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"),
70 AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
71 const AArch64Subtarget &STI)
72 : TargetLowering(TM), Subtarget(&STI) {
74 // AArch64 doesn't have comparisons which set GPRs or setcc instructions, so
75 // we have to make something up. Arbitrarily, choose ZeroOrOne.
76 setBooleanContents(ZeroOrOneBooleanContent);
77 // When comparing vectors the result sets the different elements in the
78 // vector to all-one or all-zero.
79 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
81 // Set up the register classes.
82 addRegisterClass(MVT::i32, &AArch64::GPR32allRegClass);
83 addRegisterClass(MVT::i64, &AArch64::GPR64allRegClass);
85 if (Subtarget->hasFPARMv8()) {
86 addRegisterClass(MVT::f16, &AArch64::FPR16RegClass);
87 addRegisterClass(MVT::f32, &AArch64::FPR32RegClass);
88 addRegisterClass(MVT::f64, &AArch64::FPR64RegClass);
89 addRegisterClass(MVT::f128, &AArch64::FPR128RegClass);
92 if (Subtarget->hasNEON()) {
93 addRegisterClass(MVT::v16i8, &AArch64::FPR8RegClass);
94 addRegisterClass(MVT::v8i16, &AArch64::FPR16RegClass);
95 // Someone set us up the NEON.
96 addDRTypeForNEON(MVT::v2f32);
97 addDRTypeForNEON(MVT::v8i8);
98 addDRTypeForNEON(MVT::v4i16);
99 addDRTypeForNEON(MVT::v2i32);
100 addDRTypeForNEON(MVT::v1i64);
101 addDRTypeForNEON(MVT::v1f64);
102 addDRTypeForNEON(MVT::v4f16);
104 addQRTypeForNEON(MVT::v4f32);
105 addQRTypeForNEON(MVT::v2f64);
106 addQRTypeForNEON(MVT::v16i8);
107 addQRTypeForNEON(MVT::v8i16);
108 addQRTypeForNEON(MVT::v4i32);
109 addQRTypeForNEON(MVT::v2i64);
110 addQRTypeForNEON(MVT::v8f16);
113 // Compute derived properties from the register classes
114 computeRegisterProperties();
116 // Provide all sorts of operation actions
117 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
118 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
119 setOperationAction(ISD::SETCC, MVT::i32, Custom);
120 setOperationAction(ISD::SETCC, MVT::i64, Custom);
121 setOperationAction(ISD::SETCC, MVT::f32, Custom);
122 setOperationAction(ISD::SETCC, MVT::f64, Custom);
123 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
124 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
125 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
126 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
127 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
128 setOperationAction(ISD::SELECT, MVT::i32, Custom);
129 setOperationAction(ISD::SELECT, MVT::i64, Custom);
130 setOperationAction(ISD::SELECT, MVT::f32, Custom);
131 setOperationAction(ISD::SELECT, MVT::f64, Custom);
132 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
133 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
134 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
135 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
136 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
137 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
139 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
140 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
141 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
143 setOperationAction(ISD::FREM, MVT::f32, Expand);
144 setOperationAction(ISD::FREM, MVT::f64, Expand);
145 setOperationAction(ISD::FREM, MVT::f80, Expand);
147 // Custom lowering hooks are needed for XOR
148 // to fold it into CSINC/CSINV.
149 setOperationAction(ISD::XOR, MVT::i32, Custom);
150 setOperationAction(ISD::XOR, MVT::i64, Custom);
152 // Virtually no operation on f128 is legal, but LLVM can't expand them when
153 // there's a valid register class, so we need custom operations in most cases.
154 setOperationAction(ISD::FABS, MVT::f128, Expand);
155 setOperationAction(ISD::FADD, MVT::f128, Custom);
156 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
157 setOperationAction(ISD::FCOS, MVT::f128, Expand);
158 setOperationAction(ISD::FDIV, MVT::f128, Custom);
159 setOperationAction(ISD::FMA, MVT::f128, Expand);
160 setOperationAction(ISD::FMUL, MVT::f128, Custom);
161 setOperationAction(ISD::FNEG, MVT::f128, Expand);
162 setOperationAction(ISD::FPOW, MVT::f128, Expand);
163 setOperationAction(ISD::FREM, MVT::f128, Expand);
164 setOperationAction(ISD::FRINT, MVT::f128, Expand);
165 setOperationAction(ISD::FSIN, MVT::f128, Expand);
166 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
167 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
168 setOperationAction(ISD::FSUB, MVT::f128, Custom);
169 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
170 setOperationAction(ISD::SETCC, MVT::f128, Custom);
171 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
172 setOperationAction(ISD::SELECT, MVT::f128, Custom);
173 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
174 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
176 // Lowering for many of the conversions is actually specified by the non-f128
177 // type. The LowerXXX function will be trivial when f128 isn't involved.
178 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
179 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
180 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
181 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
182 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
183 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
184 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
185 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
186 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
187 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
188 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
189 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
190 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
191 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
193 // Variable arguments.
194 setOperationAction(ISD::VASTART, MVT::Other, Custom);
195 setOperationAction(ISD::VAARG, MVT::Other, Custom);
196 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
197 setOperationAction(ISD::VAEND, MVT::Other, Expand);
199 // Variable-sized objects.
200 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
201 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
202 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
204 // Exception handling.
205 // FIXME: These are guesses. Has this been defined yet?
206 setExceptionPointerRegister(AArch64::X0);
207 setExceptionSelectorRegister(AArch64::X1);
209 // Constant pool entries
210 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
213 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
215 // Add/Sub overflow ops with MVT::Glues are lowered to NZCV dependences.
216 setOperationAction(ISD::ADDC, MVT::i32, Custom);
217 setOperationAction(ISD::ADDE, MVT::i32, Custom);
218 setOperationAction(ISD::SUBC, MVT::i32, Custom);
219 setOperationAction(ISD::SUBE, MVT::i32, Custom);
220 setOperationAction(ISD::ADDC, MVT::i64, Custom);
221 setOperationAction(ISD::ADDE, MVT::i64, Custom);
222 setOperationAction(ISD::SUBC, MVT::i64, Custom);
223 setOperationAction(ISD::SUBE, MVT::i64, Custom);
225 // AArch64 lacks both left-rotate and popcount instructions.
226 setOperationAction(ISD::ROTL, MVT::i32, Expand);
227 setOperationAction(ISD::ROTL, MVT::i64, Expand);
229 // AArch64 doesn't have {U|S}MUL_LOHI.
230 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
231 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
234 // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
235 // counterparts, which AArch64 supports directly.
236 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
237 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
238 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
239 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
241 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
242 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
244 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
245 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
246 setOperationAction(ISD::SREM, MVT::i32, Expand);
247 setOperationAction(ISD::SREM, MVT::i64, Expand);
248 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
249 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
250 setOperationAction(ISD::UREM, MVT::i32, Expand);
251 setOperationAction(ISD::UREM, MVT::i64, Expand);
253 // Custom lower Add/Sub/Mul with overflow.
254 setOperationAction(ISD::SADDO, MVT::i32, Custom);
255 setOperationAction(ISD::SADDO, MVT::i64, Custom);
256 setOperationAction(ISD::UADDO, MVT::i32, Custom);
257 setOperationAction(ISD::UADDO, MVT::i64, Custom);
258 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
259 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
260 setOperationAction(ISD::USUBO, MVT::i32, Custom);
261 setOperationAction(ISD::USUBO, MVT::i64, Custom);
262 setOperationAction(ISD::SMULO, MVT::i32, Custom);
263 setOperationAction(ISD::SMULO, MVT::i64, Custom);
264 setOperationAction(ISD::UMULO, MVT::i32, Custom);
265 setOperationAction(ISD::UMULO, MVT::i64, Custom);
267 setOperationAction(ISD::FSIN, MVT::f32, Expand);
268 setOperationAction(ISD::FSIN, MVT::f64, Expand);
269 setOperationAction(ISD::FCOS, MVT::f32, Expand);
270 setOperationAction(ISD::FCOS, MVT::f64, Expand);
271 setOperationAction(ISD::FPOW, MVT::f32, Expand);
272 setOperationAction(ISD::FPOW, MVT::f64, Expand);
273 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
274 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
276 // f16 is storage-only, so we promote operations to f32 if we know this is
277 // valid, and ignore them otherwise. The operations not mentioned here will
278 // fail to select, but this is not a major problem as no source language
279 // should be emitting native f16 operations yet.
280 setOperationAction(ISD::FADD, MVT::f16, Promote);
281 setOperationAction(ISD::FDIV, MVT::f16, Promote);
282 setOperationAction(ISD::FMUL, MVT::f16, Promote);
283 setOperationAction(ISD::FSUB, MVT::f16, Promote);
285 // v4f16 is also a storage-only type, so promote it to v4f32 when that is
287 setOperationAction(ISD::FADD, MVT::v4f16, Promote);
288 setOperationAction(ISD::FSUB, MVT::v4f16, Promote);
289 setOperationAction(ISD::FMUL, MVT::v4f16, Promote);
290 setOperationAction(ISD::FDIV, MVT::v4f16, Promote);
291 setOperationAction(ISD::FP_EXTEND, MVT::v4f16, Promote);
292 setOperationAction(ISD::FP_ROUND, MVT::v4f16, Promote);
293 AddPromotedToType(ISD::FADD, MVT::v4f16, MVT::v4f32);
294 AddPromotedToType(ISD::FSUB, MVT::v4f16, MVT::v4f32);
295 AddPromotedToType(ISD::FMUL, MVT::v4f16, MVT::v4f32);
296 AddPromotedToType(ISD::FDIV, MVT::v4f16, MVT::v4f32);
297 AddPromotedToType(ISD::FP_EXTEND, MVT::v4f16, MVT::v4f32);
298 AddPromotedToType(ISD::FP_ROUND, MVT::v4f16, MVT::v4f32);
300 // Expand all other v4f16 operations.
301 // FIXME: We could generate better code by promoting some operations to
303 setOperationAction(ISD::FABS, MVT::v4f16, Expand);
304 setOperationAction(ISD::FCEIL, MVT::v4f16, Expand);
305 setOperationAction(ISD::FCOPYSIGN, MVT::v4f16, Expand);
306 setOperationAction(ISD::FCOS, MVT::v4f16, Expand);
307 setOperationAction(ISD::FFLOOR, MVT::v4f16, Expand);
308 setOperationAction(ISD::FMA, MVT::v4f16, Expand);
309 setOperationAction(ISD::FNEARBYINT, MVT::v4f16, Expand);
310 setOperationAction(ISD::FNEG, MVT::v4f16, Expand);
311 setOperationAction(ISD::FPOW, MVT::v4f16, Expand);
312 setOperationAction(ISD::FPOWI, MVT::v4f16, Expand);
313 setOperationAction(ISD::FREM, MVT::v4f16, Expand);
314 setOperationAction(ISD::FROUND, MVT::v4f16, Expand);
315 setOperationAction(ISD::FRINT, MVT::v4f16, Expand);
316 setOperationAction(ISD::FSIN, MVT::v4f16, Expand);
317 setOperationAction(ISD::FSINCOS, MVT::v4f16, Expand);
318 setOperationAction(ISD::FSQRT, MVT::v4f16, Expand);
319 setOperationAction(ISD::FTRUNC, MVT::v4f16, Expand);
320 setOperationAction(ISD::SETCC, MVT::v4f16, Expand);
321 setOperationAction(ISD::BR_CC, MVT::v4f16, Expand);
322 setOperationAction(ISD::SELECT, MVT::v4f16, Expand);
323 setOperationAction(ISD::SELECT_CC, MVT::v4f16, Expand);
324 setOperationAction(ISD::FEXP, MVT::v4f16, Expand);
325 setOperationAction(ISD::FEXP2, MVT::v4f16, Expand);
326 setOperationAction(ISD::FLOG, MVT::v4f16, Expand);
327 setOperationAction(ISD::FLOG2, MVT::v4f16, Expand);
328 setOperationAction(ISD::FLOG10, MVT::v4f16, Expand);
331 // v8f16 is also a storage-only type, so expand it.
332 setOperationAction(ISD::FABS, MVT::v8f16, Expand);
333 setOperationAction(ISD::FADD, MVT::v8f16, Expand);
334 setOperationAction(ISD::FCEIL, MVT::v8f16, Expand);
335 setOperationAction(ISD::FCOPYSIGN, MVT::v8f16, Expand);
336 setOperationAction(ISD::FCOS, MVT::v8f16, Expand);
337 setOperationAction(ISD::FDIV, MVT::v8f16, Expand);
338 setOperationAction(ISD::FFLOOR, MVT::v8f16, Expand);
339 setOperationAction(ISD::FMA, MVT::v8f16, Expand);
340 setOperationAction(ISD::FMUL, MVT::v8f16, Expand);
341 setOperationAction(ISD::FNEARBYINT, MVT::v8f16, Expand);
342 setOperationAction(ISD::FNEG, MVT::v8f16, Expand);
343 setOperationAction(ISD::FPOW, MVT::v8f16, Expand);
344 setOperationAction(ISD::FPOWI, MVT::v8f16, Expand);
345 setOperationAction(ISD::FREM, MVT::v8f16, Expand);
346 setOperationAction(ISD::FROUND, MVT::v8f16, Expand);
347 setOperationAction(ISD::FRINT, MVT::v8f16, Expand);
348 setOperationAction(ISD::FSIN, MVT::v8f16, Expand);
349 setOperationAction(ISD::FSINCOS, MVT::v8f16, Expand);
350 setOperationAction(ISD::FSQRT, MVT::v8f16, Expand);
351 setOperationAction(ISD::FSUB, MVT::v8f16, Expand);
352 setOperationAction(ISD::FTRUNC, MVT::v8f16, Expand);
353 setOperationAction(ISD::SETCC, MVT::v8f16, Expand);
354 setOperationAction(ISD::BR_CC, MVT::v8f16, Expand);
355 setOperationAction(ISD::SELECT, MVT::v8f16, Expand);
356 setOperationAction(ISD::SELECT_CC, MVT::v8f16, Expand);
357 setOperationAction(ISD::FP_EXTEND, MVT::v8f16, Expand);
358 setOperationAction(ISD::FEXP, MVT::v8f16, Expand);
359 setOperationAction(ISD::FEXP2, MVT::v8f16, Expand);
360 setOperationAction(ISD::FLOG, MVT::v8f16, Expand);
361 setOperationAction(ISD::FLOG2, MVT::v8f16, Expand);
362 setOperationAction(ISD::FLOG10, MVT::v8f16, Expand);
364 // AArch64 has implementations of a lot of rounding-like FP operations.
365 static MVT RoundingTypes[] = { MVT::f32, MVT::f64};
366 for (unsigned I = 0; I < array_lengthof(RoundingTypes); ++I) {
367 MVT Ty = RoundingTypes[I];
368 setOperationAction(ISD::FFLOOR, Ty, Legal);
369 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
370 setOperationAction(ISD::FCEIL, Ty, Legal);
371 setOperationAction(ISD::FRINT, Ty, Legal);
372 setOperationAction(ISD::FTRUNC, Ty, Legal);
373 setOperationAction(ISD::FROUND, Ty, Legal);
376 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
378 if (Subtarget->isTargetMachO()) {
379 // For iOS, we don't want to the normal expansion of a libcall to
380 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
382 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
383 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
385 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
386 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
389 // Make floating-point constants legal for the large code model, so they don't
390 // become loads from the constant pool.
391 if (Subtarget->isTargetMachO() && TM.getCodeModel() == CodeModel::Large) {
392 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
393 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
396 // AArch64 does not have floating-point extending loads, i1 sign-extending
397 // load, floating-point truncating stores, or v2i32->v2i16 truncating store.
398 for (MVT VT : MVT::fp_valuetypes()) {
399 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
400 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
401 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f64, Expand);
402 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f80, Expand);
404 for (MVT VT : MVT::integer_valuetypes())
405 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Expand);
407 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
408 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
409 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
410 setTruncStoreAction(MVT::f128, MVT::f80, Expand);
411 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
412 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
413 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
415 setOperationAction(ISD::BITCAST, MVT::i16, Custom);
416 setOperationAction(ISD::BITCAST, MVT::f16, Custom);
418 // Indexed loads and stores are supported.
419 for (unsigned im = (unsigned)ISD::PRE_INC;
420 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
421 setIndexedLoadAction(im, MVT::i8, Legal);
422 setIndexedLoadAction(im, MVT::i16, Legal);
423 setIndexedLoadAction(im, MVT::i32, Legal);
424 setIndexedLoadAction(im, MVT::i64, Legal);
425 setIndexedLoadAction(im, MVT::f64, Legal);
426 setIndexedLoadAction(im, MVT::f32, Legal);
427 setIndexedStoreAction(im, MVT::i8, Legal);
428 setIndexedStoreAction(im, MVT::i16, Legal);
429 setIndexedStoreAction(im, MVT::i32, Legal);
430 setIndexedStoreAction(im, MVT::i64, Legal);
431 setIndexedStoreAction(im, MVT::f64, Legal);
432 setIndexedStoreAction(im, MVT::f32, Legal);
436 setOperationAction(ISD::TRAP, MVT::Other, Legal);
438 // We combine OR nodes for bitfield operations.
439 setTargetDAGCombine(ISD::OR);
441 // Vector add and sub nodes may conceal a high-half opportunity.
442 // Also, try to fold ADD into CSINC/CSINV..
443 setTargetDAGCombine(ISD::ADD);
444 setTargetDAGCombine(ISD::SUB);
446 setTargetDAGCombine(ISD::XOR);
447 setTargetDAGCombine(ISD::SINT_TO_FP);
448 setTargetDAGCombine(ISD::UINT_TO_FP);
450 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
452 setTargetDAGCombine(ISD::ANY_EXTEND);
453 setTargetDAGCombine(ISD::ZERO_EXTEND);
454 setTargetDAGCombine(ISD::SIGN_EXTEND);
455 setTargetDAGCombine(ISD::BITCAST);
456 setTargetDAGCombine(ISD::CONCAT_VECTORS);
457 setTargetDAGCombine(ISD::STORE);
459 setTargetDAGCombine(ISD::MUL);
461 setTargetDAGCombine(ISD::SELECT);
462 setTargetDAGCombine(ISD::VSELECT);
464 setTargetDAGCombine(ISD::INTRINSIC_VOID);
465 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
466 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
468 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
469 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
470 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
472 setStackPointerRegisterToSaveRestore(AArch64::SP);
474 setSchedulingPreference(Sched::Hybrid);
477 MaskAndBranchFoldingIsLegal = true;
479 setMinFunctionAlignment(2);
481 RequireStrictAlign = (Align == StrictAlign);
483 setHasExtractBitsInsn(true);
485 if (Subtarget->hasNEON()) {
486 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
487 // silliness like this:
488 setOperationAction(ISD::FABS, MVT::v1f64, Expand);
489 setOperationAction(ISD::FADD, MVT::v1f64, Expand);
490 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
491 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
492 setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
493 setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
494 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
495 setOperationAction(ISD::FMA, MVT::v1f64, Expand);
496 setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
497 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
498 setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
499 setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
500 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
501 setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
502 setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
503 setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
504 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
505 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
506 setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
507 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
508 setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
509 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
510 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
511 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
512 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);
514 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
515 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
516 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
517 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
518 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);
520 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
522 // AArch64 doesn't have a direct vector ->f32 conversion instructions for
523 // elements smaller than i32, so promote the input to i32 first.
524 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
525 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
526 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
527 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
528 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
529 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
530 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
531 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
532 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
534 // AArch64 doesn't have MUL.2d:
535 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
536 // Custom handling for some quad-vector types to detect MULL.
537 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
538 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
539 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
541 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
542 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
543 // Likewise, narrowing and extending vector loads/stores aren't handled
545 for (MVT VT : MVT::vector_valuetypes()) {
546 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
548 setOperationAction(ISD::MULHS, VT, Expand);
549 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
550 setOperationAction(ISD::MULHU, VT, Expand);
551 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
553 setOperationAction(ISD::BSWAP, VT, Expand);
555 for (MVT InnerVT : MVT::vector_valuetypes()) {
556 setTruncStoreAction(VT, InnerVT, Expand);
557 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
558 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
559 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
563 // AArch64 has implementations of a lot of rounding-like FP operations.
564 static MVT RoundingVecTypes[] = {MVT::v2f32, MVT::v4f32, MVT::v2f64 };
565 for (unsigned I = 0; I < array_lengthof(RoundingVecTypes); ++I) {
566 MVT Ty = RoundingVecTypes[I];
567 setOperationAction(ISD::FFLOOR, Ty, Legal);
568 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
569 setOperationAction(ISD::FCEIL, Ty, Legal);
570 setOperationAction(ISD::FRINT, Ty, Legal);
571 setOperationAction(ISD::FTRUNC, Ty, Legal);
572 setOperationAction(ISD::FROUND, Ty, Legal);
576 // Prefer likely predicted branches to selects on out-of-order cores.
577 if (Subtarget->isCortexA57())
578 PredictableSelectIsExpensive = true;
581 void AArch64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
582 if (VT == MVT::v2f32 || VT == MVT::v4f16) {
583 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
584 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
586 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
587 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
588 } else if (VT == MVT::v2f64 || VT == MVT::v4f32 || VT == MVT::v8f16) {
589 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
590 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
592 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
593 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
596 // Mark vector float intrinsics as expand.
597 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
598 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
599 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
600 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
601 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
602 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
603 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
604 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
605 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
606 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
609 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
610 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
611 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
612 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
613 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
614 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
615 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
616 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
617 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
618 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
619 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
620 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
622 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
623 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
624 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
625 for (MVT InnerVT : MVT::all_valuetypes())
626 setLoadExtAction(ISD::EXTLOAD, InnerVT, 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 MachineFunction *MF = MBB->getParent();
881 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
882 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
883 DebugLoc DL = MI->getDebugLoc();
884 MachineFunction::iterator It = MBB;
887 unsigned DestReg = MI->getOperand(0).getReg();
888 unsigned IfTrueReg = MI->getOperand(1).getReg();
889 unsigned IfFalseReg = MI->getOperand(2).getReg();
890 unsigned CondCode = MI->getOperand(3).getImm();
891 bool NZCVKilled = MI->getOperand(4).isKill();
893 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
894 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
895 MF->insert(It, TrueBB);
896 MF->insert(It, EndBB);
898 // Transfer rest of current basic-block to EndBB
899 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
901 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
903 BuildMI(MBB, DL, TII->get(AArch64::Bcc)).addImm(CondCode).addMBB(TrueBB);
904 BuildMI(MBB, DL, TII->get(AArch64::B)).addMBB(EndBB);
905 MBB->addSuccessor(TrueBB);
906 MBB->addSuccessor(EndBB);
908 // TrueBB falls through to the end.
909 TrueBB->addSuccessor(EndBB);
912 TrueBB->addLiveIn(AArch64::NZCV);
913 EndBB->addLiveIn(AArch64::NZCV);
916 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(AArch64::PHI), DestReg)
922 MI->eraseFromParent();
927 AArch64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
928 MachineBasicBlock *BB) const {
929 switch (MI->getOpcode()) {
934 llvm_unreachable("Unexpected instruction for custom inserter!");
936 case AArch64::F128CSEL:
937 return EmitF128CSEL(MI, BB);
939 case TargetOpcode::STACKMAP:
940 case TargetOpcode::PATCHPOINT:
941 return emitPatchPoint(MI, BB);
945 //===----------------------------------------------------------------------===//
946 // AArch64 Lowering private implementation.
947 //===----------------------------------------------------------------------===//
949 //===----------------------------------------------------------------------===//
951 //===----------------------------------------------------------------------===//
953 /// changeIntCCToAArch64CC - Convert a DAG integer condition code to an AArch64
955 static AArch64CC::CondCode changeIntCCToAArch64CC(ISD::CondCode CC) {
958 llvm_unreachable("Unknown condition code!");
960 return AArch64CC::NE;
962 return AArch64CC::EQ;
964 return AArch64CC::GT;
966 return AArch64CC::GE;
968 return AArch64CC::LT;
970 return AArch64CC::LE;
972 return AArch64CC::HI;
974 return AArch64CC::HS;
976 return AArch64CC::LO;
978 return AArch64CC::LS;
982 /// changeFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64 CC.
983 static void changeFPCCToAArch64CC(ISD::CondCode CC,
984 AArch64CC::CondCode &CondCode,
985 AArch64CC::CondCode &CondCode2) {
986 CondCode2 = AArch64CC::AL;
989 llvm_unreachable("Unknown FP condition!");
992 CondCode = AArch64CC::EQ;
996 CondCode = AArch64CC::GT;
1000 CondCode = AArch64CC::GE;
1003 CondCode = AArch64CC::MI;
1006 CondCode = AArch64CC::LS;
1009 CondCode = AArch64CC::MI;
1010 CondCode2 = AArch64CC::GT;
1013 CondCode = AArch64CC::VC;
1016 CondCode = AArch64CC::VS;
1019 CondCode = AArch64CC::EQ;
1020 CondCode2 = AArch64CC::VS;
1023 CondCode = AArch64CC::HI;
1026 CondCode = AArch64CC::PL;
1030 CondCode = AArch64CC::LT;
1034 CondCode = AArch64CC::LE;
1038 CondCode = AArch64CC::NE;
1043 /// changeVectorFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64
1044 /// CC usable with the vector instructions. Fewer operations are available
1045 /// without a real NZCV register, so we have to use less efficient combinations
1046 /// to get the same effect.
1047 static void changeVectorFPCCToAArch64CC(ISD::CondCode CC,
1048 AArch64CC::CondCode &CondCode,
1049 AArch64CC::CondCode &CondCode2,
1054 // Mostly the scalar mappings work fine.
1055 changeFPCCToAArch64CC(CC, CondCode, CondCode2);
1058 Invert = true; // Fallthrough
1060 CondCode = AArch64CC::MI;
1061 CondCode2 = AArch64CC::GE;
1068 // All of the compare-mask comparisons are ordered, but we can switch
1069 // between the two by a double inversion. E.g. ULE == !OGT.
1071 changeFPCCToAArch64CC(getSetCCInverse(CC, false), CondCode, CondCode2);
1076 static bool isLegalArithImmed(uint64_t C) {
1077 // Matches AArch64DAGToDAGISel::SelectArithImmed().
1078 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
1081 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1082 SDLoc dl, SelectionDAG &DAG) {
1083 EVT VT = LHS.getValueType();
1085 if (VT.isFloatingPoint())
1086 return DAG.getNode(AArch64ISD::FCMP, dl, VT, LHS, RHS);
1088 // The CMP instruction is just an alias for SUBS, and representing it as
1089 // SUBS means that it's possible to get CSE with subtract operations.
1090 // A later phase can perform the optimization of setting the destination
1091 // register to WZR/XZR if it ends up being unused.
1092 unsigned Opcode = AArch64ISD::SUBS;
1094 if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
1095 cast<ConstantSDNode>(RHS.getOperand(0))->getZExtValue() == 0 &&
1096 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1097 // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on
1098 // the grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags
1099 // can be set differently by this operation. It comes down to whether
1100 // "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then
1101 // everything is fine. If not then the optimization is wrong. Thus general
1102 // comparisons are only valid if op2 != 0.
1104 // So, finally, the only LLVM-native comparisons that don't mention C and V
1105 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in
1106 // the absence of information about op2.
1107 Opcode = AArch64ISD::ADDS;
1108 RHS = RHS.getOperand(1);
1109 } else if (LHS.getOpcode() == ISD::AND && isa<ConstantSDNode>(RHS) &&
1110 cast<ConstantSDNode>(RHS)->getZExtValue() == 0 &&
1111 !isUnsignedIntSetCC(CC)) {
1112 // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
1113 // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
1114 // of the signed comparisons.
1115 Opcode = AArch64ISD::ANDS;
1116 RHS = LHS.getOperand(1);
1117 LHS = LHS.getOperand(0);
1120 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS)
1124 static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1125 SDValue &AArch64cc, SelectionDAG &DAG, SDLoc dl) {
1127 AArch64CC::CondCode AArch64CC;
1128 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1129 EVT VT = RHS.getValueType();
1130 uint64_t C = RHSC->getZExtValue();
1131 if (!isLegalArithImmed(C)) {
1132 // Constant does not fit, try adjusting it by one?
1138 if ((VT == MVT::i32 && C != 0x80000000 &&
1139 isLegalArithImmed((uint32_t)(C - 1))) ||
1140 (VT == MVT::i64 && C != 0x80000000ULL &&
1141 isLegalArithImmed(C - 1ULL))) {
1142 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1143 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1144 RHS = DAG.getConstant(C, VT);
1149 if ((VT == MVT::i32 && C != 0 &&
1150 isLegalArithImmed((uint32_t)(C - 1))) ||
1151 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1152 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1153 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1154 RHS = DAG.getConstant(C, VT);
1159 if ((VT == MVT::i32 && C != INT32_MAX &&
1160 isLegalArithImmed((uint32_t)(C + 1))) ||
1161 (VT == MVT::i64 && C != INT64_MAX &&
1162 isLegalArithImmed(C + 1ULL))) {
1163 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1164 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1165 RHS = DAG.getConstant(C, VT);
1170 if ((VT == MVT::i32 && C != UINT32_MAX &&
1171 isLegalArithImmed((uint32_t)(C + 1))) ||
1172 (VT == MVT::i64 && C != UINT64_MAX &&
1173 isLegalArithImmed(C + 1ULL))) {
1174 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1175 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1176 RHS = DAG.getConstant(C, VT);
1182 // The imm operand of ADDS is an unsigned immediate, in the range 0 to 4095.
1183 // For the i8 operand, the largest immediate is 255, so this can be easily
1184 // encoded in the compare instruction. For the i16 operand, however, the
1185 // largest immediate cannot be encoded in the compare.
1186 // Therefore, use a sign extending load and cmn to avoid materializing the -1
1187 // constant. For example,
1189 // ldrh w0, [x0, #0]
1192 // ldrsh w0, [x0, #0]
1194 // Fundamental, we're relying on the property that (zext LHS) == (zext RHS)
1195 // if and only if (sext LHS) == (sext RHS). The checks are in place to ensure
1196 // both the LHS and RHS are truely zero extended and to make sure the
1197 // transformation is profitable.
1198 if ((CC == ISD::SETEQ || CC == ISD::SETNE) && isa<ConstantSDNode>(RHS)) {
1199 if ((cast<ConstantSDNode>(RHS)->getZExtValue() >> 16 == 0) &&
1200 isa<LoadSDNode>(LHS)) {
1201 if (cast<LoadSDNode>(LHS)->getExtensionType() == ISD::ZEXTLOAD &&
1202 cast<LoadSDNode>(LHS)->getMemoryVT() == MVT::i16 &&
1203 LHS.getNode()->hasNUsesOfValue(1, 0)) {
1204 int16_t ValueofRHS = cast<ConstantSDNode>(RHS)->getZExtValue();
1205 if (ValueofRHS < 0 && isLegalArithImmed(-ValueofRHS)) {
1207 DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, LHS.getValueType(), LHS,
1208 DAG.getValueType(MVT::i16));
1209 Cmp = emitComparison(SExt,
1210 DAG.getConstant(ValueofRHS, RHS.getValueType()),
1212 AArch64CC = changeIntCCToAArch64CC(CC);
1213 AArch64cc = DAG.getConstant(AArch64CC, MVT::i32);
1219 Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
1220 AArch64CC = changeIntCCToAArch64CC(CC);
1221 AArch64cc = DAG.getConstant(AArch64CC, MVT::i32);
1225 static std::pair<SDValue, SDValue>
1226 getAArch64XALUOOp(AArch64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1227 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1228 "Unsupported value type");
1229 SDValue Value, Overflow;
1231 SDValue LHS = Op.getOperand(0);
1232 SDValue RHS = Op.getOperand(1);
1234 switch (Op.getOpcode()) {
1236 llvm_unreachable("Unknown overflow instruction!");
1238 Opc = AArch64ISD::ADDS;
1242 Opc = AArch64ISD::ADDS;
1246 Opc = AArch64ISD::SUBS;
1250 Opc = AArch64ISD::SUBS;
1253 // Multiply needs a little bit extra work.
1257 bool IsSigned = (Op.getOpcode() == ISD::SMULO) ? true : false;
1258 if (Op.getValueType() == MVT::i32) {
1259 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1260 // For a 32 bit multiply with overflow check we want the instruction
1261 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1262 // need to generate the following pattern:
1263 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1264 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1265 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1266 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1267 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1268 DAG.getConstant(0, MVT::i64));
1269 // On AArch64 the upper 32 bits are always zero extended for a 32 bit
1270 // operation. We need to clear out the upper 32 bits, because we used a
1271 // widening multiply that wrote all 64 bits. In the end this should be a
1273 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1275 // The signed overflow check requires more than just a simple check for
1276 // any bit set in the upper 32 bits of the result. These bits could be
1277 // just the sign bits of a negative number. To perform the overflow
1278 // check we have to arithmetic shift right the 32nd bit of the result by
1279 // 31 bits. Then we compare the result to the upper 32 bits.
1280 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1281 DAG.getConstant(32, MVT::i64));
1282 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1283 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1284 DAG.getConstant(31, MVT::i64));
1285 // It is important that LowerBits is last, otherwise the arithmetic
1286 // shift will not be folded into the compare (SUBS).
1287 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1288 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1291 // The overflow check for unsigned multiply is easy. We only need to
1292 // check if any of the upper 32 bits are set. This can be done with a
1293 // CMP (shifted register). For that we need to generate the following
1295 // (i64 AArch64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1296 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1297 DAG.getConstant(32, MVT::i64));
1298 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1300 DAG.getNode(AArch64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1301 UpperBits).getValue(1);
1305 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1306 // For the 64 bit multiply
1307 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1309 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1310 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1311 DAG.getConstant(63, MVT::i64));
1312 // It is important that LowerBits is last, otherwise the arithmetic
1313 // shift will not be folded into the compare (SUBS).
1314 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1315 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1318 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1319 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1321 DAG.getNode(AArch64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1322 UpperBits).getValue(1);
1329 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1331 // Emit the AArch64 operation with overflow check.
1332 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1333 Overflow = Value.getValue(1);
1335 return std::make_pair(Value, Overflow);
1338 SDValue AArch64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1339 RTLIB::Libcall Call) const {
1340 SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
1341 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1345 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1346 SDValue Sel = Op.getOperand(0);
1347 SDValue Other = Op.getOperand(1);
1349 // If neither operand is a SELECT_CC, give up.
1350 if (Sel.getOpcode() != ISD::SELECT_CC)
1351 std::swap(Sel, Other);
1352 if (Sel.getOpcode() != ISD::SELECT_CC)
1355 // The folding we want to perform is:
1356 // (xor x, (select_cc a, b, cc, 0, -1) )
1358 // (csel x, (xor x, -1), cc ...)
1360 // The latter will get matched to a CSINV instruction.
1362 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1363 SDValue LHS = Sel.getOperand(0);
1364 SDValue RHS = Sel.getOperand(1);
1365 SDValue TVal = Sel.getOperand(2);
1366 SDValue FVal = Sel.getOperand(3);
1369 // FIXME: This could be generalized to non-integer comparisons.
1370 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1373 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1374 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1376 // The the values aren't constants, this isn't the pattern we're looking for.
1377 if (!CFVal || !CTVal)
1380 // We can commute the SELECT_CC by inverting the condition. This
1381 // might be needed to make this fit into a CSINV pattern.
1382 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1383 std::swap(TVal, FVal);
1384 std::swap(CTVal, CFVal);
1385 CC = ISD::getSetCCInverse(CC, true);
1388 // If the constants line up, perform the transform!
1389 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1391 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1394 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1395 DAG.getConstant(-1ULL, Other.getValueType()));
1397 return DAG.getNode(AArch64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1404 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1405 EVT VT = Op.getValueType();
1407 // Let legalize expand this if it isn't a legal type yet.
1408 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1411 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1414 bool ExtraOp = false;
1415 switch (Op.getOpcode()) {
1417 llvm_unreachable("Invalid code");
1419 Opc = AArch64ISD::ADDS;
1422 Opc = AArch64ISD::SUBS;
1425 Opc = AArch64ISD::ADCS;
1429 Opc = AArch64ISD::SBCS;
1435 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1436 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1440 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1441 // Let legalize expand this if it isn't a legal type yet.
1442 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1445 AArch64CC::CondCode CC;
1446 // The actual operation that sets the overflow or carry flag.
1447 SDValue Value, Overflow;
1448 std::tie(Value, Overflow) = getAArch64XALUOOp(CC, Op, DAG);
1450 // We use 0 and 1 as false and true values.
1451 SDValue TVal = DAG.getConstant(1, MVT::i32);
1452 SDValue FVal = DAG.getConstant(0, MVT::i32);
1454 // We use an inverted condition, because the conditional select is inverted
1455 // too. This will allow it to be selected to a single instruction:
1456 // CSINC Wd, WZR, WZR, invert(cond).
1457 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), MVT::i32);
1458 Overflow = DAG.getNode(AArch64ISD::CSEL, SDLoc(Op), MVT::i32, FVal, TVal,
1461 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1462 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
1465 // Prefetch operands are:
1466 // 1: Address to prefetch
1468 // 3: int locality (0 = no locality ... 3 = extreme locality)
1469 // 4: bool isDataCache
1470 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1472 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1473 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1474 unsigned IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1476 bool IsStream = !Locality;
1477 // When the locality number is set
1479 // The front-end should have filtered out the out-of-range values
1480 assert(Locality <= 3 && "Prefetch locality out-of-range");
1481 // The locality degree is the opposite of the cache speed.
1482 // Put the number the other way around.
1483 // The encoding starts at 0 for level 1
1484 Locality = 3 - Locality;
1487 // built the mask value encoding the expected behavior.
1488 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1489 (!IsData << 3) | // IsDataCache bit
1490 (Locality << 1) | // Cache level bits
1491 (unsigned)IsStream; // Stream bit
1492 return DAG.getNode(AArch64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1493 DAG.getConstant(PrfOp, MVT::i32), Op.getOperand(1));
1496 SDValue AArch64TargetLowering::LowerFP_EXTEND(SDValue Op,
1497 SelectionDAG &DAG) const {
1498 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1501 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1503 return LowerF128Call(Op, DAG, LC);
1506 SDValue AArch64TargetLowering::LowerFP_ROUND(SDValue Op,
1507 SelectionDAG &DAG) const {
1508 if (Op.getOperand(0).getValueType() != MVT::f128) {
1509 // It's legal except when f128 is involved
1514 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1516 // FP_ROUND node has a second operand indicating whether it is known to be
1517 // precise. That doesn't take part in the LibCall so we can't directly use
1519 SDValue SrcVal = Op.getOperand(0);
1520 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1521 /*isSigned*/ false, SDLoc(Op)).first;
1524 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1525 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1526 // Any additional optimization in this function should be recorded
1527 // in the cost tables.
1528 EVT InVT = Op.getOperand(0).getValueType();
1529 EVT VT = Op.getValueType();
1531 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1534 DAG.getNode(Op.getOpcode(), dl, InVT.changeVectorElementTypeToInteger(),
1536 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1539 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1542 MVT::getVectorVT(MVT::getFloatingPointVT(VT.getScalarSizeInBits()),
1543 VT.getVectorNumElements());
1544 SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, ExtVT, Op.getOperand(0));
1545 return DAG.getNode(Op.getOpcode(), dl, VT, Ext);
1548 // Type changing conversions are illegal.
1552 SDValue AArch64TargetLowering::LowerFP_TO_INT(SDValue Op,
1553 SelectionDAG &DAG) const {
1554 if (Op.getOperand(0).getValueType().isVector())
1555 return LowerVectorFP_TO_INT(Op, DAG);
1557 if (Op.getOperand(0).getValueType() != MVT::f128) {
1558 // It's legal except when f128 is involved
1563 if (Op.getOpcode() == ISD::FP_TO_SINT)
1564 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1566 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1568 SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
1569 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1573 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1574 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1575 // Any additional optimization in this function should be recorded
1576 // in the cost tables.
1577 EVT VT = Op.getValueType();
1579 SDValue In = Op.getOperand(0);
1580 EVT InVT = In.getValueType();
1582 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1584 MVT::getVectorVT(MVT::getFloatingPointVT(InVT.getScalarSizeInBits()),
1585 InVT.getVectorNumElements());
1586 In = DAG.getNode(Op.getOpcode(), dl, CastVT, In);
1587 return DAG.getNode(ISD::FP_ROUND, dl, VT, In, DAG.getIntPtrConstant(0));
1590 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1592 Op.getOpcode() == ISD::SINT_TO_FP ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1593 EVT CastVT = VT.changeVectorElementTypeToInteger();
1594 In = DAG.getNode(CastOpc, dl, CastVT, In);
1595 return DAG.getNode(Op.getOpcode(), dl, VT, In);
1601 SDValue AArch64TargetLowering::LowerINT_TO_FP(SDValue Op,
1602 SelectionDAG &DAG) const {
1603 if (Op.getValueType().isVector())
1604 return LowerVectorINT_TO_FP(Op, DAG);
1606 // i128 conversions are libcalls.
1607 if (Op.getOperand(0).getValueType() == MVT::i128)
1610 // Other conversions are legal, unless it's to the completely software-based
1612 if (Op.getValueType() != MVT::f128)
1616 if (Op.getOpcode() == ISD::SINT_TO_FP)
1617 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1619 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1621 return LowerF128Call(Op, DAG, LC);
1624 SDValue AArch64TargetLowering::LowerFSINCOS(SDValue Op,
1625 SelectionDAG &DAG) const {
1626 // For iOS, we want to call an alternative entry point: __sincos_stret,
1627 // which returns the values in two S / D registers.
1629 SDValue Arg = Op.getOperand(0);
1630 EVT ArgVT = Arg.getValueType();
1631 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1638 Entry.isSExt = false;
1639 Entry.isZExt = false;
1640 Args.push_back(Entry);
1642 const char *LibcallName =
1643 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1644 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
1646 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
1647 TargetLowering::CallLoweringInfo CLI(DAG);
1648 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
1649 .setCallee(CallingConv::Fast, RetTy, Callee, std::move(Args), 0);
1651 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1652 return CallResult.first;
1655 static SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) {
1656 if (Op.getValueType() != MVT::f16)
1659 assert(Op.getOperand(0).getValueType() == MVT::i16);
1662 Op = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op.getOperand(0));
1663 Op = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Op);
1665 DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, MVT::f16, Op,
1666 DAG.getTargetConstant(AArch64::hsub, MVT::i32)),
1670 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
1671 if (OrigVT.getSizeInBits() >= 64)
1674 assert(OrigVT.isSimple() && "Expecting a simple value type");
1676 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
1677 switch (OrigSimpleTy) {
1678 default: llvm_unreachable("Unexpected Vector Type");
1687 static SDValue addRequiredExtensionForVectorMULL(SDValue N, SelectionDAG &DAG,
1690 unsigned ExtOpcode) {
1691 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
1692 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
1693 // 64-bits we need to insert a new extension so that it will be 64-bits.
1694 assert(ExtTy.is128BitVector() && "Unexpected extension size");
1695 if (OrigTy.getSizeInBits() >= 64)
1698 // Must extend size to at least 64 bits to be used as an operand for VMULL.
1699 EVT NewVT = getExtensionTo64Bits(OrigTy);
1701 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
1704 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
1706 EVT VT = N->getValueType(0);
1708 if (N->getOpcode() != ISD::BUILD_VECTOR)
1711 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1712 SDNode *Elt = N->getOperand(i).getNode();
1713 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
1714 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1715 unsigned HalfSize = EltSize / 2;
1717 if (!isIntN(HalfSize, C->getSExtValue()))
1720 if (!isUIntN(HalfSize, C->getZExtValue()))
1731 static SDValue skipExtensionForVectorMULL(SDNode *N, SelectionDAG &DAG) {
1732 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
1733 return addRequiredExtensionForVectorMULL(N->getOperand(0), DAG,
1734 N->getOperand(0)->getValueType(0),
1738 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
1739 EVT VT = N->getValueType(0);
1740 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
1741 unsigned NumElts = VT.getVectorNumElements();
1742 MVT TruncVT = MVT::getIntegerVT(EltSize);
1743 SmallVector<SDValue, 8> Ops;
1744 for (unsigned i = 0; i != NumElts; ++i) {
1745 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
1746 const APInt &CInt = C->getAPIntValue();
1747 // Element types smaller than 32 bits are not legal, so use i32 elements.
1748 // The values are implicitly truncated so sext vs. zext doesn't matter.
1749 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
1751 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
1752 MVT::getVectorVT(TruncVT, NumElts), Ops);
1755 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
1756 if (N->getOpcode() == ISD::SIGN_EXTEND)
1758 if (isExtendedBUILD_VECTOR(N, DAG, true))
1763 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
1764 if (N->getOpcode() == ISD::ZERO_EXTEND)
1766 if (isExtendedBUILD_VECTOR(N, DAG, false))
1771 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
1772 unsigned Opcode = N->getOpcode();
1773 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
1774 SDNode *N0 = N->getOperand(0).getNode();
1775 SDNode *N1 = N->getOperand(1).getNode();
1776 return N0->hasOneUse() && N1->hasOneUse() &&
1777 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
1782 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
1783 unsigned Opcode = N->getOpcode();
1784 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
1785 SDNode *N0 = N->getOperand(0).getNode();
1786 SDNode *N1 = N->getOperand(1).getNode();
1787 return N0->hasOneUse() && N1->hasOneUse() &&
1788 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
1793 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
1794 // Multiplications are only custom-lowered for 128-bit vectors so that
1795 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
1796 EVT VT = Op.getValueType();
1797 assert(VT.is128BitVector() && VT.isInteger() &&
1798 "unexpected type for custom-lowering ISD::MUL");
1799 SDNode *N0 = Op.getOperand(0).getNode();
1800 SDNode *N1 = Op.getOperand(1).getNode();
1801 unsigned NewOpc = 0;
1803 bool isN0SExt = isSignExtended(N0, DAG);
1804 bool isN1SExt = isSignExtended(N1, DAG);
1805 if (isN0SExt && isN1SExt)
1806 NewOpc = AArch64ISD::SMULL;
1808 bool isN0ZExt = isZeroExtended(N0, DAG);
1809 bool isN1ZExt = isZeroExtended(N1, DAG);
1810 if (isN0ZExt && isN1ZExt)
1811 NewOpc = AArch64ISD::UMULL;
1812 else if (isN1SExt || isN1ZExt) {
1813 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
1814 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
1815 if (isN1SExt && isAddSubSExt(N0, DAG)) {
1816 NewOpc = AArch64ISD::SMULL;
1818 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
1819 NewOpc = AArch64ISD::UMULL;
1821 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
1823 NewOpc = AArch64ISD::UMULL;
1829 if (VT == MVT::v2i64)
1830 // Fall through to expand this. It is not legal.
1833 // Other vector multiplications are legal.
1838 // Legalize to a S/UMULL instruction
1841 SDValue Op1 = skipExtensionForVectorMULL(N1, DAG);
1843 Op0 = skipExtensionForVectorMULL(N0, DAG);
1844 assert(Op0.getValueType().is64BitVector() &&
1845 Op1.getValueType().is64BitVector() &&
1846 "unexpected types for extended operands to VMULL");
1847 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
1849 // Optimizing (zext A + zext B) * C, to (S/UMULL A, C) + (S/UMULL B, C) during
1850 // isel lowering to take advantage of no-stall back to back s/umul + s/umla.
1851 // This is true for CPUs with accumulate forwarding such as Cortex-A53/A57
1852 SDValue N00 = skipExtensionForVectorMULL(N0->getOperand(0).getNode(), DAG);
1853 SDValue N01 = skipExtensionForVectorMULL(N0->getOperand(1).getNode(), DAG);
1854 EVT Op1VT = Op1.getValueType();
1855 return DAG.getNode(N0->getOpcode(), DL, VT,
1856 DAG.getNode(NewOpc, DL, VT,
1857 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
1858 DAG.getNode(NewOpc, DL, VT,
1859 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
1862 SDValue AArch64TargetLowering::LowerOperation(SDValue Op,
1863 SelectionDAG &DAG) const {
1864 switch (Op.getOpcode()) {
1866 llvm_unreachable("unimplemented operand");
1869 return LowerBITCAST(Op, DAG);
1870 case ISD::GlobalAddress:
1871 return LowerGlobalAddress(Op, DAG);
1872 case ISD::GlobalTLSAddress:
1873 return LowerGlobalTLSAddress(Op, DAG);
1875 return LowerSETCC(Op, DAG);
1877 return LowerBR_CC(Op, DAG);
1879 return LowerSELECT(Op, DAG);
1880 case ISD::SELECT_CC:
1881 return LowerSELECT_CC(Op, DAG);
1882 case ISD::JumpTable:
1883 return LowerJumpTable(Op, DAG);
1884 case ISD::ConstantPool:
1885 return LowerConstantPool(Op, DAG);
1886 case ISD::BlockAddress:
1887 return LowerBlockAddress(Op, DAG);
1889 return LowerVASTART(Op, DAG);
1891 return LowerVACOPY(Op, DAG);
1893 return LowerVAARG(Op, DAG);
1898 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
1905 return LowerXALUO(Op, DAG);
1907 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
1909 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
1911 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
1913 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
1915 return LowerFP_ROUND(Op, DAG);
1916 case ISD::FP_EXTEND:
1917 return LowerFP_EXTEND(Op, DAG);
1918 case ISD::FRAMEADDR:
1919 return LowerFRAMEADDR(Op, DAG);
1920 case ISD::RETURNADDR:
1921 return LowerRETURNADDR(Op, DAG);
1922 case ISD::INSERT_VECTOR_ELT:
1923 return LowerINSERT_VECTOR_ELT(Op, DAG);
1924 case ISD::EXTRACT_VECTOR_ELT:
1925 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
1926 case ISD::BUILD_VECTOR:
1927 return LowerBUILD_VECTOR(Op, DAG);
1928 case ISD::VECTOR_SHUFFLE:
1929 return LowerVECTOR_SHUFFLE(Op, DAG);
1930 case ISD::EXTRACT_SUBVECTOR:
1931 return LowerEXTRACT_SUBVECTOR(Op, DAG);
1935 return LowerVectorSRA_SRL_SHL(Op, DAG);
1936 case ISD::SHL_PARTS:
1937 return LowerShiftLeftParts(Op, DAG);
1938 case ISD::SRL_PARTS:
1939 case ISD::SRA_PARTS:
1940 return LowerShiftRightParts(Op, DAG);
1942 return LowerCTPOP(Op, DAG);
1943 case ISD::FCOPYSIGN:
1944 return LowerFCOPYSIGN(Op, DAG);
1946 return LowerVectorAND(Op, DAG);
1948 return LowerVectorOR(Op, DAG);
1950 return LowerXOR(Op, DAG);
1952 return LowerPREFETCH(Op, DAG);
1953 case ISD::SINT_TO_FP:
1954 case ISD::UINT_TO_FP:
1955 return LowerINT_TO_FP(Op, DAG);
1956 case ISD::FP_TO_SINT:
1957 case ISD::FP_TO_UINT:
1958 return LowerFP_TO_INT(Op, DAG);
1960 return LowerFSINCOS(Op, DAG);
1962 return LowerMUL(Op, DAG);
1966 /// getFunctionAlignment - Return the Log2 alignment of this function.
1967 unsigned AArch64TargetLowering::getFunctionAlignment(const Function *F) const {
1971 //===----------------------------------------------------------------------===//
1972 // Calling Convention Implementation
1973 //===----------------------------------------------------------------------===//
1975 #include "AArch64GenCallingConv.inc"
1977 /// Selects the correct CCAssignFn for a given CallingConvention value.
1978 CCAssignFn *AArch64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
1979 bool IsVarArg) const {
1982 llvm_unreachable("Unsupported calling convention.");
1983 case CallingConv::WebKit_JS:
1984 return CC_AArch64_WebKit_JS;
1985 case CallingConv::GHC:
1986 return CC_AArch64_GHC;
1987 case CallingConv::C:
1988 case CallingConv::Fast:
1989 if (!Subtarget->isTargetDarwin())
1990 return CC_AArch64_AAPCS;
1991 return IsVarArg ? CC_AArch64_DarwinPCS_VarArg : CC_AArch64_DarwinPCS;
1995 SDValue AArch64TargetLowering::LowerFormalArguments(
1996 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1997 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
1998 SmallVectorImpl<SDValue> &InVals) const {
1999 MachineFunction &MF = DAG.getMachineFunction();
2000 MachineFrameInfo *MFI = MF.getFrameInfo();
2002 // Assign locations to all of the incoming arguments.
2003 SmallVector<CCValAssign, 16> ArgLocs;
2004 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2007 // At this point, Ins[].VT may already be promoted to i32. To correctly
2008 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2009 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2010 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
2011 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
2013 unsigned NumArgs = Ins.size();
2014 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2015 unsigned CurArgIdx = 0;
2016 for (unsigned i = 0; i != NumArgs; ++i) {
2017 MVT ValVT = Ins[i].VT;
2018 if (Ins[i].isOrigArg()) {
2019 std::advance(CurOrigArg, Ins[i].getOrigArgIndex() - CurArgIdx);
2020 CurArgIdx = Ins[i].getOrigArgIndex();
2022 // Get type of the original argument.
2023 EVT ActualVT = getValueType(CurOrigArg->getType(), /*AllowUnknown*/ true);
2024 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
2025 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2026 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2028 else if (ActualMVT == MVT::i16)
2031 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2033 AssignFn(i, ValVT, ValVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
2034 assert(!Res && "Call operand has unhandled type");
2037 assert(ArgLocs.size() == Ins.size());
2038 SmallVector<SDValue, 16> ArgValues;
2039 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2040 CCValAssign &VA = ArgLocs[i];
2042 if (Ins[i].Flags.isByVal()) {
2043 // Byval is used for HFAs in the PCS, but the system should work in a
2044 // non-compliant manner for larger structs.
2045 EVT PtrTy = getPointerTy();
2046 int Size = Ins[i].Flags.getByValSize();
2047 unsigned NumRegs = (Size + 7) / 8;
2049 // FIXME: This works on big-endian for composite byvals, which are the common
2050 // case. It should also work for fundamental types too.
2052 MFI->CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false);
2053 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrTy);
2054 InVals.push_back(FrameIdxN);
2059 if (VA.isRegLoc()) {
2060 // Arguments stored in registers.
2061 EVT RegVT = VA.getLocVT();
2064 const TargetRegisterClass *RC;
2066 if (RegVT == MVT::i32)
2067 RC = &AArch64::GPR32RegClass;
2068 else if (RegVT == MVT::i64)
2069 RC = &AArch64::GPR64RegClass;
2070 else if (RegVT == MVT::f16)
2071 RC = &AArch64::FPR16RegClass;
2072 else if (RegVT == MVT::f32)
2073 RC = &AArch64::FPR32RegClass;
2074 else if (RegVT == MVT::f64 || RegVT.is64BitVector())
2075 RC = &AArch64::FPR64RegClass;
2076 else if (RegVT == MVT::f128 || RegVT.is128BitVector())
2077 RC = &AArch64::FPR128RegClass;
2079 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2081 // Transform the arguments in physical registers into virtual ones.
2082 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2083 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2085 // If this is an 8, 16 or 32-bit value, it is really passed promoted
2086 // to 64 bits. Insert an assert[sz]ext to capture this, then
2087 // truncate to the right size.
2088 switch (VA.getLocInfo()) {
2090 llvm_unreachable("Unknown loc info!");
2091 case CCValAssign::Full:
2093 case CCValAssign::BCvt:
2094 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
2096 case CCValAssign::AExt:
2097 case CCValAssign::SExt:
2098 case CCValAssign::ZExt:
2099 // SelectionDAGBuilder will insert appropriate AssertZExt & AssertSExt
2100 // nodes after our lowering.
2101 assert(RegVT == Ins[i].VT && "incorrect register location selected");
2105 InVals.push_back(ArgValue);
2107 } else { // VA.isRegLoc()
2108 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
2109 unsigned ArgOffset = VA.getLocMemOffset();
2110 unsigned ArgSize = VA.getValVT().getSizeInBits() / 8;
2112 uint32_t BEAlign = 0;
2113 if (!Subtarget->isLittleEndian() && ArgSize < 8 &&
2114 !Ins[i].Flags.isInConsecutiveRegs())
2115 BEAlign = 8 - ArgSize;
2117 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset + BEAlign, true);
2119 // Create load nodes to retrieve arguments from the stack.
2120 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2123 // For NON_EXTLOAD, generic code in getLoad assert(ValVT == MemVT)
2124 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
2125 MVT MemVT = VA.getValVT();
2127 switch (VA.getLocInfo()) {
2130 case CCValAssign::BCvt:
2131 MemVT = VA.getLocVT();
2133 case CCValAssign::SExt:
2134 ExtType = ISD::SEXTLOAD;
2136 case CCValAssign::ZExt:
2137 ExtType = ISD::ZEXTLOAD;
2139 case CCValAssign::AExt:
2140 ExtType = ISD::EXTLOAD;
2144 ArgValue = DAG.getExtLoad(ExtType, DL, VA.getLocVT(), Chain, FIN,
2145 MachinePointerInfo::getFixedStack(FI),
2146 MemVT, false, false, false, 0);
2148 InVals.push_back(ArgValue);
2154 if (!Subtarget->isTargetDarwin()) {
2155 // The AAPCS variadic function ABI is identical to the non-variadic
2156 // one. As a result there may be more arguments in registers and we should
2157 // save them for future reference.
2158 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
2161 AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>();
2162 // This will point to the next argument passed via stack.
2163 unsigned StackOffset = CCInfo.getNextStackOffset();
2164 // We currently pass all varargs at 8-byte alignment.
2165 StackOffset = ((StackOffset + 7) & ~7);
2166 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
2169 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2170 unsigned StackArgSize = CCInfo.getNextStackOffset();
2171 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2172 if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
2173 // This is a non-standard ABI so by fiat I say we're allowed to make full
2174 // use of the stack area to be popped, which must be aligned to 16 bytes in
2176 StackArgSize = RoundUpToAlignment(StackArgSize, 16);
2178 // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
2179 // a multiple of 16.
2180 FuncInfo->setArgumentStackToRestore(StackArgSize);
2182 // This realignment carries over to the available bytes below. Our own
2183 // callers will guarantee the space is free by giving an aligned value to
2186 // Even if we're not expected to free up the space, it's useful to know how
2187 // much is there while considering tail calls (because we can reuse it).
2188 FuncInfo->setBytesInStackArgArea(StackArgSize);
2193 void AArch64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
2194 SelectionDAG &DAG, SDLoc DL,
2195 SDValue &Chain) const {
2196 MachineFunction &MF = DAG.getMachineFunction();
2197 MachineFrameInfo *MFI = MF.getFrameInfo();
2198 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2200 SmallVector<SDValue, 8> MemOps;
2202 static const MCPhysReg GPRArgRegs[] = { AArch64::X0, AArch64::X1, AArch64::X2,
2203 AArch64::X3, AArch64::X4, AArch64::X5,
2204 AArch64::X6, AArch64::X7 };
2205 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
2206 unsigned FirstVariadicGPR = CCInfo.getFirstUnallocated(GPRArgRegs);
2208 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
2210 if (GPRSaveSize != 0) {
2211 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
2213 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
2215 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
2216 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &AArch64::GPR64RegClass);
2217 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
2219 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2220 MachinePointerInfo::getStack(i * 8), false, false, 0);
2221 MemOps.push_back(Store);
2222 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2223 DAG.getConstant(8, getPointerTy()));
2226 FuncInfo->setVarArgsGPRIndex(GPRIdx);
2227 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
2229 if (Subtarget->hasFPARMv8()) {
2230 static const MCPhysReg FPRArgRegs[] = {
2231 AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3,
2232 AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7};
2233 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
2234 unsigned FirstVariadicFPR = CCInfo.getFirstUnallocated(FPRArgRegs);
2236 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
2238 if (FPRSaveSize != 0) {
2239 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
2241 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
2243 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
2244 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &AArch64::FPR128RegClass);
2245 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);
2248 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2249 MachinePointerInfo::getStack(i * 16), false, false, 0);
2250 MemOps.push_back(Store);
2251 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2252 DAG.getConstant(16, getPointerTy()));
2255 FuncInfo->setVarArgsFPRIndex(FPRIdx);
2256 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
2259 if (!MemOps.empty()) {
2260 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
2264 /// LowerCallResult - Lower the result values of a call into the
2265 /// appropriate copies out of appropriate physical registers.
2266 SDValue AArch64TargetLowering::LowerCallResult(
2267 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
2268 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2269 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
2270 SDValue ThisVal) const {
2271 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2272 ? RetCC_AArch64_WebKit_JS
2273 : RetCC_AArch64_AAPCS;
2274 // Assign locations to each value returned by this call.
2275 SmallVector<CCValAssign, 16> RVLocs;
2276 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2278 CCInfo.AnalyzeCallResult(Ins, RetCC);
2280 // Copy all of the result registers out of their specified physreg.
2281 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2282 CCValAssign VA = RVLocs[i];
2284 // Pass 'this' value directly from the argument to return value, to avoid
2285 // reg unit interference
2286 if (i == 0 && isThisReturn) {
2287 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
2288 "unexpected return calling convention register assignment");
2289 InVals.push_back(ThisVal);
2294 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
2295 Chain = Val.getValue(1);
2296 InFlag = Val.getValue(2);
2298 switch (VA.getLocInfo()) {
2300 llvm_unreachable("Unknown loc info!");
2301 case CCValAssign::Full:
2303 case CCValAssign::BCvt:
2304 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2308 InVals.push_back(Val);
2314 bool AArch64TargetLowering::isEligibleForTailCallOptimization(
2315 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
2316 bool isCalleeStructRet, bool isCallerStructRet,
2317 const SmallVectorImpl<ISD::OutputArg> &Outs,
2318 const SmallVectorImpl<SDValue> &OutVals,
2319 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
2320 // For CallingConv::C this function knows whether the ABI needs
2321 // changing. That's not true for other conventions so they will have to opt in
2323 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
2326 const MachineFunction &MF = DAG.getMachineFunction();
2327 const Function *CallerF = MF.getFunction();
2328 CallingConv::ID CallerCC = CallerF->getCallingConv();
2329 bool CCMatch = CallerCC == CalleeCC;
2331 // Byval parameters hand the function a pointer directly into the stack area
2332 // we want to reuse during a tail call. Working around this *is* possible (see
2333 // X86) but less efficient and uglier in LowerCall.
2334 for (Function::const_arg_iterator i = CallerF->arg_begin(),
2335 e = CallerF->arg_end();
2337 if (i->hasByValAttr())
2340 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
2341 if (IsTailCallConvention(CalleeCC) && CCMatch)
2346 // Externally-defined functions with weak linkage should not be
2347 // tail-called on AArch64 when the OS does not support dynamic
2348 // pre-emption of symbols, as the AAELF spec requires normal calls
2349 // to undefined weak functions to be replaced with a NOP or jump to the
2350 // next instruction. The behaviour of branch instructions in this
2351 // situation (as used for tail calls) is implementation-defined, so we
2352 // cannot rely on the linker replacing the tail call with a return.
2353 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2354 const GlobalValue *GV = G->getGlobal();
2355 const Triple TT(getTargetMachine().getTargetTriple());
2356 if (GV->hasExternalWeakLinkage() &&
2357 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2361 // Now we search for cases where we can use a tail call without changing the
2362 // ABI. Sibcall is used in some places (particularly gcc) to refer to this
2365 // I want anyone implementing a new calling convention to think long and hard
2366 // about this assert.
2367 assert((!isVarArg || CalleeCC == CallingConv::C) &&
2368 "Unexpected variadic calling convention");
2370 if (isVarArg && !Outs.empty()) {
2371 // At least two cases here: if caller is fastcc then we can't have any
2372 // memory arguments (we'd be expected to clean up the stack afterwards). If
2373 // caller is C then we could potentially use its argument area.
2375 // FIXME: for now we take the most conservative of these in both cases:
2376 // disallow all variadic memory operands.
2377 SmallVector<CCValAssign, 16> ArgLocs;
2378 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2381 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, true));
2382 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2383 if (!ArgLocs[i].isRegLoc())
2387 // If the calling conventions do not match, then we'd better make sure the
2388 // results are returned in the same way as what the caller expects.
2390 SmallVector<CCValAssign, 16> RVLocs1;
2391 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2393 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForCall(CalleeCC, isVarArg));
2395 SmallVector<CCValAssign, 16> RVLocs2;
2396 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2398 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForCall(CallerCC, isVarArg));
2400 if (RVLocs1.size() != RVLocs2.size())
2402 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2403 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2405 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2407 if (RVLocs1[i].isRegLoc()) {
2408 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2411 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2417 // Nothing more to check if the callee is taking no arguments
2421 SmallVector<CCValAssign, 16> ArgLocs;
2422 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2425 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, isVarArg));
2427 const AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2429 // If the stack arguments for this call would fit into our own save area then
2430 // the call can be made tail.
2431 return CCInfo.getNextStackOffset() <= FuncInfo->getBytesInStackArgArea();
2434 SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain,
2436 MachineFrameInfo *MFI,
2437 int ClobberedFI) const {
2438 SmallVector<SDValue, 8> ArgChains;
2439 int64_t FirstByte = MFI->getObjectOffset(ClobberedFI);
2440 int64_t LastByte = FirstByte + MFI->getObjectSize(ClobberedFI) - 1;
2442 // Include the original chain at the beginning of the list. When this is
2443 // used by target LowerCall hooks, this helps legalize find the
2444 // CALLSEQ_BEGIN node.
2445 ArgChains.push_back(Chain);
2447 // Add a chain value for each stack argument corresponding
2448 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
2449 UE = DAG.getEntryNode().getNode()->use_end();
2451 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
2452 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
2453 if (FI->getIndex() < 0) {
2454 int64_t InFirstByte = MFI->getObjectOffset(FI->getIndex());
2455 int64_t InLastByte = InFirstByte;
2456 InLastByte += MFI->getObjectSize(FI->getIndex()) - 1;
2458 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
2459 (FirstByte <= InFirstByte && InFirstByte <= LastByte))
2460 ArgChains.push_back(SDValue(L, 1));
2463 // Build a tokenfactor for all the chains.
2464 return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
2467 bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
2468 bool TailCallOpt) const {
2469 return CallCC == CallingConv::Fast && TailCallOpt;
2472 bool AArch64TargetLowering::IsTailCallConvention(CallingConv::ID CallCC) const {
2473 return CallCC == CallingConv::Fast;
2476 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
2477 /// and add input and output parameter nodes.
2479 AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
2480 SmallVectorImpl<SDValue> &InVals) const {
2481 SelectionDAG &DAG = CLI.DAG;
2483 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2484 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2485 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2486 SDValue Chain = CLI.Chain;
2487 SDValue Callee = CLI.Callee;
2488 bool &IsTailCall = CLI.IsTailCall;
2489 CallingConv::ID CallConv = CLI.CallConv;
2490 bool IsVarArg = CLI.IsVarArg;
2492 MachineFunction &MF = DAG.getMachineFunction();
2493 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2494 bool IsThisReturn = false;
2496 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2497 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2498 bool IsSibCall = false;
2501 // Check if it's really possible to do a tail call.
2502 IsTailCall = isEligibleForTailCallOptimization(
2503 Callee, CallConv, IsVarArg, IsStructRet,
2504 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2505 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
2506 report_fatal_error("failed to perform tail call elimination on a call "
2507 "site marked musttail");
2509 // A sibling call is one where we're under the usual C ABI and not planning
2510 // to change that but can still do a tail call:
2511 if (!TailCallOpt && IsTailCall)
2518 // Analyze operands of the call, assigning locations to each operand.
2519 SmallVector<CCValAssign, 16> ArgLocs;
2520 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
2524 // Handle fixed and variable vector arguments differently.
2525 // Variable vector arguments always go into memory.
2526 unsigned NumArgs = Outs.size();
2528 for (unsigned i = 0; i != NumArgs; ++i) {
2529 MVT ArgVT = Outs[i].VT;
2530 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2531 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2532 /*IsVarArg=*/ !Outs[i].IsFixed);
2533 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2534 assert(!Res && "Call operand has unhandled type");
2538 // At this point, Outs[].VT may already be promoted to i32. To correctly
2539 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2540 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2541 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2542 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2544 unsigned NumArgs = Outs.size();
2545 for (unsigned i = 0; i != NumArgs; ++i) {
2546 MVT ValVT = Outs[i].VT;
2547 // Get type of the original argument.
2548 EVT ActualVT = getValueType(CLI.getArgs()[Outs[i].OrigArgIndex].Ty,
2549 /*AllowUnknown*/ true);
2550 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2551 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2552 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2553 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2555 else if (ActualMVT == MVT::i16)
2558 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2559 bool Res = AssignFn(i, ValVT, ValVT, CCValAssign::Full, ArgFlags, CCInfo);
2560 assert(!Res && "Call operand has unhandled type");
2565 // Get a count of how many bytes are to be pushed on the stack.
2566 unsigned NumBytes = CCInfo.getNextStackOffset();
2569 // Since we're not changing the ABI to make this a tail call, the memory
2570 // operands are already available in the caller's incoming argument space.
2574 // FPDiff is the byte offset of the call's argument area from the callee's.
2575 // Stores to callee stack arguments will be placed in FixedStackSlots offset
2576 // by this amount for a tail call. In a sibling call it must be 0 because the
2577 // caller will deallocate the entire stack and the callee still expects its
2578 // arguments to begin at SP+0. Completely unused for non-tail calls.
2581 if (IsTailCall && !IsSibCall) {
2582 unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();
2584 // Since callee will pop argument stack as a tail call, we must keep the
2585 // popped size 16-byte aligned.
2586 NumBytes = RoundUpToAlignment(NumBytes, 16);
2588 // FPDiff will be negative if this tail call requires more space than we
2589 // would automatically have in our incoming argument space. Positive if we
2590 // can actually shrink the stack.
2591 FPDiff = NumReusableBytes - NumBytes;
2593 // The stack pointer must be 16-byte aligned at all times it's used for a
2594 // memory operation, which in practice means at *all* times and in
2595 // particular across call boundaries. Therefore our own arguments started at
2596 // a 16-byte aligned SP and the delta applied for the tail call should
2597 // satisfy the same constraint.
2598 assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
2601 // Adjust the stack pointer for the new arguments...
2602 // These operations are automatically eliminated by the prolog/epilog pass
2605 DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true), DL);
2607 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, AArch64::SP, getPointerTy());
2609 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2610 SmallVector<SDValue, 8> MemOpChains;
2612 // Walk the register/memloc assignments, inserting copies/loads.
2613 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2614 ++i, ++realArgIdx) {
2615 CCValAssign &VA = ArgLocs[i];
2616 SDValue Arg = OutVals[realArgIdx];
2617 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2619 // Promote the value if needed.
2620 switch (VA.getLocInfo()) {
2622 llvm_unreachable("Unknown loc info!");
2623 case CCValAssign::Full:
2625 case CCValAssign::SExt:
2626 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
2628 case CCValAssign::ZExt:
2629 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2631 case CCValAssign::AExt:
2632 if (Outs[realArgIdx].ArgVT == MVT::i1) {
2633 // AAPCS requires i1 to be zero-extended to 8-bits by the caller.
2634 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
2635 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i8, Arg);
2637 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
2639 case CCValAssign::BCvt:
2640 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2642 case CCValAssign::FPExt:
2643 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
2647 if (VA.isRegLoc()) {
2648 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
2649 assert(VA.getLocVT() == MVT::i64 &&
2650 "unexpected calling convention register assignment");
2651 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
2652 "unexpected use of 'returned'");
2653 IsThisReturn = true;
2655 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2657 assert(VA.isMemLoc());
2660 MachinePointerInfo DstInfo;
2662 // FIXME: This works on big-endian for composite byvals, which are the
2663 // common case. It should also work for fundamental types too.
2664 uint32_t BEAlign = 0;
2665 unsigned OpSize = Flags.isByVal() ? Flags.getByValSize() * 8
2666 : VA.getValVT().getSizeInBits();
2667 OpSize = (OpSize + 7) / 8;
2668 if (!Subtarget->isLittleEndian() && !Flags.isByVal() &&
2669 !Flags.isInConsecutiveRegs()) {
2671 BEAlign = 8 - OpSize;
2673 unsigned LocMemOffset = VA.getLocMemOffset();
2674 int32_t Offset = LocMemOffset + BEAlign;
2675 SDValue PtrOff = DAG.getIntPtrConstant(Offset);
2676 PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2679 Offset = Offset + FPDiff;
2680 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2682 DstAddr = DAG.getFrameIndex(FI, getPointerTy());
2683 DstInfo = MachinePointerInfo::getFixedStack(FI);
2685 // Make sure any stack arguments overlapping with where we're storing
2686 // are loaded before this eventual operation. Otherwise they'll be
2688 Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
2690 SDValue PtrOff = DAG.getIntPtrConstant(Offset);
2692 DstAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2693 DstInfo = MachinePointerInfo::getStack(LocMemOffset);
2696 if (Outs[i].Flags.isByVal()) {
2698 DAG.getConstant(Outs[i].Flags.getByValSize(), MVT::i64);
2699 SDValue Cpy = DAG.getMemcpy(
2700 Chain, DL, DstAddr, Arg, SizeNode, Outs[i].Flags.getByValAlign(),
2702 /*AlwaysInline = */ false, DstInfo, MachinePointerInfo());
2704 MemOpChains.push_back(Cpy);
2706 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
2707 // promoted to a legal register type i32, we should truncate Arg back to
2709 if (VA.getValVT() == MVT::i1 || VA.getValVT() == MVT::i8 ||
2710 VA.getValVT() == MVT::i16)
2711 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Arg);
2714 DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo, false, false, 0);
2715 MemOpChains.push_back(Store);
2720 if (!MemOpChains.empty())
2721 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
2723 // Build a sequence of copy-to-reg nodes chained together with token chain
2724 // and flag operands which copy the outgoing args into the appropriate regs.
2726 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2727 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[i].first,
2728 RegsToPass[i].second, InFlag);
2729 InFlag = Chain.getValue(1);
2732 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2733 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2734 // node so that legalize doesn't hack it.
2735 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
2736 Subtarget->isTargetMachO()) {
2737 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2738 const GlobalValue *GV = G->getGlobal();
2739 bool InternalLinkage = GV->hasInternalLinkage();
2740 if (InternalLinkage)
2741 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2743 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0,
2745 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, getPointerTy(), Callee);
2747 } else if (ExternalSymbolSDNode *S =
2748 dyn_cast<ExternalSymbolSDNode>(Callee)) {
2749 const char *Sym = S->getSymbol();
2751 DAG.getTargetExternalSymbol(Sym, getPointerTy(), AArch64II::MO_GOT);
2752 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, getPointerTy(), Callee);
2754 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2755 const GlobalValue *GV = G->getGlobal();
2756 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2757 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2758 const char *Sym = S->getSymbol();
2759 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), 0);
2762 // We don't usually want to end the call-sequence here because we would tidy
2763 // the frame up *after* the call, however in the ABI-changing tail-call case
2764 // we've carefully laid out the parameters so that when sp is reset they'll be
2765 // in the correct location.
2766 if (IsTailCall && !IsSibCall) {
2767 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2768 DAG.getIntPtrConstant(0, true), InFlag, DL);
2769 InFlag = Chain.getValue(1);
2772 std::vector<SDValue> Ops;
2773 Ops.push_back(Chain);
2774 Ops.push_back(Callee);
2777 // Each tail call may have to adjust the stack by a different amount, so
2778 // this information must travel along with the operation for eventual
2779 // consumption by emitEpilogue.
2780 Ops.push_back(DAG.getTargetConstant(FPDiff, MVT::i32));
2783 // Add argument registers to the end of the list so that they are known live
2785 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2786 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2787 RegsToPass[i].second.getValueType()));
2789 // Add a register mask operand representing the call-preserved registers.
2790 const uint32_t *Mask;
2791 const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
2793 // For 'this' returns, use the X0-preserving mask if applicable
2794 Mask = TRI->getThisReturnPreservedMask(CallConv);
2796 IsThisReturn = false;
2797 Mask = TRI->getCallPreservedMask(CallConv);
2800 Mask = TRI->getCallPreservedMask(CallConv);
2802 assert(Mask && "Missing call preserved mask for calling convention");
2803 Ops.push_back(DAG.getRegisterMask(Mask));
2805 if (InFlag.getNode())
2806 Ops.push_back(InFlag);
2808 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2810 // If we're doing a tall call, use a TC_RETURN here rather than an
2811 // actual call instruction.
2813 return DAG.getNode(AArch64ISD::TC_RETURN, DL, NodeTys, Ops);
2815 // Returns a chain and a flag for retval copy to use.
2816 Chain = DAG.getNode(AArch64ISD::CALL, DL, NodeTys, Ops);
2817 InFlag = Chain.getValue(1);
2819 uint64_t CalleePopBytes = DoesCalleeRestoreStack(CallConv, TailCallOpt)
2820 ? RoundUpToAlignment(NumBytes, 16)
2823 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2824 DAG.getIntPtrConstant(CalleePopBytes, true),
2827 InFlag = Chain.getValue(1);
2829 // Handle result values, copying them out of physregs into vregs that we
2831 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2832 InVals, IsThisReturn,
2833 IsThisReturn ? OutVals[0] : SDValue());
2836 bool AArch64TargetLowering::CanLowerReturn(
2837 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2838 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2839 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2840 ? RetCC_AArch64_WebKit_JS
2841 : RetCC_AArch64_AAPCS;
2842 SmallVector<CCValAssign, 16> RVLocs;
2843 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2844 return CCInfo.CheckReturn(Outs, RetCC);
2848 AArch64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2850 const SmallVectorImpl<ISD::OutputArg> &Outs,
2851 const SmallVectorImpl<SDValue> &OutVals,
2852 SDLoc DL, SelectionDAG &DAG) const {
2853 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2854 ? RetCC_AArch64_WebKit_JS
2855 : RetCC_AArch64_AAPCS;
2856 SmallVector<CCValAssign, 16> RVLocs;
2857 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2859 CCInfo.AnalyzeReturn(Outs, RetCC);
2861 // Copy the result values into the output registers.
2863 SmallVector<SDValue, 4> RetOps(1, Chain);
2864 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
2865 ++i, ++realRVLocIdx) {
2866 CCValAssign &VA = RVLocs[i];
2867 assert(VA.isRegLoc() && "Can only return in registers!");
2868 SDValue Arg = OutVals[realRVLocIdx];
2870 switch (VA.getLocInfo()) {
2872 llvm_unreachable("Unknown loc info!");
2873 case CCValAssign::Full:
2874 if (Outs[i].ArgVT == MVT::i1) {
2875 // AAPCS requires i1 to be zero-extended to i8 by the producer of the
2876 // value. This is strictly redundant on Darwin (which uses "zeroext
2877 // i1"), but will be optimised out before ISel.
2878 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
2879 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2882 case CCValAssign::BCvt:
2883 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2887 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
2888 Flag = Chain.getValue(1);
2889 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2892 RetOps[0] = Chain; // Update chain.
2894 // Add the flag if we have it.
2896 RetOps.push_back(Flag);
2898 return DAG.getNode(AArch64ISD::RET_FLAG, DL, MVT::Other, RetOps);
2901 //===----------------------------------------------------------------------===//
2902 // Other Lowering Code
2903 //===----------------------------------------------------------------------===//
2905 SDValue AArch64TargetLowering::LowerGlobalAddress(SDValue Op,
2906 SelectionDAG &DAG) const {
2907 EVT PtrVT = getPointerTy();
2909 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
2910 const GlobalValue *GV = GN->getGlobal();
2911 unsigned char OpFlags =
2912 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
2914 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
2915 "unexpected offset in global node");
2917 // This also catched the large code model case for Darwin.
2918 if ((OpFlags & AArch64II::MO_GOT) != 0) {
2919 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
2920 // FIXME: Once remat is capable of dealing with instructions with register
2921 // operands, expand this into two nodes instead of using a wrapper node.
2922 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
2925 if ((OpFlags & AArch64II::MO_CONSTPOOL) != 0) {
2926 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
2927 "use of MO_CONSTPOOL only supported on small model");
2928 SDValue Hi = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, AArch64II::MO_PAGE);
2929 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
2930 unsigned char LoFlags = AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
2931 SDValue Lo = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, LoFlags);
2932 SDValue PoolAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
2933 SDValue GlobalAddr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), PoolAddr,
2934 MachinePointerInfo::getConstantPool(),
2935 /*isVolatile=*/ false,
2936 /*isNonTemporal=*/ true,
2937 /*isInvariant=*/ true, 8);
2938 if (GN->getOffset() != 0)
2939 return DAG.getNode(ISD::ADD, DL, PtrVT, GlobalAddr,
2940 DAG.getConstant(GN->getOffset(), PtrVT));
2944 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2945 const unsigned char MO_NC = AArch64II::MO_NC;
2947 AArch64ISD::WrapperLarge, DL, PtrVT,
2948 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G3),
2949 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
2950 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
2951 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
2953 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
2954 // the only correct model on Darwin.
2955 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2956 OpFlags | AArch64II::MO_PAGE);
2957 unsigned char LoFlags = OpFlags | AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
2958 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
2960 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
2961 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
2965 /// \brief Convert a TLS address reference into the correct sequence of loads
2966 /// and calls to compute the variable's address (for Darwin, currently) and
2967 /// return an SDValue containing the final node.
2969 /// Darwin only has one TLS scheme which must be capable of dealing with the
2970 /// fully general situation, in the worst case. This means:
2971 /// + "extern __thread" declaration.
2972 /// + Defined in a possibly unknown dynamic library.
2974 /// The general system is that each __thread variable has a [3 x i64] descriptor
2975 /// which contains information used by the runtime to calculate the address. The
2976 /// only part of this the compiler needs to know about is the first xword, which
2977 /// contains a function pointer that must be called with the address of the
2978 /// entire descriptor in "x0".
2980 /// Since this descriptor may be in a different unit, in general even the
2981 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
2983 /// adrp x0, _var@TLVPPAGE
2984 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
2985 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
2986 /// ; the function pointer
2987 /// blr x1 ; Uses descriptor address in x0
2988 /// ; Address of _var is now in x0.
2990 /// If the address of _var's descriptor *is* known to the linker, then it can
2991 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
2992 /// a slight efficiency gain.
2994 AArch64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
2995 SelectionDAG &DAG) const {
2996 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
2999 MVT PtrVT = getPointerTy();
3000 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3003 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3004 SDValue DescAddr = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TLVPAddr);
3006 // The first entry in the descriptor is a function pointer that we must call
3007 // to obtain the address of the variable.
3008 SDValue Chain = DAG.getEntryNode();
3009 SDValue FuncTLVGet =
3010 DAG.getLoad(MVT::i64, DL, Chain, DescAddr, MachinePointerInfo::getGOT(),
3011 false, true, true, 8);
3012 Chain = FuncTLVGet.getValue(1);
3014 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3015 MFI->setAdjustsStack(true);
3017 // TLS calls preserve all registers except those that absolutely must be
3018 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
3020 const uint32_t *Mask =
3021 Subtarget->getRegisterInfo()->getTLSCallPreservedMask();
3023 // Finally, we can make the call. This is just a degenerate version of a
3024 // normal AArch64 call node: x0 takes the address of the descriptor, and
3025 // returns the address of the variable in this thread.
3026 Chain = DAG.getCopyToReg(Chain, DL, AArch64::X0, DescAddr, SDValue());
3028 DAG.getNode(AArch64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
3029 Chain, FuncTLVGet, DAG.getRegister(AArch64::X0, MVT::i64),
3030 DAG.getRegisterMask(Mask), Chain.getValue(1));
3031 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Chain.getValue(1));
3034 /// When accessing thread-local variables under either the general-dynamic or
3035 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
3036 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
3037 /// is a function pointer to carry out the resolution. This function takes the
3038 /// address of the descriptor in X0 and returns the TPIDR_EL0 offset in X0. All
3039 /// other registers (except LR, NZCV) are preserved.
3041 /// Thus, the ideal call sequence on AArch64 is:
3043 /// adrp x0, :tlsdesc:thread_var
3044 /// ldr x8, [x0, :tlsdesc_lo12:thread_var]
3045 /// add x0, x0, :tlsdesc_lo12:thread_var
3046 /// .tlsdesccall thread_var
3048 /// (TPIDR_EL0 offset now in x0).
3050 /// The ".tlsdesccall" directive instructs the assembler to insert a particular
3051 /// relocation to help the linker relax this sequence if it turns out to be too
3054 /// FIXME: we currently produce an extra, duplicated, ADRP instruction, but this
3056 SDValue AArch64TargetLowering::LowerELFTLSDescCall(SDValue SymAddr,
3057 SDValue DescAddr, SDLoc DL,
3058 SelectionDAG &DAG) const {
3059 EVT PtrVT = getPointerTy();
3061 // The function we need to call is simply the first entry in the GOT for this
3062 // descriptor, load it in preparation.
3063 SDValue Func = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, SymAddr);
3065 // TLS calls preserve all registers except those that absolutely must be
3066 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
3068 const uint32_t *Mask =
3069 Subtarget->getRegisterInfo()->getTLSCallPreservedMask();
3071 // The function takes only one argument: the address of the descriptor itself
3073 SDValue Glue, Chain;
3074 Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, AArch64::X0, DescAddr, Glue);
3075 Glue = Chain.getValue(1);
3077 // We're now ready to populate the argument list, as with a normal call:
3078 SmallVector<SDValue, 6> Ops;
3079 Ops.push_back(Chain);
3080 Ops.push_back(Func);
3081 Ops.push_back(SymAddr);
3082 Ops.push_back(DAG.getRegister(AArch64::X0, PtrVT));
3083 Ops.push_back(DAG.getRegisterMask(Mask));
3084 Ops.push_back(Glue);
3086 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3087 Chain = DAG.getNode(AArch64ISD::TLSDESC_CALL, DL, NodeTys, Ops);
3088 Glue = Chain.getValue(1);
3090 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue);
3094 AArch64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
3095 SelectionDAG &DAG) const {
3096 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
3097 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
3098 "ELF TLS only supported in small memory model");
3099 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3101 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
3104 EVT PtrVT = getPointerTy();
3106 const GlobalValue *GV = GA->getGlobal();
3108 SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);
3110 if (Model == TLSModel::LocalExec) {
3111 SDValue HiVar = DAG.getTargetGlobalAddress(
3112 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_G1);
3113 SDValue LoVar = DAG.getTargetGlobalAddress(
3115 AArch64II::MO_TLS | AArch64II::MO_G0 | AArch64II::MO_NC);
3117 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVZXi, DL, PtrVT, HiVar,
3118 DAG.getTargetConstant(16, MVT::i32)),
3120 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKXi, DL, PtrVT, TPOff, LoVar,
3121 DAG.getTargetConstant(0, MVT::i32)),
3123 } else if (Model == TLSModel::InitialExec) {
3124 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3125 TPOff = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TPOff);
3126 } else if (Model == TLSModel::LocalDynamic) {
3127 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
3128 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
3129 // the beginning of the module's TLS region, followed by a DTPREL offset
3132 // These accesses will need deduplicating if there's more than one.
3133 AArch64FunctionInfo *MFI =
3134 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
3135 MFI->incNumLocalDynamicTLSAccesses();
3137 // Accesses used in this sequence go via the TLS descriptor which lives in
3138 // the GOT. Prepare an address we can use to handle this.
3139 SDValue HiDesc = DAG.getTargetExternalSymbol(
3140 "_TLS_MODULE_BASE_", PtrVT, AArch64II::MO_TLS | AArch64II::MO_PAGE);
3141 SDValue LoDesc = DAG.getTargetExternalSymbol(
3142 "_TLS_MODULE_BASE_", PtrVT,
3143 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3145 // First argument to the descriptor call is the address of the descriptor
3147 SDValue DescAddr = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, HiDesc);
3148 DescAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
3150 // The call needs a relocation too for linker relaxation. It doesn't make
3151 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3153 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
3156 // Now we can calculate the offset from TPIDR_EL0 to this module's
3157 // thread-local area.
3158 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
3160 // Now use :dtprel_whatever: operations to calculate this variable's offset
3161 // in its thread-storage area.
3162 SDValue HiVar = DAG.getTargetGlobalAddress(
3163 GV, DL, MVT::i64, 0, AArch64II::MO_TLS | AArch64II::MO_G1);
3164 SDValue LoVar = DAG.getTargetGlobalAddress(
3165 GV, DL, MVT::i64, 0,
3166 AArch64II::MO_TLS | AArch64II::MO_G0 | AArch64II::MO_NC);
3169 SDValue(DAG.getMachineNode(AArch64::MOVZXi, DL, PtrVT, HiVar,
3170 DAG.getTargetConstant(16, MVT::i32)),
3173 SDValue(DAG.getMachineNode(AArch64::MOVKXi, DL, PtrVT, DTPOff, LoVar,
3174 DAG.getTargetConstant(0, MVT::i32)),
3177 TPOff = DAG.getNode(ISD::ADD, DL, PtrVT, TPOff, DTPOff);
3178 } else if (Model == TLSModel::GeneralDynamic) {
3179 // Accesses used in this sequence go via the TLS descriptor which lives in
3180 // the GOT. Prepare an address we can use to handle this.
3181 SDValue HiDesc = DAG.getTargetGlobalAddress(
3182 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_PAGE);
3183 SDValue LoDesc = DAG.getTargetGlobalAddress(
3185 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3187 // First argument to the descriptor call is the address of the descriptor
3189 SDValue DescAddr = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, HiDesc);
3190 DescAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
3192 // The call needs a relocation too for linker relaxation. It doesn't make
3193 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3196 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3198 // Finally we can make a call to calculate the offset from tpidr_el0.
3199 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
3201 llvm_unreachable("Unsupported ELF TLS access model");
3203 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
3206 SDValue AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
3207 SelectionDAG &DAG) const {
3208 if (Subtarget->isTargetDarwin())
3209 return LowerDarwinGlobalTLSAddress(Op, DAG);
3210 else if (Subtarget->isTargetELF())
3211 return LowerELFGlobalTLSAddress(Op, DAG);
3213 llvm_unreachable("Unexpected platform trying to use TLS");
3215 SDValue AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3216 SDValue Chain = Op.getOperand(0);
3217 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3218 SDValue LHS = Op.getOperand(2);
3219 SDValue RHS = Op.getOperand(3);
3220 SDValue Dest = Op.getOperand(4);
3223 // Handle f128 first, since lowering it will result in comparing the return
3224 // value of a libcall against zero, which is just what the rest of LowerBR_CC
3225 // is expecting to deal with.
3226 if (LHS.getValueType() == MVT::f128) {
3227 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3229 // If softenSetCCOperands returned a scalar, we need to compare the result
3230 // against zero to select between true and false values.
3231 if (!RHS.getNode()) {
3232 RHS = DAG.getConstant(0, LHS.getValueType());
3237 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
3239 unsigned Opc = LHS.getOpcode();
3240 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
3241 cast<ConstantSDNode>(RHS)->isOne() &&
3242 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3243 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3244 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
3245 "Unexpected condition code.");
3246 // Only lower legal XALUO ops.
3247 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
3250 // The actual operation with overflow check.
3251 AArch64CC::CondCode OFCC;
3252 SDValue Value, Overflow;
3253 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, LHS.getValue(0), DAG);
3255 if (CC == ISD::SETNE)
3256 OFCC = getInvertedCondCode(OFCC);
3257 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3259 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3263 if (LHS.getValueType().isInteger()) {
3264 assert((LHS.getValueType() == RHS.getValueType()) &&
3265 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3267 // If the RHS of the comparison is zero, we can potentially fold this
3268 // to a specialized branch.
3269 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
3270 if (RHSC && RHSC->getZExtValue() == 0) {
3271 if (CC == ISD::SETEQ) {
3272 // See if we can use a TBZ to fold in an AND as well.
3273 // TBZ has a smaller branch displacement than CBZ. If the offset is
3274 // out of bounds, a late MI-layer pass rewrites branches.
3275 // 403.gcc is an example that hits this case.
3276 if (LHS.getOpcode() == ISD::AND &&
3277 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3278 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3279 SDValue Test = LHS.getOperand(0);
3280 uint64_t Mask = LHS.getConstantOperandVal(1);
3281 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, Test,
3282 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3285 return DAG.getNode(AArch64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
3286 } else if (CC == ISD::SETNE) {
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::TBNZ, dl, MVT::Other, Chain, Test,
3297 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3300 return DAG.getNode(AArch64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
3301 } else if (CC == ISD::SETLT && LHS.getOpcode() != ISD::AND) {
3302 // Don't combine AND since emitComparison converts the AND to an ANDS
3303 // (a.k.a. TST) and the test in the test bit and branch instruction
3304 // becomes redundant. This would also increase register pressure.
3305 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3306 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, LHS,
3307 DAG.getConstant(Mask, MVT::i64), Dest);
3310 if (RHSC && RHSC->getSExtValue() == -1 && CC == ISD::SETGT &&
3311 LHS.getOpcode() != ISD::AND) {
3312 // Don't combine AND since emitComparison converts the AND to an ANDS
3313 // (a.k.a. TST) and the test in the test bit and branch instruction
3314 // becomes redundant. This would also increase register pressure.
3315 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3316 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, LHS,
3317 DAG.getConstant(Mask, MVT::i64), Dest);
3321 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3322 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3326 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3328 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3329 // clean. Some of them require two branches to implement.
3330 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3331 AArch64CC::CondCode CC1, CC2;
3332 changeFPCCToAArch64CC(CC, CC1, CC2);
3333 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3335 DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
3336 if (CC2 != AArch64CC::AL) {
3337 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3338 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
3345 SDValue AArch64TargetLowering::LowerFCOPYSIGN(SDValue Op,
3346 SelectionDAG &DAG) const {
3347 EVT VT = Op.getValueType();
3350 SDValue In1 = Op.getOperand(0);
3351 SDValue In2 = Op.getOperand(1);
3352 EVT SrcVT = In2.getValueType();
3354 if (SrcVT == MVT::f32 && VT == MVT::f64)
3355 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
3356 else if (SrcVT == MVT::f64 && VT == MVT::f32)
3357 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0));
3359 // FIXME: Src type is different, bail out for now. Can VT really be a
3366 SDValue EltMask, VecVal1, VecVal2;
3367 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
3370 EltMask = DAG.getConstant(0x80000000ULL, EltVT);
3372 if (!VT.isVector()) {
3373 VecVal1 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3374 DAG.getUNDEF(VecVT), In1);
3375 VecVal2 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3376 DAG.getUNDEF(VecVT), In2);
3378 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3379 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3381 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
3385 // We want to materialize a mask with the the high bit set, but the AdvSIMD
3386 // immediate moves cannot materialize that in a single instruction for
3387 // 64-bit elements. Instead, materialize zero and then negate it.
3388 EltMask = DAG.getConstant(0, EltVT);
3390 if (!VT.isVector()) {
3391 VecVal1 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3392 DAG.getUNDEF(VecVT), In1);
3393 VecVal2 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3394 DAG.getUNDEF(VecVT), In2);
3396 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3397 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3400 llvm_unreachable("Invalid type for copysign!");
3403 std::vector<SDValue> BuildVectorOps;
3404 for (unsigned i = 0; i < VecVT.getVectorNumElements(); ++i)
3405 BuildVectorOps.push_back(EltMask);
3407 SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT, BuildVectorOps);
3409 // If we couldn't materialize the mask above, then the mask vector will be
3410 // the zero vector, and we need to negate it here.
3411 if (VT == MVT::f64 || VT == MVT::v2f64) {
3412 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
3413 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
3414 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
3418 DAG.getNode(AArch64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
3421 return DAG.getTargetExtractSubreg(AArch64::ssub, DL, VT, Sel);
3422 else if (VT == MVT::f64)
3423 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, VT, Sel);
3425 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
3428 SDValue AArch64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
3429 if (DAG.getMachineFunction().getFunction()->hasFnAttribute(
3430 Attribute::NoImplicitFloat))
3433 if (!Subtarget->hasNEON())
3436 // While there is no integer popcount instruction, it can
3437 // be more efficiently lowered to the following sequence that uses
3438 // AdvSIMD registers/instructions as long as the copies to/from
3439 // the AdvSIMD registers are cheap.
3440 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
3441 // CNT V0.8B, V0.8B // 8xbyte pop-counts
3442 // ADDV B0, V0.8B // sum 8xbyte pop-counts
3443 // UMOV X0, V0.B[0] // copy byte result back to integer reg
3444 SDValue Val = Op.getOperand(0);
3446 EVT VT = Op.getValueType();
3449 Val = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Val);
3450 Val = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
3452 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, Val);
3453 SDValue UaddLV = DAG.getNode(
3454 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
3455 DAG.getConstant(Intrinsic::aarch64_neon_uaddlv, MVT::i32), CtPop);
3458 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
3462 SDValue AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3464 if (Op.getValueType().isVector())
3465 return LowerVSETCC(Op, DAG);
3467 SDValue LHS = Op.getOperand(0);
3468 SDValue RHS = Op.getOperand(1);
3469 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3472 // We chose ZeroOrOneBooleanContents, so use zero and one.
3473 EVT VT = Op.getValueType();
3474 SDValue TVal = DAG.getConstant(1, VT);
3475 SDValue FVal = DAG.getConstant(0, VT);
3477 // Handle f128 first, since one possible outcome is a normal integer
3478 // comparison which gets picked up by the next if statement.
3479 if (LHS.getValueType() == MVT::f128) {
3480 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3482 // If softenSetCCOperands returned a scalar, use it.
3483 if (!RHS.getNode()) {
3484 assert(LHS.getValueType() == Op.getValueType() &&
3485 "Unexpected setcc expansion!");
3490 if (LHS.getValueType().isInteger()) {
3493 getAArch64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
3495 // Note that we inverted the condition above, so we reverse the order of
3496 // the true and false operands here. This will allow the setcc to be
3497 // matched to a single CSINC instruction.
3498 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
3501 // Now we know we're dealing with FP values.
3502 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3504 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3505 // and do the comparison.
3506 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3508 AArch64CC::CondCode CC1, CC2;
3509 changeFPCCToAArch64CC(CC, CC1, CC2);
3510 if (CC2 == AArch64CC::AL) {
3511 changeFPCCToAArch64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
3512 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3514 // Note that we inverted the condition above, so we reverse the order of
3515 // the true and false operands here. This will allow the setcc to be
3516 // matched to a single CSINC instruction.
3517 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
3519 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't
3520 // totally clean. Some of them require two CSELs to implement. As is in
3521 // this case, we emit the first CSEL and then emit a second using the output
3522 // of the first as the RHS. We're effectively OR'ing the two CC's together.
3524 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
3525 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3527 DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3529 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3530 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3534 /// A SELECT_CC operation is really some kind of max or min if both values being
3535 /// compared are, in some sense, equal to the results in either case. However,
3536 /// it is permissible to compare f32 values and produce directly extended f64
3539 /// Extending the comparison operands would also be allowed, but is less likely
3540 /// to happen in practice since their use is right here. Note that truncate
3541 /// operations would *not* be semantically equivalent.
3542 static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) {
3546 ConstantFPSDNode *CCmp = dyn_cast<ConstantFPSDNode>(Cmp);
3547 ConstantFPSDNode *CResult = dyn_cast<ConstantFPSDNode>(Result);
3548 if (CCmp && CResult && Cmp.getValueType() == MVT::f32 &&
3549 Result.getValueType() == MVT::f64) {
3551 APFloat CmpVal = CCmp->getValueAPF();
3552 CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy);
3553 return CResult->getValueAPF().bitwiseIsEqual(CmpVal);
3556 return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp;
3559 SDValue AArch64TargetLowering::LowerSELECT(SDValue Op,
3560 SelectionDAG &DAG) const {
3561 SDValue CC = Op->getOperand(0);
3562 SDValue TVal = Op->getOperand(1);
3563 SDValue FVal = Op->getOperand(2);
3566 unsigned Opc = CC.getOpcode();
3567 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
3569 if (CC.getResNo() == 1 &&
3570 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3571 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3572 // Only lower legal XALUO ops.
3573 if (!DAG.getTargetLoweringInfo().isTypeLegal(CC->getValueType(0)))
3576 AArch64CC::CondCode OFCC;
3577 SDValue Value, Overflow;
3578 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, CC.getValue(0), DAG);
3579 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3581 return DAG.getNode(AArch64ISD::CSEL, DL, Op.getValueType(), TVal, FVal,
3585 if (CC.getOpcode() == ISD::SETCC)
3586 return DAG.getSelectCC(DL, CC.getOperand(0), CC.getOperand(1), TVal, FVal,
3587 cast<CondCodeSDNode>(CC.getOperand(2))->get());
3589 return DAG.getSelectCC(DL, CC, DAG.getConstant(0, CC.getValueType()), TVal,
3593 SDValue AArch64TargetLowering::LowerSELECT_CC(SDValue Op,
3594 SelectionDAG &DAG) const {
3595 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3596 SDValue LHS = Op.getOperand(0);
3597 SDValue RHS = Op.getOperand(1);
3598 SDValue TVal = Op.getOperand(2);
3599 SDValue FVal = Op.getOperand(3);
3602 // Handle f128 first, because it will result in a comparison of some RTLIB
3603 // call result against zero.
3604 if (LHS.getValueType() == MVT::f128) {
3605 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3607 // If softenSetCCOperands returned a scalar, we need to compare the result
3608 // against zero to select between true and false values.
3609 if (!RHS.getNode()) {
3610 RHS = DAG.getConstant(0, LHS.getValueType());
3615 // Handle integers first.
3616 if (LHS.getValueType().isInteger()) {
3617 assert((LHS.getValueType() == RHS.getValueType()) &&
3618 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3620 unsigned Opcode = AArch64ISD::CSEL;
3622 // If both the TVal and the FVal are constants, see if we can swap them in
3623 // order to for a CSINV or CSINC out of them.
3624 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3625 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3627 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3628 std::swap(TVal, FVal);
3629 std::swap(CTVal, CFVal);
3630 CC = ISD::getSetCCInverse(CC, true);
3631 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3632 std::swap(TVal, FVal);
3633 std::swap(CTVal, CFVal);
3634 CC = ISD::getSetCCInverse(CC, true);
3635 } else if (TVal.getOpcode() == ISD::XOR) {
3636 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3637 // with a CSINV rather than a CSEL.
3638 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
3640 if (CVal && CVal->isAllOnesValue()) {
3641 std::swap(TVal, FVal);
3642 std::swap(CTVal, CFVal);
3643 CC = ISD::getSetCCInverse(CC, true);
3645 } else if (TVal.getOpcode() == ISD::SUB) {
3646 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3647 // that we can match with a CSNEG rather than a CSEL.
3648 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
3650 if (CVal && CVal->isNullValue()) {
3651 std::swap(TVal, FVal);
3652 std::swap(CTVal, CFVal);
3653 CC = ISD::getSetCCInverse(CC, true);
3655 } else if (CTVal && CFVal) {
3656 const int64_t TrueVal = CTVal->getSExtValue();
3657 const int64_t FalseVal = CFVal->getSExtValue();
3660 // If both TVal and FVal are constants, see if FVal is the
3661 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3662 // instead of a CSEL in that case.
3663 if (TrueVal == ~FalseVal) {
3664 Opcode = AArch64ISD::CSINV;
3665 } else if (TrueVal == -FalseVal) {
3666 Opcode = AArch64ISD::CSNEG;
3667 } else if (TVal.getValueType() == MVT::i32) {
3668 // If our operands are only 32-bit wide, make sure we use 32-bit
3669 // arithmetic for the check whether we can use CSINC. This ensures that
3670 // the addition in the check will wrap around properly in case there is
3671 // an overflow (which would not be the case if we do the check with
3672 // 64-bit arithmetic).
3673 const uint32_t TrueVal32 = CTVal->getZExtValue();
3674 const uint32_t FalseVal32 = CFVal->getZExtValue();
3676 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3677 Opcode = AArch64ISD::CSINC;
3679 if (TrueVal32 > FalseVal32) {
3683 // 64-bit check whether we can use CSINC.
3684 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3685 Opcode = AArch64ISD::CSINC;
3687 if (TrueVal > FalseVal) {
3692 // Swap TVal and FVal if necessary.
3694 std::swap(TVal, FVal);
3695 std::swap(CTVal, CFVal);
3696 CC = ISD::getSetCCInverse(CC, true);
3699 if (Opcode != AArch64ISD::CSEL) {
3700 // Drop FVal since we can get its value by simply inverting/negating
3707 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3709 EVT VT = Op.getValueType();
3710 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3713 // Now we know we're dealing with FP values.
3714 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3715 assert(LHS.getValueType() == RHS.getValueType());
3716 EVT VT = Op.getValueType();
3718 // Try to match this select into a max/min operation, which have dedicated
3719 // opcode in the instruction set.
3720 // FIXME: This is not correct in the presence of NaNs, so we only enable this
3722 if (getTargetMachine().Options.NoNaNsFPMath) {
3723 SDValue MinMaxLHS = TVal, MinMaxRHS = FVal;
3724 if (selectCCOpsAreFMaxCompatible(LHS, MinMaxRHS) &&
3725 selectCCOpsAreFMaxCompatible(RHS, MinMaxLHS)) {
3726 CC = ISD::getSetCCSwappedOperands(CC);
3727 std::swap(MinMaxLHS, MinMaxRHS);
3730 if (selectCCOpsAreFMaxCompatible(LHS, MinMaxLHS) &&
3731 selectCCOpsAreFMaxCompatible(RHS, MinMaxRHS)) {
3741 return DAG.getNode(AArch64ISD::FMAX, dl, VT, MinMaxLHS, MinMaxRHS);
3749 return DAG.getNode(AArch64ISD::FMIN, dl, VT, MinMaxLHS, MinMaxRHS);
3755 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3756 // and do the comparison.
3757 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3759 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3760 // clean. Some of them require two CSELs to implement.
3761 AArch64CC::CondCode CC1, CC2;
3762 changeFPCCToAArch64CC(CC, CC1, CC2);
3763 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3764 SDValue CS1 = DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3766 // If we need a second CSEL, emit it, using the output of the first as the
3767 // RHS. We're effectively OR'ing the two CC's together.
3768 if (CC2 != AArch64CC::AL) {
3769 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3770 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3773 // Otherwise, return the output of the first CSEL.
3777 SDValue AArch64TargetLowering::LowerJumpTable(SDValue Op,
3778 SelectionDAG &DAG) const {
3779 // Jump table entries as PC relative offsets. No additional tweaking
3780 // is necessary here. Just get the address of the jump table.
3781 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3782 EVT PtrVT = getPointerTy();
3785 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3786 !Subtarget->isTargetMachO()) {
3787 const unsigned char MO_NC = AArch64II::MO_NC;
3789 AArch64ISD::WrapperLarge, DL, PtrVT,
3790 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G3),
3791 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G2 | MO_NC),
3792 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G1 | MO_NC),
3793 DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3794 AArch64II::MO_G0 | MO_NC));
3798 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_PAGE);
3799 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3800 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3801 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3802 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3805 SDValue AArch64TargetLowering::LowerConstantPool(SDValue Op,
3806 SelectionDAG &DAG) const {
3807 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3808 EVT PtrVT = getPointerTy();
3811 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3812 // Use the GOT for the large code model on iOS.
3813 if (Subtarget->isTargetMachO()) {
3814 SDValue GotAddr = DAG.getTargetConstantPool(
3815 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3817 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
3820 const unsigned char MO_NC = AArch64II::MO_NC;
3822 AArch64ISD::WrapperLarge, DL, PtrVT,
3823 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3824 CP->getOffset(), AArch64II::MO_G3),
3825 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3826 CP->getOffset(), AArch64II::MO_G2 | MO_NC),
3827 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3828 CP->getOffset(), AArch64II::MO_G1 | MO_NC),
3829 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3830 CP->getOffset(), AArch64II::MO_G0 | MO_NC));
3832 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
3833 // ELF, the only valid one on Darwin.
3835 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3836 CP->getOffset(), AArch64II::MO_PAGE);
3837 SDValue Lo = DAG.getTargetConstantPool(
3838 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3839 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3841 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3842 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3846 SDValue AArch64TargetLowering::LowerBlockAddress(SDValue Op,
3847 SelectionDAG &DAG) const {
3848 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
3849 EVT PtrVT = getPointerTy();
3851 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3852 !Subtarget->isTargetMachO()) {
3853 const unsigned char MO_NC = AArch64II::MO_NC;
3855 AArch64ISD::WrapperLarge, DL, PtrVT,
3856 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G3),
3857 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
3858 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
3859 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
3861 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGE);
3862 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGEOFF |
3864 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3865 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3869 SDValue AArch64TargetLowering::LowerDarwin_VASTART(SDValue Op,
3870 SelectionDAG &DAG) const {
3871 AArch64FunctionInfo *FuncInfo =
3872 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
3876 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3877 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3878 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
3879 MachinePointerInfo(SV), false, false, 0);
3882 SDValue AArch64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
3883 SelectionDAG &DAG) const {
3884 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3885 // Standard, section B.3.
3886 MachineFunction &MF = DAG.getMachineFunction();
3887 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
3890 SDValue Chain = Op.getOperand(0);
3891 SDValue VAList = Op.getOperand(1);
3892 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3893 SmallVector<SDValue, 4> MemOps;
3895 // void *__stack at offset 0
3897 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3898 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3899 MachinePointerInfo(SV), false, false, 8));
3901 // void *__gr_top at offset 8
3902 int GPRSize = FuncInfo->getVarArgsGPRSize();
3904 SDValue GRTop, GRTopAddr;
3906 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3907 DAG.getConstant(8, getPointerTy()));
3909 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), getPointerTy());
3910 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
3911 DAG.getConstant(GPRSize, getPointerTy()));
3913 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3914 MachinePointerInfo(SV, 8), false, false, 8));
3917 // void *__vr_top at offset 16
3918 int FPRSize = FuncInfo->getVarArgsFPRSize();
3920 SDValue VRTop, VRTopAddr;
3921 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3922 DAG.getConstant(16, getPointerTy()));
3924 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), getPointerTy());
3925 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
3926 DAG.getConstant(FPRSize, getPointerTy()));
3928 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3929 MachinePointerInfo(SV, 16), false, false, 8));
3932 // int __gr_offs at offset 24
3933 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3934 DAG.getConstant(24, getPointerTy()));
3935 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
3936 GROffsAddr, MachinePointerInfo(SV, 24), false,
3939 // int __vr_offs at offset 28
3940 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3941 DAG.getConstant(28, getPointerTy()));
3942 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
3943 VROffsAddr, MachinePointerInfo(SV, 28), false,
3946 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
3949 SDValue AArch64TargetLowering::LowerVASTART(SDValue Op,
3950 SelectionDAG &DAG) const {
3951 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
3952 : LowerAAPCS_VASTART(Op, DAG);
3955 SDValue AArch64TargetLowering::LowerVACOPY(SDValue Op,
3956 SelectionDAG &DAG) const {
3957 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
3959 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
3960 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3961 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
3963 return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op), Op.getOperand(1),
3964 Op.getOperand(2), DAG.getConstant(VaListSize, MVT::i32),
3965 8, false, false, MachinePointerInfo(DestSV),
3966 MachinePointerInfo(SrcSV));
3969 SDValue AArch64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
3970 assert(Subtarget->isTargetDarwin() &&
3971 "automatic va_arg instruction only works on Darwin");
3973 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3974 EVT VT = Op.getValueType();
3976 SDValue Chain = Op.getOperand(0);
3977 SDValue Addr = Op.getOperand(1);
3978 unsigned Align = Op.getConstantOperandVal(3);
3980 SDValue VAList = DAG.getLoad(getPointerTy(), DL, Chain, Addr,
3981 MachinePointerInfo(V), false, false, false, 0);
3982 Chain = VAList.getValue(1);
3985 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
3986 VAList = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3987 DAG.getConstant(Align - 1, getPointerTy()));
3988 VAList = DAG.getNode(ISD::AND, DL, getPointerTy(), VAList,
3989 DAG.getConstant(-(int64_t)Align, getPointerTy()));
3992 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
3993 uint64_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
3995 // Scalar integer and FP values smaller than 64 bits are implicitly extended
3996 // up to 64 bits. At the very least, we have to increase the striding of the
3997 // vaargs list to match this, and for FP values we need to introduce
3998 // FP_ROUND nodes as well.
3999 if (VT.isInteger() && !VT.isVector())
4001 bool NeedFPTrunc = false;
4002 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
4007 // Increment the pointer, VAList, to the next vaarg
4008 SDValue VANext = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
4009 DAG.getConstant(ArgSize, getPointerTy()));
4010 // Store the incremented VAList to the legalized pointer
4011 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
4014 // Load the actual argument out of the pointer VAList
4016 // Load the value as an f64.
4017 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
4018 MachinePointerInfo(), false, false, false, 0);
4019 // Round the value down to an f32.
4020 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
4021 DAG.getIntPtrConstant(1));
4022 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
4023 // Merge the rounded value with the chain output of the load.
4024 return DAG.getMergeValues(Ops, DL);
4027 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
4031 SDValue AArch64TargetLowering::LowerFRAMEADDR(SDValue Op,
4032 SelectionDAG &DAG) const {
4033 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4034 MFI->setFrameAddressIsTaken(true);
4036 EVT VT = Op.getValueType();
4038 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4040 DAG.getCopyFromReg(DAG.getEntryNode(), DL, AArch64::FP, VT);
4042 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
4043 MachinePointerInfo(), false, false, false, 0);
4047 // FIXME? Maybe this could be a TableGen attribute on some registers and
4048 // this table could be generated automatically from RegInfo.
4049 unsigned AArch64TargetLowering::getRegisterByName(const char* RegName,
4051 unsigned Reg = StringSwitch<unsigned>(RegName)
4052 .Case("sp", AArch64::SP)
4056 report_fatal_error("Invalid register name global variable");
4059 SDValue AArch64TargetLowering::LowerRETURNADDR(SDValue Op,
4060 SelectionDAG &DAG) const {
4061 MachineFunction &MF = DAG.getMachineFunction();
4062 MachineFrameInfo *MFI = MF.getFrameInfo();
4063 MFI->setReturnAddressIsTaken(true);
4065 EVT VT = Op.getValueType();
4067 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4069 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4070 SDValue Offset = DAG.getConstant(8, getPointerTy());
4071 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
4072 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
4073 MachinePointerInfo(), false, false, false, 0);
4076 // Return LR, which contains the return address. Mark it an implicit live-in.
4077 unsigned Reg = MF.addLiveIn(AArch64::LR, &AArch64::GPR64RegClass);
4078 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
4081 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4082 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4083 SDValue AArch64TargetLowering::LowerShiftRightParts(SDValue Op,
4084 SelectionDAG &DAG) const {
4085 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4086 EVT VT = Op.getValueType();
4087 unsigned VTBits = VT.getSizeInBits();
4089 SDValue ShOpLo = Op.getOperand(0);
4090 SDValue ShOpHi = Op.getOperand(1);
4091 SDValue ShAmt = Op.getOperand(2);
4093 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4095 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4097 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4098 DAG.getConstant(VTBits, MVT::i64), ShAmt);
4099 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4100 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4101 DAG.getConstant(VTBits, MVT::i64));
4102 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4104 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
4105 ISD::SETGE, dl, DAG);
4106 SDValue CCVal = DAG.getConstant(AArch64CC::GE, MVT::i32);
4108 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4109 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4111 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4113 // AArch64 shifts larger than the register width are wrapped rather than
4114 // clamped, so we can't just emit "hi >> x".
4115 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4116 SDValue TrueValHi = Opc == ISD::SRA
4117 ? DAG.getNode(Opc, dl, VT, ShOpHi,
4118 DAG.getConstant(VTBits - 1, MVT::i64))
4119 : DAG.getConstant(0, VT);
4121 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
4123 SDValue Ops[2] = { Lo, Hi };
4124 return DAG.getMergeValues(Ops, dl);
4127 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4128 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4129 SDValue AArch64TargetLowering::LowerShiftLeftParts(SDValue Op,
4130 SelectionDAG &DAG) const {
4131 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4132 EVT VT = Op.getValueType();
4133 unsigned VTBits = VT.getSizeInBits();
4135 SDValue ShOpLo = Op.getOperand(0);
4136 SDValue ShOpHi = Op.getOperand(1);
4137 SDValue ShAmt = Op.getOperand(2);
4140 assert(Op.getOpcode() == ISD::SHL_PARTS);
4141 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4142 DAG.getConstant(VTBits, MVT::i64), ShAmt);
4143 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4144 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4145 DAG.getConstant(VTBits, MVT::i64));
4146 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4147 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4149 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4151 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
4152 ISD::SETGE, dl, DAG);
4153 SDValue CCVal = DAG.getConstant(AArch64CC::GE, MVT::i32);
4155 DAG.getNode(AArch64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
4157 // AArch64 shifts of larger than register sizes are wrapped rather than
4158 // clamped, so we can't just emit "lo << a" if a is too big.
4159 SDValue TrueValLo = DAG.getConstant(0, VT);
4160 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4162 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4164 SDValue Ops[2] = { Lo, Hi };
4165 return DAG.getMergeValues(Ops, dl);
4168 bool AArch64TargetLowering::isOffsetFoldingLegal(
4169 const GlobalAddressSDNode *GA) const {
4170 // The AArch64 target doesn't support folding offsets into global addresses.
4174 bool AArch64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4175 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
4176 // FIXME: We should be able to handle f128 as well with a clever lowering.
4177 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
4181 return AArch64_AM::getFP64Imm(Imm) != -1;
4182 else if (VT == MVT::f32)
4183 return AArch64_AM::getFP32Imm(Imm) != -1;
4187 //===----------------------------------------------------------------------===//
4188 // AArch64 Optimization Hooks
4189 //===----------------------------------------------------------------------===//
4191 //===----------------------------------------------------------------------===//
4192 // AArch64 Inline Assembly Support
4193 //===----------------------------------------------------------------------===//
4195 // Table of Constraints
4196 // TODO: This is the current set of constraints supported by ARM for the
4197 // compiler, not all of them may make sense, e.g. S may be difficult to support.
4199 // r - A general register
4200 // w - An FP/SIMD register of some size in the range v0-v31
4201 // x - An FP/SIMD register of some size in the range v0-v15
4202 // I - Constant that can be used with an ADD instruction
4203 // J - Constant that can be used with a SUB instruction
4204 // K - Constant that can be used with a 32-bit logical instruction
4205 // L - Constant that can be used with a 64-bit logical instruction
4206 // M - Constant that can be used as a 32-bit MOV immediate
4207 // N - Constant that can be used as a 64-bit MOV immediate
4208 // Q - A memory reference with base register and no offset
4209 // S - A symbolic address
4210 // Y - Floating point constant zero
4211 // Z - Integer constant zero
4213 // Note that general register operands will be output using their 64-bit x
4214 // register name, whatever the size of the variable, unless the asm operand
4215 // is prefixed by the %w modifier. Floating-point and SIMD register operands
4216 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
4219 /// getConstraintType - Given a constraint letter, return the type of
4220 /// constraint it is for this target.
4221 AArch64TargetLowering::ConstraintType
4222 AArch64TargetLowering::getConstraintType(const std::string &Constraint) const {
4223 if (Constraint.size() == 1) {
4224 switch (Constraint[0]) {
4231 return C_RegisterClass;
4232 // An address with a single base register. Due to the way we
4233 // currently handle addresses it is the same as 'r'.
4238 return TargetLowering::getConstraintType(Constraint);
4241 /// Examine constraint type and operand type and determine a weight value.
4242 /// This object must already have been set up with the operand type
4243 /// and the current alternative constraint selected.
4244 TargetLowering::ConstraintWeight
4245 AArch64TargetLowering::getSingleConstraintMatchWeight(
4246 AsmOperandInfo &info, const char *constraint) const {
4247 ConstraintWeight weight = CW_Invalid;
4248 Value *CallOperandVal = info.CallOperandVal;
4249 // If we don't have a value, we can't do a match,
4250 // but allow it at the lowest weight.
4251 if (!CallOperandVal)
4253 Type *type = CallOperandVal->getType();
4254 // Look at the constraint type.
4255 switch (*constraint) {
4257 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
4261 if (type->isFloatingPointTy() || type->isVectorTy())
4262 weight = CW_Register;
4265 weight = CW_Constant;
4271 std::pair<unsigned, const TargetRegisterClass *>
4272 AArch64TargetLowering::getRegForInlineAsmConstraint(
4273 const std::string &Constraint, MVT VT) const {
4274 if (Constraint.size() == 1) {
4275 switch (Constraint[0]) {
4277 if (VT.getSizeInBits() == 64)
4278 return std::make_pair(0U, &AArch64::GPR64commonRegClass);
4279 return std::make_pair(0U, &AArch64::GPR32commonRegClass);
4282 return std::make_pair(0U, &AArch64::FPR32RegClass);
4283 if (VT.getSizeInBits() == 64)
4284 return std::make_pair(0U, &AArch64::FPR64RegClass);
4285 if (VT.getSizeInBits() == 128)
4286 return std::make_pair(0U, &AArch64::FPR128RegClass);
4288 // The instructions that this constraint is designed for can
4289 // only take 128-bit registers so just use that regclass.
4291 if (VT.getSizeInBits() == 128)
4292 return std::make_pair(0U, &AArch64::FPR128_loRegClass);
4296 if (StringRef("{cc}").equals_lower(Constraint))
4297 return std::make_pair(unsigned(AArch64::NZCV), &AArch64::CCRRegClass);
4299 // Use the default implementation in TargetLowering to convert the register
4300 // constraint into a member of a register class.
4301 std::pair<unsigned, const TargetRegisterClass *> Res;
4302 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4304 // Not found as a standard register?
4306 unsigned Size = Constraint.size();
4307 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
4308 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
4309 const std::string Reg =
4310 std::string(&Constraint[2], &Constraint[Size - 1]);
4311 int RegNo = atoi(Reg.c_str());
4312 if (RegNo >= 0 && RegNo <= 31) {
4313 // v0 - v31 are aliases of q0 - q31.
4314 // By default we'll emit v0-v31 for this unless there's a modifier where
4315 // we'll emit the correct register as well.
4316 Res.first = AArch64::FPR128RegClass.getRegister(RegNo);
4317 Res.second = &AArch64::FPR128RegClass;
4325 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4326 /// vector. If it is invalid, don't add anything to Ops.
4327 void AArch64TargetLowering::LowerAsmOperandForConstraint(
4328 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
4329 SelectionDAG &DAG) const {
4332 // Currently only support length 1 constraints.
4333 if (Constraint.length() != 1)
4336 char ConstraintLetter = Constraint[0];
4337 switch (ConstraintLetter) {
4341 // This set of constraints deal with valid constants for various instructions.
4342 // Validate and return a target constant for them if we can.
4344 // 'z' maps to xzr or wzr so it needs an input of 0.
4345 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4346 if (!C || C->getZExtValue() != 0)
4349 if (Op.getValueType() == MVT::i64)
4350 Result = DAG.getRegister(AArch64::XZR, MVT::i64);
4352 Result = DAG.getRegister(AArch64::WZR, MVT::i32);
4362 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4366 // Grab the value and do some validation.
4367 uint64_t CVal = C->getZExtValue();
4368 switch (ConstraintLetter) {
4369 // The I constraint applies only to simple ADD or SUB immediate operands:
4370 // i.e. 0 to 4095 with optional shift by 12
4371 // The J constraint applies only to ADD or SUB immediates that would be
4372 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
4373 // instruction [or vice versa], in other words -1 to -4095 with optional
4374 // left shift by 12.
4376 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
4380 uint64_t NVal = -C->getSExtValue();
4381 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal)) {
4382 CVal = C->getSExtValue();
4387 // The K and L constraints apply *only* to logical immediates, including
4388 // what used to be the MOVI alias for ORR (though the MOVI alias has now
4389 // been removed and MOV should be used). So these constraints have to
4390 // distinguish between bit patterns that are valid 32-bit or 64-bit
4391 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
4392 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
4395 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4399 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4402 // The M and N constraints are a superset of K and L respectively, for use
4403 // with the MOV (immediate) alias. As well as the logical immediates they
4404 // also match 32 or 64-bit immediates that can be loaded either using a
4405 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
4406 // (M) or 64-bit 0x1234000000000000 (N) etc.
4407 // As a note some of this code is liberally stolen from the asm parser.
4409 if (!isUInt<32>(CVal))
4411 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4413 if ((CVal & 0xFFFF) == CVal)
4415 if ((CVal & 0xFFFF0000ULL) == CVal)
4417 uint64_t NCVal = ~(uint32_t)CVal;
4418 if ((NCVal & 0xFFFFULL) == NCVal)
4420 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4425 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4427 if ((CVal & 0xFFFFULL) == CVal)
4429 if ((CVal & 0xFFFF0000ULL) == CVal)
4431 if ((CVal & 0xFFFF00000000ULL) == CVal)
4433 if ((CVal & 0xFFFF000000000000ULL) == CVal)
4435 uint64_t NCVal = ~CVal;
4436 if ((NCVal & 0xFFFFULL) == NCVal)
4438 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4440 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
4442 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
4450 // All assembler immediates are 64-bit integers.
4451 Result = DAG.getTargetConstant(CVal, MVT::i64);
4455 if (Result.getNode()) {
4456 Ops.push_back(Result);
4460 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4463 //===----------------------------------------------------------------------===//
4464 // AArch64 Advanced SIMD Support
4465 //===----------------------------------------------------------------------===//
4467 /// WidenVector - Given a value in the V64 register class, produce the
4468 /// equivalent value in the V128 register class.
4469 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
4470 EVT VT = V64Reg.getValueType();
4471 unsigned NarrowSize = VT.getVectorNumElements();
4472 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4473 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
4476 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
4477 V64Reg, DAG.getConstant(0, MVT::i32));
4480 /// getExtFactor - Determine the adjustment factor for the position when
4481 /// generating an "extract from vector registers" instruction.
4482 static unsigned getExtFactor(SDValue &V) {
4483 EVT EltType = V.getValueType().getVectorElementType();
4484 return EltType.getSizeInBits() / 8;
4487 /// NarrowVector - Given a value in the V128 register class, produce the
4488 /// equivalent value in the V64 register class.
4489 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
4490 EVT VT = V128Reg.getValueType();
4491 unsigned WideSize = VT.getVectorNumElements();
4492 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4493 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
4496 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, NarrowTy, V128Reg);
4499 // Gather data to see if the operation can be modelled as a
4500 // shuffle in combination with VEXTs.
4501 SDValue AArch64TargetLowering::ReconstructShuffle(SDValue Op,
4502 SelectionDAG &DAG) const {
4503 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
4505 EVT VT = Op.getValueType();
4506 unsigned NumElts = VT.getVectorNumElements();
4508 struct ShuffleSourceInfo {
4513 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
4514 // be compatible with the shuffle we intend to construct. As a result
4515 // ShuffleVec will be some sliding window into the original Vec.
4518 // Code should guarantee that element i in Vec starts at element "WindowBase
4519 // + i * WindowScale in ShuffleVec".
4523 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
4524 ShuffleSourceInfo(SDValue Vec)
4525 : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
4529 // First gather all vectors used as an immediate source for this BUILD_VECTOR
4531 SmallVector<ShuffleSourceInfo, 2> Sources;
4532 for (unsigned i = 0; i < NumElts; ++i) {
4533 SDValue V = Op.getOperand(i);
4534 if (V.getOpcode() == ISD::UNDEF)
4536 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4537 // A shuffle can only come from building a vector from various
4538 // elements of other vectors.
4542 // Add this element source to the list if it's not already there.
4543 SDValue SourceVec = V.getOperand(0);
4544 auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
4545 if (Source == Sources.end())
4546 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
4548 // Update the minimum and maximum lane number seen.
4549 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4550 Source->MinElt = std::min(Source->MinElt, EltNo);
4551 Source->MaxElt = std::max(Source->MaxElt, EltNo);
4554 // Currently only do something sane when at most two source vectors
4556 if (Sources.size() > 2)
4559 // Find out the smallest element size among result and two sources, and use
4560 // it as element size to build the shuffle_vector.
4561 EVT SmallestEltTy = VT.getVectorElementType();
4562 for (auto &Source : Sources) {
4563 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
4564 if (SrcEltTy.bitsLT(SmallestEltTy)) {
4565 SmallestEltTy = SrcEltTy;
4568 unsigned ResMultiplier =
4569 VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
4570 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
4571 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
4573 // If the source vector is too wide or too narrow, we may nevertheless be able
4574 // to construct a compatible shuffle either by concatenating it with UNDEF or
4575 // extracting a suitable range of elements.
4576 for (auto &Src : Sources) {
4577 EVT SrcVT = Src.ShuffleVec.getValueType();
4579 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
4582 // This stage of the search produces a source with the same element type as
4583 // the original, but with a total width matching the BUILD_VECTOR output.
4584 EVT EltVT = SrcVT.getVectorElementType();
4585 unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
4586 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
4588 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
4589 assert(2 * SrcVT.getSizeInBits() == VT.getSizeInBits());
4590 // We can pad out the smaller vector for free, so if it's part of a
4593 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
4594 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
4598 assert(SrcVT.getSizeInBits() == 2 * VT.getSizeInBits());
4600 if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
4601 // Span too large for a VEXT to cope
4605 if (Src.MinElt >= NumSrcElts) {
4606 // The extraction can just take the second half
4608 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4609 DAG.getConstant(NumSrcElts, MVT::i64));
4610 Src.WindowBase = -NumSrcElts;
4611 } else if (Src.MaxElt < NumSrcElts) {
4612 // The extraction can just take the first half
4614 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4615 DAG.getConstant(0, MVT::i64));
4617 // An actual VEXT is needed
4619 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4620 DAG.getConstant(0, MVT::i64));
4622 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4623 DAG.getConstant(NumSrcElts, MVT::i64));
4624 unsigned Imm = Src.MinElt * getExtFactor(VEXTSrc1);
4626 Src.ShuffleVec = DAG.getNode(AArch64ISD::EXT, dl, DestVT, VEXTSrc1,
4627 VEXTSrc2, DAG.getConstant(Imm, MVT::i32));
4628 Src.WindowBase = -Src.MinElt;
4632 // Another possible incompatibility occurs from the vector element types. We
4633 // can fix this by bitcasting the source vectors to the same type we intend
4635 for (auto &Src : Sources) {
4636 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
4637 if (SrcEltTy == SmallestEltTy)
4639 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
4640 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
4641 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
4642 Src.WindowBase *= Src.WindowScale;
4645 // Final sanity check before we try to actually produce a shuffle.
4647 for (auto Src : Sources)
4648 assert(Src.ShuffleVec.getValueType() == ShuffleVT);
4651 // The stars all align, our next step is to produce the mask for the shuffle.
4652 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
4653 int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
4654 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
4655 SDValue Entry = Op.getOperand(i);
4656 if (Entry.getOpcode() == ISD::UNDEF)
4659 auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
4660 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
4662 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
4663 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
4665 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
4666 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
4667 VT.getVectorElementType().getSizeInBits());
4668 int LanesDefined = BitsDefined / BitsPerShuffleLane;
4670 // This source is expected to fill ResMultiplier lanes of the final shuffle,
4671 // starting at the appropriate offset.
4672 int *LaneMask = &Mask[i * ResMultiplier];
4674 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
4675 ExtractBase += NumElts * (Src - Sources.begin());
4676 for (int j = 0; j < LanesDefined; ++j)
4677 LaneMask[j] = ExtractBase + j;
4680 // Final check before we try to produce nonsense...
4681 if (!isShuffleMaskLegal(Mask, ShuffleVT))
4684 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
4685 for (unsigned i = 0; i < Sources.size(); ++i)
4686 ShuffleOps[i] = Sources[i].ShuffleVec;
4688 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
4689 ShuffleOps[1], &Mask[0]);
4690 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
4693 // check if an EXT instruction can handle the shuffle mask when the
4694 // vector sources of the shuffle are the same.
4695 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4696 unsigned NumElts = VT.getVectorNumElements();
4698 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4704 // If this is a VEXT shuffle, the immediate value is the index of the first
4705 // element. The other shuffle indices must be the successive elements after
4707 unsigned ExpectedElt = Imm;
4708 for (unsigned i = 1; i < NumElts; ++i) {
4709 // Increment the expected index. If it wraps around, just follow it
4710 // back to index zero and keep going.
4712 if (ExpectedElt == NumElts)
4716 continue; // ignore UNDEF indices
4717 if (ExpectedElt != static_cast<unsigned>(M[i]))
4724 // check if an EXT instruction can handle the shuffle mask when the
4725 // vector sources of the shuffle are different.
4726 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
4728 // Look for the first non-undef element.
4729 const int *FirstRealElt = std::find_if(M.begin(), M.end(),
4730 [](int Elt) {return Elt >= 0;});
4732 // Benefit form APInt to handle overflow when calculating expected element.
4733 unsigned NumElts = VT.getVectorNumElements();
4734 unsigned MaskBits = APInt(32, NumElts * 2).logBase2();
4735 APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1);
4736 // The following shuffle indices must be the successive elements after the
4737 // first real element.
4738 const int *FirstWrongElt = std::find_if(FirstRealElt + 1, M.end(),
4739 [&](int Elt) {return Elt != ExpectedElt++ && Elt != -1;});
4740 if (FirstWrongElt != M.end())
4743 // The index of an EXT is the first element if it is not UNDEF.
4744 // Watch out for the beginning UNDEFs. The EXT index should be the expected
4745 // value of the first element. E.g.
4746 // <-1, -1, 3, ...> is treated as <1, 2, 3, ...>.
4747 // <-1, -1, 0, 1, ...> is treated as <2*NumElts-2, 2*NumElts-1, 0, 1, ...>.
4748 // ExpectedElt is the last mask index plus 1.
4749 Imm = ExpectedElt.getZExtValue();
4751 // There are two difference cases requiring to reverse input vectors.
4752 // For example, for vector <4 x i32> we have the following cases,
4753 // Case 1: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, -1, 0>)
4754 // Case 2: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, 7, 0>)
4755 // For both cases, we finally use mask <5, 6, 7, 0>, which requires
4756 // to reverse two input vectors.
4765 /// isREVMask - Check if a vector shuffle corresponds to a REV
4766 /// instruction with the specified blocksize. (The order of the elements
4767 /// within each block of the vector is reversed.)
4768 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4769 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
4770 "Only possible block sizes for REV are: 16, 32, 64");
4772 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4776 unsigned NumElts = VT.getVectorNumElements();
4777 unsigned BlockElts = M[0] + 1;
4778 // If the first shuffle index is UNDEF, be optimistic.
4780 BlockElts = BlockSize / EltSz;
4782 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4785 for (unsigned i = 0; i < NumElts; ++i) {
4787 continue; // ignore UNDEF indices
4788 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
4795 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4796 unsigned NumElts = VT.getVectorNumElements();
4797 WhichResult = (M[0] == 0 ? 0 : 1);
4798 unsigned Idx = WhichResult * NumElts / 2;
4799 for (unsigned i = 0; i != NumElts; i += 2) {
4800 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4801 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
4809 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4810 unsigned NumElts = VT.getVectorNumElements();
4811 WhichResult = (M[0] == 0 ? 0 : 1);
4812 for (unsigned i = 0; i != NumElts; ++i) {
4814 continue; // ignore UNDEF indices
4815 if ((unsigned)M[i] != 2 * i + WhichResult)
4822 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4823 unsigned NumElts = VT.getVectorNumElements();
4824 WhichResult = (M[0] == 0 ? 0 : 1);
4825 for (unsigned i = 0; i < NumElts; i += 2) {
4826 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4827 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
4833 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
4834 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4835 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4836 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4837 unsigned NumElts = VT.getVectorNumElements();
4838 WhichResult = (M[0] == 0 ? 0 : 1);
4839 unsigned Idx = WhichResult * NumElts / 2;
4840 for (unsigned i = 0; i != NumElts; i += 2) {
4841 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4842 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
4850 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
4851 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4852 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4853 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4854 unsigned Half = VT.getVectorNumElements() / 2;
4855 WhichResult = (M[0] == 0 ? 0 : 1);
4856 for (unsigned j = 0; j != 2; ++j) {
4857 unsigned Idx = WhichResult;
4858 for (unsigned i = 0; i != Half; ++i) {
4859 int MIdx = M[i + j * Half];
4860 if (MIdx >= 0 && (unsigned)MIdx != Idx)
4869 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
4870 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4871 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4872 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4873 unsigned NumElts = VT.getVectorNumElements();
4874 WhichResult = (M[0] == 0 ? 0 : 1);
4875 for (unsigned i = 0; i < NumElts; i += 2) {
4876 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4877 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
4883 static bool isINSMask(ArrayRef<int> M, int NumInputElements,
4884 bool &DstIsLeft, int &Anomaly) {
4885 if (M.size() != static_cast<size_t>(NumInputElements))
4888 int NumLHSMatch = 0, NumRHSMatch = 0;
4889 int LastLHSMismatch = -1, LastRHSMismatch = -1;
4891 for (int i = 0; i < NumInputElements; ++i) {
4901 LastLHSMismatch = i;
4903 if (M[i] == i + NumInputElements)
4906 LastRHSMismatch = i;
4909 if (NumLHSMatch == NumInputElements - 1) {
4911 Anomaly = LastLHSMismatch;
4913 } else if (NumRHSMatch == NumInputElements - 1) {
4915 Anomaly = LastRHSMismatch;
4922 static bool isConcatMask(ArrayRef<int> Mask, EVT VT, bool SplitLHS) {
4923 if (VT.getSizeInBits() != 128)
4926 unsigned NumElts = VT.getVectorNumElements();
4928 for (int I = 0, E = NumElts / 2; I != E; I++) {
4933 int Offset = NumElts / 2;
4934 for (int I = NumElts / 2, E = NumElts; I != E; I++) {
4935 if (Mask[I] != I + SplitLHS * Offset)
4942 static SDValue tryFormConcatFromShuffle(SDValue Op, SelectionDAG &DAG) {
4944 EVT VT = Op.getValueType();
4945 SDValue V0 = Op.getOperand(0);
4946 SDValue V1 = Op.getOperand(1);
4947 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();
4949 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
4950 VT.getVectorElementType() != V1.getValueType().getVectorElementType())
4953 bool SplitV0 = V0.getValueType().getSizeInBits() == 128;
4955 if (!isConcatMask(Mask, VT, SplitV0))
4958 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
4959 VT.getVectorNumElements() / 2);
4961 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
4962 DAG.getConstant(0, MVT::i64));
4964 if (V1.getValueType().getSizeInBits() == 128) {
4965 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
4966 DAG.getConstant(0, MVT::i64));
4968 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
4971 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4972 /// the specified operations to build the shuffle.
4973 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4974 SDValue RHS, SelectionDAG &DAG,
4976 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4977 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
4978 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
4981 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4990 OP_VUZPL, // VUZP, left result
4991 OP_VUZPR, // VUZP, right result
4992 OP_VZIPL, // VZIP, left result
4993 OP_VZIPR, // VZIP, right result
4994 OP_VTRNL, // VTRN, left result
4995 OP_VTRNR // VTRN, right result
4998 if (OpNum == OP_COPY) {
4999 if (LHSID == (1 * 9 + 2) * 9 + 3)
5001 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
5005 SDValue OpLHS, OpRHS;
5006 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5007 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5008 EVT VT = OpLHS.getValueType();
5012 llvm_unreachable("Unknown shuffle opcode!");
5014 // VREV divides the vector in half and swaps within the half.
5015 if (VT.getVectorElementType() == MVT::i32 ||
5016 VT.getVectorElementType() == MVT::f32)
5017 return DAG.getNode(AArch64ISD::REV64, dl, VT, OpLHS);
5018 // vrev <4 x i16> -> REV32
5019 if (VT.getVectorElementType() == MVT::i16 ||
5020 VT.getVectorElementType() == MVT::f16)
5021 return DAG.getNode(AArch64ISD::REV32, dl, VT, OpLHS);
5022 // vrev <4 x i8> -> REV16
5023 assert(VT.getVectorElementType() == MVT::i8);
5024 return DAG.getNode(AArch64ISD::REV16, dl, VT, OpLHS);
5029 EVT EltTy = VT.getVectorElementType();
5031 if (EltTy == MVT::i8)
5032 Opcode = AArch64ISD::DUPLANE8;
5033 else if (EltTy == MVT::i16)
5034 Opcode = AArch64ISD::DUPLANE16;
5035 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
5036 Opcode = AArch64ISD::DUPLANE32;
5037 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
5038 Opcode = AArch64ISD::DUPLANE64;
5040 llvm_unreachable("Invalid vector element type?");
5042 if (VT.getSizeInBits() == 64)
5043 OpLHS = WidenVector(OpLHS, DAG);
5044 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, MVT::i64);
5045 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
5050 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
5051 return DAG.getNode(AArch64ISD::EXT, dl, VT, OpLHS, OpRHS,
5052 DAG.getConstant(Imm, MVT::i32));
5055 return DAG.getNode(AArch64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS,
5058 return DAG.getNode(AArch64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS,
5061 return DAG.getNode(AArch64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS,
5064 return DAG.getNode(AArch64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS,
5067 return DAG.getNode(AArch64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS,
5070 return DAG.getNode(AArch64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS,
5075 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
5076 SelectionDAG &DAG) {
5077 // Check to see if we can use the TBL instruction.
5078 SDValue V1 = Op.getOperand(0);
5079 SDValue V2 = Op.getOperand(1);
5082 EVT EltVT = Op.getValueType().getVectorElementType();
5083 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
5085 SmallVector<SDValue, 8> TBLMask;
5086 for (int Val : ShuffleMask) {
5087 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
5088 unsigned Offset = Byte + Val * BytesPerElt;
5089 TBLMask.push_back(DAG.getConstant(Offset, MVT::i32));
5093 MVT IndexVT = MVT::v8i8;
5094 unsigned IndexLen = 8;
5095 if (Op.getValueType().getSizeInBits() == 128) {
5096 IndexVT = MVT::v16i8;
5100 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
5101 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
5104 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
5106 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
5107 Shuffle = DAG.getNode(
5108 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5109 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, MVT::i32), V1Cst,
5110 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5111 makeArrayRef(TBLMask.data(), IndexLen)));
5113 if (IndexLen == 8) {
5114 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
5115 Shuffle = DAG.getNode(
5116 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5117 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, MVT::i32), V1Cst,
5118 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5119 makeArrayRef(TBLMask.data(), IndexLen)));
5121 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
5122 // cannot currently represent the register constraints on the input
5124 // Shuffle = DAG.getNode(AArch64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
5125 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5126 // &TBLMask[0], IndexLen));
5127 Shuffle = DAG.getNode(
5128 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5129 DAG.getConstant(Intrinsic::aarch64_neon_tbl2, MVT::i32), V1Cst, V2Cst,
5130 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5131 makeArrayRef(TBLMask.data(), IndexLen)));
5134 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
5137 static unsigned getDUPLANEOp(EVT EltType) {
5138 if (EltType == MVT::i8)
5139 return AArch64ISD::DUPLANE8;
5140 if (EltType == MVT::i16 || EltType == MVT::f16)
5141 return AArch64ISD::DUPLANE16;
5142 if (EltType == MVT::i32 || EltType == MVT::f32)
5143 return AArch64ISD::DUPLANE32;
5144 if (EltType == MVT::i64 || EltType == MVT::f64)
5145 return AArch64ISD::DUPLANE64;
5147 llvm_unreachable("Invalid vector element type?");
5150 SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5151 SelectionDAG &DAG) const {
5153 EVT VT = Op.getValueType();
5155 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5157 // Convert shuffles that are directly supported on NEON to target-specific
5158 // DAG nodes, instead of keeping them as shuffles and matching them again
5159 // during code selection. This is more efficient and avoids the possibility
5160 // of inconsistencies between legalization and selection.
5161 ArrayRef<int> ShuffleMask = SVN->getMask();
5163 SDValue V1 = Op.getOperand(0);
5164 SDValue V2 = Op.getOperand(1);
5166 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
5167 V1.getValueType().getSimpleVT())) {
5168 int Lane = SVN->getSplatIndex();
5169 // If this is undef splat, generate it via "just" vdup, if possible.
5173 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
5174 return DAG.getNode(AArch64ISD::DUP, dl, V1.getValueType(),
5176 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
5177 // constant. If so, we can just reference the lane's definition directly.
5178 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
5179 !isa<ConstantSDNode>(V1.getOperand(Lane)))
5180 return DAG.getNode(AArch64ISD::DUP, dl, VT, V1.getOperand(Lane));
5182 // Otherwise, duplicate from the lane of the input vector.
5183 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
5185 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
5186 // to make a vector of the same size as this SHUFFLE. We can ignore the
5187 // extract entirely, and canonicalise the concat using WidenVector.
5188 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
5189 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
5190 V1 = V1.getOperand(0);
5191 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
5192 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
5193 Lane -= Idx * VT.getVectorNumElements() / 2;
5194 V1 = WidenVector(V1.getOperand(Idx), DAG);
5195 } else if (VT.getSizeInBits() == 64)
5196 V1 = WidenVector(V1, DAG);
5198 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, MVT::i64));
5201 if (isREVMask(ShuffleMask, VT, 64))
5202 return DAG.getNode(AArch64ISD::REV64, dl, V1.getValueType(), V1, V2);
5203 if (isREVMask(ShuffleMask, VT, 32))
5204 return DAG.getNode(AArch64ISD::REV32, dl, V1.getValueType(), V1, V2);
5205 if (isREVMask(ShuffleMask, VT, 16))
5206 return DAG.getNode(AArch64ISD::REV16, dl, V1.getValueType(), V1, V2);
5208 bool ReverseEXT = false;
5210 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
5213 Imm *= getExtFactor(V1);
5214 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V2,
5215 DAG.getConstant(Imm, MVT::i32));
5216 } else if (V2->getOpcode() == ISD::UNDEF &&
5217 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
5218 Imm *= getExtFactor(V1);
5219 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V1,
5220 DAG.getConstant(Imm, MVT::i32));
5223 unsigned WhichResult;
5224 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
5225 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5226 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5228 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
5229 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5230 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5232 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
5233 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5234 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5237 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5238 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5239 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5241 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5242 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5243 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5245 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5246 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5247 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5250 SDValue Concat = tryFormConcatFromShuffle(Op, DAG);
5251 if (Concat.getNode())
5256 int NumInputElements = V1.getValueType().getVectorNumElements();
5257 if (isINSMask(ShuffleMask, NumInputElements, DstIsLeft, Anomaly)) {
5258 SDValue DstVec = DstIsLeft ? V1 : V2;
5259 SDValue DstLaneV = DAG.getConstant(Anomaly, MVT::i64);
5261 SDValue SrcVec = V1;
5262 int SrcLane = ShuffleMask[Anomaly];
5263 if (SrcLane >= NumInputElements) {
5265 SrcLane -= VT.getVectorNumElements();
5267 SDValue SrcLaneV = DAG.getConstant(SrcLane, MVT::i64);
5269 EVT ScalarVT = VT.getVectorElementType();
5271 if (ScalarVT.getSizeInBits() < 32 && ScalarVT.isInteger())
5272 ScalarVT = MVT::i32;
5275 ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5276 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ScalarVT, SrcVec, SrcLaneV),
5280 // If the shuffle is not directly supported and it has 4 elements, use
5281 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5282 unsigned NumElts = VT.getVectorNumElements();
5284 unsigned PFIndexes[4];
5285 for (unsigned i = 0; i != 4; ++i) {
5286 if (ShuffleMask[i] < 0)
5289 PFIndexes[i] = ShuffleMask[i];
5292 // Compute the index in the perfect shuffle table.
5293 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5294 PFIndexes[2] * 9 + PFIndexes[3];
5295 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5296 unsigned Cost = (PFEntry >> 30);
5299 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5302 return GenerateTBL(Op, ShuffleMask, DAG);
5305 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
5307 EVT VT = BVN->getValueType(0);
5308 APInt SplatBits, SplatUndef;
5309 unsigned SplatBitSize;
5311 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5312 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
5314 for (unsigned i = 0; i < NumSplats; ++i) {
5315 CnstBits <<= SplatBitSize;
5316 UndefBits <<= SplatBitSize;
5317 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
5318 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
5327 SDValue AArch64TargetLowering::LowerVectorAND(SDValue Op,
5328 SelectionDAG &DAG) const {
5329 BuildVectorSDNode *BVN =
5330 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5331 SDValue LHS = Op.getOperand(0);
5333 EVT VT = Op.getValueType();
5338 APInt CnstBits(VT.getSizeInBits(), 0);
5339 APInt UndefBits(VT.getSizeInBits(), 0);
5340 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5341 // We only have BIC vector immediate instruction, which is and-not.
5342 CnstBits = ~CnstBits;
5344 // We make use of a little bit of goto ickiness in order to avoid having to
5345 // duplicate the immediate matching logic for the undef toggled case.
5346 bool SecondTry = false;
5349 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5350 CnstBits = CnstBits.zextOrTrunc(64);
5351 uint64_t CnstVal = CnstBits.getZExtValue();
5353 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5354 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5355 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5356 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5357 DAG.getConstant(CnstVal, MVT::i32),
5358 DAG.getConstant(0, MVT::i32));
5359 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5362 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5363 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5364 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5365 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5366 DAG.getConstant(CnstVal, MVT::i32),
5367 DAG.getConstant(8, MVT::i32));
5368 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5371 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5372 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5373 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5374 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5375 DAG.getConstant(CnstVal, MVT::i32),
5376 DAG.getConstant(16, MVT::i32));
5377 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5380 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5381 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5382 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5383 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5384 DAG.getConstant(CnstVal, MVT::i32),
5385 DAG.getConstant(24, MVT::i32));
5386 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5389 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5390 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5391 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5392 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5393 DAG.getConstant(CnstVal, MVT::i32),
5394 DAG.getConstant(0, MVT::i32));
5395 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5398 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5399 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5400 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5401 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5402 DAG.getConstant(CnstVal, MVT::i32),
5403 DAG.getConstant(8, MVT::i32));
5404 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5411 CnstBits = ~UndefBits;
5415 // We can always fall back to a non-immediate AND.
5420 // Specialized code to quickly find if PotentialBVec is a BuildVector that
5421 // consists of only the same constant int value, returned in reference arg
5423 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
5424 uint64_t &ConstVal) {
5425 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
5428 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
5431 EVT VT = Bvec->getValueType(0);
5432 unsigned NumElts = VT.getVectorNumElements();
5433 for (unsigned i = 1; i < NumElts; ++i)
5434 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
5436 ConstVal = FirstElt->getZExtValue();
5440 static unsigned getIntrinsicID(const SDNode *N) {
5441 unsigned Opcode = N->getOpcode();
5444 return Intrinsic::not_intrinsic;
5445 case ISD::INTRINSIC_WO_CHAIN: {
5446 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5447 if (IID < Intrinsic::num_intrinsics)
5449 return Intrinsic::not_intrinsic;
5454 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
5455 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
5456 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
5457 // Also, logical shift right -> sri, with the same structure.
5458 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
5459 EVT VT = N->getValueType(0);
5466 // Is the first op an AND?
5467 const SDValue And = N->getOperand(0);
5468 if (And.getOpcode() != ISD::AND)
5471 // Is the second op an shl or lshr?
5472 SDValue Shift = N->getOperand(1);
5473 // This will have been turned into: AArch64ISD::VSHL vector, #shift
5474 // or AArch64ISD::VLSHR vector, #shift
5475 unsigned ShiftOpc = Shift.getOpcode();
5476 if ((ShiftOpc != AArch64ISD::VSHL && ShiftOpc != AArch64ISD::VLSHR))
5478 bool IsShiftRight = ShiftOpc == AArch64ISD::VLSHR;
5480 // Is the shift amount constant?
5481 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
5485 // Is the and mask vector all constant?
5487 if (!isAllConstantBuildVector(And.getOperand(1), C1))
5490 // Is C1 == ~C2, taking into account how much one can shift elements of a
5492 uint64_t C2 = C2node->getZExtValue();
5493 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
5494 if (C2 > ElemSizeInBits)
5496 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
5497 if ((C1 & ElemMask) != (~C2 & ElemMask))
5500 SDValue X = And.getOperand(0);
5501 SDValue Y = Shift.getOperand(0);
5504 IsShiftRight ? Intrinsic::aarch64_neon_vsri : Intrinsic::aarch64_neon_vsli;
5506 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5507 DAG.getConstant(Intrin, MVT::i32), X, Y, Shift.getOperand(1));
5509 DEBUG(dbgs() << "aarch64-lower: transformed: \n");
5510 DEBUG(N->dump(&DAG));
5511 DEBUG(dbgs() << "into: \n");
5512 DEBUG(ResultSLI->dump(&DAG));
5518 SDValue AArch64TargetLowering::LowerVectorOR(SDValue Op,
5519 SelectionDAG &DAG) const {
5520 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
5521 if (EnableAArch64SlrGeneration) {
5522 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
5527 BuildVectorSDNode *BVN =
5528 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
5529 SDValue LHS = Op.getOperand(1);
5531 EVT VT = Op.getValueType();
5533 // OR commutes, so try swapping the operands.
5535 LHS = Op.getOperand(0);
5536 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5541 APInt CnstBits(VT.getSizeInBits(), 0);
5542 APInt UndefBits(VT.getSizeInBits(), 0);
5543 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5544 // We make use of a little bit of goto ickiness in order to avoid having to
5545 // duplicate the immediate matching logic for the undef toggled case.
5546 bool SecondTry = false;
5549 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5550 CnstBits = CnstBits.zextOrTrunc(64);
5551 uint64_t CnstVal = CnstBits.getZExtValue();
5553 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5554 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5555 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5556 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5557 DAG.getConstant(CnstVal, MVT::i32),
5558 DAG.getConstant(0, MVT::i32));
5559 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5562 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5563 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5564 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5565 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5566 DAG.getConstant(CnstVal, MVT::i32),
5567 DAG.getConstant(8, MVT::i32));
5568 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5571 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5572 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5573 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5574 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5575 DAG.getConstant(CnstVal, MVT::i32),
5576 DAG.getConstant(16, MVT::i32));
5577 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5580 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5581 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5582 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5583 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5584 DAG.getConstant(CnstVal, MVT::i32),
5585 DAG.getConstant(24, MVT::i32));
5586 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5589 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5590 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5591 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5592 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5593 DAG.getConstant(CnstVal, MVT::i32),
5594 DAG.getConstant(0, MVT::i32));
5595 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5598 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5599 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5600 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5601 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5602 DAG.getConstant(CnstVal, MVT::i32),
5603 DAG.getConstant(8, MVT::i32));
5604 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5611 CnstBits = UndefBits;
5615 // We can always fall back to a non-immediate OR.
5620 // Normalize the operands of BUILD_VECTOR. The value of constant operands will
5621 // be truncated to fit element width.
5622 static SDValue NormalizeBuildVector(SDValue Op,
5623 SelectionDAG &DAG) {
5624 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
5626 EVT VT = Op.getValueType();
5627 EVT EltTy= VT.getVectorElementType();
5629 if (EltTy.isFloatingPoint() || EltTy.getSizeInBits() > 16)
5632 SmallVector<SDValue, 16> Ops;
5633 for (unsigned I = 0, E = VT.getVectorNumElements(); I != E; ++I) {
5634 SDValue Lane = Op.getOperand(I);
5635 if (Lane.getOpcode() == ISD::Constant) {
5636 APInt LowBits(EltTy.getSizeInBits(),
5637 cast<ConstantSDNode>(Lane)->getZExtValue());
5638 Lane = DAG.getConstant(LowBits.getZExtValue(), MVT::i32);
5640 Ops.push_back(Lane);
5642 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5645 SDValue AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
5646 SelectionDAG &DAG) const {
5648 EVT VT = Op.getValueType();
5649 Op = NormalizeBuildVector(Op, DAG);
5650 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5652 APInt CnstBits(VT.getSizeInBits(), 0);
5653 APInt UndefBits(VT.getSizeInBits(), 0);
5654 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5655 // We make use of a little bit of goto ickiness in order to avoid having to
5656 // duplicate the immediate matching logic for the undef toggled case.
5657 bool SecondTry = false;
5660 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5661 CnstBits = CnstBits.zextOrTrunc(64);
5662 uint64_t CnstVal = CnstBits.getZExtValue();
5664 // Certain magic vector constants (used to express things like NOT
5665 // and NEG) are passed through unmodified. This allows codegen patterns
5666 // for these operations to match. Special-purpose patterns will lower
5667 // these immediates to MOVIs if it proves necessary.
5668 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
5671 // The many faces of MOVI...
5672 if (AArch64_AM::isAdvSIMDModImmType10(CnstVal)) {
5673 CnstVal = AArch64_AM::encodeAdvSIMDModImmType10(CnstVal);
5674 if (VT.getSizeInBits() == 128) {
5675 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::v2i64,
5676 DAG.getConstant(CnstVal, MVT::i32));
5677 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5680 // Support the V64 version via subregister insertion.
5681 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::f64,
5682 DAG.getConstant(CnstVal, MVT::i32));
5683 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5686 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5687 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5688 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5689 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5690 DAG.getConstant(CnstVal, MVT::i32),
5691 DAG.getConstant(0, MVT::i32));
5692 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5695 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5696 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5697 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5698 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5699 DAG.getConstant(CnstVal, MVT::i32),
5700 DAG.getConstant(8, MVT::i32));
5701 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5704 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5705 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5706 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5707 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5708 DAG.getConstant(CnstVal, MVT::i32),
5709 DAG.getConstant(16, MVT::i32));
5710 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5713 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5714 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5715 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5716 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5717 DAG.getConstant(CnstVal, MVT::i32),
5718 DAG.getConstant(24, MVT::i32));
5719 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5722 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5723 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5724 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5725 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5726 DAG.getConstant(CnstVal, MVT::i32),
5727 DAG.getConstant(0, MVT::i32));
5728 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5731 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5732 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5733 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5734 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5735 DAG.getConstant(CnstVal, MVT::i32),
5736 DAG.getConstant(8, MVT::i32));
5737 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5740 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
5741 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
5742 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5743 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
5744 DAG.getConstant(CnstVal, MVT::i32),
5745 DAG.getConstant(264, MVT::i32));
5746 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5749 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
5750 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
5751 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5752 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
5753 DAG.getConstant(CnstVal, MVT::i32),
5754 DAG.getConstant(272, MVT::i32));
5755 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5758 if (AArch64_AM::isAdvSIMDModImmType9(CnstVal)) {
5759 CnstVal = AArch64_AM::encodeAdvSIMDModImmType9(CnstVal);
5760 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
5761 SDValue Mov = DAG.getNode(AArch64ISD::MOVI, dl, MovTy,
5762 DAG.getConstant(CnstVal, MVT::i32));
5763 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5766 // The few faces of FMOV...
5767 if (AArch64_AM::isAdvSIMDModImmType11(CnstVal)) {
5768 CnstVal = AArch64_AM::encodeAdvSIMDModImmType11(CnstVal);
5769 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
5770 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MovTy,
5771 DAG.getConstant(CnstVal, MVT::i32));
5772 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5775 if (AArch64_AM::isAdvSIMDModImmType12(CnstVal) &&
5776 VT.getSizeInBits() == 128) {
5777 CnstVal = AArch64_AM::encodeAdvSIMDModImmType12(CnstVal);
5778 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MVT::v2f64,
5779 DAG.getConstant(CnstVal, MVT::i32));
5780 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5783 // The many faces of MVNI...
5785 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5786 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5787 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5788 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5789 DAG.getConstant(CnstVal, MVT::i32),
5790 DAG.getConstant(0, MVT::i32));
5791 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5794 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5795 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5796 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5797 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5798 DAG.getConstant(CnstVal, MVT::i32),
5799 DAG.getConstant(8, MVT::i32));
5800 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5803 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5804 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5805 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5806 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5807 DAG.getConstant(CnstVal, MVT::i32),
5808 DAG.getConstant(16, MVT::i32));
5809 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5812 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5813 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5814 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5815 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5816 DAG.getConstant(CnstVal, MVT::i32),
5817 DAG.getConstant(24, MVT::i32));
5818 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5821 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5822 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5823 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5824 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5825 DAG.getConstant(CnstVal, MVT::i32),
5826 DAG.getConstant(0, MVT::i32));
5827 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5830 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5831 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5832 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5833 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5834 DAG.getConstant(CnstVal, MVT::i32),
5835 DAG.getConstant(8, MVT::i32));
5836 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5839 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
5840 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
5841 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5842 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
5843 DAG.getConstant(CnstVal, MVT::i32),
5844 DAG.getConstant(264, MVT::i32));
5845 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5848 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
5849 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
5850 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5851 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
5852 DAG.getConstant(CnstVal, MVT::i32),
5853 DAG.getConstant(272, MVT::i32));
5854 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5861 CnstBits = UndefBits;
5866 // Scan through the operands to find some interesting properties we can
5868 // 1) If only one value is used, we can use a DUP, or
5869 // 2) if only the low element is not undef, we can just insert that, or
5870 // 3) if only one constant value is used (w/ some non-constant lanes),
5871 // we can splat the constant value into the whole vector then fill
5872 // in the non-constant lanes.
5873 // 4) FIXME: If different constant values are used, but we can intelligently
5874 // select the values we'll be overwriting for the non-constant
5875 // lanes such that we can directly materialize the vector
5876 // some other way (MOVI, e.g.), we can be sneaky.
5877 unsigned NumElts = VT.getVectorNumElements();
5878 bool isOnlyLowElement = true;
5879 bool usesOnlyOneValue = true;
5880 bool usesOnlyOneConstantValue = true;
5881 bool isConstant = true;
5882 unsigned NumConstantLanes = 0;
5884 SDValue ConstantValue;
5885 for (unsigned i = 0; i < NumElts; ++i) {
5886 SDValue V = Op.getOperand(i);
5887 if (V.getOpcode() == ISD::UNDEF)
5890 isOnlyLowElement = false;
5891 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5894 if (isa<ConstantSDNode>(V) || isa<ConstantFPSDNode>(V)) {
5896 if (!ConstantValue.getNode())
5898 else if (ConstantValue != V)
5899 usesOnlyOneConstantValue = false;
5902 if (!Value.getNode())
5904 else if (V != Value)
5905 usesOnlyOneValue = false;
5908 if (!Value.getNode())
5909 return DAG.getUNDEF(VT);
5911 if (isOnlyLowElement)
5912 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5914 // Use DUP for non-constant splats. For f32 constant splats, reduce to
5915 // i32 and try again.
5916 if (usesOnlyOneValue) {
5918 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5919 Value.getValueType() != VT)
5920 return DAG.getNode(AArch64ISD::DUP, dl, VT, Value);
5922 // This is actually a DUPLANExx operation, which keeps everything vectory.
5924 // DUPLANE works on 128-bit vectors, widen it if necessary.
5925 SDValue Lane = Value.getOperand(1);
5926 Value = Value.getOperand(0);
5927 if (Value.getValueType().getSizeInBits() == 64)
5928 Value = WidenVector(Value, DAG);
5930 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
5931 return DAG.getNode(Opcode, dl, VT, Value, Lane);
5934 if (VT.getVectorElementType().isFloatingPoint()) {
5935 SmallVector<SDValue, 8> Ops;
5937 (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
5938 for (unsigned i = 0; i < NumElts; ++i)
5939 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
5940 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
5941 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5942 Val = LowerBUILD_VECTOR(Val, DAG);
5944 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5948 // If there was only one constant value used and for more than one lane,
5949 // start by splatting that value, then replace the non-constant lanes. This
5950 // is better than the default, which will perform a separate initialization
5952 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
5953 SDValue Val = DAG.getNode(AArch64ISD::DUP, dl, VT, ConstantValue);
5954 // Now insert the non-constant lanes.
5955 for (unsigned i = 0; i < NumElts; ++i) {
5956 SDValue V = Op.getOperand(i);
5957 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5958 if (!isa<ConstantSDNode>(V) && !isa<ConstantFPSDNode>(V)) {
5959 // Note that type legalization likely mucked about with the VT of the
5960 // source operand, so we may have to convert it here before inserting.
5961 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
5967 // If all elements are constants and the case above didn't get hit, fall back
5968 // to the default expansion, which will generate a load from the constant
5973 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5975 SDValue shuffle = ReconstructShuffle(Op, DAG);
5976 if (shuffle != SDValue())
5980 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5981 // know the default expansion would otherwise fall back on something even
5982 // worse. For a vector with one or two non-undef values, that's
5983 // scalar_to_vector for the elements followed by a shuffle (provided the
5984 // shuffle is valid for the target) and materialization element by element
5985 // on the stack followed by a load for everything else.
5986 if (!isConstant && !usesOnlyOneValue) {
5987 SDValue Vec = DAG.getUNDEF(VT);
5988 SDValue Op0 = Op.getOperand(0);
5989 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
5991 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
5992 // a) Avoid a RMW dependency on the full vector register, and
5993 // b) Allow the register coalescer to fold away the copy if the
5994 // value is already in an S or D register.
5995 if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
5996 unsigned SubIdx = ElemSize == 32 ? AArch64::ssub : AArch64::dsub;
5998 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
5999 DAG.getTargetConstant(SubIdx, MVT::i32));
6000 Vec = SDValue(N, 0);
6003 for (; i < NumElts; ++i) {
6004 SDValue V = Op.getOperand(i);
6005 if (V.getOpcode() == ISD::UNDEF)
6007 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
6008 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
6013 // Just use the default expansion. We failed to find a better alternative.
6017 SDValue AArch64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
6018 SelectionDAG &DAG) const {
6019 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
6021 // Check for non-constant or out of range lane.
6022 EVT VT = Op.getOperand(0).getValueType();
6023 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(2));
6024 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6028 // Insertion/extraction are legal for V128 types.
6029 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6030 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6034 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6035 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6038 // For V64 types, we perform insertion by expanding the value
6039 // to a V128 type and perform the insertion on that.
6041 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6042 EVT WideTy = WideVec.getValueType();
6044 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
6045 Op.getOperand(1), Op.getOperand(2));
6046 // Re-narrow the resultant vector.
6047 return NarrowVector(Node, DAG);
6051 AArch64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6052 SelectionDAG &DAG) const {
6053 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
6055 // Check for non-constant or out of range lane.
6056 EVT VT = Op.getOperand(0).getValueType();
6057 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6058 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6062 // Insertion/extraction are legal for V128 types.
6063 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6064 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6068 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6069 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6072 // For V64 types, we perform extraction by expanding the value
6073 // to a V128 type and perform the extraction on that.
6075 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6076 EVT WideTy = WideVec.getValueType();
6078 EVT ExtrTy = WideTy.getVectorElementType();
6079 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
6082 // For extractions, we just return the result directly.
6083 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
6087 SDValue AArch64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
6088 SelectionDAG &DAG) const {
6089 EVT VT = Op.getOperand(0).getValueType();
6095 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6098 unsigned Val = Cst->getZExtValue();
6100 unsigned Size = Op.getValueType().getSizeInBits();
6104 return DAG.getTargetExtractSubreg(AArch64::bsub, dl, Op.getValueType(),
6107 return DAG.getTargetExtractSubreg(AArch64::hsub, dl, Op.getValueType(),
6110 return DAG.getTargetExtractSubreg(AArch64::ssub, dl, Op.getValueType(),
6113 return DAG.getTargetExtractSubreg(AArch64::dsub, dl, Op.getValueType(),
6116 llvm_unreachable("Unexpected vector type in extract_subvector!");
6119 // If this is extracting the upper 64-bits of a 128-bit vector, we match
6121 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
6127 bool AArch64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
6129 if (VT.getVectorNumElements() == 4 &&
6130 (VT.is128BitVector() || VT.is64BitVector())) {
6131 unsigned PFIndexes[4];
6132 for (unsigned i = 0; i != 4; ++i) {
6136 PFIndexes[i] = M[i];
6139 // Compute the index in the perfect shuffle table.
6140 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
6141 PFIndexes[2] * 9 + PFIndexes[3];
6142 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
6143 unsigned Cost = (PFEntry >> 30);
6151 unsigned DummyUnsigned;
6153 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
6154 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
6155 isEXTMask(M, VT, DummyBool, DummyUnsigned) ||
6156 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
6157 isTRNMask(M, VT, DummyUnsigned) || isUZPMask(M, VT, DummyUnsigned) ||
6158 isZIPMask(M, VT, DummyUnsigned) ||
6159 isTRN_v_undef_Mask(M, VT, DummyUnsigned) ||
6160 isUZP_v_undef_Mask(M, VT, DummyUnsigned) ||
6161 isZIP_v_undef_Mask(M, VT, DummyUnsigned) ||
6162 isINSMask(M, VT.getVectorNumElements(), DummyBool, DummyInt) ||
6163 isConcatMask(M, VT, VT.getSizeInBits() == 128));
6166 /// getVShiftImm - Check if this is a valid build_vector for the immediate
6167 /// operand of a vector shift operation, where all the elements of the
6168 /// build_vector must have the same constant integer value.
6169 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
6170 // Ignore bit_converts.
6171 while (Op.getOpcode() == ISD::BITCAST)
6172 Op = Op.getOperand(0);
6173 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
6174 APInt SplatBits, SplatUndef;
6175 unsigned SplatBitSize;
6177 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
6178 HasAnyUndefs, ElementBits) ||
6179 SplatBitSize > ElementBits)
6181 Cnt = SplatBits.getSExtValue();
6185 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
6186 /// operand of a vector shift left operation. That value must be in the range:
6187 /// 0 <= Value < ElementBits for a left shift; or
6188 /// 0 <= Value <= ElementBits for a long left shift.
6189 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
6190 assert(VT.isVector() && "vector shift count is not a vector type");
6191 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6192 if (!getVShiftImm(Op, ElementBits, Cnt))
6194 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
6197 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
6198 /// operand of a vector shift right operation. For a shift opcode, the value
6199 /// is positive, but for an intrinsic the value count must be negative. The
6200 /// absolute value must be in the range:
6201 /// 1 <= |Value| <= ElementBits for a right shift; or
6202 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
6203 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
6205 assert(VT.isVector() && "vector shift count is not a vector type");
6206 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6207 if (!getVShiftImm(Op, ElementBits, Cnt))
6211 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
6214 SDValue AArch64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
6215 SelectionDAG &DAG) const {
6216 EVT VT = Op.getValueType();
6220 if (!Op.getOperand(1).getValueType().isVector())
6222 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
6224 switch (Op.getOpcode()) {
6226 llvm_unreachable("unexpected shift opcode");
6229 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
6230 return DAG.getNode(AArch64ISD::VSHL, SDLoc(Op), VT, Op.getOperand(0),
6231 DAG.getConstant(Cnt, MVT::i32));
6232 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6233 DAG.getConstant(Intrinsic::aarch64_neon_ushl, MVT::i32),
6234 Op.getOperand(0), Op.getOperand(1));
6237 // Right shift immediate
6238 if (isVShiftRImm(Op.getOperand(1), VT, false, false, Cnt) &&
6241 (Op.getOpcode() == ISD::SRA) ? AArch64ISD::VASHR : AArch64ISD::VLSHR;
6242 return DAG.getNode(Opc, SDLoc(Op), VT, Op.getOperand(0),
6243 DAG.getConstant(Cnt, MVT::i32));
6246 // Right shift register. Note, there is not a shift right register
6247 // instruction, but the shift left register instruction takes a signed
6248 // value, where negative numbers specify a right shift.
6249 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::aarch64_neon_sshl
6250 : Intrinsic::aarch64_neon_ushl;
6251 // negate the shift amount
6252 SDValue NegShift = DAG.getNode(AArch64ISD::NEG, DL, VT, Op.getOperand(1));
6253 SDValue NegShiftLeft =
6254 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6255 DAG.getConstant(Opc, MVT::i32), Op.getOperand(0), NegShift);
6256 return NegShiftLeft;
6262 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
6263 AArch64CC::CondCode CC, bool NoNans, EVT VT,
6264 SDLoc dl, SelectionDAG &DAG) {
6265 EVT SrcVT = LHS.getValueType();
6266 assert(VT.getSizeInBits() == SrcVT.getSizeInBits() &&
6267 "function only supposed to emit natural comparisons");
6269 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
6270 APInt CnstBits(VT.getSizeInBits(), 0);
6271 APInt UndefBits(VT.getSizeInBits(), 0);
6272 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
6273 bool IsZero = IsCnst && (CnstBits == 0);
6275 if (SrcVT.getVectorElementType().isFloatingPoint()) {
6279 case AArch64CC::NE: {
6282 Fcmeq = DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6284 Fcmeq = DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6285 return DAG.getNode(AArch64ISD::NOT, dl, VT, Fcmeq);
6289 return DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6290 return DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6293 return DAG.getNode(AArch64ISD::FCMGEz, dl, VT, LHS);
6294 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, LHS, RHS);
6297 return DAG.getNode(AArch64ISD::FCMGTz, dl, VT, LHS);
6298 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, LHS, RHS);
6301 return DAG.getNode(AArch64ISD::FCMLEz, dl, VT, LHS);
6302 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, RHS, LHS);
6306 // If we ignore NaNs then we can use to the MI implementation.
6310 return DAG.getNode(AArch64ISD::FCMLTz, dl, VT, LHS);
6311 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, RHS, LHS);
6318 case AArch64CC::NE: {
6321 Cmeq = DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6323 Cmeq = DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6324 return DAG.getNode(AArch64ISD::NOT, dl, VT, Cmeq);
6328 return DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6329 return DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6332 return DAG.getNode(AArch64ISD::CMGEz, dl, VT, LHS);
6333 return DAG.getNode(AArch64ISD::CMGE, dl, VT, LHS, RHS);
6336 return DAG.getNode(AArch64ISD::CMGTz, dl, VT, LHS);
6337 return DAG.getNode(AArch64ISD::CMGT, dl, VT, LHS, RHS);
6340 return DAG.getNode(AArch64ISD::CMLEz, dl, VT, LHS);
6341 return DAG.getNode(AArch64ISD::CMGE, dl, VT, RHS, LHS);
6343 return DAG.getNode(AArch64ISD::CMHS, dl, VT, RHS, LHS);
6345 return DAG.getNode(AArch64ISD::CMHI, dl, VT, RHS, LHS);
6348 return DAG.getNode(AArch64ISD::CMLTz, dl, VT, LHS);
6349 return DAG.getNode(AArch64ISD::CMGT, dl, VT, RHS, LHS);
6351 return DAG.getNode(AArch64ISD::CMHI, dl, VT, LHS, RHS);
6353 return DAG.getNode(AArch64ISD::CMHS, dl, VT, LHS, RHS);
6357 SDValue AArch64TargetLowering::LowerVSETCC(SDValue Op,
6358 SelectionDAG &DAG) const {
6359 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6360 SDValue LHS = Op.getOperand(0);
6361 SDValue RHS = Op.getOperand(1);
6362 EVT CmpVT = LHS.getValueType().changeVectorElementTypeToInteger();
6365 if (LHS.getValueType().getVectorElementType().isInteger()) {
6366 assert(LHS.getValueType() == RHS.getValueType());
6367 AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC);
6369 EmitVectorComparison(LHS, RHS, AArch64CC, false, CmpVT, dl, DAG);
6370 return DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType());
6373 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
6374 LHS.getValueType().getVectorElementType() == MVT::f64);
6376 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
6377 // clean. Some of them require two branches to implement.
6378 AArch64CC::CondCode CC1, CC2;
6380 changeVectorFPCCToAArch64CC(CC, CC1, CC2, ShouldInvert);
6382 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
6384 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, CmpVT, dl, DAG);
6388 if (CC2 != AArch64CC::AL) {
6390 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, CmpVT, dl, DAG);
6391 if (!Cmp2.getNode())
6394 Cmp = DAG.getNode(ISD::OR, dl, CmpVT, Cmp, Cmp2);
6397 Cmp = DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType());
6400 return Cmp = DAG.getNOT(dl, Cmp, Cmp.getValueType());
6405 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
6406 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
6407 /// specified in the intrinsic calls.
6408 bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
6410 unsigned Intrinsic) const {
6411 switch (Intrinsic) {
6412 case Intrinsic::aarch64_neon_ld2:
6413 case Intrinsic::aarch64_neon_ld3:
6414 case Intrinsic::aarch64_neon_ld4:
6415 case Intrinsic::aarch64_neon_ld1x2:
6416 case Intrinsic::aarch64_neon_ld1x3:
6417 case Intrinsic::aarch64_neon_ld1x4:
6418 case Intrinsic::aarch64_neon_ld2lane:
6419 case Intrinsic::aarch64_neon_ld3lane:
6420 case Intrinsic::aarch64_neon_ld4lane:
6421 case Intrinsic::aarch64_neon_ld2r:
6422 case Intrinsic::aarch64_neon_ld3r:
6423 case Intrinsic::aarch64_neon_ld4r: {
6424 Info.opc = ISD::INTRINSIC_W_CHAIN;
6425 // Conservatively set memVT to the entire set of vectors loaded.
6426 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
6427 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6428 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6431 Info.vol = false; // volatile loads with NEON intrinsics not supported
6432 Info.readMem = true;
6433 Info.writeMem = false;
6436 case Intrinsic::aarch64_neon_st2:
6437 case Intrinsic::aarch64_neon_st3:
6438 case Intrinsic::aarch64_neon_st4:
6439 case Intrinsic::aarch64_neon_st1x2:
6440 case Intrinsic::aarch64_neon_st1x3:
6441 case Intrinsic::aarch64_neon_st1x4:
6442 case Intrinsic::aarch64_neon_st2lane:
6443 case Intrinsic::aarch64_neon_st3lane:
6444 case Intrinsic::aarch64_neon_st4lane: {
6445 Info.opc = ISD::INTRINSIC_VOID;
6446 // Conservatively set memVT to the entire set of vectors stored.
6447 unsigned NumElts = 0;
6448 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
6449 Type *ArgTy = I.getArgOperand(ArgI)->getType();
6450 if (!ArgTy->isVectorTy())
6452 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
6454 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6455 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6458 Info.vol = false; // volatile stores with NEON intrinsics not supported
6459 Info.readMem = false;
6460 Info.writeMem = true;
6463 case Intrinsic::aarch64_ldaxr:
6464 case Intrinsic::aarch64_ldxr: {
6465 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
6466 Info.opc = ISD::INTRINSIC_W_CHAIN;
6467 Info.memVT = MVT::getVT(PtrTy->getElementType());
6468 Info.ptrVal = I.getArgOperand(0);
6470 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
6472 Info.readMem = true;
6473 Info.writeMem = false;
6476 case Intrinsic::aarch64_stlxr:
6477 case Intrinsic::aarch64_stxr: {
6478 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
6479 Info.opc = ISD::INTRINSIC_W_CHAIN;
6480 Info.memVT = MVT::getVT(PtrTy->getElementType());
6481 Info.ptrVal = I.getArgOperand(1);
6483 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
6485 Info.readMem = false;
6486 Info.writeMem = true;
6489 case Intrinsic::aarch64_ldaxp:
6490 case Intrinsic::aarch64_ldxp: {
6491 Info.opc = ISD::INTRINSIC_W_CHAIN;
6492 Info.memVT = MVT::i128;
6493 Info.ptrVal = I.getArgOperand(0);
6497 Info.readMem = true;
6498 Info.writeMem = false;
6501 case Intrinsic::aarch64_stlxp:
6502 case Intrinsic::aarch64_stxp: {
6503 Info.opc = ISD::INTRINSIC_W_CHAIN;
6504 Info.memVT = MVT::i128;
6505 Info.ptrVal = I.getArgOperand(2);
6509 Info.readMem = false;
6510 Info.writeMem = true;
6520 // Truncations from 64-bit GPR to 32-bit GPR is free.
6521 bool AArch64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
6522 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6524 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6525 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6526 return NumBits1 > NumBits2;
6528 bool AArch64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
6529 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6531 unsigned NumBits1 = VT1.getSizeInBits();
6532 unsigned NumBits2 = VT2.getSizeInBits();
6533 return NumBits1 > NumBits2;
6536 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
6538 bool AArch64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
6539 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6541 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6542 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6543 return NumBits1 == 32 && NumBits2 == 64;
6545 bool AArch64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
6546 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6548 unsigned NumBits1 = VT1.getSizeInBits();
6549 unsigned NumBits2 = VT2.getSizeInBits();
6550 return NumBits1 == 32 && NumBits2 == 64;
6553 bool AArch64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
6554 EVT VT1 = Val.getValueType();
6555 if (isZExtFree(VT1, VT2)) {
6559 if (Val.getOpcode() != ISD::LOAD)
6562 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
6563 return (VT1.isSimple() && !VT1.isVector() && VT1.isInteger() &&
6564 VT2.isSimple() && !VT2.isVector() && VT2.isInteger() &&
6565 VT1.getSizeInBits() <= 32);
6568 bool AArch64TargetLowering::hasPairedLoad(Type *LoadedType,
6569 unsigned &RequiredAligment) const {
6570 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
6572 // Cyclone supports unaligned accesses.
6573 RequiredAligment = 0;
6574 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
6575 return NumBits == 32 || NumBits == 64;
6578 bool AArch64TargetLowering::hasPairedLoad(EVT LoadedType,
6579 unsigned &RequiredAligment) const {
6580 if (!LoadedType.isSimple() ||
6581 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
6583 // Cyclone supports unaligned accesses.
6584 RequiredAligment = 0;
6585 unsigned NumBits = LoadedType.getSizeInBits();
6586 return NumBits == 32 || NumBits == 64;
6589 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
6590 unsigned AlignCheck) {
6591 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
6592 (DstAlign == 0 || DstAlign % AlignCheck == 0));
6595 EVT AArch64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
6596 unsigned SrcAlign, bool IsMemset,
6599 MachineFunction &MF) const {
6600 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
6601 // instruction to materialize the v2i64 zero and one store (with restrictive
6602 // addressing mode). Just do two i64 store of zero-registers.
6604 const Function *F = MF.getFunction();
6605 if (Subtarget->hasFPARMv8() && !IsMemset && Size >= 16 &&
6606 !F->hasFnAttribute(Attribute::NoImplicitFloat) &&
6607 (memOpAlign(SrcAlign, DstAlign, 16) ||
6608 (allowsMisalignedMemoryAccesses(MVT::f128, 0, 1, &Fast) && Fast)))
6611 return Size >= 8 ? MVT::i64 : MVT::i32;
6614 // 12-bit optionally shifted immediates are legal for adds.
6615 bool AArch64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
6616 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
6621 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
6622 // immediates is the same as for an add or a sub.
6623 bool AArch64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
6626 return isLegalAddImmediate(Immed);
6629 /// isLegalAddressingMode - Return true if the addressing mode represented
6630 /// by AM is legal for this target, for a load/store of the specified type.
6631 bool AArch64TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6633 // AArch64 has five basic addressing modes:
6635 // reg + 9-bit signed offset
6636 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
6638 // reg + SIZE_IN_BYTES * reg
6640 // No global is ever allowed as a base.
6644 // No reg+reg+imm addressing.
6645 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
6648 // check reg + imm case:
6649 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
6650 uint64_t NumBytes = 0;
6651 if (Ty->isSized()) {
6652 uint64_t NumBits = getDataLayout()->getTypeSizeInBits(Ty);
6653 NumBytes = NumBits / 8;
6654 if (!isPowerOf2_64(NumBits))
6659 int64_t Offset = AM.BaseOffs;
6661 // 9-bit signed offset
6662 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
6665 // 12-bit unsigned offset
6666 unsigned shift = Log2_64(NumBytes);
6667 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
6668 // Must be a multiple of NumBytes (NumBytes is a power of 2)
6669 (Offset >> shift) << shift == Offset)
6674 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
6676 if (!AM.Scale || AM.Scale == 1 ||
6677 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
6682 int AArch64TargetLowering::getScalingFactorCost(const AddrMode &AM,
6684 // Scaling factors are not free at all.
6685 // Operands | Rt Latency
6686 // -------------------------------------------
6688 // -------------------------------------------
6689 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
6690 // Rt, [Xn, Wm, <extend> #imm] |
6691 if (isLegalAddressingMode(AM, Ty))
6692 // Scale represents reg2 * scale, thus account for 1 if
6693 // it is not equal to 0 or 1.
6694 return AM.Scale != 0 && AM.Scale != 1;
6698 bool AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
6699 VT = VT.getScalarType();
6704 switch (VT.getSimpleVT().SimpleTy) {
6716 AArch64TargetLowering::getScratchRegisters(CallingConv::ID) const {
6717 // LR is a callee-save register, but we must treat it as clobbered by any call
6718 // site. Hence we include LR in the scratch registers, which are in turn added
6719 // as implicit-defs for stackmaps and patchpoints.
6720 static const MCPhysReg ScratchRegs[] = {
6721 AArch64::X16, AArch64::X17, AArch64::LR, 0
6727 AArch64TargetLowering::isDesirableToCommuteWithShift(const SDNode *N) const {
6728 EVT VT = N->getValueType(0);
6729 // If N is unsigned bit extraction: ((x >> C) & mask), then do not combine
6730 // it with shift to let it be lowered to UBFX.
6731 if (N->getOpcode() == ISD::AND && (VT == MVT::i32 || VT == MVT::i64) &&
6732 isa<ConstantSDNode>(N->getOperand(1))) {
6733 uint64_t TruncMask = N->getConstantOperandVal(1);
6734 if (isMask_64(TruncMask) &&
6735 N->getOperand(0).getOpcode() == ISD::SRL &&
6736 isa<ConstantSDNode>(N->getOperand(0)->getOperand(1)))
6742 bool AArch64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
6744 assert(Ty->isIntegerTy());
6746 unsigned BitSize = Ty->getPrimitiveSizeInBits();
6750 int64_t Val = Imm.getSExtValue();
6751 if (Val == 0 || AArch64_AM::isLogicalImmediate(Val, BitSize))
6754 if ((int64_t)Val < 0)
6757 Val &= (1LL << 32) - 1;
6759 unsigned LZ = countLeadingZeros((uint64_t)Val);
6760 unsigned Shift = (63 - LZ) / 16;
6761 // MOVZ is free so return true for one or fewer MOVK.
6762 return (Shift < 3) ? true : false;
6765 // Generate SUBS and CSEL for integer abs.
6766 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
6767 EVT VT = N->getValueType(0);
6769 SDValue N0 = N->getOperand(0);
6770 SDValue N1 = N->getOperand(1);
6773 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
6774 // and change it to SUB and CSEL.
6775 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
6776 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
6777 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
6778 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
6779 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
6780 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
6782 // Generate SUBS & CSEL.
6784 DAG.getNode(AArch64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
6785 N0.getOperand(0), DAG.getConstant(0, VT));
6786 return DAG.getNode(AArch64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
6787 DAG.getConstant(AArch64CC::PL, MVT::i32),
6788 SDValue(Cmp.getNode(), 1));
6793 // performXorCombine - Attempts to handle integer ABS.
6794 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
6795 TargetLowering::DAGCombinerInfo &DCI,
6796 const AArch64Subtarget *Subtarget) {
6797 if (DCI.isBeforeLegalizeOps())
6800 return performIntegerAbsCombine(N, DAG);
6804 AArch64TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
6806 std::vector<SDNode *> *Created) const {
6807 // fold (sdiv X, pow2)
6808 EVT VT = N->getValueType(0);
6809 if ((VT != MVT::i32 && VT != MVT::i64) ||
6810 !(Divisor.isPowerOf2() || (-Divisor).isPowerOf2()))
6814 SDValue N0 = N->getOperand(0);
6815 unsigned Lg2 = Divisor.countTrailingZeros();
6816 SDValue Zero = DAG.getConstant(0, VT);
6817 SDValue Pow2MinusOne = DAG.getConstant((1ULL << Lg2) - 1, VT);
6819 // Add (N0 < 0) ? Pow2 - 1 : 0;
6821 SDValue Cmp = getAArch64Cmp(N0, Zero, ISD::SETLT, CCVal, DAG, DL);
6822 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Pow2MinusOne);
6823 SDValue CSel = DAG.getNode(AArch64ISD::CSEL, DL, VT, Add, N0, CCVal, Cmp);
6826 Created->push_back(Cmp.getNode());
6827 Created->push_back(Add.getNode());
6828 Created->push_back(CSel.getNode());
6833 DAG.getNode(ISD::SRA, DL, VT, CSel, DAG.getConstant(Lg2, MVT::i64));
6835 // If we're dividing by a positive value, we're done. Otherwise, we must
6836 // negate the result.
6837 if (Divisor.isNonNegative())
6841 Created->push_back(SRA.getNode());
6842 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT), SRA);
6845 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
6846 TargetLowering::DAGCombinerInfo &DCI,
6847 const AArch64Subtarget *Subtarget) {
6848 if (DCI.isBeforeLegalizeOps())
6851 // Multiplication of a power of two plus/minus one can be done more
6852 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
6853 // future CPUs have a cheaper MADD instruction, this may need to be
6854 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
6855 // 64-bit is 5 cycles, so this is always a win.
6856 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
6857 APInt Value = C->getAPIntValue();
6858 EVT VT = N->getValueType(0);
6859 if (Value.isNonNegative()) {
6860 // (mul x, 2^N + 1) => (add (shl x, N), x)
6861 APInt VM1 = Value - 1;
6862 if (VM1.isPowerOf2()) {
6863 SDValue ShiftedVal =
6864 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6865 DAG.getConstant(VM1.logBase2(), MVT::i64));
6866 return DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal,
6869 // (mul x, 2^N - 1) => (sub (shl x, N), x)
6870 APInt VP1 = Value + 1;
6871 if (VP1.isPowerOf2()) {
6872 SDValue ShiftedVal =
6873 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6874 DAG.getConstant(VP1.logBase2(), MVT::i64));
6875 return DAG.getNode(ISD::SUB, SDLoc(N), VT, ShiftedVal,
6879 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
6880 APInt VNM1 = -Value - 1;
6881 if (VNM1.isPowerOf2()) {
6882 SDValue ShiftedVal =
6883 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6884 DAG.getConstant(VNM1.logBase2(), MVT::i64));
6886 DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6887 return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), Add);
6889 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
6890 APInt VNP1 = -Value + 1;
6891 if (VNP1.isPowerOf2()) {
6892 SDValue ShiftedVal =
6893 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6894 DAG.getConstant(VNP1.logBase2(), MVT::i64));
6895 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N->getOperand(0),
6903 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
6904 SelectionDAG &DAG) {
6905 // Take advantage of vector comparisons producing 0 or -1 in each lane to
6906 // optimize away operation when it's from a constant.
6908 // The general transformation is:
6909 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
6910 // AND(VECTOR_CMP(x,y), constant2)
6911 // constant2 = UNARYOP(constant)
6913 // Early exit if this isn't a vector operation, the operand of the
6914 // unary operation isn't a bitwise AND, or if the sizes of the operations
6916 EVT VT = N->getValueType(0);
6917 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
6918 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
6919 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
6922 // Now check that the other operand of the AND is a constant. We could
6923 // make the transformation for non-constant splats as well, but it's unclear
6924 // that would be a benefit as it would not eliminate any operations, just
6925 // perform one more step in scalar code before moving to the vector unit.
6926 if (BuildVectorSDNode *BV =
6927 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
6928 // Bail out if the vector isn't a constant.
6929 if (!BV->isConstant())
6932 // Everything checks out. Build up the new and improved node.
6934 EVT IntVT = BV->getValueType(0);
6935 // Create a new constant of the appropriate type for the transformed
6937 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
6938 // The AND node needs bitcasts to/from an integer vector type around it.
6939 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
6940 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
6941 N->getOperand(0)->getOperand(0), MaskConst);
6942 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
6949 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG,
6950 const AArch64Subtarget *Subtarget) {
6951 // First try to optimize away the conversion when it's conditionally from
6952 // a constant. Vectors only.
6953 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
6954 if (Res != SDValue())
6957 EVT VT = N->getValueType(0);
6958 if (VT != MVT::f32 && VT != MVT::f64)
6961 // Only optimize when the source and destination types have the same width.
6962 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
6965 // If the result of an integer load is only used by an integer-to-float
6966 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
6967 // This eliminates an "integer-to-vector-move UOP and improve throughput.
6968 SDValue N0 = N->getOperand(0);
6969 if (Subtarget->hasNEON() && ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6970 // Do not change the width of a volatile load.
6971 !cast<LoadSDNode>(N0)->isVolatile()) {
6972 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6973 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
6974 LN0->getPointerInfo(), LN0->isVolatile(),
6975 LN0->isNonTemporal(), LN0->isInvariant(),
6976 LN0->getAlignment());
6978 // Make sure successors of the original load stay after it by updating them
6979 // to use the new Chain.
6980 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
6983 (N->getOpcode() == ISD::SINT_TO_FP) ? AArch64ISD::SITOF : AArch64ISD::UITOF;
6984 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
6990 /// An EXTR instruction is made up of two shifts, ORed together. This helper
6991 /// searches for and classifies those shifts.
6992 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
6994 if (N.getOpcode() == ISD::SHL)
6996 else if (N.getOpcode() == ISD::SRL)
7001 if (!isa<ConstantSDNode>(N.getOperand(1)))
7004 ShiftAmount = N->getConstantOperandVal(1);
7005 Src = N->getOperand(0);
7009 /// EXTR instruction extracts a contiguous chunk of bits from two existing
7010 /// registers viewed as a high/low pair. This function looks for the pattern:
7011 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
7012 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
7014 static SDValue tryCombineToEXTR(SDNode *N,
7015 TargetLowering::DAGCombinerInfo &DCI) {
7016 SelectionDAG &DAG = DCI.DAG;
7018 EVT VT = N->getValueType(0);
7020 assert(N->getOpcode() == ISD::OR && "Unexpected root");
7022 if (VT != MVT::i32 && VT != MVT::i64)
7026 uint32_t ShiftLHS = 0;
7028 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
7032 uint32_t ShiftRHS = 0;
7034 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
7037 // If they're both trying to come from the high part of the register, they're
7038 // not really an EXTR.
7039 if (LHSFromHi == RHSFromHi)
7042 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
7046 std::swap(LHS, RHS);
7047 std::swap(ShiftLHS, ShiftRHS);
7050 return DAG.getNode(AArch64ISD::EXTR, DL, VT, LHS, RHS,
7051 DAG.getConstant(ShiftRHS, MVT::i64));
7054 static SDValue tryCombineToBSL(SDNode *N,
7055 TargetLowering::DAGCombinerInfo &DCI) {
7056 EVT VT = N->getValueType(0);
7057 SelectionDAG &DAG = DCI.DAG;
7063 SDValue N0 = N->getOperand(0);
7064 if (N0.getOpcode() != ISD::AND)
7067 SDValue N1 = N->getOperand(1);
7068 if (N1.getOpcode() != ISD::AND)
7071 // We only have to look for constant vectors here since the general, variable
7072 // case can be handled in TableGen.
7073 unsigned Bits = VT.getVectorElementType().getSizeInBits();
7074 uint64_t BitMask = Bits == 64 ? -1ULL : ((1ULL << Bits) - 1);
7075 for (int i = 1; i >= 0; --i)
7076 for (int j = 1; j >= 0; --j) {
7077 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(i));
7078 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(j));
7082 bool FoundMatch = true;
7083 for (unsigned k = 0; k < VT.getVectorNumElements(); ++k) {
7084 ConstantSDNode *CN0 = dyn_cast<ConstantSDNode>(BVN0->getOperand(k));
7085 ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(BVN1->getOperand(k));
7087 CN0->getZExtValue() != (BitMask & ~CN1->getZExtValue())) {
7094 return DAG.getNode(AArch64ISD::BSL, DL, VT, SDValue(BVN0, 0),
7095 N0->getOperand(1 - i), N1->getOperand(1 - j));
7101 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
7102 const AArch64Subtarget *Subtarget) {
7103 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
7104 if (!EnableAArch64ExtrGeneration)
7106 SelectionDAG &DAG = DCI.DAG;
7107 EVT VT = N->getValueType(0);
7109 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7112 SDValue Res = tryCombineToEXTR(N, DCI);
7116 Res = tryCombineToBSL(N, DCI);
7123 static SDValue performBitcastCombine(SDNode *N,
7124 TargetLowering::DAGCombinerInfo &DCI,
7125 SelectionDAG &DAG) {
7126 // Wait 'til after everything is legalized to try this. That way we have
7127 // legal vector types and such.
7128 if (DCI.isBeforeLegalizeOps())
7131 // Remove extraneous bitcasts around an extract_subvector.
7133 // (v4i16 (bitconvert
7134 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
7136 // (extract_subvector ((v8i16 ...), (i64 4)))
7138 // Only interested in 64-bit vectors as the ultimate result.
7139 EVT VT = N->getValueType(0);
7142 if (VT.getSimpleVT().getSizeInBits() != 64)
7144 // Is the operand an extract_subvector starting at the beginning or halfway
7145 // point of the vector? A low half may also come through as an
7146 // EXTRACT_SUBREG, so look for that, too.
7147 SDValue Op0 = N->getOperand(0);
7148 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
7149 !(Op0->isMachineOpcode() &&
7150 Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG))
7152 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
7153 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
7154 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
7156 } else if (Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG) {
7157 if (idx != AArch64::dsub)
7159 // The dsub reference is equivalent to a lane zero subvector reference.
7162 // Look through the bitcast of the input to the extract.
7163 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
7165 SDValue Source = Op0->getOperand(0)->getOperand(0);
7166 // If the source type has twice the number of elements as our destination
7167 // type, we know this is an extract of the high or low half of the vector.
7168 EVT SVT = Source->getValueType(0);
7169 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
7172 DEBUG(dbgs() << "aarch64-lower: bitcast extract_subvector simplification\n");
7174 // Create the simplified form to just extract the low or high half of the
7175 // vector directly rather than bothering with the bitcasts.
7177 unsigned NumElements = VT.getVectorNumElements();
7179 SDValue HalfIdx = DAG.getConstant(NumElements, MVT::i64);
7180 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
7182 SDValue SubReg = DAG.getTargetConstant(AArch64::dsub, MVT::i32);
7183 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
7189 static SDValue performConcatVectorsCombine(SDNode *N,
7190 TargetLowering::DAGCombinerInfo &DCI,
7191 SelectionDAG &DAG) {
7192 // Wait 'til after everything is legalized to try this. That way we have
7193 // legal vector types and such.
7194 if (DCI.isBeforeLegalizeOps())
7198 EVT VT = N->getValueType(0);
7200 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
7201 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
7202 // canonicalise to that.
7203 if (N->getOperand(0) == N->getOperand(1) && VT.getVectorNumElements() == 2) {
7204 assert(VT.getVectorElementType().getSizeInBits() == 64);
7205 return DAG.getNode(AArch64ISD::DUPLANE64, dl, VT,
7206 WidenVector(N->getOperand(0), DAG),
7207 DAG.getConstant(0, MVT::i64));
7210 // Canonicalise concat_vectors so that the right-hand vector has as few
7211 // bit-casts as possible before its real operation. The primary matching
7212 // destination for these operations will be the narrowing "2" instructions,
7213 // which depend on the operation being performed on this right-hand vector.
7215 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
7217 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
7219 SDValue Op1 = N->getOperand(1);
7220 if (Op1->getOpcode() != ISD::BITCAST)
7222 SDValue RHS = Op1->getOperand(0);
7223 MVT RHSTy = RHS.getValueType().getSimpleVT();
7224 // If the RHS is not a vector, this is not the pattern we're looking for.
7225 if (!RHSTy.isVector())
7228 DEBUG(dbgs() << "aarch64-lower: concat_vectors bitcast simplification\n");
7230 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
7231 RHSTy.getVectorNumElements() * 2);
7233 ISD::BITCAST, dl, VT,
7234 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
7235 DAG.getNode(ISD::BITCAST, dl, RHSTy, N->getOperand(0)), RHS));
7238 static SDValue tryCombineFixedPointConvert(SDNode *N,
7239 TargetLowering::DAGCombinerInfo &DCI,
7240 SelectionDAG &DAG) {
7241 // Wait 'til after everything is legalized to try this. That way we have
7242 // legal vector types and such.
7243 if (DCI.isBeforeLegalizeOps())
7245 // Transform a scalar conversion of a value from a lane extract into a
7246 // lane extract of a vector conversion. E.g., from foo1 to foo2:
7247 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
7248 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
7250 // The second form interacts better with instruction selection and the
7251 // register allocator to avoid cross-class register copies that aren't
7252 // coalescable due to a lane reference.
7254 // Check the operand and see if it originates from a lane extract.
7255 SDValue Op1 = N->getOperand(1);
7256 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7257 // Yep, no additional predication needed. Perform the transform.
7258 SDValue IID = N->getOperand(0);
7259 SDValue Shift = N->getOperand(2);
7260 SDValue Vec = Op1.getOperand(0);
7261 SDValue Lane = Op1.getOperand(1);
7262 EVT ResTy = N->getValueType(0);
7266 // The vector width should be 128 bits by the time we get here, even
7267 // if it started as 64 bits (the extract_vector handling will have
7269 assert(Vec.getValueType().getSizeInBits() == 128 &&
7270 "unexpected vector size on extract_vector_elt!");
7271 if (Vec.getValueType() == MVT::v4i32)
7272 VecResTy = MVT::v4f32;
7273 else if (Vec.getValueType() == MVT::v2i64)
7274 VecResTy = MVT::v2f64;
7276 llvm_unreachable("unexpected vector type!");
7279 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
7280 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
7285 // AArch64 high-vector "long" operations are formed by performing the non-high
7286 // version on an extract_subvector of each operand which gets the high half:
7288 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
7290 // However, there are cases which don't have an extract_high explicitly, but
7291 // have another operation that can be made compatible with one for free. For
7294 // (dupv64 scalar) --> (extract_high (dup128 scalar))
7296 // This routine does the actual conversion of such DUPs, once outer routines
7297 // have determined that everything else is in order.
7298 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
7299 // We can handle most types of duplicate, but the lane ones have an extra
7300 // operand saying *which* lane, so we need to know.
7302 switch (N.getOpcode()) {
7303 case AArch64ISD::DUP:
7306 case AArch64ISD::DUPLANE8:
7307 case AArch64ISD::DUPLANE16:
7308 case AArch64ISD::DUPLANE32:
7309 case AArch64ISD::DUPLANE64:
7316 MVT NarrowTy = N.getSimpleValueType();
7317 if (!NarrowTy.is64BitVector())
7320 MVT ElementTy = NarrowTy.getVectorElementType();
7321 unsigned NumElems = NarrowTy.getVectorNumElements();
7322 MVT NewDUPVT = MVT::getVectorVT(ElementTy, NumElems * 2);
7326 NewDUP = DAG.getNode(N.getOpcode(), SDLoc(N), NewDUPVT, N.getOperand(0),
7329 NewDUP = DAG.getNode(AArch64ISD::DUP, SDLoc(N), NewDUPVT, N.getOperand(0));
7331 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N.getNode()), NarrowTy,
7332 NewDUP, DAG.getConstant(NumElems, MVT::i64));
7335 static bool isEssentiallyExtractSubvector(SDValue N) {
7336 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
7339 return N.getOpcode() == ISD::BITCAST &&
7340 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
7343 /// \brief Helper structure to keep track of ISD::SET_CC operands.
7344 struct GenericSetCCInfo {
7345 const SDValue *Opnd0;
7346 const SDValue *Opnd1;
7350 /// \brief Helper structure to keep track of a SET_CC lowered into AArch64 code.
7351 struct AArch64SetCCInfo {
7353 AArch64CC::CondCode CC;
7356 /// \brief Helper structure to keep track of SetCC information.
7358 GenericSetCCInfo Generic;
7359 AArch64SetCCInfo AArch64;
7362 /// \brief Helper structure to be able to read SetCC information. If set to
7363 /// true, IsAArch64 field, Info is a AArch64SetCCInfo, otherwise Info is a
7364 /// GenericSetCCInfo.
7365 struct SetCCInfoAndKind {
7370 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
7372 /// AArch64 lowered one.
7373 /// \p SetCCInfo is filled accordingly.
7374 /// \post SetCCInfo is meanginfull only when this function returns true.
7375 /// \return True when Op is a kind of SET_CC operation.
7376 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
7377 // If this is a setcc, this is straight forward.
7378 if (Op.getOpcode() == ISD::SETCC) {
7379 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
7380 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
7381 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7382 SetCCInfo.IsAArch64 = false;
7385 // Otherwise, check if this is a matching csel instruction.
7389 if (Op.getOpcode() != AArch64ISD::CSEL)
7391 // Set the information about the operands.
7392 // TODO: we want the operands of the Cmp not the csel
7393 SetCCInfo.Info.AArch64.Cmp = &Op.getOperand(3);
7394 SetCCInfo.IsAArch64 = true;
7395 SetCCInfo.Info.AArch64.CC = static_cast<AArch64CC::CondCode>(
7396 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
7398 // Check that the operands matches the constraints:
7399 // (1) Both operands must be constants.
7400 // (2) One must be 1 and the other must be 0.
7401 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
7402 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7405 if (!TValue || !FValue)
7409 if (!TValue->isOne()) {
7410 // Update the comparison when we are interested in !cc.
7411 std::swap(TValue, FValue);
7412 SetCCInfo.Info.AArch64.CC =
7413 AArch64CC::getInvertedCondCode(SetCCInfo.Info.AArch64.CC);
7415 return TValue->isOne() && FValue->isNullValue();
7418 // Returns true if Op is setcc or zext of setcc.
7419 static bool isSetCCOrZExtSetCC(const SDValue& Op, SetCCInfoAndKind &Info) {
7420 if (isSetCC(Op, Info))
7422 return ((Op.getOpcode() == ISD::ZERO_EXTEND) &&
7423 isSetCC(Op->getOperand(0), Info));
7426 // The folding we want to perform is:
7427 // (add x, [zext] (setcc cc ...) )
7429 // (csel x, (add x, 1), !cc ...)
7431 // The latter will get matched to a CSINC instruction.
7432 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
7433 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
7434 SDValue LHS = Op->getOperand(0);
7435 SDValue RHS = Op->getOperand(1);
7436 SetCCInfoAndKind InfoAndKind;
7438 // If neither operand is a SET_CC, give up.
7439 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind)) {
7440 std::swap(LHS, RHS);
7441 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind))
7445 // FIXME: This could be generatized to work for FP comparisons.
7446 EVT CmpVT = InfoAndKind.IsAArch64
7447 ? InfoAndKind.Info.AArch64.Cmp->getOperand(0).getValueType()
7448 : InfoAndKind.Info.Generic.Opnd0->getValueType();
7449 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
7455 if (InfoAndKind.IsAArch64) {
7456 CCVal = DAG.getConstant(
7457 AArch64CC::getInvertedCondCode(InfoAndKind.Info.AArch64.CC), MVT::i32);
7458 Cmp = *InfoAndKind.Info.AArch64.Cmp;
7460 Cmp = getAArch64Cmp(*InfoAndKind.Info.Generic.Opnd0,
7461 *InfoAndKind.Info.Generic.Opnd1,
7462 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
7465 EVT VT = Op->getValueType(0);
7466 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, VT));
7467 return DAG.getNode(AArch64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
7470 // The basic add/sub long vector instructions have variants with "2" on the end
7471 // which act on the high-half of their inputs. They are normally matched by
7474 // (add (zeroext (extract_high LHS)),
7475 // (zeroext (extract_high RHS)))
7476 // -> uaddl2 vD, vN, vM
7478 // However, if one of the extracts is something like a duplicate, this
7479 // instruction can still be used profitably. This function puts the DAG into a
7480 // more appropriate form for those patterns to trigger.
7481 static SDValue performAddSubLongCombine(SDNode *N,
7482 TargetLowering::DAGCombinerInfo &DCI,
7483 SelectionDAG &DAG) {
7484 if (DCI.isBeforeLegalizeOps())
7487 MVT VT = N->getSimpleValueType(0);
7488 if (!VT.is128BitVector()) {
7489 if (N->getOpcode() == ISD::ADD)
7490 return performSetccAddFolding(N, DAG);
7494 // Make sure both branches are extended in the same way.
7495 SDValue LHS = N->getOperand(0);
7496 SDValue RHS = N->getOperand(1);
7497 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
7498 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
7499 LHS.getOpcode() != RHS.getOpcode())
7502 unsigned ExtType = LHS.getOpcode();
7504 // It's not worth doing if at least one of the inputs isn't already an
7505 // extract, but we don't know which it'll be so we have to try both.
7506 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
7507 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
7511 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
7512 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
7513 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
7517 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
7520 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
7523 // Massage DAGs which we can use the high-half "long" operations on into
7524 // something isel will recognize better. E.g.
7526 // (aarch64_neon_umull (extract_high vec) (dupv64 scalar)) -->
7527 // (aarch64_neon_umull (extract_high (v2i64 vec)))
7528 // (extract_high (v2i64 (dup128 scalar)))))
7530 static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N,
7531 TargetLowering::DAGCombinerInfo &DCI,
7532 SelectionDAG &DAG) {
7533 if (DCI.isBeforeLegalizeOps())
7536 SDValue LHS = N->getOperand(1);
7537 SDValue RHS = N->getOperand(2);
7538 assert(LHS.getValueType().is64BitVector() &&
7539 RHS.getValueType().is64BitVector() &&
7540 "unexpected shape for long operation");
7542 // Either node could be a DUP, but it's not worth doing both of them (you'd
7543 // just as well use the non-high version) so look for a corresponding extract
7544 // operation on the other "wing".
7545 if (isEssentiallyExtractSubvector(LHS)) {
7546 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
7549 } else if (isEssentiallyExtractSubvector(RHS)) {
7550 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
7555 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
7556 N->getOperand(0), LHS, RHS);
7559 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
7560 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
7561 unsigned ElemBits = ElemTy.getSizeInBits();
7563 int64_t ShiftAmount;
7564 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
7565 APInt SplatValue, SplatUndef;
7566 unsigned SplatBitSize;
7568 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
7569 HasAnyUndefs, ElemBits) ||
7570 SplatBitSize != ElemBits)
7573 ShiftAmount = SplatValue.getSExtValue();
7574 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
7575 ShiftAmount = CVN->getSExtValue();
7583 llvm_unreachable("Unknown shift intrinsic");
7584 case Intrinsic::aarch64_neon_sqshl:
7585 Opcode = AArch64ISD::SQSHL_I;
7586 IsRightShift = false;
7588 case Intrinsic::aarch64_neon_uqshl:
7589 Opcode = AArch64ISD::UQSHL_I;
7590 IsRightShift = false;
7592 case Intrinsic::aarch64_neon_srshl:
7593 Opcode = AArch64ISD::SRSHR_I;
7594 IsRightShift = true;
7596 case Intrinsic::aarch64_neon_urshl:
7597 Opcode = AArch64ISD::URSHR_I;
7598 IsRightShift = true;
7600 case Intrinsic::aarch64_neon_sqshlu:
7601 Opcode = AArch64ISD::SQSHLU_I;
7602 IsRightShift = false;
7606 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits)
7607 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
7608 DAG.getConstant(-ShiftAmount, MVT::i32));
7609 else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount < ElemBits)
7610 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
7611 DAG.getConstant(ShiftAmount, MVT::i32));
7616 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
7617 // the intrinsics must be legal and take an i32, this means there's almost
7618 // certainly going to be a zext in the DAG which we can eliminate.
7619 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
7620 SDValue AndN = N->getOperand(2);
7621 if (AndN.getOpcode() != ISD::AND)
7624 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
7625 if (!CMask || CMask->getZExtValue() != Mask)
7628 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
7629 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
7632 static SDValue performIntrinsicCombine(SDNode *N,
7633 TargetLowering::DAGCombinerInfo &DCI,
7634 const AArch64Subtarget *Subtarget) {
7635 SelectionDAG &DAG = DCI.DAG;
7636 unsigned IID = getIntrinsicID(N);
7640 case Intrinsic::aarch64_neon_vcvtfxs2fp:
7641 case Intrinsic::aarch64_neon_vcvtfxu2fp:
7642 return tryCombineFixedPointConvert(N, DCI, DAG);
7644 case Intrinsic::aarch64_neon_fmax:
7645 return DAG.getNode(AArch64ISD::FMAX, SDLoc(N), N->getValueType(0),
7646 N->getOperand(1), N->getOperand(2));
7647 case Intrinsic::aarch64_neon_fmin:
7648 return DAG.getNode(AArch64ISD::FMIN, SDLoc(N), N->getValueType(0),
7649 N->getOperand(1), N->getOperand(2));
7650 case Intrinsic::aarch64_neon_smull:
7651 case Intrinsic::aarch64_neon_umull:
7652 case Intrinsic::aarch64_neon_pmull:
7653 case Intrinsic::aarch64_neon_sqdmull:
7654 return tryCombineLongOpWithDup(IID, N, DCI, DAG);
7655 case Intrinsic::aarch64_neon_sqshl:
7656 case Intrinsic::aarch64_neon_uqshl:
7657 case Intrinsic::aarch64_neon_sqshlu:
7658 case Intrinsic::aarch64_neon_srshl:
7659 case Intrinsic::aarch64_neon_urshl:
7660 return tryCombineShiftImm(IID, N, DAG);
7661 case Intrinsic::aarch64_crc32b:
7662 case Intrinsic::aarch64_crc32cb:
7663 return tryCombineCRC32(0xff, N, DAG);
7664 case Intrinsic::aarch64_crc32h:
7665 case Intrinsic::aarch64_crc32ch:
7666 return tryCombineCRC32(0xffff, N, DAG);
7671 static SDValue performExtendCombine(SDNode *N,
7672 TargetLowering::DAGCombinerInfo &DCI,
7673 SelectionDAG &DAG) {
7674 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
7675 // we can convert that DUP into another extract_high (of a bigger DUP), which
7676 // helps the backend to decide that an sabdl2 would be useful, saving a real
7677 // extract_high operation.
7678 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
7679 N->getOperand(0).getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
7680 SDNode *ABDNode = N->getOperand(0).getNode();
7681 unsigned IID = getIntrinsicID(ABDNode);
7682 if (IID == Intrinsic::aarch64_neon_sabd ||
7683 IID == Intrinsic::aarch64_neon_uabd) {
7684 SDValue NewABD = tryCombineLongOpWithDup(IID, ABDNode, DCI, DAG);
7685 if (!NewABD.getNode())
7688 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
7693 // This is effectively a custom type legalization for AArch64.
7695 // Type legalization will split an extend of a small, legal, type to a larger
7696 // illegal type by first splitting the destination type, often creating
7697 // illegal source types, which then get legalized in isel-confusing ways,
7698 // leading to really terrible codegen. E.g.,
7699 // %result = v8i32 sext v8i8 %value
7701 // %losrc = extract_subreg %value, ...
7702 // %hisrc = extract_subreg %value, ...
7703 // %lo = v4i32 sext v4i8 %losrc
7704 // %hi = v4i32 sext v4i8 %hisrc
7705 // Things go rapidly downhill from there.
7707 // For AArch64, the [sz]ext vector instructions can only go up one element
7708 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
7709 // take two instructions.
7711 // This implies that the most efficient way to do the extend from v8i8
7712 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
7713 // the normal splitting to happen for the v8i16->v8i32.
7715 // This is pre-legalization to catch some cases where the default
7716 // type legalization will create ill-tempered code.
7717 if (!DCI.isBeforeLegalizeOps())
7720 // We're only interested in cleaning things up for non-legal vector types
7721 // here. If both the source and destination are legal, things will just
7722 // work naturally without any fiddling.
7723 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7724 EVT ResVT = N->getValueType(0);
7725 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
7727 // If the vector type isn't a simple VT, it's beyond the scope of what
7728 // we're worried about here. Let legalization do its thing and hope for
7730 SDValue Src = N->getOperand(0);
7731 EVT SrcVT = Src->getValueType(0);
7732 if (!ResVT.isSimple() || !SrcVT.isSimple())
7735 // If the source VT is a 64-bit vector, we can play games and get the
7736 // better results we want.
7737 if (SrcVT.getSizeInBits() != 64)
7740 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
7741 unsigned ElementCount = SrcVT.getVectorNumElements();
7742 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
7744 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
7746 // Now split the rest of the operation into two halves, each with a 64
7750 unsigned NumElements = ResVT.getVectorNumElements();
7751 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
7752 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
7753 ResVT.getVectorElementType(), NumElements / 2);
7755 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
7756 LoVT.getVectorNumElements());
7757 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7758 DAG.getConstant(0, MVT::i64));
7759 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7760 DAG.getConstant(InNVT.getVectorNumElements(), MVT::i64));
7761 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
7762 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
7764 // Now combine the parts back together so we still have a single result
7765 // like the combiner expects.
7766 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
7769 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
7770 /// value. The load store optimizer pass will merge them to store pair stores.
7771 /// This has better performance than a splat of the scalar followed by a split
7772 /// vector store. Even if the stores are not merged it is four stores vs a dup,
7773 /// followed by an ext.b and two stores.
7774 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
7775 SDValue StVal = St->getValue();
7776 EVT VT = StVal.getValueType();
7778 // Don't replace floating point stores, they possibly won't be transformed to
7779 // stp because of the store pair suppress pass.
7780 if (VT.isFloatingPoint())
7783 // Check for insert vector elements.
7784 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
7787 // We can express a splat as store pair(s) for 2 or 4 elements.
7788 unsigned NumVecElts = VT.getVectorNumElements();
7789 if (NumVecElts != 4 && NumVecElts != 2)
7791 SDValue SplatVal = StVal.getOperand(1);
7792 unsigned RemainInsertElts = NumVecElts - 1;
7794 // Check that this is a splat.
7795 while (--RemainInsertElts) {
7796 SDValue NextInsertElt = StVal.getOperand(0);
7797 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
7799 if (NextInsertElt.getOperand(1) != SplatVal)
7801 StVal = NextInsertElt;
7803 unsigned OrigAlignment = St->getAlignment();
7804 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
7805 unsigned Alignment = std::min(OrigAlignment, EltOffset);
7807 // Create scalar stores. This is at least as good as the code sequence for a
7808 // split unaligned store wich is a dup.s, ext.b, and two stores.
7809 // Most of the time the three stores should be replaced by store pair
7810 // instructions (stp).
7812 SDValue BasePtr = St->getBasePtr();
7814 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
7815 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
7817 unsigned Offset = EltOffset;
7818 while (--NumVecElts) {
7819 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7820 DAG.getConstant(Offset, MVT::i64));
7821 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
7822 St->getPointerInfo(), St->isVolatile(),
7823 St->isNonTemporal(), Alignment);
7824 Offset += EltOffset;
7829 static SDValue performSTORECombine(SDNode *N,
7830 TargetLowering::DAGCombinerInfo &DCI,
7832 const AArch64Subtarget *Subtarget) {
7833 if (!DCI.isBeforeLegalize())
7836 StoreSDNode *S = cast<StoreSDNode>(N);
7837 if (S->isVolatile())
7840 // Cyclone has bad performance on unaligned 16B stores when crossing line and
7841 // page boundaries. We want to split such stores.
7842 if (!Subtarget->isCyclone())
7845 // Don't split at Oz.
7846 MachineFunction &MF = DAG.getMachineFunction();
7847 bool IsMinSize = MF.getFunction()->hasFnAttribute(Attribute::MinSize);
7851 SDValue StVal = S->getValue();
7852 EVT VT = StVal.getValueType();
7854 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
7855 // those up regresses performance on micro-benchmarks and olden/bh.
7856 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
7859 // Split unaligned 16B stores. They are terrible for performance.
7860 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
7861 // extensions can use this to mark that it does not want splitting to happen
7862 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
7863 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
7864 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
7865 S->getAlignment() <= 2)
7868 // If we get a splat of a scalar convert this vector store to a store of
7869 // scalars. They will be merged into store pairs thereby removing two
7871 SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S);
7872 if (ReplacedSplat != SDValue())
7873 return ReplacedSplat;
7876 unsigned NumElts = VT.getVectorNumElements() / 2;
7877 // Split VT into two.
7879 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
7880 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7881 DAG.getConstant(0, MVT::i64));
7882 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7883 DAG.getConstant(NumElts, MVT::i64));
7884 SDValue BasePtr = S->getBasePtr();
7886 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
7887 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
7888 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7889 DAG.getConstant(8, MVT::i64));
7890 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
7891 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
7895 /// Target-specific DAG combine function for post-increment LD1 (lane) and
7896 /// post-increment LD1R.
7897 static SDValue performPostLD1Combine(SDNode *N,
7898 TargetLowering::DAGCombinerInfo &DCI,
7900 if (DCI.isBeforeLegalizeOps())
7903 SelectionDAG &DAG = DCI.DAG;
7904 EVT VT = N->getValueType(0);
7906 unsigned LoadIdx = IsLaneOp ? 1 : 0;
7907 SDNode *LD = N->getOperand(LoadIdx).getNode();
7908 // If it is not LOAD, can not do such combine.
7909 if (LD->getOpcode() != ISD::LOAD)
7912 LoadSDNode *LoadSDN = cast<LoadSDNode>(LD);
7913 EVT MemVT = LoadSDN->getMemoryVT();
7914 // Check if memory operand is the same type as the vector element.
7915 if (MemVT != VT.getVectorElementType())
7918 // Check if there are other uses. If so, do not combine as it will introduce
7920 for (SDNode::use_iterator UI = LD->use_begin(), UE = LD->use_end(); UI != UE;
7922 if (UI.getUse().getResNo() == 1) // Ignore uses of the chain result.
7928 SDValue Addr = LD->getOperand(1);
7929 SDValue Vector = N->getOperand(0);
7930 // Search for a use of the address operand that is an increment.
7931 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), UE =
7932 Addr.getNode()->use_end(); UI != UE; ++UI) {
7934 if (User->getOpcode() != ISD::ADD
7935 || UI.getUse().getResNo() != Addr.getResNo())
7938 // Check that the add is independent of the load. Otherwise, folding it
7939 // would create a cycle.
7940 if (User->isPredecessorOf(LD) || LD->isPredecessorOf(User))
7942 // Also check that add is not used in the vector operand. This would also
7944 if (User->isPredecessorOf(Vector.getNode()))
7947 // If the increment is a constant, it must match the memory ref size.
7948 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7949 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7950 uint32_t IncVal = CInc->getZExtValue();
7951 unsigned NumBytes = VT.getScalarSizeInBits() / 8;
7952 if (IncVal != NumBytes)
7954 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
7957 SmallVector<SDValue, 8> Ops;
7958 Ops.push_back(LD->getOperand(0)); // Chain
7960 Ops.push_back(Vector); // The vector to be inserted
7961 Ops.push_back(N->getOperand(2)); // The lane to be inserted in the vector
7963 Ops.push_back(Addr);
7966 EVT Tys[3] = { VT, MVT::i64, MVT::Other };
7967 SDVTList SDTys = DAG.getVTList(Tys);
7968 unsigned NewOp = IsLaneOp ? AArch64ISD::LD1LANEpost : AArch64ISD::LD1DUPpost;
7969 SDValue UpdN = DAG.getMemIntrinsicNode(NewOp, SDLoc(N), SDTys, Ops,
7971 LoadSDN->getMemOperand());
7974 SmallVector<SDValue, 2> NewResults;
7975 NewResults.push_back(SDValue(LD, 0)); // The result of load
7976 NewResults.push_back(SDValue(UpdN.getNode(), 2)); // Chain
7977 DCI.CombineTo(LD, NewResults);
7978 DCI.CombineTo(N, SDValue(UpdN.getNode(), 0)); // Dup/Inserted Result
7979 DCI.CombineTo(User, SDValue(UpdN.getNode(), 1)); // Write back register
7986 /// Target-specific DAG combine function for NEON load/store intrinsics
7987 /// to merge base address updates.
7988 static SDValue performNEONPostLDSTCombine(SDNode *N,
7989 TargetLowering::DAGCombinerInfo &DCI,
7990 SelectionDAG &DAG) {
7991 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7994 unsigned AddrOpIdx = N->getNumOperands() - 1;
7995 SDValue Addr = N->getOperand(AddrOpIdx);
7997 // Search for a use of the address operand that is an increment.
7998 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
7999 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8001 if (User->getOpcode() != ISD::ADD ||
8002 UI.getUse().getResNo() != Addr.getResNo())
8005 // Check that the add is independent of the load/store. Otherwise, folding
8006 // it would create a cycle.
8007 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8010 // Find the new opcode for the updating load/store.
8011 bool IsStore = false;
8012 bool IsLaneOp = false;
8013 bool IsDupOp = false;
8014 unsigned NewOpc = 0;
8015 unsigned NumVecs = 0;
8016 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8018 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8019 case Intrinsic::aarch64_neon_ld2: NewOpc = AArch64ISD::LD2post;
8021 case Intrinsic::aarch64_neon_ld3: NewOpc = AArch64ISD::LD3post;
8023 case Intrinsic::aarch64_neon_ld4: NewOpc = AArch64ISD::LD4post;
8025 case Intrinsic::aarch64_neon_st2: NewOpc = AArch64ISD::ST2post;
8026 NumVecs = 2; IsStore = true; break;
8027 case Intrinsic::aarch64_neon_st3: NewOpc = AArch64ISD::ST3post;
8028 NumVecs = 3; IsStore = true; break;
8029 case Intrinsic::aarch64_neon_st4: NewOpc = AArch64ISD::ST4post;
8030 NumVecs = 4; IsStore = true; break;
8031 case Intrinsic::aarch64_neon_ld1x2: NewOpc = AArch64ISD::LD1x2post;
8033 case Intrinsic::aarch64_neon_ld1x3: NewOpc = AArch64ISD::LD1x3post;
8035 case Intrinsic::aarch64_neon_ld1x4: NewOpc = AArch64ISD::LD1x4post;
8037 case Intrinsic::aarch64_neon_st1x2: NewOpc = AArch64ISD::ST1x2post;
8038 NumVecs = 2; IsStore = true; break;
8039 case Intrinsic::aarch64_neon_st1x3: NewOpc = AArch64ISD::ST1x3post;
8040 NumVecs = 3; IsStore = true; break;
8041 case Intrinsic::aarch64_neon_st1x4: NewOpc = AArch64ISD::ST1x4post;
8042 NumVecs = 4; IsStore = true; break;
8043 case Intrinsic::aarch64_neon_ld2r: NewOpc = AArch64ISD::LD2DUPpost;
8044 NumVecs = 2; IsDupOp = true; break;
8045 case Intrinsic::aarch64_neon_ld3r: NewOpc = AArch64ISD::LD3DUPpost;
8046 NumVecs = 3; IsDupOp = true; break;
8047 case Intrinsic::aarch64_neon_ld4r: NewOpc = AArch64ISD::LD4DUPpost;
8048 NumVecs = 4; IsDupOp = true; break;
8049 case Intrinsic::aarch64_neon_ld2lane: NewOpc = AArch64ISD::LD2LANEpost;
8050 NumVecs = 2; IsLaneOp = true; break;
8051 case Intrinsic::aarch64_neon_ld3lane: NewOpc = AArch64ISD::LD3LANEpost;
8052 NumVecs = 3; IsLaneOp = true; break;
8053 case Intrinsic::aarch64_neon_ld4lane: NewOpc = AArch64ISD::LD4LANEpost;
8054 NumVecs = 4; IsLaneOp = true; break;
8055 case Intrinsic::aarch64_neon_st2lane: NewOpc = AArch64ISD::ST2LANEpost;
8056 NumVecs = 2; IsStore = true; IsLaneOp = true; break;
8057 case Intrinsic::aarch64_neon_st3lane: NewOpc = AArch64ISD::ST3LANEpost;
8058 NumVecs = 3; IsStore = true; IsLaneOp = true; break;
8059 case Intrinsic::aarch64_neon_st4lane: NewOpc = AArch64ISD::ST4LANEpost;
8060 NumVecs = 4; IsStore = true; IsLaneOp = true; break;
8065 VecTy = N->getOperand(2).getValueType();
8067 VecTy = N->getValueType(0);
8069 // If the increment is a constant, it must match the memory ref size.
8070 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8071 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8072 uint32_t IncVal = CInc->getZExtValue();
8073 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8074 if (IsLaneOp || IsDupOp)
8075 NumBytes /= VecTy.getVectorNumElements();
8076 if (IncVal != NumBytes)
8078 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
8080 SmallVector<SDValue, 8> Ops;
8081 Ops.push_back(N->getOperand(0)); // Incoming chain
8082 // Load lane and store have vector list as input.
8083 if (IsLaneOp || IsStore)
8084 for (unsigned i = 2; i < AddrOpIdx; ++i)
8085 Ops.push_back(N->getOperand(i));
8086 Ops.push_back(Addr); // Base register
8091 unsigned NumResultVecs = (IsStore ? 0 : NumVecs);
8093 for (n = 0; n < NumResultVecs; ++n)
8095 Tys[n++] = MVT::i64; // Type of write back register
8096 Tys[n] = MVT::Other; // Type of the chain
8097 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs + 2));
8099 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8100 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys, Ops,
8101 MemInt->getMemoryVT(),
8102 MemInt->getMemOperand());
8105 std::vector<SDValue> NewResults;
8106 for (unsigned i = 0; i < NumResultVecs; ++i) {
8107 NewResults.push_back(SDValue(UpdN.getNode(), i));
8109 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1));
8110 DCI.CombineTo(N, NewResults);
8111 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8118 // Checks to see if the value is the prescribed width and returns information
8119 // about its extension mode.
8121 bool checkValueWidth(SDValue V, unsigned width, ISD::LoadExtType &ExtType) {
8122 ExtType = ISD::NON_EXTLOAD;
8123 switch(V.getNode()->getOpcode()) {
8127 LoadSDNode *LoadNode = cast<LoadSDNode>(V.getNode());
8128 if ((LoadNode->getMemoryVT() == MVT::i8 && width == 8)
8129 || (LoadNode->getMemoryVT() == MVT::i16 && width == 16)) {
8130 ExtType = LoadNode->getExtensionType();
8135 case ISD::AssertSext: {
8136 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
8137 if ((TypeNode->getVT() == MVT::i8 && width == 8)
8138 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
8139 ExtType = ISD::SEXTLOAD;
8144 case ISD::AssertZext: {
8145 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
8146 if ((TypeNode->getVT() == MVT::i8 && width == 8)
8147 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
8148 ExtType = ISD::ZEXTLOAD;
8154 case ISD::TargetConstant: {
8155 if (std::abs(cast<ConstantSDNode>(V.getNode())->getSExtValue()) <
8165 // This function does a whole lot of voodoo to determine if the tests are
8166 // equivalent without and with a mask. Essentially what happens is that given a
8169 // +-------------+ +-------------+ +-------------+ +-------------+
8170 // | Input | | AddConstant | | CompConstant| | CC |
8171 // +-------------+ +-------------+ +-------------+ +-------------+
8173 // V V | +----------+
8174 // +-------------+ +----+ | |
8175 // | ADD | |0xff| | |
8176 // +-------------+ +----+ | |
8179 // +-------------+ | |
8181 // +-------------+ | |
8190 // The AND node may be safely removed for some combinations of inputs. In
8191 // particular we need to take into account the extension type of the Input,
8192 // the exact values of AddConstant, CompConstant, and CC, along with the nominal
8193 // width of the input (this can work for any width inputs, the above graph is
8194 // specific to 8 bits.
8196 // The specific equations were worked out by generating output tables for each
8197 // AArch64CC value in terms of and AddConstant (w1), CompConstant(w2). The
8198 // problem was simplified by working with 4 bit inputs, which means we only
8199 // needed to reason about 24 distinct bit patterns: 8 patterns unique to zero
8200 // extension (8,15), 8 patterns unique to sign extensions (-8,-1), and 8
8201 // patterns present in both extensions (0,7). For every distinct set of
8202 // AddConstant and CompConstants bit patterns we can consider the masked and
8203 // unmasked versions to be equivalent if the result of this function is true for
8204 // all 16 distinct bit patterns of for the current extension type of Input (w0).
8207 // and w10, w8, #0x0f
8209 // cset w9, AArch64CC
8211 // cset w11, AArch64CC
8216 // Since the above function shows when the outputs are equivalent it defines
8217 // when it is safe to remove the AND. Unfortunately it only runs on AArch64 and
8218 // would be expensive to run during compiles. The equations below were written
8219 // in a test harness that confirmed they gave equivalent outputs to the above
8220 // for all inputs function, so they can be used determine if the removal is
8223 // isEquivalentMaskless() is the code for testing if the AND can be removed
8224 // factored out of the DAG recognition as the DAG can take several forms.
8227 bool isEquivalentMaskless(unsigned CC, unsigned width,
8228 ISD::LoadExtType ExtType, signed AddConstant,
8229 signed CompConstant) {
8230 // By being careful about our equations and only writing the in term
8231 // symbolic values and well known constants (0, 1, -1, MaxUInt) we can
8232 // make them generally applicable to all bit widths.
8233 signed MaxUInt = (1 << width);
8235 // For the purposes of these comparisons sign extending the type is
8236 // equivalent to zero extending the add and displacing it by half the integer
8237 // width. Provided we are careful and make sure our equations are valid over
8238 // the whole range we can just adjust the input and avoid writing equations
8239 // for sign extended inputs.
8240 if (ExtType == ISD::SEXTLOAD)
8241 AddConstant -= (1 << (width-1));
8245 case AArch64CC::GT: {
8246 if ((AddConstant == 0) ||
8247 (CompConstant == MaxUInt - 1 && AddConstant < 0) ||
8248 (AddConstant >= 0 && CompConstant < 0) ||
8249 (AddConstant <= 0 && CompConstant <= 0 && CompConstant < AddConstant))
8253 case AArch64CC::GE: {
8254 if ((AddConstant == 0) ||
8255 (AddConstant >= 0 && CompConstant <= 0) ||
8256 (AddConstant <= 0 && CompConstant <= 0 && CompConstant <= AddConstant))
8260 case AArch64CC::LS: {
8261 if ((AddConstant >= 0 && CompConstant < 0) ||
8262 (AddConstant <= 0 && CompConstant >= -1 &&
8263 CompConstant < AddConstant + MaxUInt))
8267 case AArch64CC::MI: {
8268 if ((AddConstant == 0) ||
8269 (AddConstant > 0 && CompConstant <= 0) ||
8270 (AddConstant < 0 && CompConstant <= AddConstant))
8274 case AArch64CC::HS: {
8275 if ((AddConstant >= 0 && CompConstant <= 0) ||
8276 (AddConstant <= 0 && CompConstant >= 0 &&
8277 CompConstant <= AddConstant + MaxUInt))
8281 case AArch64CC::NE: {
8282 if ((AddConstant > 0 && CompConstant < 0) ||
8283 (AddConstant < 0 && CompConstant >= 0 &&
8284 CompConstant < AddConstant + MaxUInt) ||
8285 (AddConstant >= 0 && CompConstant >= 0 &&
8286 CompConstant >= AddConstant) ||
8287 (AddConstant <= 0 && CompConstant < 0 && CompConstant < AddConstant))
8296 case AArch64CC::Invalid:
8304 SDValue performCONDCombine(SDNode *N,
8305 TargetLowering::DAGCombinerInfo &DCI,
8306 SelectionDAG &DAG, unsigned CCIndex,
8307 unsigned CmpIndex) {
8308 unsigned CC = cast<ConstantSDNode>(N->getOperand(CCIndex))->getSExtValue();
8309 SDNode *SubsNode = N->getOperand(CmpIndex).getNode();
8310 unsigned CondOpcode = SubsNode->getOpcode();
8312 if (CondOpcode != AArch64ISD::SUBS)
8315 // There is a SUBS feeding this condition. Is it fed by a mask we can
8318 SDNode *AndNode = SubsNode->getOperand(0).getNode();
8319 unsigned MaskBits = 0;
8321 if (AndNode->getOpcode() != ISD::AND)
8324 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(AndNode->getOperand(1))) {
8325 uint32_t CNV = CN->getZExtValue();
8328 else if (CNV == 65535)
8335 SDValue AddValue = AndNode->getOperand(0);
8337 if (AddValue.getOpcode() != ISD::ADD)
8340 // The basic dag structure is correct, grab the inputs and validate them.
8342 SDValue AddInputValue1 = AddValue.getNode()->getOperand(0);
8343 SDValue AddInputValue2 = AddValue.getNode()->getOperand(1);
8344 SDValue SubsInputValue = SubsNode->getOperand(1);
8346 // The mask is present and the provenance of all the values is a smaller type,
8347 // lets see if the mask is superfluous.
8349 if (!isa<ConstantSDNode>(AddInputValue2.getNode()) ||
8350 !isa<ConstantSDNode>(SubsInputValue.getNode()))
8353 ISD::LoadExtType ExtType;
8355 if (!checkValueWidth(SubsInputValue, MaskBits, ExtType) ||
8356 !checkValueWidth(AddInputValue2, MaskBits, ExtType) ||
8357 !checkValueWidth(AddInputValue1, MaskBits, ExtType) )
8360 if(!isEquivalentMaskless(CC, MaskBits, ExtType,
8361 cast<ConstantSDNode>(AddInputValue2.getNode())->getSExtValue(),
8362 cast<ConstantSDNode>(SubsInputValue.getNode())->getSExtValue()))
8365 // The AND is not necessary, remove it.
8367 SDVTList VTs = DAG.getVTList(SubsNode->getValueType(0),
8368 SubsNode->getValueType(1));
8369 SDValue Ops[] = { AddValue, SubsNode->getOperand(1) };
8371 SDValue NewValue = DAG.getNode(CondOpcode, SDLoc(SubsNode), VTs, Ops);
8372 DAG.ReplaceAllUsesWith(SubsNode, NewValue.getNode());
8374 return SDValue(N, 0);
8377 // Optimize compare with zero and branch.
8378 static SDValue performBRCONDCombine(SDNode *N,
8379 TargetLowering::DAGCombinerInfo &DCI,
8380 SelectionDAG &DAG) {
8381 SDValue NV = performCONDCombine(N, DCI, DAG, 2, 3);
8384 SDValue Chain = N->getOperand(0);
8385 SDValue Dest = N->getOperand(1);
8386 SDValue CCVal = N->getOperand(2);
8387 SDValue Cmp = N->getOperand(3);
8389 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
8390 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
8391 if (CC != AArch64CC::EQ && CC != AArch64CC::NE)
8394 unsigned CmpOpc = Cmp.getOpcode();
8395 if (CmpOpc != AArch64ISD::ADDS && CmpOpc != AArch64ISD::SUBS)
8398 // Only attempt folding if there is only one use of the flag and no use of the
8400 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
8403 SDValue LHS = Cmp.getOperand(0);
8404 SDValue RHS = Cmp.getOperand(1);
8406 assert(LHS.getValueType() == RHS.getValueType() &&
8407 "Expected the value type to be the same for both operands!");
8408 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
8411 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
8412 std::swap(LHS, RHS);
8414 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
8417 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
8418 LHS.getOpcode() == ISD::SRL)
8421 // Fold the compare into the branch instruction.
8423 if (CC == AArch64CC::EQ)
8424 BR = DAG.getNode(AArch64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
8426 BR = DAG.getNode(AArch64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
8428 // Do not add new nodes to DAG combiner worklist.
8429 DCI.CombineTo(N, BR, false);
8434 // vselect (v1i1 setcc) ->
8435 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
8436 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
8437 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
8439 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
8440 SDValue N0 = N->getOperand(0);
8441 EVT CCVT = N0.getValueType();
8443 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
8444 CCVT.getVectorElementType() != MVT::i1)
8447 EVT ResVT = N->getValueType(0);
8448 EVT CmpVT = N0.getOperand(0).getValueType();
8449 // Only combine when the result type is of the same size as the compared
8451 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
8454 SDValue IfTrue = N->getOperand(1);
8455 SDValue IfFalse = N->getOperand(2);
8457 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
8458 N0.getOperand(0), N0.getOperand(1),
8459 cast<CondCodeSDNode>(N0.getOperand(2))->get());
8460 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
8464 /// A vector select: "(select vL, vR, (setcc LHS, RHS))" is best performed with
8465 /// the compare-mask instructions rather than going via NZCV, even if LHS and
8466 /// RHS are really scalar. This replaces any scalar setcc in the above pattern
8467 /// with a vector one followed by a DUP shuffle on the result.
8468 static SDValue performSelectCombine(SDNode *N, SelectionDAG &DAG) {
8469 SDValue N0 = N->getOperand(0);
8470 EVT ResVT = N->getValueType(0);
8472 if (N0.getOpcode() != ISD::SETCC || N0.getValueType() != MVT::i1)
8475 // If NumMaskElts == 0, the comparison is larger than select result. The
8476 // largest real NEON comparison is 64-bits per lane, which means the result is
8477 // at most 32-bits and an illegal vector. Just bail out for now.
8478 EVT SrcVT = N0.getOperand(0).getValueType();
8480 // Don't try to do this optimization when the setcc itself has i1 operands.
8481 // There are no legal vectors of i1, so this would be pointless.
8482 if (SrcVT == MVT::i1)
8485 int NumMaskElts = ResVT.getSizeInBits() / SrcVT.getSizeInBits();
8486 if (!ResVT.isVector() || NumMaskElts == 0)
8489 SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumMaskElts);
8490 EVT CCVT = SrcVT.changeVectorElementTypeToInteger();
8492 // First perform a vector comparison, where lane 0 is the one we're interested
8496 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(0));
8498 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(1));
8499 SDValue SetCC = DAG.getNode(ISD::SETCC, DL, CCVT, LHS, RHS, N0.getOperand(2));
8501 // Now duplicate the comparison mask we want across all other lanes.
8502 SmallVector<int, 8> DUPMask(CCVT.getVectorNumElements(), 0);
8503 SDValue Mask = DAG.getVectorShuffle(CCVT, DL, SetCC, SetCC, DUPMask.data());
8504 Mask = DAG.getNode(ISD::BITCAST, DL,
8505 ResVT.changeVectorElementTypeToInteger(), Mask);
8507 return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2));
8510 SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N,
8511 DAGCombinerInfo &DCI) const {
8512 SelectionDAG &DAG = DCI.DAG;
8513 switch (N->getOpcode()) {
8518 return performAddSubLongCombine(N, DCI, DAG);
8520 return performXorCombine(N, DAG, DCI, Subtarget);
8522 return performMulCombine(N, DAG, DCI, Subtarget);
8523 case ISD::SINT_TO_FP:
8524 case ISD::UINT_TO_FP:
8525 return performIntToFpCombine(N, DAG, Subtarget);
8527 return performORCombine(N, DCI, Subtarget);
8528 case ISD::INTRINSIC_WO_CHAIN:
8529 return performIntrinsicCombine(N, DCI, Subtarget);
8530 case ISD::ANY_EXTEND:
8531 case ISD::ZERO_EXTEND:
8532 case ISD::SIGN_EXTEND:
8533 return performExtendCombine(N, DCI, DAG);
8535 return performBitcastCombine(N, DCI, DAG);
8536 case ISD::CONCAT_VECTORS:
8537 return performConcatVectorsCombine(N, DCI, DAG);
8539 return performSelectCombine(N, DAG);
8541 return performVSelectCombine(N, DCI.DAG);
8543 return performSTORECombine(N, DCI, DAG, Subtarget);
8544 case AArch64ISD::BRCOND:
8545 return performBRCONDCombine(N, DCI, DAG);
8546 case AArch64ISD::CSEL:
8547 return performCONDCombine(N, DCI, DAG, 2, 3);
8548 case AArch64ISD::DUP:
8549 return performPostLD1Combine(N, DCI, false);
8550 case ISD::INSERT_VECTOR_ELT:
8551 return performPostLD1Combine(N, DCI, true);
8552 case ISD::INTRINSIC_VOID:
8553 case ISD::INTRINSIC_W_CHAIN:
8554 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8555 case Intrinsic::aarch64_neon_ld2:
8556 case Intrinsic::aarch64_neon_ld3:
8557 case Intrinsic::aarch64_neon_ld4:
8558 case Intrinsic::aarch64_neon_ld1x2:
8559 case Intrinsic::aarch64_neon_ld1x3:
8560 case Intrinsic::aarch64_neon_ld1x4:
8561 case Intrinsic::aarch64_neon_ld2lane:
8562 case Intrinsic::aarch64_neon_ld3lane:
8563 case Intrinsic::aarch64_neon_ld4lane:
8564 case Intrinsic::aarch64_neon_ld2r:
8565 case Intrinsic::aarch64_neon_ld3r:
8566 case Intrinsic::aarch64_neon_ld4r:
8567 case Intrinsic::aarch64_neon_st2:
8568 case Intrinsic::aarch64_neon_st3:
8569 case Intrinsic::aarch64_neon_st4:
8570 case Intrinsic::aarch64_neon_st1x2:
8571 case Intrinsic::aarch64_neon_st1x3:
8572 case Intrinsic::aarch64_neon_st1x4:
8573 case Intrinsic::aarch64_neon_st2lane:
8574 case Intrinsic::aarch64_neon_st3lane:
8575 case Intrinsic::aarch64_neon_st4lane:
8576 return performNEONPostLDSTCombine(N, DCI, DAG);
8584 // Check if the return value is used as only a return value, as otherwise
8585 // we can't perform a tail-call. In particular, we need to check for
8586 // target ISD nodes that are returns and any other "odd" constructs
8587 // that the generic analysis code won't necessarily catch.
8588 bool AArch64TargetLowering::isUsedByReturnOnly(SDNode *N,
8589 SDValue &Chain) const {
8590 if (N->getNumValues() != 1)
8592 if (!N->hasNUsesOfValue(1, 0))
8595 SDValue TCChain = Chain;
8596 SDNode *Copy = *N->use_begin();
8597 if (Copy->getOpcode() == ISD::CopyToReg) {
8598 // If the copy has a glue operand, we conservatively assume it isn't safe to
8599 // perform a tail call.
8600 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
8603 TCChain = Copy->getOperand(0);
8604 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
8607 bool HasRet = false;
8608 for (SDNode *Node : Copy->uses()) {
8609 if (Node->getOpcode() != AArch64ISD::RET_FLAG)
8621 // Return whether the an instruction can potentially be optimized to a tail
8622 // call. This will cause the optimizers to attempt to move, or duplicate,
8623 // return instructions to help enable tail call optimizations for this
8625 bool AArch64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
8626 if (!CI->isTailCall())
8632 bool AArch64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
8634 ISD::MemIndexedMode &AM,
8636 SelectionDAG &DAG) const {
8637 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
8640 Base = Op->getOperand(0);
8641 // All of the indexed addressing mode instructions take a signed
8642 // 9 bit immediate offset.
8643 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
8644 int64_t RHSC = (int64_t)RHS->getZExtValue();
8645 if (RHSC >= 256 || RHSC <= -256)
8647 IsInc = (Op->getOpcode() == ISD::ADD);
8648 Offset = Op->getOperand(1);
8654 bool AArch64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
8656 ISD::MemIndexedMode &AM,
8657 SelectionDAG &DAG) const {
8660 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8661 VT = LD->getMemoryVT();
8662 Ptr = LD->getBasePtr();
8663 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8664 VT = ST->getMemoryVT();
8665 Ptr = ST->getBasePtr();
8670 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
8672 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
8676 bool AArch64TargetLowering::getPostIndexedAddressParts(
8677 SDNode *N, SDNode *Op, SDValue &Base, SDValue &Offset,
8678 ISD::MemIndexedMode &AM, SelectionDAG &DAG) const {
8681 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8682 VT = LD->getMemoryVT();
8683 Ptr = LD->getBasePtr();
8684 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8685 VT = ST->getMemoryVT();
8686 Ptr = ST->getBasePtr();
8691 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
8693 // Post-indexing updates the base, so it's not a valid transform
8694 // if that's not the same as the load's pointer.
8697 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
8701 static void ReplaceBITCASTResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
8702 SelectionDAG &DAG) {
8704 SDValue Op = N->getOperand(0);
8706 if (N->getValueType(0) != MVT::i16 || Op.getValueType() != MVT::f16)
8710 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL, MVT::f32,
8711 DAG.getUNDEF(MVT::i32), Op,
8712 DAG.getTargetConstant(AArch64::hsub, MVT::i32)),
8714 Op = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op);
8715 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Op));
8718 void AArch64TargetLowering::ReplaceNodeResults(
8719 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
8720 switch (N->getOpcode()) {
8722 llvm_unreachable("Don't know how to custom expand this");
8724 ReplaceBITCASTResults(N, Results, DAG);
8726 case ISD::FP_TO_UINT:
8727 case ISD::FP_TO_SINT:
8728 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
8729 // Let normal code take care of it by not adding anything to Results.
8734 bool AArch64TargetLowering::useLoadStackGuardNode() const {
8738 bool AArch64TargetLowering::combineRepeatedFPDivisors(unsigned NumUsers) const {
8739 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
8740 // reciprocal if there are three or more FDIVs.
8741 return NumUsers > 2;
8744 TargetLoweringBase::LegalizeTypeAction
8745 AArch64TargetLowering::getPreferredVectorAction(EVT VT) const {
8746 MVT SVT = VT.getSimpleVT();
8747 // During type legalization, we prefer to widen v1i8, v1i16, v1i32 to v8i8,
8748 // v4i16, v2i32 instead of to promote.
8749 if (SVT == MVT::v1i8 || SVT == MVT::v1i16 || SVT == MVT::v1i32
8750 || SVT == MVT::v1f32)
8751 return TypeWidenVector;
8753 return TargetLoweringBase::getPreferredVectorAction(VT);
8756 // Loads and stores less than 128-bits are already atomic; ones above that
8757 // are doomed anyway, so defer to the default libcall and blame the OS when
8759 bool AArch64TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
8760 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
8764 // Loads and stores less than 128-bits are already atomic; ones above that
8765 // are doomed anyway, so defer to the default libcall and blame the OS when
8767 bool AArch64TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
8768 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
8772 // For the real atomic operations, we have ldxr/stxr up to 128 bits,
8773 bool AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
8774 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
8778 bool AArch64TargetLowering::hasLoadLinkedStoreConditional() const {
8782 Value *AArch64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
8783 AtomicOrdering Ord) const {
8784 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
8785 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
8786 bool IsAcquire = isAtLeastAcquire(Ord);
8788 // Since i128 isn't legal and intrinsics don't get type-lowered, the ldrexd
8789 // intrinsic must return {i64, i64} and we have to recombine them into a
8790 // single i128 here.
8791 if (ValTy->getPrimitiveSizeInBits() == 128) {
8793 IsAcquire ? Intrinsic::aarch64_ldaxp : Intrinsic::aarch64_ldxp;
8794 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int);
8796 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
8797 Value *LoHi = Builder.CreateCall(Ldxr, Addr, "lohi");
8799 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
8800 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
8801 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
8802 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
8803 return Builder.CreateOr(
8804 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64");
8807 Type *Tys[] = { Addr->getType() };
8809 IsAcquire ? Intrinsic::aarch64_ldaxr : Intrinsic::aarch64_ldxr;
8810 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int, Tys);
8812 return Builder.CreateTruncOrBitCast(
8813 Builder.CreateCall(Ldxr, Addr),
8814 cast<PointerType>(Addr->getType())->getElementType());
8817 Value *AArch64TargetLowering::emitStoreConditional(IRBuilder<> &Builder,
8818 Value *Val, Value *Addr,
8819 AtomicOrdering Ord) const {
8820 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
8821 bool IsRelease = isAtLeastRelease(Ord);
8823 // Since the intrinsics must have legal type, the i128 intrinsics take two
8824 // parameters: "i64, i64". We must marshal Val into the appropriate form
8826 if (Val->getType()->getPrimitiveSizeInBits() == 128) {
8828 IsRelease ? Intrinsic::aarch64_stlxp : Intrinsic::aarch64_stxp;
8829 Function *Stxr = Intrinsic::getDeclaration(M, Int);
8830 Type *Int64Ty = Type::getInt64Ty(M->getContext());
8832 Value *Lo = Builder.CreateTrunc(Val, Int64Ty, "lo");
8833 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 64), Int64Ty, "hi");
8834 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
8835 return Builder.CreateCall3(Stxr, Lo, Hi, Addr);
8839 IsRelease ? Intrinsic::aarch64_stlxr : Intrinsic::aarch64_stxr;
8840 Type *Tys[] = { Addr->getType() };
8841 Function *Stxr = Intrinsic::getDeclaration(M, Int, Tys);
8843 return Builder.CreateCall2(
8844 Stxr, Builder.CreateZExtOrBitCast(
8845 Val, Stxr->getFunctionType()->getParamType(0)),
8849 bool AArch64TargetLowering::functionArgumentNeedsConsecutiveRegisters(
8850 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
8851 return Ty->isArrayTy();