AMDGPU: Mark test as XFAIL in expensive_checks builds
[llvm-project.git] / llvm / lib / Target / ARM / ARMSubtarget.h
blob611eeac9ef712874a049be60d3d3b7d9672b6324
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.
11 //===----------------------------------------------------------------------===//
13 #ifndef LLVM_LIB_TARGET_ARM_ARMSUBTARGET_H
14 #define LLVM_LIB_TARGET_ARM_ARMSUBTARGET_H
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>
39 #define GET_SUBTARGETINFO_HEADER
40 #include "ARMGenSubtargetInfo.inc"
42 namespace llvm {
44 class ARMBaseTargetMachine;
45 class GlobalValue;
46 class StringRef;
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
56 enum ARMProcClassEnum {
57 None,
59 AClass,
60 MClass,
61 RClass
63 enum ARMArchEnum {
64 #define ARM_ARCHITECTURE(ENUM) ENUM,
65 #include "llvm/TargetParser/ARMTargetParserDef.inc"
66 #undef ARM_ARCHITECTURE
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 SingleIssuePlusExtras,
84 /// How the push and pop instructions of callee saved general-purpose
85 /// registers should be split.
86 enum PushPopSplitVariation {
87 /// All GPRs can be pushed in a single instruction.
88 /// push {r0-r12, lr}
89 /// vpush {d8-d15}
90 NoSplit,
92 /// R7 and LR must be adjacent, because R7 is the frame pointer, and must
93 /// point to a frame record consisting of the previous frame pointer and the
94 /// return address.
95 /// push {r0-r7, lr}
96 /// push {r8-r12}
97 /// vpush {d8-d15}
98 /// Note that Thumb1 changes this layout when the frame pointer is R11,
99 /// using a longer sequence of instructions because R11 can't be used by a
100 /// Thumb1 push instruction. This doesn't currently have a separate enum
101 /// value, and is handled entriely within Thumb1FrameLowering::emitPrologue.
102 SplitR7,
104 /// When the stack frame size is not known (because of variable-sized
105 /// objects or realignment), Windows SEH requires the callee-saved registers
106 /// to be stored in three regions, with R11 and LR below the floating-point
107 /// registers.
108 /// push {r0-r10, r12}
109 /// vpush {d8-d15}
110 /// push {r11, lr}
111 SplitR11WindowsSEH,
113 /// When generating AAPCS-compilant frame chains, R11 is the frame pointer,
114 /// and must be pushed adjacent to the return address (LR). Normally this
115 /// isn't a problem, because the only register between them is r12, which is
116 /// the intra-procedure-call scratch register, so doesn't need to be saved.
117 /// However, when PACBTI is in use, r12 contains the authentication code, so
118 /// does need to be saved. This means that we need a separate push for R11
119 /// and LR.
120 /// push {r0-r10, r12}
121 /// push {r11, lr}
122 /// vpush {d8-d15}
123 SplitR11AAPCSSignRA,
126 protected:
127 // Bool members corresponding to the SubtargetFeatures defined in tablegen
128 #define GET_SUBTARGETINFO_MACRO(ATTRIBUTE, DEFAULT, GETTER) \
129 bool ATTRIBUTE = DEFAULT;
130 #include "ARMGenSubtargetInfo.inc"
132 /// ARMProcFamily - ARM processor family: Cortex-A8, Cortex-A9, and others.
133 ARMProcFamilyEnum ARMProcFamily = Others;
135 /// ARMProcClass - ARM processor class: None, AClass, RClass or MClass.
136 ARMProcClassEnum ARMProcClass = None;
138 /// ARMArch - ARM architecture
139 ARMArchEnum ARMArch = ARMv4t;
141 /// UseMulOps - True if non-microcoded fused integer multiply-add and
142 /// multiply-subtract instructions should be used.
143 bool UseMulOps = false;
145 /// SupportsTailCall - True if the OS supports tail call. The dynamic linker
146 /// must be able to synthesize call stubs for interworking between ARM and
147 /// Thumb.
148 bool SupportsTailCall = false;
150 /// RestrictIT - If true, the subtarget disallows generation of complex IT
151 /// blocks.
152 bool RestrictIT = false;
154 /// UseSjLjEH - If true, the target uses SjLj exception handling (e.g. iOS).
155 bool UseSjLjEH = false;
157 /// stackAlignment - The minimum alignment known to hold of the stack frame on
158 /// entry to the function and which must be maintained by every function.
159 Align stackAlignment = Align(4);
161 /// CPUString - String name of used CPU.
162 std::string CPUString;
164 unsigned MaxInterleaveFactor = 1;
166 /// Clearance before partial register updates (in number of instructions)
167 unsigned PartialUpdateClearance = 0;
169 /// What kind of timing do load multiple/store multiple have (double issue,
170 /// single issue etc).
171 ARMLdStMultipleTiming LdStMultipleTiming = SingleIssue;
173 /// The adjustment that we need to apply to get the operand latency from the
174 /// operand cycle returned by the itinerary data for pre-ISel operands.
175 int PreISelOperandLatencyAdjustment = 2;
177 /// What alignment is preferred for loop bodies and functions, in log2(bytes).
178 unsigned PreferBranchLogAlignment = 0;
180 /// The cost factor for MVE instructions, representing the multiple beats an
181 // instruction can take. The default is 2, (set in initSubtargetFeatures so
182 // that we can use subtarget features less than 2).
183 unsigned MVEVectorCostFactor = 0;
185 /// OptMinSize - True if we're optimising for minimum code size, equal to
186 /// the function attribute.
187 bool OptMinSize = false;
189 /// IsLittle - The target is Little Endian
190 bool IsLittle;
192 /// TargetTriple - What processor and OS we're targeting.
193 Triple TargetTriple;
195 /// SchedModel - Processor specific instruction costs.
196 MCSchedModel SchedModel;
198 /// Selected instruction itineraries (one entry per itinerary class.)
199 InstrItineraryData InstrItins;
201 /// Options passed via command line that could influence the target
202 const TargetOptions &Options;
204 const ARMBaseTargetMachine &TM;
206 public:
207 /// This constructor initializes the data members to match that
208 /// of the specified triple.
210 ARMSubtarget(const Triple &TT, const std::string &CPU, const std::string &FS,
211 const ARMBaseTargetMachine &TM, bool IsLittle,
212 bool MinSize = false);
214 /// getMaxInlineSizeThreshold - Returns the maximum memset / memcpy size
215 /// that still makes it profitable to inline the call.
216 unsigned getMaxInlineSizeThreshold() const {
217 return 64;
220 /// getMaxMemcpyTPInlineSizeThreshold - Returns the maximum size
221 /// that still makes it profitable to inline a llvm.memcpy as a Tail
222 /// Predicated loop.
223 /// This threshold should only be used for constant size inputs.
224 unsigned getMaxMemcpyTPInlineSizeThreshold() const { return 128; }
226 /// ParseSubtargetFeatures - Parses features string setting specified
227 /// subtarget options. Definition of function is auto generated by tblgen.
228 void ParseSubtargetFeatures(StringRef CPU, StringRef TuneCPU, StringRef FS);
230 /// initializeSubtargetDependencies - Initializes using a CPU and feature string
231 /// so that we can use initializer lists for subtarget initialization.
232 ARMSubtarget &initializeSubtargetDependencies(StringRef CPU, StringRef FS);
234 const ARMSelectionDAGInfo *getSelectionDAGInfo() const override {
235 return &TSInfo;
238 const ARMBaseInstrInfo *getInstrInfo() const override {
239 return InstrInfo.get();
242 const ARMTargetLowering *getTargetLowering() const override {
243 return &TLInfo;
246 const ARMFrameLowering *getFrameLowering() const override {
247 return FrameLowering.get();
250 const ARMBaseRegisterInfo *getRegisterInfo() const override {
251 return &InstrInfo->getRegisterInfo();
254 const CallLowering *getCallLowering() const override;
255 InstructionSelector *getInstructionSelector() const override;
256 const LegalizerInfo *getLegalizerInfo() const override;
257 const RegisterBankInfo *getRegBankInfo() const override;
259 private:
260 ARMSelectionDAGInfo TSInfo;
261 // Either Thumb1FrameLowering or ARMFrameLowering.
262 std::unique_ptr<ARMFrameLowering> FrameLowering;
263 // Either Thumb1InstrInfo or Thumb2InstrInfo.
264 std::unique_ptr<ARMBaseInstrInfo> InstrInfo;
265 ARMTargetLowering TLInfo;
267 /// GlobalISel related APIs.
268 std::unique_ptr<CallLowering> CallLoweringInfo;
269 std::unique_ptr<InstructionSelector> InstSelector;
270 std::unique_ptr<LegalizerInfo> Legalizer;
271 std::unique_ptr<RegisterBankInfo> RegBankInfo;
273 void initializeEnvironment();
274 void initSubtargetFeatures(StringRef CPU, StringRef FS);
275 ARMFrameLowering *initializeFrameLowering(StringRef CPU, StringRef FS);
277 std::bitset<8> CoprocCDE = {};
278 public:
279 // Getters for SubtargetFeatures defined in tablegen
280 #define GET_SUBTARGETINFO_MACRO(ATTRIBUTE, DEFAULT, GETTER) \
281 bool GETTER() const { return ATTRIBUTE; }
282 #include "ARMGenSubtargetInfo.inc"
284 /// @{
285 /// These functions are obsolete, please consider adding subtarget features
286 /// or properties instead of calling them.
287 bool isCortexA5() const { return ARMProcFamily == CortexA5; }
288 bool isCortexA7() const { return ARMProcFamily == CortexA7; }
289 bool isCortexA8() const { return ARMProcFamily == CortexA8; }
290 bool isCortexA9() const { return ARMProcFamily == CortexA9; }
291 bool isCortexA15() const { return ARMProcFamily == CortexA15; }
292 bool isSwift() const { return ARMProcFamily == Swift; }
293 bool isCortexM3() const { return ARMProcFamily == CortexM3; }
294 bool isCortexM55() const { return ARMProcFamily == CortexM55; }
295 bool isCortexM7() const { return ARMProcFamily == CortexM7; }
296 bool isCortexM85() const { return ARMProcFamily == CortexM85; }
297 bool isLikeA9() const { return isCortexA9() || isCortexA15() || isKrait(); }
298 bool isCortexR5() const { return ARMProcFamily == CortexR5; }
299 bool isKrait() const { return ARMProcFamily == Krait; }
300 /// @}
302 bool hasARMOps() const { return !NoARM; }
304 bool useNEONForSinglePrecisionFP() const {
305 return hasNEON() && hasNEONForFP();
308 bool hasVFP2Base() const { return hasVFPv2SP(); }
309 bool hasVFP3Base() const { return hasVFPv3D16SP(); }
310 bool hasVFP4Base() const { return hasVFPv4D16SP(); }
311 bool hasFPARMv8Base() const { return hasFPARMv8D16SP(); }
313 bool hasAnyDataBarrier() const {
314 return HasDataBarrier || (hasV6Ops() && !isThumb());
317 bool useMulOps() const { return UseMulOps; }
318 bool useFPVMLx() const { return !SlowFPVMLx; }
319 bool useFPVFMx() const {
320 return !isTargetDarwin() && hasVFP4Base() && !SlowFPVFMx;
322 bool useFPVFMx16() const { return useFPVFMx() && hasFullFP16(); }
323 bool useFPVFMx64() const { return useFPVFMx() && hasFP64(); }
324 bool useSjLjEH() const { return UseSjLjEH; }
325 bool hasBaseDSP() const {
326 if (isThumb())
327 return hasThumb2() && hasDSP();
328 else
329 return hasV5TEOps();
332 /// Return true if the CPU supports any kind of instruction fusion.
333 bool hasFusion() const { return hasFuseAES() || hasFuseLiterals(); }
335 const Triple &getTargetTriple() const { return TargetTriple; }
337 bool isTargetDarwin() const { return TargetTriple.isOSDarwin(); }
338 bool isTargetIOS() const { return TargetTriple.isiOS(); }
339 bool isTargetWatchOS() const { return TargetTriple.isWatchOS(); }
340 bool isTargetWatchABI() const { return TargetTriple.isWatchABI(); }
341 bool isTargetDriverKit() const { return TargetTriple.isDriverKit(); }
342 bool isTargetLinux() const { return TargetTriple.isOSLinux(); }
343 bool isTargetNaCl() const { return TargetTriple.isOSNaCl(); }
344 bool isTargetNetBSD() const { return TargetTriple.isOSNetBSD(); }
345 bool isTargetWindows() const { return TargetTriple.isOSWindows(); }
347 bool isTargetCOFF() const { return TargetTriple.isOSBinFormatCOFF(); }
348 bool isTargetELF() const { return TargetTriple.isOSBinFormatELF(); }
349 bool isTargetMachO() const { return TargetTriple.isOSBinFormatMachO(); }
351 // ARM EABI is the bare-metal EABI described in ARM ABI documents and
352 // can be accessed via -target arm-none-eabi. This is NOT GNUEABI.
353 // FIXME: Add a flag for bare-metal for that target and set Triple::EABI
354 // even for GNUEABI, so we can make a distinction here and still conform to
355 // the EABI on GNU (and Android) mode. This requires change in Clang, too.
356 // FIXME: The Darwin exception is temporary, while we move users to
357 // "*-*-*-macho" triples as quickly as possible.
358 bool isTargetAEABI() const {
359 return (TargetTriple.getEnvironment() == Triple::EABI ||
360 TargetTriple.getEnvironment() == Triple::EABIHF) &&
361 !isTargetDarwin() && !isTargetWindows();
363 bool isTargetGNUAEABI() const {
364 return (TargetTriple.getEnvironment() == Triple::GNUEABI ||
365 TargetTriple.getEnvironment() == Triple::GNUEABIT64 ||
366 TargetTriple.getEnvironment() == Triple::GNUEABIHF ||
367 TargetTriple.getEnvironment() == Triple::GNUEABIHFT64) &&
368 !isTargetDarwin() && !isTargetWindows();
370 bool isTargetMuslAEABI() const {
371 return (TargetTriple.getEnvironment() == Triple::MuslEABI ||
372 TargetTriple.getEnvironment() == Triple::MuslEABIHF ||
373 TargetTriple.getEnvironment() == Triple::OpenHOS) &&
374 !isTargetDarwin() && !isTargetWindows();
377 // ARM Targets that support EHABI exception handling standard
378 // Darwin uses SjLj. Other targets might need more checks.
379 bool isTargetEHABICompatible() const {
380 return TargetTriple.isTargetEHABICompatible();
383 bool isTargetHardFloat() const;
385 bool isReadTPSoft() const {
386 return !(isReadTPTPIDRURW() || isReadTPTPIDRURO() || isReadTPTPIDRPRW());
389 bool isTargetAndroid() const { return TargetTriple.isAndroid(); }
391 bool isXRaySupported() const override;
393 bool isAPCS_ABI() const;
394 bool isAAPCS_ABI() const;
395 bool isAAPCS16_ABI() const;
397 bool isROPI() const;
398 bool isRWPI() const;
400 bool useMachineScheduler() const { return UseMISched; }
401 bool useMachinePipeliner() const { return UseMIPipeliner; }
402 bool hasMinSize() const { return OptMinSize; }
403 bool isThumb1Only() const { return isThumb() && !hasThumb2(); }
404 bool isThumb2() const { return isThumb() && hasThumb2(); }
405 bool isMClass() const { return ARMProcClass == MClass; }
406 bool isRClass() const { return ARMProcClass == RClass; }
407 bool isAClass() const { return ARMProcClass == AClass; }
409 bool isR9Reserved() const {
410 return isTargetMachO() ? (ReserveR9 || !HasV6Ops) : ReserveR9;
413 MCPhysReg getFramePointerReg() const {
414 if (isTargetDarwin() ||
415 (!isTargetWindows() && isThumb() && !createAAPCSFrameChain()))
416 return ARM::R7;
417 return ARM::R11;
420 enum PushPopSplitVariation
421 getPushPopSplitVariation(const MachineFunction &MF) const;
423 bool useStride4VFPs() const;
425 bool useMovt() const;
427 bool supportsTailCall() const { return SupportsTailCall; }
429 bool allowsUnalignedMem() const { return !StrictAlign; }
431 bool restrictIT() const { return RestrictIT; }
433 const std::string & getCPUString() const { return CPUString; }
435 bool isLittle() const { return IsLittle; }
437 unsigned getMispredictionPenalty() const;
439 /// Returns true if machine scheduler should be enabled.
440 bool enableMachineScheduler() const override;
442 /// Returns true if machine pipeliner should be enabled.
443 bool enableMachinePipeliner() const override;
444 bool useDFAforSMS() const override;
446 /// True for some subtargets at > -O0.
447 bool enablePostRAScheduler() const override;
449 /// True for some subtargets at > -O0.
450 bool enablePostRAMachineScheduler() const override;
452 /// Check whether this subtarget wants to use subregister liveness.
453 bool enableSubRegLiveness() const override;
455 /// Enable use of alias analysis during code generation (during MI
456 /// scheduling, DAGCombine, etc.).
457 bool useAA() const override { return true; }
459 /// getInstrItins - Return the instruction itineraries based on subtarget
460 /// selection.
461 const InstrItineraryData *getInstrItineraryData() const override {
462 return &InstrItins;
465 /// getStackAlignment - Returns the minimum alignment known to hold of the
466 /// stack frame on entry to the function and which must be maintained by every
467 /// function for this subtarget.
468 Align getStackAlignment() const { return stackAlignment; }
470 // Returns the required alignment for LDRD/STRD instructions
471 Align getDualLoadStoreAlignment() const {
472 return Align(hasV7Ops() || allowsUnalignedMem() ? 4 : 8);
475 unsigned getMaxInterleaveFactor() const { return MaxInterleaveFactor; }
477 unsigned getPartialUpdateClearance() const { return PartialUpdateClearance; }
479 ARMLdStMultipleTiming getLdStMultipleTiming() const {
480 return LdStMultipleTiming;
483 int getPreISelOperandLatencyAdjustment() const {
484 return PreISelOperandLatencyAdjustment;
487 /// True if the GV will be accessed via an indirect symbol.
488 bool isGVIndirectSymbol(const GlobalValue *GV) const;
490 /// Returns the constant pool modifier needed to access the GV.
491 bool isGVInGOT(const GlobalValue *GV) const;
493 /// True if fast-isel is used.
494 bool useFastISel() const;
496 /// Returns the correct return opcode for the current feature set.
497 /// Use BX if available to allow mixing thumb/arm code, but fall back
498 /// to plain mov pc,lr on ARMv4.
499 unsigned getReturnOpcode() const {
500 if (isThumb())
501 return ARM::tBX_RET;
502 if (hasV4TOps())
503 return ARM::BX_RET;
504 return ARM::MOVPCLR;
507 /// Allow movt+movw for PIC global address calculation.
508 /// ELF does not have GOT relocations for movt+movw.
509 /// ROPI does not use GOT.
510 bool allowPositionIndependentMovt() const {
511 return isROPI() || !isTargetELF();
514 unsigned getPreferBranchLogAlignment() const {
515 return PreferBranchLogAlignment;
518 unsigned
519 getMVEVectorCostFactor(TargetTransformInfo::TargetCostKind CostKind) const {
520 if (CostKind == TargetTransformInfo::TCK_CodeSize)
521 return 1;
522 return MVEVectorCostFactor;
525 bool ignoreCSRForAllocationOrder(const MachineFunction &MF,
526 unsigned PhysReg) const override;
527 unsigned getGPRAllocationOrder(const MachineFunction &MF) const;
530 } // end namespace llvm
532 #endif // LLVM_LIB_TARGET_ARM_ARMSUBTARGET_H