1 | //===-- ARMSubtarget.h - Define Subtarget for the ARM ----------*- C++ -*--===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file declares the ARM specific subclass of TargetSubtargetInfo. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #ifndef LLVM_LIB_TARGET_ARM_ARMSUBTARGET_H |
14 | #define LLVM_LIB_TARGET_ARM_ARMSUBTARGET_H |
15 | |
16 | #include "ARMBaseInstrInfo.h" |
17 | #include "ARMBaseRegisterInfo.h" |
18 | #include "ARMConstantPoolValue.h" |
19 | #include "ARMFrameLowering.h" |
20 | #include "ARMISelLowering.h" |
21 | #include "ARMMachineFunctionInfo.h" |
22 | #include "ARMSelectionDAGInfo.h" |
23 | #include "llvm/Analysis/TargetTransformInfo.h" |
24 | #include "llvm/CodeGen/GlobalISel/CallLowering.h" |
25 | #include "llvm/CodeGen/GlobalISel/InstructionSelector.h" |
26 | #include "llvm/CodeGen/GlobalISel/LegalizerInfo.h" |
27 | #include "llvm/CodeGen/MachineFunction.h" |
28 | #include "llvm/CodeGen/RegisterBankInfo.h" |
29 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
30 | #include "llvm/MC/MCInstrItineraries.h" |
31 | #include "llvm/MC/MCSchedule.h" |
32 | #include "llvm/Target/TargetMachine.h" |
33 | #include "llvm/Target/TargetOptions.h" |
34 | #include "llvm/TargetParser/Triple.h" |
35 | #include <bitset> |
36 | #include <memory> |
37 | #include <string> |
38 | |
39 | #define |
40 | #include "ARMGenSubtargetInfo.inc" |
41 | |
42 | namespace llvm { |
43 | |
44 | class ARMBaseTargetMachine; |
45 | class GlobalValue; |
46 | class StringRef; |
47 | |
48 | class ARMSubtarget : public ARMGenSubtargetInfo { |
49 | protected: |
50 | enum ARMProcFamilyEnum { |
51 | Others, |
52 | #define ARM_PROCESSOR_FAMILY(ENUM) ENUM, |
53 | #include "llvm/TargetParser/ARMTargetParserDef.inc" |
54 | #undef ARM_PROCESSOR_FAMILY |
55 | }; |
56 | enum ARMProcClassEnum { |
57 | None, |
58 | |
59 | AClass, |
60 | MClass, |
61 | RClass |
62 | }; |
63 | enum ARMArchEnum { |
64 | #define ARM_ARCHITECTURE(ENUM) ENUM, |
65 | #include "llvm/TargetParser/ARMTargetParserDef.inc" |
66 | #undef ARM_ARCHITECTURE |
67 | }; |
68 | |
69 | public: |
70 | /// What kind of timing do load multiple/store multiple instructions have. |
71 | enum ARMLdStMultipleTiming { |
72 | /// Can load/store 2 registers/cycle. |
73 | DoubleIssue, |
74 | /// Can load/store 2 registers/cycle, but needs an extra cycle if the access |
75 | /// is not 64-bit aligned. |
76 | DoubleIssueCheckUnalignedAccess, |
77 | /// Can load/store 1 register/cycle. |
78 | SingleIssue, |
79 | /// Can load/store 1 register/cycle, but needs an extra cycle for address |
80 | /// computation and potentially also for register writeback. |
81 | , |
82 | }; |
83 | |
84 | protected: |
85 | // Bool members corresponding to the SubtargetFeatures defined in tablegen |
86 | #define GET_SUBTARGETINFO_MACRO(ATTRIBUTE, DEFAULT, GETTER) \ |
87 | bool ATTRIBUTE = DEFAULT; |
88 | #include "ARMGenSubtargetInfo.inc" |
89 | |
90 | /// ARMProcFamily - ARM processor family: Cortex-A8, Cortex-A9, and others. |
91 | ARMProcFamilyEnum ARMProcFamily = Others; |
92 | |
93 | /// ARMProcClass - ARM processor class: None, AClass, RClass or MClass. |
94 | ARMProcClassEnum ARMProcClass = None; |
95 | |
96 | /// ARMArch - ARM architecture |
97 | ARMArchEnum ARMArch = ARMv4t; |
98 | |
99 | /// UseMulOps - True if non-microcoded fused integer multiply-add and |
100 | /// multiply-subtract instructions should be used. |
101 | bool UseMulOps = false; |
102 | |
103 | /// SupportsTailCall - True if the OS supports tail call. The dynamic linker |
104 | /// must be able to synthesize call stubs for interworking between ARM and |
105 | /// Thumb. |
106 | bool SupportsTailCall = false; |
107 | |
108 | /// RestrictIT - If true, the subtarget disallows generation of complex IT |
109 | /// blocks. |
110 | bool RestrictIT = false; |
111 | |
112 | /// UseSjLjEH - If true, the target uses SjLj exception handling (e.g. iOS). |
113 | bool UseSjLjEH = false; |
114 | |
115 | /// stackAlignment - The minimum alignment known to hold of the stack frame on |
116 | /// entry to the function and which must be maintained by every function. |
117 | Align stackAlignment = Align(4); |
118 | |
119 | /// CPUString - String name of used CPU. |
120 | std::string CPUString; |
121 | |
122 | unsigned MaxInterleaveFactor = 1; |
123 | |
124 | /// Clearance before partial register updates (in number of instructions) |
125 | unsigned PartialUpdateClearance = 0; |
126 | |
127 | /// What kind of timing do load multiple/store multiple have (double issue, |
128 | /// single issue etc). |
129 | ARMLdStMultipleTiming LdStMultipleTiming = SingleIssue; |
130 | |
131 | /// The adjustment that we need to apply to get the operand latency from the |
132 | /// operand cycle returned by the itinerary data for pre-ISel operands. |
133 | int PreISelOperandLatencyAdjustment = 2; |
134 | |
135 | /// What alignment is preferred for loop bodies and functions, in log2(bytes). |
136 | unsigned PrefLoopLogAlignment = 0; |
137 | |
138 | /// The cost factor for MVE instructions, representing the multiple beats an |
139 | // instruction can take. The default is 2, (set in initSubtargetFeatures so |
140 | // that we can use subtarget features less than 2). |
141 | unsigned MVEVectorCostFactor = 0; |
142 | |
143 | /// OptMinSize - True if we're optimising for minimum code size, equal to |
144 | /// the function attribute. |
145 | bool OptMinSize = false; |
146 | |
147 | /// IsLittle - The target is Little Endian |
148 | bool IsLittle; |
149 | |
150 | /// TargetTriple - What processor and OS we're targeting. |
151 | Triple TargetTriple; |
152 | |
153 | /// SchedModel - Processor specific instruction costs. |
154 | MCSchedModel SchedModel; |
155 | |
156 | /// Selected instruction itineraries (one entry per itinerary class.) |
157 | InstrItineraryData InstrItins; |
158 | |
159 | /// Options passed via command line that could influence the target |
160 | const TargetOptions &Options; |
161 | |
162 | const ARMBaseTargetMachine &TM; |
163 | |
164 | public: |
165 | /// This constructor initializes the data members to match that |
166 | /// of the specified triple. |
167 | /// |
168 | ARMSubtarget(const Triple &TT, const std::string &CPU, const std::string &FS, |
169 | const ARMBaseTargetMachine &TM, bool IsLittle, |
170 | bool MinSize = false); |
171 | |
172 | /// getMaxInlineSizeThreshold - Returns the maximum memset / memcpy size |
173 | /// that still makes it profitable to inline the call. |
174 | unsigned getMaxInlineSizeThreshold() const { |
175 | return 64; |
176 | } |
177 | |
178 | /// getMaxMemcpyTPInlineSizeThreshold - Returns the maximum size |
179 | /// that still makes it profitable to inline a llvm.memcpy as a Tail |
180 | /// Predicated loop. |
181 | /// This threshold should only be used for constant size inputs. |
182 | unsigned getMaxMemcpyTPInlineSizeThreshold() const { return 128; } |
183 | |
184 | /// ParseSubtargetFeatures - Parses features string setting specified |
185 | /// subtarget options. Definition of function is auto generated by tblgen. |
186 | void ParseSubtargetFeatures(StringRef CPU, StringRef TuneCPU, StringRef FS); |
187 | |
188 | /// initializeSubtargetDependencies - Initializes using a CPU and feature string |
189 | /// so that we can use initializer lists for subtarget initialization. |
190 | ARMSubtarget &initializeSubtargetDependencies(StringRef CPU, StringRef FS); |
191 | |
192 | const ARMSelectionDAGInfo *getSelectionDAGInfo() const override { |
193 | return &TSInfo; |
194 | } |
195 | |
196 | const ARMBaseInstrInfo *getInstrInfo() const override { |
197 | return InstrInfo.get(); |
198 | } |
199 | |
200 | const ARMTargetLowering *getTargetLowering() const override { |
201 | return &TLInfo; |
202 | } |
203 | |
204 | const ARMFrameLowering *getFrameLowering() const override { |
205 | return FrameLowering.get(); |
206 | } |
207 | |
208 | const ARMBaseRegisterInfo *getRegisterInfo() const override { |
209 | return &InstrInfo->getRegisterInfo(); |
210 | } |
211 | |
212 | /// The correct instructions have been implemented to initialize undef |
213 | /// registers, therefore the ARM Architecture is supported by the Init Undef |
214 | /// Pass. This will return true as the pass needs to be supported for all |
215 | /// types of instructions. The pass will then perform more checks to ensure it |
216 | /// should be applying the Pseudo Instructions. |
217 | bool supportsInitUndef() const override { return true; } |
218 | |
219 | const CallLowering *getCallLowering() const override; |
220 | InstructionSelector *getInstructionSelector() const override; |
221 | const LegalizerInfo *getLegalizerInfo() const override; |
222 | const RegisterBankInfo *getRegBankInfo() const override; |
223 | |
224 | private: |
225 | ARMSelectionDAGInfo TSInfo; |
226 | // Either Thumb1FrameLowering or ARMFrameLowering. |
227 | std::unique_ptr<ARMFrameLowering> FrameLowering; |
228 | // Either Thumb1InstrInfo or Thumb2InstrInfo. |
229 | std::unique_ptr<ARMBaseInstrInfo> InstrInfo; |
230 | ARMTargetLowering TLInfo; |
231 | |
232 | /// GlobalISel related APIs. |
233 | std::unique_ptr<CallLowering> CallLoweringInfo; |
234 | std::unique_ptr<InstructionSelector> InstSelector; |
235 | std::unique_ptr<LegalizerInfo> Legalizer; |
236 | std::unique_ptr<RegisterBankInfo> RegBankInfo; |
237 | |
238 | void initializeEnvironment(); |
239 | void initSubtargetFeatures(StringRef CPU, StringRef FS); |
240 | ARMFrameLowering *initializeFrameLowering(StringRef CPU, StringRef FS); |
241 | |
242 | std::bitset<8> CoprocCDE = {}; |
243 | public: |
244 | // Getters for SubtargetFeatures defined in tablegen |
245 | #define GET_SUBTARGETINFO_MACRO(ATTRIBUTE, DEFAULT, GETTER) \ |
246 | bool GETTER() const { return ATTRIBUTE; } |
247 | #include "ARMGenSubtargetInfo.inc" |
248 | |
249 | /// @{ |
250 | /// These functions are obsolete, please consider adding subtarget features |
251 | /// or properties instead of calling them. |
252 | bool isCortexA5() const { return ARMProcFamily == CortexA5; } |
253 | bool isCortexA7() const { return ARMProcFamily == CortexA7; } |
254 | bool isCortexA8() const { return ARMProcFamily == CortexA8; } |
255 | bool isCortexA9() const { return ARMProcFamily == CortexA9; } |
256 | bool isCortexA15() const { return ARMProcFamily == CortexA15; } |
257 | bool isSwift() const { return ARMProcFamily == Swift; } |
258 | bool isCortexM3() const { return ARMProcFamily == CortexM3; } |
259 | bool isCortexM7() const { return ARMProcFamily == CortexM7; } |
260 | bool isLikeA9() const { return isCortexA9() || isCortexA15() || isKrait(); } |
261 | bool isCortexR5() const { return ARMProcFamily == CortexR5; } |
262 | bool isKrait() const { return ARMProcFamily == Krait; } |
263 | /// @} |
264 | |
265 | bool hasARMOps() const { return !NoARM; } |
266 | |
267 | bool useNEONForSinglePrecisionFP() const { |
268 | return hasNEON() && hasNEONForFP(); |
269 | } |
270 | |
271 | bool hasVFP2Base() const { return hasVFPv2SP(); } |
272 | bool hasVFP3Base() const { return hasVFPv3D16SP(); } |
273 | bool hasVFP4Base() const { return hasVFPv4D16SP(); } |
274 | bool hasFPARMv8Base() const { return hasFPARMv8D16SP(); } |
275 | |
276 | bool hasAnyDataBarrier() const { |
277 | return HasDataBarrier || (hasV6Ops() && !isThumb()); |
278 | } |
279 | |
280 | bool useMulOps() const { return UseMulOps; } |
281 | bool useFPVMLx() const { return !SlowFPVMLx; } |
282 | bool useFPVFMx() const { |
283 | return !isTargetDarwin() && hasVFP4Base() && !SlowFPVFMx; |
284 | } |
285 | bool useFPVFMx16() const { return useFPVFMx() && hasFullFP16(); } |
286 | bool useFPVFMx64() const { return useFPVFMx() && hasFP64(); } |
287 | bool useSjLjEH() const { return UseSjLjEH; } |
288 | bool hasBaseDSP() const { |
289 | if (isThumb()) |
290 | return hasThumb2() && hasDSP(); |
291 | else |
292 | return hasV5TEOps(); |
293 | } |
294 | |
295 | /// Return true if the CPU supports any kind of instruction fusion. |
296 | bool hasFusion() const { return hasFuseAES() || hasFuseLiterals(); } |
297 | |
298 | const Triple &getTargetTriple() const { return TargetTriple; } |
299 | |
300 | bool isTargetDarwin() const { return TargetTriple.isOSDarwin(); } |
301 | bool isTargetIOS() const { return TargetTriple.isiOS(); } |
302 | bool isTargetWatchOS() const { return TargetTriple.isWatchOS(); } |
303 | bool isTargetWatchABI() const { return TargetTriple.isWatchABI(); } |
304 | bool isTargetDriverKit() const { return TargetTriple.isDriverKit(); } |
305 | bool isTargetLinux() const { return TargetTriple.isOSLinux(); } |
306 | bool isTargetNaCl() const { return TargetTriple.isOSNaCl(); } |
307 | bool isTargetNetBSD() const { return TargetTriple.isOSNetBSD(); } |
308 | bool isTargetWindows() const { return TargetTriple.isOSWindows(); } |
309 | |
310 | bool isTargetCOFF() const { return TargetTriple.isOSBinFormatCOFF(); } |
311 | bool isTargetELF() const { return TargetTriple.isOSBinFormatELF(); } |
312 | bool isTargetMachO() const { return TargetTriple.isOSBinFormatMachO(); } |
313 | |
314 | // ARM EABI is the bare-metal EABI described in ARM ABI documents and |
315 | // can be accessed via -target arm-none-eabi. This is NOT GNUEABI. |
316 | // FIXME: Add a flag for bare-metal for that target and set Triple::EABI |
317 | // even for GNUEABI, so we can make a distinction here and still conform to |
318 | // the EABI on GNU (and Android) mode. This requires change in Clang, too. |
319 | // FIXME: The Darwin exception is temporary, while we move users to |
320 | // "*-*-*-macho" triples as quickly as possible. |
321 | bool isTargetAEABI() const { |
322 | return (TargetTriple.getEnvironment() == Triple::EABI || |
323 | TargetTriple.getEnvironment() == Triple::EABIHF) && |
324 | !isTargetDarwin() && !isTargetWindows(); |
325 | } |
326 | bool isTargetGNUAEABI() const { |
327 | return (TargetTriple.getEnvironment() == Triple::GNUEABI || |
328 | TargetTriple.getEnvironment() == Triple::GNUEABIHF) && |
329 | !isTargetDarwin() && !isTargetWindows(); |
330 | } |
331 | bool isTargetMuslAEABI() const { |
332 | return (TargetTriple.getEnvironment() == Triple::MuslEABI || |
333 | TargetTriple.getEnvironment() == Triple::MuslEABIHF || |
334 | TargetTriple.getEnvironment() == Triple::OpenHOS) && |
335 | !isTargetDarwin() && !isTargetWindows(); |
336 | } |
337 | |
338 | // ARM Targets that support EHABI exception handling standard |
339 | // Darwin uses SjLj. Other targets might need more checks. |
340 | bool isTargetEHABICompatible() const { |
341 | return TargetTriple.isTargetEHABICompatible(); |
342 | } |
343 | |
344 | bool isTargetHardFloat() const; |
345 | |
346 | bool isReadTPSoft() const { |
347 | return !(isReadTPTPIDRURW() || isReadTPTPIDRURO() || isReadTPTPIDRPRW()); |
348 | } |
349 | |
350 | bool isTargetAndroid() const { return TargetTriple.isAndroid(); } |
351 | |
352 | bool isXRaySupported() const override; |
353 | |
354 | bool isAPCS_ABI() const; |
355 | bool isAAPCS_ABI() const; |
356 | bool isAAPCS16_ABI() const; |
357 | |
358 | bool isROPI() const; |
359 | bool isRWPI() const; |
360 | |
361 | bool useMachineScheduler() const { return UseMISched; } |
362 | bool useMachinePipeliner() const { return UseMIPipeliner; } |
363 | bool hasMinSize() const { return OptMinSize; } |
364 | bool isThumb1Only() const { return isThumb() && !hasThumb2(); } |
365 | bool isThumb2() const { return isThumb() && hasThumb2(); } |
366 | bool isMClass() const { return ARMProcClass == MClass; } |
367 | bool isRClass() const { return ARMProcClass == RClass; } |
368 | bool isAClass() const { return ARMProcClass == AClass; } |
369 | |
370 | bool isR9Reserved() const { |
371 | return isTargetMachO() ? (ReserveR9 || !HasV6Ops) : ReserveR9; |
372 | } |
373 | |
374 | MCPhysReg getFramePointerReg() const { |
375 | if (isTargetDarwin() || |
376 | (!isTargetWindows() && isThumb() && !createAAPCSFrameChain())) |
377 | return ARM::R7; |
378 | return ARM::R11; |
379 | } |
380 | |
381 | /// Returns true if the frame setup is split into two separate pushes (first |
382 | /// r0-r7,lr then r8-r11), principally so that the frame pointer is adjacent |
383 | /// to lr. This is always required on Thumb1-only targets, as the push and |
384 | /// pop instructions can't access the high registers. |
385 | bool splitFramePushPop(const MachineFunction &MF) const { |
386 | if (MF.getInfo<ARMFunctionInfo>()->shouldSignReturnAddress()) |
387 | return true; |
388 | return (getFramePointerReg() == ARM::R7 && |
389 | MF.getTarget().Options.DisableFramePointerElim(MF)) || |
390 | isThumb1Only(); |
391 | } |
392 | |
393 | bool splitFramePointerPush(const MachineFunction &MF) const; |
394 | |
395 | bool useStride4VFPs() const; |
396 | |
397 | bool useMovt() const; |
398 | |
399 | bool supportsTailCall() const { return SupportsTailCall; } |
400 | |
401 | bool allowsUnalignedMem() const { return !StrictAlign; } |
402 | |
403 | bool restrictIT() const { return RestrictIT; } |
404 | |
405 | const std::string & getCPUString() const { return CPUString; } |
406 | |
407 | bool isLittle() const { return IsLittle; } |
408 | |
409 | unsigned getMispredictionPenalty() const; |
410 | |
411 | /// Returns true if machine scheduler should be enabled. |
412 | bool enableMachineScheduler() const override; |
413 | |
414 | /// Returns true if machine pipeliner should be enabled. |
415 | bool enableMachinePipeliner() const override; |
416 | bool useDFAforSMS() const override; |
417 | |
418 | /// True for some subtargets at > -O0. |
419 | bool enablePostRAScheduler() const override; |
420 | |
421 | /// True for some subtargets at > -O0. |
422 | bool enablePostRAMachineScheduler() const override; |
423 | |
424 | /// Check whether this subtarget wants to use subregister liveness. |
425 | bool enableSubRegLiveness() const override; |
426 | |
427 | /// Enable use of alias analysis during code generation (during MI |
428 | /// scheduling, DAGCombine, etc.). |
429 | bool useAA() const override { return true; } |
430 | |
431 | /// getInstrItins - Return the instruction itineraries based on subtarget |
432 | /// selection. |
433 | const InstrItineraryData *getInstrItineraryData() const override { |
434 | return &InstrItins; |
435 | } |
436 | |
437 | /// getStackAlignment - Returns the minimum alignment known to hold of the |
438 | /// stack frame on entry to the function and which must be maintained by every |
439 | /// function for this subtarget. |
440 | Align getStackAlignment() const { return stackAlignment; } |
441 | |
442 | // Returns the required alignment for LDRD/STRD instructions |
443 | Align getDualLoadStoreAlignment() const { |
444 | return Align(hasV7Ops() || allowsUnalignedMem() ? 4 : 8); |
445 | } |
446 | |
447 | unsigned getMaxInterleaveFactor() const { return MaxInterleaveFactor; } |
448 | |
449 | unsigned getPartialUpdateClearance() const { return PartialUpdateClearance; } |
450 | |
451 | ARMLdStMultipleTiming getLdStMultipleTiming() const { |
452 | return LdStMultipleTiming; |
453 | } |
454 | |
455 | int getPreISelOperandLatencyAdjustment() const { |
456 | return PreISelOperandLatencyAdjustment; |
457 | } |
458 | |
459 | /// True if the GV will be accessed via an indirect symbol. |
460 | bool isGVIndirectSymbol(const GlobalValue *GV) const; |
461 | |
462 | /// Returns the constant pool modifier needed to access the GV. |
463 | bool isGVInGOT(const GlobalValue *GV) const; |
464 | |
465 | /// True if fast-isel is used. |
466 | bool useFastISel() const; |
467 | |
468 | /// Returns the correct return opcode for the current feature set. |
469 | /// Use BX if available to allow mixing thumb/arm code, but fall back |
470 | /// to plain mov pc,lr on ARMv4. |
471 | unsigned getReturnOpcode() const { |
472 | if (isThumb()) |
473 | return ARM::tBX_RET; |
474 | if (hasV4TOps()) |
475 | return ARM::BX_RET; |
476 | return ARM::MOVPCLR; |
477 | } |
478 | |
479 | /// Allow movt+movw for PIC global address calculation. |
480 | /// ELF does not have GOT relocations for movt+movw. |
481 | /// ROPI does not use GOT. |
482 | bool allowPositionIndependentMovt() const { |
483 | return isROPI() || !isTargetELF(); |
484 | } |
485 | |
486 | unsigned getPrefLoopLogAlignment() const { return PrefLoopLogAlignment; } |
487 | |
488 | unsigned |
489 | getMVEVectorCostFactor(TargetTransformInfo::TargetCostKind CostKind) const { |
490 | if (CostKind == TargetTransformInfo::TCK_CodeSize) |
491 | return 1; |
492 | return MVEVectorCostFactor; |
493 | } |
494 | |
495 | bool ignoreCSRForAllocationOrder(const MachineFunction &MF, |
496 | unsigned PhysReg) const override; |
497 | unsigned getGPRAllocationOrder(const MachineFunction &MF) const; |
498 | }; |
499 | |
500 | } // end namespace llvm |
501 | |
502 | #endif // LLVM_LIB_TARGET_ARM_ARMSUBTARGET_H |
503 | |