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 (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
543 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
545 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,
548 setOperationAction(ISD::MULHS, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::MULHU, (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
555 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
556 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
557 setTruncStoreAction((MVT::SimpleValueType)VT,
558 (MVT::SimpleValueType)InnerVT, Expand);
559 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
560 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
561 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
564 // AArch64 has implementations of a lot of rounding-like FP operations.
565 static MVT RoundingVecTypes[] = {MVT::v2f32, MVT::v4f32, MVT::v2f64 };
566 for (unsigned I = 0; I < array_lengthof(RoundingVecTypes); ++I) {
567 MVT Ty = RoundingVecTypes[I];
568 setOperationAction(ISD::FFLOOR, Ty, Legal);
569 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
570 setOperationAction(ISD::FCEIL, Ty, Legal);
571 setOperationAction(ISD::FRINT, Ty, Legal);
572 setOperationAction(ISD::FTRUNC, Ty, Legal);
573 setOperationAction(ISD::FROUND, Ty, Legal);
577 // Prefer likely predicted branches to selects on out-of-order cores.
578 if (Subtarget->isCortexA57())
579 PredictableSelectIsExpensive = true;
582 void AArch64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
583 if (VT == MVT::v2f32 || VT == MVT::v4f16) {
584 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
585 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
587 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
588 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
589 } else if (VT == MVT::v2f64 || VT == MVT::v4f32 || VT == MVT::v8f16) {
590 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
591 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
593 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
594 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
597 // Mark vector float intrinsics as expand.
598 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
599 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
600 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
601 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
602 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
603 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
604 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
605 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
606 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
607 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
610 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
611 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
612 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
613 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
614 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
615 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
616 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
617 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
618 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
619 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
620 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
621 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
623 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
624 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
625 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
626 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
628 // CNT supports only B element sizes.
629 if (VT != MVT::v8i8 && VT != MVT::v16i8)
630 setOperationAction(ISD::CTPOP, VT.getSimpleVT(), Expand);
632 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
633 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
634 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
635 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
636 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
638 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
639 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
641 if (Subtarget->isLittleEndian()) {
642 for (unsigned im = (unsigned)ISD::PRE_INC;
643 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
644 setIndexedLoadAction(im, VT.getSimpleVT(), Legal);
645 setIndexedStoreAction(im, VT.getSimpleVT(), Legal);
650 void AArch64TargetLowering::addDRTypeForNEON(MVT VT) {
651 addRegisterClass(VT, &AArch64::FPR64RegClass);
652 addTypeForNEON(VT, MVT::v2i32);
655 void AArch64TargetLowering::addQRTypeForNEON(MVT VT) {
656 addRegisterClass(VT, &AArch64::FPR128RegClass);
657 addTypeForNEON(VT, MVT::v4i32);
660 EVT AArch64TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
663 return VT.changeVectorElementTypeToInteger();
666 /// computeKnownBitsForTargetNode - Determine which of the bits specified in
667 /// Mask are known to be either zero or one and return them in the
668 /// KnownZero/KnownOne bitsets.
669 void AArch64TargetLowering::computeKnownBitsForTargetNode(
670 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
671 const SelectionDAG &DAG, unsigned Depth) const {
672 switch (Op.getOpcode()) {
675 case AArch64ISD::CSEL: {
676 APInt KnownZero2, KnownOne2;
677 DAG.computeKnownBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
678 DAG.computeKnownBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
679 KnownZero &= KnownZero2;
680 KnownOne &= KnownOne2;
683 case ISD::INTRINSIC_W_CHAIN: {
684 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
685 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
688 case Intrinsic::aarch64_ldaxr:
689 case Intrinsic::aarch64_ldxr: {
690 unsigned BitWidth = KnownOne.getBitWidth();
691 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
692 unsigned MemBits = VT.getScalarType().getSizeInBits();
693 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
699 case ISD::INTRINSIC_WO_CHAIN:
700 case ISD::INTRINSIC_VOID: {
701 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
705 case Intrinsic::aarch64_neon_umaxv:
706 case Intrinsic::aarch64_neon_uminv: {
707 // Figure out the datatype of the vector operand. The UMINV instruction
708 // will zero extend the result, so we can mark as known zero all the
709 // bits larger than the element datatype. 32-bit or larget doesn't need
710 // this as those are legal types and will be handled by isel directly.
711 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
712 unsigned BitWidth = KnownZero.getBitWidth();
713 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
714 assert(BitWidth >= 8 && "Unexpected width!");
715 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
717 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
718 assert(BitWidth >= 16 && "Unexpected width!");
719 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
729 MVT AArch64TargetLowering::getScalarShiftAmountTy(EVT LHSTy) const {
733 unsigned AArch64TargetLowering::getMaximalGlobalOffset() const {
734 // FIXME: On AArch64, this depends on the type.
735 // Basically, the addressable offsets are up to 4095 * Ty.getSizeInBytes().
736 // and the offset has to be a multiple of the related size in bytes.
741 AArch64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
742 const TargetLibraryInfo *libInfo) const {
743 return AArch64::createFastISel(funcInfo, libInfo);
746 const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
750 case AArch64ISD::CALL: return "AArch64ISD::CALL";
751 case AArch64ISD::ADRP: return "AArch64ISD::ADRP";
752 case AArch64ISD::ADDlow: return "AArch64ISD::ADDlow";
753 case AArch64ISD::LOADgot: return "AArch64ISD::LOADgot";
754 case AArch64ISD::RET_FLAG: return "AArch64ISD::RET_FLAG";
755 case AArch64ISD::BRCOND: return "AArch64ISD::BRCOND";
756 case AArch64ISD::CSEL: return "AArch64ISD::CSEL";
757 case AArch64ISD::FCSEL: return "AArch64ISD::FCSEL";
758 case AArch64ISD::CSINV: return "AArch64ISD::CSINV";
759 case AArch64ISD::CSNEG: return "AArch64ISD::CSNEG";
760 case AArch64ISD::CSINC: return "AArch64ISD::CSINC";
761 case AArch64ISD::THREAD_POINTER: return "AArch64ISD::THREAD_POINTER";
762 case AArch64ISD::TLSDESC_CALL: return "AArch64ISD::TLSDESC_CALL";
763 case AArch64ISD::ADC: return "AArch64ISD::ADC";
764 case AArch64ISD::SBC: return "AArch64ISD::SBC";
765 case AArch64ISD::ADDS: return "AArch64ISD::ADDS";
766 case AArch64ISD::SUBS: return "AArch64ISD::SUBS";
767 case AArch64ISD::ADCS: return "AArch64ISD::ADCS";
768 case AArch64ISD::SBCS: return "AArch64ISD::SBCS";
769 case AArch64ISD::ANDS: return "AArch64ISD::ANDS";
770 case AArch64ISD::FCMP: return "AArch64ISD::FCMP";
771 case AArch64ISD::FMIN: return "AArch64ISD::FMIN";
772 case AArch64ISD::FMAX: return "AArch64ISD::FMAX";
773 case AArch64ISD::DUP: return "AArch64ISD::DUP";
774 case AArch64ISD::DUPLANE8: return "AArch64ISD::DUPLANE8";
775 case AArch64ISD::DUPLANE16: return "AArch64ISD::DUPLANE16";
776 case AArch64ISD::DUPLANE32: return "AArch64ISD::DUPLANE32";
777 case AArch64ISD::DUPLANE64: return "AArch64ISD::DUPLANE64";
778 case AArch64ISD::MOVI: return "AArch64ISD::MOVI";
779 case AArch64ISD::MOVIshift: return "AArch64ISD::MOVIshift";
780 case AArch64ISD::MOVIedit: return "AArch64ISD::MOVIedit";
781 case AArch64ISD::MOVImsl: return "AArch64ISD::MOVImsl";
782 case AArch64ISD::FMOV: return "AArch64ISD::FMOV";
783 case AArch64ISD::MVNIshift: return "AArch64ISD::MVNIshift";
784 case AArch64ISD::MVNImsl: return "AArch64ISD::MVNImsl";
785 case AArch64ISD::BICi: return "AArch64ISD::BICi";
786 case AArch64ISD::ORRi: return "AArch64ISD::ORRi";
787 case AArch64ISD::BSL: return "AArch64ISD::BSL";
788 case AArch64ISD::NEG: return "AArch64ISD::NEG";
789 case AArch64ISD::EXTR: return "AArch64ISD::EXTR";
790 case AArch64ISD::ZIP1: return "AArch64ISD::ZIP1";
791 case AArch64ISD::ZIP2: return "AArch64ISD::ZIP2";
792 case AArch64ISD::UZP1: return "AArch64ISD::UZP1";
793 case AArch64ISD::UZP2: return "AArch64ISD::UZP2";
794 case AArch64ISD::TRN1: return "AArch64ISD::TRN1";
795 case AArch64ISD::TRN2: return "AArch64ISD::TRN2";
796 case AArch64ISD::REV16: return "AArch64ISD::REV16";
797 case AArch64ISD::REV32: return "AArch64ISD::REV32";
798 case AArch64ISD::REV64: return "AArch64ISD::REV64";
799 case AArch64ISD::EXT: return "AArch64ISD::EXT";
800 case AArch64ISD::VSHL: return "AArch64ISD::VSHL";
801 case AArch64ISD::VLSHR: return "AArch64ISD::VLSHR";
802 case AArch64ISD::VASHR: return "AArch64ISD::VASHR";
803 case AArch64ISD::CMEQ: return "AArch64ISD::CMEQ";
804 case AArch64ISD::CMGE: return "AArch64ISD::CMGE";
805 case AArch64ISD::CMGT: return "AArch64ISD::CMGT";
806 case AArch64ISD::CMHI: return "AArch64ISD::CMHI";
807 case AArch64ISD::CMHS: return "AArch64ISD::CMHS";
808 case AArch64ISD::FCMEQ: return "AArch64ISD::FCMEQ";
809 case AArch64ISD::FCMGE: return "AArch64ISD::FCMGE";
810 case AArch64ISD::FCMGT: return "AArch64ISD::FCMGT";
811 case AArch64ISD::CMEQz: return "AArch64ISD::CMEQz";
812 case AArch64ISD::CMGEz: return "AArch64ISD::CMGEz";
813 case AArch64ISD::CMGTz: return "AArch64ISD::CMGTz";
814 case AArch64ISD::CMLEz: return "AArch64ISD::CMLEz";
815 case AArch64ISD::CMLTz: return "AArch64ISD::CMLTz";
816 case AArch64ISD::FCMEQz: return "AArch64ISD::FCMEQz";
817 case AArch64ISD::FCMGEz: return "AArch64ISD::FCMGEz";
818 case AArch64ISD::FCMGTz: return "AArch64ISD::FCMGTz";
819 case AArch64ISD::FCMLEz: return "AArch64ISD::FCMLEz";
820 case AArch64ISD::FCMLTz: return "AArch64ISD::FCMLTz";
821 case AArch64ISD::NOT: return "AArch64ISD::NOT";
822 case AArch64ISD::BIT: return "AArch64ISD::BIT";
823 case AArch64ISD::CBZ: return "AArch64ISD::CBZ";
824 case AArch64ISD::CBNZ: return "AArch64ISD::CBNZ";
825 case AArch64ISD::TBZ: return "AArch64ISD::TBZ";
826 case AArch64ISD::TBNZ: return "AArch64ISD::TBNZ";
827 case AArch64ISD::TC_RETURN: return "AArch64ISD::TC_RETURN";
828 case AArch64ISD::SITOF: return "AArch64ISD::SITOF";
829 case AArch64ISD::UITOF: return "AArch64ISD::UITOF";
830 case AArch64ISD::NVCAST: return "AArch64ISD::NVCAST";
831 case AArch64ISD::SQSHL_I: return "AArch64ISD::SQSHL_I";
832 case AArch64ISD::UQSHL_I: return "AArch64ISD::UQSHL_I";
833 case AArch64ISD::SRSHR_I: return "AArch64ISD::SRSHR_I";
834 case AArch64ISD::URSHR_I: return "AArch64ISD::URSHR_I";
835 case AArch64ISD::SQSHLU_I: return "AArch64ISD::SQSHLU_I";
836 case AArch64ISD::WrapperLarge: return "AArch64ISD::WrapperLarge";
837 case AArch64ISD::LD2post: return "AArch64ISD::LD2post";
838 case AArch64ISD::LD3post: return "AArch64ISD::LD3post";
839 case AArch64ISD::LD4post: return "AArch64ISD::LD4post";
840 case AArch64ISD::ST2post: return "AArch64ISD::ST2post";
841 case AArch64ISD::ST3post: return "AArch64ISD::ST3post";
842 case AArch64ISD::ST4post: return "AArch64ISD::ST4post";
843 case AArch64ISD::LD1x2post: return "AArch64ISD::LD1x2post";
844 case AArch64ISD::LD1x3post: return "AArch64ISD::LD1x3post";
845 case AArch64ISD::LD1x4post: return "AArch64ISD::LD1x4post";
846 case AArch64ISD::ST1x2post: return "AArch64ISD::ST1x2post";
847 case AArch64ISD::ST1x3post: return "AArch64ISD::ST1x3post";
848 case AArch64ISD::ST1x4post: return "AArch64ISD::ST1x4post";
849 case AArch64ISD::LD1DUPpost: return "AArch64ISD::LD1DUPpost";
850 case AArch64ISD::LD2DUPpost: return "AArch64ISD::LD2DUPpost";
851 case AArch64ISD::LD3DUPpost: return "AArch64ISD::LD3DUPpost";
852 case AArch64ISD::LD4DUPpost: return "AArch64ISD::LD4DUPpost";
853 case AArch64ISD::LD1LANEpost: return "AArch64ISD::LD1LANEpost";
854 case AArch64ISD::LD2LANEpost: return "AArch64ISD::LD2LANEpost";
855 case AArch64ISD::LD3LANEpost: return "AArch64ISD::LD3LANEpost";
856 case AArch64ISD::LD4LANEpost: return "AArch64ISD::LD4LANEpost";
857 case AArch64ISD::ST2LANEpost: return "AArch64ISD::ST2LANEpost";
858 case AArch64ISD::ST3LANEpost: return "AArch64ISD::ST3LANEpost";
859 case AArch64ISD::ST4LANEpost: return "AArch64ISD::ST4LANEpost";
860 case AArch64ISD::SMULL: return "AArch64ISD::SMULL";
861 case AArch64ISD::UMULL: return "AArch64ISD::UMULL";
866 AArch64TargetLowering::EmitF128CSEL(MachineInstr *MI,
867 MachineBasicBlock *MBB) const {
868 // We materialise the F128CSEL pseudo-instruction as some control flow and a
872 // [... previous instrs leading to comparison ...]
878 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
880 const TargetInstrInfo *TII =
881 getTargetMachine().getSubtargetImpl()->getInstrInfo();
882 MachineFunction *MF = MBB->getParent();
883 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
884 DebugLoc DL = MI->getDebugLoc();
885 MachineFunction::iterator It = MBB;
888 unsigned DestReg = MI->getOperand(0).getReg();
889 unsigned IfTrueReg = MI->getOperand(1).getReg();
890 unsigned IfFalseReg = MI->getOperand(2).getReg();
891 unsigned CondCode = MI->getOperand(3).getImm();
892 bool NZCVKilled = MI->getOperand(4).isKill();
894 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
895 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
896 MF->insert(It, TrueBB);
897 MF->insert(It, EndBB);
899 // Transfer rest of current basic-block to EndBB
900 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
902 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
904 BuildMI(MBB, DL, TII->get(AArch64::Bcc)).addImm(CondCode).addMBB(TrueBB);
905 BuildMI(MBB, DL, TII->get(AArch64::B)).addMBB(EndBB);
906 MBB->addSuccessor(TrueBB);
907 MBB->addSuccessor(EndBB);
909 // TrueBB falls through to the end.
910 TrueBB->addSuccessor(EndBB);
913 TrueBB->addLiveIn(AArch64::NZCV);
914 EndBB->addLiveIn(AArch64::NZCV);
917 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(AArch64::PHI), DestReg)
923 MI->eraseFromParent();
928 AArch64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
929 MachineBasicBlock *BB) const {
930 switch (MI->getOpcode()) {
935 llvm_unreachable("Unexpected instruction for custom inserter!");
937 case AArch64::F128CSEL:
938 return EmitF128CSEL(MI, BB);
940 case TargetOpcode::STACKMAP:
941 case TargetOpcode::PATCHPOINT:
942 return emitPatchPoint(MI, BB);
946 //===----------------------------------------------------------------------===//
947 // AArch64 Lowering private implementation.
948 //===----------------------------------------------------------------------===//
950 //===----------------------------------------------------------------------===//
952 //===----------------------------------------------------------------------===//
954 /// changeIntCCToAArch64CC - Convert a DAG integer condition code to an AArch64
956 static AArch64CC::CondCode changeIntCCToAArch64CC(ISD::CondCode CC) {
959 llvm_unreachable("Unknown condition code!");
961 return AArch64CC::NE;
963 return AArch64CC::EQ;
965 return AArch64CC::GT;
967 return AArch64CC::GE;
969 return AArch64CC::LT;
971 return AArch64CC::LE;
973 return AArch64CC::HI;
975 return AArch64CC::HS;
977 return AArch64CC::LO;
979 return AArch64CC::LS;
983 /// changeFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64 CC.
984 static void changeFPCCToAArch64CC(ISD::CondCode CC,
985 AArch64CC::CondCode &CondCode,
986 AArch64CC::CondCode &CondCode2) {
987 CondCode2 = AArch64CC::AL;
990 llvm_unreachable("Unknown FP condition!");
993 CondCode = AArch64CC::EQ;
997 CondCode = AArch64CC::GT;
1001 CondCode = AArch64CC::GE;
1004 CondCode = AArch64CC::MI;
1007 CondCode = AArch64CC::LS;
1010 CondCode = AArch64CC::MI;
1011 CondCode2 = AArch64CC::GT;
1014 CondCode = AArch64CC::VC;
1017 CondCode = AArch64CC::VS;
1020 CondCode = AArch64CC::EQ;
1021 CondCode2 = AArch64CC::VS;
1024 CondCode = AArch64CC::HI;
1027 CondCode = AArch64CC::PL;
1031 CondCode = AArch64CC::LT;
1035 CondCode = AArch64CC::LE;
1039 CondCode = AArch64CC::NE;
1044 /// changeVectorFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64
1045 /// CC usable with the vector instructions. Fewer operations are available
1046 /// without a real NZCV register, so we have to use less efficient combinations
1047 /// to get the same effect.
1048 static void changeVectorFPCCToAArch64CC(ISD::CondCode CC,
1049 AArch64CC::CondCode &CondCode,
1050 AArch64CC::CondCode &CondCode2,
1055 // Mostly the scalar mappings work fine.
1056 changeFPCCToAArch64CC(CC, CondCode, CondCode2);
1059 Invert = true; // Fallthrough
1061 CondCode = AArch64CC::MI;
1062 CondCode2 = AArch64CC::GE;
1069 // All of the compare-mask comparisons are ordered, but we can switch
1070 // between the two by a double inversion. E.g. ULE == !OGT.
1072 changeFPCCToAArch64CC(getSetCCInverse(CC, false), CondCode, CondCode2);
1077 static bool isLegalArithImmed(uint64_t C) {
1078 // Matches AArch64DAGToDAGISel::SelectArithImmed().
1079 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
1082 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1083 SDLoc dl, SelectionDAG &DAG) {
1084 EVT VT = LHS.getValueType();
1086 if (VT.isFloatingPoint())
1087 return DAG.getNode(AArch64ISD::FCMP, dl, VT, LHS, RHS);
1089 // The CMP instruction is just an alias for SUBS, and representing it as
1090 // SUBS means that it's possible to get CSE with subtract operations.
1091 // A later phase can perform the optimization of setting the destination
1092 // register to WZR/XZR if it ends up being unused.
1093 unsigned Opcode = AArch64ISD::SUBS;
1095 if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
1096 cast<ConstantSDNode>(RHS.getOperand(0))->getZExtValue() == 0 &&
1097 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1098 // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on
1099 // the grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags
1100 // can be set differently by this operation. It comes down to whether
1101 // "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then
1102 // everything is fine. If not then the optimization is wrong. Thus general
1103 // comparisons are only valid if op2 != 0.
1105 // So, finally, the only LLVM-native comparisons that don't mention C and V
1106 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in
1107 // the absence of information about op2.
1108 Opcode = AArch64ISD::ADDS;
1109 RHS = RHS.getOperand(1);
1110 } else if (LHS.getOpcode() == ISD::AND && isa<ConstantSDNode>(RHS) &&
1111 cast<ConstantSDNode>(RHS)->getZExtValue() == 0 &&
1112 !isUnsignedIntSetCC(CC)) {
1113 // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
1114 // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
1115 // of the signed comparisons.
1116 Opcode = AArch64ISD::ANDS;
1117 RHS = LHS.getOperand(1);
1118 LHS = LHS.getOperand(0);
1121 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS)
1125 static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1126 SDValue &AArch64cc, SelectionDAG &DAG, SDLoc dl) {
1128 AArch64CC::CondCode AArch64CC;
1129 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1130 EVT VT = RHS.getValueType();
1131 uint64_t C = RHSC->getZExtValue();
1132 if (!isLegalArithImmed(C)) {
1133 // Constant does not fit, try adjusting it by one?
1139 if ((VT == MVT::i32 && C != 0x80000000 &&
1140 isLegalArithImmed((uint32_t)(C - 1))) ||
1141 (VT == MVT::i64 && C != 0x80000000ULL &&
1142 isLegalArithImmed(C - 1ULL))) {
1143 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1144 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1145 RHS = DAG.getConstant(C, VT);
1150 if ((VT == MVT::i32 && C != 0 &&
1151 isLegalArithImmed((uint32_t)(C - 1))) ||
1152 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1153 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1154 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1155 RHS = DAG.getConstant(C, VT);
1160 if ((VT == MVT::i32 && C != INT32_MAX &&
1161 isLegalArithImmed((uint32_t)(C + 1))) ||
1162 (VT == MVT::i64 && C != INT64_MAX &&
1163 isLegalArithImmed(C + 1ULL))) {
1164 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1165 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1166 RHS = DAG.getConstant(C, VT);
1171 if ((VT == MVT::i32 && C != UINT32_MAX &&
1172 isLegalArithImmed((uint32_t)(C + 1))) ||
1173 (VT == MVT::i64 && C != UINT64_MAX &&
1174 isLegalArithImmed(C + 1ULL))) {
1175 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1176 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1177 RHS = DAG.getConstant(C, VT);
1183 // The imm operand of ADDS is an unsigned immediate, in the range 0 to 4095.
1184 // For the i8 operand, the largest immediate is 255, so this can be easily
1185 // encoded in the compare instruction. For the i16 operand, however, the
1186 // largest immediate cannot be encoded in the compare.
1187 // Therefore, use a sign extending load and cmn to avoid materializing the -1
1188 // constant. For example,
1190 // ldrh w0, [x0, #0]
1193 // ldrsh w0, [x0, #0]
1195 // Fundamental, we're relying on the property that (zext LHS) == (zext RHS)
1196 // if and only if (sext LHS) == (sext RHS). The checks are in place to ensure
1197 // both the LHS and RHS are truely zero extended and to make sure the
1198 // transformation is profitable.
1199 if ((CC == ISD::SETEQ || CC == ISD::SETNE) && isa<ConstantSDNode>(RHS)) {
1200 if ((cast<ConstantSDNode>(RHS)->getZExtValue() >> 16 == 0) &&
1201 isa<LoadSDNode>(LHS)) {
1202 if (cast<LoadSDNode>(LHS)->getExtensionType() == ISD::ZEXTLOAD &&
1203 cast<LoadSDNode>(LHS)->getMemoryVT() == MVT::i16 &&
1204 LHS.getNode()->hasNUsesOfValue(1, 0)) {
1205 int16_t ValueofRHS = cast<ConstantSDNode>(RHS)->getZExtValue();
1206 if (ValueofRHS < 0 && isLegalArithImmed(-ValueofRHS)) {
1208 DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, LHS.getValueType(), LHS,
1209 DAG.getValueType(MVT::i16));
1210 Cmp = emitComparison(SExt,
1211 DAG.getConstant(ValueofRHS, RHS.getValueType()),
1213 AArch64CC = changeIntCCToAArch64CC(CC);
1214 AArch64cc = DAG.getConstant(AArch64CC, MVT::i32);
1220 Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
1221 AArch64CC = changeIntCCToAArch64CC(CC);
1222 AArch64cc = DAG.getConstant(AArch64CC, MVT::i32);
1226 static std::pair<SDValue, SDValue>
1227 getAArch64XALUOOp(AArch64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1228 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1229 "Unsupported value type");
1230 SDValue Value, Overflow;
1232 SDValue LHS = Op.getOperand(0);
1233 SDValue RHS = Op.getOperand(1);
1235 switch (Op.getOpcode()) {
1237 llvm_unreachable("Unknown overflow instruction!");
1239 Opc = AArch64ISD::ADDS;
1243 Opc = AArch64ISD::ADDS;
1247 Opc = AArch64ISD::SUBS;
1251 Opc = AArch64ISD::SUBS;
1254 // Multiply needs a little bit extra work.
1258 bool IsSigned = (Op.getOpcode() == ISD::SMULO) ? true : false;
1259 if (Op.getValueType() == MVT::i32) {
1260 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1261 // For a 32 bit multiply with overflow check we want the instruction
1262 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1263 // need to generate the following pattern:
1264 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1265 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1266 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1267 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1268 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1269 DAG.getConstant(0, MVT::i64));
1270 // On AArch64 the upper 32 bits are always zero extended for a 32 bit
1271 // operation. We need to clear out the upper 32 bits, because we used a
1272 // widening multiply that wrote all 64 bits. In the end this should be a
1274 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1276 // The signed overflow check requires more than just a simple check for
1277 // any bit set in the upper 32 bits of the result. These bits could be
1278 // just the sign bits of a negative number. To perform the overflow
1279 // check we have to arithmetic shift right the 32nd bit of the result by
1280 // 31 bits. Then we compare the result to the upper 32 bits.
1281 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1282 DAG.getConstant(32, MVT::i64));
1283 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1284 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1285 DAG.getConstant(31, MVT::i64));
1286 // It is important that LowerBits is last, otherwise the arithmetic
1287 // shift will not be folded into the compare (SUBS).
1288 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1289 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1292 // The overflow check for unsigned multiply is easy. We only need to
1293 // check if any of the upper 32 bits are set. This can be done with a
1294 // CMP (shifted register). For that we need to generate the following
1296 // (i64 AArch64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1297 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1298 DAG.getConstant(32, MVT::i64));
1299 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1301 DAG.getNode(AArch64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1302 UpperBits).getValue(1);
1306 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1307 // For the 64 bit multiply
1308 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1310 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1311 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1312 DAG.getConstant(63, MVT::i64));
1313 // It is important that LowerBits is last, otherwise the arithmetic
1314 // shift will not be folded into the compare (SUBS).
1315 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1316 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1319 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1320 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1322 DAG.getNode(AArch64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1323 UpperBits).getValue(1);
1330 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1332 // Emit the AArch64 operation with overflow check.
1333 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1334 Overflow = Value.getValue(1);
1336 return std::make_pair(Value, Overflow);
1339 SDValue AArch64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1340 RTLIB::Libcall Call) const {
1341 SmallVector<SDValue, 2> Ops;
1342 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1343 Ops.push_back(Op.getOperand(i));
1345 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1349 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1350 SDValue Sel = Op.getOperand(0);
1351 SDValue Other = Op.getOperand(1);
1353 // If neither operand is a SELECT_CC, give up.
1354 if (Sel.getOpcode() != ISD::SELECT_CC)
1355 std::swap(Sel, Other);
1356 if (Sel.getOpcode() != ISD::SELECT_CC)
1359 // The folding we want to perform is:
1360 // (xor x, (select_cc a, b, cc, 0, -1) )
1362 // (csel x, (xor x, -1), cc ...)
1364 // The latter will get matched to a CSINV instruction.
1366 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1367 SDValue LHS = Sel.getOperand(0);
1368 SDValue RHS = Sel.getOperand(1);
1369 SDValue TVal = Sel.getOperand(2);
1370 SDValue FVal = Sel.getOperand(3);
1373 // FIXME: This could be generalized to non-integer comparisons.
1374 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1377 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1378 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1380 // The the values aren't constants, this isn't the pattern we're looking for.
1381 if (!CFVal || !CTVal)
1384 // We can commute the SELECT_CC by inverting the condition. This
1385 // might be needed to make this fit into a CSINV pattern.
1386 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1387 std::swap(TVal, FVal);
1388 std::swap(CTVal, CFVal);
1389 CC = ISD::getSetCCInverse(CC, true);
1392 // If the constants line up, perform the transform!
1393 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1395 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1398 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1399 DAG.getConstant(-1ULL, Other.getValueType()));
1401 return DAG.getNode(AArch64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1408 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1409 EVT VT = Op.getValueType();
1411 // Let legalize expand this if it isn't a legal type yet.
1412 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1415 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1418 bool ExtraOp = false;
1419 switch (Op.getOpcode()) {
1421 llvm_unreachable("Invalid code");
1423 Opc = AArch64ISD::ADDS;
1426 Opc = AArch64ISD::SUBS;
1429 Opc = AArch64ISD::ADCS;
1433 Opc = AArch64ISD::SBCS;
1439 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1440 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1444 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1445 // Let legalize expand this if it isn't a legal type yet.
1446 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1449 AArch64CC::CondCode CC;
1450 // The actual operation that sets the overflow or carry flag.
1451 SDValue Value, Overflow;
1452 std::tie(Value, Overflow) = getAArch64XALUOOp(CC, Op, DAG);
1454 // We use 0 and 1 as false and true values.
1455 SDValue TVal = DAG.getConstant(1, MVT::i32);
1456 SDValue FVal = DAG.getConstant(0, MVT::i32);
1458 // We use an inverted condition, because the conditional select is inverted
1459 // too. This will allow it to be selected to a single instruction:
1460 // CSINC Wd, WZR, WZR, invert(cond).
1461 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), MVT::i32);
1462 Overflow = DAG.getNode(AArch64ISD::CSEL, SDLoc(Op), MVT::i32, FVal, TVal,
1465 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1466 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
1469 // Prefetch operands are:
1470 // 1: Address to prefetch
1472 // 3: int locality (0 = no locality ... 3 = extreme locality)
1473 // 4: bool isDataCache
1474 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1476 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1477 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1478 unsigned IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1480 bool IsStream = !Locality;
1481 // When the locality number is set
1483 // The front-end should have filtered out the out-of-range values
1484 assert(Locality <= 3 && "Prefetch locality out-of-range");
1485 // The locality degree is the opposite of the cache speed.
1486 // Put the number the other way around.
1487 // The encoding starts at 0 for level 1
1488 Locality = 3 - Locality;
1491 // built the mask value encoding the expected behavior.
1492 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1493 (!IsData << 3) | // IsDataCache bit
1494 (Locality << 1) | // Cache level bits
1495 (unsigned)IsStream; // Stream bit
1496 return DAG.getNode(AArch64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1497 DAG.getConstant(PrfOp, MVT::i32), Op.getOperand(1));
1500 SDValue AArch64TargetLowering::LowerFP_EXTEND(SDValue Op,
1501 SelectionDAG &DAG) const {
1502 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1505 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1507 return LowerF128Call(Op, DAG, LC);
1510 SDValue AArch64TargetLowering::LowerFP_ROUND(SDValue Op,
1511 SelectionDAG &DAG) const {
1512 if (Op.getOperand(0).getValueType() != MVT::f128) {
1513 // It's legal except when f128 is involved
1518 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1520 // FP_ROUND node has a second operand indicating whether it is known to be
1521 // precise. That doesn't take part in the LibCall so we can't directly use
1523 SDValue SrcVal = Op.getOperand(0);
1524 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1525 /*isSigned*/ false, SDLoc(Op)).first;
1528 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1529 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1530 // Any additional optimization in this function should be recorded
1531 // in the cost tables.
1532 EVT InVT = Op.getOperand(0).getValueType();
1533 EVT VT = Op.getValueType();
1535 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1538 DAG.getNode(Op.getOpcode(), dl, InVT.changeVectorElementTypeToInteger(),
1540 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1543 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1546 MVT::getVectorVT(MVT::getFloatingPointVT(VT.getScalarSizeInBits()),
1547 VT.getVectorNumElements());
1548 SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, ExtVT, Op.getOperand(0));
1549 return DAG.getNode(Op.getOpcode(), dl, VT, Ext);
1552 // Type changing conversions are illegal.
1556 SDValue AArch64TargetLowering::LowerFP_TO_INT(SDValue Op,
1557 SelectionDAG &DAG) const {
1558 if (Op.getOperand(0).getValueType().isVector())
1559 return LowerVectorFP_TO_INT(Op, DAG);
1561 if (Op.getOperand(0).getValueType() != MVT::f128) {
1562 // It's legal except when f128 is involved
1567 if (Op.getOpcode() == ISD::FP_TO_SINT)
1568 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1570 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1572 SmallVector<SDValue, 2> Ops;
1573 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1574 Ops.push_back(Op.getOperand(i));
1576 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1580 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1581 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1582 // Any additional optimization in this function should be recorded
1583 // in the cost tables.
1584 EVT VT = Op.getValueType();
1586 SDValue In = Op.getOperand(0);
1587 EVT InVT = In.getValueType();
1589 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1591 MVT::getVectorVT(MVT::getFloatingPointVT(InVT.getScalarSizeInBits()),
1592 InVT.getVectorNumElements());
1593 In = DAG.getNode(Op.getOpcode(), dl, CastVT, In);
1594 return DAG.getNode(ISD::FP_ROUND, dl, VT, In, DAG.getIntPtrConstant(0));
1597 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1599 Op.getOpcode() == ISD::SINT_TO_FP ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1600 EVT CastVT = VT.changeVectorElementTypeToInteger();
1601 In = DAG.getNode(CastOpc, dl, CastVT, In);
1602 return DAG.getNode(Op.getOpcode(), dl, VT, In);
1608 SDValue AArch64TargetLowering::LowerINT_TO_FP(SDValue Op,
1609 SelectionDAG &DAG) const {
1610 if (Op.getValueType().isVector())
1611 return LowerVectorINT_TO_FP(Op, DAG);
1613 // i128 conversions are libcalls.
1614 if (Op.getOperand(0).getValueType() == MVT::i128)
1617 // Other conversions are legal, unless it's to the completely software-based
1619 if (Op.getValueType() != MVT::f128)
1623 if (Op.getOpcode() == ISD::SINT_TO_FP)
1624 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1626 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1628 return LowerF128Call(Op, DAG, LC);
1631 SDValue AArch64TargetLowering::LowerFSINCOS(SDValue Op,
1632 SelectionDAG &DAG) const {
1633 // For iOS, we want to call an alternative entry point: __sincos_stret,
1634 // which returns the values in two S / D registers.
1636 SDValue Arg = Op.getOperand(0);
1637 EVT ArgVT = Arg.getValueType();
1638 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1645 Entry.isSExt = false;
1646 Entry.isZExt = false;
1647 Args.push_back(Entry);
1649 const char *LibcallName =
1650 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1651 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
1653 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
1654 TargetLowering::CallLoweringInfo CLI(DAG);
1655 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
1656 .setCallee(CallingConv::Fast, RetTy, Callee, std::move(Args), 0);
1658 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1659 return CallResult.first;
1662 static SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) {
1663 if (Op.getValueType() != MVT::f16)
1666 assert(Op.getOperand(0).getValueType() == MVT::i16);
1669 Op = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op.getOperand(0));
1670 Op = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Op);
1672 DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, MVT::f16, Op,
1673 DAG.getTargetConstant(AArch64::hsub, MVT::i32)),
1677 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
1678 if (OrigVT.getSizeInBits() >= 64)
1681 assert(OrigVT.isSimple() && "Expecting a simple value type");
1683 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
1684 switch (OrigSimpleTy) {
1685 default: llvm_unreachable("Unexpected Vector Type");
1694 static SDValue addRequiredExtensionForVectorMULL(SDValue N, SelectionDAG &DAG,
1697 unsigned ExtOpcode) {
1698 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
1699 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
1700 // 64-bits we need to insert a new extension so that it will be 64-bits.
1701 assert(ExtTy.is128BitVector() && "Unexpected extension size");
1702 if (OrigTy.getSizeInBits() >= 64)
1705 // Must extend size to at least 64 bits to be used as an operand for VMULL.
1706 EVT NewVT = getExtensionTo64Bits(OrigTy);
1708 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
1711 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
1713 EVT VT = N->getValueType(0);
1715 if (N->getOpcode() != ISD::BUILD_VECTOR)
1718 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1719 SDNode *Elt = N->getOperand(i).getNode();
1720 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
1721 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1722 unsigned HalfSize = EltSize / 2;
1724 if (!isIntN(HalfSize, C->getSExtValue()))
1727 if (!isUIntN(HalfSize, C->getZExtValue()))
1738 static SDValue skipExtensionForVectorMULL(SDNode *N, SelectionDAG &DAG) {
1739 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
1740 return addRequiredExtensionForVectorMULL(N->getOperand(0), DAG,
1741 N->getOperand(0)->getValueType(0),
1745 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
1746 EVT VT = N->getValueType(0);
1747 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
1748 unsigned NumElts = VT.getVectorNumElements();
1749 MVT TruncVT = MVT::getIntegerVT(EltSize);
1750 SmallVector<SDValue, 8> Ops;
1751 for (unsigned i = 0; i != NumElts; ++i) {
1752 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
1753 const APInt &CInt = C->getAPIntValue();
1754 // Element types smaller than 32 bits are not legal, so use i32 elements.
1755 // The values are implicitly truncated so sext vs. zext doesn't matter.
1756 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
1758 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
1759 MVT::getVectorVT(TruncVT, NumElts), Ops);
1762 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
1763 if (N->getOpcode() == ISD::SIGN_EXTEND)
1765 if (isExtendedBUILD_VECTOR(N, DAG, true))
1770 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
1771 if (N->getOpcode() == ISD::ZERO_EXTEND)
1773 if (isExtendedBUILD_VECTOR(N, DAG, false))
1778 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
1779 unsigned Opcode = N->getOpcode();
1780 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
1781 SDNode *N0 = N->getOperand(0).getNode();
1782 SDNode *N1 = N->getOperand(1).getNode();
1783 return N0->hasOneUse() && N1->hasOneUse() &&
1784 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
1789 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
1790 unsigned Opcode = N->getOpcode();
1791 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
1792 SDNode *N0 = N->getOperand(0).getNode();
1793 SDNode *N1 = N->getOperand(1).getNode();
1794 return N0->hasOneUse() && N1->hasOneUse() &&
1795 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
1800 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
1801 // Multiplications are only custom-lowered for 128-bit vectors so that
1802 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
1803 EVT VT = Op.getValueType();
1804 assert(VT.is128BitVector() && VT.isInteger() &&
1805 "unexpected type for custom-lowering ISD::MUL");
1806 SDNode *N0 = Op.getOperand(0).getNode();
1807 SDNode *N1 = Op.getOperand(1).getNode();
1808 unsigned NewOpc = 0;
1810 bool isN0SExt = isSignExtended(N0, DAG);
1811 bool isN1SExt = isSignExtended(N1, DAG);
1812 if (isN0SExt && isN1SExt)
1813 NewOpc = AArch64ISD::SMULL;
1815 bool isN0ZExt = isZeroExtended(N0, DAG);
1816 bool isN1ZExt = isZeroExtended(N1, DAG);
1817 if (isN0ZExt && isN1ZExt)
1818 NewOpc = AArch64ISD::UMULL;
1819 else if (isN1SExt || isN1ZExt) {
1820 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
1821 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
1822 if (isN1SExt && isAddSubSExt(N0, DAG)) {
1823 NewOpc = AArch64ISD::SMULL;
1825 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
1826 NewOpc = AArch64ISD::UMULL;
1828 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
1830 NewOpc = AArch64ISD::UMULL;
1836 if (VT == MVT::v2i64)
1837 // Fall through to expand this. It is not legal.
1840 // Other vector multiplications are legal.
1845 // Legalize to a S/UMULL instruction
1848 SDValue Op1 = skipExtensionForVectorMULL(N1, DAG);
1850 Op0 = skipExtensionForVectorMULL(N0, DAG);
1851 assert(Op0.getValueType().is64BitVector() &&
1852 Op1.getValueType().is64BitVector() &&
1853 "unexpected types for extended operands to VMULL");
1854 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
1856 // Optimizing (zext A + zext B) * C, to (S/UMULL A, C) + (S/UMULL B, C) during
1857 // isel lowering to take advantage of no-stall back to back s/umul + s/umla.
1858 // This is true for CPUs with accumulate forwarding such as Cortex-A53/A57
1859 SDValue N00 = skipExtensionForVectorMULL(N0->getOperand(0).getNode(), DAG);
1860 SDValue N01 = skipExtensionForVectorMULL(N0->getOperand(1).getNode(), DAG);
1861 EVT Op1VT = Op1.getValueType();
1862 return DAG.getNode(N0->getOpcode(), DL, VT,
1863 DAG.getNode(NewOpc, DL, VT,
1864 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
1865 DAG.getNode(NewOpc, DL, VT,
1866 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
1869 SDValue AArch64TargetLowering::LowerOperation(SDValue Op,
1870 SelectionDAG &DAG) const {
1871 switch (Op.getOpcode()) {
1873 llvm_unreachable("unimplemented operand");
1876 return LowerBITCAST(Op, DAG);
1877 case ISD::GlobalAddress:
1878 return LowerGlobalAddress(Op, DAG);
1879 case ISD::GlobalTLSAddress:
1880 return LowerGlobalTLSAddress(Op, DAG);
1882 return LowerSETCC(Op, DAG);
1884 return LowerBR_CC(Op, DAG);
1886 return LowerSELECT(Op, DAG);
1887 case ISD::SELECT_CC:
1888 return LowerSELECT_CC(Op, DAG);
1889 case ISD::JumpTable:
1890 return LowerJumpTable(Op, DAG);
1891 case ISD::ConstantPool:
1892 return LowerConstantPool(Op, DAG);
1893 case ISD::BlockAddress:
1894 return LowerBlockAddress(Op, DAG);
1896 return LowerVASTART(Op, DAG);
1898 return LowerVACOPY(Op, DAG);
1900 return LowerVAARG(Op, DAG);
1905 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
1912 return LowerXALUO(Op, DAG);
1914 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
1916 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
1918 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
1920 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
1922 return LowerFP_ROUND(Op, DAG);
1923 case ISD::FP_EXTEND:
1924 return LowerFP_EXTEND(Op, DAG);
1925 case ISD::FRAMEADDR:
1926 return LowerFRAMEADDR(Op, DAG);
1927 case ISD::RETURNADDR:
1928 return LowerRETURNADDR(Op, DAG);
1929 case ISD::INSERT_VECTOR_ELT:
1930 return LowerINSERT_VECTOR_ELT(Op, DAG);
1931 case ISD::EXTRACT_VECTOR_ELT:
1932 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
1933 case ISD::BUILD_VECTOR:
1934 return LowerBUILD_VECTOR(Op, DAG);
1935 case ISD::VECTOR_SHUFFLE:
1936 return LowerVECTOR_SHUFFLE(Op, DAG);
1937 case ISD::EXTRACT_SUBVECTOR:
1938 return LowerEXTRACT_SUBVECTOR(Op, DAG);
1942 return LowerVectorSRA_SRL_SHL(Op, DAG);
1943 case ISD::SHL_PARTS:
1944 return LowerShiftLeftParts(Op, DAG);
1945 case ISD::SRL_PARTS:
1946 case ISD::SRA_PARTS:
1947 return LowerShiftRightParts(Op, DAG);
1949 return LowerCTPOP(Op, DAG);
1950 case ISD::FCOPYSIGN:
1951 return LowerFCOPYSIGN(Op, DAG);
1953 return LowerVectorAND(Op, DAG);
1955 return LowerVectorOR(Op, DAG);
1957 return LowerXOR(Op, DAG);
1959 return LowerPREFETCH(Op, DAG);
1960 case ISD::SINT_TO_FP:
1961 case ISD::UINT_TO_FP:
1962 return LowerINT_TO_FP(Op, DAG);
1963 case ISD::FP_TO_SINT:
1964 case ISD::FP_TO_UINT:
1965 return LowerFP_TO_INT(Op, DAG);
1967 return LowerFSINCOS(Op, DAG);
1969 return LowerMUL(Op, DAG);
1973 /// getFunctionAlignment - Return the Log2 alignment of this function.
1974 unsigned AArch64TargetLowering::getFunctionAlignment(const Function *F) const {
1978 //===----------------------------------------------------------------------===//
1979 // Calling Convention Implementation
1980 //===----------------------------------------------------------------------===//
1982 #include "AArch64GenCallingConv.inc"
1984 /// Selects the correct CCAssignFn for a given CallingConvention value.
1985 CCAssignFn *AArch64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
1986 bool IsVarArg) const {
1989 llvm_unreachable("Unsupported calling convention.");
1990 case CallingConv::WebKit_JS:
1991 return CC_AArch64_WebKit_JS;
1992 case CallingConv::C:
1993 case CallingConv::Fast:
1994 if (!Subtarget->isTargetDarwin())
1995 return CC_AArch64_AAPCS;
1996 return IsVarArg ? CC_AArch64_DarwinPCS_VarArg : CC_AArch64_DarwinPCS;
2000 SDValue AArch64TargetLowering::LowerFormalArguments(
2001 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2002 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2003 SmallVectorImpl<SDValue> &InVals) const {
2004 MachineFunction &MF = DAG.getMachineFunction();
2005 MachineFrameInfo *MFI = MF.getFrameInfo();
2007 // Assign locations to all of the incoming arguments.
2008 SmallVector<CCValAssign, 16> ArgLocs;
2009 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2012 // At this point, Ins[].VT may already be promoted to i32. To correctly
2013 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2014 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2015 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
2016 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
2018 unsigned NumArgs = Ins.size();
2019 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2020 unsigned CurArgIdx = 0;
2021 for (unsigned i = 0; i != NumArgs; ++i) {
2022 MVT ValVT = Ins[i].VT;
2023 std::advance(CurOrigArg, Ins[i].OrigArgIndex - CurArgIdx);
2024 CurArgIdx = Ins[i].OrigArgIndex;
2026 // Get type of the original argument.
2027 EVT ActualVT = getValueType(CurOrigArg->getType(), /*AllowUnknown*/ true);
2028 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
2029 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2030 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2032 else if (ActualMVT == MVT::i16)
2035 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2037 AssignFn(i, ValVT, ValVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
2038 assert(!Res && "Call operand has unhandled type");
2041 assert(ArgLocs.size() == Ins.size());
2042 SmallVector<SDValue, 16> ArgValues;
2043 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2044 CCValAssign &VA = ArgLocs[i];
2046 if (Ins[i].Flags.isByVal()) {
2047 // Byval is used for HFAs in the PCS, but the system should work in a
2048 // non-compliant manner for larger structs.
2049 EVT PtrTy = getPointerTy();
2050 int Size = Ins[i].Flags.getByValSize();
2051 unsigned NumRegs = (Size + 7) / 8;
2053 // FIXME: This works on big-endian for composite byvals, which are the common
2054 // case. It should also work for fundamental types too.
2056 MFI->CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false);
2057 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrTy);
2058 InVals.push_back(FrameIdxN);
2063 if (VA.isRegLoc()) {
2064 // Arguments stored in registers.
2065 EVT RegVT = VA.getLocVT();
2068 const TargetRegisterClass *RC;
2070 if (RegVT == MVT::i32)
2071 RC = &AArch64::GPR32RegClass;
2072 else if (RegVT == MVT::i64)
2073 RC = &AArch64::GPR64RegClass;
2074 else if (RegVT == MVT::f16)
2075 RC = &AArch64::FPR16RegClass;
2076 else if (RegVT == MVT::f32)
2077 RC = &AArch64::FPR32RegClass;
2078 else if (RegVT == MVT::f64 || RegVT.is64BitVector())
2079 RC = &AArch64::FPR64RegClass;
2080 else if (RegVT == MVT::f128 || RegVT.is128BitVector())
2081 RC = &AArch64::FPR128RegClass;
2083 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2085 // Transform the arguments in physical registers into virtual ones.
2086 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2087 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2089 // If this is an 8, 16 or 32-bit value, it is really passed promoted
2090 // to 64 bits. Insert an assert[sz]ext to capture this, then
2091 // truncate to the right size.
2092 switch (VA.getLocInfo()) {
2094 llvm_unreachable("Unknown loc info!");
2095 case CCValAssign::Full:
2097 case CCValAssign::BCvt:
2098 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
2100 case CCValAssign::AExt:
2101 case CCValAssign::SExt:
2102 case CCValAssign::ZExt:
2103 // SelectionDAGBuilder will insert appropriate AssertZExt & AssertSExt
2104 // nodes after our lowering.
2105 assert(RegVT == Ins[i].VT && "incorrect register location selected");
2109 InVals.push_back(ArgValue);
2111 } else { // VA.isRegLoc()
2112 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
2113 unsigned ArgOffset = VA.getLocMemOffset();
2114 unsigned ArgSize = VA.getValVT().getSizeInBits() / 8;
2116 uint32_t BEAlign = 0;
2117 if (!Subtarget->isLittleEndian() && ArgSize < 8 &&
2118 !Ins[i].Flags.isInConsecutiveRegs())
2119 BEAlign = 8 - ArgSize;
2121 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset + BEAlign, true);
2123 // Create load nodes to retrieve arguments from the stack.
2124 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2127 // For NON_EXTLOAD, generic code in getLoad assert(ValVT == MemVT)
2128 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
2129 MVT MemVT = VA.getValVT();
2131 switch (VA.getLocInfo()) {
2134 case CCValAssign::BCvt:
2135 MemVT = VA.getLocVT();
2137 case CCValAssign::SExt:
2138 ExtType = ISD::SEXTLOAD;
2140 case CCValAssign::ZExt:
2141 ExtType = ISD::ZEXTLOAD;
2143 case CCValAssign::AExt:
2144 ExtType = ISD::EXTLOAD;
2148 ArgValue = DAG.getExtLoad(ExtType, DL, VA.getLocVT(), Chain, FIN,
2149 MachinePointerInfo::getFixedStack(FI),
2150 MemVT, false, false, false, 0);
2152 InVals.push_back(ArgValue);
2158 if (!Subtarget->isTargetDarwin()) {
2159 // The AAPCS variadic function ABI is identical to the non-variadic
2160 // one. As a result there may be more arguments in registers and we should
2161 // save them for future reference.
2162 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
2165 AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>();
2166 // This will point to the next argument passed via stack.
2167 unsigned StackOffset = CCInfo.getNextStackOffset();
2168 // We currently pass all varargs at 8-byte alignment.
2169 StackOffset = ((StackOffset + 7) & ~7);
2170 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
2173 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2174 unsigned StackArgSize = CCInfo.getNextStackOffset();
2175 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2176 if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
2177 // This is a non-standard ABI so by fiat I say we're allowed to make full
2178 // use of the stack area to be popped, which must be aligned to 16 bytes in
2180 StackArgSize = RoundUpToAlignment(StackArgSize, 16);
2182 // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
2183 // a multiple of 16.
2184 FuncInfo->setArgumentStackToRestore(StackArgSize);
2186 // This realignment carries over to the available bytes below. Our own
2187 // callers will guarantee the space is free by giving an aligned value to
2190 // Even if we're not expected to free up the space, it's useful to know how
2191 // much is there while considering tail calls (because we can reuse it).
2192 FuncInfo->setBytesInStackArgArea(StackArgSize);
2197 void AArch64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
2198 SelectionDAG &DAG, SDLoc DL,
2199 SDValue &Chain) const {
2200 MachineFunction &MF = DAG.getMachineFunction();
2201 MachineFrameInfo *MFI = MF.getFrameInfo();
2202 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2204 SmallVector<SDValue, 8> MemOps;
2206 static const MCPhysReg GPRArgRegs[] = { AArch64::X0, AArch64::X1, AArch64::X2,
2207 AArch64::X3, AArch64::X4, AArch64::X5,
2208 AArch64::X6, AArch64::X7 };
2209 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
2210 unsigned FirstVariadicGPR =
2211 CCInfo.getFirstUnallocated(GPRArgRegs, NumGPRArgRegs);
2213 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
2215 if (GPRSaveSize != 0) {
2216 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
2218 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
2220 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
2221 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &AArch64::GPR64RegClass);
2222 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
2224 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2225 MachinePointerInfo::getStack(i * 8), false, false, 0);
2226 MemOps.push_back(Store);
2227 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2228 DAG.getConstant(8, getPointerTy()));
2231 FuncInfo->setVarArgsGPRIndex(GPRIdx);
2232 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
2234 if (Subtarget->hasFPARMv8()) {
2235 static const MCPhysReg FPRArgRegs[] = {
2236 AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3,
2237 AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7};
2238 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
2239 unsigned FirstVariadicFPR =
2240 CCInfo.getFirstUnallocated(FPRArgRegs, NumFPRArgRegs);
2242 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
2244 if (FPRSaveSize != 0) {
2245 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
2247 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
2249 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
2250 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &AArch64::FPR128RegClass);
2251 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);
2254 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2255 MachinePointerInfo::getStack(i * 16), false, false, 0);
2256 MemOps.push_back(Store);
2257 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2258 DAG.getConstant(16, getPointerTy()));
2261 FuncInfo->setVarArgsFPRIndex(FPRIdx);
2262 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
2265 if (!MemOps.empty()) {
2266 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
2270 /// LowerCallResult - Lower the result values of a call into the
2271 /// appropriate copies out of appropriate physical registers.
2272 SDValue AArch64TargetLowering::LowerCallResult(
2273 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
2274 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2275 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
2276 SDValue ThisVal) const {
2277 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2278 ? RetCC_AArch64_WebKit_JS
2279 : RetCC_AArch64_AAPCS;
2280 // Assign locations to each value returned by this call.
2281 SmallVector<CCValAssign, 16> RVLocs;
2282 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2284 CCInfo.AnalyzeCallResult(Ins, RetCC);
2286 // Copy all of the result registers out of their specified physreg.
2287 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2288 CCValAssign VA = RVLocs[i];
2290 // Pass 'this' value directly from the argument to return value, to avoid
2291 // reg unit interference
2292 if (i == 0 && isThisReturn) {
2293 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
2294 "unexpected return calling convention register assignment");
2295 InVals.push_back(ThisVal);
2300 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
2301 Chain = Val.getValue(1);
2302 InFlag = Val.getValue(2);
2304 switch (VA.getLocInfo()) {
2306 llvm_unreachable("Unknown loc info!");
2307 case CCValAssign::Full:
2309 case CCValAssign::BCvt:
2310 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2314 InVals.push_back(Val);
2320 bool AArch64TargetLowering::isEligibleForTailCallOptimization(
2321 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
2322 bool isCalleeStructRet, bool isCallerStructRet,
2323 const SmallVectorImpl<ISD::OutputArg> &Outs,
2324 const SmallVectorImpl<SDValue> &OutVals,
2325 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
2326 // For CallingConv::C this function knows whether the ABI needs
2327 // changing. That's not true for other conventions so they will have to opt in
2329 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
2332 const MachineFunction &MF = DAG.getMachineFunction();
2333 const Function *CallerF = MF.getFunction();
2334 CallingConv::ID CallerCC = CallerF->getCallingConv();
2335 bool CCMatch = CallerCC == CalleeCC;
2337 // Byval parameters hand the function a pointer directly into the stack area
2338 // we want to reuse during a tail call. Working around this *is* possible (see
2339 // X86) but less efficient and uglier in LowerCall.
2340 for (Function::const_arg_iterator i = CallerF->arg_begin(),
2341 e = CallerF->arg_end();
2343 if (i->hasByValAttr())
2346 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
2347 if (IsTailCallConvention(CalleeCC) && CCMatch)
2352 // Externally-defined functions with weak linkage should not be
2353 // tail-called on AArch64 when the OS does not support dynamic
2354 // pre-emption of symbols, as the AAELF spec requires normal calls
2355 // to undefined weak functions to be replaced with a NOP or jump to the
2356 // next instruction. The behaviour of branch instructions in this
2357 // situation (as used for tail calls) is implementation-defined, so we
2358 // cannot rely on the linker replacing the tail call with a return.
2359 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2360 const GlobalValue *GV = G->getGlobal();
2361 const Triple TT(getTargetMachine().getTargetTriple());
2362 if (GV->hasExternalWeakLinkage() &&
2363 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2367 // Now we search for cases where we can use a tail call without changing the
2368 // ABI. Sibcall is used in some places (particularly gcc) to refer to this
2371 // I want anyone implementing a new calling convention to think long and hard
2372 // about this assert.
2373 assert((!isVarArg || CalleeCC == CallingConv::C) &&
2374 "Unexpected variadic calling convention");
2376 if (isVarArg && !Outs.empty()) {
2377 // At least two cases here: if caller is fastcc then we can't have any
2378 // memory arguments (we'd be expected to clean up the stack afterwards). If
2379 // caller is C then we could potentially use its argument area.
2381 // FIXME: for now we take the most conservative of these in both cases:
2382 // disallow all variadic memory operands.
2383 SmallVector<CCValAssign, 16> ArgLocs;
2384 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2387 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, true));
2388 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2389 if (!ArgLocs[i].isRegLoc())
2393 // If the calling conventions do not match, then we'd better make sure the
2394 // results are returned in the same way as what the caller expects.
2396 SmallVector<CCValAssign, 16> RVLocs1;
2397 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2399 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForCall(CalleeCC, isVarArg));
2401 SmallVector<CCValAssign, 16> RVLocs2;
2402 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2404 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForCall(CallerCC, isVarArg));
2406 if (RVLocs1.size() != RVLocs2.size())
2408 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2409 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2411 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2413 if (RVLocs1[i].isRegLoc()) {
2414 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2417 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2423 // Nothing more to check if the callee is taking no arguments
2427 SmallVector<CCValAssign, 16> ArgLocs;
2428 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2431 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, isVarArg));
2433 const AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2435 // If the stack arguments for this call would fit into our own save area then
2436 // the call can be made tail.
2437 return CCInfo.getNextStackOffset() <= FuncInfo->getBytesInStackArgArea();
2440 SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain,
2442 MachineFrameInfo *MFI,
2443 int ClobberedFI) const {
2444 SmallVector<SDValue, 8> ArgChains;
2445 int64_t FirstByte = MFI->getObjectOffset(ClobberedFI);
2446 int64_t LastByte = FirstByte + MFI->getObjectSize(ClobberedFI) - 1;
2448 // Include the original chain at the beginning of the list. When this is
2449 // used by target LowerCall hooks, this helps legalize find the
2450 // CALLSEQ_BEGIN node.
2451 ArgChains.push_back(Chain);
2453 // Add a chain value for each stack argument corresponding
2454 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
2455 UE = DAG.getEntryNode().getNode()->use_end();
2457 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
2458 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
2459 if (FI->getIndex() < 0) {
2460 int64_t InFirstByte = MFI->getObjectOffset(FI->getIndex());
2461 int64_t InLastByte = InFirstByte;
2462 InLastByte += MFI->getObjectSize(FI->getIndex()) - 1;
2464 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
2465 (FirstByte <= InFirstByte && InFirstByte <= LastByte))
2466 ArgChains.push_back(SDValue(L, 1));
2469 // Build a tokenfactor for all the chains.
2470 return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
2473 bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
2474 bool TailCallOpt) const {
2475 return CallCC == CallingConv::Fast && TailCallOpt;
2478 bool AArch64TargetLowering::IsTailCallConvention(CallingConv::ID CallCC) const {
2479 return CallCC == CallingConv::Fast;
2482 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
2483 /// and add input and output parameter nodes.
2485 AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
2486 SmallVectorImpl<SDValue> &InVals) const {
2487 SelectionDAG &DAG = CLI.DAG;
2489 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2490 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2491 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2492 SDValue Chain = CLI.Chain;
2493 SDValue Callee = CLI.Callee;
2494 bool &IsTailCall = CLI.IsTailCall;
2495 CallingConv::ID CallConv = CLI.CallConv;
2496 bool IsVarArg = CLI.IsVarArg;
2498 MachineFunction &MF = DAG.getMachineFunction();
2499 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2500 bool IsThisReturn = false;
2502 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2503 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2504 bool IsSibCall = false;
2507 // Check if it's really possible to do a tail call.
2508 IsTailCall = isEligibleForTailCallOptimization(
2509 Callee, CallConv, IsVarArg, IsStructRet,
2510 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2511 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
2512 report_fatal_error("failed to perform tail call elimination on a call "
2513 "site marked musttail");
2515 // A sibling call is one where we're under the usual C ABI and not planning
2516 // to change that but can still do a tail call:
2517 if (!TailCallOpt && IsTailCall)
2524 // Analyze operands of the call, assigning locations to each operand.
2525 SmallVector<CCValAssign, 16> ArgLocs;
2526 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
2530 // Handle fixed and variable vector arguments differently.
2531 // Variable vector arguments always go into memory.
2532 unsigned NumArgs = Outs.size();
2534 for (unsigned i = 0; i != NumArgs; ++i) {
2535 MVT ArgVT = Outs[i].VT;
2536 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2537 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2538 /*IsVarArg=*/ !Outs[i].IsFixed);
2539 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2540 assert(!Res && "Call operand has unhandled type");
2544 // At this point, Outs[].VT may already be promoted to i32. To correctly
2545 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2546 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2547 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2548 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2550 unsigned NumArgs = Outs.size();
2551 for (unsigned i = 0; i != NumArgs; ++i) {
2552 MVT ValVT = Outs[i].VT;
2553 // Get type of the original argument.
2554 EVT ActualVT = getValueType(CLI.getArgs()[Outs[i].OrigArgIndex].Ty,
2555 /*AllowUnknown*/ true);
2556 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2557 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2558 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2559 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2561 else if (ActualMVT == MVT::i16)
2564 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2565 bool Res = AssignFn(i, ValVT, ValVT, CCValAssign::Full, ArgFlags, CCInfo);
2566 assert(!Res && "Call operand has unhandled type");
2571 // Get a count of how many bytes are to be pushed on the stack.
2572 unsigned NumBytes = CCInfo.getNextStackOffset();
2575 // Since we're not changing the ABI to make this a tail call, the memory
2576 // operands are already available in the caller's incoming argument space.
2580 // FPDiff is the byte offset of the call's argument area from the callee's.
2581 // Stores to callee stack arguments will be placed in FixedStackSlots offset
2582 // by this amount for a tail call. In a sibling call it must be 0 because the
2583 // caller will deallocate the entire stack and the callee still expects its
2584 // arguments to begin at SP+0. Completely unused for non-tail calls.
2587 if (IsTailCall && !IsSibCall) {
2588 unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();
2590 // Since callee will pop argument stack as a tail call, we must keep the
2591 // popped size 16-byte aligned.
2592 NumBytes = RoundUpToAlignment(NumBytes, 16);
2594 // FPDiff will be negative if this tail call requires more space than we
2595 // would automatically have in our incoming argument space. Positive if we
2596 // can actually shrink the stack.
2597 FPDiff = NumReusableBytes - NumBytes;
2599 // The stack pointer must be 16-byte aligned at all times it's used for a
2600 // memory operation, which in practice means at *all* times and in
2601 // particular across call boundaries. Therefore our own arguments started at
2602 // a 16-byte aligned SP and the delta applied for the tail call should
2603 // satisfy the same constraint.
2604 assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
2607 // Adjust the stack pointer for the new arguments...
2608 // These operations are automatically eliminated by the prolog/epilog pass
2611 DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true), DL);
2613 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, AArch64::SP, getPointerTy());
2615 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2616 SmallVector<SDValue, 8> MemOpChains;
2618 // Walk the register/memloc assignments, inserting copies/loads.
2619 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2620 ++i, ++realArgIdx) {
2621 CCValAssign &VA = ArgLocs[i];
2622 SDValue Arg = OutVals[realArgIdx];
2623 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2625 // Promote the value if needed.
2626 switch (VA.getLocInfo()) {
2628 llvm_unreachable("Unknown loc info!");
2629 case CCValAssign::Full:
2631 case CCValAssign::SExt:
2632 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
2634 case CCValAssign::ZExt:
2635 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2637 case CCValAssign::AExt:
2638 if (Outs[realArgIdx].ArgVT == MVT::i1) {
2639 // AAPCS requires i1 to be zero-extended to 8-bits by the caller.
2640 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
2641 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i8, Arg);
2643 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
2645 case CCValAssign::BCvt:
2646 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2648 case CCValAssign::FPExt:
2649 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
2653 if (VA.isRegLoc()) {
2654 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
2655 assert(VA.getLocVT() == MVT::i64 &&
2656 "unexpected calling convention register assignment");
2657 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
2658 "unexpected use of 'returned'");
2659 IsThisReturn = true;
2661 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2663 assert(VA.isMemLoc());
2666 MachinePointerInfo DstInfo;
2668 // FIXME: This works on big-endian for composite byvals, which are the
2669 // common case. It should also work for fundamental types too.
2670 uint32_t BEAlign = 0;
2671 unsigned OpSize = Flags.isByVal() ? Flags.getByValSize() * 8
2672 : VA.getValVT().getSizeInBits();
2673 OpSize = (OpSize + 7) / 8;
2674 if (!Subtarget->isLittleEndian() && !Flags.isByVal() &&
2675 !Flags.isInConsecutiveRegs()) {
2677 BEAlign = 8 - OpSize;
2679 unsigned LocMemOffset = VA.getLocMemOffset();
2680 int32_t Offset = LocMemOffset + BEAlign;
2681 SDValue PtrOff = DAG.getIntPtrConstant(Offset);
2682 PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2685 Offset = Offset + FPDiff;
2686 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2688 DstAddr = DAG.getFrameIndex(FI, getPointerTy());
2689 DstInfo = MachinePointerInfo::getFixedStack(FI);
2691 // Make sure any stack arguments overlapping with where we're storing
2692 // are loaded before this eventual operation. Otherwise they'll be
2694 Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
2696 SDValue PtrOff = DAG.getIntPtrConstant(Offset);
2698 DstAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2699 DstInfo = MachinePointerInfo::getStack(LocMemOffset);
2702 if (Outs[i].Flags.isByVal()) {
2704 DAG.getConstant(Outs[i].Flags.getByValSize(), MVT::i64);
2705 SDValue Cpy = DAG.getMemcpy(
2706 Chain, DL, DstAddr, Arg, SizeNode, Outs[i].Flags.getByValAlign(),
2708 /*AlwaysInline = */ false, DstInfo, MachinePointerInfo());
2710 MemOpChains.push_back(Cpy);
2712 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
2713 // promoted to a legal register type i32, we should truncate Arg back to
2715 if (VA.getValVT() == MVT::i1 || VA.getValVT() == MVT::i8 ||
2716 VA.getValVT() == MVT::i16)
2717 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Arg);
2720 DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo, false, false, 0);
2721 MemOpChains.push_back(Store);
2726 if (!MemOpChains.empty())
2727 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
2729 // Build a sequence of copy-to-reg nodes chained together with token chain
2730 // and flag operands which copy the outgoing args into the appropriate regs.
2732 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2733 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[i].first,
2734 RegsToPass[i].second, InFlag);
2735 InFlag = Chain.getValue(1);
2738 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2739 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2740 // node so that legalize doesn't hack it.
2741 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
2742 Subtarget->isTargetMachO()) {
2743 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2744 const GlobalValue *GV = G->getGlobal();
2745 bool InternalLinkage = GV->hasInternalLinkage();
2746 if (InternalLinkage)
2747 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2749 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0,
2751 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, getPointerTy(), Callee);
2753 } else if (ExternalSymbolSDNode *S =
2754 dyn_cast<ExternalSymbolSDNode>(Callee)) {
2755 const char *Sym = S->getSymbol();
2757 DAG.getTargetExternalSymbol(Sym, getPointerTy(), AArch64II::MO_GOT);
2758 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, getPointerTy(), Callee);
2760 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2761 const GlobalValue *GV = G->getGlobal();
2762 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2763 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2764 const char *Sym = S->getSymbol();
2765 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), 0);
2768 // We don't usually want to end the call-sequence here because we would tidy
2769 // the frame up *after* the call, however in the ABI-changing tail-call case
2770 // we've carefully laid out the parameters so that when sp is reset they'll be
2771 // in the correct location.
2772 if (IsTailCall && !IsSibCall) {
2773 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2774 DAG.getIntPtrConstant(0, true), InFlag, DL);
2775 InFlag = Chain.getValue(1);
2778 std::vector<SDValue> Ops;
2779 Ops.push_back(Chain);
2780 Ops.push_back(Callee);
2783 // Each tail call may have to adjust the stack by a different amount, so
2784 // this information must travel along with the operation for eventual
2785 // consumption by emitEpilogue.
2786 Ops.push_back(DAG.getTargetConstant(FPDiff, MVT::i32));
2789 // Add argument registers to the end of the list so that they are known live
2791 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2792 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2793 RegsToPass[i].second.getValueType()));
2795 // Add a register mask operand representing the call-preserved registers.
2796 const uint32_t *Mask;
2797 const TargetRegisterInfo *TRI =
2798 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
2799 const AArch64RegisterInfo *ARI =
2800 static_cast<const AArch64RegisterInfo *>(TRI);
2802 // For 'this' returns, use the X0-preserving mask if applicable
2803 Mask = ARI->getThisReturnPreservedMask(CallConv);
2805 IsThisReturn = false;
2806 Mask = ARI->getCallPreservedMask(CallConv);
2809 Mask = ARI->getCallPreservedMask(CallConv);
2811 assert(Mask && "Missing call preserved mask for calling convention");
2812 Ops.push_back(DAG.getRegisterMask(Mask));
2814 if (InFlag.getNode())
2815 Ops.push_back(InFlag);
2817 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2819 // If we're doing a tall call, use a TC_RETURN here rather than an
2820 // actual call instruction.
2822 return DAG.getNode(AArch64ISD::TC_RETURN, DL, NodeTys, Ops);
2824 // Returns a chain and a flag for retval copy to use.
2825 Chain = DAG.getNode(AArch64ISD::CALL, DL, NodeTys, Ops);
2826 InFlag = Chain.getValue(1);
2828 uint64_t CalleePopBytes = DoesCalleeRestoreStack(CallConv, TailCallOpt)
2829 ? RoundUpToAlignment(NumBytes, 16)
2832 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2833 DAG.getIntPtrConstant(CalleePopBytes, true),
2836 InFlag = Chain.getValue(1);
2838 // Handle result values, copying them out of physregs into vregs that we
2840 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2841 InVals, IsThisReturn,
2842 IsThisReturn ? OutVals[0] : SDValue());
2845 bool AArch64TargetLowering::CanLowerReturn(
2846 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2847 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2848 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2849 ? RetCC_AArch64_WebKit_JS
2850 : RetCC_AArch64_AAPCS;
2851 SmallVector<CCValAssign, 16> RVLocs;
2852 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2853 return CCInfo.CheckReturn(Outs, RetCC);
2857 AArch64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2859 const SmallVectorImpl<ISD::OutputArg> &Outs,
2860 const SmallVectorImpl<SDValue> &OutVals,
2861 SDLoc DL, SelectionDAG &DAG) const {
2862 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2863 ? RetCC_AArch64_WebKit_JS
2864 : RetCC_AArch64_AAPCS;
2865 SmallVector<CCValAssign, 16> RVLocs;
2866 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2868 CCInfo.AnalyzeReturn(Outs, RetCC);
2870 // Copy the result values into the output registers.
2872 SmallVector<SDValue, 4> RetOps(1, Chain);
2873 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
2874 ++i, ++realRVLocIdx) {
2875 CCValAssign &VA = RVLocs[i];
2876 assert(VA.isRegLoc() && "Can only return in registers!");
2877 SDValue Arg = OutVals[realRVLocIdx];
2879 switch (VA.getLocInfo()) {
2881 llvm_unreachable("Unknown loc info!");
2882 case CCValAssign::Full:
2883 if (Outs[i].ArgVT == MVT::i1) {
2884 // AAPCS requires i1 to be zero-extended to i8 by the producer of the
2885 // value. This is strictly redundant on Darwin (which uses "zeroext
2886 // i1"), but will be optimised out before ISel.
2887 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
2888 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2891 case CCValAssign::BCvt:
2892 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2896 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
2897 Flag = Chain.getValue(1);
2898 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2901 RetOps[0] = Chain; // Update chain.
2903 // Add the flag if we have it.
2905 RetOps.push_back(Flag);
2907 return DAG.getNode(AArch64ISD::RET_FLAG, DL, MVT::Other, RetOps);
2910 //===----------------------------------------------------------------------===//
2911 // Other Lowering Code
2912 //===----------------------------------------------------------------------===//
2914 SDValue AArch64TargetLowering::LowerGlobalAddress(SDValue Op,
2915 SelectionDAG &DAG) const {
2916 EVT PtrVT = getPointerTy();
2918 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
2919 const GlobalValue *GV = GN->getGlobal();
2920 unsigned char OpFlags =
2921 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
2923 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
2924 "unexpected offset in global node");
2926 // This also catched the large code model case for Darwin.
2927 if ((OpFlags & AArch64II::MO_GOT) != 0) {
2928 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
2929 // FIXME: Once remat is capable of dealing with instructions with register
2930 // operands, expand this into two nodes instead of using a wrapper node.
2931 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
2934 if ((OpFlags & AArch64II::MO_CONSTPOOL) != 0) {
2935 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
2936 "use of MO_CONSTPOOL only supported on small model");
2937 SDValue Hi = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, AArch64II::MO_PAGE);
2938 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
2939 unsigned char LoFlags = AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
2940 SDValue Lo = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, LoFlags);
2941 SDValue PoolAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
2942 SDValue GlobalAddr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), PoolAddr,
2943 MachinePointerInfo::getConstantPool(),
2944 /*isVolatile=*/ false,
2945 /*isNonTemporal=*/ true,
2946 /*isInvariant=*/ true, 8);
2947 if (GN->getOffset() != 0)
2948 return DAG.getNode(ISD::ADD, DL, PtrVT, GlobalAddr,
2949 DAG.getConstant(GN->getOffset(), PtrVT));
2953 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2954 const unsigned char MO_NC = AArch64II::MO_NC;
2956 AArch64ISD::WrapperLarge, DL, PtrVT,
2957 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G3),
2958 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
2959 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
2960 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
2962 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
2963 // the only correct model on Darwin.
2964 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2965 OpFlags | AArch64II::MO_PAGE);
2966 unsigned char LoFlags = OpFlags | AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
2967 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
2969 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
2970 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
2974 /// \brief Convert a TLS address reference into the correct sequence of loads
2975 /// and calls to compute the variable's address (for Darwin, currently) and
2976 /// return an SDValue containing the final node.
2978 /// Darwin only has one TLS scheme which must be capable of dealing with the
2979 /// fully general situation, in the worst case. This means:
2980 /// + "extern __thread" declaration.
2981 /// + Defined in a possibly unknown dynamic library.
2983 /// The general system is that each __thread variable has a [3 x i64] descriptor
2984 /// which contains information used by the runtime to calculate the address. The
2985 /// only part of this the compiler needs to know about is the first xword, which
2986 /// contains a function pointer that must be called with the address of the
2987 /// entire descriptor in "x0".
2989 /// Since this descriptor may be in a different unit, in general even the
2990 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
2992 /// adrp x0, _var@TLVPPAGE
2993 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
2994 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
2995 /// ; the function pointer
2996 /// blr x1 ; Uses descriptor address in x0
2997 /// ; Address of _var is now in x0.
2999 /// If the address of _var's descriptor *is* known to the linker, then it can
3000 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
3001 /// a slight efficiency gain.
3003 AArch64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
3004 SelectionDAG &DAG) const {
3005 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
3008 MVT PtrVT = getPointerTy();
3009 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3012 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3013 SDValue DescAddr = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TLVPAddr);
3015 // The first entry in the descriptor is a function pointer that we must call
3016 // to obtain the address of the variable.
3017 SDValue Chain = DAG.getEntryNode();
3018 SDValue FuncTLVGet =
3019 DAG.getLoad(MVT::i64, DL, Chain, DescAddr, MachinePointerInfo::getGOT(),
3020 false, true, true, 8);
3021 Chain = FuncTLVGet.getValue(1);
3023 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3024 MFI->setAdjustsStack(true);
3026 // TLS calls preserve all registers except those that absolutely must be
3027 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
3029 const TargetRegisterInfo *TRI =
3030 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
3031 const AArch64RegisterInfo *ARI =
3032 static_cast<const AArch64RegisterInfo *>(TRI);
3033 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
3035 // Finally, we can make the call. This is just a degenerate version of a
3036 // normal AArch64 call node: x0 takes the address of the descriptor, and
3037 // returns the address of the variable in this thread.
3038 Chain = DAG.getCopyToReg(Chain, DL, AArch64::X0, DescAddr, SDValue());
3040 DAG.getNode(AArch64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
3041 Chain, FuncTLVGet, DAG.getRegister(AArch64::X0, MVT::i64),
3042 DAG.getRegisterMask(Mask), Chain.getValue(1));
3043 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Chain.getValue(1));
3046 /// When accessing thread-local variables under either the general-dynamic or
3047 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
3048 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
3049 /// is a function pointer to carry out the resolution. This function takes the
3050 /// address of the descriptor in X0 and returns the TPIDR_EL0 offset in X0. All
3051 /// other registers (except LR, NZCV) are preserved.
3053 /// Thus, the ideal call sequence on AArch64 is:
3055 /// adrp x0, :tlsdesc:thread_var
3056 /// ldr x8, [x0, :tlsdesc_lo12:thread_var]
3057 /// add x0, x0, :tlsdesc_lo12:thread_var
3058 /// .tlsdesccall thread_var
3060 /// (TPIDR_EL0 offset now in x0).
3062 /// The ".tlsdesccall" directive instructs the assembler to insert a particular
3063 /// relocation to help the linker relax this sequence if it turns out to be too
3066 /// FIXME: we currently produce an extra, duplicated, ADRP instruction, but this
3068 SDValue AArch64TargetLowering::LowerELFTLSDescCall(SDValue SymAddr,
3069 SDValue DescAddr, SDLoc DL,
3070 SelectionDAG &DAG) const {
3071 EVT PtrVT = getPointerTy();
3073 // The function we need to call is simply the first entry in the GOT for this
3074 // descriptor, load it in preparation.
3075 SDValue Func = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, SymAddr);
3077 // TLS calls preserve all registers except those that absolutely must be
3078 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
3080 const TargetRegisterInfo *TRI =
3081 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
3082 const AArch64RegisterInfo *ARI =
3083 static_cast<const AArch64RegisterInfo *>(TRI);
3084 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
3086 // The function takes only one argument: the address of the descriptor itself
3088 SDValue Glue, Chain;
3089 Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, AArch64::X0, DescAddr, Glue);
3090 Glue = Chain.getValue(1);
3092 // We're now ready to populate the argument list, as with a normal call:
3093 SmallVector<SDValue, 6> Ops;
3094 Ops.push_back(Chain);
3095 Ops.push_back(Func);
3096 Ops.push_back(SymAddr);
3097 Ops.push_back(DAG.getRegister(AArch64::X0, PtrVT));
3098 Ops.push_back(DAG.getRegisterMask(Mask));
3099 Ops.push_back(Glue);
3101 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3102 Chain = DAG.getNode(AArch64ISD::TLSDESC_CALL, DL, NodeTys, Ops);
3103 Glue = Chain.getValue(1);
3105 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue);
3109 AArch64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
3110 SelectionDAG &DAG) const {
3111 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
3112 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
3113 "ELF TLS only supported in small memory model");
3114 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3116 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
3119 EVT PtrVT = getPointerTy();
3121 const GlobalValue *GV = GA->getGlobal();
3123 SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);
3125 if (Model == TLSModel::LocalExec) {
3126 SDValue HiVar = DAG.getTargetGlobalAddress(
3127 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_G1);
3128 SDValue LoVar = DAG.getTargetGlobalAddress(
3130 AArch64II::MO_TLS | AArch64II::MO_G0 | AArch64II::MO_NC);
3132 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVZXi, DL, PtrVT, HiVar,
3133 DAG.getTargetConstant(16, MVT::i32)),
3135 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKXi, DL, PtrVT, TPOff, LoVar,
3136 DAG.getTargetConstant(0, MVT::i32)),
3138 } else if (Model == TLSModel::InitialExec) {
3139 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3140 TPOff = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TPOff);
3141 } else if (Model == TLSModel::LocalDynamic) {
3142 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
3143 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
3144 // the beginning of the module's TLS region, followed by a DTPREL offset
3147 // These accesses will need deduplicating if there's more than one.
3148 AArch64FunctionInfo *MFI =
3149 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
3150 MFI->incNumLocalDynamicTLSAccesses();
3152 // Accesses used in this sequence go via the TLS descriptor which lives in
3153 // the GOT. Prepare an address we can use to handle this.
3154 SDValue HiDesc = DAG.getTargetExternalSymbol(
3155 "_TLS_MODULE_BASE_", PtrVT, AArch64II::MO_TLS | AArch64II::MO_PAGE);
3156 SDValue LoDesc = DAG.getTargetExternalSymbol(
3157 "_TLS_MODULE_BASE_", PtrVT,
3158 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3160 // First argument to the descriptor call is the address of the descriptor
3162 SDValue DescAddr = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, HiDesc);
3163 DescAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
3165 // The call needs a relocation too for linker relaxation. It doesn't make
3166 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3168 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
3171 // Now we can calculate the offset from TPIDR_EL0 to this module's
3172 // thread-local area.
3173 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
3175 // Now use :dtprel_whatever: operations to calculate this variable's offset
3176 // in its thread-storage area.
3177 SDValue HiVar = DAG.getTargetGlobalAddress(
3178 GV, DL, MVT::i64, 0, AArch64II::MO_TLS | AArch64II::MO_G1);
3179 SDValue LoVar = DAG.getTargetGlobalAddress(
3180 GV, DL, MVT::i64, 0,
3181 AArch64II::MO_TLS | AArch64II::MO_G0 | AArch64II::MO_NC);
3184 SDValue(DAG.getMachineNode(AArch64::MOVZXi, DL, PtrVT, HiVar,
3185 DAG.getTargetConstant(16, MVT::i32)),
3188 SDValue(DAG.getMachineNode(AArch64::MOVKXi, DL, PtrVT, DTPOff, LoVar,
3189 DAG.getTargetConstant(0, MVT::i32)),
3192 TPOff = DAG.getNode(ISD::ADD, DL, PtrVT, TPOff, DTPOff);
3193 } else if (Model == TLSModel::GeneralDynamic) {
3194 // Accesses used in this sequence go via the TLS descriptor which lives in
3195 // the GOT. Prepare an address we can use to handle this.
3196 SDValue HiDesc = DAG.getTargetGlobalAddress(
3197 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_PAGE);
3198 SDValue LoDesc = DAG.getTargetGlobalAddress(
3200 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3202 // First argument to the descriptor call is the address of the descriptor
3204 SDValue DescAddr = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, HiDesc);
3205 DescAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
3207 // The call needs a relocation too for linker relaxation. It doesn't make
3208 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3211 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3213 // Finally we can make a call to calculate the offset from tpidr_el0.
3214 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
3216 llvm_unreachable("Unsupported ELF TLS access model");
3218 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
3221 SDValue AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
3222 SelectionDAG &DAG) const {
3223 if (Subtarget->isTargetDarwin())
3224 return LowerDarwinGlobalTLSAddress(Op, DAG);
3225 else if (Subtarget->isTargetELF())
3226 return LowerELFGlobalTLSAddress(Op, DAG);
3228 llvm_unreachable("Unexpected platform trying to use TLS");
3230 SDValue AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3231 SDValue Chain = Op.getOperand(0);
3232 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3233 SDValue LHS = Op.getOperand(2);
3234 SDValue RHS = Op.getOperand(3);
3235 SDValue Dest = Op.getOperand(4);
3238 // Handle f128 first, since lowering it will result in comparing the return
3239 // value of a libcall against zero, which is just what the rest of LowerBR_CC
3240 // is expecting to deal with.
3241 if (LHS.getValueType() == MVT::f128) {
3242 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3244 // If softenSetCCOperands returned a scalar, we need to compare the result
3245 // against zero to select between true and false values.
3246 if (!RHS.getNode()) {
3247 RHS = DAG.getConstant(0, LHS.getValueType());
3252 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
3254 unsigned Opc = LHS.getOpcode();
3255 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
3256 cast<ConstantSDNode>(RHS)->isOne() &&
3257 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3258 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3259 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
3260 "Unexpected condition code.");
3261 // Only lower legal XALUO ops.
3262 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
3265 // The actual operation with overflow check.
3266 AArch64CC::CondCode OFCC;
3267 SDValue Value, Overflow;
3268 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, LHS.getValue(0), DAG);
3270 if (CC == ISD::SETNE)
3271 OFCC = getInvertedCondCode(OFCC);
3272 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3274 return DAG.getNode(AArch64ISD::BRCOND, SDLoc(LHS), MVT::Other, Chain, Dest,
3278 if (LHS.getValueType().isInteger()) {
3279 assert((LHS.getValueType() == RHS.getValueType()) &&
3280 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3282 // If the RHS of the comparison is zero, we can potentially fold this
3283 // to a specialized branch.
3284 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
3285 if (RHSC && RHSC->getZExtValue() == 0) {
3286 if (CC == ISD::SETEQ) {
3287 // See if we can use a TBZ to fold in an AND as well.
3288 // TBZ has a smaller branch displacement than CBZ. If the offset is
3289 // out of bounds, a late MI-layer pass rewrites branches.
3290 // 403.gcc is an example that hits this case.
3291 if (LHS.getOpcode() == ISD::AND &&
3292 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3293 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3294 SDValue Test = LHS.getOperand(0);
3295 uint64_t Mask = LHS.getConstantOperandVal(1);
3296 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, Test,
3297 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3300 return DAG.getNode(AArch64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
3301 } else if (CC == ISD::SETNE) {
3302 // See if we can use a TBZ to fold in an AND as well.
3303 // TBZ has a smaller branch displacement than CBZ. If the offset is
3304 // out of bounds, a late MI-layer pass rewrites branches.
3305 // 403.gcc is an example that hits this case.
3306 if (LHS.getOpcode() == ISD::AND &&
3307 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3308 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3309 SDValue Test = LHS.getOperand(0);
3310 uint64_t Mask = LHS.getConstantOperandVal(1);
3311 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, Test,
3312 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3315 return DAG.getNode(AArch64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
3316 } else if (CC == ISD::SETLT && LHS.getOpcode() != ISD::AND) {
3317 // Don't combine AND since emitComparison converts the AND to an ANDS
3318 // (a.k.a. TST) and the test in the test bit and branch instruction
3319 // becomes redundant. This would also increase register pressure.
3320 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3321 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, LHS,
3322 DAG.getConstant(Mask, MVT::i64), Dest);
3325 if (RHSC && RHSC->getSExtValue() == -1 && CC == ISD::SETGT &&
3326 LHS.getOpcode() != ISD::AND) {
3327 // Don't combine AND since emitComparison converts the AND to an ANDS
3328 // (a.k.a. TST) and the test in the test bit and branch instruction
3329 // becomes redundant. This would also increase register pressure.
3330 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3331 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, LHS,
3332 DAG.getConstant(Mask, MVT::i64), Dest);
3336 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3337 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3341 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3343 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3344 // clean. Some of them require two branches to implement.
3345 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3346 AArch64CC::CondCode CC1, CC2;
3347 changeFPCCToAArch64CC(CC, CC1, CC2);
3348 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3350 DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
3351 if (CC2 != AArch64CC::AL) {
3352 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3353 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
3360 SDValue AArch64TargetLowering::LowerFCOPYSIGN(SDValue Op,
3361 SelectionDAG &DAG) const {
3362 EVT VT = Op.getValueType();
3365 SDValue In1 = Op.getOperand(0);
3366 SDValue In2 = Op.getOperand(1);
3367 EVT SrcVT = In2.getValueType();
3369 if (SrcVT == MVT::f32 && VT == MVT::f64)
3370 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
3371 else if (SrcVT == MVT::f64 && VT == MVT::f32)
3372 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0));
3374 // FIXME: Src type is different, bail out for now. Can VT really be a
3381 SDValue EltMask, VecVal1, VecVal2;
3382 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
3385 EltMask = DAG.getConstant(0x80000000ULL, EltVT);
3387 if (!VT.isVector()) {
3388 VecVal1 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3389 DAG.getUNDEF(VecVT), In1);
3390 VecVal2 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3391 DAG.getUNDEF(VecVT), In2);
3393 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3394 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3396 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
3400 // We want to materialize a mask with the the high bit set, but the AdvSIMD
3401 // immediate moves cannot materialize that in a single instruction for
3402 // 64-bit elements. Instead, materialize zero and then negate it.
3403 EltMask = DAG.getConstant(0, EltVT);
3405 if (!VT.isVector()) {
3406 VecVal1 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3407 DAG.getUNDEF(VecVT), In1);
3408 VecVal2 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3409 DAG.getUNDEF(VecVT), In2);
3411 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3412 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3415 llvm_unreachable("Invalid type for copysign!");
3418 std::vector<SDValue> BuildVectorOps;
3419 for (unsigned i = 0; i < VecVT.getVectorNumElements(); ++i)
3420 BuildVectorOps.push_back(EltMask);
3422 SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT, BuildVectorOps);
3424 // If we couldn't materialize the mask above, then the mask vector will be
3425 // the zero vector, and we need to negate it here.
3426 if (VT == MVT::f64 || VT == MVT::v2f64) {
3427 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
3428 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
3429 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
3433 DAG.getNode(AArch64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
3436 return DAG.getTargetExtractSubreg(AArch64::ssub, DL, VT, Sel);
3437 else if (VT == MVT::f64)
3438 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, VT, Sel);
3440 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
3443 SDValue AArch64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
3444 if (DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
3445 AttributeSet::FunctionIndex, Attribute::NoImplicitFloat))
3448 if (!Subtarget->hasNEON())
3451 // While there is no integer popcount instruction, it can
3452 // be more efficiently lowered to the following sequence that uses
3453 // AdvSIMD registers/instructions as long as the copies to/from
3454 // the AdvSIMD registers are cheap.
3455 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
3456 // CNT V0.8B, V0.8B // 8xbyte pop-counts
3457 // ADDV B0, V0.8B // sum 8xbyte pop-counts
3458 // UMOV X0, V0.B[0] // copy byte result back to integer reg
3459 SDValue Val = Op.getOperand(0);
3461 EVT VT = Op.getValueType();
3462 SDValue ZeroVec = DAG.getUNDEF(MVT::v8i8);
3465 if (VT == MVT::i32) {
3466 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Val);
3467 VecVal = DAG.getTargetInsertSubreg(AArch64::ssub, DL, MVT::v8i8, ZeroVec,
3470 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
3473 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, VecVal);
3474 SDValue UaddLV = DAG.getNode(
3475 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
3476 DAG.getConstant(Intrinsic::aarch64_neon_uaddlv, MVT::i32), CtPop);
3479 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
3483 SDValue AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3485 if (Op.getValueType().isVector())
3486 return LowerVSETCC(Op, DAG);
3488 SDValue LHS = Op.getOperand(0);
3489 SDValue RHS = Op.getOperand(1);
3490 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3493 // We chose ZeroOrOneBooleanContents, so use zero and one.
3494 EVT VT = Op.getValueType();
3495 SDValue TVal = DAG.getConstant(1, VT);
3496 SDValue FVal = DAG.getConstant(0, VT);
3498 // Handle f128 first, since one possible outcome is a normal integer
3499 // comparison which gets picked up by the next if statement.
3500 if (LHS.getValueType() == MVT::f128) {
3501 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3503 // If softenSetCCOperands returned a scalar, use it.
3504 if (!RHS.getNode()) {
3505 assert(LHS.getValueType() == Op.getValueType() &&
3506 "Unexpected setcc expansion!");
3511 if (LHS.getValueType().isInteger()) {
3514 getAArch64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
3516 // Note that we inverted the condition above, so we reverse the order of
3517 // the true and false operands here. This will allow the setcc to be
3518 // matched to a single CSINC instruction.
3519 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
3522 // Now we know we're dealing with FP values.
3523 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3525 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3526 // and do the comparison.
3527 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3529 AArch64CC::CondCode CC1, CC2;
3530 changeFPCCToAArch64CC(CC, CC1, CC2);
3531 if (CC2 == AArch64CC::AL) {
3532 changeFPCCToAArch64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
3533 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3535 // Note that we inverted the condition above, so we reverse the order of
3536 // the true and false operands here. This will allow the setcc to be
3537 // matched to a single CSINC instruction.
3538 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
3540 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't
3541 // totally clean. Some of them require two CSELs to implement. As is in
3542 // this case, we emit the first CSEL and then emit a second using the output
3543 // of the first as the RHS. We're effectively OR'ing the two CC's together.
3545 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
3546 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3548 DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3550 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3551 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3555 /// A SELECT_CC operation is really some kind of max or min if both values being
3556 /// compared are, in some sense, equal to the results in either case. However,
3557 /// it is permissible to compare f32 values and produce directly extended f64
3560 /// Extending the comparison operands would also be allowed, but is less likely
3561 /// to happen in practice since their use is right here. Note that truncate
3562 /// operations would *not* be semantically equivalent.
3563 static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) {
3567 ConstantFPSDNode *CCmp = dyn_cast<ConstantFPSDNode>(Cmp);
3568 ConstantFPSDNode *CResult = dyn_cast<ConstantFPSDNode>(Result);
3569 if (CCmp && CResult && Cmp.getValueType() == MVT::f32 &&
3570 Result.getValueType() == MVT::f64) {
3572 APFloat CmpVal = CCmp->getValueAPF();
3573 CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy);
3574 return CResult->getValueAPF().bitwiseIsEqual(CmpVal);
3577 return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp;
3580 SDValue AArch64TargetLowering::LowerSELECT(SDValue Op,
3581 SelectionDAG &DAG) const {
3582 SDValue CC = Op->getOperand(0);
3583 SDValue TVal = Op->getOperand(1);
3584 SDValue FVal = Op->getOperand(2);
3587 unsigned Opc = CC.getOpcode();
3588 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
3590 if (CC.getResNo() == 1 &&
3591 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3592 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3593 // Only lower legal XALUO ops.
3594 if (!DAG.getTargetLoweringInfo().isTypeLegal(CC->getValueType(0)))
3597 AArch64CC::CondCode OFCC;
3598 SDValue Value, Overflow;
3599 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, CC.getValue(0), DAG);
3600 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3602 return DAG.getNode(AArch64ISD::CSEL, DL, Op.getValueType(), TVal, FVal,
3606 if (CC.getOpcode() == ISD::SETCC)
3607 return DAG.getSelectCC(DL, CC.getOperand(0), CC.getOperand(1), TVal, FVal,
3608 cast<CondCodeSDNode>(CC.getOperand(2))->get());
3610 return DAG.getSelectCC(DL, CC, DAG.getConstant(0, CC.getValueType()), TVal,
3614 SDValue AArch64TargetLowering::LowerSELECT_CC(SDValue Op,
3615 SelectionDAG &DAG) const {
3616 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3617 SDValue LHS = Op.getOperand(0);
3618 SDValue RHS = Op.getOperand(1);
3619 SDValue TVal = Op.getOperand(2);
3620 SDValue FVal = Op.getOperand(3);
3623 // Handle f128 first, because it will result in a comparison of some RTLIB
3624 // call result against zero.
3625 if (LHS.getValueType() == MVT::f128) {
3626 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3628 // If softenSetCCOperands returned a scalar, we need to compare the result
3629 // against zero to select between true and false values.
3630 if (!RHS.getNode()) {
3631 RHS = DAG.getConstant(0, LHS.getValueType());
3636 // Handle integers first.
3637 if (LHS.getValueType().isInteger()) {
3638 assert((LHS.getValueType() == RHS.getValueType()) &&
3639 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3641 unsigned Opcode = AArch64ISD::CSEL;
3643 // If both the TVal and the FVal are constants, see if we can swap them in
3644 // order to for a CSINV or CSINC out of them.
3645 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3646 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3648 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3649 std::swap(TVal, FVal);
3650 std::swap(CTVal, CFVal);
3651 CC = ISD::getSetCCInverse(CC, true);
3652 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3653 std::swap(TVal, FVal);
3654 std::swap(CTVal, CFVal);
3655 CC = ISD::getSetCCInverse(CC, true);
3656 } else if (TVal.getOpcode() == ISD::XOR) {
3657 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3658 // with a CSINV rather than a CSEL.
3659 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
3661 if (CVal && CVal->isAllOnesValue()) {
3662 std::swap(TVal, FVal);
3663 std::swap(CTVal, CFVal);
3664 CC = ISD::getSetCCInverse(CC, true);
3666 } else if (TVal.getOpcode() == ISD::SUB) {
3667 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3668 // that we can match with a CSNEG rather than a CSEL.
3669 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
3671 if (CVal && CVal->isNullValue()) {
3672 std::swap(TVal, FVal);
3673 std::swap(CTVal, CFVal);
3674 CC = ISD::getSetCCInverse(CC, true);
3676 } else if (CTVal && CFVal) {
3677 const int64_t TrueVal = CTVal->getSExtValue();
3678 const int64_t FalseVal = CFVal->getSExtValue();
3681 // If both TVal and FVal are constants, see if FVal is the
3682 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3683 // instead of a CSEL in that case.
3684 if (TrueVal == ~FalseVal) {
3685 Opcode = AArch64ISD::CSINV;
3686 } else if (TrueVal == -FalseVal) {
3687 Opcode = AArch64ISD::CSNEG;
3688 } else if (TVal.getValueType() == MVT::i32) {
3689 // If our operands are only 32-bit wide, make sure we use 32-bit
3690 // arithmetic for the check whether we can use CSINC. This ensures that
3691 // the addition in the check will wrap around properly in case there is
3692 // an overflow (which would not be the case if we do the check with
3693 // 64-bit arithmetic).
3694 const uint32_t TrueVal32 = CTVal->getZExtValue();
3695 const uint32_t FalseVal32 = CFVal->getZExtValue();
3697 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3698 Opcode = AArch64ISD::CSINC;
3700 if (TrueVal32 > FalseVal32) {
3704 // 64-bit check whether we can use CSINC.
3705 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3706 Opcode = AArch64ISD::CSINC;
3708 if (TrueVal > FalseVal) {
3713 // Swap TVal and FVal if necessary.
3715 std::swap(TVal, FVal);
3716 std::swap(CTVal, CFVal);
3717 CC = ISD::getSetCCInverse(CC, true);
3720 if (Opcode != AArch64ISD::CSEL) {
3721 // Drop FVal since we can get its value by simply inverting/negating
3728 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3730 EVT VT = Op.getValueType();
3731 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3734 // Now we know we're dealing with FP values.
3735 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3736 assert(LHS.getValueType() == RHS.getValueType());
3737 EVT VT = Op.getValueType();
3739 // Try to match this select into a max/min operation, which have dedicated
3740 // opcode in the instruction set.
3741 // FIXME: This is not correct in the presence of NaNs, so we only enable this
3743 if (getTargetMachine().Options.NoNaNsFPMath) {
3744 SDValue MinMaxLHS = TVal, MinMaxRHS = FVal;
3745 if (selectCCOpsAreFMaxCompatible(LHS, MinMaxRHS) &&
3746 selectCCOpsAreFMaxCompatible(RHS, MinMaxLHS)) {
3747 CC = ISD::getSetCCSwappedOperands(CC);
3748 std::swap(MinMaxLHS, MinMaxRHS);
3751 if (selectCCOpsAreFMaxCompatible(LHS, MinMaxLHS) &&
3752 selectCCOpsAreFMaxCompatible(RHS, MinMaxRHS)) {
3762 return DAG.getNode(AArch64ISD::FMAX, dl, VT, MinMaxLHS, MinMaxRHS);
3770 return DAG.getNode(AArch64ISD::FMIN, dl, VT, MinMaxLHS, MinMaxRHS);
3776 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3777 // and do the comparison.
3778 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3780 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3781 // clean. Some of them require two CSELs to implement.
3782 AArch64CC::CondCode CC1, CC2;
3783 changeFPCCToAArch64CC(CC, CC1, CC2);
3784 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3785 SDValue CS1 = DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3787 // If we need a second CSEL, emit it, using the output of the first as the
3788 // RHS. We're effectively OR'ing the two CC's together.
3789 if (CC2 != AArch64CC::AL) {
3790 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3791 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3794 // Otherwise, return the output of the first CSEL.
3798 SDValue AArch64TargetLowering::LowerJumpTable(SDValue Op,
3799 SelectionDAG &DAG) const {
3800 // Jump table entries as PC relative offsets. No additional tweaking
3801 // is necessary here. Just get the address of the jump table.
3802 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3803 EVT PtrVT = getPointerTy();
3806 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3807 !Subtarget->isTargetMachO()) {
3808 const unsigned char MO_NC = AArch64II::MO_NC;
3810 AArch64ISD::WrapperLarge, DL, PtrVT,
3811 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G3),
3812 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G2 | MO_NC),
3813 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G1 | MO_NC),
3814 DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3815 AArch64II::MO_G0 | MO_NC));
3819 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_PAGE);
3820 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3821 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3822 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3823 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3826 SDValue AArch64TargetLowering::LowerConstantPool(SDValue Op,
3827 SelectionDAG &DAG) const {
3828 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3829 EVT PtrVT = getPointerTy();
3832 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3833 // Use the GOT for the large code model on iOS.
3834 if (Subtarget->isTargetMachO()) {
3835 SDValue GotAddr = DAG.getTargetConstantPool(
3836 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3838 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
3841 const unsigned char MO_NC = AArch64II::MO_NC;
3843 AArch64ISD::WrapperLarge, DL, PtrVT,
3844 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3845 CP->getOffset(), AArch64II::MO_G3),
3846 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3847 CP->getOffset(), AArch64II::MO_G2 | MO_NC),
3848 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3849 CP->getOffset(), AArch64II::MO_G1 | MO_NC),
3850 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3851 CP->getOffset(), AArch64II::MO_G0 | MO_NC));
3853 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
3854 // ELF, the only valid one on Darwin.
3856 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3857 CP->getOffset(), AArch64II::MO_PAGE);
3858 SDValue Lo = DAG.getTargetConstantPool(
3859 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3860 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3862 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3863 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3867 SDValue AArch64TargetLowering::LowerBlockAddress(SDValue Op,
3868 SelectionDAG &DAG) const {
3869 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
3870 EVT PtrVT = getPointerTy();
3872 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3873 !Subtarget->isTargetMachO()) {
3874 const unsigned char MO_NC = AArch64II::MO_NC;
3876 AArch64ISD::WrapperLarge, DL, PtrVT,
3877 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G3),
3878 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
3879 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
3880 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
3882 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGE);
3883 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGEOFF |
3885 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3886 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3890 SDValue AArch64TargetLowering::LowerDarwin_VASTART(SDValue Op,
3891 SelectionDAG &DAG) const {
3892 AArch64FunctionInfo *FuncInfo =
3893 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
3897 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3898 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3899 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
3900 MachinePointerInfo(SV), false, false, 0);
3903 SDValue AArch64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
3904 SelectionDAG &DAG) const {
3905 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3906 // Standard, section B.3.
3907 MachineFunction &MF = DAG.getMachineFunction();
3908 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
3911 SDValue Chain = Op.getOperand(0);
3912 SDValue VAList = Op.getOperand(1);
3913 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3914 SmallVector<SDValue, 4> MemOps;
3916 // void *__stack at offset 0
3918 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3919 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3920 MachinePointerInfo(SV), false, false, 8));
3922 // void *__gr_top at offset 8
3923 int GPRSize = FuncInfo->getVarArgsGPRSize();
3925 SDValue GRTop, GRTopAddr;
3927 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3928 DAG.getConstant(8, getPointerTy()));
3930 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), getPointerTy());
3931 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
3932 DAG.getConstant(GPRSize, getPointerTy()));
3934 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3935 MachinePointerInfo(SV, 8), false, false, 8));
3938 // void *__vr_top at offset 16
3939 int FPRSize = FuncInfo->getVarArgsFPRSize();
3941 SDValue VRTop, VRTopAddr;
3942 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3943 DAG.getConstant(16, getPointerTy()));
3945 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), getPointerTy());
3946 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
3947 DAG.getConstant(FPRSize, getPointerTy()));
3949 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3950 MachinePointerInfo(SV, 16), false, false, 8));
3953 // int __gr_offs at offset 24
3954 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3955 DAG.getConstant(24, getPointerTy()));
3956 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
3957 GROffsAddr, MachinePointerInfo(SV, 24), false,
3960 // int __vr_offs at offset 28
3961 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3962 DAG.getConstant(28, getPointerTy()));
3963 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
3964 VROffsAddr, MachinePointerInfo(SV, 28), false,
3967 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
3970 SDValue AArch64TargetLowering::LowerVASTART(SDValue Op,
3971 SelectionDAG &DAG) const {
3972 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
3973 : LowerAAPCS_VASTART(Op, DAG);
3976 SDValue AArch64TargetLowering::LowerVACOPY(SDValue Op,
3977 SelectionDAG &DAG) const {
3978 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
3980 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
3981 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3982 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
3984 return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op), Op.getOperand(1),
3985 Op.getOperand(2), DAG.getConstant(VaListSize, MVT::i32),
3986 8, false, false, MachinePointerInfo(DestSV),
3987 MachinePointerInfo(SrcSV));
3990 SDValue AArch64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
3991 assert(Subtarget->isTargetDarwin() &&
3992 "automatic va_arg instruction only works on Darwin");
3994 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3995 EVT VT = Op.getValueType();
3997 SDValue Chain = Op.getOperand(0);
3998 SDValue Addr = Op.getOperand(1);
3999 unsigned Align = Op.getConstantOperandVal(3);
4001 SDValue VAList = DAG.getLoad(getPointerTy(), DL, Chain, Addr,
4002 MachinePointerInfo(V), false, false, false, 0);
4003 Chain = VAList.getValue(1);
4006 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
4007 VAList = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
4008 DAG.getConstant(Align - 1, getPointerTy()));
4009 VAList = DAG.getNode(ISD::AND, DL, getPointerTy(), VAList,
4010 DAG.getConstant(-(int64_t)Align, getPointerTy()));
4013 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
4014 uint64_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
4016 // Scalar integer and FP values smaller than 64 bits are implicitly extended
4017 // up to 64 bits. At the very least, we have to increase the striding of the
4018 // vaargs list to match this, and for FP values we need to introduce
4019 // FP_ROUND nodes as well.
4020 if (VT.isInteger() && !VT.isVector())
4022 bool NeedFPTrunc = false;
4023 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
4028 // Increment the pointer, VAList, to the next vaarg
4029 SDValue VANext = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
4030 DAG.getConstant(ArgSize, getPointerTy()));
4031 // Store the incremented VAList to the legalized pointer
4032 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
4035 // Load the actual argument out of the pointer VAList
4037 // Load the value as an f64.
4038 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
4039 MachinePointerInfo(), false, false, false, 0);
4040 // Round the value down to an f32.
4041 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
4042 DAG.getIntPtrConstant(1));
4043 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
4044 // Merge the rounded value with the chain output of the load.
4045 return DAG.getMergeValues(Ops, DL);
4048 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
4052 SDValue AArch64TargetLowering::LowerFRAMEADDR(SDValue Op,
4053 SelectionDAG &DAG) const {
4054 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4055 MFI->setFrameAddressIsTaken(true);
4057 EVT VT = Op.getValueType();
4059 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4061 DAG.getCopyFromReg(DAG.getEntryNode(), DL, AArch64::FP, VT);
4063 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
4064 MachinePointerInfo(), false, false, false, 0);
4068 // FIXME? Maybe this could be a TableGen attribute on some registers and
4069 // this table could be generated automatically from RegInfo.
4070 unsigned AArch64TargetLowering::getRegisterByName(const char* RegName,
4072 unsigned Reg = StringSwitch<unsigned>(RegName)
4073 .Case("sp", AArch64::SP)
4077 report_fatal_error("Invalid register name global variable");
4080 SDValue AArch64TargetLowering::LowerRETURNADDR(SDValue Op,
4081 SelectionDAG &DAG) const {
4082 MachineFunction &MF = DAG.getMachineFunction();
4083 MachineFrameInfo *MFI = MF.getFrameInfo();
4084 MFI->setReturnAddressIsTaken(true);
4086 EVT VT = Op.getValueType();
4088 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4090 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4091 SDValue Offset = DAG.getConstant(8, getPointerTy());
4092 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
4093 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
4094 MachinePointerInfo(), false, false, false, 0);
4097 // Return LR, which contains the return address. Mark it an implicit live-in.
4098 unsigned Reg = MF.addLiveIn(AArch64::LR, &AArch64::GPR64RegClass);
4099 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
4102 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4103 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4104 SDValue AArch64TargetLowering::LowerShiftRightParts(SDValue Op,
4105 SelectionDAG &DAG) const {
4106 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4107 EVT VT = Op.getValueType();
4108 unsigned VTBits = VT.getSizeInBits();
4110 SDValue ShOpLo = Op.getOperand(0);
4111 SDValue ShOpHi = Op.getOperand(1);
4112 SDValue ShAmt = Op.getOperand(2);
4114 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4116 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4118 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4119 DAG.getConstant(VTBits, MVT::i64), ShAmt);
4120 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4121 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4122 DAG.getConstant(VTBits, MVT::i64));
4123 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4125 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
4126 ISD::SETGE, dl, DAG);
4127 SDValue CCVal = DAG.getConstant(AArch64CC::GE, MVT::i32);
4129 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4130 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4132 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4134 // AArch64 shifts larger than the register width are wrapped rather than
4135 // clamped, so we can't just emit "hi >> x".
4136 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4137 SDValue TrueValHi = Opc == ISD::SRA
4138 ? DAG.getNode(Opc, dl, VT, ShOpHi,
4139 DAG.getConstant(VTBits - 1, MVT::i64))
4140 : DAG.getConstant(0, VT);
4142 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
4144 SDValue Ops[2] = { Lo, Hi };
4145 return DAG.getMergeValues(Ops, dl);
4148 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4149 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4150 SDValue AArch64TargetLowering::LowerShiftLeftParts(SDValue Op,
4151 SelectionDAG &DAG) const {
4152 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4153 EVT VT = Op.getValueType();
4154 unsigned VTBits = VT.getSizeInBits();
4156 SDValue ShOpLo = Op.getOperand(0);
4157 SDValue ShOpHi = Op.getOperand(1);
4158 SDValue ShAmt = Op.getOperand(2);
4161 assert(Op.getOpcode() == ISD::SHL_PARTS);
4162 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4163 DAG.getConstant(VTBits, MVT::i64), ShAmt);
4164 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4165 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4166 DAG.getConstant(VTBits, MVT::i64));
4167 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4168 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4170 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4172 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
4173 ISD::SETGE, dl, DAG);
4174 SDValue CCVal = DAG.getConstant(AArch64CC::GE, MVT::i32);
4176 DAG.getNode(AArch64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
4178 // AArch64 shifts of larger than register sizes are wrapped rather than
4179 // clamped, so we can't just emit "lo << a" if a is too big.
4180 SDValue TrueValLo = DAG.getConstant(0, VT);
4181 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4183 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4185 SDValue Ops[2] = { Lo, Hi };
4186 return DAG.getMergeValues(Ops, dl);
4189 bool AArch64TargetLowering::isOffsetFoldingLegal(
4190 const GlobalAddressSDNode *GA) const {
4191 // The AArch64 target doesn't support folding offsets into global addresses.
4195 bool AArch64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4196 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
4197 // FIXME: We should be able to handle f128 as well with a clever lowering.
4198 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
4202 return AArch64_AM::getFP64Imm(Imm) != -1;
4203 else if (VT == MVT::f32)
4204 return AArch64_AM::getFP32Imm(Imm) != -1;
4208 //===----------------------------------------------------------------------===//
4209 // AArch64 Optimization Hooks
4210 //===----------------------------------------------------------------------===//
4212 //===----------------------------------------------------------------------===//
4213 // AArch64 Inline Assembly Support
4214 //===----------------------------------------------------------------------===//
4216 // Table of Constraints
4217 // TODO: This is the current set of constraints supported by ARM for the
4218 // compiler, not all of them may make sense, e.g. S may be difficult to support.
4220 // r - A general register
4221 // w - An FP/SIMD register of some size in the range v0-v31
4222 // x - An FP/SIMD register of some size in the range v0-v15
4223 // I - Constant that can be used with an ADD instruction
4224 // J - Constant that can be used with a SUB instruction
4225 // K - Constant that can be used with a 32-bit logical instruction
4226 // L - Constant that can be used with a 64-bit logical instruction
4227 // M - Constant that can be used as a 32-bit MOV immediate
4228 // N - Constant that can be used as a 64-bit MOV immediate
4229 // Q - A memory reference with base register and no offset
4230 // S - A symbolic address
4231 // Y - Floating point constant zero
4232 // Z - Integer constant zero
4234 // Note that general register operands will be output using their 64-bit x
4235 // register name, whatever the size of the variable, unless the asm operand
4236 // is prefixed by the %w modifier. Floating-point and SIMD register operands
4237 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
4240 /// getConstraintType - Given a constraint letter, return the type of
4241 /// constraint it is for this target.
4242 AArch64TargetLowering::ConstraintType
4243 AArch64TargetLowering::getConstraintType(const std::string &Constraint) const {
4244 if (Constraint.size() == 1) {
4245 switch (Constraint[0]) {
4252 return C_RegisterClass;
4253 // An address with a single base register. Due to the way we
4254 // currently handle addresses it is the same as 'r'.
4259 return TargetLowering::getConstraintType(Constraint);
4262 /// Examine constraint type and operand type and determine a weight value.
4263 /// This object must already have been set up with the operand type
4264 /// and the current alternative constraint selected.
4265 TargetLowering::ConstraintWeight
4266 AArch64TargetLowering::getSingleConstraintMatchWeight(
4267 AsmOperandInfo &info, const char *constraint) const {
4268 ConstraintWeight weight = CW_Invalid;
4269 Value *CallOperandVal = info.CallOperandVal;
4270 // If we don't have a value, we can't do a match,
4271 // but allow it at the lowest weight.
4272 if (!CallOperandVal)
4274 Type *type = CallOperandVal->getType();
4275 // Look at the constraint type.
4276 switch (*constraint) {
4278 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
4282 if (type->isFloatingPointTy() || type->isVectorTy())
4283 weight = CW_Register;
4286 weight = CW_Constant;
4292 std::pair<unsigned, const TargetRegisterClass *>
4293 AArch64TargetLowering::getRegForInlineAsmConstraint(
4294 const std::string &Constraint, MVT VT) const {
4295 if (Constraint.size() == 1) {
4296 switch (Constraint[0]) {
4298 if (VT.getSizeInBits() == 64)
4299 return std::make_pair(0U, &AArch64::GPR64commonRegClass);
4300 return std::make_pair(0U, &AArch64::GPR32commonRegClass);
4303 return std::make_pair(0U, &AArch64::FPR32RegClass);
4304 if (VT.getSizeInBits() == 64)
4305 return std::make_pair(0U, &AArch64::FPR64RegClass);
4306 if (VT.getSizeInBits() == 128)
4307 return std::make_pair(0U, &AArch64::FPR128RegClass);
4309 // The instructions that this constraint is designed for can
4310 // only take 128-bit registers so just use that regclass.
4312 if (VT.getSizeInBits() == 128)
4313 return std::make_pair(0U, &AArch64::FPR128_loRegClass);
4317 if (StringRef("{cc}").equals_lower(Constraint))
4318 return std::make_pair(unsigned(AArch64::NZCV), &AArch64::CCRRegClass);
4320 // Use the default implementation in TargetLowering to convert the register
4321 // constraint into a member of a register class.
4322 std::pair<unsigned, const TargetRegisterClass *> Res;
4323 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4325 // Not found as a standard register?
4327 unsigned Size = Constraint.size();
4328 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
4329 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
4330 const std::string Reg =
4331 std::string(&Constraint[2], &Constraint[Size - 1]);
4332 int RegNo = atoi(Reg.c_str());
4333 if (RegNo >= 0 && RegNo <= 31) {
4334 // v0 - v31 are aliases of q0 - q31.
4335 // By default we'll emit v0-v31 for this unless there's a modifier where
4336 // we'll emit the correct register as well.
4337 Res.first = AArch64::FPR128RegClass.getRegister(RegNo);
4338 Res.second = &AArch64::FPR128RegClass;
4346 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4347 /// vector. If it is invalid, don't add anything to Ops.
4348 void AArch64TargetLowering::LowerAsmOperandForConstraint(
4349 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
4350 SelectionDAG &DAG) const {
4353 // Currently only support length 1 constraints.
4354 if (Constraint.length() != 1)
4357 char ConstraintLetter = Constraint[0];
4358 switch (ConstraintLetter) {
4362 // This set of constraints deal with valid constants for various instructions.
4363 // Validate and return a target constant for them if we can.
4365 // 'z' maps to xzr or wzr so it needs an input of 0.
4366 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4367 if (!C || C->getZExtValue() != 0)
4370 if (Op.getValueType() == MVT::i64)
4371 Result = DAG.getRegister(AArch64::XZR, MVT::i64);
4373 Result = DAG.getRegister(AArch64::WZR, MVT::i32);
4383 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4387 // Grab the value and do some validation.
4388 uint64_t CVal = C->getZExtValue();
4389 switch (ConstraintLetter) {
4390 // The I constraint applies only to simple ADD or SUB immediate operands:
4391 // i.e. 0 to 4095 with optional shift by 12
4392 // The J constraint applies only to ADD or SUB immediates that would be
4393 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
4394 // instruction [or vice versa], in other words -1 to -4095 with optional
4395 // left shift by 12.
4397 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
4401 uint64_t NVal = -C->getSExtValue();
4402 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal)) {
4403 CVal = C->getSExtValue();
4408 // The K and L constraints apply *only* to logical immediates, including
4409 // what used to be the MOVI alias for ORR (though the MOVI alias has now
4410 // been removed and MOV should be used). So these constraints have to
4411 // distinguish between bit patterns that are valid 32-bit or 64-bit
4412 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
4413 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
4416 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4420 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4423 // The M and N constraints are a superset of K and L respectively, for use
4424 // with the MOV (immediate) alias. As well as the logical immediates they
4425 // also match 32 or 64-bit immediates that can be loaded either using a
4426 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
4427 // (M) or 64-bit 0x1234000000000000 (N) etc.
4428 // As a note some of this code is liberally stolen from the asm parser.
4430 if (!isUInt<32>(CVal))
4432 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4434 if ((CVal & 0xFFFF) == CVal)
4436 if ((CVal & 0xFFFF0000ULL) == CVal)
4438 uint64_t NCVal = ~(uint32_t)CVal;
4439 if ((NCVal & 0xFFFFULL) == NCVal)
4441 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4446 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4448 if ((CVal & 0xFFFFULL) == CVal)
4450 if ((CVal & 0xFFFF0000ULL) == CVal)
4452 if ((CVal & 0xFFFF00000000ULL) == CVal)
4454 if ((CVal & 0xFFFF000000000000ULL) == CVal)
4456 uint64_t NCVal = ~CVal;
4457 if ((NCVal & 0xFFFFULL) == NCVal)
4459 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4461 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
4463 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
4471 // All assembler immediates are 64-bit integers.
4472 Result = DAG.getTargetConstant(CVal, MVT::i64);
4476 if (Result.getNode()) {
4477 Ops.push_back(Result);
4481 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4484 //===----------------------------------------------------------------------===//
4485 // AArch64 Advanced SIMD Support
4486 //===----------------------------------------------------------------------===//
4488 /// WidenVector - Given a value in the V64 register class, produce the
4489 /// equivalent value in the V128 register class.
4490 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
4491 EVT VT = V64Reg.getValueType();
4492 unsigned NarrowSize = VT.getVectorNumElements();
4493 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4494 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
4497 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
4498 V64Reg, DAG.getConstant(0, MVT::i32));
4501 /// getExtFactor - Determine the adjustment factor for the position when
4502 /// generating an "extract from vector registers" instruction.
4503 static unsigned getExtFactor(SDValue &V) {
4504 EVT EltType = V.getValueType().getVectorElementType();
4505 return EltType.getSizeInBits() / 8;
4508 /// NarrowVector - Given a value in the V128 register class, produce the
4509 /// equivalent value in the V64 register class.
4510 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
4511 EVT VT = V128Reg.getValueType();
4512 unsigned WideSize = VT.getVectorNumElements();
4513 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4514 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
4517 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, NarrowTy, V128Reg);
4520 // Gather data to see if the operation can be modelled as a
4521 // shuffle in combination with VEXTs.
4522 SDValue AArch64TargetLowering::ReconstructShuffle(SDValue Op,
4523 SelectionDAG &DAG) const {
4524 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
4526 EVT VT = Op.getValueType();
4527 unsigned NumElts = VT.getVectorNumElements();
4529 struct ShuffleSourceInfo {
4534 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
4535 // be compatible with the shuffle we intend to construct. As a result
4536 // ShuffleVec will be some sliding window into the original Vec.
4539 // Code should guarantee that element i in Vec starts at element "WindowBase
4540 // + i * WindowScale in ShuffleVec".
4544 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
4545 ShuffleSourceInfo(SDValue Vec)
4546 : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
4550 // First gather all vectors used as an immediate source for this BUILD_VECTOR
4552 SmallVector<ShuffleSourceInfo, 2> Sources;
4553 for (unsigned i = 0; i < NumElts; ++i) {
4554 SDValue V = Op.getOperand(i);
4555 if (V.getOpcode() == ISD::UNDEF)
4557 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4558 // A shuffle can only come from building a vector from various
4559 // elements of other vectors.
4563 // Add this element source to the list if it's not already there.
4564 SDValue SourceVec = V.getOperand(0);
4565 auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
4566 if (Source == Sources.end())
4567 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
4569 // Update the minimum and maximum lane number seen.
4570 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4571 Source->MinElt = std::min(Source->MinElt, EltNo);
4572 Source->MaxElt = std::max(Source->MaxElt, EltNo);
4575 // Currently only do something sane when at most two source vectors
4577 if (Sources.size() > 2)
4580 // Find out the smallest element size among result and two sources, and use
4581 // it as element size to build the shuffle_vector.
4582 EVT SmallestEltTy = VT.getVectorElementType();
4583 for (auto &Source : Sources) {
4584 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
4585 if (SrcEltTy.bitsLT(SmallestEltTy)) {
4586 SmallestEltTy = SrcEltTy;
4589 unsigned ResMultiplier =
4590 VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
4591 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
4592 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
4594 // If the source vector is too wide or too narrow, we may nevertheless be able
4595 // to construct a compatible shuffle either by concatenating it with UNDEF or
4596 // extracting a suitable range of elements.
4597 for (auto &Src : Sources) {
4598 EVT SrcVT = Src.ShuffleVec.getValueType();
4600 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
4603 // This stage of the search produces a source with the same element type as
4604 // the original, but with a total width matching the BUILD_VECTOR output.
4605 EVT EltVT = SrcVT.getVectorElementType();
4606 unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
4607 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
4609 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
4610 assert(2 * SrcVT.getSizeInBits() == VT.getSizeInBits());
4611 // We can pad out the smaller vector for free, so if it's part of a
4614 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
4615 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
4619 assert(SrcVT.getSizeInBits() == 2 * VT.getSizeInBits());
4621 if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
4622 // Span too large for a VEXT to cope
4626 if (Src.MinElt >= NumSrcElts) {
4627 // The extraction can just take the second half
4629 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4630 DAG.getConstant(NumSrcElts, MVT::i64));
4631 Src.WindowBase = -NumSrcElts;
4632 } else if (Src.MaxElt < NumSrcElts) {
4633 // The extraction can just take the first half
4635 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4636 DAG.getConstant(0, MVT::i64));
4638 // An actual VEXT is needed
4640 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4641 DAG.getConstant(0, MVT::i64));
4643 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4644 DAG.getConstant(NumSrcElts, MVT::i64));
4645 unsigned Imm = Src.MinElt * getExtFactor(VEXTSrc1);
4647 Src.ShuffleVec = DAG.getNode(AArch64ISD::EXT, dl, DestVT, VEXTSrc1,
4648 VEXTSrc2, DAG.getConstant(Imm, MVT::i32));
4649 Src.WindowBase = -Src.MinElt;
4653 // Another possible incompatibility occurs from the vector element types. We
4654 // can fix this by bitcasting the source vectors to the same type we intend
4656 for (auto &Src : Sources) {
4657 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
4658 if (SrcEltTy == SmallestEltTy)
4660 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
4661 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
4662 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
4663 Src.WindowBase *= Src.WindowScale;
4666 // Final sanity check before we try to actually produce a shuffle.
4668 for (auto Src : Sources)
4669 assert(Src.ShuffleVec.getValueType() == ShuffleVT);
4672 // The stars all align, our next step is to produce the mask for the shuffle.
4673 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
4674 int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
4675 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
4676 SDValue Entry = Op.getOperand(i);
4677 if (Entry.getOpcode() == ISD::UNDEF)
4680 auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
4681 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
4683 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
4684 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
4686 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
4687 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
4688 VT.getVectorElementType().getSizeInBits());
4689 int LanesDefined = BitsDefined / BitsPerShuffleLane;
4691 // This source is expected to fill ResMultiplier lanes of the final shuffle,
4692 // starting at the appropriate offset.
4693 int *LaneMask = &Mask[i * ResMultiplier];
4695 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
4696 ExtractBase += NumElts * (Src - Sources.begin());
4697 for (int j = 0; j < LanesDefined; ++j)
4698 LaneMask[j] = ExtractBase + j;
4701 // Final check before we try to produce nonsense...
4702 if (!isShuffleMaskLegal(Mask, ShuffleVT))
4705 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
4706 for (unsigned i = 0; i < Sources.size(); ++i)
4707 ShuffleOps[i] = Sources[i].ShuffleVec;
4709 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
4710 ShuffleOps[1], &Mask[0]);
4711 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
4714 // check if an EXT instruction can handle the shuffle mask when the
4715 // vector sources of the shuffle are the same.
4716 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4717 unsigned NumElts = VT.getVectorNumElements();
4719 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4725 // If this is a VEXT shuffle, the immediate value is the index of the first
4726 // element. The other shuffle indices must be the successive elements after
4728 unsigned ExpectedElt = Imm;
4729 for (unsigned i = 1; i < NumElts; ++i) {
4730 // Increment the expected index. If it wraps around, just follow it
4731 // back to index zero and keep going.
4733 if (ExpectedElt == NumElts)
4737 continue; // ignore UNDEF indices
4738 if (ExpectedElt != static_cast<unsigned>(M[i]))
4745 // check if an EXT instruction can handle the shuffle mask when the
4746 // vector sources of the shuffle are different.
4747 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
4749 // Look for the first non-undef element.
4750 const int *FirstRealElt = std::find_if(M.begin(), M.end(),
4751 [](int Elt) {return Elt >= 0;});
4753 // Benefit form APInt to handle overflow when calculating expected element.
4754 unsigned NumElts = VT.getVectorNumElements();
4755 unsigned MaskBits = APInt(32, NumElts * 2).logBase2();
4756 APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1);
4757 // The following shuffle indices must be the successive elements after the
4758 // first real element.
4759 const int *FirstWrongElt = std::find_if(FirstRealElt + 1, M.end(),
4760 [&](int Elt) {return Elt != ExpectedElt++ && Elt != -1;});
4761 if (FirstWrongElt != M.end())
4764 // The index of an EXT is the first element if it is not UNDEF.
4765 // Watch out for the beginning UNDEFs. The EXT index should be the expected
4766 // value of the first element. E.g.
4767 // <-1, -1, 3, ...> is treated as <1, 2, 3, ...>.
4768 // <-1, -1, 0, 1, ...> is treated as <2*NumElts-2, 2*NumElts-1, 0, 1, ...>.
4769 // ExpectedElt is the last mask index plus 1.
4770 Imm = ExpectedElt.getZExtValue();
4772 // There are two difference cases requiring to reverse input vectors.
4773 // For example, for vector <4 x i32> we have the following cases,
4774 // Case 1: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, -1, 0>)
4775 // Case 2: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, 7, 0>)
4776 // For both cases, we finally use mask <5, 6, 7, 0>, which requires
4777 // to reverse two input vectors.
4786 /// isREVMask - Check if a vector shuffle corresponds to a REV
4787 /// instruction with the specified blocksize. (The order of the elements
4788 /// within each block of the vector is reversed.)
4789 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4790 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
4791 "Only possible block sizes for REV are: 16, 32, 64");
4793 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4797 unsigned NumElts = VT.getVectorNumElements();
4798 unsigned BlockElts = M[0] + 1;
4799 // If the first shuffle index is UNDEF, be optimistic.
4801 BlockElts = BlockSize / EltSz;
4803 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4806 for (unsigned i = 0; i < NumElts; ++i) {
4808 continue; // ignore UNDEF indices
4809 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
4816 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4817 unsigned NumElts = VT.getVectorNumElements();
4818 WhichResult = (M[0] == 0 ? 0 : 1);
4819 unsigned Idx = WhichResult * NumElts / 2;
4820 for (unsigned i = 0; i != NumElts; i += 2) {
4821 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4822 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
4830 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4831 unsigned NumElts = VT.getVectorNumElements();
4832 WhichResult = (M[0] == 0 ? 0 : 1);
4833 for (unsigned i = 0; i != NumElts; ++i) {
4835 continue; // ignore UNDEF indices
4836 if ((unsigned)M[i] != 2 * i + WhichResult)
4843 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4844 unsigned NumElts = VT.getVectorNumElements();
4845 WhichResult = (M[0] == 0 ? 0 : 1);
4846 for (unsigned i = 0; i < NumElts; i += 2) {
4847 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4848 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
4854 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
4855 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4856 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4857 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4858 unsigned NumElts = VT.getVectorNumElements();
4859 WhichResult = (M[0] == 0 ? 0 : 1);
4860 unsigned Idx = WhichResult * NumElts / 2;
4861 for (unsigned i = 0; i != NumElts; i += 2) {
4862 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4863 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
4871 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
4872 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4873 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4874 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4875 unsigned Half = VT.getVectorNumElements() / 2;
4876 WhichResult = (M[0] == 0 ? 0 : 1);
4877 for (unsigned j = 0; j != 2; ++j) {
4878 unsigned Idx = WhichResult;
4879 for (unsigned i = 0; i != Half; ++i) {
4880 int MIdx = M[i + j * Half];
4881 if (MIdx >= 0 && (unsigned)MIdx != Idx)
4890 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
4891 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4892 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4893 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4894 unsigned NumElts = VT.getVectorNumElements();
4895 WhichResult = (M[0] == 0 ? 0 : 1);
4896 for (unsigned i = 0; i < NumElts; i += 2) {
4897 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4898 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
4904 static bool isINSMask(ArrayRef<int> M, int NumInputElements,
4905 bool &DstIsLeft, int &Anomaly) {
4906 if (M.size() != static_cast<size_t>(NumInputElements))
4909 int NumLHSMatch = 0, NumRHSMatch = 0;
4910 int LastLHSMismatch = -1, LastRHSMismatch = -1;
4912 for (int i = 0; i < NumInputElements; ++i) {
4922 LastLHSMismatch = i;
4924 if (M[i] == i + NumInputElements)
4927 LastRHSMismatch = i;
4930 if (NumLHSMatch == NumInputElements - 1) {
4932 Anomaly = LastLHSMismatch;
4934 } else if (NumRHSMatch == NumInputElements - 1) {
4936 Anomaly = LastRHSMismatch;
4943 static bool isConcatMask(ArrayRef<int> Mask, EVT VT, bool SplitLHS) {
4944 if (VT.getSizeInBits() != 128)
4947 unsigned NumElts = VT.getVectorNumElements();
4949 for (int I = 0, E = NumElts / 2; I != E; I++) {
4954 int Offset = NumElts / 2;
4955 for (int I = NumElts / 2, E = NumElts; I != E; I++) {
4956 if (Mask[I] != I + SplitLHS * Offset)
4963 static SDValue tryFormConcatFromShuffle(SDValue Op, SelectionDAG &DAG) {
4965 EVT VT = Op.getValueType();
4966 SDValue V0 = Op.getOperand(0);
4967 SDValue V1 = Op.getOperand(1);
4968 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();
4970 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
4971 VT.getVectorElementType() != V1.getValueType().getVectorElementType())
4974 bool SplitV0 = V0.getValueType().getSizeInBits() == 128;
4976 if (!isConcatMask(Mask, VT, SplitV0))
4979 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
4980 VT.getVectorNumElements() / 2);
4982 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
4983 DAG.getConstant(0, MVT::i64));
4985 if (V1.getValueType().getSizeInBits() == 128) {
4986 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
4987 DAG.getConstant(0, MVT::i64));
4989 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
4992 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4993 /// the specified operations to build the shuffle.
4994 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4995 SDValue RHS, SelectionDAG &DAG,
4997 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4998 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
4999 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
5002 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5011 OP_VUZPL, // VUZP, left result
5012 OP_VUZPR, // VUZP, right result
5013 OP_VZIPL, // VZIP, left result
5014 OP_VZIPR, // VZIP, right result
5015 OP_VTRNL, // VTRN, left result
5016 OP_VTRNR // VTRN, right result
5019 if (OpNum == OP_COPY) {
5020 if (LHSID == (1 * 9 + 2) * 9 + 3)
5022 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
5026 SDValue OpLHS, OpRHS;
5027 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5028 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5029 EVT VT = OpLHS.getValueType();
5033 llvm_unreachable("Unknown shuffle opcode!");
5035 // VREV divides the vector in half and swaps within the half.
5036 if (VT.getVectorElementType() == MVT::i32 ||
5037 VT.getVectorElementType() == MVT::f32)
5038 return DAG.getNode(AArch64ISD::REV64, dl, VT, OpLHS);
5039 // vrev <4 x i16> -> REV32
5040 if (VT.getVectorElementType() == MVT::i16 ||
5041 VT.getVectorElementType() == MVT::f16)
5042 return DAG.getNode(AArch64ISD::REV32, dl, VT, OpLHS);
5043 // vrev <4 x i8> -> REV16
5044 assert(VT.getVectorElementType() == MVT::i8);
5045 return DAG.getNode(AArch64ISD::REV16, dl, VT, OpLHS);
5050 EVT EltTy = VT.getVectorElementType();
5052 if (EltTy == MVT::i8)
5053 Opcode = AArch64ISD::DUPLANE8;
5054 else if (EltTy == MVT::i16)
5055 Opcode = AArch64ISD::DUPLANE16;
5056 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
5057 Opcode = AArch64ISD::DUPLANE32;
5058 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
5059 Opcode = AArch64ISD::DUPLANE64;
5061 llvm_unreachable("Invalid vector element type?");
5063 if (VT.getSizeInBits() == 64)
5064 OpLHS = WidenVector(OpLHS, DAG);
5065 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, MVT::i64);
5066 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
5071 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
5072 return DAG.getNode(AArch64ISD::EXT, dl, VT, OpLHS, OpRHS,
5073 DAG.getConstant(Imm, MVT::i32));
5076 return DAG.getNode(AArch64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS,
5079 return DAG.getNode(AArch64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS,
5082 return DAG.getNode(AArch64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS,
5085 return DAG.getNode(AArch64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS,
5088 return DAG.getNode(AArch64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS,
5091 return DAG.getNode(AArch64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS,
5096 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
5097 SelectionDAG &DAG) {
5098 // Check to see if we can use the TBL instruction.
5099 SDValue V1 = Op.getOperand(0);
5100 SDValue V2 = Op.getOperand(1);
5103 EVT EltVT = Op.getValueType().getVectorElementType();
5104 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
5106 SmallVector<SDValue, 8> TBLMask;
5107 for (int Val : ShuffleMask) {
5108 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
5109 unsigned Offset = Byte + Val * BytesPerElt;
5110 TBLMask.push_back(DAG.getConstant(Offset, MVT::i32));
5114 MVT IndexVT = MVT::v8i8;
5115 unsigned IndexLen = 8;
5116 if (Op.getValueType().getSizeInBits() == 128) {
5117 IndexVT = MVT::v16i8;
5121 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
5122 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
5125 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
5127 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
5128 Shuffle = DAG.getNode(
5129 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5130 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, MVT::i32), V1Cst,
5131 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5132 makeArrayRef(TBLMask.data(), IndexLen)));
5134 if (IndexLen == 8) {
5135 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
5136 Shuffle = DAG.getNode(
5137 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5138 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, MVT::i32), V1Cst,
5139 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5140 makeArrayRef(TBLMask.data(), IndexLen)));
5142 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
5143 // cannot currently represent the register constraints on the input
5145 // Shuffle = DAG.getNode(AArch64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
5146 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5147 // &TBLMask[0], IndexLen));
5148 Shuffle = DAG.getNode(
5149 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5150 DAG.getConstant(Intrinsic::aarch64_neon_tbl2, MVT::i32), V1Cst, V2Cst,
5151 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5152 makeArrayRef(TBLMask.data(), IndexLen)));
5155 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
5158 static unsigned getDUPLANEOp(EVT EltType) {
5159 if (EltType == MVT::i8)
5160 return AArch64ISD::DUPLANE8;
5161 if (EltType == MVT::i16 || EltType == MVT::f16)
5162 return AArch64ISD::DUPLANE16;
5163 if (EltType == MVT::i32 || EltType == MVT::f32)
5164 return AArch64ISD::DUPLANE32;
5165 if (EltType == MVT::i64 || EltType == MVT::f64)
5166 return AArch64ISD::DUPLANE64;
5168 llvm_unreachable("Invalid vector element type?");
5171 SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5172 SelectionDAG &DAG) const {
5174 EVT VT = Op.getValueType();
5176 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5178 // Convert shuffles that are directly supported on NEON to target-specific
5179 // DAG nodes, instead of keeping them as shuffles and matching them again
5180 // during code selection. This is more efficient and avoids the possibility
5181 // of inconsistencies between legalization and selection.
5182 ArrayRef<int> ShuffleMask = SVN->getMask();
5184 SDValue V1 = Op.getOperand(0);
5185 SDValue V2 = Op.getOperand(1);
5187 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
5188 V1.getValueType().getSimpleVT())) {
5189 int Lane = SVN->getSplatIndex();
5190 // If this is undef splat, generate it via "just" vdup, if possible.
5194 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
5195 return DAG.getNode(AArch64ISD::DUP, dl, V1.getValueType(),
5197 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
5198 // constant. If so, we can just reference the lane's definition directly.
5199 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
5200 !isa<ConstantSDNode>(V1.getOperand(Lane)))
5201 return DAG.getNode(AArch64ISD::DUP, dl, VT, V1.getOperand(Lane));
5203 // Otherwise, duplicate from the lane of the input vector.
5204 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
5206 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
5207 // to make a vector of the same size as this SHUFFLE. We can ignore the
5208 // extract entirely, and canonicalise the concat using WidenVector.
5209 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
5210 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
5211 V1 = V1.getOperand(0);
5212 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
5213 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
5214 Lane -= Idx * VT.getVectorNumElements() / 2;
5215 V1 = WidenVector(V1.getOperand(Idx), DAG);
5216 } else if (VT.getSizeInBits() == 64)
5217 V1 = WidenVector(V1, DAG);
5219 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, MVT::i64));
5222 if (isREVMask(ShuffleMask, VT, 64))
5223 return DAG.getNode(AArch64ISD::REV64, dl, V1.getValueType(), V1, V2);
5224 if (isREVMask(ShuffleMask, VT, 32))
5225 return DAG.getNode(AArch64ISD::REV32, dl, V1.getValueType(), V1, V2);
5226 if (isREVMask(ShuffleMask, VT, 16))
5227 return DAG.getNode(AArch64ISD::REV16, dl, V1.getValueType(), V1, V2);
5229 bool ReverseEXT = false;
5231 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
5234 Imm *= getExtFactor(V1);
5235 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V2,
5236 DAG.getConstant(Imm, MVT::i32));
5237 } else if (V2->getOpcode() == ISD::UNDEF &&
5238 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
5239 Imm *= getExtFactor(V1);
5240 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V1,
5241 DAG.getConstant(Imm, MVT::i32));
5244 unsigned WhichResult;
5245 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
5246 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5247 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5249 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
5250 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5251 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5253 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
5254 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5255 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5258 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5259 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5260 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5262 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5263 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5264 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5266 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5267 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5268 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5271 SDValue Concat = tryFormConcatFromShuffle(Op, DAG);
5272 if (Concat.getNode())
5277 int NumInputElements = V1.getValueType().getVectorNumElements();
5278 if (isINSMask(ShuffleMask, NumInputElements, DstIsLeft, Anomaly)) {
5279 SDValue DstVec = DstIsLeft ? V1 : V2;
5280 SDValue DstLaneV = DAG.getConstant(Anomaly, MVT::i64);
5282 SDValue SrcVec = V1;
5283 int SrcLane = ShuffleMask[Anomaly];
5284 if (SrcLane >= NumInputElements) {
5286 SrcLane -= VT.getVectorNumElements();
5288 SDValue SrcLaneV = DAG.getConstant(SrcLane, MVT::i64);
5290 EVT ScalarVT = VT.getVectorElementType();
5292 if (ScalarVT.getSizeInBits() < 32 && ScalarVT.isInteger())
5293 ScalarVT = MVT::i32;
5296 ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5297 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ScalarVT, SrcVec, SrcLaneV),
5301 // If the shuffle is not directly supported and it has 4 elements, use
5302 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5303 unsigned NumElts = VT.getVectorNumElements();
5305 unsigned PFIndexes[4];
5306 for (unsigned i = 0; i != 4; ++i) {
5307 if (ShuffleMask[i] < 0)
5310 PFIndexes[i] = ShuffleMask[i];
5313 // Compute the index in the perfect shuffle table.
5314 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5315 PFIndexes[2] * 9 + PFIndexes[3];
5316 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5317 unsigned Cost = (PFEntry >> 30);
5320 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5323 return GenerateTBL(Op, ShuffleMask, DAG);
5326 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
5328 EVT VT = BVN->getValueType(0);
5329 APInt SplatBits, SplatUndef;
5330 unsigned SplatBitSize;
5332 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5333 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
5335 for (unsigned i = 0; i < NumSplats; ++i) {
5336 CnstBits <<= SplatBitSize;
5337 UndefBits <<= SplatBitSize;
5338 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
5339 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
5348 SDValue AArch64TargetLowering::LowerVectorAND(SDValue Op,
5349 SelectionDAG &DAG) const {
5350 BuildVectorSDNode *BVN =
5351 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5352 SDValue LHS = Op.getOperand(0);
5354 EVT VT = Op.getValueType();
5359 APInt CnstBits(VT.getSizeInBits(), 0);
5360 APInt UndefBits(VT.getSizeInBits(), 0);
5361 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5362 // We only have BIC vector immediate instruction, which is and-not.
5363 CnstBits = ~CnstBits;
5365 // We make use of a little bit of goto ickiness in order to avoid having to
5366 // duplicate the immediate matching logic for the undef toggled case.
5367 bool SecondTry = false;
5370 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5371 CnstBits = CnstBits.zextOrTrunc(64);
5372 uint64_t CnstVal = CnstBits.getZExtValue();
5374 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5375 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5376 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5377 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5378 DAG.getConstant(CnstVal, MVT::i32),
5379 DAG.getConstant(0, MVT::i32));
5380 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5383 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5384 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5385 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5386 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5387 DAG.getConstant(CnstVal, MVT::i32),
5388 DAG.getConstant(8, MVT::i32));
5389 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5392 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5393 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5394 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5395 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5396 DAG.getConstant(CnstVal, MVT::i32),
5397 DAG.getConstant(16, MVT::i32));
5398 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5401 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5402 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5403 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5404 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5405 DAG.getConstant(CnstVal, MVT::i32),
5406 DAG.getConstant(24, MVT::i32));
5407 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5410 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5411 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5412 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5413 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5414 DAG.getConstant(CnstVal, MVT::i32),
5415 DAG.getConstant(0, MVT::i32));
5416 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5419 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5420 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5421 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5422 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5423 DAG.getConstant(CnstVal, MVT::i32),
5424 DAG.getConstant(8, MVT::i32));
5425 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5432 CnstBits = ~UndefBits;
5436 // We can always fall back to a non-immediate AND.
5441 // Specialized code to quickly find if PotentialBVec is a BuildVector that
5442 // consists of only the same constant int value, returned in reference arg
5444 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
5445 uint64_t &ConstVal) {
5446 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
5449 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
5452 EVT VT = Bvec->getValueType(0);
5453 unsigned NumElts = VT.getVectorNumElements();
5454 for (unsigned i = 1; i < NumElts; ++i)
5455 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
5457 ConstVal = FirstElt->getZExtValue();
5461 static unsigned getIntrinsicID(const SDNode *N) {
5462 unsigned Opcode = N->getOpcode();
5465 return Intrinsic::not_intrinsic;
5466 case ISD::INTRINSIC_WO_CHAIN: {
5467 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5468 if (IID < Intrinsic::num_intrinsics)
5470 return Intrinsic::not_intrinsic;
5475 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
5476 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
5477 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
5478 // Also, logical shift right -> sri, with the same structure.
5479 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
5480 EVT VT = N->getValueType(0);
5487 // Is the first op an AND?
5488 const SDValue And = N->getOperand(0);
5489 if (And.getOpcode() != ISD::AND)
5492 // Is the second op an shl or lshr?
5493 SDValue Shift = N->getOperand(1);
5494 // This will have been turned into: AArch64ISD::VSHL vector, #shift
5495 // or AArch64ISD::VLSHR vector, #shift
5496 unsigned ShiftOpc = Shift.getOpcode();
5497 if ((ShiftOpc != AArch64ISD::VSHL && ShiftOpc != AArch64ISD::VLSHR))
5499 bool IsShiftRight = ShiftOpc == AArch64ISD::VLSHR;
5501 // Is the shift amount constant?
5502 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
5506 // Is the and mask vector all constant?
5508 if (!isAllConstantBuildVector(And.getOperand(1), C1))
5511 // Is C1 == ~C2, taking into account how much one can shift elements of a
5513 uint64_t C2 = C2node->getZExtValue();
5514 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
5515 if (C2 > ElemSizeInBits)
5517 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
5518 if ((C1 & ElemMask) != (~C2 & ElemMask))
5521 SDValue X = And.getOperand(0);
5522 SDValue Y = Shift.getOperand(0);
5525 IsShiftRight ? Intrinsic::aarch64_neon_vsri : Intrinsic::aarch64_neon_vsli;
5527 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5528 DAG.getConstant(Intrin, MVT::i32), X, Y, Shift.getOperand(1));
5530 DEBUG(dbgs() << "aarch64-lower: transformed: \n");
5531 DEBUG(N->dump(&DAG));
5532 DEBUG(dbgs() << "into: \n");
5533 DEBUG(ResultSLI->dump(&DAG));
5539 SDValue AArch64TargetLowering::LowerVectorOR(SDValue Op,
5540 SelectionDAG &DAG) const {
5541 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
5542 if (EnableAArch64SlrGeneration) {
5543 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
5548 BuildVectorSDNode *BVN =
5549 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
5550 SDValue LHS = Op.getOperand(1);
5552 EVT VT = Op.getValueType();
5554 // OR commutes, so try swapping the operands.
5556 LHS = Op.getOperand(0);
5557 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5562 APInt CnstBits(VT.getSizeInBits(), 0);
5563 APInt UndefBits(VT.getSizeInBits(), 0);
5564 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5565 // We make use of a little bit of goto ickiness in order to avoid having to
5566 // duplicate the immediate matching logic for the undef toggled case.
5567 bool SecondTry = false;
5570 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5571 CnstBits = CnstBits.zextOrTrunc(64);
5572 uint64_t CnstVal = CnstBits.getZExtValue();
5574 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5575 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5576 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5577 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5578 DAG.getConstant(CnstVal, MVT::i32),
5579 DAG.getConstant(0, MVT::i32));
5580 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5583 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5584 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5585 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5586 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5587 DAG.getConstant(CnstVal, MVT::i32),
5588 DAG.getConstant(8, MVT::i32));
5589 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5592 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5593 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5594 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5595 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5596 DAG.getConstant(CnstVal, MVT::i32),
5597 DAG.getConstant(16, MVT::i32));
5598 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5601 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5602 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5603 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5604 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5605 DAG.getConstant(CnstVal, MVT::i32),
5606 DAG.getConstant(24, MVT::i32));
5607 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5610 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5611 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5612 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5613 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5614 DAG.getConstant(CnstVal, MVT::i32),
5615 DAG.getConstant(0, MVT::i32));
5616 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5619 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5620 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5621 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5622 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5623 DAG.getConstant(CnstVal, MVT::i32),
5624 DAG.getConstant(8, MVT::i32));
5625 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5632 CnstBits = UndefBits;
5636 // We can always fall back to a non-immediate OR.
5641 // Normalize the operands of BUILD_VECTOR. The value of constant operands will
5642 // be truncated to fit element width.
5643 static SDValue NormalizeBuildVector(SDValue Op,
5644 SelectionDAG &DAG) {
5645 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
5647 EVT VT = Op.getValueType();
5648 EVT EltTy= VT.getVectorElementType();
5650 if (EltTy.isFloatingPoint() || EltTy.getSizeInBits() > 16)
5653 SmallVector<SDValue, 16> Ops;
5654 for (unsigned I = 0, E = VT.getVectorNumElements(); I != E; ++I) {
5655 SDValue Lane = Op.getOperand(I);
5656 if (Lane.getOpcode() == ISD::Constant) {
5657 APInt LowBits(EltTy.getSizeInBits(),
5658 cast<ConstantSDNode>(Lane)->getZExtValue());
5659 Lane = DAG.getConstant(LowBits.getZExtValue(), MVT::i32);
5661 Ops.push_back(Lane);
5663 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5666 SDValue AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
5667 SelectionDAG &DAG) const {
5669 EVT VT = Op.getValueType();
5670 Op = NormalizeBuildVector(Op, DAG);
5671 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5673 APInt CnstBits(VT.getSizeInBits(), 0);
5674 APInt UndefBits(VT.getSizeInBits(), 0);
5675 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5676 // We make use of a little bit of goto ickiness in order to avoid having to
5677 // duplicate the immediate matching logic for the undef toggled case.
5678 bool SecondTry = false;
5681 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5682 CnstBits = CnstBits.zextOrTrunc(64);
5683 uint64_t CnstVal = CnstBits.getZExtValue();
5685 // Certain magic vector constants (used to express things like NOT
5686 // and NEG) are passed through unmodified. This allows codegen patterns
5687 // for these operations to match. Special-purpose patterns will lower
5688 // these immediates to MOVIs if it proves necessary.
5689 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
5692 // The many faces of MOVI...
5693 if (AArch64_AM::isAdvSIMDModImmType10(CnstVal)) {
5694 CnstVal = AArch64_AM::encodeAdvSIMDModImmType10(CnstVal);
5695 if (VT.getSizeInBits() == 128) {
5696 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::v2i64,
5697 DAG.getConstant(CnstVal, MVT::i32));
5698 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5701 // Support the V64 version via subregister insertion.
5702 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::f64,
5703 DAG.getConstant(CnstVal, MVT::i32));
5704 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5707 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5708 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5709 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5710 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5711 DAG.getConstant(CnstVal, MVT::i32),
5712 DAG.getConstant(0, MVT::i32));
5713 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5716 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5717 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5718 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5719 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5720 DAG.getConstant(CnstVal, MVT::i32),
5721 DAG.getConstant(8, MVT::i32));
5722 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5725 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5726 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5727 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5728 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5729 DAG.getConstant(CnstVal, MVT::i32),
5730 DAG.getConstant(16, MVT::i32));
5731 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5734 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5735 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5736 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5737 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5738 DAG.getConstant(CnstVal, MVT::i32),
5739 DAG.getConstant(24, MVT::i32));
5740 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5743 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5744 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5745 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5746 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5747 DAG.getConstant(CnstVal, MVT::i32),
5748 DAG.getConstant(0, MVT::i32));
5749 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5752 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5753 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5754 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5755 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5756 DAG.getConstant(CnstVal, MVT::i32),
5757 DAG.getConstant(8, MVT::i32));
5758 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5761 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
5762 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
5763 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5764 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
5765 DAG.getConstant(CnstVal, MVT::i32),
5766 DAG.getConstant(264, MVT::i32));
5767 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5770 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
5771 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
5772 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5773 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
5774 DAG.getConstant(CnstVal, MVT::i32),
5775 DAG.getConstant(272, MVT::i32));
5776 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5779 if (AArch64_AM::isAdvSIMDModImmType9(CnstVal)) {
5780 CnstVal = AArch64_AM::encodeAdvSIMDModImmType9(CnstVal);
5781 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
5782 SDValue Mov = DAG.getNode(AArch64ISD::MOVI, dl, MovTy,
5783 DAG.getConstant(CnstVal, MVT::i32));
5784 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5787 // The few faces of FMOV...
5788 if (AArch64_AM::isAdvSIMDModImmType11(CnstVal)) {
5789 CnstVal = AArch64_AM::encodeAdvSIMDModImmType11(CnstVal);
5790 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
5791 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MovTy,
5792 DAG.getConstant(CnstVal, MVT::i32));
5793 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5796 if (AArch64_AM::isAdvSIMDModImmType12(CnstVal) &&
5797 VT.getSizeInBits() == 128) {
5798 CnstVal = AArch64_AM::encodeAdvSIMDModImmType12(CnstVal);
5799 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MVT::v2f64,
5800 DAG.getConstant(CnstVal, MVT::i32));
5801 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5804 // The many faces of MVNI...
5806 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5807 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5808 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5809 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5810 DAG.getConstant(CnstVal, MVT::i32),
5811 DAG.getConstant(0, MVT::i32));
5812 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5815 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5816 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5817 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5818 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5819 DAG.getConstant(CnstVal, MVT::i32),
5820 DAG.getConstant(8, MVT::i32));
5821 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5824 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5825 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5826 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5827 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5828 DAG.getConstant(CnstVal, MVT::i32),
5829 DAG.getConstant(16, MVT::i32));
5830 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5833 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5834 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5835 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5836 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5837 DAG.getConstant(CnstVal, MVT::i32),
5838 DAG.getConstant(24, MVT::i32));
5839 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5842 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5843 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5844 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5845 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5846 DAG.getConstant(CnstVal, MVT::i32),
5847 DAG.getConstant(0, MVT::i32));
5848 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5851 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5852 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5853 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5854 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5855 DAG.getConstant(CnstVal, MVT::i32),
5856 DAG.getConstant(8, MVT::i32));
5857 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5860 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
5861 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
5862 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5863 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
5864 DAG.getConstant(CnstVal, MVT::i32),
5865 DAG.getConstant(264, MVT::i32));
5866 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5869 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
5870 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
5871 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5872 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
5873 DAG.getConstant(CnstVal, MVT::i32),
5874 DAG.getConstant(272, MVT::i32));
5875 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5882 CnstBits = UndefBits;
5887 // Scan through the operands to find some interesting properties we can
5889 // 1) If only one value is used, we can use a DUP, or
5890 // 2) if only the low element is not undef, we can just insert that, or
5891 // 3) if only one constant value is used (w/ some non-constant lanes),
5892 // we can splat the constant value into the whole vector then fill
5893 // in the non-constant lanes.
5894 // 4) FIXME: If different constant values are used, but we can intelligently
5895 // select the values we'll be overwriting for the non-constant
5896 // lanes such that we can directly materialize the vector
5897 // some other way (MOVI, e.g.), we can be sneaky.
5898 unsigned NumElts = VT.getVectorNumElements();
5899 bool isOnlyLowElement = true;
5900 bool usesOnlyOneValue = true;
5901 bool usesOnlyOneConstantValue = true;
5902 bool isConstant = true;
5903 unsigned NumConstantLanes = 0;
5905 SDValue ConstantValue;
5906 for (unsigned i = 0; i < NumElts; ++i) {
5907 SDValue V = Op.getOperand(i);
5908 if (V.getOpcode() == ISD::UNDEF)
5911 isOnlyLowElement = false;
5912 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5915 if (isa<ConstantSDNode>(V) || isa<ConstantFPSDNode>(V)) {
5917 if (!ConstantValue.getNode())
5919 else if (ConstantValue != V)
5920 usesOnlyOneConstantValue = false;
5923 if (!Value.getNode())
5925 else if (V != Value)
5926 usesOnlyOneValue = false;
5929 if (!Value.getNode())
5930 return DAG.getUNDEF(VT);
5932 if (isOnlyLowElement)
5933 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5935 // Use DUP for non-constant splats. For f32 constant splats, reduce to
5936 // i32 and try again.
5937 if (usesOnlyOneValue) {
5939 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5940 Value.getValueType() != VT)
5941 return DAG.getNode(AArch64ISD::DUP, dl, VT, Value);
5943 // This is actually a DUPLANExx operation, which keeps everything vectory.
5945 // DUPLANE works on 128-bit vectors, widen it if necessary.
5946 SDValue Lane = Value.getOperand(1);
5947 Value = Value.getOperand(0);
5948 if (Value.getValueType().getSizeInBits() == 64)
5949 Value = WidenVector(Value, DAG);
5951 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
5952 return DAG.getNode(Opcode, dl, VT, Value, Lane);
5955 if (VT.getVectorElementType().isFloatingPoint()) {
5956 SmallVector<SDValue, 8> Ops;
5958 (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
5959 for (unsigned i = 0; i < NumElts; ++i)
5960 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
5961 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
5962 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5963 Val = LowerBUILD_VECTOR(Val, DAG);
5965 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5969 // If there was only one constant value used and for more than one lane,
5970 // start by splatting that value, then replace the non-constant lanes. This
5971 // is better than the default, which will perform a separate initialization
5973 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
5974 SDValue Val = DAG.getNode(AArch64ISD::DUP, dl, VT, ConstantValue);
5975 // Now insert the non-constant lanes.
5976 for (unsigned i = 0; i < NumElts; ++i) {
5977 SDValue V = Op.getOperand(i);
5978 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5979 if (!isa<ConstantSDNode>(V) && !isa<ConstantFPSDNode>(V)) {
5980 // Note that type legalization likely mucked about with the VT of the
5981 // source operand, so we may have to convert it here before inserting.
5982 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
5988 // If all elements are constants and the case above didn't get hit, fall back
5989 // to the default expansion, which will generate a load from the constant
5994 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5996 SDValue shuffle = ReconstructShuffle(Op, DAG);
5997 if (shuffle != SDValue())
6001 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
6002 // know the default expansion would otherwise fall back on something even
6003 // worse. For a vector with one or two non-undef values, that's
6004 // scalar_to_vector for the elements followed by a shuffle (provided the
6005 // shuffle is valid for the target) and materialization element by element
6006 // on the stack followed by a load for everything else.
6007 if (!isConstant && !usesOnlyOneValue) {
6008 SDValue Vec = DAG.getUNDEF(VT);
6009 SDValue Op0 = Op.getOperand(0);
6010 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
6012 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
6013 // a) Avoid a RMW dependency on the full vector register, and
6014 // b) Allow the register coalescer to fold away the copy if the
6015 // value is already in an S or D register.
6016 if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
6017 unsigned SubIdx = ElemSize == 32 ? AArch64::ssub : AArch64::dsub;
6019 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
6020 DAG.getTargetConstant(SubIdx, MVT::i32));
6021 Vec = SDValue(N, 0);
6024 for (; i < NumElts; ++i) {
6025 SDValue V = Op.getOperand(i);
6026 if (V.getOpcode() == ISD::UNDEF)
6028 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
6029 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
6034 // Just use the default expansion. We failed to find a better alternative.
6038 SDValue AArch64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
6039 SelectionDAG &DAG) const {
6040 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
6042 // Check for non-constant or out of range lane.
6043 EVT VT = Op.getOperand(0).getValueType();
6044 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(2));
6045 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6049 // Insertion/extraction are legal for V128 types.
6050 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6051 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6055 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6056 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6059 // For V64 types, we perform insertion by expanding the value
6060 // to a V128 type and perform the insertion on that.
6062 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6063 EVT WideTy = WideVec.getValueType();
6065 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
6066 Op.getOperand(1), Op.getOperand(2));
6067 // Re-narrow the resultant vector.
6068 return NarrowVector(Node, DAG);
6072 AArch64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6073 SelectionDAG &DAG) const {
6074 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
6076 // Check for non-constant or out of range lane.
6077 EVT VT = Op.getOperand(0).getValueType();
6078 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6079 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6083 // Insertion/extraction are legal for V128 types.
6084 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6085 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6089 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6090 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6093 // For V64 types, we perform extraction by expanding the value
6094 // to a V128 type and perform the extraction on that.
6096 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6097 EVT WideTy = WideVec.getValueType();
6099 EVT ExtrTy = WideTy.getVectorElementType();
6100 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
6103 // For extractions, we just return the result directly.
6104 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
6108 SDValue AArch64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
6109 SelectionDAG &DAG) const {
6110 EVT VT = Op.getOperand(0).getValueType();
6116 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6119 unsigned Val = Cst->getZExtValue();
6121 unsigned Size = Op.getValueType().getSizeInBits();
6125 return DAG.getTargetExtractSubreg(AArch64::bsub, dl, Op.getValueType(),
6128 return DAG.getTargetExtractSubreg(AArch64::hsub, dl, Op.getValueType(),
6131 return DAG.getTargetExtractSubreg(AArch64::ssub, dl, Op.getValueType(),
6134 return DAG.getTargetExtractSubreg(AArch64::dsub, dl, Op.getValueType(),
6137 llvm_unreachable("Unexpected vector type in extract_subvector!");
6140 // If this is extracting the upper 64-bits of a 128-bit vector, we match
6142 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
6148 bool AArch64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
6150 if (VT.getVectorNumElements() == 4 &&
6151 (VT.is128BitVector() || VT.is64BitVector())) {
6152 unsigned PFIndexes[4];
6153 for (unsigned i = 0; i != 4; ++i) {
6157 PFIndexes[i] = M[i];
6160 // Compute the index in the perfect shuffle table.
6161 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
6162 PFIndexes[2] * 9 + PFIndexes[3];
6163 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
6164 unsigned Cost = (PFEntry >> 30);
6172 unsigned DummyUnsigned;
6174 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
6175 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
6176 isEXTMask(M, VT, DummyBool, DummyUnsigned) ||
6177 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
6178 isTRNMask(M, VT, DummyUnsigned) || isUZPMask(M, VT, DummyUnsigned) ||
6179 isZIPMask(M, VT, DummyUnsigned) ||
6180 isTRN_v_undef_Mask(M, VT, DummyUnsigned) ||
6181 isUZP_v_undef_Mask(M, VT, DummyUnsigned) ||
6182 isZIP_v_undef_Mask(M, VT, DummyUnsigned) ||
6183 isINSMask(M, VT.getVectorNumElements(), DummyBool, DummyInt) ||
6184 isConcatMask(M, VT, VT.getSizeInBits() == 128));
6187 /// getVShiftImm - Check if this is a valid build_vector for the immediate
6188 /// operand of a vector shift operation, where all the elements of the
6189 /// build_vector must have the same constant integer value.
6190 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
6191 // Ignore bit_converts.
6192 while (Op.getOpcode() == ISD::BITCAST)
6193 Op = Op.getOperand(0);
6194 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
6195 APInt SplatBits, SplatUndef;
6196 unsigned SplatBitSize;
6198 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
6199 HasAnyUndefs, ElementBits) ||
6200 SplatBitSize > ElementBits)
6202 Cnt = SplatBits.getSExtValue();
6206 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
6207 /// operand of a vector shift left operation. That value must be in the range:
6208 /// 0 <= Value < ElementBits for a left shift; or
6209 /// 0 <= Value <= ElementBits for a long left shift.
6210 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
6211 assert(VT.isVector() && "vector shift count is not a vector type");
6212 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6213 if (!getVShiftImm(Op, ElementBits, Cnt))
6215 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
6218 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
6219 /// operand of a vector shift right operation. For a shift opcode, the value
6220 /// is positive, but for an intrinsic the value count must be negative. The
6221 /// absolute value must be in the range:
6222 /// 1 <= |Value| <= ElementBits for a right shift; or
6223 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
6224 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
6226 assert(VT.isVector() && "vector shift count is not a vector type");
6227 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6228 if (!getVShiftImm(Op, ElementBits, Cnt))
6232 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
6235 SDValue AArch64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
6236 SelectionDAG &DAG) const {
6237 EVT VT = Op.getValueType();
6241 if (!Op.getOperand(1).getValueType().isVector())
6243 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
6245 switch (Op.getOpcode()) {
6247 llvm_unreachable("unexpected shift opcode");
6250 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
6251 return DAG.getNode(AArch64ISD::VSHL, SDLoc(Op), VT, Op.getOperand(0),
6252 DAG.getConstant(Cnt, MVT::i32));
6253 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6254 DAG.getConstant(Intrinsic::aarch64_neon_ushl, MVT::i32),
6255 Op.getOperand(0), Op.getOperand(1));
6258 // Right shift immediate
6259 if (isVShiftRImm(Op.getOperand(1), VT, false, false, Cnt) &&
6262 (Op.getOpcode() == ISD::SRA) ? AArch64ISD::VASHR : AArch64ISD::VLSHR;
6263 return DAG.getNode(Opc, SDLoc(Op), VT, Op.getOperand(0),
6264 DAG.getConstant(Cnt, MVT::i32));
6267 // Right shift register. Note, there is not a shift right register
6268 // instruction, but the shift left register instruction takes a signed
6269 // value, where negative numbers specify a right shift.
6270 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::aarch64_neon_sshl
6271 : Intrinsic::aarch64_neon_ushl;
6272 // negate the shift amount
6273 SDValue NegShift = DAG.getNode(AArch64ISD::NEG, DL, VT, Op.getOperand(1));
6274 SDValue NegShiftLeft =
6275 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6276 DAG.getConstant(Opc, MVT::i32), Op.getOperand(0), NegShift);
6277 return NegShiftLeft;
6283 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
6284 AArch64CC::CondCode CC, bool NoNans, EVT VT,
6285 SDLoc dl, SelectionDAG &DAG) {
6286 EVT SrcVT = LHS.getValueType();
6288 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
6289 APInt CnstBits(VT.getSizeInBits(), 0);
6290 APInt UndefBits(VT.getSizeInBits(), 0);
6291 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
6292 bool IsZero = IsCnst && (CnstBits == 0);
6294 if (SrcVT.getVectorElementType().isFloatingPoint()) {
6298 case AArch64CC::NE: {
6301 Fcmeq = DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6303 Fcmeq = DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6304 return DAG.getNode(AArch64ISD::NOT, dl, VT, Fcmeq);
6308 return DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6309 return DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6312 return DAG.getNode(AArch64ISD::FCMGEz, dl, VT, LHS);
6313 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, LHS, RHS);
6316 return DAG.getNode(AArch64ISD::FCMGTz, dl, VT, LHS);
6317 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, LHS, RHS);
6320 return DAG.getNode(AArch64ISD::FCMLEz, dl, VT, LHS);
6321 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, RHS, LHS);
6325 // If we ignore NaNs then we can use to the MI implementation.
6329 return DAG.getNode(AArch64ISD::FCMLTz, dl, VT, LHS);
6330 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, RHS, LHS);
6337 case AArch64CC::NE: {
6340 Cmeq = DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6342 Cmeq = DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6343 return DAG.getNode(AArch64ISD::NOT, dl, VT, Cmeq);
6347 return DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6348 return DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6351 return DAG.getNode(AArch64ISD::CMGEz, dl, VT, LHS);
6352 return DAG.getNode(AArch64ISD::CMGE, dl, VT, LHS, RHS);
6355 return DAG.getNode(AArch64ISD::CMGTz, dl, VT, LHS);
6356 return DAG.getNode(AArch64ISD::CMGT, dl, VT, LHS, RHS);
6359 return DAG.getNode(AArch64ISD::CMLEz, dl, VT, LHS);
6360 return DAG.getNode(AArch64ISD::CMGE, dl, VT, RHS, LHS);
6362 return DAG.getNode(AArch64ISD::CMHS, dl, VT, RHS, LHS);
6364 return DAG.getNode(AArch64ISD::CMHI, dl, VT, RHS, LHS);
6367 return DAG.getNode(AArch64ISD::CMLTz, dl, VT, LHS);
6368 return DAG.getNode(AArch64ISD::CMGT, dl, VT, RHS, LHS);
6370 return DAG.getNode(AArch64ISD::CMHI, dl, VT, LHS, RHS);
6372 return DAG.getNode(AArch64ISD::CMHS, dl, VT, LHS, RHS);
6376 SDValue AArch64TargetLowering::LowerVSETCC(SDValue Op,
6377 SelectionDAG &DAG) const {
6378 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6379 SDValue LHS = Op.getOperand(0);
6380 SDValue RHS = Op.getOperand(1);
6383 if (LHS.getValueType().getVectorElementType().isInteger()) {
6384 assert(LHS.getValueType() == RHS.getValueType());
6385 AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC);
6386 return EmitVectorComparison(LHS, RHS, AArch64CC, false, Op.getValueType(),
6390 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
6391 LHS.getValueType().getVectorElementType() == MVT::f64);
6393 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
6394 // clean. Some of them require two branches to implement.
6395 AArch64CC::CondCode CC1, CC2;
6397 changeVectorFPCCToAArch64CC(CC, CC1, CC2, ShouldInvert);
6399 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
6401 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, Op.getValueType(), dl, DAG);
6405 if (CC2 != AArch64CC::AL) {
6407 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, Op.getValueType(), dl, DAG);
6408 if (!Cmp2.getNode())
6411 Cmp = DAG.getNode(ISD::OR, dl, Cmp.getValueType(), Cmp, Cmp2);
6415 return Cmp = DAG.getNOT(dl, Cmp, Cmp.getValueType());
6420 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
6421 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
6422 /// specified in the intrinsic calls.
6423 bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
6425 unsigned Intrinsic) const {
6426 switch (Intrinsic) {
6427 case Intrinsic::aarch64_neon_ld2:
6428 case Intrinsic::aarch64_neon_ld3:
6429 case Intrinsic::aarch64_neon_ld4:
6430 case Intrinsic::aarch64_neon_ld1x2:
6431 case Intrinsic::aarch64_neon_ld1x3:
6432 case Intrinsic::aarch64_neon_ld1x4:
6433 case Intrinsic::aarch64_neon_ld2lane:
6434 case Intrinsic::aarch64_neon_ld3lane:
6435 case Intrinsic::aarch64_neon_ld4lane:
6436 case Intrinsic::aarch64_neon_ld2r:
6437 case Intrinsic::aarch64_neon_ld3r:
6438 case Intrinsic::aarch64_neon_ld4r: {
6439 Info.opc = ISD::INTRINSIC_W_CHAIN;
6440 // Conservatively set memVT to the entire set of vectors loaded.
6441 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
6442 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6443 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6446 Info.vol = false; // volatile loads with NEON intrinsics not supported
6447 Info.readMem = true;
6448 Info.writeMem = false;
6451 case Intrinsic::aarch64_neon_st2:
6452 case Intrinsic::aarch64_neon_st3:
6453 case Intrinsic::aarch64_neon_st4:
6454 case Intrinsic::aarch64_neon_st1x2:
6455 case Intrinsic::aarch64_neon_st1x3:
6456 case Intrinsic::aarch64_neon_st1x4:
6457 case Intrinsic::aarch64_neon_st2lane:
6458 case Intrinsic::aarch64_neon_st3lane:
6459 case Intrinsic::aarch64_neon_st4lane: {
6460 Info.opc = ISD::INTRINSIC_VOID;
6461 // Conservatively set memVT to the entire set of vectors stored.
6462 unsigned NumElts = 0;
6463 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
6464 Type *ArgTy = I.getArgOperand(ArgI)->getType();
6465 if (!ArgTy->isVectorTy())
6467 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
6469 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6470 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6473 Info.vol = false; // volatile stores with NEON intrinsics not supported
6474 Info.readMem = false;
6475 Info.writeMem = true;
6478 case Intrinsic::aarch64_ldaxr:
6479 case Intrinsic::aarch64_ldxr: {
6480 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
6481 Info.opc = ISD::INTRINSIC_W_CHAIN;
6482 Info.memVT = MVT::getVT(PtrTy->getElementType());
6483 Info.ptrVal = I.getArgOperand(0);
6485 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
6487 Info.readMem = true;
6488 Info.writeMem = false;
6491 case Intrinsic::aarch64_stlxr:
6492 case Intrinsic::aarch64_stxr: {
6493 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
6494 Info.opc = ISD::INTRINSIC_W_CHAIN;
6495 Info.memVT = MVT::getVT(PtrTy->getElementType());
6496 Info.ptrVal = I.getArgOperand(1);
6498 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
6500 Info.readMem = false;
6501 Info.writeMem = true;
6504 case Intrinsic::aarch64_ldaxp:
6505 case Intrinsic::aarch64_ldxp: {
6506 Info.opc = ISD::INTRINSIC_W_CHAIN;
6507 Info.memVT = MVT::i128;
6508 Info.ptrVal = I.getArgOperand(0);
6512 Info.readMem = true;
6513 Info.writeMem = false;
6516 case Intrinsic::aarch64_stlxp:
6517 case Intrinsic::aarch64_stxp: {
6518 Info.opc = ISD::INTRINSIC_W_CHAIN;
6519 Info.memVT = MVT::i128;
6520 Info.ptrVal = I.getArgOperand(2);
6524 Info.readMem = false;
6525 Info.writeMem = true;
6535 // Truncations from 64-bit GPR to 32-bit GPR is free.
6536 bool AArch64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
6537 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6539 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6540 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6541 return NumBits1 > NumBits2;
6543 bool AArch64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
6544 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6546 unsigned NumBits1 = VT1.getSizeInBits();
6547 unsigned NumBits2 = VT2.getSizeInBits();
6548 return NumBits1 > NumBits2;
6551 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
6553 bool AArch64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
6554 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6556 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6557 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6558 return NumBits1 == 32 && NumBits2 == 64;
6560 bool AArch64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
6561 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6563 unsigned NumBits1 = VT1.getSizeInBits();
6564 unsigned NumBits2 = VT2.getSizeInBits();
6565 return NumBits1 == 32 && NumBits2 == 64;
6568 bool AArch64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
6569 EVT VT1 = Val.getValueType();
6570 if (isZExtFree(VT1, VT2)) {
6574 if (Val.getOpcode() != ISD::LOAD)
6577 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
6578 return (VT1.isSimple() && !VT1.isVector() && VT1.isInteger() &&
6579 VT2.isSimple() && !VT2.isVector() && VT2.isInteger() &&
6580 VT1.getSizeInBits() <= 32);
6583 bool AArch64TargetLowering::hasPairedLoad(Type *LoadedType,
6584 unsigned &RequiredAligment) const {
6585 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
6587 // Cyclone supports unaligned accesses.
6588 RequiredAligment = 0;
6589 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
6590 return NumBits == 32 || NumBits == 64;
6593 bool AArch64TargetLowering::hasPairedLoad(EVT LoadedType,
6594 unsigned &RequiredAligment) const {
6595 if (!LoadedType.isSimple() ||
6596 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
6598 // Cyclone supports unaligned accesses.
6599 RequiredAligment = 0;
6600 unsigned NumBits = LoadedType.getSizeInBits();
6601 return NumBits == 32 || NumBits == 64;
6604 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
6605 unsigned AlignCheck) {
6606 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
6607 (DstAlign == 0 || DstAlign % AlignCheck == 0));
6610 EVT AArch64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
6611 unsigned SrcAlign, bool IsMemset,
6614 MachineFunction &MF) const {
6615 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
6616 // instruction to materialize the v2i64 zero and one store (with restrictive
6617 // addressing mode). Just do two i64 store of zero-registers.
6619 const Function *F = MF.getFunction();
6620 if (Subtarget->hasFPARMv8() && !IsMemset && Size >= 16 &&
6621 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
6622 Attribute::NoImplicitFloat) &&
6623 (memOpAlign(SrcAlign, DstAlign, 16) ||
6624 (allowsMisalignedMemoryAccesses(MVT::f128, 0, 1, &Fast) && Fast)))
6627 return Size >= 8 ? MVT::i64 : MVT::i32;
6630 // 12-bit optionally shifted immediates are legal for adds.
6631 bool AArch64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
6632 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
6637 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
6638 // immediates is the same as for an add or a sub.
6639 bool AArch64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
6642 return isLegalAddImmediate(Immed);
6645 /// isLegalAddressingMode - Return true if the addressing mode represented
6646 /// by AM is legal for this target, for a load/store of the specified type.
6647 bool AArch64TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6649 // AArch64 has five basic addressing modes:
6651 // reg + 9-bit signed offset
6652 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
6654 // reg + SIZE_IN_BYTES * reg
6656 // No global is ever allowed as a base.
6660 // No reg+reg+imm addressing.
6661 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
6664 // check reg + imm case:
6665 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
6666 uint64_t NumBytes = 0;
6667 if (Ty->isSized()) {
6668 uint64_t NumBits = getDataLayout()->getTypeSizeInBits(Ty);
6669 NumBytes = NumBits / 8;
6670 if (!isPowerOf2_64(NumBits))
6675 int64_t Offset = AM.BaseOffs;
6677 // 9-bit signed offset
6678 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
6681 // 12-bit unsigned offset
6682 unsigned shift = Log2_64(NumBytes);
6683 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
6684 // Must be a multiple of NumBytes (NumBytes is a power of 2)
6685 (Offset >> shift) << shift == Offset)
6690 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
6692 if (!AM.Scale || AM.Scale == 1 ||
6693 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
6698 int AArch64TargetLowering::getScalingFactorCost(const AddrMode &AM,
6700 // Scaling factors are not free at all.
6701 // Operands | Rt Latency
6702 // -------------------------------------------
6704 // -------------------------------------------
6705 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
6706 // Rt, [Xn, Wm, <extend> #imm] |
6707 if (isLegalAddressingMode(AM, Ty))
6708 // Scale represents reg2 * scale, thus account for 1 if
6709 // it is not equal to 0 or 1.
6710 return AM.Scale != 0 && AM.Scale != 1;
6714 bool AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
6715 VT = VT.getScalarType();
6720 switch (VT.getSimpleVT().SimpleTy) {
6732 AArch64TargetLowering::getScratchRegisters(CallingConv::ID) const {
6733 // LR is a callee-save register, but we must treat it as clobbered by any call
6734 // site. Hence we include LR in the scratch registers, which are in turn added
6735 // as implicit-defs for stackmaps and patchpoints.
6736 static const MCPhysReg ScratchRegs[] = {
6737 AArch64::X16, AArch64::X17, AArch64::LR, 0
6743 AArch64TargetLowering::isDesirableToCommuteWithShift(const SDNode *N) const {
6744 EVT VT = N->getValueType(0);
6745 // If N is unsigned bit extraction: ((x >> C) & mask), then do not combine
6746 // it with shift to let it be lowered to UBFX.
6747 if (N->getOpcode() == ISD::AND && (VT == MVT::i32 || VT == MVT::i64) &&
6748 isa<ConstantSDNode>(N->getOperand(1))) {
6749 uint64_t TruncMask = N->getConstantOperandVal(1);
6750 if (isMask_64(TruncMask) &&
6751 N->getOperand(0).getOpcode() == ISD::SRL &&
6752 isa<ConstantSDNode>(N->getOperand(0)->getOperand(1)))
6758 bool AArch64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
6760 assert(Ty->isIntegerTy());
6762 unsigned BitSize = Ty->getPrimitiveSizeInBits();
6766 int64_t Val = Imm.getSExtValue();
6767 if (Val == 0 || AArch64_AM::isLogicalImmediate(Val, BitSize))
6770 if ((int64_t)Val < 0)
6773 Val &= (1LL << 32) - 1;
6775 unsigned LZ = countLeadingZeros((uint64_t)Val);
6776 unsigned Shift = (63 - LZ) / 16;
6777 // MOVZ is free so return true for one or fewer MOVK.
6778 return (Shift < 3) ? true : false;
6781 // Generate SUBS and CSEL for integer abs.
6782 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
6783 EVT VT = N->getValueType(0);
6785 SDValue N0 = N->getOperand(0);
6786 SDValue N1 = N->getOperand(1);
6789 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
6790 // and change it to SUB and CSEL.
6791 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
6792 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
6793 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
6794 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
6795 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
6796 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
6798 // Generate SUBS & CSEL.
6800 DAG.getNode(AArch64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
6801 N0.getOperand(0), DAG.getConstant(0, VT));
6802 return DAG.getNode(AArch64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
6803 DAG.getConstant(AArch64CC::PL, MVT::i32),
6804 SDValue(Cmp.getNode(), 1));
6809 // performXorCombine - Attempts to handle integer ABS.
6810 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
6811 TargetLowering::DAGCombinerInfo &DCI,
6812 const AArch64Subtarget *Subtarget) {
6813 if (DCI.isBeforeLegalizeOps())
6816 return performIntegerAbsCombine(N, DAG);
6820 AArch64TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
6822 std::vector<SDNode *> *Created) const {
6823 // fold (sdiv X, pow2)
6824 EVT VT = N->getValueType(0);
6825 if ((VT != MVT::i32 && VT != MVT::i64) ||
6826 !(Divisor.isPowerOf2() || (-Divisor).isPowerOf2()))
6830 SDValue N0 = N->getOperand(0);
6831 unsigned Lg2 = Divisor.countTrailingZeros();
6832 SDValue Zero = DAG.getConstant(0, VT);
6833 SDValue Pow2MinusOne = DAG.getConstant((1ULL << Lg2) - 1, VT);
6835 // Add (N0 < 0) ? Pow2 - 1 : 0;
6837 SDValue Cmp = getAArch64Cmp(N0, Zero, ISD::SETLT, CCVal, DAG, DL);
6838 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Pow2MinusOne);
6839 SDValue CSel = DAG.getNode(AArch64ISD::CSEL, DL, VT, Add, N0, CCVal, Cmp);
6842 Created->push_back(Cmp.getNode());
6843 Created->push_back(Add.getNode());
6844 Created->push_back(CSel.getNode());
6849 DAG.getNode(ISD::SRA, DL, VT, CSel, DAG.getConstant(Lg2, MVT::i64));
6851 // If we're dividing by a positive value, we're done. Otherwise, we must
6852 // negate the result.
6853 if (Divisor.isNonNegative())
6857 Created->push_back(SRA.getNode());
6858 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT), SRA);
6861 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
6862 TargetLowering::DAGCombinerInfo &DCI,
6863 const AArch64Subtarget *Subtarget) {
6864 if (DCI.isBeforeLegalizeOps())
6867 // Multiplication of a power of two plus/minus one can be done more
6868 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
6869 // future CPUs have a cheaper MADD instruction, this may need to be
6870 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
6871 // 64-bit is 5 cycles, so this is always a win.
6872 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
6873 APInt Value = C->getAPIntValue();
6874 EVT VT = N->getValueType(0);
6875 if (Value.isNonNegative()) {
6876 // (mul x, 2^N + 1) => (add (shl x, N), x)
6877 APInt VM1 = Value - 1;
6878 if (VM1.isPowerOf2()) {
6879 SDValue ShiftedVal =
6880 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6881 DAG.getConstant(VM1.logBase2(), MVT::i64));
6882 return DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal,
6885 // (mul x, 2^N - 1) => (sub (shl x, N), x)
6886 APInt VP1 = Value + 1;
6887 if (VP1.isPowerOf2()) {
6888 SDValue ShiftedVal =
6889 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6890 DAG.getConstant(VP1.logBase2(), MVT::i64));
6891 return DAG.getNode(ISD::SUB, SDLoc(N), VT, ShiftedVal,
6895 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
6896 APInt VNM1 = -Value - 1;
6897 if (VNM1.isPowerOf2()) {
6898 SDValue ShiftedVal =
6899 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6900 DAG.getConstant(VNM1.logBase2(), MVT::i64));
6902 DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6903 return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), Add);
6905 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
6906 APInt VNP1 = -Value + 1;
6907 if (VNP1.isPowerOf2()) {
6908 SDValue ShiftedVal =
6909 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6910 DAG.getConstant(VNP1.logBase2(), MVT::i64));
6911 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N->getOperand(0),
6919 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
6920 SelectionDAG &DAG) {
6921 // Take advantage of vector comparisons producing 0 or -1 in each lane to
6922 // optimize away operation when it's from a constant.
6924 // The general transformation is:
6925 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
6926 // AND(VECTOR_CMP(x,y), constant2)
6927 // constant2 = UNARYOP(constant)
6929 // Early exit if this isn't a vector operation, the operand of the
6930 // unary operation isn't a bitwise AND, or if the sizes of the operations
6932 EVT VT = N->getValueType(0);
6933 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
6934 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
6935 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
6938 // Now check that the other operand of the AND is a constant. We could
6939 // make the transformation for non-constant splats as well, but it's unclear
6940 // that would be a benefit as it would not eliminate any operations, just
6941 // perform one more step in scalar code before moving to the vector unit.
6942 if (BuildVectorSDNode *BV =
6943 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
6944 // Bail out if the vector isn't a constant.
6945 if (!BV->isConstant())
6948 // Everything checks out. Build up the new and improved node.
6950 EVT IntVT = BV->getValueType(0);
6951 // Create a new constant of the appropriate type for the transformed
6953 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
6954 // The AND node needs bitcasts to/from an integer vector type around it.
6955 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
6956 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
6957 N->getOperand(0)->getOperand(0), MaskConst);
6958 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
6965 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG,
6966 const AArch64Subtarget *Subtarget) {
6967 // First try to optimize away the conversion when it's conditionally from
6968 // a constant. Vectors only.
6969 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
6970 if (Res != SDValue())
6973 EVT VT = N->getValueType(0);
6974 if (VT != MVT::f32 && VT != MVT::f64)
6977 // Only optimize when the source and destination types have the same width.
6978 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
6981 // If the result of an integer load is only used by an integer-to-float
6982 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
6983 // This eliminates an "integer-to-vector-move UOP and improve throughput.
6984 SDValue N0 = N->getOperand(0);
6985 if (Subtarget->hasNEON() && ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6986 // Do not change the width of a volatile load.
6987 !cast<LoadSDNode>(N0)->isVolatile()) {
6988 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6989 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
6990 LN0->getPointerInfo(), LN0->isVolatile(),
6991 LN0->isNonTemporal(), LN0->isInvariant(),
6992 LN0->getAlignment());
6994 // Make sure successors of the original load stay after it by updating them
6995 // to use the new Chain.
6996 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
6999 (N->getOpcode() == ISD::SINT_TO_FP) ? AArch64ISD::SITOF : AArch64ISD::UITOF;
7000 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
7006 /// An EXTR instruction is made up of two shifts, ORed together. This helper
7007 /// searches for and classifies those shifts.
7008 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
7010 if (N.getOpcode() == ISD::SHL)
7012 else if (N.getOpcode() == ISD::SRL)
7017 if (!isa<ConstantSDNode>(N.getOperand(1)))
7020 ShiftAmount = N->getConstantOperandVal(1);
7021 Src = N->getOperand(0);
7025 /// EXTR instruction extracts a contiguous chunk of bits from two existing
7026 /// registers viewed as a high/low pair. This function looks for the pattern:
7027 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
7028 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
7030 static SDValue tryCombineToEXTR(SDNode *N,
7031 TargetLowering::DAGCombinerInfo &DCI) {
7032 SelectionDAG &DAG = DCI.DAG;
7034 EVT VT = N->getValueType(0);
7036 assert(N->getOpcode() == ISD::OR && "Unexpected root");
7038 if (VT != MVT::i32 && VT != MVT::i64)
7042 uint32_t ShiftLHS = 0;
7044 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
7048 uint32_t ShiftRHS = 0;
7050 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
7053 // If they're both trying to come from the high part of the register, they're
7054 // not really an EXTR.
7055 if (LHSFromHi == RHSFromHi)
7058 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
7062 std::swap(LHS, RHS);
7063 std::swap(ShiftLHS, ShiftRHS);
7066 return DAG.getNode(AArch64ISD::EXTR, DL, VT, LHS, RHS,
7067 DAG.getConstant(ShiftRHS, MVT::i64));
7070 static SDValue tryCombineToBSL(SDNode *N,
7071 TargetLowering::DAGCombinerInfo &DCI) {
7072 EVT VT = N->getValueType(0);
7073 SelectionDAG &DAG = DCI.DAG;
7079 SDValue N0 = N->getOperand(0);
7080 if (N0.getOpcode() != ISD::AND)
7083 SDValue N1 = N->getOperand(1);
7084 if (N1.getOpcode() != ISD::AND)
7087 // We only have to look for constant vectors here since the general, variable
7088 // case can be handled in TableGen.
7089 unsigned Bits = VT.getVectorElementType().getSizeInBits();
7090 uint64_t BitMask = Bits == 64 ? -1ULL : ((1ULL << Bits) - 1);
7091 for (int i = 1; i >= 0; --i)
7092 for (int j = 1; j >= 0; --j) {
7093 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(i));
7094 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(j));
7098 bool FoundMatch = true;
7099 for (unsigned k = 0; k < VT.getVectorNumElements(); ++k) {
7100 ConstantSDNode *CN0 = dyn_cast<ConstantSDNode>(BVN0->getOperand(k));
7101 ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(BVN1->getOperand(k));
7103 CN0->getZExtValue() != (BitMask & ~CN1->getZExtValue())) {
7110 return DAG.getNode(AArch64ISD::BSL, DL, VT, SDValue(BVN0, 0),
7111 N0->getOperand(1 - i), N1->getOperand(1 - j));
7117 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
7118 const AArch64Subtarget *Subtarget) {
7119 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
7120 if (!EnableAArch64ExtrGeneration)
7122 SelectionDAG &DAG = DCI.DAG;
7123 EVT VT = N->getValueType(0);
7125 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7128 SDValue Res = tryCombineToEXTR(N, DCI);
7132 Res = tryCombineToBSL(N, DCI);
7139 static SDValue performBitcastCombine(SDNode *N,
7140 TargetLowering::DAGCombinerInfo &DCI,
7141 SelectionDAG &DAG) {
7142 // Wait 'til after everything is legalized to try this. That way we have
7143 // legal vector types and such.
7144 if (DCI.isBeforeLegalizeOps())
7147 // Remove extraneous bitcasts around an extract_subvector.
7149 // (v4i16 (bitconvert
7150 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
7152 // (extract_subvector ((v8i16 ...), (i64 4)))
7154 // Only interested in 64-bit vectors as the ultimate result.
7155 EVT VT = N->getValueType(0);
7158 if (VT.getSimpleVT().getSizeInBits() != 64)
7160 // Is the operand an extract_subvector starting at the beginning or halfway
7161 // point of the vector? A low half may also come through as an
7162 // EXTRACT_SUBREG, so look for that, too.
7163 SDValue Op0 = N->getOperand(0);
7164 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
7165 !(Op0->isMachineOpcode() &&
7166 Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG))
7168 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
7169 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
7170 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
7172 } else if (Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG) {
7173 if (idx != AArch64::dsub)
7175 // The dsub reference is equivalent to a lane zero subvector reference.
7178 // Look through the bitcast of the input to the extract.
7179 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
7181 SDValue Source = Op0->getOperand(0)->getOperand(0);
7182 // If the source type has twice the number of elements as our destination
7183 // type, we know this is an extract of the high or low half of the vector.
7184 EVT SVT = Source->getValueType(0);
7185 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
7188 DEBUG(dbgs() << "aarch64-lower: bitcast extract_subvector simplification\n");
7190 // Create the simplified form to just extract the low or high half of the
7191 // vector directly rather than bothering with the bitcasts.
7193 unsigned NumElements = VT.getVectorNumElements();
7195 SDValue HalfIdx = DAG.getConstant(NumElements, MVT::i64);
7196 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
7198 SDValue SubReg = DAG.getTargetConstant(AArch64::dsub, MVT::i32);
7199 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
7205 static SDValue performConcatVectorsCombine(SDNode *N,
7206 TargetLowering::DAGCombinerInfo &DCI,
7207 SelectionDAG &DAG) {
7208 // Wait 'til after everything is legalized to try this. That way we have
7209 // legal vector types and such.
7210 if (DCI.isBeforeLegalizeOps())
7214 EVT VT = N->getValueType(0);
7216 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
7217 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
7218 // canonicalise to that.
7219 if (N->getOperand(0) == N->getOperand(1) && VT.getVectorNumElements() == 2) {
7220 assert(VT.getVectorElementType().getSizeInBits() == 64);
7221 return DAG.getNode(AArch64ISD::DUPLANE64, dl, VT,
7222 WidenVector(N->getOperand(0), DAG),
7223 DAG.getConstant(0, MVT::i64));
7226 // Canonicalise concat_vectors so that the right-hand vector has as few
7227 // bit-casts as possible before its real operation. The primary matching
7228 // destination for these operations will be the narrowing "2" instructions,
7229 // which depend on the operation being performed on this right-hand vector.
7231 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
7233 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
7235 SDValue Op1 = N->getOperand(1);
7236 if (Op1->getOpcode() != ISD::BITCAST)
7238 SDValue RHS = Op1->getOperand(0);
7239 MVT RHSTy = RHS.getValueType().getSimpleVT();
7240 // If the RHS is not a vector, this is not the pattern we're looking for.
7241 if (!RHSTy.isVector())
7244 DEBUG(dbgs() << "aarch64-lower: concat_vectors bitcast simplification\n");
7246 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
7247 RHSTy.getVectorNumElements() * 2);
7249 ISD::BITCAST, dl, VT,
7250 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
7251 DAG.getNode(ISD::BITCAST, dl, RHSTy, N->getOperand(0)), RHS));
7254 static SDValue tryCombineFixedPointConvert(SDNode *N,
7255 TargetLowering::DAGCombinerInfo &DCI,
7256 SelectionDAG &DAG) {
7257 // Wait 'til after everything is legalized to try this. That way we have
7258 // legal vector types and such.
7259 if (DCI.isBeforeLegalizeOps())
7261 // Transform a scalar conversion of a value from a lane extract into a
7262 // lane extract of a vector conversion. E.g., from foo1 to foo2:
7263 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
7264 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
7266 // The second form interacts better with instruction selection and the
7267 // register allocator to avoid cross-class register copies that aren't
7268 // coalescable due to a lane reference.
7270 // Check the operand and see if it originates from a lane extract.
7271 SDValue Op1 = N->getOperand(1);
7272 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7273 // Yep, no additional predication needed. Perform the transform.
7274 SDValue IID = N->getOperand(0);
7275 SDValue Shift = N->getOperand(2);
7276 SDValue Vec = Op1.getOperand(0);
7277 SDValue Lane = Op1.getOperand(1);
7278 EVT ResTy = N->getValueType(0);
7282 // The vector width should be 128 bits by the time we get here, even
7283 // if it started as 64 bits (the extract_vector handling will have
7285 assert(Vec.getValueType().getSizeInBits() == 128 &&
7286 "unexpected vector size on extract_vector_elt!");
7287 if (Vec.getValueType() == MVT::v4i32)
7288 VecResTy = MVT::v4f32;
7289 else if (Vec.getValueType() == MVT::v2i64)
7290 VecResTy = MVT::v2f64;
7292 llvm_unreachable("unexpected vector type!");
7295 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
7296 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
7301 // AArch64 high-vector "long" operations are formed by performing the non-high
7302 // version on an extract_subvector of each operand which gets the high half:
7304 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
7306 // However, there are cases which don't have an extract_high explicitly, but
7307 // have another operation that can be made compatible with one for free. For
7310 // (dupv64 scalar) --> (extract_high (dup128 scalar))
7312 // This routine does the actual conversion of such DUPs, once outer routines
7313 // have determined that everything else is in order.
7314 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
7315 // We can handle most types of duplicate, but the lane ones have an extra
7316 // operand saying *which* lane, so we need to know.
7318 switch (N.getOpcode()) {
7319 case AArch64ISD::DUP:
7322 case AArch64ISD::DUPLANE8:
7323 case AArch64ISD::DUPLANE16:
7324 case AArch64ISD::DUPLANE32:
7325 case AArch64ISD::DUPLANE64:
7332 MVT NarrowTy = N.getSimpleValueType();
7333 if (!NarrowTy.is64BitVector())
7336 MVT ElementTy = NarrowTy.getVectorElementType();
7337 unsigned NumElems = NarrowTy.getVectorNumElements();
7338 MVT NewDUPVT = MVT::getVectorVT(ElementTy, NumElems * 2);
7342 NewDUP = DAG.getNode(N.getOpcode(), SDLoc(N), NewDUPVT, N.getOperand(0),
7345 NewDUP = DAG.getNode(AArch64ISD::DUP, SDLoc(N), NewDUPVT, N.getOperand(0));
7347 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N.getNode()), NarrowTy,
7348 NewDUP, DAG.getConstant(NumElems, MVT::i64));
7351 static bool isEssentiallyExtractSubvector(SDValue N) {
7352 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
7355 return N.getOpcode() == ISD::BITCAST &&
7356 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
7359 /// \brief Helper structure to keep track of ISD::SET_CC operands.
7360 struct GenericSetCCInfo {
7361 const SDValue *Opnd0;
7362 const SDValue *Opnd1;
7366 /// \brief Helper structure to keep track of a SET_CC lowered into AArch64 code.
7367 struct AArch64SetCCInfo {
7369 AArch64CC::CondCode CC;
7372 /// \brief Helper structure to keep track of SetCC information.
7374 GenericSetCCInfo Generic;
7375 AArch64SetCCInfo AArch64;
7378 /// \brief Helper structure to be able to read SetCC information. If set to
7379 /// true, IsAArch64 field, Info is a AArch64SetCCInfo, otherwise Info is a
7380 /// GenericSetCCInfo.
7381 struct SetCCInfoAndKind {
7386 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
7388 /// AArch64 lowered one.
7389 /// \p SetCCInfo is filled accordingly.
7390 /// \post SetCCInfo is meanginfull only when this function returns true.
7391 /// \return True when Op is a kind of SET_CC operation.
7392 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
7393 // If this is a setcc, this is straight forward.
7394 if (Op.getOpcode() == ISD::SETCC) {
7395 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
7396 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
7397 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7398 SetCCInfo.IsAArch64 = false;
7401 // Otherwise, check if this is a matching csel instruction.
7405 if (Op.getOpcode() != AArch64ISD::CSEL)
7407 // Set the information about the operands.
7408 // TODO: we want the operands of the Cmp not the csel
7409 SetCCInfo.Info.AArch64.Cmp = &Op.getOperand(3);
7410 SetCCInfo.IsAArch64 = true;
7411 SetCCInfo.Info.AArch64.CC = static_cast<AArch64CC::CondCode>(
7412 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
7414 // Check that the operands matches the constraints:
7415 // (1) Both operands must be constants.
7416 // (2) One must be 1 and the other must be 0.
7417 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
7418 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7421 if (!TValue || !FValue)
7425 if (!TValue->isOne()) {
7426 // Update the comparison when we are interested in !cc.
7427 std::swap(TValue, FValue);
7428 SetCCInfo.Info.AArch64.CC =
7429 AArch64CC::getInvertedCondCode(SetCCInfo.Info.AArch64.CC);
7431 return TValue->isOne() && FValue->isNullValue();
7434 // Returns true if Op is setcc or zext of setcc.
7435 static bool isSetCCOrZExtSetCC(const SDValue& Op, SetCCInfoAndKind &Info) {
7436 if (isSetCC(Op, Info))
7438 return ((Op.getOpcode() == ISD::ZERO_EXTEND) &&
7439 isSetCC(Op->getOperand(0), Info));
7442 // The folding we want to perform is:
7443 // (add x, [zext] (setcc cc ...) )
7445 // (csel x, (add x, 1), !cc ...)
7447 // The latter will get matched to a CSINC instruction.
7448 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
7449 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
7450 SDValue LHS = Op->getOperand(0);
7451 SDValue RHS = Op->getOperand(1);
7452 SetCCInfoAndKind InfoAndKind;
7454 // If neither operand is a SET_CC, give up.
7455 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind)) {
7456 std::swap(LHS, RHS);
7457 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind))
7461 // FIXME: This could be generatized to work for FP comparisons.
7462 EVT CmpVT = InfoAndKind.IsAArch64
7463 ? InfoAndKind.Info.AArch64.Cmp->getOperand(0).getValueType()
7464 : InfoAndKind.Info.Generic.Opnd0->getValueType();
7465 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
7471 if (InfoAndKind.IsAArch64) {
7472 CCVal = DAG.getConstant(
7473 AArch64CC::getInvertedCondCode(InfoAndKind.Info.AArch64.CC), MVT::i32);
7474 Cmp = *InfoAndKind.Info.AArch64.Cmp;
7476 Cmp = getAArch64Cmp(*InfoAndKind.Info.Generic.Opnd0,
7477 *InfoAndKind.Info.Generic.Opnd1,
7478 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
7481 EVT VT = Op->getValueType(0);
7482 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, VT));
7483 return DAG.getNode(AArch64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
7486 // The basic add/sub long vector instructions have variants with "2" on the end
7487 // which act on the high-half of their inputs. They are normally matched by
7490 // (add (zeroext (extract_high LHS)),
7491 // (zeroext (extract_high RHS)))
7492 // -> uaddl2 vD, vN, vM
7494 // However, if one of the extracts is something like a duplicate, this
7495 // instruction can still be used profitably. This function puts the DAG into a
7496 // more appropriate form for those patterns to trigger.
7497 static SDValue performAddSubLongCombine(SDNode *N,
7498 TargetLowering::DAGCombinerInfo &DCI,
7499 SelectionDAG &DAG) {
7500 if (DCI.isBeforeLegalizeOps())
7503 MVT VT = N->getSimpleValueType(0);
7504 if (!VT.is128BitVector()) {
7505 if (N->getOpcode() == ISD::ADD)
7506 return performSetccAddFolding(N, DAG);
7510 // Make sure both branches are extended in the same way.
7511 SDValue LHS = N->getOperand(0);
7512 SDValue RHS = N->getOperand(1);
7513 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
7514 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
7515 LHS.getOpcode() != RHS.getOpcode())
7518 unsigned ExtType = LHS.getOpcode();
7520 // It's not worth doing if at least one of the inputs isn't already an
7521 // extract, but we don't know which it'll be so we have to try both.
7522 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
7523 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
7527 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
7528 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
7529 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
7533 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
7536 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
7539 // Massage DAGs which we can use the high-half "long" operations on into
7540 // something isel will recognize better. E.g.
7542 // (aarch64_neon_umull (extract_high vec) (dupv64 scalar)) -->
7543 // (aarch64_neon_umull (extract_high (v2i64 vec)))
7544 // (extract_high (v2i64 (dup128 scalar)))))
7546 static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N,
7547 TargetLowering::DAGCombinerInfo &DCI,
7548 SelectionDAG &DAG) {
7549 if (DCI.isBeforeLegalizeOps())
7552 SDValue LHS = N->getOperand(1);
7553 SDValue RHS = N->getOperand(2);
7554 assert(LHS.getValueType().is64BitVector() &&
7555 RHS.getValueType().is64BitVector() &&
7556 "unexpected shape for long operation");
7558 // Either node could be a DUP, but it's not worth doing both of them (you'd
7559 // just as well use the non-high version) so look for a corresponding extract
7560 // operation on the other "wing".
7561 if (isEssentiallyExtractSubvector(LHS)) {
7562 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
7565 } else if (isEssentiallyExtractSubvector(RHS)) {
7566 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
7571 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
7572 N->getOperand(0), LHS, RHS);
7575 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
7576 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
7577 unsigned ElemBits = ElemTy.getSizeInBits();
7579 int64_t ShiftAmount;
7580 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
7581 APInt SplatValue, SplatUndef;
7582 unsigned SplatBitSize;
7584 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
7585 HasAnyUndefs, ElemBits) ||
7586 SplatBitSize != ElemBits)
7589 ShiftAmount = SplatValue.getSExtValue();
7590 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
7591 ShiftAmount = CVN->getSExtValue();
7599 llvm_unreachable("Unknown shift intrinsic");
7600 case Intrinsic::aarch64_neon_sqshl:
7601 Opcode = AArch64ISD::SQSHL_I;
7602 IsRightShift = false;
7604 case Intrinsic::aarch64_neon_uqshl:
7605 Opcode = AArch64ISD::UQSHL_I;
7606 IsRightShift = false;
7608 case Intrinsic::aarch64_neon_srshl:
7609 Opcode = AArch64ISD::SRSHR_I;
7610 IsRightShift = true;
7612 case Intrinsic::aarch64_neon_urshl:
7613 Opcode = AArch64ISD::URSHR_I;
7614 IsRightShift = true;
7616 case Intrinsic::aarch64_neon_sqshlu:
7617 Opcode = AArch64ISD::SQSHLU_I;
7618 IsRightShift = false;
7622 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits)
7623 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
7624 DAG.getConstant(-ShiftAmount, MVT::i32));
7625 else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount < ElemBits)
7626 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
7627 DAG.getConstant(ShiftAmount, MVT::i32));
7632 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
7633 // the intrinsics must be legal and take an i32, this means there's almost
7634 // certainly going to be a zext in the DAG which we can eliminate.
7635 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
7636 SDValue AndN = N->getOperand(2);
7637 if (AndN.getOpcode() != ISD::AND)
7640 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
7641 if (!CMask || CMask->getZExtValue() != Mask)
7644 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
7645 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
7648 static SDValue performIntrinsicCombine(SDNode *N,
7649 TargetLowering::DAGCombinerInfo &DCI,
7650 const AArch64Subtarget *Subtarget) {
7651 SelectionDAG &DAG = DCI.DAG;
7652 unsigned IID = getIntrinsicID(N);
7656 case Intrinsic::aarch64_neon_vcvtfxs2fp:
7657 case Intrinsic::aarch64_neon_vcvtfxu2fp:
7658 return tryCombineFixedPointConvert(N, DCI, DAG);
7660 case Intrinsic::aarch64_neon_fmax:
7661 return DAG.getNode(AArch64ISD::FMAX, SDLoc(N), N->getValueType(0),
7662 N->getOperand(1), N->getOperand(2));
7663 case Intrinsic::aarch64_neon_fmin:
7664 return DAG.getNode(AArch64ISD::FMIN, SDLoc(N), N->getValueType(0),
7665 N->getOperand(1), N->getOperand(2));
7666 case Intrinsic::aarch64_neon_smull:
7667 case Intrinsic::aarch64_neon_umull:
7668 case Intrinsic::aarch64_neon_pmull:
7669 case Intrinsic::aarch64_neon_sqdmull:
7670 return tryCombineLongOpWithDup(IID, N, DCI, DAG);
7671 case Intrinsic::aarch64_neon_sqshl:
7672 case Intrinsic::aarch64_neon_uqshl:
7673 case Intrinsic::aarch64_neon_sqshlu:
7674 case Intrinsic::aarch64_neon_srshl:
7675 case Intrinsic::aarch64_neon_urshl:
7676 return tryCombineShiftImm(IID, N, DAG);
7677 case Intrinsic::aarch64_crc32b:
7678 case Intrinsic::aarch64_crc32cb:
7679 return tryCombineCRC32(0xff, N, DAG);
7680 case Intrinsic::aarch64_crc32h:
7681 case Intrinsic::aarch64_crc32ch:
7682 return tryCombineCRC32(0xffff, N, DAG);
7687 static SDValue performExtendCombine(SDNode *N,
7688 TargetLowering::DAGCombinerInfo &DCI,
7689 SelectionDAG &DAG) {
7690 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
7691 // we can convert that DUP into another extract_high (of a bigger DUP), which
7692 // helps the backend to decide that an sabdl2 would be useful, saving a real
7693 // extract_high operation.
7694 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
7695 N->getOperand(0).getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
7696 SDNode *ABDNode = N->getOperand(0).getNode();
7697 unsigned IID = getIntrinsicID(ABDNode);
7698 if (IID == Intrinsic::aarch64_neon_sabd ||
7699 IID == Intrinsic::aarch64_neon_uabd) {
7700 SDValue NewABD = tryCombineLongOpWithDup(IID, ABDNode, DCI, DAG);
7701 if (!NewABD.getNode())
7704 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
7709 // This is effectively a custom type legalization for AArch64.
7711 // Type legalization will split an extend of a small, legal, type to a larger
7712 // illegal type by first splitting the destination type, often creating
7713 // illegal source types, which then get legalized in isel-confusing ways,
7714 // leading to really terrible codegen. E.g.,
7715 // %result = v8i32 sext v8i8 %value
7717 // %losrc = extract_subreg %value, ...
7718 // %hisrc = extract_subreg %value, ...
7719 // %lo = v4i32 sext v4i8 %losrc
7720 // %hi = v4i32 sext v4i8 %hisrc
7721 // Things go rapidly downhill from there.
7723 // For AArch64, the [sz]ext vector instructions can only go up one element
7724 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
7725 // take two instructions.
7727 // This implies that the most efficient way to do the extend from v8i8
7728 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
7729 // the normal splitting to happen for the v8i16->v8i32.
7731 // This is pre-legalization to catch some cases where the default
7732 // type legalization will create ill-tempered code.
7733 if (!DCI.isBeforeLegalizeOps())
7736 // We're only interested in cleaning things up for non-legal vector types
7737 // here. If both the source and destination are legal, things will just
7738 // work naturally without any fiddling.
7739 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7740 EVT ResVT = N->getValueType(0);
7741 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
7743 // If the vector type isn't a simple VT, it's beyond the scope of what
7744 // we're worried about here. Let legalization do its thing and hope for
7746 SDValue Src = N->getOperand(0);
7747 EVT SrcVT = Src->getValueType(0);
7748 if (!ResVT.isSimple() || !SrcVT.isSimple())
7751 // If the source VT is a 64-bit vector, we can play games and get the
7752 // better results we want.
7753 if (SrcVT.getSizeInBits() != 64)
7756 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
7757 unsigned ElementCount = SrcVT.getVectorNumElements();
7758 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
7760 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
7762 // Now split the rest of the operation into two halves, each with a 64
7766 unsigned NumElements = ResVT.getVectorNumElements();
7767 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
7768 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
7769 ResVT.getVectorElementType(), NumElements / 2);
7771 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
7772 LoVT.getVectorNumElements());
7773 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7774 DAG.getConstant(0, MVT::i64));
7775 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7776 DAG.getConstant(InNVT.getVectorNumElements(), MVT::i64));
7777 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
7778 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
7780 // Now combine the parts back together so we still have a single result
7781 // like the combiner expects.
7782 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
7785 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
7786 /// value. The load store optimizer pass will merge them to store pair stores.
7787 /// This has better performance than a splat of the scalar followed by a split
7788 /// vector store. Even if the stores are not merged it is four stores vs a dup,
7789 /// followed by an ext.b and two stores.
7790 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
7791 SDValue StVal = St->getValue();
7792 EVT VT = StVal.getValueType();
7794 // Don't replace floating point stores, they possibly won't be transformed to
7795 // stp because of the store pair suppress pass.
7796 if (VT.isFloatingPoint())
7799 // Check for insert vector elements.
7800 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
7803 // We can express a splat as store pair(s) for 2 or 4 elements.
7804 unsigned NumVecElts = VT.getVectorNumElements();
7805 if (NumVecElts != 4 && NumVecElts != 2)
7807 SDValue SplatVal = StVal.getOperand(1);
7808 unsigned RemainInsertElts = NumVecElts - 1;
7810 // Check that this is a splat.
7811 while (--RemainInsertElts) {
7812 SDValue NextInsertElt = StVal.getOperand(0);
7813 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
7815 if (NextInsertElt.getOperand(1) != SplatVal)
7817 StVal = NextInsertElt;
7819 unsigned OrigAlignment = St->getAlignment();
7820 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
7821 unsigned Alignment = std::min(OrigAlignment, EltOffset);
7823 // Create scalar stores. This is at least as good as the code sequence for a
7824 // split unaligned store wich is a dup.s, ext.b, and two stores.
7825 // Most of the time the three stores should be replaced by store pair
7826 // instructions (stp).
7828 SDValue BasePtr = St->getBasePtr();
7830 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
7831 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
7833 unsigned Offset = EltOffset;
7834 while (--NumVecElts) {
7835 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7836 DAG.getConstant(Offset, MVT::i64));
7837 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
7838 St->getPointerInfo(), St->isVolatile(),
7839 St->isNonTemporal(), Alignment);
7840 Offset += EltOffset;
7845 static SDValue performSTORECombine(SDNode *N,
7846 TargetLowering::DAGCombinerInfo &DCI,
7848 const AArch64Subtarget *Subtarget) {
7849 if (!DCI.isBeforeLegalize())
7852 StoreSDNode *S = cast<StoreSDNode>(N);
7853 if (S->isVolatile())
7856 // Cyclone has bad performance on unaligned 16B stores when crossing line and
7857 // page boundries. We want to split such stores.
7858 if (!Subtarget->isCyclone())
7861 // Don't split at Oz.
7862 MachineFunction &MF = DAG.getMachineFunction();
7863 bool IsMinSize = MF.getFunction()->getAttributes().hasAttribute(
7864 AttributeSet::FunctionIndex, Attribute::MinSize);
7868 SDValue StVal = S->getValue();
7869 EVT VT = StVal.getValueType();
7871 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
7872 // those up regresses performance on micro-benchmarks and olden/bh.
7873 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
7876 // Split unaligned 16B stores. They are terrible for performance.
7877 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
7878 // extensions can use this to mark that it does not want splitting to happen
7879 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
7880 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
7881 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
7882 S->getAlignment() <= 2)
7885 // If we get a splat of a scalar convert this vector store to a store of
7886 // scalars. They will be merged into store pairs thereby removing two
7888 SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S);
7889 if (ReplacedSplat != SDValue())
7890 return ReplacedSplat;
7893 unsigned NumElts = VT.getVectorNumElements() / 2;
7894 // Split VT into two.
7896 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
7897 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7898 DAG.getConstant(0, MVT::i64));
7899 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7900 DAG.getConstant(NumElts, MVT::i64));
7901 SDValue BasePtr = S->getBasePtr();
7903 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
7904 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
7905 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7906 DAG.getConstant(8, MVT::i64));
7907 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
7908 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
7912 /// Target-specific DAG combine function for post-increment LD1 (lane) and
7913 /// post-increment LD1R.
7914 static SDValue performPostLD1Combine(SDNode *N,
7915 TargetLowering::DAGCombinerInfo &DCI,
7917 if (DCI.isBeforeLegalizeOps())
7920 SelectionDAG &DAG = DCI.DAG;
7921 EVT VT = N->getValueType(0);
7923 unsigned LoadIdx = IsLaneOp ? 1 : 0;
7924 SDNode *LD = N->getOperand(LoadIdx).getNode();
7925 // If it is not LOAD, can not do such combine.
7926 if (LD->getOpcode() != ISD::LOAD)
7929 LoadSDNode *LoadSDN = cast<LoadSDNode>(LD);
7930 EVT MemVT = LoadSDN->getMemoryVT();
7931 // Check if memory operand is the same type as the vector element.
7932 if (MemVT != VT.getVectorElementType())
7935 // Check if there are other uses. If so, do not combine as it will introduce
7937 for (SDNode::use_iterator UI = LD->use_begin(), UE = LD->use_end(); UI != UE;
7939 if (UI.getUse().getResNo() == 1) // Ignore uses of the chain result.
7945 SDValue Addr = LD->getOperand(1);
7946 SDValue Vector = N->getOperand(0);
7947 // Search for a use of the address operand that is an increment.
7948 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), UE =
7949 Addr.getNode()->use_end(); UI != UE; ++UI) {
7951 if (User->getOpcode() != ISD::ADD
7952 || UI.getUse().getResNo() != Addr.getResNo())
7955 // Check that the add is independent of the load. Otherwise, folding it
7956 // would create a cycle.
7957 if (User->isPredecessorOf(LD) || LD->isPredecessorOf(User))
7959 // Also check that add is not used in the vector operand. This would also
7961 if (User->isPredecessorOf(Vector.getNode()))
7964 // If the increment is a constant, it must match the memory ref size.
7965 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7966 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7967 uint32_t IncVal = CInc->getZExtValue();
7968 unsigned NumBytes = VT.getScalarSizeInBits() / 8;
7969 if (IncVal != NumBytes)
7971 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
7974 SmallVector<SDValue, 8> Ops;
7975 Ops.push_back(LD->getOperand(0)); // Chain
7977 Ops.push_back(Vector); // The vector to be inserted
7978 Ops.push_back(N->getOperand(2)); // The lane to be inserted in the vector
7980 Ops.push_back(Addr);
7983 EVT Tys[3] = { VT, MVT::i64, MVT::Other };
7984 SDVTList SDTys = DAG.getVTList(Tys);
7985 unsigned NewOp = IsLaneOp ? AArch64ISD::LD1LANEpost : AArch64ISD::LD1DUPpost;
7986 SDValue UpdN = DAG.getMemIntrinsicNode(NewOp, SDLoc(N), SDTys, Ops,
7988 LoadSDN->getMemOperand());
7991 std::vector<SDValue> NewResults;
7992 NewResults.push_back(SDValue(LD, 0)); // The result of load
7993 NewResults.push_back(SDValue(UpdN.getNode(), 2)); // Chain
7994 DCI.CombineTo(LD, NewResults);
7995 DCI.CombineTo(N, SDValue(UpdN.getNode(), 0)); // Dup/Inserted Result
7996 DCI.CombineTo(User, SDValue(UpdN.getNode(), 1)); // Write back register
8003 /// Target-specific DAG combine function for NEON load/store intrinsics
8004 /// to merge base address updates.
8005 static SDValue performNEONPostLDSTCombine(SDNode *N,
8006 TargetLowering::DAGCombinerInfo &DCI,
8007 SelectionDAG &DAG) {
8008 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8011 unsigned AddrOpIdx = N->getNumOperands() - 1;
8012 SDValue Addr = N->getOperand(AddrOpIdx);
8014 // Search for a use of the address operand that is an increment.
8015 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8016 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8018 if (User->getOpcode() != ISD::ADD ||
8019 UI.getUse().getResNo() != Addr.getResNo())
8022 // Check that the add is independent of the load/store. Otherwise, folding
8023 // it would create a cycle.
8024 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8027 // Find the new opcode for the updating load/store.
8028 bool IsStore = false;
8029 bool IsLaneOp = false;
8030 bool IsDupOp = false;
8031 unsigned NewOpc = 0;
8032 unsigned NumVecs = 0;
8033 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8035 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8036 case Intrinsic::aarch64_neon_ld2: NewOpc = AArch64ISD::LD2post;
8038 case Intrinsic::aarch64_neon_ld3: NewOpc = AArch64ISD::LD3post;
8040 case Intrinsic::aarch64_neon_ld4: NewOpc = AArch64ISD::LD4post;
8042 case Intrinsic::aarch64_neon_st2: NewOpc = AArch64ISD::ST2post;
8043 NumVecs = 2; IsStore = true; break;
8044 case Intrinsic::aarch64_neon_st3: NewOpc = AArch64ISD::ST3post;
8045 NumVecs = 3; IsStore = true; break;
8046 case Intrinsic::aarch64_neon_st4: NewOpc = AArch64ISD::ST4post;
8047 NumVecs = 4; IsStore = true; break;
8048 case Intrinsic::aarch64_neon_ld1x2: NewOpc = AArch64ISD::LD1x2post;
8050 case Intrinsic::aarch64_neon_ld1x3: NewOpc = AArch64ISD::LD1x3post;
8052 case Intrinsic::aarch64_neon_ld1x4: NewOpc = AArch64ISD::LD1x4post;
8054 case Intrinsic::aarch64_neon_st1x2: NewOpc = AArch64ISD::ST1x2post;
8055 NumVecs = 2; IsStore = true; break;
8056 case Intrinsic::aarch64_neon_st1x3: NewOpc = AArch64ISD::ST1x3post;
8057 NumVecs = 3; IsStore = true; break;
8058 case Intrinsic::aarch64_neon_st1x4: NewOpc = AArch64ISD::ST1x4post;
8059 NumVecs = 4; IsStore = true; break;
8060 case Intrinsic::aarch64_neon_ld2r: NewOpc = AArch64ISD::LD2DUPpost;
8061 NumVecs = 2; IsDupOp = true; break;
8062 case Intrinsic::aarch64_neon_ld3r: NewOpc = AArch64ISD::LD3DUPpost;
8063 NumVecs = 3; IsDupOp = true; break;
8064 case Intrinsic::aarch64_neon_ld4r: NewOpc = AArch64ISD::LD4DUPpost;
8065 NumVecs = 4; IsDupOp = true; break;
8066 case Intrinsic::aarch64_neon_ld2lane: NewOpc = AArch64ISD::LD2LANEpost;
8067 NumVecs = 2; IsLaneOp = true; break;
8068 case Intrinsic::aarch64_neon_ld3lane: NewOpc = AArch64ISD::LD3LANEpost;
8069 NumVecs = 3; IsLaneOp = true; break;
8070 case Intrinsic::aarch64_neon_ld4lane: NewOpc = AArch64ISD::LD4LANEpost;
8071 NumVecs = 4; IsLaneOp = true; break;
8072 case Intrinsic::aarch64_neon_st2lane: NewOpc = AArch64ISD::ST2LANEpost;
8073 NumVecs = 2; IsStore = true; IsLaneOp = true; break;
8074 case Intrinsic::aarch64_neon_st3lane: NewOpc = AArch64ISD::ST3LANEpost;
8075 NumVecs = 3; IsStore = true; IsLaneOp = true; break;
8076 case Intrinsic::aarch64_neon_st4lane: NewOpc = AArch64ISD::ST4LANEpost;
8077 NumVecs = 4; IsStore = true; IsLaneOp = true; break;
8082 VecTy = N->getOperand(2).getValueType();
8084 VecTy = N->getValueType(0);
8086 // If the increment is a constant, it must match the memory ref size.
8087 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8088 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8089 uint32_t IncVal = CInc->getZExtValue();
8090 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8091 if (IsLaneOp || IsDupOp)
8092 NumBytes /= VecTy.getVectorNumElements();
8093 if (IncVal != NumBytes)
8095 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
8097 SmallVector<SDValue, 8> Ops;
8098 Ops.push_back(N->getOperand(0)); // Incoming chain
8099 // Load lane and store have vector list as input.
8100 if (IsLaneOp || IsStore)
8101 for (unsigned i = 2; i < AddrOpIdx; ++i)
8102 Ops.push_back(N->getOperand(i));
8103 Ops.push_back(Addr); // Base register
8108 unsigned NumResultVecs = (IsStore ? 0 : NumVecs);
8110 for (n = 0; n < NumResultVecs; ++n)
8112 Tys[n++] = MVT::i64; // Type of write back register
8113 Tys[n] = MVT::Other; // Type of the chain
8114 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs + 2));
8116 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8117 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys, Ops,
8118 MemInt->getMemoryVT(),
8119 MemInt->getMemOperand());
8122 std::vector<SDValue> NewResults;
8123 for (unsigned i = 0; i < NumResultVecs; ++i) {
8124 NewResults.push_back(SDValue(UpdN.getNode(), i));
8126 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1));
8127 DCI.CombineTo(N, NewResults);
8128 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8135 // Checks to see if the value is the prescribed width and returns information
8136 // about its extension mode.
8138 bool checkValueWidth(SDValue V, unsigned width, ISD::LoadExtType &ExtType) {
8139 ExtType = ISD::NON_EXTLOAD;
8140 switch(V.getNode()->getOpcode()) {
8144 LoadSDNode *LoadNode = cast<LoadSDNode>(V.getNode());
8145 if ((LoadNode->getMemoryVT() == MVT::i8 && width == 8)
8146 || (LoadNode->getMemoryVT() == MVT::i16 && width == 16)) {
8147 ExtType = LoadNode->getExtensionType();
8152 case ISD::AssertSext: {
8153 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
8154 if ((TypeNode->getVT() == MVT::i8 && width == 8)
8155 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
8156 ExtType = ISD::SEXTLOAD;
8161 case ISD::AssertZext: {
8162 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
8163 if ((TypeNode->getVT() == MVT::i8 && width == 8)
8164 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
8165 ExtType = ISD::ZEXTLOAD;
8171 case ISD::TargetConstant: {
8172 if (std::abs(cast<ConstantSDNode>(V.getNode())->getSExtValue()) <
8182 // This function does a whole lot of voodoo to determine if the tests are
8183 // equivalent without and with a mask. Essentially what happens is that given a
8186 // +-------------+ +-------------+ +-------------+ +-------------+
8187 // | Input | | AddConstant | | CompConstant| | CC |
8188 // +-------------+ +-------------+ +-------------+ +-------------+
8190 // V V | +----------+
8191 // +-------------+ +----+ | |
8192 // | ADD | |0xff| | |
8193 // +-------------+ +----+ | |
8196 // +-------------+ | |
8198 // +-------------+ | |
8207 // The AND node may be safely removed for some combinations of inputs. In
8208 // particular we need to take into account the extension type of the Input,
8209 // the exact values of AddConstant, CompConstant, and CC, along with the nominal
8210 // width of the input (this can work for any width inputs, the above graph is
8211 // specific to 8 bits.
8213 // The specific equations were worked out by generating output tables for each
8214 // AArch64CC value in terms of and AddConstant (w1), CompConstant(w2). The
8215 // problem was simplified by working with 4 bit inputs, which means we only
8216 // needed to reason about 24 distinct bit patterns: 8 patterns unique to zero
8217 // extension (8,15), 8 patterns unique to sign extensions (-8,-1), and 8
8218 // patterns present in both extensions (0,7). For every distinct set of
8219 // AddConstant and CompConstants bit patterns we can consider the masked and
8220 // unmasked versions to be equivalent if the result of this function is true for
8221 // all 16 distinct bit patterns of for the current extension type of Input (w0).
8224 // and w10, w8, #0x0f
8226 // cset w9, AArch64CC
8228 // cset w11, AArch64CC
8233 // Since the above function shows when the outputs are equivalent it defines
8234 // when it is safe to remove the AND. Unfortunately it only runs on AArch64 and
8235 // would be expensive to run during compiles. The equations below were written
8236 // in a test harness that confirmed they gave equivalent outputs to the above
8237 // for all inputs function, so they can be used determine if the removal is
8240 // isEquivalentMaskless() is the code for testing if the AND can be removed
8241 // factored out of the DAG recognition as the DAG can take several forms.
8244 bool isEquivalentMaskless(unsigned CC, unsigned width,
8245 ISD::LoadExtType ExtType, signed AddConstant,
8246 signed CompConstant) {
8247 // By being careful about our equations and only writing the in term
8248 // symbolic values and well known constants (0, 1, -1, MaxUInt) we can
8249 // make them generally applicable to all bit widths.
8250 signed MaxUInt = (1 << width);
8252 // For the purposes of these comparisons sign extending the type is
8253 // equivalent to zero extending the add and displacing it by half the integer
8254 // width. Provided we are careful and make sure our equations are valid over
8255 // the whole range we can just adjust the input and avoid writing equations
8256 // for sign extended inputs.
8257 if (ExtType == ISD::SEXTLOAD)
8258 AddConstant -= (1 << (width-1));
8262 case AArch64CC::GT: {
8263 if ((AddConstant == 0) ||
8264 (CompConstant == MaxUInt - 1 && AddConstant < 0) ||
8265 (AddConstant >= 0 && CompConstant < 0) ||
8266 (AddConstant <= 0 && CompConstant <= 0 && CompConstant < AddConstant))
8270 case AArch64CC::GE: {
8271 if ((AddConstant == 0) ||
8272 (AddConstant >= 0 && CompConstant <= 0) ||
8273 (AddConstant <= 0 && CompConstant <= 0 && CompConstant <= AddConstant))
8277 case AArch64CC::LS: {
8278 if ((AddConstant >= 0 && CompConstant < 0) ||
8279 (AddConstant <= 0 && CompConstant >= -1 &&
8280 CompConstant < AddConstant + MaxUInt))
8284 case AArch64CC::MI: {
8285 if ((AddConstant == 0) ||
8286 (AddConstant > 0 && CompConstant <= 0) ||
8287 (AddConstant < 0 && CompConstant <= AddConstant))
8291 case AArch64CC::HS: {
8292 if ((AddConstant >= 0 && CompConstant <= 0) ||
8293 (AddConstant <= 0 && CompConstant >= 0 &&
8294 CompConstant <= AddConstant + MaxUInt))
8298 case AArch64CC::NE: {
8299 if ((AddConstant > 0 && CompConstant < 0) ||
8300 (AddConstant < 0 && CompConstant >= 0 &&
8301 CompConstant < AddConstant + MaxUInt) ||
8302 (AddConstant >= 0 && CompConstant >= 0 &&
8303 CompConstant >= AddConstant) ||
8304 (AddConstant <= 0 && CompConstant < 0 && CompConstant < AddConstant))
8313 case AArch64CC::Invalid:
8321 SDValue performCONDCombine(SDNode *N,
8322 TargetLowering::DAGCombinerInfo &DCI,
8323 SelectionDAG &DAG, unsigned CCIndex,
8324 unsigned CmpIndex) {
8325 unsigned CC = cast<ConstantSDNode>(N->getOperand(CCIndex))->getSExtValue();
8326 SDNode *SubsNode = N->getOperand(CmpIndex).getNode();
8327 unsigned CondOpcode = SubsNode->getOpcode();
8329 if (CondOpcode != AArch64ISD::SUBS)
8332 // There is a SUBS feeding this condition. Is it fed by a mask we can
8335 SDNode *AndNode = SubsNode->getOperand(0).getNode();
8336 unsigned MaskBits = 0;
8338 if (AndNode->getOpcode() != ISD::AND)
8341 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(AndNode->getOperand(1))) {
8342 uint32_t CNV = CN->getZExtValue();
8345 else if (CNV == 65535)
8352 SDValue AddValue = AndNode->getOperand(0);
8354 if (AddValue.getOpcode() != ISD::ADD)
8357 // The basic dag structure is correct, grab the inputs and validate them.
8359 SDValue AddInputValue1 = AddValue.getNode()->getOperand(0);
8360 SDValue AddInputValue2 = AddValue.getNode()->getOperand(1);
8361 SDValue SubsInputValue = SubsNode->getOperand(1);
8363 // The mask is present and the provenance of all the values is a smaller type,
8364 // lets see if the mask is superfluous.
8366 if (!isa<ConstantSDNode>(AddInputValue2.getNode()) ||
8367 !isa<ConstantSDNode>(SubsInputValue.getNode()))
8370 ISD::LoadExtType ExtType;
8372 if (!checkValueWidth(SubsInputValue, MaskBits, ExtType) ||
8373 !checkValueWidth(AddInputValue2, MaskBits, ExtType) ||
8374 !checkValueWidth(AddInputValue1, MaskBits, ExtType) )
8377 if(!isEquivalentMaskless(CC, MaskBits, ExtType,
8378 cast<ConstantSDNode>(AddInputValue2.getNode())->getSExtValue(),
8379 cast<ConstantSDNode>(SubsInputValue.getNode())->getSExtValue()))
8382 // The AND is not necessary, remove it.
8384 SDVTList VTs = DAG.getVTList(SubsNode->getValueType(0),
8385 SubsNode->getValueType(1));
8386 SDValue Ops[] = { AddValue, SubsNode->getOperand(1) };
8388 SDValue NewValue = DAG.getNode(CondOpcode, SDLoc(SubsNode), VTs, Ops);
8389 DAG.ReplaceAllUsesWith(SubsNode, NewValue.getNode());
8391 return SDValue(N, 0);
8394 // Optimize compare with zero and branch.
8395 static SDValue performBRCONDCombine(SDNode *N,
8396 TargetLowering::DAGCombinerInfo &DCI,
8397 SelectionDAG &DAG) {
8398 SDValue NV = performCONDCombine(N, DCI, DAG, 2, 3);
8401 SDValue Chain = N->getOperand(0);
8402 SDValue Dest = N->getOperand(1);
8403 SDValue CCVal = N->getOperand(2);
8404 SDValue Cmp = N->getOperand(3);
8406 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
8407 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
8408 if (CC != AArch64CC::EQ && CC != AArch64CC::NE)
8411 unsigned CmpOpc = Cmp.getOpcode();
8412 if (CmpOpc != AArch64ISD::ADDS && CmpOpc != AArch64ISD::SUBS)
8415 // Only attempt folding if there is only one use of the flag and no use of the
8417 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
8420 SDValue LHS = Cmp.getOperand(0);
8421 SDValue RHS = Cmp.getOperand(1);
8423 assert(LHS.getValueType() == RHS.getValueType() &&
8424 "Expected the value type to be the same for both operands!");
8425 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
8428 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
8429 std::swap(LHS, RHS);
8431 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
8434 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
8435 LHS.getOpcode() == ISD::SRL)
8438 // Fold the compare into the branch instruction.
8440 if (CC == AArch64CC::EQ)
8441 BR = DAG.getNode(AArch64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
8443 BR = DAG.getNode(AArch64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
8445 // Do not add new nodes to DAG combiner worklist.
8446 DCI.CombineTo(N, BR, false);
8451 // vselect (v1i1 setcc) ->
8452 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
8453 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
8454 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
8456 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
8457 SDValue N0 = N->getOperand(0);
8458 EVT CCVT = N0.getValueType();
8460 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
8461 CCVT.getVectorElementType() != MVT::i1)
8464 EVT ResVT = N->getValueType(0);
8465 EVT CmpVT = N0.getOperand(0).getValueType();
8466 // Only combine when the result type is of the same size as the compared
8468 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
8471 SDValue IfTrue = N->getOperand(1);
8472 SDValue IfFalse = N->getOperand(2);
8474 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
8475 N0.getOperand(0), N0.getOperand(1),
8476 cast<CondCodeSDNode>(N0.getOperand(2))->get());
8477 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
8481 /// A vector select: "(select vL, vR, (setcc LHS, RHS))" is best performed with
8482 /// the compare-mask instructions rather than going via NZCV, even if LHS and
8483 /// RHS are really scalar. This replaces any scalar setcc in the above pattern
8484 /// with a vector one followed by a DUP shuffle on the result.
8485 static SDValue performSelectCombine(SDNode *N, SelectionDAG &DAG) {
8486 SDValue N0 = N->getOperand(0);
8487 EVT ResVT = N->getValueType(0);
8489 if (N0.getOpcode() != ISD::SETCC || N0.getValueType() != MVT::i1)
8492 // If NumMaskElts == 0, the comparison is larger than select result. The
8493 // largest real NEON comparison is 64-bits per lane, which means the result is
8494 // at most 32-bits and an illegal vector. Just bail out for now.
8495 EVT SrcVT = N0.getOperand(0).getValueType();
8497 // Don't try to do this optimization when the setcc itself has i1 operands.
8498 // There are no legal vectors of i1, so this would be pointless.
8499 if (SrcVT == MVT::i1)
8502 int NumMaskElts = ResVT.getSizeInBits() / SrcVT.getSizeInBits();
8503 if (!ResVT.isVector() || NumMaskElts == 0)
8506 SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumMaskElts);
8507 EVT CCVT = SrcVT.changeVectorElementTypeToInteger();
8509 // First perform a vector comparison, where lane 0 is the one we're interested
8513 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(0));
8515 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(1));
8516 SDValue SetCC = DAG.getNode(ISD::SETCC, DL, CCVT, LHS, RHS, N0.getOperand(2));
8518 // Now duplicate the comparison mask we want across all other lanes.
8519 SmallVector<int, 8> DUPMask(CCVT.getVectorNumElements(), 0);
8520 SDValue Mask = DAG.getVectorShuffle(CCVT, DL, SetCC, SetCC, DUPMask.data());
8521 Mask = DAG.getNode(ISD::BITCAST, DL,
8522 ResVT.changeVectorElementTypeToInteger(), Mask);
8524 return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2));
8527 SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N,
8528 DAGCombinerInfo &DCI) const {
8529 SelectionDAG &DAG = DCI.DAG;
8530 switch (N->getOpcode()) {
8535 return performAddSubLongCombine(N, DCI, DAG);
8537 return performXorCombine(N, DAG, DCI, Subtarget);
8539 return performMulCombine(N, DAG, DCI, Subtarget);
8540 case ISD::SINT_TO_FP:
8541 case ISD::UINT_TO_FP:
8542 return performIntToFpCombine(N, DAG, Subtarget);
8544 return performORCombine(N, DCI, Subtarget);
8545 case ISD::INTRINSIC_WO_CHAIN:
8546 return performIntrinsicCombine(N, DCI, Subtarget);
8547 case ISD::ANY_EXTEND:
8548 case ISD::ZERO_EXTEND:
8549 case ISD::SIGN_EXTEND:
8550 return performExtendCombine(N, DCI, DAG);
8552 return performBitcastCombine(N, DCI, DAG);
8553 case ISD::CONCAT_VECTORS:
8554 return performConcatVectorsCombine(N, DCI, DAG);
8556 return performSelectCombine(N, DAG);
8558 return performVSelectCombine(N, DCI.DAG);
8560 return performSTORECombine(N, DCI, DAG, Subtarget);
8561 case AArch64ISD::BRCOND:
8562 return performBRCONDCombine(N, DCI, DAG);
8563 case AArch64ISD::CSEL:
8564 return performCONDCombine(N, DCI, DAG, 2, 3);
8565 case AArch64ISD::DUP:
8566 return performPostLD1Combine(N, DCI, false);
8567 case ISD::INSERT_VECTOR_ELT:
8568 return performPostLD1Combine(N, DCI, true);
8569 case ISD::INTRINSIC_VOID:
8570 case ISD::INTRINSIC_W_CHAIN:
8571 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8572 case Intrinsic::aarch64_neon_ld2:
8573 case Intrinsic::aarch64_neon_ld3:
8574 case Intrinsic::aarch64_neon_ld4:
8575 case Intrinsic::aarch64_neon_ld1x2:
8576 case Intrinsic::aarch64_neon_ld1x3:
8577 case Intrinsic::aarch64_neon_ld1x4:
8578 case Intrinsic::aarch64_neon_ld2lane:
8579 case Intrinsic::aarch64_neon_ld3lane:
8580 case Intrinsic::aarch64_neon_ld4lane:
8581 case Intrinsic::aarch64_neon_ld2r:
8582 case Intrinsic::aarch64_neon_ld3r:
8583 case Intrinsic::aarch64_neon_ld4r:
8584 case Intrinsic::aarch64_neon_st2:
8585 case Intrinsic::aarch64_neon_st3:
8586 case Intrinsic::aarch64_neon_st4:
8587 case Intrinsic::aarch64_neon_st1x2:
8588 case Intrinsic::aarch64_neon_st1x3:
8589 case Intrinsic::aarch64_neon_st1x4:
8590 case Intrinsic::aarch64_neon_st2lane:
8591 case Intrinsic::aarch64_neon_st3lane:
8592 case Intrinsic::aarch64_neon_st4lane:
8593 return performNEONPostLDSTCombine(N, DCI, DAG);
8601 // Check if the return value is used as only a return value, as otherwise
8602 // we can't perform a tail-call. In particular, we need to check for
8603 // target ISD nodes that are returns and any other "odd" constructs
8604 // that the generic analysis code won't necessarily catch.
8605 bool AArch64TargetLowering::isUsedByReturnOnly(SDNode *N,
8606 SDValue &Chain) const {
8607 if (N->getNumValues() != 1)
8609 if (!N->hasNUsesOfValue(1, 0))
8612 SDValue TCChain = Chain;
8613 SDNode *Copy = *N->use_begin();
8614 if (Copy->getOpcode() == ISD::CopyToReg) {
8615 // If the copy has a glue operand, we conservatively assume it isn't safe to
8616 // perform a tail call.
8617 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
8620 TCChain = Copy->getOperand(0);
8621 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
8624 bool HasRet = false;
8625 for (SDNode *Node : Copy->uses()) {
8626 if (Node->getOpcode() != AArch64ISD::RET_FLAG)
8638 // Return whether the an instruction can potentially be optimized to a tail
8639 // call. This will cause the optimizers to attempt to move, or duplicate,
8640 // return instructions to help enable tail call optimizations for this
8642 bool AArch64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
8643 if (!CI->isTailCall())
8649 bool AArch64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
8651 ISD::MemIndexedMode &AM,
8653 SelectionDAG &DAG) const {
8654 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
8657 Base = Op->getOperand(0);
8658 // All of the indexed addressing mode instructions take a signed
8659 // 9 bit immediate offset.
8660 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
8661 int64_t RHSC = (int64_t)RHS->getZExtValue();
8662 if (RHSC >= 256 || RHSC <= -256)
8664 IsInc = (Op->getOpcode() == ISD::ADD);
8665 Offset = Op->getOperand(1);
8671 bool AArch64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
8673 ISD::MemIndexedMode &AM,
8674 SelectionDAG &DAG) const {
8677 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8678 VT = LD->getMemoryVT();
8679 Ptr = LD->getBasePtr();
8680 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8681 VT = ST->getMemoryVT();
8682 Ptr = ST->getBasePtr();
8687 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
8689 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
8693 bool AArch64TargetLowering::getPostIndexedAddressParts(
8694 SDNode *N, SDNode *Op, SDValue &Base, SDValue &Offset,
8695 ISD::MemIndexedMode &AM, SelectionDAG &DAG) const {
8698 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8699 VT = LD->getMemoryVT();
8700 Ptr = LD->getBasePtr();
8701 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8702 VT = ST->getMemoryVT();
8703 Ptr = ST->getBasePtr();
8708 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
8710 // Post-indexing updates the base, so it's not a valid transform
8711 // if that's not the same as the load's pointer.
8714 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
8718 static void ReplaceBITCASTResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
8719 SelectionDAG &DAG) {
8721 SDValue Op = N->getOperand(0);
8723 if (N->getValueType(0) != MVT::i16 || Op.getValueType() != MVT::f16)
8727 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL, MVT::f32,
8728 DAG.getUNDEF(MVT::i32), Op,
8729 DAG.getTargetConstant(AArch64::hsub, MVT::i32)),
8731 Op = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op);
8732 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Op));
8735 void AArch64TargetLowering::ReplaceNodeResults(
8736 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
8737 switch (N->getOpcode()) {
8739 llvm_unreachable("Don't know how to custom expand this");
8741 ReplaceBITCASTResults(N, Results, DAG);
8743 case ISD::FP_TO_UINT:
8744 case ISD::FP_TO_SINT:
8745 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
8746 // Let normal code take care of it by not adding anything to Results.
8751 bool AArch64TargetLowering::useLoadStackGuardNode() const {
8755 bool AArch64TargetLowering::combineRepeatedFPDivisors(unsigned NumUsers) const {
8756 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
8757 // reciprocal if there are three or more FDIVs.
8758 return NumUsers > 2;
8761 TargetLoweringBase::LegalizeTypeAction
8762 AArch64TargetLowering::getPreferredVectorAction(EVT VT) const {
8763 MVT SVT = VT.getSimpleVT();
8764 // During type legalization, we prefer to widen v1i8, v1i16, v1i32 to v8i8,
8765 // v4i16, v2i32 instead of to promote.
8766 if (SVT == MVT::v1i8 || SVT == MVT::v1i16 || SVT == MVT::v1i32
8767 || SVT == MVT::v1f32)
8768 return TypeWidenVector;
8770 return TargetLoweringBase::getPreferredVectorAction(VT);
8773 // Loads and stores less than 128-bits are already atomic; ones above that
8774 // are doomed anyway, so defer to the default libcall and blame the OS when
8776 bool AArch64TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
8777 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
8781 // Loads and stores less than 128-bits are already atomic; ones above that
8782 // are doomed anyway, so defer to the default libcall and blame the OS when
8784 bool AArch64TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
8785 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
8789 // For the real atomic operations, we have ldxr/stxr up to 128 bits,
8790 bool AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
8791 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
8795 bool AArch64TargetLowering::hasLoadLinkedStoreConditional() const {
8799 Value *AArch64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
8800 AtomicOrdering Ord) const {
8801 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
8802 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
8803 bool IsAcquire = isAtLeastAcquire(Ord);
8805 // Since i128 isn't legal and intrinsics don't get type-lowered, the ldrexd
8806 // intrinsic must return {i64, i64} and we have to recombine them into a
8807 // single i128 here.
8808 if (ValTy->getPrimitiveSizeInBits() == 128) {
8810 IsAcquire ? Intrinsic::aarch64_ldaxp : Intrinsic::aarch64_ldxp;
8811 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int);
8813 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
8814 Value *LoHi = Builder.CreateCall(Ldxr, Addr, "lohi");
8816 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
8817 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
8818 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
8819 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
8820 return Builder.CreateOr(
8821 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64");
8824 Type *Tys[] = { Addr->getType() };
8826 IsAcquire ? Intrinsic::aarch64_ldaxr : Intrinsic::aarch64_ldxr;
8827 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int, Tys);
8829 return Builder.CreateTruncOrBitCast(
8830 Builder.CreateCall(Ldxr, Addr),
8831 cast<PointerType>(Addr->getType())->getElementType());
8834 Value *AArch64TargetLowering::emitStoreConditional(IRBuilder<> &Builder,
8835 Value *Val, Value *Addr,
8836 AtomicOrdering Ord) const {
8837 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
8838 bool IsRelease = isAtLeastRelease(Ord);
8840 // Since the intrinsics must have legal type, the i128 intrinsics take two
8841 // parameters: "i64, i64". We must marshal Val into the appropriate form
8843 if (Val->getType()->getPrimitiveSizeInBits() == 128) {
8845 IsRelease ? Intrinsic::aarch64_stlxp : Intrinsic::aarch64_stxp;
8846 Function *Stxr = Intrinsic::getDeclaration(M, Int);
8847 Type *Int64Ty = Type::getInt64Ty(M->getContext());
8849 Value *Lo = Builder.CreateTrunc(Val, Int64Ty, "lo");
8850 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 64), Int64Ty, "hi");
8851 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
8852 return Builder.CreateCall3(Stxr, Lo, Hi, Addr);
8856 IsRelease ? Intrinsic::aarch64_stlxr : Intrinsic::aarch64_stxr;
8857 Type *Tys[] = { Addr->getType() };
8858 Function *Stxr = Intrinsic::getDeclaration(M, Int, Tys);
8860 return Builder.CreateCall2(
8861 Stxr, Builder.CreateZExtOrBitCast(
8862 Val, Stxr->getFunctionType()->getParamType(0)),
8866 bool AArch64TargetLowering::functionArgumentNeedsConsecutiveRegisters(
8867 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
8868 return Ty->isArrayTy();