1 //===-- AArch64ISelLowering.cpp - AArch64 DAG Lowering Implementation ----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the AArch64TargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "AArch64ISelLowering.h"
15 #include "AArch64CallingConvention.h"
16 #include "AArch64MachineFunctionInfo.h"
17 #include "AArch64PerfectShuffle.h"
18 #include "AArch64Subtarget.h"
19 #include "AArch64TargetMachine.h"
20 #include "AArch64TargetObjectFile.h"
21 #include "MCTargetDesc/AArch64AddressingModes.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/Intrinsics.h"
29 #include "llvm/IR/Type.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetOptions.h"
37 #define DEBUG_TYPE "aarch64-lower"
39 STATISTIC(NumTailCalls, "Number of tail calls");
40 STATISTIC(NumShiftInserts, "Number of vector shift inserts");
49 static cl::opt<AlignMode>
50 Align(cl::desc("Load/store alignment support"),
51 cl::Hidden, cl::init(NoStrictAlign),
53 clEnumValN(StrictAlign, "aarch64-strict-align",
54 "Disallow all unaligned memory accesses"),
55 clEnumValN(NoStrictAlign, "aarch64-no-strict-align",
56 "Allow unaligned memory accesses"),
59 // Place holder until extr generation is tested fully.
61 EnableAArch64ExtrGeneration("aarch64-extr-generation", cl::Hidden,
62 cl::desc("Allow AArch64 (or (shift)(shift))->extract"),
66 EnableAArch64SlrGeneration("aarch64-shift-insert-generation", cl::Hidden,
67 cl::desc("Allow AArch64 SLI/SRI formation"),
71 AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM)
72 : TargetLowering(TM) {
73 Subtarget = &TM.getSubtarget<AArch64Subtarget>();
75 // AArch64 doesn't have comparisons which set GPRs or setcc instructions, so
76 // we have to make something up. Arbitrarily, choose ZeroOrOne.
77 setBooleanContents(ZeroOrOneBooleanContent);
78 // When comparing vectors the result sets the different elements in the
79 // vector to all-one or all-zero.
80 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
82 // Set up the register classes.
83 addRegisterClass(MVT::i32, &AArch64::GPR32allRegClass);
84 addRegisterClass(MVT::i64, &AArch64::GPR64allRegClass);
86 if (Subtarget->hasFPARMv8()) {
87 addRegisterClass(MVT::f16, &AArch64::FPR16RegClass);
88 addRegisterClass(MVT::f32, &AArch64::FPR32RegClass);
89 addRegisterClass(MVT::f64, &AArch64::FPR64RegClass);
90 addRegisterClass(MVT::f128, &AArch64::FPR128RegClass);
93 if (Subtarget->hasNEON()) {
94 addRegisterClass(MVT::v16i8, &AArch64::FPR8RegClass);
95 addRegisterClass(MVT::v8i16, &AArch64::FPR16RegClass);
96 // Someone set us up the NEON.
97 addDRTypeForNEON(MVT::v2f32);
98 addDRTypeForNEON(MVT::v8i8);
99 addDRTypeForNEON(MVT::v4i16);
100 addDRTypeForNEON(MVT::v2i32);
101 addDRTypeForNEON(MVT::v1i64);
102 addDRTypeForNEON(MVT::v1f64);
103 addDRTypeForNEON(MVT::v4f16);
105 addQRTypeForNEON(MVT::v4f32);
106 addQRTypeForNEON(MVT::v2f64);
107 addQRTypeForNEON(MVT::v16i8);
108 addQRTypeForNEON(MVT::v8i16);
109 addQRTypeForNEON(MVT::v4i32);
110 addQRTypeForNEON(MVT::v2i64);
111 addQRTypeForNEON(MVT::v8f16);
114 // Compute derived properties from the register classes
115 computeRegisterProperties();
117 // Provide all sorts of operation actions
118 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
119 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
120 setOperationAction(ISD::SETCC, MVT::i32, Custom);
121 setOperationAction(ISD::SETCC, MVT::i64, Custom);
122 setOperationAction(ISD::SETCC, MVT::f32, Custom);
123 setOperationAction(ISD::SETCC, MVT::f64, Custom);
124 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
125 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
126 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
127 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
128 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
129 setOperationAction(ISD::SELECT, MVT::i32, Custom);
130 setOperationAction(ISD::SELECT, MVT::i64, Custom);
131 setOperationAction(ISD::SELECT, MVT::f32, Custom);
132 setOperationAction(ISD::SELECT, MVT::f64, Custom);
133 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
134 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
135 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
136 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
137 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
138 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
140 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
141 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
142 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
144 setOperationAction(ISD::FREM, MVT::f32, Expand);
145 setOperationAction(ISD::FREM, MVT::f64, Expand);
146 setOperationAction(ISD::FREM, MVT::f80, Expand);
148 // Custom lowering hooks are needed for XOR
149 // to fold it into CSINC/CSINV.
150 setOperationAction(ISD::XOR, MVT::i32, Custom);
151 setOperationAction(ISD::XOR, MVT::i64, Custom);
153 // Virtually no operation on f128 is legal, but LLVM can't expand them when
154 // there's a valid register class, so we need custom operations in most cases.
155 setOperationAction(ISD::FABS, MVT::f128, Expand);
156 setOperationAction(ISD::FADD, MVT::f128, Custom);
157 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
158 setOperationAction(ISD::FCOS, MVT::f128, Expand);
159 setOperationAction(ISD::FDIV, MVT::f128, Custom);
160 setOperationAction(ISD::FMA, MVT::f128, Expand);
161 setOperationAction(ISD::FMUL, MVT::f128, Custom);
162 setOperationAction(ISD::FNEG, MVT::f128, Expand);
163 setOperationAction(ISD::FPOW, MVT::f128, Expand);
164 setOperationAction(ISD::FREM, MVT::f128, Expand);
165 setOperationAction(ISD::FRINT, MVT::f128, Expand);
166 setOperationAction(ISD::FSIN, MVT::f128, Expand);
167 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
168 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
169 setOperationAction(ISD::FSUB, MVT::f128, Custom);
170 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
171 setOperationAction(ISD::SETCC, MVT::f128, Custom);
172 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
173 setOperationAction(ISD::SELECT, MVT::f128, Custom);
174 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
175 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
177 // Lowering for many of the conversions is actually specified by the non-f128
178 // type. The LowerXXX function will be trivial when f128 isn't involved.
179 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
180 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
181 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
182 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
183 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
184 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
185 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
186 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
187 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
188 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
189 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
190 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
191 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
192 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
194 // Variable arguments.
195 setOperationAction(ISD::VASTART, MVT::Other, Custom);
196 setOperationAction(ISD::VAARG, MVT::Other, Custom);
197 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
198 setOperationAction(ISD::VAEND, MVT::Other, Expand);
200 // Variable-sized objects.
201 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
202 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
203 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
205 // Exception handling.
206 // FIXME: These are guesses. Has this been defined yet?
207 setExceptionPointerRegister(AArch64::X0);
208 setExceptionSelectorRegister(AArch64::X1);
210 // Constant pool entries
211 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
214 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
216 // Add/Sub overflow ops with MVT::Glues are lowered to NZCV dependences.
217 setOperationAction(ISD::ADDC, MVT::i32, Custom);
218 setOperationAction(ISD::ADDE, MVT::i32, Custom);
219 setOperationAction(ISD::SUBC, MVT::i32, Custom);
220 setOperationAction(ISD::SUBE, MVT::i32, Custom);
221 setOperationAction(ISD::ADDC, MVT::i64, Custom);
222 setOperationAction(ISD::ADDE, MVT::i64, Custom);
223 setOperationAction(ISD::SUBC, MVT::i64, Custom);
224 setOperationAction(ISD::SUBE, MVT::i64, Custom);
226 // AArch64 lacks both left-rotate and popcount instructions.
227 setOperationAction(ISD::ROTL, MVT::i32, Expand);
228 setOperationAction(ISD::ROTL, MVT::i64, Expand);
230 // AArch64 doesn't have {U|S}MUL_LOHI.
231 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
232 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
235 // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
236 // counterparts, which AArch64 supports directly.
237 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
238 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
239 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
240 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
242 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
243 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
245 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
246 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
247 setOperationAction(ISD::SREM, MVT::i32, Expand);
248 setOperationAction(ISD::SREM, MVT::i64, Expand);
249 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
250 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
251 setOperationAction(ISD::UREM, MVT::i32, Expand);
252 setOperationAction(ISD::UREM, MVT::i64, Expand);
254 // Custom lower Add/Sub/Mul with overflow.
255 setOperationAction(ISD::SADDO, MVT::i32, Custom);
256 setOperationAction(ISD::SADDO, MVT::i64, Custom);
257 setOperationAction(ISD::UADDO, MVT::i32, Custom);
258 setOperationAction(ISD::UADDO, MVT::i64, Custom);
259 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
260 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
261 setOperationAction(ISD::USUBO, MVT::i32, Custom);
262 setOperationAction(ISD::USUBO, MVT::i64, Custom);
263 setOperationAction(ISD::SMULO, MVT::i32, Custom);
264 setOperationAction(ISD::SMULO, MVT::i64, Custom);
265 setOperationAction(ISD::UMULO, MVT::i32, Custom);
266 setOperationAction(ISD::UMULO, MVT::i64, Custom);
268 setOperationAction(ISD::FSIN, MVT::f32, Expand);
269 setOperationAction(ISD::FSIN, MVT::f64, Expand);
270 setOperationAction(ISD::FCOS, MVT::f32, Expand);
271 setOperationAction(ISD::FCOS, MVT::f64, Expand);
272 setOperationAction(ISD::FPOW, MVT::f32, Expand);
273 setOperationAction(ISD::FPOW, MVT::f64, Expand);
274 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
275 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
277 // f16 is storage-only, so we promote operations to f32 if we know this is
278 // valid, and ignore them otherwise. The operations not mentioned here will
279 // fail to select, but this is not a major problem as no source language
280 // should be emitting native f16 operations yet.
281 setOperationAction(ISD::FADD, MVT::f16, Promote);
282 setOperationAction(ISD::FDIV, MVT::f16, Promote);
283 setOperationAction(ISD::FMUL, MVT::f16, Promote);
284 setOperationAction(ISD::FSUB, MVT::f16, Promote);
286 // v4f16 is also a storage-only type, so promote it to v4f32 when that is
288 setOperationAction(ISD::FADD, MVT::v4f16, Promote);
289 setOperationAction(ISD::FSUB, MVT::v4f16, Promote);
290 setOperationAction(ISD::FMUL, MVT::v4f16, Promote);
291 setOperationAction(ISD::FDIV, MVT::v4f16, Promote);
292 setOperationAction(ISD::FP_EXTEND, MVT::v4f16, Promote);
293 setOperationAction(ISD::FP_ROUND, MVT::v4f16, Promote);
294 AddPromotedToType(ISD::FADD, MVT::v4f16, MVT::v4f32);
295 AddPromotedToType(ISD::FSUB, MVT::v4f16, MVT::v4f32);
296 AddPromotedToType(ISD::FMUL, MVT::v4f16, MVT::v4f32);
297 AddPromotedToType(ISD::FDIV, MVT::v4f16, MVT::v4f32);
298 AddPromotedToType(ISD::FP_EXTEND, MVT::v4f16, MVT::v4f32);
299 AddPromotedToType(ISD::FP_ROUND, MVT::v4f16, MVT::v4f32);
301 // Expand all other v4f16 operations.
302 // FIXME: We could generate better code by promoting some operations to
304 setOperationAction(ISD::FABS, MVT::v4f16, Expand);
305 setOperationAction(ISD::FCEIL, MVT::v4f16, Expand);
306 setOperationAction(ISD::FCOPYSIGN, MVT::v4f16, Expand);
307 setOperationAction(ISD::FCOS, MVT::v4f16, Expand);
308 setOperationAction(ISD::FFLOOR, MVT::v4f16, Expand);
309 setOperationAction(ISD::FMA, MVT::v4f16, Expand);
310 setOperationAction(ISD::FNEARBYINT, MVT::v4f16, Expand);
311 setOperationAction(ISD::FNEG, MVT::v4f16, Expand);
312 setOperationAction(ISD::FPOW, MVT::v4f16, Expand);
313 setOperationAction(ISD::FPOWI, MVT::v4f16, Expand);
314 setOperationAction(ISD::FREM, MVT::v4f16, Expand);
315 setOperationAction(ISD::FROUND, MVT::v4f16, Expand);
316 setOperationAction(ISD::FRINT, MVT::v4f16, Expand);
317 setOperationAction(ISD::FSIN, MVT::v4f16, Expand);
318 setOperationAction(ISD::FSINCOS, MVT::v4f16, Expand);
319 setOperationAction(ISD::FSQRT, MVT::v4f16, Expand);
320 setOperationAction(ISD::FTRUNC, MVT::v4f16, Expand);
321 setOperationAction(ISD::SETCC, MVT::v4f16, Expand);
322 setOperationAction(ISD::BR_CC, MVT::v4f16, Expand);
323 setOperationAction(ISD::SELECT, MVT::v4f16, Expand);
324 setOperationAction(ISD::SELECT_CC, MVT::v4f16, Expand);
325 setOperationAction(ISD::FEXP, MVT::v4f16, Expand);
326 setOperationAction(ISD::FEXP2, MVT::v4f16, Expand);
327 setOperationAction(ISD::FLOG, MVT::v4f16, Expand);
328 setOperationAction(ISD::FLOG2, MVT::v4f16, Expand);
329 setOperationAction(ISD::FLOG10, MVT::v4f16, Expand);
332 // v8f16 is also a storage-only type, so expand it.
333 setOperationAction(ISD::FABS, MVT::v8f16, Expand);
334 setOperationAction(ISD::FADD, MVT::v8f16, Expand);
335 setOperationAction(ISD::FCEIL, MVT::v8f16, Expand);
336 setOperationAction(ISD::FCOPYSIGN, MVT::v8f16, Expand);
337 setOperationAction(ISD::FCOS, MVT::v8f16, Expand);
338 setOperationAction(ISD::FDIV, MVT::v8f16, Expand);
339 setOperationAction(ISD::FFLOOR, MVT::v8f16, Expand);
340 setOperationAction(ISD::FMA, MVT::v8f16, Expand);
341 setOperationAction(ISD::FMUL, MVT::v8f16, Expand);
342 setOperationAction(ISD::FNEARBYINT, MVT::v8f16, Expand);
343 setOperationAction(ISD::FNEG, MVT::v8f16, Expand);
344 setOperationAction(ISD::FPOW, MVT::v8f16, Expand);
345 setOperationAction(ISD::FPOWI, MVT::v8f16, Expand);
346 setOperationAction(ISD::FREM, MVT::v8f16, Expand);
347 setOperationAction(ISD::FROUND, MVT::v8f16, Expand);
348 setOperationAction(ISD::FRINT, MVT::v8f16, Expand);
349 setOperationAction(ISD::FSIN, MVT::v8f16, Expand);
350 setOperationAction(ISD::FSINCOS, MVT::v8f16, Expand);
351 setOperationAction(ISD::FSQRT, MVT::v8f16, Expand);
352 setOperationAction(ISD::FSUB, MVT::v8f16, Expand);
353 setOperationAction(ISD::FTRUNC, MVT::v8f16, Expand);
354 setOperationAction(ISD::SETCC, MVT::v8f16, Expand);
355 setOperationAction(ISD::BR_CC, MVT::v8f16, Expand);
356 setOperationAction(ISD::SELECT, MVT::v8f16, Expand);
357 setOperationAction(ISD::SELECT_CC, MVT::v8f16, Expand);
358 setOperationAction(ISD::FP_EXTEND, MVT::v8f16, Expand);
359 setOperationAction(ISD::FEXP, MVT::v8f16, Expand);
360 setOperationAction(ISD::FEXP2, MVT::v8f16, Expand);
361 setOperationAction(ISD::FLOG, MVT::v8f16, Expand);
362 setOperationAction(ISD::FLOG2, MVT::v8f16, Expand);
363 setOperationAction(ISD::FLOG10, MVT::v8f16, Expand);
365 // AArch64 has implementations of a lot of rounding-like FP operations.
366 static MVT RoundingTypes[] = { MVT::f32, MVT::f64};
367 for (unsigned I = 0; I < array_lengthof(RoundingTypes); ++I) {
368 MVT Ty = RoundingTypes[I];
369 setOperationAction(ISD::FFLOOR, Ty, Legal);
370 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
371 setOperationAction(ISD::FCEIL, Ty, Legal);
372 setOperationAction(ISD::FRINT, Ty, Legal);
373 setOperationAction(ISD::FTRUNC, Ty, Legal);
374 setOperationAction(ISD::FROUND, Ty, Legal);
377 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
379 if (Subtarget->isTargetMachO()) {
380 // For iOS, we don't want to the normal expansion of a libcall to
381 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
383 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
384 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
386 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
387 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
390 // Make floating-point constants legal for the large code model, so they don't
391 // become loads from the constant pool.
392 if (Subtarget->isTargetMachO() && TM.getCodeModel() == CodeModel::Large) {
393 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
394 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
397 // AArch64 does not have floating-point extending loads, i1 sign-extending
398 // load, floating-point truncating stores, or v2i32->v2i16 truncating store.
399 setLoadExtAction(ISD::EXTLOAD, MVT::f16, Expand);
400 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
401 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
402 setLoadExtAction(ISD::EXTLOAD, MVT::f80, Expand);
403 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Expand);
404 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
405 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
406 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
407 setTruncStoreAction(MVT::f128, MVT::f80, Expand);
408 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
409 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
410 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
412 setOperationAction(ISD::BITCAST, MVT::i16, Custom);
413 setOperationAction(ISD::BITCAST, MVT::f16, Custom);
415 // Indexed loads and stores are supported.
416 for (unsigned im = (unsigned)ISD::PRE_INC;
417 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
418 setIndexedLoadAction(im, MVT::i8, Legal);
419 setIndexedLoadAction(im, MVT::i16, Legal);
420 setIndexedLoadAction(im, MVT::i32, Legal);
421 setIndexedLoadAction(im, MVT::i64, Legal);
422 setIndexedLoadAction(im, MVT::f64, Legal);
423 setIndexedLoadAction(im, MVT::f32, Legal);
424 setIndexedStoreAction(im, MVT::i8, Legal);
425 setIndexedStoreAction(im, MVT::i16, Legal);
426 setIndexedStoreAction(im, MVT::i32, Legal);
427 setIndexedStoreAction(im, MVT::i64, Legal);
428 setIndexedStoreAction(im, MVT::f64, Legal);
429 setIndexedStoreAction(im, MVT::f32, Legal);
433 setOperationAction(ISD::TRAP, MVT::Other, Legal);
435 // We combine OR nodes for bitfield operations.
436 setTargetDAGCombine(ISD::OR);
438 // Vector add and sub nodes may conceal a high-half opportunity.
439 // Also, try to fold ADD into CSINC/CSINV..
440 setTargetDAGCombine(ISD::ADD);
441 setTargetDAGCombine(ISD::SUB);
443 setTargetDAGCombine(ISD::XOR);
444 setTargetDAGCombine(ISD::SINT_TO_FP);
445 setTargetDAGCombine(ISD::UINT_TO_FP);
447 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
449 setTargetDAGCombine(ISD::ANY_EXTEND);
450 setTargetDAGCombine(ISD::ZERO_EXTEND);
451 setTargetDAGCombine(ISD::SIGN_EXTEND);
452 setTargetDAGCombine(ISD::BITCAST);
453 setTargetDAGCombine(ISD::CONCAT_VECTORS);
454 setTargetDAGCombine(ISD::STORE);
456 setTargetDAGCombine(ISD::MUL);
458 setTargetDAGCombine(ISD::SELECT);
459 setTargetDAGCombine(ISD::VSELECT);
461 setTargetDAGCombine(ISD::INTRINSIC_VOID);
462 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
463 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
465 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
466 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
467 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
469 setStackPointerRegisterToSaveRestore(AArch64::SP);
471 setSchedulingPreference(Sched::Hybrid);
474 MaskAndBranchFoldingIsLegal = true;
476 setMinFunctionAlignment(2);
478 RequireStrictAlign = (Align == StrictAlign);
480 setHasExtractBitsInsn(true);
482 if (Subtarget->hasNEON()) {
483 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
484 // silliness like this:
485 setOperationAction(ISD::FABS, MVT::v1f64, Expand);
486 setOperationAction(ISD::FADD, MVT::v1f64, Expand);
487 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
488 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
489 setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
490 setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
491 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
492 setOperationAction(ISD::FMA, MVT::v1f64, Expand);
493 setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
494 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
495 setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
496 setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
497 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
498 setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
499 setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
500 setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
501 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
502 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
503 setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
504 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
505 setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
506 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
507 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
508 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
509 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);
511 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
512 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
513 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
514 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
515 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);
517 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
519 // AArch64 doesn't have a direct vector ->f32 conversion instructions for
520 // elements smaller than i32, so promote the input to i32 first.
521 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
522 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
523 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
524 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
525 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
526 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
527 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
528 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
529 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
531 // AArch64 doesn't have MUL.2d:
532 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
533 // Custom handling for some quad-vector types to detect MULL.
534 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
535 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
536 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
538 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
539 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
540 // Likewise, narrowing and extending vector loads/stores aren't handled
542 for (MVT VT : MVT::vector_valuetypes()) {
543 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
545 setOperationAction(ISD::MULHS, VT, Expand);
546 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
547 setOperationAction(ISD::MULHU, VT, Expand);
548 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
550 setOperationAction(ISD::BSWAP, VT, Expand);
552 for (MVT InnerVT : MVT::vector_valuetypes())
553 setTruncStoreAction(VT, InnerVT, Expand);
554 setLoadExtAction(ISD::SEXTLOAD, VT, Expand);
555 setLoadExtAction(ISD::ZEXTLOAD, VT, Expand);
556 setLoadExtAction(ISD::EXTLOAD, VT, Expand);
559 // AArch64 has implementations of a lot of rounding-like FP operations.
560 static MVT RoundingVecTypes[] = {MVT::v2f32, MVT::v4f32, MVT::v2f64 };
561 for (unsigned I = 0; I < array_lengthof(RoundingVecTypes); ++I) {
562 MVT Ty = RoundingVecTypes[I];
563 setOperationAction(ISD::FFLOOR, Ty, Legal);
564 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
565 setOperationAction(ISD::FCEIL, Ty, Legal);
566 setOperationAction(ISD::FRINT, Ty, Legal);
567 setOperationAction(ISD::FTRUNC, Ty, Legal);
568 setOperationAction(ISD::FROUND, Ty, Legal);
572 // Prefer likely predicted branches to selects on out-of-order cores.
573 if (Subtarget->isCortexA57())
574 PredictableSelectIsExpensive = true;
577 void AArch64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
578 if (VT == MVT::v2f32 || VT == MVT::v4f16) {
579 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
580 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
582 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
583 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
584 } else if (VT == MVT::v2f64 || VT == MVT::v4f32 || VT == MVT::v8f16) {
585 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
586 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
588 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
589 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
592 // Mark vector float intrinsics as expand.
593 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
594 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
595 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
596 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
597 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
598 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
599 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
600 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
601 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
602 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
605 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
606 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
607 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
608 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
609 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
610 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
611 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
612 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
613 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
614 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
615 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
616 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
618 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
619 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
620 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
621 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
623 // CNT supports only B element sizes.
624 if (VT != MVT::v8i8 && VT != MVT::v16i8)
625 setOperationAction(ISD::CTPOP, VT.getSimpleVT(), Expand);
627 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
628 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
629 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
630 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
631 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
633 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
634 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
636 if (Subtarget->isLittleEndian()) {
637 for (unsigned im = (unsigned)ISD::PRE_INC;
638 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
639 setIndexedLoadAction(im, VT.getSimpleVT(), Legal);
640 setIndexedStoreAction(im, VT.getSimpleVT(), Legal);
645 void AArch64TargetLowering::addDRTypeForNEON(MVT VT) {
646 addRegisterClass(VT, &AArch64::FPR64RegClass);
647 addTypeForNEON(VT, MVT::v2i32);
650 void AArch64TargetLowering::addQRTypeForNEON(MVT VT) {
651 addRegisterClass(VT, &AArch64::FPR128RegClass);
652 addTypeForNEON(VT, MVT::v4i32);
655 EVT AArch64TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
658 return VT.changeVectorElementTypeToInteger();
661 /// computeKnownBitsForTargetNode - Determine which of the bits specified in
662 /// Mask are known to be either zero or one and return them in the
663 /// KnownZero/KnownOne bitsets.
664 void AArch64TargetLowering::computeKnownBitsForTargetNode(
665 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
666 const SelectionDAG &DAG, unsigned Depth) const {
667 switch (Op.getOpcode()) {
670 case AArch64ISD::CSEL: {
671 APInt KnownZero2, KnownOne2;
672 DAG.computeKnownBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
673 DAG.computeKnownBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
674 KnownZero &= KnownZero2;
675 KnownOne &= KnownOne2;
678 case ISD::INTRINSIC_W_CHAIN: {
679 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
680 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
683 case Intrinsic::aarch64_ldaxr:
684 case Intrinsic::aarch64_ldxr: {
685 unsigned BitWidth = KnownOne.getBitWidth();
686 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
687 unsigned MemBits = VT.getScalarType().getSizeInBits();
688 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
694 case ISD::INTRINSIC_WO_CHAIN:
695 case ISD::INTRINSIC_VOID: {
696 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
700 case Intrinsic::aarch64_neon_umaxv:
701 case Intrinsic::aarch64_neon_uminv: {
702 // Figure out the datatype of the vector operand. The UMINV instruction
703 // will zero extend the result, so we can mark as known zero all the
704 // bits larger than the element datatype. 32-bit or larget doesn't need
705 // this as those are legal types and will be handled by isel directly.
706 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
707 unsigned BitWidth = KnownZero.getBitWidth();
708 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
709 assert(BitWidth >= 8 && "Unexpected width!");
710 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
712 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
713 assert(BitWidth >= 16 && "Unexpected width!");
714 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
724 MVT AArch64TargetLowering::getScalarShiftAmountTy(EVT LHSTy) const {
728 unsigned AArch64TargetLowering::getMaximalGlobalOffset() const {
729 // FIXME: On AArch64, this depends on the type.
730 // Basically, the addressable offsets are up to 4095 * Ty.getSizeInBytes().
731 // and the offset has to be a multiple of the related size in bytes.
736 AArch64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
737 const TargetLibraryInfo *libInfo) const {
738 return AArch64::createFastISel(funcInfo, libInfo);
741 const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
745 case AArch64ISD::CALL: return "AArch64ISD::CALL";
746 case AArch64ISD::ADRP: return "AArch64ISD::ADRP";
747 case AArch64ISD::ADDlow: return "AArch64ISD::ADDlow";
748 case AArch64ISD::LOADgot: return "AArch64ISD::LOADgot";
749 case AArch64ISD::RET_FLAG: return "AArch64ISD::RET_FLAG";
750 case AArch64ISD::BRCOND: return "AArch64ISD::BRCOND";
751 case AArch64ISD::CSEL: return "AArch64ISD::CSEL";
752 case AArch64ISD::FCSEL: return "AArch64ISD::FCSEL";
753 case AArch64ISD::CSINV: return "AArch64ISD::CSINV";
754 case AArch64ISD::CSNEG: return "AArch64ISD::CSNEG";
755 case AArch64ISD::CSINC: return "AArch64ISD::CSINC";
756 case AArch64ISD::THREAD_POINTER: return "AArch64ISD::THREAD_POINTER";
757 case AArch64ISD::TLSDESC_CALL: return "AArch64ISD::TLSDESC_CALL";
758 case AArch64ISD::ADC: return "AArch64ISD::ADC";
759 case AArch64ISD::SBC: return "AArch64ISD::SBC";
760 case AArch64ISD::ADDS: return "AArch64ISD::ADDS";
761 case AArch64ISD::SUBS: return "AArch64ISD::SUBS";
762 case AArch64ISD::ADCS: return "AArch64ISD::ADCS";
763 case AArch64ISD::SBCS: return "AArch64ISD::SBCS";
764 case AArch64ISD::ANDS: return "AArch64ISD::ANDS";
765 case AArch64ISD::FCMP: return "AArch64ISD::FCMP";
766 case AArch64ISD::FMIN: return "AArch64ISD::FMIN";
767 case AArch64ISD::FMAX: return "AArch64ISD::FMAX";
768 case AArch64ISD::DUP: return "AArch64ISD::DUP";
769 case AArch64ISD::DUPLANE8: return "AArch64ISD::DUPLANE8";
770 case AArch64ISD::DUPLANE16: return "AArch64ISD::DUPLANE16";
771 case AArch64ISD::DUPLANE32: return "AArch64ISD::DUPLANE32";
772 case AArch64ISD::DUPLANE64: return "AArch64ISD::DUPLANE64";
773 case AArch64ISD::MOVI: return "AArch64ISD::MOVI";
774 case AArch64ISD::MOVIshift: return "AArch64ISD::MOVIshift";
775 case AArch64ISD::MOVIedit: return "AArch64ISD::MOVIedit";
776 case AArch64ISD::MOVImsl: return "AArch64ISD::MOVImsl";
777 case AArch64ISD::FMOV: return "AArch64ISD::FMOV";
778 case AArch64ISD::MVNIshift: return "AArch64ISD::MVNIshift";
779 case AArch64ISD::MVNImsl: return "AArch64ISD::MVNImsl";
780 case AArch64ISD::BICi: return "AArch64ISD::BICi";
781 case AArch64ISD::ORRi: return "AArch64ISD::ORRi";
782 case AArch64ISD::BSL: return "AArch64ISD::BSL";
783 case AArch64ISD::NEG: return "AArch64ISD::NEG";
784 case AArch64ISD::EXTR: return "AArch64ISD::EXTR";
785 case AArch64ISD::ZIP1: return "AArch64ISD::ZIP1";
786 case AArch64ISD::ZIP2: return "AArch64ISD::ZIP2";
787 case AArch64ISD::UZP1: return "AArch64ISD::UZP1";
788 case AArch64ISD::UZP2: return "AArch64ISD::UZP2";
789 case AArch64ISD::TRN1: return "AArch64ISD::TRN1";
790 case AArch64ISD::TRN2: return "AArch64ISD::TRN2";
791 case AArch64ISD::REV16: return "AArch64ISD::REV16";
792 case AArch64ISD::REV32: return "AArch64ISD::REV32";
793 case AArch64ISD::REV64: return "AArch64ISD::REV64";
794 case AArch64ISD::EXT: return "AArch64ISD::EXT";
795 case AArch64ISD::VSHL: return "AArch64ISD::VSHL";
796 case AArch64ISD::VLSHR: return "AArch64ISD::VLSHR";
797 case AArch64ISD::VASHR: return "AArch64ISD::VASHR";
798 case AArch64ISD::CMEQ: return "AArch64ISD::CMEQ";
799 case AArch64ISD::CMGE: return "AArch64ISD::CMGE";
800 case AArch64ISD::CMGT: return "AArch64ISD::CMGT";
801 case AArch64ISD::CMHI: return "AArch64ISD::CMHI";
802 case AArch64ISD::CMHS: return "AArch64ISD::CMHS";
803 case AArch64ISD::FCMEQ: return "AArch64ISD::FCMEQ";
804 case AArch64ISD::FCMGE: return "AArch64ISD::FCMGE";
805 case AArch64ISD::FCMGT: return "AArch64ISD::FCMGT";
806 case AArch64ISD::CMEQz: return "AArch64ISD::CMEQz";
807 case AArch64ISD::CMGEz: return "AArch64ISD::CMGEz";
808 case AArch64ISD::CMGTz: return "AArch64ISD::CMGTz";
809 case AArch64ISD::CMLEz: return "AArch64ISD::CMLEz";
810 case AArch64ISD::CMLTz: return "AArch64ISD::CMLTz";
811 case AArch64ISD::FCMEQz: return "AArch64ISD::FCMEQz";
812 case AArch64ISD::FCMGEz: return "AArch64ISD::FCMGEz";
813 case AArch64ISD::FCMGTz: return "AArch64ISD::FCMGTz";
814 case AArch64ISD::FCMLEz: return "AArch64ISD::FCMLEz";
815 case AArch64ISD::FCMLTz: return "AArch64ISD::FCMLTz";
816 case AArch64ISD::NOT: return "AArch64ISD::NOT";
817 case AArch64ISD::BIT: return "AArch64ISD::BIT";
818 case AArch64ISD::CBZ: return "AArch64ISD::CBZ";
819 case AArch64ISD::CBNZ: return "AArch64ISD::CBNZ";
820 case AArch64ISD::TBZ: return "AArch64ISD::TBZ";
821 case AArch64ISD::TBNZ: return "AArch64ISD::TBNZ";
822 case AArch64ISD::TC_RETURN: return "AArch64ISD::TC_RETURN";
823 case AArch64ISD::SITOF: return "AArch64ISD::SITOF";
824 case AArch64ISD::UITOF: return "AArch64ISD::UITOF";
825 case AArch64ISD::NVCAST: return "AArch64ISD::NVCAST";
826 case AArch64ISD::SQSHL_I: return "AArch64ISD::SQSHL_I";
827 case AArch64ISD::UQSHL_I: return "AArch64ISD::UQSHL_I";
828 case AArch64ISD::SRSHR_I: return "AArch64ISD::SRSHR_I";
829 case AArch64ISD::URSHR_I: return "AArch64ISD::URSHR_I";
830 case AArch64ISD::SQSHLU_I: return "AArch64ISD::SQSHLU_I";
831 case AArch64ISD::WrapperLarge: return "AArch64ISD::WrapperLarge";
832 case AArch64ISD::LD2post: return "AArch64ISD::LD2post";
833 case AArch64ISD::LD3post: return "AArch64ISD::LD3post";
834 case AArch64ISD::LD4post: return "AArch64ISD::LD4post";
835 case AArch64ISD::ST2post: return "AArch64ISD::ST2post";
836 case AArch64ISD::ST3post: return "AArch64ISD::ST3post";
837 case AArch64ISD::ST4post: return "AArch64ISD::ST4post";
838 case AArch64ISD::LD1x2post: return "AArch64ISD::LD1x2post";
839 case AArch64ISD::LD1x3post: return "AArch64ISD::LD1x3post";
840 case AArch64ISD::LD1x4post: return "AArch64ISD::LD1x4post";
841 case AArch64ISD::ST1x2post: return "AArch64ISD::ST1x2post";
842 case AArch64ISD::ST1x3post: return "AArch64ISD::ST1x3post";
843 case AArch64ISD::ST1x4post: return "AArch64ISD::ST1x4post";
844 case AArch64ISD::LD1DUPpost: return "AArch64ISD::LD1DUPpost";
845 case AArch64ISD::LD2DUPpost: return "AArch64ISD::LD2DUPpost";
846 case AArch64ISD::LD3DUPpost: return "AArch64ISD::LD3DUPpost";
847 case AArch64ISD::LD4DUPpost: return "AArch64ISD::LD4DUPpost";
848 case AArch64ISD::LD1LANEpost: return "AArch64ISD::LD1LANEpost";
849 case AArch64ISD::LD2LANEpost: return "AArch64ISD::LD2LANEpost";
850 case AArch64ISD::LD3LANEpost: return "AArch64ISD::LD3LANEpost";
851 case AArch64ISD::LD4LANEpost: return "AArch64ISD::LD4LANEpost";
852 case AArch64ISD::ST2LANEpost: return "AArch64ISD::ST2LANEpost";
853 case AArch64ISD::ST3LANEpost: return "AArch64ISD::ST3LANEpost";
854 case AArch64ISD::ST4LANEpost: return "AArch64ISD::ST4LANEpost";
855 case AArch64ISD::SMULL: return "AArch64ISD::SMULL";
856 case AArch64ISD::UMULL: return "AArch64ISD::UMULL";
861 AArch64TargetLowering::EmitF128CSEL(MachineInstr *MI,
862 MachineBasicBlock *MBB) const {
863 // We materialise the F128CSEL pseudo-instruction as some control flow and a
867 // [... previous instrs leading to comparison ...]
873 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
875 const TargetInstrInfo *TII =
876 getTargetMachine().getSubtargetImpl()->getInstrInfo();
877 MachineFunction *MF = MBB->getParent();
878 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
879 DebugLoc DL = MI->getDebugLoc();
880 MachineFunction::iterator It = MBB;
883 unsigned DestReg = MI->getOperand(0).getReg();
884 unsigned IfTrueReg = MI->getOperand(1).getReg();
885 unsigned IfFalseReg = MI->getOperand(2).getReg();
886 unsigned CondCode = MI->getOperand(3).getImm();
887 bool NZCVKilled = MI->getOperand(4).isKill();
889 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
890 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
891 MF->insert(It, TrueBB);
892 MF->insert(It, EndBB);
894 // Transfer rest of current basic-block to EndBB
895 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
897 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
899 BuildMI(MBB, DL, TII->get(AArch64::Bcc)).addImm(CondCode).addMBB(TrueBB);
900 BuildMI(MBB, DL, TII->get(AArch64::B)).addMBB(EndBB);
901 MBB->addSuccessor(TrueBB);
902 MBB->addSuccessor(EndBB);
904 // TrueBB falls through to the end.
905 TrueBB->addSuccessor(EndBB);
908 TrueBB->addLiveIn(AArch64::NZCV);
909 EndBB->addLiveIn(AArch64::NZCV);
912 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(AArch64::PHI), DestReg)
918 MI->eraseFromParent();
923 AArch64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
924 MachineBasicBlock *BB) const {
925 switch (MI->getOpcode()) {
930 llvm_unreachable("Unexpected instruction for custom inserter!");
932 case AArch64::F128CSEL:
933 return EmitF128CSEL(MI, BB);
935 case TargetOpcode::STACKMAP:
936 case TargetOpcode::PATCHPOINT:
937 return emitPatchPoint(MI, BB);
941 //===----------------------------------------------------------------------===//
942 // AArch64 Lowering private implementation.
943 //===----------------------------------------------------------------------===//
945 //===----------------------------------------------------------------------===//
947 //===----------------------------------------------------------------------===//
949 /// changeIntCCToAArch64CC - Convert a DAG integer condition code to an AArch64
951 static AArch64CC::CondCode changeIntCCToAArch64CC(ISD::CondCode CC) {
954 llvm_unreachable("Unknown condition code!");
956 return AArch64CC::NE;
958 return AArch64CC::EQ;
960 return AArch64CC::GT;
962 return AArch64CC::GE;
964 return AArch64CC::LT;
966 return AArch64CC::LE;
968 return AArch64CC::HI;
970 return AArch64CC::HS;
972 return AArch64CC::LO;
974 return AArch64CC::LS;
978 /// changeFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64 CC.
979 static void changeFPCCToAArch64CC(ISD::CondCode CC,
980 AArch64CC::CondCode &CondCode,
981 AArch64CC::CondCode &CondCode2) {
982 CondCode2 = AArch64CC::AL;
985 llvm_unreachable("Unknown FP condition!");
988 CondCode = AArch64CC::EQ;
992 CondCode = AArch64CC::GT;
996 CondCode = AArch64CC::GE;
999 CondCode = AArch64CC::MI;
1002 CondCode = AArch64CC::LS;
1005 CondCode = AArch64CC::MI;
1006 CondCode2 = AArch64CC::GT;
1009 CondCode = AArch64CC::VC;
1012 CondCode = AArch64CC::VS;
1015 CondCode = AArch64CC::EQ;
1016 CondCode2 = AArch64CC::VS;
1019 CondCode = AArch64CC::HI;
1022 CondCode = AArch64CC::PL;
1026 CondCode = AArch64CC::LT;
1030 CondCode = AArch64CC::LE;
1034 CondCode = AArch64CC::NE;
1039 /// changeVectorFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64
1040 /// CC usable with the vector instructions. Fewer operations are available
1041 /// without a real NZCV register, so we have to use less efficient combinations
1042 /// to get the same effect.
1043 static void changeVectorFPCCToAArch64CC(ISD::CondCode CC,
1044 AArch64CC::CondCode &CondCode,
1045 AArch64CC::CondCode &CondCode2,
1050 // Mostly the scalar mappings work fine.
1051 changeFPCCToAArch64CC(CC, CondCode, CondCode2);
1054 Invert = true; // Fallthrough
1056 CondCode = AArch64CC::MI;
1057 CondCode2 = AArch64CC::GE;
1064 // All of the compare-mask comparisons are ordered, but we can switch
1065 // between the two by a double inversion. E.g. ULE == !OGT.
1067 changeFPCCToAArch64CC(getSetCCInverse(CC, false), CondCode, CondCode2);
1072 static bool isLegalArithImmed(uint64_t C) {
1073 // Matches AArch64DAGToDAGISel::SelectArithImmed().
1074 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
1077 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1078 SDLoc dl, SelectionDAG &DAG) {
1079 EVT VT = LHS.getValueType();
1081 if (VT.isFloatingPoint())
1082 return DAG.getNode(AArch64ISD::FCMP, dl, VT, LHS, RHS);
1084 // The CMP instruction is just an alias for SUBS, and representing it as
1085 // SUBS means that it's possible to get CSE with subtract operations.
1086 // A later phase can perform the optimization of setting the destination
1087 // register to WZR/XZR if it ends up being unused.
1088 unsigned Opcode = AArch64ISD::SUBS;
1090 if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
1091 cast<ConstantSDNode>(RHS.getOperand(0))->getZExtValue() == 0 &&
1092 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1093 // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on
1094 // the grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags
1095 // can be set differently by this operation. It comes down to whether
1096 // "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then
1097 // everything is fine. If not then the optimization is wrong. Thus general
1098 // comparisons are only valid if op2 != 0.
1100 // So, finally, the only LLVM-native comparisons that don't mention C and V
1101 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in
1102 // the absence of information about op2.
1103 Opcode = AArch64ISD::ADDS;
1104 RHS = RHS.getOperand(1);
1105 } else if (LHS.getOpcode() == ISD::AND && isa<ConstantSDNode>(RHS) &&
1106 cast<ConstantSDNode>(RHS)->getZExtValue() == 0 &&
1107 !isUnsignedIntSetCC(CC)) {
1108 // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
1109 // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
1110 // of the signed comparisons.
1111 Opcode = AArch64ISD::ANDS;
1112 RHS = LHS.getOperand(1);
1113 LHS = LHS.getOperand(0);
1116 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS)
1120 static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1121 SDValue &AArch64cc, SelectionDAG &DAG, SDLoc dl) {
1123 AArch64CC::CondCode AArch64CC;
1124 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1125 EVT VT = RHS.getValueType();
1126 uint64_t C = RHSC->getZExtValue();
1127 if (!isLegalArithImmed(C)) {
1128 // Constant does not fit, try adjusting it by one?
1134 if ((VT == MVT::i32 && C != 0x80000000 &&
1135 isLegalArithImmed((uint32_t)(C - 1))) ||
1136 (VT == MVT::i64 && C != 0x80000000ULL &&
1137 isLegalArithImmed(C - 1ULL))) {
1138 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1139 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1140 RHS = DAG.getConstant(C, VT);
1145 if ((VT == MVT::i32 && C != 0 &&
1146 isLegalArithImmed((uint32_t)(C - 1))) ||
1147 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1148 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1149 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1150 RHS = DAG.getConstant(C, VT);
1155 if ((VT == MVT::i32 && C != INT32_MAX &&
1156 isLegalArithImmed((uint32_t)(C + 1))) ||
1157 (VT == MVT::i64 && C != INT64_MAX &&
1158 isLegalArithImmed(C + 1ULL))) {
1159 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1160 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1161 RHS = DAG.getConstant(C, VT);
1166 if ((VT == MVT::i32 && C != UINT32_MAX &&
1167 isLegalArithImmed((uint32_t)(C + 1))) ||
1168 (VT == MVT::i64 && C != UINT64_MAX &&
1169 isLegalArithImmed(C + 1ULL))) {
1170 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1171 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1172 RHS = DAG.getConstant(C, VT);
1178 // The imm operand of ADDS is an unsigned immediate, in the range 0 to 4095.
1179 // For the i8 operand, the largest immediate is 255, so this can be easily
1180 // encoded in the compare instruction. For the i16 operand, however, the
1181 // largest immediate cannot be encoded in the compare.
1182 // Therefore, use a sign extending load and cmn to avoid materializing the -1
1183 // constant. For example,
1185 // ldrh w0, [x0, #0]
1188 // ldrsh w0, [x0, #0]
1190 // Fundamental, we're relying on the property that (zext LHS) == (zext RHS)
1191 // if and only if (sext LHS) == (sext RHS). The checks are in place to ensure
1192 // both the LHS and RHS are truely zero extended and to make sure the
1193 // transformation is profitable.
1194 if ((CC == ISD::SETEQ || CC == ISD::SETNE) && isa<ConstantSDNode>(RHS)) {
1195 if ((cast<ConstantSDNode>(RHS)->getZExtValue() >> 16 == 0) &&
1196 isa<LoadSDNode>(LHS)) {
1197 if (cast<LoadSDNode>(LHS)->getExtensionType() == ISD::ZEXTLOAD &&
1198 cast<LoadSDNode>(LHS)->getMemoryVT() == MVT::i16 &&
1199 LHS.getNode()->hasNUsesOfValue(1, 0)) {
1200 int16_t ValueofRHS = cast<ConstantSDNode>(RHS)->getZExtValue();
1201 if (ValueofRHS < 0 && isLegalArithImmed(-ValueofRHS)) {
1203 DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, LHS.getValueType(), LHS,
1204 DAG.getValueType(MVT::i16));
1205 Cmp = emitComparison(SExt,
1206 DAG.getConstant(ValueofRHS, RHS.getValueType()),
1208 AArch64CC = changeIntCCToAArch64CC(CC);
1209 AArch64cc = DAG.getConstant(AArch64CC, MVT::i32);
1215 Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
1216 AArch64CC = changeIntCCToAArch64CC(CC);
1217 AArch64cc = DAG.getConstant(AArch64CC, MVT::i32);
1221 static std::pair<SDValue, SDValue>
1222 getAArch64XALUOOp(AArch64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1223 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1224 "Unsupported value type");
1225 SDValue Value, Overflow;
1227 SDValue LHS = Op.getOperand(0);
1228 SDValue RHS = Op.getOperand(1);
1230 switch (Op.getOpcode()) {
1232 llvm_unreachable("Unknown overflow instruction!");
1234 Opc = AArch64ISD::ADDS;
1238 Opc = AArch64ISD::ADDS;
1242 Opc = AArch64ISD::SUBS;
1246 Opc = AArch64ISD::SUBS;
1249 // Multiply needs a little bit extra work.
1253 bool IsSigned = (Op.getOpcode() == ISD::SMULO) ? true : false;
1254 if (Op.getValueType() == MVT::i32) {
1255 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1256 // For a 32 bit multiply with overflow check we want the instruction
1257 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1258 // need to generate the following pattern:
1259 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1260 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1261 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1262 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1263 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1264 DAG.getConstant(0, MVT::i64));
1265 // On AArch64 the upper 32 bits are always zero extended for a 32 bit
1266 // operation. We need to clear out the upper 32 bits, because we used a
1267 // widening multiply that wrote all 64 bits. In the end this should be a
1269 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1271 // The signed overflow check requires more than just a simple check for
1272 // any bit set in the upper 32 bits of the result. These bits could be
1273 // just the sign bits of a negative number. To perform the overflow
1274 // check we have to arithmetic shift right the 32nd bit of the result by
1275 // 31 bits. Then we compare the result to the upper 32 bits.
1276 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1277 DAG.getConstant(32, MVT::i64));
1278 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1279 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1280 DAG.getConstant(31, MVT::i64));
1281 // It is important that LowerBits is last, otherwise the arithmetic
1282 // shift will not be folded into the compare (SUBS).
1283 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1284 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1287 // The overflow check for unsigned multiply is easy. We only need to
1288 // check if any of the upper 32 bits are set. This can be done with a
1289 // CMP (shifted register). For that we need to generate the following
1291 // (i64 AArch64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1292 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1293 DAG.getConstant(32, MVT::i64));
1294 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1296 DAG.getNode(AArch64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1297 UpperBits).getValue(1);
1301 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1302 // For the 64 bit multiply
1303 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1305 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1306 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1307 DAG.getConstant(63, MVT::i64));
1308 // It is important that LowerBits is last, otherwise the arithmetic
1309 // shift will not be folded into the compare (SUBS).
1310 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1311 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1314 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1315 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1317 DAG.getNode(AArch64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1318 UpperBits).getValue(1);
1325 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1327 // Emit the AArch64 operation with overflow check.
1328 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1329 Overflow = Value.getValue(1);
1331 return std::make_pair(Value, Overflow);
1334 SDValue AArch64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1335 RTLIB::Libcall Call) const {
1336 SmallVector<SDValue, 2> Ops;
1337 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1338 Ops.push_back(Op.getOperand(i));
1340 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1344 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1345 SDValue Sel = Op.getOperand(0);
1346 SDValue Other = Op.getOperand(1);
1348 // If neither operand is a SELECT_CC, give up.
1349 if (Sel.getOpcode() != ISD::SELECT_CC)
1350 std::swap(Sel, Other);
1351 if (Sel.getOpcode() != ISD::SELECT_CC)
1354 // The folding we want to perform is:
1355 // (xor x, (select_cc a, b, cc, 0, -1) )
1357 // (csel x, (xor x, -1), cc ...)
1359 // The latter will get matched to a CSINV instruction.
1361 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1362 SDValue LHS = Sel.getOperand(0);
1363 SDValue RHS = Sel.getOperand(1);
1364 SDValue TVal = Sel.getOperand(2);
1365 SDValue FVal = Sel.getOperand(3);
1368 // FIXME: This could be generalized to non-integer comparisons.
1369 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1372 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1373 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1375 // The the values aren't constants, this isn't the pattern we're looking for.
1376 if (!CFVal || !CTVal)
1379 // We can commute the SELECT_CC by inverting the condition. This
1380 // might be needed to make this fit into a CSINV pattern.
1381 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1382 std::swap(TVal, FVal);
1383 std::swap(CTVal, CFVal);
1384 CC = ISD::getSetCCInverse(CC, true);
1387 // If the constants line up, perform the transform!
1388 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1390 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1393 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1394 DAG.getConstant(-1ULL, Other.getValueType()));
1396 return DAG.getNode(AArch64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1403 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1404 EVT VT = Op.getValueType();
1406 // Let legalize expand this if it isn't a legal type yet.
1407 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1410 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1413 bool ExtraOp = false;
1414 switch (Op.getOpcode()) {
1416 llvm_unreachable("Invalid code");
1418 Opc = AArch64ISD::ADDS;
1421 Opc = AArch64ISD::SUBS;
1424 Opc = AArch64ISD::ADCS;
1428 Opc = AArch64ISD::SBCS;
1434 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1435 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1439 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1440 // Let legalize expand this if it isn't a legal type yet.
1441 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1444 AArch64CC::CondCode CC;
1445 // The actual operation that sets the overflow or carry flag.
1446 SDValue Value, Overflow;
1447 std::tie(Value, Overflow) = getAArch64XALUOOp(CC, Op, DAG);
1449 // We use 0 and 1 as false and true values.
1450 SDValue TVal = DAG.getConstant(1, MVT::i32);
1451 SDValue FVal = DAG.getConstant(0, MVT::i32);
1453 // We use an inverted condition, because the conditional select is inverted
1454 // too. This will allow it to be selected to a single instruction:
1455 // CSINC Wd, WZR, WZR, invert(cond).
1456 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), MVT::i32);
1457 Overflow = DAG.getNode(AArch64ISD::CSEL, SDLoc(Op), MVT::i32, FVal, TVal,
1460 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1461 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
1464 // Prefetch operands are:
1465 // 1: Address to prefetch
1467 // 3: int locality (0 = no locality ... 3 = extreme locality)
1468 // 4: bool isDataCache
1469 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1471 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1472 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1473 unsigned IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1475 bool IsStream = !Locality;
1476 // When the locality number is set
1478 // The front-end should have filtered out the out-of-range values
1479 assert(Locality <= 3 && "Prefetch locality out-of-range");
1480 // The locality degree is the opposite of the cache speed.
1481 // Put the number the other way around.
1482 // The encoding starts at 0 for level 1
1483 Locality = 3 - Locality;
1486 // built the mask value encoding the expected behavior.
1487 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1488 (!IsData << 3) | // IsDataCache bit
1489 (Locality << 1) | // Cache level bits
1490 (unsigned)IsStream; // Stream bit
1491 return DAG.getNode(AArch64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1492 DAG.getConstant(PrfOp, MVT::i32), Op.getOperand(1));
1495 SDValue AArch64TargetLowering::LowerFP_EXTEND(SDValue Op,
1496 SelectionDAG &DAG) const {
1497 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1500 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1502 return LowerF128Call(Op, DAG, LC);
1505 SDValue AArch64TargetLowering::LowerFP_ROUND(SDValue Op,
1506 SelectionDAG &DAG) const {
1507 if (Op.getOperand(0).getValueType() != MVT::f128) {
1508 // It's legal except when f128 is involved
1513 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1515 // FP_ROUND node has a second operand indicating whether it is known to be
1516 // precise. That doesn't take part in the LibCall so we can't directly use
1518 SDValue SrcVal = Op.getOperand(0);
1519 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1520 /*isSigned*/ false, SDLoc(Op)).first;
1523 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1524 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1525 // Any additional optimization in this function should be recorded
1526 // in the cost tables.
1527 EVT InVT = Op.getOperand(0).getValueType();
1528 EVT VT = Op.getValueType();
1530 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1533 DAG.getNode(Op.getOpcode(), dl, InVT.changeVectorElementTypeToInteger(),
1535 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1538 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1541 MVT::getVectorVT(MVT::getFloatingPointVT(VT.getScalarSizeInBits()),
1542 VT.getVectorNumElements());
1543 SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, ExtVT, Op.getOperand(0));
1544 return DAG.getNode(Op.getOpcode(), dl, VT, Ext);
1547 // Type changing conversions are illegal.
1551 SDValue AArch64TargetLowering::LowerFP_TO_INT(SDValue Op,
1552 SelectionDAG &DAG) const {
1553 if (Op.getOperand(0).getValueType().isVector())
1554 return LowerVectorFP_TO_INT(Op, DAG);
1556 if (Op.getOperand(0).getValueType() != MVT::f128) {
1557 // It's legal except when f128 is involved
1562 if (Op.getOpcode() == ISD::FP_TO_SINT)
1563 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1565 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1567 SmallVector<SDValue, 2> Ops;
1568 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1569 Ops.push_back(Op.getOperand(i));
1571 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1575 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1576 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1577 // Any additional optimization in this function should be recorded
1578 // in the cost tables.
1579 EVT VT = Op.getValueType();
1581 SDValue In = Op.getOperand(0);
1582 EVT InVT = In.getValueType();
1584 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1586 MVT::getVectorVT(MVT::getFloatingPointVT(InVT.getScalarSizeInBits()),
1587 InVT.getVectorNumElements());
1588 In = DAG.getNode(Op.getOpcode(), dl, CastVT, In);
1589 return DAG.getNode(ISD::FP_ROUND, dl, VT, In, DAG.getIntPtrConstant(0));
1592 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1594 Op.getOpcode() == ISD::SINT_TO_FP ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1595 EVT CastVT = VT.changeVectorElementTypeToInteger();
1596 In = DAG.getNode(CastOpc, dl, CastVT, In);
1597 return DAG.getNode(Op.getOpcode(), dl, VT, In);
1603 SDValue AArch64TargetLowering::LowerINT_TO_FP(SDValue Op,
1604 SelectionDAG &DAG) const {
1605 if (Op.getValueType().isVector())
1606 return LowerVectorINT_TO_FP(Op, DAG);
1608 // i128 conversions are libcalls.
1609 if (Op.getOperand(0).getValueType() == MVT::i128)
1612 // Other conversions are legal, unless it's to the completely software-based
1614 if (Op.getValueType() != MVT::f128)
1618 if (Op.getOpcode() == ISD::SINT_TO_FP)
1619 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1621 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1623 return LowerF128Call(Op, DAG, LC);
1626 SDValue AArch64TargetLowering::LowerFSINCOS(SDValue Op,
1627 SelectionDAG &DAG) const {
1628 // For iOS, we want to call an alternative entry point: __sincos_stret,
1629 // which returns the values in two S / D registers.
1631 SDValue Arg = Op.getOperand(0);
1632 EVT ArgVT = Arg.getValueType();
1633 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1640 Entry.isSExt = false;
1641 Entry.isZExt = false;
1642 Args.push_back(Entry);
1644 const char *LibcallName =
1645 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1646 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
1648 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
1649 TargetLowering::CallLoweringInfo CLI(DAG);
1650 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
1651 .setCallee(CallingConv::Fast, RetTy, Callee, std::move(Args), 0);
1653 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1654 return CallResult.first;
1657 static SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) {
1658 if (Op.getValueType() != MVT::f16)
1661 assert(Op.getOperand(0).getValueType() == MVT::i16);
1664 Op = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op.getOperand(0));
1665 Op = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Op);
1667 DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, MVT::f16, Op,
1668 DAG.getTargetConstant(AArch64::hsub, MVT::i32)),
1672 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
1673 if (OrigVT.getSizeInBits() >= 64)
1676 assert(OrigVT.isSimple() && "Expecting a simple value type");
1678 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
1679 switch (OrigSimpleTy) {
1680 default: llvm_unreachable("Unexpected Vector Type");
1689 static SDValue addRequiredExtensionForVectorMULL(SDValue N, SelectionDAG &DAG,
1692 unsigned ExtOpcode) {
1693 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
1694 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
1695 // 64-bits we need to insert a new extension so that it will be 64-bits.
1696 assert(ExtTy.is128BitVector() && "Unexpected extension size");
1697 if (OrigTy.getSizeInBits() >= 64)
1700 // Must extend size to at least 64 bits to be used as an operand for VMULL.
1701 EVT NewVT = getExtensionTo64Bits(OrigTy);
1703 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
1706 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
1708 EVT VT = N->getValueType(0);
1710 if (N->getOpcode() != ISD::BUILD_VECTOR)
1713 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1714 SDNode *Elt = N->getOperand(i).getNode();
1715 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
1716 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1717 unsigned HalfSize = EltSize / 2;
1719 if (!isIntN(HalfSize, C->getSExtValue()))
1722 if (!isUIntN(HalfSize, C->getZExtValue()))
1733 static SDValue skipExtensionForVectorMULL(SDNode *N, SelectionDAG &DAG) {
1734 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
1735 return addRequiredExtensionForVectorMULL(N->getOperand(0), DAG,
1736 N->getOperand(0)->getValueType(0),
1740 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
1741 EVT VT = N->getValueType(0);
1742 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
1743 unsigned NumElts = VT.getVectorNumElements();
1744 MVT TruncVT = MVT::getIntegerVT(EltSize);
1745 SmallVector<SDValue, 8> Ops;
1746 for (unsigned i = 0; i != NumElts; ++i) {
1747 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
1748 const APInt &CInt = C->getAPIntValue();
1749 // Element types smaller than 32 bits are not legal, so use i32 elements.
1750 // The values are implicitly truncated so sext vs. zext doesn't matter.
1751 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
1753 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
1754 MVT::getVectorVT(TruncVT, NumElts), Ops);
1757 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
1758 if (N->getOpcode() == ISD::SIGN_EXTEND)
1760 if (isExtendedBUILD_VECTOR(N, DAG, true))
1765 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
1766 if (N->getOpcode() == ISD::ZERO_EXTEND)
1768 if (isExtendedBUILD_VECTOR(N, DAG, false))
1773 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
1774 unsigned Opcode = N->getOpcode();
1775 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
1776 SDNode *N0 = N->getOperand(0).getNode();
1777 SDNode *N1 = N->getOperand(1).getNode();
1778 return N0->hasOneUse() && N1->hasOneUse() &&
1779 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
1784 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
1785 unsigned Opcode = N->getOpcode();
1786 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
1787 SDNode *N0 = N->getOperand(0).getNode();
1788 SDNode *N1 = N->getOperand(1).getNode();
1789 return N0->hasOneUse() && N1->hasOneUse() &&
1790 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
1795 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
1796 // Multiplications are only custom-lowered for 128-bit vectors so that
1797 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
1798 EVT VT = Op.getValueType();
1799 assert(VT.is128BitVector() && VT.isInteger() &&
1800 "unexpected type for custom-lowering ISD::MUL");
1801 SDNode *N0 = Op.getOperand(0).getNode();
1802 SDNode *N1 = Op.getOperand(1).getNode();
1803 unsigned NewOpc = 0;
1805 bool isN0SExt = isSignExtended(N0, DAG);
1806 bool isN1SExt = isSignExtended(N1, DAG);
1807 if (isN0SExt && isN1SExt)
1808 NewOpc = AArch64ISD::SMULL;
1810 bool isN0ZExt = isZeroExtended(N0, DAG);
1811 bool isN1ZExt = isZeroExtended(N1, DAG);
1812 if (isN0ZExt && isN1ZExt)
1813 NewOpc = AArch64ISD::UMULL;
1814 else if (isN1SExt || isN1ZExt) {
1815 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
1816 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
1817 if (isN1SExt && isAddSubSExt(N0, DAG)) {
1818 NewOpc = AArch64ISD::SMULL;
1820 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
1821 NewOpc = AArch64ISD::UMULL;
1823 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
1825 NewOpc = AArch64ISD::UMULL;
1831 if (VT == MVT::v2i64)
1832 // Fall through to expand this. It is not legal.
1835 // Other vector multiplications are legal.
1840 // Legalize to a S/UMULL instruction
1843 SDValue Op1 = skipExtensionForVectorMULL(N1, DAG);
1845 Op0 = skipExtensionForVectorMULL(N0, DAG);
1846 assert(Op0.getValueType().is64BitVector() &&
1847 Op1.getValueType().is64BitVector() &&
1848 "unexpected types for extended operands to VMULL");
1849 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
1851 // Optimizing (zext A + zext B) * C, to (S/UMULL A, C) + (S/UMULL B, C) during
1852 // isel lowering to take advantage of no-stall back to back s/umul + s/umla.
1853 // This is true for CPUs with accumulate forwarding such as Cortex-A53/A57
1854 SDValue N00 = skipExtensionForVectorMULL(N0->getOperand(0).getNode(), DAG);
1855 SDValue N01 = skipExtensionForVectorMULL(N0->getOperand(1).getNode(), DAG);
1856 EVT Op1VT = Op1.getValueType();
1857 return DAG.getNode(N0->getOpcode(), DL, VT,
1858 DAG.getNode(NewOpc, DL, VT,
1859 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
1860 DAG.getNode(NewOpc, DL, VT,
1861 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
1864 SDValue AArch64TargetLowering::LowerOperation(SDValue Op,
1865 SelectionDAG &DAG) const {
1866 switch (Op.getOpcode()) {
1868 llvm_unreachable("unimplemented operand");
1871 return LowerBITCAST(Op, DAG);
1872 case ISD::GlobalAddress:
1873 return LowerGlobalAddress(Op, DAG);
1874 case ISD::GlobalTLSAddress:
1875 return LowerGlobalTLSAddress(Op, DAG);
1877 return LowerSETCC(Op, DAG);
1879 return LowerBR_CC(Op, DAG);
1881 return LowerSELECT(Op, DAG);
1882 case ISD::SELECT_CC:
1883 return LowerSELECT_CC(Op, DAG);
1884 case ISD::JumpTable:
1885 return LowerJumpTable(Op, DAG);
1886 case ISD::ConstantPool:
1887 return LowerConstantPool(Op, DAG);
1888 case ISD::BlockAddress:
1889 return LowerBlockAddress(Op, DAG);
1891 return LowerVASTART(Op, DAG);
1893 return LowerVACOPY(Op, DAG);
1895 return LowerVAARG(Op, DAG);
1900 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
1907 return LowerXALUO(Op, DAG);
1909 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
1911 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
1913 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
1915 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
1917 return LowerFP_ROUND(Op, DAG);
1918 case ISD::FP_EXTEND:
1919 return LowerFP_EXTEND(Op, DAG);
1920 case ISD::FRAMEADDR:
1921 return LowerFRAMEADDR(Op, DAG);
1922 case ISD::RETURNADDR:
1923 return LowerRETURNADDR(Op, DAG);
1924 case ISD::INSERT_VECTOR_ELT:
1925 return LowerINSERT_VECTOR_ELT(Op, DAG);
1926 case ISD::EXTRACT_VECTOR_ELT:
1927 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
1928 case ISD::BUILD_VECTOR:
1929 return LowerBUILD_VECTOR(Op, DAG);
1930 case ISD::VECTOR_SHUFFLE:
1931 return LowerVECTOR_SHUFFLE(Op, DAG);
1932 case ISD::EXTRACT_SUBVECTOR:
1933 return LowerEXTRACT_SUBVECTOR(Op, DAG);
1937 return LowerVectorSRA_SRL_SHL(Op, DAG);
1938 case ISD::SHL_PARTS:
1939 return LowerShiftLeftParts(Op, DAG);
1940 case ISD::SRL_PARTS:
1941 case ISD::SRA_PARTS:
1942 return LowerShiftRightParts(Op, DAG);
1944 return LowerCTPOP(Op, DAG);
1945 case ISD::FCOPYSIGN:
1946 return LowerFCOPYSIGN(Op, DAG);
1948 return LowerVectorAND(Op, DAG);
1950 return LowerVectorOR(Op, DAG);
1952 return LowerXOR(Op, DAG);
1954 return LowerPREFETCH(Op, DAG);
1955 case ISD::SINT_TO_FP:
1956 case ISD::UINT_TO_FP:
1957 return LowerINT_TO_FP(Op, DAG);
1958 case ISD::FP_TO_SINT:
1959 case ISD::FP_TO_UINT:
1960 return LowerFP_TO_INT(Op, DAG);
1962 return LowerFSINCOS(Op, DAG);
1964 return LowerMUL(Op, DAG);
1968 /// getFunctionAlignment - Return the Log2 alignment of this function.
1969 unsigned AArch64TargetLowering::getFunctionAlignment(const Function *F) const {
1973 //===----------------------------------------------------------------------===//
1974 // Calling Convention Implementation
1975 //===----------------------------------------------------------------------===//
1977 #include "AArch64GenCallingConv.inc"
1979 /// Selects the correct CCAssignFn for a given CallingConvention value.
1980 CCAssignFn *AArch64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
1981 bool IsVarArg) const {
1984 llvm_unreachable("Unsupported calling convention.");
1985 case CallingConv::WebKit_JS:
1986 return CC_AArch64_WebKit_JS;
1987 case CallingConv::C:
1988 case CallingConv::Fast:
1989 if (!Subtarget->isTargetDarwin())
1990 return CC_AArch64_AAPCS;
1991 return IsVarArg ? CC_AArch64_DarwinPCS_VarArg : CC_AArch64_DarwinPCS;
1995 SDValue AArch64TargetLowering::LowerFormalArguments(
1996 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1997 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
1998 SmallVectorImpl<SDValue> &InVals) const {
1999 MachineFunction &MF = DAG.getMachineFunction();
2000 MachineFrameInfo *MFI = MF.getFrameInfo();
2002 // Assign locations to all of the incoming arguments.
2003 SmallVector<CCValAssign, 16> ArgLocs;
2004 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2007 // At this point, Ins[].VT may already be promoted to i32. To correctly
2008 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2009 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2010 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
2011 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
2013 unsigned NumArgs = Ins.size();
2014 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2015 unsigned CurArgIdx = 0;
2016 for (unsigned i = 0; i != NumArgs; ++i) {
2017 MVT ValVT = Ins[i].VT;
2018 std::advance(CurOrigArg, Ins[i].OrigArgIndex - CurArgIdx);
2019 CurArgIdx = Ins[i].OrigArgIndex;
2021 // Get type of the original argument.
2022 EVT ActualVT = getValueType(CurOrigArg->getType(), /*AllowUnknown*/ true);
2023 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
2024 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2025 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2027 else if (ActualMVT == MVT::i16)
2030 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2032 AssignFn(i, ValVT, ValVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
2033 assert(!Res && "Call operand has unhandled type");
2036 assert(ArgLocs.size() == Ins.size());
2037 SmallVector<SDValue, 16> ArgValues;
2038 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2039 CCValAssign &VA = ArgLocs[i];
2041 if (Ins[i].Flags.isByVal()) {
2042 // Byval is used for HFAs in the PCS, but the system should work in a
2043 // non-compliant manner for larger structs.
2044 EVT PtrTy = getPointerTy();
2045 int Size = Ins[i].Flags.getByValSize();
2046 unsigned NumRegs = (Size + 7) / 8;
2048 // FIXME: This works on big-endian for composite byvals, which are the common
2049 // case. It should also work for fundamental types too.
2051 MFI->CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false);
2052 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrTy);
2053 InVals.push_back(FrameIdxN);
2058 if (VA.isRegLoc()) {
2059 // Arguments stored in registers.
2060 EVT RegVT = VA.getLocVT();
2063 const TargetRegisterClass *RC;
2065 if (RegVT == MVT::i32)
2066 RC = &AArch64::GPR32RegClass;
2067 else if (RegVT == MVT::i64)
2068 RC = &AArch64::GPR64RegClass;
2069 else if (RegVT == MVT::f16)
2070 RC = &AArch64::FPR16RegClass;
2071 else if (RegVT == MVT::f32)
2072 RC = &AArch64::FPR32RegClass;
2073 else if (RegVT == MVT::f64 || RegVT.is64BitVector())
2074 RC = &AArch64::FPR64RegClass;
2075 else if (RegVT == MVT::f128 || RegVT.is128BitVector())
2076 RC = &AArch64::FPR128RegClass;
2078 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2080 // Transform the arguments in physical registers into virtual ones.
2081 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2082 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2084 // If this is an 8, 16 or 32-bit value, it is really passed promoted
2085 // to 64 bits. Insert an assert[sz]ext to capture this, then
2086 // truncate to the right size.
2087 switch (VA.getLocInfo()) {
2089 llvm_unreachable("Unknown loc info!");
2090 case CCValAssign::Full:
2092 case CCValAssign::BCvt:
2093 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
2095 case CCValAssign::AExt:
2096 case CCValAssign::SExt:
2097 case CCValAssign::ZExt:
2098 // SelectionDAGBuilder will insert appropriate AssertZExt & AssertSExt
2099 // nodes after our lowering.
2100 assert(RegVT == Ins[i].VT && "incorrect register location selected");
2104 InVals.push_back(ArgValue);
2106 } else { // VA.isRegLoc()
2107 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
2108 unsigned ArgOffset = VA.getLocMemOffset();
2109 unsigned ArgSize = VA.getValVT().getSizeInBits() / 8;
2111 uint32_t BEAlign = 0;
2112 if (!Subtarget->isLittleEndian() && ArgSize < 8 &&
2113 !Ins[i].Flags.isInConsecutiveRegs())
2114 BEAlign = 8 - ArgSize;
2116 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset + BEAlign, true);
2118 // Create load nodes to retrieve arguments from the stack.
2119 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2122 // For NON_EXTLOAD, generic code in getLoad assert(ValVT == MemVT)
2123 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
2124 MVT MemVT = VA.getValVT();
2126 switch (VA.getLocInfo()) {
2129 case CCValAssign::BCvt:
2130 MemVT = VA.getLocVT();
2132 case CCValAssign::SExt:
2133 ExtType = ISD::SEXTLOAD;
2135 case CCValAssign::ZExt:
2136 ExtType = ISD::ZEXTLOAD;
2138 case CCValAssign::AExt:
2139 ExtType = ISD::EXTLOAD;
2143 ArgValue = DAG.getExtLoad(ExtType, DL, VA.getLocVT(), Chain, FIN,
2144 MachinePointerInfo::getFixedStack(FI),
2145 MemVT, false, false, false, 0);
2147 InVals.push_back(ArgValue);
2153 if (!Subtarget->isTargetDarwin()) {
2154 // The AAPCS variadic function ABI is identical to the non-variadic
2155 // one. As a result there may be more arguments in registers and we should
2156 // save them for future reference.
2157 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
2160 AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>();
2161 // This will point to the next argument passed via stack.
2162 unsigned StackOffset = CCInfo.getNextStackOffset();
2163 // We currently pass all varargs at 8-byte alignment.
2164 StackOffset = ((StackOffset + 7) & ~7);
2165 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
2168 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2169 unsigned StackArgSize = CCInfo.getNextStackOffset();
2170 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2171 if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
2172 // This is a non-standard ABI so by fiat I say we're allowed to make full
2173 // use of the stack area to be popped, which must be aligned to 16 bytes in
2175 StackArgSize = RoundUpToAlignment(StackArgSize, 16);
2177 // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
2178 // a multiple of 16.
2179 FuncInfo->setArgumentStackToRestore(StackArgSize);
2181 // This realignment carries over to the available bytes below. Our own
2182 // callers will guarantee the space is free by giving an aligned value to
2185 // Even if we're not expected to free up the space, it's useful to know how
2186 // much is there while considering tail calls (because we can reuse it).
2187 FuncInfo->setBytesInStackArgArea(StackArgSize);
2192 void AArch64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
2193 SelectionDAG &DAG, SDLoc DL,
2194 SDValue &Chain) const {
2195 MachineFunction &MF = DAG.getMachineFunction();
2196 MachineFrameInfo *MFI = MF.getFrameInfo();
2197 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2199 SmallVector<SDValue, 8> MemOps;
2201 static const MCPhysReg GPRArgRegs[] = { AArch64::X0, AArch64::X1, AArch64::X2,
2202 AArch64::X3, AArch64::X4, AArch64::X5,
2203 AArch64::X6, AArch64::X7 };
2204 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
2205 unsigned FirstVariadicGPR =
2206 CCInfo.getFirstUnallocated(GPRArgRegs, NumGPRArgRegs);
2208 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
2210 if (GPRSaveSize != 0) {
2211 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
2213 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
2215 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
2216 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &AArch64::GPR64RegClass);
2217 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
2219 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2220 MachinePointerInfo::getStack(i * 8), false, false, 0);
2221 MemOps.push_back(Store);
2222 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2223 DAG.getConstant(8, getPointerTy()));
2226 FuncInfo->setVarArgsGPRIndex(GPRIdx);
2227 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
2229 if (Subtarget->hasFPARMv8()) {
2230 static const MCPhysReg FPRArgRegs[] = {
2231 AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3,
2232 AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7};
2233 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
2234 unsigned FirstVariadicFPR =
2235 CCInfo.getFirstUnallocated(FPRArgRegs, NumFPRArgRegs);
2237 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
2239 if (FPRSaveSize != 0) {
2240 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
2242 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
2244 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
2245 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &AArch64::FPR128RegClass);
2246 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);
2249 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2250 MachinePointerInfo::getStack(i * 16), false, false, 0);
2251 MemOps.push_back(Store);
2252 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2253 DAG.getConstant(16, getPointerTy()));
2256 FuncInfo->setVarArgsFPRIndex(FPRIdx);
2257 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
2260 if (!MemOps.empty()) {
2261 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
2265 /// LowerCallResult - Lower the result values of a call into the
2266 /// appropriate copies out of appropriate physical registers.
2267 SDValue AArch64TargetLowering::LowerCallResult(
2268 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
2269 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2270 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
2271 SDValue ThisVal) const {
2272 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2273 ? RetCC_AArch64_WebKit_JS
2274 : RetCC_AArch64_AAPCS;
2275 // Assign locations to each value returned by this call.
2276 SmallVector<CCValAssign, 16> RVLocs;
2277 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2279 CCInfo.AnalyzeCallResult(Ins, RetCC);
2281 // Copy all of the result registers out of their specified physreg.
2282 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2283 CCValAssign VA = RVLocs[i];
2285 // Pass 'this' value directly from the argument to return value, to avoid
2286 // reg unit interference
2287 if (i == 0 && isThisReturn) {
2288 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
2289 "unexpected return calling convention register assignment");
2290 InVals.push_back(ThisVal);
2295 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
2296 Chain = Val.getValue(1);
2297 InFlag = Val.getValue(2);
2299 switch (VA.getLocInfo()) {
2301 llvm_unreachable("Unknown loc info!");
2302 case CCValAssign::Full:
2304 case CCValAssign::BCvt:
2305 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2309 InVals.push_back(Val);
2315 bool AArch64TargetLowering::isEligibleForTailCallOptimization(
2316 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
2317 bool isCalleeStructRet, bool isCallerStructRet,
2318 const SmallVectorImpl<ISD::OutputArg> &Outs,
2319 const SmallVectorImpl<SDValue> &OutVals,
2320 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
2321 // For CallingConv::C this function knows whether the ABI needs
2322 // changing. That's not true for other conventions so they will have to opt in
2324 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
2327 const MachineFunction &MF = DAG.getMachineFunction();
2328 const Function *CallerF = MF.getFunction();
2329 CallingConv::ID CallerCC = CallerF->getCallingConv();
2330 bool CCMatch = CallerCC == CalleeCC;
2332 // Byval parameters hand the function a pointer directly into the stack area
2333 // we want to reuse during a tail call. Working around this *is* possible (see
2334 // X86) but less efficient and uglier in LowerCall.
2335 for (Function::const_arg_iterator i = CallerF->arg_begin(),
2336 e = CallerF->arg_end();
2338 if (i->hasByValAttr())
2341 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
2342 if (IsTailCallConvention(CalleeCC) && CCMatch)
2347 // Externally-defined functions with weak linkage should not be
2348 // tail-called on AArch64 when the OS does not support dynamic
2349 // pre-emption of symbols, as the AAELF spec requires normal calls
2350 // to undefined weak functions to be replaced with a NOP or jump to the
2351 // next instruction. The behaviour of branch instructions in this
2352 // situation (as used for tail calls) is implementation-defined, so we
2353 // cannot rely on the linker replacing the tail call with a return.
2354 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2355 const GlobalValue *GV = G->getGlobal();
2356 const Triple TT(getTargetMachine().getTargetTriple());
2357 if (GV->hasExternalWeakLinkage() &&
2358 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2362 // Now we search for cases where we can use a tail call without changing the
2363 // ABI. Sibcall is used in some places (particularly gcc) to refer to this
2366 // I want anyone implementing a new calling convention to think long and hard
2367 // about this assert.
2368 assert((!isVarArg || CalleeCC == CallingConv::C) &&
2369 "Unexpected variadic calling convention");
2371 if (isVarArg && !Outs.empty()) {
2372 // At least two cases here: if caller is fastcc then we can't have any
2373 // memory arguments (we'd be expected to clean up the stack afterwards). If
2374 // caller is C then we could potentially use its argument area.
2376 // FIXME: for now we take the most conservative of these in both cases:
2377 // disallow all variadic memory operands.
2378 SmallVector<CCValAssign, 16> ArgLocs;
2379 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2382 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, true));
2383 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2384 if (!ArgLocs[i].isRegLoc())
2388 // If the calling conventions do not match, then we'd better make sure the
2389 // results are returned in the same way as what the caller expects.
2391 SmallVector<CCValAssign, 16> RVLocs1;
2392 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2394 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForCall(CalleeCC, isVarArg));
2396 SmallVector<CCValAssign, 16> RVLocs2;
2397 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2399 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForCall(CallerCC, isVarArg));
2401 if (RVLocs1.size() != RVLocs2.size())
2403 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2404 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2406 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2408 if (RVLocs1[i].isRegLoc()) {
2409 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2412 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2418 // Nothing more to check if the callee is taking no arguments
2422 SmallVector<CCValAssign, 16> ArgLocs;
2423 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2426 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, isVarArg));
2428 const AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2430 // If the stack arguments for this call would fit into our own save area then
2431 // the call can be made tail.
2432 return CCInfo.getNextStackOffset() <= FuncInfo->getBytesInStackArgArea();
2435 SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain,
2437 MachineFrameInfo *MFI,
2438 int ClobberedFI) const {
2439 SmallVector<SDValue, 8> ArgChains;
2440 int64_t FirstByte = MFI->getObjectOffset(ClobberedFI);
2441 int64_t LastByte = FirstByte + MFI->getObjectSize(ClobberedFI) - 1;
2443 // Include the original chain at the beginning of the list. When this is
2444 // used by target LowerCall hooks, this helps legalize find the
2445 // CALLSEQ_BEGIN node.
2446 ArgChains.push_back(Chain);
2448 // Add a chain value for each stack argument corresponding
2449 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
2450 UE = DAG.getEntryNode().getNode()->use_end();
2452 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
2453 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
2454 if (FI->getIndex() < 0) {
2455 int64_t InFirstByte = MFI->getObjectOffset(FI->getIndex());
2456 int64_t InLastByte = InFirstByte;
2457 InLastByte += MFI->getObjectSize(FI->getIndex()) - 1;
2459 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
2460 (FirstByte <= InFirstByte && InFirstByte <= LastByte))
2461 ArgChains.push_back(SDValue(L, 1));
2464 // Build a tokenfactor for all the chains.
2465 return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
2468 bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
2469 bool TailCallOpt) const {
2470 return CallCC == CallingConv::Fast && TailCallOpt;
2473 bool AArch64TargetLowering::IsTailCallConvention(CallingConv::ID CallCC) const {
2474 return CallCC == CallingConv::Fast;
2477 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
2478 /// and add input and output parameter nodes.
2480 AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
2481 SmallVectorImpl<SDValue> &InVals) const {
2482 SelectionDAG &DAG = CLI.DAG;
2484 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2485 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2486 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2487 SDValue Chain = CLI.Chain;
2488 SDValue Callee = CLI.Callee;
2489 bool &IsTailCall = CLI.IsTailCall;
2490 CallingConv::ID CallConv = CLI.CallConv;
2491 bool IsVarArg = CLI.IsVarArg;
2493 MachineFunction &MF = DAG.getMachineFunction();
2494 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2495 bool IsThisReturn = false;
2497 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2498 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2499 bool IsSibCall = false;
2502 // Check if it's really possible to do a tail call.
2503 IsTailCall = isEligibleForTailCallOptimization(
2504 Callee, CallConv, IsVarArg, IsStructRet,
2505 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2506 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
2507 report_fatal_error("failed to perform tail call elimination on a call "
2508 "site marked musttail");
2510 // A sibling call is one where we're under the usual C ABI and not planning
2511 // to change that but can still do a tail call:
2512 if (!TailCallOpt && IsTailCall)
2519 // Analyze operands of the call, assigning locations to each operand.
2520 SmallVector<CCValAssign, 16> ArgLocs;
2521 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
2525 // Handle fixed and variable vector arguments differently.
2526 // Variable vector arguments always go into memory.
2527 unsigned NumArgs = Outs.size();
2529 for (unsigned i = 0; i != NumArgs; ++i) {
2530 MVT ArgVT = Outs[i].VT;
2531 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2532 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2533 /*IsVarArg=*/ !Outs[i].IsFixed);
2534 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2535 assert(!Res && "Call operand has unhandled type");
2539 // At this point, Outs[].VT may already be promoted to i32. To correctly
2540 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2541 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2542 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2543 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2545 unsigned NumArgs = Outs.size();
2546 for (unsigned i = 0; i != NumArgs; ++i) {
2547 MVT ValVT = Outs[i].VT;
2548 // Get type of the original argument.
2549 EVT ActualVT = getValueType(CLI.getArgs()[Outs[i].OrigArgIndex].Ty,
2550 /*AllowUnknown*/ true);
2551 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2552 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2553 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2554 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2556 else if (ActualMVT == MVT::i16)
2559 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2560 bool Res = AssignFn(i, ValVT, ValVT, CCValAssign::Full, ArgFlags, CCInfo);
2561 assert(!Res && "Call operand has unhandled type");
2566 // Get a count of how many bytes are to be pushed on the stack.
2567 unsigned NumBytes = CCInfo.getNextStackOffset();
2570 // Since we're not changing the ABI to make this a tail call, the memory
2571 // operands are already available in the caller's incoming argument space.
2575 // FPDiff is the byte offset of the call's argument area from the callee's.
2576 // Stores to callee stack arguments will be placed in FixedStackSlots offset
2577 // by this amount for a tail call. In a sibling call it must be 0 because the
2578 // caller will deallocate the entire stack and the callee still expects its
2579 // arguments to begin at SP+0. Completely unused for non-tail calls.
2582 if (IsTailCall && !IsSibCall) {
2583 unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();
2585 // Since callee will pop argument stack as a tail call, we must keep the
2586 // popped size 16-byte aligned.
2587 NumBytes = RoundUpToAlignment(NumBytes, 16);
2589 // FPDiff will be negative if this tail call requires more space than we
2590 // would automatically have in our incoming argument space. Positive if we
2591 // can actually shrink the stack.
2592 FPDiff = NumReusableBytes - NumBytes;
2594 // The stack pointer must be 16-byte aligned at all times it's used for a
2595 // memory operation, which in practice means at *all* times and in
2596 // particular across call boundaries. Therefore our own arguments started at
2597 // a 16-byte aligned SP and the delta applied for the tail call should
2598 // satisfy the same constraint.
2599 assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
2602 // Adjust the stack pointer for the new arguments...
2603 // These operations are automatically eliminated by the prolog/epilog pass
2606 DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true), DL);
2608 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, AArch64::SP, getPointerTy());
2610 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2611 SmallVector<SDValue, 8> MemOpChains;
2613 // Walk the register/memloc assignments, inserting copies/loads.
2614 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2615 ++i, ++realArgIdx) {
2616 CCValAssign &VA = ArgLocs[i];
2617 SDValue Arg = OutVals[realArgIdx];
2618 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2620 // Promote the value if needed.
2621 switch (VA.getLocInfo()) {
2623 llvm_unreachable("Unknown loc info!");
2624 case CCValAssign::Full:
2626 case CCValAssign::SExt:
2627 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
2629 case CCValAssign::ZExt:
2630 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2632 case CCValAssign::AExt:
2633 if (Outs[realArgIdx].ArgVT == MVT::i1) {
2634 // AAPCS requires i1 to be zero-extended to 8-bits by the caller.
2635 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
2636 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i8, Arg);
2638 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
2640 case CCValAssign::BCvt:
2641 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2643 case CCValAssign::FPExt:
2644 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
2648 if (VA.isRegLoc()) {
2649 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
2650 assert(VA.getLocVT() == MVT::i64 &&
2651 "unexpected calling convention register assignment");
2652 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
2653 "unexpected use of 'returned'");
2654 IsThisReturn = true;
2656 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2658 assert(VA.isMemLoc());
2661 MachinePointerInfo DstInfo;
2663 // FIXME: This works on big-endian for composite byvals, which are the
2664 // common case. It should also work for fundamental types too.
2665 uint32_t BEAlign = 0;
2666 unsigned OpSize = Flags.isByVal() ? Flags.getByValSize() * 8
2667 : VA.getValVT().getSizeInBits();
2668 OpSize = (OpSize + 7) / 8;
2669 if (!Subtarget->isLittleEndian() && !Flags.isByVal() &&
2670 !Flags.isInConsecutiveRegs()) {
2672 BEAlign = 8 - OpSize;
2674 unsigned LocMemOffset = VA.getLocMemOffset();
2675 int32_t Offset = LocMemOffset + BEAlign;
2676 SDValue PtrOff = DAG.getIntPtrConstant(Offset);
2677 PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2680 Offset = Offset + FPDiff;
2681 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2683 DstAddr = DAG.getFrameIndex(FI, getPointerTy());
2684 DstInfo = MachinePointerInfo::getFixedStack(FI);
2686 // Make sure any stack arguments overlapping with where we're storing
2687 // are loaded before this eventual operation. Otherwise they'll be
2689 Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
2691 SDValue PtrOff = DAG.getIntPtrConstant(Offset);
2693 DstAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2694 DstInfo = MachinePointerInfo::getStack(LocMemOffset);
2697 if (Outs[i].Flags.isByVal()) {
2699 DAG.getConstant(Outs[i].Flags.getByValSize(), MVT::i64);
2700 SDValue Cpy = DAG.getMemcpy(
2701 Chain, DL, DstAddr, Arg, SizeNode, Outs[i].Flags.getByValAlign(),
2703 /*AlwaysInline = */ false, DstInfo, MachinePointerInfo());
2705 MemOpChains.push_back(Cpy);
2707 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
2708 // promoted to a legal register type i32, we should truncate Arg back to
2710 if (VA.getValVT() == MVT::i1 || VA.getValVT() == MVT::i8 ||
2711 VA.getValVT() == MVT::i16)
2712 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Arg);
2715 DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo, false, false, 0);
2716 MemOpChains.push_back(Store);
2721 if (!MemOpChains.empty())
2722 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
2724 // Build a sequence of copy-to-reg nodes chained together with token chain
2725 // and flag operands which copy the outgoing args into the appropriate regs.
2727 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2728 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[i].first,
2729 RegsToPass[i].second, InFlag);
2730 InFlag = Chain.getValue(1);
2733 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2734 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2735 // node so that legalize doesn't hack it.
2736 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
2737 Subtarget->isTargetMachO()) {
2738 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2739 const GlobalValue *GV = G->getGlobal();
2740 bool InternalLinkage = GV->hasInternalLinkage();
2741 if (InternalLinkage)
2742 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2744 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0,
2746 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, getPointerTy(), Callee);
2748 } else if (ExternalSymbolSDNode *S =
2749 dyn_cast<ExternalSymbolSDNode>(Callee)) {
2750 const char *Sym = S->getSymbol();
2752 DAG.getTargetExternalSymbol(Sym, getPointerTy(), AArch64II::MO_GOT);
2753 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, getPointerTy(), Callee);
2755 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2756 const GlobalValue *GV = G->getGlobal();
2757 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2758 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2759 const char *Sym = S->getSymbol();
2760 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), 0);
2763 // We don't usually want to end the call-sequence here because we would tidy
2764 // the frame up *after* the call, however in the ABI-changing tail-call case
2765 // we've carefully laid out the parameters so that when sp is reset they'll be
2766 // in the correct location.
2767 if (IsTailCall && !IsSibCall) {
2768 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2769 DAG.getIntPtrConstant(0, true), InFlag, DL);
2770 InFlag = Chain.getValue(1);
2773 std::vector<SDValue> Ops;
2774 Ops.push_back(Chain);
2775 Ops.push_back(Callee);
2778 // Each tail call may have to adjust the stack by a different amount, so
2779 // this information must travel along with the operation for eventual
2780 // consumption by emitEpilogue.
2781 Ops.push_back(DAG.getTargetConstant(FPDiff, MVT::i32));
2784 // Add argument registers to the end of the list so that they are known live
2786 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2787 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2788 RegsToPass[i].second.getValueType()));
2790 // Add a register mask operand representing the call-preserved registers.
2791 const uint32_t *Mask;
2792 const TargetRegisterInfo *TRI =
2793 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
2794 const AArch64RegisterInfo *ARI =
2795 static_cast<const AArch64RegisterInfo *>(TRI);
2797 // For 'this' returns, use the X0-preserving mask if applicable
2798 Mask = ARI->getThisReturnPreservedMask(CallConv);
2800 IsThisReturn = false;
2801 Mask = ARI->getCallPreservedMask(CallConv);
2804 Mask = ARI->getCallPreservedMask(CallConv);
2806 assert(Mask && "Missing call preserved mask for calling convention");
2807 Ops.push_back(DAG.getRegisterMask(Mask));
2809 if (InFlag.getNode())
2810 Ops.push_back(InFlag);
2812 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2814 // If we're doing a tall call, use a TC_RETURN here rather than an
2815 // actual call instruction.
2817 return DAG.getNode(AArch64ISD::TC_RETURN, DL, NodeTys, Ops);
2819 // Returns a chain and a flag for retval copy to use.
2820 Chain = DAG.getNode(AArch64ISD::CALL, DL, NodeTys, Ops);
2821 InFlag = Chain.getValue(1);
2823 uint64_t CalleePopBytes = DoesCalleeRestoreStack(CallConv, TailCallOpt)
2824 ? RoundUpToAlignment(NumBytes, 16)
2827 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2828 DAG.getIntPtrConstant(CalleePopBytes, true),
2831 InFlag = Chain.getValue(1);
2833 // Handle result values, copying them out of physregs into vregs that we
2835 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2836 InVals, IsThisReturn,
2837 IsThisReturn ? OutVals[0] : SDValue());
2840 bool AArch64TargetLowering::CanLowerReturn(
2841 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2842 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2843 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2844 ? RetCC_AArch64_WebKit_JS
2845 : RetCC_AArch64_AAPCS;
2846 SmallVector<CCValAssign, 16> RVLocs;
2847 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2848 return CCInfo.CheckReturn(Outs, RetCC);
2852 AArch64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2854 const SmallVectorImpl<ISD::OutputArg> &Outs,
2855 const SmallVectorImpl<SDValue> &OutVals,
2856 SDLoc DL, SelectionDAG &DAG) const {
2857 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2858 ? RetCC_AArch64_WebKit_JS
2859 : RetCC_AArch64_AAPCS;
2860 SmallVector<CCValAssign, 16> RVLocs;
2861 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2863 CCInfo.AnalyzeReturn(Outs, RetCC);
2865 // Copy the result values into the output registers.
2867 SmallVector<SDValue, 4> RetOps(1, Chain);
2868 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
2869 ++i, ++realRVLocIdx) {
2870 CCValAssign &VA = RVLocs[i];
2871 assert(VA.isRegLoc() && "Can only return in registers!");
2872 SDValue Arg = OutVals[realRVLocIdx];
2874 switch (VA.getLocInfo()) {
2876 llvm_unreachable("Unknown loc info!");
2877 case CCValAssign::Full:
2878 if (Outs[i].ArgVT == MVT::i1) {
2879 // AAPCS requires i1 to be zero-extended to i8 by the producer of the
2880 // value. This is strictly redundant on Darwin (which uses "zeroext
2881 // i1"), but will be optimised out before ISel.
2882 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
2883 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2886 case CCValAssign::BCvt:
2887 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2891 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
2892 Flag = Chain.getValue(1);
2893 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2896 RetOps[0] = Chain; // Update chain.
2898 // Add the flag if we have it.
2900 RetOps.push_back(Flag);
2902 return DAG.getNode(AArch64ISD::RET_FLAG, DL, MVT::Other, RetOps);
2905 //===----------------------------------------------------------------------===//
2906 // Other Lowering Code
2907 //===----------------------------------------------------------------------===//
2909 SDValue AArch64TargetLowering::LowerGlobalAddress(SDValue Op,
2910 SelectionDAG &DAG) const {
2911 EVT PtrVT = getPointerTy();
2913 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
2914 const GlobalValue *GV = GN->getGlobal();
2915 unsigned char OpFlags =
2916 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
2918 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
2919 "unexpected offset in global node");
2921 // This also catched the large code model case for Darwin.
2922 if ((OpFlags & AArch64II::MO_GOT) != 0) {
2923 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
2924 // FIXME: Once remat is capable of dealing with instructions with register
2925 // operands, expand this into two nodes instead of using a wrapper node.
2926 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
2929 if ((OpFlags & AArch64II::MO_CONSTPOOL) != 0) {
2930 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
2931 "use of MO_CONSTPOOL only supported on small model");
2932 SDValue Hi = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, AArch64II::MO_PAGE);
2933 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
2934 unsigned char LoFlags = AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
2935 SDValue Lo = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, LoFlags);
2936 SDValue PoolAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
2937 SDValue GlobalAddr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), PoolAddr,
2938 MachinePointerInfo::getConstantPool(),
2939 /*isVolatile=*/ false,
2940 /*isNonTemporal=*/ true,
2941 /*isInvariant=*/ true, 8);
2942 if (GN->getOffset() != 0)
2943 return DAG.getNode(ISD::ADD, DL, PtrVT, GlobalAddr,
2944 DAG.getConstant(GN->getOffset(), PtrVT));
2948 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2949 const unsigned char MO_NC = AArch64II::MO_NC;
2951 AArch64ISD::WrapperLarge, DL, PtrVT,
2952 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G3),
2953 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
2954 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
2955 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
2957 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
2958 // the only correct model on Darwin.
2959 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2960 OpFlags | AArch64II::MO_PAGE);
2961 unsigned char LoFlags = OpFlags | AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
2962 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
2964 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
2965 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
2969 /// \brief Convert a TLS address reference into the correct sequence of loads
2970 /// and calls to compute the variable's address (for Darwin, currently) and
2971 /// return an SDValue containing the final node.
2973 /// Darwin only has one TLS scheme which must be capable of dealing with the
2974 /// fully general situation, in the worst case. This means:
2975 /// + "extern __thread" declaration.
2976 /// + Defined in a possibly unknown dynamic library.
2978 /// The general system is that each __thread variable has a [3 x i64] descriptor
2979 /// which contains information used by the runtime to calculate the address. The
2980 /// only part of this the compiler needs to know about is the first xword, which
2981 /// contains a function pointer that must be called with the address of the
2982 /// entire descriptor in "x0".
2984 /// Since this descriptor may be in a different unit, in general even the
2985 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
2987 /// adrp x0, _var@TLVPPAGE
2988 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
2989 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
2990 /// ; the function pointer
2991 /// blr x1 ; Uses descriptor address in x0
2992 /// ; Address of _var is now in x0.
2994 /// If the address of _var's descriptor *is* known to the linker, then it can
2995 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
2996 /// a slight efficiency gain.
2998 AArch64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
2999 SelectionDAG &DAG) const {
3000 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
3003 MVT PtrVT = getPointerTy();
3004 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3007 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3008 SDValue DescAddr = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TLVPAddr);
3010 // The first entry in the descriptor is a function pointer that we must call
3011 // to obtain the address of the variable.
3012 SDValue Chain = DAG.getEntryNode();
3013 SDValue FuncTLVGet =
3014 DAG.getLoad(MVT::i64, DL, Chain, DescAddr, MachinePointerInfo::getGOT(),
3015 false, true, true, 8);
3016 Chain = FuncTLVGet.getValue(1);
3018 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3019 MFI->setAdjustsStack(true);
3021 // TLS calls preserve all registers except those that absolutely must be
3022 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
3024 const TargetRegisterInfo *TRI =
3025 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
3026 const AArch64RegisterInfo *ARI =
3027 static_cast<const AArch64RegisterInfo *>(TRI);
3028 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
3030 // Finally, we can make the call. This is just a degenerate version of a
3031 // normal AArch64 call node: x0 takes the address of the descriptor, and
3032 // returns the address of the variable in this thread.
3033 Chain = DAG.getCopyToReg(Chain, DL, AArch64::X0, DescAddr, SDValue());
3035 DAG.getNode(AArch64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
3036 Chain, FuncTLVGet, DAG.getRegister(AArch64::X0, MVT::i64),
3037 DAG.getRegisterMask(Mask), Chain.getValue(1));
3038 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Chain.getValue(1));
3041 /// When accessing thread-local variables under either the general-dynamic or
3042 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
3043 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
3044 /// is a function pointer to carry out the resolution. This function takes the
3045 /// address of the descriptor in X0 and returns the TPIDR_EL0 offset in X0. All
3046 /// other registers (except LR, NZCV) are preserved.
3048 /// Thus, the ideal call sequence on AArch64 is:
3050 /// adrp x0, :tlsdesc:thread_var
3051 /// ldr x8, [x0, :tlsdesc_lo12:thread_var]
3052 /// add x0, x0, :tlsdesc_lo12:thread_var
3053 /// .tlsdesccall thread_var
3055 /// (TPIDR_EL0 offset now in x0).
3057 /// The ".tlsdesccall" directive instructs the assembler to insert a particular
3058 /// relocation to help the linker relax this sequence if it turns out to be too
3061 /// FIXME: we currently produce an extra, duplicated, ADRP instruction, but this
3063 SDValue AArch64TargetLowering::LowerELFTLSDescCall(SDValue SymAddr,
3064 SDValue DescAddr, SDLoc DL,
3065 SelectionDAG &DAG) const {
3066 EVT PtrVT = getPointerTy();
3068 // The function we need to call is simply the first entry in the GOT for this
3069 // descriptor, load it in preparation.
3070 SDValue Func = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, SymAddr);
3072 // TLS calls preserve all registers except those that absolutely must be
3073 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
3075 const TargetRegisterInfo *TRI =
3076 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
3077 const AArch64RegisterInfo *ARI =
3078 static_cast<const AArch64RegisterInfo *>(TRI);
3079 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
3081 // The function takes only one argument: the address of the descriptor itself
3083 SDValue Glue, Chain;
3084 Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, AArch64::X0, DescAddr, Glue);
3085 Glue = Chain.getValue(1);
3087 // We're now ready to populate the argument list, as with a normal call:
3088 SmallVector<SDValue, 6> Ops;
3089 Ops.push_back(Chain);
3090 Ops.push_back(Func);
3091 Ops.push_back(SymAddr);
3092 Ops.push_back(DAG.getRegister(AArch64::X0, PtrVT));
3093 Ops.push_back(DAG.getRegisterMask(Mask));
3094 Ops.push_back(Glue);
3096 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3097 Chain = DAG.getNode(AArch64ISD::TLSDESC_CALL, DL, NodeTys, Ops);
3098 Glue = Chain.getValue(1);
3100 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue);
3104 AArch64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
3105 SelectionDAG &DAG) const {
3106 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
3107 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
3108 "ELF TLS only supported in small memory model");
3109 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3111 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
3114 EVT PtrVT = getPointerTy();
3116 const GlobalValue *GV = GA->getGlobal();
3118 SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);
3120 if (Model == TLSModel::LocalExec) {
3121 SDValue HiVar = DAG.getTargetGlobalAddress(
3122 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_G1);
3123 SDValue LoVar = DAG.getTargetGlobalAddress(
3125 AArch64II::MO_TLS | AArch64II::MO_G0 | AArch64II::MO_NC);
3127 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVZXi, DL, PtrVT, HiVar,
3128 DAG.getTargetConstant(16, MVT::i32)),
3130 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKXi, DL, PtrVT, TPOff, LoVar,
3131 DAG.getTargetConstant(0, MVT::i32)),
3133 } else if (Model == TLSModel::InitialExec) {
3134 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3135 TPOff = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TPOff);
3136 } else if (Model == TLSModel::LocalDynamic) {
3137 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
3138 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
3139 // the beginning of the module's TLS region, followed by a DTPREL offset
3142 // These accesses will need deduplicating if there's more than one.
3143 AArch64FunctionInfo *MFI =
3144 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
3145 MFI->incNumLocalDynamicTLSAccesses();
3147 // Accesses used in this sequence go via the TLS descriptor which lives in
3148 // the GOT. Prepare an address we can use to handle this.
3149 SDValue HiDesc = DAG.getTargetExternalSymbol(
3150 "_TLS_MODULE_BASE_", PtrVT, AArch64II::MO_TLS | AArch64II::MO_PAGE);
3151 SDValue LoDesc = DAG.getTargetExternalSymbol(
3152 "_TLS_MODULE_BASE_", PtrVT,
3153 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3155 // First argument to the descriptor call is the address of the descriptor
3157 SDValue DescAddr = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, HiDesc);
3158 DescAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
3160 // The call needs a relocation too for linker relaxation. It doesn't make
3161 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3163 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
3166 // Now we can calculate the offset from TPIDR_EL0 to this module's
3167 // thread-local area.
3168 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
3170 // Now use :dtprel_whatever: operations to calculate this variable's offset
3171 // in its thread-storage area.
3172 SDValue HiVar = DAG.getTargetGlobalAddress(
3173 GV, DL, MVT::i64, 0, AArch64II::MO_TLS | AArch64II::MO_G1);
3174 SDValue LoVar = DAG.getTargetGlobalAddress(
3175 GV, DL, MVT::i64, 0,
3176 AArch64II::MO_TLS | AArch64II::MO_G0 | AArch64II::MO_NC);
3179 SDValue(DAG.getMachineNode(AArch64::MOVZXi, DL, PtrVT, HiVar,
3180 DAG.getTargetConstant(16, MVT::i32)),
3183 SDValue(DAG.getMachineNode(AArch64::MOVKXi, DL, PtrVT, DTPOff, LoVar,
3184 DAG.getTargetConstant(0, MVT::i32)),
3187 TPOff = DAG.getNode(ISD::ADD, DL, PtrVT, TPOff, DTPOff);
3188 } else if (Model == TLSModel::GeneralDynamic) {
3189 // Accesses used in this sequence go via the TLS descriptor which lives in
3190 // the GOT. Prepare an address we can use to handle this.
3191 SDValue HiDesc = DAG.getTargetGlobalAddress(
3192 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_PAGE);
3193 SDValue LoDesc = DAG.getTargetGlobalAddress(
3195 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3197 // First argument to the descriptor call is the address of the descriptor
3199 SDValue DescAddr = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, HiDesc);
3200 DescAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
3202 // The call needs a relocation too for linker relaxation. It doesn't make
3203 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3206 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3208 // Finally we can make a call to calculate the offset from tpidr_el0.
3209 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
3211 llvm_unreachable("Unsupported ELF TLS access model");
3213 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
3216 SDValue AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
3217 SelectionDAG &DAG) const {
3218 if (Subtarget->isTargetDarwin())
3219 return LowerDarwinGlobalTLSAddress(Op, DAG);
3220 else if (Subtarget->isTargetELF())
3221 return LowerELFGlobalTLSAddress(Op, DAG);
3223 llvm_unreachable("Unexpected platform trying to use TLS");
3225 SDValue AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3226 SDValue Chain = Op.getOperand(0);
3227 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3228 SDValue LHS = Op.getOperand(2);
3229 SDValue RHS = Op.getOperand(3);
3230 SDValue Dest = Op.getOperand(4);
3233 // Handle f128 first, since lowering it will result in comparing the return
3234 // value of a libcall against zero, which is just what the rest of LowerBR_CC
3235 // is expecting to deal with.
3236 if (LHS.getValueType() == MVT::f128) {
3237 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3239 // If softenSetCCOperands returned a scalar, we need to compare the result
3240 // against zero to select between true and false values.
3241 if (!RHS.getNode()) {
3242 RHS = DAG.getConstant(0, LHS.getValueType());
3247 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
3249 unsigned Opc = LHS.getOpcode();
3250 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
3251 cast<ConstantSDNode>(RHS)->isOne() &&
3252 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3253 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3254 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
3255 "Unexpected condition code.");
3256 // Only lower legal XALUO ops.
3257 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
3260 // The actual operation with overflow check.
3261 AArch64CC::CondCode OFCC;
3262 SDValue Value, Overflow;
3263 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, LHS.getValue(0), DAG);
3265 if (CC == ISD::SETNE)
3266 OFCC = getInvertedCondCode(OFCC);
3267 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3269 return DAG.getNode(AArch64ISD::BRCOND, SDLoc(LHS), MVT::Other, Chain, Dest,
3273 if (LHS.getValueType().isInteger()) {
3274 assert((LHS.getValueType() == RHS.getValueType()) &&
3275 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3277 // If the RHS of the comparison is zero, we can potentially fold this
3278 // to a specialized branch.
3279 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
3280 if (RHSC && RHSC->getZExtValue() == 0) {
3281 if (CC == ISD::SETEQ) {
3282 // See if we can use a TBZ to fold in an AND as well.
3283 // TBZ has a smaller branch displacement than CBZ. If the offset is
3284 // out of bounds, a late MI-layer pass rewrites branches.
3285 // 403.gcc is an example that hits this case.
3286 if (LHS.getOpcode() == ISD::AND &&
3287 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3288 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3289 SDValue Test = LHS.getOperand(0);
3290 uint64_t Mask = LHS.getConstantOperandVal(1);
3291 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, Test,
3292 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3295 return DAG.getNode(AArch64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
3296 } else if (CC == ISD::SETNE) {
3297 // See if we can use a TBZ to fold in an AND as well.
3298 // TBZ has a smaller branch displacement than CBZ. If the offset is
3299 // out of bounds, a late MI-layer pass rewrites branches.
3300 // 403.gcc is an example that hits this case.
3301 if (LHS.getOpcode() == ISD::AND &&
3302 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3303 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3304 SDValue Test = LHS.getOperand(0);
3305 uint64_t Mask = LHS.getConstantOperandVal(1);
3306 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, Test,
3307 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3310 return DAG.getNode(AArch64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
3311 } else if (CC == ISD::SETLT && LHS.getOpcode() != ISD::AND) {
3312 // Don't combine AND since emitComparison converts the AND to an ANDS
3313 // (a.k.a. TST) and the test in the test bit and branch instruction
3314 // becomes redundant. This would also increase register pressure.
3315 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3316 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, LHS,
3317 DAG.getConstant(Mask, MVT::i64), Dest);
3320 if (RHSC && RHSC->getSExtValue() == -1 && CC == ISD::SETGT &&
3321 LHS.getOpcode() != ISD::AND) {
3322 // Don't combine AND since emitComparison converts the AND to an ANDS
3323 // (a.k.a. TST) and the test in the test bit and branch instruction
3324 // becomes redundant. This would also increase register pressure.
3325 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3326 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, LHS,
3327 DAG.getConstant(Mask, MVT::i64), Dest);
3331 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3332 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3336 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3338 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3339 // clean. Some of them require two branches to implement.
3340 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3341 AArch64CC::CondCode CC1, CC2;
3342 changeFPCCToAArch64CC(CC, CC1, CC2);
3343 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3345 DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
3346 if (CC2 != AArch64CC::AL) {
3347 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3348 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
3355 SDValue AArch64TargetLowering::LowerFCOPYSIGN(SDValue Op,
3356 SelectionDAG &DAG) const {
3357 EVT VT = Op.getValueType();
3360 SDValue In1 = Op.getOperand(0);
3361 SDValue In2 = Op.getOperand(1);
3362 EVT SrcVT = In2.getValueType();
3364 if (SrcVT == MVT::f32 && VT == MVT::f64)
3365 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
3366 else if (SrcVT == MVT::f64 && VT == MVT::f32)
3367 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0));
3369 // FIXME: Src type is different, bail out for now. Can VT really be a
3376 SDValue EltMask, VecVal1, VecVal2;
3377 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
3380 EltMask = DAG.getConstant(0x80000000ULL, EltVT);
3382 if (!VT.isVector()) {
3383 VecVal1 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3384 DAG.getUNDEF(VecVT), In1);
3385 VecVal2 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3386 DAG.getUNDEF(VecVT), In2);
3388 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3389 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3391 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
3395 // We want to materialize a mask with the the high bit set, but the AdvSIMD
3396 // immediate moves cannot materialize that in a single instruction for
3397 // 64-bit elements. Instead, materialize zero and then negate it.
3398 EltMask = DAG.getConstant(0, EltVT);
3400 if (!VT.isVector()) {
3401 VecVal1 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3402 DAG.getUNDEF(VecVT), In1);
3403 VecVal2 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3404 DAG.getUNDEF(VecVT), In2);
3406 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3407 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3410 llvm_unreachable("Invalid type for copysign!");
3413 std::vector<SDValue> BuildVectorOps;
3414 for (unsigned i = 0; i < VecVT.getVectorNumElements(); ++i)
3415 BuildVectorOps.push_back(EltMask);
3417 SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT, BuildVectorOps);
3419 // If we couldn't materialize the mask above, then the mask vector will be
3420 // the zero vector, and we need to negate it here.
3421 if (VT == MVT::f64 || VT == MVT::v2f64) {
3422 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
3423 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
3424 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
3428 DAG.getNode(AArch64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
3431 return DAG.getTargetExtractSubreg(AArch64::ssub, DL, VT, Sel);
3432 else if (VT == MVT::f64)
3433 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, VT, Sel);
3435 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
3438 SDValue AArch64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
3439 if (DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
3440 AttributeSet::FunctionIndex, Attribute::NoImplicitFloat))
3443 if (!Subtarget->hasNEON())
3446 // While there is no integer popcount instruction, it can
3447 // be more efficiently lowered to the following sequence that uses
3448 // AdvSIMD registers/instructions as long as the copies to/from
3449 // the AdvSIMD registers are cheap.
3450 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
3451 // CNT V0.8B, V0.8B // 8xbyte pop-counts
3452 // ADDV B0, V0.8B // sum 8xbyte pop-counts
3453 // UMOV X0, V0.B[0] // copy byte result back to integer reg
3454 SDValue Val = Op.getOperand(0);
3456 EVT VT = Op.getValueType();
3457 SDValue ZeroVec = DAG.getUNDEF(MVT::v8i8);
3460 if (VT == MVT::i32) {
3461 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Val);
3462 VecVal = DAG.getTargetInsertSubreg(AArch64::ssub, DL, MVT::v8i8, ZeroVec,
3465 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
3468 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, VecVal);
3469 SDValue UaddLV = DAG.getNode(
3470 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
3471 DAG.getConstant(Intrinsic::aarch64_neon_uaddlv, MVT::i32), CtPop);
3474 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
3478 SDValue AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3480 if (Op.getValueType().isVector())
3481 return LowerVSETCC(Op, DAG);
3483 SDValue LHS = Op.getOperand(0);
3484 SDValue RHS = Op.getOperand(1);
3485 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3488 // We chose ZeroOrOneBooleanContents, so use zero and one.
3489 EVT VT = Op.getValueType();
3490 SDValue TVal = DAG.getConstant(1, VT);
3491 SDValue FVal = DAG.getConstant(0, VT);
3493 // Handle f128 first, since one possible outcome is a normal integer
3494 // comparison which gets picked up by the next if statement.
3495 if (LHS.getValueType() == MVT::f128) {
3496 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3498 // If softenSetCCOperands returned a scalar, use it.
3499 if (!RHS.getNode()) {
3500 assert(LHS.getValueType() == Op.getValueType() &&
3501 "Unexpected setcc expansion!");
3506 if (LHS.getValueType().isInteger()) {
3509 getAArch64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
3511 // Note that we inverted the condition above, so we reverse the order of
3512 // the true and false operands here. This will allow the setcc to be
3513 // matched to a single CSINC instruction.
3514 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
3517 // Now we know we're dealing with FP values.
3518 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3520 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3521 // and do the comparison.
3522 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3524 AArch64CC::CondCode CC1, CC2;
3525 changeFPCCToAArch64CC(CC, CC1, CC2);
3526 if (CC2 == AArch64CC::AL) {
3527 changeFPCCToAArch64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
3528 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3530 // Note that we inverted the condition above, so we reverse the order of
3531 // the true and false operands here. This will allow the setcc to be
3532 // matched to a single CSINC instruction.
3533 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
3535 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't
3536 // totally clean. Some of them require two CSELs to implement. As is in
3537 // this case, we emit the first CSEL and then emit a second using the output
3538 // of the first as the RHS. We're effectively OR'ing the two CC's together.
3540 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
3541 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3543 DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3545 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3546 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3550 /// A SELECT_CC operation is really some kind of max or min if both values being
3551 /// compared are, in some sense, equal to the results in either case. However,
3552 /// it is permissible to compare f32 values and produce directly extended f64
3555 /// Extending the comparison operands would also be allowed, but is less likely
3556 /// to happen in practice since their use is right here. Note that truncate
3557 /// operations would *not* be semantically equivalent.
3558 static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) {
3562 ConstantFPSDNode *CCmp = dyn_cast<ConstantFPSDNode>(Cmp);
3563 ConstantFPSDNode *CResult = dyn_cast<ConstantFPSDNode>(Result);
3564 if (CCmp && CResult && Cmp.getValueType() == MVT::f32 &&
3565 Result.getValueType() == MVT::f64) {
3567 APFloat CmpVal = CCmp->getValueAPF();
3568 CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy);
3569 return CResult->getValueAPF().bitwiseIsEqual(CmpVal);
3572 return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp;
3575 SDValue AArch64TargetLowering::LowerSELECT(SDValue Op,
3576 SelectionDAG &DAG) const {
3577 SDValue CC = Op->getOperand(0);
3578 SDValue TVal = Op->getOperand(1);
3579 SDValue FVal = Op->getOperand(2);
3582 unsigned Opc = CC.getOpcode();
3583 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
3585 if (CC.getResNo() == 1 &&
3586 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3587 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3588 // Only lower legal XALUO ops.
3589 if (!DAG.getTargetLoweringInfo().isTypeLegal(CC->getValueType(0)))
3592 AArch64CC::CondCode OFCC;
3593 SDValue Value, Overflow;
3594 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, CC.getValue(0), DAG);
3595 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3597 return DAG.getNode(AArch64ISD::CSEL, DL, Op.getValueType(), TVal, FVal,
3601 if (CC.getOpcode() == ISD::SETCC)
3602 return DAG.getSelectCC(DL, CC.getOperand(0), CC.getOperand(1), TVal, FVal,
3603 cast<CondCodeSDNode>(CC.getOperand(2))->get());
3605 return DAG.getSelectCC(DL, CC, DAG.getConstant(0, CC.getValueType()), TVal,
3609 SDValue AArch64TargetLowering::LowerSELECT_CC(SDValue Op,
3610 SelectionDAG &DAG) const {
3611 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3612 SDValue LHS = Op.getOperand(0);
3613 SDValue RHS = Op.getOperand(1);
3614 SDValue TVal = Op.getOperand(2);
3615 SDValue FVal = Op.getOperand(3);
3618 // Handle f128 first, because it will result in a comparison of some RTLIB
3619 // call result against zero.
3620 if (LHS.getValueType() == MVT::f128) {
3621 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3623 // If softenSetCCOperands returned a scalar, we need to compare the result
3624 // against zero to select between true and false values.
3625 if (!RHS.getNode()) {
3626 RHS = DAG.getConstant(0, LHS.getValueType());
3631 // Handle integers first.
3632 if (LHS.getValueType().isInteger()) {
3633 assert((LHS.getValueType() == RHS.getValueType()) &&
3634 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3636 unsigned Opcode = AArch64ISD::CSEL;
3638 // If both the TVal and the FVal are constants, see if we can swap them in
3639 // order to for a CSINV or CSINC out of them.
3640 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3641 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3643 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3644 std::swap(TVal, FVal);
3645 std::swap(CTVal, CFVal);
3646 CC = ISD::getSetCCInverse(CC, true);
3647 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3648 std::swap(TVal, FVal);
3649 std::swap(CTVal, CFVal);
3650 CC = ISD::getSetCCInverse(CC, true);
3651 } else if (TVal.getOpcode() == ISD::XOR) {
3652 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3653 // with a CSINV rather than a CSEL.
3654 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
3656 if (CVal && CVal->isAllOnesValue()) {
3657 std::swap(TVal, FVal);
3658 std::swap(CTVal, CFVal);
3659 CC = ISD::getSetCCInverse(CC, true);
3661 } else if (TVal.getOpcode() == ISD::SUB) {
3662 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3663 // that we can match with a CSNEG rather than a CSEL.
3664 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
3666 if (CVal && CVal->isNullValue()) {
3667 std::swap(TVal, FVal);
3668 std::swap(CTVal, CFVal);
3669 CC = ISD::getSetCCInverse(CC, true);
3671 } else if (CTVal && CFVal) {
3672 const int64_t TrueVal = CTVal->getSExtValue();
3673 const int64_t FalseVal = CFVal->getSExtValue();
3676 // If both TVal and FVal are constants, see if FVal is the
3677 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3678 // instead of a CSEL in that case.
3679 if (TrueVal == ~FalseVal) {
3680 Opcode = AArch64ISD::CSINV;
3681 } else if (TrueVal == -FalseVal) {
3682 Opcode = AArch64ISD::CSNEG;
3683 } else if (TVal.getValueType() == MVT::i32) {
3684 // If our operands are only 32-bit wide, make sure we use 32-bit
3685 // arithmetic for the check whether we can use CSINC. This ensures that
3686 // the addition in the check will wrap around properly in case there is
3687 // an overflow (which would not be the case if we do the check with
3688 // 64-bit arithmetic).
3689 const uint32_t TrueVal32 = CTVal->getZExtValue();
3690 const uint32_t FalseVal32 = CFVal->getZExtValue();
3692 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3693 Opcode = AArch64ISD::CSINC;
3695 if (TrueVal32 > FalseVal32) {
3699 // 64-bit check whether we can use CSINC.
3700 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3701 Opcode = AArch64ISD::CSINC;
3703 if (TrueVal > FalseVal) {
3708 // Swap TVal and FVal if necessary.
3710 std::swap(TVal, FVal);
3711 std::swap(CTVal, CFVal);
3712 CC = ISD::getSetCCInverse(CC, true);
3715 if (Opcode != AArch64ISD::CSEL) {
3716 // Drop FVal since we can get its value by simply inverting/negating
3723 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3725 EVT VT = Op.getValueType();
3726 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3729 // Now we know we're dealing with FP values.
3730 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3731 assert(LHS.getValueType() == RHS.getValueType());
3732 EVT VT = Op.getValueType();
3734 // Try to match this select into a max/min operation, which have dedicated
3735 // opcode in the instruction set.
3736 // FIXME: This is not correct in the presence of NaNs, so we only enable this
3738 if (getTargetMachine().Options.NoNaNsFPMath) {
3739 SDValue MinMaxLHS = TVal, MinMaxRHS = FVal;
3740 if (selectCCOpsAreFMaxCompatible(LHS, MinMaxRHS) &&
3741 selectCCOpsAreFMaxCompatible(RHS, MinMaxLHS)) {
3742 CC = ISD::getSetCCSwappedOperands(CC);
3743 std::swap(MinMaxLHS, MinMaxRHS);
3746 if (selectCCOpsAreFMaxCompatible(LHS, MinMaxLHS) &&
3747 selectCCOpsAreFMaxCompatible(RHS, MinMaxRHS)) {
3757 return DAG.getNode(AArch64ISD::FMAX, dl, VT, MinMaxLHS, MinMaxRHS);
3765 return DAG.getNode(AArch64ISD::FMIN, dl, VT, MinMaxLHS, MinMaxRHS);
3771 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3772 // and do the comparison.
3773 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3775 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3776 // clean. Some of them require two CSELs to implement.
3777 AArch64CC::CondCode CC1, CC2;
3778 changeFPCCToAArch64CC(CC, CC1, CC2);
3779 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3780 SDValue CS1 = DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3782 // If we need a second CSEL, emit it, using the output of the first as the
3783 // RHS. We're effectively OR'ing the two CC's together.
3784 if (CC2 != AArch64CC::AL) {
3785 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3786 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3789 // Otherwise, return the output of the first CSEL.
3793 SDValue AArch64TargetLowering::LowerJumpTable(SDValue Op,
3794 SelectionDAG &DAG) const {
3795 // Jump table entries as PC relative offsets. No additional tweaking
3796 // is necessary here. Just get the address of the jump table.
3797 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3798 EVT PtrVT = getPointerTy();
3801 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3802 !Subtarget->isTargetMachO()) {
3803 const unsigned char MO_NC = AArch64II::MO_NC;
3805 AArch64ISD::WrapperLarge, DL, PtrVT,
3806 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G3),
3807 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G2 | MO_NC),
3808 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G1 | MO_NC),
3809 DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3810 AArch64II::MO_G0 | MO_NC));
3814 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_PAGE);
3815 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3816 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3817 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3818 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3821 SDValue AArch64TargetLowering::LowerConstantPool(SDValue Op,
3822 SelectionDAG &DAG) const {
3823 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3824 EVT PtrVT = getPointerTy();
3827 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3828 // Use the GOT for the large code model on iOS.
3829 if (Subtarget->isTargetMachO()) {
3830 SDValue GotAddr = DAG.getTargetConstantPool(
3831 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3833 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
3836 const unsigned char MO_NC = AArch64II::MO_NC;
3838 AArch64ISD::WrapperLarge, DL, PtrVT,
3839 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3840 CP->getOffset(), AArch64II::MO_G3),
3841 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3842 CP->getOffset(), AArch64II::MO_G2 | MO_NC),
3843 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3844 CP->getOffset(), AArch64II::MO_G1 | MO_NC),
3845 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3846 CP->getOffset(), AArch64II::MO_G0 | MO_NC));
3848 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
3849 // ELF, the only valid one on Darwin.
3851 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3852 CP->getOffset(), AArch64II::MO_PAGE);
3853 SDValue Lo = DAG.getTargetConstantPool(
3854 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3855 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3857 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3858 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3862 SDValue AArch64TargetLowering::LowerBlockAddress(SDValue Op,
3863 SelectionDAG &DAG) const {
3864 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
3865 EVT PtrVT = getPointerTy();
3867 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3868 !Subtarget->isTargetMachO()) {
3869 const unsigned char MO_NC = AArch64II::MO_NC;
3871 AArch64ISD::WrapperLarge, DL, PtrVT,
3872 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G3),
3873 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
3874 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
3875 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
3877 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGE);
3878 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGEOFF |
3880 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3881 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3885 SDValue AArch64TargetLowering::LowerDarwin_VASTART(SDValue Op,
3886 SelectionDAG &DAG) const {
3887 AArch64FunctionInfo *FuncInfo =
3888 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
3892 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3893 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3894 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
3895 MachinePointerInfo(SV), false, false, 0);
3898 SDValue AArch64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
3899 SelectionDAG &DAG) const {
3900 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3901 // Standard, section B.3.
3902 MachineFunction &MF = DAG.getMachineFunction();
3903 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
3906 SDValue Chain = Op.getOperand(0);
3907 SDValue VAList = Op.getOperand(1);
3908 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3909 SmallVector<SDValue, 4> MemOps;
3911 // void *__stack at offset 0
3913 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3914 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3915 MachinePointerInfo(SV), false, false, 8));
3917 // void *__gr_top at offset 8
3918 int GPRSize = FuncInfo->getVarArgsGPRSize();
3920 SDValue GRTop, GRTopAddr;
3922 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3923 DAG.getConstant(8, getPointerTy()));
3925 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), getPointerTy());
3926 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
3927 DAG.getConstant(GPRSize, getPointerTy()));
3929 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3930 MachinePointerInfo(SV, 8), false, false, 8));
3933 // void *__vr_top at offset 16
3934 int FPRSize = FuncInfo->getVarArgsFPRSize();
3936 SDValue VRTop, VRTopAddr;
3937 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3938 DAG.getConstant(16, getPointerTy()));
3940 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), getPointerTy());
3941 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
3942 DAG.getConstant(FPRSize, getPointerTy()));
3944 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3945 MachinePointerInfo(SV, 16), false, false, 8));
3948 // int __gr_offs at offset 24
3949 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3950 DAG.getConstant(24, getPointerTy()));
3951 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
3952 GROffsAddr, MachinePointerInfo(SV, 24), false,
3955 // int __vr_offs at offset 28
3956 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3957 DAG.getConstant(28, getPointerTy()));
3958 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
3959 VROffsAddr, MachinePointerInfo(SV, 28), false,
3962 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
3965 SDValue AArch64TargetLowering::LowerVASTART(SDValue Op,
3966 SelectionDAG &DAG) const {
3967 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
3968 : LowerAAPCS_VASTART(Op, DAG);
3971 SDValue AArch64TargetLowering::LowerVACOPY(SDValue Op,
3972 SelectionDAG &DAG) const {
3973 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
3975 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
3976 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3977 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
3979 return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op), Op.getOperand(1),
3980 Op.getOperand(2), DAG.getConstant(VaListSize, MVT::i32),
3981 8, false, false, MachinePointerInfo(DestSV),
3982 MachinePointerInfo(SrcSV));
3985 SDValue AArch64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
3986 assert(Subtarget->isTargetDarwin() &&
3987 "automatic va_arg instruction only works on Darwin");
3989 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3990 EVT VT = Op.getValueType();
3992 SDValue Chain = Op.getOperand(0);
3993 SDValue Addr = Op.getOperand(1);
3994 unsigned Align = Op.getConstantOperandVal(3);
3996 SDValue VAList = DAG.getLoad(getPointerTy(), DL, Chain, Addr,
3997 MachinePointerInfo(V), false, false, false, 0);
3998 Chain = VAList.getValue(1);
4001 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
4002 VAList = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
4003 DAG.getConstant(Align - 1, getPointerTy()));
4004 VAList = DAG.getNode(ISD::AND, DL, getPointerTy(), VAList,
4005 DAG.getConstant(-(int64_t)Align, getPointerTy()));
4008 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
4009 uint64_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
4011 // Scalar integer and FP values smaller than 64 bits are implicitly extended
4012 // up to 64 bits. At the very least, we have to increase the striding of the
4013 // vaargs list to match this, and for FP values we need to introduce
4014 // FP_ROUND nodes as well.
4015 if (VT.isInteger() && !VT.isVector())
4017 bool NeedFPTrunc = false;
4018 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
4023 // Increment the pointer, VAList, to the next vaarg
4024 SDValue VANext = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
4025 DAG.getConstant(ArgSize, getPointerTy()));
4026 // Store the incremented VAList to the legalized pointer
4027 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
4030 // Load the actual argument out of the pointer VAList
4032 // Load the value as an f64.
4033 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
4034 MachinePointerInfo(), false, false, false, 0);
4035 // Round the value down to an f32.
4036 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
4037 DAG.getIntPtrConstant(1));
4038 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
4039 // Merge the rounded value with the chain output of the load.
4040 return DAG.getMergeValues(Ops, DL);
4043 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
4047 SDValue AArch64TargetLowering::LowerFRAMEADDR(SDValue Op,
4048 SelectionDAG &DAG) const {
4049 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4050 MFI->setFrameAddressIsTaken(true);
4052 EVT VT = Op.getValueType();
4054 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4056 DAG.getCopyFromReg(DAG.getEntryNode(), DL, AArch64::FP, VT);
4058 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
4059 MachinePointerInfo(), false, false, false, 0);
4063 // FIXME? Maybe this could be a TableGen attribute on some registers and
4064 // this table could be generated automatically from RegInfo.
4065 unsigned AArch64TargetLowering::getRegisterByName(const char* RegName,
4067 unsigned Reg = StringSwitch<unsigned>(RegName)
4068 .Case("sp", AArch64::SP)
4072 report_fatal_error("Invalid register name global variable");
4075 SDValue AArch64TargetLowering::LowerRETURNADDR(SDValue Op,
4076 SelectionDAG &DAG) const {
4077 MachineFunction &MF = DAG.getMachineFunction();
4078 MachineFrameInfo *MFI = MF.getFrameInfo();
4079 MFI->setReturnAddressIsTaken(true);
4081 EVT VT = Op.getValueType();
4083 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4085 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4086 SDValue Offset = DAG.getConstant(8, getPointerTy());
4087 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
4088 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
4089 MachinePointerInfo(), false, false, false, 0);
4092 // Return LR, which contains the return address. Mark it an implicit live-in.
4093 unsigned Reg = MF.addLiveIn(AArch64::LR, &AArch64::GPR64RegClass);
4094 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
4097 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4098 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4099 SDValue AArch64TargetLowering::LowerShiftRightParts(SDValue Op,
4100 SelectionDAG &DAG) const {
4101 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4102 EVT VT = Op.getValueType();
4103 unsigned VTBits = VT.getSizeInBits();
4105 SDValue ShOpLo = Op.getOperand(0);
4106 SDValue ShOpHi = Op.getOperand(1);
4107 SDValue ShAmt = Op.getOperand(2);
4109 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4111 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4113 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4114 DAG.getConstant(VTBits, MVT::i64), ShAmt);
4115 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4116 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4117 DAG.getConstant(VTBits, MVT::i64));
4118 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4120 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
4121 ISD::SETGE, dl, DAG);
4122 SDValue CCVal = DAG.getConstant(AArch64CC::GE, MVT::i32);
4124 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4125 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4127 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4129 // AArch64 shifts larger than the register width are wrapped rather than
4130 // clamped, so we can't just emit "hi >> x".
4131 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4132 SDValue TrueValHi = Opc == ISD::SRA
4133 ? DAG.getNode(Opc, dl, VT, ShOpHi,
4134 DAG.getConstant(VTBits - 1, MVT::i64))
4135 : DAG.getConstant(0, VT);
4137 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
4139 SDValue Ops[2] = { Lo, Hi };
4140 return DAG.getMergeValues(Ops, dl);
4143 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4144 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4145 SDValue AArch64TargetLowering::LowerShiftLeftParts(SDValue Op,
4146 SelectionDAG &DAG) const {
4147 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4148 EVT VT = Op.getValueType();
4149 unsigned VTBits = VT.getSizeInBits();
4151 SDValue ShOpLo = Op.getOperand(0);
4152 SDValue ShOpHi = Op.getOperand(1);
4153 SDValue ShAmt = Op.getOperand(2);
4156 assert(Op.getOpcode() == ISD::SHL_PARTS);
4157 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4158 DAG.getConstant(VTBits, MVT::i64), ShAmt);
4159 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4160 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4161 DAG.getConstant(VTBits, MVT::i64));
4162 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4163 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4165 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4167 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
4168 ISD::SETGE, dl, DAG);
4169 SDValue CCVal = DAG.getConstant(AArch64CC::GE, MVT::i32);
4171 DAG.getNode(AArch64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
4173 // AArch64 shifts of larger than register sizes are wrapped rather than
4174 // clamped, so we can't just emit "lo << a" if a is too big.
4175 SDValue TrueValLo = DAG.getConstant(0, VT);
4176 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4178 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4180 SDValue Ops[2] = { Lo, Hi };
4181 return DAG.getMergeValues(Ops, dl);
4184 bool AArch64TargetLowering::isOffsetFoldingLegal(
4185 const GlobalAddressSDNode *GA) const {
4186 // The AArch64 target doesn't support folding offsets into global addresses.
4190 bool AArch64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4191 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
4192 // FIXME: We should be able to handle f128 as well with a clever lowering.
4193 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
4197 return AArch64_AM::getFP64Imm(Imm) != -1;
4198 else if (VT == MVT::f32)
4199 return AArch64_AM::getFP32Imm(Imm) != -1;
4203 //===----------------------------------------------------------------------===//
4204 // AArch64 Optimization Hooks
4205 //===----------------------------------------------------------------------===//
4207 //===----------------------------------------------------------------------===//
4208 // AArch64 Inline Assembly Support
4209 //===----------------------------------------------------------------------===//
4211 // Table of Constraints
4212 // TODO: This is the current set of constraints supported by ARM for the
4213 // compiler, not all of them may make sense, e.g. S may be difficult to support.
4215 // r - A general register
4216 // w - An FP/SIMD register of some size in the range v0-v31
4217 // x - An FP/SIMD register of some size in the range v0-v15
4218 // I - Constant that can be used with an ADD instruction
4219 // J - Constant that can be used with a SUB instruction
4220 // K - Constant that can be used with a 32-bit logical instruction
4221 // L - Constant that can be used with a 64-bit logical instruction
4222 // M - Constant that can be used as a 32-bit MOV immediate
4223 // N - Constant that can be used as a 64-bit MOV immediate
4224 // Q - A memory reference with base register and no offset
4225 // S - A symbolic address
4226 // Y - Floating point constant zero
4227 // Z - Integer constant zero
4229 // Note that general register operands will be output using their 64-bit x
4230 // register name, whatever the size of the variable, unless the asm operand
4231 // is prefixed by the %w modifier. Floating-point and SIMD register operands
4232 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
4235 /// getConstraintType - Given a constraint letter, return the type of
4236 /// constraint it is for this target.
4237 AArch64TargetLowering::ConstraintType
4238 AArch64TargetLowering::getConstraintType(const std::string &Constraint) const {
4239 if (Constraint.size() == 1) {
4240 switch (Constraint[0]) {
4247 return C_RegisterClass;
4248 // An address with a single base register. Due to the way we
4249 // currently handle addresses it is the same as 'r'.
4254 return TargetLowering::getConstraintType(Constraint);
4257 /// Examine constraint type and operand type and determine a weight value.
4258 /// This object must already have been set up with the operand type
4259 /// and the current alternative constraint selected.
4260 TargetLowering::ConstraintWeight
4261 AArch64TargetLowering::getSingleConstraintMatchWeight(
4262 AsmOperandInfo &info, const char *constraint) const {
4263 ConstraintWeight weight = CW_Invalid;
4264 Value *CallOperandVal = info.CallOperandVal;
4265 // If we don't have a value, we can't do a match,
4266 // but allow it at the lowest weight.
4267 if (!CallOperandVal)
4269 Type *type = CallOperandVal->getType();
4270 // Look at the constraint type.
4271 switch (*constraint) {
4273 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
4277 if (type->isFloatingPointTy() || type->isVectorTy())
4278 weight = CW_Register;
4281 weight = CW_Constant;
4287 std::pair<unsigned, const TargetRegisterClass *>
4288 AArch64TargetLowering::getRegForInlineAsmConstraint(
4289 const std::string &Constraint, MVT VT) const {
4290 if (Constraint.size() == 1) {
4291 switch (Constraint[0]) {
4293 if (VT.getSizeInBits() == 64)
4294 return std::make_pair(0U, &AArch64::GPR64commonRegClass);
4295 return std::make_pair(0U, &AArch64::GPR32commonRegClass);
4298 return std::make_pair(0U, &AArch64::FPR32RegClass);
4299 if (VT.getSizeInBits() == 64)
4300 return std::make_pair(0U, &AArch64::FPR64RegClass);
4301 if (VT.getSizeInBits() == 128)
4302 return std::make_pair(0U, &AArch64::FPR128RegClass);
4304 // The instructions that this constraint is designed for can
4305 // only take 128-bit registers so just use that regclass.
4307 if (VT.getSizeInBits() == 128)
4308 return std::make_pair(0U, &AArch64::FPR128_loRegClass);
4312 if (StringRef("{cc}").equals_lower(Constraint))
4313 return std::make_pair(unsigned(AArch64::NZCV), &AArch64::CCRRegClass);
4315 // Use the default implementation in TargetLowering to convert the register
4316 // constraint into a member of a register class.
4317 std::pair<unsigned, const TargetRegisterClass *> Res;
4318 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4320 // Not found as a standard register?
4322 unsigned Size = Constraint.size();
4323 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
4324 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
4325 const std::string Reg =
4326 std::string(&Constraint[2], &Constraint[Size - 1]);
4327 int RegNo = atoi(Reg.c_str());
4328 if (RegNo >= 0 && RegNo <= 31) {
4329 // v0 - v31 are aliases of q0 - q31.
4330 // By default we'll emit v0-v31 for this unless there's a modifier where
4331 // we'll emit the correct register as well.
4332 Res.first = AArch64::FPR128RegClass.getRegister(RegNo);
4333 Res.second = &AArch64::FPR128RegClass;
4341 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4342 /// vector. If it is invalid, don't add anything to Ops.
4343 void AArch64TargetLowering::LowerAsmOperandForConstraint(
4344 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
4345 SelectionDAG &DAG) const {
4348 // Currently only support length 1 constraints.
4349 if (Constraint.length() != 1)
4352 char ConstraintLetter = Constraint[0];
4353 switch (ConstraintLetter) {
4357 // This set of constraints deal with valid constants for various instructions.
4358 // Validate and return a target constant for them if we can.
4360 // 'z' maps to xzr or wzr so it needs an input of 0.
4361 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4362 if (!C || C->getZExtValue() != 0)
4365 if (Op.getValueType() == MVT::i64)
4366 Result = DAG.getRegister(AArch64::XZR, MVT::i64);
4368 Result = DAG.getRegister(AArch64::WZR, MVT::i32);
4378 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4382 // Grab the value and do some validation.
4383 uint64_t CVal = C->getZExtValue();
4384 switch (ConstraintLetter) {
4385 // The I constraint applies only to simple ADD or SUB immediate operands:
4386 // i.e. 0 to 4095 with optional shift by 12
4387 // The J constraint applies only to ADD or SUB immediates that would be
4388 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
4389 // instruction [or vice versa], in other words -1 to -4095 with optional
4390 // left shift by 12.
4392 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
4396 uint64_t NVal = -C->getSExtValue();
4397 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal)) {
4398 CVal = C->getSExtValue();
4403 // The K and L constraints apply *only* to logical immediates, including
4404 // what used to be the MOVI alias for ORR (though the MOVI alias has now
4405 // been removed and MOV should be used). So these constraints have to
4406 // distinguish between bit patterns that are valid 32-bit or 64-bit
4407 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
4408 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
4411 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4415 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4418 // The M and N constraints are a superset of K and L respectively, for use
4419 // with the MOV (immediate) alias. As well as the logical immediates they
4420 // also match 32 or 64-bit immediates that can be loaded either using a
4421 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
4422 // (M) or 64-bit 0x1234000000000000 (N) etc.
4423 // As a note some of this code is liberally stolen from the asm parser.
4425 if (!isUInt<32>(CVal))
4427 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4429 if ((CVal & 0xFFFF) == CVal)
4431 if ((CVal & 0xFFFF0000ULL) == CVal)
4433 uint64_t NCVal = ~(uint32_t)CVal;
4434 if ((NCVal & 0xFFFFULL) == NCVal)
4436 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4441 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4443 if ((CVal & 0xFFFFULL) == CVal)
4445 if ((CVal & 0xFFFF0000ULL) == CVal)
4447 if ((CVal & 0xFFFF00000000ULL) == CVal)
4449 if ((CVal & 0xFFFF000000000000ULL) == CVal)
4451 uint64_t NCVal = ~CVal;
4452 if ((NCVal & 0xFFFFULL) == NCVal)
4454 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4456 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
4458 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
4466 // All assembler immediates are 64-bit integers.
4467 Result = DAG.getTargetConstant(CVal, MVT::i64);
4471 if (Result.getNode()) {
4472 Ops.push_back(Result);
4476 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4479 //===----------------------------------------------------------------------===//
4480 // AArch64 Advanced SIMD Support
4481 //===----------------------------------------------------------------------===//
4483 /// WidenVector - Given a value in the V64 register class, produce the
4484 /// equivalent value in the V128 register class.
4485 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
4486 EVT VT = V64Reg.getValueType();
4487 unsigned NarrowSize = VT.getVectorNumElements();
4488 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4489 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
4492 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
4493 V64Reg, DAG.getConstant(0, MVT::i32));
4496 /// getExtFactor - Determine the adjustment factor for the position when
4497 /// generating an "extract from vector registers" instruction.
4498 static unsigned getExtFactor(SDValue &V) {
4499 EVT EltType = V.getValueType().getVectorElementType();
4500 return EltType.getSizeInBits() / 8;
4503 /// NarrowVector - Given a value in the V128 register class, produce the
4504 /// equivalent value in the V64 register class.
4505 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
4506 EVT VT = V128Reg.getValueType();
4507 unsigned WideSize = VT.getVectorNumElements();
4508 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4509 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
4512 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, NarrowTy, V128Reg);
4515 // Gather data to see if the operation can be modelled as a
4516 // shuffle in combination with VEXTs.
4517 SDValue AArch64TargetLowering::ReconstructShuffle(SDValue Op,
4518 SelectionDAG &DAG) const {
4519 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
4521 EVT VT = Op.getValueType();
4522 unsigned NumElts = VT.getVectorNumElements();
4524 struct ShuffleSourceInfo {
4529 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
4530 // be compatible with the shuffle we intend to construct. As a result
4531 // ShuffleVec will be some sliding window into the original Vec.
4534 // Code should guarantee that element i in Vec starts at element "WindowBase
4535 // + i * WindowScale in ShuffleVec".
4539 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
4540 ShuffleSourceInfo(SDValue Vec)
4541 : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
4545 // First gather all vectors used as an immediate source for this BUILD_VECTOR
4547 SmallVector<ShuffleSourceInfo, 2> Sources;
4548 for (unsigned i = 0; i < NumElts; ++i) {
4549 SDValue V = Op.getOperand(i);
4550 if (V.getOpcode() == ISD::UNDEF)
4552 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4553 // A shuffle can only come from building a vector from various
4554 // elements of other vectors.
4558 // Add this element source to the list if it's not already there.
4559 SDValue SourceVec = V.getOperand(0);
4560 auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
4561 if (Source == Sources.end())
4562 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
4564 // Update the minimum and maximum lane number seen.
4565 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4566 Source->MinElt = std::min(Source->MinElt, EltNo);
4567 Source->MaxElt = std::max(Source->MaxElt, EltNo);
4570 // Currently only do something sane when at most two source vectors
4572 if (Sources.size() > 2)
4575 // Find out the smallest element size among result and two sources, and use
4576 // it as element size to build the shuffle_vector.
4577 EVT SmallestEltTy = VT.getVectorElementType();
4578 for (auto &Source : Sources) {
4579 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
4580 if (SrcEltTy.bitsLT(SmallestEltTy)) {
4581 SmallestEltTy = SrcEltTy;
4584 unsigned ResMultiplier =
4585 VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
4586 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
4587 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
4589 // If the source vector is too wide or too narrow, we may nevertheless be able
4590 // to construct a compatible shuffle either by concatenating it with UNDEF or
4591 // extracting a suitable range of elements.
4592 for (auto &Src : Sources) {
4593 EVT SrcVT = Src.ShuffleVec.getValueType();
4595 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
4598 // This stage of the search produces a source with the same element type as
4599 // the original, but with a total width matching the BUILD_VECTOR output.
4600 EVT EltVT = SrcVT.getVectorElementType();
4601 unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
4602 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
4604 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
4605 assert(2 * SrcVT.getSizeInBits() == VT.getSizeInBits());
4606 // We can pad out the smaller vector for free, so if it's part of a
4609 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
4610 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
4614 assert(SrcVT.getSizeInBits() == 2 * VT.getSizeInBits());
4616 if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
4617 // Span too large for a VEXT to cope
4621 if (Src.MinElt >= NumSrcElts) {
4622 // The extraction can just take the second half
4624 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4625 DAG.getConstant(NumSrcElts, MVT::i64));
4626 Src.WindowBase = -NumSrcElts;
4627 } else if (Src.MaxElt < NumSrcElts) {
4628 // The extraction can just take the first half
4630 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4631 DAG.getConstant(0, MVT::i64));
4633 // An actual VEXT is needed
4635 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4636 DAG.getConstant(0, MVT::i64));
4638 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4639 DAG.getConstant(NumSrcElts, MVT::i64));
4640 unsigned Imm = Src.MinElt * getExtFactor(VEXTSrc1);
4642 Src.ShuffleVec = DAG.getNode(AArch64ISD::EXT, dl, DestVT, VEXTSrc1,
4643 VEXTSrc2, DAG.getConstant(Imm, MVT::i32));
4644 Src.WindowBase = -Src.MinElt;
4648 // Another possible incompatibility occurs from the vector element types. We
4649 // can fix this by bitcasting the source vectors to the same type we intend
4651 for (auto &Src : Sources) {
4652 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
4653 if (SrcEltTy == SmallestEltTy)
4655 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
4656 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
4657 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
4658 Src.WindowBase *= Src.WindowScale;
4661 // Final sanity check before we try to actually produce a shuffle.
4663 for (auto Src : Sources)
4664 assert(Src.ShuffleVec.getValueType() == ShuffleVT);
4667 // The stars all align, our next step is to produce the mask for the shuffle.
4668 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
4669 int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
4670 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
4671 SDValue Entry = Op.getOperand(i);
4672 if (Entry.getOpcode() == ISD::UNDEF)
4675 auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
4676 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
4678 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
4679 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
4681 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
4682 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
4683 VT.getVectorElementType().getSizeInBits());
4684 int LanesDefined = BitsDefined / BitsPerShuffleLane;
4686 // This source is expected to fill ResMultiplier lanes of the final shuffle,
4687 // starting at the appropriate offset.
4688 int *LaneMask = &Mask[i * ResMultiplier];
4690 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
4691 ExtractBase += NumElts * (Src - Sources.begin());
4692 for (int j = 0; j < LanesDefined; ++j)
4693 LaneMask[j] = ExtractBase + j;
4696 // Final check before we try to produce nonsense...
4697 if (!isShuffleMaskLegal(Mask, ShuffleVT))
4700 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
4701 for (unsigned i = 0; i < Sources.size(); ++i)
4702 ShuffleOps[i] = Sources[i].ShuffleVec;
4704 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
4705 ShuffleOps[1], &Mask[0]);
4706 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
4709 // check if an EXT instruction can handle the shuffle mask when the
4710 // vector sources of the shuffle are the same.
4711 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4712 unsigned NumElts = VT.getVectorNumElements();
4714 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4720 // If this is a VEXT shuffle, the immediate value is the index of the first
4721 // element. The other shuffle indices must be the successive elements after
4723 unsigned ExpectedElt = Imm;
4724 for (unsigned i = 1; i < NumElts; ++i) {
4725 // Increment the expected index. If it wraps around, just follow it
4726 // back to index zero and keep going.
4728 if (ExpectedElt == NumElts)
4732 continue; // ignore UNDEF indices
4733 if (ExpectedElt != static_cast<unsigned>(M[i]))
4740 // check if an EXT instruction can handle the shuffle mask when the
4741 // vector sources of the shuffle are different.
4742 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
4744 // Look for the first non-undef element.
4745 const int *FirstRealElt = std::find_if(M.begin(), M.end(),
4746 [](int Elt) {return Elt >= 0;});
4748 // Benefit form APInt to handle overflow when calculating expected element.
4749 unsigned NumElts = VT.getVectorNumElements();
4750 unsigned MaskBits = APInt(32, NumElts * 2).logBase2();
4751 APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1);
4752 // The following shuffle indices must be the successive elements after the
4753 // first real element.
4754 const int *FirstWrongElt = std::find_if(FirstRealElt + 1, M.end(),
4755 [&](int Elt) {return Elt != ExpectedElt++ && Elt != -1;});
4756 if (FirstWrongElt != M.end())
4759 // The index of an EXT is the first element if it is not UNDEF.
4760 // Watch out for the beginning UNDEFs. The EXT index should be the expected
4761 // value of the first element. E.g.
4762 // <-1, -1, 3, ...> is treated as <1, 2, 3, ...>.
4763 // <-1, -1, 0, 1, ...> is treated as <2*NumElts-2, 2*NumElts-1, 0, 1, ...>.
4764 // ExpectedElt is the last mask index plus 1.
4765 Imm = ExpectedElt.getZExtValue();
4767 // There are two difference cases requiring to reverse input vectors.
4768 // For example, for vector <4 x i32> we have the following cases,
4769 // Case 1: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, -1, 0>)
4770 // Case 2: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, 7, 0>)
4771 // For both cases, we finally use mask <5, 6, 7, 0>, which requires
4772 // to reverse two input vectors.
4781 /// isREVMask - Check if a vector shuffle corresponds to a REV
4782 /// instruction with the specified blocksize. (The order of the elements
4783 /// within each block of the vector is reversed.)
4784 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4785 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
4786 "Only possible block sizes for REV are: 16, 32, 64");
4788 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4792 unsigned NumElts = VT.getVectorNumElements();
4793 unsigned BlockElts = M[0] + 1;
4794 // If the first shuffle index is UNDEF, be optimistic.
4796 BlockElts = BlockSize / EltSz;
4798 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4801 for (unsigned i = 0; i < NumElts; ++i) {
4803 continue; // ignore UNDEF indices
4804 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
4811 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4812 unsigned NumElts = VT.getVectorNumElements();
4813 WhichResult = (M[0] == 0 ? 0 : 1);
4814 unsigned Idx = WhichResult * NumElts / 2;
4815 for (unsigned i = 0; i != NumElts; i += 2) {
4816 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4817 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
4825 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4826 unsigned NumElts = VT.getVectorNumElements();
4827 WhichResult = (M[0] == 0 ? 0 : 1);
4828 for (unsigned i = 0; i != NumElts; ++i) {
4830 continue; // ignore UNDEF indices
4831 if ((unsigned)M[i] != 2 * i + WhichResult)
4838 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4839 unsigned NumElts = VT.getVectorNumElements();
4840 WhichResult = (M[0] == 0 ? 0 : 1);
4841 for (unsigned i = 0; i < NumElts; i += 2) {
4842 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4843 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
4849 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
4850 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4851 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4852 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4853 unsigned NumElts = VT.getVectorNumElements();
4854 WhichResult = (M[0] == 0 ? 0 : 1);
4855 unsigned Idx = WhichResult * NumElts / 2;
4856 for (unsigned i = 0; i != NumElts; i += 2) {
4857 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4858 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
4866 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
4867 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4868 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4869 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4870 unsigned Half = VT.getVectorNumElements() / 2;
4871 WhichResult = (M[0] == 0 ? 0 : 1);
4872 for (unsigned j = 0; j != 2; ++j) {
4873 unsigned Idx = WhichResult;
4874 for (unsigned i = 0; i != Half; ++i) {
4875 int MIdx = M[i + j * Half];
4876 if (MIdx >= 0 && (unsigned)MIdx != Idx)
4885 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
4886 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4887 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4888 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4889 unsigned NumElts = VT.getVectorNumElements();
4890 WhichResult = (M[0] == 0 ? 0 : 1);
4891 for (unsigned i = 0; i < NumElts; i += 2) {
4892 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4893 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
4899 static bool isINSMask(ArrayRef<int> M, int NumInputElements,
4900 bool &DstIsLeft, int &Anomaly) {
4901 if (M.size() != static_cast<size_t>(NumInputElements))
4904 int NumLHSMatch = 0, NumRHSMatch = 0;
4905 int LastLHSMismatch = -1, LastRHSMismatch = -1;
4907 for (int i = 0; i < NumInputElements; ++i) {
4917 LastLHSMismatch = i;
4919 if (M[i] == i + NumInputElements)
4922 LastRHSMismatch = i;
4925 if (NumLHSMatch == NumInputElements - 1) {
4927 Anomaly = LastLHSMismatch;
4929 } else if (NumRHSMatch == NumInputElements - 1) {
4931 Anomaly = LastRHSMismatch;
4938 static bool isConcatMask(ArrayRef<int> Mask, EVT VT, bool SplitLHS) {
4939 if (VT.getSizeInBits() != 128)
4942 unsigned NumElts = VT.getVectorNumElements();
4944 for (int I = 0, E = NumElts / 2; I != E; I++) {
4949 int Offset = NumElts / 2;
4950 for (int I = NumElts / 2, E = NumElts; I != E; I++) {
4951 if (Mask[I] != I + SplitLHS * Offset)
4958 static SDValue tryFormConcatFromShuffle(SDValue Op, SelectionDAG &DAG) {
4960 EVT VT = Op.getValueType();
4961 SDValue V0 = Op.getOperand(0);
4962 SDValue V1 = Op.getOperand(1);
4963 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();
4965 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
4966 VT.getVectorElementType() != V1.getValueType().getVectorElementType())
4969 bool SplitV0 = V0.getValueType().getSizeInBits() == 128;
4971 if (!isConcatMask(Mask, VT, SplitV0))
4974 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
4975 VT.getVectorNumElements() / 2);
4977 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
4978 DAG.getConstant(0, MVT::i64));
4980 if (V1.getValueType().getSizeInBits() == 128) {
4981 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
4982 DAG.getConstant(0, MVT::i64));
4984 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
4987 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4988 /// the specified operations to build the shuffle.
4989 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4990 SDValue RHS, SelectionDAG &DAG,
4992 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4993 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
4994 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
4997 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5006 OP_VUZPL, // VUZP, left result
5007 OP_VUZPR, // VUZP, right result
5008 OP_VZIPL, // VZIP, left result
5009 OP_VZIPR, // VZIP, right result
5010 OP_VTRNL, // VTRN, left result
5011 OP_VTRNR // VTRN, right result
5014 if (OpNum == OP_COPY) {
5015 if (LHSID == (1 * 9 + 2) * 9 + 3)
5017 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
5021 SDValue OpLHS, OpRHS;
5022 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5023 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5024 EVT VT = OpLHS.getValueType();
5028 llvm_unreachable("Unknown shuffle opcode!");
5030 // VREV divides the vector in half and swaps within the half.
5031 if (VT.getVectorElementType() == MVT::i32 ||
5032 VT.getVectorElementType() == MVT::f32)
5033 return DAG.getNode(AArch64ISD::REV64, dl, VT, OpLHS);
5034 // vrev <4 x i16> -> REV32
5035 if (VT.getVectorElementType() == MVT::i16 ||
5036 VT.getVectorElementType() == MVT::f16)
5037 return DAG.getNode(AArch64ISD::REV32, dl, VT, OpLHS);
5038 // vrev <4 x i8> -> REV16
5039 assert(VT.getVectorElementType() == MVT::i8);
5040 return DAG.getNode(AArch64ISD::REV16, dl, VT, OpLHS);
5045 EVT EltTy = VT.getVectorElementType();
5047 if (EltTy == MVT::i8)
5048 Opcode = AArch64ISD::DUPLANE8;
5049 else if (EltTy == MVT::i16)
5050 Opcode = AArch64ISD::DUPLANE16;
5051 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
5052 Opcode = AArch64ISD::DUPLANE32;
5053 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
5054 Opcode = AArch64ISD::DUPLANE64;
5056 llvm_unreachable("Invalid vector element type?");
5058 if (VT.getSizeInBits() == 64)
5059 OpLHS = WidenVector(OpLHS, DAG);
5060 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, MVT::i64);
5061 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
5066 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
5067 return DAG.getNode(AArch64ISD::EXT, dl, VT, OpLHS, OpRHS,
5068 DAG.getConstant(Imm, MVT::i32));
5071 return DAG.getNode(AArch64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS,
5074 return DAG.getNode(AArch64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS,
5077 return DAG.getNode(AArch64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS,
5080 return DAG.getNode(AArch64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS,
5083 return DAG.getNode(AArch64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS,
5086 return DAG.getNode(AArch64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS,
5091 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
5092 SelectionDAG &DAG) {
5093 // Check to see if we can use the TBL instruction.
5094 SDValue V1 = Op.getOperand(0);
5095 SDValue V2 = Op.getOperand(1);
5098 EVT EltVT = Op.getValueType().getVectorElementType();
5099 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
5101 SmallVector<SDValue, 8> TBLMask;
5102 for (int Val : ShuffleMask) {
5103 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
5104 unsigned Offset = Byte + Val * BytesPerElt;
5105 TBLMask.push_back(DAG.getConstant(Offset, MVT::i32));
5109 MVT IndexVT = MVT::v8i8;
5110 unsigned IndexLen = 8;
5111 if (Op.getValueType().getSizeInBits() == 128) {
5112 IndexVT = MVT::v16i8;
5116 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
5117 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
5120 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
5122 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
5123 Shuffle = DAG.getNode(
5124 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5125 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, MVT::i32), V1Cst,
5126 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5127 makeArrayRef(TBLMask.data(), IndexLen)));
5129 if (IndexLen == 8) {
5130 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
5131 Shuffle = DAG.getNode(
5132 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5133 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, MVT::i32), V1Cst,
5134 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5135 makeArrayRef(TBLMask.data(), IndexLen)));
5137 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
5138 // cannot currently represent the register constraints on the input
5140 // Shuffle = DAG.getNode(AArch64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
5141 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5142 // &TBLMask[0], IndexLen));
5143 Shuffle = DAG.getNode(
5144 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5145 DAG.getConstant(Intrinsic::aarch64_neon_tbl2, MVT::i32), V1Cst, V2Cst,
5146 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5147 makeArrayRef(TBLMask.data(), IndexLen)));
5150 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
5153 static unsigned getDUPLANEOp(EVT EltType) {
5154 if (EltType == MVT::i8)
5155 return AArch64ISD::DUPLANE8;
5156 if (EltType == MVT::i16 || EltType == MVT::f16)
5157 return AArch64ISD::DUPLANE16;
5158 if (EltType == MVT::i32 || EltType == MVT::f32)
5159 return AArch64ISD::DUPLANE32;
5160 if (EltType == MVT::i64 || EltType == MVT::f64)
5161 return AArch64ISD::DUPLANE64;
5163 llvm_unreachable("Invalid vector element type?");
5166 SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5167 SelectionDAG &DAG) const {
5169 EVT VT = Op.getValueType();
5171 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5173 // Convert shuffles that are directly supported on NEON to target-specific
5174 // DAG nodes, instead of keeping them as shuffles and matching them again
5175 // during code selection. This is more efficient and avoids the possibility
5176 // of inconsistencies between legalization and selection.
5177 ArrayRef<int> ShuffleMask = SVN->getMask();
5179 SDValue V1 = Op.getOperand(0);
5180 SDValue V2 = Op.getOperand(1);
5182 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
5183 V1.getValueType().getSimpleVT())) {
5184 int Lane = SVN->getSplatIndex();
5185 // If this is undef splat, generate it via "just" vdup, if possible.
5189 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
5190 return DAG.getNode(AArch64ISD::DUP, dl, V1.getValueType(),
5192 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
5193 // constant. If so, we can just reference the lane's definition directly.
5194 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
5195 !isa<ConstantSDNode>(V1.getOperand(Lane)))
5196 return DAG.getNode(AArch64ISD::DUP, dl, VT, V1.getOperand(Lane));
5198 // Otherwise, duplicate from the lane of the input vector.
5199 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
5201 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
5202 // to make a vector of the same size as this SHUFFLE. We can ignore the
5203 // extract entirely, and canonicalise the concat using WidenVector.
5204 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
5205 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
5206 V1 = V1.getOperand(0);
5207 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
5208 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
5209 Lane -= Idx * VT.getVectorNumElements() / 2;
5210 V1 = WidenVector(V1.getOperand(Idx), DAG);
5211 } else if (VT.getSizeInBits() == 64)
5212 V1 = WidenVector(V1, DAG);
5214 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, MVT::i64));
5217 if (isREVMask(ShuffleMask, VT, 64))
5218 return DAG.getNode(AArch64ISD::REV64, dl, V1.getValueType(), V1, V2);
5219 if (isREVMask(ShuffleMask, VT, 32))
5220 return DAG.getNode(AArch64ISD::REV32, dl, V1.getValueType(), V1, V2);
5221 if (isREVMask(ShuffleMask, VT, 16))
5222 return DAG.getNode(AArch64ISD::REV16, dl, V1.getValueType(), V1, V2);
5224 bool ReverseEXT = false;
5226 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
5229 Imm *= getExtFactor(V1);
5230 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V2,
5231 DAG.getConstant(Imm, MVT::i32));
5232 } else if (V2->getOpcode() == ISD::UNDEF &&
5233 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
5234 Imm *= getExtFactor(V1);
5235 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V1,
5236 DAG.getConstant(Imm, MVT::i32));
5239 unsigned WhichResult;
5240 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
5241 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5242 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5244 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
5245 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5246 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5248 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
5249 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5250 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5253 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5254 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5255 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5257 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5258 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5259 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5261 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5262 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5263 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5266 SDValue Concat = tryFormConcatFromShuffle(Op, DAG);
5267 if (Concat.getNode())
5272 int NumInputElements = V1.getValueType().getVectorNumElements();
5273 if (isINSMask(ShuffleMask, NumInputElements, DstIsLeft, Anomaly)) {
5274 SDValue DstVec = DstIsLeft ? V1 : V2;
5275 SDValue DstLaneV = DAG.getConstant(Anomaly, MVT::i64);
5277 SDValue SrcVec = V1;
5278 int SrcLane = ShuffleMask[Anomaly];
5279 if (SrcLane >= NumInputElements) {
5281 SrcLane -= VT.getVectorNumElements();
5283 SDValue SrcLaneV = DAG.getConstant(SrcLane, MVT::i64);
5285 EVT ScalarVT = VT.getVectorElementType();
5287 if (ScalarVT.getSizeInBits() < 32 && ScalarVT.isInteger())
5288 ScalarVT = MVT::i32;
5291 ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5292 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ScalarVT, SrcVec, SrcLaneV),
5296 // If the shuffle is not directly supported and it has 4 elements, use
5297 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5298 unsigned NumElts = VT.getVectorNumElements();
5300 unsigned PFIndexes[4];
5301 for (unsigned i = 0; i != 4; ++i) {
5302 if (ShuffleMask[i] < 0)
5305 PFIndexes[i] = ShuffleMask[i];
5308 // Compute the index in the perfect shuffle table.
5309 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5310 PFIndexes[2] * 9 + PFIndexes[3];
5311 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5312 unsigned Cost = (PFEntry >> 30);
5315 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5318 return GenerateTBL(Op, ShuffleMask, DAG);
5321 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
5323 EVT VT = BVN->getValueType(0);
5324 APInt SplatBits, SplatUndef;
5325 unsigned SplatBitSize;
5327 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5328 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
5330 for (unsigned i = 0; i < NumSplats; ++i) {
5331 CnstBits <<= SplatBitSize;
5332 UndefBits <<= SplatBitSize;
5333 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
5334 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
5343 SDValue AArch64TargetLowering::LowerVectorAND(SDValue Op,
5344 SelectionDAG &DAG) const {
5345 BuildVectorSDNode *BVN =
5346 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5347 SDValue LHS = Op.getOperand(0);
5349 EVT VT = Op.getValueType();
5354 APInt CnstBits(VT.getSizeInBits(), 0);
5355 APInt UndefBits(VT.getSizeInBits(), 0);
5356 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5357 // We only have BIC vector immediate instruction, which is and-not.
5358 CnstBits = ~CnstBits;
5360 // We make use of a little bit of goto ickiness in order to avoid having to
5361 // duplicate the immediate matching logic for the undef toggled case.
5362 bool SecondTry = false;
5365 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5366 CnstBits = CnstBits.zextOrTrunc(64);
5367 uint64_t CnstVal = CnstBits.getZExtValue();
5369 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5370 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5371 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5372 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5373 DAG.getConstant(CnstVal, MVT::i32),
5374 DAG.getConstant(0, MVT::i32));
5375 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5378 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5379 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5380 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5381 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5382 DAG.getConstant(CnstVal, MVT::i32),
5383 DAG.getConstant(8, MVT::i32));
5384 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5387 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5388 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5389 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5390 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5391 DAG.getConstant(CnstVal, MVT::i32),
5392 DAG.getConstant(16, MVT::i32));
5393 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5396 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5397 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5398 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5399 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5400 DAG.getConstant(CnstVal, MVT::i32),
5401 DAG.getConstant(24, MVT::i32));
5402 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5405 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5406 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5407 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5408 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5409 DAG.getConstant(CnstVal, MVT::i32),
5410 DAG.getConstant(0, MVT::i32));
5411 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5414 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5415 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5416 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5417 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5418 DAG.getConstant(CnstVal, MVT::i32),
5419 DAG.getConstant(8, MVT::i32));
5420 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5427 CnstBits = ~UndefBits;
5431 // We can always fall back to a non-immediate AND.
5436 // Specialized code to quickly find if PotentialBVec is a BuildVector that
5437 // consists of only the same constant int value, returned in reference arg
5439 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
5440 uint64_t &ConstVal) {
5441 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
5444 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
5447 EVT VT = Bvec->getValueType(0);
5448 unsigned NumElts = VT.getVectorNumElements();
5449 for (unsigned i = 1; i < NumElts; ++i)
5450 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
5452 ConstVal = FirstElt->getZExtValue();
5456 static unsigned getIntrinsicID(const SDNode *N) {
5457 unsigned Opcode = N->getOpcode();
5460 return Intrinsic::not_intrinsic;
5461 case ISD::INTRINSIC_WO_CHAIN: {
5462 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5463 if (IID < Intrinsic::num_intrinsics)
5465 return Intrinsic::not_intrinsic;
5470 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
5471 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
5472 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
5473 // Also, logical shift right -> sri, with the same structure.
5474 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
5475 EVT VT = N->getValueType(0);
5482 // Is the first op an AND?
5483 const SDValue And = N->getOperand(0);
5484 if (And.getOpcode() != ISD::AND)
5487 // Is the second op an shl or lshr?
5488 SDValue Shift = N->getOperand(1);
5489 // This will have been turned into: AArch64ISD::VSHL vector, #shift
5490 // or AArch64ISD::VLSHR vector, #shift
5491 unsigned ShiftOpc = Shift.getOpcode();
5492 if ((ShiftOpc != AArch64ISD::VSHL && ShiftOpc != AArch64ISD::VLSHR))
5494 bool IsShiftRight = ShiftOpc == AArch64ISD::VLSHR;
5496 // Is the shift amount constant?
5497 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
5501 // Is the and mask vector all constant?
5503 if (!isAllConstantBuildVector(And.getOperand(1), C1))
5506 // Is C1 == ~C2, taking into account how much one can shift elements of a
5508 uint64_t C2 = C2node->getZExtValue();
5509 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
5510 if (C2 > ElemSizeInBits)
5512 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
5513 if ((C1 & ElemMask) != (~C2 & ElemMask))
5516 SDValue X = And.getOperand(0);
5517 SDValue Y = Shift.getOperand(0);
5520 IsShiftRight ? Intrinsic::aarch64_neon_vsri : Intrinsic::aarch64_neon_vsli;
5522 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5523 DAG.getConstant(Intrin, MVT::i32), X, Y, Shift.getOperand(1));
5525 DEBUG(dbgs() << "aarch64-lower: transformed: \n");
5526 DEBUG(N->dump(&DAG));
5527 DEBUG(dbgs() << "into: \n");
5528 DEBUG(ResultSLI->dump(&DAG));
5534 SDValue AArch64TargetLowering::LowerVectorOR(SDValue Op,
5535 SelectionDAG &DAG) const {
5536 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
5537 if (EnableAArch64SlrGeneration) {
5538 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
5543 BuildVectorSDNode *BVN =
5544 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
5545 SDValue LHS = Op.getOperand(1);
5547 EVT VT = Op.getValueType();
5549 // OR commutes, so try swapping the operands.
5551 LHS = Op.getOperand(0);
5552 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5557 APInt CnstBits(VT.getSizeInBits(), 0);
5558 APInt UndefBits(VT.getSizeInBits(), 0);
5559 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5560 // We make use of a little bit of goto ickiness in order to avoid having to
5561 // duplicate the immediate matching logic for the undef toggled case.
5562 bool SecondTry = false;
5565 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5566 CnstBits = CnstBits.zextOrTrunc(64);
5567 uint64_t CnstVal = CnstBits.getZExtValue();
5569 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5570 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5571 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5572 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5573 DAG.getConstant(CnstVal, MVT::i32),
5574 DAG.getConstant(0, MVT::i32));
5575 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5578 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5579 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5580 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5581 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5582 DAG.getConstant(CnstVal, MVT::i32),
5583 DAG.getConstant(8, MVT::i32));
5584 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5587 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5588 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5589 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5590 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5591 DAG.getConstant(CnstVal, MVT::i32),
5592 DAG.getConstant(16, MVT::i32));
5593 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5596 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5597 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5598 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5599 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5600 DAG.getConstant(CnstVal, MVT::i32),
5601 DAG.getConstant(24, MVT::i32));
5602 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5605 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5606 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5607 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5608 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5609 DAG.getConstant(CnstVal, MVT::i32),
5610 DAG.getConstant(0, MVT::i32));
5611 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5614 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5615 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5616 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5617 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5618 DAG.getConstant(CnstVal, MVT::i32),
5619 DAG.getConstant(8, MVT::i32));
5620 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5627 CnstBits = UndefBits;
5631 // We can always fall back to a non-immediate OR.
5636 // Normalize the operands of BUILD_VECTOR. The value of constant operands will
5637 // be truncated to fit element width.
5638 static SDValue NormalizeBuildVector(SDValue Op,
5639 SelectionDAG &DAG) {
5640 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
5642 EVT VT = Op.getValueType();
5643 EVT EltTy= VT.getVectorElementType();
5645 if (EltTy.isFloatingPoint() || EltTy.getSizeInBits() > 16)
5648 SmallVector<SDValue, 16> Ops;
5649 for (unsigned I = 0, E = VT.getVectorNumElements(); I != E; ++I) {
5650 SDValue Lane = Op.getOperand(I);
5651 if (Lane.getOpcode() == ISD::Constant) {
5652 APInt LowBits(EltTy.getSizeInBits(),
5653 cast<ConstantSDNode>(Lane)->getZExtValue());
5654 Lane = DAG.getConstant(LowBits.getZExtValue(), MVT::i32);
5656 Ops.push_back(Lane);
5658 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5661 SDValue AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
5662 SelectionDAG &DAG) const {
5664 EVT VT = Op.getValueType();
5665 Op = NormalizeBuildVector(Op, DAG);
5666 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5668 APInt CnstBits(VT.getSizeInBits(), 0);
5669 APInt UndefBits(VT.getSizeInBits(), 0);
5670 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5671 // We make use of a little bit of goto ickiness in order to avoid having to
5672 // duplicate the immediate matching logic for the undef toggled case.
5673 bool SecondTry = false;
5676 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5677 CnstBits = CnstBits.zextOrTrunc(64);
5678 uint64_t CnstVal = CnstBits.getZExtValue();
5680 // Certain magic vector constants (used to express things like NOT
5681 // and NEG) are passed through unmodified. This allows codegen patterns
5682 // for these operations to match. Special-purpose patterns will lower
5683 // these immediates to MOVIs if it proves necessary.
5684 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
5687 // The many faces of MOVI...
5688 if (AArch64_AM::isAdvSIMDModImmType10(CnstVal)) {
5689 CnstVal = AArch64_AM::encodeAdvSIMDModImmType10(CnstVal);
5690 if (VT.getSizeInBits() == 128) {
5691 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::v2i64,
5692 DAG.getConstant(CnstVal, MVT::i32));
5693 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5696 // Support the V64 version via subregister insertion.
5697 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::f64,
5698 DAG.getConstant(CnstVal, MVT::i32));
5699 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5702 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5703 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5704 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5705 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5706 DAG.getConstant(CnstVal, MVT::i32),
5707 DAG.getConstant(0, MVT::i32));
5708 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5711 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5712 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5713 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5714 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5715 DAG.getConstant(CnstVal, MVT::i32),
5716 DAG.getConstant(8, MVT::i32));
5717 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5720 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5721 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5722 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5723 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5724 DAG.getConstant(CnstVal, MVT::i32),
5725 DAG.getConstant(16, MVT::i32));
5726 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5729 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5730 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5731 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5732 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5733 DAG.getConstant(CnstVal, MVT::i32),
5734 DAG.getConstant(24, MVT::i32));
5735 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5738 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5739 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5740 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5741 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5742 DAG.getConstant(CnstVal, MVT::i32),
5743 DAG.getConstant(0, MVT::i32));
5744 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5747 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5748 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5749 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5750 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5751 DAG.getConstant(CnstVal, MVT::i32),
5752 DAG.getConstant(8, MVT::i32));
5753 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5756 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
5757 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
5758 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5759 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
5760 DAG.getConstant(CnstVal, MVT::i32),
5761 DAG.getConstant(264, MVT::i32));
5762 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5765 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
5766 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
5767 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5768 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
5769 DAG.getConstant(CnstVal, MVT::i32),
5770 DAG.getConstant(272, MVT::i32));
5771 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5774 if (AArch64_AM::isAdvSIMDModImmType9(CnstVal)) {
5775 CnstVal = AArch64_AM::encodeAdvSIMDModImmType9(CnstVal);
5776 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
5777 SDValue Mov = DAG.getNode(AArch64ISD::MOVI, dl, MovTy,
5778 DAG.getConstant(CnstVal, MVT::i32));
5779 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5782 // The few faces of FMOV...
5783 if (AArch64_AM::isAdvSIMDModImmType11(CnstVal)) {
5784 CnstVal = AArch64_AM::encodeAdvSIMDModImmType11(CnstVal);
5785 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
5786 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MovTy,
5787 DAG.getConstant(CnstVal, MVT::i32));
5788 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5791 if (AArch64_AM::isAdvSIMDModImmType12(CnstVal) &&
5792 VT.getSizeInBits() == 128) {
5793 CnstVal = AArch64_AM::encodeAdvSIMDModImmType12(CnstVal);
5794 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MVT::v2f64,
5795 DAG.getConstant(CnstVal, MVT::i32));
5796 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5799 // The many faces of MVNI...
5801 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5802 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5803 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5804 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5805 DAG.getConstant(CnstVal, MVT::i32),
5806 DAG.getConstant(0, MVT::i32));
5807 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5810 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5811 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5812 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5813 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5814 DAG.getConstant(CnstVal, MVT::i32),
5815 DAG.getConstant(8, MVT::i32));
5816 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5819 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5820 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5821 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5822 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5823 DAG.getConstant(CnstVal, MVT::i32),
5824 DAG.getConstant(16, MVT::i32));
5825 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5828 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5829 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5830 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5831 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5832 DAG.getConstant(CnstVal, MVT::i32),
5833 DAG.getConstant(24, MVT::i32));
5834 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5837 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5838 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5839 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5840 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5841 DAG.getConstant(CnstVal, MVT::i32),
5842 DAG.getConstant(0, MVT::i32));
5843 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5846 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5847 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5848 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5849 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5850 DAG.getConstant(CnstVal, MVT::i32),
5851 DAG.getConstant(8, MVT::i32));
5852 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5855 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
5856 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
5857 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5858 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
5859 DAG.getConstant(CnstVal, MVT::i32),
5860 DAG.getConstant(264, MVT::i32));
5861 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5864 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
5865 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
5866 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5867 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
5868 DAG.getConstant(CnstVal, MVT::i32),
5869 DAG.getConstant(272, MVT::i32));
5870 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5877 CnstBits = UndefBits;
5882 // Scan through the operands to find some interesting properties we can
5884 // 1) If only one value is used, we can use a DUP, or
5885 // 2) if only the low element is not undef, we can just insert that, or
5886 // 3) if only one constant value is used (w/ some non-constant lanes),
5887 // we can splat the constant value into the whole vector then fill
5888 // in the non-constant lanes.
5889 // 4) FIXME: If different constant values are used, but we can intelligently
5890 // select the values we'll be overwriting for the non-constant
5891 // lanes such that we can directly materialize the vector
5892 // some other way (MOVI, e.g.), we can be sneaky.
5893 unsigned NumElts = VT.getVectorNumElements();
5894 bool isOnlyLowElement = true;
5895 bool usesOnlyOneValue = true;
5896 bool usesOnlyOneConstantValue = true;
5897 bool isConstant = true;
5898 unsigned NumConstantLanes = 0;
5900 SDValue ConstantValue;
5901 for (unsigned i = 0; i < NumElts; ++i) {
5902 SDValue V = Op.getOperand(i);
5903 if (V.getOpcode() == ISD::UNDEF)
5906 isOnlyLowElement = false;
5907 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5910 if (isa<ConstantSDNode>(V) || isa<ConstantFPSDNode>(V)) {
5912 if (!ConstantValue.getNode())
5914 else if (ConstantValue != V)
5915 usesOnlyOneConstantValue = false;
5918 if (!Value.getNode())
5920 else if (V != Value)
5921 usesOnlyOneValue = false;
5924 if (!Value.getNode())
5925 return DAG.getUNDEF(VT);
5927 if (isOnlyLowElement)
5928 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5930 // Use DUP for non-constant splats. For f32 constant splats, reduce to
5931 // i32 and try again.
5932 if (usesOnlyOneValue) {
5934 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5935 Value.getValueType() != VT)
5936 return DAG.getNode(AArch64ISD::DUP, dl, VT, Value);
5938 // This is actually a DUPLANExx operation, which keeps everything vectory.
5940 // DUPLANE works on 128-bit vectors, widen it if necessary.
5941 SDValue Lane = Value.getOperand(1);
5942 Value = Value.getOperand(0);
5943 if (Value.getValueType().getSizeInBits() == 64)
5944 Value = WidenVector(Value, DAG);
5946 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
5947 return DAG.getNode(Opcode, dl, VT, Value, Lane);
5950 if (VT.getVectorElementType().isFloatingPoint()) {
5951 SmallVector<SDValue, 8> Ops;
5953 (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
5954 for (unsigned i = 0; i < NumElts; ++i)
5955 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
5956 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
5957 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5958 Val = LowerBUILD_VECTOR(Val, DAG);
5960 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5964 // If there was only one constant value used and for more than one lane,
5965 // start by splatting that value, then replace the non-constant lanes. This
5966 // is better than the default, which will perform a separate initialization
5968 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
5969 SDValue Val = DAG.getNode(AArch64ISD::DUP, dl, VT, ConstantValue);
5970 // Now insert the non-constant lanes.
5971 for (unsigned i = 0; i < NumElts; ++i) {
5972 SDValue V = Op.getOperand(i);
5973 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5974 if (!isa<ConstantSDNode>(V) && !isa<ConstantFPSDNode>(V)) {
5975 // Note that type legalization likely mucked about with the VT of the
5976 // source operand, so we may have to convert it here before inserting.
5977 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
5983 // If all elements are constants and the case above didn't get hit, fall back
5984 // to the default expansion, which will generate a load from the constant
5989 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5991 SDValue shuffle = ReconstructShuffle(Op, DAG);
5992 if (shuffle != SDValue())
5996 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5997 // know the default expansion would otherwise fall back on something even
5998 // worse. For a vector with one or two non-undef values, that's
5999 // scalar_to_vector for the elements followed by a shuffle (provided the
6000 // shuffle is valid for the target) and materialization element by element
6001 // on the stack followed by a load for everything else.
6002 if (!isConstant && !usesOnlyOneValue) {
6003 SDValue Vec = DAG.getUNDEF(VT);
6004 SDValue Op0 = Op.getOperand(0);
6005 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
6007 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
6008 // a) Avoid a RMW dependency on the full vector register, and
6009 // b) Allow the register coalescer to fold away the copy if the
6010 // value is already in an S or D register.
6011 if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
6012 unsigned SubIdx = ElemSize == 32 ? AArch64::ssub : AArch64::dsub;
6014 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
6015 DAG.getTargetConstant(SubIdx, MVT::i32));
6016 Vec = SDValue(N, 0);
6019 for (; i < NumElts; ++i) {
6020 SDValue V = Op.getOperand(i);
6021 if (V.getOpcode() == ISD::UNDEF)
6023 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
6024 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
6029 // Just use the default expansion. We failed to find a better alternative.
6033 SDValue AArch64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
6034 SelectionDAG &DAG) const {
6035 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
6037 // Check for non-constant or out of range lane.
6038 EVT VT = Op.getOperand(0).getValueType();
6039 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(2));
6040 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6044 // Insertion/extraction are legal for V128 types.
6045 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6046 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6050 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6051 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6054 // For V64 types, we perform insertion by expanding the value
6055 // to a V128 type and perform the insertion on that.
6057 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6058 EVT WideTy = WideVec.getValueType();
6060 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
6061 Op.getOperand(1), Op.getOperand(2));
6062 // Re-narrow the resultant vector.
6063 return NarrowVector(Node, DAG);
6067 AArch64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6068 SelectionDAG &DAG) const {
6069 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
6071 // Check for non-constant or out of range lane.
6072 EVT VT = Op.getOperand(0).getValueType();
6073 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6074 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6078 // Insertion/extraction are legal for V128 types.
6079 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6080 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6084 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6085 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6088 // For V64 types, we perform extraction by expanding the value
6089 // to a V128 type and perform the extraction on that.
6091 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6092 EVT WideTy = WideVec.getValueType();
6094 EVT ExtrTy = WideTy.getVectorElementType();
6095 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
6098 // For extractions, we just return the result directly.
6099 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
6103 SDValue AArch64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
6104 SelectionDAG &DAG) const {
6105 EVT VT = Op.getOperand(0).getValueType();
6111 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6114 unsigned Val = Cst->getZExtValue();
6116 unsigned Size = Op.getValueType().getSizeInBits();
6120 return DAG.getTargetExtractSubreg(AArch64::bsub, dl, Op.getValueType(),
6123 return DAG.getTargetExtractSubreg(AArch64::hsub, dl, Op.getValueType(),
6126 return DAG.getTargetExtractSubreg(AArch64::ssub, dl, Op.getValueType(),
6129 return DAG.getTargetExtractSubreg(AArch64::dsub, dl, Op.getValueType(),
6132 llvm_unreachable("Unexpected vector type in extract_subvector!");
6135 // If this is extracting the upper 64-bits of a 128-bit vector, we match
6137 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
6143 bool AArch64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
6145 if (VT.getVectorNumElements() == 4 &&
6146 (VT.is128BitVector() || VT.is64BitVector())) {
6147 unsigned PFIndexes[4];
6148 for (unsigned i = 0; i != 4; ++i) {
6152 PFIndexes[i] = M[i];
6155 // Compute the index in the perfect shuffle table.
6156 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
6157 PFIndexes[2] * 9 + PFIndexes[3];
6158 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
6159 unsigned Cost = (PFEntry >> 30);
6167 unsigned DummyUnsigned;
6169 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
6170 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
6171 isEXTMask(M, VT, DummyBool, DummyUnsigned) ||
6172 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
6173 isTRNMask(M, VT, DummyUnsigned) || isUZPMask(M, VT, DummyUnsigned) ||
6174 isZIPMask(M, VT, DummyUnsigned) ||
6175 isTRN_v_undef_Mask(M, VT, DummyUnsigned) ||
6176 isUZP_v_undef_Mask(M, VT, DummyUnsigned) ||
6177 isZIP_v_undef_Mask(M, VT, DummyUnsigned) ||
6178 isINSMask(M, VT.getVectorNumElements(), DummyBool, DummyInt) ||
6179 isConcatMask(M, VT, VT.getSizeInBits() == 128));
6182 /// getVShiftImm - Check if this is a valid build_vector for the immediate
6183 /// operand of a vector shift operation, where all the elements of the
6184 /// build_vector must have the same constant integer value.
6185 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
6186 // Ignore bit_converts.
6187 while (Op.getOpcode() == ISD::BITCAST)
6188 Op = Op.getOperand(0);
6189 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
6190 APInt SplatBits, SplatUndef;
6191 unsigned SplatBitSize;
6193 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
6194 HasAnyUndefs, ElementBits) ||
6195 SplatBitSize > ElementBits)
6197 Cnt = SplatBits.getSExtValue();
6201 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
6202 /// operand of a vector shift left operation. That value must be in the range:
6203 /// 0 <= Value < ElementBits for a left shift; or
6204 /// 0 <= Value <= ElementBits for a long left shift.
6205 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
6206 assert(VT.isVector() && "vector shift count is not a vector type");
6207 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6208 if (!getVShiftImm(Op, ElementBits, Cnt))
6210 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
6213 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
6214 /// operand of a vector shift right operation. For a shift opcode, the value
6215 /// is positive, but for an intrinsic the value count must be negative. The
6216 /// absolute value must be in the range:
6217 /// 1 <= |Value| <= ElementBits for a right shift; or
6218 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
6219 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
6221 assert(VT.isVector() && "vector shift count is not a vector type");
6222 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6223 if (!getVShiftImm(Op, ElementBits, Cnt))
6227 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
6230 SDValue AArch64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
6231 SelectionDAG &DAG) const {
6232 EVT VT = Op.getValueType();
6236 if (!Op.getOperand(1).getValueType().isVector())
6238 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
6240 switch (Op.getOpcode()) {
6242 llvm_unreachable("unexpected shift opcode");
6245 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
6246 return DAG.getNode(AArch64ISD::VSHL, SDLoc(Op), VT, Op.getOperand(0),
6247 DAG.getConstant(Cnt, MVT::i32));
6248 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6249 DAG.getConstant(Intrinsic::aarch64_neon_ushl, MVT::i32),
6250 Op.getOperand(0), Op.getOperand(1));
6253 // Right shift immediate
6254 if (isVShiftRImm(Op.getOperand(1), VT, false, false, Cnt) &&
6257 (Op.getOpcode() == ISD::SRA) ? AArch64ISD::VASHR : AArch64ISD::VLSHR;
6258 return DAG.getNode(Opc, SDLoc(Op), VT, Op.getOperand(0),
6259 DAG.getConstant(Cnt, MVT::i32));
6262 // Right shift register. Note, there is not a shift right register
6263 // instruction, but the shift left register instruction takes a signed
6264 // value, where negative numbers specify a right shift.
6265 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::aarch64_neon_sshl
6266 : Intrinsic::aarch64_neon_ushl;
6267 // negate the shift amount
6268 SDValue NegShift = DAG.getNode(AArch64ISD::NEG, DL, VT, Op.getOperand(1));
6269 SDValue NegShiftLeft =
6270 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6271 DAG.getConstant(Opc, MVT::i32), Op.getOperand(0), NegShift);
6272 return NegShiftLeft;
6278 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
6279 AArch64CC::CondCode CC, bool NoNans, EVT VT,
6280 SDLoc dl, SelectionDAG &DAG) {
6281 EVT SrcVT = LHS.getValueType();
6283 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
6284 APInt CnstBits(VT.getSizeInBits(), 0);
6285 APInt UndefBits(VT.getSizeInBits(), 0);
6286 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
6287 bool IsZero = IsCnst && (CnstBits == 0);
6289 if (SrcVT.getVectorElementType().isFloatingPoint()) {
6293 case AArch64CC::NE: {
6296 Fcmeq = DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6298 Fcmeq = DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6299 return DAG.getNode(AArch64ISD::NOT, dl, VT, Fcmeq);
6303 return DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6304 return DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6307 return DAG.getNode(AArch64ISD::FCMGEz, dl, VT, LHS);
6308 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, LHS, RHS);
6311 return DAG.getNode(AArch64ISD::FCMGTz, dl, VT, LHS);
6312 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, LHS, RHS);
6315 return DAG.getNode(AArch64ISD::FCMLEz, dl, VT, LHS);
6316 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, RHS, LHS);
6320 // If we ignore NaNs then we can use to the MI implementation.
6324 return DAG.getNode(AArch64ISD::FCMLTz, dl, VT, LHS);
6325 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, RHS, LHS);
6332 case AArch64CC::NE: {
6335 Cmeq = DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6337 Cmeq = DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6338 return DAG.getNode(AArch64ISD::NOT, dl, VT, Cmeq);
6342 return DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6343 return DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6346 return DAG.getNode(AArch64ISD::CMGEz, dl, VT, LHS);
6347 return DAG.getNode(AArch64ISD::CMGE, dl, VT, LHS, RHS);
6350 return DAG.getNode(AArch64ISD::CMGTz, dl, VT, LHS);
6351 return DAG.getNode(AArch64ISD::CMGT, dl, VT, LHS, RHS);
6354 return DAG.getNode(AArch64ISD::CMLEz, dl, VT, LHS);
6355 return DAG.getNode(AArch64ISD::CMGE, dl, VT, RHS, LHS);
6357 return DAG.getNode(AArch64ISD::CMHS, dl, VT, RHS, LHS);
6359 return DAG.getNode(AArch64ISD::CMHI, dl, VT, RHS, LHS);
6362 return DAG.getNode(AArch64ISD::CMLTz, dl, VT, LHS);
6363 return DAG.getNode(AArch64ISD::CMGT, dl, VT, RHS, LHS);
6365 return DAG.getNode(AArch64ISD::CMHI, dl, VT, LHS, RHS);
6367 return DAG.getNode(AArch64ISD::CMHS, dl, VT, LHS, RHS);
6371 SDValue AArch64TargetLowering::LowerVSETCC(SDValue Op,
6372 SelectionDAG &DAG) const {
6373 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6374 SDValue LHS = Op.getOperand(0);
6375 SDValue RHS = Op.getOperand(1);
6378 if (LHS.getValueType().getVectorElementType().isInteger()) {
6379 assert(LHS.getValueType() == RHS.getValueType());
6380 AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC);
6381 return EmitVectorComparison(LHS, RHS, AArch64CC, false, Op.getValueType(),
6385 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
6386 LHS.getValueType().getVectorElementType() == MVT::f64);
6388 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
6389 // clean. Some of them require two branches to implement.
6390 AArch64CC::CondCode CC1, CC2;
6392 changeVectorFPCCToAArch64CC(CC, CC1, CC2, ShouldInvert);
6394 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
6396 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, Op.getValueType(), dl, DAG);
6400 if (CC2 != AArch64CC::AL) {
6402 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, Op.getValueType(), dl, DAG);
6403 if (!Cmp2.getNode())
6406 Cmp = DAG.getNode(ISD::OR, dl, Cmp.getValueType(), Cmp, Cmp2);
6410 return Cmp = DAG.getNOT(dl, Cmp, Cmp.getValueType());
6415 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
6416 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
6417 /// specified in the intrinsic calls.
6418 bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
6420 unsigned Intrinsic) const {
6421 switch (Intrinsic) {
6422 case Intrinsic::aarch64_neon_ld2:
6423 case Intrinsic::aarch64_neon_ld3:
6424 case Intrinsic::aarch64_neon_ld4:
6425 case Intrinsic::aarch64_neon_ld1x2:
6426 case Intrinsic::aarch64_neon_ld1x3:
6427 case Intrinsic::aarch64_neon_ld1x4:
6428 case Intrinsic::aarch64_neon_ld2lane:
6429 case Intrinsic::aarch64_neon_ld3lane:
6430 case Intrinsic::aarch64_neon_ld4lane:
6431 case Intrinsic::aarch64_neon_ld2r:
6432 case Intrinsic::aarch64_neon_ld3r:
6433 case Intrinsic::aarch64_neon_ld4r: {
6434 Info.opc = ISD::INTRINSIC_W_CHAIN;
6435 // Conservatively set memVT to the entire set of vectors loaded.
6436 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
6437 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6438 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6441 Info.vol = false; // volatile loads with NEON intrinsics not supported
6442 Info.readMem = true;
6443 Info.writeMem = false;
6446 case Intrinsic::aarch64_neon_st2:
6447 case Intrinsic::aarch64_neon_st3:
6448 case Intrinsic::aarch64_neon_st4:
6449 case Intrinsic::aarch64_neon_st1x2:
6450 case Intrinsic::aarch64_neon_st1x3:
6451 case Intrinsic::aarch64_neon_st1x4:
6452 case Intrinsic::aarch64_neon_st2lane:
6453 case Intrinsic::aarch64_neon_st3lane:
6454 case Intrinsic::aarch64_neon_st4lane: {
6455 Info.opc = ISD::INTRINSIC_VOID;
6456 // Conservatively set memVT to the entire set of vectors stored.
6457 unsigned NumElts = 0;
6458 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
6459 Type *ArgTy = I.getArgOperand(ArgI)->getType();
6460 if (!ArgTy->isVectorTy())
6462 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
6464 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6465 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6468 Info.vol = false; // volatile stores with NEON intrinsics not supported
6469 Info.readMem = false;
6470 Info.writeMem = true;
6473 case Intrinsic::aarch64_ldaxr:
6474 case Intrinsic::aarch64_ldxr: {
6475 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
6476 Info.opc = ISD::INTRINSIC_W_CHAIN;
6477 Info.memVT = MVT::getVT(PtrTy->getElementType());
6478 Info.ptrVal = I.getArgOperand(0);
6480 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
6482 Info.readMem = true;
6483 Info.writeMem = false;
6486 case Intrinsic::aarch64_stlxr:
6487 case Intrinsic::aarch64_stxr: {
6488 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
6489 Info.opc = ISD::INTRINSIC_W_CHAIN;
6490 Info.memVT = MVT::getVT(PtrTy->getElementType());
6491 Info.ptrVal = I.getArgOperand(1);
6493 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
6495 Info.readMem = false;
6496 Info.writeMem = true;
6499 case Intrinsic::aarch64_ldaxp:
6500 case Intrinsic::aarch64_ldxp: {
6501 Info.opc = ISD::INTRINSIC_W_CHAIN;
6502 Info.memVT = MVT::i128;
6503 Info.ptrVal = I.getArgOperand(0);
6507 Info.readMem = true;
6508 Info.writeMem = false;
6511 case Intrinsic::aarch64_stlxp:
6512 case Intrinsic::aarch64_stxp: {
6513 Info.opc = ISD::INTRINSIC_W_CHAIN;
6514 Info.memVT = MVT::i128;
6515 Info.ptrVal = I.getArgOperand(2);
6519 Info.readMem = false;
6520 Info.writeMem = true;
6530 // Truncations from 64-bit GPR to 32-bit GPR is free.
6531 bool AArch64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
6532 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6534 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6535 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6536 return NumBits1 > NumBits2;
6538 bool AArch64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
6539 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6541 unsigned NumBits1 = VT1.getSizeInBits();
6542 unsigned NumBits2 = VT2.getSizeInBits();
6543 return NumBits1 > NumBits2;
6546 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
6548 bool AArch64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
6549 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6551 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6552 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6553 return NumBits1 == 32 && NumBits2 == 64;
6555 bool AArch64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
6556 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6558 unsigned NumBits1 = VT1.getSizeInBits();
6559 unsigned NumBits2 = VT2.getSizeInBits();
6560 return NumBits1 == 32 && NumBits2 == 64;
6563 bool AArch64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
6564 EVT VT1 = Val.getValueType();
6565 if (isZExtFree(VT1, VT2)) {
6569 if (Val.getOpcode() != ISD::LOAD)
6572 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
6573 return (VT1.isSimple() && !VT1.isVector() && VT1.isInteger() &&
6574 VT2.isSimple() && !VT2.isVector() && VT2.isInteger() &&
6575 VT1.getSizeInBits() <= 32);
6578 bool AArch64TargetLowering::hasPairedLoad(Type *LoadedType,
6579 unsigned &RequiredAligment) const {
6580 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
6582 // Cyclone supports unaligned accesses.
6583 RequiredAligment = 0;
6584 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
6585 return NumBits == 32 || NumBits == 64;
6588 bool AArch64TargetLowering::hasPairedLoad(EVT LoadedType,
6589 unsigned &RequiredAligment) const {
6590 if (!LoadedType.isSimple() ||
6591 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
6593 // Cyclone supports unaligned accesses.
6594 RequiredAligment = 0;
6595 unsigned NumBits = LoadedType.getSizeInBits();
6596 return NumBits == 32 || NumBits == 64;
6599 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
6600 unsigned AlignCheck) {
6601 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
6602 (DstAlign == 0 || DstAlign % AlignCheck == 0));
6605 EVT AArch64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
6606 unsigned SrcAlign, bool IsMemset,
6609 MachineFunction &MF) const {
6610 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
6611 // instruction to materialize the v2i64 zero and one store (with restrictive
6612 // addressing mode). Just do two i64 store of zero-registers.
6614 const Function *F = MF.getFunction();
6615 if (Subtarget->hasFPARMv8() && !IsMemset && Size >= 16 &&
6616 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
6617 Attribute::NoImplicitFloat) &&
6618 (memOpAlign(SrcAlign, DstAlign, 16) ||
6619 (allowsMisalignedMemoryAccesses(MVT::f128, 0, 1, &Fast) && Fast)))
6622 return Size >= 8 ? MVT::i64 : MVT::i32;
6625 // 12-bit optionally shifted immediates are legal for adds.
6626 bool AArch64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
6627 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
6632 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
6633 // immediates is the same as for an add or a sub.
6634 bool AArch64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
6637 return isLegalAddImmediate(Immed);
6640 /// isLegalAddressingMode - Return true if the addressing mode represented
6641 /// by AM is legal for this target, for a load/store of the specified type.
6642 bool AArch64TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6644 // AArch64 has five basic addressing modes:
6646 // reg + 9-bit signed offset
6647 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
6649 // reg + SIZE_IN_BYTES * reg
6651 // No global is ever allowed as a base.
6655 // No reg+reg+imm addressing.
6656 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
6659 // check reg + imm case:
6660 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
6661 uint64_t NumBytes = 0;
6662 if (Ty->isSized()) {
6663 uint64_t NumBits = getDataLayout()->getTypeSizeInBits(Ty);
6664 NumBytes = NumBits / 8;
6665 if (!isPowerOf2_64(NumBits))
6670 int64_t Offset = AM.BaseOffs;
6672 // 9-bit signed offset
6673 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
6676 // 12-bit unsigned offset
6677 unsigned shift = Log2_64(NumBytes);
6678 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
6679 // Must be a multiple of NumBytes (NumBytes is a power of 2)
6680 (Offset >> shift) << shift == Offset)
6685 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
6687 if (!AM.Scale || AM.Scale == 1 ||
6688 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
6693 int AArch64TargetLowering::getScalingFactorCost(const AddrMode &AM,
6695 // Scaling factors are not free at all.
6696 // Operands | Rt Latency
6697 // -------------------------------------------
6699 // -------------------------------------------
6700 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
6701 // Rt, [Xn, Wm, <extend> #imm] |
6702 if (isLegalAddressingMode(AM, Ty))
6703 // Scale represents reg2 * scale, thus account for 1 if
6704 // it is not equal to 0 or 1.
6705 return AM.Scale != 0 && AM.Scale != 1;
6709 bool AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
6710 VT = VT.getScalarType();
6715 switch (VT.getSimpleVT().SimpleTy) {
6727 AArch64TargetLowering::getScratchRegisters(CallingConv::ID) const {
6728 // LR is a callee-save register, but we must treat it as clobbered by any call
6729 // site. Hence we include LR in the scratch registers, which are in turn added
6730 // as implicit-defs for stackmaps and patchpoints.
6731 static const MCPhysReg ScratchRegs[] = {
6732 AArch64::X16, AArch64::X17, AArch64::LR, 0
6738 AArch64TargetLowering::isDesirableToCommuteWithShift(const SDNode *N) const {
6739 EVT VT = N->getValueType(0);
6740 // If N is unsigned bit extraction: ((x >> C) & mask), then do not combine
6741 // it with shift to let it be lowered to UBFX.
6742 if (N->getOpcode() == ISD::AND && (VT == MVT::i32 || VT == MVT::i64) &&
6743 isa<ConstantSDNode>(N->getOperand(1))) {
6744 uint64_t TruncMask = N->getConstantOperandVal(1);
6745 if (isMask_64(TruncMask) &&
6746 N->getOperand(0).getOpcode() == ISD::SRL &&
6747 isa<ConstantSDNode>(N->getOperand(0)->getOperand(1)))
6753 bool AArch64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
6755 assert(Ty->isIntegerTy());
6757 unsigned BitSize = Ty->getPrimitiveSizeInBits();
6761 int64_t Val = Imm.getSExtValue();
6762 if (Val == 0 || AArch64_AM::isLogicalImmediate(Val, BitSize))
6765 if ((int64_t)Val < 0)
6768 Val &= (1LL << 32) - 1;
6770 unsigned LZ = countLeadingZeros((uint64_t)Val);
6771 unsigned Shift = (63 - LZ) / 16;
6772 // MOVZ is free so return true for one or fewer MOVK.
6773 return (Shift < 3) ? true : false;
6776 // Generate SUBS and CSEL for integer abs.
6777 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
6778 EVT VT = N->getValueType(0);
6780 SDValue N0 = N->getOperand(0);
6781 SDValue N1 = N->getOperand(1);
6784 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
6785 // and change it to SUB and CSEL.
6786 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
6787 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
6788 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
6789 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
6790 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
6791 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
6793 // Generate SUBS & CSEL.
6795 DAG.getNode(AArch64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
6796 N0.getOperand(0), DAG.getConstant(0, VT));
6797 return DAG.getNode(AArch64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
6798 DAG.getConstant(AArch64CC::PL, MVT::i32),
6799 SDValue(Cmp.getNode(), 1));
6804 // performXorCombine - Attempts to handle integer ABS.
6805 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
6806 TargetLowering::DAGCombinerInfo &DCI,
6807 const AArch64Subtarget *Subtarget) {
6808 if (DCI.isBeforeLegalizeOps())
6811 return performIntegerAbsCombine(N, DAG);
6815 AArch64TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
6817 std::vector<SDNode *> *Created) const {
6818 // fold (sdiv X, pow2)
6819 EVT VT = N->getValueType(0);
6820 if ((VT != MVT::i32 && VT != MVT::i64) ||
6821 !(Divisor.isPowerOf2() || (-Divisor).isPowerOf2()))
6825 SDValue N0 = N->getOperand(0);
6826 unsigned Lg2 = Divisor.countTrailingZeros();
6827 SDValue Zero = DAG.getConstant(0, VT);
6828 SDValue Pow2MinusOne = DAG.getConstant((1ULL << Lg2) - 1, VT);
6830 // Add (N0 < 0) ? Pow2 - 1 : 0;
6832 SDValue Cmp = getAArch64Cmp(N0, Zero, ISD::SETLT, CCVal, DAG, DL);
6833 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Pow2MinusOne);
6834 SDValue CSel = DAG.getNode(AArch64ISD::CSEL, DL, VT, Add, N0, CCVal, Cmp);
6837 Created->push_back(Cmp.getNode());
6838 Created->push_back(Add.getNode());
6839 Created->push_back(CSel.getNode());
6844 DAG.getNode(ISD::SRA, DL, VT, CSel, DAG.getConstant(Lg2, MVT::i64));
6846 // If we're dividing by a positive value, we're done. Otherwise, we must
6847 // negate the result.
6848 if (Divisor.isNonNegative())
6852 Created->push_back(SRA.getNode());
6853 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT), SRA);
6856 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
6857 TargetLowering::DAGCombinerInfo &DCI,
6858 const AArch64Subtarget *Subtarget) {
6859 if (DCI.isBeforeLegalizeOps())
6862 // Multiplication of a power of two plus/minus one can be done more
6863 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
6864 // future CPUs have a cheaper MADD instruction, this may need to be
6865 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
6866 // 64-bit is 5 cycles, so this is always a win.
6867 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
6868 APInt Value = C->getAPIntValue();
6869 EVT VT = N->getValueType(0);
6870 if (Value.isNonNegative()) {
6871 // (mul x, 2^N + 1) => (add (shl x, N), x)
6872 APInt VM1 = Value - 1;
6873 if (VM1.isPowerOf2()) {
6874 SDValue ShiftedVal =
6875 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6876 DAG.getConstant(VM1.logBase2(), MVT::i64));
6877 return DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal,
6880 // (mul x, 2^N - 1) => (sub (shl x, N), x)
6881 APInt VP1 = Value + 1;
6882 if (VP1.isPowerOf2()) {
6883 SDValue ShiftedVal =
6884 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6885 DAG.getConstant(VP1.logBase2(), MVT::i64));
6886 return DAG.getNode(ISD::SUB, SDLoc(N), VT, ShiftedVal,
6890 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
6891 APInt VNM1 = -Value - 1;
6892 if (VNM1.isPowerOf2()) {
6893 SDValue ShiftedVal =
6894 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6895 DAG.getConstant(VNM1.logBase2(), MVT::i64));
6897 DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6898 return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), Add);
6900 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
6901 APInt VNP1 = -Value + 1;
6902 if (VNP1.isPowerOf2()) {
6903 SDValue ShiftedVal =
6904 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6905 DAG.getConstant(VNP1.logBase2(), MVT::i64));
6906 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N->getOperand(0),
6914 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
6915 SelectionDAG &DAG) {
6916 // Take advantage of vector comparisons producing 0 or -1 in each lane to
6917 // optimize away operation when it's from a constant.
6919 // The general transformation is:
6920 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
6921 // AND(VECTOR_CMP(x,y), constant2)
6922 // constant2 = UNARYOP(constant)
6924 // Early exit if this isn't a vector operation, the operand of the
6925 // unary operation isn't a bitwise AND, or if the sizes of the operations
6927 EVT VT = N->getValueType(0);
6928 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
6929 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
6930 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
6933 // Now check that the other operand of the AND is a constant. We could
6934 // make the transformation for non-constant splats as well, but it's unclear
6935 // that would be a benefit as it would not eliminate any operations, just
6936 // perform one more step in scalar code before moving to the vector unit.
6937 if (BuildVectorSDNode *BV =
6938 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
6939 // Bail out if the vector isn't a constant.
6940 if (!BV->isConstant())
6943 // Everything checks out. Build up the new and improved node.
6945 EVT IntVT = BV->getValueType(0);
6946 // Create a new constant of the appropriate type for the transformed
6948 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
6949 // The AND node needs bitcasts to/from an integer vector type around it.
6950 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
6951 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
6952 N->getOperand(0)->getOperand(0), MaskConst);
6953 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
6960 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG,
6961 const AArch64Subtarget *Subtarget) {
6962 // First try to optimize away the conversion when it's conditionally from
6963 // a constant. Vectors only.
6964 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
6965 if (Res != SDValue())
6968 EVT VT = N->getValueType(0);
6969 if (VT != MVT::f32 && VT != MVT::f64)
6972 // Only optimize when the source and destination types have the same width.
6973 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
6976 // If the result of an integer load is only used by an integer-to-float
6977 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
6978 // This eliminates an "integer-to-vector-move UOP and improve throughput.
6979 SDValue N0 = N->getOperand(0);
6980 if (Subtarget->hasNEON() && ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6981 // Do not change the width of a volatile load.
6982 !cast<LoadSDNode>(N0)->isVolatile()) {
6983 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6984 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
6985 LN0->getPointerInfo(), LN0->isVolatile(),
6986 LN0->isNonTemporal(), LN0->isInvariant(),
6987 LN0->getAlignment());
6989 // Make sure successors of the original load stay after it by updating them
6990 // to use the new Chain.
6991 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
6994 (N->getOpcode() == ISD::SINT_TO_FP) ? AArch64ISD::SITOF : AArch64ISD::UITOF;
6995 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
7001 /// An EXTR instruction is made up of two shifts, ORed together. This helper
7002 /// searches for and classifies those shifts.
7003 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
7005 if (N.getOpcode() == ISD::SHL)
7007 else if (N.getOpcode() == ISD::SRL)
7012 if (!isa<ConstantSDNode>(N.getOperand(1)))
7015 ShiftAmount = N->getConstantOperandVal(1);
7016 Src = N->getOperand(0);
7020 /// EXTR instruction extracts a contiguous chunk of bits from two existing
7021 /// registers viewed as a high/low pair. This function looks for the pattern:
7022 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
7023 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
7025 static SDValue tryCombineToEXTR(SDNode *N,
7026 TargetLowering::DAGCombinerInfo &DCI) {
7027 SelectionDAG &DAG = DCI.DAG;
7029 EVT VT = N->getValueType(0);
7031 assert(N->getOpcode() == ISD::OR && "Unexpected root");
7033 if (VT != MVT::i32 && VT != MVT::i64)
7037 uint32_t ShiftLHS = 0;
7039 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
7043 uint32_t ShiftRHS = 0;
7045 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
7048 // If they're both trying to come from the high part of the register, they're
7049 // not really an EXTR.
7050 if (LHSFromHi == RHSFromHi)
7053 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
7057 std::swap(LHS, RHS);
7058 std::swap(ShiftLHS, ShiftRHS);
7061 return DAG.getNode(AArch64ISD::EXTR, DL, VT, LHS, RHS,
7062 DAG.getConstant(ShiftRHS, MVT::i64));
7065 static SDValue tryCombineToBSL(SDNode *N,
7066 TargetLowering::DAGCombinerInfo &DCI) {
7067 EVT VT = N->getValueType(0);
7068 SelectionDAG &DAG = DCI.DAG;
7074 SDValue N0 = N->getOperand(0);
7075 if (N0.getOpcode() != ISD::AND)
7078 SDValue N1 = N->getOperand(1);
7079 if (N1.getOpcode() != ISD::AND)
7082 // We only have to look for constant vectors here since the general, variable
7083 // case can be handled in TableGen.
7084 unsigned Bits = VT.getVectorElementType().getSizeInBits();
7085 uint64_t BitMask = Bits == 64 ? -1ULL : ((1ULL << Bits) - 1);
7086 for (int i = 1; i >= 0; --i)
7087 for (int j = 1; j >= 0; --j) {
7088 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(i));
7089 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(j));
7093 bool FoundMatch = true;
7094 for (unsigned k = 0; k < VT.getVectorNumElements(); ++k) {
7095 ConstantSDNode *CN0 = dyn_cast<ConstantSDNode>(BVN0->getOperand(k));
7096 ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(BVN1->getOperand(k));
7098 CN0->getZExtValue() != (BitMask & ~CN1->getZExtValue())) {
7105 return DAG.getNode(AArch64ISD::BSL, DL, VT, SDValue(BVN0, 0),
7106 N0->getOperand(1 - i), N1->getOperand(1 - j));
7112 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
7113 const AArch64Subtarget *Subtarget) {
7114 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
7115 if (!EnableAArch64ExtrGeneration)
7117 SelectionDAG &DAG = DCI.DAG;
7118 EVT VT = N->getValueType(0);
7120 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7123 SDValue Res = tryCombineToEXTR(N, DCI);
7127 Res = tryCombineToBSL(N, DCI);
7134 static SDValue performBitcastCombine(SDNode *N,
7135 TargetLowering::DAGCombinerInfo &DCI,
7136 SelectionDAG &DAG) {
7137 // Wait 'til after everything is legalized to try this. That way we have
7138 // legal vector types and such.
7139 if (DCI.isBeforeLegalizeOps())
7142 // Remove extraneous bitcasts around an extract_subvector.
7144 // (v4i16 (bitconvert
7145 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
7147 // (extract_subvector ((v8i16 ...), (i64 4)))
7149 // Only interested in 64-bit vectors as the ultimate result.
7150 EVT VT = N->getValueType(0);
7153 if (VT.getSimpleVT().getSizeInBits() != 64)
7155 // Is the operand an extract_subvector starting at the beginning or halfway
7156 // point of the vector? A low half may also come through as an
7157 // EXTRACT_SUBREG, so look for that, too.
7158 SDValue Op0 = N->getOperand(0);
7159 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
7160 !(Op0->isMachineOpcode() &&
7161 Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG))
7163 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
7164 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
7165 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
7167 } else if (Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG) {
7168 if (idx != AArch64::dsub)
7170 // The dsub reference is equivalent to a lane zero subvector reference.
7173 // Look through the bitcast of the input to the extract.
7174 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
7176 SDValue Source = Op0->getOperand(0)->getOperand(0);
7177 // If the source type has twice the number of elements as our destination
7178 // type, we know this is an extract of the high or low half of the vector.
7179 EVT SVT = Source->getValueType(0);
7180 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
7183 DEBUG(dbgs() << "aarch64-lower: bitcast extract_subvector simplification\n");
7185 // Create the simplified form to just extract the low or high half of the
7186 // vector directly rather than bothering with the bitcasts.
7188 unsigned NumElements = VT.getVectorNumElements();
7190 SDValue HalfIdx = DAG.getConstant(NumElements, MVT::i64);
7191 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
7193 SDValue SubReg = DAG.getTargetConstant(AArch64::dsub, MVT::i32);
7194 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
7200 static SDValue performConcatVectorsCombine(SDNode *N,
7201 TargetLowering::DAGCombinerInfo &DCI,
7202 SelectionDAG &DAG) {
7203 // Wait 'til after everything is legalized to try this. That way we have
7204 // legal vector types and such.
7205 if (DCI.isBeforeLegalizeOps())
7209 EVT VT = N->getValueType(0);
7211 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
7212 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
7213 // canonicalise to that.
7214 if (N->getOperand(0) == N->getOperand(1) && VT.getVectorNumElements() == 2) {
7215 assert(VT.getVectorElementType().getSizeInBits() == 64);
7216 return DAG.getNode(AArch64ISD::DUPLANE64, dl, VT,
7217 WidenVector(N->getOperand(0), DAG),
7218 DAG.getConstant(0, MVT::i64));
7221 // Canonicalise concat_vectors so that the right-hand vector has as few
7222 // bit-casts as possible before its real operation. The primary matching
7223 // destination for these operations will be the narrowing "2" instructions,
7224 // which depend on the operation being performed on this right-hand vector.
7226 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
7228 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
7230 SDValue Op1 = N->getOperand(1);
7231 if (Op1->getOpcode() != ISD::BITCAST)
7233 SDValue RHS = Op1->getOperand(0);
7234 MVT RHSTy = RHS.getValueType().getSimpleVT();
7235 // If the RHS is not a vector, this is not the pattern we're looking for.
7236 if (!RHSTy.isVector())
7239 DEBUG(dbgs() << "aarch64-lower: concat_vectors bitcast simplification\n");
7241 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
7242 RHSTy.getVectorNumElements() * 2);
7244 ISD::BITCAST, dl, VT,
7245 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
7246 DAG.getNode(ISD::BITCAST, dl, RHSTy, N->getOperand(0)), RHS));
7249 static SDValue tryCombineFixedPointConvert(SDNode *N,
7250 TargetLowering::DAGCombinerInfo &DCI,
7251 SelectionDAG &DAG) {
7252 // Wait 'til after everything is legalized to try this. That way we have
7253 // legal vector types and such.
7254 if (DCI.isBeforeLegalizeOps())
7256 // Transform a scalar conversion of a value from a lane extract into a
7257 // lane extract of a vector conversion. E.g., from foo1 to foo2:
7258 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
7259 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
7261 // The second form interacts better with instruction selection and the
7262 // register allocator to avoid cross-class register copies that aren't
7263 // coalescable due to a lane reference.
7265 // Check the operand and see if it originates from a lane extract.
7266 SDValue Op1 = N->getOperand(1);
7267 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7268 // Yep, no additional predication needed. Perform the transform.
7269 SDValue IID = N->getOperand(0);
7270 SDValue Shift = N->getOperand(2);
7271 SDValue Vec = Op1.getOperand(0);
7272 SDValue Lane = Op1.getOperand(1);
7273 EVT ResTy = N->getValueType(0);
7277 // The vector width should be 128 bits by the time we get here, even
7278 // if it started as 64 bits (the extract_vector handling will have
7280 assert(Vec.getValueType().getSizeInBits() == 128 &&
7281 "unexpected vector size on extract_vector_elt!");
7282 if (Vec.getValueType() == MVT::v4i32)
7283 VecResTy = MVT::v4f32;
7284 else if (Vec.getValueType() == MVT::v2i64)
7285 VecResTy = MVT::v2f64;
7287 llvm_unreachable("unexpected vector type!");
7290 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
7291 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
7296 // AArch64 high-vector "long" operations are formed by performing the non-high
7297 // version on an extract_subvector of each operand which gets the high half:
7299 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
7301 // However, there are cases which don't have an extract_high explicitly, but
7302 // have another operation that can be made compatible with one for free. For
7305 // (dupv64 scalar) --> (extract_high (dup128 scalar))
7307 // This routine does the actual conversion of such DUPs, once outer routines
7308 // have determined that everything else is in order.
7309 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
7310 // We can handle most types of duplicate, but the lane ones have an extra
7311 // operand saying *which* lane, so we need to know.
7313 switch (N.getOpcode()) {
7314 case AArch64ISD::DUP:
7317 case AArch64ISD::DUPLANE8:
7318 case AArch64ISD::DUPLANE16:
7319 case AArch64ISD::DUPLANE32:
7320 case AArch64ISD::DUPLANE64:
7327 MVT NarrowTy = N.getSimpleValueType();
7328 if (!NarrowTy.is64BitVector())
7331 MVT ElementTy = NarrowTy.getVectorElementType();
7332 unsigned NumElems = NarrowTy.getVectorNumElements();
7333 MVT NewDUPVT = MVT::getVectorVT(ElementTy, NumElems * 2);
7337 NewDUP = DAG.getNode(N.getOpcode(), SDLoc(N), NewDUPVT, N.getOperand(0),
7340 NewDUP = DAG.getNode(AArch64ISD::DUP, SDLoc(N), NewDUPVT, N.getOperand(0));
7342 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N.getNode()), NarrowTy,
7343 NewDUP, DAG.getConstant(NumElems, MVT::i64));
7346 static bool isEssentiallyExtractSubvector(SDValue N) {
7347 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
7350 return N.getOpcode() == ISD::BITCAST &&
7351 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
7354 /// \brief Helper structure to keep track of ISD::SET_CC operands.
7355 struct GenericSetCCInfo {
7356 const SDValue *Opnd0;
7357 const SDValue *Opnd1;
7361 /// \brief Helper structure to keep track of a SET_CC lowered into AArch64 code.
7362 struct AArch64SetCCInfo {
7364 AArch64CC::CondCode CC;
7367 /// \brief Helper structure to keep track of SetCC information.
7369 GenericSetCCInfo Generic;
7370 AArch64SetCCInfo AArch64;
7373 /// \brief Helper structure to be able to read SetCC information. If set to
7374 /// true, IsAArch64 field, Info is a AArch64SetCCInfo, otherwise Info is a
7375 /// GenericSetCCInfo.
7376 struct SetCCInfoAndKind {
7381 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
7383 /// AArch64 lowered one.
7384 /// \p SetCCInfo is filled accordingly.
7385 /// \post SetCCInfo is meanginfull only when this function returns true.
7386 /// \return True when Op is a kind of SET_CC operation.
7387 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
7388 // If this is a setcc, this is straight forward.
7389 if (Op.getOpcode() == ISD::SETCC) {
7390 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
7391 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
7392 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7393 SetCCInfo.IsAArch64 = false;
7396 // Otherwise, check if this is a matching csel instruction.
7400 if (Op.getOpcode() != AArch64ISD::CSEL)
7402 // Set the information about the operands.
7403 // TODO: we want the operands of the Cmp not the csel
7404 SetCCInfo.Info.AArch64.Cmp = &Op.getOperand(3);
7405 SetCCInfo.IsAArch64 = true;
7406 SetCCInfo.Info.AArch64.CC = static_cast<AArch64CC::CondCode>(
7407 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
7409 // Check that the operands matches the constraints:
7410 // (1) Both operands must be constants.
7411 // (2) One must be 1 and the other must be 0.
7412 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
7413 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7416 if (!TValue || !FValue)
7420 if (!TValue->isOne()) {
7421 // Update the comparison when we are interested in !cc.
7422 std::swap(TValue, FValue);
7423 SetCCInfo.Info.AArch64.CC =
7424 AArch64CC::getInvertedCondCode(SetCCInfo.Info.AArch64.CC);
7426 return TValue->isOne() && FValue->isNullValue();
7429 // Returns true if Op is setcc or zext of setcc.
7430 static bool isSetCCOrZExtSetCC(const SDValue& Op, SetCCInfoAndKind &Info) {
7431 if (isSetCC(Op, Info))
7433 return ((Op.getOpcode() == ISD::ZERO_EXTEND) &&
7434 isSetCC(Op->getOperand(0), Info));
7437 // The folding we want to perform is:
7438 // (add x, [zext] (setcc cc ...) )
7440 // (csel x, (add x, 1), !cc ...)
7442 // The latter will get matched to a CSINC instruction.
7443 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
7444 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
7445 SDValue LHS = Op->getOperand(0);
7446 SDValue RHS = Op->getOperand(1);
7447 SetCCInfoAndKind InfoAndKind;
7449 // If neither operand is a SET_CC, give up.
7450 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind)) {
7451 std::swap(LHS, RHS);
7452 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind))
7456 // FIXME: This could be generatized to work for FP comparisons.
7457 EVT CmpVT = InfoAndKind.IsAArch64
7458 ? InfoAndKind.Info.AArch64.Cmp->getOperand(0).getValueType()
7459 : InfoAndKind.Info.Generic.Opnd0->getValueType();
7460 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
7466 if (InfoAndKind.IsAArch64) {
7467 CCVal = DAG.getConstant(
7468 AArch64CC::getInvertedCondCode(InfoAndKind.Info.AArch64.CC), MVT::i32);
7469 Cmp = *InfoAndKind.Info.AArch64.Cmp;
7471 Cmp = getAArch64Cmp(*InfoAndKind.Info.Generic.Opnd0,
7472 *InfoAndKind.Info.Generic.Opnd1,
7473 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
7476 EVT VT = Op->getValueType(0);
7477 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, VT));
7478 return DAG.getNode(AArch64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
7481 // The basic add/sub long vector instructions have variants with "2" on the end
7482 // which act on the high-half of their inputs. They are normally matched by
7485 // (add (zeroext (extract_high LHS)),
7486 // (zeroext (extract_high RHS)))
7487 // -> uaddl2 vD, vN, vM
7489 // However, if one of the extracts is something like a duplicate, this
7490 // instruction can still be used profitably. This function puts the DAG into a
7491 // more appropriate form for those patterns to trigger.
7492 static SDValue performAddSubLongCombine(SDNode *N,
7493 TargetLowering::DAGCombinerInfo &DCI,
7494 SelectionDAG &DAG) {
7495 if (DCI.isBeforeLegalizeOps())
7498 MVT VT = N->getSimpleValueType(0);
7499 if (!VT.is128BitVector()) {
7500 if (N->getOpcode() == ISD::ADD)
7501 return performSetccAddFolding(N, DAG);
7505 // Make sure both branches are extended in the same way.
7506 SDValue LHS = N->getOperand(0);
7507 SDValue RHS = N->getOperand(1);
7508 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
7509 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
7510 LHS.getOpcode() != RHS.getOpcode())
7513 unsigned ExtType = LHS.getOpcode();
7515 // It's not worth doing if at least one of the inputs isn't already an
7516 // extract, but we don't know which it'll be so we have to try both.
7517 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
7518 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
7522 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
7523 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
7524 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
7528 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
7531 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
7534 // Massage DAGs which we can use the high-half "long" operations on into
7535 // something isel will recognize better. E.g.
7537 // (aarch64_neon_umull (extract_high vec) (dupv64 scalar)) -->
7538 // (aarch64_neon_umull (extract_high (v2i64 vec)))
7539 // (extract_high (v2i64 (dup128 scalar)))))
7541 static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N,
7542 TargetLowering::DAGCombinerInfo &DCI,
7543 SelectionDAG &DAG) {
7544 if (DCI.isBeforeLegalizeOps())
7547 SDValue LHS = N->getOperand(1);
7548 SDValue RHS = N->getOperand(2);
7549 assert(LHS.getValueType().is64BitVector() &&
7550 RHS.getValueType().is64BitVector() &&
7551 "unexpected shape for long operation");
7553 // Either node could be a DUP, but it's not worth doing both of them (you'd
7554 // just as well use the non-high version) so look for a corresponding extract
7555 // operation on the other "wing".
7556 if (isEssentiallyExtractSubvector(LHS)) {
7557 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
7560 } else if (isEssentiallyExtractSubvector(RHS)) {
7561 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
7566 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
7567 N->getOperand(0), LHS, RHS);
7570 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
7571 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
7572 unsigned ElemBits = ElemTy.getSizeInBits();
7574 int64_t ShiftAmount;
7575 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
7576 APInt SplatValue, SplatUndef;
7577 unsigned SplatBitSize;
7579 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
7580 HasAnyUndefs, ElemBits) ||
7581 SplatBitSize != ElemBits)
7584 ShiftAmount = SplatValue.getSExtValue();
7585 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
7586 ShiftAmount = CVN->getSExtValue();
7594 llvm_unreachable("Unknown shift intrinsic");
7595 case Intrinsic::aarch64_neon_sqshl:
7596 Opcode = AArch64ISD::SQSHL_I;
7597 IsRightShift = false;
7599 case Intrinsic::aarch64_neon_uqshl:
7600 Opcode = AArch64ISD::UQSHL_I;
7601 IsRightShift = false;
7603 case Intrinsic::aarch64_neon_srshl:
7604 Opcode = AArch64ISD::SRSHR_I;
7605 IsRightShift = true;
7607 case Intrinsic::aarch64_neon_urshl:
7608 Opcode = AArch64ISD::URSHR_I;
7609 IsRightShift = true;
7611 case Intrinsic::aarch64_neon_sqshlu:
7612 Opcode = AArch64ISD::SQSHLU_I;
7613 IsRightShift = false;
7617 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits)
7618 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
7619 DAG.getConstant(-ShiftAmount, MVT::i32));
7620 else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount < ElemBits)
7621 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
7622 DAG.getConstant(ShiftAmount, MVT::i32));
7627 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
7628 // the intrinsics must be legal and take an i32, this means there's almost
7629 // certainly going to be a zext in the DAG which we can eliminate.
7630 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
7631 SDValue AndN = N->getOperand(2);
7632 if (AndN.getOpcode() != ISD::AND)
7635 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
7636 if (!CMask || CMask->getZExtValue() != Mask)
7639 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
7640 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
7643 static SDValue performIntrinsicCombine(SDNode *N,
7644 TargetLowering::DAGCombinerInfo &DCI,
7645 const AArch64Subtarget *Subtarget) {
7646 SelectionDAG &DAG = DCI.DAG;
7647 unsigned IID = getIntrinsicID(N);
7651 case Intrinsic::aarch64_neon_vcvtfxs2fp:
7652 case Intrinsic::aarch64_neon_vcvtfxu2fp:
7653 return tryCombineFixedPointConvert(N, DCI, DAG);
7655 case Intrinsic::aarch64_neon_fmax:
7656 return DAG.getNode(AArch64ISD::FMAX, SDLoc(N), N->getValueType(0),
7657 N->getOperand(1), N->getOperand(2));
7658 case Intrinsic::aarch64_neon_fmin:
7659 return DAG.getNode(AArch64ISD::FMIN, SDLoc(N), N->getValueType(0),
7660 N->getOperand(1), N->getOperand(2));
7661 case Intrinsic::aarch64_neon_smull:
7662 case Intrinsic::aarch64_neon_umull:
7663 case Intrinsic::aarch64_neon_pmull:
7664 case Intrinsic::aarch64_neon_sqdmull:
7665 return tryCombineLongOpWithDup(IID, N, DCI, DAG);
7666 case Intrinsic::aarch64_neon_sqshl:
7667 case Intrinsic::aarch64_neon_uqshl:
7668 case Intrinsic::aarch64_neon_sqshlu:
7669 case Intrinsic::aarch64_neon_srshl:
7670 case Intrinsic::aarch64_neon_urshl:
7671 return tryCombineShiftImm(IID, N, DAG);
7672 case Intrinsic::aarch64_crc32b:
7673 case Intrinsic::aarch64_crc32cb:
7674 return tryCombineCRC32(0xff, N, DAG);
7675 case Intrinsic::aarch64_crc32h:
7676 case Intrinsic::aarch64_crc32ch:
7677 return tryCombineCRC32(0xffff, N, DAG);
7682 static SDValue performExtendCombine(SDNode *N,
7683 TargetLowering::DAGCombinerInfo &DCI,
7684 SelectionDAG &DAG) {
7685 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
7686 // we can convert that DUP into another extract_high (of a bigger DUP), which
7687 // helps the backend to decide that an sabdl2 would be useful, saving a real
7688 // extract_high operation.
7689 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
7690 N->getOperand(0).getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
7691 SDNode *ABDNode = N->getOperand(0).getNode();
7692 unsigned IID = getIntrinsicID(ABDNode);
7693 if (IID == Intrinsic::aarch64_neon_sabd ||
7694 IID == Intrinsic::aarch64_neon_uabd) {
7695 SDValue NewABD = tryCombineLongOpWithDup(IID, ABDNode, DCI, DAG);
7696 if (!NewABD.getNode())
7699 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
7704 // This is effectively a custom type legalization for AArch64.
7706 // Type legalization will split an extend of a small, legal, type to a larger
7707 // illegal type by first splitting the destination type, often creating
7708 // illegal source types, which then get legalized in isel-confusing ways,
7709 // leading to really terrible codegen. E.g.,
7710 // %result = v8i32 sext v8i8 %value
7712 // %losrc = extract_subreg %value, ...
7713 // %hisrc = extract_subreg %value, ...
7714 // %lo = v4i32 sext v4i8 %losrc
7715 // %hi = v4i32 sext v4i8 %hisrc
7716 // Things go rapidly downhill from there.
7718 // For AArch64, the [sz]ext vector instructions can only go up one element
7719 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
7720 // take two instructions.
7722 // This implies that the most efficient way to do the extend from v8i8
7723 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
7724 // the normal splitting to happen for the v8i16->v8i32.
7726 // This is pre-legalization to catch some cases where the default
7727 // type legalization will create ill-tempered code.
7728 if (!DCI.isBeforeLegalizeOps())
7731 // We're only interested in cleaning things up for non-legal vector types
7732 // here. If both the source and destination are legal, things will just
7733 // work naturally without any fiddling.
7734 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7735 EVT ResVT = N->getValueType(0);
7736 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
7738 // If the vector type isn't a simple VT, it's beyond the scope of what
7739 // we're worried about here. Let legalization do its thing and hope for
7741 SDValue Src = N->getOperand(0);
7742 EVT SrcVT = Src->getValueType(0);
7743 if (!ResVT.isSimple() || !SrcVT.isSimple())
7746 // If the source VT is a 64-bit vector, we can play games and get the
7747 // better results we want.
7748 if (SrcVT.getSizeInBits() != 64)
7751 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
7752 unsigned ElementCount = SrcVT.getVectorNumElements();
7753 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
7755 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
7757 // Now split the rest of the operation into two halves, each with a 64
7761 unsigned NumElements = ResVT.getVectorNumElements();
7762 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
7763 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
7764 ResVT.getVectorElementType(), NumElements / 2);
7766 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
7767 LoVT.getVectorNumElements());
7768 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7769 DAG.getConstant(0, MVT::i64));
7770 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7771 DAG.getConstant(InNVT.getVectorNumElements(), MVT::i64));
7772 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
7773 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
7775 // Now combine the parts back together so we still have a single result
7776 // like the combiner expects.
7777 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
7780 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
7781 /// value. The load store optimizer pass will merge them to store pair stores.
7782 /// This has better performance than a splat of the scalar followed by a split
7783 /// vector store. Even if the stores are not merged it is four stores vs a dup,
7784 /// followed by an ext.b and two stores.
7785 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
7786 SDValue StVal = St->getValue();
7787 EVT VT = StVal.getValueType();
7789 // Don't replace floating point stores, they possibly won't be transformed to
7790 // stp because of the store pair suppress pass.
7791 if (VT.isFloatingPoint())
7794 // Check for insert vector elements.
7795 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
7798 // We can express a splat as store pair(s) for 2 or 4 elements.
7799 unsigned NumVecElts = VT.getVectorNumElements();
7800 if (NumVecElts != 4 && NumVecElts != 2)
7802 SDValue SplatVal = StVal.getOperand(1);
7803 unsigned RemainInsertElts = NumVecElts - 1;
7805 // Check that this is a splat.
7806 while (--RemainInsertElts) {
7807 SDValue NextInsertElt = StVal.getOperand(0);
7808 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
7810 if (NextInsertElt.getOperand(1) != SplatVal)
7812 StVal = NextInsertElt;
7814 unsigned OrigAlignment = St->getAlignment();
7815 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
7816 unsigned Alignment = std::min(OrigAlignment, EltOffset);
7818 // Create scalar stores. This is at least as good as the code sequence for a
7819 // split unaligned store wich is a dup.s, ext.b, and two stores.
7820 // Most of the time the three stores should be replaced by store pair
7821 // instructions (stp).
7823 SDValue BasePtr = St->getBasePtr();
7825 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
7826 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
7828 unsigned Offset = EltOffset;
7829 while (--NumVecElts) {
7830 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7831 DAG.getConstant(Offset, MVT::i64));
7832 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
7833 St->getPointerInfo(), St->isVolatile(),
7834 St->isNonTemporal(), Alignment);
7835 Offset += EltOffset;
7840 static SDValue performSTORECombine(SDNode *N,
7841 TargetLowering::DAGCombinerInfo &DCI,
7843 const AArch64Subtarget *Subtarget) {
7844 if (!DCI.isBeforeLegalize())
7847 StoreSDNode *S = cast<StoreSDNode>(N);
7848 if (S->isVolatile())
7851 // Cyclone has bad performance on unaligned 16B stores when crossing line and
7852 // page boundries. We want to split such stores.
7853 if (!Subtarget->isCyclone())
7856 // Don't split at Oz.
7857 MachineFunction &MF = DAG.getMachineFunction();
7858 bool IsMinSize = MF.getFunction()->getAttributes().hasAttribute(
7859 AttributeSet::FunctionIndex, Attribute::MinSize);
7863 SDValue StVal = S->getValue();
7864 EVT VT = StVal.getValueType();
7866 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
7867 // those up regresses performance on micro-benchmarks and olden/bh.
7868 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
7871 // Split unaligned 16B stores. They are terrible for performance.
7872 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
7873 // extensions can use this to mark that it does not want splitting to happen
7874 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
7875 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
7876 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
7877 S->getAlignment() <= 2)
7880 // If we get a splat of a scalar convert this vector store to a store of
7881 // scalars. They will be merged into store pairs thereby removing two
7883 SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S);
7884 if (ReplacedSplat != SDValue())
7885 return ReplacedSplat;
7888 unsigned NumElts = VT.getVectorNumElements() / 2;
7889 // Split VT into two.
7891 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
7892 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7893 DAG.getConstant(0, MVT::i64));
7894 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7895 DAG.getConstant(NumElts, MVT::i64));
7896 SDValue BasePtr = S->getBasePtr();
7898 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
7899 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
7900 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7901 DAG.getConstant(8, MVT::i64));
7902 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
7903 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
7907 /// Target-specific DAG combine function for post-increment LD1 (lane) and
7908 /// post-increment LD1R.
7909 static SDValue performPostLD1Combine(SDNode *N,
7910 TargetLowering::DAGCombinerInfo &DCI,
7912 if (DCI.isBeforeLegalizeOps())
7915 SelectionDAG &DAG = DCI.DAG;
7916 EVT VT = N->getValueType(0);
7918 unsigned LoadIdx = IsLaneOp ? 1 : 0;
7919 SDNode *LD = N->getOperand(LoadIdx).getNode();
7920 // If it is not LOAD, can not do such combine.
7921 if (LD->getOpcode() != ISD::LOAD)
7924 LoadSDNode *LoadSDN = cast<LoadSDNode>(LD);
7925 EVT MemVT = LoadSDN->getMemoryVT();
7926 // Check if memory operand is the same type as the vector element.
7927 if (MemVT != VT.getVectorElementType())
7930 // Check if there are other uses. If so, do not combine as it will introduce
7932 for (SDNode::use_iterator UI = LD->use_begin(), UE = LD->use_end(); UI != UE;
7934 if (UI.getUse().getResNo() == 1) // Ignore uses of the chain result.
7940 SDValue Addr = LD->getOperand(1);
7941 SDValue Vector = N->getOperand(0);
7942 // Search for a use of the address operand that is an increment.
7943 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), UE =
7944 Addr.getNode()->use_end(); UI != UE; ++UI) {
7946 if (User->getOpcode() != ISD::ADD
7947 || UI.getUse().getResNo() != Addr.getResNo())
7950 // Check that the add is independent of the load. Otherwise, folding it
7951 // would create a cycle.
7952 if (User->isPredecessorOf(LD) || LD->isPredecessorOf(User))
7954 // Also check that add is not used in the vector operand. This would also
7956 if (User->isPredecessorOf(Vector.getNode()))
7959 // If the increment is a constant, it must match the memory ref size.
7960 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7961 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7962 uint32_t IncVal = CInc->getZExtValue();
7963 unsigned NumBytes = VT.getScalarSizeInBits() / 8;
7964 if (IncVal != NumBytes)
7966 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
7969 SmallVector<SDValue, 8> Ops;
7970 Ops.push_back(LD->getOperand(0)); // Chain
7972 Ops.push_back(Vector); // The vector to be inserted
7973 Ops.push_back(N->getOperand(2)); // The lane to be inserted in the vector
7975 Ops.push_back(Addr);
7978 EVT Tys[3] = { VT, MVT::i64, MVT::Other };
7979 SDVTList SDTys = DAG.getVTList(Tys);
7980 unsigned NewOp = IsLaneOp ? AArch64ISD::LD1LANEpost : AArch64ISD::LD1DUPpost;
7981 SDValue UpdN = DAG.getMemIntrinsicNode(NewOp, SDLoc(N), SDTys, Ops,
7983 LoadSDN->getMemOperand());
7986 std::vector<SDValue> NewResults;
7987 NewResults.push_back(SDValue(LD, 0)); // The result of load
7988 NewResults.push_back(SDValue(UpdN.getNode(), 2)); // Chain
7989 DCI.CombineTo(LD, NewResults);
7990 DCI.CombineTo(N, SDValue(UpdN.getNode(), 0)); // Dup/Inserted Result
7991 DCI.CombineTo(User, SDValue(UpdN.getNode(), 1)); // Write back register
7998 /// Target-specific DAG combine function for NEON load/store intrinsics
7999 /// to merge base address updates.
8000 static SDValue performNEONPostLDSTCombine(SDNode *N,
8001 TargetLowering::DAGCombinerInfo &DCI,
8002 SelectionDAG &DAG) {
8003 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8006 unsigned AddrOpIdx = N->getNumOperands() - 1;
8007 SDValue Addr = N->getOperand(AddrOpIdx);
8009 // Search for a use of the address operand that is an increment.
8010 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8011 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8013 if (User->getOpcode() != ISD::ADD ||
8014 UI.getUse().getResNo() != Addr.getResNo())
8017 // Check that the add is independent of the load/store. Otherwise, folding
8018 // it would create a cycle.
8019 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8022 // Find the new opcode for the updating load/store.
8023 bool IsStore = false;
8024 bool IsLaneOp = false;
8025 bool IsDupOp = false;
8026 unsigned NewOpc = 0;
8027 unsigned NumVecs = 0;
8028 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8030 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8031 case Intrinsic::aarch64_neon_ld2: NewOpc = AArch64ISD::LD2post;
8033 case Intrinsic::aarch64_neon_ld3: NewOpc = AArch64ISD::LD3post;
8035 case Intrinsic::aarch64_neon_ld4: NewOpc = AArch64ISD::LD4post;
8037 case Intrinsic::aarch64_neon_st2: NewOpc = AArch64ISD::ST2post;
8038 NumVecs = 2; IsStore = true; break;
8039 case Intrinsic::aarch64_neon_st3: NewOpc = AArch64ISD::ST3post;
8040 NumVecs = 3; IsStore = true; break;
8041 case Intrinsic::aarch64_neon_st4: NewOpc = AArch64ISD::ST4post;
8042 NumVecs = 4; IsStore = true; break;
8043 case Intrinsic::aarch64_neon_ld1x2: NewOpc = AArch64ISD::LD1x2post;
8045 case Intrinsic::aarch64_neon_ld1x3: NewOpc = AArch64ISD::LD1x3post;
8047 case Intrinsic::aarch64_neon_ld1x4: NewOpc = AArch64ISD::LD1x4post;
8049 case Intrinsic::aarch64_neon_st1x2: NewOpc = AArch64ISD::ST1x2post;
8050 NumVecs = 2; IsStore = true; break;
8051 case Intrinsic::aarch64_neon_st1x3: NewOpc = AArch64ISD::ST1x3post;
8052 NumVecs = 3; IsStore = true; break;
8053 case Intrinsic::aarch64_neon_st1x4: NewOpc = AArch64ISD::ST1x4post;
8054 NumVecs = 4; IsStore = true; break;
8055 case Intrinsic::aarch64_neon_ld2r: NewOpc = AArch64ISD::LD2DUPpost;
8056 NumVecs = 2; IsDupOp = true; break;
8057 case Intrinsic::aarch64_neon_ld3r: NewOpc = AArch64ISD::LD3DUPpost;
8058 NumVecs = 3; IsDupOp = true; break;
8059 case Intrinsic::aarch64_neon_ld4r: NewOpc = AArch64ISD::LD4DUPpost;
8060 NumVecs = 4; IsDupOp = true; break;
8061 case Intrinsic::aarch64_neon_ld2lane: NewOpc = AArch64ISD::LD2LANEpost;
8062 NumVecs = 2; IsLaneOp = true; break;
8063 case Intrinsic::aarch64_neon_ld3lane: NewOpc = AArch64ISD::LD3LANEpost;
8064 NumVecs = 3; IsLaneOp = true; break;
8065 case Intrinsic::aarch64_neon_ld4lane: NewOpc = AArch64ISD::LD4LANEpost;
8066 NumVecs = 4; IsLaneOp = true; break;
8067 case Intrinsic::aarch64_neon_st2lane: NewOpc = AArch64ISD::ST2LANEpost;
8068 NumVecs = 2; IsStore = true; IsLaneOp = true; break;
8069 case Intrinsic::aarch64_neon_st3lane: NewOpc = AArch64ISD::ST3LANEpost;
8070 NumVecs = 3; IsStore = true; IsLaneOp = true; break;
8071 case Intrinsic::aarch64_neon_st4lane: NewOpc = AArch64ISD::ST4LANEpost;
8072 NumVecs = 4; IsStore = true; IsLaneOp = true; break;
8077 VecTy = N->getOperand(2).getValueType();
8079 VecTy = N->getValueType(0);
8081 // If the increment is a constant, it must match the memory ref size.
8082 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8083 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8084 uint32_t IncVal = CInc->getZExtValue();
8085 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8086 if (IsLaneOp || IsDupOp)
8087 NumBytes /= VecTy.getVectorNumElements();
8088 if (IncVal != NumBytes)
8090 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
8092 SmallVector<SDValue, 8> Ops;
8093 Ops.push_back(N->getOperand(0)); // Incoming chain
8094 // Load lane and store have vector list as input.
8095 if (IsLaneOp || IsStore)
8096 for (unsigned i = 2; i < AddrOpIdx; ++i)
8097 Ops.push_back(N->getOperand(i));
8098 Ops.push_back(Addr); // Base register
8103 unsigned NumResultVecs = (IsStore ? 0 : NumVecs);
8105 for (n = 0; n < NumResultVecs; ++n)
8107 Tys[n++] = MVT::i64; // Type of write back register
8108 Tys[n] = MVT::Other; // Type of the chain
8109 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs + 2));
8111 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8112 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys, Ops,
8113 MemInt->getMemoryVT(),
8114 MemInt->getMemOperand());
8117 std::vector<SDValue> NewResults;
8118 for (unsigned i = 0; i < NumResultVecs; ++i) {
8119 NewResults.push_back(SDValue(UpdN.getNode(), i));
8121 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1));
8122 DCI.CombineTo(N, NewResults);
8123 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8130 // Checks to see if the value is the prescribed width and returns information
8131 // about its extension mode.
8133 bool checkValueWidth(SDValue V, unsigned width, ISD::LoadExtType &ExtType) {
8134 ExtType = ISD::NON_EXTLOAD;
8135 switch(V.getNode()->getOpcode()) {
8139 LoadSDNode *LoadNode = cast<LoadSDNode>(V.getNode());
8140 if ((LoadNode->getMemoryVT() == MVT::i8 && width == 8)
8141 || (LoadNode->getMemoryVT() == MVT::i16 && width == 16)) {
8142 ExtType = LoadNode->getExtensionType();
8147 case ISD::AssertSext: {
8148 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
8149 if ((TypeNode->getVT() == MVT::i8 && width == 8)
8150 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
8151 ExtType = ISD::SEXTLOAD;
8156 case ISD::AssertZext: {
8157 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
8158 if ((TypeNode->getVT() == MVT::i8 && width == 8)
8159 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
8160 ExtType = ISD::ZEXTLOAD;
8166 case ISD::TargetConstant: {
8167 if (std::abs(cast<ConstantSDNode>(V.getNode())->getSExtValue()) <
8177 // This function does a whole lot of voodoo to determine if the tests are
8178 // equivalent without and with a mask. Essentially what happens is that given a
8181 // +-------------+ +-------------+ +-------------+ +-------------+
8182 // | Input | | AddConstant | | CompConstant| | CC |
8183 // +-------------+ +-------------+ +-------------+ +-------------+
8185 // V V | +----------+
8186 // +-------------+ +----+ | |
8187 // | ADD | |0xff| | |
8188 // +-------------+ +----+ | |
8191 // +-------------+ | |
8193 // +-------------+ | |
8202 // The AND node may be safely removed for some combinations of inputs. In
8203 // particular we need to take into account the extension type of the Input,
8204 // the exact values of AddConstant, CompConstant, and CC, along with the nominal
8205 // width of the input (this can work for any width inputs, the above graph is
8206 // specific to 8 bits.
8208 // The specific equations were worked out by generating output tables for each
8209 // AArch64CC value in terms of and AddConstant (w1), CompConstant(w2). The
8210 // problem was simplified by working with 4 bit inputs, which means we only
8211 // needed to reason about 24 distinct bit patterns: 8 patterns unique to zero
8212 // extension (8,15), 8 patterns unique to sign extensions (-8,-1), and 8
8213 // patterns present in both extensions (0,7). For every distinct set of
8214 // AddConstant and CompConstants bit patterns we can consider the masked and
8215 // unmasked versions to be equivalent if the result of this function is true for
8216 // all 16 distinct bit patterns of for the current extension type of Input (w0).
8219 // and w10, w8, #0x0f
8221 // cset w9, AArch64CC
8223 // cset w11, AArch64CC
8228 // Since the above function shows when the outputs are equivalent it defines
8229 // when it is safe to remove the AND. Unfortunately it only runs on AArch64 and
8230 // would be expensive to run during compiles. The equations below were written
8231 // in a test harness that confirmed they gave equivalent outputs to the above
8232 // for all inputs function, so they can be used determine if the removal is
8235 // isEquivalentMaskless() is the code for testing if the AND can be removed
8236 // factored out of the DAG recognition as the DAG can take several forms.
8239 bool isEquivalentMaskless(unsigned CC, unsigned width,
8240 ISD::LoadExtType ExtType, signed AddConstant,
8241 signed CompConstant) {
8242 // By being careful about our equations and only writing the in term
8243 // symbolic values and well known constants (0, 1, -1, MaxUInt) we can
8244 // make them generally applicable to all bit widths.
8245 signed MaxUInt = (1 << width);
8247 // For the purposes of these comparisons sign extending the type is
8248 // equivalent to zero extending the add and displacing it by half the integer
8249 // width. Provided we are careful and make sure our equations are valid over
8250 // the whole range we can just adjust the input and avoid writing equations
8251 // for sign extended inputs.
8252 if (ExtType == ISD::SEXTLOAD)
8253 AddConstant -= (1 << (width-1));
8257 case AArch64CC::GT: {
8258 if ((AddConstant == 0) ||
8259 (CompConstant == MaxUInt - 1 && AddConstant < 0) ||
8260 (AddConstant >= 0 && CompConstant < 0) ||
8261 (AddConstant <= 0 && CompConstant <= 0 && CompConstant < AddConstant))
8265 case AArch64CC::GE: {
8266 if ((AddConstant == 0) ||
8267 (AddConstant >= 0 && CompConstant <= 0) ||
8268 (AddConstant <= 0 && CompConstant <= 0 && CompConstant <= AddConstant))
8272 case AArch64CC::LS: {
8273 if ((AddConstant >= 0 && CompConstant < 0) ||
8274 (AddConstant <= 0 && CompConstant >= -1 &&
8275 CompConstant < AddConstant + MaxUInt))
8279 case AArch64CC::MI: {
8280 if ((AddConstant == 0) ||
8281 (AddConstant > 0 && CompConstant <= 0) ||
8282 (AddConstant < 0 && CompConstant <= AddConstant))
8286 case AArch64CC::HS: {
8287 if ((AddConstant >= 0 && CompConstant <= 0) ||
8288 (AddConstant <= 0 && CompConstant >= 0 &&
8289 CompConstant <= AddConstant + MaxUInt))
8293 case AArch64CC::NE: {
8294 if ((AddConstant > 0 && CompConstant < 0) ||
8295 (AddConstant < 0 && CompConstant >= 0 &&
8296 CompConstant < AddConstant + MaxUInt) ||
8297 (AddConstant >= 0 && CompConstant >= 0 &&
8298 CompConstant >= AddConstant) ||
8299 (AddConstant <= 0 && CompConstant < 0 && CompConstant < AddConstant))
8308 case AArch64CC::Invalid:
8316 SDValue performCONDCombine(SDNode *N,
8317 TargetLowering::DAGCombinerInfo &DCI,
8318 SelectionDAG &DAG, unsigned CCIndex,
8319 unsigned CmpIndex) {
8320 unsigned CC = cast<ConstantSDNode>(N->getOperand(CCIndex))->getSExtValue();
8321 SDNode *SubsNode = N->getOperand(CmpIndex).getNode();
8322 unsigned CondOpcode = SubsNode->getOpcode();
8324 if (CondOpcode != AArch64ISD::SUBS)
8327 // There is a SUBS feeding this condition. Is it fed by a mask we can
8330 SDNode *AndNode = SubsNode->getOperand(0).getNode();
8331 unsigned MaskBits = 0;
8333 if (AndNode->getOpcode() != ISD::AND)
8336 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(AndNode->getOperand(1))) {
8337 uint32_t CNV = CN->getZExtValue();
8340 else if (CNV == 65535)
8347 SDValue AddValue = AndNode->getOperand(0);
8349 if (AddValue.getOpcode() != ISD::ADD)
8352 // The basic dag structure is correct, grab the inputs and validate them.
8354 SDValue AddInputValue1 = AddValue.getNode()->getOperand(0);
8355 SDValue AddInputValue2 = AddValue.getNode()->getOperand(1);
8356 SDValue SubsInputValue = SubsNode->getOperand(1);
8358 // The mask is present and the provenance of all the values is a smaller type,
8359 // lets see if the mask is superfluous.
8361 if (!isa<ConstantSDNode>(AddInputValue2.getNode()) ||
8362 !isa<ConstantSDNode>(SubsInputValue.getNode()))
8365 ISD::LoadExtType ExtType;
8367 if (!checkValueWidth(SubsInputValue, MaskBits, ExtType) ||
8368 !checkValueWidth(AddInputValue2, MaskBits, ExtType) ||
8369 !checkValueWidth(AddInputValue1, MaskBits, ExtType) )
8372 if(!isEquivalentMaskless(CC, MaskBits, ExtType,
8373 cast<ConstantSDNode>(AddInputValue2.getNode())->getSExtValue(),
8374 cast<ConstantSDNode>(SubsInputValue.getNode())->getSExtValue()))
8377 // The AND is not necessary, remove it.
8379 SDVTList VTs = DAG.getVTList(SubsNode->getValueType(0),
8380 SubsNode->getValueType(1));
8381 SDValue Ops[] = { AddValue, SubsNode->getOperand(1) };
8383 SDValue NewValue = DAG.getNode(CondOpcode, SDLoc(SubsNode), VTs, Ops);
8384 DAG.ReplaceAllUsesWith(SubsNode, NewValue.getNode());
8386 return SDValue(N, 0);
8389 // Optimize compare with zero and branch.
8390 static SDValue performBRCONDCombine(SDNode *N,
8391 TargetLowering::DAGCombinerInfo &DCI,
8392 SelectionDAG &DAG) {
8393 SDValue NV = performCONDCombine(N, DCI, DAG, 2, 3);
8396 SDValue Chain = N->getOperand(0);
8397 SDValue Dest = N->getOperand(1);
8398 SDValue CCVal = N->getOperand(2);
8399 SDValue Cmp = N->getOperand(3);
8401 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
8402 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
8403 if (CC != AArch64CC::EQ && CC != AArch64CC::NE)
8406 unsigned CmpOpc = Cmp.getOpcode();
8407 if (CmpOpc != AArch64ISD::ADDS && CmpOpc != AArch64ISD::SUBS)
8410 // Only attempt folding if there is only one use of the flag and no use of the
8412 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
8415 SDValue LHS = Cmp.getOperand(0);
8416 SDValue RHS = Cmp.getOperand(1);
8418 assert(LHS.getValueType() == RHS.getValueType() &&
8419 "Expected the value type to be the same for both operands!");
8420 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
8423 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
8424 std::swap(LHS, RHS);
8426 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
8429 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
8430 LHS.getOpcode() == ISD::SRL)
8433 // Fold the compare into the branch instruction.
8435 if (CC == AArch64CC::EQ)
8436 BR = DAG.getNode(AArch64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
8438 BR = DAG.getNode(AArch64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
8440 // Do not add new nodes to DAG combiner worklist.
8441 DCI.CombineTo(N, BR, false);
8446 // vselect (v1i1 setcc) ->
8447 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
8448 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
8449 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
8451 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
8452 SDValue N0 = N->getOperand(0);
8453 EVT CCVT = N0.getValueType();
8455 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
8456 CCVT.getVectorElementType() != MVT::i1)
8459 EVT ResVT = N->getValueType(0);
8460 EVT CmpVT = N0.getOperand(0).getValueType();
8461 // Only combine when the result type is of the same size as the compared
8463 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
8466 SDValue IfTrue = N->getOperand(1);
8467 SDValue IfFalse = N->getOperand(2);
8469 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
8470 N0.getOperand(0), N0.getOperand(1),
8471 cast<CondCodeSDNode>(N0.getOperand(2))->get());
8472 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
8476 /// A vector select: "(select vL, vR, (setcc LHS, RHS))" is best performed with
8477 /// the compare-mask instructions rather than going via NZCV, even if LHS and
8478 /// RHS are really scalar. This replaces any scalar setcc in the above pattern
8479 /// with a vector one followed by a DUP shuffle on the result.
8480 static SDValue performSelectCombine(SDNode *N, SelectionDAG &DAG) {
8481 SDValue N0 = N->getOperand(0);
8482 EVT ResVT = N->getValueType(0);
8484 if (N0.getOpcode() != ISD::SETCC || N0.getValueType() != MVT::i1)
8487 // If NumMaskElts == 0, the comparison is larger than select result. The
8488 // largest real NEON comparison is 64-bits per lane, which means the result is
8489 // at most 32-bits and an illegal vector. Just bail out for now.
8490 EVT SrcVT = N0.getOperand(0).getValueType();
8492 // Don't try to do this optimization when the setcc itself has i1 operands.
8493 // There are no legal vectors of i1, so this would be pointless.
8494 if (SrcVT == MVT::i1)
8497 int NumMaskElts = ResVT.getSizeInBits() / SrcVT.getSizeInBits();
8498 if (!ResVT.isVector() || NumMaskElts == 0)
8501 SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumMaskElts);
8502 EVT CCVT = SrcVT.changeVectorElementTypeToInteger();
8504 // First perform a vector comparison, where lane 0 is the one we're interested
8508 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(0));
8510 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(1));
8511 SDValue SetCC = DAG.getNode(ISD::SETCC, DL, CCVT, LHS, RHS, N0.getOperand(2));
8513 // Now duplicate the comparison mask we want across all other lanes.
8514 SmallVector<int, 8> DUPMask(CCVT.getVectorNumElements(), 0);
8515 SDValue Mask = DAG.getVectorShuffle(CCVT, DL, SetCC, SetCC, DUPMask.data());
8516 Mask = DAG.getNode(ISD::BITCAST, DL,
8517 ResVT.changeVectorElementTypeToInteger(), Mask);
8519 return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2));
8522 SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N,
8523 DAGCombinerInfo &DCI) const {
8524 SelectionDAG &DAG = DCI.DAG;
8525 switch (N->getOpcode()) {
8530 return performAddSubLongCombine(N, DCI, DAG);
8532 return performXorCombine(N, DAG, DCI, Subtarget);
8534 return performMulCombine(N, DAG, DCI, Subtarget);
8535 case ISD::SINT_TO_FP:
8536 case ISD::UINT_TO_FP:
8537 return performIntToFpCombine(N, DAG, Subtarget);
8539 return performORCombine(N, DCI, Subtarget);
8540 case ISD::INTRINSIC_WO_CHAIN:
8541 return performIntrinsicCombine(N, DCI, Subtarget);
8542 case ISD::ANY_EXTEND:
8543 case ISD::ZERO_EXTEND:
8544 case ISD::SIGN_EXTEND:
8545 return performExtendCombine(N, DCI, DAG);
8547 return performBitcastCombine(N, DCI, DAG);
8548 case ISD::CONCAT_VECTORS:
8549 return performConcatVectorsCombine(N, DCI, DAG);
8551 return performSelectCombine(N, DAG);
8553 return performVSelectCombine(N, DCI.DAG);
8555 return performSTORECombine(N, DCI, DAG, Subtarget);
8556 case AArch64ISD::BRCOND:
8557 return performBRCONDCombine(N, DCI, DAG);
8558 case AArch64ISD::CSEL:
8559 return performCONDCombine(N, DCI, DAG, 2, 3);
8560 case AArch64ISD::DUP:
8561 return performPostLD1Combine(N, DCI, false);
8562 case ISD::INSERT_VECTOR_ELT:
8563 return performPostLD1Combine(N, DCI, true);
8564 case ISD::INTRINSIC_VOID:
8565 case ISD::INTRINSIC_W_CHAIN:
8566 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8567 case Intrinsic::aarch64_neon_ld2:
8568 case Intrinsic::aarch64_neon_ld3:
8569 case Intrinsic::aarch64_neon_ld4:
8570 case Intrinsic::aarch64_neon_ld1x2:
8571 case Intrinsic::aarch64_neon_ld1x3:
8572 case Intrinsic::aarch64_neon_ld1x4:
8573 case Intrinsic::aarch64_neon_ld2lane:
8574 case Intrinsic::aarch64_neon_ld3lane:
8575 case Intrinsic::aarch64_neon_ld4lane:
8576 case Intrinsic::aarch64_neon_ld2r:
8577 case Intrinsic::aarch64_neon_ld3r:
8578 case Intrinsic::aarch64_neon_ld4r:
8579 case Intrinsic::aarch64_neon_st2:
8580 case Intrinsic::aarch64_neon_st3:
8581 case Intrinsic::aarch64_neon_st4:
8582 case Intrinsic::aarch64_neon_st1x2:
8583 case Intrinsic::aarch64_neon_st1x3:
8584 case Intrinsic::aarch64_neon_st1x4:
8585 case Intrinsic::aarch64_neon_st2lane:
8586 case Intrinsic::aarch64_neon_st3lane:
8587 case Intrinsic::aarch64_neon_st4lane:
8588 return performNEONPostLDSTCombine(N, DCI, DAG);
8596 // Check if the return value is used as only a return value, as otherwise
8597 // we can't perform a tail-call. In particular, we need to check for
8598 // target ISD nodes that are returns and any other "odd" constructs
8599 // that the generic analysis code won't necessarily catch.
8600 bool AArch64TargetLowering::isUsedByReturnOnly(SDNode *N,
8601 SDValue &Chain) const {
8602 if (N->getNumValues() != 1)
8604 if (!N->hasNUsesOfValue(1, 0))
8607 SDValue TCChain = Chain;
8608 SDNode *Copy = *N->use_begin();
8609 if (Copy->getOpcode() == ISD::CopyToReg) {
8610 // If the copy has a glue operand, we conservatively assume it isn't safe to
8611 // perform a tail call.
8612 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
8615 TCChain = Copy->getOperand(0);
8616 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
8619 bool HasRet = false;
8620 for (SDNode *Node : Copy->uses()) {
8621 if (Node->getOpcode() != AArch64ISD::RET_FLAG)
8633 // Return whether the an instruction can potentially be optimized to a tail
8634 // call. This will cause the optimizers to attempt to move, or duplicate,
8635 // return instructions to help enable tail call optimizations for this
8637 bool AArch64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
8638 if (!CI->isTailCall())
8644 bool AArch64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
8646 ISD::MemIndexedMode &AM,
8648 SelectionDAG &DAG) const {
8649 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
8652 Base = Op->getOperand(0);
8653 // All of the indexed addressing mode instructions take a signed
8654 // 9 bit immediate offset.
8655 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
8656 int64_t RHSC = (int64_t)RHS->getZExtValue();
8657 if (RHSC >= 256 || RHSC <= -256)
8659 IsInc = (Op->getOpcode() == ISD::ADD);
8660 Offset = Op->getOperand(1);
8666 bool AArch64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
8668 ISD::MemIndexedMode &AM,
8669 SelectionDAG &DAG) const {
8672 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8673 VT = LD->getMemoryVT();
8674 Ptr = LD->getBasePtr();
8675 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8676 VT = ST->getMemoryVT();
8677 Ptr = ST->getBasePtr();
8682 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
8684 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
8688 bool AArch64TargetLowering::getPostIndexedAddressParts(
8689 SDNode *N, SDNode *Op, SDValue &Base, SDValue &Offset,
8690 ISD::MemIndexedMode &AM, SelectionDAG &DAG) const {
8693 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8694 VT = LD->getMemoryVT();
8695 Ptr = LD->getBasePtr();
8696 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8697 VT = ST->getMemoryVT();
8698 Ptr = ST->getBasePtr();
8703 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
8705 // Post-indexing updates the base, so it's not a valid transform
8706 // if that's not the same as the load's pointer.
8709 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
8713 static void ReplaceBITCASTResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
8714 SelectionDAG &DAG) {
8716 SDValue Op = N->getOperand(0);
8718 if (N->getValueType(0) != MVT::i16 || Op.getValueType() != MVT::f16)
8722 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL, MVT::f32,
8723 DAG.getUNDEF(MVT::i32), Op,
8724 DAG.getTargetConstant(AArch64::hsub, MVT::i32)),
8726 Op = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op);
8727 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Op));
8730 void AArch64TargetLowering::ReplaceNodeResults(
8731 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
8732 switch (N->getOpcode()) {
8734 llvm_unreachable("Don't know how to custom expand this");
8736 ReplaceBITCASTResults(N, Results, DAG);
8738 case ISD::FP_TO_UINT:
8739 case ISD::FP_TO_SINT:
8740 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
8741 // Let normal code take care of it by not adding anything to Results.
8746 bool AArch64TargetLowering::useLoadStackGuardNode() const {
8750 bool AArch64TargetLowering::combineRepeatedFPDivisors(unsigned NumUsers) const {
8751 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
8752 // reciprocal if there are three or more FDIVs.
8753 return NumUsers > 2;
8756 TargetLoweringBase::LegalizeTypeAction
8757 AArch64TargetLowering::getPreferredVectorAction(EVT VT) const {
8758 MVT SVT = VT.getSimpleVT();
8759 // During type legalization, we prefer to widen v1i8, v1i16, v1i32 to v8i8,
8760 // v4i16, v2i32 instead of to promote.
8761 if (SVT == MVT::v1i8 || SVT == MVT::v1i16 || SVT == MVT::v1i32
8762 || SVT == MVT::v1f32)
8763 return TypeWidenVector;
8765 return TargetLoweringBase::getPreferredVectorAction(VT);
8768 // Loads and stores less than 128-bits are already atomic; ones above that
8769 // are doomed anyway, so defer to the default libcall and blame the OS when
8771 bool AArch64TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
8772 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
8776 // Loads and stores less than 128-bits are already atomic; ones above that
8777 // are doomed anyway, so defer to the default libcall and blame the OS when
8779 bool AArch64TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
8780 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
8784 // For the real atomic operations, we have ldxr/stxr up to 128 bits,
8785 bool AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
8786 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
8790 bool AArch64TargetLowering::hasLoadLinkedStoreConditional() const {
8794 Value *AArch64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
8795 AtomicOrdering Ord) const {
8796 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
8797 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
8798 bool IsAcquire = isAtLeastAcquire(Ord);
8800 // Since i128 isn't legal and intrinsics don't get type-lowered, the ldrexd
8801 // intrinsic must return {i64, i64} and we have to recombine them into a
8802 // single i128 here.
8803 if (ValTy->getPrimitiveSizeInBits() == 128) {
8805 IsAcquire ? Intrinsic::aarch64_ldaxp : Intrinsic::aarch64_ldxp;
8806 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int);
8808 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
8809 Value *LoHi = Builder.CreateCall(Ldxr, Addr, "lohi");
8811 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
8812 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
8813 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
8814 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
8815 return Builder.CreateOr(
8816 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64");
8819 Type *Tys[] = { Addr->getType() };
8821 IsAcquire ? Intrinsic::aarch64_ldaxr : Intrinsic::aarch64_ldxr;
8822 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int, Tys);
8824 return Builder.CreateTruncOrBitCast(
8825 Builder.CreateCall(Ldxr, Addr),
8826 cast<PointerType>(Addr->getType())->getElementType());
8829 Value *AArch64TargetLowering::emitStoreConditional(IRBuilder<> &Builder,
8830 Value *Val, Value *Addr,
8831 AtomicOrdering Ord) const {
8832 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
8833 bool IsRelease = isAtLeastRelease(Ord);
8835 // Since the intrinsics must have legal type, the i128 intrinsics take two
8836 // parameters: "i64, i64". We must marshal Val into the appropriate form
8838 if (Val->getType()->getPrimitiveSizeInBits() == 128) {
8840 IsRelease ? Intrinsic::aarch64_stlxp : Intrinsic::aarch64_stxp;
8841 Function *Stxr = Intrinsic::getDeclaration(M, Int);
8842 Type *Int64Ty = Type::getInt64Ty(M->getContext());
8844 Value *Lo = Builder.CreateTrunc(Val, Int64Ty, "lo");
8845 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 64), Int64Ty, "hi");
8846 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
8847 return Builder.CreateCall3(Stxr, Lo, Hi, Addr);
8851 IsRelease ? Intrinsic::aarch64_stlxr : Intrinsic::aarch64_stxr;
8852 Type *Tys[] = { Addr->getType() };
8853 Function *Stxr = Intrinsic::getDeclaration(M, Int, Tys);
8855 return Builder.CreateCall2(
8856 Stxr, Builder.CreateZExtOrBitCast(
8857 Val, Stxr->getFunctionType()->getParamType(0)),
8861 bool AArch64TargetLowering::functionArgumentNeedsConsecutiveRegisters(
8862 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
8863 return Ty->isArrayTy();