File: | lib/Target/AMDGPU/SIISelLowering.cpp |
Location: | line 1359, column 5 |
Description: | Value stored to 'BR' is never read |
1 | //===-- SIISelLowering.cpp - SI DAG Lowering Implementation ---------------===// |
2 | // |
3 | // The LLVM Compiler Infrastructure |
4 | // |
5 | // This file is distributed under the University of Illinois Open Source |
6 | // License. See LICENSE.TXT for details. |
7 | // |
8 | //===----------------------------------------------------------------------===// |
9 | // |
10 | /// \file |
11 | /// \brief Custom DAG lowering for SI |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #ifdef _MSC_VER |
16 | // Provide M_PI. |
17 | #define _USE_MATH_DEFINES |
18 | #include <cmath> |
19 | #endif |
20 | |
21 | #include "AMDGPU.h" |
22 | #include "AMDGPUIntrinsicInfo.h" |
23 | #include "AMDGPUSubtarget.h" |
24 | #include "SIISelLowering.h" |
25 | #include "SIInstrInfo.h" |
26 | #include "SIMachineFunctionInfo.h" |
27 | #include "SIRegisterInfo.h" |
28 | #include "llvm/ADT/BitVector.h" |
29 | #include "llvm/ADT/StringSwitch.h" |
30 | #include "llvm/CodeGen/CallingConvLower.h" |
31 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
32 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
33 | #include "llvm/CodeGen/SelectionDAG.h" |
34 | #include "llvm/IR/DiagnosticInfo.h" |
35 | #include "llvm/IR/Function.h" |
36 | |
37 | using namespace llvm; |
38 | |
39 | static unsigned findFirstFreeSGPR(CCState &CCInfo) { |
40 | unsigned NumSGPRs = AMDGPU::SGPR_32RegClass.getNumRegs(); |
41 | for (unsigned Reg = 0; Reg < NumSGPRs; ++Reg) { |
42 | if (!CCInfo.isAllocated(AMDGPU::SGPR0 + Reg)) { |
43 | return AMDGPU::SGPR0 + Reg; |
44 | } |
45 | } |
46 | llvm_unreachable("Cannot allocate sgpr")::llvm::llvm_unreachable_internal("Cannot allocate sgpr", "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 46); |
47 | } |
48 | |
49 | SITargetLowering::SITargetLowering(TargetMachine &TM, |
50 | const AMDGPUSubtarget &STI) |
51 | : AMDGPUTargetLowering(TM, STI) { |
52 | addRegisterClass(MVT::i1, &AMDGPU::VReg_1RegClass); |
53 | addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass); |
54 | |
55 | addRegisterClass(MVT::i32, &AMDGPU::SReg_32RegClass); |
56 | addRegisterClass(MVT::f32, &AMDGPU::VGPR_32RegClass); |
57 | |
58 | addRegisterClass(MVT::f64, &AMDGPU::VReg_64RegClass); |
59 | addRegisterClass(MVT::v2i32, &AMDGPU::SReg_64RegClass); |
60 | addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass); |
61 | |
62 | addRegisterClass(MVT::v2i64, &AMDGPU::SReg_128RegClass); |
63 | addRegisterClass(MVT::v2f64, &AMDGPU::SReg_128RegClass); |
64 | |
65 | addRegisterClass(MVT::v4i32, &AMDGPU::SReg_128RegClass); |
66 | addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass); |
67 | |
68 | addRegisterClass(MVT::v8i32, &AMDGPU::SReg_256RegClass); |
69 | addRegisterClass(MVT::v8f32, &AMDGPU::VReg_256RegClass); |
70 | |
71 | addRegisterClass(MVT::v16i32, &AMDGPU::SReg_512RegClass); |
72 | addRegisterClass(MVT::v16f32, &AMDGPU::VReg_512RegClass); |
73 | |
74 | computeRegisterProperties(STI.getRegisterInfo()); |
75 | |
76 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i32, Expand); |
77 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Expand); |
78 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i32, Expand); |
79 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16f32, Expand); |
80 | |
81 | setOperationAction(ISD::ADD, MVT::i32, Legal); |
82 | setOperationAction(ISD::ADDC, MVT::i32, Legal); |
83 | setOperationAction(ISD::ADDE, MVT::i32, Legal); |
84 | setOperationAction(ISD::SUBC, MVT::i32, Legal); |
85 | setOperationAction(ISD::SUBE, MVT::i32, Legal); |
86 | |
87 | setOperationAction(ISD::FSIN, MVT::f32, Custom); |
88 | setOperationAction(ISD::FCOS, MVT::f32, Custom); |
89 | |
90 | setOperationAction(ISD::FMINNUM, MVT::f64, Legal); |
91 | setOperationAction(ISD::FMAXNUM, MVT::f64, Legal); |
92 | |
93 | // We need to custom lower vector stores from local memory |
94 | setOperationAction(ISD::LOAD, MVT::v4i32, Custom); |
95 | setOperationAction(ISD::LOAD, MVT::v8i32, Custom); |
96 | setOperationAction(ISD::LOAD, MVT::v16i32, Custom); |
97 | |
98 | setOperationAction(ISD::STORE, MVT::v8i32, Custom); |
99 | setOperationAction(ISD::STORE, MVT::v16i32, Custom); |
100 | |
101 | setOperationAction(ISD::STORE, MVT::i1, Custom); |
102 | setOperationAction(ISD::STORE, MVT::v4i32, Custom); |
103 | |
104 | setOperationAction(ISD::SELECT, MVT::i64, Custom); |
105 | setOperationAction(ISD::SELECT, MVT::f64, Promote); |
106 | AddPromotedToType(ISD::SELECT, MVT::f64, MVT::i64); |
107 | |
108 | setOperationAction(ISD::SELECT_CC, MVT::f32, Expand); |
109 | setOperationAction(ISD::SELECT_CC, MVT::i32, Expand); |
110 | setOperationAction(ISD::SELECT_CC, MVT::i64, Expand); |
111 | setOperationAction(ISD::SELECT_CC, MVT::f64, Expand); |
112 | |
113 | setOperationAction(ISD::SETCC, MVT::i1, Promote); |
114 | setOperationAction(ISD::SETCC, MVT::v2i1, Expand); |
115 | setOperationAction(ISD::SETCC, MVT::v4i1, Expand); |
116 | |
117 | setOperationAction(ISD::BSWAP, MVT::i32, Legal); |
118 | setOperationAction(ISD::BITREVERSE, MVT::i32, Legal); |
119 | |
120 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Legal); |
121 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Custom); |
122 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Custom); |
123 | |
124 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Legal); |
125 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom); |
126 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Custom); |
127 | |
128 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal); |
129 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom); |
130 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Custom); |
131 | |
132 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal); |
133 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Custom); |
134 | |
135 | setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); |
136 | setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::f32, Custom); |
137 | setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v16i8, Custom); |
138 | setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v4f32, Custom); |
139 | |
140 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom); |
141 | |
142 | setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom); |
143 | setOperationAction(ISD::BRCOND, MVT::Other, Custom); |
144 | setOperationAction(ISD::BR_CC, MVT::i32, Expand); |
145 | setOperationAction(ISD::BR_CC, MVT::i64, Expand); |
146 | setOperationAction(ISD::BR_CC, MVT::f32, Expand); |
147 | setOperationAction(ISD::BR_CC, MVT::f64, Expand); |
148 | |
149 | // On SI this is s_memtime and s_memrealtime on VI. |
150 | setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal); |
151 | |
152 | for (MVT VT : MVT::integer_valuetypes()) { |
153 | if (VT == MVT::i64) |
154 | continue; |
155 | |
156 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote); |
157 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Legal); |
158 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i16, Legal); |
159 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i32, Expand); |
160 | |
161 | setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote); |
162 | setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i8, Legal); |
163 | setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i16, Legal); |
164 | setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i32, Expand); |
165 | |
166 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote); |
167 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::i8, Legal); |
168 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::i16, Legal); |
169 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::i32, Expand); |
170 | } |
171 | |
172 | for (MVT VT : MVT::integer_vector_valuetypes()) { |
173 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i16, Expand); |
174 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v16i16, Expand); |
175 | } |
176 | |
177 | for (MVT VT : MVT::fp_valuetypes()) |
178 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand); |
179 | |
180 | setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f16, Expand); |
181 | setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f32, Expand); |
182 | |
183 | setTruncStoreAction(MVT::i64, MVT::i32, Expand); |
184 | setTruncStoreAction(MVT::v8i32, MVT::v8i16, Expand); |
185 | setTruncStoreAction(MVT::v16i32, MVT::v16i8, Expand); |
186 | setTruncStoreAction(MVT::v16i32, MVT::v16i16, Expand); |
187 | |
188 | |
189 | setTruncStoreAction(MVT::v2i64, MVT::v2i32, Expand); |
190 | |
191 | setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand); |
192 | setTruncStoreAction(MVT::v2f64, MVT::v2f16, Expand); |
193 | |
194 | setOperationAction(ISD::LOAD, MVT::i1, Custom); |
195 | |
196 | setOperationAction(ISD::LOAD, MVT::v2i64, Promote); |
197 | AddPromotedToType(ISD::LOAD, MVT::v2i64, MVT::v4i32); |
198 | |
199 | setOperationAction(ISD::STORE, MVT::v2i64, Promote); |
200 | AddPromotedToType(ISD::STORE, MVT::v2i64, MVT::v4i32); |
201 | |
202 | setOperationAction(ISD::ConstantPool, MVT::v2i64, Expand); |
203 | |
204 | setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); |
205 | setOperationAction(ISD::GlobalAddress, MVT::i64, Custom); |
206 | setOperationAction(ISD::FrameIndex, MVT::i32, Custom); |
207 | |
208 | // These should use UDIVREM, so set them to expand |
209 | setOperationAction(ISD::UDIV, MVT::i64, Expand); |
210 | setOperationAction(ISD::UREM, MVT::i64, Expand); |
211 | |
212 | setOperationAction(ISD::SELECT_CC, MVT::i1, Expand); |
213 | setOperationAction(ISD::SELECT, MVT::i1, Promote); |
214 | |
215 | setOperationAction(ISD::TRUNCATE, MVT::v2i32, Expand); |
216 | |
217 | |
218 | setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand); |
219 | |
220 | // We only support LOAD/STORE and vector manipulation ops for vectors |
221 | // with > 4 elements. |
222 | for (MVT VT : {MVT::v8i32, MVT::v8f32, MVT::v16i32, MVT::v16f32, MVT::v2i64, MVT::v2f64}) { |
223 | for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) { |
224 | switch(Op) { |
225 | case ISD::LOAD: |
226 | case ISD::STORE: |
227 | case ISD::BUILD_VECTOR: |
228 | case ISD::BITCAST: |
229 | case ISD::EXTRACT_VECTOR_ELT: |
230 | case ISD::INSERT_VECTOR_ELT: |
231 | case ISD::INSERT_SUBVECTOR: |
232 | case ISD::EXTRACT_SUBVECTOR: |
233 | case ISD::SCALAR_TO_VECTOR: |
234 | break; |
235 | case ISD::CONCAT_VECTORS: |
236 | setOperationAction(Op, VT, Custom); |
237 | break; |
238 | default: |
239 | setOperationAction(Op, VT, Expand); |
240 | break; |
241 | } |
242 | } |
243 | } |
244 | |
245 | // Most operations are naturally 32-bit vector operations. We only support |
246 | // load and store of i64 vectors, so promote v2i64 vector operations to v4i32. |
247 | for (MVT Vec64 : { MVT::v2i64, MVT::v2f64 }) { |
248 | setOperationAction(ISD::BUILD_VECTOR, Vec64, Promote); |
249 | AddPromotedToType(ISD::BUILD_VECTOR, Vec64, MVT::v4i32); |
250 | |
251 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, Vec64, Promote); |
252 | AddPromotedToType(ISD::EXTRACT_VECTOR_ELT, Vec64, MVT::v4i32); |
253 | |
254 | setOperationAction(ISD::INSERT_VECTOR_ELT, Vec64, Promote); |
255 | AddPromotedToType(ISD::INSERT_VECTOR_ELT, Vec64, MVT::v4i32); |
256 | |
257 | setOperationAction(ISD::SCALAR_TO_VECTOR, Vec64, Promote); |
258 | AddPromotedToType(ISD::SCALAR_TO_VECTOR, Vec64, MVT::v4i32); |
259 | } |
260 | |
261 | if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) { |
262 | setOperationAction(ISD::FTRUNC, MVT::f64, Legal); |
263 | setOperationAction(ISD::FCEIL, MVT::f64, Legal); |
264 | setOperationAction(ISD::FRINT, MVT::f64, Legal); |
265 | } |
266 | |
267 | setOperationAction(ISD::FFLOOR, MVT::f64, Legal); |
268 | setOperationAction(ISD::FDIV, MVT::f32, Custom); |
269 | setOperationAction(ISD::FDIV, MVT::f64, Custom); |
270 | |
271 | // BUFFER/FLAT_ATOMIC_CMP_SWAP on GCN GPUs needs input marshalling, |
272 | // and output demarshalling |
273 | setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom); |
274 | setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom); |
275 | |
276 | // We can't return success/failure, only the old value, |
277 | // let LLVM add the comparison |
278 | setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i32, Expand); |
279 | setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i64, Expand); |
280 | |
281 | setTargetDAGCombine(ISD::FADD); |
282 | setTargetDAGCombine(ISD::FSUB); |
283 | setTargetDAGCombine(ISD::FMINNUM); |
284 | setTargetDAGCombine(ISD::FMAXNUM); |
285 | setTargetDAGCombine(ISD::SMIN); |
286 | setTargetDAGCombine(ISD::SMAX); |
287 | setTargetDAGCombine(ISD::UMIN); |
288 | setTargetDAGCombine(ISD::UMAX); |
289 | setTargetDAGCombine(ISD::SETCC); |
290 | setTargetDAGCombine(ISD::AND); |
291 | setTargetDAGCombine(ISD::OR); |
292 | setTargetDAGCombine(ISD::UINT_TO_FP); |
293 | setTargetDAGCombine(ISD::FCANONICALIZE); |
294 | |
295 | // All memory operations. Some folding on the pointer operand is done to help |
296 | // matching the constant offsets in the addressing modes. |
297 | setTargetDAGCombine(ISD::LOAD); |
298 | setTargetDAGCombine(ISD::STORE); |
299 | setTargetDAGCombine(ISD::ATOMIC_LOAD); |
300 | setTargetDAGCombine(ISD::ATOMIC_STORE); |
301 | setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP); |
302 | setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS); |
303 | setTargetDAGCombine(ISD::ATOMIC_SWAP); |
304 | setTargetDAGCombine(ISD::ATOMIC_LOAD_ADD); |
305 | setTargetDAGCombine(ISD::ATOMIC_LOAD_SUB); |
306 | setTargetDAGCombine(ISD::ATOMIC_LOAD_AND); |
307 | setTargetDAGCombine(ISD::ATOMIC_LOAD_OR); |
308 | setTargetDAGCombine(ISD::ATOMIC_LOAD_XOR); |
309 | setTargetDAGCombine(ISD::ATOMIC_LOAD_NAND); |
310 | setTargetDAGCombine(ISD::ATOMIC_LOAD_MIN); |
311 | setTargetDAGCombine(ISD::ATOMIC_LOAD_MAX); |
312 | setTargetDAGCombine(ISD::ATOMIC_LOAD_UMIN); |
313 | setTargetDAGCombine(ISD::ATOMIC_LOAD_UMAX); |
314 | |
315 | setSchedulingPreference(Sched::RegPressure); |
316 | } |
317 | |
318 | //===----------------------------------------------------------------------===// |
319 | // TargetLowering queries |
320 | //===----------------------------------------------------------------------===// |
321 | |
322 | bool SITargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info, |
323 | const CallInst &CI, |
324 | unsigned IntrID) const { |
325 | switch (IntrID) { |
326 | case Intrinsic::amdgcn_atomic_inc: |
327 | case Intrinsic::amdgcn_atomic_dec: |
328 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
329 | Info.memVT = MVT::getVT(CI.getType()); |
330 | Info.ptrVal = CI.getOperand(0); |
331 | Info.align = 0; |
332 | Info.vol = false; |
333 | Info.readMem = true; |
334 | Info.writeMem = true; |
335 | return true; |
336 | default: |
337 | return false; |
338 | } |
339 | } |
340 | |
341 | bool SITargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &, |
342 | EVT) const { |
343 | // SI has some legal vector types, but no legal vector operations. Say no |
344 | // shuffles are legal in order to prefer scalarizing some vector operations. |
345 | return false; |
346 | } |
347 | |
348 | bool SITargetLowering::isLegalFlatAddressingMode(const AddrMode &AM) const { |
349 | // Flat instructions do not have offsets, and only have the register |
350 | // address. |
351 | return AM.BaseOffs == 0 && (AM.Scale == 0 || AM.Scale == 1); |
352 | } |
353 | |
354 | bool SITargetLowering::isLegalMUBUFAddressingMode(const AddrMode &AM) const { |
355 | // MUBUF / MTBUF instructions have a 12-bit unsigned byte offset, and |
356 | // additionally can do r + r + i with addr64. 32-bit has more addressing |
357 | // mode options. Depending on the resource constant, it can also do |
358 | // (i64 r0) + (i32 r1) * (i14 i). |
359 | // |
360 | // Private arrays end up using a scratch buffer most of the time, so also |
361 | // assume those use MUBUF instructions. Scratch loads / stores are currently |
362 | // implemented as mubuf instructions with offen bit set, so slightly |
363 | // different than the normal addr64. |
364 | if (!isUInt<12>(AM.BaseOffs)) |
365 | return false; |
366 | |
367 | // FIXME: Since we can split immediate into soffset and immediate offset, |
368 | // would it make sense to allow any immediate? |
369 | |
370 | switch (AM.Scale) { |
371 | case 0: // r + i or just i, depending on HasBaseReg. |
372 | return true; |
373 | case 1: |
374 | return true; // We have r + r or r + i. |
375 | case 2: |
376 | if (AM.HasBaseReg) { |
377 | // Reject 2 * r + r. |
378 | return false; |
379 | } |
380 | |
381 | // Allow 2 * r as r + r |
382 | // Or 2 * r + i is allowed as r + r + i. |
383 | return true; |
384 | default: // Don't allow n * r |
385 | return false; |
386 | } |
387 | } |
388 | |
389 | bool SITargetLowering::isLegalAddressingMode(const DataLayout &DL, |
390 | const AddrMode &AM, Type *Ty, |
391 | unsigned AS) const { |
392 | // No global is ever allowed as a base. |
393 | if (AM.BaseGV) |
394 | return false; |
395 | |
396 | switch (AS) { |
397 | case AMDGPUAS::GLOBAL_ADDRESS: { |
398 | if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) { |
399 | // Assume the we will use FLAT for all global memory accesses |
400 | // on VI. |
401 | // FIXME: This assumption is currently wrong. On VI we still use |
402 | // MUBUF instructions for the r + i addressing mode. As currently |
403 | // implemented, the MUBUF instructions only work on buffer < 4GB. |
404 | // It may be possible to support > 4GB buffers with MUBUF instructions, |
405 | // by setting the stride value in the resource descriptor which would |
406 | // increase the size limit to (stride * 4GB). However, this is risky, |
407 | // because it has never been validated. |
408 | return isLegalFlatAddressingMode(AM); |
409 | } |
410 | |
411 | return isLegalMUBUFAddressingMode(AM); |
412 | } |
413 | case AMDGPUAS::CONSTANT_ADDRESS: { |
414 | // If the offset isn't a multiple of 4, it probably isn't going to be |
415 | // correctly aligned. |
416 | if (AM.BaseOffs % 4 != 0) |
417 | return isLegalMUBUFAddressingMode(AM); |
418 | |
419 | // There are no SMRD extloads, so if we have to do a small type access we |
420 | // will use a MUBUF load. |
421 | // FIXME?: We also need to do this if unaligned, but we don't know the |
422 | // alignment here. |
423 | if (DL.getTypeStoreSize(Ty) < 4) |
424 | return isLegalMUBUFAddressingMode(AM); |
425 | |
426 | if (Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS) { |
427 | // SMRD instructions have an 8-bit, dword offset on SI. |
428 | if (!isUInt<8>(AM.BaseOffs / 4)) |
429 | return false; |
430 | } else if (Subtarget->getGeneration() == AMDGPUSubtarget::SEA_ISLANDS) { |
431 | // On CI+, this can also be a 32-bit literal constant offset. If it fits |
432 | // in 8-bits, it can use a smaller encoding. |
433 | if (!isUInt<32>(AM.BaseOffs / 4)) |
434 | return false; |
435 | } else if (Subtarget->getGeneration() == AMDGPUSubtarget::VOLCANIC_ISLANDS) { |
436 | // On VI, these use the SMEM format and the offset is 20-bit in bytes. |
437 | if (!isUInt<20>(AM.BaseOffs)) |
438 | return false; |
439 | } else |
440 | llvm_unreachable("unhandled generation")::llvm::llvm_unreachable_internal("unhandled generation", "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 440); |
441 | |
442 | if (AM.Scale == 0) // r + i or just i, depending on HasBaseReg. |
443 | return true; |
444 | |
445 | if (AM.Scale == 1 && AM.HasBaseReg) |
446 | return true; |
447 | |
448 | return false; |
449 | } |
450 | |
451 | case AMDGPUAS::PRIVATE_ADDRESS: |
452 | case AMDGPUAS::UNKNOWN_ADDRESS_SPACE: |
453 | return isLegalMUBUFAddressingMode(AM); |
454 | |
455 | case AMDGPUAS::LOCAL_ADDRESS: |
456 | case AMDGPUAS::REGION_ADDRESS: { |
457 | // Basic, single offset DS instructions allow a 16-bit unsigned immediate |
458 | // field. |
459 | // XXX - If doing a 4-byte aligned 8-byte type access, we effectively have |
460 | // an 8-bit dword offset but we don't know the alignment here. |
461 | if (!isUInt<16>(AM.BaseOffs)) |
462 | return false; |
463 | |
464 | if (AM.Scale == 0) // r + i or just i, depending on HasBaseReg. |
465 | return true; |
466 | |
467 | if (AM.Scale == 1 && AM.HasBaseReg) |
468 | return true; |
469 | |
470 | return false; |
471 | } |
472 | case AMDGPUAS::FLAT_ADDRESS: |
473 | return isLegalFlatAddressingMode(AM); |
474 | |
475 | default: |
476 | llvm_unreachable("unhandled address space")::llvm::llvm_unreachable_internal("unhandled address space", "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 476); |
477 | } |
478 | } |
479 | |
480 | bool SITargetLowering::allowsMisalignedMemoryAccesses(EVT VT, |
481 | unsigned AddrSpace, |
482 | unsigned Align, |
483 | bool *IsFast) const { |
484 | if (IsFast) |
485 | *IsFast = false; |
486 | |
487 | // TODO: I think v3i32 should allow unaligned accesses on CI with DS_READ_B96, |
488 | // which isn't a simple VT. |
489 | if (!VT.isSimple() || VT == MVT::Other) |
490 | return false; |
491 | |
492 | // TODO - CI+ supports unaligned memory accesses, but this requires driver |
493 | // support. |
494 | |
495 | // XXX - The only mention I see of this in the ISA manual is for LDS direct |
496 | // reads the "byte address and must be dword aligned". Is it also true for the |
497 | // normal loads and stores? |
498 | if (AddrSpace == AMDGPUAS::LOCAL_ADDRESS) { |
499 | // ds_read/write_b64 require 8-byte alignment, but we can do a 4 byte |
500 | // aligned, 8 byte access in a single operation using ds_read2/write2_b32 |
501 | // with adjacent offsets. |
502 | bool AlignedBy4 = (Align % 4 == 0); |
503 | if (IsFast) |
504 | *IsFast = AlignedBy4; |
505 | return AlignedBy4; |
506 | } |
507 | |
508 | // Smaller than dword value must be aligned. |
509 | // FIXME: This should be allowed on CI+ |
510 | if (VT.bitsLT(MVT::i32)) |
511 | return false; |
512 | |
513 | // 8.1.6 - For Dword or larger reads or writes, the two LSBs of the |
514 | // byte-address are ignored, thus forcing Dword alignment. |
515 | // This applies to private, global, and constant memory. |
516 | if (IsFast) |
517 | *IsFast = true; |
518 | |
519 | return VT.bitsGT(MVT::i32) && Align % 4 == 0; |
520 | } |
521 | |
522 | EVT SITargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign, |
523 | unsigned SrcAlign, bool IsMemset, |
524 | bool ZeroMemset, |
525 | bool MemcpyStrSrc, |
526 | MachineFunction &MF) const { |
527 | // FIXME: Should account for address space here. |
528 | |
529 | // The default fallback uses the private pointer size as a guess for a type to |
530 | // use. Make sure we switch these to 64-bit accesses. |
531 | |
532 | if (Size >= 16 && DstAlign >= 4) // XXX: Should only do for global |
533 | return MVT::v4i32; |
534 | |
535 | if (Size >= 8 && DstAlign >= 4) |
536 | return MVT::v2i32; |
537 | |
538 | // Use the default. |
539 | return MVT::Other; |
540 | } |
541 | |
542 | static bool isFlatGlobalAddrSpace(unsigned AS) { |
543 | return AS == AMDGPUAS::GLOBAL_ADDRESS || |
544 | AS == AMDGPUAS::FLAT_ADDRESS || |
545 | AS == AMDGPUAS::CONSTANT_ADDRESS; |
546 | } |
547 | |
548 | bool SITargetLowering::isNoopAddrSpaceCast(unsigned SrcAS, |
549 | unsigned DestAS) const { |
550 | return isFlatGlobalAddrSpace(SrcAS) && isFlatGlobalAddrSpace(DestAS); |
551 | } |
552 | |
553 | |
554 | bool SITargetLowering::isMemOpUniform(const SDNode *N) const { |
555 | const MemSDNode *MemNode = cast<MemSDNode>(N); |
556 | const Value *Ptr = MemNode->getMemOperand()->getValue(); |
557 | |
558 | // UndefValue means this is a load of a kernel input. These are uniform. |
559 | // Sometimes LDS instructions have constant pointers |
560 | if (isa<UndefValue>(Ptr) || isa<Argument>(Ptr) || isa<Constant>(Ptr) || |
561 | isa<GlobalValue>(Ptr)) |
562 | return true; |
563 | |
564 | const Instruction *I = dyn_cast_or_null<Instruction>(Ptr); |
565 | return I && I->getMetadata("amdgpu.uniform"); |
566 | } |
567 | |
568 | TargetLoweringBase::LegalizeTypeAction |
569 | SITargetLowering::getPreferredVectorAction(EVT VT) const { |
570 | if (VT.getVectorNumElements() != 1 && VT.getScalarType().bitsLE(MVT::i16)) |
571 | return TypeSplitVector; |
572 | |
573 | return TargetLoweringBase::getPreferredVectorAction(VT); |
574 | } |
575 | |
576 | bool SITargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm, |
577 | Type *Ty) const { |
578 | const SIInstrInfo *TII = |
579 | static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo()); |
580 | return TII->isInlineConstant(Imm); |
581 | } |
582 | |
583 | bool SITargetLowering::isTypeDesirableForOp(unsigned Op, EVT VT) const { |
584 | |
585 | // SimplifySetCC uses this function to determine whether or not it should |
586 | // create setcc with i1 operands. We don't have instructions for i1 setcc. |
587 | if (VT == MVT::i1 && Op == ISD::SETCC) |
588 | return false; |
589 | |
590 | return TargetLowering::isTypeDesirableForOp(Op, VT); |
591 | } |
592 | |
593 | SDValue SITargetLowering::LowerParameter(SelectionDAG &DAG, EVT VT, EVT MemVT, |
594 | SDLoc SL, SDValue Chain, |
595 | unsigned Offset, bool Signed) const { |
596 | const DataLayout &DL = DAG.getDataLayout(); |
597 | MachineFunction &MF = DAG.getMachineFunction(); |
598 | const SIRegisterInfo *TRI = |
599 | static_cast<const SIRegisterInfo*>(Subtarget->getRegisterInfo()); |
600 | unsigned InputPtrReg = TRI->getPreloadedValue(MF, SIRegisterInfo::KERNARG_SEGMENT_PTR); |
601 | |
602 | Type *Ty = VT.getTypeForEVT(*DAG.getContext()); |
603 | |
604 | MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo(); |
605 | MVT PtrVT = getPointerTy(DL, AMDGPUAS::CONSTANT_ADDRESS); |
606 | PointerType *PtrTy = PointerType::get(Ty, AMDGPUAS::CONSTANT_ADDRESS); |
607 | SDValue BasePtr = DAG.getCopyFromReg(Chain, SL, |
608 | MRI.getLiveInVirtReg(InputPtrReg), PtrVT); |
609 | SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr, |
610 | DAG.getConstant(Offset, SL, PtrVT)); |
611 | SDValue PtrOffset = DAG.getUNDEF(PtrVT); |
612 | MachinePointerInfo PtrInfo(UndefValue::get(PtrTy)); |
613 | |
614 | unsigned Align = DL.getABITypeAlignment(Ty); |
615 | |
616 | ISD::LoadExtType ExtTy = Signed ? ISD::SEXTLOAD : ISD::ZEXTLOAD; |
617 | if (MemVT.isFloatingPoint()) |
618 | ExtTy = ISD::EXTLOAD; |
619 | |
620 | return DAG.getLoad(ISD::UNINDEXED, ExtTy, |
621 | VT, SL, Chain, Ptr, PtrOffset, PtrInfo, MemVT, |
622 | false, // isVolatile |
623 | true, // isNonTemporal |
624 | true, // isInvariant |
625 | Align); // Alignment |
626 | } |
627 | |
628 | SDValue SITargetLowering::LowerFormalArguments( |
629 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
630 | const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG, |
631 | SmallVectorImpl<SDValue> &InVals) const { |
632 | const SIRegisterInfo *TRI = |
633 | static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo()); |
634 | |
635 | MachineFunction &MF = DAG.getMachineFunction(); |
636 | FunctionType *FType = MF.getFunction()->getFunctionType(); |
637 | SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>(); |
638 | const AMDGPUSubtarget &ST = MF.getSubtarget<AMDGPUSubtarget>(); |
639 | |
640 | if (Subtarget->isAmdHsaOS() && AMDGPU::isShader(CallConv)) { |
641 | const Function *Fn = MF.getFunction(); |
642 | DiagnosticInfoUnsupported NoGraphicsHSA( |
643 | *Fn, "unsupported non-compute shaders with HSA", DL.getDebugLoc()); |
644 | DAG.getContext()->diagnose(NoGraphicsHSA); |
645 | return SDValue(); |
646 | } |
647 | |
648 | SmallVector<ISD::InputArg, 16> Splits; |
649 | BitVector Skipped(Ins.size()); |
650 | |
651 | for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) { |
652 | const ISD::InputArg &Arg = Ins[i]; |
653 | |
654 | // First check if it's a PS input addr |
655 | if (CallConv == CallingConv::AMDGPU_PS && !Arg.Flags.isInReg() && |
656 | !Arg.Flags.isByVal() && PSInputNum <= 15) { |
657 | |
658 | if (!Arg.Used && !Info->isPSInputAllocated(PSInputNum)) { |
659 | // We can safely skip PS inputs |
660 | Skipped.set(i); |
661 | ++PSInputNum; |
662 | continue; |
663 | } |
664 | |
665 | Info->markPSInputAllocated(PSInputNum); |
666 | if (Arg.Used) |
667 | Info->PSInputEna |= 1 << PSInputNum; |
668 | |
669 | ++PSInputNum; |
670 | } |
671 | |
672 | // Second split vertices into their elements |
673 | if (AMDGPU::isShader(CallConv) && |
674 | Arg.VT.isVector()) { |
675 | ISD::InputArg NewArg = Arg; |
676 | NewArg.Flags.setSplit(); |
677 | NewArg.VT = Arg.VT.getVectorElementType(); |
678 | |
679 | // We REALLY want the ORIGINAL number of vertex elements here, e.g. a |
680 | // three or five element vertex only needs three or five registers, |
681 | // NOT four or eight. |
682 | Type *ParamType = FType->getParamType(Arg.getOrigArgIndex()); |
683 | unsigned NumElements = ParamType->getVectorNumElements(); |
684 | |
685 | for (unsigned j = 0; j != NumElements; ++j) { |
686 | Splits.push_back(NewArg); |
687 | NewArg.PartOffset += NewArg.VT.getStoreSize(); |
688 | } |
689 | |
690 | } else if (AMDGPU::isShader(CallConv)) { |
691 | Splits.push_back(Arg); |
692 | } |
693 | } |
694 | |
695 | SmallVector<CCValAssign, 16> ArgLocs; |
696 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs, |
697 | *DAG.getContext()); |
698 | |
699 | // At least one interpolation mode must be enabled or else the GPU will hang. |
700 | // |
701 | // Check PSInputAddr instead of PSInputEna. The idea is that if the user set |
702 | // PSInputAddr, the user wants to enable some bits after the compilation |
703 | // based on run-time states. Since we can't know what the final PSInputEna |
704 | // will look like, so we shouldn't do anything here and the user should take |
705 | // responsibility for the correct programming. |
706 | // |
707 | // Otherwise, the following restrictions apply: |
708 | // - At least one of PERSP_* (0xF) or LINEAR_* (0x70) must be enabled. |
709 | // - If POS_W_FLOAT (11) is enabled, at least one of PERSP_* must be |
710 | // enabled too. |
711 | if (CallConv == CallingConv::AMDGPU_PS && |
712 | ((Info->getPSInputAddr() & 0x7F) == 0 || |
713 | ((Info->getPSInputAddr() & 0xF) == 0 && |
714 | Info->isPSInputAllocated(11)))) { |
715 | CCInfo.AllocateReg(AMDGPU::VGPR0); |
716 | CCInfo.AllocateReg(AMDGPU::VGPR1); |
717 | Info->markPSInputAllocated(0); |
718 | Info->PSInputEna |= 1; |
719 | } |
720 | |
721 | if (!AMDGPU::isShader(CallConv)) { |
722 | getOriginalFunctionArgs(DAG, DAG.getMachineFunction().getFunction(), Ins, |
723 | Splits); |
724 | |
725 | assert(Info->hasWorkGroupIDX() && Info->hasWorkItemIDX())((Info->hasWorkGroupIDX() && Info->hasWorkItemIDX ()) ? static_cast<void> (0) : __assert_fail ("Info->hasWorkGroupIDX() && Info->hasWorkItemIDX()" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 725, __PRETTY_FUNCTION__)); |
726 | } else { |
727 | assert(!Info->hasPrivateSegmentBuffer() && !Info->hasDispatchPtr() &&((!Info->hasPrivateSegmentBuffer() && !Info->hasDispatchPtr () && !Info->hasKernargSegmentPtr() && !Info ->hasFlatScratchInit() && !Info->hasWorkGroupIDX () && !Info->hasWorkGroupIDY() && !Info-> hasWorkGroupIDZ() && !Info->hasWorkGroupInfo() && !Info->hasWorkItemIDX() && !Info->hasWorkItemIDY () && !Info->hasWorkItemIDZ()) ? static_cast<void > (0) : __assert_fail ("!Info->hasPrivateSegmentBuffer() && !Info->hasDispatchPtr() && !Info->hasKernargSegmentPtr() && !Info->hasFlatScratchInit() && !Info->hasWorkGroupIDX() && !Info->hasWorkGroupIDY() && !Info->hasWorkGroupIDZ() && !Info->hasWorkGroupInfo() && !Info->hasWorkItemIDX() && !Info->hasWorkItemIDY() && !Info->hasWorkItemIDZ()" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 732, __PRETTY_FUNCTION__)) |
728 | !Info->hasKernargSegmentPtr() && !Info->hasFlatScratchInit() &&((!Info->hasPrivateSegmentBuffer() && !Info->hasDispatchPtr () && !Info->hasKernargSegmentPtr() && !Info ->hasFlatScratchInit() && !Info->hasWorkGroupIDX () && !Info->hasWorkGroupIDY() && !Info-> hasWorkGroupIDZ() && !Info->hasWorkGroupInfo() && !Info->hasWorkItemIDX() && !Info->hasWorkItemIDY () && !Info->hasWorkItemIDZ()) ? static_cast<void > (0) : __assert_fail ("!Info->hasPrivateSegmentBuffer() && !Info->hasDispatchPtr() && !Info->hasKernargSegmentPtr() && !Info->hasFlatScratchInit() && !Info->hasWorkGroupIDX() && !Info->hasWorkGroupIDY() && !Info->hasWorkGroupIDZ() && !Info->hasWorkGroupInfo() && !Info->hasWorkItemIDX() && !Info->hasWorkItemIDY() && !Info->hasWorkItemIDZ()" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 732, __PRETTY_FUNCTION__)) |
729 | !Info->hasWorkGroupIDX() && !Info->hasWorkGroupIDY() &&((!Info->hasPrivateSegmentBuffer() && !Info->hasDispatchPtr () && !Info->hasKernargSegmentPtr() && !Info ->hasFlatScratchInit() && !Info->hasWorkGroupIDX () && !Info->hasWorkGroupIDY() && !Info-> hasWorkGroupIDZ() && !Info->hasWorkGroupInfo() && !Info->hasWorkItemIDX() && !Info->hasWorkItemIDY () && !Info->hasWorkItemIDZ()) ? static_cast<void > (0) : __assert_fail ("!Info->hasPrivateSegmentBuffer() && !Info->hasDispatchPtr() && !Info->hasKernargSegmentPtr() && !Info->hasFlatScratchInit() && !Info->hasWorkGroupIDX() && !Info->hasWorkGroupIDY() && !Info->hasWorkGroupIDZ() && !Info->hasWorkGroupInfo() && !Info->hasWorkItemIDX() && !Info->hasWorkItemIDY() && !Info->hasWorkItemIDZ()" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 732, __PRETTY_FUNCTION__)) |
730 | !Info->hasWorkGroupIDZ() && !Info->hasWorkGroupInfo() &&((!Info->hasPrivateSegmentBuffer() && !Info->hasDispatchPtr () && !Info->hasKernargSegmentPtr() && !Info ->hasFlatScratchInit() && !Info->hasWorkGroupIDX () && !Info->hasWorkGroupIDY() && !Info-> hasWorkGroupIDZ() && !Info->hasWorkGroupInfo() && !Info->hasWorkItemIDX() && !Info->hasWorkItemIDY () && !Info->hasWorkItemIDZ()) ? static_cast<void > (0) : __assert_fail ("!Info->hasPrivateSegmentBuffer() && !Info->hasDispatchPtr() && !Info->hasKernargSegmentPtr() && !Info->hasFlatScratchInit() && !Info->hasWorkGroupIDX() && !Info->hasWorkGroupIDY() && !Info->hasWorkGroupIDZ() && !Info->hasWorkGroupInfo() && !Info->hasWorkItemIDX() && !Info->hasWorkItemIDY() && !Info->hasWorkItemIDZ()" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 732, __PRETTY_FUNCTION__)) |
731 | !Info->hasWorkItemIDX() && !Info->hasWorkItemIDY() &&((!Info->hasPrivateSegmentBuffer() && !Info->hasDispatchPtr () && !Info->hasKernargSegmentPtr() && !Info ->hasFlatScratchInit() && !Info->hasWorkGroupIDX () && !Info->hasWorkGroupIDY() && !Info-> hasWorkGroupIDZ() && !Info->hasWorkGroupInfo() && !Info->hasWorkItemIDX() && !Info->hasWorkItemIDY () && !Info->hasWorkItemIDZ()) ? static_cast<void > (0) : __assert_fail ("!Info->hasPrivateSegmentBuffer() && !Info->hasDispatchPtr() && !Info->hasKernargSegmentPtr() && !Info->hasFlatScratchInit() && !Info->hasWorkGroupIDX() && !Info->hasWorkGroupIDY() && !Info->hasWorkGroupIDZ() && !Info->hasWorkGroupInfo() && !Info->hasWorkItemIDX() && !Info->hasWorkItemIDY() && !Info->hasWorkItemIDZ()" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 732, __PRETTY_FUNCTION__)) |
732 | !Info->hasWorkItemIDZ())((!Info->hasPrivateSegmentBuffer() && !Info->hasDispatchPtr () && !Info->hasKernargSegmentPtr() && !Info ->hasFlatScratchInit() && !Info->hasWorkGroupIDX () && !Info->hasWorkGroupIDY() && !Info-> hasWorkGroupIDZ() && !Info->hasWorkGroupInfo() && !Info->hasWorkItemIDX() && !Info->hasWorkItemIDY () && !Info->hasWorkItemIDZ()) ? static_cast<void > (0) : __assert_fail ("!Info->hasPrivateSegmentBuffer() && !Info->hasDispatchPtr() && !Info->hasKernargSegmentPtr() && !Info->hasFlatScratchInit() && !Info->hasWorkGroupIDX() && !Info->hasWorkGroupIDY() && !Info->hasWorkGroupIDZ() && !Info->hasWorkGroupInfo() && !Info->hasWorkItemIDX() && !Info->hasWorkItemIDY() && !Info->hasWorkItemIDZ()" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 732, __PRETTY_FUNCTION__)); |
733 | } |
734 | |
735 | // FIXME: How should these inputs interact with inreg / custom SGPR inputs? |
736 | if (Info->hasPrivateSegmentBuffer()) { |
737 | unsigned PrivateSegmentBufferReg = Info->addPrivateSegmentBuffer(*TRI); |
738 | MF.addLiveIn(PrivateSegmentBufferReg, &AMDGPU::SReg_128RegClass); |
739 | CCInfo.AllocateReg(PrivateSegmentBufferReg); |
740 | } |
741 | |
742 | if (Info->hasDispatchPtr()) { |
743 | unsigned DispatchPtrReg = Info->addDispatchPtr(*TRI); |
744 | MF.addLiveIn(DispatchPtrReg, &AMDGPU::SReg_64RegClass); |
745 | CCInfo.AllocateReg(DispatchPtrReg); |
746 | } |
747 | |
748 | if (Info->hasKernargSegmentPtr()) { |
749 | unsigned InputPtrReg = Info->addKernargSegmentPtr(*TRI); |
750 | MF.addLiveIn(InputPtrReg, &AMDGPU::SReg_64RegClass); |
751 | CCInfo.AllocateReg(InputPtrReg); |
752 | } |
753 | |
754 | if (Info->hasFlatScratchInit()) { |
755 | unsigned FlatScratchInitReg = Info->addFlatScratchInit(*TRI); |
756 | MF.addLiveIn(FlatScratchInitReg, &AMDGPU::SReg_64RegClass); |
757 | CCInfo.AllocateReg(FlatScratchInitReg); |
758 | } |
759 | |
760 | AnalyzeFormalArguments(CCInfo, Splits); |
761 | |
762 | SmallVector<SDValue, 16> Chains; |
763 | |
764 | for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) { |
765 | |
766 | const ISD::InputArg &Arg = Ins[i]; |
767 | if (Skipped[i]) { |
768 | InVals.push_back(DAG.getUNDEF(Arg.VT)); |
769 | continue; |
770 | } |
771 | |
772 | CCValAssign &VA = ArgLocs[ArgIdx++]; |
773 | MVT VT = VA.getLocVT(); |
774 | |
775 | if (VA.isMemLoc()) { |
776 | VT = Ins[i].VT; |
777 | EVT MemVT = Splits[i].VT; |
778 | const unsigned Offset = Subtarget->getExplicitKernelArgOffset() + |
779 | VA.getLocMemOffset(); |
780 | // The first 36 bytes of the input buffer contains information about |
781 | // thread group and global sizes. |
782 | SDValue Arg = LowerParameter(DAG, VT, MemVT, DL, Chain, |
783 | Offset, Ins[i].Flags.isSExt()); |
784 | Chains.push_back(Arg.getValue(1)); |
785 | |
786 | auto *ParamTy = |
787 | dyn_cast<PointerType>(FType->getParamType(Ins[i].getOrigArgIndex())); |
788 | if (Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS && |
789 | ParamTy && ParamTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS) { |
790 | // On SI local pointers are just offsets into LDS, so they are always |
791 | // less than 16-bits. On CI and newer they could potentially be |
792 | // real pointers, so we can't guarantee their size. |
793 | Arg = DAG.getNode(ISD::AssertZext, DL, Arg.getValueType(), Arg, |
794 | DAG.getValueType(MVT::i16)); |
795 | } |
796 | |
797 | InVals.push_back(Arg); |
798 | Info->ABIArgOffset = Offset + MemVT.getStoreSize(); |
799 | continue; |
800 | } |
801 | assert(VA.isRegLoc() && "Parameter must be in a register!")((VA.isRegLoc() && "Parameter must be in a register!" ) ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && \"Parameter must be in a register!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 801, __PRETTY_FUNCTION__)); |
802 | |
803 | unsigned Reg = VA.getLocReg(); |
804 | |
805 | if (VT == MVT::i64) { |
806 | // For now assume it is a pointer |
807 | Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0, |
808 | &AMDGPU::SReg_64RegClass); |
809 | Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass); |
810 | SDValue Copy = DAG.getCopyFromReg(Chain, DL, Reg, VT); |
811 | InVals.push_back(Copy); |
812 | continue; |
813 | } |
814 | |
815 | const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT); |
816 | |
817 | Reg = MF.addLiveIn(Reg, RC); |
818 | SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT); |
819 | |
820 | if (Arg.VT.isVector()) { |
821 | |
822 | // Build a vector from the registers |
823 | Type *ParamType = FType->getParamType(Arg.getOrigArgIndex()); |
824 | unsigned NumElements = ParamType->getVectorNumElements(); |
825 | |
826 | SmallVector<SDValue, 4> Regs; |
827 | Regs.push_back(Val); |
828 | for (unsigned j = 1; j != NumElements; ++j) { |
829 | Reg = ArgLocs[ArgIdx++].getLocReg(); |
830 | Reg = MF.addLiveIn(Reg, RC); |
831 | |
832 | SDValue Copy = DAG.getCopyFromReg(Chain, DL, Reg, VT); |
833 | Regs.push_back(Copy); |
834 | } |
835 | |
836 | // Fill up the missing vector elements |
837 | NumElements = Arg.VT.getVectorNumElements() - NumElements; |
838 | Regs.append(NumElements, DAG.getUNDEF(VT)); |
839 | |
840 | InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, DL, Arg.VT, Regs)); |
841 | continue; |
842 | } |
843 | |
844 | InVals.push_back(Val); |
845 | } |
846 | |
847 | // TODO: Add GridWorkGroupCount user SGPRs when used. For now with HSA we read |
848 | // these from the dispatch pointer. |
849 | |
850 | // Start adding system SGPRs. |
851 | if (Info->hasWorkGroupIDX()) { |
852 | unsigned Reg = Info->addWorkGroupIDX(); |
853 | MF.addLiveIn(Reg, &AMDGPU::SReg_32RegClass); |
854 | CCInfo.AllocateReg(Reg); |
855 | } |
856 | |
857 | if (Info->hasWorkGroupIDY()) { |
858 | unsigned Reg = Info->addWorkGroupIDY(); |
859 | MF.addLiveIn(Reg, &AMDGPU::SReg_32RegClass); |
860 | CCInfo.AllocateReg(Reg); |
861 | } |
862 | |
863 | if (Info->hasWorkGroupIDZ()) { |
864 | unsigned Reg = Info->addWorkGroupIDZ(); |
865 | MF.addLiveIn(Reg, &AMDGPU::SReg_32RegClass); |
866 | CCInfo.AllocateReg(Reg); |
867 | } |
868 | |
869 | if (Info->hasWorkGroupInfo()) { |
870 | unsigned Reg = Info->addWorkGroupInfo(); |
871 | MF.addLiveIn(Reg, &AMDGPU::SReg_32RegClass); |
872 | CCInfo.AllocateReg(Reg); |
873 | } |
874 | |
875 | if (Info->hasPrivateSegmentWaveByteOffset()) { |
876 | // Scratch wave offset passed in system SGPR. |
877 | unsigned PrivateSegmentWaveByteOffsetReg; |
878 | |
879 | if (AMDGPU::isShader(CallConv)) { |
880 | PrivateSegmentWaveByteOffsetReg = findFirstFreeSGPR(CCInfo); |
881 | Info->setPrivateSegmentWaveByteOffset(PrivateSegmentWaveByteOffsetReg); |
882 | } else |
883 | PrivateSegmentWaveByteOffsetReg = Info->addPrivateSegmentWaveByteOffset(); |
884 | |
885 | MF.addLiveIn(PrivateSegmentWaveByteOffsetReg, &AMDGPU::SGPR_32RegClass); |
886 | CCInfo.AllocateReg(PrivateSegmentWaveByteOffsetReg); |
887 | } |
888 | |
889 | // Now that we've figured out where the scratch register inputs are, see if |
890 | // should reserve the arguments and use them directly. |
891 | bool HasStackObjects = MF.getFrameInfo()->hasStackObjects(); |
892 | // Record that we know we have non-spill stack objects so we don't need to |
893 | // check all stack objects later. |
894 | if (HasStackObjects) |
895 | Info->setHasNonSpillStackObjects(true); |
896 | |
897 | if (ST.isAmdHsaOS()) { |
898 | // TODO: Assume we will spill without optimizations. |
899 | if (HasStackObjects) { |
900 | // If we have stack objects, we unquestionably need the private buffer |
901 | // resource. For the HSA ABI, this will be the first 4 user SGPR |
902 | // inputs. We can reserve those and use them directly. |
903 | |
904 | unsigned PrivateSegmentBufferReg = TRI->getPreloadedValue( |
905 | MF, SIRegisterInfo::PRIVATE_SEGMENT_BUFFER); |
906 | Info->setScratchRSrcReg(PrivateSegmentBufferReg); |
907 | |
908 | unsigned PrivateSegmentWaveByteOffsetReg = TRI->getPreloadedValue( |
909 | MF, SIRegisterInfo::PRIVATE_SEGMENT_WAVE_BYTE_OFFSET); |
910 | Info->setScratchWaveOffsetReg(PrivateSegmentWaveByteOffsetReg); |
911 | } else { |
912 | unsigned ReservedBufferReg |
913 | = TRI->reservedPrivateSegmentBufferReg(MF); |
914 | unsigned ReservedOffsetReg |
915 | = TRI->reservedPrivateSegmentWaveByteOffsetReg(MF); |
916 | |
917 | // We tentatively reserve the last registers (skipping the last two |
918 | // which may contain VCC). After register allocation, we'll replace |
919 | // these with the ones immediately after those which were really |
920 | // allocated. In the prologue copies will be inserted from the argument |
921 | // to these reserved registers. |
922 | Info->setScratchRSrcReg(ReservedBufferReg); |
923 | Info->setScratchWaveOffsetReg(ReservedOffsetReg); |
924 | } |
925 | } else { |
926 | unsigned ReservedBufferReg = TRI->reservedPrivateSegmentBufferReg(MF); |
927 | |
928 | // Without HSA, relocations are used for the scratch pointer and the |
929 | // buffer resource setup is always inserted in the prologue. Scratch wave |
930 | // offset is still in an input SGPR. |
931 | Info->setScratchRSrcReg(ReservedBufferReg); |
932 | |
933 | if (HasStackObjects) { |
934 | unsigned ScratchWaveOffsetReg = TRI->getPreloadedValue( |
935 | MF, SIRegisterInfo::PRIVATE_SEGMENT_WAVE_BYTE_OFFSET); |
936 | Info->setScratchWaveOffsetReg(ScratchWaveOffsetReg); |
937 | } else { |
938 | unsigned ReservedOffsetReg |
939 | = TRI->reservedPrivateSegmentWaveByteOffsetReg(MF); |
940 | Info->setScratchWaveOffsetReg(ReservedOffsetReg); |
941 | } |
942 | } |
943 | |
944 | if (Info->hasWorkItemIDX()) { |
945 | unsigned Reg = TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_X); |
946 | MF.addLiveIn(Reg, &AMDGPU::VGPR_32RegClass); |
947 | CCInfo.AllocateReg(Reg); |
948 | } |
949 | |
950 | if (Info->hasWorkItemIDY()) { |
951 | unsigned Reg = TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Y); |
952 | MF.addLiveIn(Reg, &AMDGPU::VGPR_32RegClass); |
953 | CCInfo.AllocateReg(Reg); |
954 | } |
955 | |
956 | if (Info->hasWorkItemIDZ()) { |
957 | unsigned Reg = TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Z); |
958 | MF.addLiveIn(Reg, &AMDGPU::VGPR_32RegClass); |
959 | CCInfo.AllocateReg(Reg); |
960 | } |
961 | |
962 | if (Chains.empty()) |
963 | return Chain; |
964 | |
965 | return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains); |
966 | } |
967 | |
968 | SDValue SITargetLowering::LowerReturn(SDValue Chain, |
969 | CallingConv::ID CallConv, |
970 | bool isVarArg, |
971 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
972 | const SmallVectorImpl<SDValue> &OutVals, |
973 | SDLoc DL, SelectionDAG &DAG) const { |
974 | MachineFunction &MF = DAG.getMachineFunction(); |
975 | SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>(); |
976 | |
977 | if (!AMDGPU::isShader(CallConv)) |
978 | return AMDGPUTargetLowering::LowerReturn(Chain, CallConv, isVarArg, Outs, |
979 | OutVals, DL, DAG); |
980 | |
981 | Info->setIfReturnsVoid(Outs.size() == 0); |
982 | |
983 | SmallVector<ISD::OutputArg, 48> Splits; |
984 | SmallVector<SDValue, 48> SplitVals; |
985 | |
986 | // Split vectors into their elements. |
987 | for (unsigned i = 0, e = Outs.size(); i != e; ++i) { |
988 | const ISD::OutputArg &Out = Outs[i]; |
989 | |
990 | if (Out.VT.isVector()) { |
991 | MVT VT = Out.VT.getVectorElementType(); |
992 | ISD::OutputArg NewOut = Out; |
993 | NewOut.Flags.setSplit(); |
994 | NewOut.VT = VT; |
995 | |
996 | // We want the original number of vector elements here, e.g. |
997 | // three or five, not four or eight. |
998 | unsigned NumElements = Out.ArgVT.getVectorNumElements(); |
999 | |
1000 | for (unsigned j = 0; j != NumElements; ++j) { |
1001 | SDValue Elem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, OutVals[i], |
1002 | DAG.getConstant(j, DL, MVT::i32)); |
1003 | SplitVals.push_back(Elem); |
1004 | Splits.push_back(NewOut); |
1005 | NewOut.PartOffset += NewOut.VT.getStoreSize(); |
1006 | } |
1007 | } else { |
1008 | SplitVals.push_back(OutVals[i]); |
1009 | Splits.push_back(Out); |
1010 | } |
1011 | } |
1012 | |
1013 | // CCValAssign - represent the assignment of the return value to a location. |
1014 | SmallVector<CCValAssign, 48> RVLocs; |
1015 | |
1016 | // CCState - Info about the registers and stack slots. |
1017 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, |
1018 | *DAG.getContext()); |
1019 | |
1020 | // Analyze outgoing return values. |
1021 | AnalyzeReturn(CCInfo, Splits); |
1022 | |
1023 | SDValue Flag; |
1024 | SmallVector<SDValue, 48> RetOps; |
1025 | RetOps.push_back(Chain); // Operand #0 = Chain (updated below) |
1026 | |
1027 | // Copy the result values into the output registers. |
1028 | for (unsigned i = 0, realRVLocIdx = 0; |
1029 | i != RVLocs.size(); |
1030 | ++i, ++realRVLocIdx) { |
1031 | CCValAssign &VA = RVLocs[i]; |
1032 | assert(VA.isRegLoc() && "Can only return in registers!")((VA.isRegLoc() && "Can only return in registers!") ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && \"Can only return in registers!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 1032, __PRETTY_FUNCTION__)); |
1033 | |
1034 | SDValue Arg = SplitVals[realRVLocIdx]; |
1035 | |
1036 | // Copied from other backends. |
1037 | switch (VA.getLocInfo()) { |
1038 | default: llvm_unreachable("Unknown loc info!")::llvm::llvm_unreachable_internal("Unknown loc info!", "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 1038); |
1039 | case CCValAssign::Full: |
1040 | break; |
1041 | case CCValAssign::BCvt: |
1042 | Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg); |
1043 | break; |
1044 | } |
1045 | |
1046 | Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag); |
1047 | Flag = Chain.getValue(1); |
1048 | RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); |
1049 | } |
1050 | |
1051 | // Update chain and glue. |
1052 | RetOps[0] = Chain; |
1053 | if (Flag.getNode()) |
1054 | RetOps.push_back(Flag); |
1055 | |
1056 | return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, RetOps); |
1057 | } |
1058 | |
1059 | unsigned SITargetLowering::getRegisterByName(const char* RegName, EVT VT, |
1060 | SelectionDAG &DAG) const { |
1061 | unsigned Reg = StringSwitch<unsigned>(RegName) |
1062 | .Case("m0", AMDGPU::M0) |
1063 | .Case("exec", AMDGPU::EXEC) |
1064 | .Case("exec_lo", AMDGPU::EXEC_LO) |
1065 | .Case("exec_hi", AMDGPU::EXEC_HI) |
1066 | .Case("flat_scratch", AMDGPU::FLAT_SCR) |
1067 | .Case("flat_scratch_lo", AMDGPU::FLAT_SCR_LO) |
1068 | .Case("flat_scratch_hi", AMDGPU::FLAT_SCR_HI) |
1069 | .Default(AMDGPU::NoRegister); |
1070 | |
1071 | if (Reg == AMDGPU::NoRegister) { |
1072 | report_fatal_error(Twine("invalid register name \"" |
1073 | + StringRef(RegName) + "\".")); |
1074 | |
1075 | } |
1076 | |
1077 | if (Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS && |
1078 | Subtarget->getRegisterInfo()->regsOverlap(Reg, AMDGPU::FLAT_SCR)) { |
1079 | report_fatal_error(Twine("invalid register \"" |
1080 | + StringRef(RegName) + "\" for subtarget.")); |
1081 | } |
1082 | |
1083 | switch (Reg) { |
1084 | case AMDGPU::M0: |
1085 | case AMDGPU::EXEC_LO: |
1086 | case AMDGPU::EXEC_HI: |
1087 | case AMDGPU::FLAT_SCR_LO: |
1088 | case AMDGPU::FLAT_SCR_HI: |
1089 | if (VT.getSizeInBits() == 32) |
1090 | return Reg; |
1091 | break; |
1092 | case AMDGPU::EXEC: |
1093 | case AMDGPU::FLAT_SCR: |
1094 | if (VT.getSizeInBits() == 64) |
1095 | return Reg; |
1096 | break; |
1097 | default: |
1098 | llvm_unreachable("missing register type checking")::llvm::llvm_unreachable_internal("missing register type checking" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 1098); |
1099 | } |
1100 | |
1101 | report_fatal_error(Twine("invalid type for register \"" |
1102 | + StringRef(RegName) + "\".")); |
1103 | } |
1104 | |
1105 | MachineBasicBlock *SITargetLowering::EmitInstrWithCustomInserter( |
1106 | MachineInstr *MI, MachineBasicBlock *BB) const { |
1107 | switch (MI->getOpcode()) { |
1108 | case AMDGPU::SI_INIT_M0: { |
1109 | const SIInstrInfo *TII = |
1110 | static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo()); |
1111 | BuildMI(*BB, MI->getIterator(), MI->getDebugLoc(), |
1112 | TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0) |
1113 | .addOperand(MI->getOperand(0)); |
1114 | MI->eraseFromParent(); |
1115 | break; |
1116 | } |
1117 | case AMDGPU::BRANCH: |
1118 | return BB; |
1119 | case AMDGPU::GET_GROUPSTATICSIZE: { |
1120 | const SIInstrInfo *TII = |
1121 | static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo()); |
1122 | MachineFunction *MF = BB->getParent(); |
1123 | SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>(); |
1124 | DebugLoc DL = MI->getDebugLoc(); |
1125 | BuildMI (*BB, MI, DL, TII->get(AMDGPU::S_MOVK_I32)) |
1126 | .addOperand(MI->getOperand(0)) |
1127 | .addImm(MFI->LDSSize); |
1128 | MI->eraseFromParent(); |
1129 | return BB; |
1130 | } |
1131 | default: |
1132 | return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB); |
1133 | } |
1134 | return BB; |
1135 | } |
1136 | |
1137 | bool SITargetLowering::enableAggressiveFMAFusion(EVT VT) const { |
1138 | // This currently forces unfolding various combinations of fsub into fma with |
1139 | // free fneg'd operands. As long as we have fast FMA (controlled by |
1140 | // isFMAFasterThanFMulAndFAdd), we should perform these. |
1141 | |
1142 | // When fma is quarter rate, for f64 where add / sub are at best half rate, |
1143 | // most of these combines appear to be cycle neutral but save on instruction |
1144 | // count / code size. |
1145 | return true; |
1146 | } |
1147 | |
1148 | EVT SITargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &Ctx, |
1149 | EVT VT) const { |
1150 | if (!VT.isVector()) { |
1151 | return MVT::i1; |
1152 | } |
1153 | return EVT::getVectorVT(Ctx, MVT::i1, VT.getVectorNumElements()); |
1154 | } |
1155 | |
1156 | MVT SITargetLowering::getScalarShiftAmountTy(const DataLayout &, EVT) const { |
1157 | return MVT::i32; |
1158 | } |
1159 | |
1160 | // Answering this is somewhat tricky and depends on the specific device which |
1161 | // have different rates for fma or all f64 operations. |
1162 | // |
1163 | // v_fma_f64 and v_mul_f64 always take the same number of cycles as each other |
1164 | // regardless of which device (although the number of cycles differs between |
1165 | // devices), so it is always profitable for f64. |
1166 | // |
1167 | // v_fma_f32 takes 4 or 16 cycles depending on the device, so it is profitable |
1168 | // only on full rate devices. Normally, we should prefer selecting v_mad_f32 |
1169 | // which we can always do even without fused FP ops since it returns the same |
1170 | // result as the separate operations and since it is always full |
1171 | // rate. Therefore, we lie and report that it is not faster for f32. v_mad_f32 |
1172 | // however does not support denormals, so we do report fma as faster if we have |
1173 | // a fast fma device and require denormals. |
1174 | // |
1175 | bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const { |
1176 | VT = VT.getScalarType(); |
1177 | |
1178 | if (!VT.isSimple()) |
1179 | return false; |
1180 | |
1181 | switch (VT.getSimpleVT().SimpleTy) { |
1182 | case MVT::f32: |
1183 | // This is as fast on some subtargets. However, we always have full rate f32 |
1184 | // mad available which returns the same result as the separate operations |
1185 | // which we should prefer over fma. We can't use this if we want to support |
1186 | // denormals, so only report this in these cases. |
1187 | return Subtarget->hasFP32Denormals() && Subtarget->hasFastFMAF32(); |
1188 | case MVT::f64: |
1189 | return true; |
1190 | default: |
1191 | break; |
1192 | } |
1193 | |
1194 | return false; |
1195 | } |
1196 | |
1197 | //===----------------------------------------------------------------------===// |
1198 | // Custom DAG Lowering Operations |
1199 | //===----------------------------------------------------------------------===// |
1200 | |
1201 | SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { |
1202 | switch (Op.getOpcode()) { |
1203 | default: return AMDGPUTargetLowering::LowerOperation(Op, DAG); |
1204 | case ISD::FrameIndex: return LowerFrameIndex(Op, DAG); |
1205 | case ISD::BRCOND: return LowerBRCOND(Op, DAG); |
1206 | case ISD::LOAD: { |
1207 | SDValue Result = LowerLOAD(Op, DAG); |
1208 | assert((!Result.getNode() ||(((!Result.getNode() || Result.getNode()->getNumValues() == 2) && "Load should return a value and a chain") ? static_cast <void> (0) : __assert_fail ("(!Result.getNode() || Result.getNode()->getNumValues() == 2) && \"Load should return a value and a chain\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 1210, __PRETTY_FUNCTION__)) |
1209 | Result.getNode()->getNumValues() == 2) &&(((!Result.getNode() || Result.getNode()->getNumValues() == 2) && "Load should return a value and a chain") ? static_cast <void> (0) : __assert_fail ("(!Result.getNode() || Result.getNode()->getNumValues() == 2) && \"Load should return a value and a chain\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 1210, __PRETTY_FUNCTION__)) |
1210 | "Load should return a value and a chain")(((!Result.getNode() || Result.getNode()->getNumValues() == 2) && "Load should return a value and a chain") ? static_cast <void> (0) : __assert_fail ("(!Result.getNode() || Result.getNode()->getNumValues() == 2) && \"Load should return a value and a chain\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 1210, __PRETTY_FUNCTION__)); |
1211 | return Result; |
1212 | } |
1213 | |
1214 | case ISD::FSIN: |
1215 | case ISD::FCOS: |
1216 | return LowerTrig(Op, DAG); |
1217 | case ISD::SELECT: return LowerSELECT(Op, DAG); |
1218 | case ISD::FDIV: return LowerFDIV(Op, DAG); |
1219 | case ISD::ATOMIC_CMP_SWAP: return LowerATOMIC_CMP_SWAP(Op, DAG); |
1220 | case ISD::STORE: return LowerSTORE(Op, DAG); |
1221 | case ISD::GlobalAddress: { |
1222 | MachineFunction &MF = DAG.getMachineFunction(); |
1223 | SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
1224 | return LowerGlobalAddress(MFI, Op, DAG); |
1225 | } |
1226 | case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG); |
1227 | case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG); |
1228 | case ISD::INTRINSIC_VOID: return LowerINTRINSIC_VOID(Op, DAG); |
1229 | } |
1230 | return SDValue(); |
1231 | } |
1232 | |
1233 | /// \brief Helper function for LowerBRCOND |
1234 | static SDNode *findUser(SDValue Value, unsigned Opcode) { |
1235 | |
1236 | SDNode *Parent = Value.getNode(); |
1237 | for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end(); |
1238 | I != E; ++I) { |
1239 | |
1240 | if (I.getUse().get() != Value) |
1241 | continue; |
1242 | |
1243 | if (I->getOpcode() == Opcode) |
1244 | return *I; |
1245 | } |
1246 | return nullptr; |
1247 | } |
1248 | |
1249 | SDValue SITargetLowering::LowerFrameIndex(SDValue Op, SelectionDAG &DAG) const { |
1250 | |
1251 | SDLoc SL(Op); |
1252 | FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Op); |
1253 | unsigned FrameIndex = FINode->getIndex(); |
1254 | |
1255 | // A FrameIndex node represents a 32-bit offset into scratch memory. If the |
1256 | // high bit of a frame index offset were to be set, this would mean that it |
1257 | // represented an offset of ~2GB * 64 = ~128GB from the start of the scratch |
1258 | // buffer, with 64 being the number of threads per wave. |
1259 | // |
1260 | // The maximum private allocation for the entire GPU is 4G, and we are |
1261 | // concerned with the largest the index could ever be for an individual |
1262 | // workitem. This will occur with the minmum dispatch size. If a program |
1263 | // requires more, the dispatch size will be reduced. |
1264 | // |
1265 | // With this limit, we can mark the high bit of the FrameIndex node as known |
1266 | // zero, which is important, because it means in most situations we can prove |
1267 | // that values derived from FrameIndex nodes are non-negative. This enables us |
1268 | // to take advantage of more addressing modes when accessing scratch buffers, |
1269 | // since for scratch reads/writes, the register offset must always be |
1270 | // positive. |
1271 | |
1272 | uint64_t MaxGPUAlloc = UINT64_C(4)4UL * 1024 * 1024 * 1024; |
1273 | |
1274 | // XXX - It is unclear if partial dispatch works. Assume it works at half wave |
1275 | // granularity. It is probably a full wave. |
1276 | uint64_t MinGranularity = 32; |
1277 | |
1278 | unsigned KnownBits = Log2_64(MaxGPUAlloc / MinGranularity); |
1279 | EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), KnownBits); |
1280 | |
1281 | SDValue TFI = DAG.getTargetFrameIndex(FrameIndex, MVT::i32); |
1282 | return DAG.getNode(ISD::AssertZext, SL, MVT::i32, TFI, |
1283 | DAG.getValueType(ExtVT)); |
1284 | } |
1285 | |
1286 | bool SITargetLowering::isCFIntrinsic(const SDNode *Intr) const { |
1287 | if (Intr->getOpcode() != ISD::INTRINSIC_W_CHAIN) |
1288 | return false; |
1289 | |
1290 | switch (cast<ConstantSDNode>(Intr->getOperand(1))->getZExtValue()) { |
1291 | default: return false; |
1292 | case AMDGPUIntrinsic::amdgcn_if: |
1293 | case AMDGPUIntrinsic::amdgcn_else: |
1294 | case AMDGPUIntrinsic::amdgcn_break: |
1295 | case AMDGPUIntrinsic::amdgcn_if_break: |
1296 | case AMDGPUIntrinsic::amdgcn_else_break: |
1297 | case AMDGPUIntrinsic::amdgcn_loop: |
1298 | case AMDGPUIntrinsic::amdgcn_end_cf: |
1299 | return true; |
1300 | } |
1301 | } |
1302 | |
1303 | /// This transforms the control flow intrinsics to get the branch destination as |
1304 | /// last parameter, also switches branch target with BR if the need arise |
1305 | SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND, |
1306 | SelectionDAG &DAG) const { |
1307 | |
1308 | SDLoc DL(BRCOND); |
1309 | |
1310 | SDNode *Intr = BRCOND.getOperand(1).getNode(); |
1311 | SDValue Target = BRCOND.getOperand(2); |
1312 | SDNode *BR = nullptr; |
1313 | SDNode *SetCC = nullptr; |
1314 | |
1315 | if (Intr->getOpcode() == ISD::SETCC) { |
1316 | // As long as we negate the condition everything is fine |
1317 | SetCC = Intr; |
1318 | Intr = SetCC->getOperand(0).getNode(); |
1319 | |
1320 | } else { |
1321 | // Get the target from BR if we don't negate the condition |
1322 | BR = findUser(BRCOND, ISD::BR); |
1323 | Target = BR->getOperand(1); |
1324 | } |
1325 | |
1326 | if (Intr->getOpcode() != ISD::INTRINSIC_W_CHAIN) { |
1327 | // This is a uniform branch so we don't need to legalize. |
1328 | return BRCOND; |
1329 | } |
1330 | |
1331 | assert(!SetCC ||((!SetCC || (SetCC->getConstantOperandVal(1) == 1 && isCFIntrinsic(Intr) && cast<CondCodeSDNode>(SetCC ->getOperand(2).getNode())->get() == ISD::SETNE)) ? static_cast <void> (0) : __assert_fail ("!SetCC || (SetCC->getConstantOperandVal(1) == 1 && isCFIntrinsic(Intr) && cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() == ISD::SETNE)" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 1335, __PRETTY_FUNCTION__)) |
1332 | (SetCC->getConstantOperandVal(1) == 1 &&((!SetCC || (SetCC->getConstantOperandVal(1) == 1 && isCFIntrinsic(Intr) && cast<CondCodeSDNode>(SetCC ->getOperand(2).getNode())->get() == ISD::SETNE)) ? static_cast <void> (0) : __assert_fail ("!SetCC || (SetCC->getConstantOperandVal(1) == 1 && isCFIntrinsic(Intr) && cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() == ISD::SETNE)" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 1335, __PRETTY_FUNCTION__)) |
1333 | isCFIntrinsic(Intr) &&((!SetCC || (SetCC->getConstantOperandVal(1) == 1 && isCFIntrinsic(Intr) && cast<CondCodeSDNode>(SetCC ->getOperand(2).getNode())->get() == ISD::SETNE)) ? static_cast <void> (0) : __assert_fail ("!SetCC || (SetCC->getConstantOperandVal(1) == 1 && isCFIntrinsic(Intr) && cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() == ISD::SETNE)" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 1335, __PRETTY_FUNCTION__)) |
1334 | cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==((!SetCC || (SetCC->getConstantOperandVal(1) == 1 && isCFIntrinsic(Intr) && cast<CondCodeSDNode>(SetCC ->getOperand(2).getNode())->get() == ISD::SETNE)) ? static_cast <void> (0) : __assert_fail ("!SetCC || (SetCC->getConstantOperandVal(1) == 1 && isCFIntrinsic(Intr) && cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() == ISD::SETNE)" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 1335, __PRETTY_FUNCTION__)) |
1335 | ISD::SETNE))((!SetCC || (SetCC->getConstantOperandVal(1) == 1 && isCFIntrinsic(Intr) && cast<CondCodeSDNode>(SetCC ->getOperand(2).getNode())->get() == ISD::SETNE)) ? static_cast <void> (0) : __assert_fail ("!SetCC || (SetCC->getConstantOperandVal(1) == 1 && isCFIntrinsic(Intr) && cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() == ISD::SETNE)" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 1335, __PRETTY_FUNCTION__)); |
1336 | |
1337 | // Build the result and |
1338 | ArrayRef<EVT> Res(Intr->value_begin() + 1, Intr->value_end()); |
1339 | |
1340 | // operands of the new intrinsic call |
1341 | SmallVector<SDValue, 4> Ops; |
1342 | Ops.push_back(BRCOND.getOperand(0)); |
1343 | Ops.append(Intr->op_begin() + 1, Intr->op_end()); |
1344 | Ops.push_back(Target); |
1345 | |
1346 | // build the new intrinsic call |
1347 | SDNode *Result = DAG.getNode( |
1348 | Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL, |
1349 | DAG.getVTList(Res), Ops).getNode(); |
1350 | |
1351 | if (BR) { |
1352 | // Give the branch instruction our target |
1353 | SDValue Ops[] = { |
1354 | BR->getOperand(0), |
1355 | BRCOND.getOperand(2) |
1356 | }; |
1357 | SDValue NewBR = DAG.getNode(ISD::BR, DL, BR->getVTList(), Ops); |
1358 | DAG.ReplaceAllUsesWith(BR, NewBR.getNode()); |
1359 | BR = NewBR.getNode(); |
Value stored to 'BR' is never read | |
1360 | } |
1361 | |
1362 | SDValue Chain = SDValue(Result, Result->getNumValues() - 1); |
1363 | |
1364 | // Copy the intrinsic results to registers |
1365 | for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) { |
1366 | SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg); |
1367 | if (!CopyToReg) |
1368 | continue; |
1369 | |
1370 | Chain = DAG.getCopyToReg( |
1371 | Chain, DL, |
1372 | CopyToReg->getOperand(1), |
1373 | SDValue(Result, i - 1), |
1374 | SDValue()); |
1375 | |
1376 | DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0)); |
1377 | } |
1378 | |
1379 | // Remove the old intrinsic from the chain |
1380 | DAG.ReplaceAllUsesOfValueWith( |
1381 | SDValue(Intr, Intr->getNumValues() - 1), |
1382 | Intr->getOperand(0)); |
1383 | |
1384 | return Chain; |
1385 | } |
1386 | |
1387 | SDValue SITargetLowering::LowerGlobalAddress(AMDGPUMachineFunction *MFI, |
1388 | SDValue Op, |
1389 | SelectionDAG &DAG) const { |
1390 | GlobalAddressSDNode *GSD = cast<GlobalAddressSDNode>(Op); |
1391 | |
1392 | if (GSD->getAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS) |
1393 | return AMDGPUTargetLowering::LowerGlobalAddress(MFI, Op, DAG); |
1394 | |
1395 | SDLoc DL(GSD); |
1396 | const GlobalValue *GV = GSD->getGlobal(); |
1397 | MVT PtrVT = getPointerTy(DAG.getDataLayout(), GSD->getAddressSpace()); |
1398 | |
1399 | SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i32); |
1400 | return DAG.getNode(AMDGPUISD::CONST_DATA_PTR, DL, PtrVT, GA); |
1401 | } |
1402 | |
1403 | SDValue SITargetLowering::copyToM0(SelectionDAG &DAG, SDValue Chain, SDLoc DL, |
1404 | SDValue V) const { |
1405 | // We can't use S_MOV_B32 directly, because there is no way to specify m0 as |
1406 | // the destination register. |
1407 | // |
1408 | // We can't use CopyToReg, because MachineCSE won't combine COPY instructions, |
1409 | // so we will end up with redundant moves to m0. |
1410 | // |
1411 | // We use a pseudo to ensure we emit s_mov_b32 with m0 as the direct result. |
1412 | |
1413 | // A Null SDValue creates a glue result. |
1414 | SDNode *M0 = DAG.getMachineNode(AMDGPU::SI_INIT_M0, DL, MVT::Other, MVT::Glue, |
1415 | V, Chain); |
1416 | return SDValue(M0, 0); |
1417 | } |
1418 | |
1419 | SDValue SITargetLowering::lowerImplicitZextParam(SelectionDAG &DAG, |
1420 | SDValue Op, |
1421 | MVT VT, |
1422 | unsigned Offset) const { |
1423 | SDLoc SL(Op); |
1424 | SDValue Param = LowerParameter(DAG, MVT::i32, MVT::i32, SL, |
1425 | DAG.getEntryNode(), Offset, false); |
1426 | // The local size values will have the hi 16-bits as zero. |
1427 | return DAG.getNode(ISD::AssertZext, SL, MVT::i32, Param, |
1428 | DAG.getValueType(VT)); |
1429 | } |
1430 | |
1431 | static SDValue emitNonHSAIntrinsicError(SelectionDAG& DAG, EVT VT) { |
1432 | DiagnosticInfoUnsupported BadIntrin(*DAG.getMachineFunction().getFunction(), |
1433 | "non-hsa intrinsic with hsa target"); |
1434 | DAG.getContext()->diagnose(BadIntrin); |
1435 | return DAG.getUNDEF(VT); |
1436 | } |
1437 | |
1438 | SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, |
1439 | SelectionDAG &DAG) const { |
1440 | MachineFunction &MF = DAG.getMachineFunction(); |
1441 | auto MFI = MF.getInfo<SIMachineFunctionInfo>(); |
1442 | const SIRegisterInfo *TRI = |
1443 | static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo()); |
1444 | |
1445 | EVT VT = Op.getValueType(); |
1446 | SDLoc DL(Op); |
1447 | unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); |
1448 | |
1449 | // TODO: Should this propagate fast-math-flags? |
1450 | |
1451 | switch (IntrinsicID) { |
1452 | case Intrinsic::amdgcn_dispatch_ptr: |
1453 | if (!Subtarget->isAmdHsaOS()) { |
1454 | DiagnosticInfoUnsupported BadIntrin( |
1455 | *MF.getFunction(), "unsupported hsa intrinsic without hsa target", |
1456 | DL.getDebugLoc()); |
1457 | DAG.getContext()->diagnose(BadIntrin); |
1458 | return DAG.getUNDEF(VT); |
1459 | } |
1460 | |
1461 | return CreateLiveInRegister(DAG, &AMDGPU::SReg_64RegClass, |
1462 | TRI->getPreloadedValue(MF, SIRegisterInfo::DISPATCH_PTR), VT); |
1463 | case Intrinsic::amdgcn_rcp: |
1464 | return DAG.getNode(AMDGPUISD::RCP, DL, VT, Op.getOperand(1)); |
1465 | case Intrinsic::amdgcn_rsq: |
1466 | case AMDGPUIntrinsic::AMDGPU_rsq: // Legacy name |
1467 | return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1)); |
1468 | case Intrinsic::amdgcn_rsq_clamp: |
1469 | case AMDGPUIntrinsic::AMDGPU_rsq_clamped: { // Legacy name |
1470 | if (Subtarget->getGeneration() < AMDGPUSubtarget::VOLCANIC_ISLANDS) |
1471 | return DAG.getNode(AMDGPUISD::RSQ_CLAMP, DL, VT, Op.getOperand(1)); |
1472 | |
1473 | Type *Type = VT.getTypeForEVT(*DAG.getContext()); |
1474 | APFloat Max = APFloat::getLargest(Type->getFltSemantics()); |
1475 | APFloat Min = APFloat::getLargest(Type->getFltSemantics(), true); |
1476 | |
1477 | SDValue Rsq = DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1)); |
1478 | SDValue Tmp = DAG.getNode(ISD::FMINNUM, DL, VT, Rsq, |
1479 | DAG.getConstantFP(Max, DL, VT)); |
1480 | return DAG.getNode(ISD::FMAXNUM, DL, VT, Tmp, |
1481 | DAG.getConstantFP(Min, DL, VT)); |
1482 | } |
1483 | case Intrinsic::r600_read_ngroups_x: |
1484 | if (Subtarget->isAmdHsaOS()) |
1485 | return emitNonHSAIntrinsicError(DAG, VT); |
1486 | |
1487 | return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), |
1488 | SI::KernelInputOffsets::NGROUPS_X, false); |
1489 | case Intrinsic::r600_read_ngroups_y: |
1490 | if (Subtarget->isAmdHsaOS()) |
1491 | return emitNonHSAIntrinsicError(DAG, VT); |
1492 | |
1493 | return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), |
1494 | SI::KernelInputOffsets::NGROUPS_Y, false); |
1495 | case Intrinsic::r600_read_ngroups_z: |
1496 | if (Subtarget->isAmdHsaOS()) |
1497 | return emitNonHSAIntrinsicError(DAG, VT); |
1498 | |
1499 | return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), |
1500 | SI::KernelInputOffsets::NGROUPS_Z, false); |
1501 | case Intrinsic::r600_read_global_size_x: |
1502 | if (Subtarget->isAmdHsaOS()) |
1503 | return emitNonHSAIntrinsicError(DAG, VT); |
1504 | |
1505 | return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), |
1506 | SI::KernelInputOffsets::GLOBAL_SIZE_X, false); |
1507 | case Intrinsic::r600_read_global_size_y: |
1508 | if (Subtarget->isAmdHsaOS()) |
1509 | return emitNonHSAIntrinsicError(DAG, VT); |
1510 | |
1511 | return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), |
1512 | SI::KernelInputOffsets::GLOBAL_SIZE_Y, false); |
1513 | case Intrinsic::r600_read_global_size_z: |
1514 | if (Subtarget->isAmdHsaOS()) |
1515 | return emitNonHSAIntrinsicError(DAG, VT); |
1516 | |
1517 | return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), |
1518 | SI::KernelInputOffsets::GLOBAL_SIZE_Z, false); |
1519 | case Intrinsic::r600_read_local_size_x: |
1520 | if (Subtarget->isAmdHsaOS()) |
1521 | return emitNonHSAIntrinsicError(DAG, VT); |
1522 | |
1523 | return lowerImplicitZextParam(DAG, Op, MVT::i16, |
1524 | SI::KernelInputOffsets::LOCAL_SIZE_X); |
1525 | case Intrinsic::r600_read_local_size_y: |
1526 | if (Subtarget->isAmdHsaOS()) |
1527 | return emitNonHSAIntrinsicError(DAG, VT); |
1528 | |
1529 | return lowerImplicitZextParam(DAG, Op, MVT::i16, |
1530 | SI::KernelInputOffsets::LOCAL_SIZE_Y); |
1531 | case Intrinsic::r600_read_local_size_z: |
1532 | if (Subtarget->isAmdHsaOS()) |
1533 | return emitNonHSAIntrinsicError(DAG, VT); |
1534 | |
1535 | return lowerImplicitZextParam(DAG, Op, MVT::i16, |
1536 | SI::KernelInputOffsets::LOCAL_SIZE_Z); |
1537 | case Intrinsic::amdgcn_read_workdim: |
1538 | case AMDGPUIntrinsic::AMDGPU_read_workdim: // Legacy name. |
1539 | // Really only 2 bits. |
1540 | return lowerImplicitZextParam(DAG, Op, MVT::i8, |
1541 | getImplicitParameterOffset(MFI, GRID_DIM)); |
1542 | case Intrinsic::amdgcn_workgroup_id_x: |
1543 | case Intrinsic::r600_read_tgid_x: |
1544 | return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass, |
1545 | TRI->getPreloadedValue(MF, SIRegisterInfo::WORKGROUP_ID_X), VT); |
1546 | case Intrinsic::amdgcn_workgroup_id_y: |
1547 | case Intrinsic::r600_read_tgid_y: |
1548 | return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass, |
1549 | TRI->getPreloadedValue(MF, SIRegisterInfo::WORKGROUP_ID_Y), VT); |
1550 | case Intrinsic::amdgcn_workgroup_id_z: |
1551 | case Intrinsic::r600_read_tgid_z: |
1552 | return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass, |
1553 | TRI->getPreloadedValue(MF, SIRegisterInfo::WORKGROUP_ID_Z), VT); |
1554 | case Intrinsic::amdgcn_workitem_id_x: |
1555 | case Intrinsic::r600_read_tidig_x: |
1556 | return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass, |
1557 | TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_X), VT); |
1558 | case Intrinsic::amdgcn_workitem_id_y: |
1559 | case Intrinsic::r600_read_tidig_y: |
1560 | return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass, |
1561 | TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Y), VT); |
1562 | case Intrinsic::amdgcn_workitem_id_z: |
1563 | case Intrinsic::r600_read_tidig_z: |
1564 | return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass, |
1565 | TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Z), VT); |
1566 | case AMDGPUIntrinsic::SI_load_const: { |
1567 | SDValue Ops[] = { |
1568 | Op.getOperand(1), |
1569 | Op.getOperand(2) |
1570 | }; |
1571 | |
1572 | MachineMemOperand *MMO = MF.getMachineMemOperand( |
1573 | MachinePointerInfo(), |
1574 | MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant, |
1575 | VT.getStoreSize(), 4); |
1576 | return DAG.getMemIntrinsicNode(AMDGPUISD::LOAD_CONSTANT, DL, |
1577 | Op->getVTList(), Ops, VT, MMO); |
1578 | } |
1579 | case AMDGPUIntrinsic::SI_vs_load_input: |
1580 | return DAG.getNode(AMDGPUISD::LOAD_INPUT, DL, VT, |
1581 | Op.getOperand(1), |
1582 | Op.getOperand(2), |
1583 | Op.getOperand(3)); |
1584 | |
1585 | case AMDGPUIntrinsic::SI_fs_constant: { |
1586 | SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(3)); |
1587 | SDValue Glue = M0.getValue(1); |
1588 | return DAG.getNode(AMDGPUISD::INTERP_MOV, DL, MVT::f32, |
1589 | DAG.getConstant(2, DL, MVT::i32), // P0 |
1590 | Op.getOperand(1), Op.getOperand(2), Glue); |
1591 | } |
1592 | case AMDGPUIntrinsic::SI_packf16: |
1593 | if (Op.getOperand(1).isUndef() && Op.getOperand(2).isUndef()) |
1594 | return DAG.getUNDEF(MVT::i32); |
1595 | return Op; |
1596 | case AMDGPUIntrinsic::SI_fs_interp: { |
1597 | SDValue IJ = Op.getOperand(4); |
1598 | SDValue I = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, IJ, |
1599 | DAG.getConstant(0, DL, MVT::i32)); |
1600 | SDValue J = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, IJ, |
1601 | DAG.getConstant(1, DL, MVT::i32)); |
1602 | SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(3)); |
1603 | SDValue Glue = M0.getValue(1); |
1604 | SDValue P1 = DAG.getNode(AMDGPUISD::INTERP_P1, DL, |
1605 | DAG.getVTList(MVT::f32, MVT::Glue), |
1606 | I, Op.getOperand(1), Op.getOperand(2), Glue); |
1607 | Glue = SDValue(P1.getNode(), 1); |
1608 | return DAG.getNode(AMDGPUISD::INTERP_P2, DL, MVT::f32, P1, J, |
1609 | Op.getOperand(1), Op.getOperand(2), Glue); |
1610 | } |
1611 | case Intrinsic::amdgcn_interp_p1: { |
1612 | SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(4)); |
1613 | SDValue Glue = M0.getValue(1); |
1614 | return DAG.getNode(AMDGPUISD::INTERP_P1, DL, MVT::f32, Op.getOperand(1), |
1615 | Op.getOperand(2), Op.getOperand(3), Glue); |
1616 | } |
1617 | case Intrinsic::amdgcn_interp_p2: { |
1618 | SDValue M0 = copyToM0(DAG, DAG.getEntryNode(), DL, Op.getOperand(5)); |
1619 | SDValue Glue = SDValue(M0.getNode(), 1); |
1620 | return DAG.getNode(AMDGPUISD::INTERP_P2, DL, MVT::f32, Op.getOperand(1), |
1621 | Op.getOperand(2), Op.getOperand(3), Op.getOperand(4), |
1622 | Glue); |
1623 | } |
1624 | case Intrinsic::amdgcn_sin: |
1625 | return DAG.getNode(AMDGPUISD::SIN_HW, DL, VT, Op.getOperand(1)); |
1626 | |
1627 | case Intrinsic::amdgcn_cos: |
1628 | return DAG.getNode(AMDGPUISD::COS_HW, DL, VT, Op.getOperand(1)); |
1629 | |
1630 | case Intrinsic::amdgcn_log_clamp: { |
1631 | if (Subtarget->getGeneration() < AMDGPUSubtarget::VOLCANIC_ISLANDS) |
1632 | return SDValue(); |
1633 | |
1634 | DiagnosticInfoUnsupported BadIntrin( |
1635 | *MF.getFunction(), "intrinsic not supported on subtarget", |
1636 | DL.getDebugLoc()); |
1637 | DAG.getContext()->diagnose(BadIntrin); |
1638 | return DAG.getUNDEF(VT); |
1639 | } |
1640 | case Intrinsic::amdgcn_ldexp: |
1641 | return DAG.getNode(AMDGPUISD::LDEXP, DL, VT, |
1642 | Op.getOperand(1), Op.getOperand(2)); |
1643 | case Intrinsic::amdgcn_class: |
1644 | return DAG.getNode(AMDGPUISD::FP_CLASS, DL, VT, |
1645 | Op.getOperand(1), Op.getOperand(2)); |
1646 | case Intrinsic::amdgcn_div_fmas: |
1647 | return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT, |
1648 | Op.getOperand(1), Op.getOperand(2), Op.getOperand(3), |
1649 | Op.getOperand(4)); |
1650 | |
1651 | case Intrinsic::amdgcn_div_fixup: |
1652 | return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT, |
1653 | Op.getOperand(1), Op.getOperand(2), Op.getOperand(3)); |
1654 | |
1655 | case Intrinsic::amdgcn_trig_preop: |
1656 | return DAG.getNode(AMDGPUISD::TRIG_PREOP, DL, VT, |
1657 | Op.getOperand(1), Op.getOperand(2)); |
1658 | case Intrinsic::amdgcn_div_scale: { |
1659 | // 3rd parameter required to be a constant. |
1660 | const ConstantSDNode *Param = dyn_cast<ConstantSDNode>(Op.getOperand(3)); |
1661 | if (!Param) |
1662 | return DAG.getUNDEF(VT); |
1663 | |
1664 | // Translate to the operands expected by the machine instruction. The |
1665 | // first parameter must be the same as the first instruction. |
1666 | SDValue Numerator = Op.getOperand(1); |
1667 | SDValue Denominator = Op.getOperand(2); |
1668 | |
1669 | // Note this order is opposite of the machine instruction's operations, |
1670 | // which is s0.f = Quotient, s1.f = Denominator, s2.f = Numerator. The |
1671 | // intrinsic has the numerator as the first operand to match a normal |
1672 | // division operation. |
1673 | |
1674 | SDValue Src0 = Param->isAllOnesValue() ? Numerator : Denominator; |
1675 | |
1676 | return DAG.getNode(AMDGPUISD::DIV_SCALE, DL, Op->getVTList(), Src0, |
1677 | Denominator, Numerator); |
1678 | } |
1679 | case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0: |
1680 | return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1)); |
1681 | case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1: |
1682 | return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1)); |
1683 | case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2: |
1684 | return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1)); |
1685 | case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3: |
1686 | return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1)); |
1687 | default: |
1688 | return AMDGPUTargetLowering::LowerOperation(Op, DAG); |
1689 | } |
1690 | } |
1691 | |
1692 | SDValue SITargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op, |
1693 | SelectionDAG &DAG) const { |
1694 | unsigned IntrID = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue(); |
1695 | switch (IntrID) { |
1696 | case Intrinsic::amdgcn_atomic_inc: |
1697 | case Intrinsic::amdgcn_atomic_dec: { |
1698 | MemSDNode *M = cast<MemSDNode>(Op); |
1699 | unsigned Opc = (IntrID == Intrinsic::amdgcn_atomic_inc) ? |
1700 | AMDGPUISD::ATOMIC_INC : AMDGPUISD::ATOMIC_DEC; |
1701 | SDValue Ops[] = { |
1702 | M->getOperand(0), // Chain |
1703 | M->getOperand(2), // Ptr |
1704 | M->getOperand(3) // Value |
1705 | }; |
1706 | |
1707 | return DAG.getMemIntrinsicNode(Opc, SDLoc(Op), M->getVTList(), Ops, |
1708 | M->getMemoryVT(), M->getMemOperand()); |
1709 | } |
1710 | default: |
1711 | return SDValue(); |
1712 | } |
1713 | } |
1714 | |
1715 | SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op, |
1716 | SelectionDAG &DAG) const { |
1717 | MachineFunction &MF = DAG.getMachineFunction(); |
1718 | SDLoc DL(Op); |
1719 | SDValue Chain = Op.getOperand(0); |
1720 | unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue(); |
1721 | |
1722 | switch (IntrinsicID) { |
1723 | case AMDGPUIntrinsic::SI_sendmsg: { |
1724 | Chain = copyToM0(DAG, Chain, DL, Op.getOperand(3)); |
1725 | SDValue Glue = Chain.getValue(1); |
1726 | return DAG.getNode(AMDGPUISD::SENDMSG, DL, MVT::Other, Chain, |
1727 | Op.getOperand(2), Glue); |
1728 | } |
1729 | case AMDGPUIntrinsic::SI_tbuffer_store: { |
1730 | SDValue Ops[] = { |
1731 | Chain, |
1732 | Op.getOperand(2), |
1733 | Op.getOperand(3), |
1734 | Op.getOperand(4), |
1735 | Op.getOperand(5), |
1736 | Op.getOperand(6), |
1737 | Op.getOperand(7), |
1738 | Op.getOperand(8), |
1739 | Op.getOperand(9), |
1740 | Op.getOperand(10), |
1741 | Op.getOperand(11), |
1742 | Op.getOperand(12), |
1743 | Op.getOperand(13), |
1744 | Op.getOperand(14) |
1745 | }; |
1746 | |
1747 | EVT VT = Op.getOperand(3).getValueType(); |
1748 | |
1749 | MachineMemOperand *MMO = MF.getMachineMemOperand( |
1750 | MachinePointerInfo(), |
1751 | MachineMemOperand::MOStore, |
1752 | VT.getStoreSize(), 4); |
1753 | return DAG.getMemIntrinsicNode(AMDGPUISD::TBUFFER_STORE_FORMAT, DL, |
1754 | Op->getVTList(), Ops, VT, MMO); |
1755 | } |
1756 | default: |
1757 | return SDValue(); |
1758 | } |
1759 | } |
1760 | |
1761 | SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const { |
1762 | SDLoc DL(Op); |
1763 | LoadSDNode *Load = cast<LoadSDNode>(Op); |
1764 | ISD::LoadExtType ExtType = Load->getExtensionType(); |
1765 | EVT MemVT = Load->getMemoryVT(); |
1766 | |
1767 | if (ExtType == ISD::NON_EXTLOAD && MemVT.getSizeInBits() < 32) { |
1768 | assert(MemVT == MVT::i1 && "Only i1 non-extloads expected")((MemVT == MVT::i1 && "Only i1 non-extloads expected" ) ? static_cast<void> (0) : __assert_fail ("MemVT == MVT::i1 && \"Only i1 non-extloads expected\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 1768, __PRETTY_FUNCTION__)); |
1769 | // FIXME: Copied from PPC |
1770 | // First, load into 32 bits, then truncate to 1 bit. |
1771 | |
1772 | SDValue Chain = Load->getChain(); |
1773 | SDValue BasePtr = Load->getBasePtr(); |
1774 | MachineMemOperand *MMO = Load->getMemOperand(); |
1775 | |
1776 | SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain, |
1777 | BasePtr, MVT::i8, MMO); |
1778 | |
1779 | SDValue Ops[] = { |
1780 | DAG.getNode(ISD::TRUNCATE, DL, MemVT, NewLD), |
1781 | NewLD.getValue(1) |
1782 | }; |
1783 | |
1784 | return DAG.getMergeValues(Ops, DL); |
1785 | } |
1786 | |
1787 | if (!MemVT.isVector()) |
1788 | return SDValue(); |
1789 | |
1790 | assert(Op.getValueType().getVectorElementType() == MVT::i32 &&((Op.getValueType().getVectorElementType() == MVT::i32 && "Custom lowering for non-i32 vectors hasn't been implemented." ) ? static_cast<void> (0) : __assert_fail ("Op.getValueType().getVectorElementType() == MVT::i32 && \"Custom lowering for non-i32 vectors hasn't been implemented.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 1791, __PRETTY_FUNCTION__)) |
1791 | "Custom lowering for non-i32 vectors hasn't been implemented.")((Op.getValueType().getVectorElementType() == MVT::i32 && "Custom lowering for non-i32 vectors hasn't been implemented." ) ? static_cast<void> (0) : __assert_fail ("Op.getValueType().getVectorElementType() == MVT::i32 && \"Custom lowering for non-i32 vectors hasn't been implemented.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 1791, __PRETTY_FUNCTION__)); |
1792 | unsigned NumElements = MemVT.getVectorNumElements(); |
1793 | assert(NumElements != 2 && "v2 loads are supported for all address spaces.")((NumElements != 2 && "v2 loads are supported for all address spaces." ) ? static_cast<void> (0) : __assert_fail ("NumElements != 2 && \"v2 loads are supported for all address spaces.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 1793, __PRETTY_FUNCTION__)); |
1794 | |
1795 | switch (Load->getAddressSpace()) { |
1796 | case AMDGPUAS::CONSTANT_ADDRESS: |
1797 | if (isMemOpUniform(Load)) |
1798 | return SDValue(); |
1799 | // Non-uniform loads will be selected to MUBUF instructions, so they |
1800 | // have the same legalization requires ments as global and private |
1801 | // loads. |
1802 | // |
1803 | // Fall-through |
1804 | case AMDGPUAS::GLOBAL_ADDRESS: |
1805 | case AMDGPUAS::FLAT_ADDRESS: |
1806 | if (NumElements > 4) |
1807 | return SplitVectorLoad(Op, DAG); |
1808 | // v4 loads are supported for private and global memory. |
1809 | return SDValue(); |
1810 | case AMDGPUAS::PRIVATE_ADDRESS: { |
1811 | // Depending on the setting of the private_element_size field in the |
1812 | // resource descriptor, we can only make private accesses up to a certain |
1813 | // size. |
1814 | switch (Subtarget->getMaxPrivateElementSize()) { |
1815 | case 4: |
1816 | return scalarizeVectorLoad(Load, DAG); |
1817 | case 8: |
1818 | if (NumElements > 2) |
1819 | return SplitVectorLoad(Op, DAG); |
1820 | return SDValue(); |
1821 | case 16: |
1822 | // Same as global/flat |
1823 | if (NumElements > 4) |
1824 | return SplitVectorLoad(Op, DAG); |
1825 | return SDValue(); |
1826 | default: |
1827 | llvm_unreachable("unsupported private_element_size")::llvm::llvm_unreachable_internal("unsupported private_element_size" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 1827); |
1828 | } |
1829 | } |
1830 | case AMDGPUAS::LOCAL_ADDRESS: |
1831 | // If properly aligned, if we split we might be able to use ds_read_b64. |
1832 | return SplitVectorLoad(Op, DAG); |
1833 | default: |
1834 | return SDValue(); |
1835 | } |
1836 | } |
1837 | |
1838 | SDValue SITargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const { |
1839 | if (Op.getValueType() != MVT::i64) |
1840 | return SDValue(); |
1841 | |
1842 | SDLoc DL(Op); |
1843 | SDValue Cond = Op.getOperand(0); |
1844 | |
1845 | SDValue Zero = DAG.getConstant(0, DL, MVT::i32); |
1846 | SDValue One = DAG.getConstant(1, DL, MVT::i32); |
1847 | |
1848 | SDValue LHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(1)); |
1849 | SDValue RHS = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Op.getOperand(2)); |
1850 | |
1851 | SDValue Lo0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, Zero); |
1852 | SDValue Lo1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, Zero); |
1853 | |
1854 | SDValue Lo = DAG.getSelect(DL, MVT::i32, Cond, Lo0, Lo1); |
1855 | |
1856 | SDValue Hi0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, LHS, One); |
1857 | SDValue Hi1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, RHS, One); |
1858 | |
1859 | SDValue Hi = DAG.getSelect(DL, MVT::i32, Cond, Hi0, Hi1); |
1860 | |
1861 | SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v2i32, Lo, Hi); |
1862 | return DAG.getNode(ISD::BITCAST, DL, MVT::i64, Res); |
1863 | } |
1864 | |
1865 | // Catch division cases where we can use shortcuts with rcp and rsq |
1866 | // instructions. |
1867 | SDValue SITargetLowering::LowerFastFDIV(SDValue Op, SelectionDAG &DAG) const { |
1868 | SDLoc SL(Op); |
1869 | SDValue LHS = Op.getOperand(0); |
1870 | SDValue RHS = Op.getOperand(1); |
1871 | EVT VT = Op.getValueType(); |
1872 | bool Unsafe = DAG.getTarget().Options.UnsafeFPMath; |
1873 | |
1874 | if (const ConstantFPSDNode *CLHS = dyn_cast<ConstantFPSDNode>(LHS)) { |
1875 | if ((Unsafe || (VT == MVT::f32 && !Subtarget->hasFP32Denormals())) && |
1876 | CLHS->isExactlyValue(1.0)) { |
1877 | // v_rcp_f32 and v_rsq_f32 do not support denormals, and according to |
1878 | // the CI documentation has a worst case error of 1 ulp. |
1879 | // OpenCL requires <= 2.5 ulp for 1.0 / x, so it should always be OK to |
1880 | // use it as long as we aren't trying to use denormals. |
1881 | |
1882 | // 1.0 / sqrt(x) -> rsq(x) |
1883 | // |
1884 | // XXX - Is UnsafeFPMath sufficient to do this for f64? The maximum ULP |
1885 | // error seems really high at 2^29 ULP. |
1886 | if (RHS.getOpcode() == ISD::FSQRT) |
1887 | return DAG.getNode(AMDGPUISD::RSQ, SL, VT, RHS.getOperand(0)); |
1888 | |
1889 | // 1.0 / x -> rcp(x) |
1890 | return DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS); |
1891 | } |
1892 | } |
1893 | |
1894 | if (Unsafe) { |
1895 | // Turn into multiply by the reciprocal. |
1896 | // x / y -> x * (1.0 / y) |
1897 | SDNodeFlags Flags; |
1898 | Flags.setUnsafeAlgebra(true); |
1899 | SDValue Recip = DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS); |
1900 | return DAG.getNode(ISD::FMUL, SL, VT, LHS, Recip, &Flags); |
1901 | } |
1902 | |
1903 | return SDValue(); |
1904 | } |
1905 | |
1906 | SDValue SITargetLowering::LowerFDIV32(SDValue Op, SelectionDAG &DAG) const { |
1907 | if (SDValue FastLowered = LowerFastFDIV(Op, DAG)) |
1908 | return FastLowered; |
1909 | |
1910 | // This uses v_rcp_f32 which does not handle denormals. Let this hit a |
1911 | // selection error for now rather than do something incorrect. |
1912 | if (Subtarget->hasFP32Denormals()) |
1913 | return SDValue(); |
1914 | |
1915 | SDLoc SL(Op); |
1916 | SDValue LHS = Op.getOperand(0); |
1917 | SDValue RHS = Op.getOperand(1); |
1918 | |
1919 | SDValue r1 = DAG.getNode(ISD::FABS, SL, MVT::f32, RHS); |
1920 | |
1921 | const APFloat K0Val(BitsToFloat(0x6f800000)); |
1922 | const SDValue K0 = DAG.getConstantFP(K0Val, SL, MVT::f32); |
1923 | |
1924 | const APFloat K1Val(BitsToFloat(0x2f800000)); |
1925 | const SDValue K1 = DAG.getConstantFP(K1Val, SL, MVT::f32); |
1926 | |
1927 | const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f32); |
1928 | |
1929 | EVT SetCCVT = |
1930 | getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f32); |
1931 | |
1932 | SDValue r2 = DAG.getSetCC(SL, SetCCVT, r1, K0, ISD::SETOGT); |
1933 | |
1934 | SDValue r3 = DAG.getNode(ISD::SELECT, SL, MVT::f32, r2, K1, One); |
1935 | |
1936 | // TODO: Should this propagate fast-math-flags? |
1937 | |
1938 | r1 = DAG.getNode(ISD::FMUL, SL, MVT::f32, RHS, r3); |
1939 | |
1940 | SDValue r0 = DAG.getNode(AMDGPUISD::RCP, SL, MVT::f32, r1); |
1941 | |
1942 | SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f32, LHS, r0); |
1943 | |
1944 | return DAG.getNode(ISD::FMUL, SL, MVT::f32, r3, Mul); |
1945 | } |
1946 | |
1947 | SDValue SITargetLowering::LowerFDIV64(SDValue Op, SelectionDAG &DAG) const { |
1948 | if (DAG.getTarget().Options.UnsafeFPMath) |
1949 | return LowerFastFDIV(Op, DAG); |
1950 | |
1951 | SDLoc SL(Op); |
1952 | SDValue X = Op.getOperand(0); |
1953 | SDValue Y = Op.getOperand(1); |
1954 | |
1955 | const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f64); |
1956 | |
1957 | SDVTList ScaleVT = DAG.getVTList(MVT::f64, MVT::i1); |
1958 | |
1959 | SDValue DivScale0 = DAG.getNode(AMDGPUISD::DIV_SCALE, SL, ScaleVT, Y, Y, X); |
1960 | |
1961 | SDValue NegDivScale0 = DAG.getNode(ISD::FNEG, SL, MVT::f64, DivScale0); |
1962 | |
1963 | SDValue Rcp = DAG.getNode(AMDGPUISD::RCP, SL, MVT::f64, DivScale0); |
1964 | |
1965 | SDValue Fma0 = DAG.getNode(ISD::FMA, SL, MVT::f64, NegDivScale0, Rcp, One); |
1966 | |
1967 | SDValue Fma1 = DAG.getNode(ISD::FMA, SL, MVT::f64, Rcp, Fma0, Rcp); |
1968 | |
1969 | SDValue Fma2 = DAG.getNode(ISD::FMA, SL, MVT::f64, NegDivScale0, Fma1, One); |
1970 | |
1971 | SDValue DivScale1 = DAG.getNode(AMDGPUISD::DIV_SCALE, SL, ScaleVT, X, Y, X); |
1972 | |
1973 | SDValue Fma3 = DAG.getNode(ISD::FMA, SL, MVT::f64, Fma1, Fma2, Fma1); |
1974 | SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f64, DivScale1, Fma3); |
1975 | |
1976 | SDValue Fma4 = DAG.getNode(ISD::FMA, SL, MVT::f64, |
1977 | NegDivScale0, Mul, DivScale1); |
1978 | |
1979 | SDValue Scale; |
1980 | |
1981 | if (Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS) { |
1982 | // Workaround a hardware bug on SI where the condition output from div_scale |
1983 | // is not usable. |
1984 | |
1985 | const SDValue Hi = DAG.getConstant(1, SL, MVT::i32); |
1986 | |
1987 | // Figure out if the scale to use for div_fmas. |
1988 | SDValue NumBC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, X); |
1989 | SDValue DenBC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Y); |
1990 | SDValue Scale0BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, DivScale0); |
1991 | SDValue Scale1BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, DivScale1); |
1992 | |
1993 | SDValue NumHi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, NumBC, Hi); |
1994 | SDValue DenHi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, DenBC, Hi); |
1995 | |
1996 | SDValue Scale0Hi |
1997 | = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Scale0BC, Hi); |
1998 | SDValue Scale1Hi |
1999 | = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Scale1BC, Hi); |
2000 | |
2001 | SDValue CmpDen = DAG.getSetCC(SL, MVT::i1, DenHi, Scale0Hi, ISD::SETEQ); |
2002 | SDValue CmpNum = DAG.getSetCC(SL, MVT::i1, NumHi, Scale1Hi, ISD::SETEQ); |
2003 | Scale = DAG.getNode(ISD::XOR, SL, MVT::i1, CmpNum, CmpDen); |
2004 | } else { |
2005 | Scale = DivScale1.getValue(1); |
2006 | } |
2007 | |
2008 | SDValue Fmas = DAG.getNode(AMDGPUISD::DIV_FMAS, SL, MVT::f64, |
2009 | Fma4, Fma3, Mul, Scale); |
2010 | |
2011 | return DAG.getNode(AMDGPUISD::DIV_FIXUP, SL, MVT::f64, Fmas, Y, X); |
2012 | } |
2013 | |
2014 | SDValue SITargetLowering::LowerFDIV(SDValue Op, SelectionDAG &DAG) const { |
2015 | EVT VT = Op.getValueType(); |
2016 | |
2017 | if (VT == MVT::f32) |
2018 | return LowerFDIV32(Op, DAG); |
2019 | |
2020 | if (VT == MVT::f64) |
2021 | return LowerFDIV64(Op, DAG); |
2022 | |
2023 | llvm_unreachable("Unexpected type for fdiv")::llvm::llvm_unreachable_internal("Unexpected type for fdiv", "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 2023); |
2024 | } |
2025 | |
2026 | SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const { |
2027 | SDLoc DL(Op); |
2028 | StoreSDNode *Store = cast<StoreSDNode>(Op); |
2029 | EVT VT = Store->getMemoryVT(); |
2030 | |
2031 | if (VT == MVT::i1) { |
2032 | return DAG.getTruncStore(Store->getChain(), DL, |
2033 | DAG.getSExtOrTrunc(Store->getValue(), DL, MVT::i32), |
2034 | Store->getBasePtr(), MVT::i1, Store->getMemOperand()); |
2035 | } |
2036 | |
2037 | assert(Store->getValue().getValueType().getScalarType() == MVT::i32)((Store->getValue().getValueType().getScalarType() == MVT:: i32) ? static_cast<void> (0) : __assert_fail ("Store->getValue().getValueType().getScalarType() == MVT::i32" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 2037, __PRETTY_FUNCTION__)); |
2038 | |
2039 | unsigned NumElements = VT.getVectorNumElements(); |
2040 | switch (Store->getAddressSpace()) { |
2041 | case AMDGPUAS::GLOBAL_ADDRESS: |
2042 | case AMDGPUAS::FLAT_ADDRESS: |
2043 | if (NumElements > 4) |
2044 | return SplitVectorStore(Op, DAG); |
2045 | return SDValue(); |
2046 | case AMDGPUAS::PRIVATE_ADDRESS: { |
2047 | switch (Subtarget->getMaxPrivateElementSize()) { |
2048 | case 4: |
2049 | return scalarizeVectorStore(Store, DAG); |
2050 | case 8: |
2051 | if (NumElements > 2) |
2052 | return SplitVectorStore(Op, DAG); |
2053 | return SDValue(); |
2054 | case 16: |
2055 | if (NumElements > 4) |
2056 | return SplitVectorStore(Op, DAG); |
2057 | return SDValue(); |
2058 | default: |
2059 | llvm_unreachable("unsupported private_element_size")::llvm::llvm_unreachable_internal("unsupported private_element_size" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 2059); |
2060 | } |
2061 | } |
2062 | case AMDGPUAS::LOCAL_ADDRESS: |
2063 | // If properly aligned, if we split we might be able to use ds_write_b64. |
2064 | return SplitVectorStore(Op, DAG); |
2065 | default: |
2066 | llvm_unreachable("unhandled address space")::llvm::llvm_unreachable_internal("unhandled address space", "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 2066); |
2067 | } |
2068 | } |
2069 | |
2070 | SDValue SITargetLowering::LowerTrig(SDValue Op, SelectionDAG &DAG) const { |
2071 | SDLoc DL(Op); |
2072 | EVT VT = Op.getValueType(); |
2073 | SDValue Arg = Op.getOperand(0); |
2074 | // TODO: Should this propagate fast-math-flags? |
2075 | SDValue FractPart = DAG.getNode(AMDGPUISD::FRACT, DL, VT, |
2076 | DAG.getNode(ISD::FMUL, DL, VT, Arg, |
2077 | DAG.getConstantFP(0.5/M_PI3.14159265358979323846, DL, |
2078 | VT))); |
2079 | |
2080 | switch (Op.getOpcode()) { |
2081 | case ISD::FCOS: |
2082 | return DAG.getNode(AMDGPUISD::COS_HW, SDLoc(Op), VT, FractPart); |
2083 | case ISD::FSIN: |
2084 | return DAG.getNode(AMDGPUISD::SIN_HW, SDLoc(Op), VT, FractPart); |
2085 | default: |
2086 | llvm_unreachable("Wrong trig opcode")::llvm::llvm_unreachable_internal("Wrong trig opcode", "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 2086); |
2087 | } |
2088 | } |
2089 | |
2090 | SDValue SITargetLowering::LowerATOMIC_CMP_SWAP(SDValue Op, SelectionDAG &DAG) const { |
2091 | AtomicSDNode *AtomicNode = cast<AtomicSDNode>(Op); |
2092 | assert(AtomicNode->isCompareAndSwap())((AtomicNode->isCompareAndSwap()) ? static_cast<void> (0) : __assert_fail ("AtomicNode->isCompareAndSwap()", "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 2092, __PRETTY_FUNCTION__)); |
2093 | unsigned AS = AtomicNode->getAddressSpace(); |
2094 | |
2095 | // No custom lowering required for local address space |
2096 | if (!isFlatGlobalAddrSpace(AS)) |
2097 | return Op; |
2098 | |
2099 | // Non-local address space requires custom lowering for atomic compare |
2100 | // and swap; cmp and swap should be in a v2i32 or v2i64 in case of _X2 |
2101 | SDLoc DL(Op); |
2102 | SDValue ChainIn = Op.getOperand(0); |
2103 | SDValue Addr = Op.getOperand(1); |
2104 | SDValue Old = Op.getOperand(2); |
2105 | SDValue New = Op.getOperand(3); |
2106 | EVT VT = Op.getValueType(); |
2107 | MVT SimpleVT = VT.getSimpleVT(); |
2108 | MVT VecType = MVT::getVectorVT(SimpleVT, 2); |
2109 | |
2110 | SDValue NewOld = DAG.getNode(ISD::BUILD_VECTOR, DL, VecType, |
2111 | New, Old); |
2112 | SDValue Ops[] = { ChainIn, Addr, NewOld }; |
2113 | SDVTList VTList = DAG.getVTList(VT, MVT::Other); |
2114 | return DAG.getMemIntrinsicNode(AMDGPUISD::ATOMIC_CMP_SWAP, DL, |
2115 | VTList, Ops, VT, AtomicNode->getMemOperand()); |
2116 | } |
2117 | |
2118 | //===----------------------------------------------------------------------===// |
2119 | // Custom DAG optimizations |
2120 | //===----------------------------------------------------------------------===// |
2121 | |
2122 | SDValue SITargetLowering::performUCharToFloatCombine(SDNode *N, |
2123 | DAGCombinerInfo &DCI) const { |
2124 | EVT VT = N->getValueType(0); |
2125 | EVT ScalarVT = VT.getScalarType(); |
2126 | if (ScalarVT != MVT::f32) |
2127 | return SDValue(); |
2128 | |
2129 | SelectionDAG &DAG = DCI.DAG; |
2130 | SDLoc DL(N); |
2131 | |
2132 | SDValue Src = N->getOperand(0); |
2133 | EVT SrcVT = Src.getValueType(); |
2134 | |
2135 | // TODO: We could try to match extracting the higher bytes, which would be |
2136 | // easier if i8 vectors weren't promoted to i32 vectors, particularly after |
2137 | // types are legalized. v4i8 -> v4f32 is probably the only case to worry |
2138 | // about in practice. |
2139 | if (DCI.isAfterLegalizeVectorOps() && SrcVT == MVT::i32) { |
2140 | if (DAG.MaskedValueIsZero(Src, APInt::getHighBitsSet(32, 24))) { |
2141 | SDValue Cvt = DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Src); |
2142 | DCI.AddToWorklist(Cvt.getNode()); |
2143 | return Cvt; |
2144 | } |
2145 | } |
2146 | |
2147 | // We are primarily trying to catch operations on illegal vector types |
2148 | // before they are expanded. |
2149 | // For scalars, we can use the more flexible method of checking masked bits |
2150 | // after legalization. |
2151 | if (!DCI.isBeforeLegalize() || |
2152 | !SrcVT.isVector() || |
2153 | SrcVT.getVectorElementType() != MVT::i8) { |
2154 | return SDValue(); |
2155 | } |
2156 | |
2157 | assert(DCI.isBeforeLegalize() && "Unexpected legal type")((DCI.isBeforeLegalize() && "Unexpected legal type") ? static_cast<void> (0) : __assert_fail ("DCI.isBeforeLegalize() && \"Unexpected legal type\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 2157, __PRETTY_FUNCTION__)); |
2158 | |
2159 | // Weird sized vectors are a pain to handle, but we know 3 is really the same |
2160 | // size as 4. |
2161 | unsigned NElts = SrcVT.getVectorNumElements(); |
2162 | if (!SrcVT.isSimple() && NElts != 3) |
2163 | return SDValue(); |
2164 | |
2165 | // Handle v4i8 -> v4f32 extload. Replace the v4i8 with a legal i32 load to |
2166 | // prevent a mess from expanding to v4i32 and repacking. |
2167 | if (ISD::isNormalLoad(Src.getNode()) && Src.hasOneUse()) { |
2168 | EVT LoadVT = getEquivalentMemType(*DAG.getContext(), SrcVT); |
2169 | EVT RegVT = getEquivalentLoadRegType(*DAG.getContext(), SrcVT); |
2170 | EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f32, NElts); |
2171 | LoadSDNode *Load = cast<LoadSDNode>(Src); |
2172 | |
2173 | unsigned AS = Load->getAddressSpace(); |
2174 | unsigned Align = Load->getAlignment(); |
2175 | Type *Ty = LoadVT.getTypeForEVT(*DAG.getContext()); |
2176 | unsigned ABIAlignment = DAG.getDataLayout().getABITypeAlignment(Ty); |
2177 | |
2178 | // Don't try to replace the load if we have to expand it due to alignment |
2179 | // problems. Otherwise we will end up scalarizing the load, and trying to |
2180 | // repack into the vector for no real reason. |
2181 | if (Align < ABIAlignment && |
2182 | !allowsMisalignedMemoryAccesses(LoadVT, AS, Align, nullptr)) { |
2183 | return SDValue(); |
2184 | } |
2185 | |
2186 | SDValue NewLoad = DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegVT, |
2187 | Load->getChain(), |
2188 | Load->getBasePtr(), |
2189 | LoadVT, |
2190 | Load->getMemOperand()); |
2191 | |
2192 | // Make sure successors of the original load stay after it by updating |
2193 | // them to use the new Chain. |
2194 | DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), NewLoad.getValue(1)); |
2195 | |
2196 | SmallVector<SDValue, 4> Elts; |
2197 | if (RegVT.isVector()) |
2198 | DAG.ExtractVectorElements(NewLoad, Elts); |
2199 | else |
2200 | Elts.push_back(NewLoad); |
2201 | |
2202 | SmallVector<SDValue, 4> Ops; |
2203 | |
2204 | unsigned EltIdx = 0; |
2205 | for (SDValue Elt : Elts) { |
2206 | unsigned ComponentsInElt = std::min(4u, NElts - 4 * EltIdx); |
2207 | for (unsigned I = 0; I < ComponentsInElt; ++I) { |
2208 | unsigned Opc = AMDGPUISD::CVT_F32_UBYTE0 + I; |
2209 | SDValue Cvt = DAG.getNode(Opc, DL, MVT::f32, Elt); |
2210 | DCI.AddToWorklist(Cvt.getNode()); |
2211 | Ops.push_back(Cvt); |
2212 | } |
2213 | |
2214 | ++EltIdx; |
2215 | } |
2216 | |
2217 | assert(Ops.size() == NElts)((Ops.size() == NElts) ? static_cast<void> (0) : __assert_fail ("Ops.size() == NElts", "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 2217, __PRETTY_FUNCTION__)); |
2218 | |
2219 | return DAG.getNode(ISD::BUILD_VECTOR, DL, FloatVT, Ops); |
2220 | } |
2221 | |
2222 | return SDValue(); |
2223 | } |
2224 | |
2225 | /// \brief Return true if the given offset Size in bytes can be folded into |
2226 | /// the immediate offsets of a memory instruction for the given address space. |
2227 | static bool canFoldOffset(unsigned OffsetSize, unsigned AS, |
2228 | const AMDGPUSubtarget &STI) { |
2229 | switch (AS) { |
2230 | case AMDGPUAS::GLOBAL_ADDRESS: { |
2231 | // MUBUF instructions a 12-bit offset in bytes. |
2232 | return isUInt<12>(OffsetSize); |
2233 | } |
2234 | case AMDGPUAS::CONSTANT_ADDRESS: { |
2235 | // SMRD instructions have an 8-bit offset in dwords on SI and |
2236 | // a 20-bit offset in bytes on VI. |
2237 | if (STI.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) |
2238 | return isUInt<20>(OffsetSize); |
2239 | else |
2240 | return (OffsetSize % 4 == 0) && isUInt<8>(OffsetSize / 4); |
2241 | } |
2242 | case AMDGPUAS::LOCAL_ADDRESS: |
2243 | case AMDGPUAS::REGION_ADDRESS: { |
2244 | // The single offset versions have a 16-bit offset in bytes. |
2245 | return isUInt<16>(OffsetSize); |
2246 | } |
2247 | case AMDGPUAS::PRIVATE_ADDRESS: |
2248 | // Indirect register addressing does not use any offsets. |
2249 | default: |
2250 | return 0; |
2251 | } |
2252 | } |
2253 | |
2254 | // (shl (add x, c1), c2) -> add (shl x, c2), (shl c1, c2) |
2255 | |
2256 | // This is a variant of |
2257 | // (mul (add x, c1), c2) -> add (mul x, c2), (mul c1, c2), |
2258 | // |
2259 | // The normal DAG combiner will do this, but only if the add has one use since |
2260 | // that would increase the number of instructions. |
2261 | // |
2262 | // This prevents us from seeing a constant offset that can be folded into a |
2263 | // memory instruction's addressing mode. If we know the resulting add offset of |
2264 | // a pointer can be folded into an addressing offset, we can replace the pointer |
2265 | // operand with the add of new constant offset. This eliminates one of the uses, |
2266 | // and may allow the remaining use to also be simplified. |
2267 | // |
2268 | SDValue SITargetLowering::performSHLPtrCombine(SDNode *N, |
2269 | unsigned AddrSpace, |
2270 | DAGCombinerInfo &DCI) const { |
2271 | SDValue N0 = N->getOperand(0); |
2272 | SDValue N1 = N->getOperand(1); |
2273 | |
2274 | if (N0.getOpcode() != ISD::ADD) |
2275 | return SDValue(); |
2276 | |
2277 | const ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(N1); |
2278 | if (!CN1) |
2279 | return SDValue(); |
2280 | |
2281 | const ConstantSDNode *CAdd = dyn_cast<ConstantSDNode>(N0.getOperand(1)); |
2282 | if (!CAdd) |
2283 | return SDValue(); |
2284 | |
2285 | // If the resulting offset is too large, we can't fold it into the addressing |
2286 | // mode offset. |
2287 | APInt Offset = CAdd->getAPIntValue() << CN1->getAPIntValue(); |
2288 | if (!canFoldOffset(Offset.getZExtValue(), AddrSpace, *Subtarget)) |
2289 | return SDValue(); |
2290 | |
2291 | SelectionDAG &DAG = DCI.DAG; |
2292 | SDLoc SL(N); |
2293 | EVT VT = N->getValueType(0); |
2294 | |
2295 | SDValue ShlX = DAG.getNode(ISD::SHL, SL, VT, N0.getOperand(0), N1); |
2296 | SDValue COffset = DAG.getConstant(Offset, SL, MVT::i32); |
2297 | |
2298 | return DAG.getNode(ISD::ADD, SL, VT, ShlX, COffset); |
2299 | } |
2300 | |
2301 | SDValue SITargetLowering::performAndCombine(SDNode *N, |
2302 | DAGCombinerInfo &DCI) const { |
2303 | if (DCI.isBeforeLegalize()) |
2304 | return SDValue(); |
2305 | |
2306 | if (SDValue Base = AMDGPUTargetLowering::performAndCombine(N, DCI)) |
2307 | return Base; |
2308 | |
2309 | SelectionDAG &DAG = DCI.DAG; |
2310 | |
2311 | // (and (fcmp ord x, x), (fcmp une (fabs x), inf)) -> |
2312 | // fp_class x, ~(s_nan | q_nan | n_infinity | p_infinity) |
2313 | SDValue LHS = N->getOperand(0); |
2314 | SDValue RHS = N->getOperand(1); |
2315 | |
2316 | if (LHS.getOpcode() == ISD::SETCC && |
2317 | RHS.getOpcode() == ISD::SETCC) { |
2318 | ISD::CondCode LCC = cast<CondCodeSDNode>(LHS.getOperand(2))->get(); |
2319 | ISD::CondCode RCC = cast<CondCodeSDNode>(RHS.getOperand(2))->get(); |
2320 | |
2321 | SDValue X = LHS.getOperand(0); |
2322 | SDValue Y = RHS.getOperand(0); |
2323 | if (Y.getOpcode() != ISD::FABS || Y.getOperand(0) != X) |
2324 | return SDValue(); |
2325 | |
2326 | if (LCC == ISD::SETO) { |
2327 | if (X != LHS.getOperand(1)) |
2328 | return SDValue(); |
2329 | |
2330 | if (RCC == ISD::SETUNE) { |
2331 | const ConstantFPSDNode *C1 = dyn_cast<ConstantFPSDNode>(RHS.getOperand(1)); |
2332 | if (!C1 || !C1->isInfinity() || C1->isNegative()) |
2333 | return SDValue(); |
2334 | |
2335 | const uint32_t Mask = SIInstrFlags::N_NORMAL | |
2336 | SIInstrFlags::N_SUBNORMAL | |
2337 | SIInstrFlags::N_ZERO | |
2338 | SIInstrFlags::P_ZERO | |
2339 | SIInstrFlags::P_SUBNORMAL | |
2340 | SIInstrFlags::P_NORMAL; |
2341 | |
2342 | static_assert(((~(SIInstrFlags::S_NAN | |
2343 | SIInstrFlags::Q_NAN | |
2344 | SIInstrFlags::N_INFINITY | |
2345 | SIInstrFlags::P_INFINITY)) & 0x3ff) == Mask, |
2346 | "mask not equal"); |
2347 | |
2348 | SDLoc DL(N); |
2349 | return DAG.getNode(AMDGPUISD::FP_CLASS, DL, MVT::i1, |
2350 | X, DAG.getConstant(Mask, DL, MVT::i32)); |
2351 | } |
2352 | } |
2353 | } |
2354 | |
2355 | return SDValue(); |
2356 | } |
2357 | |
2358 | SDValue SITargetLowering::performOrCombine(SDNode *N, |
2359 | DAGCombinerInfo &DCI) const { |
2360 | SelectionDAG &DAG = DCI.DAG; |
2361 | SDValue LHS = N->getOperand(0); |
2362 | SDValue RHS = N->getOperand(1); |
2363 | |
2364 | EVT VT = N->getValueType(0); |
2365 | if (VT == MVT::i64) { |
2366 | // TODO: This could be a generic combine with a predicate for extracting the |
2367 | // high half of an integer being free. |
2368 | |
2369 | // (or i64:x, (zero_extend i32:y)) -> |
2370 | // i64 (bitcast (v2i32 build_vector (or i32:y, lo_32(x)), hi_32(x))) |
2371 | if (LHS.getOpcode() == ISD::ZERO_EXTEND && |
2372 | RHS.getOpcode() != ISD::ZERO_EXTEND) |
2373 | std::swap(LHS, RHS); |
2374 | |
2375 | if (RHS.getOpcode() == ISD::ZERO_EXTEND) { |
2376 | SDValue ExtSrc = RHS.getOperand(0); |
2377 | EVT SrcVT = ExtSrc.getValueType(); |
2378 | if (SrcVT == MVT::i32) { |
2379 | SDLoc SL(N); |
2380 | SDValue LowLHS, HiBits; |
2381 | std::tie(LowLHS, HiBits) = split64BitValue(LHS, DAG); |
2382 | SDValue LowOr = DAG.getNode(ISD::OR, SL, MVT::i32, LowLHS, ExtSrc); |
2383 | |
2384 | DCI.AddToWorklist(LowOr.getNode()); |
2385 | DCI.AddToWorklist(HiBits.getNode()); |
2386 | |
2387 | SDValue Vec = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32, |
2388 | LowOr, HiBits); |
2389 | return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Vec); |
2390 | } |
2391 | } |
2392 | } |
2393 | |
2394 | // or (fp_class x, c1), (fp_class x, c2) -> fp_class x, (c1 | c2) |
2395 | if (LHS.getOpcode() == AMDGPUISD::FP_CLASS && |
2396 | RHS.getOpcode() == AMDGPUISD::FP_CLASS) { |
2397 | SDValue Src = LHS.getOperand(0); |
2398 | if (Src != RHS.getOperand(0)) |
2399 | return SDValue(); |
2400 | |
2401 | const ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(LHS.getOperand(1)); |
2402 | const ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(RHS.getOperand(1)); |
2403 | if (!CLHS || !CRHS) |
2404 | return SDValue(); |
2405 | |
2406 | // Only 10 bits are used. |
2407 | static const uint32_t MaxMask = 0x3ff; |
2408 | |
2409 | uint32_t NewMask = (CLHS->getZExtValue() | CRHS->getZExtValue()) & MaxMask; |
2410 | SDLoc DL(N); |
2411 | return DAG.getNode(AMDGPUISD::FP_CLASS, DL, MVT::i1, |
2412 | Src, DAG.getConstant(NewMask, DL, MVT::i32)); |
2413 | } |
2414 | |
2415 | return SDValue(); |
2416 | } |
2417 | |
2418 | SDValue SITargetLowering::performClassCombine(SDNode *N, |
2419 | DAGCombinerInfo &DCI) const { |
2420 | SelectionDAG &DAG = DCI.DAG; |
2421 | SDValue Mask = N->getOperand(1); |
2422 | |
2423 | // fp_class x, 0 -> false |
2424 | if (const ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(Mask)) { |
2425 | if (CMask->isNullValue()) |
2426 | return DAG.getConstant(0, SDLoc(N), MVT::i1); |
2427 | } |
2428 | |
2429 | return SDValue(); |
2430 | } |
2431 | |
2432 | // Constant fold canonicalize. |
2433 | SDValue SITargetLowering::performFCanonicalizeCombine( |
2434 | SDNode *N, |
2435 | DAGCombinerInfo &DCI) const { |
2436 | ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)); |
2437 | if (!CFP) |
2438 | return SDValue(); |
2439 | |
2440 | SelectionDAG &DAG = DCI.DAG; |
2441 | const APFloat &C = CFP->getValueAPF(); |
2442 | |
2443 | // Flush denormals to 0 if not enabled. |
2444 | if (C.isDenormal()) { |
2445 | EVT VT = N->getValueType(0); |
2446 | if (VT == MVT::f32 && !Subtarget->hasFP32Denormals()) |
2447 | return DAG.getConstantFP(0.0, SDLoc(N), VT); |
2448 | |
2449 | if (VT == MVT::f64 && !Subtarget->hasFP64Denormals()) |
2450 | return DAG.getConstantFP(0.0, SDLoc(N), VT); |
2451 | } |
2452 | |
2453 | if (C.isNaN()) { |
2454 | EVT VT = N->getValueType(0); |
2455 | APFloat CanonicalQNaN = APFloat::getQNaN(C.getSemantics()); |
2456 | if (C.isSignaling()) { |
2457 | // Quiet a signaling NaN. |
2458 | return DAG.getConstantFP(CanonicalQNaN, SDLoc(N), VT); |
2459 | } |
2460 | |
2461 | // Make sure it is the canonical NaN bitpattern. |
2462 | // |
2463 | // TODO: Can we use -1 as the canonical NaN value since it's an inline |
2464 | // immediate? |
2465 | if (C.bitcastToAPInt() != CanonicalQNaN.bitcastToAPInt()) |
2466 | return DAG.getConstantFP(CanonicalQNaN, SDLoc(N), VT); |
2467 | } |
2468 | |
2469 | return SDValue(CFP, 0); |
2470 | } |
2471 | |
2472 | static unsigned minMaxOpcToMin3Max3Opc(unsigned Opc) { |
2473 | switch (Opc) { |
2474 | case ISD::FMAXNUM: |
2475 | return AMDGPUISD::FMAX3; |
2476 | case ISD::SMAX: |
2477 | return AMDGPUISD::SMAX3; |
2478 | case ISD::UMAX: |
2479 | return AMDGPUISD::UMAX3; |
2480 | case ISD::FMINNUM: |
2481 | return AMDGPUISD::FMIN3; |
2482 | case ISD::SMIN: |
2483 | return AMDGPUISD::SMIN3; |
2484 | case ISD::UMIN: |
2485 | return AMDGPUISD::UMIN3; |
2486 | default: |
2487 | llvm_unreachable("Not a min/max opcode")::llvm::llvm_unreachable_internal("Not a min/max opcode", "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 2487); |
2488 | } |
2489 | } |
2490 | |
2491 | static SDValue performIntMed3ImmCombine(SelectionDAG &DAG, |
2492 | SDLoc SL, |
2493 | SDValue Op0, |
2494 | SDValue Op1, |
2495 | bool Signed) { |
2496 | ConstantSDNode *K1 = dyn_cast<ConstantSDNode>(Op1); |
2497 | if (!K1) |
2498 | return SDValue(); |
2499 | |
2500 | ConstantSDNode *K0 = dyn_cast<ConstantSDNode>(Op0.getOperand(1)); |
2501 | if (!K0) |
2502 | return SDValue(); |
2503 | |
2504 | |
2505 | if (Signed) { |
2506 | if (K0->getAPIntValue().sge(K1->getAPIntValue())) |
2507 | return SDValue(); |
2508 | } else { |
2509 | if (K0->getAPIntValue().uge(K1->getAPIntValue())) |
2510 | return SDValue(); |
2511 | } |
2512 | |
2513 | EVT VT = K0->getValueType(0); |
2514 | return DAG.getNode(Signed ? AMDGPUISD::SMED3 : AMDGPUISD::UMED3, SL, VT, |
2515 | Op0.getOperand(0), SDValue(K0, 0), SDValue(K1, 0)); |
2516 | } |
2517 | |
2518 | static bool isKnownNeverSNan(SelectionDAG &DAG, SDValue Op) { |
2519 | if (!DAG.getTargetLoweringInfo().hasFloatingPointExceptions()) |
2520 | return true; |
2521 | |
2522 | return DAG.isKnownNeverNaN(Op); |
2523 | } |
2524 | |
2525 | static SDValue performFPMed3ImmCombine(SelectionDAG &DAG, |
2526 | SDLoc SL, |
2527 | SDValue Op0, |
2528 | SDValue Op1) { |
2529 | ConstantFPSDNode *K1 = dyn_cast<ConstantFPSDNode>(Op1); |
2530 | if (!K1) |
2531 | return SDValue(); |
2532 | |
2533 | ConstantFPSDNode *K0 = dyn_cast<ConstantFPSDNode>(Op0.getOperand(1)); |
2534 | if (!K0) |
2535 | return SDValue(); |
2536 | |
2537 | // Ordered >= (although NaN inputs should have folded away by now). |
2538 | APFloat::cmpResult Cmp = K0->getValueAPF().compare(K1->getValueAPF()); |
2539 | if (Cmp == APFloat::cmpGreaterThan) |
2540 | return SDValue(); |
2541 | |
2542 | // This isn't safe with signaling NaNs because in IEEE mode, min/max on a |
2543 | // signaling NaN gives a quiet NaN. The quiet NaN input to the min would then |
2544 | // give the other result, which is different from med3 with a NaN input. |
2545 | SDValue Var = Op0.getOperand(0); |
2546 | if (!isKnownNeverSNan(DAG, Var)) |
2547 | return SDValue(); |
2548 | |
2549 | return DAG.getNode(AMDGPUISD::FMED3, SL, K0->getValueType(0), |
2550 | Var, SDValue(K0, 0), SDValue(K1, 0)); |
2551 | } |
2552 | |
2553 | SDValue SITargetLowering::performMinMaxCombine(SDNode *N, |
2554 | DAGCombinerInfo &DCI) const { |
2555 | SelectionDAG &DAG = DCI.DAG; |
2556 | |
2557 | unsigned Opc = N->getOpcode(); |
2558 | SDValue Op0 = N->getOperand(0); |
2559 | SDValue Op1 = N->getOperand(1); |
2560 | |
2561 | // Only do this if the inner op has one use since this will just increases |
2562 | // register pressure for no benefit. |
2563 | |
2564 | if (Opc != AMDGPUISD::FMIN_LEGACY && Opc != AMDGPUISD::FMAX_LEGACY) { |
2565 | // max(max(a, b), c) -> max3(a, b, c) |
2566 | // min(min(a, b), c) -> min3(a, b, c) |
2567 | if (Op0.getOpcode() == Opc && Op0.hasOneUse()) { |
2568 | SDLoc DL(N); |
2569 | return DAG.getNode(minMaxOpcToMin3Max3Opc(Opc), |
2570 | DL, |
2571 | N->getValueType(0), |
2572 | Op0.getOperand(0), |
2573 | Op0.getOperand(1), |
2574 | Op1); |
2575 | } |
2576 | |
2577 | // Try commuted. |
2578 | // max(a, max(b, c)) -> max3(a, b, c) |
2579 | // min(a, min(b, c)) -> min3(a, b, c) |
2580 | if (Op1.getOpcode() == Opc && Op1.hasOneUse()) { |
2581 | SDLoc DL(N); |
2582 | return DAG.getNode(minMaxOpcToMin3Max3Opc(Opc), |
2583 | DL, |
2584 | N->getValueType(0), |
2585 | Op0, |
2586 | Op1.getOperand(0), |
2587 | Op1.getOperand(1)); |
2588 | } |
2589 | } |
2590 | |
2591 | // min(max(x, K0), K1), K0 < K1 -> med3(x, K0, K1) |
2592 | if (Opc == ISD::SMIN && Op0.getOpcode() == ISD::SMAX && Op0.hasOneUse()) { |
2593 | if (SDValue Med3 = performIntMed3ImmCombine(DAG, SDLoc(N), Op0, Op1, true)) |
2594 | return Med3; |
2595 | } |
2596 | |
2597 | if (Opc == ISD::UMIN && Op0.getOpcode() == ISD::UMAX && Op0.hasOneUse()) { |
2598 | if (SDValue Med3 = performIntMed3ImmCombine(DAG, SDLoc(N), Op0, Op1, false)) |
2599 | return Med3; |
2600 | } |
2601 | |
2602 | // fminnum(fmaxnum(x, K0), K1), K0 < K1 && !is_snan(x) -> fmed3(x, K0, K1) |
2603 | if (((Opc == ISD::FMINNUM && Op0.getOpcode() == ISD::FMAXNUM) || |
2604 | (Opc == AMDGPUISD::FMIN_LEGACY && |
2605 | Op0.getOpcode() == AMDGPUISD::FMAX_LEGACY)) && |
2606 | N->getValueType(0) == MVT::f32 && Op0.hasOneUse()) { |
2607 | if (SDValue Res = performFPMed3ImmCombine(DAG, SDLoc(N), Op0, Op1)) |
2608 | return Res; |
2609 | } |
2610 | |
2611 | return SDValue(); |
2612 | } |
2613 | |
2614 | SDValue SITargetLowering::performSetCCCombine(SDNode *N, |
2615 | DAGCombinerInfo &DCI) const { |
2616 | SelectionDAG &DAG = DCI.DAG; |
2617 | SDLoc SL(N); |
2618 | |
2619 | SDValue LHS = N->getOperand(0); |
2620 | SDValue RHS = N->getOperand(1); |
2621 | EVT VT = LHS.getValueType(); |
2622 | |
2623 | if (VT != MVT::f32 && VT != MVT::f64) |
2624 | return SDValue(); |
2625 | |
2626 | // Match isinf pattern |
2627 | // (fcmp oeq (fabs x), inf) -> (fp_class x, (p_infinity | n_infinity)) |
2628 | ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get(); |
2629 | if (CC == ISD::SETOEQ && LHS.getOpcode() == ISD::FABS) { |
2630 | const ConstantFPSDNode *CRHS = dyn_cast<ConstantFPSDNode>(RHS); |
2631 | if (!CRHS) |
2632 | return SDValue(); |
2633 | |
2634 | const APFloat &APF = CRHS->getValueAPF(); |
2635 | if (APF.isInfinity() && !APF.isNegative()) { |
2636 | unsigned Mask = SIInstrFlags::P_INFINITY | SIInstrFlags::N_INFINITY; |
2637 | return DAG.getNode(AMDGPUISD::FP_CLASS, SL, MVT::i1, LHS.getOperand(0), |
2638 | DAG.getConstant(Mask, SL, MVT::i32)); |
2639 | } |
2640 | } |
2641 | |
2642 | return SDValue(); |
2643 | } |
2644 | |
2645 | SDValue SITargetLowering::PerformDAGCombine(SDNode *N, |
2646 | DAGCombinerInfo &DCI) const { |
2647 | SelectionDAG &DAG = DCI.DAG; |
2648 | SDLoc DL(N); |
2649 | |
2650 | switch (N->getOpcode()) { |
2651 | default: |
2652 | return AMDGPUTargetLowering::PerformDAGCombine(N, DCI); |
2653 | case ISD::SETCC: |
2654 | return performSetCCCombine(N, DCI); |
2655 | case ISD::FMAXNUM: |
2656 | case ISD::FMINNUM: |
2657 | case ISD::SMAX: |
2658 | case ISD::SMIN: |
2659 | case ISD::UMAX: |
2660 | case ISD::UMIN: |
2661 | case AMDGPUISD::FMIN_LEGACY: |
2662 | case AMDGPUISD::FMAX_LEGACY: { |
2663 | if (DCI.getDAGCombineLevel() >= AfterLegalizeDAG && |
2664 | N->getValueType(0) != MVT::f64 && |
2665 | getTargetMachine().getOptLevel() > CodeGenOpt::None) |
2666 | return performMinMaxCombine(N, DCI); |
2667 | break; |
2668 | } |
2669 | |
2670 | case AMDGPUISD::CVT_F32_UBYTE0: |
2671 | case AMDGPUISD::CVT_F32_UBYTE1: |
2672 | case AMDGPUISD::CVT_F32_UBYTE2: |
2673 | case AMDGPUISD::CVT_F32_UBYTE3: { |
2674 | unsigned Offset = N->getOpcode() - AMDGPUISD::CVT_F32_UBYTE0; |
2675 | |
2676 | SDValue Src = N->getOperand(0); |
2677 | APInt Demanded = APInt::getBitsSet(32, 8 * Offset, 8 * Offset + 8); |
2678 | |
2679 | APInt KnownZero, KnownOne; |
2680 | TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(), |
2681 | !DCI.isBeforeLegalizeOps()); |
2682 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
2683 | if (TLO.ShrinkDemandedConstant(Src, Demanded) || |
2684 | TLI.SimplifyDemandedBits(Src, Demanded, KnownZero, KnownOne, TLO)) { |
2685 | DCI.CommitTargetLoweringOpt(TLO); |
2686 | } |
2687 | |
2688 | break; |
2689 | } |
2690 | |
2691 | case ISD::UINT_TO_FP: { |
2692 | return performUCharToFloatCombine(N, DCI); |
2693 | } |
2694 | case ISD::FADD: { |
2695 | if (DCI.getDAGCombineLevel() < AfterLegalizeDAG) |
2696 | break; |
2697 | |
2698 | EVT VT = N->getValueType(0); |
2699 | if (VT != MVT::f32) |
2700 | break; |
2701 | |
2702 | // Only do this if we are not trying to support denormals. v_mad_f32 does |
2703 | // not support denormals ever. |
2704 | if (Subtarget->hasFP32Denormals()) |
2705 | break; |
2706 | |
2707 | SDValue LHS = N->getOperand(0); |
2708 | SDValue RHS = N->getOperand(1); |
2709 | |
2710 | // These should really be instruction patterns, but writing patterns with |
2711 | // source modiifiers is a pain. |
2712 | |
2713 | // fadd (fadd (a, a), b) -> mad 2.0, a, b |
2714 | if (LHS.getOpcode() == ISD::FADD) { |
2715 | SDValue A = LHS.getOperand(0); |
2716 | if (A == LHS.getOperand(1)) { |
2717 | const SDValue Two = DAG.getConstantFP(2.0, DL, MVT::f32); |
2718 | return DAG.getNode(ISD::FMAD, DL, VT, Two, A, RHS); |
2719 | } |
2720 | } |
2721 | |
2722 | // fadd (b, fadd (a, a)) -> mad 2.0, a, b |
2723 | if (RHS.getOpcode() == ISD::FADD) { |
2724 | SDValue A = RHS.getOperand(0); |
2725 | if (A == RHS.getOperand(1)) { |
2726 | const SDValue Two = DAG.getConstantFP(2.0, DL, MVT::f32); |
2727 | return DAG.getNode(ISD::FMAD, DL, VT, Two, A, LHS); |
2728 | } |
2729 | } |
2730 | |
2731 | return SDValue(); |
2732 | } |
2733 | case ISD::FSUB: { |
2734 | if (DCI.getDAGCombineLevel() < AfterLegalizeDAG) |
2735 | break; |
2736 | |
2737 | EVT VT = N->getValueType(0); |
2738 | |
2739 | // Try to get the fneg to fold into the source modifier. This undoes generic |
2740 | // DAG combines and folds them into the mad. |
2741 | // |
2742 | // Only do this if we are not trying to support denormals. v_mad_f32 does |
2743 | // not support denormals ever. |
2744 | if (VT == MVT::f32 && |
2745 | !Subtarget->hasFP32Denormals()) { |
2746 | SDValue LHS = N->getOperand(0); |
2747 | SDValue RHS = N->getOperand(1); |
2748 | if (LHS.getOpcode() == ISD::FADD) { |
2749 | // (fsub (fadd a, a), c) -> mad 2.0, a, (fneg c) |
2750 | |
2751 | SDValue A = LHS.getOperand(0); |
2752 | if (A == LHS.getOperand(1)) { |
2753 | const SDValue Two = DAG.getConstantFP(2.0, DL, MVT::f32); |
2754 | SDValue NegRHS = DAG.getNode(ISD::FNEG, DL, VT, RHS); |
2755 | |
2756 | return DAG.getNode(ISD::FMAD, DL, VT, Two, A, NegRHS); |
2757 | } |
2758 | } |
2759 | |
2760 | if (RHS.getOpcode() == ISD::FADD) { |
2761 | // (fsub c, (fadd a, a)) -> mad -2.0, a, c |
2762 | |
2763 | SDValue A = RHS.getOperand(0); |
2764 | if (A == RHS.getOperand(1)) { |
2765 | const SDValue NegTwo = DAG.getConstantFP(-2.0, DL, MVT::f32); |
2766 | return DAG.getNode(ISD::FMAD, DL, VT, NegTwo, A, LHS); |
2767 | } |
2768 | } |
2769 | |
2770 | return SDValue(); |
2771 | } |
2772 | |
2773 | break; |
2774 | } |
2775 | case ISD::LOAD: |
2776 | case ISD::STORE: |
2777 | case ISD::ATOMIC_LOAD: |
2778 | case ISD::ATOMIC_STORE: |
2779 | case ISD::ATOMIC_CMP_SWAP: |
2780 | case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: |
2781 | case ISD::ATOMIC_SWAP: |
2782 | case ISD::ATOMIC_LOAD_ADD: |
2783 | case ISD::ATOMIC_LOAD_SUB: |
2784 | case ISD::ATOMIC_LOAD_AND: |
2785 | case ISD::ATOMIC_LOAD_OR: |
2786 | case ISD::ATOMIC_LOAD_XOR: |
2787 | case ISD::ATOMIC_LOAD_NAND: |
2788 | case ISD::ATOMIC_LOAD_MIN: |
2789 | case ISD::ATOMIC_LOAD_MAX: |
2790 | case ISD::ATOMIC_LOAD_UMIN: |
2791 | case ISD::ATOMIC_LOAD_UMAX: |
2792 | case AMDGPUISD::ATOMIC_INC: |
2793 | case AMDGPUISD::ATOMIC_DEC: { // TODO: Target mem intrinsics. |
2794 | if (DCI.isBeforeLegalize()) |
2795 | break; |
2796 | |
2797 | MemSDNode *MemNode = cast<MemSDNode>(N); |
2798 | SDValue Ptr = MemNode->getBasePtr(); |
2799 | |
2800 | // TODO: We could also do this for multiplies. |
2801 | unsigned AS = MemNode->getAddressSpace(); |
2802 | if (Ptr.getOpcode() == ISD::SHL && AS != AMDGPUAS::PRIVATE_ADDRESS) { |
2803 | SDValue NewPtr = performSHLPtrCombine(Ptr.getNode(), AS, DCI); |
2804 | if (NewPtr) { |
2805 | SmallVector<SDValue, 8> NewOps(MemNode->op_begin(), MemNode->op_end()); |
2806 | |
2807 | NewOps[N->getOpcode() == ISD::STORE ? 2 : 1] = NewPtr; |
2808 | return SDValue(DAG.UpdateNodeOperands(MemNode, NewOps), 0); |
2809 | } |
2810 | } |
2811 | break; |
2812 | } |
2813 | case ISD::AND: |
2814 | return performAndCombine(N, DCI); |
2815 | case ISD::OR: |
2816 | return performOrCombine(N, DCI); |
2817 | case AMDGPUISD::FP_CLASS: |
2818 | return performClassCombine(N, DCI); |
2819 | case ISD::FCANONICALIZE: |
2820 | return performFCanonicalizeCombine(N, DCI); |
2821 | } |
2822 | return AMDGPUTargetLowering::PerformDAGCombine(N, DCI); |
2823 | } |
2824 | |
2825 | /// \brief Analyze the possible immediate value Op |
2826 | /// |
2827 | /// Returns -1 if it isn't an immediate, 0 if it's and inline immediate |
2828 | /// and the immediate value if it's a literal immediate |
2829 | int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const { |
2830 | |
2831 | const SIInstrInfo *TII = |
2832 | static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo()); |
2833 | |
2834 | if (const ConstantSDNode *Node = dyn_cast<ConstantSDNode>(N)) { |
2835 | if (TII->isInlineConstant(Node->getAPIntValue())) |
2836 | return 0; |
2837 | |
2838 | uint64_t Val = Node->getZExtValue(); |
2839 | return isUInt<32>(Val) ? Val : -1; |
2840 | } |
2841 | |
2842 | if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N)) { |
2843 | if (TII->isInlineConstant(Node->getValueAPF().bitcastToAPInt())) |
2844 | return 0; |
2845 | |
2846 | if (Node->getValueType(0) == MVT::f32) |
2847 | return FloatToBits(Node->getValueAPF().convertToFloat()); |
2848 | |
2849 | return -1; |
2850 | } |
2851 | |
2852 | return -1; |
2853 | } |
2854 | |
2855 | /// \brief Helper function for adjustWritemask |
2856 | static unsigned SubIdx2Lane(unsigned Idx) { |
2857 | switch (Idx) { |
2858 | default: return 0; |
2859 | case AMDGPU::sub0: return 0; |
2860 | case AMDGPU::sub1: return 1; |
2861 | case AMDGPU::sub2: return 2; |
2862 | case AMDGPU::sub3: return 3; |
2863 | } |
2864 | } |
2865 | |
2866 | /// \brief Adjust the writemask of MIMG instructions |
2867 | void SITargetLowering::adjustWritemask(MachineSDNode *&Node, |
2868 | SelectionDAG &DAG) const { |
2869 | SDNode *Users[4] = { }; |
2870 | unsigned Lane = 0; |
2871 | unsigned DmaskIdx = (Node->getNumOperands() - Node->getNumValues() == 9) ? 2 : 3; |
2872 | unsigned OldDmask = Node->getConstantOperandVal(DmaskIdx); |
2873 | unsigned NewDmask = 0; |
2874 | |
2875 | // Try to figure out the used register components |
2876 | for (SDNode::use_iterator I = Node->use_begin(), E = Node->use_end(); |
2877 | I != E; ++I) { |
2878 | |
2879 | // Abort if we can't understand the usage |
2880 | if (!I->isMachineOpcode() || |
2881 | I->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG) |
2882 | return; |
2883 | |
2884 | // Lane means which subreg of %VGPRa_VGPRb_VGPRc_VGPRd is used. |
2885 | // Note that subregs are packed, i.e. Lane==0 is the first bit set |
2886 | // in OldDmask, so it can be any of X,Y,Z,W; Lane==1 is the second bit |
2887 | // set, etc. |
2888 | Lane = SubIdx2Lane(I->getConstantOperandVal(1)); |
2889 | |
2890 | // Set which texture component corresponds to the lane. |
2891 | unsigned Comp; |
2892 | for (unsigned i = 0, Dmask = OldDmask; i <= Lane; i++) { |
2893 | assert(Dmask)((Dmask) ? static_cast<void> (0) : __assert_fail ("Dmask" , "/tmp/buildd/llvm-toolchain-snapshot-3.9~svn267387/lib/Target/AMDGPU/SIISelLowering.cpp" , 2893, __PRETTY_FUNCTION__)); |
2894 | Comp = countTrailingZeros(Dmask); |
2895 | Dmask &= ~(1 << Comp); |
2896 | } |
2897 | |
2898 | // Abort if we have more than one user per component |
2899 | if (Users[Lane]) |
2900 | return; |
2901 | |
2902 | Users[Lane] = *I; |
2903 | NewDmask |= 1 << Comp; |
2904 | } |
2905 | |
2906 | // Abort if there's no change |
2907 | if (NewDmask == OldDmask) |
2908 | return; |
2909 | |
2910 | // Adjust the writemask in the node |
2911 | std::vector<SDValue> Ops; |
2912 | Ops.insert(Ops.end(), Node->op_begin(), Node->op_begin() + DmaskIdx); |
2913 | Ops.push_back(DAG.getTargetConstant(NewDmask, SDLoc(Node), MVT::i32)); |
2914 | Ops.insert(Ops.end(), Node->op_begin() + DmaskIdx + 1, Node->op_end()); |
2915 | Node = (MachineSDNode*)DAG.UpdateNodeOperands(Node, Ops); |
2916 | |
2917 | // If we only got one lane, replace it with a copy |
2918 | // (if NewDmask has only one bit set...) |
2919 | if (NewDmask && (NewDmask & (NewDmask-1)) == 0) { |
2920 | SDValue RC = DAG.getTargetConstant(AMDGPU::VGPR_32RegClassID, SDLoc(), |
2921 | MVT::i32); |
2922 | SDNode *Copy = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS, |
2923 | SDLoc(), Users[Lane]->getValueType(0), |
2924 | SDValue(Node, 0), RC); |
2925 | DAG.ReplaceAllUsesWith(Users[Lane], Copy); |
2926 | return; |
2927 | } |
2928 | |
2929 | // Update the users of the node with the new indices |
2930 | for (unsigned i = 0, Idx = AMDGPU::sub0; i < 4; ++i) { |
2931 | |
2932 | SDNode *User = Users[i]; |
2933 | if (!User) |
2934 | continue; |
2935 | |
2936 | SDValue Op = DAG.getTargetConstant(Idx, SDLoc(User), MVT::i32); |
2937 | DAG.UpdateNodeOperands(User, User->getOperand(0), Op); |
2938 | |
2939 | switch (Idx) { |
2940 | default: break; |
2941 | case AMDGPU::sub0: Idx = AMDGPU::sub1; break; |
2942 | case AMDGPU::sub1: Idx = AMDGPU::sub2; break; |
2943 | case AMDGPU::sub2: Idx = AMDGPU::sub3; break; |
2944 | } |
2945 | } |
2946 | } |
2947 | |
2948 | static bool isFrameIndexOp(SDValue Op) { |
2949 | if (Op.getOpcode() == ISD::AssertZext) |
2950 | Op = Op.getOperand(0); |
2951 | |
2952 | return isa<FrameIndexSDNode>(Op); |
2953 | } |
2954 | |
2955 | /// \brief Legalize target independent instructions (e.g. INSERT_SUBREG) |
2956 | /// with frame index operands. |
2957 | /// LLVM assumes that inputs are to these instructions are registers. |
2958 | void SITargetLowering::legalizeTargetIndependentNode(SDNode *Node, |
2959 | SelectionDAG &DAG) const { |
2960 | |
2961 | SmallVector<SDValue, 8> Ops; |
2962 | for (unsigned i = 0; i < Node->getNumOperands(); ++i) { |
2963 | if (!isFrameIndexOp(Node->getOperand(i))) { |
2964 | Ops.push_back(Node->getOperand(i)); |
2965 | continue; |
2966 | } |
2967 | |
2968 | SDLoc DL(Node); |
2969 | Ops.push_back(SDValue(DAG.getMachineNode(AMDGPU::S_MOV_B32, DL, |
2970 | Node->getOperand(i).getValueType(), |
2971 | Node->getOperand(i)), 0)); |
2972 | } |
2973 | |
2974 | DAG.UpdateNodeOperands(Node, Ops); |
2975 | } |
2976 | |
2977 | /// \brief Fold the instructions after selecting them. |
2978 | SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node, |
2979 | SelectionDAG &DAG) const { |
2980 | const SIInstrInfo *TII = |
2981 | static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo()); |
2982 | unsigned Opcode = Node->getMachineOpcode(); |
2983 | |
2984 | if (TII->isMIMG(Opcode) && !TII->get(Opcode).mayStore()) |
2985 | adjustWritemask(Node, DAG); |
2986 | |
2987 | if (Opcode == AMDGPU::INSERT_SUBREG || |
2988 | Opcode == AMDGPU::REG_SEQUENCE) { |
2989 | legalizeTargetIndependentNode(Node, DAG); |
2990 | return Node; |
2991 | } |
2992 | return Node; |
2993 | } |
2994 | |
2995 | /// \brief Assign the register class depending on the number of |
2996 | /// bits set in the writemask |
2997 | void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI, |
2998 | SDNode *Node) const { |
2999 | const SIInstrInfo *TII = |
3000 | static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo()); |
3001 | |
3002 | MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo(); |
3003 | |
3004 | if (TII->isVOP3(MI->getOpcode())) { |
3005 | // Make sure constant bus requirements are respected. |
3006 | TII->legalizeOperandsVOP3(MRI, MI); |
3007 | return; |
3008 | } |
3009 | |
3010 | if (TII->isMIMG(*MI)) { |
3011 | unsigned VReg = MI->getOperand(0).getReg(); |
3012 | unsigned DmaskIdx = MI->getNumOperands() == 12 ? 3 : 4; |
3013 | unsigned Writemask = MI->getOperand(DmaskIdx).getImm(); |
3014 | unsigned BitsSet = 0; |
3015 | for (unsigned i = 0; i < 4; ++i) |
3016 | BitsSet += Writemask & (1 << i) ? 1 : 0; |
3017 | |
3018 | const TargetRegisterClass *RC; |
3019 | switch (BitsSet) { |
3020 | default: return; |
3021 | case 1: RC = &AMDGPU::VGPR_32RegClass; break; |
3022 | case 2: RC = &AMDGPU::VReg_64RegClass; break; |
3023 | case 3: RC = &AMDGPU::VReg_96RegClass; break; |
3024 | } |
3025 | |
3026 | unsigned NewOpcode = TII->getMaskedMIMGOp(MI->getOpcode(), BitsSet); |
3027 | MI->setDesc(TII->get(NewOpcode)); |
3028 | MRI.setRegClass(VReg, RC); |
3029 | return; |
3030 | } |
3031 | |
3032 | // Replace unused atomics with the no return version. |
3033 | int NoRetAtomicOp = AMDGPU::getAtomicNoRetOp(MI->getOpcode()); |
3034 | if (NoRetAtomicOp != -1) { |
3035 | if (!Node->hasAnyUseOfValue(0)) { |
3036 | MI->setDesc(TII->get(NoRetAtomicOp)); |
3037 | MI->RemoveOperand(0); |
3038 | return; |
3039 | } |
3040 | |
3041 | // For mubuf_atomic_cmpswap, we need to have tablegen use an extract_subreg |
3042 | // instruction, because the return type of these instructions is a vec2 of |
3043 | // the memory type, so it can be tied to the input operand. |
3044 | // This means these instructions always have a use, so we need to add a |
3045 | // special case to check if the atomic has only one extract_subreg use, |
3046 | // which itself has no uses. |
3047 | if ((Node->hasNUsesOfValue(1, 0) && |
3048 | Node->use_begin()->isMachineOpcode() && |
3049 | Node->use_begin()->getMachineOpcode() == AMDGPU::EXTRACT_SUBREG && |
3050 | !Node->use_begin()->hasAnyUseOfValue(0))) { |
3051 | unsigned Def = MI->getOperand(0).getReg(); |
3052 | |
3053 | // Change this into a noret atomic. |
3054 | MI->setDesc(TII->get(NoRetAtomicOp)); |
3055 | MI->RemoveOperand(0); |
3056 | |
3057 | // If we only remove the def operand from the atomic instruction, the |
3058 | // extract_subreg will be left with a use of a vreg without a def. |
3059 | // So we need to insert an implicit_def to avoid machine verifier |
3060 | // errors. |
3061 | BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), |
3062 | TII->get(AMDGPU::IMPLICIT_DEF), Def); |
3063 | } |
3064 | return; |
3065 | } |
3066 | } |
3067 | |
3068 | static SDValue buildSMovImm32(SelectionDAG &DAG, SDLoc DL, uint64_t Val) { |
3069 | SDValue K = DAG.getTargetConstant(Val, DL, MVT::i32); |
3070 | return SDValue(DAG.getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, K), 0); |
3071 | } |
3072 | |
3073 | MachineSDNode *SITargetLowering::wrapAddr64Rsrc(SelectionDAG &DAG, |
3074 | SDLoc DL, |
3075 | SDValue Ptr) const { |
3076 | const SIInstrInfo *TII = |
3077 | static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo()); |
3078 | |
3079 | // Build the half of the subregister with the constants before building the |
3080 | // full 128-bit register. If we are building multiple resource descriptors, |
3081 | // this will allow CSEing of the 2-component register. |
3082 | const SDValue Ops0[] = { |
3083 | DAG.getTargetConstant(AMDGPU::SGPR_64RegClassID, DL, MVT::i32), |
3084 | buildSMovImm32(DAG, DL, 0), |
3085 | DAG.getTargetConstant(AMDGPU::sub0, DL, MVT::i32), |
3086 | buildSMovImm32(DAG, DL, TII->getDefaultRsrcDataFormat() >> 32), |
3087 | DAG.getTargetConstant(AMDGPU::sub1, DL, MVT::i32) |
3088 | }; |
3089 | |
3090 | SDValue SubRegHi = SDValue(DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL, |
3091 | MVT::v2i32, Ops0), 0); |
3092 | |
3093 | // Combine the constants and the pointer. |
3094 | const SDValue Ops1[] = { |
3095 | DAG.getTargetConstant(AMDGPU::SReg_128RegClassID, DL, MVT::i32), |
3096 | Ptr, |
3097 | DAG.getTargetConstant(AMDGPU::sub0_sub1, DL, MVT::i32), |
3098 | SubRegHi, |
3099 | DAG.getTargetConstant(AMDGPU::sub2_sub3, DL, MVT::i32) |
3100 | }; |
3101 | |
3102 | return DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL, MVT::v4i32, Ops1); |
3103 | } |
3104 | |
3105 | /// \brief Return a resource descriptor with the 'Add TID' bit enabled |
3106 | /// The TID (Thread ID) is multiplied by the stride value (bits [61:48] |
3107 | /// of the resource descriptor) to create an offset, which is added to |
3108 | /// the resource pointer. |
3109 | MachineSDNode *SITargetLowering::buildRSRC(SelectionDAG &DAG, |
3110 | SDLoc DL, |
3111 | SDValue Ptr, |
3112 | uint32_t RsrcDword1, |
3113 | uint64_t RsrcDword2And3) const { |
3114 | SDValue PtrLo = DAG.getTargetExtractSubreg(AMDGPU::sub0, DL, MVT::i32, Ptr); |
3115 | SDValue PtrHi = DAG.getTargetExtractSubreg(AMDGPU::sub1, DL, MVT::i32, Ptr); |
3116 | if (RsrcDword1) { |
3117 | PtrHi = SDValue(DAG.getMachineNode(AMDGPU::S_OR_B32, DL, MVT::i32, PtrHi, |
3118 | DAG.getConstant(RsrcDword1, DL, MVT::i32)), |
3119 | 0); |
3120 | } |
3121 | |
3122 | SDValue DataLo = buildSMovImm32(DAG, DL, |
3123 | RsrcDword2And3 & UINT64_C(0xFFFFFFFF)0xFFFFFFFFUL); |
3124 | SDValue DataHi = buildSMovImm32(DAG, DL, RsrcDword2And3 >> 32); |
3125 | |
3126 | const SDValue Ops[] = { |
3127 | DAG.getTargetConstant(AMDGPU::SReg_128RegClassID, DL, MVT::i32), |
3128 | PtrLo, |
3129 | DAG.getTargetConstant(AMDGPU::sub0, DL, MVT::i32), |
3130 | PtrHi, |
3131 | DAG.getTargetConstant(AMDGPU::sub1, DL, MVT::i32), |
3132 | DataLo, |
3133 | DAG.getTargetConstant(AMDGPU::sub2, DL, MVT::i32), |
3134 | DataHi, |
3135 | DAG.getTargetConstant(AMDGPU::sub3, DL, MVT::i32) |
3136 | }; |
3137 | |
3138 | return DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL, MVT::v4i32, Ops); |
3139 | } |
3140 | |
3141 | SDValue SITargetLowering::CreateLiveInRegister(SelectionDAG &DAG, |
3142 | const TargetRegisterClass *RC, |
3143 | unsigned Reg, EVT VT) const { |
3144 | SDValue VReg = AMDGPUTargetLowering::CreateLiveInRegister(DAG, RC, Reg, VT); |
3145 | |
3146 | return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(DAG.getEntryNode()), |
3147 | cast<RegisterSDNode>(VReg)->getReg(), VT); |
3148 | } |
3149 | |
3150 | //===----------------------------------------------------------------------===// |
3151 | // SI Inline Assembly Support |
3152 | //===----------------------------------------------------------------------===// |
3153 | |
3154 | std::pair<unsigned, const TargetRegisterClass *> |
3155 | SITargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, |
3156 | StringRef Constraint, |
3157 | MVT VT) const { |
3158 | |
3159 | if (Constraint.size() == 1) { |
3160 | switch (Constraint[0]) { |
3161 | case 's': |
3162 | case 'r': |
3163 | switch (VT.getSizeInBits()) { |
3164 | default: |
3165 | return std::make_pair(0U, nullptr); |
3166 | case 32: |
3167 | return std::make_pair(0U, &AMDGPU::SGPR_32RegClass); |
3168 | case 64: |
3169 | return std::make_pair(0U, &AMDGPU::SGPR_64RegClass); |
3170 | case 128: |
3171 | return std::make_pair(0U, &AMDGPU::SReg_128RegClass); |
3172 | case 256: |
3173 | return std::make_pair(0U, &AMDGPU::SReg_256RegClass); |
3174 | } |
3175 | |
3176 | case 'v': |
3177 | switch (VT.getSizeInBits()) { |
3178 | default: |
3179 | return std::make_pair(0U, nullptr); |
3180 | case 32: |
3181 | return std::make_pair(0U, &AMDGPU::VGPR_32RegClass); |
3182 | case 64: |
3183 | return std::make_pair(0U, &AMDGPU::VReg_64RegClass); |
3184 | case 96: |
3185 | return std::make_pair(0U, &AMDGPU::VReg_96RegClass); |
3186 | case 128: |
3187 | return std::make_pair(0U, &AMDGPU::VReg_128RegClass); |
3188 | case 256: |
3189 | return std::make_pair(0U, &AMDGPU::VReg_256RegClass); |
3190 | case 512: |
3191 | return std::make_pair(0U, &AMDGPU::VReg_512RegClass); |
3192 | } |
3193 | } |
3194 | } |
3195 | |
3196 | if (Constraint.size() > 1) { |
3197 | const TargetRegisterClass *RC = nullptr; |
3198 | if (Constraint[1] == 'v') { |
3199 | RC = &AMDGPU::VGPR_32RegClass; |
3200 | } else if (Constraint[1] == 's') { |
3201 | RC = &AMDGPU::SGPR_32RegClass; |
3202 | } |
3203 | |
3204 | if (RC) { |
3205 | uint32_t Idx; |
3206 | bool Failed = Constraint.substr(2).getAsInteger(10, Idx); |
3207 | if (!Failed && Idx < RC->getNumRegs()) |
3208 | return std::make_pair(RC->getRegister(Idx), RC); |
3209 | } |
3210 | } |
3211 | return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); |
3212 | } |
3213 | |
3214 | SITargetLowering::ConstraintType |
3215 | SITargetLowering::getConstraintType(StringRef Constraint) const { |
3216 | if (Constraint.size() == 1) { |
3217 | switch (Constraint[0]) { |
3218 | default: break; |
3219 | case 's': |
3220 | case 'v': |
3221 | return C_RegisterClass; |
3222 | } |
3223 | } |
3224 | return TargetLowering::getConstraintType(Constraint); |
3225 | } |