1 | //===--- AArch64.cpp - Implement AArch64 target feature support -----------===// |
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 AArch64 TargetInfo objects. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "AArch64.h" |
14 | #include "clang/Basic/Diagnostic.h" |
15 | #include "clang/Basic/LangOptions.h" |
16 | #include "clang/Basic/TargetBuiltins.h" |
17 | #include "clang/Basic/TargetInfo.h" |
18 | #include "llvm/ADT/APSInt.h" |
19 | #include "llvm/ADT/ArrayRef.h" |
20 | #include "llvm/ADT/StringExtras.h" |
21 | #include "llvm/ADT/StringSwitch.h" |
22 | #include "llvm/TargetParser/AArch64TargetParser.h" |
23 | #include "llvm/TargetParser/ARMTargetParserCommon.h" |
24 | #include <optional> |
25 | |
26 | using namespace clang; |
27 | using namespace clang::targets; |
28 | |
29 | static constexpr Builtin::Info BuiltinInfo[] = { |
30 | #define BUILTIN(ID, TYPE, ATTRS) \ |
31 | {#ID, TYPE, ATTRS, nullptr, HeaderDesc::NO_HEADER, ALL_LANGUAGES}, |
32 | #define TARGET_BUILTIN(ID, TYPE, ATTRS, FEATURE) \ |
33 | {#ID, TYPE, ATTRS, FEATURE, HeaderDesc::NO_HEADER, ALL_LANGUAGES}, |
34 | #include "clang/Basic/BuiltinsNEON.def" |
35 | |
36 | #define BUILTIN(ID, TYPE, ATTRS) \ |
37 | {#ID, TYPE, ATTRS, nullptr, HeaderDesc::NO_HEADER, ALL_LANGUAGES}, |
38 | #define TARGET_BUILTIN(ID, TYPE, ATTRS, FEATURE) \ |
39 | {#ID, TYPE, ATTRS, FEATURE, HeaderDesc::NO_HEADER, ALL_LANGUAGES}, |
40 | #include "clang/Basic/BuiltinsSVE.def" |
41 | |
42 | #define BUILTIN(ID, TYPE, ATTRS) \ |
43 | {#ID, TYPE, ATTRS, nullptr, HeaderDesc::NO_HEADER, ALL_LANGUAGES}, |
44 | #define TARGET_BUILTIN(ID, TYPE, ATTRS, FEATURE) \ |
45 | {#ID, TYPE, ATTRS, FEATURE, HeaderDesc::NO_HEADER, ALL_LANGUAGES}, |
46 | #include "clang/Basic/BuiltinsSME.def" |
47 | |
48 | #define BUILTIN(ID, TYPE, ATTRS) \ |
49 | {#ID, TYPE, ATTRS, nullptr, HeaderDesc::NO_HEADER, ALL_LANGUAGES}, |
50 | #define LANGBUILTIN(ID, TYPE, ATTRS, LANG) \ |
51 | {#ID, TYPE, ATTRS, nullptr, HeaderDesc::NO_HEADER, LANG}, |
52 | #define TARGET_BUILTIN(ID, TYPE, ATTRS, FEATURE) \ |
53 | {#ID, TYPE, ATTRS, FEATURE, HeaderDesc::NO_HEADER, ALL_LANGUAGES}, |
54 | #define (ID, TYPE, ATTRS, HEADER, LANGS, FEATURE) \ |
55 | {#ID, TYPE, ATTRS, FEATURE, HeaderDesc::HEADER, LANGS}, |
56 | #include "clang/Basic/BuiltinsAArch64.def" |
57 | }; |
58 | |
59 | void AArch64TargetInfo::setArchFeatures() { |
60 | if (*ArchInfo == llvm::AArch64::ARMV8R) { |
61 | HasDotProd = true; |
62 | HasDIT = true; |
63 | HasFlagM = true; |
64 | HasRCPC = true; |
65 | FPU |= NeonMode; |
66 | HasCCPP = true; |
67 | HasCRC = true; |
68 | HasLSE = true; |
69 | HasRDM = true; |
70 | } else if (ArchInfo->Version.getMajor() == 8) { |
71 | if (ArchInfo->Version.getMinor() >= 7u) { |
72 | HasWFxT = true; |
73 | } |
74 | if (ArchInfo->Version.getMinor() >= 6u) { |
75 | HasBFloat16 = true; |
76 | HasMatMul = true; |
77 | } |
78 | if (ArchInfo->Version.getMinor() >= 5u) { |
79 | HasAlternativeNZCV = true; |
80 | HasFRInt3264 = true; |
81 | HasSSBS = true; |
82 | HasSB = true; |
83 | HasPredRes = true; |
84 | HasBTI = true; |
85 | } |
86 | if (ArchInfo->Version.getMinor() >= 4u) { |
87 | HasDotProd = true; |
88 | HasDIT = true; |
89 | HasFlagM = true; |
90 | } |
91 | if (ArchInfo->Version.getMinor() >= 3u) { |
92 | HasRCPC = true; |
93 | FPU |= NeonMode; |
94 | } |
95 | if (ArchInfo->Version.getMinor() >= 2u) { |
96 | HasCCPP = true; |
97 | } |
98 | if (ArchInfo->Version.getMinor() >= 1u) { |
99 | HasCRC = true; |
100 | HasLSE = true; |
101 | HasRDM = true; |
102 | } |
103 | } else if (ArchInfo->Version.getMajor() == 9) { |
104 | if (ArchInfo->Version.getMinor() >= 2u) { |
105 | HasWFxT = true; |
106 | } |
107 | if (ArchInfo->Version.getMinor() >= 1u) { |
108 | HasBFloat16 = true; |
109 | HasMatMul = true; |
110 | } |
111 | FPU |= SveMode; |
112 | HasSVE2 = true; |
113 | HasFullFP16 = true; |
114 | HasAlternativeNZCV = true; |
115 | HasFRInt3264 = true; |
116 | HasSSBS = true; |
117 | HasSB = true; |
118 | HasPredRes = true; |
119 | HasBTI = true; |
120 | HasDotProd = true; |
121 | HasDIT = true; |
122 | HasFlagM = true; |
123 | HasRCPC = true; |
124 | FPU |= NeonMode; |
125 | HasCCPP = true; |
126 | HasCRC = true; |
127 | HasLSE = true; |
128 | HasRDM = true; |
129 | } |
130 | } |
131 | |
132 | AArch64TargetInfo::AArch64TargetInfo(const llvm::Triple &Triple, |
133 | const TargetOptions &Opts) |
134 | : TargetInfo(Triple), ABI("aapcs" ) { |
135 | if (getTriple().isOSOpenBSD()) { |
136 | Int64Type = SignedLongLong; |
137 | IntMaxType = SignedLongLong; |
138 | } else { |
139 | if (!getTriple().isOSDarwin() && !getTriple().isOSNetBSD()) |
140 | WCharType = UnsignedInt; |
141 | |
142 | Int64Type = SignedLong; |
143 | IntMaxType = SignedLong; |
144 | } |
145 | |
146 | // All AArch64 implementations support ARMv8 FP, which makes half a legal type. |
147 | HasLegalHalfType = true; |
148 | HalfArgsAndReturns = true; |
149 | HasFloat16 = true; |
150 | HasStrictFP = true; |
151 | |
152 | if (Triple.isArch64Bit()) |
153 | LongWidth = LongAlign = PointerWidth = PointerAlign = 64; |
154 | else |
155 | LongWidth = LongAlign = PointerWidth = PointerAlign = 32; |
156 | |
157 | BitIntMaxAlign = 128; |
158 | MaxVectorAlign = 128; |
159 | MaxAtomicInlineWidth = 128; |
160 | MaxAtomicPromoteWidth = 128; |
161 | |
162 | LongDoubleWidth = LongDoubleAlign = SuitableAlign = 128; |
163 | LongDoubleFormat = &llvm::APFloat::IEEEquad(); |
164 | |
165 | BFloat16Width = BFloat16Align = 16; |
166 | BFloat16Format = &llvm::APFloat::BFloat(); |
167 | |
168 | // Make __builtin_ms_va_list available. |
169 | HasBuiltinMSVaList = true; |
170 | |
171 | // Make the SVE types available. Note that this deliberately doesn't |
172 | // depend on SveMode, since in principle it should be possible to turn |
173 | // SVE on and off within a translation unit. It should also be possible |
174 | // to compile the global declaration: |
175 | // |
176 | // __SVInt8_t *ptr; |
177 | // |
178 | // even without SVE. |
179 | HasAArch64SVETypes = true; |
180 | |
181 | // {} in inline assembly are neon specifiers, not assembly variant |
182 | // specifiers. |
183 | NoAsmVariants = true; |
184 | |
185 | // AAPCS gives rules for bitfields. 7.1.7 says: "The container type |
186 | // contributes to the alignment of the containing aggregate in the same way |
187 | // a plain (non bit-field) member of that type would, without exception for |
188 | // zero-sized or anonymous bit-fields." |
189 | assert(UseBitFieldTypeAlignment && "bitfields affect type alignment" ); |
190 | UseZeroLengthBitfieldAlignment = true; |
191 | |
192 | HasUnalignedAccess = true; |
193 | |
194 | // AArch64 targets default to using the ARM C++ ABI. |
195 | TheCXXABI.set(TargetCXXABI::GenericAArch64); |
196 | |
197 | if (Triple.getOS() == llvm::Triple::Linux) |
198 | this->MCountName = "\01_mcount" ; |
199 | else if (Triple.getOS() == llvm::Triple::UnknownOS) |
200 | this->MCountName = |
201 | Opts.EABIVersion == llvm::EABI::GNU ? "\01_mcount" : "mcount" ; |
202 | } |
203 | |
204 | StringRef AArch64TargetInfo::getABI() const { return ABI; } |
205 | |
206 | bool AArch64TargetInfo::setABI(const std::string &Name) { |
207 | if (Name != "aapcs" && Name != "aapcs-soft" && Name != "darwinpcs" && |
208 | Name != "pauthtest" ) |
209 | return false; |
210 | |
211 | ABI = Name; |
212 | return true; |
213 | } |
214 | |
215 | bool AArch64TargetInfo::validateTarget(DiagnosticsEngine &Diags) const { |
216 | if (hasFeature(Feature: "fp" ) && ABI == "aapcs-soft" ) { |
217 | // aapcs-soft is not allowed for targets with an FPU, to avoid there being |
218 | // two incomatible ABIs. |
219 | Diags.Report(DiagID: diag::err_target_unsupported_abi_with_fpu) << ABI; |
220 | return false; |
221 | } |
222 | if (getTriple().getEnvironment() == llvm::Triple::PAuthTest && |
223 | getTriple().getOS() != llvm::Triple::Linux) { |
224 | Diags.Report(DiagID: diag::err_target_unsupported_abi_for_triple) |
225 | << getTriple().getEnvironmentName() << getTriple().getTriple(); |
226 | return false; |
227 | } |
228 | return true; |
229 | } |
230 | |
231 | bool AArch64TargetInfo::validateGlobalRegisterVariable( |
232 | StringRef RegName, unsigned RegSize, bool &HasSizeMismatch) const { |
233 | if ((RegName == "sp" ) || RegName.starts_with(Prefix: "x" )) { |
234 | HasSizeMismatch = RegSize != 64; |
235 | return true; |
236 | } else if (RegName.starts_with(Prefix: "w" )) { |
237 | HasSizeMismatch = RegSize != 32; |
238 | return true; |
239 | } |
240 | return false; |
241 | } |
242 | |
243 | bool AArch64TargetInfo::validateBranchProtection(StringRef Spec, StringRef, |
244 | BranchProtectionInfo &BPI, |
245 | StringRef &Err) const { |
246 | llvm::ARM::ParsedBranchProtection PBP; |
247 | if (!llvm::ARM::parseBranchProtection(Spec, PBP, Err, EnablePAuthLR: HasPAuthLR)) |
248 | return false; |
249 | |
250 | BPI.SignReturnAddr = |
251 | llvm::StringSwitch<LangOptions::SignReturnAddressScopeKind>(PBP.Scope) |
252 | .Case(S: "non-leaf" , Value: LangOptions::SignReturnAddressScopeKind::NonLeaf) |
253 | .Case(S: "all" , Value: LangOptions::SignReturnAddressScopeKind::All) |
254 | .Default(Value: LangOptions::SignReturnAddressScopeKind::None); |
255 | |
256 | if (PBP.Key == "a_key" ) |
257 | BPI.SignKey = LangOptions::SignReturnAddressKeyKind::AKey; |
258 | else |
259 | BPI.SignKey = LangOptions::SignReturnAddressKeyKind::BKey; |
260 | |
261 | BPI.BranchTargetEnforcement = PBP.BranchTargetEnforcement; |
262 | BPI.BranchProtectionPAuthLR = PBP.BranchProtectionPAuthLR; |
263 | BPI.GuardedControlStack = PBP.GuardedControlStack; |
264 | return true; |
265 | } |
266 | |
267 | bool AArch64TargetInfo::isValidCPUName(StringRef Name) const { |
268 | return llvm::AArch64::parseCpu(Name).has_value(); |
269 | } |
270 | |
271 | bool AArch64TargetInfo::setCPU(const std::string &Name) { |
272 | return isValidCPUName(Name); |
273 | } |
274 | |
275 | void AArch64TargetInfo::fillValidCPUList( |
276 | SmallVectorImpl<StringRef> &Values) const { |
277 | llvm::AArch64::fillValidCPUArchList(Values); |
278 | } |
279 | |
280 | void AArch64TargetInfo::getTargetDefinesARMV81A(const LangOptions &Opts, |
281 | MacroBuilder &Builder) const { |
282 | Builder.defineMacro(Name: "__ARM_FEATURE_QRDMX" , Value: "1" ); |
283 | } |
284 | |
285 | void AArch64TargetInfo::getTargetDefinesARMV82A(const LangOptions &Opts, |
286 | MacroBuilder &Builder) const { |
287 | // Also include the ARMv8.1 defines |
288 | getTargetDefinesARMV81A(Opts, Builder); |
289 | } |
290 | |
291 | void AArch64TargetInfo::getTargetDefinesARMV83A(const LangOptions &Opts, |
292 | MacroBuilder &Builder) const { |
293 | Builder.defineMacro(Name: "__ARM_FEATURE_COMPLEX" , Value: "1" ); |
294 | Builder.defineMacro(Name: "__ARM_FEATURE_JCVT" , Value: "1" ); |
295 | // Also include the Armv8.2 defines |
296 | getTargetDefinesARMV82A(Opts, Builder); |
297 | } |
298 | |
299 | void AArch64TargetInfo::getTargetDefinesARMV84A(const LangOptions &Opts, |
300 | MacroBuilder &Builder) const { |
301 | // Also include the Armv8.3 defines |
302 | getTargetDefinesARMV83A(Opts, Builder); |
303 | } |
304 | |
305 | void AArch64TargetInfo::getTargetDefinesARMV85A(const LangOptions &Opts, |
306 | MacroBuilder &Builder) const { |
307 | Builder.defineMacro(Name: "__ARM_FEATURE_FRINT" , Value: "1" ); |
308 | // Also include the Armv8.4 defines |
309 | getTargetDefinesARMV84A(Opts, Builder); |
310 | } |
311 | |
312 | void AArch64TargetInfo::getTargetDefinesARMV86A(const LangOptions &Opts, |
313 | MacroBuilder &Builder) const { |
314 | // Also include the Armv8.5 defines |
315 | // FIXME: Armv8.6 makes the following extensions mandatory: |
316 | // - __ARM_FEATURE_BF16 |
317 | // - __ARM_FEATURE_MATMUL_INT8 |
318 | // Handle them here. |
319 | getTargetDefinesARMV85A(Opts, Builder); |
320 | } |
321 | |
322 | void AArch64TargetInfo::getTargetDefinesARMV87A(const LangOptions &Opts, |
323 | MacroBuilder &Builder) const { |
324 | // Also include the Armv8.6 defines |
325 | getTargetDefinesARMV86A(Opts, Builder); |
326 | } |
327 | |
328 | void AArch64TargetInfo::getTargetDefinesARMV88A(const LangOptions &Opts, |
329 | MacroBuilder &Builder) const { |
330 | // Also include the Armv8.7 defines |
331 | getTargetDefinesARMV87A(Opts, Builder); |
332 | } |
333 | |
334 | void AArch64TargetInfo::getTargetDefinesARMV89A(const LangOptions &Opts, |
335 | MacroBuilder &Builder) const { |
336 | // Also include the Armv8.8 defines |
337 | getTargetDefinesARMV88A(Opts, Builder); |
338 | } |
339 | |
340 | void AArch64TargetInfo::getTargetDefinesARMV9A(const LangOptions &Opts, |
341 | MacroBuilder &Builder) const { |
342 | // Armv9-A maps to Armv8.5-A |
343 | getTargetDefinesARMV85A(Opts, Builder); |
344 | } |
345 | |
346 | void AArch64TargetInfo::getTargetDefinesARMV91A(const LangOptions &Opts, |
347 | MacroBuilder &Builder) const { |
348 | // Armv9.1-A maps to Armv8.6-A |
349 | getTargetDefinesARMV86A(Opts, Builder); |
350 | } |
351 | |
352 | void AArch64TargetInfo::getTargetDefinesARMV92A(const LangOptions &Opts, |
353 | MacroBuilder &Builder) const { |
354 | // Armv9.2-A maps to Armv8.7-A |
355 | getTargetDefinesARMV87A(Opts, Builder); |
356 | } |
357 | |
358 | void AArch64TargetInfo::getTargetDefinesARMV93A(const LangOptions &Opts, |
359 | MacroBuilder &Builder) const { |
360 | // Armv9.3-A maps to Armv8.8-A |
361 | getTargetDefinesARMV88A(Opts, Builder); |
362 | } |
363 | |
364 | void AArch64TargetInfo::getTargetDefinesARMV94A(const LangOptions &Opts, |
365 | MacroBuilder &Builder) const { |
366 | // Armv9.4-A maps to Armv8.9-A |
367 | getTargetDefinesARMV89A(Opts, Builder); |
368 | } |
369 | |
370 | void AArch64TargetInfo::getTargetDefinesARMV95A(const LangOptions &Opts, |
371 | MacroBuilder &Builder) const { |
372 | // Armv9.5-A does not have a v8.* equivalent, but is a superset of v9.4-A. |
373 | getTargetDefinesARMV94A(Opts, Builder); |
374 | } |
375 | |
376 | void AArch64TargetInfo::getTargetDefines(const LangOptions &Opts, |
377 | MacroBuilder &Builder) const { |
378 | // Target identification. |
379 | if (getTriple().isWindowsArm64EC()) { |
380 | // Define the same set of macros as would be defined on x86_64 to ensure that |
381 | // ARM64EC datatype layouts match those of x86_64 compiled code |
382 | Builder.defineMacro(Name: "__amd64__" ); |
383 | Builder.defineMacro(Name: "__amd64" ); |
384 | Builder.defineMacro(Name: "__x86_64" ); |
385 | Builder.defineMacro(Name: "__x86_64__" ); |
386 | Builder.defineMacro(Name: "__arm64ec__" ); |
387 | } else { |
388 | Builder.defineMacro(Name: "__aarch64__" ); |
389 | } |
390 | |
391 | // Inline assembly supports AArch64 flag outputs. |
392 | Builder.defineMacro(Name: "__GCC_ASM_FLAG_OUTPUTS__" ); |
393 | |
394 | std::string CodeModel = getTargetOpts().CodeModel; |
395 | if (CodeModel == "default" ) |
396 | CodeModel = "small" ; |
397 | for (char &c : CodeModel) |
398 | c = toupper(c: c); |
399 | Builder.defineMacro(Name: "__AARCH64_CMODEL_" + CodeModel + "__" ); |
400 | |
401 | // ACLE predefines. Many can only have one possible value on v8 AArch64. |
402 | Builder.defineMacro(Name: "__ARM_ACLE" , Value: "200" ); |
403 | Builder.defineMacro(Name: "__ARM_ARCH" , |
404 | Value: std::to_string(val: ArchInfo->Version.getMajor())); |
405 | Builder.defineMacro(Name: "__ARM_ARCH_PROFILE" , |
406 | Value: std::string("'" ) + (char)ArchInfo->Profile + "'" ); |
407 | |
408 | Builder.defineMacro(Name: "__ARM_64BIT_STATE" , Value: "1" ); |
409 | Builder.defineMacro(Name: "__ARM_PCS_AAPCS64" , Value: "1" ); |
410 | Builder.defineMacro(Name: "__ARM_ARCH_ISA_A64" , Value: "1" ); |
411 | |
412 | Builder.defineMacro(Name: "__ARM_FEATURE_CLZ" , Value: "1" ); |
413 | Builder.defineMacro(Name: "__ARM_FEATURE_FMA" , Value: "1" ); |
414 | Builder.defineMacro(Name: "__ARM_FEATURE_LDREX" , Value: "0xF" ); |
415 | Builder.defineMacro(Name: "__ARM_FEATURE_IDIV" , Value: "1" ); // As specified in ACLE |
416 | Builder.defineMacro(Name: "__ARM_FEATURE_DIV" ); // For backwards compatibility |
417 | Builder.defineMacro(Name: "__ARM_FEATURE_NUMERIC_MAXMIN" , Value: "1" ); |
418 | Builder.defineMacro(Name: "__ARM_FEATURE_DIRECTED_ROUNDING" , Value: "1" ); |
419 | |
420 | Builder.defineMacro(Name: "__ARM_ALIGN_MAX_STACK_PWR" , Value: "4" ); |
421 | |
422 | // These macros are set when Clang can parse declarations with these |
423 | // attributes. |
424 | Builder.defineMacro(Name: "__ARM_STATE_ZA" , Value: "1" ); |
425 | Builder.defineMacro(Name: "__ARM_STATE_ZT0" , Value: "1" ); |
426 | |
427 | // 0xe implies support for half, single and double precision operations. |
428 | if (FPU & FPUMode) |
429 | Builder.defineMacro(Name: "__ARM_FP" , Value: "0xE" ); |
430 | |
431 | // PCS specifies this for SysV variants, which is all we support. Other ABIs |
432 | // may choose __ARM_FP16_FORMAT_ALTERNATIVE. |
433 | Builder.defineMacro(Name: "__ARM_FP16_FORMAT_IEEE" , Value: "1" ); |
434 | Builder.defineMacro(Name: "__ARM_FP16_ARGS" , Value: "1" ); |
435 | |
436 | if (Opts.UnsafeFPMath) |
437 | Builder.defineMacro(Name: "__ARM_FP_FAST" , Value: "1" ); |
438 | |
439 | Builder.defineMacro(Name: "__ARM_SIZEOF_WCHAR_T" , |
440 | Value: Twine(Opts.WCharSize ? Opts.WCharSize : 4)); |
441 | |
442 | Builder.defineMacro(Name: "__ARM_SIZEOF_MINIMAL_ENUM" , Value: Opts.ShortEnums ? "1" : "4" ); |
443 | |
444 | if (FPU & NeonMode) { |
445 | Builder.defineMacro(Name: "__ARM_NEON" , Value: "1" ); |
446 | // 64-bit NEON supports half, single and double precision operations. |
447 | Builder.defineMacro(Name: "__ARM_NEON_FP" , Value: "0xE" ); |
448 | } |
449 | |
450 | if (FPU & SveMode) |
451 | Builder.defineMacro(Name: "__ARM_FEATURE_SVE" , Value: "1" ); |
452 | |
453 | if ((FPU & NeonMode) && (FPU & SveMode)) |
454 | Builder.defineMacro(Name: "__ARM_NEON_SVE_BRIDGE" , Value: "1" ); |
455 | |
456 | if (HasSVE2) |
457 | Builder.defineMacro(Name: "__ARM_FEATURE_SVE2" , Value: "1" ); |
458 | |
459 | if (HasSVE2p1) |
460 | Builder.defineMacro(Name: "__ARM_FEATURE_SVE2p1" , Value: "1" ); |
461 | |
462 | if (HasSVE2 && HasSVE2AES) |
463 | Builder.defineMacro(Name: "__ARM_FEATURE_SVE2_AES" , Value: "1" ); |
464 | |
465 | if (HasSVE2 && HasSVE2BitPerm) |
466 | Builder.defineMacro(Name: "__ARM_FEATURE_SVE2_BITPERM" , Value: "1" ); |
467 | |
468 | if (HasSVE2 && HasSVE2SHA3) |
469 | Builder.defineMacro(Name: "__ARM_FEATURE_SVE2_SHA3" , Value: "1" ); |
470 | |
471 | if (HasSVE2 && HasSVE2SM4) |
472 | Builder.defineMacro(Name: "__ARM_FEATURE_SVE2_SM4" , Value: "1" ); |
473 | |
474 | if (HasSVEB16B16) |
475 | Builder.defineMacro(Name: "__ARM_FEATURE_SVE_B16B16" , Value: "1" ); |
476 | |
477 | if (HasSME) { |
478 | Builder.defineMacro(Name: "__ARM_FEATURE_SME" ); |
479 | Builder.defineMacro(Name: "__ARM_FEATURE_LOCALLY_STREAMING" , Value: "1" ); |
480 | } |
481 | |
482 | if (HasSME2) |
483 | Builder.defineMacro(Name: "__ARM_FEATURE_SME2" , Value: "1" ); |
484 | |
485 | if (HasSME2p1) |
486 | Builder.defineMacro(Name: "__ARM_FEATURE_SME2p1" , Value: "1" ); |
487 | |
488 | if (HasSMEF16F16) |
489 | Builder.defineMacro(Name: "__ARM_FEATURE_SME_F16F16" , Value: "1" ); |
490 | |
491 | if (HasSMEB16B16) |
492 | Builder.defineMacro(Name: "__ARM_FEATURE_SME_B16B16" , Value: "1" ); |
493 | |
494 | if (HasCRC) |
495 | Builder.defineMacro(Name: "__ARM_FEATURE_CRC32" , Value: "1" ); |
496 | |
497 | if (HasRCPC3) |
498 | Builder.defineMacro(Name: "__ARM_FEATURE_RCPC" , Value: "3" ); |
499 | else if (HasRCPC) |
500 | Builder.defineMacro(Name: "__ARM_FEATURE_RCPC" , Value: "1" ); |
501 | |
502 | if (HasFMV) |
503 | Builder.defineMacro(Name: "__HAVE_FUNCTION_MULTI_VERSIONING" , Value: "1" ); |
504 | |
505 | // The __ARM_FEATURE_CRYPTO is deprecated in favor of finer grained feature |
506 | // macros for AES, SHA2, SHA3 and SM4 |
507 | if (HasAES && HasSHA2) |
508 | Builder.defineMacro(Name: "__ARM_FEATURE_CRYPTO" , Value: "1" ); |
509 | |
510 | if (HasAES) |
511 | Builder.defineMacro(Name: "__ARM_FEATURE_AES" , Value: "1" ); |
512 | |
513 | if (HasSHA2) |
514 | Builder.defineMacro(Name: "__ARM_FEATURE_SHA2" , Value: "1" ); |
515 | |
516 | if (HasSHA3) { |
517 | Builder.defineMacro(Name: "__ARM_FEATURE_SHA3" , Value: "1" ); |
518 | Builder.defineMacro(Name: "__ARM_FEATURE_SHA512" , Value: "1" ); |
519 | } |
520 | |
521 | if (HasSM4) { |
522 | Builder.defineMacro(Name: "__ARM_FEATURE_SM3" , Value: "1" ); |
523 | Builder.defineMacro(Name: "__ARM_FEATURE_SM4" , Value: "1" ); |
524 | } |
525 | |
526 | if (HasPAuth) |
527 | Builder.defineMacro(Name: "__ARM_FEATURE_PAUTH" , Value: "1" ); |
528 | |
529 | if (HasPAuthLR) |
530 | Builder.defineMacro(Name: "__ARM_FEATURE_PAUTH_LR" , Value: "1" ); |
531 | |
532 | if (HasBTI) |
533 | Builder.defineMacro(Name: "__ARM_FEATURE_BTI" , Value: "1" ); |
534 | |
535 | if (HasUnalignedAccess) |
536 | Builder.defineMacro(Name: "__ARM_FEATURE_UNALIGNED" , Value: "1" ); |
537 | |
538 | if ((FPU & NeonMode) && HasFullFP16) |
539 | Builder.defineMacro(Name: "__ARM_FEATURE_FP16_VECTOR_ARITHMETIC" , Value: "1" ); |
540 | if (HasFullFP16) |
541 | Builder.defineMacro(Name: "__ARM_FEATURE_FP16_SCALAR_ARITHMETIC" , Value: "1" ); |
542 | |
543 | if (HasDotProd) |
544 | Builder.defineMacro(Name: "__ARM_FEATURE_DOTPROD" , Value: "1" ); |
545 | |
546 | if (HasMTE) |
547 | Builder.defineMacro(Name: "__ARM_FEATURE_MEMORY_TAGGING" , Value: "1" ); |
548 | |
549 | if (HasTME) |
550 | Builder.defineMacro(Name: "__ARM_FEATURE_TME" , Value: "1" ); |
551 | |
552 | if (HasMatMul) |
553 | Builder.defineMacro(Name: "__ARM_FEATURE_MATMUL_INT8" , Value: "1" ); |
554 | |
555 | if (HasLSE) |
556 | Builder.defineMacro(Name: "__ARM_FEATURE_ATOMICS" , Value: "1" ); |
557 | |
558 | if (HasBFloat16) { |
559 | Builder.defineMacro(Name: "__ARM_FEATURE_BF16" , Value: "1" ); |
560 | Builder.defineMacro(Name: "__ARM_FEATURE_BF16_VECTOR_ARITHMETIC" , Value: "1" ); |
561 | Builder.defineMacro(Name: "__ARM_BF16_FORMAT_ALTERNATIVE" , Value: "1" ); |
562 | Builder.defineMacro(Name: "__ARM_FEATURE_BF16_SCALAR_ARITHMETIC" , Value: "1" ); |
563 | } |
564 | |
565 | if ((FPU & SveMode) && HasBFloat16) { |
566 | Builder.defineMacro(Name: "__ARM_FEATURE_SVE_BF16" , Value: "1" ); |
567 | } |
568 | |
569 | if ((FPU & SveMode) && HasMatmulFP64) |
570 | Builder.defineMacro(Name: "__ARM_FEATURE_SVE_MATMUL_FP64" , Value: "1" ); |
571 | |
572 | if ((FPU & SveMode) && HasMatmulFP32) |
573 | Builder.defineMacro(Name: "__ARM_FEATURE_SVE_MATMUL_FP32" , Value: "1" ); |
574 | |
575 | if ((FPU & SveMode) && HasMatMul) |
576 | Builder.defineMacro(Name: "__ARM_FEATURE_SVE_MATMUL_INT8" , Value: "1" ); |
577 | |
578 | if ((FPU & NeonMode) && HasFP16FML) |
579 | Builder.defineMacro(Name: "__ARM_FEATURE_FP16_FML" , Value: "1" ); |
580 | |
581 | if (Opts.hasSignReturnAddress()) { |
582 | // Bitmask: |
583 | // 0: Protection using the A key |
584 | // 1: Protection using the B key |
585 | // 2: Protection including leaf functions |
586 | // 3: Protection using PC as a diversifier |
587 | unsigned Value = 0; |
588 | |
589 | if (Opts.isSignReturnAddressWithAKey()) |
590 | Value |= (1 << 0); |
591 | else |
592 | Value |= (1 << 1); |
593 | |
594 | if (Opts.isSignReturnAddressScopeAll()) |
595 | Value |= (1 << 2); |
596 | |
597 | if (Opts.BranchProtectionPAuthLR) |
598 | Value |= (1 << 3); |
599 | |
600 | Builder.defineMacro(Name: "__ARM_FEATURE_PAC_DEFAULT" , Value: std::to_string(val: Value)); |
601 | } |
602 | |
603 | if (Opts.BranchTargetEnforcement) |
604 | Builder.defineMacro(Name: "__ARM_FEATURE_BTI_DEFAULT" , Value: "1" ); |
605 | |
606 | if (Opts.GuardedControlStack) |
607 | Builder.defineMacro(Name: "__ARM_FEATURE_GCS_DEFAULT" , Value: "1" ); |
608 | |
609 | if (HasLS64) |
610 | Builder.defineMacro(Name: "__ARM_FEATURE_LS64" , Value: "1" ); |
611 | |
612 | if (HasRandGen) |
613 | Builder.defineMacro(Name: "__ARM_FEATURE_RNG" , Value: "1" ); |
614 | |
615 | if (HasMOPS) |
616 | Builder.defineMacro(Name: "__ARM_FEATURE_MOPS" , Value: "1" ); |
617 | |
618 | if (HasD128) |
619 | Builder.defineMacro(Name: "__ARM_FEATURE_SYSREG128" , Value: "1" ); |
620 | |
621 | if (HasGCS) |
622 | Builder.defineMacro(Name: "__ARM_FEATURE_GCS" , Value: "1" ); |
623 | |
624 | if (*ArchInfo == llvm::AArch64::ARMV8_1A) |
625 | getTargetDefinesARMV81A(Opts, Builder); |
626 | else if (*ArchInfo == llvm::AArch64::ARMV8_2A) |
627 | getTargetDefinesARMV82A(Opts, Builder); |
628 | else if (*ArchInfo == llvm::AArch64::ARMV8_3A) |
629 | getTargetDefinesARMV83A(Opts, Builder); |
630 | else if (*ArchInfo == llvm::AArch64::ARMV8_4A) |
631 | getTargetDefinesARMV84A(Opts, Builder); |
632 | else if (*ArchInfo == llvm::AArch64::ARMV8_5A) |
633 | getTargetDefinesARMV85A(Opts, Builder); |
634 | else if (*ArchInfo == llvm::AArch64::ARMV8_6A) |
635 | getTargetDefinesARMV86A(Opts, Builder); |
636 | else if (*ArchInfo == llvm::AArch64::ARMV8_7A) |
637 | getTargetDefinesARMV87A(Opts, Builder); |
638 | else if (*ArchInfo == llvm::AArch64::ARMV8_8A) |
639 | getTargetDefinesARMV88A(Opts, Builder); |
640 | else if (*ArchInfo == llvm::AArch64::ARMV8_9A) |
641 | getTargetDefinesARMV89A(Opts, Builder); |
642 | else if (*ArchInfo == llvm::AArch64::ARMV9A) |
643 | getTargetDefinesARMV9A(Opts, Builder); |
644 | else if (*ArchInfo == llvm::AArch64::ARMV9_1A) |
645 | getTargetDefinesARMV91A(Opts, Builder); |
646 | else if (*ArchInfo == llvm::AArch64::ARMV9_2A) |
647 | getTargetDefinesARMV92A(Opts, Builder); |
648 | else if (*ArchInfo == llvm::AArch64::ARMV9_3A) |
649 | getTargetDefinesARMV93A(Opts, Builder); |
650 | else if (*ArchInfo == llvm::AArch64::ARMV9_4A) |
651 | getTargetDefinesARMV94A(Opts, Builder); |
652 | else if (*ArchInfo == llvm::AArch64::ARMV9_5A) |
653 | getTargetDefinesARMV95A(Opts, Builder); |
654 | |
655 | // All of the __sync_(bool|val)_compare_and_swap_(1|2|4|8|16) builtins work. |
656 | Builder.defineMacro(Name: "__GCC_HAVE_SYNC_COMPARE_AND_SWAP_1" ); |
657 | Builder.defineMacro(Name: "__GCC_HAVE_SYNC_COMPARE_AND_SWAP_2" ); |
658 | Builder.defineMacro(Name: "__GCC_HAVE_SYNC_COMPARE_AND_SWAP_4" ); |
659 | Builder.defineMacro(Name: "__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8" ); |
660 | Builder.defineMacro(Name: "__GCC_HAVE_SYNC_COMPARE_AND_SWAP_16" ); |
661 | |
662 | // Allow detection of fast FMA support. |
663 | Builder.defineMacro(Name: "__FP_FAST_FMA" , Value: "1" ); |
664 | Builder.defineMacro(Name: "__FP_FAST_FMAF" , Value: "1" ); |
665 | |
666 | // C/C++ operators work on both VLS and VLA SVE types |
667 | if (FPU & SveMode) |
668 | Builder.defineMacro(Name: "__ARM_FEATURE_SVE_VECTOR_OPERATORS" , Value: "2" ); |
669 | |
670 | if (Opts.VScaleMin && Opts.VScaleMin == Opts.VScaleMax) { |
671 | Builder.defineMacro(Name: "__ARM_FEATURE_SVE_BITS" , Value: Twine(Opts.VScaleMin * 128)); |
672 | } |
673 | } |
674 | |
675 | ArrayRef<Builtin::Info> AArch64TargetInfo::getTargetBuiltins() const { |
676 | return llvm::ArrayRef(BuiltinInfo, clang::AArch64::LastTSBuiltin - |
677 | Builtin::FirstTSBuiltin); |
678 | } |
679 | |
680 | std::optional<std::pair<unsigned, unsigned>> |
681 | AArch64TargetInfo::getVScaleRange(const LangOptions &LangOpts) const { |
682 | if (LangOpts.VScaleMin || LangOpts.VScaleMax) |
683 | return std::pair<unsigned, unsigned>( |
684 | LangOpts.VScaleMin ? LangOpts.VScaleMin : 1, LangOpts.VScaleMax); |
685 | |
686 | if (hasFeature(Feature: "sve" )) |
687 | return std::pair<unsigned, unsigned>(1, 16); |
688 | |
689 | return std::nullopt; |
690 | } |
691 | |
692 | unsigned AArch64TargetInfo::multiVersionSortPriority(StringRef Name) const { |
693 | if (Name == "default" ) |
694 | return 0; |
695 | if (auto Ext = llvm::AArch64::parseFMVExtension(Extension: Name)) |
696 | return Ext->Priority; |
697 | return 0; |
698 | } |
699 | |
700 | unsigned AArch64TargetInfo::multiVersionFeatureCost() const { |
701 | // Take the maximum priority as per feature cost, so more features win. |
702 | constexpr unsigned MaxFMVPriority = 1000; |
703 | return MaxFMVPriority; |
704 | } |
705 | |
706 | bool AArch64TargetInfo::doesFeatureAffectCodeGen(StringRef Name) const { |
707 | // FMV extensions which imply no backend features do not affect codegen. |
708 | if (auto Ext = llvm::AArch64::parseFMVExtension(Extension: Name)) |
709 | return !Ext->Features.empty(); |
710 | return false; |
711 | } |
712 | |
713 | bool AArch64TargetInfo::validateCpuSupports(StringRef FeatureStr) const { |
714 | // CPU features might be separated by '+', extract them and check |
715 | llvm::SmallVector<StringRef, 8> Features; |
716 | FeatureStr.split(A&: Features, Separator: "+" ); |
717 | for (auto &Feature : Features) |
718 | if (!llvm::AArch64::parseFMVExtension(Extension: Feature.trim()).has_value()) |
719 | return false; |
720 | return true; |
721 | } |
722 | |
723 | bool AArch64TargetInfo::hasFeature(StringRef Feature) const { |
724 | return llvm::StringSwitch<bool>(Feature) |
725 | .Cases(S0: "aarch64" , S1: "arm64" , S2: "arm" , Value: true) |
726 | .Case(S: "fmv" , Value: HasFMV) |
727 | .Case(S: "fp" , Value: FPU & FPUMode) |
728 | .Cases(S0: "neon" , S1: "simd" , Value: FPU & NeonMode) |
729 | .Case(S: "jscvt" , Value: HasJSCVT) |
730 | .Case(S: "fcma" , Value: HasFCMA) |
731 | .Case(S: "rng" , Value: HasRandGen) |
732 | .Case(S: "flagm" , Value: HasFlagM) |
733 | .Case(S: "flagm2" , Value: HasAlternativeNZCV) |
734 | .Case(S: "fp16fml" , Value: HasFP16FML) |
735 | .Case(S: "dotprod" , Value: HasDotProd) |
736 | .Case(S: "sm4" , Value: HasSM4) |
737 | .Case(S: "rdm" , Value: HasRDM) |
738 | .Case(S: "lse" , Value: HasLSE) |
739 | .Case(S: "crc" , Value: HasCRC) |
740 | .Case(S: "sha2" , Value: HasSHA2) |
741 | .Case(S: "sha3" , Value: HasSHA3) |
742 | .Cases(S0: "aes" , S1: "pmull" , Value: HasAES) |
743 | .Cases(S0: "fp16" , S1: "fullfp16" , Value: HasFullFP16) |
744 | .Case(S: "dit" , Value: HasDIT) |
745 | .Case(S: "dpb" , Value: HasCCPP) |
746 | .Case(S: "dpb2" , Value: HasCCDP) |
747 | .Case(S: "rcpc" , Value: HasRCPC) |
748 | .Case(S: "frintts" , Value: HasFRInt3264) |
749 | .Case(S: "i8mm" , Value: HasMatMul) |
750 | .Case(S: "bf16" , Value: HasBFloat16) |
751 | .Case(S: "sve" , Value: FPU & SveMode) |
752 | .Case(S: "sve-bf16" , Value: FPU & SveMode && HasBFloat16) |
753 | .Case(S: "sve-i8mm" , Value: FPU & SveMode && HasMatMul) |
754 | .Case(S: "sve-b16b16" , Value: HasSVEB16B16) |
755 | .Case(S: "f32mm" , Value: FPU & SveMode && HasMatmulFP32) |
756 | .Case(S: "f64mm" , Value: FPU & SveMode && HasMatmulFP64) |
757 | .Case(S: "sve2" , Value: FPU & SveMode && HasSVE2) |
758 | .Case(S: "sve2-pmull128" , Value: FPU & SveMode && HasSVE2AES) |
759 | .Case(S: "sve2-bitperm" , Value: FPU & SveMode && HasSVE2BitPerm) |
760 | .Case(S: "sve2-sha3" , Value: FPU & SveMode && HasSVE2SHA3) |
761 | .Case(S: "sve2-sm4" , Value: FPU & SveMode && HasSVE2SM4) |
762 | .Case(S: "sve2p1" , Value: FPU & SveMode && HasSVE2p1) |
763 | .Case(S: "sme" , Value: HasSME) |
764 | .Case(S: "sme2" , Value: HasSME2) |
765 | .Case(S: "sme2p1" , Value: HasSME2p1) |
766 | .Case(S: "sme-f64f64" , Value: HasSMEF64F64) |
767 | .Case(S: "sme-i16i64" , Value: HasSMEI16I64) |
768 | .Case(S: "sme-fa64" , Value: HasSMEFA64) |
769 | .Case(S: "sme-f16f16" , Value: HasSMEF16F16) |
770 | .Case(S: "sme-b16b16" , Value: HasSMEB16B16) |
771 | .Cases(S0: "memtag" , S1: "memtag2" , Value: HasMTE) |
772 | .Case(S: "sb" , Value: HasSB) |
773 | .Case(S: "predres" , Value: HasPredRes) |
774 | .Cases(S0: "ssbs" , S1: "ssbs2" , Value: HasSSBS) |
775 | .Case(S: "bti" , Value: HasBTI) |
776 | .Cases(S0: "ls64" , S1: "ls64_v" , S2: "ls64_accdata" , Value: HasLS64) |
777 | .Case(S: "wfxt" , Value: HasWFxT) |
778 | .Case(S: "rcpc3" , Value: HasRCPC3) |
779 | .Default(Value: false); |
780 | } |
781 | |
782 | void AArch64TargetInfo::setFeatureEnabled(llvm::StringMap<bool> &Features, |
783 | StringRef Name, bool Enabled) const { |
784 | Features[Name] = Enabled; |
785 | // If the feature is an architecture feature (like v8.2a), add all previous |
786 | // architecture versions and any dependant target features. |
787 | const std::optional<llvm::AArch64::ArchInfo> ArchInfo = |
788 | llvm::AArch64::ArchInfo::findBySubArch(SubArch: Name); |
789 | |
790 | if (!ArchInfo) |
791 | return; // Not an architecture, nothing more to do. |
792 | |
793 | // Disabling an architecture feature does not affect dependent features |
794 | if (!Enabled) |
795 | return; |
796 | |
797 | for (const auto *OtherArch : llvm::AArch64::ArchInfos) |
798 | if (ArchInfo->implies(Other: *OtherArch)) |
799 | Features[OtherArch->getSubArch()] = true; |
800 | |
801 | // Set any features implied by the architecture |
802 | std::vector<StringRef> CPUFeats; |
803 | if (llvm::AArch64::getExtensionFeatures(Extensions: ArchInfo->DefaultExts, Features&: CPUFeats)) { |
804 | for (auto F : CPUFeats) { |
805 | assert(F[0] == '+' && "Expected + in target feature!" ); |
806 | Features[F.drop_front(N: 1)] = true; |
807 | } |
808 | } |
809 | } |
810 | |
811 | bool AArch64TargetInfo::handleTargetFeatures(std::vector<std::string> &Features, |
812 | DiagnosticsEngine &Diags) { |
813 | for (const auto &Feature : Features) { |
814 | if (Feature == "-fp-armv8" ) |
815 | HasNoFP = true; |
816 | if (Feature == "-neon" ) |
817 | HasNoNeon = true; |
818 | if (Feature == "-sve" ) |
819 | HasNoSVE = true; |
820 | |
821 | if (Feature == "+neon" || Feature == "+fp-armv8" ) |
822 | FPU |= NeonMode; |
823 | if (Feature == "+jscvt" ) { |
824 | HasJSCVT = true; |
825 | FPU |= NeonMode; |
826 | } |
827 | if (Feature == "+fcma" ) { |
828 | HasFCMA = true; |
829 | FPU |= NeonMode; |
830 | } |
831 | |
832 | if (Feature == "+sve" ) { |
833 | FPU |= NeonMode; |
834 | FPU |= SveMode; |
835 | HasFullFP16 = true; |
836 | } |
837 | if (Feature == "+sve2" ) { |
838 | FPU |= NeonMode; |
839 | FPU |= SveMode; |
840 | HasFullFP16 = true; |
841 | HasSVE2 = true; |
842 | } |
843 | if (Feature == "+sve2p1" ) { |
844 | FPU |= NeonMode; |
845 | FPU |= SveMode; |
846 | HasFullFP16 = true; |
847 | HasSVE2 = true; |
848 | HasSVE2p1 = true; |
849 | } |
850 | if (Feature == "+sve2-aes" ) { |
851 | FPU |= NeonMode; |
852 | FPU |= SveMode; |
853 | HasFullFP16 = true; |
854 | HasSVE2 = true; |
855 | HasSVE2AES = true; |
856 | } |
857 | if (Feature == "+sve2-sha3" ) { |
858 | FPU |= NeonMode; |
859 | FPU |= SveMode; |
860 | HasFullFP16 = true; |
861 | HasSVE2 = true; |
862 | HasSVE2SHA3 = true; |
863 | } |
864 | if (Feature == "+sve2-sm4" ) { |
865 | FPU |= NeonMode; |
866 | FPU |= SveMode; |
867 | HasFullFP16 = true; |
868 | HasSVE2 = true; |
869 | HasSVE2SM4 = true; |
870 | } |
871 | if (Feature == "+sve-b16b16" ) |
872 | HasSVEB16B16 = true; |
873 | if (Feature == "+sve2-bitperm" ) { |
874 | FPU |= NeonMode; |
875 | FPU |= SveMode; |
876 | HasFullFP16 = true; |
877 | HasSVE2 = true; |
878 | HasSVE2BitPerm = true; |
879 | } |
880 | if (Feature == "+f32mm" ) { |
881 | FPU |= NeonMode; |
882 | FPU |= SveMode; |
883 | HasFullFP16 = true; |
884 | HasMatmulFP32 = true; |
885 | } |
886 | if (Feature == "+f64mm" ) { |
887 | FPU |= NeonMode; |
888 | FPU |= SveMode; |
889 | HasFullFP16 = true; |
890 | HasMatmulFP64 = true; |
891 | } |
892 | if (Feature == "+sme" ) { |
893 | HasSME = true; |
894 | HasBFloat16 = true; |
895 | HasFullFP16 = true; |
896 | } |
897 | if (Feature == "+sme2" ) { |
898 | HasSME = true; |
899 | HasSME2 = true; |
900 | HasBFloat16 = true; |
901 | HasFullFP16 = true; |
902 | } |
903 | if (Feature == "+sme2p1" ) { |
904 | HasSME = true; |
905 | HasSME2 = true; |
906 | HasSME2p1 = true; |
907 | HasBFloat16 = true; |
908 | HasFullFP16 = true; |
909 | } |
910 | if (Feature == "+sme-f64f64" ) { |
911 | HasSME = true; |
912 | HasSMEF64F64 = true; |
913 | HasBFloat16 = true; |
914 | HasFullFP16 = true; |
915 | } |
916 | if (Feature == "+sme-i16i64" ) { |
917 | HasSME = true; |
918 | HasSMEI16I64 = true; |
919 | HasBFloat16 = true; |
920 | HasFullFP16 = true; |
921 | } |
922 | if (Feature == "+sme-fa64" ) { |
923 | FPU |= NeonMode; |
924 | FPU |= SveMode; |
925 | HasSME = true; |
926 | HasSVE2 = true; |
927 | HasSMEFA64 = true; |
928 | } |
929 | if (Feature == "+sme-f16f16" ) { |
930 | HasSME = true; |
931 | HasSME2 = true; |
932 | HasBFloat16 = true; |
933 | HasFullFP16 = true; |
934 | HasSMEF16F16 = true; |
935 | } |
936 | if (Feature == "+sme-b16b16" ) { |
937 | HasSME = true; |
938 | HasSME2 = true; |
939 | HasBFloat16 = true; |
940 | HasFullFP16 = true; |
941 | HasSVEB16B16 = true; |
942 | HasSMEB16B16 = true; |
943 | } |
944 | if (Feature == "+sb" ) |
945 | HasSB = true; |
946 | if (Feature == "+predres" ) |
947 | HasPredRes = true; |
948 | if (Feature == "+ssbs" ) |
949 | HasSSBS = true; |
950 | if (Feature == "+bti" ) |
951 | HasBTI = true; |
952 | if (Feature == "+wfxt" ) |
953 | HasWFxT = true; |
954 | if (Feature == "-fmv" ) |
955 | HasFMV = false; |
956 | if (Feature == "+crc" ) |
957 | HasCRC = true; |
958 | if (Feature == "+rcpc" ) |
959 | HasRCPC = true; |
960 | if (Feature == "+aes" ) { |
961 | FPU |= NeonMode; |
962 | HasAES = true; |
963 | } |
964 | if (Feature == "+sha2" ) { |
965 | FPU |= NeonMode; |
966 | HasSHA2 = true; |
967 | } |
968 | if (Feature == "+sha3" ) { |
969 | FPU |= NeonMode; |
970 | HasSHA2 = true; |
971 | HasSHA3 = true; |
972 | } |
973 | if (Feature == "+rdm" ) { |
974 | FPU |= NeonMode; |
975 | HasRDM = true; |
976 | } |
977 | if (Feature == "+dit" ) |
978 | HasDIT = true; |
979 | if (Feature == "+cccp" ) |
980 | HasCCPP = true; |
981 | if (Feature == "+ccdp" ) { |
982 | HasCCPP = true; |
983 | HasCCDP = true; |
984 | } |
985 | if (Feature == "+fptoint" ) |
986 | HasFRInt3264 = true; |
987 | if (Feature == "+sm4" ) { |
988 | FPU |= NeonMode; |
989 | HasSM4 = true; |
990 | } |
991 | if (Feature == "+strict-align" ) |
992 | HasUnalignedAccess = false; |
993 | |
994 | // All predecessor archs are added but select the latest one for ArchKind. |
995 | if (Feature == "+v8a" && ArchInfo->Version < llvm::AArch64::ARMV8A.Version) |
996 | ArchInfo = &llvm::AArch64::ARMV8A; |
997 | if (Feature == "+v8.1a" && |
998 | ArchInfo->Version < llvm::AArch64::ARMV8_1A.Version) |
999 | ArchInfo = &llvm::AArch64::ARMV8_1A; |
1000 | if (Feature == "+v8.2a" && |
1001 | ArchInfo->Version < llvm::AArch64::ARMV8_2A.Version) |
1002 | ArchInfo = &llvm::AArch64::ARMV8_2A; |
1003 | if (Feature == "+v8.3a" && |
1004 | ArchInfo->Version < llvm::AArch64::ARMV8_3A.Version) |
1005 | ArchInfo = &llvm::AArch64::ARMV8_3A; |
1006 | if (Feature == "+v8.4a" && |
1007 | ArchInfo->Version < llvm::AArch64::ARMV8_4A.Version) |
1008 | ArchInfo = &llvm::AArch64::ARMV8_4A; |
1009 | if (Feature == "+v8.5a" && |
1010 | ArchInfo->Version < llvm::AArch64::ARMV8_5A.Version) |
1011 | ArchInfo = &llvm::AArch64::ARMV8_5A; |
1012 | if (Feature == "+v8.6a" && |
1013 | ArchInfo->Version < llvm::AArch64::ARMV8_6A.Version) |
1014 | ArchInfo = &llvm::AArch64::ARMV8_6A; |
1015 | if (Feature == "+v8.7a" && |
1016 | ArchInfo->Version < llvm::AArch64::ARMV8_7A.Version) |
1017 | ArchInfo = &llvm::AArch64::ARMV8_7A; |
1018 | if (Feature == "+v8.8a" && |
1019 | ArchInfo->Version < llvm::AArch64::ARMV8_8A.Version) |
1020 | ArchInfo = &llvm::AArch64::ARMV8_8A; |
1021 | if (Feature == "+v8.9a" && |
1022 | ArchInfo->Version < llvm::AArch64::ARMV8_9A.Version) |
1023 | ArchInfo = &llvm::AArch64::ARMV8_9A; |
1024 | if (Feature == "+v9a" && ArchInfo->Version < llvm::AArch64::ARMV9A.Version) |
1025 | ArchInfo = &llvm::AArch64::ARMV9A; |
1026 | if (Feature == "+v9.1a" && |
1027 | ArchInfo->Version < llvm::AArch64::ARMV9_1A.Version) |
1028 | ArchInfo = &llvm::AArch64::ARMV9_1A; |
1029 | if (Feature == "+v9.2a" && |
1030 | ArchInfo->Version < llvm::AArch64::ARMV9_2A.Version) |
1031 | ArchInfo = &llvm::AArch64::ARMV9_2A; |
1032 | if (Feature == "+v9.3a" && |
1033 | ArchInfo->Version < llvm::AArch64::ARMV9_3A.Version) |
1034 | ArchInfo = &llvm::AArch64::ARMV9_3A; |
1035 | if (Feature == "+v9.4a" && |
1036 | ArchInfo->Version < llvm::AArch64::ARMV9_4A.Version) |
1037 | ArchInfo = &llvm::AArch64::ARMV9_4A; |
1038 | if (Feature == "+v9.5a" && |
1039 | ArchInfo->Version < llvm::AArch64::ARMV9_5A.Version) |
1040 | ArchInfo = &llvm::AArch64::ARMV9_5A; |
1041 | if (Feature == "+v8r" ) |
1042 | ArchInfo = &llvm::AArch64::ARMV8R; |
1043 | if (Feature == "+fullfp16" ) { |
1044 | FPU |= NeonMode; |
1045 | HasFullFP16 = true; |
1046 | } |
1047 | if (Feature == "+dotprod" ) { |
1048 | FPU |= NeonMode; |
1049 | HasDotProd = true; |
1050 | } |
1051 | if (Feature == "+fp16fml" ) { |
1052 | FPU |= NeonMode; |
1053 | HasFullFP16 = true; |
1054 | HasFP16FML = true; |
1055 | } |
1056 | if (Feature == "+mte" ) |
1057 | HasMTE = true; |
1058 | if (Feature == "+tme" ) |
1059 | HasTME = true; |
1060 | if (Feature == "+pauth" ) |
1061 | HasPAuth = true; |
1062 | if (Feature == "+i8mm" ) |
1063 | HasMatMul = true; |
1064 | if (Feature == "+bf16" ) |
1065 | HasBFloat16 = true; |
1066 | if (Feature == "+lse" ) |
1067 | HasLSE = true; |
1068 | if (Feature == "+ls64" ) |
1069 | HasLS64 = true; |
1070 | if (Feature == "+rand" ) |
1071 | HasRandGen = true; |
1072 | if (Feature == "+flagm" ) |
1073 | HasFlagM = true; |
1074 | if (Feature == "+altnzcv" ) { |
1075 | HasFlagM = true; |
1076 | HasAlternativeNZCV = true; |
1077 | } |
1078 | if (Feature == "+mops" ) |
1079 | HasMOPS = true; |
1080 | if (Feature == "+d128" ) |
1081 | HasD128 = true; |
1082 | if (Feature == "+gcs" ) |
1083 | HasGCS = true; |
1084 | if (Feature == "+rcpc3" ) |
1085 | HasRCPC3 = true; |
1086 | if (Feature == "+pauth-lr" ) { |
1087 | HasPAuthLR = true; |
1088 | HasPAuth = true; |
1089 | } |
1090 | } |
1091 | |
1092 | // Check features that are manually disabled by command line options. |
1093 | // This needs to be checked after architecture-related features are handled, |
1094 | // making sure they are properly disabled when required. |
1095 | for (const auto &Feature : Features) { |
1096 | if (Feature == "-d128" ) |
1097 | HasD128 = false; |
1098 | } |
1099 | |
1100 | setDataLayout(); |
1101 | setArchFeatures(); |
1102 | |
1103 | if (HasNoFP) { |
1104 | FPU &= ~FPUMode; |
1105 | FPU &= ~NeonMode; |
1106 | FPU &= ~SveMode; |
1107 | } |
1108 | if (HasNoNeon) { |
1109 | FPU &= ~NeonMode; |
1110 | FPU &= ~SveMode; |
1111 | } |
1112 | if (HasNoSVE) |
1113 | FPU &= ~SveMode; |
1114 | |
1115 | return true; |
1116 | } |
1117 | |
1118 | // Parse AArch64 Target attributes, which are a comma separated list of: |
1119 | // "arch=<arch>" - parsed to features as per -march=.. |
1120 | // "cpu=<cpu>" - parsed to features as per -mcpu=.., with CPU set to <cpu> |
1121 | // "tune=<cpu>" - TuneCPU set to <cpu> |
1122 | // "feature", "no-feature" - Add (or remove) feature. |
1123 | // "+feature", "+nofeature" - Add (or remove) feature. |
1124 | // |
1125 | // A feature may correspond to an Extension (anything with a corresponding |
1126 | // AEK_), in which case an ExtensionSet is used to parse it and expand its |
1127 | // dependencies. If the feature does not yield a successful parse then it |
1128 | // is passed through. |
1129 | ParsedTargetAttr AArch64TargetInfo::parseTargetAttr(StringRef Features) const { |
1130 | ParsedTargetAttr Ret; |
1131 | if (Features == "default" ) |
1132 | return Ret; |
1133 | SmallVector<StringRef, 1> AttrFeatures; |
1134 | Features.split(A&: AttrFeatures, Separator: "," ); |
1135 | bool FoundArch = false; |
1136 | |
1137 | auto SplitAndAddFeatures = [](StringRef FeatString, |
1138 | std::vector<std::string> &Features, |
1139 | llvm::AArch64::ExtensionSet &FeatureBits) { |
1140 | SmallVector<StringRef, 8> SplitFeatures; |
1141 | FeatString.split(A&: SplitFeatures, Separator: StringRef("+" ), MaxSplit: -1, KeepEmpty: false); |
1142 | for (StringRef Feature : SplitFeatures) { |
1143 | if (FeatureBits.parseModifier(Modifier: Feature)) |
1144 | continue; |
1145 | // Pass through anything that failed to parse so that we can emit |
1146 | // diagnostics, as well as valid internal feature names. |
1147 | // |
1148 | // FIXME: We should consider rejecting internal feature names like |
1149 | // neon, v8a, etc. |
1150 | // FIXME: We should consider emitting diagnostics here. |
1151 | if (Feature.starts_with(Prefix: "no" )) |
1152 | Features.push_back(x: "-" + Feature.drop_front(N: 2).str()); |
1153 | else |
1154 | Features.push_back(x: "+" + Feature.str()); |
1155 | } |
1156 | }; |
1157 | |
1158 | llvm::AArch64::ExtensionSet FeatureBits; |
1159 | // Reconstruct the bitset from the command line option features. |
1160 | FeatureBits.reconstructFromParsedFeatures(Features: getTargetOpts().FeaturesAsWritten, |
1161 | NonExtensions&: Ret.Features); |
1162 | |
1163 | for (auto &Feature : AttrFeatures) { |
1164 | Feature = Feature.trim(); |
1165 | if (Feature.starts_with(Prefix: "fpmath=" )) |
1166 | continue; |
1167 | |
1168 | if (Feature.starts_with(Prefix: "branch-protection=" )) { |
1169 | Ret.BranchProtection = Feature.split(Separator: '=').second.trim(); |
1170 | continue; |
1171 | } |
1172 | |
1173 | if (Feature.starts_with(Prefix: "arch=" )) { |
1174 | if (FoundArch) |
1175 | Ret.Duplicate = "arch=" ; |
1176 | FoundArch = true; |
1177 | std::pair<StringRef, StringRef> Split = |
1178 | Feature.split(Separator: "=" ).second.trim().split(Separator: "+" ); |
1179 | const llvm::AArch64::ArchInfo *AI = llvm::AArch64::parseArch(Arch: Split.first); |
1180 | |
1181 | // Parse the architecture version, adding the required features to |
1182 | // Ret.Features. |
1183 | if (!AI) |
1184 | continue; |
1185 | FeatureBits.addArchDefaults(Arch: *AI); |
1186 | // Add any extra features, after the + |
1187 | SplitAndAddFeatures(Split.second, Ret.Features, FeatureBits); |
1188 | } else if (Feature.starts_with(Prefix: "cpu=" )) { |
1189 | if (!Ret.CPU.empty()) |
1190 | Ret.Duplicate = "cpu=" ; |
1191 | else { |
1192 | // Split the cpu string into "cpu=", "cortex-a710" and any remaining |
1193 | // "+feat" features. |
1194 | std::pair<StringRef, StringRef> Split = |
1195 | Feature.split(Separator: "=" ).second.trim().split(Separator: "+" ); |
1196 | Ret.CPU = Split.first; |
1197 | if (auto CpuInfo = llvm::AArch64::parseCpu(Name: Ret.CPU)) { |
1198 | FeatureBits.addCPUDefaults(CPU: *CpuInfo); |
1199 | SplitAndAddFeatures(Split.second, Ret.Features, FeatureBits); |
1200 | } |
1201 | } |
1202 | } else if (Feature.starts_with(Prefix: "tune=" )) { |
1203 | if (!Ret.Tune.empty()) |
1204 | Ret.Duplicate = "tune=" ; |
1205 | else |
1206 | Ret.Tune = Feature.split(Separator: "=" ).second.trim(); |
1207 | } else if (Feature.starts_with(Prefix: "+" )) { |
1208 | SplitAndAddFeatures(Feature, Ret.Features, FeatureBits); |
1209 | } else { |
1210 | if (FeatureBits.parseModifier(Modifier: Feature, /* AllowNoDashForm = */ true)) |
1211 | continue; |
1212 | // Pass through anything that failed to parse so that we can emit |
1213 | // diagnostics, as well as valid internal feature names. |
1214 | // |
1215 | // FIXME: We should consider rejecting internal feature names like |
1216 | // neon, v8a, etc. |
1217 | // FIXME: We should consider emitting diagnostics here. |
1218 | if (Feature.starts_with(Prefix: "no-" )) |
1219 | Ret.Features.push_back(x: "-" + Feature.drop_front(N: 3).str()); |
1220 | else |
1221 | Ret.Features.push_back(x: "+" + Feature.str()); |
1222 | } |
1223 | } |
1224 | FeatureBits.toLLVMFeatureList(Features&: Ret.Features); |
1225 | return Ret; |
1226 | } |
1227 | |
1228 | bool AArch64TargetInfo::hasBFloat16Type() const { |
1229 | return true; |
1230 | } |
1231 | |
1232 | TargetInfo::CallingConvCheckResult |
1233 | AArch64TargetInfo::checkCallingConvention(CallingConv CC) const { |
1234 | switch (CC) { |
1235 | case CC_C: |
1236 | case CC_Swift: |
1237 | case CC_SwiftAsync: |
1238 | case CC_PreserveMost: |
1239 | case CC_PreserveAll: |
1240 | case CC_PreserveNone: |
1241 | case CC_OpenCLKernel: |
1242 | case CC_AArch64VectorCall: |
1243 | case CC_AArch64SVEPCS: |
1244 | case CC_Win64: |
1245 | return CCCR_OK; |
1246 | default: |
1247 | return CCCR_Warning; |
1248 | } |
1249 | } |
1250 | |
1251 | bool AArch64TargetInfo::isCLZForZeroUndef() const { return false; } |
1252 | |
1253 | TargetInfo::BuiltinVaListKind AArch64TargetInfo::getBuiltinVaListKind() const { |
1254 | return TargetInfo::AArch64ABIBuiltinVaList; |
1255 | } |
1256 | |
1257 | const char *const AArch64TargetInfo::GCCRegNames[] = { |
1258 | // clang-format off |
1259 | |
1260 | // 32-bit Integer registers |
1261 | "w0" , "w1" , "w2" , "w3" , "w4" , "w5" , "w6" , "w7" , "w8" , "w9" , "w10" , "w11" , |
1262 | "w12" , "w13" , "w14" , "w15" , "w16" , "w17" , "w18" , "w19" , "w20" , "w21" , "w22" , |
1263 | "w23" , "w24" , "w25" , "w26" , "w27" , "w28" , "w29" , "w30" , "wsp" , |
1264 | |
1265 | // 64-bit Integer registers |
1266 | "x0" , "x1" , "x2" , "x3" , "x4" , "x5" , "x6" , "x7" , "x8" , "x9" , "x10" , "x11" , |
1267 | "x12" , "x13" , "x14" , "x15" , "x16" , "x17" , "x18" , "x19" , "x20" , "x21" , "x22" , |
1268 | "x23" , "x24" , "x25" , "x26" , "x27" , "x28" , "fp" , "lr" , "sp" , |
1269 | |
1270 | // 32-bit floating point regsisters |
1271 | "s0" , "s1" , "s2" , "s3" , "s4" , "s5" , "s6" , "s7" , "s8" , "s9" , "s10" , "s11" , |
1272 | "s12" , "s13" , "s14" , "s15" , "s16" , "s17" , "s18" , "s19" , "s20" , "s21" , "s22" , |
1273 | "s23" , "s24" , "s25" , "s26" , "s27" , "s28" , "s29" , "s30" , "s31" , |
1274 | |
1275 | // 64-bit floating point regsisters |
1276 | "d0" , "d1" , "d2" , "d3" , "d4" , "d5" , "d6" , "d7" , "d8" , "d9" , "d10" , "d11" , |
1277 | "d12" , "d13" , "d14" , "d15" , "d16" , "d17" , "d18" , "d19" , "d20" , "d21" , "d22" , |
1278 | "d23" , "d24" , "d25" , "d26" , "d27" , "d28" , "d29" , "d30" , "d31" , |
1279 | |
1280 | // Neon vector registers |
1281 | "v0" , "v1" , "v2" , "v3" , "v4" , "v5" , "v6" , "v7" , "v8" , "v9" , "v10" , "v11" , |
1282 | "v12" , "v13" , "v14" , "v15" , "v16" , "v17" , "v18" , "v19" , "v20" , "v21" , "v22" , |
1283 | "v23" , "v24" , "v25" , "v26" , "v27" , "v28" , "v29" , "v30" , "v31" , |
1284 | |
1285 | // SVE vector registers |
1286 | "z0" , "z1" , "z2" , "z3" , "z4" , "z5" , "z6" , "z7" , "z8" , "z9" , "z10" , |
1287 | "z11" , "z12" , "z13" , "z14" , "z15" , "z16" , "z17" , "z18" , "z19" , "z20" , "z21" , |
1288 | "z22" , "z23" , "z24" , "z25" , "z26" , "z27" , "z28" , "z29" , "z30" , "z31" , |
1289 | |
1290 | // SVE predicate registers |
1291 | "p0" , "p1" , "p2" , "p3" , "p4" , "p5" , "p6" , "p7" , "p8" , "p9" , "p10" , |
1292 | "p11" , "p12" , "p13" , "p14" , "p15" , |
1293 | |
1294 | // SVE predicate-as-counter registers |
1295 | "pn0" , "pn1" , "pn2" , "pn3" , "pn4" , "pn5" , "pn6" , "pn7" , "pn8" , |
1296 | "pn9" , "pn10" , "pn11" , "pn12" , "pn13" , "pn14" , "pn15" , |
1297 | |
1298 | // SME registers |
1299 | "za" , "zt0" , |
1300 | |
1301 | // clang-format on |
1302 | }; |
1303 | |
1304 | ArrayRef<const char *> AArch64TargetInfo::getGCCRegNames() const { |
1305 | return llvm::ArrayRef(GCCRegNames); |
1306 | } |
1307 | |
1308 | const TargetInfo::GCCRegAlias AArch64TargetInfo::GCCRegAliases[] = { |
1309 | {.Aliases: {"w31" }, .Register: "wsp" }, |
1310 | {.Aliases: {"x31" }, .Register: "sp" }, |
1311 | // GCC rN registers are aliases of xN registers. |
1312 | {.Aliases: {"r0" }, .Register: "x0" }, |
1313 | {.Aliases: {"r1" }, .Register: "x1" }, |
1314 | {.Aliases: {"r2" }, .Register: "x2" }, |
1315 | {.Aliases: {"r3" }, .Register: "x3" }, |
1316 | {.Aliases: {"r4" }, .Register: "x4" }, |
1317 | {.Aliases: {"r5" }, .Register: "x5" }, |
1318 | {.Aliases: {"r6" }, .Register: "x6" }, |
1319 | {.Aliases: {"r7" }, .Register: "x7" }, |
1320 | {.Aliases: {"r8" }, .Register: "x8" }, |
1321 | {.Aliases: {"r9" }, .Register: "x9" }, |
1322 | {.Aliases: {"r10" }, .Register: "x10" }, |
1323 | {.Aliases: {"r11" }, .Register: "x11" }, |
1324 | {.Aliases: {"r12" }, .Register: "x12" }, |
1325 | {.Aliases: {"r13" }, .Register: "x13" }, |
1326 | {.Aliases: {"r14" }, .Register: "x14" }, |
1327 | {.Aliases: {"r15" }, .Register: "x15" }, |
1328 | {.Aliases: {"r16" }, .Register: "x16" }, |
1329 | {.Aliases: {"r17" }, .Register: "x17" }, |
1330 | {.Aliases: {"r18" }, .Register: "x18" }, |
1331 | {.Aliases: {"r19" }, .Register: "x19" }, |
1332 | {.Aliases: {"r20" }, .Register: "x20" }, |
1333 | {.Aliases: {"r21" }, .Register: "x21" }, |
1334 | {.Aliases: {"r22" }, .Register: "x22" }, |
1335 | {.Aliases: {"r23" }, .Register: "x23" }, |
1336 | {.Aliases: {"r24" }, .Register: "x24" }, |
1337 | {.Aliases: {"r25" }, .Register: "x25" }, |
1338 | {.Aliases: {"r26" }, .Register: "x26" }, |
1339 | {.Aliases: {"r27" }, .Register: "x27" }, |
1340 | {.Aliases: {"r28" }, .Register: "x28" }, |
1341 | {.Aliases: {"r29" , "x29" }, .Register: "fp" }, |
1342 | {.Aliases: {"r30" , "x30" }, .Register: "lr" }, |
1343 | // The S/D/Q and W/X registers overlap, but aren't really aliases; we |
1344 | // don't want to substitute one of these for a different-sized one. |
1345 | }; |
1346 | |
1347 | ArrayRef<TargetInfo::GCCRegAlias> AArch64TargetInfo::getGCCRegAliases() const { |
1348 | return llvm::ArrayRef(GCCRegAliases); |
1349 | } |
1350 | |
1351 | // Returns the length of cc constraint. |
1352 | static unsigned matchAsmCCConstraint(const char *Name) { |
1353 | constexpr unsigned len = 5; |
1354 | auto RV = llvm::StringSwitch<unsigned>(Name) |
1355 | .Case(S: "@cceq" , Value: len) |
1356 | .Case(S: "@ccne" , Value: len) |
1357 | .Case(S: "@cchs" , Value: len) |
1358 | .Case(S: "@cccs" , Value: len) |
1359 | .Case(S: "@cccc" , Value: len) |
1360 | .Case(S: "@cclo" , Value: len) |
1361 | .Case(S: "@ccmi" , Value: len) |
1362 | .Case(S: "@ccpl" , Value: len) |
1363 | .Case(S: "@ccvs" , Value: len) |
1364 | .Case(S: "@ccvc" , Value: len) |
1365 | .Case(S: "@cchi" , Value: len) |
1366 | .Case(S: "@ccls" , Value: len) |
1367 | .Case(S: "@ccge" , Value: len) |
1368 | .Case(S: "@cclt" , Value: len) |
1369 | .Case(S: "@ccgt" , Value: len) |
1370 | .Case(S: "@ccle" , Value: len) |
1371 | .Default(Value: 0); |
1372 | return RV; |
1373 | } |
1374 | |
1375 | std::string |
1376 | AArch64TargetInfo::convertConstraint(const char *&Constraint) const { |
1377 | std::string R; |
1378 | switch (*Constraint) { |
1379 | case 'U': // Three-character constraint; add "@3" hint for later parsing. |
1380 | R = std::string("@3" ) + std::string(Constraint, 3); |
1381 | Constraint += 2; |
1382 | break; |
1383 | case '@': |
1384 | if (const unsigned Len = matchAsmCCConstraint(Name: Constraint)) { |
1385 | std::string Converted = "{" + std::string(Constraint, Len) + "}" ; |
1386 | Constraint += Len - 1; |
1387 | return Converted; |
1388 | } |
1389 | return std::string(1, *Constraint); |
1390 | default: |
1391 | R = TargetInfo::convertConstraint(Constraint); |
1392 | break; |
1393 | } |
1394 | return R; |
1395 | } |
1396 | |
1397 | bool AArch64TargetInfo::validateAsmConstraint( |
1398 | const char *&Name, TargetInfo::ConstraintInfo &Info) const { |
1399 | switch (*Name) { |
1400 | default: |
1401 | return false; |
1402 | case 'w': // Floating point and SIMD registers (V0-V31) |
1403 | Info.setAllowsRegister(); |
1404 | return true; |
1405 | case 'I': // Constant that can be used with an ADD instruction |
1406 | case 'J': // Constant that can be used with a SUB instruction |
1407 | case 'K': // Constant that can be used with a 32-bit logical instruction |
1408 | case 'L': // Constant that can be used with a 64-bit logical instruction |
1409 | case 'M': // Constant that can be used as a 32-bit MOV immediate |
1410 | case 'N': // Constant that can be used as a 64-bit MOV immediate |
1411 | case 'Y': // Floating point constant zero |
1412 | case 'Z': // Integer constant zero |
1413 | return true; |
1414 | case 'Q': // A memory reference with base register and no offset |
1415 | Info.setAllowsMemory(); |
1416 | return true; |
1417 | case 'S': // A symbolic address |
1418 | Info.setAllowsRegister(); |
1419 | return true; |
1420 | case 'U': |
1421 | if (Name[1] == 'p' && |
1422 | (Name[2] == 'l' || Name[2] == 'a' || Name[2] == 'h')) { |
1423 | // SVE predicate registers ("Upa"=P0-15, "Upl"=P0-P7, "Uph"=P8-P15) |
1424 | Info.setAllowsRegister(); |
1425 | Name += 2; |
1426 | return true; |
1427 | } |
1428 | if (Name[1] == 'c' && (Name[2] == 'i' || Name[2] == 'j')) { |
1429 | // Gpr registers ("Uci"=w8-11, "Ucj"=w12-15) |
1430 | Info.setAllowsRegister(); |
1431 | Name += 2; |
1432 | return true; |
1433 | } |
1434 | // Ump: A memory address suitable for ldp/stp in SI, DI, SF and DF modes. |
1435 | // Utf: A memory address suitable for ldp/stp in TF mode. |
1436 | // Usa: An absolute symbolic address. |
1437 | // Ush: The high part (bits 32:12) of a pc-relative symbolic address. |
1438 | |
1439 | // Better to return an error saying that it's an unrecognised constraint |
1440 | // even if this is a valid constraint in gcc. |
1441 | return false; |
1442 | case 'z': // Zero register, wzr or xzr |
1443 | Info.setAllowsRegister(); |
1444 | return true; |
1445 | case 'x': // Floating point and SIMD registers (V0-V15) |
1446 | Info.setAllowsRegister(); |
1447 | return true; |
1448 | case 'y': // SVE registers (V0-V7) |
1449 | Info.setAllowsRegister(); |
1450 | return true; |
1451 | case '@': |
1452 | // CC condition |
1453 | if (const unsigned Len = matchAsmCCConstraint(Name)) { |
1454 | Name += Len - 1; |
1455 | Info.setAllowsRegister(); |
1456 | return true; |
1457 | } |
1458 | } |
1459 | return false; |
1460 | } |
1461 | |
1462 | bool AArch64TargetInfo::validateConstraintModifier( |
1463 | StringRef Constraint, char Modifier, unsigned Size, |
1464 | std::string &SuggestedModifier) const { |
1465 | // Strip off constraint modifiers. |
1466 | Constraint = Constraint.ltrim(Chars: "=+&" ); |
1467 | |
1468 | switch (Constraint[0]) { |
1469 | default: |
1470 | return true; |
1471 | case 'z': |
1472 | case 'r': { |
1473 | switch (Modifier) { |
1474 | case 'x': |
1475 | case 'w': |
1476 | // For now assume that the person knows what they're |
1477 | // doing with the modifier. |
1478 | return true; |
1479 | default: |
1480 | // By default an 'r' constraint will be in the 'x' |
1481 | // registers. |
1482 | if (Size == 64) |
1483 | return true; |
1484 | |
1485 | if (Size == 512) |
1486 | return HasLS64; |
1487 | |
1488 | SuggestedModifier = "w" ; |
1489 | return false; |
1490 | } |
1491 | } |
1492 | } |
1493 | } |
1494 | |
1495 | std::string_view AArch64TargetInfo::getClobbers() const { return "" ; } |
1496 | |
1497 | int AArch64TargetInfo::getEHDataRegisterNumber(unsigned RegNo) const { |
1498 | if (RegNo == 0) |
1499 | return 0; |
1500 | if (RegNo == 1) |
1501 | return 1; |
1502 | return -1; |
1503 | } |
1504 | |
1505 | bool AArch64TargetInfo::validatePointerAuthKey( |
1506 | const llvm::APSInt &value) const { |
1507 | return 0 <= value && value <= 3; |
1508 | } |
1509 | |
1510 | bool AArch64TargetInfo::hasInt128Type() const { return true; } |
1511 | |
1512 | AArch64leTargetInfo::AArch64leTargetInfo(const llvm::Triple &Triple, |
1513 | const TargetOptions &Opts) |
1514 | : AArch64TargetInfo(Triple, Opts) {} |
1515 | |
1516 | void AArch64leTargetInfo::setDataLayout() { |
1517 | if (getTriple().isOSBinFormatMachO()) { |
1518 | if(getTriple().isArch32Bit()) |
1519 | resetDataLayout(DL: "e-m:o-p:32:32-i64:64-i128:128-n32:64-S128-Fn32" , UserLabelPrefix: "_" ); |
1520 | else |
1521 | resetDataLayout(DL: "e-m:o-i64:64-i128:128-n32:64-S128-Fn32" , UserLabelPrefix: "_" ); |
1522 | } else |
1523 | resetDataLayout(DL: "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128-Fn32" ); |
1524 | } |
1525 | |
1526 | void AArch64leTargetInfo::getTargetDefines(const LangOptions &Opts, |
1527 | MacroBuilder &Builder) const { |
1528 | Builder.defineMacro(Name: "__AARCH64EL__" ); |
1529 | AArch64TargetInfo::getTargetDefines(Opts, Builder); |
1530 | } |
1531 | |
1532 | AArch64beTargetInfo::AArch64beTargetInfo(const llvm::Triple &Triple, |
1533 | const TargetOptions &Opts) |
1534 | : AArch64TargetInfo(Triple, Opts) {} |
1535 | |
1536 | void AArch64beTargetInfo::getTargetDefines(const LangOptions &Opts, |
1537 | MacroBuilder &Builder) const { |
1538 | Builder.defineMacro(Name: "__AARCH64EB__" ); |
1539 | Builder.defineMacro(Name: "__AARCH_BIG_ENDIAN" ); |
1540 | Builder.defineMacro(Name: "__ARM_BIG_ENDIAN" ); |
1541 | AArch64TargetInfo::getTargetDefines(Opts, Builder); |
1542 | } |
1543 | |
1544 | void AArch64beTargetInfo::setDataLayout() { |
1545 | assert(!getTriple().isOSBinFormatMachO()); |
1546 | resetDataLayout(DL: "E-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128-Fn32" ); |
1547 | } |
1548 | |
1549 | WindowsARM64TargetInfo::WindowsARM64TargetInfo(const llvm::Triple &Triple, |
1550 | const TargetOptions &Opts) |
1551 | : WindowsTargetInfo<AArch64leTargetInfo>(Triple, Opts), Triple(Triple) { |
1552 | |
1553 | // This is an LLP64 platform. |
1554 | // int:4, long:4, long long:8, long double:8. |
1555 | IntWidth = IntAlign = 32; |
1556 | LongWidth = LongAlign = 32; |
1557 | DoubleAlign = LongLongAlign = 64; |
1558 | LongDoubleWidth = LongDoubleAlign = 64; |
1559 | LongDoubleFormat = &llvm::APFloat::IEEEdouble(); |
1560 | IntMaxType = SignedLongLong; |
1561 | Int64Type = SignedLongLong; |
1562 | SizeType = UnsignedLongLong; |
1563 | PtrDiffType = SignedLongLong; |
1564 | IntPtrType = SignedLongLong; |
1565 | } |
1566 | |
1567 | void WindowsARM64TargetInfo::setDataLayout() { |
1568 | resetDataLayout(DL: Triple.isOSBinFormatMachO() |
1569 | ? "e-m:o-i64:64-i128:128-n32:64-S128-Fn32" |
1570 | : "e-m:w-p:64:64-i32:32-i64:64-i128:128-n32:64-S128-Fn32" , |
1571 | UserLabelPrefix: Triple.isOSBinFormatMachO() ? "_" : "" ); |
1572 | } |
1573 | |
1574 | TargetInfo::BuiltinVaListKind |
1575 | WindowsARM64TargetInfo::getBuiltinVaListKind() const { |
1576 | return TargetInfo::CharPtrBuiltinVaList; |
1577 | } |
1578 | |
1579 | TargetInfo::CallingConvCheckResult |
1580 | WindowsARM64TargetInfo::checkCallingConvention(CallingConv CC) const { |
1581 | switch (CC) { |
1582 | case CC_X86VectorCall: |
1583 | if (getTriple().isWindowsArm64EC()) |
1584 | return CCCR_OK; |
1585 | return CCCR_Ignore; |
1586 | case CC_X86StdCall: |
1587 | case CC_X86ThisCall: |
1588 | case CC_X86FastCall: |
1589 | return CCCR_Ignore; |
1590 | case CC_C: |
1591 | case CC_OpenCLKernel: |
1592 | case CC_PreserveMost: |
1593 | case CC_PreserveAll: |
1594 | case CC_PreserveNone: |
1595 | case CC_Swift: |
1596 | case CC_SwiftAsync: |
1597 | case CC_Win64: |
1598 | return CCCR_OK; |
1599 | default: |
1600 | return CCCR_Warning; |
1601 | } |
1602 | } |
1603 | |
1604 | MicrosoftARM64TargetInfo::MicrosoftARM64TargetInfo(const llvm::Triple &Triple, |
1605 | const TargetOptions &Opts) |
1606 | : WindowsARM64TargetInfo(Triple, Opts) { |
1607 | TheCXXABI.set(TargetCXXABI::Microsoft); |
1608 | } |
1609 | |
1610 | void MicrosoftARM64TargetInfo::getTargetDefines(const LangOptions &Opts, |
1611 | MacroBuilder &Builder) const { |
1612 | WindowsARM64TargetInfo::getTargetDefines(Opts, Builder); |
1613 | if (getTriple().isWindowsArm64EC()) { |
1614 | Builder.defineMacro(Name: "_M_X64" , Value: "100" ); |
1615 | Builder.defineMacro(Name: "_M_AMD64" , Value: "100" ); |
1616 | Builder.defineMacro(Name: "_M_ARM64EC" , Value: "1" ); |
1617 | } else { |
1618 | Builder.defineMacro(Name: "_M_ARM64" , Value: "1" ); |
1619 | } |
1620 | } |
1621 | |
1622 | TargetInfo::CallingConvKind |
1623 | MicrosoftARM64TargetInfo::getCallingConvKind(bool ClangABICompat4) const { |
1624 | return CCK_MicrosoftWin64; |
1625 | } |
1626 | |
1627 | unsigned MicrosoftARM64TargetInfo::getMinGlobalAlign(uint64_t TypeSize, |
1628 | bool HasNonWeakDef) const { |
1629 | unsigned Align = |
1630 | WindowsARM64TargetInfo::getMinGlobalAlign(Size: TypeSize, HasNonWeakDef); |
1631 | |
1632 | // MSVC does size based alignment for arm64 based on alignment section in |
1633 | // below document, replicate that to keep alignment consistent with object |
1634 | // files compiled by MSVC. |
1635 | // https://docs.microsoft.com/en-us/cpp/build/arm64-windows-abi-conventions |
1636 | if (TypeSize >= 512) { // TypeSize >= 64 bytes |
1637 | Align = std::max(a: Align, b: 128u); // align type at least 16 bytes |
1638 | } else if (TypeSize >= 64) { // TypeSize >= 8 bytes |
1639 | Align = std::max(a: Align, b: 64u); // align type at least 8 butes |
1640 | } else if (TypeSize >= 16) { // TypeSize >= 2 bytes |
1641 | Align = std::max(a: Align, b: 32u); // align type at least 4 bytes |
1642 | } |
1643 | return Align; |
1644 | } |
1645 | |
1646 | MinGWARM64TargetInfo::MinGWARM64TargetInfo(const llvm::Triple &Triple, |
1647 | const TargetOptions &Opts) |
1648 | : WindowsARM64TargetInfo(Triple, Opts) { |
1649 | TheCXXABI.set(TargetCXXABI::GenericAArch64); |
1650 | } |
1651 | |
1652 | DarwinAArch64TargetInfo::DarwinAArch64TargetInfo(const llvm::Triple &Triple, |
1653 | const TargetOptions &Opts) |
1654 | : DarwinTargetInfo<AArch64leTargetInfo>(Triple, Opts) { |
1655 | Int64Type = SignedLongLong; |
1656 | if (getTriple().isArch32Bit()) |
1657 | IntMaxType = SignedLongLong; |
1658 | |
1659 | WCharType = SignedInt; |
1660 | UseSignedCharForObjCBool = false; |
1661 | |
1662 | LongDoubleWidth = LongDoubleAlign = SuitableAlign = 64; |
1663 | LongDoubleFormat = &llvm::APFloat::IEEEdouble(); |
1664 | |
1665 | UseZeroLengthBitfieldAlignment = false; |
1666 | |
1667 | if (getTriple().isArch32Bit()) { |
1668 | UseBitFieldTypeAlignment = false; |
1669 | ZeroLengthBitfieldBoundary = 32; |
1670 | UseZeroLengthBitfieldAlignment = true; |
1671 | TheCXXABI.set(TargetCXXABI::WatchOS); |
1672 | } else |
1673 | TheCXXABI.set(TargetCXXABI::AppleARM64); |
1674 | } |
1675 | |
1676 | void DarwinAArch64TargetInfo::getOSDefines(const LangOptions &Opts, |
1677 | const llvm::Triple &Triple, |
1678 | MacroBuilder &Builder) const { |
1679 | Builder.defineMacro(Name: "__AARCH64_SIMD__" ); |
1680 | if (Triple.isArch32Bit()) |
1681 | Builder.defineMacro(Name: "__ARM64_ARCH_8_32__" ); |
1682 | else |
1683 | Builder.defineMacro(Name: "__ARM64_ARCH_8__" ); |
1684 | Builder.defineMacro(Name: "__ARM_NEON__" ); |
1685 | Builder.defineMacro(Name: "__REGISTER_PREFIX__" , Value: "" ); |
1686 | Builder.defineMacro(Name: "__arm64" , Value: "1" ); |
1687 | Builder.defineMacro(Name: "__arm64__" , Value: "1" ); |
1688 | |
1689 | if (Triple.isArm64e()) |
1690 | Builder.defineMacro(Name: "__arm64e__" , Value: "1" ); |
1691 | |
1692 | getDarwinDefines(Builder, Opts, Triple, PlatformName, PlatformMinVersion); |
1693 | } |
1694 | |
1695 | TargetInfo::BuiltinVaListKind |
1696 | DarwinAArch64TargetInfo::getBuiltinVaListKind() const { |
1697 | return TargetInfo::CharPtrBuiltinVaList; |
1698 | } |
1699 | |
1700 | // 64-bit RenderScript is aarch64 |
1701 | RenderScript64TargetInfo::RenderScript64TargetInfo(const llvm::Triple &Triple, |
1702 | const TargetOptions &Opts) |
1703 | : AArch64leTargetInfo(llvm::Triple("aarch64" , Triple.getVendorName(), |
1704 | Triple.getOSName(), |
1705 | Triple.getEnvironmentName()), |
1706 | Opts) { |
1707 | IsRenderScriptTarget = true; |
1708 | } |
1709 | |
1710 | void RenderScript64TargetInfo::getTargetDefines(const LangOptions &Opts, |
1711 | MacroBuilder &Builder) const { |
1712 | Builder.defineMacro(Name: "__RENDERSCRIPT__" ); |
1713 | AArch64leTargetInfo::getTargetDefines(Opts, Builder); |
1714 | } |
1715 | |