/Users/buildslave/jenkins/sharedspace/clang-stage2-coverage-R@2/llvm/lib/Target/NVPTX/NVPTXISelLowering.h
Line | Count | Source (jump to first uncovered line) |
1 | | //===-- NVPTXISelLowering.h - NVPTX DAG Lowering Interface ------*- C++ -*-===// |
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 | | // This file defines the interfaces that NVPTX uses to lower LLVM code into a |
11 | | // selection DAG. |
12 | | // |
13 | | //===----------------------------------------------------------------------===// |
14 | | |
15 | | #ifndef LLVM_LIB_TARGET_NVPTX_NVPTXISELLOWERING_H |
16 | | #define LLVM_LIB_TARGET_NVPTX_NVPTXISELLOWERING_H |
17 | | |
18 | | #include "NVPTX.h" |
19 | | #include "llvm/CodeGen/SelectionDAG.h" |
20 | | #include "llvm/Target/TargetLowering.h" |
21 | | |
22 | | namespace llvm { |
23 | | namespace NVPTXISD { |
24 | | enum NodeType : unsigned { |
25 | | // Start the numbering from where ISD NodeType finishes. |
26 | | FIRST_NUMBER = ISD::BUILTIN_OP_END, |
27 | | Wrapper, |
28 | | CALL, |
29 | | RET_FLAG, |
30 | | LOAD_PARAM, |
31 | | DeclareParam, |
32 | | DeclareScalarParam, |
33 | | DeclareRetParam, |
34 | | DeclareRet, |
35 | | DeclareScalarRet, |
36 | | PrintCall, |
37 | | PrintConvergentCall, |
38 | | PrintCallUni, |
39 | | PrintConvergentCallUni, |
40 | | CallArgBegin, |
41 | | CallArg, |
42 | | LastCallArg, |
43 | | CallArgEnd, |
44 | | CallVoid, |
45 | | CallVal, |
46 | | CallSymbol, |
47 | | Prototype, |
48 | | MoveParam, |
49 | | PseudoUseParam, |
50 | | RETURN, |
51 | | CallSeqBegin, |
52 | | CallSeqEnd, |
53 | | CallPrototype, |
54 | | FUN_SHFL_CLAMP, |
55 | | FUN_SHFR_CLAMP, |
56 | | MUL_WIDE_SIGNED, |
57 | | MUL_WIDE_UNSIGNED, |
58 | | IMAD, |
59 | | SETP_F16X2, |
60 | | Dummy, |
61 | | |
62 | | LoadV2 = ISD::FIRST_TARGET_MEMORY_OPCODE, |
63 | | LoadV4, |
64 | | LDGV2, // LDG.v2 |
65 | | LDGV4, // LDG.v4 |
66 | | LDUV2, // LDU.v2 |
67 | | LDUV4, // LDU.v4 |
68 | | StoreV2, |
69 | | StoreV4, |
70 | | LoadParam, |
71 | | LoadParamV2, |
72 | | LoadParamV4, |
73 | | StoreParam, |
74 | | StoreParamV2, |
75 | | StoreParamV4, |
76 | | StoreParamS32, // to sext and store a <32bit value, not used currently |
77 | | StoreParamU32, // to zext and store a <32bit value, not used currently |
78 | | StoreRetval, |
79 | | StoreRetvalV2, |
80 | | StoreRetvalV4, |
81 | | |
82 | | // Texture intrinsics |
83 | | Tex1DFloatS32, |
84 | | Tex1DFloatFloat, |
85 | | Tex1DFloatFloatLevel, |
86 | | Tex1DFloatFloatGrad, |
87 | | Tex1DS32S32, |
88 | | Tex1DS32Float, |
89 | | Tex1DS32FloatLevel, |
90 | | Tex1DS32FloatGrad, |
91 | | Tex1DU32S32, |
92 | | Tex1DU32Float, |
93 | | Tex1DU32FloatLevel, |
94 | | Tex1DU32FloatGrad, |
95 | | Tex1DArrayFloatS32, |
96 | | Tex1DArrayFloatFloat, |
97 | | Tex1DArrayFloatFloatLevel, |
98 | | Tex1DArrayFloatFloatGrad, |
99 | | Tex1DArrayS32S32, |
100 | | Tex1DArrayS32Float, |
101 | | Tex1DArrayS32FloatLevel, |
102 | | Tex1DArrayS32FloatGrad, |
103 | | Tex1DArrayU32S32, |
104 | | Tex1DArrayU32Float, |
105 | | Tex1DArrayU32FloatLevel, |
106 | | Tex1DArrayU32FloatGrad, |
107 | | Tex2DFloatS32, |
108 | | Tex2DFloatFloat, |
109 | | Tex2DFloatFloatLevel, |
110 | | Tex2DFloatFloatGrad, |
111 | | Tex2DS32S32, |
112 | | Tex2DS32Float, |
113 | | Tex2DS32FloatLevel, |
114 | | Tex2DS32FloatGrad, |
115 | | Tex2DU32S32, |
116 | | Tex2DU32Float, |
117 | | Tex2DU32FloatLevel, |
118 | | Tex2DU32FloatGrad, |
119 | | Tex2DArrayFloatS32, |
120 | | Tex2DArrayFloatFloat, |
121 | | Tex2DArrayFloatFloatLevel, |
122 | | Tex2DArrayFloatFloatGrad, |
123 | | Tex2DArrayS32S32, |
124 | | Tex2DArrayS32Float, |
125 | | Tex2DArrayS32FloatLevel, |
126 | | Tex2DArrayS32FloatGrad, |
127 | | Tex2DArrayU32S32, |
128 | | Tex2DArrayU32Float, |
129 | | Tex2DArrayU32FloatLevel, |
130 | | Tex2DArrayU32FloatGrad, |
131 | | Tex3DFloatS32, |
132 | | Tex3DFloatFloat, |
133 | | Tex3DFloatFloatLevel, |
134 | | Tex3DFloatFloatGrad, |
135 | | Tex3DS32S32, |
136 | | Tex3DS32Float, |
137 | | Tex3DS32FloatLevel, |
138 | | Tex3DS32FloatGrad, |
139 | | Tex3DU32S32, |
140 | | Tex3DU32Float, |
141 | | Tex3DU32FloatLevel, |
142 | | Tex3DU32FloatGrad, |
143 | | TexCubeFloatFloat, |
144 | | TexCubeFloatFloatLevel, |
145 | | TexCubeS32Float, |
146 | | TexCubeS32FloatLevel, |
147 | | TexCubeU32Float, |
148 | | TexCubeU32FloatLevel, |
149 | | TexCubeArrayFloatFloat, |
150 | | TexCubeArrayFloatFloatLevel, |
151 | | TexCubeArrayS32Float, |
152 | | TexCubeArrayS32FloatLevel, |
153 | | TexCubeArrayU32Float, |
154 | | TexCubeArrayU32FloatLevel, |
155 | | Tld4R2DFloatFloat, |
156 | | Tld4G2DFloatFloat, |
157 | | Tld4B2DFloatFloat, |
158 | | Tld4A2DFloatFloat, |
159 | | Tld4R2DS64Float, |
160 | | Tld4G2DS64Float, |
161 | | Tld4B2DS64Float, |
162 | | Tld4A2DS64Float, |
163 | | Tld4R2DU64Float, |
164 | | Tld4G2DU64Float, |
165 | | Tld4B2DU64Float, |
166 | | Tld4A2DU64Float, |
167 | | TexUnified1DFloatS32, |
168 | | TexUnified1DFloatFloat, |
169 | | TexUnified1DFloatFloatLevel, |
170 | | TexUnified1DFloatFloatGrad, |
171 | | TexUnified1DS32S32, |
172 | | TexUnified1DS32Float, |
173 | | TexUnified1DS32FloatLevel, |
174 | | TexUnified1DS32FloatGrad, |
175 | | TexUnified1DU32S32, |
176 | | TexUnified1DU32Float, |
177 | | TexUnified1DU32FloatLevel, |
178 | | TexUnified1DU32FloatGrad, |
179 | | TexUnified1DArrayFloatS32, |
180 | | TexUnified1DArrayFloatFloat, |
181 | | TexUnified1DArrayFloatFloatLevel, |
182 | | TexUnified1DArrayFloatFloatGrad, |
183 | | TexUnified1DArrayS32S32, |
184 | | TexUnified1DArrayS32Float, |
185 | | TexUnified1DArrayS32FloatLevel, |
186 | | TexUnified1DArrayS32FloatGrad, |
187 | | TexUnified1DArrayU32S32, |
188 | | TexUnified1DArrayU32Float, |
189 | | TexUnified1DArrayU32FloatLevel, |
190 | | TexUnified1DArrayU32FloatGrad, |
191 | | TexUnified2DFloatS32, |
192 | | TexUnified2DFloatFloat, |
193 | | TexUnified2DFloatFloatLevel, |
194 | | TexUnified2DFloatFloatGrad, |
195 | | TexUnified2DS32S32, |
196 | | TexUnified2DS32Float, |
197 | | TexUnified2DS32FloatLevel, |
198 | | TexUnified2DS32FloatGrad, |
199 | | TexUnified2DU32S32, |
200 | | TexUnified2DU32Float, |
201 | | TexUnified2DU32FloatLevel, |
202 | | TexUnified2DU32FloatGrad, |
203 | | TexUnified2DArrayFloatS32, |
204 | | TexUnified2DArrayFloatFloat, |
205 | | TexUnified2DArrayFloatFloatLevel, |
206 | | TexUnified2DArrayFloatFloatGrad, |
207 | | TexUnified2DArrayS32S32, |
208 | | TexUnified2DArrayS32Float, |
209 | | TexUnified2DArrayS32FloatLevel, |
210 | | TexUnified2DArrayS32FloatGrad, |
211 | | TexUnified2DArrayU32S32, |
212 | | TexUnified2DArrayU32Float, |
213 | | TexUnified2DArrayU32FloatLevel, |
214 | | TexUnified2DArrayU32FloatGrad, |
215 | | TexUnified3DFloatS32, |
216 | | TexUnified3DFloatFloat, |
217 | | TexUnified3DFloatFloatLevel, |
218 | | TexUnified3DFloatFloatGrad, |
219 | | TexUnified3DS32S32, |
220 | | TexUnified3DS32Float, |
221 | | TexUnified3DS32FloatLevel, |
222 | | TexUnified3DS32FloatGrad, |
223 | | TexUnified3DU32S32, |
224 | | TexUnified3DU32Float, |
225 | | TexUnified3DU32FloatLevel, |
226 | | TexUnified3DU32FloatGrad, |
227 | | TexUnifiedCubeFloatFloat, |
228 | | TexUnifiedCubeFloatFloatLevel, |
229 | | TexUnifiedCubeS32Float, |
230 | | TexUnifiedCubeS32FloatLevel, |
231 | | TexUnifiedCubeU32Float, |
232 | | TexUnifiedCubeU32FloatLevel, |
233 | | TexUnifiedCubeArrayFloatFloat, |
234 | | TexUnifiedCubeArrayFloatFloatLevel, |
235 | | TexUnifiedCubeArrayS32Float, |
236 | | TexUnifiedCubeArrayS32FloatLevel, |
237 | | TexUnifiedCubeArrayU32Float, |
238 | | TexUnifiedCubeArrayU32FloatLevel, |
239 | | Tld4UnifiedR2DFloatFloat, |
240 | | Tld4UnifiedG2DFloatFloat, |
241 | | Tld4UnifiedB2DFloatFloat, |
242 | | Tld4UnifiedA2DFloatFloat, |
243 | | Tld4UnifiedR2DS64Float, |
244 | | Tld4UnifiedG2DS64Float, |
245 | | Tld4UnifiedB2DS64Float, |
246 | | Tld4UnifiedA2DS64Float, |
247 | | Tld4UnifiedR2DU64Float, |
248 | | Tld4UnifiedG2DU64Float, |
249 | | Tld4UnifiedB2DU64Float, |
250 | | Tld4UnifiedA2DU64Float, |
251 | | |
252 | | // Surface intrinsics |
253 | | Suld1DI8Clamp, |
254 | | Suld1DI16Clamp, |
255 | | Suld1DI32Clamp, |
256 | | Suld1DI64Clamp, |
257 | | Suld1DV2I8Clamp, |
258 | | Suld1DV2I16Clamp, |
259 | | Suld1DV2I32Clamp, |
260 | | Suld1DV2I64Clamp, |
261 | | Suld1DV4I8Clamp, |
262 | | Suld1DV4I16Clamp, |
263 | | Suld1DV4I32Clamp, |
264 | | |
265 | | Suld1DArrayI8Clamp, |
266 | | Suld1DArrayI16Clamp, |
267 | | Suld1DArrayI32Clamp, |
268 | | Suld1DArrayI64Clamp, |
269 | | Suld1DArrayV2I8Clamp, |
270 | | Suld1DArrayV2I16Clamp, |
271 | | Suld1DArrayV2I32Clamp, |
272 | | Suld1DArrayV2I64Clamp, |
273 | | Suld1DArrayV4I8Clamp, |
274 | | Suld1DArrayV4I16Clamp, |
275 | | Suld1DArrayV4I32Clamp, |
276 | | |
277 | | Suld2DI8Clamp, |
278 | | Suld2DI16Clamp, |
279 | | Suld2DI32Clamp, |
280 | | Suld2DI64Clamp, |
281 | | Suld2DV2I8Clamp, |
282 | | Suld2DV2I16Clamp, |
283 | | Suld2DV2I32Clamp, |
284 | | Suld2DV2I64Clamp, |
285 | | Suld2DV4I8Clamp, |
286 | | Suld2DV4I16Clamp, |
287 | | Suld2DV4I32Clamp, |
288 | | |
289 | | Suld2DArrayI8Clamp, |
290 | | Suld2DArrayI16Clamp, |
291 | | Suld2DArrayI32Clamp, |
292 | | Suld2DArrayI64Clamp, |
293 | | Suld2DArrayV2I8Clamp, |
294 | | Suld2DArrayV2I16Clamp, |
295 | | Suld2DArrayV2I32Clamp, |
296 | | Suld2DArrayV2I64Clamp, |
297 | | Suld2DArrayV4I8Clamp, |
298 | | Suld2DArrayV4I16Clamp, |
299 | | Suld2DArrayV4I32Clamp, |
300 | | |
301 | | Suld3DI8Clamp, |
302 | | Suld3DI16Clamp, |
303 | | Suld3DI32Clamp, |
304 | | Suld3DI64Clamp, |
305 | | Suld3DV2I8Clamp, |
306 | | Suld3DV2I16Clamp, |
307 | | Suld3DV2I32Clamp, |
308 | | Suld3DV2I64Clamp, |
309 | | Suld3DV4I8Clamp, |
310 | | Suld3DV4I16Clamp, |
311 | | Suld3DV4I32Clamp, |
312 | | |
313 | | Suld1DI8Trap, |
314 | | Suld1DI16Trap, |
315 | | Suld1DI32Trap, |
316 | | Suld1DI64Trap, |
317 | | Suld1DV2I8Trap, |
318 | | Suld1DV2I16Trap, |
319 | | Suld1DV2I32Trap, |
320 | | Suld1DV2I64Trap, |
321 | | Suld1DV4I8Trap, |
322 | | Suld1DV4I16Trap, |
323 | | Suld1DV4I32Trap, |
324 | | |
325 | | Suld1DArrayI8Trap, |
326 | | Suld1DArrayI16Trap, |
327 | | Suld1DArrayI32Trap, |
328 | | Suld1DArrayI64Trap, |
329 | | Suld1DArrayV2I8Trap, |
330 | | Suld1DArrayV2I16Trap, |
331 | | Suld1DArrayV2I32Trap, |
332 | | Suld1DArrayV2I64Trap, |
333 | | Suld1DArrayV4I8Trap, |
334 | | Suld1DArrayV4I16Trap, |
335 | | Suld1DArrayV4I32Trap, |
336 | | |
337 | | Suld2DI8Trap, |
338 | | Suld2DI16Trap, |
339 | | Suld2DI32Trap, |
340 | | Suld2DI64Trap, |
341 | | Suld2DV2I8Trap, |
342 | | Suld2DV2I16Trap, |
343 | | Suld2DV2I32Trap, |
344 | | Suld2DV2I64Trap, |
345 | | Suld2DV4I8Trap, |
346 | | Suld2DV4I16Trap, |
347 | | Suld2DV4I32Trap, |
348 | | |
349 | | Suld2DArrayI8Trap, |
350 | | Suld2DArrayI16Trap, |
351 | | Suld2DArrayI32Trap, |
352 | | Suld2DArrayI64Trap, |
353 | | Suld2DArrayV2I8Trap, |
354 | | Suld2DArrayV2I16Trap, |
355 | | Suld2DArrayV2I32Trap, |
356 | | Suld2DArrayV2I64Trap, |
357 | | Suld2DArrayV4I8Trap, |
358 | | Suld2DArrayV4I16Trap, |
359 | | Suld2DArrayV4I32Trap, |
360 | | |
361 | | Suld3DI8Trap, |
362 | | Suld3DI16Trap, |
363 | | Suld3DI32Trap, |
364 | | Suld3DI64Trap, |
365 | | Suld3DV2I8Trap, |
366 | | Suld3DV2I16Trap, |
367 | | Suld3DV2I32Trap, |
368 | | Suld3DV2I64Trap, |
369 | | Suld3DV4I8Trap, |
370 | | Suld3DV4I16Trap, |
371 | | Suld3DV4I32Trap, |
372 | | |
373 | | Suld1DI8Zero, |
374 | | Suld1DI16Zero, |
375 | | Suld1DI32Zero, |
376 | | Suld1DI64Zero, |
377 | | Suld1DV2I8Zero, |
378 | | Suld1DV2I16Zero, |
379 | | Suld1DV2I32Zero, |
380 | | Suld1DV2I64Zero, |
381 | | Suld1DV4I8Zero, |
382 | | Suld1DV4I16Zero, |
383 | | Suld1DV4I32Zero, |
384 | | |
385 | | Suld1DArrayI8Zero, |
386 | | Suld1DArrayI16Zero, |
387 | | Suld1DArrayI32Zero, |
388 | | Suld1DArrayI64Zero, |
389 | | Suld1DArrayV2I8Zero, |
390 | | Suld1DArrayV2I16Zero, |
391 | | Suld1DArrayV2I32Zero, |
392 | | Suld1DArrayV2I64Zero, |
393 | | Suld1DArrayV4I8Zero, |
394 | | Suld1DArrayV4I16Zero, |
395 | | Suld1DArrayV4I32Zero, |
396 | | |
397 | | Suld2DI8Zero, |
398 | | Suld2DI16Zero, |
399 | | Suld2DI32Zero, |
400 | | Suld2DI64Zero, |
401 | | Suld2DV2I8Zero, |
402 | | Suld2DV2I16Zero, |
403 | | Suld2DV2I32Zero, |
404 | | Suld2DV2I64Zero, |
405 | | Suld2DV4I8Zero, |
406 | | Suld2DV4I16Zero, |
407 | | Suld2DV4I32Zero, |
408 | | |
409 | | Suld2DArrayI8Zero, |
410 | | Suld2DArrayI16Zero, |
411 | | Suld2DArrayI32Zero, |
412 | | Suld2DArrayI64Zero, |
413 | | Suld2DArrayV2I8Zero, |
414 | | Suld2DArrayV2I16Zero, |
415 | | Suld2DArrayV2I32Zero, |
416 | | Suld2DArrayV2I64Zero, |
417 | | Suld2DArrayV4I8Zero, |
418 | | Suld2DArrayV4I16Zero, |
419 | | Suld2DArrayV4I32Zero, |
420 | | |
421 | | Suld3DI8Zero, |
422 | | Suld3DI16Zero, |
423 | | Suld3DI32Zero, |
424 | | Suld3DI64Zero, |
425 | | Suld3DV2I8Zero, |
426 | | Suld3DV2I16Zero, |
427 | | Suld3DV2I32Zero, |
428 | | Suld3DV2I64Zero, |
429 | | Suld3DV4I8Zero, |
430 | | Suld3DV4I16Zero, |
431 | | Suld3DV4I32Zero |
432 | | }; |
433 | | } |
434 | | |
435 | | class NVPTXSubtarget; |
436 | | |
437 | | //===--------------------------------------------------------------------===// |
438 | | // TargetLowering Implementation |
439 | | //===--------------------------------------------------------------------===// |
440 | | class NVPTXTargetLowering : public TargetLowering { |
441 | | public: |
442 | | explicit NVPTXTargetLowering(const NVPTXTargetMachine &TM, |
443 | | const NVPTXSubtarget &STI); |
444 | | SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override; |
445 | | |
446 | | SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const; |
447 | | |
448 | | const char *getTargetNodeName(unsigned Opcode) const override; |
449 | | |
450 | | bool getTgtMemIntrinsic(IntrinsicInfo &Info, const CallInst &I, |
451 | | unsigned Intrinsic) const override; |
452 | | |
453 | | /// isLegalAddressingMode - Return true if the addressing mode represented |
454 | | /// by AM is legal for this target, for a load/store of the specified type |
455 | | /// Used to guide target specific optimizations, like loop strength |
456 | | /// reduction (LoopStrengthReduce.cpp) and memory optimization for |
457 | | /// address mode (CodeGenPrepare.cpp) |
458 | | bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, Type *Ty, |
459 | | unsigned AS, |
460 | | Instruction *I = nullptr) const override; |
461 | | |
462 | 98 | bool isTruncateFree(Type *SrcTy, Type *DstTy) const override { |
463 | 98 | // Truncating 64-bit to 32-bit is free in SASS. |
464 | 98 | if (!SrcTy->isIntegerTy() || 98 !DstTy->isIntegerTy()98 ) |
465 | 0 | return false; |
466 | 98 | return SrcTy->getPrimitiveSizeInBits() == 64 && |
467 | 51 | DstTy->getPrimitiveSizeInBits() == 32; |
468 | 98 | } |
469 | | |
470 | | EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Ctx, |
471 | 620 | EVT VT) const override { |
472 | 620 | if (VT.isVector()) |
473 | 0 | return EVT::getVectorVT(Ctx, MVT::i1, VT.getVectorNumElements()); |
474 | 620 | return MVT::i1; |
475 | 620 | } |
476 | | |
477 | | ConstraintType getConstraintType(StringRef Constraint) const override; |
478 | | std::pair<unsigned, const TargetRegisterClass *> |
479 | | getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, |
480 | | StringRef Constraint, MVT VT) const override; |
481 | | |
482 | | SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, |
483 | | bool isVarArg, |
484 | | const SmallVectorImpl<ISD::InputArg> &Ins, |
485 | | const SDLoc &dl, SelectionDAG &DAG, |
486 | | SmallVectorImpl<SDValue> &InVals) const override; |
487 | | |
488 | | SDValue LowerCall(CallLoweringInfo &CLI, |
489 | | SmallVectorImpl<SDValue> &InVals) const override; |
490 | | |
491 | | std::string getPrototype(const DataLayout &DL, Type *, const ArgListTy &, |
492 | | const SmallVectorImpl<ISD::OutputArg> &, |
493 | | unsigned retAlignment, |
494 | | ImmutableCallSite CS) const; |
495 | | |
496 | | SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
497 | | const SmallVectorImpl<ISD::OutputArg> &Outs, |
498 | | const SmallVectorImpl<SDValue> &OutVals, const SDLoc &dl, |
499 | | SelectionDAG &DAG) const override; |
500 | | |
501 | | void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint, |
502 | | std::vector<SDValue> &Ops, |
503 | | SelectionDAG &DAG) const override; |
504 | | |
505 | | const NVPTXTargetMachine *nvTM; |
506 | | |
507 | | // PTX always uses 32-bit shift amounts |
508 | 632 | MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override { |
509 | 632 | return MVT::i32; |
510 | 632 | } |
511 | | |
512 | | TargetLoweringBase::LegalizeTypeAction |
513 | | getPreferredVectorAction(EVT VT) const override; |
514 | | |
515 | | // Get the degree of precision we want from 32-bit floating point division |
516 | | // operations. |
517 | | // |
518 | | // 0 - Use ptx div.approx |
519 | | // 1 - Use ptx.div.full (approximate, but less so than div.approx) |
520 | | // 2 - Use IEEE-compliant div instructions, if available. |
521 | | int getDivF32Level() const; |
522 | | |
523 | | // Get whether we should use a precise or approximate 32-bit floating point |
524 | | // sqrt instruction. |
525 | | bool usePrecSqrtF32() const; |
526 | | |
527 | | // Get whether we should use instructions that flush floating-point denormals |
528 | | // to sign-preserving zero. |
529 | | bool useF32FTZ(const MachineFunction &MF) const; |
530 | | |
531 | | SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled, |
532 | | int &ExtraSteps, bool &UseOneConst, |
533 | | bool Reciprocal) const override; |
534 | | |
535 | 16 | unsigned combineRepeatedFPDivisors() const override { return 2; } |
536 | | |
537 | | bool allowFMA(MachineFunction &MF, CodeGenOpt::Level OptLevel) const; |
538 | | bool allowUnsafeFPMath(MachineFunction &MF) const; |
539 | | |
540 | 450 | bool isFMAFasterThanFMulAndFAdd(EVT) const override { return true; } |
541 | | |
542 | 73 | bool enableAggressiveFMAFusion(EVT VT) const override { return true; } |
543 | | |
544 | | // The default is to transform llvm.ctlz(x, false) (where false indicates that |
545 | | // x == 0 is not undefined behavior) into a branch that checks whether x is 0 |
546 | | // and avoids calling ctlz in that case. We have a dedicated ctlz |
547 | | // instruction, so we say that ctlz is cheap to speculate. |
548 | 7 | bool isCheapToSpeculateCtlz() const override { return true; } |
549 | | |
550 | | private: |
551 | | const NVPTXSubtarget &STI; // cache the subtarget here |
552 | | SDValue getParamSymbol(SelectionDAG &DAG, int idx, EVT) const; |
553 | | |
554 | | SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const; |
555 | | SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const; |
556 | | SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const; |
557 | | |
558 | | SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const; |
559 | | SDValue LowerLOADi1(SDValue Op, SelectionDAG &DAG) const; |
560 | | |
561 | | SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const; |
562 | | SDValue LowerSTOREi1(SDValue Op, SelectionDAG &DAG) const; |
563 | | SDValue LowerSTOREVector(SDValue Op, SelectionDAG &DAG) const; |
564 | | |
565 | | SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const; |
566 | | SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const; |
567 | | |
568 | | SDValue LowerSelect(SDValue Op, SelectionDAG &DAG) const; |
569 | | |
570 | | void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results, |
571 | | SelectionDAG &DAG) const override; |
572 | | SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override; |
573 | | |
574 | | unsigned getArgumentAlignment(SDValue Callee, ImmutableCallSite CS, Type *Ty, |
575 | | unsigned Idx, const DataLayout &DL) const; |
576 | | }; |
577 | | } // namespace llvm |
578 | | |
579 | | #endif |