1 //===-- ARM64ISelLowering.cpp - ARM64 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 ARM64TargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "arm64-lower"
16 #include "ARM64ISelLowering.h"
17 #include "ARM64PerfectShuffle.h"
18 #include "ARM64Subtarget.h"
19 #include "ARM64CallingConv.h"
20 #include "ARM64MachineFunctionInfo.h"
21 #include "ARM64TargetMachine.h"
22 #include "ARM64TargetObjectFile.h"
23 #include "MCTargetDesc/ARM64AddressingModes.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/Intrinsics.h"
31 #include "llvm/IR/Type.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetOptions.h"
39 STATISTIC(NumTailCalls, "Number of tail calls");
40 STATISTIC(NumShiftInserts, "Number of vector shift inserts");
42 // This option should go away when tail calls fully work.
44 EnableARM64TailCalls("arm64-tail-calls", cl::Hidden,
45 cl::desc("Generate ARM64 tail calls (TEMPORARY OPTION)."),
49 StrictAlign("arm64-strict-align", cl::Hidden,
50 cl::desc("Disallow all unaligned memory accesses"));
52 // Place holder until extr generation is tested fully.
54 EnableARM64ExtrGeneration("arm64-extr-generation", cl::Hidden,
55 cl::desc("Allow ARM64 (or (shift)(shift))->extract"),
59 EnableARM64SlrGeneration("arm64-shift-insert-generation", cl::Hidden,
60 cl::desc("Allow ARM64 SLI/SRI formation"),
63 //===----------------------------------------------------------------------===//
64 // ARM64 Lowering public interface.
65 //===----------------------------------------------------------------------===//
66 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
67 if (TM.getSubtarget<ARM64Subtarget>().isTargetDarwin())
68 return new ARM64_MachoTargetObjectFile();
70 return new ARM64_ELFTargetObjectFile();
73 ARM64TargetLowering::ARM64TargetLowering(ARM64TargetMachine &TM)
74 : TargetLowering(TM, createTLOF(TM)) {
75 Subtarget = &TM.getSubtarget<ARM64Subtarget>();
77 // ARM64 doesn't have comparisons which set GPRs or setcc instructions, so
78 // we have to make something up. Arbitrarily, choose ZeroOrOne.
79 setBooleanContents(ZeroOrOneBooleanContent);
80 // When comparing vectors the result sets the different elements in the
81 // vector to all-one or all-zero.
82 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
84 // Set up the register classes.
85 addRegisterClass(MVT::i32, &ARM64::GPR32allRegClass);
86 addRegisterClass(MVT::i64, &ARM64::GPR64allRegClass);
87 addRegisterClass(MVT::f16, &ARM64::FPR16RegClass);
88 addRegisterClass(MVT::f32, &ARM64::FPR32RegClass);
89 addRegisterClass(MVT::f64, &ARM64::FPR64RegClass);
90 addRegisterClass(MVT::f128, &ARM64::FPR128RegClass);
91 addRegisterClass(MVT::v16i8, &ARM64::FPR8RegClass);
92 addRegisterClass(MVT::v8i16, &ARM64::FPR16RegClass);
94 // Someone set us up the NEON.
95 addDRTypeForNEON(MVT::v2f32);
96 addDRTypeForNEON(MVT::v8i8);
97 addDRTypeForNEON(MVT::v4i16);
98 addDRTypeForNEON(MVT::v2i32);
99 addDRTypeForNEON(MVT::v1i64);
100 addDRTypeForNEON(MVT::v1f64);
102 addQRTypeForNEON(MVT::v4f32);
103 addQRTypeForNEON(MVT::v2f64);
104 addQRTypeForNEON(MVT::v16i8);
105 addQRTypeForNEON(MVT::v8i16);
106 addQRTypeForNEON(MVT::v4i32);
107 addQRTypeForNEON(MVT::v2i64);
109 // Compute derived properties from the register classes
110 computeRegisterProperties();
112 // Provide all sorts of operation actions
113 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
114 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
115 setOperationAction(ISD::SETCC, MVT::i32, Custom);
116 setOperationAction(ISD::SETCC, MVT::i64, Custom);
117 setOperationAction(ISD::SETCC, MVT::f32, Custom);
118 setOperationAction(ISD::SETCC, MVT::f64, Custom);
119 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
120 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
121 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
122 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
123 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
124 setOperationAction(ISD::SELECT, MVT::i32, Custom);
125 setOperationAction(ISD::SELECT, MVT::i64, Custom);
126 setOperationAction(ISD::SELECT, MVT::f32, Custom);
127 setOperationAction(ISD::SELECT, MVT::f64, Custom);
128 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
129 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
130 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
131 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
132 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
133 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
135 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
136 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
137 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
139 setOperationAction(ISD::FREM, MVT::f32, Expand);
140 setOperationAction(ISD::FREM, MVT::f64, Expand);
141 setOperationAction(ISD::FREM, MVT::f80, Expand);
143 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
144 // silliness like this:
145 setOperationAction(ISD::FABS, MVT::v1f64, Expand);
146 setOperationAction(ISD::FADD, MVT::v1f64, Expand);
147 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
148 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
149 setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
150 setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
151 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
152 setOperationAction(ISD::FMA, MVT::v1f64, Expand);
153 setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
154 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
155 setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
156 setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
157 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
158 setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
159 setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
160 setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
161 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
162 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
163 setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
164 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
165 setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
166 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
167 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
168 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
169 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);
171 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
172 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
173 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
174 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
175 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);
177 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
179 // Custom lowering hooks are needed for XOR
180 // to fold it into CSINC/CSINV.
181 setOperationAction(ISD::XOR, MVT::i32, Custom);
182 setOperationAction(ISD::XOR, MVT::i64, Custom);
184 // Virtually no operation on f128 is legal, but LLVM can't expand them when
185 // there's a valid register class, so we need custom operations in most cases.
186 setOperationAction(ISD::FABS, MVT::f128, Expand);
187 setOperationAction(ISD::FADD, MVT::f128, Custom);
188 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
189 setOperationAction(ISD::FCOS, MVT::f128, Expand);
190 setOperationAction(ISD::FDIV, MVT::f128, Custom);
191 setOperationAction(ISD::FMA, MVT::f128, Expand);
192 setOperationAction(ISD::FMUL, MVT::f128, Custom);
193 setOperationAction(ISD::FNEG, MVT::f128, Expand);
194 setOperationAction(ISD::FPOW, MVT::f128, Expand);
195 setOperationAction(ISD::FREM, MVT::f128, Expand);
196 setOperationAction(ISD::FRINT, MVT::f128, Expand);
197 setOperationAction(ISD::FSIN, MVT::f128, Expand);
198 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
199 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
200 setOperationAction(ISD::FSUB, MVT::f128, Custom);
201 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
202 setOperationAction(ISD::SETCC, MVT::f128, Custom);
203 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
204 setOperationAction(ISD::SELECT, MVT::f128, Custom);
205 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
206 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
208 // Lowering for many of the conversions is actually specified by the non-f128
209 // type. The LowerXXX function will be trivial when f128 isn't involved.
210 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
211 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
212 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
213 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
214 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
215 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
216 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
217 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
218 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
219 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
220 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
221 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
222 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
223 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
226 setOperationAction(ISD::ATOMIC_SWAP, MVT::i128, Custom);
227 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i128, Custom);
228 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i128, Custom);
229 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i128, Custom);
230 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i128, Custom);
231 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i128, Custom);
232 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i128, Custom);
233 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i128, Custom);
234 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i128, Custom);
235 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i128, Custom);
236 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i128, Custom);
237 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i128, Custom);
238 // These are surprisingly difficult. The only single-copy atomic 128-bit
239 // instruction on AArch64 is stxp (when it succeeds). So a store can safely
240 // become a simple swap, but a load can only be determined to have been atomic
241 // if storing the same value back succeeds.
242 setOperationAction(ISD::ATOMIC_LOAD, MVT::i128, Custom);
243 setOperationAction(ISD::ATOMIC_STORE, MVT::i128, Expand);
245 // Variable arguments.
246 setOperationAction(ISD::VASTART, MVT::Other, Custom);
247 setOperationAction(ISD::VAARG, MVT::Other, Custom);
248 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
249 setOperationAction(ISD::VAEND, MVT::Other, Expand);
251 // Variable-sized objects.
252 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
253 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
254 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
256 // Exception handling.
257 // FIXME: These are guesses. Has this been defined yet?
258 setExceptionPointerRegister(ARM64::X0);
259 setExceptionSelectorRegister(ARM64::X1);
261 // Constant pool entries
262 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
265 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
267 // Add/Sub overflow ops with MVT::Glues are lowered to CPSR dependences.
268 setOperationAction(ISD::ADDC, MVT::i32, Custom);
269 setOperationAction(ISD::ADDE, MVT::i32, Custom);
270 setOperationAction(ISD::SUBC, MVT::i32, Custom);
271 setOperationAction(ISD::SUBE, MVT::i32, Custom);
272 setOperationAction(ISD::ADDC, MVT::i64, Custom);
273 setOperationAction(ISD::ADDE, MVT::i64, Custom);
274 setOperationAction(ISD::SUBC, MVT::i64, Custom);
275 setOperationAction(ISD::SUBE, MVT::i64, Custom);
277 // ARM64 lacks both left-rotate and popcount instructions.
278 setOperationAction(ISD::ROTL, MVT::i32, Expand);
279 setOperationAction(ISD::ROTL, MVT::i64, Expand);
281 // ARM64 doesn't have a direct vector ->f32 conversion instructions for
282 // elements smaller than i32, so promote the input to i32 first.
283 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
284 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
285 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
286 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
287 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
288 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
289 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
290 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
291 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
293 // ARM64 doesn't have {U|S}MUL_LOHI.
294 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
295 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
297 // ARM64 doesn't have MUL.2d:
298 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
300 // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
301 // counterparts, which ARM64 supports directly.
302 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
303 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
304 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
305 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
307 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
308 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
310 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
311 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
312 setOperationAction(ISD::SREM, MVT::i32, Expand);
313 setOperationAction(ISD::SREM, MVT::i64, Expand);
314 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
315 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
316 setOperationAction(ISD::UREM, MVT::i32, Expand);
317 setOperationAction(ISD::UREM, MVT::i64, Expand);
319 // Custom lower Add/Sub/Mul with overflow.
320 setOperationAction(ISD::SADDO, MVT::i32, Custom);
321 setOperationAction(ISD::SADDO, MVT::i64, Custom);
322 setOperationAction(ISD::UADDO, MVT::i32, Custom);
323 setOperationAction(ISD::UADDO, MVT::i64, Custom);
324 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
325 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
326 setOperationAction(ISD::USUBO, MVT::i32, Custom);
327 setOperationAction(ISD::USUBO, MVT::i64, Custom);
328 setOperationAction(ISD::SMULO, MVT::i32, Custom);
329 setOperationAction(ISD::SMULO, MVT::i64, Custom);
330 setOperationAction(ISD::UMULO, MVT::i32, Custom);
331 setOperationAction(ISD::UMULO, MVT::i64, Custom);
333 setOperationAction(ISD::FSIN, MVT::f32, Expand);
334 setOperationAction(ISD::FSIN, MVT::f64, Expand);
335 setOperationAction(ISD::FCOS, MVT::f32, Expand);
336 setOperationAction(ISD::FCOS, MVT::f64, Expand);
337 setOperationAction(ISD::FPOW, MVT::f32, Expand);
338 setOperationAction(ISD::FPOW, MVT::f64, Expand);
339 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
340 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
342 // ARM64 has implementations of a lot of rounding-like FP operations.
343 static MVT RoundingTypes[] = { MVT::f32, MVT::f64, MVT::v2f32,
344 MVT::v4f32, MVT::v2f64 };
345 for (unsigned I = 0; I < array_lengthof(RoundingTypes); ++I) {
346 MVT Ty = RoundingTypes[I];
347 setOperationAction(ISD::FFLOOR, Ty, Legal);
348 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
349 setOperationAction(ISD::FCEIL, Ty, Legal);
350 setOperationAction(ISD::FRINT, Ty, Legal);
351 setOperationAction(ISD::FTRUNC, Ty, Legal);
352 setOperationAction(ISD::FROUND, Ty, Legal);
355 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
357 if (Subtarget->isTargetMachO()) {
358 // For iOS, we don't want to the normal expansion of a libcall to
359 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
361 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
362 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
364 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
365 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
368 // ARM64 does not have floating-point extending loads, i1 sign-extending load,
369 // floating-point truncating stores, or v2i32->v2i16 truncating store.
370 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
371 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
372 setLoadExtAction(ISD::EXTLOAD, MVT::f80, Expand);
373 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Expand);
374 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
375 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
376 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
377 setTruncStoreAction(MVT::f128, MVT::f80, Expand);
378 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
379 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
380 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
381 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
382 // Indexed loads and stores are supported.
383 for (unsigned im = (unsigned)ISD::PRE_INC;
384 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
385 setIndexedLoadAction(im, MVT::i8, Legal);
386 setIndexedLoadAction(im, MVT::i16, Legal);
387 setIndexedLoadAction(im, MVT::i32, Legal);
388 setIndexedLoadAction(im, MVT::i64, Legal);
389 setIndexedLoadAction(im, MVT::f64, Legal);
390 setIndexedLoadAction(im, MVT::f32, Legal);
391 setIndexedStoreAction(im, MVT::i8, Legal);
392 setIndexedStoreAction(im, MVT::i16, Legal);
393 setIndexedStoreAction(im, MVT::i32, Legal);
394 setIndexedStoreAction(im, MVT::i64, Legal);
395 setIndexedStoreAction(im, MVT::f64, Legal);
396 setIndexedStoreAction(im, MVT::f32, Legal);
399 // Likewise, narrowing and extending vector loads/stores aren't handled
401 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
402 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
404 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,
407 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
408 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
409 setTruncStoreAction((MVT::SimpleValueType)VT,
410 (MVT::SimpleValueType)InnerVT, Expand);
411 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
412 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
413 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
417 setOperationAction(ISD::TRAP, MVT::Other, Legal);
418 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
420 // We combine OR nodes for bitfield operations.
421 setTargetDAGCombine(ISD::OR);
423 // Vector add and sub nodes may conceal a high-half opportunity.
424 // Also, try to fold ADD into CSINC/CSINV..
425 setTargetDAGCombine(ISD::ADD);
426 setTargetDAGCombine(ISD::SUB);
428 setTargetDAGCombine(ISD::XOR);
429 setTargetDAGCombine(ISD::SINT_TO_FP);
430 setTargetDAGCombine(ISD::UINT_TO_FP);
432 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
434 setTargetDAGCombine(ISD::ANY_EXTEND);
435 setTargetDAGCombine(ISD::ZERO_EXTEND);
436 setTargetDAGCombine(ISD::SIGN_EXTEND);
437 setTargetDAGCombine(ISD::BITCAST);
438 setTargetDAGCombine(ISD::CONCAT_VECTORS);
439 setTargetDAGCombine(ISD::STORE);
441 setTargetDAGCombine(ISD::MUL);
443 setTargetDAGCombine(ISD::VSELECT);
445 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
446 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
447 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
449 setStackPointerRegisterToSaveRestore(ARM64::SP);
451 setSchedulingPreference(Sched::Hybrid);
454 MaskAndBranchFoldingIsLegal = true;
456 setMinFunctionAlignment(2);
458 RequireStrictAlign = StrictAlign;
461 void ARM64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
462 if (VT == MVT::v2f32) {
463 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
464 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
466 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
467 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
468 } else if (VT == MVT::v2f64 || VT == MVT::v4f32) {
469 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
470 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
472 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
473 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
476 // Mark vector float intrinsics as expand.
477 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
478 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
479 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
480 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
481 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
482 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
483 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
484 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
485 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
486 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
489 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
490 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
491 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
492 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
493 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
494 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
495 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
496 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
497 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
498 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
499 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
500 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
502 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
503 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
504 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
505 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
507 // CNT supports only B element sizes.
508 if (VT != MVT::v8i8 && VT != MVT::v16i8)
509 setOperationAction(ISD::CTPOP, VT.getSimpleVT(), Expand);
511 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
512 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
513 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
514 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
515 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
517 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
518 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
521 void ARM64TargetLowering::addDRTypeForNEON(MVT VT) {
522 addRegisterClass(VT, &ARM64::FPR64RegClass);
523 addTypeForNEON(VT, MVT::v2i32);
526 void ARM64TargetLowering::addQRTypeForNEON(MVT VT) {
527 addRegisterClass(VT, &ARM64::FPR128RegClass);
528 addTypeForNEON(VT, MVT::v4i32);
531 EVT ARM64TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
534 return VT.changeVectorElementTypeToInteger();
537 /// computeMaskedBitsForTargetNode - Determine which of the bits specified in
538 /// Mask are known to be either zero or one and return them in the
539 /// KnownZero/KnownOne bitsets.
540 void ARM64TargetLowering::computeMaskedBitsForTargetNode(
541 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
542 const SelectionDAG &DAG, unsigned Depth) const {
543 switch (Op.getOpcode()) {
546 case ARM64ISD::CSEL: {
547 APInt KnownZero2, KnownOne2;
548 DAG.ComputeMaskedBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
549 DAG.ComputeMaskedBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
550 KnownZero &= KnownZero2;
551 KnownOne &= KnownOne2;
554 case ISD::INTRINSIC_W_CHAIN: {
555 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
556 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
559 case Intrinsic::arm64_ldaxr:
560 case Intrinsic::arm64_ldxr: {
561 unsigned BitWidth = KnownOne.getBitWidth();
562 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
563 unsigned MemBits = VT.getScalarType().getSizeInBits();
564 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
570 case ISD::INTRINSIC_WO_CHAIN:
571 case ISD::INTRINSIC_VOID: {
572 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
576 case Intrinsic::arm64_neon_umaxv:
577 case Intrinsic::arm64_neon_uminv: {
578 // Figure out the datatype of the vector operand. The UMINV instruction
579 // will zero extend the result, so we can mark as known zero all the
580 // bits larger than the element datatype. 32-bit or larget doesn't need
581 // this as those are legal types and will be handled by isel directly.
582 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
583 unsigned BitWidth = KnownZero.getBitWidth();
584 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
585 assert(BitWidth >= 8 && "Unexpected width!");
586 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
588 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
589 assert(BitWidth >= 16 && "Unexpected width!");
590 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
600 MVT ARM64TargetLowering::getScalarShiftAmountTy(EVT LHSTy) const {
604 unsigned ARM64TargetLowering::getMaximalGlobalOffset() const {
605 // FIXME: On ARM64, this depends on the type.
606 // Basically, the addressable offsets are o to 4095 * Ty.getSizeInBytes().
607 // and the offset has to be a multiple of the related size in bytes.
612 ARM64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
613 const TargetLibraryInfo *libInfo) const {
614 return ARM64::createFastISel(funcInfo, libInfo);
617 const char *ARM64TargetLowering::getTargetNodeName(unsigned Opcode) const {
621 case ARM64ISD::CALL: return "ARM64ISD::CALL";
622 case ARM64ISD::ADRP: return "ARM64ISD::ADRP";
623 case ARM64ISD::ADDlow: return "ARM64ISD::ADDlow";
624 case ARM64ISD::LOADgot: return "ARM64ISD::LOADgot";
625 case ARM64ISD::RET_FLAG: return "ARM64ISD::RET_FLAG";
626 case ARM64ISD::BRCOND: return "ARM64ISD::BRCOND";
627 case ARM64ISD::CSEL: return "ARM64ISD::CSEL";
628 case ARM64ISD::FCSEL: return "ARM64ISD::FCSEL";
629 case ARM64ISD::CSINV: return "ARM64ISD::CSINV";
630 case ARM64ISD::CSNEG: return "ARM64ISD::CSNEG";
631 case ARM64ISD::CSINC: return "ARM64ISD::CSINC";
632 case ARM64ISD::THREAD_POINTER: return "ARM64ISD::THREAD_POINTER";
633 case ARM64ISD::TLSDESC_CALL: return "ARM64ISD::TLSDESC_CALL";
634 case ARM64ISD::ADC: return "ARM64ISD::ADC";
635 case ARM64ISD::SBC: return "ARM64ISD::SBC";
636 case ARM64ISD::ADDS: return "ARM64ISD::ADDS";
637 case ARM64ISD::SUBS: return "ARM64ISD::SUBS";
638 case ARM64ISD::ADCS: return "ARM64ISD::ADCS";
639 case ARM64ISD::SBCS: return "ARM64ISD::SBCS";
640 case ARM64ISD::ANDS: return "ARM64ISD::ANDS";
641 case ARM64ISD::FCMP: return "ARM64ISD::FCMP";
642 case ARM64ISD::FMIN: return "ARM64ISD::FMIN";
643 case ARM64ISD::FMAX: return "ARM64ISD::FMAX";
644 case ARM64ISD::DUP: return "ARM64ISD::DUP";
645 case ARM64ISD::DUPLANE8: return "ARM64ISD::DUPLANE8";
646 case ARM64ISD::DUPLANE16: return "ARM64ISD::DUPLANE16";
647 case ARM64ISD::DUPLANE32: return "ARM64ISD::DUPLANE32";
648 case ARM64ISD::DUPLANE64: return "ARM64ISD::DUPLANE64";
649 case ARM64ISD::MOVI: return "ARM64ISD::MOVI";
650 case ARM64ISD::MOVIshift: return "ARM64ISD::MOVIshift";
651 case ARM64ISD::MOVIedit: return "ARM64ISD::MOVIedit";
652 case ARM64ISD::MOVImsl: return "ARM64ISD::MOVImsl";
653 case ARM64ISD::FMOV: return "ARM64ISD::FMOV";
654 case ARM64ISD::MVNIshift: return "ARM64ISD::MVNIshift";
655 case ARM64ISD::MVNImsl: return "ARM64ISD::MVNImsl";
656 case ARM64ISD::BICi: return "ARM64ISD::BICi";
657 case ARM64ISD::ORRi: return "ARM64ISD::ORRi";
658 case ARM64ISD::NEG: return "ARM64ISD::NEG";
659 case ARM64ISD::EXTR: return "ARM64ISD::EXTR";
660 case ARM64ISD::ZIP1: return "ARM64ISD::ZIP1";
661 case ARM64ISD::ZIP2: return "ARM64ISD::ZIP2";
662 case ARM64ISD::UZP1: return "ARM64ISD::UZP1";
663 case ARM64ISD::UZP2: return "ARM64ISD::UZP2";
664 case ARM64ISD::TRN1: return "ARM64ISD::TRN1";
665 case ARM64ISD::TRN2: return "ARM64ISD::TRN2";
666 case ARM64ISD::REV16: return "ARM64ISD::REV16";
667 case ARM64ISD::REV32: return "ARM64ISD::REV32";
668 case ARM64ISD::REV64: return "ARM64ISD::REV64";
669 case ARM64ISD::EXT: return "ARM64ISD::EXT";
670 case ARM64ISD::VSHL: return "ARM64ISD::VSHL";
671 case ARM64ISD::VLSHR: return "ARM64ISD::VLSHR";
672 case ARM64ISD::VASHR: return "ARM64ISD::VASHR";
673 case ARM64ISD::CMEQ: return "ARM64ISD::CMEQ";
674 case ARM64ISD::CMGE: return "ARM64ISD::CMGE";
675 case ARM64ISD::CMGT: return "ARM64ISD::CMGT";
676 case ARM64ISD::CMHI: return "ARM64ISD::CMHI";
677 case ARM64ISD::CMHS: return "ARM64ISD::CMHS";
678 case ARM64ISD::FCMEQ: return "ARM64ISD::FCMEQ";
679 case ARM64ISD::FCMGE: return "ARM64ISD::FCMGE";
680 case ARM64ISD::FCMGT: return "ARM64ISD::FCMGT";
681 case ARM64ISD::CMEQz: return "ARM64ISD::CMEQz";
682 case ARM64ISD::CMGEz: return "ARM64ISD::CMGEz";
683 case ARM64ISD::CMGTz: return "ARM64ISD::CMGTz";
684 case ARM64ISD::CMLEz: return "ARM64ISD::CMLEz";
685 case ARM64ISD::CMLTz: return "ARM64ISD::CMLTz";
686 case ARM64ISD::FCMEQz: return "ARM64ISD::FCMEQz";
687 case ARM64ISD::FCMGEz: return "ARM64ISD::FCMGEz";
688 case ARM64ISD::FCMGTz: return "ARM64ISD::FCMGTz";
689 case ARM64ISD::FCMLEz: return "ARM64ISD::FCMLEz";
690 case ARM64ISD::FCMLTz: return "ARM64ISD::FCMLTz";
691 case ARM64ISD::NOT: return "ARM64ISD::NOT";
692 case ARM64ISD::BIT: return "ARM64ISD::BIT";
693 case ARM64ISD::CBZ: return "ARM64ISD::CBZ";
694 case ARM64ISD::CBNZ: return "ARM64ISD::CBNZ";
695 case ARM64ISD::TBZ: return "ARM64ISD::TBZ";
696 case ARM64ISD::TBNZ: return "ARM64ISD::TBNZ";
697 case ARM64ISD::TC_RETURN: return "ARM64ISD::TC_RETURN";
698 case ARM64ISD::SITOF: return "ARM64ISD::SITOF";
699 case ARM64ISD::UITOF: return "ARM64ISD::UITOF";
700 case ARM64ISD::SQSHL_I: return "ARM64ISD::SQSHL_I";
701 case ARM64ISD::UQSHL_I: return "ARM64ISD::UQSHL_I";
702 case ARM64ISD::SRSHR_I: return "ARM64ISD::SRSHR_I";
703 case ARM64ISD::URSHR_I: return "ARM64ISD::URSHR_I";
704 case ARM64ISD::SQSHLU_I: return "ARM64ISD::SQSHLU_I";
705 case ARM64ISD::WrapperLarge: return "ARM64ISD::WrapperLarge";
709 static void getExclusiveOperation(unsigned Size, AtomicOrdering Ord,
710 unsigned &LdrOpc, unsigned &StrOpc) {
711 static unsigned LoadBares[] = { ARM64::LDXRB, ARM64::LDXRH, ARM64::LDXRW,
712 ARM64::LDXRX, ARM64::LDXPX };
713 static unsigned LoadAcqs[] = { ARM64::LDAXRB, ARM64::LDAXRH, ARM64::LDAXRW,
714 ARM64::LDAXRX, ARM64::LDAXPX };
715 static unsigned StoreBares[] = { ARM64::STXRB, ARM64::STXRH, ARM64::STXRW,
716 ARM64::STXRX, ARM64::STXPX };
717 static unsigned StoreRels[] = { ARM64::STLXRB, ARM64::STLXRH, ARM64::STLXRW,
718 ARM64::STLXRX, ARM64::STLXPX };
720 unsigned *LoadOps, *StoreOps;
721 if (Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent)
726 if (Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent)
727 StoreOps = StoreRels;
729 StoreOps = StoreBares;
731 assert(isPowerOf2_32(Size) && Size <= 16 &&
732 "unsupported size for atomic binary op!");
734 LdrOpc = LoadOps[Log2_32(Size)];
735 StrOpc = StoreOps[Log2_32(Size)];
738 MachineBasicBlock *ARM64TargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
739 MachineBasicBlock *BB,
740 unsigned Size) const {
741 unsigned dest = MI->getOperand(0).getReg();
742 unsigned ptr = MI->getOperand(1).getReg();
743 unsigned oldval = MI->getOperand(2).getReg();
744 unsigned newval = MI->getOperand(3).getReg();
745 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(4).getImm());
746 unsigned scratch = BB->getParent()->getRegInfo().createVirtualRegister(
747 &ARM64::GPR32RegClass);
748 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
749 DebugLoc dl = MI->getDebugLoc();
751 // FIXME: We currently always generate a seq_cst operation; we should
752 // be able to relax this in some cases.
753 unsigned ldrOpc, strOpc;
754 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
756 MachineFunction *MF = BB->getParent();
757 const BasicBlock *LLVM_BB = BB->getBasicBlock();
758 MachineFunction::iterator It = BB;
759 ++It; // insert the new blocks after the current block
761 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
762 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
763 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
764 MF->insert(It, loop1MBB);
765 MF->insert(It, loop2MBB);
766 MF->insert(It, exitMBB);
768 // Transfer the remainder of BB and its successor edges to exitMBB.
769 exitMBB->splice(exitMBB->begin(), BB,
770 std::next(MachineBasicBlock::iterator(MI)), BB->end());
771 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
775 // fallthrough --> loop1MBB
776 BB->addSuccessor(loop1MBB);
783 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
784 BuildMI(BB, dl, TII->get(Size == 8 ? ARM64::SUBSXrr : ARM64::SUBSWrr))
785 .addReg(Size == 8 ? ARM64::XZR : ARM64::WZR, RegState::Define)
788 BuildMI(BB, dl, TII->get(ARM64::Bcc)).addImm(ARM64CC::NE).addMBB(exitMBB);
789 BB->addSuccessor(loop2MBB);
790 BB->addSuccessor(exitMBB);
793 // strex scratch, newval, [ptr]
797 BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
798 BuildMI(BB, dl, TII->get(ARM64::CBNZW)).addReg(scratch).addMBB(loop1MBB);
799 BB->addSuccessor(loop1MBB);
800 BB->addSuccessor(exitMBB);
806 MI->eraseFromParent(); // The instruction is gone now.
812 ARM64TargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
813 unsigned Size, unsigned BinOpcode) const {
814 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
815 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
817 const BasicBlock *LLVM_BB = BB->getBasicBlock();
818 MachineFunction *MF = BB->getParent();
819 MachineFunction::iterator It = BB;
822 unsigned dest = MI->getOperand(0).getReg();
823 unsigned ptr = MI->getOperand(1).getReg();
824 unsigned incr = MI->getOperand(2).getReg();
825 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
826 DebugLoc dl = MI->getDebugLoc();
828 unsigned ldrOpc, strOpc;
829 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
831 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
832 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
833 MF->insert(It, loopMBB);
834 MF->insert(It, exitMBB);
836 // Transfer the remainder of BB and its successor edges to exitMBB.
837 exitMBB->splice(exitMBB->begin(), BB,
838 std::next(MachineBasicBlock::iterator(MI)), BB->end());
839 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
841 MachineRegisterInfo &RegInfo = MF->getRegInfo();
842 unsigned scratch = RegInfo.createVirtualRegister(&ARM64::GPR32RegClass);
846 : RegInfo.createVirtualRegister(Size == 8 ? &ARM64::GPR64RegClass
847 : &ARM64::GPR32RegClass);
851 // fallthrough --> loopMBB
852 BB->addSuccessor(loopMBB);
856 // <binop> scratch2, dest, incr
857 // stxr scratch, scratch2, ptr
858 // cbnz scratch, loopMBB
859 // fallthrough --> exitMBB
861 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
863 // operand order needs to go the other way for NAND
864 if (BinOpcode == ARM64::BICWrr || BinOpcode == ARM64::BICXrr)
865 BuildMI(BB, dl, TII->get(BinOpcode), scratch2).addReg(incr).addReg(dest);
867 BuildMI(BB, dl, TII->get(BinOpcode), scratch2).addReg(dest).addReg(incr);
870 BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
871 BuildMI(BB, dl, TII->get(ARM64::CBNZW)).addReg(scratch).addMBB(loopMBB);
873 BB->addSuccessor(loopMBB);
874 BB->addSuccessor(exitMBB);
880 MI->eraseFromParent(); // The instruction is gone now.
885 MachineBasicBlock *ARM64TargetLowering::EmitAtomicBinary128(
886 MachineInstr *MI, MachineBasicBlock *BB, unsigned BinOpcodeLo,
887 unsigned BinOpcodeHi) const {
888 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
889 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
891 const BasicBlock *LLVM_BB = BB->getBasicBlock();
892 MachineFunction *MF = BB->getParent();
893 MachineFunction::iterator It = BB;
896 unsigned DestLo = MI->getOperand(0).getReg();
897 unsigned DestHi = MI->getOperand(1).getReg();
898 unsigned Ptr = MI->getOperand(2).getReg();
899 unsigned IncrLo = MI->getOperand(3).getReg();
900 unsigned IncrHi = MI->getOperand(4).getReg();
901 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(5).getImm());
902 DebugLoc DL = MI->getDebugLoc();
904 unsigned LdrOpc, StrOpc;
905 getExclusiveOperation(16, Ord, LdrOpc, StrOpc);
907 MachineBasicBlock *LoopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
908 MachineBasicBlock *ExitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
909 MF->insert(It, LoopMBB);
910 MF->insert(It, ExitMBB);
912 // Transfer the remainder of BB and its successor edges to exitMBB.
913 ExitMBB->splice(ExitMBB->begin(), BB,
914 std::next(MachineBasicBlock::iterator(MI)), BB->end());
915 ExitMBB->transferSuccessorsAndUpdatePHIs(BB);
917 MachineRegisterInfo &RegInfo = MF->getRegInfo();
918 unsigned ScratchRes = RegInfo.createVirtualRegister(&ARM64::GPR32RegClass);
919 unsigned ScratchLo = IncrLo, ScratchHi = IncrHi;
921 assert(BinOpcodeHi && "Expect neither or both opcodes to be defined");
922 ScratchLo = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
923 ScratchHi = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
928 // fallthrough --> LoopMBB
929 BB->addSuccessor(LoopMBB);
932 // ldxp DestLo, DestHi, Ptr
933 // <binoplo> ScratchLo, DestLo, IncrLo
934 // <binophi> ScratchHi, DestHi, IncrHi
935 // stxp ScratchRes, ScratchLo, ScratchHi, ptr
936 // cbnz ScratchRes, LoopMBB
937 // fallthrough --> ExitMBB
939 BuildMI(BB, DL, TII->get(LdrOpc), DestLo)
940 .addReg(DestHi, RegState::Define)
943 // operand order needs to go the other way for NAND
944 if (BinOpcodeLo == ARM64::BICXrr) {
945 std::swap(IncrLo, DestLo);
946 std::swap(IncrHi, DestHi);
949 BuildMI(BB, DL, TII->get(BinOpcodeLo), ScratchLo).addReg(DestLo).addReg(
951 BuildMI(BB, DL, TII->get(BinOpcodeHi), ScratchHi).addReg(DestHi).addReg(
955 BuildMI(BB, DL, TII->get(StrOpc), ScratchRes)
959 BuildMI(BB, DL, TII->get(ARM64::CBNZW)).addReg(ScratchRes).addMBB(LoopMBB);
961 BB->addSuccessor(LoopMBB);
962 BB->addSuccessor(ExitMBB);
968 MI->eraseFromParent(); // The instruction is gone now.
974 ARM64TargetLowering::EmitAtomicCmpSwap128(MachineInstr *MI,
975 MachineBasicBlock *BB) const {
976 unsigned DestLo = MI->getOperand(0).getReg();
977 unsigned DestHi = MI->getOperand(1).getReg();
978 unsigned Ptr = MI->getOperand(2).getReg();
979 unsigned OldValLo = MI->getOperand(3).getReg();
980 unsigned OldValHi = MI->getOperand(4).getReg();
981 unsigned NewValLo = MI->getOperand(5).getReg();
982 unsigned NewValHi = MI->getOperand(6).getReg();
983 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(7).getImm());
984 unsigned ScratchRes = BB->getParent()->getRegInfo().createVirtualRegister(
985 &ARM64::GPR32RegClass);
986 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
987 DebugLoc DL = MI->getDebugLoc();
989 unsigned LdrOpc, StrOpc;
990 getExclusiveOperation(16, Ord, LdrOpc, StrOpc);
992 MachineFunction *MF = BB->getParent();
993 const BasicBlock *LLVM_BB = BB->getBasicBlock();
994 MachineFunction::iterator It = BB;
995 ++It; // insert the new blocks after the current block
997 MachineBasicBlock *Loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
998 MachineBasicBlock *Loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
999 MachineBasicBlock *ExitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1000 MF->insert(It, Loop1MBB);
1001 MF->insert(It, Loop2MBB);
1002 MF->insert(It, ExitMBB);
1004 // Transfer the remainder of BB and its successor edges to exitMBB.
1005 ExitMBB->splice(ExitMBB->begin(), BB,
1006 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1007 ExitMBB->transferSuccessorsAndUpdatePHIs(BB);
1011 // fallthrough --> Loop1MBB
1012 BB->addSuccessor(Loop1MBB);
1015 // ldxp DestLo, DestHi, [Ptr]
1016 // cmp DestLo, OldValLo
1017 // sbc xzr, DestHi, OldValHi
1020 BuildMI(BB, DL, TII->get(LdrOpc), DestLo)
1021 .addReg(DestHi, RegState::Define)
1023 BuildMI(BB, DL, TII->get(ARM64::SUBSXrr), ARM64::XZR).addReg(DestLo).addReg(
1025 BuildMI(BB, DL, TII->get(ARM64::SBCXr), ARM64::XZR).addReg(DestHi).addReg(
1028 BuildMI(BB, DL, TII->get(ARM64::Bcc)).addImm(ARM64CC::NE).addMBB(ExitMBB);
1029 BB->addSuccessor(Loop2MBB);
1030 BB->addSuccessor(ExitMBB);
1033 // stxp ScratchRes, NewValLo, NewValHi, [Ptr]
1034 // cbnz ScratchRes, Loop1MBB
1036 BuildMI(BB, DL, TII->get(StrOpc), ScratchRes)
1040 BuildMI(BB, DL, TII->get(ARM64::CBNZW)).addReg(ScratchRes).addMBB(Loop1MBB);
1041 BB->addSuccessor(Loop1MBB);
1042 BB->addSuccessor(ExitMBB);
1048 MI->eraseFromParent(); // The instruction is gone now.
1053 MachineBasicBlock *ARM64TargetLowering::EmitAtomicMinMax128(
1054 MachineInstr *MI, MachineBasicBlock *BB, unsigned CondCode) const {
1055 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
1056 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1058 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1059 MachineFunction *MF = BB->getParent();
1060 MachineFunction::iterator It = BB;
1063 unsigned DestLo = MI->getOperand(0).getReg();
1064 unsigned DestHi = MI->getOperand(1).getReg();
1065 unsigned Ptr = MI->getOperand(2).getReg();
1066 unsigned IncrLo = MI->getOperand(3).getReg();
1067 unsigned IncrHi = MI->getOperand(4).getReg();
1068 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(5).getImm());
1069 DebugLoc DL = MI->getDebugLoc();
1071 unsigned LdrOpc, StrOpc;
1072 getExclusiveOperation(16, Ord, LdrOpc, StrOpc);
1074 MachineBasicBlock *LoopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1075 MachineBasicBlock *ExitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1076 MF->insert(It, LoopMBB);
1077 MF->insert(It, ExitMBB);
1079 // Transfer the remainder of BB and its successor edges to exitMBB.
1080 ExitMBB->splice(ExitMBB->begin(), BB,
1081 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1082 ExitMBB->transferSuccessorsAndUpdatePHIs(BB);
1084 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1085 unsigned ScratchRes = RegInfo.createVirtualRegister(&ARM64::GPR32RegClass);
1086 unsigned ScratchLo = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
1087 unsigned ScratchHi = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
1091 // fallthrough --> LoopMBB
1092 BB->addSuccessor(LoopMBB);
1095 // ldxp DestLo, DestHi, Ptr
1096 // cmp ScratchLo, DestLo, IncrLo
1097 // sbc xzr, ScratchHi, DestHi, IncrHi
1098 // csel ScratchLo, DestLo, IncrLo, <cmp-op>
1099 // csel ScratchHi, DestHi, IncrHi, <cmp-op>
1100 // stxp ScratchRes, ScratchLo, ScratchHi, ptr
1101 // cbnz ScratchRes, LoopMBB
1102 // fallthrough --> ExitMBB
1104 BuildMI(BB, DL, TII->get(LdrOpc), DestLo)
1105 .addReg(DestHi, RegState::Define)
1108 BuildMI(BB, DL, TII->get(ARM64::SUBSXrr), ARM64::XZR).addReg(DestLo).addReg(
1110 BuildMI(BB, DL, TII->get(ARM64::SBCXr), ARM64::XZR).addReg(DestHi).addReg(
1113 BuildMI(BB, DL, TII->get(ARM64::CSELXr), ScratchLo)
1117 BuildMI(BB, DL, TII->get(ARM64::CSELXr), ScratchHi)
1122 BuildMI(BB, DL, TII->get(StrOpc), ScratchRes)
1126 BuildMI(BB, DL, TII->get(ARM64::CBNZW)).addReg(ScratchRes).addMBB(LoopMBB);
1128 BB->addSuccessor(LoopMBB);
1129 BB->addSuccessor(ExitMBB);
1135 MI->eraseFromParent(); // The instruction is gone now.
1141 ARM64TargetLowering::EmitF128CSEL(MachineInstr *MI,
1142 MachineBasicBlock *MBB) const {
1143 // We materialise the F128CSEL pseudo-instruction as some control flow and a
1147 // [... previous instrs leading to comparison ...]
1153 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
1155 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1156 MachineFunction *MF = MBB->getParent();
1157 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
1158 DebugLoc DL = MI->getDebugLoc();
1159 MachineFunction::iterator It = MBB;
1162 unsigned DestReg = MI->getOperand(0).getReg();
1163 unsigned IfTrueReg = MI->getOperand(1).getReg();
1164 unsigned IfFalseReg = MI->getOperand(2).getReg();
1165 unsigned CondCode = MI->getOperand(3).getImm();
1166 bool CPSRKilled = MI->getOperand(4).isKill();
1168 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
1169 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
1170 MF->insert(It, TrueBB);
1171 MF->insert(It, EndBB);
1173 // Transfer rest of current basic-block to EndBB
1174 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
1176 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
1178 BuildMI(MBB, DL, TII->get(ARM64::Bcc)).addImm(CondCode).addMBB(TrueBB);
1179 BuildMI(MBB, DL, TII->get(ARM64::B)).addMBB(EndBB);
1180 MBB->addSuccessor(TrueBB);
1181 MBB->addSuccessor(EndBB);
1183 // TrueBB falls through to the end.
1184 TrueBB->addSuccessor(EndBB);
1187 TrueBB->addLiveIn(ARM64::CPSR);
1188 EndBB->addLiveIn(ARM64::CPSR);
1191 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(ARM64::PHI), DestReg)
1197 MI->eraseFromParent();
1202 ARM64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
1203 MachineBasicBlock *BB) const {
1204 switch (MI->getOpcode()) {
1209 assert(0 && "Unexpected instruction for custom inserter!");
1212 case ARM64::ATOMIC_LOAD_ADD_I8:
1213 return EmitAtomicBinary(MI, BB, 1, ARM64::ADDWrr);
1214 case ARM64::ATOMIC_LOAD_ADD_I16:
1215 return EmitAtomicBinary(MI, BB, 2, ARM64::ADDWrr);
1216 case ARM64::ATOMIC_LOAD_ADD_I32:
1217 return EmitAtomicBinary(MI, BB, 4, ARM64::ADDWrr);
1218 case ARM64::ATOMIC_LOAD_ADD_I64:
1219 return EmitAtomicBinary(MI, BB, 8, ARM64::ADDXrr);
1220 case ARM64::ATOMIC_LOAD_ADD_I128:
1221 return EmitAtomicBinary128(MI, BB, ARM64::ADDSXrr, ARM64::ADCXr);
1223 case ARM64::ATOMIC_LOAD_AND_I8:
1224 return EmitAtomicBinary(MI, BB, 1, ARM64::ANDWrr);
1225 case ARM64::ATOMIC_LOAD_AND_I16:
1226 return EmitAtomicBinary(MI, BB, 2, ARM64::ANDWrr);
1227 case ARM64::ATOMIC_LOAD_AND_I32:
1228 return EmitAtomicBinary(MI, BB, 4, ARM64::ANDWrr);
1229 case ARM64::ATOMIC_LOAD_AND_I64:
1230 return EmitAtomicBinary(MI, BB, 8, ARM64::ANDXrr);
1231 case ARM64::ATOMIC_LOAD_AND_I128:
1232 return EmitAtomicBinary128(MI, BB, ARM64::ANDXrr, ARM64::ANDXrr);
1234 case ARM64::ATOMIC_LOAD_OR_I8:
1235 return EmitAtomicBinary(MI, BB, 1, ARM64::ORRWrr);
1236 case ARM64::ATOMIC_LOAD_OR_I16:
1237 return EmitAtomicBinary(MI, BB, 2, ARM64::ORRWrr);
1238 case ARM64::ATOMIC_LOAD_OR_I32:
1239 return EmitAtomicBinary(MI, BB, 4, ARM64::ORRWrr);
1240 case ARM64::ATOMIC_LOAD_OR_I64:
1241 return EmitAtomicBinary(MI, BB, 8, ARM64::ORRXrr);
1242 case ARM64::ATOMIC_LOAD_OR_I128:
1243 return EmitAtomicBinary128(MI, BB, ARM64::ORRXrr, ARM64::ORRXrr);
1245 case ARM64::ATOMIC_LOAD_XOR_I8:
1246 return EmitAtomicBinary(MI, BB, 1, ARM64::EORWrr);
1247 case ARM64::ATOMIC_LOAD_XOR_I16:
1248 return EmitAtomicBinary(MI, BB, 2, ARM64::EORWrr);
1249 case ARM64::ATOMIC_LOAD_XOR_I32:
1250 return EmitAtomicBinary(MI, BB, 4, ARM64::EORWrr);
1251 case ARM64::ATOMIC_LOAD_XOR_I64:
1252 return EmitAtomicBinary(MI, BB, 8, ARM64::EORXrr);
1253 case ARM64::ATOMIC_LOAD_XOR_I128:
1254 return EmitAtomicBinary128(MI, BB, ARM64::EORXrr, ARM64::EORXrr);
1256 case ARM64::ATOMIC_LOAD_NAND_I8:
1257 return EmitAtomicBinary(MI, BB, 1, ARM64::BICWrr);
1258 case ARM64::ATOMIC_LOAD_NAND_I16:
1259 return EmitAtomicBinary(MI, BB, 2, ARM64::BICWrr);
1260 case ARM64::ATOMIC_LOAD_NAND_I32:
1261 return EmitAtomicBinary(MI, BB, 4, ARM64::BICWrr);
1262 case ARM64::ATOMIC_LOAD_NAND_I64:
1263 return EmitAtomicBinary(MI, BB, 8, ARM64::BICXrr);
1264 case ARM64::ATOMIC_LOAD_NAND_I128:
1265 return EmitAtomicBinary128(MI, BB, ARM64::BICXrr, ARM64::BICXrr);
1267 case ARM64::ATOMIC_LOAD_SUB_I8:
1268 return EmitAtomicBinary(MI, BB, 1, ARM64::SUBWrr);
1269 case ARM64::ATOMIC_LOAD_SUB_I16:
1270 return EmitAtomicBinary(MI, BB, 2, ARM64::SUBWrr);
1271 case ARM64::ATOMIC_LOAD_SUB_I32:
1272 return EmitAtomicBinary(MI, BB, 4, ARM64::SUBWrr);
1273 case ARM64::ATOMIC_LOAD_SUB_I64:
1274 return EmitAtomicBinary(MI, BB, 8, ARM64::SUBXrr);
1275 case ARM64::ATOMIC_LOAD_SUB_I128:
1276 return EmitAtomicBinary128(MI, BB, ARM64::SUBSXrr, ARM64::SBCXr);
1278 case ARM64::ATOMIC_LOAD_MIN_I128:
1279 return EmitAtomicMinMax128(MI, BB, ARM64CC::LT);
1281 case ARM64::ATOMIC_LOAD_MAX_I128:
1282 return EmitAtomicMinMax128(MI, BB, ARM64CC::GT);
1284 case ARM64::ATOMIC_LOAD_UMIN_I128:
1285 return EmitAtomicMinMax128(MI, BB, ARM64CC::CC);
1287 case ARM64::ATOMIC_LOAD_UMAX_I128:
1288 return EmitAtomicMinMax128(MI, BB, ARM64CC::HI);
1290 case ARM64::ATOMIC_SWAP_I8:
1291 return EmitAtomicBinary(MI, BB, 1, 0);
1292 case ARM64::ATOMIC_SWAP_I16:
1293 return EmitAtomicBinary(MI, BB, 2, 0);
1294 case ARM64::ATOMIC_SWAP_I32:
1295 return EmitAtomicBinary(MI, BB, 4, 0);
1296 case ARM64::ATOMIC_SWAP_I64:
1297 return EmitAtomicBinary(MI, BB, 8, 0);
1298 case ARM64::ATOMIC_SWAP_I128:
1299 return EmitAtomicBinary128(MI, BB, 0, 0);
1301 case ARM64::ATOMIC_CMP_SWAP_I8:
1302 return EmitAtomicCmpSwap(MI, BB, 1);
1303 case ARM64::ATOMIC_CMP_SWAP_I16:
1304 return EmitAtomicCmpSwap(MI, BB, 2);
1305 case ARM64::ATOMIC_CMP_SWAP_I32:
1306 return EmitAtomicCmpSwap(MI, BB, 4);
1307 case ARM64::ATOMIC_CMP_SWAP_I64:
1308 return EmitAtomicCmpSwap(MI, BB, 8);
1309 case ARM64::ATOMIC_CMP_SWAP_I128:
1310 return EmitAtomicCmpSwap128(MI, BB);
1312 case ARM64::F128CSEL:
1313 return EmitF128CSEL(MI, BB);
1315 case TargetOpcode::STACKMAP:
1316 case TargetOpcode::PATCHPOINT:
1317 return emitPatchPoint(MI, BB);
1319 llvm_unreachable("Unexpected instruction for custom inserter!");
1322 //===----------------------------------------------------------------------===//
1323 // ARM64 Lowering private implementation.
1324 //===----------------------------------------------------------------------===//
1326 //===----------------------------------------------------------------------===//
1328 //===----------------------------------------------------------------------===//
1330 /// changeIntCCToARM64CC - Convert a DAG integer condition code to an ARM64 CC
1331 static ARM64CC::CondCode changeIntCCToARM64CC(ISD::CondCode CC) {
1334 llvm_unreachable("Unknown condition code!");
1358 /// changeFPCCToARM64CC - Convert a DAG fp condition code to an ARM64 CC.
1359 static void changeFPCCToARM64CC(ISD::CondCode CC, ARM64CC::CondCode &CondCode,
1360 ARM64CC::CondCode &CondCode2) {
1361 CondCode2 = ARM64CC::AL;
1364 llvm_unreachable("Unknown FP condition!");
1367 CondCode = ARM64CC::EQ;
1371 CondCode = ARM64CC::GT;
1375 CondCode = ARM64CC::GE;
1378 CondCode = ARM64CC::MI;
1381 CondCode = ARM64CC::LS;
1384 CondCode = ARM64CC::MI;
1385 CondCode2 = ARM64CC::GT;
1388 CondCode = ARM64CC::VC;
1391 CondCode = ARM64CC::VS;
1394 CondCode = ARM64CC::EQ;
1395 CondCode2 = ARM64CC::VS;
1398 CondCode = ARM64CC::HI;
1401 CondCode = ARM64CC::PL;
1405 CondCode = ARM64CC::LT;
1409 CondCode = ARM64CC::LE;
1413 CondCode = ARM64CC::NE;
1418 static bool isLegalArithImmed(uint64_t C) {
1419 // Matches ARM64DAGToDAGISel::SelectArithImmed().
1420 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
1423 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1424 SDLoc dl, SelectionDAG &DAG) {
1425 EVT VT = LHS.getValueType();
1427 if (VT.isFloatingPoint())
1428 return DAG.getNode(ARM64ISD::FCMP, dl, VT, LHS, RHS);
1430 // The CMP instruction is just an alias for SUBS, and representing it as
1431 // SUBS means that it's possible to get CSE with subtract operations.
1432 // A later phase can perform the optimization of setting the destination
1433 // register to WZR/XZR if it ends up being unused.
1435 // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on the
1436 // grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags can be
1437 // set differently by this operation. It comes down to whether "SInt(~op2)+1
1438 // == SInt(~op2+1)" (and the same for UInt). If they are then everything is
1439 // fine. If not then the optimization is wrong. Thus general comparisons are
1440 // only valid if op2 != 0.
1442 // So, finally, the only LLVM-native comparisons that don't mention C and V
1443 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in the
1444 // absence of information about op2.
1445 unsigned Opcode = ARM64ISD::SUBS;
1446 if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
1447 cast<ConstantSDNode>(RHS.getOperand(0))->getZExtValue() == 0 &&
1448 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1449 Opcode = ARM64ISD::ADDS;
1450 RHS = RHS.getOperand(1);
1453 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS)
1457 static SDValue getARM64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1458 SDValue &ARM64cc, SelectionDAG &DAG, SDLoc dl) {
1459 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1460 EVT VT = RHS.getValueType();
1461 uint64_t C = RHSC->getZExtValue();
1462 if (!isLegalArithImmed(C)) {
1463 // Constant does not fit, try adjusting it by one?
1469 if ((VT == MVT::i32 && C != 0x80000000 &&
1470 isLegalArithImmed((uint32_t)(C - 1))) ||
1471 (VT == MVT::i64 && C != 0x80000000ULL &&
1472 isLegalArithImmed(C - 1ULL))) {
1473 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1474 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1475 RHS = DAG.getConstant(C, VT);
1480 if ((VT == MVT::i32 && C != 0 &&
1481 isLegalArithImmed((uint32_t)(C - 1))) ||
1482 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1483 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1484 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1485 RHS = DAG.getConstant(C, VT);
1490 if ((VT == MVT::i32 && C != 0x7fffffff &&
1491 isLegalArithImmed((uint32_t)(C + 1))) ||
1492 (VT == MVT::i64 && C != 0x7ffffffffffffffULL &&
1493 isLegalArithImmed(C + 1ULL))) {
1494 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1495 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1496 RHS = DAG.getConstant(C, VT);
1501 if ((VT == MVT::i32 && C != 0xffffffff &&
1502 isLegalArithImmed((uint32_t)(C + 1))) ||
1503 (VT == MVT::i64 && C != 0xfffffffffffffffULL &&
1504 isLegalArithImmed(C + 1ULL))) {
1505 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1506 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1507 RHS = DAG.getConstant(C, VT);
1514 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
1515 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
1516 ARM64cc = DAG.getConstant(ARM64CC, MVT::i32);
1520 static std::pair<SDValue, SDValue>
1521 getARM64XALUOOp(ARM64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1522 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1523 "Unsupported value type");
1524 SDValue Value, Overflow;
1526 SDValue LHS = Op.getOperand(0);
1527 SDValue RHS = Op.getOperand(1);
1529 switch (Op.getOpcode()) {
1531 llvm_unreachable("Unknown overflow instruction!");
1533 Opc = ARM64ISD::ADDS;
1537 Opc = ARM64ISD::ADDS;
1541 Opc = ARM64ISD::SUBS;
1545 Opc = ARM64ISD::SUBS;
1548 // Multiply needs a little bit extra work.
1552 bool IsSigned = (Op.getOpcode() == ISD::SMULO) ? true : false;
1553 if (Op.getValueType() == MVT::i32) {
1554 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1555 // For a 32 bit multiply with overflow check we want the instruction
1556 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1557 // need to generate the following pattern:
1558 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1559 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1560 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1561 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1562 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1563 DAG.getConstant(0, MVT::i64));
1564 // On ARM64 the upper 32 bits are always zero extended for a 32 bit
1565 // operation. We need to clear out the upper 32 bits, because we used a
1566 // widening multiply that wrote all 64 bits. In the end this should be a
1568 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1570 // The signed overflow check requires more than just a simple check for
1571 // any bit set in the upper 32 bits of the result. These bits could be
1572 // just the sign bits of a negative number. To perform the overflow
1573 // check we have to arithmetic shift right the 32nd bit of the result by
1574 // 31 bits. Then we compare the result to the upper 32 bits.
1575 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1576 DAG.getConstant(32, MVT::i64));
1577 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1578 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1579 DAG.getConstant(31, MVT::i64));
1580 // It is important that LowerBits is last, otherwise the arithmetic
1581 // shift will not be folded into the compare (SUBS).
1582 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1583 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1586 // The overflow check for unsigned multiply is easy. We only need to
1587 // check if any of the upper 32 bits are set. This can be done with a
1588 // CMP (shifted register). For that we need to generate the following
1590 // (i64 ARM64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1591 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1592 DAG.getConstant(32, MVT::i64));
1593 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1595 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1596 UpperBits).getValue(1);
1600 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1601 // For the 64 bit multiply
1602 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1604 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1605 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1606 DAG.getConstant(63, MVT::i64));
1607 // It is important that LowerBits is last, otherwise the arithmetic
1608 // shift will not be folded into the compare (SUBS).
1609 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1610 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1613 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1614 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1616 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1617 UpperBits).getValue(1);
1624 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1626 // Emit the ARM64 operation with overflow check.
1627 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1628 Overflow = Value.getValue(1);
1630 return std::make_pair(Value, Overflow);
1633 SDValue ARM64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1634 RTLIB::Libcall Call) const {
1635 SmallVector<SDValue, 2> Ops;
1636 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1637 Ops.push_back(Op.getOperand(i));
1639 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1643 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1644 SDValue Sel = Op.getOperand(0);
1645 SDValue Other = Op.getOperand(1);
1647 // If neither operand is a SELECT_CC, give up.
1648 if (Sel.getOpcode() != ISD::SELECT_CC)
1649 std::swap(Sel, Other);
1650 if (Sel.getOpcode() != ISD::SELECT_CC)
1653 // The folding we want to perform is:
1654 // (xor x, (select_cc a, b, cc, 0, -1) )
1656 // (csel x, (xor x, -1), cc ...)
1658 // The latter will get matched to a CSINV instruction.
1660 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1661 SDValue LHS = Sel.getOperand(0);
1662 SDValue RHS = Sel.getOperand(1);
1663 SDValue TVal = Sel.getOperand(2);
1664 SDValue FVal = Sel.getOperand(3);
1667 // FIXME: This could be generalized to non-integer comparisons.
1668 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1671 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1672 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1674 // The the values aren't constants, this isn't the pattern we're looking for.
1675 if (!CFVal || !CTVal)
1678 // We can commute the SELECT_CC by inverting the condition. This
1679 // might be needed to make this fit into a CSINV pattern.
1680 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1681 std::swap(TVal, FVal);
1682 std::swap(CTVal, CFVal);
1683 CC = ISD::getSetCCInverse(CC, true);
1686 // If the constants line up, perform the transform!
1687 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1689 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1692 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1693 DAG.getConstant(-1ULL, Other.getValueType()));
1695 return DAG.getNode(ARM64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1702 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1703 EVT VT = Op.getValueType();
1705 // Let legalize expand this if it isn't a legal type yet.
1706 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1709 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1712 bool ExtraOp = false;
1713 switch (Op.getOpcode()) {
1715 assert(0 && "Invalid code");
1717 Opc = ARM64ISD::ADDS;
1720 Opc = ARM64ISD::SUBS;
1723 Opc = ARM64ISD::ADCS;
1727 Opc = ARM64ISD::SBCS;
1733 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1734 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1738 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1739 // Let legalize expand this if it isn't a legal type yet.
1740 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1743 ARM64CC::CondCode CC;
1744 // The actual operation that sets the overflow or carry flag.
1745 SDValue Value, Overflow;
1746 std::tie(Value, Overflow) = getARM64XALUOOp(CC, Op, DAG);
1748 // We use 0 and 1 as false and true values.
1749 SDValue TVal = DAG.getConstant(1, MVT::i32);
1750 SDValue FVal = DAG.getConstant(0, MVT::i32);
1752 // We use an inverted condition, because the conditional select is inverted
1753 // too. This will allow it to be selected to a single instruction:
1754 // CSINC Wd, WZR, WZR, invert(cond).
1755 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), MVT::i32);
1756 Overflow = DAG.getNode(ARM64ISD::CSEL, SDLoc(Op), MVT::i32, FVal, TVal, CCVal,
1759 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1760 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
1763 // Prefetch operands are:
1764 // 1: Address to prefetch
1766 // 3: int locality (0 = no locality ... 3 = extreme locality)
1767 // 4: bool isDataCache
1768 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1770 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1771 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1772 // The data thing is not used.
1773 // unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1775 bool IsStream = !Locality;
1776 // When the locality number is set
1778 // The front-end should have filtered out the out-of-range values
1779 assert(Locality <= 3 && "Prefetch locality out-of-range");
1780 // The locality degree is the opposite of the cache speed.
1781 // Put the number the other way around.
1782 // The encoding starts at 0 for level 1
1783 Locality = 3 - Locality;
1786 // built the mask value encoding the expected behavior.
1787 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1788 (Locality << 1) | // Cache level bits
1789 (unsigned)IsStream; // Stream bit
1790 return DAG.getNode(ARM64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1791 DAG.getConstant(PrfOp, MVT::i32), Op.getOperand(1));
1794 SDValue ARM64TargetLowering::LowerFP_EXTEND(SDValue Op,
1795 SelectionDAG &DAG) const {
1796 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1799 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1801 return LowerF128Call(Op, DAG, LC);
1804 SDValue ARM64TargetLowering::LowerFP_ROUND(SDValue Op,
1805 SelectionDAG &DAG) const {
1806 if (Op.getOperand(0).getValueType() != MVT::f128) {
1807 // It's legal except when f128 is involved
1812 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1814 // FP_ROUND node has a second operand indicating whether it is known to be
1815 // precise. That doesn't take part in the LibCall so we can't directly use
1817 SDValue SrcVal = Op.getOperand(0);
1818 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1819 /*isSigned*/ false, SDLoc(Op)).first;
1822 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1823 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1824 // Any additional optimization in this function should be recorded
1825 // in the cost tables.
1826 EVT InVT = Op.getOperand(0).getValueType();
1827 EVT VT = Op.getValueType();
1829 // FP_TO_XINT conversion from the same type are legal.
1830 if (VT.getSizeInBits() == InVT.getSizeInBits())
1833 if (InVT == MVT::v2f64) {
1835 SDValue Cv = DAG.getNode(Op.getOpcode(), dl, MVT::v2i64, Op.getOperand(0));
1836 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1839 // Type changing conversions are illegal.
1843 SDValue ARM64TargetLowering::LowerFP_TO_INT(SDValue Op,
1844 SelectionDAG &DAG) const {
1845 if (Op.getOperand(0).getValueType().isVector())
1846 return LowerVectorFP_TO_INT(Op, DAG);
1848 if (Op.getOperand(0).getValueType() != MVT::f128) {
1849 // It's legal except when f128 is involved
1854 if (Op.getOpcode() == ISD::FP_TO_SINT)
1855 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1857 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1859 SmallVector<SDValue, 2> Ops;
1860 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1861 Ops.push_back(Op.getOperand(i));
1863 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1867 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1868 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1869 // Any additional optimization in this function should be recorded
1870 // in the cost tables.
1871 EVT VT = Op.getValueType();
1873 SDValue In = Op.getOperand(0);
1874 EVT InVT = In.getValueType();
1876 // v2i32 to v2f32 is legal.
1877 if (VT == MVT::v2f32 && InVT == MVT::v2i32)
1880 // This function only handles v2f64 outputs.
1881 if (VT == MVT::v2f64) {
1882 // Extend the input argument to a v2i64 that we can feed into the
1883 // floating point conversion. Zero or sign extend based on whether
1884 // we're doing a signed or unsigned float conversion.
1886 Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
1887 assert(Op.getNumOperands() == 1 && "FP conversions take one argument");
1888 SDValue Promoted = DAG.getNode(Opc, dl, MVT::v2i64, Op.getOperand(0));
1889 return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), Promoted);
1892 // Scalarize v2i64 to v2f32 conversions.
1893 std::vector<SDValue> BuildVectorOps;
1894 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
1895 SDValue Sclr = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, In,
1896 DAG.getConstant(i, MVT::i64));
1897 Sclr = DAG.getNode(Op->getOpcode(), dl, MVT::f32, Sclr);
1898 BuildVectorOps.push_back(Sclr);
1901 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &BuildVectorOps[0],
1902 BuildVectorOps.size());
1905 SDValue ARM64TargetLowering::LowerINT_TO_FP(SDValue Op,
1906 SelectionDAG &DAG) const {
1907 if (Op.getValueType().isVector())
1908 return LowerVectorINT_TO_FP(Op, DAG);
1910 // i128 conversions are libcalls.
1911 if (Op.getOperand(0).getValueType() == MVT::i128)
1914 // Other conversions are legal, unless it's to the completely software-based
1916 if (Op.getValueType() != MVT::f128)
1920 if (Op.getOpcode() == ISD::SINT_TO_FP)
1921 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1923 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1925 return LowerF128Call(Op, DAG, LC);
1928 SDValue ARM64TargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
1929 // For iOS, we want to call an alternative entry point: __sincos_stret,
1930 // which returns the values in two S / D registers.
1932 SDValue Arg = Op.getOperand(0);
1933 EVT ArgVT = Arg.getValueType();
1934 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1941 Entry.isSExt = false;
1942 Entry.isZExt = false;
1943 Args.push_back(Entry);
1945 const char *LibcallName =
1946 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1947 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
1949 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
1950 TargetLowering::CallLoweringInfo CLI(
1951 DAG.getEntryNode(), RetTy, false, false, false, false, 0,
1952 CallingConv::Fast, /*isTaillCall=*/false,
1953 /*doesNotRet=*/false, /*isReturnValueUsed*/ true, Callee, Args, DAG, dl);
1954 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1955 return CallResult.first;
1958 SDValue ARM64TargetLowering::LowerOperation(SDValue Op,
1959 SelectionDAG &DAG) const {
1960 switch (Op.getOpcode()) {
1962 llvm_unreachable("unimplemented operand");
1964 case ISD::GlobalAddress:
1965 return LowerGlobalAddress(Op, DAG);
1966 case ISD::GlobalTLSAddress:
1967 return LowerGlobalTLSAddress(Op, DAG);
1969 return LowerSETCC(Op, DAG);
1971 return LowerBR_CC(Op, DAG);
1973 return LowerSELECT(Op, DAG);
1974 case ISD::SELECT_CC:
1975 return LowerSELECT_CC(Op, DAG);
1976 case ISD::JumpTable:
1977 return LowerJumpTable(Op, DAG);
1978 case ISD::ConstantPool:
1979 return LowerConstantPool(Op, DAG);
1980 case ISD::BlockAddress:
1981 return LowerBlockAddress(Op, DAG);
1983 return LowerVASTART(Op, DAG);
1985 return LowerVACOPY(Op, DAG);
1987 return LowerVAARG(Op, DAG);
1992 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
1999 return LowerXALUO(Op, DAG);
2001 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
2003 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
2005 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
2007 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
2009 return LowerFP_ROUND(Op, DAG);
2010 case ISD::FP_EXTEND:
2011 return LowerFP_EXTEND(Op, DAG);
2012 case ISD::FRAMEADDR:
2013 return LowerFRAMEADDR(Op, DAG);
2014 case ISD::RETURNADDR:
2015 return LowerRETURNADDR(Op, DAG);
2016 case ISD::INSERT_VECTOR_ELT:
2017 return LowerINSERT_VECTOR_ELT(Op, DAG);
2018 case ISD::EXTRACT_VECTOR_ELT:
2019 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2020 case ISD::BUILD_VECTOR:
2021 return LowerBUILD_VECTOR(Op, DAG);
2022 case ISD::VECTOR_SHUFFLE:
2023 return LowerVECTOR_SHUFFLE(Op, DAG);
2024 case ISD::EXTRACT_SUBVECTOR:
2025 return LowerEXTRACT_SUBVECTOR(Op, DAG);
2029 return LowerVectorSRA_SRL_SHL(Op, DAG);
2030 case ISD::SHL_PARTS:
2031 return LowerShiftLeftParts(Op, DAG);
2032 case ISD::SRL_PARTS:
2033 case ISD::SRA_PARTS:
2034 return LowerShiftRightParts(Op, DAG);
2036 return LowerCTPOP(Op, DAG);
2037 case ISD::FCOPYSIGN:
2038 return LowerFCOPYSIGN(Op, DAG);
2040 return LowerVectorAND(Op, DAG);
2042 return LowerVectorOR(Op, DAG);
2044 return LowerXOR(Op, DAG);
2046 return LowerPREFETCH(Op, DAG);
2047 case ISD::SINT_TO_FP:
2048 case ISD::UINT_TO_FP:
2049 return LowerINT_TO_FP(Op, DAG);
2050 case ISD::FP_TO_SINT:
2051 case ISD::FP_TO_UINT:
2052 return LowerFP_TO_INT(Op, DAG);
2054 return LowerFSINCOS(Op, DAG);
2058 /// getFunctionAlignment - Return the Log2 alignment of this function.
2059 unsigned ARM64TargetLowering::getFunctionAlignment(const Function *F) const {
2063 //===----------------------------------------------------------------------===//
2064 // Calling Convention Implementation
2065 //===----------------------------------------------------------------------===//
2067 #include "ARM64GenCallingConv.inc"
2069 /// Selects the correct CCAssignFn for a the given CallingConvention
2071 CCAssignFn *ARM64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
2072 bool IsVarArg) const {
2075 llvm_unreachable("Unsupported calling convention.");
2076 case CallingConv::WebKit_JS:
2077 return CC_ARM64_WebKit_JS;
2078 case CallingConv::C:
2079 case CallingConv::Fast:
2080 if (!Subtarget->isTargetDarwin())
2081 return CC_ARM64_AAPCS;
2082 return IsVarArg ? CC_ARM64_DarwinPCS_VarArg : CC_ARM64_DarwinPCS;
2086 SDValue ARM64TargetLowering::LowerFormalArguments(
2087 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2088 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2089 SmallVectorImpl<SDValue> &InVals) const {
2090 MachineFunction &MF = DAG.getMachineFunction();
2091 MachineFrameInfo *MFI = MF.getFrameInfo();
2093 // Assign locations to all of the incoming arguments.
2094 SmallVector<CCValAssign, 16> ArgLocs;
2095 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2096 getTargetMachine(), ArgLocs, *DAG.getContext());
2098 // At this point, Ins[].VT may already be promoted to i32. To correctly
2099 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2100 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
2101 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
2102 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
2104 unsigned NumArgs = Ins.size();
2105 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2106 unsigned CurArgIdx = 0;
2107 for (unsigned i = 0; i != NumArgs; ++i) {
2108 MVT ValVT = Ins[i].VT;
2109 std::advance(CurOrigArg, Ins[i].OrigArgIndex - CurArgIdx);
2110 CurArgIdx = Ins[i].OrigArgIndex;
2112 // Get type of the original argument.
2113 EVT ActualVT = getValueType(CurOrigArg->getType(), /*AllowUnknown*/ true);
2114 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
2115 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2117 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2119 else if (ActualMVT == MVT::i16)
2122 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2124 AssignFn(i, ValVT, LocVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
2125 assert(!Res && "Call operand has unhandled type");
2129 SmallVector<SDValue, 16> ArgValues;
2130 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2131 CCValAssign &VA = ArgLocs[i];
2133 // Arguments stored in registers.
2134 if (VA.isRegLoc()) {
2135 EVT RegVT = VA.getLocVT();
2138 const TargetRegisterClass *RC;
2140 if (RegVT == MVT::i32)
2141 RC = &ARM64::GPR32RegClass;
2142 else if (RegVT == MVT::i64)
2143 RC = &ARM64::GPR64RegClass;
2144 else if (RegVT == MVT::f32)
2145 RC = &ARM64::FPR32RegClass;
2146 else if (RegVT == MVT::f64 || RegVT == MVT::v1i64 ||
2147 RegVT == MVT::v1f64 || RegVT == MVT::v2i32 ||
2148 RegVT == MVT::v4i16 || RegVT == MVT::v8i8)
2149 RC = &ARM64::FPR64RegClass;
2150 else if (RegVT == MVT::v2i64 || RegVT == MVT::v4i32 ||
2151 RegVT == MVT::v8i16 || RegVT == MVT::v16i8)
2152 RC = &ARM64::FPR128RegClass;
2154 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2156 // Transform the arguments in physical registers into virtual ones.
2157 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2158 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2160 // If this is an 8, 16 or 32-bit value, it is really passed promoted
2161 // to 64 bits. Insert an assert[sz]ext to capture this, then
2162 // truncate to the right size.
2163 switch (VA.getLocInfo()) {
2165 llvm_unreachable("Unknown loc info!");
2166 case CCValAssign::Full:
2168 case CCValAssign::BCvt:
2169 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
2171 case CCValAssign::SExt:
2172 ArgValue = DAG.getNode(ISD::AssertSext, DL, RegVT, ArgValue,
2173 DAG.getValueType(VA.getValVT()));
2174 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
2176 case CCValAssign::ZExt:
2177 ArgValue = DAG.getNode(ISD::AssertZext, DL, RegVT, ArgValue,
2178 DAG.getValueType(VA.getValVT()));
2179 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
2183 InVals.push_back(ArgValue);
2185 } else { // VA.isRegLoc()
2186 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
2187 unsigned ArgOffset = VA.getLocMemOffset();
2188 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
2189 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset, true);
2191 // Create load nodes to retrieve arguments from the stack.
2192 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2193 InVals.push_back(DAG.getLoad(VA.getValVT(), DL, Chain, FIN,
2194 MachinePointerInfo::getFixedStack(FI), false,
2201 if (!Subtarget->isTargetDarwin()) {
2202 // The AAPCS variadic function ABI is identical to the non-variadic
2203 // one. As a result there may be more arguments in registers and we should
2204 // save them for future reference.
2205 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
2208 ARM64FunctionInfo *AFI = MF.getInfo<ARM64FunctionInfo>();
2209 // This will point to the next argument passed via stack.
2210 unsigned StackOffset = CCInfo.getNextStackOffset();
2211 // We currently pass all varargs at 8-byte alignment.
2212 StackOffset = ((StackOffset + 7) & ~7);
2213 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
2219 void ARM64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
2220 SelectionDAG &DAG, SDLoc DL,
2221 SDValue &Chain) const {
2222 MachineFunction &MF = DAG.getMachineFunction();
2223 MachineFrameInfo *MFI = MF.getFrameInfo();
2224 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
2226 SmallVector<SDValue, 8> MemOps;
2228 static const MCPhysReg GPRArgRegs[] = { ARM64::X0, ARM64::X1, ARM64::X2,
2229 ARM64::X3, ARM64::X4, ARM64::X5,
2230 ARM64::X6, ARM64::X7 };
2231 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
2232 unsigned FirstVariadicGPR =
2233 CCInfo.getFirstUnallocated(GPRArgRegs, NumGPRArgRegs);
2235 static const MCPhysReg FPRArgRegs[] = { ARM64::Q0, ARM64::Q1, ARM64::Q2,
2236 ARM64::Q3, ARM64::Q4, ARM64::Q5,
2237 ARM64::Q6, ARM64::Q7 };
2238 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
2239 unsigned FirstVariadicFPR =
2240 CCInfo.getFirstUnallocated(FPRArgRegs, NumFPRArgRegs);
2242 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
2244 if (GPRSaveSize != 0) {
2245 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
2247 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
2249 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
2250 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &ARM64::GPR64RegClass);
2251 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
2253 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2254 MachinePointerInfo::getStack(i * 8), false, false, 0);
2255 MemOps.push_back(Store);
2256 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2257 DAG.getConstant(8, getPointerTy()));
2261 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
2263 if (FPRSaveSize != 0) {
2264 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
2266 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
2268 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
2269 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &ARM64::FPR128RegClass);
2270 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::v2i64);
2272 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2273 MachinePointerInfo::getStack(i * 16), false, false, 0);
2274 MemOps.push_back(Store);
2275 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2276 DAG.getConstant(16, getPointerTy()));
2280 FuncInfo->setVarArgsGPRIndex(GPRIdx);
2281 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
2282 FuncInfo->setVarArgsFPRIndex(FPRIdx);
2283 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
2285 if (!MemOps.empty()) {
2286 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
2291 /// LowerCallResult - Lower the result values of a call into the
2292 /// appropriate copies out of appropriate physical registers.
2293 SDValue ARM64TargetLowering::LowerCallResult(
2294 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
2295 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2296 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
2297 SDValue ThisVal) const {
2298 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2299 : RetCC_ARM64_AAPCS;
2300 // Assign locations to each value returned by this call.
2301 SmallVector<CCValAssign, 16> RVLocs;
2302 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2303 getTargetMachine(), RVLocs, *DAG.getContext());
2304 CCInfo.AnalyzeCallResult(Ins, RetCC);
2306 // Copy all of the result registers out of their specified physreg.
2307 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2308 CCValAssign VA = RVLocs[i];
2310 // Pass 'this' value directly from the argument to return value, to avoid
2311 // reg unit interference
2312 if (i == 0 && isThisReturn) {
2313 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
2314 "unexpected return calling convention register assignment");
2315 InVals.push_back(ThisVal);
2320 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
2321 Chain = Val.getValue(1);
2322 InFlag = Val.getValue(2);
2324 switch (VA.getLocInfo()) {
2326 llvm_unreachable("Unknown loc info!");
2327 case CCValAssign::Full:
2329 case CCValAssign::BCvt:
2330 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2334 InVals.push_back(Val);
2340 bool ARM64TargetLowering::isEligibleForTailCallOptimization(
2341 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
2342 bool isCalleeStructRet, bool isCallerStructRet,
2343 const SmallVectorImpl<ISD::OutputArg> &Outs,
2344 const SmallVectorImpl<SDValue> &OutVals,
2345 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
2346 // Look for obvious safe cases to perform tail call optimization that do not
2347 // require ABI changes. This is what gcc calls sibcall.
2349 // Do not sibcall optimize vararg calls unless the call site is not passing
2351 if (isVarArg && !Outs.empty())
2354 // Also avoid sibcall optimization if either caller or callee uses struct
2355 // return semantics.
2356 if (isCalleeStructRet || isCallerStructRet)
2359 // Note that currently ARM64 "C" calling convention and "Fast" calling
2360 // convention are compatible. If/when that ever changes, we'll need to
2361 // add checks here to make sure any interactions are OK.
2363 // If the callee takes no arguments then go on to check the results of the
2365 if (!Outs.empty()) {
2366 // Check if stack adjustment is needed. For now, do not do this if any
2367 // argument is passed on the stack.
2368 SmallVector<CCValAssign, 16> ArgLocs;
2369 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2370 getTargetMachine(), ArgLocs, *DAG.getContext());
2371 CCAssignFn *AssignFn = CCAssignFnForCall(CalleeCC, /*IsVarArg=*/false);
2372 CCInfo.AnalyzeCallOperands(Outs, AssignFn);
2373 if (CCInfo.getNextStackOffset()) {
2374 // Check if the arguments are already laid out in the right way as
2375 // the caller's fixed stack objects.
2376 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2377 ++i, ++realArgIdx) {
2378 CCValAssign &VA = ArgLocs[i];
2379 if (VA.getLocInfo() == CCValAssign::Indirect)
2381 if (VA.needsCustom()) {
2382 // Just don't handle anything that needs custom adjustments for now.
2383 // If need be, we can revisit later, but we shouldn't ever end up
2386 } else if (!VA.isRegLoc()) {
2387 // Likewise, don't try to handle stack based arguments for the
2397 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
2398 /// and add input and output parameter nodes.
2399 SDValue ARM64TargetLowering::LowerCall(CallLoweringInfo &CLI,
2400 SmallVectorImpl<SDValue> &InVals) const {
2401 SelectionDAG &DAG = CLI.DAG;
2403 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2404 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2405 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2406 SDValue Chain = CLI.Chain;
2407 SDValue Callee = CLI.Callee;
2408 bool &IsTailCall = CLI.IsTailCall;
2409 CallingConv::ID CallConv = CLI.CallConv;
2410 bool IsVarArg = CLI.IsVarArg;
2412 MachineFunction &MF = DAG.getMachineFunction();
2413 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2414 bool IsThisReturn = false;
2416 // If tail calls are explicitly disabled, make sure not to use them.
2417 if (!EnableARM64TailCalls)
2421 // Check if it's really possible to do a tail call.
2422 IsTailCall = isEligibleForTailCallOptimization(
2423 Callee, CallConv, IsVarArg, IsStructRet,
2424 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2425 // We don't support GuaranteedTailCallOpt, only automatically
2426 // detected sibcalls.
2427 // FIXME: Re-evaluate. Is this true? Should it be true?
2432 // Analyze operands of the call, assigning locations to each operand.
2433 SmallVector<CCValAssign, 16> ArgLocs;
2434 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2435 getTargetMachine(), ArgLocs, *DAG.getContext());
2438 // Handle fixed and variable vector arguments differently.
2439 // Variable vector arguments always go into memory.
2440 unsigned NumArgs = Outs.size();
2442 for (unsigned i = 0; i != NumArgs; ++i) {
2443 MVT ArgVT = Outs[i].VT;
2444 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2445 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2446 /*IsVarArg=*/ !Outs[i].IsFixed);
2447 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2448 assert(!Res && "Call operand has unhandled type");
2452 // At this point, Outs[].VT may already be promoted to i32. To correctly
2453 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2454 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
2455 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2456 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2458 unsigned NumArgs = Outs.size();
2459 for (unsigned i = 0; i != NumArgs; ++i) {
2460 MVT ValVT = Outs[i].VT;
2461 // Get type of the original argument.
2462 EVT ActualVT = getValueType(CLI.Args[Outs[i].OrigArgIndex].Ty,
2463 /*AllowUnknown*/ true);
2464 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2465 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2466 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2468 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2470 else if (ActualMVT == MVT::i16)
2473 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2474 bool Res = AssignFn(i, ValVT, LocVT, CCValAssign::Full, ArgFlags, CCInfo);
2475 assert(!Res && "Call operand has unhandled type");
2480 // Get a count of how many bytes are to be pushed on the stack.
2481 unsigned NumBytes = CCInfo.getNextStackOffset();
2483 // Adjust the stack pointer for the new arguments...
2484 // These operations are automatically eliminated by the prolog/epilog pass
2487 DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true), DL);
2489 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, ARM64::SP, getPointerTy());
2491 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2492 SmallVector<SDValue, 8> MemOpChains;
2494 // Walk the register/memloc assignments, inserting copies/loads.
2495 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2496 ++i, ++realArgIdx) {
2497 CCValAssign &VA = ArgLocs[i];
2498 SDValue Arg = OutVals[realArgIdx];
2499 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2501 // Promote the value if needed.
2502 switch (VA.getLocInfo()) {
2504 llvm_unreachable("Unknown loc info!");
2505 case CCValAssign::Full:
2507 case CCValAssign::SExt:
2508 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
2510 case CCValAssign::ZExt:
2511 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2513 case CCValAssign::AExt:
2514 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
2516 case CCValAssign::BCvt:
2517 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2519 case CCValAssign::FPExt:
2520 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
2524 if (VA.isRegLoc()) {
2525 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
2526 assert(VA.getLocVT() == MVT::i64 &&
2527 "unexpected calling convention register assignment");
2528 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
2529 "unexpected use of 'returned'");
2530 IsThisReturn = true;
2532 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2534 assert(VA.isMemLoc());
2535 // There's no reason we can't support stack args w/ tailcall, but
2536 // we currently don't, so assert if we see one.
2537 assert(!IsTailCall && "stack argument with tail call!?");
2538 unsigned LocMemOffset = VA.getLocMemOffset();
2539 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2540 PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2542 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
2543 // promoted to a legal register type i32, we should truncate Arg back to
2545 if (Arg.getValueType().isSimple() &&
2546 Arg.getValueType().getSimpleVT() == MVT::i32 &&
2547 (VA.getLocVT() == MVT::i1 || VA.getLocVT() == MVT::i8 ||
2548 VA.getLocVT() == MVT::i16))
2549 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getLocVT(), Arg);
2551 SDValue Store = DAG.getStore(Chain, DL, Arg, PtrOff,
2552 MachinePointerInfo::getStack(LocMemOffset),
2554 MemOpChains.push_back(Store);
2558 if (!MemOpChains.empty())
2559 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOpChains[0],
2560 MemOpChains.size());
2562 // Build a sequence of copy-to-reg nodes chained together with token chain
2563 // and flag operands which copy the outgoing args into the appropriate regs.
2565 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2566 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[i].first,
2567 RegsToPass[i].second, InFlag);
2568 InFlag = Chain.getValue(1);
2571 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2572 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2573 // node so that legalize doesn't hack it.
2574 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
2575 Subtarget->isTargetMachO()) {
2576 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2577 const GlobalValue *GV = G->getGlobal();
2578 bool InternalLinkage = GV->hasInternalLinkage();
2579 if (InternalLinkage)
2580 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2582 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0,
2584 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2586 } else if (ExternalSymbolSDNode *S =
2587 dyn_cast<ExternalSymbolSDNode>(Callee)) {
2588 const char *Sym = S->getSymbol();
2590 DAG.getTargetExternalSymbol(Sym, getPointerTy(), ARM64II::MO_GOT);
2591 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2593 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2594 const GlobalValue *GV = G->getGlobal();
2595 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2596 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2597 const char *Sym = S->getSymbol();
2598 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), 0);
2601 std::vector<SDValue> Ops;
2602 Ops.push_back(Chain);
2603 Ops.push_back(Callee);
2605 // Add argument registers to the end of the list so that they are known live
2607 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2608 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2609 RegsToPass[i].second.getValueType()));
2611 // Add a register mask operand representing the call-preserved registers.
2612 const uint32_t *Mask;
2613 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2614 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2616 // For 'this' returns, use the X0-preserving mask if applicable
2617 Mask = ARI->getThisReturnPreservedMask(CallConv);
2619 IsThisReturn = false;
2620 Mask = ARI->getCallPreservedMask(CallConv);
2623 Mask = ARI->getCallPreservedMask(CallConv);
2625 assert(Mask && "Missing call preserved mask for calling convention");
2626 Ops.push_back(DAG.getRegisterMask(Mask));
2628 if (InFlag.getNode())
2629 Ops.push_back(InFlag);
2631 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2633 // If we're doing a tall call, use a TC_RETURN here rather than an
2634 // actual call instruction.
2636 return DAG.getNode(ARM64ISD::TC_RETURN, DL, NodeTys, &Ops[0], Ops.size());
2638 // Returns a chain and a flag for retval copy to use.
2639 Chain = DAG.getNode(ARM64ISD::CALL, DL, NodeTys, &Ops[0], Ops.size());
2640 InFlag = Chain.getValue(1);
2642 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2643 DAG.getIntPtrConstant(0, true), InFlag, DL);
2645 InFlag = Chain.getValue(1);
2647 // Handle result values, copying them out of physregs into vregs that we
2649 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2650 InVals, IsThisReturn,
2651 IsThisReturn ? OutVals[0] : SDValue());
2654 bool ARM64TargetLowering::CanLowerReturn(
2655 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2656 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2657 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2658 : RetCC_ARM64_AAPCS;
2659 SmallVector<CCValAssign, 16> RVLocs;
2660 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2661 return CCInfo.CheckReturn(Outs, RetCC);
2665 ARM64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2667 const SmallVectorImpl<ISD::OutputArg> &Outs,
2668 const SmallVectorImpl<SDValue> &OutVals,
2669 SDLoc DL, SelectionDAG &DAG) const {
2670 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2671 : RetCC_ARM64_AAPCS;
2672 SmallVector<CCValAssign, 16> RVLocs;
2673 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2674 getTargetMachine(), RVLocs, *DAG.getContext());
2675 CCInfo.AnalyzeReturn(Outs, RetCC);
2677 // Copy the result values into the output registers.
2679 SmallVector<SDValue, 4> RetOps(1, Chain);
2680 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
2681 ++i, ++realRVLocIdx) {
2682 CCValAssign &VA = RVLocs[i];
2683 assert(VA.isRegLoc() && "Can only return in registers!");
2684 SDValue Arg = OutVals[realRVLocIdx];
2686 switch (VA.getLocInfo()) {
2688 llvm_unreachable("Unknown loc info!");
2689 case CCValAssign::Full:
2691 case CCValAssign::BCvt:
2692 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2696 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
2697 Flag = Chain.getValue(1);
2698 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2701 RetOps[0] = Chain; // Update chain.
2703 // Add the flag if we have it.
2705 RetOps.push_back(Flag);
2707 return DAG.getNode(ARM64ISD::RET_FLAG, DL, MVT::Other, &RetOps[0],
2711 //===----------------------------------------------------------------------===//
2712 // Other Lowering Code
2713 //===----------------------------------------------------------------------===//
2715 SDValue ARM64TargetLowering::LowerGlobalAddress(SDValue Op,
2716 SelectionDAG &DAG) const {
2717 EVT PtrVT = getPointerTy();
2719 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2720 unsigned char OpFlags =
2721 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
2723 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
2724 "unexpected offset in global node");
2726 // This also catched the large code model case for Darwin.
2727 if ((OpFlags & ARM64II::MO_GOT) != 0) {
2728 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
2729 // FIXME: Once remat is capable of dealing with instructions with register
2730 // operands, expand this into two nodes instead of using a wrapper node.
2731 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
2734 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2735 const unsigned char MO_NC = ARM64II::MO_NC;
2737 ARM64ISD::WrapperLarge, DL, PtrVT,
2738 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G3),
2739 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
2740 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
2741 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
2743 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
2744 // the only correct model on Darwin.
2745 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2746 OpFlags | ARM64II::MO_PAGE);
2747 unsigned char LoFlags = OpFlags | ARM64II::MO_PAGEOFF | ARM64II::MO_NC;
2748 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
2750 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
2751 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
2755 /// \brief Convert a TLS address reference into the correct sequence of loads
2756 /// and calls to compute the variable's address (for Darwin, currently) and
2757 /// return an SDValue containing the final node.
2759 /// Darwin only has one TLS scheme which must be capable of dealing with the
2760 /// fully general situation, in the worst case. This means:
2761 /// + "extern __thread" declaration.
2762 /// + Defined in a possibly unknown dynamic library.
2764 /// The general system is that each __thread variable has a [3 x i64] descriptor
2765 /// which contains information used by the runtime to calculate the address. The
2766 /// only part of this the compiler needs to know about is the first xword, which
2767 /// contains a function pointer that must be called with the address of the
2768 /// entire descriptor in "x0".
2770 /// Since this descriptor may be in a different unit, in general even the
2771 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
2773 /// adrp x0, _var@TLVPPAGE
2774 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
2775 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
2776 /// ; the function pointer
2777 /// blr x1 ; Uses descriptor address in x0
2778 /// ; Address of _var is now in x0.
2780 /// If the address of _var's descriptor *is* known to the linker, then it can
2781 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
2782 /// a slight efficiency gain.
2784 ARM64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
2785 SelectionDAG &DAG) const {
2786 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
2789 MVT PtrVT = getPointerTy();
2790 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2793 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2794 SDValue DescAddr = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TLVPAddr);
2796 // The first entry in the descriptor is a function pointer that we must call
2797 // to obtain the address of the variable.
2798 SDValue Chain = DAG.getEntryNode();
2799 SDValue FuncTLVGet =
2800 DAG.getLoad(MVT::i64, DL, Chain, DescAddr, MachinePointerInfo::getGOT(),
2801 false, true, true, 8);
2802 Chain = FuncTLVGet.getValue(1);
2804 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2805 MFI->setAdjustsStack(true);
2807 // TLS calls preserve all registers except those that absolutely must be
2808 // trashed: X0 (it takes an argument), LR (it's a call) and CPSR (let's not be
2810 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2811 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2812 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2814 // Finally, we can make the call. This is just a degenerate version of a
2815 // normal ARM64 call node: x0 takes the address of the descriptor, and returns
2816 // the address of the variable in this thread.
2817 Chain = DAG.getCopyToReg(Chain, DL, ARM64::X0, DescAddr, SDValue());
2818 Chain = DAG.getNode(ARM64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
2819 Chain, FuncTLVGet, DAG.getRegister(ARM64::X0, MVT::i64),
2820 DAG.getRegisterMask(Mask), Chain.getValue(1));
2821 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Chain.getValue(1));
2824 /// When accessing thread-local variables under either the general-dynamic or
2825 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
2826 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
2827 /// is a function pointer to carry out the resolution. This function takes the
2828 /// address of the descriptor in X0 and returns the TPIDR_EL0 offset in X0. All
2829 /// other registers (except LR, CPSR) are preserved.
2831 /// Thus, the ideal call sequence on AArch64 is:
2833 /// adrp x0, :tlsdesc:thread_var
2834 /// ldr x8, [x0, :tlsdesc_lo12:thread_var]
2835 /// add x0, x0, :tlsdesc_lo12:thread_var
2836 /// .tlsdesccall thread_var
2838 /// (TPIDR_EL0 offset now in x0).
2840 /// The ".tlsdesccall" directive instructs the assembler to insert a particular
2841 /// relocation to help the linker relax this sequence if it turns out to be too
2844 /// FIXME: we currently produce an extra, duplicated, ADRP instruction, but this
2846 SDValue ARM64TargetLowering::LowerELFTLSDescCall(SDValue SymAddr,
2847 SDValue DescAddr, SDLoc DL,
2848 SelectionDAG &DAG) const {
2849 EVT PtrVT = getPointerTy();
2851 // The function we need to call is simply the first entry in the GOT for this
2852 // descriptor, load it in preparation.
2853 SDValue Func = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, SymAddr);
2855 // TLS calls preserve all registers except those that absolutely must be
2856 // trashed: X0 (it takes an argument), LR (it's a call) and CPSR (let's not be
2858 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2859 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2860 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2862 // The function takes only one argument: the address of the descriptor itself
2864 SDValue Glue, Chain;
2865 Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, ARM64::X0, DescAddr, Glue);
2866 Glue = Chain.getValue(1);
2868 // We're now ready to populate the argument list, as with a normal call:
2869 SmallVector<SDValue, 6> Ops;
2870 Ops.push_back(Chain);
2871 Ops.push_back(Func);
2872 Ops.push_back(SymAddr);
2873 Ops.push_back(DAG.getRegister(ARM64::X0, PtrVT));
2874 Ops.push_back(DAG.getRegisterMask(Mask));
2875 Ops.push_back(Glue);
2877 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2878 Chain = DAG.getNode(ARM64ISD::TLSDESC_CALL, DL, NodeTys, &Ops[0], Ops.size());
2879 Glue = Chain.getValue(1);
2881 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Glue);
2884 SDValue ARM64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
2885 SelectionDAG &DAG) const {
2886 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
2887 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
2888 "ELF TLS only supported in small memory model");
2889 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2891 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
2894 EVT PtrVT = getPointerTy();
2896 const GlobalValue *GV = GA->getGlobal();
2898 SDValue ThreadBase = DAG.getNode(ARM64ISD::THREAD_POINTER, DL, PtrVT);
2900 if (Model == TLSModel::LocalExec) {
2901 SDValue HiVar = DAG.getTargetGlobalAddress(
2902 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2903 SDValue LoVar = DAG.getTargetGlobalAddress(
2904 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2906 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2907 DAG.getTargetConstant(16, MVT::i32)),
2909 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, TPOff, LoVar,
2910 DAG.getTargetConstant(0, MVT::i32)),
2912 } else if (Model == TLSModel::InitialExec) {
2913 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2914 TPOff = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TPOff);
2915 } else if (Model == TLSModel::LocalDynamic) {
2916 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
2917 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
2918 // the beginning of the module's TLS region, followed by a DTPREL offset
2921 // These accesses will need deduplicating if there's more than one.
2922 ARM64FunctionInfo *MFI =
2923 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
2924 MFI->incNumLocalDynamicTLSAccesses();
2926 // Accesses used in this sequence go via the TLS descriptor which lives in
2927 // the GOT. Prepare an address we can use to handle this.
2928 SDValue HiDesc = DAG.getTargetExternalSymbol(
2929 "_TLS_MODULE_BASE_", PtrVT, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2930 SDValue LoDesc = DAG.getTargetExternalSymbol(
2931 "_TLS_MODULE_BASE_", PtrVT,
2932 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2934 // First argument to the descriptor call is the address of the descriptor
2936 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2937 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2939 // The call needs a relocation too for linker relaxation. It doesn't make
2940 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2942 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2945 // Now we can calculate the offset from TPIDR_EL0 to this module's
2946 // thread-local area.
2947 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2949 // Now use :dtprel_whatever: operations to calculate this variable's offset
2950 // in its thread-storage area.
2951 SDValue HiVar = DAG.getTargetGlobalAddress(
2952 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2953 SDValue LoVar = DAG.getTargetGlobalAddress(
2954 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2957 SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2958 DAG.getTargetConstant(16, MVT::i32)),
2960 DTPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, DTPOff, LoVar,
2961 DAG.getTargetConstant(0, MVT::i32)),
2964 TPOff = DAG.getNode(ISD::ADD, DL, PtrVT, TPOff, DTPOff);
2965 } else if (Model == TLSModel::GeneralDynamic) {
2966 // Accesses used in this sequence go via the TLS descriptor which lives in
2967 // the GOT. Prepare an address we can use to handle this.
2968 SDValue HiDesc = DAG.getTargetGlobalAddress(
2969 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2970 SDValue LoDesc = DAG.getTargetGlobalAddress(
2972 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2974 // First argument to the descriptor call is the address of the descriptor
2976 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2977 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2979 // The call needs a relocation too for linker relaxation. It doesn't make
2980 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2983 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2985 // Finally we can make a call to calculate the offset from tpidr_el0.
2986 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2988 llvm_unreachable("Unsupported ELF TLS access model");
2990 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
2993 SDValue ARM64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
2994 SelectionDAG &DAG) const {
2995 if (Subtarget->isTargetDarwin())
2996 return LowerDarwinGlobalTLSAddress(Op, DAG);
2997 else if (Subtarget->isTargetELF())
2998 return LowerELFGlobalTLSAddress(Op, DAG);
3000 llvm_unreachable("Unexpected platform trying to use TLS");
3002 SDValue ARM64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3003 SDValue Chain = Op.getOperand(0);
3004 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3005 SDValue LHS = Op.getOperand(2);
3006 SDValue RHS = Op.getOperand(3);
3007 SDValue Dest = Op.getOperand(4);
3010 // Handle f128 first, since lowering it will result in comparing the return
3011 // value of a libcall against zero, which is just what the rest of LowerBR_CC
3012 // is expecting to deal with.
3013 if (LHS.getValueType() == MVT::f128) {
3014 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3016 // If softenSetCCOperands returned a scalar, we need to compare the result
3017 // against zero to select between true and false values.
3018 if (RHS.getNode() == 0) {
3019 RHS = DAG.getConstant(0, LHS.getValueType());
3024 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
3026 unsigned Opc = LHS.getOpcode();
3027 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
3028 cast<ConstantSDNode>(RHS)->isOne() &&
3029 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3030 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3031 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
3032 "Unexpected condition code.");
3033 // Only lower legal XALUO ops.
3034 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
3037 // The actual operation with overflow check.
3038 ARM64CC::CondCode OFCC;
3039 SDValue Value, Overflow;
3040 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, LHS.getValue(0), DAG);
3042 if (CC == ISD::SETNE)
3043 OFCC = getInvertedCondCode(OFCC);
3044 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3046 return DAG.getNode(ARM64ISD::BRCOND, SDLoc(LHS), MVT::Other, Chain, Dest,
3050 if (LHS.getValueType().isInteger()) {
3051 assert((LHS.getValueType() == RHS.getValueType()) &&
3052 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3054 // If the RHS of the comparison is zero, we can potentially fold this
3055 // to a specialized branch.
3056 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
3057 if (RHSC && RHSC->getZExtValue() == 0) {
3058 if (CC == ISD::SETEQ) {
3059 // See if we can use a TBZ to fold in an AND as well.
3060 // TBZ has a smaller branch displacement than CBZ. If the offset is
3061 // out of bounds, a late MI-layer pass rewrites branches.
3062 // 403.gcc is an example that hits this case.
3063 if (LHS.getOpcode() == ISD::AND &&
3064 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3065 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3066 SDValue Test = LHS.getOperand(0);
3067 uint64_t Mask = LHS.getConstantOperandVal(1);
3069 // TBZ only operates on i64's, but the ext should be free.
3070 if (Test.getValueType() == MVT::i32)
3071 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
3073 return DAG.getNode(ARM64ISD::TBZ, dl, MVT::Other, Chain, Test,
3074 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3077 return DAG.getNode(ARM64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
3078 } else if (CC == ISD::SETNE) {
3079 // See if we can use a TBZ to fold in an AND as well.
3080 // TBZ has a smaller branch displacement than CBZ. If the offset is
3081 // out of bounds, a late MI-layer pass rewrites branches.
3082 // 403.gcc is an example that hits this case.
3083 if (LHS.getOpcode() == ISD::AND &&
3084 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3085 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3086 SDValue Test = LHS.getOperand(0);
3087 uint64_t Mask = LHS.getConstantOperandVal(1);
3089 // TBNZ only operates on i64's, but the ext should be free.
3090 if (Test.getValueType() == MVT::i32)
3091 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
3093 return DAG.getNode(ARM64ISD::TBNZ, dl, MVT::Other, Chain, Test,
3094 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3097 return DAG.getNode(ARM64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
3102 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3103 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3107 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3109 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3110 // clean. Some of them require two branches to implement.
3111 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3112 ARM64CC::CondCode CC1, CC2;
3113 changeFPCCToARM64CC(CC, CC1, CC2);
3114 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3116 DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
3117 if (CC2 != ARM64CC::AL) {
3118 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3119 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
3126 SDValue ARM64TargetLowering::LowerFCOPYSIGN(SDValue Op,
3127 SelectionDAG &DAG) const {
3128 EVT VT = Op.getValueType();
3131 SDValue In1 = Op.getOperand(0);
3132 SDValue In2 = Op.getOperand(1);
3133 EVT SrcVT = In2.getValueType();
3135 if (SrcVT == MVT::f32 && VT == MVT::f64)
3136 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
3137 else if (SrcVT == MVT::f64 && VT == MVT::f32)
3138 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0));
3140 // FIXME: Src type is different, bail out for now. Can VT really be a
3147 SDValue EltMask, VecVal1, VecVal2;
3148 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
3151 EltMask = DAG.getConstant(0x80000000ULL, EltVT);
3153 if (!VT.isVector()) {
3154 VecVal1 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
3155 DAG.getUNDEF(VecVT), In1);
3156 VecVal2 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
3157 DAG.getUNDEF(VecVT), In2);
3159 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3160 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3162 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
3166 // We want to materialize a mask with the the high bit set, but the AdvSIMD
3167 // immediate moves cannot materialize that in a single instruction for
3168 // 64-bit elements. Instead, materialize zero and then negate it.
3169 EltMask = DAG.getConstant(0, EltVT);
3171 if (!VT.isVector()) {
3172 VecVal1 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
3173 DAG.getUNDEF(VecVT), In1);
3174 VecVal2 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
3175 DAG.getUNDEF(VecVT), In2);
3177 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3178 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3181 llvm_unreachable("Invalid type for copysign!");
3184 std::vector<SDValue> BuildVectorOps;
3185 for (unsigned i = 0; i < VecVT.getVectorNumElements(); ++i)
3186 BuildVectorOps.push_back(EltMask);
3188 SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT,
3189 &BuildVectorOps[0], BuildVectorOps.size());
3191 // If we couldn't materialize the mask above, then the mask vector will be
3192 // the zero vector, and we need to negate it here.
3193 if (VT == MVT::f64 || VT == MVT::v2f64) {
3194 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
3195 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
3196 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
3200 DAG.getNode(ARM64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
3203 return DAG.getTargetExtractSubreg(ARM64::ssub, DL, VT, Sel);
3204 else if (VT == MVT::f64)
3205 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, VT, Sel);
3207 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
3210 SDValue ARM64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
3211 if (DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
3212 AttributeSet::FunctionIndex, Attribute::NoImplicitFloat))
3215 // While there is no integer popcount instruction, it can
3216 // be more efficiently lowered to the following sequence that uses
3217 // AdvSIMD registers/instructions as long as the copies to/from
3218 // the AdvSIMD registers are cheap.
3219 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
3220 // CNT V0.8B, V0.8B // 8xbyte pop-counts
3221 // ADDV B0, V0.8B // sum 8xbyte pop-counts
3222 // UMOV X0, V0.B[0] // copy byte result back to integer reg
3223 SDValue Val = Op.getOperand(0);
3225 EVT VT = Op.getValueType();
3226 SDValue ZeroVec = DAG.getUNDEF(MVT::v8i8);
3229 if (VT == MVT::i32) {
3230 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Val);
3232 DAG.getTargetInsertSubreg(ARM64::ssub, DL, MVT::v8i8, ZeroVec, VecVal);
3234 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
3237 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, VecVal);
3238 SDValue UaddLV = DAG.getNode(
3239 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
3240 DAG.getConstant(Intrinsic::arm64_neon_uaddlv, MVT::i32), CtPop);
3243 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
3247 SDValue ARM64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3249 if (Op.getValueType().isVector())
3250 return LowerVSETCC(Op, DAG);
3252 SDValue LHS = Op.getOperand(0);
3253 SDValue RHS = Op.getOperand(1);
3254 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3257 // We chose ZeroOrOneBooleanContents, so use zero and one.
3258 EVT VT = Op.getValueType();
3259 SDValue TVal = DAG.getConstant(1, VT);
3260 SDValue FVal = DAG.getConstant(0, VT);
3262 // Handle f128 first, since one possible outcome is a normal integer
3263 // comparison which gets picked up by the next if statement.
3264 if (LHS.getValueType() == MVT::f128) {
3265 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3267 // If softenSetCCOperands returned a scalar, use it.
3268 if (RHS.getNode() == 0) {
3269 assert(LHS.getValueType() == Op.getValueType() &&
3270 "Unexpected setcc expansion!");
3275 if (LHS.getValueType().isInteger()) {
3278 getARM64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
3280 // Note that we inverted the condition above, so we reverse the order of
3281 // the true and false operands here. This will allow the setcc to be
3282 // matched to a single CSINC instruction.
3283 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
3286 // Now we know we're dealing with FP values.
3287 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3289 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3290 // and do the comparison.
3291 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3293 ARM64CC::CondCode CC1, CC2;
3294 changeFPCCToARM64CC(CC, CC1, CC2);
3295 if (CC2 == ARM64CC::AL) {
3296 changeFPCCToARM64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
3297 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3299 // Note that we inverted the condition above, so we reverse the order of
3300 // the true and false operands here. This will allow the setcc to be
3301 // matched to a single CSINC instruction.
3302 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
3304 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3305 // clean. Some of them require two CSELs to implement. As is in this case,
3306 // we emit the first CSEL and then emit a second using the output of the
3307 // first as the RHS. We're effectively OR'ing the two CC's together.
3309 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
3310 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3311 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3313 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3314 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3318 /// A SELECT_CC operation is really some kind of max or min if both values being
3319 /// compared are, in some sense, equal to the results in either case. However,
3320 /// it is permissible to compare f32 values and produce directly extended f64
3323 /// Extending the comparison operands would also be allowed, but is less likely
3324 /// to happen in practice since their use is right here. Note that truncate
3325 /// operations would *not* be semantically equivalent.
3326 static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) {
3330 ConstantFPSDNode *CCmp = dyn_cast<ConstantFPSDNode>(Cmp);
3331 ConstantFPSDNode *CResult = dyn_cast<ConstantFPSDNode>(Result);
3332 if (CCmp && CResult && Cmp.getValueType() == MVT::f32 &&
3333 Result.getValueType() == MVT::f64) {
3335 APFloat CmpVal = CCmp->getValueAPF();
3336 CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy);
3337 return CResult->getValueAPF().bitwiseIsEqual(CmpVal);
3340 return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp;
3343 SDValue ARM64TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3344 SDValue CC = Op->getOperand(0);
3345 SDValue TVal = Op->getOperand(1);
3346 SDValue FVal = Op->getOperand(2);
3349 unsigned Opc = CC.getOpcode();
3350 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
3352 if (CC.getResNo() == 1 &&
3353 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3354 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3355 // Only lower legal XALUO ops.
3356 if (!DAG.getTargetLoweringInfo().isTypeLegal(CC->getValueType(0)))
3359 ARM64CC::CondCode OFCC;
3360 SDValue Value, Overflow;
3361 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, CC.getValue(0), DAG);
3362 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3364 return DAG.getNode(ARM64ISD::CSEL, DL, Op.getValueType(), TVal, FVal, CCVal,
3368 if (CC.getOpcode() == ISD::SETCC)
3369 return DAG.getSelectCC(DL, CC.getOperand(0), CC.getOperand(1), TVal, FVal,
3370 cast<CondCodeSDNode>(CC.getOperand(2))->get());
3372 return DAG.getSelectCC(DL, CC, DAG.getConstant(0, CC.getValueType()), TVal,
3376 SDValue ARM64TargetLowering::LowerSELECT_CC(SDValue Op,
3377 SelectionDAG &DAG) const {
3378 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3379 SDValue LHS = Op.getOperand(0);
3380 SDValue RHS = Op.getOperand(1);
3381 SDValue TVal = Op.getOperand(2);
3382 SDValue FVal = Op.getOperand(3);
3385 // Handle f128 first, because it will result in a comparison of some RTLIB
3386 // call result against zero.
3387 if (LHS.getValueType() == MVT::f128) {
3388 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3390 // If softenSetCCOperands returned a scalar, we need to compare the result
3391 // against zero to select between true and false values.
3392 if (RHS.getNode() == 0) {
3393 RHS = DAG.getConstant(0, LHS.getValueType());
3398 // Handle integers first.
3399 if (LHS.getValueType().isInteger()) {
3400 assert((LHS.getValueType() == RHS.getValueType()) &&
3401 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3403 unsigned Opcode = ARM64ISD::CSEL;
3405 // If both the TVal and the FVal are constants, see if we can swap them in
3406 // order to for a CSINV or CSINC out of them.
3407 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3408 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3410 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3411 std::swap(TVal, FVal);
3412 std::swap(CTVal, CFVal);
3413 CC = ISD::getSetCCInverse(CC, true);
3414 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3415 std::swap(TVal, FVal);
3416 std::swap(CTVal, CFVal);
3417 CC = ISD::getSetCCInverse(CC, true);
3418 } else if (TVal.getOpcode() == ISD::XOR) {
3419 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3420 // with a CSINV rather than a CSEL.
3421 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
3423 if (CVal && CVal->isAllOnesValue()) {
3424 std::swap(TVal, FVal);
3425 std::swap(CTVal, CFVal);
3426 CC = ISD::getSetCCInverse(CC, true);
3428 } else if (TVal.getOpcode() == ISD::SUB) {
3429 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3430 // that we can match with a CSNEG rather than a CSEL.
3431 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
3433 if (CVal && CVal->isNullValue()) {
3434 std::swap(TVal, FVal);
3435 std::swap(CTVal, CFVal);
3436 CC = ISD::getSetCCInverse(CC, true);
3438 } else if (CTVal && CFVal) {
3439 const int64_t TrueVal = CTVal->getSExtValue();
3440 const int64_t FalseVal = CFVal->getSExtValue();
3443 // If both TVal and FVal are constants, see if FVal is the
3444 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3445 // instead of a CSEL in that case.
3446 if (TrueVal == ~FalseVal) {
3447 Opcode = ARM64ISD::CSINV;
3448 } else if (TrueVal == -FalseVal) {
3449 Opcode = ARM64ISD::CSNEG;
3450 } else if (TVal.getValueType() == MVT::i32) {
3451 // If our operands are only 32-bit wide, make sure we use 32-bit
3452 // arithmetic for the check whether we can use CSINC. This ensures that
3453 // the addition in the check will wrap around properly in case there is
3454 // an overflow (which would not be the case if we do the check with
3455 // 64-bit arithmetic).
3456 const uint32_t TrueVal32 = CTVal->getZExtValue();
3457 const uint32_t FalseVal32 = CFVal->getZExtValue();
3459 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3460 Opcode = ARM64ISD::CSINC;
3462 if (TrueVal32 > FalseVal32) {
3466 // 64-bit check whether we can use CSINC.
3467 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3468 Opcode = ARM64ISD::CSINC;
3470 if (TrueVal > FalseVal) {
3475 // Swap TVal and FVal if necessary.
3477 std::swap(TVal, FVal);
3478 std::swap(CTVal, CFVal);
3479 CC = ISD::getSetCCInverse(CC, true);
3482 if (Opcode != ARM64ISD::CSEL) {
3483 // Drop FVal since we can get its value by simply inverting/negating
3490 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3492 EVT VT = Op.getValueType();
3493 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3496 // Now we know we're dealing with FP values.
3497 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3498 assert(LHS.getValueType() == RHS.getValueType());
3499 EVT VT = Op.getValueType();
3501 // Try to match this select into a max/min operation, which have dedicated
3502 // opcode in the instruction set.
3503 // NOTE: This is not correct in the presence of NaNs, so we only enable this
3505 if (getTargetMachine().Options.NoNaNsFPMath) {
3506 if (selectCCOpsAreFMaxCompatible(LHS, FVal) &&
3507 selectCCOpsAreFMaxCompatible(RHS, TVal)) {
3508 CC = ISD::getSetCCSwappedOperands(CC);
3509 std::swap(TVal, FVal);
3512 if (selectCCOpsAreFMaxCompatible(LHS, TVal) &&
3513 selectCCOpsAreFMaxCompatible(RHS, FVal)) {
3523 return DAG.getNode(ARM64ISD::FMAX, dl, VT, TVal, FVal);
3531 return DAG.getNode(ARM64ISD::FMIN, dl, VT, TVal, FVal);
3537 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3538 // and do the comparison.
3539 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3541 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3542 // clean. Some of them require two CSELs to implement.
3543 ARM64CC::CondCode CC1, CC2;
3544 changeFPCCToARM64CC(CC, CC1, CC2);
3545 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3546 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3548 // If we need a second CSEL, emit it, using the output of the first as the
3549 // RHS. We're effectively OR'ing the two CC's together.
3550 if (CC2 != ARM64CC::AL) {
3551 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3552 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3555 // Otherwise, return the output of the first CSEL.
3559 SDValue ARM64TargetLowering::LowerJumpTable(SDValue Op,
3560 SelectionDAG &DAG) const {
3561 // Jump table entries as PC relative offsets. No additional tweaking
3562 // is necessary here. Just get the address of the jump table.
3563 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3564 EVT PtrVT = getPointerTy();
3567 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3568 !Subtarget->isTargetMachO()) {
3569 const unsigned char MO_NC = ARM64II::MO_NC;
3571 ARM64ISD::WrapperLarge, DL, PtrVT,
3572 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G3),
3573 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G2 | MO_NC),
3574 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G1 | MO_NC),
3575 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G0 | MO_NC));
3578 SDValue Hi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_PAGE);
3579 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3580 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3581 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3582 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3585 SDValue ARM64TargetLowering::LowerConstantPool(SDValue Op,
3586 SelectionDAG &DAG) const {
3587 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3588 EVT PtrVT = getPointerTy();
3591 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3592 // Use the GOT for the large code model on iOS.
3593 if (Subtarget->isTargetMachO()) {
3594 SDValue GotAddr = DAG.getTargetConstantPool(
3595 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3597 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
3600 const unsigned char MO_NC = ARM64II::MO_NC;
3602 ARM64ISD::WrapperLarge, DL, PtrVT,
3603 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3604 CP->getOffset(), ARM64II::MO_G3),
3605 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3606 CP->getOffset(), ARM64II::MO_G2 | MO_NC),
3607 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3608 CP->getOffset(), ARM64II::MO_G1 | MO_NC),
3609 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3610 CP->getOffset(), ARM64II::MO_G0 | MO_NC));
3612 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
3613 // ELF, the only valid one on Darwin.
3615 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3616 CP->getOffset(), ARM64II::MO_PAGE);
3617 SDValue Lo = DAG.getTargetConstantPool(
3618 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3619 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3621 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3622 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3626 SDValue ARM64TargetLowering::LowerBlockAddress(SDValue Op,
3627 SelectionDAG &DAG) const {
3628 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
3629 EVT PtrVT = getPointerTy();
3631 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3632 !Subtarget->isTargetMachO()) {
3633 const unsigned char MO_NC = ARM64II::MO_NC;
3635 ARM64ISD::WrapperLarge, DL, PtrVT,
3636 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G3),
3637 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
3638 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
3639 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
3641 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGE);
3642 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGEOFF |
3644 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3645 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3649 SDValue ARM64TargetLowering::LowerDarwin_VASTART(SDValue Op,
3650 SelectionDAG &DAG) const {
3651 ARM64FunctionInfo *FuncInfo =
3652 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
3656 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3657 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3658 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
3659 MachinePointerInfo(SV), false, false, 0);
3662 SDValue ARM64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
3663 SelectionDAG &DAG) const {
3664 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3665 // Standard, section B.3.
3666 MachineFunction &MF = DAG.getMachineFunction();
3667 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
3670 SDValue Chain = Op.getOperand(0);
3671 SDValue VAList = Op.getOperand(1);
3672 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3673 SmallVector<SDValue, 4> MemOps;
3675 // void *__stack at offset 0
3677 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3678 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3679 MachinePointerInfo(SV), false, false, 8));
3681 // void *__gr_top at offset 8
3682 int GPRSize = FuncInfo->getVarArgsGPRSize();
3684 SDValue GRTop, GRTopAddr;
3686 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3687 DAG.getConstant(8, getPointerTy()));
3689 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), getPointerTy());
3690 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
3691 DAG.getConstant(GPRSize, getPointerTy()));
3693 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3694 MachinePointerInfo(SV, 8), false, false, 8));
3697 // void *__vr_top at offset 16
3698 int FPRSize = FuncInfo->getVarArgsFPRSize();
3700 SDValue VRTop, VRTopAddr;
3701 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3702 DAG.getConstant(16, getPointerTy()));
3704 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), getPointerTy());
3705 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
3706 DAG.getConstant(FPRSize, getPointerTy()));
3708 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3709 MachinePointerInfo(SV, 16), false, false, 8));
3712 // int __gr_offs at offset 24
3713 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3714 DAG.getConstant(24, getPointerTy()));
3715 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
3716 GROffsAddr, MachinePointerInfo(SV, 24), false,
3719 // int __vr_offs at offset 28
3720 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3721 DAG.getConstant(28, getPointerTy()));
3722 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
3723 VROffsAddr, MachinePointerInfo(SV, 28), false,
3726 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
3730 SDValue ARM64TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
3731 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
3732 : LowerAAPCS_VASTART(Op, DAG);
3735 SDValue ARM64TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
3736 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
3738 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
3739 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3740 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
3742 return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op), Op.getOperand(1),
3743 Op.getOperand(2), DAG.getConstant(VaListSize, MVT::i32),
3744 8, false, false, MachinePointerInfo(DestSV),
3745 MachinePointerInfo(SrcSV));
3748 SDValue ARM64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
3749 assert(Subtarget->isTargetDarwin() &&
3750 "automatic va_arg instruction only works on Darwin");
3752 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3753 EVT VT = Op.getValueType();
3755 SDValue Chain = Op.getOperand(0);
3756 SDValue Addr = Op.getOperand(1);
3757 unsigned Align = Op.getConstantOperandVal(3);
3759 SDValue VAList = DAG.getLoad(getPointerTy(), DL, Chain, Addr,
3760 MachinePointerInfo(V), false, false, false, 0);
3761 Chain = VAList.getValue(1);
3764 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
3765 VAList = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3766 DAG.getConstant(Align - 1, getPointerTy()));
3767 VAList = DAG.getNode(ISD::AND, DL, getPointerTy(), VAList,
3768 DAG.getConstant(-(int64_t)Align, getPointerTy()));
3771 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
3772 uint64_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
3774 // Scalar integer and FP values smaller than 64 bits are implicitly extended
3775 // up to 64 bits. At the very least, we have to increase the striding of the
3776 // vaargs list to match this, and for FP values we need to introduce
3777 // FP_ROUND nodes as well.
3778 if (VT.isInteger() && !VT.isVector())
3780 bool NeedFPTrunc = false;
3781 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
3786 // Increment the pointer, VAList, to the next vaarg
3787 SDValue VANext = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3788 DAG.getConstant(ArgSize, getPointerTy()));
3789 // Store the incremented VAList to the legalized pointer
3790 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
3793 // Load the actual argument out of the pointer VAList
3795 // Load the value as an f64.
3796 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
3797 MachinePointerInfo(), false, false, false, 0);
3798 // Round the value down to an f32.
3799 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
3800 DAG.getIntPtrConstant(1));
3801 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
3802 // Merge the rounded value with the chain output of the load.
3803 return DAG.getMergeValues(Ops, 2, DL);
3806 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
3810 SDValue ARM64TargetLowering::LowerFRAMEADDR(SDValue Op,
3811 SelectionDAG &DAG) const {
3812 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3813 MFI->setFrameAddressIsTaken(true);
3815 EVT VT = Op.getValueType();
3817 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3818 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, ARM64::FP, VT);
3820 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
3821 MachinePointerInfo(), false, false, false, 0);
3825 SDValue ARM64TargetLowering::LowerRETURNADDR(SDValue Op,
3826 SelectionDAG &DAG) const {
3827 MachineFunction &MF = DAG.getMachineFunction();
3828 MachineFrameInfo *MFI = MF.getFrameInfo();
3829 MFI->setReturnAddressIsTaken(true);
3831 EVT VT = Op.getValueType();
3833 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3835 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3836 SDValue Offset = DAG.getConstant(8, getPointerTy());
3837 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
3838 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
3839 MachinePointerInfo(), false, false, false, 0);
3842 // Return LR, which contains the return address. Mark it an implicit live-in.
3843 unsigned Reg = MF.addLiveIn(ARM64::LR, &ARM64::GPR64RegClass);
3844 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
3847 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3848 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3849 SDValue ARM64TargetLowering::LowerShiftRightParts(SDValue Op,
3850 SelectionDAG &DAG) const {
3851 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3852 EVT VT = Op.getValueType();
3853 unsigned VTBits = VT.getSizeInBits();
3855 SDValue ShOpLo = Op.getOperand(0);
3856 SDValue ShOpHi = Op.getOperand(1);
3857 SDValue ShAmt = Op.getOperand(2);
3859 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3861 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3863 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3864 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3865 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3866 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3867 DAG.getConstant(VTBits, MVT::i64));
3868 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3870 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
3871 ISD::SETGE, dl, DAG);
3872 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3874 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3875 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3877 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3879 // ARM64 shifts larger than the register width are wrapped rather than
3880 // clamped, so we can't just emit "hi >> x".
3881 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3882 SDValue TrueValHi = Opc == ISD::SRA
3883 ? DAG.getNode(Opc, dl, VT, ShOpHi,
3884 DAG.getConstant(VTBits - 1, MVT::i64))
3885 : DAG.getConstant(0, VT);
3887 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
3889 SDValue Ops[2] = { Lo, Hi };
3890 return DAG.getMergeValues(Ops, 2, dl);
3893 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3894 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3895 SDValue ARM64TargetLowering::LowerShiftLeftParts(SDValue Op,
3896 SelectionDAG &DAG) const {
3897 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3898 EVT VT = Op.getValueType();
3899 unsigned VTBits = VT.getSizeInBits();
3901 SDValue ShOpLo = Op.getOperand(0);
3902 SDValue ShOpHi = Op.getOperand(1);
3903 SDValue ShAmt = Op.getOperand(2);
3906 assert(Op.getOpcode() == ISD::SHL_PARTS);
3907 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3908 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3909 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3910 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3911 DAG.getConstant(VTBits, MVT::i64));
3912 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3913 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3915 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3917 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
3918 ISD::SETGE, dl, DAG);
3919 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3920 SDValue Hi = DAG.getNode(ARM64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
3922 // ARM64 shifts of larger than register sizes are wrapped rather than clamped,
3923 // so we can't just emit "lo << a" if a is too big.
3924 SDValue TrueValLo = DAG.getConstant(0, VT);
3925 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3927 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3929 SDValue Ops[2] = { Lo, Hi };
3930 return DAG.getMergeValues(Ops, 2, dl);
3934 ARM64TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3935 // The ARM64 target doesn't support folding offsets into global addresses.
3939 bool ARM64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3940 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
3941 // FIXME: We should be able to handle f128 as well with a clever lowering.
3942 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
3946 return ARM64_AM::getFP64Imm(Imm) != -1;
3947 else if (VT == MVT::f32)
3948 return ARM64_AM::getFP32Imm(Imm) != -1;
3952 //===----------------------------------------------------------------------===//
3953 // ARM64 Optimization Hooks
3954 //===----------------------------------------------------------------------===//
3956 //===----------------------------------------------------------------------===//
3957 // ARM64 Inline Assembly Support
3958 //===----------------------------------------------------------------------===//
3960 // Table of Constraints
3961 // TODO: This is the current set of constraints supported by ARM for the
3962 // compiler, not all of them may make sense, e.g. S may be difficult to support.
3964 // r - A general register
3965 // w - An FP/SIMD register of some size in the range v0-v31
3966 // x - An FP/SIMD register of some size in the range v0-v15
3967 // I - Constant that can be used with an ADD instruction
3968 // J - Constant that can be used with a SUB instruction
3969 // K - Constant that can be used with a 32-bit logical instruction
3970 // L - Constant that can be used with a 64-bit logical instruction
3971 // M - Constant that can be used as a 32-bit MOV immediate
3972 // N - Constant that can be used as a 64-bit MOV immediate
3973 // Q - A memory reference with base register and no offset
3974 // S - A symbolic address
3975 // Y - Floating point constant zero
3976 // Z - Integer constant zero
3978 // Note that general register operands will be output using their 64-bit x
3979 // register name, whatever the size of the variable, unless the asm operand
3980 // is prefixed by the %w modifier. Floating-point and SIMD register operands
3981 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
3984 /// getConstraintType - Given a constraint letter, return the type of
3985 /// constraint it is for this target.
3986 ARM64TargetLowering::ConstraintType
3987 ARM64TargetLowering::getConstraintType(const std::string &Constraint) const {
3988 if (Constraint.size() == 1) {
3989 switch (Constraint[0]) {
3996 return C_RegisterClass;
3997 // An address with a single base register. Due to the way we
3998 // currently handle addresses it is the same as 'r'.
4003 return TargetLowering::getConstraintType(Constraint);
4006 /// Examine constraint type and operand type and determine a weight value.
4007 /// This object must already have been set up with the operand type
4008 /// and the current alternative constraint selected.
4009 TargetLowering::ConstraintWeight
4010 ARM64TargetLowering::getSingleConstraintMatchWeight(
4011 AsmOperandInfo &info, const char *constraint) const {
4012 ConstraintWeight weight = CW_Invalid;
4013 Value *CallOperandVal = info.CallOperandVal;
4014 // If we don't have a value, we can't do a match,
4015 // but allow it at the lowest weight.
4016 if (CallOperandVal == NULL)
4018 Type *type = CallOperandVal->getType();
4019 // Look at the constraint type.
4020 switch (*constraint) {
4022 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
4026 if (type->isFloatingPointTy() || type->isVectorTy())
4027 weight = CW_Register;
4030 weight = CW_Constant;
4036 std::pair<unsigned, const TargetRegisterClass *>
4037 ARM64TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4039 if (Constraint.size() == 1) {
4040 switch (Constraint[0]) {
4042 if (VT.getSizeInBits() == 64)
4043 return std::make_pair(0U, &ARM64::GPR64commonRegClass);
4044 return std::make_pair(0U, &ARM64::GPR32commonRegClass);
4047 return std::make_pair(0U, &ARM64::FPR32RegClass);
4048 if (VT.getSizeInBits() == 64)
4049 return std::make_pair(0U, &ARM64::FPR64RegClass);
4050 if (VT.getSizeInBits() == 128)
4051 return std::make_pair(0U, &ARM64::FPR128RegClass);
4053 // The instructions that this constraint is designed for can
4054 // only take 128-bit registers so just use that regclass.
4056 if (VT.getSizeInBits() == 128)
4057 return std::make_pair(0U, &ARM64::FPR128_loRegClass);
4061 if (StringRef("{cc}").equals_lower(Constraint))
4062 return std::make_pair(unsigned(ARM64::CPSR), &ARM64::CCRRegClass);
4064 // Use the default implementation in TargetLowering to convert the register
4065 // constraint into a member of a register class.
4066 std::pair<unsigned, const TargetRegisterClass *> Res;
4067 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4069 // Not found as a standard register?
4070 if (Res.second == 0) {
4071 unsigned Size = Constraint.size();
4072 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
4073 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
4074 const std::string Reg =
4075 std::string(&Constraint[2], &Constraint[Size - 1]);
4076 int RegNo = atoi(Reg.c_str());
4077 if (RegNo >= 0 && RegNo <= 31) {
4078 // v0 - v31 are aliases of q0 - q31.
4079 // By default we'll emit v0-v31 for this unless there's a modifier where
4080 // we'll emit the correct register as well.
4081 Res.first = ARM64::FPR128RegClass.getRegister(RegNo);
4082 Res.second = &ARM64::FPR128RegClass;
4090 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4091 /// vector. If it is invalid, don't add anything to Ops.
4092 void ARM64TargetLowering::LowerAsmOperandForConstraint(
4093 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
4094 SelectionDAG &DAG) const {
4095 SDValue Result(0, 0);
4097 // Currently only support length 1 constraints.
4098 if (Constraint.length() != 1)
4101 char ConstraintLetter = Constraint[0];
4102 switch (ConstraintLetter) {
4106 // This set of constraints deal with valid constants for various instructions.
4107 // Validate and return a target constant for them if we can.
4109 // 'z' maps to xzr or wzr so it needs an input of 0.
4110 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4111 if (!C || C->getZExtValue() != 0)
4114 if (Op.getValueType() == MVT::i64)
4115 Result = DAG.getRegister(ARM64::XZR, MVT::i64);
4117 Result = DAG.getRegister(ARM64::WZR, MVT::i32);
4127 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4131 // Grab the value and do some validation.
4132 uint64_t CVal = C->getZExtValue();
4133 switch (ConstraintLetter) {
4134 // The I constraint applies only to simple ADD or SUB immediate operands:
4135 // i.e. 0 to 4095 with optional shift by 12
4136 // The J constraint applies only to ADD or SUB immediates that would be
4137 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
4138 // instruction [or vice versa], in other words -1 to -4095 with optional
4139 // left shift by 12.
4141 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
4145 uint64_t NVal = -C->getSExtValue();
4146 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal))
4150 // The K and L constraints apply *only* to logical immediates, including
4151 // what used to be the MOVI alias for ORR (though the MOVI alias has now
4152 // been removed and MOV should be used). So these constraints have to
4153 // distinguish between bit patterns that are valid 32-bit or 64-bit
4154 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
4155 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
4158 if (ARM64_AM::isLogicalImmediate(CVal, 32))
4162 if (ARM64_AM::isLogicalImmediate(CVal, 64))
4165 // The M and N constraints are a superset of K and L respectively, for use
4166 // with the MOV (immediate) alias. As well as the logical immediates they
4167 // also match 32 or 64-bit immediates that can be loaded either using a
4168 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
4169 // (M) or 64-bit 0x1234000000000000 (N) etc.
4170 // As a note some of this code is liberally stolen from the asm parser.
4172 if (!isUInt<32>(CVal))
4174 if (ARM64_AM::isLogicalImmediate(CVal, 32))
4176 if ((CVal & 0xFFFF) == CVal)
4178 if ((CVal & 0xFFFF0000ULL) == CVal)
4180 uint64_t NCVal = ~(uint32_t)CVal;
4181 if ((NCVal & 0xFFFFULL) == NCVal)
4183 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4188 if (ARM64_AM::isLogicalImmediate(CVal, 64))
4190 if ((CVal & 0xFFFFULL) == CVal)
4192 if ((CVal & 0xFFFF0000ULL) == CVal)
4194 if ((CVal & 0xFFFF00000000ULL) == CVal)
4196 if ((CVal & 0xFFFF000000000000ULL) == CVal)
4198 uint64_t NCVal = ~CVal;
4199 if ((NCVal & 0xFFFFULL) == NCVal)
4201 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4203 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
4205 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
4213 // All assembler immediates are 64-bit integers.
4214 Result = DAG.getTargetConstant(CVal, MVT::i64);
4218 if (Result.getNode()) {
4219 Ops.push_back(Result);
4223 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4226 //===----------------------------------------------------------------------===//
4227 // ARM64 Advanced SIMD Support
4228 //===----------------------------------------------------------------------===//
4230 /// WidenVector - Given a value in the V64 register class, produce the
4231 /// equivalent value in the V128 register class.
4232 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
4233 EVT VT = V64Reg.getValueType();
4234 unsigned NarrowSize = VT.getVectorNumElements();
4235 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4236 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
4239 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
4240 V64Reg, DAG.getConstant(0, MVT::i32));
4243 /// getExtFactor - Determine the adjustment factor for the position when
4244 /// generating an "extract from vector registers" instruction.
4245 static unsigned getExtFactor(SDValue &V) {
4246 EVT EltType = V.getValueType().getVectorElementType();
4247 return EltType.getSizeInBits() / 8;
4250 /// NarrowVector - Given a value in the V128 register class, produce the
4251 /// equivalent value in the V64 register class.
4252 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
4253 EVT VT = V128Reg.getValueType();
4254 unsigned WideSize = VT.getVectorNumElements();
4255 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4256 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
4259 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, NarrowTy, V128Reg);
4262 // Gather data to see if the operation can be modelled as a
4263 // shuffle in combination with VEXTs.
4264 SDValue ARM64TargetLowering::ReconstructShuffle(SDValue Op,
4265 SelectionDAG &DAG) const {
4267 EVT VT = Op.getValueType();
4268 unsigned NumElts = VT.getVectorNumElements();
4270 SmallVector<SDValue, 2> SourceVecs;
4271 SmallVector<unsigned, 2> MinElts;
4272 SmallVector<unsigned, 2> MaxElts;
4274 for (unsigned i = 0; i < NumElts; ++i) {
4275 SDValue V = Op.getOperand(i);
4276 if (V.getOpcode() == ISD::UNDEF)
4278 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4279 // A shuffle can only come from building a vector from various
4280 // elements of other vectors.
4284 // Record this extraction against the appropriate vector if possible...
4285 SDValue SourceVec = V.getOperand(0);
4286 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4287 bool FoundSource = false;
4288 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4289 if (SourceVecs[j] == SourceVec) {
4290 if (MinElts[j] > EltNo)
4292 if (MaxElts[j] < EltNo)
4299 // Or record a new source if not...
4301 SourceVecs.push_back(SourceVec);
4302 MinElts.push_back(EltNo);
4303 MaxElts.push_back(EltNo);
4307 // Currently only do something sane when at most two source vectors
4309 if (SourceVecs.size() > 2)
4312 SDValue ShuffleSrcs[2] = { DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4313 int VEXTOffsets[2] = { 0, 0 };
4315 // This loop extracts the usage patterns of the source vectors
4316 // and prepares appropriate SDValues for a shuffle if possible.
4317 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4318 if (SourceVecs[i].getValueType() == VT) {
4319 // No VEXT necessary
4320 ShuffleSrcs[i] = SourceVecs[i];
4323 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4324 // It probably isn't worth padding out a smaller vector just to
4325 // break it down again in a shuffle.
4329 // Don't attempt to extract subvectors from BUILD_VECTOR sources
4330 // that expand or trunc the original value.
4331 // TODO: We can try to bitcast and ANY_EXTEND the result but
4332 // we need to consider the cost of vector ANY_EXTEND, and the
4333 // legality of all the types.
4334 if (SourceVecs[i].getValueType().getVectorElementType() !=
4335 VT.getVectorElementType())
4338 // Since only 64-bit and 128-bit vectors are legal on ARM and
4339 // we've eliminated the other cases...
4340 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2 * NumElts &&
4341 "unexpected vector sizes in ReconstructShuffle");
4343 if (MaxElts[i] - MinElts[i] >= NumElts) {
4344 // Span too large for a VEXT to cope
4348 if (MinElts[i] >= NumElts) {
4349 // The extraction can just take the second half
4350 VEXTOffsets[i] = NumElts;
4352 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
4353 DAG.getIntPtrConstant(NumElts));
4354 } else if (MaxElts[i] < NumElts) {
4355 // The extraction can just take the first half
4357 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4358 SourceVecs[i], DAG.getIntPtrConstant(0));
4360 // An actual VEXT is needed
4361 VEXTOffsets[i] = MinElts[i];
4362 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4363 SourceVecs[i], DAG.getIntPtrConstant(0));
4365 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
4366 DAG.getIntPtrConstant(NumElts));
4367 unsigned Imm = VEXTOffsets[i] * getExtFactor(VEXTSrc1);
4368 ShuffleSrcs[i] = DAG.getNode(ARM64ISD::EXT, dl, VT, VEXTSrc1, VEXTSrc2,
4369 DAG.getConstant(Imm, MVT::i32));
4373 SmallVector<int, 8> Mask;
4375 for (unsigned i = 0; i < NumElts; ++i) {
4376 SDValue Entry = Op.getOperand(i);
4377 if (Entry.getOpcode() == ISD::UNDEF) {
4382 SDValue ExtractVec = Entry.getOperand(0);
4384 cast<ConstantSDNode>(Op.getOperand(i).getOperand(1))->getSExtValue();
4385 if (ExtractVec == SourceVecs[0]) {
4386 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4388 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4392 // Final check before we try to produce nonsense...
4393 if (isShuffleMaskLegal(Mask, VT))
4394 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4400 // check if an EXT instruction can handle the shuffle mask when the
4401 // vector sources of the shuffle are the same.
4402 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4403 unsigned NumElts = VT.getVectorNumElements();
4405 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4411 // If this is a VEXT shuffle, the immediate value is the index of the first
4412 // element. The other shuffle indices must be the successive elements after
4414 unsigned ExpectedElt = Imm;
4415 for (unsigned i = 1; i < NumElts; ++i) {
4416 // Increment the expected index. If it wraps around, just follow it
4417 // back to index zero and keep going.
4419 if (ExpectedElt == NumElts)
4423 continue; // ignore UNDEF indices
4424 if (ExpectedElt != static_cast<unsigned>(M[i]))
4431 // check if an EXT instruction can handle the shuffle mask when the
4432 // vector sources of the shuffle are different.
4433 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
4435 unsigned NumElts = VT.getVectorNumElements();
4438 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4444 // If this is a VEXT shuffle, the immediate value is the index of the first
4445 // element. The other shuffle indices must be the successive elements after
4447 unsigned ExpectedElt = Imm;
4448 for (unsigned i = 1; i < NumElts; ++i) {
4449 // Increment the expected index. If it wraps around, it may still be
4450 // a VEXT but the source vectors must be swapped.
4452 if (ExpectedElt == NumElts * 2) {
4458 continue; // ignore UNDEF indices
4459 if (ExpectedElt != static_cast<unsigned>(M[i]))
4463 // Adjust the index value if the source operands will be swapped.
4470 /// isREVMask - Check if a vector shuffle corresponds to a REV
4471 /// instruction with the specified blocksize. (The order of the elements
4472 /// within each block of the vector is reversed.)
4473 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4474 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
4475 "Only possible block sizes for REV are: 16, 32, 64");
4477 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4481 unsigned NumElts = VT.getVectorNumElements();
4482 unsigned BlockElts = M[0] + 1;
4483 // If the first shuffle index is UNDEF, be optimistic.
4485 BlockElts = BlockSize / EltSz;
4487 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4490 for (unsigned i = 0; i < NumElts; ++i) {
4492 continue; // ignore UNDEF indices
4493 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
4500 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4501 unsigned NumElts = VT.getVectorNumElements();
4502 WhichResult = (M[0] == 0 ? 0 : 1);
4503 unsigned Idx = WhichResult * NumElts / 2;
4504 for (unsigned i = 0; i != NumElts; i += 2) {
4505 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4506 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
4514 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4515 unsigned NumElts = VT.getVectorNumElements();
4516 WhichResult = (M[0] == 0 ? 0 : 1);
4517 for (unsigned i = 0; i != NumElts; ++i) {
4519 continue; // ignore UNDEF indices
4520 if ((unsigned)M[i] != 2 * i + WhichResult)
4527 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4528 unsigned NumElts = VT.getVectorNumElements();
4529 WhichResult = (M[0] == 0 ? 0 : 1);
4530 for (unsigned i = 0; i < NumElts; i += 2) {
4531 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4532 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
4538 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
4539 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4540 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4541 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4542 unsigned NumElts = VT.getVectorNumElements();
4543 WhichResult = (M[0] == 0 ? 0 : 1);
4544 unsigned Idx = WhichResult * NumElts / 2;
4545 for (unsigned i = 0; i != NumElts; i += 2) {
4546 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4547 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
4555 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
4556 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4557 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4558 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4559 unsigned Half = VT.getVectorNumElements() / 2;
4560 WhichResult = (M[0] == 0 ? 0 : 1);
4561 for (unsigned j = 0; j != 2; ++j) {
4562 unsigned Idx = WhichResult;
4563 for (unsigned i = 0; i != Half; ++i) {
4564 int MIdx = M[i + j * Half];
4565 if (MIdx >= 0 && (unsigned)MIdx != Idx)
4574 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
4575 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4576 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4577 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4578 unsigned NumElts = VT.getVectorNumElements();
4579 WhichResult = (M[0] == 0 ? 0 : 1);
4580 for (unsigned i = 0; i < NumElts; i += 2) {
4581 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4582 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
4588 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4589 /// the specified operations to build the shuffle.
4590 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4591 SDValue RHS, SelectionDAG &DAG,
4593 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4594 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
4595 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
4598 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4607 OP_VUZPL, // VUZP, left result
4608 OP_VUZPR, // VUZP, right result
4609 OP_VZIPL, // VZIP, left result
4610 OP_VZIPR, // VZIP, right result
4611 OP_VTRNL, // VTRN, left result
4612 OP_VTRNR // VTRN, right result
4615 if (OpNum == OP_COPY) {
4616 if (LHSID == (1 * 9 + 2) * 9 + 3)
4618 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
4622 SDValue OpLHS, OpRHS;
4623 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4624 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4625 EVT VT = OpLHS.getValueType();
4629 llvm_unreachable("Unknown shuffle opcode!");
4631 // VREV divides the vector in half and swaps within the half.
4632 if (VT.getVectorElementType() == MVT::i32 ||
4633 VT.getVectorElementType() == MVT::f32)
4634 return DAG.getNode(ARM64ISD::REV64, dl, VT, OpLHS);
4635 // vrev <4 x i16> -> REV32
4636 if (VT.getVectorElementType() == MVT::i16)
4637 return DAG.getNode(ARM64ISD::REV32, dl, VT, OpLHS);
4638 // vrev <4 x i8> -> REV16
4639 assert(VT.getVectorElementType() == MVT::i8);
4640 return DAG.getNode(ARM64ISD::REV16, dl, VT, OpLHS);
4645 EVT EltTy = VT.getVectorElementType();
4647 if (EltTy == MVT::i8)
4648 Opcode = ARM64ISD::DUPLANE8;
4649 else if (EltTy == MVT::i16)
4650 Opcode = ARM64ISD::DUPLANE16;
4651 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
4652 Opcode = ARM64ISD::DUPLANE32;
4653 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
4654 Opcode = ARM64ISD::DUPLANE64;
4656 llvm_unreachable("Invalid vector element type?");
4658 if (VT.getSizeInBits() == 64)
4659 OpLHS = WidenVector(OpLHS, DAG);
4660 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, MVT::i64);
4661 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
4666 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
4667 return DAG.getNode(ARM64ISD::EXT, dl, VT, OpLHS, OpRHS,
4668 DAG.getConstant(Imm, MVT::i32));
4671 return DAG.getNode(ARM64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4673 return DAG.getNode(ARM64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4675 return DAG.getNode(ARM64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4677 return DAG.getNode(ARM64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4679 return DAG.getNode(ARM64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4681 return DAG.getNode(ARM64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4685 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
4686 SelectionDAG &DAG) {
4687 // Check to see if we can use the TBL instruction.
4688 SDValue V1 = Op.getOperand(0);
4689 SDValue V2 = Op.getOperand(1);
4692 EVT EltVT = Op.getValueType().getVectorElementType();
4693 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
4695 SmallVector<SDValue, 8> TBLMask;
4696 for (int Val : ShuffleMask) {
4697 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
4698 unsigned Offset = Byte + Val * BytesPerElt;
4699 TBLMask.push_back(DAG.getConstant(Offset, MVT::i32));
4703 MVT IndexVT = MVT::v8i8;
4704 unsigned IndexLen = 8;
4705 if (Op.getValueType().getSizeInBits() == 128) {
4706 IndexVT = MVT::v16i8;
4710 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
4711 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
4714 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
4716 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
4717 Shuffle = DAG.getNode(
4718 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4719 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4720 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4722 if (IndexLen == 8) {
4723 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
4724 Shuffle = DAG.getNode(
4725 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4726 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4727 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4729 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
4730 // cannot currently represent the register constraints on the input
4732 // Shuffle = DAG.getNode(ARM64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
4733 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4734 // &TBLMask[0], IndexLen));
4735 Shuffle = DAG.getNode(
4736 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4737 DAG.getConstant(Intrinsic::arm64_neon_tbl2, MVT::i32), V1Cst, V2Cst,
4738 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4741 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
4744 static unsigned getDUPLANEOp(EVT EltType) {
4745 if (EltType == MVT::i8)
4746 return ARM64ISD::DUPLANE8;
4747 if (EltType == MVT::i16)
4748 return ARM64ISD::DUPLANE16;
4749 if (EltType == MVT::i32 || EltType == MVT::f32)
4750 return ARM64ISD::DUPLANE32;
4751 if (EltType == MVT::i64 || EltType == MVT::f64)
4752 return ARM64ISD::DUPLANE64;
4754 llvm_unreachable("Invalid vector element type?");
4757 SDValue ARM64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4758 SelectionDAG &DAG) const {
4760 EVT VT = Op.getValueType();
4762 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4764 // Convert shuffles that are directly supported on NEON to target-specific
4765 // DAG nodes, instead of keeping them as shuffles and matching them again
4766 // during code selection. This is more efficient and avoids the possibility
4767 // of inconsistencies between legalization and selection.
4768 ArrayRef<int> ShuffleMask = SVN->getMask();
4770 SDValue V1 = Op.getOperand(0);
4771 SDValue V2 = Op.getOperand(1);
4773 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
4774 V1.getValueType().getSimpleVT())) {
4775 int Lane = SVN->getSplatIndex();
4776 // If this is undef splat, generate it via "just" vdup, if possible.
4780 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
4781 return DAG.getNode(ARM64ISD::DUP, dl, V1.getValueType(),
4783 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
4784 // constant. If so, we can just reference the lane's definition directly.
4785 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
4786 !isa<ConstantSDNode>(V1.getOperand(Lane)))
4787 return DAG.getNode(ARM64ISD::DUP, dl, VT, V1.getOperand(Lane));
4789 // Otherwise, duplicate from the lane of the input vector.
4790 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
4792 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
4793 // to make a vector of the same size as this SHUFFLE. We can ignore the
4794 // extract entirely, and canonicalise the concat using WidenVector.
4795 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
4796 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
4797 V1 = V1.getOperand(0);
4798 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
4799 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
4800 Lane -= Idx * VT.getVectorNumElements() / 2;
4801 V1 = WidenVector(V1.getOperand(Idx), DAG);
4802 } else if (VT.getSizeInBits() == 64)
4803 V1 = WidenVector(V1, DAG);
4805 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, MVT::i64));
4808 if (isREVMask(ShuffleMask, VT, 64))
4809 return DAG.getNode(ARM64ISD::REV64, dl, V1.getValueType(), V1, V2);
4810 if (isREVMask(ShuffleMask, VT, 32))
4811 return DAG.getNode(ARM64ISD::REV32, dl, V1.getValueType(), V1, V2);
4812 if (isREVMask(ShuffleMask, VT, 16))
4813 return DAG.getNode(ARM64ISD::REV16, dl, V1.getValueType(), V1, V2);
4815 bool ReverseEXT = false;
4817 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
4820 Imm *= getExtFactor(V1);
4821 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V2,
4822 DAG.getConstant(Imm, MVT::i32));
4823 } else if (V2->getOpcode() == ISD::UNDEF &&
4824 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
4825 Imm *= getExtFactor(V1);
4826 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V1,
4827 DAG.getConstant(Imm, MVT::i32));
4830 unsigned WhichResult;
4831 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
4832 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4833 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4835 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
4836 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4837 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4839 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
4840 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4841 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4844 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4845 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4846 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4848 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4849 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4850 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4852 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4853 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4854 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4857 // If the shuffle is not directly supported and it has 4 elements, use
4858 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4859 unsigned NumElts = VT.getVectorNumElements();
4861 unsigned PFIndexes[4];
4862 for (unsigned i = 0; i != 4; ++i) {
4863 if (ShuffleMask[i] < 0)
4866 PFIndexes[i] = ShuffleMask[i];
4869 // Compute the index in the perfect shuffle table.
4870 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
4871 PFIndexes[2] * 9 + PFIndexes[3];
4872 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4873 unsigned Cost = (PFEntry >> 30);
4876 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4879 return GenerateTBL(Op, ShuffleMask, DAG);
4882 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
4884 EVT VT = BVN->getValueType(0);
4885 APInt SplatBits, SplatUndef;
4886 unsigned SplatBitSize;
4888 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4889 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
4891 for (unsigned i = 0; i < NumSplats; ++i) {
4892 CnstBits <<= SplatBitSize;
4893 UndefBits <<= SplatBitSize;
4894 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
4895 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
4904 SDValue ARM64TargetLowering::LowerVectorAND(SDValue Op,
4905 SelectionDAG &DAG) const {
4906 BuildVectorSDNode *BVN =
4907 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
4908 SDValue LHS = Op.getOperand(0);
4910 EVT VT = Op.getValueType();
4915 APInt CnstBits(VT.getSizeInBits(), 0);
4916 APInt UndefBits(VT.getSizeInBits(), 0);
4917 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
4918 // We only have BIC vector immediate instruction, which is and-not.
4919 CnstBits = ~CnstBits;
4921 // We make use of a little bit of goto ickiness in order to avoid having to
4922 // duplicate the immediate matching logic for the undef toggled case.
4923 bool SecondTry = false;
4926 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
4927 CnstBits = CnstBits.zextOrTrunc(64);
4928 uint64_t CnstVal = CnstBits.getZExtValue();
4930 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
4931 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
4932 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4933 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4934 DAG.getConstant(CnstVal, MVT::i32),
4935 DAG.getConstant(0, MVT::i32));
4936 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4939 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
4940 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
4941 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4942 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4943 DAG.getConstant(CnstVal, MVT::i32),
4944 DAG.getConstant(8, MVT::i32));
4945 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4948 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
4949 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
4950 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4951 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4952 DAG.getConstant(CnstVal, MVT::i32),
4953 DAG.getConstant(16, MVT::i32));
4954 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4957 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
4958 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
4959 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4960 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4961 DAG.getConstant(CnstVal, MVT::i32),
4962 DAG.getConstant(24, MVT::i32));
4963 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4966 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
4967 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
4968 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4969 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4970 DAG.getConstant(CnstVal, MVT::i32),
4971 DAG.getConstant(0, MVT::i32));
4972 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4975 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
4976 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
4977 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4978 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4979 DAG.getConstant(CnstVal, MVT::i32),
4980 DAG.getConstant(8, MVT::i32));
4981 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4988 CnstBits = ~UndefBits;
4992 // We can always fall back to a non-immediate AND.
4997 // Specialized code to quickly find if PotentialBVec is a BuildVector that
4998 // consists of only the same constant int value, returned in reference arg
5000 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
5001 uint64_t &ConstVal) {
5002 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
5005 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
5008 EVT VT = Bvec->getValueType(0);
5009 unsigned NumElts = VT.getVectorNumElements();
5010 for (unsigned i = 1; i < NumElts; ++i)
5011 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
5013 ConstVal = FirstElt->getZExtValue();
5017 static unsigned getIntrinsicID(const SDNode *N) {
5018 unsigned Opcode = N->getOpcode();
5021 return Intrinsic::not_intrinsic;
5022 case ISD::INTRINSIC_WO_CHAIN: {
5023 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5024 if (IID < Intrinsic::num_intrinsics)
5026 return Intrinsic::not_intrinsic;
5031 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
5032 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
5033 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
5034 // Also, logical shift right -> sri, with the same structure.
5035 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
5036 EVT VT = N->getValueType(0);
5043 // Is the first op an AND?
5044 const SDValue And = N->getOperand(0);
5045 if (And.getOpcode() != ISD::AND)
5048 // Is the second op an shl or lshr?
5049 SDValue Shift = N->getOperand(1);
5050 // This will have been turned into: ARM64ISD::VSHL vector, #shift
5051 // or ARM64ISD::VLSHR vector, #shift
5052 unsigned ShiftOpc = Shift.getOpcode();
5053 if ((ShiftOpc != ARM64ISD::VSHL && ShiftOpc != ARM64ISD::VLSHR))
5055 bool IsShiftRight = ShiftOpc == ARM64ISD::VLSHR;
5057 // Is the shift amount constant?
5058 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
5062 // Is the and mask vector all constant?
5064 if (!isAllConstantBuildVector(And.getOperand(1), C1))
5067 // Is C1 == ~C2, taking into account how much one can shift elements of a
5069 uint64_t C2 = C2node->getZExtValue();
5070 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
5071 if (C2 > ElemSizeInBits)
5073 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
5074 if ((C1 & ElemMask) != (~C2 & ElemMask))
5077 SDValue X = And.getOperand(0);
5078 SDValue Y = Shift.getOperand(0);
5081 IsShiftRight ? Intrinsic::arm64_neon_vsri : Intrinsic::arm64_neon_vsli;
5083 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5084 DAG.getConstant(Intrin, MVT::i32), X, Y, Shift.getOperand(1));
5086 DEBUG(dbgs() << "arm64-lower: transformed: \n");
5087 DEBUG(N->dump(&DAG));
5088 DEBUG(dbgs() << "into: \n");
5089 DEBUG(ResultSLI->dump(&DAG));
5095 SDValue ARM64TargetLowering::LowerVectorOR(SDValue Op,
5096 SelectionDAG &DAG) const {
5097 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
5098 if (EnableARM64SlrGeneration) {
5099 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
5104 BuildVectorSDNode *BVN =
5105 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
5106 SDValue LHS = Op.getOperand(1);
5108 EVT VT = Op.getValueType();
5110 // OR commutes, so try swapping the operands.
5112 LHS = Op.getOperand(0);
5113 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5118 APInt CnstBits(VT.getSizeInBits(), 0);
5119 APInt UndefBits(VT.getSizeInBits(), 0);
5120 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5121 // We make use of a little bit of goto ickiness in order to avoid having to
5122 // duplicate the immediate matching logic for the undef toggled case.
5123 bool SecondTry = false;
5126 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5127 CnstBits = CnstBits.zextOrTrunc(64);
5128 uint64_t CnstVal = CnstBits.getZExtValue();
5130 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5131 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5132 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5133 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5134 DAG.getConstant(CnstVal, MVT::i32),
5135 DAG.getConstant(0, MVT::i32));
5136 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5139 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5140 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5141 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5142 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5143 DAG.getConstant(CnstVal, MVT::i32),
5144 DAG.getConstant(8, MVT::i32));
5145 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5148 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5149 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5150 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5151 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5152 DAG.getConstant(CnstVal, MVT::i32),
5153 DAG.getConstant(16, MVT::i32));
5154 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5157 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5158 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5159 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5160 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5161 DAG.getConstant(CnstVal, MVT::i32),
5162 DAG.getConstant(24, MVT::i32));
5163 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5166 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5167 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5168 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5169 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5170 DAG.getConstant(CnstVal, MVT::i32),
5171 DAG.getConstant(0, MVT::i32));
5172 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5175 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5176 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5177 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5178 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5179 DAG.getConstant(CnstVal, MVT::i32),
5180 DAG.getConstant(8, MVT::i32));
5181 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5188 CnstBits = UndefBits;
5192 // We can always fall back to a non-immediate OR.
5197 SDValue ARM64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
5198 SelectionDAG &DAG) const {
5199 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5201 EVT VT = Op.getValueType();
5203 APInt CnstBits(VT.getSizeInBits(), 0);
5204 APInt UndefBits(VT.getSizeInBits(), 0);
5205 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5206 // We make use of a little bit of goto ickiness in order to avoid having to
5207 // duplicate the immediate matching logic for the undef toggled case.
5208 bool SecondTry = false;
5211 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5212 CnstBits = CnstBits.zextOrTrunc(64);
5213 uint64_t CnstVal = CnstBits.getZExtValue();
5215 // Certain magic vector constants (used to express things like NOT
5216 // and NEG) are passed through unmodified. This allows codegen patterns
5217 // for these operations to match. Special-purpose patterns will lower
5218 // these immediates to MOVIs if it proves necessary.
5219 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
5222 // The many faces of MOVI...
5223 if (ARM64_AM::isAdvSIMDModImmType10(CnstVal)) {
5224 CnstVal = ARM64_AM::encodeAdvSIMDModImmType10(CnstVal);
5225 if (VT.getSizeInBits() == 128) {
5226 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::v2i64,
5227 DAG.getConstant(CnstVal, MVT::i32));
5228 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5231 // Support the V64 version via subregister insertion.
5232 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::f64,
5233 DAG.getConstant(CnstVal, MVT::i32));
5234 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5237 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5238 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5239 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5240 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5241 DAG.getConstant(CnstVal, MVT::i32),
5242 DAG.getConstant(0, MVT::i32));
5243 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5246 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5247 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5248 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5249 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5250 DAG.getConstant(CnstVal, MVT::i32),
5251 DAG.getConstant(8, MVT::i32));
5252 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5255 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5256 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5257 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5258 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5259 DAG.getConstant(CnstVal, MVT::i32),
5260 DAG.getConstant(16, MVT::i32));
5261 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5264 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5265 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5266 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5267 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5268 DAG.getConstant(CnstVal, MVT::i32),
5269 DAG.getConstant(24, MVT::i32));
5270 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5273 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5274 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5275 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5276 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5277 DAG.getConstant(CnstVal, MVT::i32),
5278 DAG.getConstant(0, MVT::i32));
5279 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5282 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5283 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5284 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5285 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5286 DAG.getConstant(CnstVal, MVT::i32),
5287 DAG.getConstant(8, MVT::i32));
5288 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5291 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
5292 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
5293 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5294 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
5295 DAG.getConstant(CnstVal, MVT::i32),
5296 DAG.getConstant(264, MVT::i32));
5297 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5300 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
5301 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
5302 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5303 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
5304 DAG.getConstant(CnstVal, MVT::i32),
5305 DAG.getConstant(272, MVT::i32));
5306 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5309 if (ARM64_AM::isAdvSIMDModImmType9(CnstVal)) {
5310 CnstVal = ARM64_AM::encodeAdvSIMDModImmType9(CnstVal);
5311 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
5312 SDValue Mov = DAG.getNode(ARM64ISD::MOVI, dl, MovTy,
5313 DAG.getConstant(CnstVal, MVT::i32));
5314 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5317 // The few faces of FMOV...
5318 if (ARM64_AM::isAdvSIMDModImmType11(CnstVal)) {
5319 CnstVal = ARM64_AM::encodeAdvSIMDModImmType11(CnstVal);
5320 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
5321 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MovTy,
5322 DAG.getConstant(CnstVal, MVT::i32));
5323 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5326 if (ARM64_AM::isAdvSIMDModImmType12(CnstVal) &&
5327 VT.getSizeInBits() == 128) {
5328 CnstVal = ARM64_AM::encodeAdvSIMDModImmType12(CnstVal);
5329 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MVT::v2f64,
5330 DAG.getConstant(CnstVal, MVT::i32));
5331 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5334 // The many faces of MVNI...
5336 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5337 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5338 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5339 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5340 DAG.getConstant(CnstVal, MVT::i32),
5341 DAG.getConstant(0, MVT::i32));
5342 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5345 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5346 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5347 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5348 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5349 DAG.getConstant(CnstVal, MVT::i32),
5350 DAG.getConstant(8, MVT::i32));
5351 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5354 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5355 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5356 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5357 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5358 DAG.getConstant(CnstVal, MVT::i32),
5359 DAG.getConstant(16, MVT::i32));
5360 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5363 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5364 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5365 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5366 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5367 DAG.getConstant(CnstVal, MVT::i32),
5368 DAG.getConstant(24, MVT::i32));
5369 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5372 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5373 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5374 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5375 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5376 DAG.getConstant(CnstVal, MVT::i32),
5377 DAG.getConstant(0, MVT::i32));
5378 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5381 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5382 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5383 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5384 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5385 DAG.getConstant(CnstVal, MVT::i32),
5386 DAG.getConstant(8, MVT::i32));
5387 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5390 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
5391 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
5392 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5393 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
5394 DAG.getConstant(CnstVal, MVT::i32),
5395 DAG.getConstant(264, MVT::i32));
5396 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5399 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
5400 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
5401 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5402 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
5403 DAG.getConstant(CnstVal, MVT::i32),
5404 DAG.getConstant(272, MVT::i32));
5405 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5412 CnstBits = UndefBits;
5417 // Scan through the operands to find some interesting properties we can
5419 // 1) If only one value is used, we can use a DUP, or
5420 // 2) if only the low element is not undef, we can just insert that, or
5421 // 3) if only one constant value is used (w/ some non-constant lanes),
5422 // we can splat the constant value into the whole vector then fill
5423 // in the non-constant lanes.
5424 // 4) FIXME: If different constant values are used, but we can intelligently
5425 // select the values we'll be overwriting for the non-constant
5426 // lanes such that we can directly materialize the vector
5427 // some other way (MOVI, e.g.), we can be sneaky.
5428 unsigned NumElts = VT.getVectorNumElements();
5429 bool isOnlyLowElement = true;
5430 bool usesOnlyOneValue = true;
5431 bool usesOnlyOneConstantValue = true;
5432 bool isConstant = true;
5433 unsigned NumConstantLanes = 0;
5435 SDValue ConstantValue;
5436 for (unsigned i = 0; i < NumElts; ++i) {
5437 SDValue V = Op.getOperand(i);
5438 if (V.getOpcode() == ISD::UNDEF)
5441 isOnlyLowElement = false;
5442 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5445 if (isa<ConstantSDNode>(V)) {
5447 if (!ConstantValue.getNode())
5449 else if (ConstantValue != V)
5450 usesOnlyOneConstantValue = false;
5453 if (!Value.getNode())
5455 else if (V != Value)
5456 usesOnlyOneValue = false;
5459 if (!Value.getNode())
5460 return DAG.getUNDEF(VT);
5462 if (isOnlyLowElement)
5463 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5465 // Use DUP for non-constant splats. For f32 constant splats, reduce to
5466 // i32 and try again.
5467 if (usesOnlyOneValue) {
5469 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5470 Value.getValueType() != VT)
5471 return DAG.getNode(ARM64ISD::DUP, dl, VT, Value);
5473 // This is actually a DUPLANExx operation, which keeps everything vectory.
5475 // DUPLANE works on 128-bit vectors, widen it if necessary.
5476 SDValue Lane = Value.getOperand(1);
5477 Value = Value.getOperand(0);
5478 if (Value.getValueType().getSizeInBits() == 64)
5479 Value = WidenVector(Value, DAG);
5481 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
5482 return DAG.getNode(Opcode, dl, VT, Value, Lane);
5485 if (VT.getVectorElementType().isFloatingPoint()) {
5486 SmallVector<SDValue, 8> Ops;
5488 (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
5489 for (unsigned i = 0; i < NumElts; ++i)
5490 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
5491 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
5492 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
5493 Val = LowerBUILD_VECTOR(Val, DAG);
5495 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5499 // If there was only one constant value used and for more than one lane,
5500 // start by splatting that value, then replace the non-constant lanes. This
5501 // is better than the default, which will perform a separate initialization
5503 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
5504 SDValue Val = DAG.getNode(ARM64ISD::DUP, dl, VT, ConstantValue);
5505 // Now insert the non-constant lanes.
5506 for (unsigned i = 0; i < NumElts; ++i) {
5507 SDValue V = Op.getOperand(i);
5508 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5509 if (!isa<ConstantSDNode>(V)) {
5510 // Note that type legalization likely mucked about with the VT of the
5511 // source operand, so we may have to convert it here before inserting.
5512 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
5518 // If all elements are constants and the case above didn't get hit, fall back
5519 // to the default expansion, which will generate a load from the constant
5524 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5526 SDValue shuffle = ReconstructShuffle(Op, DAG);
5527 if (shuffle != SDValue())
5531 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5532 // know the default expansion would otherwise fall back on something even
5533 // worse. For a vector with one or two non-undef values, that's
5534 // scalar_to_vector for the elements followed by a shuffle (provided the
5535 // shuffle is valid for the target) and materialization element by element
5536 // on the stack followed by a load for everything else.
5537 if (!isConstant && !usesOnlyOneValue) {
5538 SDValue Vec = DAG.getUNDEF(VT);
5539 SDValue Op0 = Op.getOperand(0);
5540 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
5542 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
5543 // a) Avoid a RMW dependency on the full vector register, and
5544 // b) Allow the register coalescer to fold away the copy if the
5545 // value is already in an S or D register.
5546 if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
5547 unsigned SubIdx = ElemSize == 32 ? ARM64::ssub : ARM64::dsub;
5549 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
5550 DAG.getTargetConstant(SubIdx, MVT::i32));
5551 Vec = SDValue(N, 0);
5554 for (; i < NumElts; ++i) {
5555 SDValue V = Op.getOperand(i);
5556 if (V.getOpcode() == ISD::UNDEF)
5558 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5559 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5564 // Just use the default expansion. We failed to find a better alternative.
5568 SDValue ARM64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
5569 SelectionDAG &DAG) const {
5570 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
5572 // Check for non-constant lane.
5573 if (!isa<ConstantSDNode>(Op.getOperand(2)))
5576 EVT VT = Op.getOperand(0).getValueType();
5578 // Insertion/extraction are legal for V128 types.
5579 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5580 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5583 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5584 VT != MVT::v1i64 && VT != MVT::v2f32)
5587 // For V64 types, we perform insertion by expanding the value
5588 // to a V128 type and perform the insertion on that.
5590 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5591 EVT WideTy = WideVec.getValueType();
5593 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
5594 Op.getOperand(1), Op.getOperand(2));
5595 // Re-narrow the resultant vector.
5596 return NarrowVector(Node, DAG);
5599 SDValue ARM64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5600 SelectionDAG &DAG) const {
5601 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
5603 // Check for non-constant lane.
5604 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5607 EVT VT = Op.getOperand(0).getValueType();
5609 // Insertion/extraction are legal for V128 types.
5610 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5611 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5614 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5615 VT != MVT::v1i64 && VT != MVT::v2f32)
5618 // For V64 types, we perform extraction by expanding the value
5619 // to a V128 type and perform the extraction on that.
5621 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5622 EVT WideTy = WideVec.getValueType();
5624 EVT ExtrTy = WideTy.getVectorElementType();
5625 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
5628 // For extractions, we just return the result directly.
5629 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
5633 SDValue ARM64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
5634 SelectionDAG &DAG) const {
5635 EVT VT = Op.getOperand(0).getValueType();
5641 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5644 unsigned Val = Cst->getZExtValue();
5646 unsigned Size = Op.getValueType().getSizeInBits();
5650 return DAG.getTargetExtractSubreg(ARM64::bsub, dl, Op.getValueType(),
5653 return DAG.getTargetExtractSubreg(ARM64::hsub, dl, Op.getValueType(),
5656 return DAG.getTargetExtractSubreg(ARM64::ssub, dl, Op.getValueType(),
5659 return DAG.getTargetExtractSubreg(ARM64::dsub, dl, Op.getValueType(),
5662 llvm_unreachable("Unexpected vector type in extract_subvector!");
5665 // If this is extracting the upper 64-bits of a 128-bit vector, we match
5667 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
5673 bool ARM64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5675 if (VT.getVectorNumElements() == 4 &&
5676 (VT.is128BitVector() || VT.is64BitVector())) {
5677 unsigned PFIndexes[4];
5678 for (unsigned i = 0; i != 4; ++i) {
5682 PFIndexes[i] = M[i];
5685 // Compute the index in the perfect shuffle table.
5686 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5687 PFIndexes[2] * 9 + PFIndexes[3];
5688 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5689 unsigned Cost = (PFEntry >> 30);
5696 unsigned Imm, WhichResult;
5698 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
5699 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
5700 isEXTMask(M, VT, ReverseVEXT, Imm) ||
5701 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
5702 isTRNMask(M, VT, WhichResult) || isUZPMask(M, VT, WhichResult) ||
5703 isZIPMask(M, VT, WhichResult) ||
5704 isTRN_v_undef_Mask(M, VT, WhichResult) ||
5705 isUZP_v_undef_Mask(M, VT, WhichResult) ||
5706 isZIP_v_undef_Mask(M, VT, WhichResult));
5709 /// getVShiftImm - Check if this is a valid build_vector for the immediate
5710 /// operand of a vector shift operation, where all the elements of the
5711 /// build_vector must have the same constant integer value.
5712 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
5713 // Ignore bit_converts.
5714 while (Op.getOpcode() == ISD::BITCAST)
5715 Op = Op.getOperand(0);
5716 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
5717 APInt SplatBits, SplatUndef;
5718 unsigned SplatBitSize;
5720 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
5721 HasAnyUndefs, ElementBits) ||
5722 SplatBitSize > ElementBits)
5724 Cnt = SplatBits.getSExtValue();
5728 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
5729 /// operand of a vector shift left operation. That value must be in the range:
5730 /// 0 <= Value < ElementBits for a left shift; or
5731 /// 0 <= Value <= ElementBits for a long left shift.
5732 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
5733 assert(VT.isVector() && "vector shift count is not a vector type");
5734 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5735 if (!getVShiftImm(Op, ElementBits, Cnt))
5737 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
5740 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
5741 /// operand of a vector shift right operation. For a shift opcode, the value
5742 /// is positive, but for an intrinsic the value count must be negative. The
5743 /// absolute value must be in the range:
5744 /// 1 <= |Value| <= ElementBits for a right shift; or
5745 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
5746 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
5748 assert(VT.isVector() && "vector shift count is not a vector type");
5749 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5750 if (!getVShiftImm(Op, ElementBits, Cnt))
5754 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
5757 SDValue ARM64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
5758 SelectionDAG &DAG) const {
5759 EVT VT = Op.getValueType();
5763 if (!Op.getOperand(1).getValueType().isVector())
5765 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5767 switch (Op.getOpcode()) {
5769 llvm_unreachable("unexpected shift opcode");
5772 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
5773 return DAG.getNode(ARM64ISD::VSHL, SDLoc(Op), VT, Op.getOperand(0),
5774 DAG.getConstant(Cnt, MVT::i32));
5775 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5776 DAG.getConstant(Intrinsic::arm64_neon_ushl, MVT::i32),
5777 Op.getOperand(0), Op.getOperand(1));
5780 // Right shift immediate
5781 if (isVShiftRImm(Op.getOperand(1), VT, false, false, Cnt) &&
5784 (Op.getOpcode() == ISD::SRA) ? ARM64ISD::VASHR : ARM64ISD::VLSHR;
5785 return DAG.getNode(Opc, SDLoc(Op), VT, Op.getOperand(0),
5786 DAG.getConstant(Cnt, MVT::i32));
5789 // Right shift register. Note, there is not a shift right register
5790 // instruction, but the shift left register instruction takes a signed
5791 // value, where negative numbers specify a right shift.
5792 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::arm64_neon_sshl
5793 : Intrinsic::arm64_neon_ushl;
5794 // negate the shift amount
5795 SDValue NegShift = DAG.getNode(ARM64ISD::NEG, DL, VT, Op.getOperand(1));
5796 SDValue NegShiftLeft =
5797 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5798 DAG.getConstant(Opc, MVT::i32), Op.getOperand(0), NegShift);
5799 return NegShiftLeft;
5805 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
5806 ARM64CC::CondCode CC, bool NoNans, EVT VT,
5807 SDLoc dl, SelectionDAG &DAG) {
5808 EVT SrcVT = LHS.getValueType();
5810 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
5811 APInt CnstBits(VT.getSizeInBits(), 0);
5812 APInt UndefBits(VT.getSizeInBits(), 0);
5813 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
5814 bool IsZero = IsCnst && (CnstBits == 0);
5816 if (SrcVT.getVectorElementType().isFloatingPoint()) {
5823 Fcmeq = DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5825 Fcmeq = DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5826 return DAG.getNode(ARM64ISD::NOT, dl, VT, Fcmeq);
5830 return DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5831 return DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5834 return DAG.getNode(ARM64ISD::FCMGEz, dl, VT, LHS);
5835 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, LHS, RHS);
5838 return DAG.getNode(ARM64ISD::FCMGTz, dl, VT, LHS);
5839 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, LHS, RHS);
5842 return DAG.getNode(ARM64ISD::FCMLEz, dl, VT, LHS);
5843 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, RHS, LHS);
5847 // If we ignore NaNs then we can use to the MI implementation.
5851 return DAG.getNode(ARM64ISD::FCMLTz, dl, VT, LHS);
5852 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, RHS, LHS);
5862 Cmeq = DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5864 Cmeq = DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5865 return DAG.getNode(ARM64ISD::NOT, dl, VT, Cmeq);
5869 return DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5870 return DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5873 return DAG.getNode(ARM64ISD::CMGEz, dl, VT, LHS);
5874 return DAG.getNode(ARM64ISD::CMGE, dl, VT, LHS, RHS);
5877 return DAG.getNode(ARM64ISD::CMGTz, dl, VT, LHS);
5878 return DAG.getNode(ARM64ISD::CMGT, dl, VT, LHS, RHS);
5881 return DAG.getNode(ARM64ISD::CMLEz, dl, VT, LHS);
5882 return DAG.getNode(ARM64ISD::CMGE, dl, VT, RHS, LHS);
5884 return DAG.getNode(ARM64ISD::CMHS, dl, VT, RHS, LHS);
5886 return DAG.getNode(ARM64ISD::CMHI, dl, VT, RHS, LHS);
5889 return DAG.getNode(ARM64ISD::CMLTz, dl, VT, LHS);
5890 return DAG.getNode(ARM64ISD::CMGT, dl, VT, RHS, LHS);
5892 return DAG.getNode(ARM64ISD::CMHI, dl, VT, LHS, RHS);
5894 return DAG.getNode(ARM64ISD::CMHS, dl, VT, LHS, RHS);
5898 SDValue ARM64TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
5899 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5900 SDValue LHS = Op.getOperand(0);
5901 SDValue RHS = Op.getOperand(1);
5904 if (LHS.getValueType().getVectorElementType().isInteger()) {
5905 assert(LHS.getValueType() == RHS.getValueType());
5906 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
5907 return EmitVectorComparison(LHS, RHS, ARM64CC, false, Op.getValueType(), dl,
5911 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
5912 LHS.getValueType().getVectorElementType() == MVT::f64);
5914 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
5915 // clean. Some of them require two branches to implement.
5916 ARM64CC::CondCode CC1, CC2;
5917 changeFPCCToARM64CC(CC, CC1, CC2);
5919 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
5921 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, Op.getValueType(), dl, DAG);
5922 if (!Cmp1.getNode())
5925 if (CC2 != ARM64CC::AL) {
5927 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, Op.getValueType(), dl, DAG);
5928 if (!Cmp2.getNode())
5931 return DAG.getNode(ISD::OR, dl, Cmp1.getValueType(), Cmp1, Cmp2);
5937 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
5938 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
5939 /// specified in the intrinsic calls.
5940 bool ARM64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5942 unsigned Intrinsic) const {
5943 switch (Intrinsic) {
5944 case Intrinsic::arm64_neon_ld2:
5945 case Intrinsic::arm64_neon_ld3:
5946 case Intrinsic::arm64_neon_ld4:
5947 case Intrinsic::arm64_neon_ld2lane:
5948 case Intrinsic::arm64_neon_ld3lane:
5949 case Intrinsic::arm64_neon_ld4lane:
5950 case Intrinsic::arm64_neon_ld2r:
5951 case Intrinsic::arm64_neon_ld3r:
5952 case Intrinsic::arm64_neon_ld4r: {
5953 Info.opc = ISD::INTRINSIC_W_CHAIN;
5954 // Conservatively set memVT to the entire set of vectors loaded.
5955 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
5956 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5957 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5960 Info.vol = false; // volatile loads with NEON intrinsics not supported
5961 Info.readMem = true;
5962 Info.writeMem = false;
5965 case Intrinsic::arm64_neon_st2:
5966 case Intrinsic::arm64_neon_st3:
5967 case Intrinsic::arm64_neon_st4:
5968 case Intrinsic::arm64_neon_st2lane:
5969 case Intrinsic::arm64_neon_st3lane:
5970 case Intrinsic::arm64_neon_st4lane: {
5971 Info.opc = ISD::INTRINSIC_VOID;
5972 // Conservatively set memVT to the entire set of vectors stored.
5973 unsigned NumElts = 0;
5974 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5975 Type *ArgTy = I.getArgOperand(ArgI)->getType();
5976 if (!ArgTy->isVectorTy())
5978 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
5980 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5981 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5984 Info.vol = false; // volatile stores with NEON intrinsics not supported
5985 Info.readMem = false;
5986 Info.writeMem = true;
5989 case Intrinsic::arm64_ldaxr:
5990 case Intrinsic::arm64_ldxr: {
5991 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
5992 Info.opc = ISD::INTRINSIC_W_CHAIN;
5993 Info.memVT = MVT::getVT(PtrTy->getElementType());
5994 Info.ptrVal = I.getArgOperand(0);
5996 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
5998 Info.readMem = true;
5999 Info.writeMem = false;
6002 case Intrinsic::arm64_stlxr:
6003 case Intrinsic::arm64_stxr: {
6004 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
6005 Info.opc = ISD::INTRINSIC_W_CHAIN;
6006 Info.memVT = MVT::getVT(PtrTy->getElementType());
6007 Info.ptrVal = I.getArgOperand(1);
6009 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
6011 Info.readMem = false;
6012 Info.writeMem = true;
6015 case Intrinsic::arm64_ldaxp:
6016 case Intrinsic::arm64_ldxp: {
6017 Info.opc = ISD::INTRINSIC_W_CHAIN;
6018 Info.memVT = MVT::i128;
6019 Info.ptrVal = I.getArgOperand(0);
6023 Info.readMem = true;
6024 Info.writeMem = false;
6027 case Intrinsic::arm64_stlxp:
6028 case Intrinsic::arm64_stxp: {
6029 Info.opc = ISD::INTRINSIC_W_CHAIN;
6030 Info.memVT = MVT::i128;
6031 Info.ptrVal = I.getArgOperand(2);
6035 Info.readMem = false;
6036 Info.writeMem = true;
6046 // Truncations from 64-bit GPR to 32-bit GPR is free.
6047 bool ARM64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
6048 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6050 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6051 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6052 if (NumBits1 <= NumBits2)
6056 bool ARM64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
6057 if (!VT1.isInteger() || !VT2.isInteger())
6059 unsigned NumBits1 = VT1.getSizeInBits();
6060 unsigned NumBits2 = VT2.getSizeInBits();
6061 if (NumBits1 <= NumBits2)
6066 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
6068 bool ARM64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
6069 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6071 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6072 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6073 if (NumBits1 == 32 && NumBits2 == 64)
6077 bool ARM64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
6078 if (!VT1.isInteger() || !VT2.isInteger())
6080 unsigned NumBits1 = VT1.getSizeInBits();
6081 unsigned NumBits2 = VT2.getSizeInBits();
6082 if (NumBits1 == 32 && NumBits2 == 64)
6087 bool ARM64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
6088 EVT VT1 = Val.getValueType();
6089 if (isZExtFree(VT1, VT2)) {
6093 if (Val.getOpcode() != ISD::LOAD)
6096 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
6097 return (VT1.isSimple() && VT1.isInteger() && VT2.isSimple() &&
6098 VT2.isInteger() && VT1.getSizeInBits() <= 32);
6101 bool ARM64TargetLowering::hasPairedLoad(Type *LoadedType,
6102 unsigned &RequiredAligment) const {
6103 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
6105 // Cyclone supports unaligned accesses.
6106 RequiredAligment = 0;
6107 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
6108 return NumBits == 32 || NumBits == 64;
6111 bool ARM64TargetLowering::hasPairedLoad(EVT LoadedType,
6112 unsigned &RequiredAligment) const {
6113 if (!LoadedType.isSimple() ||
6114 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
6116 // Cyclone supports unaligned accesses.
6117 RequiredAligment = 0;
6118 unsigned NumBits = LoadedType.getSizeInBits();
6119 return NumBits == 32 || NumBits == 64;
6122 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
6123 unsigned AlignCheck) {
6124 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
6125 (DstAlign == 0 || DstAlign % AlignCheck == 0));
6128 EVT ARM64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
6129 unsigned SrcAlign, bool IsMemset,
6130 bool ZeroMemset, bool MemcpyStrSrc,
6131 MachineFunction &MF) const {
6132 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
6133 // instruction to materialize the v2i64 zero and one store (with restrictive
6134 // addressing mode). Just do two i64 store of zero-registers.
6136 const Function *F = MF.getFunction();
6137 if (!IsMemset && Size >= 16 &&
6138 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
6139 Attribute::NoImplicitFloat) &&
6140 (memOpAlign(SrcAlign, DstAlign, 16) ||
6141 (allowsUnalignedMemoryAccesses(MVT::v2i64, 0, &Fast) && Fast)))
6144 return Size >= 8 ? MVT::i64 : MVT::i32;
6147 // 12-bit optionally shifted immediates are legal for adds.
6148 bool ARM64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
6149 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
6154 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
6155 // immediates is the same as for an add or a sub.
6156 bool ARM64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
6159 return isLegalAddImmediate(Immed);
6162 /// isLegalAddressingMode - Return true if the addressing mode represented
6163 /// by AM is legal for this target, for a load/store of the specified type.
6164 bool ARM64TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6166 // ARM64 has five basic addressing modes:
6168 // reg + 9-bit signed offset
6169 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
6171 // reg + SIZE_IN_BYTES * reg
6173 // No global is ever allowed as a base.
6177 // No reg+reg+imm addressing.
6178 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
6181 // check reg + imm case:
6182 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
6183 uint64_t NumBytes = 0;
6184 if (Ty->isSized()) {
6185 uint64_t NumBits = getDataLayout()->getTypeSizeInBits(Ty);
6186 NumBytes = NumBits / 8;
6187 if (!isPowerOf2_64(NumBits))
6192 int64_t Offset = AM.BaseOffs;
6194 // 9-bit signed offset
6195 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
6198 // 12-bit unsigned offset
6199 unsigned shift = Log2_64(NumBytes);
6200 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
6201 // Must be a multiple of NumBytes (NumBytes is a power of 2)
6202 (Offset >> shift) << shift == Offset)
6207 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
6209 if (!AM.Scale || AM.Scale == 1 ||
6210 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
6215 int ARM64TargetLowering::getScalingFactorCost(const AddrMode &AM,
6217 // Scaling factors are not free at all.
6218 // Operands | Rt Latency
6219 // -------------------------------------------
6221 // -------------------------------------------
6222 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
6223 // Rt, [Xn, Wm, <extend> #imm] |
6224 if (isLegalAddressingMode(AM, Ty))
6225 // Scale represents reg2 * scale, thus account for 1 if
6226 // it is not equal to 0 or 1.
6227 return AM.Scale != 0 && AM.Scale != 1;
6231 bool ARM64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
6232 VT = VT.getScalarType();
6237 switch (VT.getSimpleVT().SimpleTy) {
6249 ARM64TargetLowering::getScratchRegisters(CallingConv::ID) const {
6250 // LR is a callee-save register, but we must treat it as clobbered by any call
6251 // site. Hence we include LR in the scratch registers, which are in turn added
6252 // as implicit-defs for stackmaps and patchpoints.
6253 static const MCPhysReg ScratchRegs[] = {
6254 ARM64::X16, ARM64::X17, ARM64::LR, 0
6259 bool ARM64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
6261 assert(Ty->isIntegerTy());
6263 unsigned BitSize = Ty->getPrimitiveSizeInBits();
6267 int64_t Val = Imm.getSExtValue();
6268 if (Val == 0 || ARM64_AM::isLogicalImmediate(Val, BitSize))
6271 if ((int64_t)Val < 0)
6274 Val &= (1LL << 32) - 1;
6276 unsigned LZ = countLeadingZeros((uint64_t)Val);
6277 unsigned Shift = (63 - LZ) / 16;
6278 // MOVZ is free so return true for one or fewer MOVK.
6279 return (Shift < 3) ? true : false;
6282 // Generate SUBS and CSEL for integer abs.
6283 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
6284 EVT VT = N->getValueType(0);
6286 SDValue N0 = N->getOperand(0);
6287 SDValue N1 = N->getOperand(1);
6290 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
6291 // and change it to SUB and CSEL.
6292 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
6293 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
6294 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
6295 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
6296 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
6297 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
6299 // Generate SUBS & CSEL.
6301 DAG.getNode(ARM64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
6302 N0.getOperand(0), DAG.getConstant(0, VT));
6303 return DAG.getNode(ARM64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
6304 DAG.getConstant(ARM64CC::PL, MVT::i32),
6305 SDValue(Cmp.getNode(), 1));
6310 // performXorCombine - Attempts to handle integer ABS.
6311 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
6312 TargetLowering::DAGCombinerInfo &DCI,
6313 const ARM64Subtarget *Subtarget) {
6314 if (DCI.isBeforeLegalizeOps())
6317 return performIntegerAbsCombine(N, DAG);
6320 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
6321 TargetLowering::DAGCombinerInfo &DCI,
6322 const ARM64Subtarget *Subtarget) {
6323 if (DCI.isBeforeLegalizeOps())
6326 // Multiplication of a power of two plus/minus one can be done more
6327 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
6328 // future CPUs have a cheaper MADD instruction, this may need to be
6329 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
6330 // 64-bit is 5 cycles, so this is always a win.
6331 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
6332 APInt Value = C->getAPIntValue();
6333 EVT VT = N->getValueType(0);
6334 APInt VP1 = Value + 1;
6335 if (VP1.isPowerOf2()) {
6336 // Multiplying by one less than a power of two, replace with a shift
6338 SDValue ShiftedVal =
6339 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6340 DAG.getConstant(VP1.logBase2(), MVT::i64));
6341 return DAG.getNode(ISD::SUB, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6343 APInt VM1 = Value - 1;
6344 if (VM1.isPowerOf2()) {
6345 // Multiplying by one more than a power of two, replace with a shift
6347 SDValue ShiftedVal =
6348 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6349 DAG.getConstant(VM1.logBase2(), MVT::i64));
6350 return DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6356 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG) {
6357 EVT VT = N->getValueType(0);
6358 if (VT != MVT::f32 && VT != MVT::f64)
6360 // Only optimize when the source and destination types have the same width.
6361 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
6364 // If the result of an integer load is only used by an integer-to-float
6365 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
6366 // This eliminates an "integer-to-vector-move UOP and improve throughput.
6367 SDValue N0 = N->getOperand(0);
6368 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6369 // Do not change the width of a volatile load.
6370 !cast<LoadSDNode>(N0)->isVolatile()) {
6371 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6372 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
6373 LN0->getPointerInfo(), LN0->isVolatile(),
6374 LN0->isNonTemporal(), LN0->isInvariant(),
6375 LN0->getAlignment());
6377 // Make sure successors of the original load stay after it by updating them
6378 // to use the new Chain.
6379 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
6382 (N->getOpcode() == ISD::SINT_TO_FP) ? ARM64ISD::SITOF : ARM64ISD::UITOF;
6383 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
6389 /// An EXTR instruction is made up of two shifts, ORed together. This helper
6390 /// searches for and classifies those shifts.
6391 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
6393 if (N.getOpcode() == ISD::SHL)
6395 else if (N.getOpcode() == ISD::SRL)
6400 if (!isa<ConstantSDNode>(N.getOperand(1)))
6403 ShiftAmount = N->getConstantOperandVal(1);
6404 Src = N->getOperand(0);
6408 /// EXTR instruction extracts a contiguous chunk of bits from two existing
6409 /// registers viewed as a high/low pair. This function looks for the pattern:
6410 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
6411 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
6413 static SDValue tryCombineToEXTR(SDNode *N,
6414 TargetLowering::DAGCombinerInfo &DCI) {
6415 SelectionDAG &DAG = DCI.DAG;
6417 EVT VT = N->getValueType(0);
6419 assert(N->getOpcode() == ISD::OR && "Unexpected root");
6421 if (VT != MVT::i32 && VT != MVT::i64)
6425 uint32_t ShiftLHS = 0;
6427 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
6431 uint32_t ShiftRHS = 0;
6433 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
6436 // If they're both trying to come from the high part of the register, they're
6437 // not really an EXTR.
6438 if (LHSFromHi == RHSFromHi)
6441 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
6445 std::swap(LHS, RHS);
6446 std::swap(ShiftLHS, ShiftRHS);
6449 return DAG.getNode(ARM64ISD::EXTR, DL, VT, LHS, RHS,
6450 DAG.getConstant(ShiftRHS, MVT::i64));
6453 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
6454 const ARM64Subtarget *Subtarget) {
6455 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
6456 if (!EnableARM64ExtrGeneration)
6458 SelectionDAG &DAG = DCI.DAG;
6459 EVT VT = N->getValueType(0);
6461 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6464 SDValue Res = tryCombineToEXTR(N, DCI);
6471 static SDValue performBitcastCombine(SDNode *N,
6472 TargetLowering::DAGCombinerInfo &DCI,
6473 SelectionDAG &DAG) {
6474 // Wait 'til after everything is legalized to try this. That way we have
6475 // legal vector types and such.
6476 if (DCI.isBeforeLegalizeOps())
6479 // Remove extraneous bitcasts around an extract_subvector.
6481 // (v4i16 (bitconvert
6482 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
6484 // (extract_subvector ((v8i16 ...), (i64 4)))
6486 // Only interested in 64-bit vectors as the ultimate result.
6487 EVT VT = N->getValueType(0);
6490 if (VT.getSimpleVT().getSizeInBits() != 64)
6492 // Is the operand an extract_subvector starting at the beginning or halfway
6493 // point of the vector? A low half may also come through as an
6494 // EXTRACT_SUBREG, so look for that, too.
6495 SDValue Op0 = N->getOperand(0);
6496 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
6497 !(Op0->isMachineOpcode() &&
6498 Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG))
6500 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
6501 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
6502 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
6504 } else if (Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG) {
6505 if (idx != ARM64::dsub)
6507 // The dsub reference is equivalent to a lane zero subvector reference.
6510 // Look through the bitcast of the input to the extract.
6511 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
6513 SDValue Source = Op0->getOperand(0)->getOperand(0);
6514 // If the source type has twice the number of elements as our destination
6515 // type, we know this is an extract of the high or low half of the vector.
6516 EVT SVT = Source->getValueType(0);
6517 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
6520 DEBUG(dbgs() << "arm64-lower: bitcast extract_subvector simplification\n");
6522 // Create the simplified form to just extract the low or high half of the
6523 // vector directly rather than bothering with the bitcasts.
6525 unsigned NumElements = VT.getVectorNumElements();
6527 SDValue HalfIdx = DAG.getConstant(NumElements, MVT::i64);
6528 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
6530 SDValue SubReg = DAG.getTargetConstant(ARM64::dsub, MVT::i32);
6531 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
6537 static SDValue performConcatVectorsCombine(SDNode *N,
6538 TargetLowering::DAGCombinerInfo &DCI,
6539 SelectionDAG &DAG) {
6540 // Wait 'til after everything is legalized to try this. That way we have
6541 // legal vector types and such.
6542 if (DCI.isBeforeLegalizeOps())
6546 EVT VT = N->getValueType(0);
6548 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
6549 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
6550 // canonicalise to that.
6551 if (N->getOperand(0) == N->getOperand(1) && VT.getVectorNumElements() == 2) {
6552 assert(VT.getVectorElementType().getSizeInBits() == 64);
6553 return DAG.getNode(ARM64ISD::DUPLANE64, dl, VT,
6554 WidenVector(N->getOperand(0), DAG),
6555 DAG.getConstant(0, MVT::i64));
6558 // Canonicalise concat_vectors so that the right-hand vector has as few
6559 // bit-casts as possible before its real operation. The primary matching
6560 // destination for these operations will be the narrowing "2" instructions,
6561 // which depend on the operation being performed on this right-hand vector.
6563 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
6565 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
6567 SDValue Op1 = N->getOperand(1);
6568 if (Op1->getOpcode() != ISD::BITCAST)
6570 SDValue RHS = Op1->getOperand(0);
6571 MVT RHSTy = RHS.getValueType().getSimpleVT();
6572 // If the RHS is not a vector, this is not the pattern we're looking for.
6573 if (!RHSTy.isVector())
6576 DEBUG(dbgs() << "arm64-lower: concat_vectors bitcast simplification\n");
6578 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
6579 RHSTy.getVectorNumElements() * 2);
6581 ISD::BITCAST, dl, VT,
6582 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
6583 DAG.getNode(ISD::BITCAST, dl, RHSTy, N->getOperand(0)), RHS));
6586 static SDValue tryCombineFixedPointConvert(SDNode *N,
6587 TargetLowering::DAGCombinerInfo &DCI,
6588 SelectionDAG &DAG) {
6589 // Wait 'til after everything is legalized to try this. That way we have
6590 // legal vector types and such.
6591 if (DCI.isBeforeLegalizeOps())
6593 // Transform a scalar conversion of a value from a lane extract into a
6594 // lane extract of a vector conversion. E.g., from foo1 to foo2:
6595 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
6596 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
6598 // The second form interacts better with instruction selection and the
6599 // register allocator to avoid cross-class register copies that aren't
6600 // coalescable due to a lane reference.
6602 // Check the operand and see if it originates from a lane extract.
6603 SDValue Op1 = N->getOperand(1);
6604 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6605 // Yep, no additional predication needed. Perform the transform.
6606 SDValue IID = N->getOperand(0);
6607 SDValue Shift = N->getOperand(2);
6608 SDValue Vec = Op1.getOperand(0);
6609 SDValue Lane = Op1.getOperand(1);
6610 EVT ResTy = N->getValueType(0);
6614 // The vector width should be 128 bits by the time we get here, even
6615 // if it started as 64 bits (the extract_vector handling will have
6617 assert(Vec.getValueType().getSizeInBits() == 128 &&
6618 "unexpected vector size on extract_vector_elt!");
6619 if (Vec.getValueType() == MVT::v4i32)
6620 VecResTy = MVT::v4f32;
6621 else if (Vec.getValueType() == MVT::v2i64)
6622 VecResTy = MVT::v2f64;
6624 assert(0 && "unexpected vector type!");
6627 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
6628 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
6633 // AArch64 high-vector "long" operations are formed by performing the non-high
6634 // version on an extract_subvector of each operand which gets the high half:
6636 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
6638 // However, there are cases which don't have an extract_high explicitly, but
6639 // have another operation that can be made compatible with one for free. For
6642 // (dupv64 scalar) --> (extract_high (dup128 scalar))
6644 // This routine does the actual conversion of such DUPs, once outer routines
6645 // have determined that everything else is in order.
6646 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
6647 // We can handle most types of duplicate, but the lane ones have an extra
6648 // operand saying *which* lane, so we need to know.
6650 switch (N.getOpcode()) {
6654 case ARM64ISD::DUPLANE8:
6655 case ARM64ISD::DUPLANE16:
6656 case ARM64ISD::DUPLANE32:
6657 case ARM64ISD::DUPLANE64:
6664 MVT NarrowTy = N.getSimpleValueType();
6665 if (!NarrowTy.is64BitVector())
6668 MVT ElementTy = NarrowTy.getVectorElementType();
6669 unsigned NumElems = NarrowTy.getVectorNumElements();
6670 MVT NewDUPVT = MVT::getVectorVT(ElementTy, NumElems * 2);
6674 NewDUP = DAG.getNode(N.getOpcode(), SDLoc(N), NewDUPVT, N.getOperand(0),
6677 NewDUP = DAG.getNode(ARM64ISD::DUP, SDLoc(N), NewDUPVT, N.getOperand(0));
6679 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N.getNode()), NarrowTy,
6680 NewDUP, DAG.getConstant(NumElems, MVT::i64));
6683 static bool isEssentiallyExtractSubvector(SDValue N) {
6684 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
6687 return N.getOpcode() == ISD::BITCAST &&
6688 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
6691 /// \brief Helper structure to keep track of ISD::SET_CC operands.
6692 struct GenericSetCCInfo {
6693 const SDValue *Opnd0;
6694 const SDValue *Opnd1;
6698 /// \brief Helper structure to keep track of a SET_CC lowered into ARM64 code.
6699 struct ARM64SetCCInfo {
6701 ARM64CC::CondCode CC;
6704 /// \brief Helper structure to keep track of SetCC information.
6706 GenericSetCCInfo Generic;
6707 ARM64SetCCInfo ARM64;
6710 /// \brief Helper structure to be able to read SetCC information.
6711 /// If set to true, IsARM64 field, Info is a ARM64SetCCInfo, otherwise Info is
6712 /// a GenericSetCCInfo.
6713 struct SetCCInfoAndKind {
6718 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
6720 /// ARM64 lowered one.
6721 /// \p SetCCInfo is filled accordingly.
6722 /// \post SetCCInfo is meanginfull only when this function returns true.
6723 /// \return True when Op is a kind of SET_CC operation.
6724 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
6725 // If this is a setcc, this is straight forward.
6726 if (Op.getOpcode() == ISD::SETCC) {
6727 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
6728 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
6729 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6730 SetCCInfo.IsARM64 = false;
6733 // Otherwise, check if this is a matching csel instruction.
6737 if (Op.getOpcode() != ARM64ISD::CSEL)
6739 // Set the information about the operands.
6740 // TODO: we want the operands of the Cmp not the csel
6741 SetCCInfo.Info.ARM64.Cmp = &Op.getOperand(3);
6742 SetCCInfo.IsARM64 = true;
6743 SetCCInfo.Info.ARM64.CC = static_cast<ARM64CC::CondCode>(
6744 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
6746 // Check that the operands matches the constraints:
6747 // (1) Both operands must be constants.
6748 // (2) One must be 1 and the other must be 0.
6749 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
6750 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6753 if (!TValue || !FValue)
6757 if (!TValue->isOne()) {
6758 // Update the comparison when we are interested in !cc.
6759 std::swap(TValue, FValue);
6760 SetCCInfo.Info.ARM64.CC =
6761 ARM64CC::getInvertedCondCode(SetCCInfo.Info.ARM64.CC);
6763 return TValue->isOne() && FValue->isNullValue();
6766 // The folding we want to perform is:
6767 // (add x, (setcc cc ...) )
6769 // (csel x, (add x, 1), !cc ...)
6771 // The latter will get matched to a CSINC instruction.
6772 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
6773 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
6774 SDValue LHS = Op->getOperand(0);
6775 SDValue RHS = Op->getOperand(1);
6776 SetCCInfoAndKind InfoAndKind;
6778 // If neither operand is a SET_CC, give up.
6779 if (!isSetCC(LHS, InfoAndKind)) {
6780 std::swap(LHS, RHS);
6781 if (!isSetCC(LHS, InfoAndKind))
6785 // FIXME: This could be generatized to work for FP comparisons.
6786 EVT CmpVT = InfoAndKind.IsARM64
6787 ? InfoAndKind.Info.ARM64.Cmp->getOperand(0).getValueType()
6788 : InfoAndKind.Info.Generic.Opnd0->getValueType();
6789 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
6795 if (InfoAndKind.IsARM64) {
6796 CCVal = DAG.getConstant(
6797 ARM64CC::getInvertedCondCode(InfoAndKind.Info.ARM64.CC), MVT::i32);
6798 Cmp = *InfoAndKind.Info.ARM64.Cmp;
6800 Cmp = getARM64Cmp(*InfoAndKind.Info.Generic.Opnd0,
6801 *InfoAndKind.Info.Generic.Opnd1,
6802 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
6805 EVT VT = Op->getValueType(0);
6806 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, VT));
6807 return DAG.getNode(ARM64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
6810 // The basic add/sub long vector instructions have variants with "2" on the end
6811 // which act on the high-half of their inputs. They are normally matched by
6814 // (add (zeroext (extract_high LHS)),
6815 // (zeroext (extract_high RHS)))
6816 // -> uaddl2 vD, vN, vM
6818 // However, if one of the extracts is something like a duplicate, this
6819 // instruction can still be used profitably. This function puts the DAG into a
6820 // more appropriate form for those patterns to trigger.
6821 static SDValue performAddSubLongCombine(SDNode *N,
6822 TargetLowering::DAGCombinerInfo &DCI,
6823 SelectionDAG &DAG) {
6824 if (DCI.isBeforeLegalizeOps())
6827 MVT VT = N->getSimpleValueType(0);
6828 if (!VT.is128BitVector()) {
6829 if (N->getOpcode() == ISD::ADD)
6830 return performSetccAddFolding(N, DAG);
6834 // Make sure both branches are extended in the same way.
6835 SDValue LHS = N->getOperand(0);
6836 SDValue RHS = N->getOperand(1);
6837 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
6838 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
6839 LHS.getOpcode() != RHS.getOpcode())
6842 unsigned ExtType = LHS.getOpcode();
6844 // It's not worth doing if at least one of the inputs isn't already an
6845 // extract, but we don't know which it'll be so we have to try both.
6846 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
6847 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
6851 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
6852 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
6853 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
6857 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
6860 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
6863 // Massage DAGs which we can use the high-half "long" operations on into
6864 // something isel will recognize better. E.g.
6866 // (arm64_neon_umull (extract_high vec) (dupv64 scalar)) -->
6867 // (arm64_neon_umull (extract_high (v2i64 vec)))
6868 // (extract_high (v2i64 (dup128 scalar)))))
6870 static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N,
6871 TargetLowering::DAGCombinerInfo &DCI,
6872 SelectionDAG &DAG) {
6873 if (DCI.isBeforeLegalizeOps())
6876 SDValue LHS = N->getOperand(1);
6877 SDValue RHS = N->getOperand(2);
6878 assert(LHS.getValueType().is64BitVector() &&
6879 RHS.getValueType().is64BitVector() &&
6880 "unexpected shape for long operation");
6882 // Either node could be a DUP, but it's not worth doing both of them (you'd
6883 // just as well use the non-high version) so look for a corresponding extract
6884 // operation on the other "wing".
6885 if (isEssentiallyExtractSubvector(LHS)) {
6886 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
6889 } else if (isEssentiallyExtractSubvector(RHS)) {
6890 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
6895 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
6896 N->getOperand(0), LHS, RHS);
6899 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
6900 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
6901 unsigned ElemBits = ElemTy.getSizeInBits();
6903 int64_t ShiftAmount;
6904 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
6905 APInt SplatValue, SplatUndef;
6906 unsigned SplatBitSize;
6908 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
6909 HasAnyUndefs, ElemBits) ||
6910 SplatBitSize != ElemBits)
6913 ShiftAmount = SplatValue.getSExtValue();
6914 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
6915 ShiftAmount = CVN->getSExtValue();
6923 llvm_unreachable("Unknown shift intrinsic");
6924 case Intrinsic::arm64_neon_sqshl:
6925 Opcode = ARM64ISD::SQSHL_I;
6926 IsRightShift = false;
6928 case Intrinsic::arm64_neon_uqshl:
6929 Opcode = ARM64ISD::UQSHL_I;
6930 IsRightShift = false;
6932 case Intrinsic::arm64_neon_srshl:
6933 Opcode = ARM64ISD::SRSHR_I;
6934 IsRightShift = true;
6936 case Intrinsic::arm64_neon_urshl:
6937 Opcode = ARM64ISD::URSHR_I;
6938 IsRightShift = true;
6940 case Intrinsic::arm64_neon_sqshlu:
6941 Opcode = ARM64ISD::SQSHLU_I;
6942 IsRightShift = false;
6946 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits)
6947 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6948 DAG.getConstant(-ShiftAmount, MVT::i32));
6949 else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount <= ElemBits)
6950 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6951 DAG.getConstant(ShiftAmount, MVT::i32));
6956 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
6957 // the intrinsics must be legal and take an i32, this means there's almost
6958 // certainly going to be a zext in the DAG which we can eliminate.
6959 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
6960 SDValue AndN = N->getOperand(2);
6961 if (AndN.getOpcode() != ISD::AND)
6964 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
6965 if (!CMask || CMask->getZExtValue() != Mask)
6968 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
6969 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
6972 static SDValue performIntrinsicCombine(SDNode *N,
6973 TargetLowering::DAGCombinerInfo &DCI,
6974 const ARM64Subtarget *Subtarget) {
6975 SelectionDAG &DAG = DCI.DAG;
6976 unsigned IID = getIntrinsicID(N);
6980 case Intrinsic::arm64_neon_vcvtfxs2fp:
6981 case Intrinsic::arm64_neon_vcvtfxu2fp:
6982 return tryCombineFixedPointConvert(N, DCI, DAG);
6984 case Intrinsic::arm64_neon_fmax:
6985 return DAG.getNode(ARM64ISD::FMAX, SDLoc(N), N->getValueType(0),
6986 N->getOperand(1), N->getOperand(2));
6987 case Intrinsic::arm64_neon_fmin:
6988 return DAG.getNode(ARM64ISD::FMIN, SDLoc(N), N->getValueType(0),
6989 N->getOperand(1), N->getOperand(2));
6990 case Intrinsic::arm64_neon_smull:
6991 case Intrinsic::arm64_neon_umull:
6992 case Intrinsic::arm64_neon_pmull:
6993 case Intrinsic::arm64_neon_sqdmull:
6994 return tryCombineLongOpWithDup(IID, N, DCI, DAG);
6995 case Intrinsic::arm64_neon_sqshl:
6996 case Intrinsic::arm64_neon_uqshl:
6997 case Intrinsic::arm64_neon_sqshlu:
6998 case Intrinsic::arm64_neon_srshl:
6999 case Intrinsic::arm64_neon_urshl:
7000 return tryCombineShiftImm(IID, N, DAG);
7001 case Intrinsic::arm64_crc32b:
7002 case Intrinsic::arm64_crc32cb:
7003 return tryCombineCRC32(0xff, N, DAG);
7004 case Intrinsic::arm64_crc32h:
7005 case Intrinsic::arm64_crc32ch:
7006 return tryCombineCRC32(0xffff, N, DAG);
7011 static SDValue performExtendCombine(SDNode *N,
7012 TargetLowering::DAGCombinerInfo &DCI,
7013 SelectionDAG &DAG) {
7014 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
7015 // we can convert that DUP into another extract_high (of a bigger DUP), which
7016 // helps the backend to decide that an sabdl2 would be useful, saving a real
7017 // extract_high operation.
7018 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
7019 N->getOperand(0).getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
7020 SDNode *ABDNode = N->getOperand(0).getNode();
7021 unsigned IID = getIntrinsicID(ABDNode);
7022 if (IID == Intrinsic::arm64_neon_sabd ||
7023 IID == Intrinsic::arm64_neon_uabd) {
7024 SDValue NewABD = tryCombineLongOpWithDup(IID, ABDNode, DCI, DAG);
7025 if (!NewABD.getNode())
7028 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
7033 // This is effectively a custom type legalization for ARM64.
7035 // Type legalization will split an extend of a small, legal, type to a larger
7036 // illegal type by first splitting the destination type, often creating
7037 // illegal source types, which then get legalized in isel-confusing ways,
7038 // leading to really terrible codegen. E.g.,
7039 // %result = v8i32 sext v8i8 %value
7041 // %losrc = extract_subreg %value, ...
7042 // %hisrc = extract_subreg %value, ...
7043 // %lo = v4i32 sext v4i8 %losrc
7044 // %hi = v4i32 sext v4i8 %hisrc
7045 // Things go rapidly downhill from there.
7047 // For ARM64, the [sz]ext vector instructions can only go up one element
7048 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
7049 // take two instructions.
7051 // This implies that the most efficient way to do the extend from v8i8
7052 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
7053 // the normal splitting to happen for the v8i16->v8i32.
7055 // This is pre-legalization to catch some cases where the default
7056 // type legalization will create ill-tempered code.
7057 if (!DCI.isBeforeLegalizeOps())
7060 // We're only interested in cleaning things up for non-legal vector types
7061 // here. If both the source and destination are legal, things will just
7062 // work naturally without any fiddling.
7063 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7064 EVT ResVT = N->getValueType(0);
7065 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
7067 // If the vector type isn't a simple VT, it's beyond the scope of what
7068 // we're worried about here. Let legalization do its thing and hope for
7070 if (!ResVT.isSimple())
7073 SDValue Src = N->getOperand(0);
7074 MVT SrcVT = Src->getValueType(0).getSimpleVT();
7075 // If the source VT is a 64-bit vector, we can play games and get the
7076 // better results we want.
7077 if (SrcVT.getSizeInBits() != 64)
7080 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
7081 unsigned ElementCount = SrcVT.getVectorNumElements();
7082 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
7084 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
7086 // Now split the rest of the operation into two halves, each with a 64
7090 unsigned NumElements = ResVT.getVectorNumElements();
7091 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
7092 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
7093 ResVT.getVectorElementType(), NumElements / 2);
7095 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
7096 LoVT.getVectorNumElements());
7097 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7098 DAG.getIntPtrConstant(0));
7099 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7100 DAG.getIntPtrConstant(InNVT.getVectorNumElements()));
7101 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
7102 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
7104 // Now combine the parts back together so we still have a single result
7105 // like the combiner expects.
7106 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
7109 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
7110 /// value. The load store optimizer pass will merge them to store pair stores.
7111 /// This has better performance than a splat of the scalar followed by a split
7112 /// vector store. Even if the stores are not merged it is four stores vs a dup,
7113 /// followed by an ext.b and two stores.
7114 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
7115 SDValue StVal = St->getValue();
7116 EVT VT = StVal.getValueType();
7118 // Don't replace floating point stores, they possibly won't be transformed to
7119 // stp because of the store pair suppress pass.
7120 if (VT.isFloatingPoint())
7123 // Check for insert vector elements.
7124 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
7127 // We can express a splat as store pair(s) for 2 or 4 elements.
7128 unsigned NumVecElts = VT.getVectorNumElements();
7129 if (NumVecElts != 4 && NumVecElts != 2)
7131 SDValue SplatVal = StVal.getOperand(1);
7132 unsigned RemainInsertElts = NumVecElts - 1;
7134 // Check that this is a splat.
7135 while (--RemainInsertElts) {
7136 SDValue NextInsertElt = StVal.getOperand(0);
7137 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
7139 if (NextInsertElt.getOperand(1) != SplatVal)
7141 StVal = NextInsertElt;
7143 unsigned OrigAlignment = St->getAlignment();
7144 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
7145 unsigned Alignment = std::min(OrigAlignment, EltOffset);
7147 // Create scalar stores. This is at least as good as the code sequence for a
7148 // split unaligned store wich is a dup.s, ext.b, and two stores.
7149 // Most of the time the three stores should be replaced by store pair
7150 // instructions (stp).
7152 SDValue BasePtr = St->getBasePtr();
7154 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
7155 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
7157 unsigned Offset = EltOffset;
7158 while (--NumVecElts) {
7159 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7160 DAG.getConstant(Offset, MVT::i64));
7161 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
7162 St->getPointerInfo(), St->isVolatile(),
7163 St->isNonTemporal(), Alignment);
7164 Offset += EltOffset;
7169 static SDValue performSTORECombine(SDNode *N,
7170 TargetLowering::DAGCombinerInfo &DCI,
7172 const ARM64Subtarget *Subtarget) {
7173 if (!DCI.isBeforeLegalize())
7176 StoreSDNode *S = cast<StoreSDNode>(N);
7177 if (S->isVolatile())
7180 // Cyclone has bad performance on unaligned 16B stores when crossing line and
7181 // page boundries. We want to split such stores.
7182 if (!Subtarget->isCyclone())
7185 // Don't split at Oz.
7186 MachineFunction &MF = DAG.getMachineFunction();
7187 bool IsMinSize = MF.getFunction()->getAttributes().hasAttribute(
7188 AttributeSet::FunctionIndex, Attribute::MinSize);
7192 SDValue StVal = S->getValue();
7193 EVT VT = StVal.getValueType();
7195 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
7196 // those up regresses performance on micro-benchmarks and olden/bh.
7197 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
7200 // Split unaligned 16B stores. They are terrible for performance.
7201 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
7202 // extensions can use this to mark that it does not want splitting to happen
7203 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
7204 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
7205 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
7206 S->getAlignment() <= 2)
7209 // If we get a splat of a scalar convert this vector store to a store of
7210 // scalars. They will be merged into store pairs thereby removing two
7212 SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S);
7213 if (ReplacedSplat != SDValue())
7214 return ReplacedSplat;
7217 unsigned NumElts = VT.getVectorNumElements() / 2;
7218 // Split VT into two.
7220 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
7221 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7222 DAG.getIntPtrConstant(0));
7223 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7224 DAG.getIntPtrConstant(NumElts));
7225 SDValue BasePtr = S->getBasePtr();
7227 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
7228 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
7229 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7230 DAG.getConstant(8, MVT::i64));
7231 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
7232 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
7236 // Optimize compare with zero and branch.
7237 static SDValue performBRCONDCombine(SDNode *N,
7238 TargetLowering::DAGCombinerInfo &DCI,
7239 SelectionDAG &DAG) {
7240 SDValue Chain = N->getOperand(0);
7241 SDValue Dest = N->getOperand(1);
7242 SDValue CCVal = N->getOperand(2);
7243 SDValue Cmp = N->getOperand(3);
7245 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
7246 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
7247 if (CC != ARM64CC::EQ && CC != ARM64CC::NE)
7250 unsigned CmpOpc = Cmp.getOpcode();
7251 if (CmpOpc != ARM64ISD::ADDS && CmpOpc != ARM64ISD::SUBS)
7254 // Only attempt folding if there is only one use of the flag and no use of the
7256 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
7259 SDValue LHS = Cmp.getOperand(0);
7260 SDValue RHS = Cmp.getOperand(1);
7262 assert(LHS.getValueType() == RHS.getValueType() &&
7263 "Expected the value type to be the same for both operands!");
7264 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
7267 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
7268 std::swap(LHS, RHS);
7270 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
7273 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
7274 LHS.getOpcode() == ISD::SRL)
7277 // Fold the compare into the branch instruction.
7279 if (CC == ARM64CC::EQ)
7280 BR = DAG.getNode(ARM64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7282 BR = DAG.getNode(ARM64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7284 // Do not add new nodes to DAG combiner worklist.
7285 DCI.CombineTo(N, BR, false);
7290 // vselect (v1i1 setcc) ->
7291 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
7292 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
7293 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
7295 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
7296 SDValue N0 = N->getOperand(0);
7297 EVT CCVT = N0.getValueType();
7299 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
7300 CCVT.getVectorElementType() != MVT::i1)
7303 EVT ResVT = N->getValueType(0);
7304 EVT CmpVT = N0.getOperand(0).getValueType();
7305 // Only combine when the result type is of the same size as the compared
7307 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
7310 SDValue IfTrue = N->getOperand(1);
7311 SDValue IfFalse = N->getOperand(2);
7313 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
7314 N0.getOperand(0), N0.getOperand(1),
7315 cast<CondCodeSDNode>(N0.getOperand(2))->get());
7316 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
7320 SDValue ARM64TargetLowering::PerformDAGCombine(SDNode *N,
7321 DAGCombinerInfo &DCI) const {
7322 SelectionDAG &DAG = DCI.DAG;
7323 switch (N->getOpcode()) {
7328 return performAddSubLongCombine(N, DCI, DAG);
7330 return performXorCombine(N, DAG, DCI, Subtarget);
7332 return performMulCombine(N, DAG, DCI, Subtarget);
7333 case ISD::SINT_TO_FP:
7334 case ISD::UINT_TO_FP:
7335 return performIntToFpCombine(N, DAG);
7337 return performORCombine(N, DCI, Subtarget);
7338 case ISD::INTRINSIC_WO_CHAIN:
7339 return performIntrinsicCombine(N, DCI, Subtarget);
7340 case ISD::ANY_EXTEND:
7341 case ISD::ZERO_EXTEND:
7342 case ISD::SIGN_EXTEND:
7343 return performExtendCombine(N, DCI, DAG);
7345 return performBitcastCombine(N, DCI, DAG);
7346 case ISD::CONCAT_VECTORS:
7347 return performConcatVectorsCombine(N, DCI, DAG);
7349 return performVSelectCombine(N, DCI.DAG);
7351 return performSTORECombine(N, DCI, DAG, Subtarget);
7352 case ARM64ISD::BRCOND:
7353 return performBRCONDCombine(N, DCI, DAG);
7358 // Check if the return value is used as only a return value, as otherwise
7359 // we can't perform a tail-call. In particular, we need to check for
7360 // target ISD nodes that are returns and any other "odd" constructs
7361 // that the generic analysis code won't necessarily catch.
7362 bool ARM64TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
7363 if (N->getNumValues() != 1)
7365 if (!N->hasNUsesOfValue(1, 0))
7368 SDValue TCChain = Chain;
7369 SDNode *Copy = *N->use_begin();
7370 if (Copy->getOpcode() == ISD::CopyToReg) {
7371 // If the copy has a glue operand, we conservatively assume it isn't safe to
7372 // perform a tail call.
7373 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
7376 TCChain = Copy->getOperand(0);
7377 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
7380 bool HasRet = false;
7381 for (SDNode *Node : Copy->uses()) {
7382 if (Node->getOpcode() != ARM64ISD::RET_FLAG)
7394 // Return whether the an instruction can potentially be optimized to a tail
7395 // call. This will cause the optimizers to attempt to move, or duplicate,
7396 // return instructions to help enable tail call optimizations for this
7398 bool ARM64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
7399 if (!EnableARM64TailCalls)
7402 if (!CI->isTailCall())
7408 bool ARM64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
7410 ISD::MemIndexedMode &AM,
7412 SelectionDAG &DAG) const {
7413 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
7416 Base = Op->getOperand(0);
7417 // All of the indexed addressing mode instructions take a signed
7418 // 9 bit immediate offset.
7419 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
7420 int64_t RHSC = (int64_t)RHS->getZExtValue();
7421 if (RHSC >= 256 || RHSC <= -256)
7423 IsInc = (Op->getOpcode() == ISD::ADD);
7424 Offset = Op->getOperand(1);
7430 bool ARM64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
7432 ISD::MemIndexedMode &AM,
7433 SelectionDAG &DAG) const {
7436 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7437 VT = LD->getMemoryVT();
7438 Ptr = LD->getBasePtr();
7439 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7440 VT = ST->getMemoryVT();
7441 Ptr = ST->getBasePtr();
7446 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
7448 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
7452 bool ARM64TargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
7455 ISD::MemIndexedMode &AM,
7456 SelectionDAG &DAG) const {
7459 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7460 VT = LD->getMemoryVT();
7461 Ptr = LD->getBasePtr();
7462 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7463 VT = ST->getMemoryVT();
7464 Ptr = ST->getBasePtr();
7469 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
7471 // Post-indexing updates the base, so it's not a valid transform
7472 // if that's not the same as the load's pointer.
7475 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
7479 /// The only 128-bit atomic operation is an stxp that succeeds. In particular
7480 /// neither ldp nor ldxp are atomic. So the canonical sequence for an atomic
7483 /// ldxp x0, x1, [x8]
7484 /// stxp w2, x0, x1, [x8]
7486 /// If the stxp succeeds then the ldxp managed to get both halves without an
7487 /// intervening stxp from a different thread and the read was atomic.
7488 static void ReplaceATOMIC_LOAD_128(SDNode *N, SmallVectorImpl<SDValue> &Results,
7489 SelectionDAG &DAG) {
7491 AtomicSDNode *AN = cast<AtomicSDNode>(N);
7492 EVT VT = AN->getMemoryVT();
7493 SDValue Zero = DAG.getConstant(0, VT);
7495 // FIXME: Really want ATOMIC_LOAD_NOP but that doesn't fit into the existing
7496 // scheme very well. Given the complexity of what we're already generating, an
7497 // extra couple of ORRs probably won't make much difference.
7498 SDValue Result = DAG.getAtomic(ISD::ATOMIC_LOAD_OR, DL, AN->getMemoryVT(),
7499 N->getOperand(0), N->getOperand(1), Zero,
7500 AN->getMemOperand(), AN->getOrdering(),
7501 AN->getSynchScope());
7503 Results.push_back(Result.getValue(0)); // Value
7504 Results.push_back(Result.getValue(1)); // Chain
7507 static void ReplaceATOMIC_OP_128(SDNode *N, SmallVectorImpl<SDValue> &Results,
7508 SelectionDAG &DAG, unsigned NewOp) {
7510 AtomicOrdering Ordering = cast<AtomicSDNode>(N)->getOrdering();
7511 assert(N->getValueType(0) == MVT::i128 &&
7512 "Only know how to expand i128 atomics");
7514 SmallVector<SDValue, 6> Ops;
7515 Ops.push_back(N->getOperand(1)); // Ptr
7517 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7518 N->getOperand(2), DAG.getIntPtrConstant(0)));
7519 // High part of Val1
7520 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7521 N->getOperand(2), DAG.getIntPtrConstant(1)));
7522 if (NewOp == ARM64::ATOMIC_CMP_SWAP_I128) {
7524 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7525 N->getOperand(3), DAG.getIntPtrConstant(0)));
7526 // High part of Val2
7527 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7528 N->getOperand(3), DAG.getIntPtrConstant(1)));
7531 Ops.push_back(DAG.getTargetConstant(Ordering, MVT::i32));
7532 Ops.push_back(N->getOperand(0)); // Chain
7534 SDVTList Tys = DAG.getVTList(MVT::i64, MVT::i64, MVT::Other);
7535 SDNode *Result = DAG.getMachineNode(NewOp, DL, Tys, Ops);
7536 SDValue OpsF[] = { SDValue(Result, 0), SDValue(Result, 1) };
7537 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i128, OpsF, 2));
7538 Results.push_back(SDValue(Result, 2));
7541 void ARM64TargetLowering::ReplaceNodeResults(SDNode *N,
7542 SmallVectorImpl<SDValue> &Results,
7543 SelectionDAG &DAG) const {
7544 switch (N->getOpcode()) {
7546 llvm_unreachable("Don't know how to custom expand this");
7547 case ISD::ATOMIC_LOAD:
7548 ReplaceATOMIC_LOAD_128(N, Results, DAG);
7550 case ISD::ATOMIC_LOAD_ADD:
7551 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_ADD_I128);
7553 case ISD::ATOMIC_LOAD_SUB:
7554 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_SUB_I128);
7556 case ISD::ATOMIC_LOAD_AND:
7557 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_AND_I128);
7559 case ISD::ATOMIC_LOAD_OR:
7560 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_OR_I128);
7562 case ISD::ATOMIC_LOAD_XOR:
7563 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_XOR_I128);
7565 case ISD::ATOMIC_LOAD_NAND:
7566 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_NAND_I128);
7568 case ISD::ATOMIC_SWAP:
7569 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_SWAP_I128);
7571 case ISD::ATOMIC_LOAD_MIN:
7572 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_MIN_I128);
7574 case ISD::ATOMIC_LOAD_MAX:
7575 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_MAX_I128);
7577 case ISD::ATOMIC_LOAD_UMIN:
7578 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_UMIN_I128);
7580 case ISD::ATOMIC_LOAD_UMAX:
7581 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_UMAX_I128);
7583 case ISD::ATOMIC_CMP_SWAP:
7584 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_CMP_SWAP_I128);
7586 case ISD::FP_TO_UINT:
7587 case ISD::FP_TO_SINT:
7588 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
7589 // Let normal code take care of it by not adding anything to Results.