 }
2776
2777 APInt Demanded = APInt::getLowBitsSet(LHS.getValueSizeInBits(), 24);
2778
2779 // First try to simplify using GetDemandedBits which allows the operands to
2780 // have other uses, but will only perform simplifications that involve
2781 // bypassing some nodes for this user.
2782 SDValue DemandedLHS = DAG.GetDemandedBits(LHS, Demanded);
2783 SDValue DemandedRHS = DAG.GetDemandedBits(RHS, Demanded);
2784 if (DemandedLHS || DemandedRHS)
2785 return DAG.getNode(NewOpcode, SDLoc(Node24), Node24->getVTList(),
2786 DemandedLHS ? DemandedLHS : LHS,
2787 DemandedRHS ? DemandedRHS : RHS);
2788
2789 // Now try SimplifyDemandedBits which can simplify the nodes used by our
2790 // operands if this node is the only user.
2791 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2792 if (TLI.SimplifyDemandedBits(LHS, Demanded, DCI))
2793 return SDValue(Node24, 0);
2794 if (TLI.SimplifyDemandedBits(RHS, Demanded, DCI))
2795 return SDValue(Node24, 0);
2796
2797 return SDValue();
2798 }
2799
2800 template <typename IntTy>
2801 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0, uint32_t Offset,
2802 uint32_t Width, const SDLoc &DL) {
2803 if (Width + Offset < 32) {
2804 uint32_t Shl = static_cast<uint32_t>(Src0) << (32 - Offset - Width);
2805 IntTy Result = static_cast<IntTy>(Shl) >> (32 - Width);
2806 return DAG.getConstant(Result, DL, MVT::i32);
2807 }
2808
2809 return DAG.getConstant(Src0 >> Offset, DL, MVT::i32);
2810 }
2811
2812 static bool hasVolatileUser(SDNode *Val) {
2813 for (SDNode *U : Val->uses()) {
2814 if (MemSDNode *M = dyn_cast<MemSDNode>(U)) {
2815 if (M->isVolatile())
2816 return true;
2817 }
2818 }
2819
2820 return false;
2821 }
2822
2823 bool AMDGPUTargetLowering::shouldCombineMemoryType(EVT VT) const {
2824 // i32 vectors are the canonical memory type.
2825 if (VT.getScalarType() == MVT::i32 || isTypeLegal(VT))
2826 return false;
2827
2828 if (!VT.isByteSized())
2829 return false;
2830
2831 unsigned Size = VT.getStoreSize();
2832
2833 if ((Size == 1 || Size == 2 || Size == 4) && !VT.isVector())
2834 return false;
2835
2836 if (Size == 3 || (Size > 4 && (Size % 4 != 0)))
2837 return false;
2838
2839 return true;
2840 }
2841
2842 // Replace load of an illegal type with a store of a bitcast to a friendlier
2843 // type.
2844 SDValue AMDGPUTargetLowering::performLoadCombine(SDNode *N,
2845 DAGCombinerInfo &DCI) const {
2846 if (!DCI.isBeforeLegalize())
2847 return SDValue();
2848
2849 LoadSDNode *LN = cast<LoadSDNode>(N);
2850 if (LN->isVolatile() || !ISD::isNormalLoad(LN) || hasVolatileUser(LN))
2851 return SDValue();
2852
2853 SDLoc SL(N);
2854 SelectionDAG &DAG = DCI.DAG;
2855 EVT VT = LN->getMemoryVT();
2856
2857 unsigned Size = VT.getStoreSize();
2858 unsigned Align = LN->getAlignment();
2859 if (Align < Size && isTypeLegal(VT)) {
2860 bool IsFast;
2861 unsigned AS = LN->getAddressSpace();
2862
2863 // Expand unaligned loads earlier than legalization. Due to visitation order
2864 // problems during legalization, the emitted instructions to pack and unpack
2865 // the bytes again are not eliminated in the case of an unaligned copy.
2866 if (!allowsMisalignedMemoryAccesses(
2867 VT, AS, Align, LN->getMemOperand()->getFlags(), &IsFast)) {
2868 if (VT.isVector())
2869 return scalarizeVectorLoad(LN, DAG);
2870
2871 SDValue Ops[2];
2872 std::tie(Ops[0], Ops[1]) = expandUnalignedLoad(LN, DAG);
2873 return DAG.getMergeValues(Ops, SDLoc(N));
2874 }
2875
2876 if (!IsFast)
2877 return SDValue();
2878 }
2879
2880 if (!shouldCombineMemoryType(VT))
2881 return SDValue();
2882
2883 EVT NewVT = getEquivalentMemType(*DAG.getContext(), VT);
2884
2885 SDValue NewLoad
2886 = DAG.getLoad(NewVT, SL, LN->getChain(),
2887 LN->getBasePtr(), LN->getMemOperand());
2888
2889 SDValue BC = DAG.getNode(ISD::BITCAST, SL, VT, NewLoad);
2890 DCI.CombineTo(N, BC, NewLoad.getValue(1));
2891 return SDValue(N, 0);
2892 }
2893
2894 // Replace store of an illegal type with a store of a bitcast to a friendlier
2895 // type.
2896 SDValue AMDGPUTargetLowering::performStoreCombine(SDNode *N,
2897 DAGCombinerInfo &DCI) const {
2898 if (!DCI.isBeforeLegalize())
2899 return SDValue();
2900
2901 StoreSDNode *SN = cast<StoreSDNode>(N);
2902 if (SN->isVolatile() || !ISD::isNormalStore(SN))
2903 return SDValue();
2904
2905 EVT VT = SN->getMemoryVT();
2906 unsigned Size = VT.getStoreSize();
2907
2908 SDLoc SL(N);
2909 SelectionDAG &DAG = DCI.DAG;
2910 unsigned Align = SN->getAlignment();
2911 if (Align < Size && isTypeLegal(VT)) {
2912 bool IsFast;
2913 unsigned AS = SN->getAddressSpace();
2914
2915 // Expand unaligned stores earlier than legalization. Due to visitation
2916 // order problems during legalization, the emitted instructions to pack and
2917 // unpack the bytes again are not eliminated in the case of an unaligned
2918 // copy.
2919 if (!allowsMisalignedMemoryAccesses(
2920 VT, AS, Align, SN->getMemOperand()->getFlags(), &IsFast)) {
2921 if (VT.isVector())
2922 return scalarizeVectorStore(SN, DAG);
2923
2924 return expandUnalignedStore(SN, DAG);
2925 }
2926
2927 if (!IsFast)
2928 return SDValue();
2929 }
2930
2931 if (!shouldCombineMemoryType(VT))
2932 return SDValue();
2933
2934 EVT NewVT = getEquivalentMemType(*DAG.getContext(), VT);
2935 SDValue Val = SN->getValue();
2936
2937 //DCI.AddToWorklist(Val.getNode());
2938
2939 bool OtherUses = !Val.hasOneUse();
2940 SDValue CastVal = DAG.getNode(ISD::BITCAST, SL, NewVT, Val);
2941 if (OtherUses) {
2942 SDValue CastBack = DAG.getNode(ISD::BITCAST, SL, VT, CastVal);
2943 DAG.ReplaceAllUsesOfValueWith(Val, CastBack);
2944 }
2945
2946 return DAG.getStore(SN->getChain(), SL, CastVal,
2947 SN->getBasePtr(), SN->getMemOperand());
2948 }
2949
2950 // FIXME: This should go in generic DAG combiner with an isTruncateFree check,
2951 // but isTruncateFree is inaccurate for i16 now because of SALU vs. VALU
2952 // issues.
2953 SDValue AMDGPUTargetLowering::performAssertSZExtCombine(SDNode *N,
2954 DAGCombinerInfo &DCI) const {
2955 SelectionDAG &DAG = DCI.DAG;
2956 SDValue N0 = N->getOperand(0);
2957
2958 // (vt2 (assertzext (truncate vt0:x), vt1)) ->
2959 // (vt2 (truncate (assertzext vt0:x, vt1)))
2960 if (N0.getOpcode() == ISD::TRUNCATE) {
2961 SDValue N1 = N->getOperand(1);
2962 EVT ExtVT = cast<VTSDNode>(N1)->getVT();
2963 SDLoc SL(N);
2964
2965 SDValue Src = N0.getOperand(0);
2966 EVT SrcVT = Src.getValueType();
2967 if (SrcVT.bitsGE(ExtVT)) {
2968 SDValue NewInReg = DAG.getNode(N->getOpcode(), SL, SrcVT, Src, N1);
2969 return DAG.getNode(ISD::TRUNCATE, SL, N->getValueType(0), NewInReg);
2970 }
2971 }
2972
2973 return SDValue();
2974 }
2975
2976 SDValue AMDGPUTargetLowering::performIntrinsicWOChainCombine(
2977 SDNode *N, DAGCombinerInfo &DCI) const {
2978 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
2979 switch (IID) {
2980 case Intrinsic::amdgcn_mul_i24:
2981 case Intrinsic::amdgcn_mul_u24:
2982 return simplifyI24(N, DCI);
2983 default:
2984 return SDValue();
2985 }
2986 }
2987
2988 /// Split the 64-bit value \p LHS into two 32-bit components, and perform the
2989 /// binary operation \p Opc to it with the corresponding constant operands.
2990 SDValue AMDGPUTargetLowering::splitBinaryBitConstantOpImpl(
2991 DAGCombinerInfo &DCI, const SDLoc &SL,
2992 unsigned Opc, SDValue LHS,
2993 uint32_t ValLo, uint32_t ValHi) const {
2994 SelectionDAG &DAG = DCI.DAG;
2995 SDValue Lo, Hi;
2996 std::tie(Lo, Hi) = split64BitValue(LHS, DAG);
2997
2998 SDValue LoRHS = DAG.getConstant(ValLo, SL, MVT::i32);
2999 SDValue HiRHS = DAG.getConstant(ValHi, SL, MVT::i32);
3000
3001 SDValue LoAnd = DAG.getNode(Opc, SL, MVT::i32, Lo, LoRHS);
3002 SDValue HiAnd = DAG.getNode(Opc, SL, MVT::i32, Hi, HiRHS);
3003
3004 // Re-visit the ands. It's possible we eliminated one of them and it could
3005 // simplify the vector.
3006 DCI.AddToWorklist(Lo.getNode());
3007 DCI.AddToWorklist(Hi.getNode());
3008
3009 SDValue Vec = DAG.getBuildVector(MVT::v2i32, SL, {LoAnd, HiAnd});
3010 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Vec);
3011 }
3012
3013 SDValue AMDGPUTargetLowering::performShlCombine(SDNode *N,
3014 DAGCombinerInfo &DCI) const {
3015 EVT VT = N->getValueType(0);
3016
3017 ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
3018 if (!RHS)
3019 return SDValue();
3020
3021 SDValue LHS = N->getOperand(0);
3022 unsigned RHSVal = RHS->getZExtValue();
3023 if (!RHSVal)
3024 return LHS;
3025
3026 SDLoc SL(N);
3027 SelectionDAG &DAG = DCI.DAG;
3028
3029 switch (LHS->getOpcode()) {
3030 default:
3031 break;
3032 case ISD::ZERO_EXTEND:
3033 case ISD::SIGN_EXTEND:
3034 case ISD::ANY_EXTEND: {
3035 SDValue X = LHS->getOperand(0);
3036
3037 if (VT == MVT::i32 && RHSVal == 16 && X.getValueType() == MVT::i16 &&
3038 isOperationLegal(ISD::BUILD_VECTOR, MVT::v2i16)) {
3039 // Prefer build_vector as the canonical form if packed types are legal.
3040 // (shl ([asz]ext i16:x), 16 -> build_vector 0, x
3041 SDValue Vec = DAG.getBuildVector(MVT::v2i16, SL,
3042 { DAG.getConstant(0, SL, MVT::i16), LHS->getOperand(0) });
3043 return DAG.getNode(ISD::BITCAST, SL, MVT::i32, Vec);
3044 }
3045
3046 // shl (ext x) => zext (shl x), if shift does not overflow int
3047 if (VT != MVT::i64)
3048 break;
3049 KnownBits Known = DAG.computeKnownBits(X);
3050 unsigned LZ = Known.countMinLeadingZeros();
3051 if (LZ < RHSVal)
3052 break;
3053 EVT XVT = X.getValueType();
3054 SDValue Shl = DAG.getNode(ISD::SHL, SL, XVT, X, SDValue(RHS, 0));
3055 return DAG.getZExtOrTrunc(Shl, SL, VT);
3056 }
3057 }
3058
3059 if (VT != MVT::i64)
3060 return SDValue();
3061
3062 // i64 (shl x, C) -> (build_pair 0, (shl x, C -32))
3063
3064 // On some subtargets, 64-bit shift is a quarter rate instruction. In the
3065 // common case, splitting this into a move and a 32-bit shift is faster and
3066 // the same code size.
3067 if (RHSVal < 32)
3068 return SDValue();
3069
3070 SDValue ShiftAmt = DAG.getConstant(RHSVal - 32, SL, MVT::i32);
3071
3072 SDValue Lo = DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, LHS);
3073 SDValue NewShift = DAG.getNode(ISD::SHL, SL, MVT::i32, Lo, ShiftAmt);
3074
3075 const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
3076
3077 SDValue Vec = DAG.getBuildVector(MVT::v2i32, SL, {Zero, NewShift});
3078 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Vec);
3079 }
3080
3081 SDValue AMDGPUTargetLowering::performSraCombine(SDNode *N,
3082 DAGCombinerInfo &DCI) const {
3083 if (N->getValueType(0) != MVT::i64)
3084 return SDValue();
3085
3086 const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
3087 if (!RHS)
3088 return SDValue();
3089
3090 SelectionDAG &DAG = DCI.DAG;
3091 SDLoc SL(N);
3092 unsigned RHSVal = RHS->getZExtValue();
3093
3094 // (sra i64:x, 32) -> build_pair x, (sra hi_32(x), 31)
3095 if (RHSVal == 32) {
3096 SDValue Hi = getHiHalf64(N->getOperand(0), DAG);
3097 SDValue NewShift = DAG.getNode(ISD::SRA, SL, MVT::i32, Hi,
3098 DAG.getConstant(31, SL, MVT::i32));
3099
3100 SDValue BuildVec = DAG.getBuildVector(MVT::v2i32, SL, {Hi, NewShift});
3101 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildVec);
3102 }
3103
3104 // (sra i64:x, 63) -> build_pair (sra hi_32(x), 31), (sra hi_32(x), 31)
3105 if (RHSVal == 63) {
3106 SDValue Hi = getHiHalf64(N->getOperand(0), DAG);
3107 SDValue NewShift = DAG.getNode(ISD::SRA, SL, MVT::i32, Hi,
3108 DAG.getConstant(31, SL, MVT::i32));
3109 SDValue BuildVec = DAG.getBuildVector(MVT::v2i32, SL, {NewShift, NewShift});
3110 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildVec);
3111 }
3112
3113 return SDValue();
3114 }
3115
3116 SDValue AMDGPUTargetLowering::performSrlCombine(SDNode *N,
3117 DAGCombinerInfo &DCI) const {
3118 auto *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
3119 if (!RHS)
3120 return SDValue();
3121
3122 EVT VT = N->getValueType(0);
3123 SDValue LHS = N->getOperand(0);
3124 unsigned ShiftAmt = RHS->getZExtValue();
3125 SelectionDAG &DAG = DCI.DAG;
3126 SDLoc SL(N);
3127
3128 // fold (srl (and x, c1 << c2), c2) -> (and (srl(x, c2), c1)
3129 // this improves the ability to match BFE patterns in isel.
3130 if (LHS.getOpcode() == ISD::AND) {
3131 if (auto *Mask = dyn_cast<ConstantSDNode>(LHS.getOperand(1))) {
3132 if (Mask->getAPIntValue().isShiftedMask() &&
3133 Mask->getAPIntValue().countTrailingZeros() == ShiftAmt) {
3134 return DAG.getNode(
3135 ISD::AND, SL, VT,
3136 DAG.getNode(ISD::SRL, SL, VT, LHS.getOperand(0), N->getOperand(1)),
3137 DAG.getNode(ISD::SRL, SL, VT, LHS.getOperand(1), N->getOperand(1)));
3138 }
3139 }
3140 }
3141
3142 if (VT != MVT::i64)
3143 return SDValue();
3144
3145 if (ShiftAmt < 32)
3146 return SDValue();
3147
3148 // srl i64:x, C for C >= 32
3149 // =>
3150 // build_pair (srl hi_32(x), C - 32), 0
3151 SDValue One = DAG.getConstant(1, SL, MVT::i32);
3152 SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
3153
3154 SDValue VecOp = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, LHS);
3155 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecOp, One);
3156
3157 SDValue NewConst = DAG.getConstant(ShiftAmt - 32, SL, MVT::i32);
3158 SDValue NewShift = DAG.getNode(ISD::SRL, SL, MVT::i32, Hi, NewConst);
3159
3160 SDValue BuildPair = DAG.getBuildVector(MVT::v2i32, SL, {NewShift, Zero});
3161
3162 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildPair);
3163 }
3164
3165 SDValue AMDGPUTargetLowering::performTruncateCombine(
3166 SDNode *N, DAGCombinerInfo &DCI) const {
3167 SDLoc SL(N);
3168 SelectionDAG &DAG = DCI.DAG;
3169 EVT VT = N->getValueType(0);
3170 SDValue Src = N->getOperand(0);
3171
3172 // vt1 (truncate (bitcast (build_vector vt0:x, ...))) -> vt1 (bitcast vt0:x)
3173 if (Src.getOpcode() == ISD::BITCAST && !VT.isVector()) {
3174 SDValue Vec = Src.getOperand(0);
3175 if (Vec.getOpcode() == ISD::BUILD_VECTOR) {
3176 SDValue Elt0 = Vec.getOperand(0);
3177 EVT EltVT = Elt0.getValueType();
3178 if (VT.getSizeInBits() <= EltVT.getSizeInBits()) {
3179 if (EltVT.isFloatingPoint()) {
3180 Elt0 = DAG.getNode(ISD::BITCAST, SL,
3181 EltVT.changeTypeToInteger(), Elt0);
3182 }
3183
3184 return DAG.getNode(ISD::TRUNCATE, SL, VT, Elt0);
3185 }
3186 }
3187 }
3188
3189 // Equivalent of above for accessing the high element of a vector as an
3190 // integer operation.
3191 // trunc (srl (bitcast (build_vector x, y))), 16 -> trunc (bitcast y)
3192 if (Src.getOpcode() == ISD::SRL && !VT.isVector()) {
3193 if (auto K = isConstOrConstSplat(Src.getOperand(1))) {
3194 if (2 * K->getZExtValue() == Src.getValueType().getScalarSizeInBits()) {
3195 SDValue BV = stripBitcast(Src.getOperand(0));
3196 if (BV.getOpcode() == ISD::BUILD_VECTOR &&
3197 BV.getValueType().getVectorNumElements() == 2) {
3198 SDValue SrcElt = BV.getOperand(1);
3199 EVT SrcEltVT = SrcElt.getValueType();
3200 if (SrcEltVT.isFloatingPoint()) {
3201 SrcElt = DAG.getNode(ISD::BITCAST, SL,
3202 SrcEltVT.changeTypeToInteger(), SrcElt);
3203 }
3204
3205 return DAG.getNode(ISD::TRUNCATE, SL, VT, SrcElt);
3206 }
3207 }
3208 }
3209 }
3210
3211 // Partially shrink 64-bit shifts to 32-bit if reduced to 16-bit.
3212 //
3213 // i16 (trunc (srl i64:x, K)), K <= 16 ->
3214 // i16 (trunc (srl (i32 (trunc x), K)))
3215 if (VT.getScalarSizeInBits() < 32) {
3216 EVT SrcVT = Src.getValueType();
3217 if (SrcVT.getScalarSizeInBits() > 32 &&
3218 (Src.getOpcode() == ISD::SRL ||
3219 Src.getOpcode() == ISD::SRA ||
3220 Src.getOpcode() == ISD::SHL)) {
3221 SDValue Amt = Src.getOperand(1);
3222 KnownBits Known = DAG.computeKnownBits(Amt);
3223 unsigned Size = VT.getScalarSizeInBits();
3224 if ((Known.isConstant() && Known.getConstant().ule(Size)) ||
3225 (Known.getBitWidth() - Known.countMinLeadingZeros() <= Log2_32(Size))) {
3226 EVT MidVT = VT.isVector() ?
3227 EVT::getVectorVT(*DAG.getContext(), MVT::i32,
3228 VT.getVectorNumElements()) : MVT::i32;
3229
3230 EVT NewShiftVT = getShiftAmountTy(MidVT, DAG.getDataLayout());
3231 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SL, MidVT,
3232 Src.getOperand(0));
3233 DCI.AddToWorklist(Trunc.getNode());
3234
3235 if (Amt.getValueType() != NewShiftVT) {
3236 Amt = DAG.getZExtOrTrunc(Amt, SL, NewShiftVT);
3237 DCI.AddToWorklist(Amt.getNode());
3238 }
3239
3240 SDValue ShrunkShift = DAG.getNode(Src.getOpcode(), SL, MidVT,
3241 Trunc, Amt);
3242 return DAG.getNode(ISD::TRUNCATE, SL, VT, ShrunkShift);
3243 }
3244 }
3245 }
3246
3247 return SDValue();
3248 }
3249
3250 // We need to specifically handle i64 mul here to avoid unnecessary conversion
3251 // instructions. If we only match on the legalized i64 mul expansion,
3252 // SimplifyDemandedBits will be unable to remove them because there will be
3253 // multiple uses due to the separate mul + mulh[su].
3254 static SDValue getMul24(SelectionDAG &DAG, const SDLoc &SL,
3255 SDValue N0, SDValue N1, unsigned Size, bool Signed) {
3256 if (Size <= 32) {
3257 unsigned MulOpc = Signed ? AMDGPUISD::MUL_I24 : AMDGPUISD::MUL_U24;
3258 return DAG.getNode(MulOpc, SL, MVT::i32, N0, N1);
3259 }
3260
3261 // Because we want to eliminate extension instructions before the
3262 // operation, we need to create a single user here (i.e. not the separate
3263 // mul_lo + mul_hi) so that SimplifyDemandedBits will deal with it.
3264
3265 unsigned MulOpc = Signed ? AMDGPUISD::MUL_LOHI_I24 : AMDGPUISD::MUL_LOHI_U24;
3266
3267 SDValue Mul = DAG.getNode(MulOpc, SL,
3268 DAG.getVTList(MVT::i32, MVT::i32), N0, N1);
3269
3270 return DAG.getNode(ISD::BUILD_PAIR, SL, MVT::i64,
3271 Mul.getValue(0), Mul.getValue(1));
3272 }
3273
3274 SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N,
3275 DAGCombinerInfo &DCI) const {
3276 EVT VT = N->getValueType(0);
3277
3278 unsigned Size = VT.getSizeInBits();
3279 if (VT.isVector() || Size > 64)
3280 return SDValue();
3281
3282 // There are i16 integer mul/mad.
3283 if (Subtarget->has16BitInsts() && VT.getScalarType().bitsLE(MVT::i16))
3284 return SDValue();
3285
3286 SelectionDAG &DAG = DCI.DAG;
3287 SDLoc DL(N);
3288
3289 SDValue N0 = N->getOperand(0);
3290 SDValue N1 = N->getOperand(1);
3291
3292 // SimplifyDemandedBits has the annoying habit of turning useful zero_extends
3293 // in the source into any_extends if the result of the mul is truncated. Since
3294 // we can assume the high bits are whatever we want, use the underlying value
3295 // to avoid the unknown high bits from interfering.
3296 if (N0.getOpcode() == ISD::ANY_EXTEND)
3297 N0 = N0.getOperand(0);
3298
3299 if (N1.getOpcode() == ISD::ANY_EXTEND)
3300 N1 = N1.getOperand(0);
3301
3302 SDValue Mul;
3303
3304 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
3305 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
3306 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
3307 Mul = getMul24(DAG, DL, N0, N1, Size, false);
3308 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
3309 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
3310 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
3311 Mul = getMul24(DAG, DL, N0, N1, Size, true);
3312 } else {
3313 return SDValue();
3314 }
3315
3316 // We need to use sext even for MUL_U24, because MUL_U24 is used
3317 // for signed multiply of 8 and 16-bit types.
3318 return DAG.getSExtOrTrunc(Mul, DL, VT);
3319 }
3320
3321 SDValue AMDGPUTargetLowering::performMulhsCombine(SDNode *N,
3322 DAGCombinerInfo &DCI) const {
3323 EVT VT = N->getValueType(0);
3324
3325 if (!Subtarget->hasMulI24() || VT.isVector())
3326 return SDValue();
3327
3328 SelectionDAG &DAG = DCI.DAG;
3329 SDLoc DL(N);
3330
3331 SDValue N0 = N->getOperand(0);
3332 SDValue N1 = N->getOperand(1);
3333
3334 if (!isI24(N0, DAG) || !isI24(N1, DAG))
3335 return SDValue();
3336
3337 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
3338 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
3339
3340 SDValue Mulhi = DAG.getNode(AMDGPUISD::MULHI_I24, DL, MVT::i32, N0, N1);
3341 DCI.AddToWorklist(Mulhi.getNode());
3342 return DAG.getSExtOrTrunc(Mulhi, DL, VT);
3343 }
3344
3345 SDValue AMDGPUTargetLowering::performMulhuCombine(SDNode *N,
3346 DAGCombinerInfo &DCI) const {
3347 EVT VT = N->getValueType(0);
3348
3349 if (!Subtarget->hasMulU24() || VT.isVector() || VT.getSizeInBits() > 32)
3350 return SDValue();
3351
3352 SelectionDAG &DAG = DCI.DAG;
3353 SDLoc DL(N);
3354
3355 SDValue N0 = N->getOperand(0);
3356 SDValue N1 = N->getOperand(1);
3357
3358 if (!isU24(N0, DAG) || !isU24(N1, DAG))
3359 return SDValue();
3360
3361 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
3362 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
3363
3364 SDValue Mulhi = DAG.getNode(AMDGPUISD::MULHI_U24, DL, MVT::i32, N0, N1);
3365 DCI.AddToWorklist(Mulhi.getNode());
3366 return DAG.getZExtOrTrunc(Mulhi, DL, VT);
3367 }
3368
3369 SDValue AMDGPUTargetLowering::performMulLoHi24Combine(
3370 SDNode *N, DAGCombinerInfo &DCI) const {
3371 SelectionDAG &DAG = DCI.DAG;
3372
3373 // Simplify demanded bits before splitting into multiple users.
3374 if (SDValue V = simplifyI24(N, DCI))
3375 return V;
3376
3377 SDValue N0 = N->getOperand(0);
3378 SDValue N1 = N->getOperand(1);
3379
3380 bool Signed = (N->getOpcode() == AMDGPUISD::MUL_LOHI_I24);
3381
3382 unsigned MulLoOpc = Signed ? AMDGPUISD::MUL_I24 : AMDGPUISD::MUL_U24;
3383 unsigned MulHiOpc = Signed ? AMDGPUISD::MULHI_I24 : AMDGPUISD::MULHI_U24;
3384
3385 SDLoc SL(N);
3386
3387 SDValue MulLo = DAG.getNode(MulLoOpc, SL, MVT::i32, N0, N1);
3388 SDValue MulHi = DAG.getNode(MulHiOpc, SL, MVT::i32, N0, N1);
3389 return DAG.getMergeValues({ MulLo, MulHi }, SL);
3390 }
3391
3392 static bool isNegativeOne(SDValue Val) {
3393 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val))
3394 return C->isAllOnesValue();
3395 return false;
3396 }
3397
3398 SDValue AMDGPUTargetLowering::getFFBX_U32(SelectionDAG &DAG,
3399 SDValue Op,
3400 const SDLoc &DL,
3401 unsigned Opc) const {
3402 EVT VT = Op.getValueType();
3403 EVT LegalVT = getTypeToTransformTo(*DAG.getContext(), VT);
3404 if (LegalVT != MVT::i32 && (Subtarget->has16BitInsts() &&
3405 LegalVT != MVT::i16))
3406 return SDValue();
3407
3408 if (VT != MVT::i32)
3409 Op = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, Op);
3410
3411 SDValue FFBX = DAG.getNode(Opc, DL, MVT::i32, Op);
3412 if (VT != MVT::i32)
3413 FFBX = DAG.getNode(ISD::TRUNCATE, DL, VT, FFBX);
3414
3415 return FFBX;
3416 }
3417
3418 // The native instructions return -1 on 0 input. Optimize out a select that
3419 // produces -1 on 0.
3420 //
3421 // TODO: If zero is not undef, we could also do this if the output is compared
3422 // against the bitwidth.
3423 //
3424 // TODO: Should probably combine against FFBH_U32 instead of ctlz directly.
3425 SDValue AMDGPUTargetLowering::performCtlz_CttzCombine(const SDLoc &SL, SDValue Cond,
3426 SDValue LHS, SDValue RHS,
3427 DAGCombinerInfo &DCI) const {
3428 ConstantSDNode *CmpRhs = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3429 if (!CmpRhs || !CmpRhs->isNullValue())
3430 return SDValue();
3431
3432 SelectionDAG &DAG = DCI.DAG;
3433 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
3434 SDValue CmpLHS = Cond.getOperand(0);
3435
3436 unsigned Opc = isCttzOpc(RHS.getOpcode()) ? AMDGPUISD::FFBL_B32 :
3437 AMDGPUISD::FFBH_U32;
3438
3439 // select (setcc x, 0, eq), -1, (ctlz_zero_undef x) -> ffbh_u32 x
3440 // select (setcc x, 0, eq), -1, (cttz_zero_undef x) -> ffbl_u32 x
3441 if (CCOpcode == ISD::SETEQ &&
3442 (isCtlzOpc(RHS.getOpcode()) || isCttzOpc(RHS.getOpcode())) &&
3443 RHS.getOperand(0) == CmpLHS &&
3444 isNegativeOne(LHS)) {
3445 return getFFBX_U32(DAG, CmpLHS, SL, Opc);
3446 }
3447
3448 // select (setcc x, 0, ne), (ctlz_zero_undef x), -1 -> ffbh_u32 x
3449 // select (setcc x, 0, ne), (cttz_zero_undef x), -1 -> ffbl_u32 x
3450 if (CCOpcode == ISD::SETNE &&
3451 (isCtlzOpc(LHS.getOpcode()) || isCttzOpc(RHS.getOpcode())) &&
3452 LHS.getOperand(0) == CmpLHS &&
3453 isNegativeOne(RHS)) {
3454 return getFFBX_U32(DAG, CmpLHS, SL, Opc);
3455 }
3456
3457 return SDValue();
3458 }
3459
3460 static SDValue distributeOpThroughSelect(TargetLowering::DAGCombinerInfo &DCI,
3461 unsigned Op,
3462 const SDLoc &SL,
3463 SDValue Cond,
3464 SDValue N1,
3465 SDValue N2) {
3466 SelectionDAG &DAG = DCI.DAG;
3467 EVT VT = N1.getValueType();
3468
3469 SDValue NewSelect = DAG.getNode(ISD::SELECT, SL, VT, Cond,
3470 N1.getOperand(0), N2.getOperand(0));
3471 DCI.AddToWorklist(NewSelect.getNode());
3472 return DAG.getNode(Op, SL, VT, NewSelect);
3473 }
3474
3475 // Pull a free FP operation out of a select so it may fold into uses.
3476 //
3477 // select c, (fneg x), (fneg y) -> fneg (select c, x, y)
3478 // select c, (fneg x), k -> fneg (select c, x, (fneg k))
3479 //
3480 // select c, (fabs x), (fabs y) -> fabs (select c, x, y)
3481 // select c, (fabs x), +k -> fabs (select c, x, k)
3482 static SDValue foldFreeOpFromSelect(TargetLowering::DAGCombinerInfo &DCI,
3483 SDValue N) {
3484 SelectionDAG &DAG = DCI.DAG;
3485 SDValue Cond = N.getOperand(0);
3486 SDValue LHS = N.getOperand(1);
3487 SDValue RHS = N.getOperand(2);
3488
3489 EVT VT = N.getValueType();
3490 if ((LHS.getOpcode() == ISD::FABS && RHS.getOpcode() == ISD::FABS) ||
3491 (LHS.getOpcode() == ISD::FNEG && RHS.getOpcode() == ISD::FNEG)) {
3492 return distributeOpThroughSelect(DCI, LHS.getOpcode(),
3493 SDLoc(N), Cond, LHS, RHS);
3494 }
3495
3496 bool Inv = false;
3497 if (RHS.getOpcode() == ISD::FABS || RHS.getOpcode() == ISD::FNEG) {
3498 std::swap(LHS, RHS);
3499 Inv = true;
3500 }
3501
3502 // TODO: Support vector constants.
3503 ConstantFPSDNode *CRHS = dyn_cast<ConstantFPSDNode>(RHS);
3504 if ((LHS.getOpcode() == ISD::FNEG || LHS.getOpcode() == ISD::FABS) && CRHS) {
3505 SDLoc SL(N);
3506 // If one side is an fneg/fabs and the other is a constant, we can push the
3507 // fneg/fabs down. If it's an fabs, the constant needs to be non-negative.
3508 SDValue NewLHS = LHS.getOperand(0);
3509 SDValue NewRHS = RHS;
3510
3511 // Careful: if the neg can be folded up, don't try to pull it back down.
3512 bool ShouldFoldNeg = true;
3513
3514 if (NewLHS.hasOneUse()) {
3515 unsigned Opc = NewLHS.getOpcode();
3516 if (LHS.getOpcode() == ISD::FNEG && fnegFoldsIntoOp(Opc))
3517 ShouldFoldNeg = false;
3518 if (LHS.getOpcode() == ISD::FABS && Opc == ISD::FMUL)
3519 ShouldFoldNeg = false;
3520 }
3521
3522 if (ShouldFoldNeg) {
3523 if (LHS.getOpcode() == ISD::FNEG)
3524 NewRHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
3525 else if (CRHS->isNegative())
3526 return SDValue();
3527
3528 if (Inv)
3529 std::swap(NewLHS, NewRHS);
3530
3531 SDValue NewSelect = DAG.getNode(ISD::SELECT, SL, VT,
3532 Cond, NewLHS, NewRHS);
3533 DCI.AddToWorklist(NewSelect.getNode());
3534 return DAG.getNode(LHS.getOpcode(), SL, VT, NewSelect);
3535 }
3536 }
3537
3538 return SDValue();
3539 }
3540
3541
3542 SDValue AMDGPUTargetLowering::performSelectCombine(SDNode *N,
3543 DAGCombinerInfo &DCI) const {
3544 if (SDValue Folded = foldFreeOpFromSelect(DCI, SDValue(N, 0)))
3545 return Folded;
3546
3547 SDValue Cond = N->getOperand(0);
3548 if (Cond.getOpcode() != ISD::SETCC)
3549 return SDValue();
3550
3551 EVT VT = N->getValueType(0);
3552 SDValue LHS = Cond.getOperand(0);
3553 SDValue RHS = Cond.getOperand(1);
3554 SDValue CC = Cond.getOperand(2);
3555
3556 SDValue True = N->getOperand(1);
3557 SDValue False = N->getOperand(2);
3558
3559 if (Cond.hasOneUse()) { // TODO: Look for multiple select uses.
3560 SelectionDAG &DAG = DCI.DAG;
3561 if (DAG.isConstantValueOfAnyType(True) &&
3562 !DAG.isConstantValueOfAnyType(False)) {
3563 // Swap cmp + select pair to move constant to false input.
3564 // This will allow using VOPC cndmasks more often.
3565 // select (setcc x, y), k, x -> select (setccinv x, y), x, k
3566
3567 SDLoc SL(N);
3568 ISD::CondCode NewCC = getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3569 LHS.getValueType().isInteger());
3570
3571 SDValue NewCond = DAG.getSetCC(SL, Cond.getValueType(), LHS, RHS, NewCC);
3572 return DAG.getNode(ISD::SELECT, SL, VT, NewCond, False, True);
3573 }
3574
3575 if (VT == MVT::f32 && Subtarget->hasFminFmaxLegacy()) {
3576 SDValue MinMax
3577 = combineFMinMaxLegacy(SDLoc(N), VT, LHS, RHS, True, False, CC, DCI);
3578 // Revisit this node so we can catch min3/max3/med3 patterns.
3579 //DCI.AddToWorklist(MinMax.getNode());
3580 return MinMax;
3581 }
3582 }
3583
3584 // There's no reason to not do this if the condition has other uses.
3585 return performCtlz_CttzCombine(SDLoc(N), Cond, True, False, DCI);
3586 }
3587
3588 static bool isInv2Pi(const APFloat &APF) {
3589 static const APFloat KF16(APFloat::IEEEhalf(), APInt(16, 0x3118));
3590 static const APFloat KF32(APFloat::IEEEsingle(), APInt(32, 0x3e22f983));
3591 static const APFloat KF64(APFloat::IEEEdouble(), APInt(64, 0x3fc45f306dc9c882));
3592
3593 return APF.bitwiseIsEqual(KF16) ||
3594 APF.bitwiseIsEqual(KF32) ||
3595 APF.bitwiseIsEqual(KF64);
3596 }
3597
3598 // 0 and 1.0 / (0.5 * pi) do not have inline immmediates, so there is an
3599 // additional cost to negate them.
3600 bool AMDGPUTargetLowering::isConstantCostlierToNegate(SDValue N) const {
3601 if (const ConstantFPSDNode *C = isConstOrConstSplatFP(N)) {
3602 if (C->isZero() && !C->isNegative())
3603 return true;
3604
3605 if (Subtarget->hasInv2PiInlineImm() && isInv2Pi(C->getValueAPF()))
3606 return true;
3607 }
3608
3609 return false;
3610 }
3611
3612 static unsigned inverseMinMax(unsigned Opc) {
3613 switch (Opc) {
3614 case ISD::FMAXNUM:
3615 return ISD::FMINNUM;
3616 case ISD::FMINNUM:
3617 return ISD::FMAXNUM;
3618 case ISD::FMAXNUM_IEEE:
3619 return ISD::FMINNUM_IEEE;
3620 case ISD::FMINNUM_IEEE:
3621 return ISD::FMAXNUM_IEEE;
3622 case AMDGPUISD::FMAX_LEGACY:
3623 return AMDGPUISD::FMIN_LEGACY;
3624 case AMDGPUISD::FMIN_LEGACY:
3625 return AMDGPUISD::FMAX_LEGACY;
3626 default:
3627 llvm_unreachable("invalid min/max opcode");
3628 }
3629 }
3630
3631 SDValue AMDGPUTargetLowering::performFNegCombine(SDNode *N,
3632 DAGCombinerInfo &DCI) const {
3633 SelectionDAG &DAG = DCI.DAG;
3634 SDValue N0 = N->getOperand(0);
3635 EVT VT = N->getValueType(0);
3636
3637 unsigned Opc = N0.getOpcode();
3638
3639 // If the input has multiple uses and we can either fold the negate down, or
3640 // the other uses cannot, give up. This both prevents unprofitable
3641 // transformations and infinite loops: we won't repeatedly try to fold around
3642 // a negate that has no 'good' form.
3643 if (N0.hasOneUse()) {
3644 // This may be able to fold into the source, but at a code size cost. Don't
3645 // fold if the fold into the user is free.
3646 if (allUsesHaveSourceMods(N, 0))
3647 return SDValue();
3648 } else {
3649 if (fnegFoldsIntoOp(Opc) &&
3650 (allUsesHaveSourceMods(N) || !allUsesHaveSourceMods(N0.getNode())))
3651 return SDValue();
3652 }
3653
3654 SDLoc SL(N);
3655 switch (Opc) {
3656 case ISD::FADD: {
3657 if (!mayIgnoreSignedZero(N0))
3658 return SDValue();
3659
3660 // (fneg (fadd x, y)) -> (fadd (fneg x), (fneg y))
3661 SDValue LHS = N0.getOperand(0);
3662 SDValue RHS = N0.getOperand(1);
3663
3664 if (LHS.getOpcode() != ISD::FNEG)
3665 LHS = DAG.getNode(ISD::FNEG, SL, VT, LHS);
3666 else
3667 LHS = LHS.getOperand(0);
3668
3669 if (RHS.getOpcode() != ISD::FNEG)
3670 RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
3671 else
3672 RHS = RHS.getOperand(0);
3673
3674 SDValue Res = DAG.getNode(ISD::FADD, SL, VT, LHS, RHS, N0->getFlags());
3675 if (Res.getOpcode() != ISD::FADD)
3676 return SDValue(); // Op got folded away.
3677 if (!N0.hasOneUse())
3678 DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
3679 return Res;
3680 }
3681 case ISD::FMUL:
3682 case AMDGPUISD::FMUL_LEGACY: {
3683 // (fneg (fmul x, y)) -> (fmul x, (fneg y))
3684 // (fneg (fmul_legacy x, y)) -> (fmul_legacy x, (fneg y))
3685 SDValue LHS = N0.getOperand(0);
3686 SDValue RHS = N0.getOperand(1);
3687
3688 if (LHS.getOpcode() == ISD::FNEG)
3689 LHS = LHS.getOperand(0);
3690 else if (RHS.getOpcode() == ISD::FNEG)
3691 RHS = RHS.getOperand(0);
3692 else
3693 RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
3694
3695 SDValue Res = DAG.getNode(Opc, SL, VT, LHS, RHS, N0->getFlags());
3696 if (Res.getOpcode() != Opc)
3697 return SDValue(); // Op got folded away.
3698 if (!N0.hasOneUse())
3699 DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
3700 return Res;
3701 }
3702 case ISD::FMA:
3703 case ISD::FMAD: {
3704 if (!mayIgnoreSignedZero(N0))
3705 return SDValue();
3706
3707 // (fneg (fma x, y, z)) -> (fma x, (fneg y), (fneg z))
3708 SDValue LHS = N0.getOperand(0);
3709 SDValue MHS = N0.getOperand(1);
3710 SDValue RHS = N0.getOperand(2);
3711
3712 if (LHS.getOpcode() == ISD::FNEG)
3713 LHS = LHS.getOperand(0);
3714 else if (MHS.getOpcode() == ISD::FNEG)
3715 MHS = MHS.getOperand(0);
3716 else
3717 MHS = DAG.getNode(ISD::FNEG, SL, VT, MHS);
3718
3719 if (RHS.getOpcode() != ISD::FNEG)
3720 RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
3721 else
3722 RHS = RHS.getOperand(0);
3723
3724 SDValue Res = DAG.getNode(Opc, SL, VT, LHS, MHS, RHS);
3725 if (Res.getOpcode() != Opc)
3726 return SDValue(); // Op got folded away.
3727 if (!N0.hasOneUse())
3728 DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
3729 return Res;
3730 }
3731 case ISD::FMAXNUM:
3732 case ISD::FMINNUM:
3733 case ISD::FMAXNUM_IEEE:
3734 case ISD::FMINNUM_IEEE:
3735 case AMDGPUISD::FMAX_LEGACY:
3736 case AMDGPUISD::FMIN_LEGACY: {
3737 // fneg (fmaxnum x, y) -> fminnum (fneg x), (fneg y)
3738 // fneg (fminnum x, y) -> fmaxnum (fneg x), (fneg y)
3739 // fneg (fmax_legacy x, y) -> fmin_legacy (fneg x), (fneg y)
3740 // fneg (fmin_legacy x, y) -> fmax_legacy (fneg x), (fneg y)
3741
3742 SDValue LHS = N0.getOperand(0);
3743 SDValue RHS = N0.getOperand(1);
3744
3745 // 0 doesn't have a negated inline immediate.
3746 // TODO: This constant check should be generalized to other operations.
3747 if (isConstantCostlierToNegate(RHS))
3748 return SDValue();
3749
3750 SDValue NegLHS = DAG.getNode(ISD::FNEG, SL, VT, LHS);
3751 SDValue NegRHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
3752 unsigned Opposite = inverseMinMax(Opc);
3753
3754 SDValue Res = DAG.getNode(Opposite, SL, VT, NegLHS, NegRHS, N0->getFlags());
3755 if (Res.getOpcode() != Opposite)
3756 return SDValue(); // Op got folded away.
3757 if (!N0.hasOneUse())
3758 DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
3759 return Res;
3760 }
3761 case AMDGPUISD::FMED3: {
3762 SDValue Ops[3];
3763 for (unsigned I = 0; I < 3; ++I)
3764 Ops[I] = DAG.getNode(ISD::FNEG, SL, VT, N0->getOperand(I), N0->getFlags());
3765
3766 SDValue Res = DAG.getNode(AMDGPUISD::FMED3, SL, VT, Ops, N0->getFlags());
3767 if (Res.getOpcode() != AMDGPUISD::FMED3)
3768 return SDValue(); // Op got folded away.
3769 if (!N0.hasOneUse())
3770 DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
3771 return Res;
3772 }
3773 case ISD::FP_EXTEND:
3774 case ISD::FTRUNC:
3775 case ISD::FRINT:
3776 case ISD::FNEARBYINT: // XXX - Should fround be handled?
3777 case ISD::FSIN:
3778 case ISD::FCANONICALIZE:
3779 case AMDGPUISD::RCP:
3780 case AMDGPUISD::RCP_LEGACY:
3781 case AMDGPUISD::RCP_IFLAG:
3782 case AMDGPUISD::SIN_HW: {
3783 SDValue CvtSrc = N0.getOperand(0);
3784 if (CvtSrc.getOpcode() == ISD::FNEG) {
3785 // (fneg (fp_extend (fneg x))) -> (fp_extend x)
3786 // (fneg (rcp (fneg x))) -> (rcp x)
3787 return DAG.getNode(Opc, SL, VT, CvtSrc.getOperand(0));
3788 }
3789
3790 if (!N0.hasOneUse())
3791 return SDValue();
3792
3793 // (fneg (fp_extend x)) -> (fp_extend (fneg x))
3794 // (fneg (rcp x)) -> (rcp (fneg x))
3795 SDValue Neg = DAG.getNode(ISD::FNEG, SL, CvtSrc.getValueType(), CvtSrc);
3796 return DAG.getNode(Opc, SL, VT, Neg, N0->getFlags());
3797 }
3798 case ISD::FP_ROUND: {
3799 SDValue CvtSrc = N0.getOperand(0);
3800
3801 if (CvtSrc.getOpcode() == ISD::FNEG) {
3802 // (fneg (fp_round (fneg x))) -> (fp_round x)
3803 return DAG.getNode(ISD::FP_ROUND, SL, VT,
3804 CvtSrc.getOperand(0), N0.getOperand(1));
3805 }
3806
3807 if (!N0.hasOneUse())
3808 return SDValue();
3809
3810 // (fneg (fp_round x)) -> (fp_round (fneg x))
3811 SDValue Neg = DAG.getNode(ISD::FNEG, SL, CvtSrc.getValueType(), CvtSrc);
3812 return DAG.getNode(ISD::FP_ROUND, SL, VT, Neg, N0.getOperand(1));
3813 }
3814 case ISD::FP16_TO_FP: {
3815 // v_cvt_f32_f16 supports source modifiers on pre-VI targets without legal
3816 // f16, but legalization of f16 fneg ends up pulling it out of the source.
3817 // Put the fneg back as a legal source operation that can be matched later.
3818 SDLoc SL(N);
3819
3820 SDValue Src = N0.getOperand(0);
3821 EVT SrcVT = Src.getValueType();
3822
3823 // fneg (fp16_to_fp x) -> fp16_to_fp (xor x, 0x8000)
3824 SDValue IntFNeg = DAG.getNode(ISD::XOR, SL, SrcVT, Src,
3825 DAG.getConstant(0x8000, SL, SrcVT));
3826 return DAG.getNode(ISD::FP16_TO_FP, SL, N->getValueType(0), IntFNeg);
3827 }
3828 default:
3829 return SDValue();
3830 }
3831 }
3832
3833 SDValue AMDGPUTargetLowering::performFAbsCombine(SDNode *N,
3834 DAGCombinerInfo &DCI) const {
3835 SelectionDAG &DAG = DCI.DAG;
3836 SDValue N0 = N->getOperand(0);
3837
3838 if (!N0.hasOneUse())
3839 return SDValue();
3840
3841 switch (N0.getOpcode()) {
3842 case ISD::FP16_TO_FP: {
3843 assert(!Subtarget->has16BitInsts() && "should only see if f16 is illegal");
3844 SDLoc SL(N);
3845 SDValue Src = N0.getOperand(0);
3846 EVT SrcVT = Src.getValueType();
3847
3848 // fabs (fp16_to_fp x) -> fp16_to_fp (and x, 0x7fff)
3849 SDValue IntFAbs = DAG.getNode(ISD::AND, SL, SrcVT, Src,
3850 DAG.getConstant(0x7fff, SL, SrcVT));
3851 return DAG.getNode(ISD::FP16_TO_FP, SL, N->getValueType(0), IntFAbs);
3852 }
3853 default:
3854 return SDValue();
3855 }
3856 }
3857
3858 SDValue AMDGPUTargetLowering::performRcpCombine(SDNode *N,
3859 DAGCombinerInfo &DCI) const {
3860 const auto *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0));
3861 if (!CFP)
3862 return SDValue();
3863
3864 // XXX - Should this flush denormals?
3865 const APFloat &Val = CFP->getValueAPF();
3866 APFloat One(Val.getSemantics(), "1.0");
3867 return DCI.DAG.getConstantFP(One / Val, SDLoc(N), N->getValueType(0));
3868 }
3869
3870 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
3871 DAGCombinerInfo &DCI) const {
3872 SelectionDAG &DAG = DCI.DAG;
3873 SDLoc DL(N);
3874
3875 switch(N->getOpcode()) {
3876 default:
3877 break;
3878 case ISD::BITCAST: {
3879 EVT DestVT = N->getValueType(0);
3880
3881 // Push casts through vector builds. This helps avoid emitting a large
3882 // number of copies when materializing floating point vector constants.
3883 //
3884 // vNt1 bitcast (vNt0 (build_vector t0:x, t0:y)) =>
3885 // vnt1 = build_vector (t1 (bitcast t0:x)), (t1 (bitcast t0:y))
3886 if (DestVT.isVector()) {
3887 SDValue Src = N->getOperand(0);
3888 if (Src.getOpcode() == ISD::BUILD_VECTOR) {
3889 EVT SrcVT = Src.getValueType();
3890 unsigned NElts = DestVT.getVectorNumElements();
3891
3892 if (SrcVT.getVectorNumElements() == NElts) {
3893 EVT DestEltVT = DestVT.getVectorElementType();
3894
3895 SmallVector<SDValue, 8> CastedElts;
3896 SDLoc SL(N);
3897 for (unsigned I = 0, E = SrcVT.getVectorNumElements(); I != E; ++I) {
3898 SDValue Elt = Src.getOperand(I);
3899 CastedElts.push_back(DAG.getNode(ISD::BITCAST, DL, DestEltVT, Elt));
3900 }
3901
3902 return DAG.getBuildVector(DestVT, SL, CastedElts);
3903 }
3904 }
3905 }
3906
3907 if (DestVT.getSizeInBits() != 64 && !DestVT.isVector())
3908 break;
3909
3910 // Fold bitcasts of constants.
3911 //
3912 // v2i32 (bitcast i64:k) -> build_vector lo_32(k), hi_32(k)
3913 // TODO: Generalize and move to DAGCombiner
3914 SDValue Src = N->getOperand(0);
3915 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Src)) {
3916 if (Src.getValueType() == MVT::i64) {
3917 SDLoc SL(N);
3918 uint64_t CVal = C->getZExtValue();
3919 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
3920 DAG.getConstant(Lo_32(CVal), SL, MVT::i32),
3921 DAG.getConstant(Hi_32(CVal), SL, MVT::i32));
3922 return DAG.getNode(ISD::BITCAST, SL, DestVT, BV);
3923 }
3924 }
3925
3926 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Src)) {
3927 const APInt &Val = C->getValueAPF().bitcastToAPInt();
3928 SDLoc SL(N);
3929 uint64_t CVal = Val.getZExtValue();
3930 SDValue Vec = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
3931 DAG.getConstant(Lo_32(CVal), SL, MVT::i32),
3932 DAG.getConstant(Hi_32(CVal), SL, MVT::i32));
3933
3934 return DAG.getNode(ISD::BITCAST, SL, DestVT, Vec);
3935 }
3936
3937 break;
3938 }
3939 case ISD::SHL: {
3940 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
3941 break;
3942
3943 return performShlCombine(N, DCI);
3944 }
3945 case ISD::SRL: {
3946 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
3947 break;
3948
3949 return performSrlCombine(N, DCI);
3950 }
3951 case ISD::SRA: {
3952 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
3953 break;
3954
3955 return performSraCombine(N, DCI);
3956 }
3957 case ISD::TRUNCATE:
3958 return performTruncateCombine(N, DCI);
3959 case ISD::MUL:
3960 return performMulCombine(N, DCI);
3961 case ISD::MULHS:
3962 return performMulhsCombine(N, DCI);
3963 case ISD::MULHU:
3964 return performMulhuCombine(N, DCI);
3965 case AMDGPUISD::MUL_I24:
3966 case AMDGPUISD::MUL_U24:
3967 case AMDGPUISD::MULHI_I24:
3968 case AMDGPUISD::MULHI_U24: {
3969 if (SDValue V = simplifyI24(N, DCI))
3970 return V;
3971 return SDValue();
3972 }
3973 case AMDGPUISD::MUL_LOHI_I24:
3974 case AMDGPUISD::MUL_LOHI_U24:
3975 return performMulLoHi24Combine(N, DCI);
3976 case ISD::SELECT:
3977 return performSelectCombine(N, DCI);
3978 case ISD::FNEG:
3979 return performFNegCombine(N, DCI);
3980 case ISD::FABS:
3981 return performFAbsCombine(N, DCI);
3982 case AMDGPUISD::BFE_I32:
3983 case AMDGPUISD::BFE_U32: {
3984 assert(!N->getValueType(0).isVector() &&
3985 "Vector handling of BFE not implemented");
3986 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
3987 if (!Width)
3988 break;
3989
3990 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
3991 if (WidthVal == 0)
3992 return DAG.getConstant(0, DL, MVT::i32);
3993
3994 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
3995 if (!Offset)
3996 break;
3997
3998 SDValue BitsFrom = N->getOperand(0);
3999 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
4000
4001 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
4002
4003 if (OffsetVal == 0) {
4004 // This is already sign / zero extended, so try to fold away extra BFEs.
4005 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
4006
4007 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
4008 if (OpSignBits >= SignBits)
4009 return BitsFrom;
4010
4011 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
4012 if (Signed) {
4013 // This is a sign_extend_inreg. Replace it to take advantage of existing
4014 // DAG Combines. If not eliminated, we will match back to BFE during
4015 // selection.
4016
4017 // TODO: The sext_inreg of extended types ends, although we can could
4018 // handle them in a single BFE.
4019 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
4020 DAG.getValueType(SmallVT));
4021 }
4022
4023 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
4024 }
4025
4026 if (ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(BitsFrom)) {
4027 if (Signed) {
4028 return constantFoldBFE<int32_t>(DAG,
4029 CVal->getSExtValue(),
4030 OffsetVal,
4031 WidthVal,
4032 DL);
4033 }
4034
4035 return constantFoldBFE<uint32_t>(DAG,
4036 CVal->getZExtValue(),
4037 OffsetVal,
4038 WidthVal,
4039 DL);
4040 }
4041
4042 if ((OffsetVal + WidthVal) >= 32 &&
4043 !(Subtarget->hasSDWA() && OffsetVal == 16 && WidthVal == 16)) {
4044 SDValue ShiftVal = DAG.getConstant(OffsetVal, DL, MVT::i32);
4045 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
4046 BitsFrom, ShiftVal);
4047 }
4048
4049 if (BitsFrom.hasOneUse()) {
4050 APInt Demanded = APInt::getBitsSet(32,
4051 OffsetVal,
4052 OffsetVal + WidthVal);
4053
4054 KnownBits Known;
4055 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
4056 !DCI.isBeforeLegalizeOps());
4057 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4058 if (TLI.ShrinkDemandedConstant(BitsFrom, Demanded, TLO) ||
4059 TLI.SimplifyDemandedBits(BitsFrom, Demanded, Known, TLO)) {
4060 DCI.CommitTargetLoweringOpt(TLO);
4061 }
4062 }
4063
4064 break;
4065 }
4066 case ISD::LOAD:
4067 return performLoadCombine(N, DCI);
4068 case ISD::STORE:
4069 return performStoreCombine(N, DCI);
4070 case AMDGPUISD::RCP:
4071 case AMDGPUISD::RCP_IFLAG:
4072 return performRcpCombine(N, DCI);
4073 case ISD::AssertZext:
4074 case ISD::AssertSext:
4075 return performAssertSZExtCombine(N, DCI);
4076 case ISD::INTRINSIC_WO_CHAIN:
4077 return performIntrinsicWOChainCombine(N, DCI);
4078 }
4079 return SDValue();
4080 }
4081
4082 //===----------------------------------------------------------------------===//
4083 // Helper functions
4084 //===----------------------------------------------------------------------===//
4085
4086 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
4087 const TargetRegisterClass *RC,
4088 unsigned Reg, EVT VT,
4089 const SDLoc &SL,
4090 bool RawReg) const {
4091 MachineFunction &MF = DAG.getMachineFunction();
4092 MachineRegisterInfo &MRI = MF.getRegInfo();
4093 unsigned VReg;
4094
4095 if (!MRI.isLiveIn(Reg)) {
4096 VReg = MRI.createVirtualRegister(RC);
4097 MRI.addLiveIn(Reg, VReg);
4098 } else {
4099 VReg = MRI.getLiveInVirtReg(Reg);
4100 }
4101
4102 if (RawReg)
4103 return DAG.getRegister(VReg, VT);
4104
4105 return DAG.getCopyFromReg(DAG.getEntryNode(), SL, VReg, VT);
4106 }
4107
4108 // This may be called multiple times, and nothing prevents creating multiple
4109 // objects at the same offset. See if we already defined this object.
4110 static int getOrCreateFixedStackObject(MachineFrameInfo &MFI, unsigned Size,
4111 int64_t Offset) {
4112 for (int I = MFI.getObjectIndexBegin(); I < 0; ++I) {
4113 if (MFI.getObjectOffset(I) == Offset) {
4114 assert(MFI.getObjectSize(I) == Size);
4115 return I;
4116 }
4117 }
4118
4119 return MFI.CreateFixedObject(Size, Offset, true);
4120 }
4121
4122 SDValue AMDGPUTargetLowering::loadStackInputValue(SelectionDAG &DAG,
4123 EVT VT,
4124 const SDLoc &SL,
4125 int64_t Offset) const {
4126 MachineFunction &MF = DAG.getMachineFunction();
4127 MachineFrameInfo &MFI = MF.getFrameInfo();
4128 int FI = getOrCreateFixedStackObject(MFI, VT.getStoreSize(), Offset);
4129
4130 auto SrcPtrInfo = MachinePointerInfo::getStack(MF, Offset);
4131 SDValue Ptr = DAG.getFrameIndex(FI, MVT::i32);
4132
4133 return DAG.getLoad(VT, SL, DAG.getEntryNode(), Ptr, SrcPtrInfo, 4,
4134 MachineMemOperand::MODereferenceable |
4135 MachineMemOperand::MOInvariant);
4136 }
4137
4138 SDValue AMDGPUTargetLowering::storeStackInputValue(SelectionDAG &DAG,
4139 const SDLoc &SL,
4140 SDValue Chain,
4141 SDValue ArgVal,
4142 int64_t Offset) const {
4143 MachineFunction &MF = DAG.getMachineFunction();
4144 MachinePointerInfo DstInfo = MachinePointerInfo::getStack(MF, Offset);
4145
4146 SDValue Ptr = DAG.getConstant(Offset, SL, MVT::i32);
4147 SDValue Store = DAG.getStore(Chain, SL, ArgVal, Ptr, DstInfo, 4,
4148 MachineMemOperand::MODereferenceable);
4149 return Store;
4150 }
4151
4152 SDValue AMDGPUTargetLowering::loadInputValue(SelectionDAG &DAG,
4153 const TargetRegisterClass *RC,
4154 EVT VT, const SDLoc &SL,
4155 const ArgDescriptor &Arg) const {
4156 assert(Arg && "Attempting to load missing argument");
4157
4158 SDValue V = Arg.isRegister() ?
4159 CreateLiveInRegister(DAG, RC, Arg.getRegister(), VT, SL) :
4160 loadStackInputValue(DAG, VT, SL, Arg.getStackOffset());
4161
4162 if (!Arg.isMasked())
4163 return V;
4164
4165 unsigned Mask = Arg.getMask();
4166 unsigned Shift = countTrailingZeros<unsigned>(Mask);
4167 V = DAG.getNode(ISD::SRL, SL, VT, V,
4168 DAG.getShiftAmountConstant(Shift, VT, SL));
4169 return DAG.getNode(ISD::AND, SL, VT, V,
4170 DAG.getConstant(Mask >> Shift, SL, VT));
4171 }
4172
4173 uint32_t AMDGPUTargetLowering::getImplicitParameterOffset(
4174 const MachineFunction &MF, const ImplicitParameter Param) const {
4175 const AMDGPUMachineFunction *MFI = MF.getInfo<AMDGPUMachineFunction>();
4176 const AMDGPUSubtarget &ST =
4177 AMDGPUSubtarget::get(getTargetMachine(), MF.getFunction());
4178 unsigned ExplicitArgOffset = ST.getExplicitKernelArgOffset(MF.getFunction());
4179 const Align Alignment = ST.getAlignmentForImplicitArgPtr();
4180 uint64_t ArgOffset = alignTo(MFI->getExplicitKernArgSize(), Alignment) +
4181 ExplicitArgOffset;
4182 switch (Param) {
4183 case GRID_DIM:
4184 return ArgOffset;
4185 case GRID_OFFSET:
4186 return ArgOffset + 4;
4187 }
4188 llvm_unreachable("unexpected implicit parameter type");
4189 }
4190
4191 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
4192
4193 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
4194 switch ((AMDGPUISD::NodeType)Opcode) {
4195 case AMDGPUISD::FIRST_NUMBER: break;
4196 // AMDIL DAG nodes
4197 NODE_NAME_CASE(UMUL);
4198 NODE_NAME_CASE(BRANCH_COND);
4199
4200 // AMDGPU DAG nodes
4201 NODE_NAME_CASE(IF)
4202 NODE_NAME_CASE(ELSE)
4203 NODE_NAME_CASE(LOOP)
4204 NODE_NAME_CASE(CALL)
4205 NODE_NAME_CASE(TC_RETURN)
4206 NODE_NAME_CASE(TRAP)
4207 NODE_NAME_CASE(RET_FLAG)
4208 NODE_NAME_CASE(RETURN_TO_EPILOG)
4209 NODE_NAME_CASE(ENDPGM)
4210 NODE_NAME_CASE(DWORDADDR)
4211 NODE_NAME_CASE(FRACT)
4212 NODE_NAME_CASE(SETCC)
4213 NODE_NAME_CASE(SETREG)
4214 NODE_NAME_CASE(DENORM_MODE)
4215 NODE_NAME_CASE(FMA_W_CHAIN)
4216 NODE_NAME_CASE(FMUL_W_CHAIN)
4217 NODE_NAME_CASE(CLAMP)
4218 NODE_NAME_CASE(COS_HW)
4219 NODE_NAME_CASE(SIN_HW)
4220 NODE_NAME_CASE(FMAX_LEGACY)
4221 NODE_NAME_CASE(FMIN_LEGACY)
4222 NODE_NAME_CASE(FMAX3)
4223 NODE_NAME_CASE(SMAX3)
4224 NODE_NAME_CASE(UMAX3)
4225 NODE_NAME_CASE(FMIN3)
4226 NODE_NAME_CASE(SMIN3)
4227 NODE_NAME_CASE(UMIN3)
4228 NODE_NAME_CASE(FMED3)
4229 NODE_NAME_CASE(SMED3)
4230 NODE_NAME_CASE(UMED3)
4231 NODE_NAME_CASE(FDOT2)
4232 NODE_NAME_CASE(URECIP)
4233 NODE_NAME_CASE(DIV_SCALE)
4234 NODE_NAME_CASE(DIV_FMAS)
4235 NODE_NAME_CASE(DIV_FIXUP)
4236 NODE_NAME_CASE(FMAD_FTZ)
4237 NODE_NAME_CASE(TRIG_PREOP)
4238 NODE_NAME_CASE(RCP)
4239 NODE_NAME_CASE(RSQ)
4240 NODE_NAME_CASE(RCP_LEGACY)
4241 NODE_NAME_CASE(RSQ_LEGACY)
4242 NODE_NAME_CASE(RCP_IFLAG)
4243 NODE_NAME_CASE(FMUL_LEGACY)
4244 NODE_NAME_CASE(RSQ_CLAMP)
4245 NODE_NAME_CASE(LDEXP)
4246 NODE_NAME_CASE(FP_CLASS)
4247 NODE_NAME_CASE(DOT4)
4248 NODE_NAME_CASE(CARRY)
4249 NODE_NAME_CASE(BORROW)
4250 NODE_NAME_CASE(BFE_U32)
4251 NODE_NAME_CASE(BFE_I32)
4252 NODE_NAME_CASE(BFI)
4253 NODE_NAME_CASE(BFM)
4254 NODE_NAME_CASE(FFBH_U32)
4255 NODE_NAME_CASE(FFBH_I32)
4256 NODE_NAME_CASE(FFBL_B32)
4257 NODE_NAME_CASE(MUL_U24)
4258 NODE_NAME_CASE(MUL_I24)
4259 NODE_NAME_CASE(MULHI_U24)
4260 NODE_NAME_CASE(MULHI_I24)
4261 NODE_NAME_CASE(MUL_LOHI_U24)
4262 NODE_NAME_CASE(MUL_LOHI_I24)
4263 NODE_NAME_CASE(MAD_U24)
4264 NODE_NAME_CASE(MAD_I24)
4265 NODE_NAME_CASE(MAD_I64_I32)
4266 NODE_NAME_CASE(MAD_U64_U32)
4267 NODE_NAME_CASE(PERM)
4268 NODE_NAME_CASE(TEXTURE_FETCH)
4269 NODE_NAME_CASE(EXPORT)
4270 NODE_NAME_CASE(EXPORT_DONE)
4271 NODE_NAME_CASE(R600_EXPORT)
4272 NODE_NAME_CASE(CONST_ADDRESS)
4273 NODE_NAME_CASE(REGISTER_LOAD)
4274 NODE_NAME_CASE(REGISTER_STORE)
4275 NODE_NAME_CASE(SAMPLE)
4276 NODE_NAME_CASE(SAMPLEB)
4277 NODE_NAME_CASE(SAMPLED)
4278 NODE_NAME_CASE(SAMPLEL)
4279 NODE_NAME_CASE(CVT_F32_UBYTE0)
4280 NODE_NAME_CASE(CVT_F32_UBYTE1)
4281 NODE_NAME_CASE(CVT_F32_UBYTE2)
4282 NODE_NAME_CASE(CVT_F32_UBYTE3)
4283 NODE_NAME_CASE(CVT_PKRTZ_F16_F32)
4284 NODE_NAME_CASE(CVT_PKNORM_I16_F32)
4285 NODE_NAME_CASE(CVT_PKNORM_U16_F32)
4286 NODE_NAME_CASE(CVT_PK_I16_I32)
4287 NODE_NAME_CASE(CVT_PK_U16_U32)
4288 NODE_NAME_CASE(FP_TO_FP16)
4289 NODE_NAME_CASE(FP16_ZEXT)
4290 NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
4291 NODE_NAME_CASE(CONST_DATA_PTR)
4292 NODE_NAME_CASE(PC_ADD_REL_OFFSET)
4293 NODE_NAME_CASE(LDS)
4294 NODE_NAME_CASE(KILL)
4295 NODE_NAME_CASE(DUMMY_CHAIN)
4296 case AMDGPUISD::FIRST_MEM_OPCODE_NUMBER: break;
4297 NODE_NAME_CASE(INTERP_P1LL_F16)
4298 NODE_NAME_CASE(INTERP_P1LV_F16)
4299 NODE_NAME_CASE(INTERP_P2_F16)
4300 NODE_NAME_CASE(LOAD_D16_HI)
4301 NODE_NAME_CASE(LOAD_D16_LO)
4302 NODE_NAME_CASE(LOAD_D16_HI_I8)
4303 NODE_NAME_CASE(LOAD_D16_HI_U8)
4304 NODE_NAME_CASE(LOAD_D16_LO_I8)
4305 NODE_NAME_CASE(LOAD_D16_LO_U8)
4306 NODE_NAME_CASE(STORE_MSKOR)
4307 NODE_NAME_CASE(LOAD_CONSTANT)
4308 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
4309 NODE_NAME_CASE(TBUFFER_STORE_FORMAT_D16)
4310 NODE_NAME_CASE(TBUFFER_LOAD_FORMAT)
4311 NODE_NAME_CASE(TBUFFER_LOAD_FORMAT_D16)
4312 NODE_NAME_CASE(DS_ORDERED_COUNT)
4313 NODE_NAME_CASE(ATOMIC_CMP_SWAP)
4314 NODE_NAME_CASE(ATOMIC_INC)
4315 NODE_NAME_CASE(ATOMIC_DEC)
4316 NODE_NAME_CASE(ATOMIC_LOAD_FMIN)
4317 NODE_NAME_CASE(ATOMIC_LOAD_FMAX)
4318 NODE_NAME_CASE(BUFFER_LOAD)
4319 NODE_NAME_CASE(BUFFER_LOAD_UBYTE)
4320 NODE_NAME_CASE(BUFFER_LOAD_USHORT)
4321 NODE_NAME_CASE(BUFFER_LOAD_BYTE)
4322 NODE_NAME_CASE(BUFFER_LOAD_SHORT)
4323 NODE_NAME_CASE(BUFFER_LOAD_FORMAT)
4324 NODE_NAME_CASE(BUFFER_LOAD_FORMAT_D16)
4325 NODE_NAME_CASE(SBUFFER_LOAD)
4326 NODE_NAME_CASE(BUFFER_STORE)
4327 NODE_NAME_CASE(BUFFER_STORE_BYTE)
4328 NODE_NAME_CASE(BUFFER_STORE_SHORT)
4329 NODE_NAME_CASE(BUFFER_STORE_FORMAT)
4330 NODE_NAME_CASE(BUFFER_STORE_FORMAT_D16)
4331 NODE_NAME_CASE(BUFFER_ATOMIC_SWAP)
4332 NODE_NAME_CASE(BUFFER_ATOMIC_ADD)
4333 NODE_NAME_CASE(BUFFER_ATOMIC_SUB)
4334 NODE_NAME_CASE(BUFFER_ATOMIC_SMIN)
4335 NODE_NAME_CASE(BUFFER_ATOMIC_UMIN)
4336 NODE_NAME_CASE(BUFFER_ATOMIC_SMAX)
4337 NODE_NAME_CASE(BUFFER_ATOMIC_UMAX)
4338 NODE_NAME_CASE(BUFFER_ATOMIC_AND)
4339 NODE_NAME_CASE(BUFFER_ATOMIC_OR)
4340 NODE_NAME_CASE(BUFFER_ATOMIC_XOR)
4341 NODE_NAME_CASE(BUFFER_ATOMIC_INC)
4342 NODE_NAME_CASE(BUFFER_ATOMIC_DEC)
4343 NODE_NAME_CASE(BUFFER_ATOMIC_CMPSWAP)
4344 NODE_NAME_CASE(BUFFER_ATOMIC_FADD)
4345 NODE_NAME_CASE(BUFFER_ATOMIC_PK_FADD)
4346 NODE_NAME_CASE(ATOMIC_FADD)
4347 NODE_NAME_CASE(ATOMIC_PK_FADD)
4348
4349 case AMDGPUISD::LAST_AMDGPU_ISD_NUMBER: break;
4350 }
4351 return nullptr;
4352 }
4353
4354 SDValue AMDGPUTargetLowering::getSqrtEstimate(SDValue Operand,
4355 SelectionDAG &DAG, int Enabled,
4356 int &RefinementSteps,
4357 bool &UseOneConstNR,
4358 bool Reciprocal) const {
4359 EVT VT = Operand.getValueType();
4360
4361 if (VT == MVT::f32) {
4362 RefinementSteps = 0;
4363 return DAG.getNode(AMDGPUISD::RSQ, SDLoc(Operand), VT, Operand);
4364 }
4365
4366 // TODO: There is also f64 rsq instruction, but the documentation is less
4367 // clear on its precision.
4368
4369 return SDValue();
4370 }
4371
4372 SDValue AMDGPUTargetLowering::getRecipEstimate(SDValue Operand,
4373 SelectionDAG &DAG, int Enabled,
4374 int &RefinementSteps) const {
4375 EVT VT = Operand.getValueType();
4376
4377 if (VT == MVT::f32) {
4378 // Reciprocal, < 1 ulp error.
4379 //
4380 // This reciprocal approximation converges to < 0.5 ulp error with one
4381 // newton rhapson performed with two fused multiple adds (FMAs).
4382
4383 RefinementSteps = 0;
4384 return DAG.getNode(AMDGPUISD::RCP, SDLoc(Operand), VT, Operand);
4385 }
4386
4387 // TODO: There is also f64 rcp instruction, but the documentation is less
4388 // clear on its precision.
4389
4390 return SDValue();
4391 }
4392
4393 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
4394 const SDValue Op, KnownBits &Known,
4395 const APInt &DemandedElts, const SelectionDAG &DAG, unsigned Depth) const {
4396
4397 Known.resetAll(); // Don't know anything.
4398
4399 unsigned Opc = Op.getOpcode();
4400
4401 switch (Opc) {
4402 default:
4403 break;
4404 case AMDGPUISD::CARRY:
4405 case AMDGPUISD::BORROW: {
4406 Known.Zero = APInt::getHighBitsSet(32, 31);
4407 break;
4408 }
4409
4410 case AMDGPUISD::BFE_I32:
4411 case AMDGPUISD::BFE_U32: {
4412 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4413 if (!CWidth)
4414 return;
4415
4416 uint32_t Width = CWidth->getZExtValue() & 0x1f;
4417
4418 if (Opc == AMDGPUISD::BFE_U32)
4419 Known.Zero = APInt::getHighBitsSet(32, 32 - Width);
4420
4421 break;
4422 }
4423 case AMDGPUISD::FP_TO_FP16:
4424 case AMDGPUISD::FP16_ZEXT: {
4425 unsigned BitWidth = Known.getBitWidth();
4426
4427 // High bits are zero.
4428 Known.Zero = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
4429 break;
4430 }
4431 case AMDGPUISD::MUL_U24:
4432 case AMDGPUISD::MUL_I24: {
4433 KnownBits LHSKnown = DAG.computeKnownBits(Op.getOperand(0), Depth + 1);
4434 KnownBits RHSKnown = DAG.computeKnownBits(Op.getOperand(1), Depth + 1);
4435 unsigned TrailZ = LHSKnown.countMinTrailingZeros() +
4436 RHSKnown.countMinTrailingZeros();
4437 Known.Zero.setLowBits(std::min(TrailZ, 32u));
4438
4439 // Truncate to 24 bits.
4440 LHSKnown = LHSKnown.trunc(24);
4441 RHSKnown = RHSKnown.trunc(24);
4442
4443 bool Negative = false;
4444 if (Opc == AMDGPUISD::MUL_I24) {
4445 unsigned LHSValBits = 24 - LHSKnown.countMinSignBits();
4446 unsigned RHSValBits = 24 - RHSKnown.countMinSignBits();
4447 unsigned MaxValBits = std::min(LHSValBits + RHSValBits, 32u);
4448 if (MaxValBits >= 32)
4449 break;
4450 bool LHSNegative = LHSKnown.isNegative();
4451 bool LHSPositive = LHSKnown.isNonNegative();
4452 bool RHSNegative = RHSKnown.isNegative();
4453 bool RHSPositive = RHSKnown.isNonNegative();
4454 if ((!LHSNegative && !LHSPositive) || (!RHSNegative && !RHSPositive))
4455 break;
4456 Negative = (LHSNegative && RHSPositive) || (LHSPositive && RHSNegative);
4457 if (Negative)
4458 Known.One.setHighBits(32 - MaxValBits);
4459 else
4460 Known.Zero.setHighBits(32 - MaxValBits);
4461 } else {
4462 unsigned LHSValBits = 24 - LHSKnown.countMinLeadingZeros();
4463 unsigned RHSValBits = 24 - RHSKnown.countMinLeadingZeros();
4464 unsigned MaxValBits = std::min(LHSValBits + RHSValBits, 32u);
4465 if (MaxValBits >= 32)
4466 break;
4467 Known.Zero.setHighBits(32 - MaxValBits);
4468 }
4469 break;
4470 }
4471 case AMDGPUISD::PERM: {
4472 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4473 if (!CMask)
4474 return;
4475
4476 KnownBits LHSKnown = DAG.computeKnownBits(Op.getOperand(0), Depth + 1);
4477 KnownBits RHSKnown = DAG.computeKnownBits(Op.getOperand(1), Depth + 1);
4478 unsigned Sel = CMask->getZExtValue();
4479
4480 for (unsigned I = 0; I < 32; I += 8) {
4481 unsigned SelBits = Sel & 0xff;
4482 if (SelBits < 4) {
4483 SelBits *= 8;
4484 Known.One |= ((RHSKnown.One.getZExtValue() >> SelBits) & 0xff) << I;
4485 Known.Zero |= ((RHSKnown.Zero.getZExtValue() >> SelBits) & 0xff) << I;
4486 } else if (SelBits < 7) {
4487 SelBits = (SelBits & 3) * 8;
4488 Known.One |= ((LHSKnown.One.getZExtValue() >> SelBits) & 0xff) << I;
4489 Known.Zero |= ((LHSKnown.Zero.getZExtValue() >> SelBits) & 0xff) << I;
4490 } else if (SelBits == 0x0c) {
4491 Known.Zero |= 0xFFull << I;
4492 } else if (SelBits > 0x0c) {
4493 Known.One |= 0xFFull << I;
4494 }
4495 Sel >>= 8;
4496 }
4497 break;
4498 }
4499 case AMDGPUISD::BUFFER_LOAD_UBYTE: {
4500 Known.Zero.setHighBits(24);
4501 break;
4502 }
4503 case AMDGPUISD::BUFFER_LOAD_USHORT: {
4504 Known.Zero.setHighBits(16);
4505 break;
4506 }
4507 case AMDGPUISD::LDS: {
4508 auto GA = cast<GlobalAddressSDNode>(Op.getOperand(0).getNode());
4509 unsigned Align = GA->getGlobal()->getAlignment();
4510
4511 Known.Zero.setHighBits(16);
4512 if (Align)
4513 Known.Zero.setLowBits(Log2_32(Align));
4514 break;
4515 }
4516 case ISD::INTRINSIC_WO_CHAIN: {
4517 unsigned IID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4518 switch (IID) {
4519 case Intrinsic::amdgcn_mbcnt_lo:
4520 case Intrinsic::amdgcn_mbcnt_hi: {
4521 const GCNSubtarget &ST =
4522 DAG.getMachineFunction().getSubtarget<GCNSubtarget>();
4523 // These return at most the wavefront size - 1.
4524 unsigned Size = Op.getValueType().getSizeInBits();
4525 Known.Zero.setHighBits(Size - ST.getWavefrontSizeLog2());
4526 break;
4527 }
4528 default:
4529 break;
4530 }
4531 }
4532 }
4533 }
4534
4535 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
4536 SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
4537 unsigned Depth) const {
4538 switch (Op.getOpcode()) {
4539 case AMDGPUISD::BFE_I32: {
4540 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4541 if (!Width)
4542 return 1;
4543
4544 unsigned SignBits = 32 - Width->getZExtValue() + 1;
4545 if (!isNullConstant(Op.getOperand(1)))
4546 return SignBits;
4547
4548 // TODO: Could probably figure something out with non-0 offsets.
4549 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
4550 return std::max(SignBits, Op0SignBits);
4551 }
4552
4553 case AMDGPUISD::BFE_U32: {
4554 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4555 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;
4556 }
4557
4558 case AMDGPUISD::CARRY:
4559 case AMDGPUISD::BORROW:
4560 return 31;
4561 case AMDGPUISD::BUFFER_LOAD_BYTE:
4562 return 25;
4563 case AMDGPUISD::BUFFER_LOAD_SHORT:
4564 return 17;
4565 case AMDGPUISD::BUFFER_LOAD_UBYTE:
4566 return 24;
4567 case AMDGPUISD::BUFFER_LOAD_USHORT:
4568 return 16;
4569 case AMDGPUISD::FP_TO_FP16:
4570 case AMDGPUISD::FP16_ZEXT:
4571 return 16;
4572 default:
4573 return 1;
4574 }
4575 }
4576
4577 bool AMDGPUTargetLowering::isKnownNeverNaNForTargetNode(SDValue Op,
4578 const SelectionDAG &DAG,
4579 bool SNaN,
4580 unsigned Depth) const {
4581 unsigned Opcode = Op.getOpcode();
4582 switch (Opcode) {
4583 case AMDGPUISD::FMIN_LEGACY:
4584 case AMDGPUISD::FMAX_LEGACY: {
4585 if (SNaN)
4586 return true;
4587
4588 // TODO: Can check no nans on one of the operands for each one, but which
4589 // one?
4590 return false;
4591 }
4592 case AMDGPUISD::FMUL_LEGACY:
4593 case AMDGPUISD::CVT_PKRTZ_F16_F32: {
4594 if (SNaN)
4595 return true;
4596 return DAG.isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1) &&
4597 DAG.isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1);
4598 }
4599 case AMDGPUISD::FMED3:
4600 case AMDGPUISD::FMIN3:
4601 case AMDGPUISD::FMAX3:
4602 case AMDGPUISD::FMAD_FTZ: {
4603 if (SNaN)
4604 return true;
4605 return DAG.isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1) &&
4606 DAG.isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1) &&
4607 DAG.isKnownNeverNaN(Op.getOperand(2), SNaN, Depth + 1);
4608 }
4609 case AMDGPUISD::CVT_F32_UBYTE0:
4610 case AMDGPUISD::CVT_F32_UBYTE1:
4611 case AMDGPUISD::CVT_F32_UBYTE2:
4612 case AMDGPUISD::CVT_F32_UBYTE3:
4613 return true;
4614
4615 case AMDGPUISD::RCP:
4616 case AMDGPUISD::RSQ:
4617 case AMDGPUISD::RCP_LEGACY:
4618 case AMDGPUISD::RSQ_LEGACY:
4619 case AMDGPUISD::RSQ_CLAMP: {
4620 if (SNaN)
4621 return true;
4622
4623 // TODO: Need is known positive check.
4624 return false;
4625 }
4626 case AMDGPUISD::LDEXP:
4627 case AMDGPUISD::FRACT: {
4628 if (SNaN)
4629 return true;
4630 return DAG.isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1);
4631 }
4632 case AMDGPUISD::DIV_SCALE:
4633 case AMDGPUISD::DIV_FMAS:
4634 case AMDGPUISD::DIV_FIXUP:
4635 case AMDGPUISD::TRIG_PREOP:
4636 // TODO: Refine on operands.
4637 return SNaN;
4638 case AMDGPUISD::SIN_HW:
4639 case AMDGPUISD::COS_HW: {
4640 // TODO: Need check for infinity
4641 return SNaN;
4642 }
4643 case ISD::INTRINSIC_WO_CHAIN: {
4644 unsigned IntrinsicID
4645 = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4646 // TODO: Handle more intrinsics
4647 switch (IntrinsicID) {
4648 case Intrinsic::amdgcn_cubeid:
4649 return true;
4650
4651 case Intrinsic::amdgcn_frexp_mant: {
4652 if (SNaN)
4653 return true;
4654 return DAG.isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1);
4655 }
4656 case Intrinsic::amdgcn_cvt_pkrtz: {
4657 if (SNaN)
4658 return true;
4659 return DAG.isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1) &&
4660 DAG.isKnownNeverNaN(Op.getOperand(2), SNaN, Depth + 1);
4661 }
4662 case Intrinsic::amdgcn_fdot2:
4663 // TODO: Refine on operand
4664 return SNaN;
4665 default:
4666 return false;
4667 }
4668 }
4669 default:
4670 return false;
4671 }
4672 }
4673
4674 TargetLowering::AtomicExpansionKind
4675 AMDGPUTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const {
4676 switch (RMW->getOperation()) {
4677 case AtomicRMWInst::Nand:
4678 case AtomicRMWInst::FAdd:
4679 case AtomicRMWInst::FSub:
4680 return AtomicExpansionKind::CmpXChg;
4681 default:
4682 return AtomicExpansionKind::None;
4683 }
4684 }
The LLVM Project is a collection of modular and reusable compiler and toolchain technologies.