File: | build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp |
Warning: | line 4363, column 43 The result of the right shift is undefined due to shifting by '32', which is greater or equal to the width of type 'unsigned int' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===-- AMDGPUISelLowering.cpp - AMDGPU Common DAG lowering functions -----===// | |||
2 | // | |||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | |||
4 | // See https://llvm.org/LICENSE.txt for license information. | |||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | |||
6 | // | |||
7 | //===----------------------------------------------------------------------===// | |||
8 | // | |||
9 | /// \file | |||
10 | /// This is the parent TargetLowering class for hardware code gen | |||
11 | /// targets. | |||
12 | // | |||
13 | //===----------------------------------------------------------------------===// | |||
14 | ||||
15 | #include "AMDGPUISelLowering.h" | |||
16 | #include "AMDGPU.h" | |||
17 | #include "AMDGPUInstrInfo.h" | |||
18 | #include "AMDGPUMachineFunction.h" | |||
19 | #include "GCNSubtarget.h" | |||
20 | #include "SIMachineFunctionInfo.h" | |||
21 | #include "llvm/CodeGen/Analysis.h" | |||
22 | #include "llvm/CodeGen/MachineFrameInfo.h" | |||
23 | #include "llvm/IR/DiagnosticInfo.h" | |||
24 | #include "llvm/IR/IntrinsicsAMDGPU.h" | |||
25 | #include "llvm/Support/CommandLine.h" | |||
26 | #include "llvm/Support/KnownBits.h" | |||
27 | #include "llvm/Target/TargetMachine.h" | |||
28 | ||||
29 | using namespace llvm; | |||
30 | ||||
31 | #include "AMDGPUGenCallingConv.inc" | |||
32 | ||||
33 | static cl::opt<bool> AMDGPUBypassSlowDiv( | |||
34 | "amdgpu-bypass-slow-div", | |||
35 | cl::desc("Skip 64-bit divide for dynamic 32-bit values"), | |||
36 | cl::init(true)); | |||
37 | ||||
38 | // Find a larger type to do a load / store of a vector with. | |||
39 | EVT AMDGPUTargetLowering::getEquivalentMemType(LLVMContext &Ctx, EVT VT) { | |||
40 | unsigned StoreSize = VT.getStoreSizeInBits(); | |||
41 | if (StoreSize <= 32) | |||
42 | return EVT::getIntegerVT(Ctx, StoreSize); | |||
43 | ||||
44 | assert(StoreSize % 32 == 0 && "Store size not a multiple of 32")(static_cast <bool> (StoreSize % 32 == 0 && "Store size not a multiple of 32" ) ? void (0) : __assert_fail ("StoreSize % 32 == 0 && \"Store size not a multiple of 32\"" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 44, __extension__ __PRETTY_FUNCTION__)); | |||
45 | return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32); | |||
46 | } | |||
47 | ||||
48 | unsigned AMDGPUTargetLowering::numBitsUnsigned(SDValue Op, SelectionDAG &DAG) { | |||
49 | return DAG.computeKnownBits(Op).countMaxActiveBits(); | |||
50 | } | |||
51 | ||||
52 | unsigned AMDGPUTargetLowering::numBitsSigned(SDValue Op, SelectionDAG &DAG) { | |||
53 | // In order for this to be a signed 24-bit value, bit 23, must | |||
54 | // be a sign bit. | |||
55 | return DAG.ComputeMaxSignificantBits(Op); | |||
56 | } | |||
57 | ||||
58 | AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM, | |||
59 | const AMDGPUSubtarget &STI) | |||
60 | : TargetLowering(TM), Subtarget(&STI) { | |||
61 | // Lower floating point store/load to integer store/load to reduce the number | |||
62 | // of patterns in tablegen. | |||
63 | setOperationAction(ISD::LOAD, MVT::f32, Promote); | |||
64 | AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32); | |||
65 | ||||
66 | setOperationAction(ISD::LOAD, MVT::v2f32, Promote); | |||
67 | AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32); | |||
68 | ||||
69 | setOperationAction(ISD::LOAD, MVT::v3f32, Promote); | |||
70 | AddPromotedToType(ISD::LOAD, MVT::v3f32, MVT::v3i32); | |||
71 | ||||
72 | setOperationAction(ISD::LOAD, MVT::v4f32, Promote); | |||
73 | AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32); | |||
74 | ||||
75 | setOperationAction(ISD::LOAD, MVT::v5f32, Promote); | |||
76 | AddPromotedToType(ISD::LOAD, MVT::v5f32, MVT::v5i32); | |||
77 | ||||
78 | setOperationAction(ISD::LOAD, MVT::v6f32, Promote); | |||
79 | AddPromotedToType(ISD::LOAD, MVT::v6f32, MVT::v6i32); | |||
80 | ||||
81 | setOperationAction(ISD::LOAD, MVT::v7f32, Promote); | |||
82 | AddPromotedToType(ISD::LOAD, MVT::v7f32, MVT::v7i32); | |||
83 | ||||
84 | setOperationAction(ISD::LOAD, MVT::v8f32, Promote); | |||
85 | AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32); | |||
86 | ||||
87 | setOperationAction(ISD::LOAD, MVT::v16f32, Promote); | |||
88 | AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32); | |||
89 | ||||
90 | setOperationAction(ISD::LOAD, MVT::v32f32, Promote); | |||
91 | AddPromotedToType(ISD::LOAD, MVT::v32f32, MVT::v32i32); | |||
92 | ||||
93 | setOperationAction(ISD::LOAD, MVT::i64, Promote); | |||
94 | AddPromotedToType(ISD::LOAD, MVT::i64, MVT::v2i32); | |||
95 | ||||
96 | setOperationAction(ISD::LOAD, MVT::v2i64, Promote); | |||
97 | AddPromotedToType(ISD::LOAD, MVT::v2i64, MVT::v4i32); | |||
98 | ||||
99 | setOperationAction(ISD::LOAD, MVT::f64, Promote); | |||
100 | AddPromotedToType(ISD::LOAD, MVT::f64, MVT::v2i32); | |||
101 | ||||
102 | setOperationAction(ISD::LOAD, MVT::v2f64, Promote); | |||
103 | AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v4i32); | |||
104 | ||||
105 | setOperationAction(ISD::LOAD, MVT::v3i64, Promote); | |||
106 | AddPromotedToType(ISD::LOAD, MVT::v3i64, MVT::v6i32); | |||
107 | ||||
108 | setOperationAction(ISD::LOAD, MVT::v4i64, Promote); | |||
109 | AddPromotedToType(ISD::LOAD, MVT::v4i64, MVT::v8i32); | |||
110 | ||||
111 | setOperationAction(ISD::LOAD, MVT::v3f64, Promote); | |||
112 | AddPromotedToType(ISD::LOAD, MVT::v3f64, MVT::v6i32); | |||
113 | ||||
114 | setOperationAction(ISD::LOAD, MVT::v4f64, Promote); | |||
115 | AddPromotedToType(ISD::LOAD, MVT::v4f64, MVT::v8i32); | |||
116 | ||||
117 | setOperationAction(ISD::LOAD, MVT::v8i64, Promote); | |||
118 | AddPromotedToType(ISD::LOAD, MVT::v8i64, MVT::v16i32); | |||
119 | ||||
120 | setOperationAction(ISD::LOAD, MVT::v8f64, Promote); | |||
121 | AddPromotedToType(ISD::LOAD, MVT::v8f64, MVT::v16i32); | |||
122 | ||||
123 | setOperationAction(ISD::LOAD, MVT::v16i64, Promote); | |||
124 | AddPromotedToType(ISD::LOAD, MVT::v16i64, MVT::v32i32); | |||
125 | ||||
126 | setOperationAction(ISD::LOAD, MVT::v16f64, Promote); | |||
127 | AddPromotedToType(ISD::LOAD, MVT::v16f64, MVT::v32i32); | |||
128 | ||||
129 | // There are no 64-bit extloads. These should be done as a 32-bit extload and | |||
130 | // an extension to 64-bit. | |||
131 | for (MVT VT : MVT::integer_valuetypes()) { | |||
132 | setLoadExtAction(ISD::EXTLOAD, MVT::i64, VT, Expand); | |||
133 | setLoadExtAction(ISD::SEXTLOAD, MVT::i64, VT, Expand); | |||
134 | setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, VT, Expand); | |||
135 | } | |||
136 | ||||
137 | for (MVT VT : MVT::integer_valuetypes()) { | |||
138 | if (VT == MVT::i64) | |||
139 | continue; | |||
140 | ||||
141 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote); | |||
142 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Legal); | |||
143 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i16, Legal); | |||
144 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i32, Expand); | |||
145 | ||||
146 | setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote); | |||
147 | setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i8, Legal); | |||
148 | setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i16, Legal); | |||
149 | setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i32, Expand); | |||
150 | ||||
151 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote); | |||
152 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::i8, Legal); | |||
153 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::i16, Legal); | |||
154 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::i32, Expand); | |||
155 | } | |||
156 | ||||
157 | for (MVT VT : MVT::integer_fixedlen_vector_valuetypes()) { | |||
158 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Expand); | |||
159 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Expand); | |||
160 | setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i8, Expand); | |||
161 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Expand); | |||
162 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Expand); | |||
163 | setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i8, Expand); | |||
164 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Expand); | |||
165 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Expand); | |||
166 | setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i16, Expand); | |||
167 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v3i16, Expand); | |||
168 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v3i16, Expand); | |||
169 | setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v3i16, Expand); | |||
170 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Expand); | |||
171 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Expand); | |||
172 | setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i16, Expand); | |||
173 | } | |||
174 | ||||
175 | setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand); | |||
176 | setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, MVT::v2f16, Expand); | |||
177 | setLoadExtAction(ISD::EXTLOAD, MVT::v3f32, MVT::v3f16, Expand); | |||
178 | setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, MVT::v4f16, Expand); | |||
179 | setLoadExtAction(ISD::EXTLOAD, MVT::v8f32, MVT::v8f16, Expand); | |||
180 | setLoadExtAction(ISD::EXTLOAD, MVT::v16f32, MVT::v16f16, Expand); | |||
181 | setLoadExtAction(ISD::EXTLOAD, MVT::v32f32, MVT::v32f16, Expand); | |||
182 | ||||
183 | setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand); | |||
184 | setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f32, Expand); | |||
185 | setLoadExtAction(ISD::EXTLOAD, MVT::v3f64, MVT::v3f32, Expand); | |||
186 | setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f32, Expand); | |||
187 | setLoadExtAction(ISD::EXTLOAD, MVT::v8f64, MVT::v8f32, Expand); | |||
188 | setLoadExtAction(ISD::EXTLOAD, MVT::v16f64, MVT::v16f32, Expand); | |||
189 | ||||
190 | setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand); | |||
191 | setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f16, Expand); | |||
192 | setLoadExtAction(ISD::EXTLOAD, MVT::v3f64, MVT::v3f16, Expand); | |||
193 | setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f16, Expand); | |||
194 | setLoadExtAction(ISD::EXTLOAD, MVT::v8f64, MVT::v8f16, Expand); | |||
195 | setLoadExtAction(ISD::EXTLOAD, MVT::v16f64, MVT::v16f16, Expand); | |||
196 | ||||
197 | setOperationAction(ISD::STORE, MVT::f32, Promote); | |||
198 | AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32); | |||
199 | ||||
200 | setOperationAction(ISD::STORE, MVT::v2f32, Promote); | |||
201 | AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32); | |||
202 | ||||
203 | setOperationAction(ISD::STORE, MVT::v3f32, Promote); | |||
204 | AddPromotedToType(ISD::STORE, MVT::v3f32, MVT::v3i32); | |||
205 | ||||
206 | setOperationAction(ISD::STORE, MVT::v4f32, Promote); | |||
207 | AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32); | |||
208 | ||||
209 | setOperationAction(ISD::STORE, MVT::v5f32, Promote); | |||
210 | AddPromotedToType(ISD::STORE, MVT::v5f32, MVT::v5i32); | |||
211 | ||||
212 | setOperationAction(ISD::STORE, MVT::v6f32, Promote); | |||
213 | AddPromotedToType(ISD::STORE, MVT::v6f32, MVT::v6i32); | |||
214 | ||||
215 | setOperationAction(ISD::STORE, MVT::v7f32, Promote); | |||
216 | AddPromotedToType(ISD::STORE, MVT::v7f32, MVT::v7i32); | |||
217 | ||||
218 | setOperationAction(ISD::STORE, MVT::v8f32, Promote); | |||
219 | AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32); | |||
220 | ||||
221 | setOperationAction(ISD::STORE, MVT::v16f32, Promote); | |||
222 | AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32); | |||
223 | ||||
224 | setOperationAction(ISD::STORE, MVT::v32f32, Promote); | |||
225 | AddPromotedToType(ISD::STORE, MVT::v32f32, MVT::v32i32); | |||
226 | ||||
227 | setOperationAction(ISD::STORE, MVT::i64, Promote); | |||
228 | AddPromotedToType(ISD::STORE, MVT::i64, MVT::v2i32); | |||
229 | ||||
230 | setOperationAction(ISD::STORE, MVT::v2i64, Promote); | |||
231 | AddPromotedToType(ISD::STORE, MVT::v2i64, MVT::v4i32); | |||
232 | ||||
233 | setOperationAction(ISD::STORE, MVT::f64, Promote); | |||
234 | AddPromotedToType(ISD::STORE, MVT::f64, MVT::v2i32); | |||
235 | ||||
236 | setOperationAction(ISD::STORE, MVT::v2f64, Promote); | |||
237 | AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v4i32); | |||
238 | ||||
239 | setOperationAction(ISD::STORE, MVT::v3i64, Promote); | |||
240 | AddPromotedToType(ISD::STORE, MVT::v3i64, MVT::v6i32); | |||
241 | ||||
242 | setOperationAction(ISD::STORE, MVT::v3f64, Promote); | |||
243 | AddPromotedToType(ISD::STORE, MVT::v3f64, MVT::v6i32); | |||
244 | ||||
245 | setOperationAction(ISD::STORE, MVT::v4i64, Promote); | |||
246 | AddPromotedToType(ISD::STORE, MVT::v4i64, MVT::v8i32); | |||
247 | ||||
248 | setOperationAction(ISD::STORE, MVT::v4f64, Promote); | |||
249 | AddPromotedToType(ISD::STORE, MVT::v4f64, MVT::v8i32); | |||
250 | ||||
251 | setOperationAction(ISD::STORE, MVT::v8i64, Promote); | |||
252 | AddPromotedToType(ISD::STORE, MVT::v8i64, MVT::v16i32); | |||
253 | ||||
254 | setOperationAction(ISD::STORE, MVT::v8f64, Promote); | |||
255 | AddPromotedToType(ISD::STORE, MVT::v8f64, MVT::v16i32); | |||
256 | ||||
257 | setOperationAction(ISD::STORE, MVT::v16i64, Promote); | |||
258 | AddPromotedToType(ISD::STORE, MVT::v16i64, MVT::v32i32); | |||
259 | ||||
260 | setOperationAction(ISD::STORE, MVT::v16f64, Promote); | |||
261 | AddPromotedToType(ISD::STORE, MVT::v16f64, MVT::v32i32); | |||
262 | ||||
263 | setTruncStoreAction(MVT::i64, MVT::i1, Expand); | |||
264 | setTruncStoreAction(MVT::i64, MVT::i8, Expand); | |||
265 | setTruncStoreAction(MVT::i64, MVT::i16, Expand); | |||
266 | setTruncStoreAction(MVT::i64, MVT::i32, Expand); | |||
267 | ||||
268 | setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand); | |||
269 | setTruncStoreAction(MVT::v2i64, MVT::v2i8, Expand); | |||
270 | setTruncStoreAction(MVT::v2i64, MVT::v2i16, Expand); | |||
271 | setTruncStoreAction(MVT::v2i64, MVT::v2i32, Expand); | |||
272 | ||||
273 | setTruncStoreAction(MVT::f32, MVT::f16, Expand); | |||
274 | setTruncStoreAction(MVT::v2f32, MVT::v2f16, Expand); | |||
275 | setTruncStoreAction(MVT::v3f32, MVT::v3f16, Expand); | |||
276 | setTruncStoreAction(MVT::v4f32, MVT::v4f16, Expand); | |||
277 | setTruncStoreAction(MVT::v8f32, MVT::v8f16, Expand); | |||
278 | setTruncStoreAction(MVT::v16f32, MVT::v16f16, Expand); | |||
279 | setTruncStoreAction(MVT::v32f32, MVT::v32f16, Expand); | |||
280 | ||||
281 | setTruncStoreAction(MVT::f64, MVT::f16, Expand); | |||
282 | setTruncStoreAction(MVT::f64, MVT::f32, Expand); | |||
283 | ||||
284 | setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand); | |||
285 | setTruncStoreAction(MVT::v2f64, MVT::v2f16, Expand); | |||
286 | ||||
287 | setTruncStoreAction(MVT::v3i64, MVT::v3i32, Expand); | |||
288 | setTruncStoreAction(MVT::v3i64, MVT::v3i16, Expand); | |||
289 | setTruncStoreAction(MVT::v3f64, MVT::v3f32, Expand); | |||
290 | setTruncStoreAction(MVT::v3f64, MVT::v3f16, Expand); | |||
291 | ||||
292 | setTruncStoreAction(MVT::v4i64, MVT::v4i32, Expand); | |||
293 | setTruncStoreAction(MVT::v4i64, MVT::v4i16, Expand); | |||
294 | setTruncStoreAction(MVT::v4f64, MVT::v4f32, Expand); | |||
295 | setTruncStoreAction(MVT::v4f64, MVT::v4f16, Expand); | |||
296 | ||||
297 | setTruncStoreAction(MVT::v8f64, MVT::v8f32, Expand); | |||
298 | setTruncStoreAction(MVT::v8f64, MVT::v8f16, Expand); | |||
299 | ||||
300 | setTruncStoreAction(MVT::v16f64, MVT::v16f32, Expand); | |||
301 | setTruncStoreAction(MVT::v16f64, MVT::v16f16, Expand); | |||
302 | setTruncStoreAction(MVT::v16i64, MVT::v16i16, Expand); | |||
303 | setTruncStoreAction(MVT::v16i64, MVT::v16i16, Expand); | |||
304 | setTruncStoreAction(MVT::v16i64, MVT::v16i8, Expand); | |||
305 | setTruncStoreAction(MVT::v16i64, MVT::v16i8, Expand); | |||
306 | setTruncStoreAction(MVT::v16i64, MVT::v16i1, Expand); | |||
307 | ||||
308 | setOperationAction(ISD::Constant, MVT::i32, Legal); | |||
309 | setOperationAction(ISD::Constant, MVT::i64, Legal); | |||
310 | setOperationAction(ISD::ConstantFP, MVT::f32, Legal); | |||
311 | setOperationAction(ISD::ConstantFP, MVT::f64, Legal); | |||
312 | ||||
313 | setOperationAction(ISD::BR_JT, MVT::Other, Expand); | |||
314 | setOperationAction(ISD::BRIND, MVT::Other, Expand); | |||
315 | ||||
316 | // This is totally unsupported, just custom lower to produce an error. | |||
317 | setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom); | |||
318 | ||||
319 | // Library functions. These default to Expand, but we have instructions | |||
320 | // for them. | |||
321 | setOperationAction(ISD::FCEIL, MVT::f32, Legal); | |||
322 | setOperationAction(ISD::FEXP2, MVT::f32, Legal); | |||
323 | setOperationAction(ISD::FPOW, MVT::f32, Legal); | |||
324 | setOperationAction(ISD::FLOG2, MVT::f32, Legal); | |||
325 | setOperationAction(ISD::FABS, MVT::f32, Legal); | |||
326 | setOperationAction(ISD::FFLOOR, MVT::f32, Legal); | |||
327 | setOperationAction(ISD::FRINT, MVT::f32, Legal); | |||
328 | setOperationAction(ISD::FTRUNC, MVT::f32, Legal); | |||
329 | setOperationAction(ISD::FMINNUM, MVT::f32, Legal); | |||
330 | setOperationAction(ISD::FMAXNUM, MVT::f32, Legal); | |||
331 | ||||
332 | setOperationAction(ISD::FROUND, MVT::f32, Custom); | |||
333 | setOperationAction(ISD::FROUND, MVT::f64, Custom); | |||
334 | ||||
335 | setOperationAction(ISD::FLOG, MVT::f32, Custom); | |||
336 | setOperationAction(ISD::FLOG10, MVT::f32, Custom); | |||
337 | setOperationAction(ISD::FEXP, MVT::f32, Custom); | |||
338 | ||||
339 | ||||
340 | setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom); | |||
341 | setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom); | |||
342 | ||||
343 | setOperationAction(ISD::FREM, MVT::f16, Custom); | |||
344 | setOperationAction(ISD::FREM, MVT::f32, Custom); | |||
345 | setOperationAction(ISD::FREM, MVT::f64, Custom); | |||
346 | ||||
347 | // Expand to fneg + fadd. | |||
348 | setOperationAction(ISD::FSUB, MVT::f64, Expand); | |||
349 | ||||
350 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v3i32, Custom); | |||
351 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v3f32, Custom); | |||
352 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom); | |||
353 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom); | |||
354 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v5i32, Custom); | |||
355 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v5f32, Custom); | |||
356 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v6i32, Custom); | |||
357 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v6f32, Custom); | |||
358 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v7i32, Custom); | |||
359 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v7f32, Custom); | |||
360 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom); | |||
361 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom); | |||
362 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f16, Custom); | |||
363 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i16, Custom); | |||
364 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f16, Custom); | |||
365 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i16, Custom); | |||
366 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom); | |||
367 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom); | |||
368 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v3f32, Custom); | |||
369 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v3i32, Custom); | |||
370 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom); | |||
371 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom); | |||
372 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v5f32, Custom); | |||
373 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v5i32, Custom); | |||
374 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v6f32, Custom); | |||
375 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v6i32, Custom); | |||
376 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v7f32, Custom); | |||
377 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v7i32, Custom); | |||
378 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom); | |||
379 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom); | |||
380 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v16f32, Custom); | |||
381 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v16i32, Custom); | |||
382 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v32f32, Custom); | |||
383 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v32i32, Custom); | |||
384 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f64, Custom); | |||
385 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i64, Custom); | |||
386 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v3f64, Custom); | |||
387 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v3i64, Custom); | |||
388 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f64, Custom); | |||
389 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i64, Custom); | |||
390 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f64, Custom); | |||
391 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i64, Custom); | |||
392 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v16f64, Custom); | |||
393 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v16i64, Custom); | |||
394 | ||||
395 | setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand); | |||
396 | setOperationAction(ISD::FP_TO_FP16, MVT::f64, Custom); | |||
397 | setOperationAction(ISD::FP_TO_FP16, MVT::f32, Custom); | |||
398 | ||||
399 | const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 }; | |||
400 | for (MVT VT : ScalarIntVTs) { | |||
401 | // These should use [SU]DIVREM, so set them to expand | |||
402 | setOperationAction(ISD::SDIV, VT, Expand); | |||
403 | setOperationAction(ISD::UDIV, VT, Expand); | |||
404 | setOperationAction(ISD::SREM, VT, Expand); | |||
405 | setOperationAction(ISD::UREM, VT, Expand); | |||
406 | ||||
407 | // GPU does not have divrem function for signed or unsigned. | |||
408 | setOperationAction(ISD::SDIVREM, VT, Custom); | |||
409 | setOperationAction(ISD::UDIVREM, VT, Custom); | |||
410 | ||||
411 | // GPU does not have [S|U]MUL_LOHI functions as a single instruction. | |||
412 | setOperationAction(ISD::SMUL_LOHI, VT, Expand); | |||
413 | setOperationAction(ISD::UMUL_LOHI, VT, Expand); | |||
414 | ||||
415 | setOperationAction(ISD::BSWAP, VT, Expand); | |||
416 | setOperationAction(ISD::CTTZ, VT, Expand); | |||
417 | setOperationAction(ISD::CTLZ, VT, Expand); | |||
418 | ||||
419 | // AMDGPU uses ADDC/SUBC/ADDE/SUBE | |||
420 | setOperationAction(ISD::ADDC, VT, Legal); | |||
421 | setOperationAction(ISD::SUBC, VT, Legal); | |||
422 | setOperationAction(ISD::ADDE, VT, Legal); | |||
423 | setOperationAction(ISD::SUBE, VT, Legal); | |||
424 | } | |||
425 | ||||
426 | // The hardware supports 32-bit FSHR, but not FSHL. | |||
427 | setOperationAction(ISD::FSHR, MVT::i32, Legal); | |||
428 | ||||
429 | // The hardware supports 32-bit ROTR, but not ROTL. | |||
430 | setOperationAction(ISD::ROTL, MVT::i32, Expand); | |||
431 | setOperationAction(ISD::ROTL, MVT::i64, Expand); | |||
432 | setOperationAction(ISD::ROTR, MVT::i64, Expand); | |||
433 | ||||
434 | setOperationAction(ISD::MULHU, MVT::i16, Expand); | |||
435 | setOperationAction(ISD::MULHS, MVT::i16, Expand); | |||
436 | ||||
437 | setOperationAction(ISD::MUL, MVT::i64, Expand); | |||
438 | setOperationAction(ISD::MULHU, MVT::i64, Expand); | |||
439 | setOperationAction(ISD::MULHS, MVT::i64, Expand); | |||
440 | setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom); | |||
441 | setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom); | |||
442 | setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom); | |||
443 | setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom); | |||
444 | setOperationAction(ISD::SELECT_CC, MVT::i64, Expand); | |||
445 | ||||
446 | setOperationAction(ISD::SMIN, MVT::i32, Legal); | |||
447 | setOperationAction(ISD::UMIN, MVT::i32, Legal); | |||
448 | setOperationAction(ISD::SMAX, MVT::i32, Legal); | |||
449 | setOperationAction(ISD::UMAX, MVT::i32, Legal); | |||
450 | ||||
451 | setOperationAction(ISD::CTTZ, MVT::i64, Custom); | |||
452 | setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Custom); | |||
453 | setOperationAction(ISD::CTLZ, MVT::i64, Custom); | |||
454 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom); | |||
455 | ||||
456 | static const MVT::SimpleValueType VectorIntTypes[] = { | |||
457 | MVT::v2i32, MVT::v3i32, MVT::v4i32, MVT::v5i32, MVT::v6i32, MVT::v7i32}; | |||
458 | ||||
459 | for (MVT VT : VectorIntTypes) { | |||
460 | // Expand the following operations for the current type by default. | |||
461 | setOperationAction(ISD::ADD, VT, Expand); | |||
462 | setOperationAction(ISD::AND, VT, Expand); | |||
463 | setOperationAction(ISD::FP_TO_SINT, VT, Expand); | |||
464 | setOperationAction(ISD::FP_TO_UINT, VT, Expand); | |||
465 | setOperationAction(ISD::MUL, VT, Expand); | |||
466 | setOperationAction(ISD::MULHU, VT, Expand); | |||
467 | setOperationAction(ISD::MULHS, VT, Expand); | |||
468 | setOperationAction(ISD::OR, VT, Expand); | |||
469 | setOperationAction(ISD::SHL, VT, Expand); | |||
470 | setOperationAction(ISD::SRA, VT, Expand); | |||
471 | setOperationAction(ISD::SRL, VT, Expand); | |||
472 | setOperationAction(ISD::ROTL, VT, Expand); | |||
473 | setOperationAction(ISD::ROTR, VT, Expand); | |||
474 | setOperationAction(ISD::SUB, VT, Expand); | |||
475 | setOperationAction(ISD::SINT_TO_FP, VT, Expand); | |||
476 | setOperationAction(ISD::UINT_TO_FP, VT, Expand); | |||
477 | setOperationAction(ISD::SDIV, VT, Expand); | |||
478 | setOperationAction(ISD::UDIV, VT, Expand); | |||
479 | setOperationAction(ISD::SREM, VT, Expand); | |||
480 | setOperationAction(ISD::UREM, VT, Expand); | |||
481 | setOperationAction(ISD::SMUL_LOHI, VT, Expand); | |||
482 | setOperationAction(ISD::UMUL_LOHI, VT, Expand); | |||
483 | setOperationAction(ISD::SDIVREM, VT, Expand); | |||
484 | setOperationAction(ISD::UDIVREM, VT, Expand); | |||
485 | setOperationAction(ISD::SELECT, VT, Expand); | |||
486 | setOperationAction(ISD::VSELECT, VT, Expand); | |||
487 | setOperationAction(ISD::SELECT_CC, VT, Expand); | |||
488 | setOperationAction(ISD::XOR, VT, Expand); | |||
489 | setOperationAction(ISD::BSWAP, VT, Expand); | |||
490 | setOperationAction(ISD::CTPOP, VT, Expand); | |||
491 | setOperationAction(ISD::CTTZ, VT, Expand); | |||
492 | setOperationAction(ISD::CTLZ, VT, Expand); | |||
493 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand); | |||
494 | setOperationAction(ISD::SETCC, VT, Expand); | |||
495 | } | |||
496 | ||||
497 | static const MVT::SimpleValueType FloatVectorTypes[] = { | |||
498 | MVT::v2f32, MVT::v3f32, MVT::v4f32, MVT::v5f32, MVT::v6f32, MVT::v7f32}; | |||
499 | ||||
500 | for (MVT VT : FloatVectorTypes) { | |||
501 | setOperationAction(ISD::FABS, VT, Expand); | |||
502 | setOperationAction(ISD::FMINNUM, VT, Expand); | |||
503 | setOperationAction(ISD::FMAXNUM, VT, Expand); | |||
504 | setOperationAction(ISD::FADD, VT, Expand); | |||
505 | setOperationAction(ISD::FCEIL, VT, Expand); | |||
506 | setOperationAction(ISD::FCOS, VT, Expand); | |||
507 | setOperationAction(ISD::FDIV, VT, Expand); | |||
508 | setOperationAction(ISD::FEXP2, VT, Expand); | |||
509 | setOperationAction(ISD::FEXP, VT, Expand); | |||
510 | setOperationAction(ISD::FLOG2, VT, Expand); | |||
511 | setOperationAction(ISD::FREM, VT, Expand); | |||
512 | setOperationAction(ISD::FLOG, VT, Expand); | |||
513 | setOperationAction(ISD::FLOG10, VT, Expand); | |||
514 | setOperationAction(ISD::FPOW, VT, Expand); | |||
515 | setOperationAction(ISD::FFLOOR, VT, Expand); | |||
516 | setOperationAction(ISD::FTRUNC, VT, Expand); | |||
517 | setOperationAction(ISD::FMUL, VT, Expand); | |||
518 | setOperationAction(ISD::FMA, VT, Expand); | |||
519 | setOperationAction(ISD::FRINT, VT, Expand); | |||
520 | setOperationAction(ISD::FNEARBYINT, VT, Expand); | |||
521 | setOperationAction(ISD::FSQRT, VT, Expand); | |||
522 | setOperationAction(ISD::FSIN, VT, Expand); | |||
523 | setOperationAction(ISD::FSUB, VT, Expand); | |||
524 | setOperationAction(ISD::FNEG, VT, Expand); | |||
525 | setOperationAction(ISD::VSELECT, VT, Expand); | |||
526 | setOperationAction(ISD::SELECT_CC, VT, Expand); | |||
527 | setOperationAction(ISD::FCOPYSIGN, VT, Expand); | |||
528 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand); | |||
529 | setOperationAction(ISD::SETCC, VT, Expand); | |||
530 | setOperationAction(ISD::FCANONICALIZE, VT, Expand); | |||
531 | } | |||
532 | ||||
533 | // This causes using an unrolled select operation rather than expansion with | |||
534 | // bit operations. This is in general better, but the alternative using BFI | |||
535 | // instructions may be better if the select sources are SGPRs. | |||
536 | setOperationAction(ISD::SELECT, MVT::v2f32, Promote); | |||
537 | AddPromotedToType(ISD::SELECT, MVT::v2f32, MVT::v2i32); | |||
538 | ||||
539 | setOperationAction(ISD::SELECT, MVT::v3f32, Promote); | |||
540 | AddPromotedToType(ISD::SELECT, MVT::v3f32, MVT::v3i32); | |||
541 | ||||
542 | setOperationAction(ISD::SELECT, MVT::v4f32, Promote); | |||
543 | AddPromotedToType(ISD::SELECT, MVT::v4f32, MVT::v4i32); | |||
544 | ||||
545 | setOperationAction(ISD::SELECT, MVT::v5f32, Promote); | |||
546 | AddPromotedToType(ISD::SELECT, MVT::v5f32, MVT::v5i32); | |||
547 | ||||
548 | setOperationAction(ISD::SELECT, MVT::v6f32, Promote); | |||
549 | AddPromotedToType(ISD::SELECT, MVT::v6f32, MVT::v6i32); | |||
550 | ||||
551 | setOperationAction(ISD::SELECT, MVT::v7f32, Promote); | |||
552 | AddPromotedToType(ISD::SELECT, MVT::v7f32, MVT::v7i32); | |||
553 | ||||
554 | // There are no libcalls of any kind. | |||
555 | for (int I = 0; I < RTLIB::UNKNOWN_LIBCALL; ++I) | |||
556 | setLibcallName(static_cast<RTLIB::Libcall>(I), nullptr); | |||
557 | ||||
558 | setSchedulingPreference(Sched::RegPressure); | |||
559 | setJumpIsExpensive(true); | |||
560 | ||||
561 | // FIXME: This is only partially true. If we have to do vector compares, any | |||
562 | // SGPR pair can be a condition register. If we have a uniform condition, we | |||
563 | // are better off doing SALU operations, where there is only one SCC. For now, | |||
564 | // we don't have a way of knowing during instruction selection if a condition | |||
565 | // will be uniform and we always use vector compares. Assume we are using | |||
566 | // vector compares until that is fixed. | |||
567 | setHasMultipleConditionRegisters(true); | |||
568 | ||||
569 | setMinCmpXchgSizeInBits(32); | |||
570 | setSupportsUnalignedAtomics(false); | |||
571 | ||||
572 | PredictableSelectIsExpensive = false; | |||
573 | ||||
574 | // We want to find all load dependencies for long chains of stores to enable | |||
575 | // merging into very wide vectors. The problem is with vectors with > 4 | |||
576 | // elements. MergeConsecutiveStores will attempt to merge these because x8/x16 | |||
577 | // vectors are a legal type, even though we have to split the loads | |||
578 | // usually. When we can more precisely specify load legality per address | |||
579 | // space, we should be able to make FindBetterChain/MergeConsecutiveStores | |||
580 | // smarter so that they can figure out what to do in 2 iterations without all | |||
581 | // N > 4 stores on the same chain. | |||
582 | GatherAllAliasesMaxDepth = 16; | |||
583 | ||||
584 | // memcpy/memmove/memset are expanded in the IR, so we shouldn't need to worry | |||
585 | // about these during lowering. | |||
586 | MaxStoresPerMemcpy = 0xffffffff; | |||
587 | MaxStoresPerMemmove = 0xffffffff; | |||
588 | MaxStoresPerMemset = 0xffffffff; | |||
589 | ||||
590 | // The expansion for 64-bit division is enormous. | |||
591 | if (AMDGPUBypassSlowDiv) | |||
592 | addBypassSlowDiv(64, 32); | |||
593 | ||||
594 | setTargetDAGCombine({ISD::BITCAST, ISD::SHL, | |||
595 | ISD::SRA, ISD::SRL, | |||
596 | ISD::TRUNCATE, ISD::MUL, | |||
597 | ISD::SMUL_LOHI, ISD::UMUL_LOHI, | |||
598 | ISD::MULHU, ISD::MULHS, | |||
599 | ISD::SELECT, ISD::SELECT_CC, | |||
600 | ISD::STORE, ISD::FADD, | |||
601 | ISD::FSUB, ISD::FNEG, | |||
602 | ISD::FABS, ISD::AssertZext, | |||
603 | ISD::AssertSext, ISD::INTRINSIC_WO_CHAIN}); | |||
604 | } | |||
605 | ||||
606 | bool AMDGPUTargetLowering::mayIgnoreSignedZero(SDValue Op) const { | |||
607 | if (getTargetMachine().Options.NoSignedZerosFPMath) | |||
608 | return true; | |||
609 | ||||
610 | const auto Flags = Op.getNode()->getFlags(); | |||
611 | if (Flags.hasNoSignedZeros()) | |||
612 | return true; | |||
613 | ||||
614 | return false; | |||
615 | } | |||
616 | ||||
617 | //===----------------------------------------------------------------------===// | |||
618 | // Target Information | |||
619 | //===----------------------------------------------------------------------===// | |||
620 | ||||
621 | LLVM_READNONE__attribute__((__const__)) | |||
622 | static bool fnegFoldsIntoOp(unsigned Opc) { | |||
623 | switch (Opc) { | |||
624 | case ISD::FADD: | |||
625 | case ISD::FSUB: | |||
626 | case ISD::FMUL: | |||
627 | case ISD::FMA: | |||
628 | case ISD::FMAD: | |||
629 | case ISD::FMINNUM: | |||
630 | case ISD::FMAXNUM: | |||
631 | case ISD::FMINNUM_IEEE: | |||
632 | case ISD::FMAXNUM_IEEE: | |||
633 | case ISD::FSIN: | |||
634 | case ISD::FTRUNC: | |||
635 | case ISD::FRINT: | |||
636 | case ISD::FNEARBYINT: | |||
637 | case ISD::FCANONICALIZE: | |||
638 | case AMDGPUISD::RCP: | |||
639 | case AMDGPUISD::RCP_LEGACY: | |||
640 | case AMDGPUISD::RCP_IFLAG: | |||
641 | case AMDGPUISD::SIN_HW: | |||
642 | case AMDGPUISD::FMUL_LEGACY: | |||
643 | case AMDGPUISD::FMIN_LEGACY: | |||
644 | case AMDGPUISD::FMAX_LEGACY: | |||
645 | case AMDGPUISD::FMED3: | |||
646 | // TODO: handle llvm.amdgcn.fma.legacy | |||
647 | return true; | |||
648 | default: | |||
649 | return false; | |||
650 | } | |||
651 | } | |||
652 | ||||
653 | /// \p returns true if the operation will definitely need to use a 64-bit | |||
654 | /// encoding, and thus will use a VOP3 encoding regardless of the source | |||
655 | /// modifiers. | |||
656 | LLVM_READONLY__attribute__((__pure__)) | |||
657 | static bool opMustUseVOP3Encoding(const SDNode *N, MVT VT) { | |||
658 | return N->getNumOperands() > 2 || VT == MVT::f64; | |||
659 | } | |||
660 | ||||
661 | // Most FP instructions support source modifiers, but this could be refined | |||
662 | // slightly. | |||
663 | LLVM_READONLY__attribute__((__pure__)) | |||
664 | static bool hasSourceMods(const SDNode *N) { | |||
665 | if (isa<MemSDNode>(N)) | |||
666 | return false; | |||
667 | ||||
668 | switch (N->getOpcode()) { | |||
669 | case ISD::CopyToReg: | |||
670 | case ISD::SELECT: | |||
671 | case ISD::FDIV: | |||
672 | case ISD::FREM: | |||
673 | case ISD::INLINEASM: | |||
674 | case ISD::INLINEASM_BR: | |||
675 | case AMDGPUISD::DIV_SCALE: | |||
676 | case ISD::INTRINSIC_W_CHAIN: | |||
677 | ||||
678 | // TODO: Should really be looking at the users of the bitcast. These are | |||
679 | // problematic because bitcasts are used to legalize all stores to integer | |||
680 | // types. | |||
681 | case ISD::BITCAST: | |||
682 | return false; | |||
683 | case ISD::INTRINSIC_WO_CHAIN: { | |||
684 | switch (cast<ConstantSDNode>(N->getOperand(0))->getZExtValue()) { | |||
685 | case Intrinsic::amdgcn_interp_p1: | |||
686 | case Intrinsic::amdgcn_interp_p2: | |||
687 | case Intrinsic::amdgcn_interp_mov: | |||
688 | case Intrinsic::amdgcn_interp_p1_f16: | |||
689 | case Intrinsic::amdgcn_interp_p2_f16: | |||
690 | return false; | |||
691 | default: | |||
692 | return true; | |||
693 | } | |||
694 | } | |||
695 | default: | |||
696 | return true; | |||
697 | } | |||
698 | } | |||
699 | ||||
700 | bool AMDGPUTargetLowering::allUsesHaveSourceMods(const SDNode *N, | |||
701 | unsigned CostThreshold) { | |||
702 | // Some users (such as 3-operand FMA/MAD) must use a VOP3 encoding, and thus | |||
703 | // it is truly free to use a source modifier in all cases. If there are | |||
704 | // multiple users but for each one will necessitate using VOP3, there will be | |||
705 | // a code size increase. Try to avoid increasing code size unless we know it | |||
706 | // will save on the instruction count. | |||
707 | unsigned NumMayIncreaseSize = 0; | |||
708 | MVT VT = N->getValueType(0).getScalarType().getSimpleVT(); | |||
709 | ||||
710 | // XXX - Should this limit number of uses to check? | |||
711 | for (const SDNode *U : N->uses()) { | |||
712 | if (!hasSourceMods(U)) | |||
713 | return false; | |||
714 | ||||
715 | if (!opMustUseVOP3Encoding(U, VT)) { | |||
716 | if (++NumMayIncreaseSize > CostThreshold) | |||
717 | return false; | |||
718 | } | |||
719 | } | |||
720 | ||||
721 | return true; | |||
722 | } | |||
723 | ||||
724 | EVT AMDGPUTargetLowering::getTypeForExtReturn(LLVMContext &Context, EVT VT, | |||
725 | ISD::NodeType ExtendKind) const { | |||
726 | assert(!VT.isVector() && "only scalar expected")(static_cast <bool> (!VT.isVector() && "only scalar expected" ) ? void (0) : __assert_fail ("!VT.isVector() && \"only scalar expected\"" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 726, __extension__ __PRETTY_FUNCTION__)); | |||
727 | ||||
728 | // Round to the next multiple of 32-bits. | |||
729 | unsigned Size = VT.getSizeInBits(); | |||
730 | if (Size <= 32) | |||
731 | return MVT::i32; | |||
732 | return EVT::getIntegerVT(Context, 32 * ((Size + 31) / 32)); | |||
733 | } | |||
734 | ||||
735 | MVT AMDGPUTargetLowering::getVectorIdxTy(const DataLayout &) const { | |||
736 | return MVT::i32; | |||
737 | } | |||
738 | ||||
739 | bool AMDGPUTargetLowering::isSelectSupported(SelectSupportKind SelType) const { | |||
740 | return true; | |||
741 | } | |||
742 | ||||
743 | // The backend supports 32 and 64 bit floating point immediates. | |||
744 | // FIXME: Why are we reporting vectors of FP immediates as legal? | |||
745 | bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT, | |||
746 | bool ForCodeSize) const { | |||
747 | EVT ScalarVT = VT.getScalarType(); | |||
748 | return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64 || | |||
749 | (ScalarVT == MVT::f16 && Subtarget->has16BitInsts())); | |||
750 | } | |||
751 | ||||
752 | // We don't want to shrink f64 / f32 constants. | |||
753 | bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const { | |||
754 | EVT ScalarVT = VT.getScalarType(); | |||
755 | return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64); | |||
756 | } | |||
757 | ||||
758 | bool AMDGPUTargetLowering::shouldReduceLoadWidth(SDNode *N, | |||
759 | ISD::LoadExtType ExtTy, | |||
760 | EVT NewVT) const { | |||
761 | // TODO: This may be worth removing. Check regression tests for diffs. | |||
762 | if (!TargetLoweringBase::shouldReduceLoadWidth(N, ExtTy, NewVT)) | |||
763 | return false; | |||
764 | ||||
765 | unsigned NewSize = NewVT.getStoreSizeInBits(); | |||
766 | ||||
767 | // If we are reducing to a 32-bit load or a smaller multi-dword load, | |||
768 | // this is always better. | |||
769 | if (NewSize >= 32) | |||
770 | return true; | |||
771 | ||||
772 | EVT OldVT = N->getValueType(0); | |||
773 | unsigned OldSize = OldVT.getStoreSizeInBits(); | |||
774 | ||||
775 | MemSDNode *MN = cast<MemSDNode>(N); | |||
776 | unsigned AS = MN->getAddressSpace(); | |||
777 | // Do not shrink an aligned scalar load to sub-dword. | |||
778 | // Scalar engine cannot do sub-dword loads. | |||
779 | if (OldSize >= 32 && NewSize < 32 && MN->getAlignment() >= 4 && | |||
780 | (AS == AMDGPUAS::CONSTANT_ADDRESS || | |||
781 | AS == AMDGPUAS::CONSTANT_ADDRESS_32BIT || | |||
782 | (isa<LoadSDNode>(N) && | |||
783 | AS == AMDGPUAS::GLOBAL_ADDRESS && MN->isInvariant())) && | |||
784 | AMDGPUInstrInfo::isUniformMMO(MN->getMemOperand())) | |||
785 | return false; | |||
786 | ||||
787 | // Don't produce extloads from sub 32-bit types. SI doesn't have scalar | |||
788 | // extloads, so doing one requires using a buffer_load. In cases where we | |||
789 | // still couldn't use a scalar load, using the wider load shouldn't really | |||
790 | // hurt anything. | |||
791 | ||||
792 | // If the old size already had to be an extload, there's no harm in continuing | |||
793 | // to reduce the width. | |||
794 | return (OldSize < 32); | |||
795 | } | |||
796 | ||||
797 | bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy, EVT CastTy, | |||
798 | const SelectionDAG &DAG, | |||
799 | const MachineMemOperand &MMO) const { | |||
800 | ||||
801 | assert(LoadTy.getSizeInBits() == CastTy.getSizeInBits())(static_cast <bool> (LoadTy.getSizeInBits() == CastTy.getSizeInBits ()) ? void (0) : __assert_fail ("LoadTy.getSizeInBits() == CastTy.getSizeInBits()" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 801, __extension__ __PRETTY_FUNCTION__)); | |||
802 | ||||
803 | if (LoadTy.getScalarType() == MVT::i32) | |||
804 | return false; | |||
805 | ||||
806 | unsigned LScalarSize = LoadTy.getScalarSizeInBits(); | |||
807 | unsigned CastScalarSize = CastTy.getScalarSizeInBits(); | |||
808 | ||||
809 | if ((LScalarSize >= CastScalarSize) && (CastScalarSize < 32)) | |||
810 | return false; | |||
811 | ||||
812 | bool Fast = false; | |||
813 | return allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(), | |||
814 | CastTy, MMO, &Fast) && | |||
815 | Fast; | |||
816 | } | |||
817 | ||||
818 | // SI+ has instructions for cttz / ctlz for 32-bit values. This is probably also | |||
819 | // profitable with the expansion for 64-bit since it's generally good to | |||
820 | // speculate things. | |||
821 | // FIXME: These should really have the size as a parameter. | |||
822 | bool AMDGPUTargetLowering::isCheapToSpeculateCttz() const { | |||
823 | return true; | |||
824 | } | |||
825 | ||||
826 | bool AMDGPUTargetLowering::isCheapToSpeculateCtlz() const { | |||
827 | return true; | |||
828 | } | |||
829 | ||||
830 | bool AMDGPUTargetLowering::isSDNodeAlwaysUniform(const SDNode *N) const { | |||
831 | switch (N->getOpcode()) { | |||
832 | case ISD::EntryToken: | |||
833 | case ISD::TokenFactor: | |||
834 | return true; | |||
835 | case ISD::INTRINSIC_WO_CHAIN: { | |||
836 | unsigned IntrID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); | |||
837 | switch (IntrID) { | |||
838 | case Intrinsic::amdgcn_readfirstlane: | |||
839 | case Intrinsic::amdgcn_readlane: | |||
840 | return true; | |||
841 | } | |||
842 | return false; | |||
843 | } | |||
844 | case ISD::LOAD: | |||
845 | if (cast<LoadSDNode>(N)->getMemOperand()->getAddrSpace() == | |||
846 | AMDGPUAS::CONSTANT_ADDRESS_32BIT) | |||
847 | return true; | |||
848 | return false; | |||
849 | case AMDGPUISD::SETCC: // ballot-style instruction | |||
850 | return true; | |||
851 | } | |||
852 | return false; | |||
853 | } | |||
854 | ||||
855 | SDValue AMDGPUTargetLowering::getNegatedExpression( | |||
856 | SDValue Op, SelectionDAG &DAG, bool LegalOperations, bool ForCodeSize, | |||
857 | NegatibleCost &Cost, unsigned Depth) const { | |||
858 | ||||
859 | switch (Op.getOpcode()) { | |||
860 | case ISD::FMA: | |||
861 | case ISD::FMAD: { | |||
862 | // Negating a fma is not free if it has users without source mods. | |||
863 | if (!allUsesHaveSourceMods(Op.getNode())) | |||
864 | return SDValue(); | |||
865 | break; | |||
866 | } | |||
867 | default: | |||
868 | break; | |||
869 | } | |||
870 | ||||
871 | return TargetLowering::getNegatedExpression(Op, DAG, LegalOperations, | |||
872 | ForCodeSize, Cost, Depth); | |||
873 | } | |||
874 | ||||
875 | //===---------------------------------------------------------------------===// | |||
876 | // Target Properties | |||
877 | //===---------------------------------------------------------------------===// | |||
878 | ||||
879 | bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const { | |||
880 | assert(VT.isFloatingPoint())(static_cast <bool> (VT.isFloatingPoint()) ? void (0) : __assert_fail ("VT.isFloatingPoint()", "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp" , 880, __extension__ __PRETTY_FUNCTION__)); | |||
881 | ||||
882 | // Packed operations do not have a fabs modifier. | |||
883 | return VT == MVT::f32 || VT == MVT::f64 || | |||
884 | (Subtarget->has16BitInsts() && VT == MVT::f16); | |||
885 | } | |||
886 | ||||
887 | bool AMDGPUTargetLowering::isFNegFree(EVT VT) const { | |||
888 | assert(VT.isFloatingPoint())(static_cast <bool> (VT.isFloatingPoint()) ? void (0) : __assert_fail ("VT.isFloatingPoint()", "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp" , 888, __extension__ __PRETTY_FUNCTION__)); | |||
889 | // Report this based on the end legalized type. | |||
890 | VT = VT.getScalarType(); | |||
891 | return VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f16; | |||
892 | } | |||
893 | ||||
894 | bool AMDGPUTargetLowering:: storeOfVectorConstantIsCheap(EVT MemVT, | |||
895 | unsigned NumElem, | |||
896 | unsigned AS) const { | |||
897 | return true; | |||
898 | } | |||
899 | ||||
900 | bool AMDGPUTargetLowering::aggressivelyPreferBuildVectorSources(EVT VecVT) const { | |||
901 | // There are few operations which truly have vector input operands. Any vector | |||
902 | // operation is going to involve operations on each component, and a | |||
903 | // build_vector will be a copy per element, so it always makes sense to use a | |||
904 | // build_vector input in place of the extracted element to avoid a copy into a | |||
905 | // super register. | |||
906 | // | |||
907 | // We should probably only do this if all users are extracts only, but this | |||
908 | // should be the common case. | |||
909 | return true; | |||
910 | } | |||
911 | ||||
912 | bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const { | |||
913 | // Truncate is just accessing a subregister. | |||
914 | ||||
915 | unsigned SrcSize = Source.getSizeInBits(); | |||
916 | unsigned DestSize = Dest.getSizeInBits(); | |||
917 | ||||
918 | return DestSize < SrcSize && DestSize % 32 == 0 ; | |||
919 | } | |||
920 | ||||
921 | bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const { | |||
922 | // Truncate is just accessing a subregister. | |||
923 | ||||
924 | unsigned SrcSize = Source->getScalarSizeInBits(); | |||
925 | unsigned DestSize = Dest->getScalarSizeInBits(); | |||
926 | ||||
927 | if (DestSize== 16 && Subtarget->has16BitInsts()) | |||
928 | return SrcSize >= 32; | |||
929 | ||||
930 | return DestSize < SrcSize && DestSize % 32 == 0; | |||
931 | } | |||
932 | ||||
933 | bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const { | |||
934 | unsigned SrcSize = Src->getScalarSizeInBits(); | |||
935 | unsigned DestSize = Dest->getScalarSizeInBits(); | |||
936 | ||||
937 | if (SrcSize == 16 && Subtarget->has16BitInsts()) | |||
938 | return DestSize >= 32; | |||
939 | ||||
940 | return SrcSize == 32 && DestSize == 64; | |||
941 | } | |||
942 | ||||
943 | bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const { | |||
944 | // Any register load of a 64-bit value really requires 2 32-bit moves. For all | |||
945 | // practical purposes, the extra mov 0 to load a 64-bit is free. As used, | |||
946 | // this will enable reducing 64-bit operations the 32-bit, which is always | |||
947 | // good. | |||
948 | ||||
949 | if (Src == MVT::i16) | |||
950 | return Dest == MVT::i32 ||Dest == MVT::i64 ; | |||
951 | ||||
952 | return Src == MVT::i32 && Dest == MVT::i64; | |||
953 | } | |||
954 | ||||
955 | bool AMDGPUTargetLowering::isZExtFree(SDValue Val, EVT VT2) const { | |||
956 | return isZExtFree(Val.getValueType(), VT2); | |||
957 | } | |||
958 | ||||
959 | bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const { | |||
960 | // There aren't really 64-bit registers, but pairs of 32-bit ones and only a | |||
961 | // limited number of native 64-bit operations. Shrinking an operation to fit | |||
962 | // in a single 32-bit register should always be helpful. As currently used, | |||
963 | // this is much less general than the name suggests, and is only used in | |||
964 | // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is | |||
965 | // not profitable, and may actually be harmful. | |||
966 | return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32; | |||
967 | } | |||
968 | ||||
969 | //===---------------------------------------------------------------------===// | |||
970 | // TargetLowering Callbacks | |||
971 | //===---------------------------------------------------------------------===// | |||
972 | ||||
973 | CCAssignFn *AMDGPUCallLowering::CCAssignFnForCall(CallingConv::ID CC, | |||
974 | bool IsVarArg) { | |||
975 | switch (CC) { | |||
976 | case CallingConv::AMDGPU_VS: | |||
977 | case CallingConv::AMDGPU_GS: | |||
978 | case CallingConv::AMDGPU_PS: | |||
979 | case CallingConv::AMDGPU_CS: | |||
980 | case CallingConv::AMDGPU_HS: | |||
981 | case CallingConv::AMDGPU_ES: | |||
982 | case CallingConv::AMDGPU_LS: | |||
983 | return CC_AMDGPU; | |||
984 | case CallingConv::C: | |||
985 | case CallingConv::Fast: | |||
986 | case CallingConv::Cold: | |||
987 | return CC_AMDGPU_Func; | |||
988 | case CallingConv::AMDGPU_Gfx: | |||
989 | return CC_SI_Gfx; | |||
990 | case CallingConv::AMDGPU_KERNEL: | |||
991 | case CallingConv::SPIR_KERNEL: | |||
992 | default: | |||
993 | report_fatal_error("Unsupported calling convention for call"); | |||
994 | } | |||
995 | } | |||
996 | ||||
997 | CCAssignFn *AMDGPUCallLowering::CCAssignFnForReturn(CallingConv::ID CC, | |||
998 | bool IsVarArg) { | |||
999 | switch (CC) { | |||
1000 | case CallingConv::AMDGPU_KERNEL: | |||
1001 | case CallingConv::SPIR_KERNEL: | |||
1002 | llvm_unreachable("kernels should not be handled here")::llvm::llvm_unreachable_internal("kernels should not be handled here" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 1002); | |||
1003 | case CallingConv::AMDGPU_VS: | |||
1004 | case CallingConv::AMDGPU_GS: | |||
1005 | case CallingConv::AMDGPU_PS: | |||
1006 | case CallingConv::AMDGPU_CS: | |||
1007 | case CallingConv::AMDGPU_HS: | |||
1008 | case CallingConv::AMDGPU_ES: | |||
1009 | case CallingConv::AMDGPU_LS: | |||
1010 | return RetCC_SI_Shader; | |||
1011 | case CallingConv::AMDGPU_Gfx: | |||
1012 | return RetCC_SI_Gfx; | |||
1013 | case CallingConv::C: | |||
1014 | case CallingConv::Fast: | |||
1015 | case CallingConv::Cold: | |||
1016 | return RetCC_AMDGPU_Func; | |||
1017 | default: | |||
1018 | report_fatal_error("Unsupported calling convention."); | |||
1019 | } | |||
1020 | } | |||
1021 | ||||
1022 | /// The SelectionDAGBuilder will automatically promote function arguments | |||
1023 | /// with illegal types. However, this does not work for the AMDGPU targets | |||
1024 | /// since the function arguments are stored in memory as these illegal types. | |||
1025 | /// In order to handle this properly we need to get the original types sizes | |||
1026 | /// from the LLVM IR Function and fixup the ISD:InputArg values before | |||
1027 | /// passing them to AnalyzeFormalArguments() | |||
1028 | ||||
1029 | /// When the SelectionDAGBuilder computes the Ins, it takes care of splitting | |||
1030 | /// input values across multiple registers. Each item in the Ins array | |||
1031 | /// represents a single value that will be stored in registers. Ins[x].VT is | |||
1032 | /// the value type of the value that will be stored in the register, so | |||
1033 | /// whatever SDNode we lower the argument to needs to be this type. | |||
1034 | /// | |||
1035 | /// In order to correctly lower the arguments we need to know the size of each | |||
1036 | /// argument. Since Ins[x].VT gives us the size of the register that will | |||
1037 | /// hold the value, we need to look at Ins[x].ArgVT to see the 'real' type | |||
1038 | /// for the original function argument so that we can deduce the correct memory | |||
1039 | /// type to use for Ins[x]. In most cases the correct memory type will be | |||
1040 | /// Ins[x].ArgVT. However, this will not always be the case. If, for example, | |||
1041 | /// we have a kernel argument of type v8i8, this argument will be split into | |||
1042 | /// 8 parts and each part will be represented by its own item in the Ins array. | |||
1043 | /// For each part the Ins[x].ArgVT will be the v8i8, which is the full type of | |||
1044 | /// the argument before it was split. From this, we deduce that the memory type | |||
1045 | /// for each individual part is i8. We pass the memory type as LocVT to the | |||
1046 | /// calling convention analysis function and the register type (Ins[x].VT) as | |||
1047 | /// the ValVT. | |||
1048 | void AMDGPUTargetLowering::analyzeFormalArgumentsCompute( | |||
1049 | CCState &State, | |||
1050 | const SmallVectorImpl<ISD::InputArg> &Ins) const { | |||
1051 | const MachineFunction &MF = State.getMachineFunction(); | |||
1052 | const Function &Fn = MF.getFunction(); | |||
1053 | LLVMContext &Ctx = Fn.getParent()->getContext(); | |||
1054 | const AMDGPUSubtarget &ST = AMDGPUSubtarget::get(MF); | |||
1055 | const unsigned ExplicitOffset = ST.getExplicitKernelArgOffset(Fn); | |||
1056 | CallingConv::ID CC = Fn.getCallingConv(); | |||
1057 | ||||
1058 | Align MaxAlign = Align(1); | |||
1059 | uint64_t ExplicitArgOffset = 0; | |||
1060 | const DataLayout &DL = Fn.getParent()->getDataLayout(); | |||
1061 | ||||
1062 | unsigned InIndex = 0; | |||
1063 | ||||
1064 | for (const Argument &Arg : Fn.args()) { | |||
1065 | const bool IsByRef = Arg.hasByRefAttr(); | |||
1066 | Type *BaseArgTy = Arg.getType(); | |||
1067 | Type *MemArgTy = IsByRef ? Arg.getParamByRefType() : BaseArgTy; | |||
1068 | MaybeAlign Alignment = IsByRef ? Arg.getParamAlign() : None; | |||
1069 | if (!Alignment) | |||
1070 | Alignment = DL.getABITypeAlign(MemArgTy); | |||
1071 | MaxAlign = max(Alignment, MaxAlign); | |||
1072 | uint64_t AllocSize = DL.getTypeAllocSize(MemArgTy); | |||
1073 | ||||
1074 | uint64_t ArgOffset = alignTo(ExplicitArgOffset, Alignment) + ExplicitOffset; | |||
1075 | ExplicitArgOffset = alignTo(ExplicitArgOffset, Alignment) + AllocSize; | |||
1076 | ||||
1077 | // We're basically throwing away everything passed into us and starting over | |||
1078 | // to get accurate in-memory offsets. The "PartOffset" is completely useless | |||
1079 | // to us as computed in Ins. | |||
1080 | // | |||
1081 | // We also need to figure out what type legalization is trying to do to get | |||
1082 | // the correct memory offsets. | |||
1083 | ||||
1084 | SmallVector<EVT, 16> ValueVTs; | |||
1085 | SmallVector<uint64_t, 16> Offsets; | |||
1086 | ComputeValueVTs(*this, DL, BaseArgTy, ValueVTs, &Offsets, ArgOffset); | |||
1087 | ||||
1088 | for (unsigned Value = 0, NumValues = ValueVTs.size(); | |||
1089 | Value != NumValues; ++Value) { | |||
1090 | uint64_t BasePartOffset = Offsets[Value]; | |||
1091 | ||||
1092 | EVT ArgVT = ValueVTs[Value]; | |||
1093 | EVT MemVT = ArgVT; | |||
1094 | MVT RegisterVT = getRegisterTypeForCallingConv(Ctx, CC, ArgVT); | |||
1095 | unsigned NumRegs = getNumRegistersForCallingConv(Ctx, CC, ArgVT); | |||
1096 | ||||
1097 | if (NumRegs == 1) { | |||
1098 | // This argument is not split, so the IR type is the memory type. | |||
1099 | if (ArgVT.isExtended()) { | |||
1100 | // We have an extended type, like i24, so we should just use the | |||
1101 | // register type. | |||
1102 | MemVT = RegisterVT; | |||
1103 | } else { | |||
1104 | MemVT = ArgVT; | |||
1105 | } | |||
1106 | } else if (ArgVT.isVector() && RegisterVT.isVector() && | |||
1107 | ArgVT.getScalarType() == RegisterVT.getScalarType()) { | |||
1108 | assert(ArgVT.getVectorNumElements() > RegisterVT.getVectorNumElements())(static_cast <bool> (ArgVT.getVectorNumElements() > RegisterVT .getVectorNumElements()) ? void (0) : __assert_fail ("ArgVT.getVectorNumElements() > RegisterVT.getVectorNumElements()" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 1108, __extension__ __PRETTY_FUNCTION__)); | |||
1109 | // We have a vector value which has been split into a vector with | |||
1110 | // the same scalar type, but fewer elements. This should handle | |||
1111 | // all the floating-point vector types. | |||
1112 | MemVT = RegisterVT; | |||
1113 | } else if (ArgVT.isVector() && | |||
1114 | ArgVT.getVectorNumElements() == NumRegs) { | |||
1115 | // This arg has been split so that each element is stored in a separate | |||
1116 | // register. | |||
1117 | MemVT = ArgVT.getScalarType(); | |||
1118 | } else if (ArgVT.isExtended()) { | |||
1119 | // We have an extended type, like i65. | |||
1120 | MemVT = RegisterVT; | |||
1121 | } else { | |||
1122 | unsigned MemoryBits = ArgVT.getStoreSizeInBits() / NumRegs; | |||
1123 | assert(ArgVT.getStoreSizeInBits() % NumRegs == 0)(static_cast <bool> (ArgVT.getStoreSizeInBits() % NumRegs == 0) ? void (0) : __assert_fail ("ArgVT.getStoreSizeInBits() % NumRegs == 0" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 1123, __extension__ __PRETTY_FUNCTION__)); | |||
1124 | if (RegisterVT.isInteger()) { | |||
1125 | MemVT = EVT::getIntegerVT(State.getContext(), MemoryBits); | |||
1126 | } else if (RegisterVT.isVector()) { | |||
1127 | assert(!RegisterVT.getScalarType().isFloatingPoint())(static_cast <bool> (!RegisterVT.getScalarType().isFloatingPoint ()) ? void (0) : __assert_fail ("!RegisterVT.getScalarType().isFloatingPoint()" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 1127, __extension__ __PRETTY_FUNCTION__)); | |||
1128 | unsigned NumElements = RegisterVT.getVectorNumElements(); | |||
1129 | assert(MemoryBits % NumElements == 0)(static_cast <bool> (MemoryBits % NumElements == 0) ? void (0) : __assert_fail ("MemoryBits % NumElements == 0", "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp" , 1129, __extension__ __PRETTY_FUNCTION__)); | |||
1130 | // This vector type has been split into another vector type with | |||
1131 | // a different elements size. | |||
1132 | EVT ScalarVT = EVT::getIntegerVT(State.getContext(), | |||
1133 | MemoryBits / NumElements); | |||
1134 | MemVT = EVT::getVectorVT(State.getContext(), ScalarVT, NumElements); | |||
1135 | } else { | |||
1136 | llvm_unreachable("cannot deduce memory type.")::llvm::llvm_unreachable_internal("cannot deduce memory type." , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 1136); | |||
1137 | } | |||
1138 | } | |||
1139 | ||||
1140 | // Convert one element vectors to scalar. | |||
1141 | if (MemVT.isVector() && MemVT.getVectorNumElements() == 1) | |||
1142 | MemVT = MemVT.getScalarType(); | |||
1143 | ||||
1144 | // Round up vec3/vec5 argument. | |||
1145 | if (MemVT.isVector() && !MemVT.isPow2VectorType()) { | |||
1146 | assert(MemVT.getVectorNumElements() == 3 ||(static_cast <bool> (MemVT.getVectorNumElements() == 3 || MemVT.getVectorNumElements() == 5) ? void (0) : __assert_fail ("MemVT.getVectorNumElements() == 3 || MemVT.getVectorNumElements() == 5" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 1147, __extension__ __PRETTY_FUNCTION__)) | |||
1147 | MemVT.getVectorNumElements() == 5)(static_cast <bool> (MemVT.getVectorNumElements() == 3 || MemVT.getVectorNumElements() == 5) ? void (0) : __assert_fail ("MemVT.getVectorNumElements() == 3 || MemVT.getVectorNumElements() == 5" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 1147, __extension__ __PRETTY_FUNCTION__)); | |||
1148 | MemVT = MemVT.getPow2VectorType(State.getContext()); | |||
1149 | } else if (!MemVT.isSimple() && !MemVT.isVector()) { | |||
1150 | MemVT = MemVT.getRoundIntegerType(State.getContext()); | |||
1151 | } | |||
1152 | ||||
1153 | unsigned PartOffset = 0; | |||
1154 | for (unsigned i = 0; i != NumRegs; ++i) { | |||
1155 | State.addLoc(CCValAssign::getCustomMem(InIndex++, RegisterVT, | |||
1156 | BasePartOffset + PartOffset, | |||
1157 | MemVT.getSimpleVT(), | |||
1158 | CCValAssign::Full)); | |||
1159 | PartOffset += MemVT.getStoreSize(); | |||
1160 | } | |||
1161 | } | |||
1162 | } | |||
1163 | } | |||
1164 | ||||
1165 | SDValue AMDGPUTargetLowering::LowerReturn( | |||
1166 | SDValue Chain, CallingConv::ID CallConv, | |||
1167 | bool isVarArg, | |||
1168 | const SmallVectorImpl<ISD::OutputArg> &Outs, | |||
1169 | const SmallVectorImpl<SDValue> &OutVals, | |||
1170 | const SDLoc &DL, SelectionDAG &DAG) const { | |||
1171 | // FIXME: Fails for r600 tests | |||
1172 | //assert(!isVarArg && Outs.empty() && OutVals.empty() && | |||
1173 | // "wave terminate should not have return values"); | |||
1174 | return DAG.getNode(AMDGPUISD::ENDPGM, DL, MVT::Other, Chain); | |||
1175 | } | |||
1176 | ||||
1177 | //===---------------------------------------------------------------------===// | |||
1178 | // Target specific lowering | |||
1179 | //===---------------------------------------------------------------------===// | |||
1180 | ||||
1181 | /// Selects the correct CCAssignFn for a given CallingConvention value. | |||
1182 | CCAssignFn *AMDGPUTargetLowering::CCAssignFnForCall(CallingConv::ID CC, | |||
1183 | bool IsVarArg) { | |||
1184 | return AMDGPUCallLowering::CCAssignFnForCall(CC, IsVarArg); | |||
1185 | } | |||
1186 | ||||
1187 | CCAssignFn *AMDGPUTargetLowering::CCAssignFnForReturn(CallingConv::ID CC, | |||
1188 | bool IsVarArg) { | |||
1189 | return AMDGPUCallLowering::CCAssignFnForReturn(CC, IsVarArg); | |||
1190 | } | |||
1191 | ||||
1192 | SDValue AMDGPUTargetLowering::addTokenForArgument(SDValue Chain, | |||
1193 | SelectionDAG &DAG, | |||
1194 | MachineFrameInfo &MFI, | |||
1195 | int ClobberedFI) const { | |||
1196 | SmallVector<SDValue, 8> ArgChains; | |||
1197 | int64_t FirstByte = MFI.getObjectOffset(ClobberedFI); | |||
1198 | int64_t LastByte = FirstByte + MFI.getObjectSize(ClobberedFI) - 1; | |||
1199 | ||||
1200 | // Include the original chain at the beginning of the list. When this is | |||
1201 | // used by target LowerCall hooks, this helps legalize find the | |||
1202 | // CALLSEQ_BEGIN node. | |||
1203 | ArgChains.push_back(Chain); | |||
1204 | ||||
1205 | // Add a chain value for each stack argument corresponding | |||
1206 | for (SDNode *U : DAG.getEntryNode().getNode()->uses()) { | |||
1207 | if (LoadSDNode *L = dyn_cast<LoadSDNode>(U)) { | |||
1208 | if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr())) { | |||
1209 | if (FI->getIndex() < 0) { | |||
1210 | int64_t InFirstByte = MFI.getObjectOffset(FI->getIndex()); | |||
1211 | int64_t InLastByte = InFirstByte; | |||
1212 | InLastByte += MFI.getObjectSize(FI->getIndex()) - 1; | |||
1213 | ||||
1214 | if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) || | |||
1215 | (FirstByte <= InFirstByte && InFirstByte <= LastByte)) | |||
1216 | ArgChains.push_back(SDValue(L, 1)); | |||
1217 | } | |||
1218 | } | |||
1219 | } | |||
1220 | } | |||
1221 | ||||
1222 | // Build a tokenfactor for all the chains. | |||
1223 | return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains); | |||
1224 | } | |||
1225 | ||||
1226 | SDValue AMDGPUTargetLowering::lowerUnhandledCall(CallLoweringInfo &CLI, | |||
1227 | SmallVectorImpl<SDValue> &InVals, | |||
1228 | StringRef Reason) const { | |||
1229 | SDValue Callee = CLI.Callee; | |||
1230 | SelectionDAG &DAG = CLI.DAG; | |||
1231 | ||||
1232 | const Function &Fn = DAG.getMachineFunction().getFunction(); | |||
1233 | ||||
1234 | StringRef FuncName("<unknown>"); | |||
1235 | ||||
1236 | if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee)) | |||
1237 | FuncName = G->getSymbol(); | |||
1238 | else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) | |||
1239 | FuncName = G->getGlobal()->getName(); | |||
1240 | ||||
1241 | DiagnosticInfoUnsupported NoCalls( | |||
1242 | Fn, Reason + FuncName, CLI.DL.getDebugLoc()); | |||
1243 | DAG.getContext()->diagnose(NoCalls); | |||
1244 | ||||
1245 | if (!CLI.IsTailCall) { | |||
1246 | for (unsigned I = 0, E = CLI.Ins.size(); I != E; ++I) | |||
1247 | InVals.push_back(DAG.getUNDEF(CLI.Ins[I].VT)); | |||
1248 | } | |||
1249 | ||||
1250 | return DAG.getEntryNode(); | |||
1251 | } | |||
1252 | ||||
1253 | SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI, | |||
1254 | SmallVectorImpl<SDValue> &InVals) const { | |||
1255 | return lowerUnhandledCall(CLI, InVals, "unsupported call to function "); | |||
1256 | } | |||
1257 | ||||
1258 | SDValue AMDGPUTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, | |||
1259 | SelectionDAG &DAG) const { | |||
1260 | const Function &Fn = DAG.getMachineFunction().getFunction(); | |||
1261 | ||||
1262 | DiagnosticInfoUnsupported NoDynamicAlloca(Fn, "unsupported dynamic alloca", | |||
1263 | SDLoc(Op).getDebugLoc()); | |||
1264 | DAG.getContext()->diagnose(NoDynamicAlloca); | |||
1265 | auto Ops = {DAG.getConstant(0, SDLoc(), Op.getValueType()), Op.getOperand(0)}; | |||
1266 | return DAG.getMergeValues(Ops, SDLoc()); | |||
1267 | } | |||
1268 | ||||
1269 | SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, | |||
1270 | SelectionDAG &DAG) const { | |||
1271 | switch (Op.getOpcode()) { | |||
1272 | default: | |||
1273 | Op->print(errs(), &DAG); | |||
1274 | llvm_unreachable("Custom lowering code for this "::llvm::llvm_unreachable_internal("Custom lowering code for this " "instruction is not implemented yet!", "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp" , 1275) | |||
1275 | "instruction is not implemented yet!")::llvm::llvm_unreachable_internal("Custom lowering code for this " "instruction is not implemented yet!", "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp" , 1275); | |||
1276 | break; | |||
1277 | case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG); | |||
1278 | case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG); | |||
1279 | case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG); | |||
1280 | case ISD::UDIVREM: return LowerUDIVREM(Op, DAG); | |||
1281 | case ISD::SDIVREM: return LowerSDIVREM(Op, DAG); | |||
1282 | case ISD::FREM: return LowerFREM(Op, DAG); | |||
1283 | case ISD::FCEIL: return LowerFCEIL(Op, DAG); | |||
1284 | case ISD::FTRUNC: return LowerFTRUNC(Op, DAG); | |||
1285 | case ISD::FRINT: return LowerFRINT(Op, DAG); | |||
1286 | case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG); | |||
1287 | case ISD::FROUND: return LowerFROUND(Op, DAG); | |||
1288 | case ISD::FFLOOR: return LowerFFLOOR(Op, DAG); | |||
1289 | case ISD::FLOG: | |||
1290 | return LowerFLOG(Op, DAG, numbers::ln2f); | |||
1291 | case ISD::FLOG10: | |||
1292 | return LowerFLOG(Op, DAG, numbers::ln2f / numbers::ln10f); | |||
1293 | case ISD::FEXP: | |||
1294 | return lowerFEXP(Op, DAG); | |||
1295 | case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG); | |||
1296 | case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG); | |||
1297 | case ISD::FP_TO_FP16: return LowerFP_TO_FP16(Op, DAG); | |||
1298 | case ISD::FP_TO_SINT: | |||
1299 | case ISD::FP_TO_UINT: | |||
1300 | return LowerFP_TO_INT(Op, DAG); | |||
1301 | case ISD::CTTZ: | |||
1302 | case ISD::CTTZ_ZERO_UNDEF: | |||
1303 | case ISD::CTLZ: | |||
1304 | case ISD::CTLZ_ZERO_UNDEF: | |||
1305 | return LowerCTLZ_CTTZ(Op, DAG); | |||
1306 | case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG); | |||
1307 | } | |||
1308 | return Op; | |||
1309 | } | |||
1310 | ||||
1311 | void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N, | |||
1312 | SmallVectorImpl<SDValue> &Results, | |||
1313 | SelectionDAG &DAG) const { | |||
1314 | switch (N->getOpcode()) { | |||
1315 | case ISD::SIGN_EXTEND_INREG: | |||
1316 | // Different parts of legalization seem to interpret which type of | |||
1317 | // sign_extend_inreg is the one to check for custom lowering. The extended | |||
1318 | // from type is what really matters, but some places check for custom | |||
1319 | // lowering of the result type. This results in trying to use | |||
1320 | // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do | |||
1321 | // nothing here and let the illegal result integer be handled normally. | |||
1322 | return; | |||
1323 | default: | |||
1324 | return; | |||
1325 | } | |||
1326 | } | |||
1327 | ||||
1328 | SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI, | |||
1329 | SDValue Op, | |||
1330 | SelectionDAG &DAG) const { | |||
1331 | ||||
1332 | const DataLayout &DL = DAG.getDataLayout(); | |||
1333 | GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op); | |||
1334 | const GlobalValue *GV = G->getGlobal(); | |||
1335 | ||||
1336 | if (G->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS || | |||
1337 | G->getAddressSpace() == AMDGPUAS::REGION_ADDRESS) { | |||
1338 | if (!MFI->isModuleEntryFunction() && | |||
1339 | !GV->getName().equals("llvm.amdgcn.module.lds")) { | |||
1340 | SDLoc DL(Op); | |||
1341 | const Function &Fn = DAG.getMachineFunction().getFunction(); | |||
1342 | DiagnosticInfoUnsupported BadLDSDecl( | |||
1343 | Fn, "local memory global used by non-kernel function", | |||
1344 | DL.getDebugLoc(), DS_Warning); | |||
1345 | DAG.getContext()->diagnose(BadLDSDecl); | |||
1346 | ||||
1347 | // We currently don't have a way to correctly allocate LDS objects that | |||
1348 | // aren't directly associated with a kernel. We do force inlining of | |||
1349 | // functions that use local objects. However, if these dead functions are | |||
1350 | // not eliminated, we don't want a compile time error. Just emit a warning | |||
1351 | // and a trap, since there should be no callable path here. | |||
1352 | SDValue Trap = DAG.getNode(ISD::TRAP, DL, MVT::Other, DAG.getEntryNode()); | |||
1353 | SDValue OutputChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, | |||
1354 | Trap, DAG.getRoot()); | |||
1355 | DAG.setRoot(OutputChain); | |||
1356 | return DAG.getUNDEF(Op.getValueType()); | |||
1357 | } | |||
1358 | ||||
1359 | // XXX: What does the value of G->getOffset() mean? | |||
1360 | assert(G->getOffset() == 0 &&(static_cast <bool> (G->getOffset() == 0 && "Do not know what to do with an non-zero offset" ) ? void (0) : __assert_fail ("G->getOffset() == 0 && \"Do not know what to do with an non-zero offset\"" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 1361, __extension__ __PRETTY_FUNCTION__)) | |||
1361 | "Do not know what to do with an non-zero offset")(static_cast <bool> (G->getOffset() == 0 && "Do not know what to do with an non-zero offset" ) ? void (0) : __assert_fail ("G->getOffset() == 0 && \"Do not know what to do with an non-zero offset\"" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 1361, __extension__ __PRETTY_FUNCTION__)); | |||
1362 | ||||
1363 | // TODO: We could emit code to handle the initialization somewhere. | |||
1364 | // We ignore the initializer for now and legalize it to allow selection. | |||
1365 | // The initializer will anyway get errored out during assembly emission. | |||
1366 | unsigned Offset = MFI->allocateLDSGlobal(DL, *cast<GlobalVariable>(GV)); | |||
1367 | return DAG.getConstant(Offset, SDLoc(Op), Op.getValueType()); | |||
1368 | } | |||
1369 | return SDValue(); | |||
1370 | } | |||
1371 | ||||
1372 | SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op, | |||
1373 | SelectionDAG &DAG) const { | |||
1374 | SmallVector<SDValue, 8> Args; | |||
1375 | ||||
1376 | EVT VT = Op.getValueType(); | |||
1377 | if (VT == MVT::v4i16 || VT == MVT::v4f16) { | |||
1378 | SDLoc SL(Op); | |||
1379 | SDValue Lo = DAG.getNode(ISD::BITCAST, SL, MVT::i32, Op.getOperand(0)); | |||
1380 | SDValue Hi = DAG.getNode(ISD::BITCAST, SL, MVT::i32, Op.getOperand(1)); | |||
1381 | ||||
1382 | SDValue BV = DAG.getBuildVector(MVT::v2i32, SL, { Lo, Hi }); | |||
1383 | return DAG.getNode(ISD::BITCAST, SL, VT, BV); | |||
1384 | } | |||
1385 | ||||
1386 | for (const SDUse &U : Op->ops()) | |||
1387 | DAG.ExtractVectorElements(U.get(), Args); | |||
1388 | ||||
1389 | return DAG.getBuildVector(Op.getValueType(), SDLoc(Op), Args); | |||
1390 | } | |||
1391 | ||||
1392 | SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op, | |||
1393 | SelectionDAG &DAG) const { | |||
1394 | ||||
1395 | SmallVector<SDValue, 8> Args; | |||
1396 | unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue(); | |||
1397 | EVT VT = Op.getValueType(); | |||
1398 | EVT SrcVT = Op.getOperand(0).getValueType(); | |||
1399 | ||||
1400 | // For these types, we have some TableGen patterns except if the index is 1 | |||
1401 | if (((SrcVT == MVT::v4f16 && VT == MVT::v2f16) || | |||
1402 | (SrcVT == MVT::v4i16 && VT == MVT::v2i16)) && | |||
1403 | Start != 1) | |||
1404 | return Op; | |||
1405 | ||||
1406 | if (((SrcVT == MVT::v8f16 && VT == MVT::v4f16) || | |||
1407 | (SrcVT == MVT::v8i16 && VT == MVT::v4i16)) && | |||
1408 | (Start == 0 || Start == 4)) | |||
1409 | return Op; | |||
1410 | ||||
1411 | DAG.ExtractVectorElements(Op.getOperand(0), Args, Start, | |||
1412 | VT.getVectorNumElements()); | |||
1413 | ||||
1414 | return DAG.getBuildVector(Op.getValueType(), SDLoc(Op), Args); | |||
1415 | } | |||
1416 | ||||
1417 | /// Generate Min/Max node | |||
1418 | SDValue AMDGPUTargetLowering::combineFMinMaxLegacy(const SDLoc &DL, EVT VT, | |||
1419 | SDValue LHS, SDValue RHS, | |||
1420 | SDValue True, SDValue False, | |||
1421 | SDValue CC, | |||
1422 | DAGCombinerInfo &DCI) const { | |||
1423 | if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True)) | |||
1424 | return SDValue(); | |||
1425 | ||||
1426 | SelectionDAG &DAG = DCI.DAG; | |||
1427 | ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get(); | |||
1428 | switch (CCOpcode) { | |||
1429 | case ISD::SETOEQ: | |||
1430 | case ISD::SETONE: | |||
1431 | case ISD::SETUNE: | |||
1432 | case ISD::SETNE: | |||
1433 | case ISD::SETUEQ: | |||
1434 | case ISD::SETEQ: | |||
1435 | case ISD::SETFALSE: | |||
1436 | case ISD::SETFALSE2: | |||
1437 | case ISD::SETTRUE: | |||
1438 | case ISD::SETTRUE2: | |||
1439 | case ISD::SETUO: | |||
1440 | case ISD::SETO: | |||
1441 | break; | |||
1442 | case ISD::SETULE: | |||
1443 | case ISD::SETULT: { | |||
1444 | if (LHS == True) | |||
1445 | return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS); | |||
1446 | return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS); | |||
1447 | } | |||
1448 | case ISD::SETOLE: | |||
1449 | case ISD::SETOLT: | |||
1450 | case ISD::SETLE: | |||
1451 | case ISD::SETLT: { | |||
1452 | // Ordered. Assume ordered for undefined. | |||
1453 | ||||
1454 | // Only do this after legalization to avoid interfering with other combines | |||
1455 | // which might occur. | |||
1456 | if (DCI.getDAGCombineLevel() < AfterLegalizeDAG && | |||
1457 | !DCI.isCalledByLegalizer()) | |||
1458 | return SDValue(); | |||
1459 | ||||
1460 | // We need to permute the operands to get the correct NaN behavior. The | |||
1461 | // selected operand is the second one based on the failing compare with NaN, | |||
1462 | // so permute it based on the compare type the hardware uses. | |||
1463 | if (LHS == True) | |||
1464 | return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS); | |||
1465 | return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS); | |||
1466 | } | |||
1467 | case ISD::SETUGE: | |||
1468 | case ISD::SETUGT: { | |||
1469 | if (LHS == True) | |||
1470 | return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS); | |||
1471 | return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS); | |||
1472 | } | |||
1473 | case ISD::SETGT: | |||
1474 | case ISD::SETGE: | |||
1475 | case ISD::SETOGE: | |||
1476 | case ISD::SETOGT: { | |||
1477 | if (DCI.getDAGCombineLevel() < AfterLegalizeDAG && | |||
1478 | !DCI.isCalledByLegalizer()) | |||
1479 | return SDValue(); | |||
1480 | ||||
1481 | if (LHS == True) | |||
1482 | return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS); | |||
1483 | return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS); | |||
1484 | } | |||
1485 | case ISD::SETCC_INVALID: | |||
1486 | llvm_unreachable("Invalid setcc condcode!")::llvm::llvm_unreachable_internal("Invalid setcc condcode!", "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp" , 1486); | |||
1487 | } | |||
1488 | return SDValue(); | |||
1489 | } | |||
1490 | ||||
1491 | std::pair<SDValue, SDValue> | |||
1492 | AMDGPUTargetLowering::split64BitValue(SDValue Op, SelectionDAG &DAG) const { | |||
1493 | SDLoc SL(Op); | |||
1494 | ||||
1495 | SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op); | |||
1496 | ||||
1497 | const SDValue Zero = DAG.getConstant(0, SL, MVT::i32); | |||
1498 | const SDValue One = DAG.getConstant(1, SL, MVT::i32); | |||
1499 | ||||
1500 | SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero); | |||
1501 | SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One); | |||
1502 | ||||
1503 | return std::make_pair(Lo, Hi); | |||
1504 | } | |||
1505 | ||||
1506 | SDValue AMDGPUTargetLowering::getLoHalf64(SDValue Op, SelectionDAG &DAG) const { | |||
1507 | SDLoc SL(Op); | |||
1508 | ||||
1509 | SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op); | |||
1510 | const SDValue Zero = DAG.getConstant(0, SL, MVT::i32); | |||
1511 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero); | |||
1512 | } | |||
1513 | ||||
1514 | SDValue AMDGPUTargetLowering::getHiHalf64(SDValue Op, SelectionDAG &DAG) const { | |||
1515 | SDLoc SL(Op); | |||
1516 | ||||
1517 | SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op); | |||
1518 | const SDValue One = DAG.getConstant(1, SL, MVT::i32); | |||
1519 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One); | |||
1520 | } | |||
1521 | ||||
1522 | // Split a vector type into two parts. The first part is a power of two vector. | |||
1523 | // The second part is whatever is left over, and is a scalar if it would | |||
1524 | // otherwise be a 1-vector. | |||
1525 | std::pair<EVT, EVT> | |||
1526 | AMDGPUTargetLowering::getSplitDestVTs(const EVT &VT, SelectionDAG &DAG) const { | |||
1527 | EVT LoVT, HiVT; | |||
1528 | EVT EltVT = VT.getVectorElementType(); | |||
1529 | unsigned NumElts = VT.getVectorNumElements(); | |||
1530 | unsigned LoNumElts = PowerOf2Ceil((NumElts + 1) / 2); | |||
1531 | LoVT = EVT::getVectorVT(*DAG.getContext(), EltVT, LoNumElts); | |||
1532 | HiVT = NumElts - LoNumElts == 1 | |||
1533 | ? EltVT | |||
1534 | : EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts - LoNumElts); | |||
1535 | return std::make_pair(LoVT, HiVT); | |||
1536 | } | |||
1537 | ||||
1538 | // Split a vector value into two parts of types LoVT and HiVT. HiVT could be | |||
1539 | // scalar. | |||
1540 | std::pair<SDValue, SDValue> | |||
1541 | AMDGPUTargetLowering::splitVector(const SDValue &N, const SDLoc &DL, | |||
1542 | const EVT &LoVT, const EVT &HiVT, | |||
1543 | SelectionDAG &DAG) const { | |||
1544 | assert(LoVT.getVectorNumElements() +(static_cast <bool> (LoVT.getVectorNumElements() + (HiVT .isVector() ? HiVT.getVectorNumElements() : 1) <= N.getValueType ().getVectorNumElements() && "More vector elements requested than available!" ) ? void (0) : __assert_fail ("LoVT.getVectorNumElements() + (HiVT.isVector() ? HiVT.getVectorNumElements() : 1) <= N.getValueType().getVectorNumElements() && \"More vector elements requested than available!\"" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 1547, __extension__ __PRETTY_FUNCTION__)) | |||
1545 | (HiVT.isVector() ? HiVT.getVectorNumElements() : 1) <=(static_cast <bool> (LoVT.getVectorNumElements() + (HiVT .isVector() ? HiVT.getVectorNumElements() : 1) <= N.getValueType ().getVectorNumElements() && "More vector elements requested than available!" ) ? void (0) : __assert_fail ("LoVT.getVectorNumElements() + (HiVT.isVector() ? HiVT.getVectorNumElements() : 1) <= N.getValueType().getVectorNumElements() && \"More vector elements requested than available!\"" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 1547, __extension__ __PRETTY_FUNCTION__)) | |||
1546 | N.getValueType().getVectorNumElements() &&(static_cast <bool> (LoVT.getVectorNumElements() + (HiVT .isVector() ? HiVT.getVectorNumElements() : 1) <= N.getValueType ().getVectorNumElements() && "More vector elements requested than available!" ) ? void (0) : __assert_fail ("LoVT.getVectorNumElements() + (HiVT.isVector() ? HiVT.getVectorNumElements() : 1) <= N.getValueType().getVectorNumElements() && \"More vector elements requested than available!\"" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 1547, __extension__ __PRETTY_FUNCTION__)) | |||
1547 | "More vector elements requested than available!")(static_cast <bool> (LoVT.getVectorNumElements() + (HiVT .isVector() ? HiVT.getVectorNumElements() : 1) <= N.getValueType ().getVectorNumElements() && "More vector elements requested than available!" ) ? void (0) : __assert_fail ("LoVT.getVectorNumElements() + (HiVT.isVector() ? HiVT.getVectorNumElements() : 1) <= N.getValueType().getVectorNumElements() && \"More vector elements requested than available!\"" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 1547, __extension__ __PRETTY_FUNCTION__)); | |||
1548 | SDValue Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, LoVT, N, | |||
1549 | DAG.getVectorIdxConstant(0, DL)); | |||
1550 | SDValue Hi = DAG.getNode( | |||
1551 | HiVT.isVector() ? ISD::EXTRACT_SUBVECTOR : ISD::EXTRACT_VECTOR_ELT, DL, | |||
1552 | HiVT, N, DAG.getVectorIdxConstant(LoVT.getVectorNumElements(), DL)); | |||
1553 | return std::make_pair(Lo, Hi); | |||
1554 | } | |||
1555 | ||||
1556 | SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue Op, | |||
1557 | SelectionDAG &DAG) const { | |||
1558 | LoadSDNode *Load = cast<LoadSDNode>(Op); | |||
1559 | EVT VT = Op.getValueType(); | |||
1560 | SDLoc SL(Op); | |||
1561 | ||||
1562 | ||||
1563 | // If this is a 2 element vector, we really want to scalarize and not create | |||
1564 | // weird 1 element vectors. | |||
1565 | if (VT.getVectorNumElements() == 2) { | |||
1566 | SDValue Ops[2]; | |||
1567 | std::tie(Ops[0], Ops[1]) = scalarizeVectorLoad(Load, DAG); | |||
1568 | return DAG.getMergeValues(Ops, SL); | |||
1569 | } | |||
1570 | ||||
1571 | SDValue BasePtr = Load->getBasePtr(); | |||
1572 | EVT MemVT = Load->getMemoryVT(); | |||
1573 | ||||
1574 | const MachinePointerInfo &SrcValue = Load->getMemOperand()->getPointerInfo(); | |||
1575 | ||||
1576 | EVT LoVT, HiVT; | |||
1577 | EVT LoMemVT, HiMemVT; | |||
1578 | SDValue Lo, Hi; | |||
1579 | ||||
1580 | std::tie(LoVT, HiVT) = getSplitDestVTs(VT, DAG); | |||
1581 | std::tie(LoMemVT, HiMemVT) = getSplitDestVTs(MemVT, DAG); | |||
1582 | std::tie(Lo, Hi) = splitVector(Op, SL, LoVT, HiVT, DAG); | |||
1583 | ||||
1584 | unsigned Size = LoMemVT.getStoreSize(); | |||
1585 | unsigned BaseAlign = Load->getAlignment(); | |||
1586 | unsigned HiAlign = MinAlign(BaseAlign, Size); | |||
1587 | ||||
1588 | SDValue LoLoad = DAG.getExtLoad(Load->getExtensionType(), SL, LoVT, | |||
1589 | Load->getChain(), BasePtr, SrcValue, LoMemVT, | |||
1590 | BaseAlign, Load->getMemOperand()->getFlags()); | |||
1591 | SDValue HiPtr = DAG.getObjectPtrOffset(SL, BasePtr, TypeSize::Fixed(Size)); | |||
1592 | SDValue HiLoad = | |||
1593 | DAG.getExtLoad(Load->getExtensionType(), SL, HiVT, Load->getChain(), | |||
1594 | HiPtr, SrcValue.getWithOffset(LoMemVT.getStoreSize()), | |||
1595 | HiMemVT, HiAlign, Load->getMemOperand()->getFlags()); | |||
1596 | ||||
1597 | SDValue Join; | |||
1598 | if (LoVT == HiVT) { | |||
1599 | // This is the case that the vector is power of two so was evenly split. | |||
1600 | Join = DAG.getNode(ISD::CONCAT_VECTORS, SL, VT, LoLoad, HiLoad); | |||
1601 | } else { | |||
1602 | Join = DAG.getNode(ISD::INSERT_SUBVECTOR, SL, VT, DAG.getUNDEF(VT), LoLoad, | |||
1603 | DAG.getVectorIdxConstant(0, SL)); | |||
1604 | Join = DAG.getNode( | |||
1605 | HiVT.isVector() ? ISD::INSERT_SUBVECTOR : ISD::INSERT_VECTOR_ELT, SL, | |||
1606 | VT, Join, HiLoad, | |||
1607 | DAG.getVectorIdxConstant(LoVT.getVectorNumElements(), SL)); | |||
1608 | } | |||
1609 | ||||
1610 | SDValue Ops[] = {Join, DAG.getNode(ISD::TokenFactor, SL, MVT::Other, | |||
1611 | LoLoad.getValue(1), HiLoad.getValue(1))}; | |||
1612 | ||||
1613 | return DAG.getMergeValues(Ops, SL); | |||
1614 | } | |||
1615 | ||||
1616 | SDValue AMDGPUTargetLowering::WidenOrSplitVectorLoad(SDValue Op, | |||
1617 | SelectionDAG &DAG) const { | |||
1618 | LoadSDNode *Load = cast<LoadSDNode>(Op); | |||
1619 | EVT VT = Op.getValueType(); | |||
1620 | SDValue BasePtr = Load->getBasePtr(); | |||
1621 | EVT MemVT = Load->getMemoryVT(); | |||
1622 | SDLoc SL(Op); | |||
1623 | const MachinePointerInfo &SrcValue = Load->getMemOperand()->getPointerInfo(); | |||
1624 | unsigned BaseAlign = Load->getAlignment(); | |||
1625 | unsigned NumElements = MemVT.getVectorNumElements(); | |||
1626 | ||||
1627 | // Widen from vec3 to vec4 when the load is at least 8-byte aligned | |||
1628 | // or 16-byte fully dereferenceable. Otherwise, split the vector load. | |||
1629 | if (NumElements != 3 || | |||
1630 | (BaseAlign < 8 && | |||
1631 | !SrcValue.isDereferenceable(16, *DAG.getContext(), DAG.getDataLayout()))) | |||
1632 | return SplitVectorLoad(Op, DAG); | |||
1633 | ||||
1634 | assert(NumElements == 3)(static_cast <bool> (NumElements == 3) ? void (0) : __assert_fail ("NumElements == 3", "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp" , 1634, __extension__ __PRETTY_FUNCTION__)); | |||
1635 | ||||
1636 | EVT WideVT = | |||
1637 | EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), 4); | |||
1638 | EVT WideMemVT = | |||
1639 | EVT::getVectorVT(*DAG.getContext(), MemVT.getVectorElementType(), 4); | |||
1640 | SDValue WideLoad = DAG.getExtLoad( | |||
1641 | Load->getExtensionType(), SL, WideVT, Load->getChain(), BasePtr, SrcValue, | |||
1642 | WideMemVT, BaseAlign, Load->getMemOperand()->getFlags()); | |||
1643 | return DAG.getMergeValues( | |||
1644 | {DAG.getNode(ISD::EXTRACT_SUBVECTOR, SL, VT, WideLoad, | |||
1645 | DAG.getVectorIdxConstant(0, SL)), | |||
1646 | WideLoad.getValue(1)}, | |||
1647 | SL); | |||
1648 | } | |||
1649 | ||||
1650 | SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op, | |||
1651 | SelectionDAG &DAG) const { | |||
1652 | StoreSDNode *Store = cast<StoreSDNode>(Op); | |||
1653 | SDValue Val = Store->getValue(); | |||
1654 | EVT VT = Val.getValueType(); | |||
1655 | ||||
1656 | // If this is a 2 element vector, we really want to scalarize and not create | |||
1657 | // weird 1 element vectors. | |||
1658 | if (VT.getVectorNumElements() == 2) | |||
1659 | return scalarizeVectorStore(Store, DAG); | |||
1660 | ||||
1661 | EVT MemVT = Store->getMemoryVT(); | |||
1662 | SDValue Chain = Store->getChain(); | |||
1663 | SDValue BasePtr = Store->getBasePtr(); | |||
1664 | SDLoc SL(Op); | |||
1665 | ||||
1666 | EVT LoVT, HiVT; | |||
1667 | EVT LoMemVT, HiMemVT; | |||
1668 | SDValue Lo, Hi; | |||
1669 | ||||
1670 | std::tie(LoVT, HiVT) = getSplitDestVTs(VT, DAG); | |||
1671 | std::tie(LoMemVT, HiMemVT) = getSplitDestVTs(MemVT, DAG); | |||
1672 | std::tie(Lo, Hi) = splitVector(Val, SL, LoVT, HiVT, DAG); | |||
1673 | ||||
1674 | SDValue HiPtr = DAG.getObjectPtrOffset(SL, BasePtr, LoMemVT.getStoreSize()); | |||
1675 | ||||
1676 | const MachinePointerInfo &SrcValue = Store->getMemOperand()->getPointerInfo(); | |||
1677 | unsigned BaseAlign = Store->getAlignment(); | |||
1678 | unsigned Size = LoMemVT.getStoreSize(); | |||
1679 | unsigned HiAlign = MinAlign(BaseAlign, Size); | |||
1680 | ||||
1681 | SDValue LoStore = | |||
1682 | DAG.getTruncStore(Chain, SL, Lo, BasePtr, SrcValue, LoMemVT, BaseAlign, | |||
1683 | Store->getMemOperand()->getFlags()); | |||
1684 | SDValue HiStore = | |||
1685 | DAG.getTruncStore(Chain, SL, Hi, HiPtr, SrcValue.getWithOffset(Size), | |||
1686 | HiMemVT, HiAlign, Store->getMemOperand()->getFlags()); | |||
1687 | ||||
1688 | return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoStore, HiStore); | |||
1689 | } | |||
1690 | ||||
1691 | // This is a shortcut for integer division because we have fast i32<->f32 | |||
1692 | // conversions, and fast f32 reciprocal instructions. The fractional part of a | |||
1693 | // float is enough to accurately represent up to a 24-bit signed integer. | |||
1694 | SDValue AMDGPUTargetLowering::LowerDIVREM24(SDValue Op, SelectionDAG &DAG, | |||
1695 | bool Sign) const { | |||
1696 | SDLoc DL(Op); | |||
1697 | EVT VT = Op.getValueType(); | |||
1698 | SDValue LHS = Op.getOperand(0); | |||
1699 | SDValue RHS = Op.getOperand(1); | |||
1700 | MVT IntVT = MVT::i32; | |||
1701 | MVT FltVT = MVT::f32; | |||
1702 | ||||
1703 | unsigned LHSSignBits = DAG.ComputeNumSignBits(LHS); | |||
1704 | if (LHSSignBits < 9) | |||
1705 | return SDValue(); | |||
1706 | ||||
1707 | unsigned RHSSignBits = DAG.ComputeNumSignBits(RHS); | |||
1708 | if (RHSSignBits < 9) | |||
1709 | return SDValue(); | |||
1710 | ||||
1711 | unsigned BitSize = VT.getSizeInBits(); | |||
1712 | unsigned SignBits = std::min(LHSSignBits, RHSSignBits); | |||
1713 | unsigned DivBits = BitSize - SignBits; | |||
1714 | if (Sign) | |||
1715 | ++DivBits; | |||
1716 | ||||
1717 | ISD::NodeType ToFp = Sign ? ISD::SINT_TO_FP : ISD::UINT_TO_FP; | |||
1718 | ISD::NodeType ToInt = Sign ? ISD::FP_TO_SINT : ISD::FP_TO_UINT; | |||
1719 | ||||
1720 | SDValue jq = DAG.getConstant(1, DL, IntVT); | |||
1721 | ||||
1722 | if (Sign) { | |||
1723 | // char|short jq = ia ^ ib; | |||
1724 | jq = DAG.getNode(ISD::XOR, DL, VT, LHS, RHS); | |||
1725 | ||||
1726 | // jq = jq >> (bitsize - 2) | |||
1727 | jq = DAG.getNode(ISD::SRA, DL, VT, jq, | |||
1728 | DAG.getConstant(BitSize - 2, DL, VT)); | |||
1729 | ||||
1730 | // jq = jq | 0x1 | |||
1731 | jq = DAG.getNode(ISD::OR, DL, VT, jq, DAG.getConstant(1, DL, VT)); | |||
1732 | } | |||
1733 | ||||
1734 | // int ia = (int)LHS; | |||
1735 | SDValue ia = LHS; | |||
1736 | ||||
1737 | // int ib, (int)RHS; | |||
1738 | SDValue ib = RHS; | |||
1739 | ||||
1740 | // float fa = (float)ia; | |||
1741 | SDValue fa = DAG.getNode(ToFp, DL, FltVT, ia); | |||
1742 | ||||
1743 | // float fb = (float)ib; | |||
1744 | SDValue fb = DAG.getNode(ToFp, DL, FltVT, ib); | |||
1745 | ||||
1746 | SDValue fq = DAG.getNode(ISD::FMUL, DL, FltVT, | |||
1747 | fa, DAG.getNode(AMDGPUISD::RCP, DL, FltVT, fb)); | |||
1748 | ||||
1749 | // fq = trunc(fq); | |||
1750 | fq = DAG.getNode(ISD::FTRUNC, DL, FltVT, fq); | |||
1751 | ||||
1752 | // float fqneg = -fq; | |||
1753 | SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FltVT, fq); | |||
1754 | ||||
1755 | MachineFunction &MF = DAG.getMachineFunction(); | |||
1756 | const AMDGPUMachineFunction *MFI = MF.getInfo<AMDGPUMachineFunction>(); | |||
1757 | ||||
1758 | // float fr = mad(fqneg, fb, fa); | |||
1759 | unsigned OpCode = !Subtarget->hasMadMacF32Insts() ? | |||
1760 | (unsigned)ISD::FMA : | |||
1761 | !MFI->getMode().allFP32Denormals() ? | |||
1762 | (unsigned)ISD::FMAD : | |||
1763 | (unsigned)AMDGPUISD::FMAD_FTZ; | |||
1764 | SDValue fr = DAG.getNode(OpCode, DL, FltVT, fqneg, fb, fa); | |||
1765 | ||||
1766 | // int iq = (int)fq; | |||
1767 | SDValue iq = DAG.getNode(ToInt, DL, IntVT, fq); | |||
1768 | ||||
1769 | // fr = fabs(fr); | |||
1770 | fr = DAG.getNode(ISD::FABS, DL, FltVT, fr); | |||
1771 | ||||
1772 | // fb = fabs(fb); | |||
1773 | fb = DAG.getNode(ISD::FABS, DL, FltVT, fb); | |||
1774 | ||||
1775 | EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); | |||
1776 | ||||
1777 | // int cv = fr >= fb; | |||
1778 | SDValue cv = DAG.getSetCC(DL, SetCCVT, fr, fb, ISD::SETOGE); | |||
1779 | ||||
1780 | // jq = (cv ? jq : 0); | |||
1781 | jq = DAG.getNode(ISD::SELECT, DL, VT, cv, jq, DAG.getConstant(0, DL, VT)); | |||
1782 | ||||
1783 | // dst = iq + jq; | |||
1784 | SDValue Div = DAG.getNode(ISD::ADD, DL, VT, iq, jq); | |||
1785 | ||||
1786 | // Rem needs compensation, it's easier to recompute it | |||
1787 | SDValue Rem = DAG.getNode(ISD::MUL, DL, VT, Div, RHS); | |||
1788 | Rem = DAG.getNode(ISD::SUB, DL, VT, LHS, Rem); | |||
1789 | ||||
1790 | // Truncate to number of bits this divide really is. | |||
1791 | if (Sign) { | |||
1792 | SDValue InRegSize | |||
1793 | = DAG.getValueType(EVT::getIntegerVT(*DAG.getContext(), DivBits)); | |||
1794 | Div = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Div, InRegSize); | |||
1795 | Rem = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Rem, InRegSize); | |||
1796 | } else { | |||
1797 | SDValue TruncMask = DAG.getConstant((UINT64_C(1)1UL << DivBits) - 1, DL, VT); | |||
1798 | Div = DAG.getNode(ISD::AND, DL, VT, Div, TruncMask); | |||
1799 | Rem = DAG.getNode(ISD::AND, DL, VT, Rem, TruncMask); | |||
1800 | } | |||
1801 | ||||
1802 | return DAG.getMergeValues({ Div, Rem }, DL); | |||
1803 | } | |||
1804 | ||||
1805 | void AMDGPUTargetLowering::LowerUDIVREM64(SDValue Op, | |||
1806 | SelectionDAG &DAG, | |||
1807 | SmallVectorImpl<SDValue> &Results) const { | |||
1808 | SDLoc DL(Op); | |||
1809 | EVT VT = Op.getValueType(); | |||
1810 | ||||
1811 | assert(VT == MVT::i64 && "LowerUDIVREM64 expects an i64")(static_cast <bool> (VT == MVT::i64 && "LowerUDIVREM64 expects an i64" ) ? void (0) : __assert_fail ("VT == MVT::i64 && \"LowerUDIVREM64 expects an i64\"" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 1811, __extension__ __PRETTY_FUNCTION__)); | |||
1812 | ||||
1813 | EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext()); | |||
1814 | ||||
1815 | SDValue One = DAG.getConstant(1, DL, HalfVT); | |||
1816 | SDValue Zero = DAG.getConstant(0, DL, HalfVT); | |||
1817 | ||||
1818 | //HiLo split | |||
1819 | SDValue LHS = Op.getOperand(0); | |||
1820 | SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero); | |||
1821 | SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, One); | |||
1822 | ||||
1823 | SDValue RHS = Op.getOperand(1); | |||
1824 | SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero); | |||
1825 | SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, One); | |||
1826 | ||||
1827 | if (DAG.MaskedValueIsZero(RHS, APInt::getHighBitsSet(64, 32)) && | |||
1828 | DAG.MaskedValueIsZero(LHS, APInt::getHighBitsSet(64, 32))) { | |||
1829 | ||||
1830 | SDValue Res = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(HalfVT, HalfVT), | |||
1831 | LHS_Lo, RHS_Lo); | |||
1832 | ||||
1833 | SDValue DIV = DAG.getBuildVector(MVT::v2i32, DL, {Res.getValue(0), Zero}); | |||
1834 | SDValue REM = DAG.getBuildVector(MVT::v2i32, DL, {Res.getValue(1), Zero}); | |||
1835 | ||||
1836 | Results.push_back(DAG.getNode(ISD::BITCAST, DL, MVT::i64, DIV)); | |||
1837 | Results.push_back(DAG.getNode(ISD::BITCAST, DL, MVT::i64, REM)); | |||
1838 | return; | |||
1839 | } | |||
1840 | ||||
1841 | if (isTypeLegal(MVT::i64)) { | |||
1842 | // The algorithm here is based on ideas from "Software Integer Division", | |||
1843 | // Tom Rodeheffer, August 2008. | |||
1844 | ||||
1845 | MachineFunction &MF = DAG.getMachineFunction(); | |||
1846 | const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); | |||
1847 | ||||
1848 | // Compute denominator reciprocal. | |||
1849 | unsigned FMAD = !Subtarget->hasMadMacF32Insts() ? | |||
1850 | (unsigned)ISD::FMA : | |||
1851 | !MFI->getMode().allFP32Denormals() ? | |||
1852 | (unsigned)ISD::FMAD : | |||
1853 | (unsigned)AMDGPUISD::FMAD_FTZ; | |||
1854 | ||||
1855 | SDValue Cvt_Lo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, RHS_Lo); | |||
1856 | SDValue Cvt_Hi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, RHS_Hi); | |||
1857 | SDValue Mad1 = DAG.getNode(FMAD, DL, MVT::f32, Cvt_Hi, | |||
1858 | DAG.getConstantFP(APInt(32, 0x4f800000).bitsToFloat(), DL, MVT::f32), | |||
1859 | Cvt_Lo); | |||
1860 | SDValue Rcp = DAG.getNode(AMDGPUISD::RCP, DL, MVT::f32, Mad1); | |||
1861 | SDValue Mul1 = DAG.getNode(ISD::FMUL, DL, MVT::f32, Rcp, | |||
1862 | DAG.getConstantFP(APInt(32, 0x5f7ffffc).bitsToFloat(), DL, MVT::f32)); | |||
1863 | SDValue Mul2 = DAG.getNode(ISD::FMUL, DL, MVT::f32, Mul1, | |||
1864 | DAG.getConstantFP(APInt(32, 0x2f800000).bitsToFloat(), DL, MVT::f32)); | |||
1865 | SDValue Trunc = DAG.getNode(ISD::FTRUNC, DL, MVT::f32, Mul2); | |||
1866 | SDValue Mad2 = DAG.getNode(FMAD, DL, MVT::f32, Trunc, | |||
1867 | DAG.getConstantFP(APInt(32, 0xcf800000).bitsToFloat(), DL, MVT::f32), | |||
1868 | Mul1); | |||
1869 | SDValue Rcp_Lo = DAG.getNode(ISD::FP_TO_UINT, DL, HalfVT, Mad2); | |||
1870 | SDValue Rcp_Hi = DAG.getNode(ISD::FP_TO_UINT, DL, HalfVT, Trunc); | |||
1871 | SDValue Rcp64 = DAG.getBitcast(VT, | |||
1872 | DAG.getBuildVector(MVT::v2i32, DL, {Rcp_Lo, Rcp_Hi})); | |||
1873 | ||||
1874 | SDValue Zero64 = DAG.getConstant(0, DL, VT); | |||
1875 | SDValue One64 = DAG.getConstant(1, DL, VT); | |||
1876 | SDValue Zero1 = DAG.getConstant(0, DL, MVT::i1); | |||
1877 | SDVTList HalfCarryVT = DAG.getVTList(HalfVT, MVT::i1); | |||
1878 | ||||
1879 | // First round of UNR (Unsigned integer Newton-Raphson). | |||
1880 | SDValue Neg_RHS = DAG.getNode(ISD::SUB, DL, VT, Zero64, RHS); | |||
1881 | SDValue Mullo1 = DAG.getNode(ISD::MUL, DL, VT, Neg_RHS, Rcp64); | |||
1882 | SDValue Mulhi1 = DAG.getNode(ISD::MULHU, DL, VT, Rcp64, Mullo1); | |||
1883 | SDValue Mulhi1_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi1, | |||
1884 | Zero); | |||
1885 | SDValue Mulhi1_Hi = | |||
1886 | DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi1, One); | |||
1887 | SDValue Add1_Lo = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Rcp_Lo, | |||
1888 | Mulhi1_Lo, Zero1); | |||
1889 | SDValue Add1_Hi = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Rcp_Hi, | |||
1890 | Mulhi1_Hi, Add1_Lo.getValue(1)); | |||
1891 | SDValue Add1 = DAG.getBitcast(VT, | |||
1892 | DAG.getBuildVector(MVT::v2i32, DL, {Add1_Lo, Add1_Hi})); | |||
1893 | ||||
1894 | // Second round of UNR. | |||
1895 | SDValue Mullo2 = DAG.getNode(ISD::MUL, DL, VT, Neg_RHS, Add1); | |||
1896 | SDValue Mulhi2 = DAG.getNode(ISD::MULHU, DL, VT, Add1, Mullo2); | |||
1897 | SDValue Mulhi2_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi2, | |||
1898 | Zero); | |||
1899 | SDValue Mulhi2_Hi = | |||
1900 | DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi2, One); | |||
1901 | SDValue Add2_Lo = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Add1_Lo, | |||
1902 | Mulhi2_Lo, Zero1); | |||
1903 | SDValue Add2_Hi = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Add1_Hi, | |||
1904 | Mulhi2_Hi, Add2_Lo.getValue(1)); | |||
1905 | SDValue Add2 = DAG.getBitcast(VT, | |||
1906 | DAG.getBuildVector(MVT::v2i32, DL, {Add2_Lo, Add2_Hi})); | |||
1907 | ||||
1908 | SDValue Mulhi3 = DAG.getNode(ISD::MULHU, DL, VT, LHS, Add2); | |||
1909 | ||||
1910 | SDValue Mul3 = DAG.getNode(ISD::MUL, DL, VT, RHS, Mulhi3); | |||
1911 | ||||
1912 | SDValue Mul3_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mul3, Zero); | |||
1913 | SDValue Mul3_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mul3, One); | |||
1914 | SDValue Sub1_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, LHS_Lo, | |||
1915 | Mul3_Lo, Zero1); | |||
1916 | SDValue Sub1_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, LHS_Hi, | |||
1917 | Mul3_Hi, Sub1_Lo.getValue(1)); | |||
1918 | SDValue Sub1_Mi = DAG.getNode(ISD::SUB, DL, HalfVT, LHS_Hi, Mul3_Hi); | |||
1919 | SDValue Sub1 = DAG.getBitcast(VT, | |||
1920 | DAG.getBuildVector(MVT::v2i32, DL, {Sub1_Lo, Sub1_Hi})); | |||
1921 | ||||
1922 | SDValue MinusOne = DAG.getConstant(0xffffffffu, DL, HalfVT); | |||
1923 | SDValue C1 = DAG.getSelectCC(DL, Sub1_Hi, RHS_Hi, MinusOne, Zero, | |||
1924 | ISD::SETUGE); | |||
1925 | SDValue C2 = DAG.getSelectCC(DL, Sub1_Lo, RHS_Lo, MinusOne, Zero, | |||
1926 | ISD::SETUGE); | |||
1927 | SDValue C3 = DAG.getSelectCC(DL, Sub1_Hi, RHS_Hi, C2, C1, ISD::SETEQ); | |||
1928 | ||||
1929 | // TODO: Here and below portions of the code can be enclosed into if/endif. | |||
1930 | // Currently control flow is unconditional and we have 4 selects after | |||
1931 | // potential endif to substitute PHIs. | |||
1932 | ||||
1933 | // if C3 != 0 ... | |||
1934 | SDValue Sub2_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub1_Lo, | |||
1935 | RHS_Lo, Zero1); | |||
1936 | SDValue Sub2_Mi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub1_Mi, | |||
1937 | RHS_Hi, Sub1_Lo.getValue(1)); | |||
1938 | SDValue Sub2_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Mi, | |||
1939 | Zero, Sub2_Lo.getValue(1)); | |||
1940 | SDValue Sub2 = DAG.getBitcast(VT, | |||
1941 | DAG.getBuildVector(MVT::v2i32, DL, {Sub2_Lo, Sub2_Hi})); | |||
1942 | ||||
1943 | SDValue Add3 = DAG.getNode(ISD::ADD, DL, VT, Mulhi3, One64); | |||
1944 | ||||
1945 | SDValue C4 = DAG.getSelectCC(DL, Sub2_Hi, RHS_Hi, MinusOne, Zero, | |||
1946 | ISD::SETUGE); | |||
1947 | SDValue C5 = DAG.getSelectCC(DL, Sub2_Lo, RHS_Lo, MinusOne, Zero, | |||
1948 | ISD::SETUGE); | |||
1949 | SDValue C6 = DAG.getSelectCC(DL, Sub2_Hi, RHS_Hi, C5, C4, ISD::SETEQ); | |||
1950 | ||||
1951 | // if (C6 != 0) | |||
1952 | SDValue Add4 = DAG.getNode(ISD::ADD, DL, VT, Add3, One64); | |||
1953 | ||||
1954 | SDValue Sub3_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Lo, | |||
1955 | RHS_Lo, Zero1); | |||
1956 | SDValue Sub3_Mi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Mi, | |||
1957 | RHS_Hi, Sub2_Lo.getValue(1)); | |||
1958 | SDValue Sub3_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub3_Mi, | |||
1959 | Zero, Sub3_Lo.getValue(1)); | |||
1960 | SDValue Sub3 = DAG.getBitcast(VT, | |||
1961 | DAG.getBuildVector(MVT::v2i32, DL, {Sub3_Lo, Sub3_Hi})); | |||
1962 | ||||
1963 | // endif C6 | |||
1964 | // endif C3 | |||
1965 | ||||
1966 | SDValue Sel1 = DAG.getSelectCC(DL, C6, Zero, Add4, Add3, ISD::SETNE); | |||
1967 | SDValue Div = DAG.getSelectCC(DL, C3, Zero, Sel1, Mulhi3, ISD::SETNE); | |||
1968 | ||||
1969 | SDValue Sel2 = DAG.getSelectCC(DL, C6, Zero, Sub3, Sub2, ISD::SETNE); | |||
1970 | SDValue Rem = DAG.getSelectCC(DL, C3, Zero, Sel2, Sub1, ISD::SETNE); | |||
1971 | ||||
1972 | Results.push_back(Div); | |||
1973 | Results.push_back(Rem); | |||
1974 | ||||
1975 | return; | |||
1976 | } | |||
1977 | ||||
1978 | // r600 expandion. | |||
1979 | // Get Speculative values | |||
1980 | SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo); | |||
1981 | SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo); | |||
1982 | ||||
1983 | SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, Zero, REM_Part, LHS_Hi, ISD::SETEQ); | |||
1984 | SDValue REM = DAG.getBuildVector(MVT::v2i32, DL, {REM_Lo, Zero}); | |||
1985 | REM = DAG.getNode(ISD::BITCAST, DL, MVT::i64, REM); | |||
1986 | ||||
1987 | SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, Zero, DIV_Part, Zero, ISD::SETEQ); | |||
1988 | SDValue DIV_Lo = Zero; | |||
1989 | ||||
1990 | const unsigned halfBitWidth = HalfVT.getSizeInBits(); | |||
1991 | ||||
1992 | for (unsigned i = 0; i < halfBitWidth; ++i) { | |||
1993 | const unsigned bitPos = halfBitWidth - i - 1; | |||
1994 | SDValue POS = DAG.getConstant(bitPos, DL, HalfVT); | |||
1995 | // Get value of high bit | |||
1996 | SDValue HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS); | |||
1997 | HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, One); | |||
1998 | HBit = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, HBit); | |||
1999 | ||||
2000 | // Shift | |||
2001 | REM = DAG.getNode(ISD::SHL, DL, VT, REM, DAG.getConstant(1, DL, VT)); | |||
2002 | // Add LHS high bit | |||
2003 | REM = DAG.getNode(ISD::OR, DL, VT, REM, HBit); | |||
2004 | ||||
2005 | SDValue BIT = DAG.getConstant(1ULL << bitPos, DL, HalfVT); | |||
2006 | SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, Zero, ISD::SETUGE); | |||
2007 | ||||
2008 | DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT); | |||
2009 | ||||
2010 | // Update REM | |||
2011 | SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS); | |||
2012 | REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETUGE); | |||
2013 | } | |||
2014 | ||||
2015 | SDValue DIV = DAG.getBuildVector(MVT::v2i32, DL, {DIV_Lo, DIV_Hi}); | |||
2016 | DIV = DAG.getNode(ISD::BITCAST, DL, MVT::i64, DIV); | |||
2017 | Results.push_back(DIV); | |||
2018 | Results.push_back(REM); | |||
2019 | } | |||
2020 | ||||
2021 | SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op, | |||
2022 | SelectionDAG &DAG) const { | |||
2023 | SDLoc DL(Op); | |||
2024 | EVT VT = Op.getValueType(); | |||
2025 | ||||
2026 | if (VT == MVT::i64) { | |||
2027 | SmallVector<SDValue, 2> Results; | |||
2028 | LowerUDIVREM64(Op, DAG, Results); | |||
2029 | return DAG.getMergeValues(Results, DL); | |||
2030 | } | |||
2031 | ||||
2032 | if (VT == MVT::i32) { | |||
2033 | if (SDValue Res = LowerDIVREM24(Op, DAG, false)) | |||
2034 | return Res; | |||
2035 | } | |||
2036 | ||||
2037 | SDValue X = Op.getOperand(0); | |||
2038 | SDValue Y = Op.getOperand(1); | |||
2039 | ||||
2040 | // See AMDGPUCodeGenPrepare::expandDivRem32 for a description of the | |||
2041 | // algorithm used here. | |||
2042 | ||||
2043 | // Initial estimate of inv(y). | |||
2044 | SDValue Z = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Y); | |||
2045 | ||||
2046 | // One round of UNR. | |||
2047 | SDValue NegY = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Y); | |||
2048 | SDValue NegYZ = DAG.getNode(ISD::MUL, DL, VT, NegY, Z); | |||
2049 | Z = DAG.getNode(ISD::ADD, DL, VT, Z, | |||
2050 | DAG.getNode(ISD::MULHU, DL, VT, Z, NegYZ)); | |||
2051 | ||||
2052 | // Quotient/remainder estimate. | |||
2053 | SDValue Q = DAG.getNode(ISD::MULHU, DL, VT, X, Z); | |||
2054 | SDValue R = | |||
2055 | DAG.getNode(ISD::SUB, DL, VT, X, DAG.getNode(ISD::MUL, DL, VT, Q, Y)); | |||
2056 | ||||
2057 | // First quotient/remainder refinement. | |||
2058 | EVT CCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); | |||
2059 | SDValue One = DAG.getConstant(1, DL, VT); | |||
2060 | SDValue Cond = DAG.getSetCC(DL, CCVT, R, Y, ISD::SETUGE); | |||
2061 | Q = DAG.getNode(ISD::SELECT, DL, VT, Cond, | |||
2062 | DAG.getNode(ISD::ADD, DL, VT, Q, One), Q); | |||
2063 | R = DAG.getNode(ISD::SELECT, DL, VT, Cond, | |||
2064 | DAG.getNode(ISD::SUB, DL, VT, R, Y), R); | |||
2065 | ||||
2066 | // Second quotient/remainder refinement. | |||
2067 | Cond = DAG.getSetCC(DL, CCVT, R, Y, ISD::SETUGE); | |||
2068 | Q = DAG.getNode(ISD::SELECT, DL, VT, Cond, | |||
2069 | DAG.getNode(ISD::ADD, DL, VT, Q, One), Q); | |||
2070 | R = DAG.getNode(ISD::SELECT, DL, VT, Cond, | |||
2071 | DAG.getNode(ISD::SUB, DL, VT, R, Y), R); | |||
2072 | ||||
2073 | return DAG.getMergeValues({Q, R}, DL); | |||
2074 | } | |||
2075 | ||||
2076 | SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op, | |||
2077 | SelectionDAG &DAG) const { | |||
2078 | SDLoc DL(Op); | |||
2079 | EVT VT = Op.getValueType(); | |||
2080 | ||||
2081 | SDValue LHS = Op.getOperand(0); | |||
2082 | SDValue RHS = Op.getOperand(1); | |||
2083 | ||||
2084 | SDValue Zero = DAG.getConstant(0, DL, VT); | |||
2085 | SDValue NegOne = DAG.getConstant(-1, DL, VT); | |||
2086 | ||||
2087 | if (VT == MVT::i32) { | |||
2088 | if (SDValue Res = LowerDIVREM24(Op, DAG, true)) | |||
2089 | return Res; | |||
2090 | } | |||
2091 | ||||
2092 | if (VT == MVT::i64 && | |||
2093 | DAG.ComputeNumSignBits(LHS) > 32 && | |||
2094 | DAG.ComputeNumSignBits(RHS) > 32) { | |||
2095 | EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext()); | |||
2096 | ||||
2097 | //HiLo split | |||
2098 | SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero); | |||
2099 | SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero); | |||
2100 | SDValue DIVREM = DAG.getNode(ISD::SDIVREM, DL, DAG.getVTList(HalfVT, HalfVT), | |||
2101 | LHS_Lo, RHS_Lo); | |||
2102 | SDValue Res[2] = { | |||
2103 | DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(0)), | |||
2104 | DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(1)) | |||
2105 | }; | |||
2106 | return DAG.getMergeValues(Res, DL); | |||
2107 | } | |||
2108 | ||||
2109 | SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT); | |||
2110 | SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT); | |||
2111 | SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign); | |||
2112 | SDValue RSign = LHSign; // Remainder sign is the same as LHS | |||
2113 | ||||
2114 | LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign); | |||
2115 | RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign); | |||
2116 | ||||
2117 | LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign); | |||
2118 | RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign); | |||
2119 | ||||
2120 | SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS); | |||
2121 | SDValue Rem = Div.getValue(1); | |||
2122 | ||||
2123 | Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign); | |||
2124 | Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign); | |||
2125 | ||||
2126 | Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign); | |||
2127 | Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign); | |||
2128 | ||||
2129 | SDValue Res[2] = { | |||
2130 | Div, | |||
2131 | Rem | |||
2132 | }; | |||
2133 | return DAG.getMergeValues(Res, DL); | |||
2134 | } | |||
2135 | ||||
2136 | // (frem x, y) -> (fma (fneg (ftrunc (fdiv x, y))), y, x) | |||
2137 | SDValue AMDGPUTargetLowering::LowerFREM(SDValue Op, SelectionDAG &DAG) const { | |||
2138 | SDLoc SL(Op); | |||
2139 | EVT VT = Op.getValueType(); | |||
2140 | auto Flags = Op->getFlags(); | |||
2141 | SDValue X = Op.getOperand(0); | |||
2142 | SDValue Y = Op.getOperand(1); | |||
2143 | ||||
2144 | SDValue Div = DAG.getNode(ISD::FDIV, SL, VT, X, Y, Flags); | |||
2145 | SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, VT, Div, Flags); | |||
2146 | SDValue Neg = DAG.getNode(ISD::FNEG, SL, VT, Trunc, Flags); | |||
2147 | // TODO: For f32 use FMAD instead if !hasFastFMA32? | |||
2148 | return DAG.getNode(ISD::FMA, SL, VT, Neg, Y, X, Flags); | |||
2149 | } | |||
2150 | ||||
2151 | SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const { | |||
2152 | SDLoc SL(Op); | |||
2153 | SDValue Src = Op.getOperand(0); | |||
2154 | ||||
2155 | // result = trunc(src) | |||
2156 | // if (src > 0.0 && src != result) | |||
2157 | // result += 1.0 | |||
2158 | ||||
2159 | SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src); | |||
2160 | ||||
2161 | const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64); | |||
2162 | const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f64); | |||
2163 | ||||
2164 | EVT SetCCVT = | |||
2165 | getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64); | |||
2166 | ||||
2167 | SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT); | |||
2168 | SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE); | |||
2169 | SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc); | |||
2170 | ||||
2171 | SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero); | |||
2172 | // TODO: Should this propagate fast-math-flags? | |||
2173 | return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add); | |||
2174 | } | |||
2175 | ||||
2176 | static SDValue extractF64Exponent(SDValue Hi, const SDLoc &SL, | |||
2177 | SelectionDAG &DAG) { | |||
2178 | const unsigned FractBits = 52; | |||
2179 | const unsigned ExpBits = 11; | |||
2180 | ||||
2181 | SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_U32, SL, MVT::i32, | |||
2182 | Hi, | |||
2183 | DAG.getConstant(FractBits - 32, SL, MVT::i32), | |||
2184 | DAG.getConstant(ExpBits, SL, MVT::i32)); | |||
2185 | SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart, | |||
2186 | DAG.getConstant(1023, SL, MVT::i32)); | |||
2187 | ||||
2188 | return Exp; | |||
2189 | } | |||
2190 | ||||
2191 | SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const { | |||
2192 | SDLoc SL(Op); | |||
2193 | SDValue Src = Op.getOperand(0); | |||
2194 | ||||
2195 | assert(Op.getValueType() == MVT::f64)(static_cast <bool> (Op.getValueType() == MVT::f64) ? void (0) : __assert_fail ("Op.getValueType() == MVT::f64", "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp" , 2195, __extension__ __PRETTY_FUNCTION__)); | |||
2196 | ||||
2197 | const SDValue Zero = DAG.getConstant(0, SL, MVT::i32); | |||
2198 | ||||
2199 | // Extract the upper half, since this is where we will find the sign and | |||
2200 | // exponent. | |||
2201 | SDValue Hi = getHiHalf64(Src, DAG); | |||
2202 | ||||
2203 | SDValue Exp = extractF64Exponent(Hi, SL, DAG); | |||
2204 | ||||
2205 | const unsigned FractBits = 52; | |||
2206 | ||||
2207 | // Extract the sign bit. | |||
2208 | const SDValue SignBitMask = DAG.getConstant(UINT32_C(1)1U << 31, SL, MVT::i32); | |||
2209 | SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask); | |||
2210 | ||||
2211 | // Extend back to 64-bits. | |||
2212 | SDValue SignBit64 = DAG.getBuildVector(MVT::v2i32, SL, {Zero, SignBit}); | |||
2213 | SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64); | |||
2214 | ||||
2215 | SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src); | |||
2216 | const SDValue FractMask | |||
2217 | = DAG.getConstant((UINT64_C(1)1UL << FractBits) - 1, SL, MVT::i64); | |||
2218 | ||||
2219 | SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp); | |||
2220 | SDValue Not = DAG.getNOT(SL, Shr, MVT::i64); | |||
2221 | SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not); | |||
2222 | ||||
2223 | EVT SetCCVT = | |||
2224 | getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i32); | |||
2225 | ||||
2226 | const SDValue FiftyOne = DAG.getConstant(FractBits - 1, SL, MVT::i32); | |||
2227 | ||||
2228 | SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT); | |||
2229 | SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT); | |||
2230 | ||||
2231 | SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0); | |||
2232 | SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1); | |||
2233 | ||||
2234 | return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2); | |||
2235 | } | |||
2236 | ||||
2237 | SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const { | |||
2238 | SDLoc SL(Op); | |||
2239 | SDValue Src = Op.getOperand(0); | |||
2240 | ||||
2241 | assert(Op.getValueType() == MVT::f64)(static_cast <bool> (Op.getValueType() == MVT::f64) ? void (0) : __assert_fail ("Op.getValueType() == MVT::f64", "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp" , 2241, __extension__ __PRETTY_FUNCTION__)); | |||
2242 | ||||
2243 | APFloat C1Val(APFloat::IEEEdouble(), "0x1.0p+52"); | |||
2244 | SDValue C1 = DAG.getConstantFP(C1Val, SL, MVT::f64); | |||
2245 | SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src); | |||
2246 | ||||
2247 | // TODO: Should this propagate fast-math-flags? | |||
2248 | ||||
2249 | SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign); | |||
2250 | SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign); | |||
2251 | ||||
2252 | SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src); | |||
2253 | ||||
2254 | APFloat C2Val(APFloat::IEEEdouble(), "0x1.fffffffffffffp+51"); | |||
2255 | SDValue C2 = DAG.getConstantFP(C2Val, SL, MVT::f64); | |||
2256 | ||||
2257 | EVT SetCCVT = | |||
2258 | getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64); | |||
2259 | SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT); | |||
2260 | ||||
2261 | return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2); | |||
2262 | } | |||
2263 | ||||
2264 | SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const { | |||
2265 | // FNEARBYINT and FRINT are the same, except in their handling of FP | |||
2266 | // exceptions. Those aren't really meaningful for us, and OpenCL only has | |||
2267 | // rint, so just treat them as equivalent. | |||
2268 | return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0)); | |||
2269 | } | |||
2270 | ||||
2271 | // XXX - May require not supporting f32 denormals? | |||
2272 | ||||
2273 | // Don't handle v2f16. The extra instructions to scalarize and repack around the | |||
2274 | // compare and vselect end up producing worse code than scalarizing the whole | |||
2275 | // operation. | |||
2276 | SDValue AMDGPUTargetLowering::LowerFROUND(SDValue Op, SelectionDAG &DAG) const { | |||
2277 | SDLoc SL(Op); | |||
2278 | SDValue X = Op.getOperand(0); | |||
2279 | EVT VT = Op.getValueType(); | |||
2280 | ||||
2281 | SDValue T = DAG.getNode(ISD::FTRUNC, SL, VT, X); | |||
2282 | ||||
2283 | // TODO: Should this propagate fast-math-flags? | |||
2284 | ||||
2285 | SDValue Diff = DAG.getNode(ISD::FSUB, SL, VT, X, T); | |||
2286 | ||||
2287 | SDValue AbsDiff = DAG.getNode(ISD::FABS, SL, VT, Diff); | |||
2288 | ||||
2289 | const SDValue Zero = DAG.getConstantFP(0.0, SL, VT); | |||
2290 | const SDValue One = DAG.getConstantFP(1.0, SL, VT); | |||
2291 | const SDValue Half = DAG.getConstantFP(0.5, SL, VT); | |||
2292 | ||||
2293 | SDValue SignOne = DAG.getNode(ISD::FCOPYSIGN, SL, VT, One, X); | |||
2294 | ||||
2295 | EVT SetCCVT = | |||
2296 | getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); | |||
2297 | ||||
2298 | SDValue Cmp = DAG.getSetCC(SL, SetCCVT, AbsDiff, Half, ISD::SETOGE); | |||
2299 | ||||
2300 | SDValue Sel = DAG.getNode(ISD::SELECT, SL, VT, Cmp, SignOne, Zero); | |||
2301 | ||||
2302 | return DAG.getNode(ISD::FADD, SL, VT, T, Sel); | |||
2303 | } | |||
2304 | ||||
2305 | SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const { | |||
2306 | SDLoc SL(Op); | |||
2307 | SDValue Src = Op.getOperand(0); | |||
2308 | ||||
2309 | // result = trunc(src); | |||
2310 | // if (src < 0.0 && src != result) | |||
2311 | // result += -1.0. | |||
2312 | ||||
2313 | SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src); | |||
2314 | ||||
2315 | const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64); | |||
2316 | const SDValue NegOne = DAG.getConstantFP(-1.0, SL, MVT::f64); | |||
2317 | ||||
2318 | EVT SetCCVT = | |||
2319 | getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64); | |||
2320 | ||||
2321 | SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT); | |||
2322 | SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE); | |||
2323 | SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc); | |||
2324 | ||||
2325 | SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero); | |||
2326 | // TODO: Should this propagate fast-math-flags? | |||
2327 | return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add); | |||
2328 | } | |||
2329 | ||||
2330 | SDValue AMDGPUTargetLowering::LowerFLOG(SDValue Op, SelectionDAG &DAG, | |||
2331 | double Log2BaseInverted) const { | |||
2332 | EVT VT = Op.getValueType(); | |||
2333 | ||||
2334 | SDLoc SL(Op); | |||
2335 | SDValue Operand = Op.getOperand(0); | |||
2336 | SDValue Log2Operand = DAG.getNode(ISD::FLOG2, SL, VT, Operand); | |||
2337 | SDValue Log2BaseInvertedOperand = DAG.getConstantFP(Log2BaseInverted, SL, VT); | |||
2338 | ||||
2339 | return DAG.getNode(ISD::FMUL, SL, VT, Log2Operand, Log2BaseInvertedOperand); | |||
2340 | } | |||
2341 | ||||
2342 | // exp2(M_LOG2E_F * f); | |||
2343 | SDValue AMDGPUTargetLowering::lowerFEXP(SDValue Op, SelectionDAG &DAG) const { | |||
2344 | EVT VT = Op.getValueType(); | |||
2345 | SDLoc SL(Op); | |||
2346 | SDValue Src = Op.getOperand(0); | |||
2347 | ||||
2348 | const SDValue K = DAG.getConstantFP(numbers::log2e, SL, VT); | |||
2349 | SDValue Mul = DAG.getNode(ISD::FMUL, SL, VT, Src, K, Op->getFlags()); | |||
2350 | return DAG.getNode(ISD::FEXP2, SL, VT, Mul, Op->getFlags()); | |||
2351 | } | |||
2352 | ||||
2353 | static bool isCtlzOpc(unsigned Opc) { | |||
2354 | return Opc == ISD::CTLZ || Opc == ISD::CTLZ_ZERO_UNDEF; | |||
2355 | } | |||
2356 | ||||
2357 | static bool isCttzOpc(unsigned Opc) { | |||
2358 | return Opc == ISD::CTTZ || Opc == ISD::CTTZ_ZERO_UNDEF; | |||
2359 | } | |||
2360 | ||||
2361 | SDValue AMDGPUTargetLowering::LowerCTLZ_CTTZ(SDValue Op, SelectionDAG &DAG) const { | |||
2362 | SDLoc SL(Op); | |||
2363 | SDValue Src = Op.getOperand(0); | |||
2364 | ||||
2365 | assert(isCtlzOpc(Op.getOpcode()) || isCttzOpc(Op.getOpcode()))(static_cast <bool> (isCtlzOpc(Op.getOpcode()) || isCttzOpc (Op.getOpcode())) ? void (0) : __assert_fail ("isCtlzOpc(Op.getOpcode()) || isCttzOpc(Op.getOpcode())" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 2365, __extension__ __PRETTY_FUNCTION__)); | |||
2366 | bool Ctlz = isCtlzOpc(Op.getOpcode()); | |||
2367 | unsigned NewOpc = Ctlz ? AMDGPUISD::FFBH_U32 : AMDGPUISD::FFBL_B32; | |||
2368 | ||||
2369 | bool ZeroUndef = Op.getOpcode() == ISD::CTLZ_ZERO_UNDEF || | |||
2370 | Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF; | |||
2371 | ||||
2372 | if (Src.getValueType() == MVT::i32) { | |||
2373 | // (ctlz hi:lo) -> (umin (ffbh src), 32) | |||
2374 | // (cttz hi:lo) -> (umin (ffbl src), 32) | |||
2375 | // (ctlz_zero_undef src) -> (ffbh src) | |||
2376 | // (cttz_zero_undef src) -> (ffbl src) | |||
2377 | SDValue NewOpr = DAG.getNode(NewOpc, SL, MVT::i32, Src); | |||
2378 | if (!ZeroUndef) { | |||
2379 | const SDValue Const32 = DAG.getConstant(32, SL, MVT::i32); | |||
2380 | NewOpr = DAG.getNode(ISD::UMIN, SL, MVT::i32, NewOpr, Const32); | |||
2381 | } | |||
2382 | return NewOpr; | |||
2383 | } | |||
2384 | ||||
2385 | SDValue Lo, Hi; | |||
2386 | std::tie(Lo, Hi) = split64BitValue(Src, DAG); | |||
2387 | ||||
2388 | SDValue OprLo = DAG.getNode(NewOpc, SL, MVT::i32, Lo); | |||
2389 | SDValue OprHi = DAG.getNode(NewOpc, SL, MVT::i32, Hi); | |||
2390 | ||||
2391 | // (ctlz hi:lo) -> (umin3 (ffbh hi), (uaddsat (ffbh lo), 32), 64) | |||
2392 | // (cttz hi:lo) -> (umin3 (uaddsat (ffbl hi), 32), (ffbl lo), 64) | |||
2393 | // (ctlz_zero_undef hi:lo) -> (umin (ffbh hi), (add (ffbh lo), 32)) | |||
2394 | // (cttz_zero_undef hi:lo) -> (umin (add (ffbl hi), 32), (ffbl lo)) | |||
2395 | ||||
2396 | unsigned AddOpc = ZeroUndef ? ISD::ADD : ISD::UADDSAT; | |||
2397 | const SDValue Const32 = DAG.getConstant(32, SL, MVT::i32); | |||
2398 | if (Ctlz) | |||
2399 | OprLo = DAG.getNode(AddOpc, SL, MVT::i32, OprLo, Const32); | |||
2400 | else | |||
2401 | OprHi = DAG.getNode(AddOpc, SL, MVT::i32, OprHi, Const32); | |||
2402 | ||||
2403 | SDValue NewOpr; | |||
2404 | NewOpr = DAG.getNode(ISD::UMIN, SL, MVT::i32, OprLo, OprHi); | |||
2405 | if (!ZeroUndef) { | |||
2406 | const SDValue Const64 = DAG.getConstant(64, SL, MVT::i32); | |||
2407 | NewOpr = DAG.getNode(ISD::UMIN, SL, MVT::i32, NewOpr, Const64); | |||
2408 | } | |||
2409 | ||||
2410 | return DAG.getNode(ISD::ZERO_EXTEND, SL, MVT::i64, NewOpr); | |||
2411 | } | |||
2412 | ||||
2413 | SDValue AMDGPUTargetLowering::LowerINT_TO_FP32(SDValue Op, SelectionDAG &DAG, | |||
2414 | bool Signed) const { | |||
2415 | // The regular method converting a 64-bit integer to float roughly consists of | |||
2416 | // 2 steps: normalization and rounding. In fact, after normalization, the | |||
2417 | // conversion from a 64-bit integer to a float is essentially the same as the | |||
2418 | // one from a 32-bit integer. The only difference is that it has more | |||
2419 | // trailing bits to be rounded. To leverage the native 32-bit conversion, a | |||
2420 | // 64-bit integer could be preprocessed and fit into a 32-bit integer then | |||
2421 | // converted into the correct float number. The basic steps for the unsigned | |||
2422 | // conversion are illustrated in the following pseudo code: | |||
2423 | // | |||
2424 | // f32 uitofp(i64 u) { | |||
2425 | // i32 hi, lo = split(u); | |||
2426 | // // Only count the leading zeros in hi as we have native support of the | |||
2427 | // // conversion from i32 to f32. If hi is all 0s, the conversion is | |||
2428 | // // reduced to a 32-bit one automatically. | |||
2429 | // i32 shamt = clz(hi); // Return 32 if hi is all 0s. | |||
2430 | // u <<= shamt; | |||
2431 | // hi, lo = split(u); | |||
2432 | // hi |= (lo != 0) ? 1 : 0; // Adjust rounding bit in hi based on lo. | |||
2433 | // // convert it as a 32-bit integer and scale the result back. | |||
2434 | // return uitofp(hi) * 2^(32 - shamt); | |||
2435 | // } | |||
2436 | // | |||
2437 | // The signed one follows the same principle but uses 'ffbh_i32' to count its | |||
2438 | // sign bits instead. If 'ffbh_i32' is not available, its absolute value is | |||
2439 | // converted instead followed by negation based its sign bit. | |||
2440 | ||||
2441 | SDLoc SL(Op); | |||
2442 | SDValue Src = Op.getOperand(0); | |||
2443 | ||||
2444 | SDValue Lo, Hi; | |||
2445 | std::tie(Lo, Hi) = split64BitValue(Src, DAG); | |||
2446 | SDValue Sign; | |||
2447 | SDValue ShAmt; | |||
2448 | if (Signed && Subtarget->isGCN()) { | |||
2449 | // We also need to consider the sign bit in Lo if Hi has just sign bits, | |||
2450 | // i.e. Hi is 0 or -1. However, that only needs to take the MSB into | |||
2451 | // account. That is, the maximal shift is | |||
2452 | // - 32 if Lo and Hi have opposite signs; | |||
2453 | // - 33 if Lo and Hi have the same sign. | |||
2454 | // | |||
2455 | // Or, MaxShAmt = 33 + OppositeSign, where | |||
2456 | // | |||
2457 | // OppositeSign is defined as ((Lo ^ Hi) >> 31), which is | |||
2458 | // - -1 if Lo and Hi have opposite signs; and | |||
2459 | // - 0 otherwise. | |||
2460 | // | |||
2461 | // All in all, ShAmt is calculated as | |||
2462 | // | |||
2463 | // umin(sffbh(Hi), 33 + (Lo^Hi)>>31) - 1. | |||
2464 | // | |||
2465 | // or | |||
2466 | // | |||
2467 | // umin(sffbh(Hi) - 1, 32 + (Lo^Hi)>>31). | |||
2468 | // | |||
2469 | // to reduce the critical path. | |||
2470 | SDValue OppositeSign = DAG.getNode( | |||
2471 | ISD::SRA, SL, MVT::i32, DAG.getNode(ISD::XOR, SL, MVT::i32, Lo, Hi), | |||
2472 | DAG.getConstant(31, SL, MVT::i32)); | |||
2473 | SDValue MaxShAmt = | |||
2474 | DAG.getNode(ISD::ADD, SL, MVT::i32, DAG.getConstant(32, SL, MVT::i32), | |||
2475 | OppositeSign); | |||
2476 | // Count the leading sign bits. | |||
2477 | ShAmt = DAG.getNode(AMDGPUISD::FFBH_I32, SL, MVT::i32, Hi); | |||
2478 | // Different from unsigned conversion, the shift should be one bit less to | |||
2479 | // preserve the sign bit. | |||
2480 | ShAmt = DAG.getNode(ISD::SUB, SL, MVT::i32, ShAmt, | |||
2481 | DAG.getConstant(1, SL, MVT::i32)); | |||
2482 | ShAmt = DAG.getNode(ISD::UMIN, SL, MVT::i32, ShAmt, MaxShAmt); | |||
2483 | } else { | |||
2484 | if (Signed) { | |||
2485 | // Without 'ffbh_i32', only leading zeros could be counted. Take the | |||
2486 | // absolute value first. | |||
2487 | Sign = DAG.getNode(ISD::SRA, SL, MVT::i64, Src, | |||
2488 | DAG.getConstant(63, SL, MVT::i64)); | |||
2489 | SDValue Abs = | |||
2490 | DAG.getNode(ISD::XOR, SL, MVT::i64, | |||
2491 | DAG.getNode(ISD::ADD, SL, MVT::i64, Src, Sign), Sign); | |||
2492 | std::tie(Lo, Hi) = split64BitValue(Abs, DAG); | |||
2493 | } | |||
2494 | // Count the leading zeros. | |||
2495 | ShAmt = DAG.getNode(ISD::CTLZ, SL, MVT::i32, Hi); | |||
2496 | // The shift amount for signed integers is [0, 32]. | |||
2497 | } | |||
2498 | // Normalize the given 64-bit integer. | |||
2499 | SDValue Norm = DAG.getNode(ISD::SHL, SL, MVT::i64, Src, ShAmt); | |||
2500 | // Split it again. | |||
2501 | std::tie(Lo, Hi) = split64BitValue(Norm, DAG); | |||
2502 | // Calculate the adjust bit for rounding. | |||
2503 | // (lo != 0) ? 1 : 0 => (lo >= 1) ? 1 : 0 => umin(1, lo) | |||
2504 | SDValue Adjust = DAG.getNode(ISD::UMIN, SL, MVT::i32, | |||
2505 | DAG.getConstant(1, SL, MVT::i32), Lo); | |||
2506 | // Get the 32-bit normalized integer. | |||
2507 | Norm = DAG.getNode(ISD::OR, SL, MVT::i32, Hi, Adjust); | |||
2508 | // Convert the normalized 32-bit integer into f32. | |||
2509 | unsigned Opc = | |||
2510 | (Signed && Subtarget->isGCN()) ? ISD::SINT_TO_FP : ISD::UINT_TO_FP; | |||
2511 | SDValue FVal = DAG.getNode(Opc, SL, MVT::f32, Norm); | |||
2512 | ||||
2513 | // Finally, need to scale back the converted floating number as the original | |||
2514 | // 64-bit integer is converted as a 32-bit one. | |||
2515 | ShAmt = DAG.getNode(ISD::SUB, SL, MVT::i32, DAG.getConstant(32, SL, MVT::i32), | |||
2516 | ShAmt); | |||
2517 | // On GCN, use LDEXP directly. | |||
2518 | if (Subtarget->isGCN()) | |||
2519 | return DAG.getNode(AMDGPUISD::LDEXP, SL, MVT::f32, FVal, ShAmt); | |||
2520 | ||||
2521 | // Otherwise, align 'ShAmt' to the exponent part and add it into the exponent | |||
2522 | // part directly to emulate the multiplication of 2^ShAmt. That 8-bit | |||
2523 | // exponent is enough to avoid overflowing into the sign bit. | |||
2524 | SDValue Exp = DAG.getNode(ISD::SHL, SL, MVT::i32, ShAmt, | |||
2525 | DAG.getConstant(23, SL, MVT::i32)); | |||
2526 | SDValue IVal = | |||
2527 | DAG.getNode(ISD::ADD, SL, MVT::i32, | |||
2528 | DAG.getNode(ISD::BITCAST, SL, MVT::i32, FVal), Exp); | |||
2529 | if (Signed) { | |||
2530 | // Set the sign bit. | |||
2531 | Sign = DAG.getNode(ISD::SHL, SL, MVT::i32, | |||
2532 | DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, Sign), | |||
2533 | DAG.getConstant(31, SL, MVT::i32)); | |||
2534 | IVal = DAG.getNode(ISD::OR, SL, MVT::i32, IVal, Sign); | |||
2535 | } | |||
2536 | return DAG.getNode(ISD::BITCAST, SL, MVT::f32, IVal); | |||
2537 | } | |||
2538 | ||||
2539 | SDValue AMDGPUTargetLowering::LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG, | |||
2540 | bool Signed) const { | |||
2541 | SDLoc SL(Op); | |||
2542 | SDValue Src = Op.getOperand(0); | |||
2543 | ||||
2544 | SDValue Lo, Hi; | |||
2545 | std::tie(Lo, Hi) = split64BitValue(Src, DAG); | |||
2546 | ||||
2547 | SDValue CvtHi = DAG.getNode(Signed ? ISD::SINT_TO_FP : ISD::UINT_TO_FP, | |||
2548 | SL, MVT::f64, Hi); | |||
2549 | ||||
2550 | SDValue CvtLo = DAG.getNode(ISD::UINT_TO_FP, SL, MVT::f64, Lo); | |||
2551 | ||||
2552 | SDValue LdExp = DAG.getNode(AMDGPUISD::LDEXP, SL, MVT::f64, CvtHi, | |||
2553 | DAG.getConstant(32, SL, MVT::i32)); | |||
2554 | // TODO: Should this propagate fast-math-flags? | |||
2555 | return DAG.getNode(ISD::FADD, SL, MVT::f64, LdExp, CvtLo); | |||
2556 | } | |||
2557 | ||||
2558 | SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op, | |||
2559 | SelectionDAG &DAG) const { | |||
2560 | // TODO: Factor out code common with LowerSINT_TO_FP. | |||
2561 | EVT DestVT = Op.getValueType(); | |||
2562 | SDValue Src = Op.getOperand(0); | |||
2563 | EVT SrcVT = Src.getValueType(); | |||
2564 | ||||
2565 | if (SrcVT == MVT::i16) { | |||
2566 | if (DestVT == MVT::f16) | |||
2567 | return Op; | |||
2568 | SDLoc DL(Op); | |||
2569 | ||||
2570 | // Promote src to i32 | |||
2571 | SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, Src); | |||
2572 | return DAG.getNode(ISD::UINT_TO_FP, DL, DestVT, Ext); | |||
2573 | } | |||
2574 | ||||
2575 | assert(SrcVT == MVT::i64 && "operation should be legal")(static_cast <bool> (SrcVT == MVT::i64 && "operation should be legal" ) ? void (0) : __assert_fail ("SrcVT == MVT::i64 && \"operation should be legal\"" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 2575, __extension__ __PRETTY_FUNCTION__)); | |||
2576 | ||||
2577 | if (Subtarget->has16BitInsts() && DestVT == MVT::f16) { | |||
2578 | SDLoc DL(Op); | |||
2579 | ||||
2580 | SDValue IntToFp32 = DAG.getNode(Op.getOpcode(), DL, MVT::f32, Src); | |||
2581 | SDValue FPRoundFlag = DAG.getIntPtrConstant(0, SDLoc(Op)); | |||
2582 | SDValue FPRound = | |||
2583 | DAG.getNode(ISD::FP_ROUND, DL, MVT::f16, IntToFp32, FPRoundFlag); | |||
2584 | ||||
2585 | return FPRound; | |||
2586 | } | |||
2587 | ||||
2588 | if (DestVT == MVT::f32) | |||
2589 | return LowerINT_TO_FP32(Op, DAG, false); | |||
2590 | ||||
2591 | assert(DestVT == MVT::f64)(static_cast <bool> (DestVT == MVT::f64) ? void (0) : __assert_fail ("DestVT == MVT::f64", "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp" , 2591, __extension__ __PRETTY_FUNCTION__)); | |||
2592 | return LowerINT_TO_FP64(Op, DAG, false); | |||
2593 | } | |||
2594 | ||||
2595 | SDValue AMDGPUTargetLowering::LowerSINT_TO_FP(SDValue Op, | |||
2596 | SelectionDAG &DAG) const { | |||
2597 | EVT DestVT = Op.getValueType(); | |||
2598 | ||||
2599 | SDValue Src = Op.getOperand(0); | |||
2600 | EVT SrcVT = Src.getValueType(); | |||
2601 | ||||
2602 | if (SrcVT == MVT::i16) { | |||
2603 | if (DestVT == MVT::f16) | |||
2604 | return Op; | |||
2605 | ||||
2606 | SDLoc DL(Op); | |||
2607 | // Promote src to i32 | |||
2608 | SDValue Ext = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32, Src); | |||
2609 | return DAG.getNode(ISD::SINT_TO_FP, DL, DestVT, Ext); | |||
2610 | } | |||
2611 | ||||
2612 | assert(SrcVT == MVT::i64 && "operation should be legal")(static_cast <bool> (SrcVT == MVT::i64 && "operation should be legal" ) ? void (0) : __assert_fail ("SrcVT == MVT::i64 && \"operation should be legal\"" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 2612, __extension__ __PRETTY_FUNCTION__)); | |||
2613 | ||||
2614 | // TODO: Factor out code common with LowerUINT_TO_FP. | |||
2615 | ||||
2616 | if (Subtarget->has16BitInsts() && DestVT == MVT::f16) { | |||
2617 | SDLoc DL(Op); | |||
2618 | SDValue Src = Op.getOperand(0); | |||
2619 | ||||
2620 | SDValue IntToFp32 = DAG.getNode(Op.getOpcode(), DL, MVT::f32, Src); | |||
2621 | SDValue FPRoundFlag = DAG.getIntPtrConstant(0, SDLoc(Op)); | |||
2622 | SDValue FPRound = | |||
2623 | DAG.getNode(ISD::FP_ROUND, DL, MVT::f16, IntToFp32, FPRoundFlag); | |||
2624 | ||||
2625 | return FPRound; | |||
2626 | } | |||
2627 | ||||
2628 | if (DestVT == MVT::f32) | |||
2629 | return LowerINT_TO_FP32(Op, DAG, true); | |||
2630 | ||||
2631 | assert(DestVT == MVT::f64)(static_cast <bool> (DestVT == MVT::f64) ? void (0) : __assert_fail ("DestVT == MVT::f64", "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp" , 2631, __extension__ __PRETTY_FUNCTION__)); | |||
2632 | return LowerINT_TO_FP64(Op, DAG, true); | |||
2633 | } | |||
2634 | ||||
2635 | SDValue AMDGPUTargetLowering::LowerFP_TO_INT64(SDValue Op, SelectionDAG &DAG, | |||
2636 | bool Signed) const { | |||
2637 | SDLoc SL(Op); | |||
2638 | ||||
2639 | SDValue Src = Op.getOperand(0); | |||
2640 | EVT SrcVT = Src.getValueType(); | |||
2641 | ||||
2642 | assert(SrcVT == MVT::f32 || SrcVT == MVT::f64)(static_cast <bool> (SrcVT == MVT::f32 || SrcVT == MVT:: f64) ? void (0) : __assert_fail ("SrcVT == MVT::f32 || SrcVT == MVT::f64" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 2642, __extension__ __PRETTY_FUNCTION__)); | |||
2643 | ||||
2644 | // The basic idea of converting a floating point number into a pair of 32-bit | |||
2645 | // integers is illustrated as follows: | |||
2646 | // | |||
2647 | // tf := trunc(val); | |||
2648 | // hif := floor(tf * 2^-32); | |||
2649 | // lof := tf - hif * 2^32; // lof is always positive due to floor. | |||
2650 | // hi := fptoi(hif); | |||
2651 | // lo := fptoi(lof); | |||
2652 | // | |||
2653 | SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, SrcVT, Src); | |||
2654 | SDValue Sign; | |||
2655 | if (Signed && SrcVT == MVT::f32) { | |||
2656 | // However, a 32-bit floating point number has only 23 bits mantissa and | |||
2657 | // it's not enough to hold all the significant bits of `lof` if val is | |||
2658 | // negative. To avoid the loss of precision, We need to take the absolute | |||
2659 | // value after truncating and flip the result back based on the original | |||
2660 | // signedness. | |||
2661 | Sign = DAG.getNode(ISD::SRA, SL, MVT::i32, | |||
2662 | DAG.getNode(ISD::BITCAST, SL, MVT::i32, Trunc), | |||
2663 | DAG.getConstant(31, SL, MVT::i32)); | |||
2664 | Trunc = DAG.getNode(ISD::FABS, SL, SrcVT, Trunc); | |||
2665 | } | |||
2666 | ||||
2667 | SDValue K0, K1; | |||
2668 | if (SrcVT == MVT::f64) { | |||
2669 | K0 = DAG.getConstantFP(BitsToDouble(UINT64_C(/*2^-32*/ 0x3df0000000000000)0x3df0000000000000UL), | |||
2670 | SL, SrcVT); | |||
2671 | K1 = DAG.getConstantFP(BitsToDouble(UINT64_C(/*-2^32*/ 0xc1f0000000000000)0xc1f0000000000000UL), | |||
2672 | SL, SrcVT); | |||
2673 | } else { | |||
2674 | K0 = DAG.getConstantFP(BitsToFloat(UINT32_C(/*2^-32*/ 0x2f800000)0x2f800000U), SL, | |||
2675 | SrcVT); | |||
2676 | K1 = DAG.getConstantFP(BitsToFloat(UINT32_C(/*-2^32*/ 0xcf800000)0xcf800000U), SL, | |||
2677 | SrcVT); | |||
2678 | } | |||
2679 | // TODO: Should this propagate fast-math-flags? | |||
2680 | SDValue Mul = DAG.getNode(ISD::FMUL, SL, SrcVT, Trunc, K0); | |||
2681 | ||||
2682 | SDValue FloorMul = DAG.getNode(ISD::FFLOOR, SL, SrcVT, Mul); | |||
2683 | ||||
2684 | SDValue Fma = DAG.getNode(ISD::FMA, SL, SrcVT, FloorMul, K1, Trunc); | |||
2685 | ||||
2686 | SDValue Hi = DAG.getNode((Signed && SrcVT == MVT::f64) ? ISD::FP_TO_SINT | |||
2687 | : ISD::FP_TO_UINT, | |||
2688 | SL, MVT::i32, FloorMul); | |||
2689 | SDValue Lo = DAG.getNode(ISD::FP_TO_UINT, SL, MVT::i32, Fma); | |||
2690 | ||||
2691 | SDValue Result = DAG.getNode(ISD::BITCAST, SL, MVT::i64, | |||
2692 | DAG.getBuildVector(MVT::v2i32, SL, {Lo, Hi})); | |||
2693 | ||||
2694 | if (Signed && SrcVT == MVT::f32) { | |||
2695 | assert(Sign)(static_cast <bool> (Sign) ? void (0) : __assert_fail ( "Sign", "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 2695 , __extension__ __PRETTY_FUNCTION__)); | |||
2696 | // Flip the result based on the signedness, which is either all 0s or 1s. | |||
2697 | Sign = DAG.getNode(ISD::BITCAST, SL, MVT::i64, | |||
2698 | DAG.getBuildVector(MVT::v2i32, SL, {Sign, Sign})); | |||
2699 | // r := xor(r, sign) - sign; | |||
2700 | Result = | |||
2701 | DAG.getNode(ISD::SUB, SL, MVT::i64, | |||
2702 | DAG.getNode(ISD::XOR, SL, MVT::i64, Result, Sign), Sign); | |||
2703 | } | |||
2704 | ||||
2705 | return Result; | |||
2706 | } | |||
2707 | ||||
2708 | SDValue AMDGPUTargetLowering::LowerFP_TO_FP16(SDValue Op, SelectionDAG &DAG) const { | |||
2709 | SDLoc DL(Op); | |||
2710 | SDValue N0 = Op.getOperand(0); | |||
2711 | ||||
2712 | // Convert to target node to get known bits | |||
2713 | if (N0.getValueType() == MVT::f32) | |||
2714 | return DAG.getNode(AMDGPUISD::FP_TO_FP16, DL, Op.getValueType(), N0); | |||
2715 | ||||
2716 | if (getTargetMachine().Options.UnsafeFPMath) { | |||
2717 | // There is a generic expand for FP_TO_FP16 with unsafe fast math. | |||
2718 | return SDValue(); | |||
2719 | } | |||
2720 | ||||
2721 | assert(N0.getSimpleValueType() == MVT::f64)(static_cast <bool> (N0.getSimpleValueType() == MVT::f64 ) ? void (0) : __assert_fail ("N0.getSimpleValueType() == MVT::f64" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 2721, __extension__ __PRETTY_FUNCTION__)); | |||
2722 | ||||
2723 | // f64 -> f16 conversion using round-to-nearest-even rounding mode. | |||
2724 | const unsigned ExpMask = 0x7ff; | |||
2725 | const unsigned ExpBiasf64 = 1023; | |||
2726 | const unsigned ExpBiasf16 = 15; | |||
2727 | SDValue Zero = DAG.getConstant(0, DL, MVT::i32); | |||
2728 | SDValue One = DAG.getConstant(1, DL, MVT::i32); | |||
2729 | SDValue U = DAG.getNode(ISD::BITCAST, DL, MVT::i64, N0); | |||
2730 | SDValue UH = DAG.getNode(ISD::SRL, DL, MVT::i64, U, | |||
2731 | DAG.getConstant(32, DL, MVT::i64)); | |||
2732 | UH = DAG.getZExtOrTrunc(UH, DL, MVT::i32); | |||
2733 | U = DAG.getZExtOrTrunc(U, DL, MVT::i32); | |||
2734 | SDValue E = DAG.getNode(ISD::SRL, DL, MVT::i32, UH, | |||
2735 | DAG.getConstant(20, DL, MVT::i64)); | |||
2736 | E = DAG.getNode(ISD::AND, DL, MVT::i32, E, | |||
2737 | DAG.getConstant(ExpMask, DL, MVT::i32)); | |||
2738 | // Subtract the fp64 exponent bias (1023) to get the real exponent and | |||
2739 | // add the f16 bias (15) to get the biased exponent for the f16 format. | |||
2740 | E = DAG.getNode(ISD::ADD, DL, MVT::i32, E, | |||
2741 | DAG.getConstant(-ExpBiasf64 + ExpBiasf16, DL, MVT::i32)); | |||
2742 | ||||
2743 | SDValue M = DAG.getNode(ISD::SRL, DL, MVT::i32, UH, | |||
2744 | DAG.getConstant(8, DL, MVT::i32)); | |||
2745 | M = DAG.getNode(ISD::AND, DL, MVT::i32, M, | |||
2746 | DAG.getConstant(0xffe, DL, MVT::i32)); | |||
2747 | ||||
2748 | SDValue MaskedSig = DAG.getNode(ISD::AND, DL, MVT::i32, UH, | |||
2749 | DAG.getConstant(0x1ff, DL, MVT::i32)); | |||
2750 | MaskedSig = DAG.getNode(ISD::OR, DL, MVT::i32, MaskedSig, U); | |||
2751 | ||||
2752 | SDValue Lo40Set = DAG.getSelectCC(DL, MaskedSig, Zero, Zero, One, ISD::SETEQ); | |||
2753 | M = DAG.getNode(ISD::OR, DL, MVT::i32, M, Lo40Set); | |||
2754 | ||||
2755 | // (M != 0 ? 0x0200 : 0) | 0x7c00; | |||
2756 | SDValue I = DAG.getNode(ISD::OR, DL, MVT::i32, | |||
2757 | DAG.getSelectCC(DL, M, Zero, DAG.getConstant(0x0200, DL, MVT::i32), | |||
2758 | Zero, ISD::SETNE), DAG.getConstant(0x7c00, DL, MVT::i32)); | |||
2759 | ||||
2760 | // N = M | (E << 12); | |||
2761 | SDValue N = DAG.getNode(ISD::OR, DL, MVT::i32, M, | |||
2762 | DAG.getNode(ISD::SHL, DL, MVT::i32, E, | |||
2763 | DAG.getConstant(12, DL, MVT::i32))); | |||
2764 | ||||
2765 | // B = clamp(1-E, 0, 13); | |||
2766 | SDValue OneSubExp = DAG.getNode(ISD::SUB, DL, MVT::i32, | |||
2767 | One, E); | |||
2768 | SDValue B = DAG.getNode(ISD::SMAX, DL, MVT::i32, OneSubExp, Zero); | |||
2769 | B = DAG.getNode(ISD::SMIN, DL, MVT::i32, B, | |||
2770 | DAG.getConstant(13, DL, MVT::i32)); | |||
2771 | ||||
2772 | SDValue SigSetHigh = DAG.getNode(ISD::OR, DL, MVT::i32, M, | |||
2773 | DAG.getConstant(0x1000, DL, MVT::i32)); | |||
2774 | ||||
2775 | SDValue D = DAG.getNode(ISD::SRL, DL, MVT::i32, SigSetHigh, B); | |||
2776 | SDValue D0 = DAG.getNode(ISD::SHL, DL, MVT::i32, D, B); | |||
2777 | SDValue D1 = DAG.getSelectCC(DL, D0, SigSetHigh, One, Zero, ISD::SETNE); | |||
2778 | D = DAG.getNode(ISD::OR, DL, MVT::i32, D, D1); | |||
2779 | ||||
2780 | SDValue V = DAG.getSelectCC(DL, E, One, D, N, ISD::SETLT); | |||
2781 | SDValue VLow3 = DAG.getNode(ISD::AND, DL, MVT::i32, V, | |||
2782 | DAG.getConstant(0x7, DL, MVT::i32)); | |||
2783 | V = DAG.getNode(ISD::SRL, DL, MVT::i32, V, | |||
2784 | DAG.getConstant(2, DL, MVT::i32)); | |||
2785 | SDValue V0 = DAG.getSelectCC(DL, VLow3, DAG.getConstant(3, DL, MVT::i32), | |||
2786 | One, Zero, ISD::SETEQ); | |||
2787 | SDValue V1 = DAG.getSelectCC(DL, VLow3, DAG.getConstant(5, DL, MVT::i32), | |||
2788 | One, Zero, ISD::SETGT); | |||
2789 | V1 = DAG.getNode(ISD::OR, DL, MVT::i32, V0, V1); | |||
2790 | V = DAG.getNode(ISD::ADD, DL, MVT::i32, V, V1); | |||
2791 | ||||
2792 | V = DAG.getSelectCC(DL, E, DAG.getConstant(30, DL, MVT::i32), | |||
2793 | DAG.getConstant(0x7c00, DL, MVT::i32), V, ISD::SETGT); | |||
2794 | V = DAG.getSelectCC(DL, E, DAG.getConstant(1039, DL, MVT::i32), | |||
2795 | I, V, ISD::SETEQ); | |||
2796 | ||||
2797 | // Extract the sign bit. | |||
2798 | SDValue Sign = DAG.getNode(ISD::SRL, DL, MVT::i32, UH, | |||
2799 | DAG.getConstant(16, DL, MVT::i32)); | |||
2800 | Sign = DAG.getNode(ISD::AND, DL, MVT::i32, Sign, | |||
2801 | DAG.getConstant(0x8000, DL, MVT::i32)); | |||
2802 | ||||
2803 | V = DAG.getNode(ISD::OR, DL, MVT::i32, Sign, V); | |||
2804 | return DAG.getZExtOrTrunc(V, DL, Op.getValueType()); | |||
2805 | } | |||
2806 | ||||
2807 | SDValue AMDGPUTargetLowering::LowerFP_TO_INT(SDValue Op, | |||
2808 | SelectionDAG &DAG) const { | |||
2809 | SDValue Src = Op.getOperand(0); | |||
2810 | unsigned OpOpcode = Op.getOpcode(); | |||
2811 | EVT SrcVT = Src.getValueType(); | |||
2812 | EVT DestVT = Op.getValueType(); | |||
2813 | ||||
2814 | // Will be selected natively | |||
2815 | if (SrcVT == MVT::f16 && DestVT == MVT::i16) | |||
2816 | return Op; | |||
2817 | ||||
2818 | // Promote i16 to i32 | |||
2819 | if (DestVT == MVT::i16 && (SrcVT == MVT::f32 || SrcVT == MVT::f64)) { | |||
2820 | SDLoc DL(Op); | |||
2821 | ||||
2822 | SDValue FpToInt32 = DAG.getNode(OpOpcode, DL, MVT::i32, Src); | |||
2823 | return DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, FpToInt32); | |||
2824 | } | |||
2825 | ||||
2826 | if (SrcVT == MVT::f16 || | |||
2827 | (SrcVT == MVT::f32 && Src.getOpcode() == ISD::FP16_TO_FP)) { | |||
2828 | SDLoc DL(Op); | |||
2829 | ||||
2830 | SDValue FpToInt32 = DAG.getNode(OpOpcode, DL, MVT::i32, Src); | |||
2831 | unsigned Ext = | |||
2832 | OpOpcode == ISD::FP_TO_SINT ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; | |||
2833 | return DAG.getNode(Ext, DL, MVT::i64, FpToInt32); | |||
2834 | } | |||
2835 | ||||
2836 | if (DestVT == MVT::i64 && (SrcVT == MVT::f32 || SrcVT == MVT::f64)) | |||
2837 | return LowerFP_TO_INT64(Op, DAG, OpOpcode == ISD::FP_TO_SINT); | |||
2838 | ||||
2839 | return SDValue(); | |||
2840 | } | |||
2841 | ||||
2842 | SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op, | |||
2843 | SelectionDAG &DAG) const { | |||
2844 | EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); | |||
2845 | MVT VT = Op.getSimpleValueType(); | |||
2846 | MVT ScalarVT = VT.getScalarType(); | |||
2847 | ||||
2848 | assert(VT.isVector())(static_cast <bool> (VT.isVector()) ? void (0) : __assert_fail ("VT.isVector()", "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp" , 2848, __extension__ __PRETTY_FUNCTION__)); | |||
2849 | ||||
2850 | SDValue Src = Op.getOperand(0); | |||
2851 | SDLoc DL(Op); | |||
2852 | ||||
2853 | // TODO: Don't scalarize on Evergreen? | |||
2854 | unsigned NElts = VT.getVectorNumElements(); | |||
2855 | SmallVector<SDValue, 8> Args; | |||
2856 | DAG.ExtractVectorElements(Src, Args, 0, NElts); | |||
2857 | ||||
2858 | SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType()); | |||
2859 | for (unsigned I = 0; I < NElts; ++I) | |||
2860 | Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp); | |||
2861 | ||||
2862 | return DAG.getBuildVector(VT, DL, Args); | |||
2863 | } | |||
2864 | ||||
2865 | //===----------------------------------------------------------------------===// | |||
2866 | // Custom DAG optimizations | |||
2867 | //===----------------------------------------------------------------------===// | |||
2868 | ||||
2869 | static bool isU24(SDValue Op, SelectionDAG &DAG) { | |||
2870 | return AMDGPUTargetLowering::numBitsUnsigned(Op, DAG) <= 24; | |||
2871 | } | |||
2872 | ||||
2873 | static bool isI24(SDValue Op, SelectionDAG &DAG) { | |||
2874 | EVT VT = Op.getValueType(); | |||
2875 | return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated | |||
2876 | // as unsigned 24-bit values. | |||
2877 | AMDGPUTargetLowering::numBitsSigned(Op, DAG) <= 24; | |||
2878 | } | |||
2879 | ||||
2880 | static SDValue simplifyMul24(SDNode *Node24, | |||
2881 | TargetLowering::DAGCombinerInfo &DCI) { | |||
2882 | SelectionDAG &DAG = DCI.DAG; | |||
2883 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
2884 | bool IsIntrin = Node24->getOpcode() == ISD::INTRINSIC_WO_CHAIN; | |||
2885 | ||||
2886 | SDValue LHS = IsIntrin ? Node24->getOperand(1) : Node24->getOperand(0); | |||
2887 | SDValue RHS = IsIntrin ? Node24->getOperand(2) : Node24->getOperand(1); | |||
2888 | unsigned NewOpcode = Node24->getOpcode(); | |||
2889 | if (IsIntrin) { | |||
2890 | unsigned IID = cast<ConstantSDNode>(Node24->getOperand(0))->getZExtValue(); | |||
2891 | switch (IID) { | |||
2892 | case Intrinsic::amdgcn_mul_i24: | |||
2893 | NewOpcode = AMDGPUISD::MUL_I24; | |||
2894 | break; | |||
2895 | case Intrinsic::amdgcn_mul_u24: | |||
2896 | NewOpcode = AMDGPUISD::MUL_U24; | |||
2897 | break; | |||
2898 | case Intrinsic::amdgcn_mulhi_i24: | |||
2899 | NewOpcode = AMDGPUISD::MULHI_I24; | |||
2900 | break; | |||
2901 | case Intrinsic::amdgcn_mulhi_u24: | |||
2902 | NewOpcode = AMDGPUISD::MULHI_U24; | |||
2903 | break; | |||
2904 | default: | |||
2905 | llvm_unreachable("Expected 24-bit mul intrinsic")::llvm::llvm_unreachable_internal("Expected 24-bit mul intrinsic" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 2905); | |||
2906 | } | |||
2907 | } | |||
2908 | ||||
2909 | APInt Demanded = APInt::getLowBitsSet(LHS.getValueSizeInBits(), 24); | |||
2910 | ||||
2911 | // First try to simplify using SimplifyMultipleUseDemandedBits which allows | |||
2912 | // the operands to have other uses, but will only perform simplifications that | |||
2913 | // involve bypassing some nodes for this user. | |||
2914 | SDValue DemandedLHS = TLI.SimplifyMultipleUseDemandedBits(LHS, Demanded, DAG); | |||
2915 | SDValue DemandedRHS = TLI.SimplifyMultipleUseDemandedBits(RHS, Demanded, DAG); | |||
2916 | if (DemandedLHS || DemandedRHS) | |||
2917 | return DAG.getNode(NewOpcode, SDLoc(Node24), Node24->getVTList(), | |||
2918 | DemandedLHS ? DemandedLHS : LHS, | |||
2919 | DemandedRHS ? DemandedRHS : RHS); | |||
2920 | ||||
2921 | // Now try SimplifyDemandedBits which can simplify the nodes used by our | |||
2922 | // operands if this node is the only user. | |||
2923 | if (TLI.SimplifyDemandedBits(LHS, Demanded, DCI)) | |||
2924 | return SDValue(Node24, 0); | |||
2925 | if (TLI.SimplifyDemandedBits(RHS, Demanded, DCI)) | |||
2926 | return SDValue(Node24, 0); | |||
2927 | ||||
2928 | return SDValue(); | |||
2929 | } | |||
2930 | ||||
2931 | template <typename IntTy> | |||
2932 | static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0, uint32_t Offset, | |||
2933 | uint32_t Width, const SDLoc &DL) { | |||
2934 | if (Width + Offset < 32) { | |||
2935 | uint32_t Shl = static_cast<uint32_t>(Src0) << (32 - Offset - Width); | |||
2936 | IntTy Result = static_cast<IntTy>(Shl) >> (32 - Width); | |||
2937 | return DAG.getConstant(Result, DL, MVT::i32); | |||
2938 | } | |||
2939 | ||||
2940 | return DAG.getConstant(Src0 >> Offset, DL, MVT::i32); | |||
2941 | } | |||
2942 | ||||
2943 | static bool hasVolatileUser(SDNode *Val) { | |||
2944 | for (SDNode *U : Val->uses()) { | |||
2945 | if (MemSDNode *M = dyn_cast<MemSDNode>(U)) { | |||
2946 | if (M->isVolatile()) | |||
2947 | return true; | |||
2948 | } | |||
2949 | } | |||
2950 | ||||
2951 | return false; | |||
2952 | } | |||
2953 | ||||
2954 | bool AMDGPUTargetLowering::shouldCombineMemoryType(EVT VT) const { | |||
2955 | // i32 vectors are the canonical memory type. | |||
2956 | if (VT.getScalarType() == MVT::i32 || isTypeLegal(VT)) | |||
2957 | return false; | |||
2958 | ||||
2959 | if (!VT.isByteSized()) | |||
2960 | return false; | |||
2961 | ||||
2962 | unsigned Size = VT.getStoreSize(); | |||
2963 | ||||
2964 | if ((Size == 1 || Size == 2 || Size == 4) && !VT.isVector()) | |||
2965 | return false; | |||
2966 | ||||
2967 | if (Size == 3 || (Size > 4 && (Size % 4 != 0))) | |||
2968 | return false; | |||
2969 | ||||
2970 | return true; | |||
2971 | } | |||
2972 | ||||
2973 | // Replace load of an illegal type with a store of a bitcast to a friendlier | |||
2974 | // type. | |||
2975 | SDValue AMDGPUTargetLowering::performLoadCombine(SDNode *N, | |||
2976 | DAGCombinerInfo &DCI) const { | |||
2977 | if (!DCI.isBeforeLegalize()) | |||
2978 | return SDValue(); | |||
2979 | ||||
2980 | LoadSDNode *LN = cast<LoadSDNode>(N); | |||
2981 | if (!LN->isSimple() || !ISD::isNormalLoad(LN) || hasVolatileUser(LN)) | |||
2982 | return SDValue(); | |||
2983 | ||||
2984 | SDLoc SL(N); | |||
2985 | SelectionDAG &DAG = DCI.DAG; | |||
2986 | EVT VT = LN->getMemoryVT(); | |||
2987 | ||||
2988 | unsigned Size = VT.getStoreSize(); | |||
2989 | Align Alignment = LN->getAlign(); | |||
2990 | if (Alignment < Size && isTypeLegal(VT)) { | |||
2991 | bool IsFast; | |||
2992 | unsigned AS = LN->getAddressSpace(); | |||
2993 | ||||
2994 | // Expand unaligned loads earlier than legalization. Due to visitation order | |||
2995 | // problems during legalization, the emitted instructions to pack and unpack | |||
2996 | // the bytes again are not eliminated in the case of an unaligned copy. | |||
2997 | if (!allowsMisalignedMemoryAccesses( | |||
2998 | VT, AS, Alignment, LN->getMemOperand()->getFlags(), &IsFast)) { | |||
2999 | if (VT.isVector()) | |||
3000 | return SplitVectorLoad(SDValue(LN, 0), DAG); | |||
3001 | ||||
3002 | SDValue Ops[2]; | |||
3003 | std::tie(Ops[0], Ops[1]) = expandUnalignedLoad(LN, DAG); | |||
3004 | ||||
3005 | return DAG.getMergeValues(Ops, SDLoc(N)); | |||
3006 | } | |||
3007 | ||||
3008 | if (!IsFast) | |||
3009 | return SDValue(); | |||
3010 | } | |||
3011 | ||||
3012 | if (!shouldCombineMemoryType(VT)) | |||
3013 | return SDValue(); | |||
3014 | ||||
3015 | EVT NewVT = getEquivalentMemType(*DAG.getContext(), VT); | |||
3016 | ||||
3017 | SDValue NewLoad | |||
3018 | = DAG.getLoad(NewVT, SL, LN->getChain(), | |||
3019 | LN->getBasePtr(), LN->getMemOperand()); | |||
3020 | ||||
3021 | SDValue BC = DAG.getNode(ISD::BITCAST, SL, VT, NewLoad); | |||
3022 | DCI.CombineTo(N, BC, NewLoad.getValue(1)); | |||
3023 | return SDValue(N, 0); | |||
3024 | } | |||
3025 | ||||
3026 | // Replace store of an illegal type with a store of a bitcast to a friendlier | |||
3027 | // type. | |||
3028 | SDValue AMDGPUTargetLowering::performStoreCombine(SDNode *N, | |||
3029 | DAGCombinerInfo &DCI) const { | |||
3030 | if (!DCI.isBeforeLegalize()) | |||
3031 | return SDValue(); | |||
3032 | ||||
3033 | StoreSDNode *SN = cast<StoreSDNode>(N); | |||
3034 | if (!SN->isSimple() || !ISD::isNormalStore(SN)) | |||
3035 | return SDValue(); | |||
3036 | ||||
3037 | EVT VT = SN->getMemoryVT(); | |||
3038 | unsigned Size = VT.getStoreSize(); | |||
3039 | ||||
3040 | SDLoc SL(N); | |||
3041 | SelectionDAG &DAG = DCI.DAG; | |||
3042 | Align Alignment = SN->getAlign(); | |||
3043 | if (Alignment < Size && isTypeLegal(VT)) { | |||
3044 | bool IsFast; | |||
3045 | unsigned AS = SN->getAddressSpace(); | |||
3046 | ||||
3047 | // Expand unaligned stores earlier than legalization. Due to visitation | |||
3048 | // order problems during legalization, the emitted instructions to pack and | |||
3049 | // unpack the bytes again are not eliminated in the case of an unaligned | |||
3050 | // copy. | |||
3051 | if (!allowsMisalignedMemoryAccesses( | |||
3052 | VT, AS, Alignment, SN->getMemOperand()->getFlags(), &IsFast)) { | |||
3053 | if (VT.isVector()) | |||
3054 | return SplitVectorStore(SDValue(SN, 0), DAG); | |||
3055 | ||||
3056 | return expandUnalignedStore(SN, DAG); | |||
3057 | } | |||
3058 | ||||
3059 | if (!IsFast) | |||
3060 | return SDValue(); | |||
3061 | } | |||
3062 | ||||
3063 | if (!shouldCombineMemoryType(VT)) | |||
3064 | return SDValue(); | |||
3065 | ||||
3066 | EVT NewVT = getEquivalentMemType(*DAG.getContext(), VT); | |||
3067 | SDValue Val = SN->getValue(); | |||
3068 | ||||
3069 | //DCI.AddToWorklist(Val.getNode()); | |||
3070 | ||||
3071 | bool OtherUses = !Val.hasOneUse(); | |||
3072 | SDValue CastVal = DAG.getNode(ISD::BITCAST, SL, NewVT, Val); | |||
3073 | if (OtherUses) { | |||
3074 | SDValue CastBack = DAG.getNode(ISD::BITCAST, SL, VT, CastVal); | |||
3075 | DAG.ReplaceAllUsesOfValueWith(Val, CastBack); | |||
3076 | } | |||
3077 | ||||
3078 | return DAG.getStore(SN->getChain(), SL, CastVal, | |||
3079 | SN->getBasePtr(), SN->getMemOperand()); | |||
3080 | } | |||
3081 | ||||
3082 | // FIXME: This should go in generic DAG combiner with an isTruncateFree check, | |||
3083 | // but isTruncateFree is inaccurate for i16 now because of SALU vs. VALU | |||
3084 | // issues. | |||
3085 | SDValue AMDGPUTargetLowering::performAssertSZExtCombine(SDNode *N, | |||
3086 | DAGCombinerInfo &DCI) const { | |||
3087 | SelectionDAG &DAG = DCI.DAG; | |||
3088 | SDValue N0 = N->getOperand(0); | |||
3089 | ||||
3090 | // (vt2 (assertzext (truncate vt0:x), vt1)) -> | |||
3091 | // (vt2 (truncate (assertzext vt0:x, vt1))) | |||
3092 | if (N0.getOpcode() == ISD::TRUNCATE) { | |||
3093 | SDValue N1 = N->getOperand(1); | |||
3094 | EVT ExtVT = cast<VTSDNode>(N1)->getVT(); | |||
3095 | SDLoc SL(N); | |||
3096 | ||||
3097 | SDValue Src = N0.getOperand(0); | |||
3098 | EVT SrcVT = Src.getValueType(); | |||
3099 | if (SrcVT.bitsGE(ExtVT)) { | |||
3100 | SDValue NewInReg = DAG.getNode(N->getOpcode(), SL, SrcVT, Src, N1); | |||
3101 | return DAG.getNode(ISD::TRUNCATE, SL, N->getValueType(0), NewInReg); | |||
3102 | } | |||
3103 | } | |||
3104 | ||||
3105 | return SDValue(); | |||
3106 | } | |||
3107 | ||||
3108 | SDValue AMDGPUTargetLowering::performIntrinsicWOChainCombine( | |||
3109 | SDNode *N, DAGCombinerInfo &DCI) const { | |||
3110 | unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); | |||
3111 | switch (IID) { | |||
3112 | case Intrinsic::amdgcn_mul_i24: | |||
3113 | case Intrinsic::amdgcn_mul_u24: | |||
3114 | case Intrinsic::amdgcn_mulhi_i24: | |||
3115 | case Intrinsic::amdgcn_mulhi_u24: | |||
3116 | return simplifyMul24(N, DCI); | |||
3117 | case Intrinsic::amdgcn_fract: | |||
3118 | case Intrinsic::amdgcn_rsq: | |||
3119 | case Intrinsic::amdgcn_rcp_legacy: | |||
3120 | case Intrinsic::amdgcn_rsq_legacy: | |||
3121 | case Intrinsic::amdgcn_rsq_clamp: | |||
3122 | case Intrinsic::amdgcn_ldexp: { | |||
3123 | // FIXME: This is probably wrong. If src is an sNaN, it won't be quieted | |||
3124 | SDValue Src = N->getOperand(1); | |||
3125 | return Src.isUndef() ? Src : SDValue(); | |||
3126 | } | |||
3127 | default: | |||
3128 | return SDValue(); | |||
3129 | } | |||
3130 | } | |||
3131 | ||||
3132 | /// Split the 64-bit value \p LHS into two 32-bit components, and perform the | |||
3133 | /// binary operation \p Opc to it with the corresponding constant operands. | |||
3134 | SDValue AMDGPUTargetLowering::splitBinaryBitConstantOpImpl( | |||
3135 | DAGCombinerInfo &DCI, const SDLoc &SL, | |||
3136 | unsigned Opc, SDValue LHS, | |||
3137 | uint32_t ValLo, uint32_t ValHi) const { | |||
3138 | SelectionDAG &DAG = DCI.DAG; | |||
3139 | SDValue Lo, Hi; | |||
3140 | std::tie(Lo, Hi) = split64BitValue(LHS, DAG); | |||
3141 | ||||
3142 | SDValue LoRHS = DAG.getConstant(ValLo, SL, MVT::i32); | |||
3143 | SDValue HiRHS = DAG.getConstant(ValHi, SL, MVT::i32); | |||
3144 | ||||
3145 | SDValue LoAnd = DAG.getNode(Opc, SL, MVT::i32, Lo, LoRHS); | |||
3146 | SDValue HiAnd = DAG.getNode(Opc, SL, MVT::i32, Hi, HiRHS); | |||
3147 | ||||
3148 | // Re-visit the ands. It's possible we eliminated one of them and it could | |||
3149 | // simplify the vector. | |||
3150 | DCI.AddToWorklist(Lo.getNode()); | |||
3151 | DCI.AddToWorklist(Hi.getNode()); | |||
3152 | ||||
3153 | SDValue Vec = DAG.getBuildVector(MVT::v2i32, SL, {LoAnd, HiAnd}); | |||
3154 | return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Vec); | |||
3155 | } | |||
3156 | ||||
3157 | SDValue AMDGPUTargetLowering::performShlCombine(SDNode *N, | |||
3158 | DAGCombinerInfo &DCI) const { | |||
3159 | EVT VT = N->getValueType(0); | |||
3160 | ||||
3161 | ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1)); | |||
3162 | if (!RHS) | |||
3163 | return SDValue(); | |||
3164 | ||||
3165 | SDValue LHS = N->getOperand(0); | |||
3166 | unsigned RHSVal = RHS->getZExtValue(); | |||
3167 | if (!RHSVal) | |||
3168 | return LHS; | |||
3169 | ||||
3170 | SDLoc SL(N); | |||
3171 | SelectionDAG &DAG = DCI.DAG; | |||
3172 | ||||
3173 | switch (LHS->getOpcode()) { | |||
3174 | default: | |||
3175 | break; | |||
3176 | case ISD::ZERO_EXTEND: | |||
3177 | case ISD::SIGN_EXTEND: | |||
3178 | case ISD::ANY_EXTEND: { | |||
3179 | SDValue X = LHS->getOperand(0); | |||
3180 | ||||
3181 | if (VT == MVT::i32 && RHSVal == 16 && X.getValueType() == MVT::i16 && | |||
3182 | isOperationLegal(ISD::BUILD_VECTOR, MVT::v2i16)) { | |||
3183 | // Prefer build_vector as the canonical form if packed types are legal. | |||
3184 | // (shl ([asz]ext i16:x), 16 -> build_vector 0, x | |||
3185 | SDValue Vec = DAG.getBuildVector(MVT::v2i16, SL, | |||
3186 | { DAG.getConstant(0, SL, MVT::i16), LHS->getOperand(0) }); | |||
3187 | return DAG.getNode(ISD::BITCAST, SL, MVT::i32, Vec); | |||
3188 | } | |||
3189 | ||||
3190 | // shl (ext x) => zext (shl x), if shift does not overflow int | |||
3191 | if (VT != MVT::i64) | |||
3192 | break; | |||
3193 | KnownBits Known = DAG.computeKnownBits(X); | |||
3194 | unsigned LZ = Known.countMinLeadingZeros(); | |||
3195 | if (LZ < RHSVal) | |||
3196 | break; | |||
3197 | EVT XVT = X.getValueType(); | |||
3198 | SDValue Shl = DAG.getNode(ISD::SHL, SL, XVT, X, SDValue(RHS, 0)); | |||
3199 | return DAG.getZExtOrTrunc(Shl, SL, VT); | |||
3200 | } | |||
3201 | } | |||
3202 | ||||
3203 | if (VT != MVT::i64) | |||
3204 | return SDValue(); | |||
3205 | ||||
3206 | // i64 (shl x, C) -> (build_pair 0, (shl x, C -32)) | |||
3207 | ||||
3208 | // On some subtargets, 64-bit shift is a quarter rate instruction. In the | |||
3209 | // common case, splitting this into a move and a 32-bit shift is faster and | |||
3210 | // the same code size. | |||
3211 | if (RHSVal < 32) | |||
3212 | return SDValue(); | |||
3213 | ||||
3214 | SDValue ShiftAmt = DAG.getConstant(RHSVal - 32, SL, MVT::i32); | |||
3215 | ||||
3216 | SDValue Lo = DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, LHS); | |||
3217 | SDValue NewShift = DAG.getNode(ISD::SHL, SL, MVT::i32, Lo, ShiftAmt); | |||
3218 | ||||
3219 | const SDValue Zero = DAG.getConstant(0, SL, MVT::i32); | |||
3220 | ||||
3221 | SDValue Vec = DAG.getBuildVector(MVT::v2i32, SL, {Zero, NewShift}); | |||
3222 | return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Vec); | |||
3223 | } | |||
3224 | ||||
3225 | SDValue AMDGPUTargetLowering::performSraCombine(SDNode *N, | |||
3226 | DAGCombinerInfo &DCI) const { | |||
3227 | if (N->getValueType(0) != MVT::i64) | |||
3228 | return SDValue(); | |||
3229 | ||||
3230 | const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1)); | |||
3231 | if (!RHS) | |||
3232 | return SDValue(); | |||
3233 | ||||
3234 | SelectionDAG &DAG = DCI.DAG; | |||
3235 | SDLoc SL(N); | |||
3236 | unsigned RHSVal = RHS->getZExtValue(); | |||
3237 | ||||
3238 | // (sra i64:x, 32) -> build_pair x, (sra hi_32(x), 31) | |||
3239 | if (RHSVal == 32) { | |||
3240 | SDValue Hi = getHiHalf64(N->getOperand(0), DAG); | |||
3241 | SDValue NewShift = DAG.getNode(ISD::SRA, SL, MVT::i32, Hi, | |||
3242 | DAG.getConstant(31, SL, MVT::i32)); | |||
3243 | ||||
3244 | SDValue BuildVec = DAG.getBuildVector(MVT::v2i32, SL, {Hi, NewShift}); | |||
3245 | return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildVec); | |||
3246 | } | |||
3247 | ||||
3248 | // (sra i64:x, 63) -> build_pair (sra hi_32(x), 31), (sra hi_32(x), 31) | |||
3249 | if (RHSVal == 63) { | |||
3250 | SDValue Hi = getHiHalf64(N->getOperand(0), DAG); | |||
3251 | SDValue NewShift = DAG.getNode(ISD::SRA, SL, MVT::i32, Hi, | |||
3252 | DAG.getConstant(31, SL, MVT::i32)); | |||
3253 | SDValue BuildVec = DAG.getBuildVector(MVT::v2i32, SL, {NewShift, NewShift}); | |||
3254 | return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildVec); | |||
3255 | } | |||
3256 | ||||
3257 | return SDValue(); | |||
3258 | } | |||
3259 | ||||
3260 | SDValue AMDGPUTargetLowering::performSrlCombine(SDNode *N, | |||
3261 | DAGCombinerInfo &DCI) const { | |||
3262 | auto *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1)); | |||
3263 | if (!RHS) | |||
3264 | return SDValue(); | |||
3265 | ||||
3266 | EVT VT = N->getValueType(0); | |||
3267 | SDValue LHS = N->getOperand(0); | |||
3268 | unsigned ShiftAmt = RHS->getZExtValue(); | |||
3269 | SelectionDAG &DAG = DCI.DAG; | |||
3270 | SDLoc SL(N); | |||
3271 | ||||
3272 | // fold (srl (and x, c1 << c2), c2) -> (and (srl(x, c2), c1) | |||
3273 | // this improves the ability to match BFE patterns in isel. | |||
3274 | if (LHS.getOpcode() == ISD::AND) { | |||
3275 | if (auto *Mask = dyn_cast<ConstantSDNode>(LHS.getOperand(1))) { | |||
3276 | unsigned MaskIdx, MaskLen; | |||
3277 | if (Mask->getAPIntValue().isShiftedMask(MaskIdx, MaskLen) && | |||
3278 | MaskIdx == ShiftAmt) { | |||
3279 | return DAG.getNode( | |||
3280 | ISD::AND, SL, VT, | |||
3281 | DAG.getNode(ISD::SRL, SL, VT, LHS.getOperand(0), N->getOperand(1)), | |||
3282 | DAG.getNode(ISD::SRL, SL, VT, LHS.getOperand(1), N->getOperand(1))); | |||
3283 | } | |||
3284 | } | |||
3285 | } | |||
3286 | ||||
3287 | if (VT != MVT::i64) | |||
3288 | return SDValue(); | |||
3289 | ||||
3290 | if (ShiftAmt < 32) | |||
3291 | return SDValue(); | |||
3292 | ||||
3293 | // srl i64:x, C for C >= 32 | |||
3294 | // => | |||
3295 | // build_pair (srl hi_32(x), C - 32), 0 | |||
3296 | SDValue Zero = DAG.getConstant(0, SL, MVT::i32); | |||
3297 | ||||
3298 | SDValue Hi = getHiHalf64(LHS, DAG); | |||
3299 | ||||
3300 | SDValue NewConst = DAG.getConstant(ShiftAmt - 32, SL, MVT::i32); | |||
3301 | SDValue NewShift = DAG.getNode(ISD::SRL, SL, MVT::i32, Hi, NewConst); | |||
3302 | ||||
3303 | SDValue BuildPair = DAG.getBuildVector(MVT::v2i32, SL, {NewShift, Zero}); | |||
3304 | ||||
3305 | return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildPair); | |||
3306 | } | |||
3307 | ||||
3308 | SDValue AMDGPUTargetLowering::performTruncateCombine( | |||
3309 | SDNode *N, DAGCombinerInfo &DCI) const { | |||
3310 | SDLoc SL(N); | |||
3311 | SelectionDAG &DAG = DCI.DAG; | |||
3312 | EVT VT = N->getValueType(0); | |||
3313 | SDValue Src = N->getOperand(0); | |||
3314 | ||||
3315 | // vt1 (truncate (bitcast (build_vector vt0:x, ...))) -> vt1 (bitcast vt0:x) | |||
3316 | if (Src.getOpcode() == ISD::BITCAST && !VT.isVector()) { | |||
3317 | SDValue Vec = Src.getOperand(0); | |||
3318 | if (Vec.getOpcode() == ISD::BUILD_VECTOR) { | |||
3319 | SDValue Elt0 = Vec.getOperand(0); | |||
3320 | EVT EltVT = Elt0.getValueType(); | |||
3321 | if (VT.getFixedSizeInBits() <= EltVT.getFixedSizeInBits()) { | |||
3322 | if (EltVT.isFloatingPoint()) { | |||
3323 | Elt0 = DAG.getNode(ISD::BITCAST, SL, | |||
3324 | EltVT.changeTypeToInteger(), Elt0); | |||
3325 | } | |||
3326 | ||||
3327 | return DAG.getNode(ISD::TRUNCATE, SL, VT, Elt0); | |||
3328 | } | |||
3329 | } | |||
3330 | } | |||
3331 | ||||
3332 | // Equivalent of above for accessing the high element of a vector as an | |||
3333 | // integer operation. | |||
3334 | // trunc (srl (bitcast (build_vector x, y))), 16 -> trunc (bitcast y) | |||
3335 | if (Src.getOpcode() == ISD::SRL && !VT.isVector()) { | |||
3336 | if (auto K = isConstOrConstSplat(Src.getOperand(1))) { | |||
3337 | if (2 * K->getZExtValue() == Src.getValueType().getScalarSizeInBits()) { | |||
3338 | SDValue BV = stripBitcast(Src.getOperand(0)); | |||
3339 | if (BV.getOpcode() == ISD::BUILD_VECTOR && | |||
3340 | BV.getValueType().getVectorNumElements() == 2) { | |||
3341 | SDValue SrcElt = BV.getOperand(1); | |||
3342 | EVT SrcEltVT = SrcElt.getValueType(); | |||
3343 | if (SrcEltVT.isFloatingPoint()) { | |||
3344 | SrcElt = DAG.getNode(ISD::BITCAST, SL, | |||
3345 | SrcEltVT.changeTypeToInteger(), SrcElt); | |||
3346 | } | |||
3347 | ||||
3348 | return DAG.getNode(ISD::TRUNCATE, SL, VT, SrcElt); | |||
3349 | } | |||
3350 | } | |||
3351 | } | |||
3352 | } | |||
3353 | ||||
3354 | // Partially shrink 64-bit shifts to 32-bit if reduced to 16-bit. | |||
3355 | // | |||
3356 | // i16 (trunc (srl i64:x, K)), K <= 16 -> | |||
3357 | // i16 (trunc (srl (i32 (trunc x), K))) | |||
3358 | if (VT.getScalarSizeInBits() < 32) { | |||
3359 | EVT SrcVT = Src.getValueType(); | |||
3360 | if (SrcVT.getScalarSizeInBits() > 32 && | |||
3361 | (Src.getOpcode() == ISD::SRL || | |||
3362 | Src.getOpcode() == ISD::SRA || | |||
3363 | Src.getOpcode() == ISD::SHL)) { | |||
3364 | SDValue Amt = Src.getOperand(1); | |||
3365 | KnownBits Known = DAG.computeKnownBits(Amt); | |||
3366 | unsigned Size = VT.getScalarSizeInBits(); | |||
3367 | if ((Known.isConstant() && Known.getConstant().ule(Size)) || | |||
3368 | (Known.countMaxActiveBits() <= Log2_32(Size))) { | |||
3369 | EVT MidVT = VT.isVector() ? | |||
3370 | EVT::getVectorVT(*DAG.getContext(), MVT::i32, | |||
3371 | VT.getVectorNumElements()) : MVT::i32; | |||
3372 | ||||
3373 | EVT NewShiftVT = getShiftAmountTy(MidVT, DAG.getDataLayout()); | |||
3374 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SL, MidVT, | |||
3375 | Src.getOperand(0)); | |||
3376 | DCI.AddToWorklist(Trunc.getNode()); | |||
3377 | ||||
3378 | if (Amt.getValueType() != NewShiftVT) { | |||
3379 | Amt = DAG.getZExtOrTrunc(Amt, SL, NewShiftVT); | |||
3380 | DCI.AddToWorklist(Amt.getNode()); | |||
3381 | } | |||
3382 | ||||
3383 | SDValue ShrunkShift = DAG.getNode(Src.getOpcode(), SL, MidVT, | |||
3384 | Trunc, Amt); | |||
3385 | return DAG.getNode(ISD::TRUNCATE, SL, VT, ShrunkShift); | |||
3386 | } | |||
3387 | } | |||
3388 | } | |||
3389 | ||||
3390 | return SDValue(); | |||
3391 | } | |||
3392 | ||||
3393 | // We need to specifically handle i64 mul here to avoid unnecessary conversion | |||
3394 | // instructions. If we only match on the legalized i64 mul expansion, | |||
3395 | // SimplifyDemandedBits will be unable to remove them because there will be | |||
3396 | // multiple uses due to the separate mul + mulh[su]. | |||
3397 | static SDValue getMul24(SelectionDAG &DAG, const SDLoc &SL, | |||
3398 | SDValue N0, SDValue N1, unsigned Size, bool Signed) { | |||
3399 | if (Size <= 32) { | |||
3400 | unsigned MulOpc = Signed ? AMDGPUISD::MUL_I24 : AMDGPUISD::MUL_U24; | |||
3401 | return DAG.getNode(MulOpc, SL, MVT::i32, N0, N1); | |||
3402 | } | |||
3403 | ||||
3404 | unsigned MulLoOpc = Signed ? AMDGPUISD::MUL_I24 : AMDGPUISD::MUL_U24; | |||
3405 | unsigned MulHiOpc = Signed ? AMDGPUISD::MULHI_I24 : AMDGPUISD::MULHI_U24; | |||
3406 | ||||
3407 | SDValue MulLo = DAG.getNode(MulLoOpc, SL, MVT::i32, N0, N1); | |||
3408 | SDValue MulHi = DAG.getNode(MulHiOpc, SL, MVT::i32, N0, N1); | |||
3409 | ||||
3410 | return DAG.getNode(ISD::BUILD_PAIR, SL, MVT::i64, MulLo, MulHi); | |||
3411 | } | |||
3412 | ||||
3413 | SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N, | |||
3414 | DAGCombinerInfo &DCI) const { | |||
3415 | EVT VT = N->getValueType(0); | |||
3416 | ||||
3417 | // Don't generate 24-bit multiplies on values that are in SGPRs, since | |||
3418 | // we only have a 32-bit scalar multiply (avoid values being moved to VGPRs | |||
3419 | // unnecessarily). isDivergent() is used as an approximation of whether the | |||
3420 | // value is in an SGPR. | |||
3421 | if (!N->isDivergent()) | |||
3422 | return SDValue(); | |||
3423 | ||||
3424 | unsigned Size = VT.getSizeInBits(); | |||
3425 | if (VT.isVector() || Size > 64) | |||
3426 | return SDValue(); | |||
3427 | ||||
3428 | // There are i16 integer mul/mad. | |||
3429 | if (Subtarget->has16BitInsts() && VT.getScalarType().bitsLE(MVT::i16)) | |||
3430 | return SDValue(); | |||
3431 | ||||
3432 | SelectionDAG &DAG = DCI.DAG; | |||
3433 | SDLoc DL(N); | |||
3434 | ||||
3435 | SDValue N0 = N->getOperand(0); | |||
3436 | SDValue N1 = N->getOperand(1); | |||
3437 | ||||
3438 | // SimplifyDemandedBits has the annoying habit of turning useful zero_extends | |||
3439 | // in the source into any_extends if the result of the mul is truncated. Since | |||
3440 | // we can assume the high bits are whatever we want, use the underlying value | |||
3441 | // to avoid the unknown high bits from interfering. | |||
3442 | if (N0.getOpcode() == ISD::ANY_EXTEND) | |||
3443 | N0 = N0.getOperand(0); | |||
3444 | ||||
3445 | if (N1.getOpcode() == ISD::ANY_EXTEND) | |||
3446 | N1 = N1.getOperand(0); | |||
3447 | ||||
3448 | SDValue Mul; | |||
3449 | ||||
3450 | if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) { | |||
3451 | N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32); | |||
3452 | N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32); | |||
3453 | Mul = getMul24(DAG, DL, N0, N1, Size, false); | |||
3454 | } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) { | |||
3455 | N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32); | |||
3456 | N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32); | |||
3457 | Mul = getMul24(DAG, DL, N0, N1, Size, true); | |||
3458 | } else { | |||
3459 | return SDValue(); | |||
3460 | } | |||
3461 | ||||
3462 | // We need to use sext even for MUL_U24, because MUL_U24 is used | |||
3463 | // for signed multiply of 8 and 16-bit types. | |||
3464 | return DAG.getSExtOrTrunc(Mul, DL, VT); | |||
3465 | } | |||
3466 | ||||
3467 | SDValue | |||
3468 | AMDGPUTargetLowering::performMulLoHiCombine(SDNode *N, | |||
3469 | DAGCombinerInfo &DCI) const { | |||
3470 | if (N->getValueType(0) != MVT::i32) | |||
3471 | return SDValue(); | |||
3472 | ||||
3473 | SelectionDAG &DAG = DCI.DAG; | |||
3474 | SDLoc DL(N); | |||
3475 | ||||
3476 | SDValue N0 = N->getOperand(0); | |||
3477 | SDValue N1 = N->getOperand(1); | |||
3478 | ||||
3479 | // SimplifyDemandedBits has the annoying habit of turning useful zero_extends | |||
3480 | // in the source into any_extends if the result of the mul is truncated. Since | |||
3481 | // we can assume the high bits are whatever we want, use the underlying value | |||
3482 | // to avoid the unknown high bits from interfering. | |||
3483 | if (N0.getOpcode() == ISD::ANY_EXTEND) | |||
3484 | N0 = N0.getOperand(0); | |||
3485 | if (N1.getOpcode() == ISD::ANY_EXTEND) | |||
3486 | N1 = N1.getOperand(0); | |||
3487 | ||||
3488 | // Try to use two fast 24-bit multiplies (one for each half of the result) | |||
3489 | // instead of one slow extending multiply. | |||
3490 | unsigned LoOpcode, HiOpcode; | |||
3491 | if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) { | |||
3492 | N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32); | |||
3493 | N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32); | |||
3494 | LoOpcode = AMDGPUISD::MUL_U24; | |||
3495 | HiOpcode = AMDGPUISD::MULHI_U24; | |||
3496 | } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) { | |||
3497 | N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32); | |||
3498 | N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32); | |||
3499 | LoOpcode = AMDGPUISD::MUL_I24; | |||
3500 | HiOpcode = AMDGPUISD::MULHI_I24; | |||
3501 | } else { | |||
3502 | return SDValue(); | |||
3503 | } | |||
3504 | ||||
3505 | SDValue Lo = DAG.getNode(LoOpcode, DL, MVT::i32, N0, N1); | |||
3506 | SDValue Hi = DAG.getNode(HiOpcode, DL, MVT::i32, N0, N1); | |||
3507 | DCI.CombineTo(N, Lo, Hi); | |||
3508 | return SDValue(N, 0); | |||
3509 | } | |||
3510 | ||||
3511 | SDValue AMDGPUTargetLowering::performMulhsCombine(SDNode *N, | |||
3512 | DAGCombinerInfo &DCI) const { | |||
3513 | EVT VT = N->getValueType(0); | |||
3514 | ||||
3515 | if (!Subtarget->hasMulI24() || VT.isVector()) | |||
3516 | return SDValue(); | |||
3517 | ||||
3518 | // Don't generate 24-bit multiplies on values that are in SGPRs, since | |||
3519 | // we only have a 32-bit scalar multiply (avoid values being moved to VGPRs | |||
3520 | // unnecessarily). isDivergent() is used as an approximation of whether the | |||
3521 | // value is in an SGPR. | |||
3522 | // This doesn't apply if no s_mul_hi is available (since we'll end up with a | |||
3523 | // valu op anyway) | |||
3524 | if (Subtarget->hasSMulHi() && !N->isDivergent()) | |||
3525 | return SDValue(); | |||
3526 | ||||
3527 | SelectionDAG &DAG = DCI.DAG; | |||
3528 | SDLoc DL(N); | |||
3529 | ||||
3530 | SDValue N0 = N->getOperand(0); | |||
3531 | SDValue N1 = N->getOperand(1); | |||
3532 | ||||
3533 | if (!isI24(N0, DAG) || !isI24(N1, DAG)) | |||
3534 | return SDValue(); | |||
3535 | ||||
3536 | N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32); | |||
3537 | N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32); | |||
3538 | ||||
3539 | SDValue Mulhi = DAG.getNode(AMDGPUISD::MULHI_I24, DL, MVT::i32, N0, N1); | |||
3540 | DCI.AddToWorklist(Mulhi.getNode()); | |||
3541 | return DAG.getSExtOrTrunc(Mulhi, DL, VT); | |||
3542 | } | |||
3543 | ||||
3544 | SDValue AMDGPUTargetLowering::performMulhuCombine(SDNode *N, | |||
3545 | DAGCombinerInfo &DCI) const { | |||
3546 | EVT VT = N->getValueType(0); | |||
3547 | ||||
3548 | if (!Subtarget->hasMulU24() || VT.isVector() || VT.getSizeInBits() > 32) | |||
3549 | return SDValue(); | |||
3550 | ||||
3551 | // Don't generate 24-bit multiplies on values that are in SGPRs, since | |||
3552 | // we only have a 32-bit scalar multiply (avoid values being moved to VGPRs | |||
3553 | // unnecessarily). isDivergent() is used as an approximation of whether the | |||
3554 | // value is in an SGPR. | |||
3555 | // This doesn't apply if no s_mul_hi is available (since we'll end up with a | |||
3556 | // valu op anyway) | |||
3557 | if (Subtarget->hasSMulHi() && !N->isDivergent()) | |||
3558 | return SDValue(); | |||
3559 | ||||
3560 | SelectionDAG &DAG = DCI.DAG; | |||
3561 | SDLoc DL(N); | |||
3562 | ||||
3563 | SDValue N0 = N->getOperand(0); | |||
3564 | SDValue N1 = N->getOperand(1); | |||
3565 | ||||
3566 | if (!isU24(N0, DAG) || !isU24(N1, DAG)) | |||
3567 | return SDValue(); | |||
3568 | ||||
3569 | N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32); | |||
3570 | N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32); | |||
3571 | ||||
3572 | SDValue Mulhi = DAG.getNode(AMDGPUISD::MULHI_U24, DL, MVT::i32, N0, N1); | |||
3573 | DCI.AddToWorklist(Mulhi.getNode()); | |||
3574 | return DAG.getZExtOrTrunc(Mulhi, DL, VT); | |||
3575 | } | |||
3576 | ||||
3577 | static bool isNegativeOne(SDValue Val) { | |||
3578 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) | |||
3579 | return C->isAllOnes(); | |||
3580 | return false; | |||
3581 | } | |||
3582 | ||||
3583 | SDValue AMDGPUTargetLowering::getFFBX_U32(SelectionDAG &DAG, | |||
3584 | SDValue Op, | |||
3585 | const SDLoc &DL, | |||
3586 | unsigned Opc) const { | |||
3587 | EVT VT = Op.getValueType(); | |||
3588 | EVT LegalVT = getTypeToTransformTo(*DAG.getContext(), VT); | |||
3589 | if (LegalVT != MVT::i32 && (Subtarget->has16BitInsts() && | |||
3590 | LegalVT != MVT::i16)) | |||
3591 | return SDValue(); | |||
3592 | ||||
3593 | if (VT != MVT::i32) | |||
3594 | Op = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, Op); | |||
3595 | ||||
3596 | SDValue FFBX = DAG.getNode(Opc, DL, MVT::i32, Op); | |||
3597 | if (VT != MVT::i32) | |||
3598 | FFBX = DAG.getNode(ISD::TRUNCATE, DL, VT, FFBX); | |||
3599 | ||||
3600 | return FFBX; | |||
3601 | } | |||
3602 | ||||
3603 | // The native instructions return -1 on 0 input. Optimize out a select that | |||
3604 | // produces -1 on 0. | |||
3605 | // | |||
3606 | // TODO: If zero is not undef, we could also do this if the output is compared | |||
3607 | // against the bitwidth. | |||
3608 | // | |||
3609 | // TODO: Should probably combine against FFBH_U32 instead of ctlz directly. | |||
3610 | SDValue AMDGPUTargetLowering::performCtlz_CttzCombine(const SDLoc &SL, SDValue Cond, | |||
3611 | SDValue LHS, SDValue RHS, | |||
3612 | DAGCombinerInfo &DCI) const { | |||
3613 | ConstantSDNode *CmpRhs = dyn_cast<ConstantSDNode>(Cond.getOperand(1)); | |||
3614 | if (!CmpRhs || !CmpRhs->isZero()) | |||
3615 | return SDValue(); | |||
3616 | ||||
3617 | SelectionDAG &DAG = DCI.DAG; | |||
3618 | ISD::CondCode CCOpcode = cast<CondCodeSDNode>(Cond.getOperand(2))->get(); | |||
3619 | SDValue CmpLHS = Cond.getOperand(0); | |||
3620 | ||||
3621 | // select (setcc x, 0, eq), -1, (ctlz_zero_undef x) -> ffbh_u32 x | |||
3622 | // select (setcc x, 0, eq), -1, (cttz_zero_undef x) -> ffbl_u32 x | |||
3623 | if (CCOpcode == ISD::SETEQ && | |||
3624 | (isCtlzOpc(RHS.getOpcode()) || isCttzOpc(RHS.getOpcode())) && | |||
3625 | RHS.getOperand(0) == CmpLHS && isNegativeOne(LHS)) { | |||
3626 | unsigned Opc = | |||
3627 | isCttzOpc(RHS.getOpcode()) ? AMDGPUISD::FFBL_B32 : AMDGPUISD::FFBH_U32; | |||
3628 | return getFFBX_U32(DAG, CmpLHS, SL, Opc); | |||
3629 | } | |||
3630 | ||||
3631 | // select (setcc x, 0, ne), (ctlz_zero_undef x), -1 -> ffbh_u32 x | |||
3632 | // select (setcc x, 0, ne), (cttz_zero_undef x), -1 -> ffbl_u32 x | |||
3633 | if (CCOpcode == ISD::SETNE && | |||
3634 | (isCtlzOpc(LHS.getOpcode()) || isCttzOpc(LHS.getOpcode())) && | |||
3635 | LHS.getOperand(0) == CmpLHS && isNegativeOne(RHS)) { | |||
3636 | unsigned Opc = | |||
3637 | isCttzOpc(LHS.getOpcode()) ? AMDGPUISD::FFBL_B32 : AMDGPUISD::FFBH_U32; | |||
3638 | ||||
3639 | return getFFBX_U32(DAG, CmpLHS, SL, Opc); | |||
3640 | } | |||
3641 | ||||
3642 | return SDValue(); | |||
3643 | } | |||
3644 | ||||
3645 | static SDValue distributeOpThroughSelect(TargetLowering::DAGCombinerInfo &DCI, | |||
3646 | unsigned Op, | |||
3647 | const SDLoc &SL, | |||
3648 | SDValue Cond, | |||
3649 | SDValue N1, | |||
3650 | SDValue N2) { | |||
3651 | SelectionDAG &DAG = DCI.DAG; | |||
3652 | EVT VT = N1.getValueType(); | |||
3653 | ||||
3654 | SDValue NewSelect = DAG.getNode(ISD::SELECT, SL, VT, Cond, | |||
3655 | N1.getOperand(0), N2.getOperand(0)); | |||
3656 | DCI.AddToWorklist(NewSelect.getNode()); | |||
3657 | return DAG.getNode(Op, SL, VT, NewSelect); | |||
3658 | } | |||
3659 | ||||
3660 | // Pull a free FP operation out of a select so it may fold into uses. | |||
3661 | // | |||
3662 | // select c, (fneg x), (fneg y) -> fneg (select c, x, y) | |||
3663 | // select c, (fneg x), k -> fneg (select c, x, (fneg k)) | |||
3664 | // | |||
3665 | // select c, (fabs x), (fabs y) -> fabs (select c, x, y) | |||
3666 | // select c, (fabs x), +k -> fabs (select c, x, k) | |||
3667 | static SDValue foldFreeOpFromSelect(TargetLowering::DAGCombinerInfo &DCI, | |||
3668 | SDValue N) { | |||
3669 | SelectionDAG &DAG = DCI.DAG; | |||
3670 | SDValue Cond = N.getOperand(0); | |||
3671 | SDValue LHS = N.getOperand(1); | |||
3672 | SDValue RHS = N.getOperand(2); | |||
3673 | ||||
3674 | EVT VT = N.getValueType(); | |||
3675 | if ((LHS.getOpcode() == ISD::FABS && RHS.getOpcode() == ISD::FABS) || | |||
3676 | (LHS.getOpcode() == ISD::FNEG && RHS.getOpcode() == ISD::FNEG)) { | |||
3677 | return distributeOpThroughSelect(DCI, LHS.getOpcode(), | |||
3678 | SDLoc(N), Cond, LHS, RHS); | |||
3679 | } | |||
3680 | ||||
3681 | bool Inv = false; | |||
3682 | if (RHS.getOpcode() == ISD::FABS || RHS.getOpcode() == ISD::FNEG) { | |||
3683 | std::swap(LHS, RHS); | |||
3684 | Inv = true; | |||
3685 | } | |||
3686 | ||||
3687 | // TODO: Support vector constants. | |||
3688 | ConstantFPSDNode *CRHS = dyn_cast<ConstantFPSDNode>(RHS); | |||
3689 | if ((LHS.getOpcode() == ISD::FNEG || LHS.getOpcode() == ISD::FABS) && CRHS) { | |||
3690 | SDLoc SL(N); | |||
3691 | // If one side is an fneg/fabs and the other is a constant, we can push the | |||
3692 | // fneg/fabs down. If it's an fabs, the constant needs to be non-negative. | |||
3693 | SDValue NewLHS = LHS.getOperand(0); | |||
3694 | SDValue NewRHS = RHS; | |||
3695 | ||||
3696 | // Careful: if the neg can be folded up, don't try to pull it back down. | |||
3697 | bool ShouldFoldNeg = true; | |||
3698 | ||||
3699 | if (NewLHS.hasOneUse()) { | |||
3700 | unsigned Opc = NewLHS.getOpcode(); | |||
3701 | if (LHS.getOpcode() == ISD::FNEG && fnegFoldsIntoOp(Opc)) | |||
3702 | ShouldFoldNeg = false; | |||
3703 | if (LHS.getOpcode() == ISD::FABS && Opc == ISD::FMUL) | |||
3704 | ShouldFoldNeg = false; | |||
3705 | } | |||
3706 | ||||
3707 | if (ShouldFoldNeg) { | |||
3708 | if (LHS.getOpcode() == ISD::FNEG) | |||
3709 | NewRHS = DAG.getNode(ISD::FNEG, SL, VT, RHS); | |||
3710 | else if (CRHS->isNegative()) | |||
3711 | return SDValue(); | |||
3712 | ||||
3713 | if (Inv) | |||
3714 | std::swap(NewLHS, NewRHS); | |||
3715 | ||||
3716 | SDValue NewSelect = DAG.getNode(ISD::SELECT, SL, VT, | |||
3717 | Cond, NewLHS, NewRHS); | |||
3718 | DCI.AddToWorklist(NewSelect.getNode()); | |||
3719 | return DAG.getNode(LHS.getOpcode(), SL, VT, NewSelect); | |||
3720 | } | |||
3721 | } | |||
3722 | ||||
3723 | return SDValue(); | |||
3724 | } | |||
3725 | ||||
3726 | ||||
3727 | SDValue AMDGPUTargetLowering::performSelectCombine(SDNode *N, | |||
3728 | DAGCombinerInfo &DCI) const { | |||
3729 | if (SDValue Folded = foldFreeOpFromSelect(DCI, SDValue(N, 0))) | |||
3730 | return Folded; | |||
3731 | ||||
3732 | SDValue Cond = N->getOperand(0); | |||
3733 | if (Cond.getOpcode() != ISD::SETCC) | |||
3734 | return SDValue(); | |||
3735 | ||||
3736 | EVT VT = N->getValueType(0); | |||
3737 | SDValue LHS = Cond.getOperand(0); | |||
3738 | SDValue RHS = Cond.getOperand(1); | |||
3739 | SDValue CC = Cond.getOperand(2); | |||
3740 | ||||
3741 | SDValue True = N->getOperand(1); | |||
3742 | SDValue False = N->getOperand(2); | |||
3743 | ||||
3744 | if (Cond.hasOneUse()) { // TODO: Look for multiple select uses. | |||
3745 | SelectionDAG &DAG = DCI.DAG; | |||
3746 | if (DAG.isConstantValueOfAnyType(True) && | |||
3747 | !DAG.isConstantValueOfAnyType(False)) { | |||
3748 | // Swap cmp + select pair to move constant to false input. | |||
3749 | // This will allow using VOPC cndmasks more often. | |||
3750 | // select (setcc x, y), k, x -> select (setccinv x, y), x, k | |||
3751 | ||||
3752 | SDLoc SL(N); | |||
3753 | ISD::CondCode NewCC = | |||
3754 | getSetCCInverse(cast<CondCodeSDNode>(CC)->get(), LHS.getValueType()); | |||
3755 | ||||
3756 | SDValue NewCond = DAG.getSetCC(SL, Cond.getValueType(), LHS, RHS, NewCC); | |||
3757 | return DAG.getNode(ISD::SELECT, SL, VT, NewCond, False, True); | |||
3758 | } | |||
3759 | ||||
3760 | if (VT == MVT::f32 && Subtarget->hasFminFmaxLegacy()) { | |||
3761 | SDValue MinMax | |||
3762 | = combineFMinMaxLegacy(SDLoc(N), VT, LHS, RHS, True, False, CC, DCI); | |||
3763 | // Revisit this node so we can catch min3/max3/med3 patterns. | |||
3764 | //DCI.AddToWorklist(MinMax.getNode()); | |||
3765 | return MinMax; | |||
3766 | } | |||
3767 | } | |||
3768 | ||||
3769 | // There's no reason to not do this if the condition has other uses. | |||
3770 | return performCtlz_CttzCombine(SDLoc(N), Cond, True, False, DCI); | |||
3771 | } | |||
3772 | ||||
3773 | static bool isInv2Pi(const APFloat &APF) { | |||
3774 | static const APFloat KF16(APFloat::IEEEhalf(), APInt(16, 0x3118)); | |||
3775 | static const APFloat KF32(APFloat::IEEEsingle(), APInt(32, 0x3e22f983)); | |||
3776 | static const APFloat KF64(APFloat::IEEEdouble(), APInt(64, 0x3fc45f306dc9c882)); | |||
3777 | ||||
3778 | return APF.bitwiseIsEqual(KF16) || | |||
3779 | APF.bitwiseIsEqual(KF32) || | |||
3780 | APF.bitwiseIsEqual(KF64); | |||
3781 | } | |||
3782 | ||||
3783 | // 0 and 1.0 / (0.5 * pi) do not have inline immmediates, so there is an | |||
3784 | // additional cost to negate them. | |||
3785 | bool AMDGPUTargetLowering::isConstantCostlierToNegate(SDValue N) const { | |||
3786 | if (const ConstantFPSDNode *C = isConstOrConstSplatFP(N)) { | |||
3787 | if (C->isZero() && !C->isNegative()) | |||
3788 | return true; | |||
3789 | ||||
3790 | if (Subtarget->hasInv2PiInlineImm() && isInv2Pi(C->getValueAPF())) | |||
3791 | return true; | |||
3792 | } | |||
3793 | ||||
3794 | return false; | |||
3795 | } | |||
3796 | ||||
3797 | static unsigned inverseMinMax(unsigned Opc) { | |||
3798 | switch (Opc) { | |||
3799 | case ISD::FMAXNUM: | |||
3800 | return ISD::FMINNUM; | |||
3801 | case ISD::FMINNUM: | |||
3802 | return ISD::FMAXNUM; | |||
3803 | case ISD::FMAXNUM_IEEE: | |||
3804 | return ISD::FMINNUM_IEEE; | |||
3805 | case ISD::FMINNUM_IEEE: | |||
3806 | return ISD::FMAXNUM_IEEE; | |||
3807 | case AMDGPUISD::FMAX_LEGACY: | |||
3808 | return AMDGPUISD::FMIN_LEGACY; | |||
3809 | case AMDGPUISD::FMIN_LEGACY: | |||
3810 | return AMDGPUISD::FMAX_LEGACY; | |||
3811 | default: | |||
3812 | llvm_unreachable("invalid min/max opcode")::llvm::llvm_unreachable_internal("invalid min/max opcode", "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp" , 3812); | |||
3813 | } | |||
3814 | } | |||
3815 | ||||
3816 | SDValue AMDGPUTargetLowering::performFNegCombine(SDNode *N, | |||
3817 | DAGCombinerInfo &DCI) const { | |||
3818 | SelectionDAG &DAG = DCI.DAG; | |||
3819 | SDValue N0 = N->getOperand(0); | |||
3820 | EVT VT = N->getValueType(0); | |||
3821 | ||||
3822 | unsigned Opc = N0.getOpcode(); | |||
3823 | ||||
3824 | // If the input has multiple uses and we can either fold the negate down, or | |||
3825 | // the other uses cannot, give up. This both prevents unprofitable | |||
3826 | // transformations and infinite loops: we won't repeatedly try to fold around | |||
3827 | // a negate that has no 'good' form. | |||
3828 | if (N0.hasOneUse()) { | |||
3829 | // This may be able to fold into the source, but at a code size cost. Don't | |||
3830 | // fold if the fold into the user is free. | |||
3831 | if (allUsesHaveSourceMods(N, 0)) | |||
3832 | return SDValue(); | |||
3833 | } else { | |||
3834 | if (fnegFoldsIntoOp(Opc) && | |||
3835 | (allUsesHaveSourceMods(N) || !allUsesHaveSourceMods(N0.getNode()))) | |||
3836 | return SDValue(); | |||
3837 | } | |||
3838 | ||||
3839 | SDLoc SL(N); | |||
3840 | switch (Opc) { | |||
3841 | case ISD::FADD: { | |||
3842 | if (!mayIgnoreSignedZero(N0)) | |||
3843 | return SDValue(); | |||
3844 | ||||
3845 | // (fneg (fadd x, y)) -> (fadd (fneg x), (fneg y)) | |||
3846 | SDValue LHS = N0.getOperand(0); | |||
3847 | SDValue RHS = N0.getOperand(1); | |||
3848 | ||||
3849 | if (LHS.getOpcode() != ISD::FNEG) | |||
3850 | LHS = DAG.getNode(ISD::FNEG, SL, VT, LHS); | |||
3851 | else | |||
3852 | LHS = LHS.getOperand(0); | |||
3853 | ||||
3854 | if (RHS.getOpcode() != ISD::FNEG) | |||
3855 | RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS); | |||
3856 | else | |||
3857 | RHS = RHS.getOperand(0); | |||
3858 | ||||
3859 | SDValue Res = DAG.getNode(ISD::FADD, SL, VT, LHS, RHS, N0->getFlags()); | |||
3860 | if (Res.getOpcode() != ISD::FADD) | |||
3861 | return SDValue(); // Op got folded away. | |||
3862 | if (!N0.hasOneUse()) | |||
3863 | DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res)); | |||
3864 | return Res; | |||
3865 | } | |||
3866 | case ISD::FMUL: | |||
3867 | case AMDGPUISD::FMUL_LEGACY: { | |||
3868 | // (fneg (fmul x, y)) -> (fmul x, (fneg y)) | |||
3869 | // (fneg (fmul_legacy x, y)) -> (fmul_legacy x, (fneg y)) | |||
3870 | SDValue LHS = N0.getOperand(0); | |||
3871 | SDValue RHS = N0.getOperand(1); | |||
3872 | ||||
3873 | if (LHS.getOpcode() == ISD::FNEG) | |||
3874 | LHS = LHS.getOperand(0); | |||
3875 | else if (RHS.getOpcode() == ISD::FNEG) | |||
3876 | RHS = RHS.getOperand(0); | |||
3877 | else | |||
3878 | RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS); | |||
3879 | ||||
3880 | SDValue Res = DAG.getNode(Opc, SL, VT, LHS, RHS, N0->getFlags()); | |||
3881 | if (Res.getOpcode() != Opc) | |||
3882 | return SDValue(); // Op got folded away. | |||
3883 | if (!N0.hasOneUse()) | |||
3884 | DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res)); | |||
3885 | return Res; | |||
3886 | } | |||
3887 | case ISD::FMA: | |||
3888 | case ISD::FMAD: { | |||
3889 | // TODO: handle llvm.amdgcn.fma.legacy | |||
3890 | if (!mayIgnoreSignedZero(N0)) | |||
3891 | return SDValue(); | |||
3892 | ||||
3893 | // (fneg (fma x, y, z)) -> (fma x, (fneg y), (fneg z)) | |||
3894 | SDValue LHS = N0.getOperand(0); | |||
3895 | SDValue MHS = N0.getOperand(1); | |||
3896 | SDValue RHS = N0.getOperand(2); | |||
3897 | ||||
3898 | if (LHS.getOpcode() == ISD::FNEG) | |||
3899 | LHS = LHS.getOperand(0); | |||
3900 | else if (MHS.getOpcode() == ISD::FNEG) | |||
3901 | MHS = MHS.getOperand(0); | |||
3902 | else | |||
3903 | MHS = DAG.getNode(ISD::FNEG, SL, VT, MHS); | |||
3904 | ||||
3905 | if (RHS.getOpcode() != ISD::FNEG) | |||
3906 | RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS); | |||
3907 | else | |||
3908 | RHS = RHS.getOperand(0); | |||
3909 | ||||
3910 | SDValue Res = DAG.getNode(Opc, SL, VT, LHS, MHS, RHS); | |||
3911 | if (Res.getOpcode() != Opc) | |||
3912 | return SDValue(); // Op got folded away. | |||
3913 | if (!N0.hasOneUse()) | |||
3914 | DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res)); | |||
3915 | return Res; | |||
3916 | } | |||
3917 | case ISD::FMAXNUM: | |||
3918 | case ISD::FMINNUM: | |||
3919 | case ISD::FMAXNUM_IEEE: | |||
3920 | case ISD::FMINNUM_IEEE: | |||
3921 | case AMDGPUISD::FMAX_LEGACY: | |||
3922 | case AMDGPUISD::FMIN_LEGACY: { | |||
3923 | // fneg (fmaxnum x, y) -> fminnum (fneg x), (fneg y) | |||
3924 | // fneg (fminnum x, y) -> fmaxnum (fneg x), (fneg y) | |||
3925 | // fneg (fmax_legacy x, y) -> fmin_legacy (fneg x), (fneg y) | |||
3926 | // fneg (fmin_legacy x, y) -> fmax_legacy (fneg x), (fneg y) | |||
3927 | ||||
3928 | SDValue LHS = N0.getOperand(0); | |||
3929 | SDValue RHS = N0.getOperand(1); | |||
3930 | ||||
3931 | // 0 doesn't have a negated inline immediate. | |||
3932 | // TODO: This constant check should be generalized to other operations. | |||
3933 | if (isConstantCostlierToNegate(RHS)) | |||
3934 | return SDValue(); | |||
3935 | ||||
3936 | SDValue NegLHS = DAG.getNode(ISD::FNEG, SL, VT, LHS); | |||
3937 | SDValue NegRHS = DAG.getNode(ISD::FNEG, SL, VT, RHS); | |||
3938 | unsigned Opposite = inverseMinMax(Opc); | |||
3939 | ||||
3940 | SDValue Res = DAG.getNode(Opposite, SL, VT, NegLHS, NegRHS, N0->getFlags()); | |||
3941 | if (Res.getOpcode() != Opposite) | |||
3942 | return SDValue(); // Op got folded away. | |||
3943 | if (!N0.hasOneUse()) | |||
3944 | DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res)); | |||
3945 | return Res; | |||
3946 | } | |||
3947 | case AMDGPUISD::FMED3: { | |||
3948 | SDValue Ops[3]; | |||
3949 | for (unsigned I = 0; I < 3; ++I) | |||
3950 | Ops[I] = DAG.getNode(ISD::FNEG, SL, VT, N0->getOperand(I), N0->getFlags()); | |||
3951 | ||||
3952 | SDValue Res = DAG.getNode(AMDGPUISD::FMED3, SL, VT, Ops, N0->getFlags()); | |||
3953 | if (Res.getOpcode() != AMDGPUISD::FMED3) | |||
3954 | return SDValue(); // Op got folded away. | |||
3955 | ||||
3956 | if (!N0.hasOneUse()) { | |||
3957 | SDValue Neg = DAG.getNode(ISD::FNEG, SL, VT, Res); | |||
3958 | DAG.ReplaceAllUsesWith(N0, Neg); | |||
3959 | ||||
3960 | for (SDNode *U : Neg->uses()) | |||
3961 | DCI.AddToWorklist(U); | |||
3962 | } | |||
3963 | ||||
3964 | return Res; | |||
3965 | } | |||
3966 | case ISD::FP_EXTEND: | |||
3967 | case ISD::FTRUNC: | |||
3968 | case ISD::FRINT: | |||
3969 | case ISD::FNEARBYINT: // XXX - Should fround be handled? | |||
3970 | case ISD::FSIN: | |||
3971 | case ISD::FCANONICALIZE: | |||
3972 | case AMDGPUISD::RCP: | |||
3973 | case AMDGPUISD::RCP_LEGACY: | |||
3974 | case AMDGPUISD::RCP_IFLAG: | |||
3975 | case AMDGPUISD::SIN_HW: { | |||
3976 | SDValue CvtSrc = N0.getOperand(0); | |||
3977 | if (CvtSrc.getOpcode() == ISD::FNEG) { | |||
3978 | // (fneg (fp_extend (fneg x))) -> (fp_extend x) | |||
3979 | // (fneg (rcp (fneg x))) -> (rcp x) | |||
3980 | return DAG.getNode(Opc, SL, VT, CvtSrc.getOperand(0)); | |||
3981 | } | |||
3982 | ||||
3983 | if (!N0.hasOneUse()) | |||
3984 | return SDValue(); | |||
3985 | ||||
3986 | // (fneg (fp_extend x)) -> (fp_extend (fneg x)) | |||
3987 | // (fneg (rcp x)) -> (rcp (fneg x)) | |||
3988 | SDValue Neg = DAG.getNode(ISD::FNEG, SL, CvtSrc.getValueType(), CvtSrc); | |||
3989 | return DAG.getNode(Opc, SL, VT, Neg, N0->getFlags()); | |||
3990 | } | |||
3991 | case ISD::FP_ROUND: { | |||
3992 | SDValue CvtSrc = N0.getOperand(0); | |||
3993 | ||||
3994 | if (CvtSrc.getOpcode() == ISD::FNEG) { | |||
3995 | // (fneg (fp_round (fneg x))) -> (fp_round x) | |||
3996 | return DAG.getNode(ISD::FP_ROUND, SL, VT, | |||
3997 | CvtSrc.getOperand(0), N0.getOperand(1)); | |||
3998 | } | |||
3999 | ||||
4000 | if (!N0.hasOneUse()) | |||
4001 | return SDValue(); | |||
4002 | ||||
4003 | // (fneg (fp_round x)) -> (fp_round (fneg x)) | |||
4004 | SDValue Neg = DAG.getNode(ISD::FNEG, SL, CvtSrc.getValueType(), CvtSrc); | |||
4005 | return DAG.getNode(ISD::FP_ROUND, SL, VT, Neg, N0.getOperand(1)); | |||
4006 | } | |||
4007 | case ISD::FP16_TO_FP: { | |||
4008 | // v_cvt_f32_f16 supports source modifiers on pre-VI targets without legal | |||
4009 | // f16, but legalization of f16 fneg ends up pulling it out of the source. | |||
4010 | // Put the fneg back as a legal source operation that can be matched later. | |||
4011 | SDLoc SL(N); | |||
4012 | ||||
4013 | SDValue Src = N0.getOperand(0); | |||
4014 | EVT SrcVT = Src.getValueType(); | |||
4015 | ||||
4016 | // fneg (fp16_to_fp x) -> fp16_to_fp (xor x, 0x8000) | |||
4017 | SDValue IntFNeg = DAG.getNode(ISD::XOR, SL, SrcVT, Src, | |||
4018 | DAG.getConstant(0x8000, SL, SrcVT)); | |||
4019 | return DAG.getNode(ISD::FP16_TO_FP, SL, N->getValueType(0), IntFNeg); | |||
4020 | } | |||
4021 | default: | |||
4022 | return SDValue(); | |||
4023 | } | |||
4024 | } | |||
4025 | ||||
4026 | SDValue AMDGPUTargetLowering::performFAbsCombine(SDNode *N, | |||
4027 | DAGCombinerInfo &DCI) const { | |||
4028 | SelectionDAG &DAG = DCI.DAG; | |||
4029 | SDValue N0 = N->getOperand(0); | |||
4030 | ||||
4031 | if (!N0.hasOneUse()) | |||
4032 | return SDValue(); | |||
4033 | ||||
4034 | switch (N0.getOpcode()) { | |||
4035 | case ISD::FP16_TO_FP: { | |||
4036 | assert(!Subtarget->has16BitInsts() && "should only see if f16 is illegal")(static_cast <bool> (!Subtarget->has16BitInsts() && "should only see if f16 is illegal") ? void (0) : __assert_fail ("!Subtarget->has16BitInsts() && \"should only see if f16 is illegal\"" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 4036, __extension__ __PRETTY_FUNCTION__)); | |||
4037 | SDLoc SL(N); | |||
4038 | SDValue Src = N0.getOperand(0); | |||
4039 | EVT SrcVT = Src.getValueType(); | |||
4040 | ||||
4041 | // fabs (fp16_to_fp x) -> fp16_to_fp (and x, 0x7fff) | |||
4042 | SDValue IntFAbs = DAG.getNode(ISD::AND, SL, SrcVT, Src, | |||
4043 | DAG.getConstant(0x7fff, SL, SrcVT)); | |||
4044 | return DAG.getNode(ISD::FP16_TO_FP, SL, N->getValueType(0), IntFAbs); | |||
4045 | } | |||
4046 | default: | |||
4047 | return SDValue(); | |||
4048 | } | |||
4049 | } | |||
4050 | ||||
4051 | SDValue AMDGPUTargetLowering::performRcpCombine(SDNode *N, | |||
4052 | DAGCombinerInfo &DCI) const { | |||
4053 | const auto *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)); | |||
4054 | if (!CFP) | |||
4055 | return SDValue(); | |||
4056 | ||||
4057 | // XXX - Should this flush denormals? | |||
4058 | const APFloat &Val = CFP->getValueAPF(); | |||
4059 | APFloat One(Val.getSemantics(), "1.0"); | |||
4060 | return DCI.DAG.getConstantFP(One / Val, SDLoc(N), N->getValueType(0)); | |||
4061 | } | |||
4062 | ||||
4063 | SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N, | |||
4064 | DAGCombinerInfo &DCI) const { | |||
4065 | SelectionDAG &DAG = DCI.DAG; | |||
4066 | SDLoc DL(N); | |||
4067 | ||||
4068 | switch(N->getOpcode()) { | |||
4069 | default: | |||
4070 | break; | |||
4071 | case ISD::BITCAST: { | |||
4072 | EVT DestVT = N->getValueType(0); | |||
4073 | ||||
4074 | // Push casts through vector builds. This helps avoid emitting a large | |||
4075 | // number of copies when materializing floating point vector constants. | |||
4076 | // | |||
4077 | // vNt1 bitcast (vNt0 (build_vector t0:x, t0:y)) => | |||
4078 | // vnt1 = build_vector (t1 (bitcast t0:x)), (t1 (bitcast t0:y)) | |||
4079 | if (DestVT.isVector()) { | |||
4080 | SDValue Src = N->getOperand(0); | |||
4081 | if (Src.getOpcode() == ISD::BUILD_VECTOR) { | |||
4082 | EVT SrcVT = Src.getValueType(); | |||
4083 | unsigned NElts = DestVT.getVectorNumElements(); | |||
4084 | ||||
4085 | if (SrcVT.getVectorNumElements() == NElts) { | |||
4086 | EVT DestEltVT = DestVT.getVectorElementType(); | |||
4087 | ||||
4088 | SmallVector<SDValue, 8> CastedElts; | |||
4089 | SDLoc SL(N); | |||
4090 | for (unsigned I = 0, E = SrcVT.getVectorNumElements(); I != E; ++I) { | |||
4091 | SDValue Elt = Src.getOperand(I); | |||
4092 | CastedElts.push_back(DAG.getNode(ISD::BITCAST, DL, DestEltVT, Elt)); | |||
4093 | } | |||
4094 | ||||
4095 | return DAG.getBuildVector(DestVT, SL, CastedElts); | |||
4096 | } | |||
4097 | } | |||
4098 | } | |||
4099 | ||||
4100 | if (DestVT.getSizeInBits() != 64 || !DestVT.isVector()) | |||
4101 | break; | |||
4102 | ||||
4103 | // Fold bitcasts of constants. | |||
4104 | // | |||
4105 | // v2i32 (bitcast i64:k) -> build_vector lo_32(k), hi_32(k) | |||
4106 | // TODO: Generalize and move to DAGCombiner | |||
4107 | SDValue Src = N->getOperand(0); | |||
4108 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Src)) { | |||
4109 | SDLoc SL(N); | |||
4110 | uint64_t CVal = C->getZExtValue(); | |||
4111 | SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32, | |||
4112 | DAG.getConstant(Lo_32(CVal), SL, MVT::i32), | |||
4113 | DAG.getConstant(Hi_32(CVal), SL, MVT::i32)); | |||
4114 | return DAG.getNode(ISD::BITCAST, SL, DestVT, BV); | |||
4115 | } | |||
4116 | ||||
4117 | if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Src)) { | |||
4118 | const APInt &Val = C->getValueAPF().bitcastToAPInt(); | |||
4119 | SDLoc SL(N); | |||
4120 | uint64_t CVal = Val.getZExtValue(); | |||
4121 | SDValue Vec = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32, | |||
4122 | DAG.getConstant(Lo_32(CVal), SL, MVT::i32), | |||
4123 | DAG.getConstant(Hi_32(CVal), SL, MVT::i32)); | |||
4124 | ||||
4125 | return DAG.getNode(ISD::BITCAST, SL, DestVT, Vec); | |||
4126 | } | |||
4127 | ||||
4128 | break; | |||
4129 | } | |||
4130 | case ISD::SHL: { | |||
4131 | if (DCI.getDAGCombineLevel() < AfterLegalizeDAG) | |||
4132 | break; | |||
4133 | ||||
4134 | return performShlCombine(N, DCI); | |||
4135 | } | |||
4136 | case ISD::SRL: { | |||
4137 | if (DCI.getDAGCombineLevel() < AfterLegalizeDAG) | |||
4138 | break; | |||
4139 | ||||
4140 | return performSrlCombine(N, DCI); | |||
4141 | } | |||
4142 | case ISD::SRA: { | |||
4143 | if (DCI.getDAGCombineLevel() < AfterLegalizeDAG) | |||
4144 | break; | |||
4145 | ||||
4146 | return performSraCombine(N, DCI); | |||
4147 | } | |||
4148 | case ISD::TRUNCATE: | |||
4149 | return performTruncateCombine(N, DCI); | |||
4150 | case ISD::MUL: | |||
4151 | return performMulCombine(N, DCI); | |||
4152 | case ISD::SMUL_LOHI: | |||
4153 | case ISD::UMUL_LOHI: | |||
4154 | return performMulLoHiCombine(N, DCI); | |||
4155 | case ISD::MULHS: | |||
4156 | return performMulhsCombine(N, DCI); | |||
4157 | case ISD::MULHU: | |||
4158 | return performMulhuCombine(N, DCI); | |||
4159 | case AMDGPUISD::MUL_I24: | |||
4160 | case AMDGPUISD::MUL_U24: | |||
4161 | case AMDGPUISD::MULHI_I24: | |||
4162 | case AMDGPUISD::MULHI_U24: | |||
4163 | return simplifyMul24(N, DCI); | |||
4164 | case ISD::SELECT: | |||
4165 | return performSelectCombine(N, DCI); | |||
4166 | case ISD::FNEG: | |||
4167 | return performFNegCombine(N, DCI); | |||
4168 | case ISD::FABS: | |||
4169 | return performFAbsCombine(N, DCI); | |||
4170 | case AMDGPUISD::BFE_I32: | |||
4171 | case AMDGPUISD::BFE_U32: { | |||
4172 | assert(!N->getValueType(0).isVector() &&(static_cast <bool> (!N->getValueType(0).isVector() && "Vector handling of BFE not implemented") ? void (0) : __assert_fail ("!N->getValueType(0).isVector() && \"Vector handling of BFE not implemented\"" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 4173, __extension__ __PRETTY_FUNCTION__)) | |||
4173 | "Vector handling of BFE not implemented")(static_cast <bool> (!N->getValueType(0).isVector() && "Vector handling of BFE not implemented") ? void (0) : __assert_fail ("!N->getValueType(0).isVector() && \"Vector handling of BFE not implemented\"" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 4173, __extension__ __PRETTY_FUNCTION__)); | |||
4174 | ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2)); | |||
4175 | if (!Width) | |||
4176 | break; | |||
4177 | ||||
4178 | uint32_t WidthVal = Width->getZExtValue() & 0x1f; | |||
4179 | if (WidthVal == 0) | |||
4180 | return DAG.getConstant(0, DL, MVT::i32); | |||
4181 | ||||
4182 | ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1)); | |||
4183 | if (!Offset) | |||
4184 | break; | |||
4185 | ||||
4186 | SDValue BitsFrom = N->getOperand(0); | |||
4187 | uint32_t OffsetVal = Offset->getZExtValue() & 0x1f; | |||
4188 | ||||
4189 | bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32; | |||
4190 | ||||
4191 | if (OffsetVal == 0) { | |||
4192 | // This is already sign / zero extended, so try to fold away extra BFEs. | |||
4193 | unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal); | |||
4194 | ||||
4195 | unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom); | |||
4196 | if (OpSignBits >= SignBits) | |||
4197 | return BitsFrom; | |||
4198 | ||||
4199 | EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal); | |||
4200 | if (Signed) { | |||
4201 | // This is a sign_extend_inreg. Replace it to take advantage of existing | |||
4202 | // DAG Combines. If not eliminated, we will match back to BFE during | |||
4203 | // selection. | |||
4204 | ||||
4205 | // TODO: The sext_inreg of extended types ends, although we can could | |||
4206 | // handle them in a single BFE. | |||
4207 | return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom, | |||
4208 | DAG.getValueType(SmallVT)); | |||
4209 | } | |||
4210 | ||||
4211 | return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT); | |||
4212 | } | |||
4213 | ||||
4214 | if (ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(BitsFrom)) { | |||
4215 | if (Signed) { | |||
4216 | return constantFoldBFE<int32_t>(DAG, | |||
4217 | CVal->getSExtValue(), | |||
4218 | OffsetVal, | |||
4219 | WidthVal, | |||
4220 | DL); | |||
4221 | } | |||
4222 | ||||
4223 | return constantFoldBFE<uint32_t>(DAG, | |||
4224 | CVal->getZExtValue(), | |||
4225 | OffsetVal, | |||
4226 | WidthVal, | |||
4227 | DL); | |||
4228 | } | |||
4229 | ||||
4230 | if ((OffsetVal + WidthVal) >= 32 && | |||
4231 | !(Subtarget->hasSDWA() && OffsetVal == 16 && WidthVal == 16)) { | |||
4232 | SDValue ShiftVal = DAG.getConstant(OffsetVal, DL, MVT::i32); | |||
4233 | return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32, | |||
4234 | BitsFrom, ShiftVal); | |||
4235 | } | |||
4236 | ||||
4237 | if (BitsFrom.hasOneUse()) { | |||
4238 | APInt Demanded = APInt::getBitsSet(32, | |||
4239 | OffsetVal, | |||
4240 | OffsetVal + WidthVal); | |||
4241 | ||||
4242 | KnownBits Known; | |||
4243 | TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(), | |||
4244 | !DCI.isBeforeLegalizeOps()); | |||
4245 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
4246 | if (TLI.ShrinkDemandedConstant(BitsFrom, Demanded, TLO) || | |||
4247 | TLI.SimplifyDemandedBits(BitsFrom, Demanded, Known, TLO)) { | |||
4248 | DCI.CommitTargetLoweringOpt(TLO); | |||
4249 | } | |||
4250 | } | |||
4251 | ||||
4252 | break; | |||
4253 | } | |||
4254 | case ISD::LOAD: | |||
4255 | return performLoadCombine(N, DCI); | |||
4256 | case ISD::STORE: | |||
4257 | return performStoreCombine(N, DCI); | |||
4258 | case AMDGPUISD::RCP: | |||
4259 | case AMDGPUISD::RCP_IFLAG: | |||
4260 | return performRcpCombine(N, DCI); | |||
4261 | case ISD::AssertZext: | |||
4262 | case ISD::AssertSext: | |||
4263 | return performAssertSZExtCombine(N, DCI); | |||
4264 | case ISD::INTRINSIC_WO_CHAIN: | |||
4265 | return performIntrinsicWOChainCombine(N, DCI); | |||
4266 | } | |||
4267 | return SDValue(); | |||
4268 | } | |||
4269 | ||||
4270 | //===----------------------------------------------------------------------===// | |||
4271 | // Helper functions | |||
4272 | //===----------------------------------------------------------------------===// | |||
4273 | ||||
4274 | SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG, | |||
4275 | const TargetRegisterClass *RC, | |||
4276 | Register Reg, EVT VT, | |||
4277 | const SDLoc &SL, | |||
4278 | bool RawReg) const { | |||
4279 | MachineFunction &MF = DAG.getMachineFunction(); | |||
4280 | MachineRegisterInfo &MRI = MF.getRegInfo(); | |||
4281 | Register VReg; | |||
4282 | ||||
4283 | if (!MRI.isLiveIn(Reg)) { | |||
4284 | VReg = MRI.createVirtualRegister(RC); | |||
4285 | MRI.addLiveIn(Reg, VReg); | |||
4286 | } else { | |||
4287 | VReg = MRI.getLiveInVirtReg(Reg); | |||
4288 | } | |||
4289 | ||||
4290 | if (RawReg) | |||
4291 | return DAG.getRegister(VReg, VT); | |||
4292 | ||||
4293 | return DAG.getCopyFromReg(DAG.getEntryNode(), SL, VReg, VT); | |||
4294 | } | |||
4295 | ||||
4296 | // This may be called multiple times, and nothing prevents creating multiple | |||
4297 | // objects at the same offset. See if we already defined this object. | |||
4298 | static int getOrCreateFixedStackObject(MachineFrameInfo &MFI, unsigned Size, | |||
4299 | int64_t Offset) { | |||
4300 | for (int I = MFI.getObjectIndexBegin(); I < 0; ++I) { | |||
4301 | if (MFI.getObjectOffset(I) == Offset) { | |||
4302 | assert(MFI.getObjectSize(I) == Size)(static_cast <bool> (MFI.getObjectSize(I) == Size) ? void (0) : __assert_fail ("MFI.getObjectSize(I) == Size", "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp" , 4302, __extension__ __PRETTY_FUNCTION__)); | |||
4303 | return I; | |||
4304 | } | |||
4305 | } | |||
4306 | ||||
4307 | return MFI.CreateFixedObject(Size, Offset, true); | |||
4308 | } | |||
4309 | ||||
4310 | SDValue AMDGPUTargetLowering::loadStackInputValue(SelectionDAG &DAG, | |||
4311 | EVT VT, | |||
4312 | const SDLoc &SL, | |||
4313 | int64_t Offset) const { | |||
4314 | MachineFunction &MF = DAG.getMachineFunction(); | |||
4315 | MachineFrameInfo &MFI = MF.getFrameInfo(); | |||
4316 | int FI = getOrCreateFixedStackObject(MFI, VT.getStoreSize(), Offset); | |||
4317 | ||||
4318 | auto SrcPtrInfo = MachinePointerInfo::getStack(MF, Offset); | |||
4319 | SDValue Ptr = DAG.getFrameIndex(FI, MVT::i32); | |||
4320 | ||||
4321 | return DAG.getLoad(VT, SL, DAG.getEntryNode(), Ptr, SrcPtrInfo, Align(4), | |||
4322 | MachineMemOperand::MODereferenceable | | |||
4323 | MachineMemOperand::MOInvariant); | |||
4324 | } | |||
4325 | ||||
4326 | SDValue AMDGPUTargetLowering::storeStackInputValue(SelectionDAG &DAG, | |||
4327 | const SDLoc &SL, | |||
4328 | SDValue Chain, | |||
4329 | SDValue ArgVal, | |||
4330 | int64_t Offset) const { | |||
4331 | MachineFunction &MF = DAG.getMachineFunction(); | |||
4332 | MachinePointerInfo DstInfo = MachinePointerInfo::getStack(MF, Offset); | |||
4333 | const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>(); | |||
4334 | ||||
4335 | SDValue Ptr = DAG.getConstant(Offset, SL, MVT::i32); | |||
4336 | // Stores to the argument stack area are relative to the stack pointer. | |||
4337 | SDValue SP = | |||
4338 | DAG.getCopyFromReg(Chain, SL, Info->getStackPtrOffsetReg(), MVT::i32); | |||
4339 | Ptr = DAG.getNode(ISD::ADD, SL, MVT::i32, SP, Ptr); | |||
4340 | SDValue Store = DAG.getStore(Chain, SL, ArgVal, Ptr, DstInfo, Align(4), | |||
4341 | MachineMemOperand::MODereferenceable); | |||
4342 | return Store; | |||
4343 | } | |||
4344 | ||||
4345 | SDValue AMDGPUTargetLowering::loadInputValue(SelectionDAG &DAG, | |||
4346 | const TargetRegisterClass *RC, | |||
4347 | EVT VT, const SDLoc &SL, | |||
4348 | const ArgDescriptor &Arg) const { | |||
4349 | assert(Arg && "Attempting to load missing argument")(static_cast <bool> (Arg && "Attempting to load missing argument" ) ? void (0) : __assert_fail ("Arg && \"Attempting to load missing argument\"" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 4349, __extension__ __PRETTY_FUNCTION__)); | |||
| ||||
4350 | ||||
4351 | SDValue V = Arg.isRegister() ? | |||
4352 | CreateLiveInRegister(DAG, RC, Arg.getRegister(), VT, SL) : | |||
4353 | loadStackInputValue(DAG, VT, SL, Arg.getStackOffset()); | |||
4354 | ||||
4355 | if (!Arg.isMasked()) | |||
4356 | return V; | |||
4357 | ||||
4358 | unsigned Mask = Arg.getMask(); | |||
4359 | unsigned Shift = countTrailingZeros<unsigned>(Mask); | |||
4360 | V = DAG.getNode(ISD::SRL, SL, VT, V, | |||
4361 | DAG.getShiftAmountConstant(Shift, VT, SL)); | |||
4362 | return DAG.getNode(ISD::AND, SL, VT, V, | |||
4363 | DAG.getConstant(Mask >> Shift, SL, VT)); | |||
| ||||
4364 | } | |||
4365 | ||||
4366 | uint32_t AMDGPUTargetLowering::getImplicitParameterOffset( | |||
4367 | const MachineFunction &MF, const ImplicitParameter Param) const { | |||
4368 | const AMDGPUMachineFunction *MFI = MF.getInfo<AMDGPUMachineFunction>(); | |||
4369 | const AMDGPUSubtarget &ST = | |||
4370 | AMDGPUSubtarget::get(getTargetMachine(), MF.getFunction()); | |||
4371 | unsigned ExplicitArgOffset = ST.getExplicitKernelArgOffset(MF.getFunction()); | |||
4372 | const Align Alignment = ST.getAlignmentForImplicitArgPtr(); | |||
4373 | uint64_t ArgOffset = alignTo(MFI->getExplicitKernArgSize(), Alignment) + | |||
4374 | ExplicitArgOffset; | |||
4375 | switch (Param) { | |||
4376 | case FIRST_IMPLICIT: | |||
4377 | return ArgOffset; | |||
4378 | case PRIVATE_BASE: | |||
4379 | return ArgOffset + AMDGPU::ImplicitArg::PRIVATE_BASE_OFFSET; | |||
4380 | case SHARED_BASE: | |||
4381 | return ArgOffset + AMDGPU::ImplicitArg::SHARED_BASE_OFFSET; | |||
4382 | case QUEUE_PTR: | |||
4383 | return ArgOffset + AMDGPU::ImplicitArg::QUEUE_PTR_OFFSET; | |||
4384 | } | |||
4385 | llvm_unreachable("unexpected implicit parameter type")::llvm::llvm_unreachable_internal("unexpected implicit parameter type" , "llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp", 4385); | |||
4386 | } | |||
4387 | ||||
4388 | #define NODE_NAME_CASE(node)case AMDGPUISD::node: return "node"; case AMDGPUISD::node: return #node; | |||
4389 | ||||
4390 | const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const { | |||
4391 | switch ((AMDGPUISD::NodeType)Opcode) { | |||
4392 | case AMDGPUISD::FIRST_NUMBER: break; | |||
4393 | // AMDIL DAG nodes | |||
4394 | NODE_NAME_CASE(UMUL)case AMDGPUISD::UMUL: return "UMUL";; | |||
4395 | NODE_NAME_CASE(BRANCH_COND)case AMDGPUISD::BRANCH_COND: return "BRANCH_COND";; | |||
4396 | ||||
4397 | // AMDGPU DAG nodes | |||
4398 | NODE_NAME_CASE(IF)case AMDGPUISD::IF: return "IF"; | |||
4399 | NODE_NAME_CASE(ELSE)case AMDGPUISD::ELSE: return "ELSE"; | |||
4400 | NODE_NAME_CASE(LOOP)case AMDGPUISD::LOOP: return "LOOP"; | |||
4401 | NODE_NAME_CASE(CALL)case AMDGPUISD::CALL: return "CALL"; | |||
4402 | NODE_NAME_CASE(TC_RETURN)case AMDGPUISD::TC_RETURN: return "TC_RETURN"; | |||
4403 | NODE_NAME_CASE(TRAP)case AMDGPUISD::TRAP: return "TRAP"; | |||
4404 | NODE_NAME_CASE(RET_FLAG)case AMDGPUISD::RET_FLAG: return "RET_FLAG"; | |||
4405 | NODE_NAME_CASE(RETURN_TO_EPILOG)case AMDGPUISD::RETURN_TO_EPILOG: return "RETURN_TO_EPILOG"; | |||
4406 | NODE_NAME_CASE(ENDPGM)case AMDGPUISD::ENDPGM: return "ENDPGM"; | |||
4407 | NODE_NAME_CASE(DWORDADDR)case AMDGPUISD::DWORDADDR: return "DWORDADDR"; | |||
4408 | NODE_NAME_CASE(FRACT)case AMDGPUISD::FRACT: return "FRACT"; | |||
4409 | NODE_NAME_CASE(SETCC)case AMDGPUISD::SETCC: return "SETCC"; | |||
4410 | NODE_NAME_CASE(SETREG)case AMDGPUISD::SETREG: return "SETREG"; | |||
4411 | NODE_NAME_CASE(DENORM_MODE)case AMDGPUISD::DENORM_MODE: return "DENORM_MODE"; | |||
4412 | NODE_NAME_CASE(FMA_W_CHAIN)case AMDGPUISD::FMA_W_CHAIN: return "FMA_W_CHAIN"; | |||
4413 | NODE_NAME_CASE(FMUL_W_CHAIN)case AMDGPUISD::FMUL_W_CHAIN: return "FMUL_W_CHAIN"; | |||
4414 | NODE_NAME_CASE(CLAMP)case AMDGPUISD::CLAMP: return "CLAMP"; | |||
4415 | NODE_NAME_CASE(COS_HW)case AMDGPUISD::COS_HW: return "COS_HW"; | |||
4416 | NODE_NAME_CASE(SIN_HW)case AMDGPUISD::SIN_HW: return "SIN_HW"; | |||
4417 | NODE_NAME_CASE(FMAX_LEGACY)case AMDGPUISD::FMAX_LEGACY: return "FMAX_LEGACY"; | |||
4418 | NODE_NAME_CASE(FMIN_LEGACY)case AMDGPUISD::FMIN_LEGACY: return "FMIN_LEGACY"; | |||
4419 | NODE_NAME_CASE(FMAX3)case AMDGPUISD::FMAX3: return "FMAX3"; | |||
4420 | NODE_NAME_CASE(SMAX3)case AMDGPUISD::SMAX3: return "SMAX3"; | |||
4421 | NODE_NAME_CASE(UMAX3)case AMDGPUISD::UMAX3: return "UMAX3"; | |||
4422 | NODE_NAME_CASE(FMIN3)case AMDGPUISD::FMIN3: return "FMIN3"; | |||
4423 | NODE_NAME_CASE(SMIN3)case AMDGPUISD::SMIN3: return "SMIN3"; | |||
4424 | NODE_NAME_CASE(UMIN3)case AMDGPUISD::UMIN3: return "UMIN3"; | |||
4425 | NODE_NAME_CASE(FMED3)case AMDGPUISD::FMED3: return "FMED3"; | |||
4426 | NODE_NAME_CASE(SMED3)case AMDGPUISD::SMED3: return "SMED3"; | |||
4427 | NODE_NAME_CASE(UMED3)case AMDGPUISD::UMED3: return "UMED3"; | |||
4428 | NODE_NAME_CASE(FDOT2)case AMDGPUISD::FDOT2: return "FDOT2"; | |||
4429 | NODE_NAME_CASE(URECIP)case AMDGPUISD::URECIP: return "URECIP"; | |||
4430 | NODE_NAME_CASE(DIV_SCALE)case AMDGPUISD::DIV_SCALE: return "DIV_SCALE"; | |||
4431 | NODE_NAME_CASE(DIV_FMAS)case AMDGPUISD::DIV_FMAS: return "DIV_FMAS"; | |||
4432 | NODE_NAME_CASE(DIV_FIXUP)case AMDGPUISD::DIV_FIXUP: return "DIV_FIXUP"; | |||
4433 | NODE_NAME_CASE(FMAD_FTZ)case AMDGPUISD::FMAD_FTZ: return "FMAD_FTZ"; | |||
4434 | NODE_NAME_CASE(RCP)case AMDGPUISD::RCP: return "RCP"; | |||
4435 | NODE_NAME_CASE(RSQ)case AMDGPUISD::RSQ: return "RSQ"; | |||
4436 | NODE_NAME_CASE(RCP_LEGACY)case AMDGPUISD::RCP_LEGACY: return "RCP_LEGACY"; | |||
4437 | NODE_NAME_CASE(RCP_IFLAG)case AMDGPUISD::RCP_IFLAG: return "RCP_IFLAG"; | |||
4438 | NODE_NAME_CASE(FMUL_LEGACY)case AMDGPUISD::FMUL_LEGACY: return "FMUL_LEGACY"; | |||
4439 | NODE_NAME_CASE(RSQ_CLAMP)case AMDGPUISD::RSQ_CLAMP: return "RSQ_CLAMP"; | |||
4440 | NODE_NAME_CASE(LDEXP)case AMDGPUISD::LDEXP: return "LDEXP"; | |||
4441 | NODE_NAME_CASE(FP_CLASS)case AMDGPUISD::FP_CLASS: return "FP_CLASS"; | |||
4442 | NODE_NAME_CASE(DOT4)case AMDGPUISD::DOT4: return "DOT4"; | |||
4443 | NODE_NAME_CASE(CARRY)case AMDGPUISD::CARRY: return "CARRY"; | |||
4444 | NODE_NAME_CASE(BORROW)case AMDGPUISD::BORROW: return "BORROW"; | |||
4445 | NODE_NAME_CASE(BFE_U32)case AMDGPUISD::BFE_U32: return "BFE_U32"; | |||
4446 | NODE_NAME_CASE(BFE_I32)case AMDGPUISD::BFE_I32: return "BFE_I32"; | |||
4447 | NODE_NAME_CASE(BFI)case AMDGPUISD::BFI: return "BFI"; | |||
4448 | NODE_NAME_CASE(BFM)case AMDGPUISD::BFM: return "BFM"; | |||
4449 | NODE_NAME_CASE(FFBH_U32)case AMDGPUISD::FFBH_U32: return "FFBH_U32"; | |||
4450 | NODE_NAME_CASE(FFBH_I32)case AMDGPUISD::FFBH_I32: return "FFBH_I32"; | |||
4451 | NODE_NAME_CASE(FFBL_B32)case AMDGPUISD::FFBL_B32: return "FFBL_B32"; | |||
4452 | NODE_NAME_CASE(MUL_U24)case AMDGPUISD::MUL_U24: return "MUL_U24"; | |||
4453 | NODE_NAME_CASE(MUL_I24)case AMDGPUISD::MUL_I24: return "MUL_I24"; | |||
4454 | NODE_NAME_CASE(MULHI_U24)case AMDGPUISD::MULHI_U24: return "MULHI_U24"; | |||
4455 | NODE_NAME_CASE(MULHI_I24)case AMDGPUISD::MULHI_I24: return "MULHI_I24"; | |||
4456 | NODE_NAME_CASE(MAD_U24)case AMDGPUISD::MAD_U24: return "MAD_U24"; | |||
4457 | NODE_NAME_CASE(MAD_I24)case AMDGPUISD::MAD_I24: return "MAD_I24"; | |||
4458 | NODE_NAME_CASE(MAD_I64_I32)case AMDGPUISD::MAD_I64_I32: return "MAD_I64_I32"; | |||
4459 | NODE_NAME_CASE(MAD_U64_U32)case AMDGPUISD::MAD_U64_U32: return "MAD_U64_U32"; | |||
4460 | NODE_NAME_CASE(PERM)case AMDGPUISD::PERM: return "PERM"; | |||
4461 | NODE_NAME_CASE(TEXTURE_FETCH)case AMDGPUISD::TEXTURE_FETCH: return "TEXTURE_FETCH"; | |||
4462 | NODE_NAME_CASE(R600_EXPORT)case AMDGPUISD::R600_EXPORT: return "R600_EXPORT"; | |||
4463 | NODE_NAME_CASE(CONST_ADDRESS)case AMDGPUISD::CONST_ADDRESS: return "CONST_ADDRESS"; | |||
4464 | NODE_NAME_CASE(REGISTER_LOAD)case AMDGPUISD::REGISTER_LOAD: return "REGISTER_LOAD"; | |||
4465 | NODE_NAME_CASE(REGISTER_STORE)case AMDGPUISD::REGISTER_STORE: return "REGISTER_STORE"; | |||
4466 | NODE_NAME_CASE(SAMPLE)case AMDGPUISD::SAMPLE: return "SAMPLE"; | |||
4467 | NODE_NAME_CASE(SAMPLEB)case AMDGPUISD::SAMPLEB: return "SAMPLEB"; | |||
4468 | NODE_NAME_CASE(SAMPLED)case AMDGPUISD::SAMPLED: return "SAMPLED"; | |||
4469 | NODE_NAME_CASE(SAMPLEL)case AMDGPUISD::SAMPLEL: return "SAMPLEL"; | |||
4470 | NODE_NAME_CASE(CVT_F32_UBYTE0)case AMDGPUISD::CVT_F32_UBYTE0: return "CVT_F32_UBYTE0"; | |||
4471 | NODE_NAME_CASE(CVT_F32_UBYTE1)case AMDGPUISD::CVT_F32_UBYTE1: return "CVT_F32_UBYTE1"; | |||
4472 | NODE_NAME_CASE(CVT_F32_UBYTE2)case AMDGPUISD::CVT_F32_UBYTE2: return "CVT_F32_UBYTE2"; | |||
4473 | NODE_NAME_CASE(CVT_F32_UBYTE3)case AMDGPUISD::CVT_F32_UBYTE3: return "CVT_F32_UBYTE3"; | |||
4474 | NODE_NAME_CASE(CVT_PKRTZ_F16_F32)case AMDGPUISD::CVT_PKRTZ_F16_F32: return "CVT_PKRTZ_F16_F32" ; | |||
4475 | NODE_NAME_CASE(CVT_PKNORM_I16_F32)case AMDGPUISD::CVT_PKNORM_I16_F32: return "CVT_PKNORM_I16_F32" ; | |||
4476 | NODE_NAME_CASE(CVT_PKNORM_U16_F32)case AMDGPUISD::CVT_PKNORM_U16_F32: return "CVT_PKNORM_U16_F32" ; | |||
4477 | NODE_NAME_CASE(CVT_PK_I16_I32)case AMDGPUISD::CVT_PK_I16_I32: return "CVT_PK_I16_I32"; | |||
4478 | NODE_NAME_CASE(CVT_PK_U16_U32)case AMDGPUISD::CVT_PK_U16_U32: return "CVT_PK_U16_U32"; | |||
4479 | NODE_NAME_CASE(FP_TO_FP16)case AMDGPUISD::FP_TO_FP16: return "FP_TO_FP16"; | |||
4480 | NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)case AMDGPUISD::BUILD_VERTICAL_VECTOR: return "BUILD_VERTICAL_VECTOR" ; | |||
4481 | NODE_NAME_CASE(CONST_DATA_PTR)case AMDGPUISD::CONST_DATA_PTR: return "CONST_DATA_PTR"; | |||
4482 | NODE_NAME_CASE(PC_ADD_REL_OFFSET)case AMDGPUISD::PC_ADD_REL_OFFSET: return "PC_ADD_REL_OFFSET" ; | |||
4483 | NODE_NAME_CASE(LDS)case AMDGPUISD::LDS: return "LDS"; | |||
4484 | NODE_NAME_CASE(FPTRUNC_ROUND_UPWARD)case AMDGPUISD::FPTRUNC_ROUND_UPWARD: return "FPTRUNC_ROUND_UPWARD" ; | |||
4485 | NODE_NAME_CASE(FPTRUNC_ROUND_DOWNWARD)case AMDGPUISD::FPTRUNC_ROUND_DOWNWARD: return "FPTRUNC_ROUND_DOWNWARD" ; | |||
4486 | NODE_NAME_CASE(DUMMY_CHAIN)case AMDGPUISD::DUMMY_CHAIN: return "DUMMY_CHAIN"; | |||
4487 | case AMDGPUISD::FIRST_MEM_OPCODE_NUMBER: break; | |||
4488 | NODE_NAME_CASE(LOAD_D16_HI)case AMDGPUISD::LOAD_D16_HI: return "LOAD_D16_HI"; | |||
4489 | NODE_NAME_CASE(LOAD_D16_LO)case AMDGPUISD::LOAD_D16_LO: return "LOAD_D16_LO"; | |||
4490 | NODE_NAME_CASE(LOAD_D16_HI_I8)case AMDGPUISD::LOAD_D16_HI_I8: return "LOAD_D16_HI_I8"; | |||
4491 | NODE_NAME_CASE(LOAD_D16_HI_U8)case AMDGPUISD::LOAD_D16_HI_U8: return "LOAD_D16_HI_U8"; | |||
4492 | NODE_NAME_CASE(LOAD_D16_LO_I8)case AMDGPUISD::LOAD_D16_LO_I8: return "LOAD_D16_LO_I8"; | |||
4493 | NODE_NAME_CASE(LOAD_D16_LO_U8)case AMDGPUISD::LOAD_D16_LO_U8: return "LOAD_D16_LO_U8"; | |||
4494 | NODE_NAME_CASE(STORE_MSKOR)case AMDGPUISD::STORE_MSKOR: return "STORE_MSKOR"; | |||
4495 | NODE_NAME_CASE(LOAD_CONSTANT)case AMDGPUISD::LOAD_CONSTANT: return "LOAD_CONSTANT"; | |||
4496 | NODE_NAME_CASE(TBUFFER_STORE_FORMAT)case AMDGPUISD::TBUFFER_STORE_FORMAT: return "TBUFFER_STORE_FORMAT" ; | |||
4497 | NODE_NAME_CASE(TBUFFER_STORE_FORMAT_D16)case AMDGPUISD::TBUFFER_STORE_FORMAT_D16: return "TBUFFER_STORE_FORMAT_D16" ; | |||
4498 | NODE_NAME_CASE(TBUFFER_LOAD_FORMAT)case AMDGPUISD::TBUFFER_LOAD_FORMAT: return "TBUFFER_LOAD_FORMAT" ; | |||
4499 | NODE_NAME_CASE(TBUFFER_LOAD_FORMAT_D16)case AMDGPUISD::TBUFFER_LOAD_FORMAT_D16: return "TBUFFER_LOAD_FORMAT_D16" ; | |||
4500 | NODE_NAME_CASE(DS_ORDERED_COUNT)case AMDGPUISD::DS_ORDERED_COUNT: return "DS_ORDERED_COUNT"; | |||
4501 | NODE_NAME_CASE(ATOMIC_CMP_SWAP)case AMDGPUISD::ATOMIC_CMP_SWAP: return "ATOMIC_CMP_SWAP"; | |||
4502 | NODE_NAME_CASE(ATOMIC_INC)case AMDGPUISD::ATOMIC_INC: return "ATOMIC_INC"; | |||
4503 | NODE_NAME_CASE(ATOMIC_DEC)case AMDGPUISD::ATOMIC_DEC: return "ATOMIC_DEC"; | |||
4504 | NODE_NAME_CASE(ATOMIC_LOAD_FMIN)case AMDGPUISD::ATOMIC_LOAD_FMIN: return "ATOMIC_LOAD_FMIN"; | |||
4505 | NODE_NAME_CASE(ATOMIC_LOAD_FMAX)case AMDGPUISD::ATOMIC_LOAD_FMAX: return "ATOMIC_LOAD_FMAX"; | |||
4506 | NODE_NAME_CASE(BUFFER_LOAD)case AMDGPUISD::BUFFER_LOAD: return "BUFFER_LOAD"; | |||
4507 | NODE_NAME_CASE(BUFFER_LOAD_UBYTE)case AMDGPUISD::BUFFER_LOAD_UBYTE: return "BUFFER_LOAD_UBYTE" ; | |||
4508 | NODE_NAME_CASE(BUFFER_LOAD_USHORT)case AMDGPUISD::BUFFER_LOAD_USHORT: return "BUFFER_LOAD_USHORT" ; | |||
4509 | NODE_NAME_CASE(BUFFER_LOAD_BYTE)case AMDGPUISD::BUFFER_LOAD_BYTE: return "BUFFER_LOAD_BYTE"; | |||
4510 | NODE_NAME_CASE(BUFFER_LOAD_SHORT)case AMDGPUISD::BUFFER_LOAD_SHORT: return "BUFFER_LOAD_SHORT" ; | |||
4511 | NODE_NAME_CASE(BUFFER_LOAD_FORMAT)case AMDGPUISD::BUFFER_LOAD_FORMAT: return "BUFFER_LOAD_FORMAT" ; | |||
4512 | NODE_NAME_CASE(BUFFER_LOAD_FORMAT_D16)case AMDGPUISD::BUFFER_LOAD_FORMAT_D16: return "BUFFER_LOAD_FORMAT_D16" ; | |||
4513 | NODE_NAME_CASE(SBUFFER_LOAD)case AMDGPUISD::SBUFFER_LOAD: return "SBUFFER_LOAD"; | |||
4514 | NODE_NAME_CASE(BUFFER_STORE)case AMDGPUISD::BUFFER_STORE: return "BUFFER_STORE"; | |||
4515 | NODE_NAME_CASE(BUFFER_STORE_BYTE)case AMDGPUISD::BUFFER_STORE_BYTE: return "BUFFER_STORE_BYTE" ; | |||
4516 | NODE_NAME_CASE(BUFFER_STORE_SHORT)case AMDGPUISD::BUFFER_STORE_SHORT: return "BUFFER_STORE_SHORT" ; | |||
4517 | NODE_NAME_CASE(BUFFER_STORE_FORMAT)case AMDGPUISD::BUFFER_STORE_FORMAT: return "BUFFER_STORE_FORMAT" ; | |||
4518 | NODE_NAME_CASE(BUFFER_STORE_FORMAT_D16)case AMDGPUISD::BUFFER_STORE_FORMAT_D16: return "BUFFER_STORE_FORMAT_D16" ; | |||
4519 | NODE_NAME_CASE(BUFFER_ATOMIC_SWAP)case AMDGPUISD::BUFFER_ATOMIC_SWAP: return "BUFFER_ATOMIC_SWAP" ; | |||
4520 | NODE_NAME_CASE(BUFFER_ATOMIC_ADD)case AMDGPUISD::BUFFER_ATOMIC_ADD: return "BUFFER_ATOMIC_ADD" ; | |||
4521 | NODE_NAME_CASE(BUFFER_ATOMIC_SUB)case AMDGPUISD::BUFFER_ATOMIC_SUB: return "BUFFER_ATOMIC_SUB" ; | |||
4522 | NODE_NAME_CASE(BUFFER_ATOMIC_SMIN)case AMDGPUISD::BUFFER_ATOMIC_SMIN: return "BUFFER_ATOMIC_SMIN" ; | |||
4523 | NODE_NAME_CASE(BUFFER_ATOMIC_UMIN)case AMDGPUISD::BUFFER_ATOMIC_UMIN: return "BUFFER_ATOMIC_UMIN" ; | |||
4524 | NODE_NAME_CASE(BUFFER_ATOMIC_SMAX)case AMDGPUISD::BUFFER_ATOMIC_SMAX: return "BUFFER_ATOMIC_SMAX" ; | |||
4525 | NODE_NAME_CASE(BUFFER_ATOMIC_UMAX)case AMDGPUISD::BUFFER_ATOMIC_UMAX: return "BUFFER_ATOMIC_UMAX" ; | |||
4526 | NODE_NAME_CASE(BUFFER_ATOMIC_AND)case AMDGPUISD::BUFFER_ATOMIC_AND: return "BUFFER_ATOMIC_AND" ; | |||
4527 | NODE_NAME_CASE(BUFFER_ATOMIC_OR)case AMDGPUISD::BUFFER_ATOMIC_OR: return "BUFFER_ATOMIC_OR"; | |||
4528 | NODE_NAME_CASE(BUFFER_ATOMIC_XOR)case AMDGPUISD::BUFFER_ATOMIC_XOR: return "BUFFER_ATOMIC_XOR" ; | |||
4529 | NODE_NAME_CASE(BUFFER_ATOMIC_INC)case AMDGPUISD::BUFFER_ATOMIC_INC: return "BUFFER_ATOMIC_INC" ; | |||
4530 | NODE_NAME_CASE(BUFFER_ATOMIC_DEC)case AMDGPUISD::BUFFER_ATOMIC_DEC: return "BUFFER_ATOMIC_DEC" ; | |||
4531 | NODE_NAME_CASE(BUFFER_ATOMIC_CMPSWAP)case AMDGPUISD::BUFFER_ATOMIC_CMPSWAP: return "BUFFER_ATOMIC_CMPSWAP" ; | |||
4532 | NODE_NAME_CASE(BUFFER_ATOMIC_CSUB)case AMDGPUISD::BUFFER_ATOMIC_CSUB: return "BUFFER_ATOMIC_CSUB" ; | |||
4533 | NODE_NAME_CASE(BUFFER_ATOMIC_FADD)case AMDGPUISD::BUFFER_ATOMIC_FADD: return "BUFFER_ATOMIC_FADD" ; | |||
4534 | NODE_NAME_CASE(BUFFER_ATOMIC_FMIN)case AMDGPUISD::BUFFER_ATOMIC_FMIN: return "BUFFER_ATOMIC_FMIN" ; | |||
4535 | NODE_NAME_CASE(BUFFER_ATOMIC_FMAX)case AMDGPUISD::BUFFER_ATOMIC_FMAX: return "BUFFER_ATOMIC_FMAX" ; | |||
4536 | ||||
4537 | case AMDGPUISD::LAST_AMDGPU_ISD_NUMBER: break; | |||
4538 | } | |||
4539 | return nullptr; | |||
4540 | } | |||
4541 | ||||
4542 | SDValue AMDGPUTargetLowering::getSqrtEstimate(SDValue Operand, | |||
4543 | SelectionDAG &DAG, int Enabled, | |||
4544 | int &RefinementSteps, | |||
4545 | bool &UseOneConstNR, | |||
4546 | bool Reciprocal) const { | |||
4547 | EVT VT = Operand.getValueType(); | |||
4548 | ||||
4549 | if (VT == MVT::f32) { | |||
4550 | RefinementSteps = 0; | |||
4551 | return DAG.getNode(AMDGPUISD::RSQ, SDLoc(Operand), VT, Operand); | |||
4552 | } | |||
4553 | ||||
4554 | // TODO: There is also f64 rsq instruction, but the documentation is less | |||
4555 | // clear on its precision. | |||
4556 | ||||
4557 | return SDValue(); | |||
4558 | } | |||
4559 | ||||
4560 | SDValue AMDGPUTargetLowering::getRecipEstimate(SDValue Operand, | |||
4561 | SelectionDAG &DAG, int Enabled, | |||
4562 | int &RefinementSteps) const { | |||
4563 | EVT VT = Operand.getValueType(); | |||
4564 | ||||
4565 | if (VT == MVT::f32) { | |||
4566 | // Reciprocal, < 1 ulp error. | |||
4567 | // | |||
4568 | // This reciprocal approximation converges to < 0.5 ulp error with one | |||
4569 | // newton rhapson performed with two fused multiple adds (FMAs). | |||
4570 | ||||
4571 | RefinementSteps = 0; | |||
4572 | return DAG.getNode(AMDGPUISD::RCP, SDLoc(Operand), VT, Operand); | |||
4573 | } | |||
4574 | ||||
4575 | // TODO: There is also f64 rcp instruction, but the documentation is less | |||
4576 | // clear on its precision. | |||
4577 | ||||
4578 | return SDValue(); | |||
4579 | } | |||
4580 | ||||
4581 | void AMDGPUTargetLowering::computeKnownBitsForTargetNode( | |||
4582 | const SDValue Op, KnownBits &Known, | |||
4583 | const APInt &DemandedElts, const SelectionDAG &DAG, unsigned Depth) const { | |||
4584 | ||||
4585 | Known.resetAll(); // Don't know anything. | |||
4586 | ||||
4587 | unsigned Opc = Op.getOpcode(); | |||
4588 | ||||
4589 | switch (Opc) { | |||
4590 | default: | |||
4591 | break; | |||
4592 | case AMDGPUISD::CARRY: | |||
4593 | case AMDGPUISD::BORROW: { | |||
4594 | Known.Zero = APInt::getHighBitsSet(32, 31); | |||
4595 | break; | |||
4596 | } | |||
4597 | ||||
4598 | case AMDGPUISD::BFE_I32: | |||
4599 | case AMDGPUISD::BFE_U32: { | |||
4600 | ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2)); | |||
4601 | if (!CWidth) | |||
4602 | return; | |||
4603 | ||||
4604 | uint32_t Width = CWidth->getZExtValue() & 0x1f; | |||
4605 | ||||
4606 | if (Opc == AMDGPUISD::BFE_U32) | |||
4607 | Known.Zero = APInt::getHighBitsSet(32, 32 - Width); | |||
4608 | ||||
4609 | break; | |||
4610 | } | |||
4611 | case AMDGPUISD::FP_TO_FP16: { | |||
4612 | unsigned BitWidth = Known.getBitWidth(); | |||
4613 | ||||
4614 | // High bits are zero. | |||
4615 | Known.Zero = APInt::getHighBitsSet(BitWidth, BitWidth - 16); | |||
4616 | break; | |||
4617 | } | |||
4618 | case AMDGPUISD::MUL_U24: | |||
4619 | case AMDGPUISD::MUL_I24: { | |||
4620 | KnownBits LHSKnown = DAG.computeKnownBits(Op.getOperand(0), Depth + 1); | |||
4621 | KnownBits RHSKnown = DAG.computeKnownBits(Op.getOperand(1), Depth + 1); | |||
4622 | unsigned TrailZ = LHSKnown.countMinTrailingZeros() + | |||
4623 | RHSKnown.countMinTrailingZeros(); | |||
4624 | Known.Zero.setLowBits(std::min(TrailZ, 32u)); | |||
4625 | // Skip extra check if all bits are known zeros. | |||
4626 | if (TrailZ >= 32) | |||
4627 | break; | |||
4628 | ||||
4629 | // Truncate to 24 bits. | |||
4630 | LHSKnown = LHSKnown.trunc(24); | |||
4631 | RHSKnown = RHSKnown.trunc(24); | |||
4632 | ||||
4633 | if (Opc == AMDGPUISD::MUL_I24) { | |||
4634 | unsigned LHSValBits = LHSKnown.countMaxSignificantBits(); | |||
4635 | unsigned RHSValBits = RHSKnown.countMaxSignificantBits(); | |||
4636 | unsigned MaxValBits = LHSValBits + RHSValBits; | |||
4637 | if (MaxValBits > 32) | |||
4638 | break; | |||
4639 | unsigned SignBits = 32 - MaxValBits + 1; | |||
4640 | bool LHSNegative = LHSKnown.isNegative(); | |||
4641 | bool LHSNonNegative = LHSKnown.isNonNegative(); | |||
4642 | bool LHSPositive = LHSKnown.isStrictlyPositive(); | |||
4643 | bool RHSNegative = RHSKnown.isNegative(); | |||
4644 | bool RHSNonNegative = RHSKnown.isNonNegative(); | |||
4645 | bool RHSPositive = RHSKnown.isStrictlyPositive(); | |||
4646 | ||||
4647 | if ((LHSNonNegative && RHSNonNegative) || (LHSNegative && RHSNegative)) | |||
4648 | Known.Zero.setHighBits(SignBits); | |||
4649 | else if ((LHSNegative && RHSPositive) || (LHSPositive && RHSNegative)) | |||
4650 | Known.One.setHighBits(SignBits); | |||
4651 | } else { | |||
4652 | unsigned LHSValBits = LHSKnown.countMaxActiveBits(); | |||
4653 | unsigned RHSValBits = RHSKnown.countMaxActiveBits(); | |||
4654 | unsigned MaxValBits = LHSValBits + RHSValBits; | |||
4655 | if (MaxValBits >= 32) | |||
4656 | break; | |||
4657 | Known.Zero.setBitsFrom(MaxValBits); | |||
4658 | } | |||
4659 | break; | |||
4660 | } | |||
4661 | case AMDGPUISD::PERM: { | |||
4662 | ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(Op.getOperand(2)); | |||
4663 | if (!CMask) | |||
4664 | return; | |||
4665 | ||||
4666 | KnownBits LHSKnown = DAG.computeKnownBits(Op.getOperand(0), Depth + 1); | |||
4667 | KnownBits RHSKnown = DAG.computeKnownBits(Op.getOperand(1), Depth + 1); | |||
4668 | unsigned Sel = CMask->getZExtValue(); | |||
4669 | ||||
4670 | for (unsigned I = 0; I < 32; I += 8) { | |||
4671 | unsigned SelBits = Sel & 0xff; | |||
4672 | if (SelBits < 4) { | |||
4673 | SelBits *= 8; | |||
4674 | Known.One |= ((RHSKnown.One.getZExtValue() >> SelBits) & 0xff) << I; | |||
4675 | Known.Zero |= ((RHSKnown.Zero.getZExtValue() >> SelBits) & 0xff) << I; | |||
4676 | } else if (SelBits < 7) { | |||
4677 | SelBits = (SelBits & 3) * 8; | |||
4678 | Known.One |= ((LHSKnown.One.getZExtValue() >> SelBits) & 0xff) << I; | |||
4679 | Known.Zero |= ((LHSKnown.Zero.getZExtValue() >> SelBits) & 0xff) << I; | |||
4680 | } else if (SelBits == 0x0c) { | |||
4681 | Known.Zero |= 0xFFull << I; | |||
4682 | } else if (SelBits > 0x0c) { | |||
4683 | Known.One |= 0xFFull << I; | |||
4684 | } | |||
4685 | Sel >>= 8; | |||
4686 | } | |||
4687 | break; | |||
4688 | } | |||
4689 | case AMDGPUISD::BUFFER_LOAD_UBYTE: { | |||
4690 | Known.Zero.setHighBits(24); | |||
4691 | break; | |||
4692 | } | |||
4693 | case AMDGPUISD::BUFFER_LOAD_USHORT: { | |||
4694 | Known.Zero.setHighBits(16); | |||
4695 | break; | |||
4696 | } | |||
4697 | case AMDGPUISD::LDS: { | |||
4698 | auto GA = cast<GlobalAddressSDNode>(Op.getOperand(0).getNode()); | |||
4699 | Align Alignment = GA->getGlobal()->getPointerAlignment(DAG.getDataLayout()); | |||
4700 | ||||
4701 | Known.Zero.setHighBits(16); | |||
4702 | Known.Zero.setLowBits(Log2(Alignment)); | |||
4703 | break; | |||
4704 | } | |||
4705 | case ISD::INTRINSIC_WO_CHAIN: { | |||
4706 | unsigned IID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); | |||
4707 | switch (IID) { | |||
4708 | case Intrinsic::amdgcn_mbcnt_lo: | |||
4709 | case Intrinsic::amdgcn_mbcnt_hi: { | |||
4710 | const GCNSubtarget &ST = | |||
4711 | DAG.getMachineFunction().getSubtarget<GCNSubtarget>(); | |||
4712 | // These return at most the wavefront size - 1. | |||
4713 | unsigned Size = Op.getValueType().getSizeInBits(); | |||
4714 | Known.Zero.setHighBits(Size - ST.getWavefrontSizeLog2()); | |||
4715 | break; | |||
4716 | } | |||
4717 | default: | |||
4718 | break; | |||
4719 | } | |||
4720 | } | |||
4721 | } | |||
4722 | } | |||
4723 | ||||
4724 | unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode( | |||
4725 | SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG, | |||
4726 | unsigned Depth) const { | |||
4727 | switch (Op.getOpcode()) { | |||
4728 | case AMDGPUISD::BFE_I32: { | |||
4729 | ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2)); | |||
4730 | if (!Width) | |||
4731 | return 1; | |||
4732 | ||||
4733 | unsigned SignBits = 32 - Width->getZExtValue() + 1; | |||
4734 | if (!isNullConstant(Op.getOperand(1))) | |||
4735 | return SignBits; | |||
4736 | ||||
4737 | // TODO: Could probably figure something out with non-0 offsets. | |||
4738 | unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1); | |||
4739 | return std::max(SignBits, Op0SignBits); | |||
4740 | } | |||
4741 | ||||
4742 | case AMDGPUISD::BFE_U32: { | |||
4743 | ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2)); | |||
4744 | return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1; | |||
4745 | } | |||
4746 | ||||
4747 | case AMDGPUISD::CARRY: | |||
4748 | case AMDGPUISD::BORROW: | |||
4749 | return 31; | |||
4750 | case AMDGPUISD::BUFFER_LOAD_BYTE: | |||
4751 | return 25; | |||
4752 | case AMDGPUISD::BUFFER_LOAD_SHORT: | |||
4753 | return 17; | |||
4754 | case AMDGPUISD::BUFFER_LOAD_UBYTE: | |||
4755 | return 24; | |||
4756 | case AMDGPUISD::BUFFER_LOAD_USHORT: | |||
4757 | return 16; | |||
4758 | case AMDGPUISD::FP_TO_FP16: | |||
4759 | return 16; | |||
4760 | default: | |||
4761 | return 1; | |||
4762 | } | |||
4763 | } | |||
4764 | ||||
4765 | unsigned AMDGPUTargetLowering::computeNumSignBitsForTargetInstr( | |||
4766 | GISelKnownBits &Analysis, Register R, | |||
4767 | const APInt &DemandedElts, const MachineRegisterInfo &MRI, | |||
4768 | unsigned Depth) const { | |||
4769 | const MachineInstr *MI = MRI.getVRegDef(R); | |||
4770 | if (!MI) | |||
4771 | return 1; | |||
4772 | ||||
4773 | // TODO: Check range metadata on MMO. | |||
4774 | switch (MI->getOpcode()) { | |||
4775 | case AMDGPU::G_AMDGPU_BUFFER_LOAD_SBYTE: | |||
4776 | return 25; | |||
4777 | case AMDGPU::G_AMDGPU_BUFFER_LOAD_SSHORT: | |||
4778 | return 17; | |||
4779 | case AMDGPU::G_AMDGPU_BUFFER_LOAD_UBYTE: | |||
4780 | return 24; | |||
4781 | case AMDGPU::G_AMDGPU_BUFFER_LOAD_USHORT: | |||
4782 | return 16; | |||
4783 | default: | |||
4784 | return 1; | |||
4785 | } | |||
4786 | } | |||
4787 | ||||
4788 | bool AMDGPUTargetLowering::isKnownNeverNaNForTargetNode(SDValue Op, | |||
4789 | const SelectionDAG &DAG, | |||
4790 | bool SNaN, | |||
4791 | unsigned Depth) const { | |||
4792 | unsigned Opcode = Op.getOpcode(); | |||
4793 | switch (Opcode) { | |||
4794 | case AMDGPUISD::FMIN_LEGACY: | |||
4795 | case AMDGPUISD::FMAX_LEGACY: { | |||
4796 | if (SNaN) | |||
4797 | return true; | |||
4798 | ||||
4799 | // TODO: Can check no nans on one of the operands for each one, but which | |||
4800 | // one? | |||
4801 | return false; | |||
4802 | } | |||
4803 | case AMDGPUISD::FMUL_LEGACY: | |||
4804 | case AMDGPUISD::CVT_PKRTZ_F16_F32: { | |||
4805 | if (SNaN) | |||
4806 | return true; | |||
4807 | return DAG.isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1) && | |||
4808 | DAG.isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1); | |||
4809 | } | |||
4810 | case AMDGPUISD::FMED3: | |||
4811 | case AMDGPUISD::FMIN3: | |||
4812 | case AMDGPUISD::FMAX3: | |||
4813 | case AMDGPUISD::FMAD_FTZ: { | |||
4814 | if (SNaN) | |||
4815 | return true; | |||
4816 | return DAG.isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1) && | |||
4817 | DAG.isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1) && | |||
4818 | DAG.isKnownNeverNaN(Op.getOperand(2), SNaN, Depth + 1); | |||
4819 | } | |||
4820 | case AMDGPUISD::CVT_F32_UBYTE0: | |||
4821 | case AMDGPUISD::CVT_F32_UBYTE1: | |||
4822 | case AMDGPUISD::CVT_F32_UBYTE2: | |||
4823 | case AMDGPUISD::CVT_F32_UBYTE3: | |||
4824 | return true; | |||
4825 | ||||
4826 | case AMDGPUISD::RCP: | |||
4827 | case AMDGPUISD::RSQ: | |||
4828 | case AMDGPUISD::RCP_LEGACY: | |||
4829 | case AMDGPUISD::RSQ_CLAMP: { | |||
4830 | if (SNaN) | |||
4831 | return true; | |||
4832 | ||||
4833 | // TODO: Need is known positive check. | |||
4834 | return false; | |||
4835 | } | |||
4836 | case AMDGPUISD::LDEXP: | |||
4837 | case AMDGPUISD::FRACT: { | |||
4838 | if (SNaN) | |||
4839 | return true; | |||
4840 | return DAG.isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1); | |||
4841 | } | |||
4842 | case AMDGPUISD::DIV_SCALE: | |||
4843 | case AMDGPUISD::DIV_FMAS: | |||
4844 | case AMDGPUISD::DIV_FIXUP: | |||
4845 | // TODO: Refine on operands. | |||
4846 | return SNaN; | |||
4847 | case AMDGPUISD::SIN_HW: | |||
4848 | case AMDGPUISD::COS_HW: { | |||
4849 | // TODO: Need check for infinity | |||
4850 | return SNaN; | |||
4851 | } | |||
4852 | case ISD::INTRINSIC_WO_CHAIN: { | |||
4853 | unsigned IntrinsicID | |||
4854 | = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); | |||
4855 | // TODO: Handle more intrinsics | |||
4856 | switch (IntrinsicID) { | |||
4857 | case Intrinsic::amdgcn_cubeid: | |||
4858 | return true; | |||
4859 | ||||
4860 | case Intrinsic::amdgcn_frexp_mant: { | |||
4861 | if (SNaN) | |||
4862 | return true; | |||
4863 | return DAG.isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1); | |||
4864 | } | |||
4865 | case Intrinsic::amdgcn_cvt_pkrtz: { | |||
4866 | if (SNaN) | |||
4867 | return true; | |||
4868 | return DAG.isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1) && | |||
4869 | DAG.isKnownNeverNaN(Op.getOperand(2), SNaN, Depth + 1); | |||
4870 | } | |||
4871 | case Intrinsic::amdgcn_rcp: | |||
4872 | case Intrinsic::amdgcn_rsq: | |||
4873 | case Intrinsic::amdgcn_rcp_legacy: | |||
4874 | case Intrinsic::amdgcn_rsq_legacy: | |||
4875 | case Intrinsic::amdgcn_rsq_clamp: { | |||
4876 | if (SNaN) | |||
4877 | return true; | |||
4878 | ||||
4879 | // TODO: Need is known positive check. | |||
4880 | return false; | |||
4881 | } | |||
4882 | case Intrinsic::amdgcn_trig_preop: | |||
4883 | case Intrinsic::amdgcn_fdot2: | |||
4884 | // TODO: Refine on operand | |||
4885 | return SNaN; | |||
4886 | case Intrinsic::amdgcn_fma_legacy: | |||
4887 | if (SNaN) | |||
4888 | return true; | |||
4889 | return DAG.isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1) && | |||
4890 | DAG.isKnownNeverNaN(Op.getOperand(2), SNaN, Depth + 1) && | |||
4891 | DAG.isKnownNeverNaN(Op.getOperand(3), SNaN, Depth + 1); | |||
4892 | default: | |||
4893 | return false; | |||
4894 | } | |||
4895 | } | |||
4896 | default: | |||
4897 | return false; | |||
4898 | } | |||
4899 | } | |||
4900 | ||||
4901 | TargetLowering::AtomicExpansionKind | |||
4902 | AMDGPUTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const { | |||
4903 | switch (RMW->getOperation()) { | |||
4904 | case AtomicRMWInst::Nand: | |||
4905 | case AtomicRMWInst::FAdd: | |||
4906 | case AtomicRMWInst::FSub: | |||
4907 | return AtomicExpansionKind::CmpXChg; | |||
4908 | default: | |||
4909 | return AtomicExpansionKind::None; | |||
4910 | } | |||
4911 | } | |||
4912 | ||||
4913 | bool AMDGPUTargetLowering::isConstantUnsignedBitfieldExtractLegal( | |||
4914 | unsigned Opc, LLT Ty1, LLT Ty2) const { | |||
4915 | return (Ty1 == LLT::scalar(32) || Ty1 == LLT::scalar(64)) && | |||
4916 | Ty2 == LLT::scalar(32); | |||
4917 | } |
1 | //===-- llvm/Support/MathExtras.h - Useful math functions -------*- C++ -*-===// | ||||
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 | // This file contains some functions that are useful for math stuff. | ||||
10 | // | ||||
11 | //===----------------------------------------------------------------------===// | ||||
12 | |||||
13 | #ifndef LLVM_SUPPORT_MATHEXTRAS_H | ||||
14 | #define LLVM_SUPPORT_MATHEXTRAS_H | ||||
15 | |||||
16 | #include "llvm/Support/Compiler.h" | ||||
17 | #include <cassert> | ||||
18 | #include <climits> | ||||
19 | #include <cmath> | ||||
20 | #include <cstdint> | ||||
21 | #include <cstring> | ||||
22 | #include <limits> | ||||
23 | #include <type_traits> | ||||
24 | |||||
25 | #ifdef __ANDROID_NDK__ | ||||
26 | #include <android/api-level.h> | ||||
27 | #endif | ||||
28 | |||||
29 | #ifdef _MSC_VER | ||||
30 | // Declare these intrinsics manually rather including intrin.h. It's very | ||||
31 | // expensive, and MathExtras.h is popular. | ||||
32 | // #include <intrin.h> | ||||
33 | extern "C" { | ||||
34 | unsigned char _BitScanForward(unsigned long *_Index, unsigned long _Mask); | ||||
35 | unsigned char _BitScanForward64(unsigned long *_Index, unsigned __int64 _Mask); | ||||
36 | unsigned char _BitScanReverse(unsigned long *_Index, unsigned long _Mask); | ||||
37 | unsigned char _BitScanReverse64(unsigned long *_Index, unsigned __int64 _Mask); | ||||
38 | } | ||||
39 | #endif | ||||
40 | |||||
41 | namespace llvm { | ||||
42 | |||||
43 | /// The behavior an operation has on an input of 0. | ||||
44 | enum ZeroBehavior { | ||||
45 | /// The returned value is undefined. | ||||
46 | ZB_Undefined, | ||||
47 | /// The returned value is numeric_limits<T>::max() | ||||
48 | ZB_Max, | ||||
49 | /// The returned value is numeric_limits<T>::digits | ||||
50 | ZB_Width | ||||
51 | }; | ||||
52 | |||||
53 | /// Mathematical constants. | ||||
54 | namespace numbers { | ||||
55 | // TODO: Track C++20 std::numbers. | ||||
56 | // TODO: Favor using the hexadecimal FP constants (requires C++17). | ||||
57 | constexpr double e = 2.7182818284590452354, // (0x1.5bf0a8b145749P+1) https://oeis.org/A001113 | ||||
58 | egamma = .57721566490153286061, // (0x1.2788cfc6fb619P-1) https://oeis.org/A001620 | ||||
59 | ln2 = .69314718055994530942, // (0x1.62e42fefa39efP-1) https://oeis.org/A002162 | ||||
60 | ln10 = 2.3025850929940456840, // (0x1.24bb1bbb55516P+1) https://oeis.org/A002392 | ||||
61 | log2e = 1.4426950408889634074, // (0x1.71547652b82feP+0) | ||||
62 | log10e = .43429448190325182765, // (0x1.bcb7b1526e50eP-2) | ||||
63 | pi = 3.1415926535897932385, // (0x1.921fb54442d18P+1) https://oeis.org/A000796 | ||||
64 | inv_pi = .31830988618379067154, // (0x1.45f306bc9c883P-2) https://oeis.org/A049541 | ||||
65 | sqrtpi = 1.7724538509055160273, // (0x1.c5bf891b4ef6bP+0) https://oeis.org/A002161 | ||||
66 | inv_sqrtpi = .56418958354775628695, // (0x1.20dd750429b6dP-1) https://oeis.org/A087197 | ||||
67 | sqrt2 = 1.4142135623730950488, // (0x1.6a09e667f3bcdP+0) https://oeis.org/A00219 | ||||
68 | inv_sqrt2 = .70710678118654752440, // (0x1.6a09e667f3bcdP-1) | ||||
69 | sqrt3 = 1.7320508075688772935, // (0x1.bb67ae8584caaP+0) https://oeis.org/A002194 | ||||
70 | inv_sqrt3 = .57735026918962576451, // (0x1.279a74590331cP-1) | ||||
71 | phi = 1.6180339887498948482; // (0x1.9e3779b97f4a8P+0) https://oeis.org/A001622 | ||||
72 | constexpr float ef = 2.71828183F, // (0x1.5bf0a8P+1) https://oeis.org/A001113 | ||||
73 | egammaf = .577215665F, // (0x1.2788d0P-1) https://oeis.org/A001620 | ||||
74 | ln2f = .693147181F, // (0x1.62e430P-1) https://oeis.org/A002162 | ||||
75 | ln10f = 2.30258509F, // (0x1.26bb1cP+1) https://oeis.org/A002392 | ||||
76 | log2ef = 1.44269504F, // (0x1.715476P+0) | ||||
77 | log10ef = .434294482F, // (0x1.bcb7b2P-2) | ||||
78 | pif = 3.14159265F, // (0x1.921fb6P+1) https://oeis.org/A000796 | ||||
79 | inv_pif = .318309886F, // (0x1.45f306P-2) https://oeis.org/A049541 | ||||
80 | sqrtpif = 1.77245385F, // (0x1.c5bf8aP+0) https://oeis.org/A002161 | ||||
81 | inv_sqrtpif = .564189584F, // (0x1.20dd76P-1) https://oeis.org/A087197 | ||||
82 | sqrt2f = 1.41421356F, // (0x1.6a09e6P+0) https://oeis.org/A002193 | ||||
83 | inv_sqrt2f = .707106781F, // (0x1.6a09e6P-1) | ||||
84 | sqrt3f = 1.73205081F, // (0x1.bb67aeP+0) https://oeis.org/A002194 | ||||
85 | inv_sqrt3f = .577350269F, // (0x1.279a74P-1) | ||||
86 | phif = 1.61803399F; // (0x1.9e377aP+0) https://oeis.org/A001622 | ||||
87 | } // namespace numbers | ||||
88 | |||||
89 | namespace detail { | ||||
90 | template <typename T, std::size_t SizeOfT> struct TrailingZerosCounter { | ||||
91 | static unsigned count(T Val, ZeroBehavior) { | ||||
92 | if (!Val) | ||||
93 | return std::numeric_limits<T>::digits; | ||||
94 | if (Val & 0x1) | ||||
95 | return 0; | ||||
96 | |||||
97 | // Bisection method. | ||||
98 | unsigned ZeroBits = 0; | ||||
99 | T Shift = std::numeric_limits<T>::digits >> 1; | ||||
100 | T Mask = std::numeric_limits<T>::max() >> Shift; | ||||
101 | while (Shift) { | ||||
102 | if ((Val & Mask) == 0) { | ||||
103 | Val >>= Shift; | ||||
104 | ZeroBits |= Shift; | ||||
105 | } | ||||
106 | Shift >>= 1; | ||||
107 | Mask >>= Shift; | ||||
108 | } | ||||
109 | return ZeroBits; | ||||
110 | } | ||||
111 | }; | ||||
112 | |||||
113 | #if defined(__GNUC__4) || defined(_MSC_VER) | ||||
114 | template <typename T> struct TrailingZerosCounter<T, 4> { | ||||
115 | static unsigned count(T Val, ZeroBehavior ZB) { | ||||
116 | if (ZB
| ||||
117 | return 32; | ||||
118 | |||||
119 | #if __has_builtin(__builtin_ctz)1 || defined(__GNUC__4) | ||||
120 | return __builtin_ctz(Val); | ||||
121 | #elif defined(_MSC_VER) | ||||
122 | unsigned long Index; | ||||
123 | _BitScanForward(&Index, Val); | ||||
124 | return Index; | ||||
125 | #endif | ||||
126 | } | ||||
127 | }; | ||||
128 | |||||
129 | #if !defined(_MSC_VER) || defined(_M_X64) | ||||
130 | template <typename T> struct TrailingZerosCounter<T, 8> { | ||||
131 | static unsigned count(T Val, ZeroBehavior ZB) { | ||||
132 | if (ZB != ZB_Undefined && Val == 0) | ||||
133 | return 64; | ||||
134 | |||||
135 | #if __has_builtin(__builtin_ctzll)1 || defined(__GNUC__4) | ||||
136 | return __builtin_ctzll(Val); | ||||
137 | #elif defined(_MSC_VER) | ||||
138 | unsigned long Index; | ||||
139 | _BitScanForward64(&Index, Val); | ||||
140 | return Index; | ||||
141 | #endif | ||||
142 | } | ||||
143 | }; | ||||
144 | #endif | ||||
145 | #endif | ||||
146 | } // namespace detail | ||||
147 | |||||
148 | /// Count number of 0's from the least significant bit to the most | ||||
149 | /// stopping at the first 1. | ||||
150 | /// | ||||
151 | /// Only unsigned integral types are allowed. | ||||
152 | /// | ||||
153 | /// \param ZB the behavior on an input of 0. Only ZB_Width and ZB_Undefined are | ||||
154 | /// valid arguments. | ||||
155 | template <typename T> | ||||
156 | unsigned countTrailingZeros(T Val, ZeroBehavior ZB = ZB_Width) { | ||||
157 | static_assert(std::numeric_limits<T>::is_integer && | ||||
158 | !std::numeric_limits<T>::is_signed, | ||||
159 | "Only unsigned integral types are allowed."); | ||||
160 | return llvm::detail::TrailingZerosCounter<T, sizeof(T)>::count(Val, ZB); | ||||
161 | } | ||||
162 | |||||
163 | namespace detail { | ||||
164 | template <typename T, std::size_t SizeOfT> struct LeadingZerosCounter { | ||||
165 | static unsigned count(T Val, ZeroBehavior) { | ||||
166 | if (!Val) | ||||
167 | return std::numeric_limits<T>::digits; | ||||
168 | |||||
169 | // Bisection method. | ||||
170 | unsigned ZeroBits = 0; | ||||
171 | for (T Shift = std::numeric_limits<T>::digits >> 1; Shift; Shift >>= 1) { | ||||
172 | T Tmp = Val >> Shift; | ||||
173 | if (Tmp) | ||||
174 | Val = Tmp; | ||||
175 | else | ||||
176 | ZeroBits |= Shift; | ||||
177 | } | ||||
178 | return ZeroBits; | ||||
179 | } | ||||
180 | }; | ||||
181 | |||||
182 | #if defined(__GNUC__4) || defined(_MSC_VER) | ||||
183 | template <typename T> struct LeadingZerosCounter<T, 4> { | ||||
184 | static unsigned count(T Val, ZeroBehavior ZB) { | ||||
185 | if (ZB != ZB_Undefined && Val == 0) | ||||
186 | return 32; | ||||
187 | |||||
188 | #if __has_builtin(__builtin_clz)1 || defined(__GNUC__4) | ||||
189 | return __builtin_clz(Val); | ||||
190 | #elif defined(_MSC_VER) | ||||
191 | unsigned long Index; | ||||
192 | _BitScanReverse(&Index, Val); | ||||
193 | return Index ^ 31; | ||||
194 | #endif | ||||
195 | } | ||||
196 | }; | ||||
197 | |||||
198 | #if !defined(_MSC_VER) || defined(_M_X64) | ||||
199 | template <typename T> struct LeadingZerosCounter<T, 8> { | ||||
200 | static unsigned count(T Val, ZeroBehavior ZB) { | ||||
201 | if (ZB != ZB_Undefined && Val == 0) | ||||
202 | return 64; | ||||
203 | |||||
204 | #if __has_builtin(__builtin_clzll)1 || defined(__GNUC__4) | ||||
205 | return __builtin_clzll(Val); | ||||
206 | #elif defined(_MSC_VER) | ||||
207 | unsigned long Index; | ||||
208 | _BitScanReverse64(&Index, Val); | ||||
209 | return Index ^ 63; | ||||
210 | #endif | ||||
211 | } | ||||
212 | }; | ||||
213 | #endif | ||||
214 | #endif | ||||
215 | } // namespace detail | ||||
216 | |||||
217 | /// Count number of 0's from the most significant bit to the least | ||||
218 | /// stopping at the first 1. | ||||
219 | /// | ||||
220 | /// Only unsigned integral types are allowed. | ||||
221 | /// | ||||
222 | /// \param ZB the behavior on an input of 0. Only ZB_Width and ZB_Undefined are | ||||
223 | /// valid arguments. | ||||
224 | template <typename T> | ||||
225 | unsigned countLeadingZeros(T Val, ZeroBehavior ZB = ZB_Width) { | ||||
226 | static_assert(std::numeric_limits<T>::is_integer && | ||||
227 | !std::numeric_limits<T>::is_signed, | ||||
228 | "Only unsigned integral types are allowed."); | ||||
229 | return llvm::detail::LeadingZerosCounter<T, sizeof(T)>::count(Val, ZB); | ||||
230 | } | ||||
231 | |||||
232 | /// Get the index of the first set bit starting from the least | ||||
233 | /// significant bit. | ||||
234 | /// | ||||
235 | /// Only unsigned integral types are allowed. | ||||
236 | /// | ||||
237 | /// \param ZB the behavior on an input of 0. Only ZB_Max and ZB_Undefined are | ||||
238 | /// valid arguments. | ||||
239 | template <typename T> T findFirstSet(T Val, ZeroBehavior ZB = ZB_Max) { | ||||
240 | if (ZB == ZB_Max && Val == 0) | ||||
241 | return std::numeric_limits<T>::max(); | ||||
242 | |||||
243 | return countTrailingZeros(Val, ZB_Undefined); | ||||
244 | } | ||||
245 | |||||
246 | /// Create a bitmask with the N right-most bits set to 1, and all other | ||||
247 | /// bits set to 0. Only unsigned types are allowed. | ||||
248 | template <typename T> T maskTrailingOnes(unsigned N) { | ||||
249 | static_assert(std::is_unsigned<T>::value, "Invalid type!"); | ||||
250 | const unsigned Bits = CHAR_BIT8 * sizeof(T); | ||||
251 | assert(N <= Bits && "Invalid bit index")(static_cast <bool> (N <= Bits && "Invalid bit index" ) ? void (0) : __assert_fail ("N <= Bits && \"Invalid bit index\"" , "llvm/include/llvm/Support/MathExtras.h", 251, __extension__ __PRETTY_FUNCTION__)); | ||||
252 | return N == 0 ? 0 : (T(-1) >> (Bits - N)); | ||||
253 | } | ||||
254 | |||||
255 | /// Create a bitmask with the N left-most bits set to 1, and all other | ||||
256 | /// bits set to 0. Only unsigned types are allowed. | ||||
257 | template <typename T> T maskLeadingOnes(unsigned N) { | ||||
258 | return ~maskTrailingOnes<T>(CHAR_BIT8 * sizeof(T) - N); | ||||
259 | } | ||||
260 | |||||
261 | /// Create a bitmask with the N right-most bits set to 0, and all other | ||||
262 | /// bits set to 1. Only unsigned types are allowed. | ||||
263 | template <typename T> T maskTrailingZeros(unsigned N) { | ||||
264 | return maskLeadingOnes<T>(CHAR_BIT8 * sizeof(T) - N); | ||||
265 | } | ||||
266 | |||||
267 | /// Create a bitmask with the N left-most bits set to 0, and all other | ||||
268 | /// bits set to 1. Only unsigned types are allowed. | ||||
269 | template <typename T> T maskLeadingZeros(unsigned N) { | ||||
270 | return maskTrailingOnes<T>(CHAR_BIT8 * sizeof(T) - N); | ||||
271 | } | ||||
272 | |||||
273 | /// Get the index of the last set bit starting from the least | ||||
274 | /// significant bit. | ||||
275 | /// | ||||
276 | /// Only unsigned integral types are allowed. | ||||
277 | /// | ||||
278 | /// \param ZB the behavior on an input of 0. Only ZB_Max and ZB_Undefined are | ||||
279 | /// valid arguments. | ||||
280 | template <typename T> T findLastSet(T Val, ZeroBehavior ZB = ZB_Max) { | ||||
281 | if (ZB == ZB_Max && Val == 0) | ||||
282 | return std::numeric_limits<T>::max(); | ||||
283 | |||||
284 | // Use ^ instead of - because both gcc and llvm can remove the associated ^ | ||||
285 | // in the __builtin_clz intrinsic on x86. | ||||
286 | return countLeadingZeros(Val, ZB_Undefined) ^ | ||||
287 | (std::numeric_limits<T>::digits - 1); | ||||
288 | } | ||||
289 | |||||
290 | /// Macro compressed bit reversal table for 256 bits. | ||||
291 | /// | ||||
292 | /// http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable | ||||
293 | static const unsigned char BitReverseTable256[256] = { | ||||
294 | #define R2(n) n, n + 2 * 64, n + 1 * 64, n + 3 * 64 | ||||
295 | #define R4(n) R2(n), R2(n + 2 * 16), R2(n + 1 * 16), R2(n + 3 * 16) | ||||
296 | #define R6(n) R4(n), R4(n + 2 * 4), R4(n + 1 * 4), R4(n + 3 * 4) | ||||
297 | R6(0), R6(2), R6(1), R6(3) | ||||
298 | #undef R2 | ||||
299 | #undef R4 | ||||
300 | #undef R6 | ||||
301 | }; | ||||
302 | |||||
303 | /// Reverse the bits in \p Val. | ||||
304 | template <typename T> | ||||
305 | T reverseBits(T Val) { | ||||
306 | unsigned char in[sizeof(Val)]; | ||||
307 | unsigned char out[sizeof(Val)]; | ||||
308 | std::memcpy(in, &Val, sizeof(Val)); | ||||
309 | for (unsigned i = 0; i < sizeof(Val); ++i) | ||||
310 | out[(sizeof(Val) - i) - 1] = BitReverseTable256[in[i]]; | ||||
311 | std::memcpy(&Val, out, sizeof(Val)); | ||||
312 | return Val; | ||||
313 | } | ||||
314 | |||||
315 | #if __has_builtin(__builtin_bitreverse8)1 | ||||
316 | template<> | ||||
317 | inline uint8_t reverseBits<uint8_t>(uint8_t Val) { | ||||
318 | return __builtin_bitreverse8(Val); | ||||
319 | } | ||||
320 | #endif | ||||
321 | |||||
322 | #if __has_builtin(__builtin_bitreverse16)1 | ||||
323 | template<> | ||||
324 | inline uint16_t reverseBits<uint16_t>(uint16_t Val) { | ||||
325 | return __builtin_bitreverse16(Val); | ||||
326 | } | ||||
327 | #endif | ||||
328 | |||||
329 | #if __has_builtin(__builtin_bitreverse32)1 | ||||
330 | template<> | ||||
331 | inline uint32_t reverseBits<uint32_t>(uint32_t Val) { | ||||
332 | return __builtin_bitreverse32(Val); | ||||
333 | } | ||||
334 | #endif | ||||
335 | |||||
336 | #if __has_builtin(__builtin_bitreverse64)1 | ||||
337 | template<> | ||||
338 | inline uint64_t reverseBits<uint64_t>(uint64_t Val) { | ||||
339 | return __builtin_bitreverse64(Val); | ||||
340 | } | ||||
341 | #endif | ||||
342 | |||||
343 | // NOTE: The following support functions use the _32/_64 extensions instead of | ||||
344 | // type overloading so that signed and unsigned integers can be used without | ||||
345 | // ambiguity. | ||||
346 | |||||
347 | /// Return the high 32 bits of a 64 bit value. | ||||
348 | constexpr inline uint32_t Hi_32(uint64_t Value) { | ||||
349 | return static_cast<uint32_t>(Value >> 32); | ||||
350 | } | ||||
351 | |||||
352 | /// Return the low 32 bits of a 64 bit value. | ||||
353 | constexpr inline uint32_t Lo_32(uint64_t Value) { | ||||
354 | return static_cast<uint32_t>(Value); | ||||
355 | } | ||||
356 | |||||
357 | /// Make a 64-bit integer from a high / low pair of 32-bit integers. | ||||
358 | constexpr inline uint64_t Make_64(uint32_t High, uint32_t Low) { | ||||
359 | return ((uint64_t)High << 32) | (uint64_t)Low; | ||||
360 | } | ||||
361 | |||||
362 | /// Checks if an integer fits into the given bit width. | ||||
363 | template <unsigned N> constexpr inline bool isInt(int64_t x) { | ||||
364 | return N >= 64 || (-(INT64_C(1)1L<<(N-1)) <= x && x < (INT64_C(1)1L<<(N-1))); | ||||
365 | } | ||||
366 | // Template specializations to get better code for common cases. | ||||
367 | template <> constexpr inline bool isInt<8>(int64_t x) { | ||||
368 | return static_cast<int8_t>(x) == x; | ||||
369 | } | ||||
370 | template <> constexpr inline bool isInt<16>(int64_t x) { | ||||
371 | return static_cast<int16_t>(x) == x; | ||||
372 | } | ||||
373 | template <> constexpr inline bool isInt<32>(int64_t x) { | ||||
374 | return static_cast<int32_t>(x) == x; | ||||
375 | } | ||||
376 | |||||
377 | /// Checks if a signed integer is an N bit number shifted left by S. | ||||
378 | template <unsigned N, unsigned S> | ||||
379 | constexpr inline bool isShiftedInt(int64_t x) { | ||||
380 | static_assert( | ||||
381 | N > 0, "isShiftedInt<0> doesn't make sense (refers to a 0-bit number."); | ||||
382 | static_assert(N + S <= 64, "isShiftedInt<N, S> with N + S > 64 is too wide."); | ||||
383 | return isInt<N + S>(x) && (x % (UINT64_C(1)1UL << S) == 0); | ||||
384 | } | ||||
385 | |||||
386 | /// Checks if an unsigned integer fits into the given bit width. | ||||
387 | /// | ||||
388 | /// This is written as two functions rather than as simply | ||||
389 | /// | ||||
390 | /// return N >= 64 || X < (UINT64_C(1) << N); | ||||
391 | /// | ||||
392 | /// to keep MSVC from (incorrectly) warning on isUInt<64> that we're shifting | ||||
393 | /// left too many places. | ||||
394 | template <unsigned N> | ||||
395 | constexpr inline std::enable_if_t<(N < 64), bool> isUInt(uint64_t X) { | ||||
396 | static_assert(N > 0, "isUInt<0> doesn't make sense"); | ||||
397 | return X < (UINT64_C(1)1UL << (N)); | ||||
398 | } | ||||
399 | template <unsigned N> | ||||
400 | constexpr inline std::enable_if_t<N >= 64, bool> isUInt(uint64_t) { | ||||
401 | return true; | ||||
402 | } | ||||
403 | |||||
404 | // Template specializations to get better code for common cases. | ||||
405 | template <> constexpr inline bool isUInt<8>(uint64_t x) { | ||||
406 | return static_cast<uint8_t>(x) == x; | ||||
407 | } | ||||
408 | template <> constexpr inline bool isUInt<16>(uint64_t x) { | ||||
409 | return static_cast<uint16_t>(x) == x; | ||||
410 | } | ||||
411 | template <> constexpr inline bool isUInt<32>(uint64_t x) { | ||||
412 | return static_cast<uint32_t>(x) == x; | ||||
413 | } | ||||
414 | |||||
415 | /// Checks if a unsigned integer is an N bit number shifted left by S. | ||||
416 | template <unsigned N, unsigned S> | ||||
417 | constexpr inline bool isShiftedUInt(uint64_t x) { | ||||
418 | static_assert( | ||||
419 | N > 0, "isShiftedUInt<0> doesn't make sense (refers to a 0-bit number)"); | ||||
420 | static_assert(N + S <= 64, | ||||
421 | "isShiftedUInt<N, S> with N + S > 64 is too wide."); | ||||
422 | // Per the two static_asserts above, S must be strictly less than 64. So | ||||
423 | // 1 << S is not undefined behavior. | ||||
424 | return isUInt<N + S>(x) && (x % (UINT64_C(1)1UL << S) == 0); | ||||
425 | } | ||||
426 | |||||
427 | /// Gets the maximum value for a N-bit unsigned integer. | ||||
428 | inline uint64_t maxUIntN(uint64_t N) { | ||||
429 | assert(N > 0 && N <= 64 && "integer width out of range")(static_cast <bool> (N > 0 && N <= 64 && "integer width out of range") ? void (0) : __assert_fail ("N > 0 && N <= 64 && \"integer width out of range\"" , "llvm/include/llvm/Support/MathExtras.h", 429, __extension__ __PRETTY_FUNCTION__)); | ||||
430 | |||||
431 | // uint64_t(1) << 64 is undefined behavior, so we can't do | ||||
432 | // (uint64_t(1) << N) - 1 | ||||
433 | // without checking first that N != 64. But this works and doesn't have a | ||||
434 | // branch. | ||||
435 | return UINT64_MAX(18446744073709551615UL) >> (64 - N); | ||||
436 | } | ||||
437 | |||||
438 | /// Gets the minimum value for a N-bit signed integer. | ||||
439 | inline int64_t minIntN(int64_t N) { | ||||
440 | assert(N > 0 && N <= 64 && "integer width out of range")(static_cast <bool> (N > 0 && N <= 64 && "integer width out of range") ? void (0) : __assert_fail ("N > 0 && N <= 64 && \"integer width out of range\"" , "llvm/include/llvm/Support/MathExtras.h", 440, __extension__ __PRETTY_FUNCTION__)); | ||||
441 | |||||
442 | return UINT64_C(1)1UL + ~(UINT64_C(1)1UL << (N - 1)); | ||||
443 | } | ||||
444 | |||||
445 | /// Gets the maximum value for a N-bit signed integer. | ||||
446 | inline int64_t maxIntN(int64_t N) { | ||||
447 | assert(N > 0 && N <= 64 && "integer width out of range")(static_cast <bool> (N > 0 && N <= 64 && "integer width out of range") ? void (0) : __assert_fail ("N > 0 && N <= 64 && \"integer width out of range\"" , "llvm/include/llvm/Support/MathExtras.h", 447, __extension__ __PRETTY_FUNCTION__)); | ||||
448 | |||||
449 | // This relies on two's complement wraparound when N == 64, so we convert to | ||||
450 | // int64_t only at the very end to avoid UB. | ||||
451 | return (UINT64_C(1)1UL << (N - 1)) - 1; | ||||
452 | } | ||||
453 | |||||
454 | /// Checks if an unsigned integer fits into the given (dynamic) bit width. | ||||
455 | inline bool isUIntN(unsigned N, uint64_t x) { | ||||
456 | return N >= 64 || x <= maxUIntN(N); | ||||
457 | } | ||||
458 | |||||
459 | /// Checks if an signed integer fits into the given (dynamic) bit width. | ||||
460 | inline bool isIntN(unsigned N, int64_t x) { | ||||
461 | return N >= 64 || (minIntN(N) <= x && x <= maxIntN(N)); | ||||
462 | } | ||||
463 | |||||
464 | /// Return true if the argument is a non-empty sequence of ones starting at the | ||||
465 | /// least significant bit with the remainder zero (32 bit version). | ||||
466 | /// Ex. isMask_32(0x0000FFFFU) == true. | ||||
467 | constexpr inline bool isMask_32(uint32_t Value) { | ||||
468 | return Value && ((Value + 1) & Value) == 0; | ||||
469 | } | ||||
470 | |||||
471 | /// Return true if the argument is a non-empty sequence of ones starting at the | ||||
472 | /// least significant bit with the remainder zero (64 bit version). | ||||
473 | constexpr inline bool isMask_64(uint64_t Value) { | ||||
474 | return Value && ((Value + 1) & Value) == 0; | ||||
475 | } | ||||
476 | |||||
477 | /// Return true if the argument contains a non-empty sequence of ones with the | ||||
478 | /// remainder zero (32 bit version.) Ex. isShiftedMask_32(0x0000FF00U) == true. | ||||
479 | constexpr inline bool isShiftedMask_32(uint32_t Value) { | ||||
480 | return Value && isMask_32((Value - 1) | Value); | ||||
481 | } | ||||
482 | |||||
483 | /// Return true if the argument contains a non-empty sequence of ones with the | ||||
484 | /// remainder zero (64 bit version.) | ||||
485 | constexpr inline bool isShiftedMask_64(uint64_t Value) { | ||||
486 | return Value && isMask_64((Value - 1) | Value); | ||||
487 | } | ||||
488 | |||||
489 | /// Return true if the argument is a power of two > 0. | ||||
490 | /// Ex. isPowerOf2_32(0x00100000U) == true (32 bit edition.) | ||||
491 | constexpr inline bool isPowerOf2_32(uint32_t Value) { | ||||
492 | return Value && !(Value & (Value - 1)); | ||||
493 | } | ||||
494 | |||||
495 | /// Return true if the argument is a power of two > 0 (64 bit edition.) | ||||
496 | constexpr inline bool isPowerOf2_64(uint64_t Value) { | ||||
497 | return Value && !(Value & (Value - 1)); | ||||
498 | } | ||||
499 | |||||
500 | /// Count the number of ones from the most significant bit to the first | ||||
501 | /// zero bit. | ||||
502 | /// | ||||
503 | /// Ex. countLeadingOnes(0xFF0FFF00) == 8. | ||||
504 | /// Only unsigned integral types are allowed. | ||||
505 | /// | ||||
506 | /// \param ZB the behavior on an input of all ones. Only ZB_Width and | ||||
507 | /// ZB_Undefined are valid arguments. | ||||
508 | template <typename T> | ||||
509 | unsigned countLeadingOnes(T Value, ZeroBehavior ZB = ZB_Width) { | ||||
510 | static_assert(std::numeric_limits<T>::is_integer && | ||||
511 | !std::numeric_limits<T>::is_signed, | ||||
512 | "Only unsigned integral types are allowed."); | ||||
513 | return countLeadingZeros<T>(~Value, ZB); | ||||
514 | } | ||||
515 | |||||
516 | /// Count the number of ones from the least significant bit to the first | ||||
517 | /// zero bit. | ||||
518 | /// | ||||
519 | /// Ex. countTrailingOnes(0x00FF00FF) == 8. | ||||
520 | /// Only unsigned integral types are allowed. | ||||
521 | /// | ||||
522 | /// \param ZB the behavior on an input of all ones. Only ZB_Width and | ||||
523 | /// ZB_Undefined are valid arguments. | ||||
524 | template <typename T> | ||||
525 | unsigned countTrailingOnes(T Value, ZeroBehavior ZB = ZB_Width) { | ||||
526 | static_assert(std::numeric_limits<T>::is_integer && | ||||
527 | !std::numeric_limits<T>::is_signed, | ||||
528 | "Only unsigned integral types are allowed."); | ||||
529 | return countTrailingZeros<T>(~Value, ZB); | ||||
530 | } | ||||
531 | |||||
532 | namespace detail { | ||||
533 | template <typename T, std::size_t SizeOfT> struct PopulationCounter { | ||||
534 | static unsigned count(T Value) { | ||||
535 | // Generic version, forward to 32 bits. | ||||
536 | static_assert(SizeOfT <= 4, "Not implemented!"); | ||||
537 | #if defined(__GNUC__4) | ||||
538 | return __builtin_popcount(Value); | ||||
539 | #else | ||||
540 | uint32_t v = Value; | ||||
541 | v = v - ((v >> 1) & 0x55555555); | ||||
542 | v = (v & 0x33333333) + ((v >> 2) & 0x33333333); | ||||
543 | return ((v + (v >> 4) & 0xF0F0F0F) * 0x1010101) >> 24; | ||||
544 | #endif | ||||
545 | } | ||||
546 | }; | ||||
547 | |||||
548 | template <typename T> struct PopulationCounter<T, 8> { | ||||
549 | static unsigned count(T Value) { | ||||
550 | #if defined(__GNUC__4) | ||||
551 | return __builtin_popcountll(Value); | ||||
552 | #else | ||||
553 | uint64_t v = Value; | ||||
554 | v = v - ((v >> 1) & 0x5555555555555555ULL); | ||||
555 | v = (v & 0x3333333333333333ULL) + ((v >> 2) & 0x3333333333333333ULL); | ||||
556 | v = (v + (v >> 4)) & 0x0F0F0F0F0F0F0F0FULL; | ||||
557 | return unsigned((uint64_t)(v * 0x0101010101010101ULL) >> 56); | ||||
558 | #endif | ||||
559 | } | ||||
560 | }; | ||||
561 | } // namespace detail | ||||
562 | |||||
563 | /// Count the number of set bits in a value. | ||||
564 | /// Ex. countPopulation(0xF000F000) = 8 | ||||
565 | /// Returns 0 if the word is zero. | ||||
566 | template <typename T> | ||||
567 | inline unsigned countPopulation(T Value) { | ||||
568 | static_assert(std::numeric_limits<T>::is_integer && | ||||
569 | !std::numeric_limits<T>::is_signed, | ||||
570 | "Only unsigned integral types are allowed."); | ||||
571 | return detail::PopulationCounter<T, sizeof(T)>::count(Value); | ||||
572 | } | ||||
573 | |||||
574 | /// Return true if the argument contains a non-empty sequence of ones with the | ||||
575 | /// remainder zero (32 bit version.) Ex. isShiftedMask_32(0x0000FF00U) == true. | ||||
576 | /// If true, \p MaskIdx will specify the index of the lowest set bit and \p | ||||
577 | /// MaskLen is updated to specify the length of the mask, else neither are | ||||
578 | /// updated. | ||||
579 | inline bool isShiftedMask_32(uint32_t Value, unsigned &MaskIdx, | ||||
580 | unsigned &MaskLen) { | ||||
581 | if (!isShiftedMask_32(Value)) | ||||
582 | return false; | ||||
583 | MaskIdx = countTrailingZeros(Value); | ||||
584 | MaskLen = countPopulation(Value); | ||||
585 | return true; | ||||
586 | } | ||||
587 | |||||
588 | /// Return true if the argument contains a non-empty sequence of ones with the | ||||
589 | /// remainder zero (64 bit version.) If true, \p MaskIdx will specify the index | ||||
590 | /// of the lowest set bit and \p MaskLen is updated to specify the length of the | ||||
591 | /// mask, else neither are updated. | ||||
592 | inline bool isShiftedMask_64(uint64_t Value, unsigned &MaskIdx, | ||||
593 | unsigned &MaskLen) { | ||||
594 | if (!isShiftedMask_64(Value)) | ||||
595 | return false; | ||||
596 | MaskIdx = countTrailingZeros(Value); | ||||
597 | MaskLen = countPopulation(Value); | ||||
598 | return true; | ||||
599 | } | ||||
600 | |||||
601 | /// Compile time Log2. | ||||
602 | /// Valid only for positive powers of two. | ||||
603 | template <size_t kValue> constexpr inline size_t CTLog2() { | ||||
604 | static_assert(kValue > 0 && llvm::isPowerOf2_64(kValue), | ||||
605 | "Value is not a valid power of 2"); | ||||
606 | return 1 + CTLog2<kValue / 2>(); | ||||
607 | } | ||||
608 | |||||
609 | template <> constexpr inline size_t CTLog2<1>() { return 0; } | ||||
610 | |||||
611 | /// Return the log base 2 of the specified value. | ||||
612 | inline double Log2(double Value) { | ||||
613 | #if defined(__ANDROID_API__) && __ANDROID_API__ < 18 | ||||
614 | return __builtin_log(Value) / __builtin_log(2.0); | ||||
615 | #else | ||||
616 | return log2(Value); | ||||
617 | #endif | ||||
618 | } | ||||
619 | |||||
620 | /// Return the floor log base 2 of the specified value, -1 if the value is zero. | ||||
621 | /// (32 bit edition.) | ||||
622 | /// Ex. Log2_32(32) == 5, Log2_32(1) == 0, Log2_32(0) == -1, Log2_32(6) == 2 | ||||
623 | inline unsigned Log2_32(uint32_t Value) { | ||||
624 | return 31 - countLeadingZeros(Value); | ||||
625 | } | ||||
626 | |||||
627 | /// Return the floor log base 2 of the specified value, -1 if the value is zero. | ||||
628 | /// (64 bit edition.) | ||||
629 | inline unsigned Log2_64(uint64_t Value) { | ||||
630 | return 63 - countLeadingZeros(Value); | ||||
631 | } | ||||
632 | |||||
633 | /// Return the ceil log base 2 of the specified value, 32 if the value is zero. | ||||
634 | /// (32 bit edition). | ||||
635 | /// Ex. Log2_32_Ceil(32) == 5, Log2_32_Ceil(1) == 0, Log2_32_Ceil(6) == 3 | ||||
636 | inline unsigned Log2_32_Ceil(uint32_t Value) { | ||||
637 | return 32 - countLeadingZeros(Value - 1); | ||||
638 | } | ||||
639 | |||||
640 | /// Return the ceil log base 2 of the specified value, 64 if the value is zero. | ||||
641 | /// (64 bit edition.) | ||||
642 | inline unsigned Log2_64_Ceil(uint64_t Value) { | ||||
643 | return 64 - countLeadingZeros(Value - 1); | ||||
644 | } | ||||
645 | |||||
646 | /// Return the greatest common divisor of the values using Euclid's algorithm. | ||||
647 | template <typename T> | ||||
648 | inline T greatestCommonDivisor(T A, T B) { | ||||
649 | while (B) { | ||||
650 | T Tmp = B; | ||||
651 | B = A % B; | ||||
652 | A = Tmp; | ||||
653 | } | ||||
654 | return A; | ||||
655 | } | ||||
656 | |||||
657 | inline uint64_t GreatestCommonDivisor64(uint64_t A, uint64_t B) { | ||||
658 | return greatestCommonDivisor<uint64_t>(A, B); | ||||
659 | } | ||||
660 | |||||
661 | /// This function takes a 64-bit integer and returns the bit equivalent double. | ||||
662 | inline double BitsToDouble(uint64_t Bits) { | ||||
663 | double D; | ||||
664 | static_assert(sizeof(uint64_t) == sizeof(double), "Unexpected type sizes"); | ||||
665 | memcpy(&D, &Bits, sizeof(Bits)); | ||||
666 | return D; | ||||
667 | } | ||||
668 | |||||
669 | /// This function takes a 32-bit integer and returns the bit equivalent float. | ||||
670 | inline float BitsToFloat(uint32_t Bits) { | ||||
671 | float F; | ||||
672 | static_assert(sizeof(uint32_t) == sizeof(float), "Unexpected type sizes"); | ||||
673 | memcpy(&F, &Bits, sizeof(Bits)); | ||||
674 | return F; | ||||
675 | } | ||||
676 | |||||
677 | /// This function takes a double and returns the bit equivalent 64-bit integer. | ||||
678 | /// Note that copying doubles around changes the bits of NaNs on some hosts, | ||||
679 | /// notably x86, so this routine cannot be used if these bits are needed. | ||||
680 | inline uint64_t DoubleToBits(double Double) { | ||||
681 | uint64_t Bits; | ||||
682 | static_assert(sizeof(uint64_t) == sizeof(double), "Unexpected type sizes"); | ||||
683 | memcpy(&Bits, &Double, sizeof(Double)); | ||||
684 | return Bits; | ||||
685 | } | ||||
686 | |||||
687 | /// This function takes a float and returns the bit equivalent 32-bit integer. | ||||
688 | /// Note that copying floats around changes the bits of NaNs on some hosts, | ||||
689 | /// notably x86, so this routine cannot be used if these bits are needed. | ||||
690 | inline uint32_t FloatToBits(float Float) { | ||||
691 | uint32_t Bits; | ||||
692 | static_assert(sizeof(uint32_t) == sizeof(float), "Unexpected type sizes"); | ||||
693 | memcpy(&Bits, &Float, sizeof(Float)); | ||||
694 | return Bits; | ||||
695 | } | ||||
696 | |||||
697 | /// A and B are either alignments or offsets. Return the minimum alignment that | ||||
698 | /// may be assumed after adding the two together. | ||||
699 | constexpr inline uint64_t MinAlign(uint64_t A, uint64_t B) { | ||||
700 | // The largest power of 2 that divides both A and B. | ||||
701 | // | ||||
702 | // Replace "-Value" by "1+~Value" in the following commented code to avoid | ||||
703 | // MSVC warning C4146 | ||||
704 | // return (A | B) & -(A | B); | ||||
705 | return (A | B) & (1 + ~(A | B)); | ||||
706 | } | ||||
707 | |||||
708 | /// Returns the next power of two (in 64-bits) that is strictly greater than A. | ||||
709 | /// Returns zero on overflow. | ||||
710 | constexpr inline uint64_t NextPowerOf2(uint64_t A) { | ||||
711 | A |= (A >> 1); | ||||
712 | A |= (A >> 2); | ||||
713 | A |= (A >> 4); | ||||
714 | A |= (A >> 8); | ||||
715 | A |= (A >> 16); | ||||
716 | A |= (A >> 32); | ||||
717 | return A + 1; | ||||
718 | } | ||||
719 | |||||
720 | /// Returns the power of two which is less than or equal to the given value. | ||||
721 | /// Essentially, it is a floor operation across the domain of powers of two. | ||||
722 | inline uint64_t PowerOf2Floor(uint64_t A) { | ||||
723 | if (!A) return 0; | ||||
724 | return 1ull << (63 - countLeadingZeros(A, ZB_Undefined)); | ||||
725 | } | ||||
726 | |||||
727 | /// Returns the power of two which is greater than or equal to the given value. | ||||
728 | /// Essentially, it is a ceil operation across the domain of powers of two. | ||||
729 | inline uint64_t PowerOf2Ceil(uint64_t A) { | ||||
730 | if (!A) | ||||
731 | return 0; | ||||
732 | return NextPowerOf2(A - 1); | ||||
733 | } | ||||
734 | |||||
735 | /// Returns the next integer (mod 2**64) that is greater than or equal to | ||||
736 | /// \p Value and is a multiple of \p Align. \p Align must be non-zero. | ||||
737 | /// | ||||
738 | /// If non-zero \p Skew is specified, the return value will be a minimal | ||||
739 | /// integer that is greater than or equal to \p Value and equal to | ||||
740 | /// \p Align * N + \p Skew for some integer N. If \p Skew is larger than | ||||
741 | /// \p Align, its value is adjusted to '\p Skew mod \p Align'. | ||||
742 | /// | ||||
743 | /// Examples: | ||||
744 | /// \code | ||||
745 | /// alignTo(5, 8) = 8 | ||||
746 | /// alignTo(17, 8) = 24 | ||||
747 | /// alignTo(~0LL, 8) = 0 | ||||
748 | /// alignTo(321, 255) = 510 | ||||
749 | /// | ||||
750 | /// alignTo(5, 8, 7) = 7 | ||||
751 | /// alignTo(17, 8, 1) = 17 | ||||
752 | /// alignTo(~0LL, 8, 3) = 3 | ||||
753 | /// alignTo(321, 255, 42) = 552 | ||||
754 | /// \endcode | ||||
755 | inline uint64_t alignTo(uint64_t Value, uint64_t Align, uint64_t Skew = 0) { | ||||
756 | assert(Align != 0u && "Align can't be 0.")(static_cast <bool> (Align != 0u && "Align can't be 0." ) ? void (0) : __assert_fail ("Align != 0u && \"Align can't be 0.\"" , "llvm/include/llvm/Support/MathExtras.h", 756, __extension__ __PRETTY_FUNCTION__)); | ||||
757 | Skew %= Align; | ||||
758 | return (Value + Align - 1 - Skew) / Align * Align + Skew; | ||||
759 | } | ||||
760 | |||||
761 | /// Returns the next integer (mod 2**64) that is greater than or equal to | ||||
762 | /// \p Value and is a multiple of \c Align. \c Align must be non-zero. | ||||
763 | template <uint64_t Align> constexpr inline uint64_t alignTo(uint64_t Value) { | ||||
764 | static_assert(Align != 0u, "Align must be non-zero"); | ||||
765 | return (Value + Align - 1) / Align * Align; | ||||
766 | } | ||||
767 | |||||
768 | /// Returns the integer ceil(Numerator / Denominator). | ||||
769 | inline uint64_t divideCeil(uint64_t Numerator, uint64_t Denominator) { | ||||
770 | return alignTo(Numerator, Denominator) / Denominator; | ||||
771 | } | ||||
772 | |||||
773 | /// Returns the integer nearest(Numerator / Denominator). | ||||
774 | inline uint64_t divideNearest(uint64_t Numerator, uint64_t Denominator) { | ||||
775 | return (Numerator + (Denominator / 2)) / Denominator; | ||||
776 | } | ||||
777 | |||||
778 | /// Returns the largest uint64_t less than or equal to \p Value and is | ||||
779 | /// \p Skew mod \p Align. \p Align must be non-zero | ||||
780 | inline uint64_t alignDown(uint64_t Value, uint64_t Align, uint64_t Skew = 0) { | ||||
781 | assert(Align != 0u && "Align can't be 0.")(static_cast <bool> (Align != 0u && "Align can't be 0." ) ? void (0) : __assert_fail ("Align != 0u && \"Align can't be 0.\"" , "llvm/include/llvm/Support/MathExtras.h", 781, __extension__ __PRETTY_FUNCTION__)); | ||||
782 | Skew %= Align; | ||||
783 | return (Value - Skew) / Align * Align + Skew; | ||||
784 | } | ||||
785 | |||||
786 | /// Sign-extend the number in the bottom B bits of X to a 32-bit integer. | ||||
787 | /// Requires 0 < B <= 32. | ||||
788 | template <unsigned B> constexpr inline int32_t SignExtend32(uint32_t X) { | ||||
789 | static_assert(B > 0, "Bit width can't be 0."); | ||||
790 | static_assert(B <= 32, "Bit width out of range."); | ||||
791 | return int32_t(X << (32 - B)) >> (32 - B); | ||||
792 | } | ||||
793 | |||||
794 | /// Sign-extend the number in the bottom B bits of X to a 32-bit integer. | ||||
795 | /// Requires 0 < B <= 32. | ||||
796 | inline int32_t SignExtend32(uint32_t X, unsigned B) { | ||||
797 | assert(B > 0 && "Bit width can't be 0.")(static_cast <bool> (B > 0 && "Bit width can't be 0." ) ? void (0) : __assert_fail ("B > 0 && \"Bit width can't be 0.\"" , "llvm/include/llvm/Support/MathExtras.h", 797, __extension__ __PRETTY_FUNCTION__)); | ||||
798 | assert(B <= 32 && "Bit width out of range.")(static_cast <bool> (B <= 32 && "Bit width out of range." ) ? void (0) : __assert_fail ("B <= 32 && \"Bit width out of range.\"" , "llvm/include/llvm/Support/MathExtras.h", 798, __extension__ __PRETTY_FUNCTION__)); | ||||
799 | return int32_t(X << (32 - B)) >> (32 - B); | ||||
800 | } | ||||
801 | |||||
802 | /// Sign-extend the number in the bottom B bits of X to a 64-bit integer. | ||||
803 | /// Requires 0 < B <= 64. | ||||
804 | template <unsigned B> constexpr inline int64_t SignExtend64(uint64_t x) { | ||||
805 | static_assert(B > 0, "Bit width can't be 0."); | ||||
806 | static_assert(B <= 64, "Bit width out of range."); | ||||
807 | return int64_t(x << (64 - B)) >> (64 - B); | ||||
808 | } | ||||
809 | |||||
810 | /// Sign-extend the number in the bottom B bits of X to a 64-bit integer. | ||||
811 | /// Requires 0 < B <= 64. | ||||
812 | inline int64_t SignExtend64(uint64_t X, unsigned B) { | ||||
813 | assert(B > 0 && "Bit width can't be 0.")(static_cast <bool> (B > 0 && "Bit width can't be 0." ) ? void (0) : __assert_fail ("B > 0 && \"Bit width can't be 0.\"" , "llvm/include/llvm/Support/MathExtras.h", 813, __extension__ __PRETTY_FUNCTION__)); | ||||
814 | assert(B <= 64 && "Bit width out of range.")(static_cast <bool> (B <= 64 && "Bit width out of range." ) ? void (0) : __assert_fail ("B <= 64 && \"Bit width out of range.\"" , "llvm/include/llvm/Support/MathExtras.h", 814, __extension__ __PRETTY_FUNCTION__)); | ||||
815 | return int64_t(X << (64 - B)) >> (64 - B); | ||||
816 | } | ||||
817 | |||||
818 | /// Subtract two unsigned integers, X and Y, of type T and return the absolute | ||||
819 | /// value of the result. | ||||
820 | template <typename T> | ||||
821 | std::enable_if_t<std::is_unsigned<T>::value, T> AbsoluteDifference(T X, T Y) { | ||||
822 | return X > Y ? (X - Y) : (Y - X); | ||||
823 | } | ||||
824 | |||||
825 | /// Add two unsigned integers, X and Y, of type T. Clamp the result to the | ||||
826 | /// maximum representable value of T on overflow. ResultOverflowed indicates if | ||||
827 | /// the result is larger than the maximum representable value of type T. | ||||
828 | template <typename T> | ||||
829 | std::enable_if_t<std::is_unsigned<T>::value, T> | ||||
830 | SaturatingAdd(T X, T Y, bool *ResultOverflowed = nullptr) { | ||||
831 | bool Dummy; | ||||
832 | bool &Overflowed = ResultOverflowed ? *ResultOverflowed : Dummy; | ||||
833 | // Hacker's Delight, p. 29 | ||||
834 | T Z = X + Y; | ||||
835 | Overflowed = (Z < X || Z < Y); | ||||
836 | if (Overflowed) | ||||
837 | return std::numeric_limits<T>::max(); | ||||
838 | else | ||||
839 | return Z; | ||||
840 | } | ||||
841 | |||||
842 | /// Multiply two unsigned integers, X and Y, of type T. Clamp the result to the | ||||
843 | /// maximum representable value of T on overflow. ResultOverflowed indicates if | ||||
844 | /// the result is larger than the maximum representable value of type T. | ||||
845 | template <typename T> | ||||
846 | std::enable_if_t<std::is_unsigned<T>::value, T> | ||||
847 | SaturatingMultiply(T X, T Y, bool *ResultOverflowed = nullptr) { | ||||
848 | bool Dummy; | ||||
849 | bool &Overflowed = ResultOverflowed ? *ResultOverflowed : Dummy; | ||||
850 | |||||
851 | // Hacker's Delight, p. 30 has a different algorithm, but we don't use that | ||||
852 | // because it fails for uint16_t (where multiplication can have undefined | ||||
853 | // behavior due to promotion to int), and requires a division in addition | ||||
854 | // to the multiplication. | ||||
855 | |||||
856 | Overflowed = false; | ||||
857 | |||||
858 | // Log2(Z) would be either Log2Z or Log2Z + 1. | ||||
859 | // Special case: if X or Y is 0, Log2_64 gives -1, and Log2Z | ||||
860 | // will necessarily be less than Log2Max as desired. | ||||
861 | int Log2Z = Log2_64(X) + Log2_64(Y); | ||||
862 | const T Max = std::numeric_limits<T>::max(); | ||||
863 | int Log2Max = Log2_64(Max); | ||||
864 | if (Log2Z < Log2Max) { | ||||
865 | return X * Y; | ||||
866 | } | ||||
867 | if (Log2Z > Log2Max) { | ||||
868 | Overflowed = true; | ||||
869 | return Max; | ||||
870 | } | ||||
871 | |||||
872 | // We're going to use the top bit, and maybe overflow one | ||||
873 | // bit past it. Multiply all but the bottom bit then add | ||||
874 | // that on at the end. | ||||
875 | T Z = (X >> 1) * Y; | ||||
876 | if (Z & ~(Max >> 1)) { | ||||
877 | Overflowed = true; | ||||
878 | return Max; | ||||
879 | } | ||||
880 | Z <<= 1; | ||||
881 | if (X & 1) | ||||
882 | return SaturatingAdd(Z, Y, ResultOverflowed); | ||||
883 | |||||
884 | return Z; | ||||
885 | } | ||||
886 | |||||
887 | /// Multiply two unsigned integers, X and Y, and add the unsigned integer, A to | ||||
888 | /// the product. Clamp the result to the maximum representable value of T on | ||||
889 | /// overflow. ResultOverflowed indicates if the result is larger than the | ||||
890 | /// maximum representable value of type T. | ||||
891 | template <typename T> | ||||
892 | std::enable_if_t<std::is_unsigned<T>::value, T> | ||||
893 | SaturatingMultiplyAdd(T X, T Y, T A, bool *ResultOverflowed = nullptr) { | ||||
894 | bool Dummy; | ||||
895 | bool &Overflowed = ResultOverflowed ? *ResultOverflowed : Dummy; | ||||
896 | |||||
897 | T Product = SaturatingMultiply(X, Y, &Overflowed); | ||||
898 | if (Overflowed) | ||||
899 | return Product; | ||||
900 | |||||
901 | return SaturatingAdd(A, Product, &Overflowed); | ||||
902 | } | ||||
903 | |||||
904 | /// Use this rather than HUGE_VALF; the latter causes warnings on MSVC. | ||||
905 | extern const float huge_valf; | ||||
906 | |||||
907 | |||||
908 | /// Add two signed integers, computing the two's complement truncated result, | ||||
909 | /// returning true if overflow occurred. | ||||
910 | template <typename T> | ||||
911 | std::enable_if_t<std::is_signed<T>::value, T> AddOverflow(T X, T Y, T &Result) { | ||||
912 | #if __has_builtin(__builtin_add_overflow)1 | ||||
913 | return __builtin_add_overflow(X, Y, &Result); | ||||
914 | #else | ||||
915 | // Perform the unsigned addition. | ||||
916 | using U = std::make_unsigned_t<T>; | ||||
917 | const U UX = static_cast<U>(X); | ||||
918 | const U UY = static_cast<U>(Y); | ||||
919 | const U UResult = UX + UY; | ||||
920 | |||||
921 | // Convert to signed. | ||||
922 | Result = static_cast<T>(UResult); | ||||
923 | |||||
924 | // Adding two positive numbers should result in a positive number. | ||||
925 | if (X > 0 && Y > 0) | ||||
926 | return Result <= 0; | ||||
927 | // Adding two negatives should result in a negative number. | ||||
928 | if (X < 0 && Y < 0) | ||||
929 | return Result >= 0; | ||||
930 | return false; | ||||
931 | #endif | ||||
932 | } | ||||
933 | |||||
934 | /// Subtract two signed integers, computing the two's complement truncated | ||||
935 | /// result, returning true if an overflow ocurred. | ||||
936 | template <typename T> | ||||
937 | std::enable_if_t<std::is_signed<T>::value, T> SubOverflow(T X, T Y, T &Result) { | ||||
938 | #if __has_builtin(__builtin_sub_overflow)1 | ||||
939 | return __builtin_sub_overflow(X, Y, &Result); | ||||
940 | #else | ||||
941 | // Perform the unsigned addition. | ||||
942 | using U = std::make_unsigned_t<T>; | ||||
943 | const U UX = static_cast<U>(X); | ||||
944 | const U UY = static_cast<U>(Y); | ||||
945 | const U UResult = UX - UY; | ||||
946 | |||||
947 | // Convert to signed. | ||||
948 | Result = static_cast<T>(UResult); | ||||
949 | |||||
950 | // Subtracting a positive number from a negative results in a negative number. | ||||
951 | if (X <= 0 && Y > 0) | ||||
952 | return Result >= 0; | ||||
953 | // Subtracting a negative number from a positive results in a positive number. | ||||
954 | if (X >= 0 && Y < 0) | ||||
955 | return Result <= 0; | ||||
956 | return false; | ||||
957 | #endif | ||||
958 | } | ||||
959 | |||||
960 | /// Multiply two signed integers, computing the two's complement truncated | ||||
961 | /// result, returning true if an overflow ocurred. | ||||
962 | template <typename T> | ||||
963 | std::enable_if_t<std::is_signed<T>::value, T> MulOverflow(T X, T Y, T &Result) { | ||||
964 | // Perform the unsigned multiplication on absolute values. | ||||
965 | using U = std::make_unsigned_t<T>; | ||||
966 | const U UX = X < 0 ? (0 - static_cast<U>(X)) : static_cast<U>(X); | ||||
967 | const U UY = Y < 0 ? (0 - static_cast<U>(Y)) : static_cast<U>(Y); | ||||
968 | const U UResult = UX * UY; | ||||
969 | |||||
970 | // Convert to signed. | ||||
971 | const bool IsNegative = (X < 0) ^ (Y < 0); | ||||
972 | Result = IsNegative ? (0 - UResult) : UResult; | ||||
973 | |||||
974 | // If any of the args was 0, result is 0 and no overflow occurs. | ||||
975 | if (UX == 0 || UY == 0) | ||||
976 | return false; | ||||
977 | |||||
978 | // UX and UY are in [1, 2^n], where n is the number of digits. | ||||
979 | // Check how the max allowed absolute value (2^n for negative, 2^(n-1) for | ||||
980 | // positive) divided by an argument compares to the other. | ||||
981 | if (IsNegative) | ||||
982 | return UX > (static_cast<U>(std::numeric_limits<T>::max()) + U(1)) / UY; | ||||
983 | else | ||||
984 | return UX > (static_cast<U>(std::numeric_limits<T>::max())) / UY; | ||||
985 | } | ||||
986 | |||||
987 | } // End llvm namespace | ||||
988 | |||||
989 | #endif |