1 | //===-- ARMSubtarget.cpp - ARM Subtarget Information ----------------------===// |
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 implements the ARM specific subclass of TargetSubtargetInfo. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "ARM.h" |
14 | |
15 | #include "ARMCallLowering.h" |
16 | #include "ARMFrameLowering.h" |
17 | #include "ARMInstrInfo.h" |
18 | #include "ARMLegalizerInfo.h" |
19 | #include "ARMRegisterBankInfo.h" |
20 | #include "ARMSubtarget.h" |
21 | #include "ARMTargetMachine.h" |
22 | #include "MCTargetDesc/ARMMCTargetDesc.h" |
23 | #include "Thumb1FrameLowering.h" |
24 | #include "Thumb1InstrInfo.h" |
25 | #include "Thumb2InstrInfo.h" |
26 | #include "llvm/ADT/StringRef.h" |
27 | #include "llvm/ADT/Twine.h" |
28 | #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" |
29 | #include "llvm/CodeGen/MachineFrameInfo.h" |
30 | #include "llvm/CodeGen/MachineFunction.h" |
31 | #include "llvm/IR/Function.h" |
32 | #include "llvm/IR/GlobalValue.h" |
33 | #include "llvm/MC/MCAsmInfo.h" |
34 | #include "llvm/MC/MCTargetOptions.h" |
35 | #include "llvm/Support/CodeGen.h" |
36 | #include "llvm/Support/CommandLine.h" |
37 | #include "llvm/Target/TargetOptions.h" |
38 | #include "llvm/TargetParser/ARMTargetParser.h" |
39 | #include "llvm/TargetParser/Triple.h" |
40 | |
41 | using namespace llvm; |
42 | |
43 | #define DEBUG_TYPE "arm-subtarget" |
44 | |
45 | #define GET_SUBTARGETINFO_TARGET_DESC |
46 | #define GET_SUBTARGETINFO_CTOR |
47 | #include "ARMGenSubtargetInfo.inc" |
48 | |
49 | static cl::opt<bool> |
50 | UseFusedMulOps("arm-use-mulops" , |
51 | cl::init(Val: true), cl::Hidden); |
52 | |
53 | enum ITMode { |
54 | DefaultIT, |
55 | RestrictedIT |
56 | }; |
57 | |
58 | static cl::opt<ITMode> |
59 | IT(cl::desc("IT block support" ), cl::Hidden, cl::init(Val: DefaultIT), |
60 | cl::values(clEnumValN(DefaultIT, "arm-default-it" , |
61 | "Generate any type of IT block" ), |
62 | clEnumValN(RestrictedIT, "arm-restrict-it" , |
63 | "Disallow complex IT blocks" ))); |
64 | |
65 | /// ForceFastISel - Use the fast-isel, even for subtargets where it is not |
66 | /// currently supported (for testing only). |
67 | static cl::opt<bool> |
68 | ForceFastISel("arm-force-fast-isel" , |
69 | cl::init(Val: false), cl::Hidden); |
70 | |
71 | /// initializeSubtargetDependencies - Initializes using a CPU and feature string |
72 | /// so that we can use initializer lists for subtarget initialization. |
73 | ARMSubtarget &ARMSubtarget::initializeSubtargetDependencies(StringRef CPU, |
74 | StringRef FS) { |
75 | initializeEnvironment(); |
76 | initSubtargetFeatures(CPU, FS); |
77 | return *this; |
78 | } |
79 | |
80 | ARMFrameLowering *ARMSubtarget::initializeFrameLowering(StringRef CPU, |
81 | StringRef FS) { |
82 | ARMSubtarget &STI = initializeSubtargetDependencies(CPU, FS); |
83 | if (STI.isThumb1Only()) |
84 | return (ARMFrameLowering *)new Thumb1FrameLowering(STI); |
85 | |
86 | return new ARMFrameLowering(STI); |
87 | } |
88 | |
89 | ARMSubtarget::ARMSubtarget(const Triple &TT, const std::string &CPU, |
90 | const std::string &FS, |
91 | const ARMBaseTargetMachine &TM, bool IsLittle, |
92 | bool MinSize) |
93 | : ARMGenSubtargetInfo(TT, CPU, /*TuneCPU*/ CPU, FS), |
94 | UseMulOps(UseFusedMulOps), CPUString(CPU), OptMinSize(MinSize), |
95 | IsLittle(IsLittle), TargetTriple(TT), Options(TM.Options), TM(TM), |
96 | FrameLowering(initializeFrameLowering(CPU, FS)), |
97 | // At this point initializeSubtargetDependencies has been called so |
98 | // we can query directly. |
99 | InstrInfo(isThumb1Only() |
100 | ? (ARMBaseInstrInfo *)new Thumb1InstrInfo(*this) |
101 | : !isThumb() |
102 | ? (ARMBaseInstrInfo *)new ARMInstrInfo(*this) |
103 | : (ARMBaseInstrInfo *)new Thumb2InstrInfo(*this)), |
104 | TLInfo(TM, *this) { |
105 | |
106 | CallLoweringInfo.reset(p: new ARMCallLowering(*getTargetLowering())); |
107 | Legalizer.reset(p: new ARMLegalizerInfo(*this)); |
108 | |
109 | auto *RBI = new ARMRegisterBankInfo(*getRegisterInfo()); |
110 | |
111 | // FIXME: At this point, we can't rely on Subtarget having RBI. |
112 | // It's awkward to mix passing RBI and the Subtarget; should we pass |
113 | // TII/TRI as well? |
114 | InstSelector.reset(p: createARMInstructionSelector(TM, STI: *this, RBI: *RBI)); |
115 | |
116 | RegBankInfo.reset(p: RBI); |
117 | } |
118 | |
119 | const CallLowering *ARMSubtarget::getCallLowering() const { |
120 | return CallLoweringInfo.get(); |
121 | } |
122 | |
123 | InstructionSelector *ARMSubtarget::getInstructionSelector() const { |
124 | return InstSelector.get(); |
125 | } |
126 | |
127 | const LegalizerInfo *ARMSubtarget::getLegalizerInfo() const { |
128 | return Legalizer.get(); |
129 | } |
130 | |
131 | const RegisterBankInfo *ARMSubtarget::getRegBankInfo() const { |
132 | return RegBankInfo.get(); |
133 | } |
134 | |
135 | bool ARMSubtarget::isXRaySupported() const { |
136 | // We don't currently suppport Thumb, but Windows requires Thumb. |
137 | return hasV6Ops() && hasARMOps() && !isTargetWindows(); |
138 | } |
139 | |
140 | void ARMSubtarget::initializeEnvironment() { |
141 | // MCAsmInfo isn't always present (e.g. in opt) so we can't initialize this |
142 | // directly from it, but we can try to make sure they're consistent when both |
143 | // available. |
144 | UseSjLjEH = (isTargetDarwin() && !isTargetWatchABI() && |
145 | Options.ExceptionModel == ExceptionHandling::None) || |
146 | Options.ExceptionModel == ExceptionHandling::SjLj; |
147 | assert((!TM.getMCAsmInfo() || |
148 | (TM.getMCAsmInfo()->getExceptionHandlingType() == |
149 | ExceptionHandling::SjLj) == UseSjLjEH) && |
150 | "inconsistent sjlj choice between CodeGen and MC" ); |
151 | } |
152 | |
153 | void ARMSubtarget::initSubtargetFeatures(StringRef CPU, StringRef FS) { |
154 | if (CPUString.empty()) { |
155 | CPUString = "generic" ; |
156 | |
157 | if (isTargetDarwin()) { |
158 | StringRef ArchName = TargetTriple.getArchName(); |
159 | ARM::ArchKind AK = ARM::parseArch(Arch: ArchName); |
160 | if (AK == ARM::ArchKind::ARMV7S) |
161 | // Default to the Swift CPU when targeting armv7s/thumbv7s. |
162 | CPUString = "swift" ; |
163 | else if (AK == ARM::ArchKind::ARMV7K) |
164 | // Default to the Cortex-a7 CPU when targeting armv7k/thumbv7k. |
165 | // ARMv7k does not use SjLj exception handling. |
166 | CPUString = "cortex-a7" ; |
167 | } |
168 | } |
169 | |
170 | // Insert the architecture feature derived from the target triple into the |
171 | // feature string. This is important for setting features that are implied |
172 | // based on the architecture version. |
173 | std::string ArchFS = ARM_MC::ParseARMTriple(TT: TargetTriple, CPU: CPUString); |
174 | if (!FS.empty()) { |
175 | if (!ArchFS.empty()) |
176 | ArchFS = (Twine(ArchFS) + "," + FS).str(); |
177 | else |
178 | ArchFS = std::string(FS); |
179 | } |
180 | ParseSubtargetFeatures(CPU: CPUString, /*TuneCPU*/ CPUString, FS: ArchFS); |
181 | |
182 | // FIXME: This used enable V6T2 support implicitly for Thumb2 mode. |
183 | // Assert this for now to make the change obvious. |
184 | assert(hasV6T2Ops() || !hasThumb2()); |
185 | |
186 | if (genExecuteOnly()) { |
187 | // Execute only support for >= v8-M Baseline requires movt support |
188 | if (hasV8MBaselineOps()) |
189 | NoMovt = false; |
190 | if (!hasV6MOps()) |
191 | report_fatal_error(reason: "Cannot generate execute-only code for this target" ); |
192 | } |
193 | |
194 | // Keep a pointer to static instruction cost data for the specified CPU. |
195 | SchedModel = getSchedModelForCPU(CPU: CPUString); |
196 | |
197 | // Initialize scheduling itinerary for the specified CPU. |
198 | InstrItins = getInstrItineraryForCPU(CPU: CPUString); |
199 | |
200 | // FIXME: this is invalid for WindowsCE |
201 | if (isTargetWindows()) |
202 | NoARM = true; |
203 | |
204 | if (isAAPCS_ABI()) |
205 | stackAlignment = Align(8); |
206 | if (isTargetNaCl() || isAAPCS16_ABI()) |
207 | stackAlignment = Align(16); |
208 | |
209 | // FIXME: Completely disable sibcall for Thumb1 since ThumbRegisterInfo:: |
210 | // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as |
211 | // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation |
212 | // support in the assembler and linker to be used. This would need to be |
213 | // fixed to fully support tail calls in Thumb1. |
214 | // |
215 | // For ARMv8-M, we /do/ implement tail calls. Doing this is tricky for v8-M |
216 | // baseline, since the LDM/POP instruction on Thumb doesn't take LR. This |
217 | // means if we need to reload LR, it takes extra instructions, which outweighs |
218 | // the value of the tail call; but here we don't know yet whether LR is going |
219 | // to be used. We take the optimistic approach of generating the tail call and |
220 | // perhaps taking a hit if we need to restore the LR. |
221 | |
222 | // Thumb1 PIC calls to external symbols use BX, so they can be tail calls, |
223 | // but we need to make sure there are enough registers; the only valid |
224 | // registers are the 4 used for parameters. We don't currently do this |
225 | // case. |
226 | |
227 | SupportsTailCall = !isThumb1Only() || hasV8MBaselineOps(); |
228 | |
229 | if (isTargetMachO() && isTargetIOS() && getTargetTriple().isOSVersionLT(Major: 5, Minor: 0)) |
230 | SupportsTailCall = false; |
231 | |
232 | switch (IT) { |
233 | case DefaultIT: |
234 | RestrictIT = false; |
235 | break; |
236 | case RestrictedIT: |
237 | RestrictIT = true; |
238 | break; |
239 | } |
240 | |
241 | // NEON f32 ops are non-IEEE 754 compliant. Darwin is ok with it by default. |
242 | const FeatureBitset &Bits = getFeatureBits(); |
243 | if ((Bits[ARM::ProcA5] || Bits[ARM::ProcA8]) && // Where this matters |
244 | (Options.UnsafeFPMath || isTargetDarwin())) |
245 | HasNEONForFP = true; |
246 | |
247 | if (isRWPI()) |
248 | ReserveR9 = true; |
249 | |
250 | // If MVEVectorCostFactor is still 0 (has not been set to anything else), default it to 2 |
251 | if (MVEVectorCostFactor == 0) |
252 | MVEVectorCostFactor = 2; |
253 | |
254 | // FIXME: Teach TableGen to deal with these instead of doing it manually here. |
255 | switch (ARMProcFamily) { |
256 | case Others: |
257 | case CortexA5: |
258 | break; |
259 | case CortexA7: |
260 | LdStMultipleTiming = DoubleIssue; |
261 | break; |
262 | case CortexA8: |
263 | LdStMultipleTiming = DoubleIssue; |
264 | break; |
265 | case CortexA9: |
266 | LdStMultipleTiming = DoubleIssueCheckUnalignedAccess; |
267 | PreISelOperandLatencyAdjustment = 1; |
268 | break; |
269 | case CortexA12: |
270 | break; |
271 | case CortexA15: |
272 | MaxInterleaveFactor = 2; |
273 | PreISelOperandLatencyAdjustment = 1; |
274 | PartialUpdateClearance = 12; |
275 | break; |
276 | case CortexA17: |
277 | case CortexA32: |
278 | case CortexA35: |
279 | case CortexA53: |
280 | case CortexA55: |
281 | case CortexA57: |
282 | case CortexA72: |
283 | case CortexA73: |
284 | case CortexA75: |
285 | case CortexA76: |
286 | case CortexA77: |
287 | case CortexA78: |
288 | case CortexA78AE: |
289 | case CortexA78C: |
290 | case CortexA710: |
291 | case CortexR4: |
292 | case CortexR5: |
293 | case CortexR7: |
294 | case CortexM3: |
295 | case CortexM7: |
296 | case CortexR52: |
297 | case CortexR52plus: |
298 | case CortexX1: |
299 | case CortexX1C: |
300 | break; |
301 | case Exynos: |
302 | LdStMultipleTiming = SingleIssuePlusExtras; |
303 | MaxInterleaveFactor = 4; |
304 | if (!isThumb()) |
305 | PrefLoopLogAlignment = 3; |
306 | break; |
307 | case Kryo: |
308 | break; |
309 | case Krait: |
310 | PreISelOperandLatencyAdjustment = 1; |
311 | break; |
312 | case NeoverseV1: |
313 | break; |
314 | case Swift: |
315 | MaxInterleaveFactor = 2; |
316 | LdStMultipleTiming = SingleIssuePlusExtras; |
317 | PreISelOperandLatencyAdjustment = 1; |
318 | PartialUpdateClearance = 12; |
319 | break; |
320 | } |
321 | } |
322 | |
323 | bool ARMSubtarget::isTargetHardFloat() const { return TM.isTargetHardFloat(); } |
324 | |
325 | bool ARMSubtarget::isAPCS_ABI() const { |
326 | assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); |
327 | return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_APCS; |
328 | } |
329 | bool ARMSubtarget::isAAPCS_ABI() const { |
330 | assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); |
331 | return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS || |
332 | TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS16; |
333 | } |
334 | bool ARMSubtarget::isAAPCS16_ABI() const { |
335 | assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); |
336 | return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS16; |
337 | } |
338 | |
339 | bool ARMSubtarget::isROPI() const { |
340 | return TM.getRelocationModel() == Reloc::ROPI || |
341 | TM.getRelocationModel() == Reloc::ROPI_RWPI; |
342 | } |
343 | bool ARMSubtarget::isRWPI() const { |
344 | return TM.getRelocationModel() == Reloc::RWPI || |
345 | TM.getRelocationModel() == Reloc::ROPI_RWPI; |
346 | } |
347 | |
348 | bool ARMSubtarget::isGVIndirectSymbol(const GlobalValue *GV) const { |
349 | if (!TM.shouldAssumeDSOLocal(GV)) |
350 | return true; |
351 | |
352 | // 32 bit macho has no relocation for a-b if a is undefined, even if b is in |
353 | // the section that is being relocated. This means we have to use o load even |
354 | // for GVs that are known to be local to the dso. |
355 | if (isTargetMachO() && TM.isPositionIndependent() && |
356 | (GV->isDeclarationForLinker() || GV->hasCommonLinkage())) |
357 | return true; |
358 | |
359 | return false; |
360 | } |
361 | |
362 | bool ARMSubtarget::isGVInGOT(const GlobalValue *GV) const { |
363 | return isTargetELF() && TM.isPositionIndependent() && !GV->isDSOLocal(); |
364 | } |
365 | |
366 | unsigned ARMSubtarget::getMispredictionPenalty() const { |
367 | return SchedModel.MispredictPenalty; |
368 | } |
369 | |
370 | bool ARMSubtarget::enableMachineScheduler() const { |
371 | // The MachineScheduler can increase register usage, so we use more high |
372 | // registers and end up with more T2 instructions that cannot be converted to |
373 | // T1 instructions. At least until we do better at converting to thumb1 |
374 | // instructions, on cortex-m at Oz where we are size-paranoid, don't use the |
375 | // Machine scheduler, relying on the DAG register pressure scheduler instead. |
376 | if (isMClass() && hasMinSize()) |
377 | return false; |
378 | // Enable the MachineScheduler before register allocation for subtargets |
379 | // with the use-misched feature. |
380 | return useMachineScheduler(); |
381 | } |
382 | |
383 | bool ARMSubtarget::enableSubRegLiveness() const { |
384 | // Enable SubRegLiveness for MVE to better optimize s subregs for mqpr regs |
385 | // and q subregs for qqqqpr regs. |
386 | return hasMVEIntegerOps(); |
387 | } |
388 | |
389 | bool ARMSubtarget::enableMachinePipeliner() const { |
390 | // Enable the MachinePipeliner before register allocation for subtargets |
391 | // with the use-mipipeliner feature. |
392 | return getSchedModel().hasInstrSchedModel() && useMachinePipeliner(); |
393 | } |
394 | |
395 | bool ARMSubtarget::useDFAforSMS() const { return false; } |
396 | |
397 | // This overrides the PostRAScheduler bit in the SchedModel for any CPU. |
398 | bool ARMSubtarget::enablePostRAScheduler() const { |
399 | if (enableMachineScheduler()) |
400 | return false; |
401 | if (disablePostRAScheduler()) |
402 | return false; |
403 | // Thumb1 cores will generally not benefit from post-ra scheduling |
404 | return !isThumb1Only(); |
405 | } |
406 | |
407 | bool ARMSubtarget::enablePostRAMachineScheduler() const { |
408 | if (!enableMachineScheduler()) |
409 | return false; |
410 | if (disablePostRAScheduler()) |
411 | return false; |
412 | return !isThumb1Only(); |
413 | } |
414 | |
415 | bool ARMSubtarget::useStride4VFPs() const { |
416 | // For general targets, the prologue can grow when VFPs are allocated with |
417 | // stride 4 (more vpush instructions). But WatchOS uses a compact unwind |
418 | // format which it's more important to get right. |
419 | return isTargetWatchABI() || |
420 | (useWideStrideVFP() && !OptMinSize); |
421 | } |
422 | |
423 | bool ARMSubtarget::useMovt() const { |
424 | // NOTE Windows on ARM needs to use mov.w/mov.t pairs to materialise 32-bit |
425 | // immediates as it is inherently position independent, and may be out of |
426 | // range otherwise. |
427 | return !NoMovt && hasV8MBaselineOps() && |
428 | (isTargetWindows() || !OptMinSize || genExecuteOnly()); |
429 | } |
430 | |
431 | bool ARMSubtarget::useFastISel() const { |
432 | // Enable fast-isel for any target, for testing only. |
433 | if (ForceFastISel) |
434 | return true; |
435 | |
436 | // Limit fast-isel to the targets that are or have been tested. |
437 | if (!hasV6Ops()) |
438 | return false; |
439 | |
440 | // Thumb2 support on iOS; ARM support on iOS, Linux and NaCl. |
441 | return TM.Options.EnableFastISel && |
442 | ((isTargetMachO() && !isThumb1Only()) || |
443 | (isTargetLinux() && !isThumb()) || (isTargetNaCl() && !isThumb())); |
444 | } |
445 | |
446 | unsigned ARMSubtarget::getGPRAllocationOrder(const MachineFunction &MF) const { |
447 | // The GPR register class has multiple possible allocation orders, with |
448 | // tradeoffs preferred by different sub-architectures and optimisation goals. |
449 | // The allocation orders are: |
450 | // 0: (the default tablegen order, not used) |
451 | // 1: r14, r0-r13 |
452 | // 2: r0-r7 |
453 | // 3: r0-r7, r12, lr, r8-r11 |
454 | // Note that the register allocator will change this order so that |
455 | // callee-saved registers are used later, as they require extra work in the |
456 | // prologue/epilogue (though we sometimes override that). |
457 | |
458 | // For thumb1-only targets, only the low registers are allocatable. |
459 | if (isThumb1Only()) |
460 | return 2; |
461 | |
462 | // Allocate low registers first, so we can select more 16-bit instructions. |
463 | // We also (in ignoreCSRForAllocationOrder) override the default behaviour |
464 | // with regards to callee-saved registers, because pushing extra registers is |
465 | // much cheaper (in terms of code size) than using high registers. After |
466 | // that, we allocate r12 (doesn't need to be saved), lr (saving it means we |
467 | // can return with the pop, don't need an extra "bx lr") and then the rest of |
468 | // the high registers. |
469 | if (isThumb2() && MF.getFunction().hasMinSize()) |
470 | return 3; |
471 | |
472 | // Otherwise, allocate in the default order, using LR first because saving it |
473 | // allows a shorter epilogue sequence. |
474 | return 1; |
475 | } |
476 | |
477 | bool ARMSubtarget::ignoreCSRForAllocationOrder(const MachineFunction &MF, |
478 | unsigned PhysReg) const { |
479 | // To minimize code size in Thumb2, we prefer the usage of low regs (lower |
480 | // cost per use) so we can use narrow encoding. By default, caller-saved |
481 | // registers (e.g. lr, r12) are always allocated first, regardless of |
482 | // their cost per use. When optForMinSize, we prefer the low regs even if |
483 | // they are CSR because usually push/pop can be folded into existing ones. |
484 | return isThumb2() && MF.getFunction().hasMinSize() && |
485 | ARM::GPRRegClass.contains(Reg: PhysReg); |
486 | } |
487 | |
488 | bool ARMSubtarget::splitFramePointerPush(const MachineFunction &MF) const { |
489 | const Function &F = MF.getFunction(); |
490 | if (!MF.getTarget().getMCAsmInfo()->usesWindowsCFI() || |
491 | !F.needsUnwindTableEntry()) |
492 | return false; |
493 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
494 | return MFI.hasVarSizedObjects() || getRegisterInfo()->hasStackRealignment(MF); |
495 | } |
496 | |