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:
556 case ISD::INTRINSIC_WO_CHAIN:
557 case ISD::INTRINSIC_VOID: {
558 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
562 case Intrinsic::arm64_neon_umaxv:
563 case Intrinsic::arm64_neon_uminv: {
564 // Figure out the datatype of the vector operand. The UMINV instruction
565 // will zero extend the result, so we can mark as known zero all the
566 // bits larger than the element datatype. 32-bit or larget doesn't need
567 // this as those are legal types and will be handled by isel directly.
568 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
569 unsigned BitWidth = KnownZero.getBitWidth();
570 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
571 assert(BitWidth >= 8 && "Unexpected width!");
572 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
574 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
575 assert(BitWidth >= 16 && "Unexpected width!");
576 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
586 MVT ARM64TargetLowering::getScalarShiftAmountTy(EVT LHSTy) const {
590 unsigned ARM64TargetLowering::getMaximalGlobalOffset() const {
591 // FIXME: On ARM64, this depends on the type.
592 // Basically, the addressable offsets are o to 4095 * Ty.getSizeInBytes().
593 // and the offset has to be a multiple of the related size in bytes.
598 ARM64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
599 const TargetLibraryInfo *libInfo) const {
600 return ARM64::createFastISel(funcInfo, libInfo);
603 const char *ARM64TargetLowering::getTargetNodeName(unsigned Opcode) const {
607 case ARM64ISD::CALL: return "ARM64ISD::CALL";
608 case ARM64ISD::ADRP: return "ARM64ISD::ADRP";
609 case ARM64ISD::ADDlow: return "ARM64ISD::ADDlow";
610 case ARM64ISD::LOADgot: return "ARM64ISD::LOADgot";
611 case ARM64ISD::RET_FLAG: return "ARM64ISD::RET_FLAG";
612 case ARM64ISD::BRCOND: return "ARM64ISD::BRCOND";
613 case ARM64ISD::CSEL: return "ARM64ISD::CSEL";
614 case ARM64ISD::FCSEL: return "ARM64ISD::FCSEL";
615 case ARM64ISD::CSINV: return "ARM64ISD::CSINV";
616 case ARM64ISD::CSNEG: return "ARM64ISD::CSNEG";
617 case ARM64ISD::CSINC: return "ARM64ISD::CSINC";
618 case ARM64ISD::THREAD_POINTER: return "ARM64ISD::THREAD_POINTER";
619 case ARM64ISD::TLSDESC_CALL: return "ARM64ISD::TLSDESC_CALL";
620 case ARM64ISD::ADC: return "ARM64ISD::ADC";
621 case ARM64ISD::SBC: return "ARM64ISD::SBC";
622 case ARM64ISD::ADDS: return "ARM64ISD::ADDS";
623 case ARM64ISD::SUBS: return "ARM64ISD::SUBS";
624 case ARM64ISD::ADCS: return "ARM64ISD::ADCS";
625 case ARM64ISD::SBCS: return "ARM64ISD::SBCS";
626 case ARM64ISD::ANDS: return "ARM64ISD::ANDS";
627 case ARM64ISD::FCMP: return "ARM64ISD::FCMP";
628 case ARM64ISD::FMIN: return "ARM64ISD::FMIN";
629 case ARM64ISD::FMAX: return "ARM64ISD::FMAX";
630 case ARM64ISD::DUP: return "ARM64ISD::DUP";
631 case ARM64ISD::DUPLANE8: return "ARM64ISD::DUPLANE8";
632 case ARM64ISD::DUPLANE16: return "ARM64ISD::DUPLANE16";
633 case ARM64ISD::DUPLANE32: return "ARM64ISD::DUPLANE32";
634 case ARM64ISD::DUPLANE64: return "ARM64ISD::DUPLANE64";
635 case ARM64ISD::MOVI: return "ARM64ISD::MOVI";
636 case ARM64ISD::MOVIshift: return "ARM64ISD::MOVIshift";
637 case ARM64ISD::MOVIedit: return "ARM64ISD::MOVIedit";
638 case ARM64ISD::MOVImsl: return "ARM64ISD::MOVImsl";
639 case ARM64ISD::FMOV: return "ARM64ISD::FMOV";
640 case ARM64ISD::MVNIshift: return "ARM64ISD::MVNIshift";
641 case ARM64ISD::MVNImsl: return "ARM64ISD::MVNImsl";
642 case ARM64ISD::BICi: return "ARM64ISD::BICi";
643 case ARM64ISD::ORRi: return "ARM64ISD::ORRi";
644 case ARM64ISD::NEG: return "ARM64ISD::NEG";
645 case ARM64ISD::EXTR: return "ARM64ISD::EXTR";
646 case ARM64ISD::ZIP1: return "ARM64ISD::ZIP1";
647 case ARM64ISD::ZIP2: return "ARM64ISD::ZIP2";
648 case ARM64ISD::UZP1: return "ARM64ISD::UZP1";
649 case ARM64ISD::UZP2: return "ARM64ISD::UZP2";
650 case ARM64ISD::TRN1: return "ARM64ISD::TRN1";
651 case ARM64ISD::TRN2: return "ARM64ISD::TRN2";
652 case ARM64ISD::REV16: return "ARM64ISD::REV16";
653 case ARM64ISD::REV32: return "ARM64ISD::REV32";
654 case ARM64ISD::REV64: return "ARM64ISD::REV64";
655 case ARM64ISD::EXT: return "ARM64ISD::EXT";
656 case ARM64ISD::VSHL: return "ARM64ISD::VSHL";
657 case ARM64ISD::VLSHR: return "ARM64ISD::VLSHR";
658 case ARM64ISD::VASHR: return "ARM64ISD::VASHR";
659 case ARM64ISD::CMEQ: return "ARM64ISD::CMEQ";
660 case ARM64ISD::CMGE: return "ARM64ISD::CMGE";
661 case ARM64ISD::CMGT: return "ARM64ISD::CMGT";
662 case ARM64ISD::CMHI: return "ARM64ISD::CMHI";
663 case ARM64ISD::CMHS: return "ARM64ISD::CMHS";
664 case ARM64ISD::FCMEQ: return "ARM64ISD::FCMEQ";
665 case ARM64ISD::FCMGE: return "ARM64ISD::FCMGE";
666 case ARM64ISD::FCMGT: return "ARM64ISD::FCMGT";
667 case ARM64ISD::CMEQz: return "ARM64ISD::CMEQz";
668 case ARM64ISD::CMGEz: return "ARM64ISD::CMGEz";
669 case ARM64ISD::CMGTz: return "ARM64ISD::CMGTz";
670 case ARM64ISD::CMLEz: return "ARM64ISD::CMLEz";
671 case ARM64ISD::CMLTz: return "ARM64ISD::CMLTz";
672 case ARM64ISD::FCMEQz: return "ARM64ISD::FCMEQz";
673 case ARM64ISD::FCMGEz: return "ARM64ISD::FCMGEz";
674 case ARM64ISD::FCMGTz: return "ARM64ISD::FCMGTz";
675 case ARM64ISD::FCMLEz: return "ARM64ISD::FCMLEz";
676 case ARM64ISD::FCMLTz: return "ARM64ISD::FCMLTz";
677 case ARM64ISD::NOT: return "ARM64ISD::NOT";
678 case ARM64ISD::BIT: return "ARM64ISD::BIT";
679 case ARM64ISD::CBZ: return "ARM64ISD::CBZ";
680 case ARM64ISD::CBNZ: return "ARM64ISD::CBNZ";
681 case ARM64ISD::TBZ: return "ARM64ISD::TBZ";
682 case ARM64ISD::TBNZ: return "ARM64ISD::TBNZ";
683 case ARM64ISD::TC_RETURN: return "ARM64ISD::TC_RETURN";
684 case ARM64ISD::SITOF: return "ARM64ISD::SITOF";
685 case ARM64ISD::UITOF: return "ARM64ISD::UITOF";
686 case ARM64ISD::SQSHL_I: return "ARM64ISD::SQSHL_I";
687 case ARM64ISD::UQSHL_I: return "ARM64ISD::UQSHL_I";
688 case ARM64ISD::SRSHR_I: return "ARM64ISD::SRSHR_I";
689 case ARM64ISD::URSHR_I: return "ARM64ISD::URSHR_I";
690 case ARM64ISD::SQSHLU_I: return "ARM64ISD::SQSHLU_I";
691 case ARM64ISD::WrapperLarge: return "ARM64ISD::WrapperLarge";
695 static void getExclusiveOperation(unsigned Size, AtomicOrdering Ord,
696 unsigned &LdrOpc, unsigned &StrOpc) {
697 static unsigned LoadBares[] = { ARM64::LDXRB, ARM64::LDXRH, ARM64::LDXRW,
698 ARM64::LDXRX, ARM64::LDXPX };
699 static unsigned LoadAcqs[] = { ARM64::LDAXRB, ARM64::LDAXRH, ARM64::LDAXRW,
700 ARM64::LDAXRX, ARM64::LDAXPX };
701 static unsigned StoreBares[] = { ARM64::STXRB, ARM64::STXRH, ARM64::STXRW,
702 ARM64::STXRX, ARM64::STXPX };
703 static unsigned StoreRels[] = { ARM64::STLXRB, ARM64::STLXRH, ARM64::STLXRW,
704 ARM64::STLXRX, ARM64::STLXPX };
706 unsigned *LoadOps, *StoreOps;
707 if (Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent)
712 if (Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent)
713 StoreOps = StoreRels;
715 StoreOps = StoreBares;
717 assert(isPowerOf2_32(Size) && Size <= 16 &&
718 "unsupported size for atomic binary op!");
720 LdrOpc = LoadOps[Log2_32(Size)];
721 StrOpc = StoreOps[Log2_32(Size)];
724 MachineBasicBlock *ARM64TargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
725 MachineBasicBlock *BB,
726 unsigned Size) const {
727 unsigned dest = MI->getOperand(0).getReg();
728 unsigned ptr = MI->getOperand(1).getReg();
729 unsigned oldval = MI->getOperand(2).getReg();
730 unsigned newval = MI->getOperand(3).getReg();
731 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(4).getImm());
732 unsigned scratch = BB->getParent()->getRegInfo().createVirtualRegister(
733 &ARM64::GPR32RegClass);
734 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
735 DebugLoc dl = MI->getDebugLoc();
737 // FIXME: We currently always generate a seq_cst operation; we should
738 // be able to relax this in some cases.
739 unsigned ldrOpc, strOpc;
740 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
742 MachineFunction *MF = BB->getParent();
743 const BasicBlock *LLVM_BB = BB->getBasicBlock();
744 MachineFunction::iterator It = BB;
745 ++It; // insert the new blocks after the current block
747 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
748 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
749 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
750 MF->insert(It, loop1MBB);
751 MF->insert(It, loop2MBB);
752 MF->insert(It, exitMBB);
754 // Transfer the remainder of BB and its successor edges to exitMBB.
755 exitMBB->splice(exitMBB->begin(), BB,
756 std::next(MachineBasicBlock::iterator(MI)), BB->end());
757 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
761 // fallthrough --> loop1MBB
762 BB->addSuccessor(loop1MBB);
769 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
770 BuildMI(BB, dl, TII->get(Size == 8 ? ARM64::SUBSXrr : ARM64::SUBSWrr))
771 .addReg(Size == 8 ? ARM64::XZR : ARM64::WZR, RegState::Define)
774 BuildMI(BB, dl, TII->get(ARM64::Bcc)).addImm(ARM64CC::NE).addMBB(exitMBB);
775 BB->addSuccessor(loop2MBB);
776 BB->addSuccessor(exitMBB);
779 // strex scratch, newval, [ptr]
783 BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
784 BuildMI(BB, dl, TII->get(ARM64::CBNZW)).addReg(scratch).addMBB(loop1MBB);
785 BB->addSuccessor(loop1MBB);
786 BB->addSuccessor(exitMBB);
792 MI->eraseFromParent(); // The instruction is gone now.
798 ARM64TargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
799 unsigned Size, unsigned BinOpcode) const {
800 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
801 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
803 const BasicBlock *LLVM_BB = BB->getBasicBlock();
804 MachineFunction *MF = BB->getParent();
805 MachineFunction::iterator It = BB;
808 unsigned dest = MI->getOperand(0).getReg();
809 unsigned ptr = MI->getOperand(1).getReg();
810 unsigned incr = MI->getOperand(2).getReg();
811 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
812 DebugLoc dl = MI->getDebugLoc();
814 unsigned ldrOpc, strOpc;
815 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
817 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
818 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
819 MF->insert(It, loopMBB);
820 MF->insert(It, exitMBB);
822 // Transfer the remainder of BB and its successor edges to exitMBB.
823 exitMBB->splice(exitMBB->begin(), BB,
824 std::next(MachineBasicBlock::iterator(MI)), BB->end());
825 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
827 MachineRegisterInfo &RegInfo = MF->getRegInfo();
828 unsigned scratch = RegInfo.createVirtualRegister(&ARM64::GPR32RegClass);
832 : RegInfo.createVirtualRegister(Size == 8 ? &ARM64::GPR64RegClass
833 : &ARM64::GPR32RegClass);
837 // fallthrough --> loopMBB
838 BB->addSuccessor(loopMBB);
842 // <binop> scratch2, dest, incr
843 // stxr scratch, scratch2, ptr
844 // cbnz scratch, loopMBB
845 // fallthrough --> exitMBB
847 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
849 // operand order needs to go the other way for NAND
850 if (BinOpcode == ARM64::BICWrr || BinOpcode == ARM64::BICXrr)
851 BuildMI(BB, dl, TII->get(BinOpcode), scratch2).addReg(incr).addReg(dest);
853 BuildMI(BB, dl, TII->get(BinOpcode), scratch2).addReg(dest).addReg(incr);
856 BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
857 BuildMI(BB, dl, TII->get(ARM64::CBNZW)).addReg(scratch).addMBB(loopMBB);
859 BB->addSuccessor(loopMBB);
860 BB->addSuccessor(exitMBB);
866 MI->eraseFromParent(); // The instruction is gone now.
871 MachineBasicBlock *ARM64TargetLowering::EmitAtomicBinary128(
872 MachineInstr *MI, MachineBasicBlock *BB, unsigned BinOpcodeLo,
873 unsigned BinOpcodeHi) const {
874 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
875 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
877 const BasicBlock *LLVM_BB = BB->getBasicBlock();
878 MachineFunction *MF = BB->getParent();
879 MachineFunction::iterator It = BB;
882 unsigned DestLo = MI->getOperand(0).getReg();
883 unsigned DestHi = MI->getOperand(1).getReg();
884 unsigned Ptr = MI->getOperand(2).getReg();
885 unsigned IncrLo = MI->getOperand(3).getReg();
886 unsigned IncrHi = MI->getOperand(4).getReg();
887 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(5).getImm());
888 DebugLoc DL = MI->getDebugLoc();
890 unsigned LdrOpc, StrOpc;
891 getExclusiveOperation(16, Ord, LdrOpc, StrOpc);
893 MachineBasicBlock *LoopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
894 MachineBasicBlock *ExitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
895 MF->insert(It, LoopMBB);
896 MF->insert(It, ExitMBB);
898 // Transfer the remainder of BB and its successor edges to exitMBB.
899 ExitMBB->splice(ExitMBB->begin(), BB,
900 std::next(MachineBasicBlock::iterator(MI)), BB->end());
901 ExitMBB->transferSuccessorsAndUpdatePHIs(BB);
903 MachineRegisterInfo &RegInfo = MF->getRegInfo();
904 unsigned ScratchRes = RegInfo.createVirtualRegister(&ARM64::GPR32RegClass);
905 unsigned ScratchLo = IncrLo, ScratchHi = IncrHi;
907 assert(BinOpcodeHi && "Expect neither or both opcodes to be defined");
908 ScratchLo = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
909 ScratchHi = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
914 // fallthrough --> LoopMBB
915 BB->addSuccessor(LoopMBB);
918 // ldxp DestLo, DestHi, Ptr
919 // <binoplo> ScratchLo, DestLo, IncrLo
920 // <binophi> ScratchHi, DestHi, IncrHi
921 // stxp ScratchRes, ScratchLo, ScratchHi, ptr
922 // cbnz ScratchRes, LoopMBB
923 // fallthrough --> ExitMBB
925 BuildMI(BB, DL, TII->get(LdrOpc), DestLo)
926 .addReg(DestHi, RegState::Define)
929 // operand order needs to go the other way for NAND
930 if (BinOpcodeLo == ARM64::BICXrr) {
931 std::swap(IncrLo, DestLo);
932 std::swap(IncrHi, DestHi);
935 BuildMI(BB, DL, TII->get(BinOpcodeLo), ScratchLo).addReg(DestLo).addReg(
937 BuildMI(BB, DL, TII->get(BinOpcodeHi), ScratchHi).addReg(DestHi).addReg(
941 BuildMI(BB, DL, TII->get(StrOpc), ScratchRes)
945 BuildMI(BB, DL, TII->get(ARM64::CBNZW)).addReg(ScratchRes).addMBB(LoopMBB);
947 BB->addSuccessor(LoopMBB);
948 BB->addSuccessor(ExitMBB);
954 MI->eraseFromParent(); // The instruction is gone now.
960 ARM64TargetLowering::EmitAtomicCmpSwap128(MachineInstr *MI,
961 MachineBasicBlock *BB) const {
962 unsigned DestLo = MI->getOperand(0).getReg();
963 unsigned DestHi = MI->getOperand(1).getReg();
964 unsigned Ptr = MI->getOperand(2).getReg();
965 unsigned OldValLo = MI->getOperand(3).getReg();
966 unsigned OldValHi = MI->getOperand(4).getReg();
967 unsigned NewValLo = MI->getOperand(5).getReg();
968 unsigned NewValHi = MI->getOperand(6).getReg();
969 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(7).getImm());
970 unsigned ScratchRes = BB->getParent()->getRegInfo().createVirtualRegister(
971 &ARM64::GPR32RegClass);
972 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
973 DebugLoc DL = MI->getDebugLoc();
975 unsigned LdrOpc, StrOpc;
976 getExclusiveOperation(16, Ord, LdrOpc, StrOpc);
978 MachineFunction *MF = BB->getParent();
979 const BasicBlock *LLVM_BB = BB->getBasicBlock();
980 MachineFunction::iterator It = BB;
981 ++It; // insert the new blocks after the current block
983 MachineBasicBlock *Loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
984 MachineBasicBlock *Loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
985 MachineBasicBlock *ExitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
986 MF->insert(It, Loop1MBB);
987 MF->insert(It, Loop2MBB);
988 MF->insert(It, ExitMBB);
990 // Transfer the remainder of BB and its successor edges to exitMBB.
991 ExitMBB->splice(ExitMBB->begin(), BB,
992 std::next(MachineBasicBlock::iterator(MI)), BB->end());
993 ExitMBB->transferSuccessorsAndUpdatePHIs(BB);
997 // fallthrough --> Loop1MBB
998 BB->addSuccessor(Loop1MBB);
1001 // ldxp DestLo, DestHi, [Ptr]
1002 // cmp DestLo, OldValLo
1003 // sbc xzr, DestHi, OldValHi
1006 BuildMI(BB, DL, TII->get(LdrOpc), DestLo)
1007 .addReg(DestHi, RegState::Define)
1009 BuildMI(BB, DL, TII->get(ARM64::SUBSXrr), ARM64::XZR).addReg(DestLo).addReg(
1011 BuildMI(BB, DL, TII->get(ARM64::SBCXr), ARM64::XZR).addReg(DestHi).addReg(
1014 BuildMI(BB, DL, TII->get(ARM64::Bcc)).addImm(ARM64CC::NE).addMBB(ExitMBB);
1015 BB->addSuccessor(Loop2MBB);
1016 BB->addSuccessor(ExitMBB);
1019 // stxp ScratchRes, NewValLo, NewValHi, [Ptr]
1020 // cbnz ScratchRes, Loop1MBB
1022 BuildMI(BB, DL, TII->get(StrOpc), ScratchRes)
1026 BuildMI(BB, DL, TII->get(ARM64::CBNZW)).addReg(ScratchRes).addMBB(Loop1MBB);
1027 BB->addSuccessor(Loop1MBB);
1028 BB->addSuccessor(ExitMBB);
1034 MI->eraseFromParent(); // The instruction is gone now.
1039 MachineBasicBlock *ARM64TargetLowering::EmitAtomicMinMax128(
1040 MachineInstr *MI, MachineBasicBlock *BB, unsigned CondCode) const {
1041 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
1042 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1044 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1045 MachineFunction *MF = BB->getParent();
1046 MachineFunction::iterator It = BB;
1049 unsigned DestLo = MI->getOperand(0).getReg();
1050 unsigned DestHi = MI->getOperand(1).getReg();
1051 unsigned Ptr = MI->getOperand(2).getReg();
1052 unsigned IncrLo = MI->getOperand(3).getReg();
1053 unsigned IncrHi = MI->getOperand(4).getReg();
1054 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(5).getImm());
1055 DebugLoc DL = MI->getDebugLoc();
1057 unsigned LdrOpc, StrOpc;
1058 getExclusiveOperation(16, Ord, LdrOpc, StrOpc);
1060 MachineBasicBlock *LoopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1061 MachineBasicBlock *ExitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1062 MF->insert(It, LoopMBB);
1063 MF->insert(It, ExitMBB);
1065 // Transfer the remainder of BB and its successor edges to exitMBB.
1066 ExitMBB->splice(ExitMBB->begin(), BB,
1067 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1068 ExitMBB->transferSuccessorsAndUpdatePHIs(BB);
1070 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1071 unsigned ScratchRes = RegInfo.createVirtualRegister(&ARM64::GPR32RegClass);
1072 unsigned ScratchLo = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
1073 unsigned ScratchHi = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
1077 // fallthrough --> LoopMBB
1078 BB->addSuccessor(LoopMBB);
1081 // ldxp DestLo, DestHi, Ptr
1082 // cmp ScratchLo, DestLo, IncrLo
1083 // sbc xzr, ScratchHi, DestHi, IncrHi
1084 // csel ScratchLo, DestLo, IncrLo, <cmp-op>
1085 // csel ScratchHi, DestHi, IncrHi, <cmp-op>
1086 // stxp ScratchRes, ScratchLo, ScratchHi, ptr
1087 // cbnz ScratchRes, LoopMBB
1088 // fallthrough --> ExitMBB
1090 BuildMI(BB, DL, TII->get(LdrOpc), DestLo)
1091 .addReg(DestHi, RegState::Define)
1094 BuildMI(BB, DL, TII->get(ARM64::SUBSXrr), ARM64::XZR).addReg(DestLo).addReg(
1096 BuildMI(BB, DL, TII->get(ARM64::SBCXr), ARM64::XZR).addReg(DestHi).addReg(
1099 BuildMI(BB, DL, TII->get(ARM64::CSELXr), ScratchLo)
1103 BuildMI(BB, DL, TII->get(ARM64::CSELXr), ScratchHi)
1108 BuildMI(BB, DL, TII->get(StrOpc), ScratchRes)
1112 BuildMI(BB, DL, TII->get(ARM64::CBNZW)).addReg(ScratchRes).addMBB(LoopMBB);
1114 BB->addSuccessor(LoopMBB);
1115 BB->addSuccessor(ExitMBB);
1121 MI->eraseFromParent(); // The instruction is gone now.
1127 ARM64TargetLowering::EmitF128CSEL(MachineInstr *MI,
1128 MachineBasicBlock *MBB) const {
1129 // We materialise the F128CSEL pseudo-instruction as some control flow and a
1133 // [... previous instrs leading to comparison ...]
1139 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
1141 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1142 MachineFunction *MF = MBB->getParent();
1143 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
1144 DebugLoc DL = MI->getDebugLoc();
1145 MachineFunction::iterator It = MBB;
1148 unsigned DestReg = MI->getOperand(0).getReg();
1149 unsigned IfTrueReg = MI->getOperand(1).getReg();
1150 unsigned IfFalseReg = MI->getOperand(2).getReg();
1151 unsigned CondCode = MI->getOperand(3).getImm();
1152 bool CPSRKilled = MI->getOperand(4).isKill();
1154 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
1155 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
1156 MF->insert(It, TrueBB);
1157 MF->insert(It, EndBB);
1159 // Transfer rest of current basic-block to EndBB
1160 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
1162 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
1164 BuildMI(MBB, DL, TII->get(ARM64::Bcc)).addImm(CondCode).addMBB(TrueBB);
1165 BuildMI(MBB, DL, TII->get(ARM64::B)).addMBB(EndBB);
1166 MBB->addSuccessor(TrueBB);
1167 MBB->addSuccessor(EndBB);
1169 // TrueBB falls through to the end.
1170 TrueBB->addSuccessor(EndBB);
1173 TrueBB->addLiveIn(ARM64::CPSR);
1174 EndBB->addLiveIn(ARM64::CPSR);
1177 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(ARM64::PHI), DestReg)
1183 MI->eraseFromParent();
1188 ARM64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
1189 MachineBasicBlock *BB) const {
1190 switch (MI->getOpcode()) {
1195 assert(0 && "Unexpected instruction for custom inserter!");
1198 case ARM64::ATOMIC_LOAD_ADD_I8:
1199 return EmitAtomicBinary(MI, BB, 1, ARM64::ADDWrr);
1200 case ARM64::ATOMIC_LOAD_ADD_I16:
1201 return EmitAtomicBinary(MI, BB, 2, ARM64::ADDWrr);
1202 case ARM64::ATOMIC_LOAD_ADD_I32:
1203 return EmitAtomicBinary(MI, BB, 4, ARM64::ADDWrr);
1204 case ARM64::ATOMIC_LOAD_ADD_I64:
1205 return EmitAtomicBinary(MI, BB, 8, ARM64::ADDXrr);
1206 case ARM64::ATOMIC_LOAD_ADD_I128:
1207 return EmitAtomicBinary128(MI, BB, ARM64::ADDSXrr, ARM64::ADCXr);
1209 case ARM64::ATOMIC_LOAD_AND_I8:
1210 return EmitAtomicBinary(MI, BB, 1, ARM64::ANDWrr);
1211 case ARM64::ATOMIC_LOAD_AND_I16:
1212 return EmitAtomicBinary(MI, BB, 2, ARM64::ANDWrr);
1213 case ARM64::ATOMIC_LOAD_AND_I32:
1214 return EmitAtomicBinary(MI, BB, 4, ARM64::ANDWrr);
1215 case ARM64::ATOMIC_LOAD_AND_I64:
1216 return EmitAtomicBinary(MI, BB, 8, ARM64::ANDXrr);
1217 case ARM64::ATOMIC_LOAD_AND_I128:
1218 return EmitAtomicBinary128(MI, BB, ARM64::ANDXrr, ARM64::ANDXrr);
1220 case ARM64::ATOMIC_LOAD_OR_I8:
1221 return EmitAtomicBinary(MI, BB, 1, ARM64::ORRWrr);
1222 case ARM64::ATOMIC_LOAD_OR_I16:
1223 return EmitAtomicBinary(MI, BB, 2, ARM64::ORRWrr);
1224 case ARM64::ATOMIC_LOAD_OR_I32:
1225 return EmitAtomicBinary(MI, BB, 4, ARM64::ORRWrr);
1226 case ARM64::ATOMIC_LOAD_OR_I64:
1227 return EmitAtomicBinary(MI, BB, 8, ARM64::ORRXrr);
1228 case ARM64::ATOMIC_LOAD_OR_I128:
1229 return EmitAtomicBinary128(MI, BB, ARM64::ORRXrr, ARM64::ORRXrr);
1231 case ARM64::ATOMIC_LOAD_XOR_I8:
1232 return EmitAtomicBinary(MI, BB, 1, ARM64::EORWrr);
1233 case ARM64::ATOMIC_LOAD_XOR_I16:
1234 return EmitAtomicBinary(MI, BB, 2, ARM64::EORWrr);
1235 case ARM64::ATOMIC_LOAD_XOR_I32:
1236 return EmitAtomicBinary(MI, BB, 4, ARM64::EORWrr);
1237 case ARM64::ATOMIC_LOAD_XOR_I64:
1238 return EmitAtomicBinary(MI, BB, 8, ARM64::EORXrr);
1239 case ARM64::ATOMIC_LOAD_XOR_I128:
1240 return EmitAtomicBinary128(MI, BB, ARM64::EORXrr, ARM64::EORXrr);
1242 case ARM64::ATOMIC_LOAD_NAND_I8:
1243 return EmitAtomicBinary(MI, BB, 1, ARM64::BICWrr);
1244 case ARM64::ATOMIC_LOAD_NAND_I16:
1245 return EmitAtomicBinary(MI, BB, 2, ARM64::BICWrr);
1246 case ARM64::ATOMIC_LOAD_NAND_I32:
1247 return EmitAtomicBinary(MI, BB, 4, ARM64::BICWrr);
1248 case ARM64::ATOMIC_LOAD_NAND_I64:
1249 return EmitAtomicBinary(MI, BB, 8, ARM64::BICXrr);
1250 case ARM64::ATOMIC_LOAD_NAND_I128:
1251 return EmitAtomicBinary128(MI, BB, ARM64::BICXrr, ARM64::BICXrr);
1253 case ARM64::ATOMIC_LOAD_SUB_I8:
1254 return EmitAtomicBinary(MI, BB, 1, ARM64::SUBWrr);
1255 case ARM64::ATOMIC_LOAD_SUB_I16:
1256 return EmitAtomicBinary(MI, BB, 2, ARM64::SUBWrr);
1257 case ARM64::ATOMIC_LOAD_SUB_I32:
1258 return EmitAtomicBinary(MI, BB, 4, ARM64::SUBWrr);
1259 case ARM64::ATOMIC_LOAD_SUB_I64:
1260 return EmitAtomicBinary(MI, BB, 8, ARM64::SUBXrr);
1261 case ARM64::ATOMIC_LOAD_SUB_I128:
1262 return EmitAtomicBinary128(MI, BB, ARM64::SUBSXrr, ARM64::SBCXr);
1264 case ARM64::ATOMIC_LOAD_MIN_I128:
1265 return EmitAtomicMinMax128(MI, BB, ARM64CC::LT);
1267 case ARM64::ATOMIC_LOAD_MAX_I128:
1268 return EmitAtomicMinMax128(MI, BB, ARM64CC::GT);
1270 case ARM64::ATOMIC_LOAD_UMIN_I128:
1271 return EmitAtomicMinMax128(MI, BB, ARM64CC::CC);
1273 case ARM64::ATOMIC_LOAD_UMAX_I128:
1274 return EmitAtomicMinMax128(MI, BB, ARM64CC::HI);
1276 case ARM64::ATOMIC_SWAP_I8:
1277 return EmitAtomicBinary(MI, BB, 1, 0);
1278 case ARM64::ATOMIC_SWAP_I16:
1279 return EmitAtomicBinary(MI, BB, 2, 0);
1280 case ARM64::ATOMIC_SWAP_I32:
1281 return EmitAtomicBinary(MI, BB, 4, 0);
1282 case ARM64::ATOMIC_SWAP_I64:
1283 return EmitAtomicBinary(MI, BB, 8, 0);
1284 case ARM64::ATOMIC_SWAP_I128:
1285 return EmitAtomicBinary128(MI, BB, 0, 0);
1287 case ARM64::ATOMIC_CMP_SWAP_I8:
1288 return EmitAtomicCmpSwap(MI, BB, 1);
1289 case ARM64::ATOMIC_CMP_SWAP_I16:
1290 return EmitAtomicCmpSwap(MI, BB, 2);
1291 case ARM64::ATOMIC_CMP_SWAP_I32:
1292 return EmitAtomicCmpSwap(MI, BB, 4);
1293 case ARM64::ATOMIC_CMP_SWAP_I64:
1294 return EmitAtomicCmpSwap(MI, BB, 8);
1295 case ARM64::ATOMIC_CMP_SWAP_I128:
1296 return EmitAtomicCmpSwap128(MI, BB);
1298 case ARM64::F128CSEL:
1299 return EmitF128CSEL(MI, BB);
1301 case TargetOpcode::STACKMAP:
1302 case TargetOpcode::PATCHPOINT:
1303 return emitPatchPoint(MI, BB);
1305 llvm_unreachable("Unexpected instruction for custom inserter!");
1308 //===----------------------------------------------------------------------===//
1309 // ARM64 Lowering private implementation.
1310 //===----------------------------------------------------------------------===//
1312 //===----------------------------------------------------------------------===//
1314 //===----------------------------------------------------------------------===//
1316 /// changeIntCCToARM64CC - Convert a DAG integer condition code to an ARM64 CC
1317 static ARM64CC::CondCode changeIntCCToARM64CC(ISD::CondCode CC) {
1320 llvm_unreachable("Unknown condition code!");
1344 /// changeFPCCToARM64CC - Convert a DAG fp condition code to an ARM64 CC.
1345 static void changeFPCCToARM64CC(ISD::CondCode CC, ARM64CC::CondCode &CondCode,
1346 ARM64CC::CondCode &CondCode2) {
1347 CondCode2 = ARM64CC::AL;
1350 llvm_unreachable("Unknown FP condition!");
1353 CondCode = ARM64CC::EQ;
1357 CondCode = ARM64CC::GT;
1361 CondCode = ARM64CC::GE;
1364 CondCode = ARM64CC::MI;
1367 CondCode = ARM64CC::LS;
1370 CondCode = ARM64CC::MI;
1371 CondCode2 = ARM64CC::GT;
1374 CondCode = ARM64CC::VC;
1377 CondCode = ARM64CC::VS;
1380 CondCode = ARM64CC::EQ;
1381 CondCode2 = ARM64CC::VS;
1384 CondCode = ARM64CC::HI;
1387 CondCode = ARM64CC::PL;
1391 CondCode = ARM64CC::LT;
1395 CondCode = ARM64CC::LE;
1399 CondCode = ARM64CC::NE;
1404 static bool isLegalArithImmed(uint64_t C) {
1405 // Matches ARM64DAGToDAGISel::SelectArithImmed().
1406 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
1409 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1410 SDLoc dl, SelectionDAG &DAG) {
1411 EVT VT = LHS.getValueType();
1413 if (VT.isFloatingPoint())
1414 return DAG.getNode(ARM64ISD::FCMP, dl, VT, LHS, RHS);
1416 // The CMP instruction is just an alias for SUBS, and representing it as
1417 // SUBS means that it's possible to get CSE with subtract operations.
1418 // A later phase can perform the optimization of setting the destination
1419 // register to WZR/XZR if it ends up being unused.
1421 // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on the
1422 // grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags can be
1423 // set differently by this operation. It comes down to whether "SInt(~op2)+1
1424 // == SInt(~op2+1)" (and the same for UInt). If they are then everything is
1425 // fine. If not then the optimization is wrong. Thus general comparisons are
1426 // only valid if op2 != 0.
1428 // So, finally, the only LLVM-native comparisons that don't mention C and V
1429 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in the
1430 // absence of information about op2.
1431 unsigned Opcode = ARM64ISD::SUBS;
1432 if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
1433 cast<ConstantSDNode>(RHS.getOperand(0))->getZExtValue() == 0 &&
1434 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1435 Opcode = ARM64ISD::ADDS;
1436 RHS = RHS.getOperand(1);
1439 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS)
1443 static SDValue getARM64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1444 SDValue &ARM64cc, SelectionDAG &DAG, SDLoc dl) {
1445 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1446 EVT VT = RHS.getValueType();
1447 uint64_t C = RHSC->getZExtValue();
1448 if (!isLegalArithImmed(C)) {
1449 // Constant does not fit, try adjusting it by one?
1455 if ((VT == MVT::i32 && C != 0x80000000 &&
1456 isLegalArithImmed((uint32_t)(C - 1))) ||
1457 (VT == MVT::i64 && C != 0x80000000ULL &&
1458 isLegalArithImmed(C - 1ULL))) {
1459 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1460 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1461 RHS = DAG.getConstant(C, VT);
1466 if ((VT == MVT::i32 && C != 0 &&
1467 isLegalArithImmed((uint32_t)(C - 1))) ||
1468 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1469 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1470 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1471 RHS = DAG.getConstant(C, VT);
1476 if ((VT == MVT::i32 && C != 0x7fffffff &&
1477 isLegalArithImmed((uint32_t)(C + 1))) ||
1478 (VT == MVT::i64 && C != 0x7ffffffffffffffULL &&
1479 isLegalArithImmed(C + 1ULL))) {
1480 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1481 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1482 RHS = DAG.getConstant(C, VT);
1487 if ((VT == MVT::i32 && C != 0xffffffff &&
1488 isLegalArithImmed((uint32_t)(C + 1))) ||
1489 (VT == MVT::i64 && C != 0xfffffffffffffffULL &&
1490 isLegalArithImmed(C + 1ULL))) {
1491 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1492 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1493 RHS = DAG.getConstant(C, VT);
1500 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
1501 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
1502 ARM64cc = DAG.getConstant(ARM64CC, MVT::i32);
1506 static std::pair<SDValue, SDValue>
1507 getARM64XALUOOp(ARM64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1508 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1509 "Unsupported value type");
1510 SDValue Value, Overflow;
1512 SDValue LHS = Op.getOperand(0);
1513 SDValue RHS = Op.getOperand(1);
1515 switch (Op.getOpcode()) {
1517 llvm_unreachable("Unknown overflow instruction!");
1519 Opc = ARM64ISD::ADDS;
1523 Opc = ARM64ISD::ADDS;
1527 Opc = ARM64ISD::SUBS;
1531 Opc = ARM64ISD::SUBS;
1534 // Multiply needs a little bit extra work.
1538 bool IsSigned = (Op.getOpcode() == ISD::SMULO) ? true : false;
1539 if (Op.getValueType() == MVT::i32) {
1540 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1541 // For a 32 bit multiply with overflow check we want the instruction
1542 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1543 // need to generate the following pattern:
1544 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1545 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1546 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1547 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1548 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1549 DAG.getConstant(0, MVT::i64));
1550 // On ARM64 the upper 32 bits are always zero extended for a 32 bit
1551 // operation. We need to clear out the upper 32 bits, because we used a
1552 // widening multiply that wrote all 64 bits. In the end this should be a
1554 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1556 // The signed overflow check requires more than just a simple check for
1557 // any bit set in the upper 32 bits of the result. These bits could be
1558 // just the sign bits of a negative number. To perform the overflow
1559 // check we have to arithmetic shift right the 32nd bit of the result by
1560 // 31 bits. Then we compare the result to the upper 32 bits.
1561 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1562 DAG.getConstant(32, MVT::i64));
1563 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1564 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1565 DAG.getConstant(31, MVT::i64));
1566 // It is important that LowerBits is last, otherwise the arithmetic
1567 // shift will not be folded into the compare (SUBS).
1568 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1569 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1572 // The overflow check for unsigned multiply is easy. We only need to
1573 // check if any of the upper 32 bits are set. This can be done with a
1574 // CMP (shifted register). For that we need to generate the following
1576 // (i64 ARM64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1577 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1578 DAG.getConstant(32, MVT::i64));
1579 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1581 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1582 UpperBits).getValue(1);
1586 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1587 // For the 64 bit multiply
1588 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1590 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1591 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1592 DAG.getConstant(63, MVT::i64));
1593 // It is important that LowerBits is last, otherwise the arithmetic
1594 // shift will not be folded into the compare (SUBS).
1595 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1596 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1599 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1600 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1602 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1603 UpperBits).getValue(1);
1610 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1612 // Emit the ARM64 operation with overflow check.
1613 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1614 Overflow = Value.getValue(1);
1616 return std::make_pair(Value, Overflow);
1619 SDValue ARM64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1620 RTLIB::Libcall Call) const {
1621 SmallVector<SDValue, 2> Ops;
1622 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1623 Ops.push_back(Op.getOperand(i));
1625 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1629 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1630 SDValue Sel = Op.getOperand(0);
1631 SDValue Other = Op.getOperand(1);
1633 // If neither operand is a SELECT_CC, give up.
1634 if (Sel.getOpcode() != ISD::SELECT_CC)
1635 std::swap(Sel, Other);
1636 if (Sel.getOpcode() != ISD::SELECT_CC)
1639 // The folding we want to perform is:
1640 // (xor x, (select_cc a, b, cc, 0, -1) )
1642 // (csel x, (xor x, -1), cc ...)
1644 // The latter will get matched to a CSINV instruction.
1646 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1647 SDValue LHS = Sel.getOperand(0);
1648 SDValue RHS = Sel.getOperand(1);
1649 SDValue TVal = Sel.getOperand(2);
1650 SDValue FVal = Sel.getOperand(3);
1653 // FIXME: This could be generalized to non-integer comparisons.
1654 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1657 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1658 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1660 // The the values aren't constants, this isn't the pattern we're looking for.
1661 if (!CFVal || !CTVal)
1664 // We can commute the SELECT_CC by inverting the condition. This
1665 // might be needed to make this fit into a CSINV pattern.
1666 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1667 std::swap(TVal, FVal);
1668 std::swap(CTVal, CFVal);
1669 CC = ISD::getSetCCInverse(CC, true);
1672 // If the constants line up, perform the transform!
1673 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1675 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1678 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1679 DAG.getConstant(-1ULL, Other.getValueType()));
1681 return DAG.getNode(ARM64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1688 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1689 EVT VT = Op.getValueType();
1691 // Let legalize expand this if it isn't a legal type yet.
1692 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1695 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1698 bool ExtraOp = false;
1699 switch (Op.getOpcode()) {
1701 assert(0 && "Invalid code");
1703 Opc = ARM64ISD::ADDS;
1706 Opc = ARM64ISD::SUBS;
1709 Opc = ARM64ISD::ADCS;
1713 Opc = ARM64ISD::SBCS;
1719 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1720 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1724 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1725 // Let legalize expand this if it isn't a legal type yet.
1726 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1729 ARM64CC::CondCode CC;
1730 // The actual operation that sets the overflow or carry flag.
1731 SDValue Value, Overflow;
1732 std::tie(Value, Overflow) = getARM64XALUOOp(CC, Op, DAG);
1734 // We use 0 and 1 as false and true values.
1735 SDValue TVal = DAG.getConstant(1, MVT::i32);
1736 SDValue FVal = DAG.getConstant(0, MVT::i32);
1738 // We use an inverted condition, because the conditional select is inverted
1739 // too. This will allow it to be selected to a single instruction:
1740 // CSINC Wd, WZR, WZR, invert(cond).
1741 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), MVT::i32);
1742 Overflow = DAG.getNode(ARM64ISD::CSEL, SDLoc(Op), MVT::i32, FVal, TVal, CCVal,
1745 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1746 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
1749 // Prefetch operands are:
1750 // 1: Address to prefetch
1752 // 3: int locality (0 = no locality ... 3 = extreme locality)
1753 // 4: bool isDataCache
1754 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1756 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1757 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1758 // The data thing is not used.
1759 // unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1761 bool IsStream = !Locality;
1762 // When the locality number is set
1764 // The front-end should have filtered out the out-of-range values
1765 assert(Locality <= 3 && "Prefetch locality out-of-range");
1766 // The locality degree is the opposite of the cache speed.
1767 // Put the number the other way around.
1768 // The encoding starts at 0 for level 1
1769 Locality = 3 - Locality;
1772 // built the mask value encoding the expected behavior.
1773 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1774 (Locality << 1) | // Cache level bits
1775 (unsigned)IsStream; // Stream bit
1776 return DAG.getNode(ARM64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1777 DAG.getConstant(PrfOp, MVT::i32), Op.getOperand(1));
1780 SDValue ARM64TargetLowering::LowerFP_EXTEND(SDValue Op,
1781 SelectionDAG &DAG) const {
1782 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1785 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1787 return LowerF128Call(Op, DAG, LC);
1790 SDValue ARM64TargetLowering::LowerFP_ROUND(SDValue Op,
1791 SelectionDAG &DAG) const {
1792 if (Op.getOperand(0).getValueType() != MVT::f128) {
1793 // It's legal except when f128 is involved
1798 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1800 // FP_ROUND node has a second operand indicating whether it is known to be
1801 // precise. That doesn't take part in the LibCall so we can't directly use
1803 SDValue SrcVal = Op.getOperand(0);
1804 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1805 /*isSigned*/ false, SDLoc(Op)).first;
1808 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1809 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1810 // Any additional optimization in this function should be recorded
1811 // in the cost tables.
1812 EVT InVT = Op.getOperand(0).getValueType();
1813 EVT VT = Op.getValueType();
1815 // FP_TO_XINT conversion from the same type are legal.
1816 if (VT.getSizeInBits() == InVT.getSizeInBits())
1819 if (InVT == MVT::v2f64) {
1821 SDValue Cv = DAG.getNode(Op.getOpcode(), dl, MVT::v2i64, Op.getOperand(0));
1822 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1825 // Type changing conversions are illegal.
1829 SDValue ARM64TargetLowering::LowerFP_TO_INT(SDValue Op,
1830 SelectionDAG &DAG) const {
1831 if (Op.getOperand(0).getValueType().isVector())
1832 return LowerVectorFP_TO_INT(Op, DAG);
1834 if (Op.getOperand(0).getValueType() != MVT::f128) {
1835 // It's legal except when f128 is involved
1840 if (Op.getOpcode() == ISD::FP_TO_SINT)
1841 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1843 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1845 SmallVector<SDValue, 2> Ops;
1846 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1847 Ops.push_back(Op.getOperand(i));
1849 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1853 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1854 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1855 // Any additional optimization in this function should be recorded
1856 // in the cost tables.
1857 EVT VT = Op.getValueType();
1859 SDValue In = Op.getOperand(0);
1860 EVT InVT = In.getValueType();
1862 // v2i32 to v2f32 is legal.
1863 if (VT == MVT::v2f32 && InVT == MVT::v2i32)
1866 // This function only handles v2f64 outputs.
1867 if (VT == MVT::v2f64) {
1868 // Extend the input argument to a v2i64 that we can feed into the
1869 // floating point conversion. Zero or sign extend based on whether
1870 // we're doing a signed or unsigned float conversion.
1872 Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
1873 assert(Op.getNumOperands() == 1 && "FP conversions take one argument");
1874 SDValue Promoted = DAG.getNode(Opc, dl, MVT::v2i64, Op.getOperand(0));
1875 return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), Promoted);
1878 // Scalarize v2i64 to v2f32 conversions.
1879 std::vector<SDValue> BuildVectorOps;
1880 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
1881 SDValue Sclr = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, In,
1882 DAG.getConstant(i, MVT::i64));
1883 Sclr = DAG.getNode(Op->getOpcode(), dl, MVT::f32, Sclr);
1884 BuildVectorOps.push_back(Sclr);
1887 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &BuildVectorOps[0],
1888 BuildVectorOps.size());
1891 SDValue ARM64TargetLowering::LowerINT_TO_FP(SDValue Op,
1892 SelectionDAG &DAG) const {
1893 if (Op.getValueType().isVector())
1894 return LowerVectorINT_TO_FP(Op, DAG);
1896 // i128 conversions are libcalls.
1897 if (Op.getOperand(0).getValueType() == MVT::i128)
1900 // Other conversions are legal, unless it's to the completely software-based
1902 if (Op.getValueType() != MVT::f128)
1906 if (Op.getOpcode() == ISD::SINT_TO_FP)
1907 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1909 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1911 return LowerF128Call(Op, DAG, LC);
1914 SDValue ARM64TargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
1915 // For iOS, we want to call an alternative entry point: __sincos_stret,
1916 // which returns the values in two S / D registers.
1918 SDValue Arg = Op.getOperand(0);
1919 EVT ArgVT = Arg.getValueType();
1920 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1927 Entry.isSExt = false;
1928 Entry.isZExt = false;
1929 Args.push_back(Entry);
1931 const char *LibcallName =
1932 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1933 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
1935 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
1936 TargetLowering::CallLoweringInfo CLI(
1937 DAG.getEntryNode(), RetTy, false, false, false, false, 0,
1938 CallingConv::Fast, /*isTaillCall=*/false,
1939 /*doesNotRet=*/false, /*isReturnValueUsed*/ true, Callee, Args, DAG, dl);
1940 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1941 return CallResult.first;
1944 SDValue ARM64TargetLowering::LowerOperation(SDValue Op,
1945 SelectionDAG &DAG) const {
1946 switch (Op.getOpcode()) {
1948 llvm_unreachable("unimplemented operand");
1950 case ISD::GlobalAddress:
1951 return LowerGlobalAddress(Op, DAG);
1952 case ISD::GlobalTLSAddress:
1953 return LowerGlobalTLSAddress(Op, DAG);
1955 return LowerSETCC(Op, DAG);
1957 return LowerBR_CC(Op, DAG);
1959 return LowerSELECT(Op, DAG);
1960 case ISD::SELECT_CC:
1961 return LowerSELECT_CC(Op, DAG);
1962 case ISD::JumpTable:
1963 return LowerJumpTable(Op, DAG);
1964 case ISD::ConstantPool:
1965 return LowerConstantPool(Op, DAG);
1966 case ISD::BlockAddress:
1967 return LowerBlockAddress(Op, DAG);
1969 return LowerVASTART(Op, DAG);
1971 return LowerVACOPY(Op, DAG);
1973 return LowerVAARG(Op, DAG);
1978 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
1985 return LowerXALUO(Op, DAG);
1987 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
1989 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
1991 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
1993 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
1995 return LowerFP_ROUND(Op, DAG);
1996 case ISD::FP_EXTEND:
1997 return LowerFP_EXTEND(Op, DAG);
1998 case ISD::FRAMEADDR:
1999 return LowerFRAMEADDR(Op, DAG);
2000 case ISD::RETURNADDR:
2001 return LowerRETURNADDR(Op, DAG);
2002 case ISD::INSERT_VECTOR_ELT:
2003 return LowerINSERT_VECTOR_ELT(Op, DAG);
2004 case ISD::EXTRACT_VECTOR_ELT:
2005 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2006 case ISD::BUILD_VECTOR:
2007 return LowerBUILD_VECTOR(Op, DAG);
2008 case ISD::VECTOR_SHUFFLE:
2009 return LowerVECTOR_SHUFFLE(Op, DAG);
2010 case ISD::EXTRACT_SUBVECTOR:
2011 return LowerEXTRACT_SUBVECTOR(Op, DAG);
2015 return LowerVectorSRA_SRL_SHL(Op, DAG);
2016 case ISD::SHL_PARTS:
2017 return LowerShiftLeftParts(Op, DAG);
2018 case ISD::SRL_PARTS:
2019 case ISD::SRA_PARTS:
2020 return LowerShiftRightParts(Op, DAG);
2022 return LowerCTPOP(Op, DAG);
2023 case ISD::FCOPYSIGN:
2024 return LowerFCOPYSIGN(Op, DAG);
2026 return LowerVectorAND(Op, DAG);
2028 return LowerVectorOR(Op, DAG);
2030 return LowerXOR(Op, DAG);
2032 return LowerPREFETCH(Op, DAG);
2033 case ISD::SINT_TO_FP:
2034 case ISD::UINT_TO_FP:
2035 return LowerINT_TO_FP(Op, DAG);
2036 case ISD::FP_TO_SINT:
2037 case ISD::FP_TO_UINT:
2038 return LowerFP_TO_INT(Op, DAG);
2040 return LowerFSINCOS(Op, DAG);
2044 /// getFunctionAlignment - Return the Log2 alignment of this function.
2045 unsigned ARM64TargetLowering::getFunctionAlignment(const Function *F) const {
2049 //===----------------------------------------------------------------------===//
2050 // Calling Convention Implementation
2051 //===----------------------------------------------------------------------===//
2053 #include "ARM64GenCallingConv.inc"
2055 /// Selects the correct CCAssignFn for a the given CallingConvention
2057 CCAssignFn *ARM64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
2058 bool IsVarArg) const {
2061 llvm_unreachable("Unsupported calling convention.");
2062 case CallingConv::WebKit_JS:
2063 return CC_ARM64_WebKit_JS;
2064 case CallingConv::C:
2065 case CallingConv::Fast:
2066 if (!Subtarget->isTargetDarwin())
2067 return CC_ARM64_AAPCS;
2068 return IsVarArg ? CC_ARM64_DarwinPCS_VarArg : CC_ARM64_DarwinPCS;
2072 SDValue ARM64TargetLowering::LowerFormalArguments(
2073 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2074 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2075 SmallVectorImpl<SDValue> &InVals) const {
2076 MachineFunction &MF = DAG.getMachineFunction();
2077 MachineFrameInfo *MFI = MF.getFrameInfo();
2079 // Assign locations to all of the incoming arguments.
2080 SmallVector<CCValAssign, 16> ArgLocs;
2081 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2082 getTargetMachine(), ArgLocs, *DAG.getContext());
2084 // At this point, Ins[].VT may already be promoted to i32. To correctly
2085 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2086 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
2087 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
2088 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
2090 unsigned NumArgs = Ins.size();
2091 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2092 unsigned CurArgIdx = 0;
2093 for (unsigned i = 0; i != NumArgs; ++i) {
2094 MVT ValVT = Ins[i].VT;
2095 std::advance(CurOrigArg, Ins[i].OrigArgIndex - CurArgIdx);
2096 CurArgIdx = Ins[i].OrigArgIndex;
2098 // Get type of the original argument.
2099 EVT ActualVT = getValueType(CurOrigArg->getType(), /*AllowUnknown*/ true);
2100 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
2101 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2103 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2105 else if (ActualMVT == MVT::i16)
2108 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2110 AssignFn(i, ValVT, LocVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
2111 assert(!Res && "Call operand has unhandled type");
2115 SmallVector<SDValue, 16> ArgValues;
2116 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2117 CCValAssign &VA = ArgLocs[i];
2119 // Arguments stored in registers.
2120 if (VA.isRegLoc()) {
2121 EVT RegVT = VA.getLocVT();
2124 const TargetRegisterClass *RC;
2126 if (RegVT == MVT::i32)
2127 RC = &ARM64::GPR32RegClass;
2128 else if (RegVT == MVT::i64)
2129 RC = &ARM64::GPR64RegClass;
2130 else if (RegVT == MVT::f32)
2131 RC = &ARM64::FPR32RegClass;
2132 else if (RegVT == MVT::f64 || RegVT == MVT::v1i64 ||
2133 RegVT == MVT::v1f64 || RegVT == MVT::v2i32 ||
2134 RegVT == MVT::v4i16 || RegVT == MVT::v8i8)
2135 RC = &ARM64::FPR64RegClass;
2136 else if (RegVT == MVT::v2i64 || RegVT == MVT::v4i32 ||
2137 RegVT == MVT::v8i16 || RegVT == MVT::v16i8)
2138 RC = &ARM64::FPR128RegClass;
2140 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2142 // Transform the arguments in physical registers into virtual ones.
2143 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2144 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2146 // If this is an 8, 16 or 32-bit value, it is really passed promoted
2147 // to 64 bits. Insert an assert[sz]ext to capture this, then
2148 // truncate to the right size.
2149 switch (VA.getLocInfo()) {
2151 llvm_unreachable("Unknown loc info!");
2152 case CCValAssign::Full:
2154 case CCValAssign::BCvt:
2155 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
2157 case CCValAssign::SExt:
2158 ArgValue = DAG.getNode(ISD::AssertSext, DL, RegVT, ArgValue,
2159 DAG.getValueType(VA.getValVT()));
2160 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
2162 case CCValAssign::ZExt:
2163 ArgValue = DAG.getNode(ISD::AssertZext, DL, RegVT, ArgValue,
2164 DAG.getValueType(VA.getValVT()));
2165 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
2169 InVals.push_back(ArgValue);
2171 } else { // VA.isRegLoc()
2172 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
2173 unsigned ArgOffset = VA.getLocMemOffset();
2174 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
2175 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset, true);
2177 // Create load nodes to retrieve arguments from the stack.
2178 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2179 InVals.push_back(DAG.getLoad(VA.getValVT(), DL, Chain, FIN,
2180 MachinePointerInfo::getFixedStack(FI), false,
2187 if (!Subtarget->isTargetDarwin()) {
2188 // The AAPCS variadic function ABI is identical to the non-variadic
2189 // one. As a result there may be more arguments in registers and we should
2190 // save them for future reference.
2191 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
2194 ARM64FunctionInfo *AFI = MF.getInfo<ARM64FunctionInfo>();
2195 // This will point to the next argument passed via stack.
2196 unsigned StackOffset = CCInfo.getNextStackOffset();
2197 // We currently pass all varargs at 8-byte alignment.
2198 StackOffset = ((StackOffset + 7) & ~7);
2199 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
2205 void ARM64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
2206 SelectionDAG &DAG, SDLoc DL,
2207 SDValue &Chain) const {
2208 MachineFunction &MF = DAG.getMachineFunction();
2209 MachineFrameInfo *MFI = MF.getFrameInfo();
2210 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
2212 SmallVector<SDValue, 8> MemOps;
2214 static const MCPhysReg GPRArgRegs[] = { ARM64::X0, ARM64::X1, ARM64::X2,
2215 ARM64::X3, ARM64::X4, ARM64::X5,
2216 ARM64::X6, ARM64::X7 };
2217 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
2218 unsigned FirstVariadicGPR =
2219 CCInfo.getFirstUnallocated(GPRArgRegs, NumGPRArgRegs);
2221 static const MCPhysReg FPRArgRegs[] = { ARM64::Q0, ARM64::Q1, ARM64::Q2,
2222 ARM64::Q3, ARM64::Q4, ARM64::Q5,
2223 ARM64::Q6, ARM64::Q7 };
2224 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
2225 unsigned FirstVariadicFPR =
2226 CCInfo.getFirstUnallocated(FPRArgRegs, NumFPRArgRegs);
2228 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
2230 if (GPRSaveSize != 0) {
2231 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
2233 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
2235 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
2236 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &ARM64::GPR64RegClass);
2237 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
2239 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2240 MachinePointerInfo::getStack(i * 8), false, false, 0);
2241 MemOps.push_back(Store);
2242 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2243 DAG.getConstant(8, getPointerTy()));
2247 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
2249 if (FPRSaveSize != 0) {
2250 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
2252 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
2254 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
2255 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &ARM64::FPR128RegClass);
2256 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::v2i64);
2258 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2259 MachinePointerInfo::getStack(i * 16), false, false, 0);
2260 MemOps.push_back(Store);
2261 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2262 DAG.getConstant(16, getPointerTy()));
2266 FuncInfo->setVarArgsGPRIndex(GPRIdx);
2267 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
2268 FuncInfo->setVarArgsFPRIndex(FPRIdx);
2269 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
2271 if (!MemOps.empty()) {
2272 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
2277 /// LowerCallResult - Lower the result values of a call into the
2278 /// appropriate copies out of appropriate physical registers.
2279 SDValue ARM64TargetLowering::LowerCallResult(
2280 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
2281 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2282 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
2283 SDValue ThisVal) const {
2284 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2285 : RetCC_ARM64_AAPCS;
2286 // Assign locations to each value returned by this call.
2287 SmallVector<CCValAssign, 16> RVLocs;
2288 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2289 getTargetMachine(), RVLocs, *DAG.getContext());
2290 CCInfo.AnalyzeCallResult(Ins, RetCC);
2292 // Copy all of the result registers out of their specified physreg.
2293 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2294 CCValAssign VA = RVLocs[i];
2296 // Pass 'this' value directly from the argument to return value, to avoid
2297 // reg unit interference
2298 if (i == 0 && isThisReturn) {
2299 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
2300 "unexpected return calling convention register assignment");
2301 InVals.push_back(ThisVal);
2306 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
2307 Chain = Val.getValue(1);
2308 InFlag = Val.getValue(2);
2310 switch (VA.getLocInfo()) {
2312 llvm_unreachable("Unknown loc info!");
2313 case CCValAssign::Full:
2315 case CCValAssign::BCvt:
2316 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2320 InVals.push_back(Val);
2326 bool ARM64TargetLowering::isEligibleForTailCallOptimization(
2327 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
2328 bool isCalleeStructRet, bool isCallerStructRet,
2329 const SmallVectorImpl<ISD::OutputArg> &Outs,
2330 const SmallVectorImpl<SDValue> &OutVals,
2331 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
2332 // Look for obvious safe cases to perform tail call optimization that do not
2333 // require ABI changes. This is what gcc calls sibcall.
2335 // Do not sibcall optimize vararg calls unless the call site is not passing
2337 if (isVarArg && !Outs.empty())
2340 // Also avoid sibcall optimization if either caller or callee uses struct
2341 // return semantics.
2342 if (isCalleeStructRet || isCallerStructRet)
2345 // Note that currently ARM64 "C" calling convention and "Fast" calling
2346 // convention are compatible. If/when that ever changes, we'll need to
2347 // add checks here to make sure any interactions are OK.
2349 // If the callee takes no arguments then go on to check the results of the
2351 if (!Outs.empty()) {
2352 // Check if stack adjustment is needed. For now, do not do this if any
2353 // argument is passed on the stack.
2354 SmallVector<CCValAssign, 16> ArgLocs;
2355 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2356 getTargetMachine(), ArgLocs, *DAG.getContext());
2357 CCAssignFn *AssignFn = CCAssignFnForCall(CalleeCC, /*IsVarArg=*/false);
2358 CCInfo.AnalyzeCallOperands(Outs, AssignFn);
2359 if (CCInfo.getNextStackOffset()) {
2360 // Check if the arguments are already laid out in the right way as
2361 // the caller's fixed stack objects.
2362 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2363 ++i, ++realArgIdx) {
2364 CCValAssign &VA = ArgLocs[i];
2365 if (VA.getLocInfo() == CCValAssign::Indirect)
2367 if (VA.needsCustom()) {
2368 // Just don't handle anything that needs custom adjustments for now.
2369 // If need be, we can revisit later, but we shouldn't ever end up
2372 } else if (!VA.isRegLoc()) {
2373 // Likewise, don't try to handle stack based arguments for the
2383 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
2384 /// and add input and output parameter nodes.
2385 SDValue ARM64TargetLowering::LowerCall(CallLoweringInfo &CLI,
2386 SmallVectorImpl<SDValue> &InVals) const {
2387 SelectionDAG &DAG = CLI.DAG;
2389 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2390 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2391 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2392 SDValue Chain = CLI.Chain;
2393 SDValue Callee = CLI.Callee;
2394 bool &IsTailCall = CLI.IsTailCall;
2395 CallingConv::ID CallConv = CLI.CallConv;
2396 bool IsVarArg = CLI.IsVarArg;
2398 MachineFunction &MF = DAG.getMachineFunction();
2399 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2400 bool IsThisReturn = false;
2402 // If tail calls are explicitly disabled, make sure not to use them.
2403 if (!EnableARM64TailCalls)
2407 // Check if it's really possible to do a tail call.
2408 IsTailCall = isEligibleForTailCallOptimization(
2409 Callee, CallConv, IsVarArg, IsStructRet,
2410 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2411 // We don't support GuaranteedTailCallOpt, only automatically
2412 // detected sibcalls.
2413 // FIXME: Re-evaluate. Is this true? Should it be true?
2418 // Analyze operands of the call, assigning locations to each operand.
2419 SmallVector<CCValAssign, 16> ArgLocs;
2420 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2421 getTargetMachine(), ArgLocs, *DAG.getContext());
2424 // Handle fixed and variable vector arguments differently.
2425 // Variable vector arguments always go into memory.
2426 unsigned NumArgs = Outs.size();
2428 for (unsigned i = 0; i != NumArgs; ++i) {
2429 MVT ArgVT = Outs[i].VT;
2430 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2431 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2432 /*IsVarArg=*/ !Outs[i].IsFixed);
2433 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2434 assert(!Res && "Call operand has unhandled type");
2438 // At this point, Outs[].VT may already be promoted to i32. To correctly
2439 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2440 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
2441 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2442 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2444 unsigned NumArgs = Outs.size();
2445 for (unsigned i = 0; i != NumArgs; ++i) {
2446 MVT ValVT = Outs[i].VT;
2447 // Get type of the original argument.
2448 EVT ActualVT = getValueType(CLI.Args[Outs[i].OrigArgIndex].Ty,
2449 /*AllowUnknown*/ true);
2450 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2451 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2452 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2454 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2456 else if (ActualMVT == MVT::i16)
2459 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2460 bool Res = AssignFn(i, ValVT, LocVT, CCValAssign::Full, ArgFlags, CCInfo);
2461 assert(!Res && "Call operand has unhandled type");
2466 // Get a count of how many bytes are to be pushed on the stack.
2467 unsigned NumBytes = CCInfo.getNextStackOffset();
2469 // Adjust the stack pointer for the new arguments...
2470 // These operations are automatically eliminated by the prolog/epilog pass
2473 DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true), DL);
2475 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, ARM64::SP, getPointerTy());
2477 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2478 SmallVector<SDValue, 8> MemOpChains;
2480 // Walk the register/memloc assignments, inserting copies/loads.
2481 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2482 ++i, ++realArgIdx) {
2483 CCValAssign &VA = ArgLocs[i];
2484 SDValue Arg = OutVals[realArgIdx];
2485 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2487 // Promote the value if needed.
2488 switch (VA.getLocInfo()) {
2490 llvm_unreachable("Unknown loc info!");
2491 case CCValAssign::Full:
2493 case CCValAssign::SExt:
2494 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
2496 case CCValAssign::ZExt:
2497 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2499 case CCValAssign::AExt:
2500 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
2502 case CCValAssign::BCvt:
2503 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2505 case CCValAssign::FPExt:
2506 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
2510 if (VA.isRegLoc()) {
2511 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
2512 assert(VA.getLocVT() == MVT::i64 &&
2513 "unexpected calling convention register assignment");
2514 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
2515 "unexpected use of 'returned'");
2516 IsThisReturn = true;
2518 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2520 assert(VA.isMemLoc());
2521 // There's no reason we can't support stack args w/ tailcall, but
2522 // we currently don't, so assert if we see one.
2523 assert(!IsTailCall && "stack argument with tail call!?");
2524 unsigned LocMemOffset = VA.getLocMemOffset();
2525 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2526 PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2528 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
2529 // promoted to a legal register type i32, we should truncate Arg back to
2531 if (Arg.getValueType().isSimple() &&
2532 Arg.getValueType().getSimpleVT() == MVT::i32 &&
2533 (VA.getLocVT() == MVT::i1 || VA.getLocVT() == MVT::i8 ||
2534 VA.getLocVT() == MVT::i16))
2535 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getLocVT(), Arg);
2537 SDValue Store = DAG.getStore(Chain, DL, Arg, PtrOff,
2538 MachinePointerInfo::getStack(LocMemOffset),
2540 MemOpChains.push_back(Store);
2544 if (!MemOpChains.empty())
2545 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOpChains[0],
2546 MemOpChains.size());
2548 // Build a sequence of copy-to-reg nodes chained together with token chain
2549 // and flag operands which copy the outgoing args into the appropriate regs.
2551 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2552 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[i].first,
2553 RegsToPass[i].second, InFlag);
2554 InFlag = Chain.getValue(1);
2557 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2558 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2559 // node so that legalize doesn't hack it.
2560 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
2561 Subtarget->isTargetMachO()) {
2562 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2563 const GlobalValue *GV = G->getGlobal();
2564 bool InternalLinkage = GV->hasInternalLinkage();
2565 if (InternalLinkage)
2566 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2568 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0,
2570 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2572 } else if (ExternalSymbolSDNode *S =
2573 dyn_cast<ExternalSymbolSDNode>(Callee)) {
2574 const char *Sym = S->getSymbol();
2576 DAG.getTargetExternalSymbol(Sym, getPointerTy(), ARM64II::MO_GOT);
2577 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2579 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2580 const GlobalValue *GV = G->getGlobal();
2581 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2582 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2583 const char *Sym = S->getSymbol();
2584 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), 0);
2587 std::vector<SDValue> Ops;
2588 Ops.push_back(Chain);
2589 Ops.push_back(Callee);
2591 // Add argument registers to the end of the list so that they are known live
2593 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2594 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2595 RegsToPass[i].second.getValueType()));
2597 // Add a register mask operand representing the call-preserved registers.
2598 const uint32_t *Mask;
2599 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2600 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2602 // For 'this' returns, use the X0-preserving mask if applicable
2603 Mask = ARI->getThisReturnPreservedMask(CallConv);
2605 IsThisReturn = false;
2606 Mask = ARI->getCallPreservedMask(CallConv);
2609 Mask = ARI->getCallPreservedMask(CallConv);
2611 assert(Mask && "Missing call preserved mask for calling convention");
2612 Ops.push_back(DAG.getRegisterMask(Mask));
2614 if (InFlag.getNode())
2615 Ops.push_back(InFlag);
2617 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2619 // If we're doing a tall call, use a TC_RETURN here rather than an
2620 // actual call instruction.
2622 return DAG.getNode(ARM64ISD::TC_RETURN, DL, NodeTys, &Ops[0], Ops.size());
2624 // Returns a chain and a flag for retval copy to use.
2625 Chain = DAG.getNode(ARM64ISD::CALL, DL, NodeTys, &Ops[0], Ops.size());
2626 InFlag = Chain.getValue(1);
2628 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2629 DAG.getIntPtrConstant(0, true), InFlag, DL);
2631 InFlag = Chain.getValue(1);
2633 // Handle result values, copying them out of physregs into vregs that we
2635 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2636 InVals, IsThisReturn,
2637 IsThisReturn ? OutVals[0] : SDValue());
2640 bool ARM64TargetLowering::CanLowerReturn(
2641 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2642 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2643 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2644 : RetCC_ARM64_AAPCS;
2645 SmallVector<CCValAssign, 16> RVLocs;
2646 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2647 return CCInfo.CheckReturn(Outs, RetCC);
2651 ARM64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2653 const SmallVectorImpl<ISD::OutputArg> &Outs,
2654 const SmallVectorImpl<SDValue> &OutVals,
2655 SDLoc DL, SelectionDAG &DAG) const {
2656 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2657 : RetCC_ARM64_AAPCS;
2658 SmallVector<CCValAssign, 16> RVLocs;
2659 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2660 getTargetMachine(), RVLocs, *DAG.getContext());
2661 CCInfo.AnalyzeReturn(Outs, RetCC);
2663 // Copy the result values into the output registers.
2665 SmallVector<SDValue, 4> RetOps(1, Chain);
2666 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
2667 ++i, ++realRVLocIdx) {
2668 CCValAssign &VA = RVLocs[i];
2669 assert(VA.isRegLoc() && "Can only return in registers!");
2670 SDValue Arg = OutVals[realRVLocIdx];
2672 switch (VA.getLocInfo()) {
2674 llvm_unreachable("Unknown loc info!");
2675 case CCValAssign::Full:
2677 case CCValAssign::BCvt:
2678 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2682 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
2683 Flag = Chain.getValue(1);
2684 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2687 RetOps[0] = Chain; // Update chain.
2689 // Add the flag if we have it.
2691 RetOps.push_back(Flag);
2693 return DAG.getNode(ARM64ISD::RET_FLAG, DL, MVT::Other, &RetOps[0],
2697 //===----------------------------------------------------------------------===//
2698 // Other Lowering Code
2699 //===----------------------------------------------------------------------===//
2701 SDValue ARM64TargetLowering::LowerGlobalAddress(SDValue Op,
2702 SelectionDAG &DAG) const {
2703 EVT PtrVT = getPointerTy();
2705 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2706 unsigned char OpFlags =
2707 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
2709 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
2710 "unexpected offset in global node");
2712 // This also catched the large code model case for Darwin.
2713 if ((OpFlags & ARM64II::MO_GOT) != 0) {
2714 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
2715 // FIXME: Once remat is capable of dealing with instructions with register
2716 // operands, expand this into two nodes instead of using a wrapper node.
2717 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
2720 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2721 const unsigned char MO_NC = ARM64II::MO_NC;
2723 ARM64ISD::WrapperLarge, DL, PtrVT,
2724 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G3),
2725 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
2726 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
2727 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
2729 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
2730 // the only correct model on Darwin.
2731 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2732 OpFlags | ARM64II::MO_PAGE);
2733 unsigned char LoFlags = OpFlags | ARM64II::MO_PAGEOFF | ARM64II::MO_NC;
2734 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
2736 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
2737 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
2741 /// \brief Convert a TLS address reference into the correct sequence of loads
2742 /// and calls to compute the variable's address (for Darwin, currently) and
2743 /// return an SDValue containing the final node.
2745 /// Darwin only has one TLS scheme which must be capable of dealing with the
2746 /// fully general situation, in the worst case. This means:
2747 /// + "extern __thread" declaration.
2748 /// + Defined in a possibly unknown dynamic library.
2750 /// The general system is that each __thread variable has a [3 x i64] descriptor
2751 /// which contains information used by the runtime to calculate the address. The
2752 /// only part of this the compiler needs to know about is the first xword, which
2753 /// contains a function pointer that must be called with the address of the
2754 /// entire descriptor in "x0".
2756 /// Since this descriptor may be in a different unit, in general even the
2757 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
2759 /// adrp x0, _var@TLVPPAGE
2760 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
2761 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
2762 /// ; the function pointer
2763 /// blr x1 ; Uses descriptor address in x0
2764 /// ; Address of _var is now in x0.
2766 /// If the address of _var's descriptor *is* known to the linker, then it can
2767 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
2768 /// a slight efficiency gain.
2770 ARM64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
2771 SelectionDAG &DAG) const {
2772 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
2775 MVT PtrVT = getPointerTy();
2776 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2779 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2780 SDValue DescAddr = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TLVPAddr);
2782 // The first entry in the descriptor is a function pointer that we must call
2783 // to obtain the address of the variable.
2784 SDValue Chain = DAG.getEntryNode();
2785 SDValue FuncTLVGet =
2786 DAG.getLoad(MVT::i64, DL, Chain, DescAddr, MachinePointerInfo::getGOT(),
2787 false, true, true, 8);
2788 Chain = FuncTLVGet.getValue(1);
2790 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2791 MFI->setAdjustsStack(true);
2793 // TLS calls preserve all registers except those that absolutely must be
2794 // trashed: X0 (it takes an argument), LR (it's a call) and CPSR (let's not be
2796 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2797 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2798 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2800 // Finally, we can make the call. This is just a degenerate version of a
2801 // normal ARM64 call node: x0 takes the address of the descriptor, and returns
2802 // the address of the variable in this thread.
2803 Chain = DAG.getCopyToReg(Chain, DL, ARM64::X0, DescAddr, SDValue());
2804 Chain = DAG.getNode(ARM64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
2805 Chain, FuncTLVGet, DAG.getRegister(ARM64::X0, MVT::i64),
2806 DAG.getRegisterMask(Mask), Chain.getValue(1));
2807 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Chain.getValue(1));
2810 /// When accessing thread-local variables under either the general-dynamic or
2811 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
2812 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
2813 /// is a function pointer to carry out the resolution. This function takes the
2814 /// address of the descriptor in X0 and returns the TPIDR_EL0 offset in X0. All
2815 /// other registers (except LR, CPSR) are preserved.
2817 /// Thus, the ideal call sequence on AArch64 is:
2819 /// adrp x0, :tlsdesc:thread_var
2820 /// ldr x8, [x0, :tlsdesc_lo12:thread_var]
2821 /// add x0, x0, :tlsdesc_lo12:thread_var
2822 /// .tlsdesccall thread_var
2824 /// (TPIDR_EL0 offset now in x0).
2826 /// The ".tlsdesccall" directive instructs the assembler to insert a particular
2827 /// relocation to help the linker relax this sequence if it turns out to be too
2830 /// FIXME: we currently produce an extra, duplicated, ADRP instruction, but this
2832 SDValue ARM64TargetLowering::LowerELFTLSDescCall(SDValue SymAddr,
2833 SDValue DescAddr, SDLoc DL,
2834 SelectionDAG &DAG) const {
2835 EVT PtrVT = getPointerTy();
2837 // The function we need to call is simply the first entry in the GOT for this
2838 // descriptor, load it in preparation.
2839 SDValue Func = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, SymAddr);
2841 // TLS calls preserve all registers except those that absolutely must be
2842 // trashed: X0 (it takes an argument), LR (it's a call) and CPSR (let's not be
2844 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2845 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2846 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2848 // The function takes only one argument: the address of the descriptor itself
2850 SDValue Glue, Chain;
2851 Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, ARM64::X0, DescAddr, Glue);
2852 Glue = Chain.getValue(1);
2854 // We're now ready to populate the argument list, as with a normal call:
2855 SmallVector<SDValue, 6> Ops;
2856 Ops.push_back(Chain);
2857 Ops.push_back(Func);
2858 Ops.push_back(SymAddr);
2859 Ops.push_back(DAG.getRegister(ARM64::X0, PtrVT));
2860 Ops.push_back(DAG.getRegisterMask(Mask));
2861 Ops.push_back(Glue);
2863 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2864 Chain = DAG.getNode(ARM64ISD::TLSDESC_CALL, DL, NodeTys, &Ops[0], Ops.size());
2865 Glue = Chain.getValue(1);
2867 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Glue);
2870 SDValue ARM64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
2871 SelectionDAG &DAG) const {
2872 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
2873 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
2874 "ELF TLS only supported in small memory model");
2875 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2877 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
2880 EVT PtrVT = getPointerTy();
2882 const GlobalValue *GV = GA->getGlobal();
2884 SDValue ThreadBase = DAG.getNode(ARM64ISD::THREAD_POINTER, DL, PtrVT);
2886 if (Model == TLSModel::LocalExec) {
2887 SDValue HiVar = DAG.getTargetGlobalAddress(
2888 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2889 SDValue LoVar = DAG.getTargetGlobalAddress(
2890 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2892 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2893 DAG.getTargetConstant(16, MVT::i32)),
2895 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, TPOff, LoVar,
2896 DAG.getTargetConstant(0, MVT::i32)),
2898 } else if (Model == TLSModel::InitialExec) {
2899 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2900 TPOff = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TPOff);
2901 } else if (Model == TLSModel::LocalDynamic) {
2902 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
2903 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
2904 // the beginning of the module's TLS region, followed by a DTPREL offset
2907 // These accesses will need deduplicating if there's more than one.
2908 ARM64FunctionInfo *MFI =
2909 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
2910 MFI->incNumLocalDynamicTLSAccesses();
2912 // Accesses used in this sequence go via the TLS descriptor which lives in
2913 // the GOT. Prepare an address we can use to handle this.
2914 SDValue HiDesc = DAG.getTargetExternalSymbol(
2915 "_TLS_MODULE_BASE_", PtrVT, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2916 SDValue LoDesc = DAG.getTargetExternalSymbol(
2917 "_TLS_MODULE_BASE_", PtrVT,
2918 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2920 // First argument to the descriptor call is the address of the descriptor
2922 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2923 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2925 // The call needs a relocation too for linker relaxation. It doesn't make
2926 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2928 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2931 // Now we can calculate the offset from TPIDR_EL0 to this module's
2932 // thread-local area.
2933 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2935 // Now use :dtprel_whatever: operations to calculate this variable's offset
2936 // in its thread-storage area.
2937 SDValue HiVar = DAG.getTargetGlobalAddress(
2938 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2939 SDValue LoVar = DAG.getTargetGlobalAddress(
2940 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2943 SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2944 DAG.getTargetConstant(16, MVT::i32)),
2946 DTPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, DTPOff, LoVar,
2947 DAG.getTargetConstant(0, MVT::i32)),
2950 TPOff = DAG.getNode(ISD::ADD, DL, PtrVT, TPOff, DTPOff);
2951 } else if (Model == TLSModel::GeneralDynamic) {
2952 // Accesses used in this sequence go via the TLS descriptor which lives in
2953 // the GOT. Prepare an address we can use to handle this.
2954 SDValue HiDesc = DAG.getTargetGlobalAddress(
2955 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2956 SDValue LoDesc = DAG.getTargetGlobalAddress(
2958 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2960 // First argument to the descriptor call is the address of the descriptor
2962 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2963 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2965 // The call needs a relocation too for linker relaxation. It doesn't make
2966 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2969 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2971 // Finally we can make a call to calculate the offset from tpidr_el0.
2972 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2974 llvm_unreachable("Unsupported ELF TLS access model");
2976 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
2979 SDValue ARM64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
2980 SelectionDAG &DAG) const {
2981 if (Subtarget->isTargetDarwin())
2982 return LowerDarwinGlobalTLSAddress(Op, DAG);
2983 else if (Subtarget->isTargetELF())
2984 return LowerELFGlobalTLSAddress(Op, DAG);
2986 llvm_unreachable("Unexpected platform trying to use TLS");
2988 SDValue ARM64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2989 SDValue Chain = Op.getOperand(0);
2990 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2991 SDValue LHS = Op.getOperand(2);
2992 SDValue RHS = Op.getOperand(3);
2993 SDValue Dest = Op.getOperand(4);
2996 // Handle f128 first, since lowering it will result in comparing the return
2997 // value of a libcall against zero, which is just what the rest of LowerBR_CC
2998 // is expecting to deal with.
2999 if (LHS.getValueType() == MVT::f128) {
3000 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3002 // If softenSetCCOperands returned a scalar, we need to compare the result
3003 // against zero to select between true and false values.
3004 if (RHS.getNode() == 0) {
3005 RHS = DAG.getConstant(0, LHS.getValueType());
3010 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
3012 unsigned Opc = LHS.getOpcode();
3013 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
3014 cast<ConstantSDNode>(RHS)->isOne() &&
3015 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3016 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3017 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
3018 "Unexpected condition code.");
3019 // Only lower legal XALUO ops.
3020 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
3023 // The actual operation with overflow check.
3024 ARM64CC::CondCode OFCC;
3025 SDValue Value, Overflow;
3026 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, LHS.getValue(0), DAG);
3028 if (CC == ISD::SETNE)
3029 OFCC = getInvertedCondCode(OFCC);
3030 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3032 return DAG.getNode(ARM64ISD::BRCOND, SDLoc(LHS), MVT::Other, Chain, Dest,
3036 if (LHS.getValueType().isInteger()) {
3037 assert((LHS.getValueType() == RHS.getValueType()) &&
3038 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3040 // If the RHS of the comparison is zero, we can potentially fold this
3041 // to a specialized branch.
3042 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
3043 if (RHSC && RHSC->getZExtValue() == 0) {
3044 if (CC == ISD::SETEQ) {
3045 // See if we can use a TBZ to fold in an AND as well.
3046 // TBZ has a smaller branch displacement than CBZ. If the offset is
3047 // out of bounds, a late MI-layer pass rewrites branches.
3048 // 403.gcc is an example that hits this case.
3049 if (LHS.getOpcode() == ISD::AND &&
3050 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3051 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3052 SDValue Test = LHS.getOperand(0);
3053 uint64_t Mask = LHS.getConstantOperandVal(1);
3055 // TBZ only operates on i64's, but the ext should be free.
3056 if (Test.getValueType() == MVT::i32)
3057 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
3059 return DAG.getNode(ARM64ISD::TBZ, dl, MVT::Other, Chain, Test,
3060 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3063 return DAG.getNode(ARM64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
3064 } else if (CC == ISD::SETNE) {
3065 // See if we can use a TBZ to fold in an AND as well.
3066 // TBZ has a smaller branch displacement than CBZ. If the offset is
3067 // out of bounds, a late MI-layer pass rewrites branches.
3068 // 403.gcc is an example that hits this case.
3069 if (LHS.getOpcode() == ISD::AND &&
3070 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3071 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3072 SDValue Test = LHS.getOperand(0);
3073 uint64_t Mask = LHS.getConstantOperandVal(1);
3075 // TBNZ only operates on i64's, but the ext should be free.
3076 if (Test.getValueType() == MVT::i32)
3077 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
3079 return DAG.getNode(ARM64ISD::TBNZ, dl, MVT::Other, Chain, Test,
3080 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3083 return DAG.getNode(ARM64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
3088 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3089 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3093 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3095 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3096 // clean. Some of them require two branches to implement.
3097 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3098 ARM64CC::CondCode CC1, CC2;
3099 changeFPCCToARM64CC(CC, CC1, CC2);
3100 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3102 DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
3103 if (CC2 != ARM64CC::AL) {
3104 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3105 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
3112 SDValue ARM64TargetLowering::LowerFCOPYSIGN(SDValue Op,
3113 SelectionDAG &DAG) const {
3114 EVT VT = Op.getValueType();
3117 SDValue In1 = Op.getOperand(0);
3118 SDValue In2 = Op.getOperand(1);
3119 EVT SrcVT = In2.getValueType();
3121 if (SrcVT == MVT::f32 && VT == MVT::f64)
3122 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
3123 else if (SrcVT == MVT::f64 && VT == MVT::f32)
3124 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0));
3126 // FIXME: Src type is different, bail out for now. Can VT really be a
3133 SDValue EltMask, VecVal1, VecVal2;
3134 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
3137 EltMask = DAG.getConstant(0x80000000ULL, EltVT);
3139 if (!VT.isVector()) {
3140 VecVal1 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
3141 DAG.getUNDEF(VecVT), In1);
3142 VecVal2 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
3143 DAG.getUNDEF(VecVT), In2);
3145 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3146 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3148 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
3152 // We want to materialize a mask with the the high bit set, but the AdvSIMD
3153 // immediate moves cannot materialize that in a single instruction for
3154 // 64-bit elements. Instead, materialize zero and then negate it.
3155 EltMask = DAG.getConstant(0, EltVT);
3157 if (!VT.isVector()) {
3158 VecVal1 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
3159 DAG.getUNDEF(VecVT), In1);
3160 VecVal2 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
3161 DAG.getUNDEF(VecVT), In2);
3163 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3164 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3167 llvm_unreachable("Invalid type for copysign!");
3170 std::vector<SDValue> BuildVectorOps;
3171 for (unsigned i = 0; i < VecVT.getVectorNumElements(); ++i)
3172 BuildVectorOps.push_back(EltMask);
3174 SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT,
3175 &BuildVectorOps[0], BuildVectorOps.size());
3177 // If we couldn't materialize the mask above, then the mask vector will be
3178 // the zero vector, and we need to negate it here.
3179 if (VT == MVT::f64 || VT == MVT::v2f64) {
3180 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
3181 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
3182 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
3186 DAG.getNode(ARM64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
3189 return DAG.getTargetExtractSubreg(ARM64::ssub, DL, VT, Sel);
3190 else if (VT == MVT::f64)
3191 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, VT, Sel);
3193 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
3196 SDValue ARM64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
3197 if (DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
3198 AttributeSet::FunctionIndex, Attribute::NoImplicitFloat))
3201 // While there is no integer popcount instruction, it can
3202 // be more efficiently lowered to the following sequence that uses
3203 // AdvSIMD registers/instructions as long as the copies to/from
3204 // the AdvSIMD registers are cheap.
3205 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
3206 // CNT V0.8B, V0.8B // 8xbyte pop-counts
3207 // ADDV B0, V0.8B // sum 8xbyte pop-counts
3208 // UMOV X0, V0.B[0] // copy byte result back to integer reg
3209 SDValue Val = Op.getOperand(0);
3211 EVT VT = Op.getValueType();
3212 SDValue ZeroVec = DAG.getUNDEF(MVT::v8i8);
3215 if (VT == MVT::i32) {
3216 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Val);
3218 DAG.getTargetInsertSubreg(ARM64::ssub, DL, MVT::v8i8, ZeroVec, VecVal);
3220 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
3223 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, VecVal);
3224 SDValue UaddLV = DAG.getNode(
3225 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
3226 DAG.getConstant(Intrinsic::arm64_neon_uaddlv, MVT::i32), CtPop);
3229 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
3233 SDValue ARM64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3235 if (Op.getValueType().isVector())
3236 return LowerVSETCC(Op, DAG);
3238 SDValue LHS = Op.getOperand(0);
3239 SDValue RHS = Op.getOperand(1);
3240 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3243 // We chose ZeroOrOneBooleanContents, so use zero and one.
3244 EVT VT = Op.getValueType();
3245 SDValue TVal = DAG.getConstant(1, VT);
3246 SDValue FVal = DAG.getConstant(0, VT);
3248 // Handle f128 first, since one possible outcome is a normal integer
3249 // comparison which gets picked up by the next if statement.
3250 if (LHS.getValueType() == MVT::f128) {
3251 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3253 // If softenSetCCOperands returned a scalar, use it.
3254 if (RHS.getNode() == 0) {
3255 assert(LHS.getValueType() == Op.getValueType() &&
3256 "Unexpected setcc expansion!");
3261 if (LHS.getValueType().isInteger()) {
3264 getARM64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
3266 // Note that we inverted the condition above, so we reverse the order of
3267 // the true and false operands here. This will allow the setcc to be
3268 // matched to a single CSINC instruction.
3269 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
3272 // Now we know we're dealing with FP values.
3273 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3275 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3276 // and do the comparison.
3277 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3279 ARM64CC::CondCode CC1, CC2;
3280 changeFPCCToARM64CC(CC, CC1, CC2);
3281 if (CC2 == ARM64CC::AL) {
3282 changeFPCCToARM64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
3283 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3285 // Note that we inverted the condition above, so we reverse the order of
3286 // the true and false operands here. This will allow the setcc to be
3287 // matched to a single CSINC instruction.
3288 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
3290 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3291 // clean. Some of them require two CSELs to implement. As is in this case,
3292 // we emit the first CSEL and then emit a second using the output of the
3293 // first as the RHS. We're effectively OR'ing the two CC's together.
3295 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
3296 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3297 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3299 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3300 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3304 /// A SELECT_CC operation is really some kind of max or min if both values being
3305 /// compared are, in some sense, equal to the results in either case. However,
3306 /// it is permissible to compare f32 values and produce directly extended f64
3309 /// Extending the comparison operands would also be allowed, but is less likely
3310 /// to happen in practice since their use is right here. Note that truncate
3311 /// operations would *not* be semantically equivalent.
3312 static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) {
3316 ConstantFPSDNode *CCmp = dyn_cast<ConstantFPSDNode>(Cmp);
3317 ConstantFPSDNode *CResult = dyn_cast<ConstantFPSDNode>(Result);
3318 if (CCmp && CResult && Cmp.getValueType() == MVT::f32 &&
3319 Result.getValueType() == MVT::f64) {
3321 APFloat CmpVal = CCmp->getValueAPF();
3322 CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy);
3323 return CResult->getValueAPF().bitwiseIsEqual(CmpVal);
3326 return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp;
3329 SDValue ARM64TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3330 SDValue CC = Op->getOperand(0);
3331 SDValue TVal = Op->getOperand(1);
3332 SDValue FVal = Op->getOperand(2);
3335 unsigned Opc = CC.getOpcode();
3336 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
3338 if (CC.getResNo() == 1 &&
3339 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3340 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3341 // Only lower legal XALUO ops.
3342 if (!DAG.getTargetLoweringInfo().isTypeLegal(CC->getValueType(0)))
3345 ARM64CC::CondCode OFCC;
3346 SDValue Value, Overflow;
3347 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, CC.getValue(0), DAG);
3348 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3350 return DAG.getNode(ARM64ISD::CSEL, DL, Op.getValueType(), TVal, FVal, CCVal,
3354 if (CC.getOpcode() == ISD::SETCC)
3355 return DAG.getSelectCC(DL, CC.getOperand(0), CC.getOperand(1), TVal, FVal,
3356 cast<CondCodeSDNode>(CC.getOperand(2))->get());
3358 return DAG.getSelectCC(DL, CC, DAG.getConstant(0, CC.getValueType()), TVal,
3362 SDValue ARM64TargetLowering::LowerSELECT_CC(SDValue Op,
3363 SelectionDAG &DAG) const {
3364 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3365 SDValue LHS = Op.getOperand(0);
3366 SDValue RHS = Op.getOperand(1);
3367 SDValue TVal = Op.getOperand(2);
3368 SDValue FVal = Op.getOperand(3);
3371 // Handle f128 first, because it will result in a comparison of some RTLIB
3372 // call result against zero.
3373 if (LHS.getValueType() == MVT::f128) {
3374 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3376 // If softenSetCCOperands returned a scalar, we need to compare the result
3377 // against zero to select between true and false values.
3378 if (RHS.getNode() == 0) {
3379 RHS = DAG.getConstant(0, LHS.getValueType());
3384 // Handle integers first.
3385 if (LHS.getValueType().isInteger()) {
3386 assert((LHS.getValueType() == RHS.getValueType()) &&
3387 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3389 unsigned Opcode = ARM64ISD::CSEL;
3391 // If both the TVal and the FVal are constants, see if we can swap them in
3392 // order to for a CSINV or CSINC out of them.
3393 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3394 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3396 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3397 std::swap(TVal, FVal);
3398 std::swap(CTVal, CFVal);
3399 CC = ISD::getSetCCInverse(CC, true);
3400 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3401 std::swap(TVal, FVal);
3402 std::swap(CTVal, CFVal);
3403 CC = ISD::getSetCCInverse(CC, true);
3404 } else if (TVal.getOpcode() == ISD::XOR) {
3405 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3406 // with a CSINV rather than a CSEL.
3407 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
3409 if (CVal && CVal->isAllOnesValue()) {
3410 std::swap(TVal, FVal);
3411 std::swap(CTVal, CFVal);
3412 CC = ISD::getSetCCInverse(CC, true);
3414 } else if (TVal.getOpcode() == ISD::SUB) {
3415 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3416 // that we can match with a CSNEG rather than a CSEL.
3417 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
3419 if (CVal && CVal->isNullValue()) {
3420 std::swap(TVal, FVal);
3421 std::swap(CTVal, CFVal);
3422 CC = ISD::getSetCCInverse(CC, true);
3424 } else if (CTVal && CFVal) {
3425 const int64_t TrueVal = CTVal->getSExtValue();
3426 const int64_t FalseVal = CFVal->getSExtValue();
3429 // If both TVal and FVal are constants, see if FVal is the
3430 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3431 // instead of a CSEL in that case.
3432 if (TrueVal == ~FalseVal) {
3433 Opcode = ARM64ISD::CSINV;
3434 } else if (TrueVal == -FalseVal) {
3435 Opcode = ARM64ISD::CSNEG;
3436 } else if (TVal.getValueType() == MVT::i32) {
3437 // If our operands are only 32-bit wide, make sure we use 32-bit
3438 // arithmetic for the check whether we can use CSINC. This ensures that
3439 // the addition in the check will wrap around properly in case there is
3440 // an overflow (which would not be the case if we do the check with
3441 // 64-bit arithmetic).
3442 const uint32_t TrueVal32 = CTVal->getZExtValue();
3443 const uint32_t FalseVal32 = CFVal->getZExtValue();
3445 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3446 Opcode = ARM64ISD::CSINC;
3448 if (TrueVal32 > FalseVal32) {
3452 // 64-bit check whether we can use CSINC.
3453 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3454 Opcode = ARM64ISD::CSINC;
3456 if (TrueVal > FalseVal) {
3461 // Swap TVal and FVal if necessary.
3463 std::swap(TVal, FVal);
3464 std::swap(CTVal, CFVal);
3465 CC = ISD::getSetCCInverse(CC, true);
3468 if (Opcode != ARM64ISD::CSEL) {
3469 // Drop FVal since we can get its value by simply inverting/negating
3476 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3478 EVT VT = Op.getValueType();
3479 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3482 // Now we know we're dealing with FP values.
3483 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3484 assert(LHS.getValueType() == RHS.getValueType());
3485 EVT VT = Op.getValueType();
3487 // Try to match this select into a max/min operation, which have dedicated
3488 // opcode in the instruction set.
3489 // NOTE: This is not correct in the presence of NaNs, so we only enable this
3491 if (getTargetMachine().Options.NoNaNsFPMath) {
3492 if (selectCCOpsAreFMaxCompatible(LHS, FVal) &&
3493 selectCCOpsAreFMaxCompatible(RHS, TVal)) {
3494 CC = ISD::getSetCCSwappedOperands(CC);
3495 std::swap(TVal, FVal);
3498 if (selectCCOpsAreFMaxCompatible(LHS, TVal) &&
3499 selectCCOpsAreFMaxCompatible(RHS, FVal)) {
3509 return DAG.getNode(ARM64ISD::FMAX, dl, VT, TVal, FVal);
3517 return DAG.getNode(ARM64ISD::FMIN, dl, VT, TVal, FVal);
3523 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3524 // and do the comparison.
3525 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3527 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3528 // clean. Some of them require two CSELs to implement.
3529 ARM64CC::CondCode CC1, CC2;
3530 changeFPCCToARM64CC(CC, CC1, CC2);
3531 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3532 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3534 // If we need a second CSEL, emit it, using the output of the first as the
3535 // RHS. We're effectively OR'ing the two CC's together.
3536 if (CC2 != ARM64CC::AL) {
3537 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3538 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3541 // Otherwise, return the output of the first CSEL.
3545 SDValue ARM64TargetLowering::LowerJumpTable(SDValue Op,
3546 SelectionDAG &DAG) const {
3547 // Jump table entries as PC relative offsets. No additional tweaking
3548 // is necessary here. Just get the address of the jump table.
3549 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3550 EVT PtrVT = getPointerTy();
3553 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3554 !Subtarget->isTargetMachO()) {
3555 const unsigned char MO_NC = ARM64II::MO_NC;
3557 ARM64ISD::WrapperLarge, DL, PtrVT,
3558 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G3),
3559 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G2 | MO_NC),
3560 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G1 | MO_NC),
3561 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G0 | MO_NC));
3564 SDValue Hi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_PAGE);
3565 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3566 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3567 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3568 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3571 SDValue ARM64TargetLowering::LowerConstantPool(SDValue Op,
3572 SelectionDAG &DAG) const {
3573 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3574 EVT PtrVT = getPointerTy();
3577 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3578 // Use the GOT for the large code model on iOS.
3579 if (Subtarget->isTargetMachO()) {
3580 SDValue GotAddr = DAG.getTargetConstantPool(
3581 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3583 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
3586 const unsigned char MO_NC = ARM64II::MO_NC;
3588 ARM64ISD::WrapperLarge, DL, PtrVT,
3589 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3590 CP->getOffset(), ARM64II::MO_G3),
3591 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3592 CP->getOffset(), ARM64II::MO_G2 | MO_NC),
3593 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3594 CP->getOffset(), ARM64II::MO_G1 | MO_NC),
3595 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3596 CP->getOffset(), ARM64II::MO_G0 | MO_NC));
3598 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
3599 // ELF, the only valid one on Darwin.
3601 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3602 CP->getOffset(), ARM64II::MO_PAGE);
3603 SDValue Lo = DAG.getTargetConstantPool(
3604 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3605 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3607 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3608 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3612 SDValue ARM64TargetLowering::LowerBlockAddress(SDValue Op,
3613 SelectionDAG &DAG) const {
3614 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
3615 EVT PtrVT = getPointerTy();
3617 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3618 !Subtarget->isTargetMachO()) {
3619 const unsigned char MO_NC = ARM64II::MO_NC;
3621 ARM64ISD::WrapperLarge, DL, PtrVT,
3622 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G3),
3623 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
3624 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
3625 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
3627 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGE);
3628 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGEOFF |
3630 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3631 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3635 SDValue ARM64TargetLowering::LowerDarwin_VASTART(SDValue Op,
3636 SelectionDAG &DAG) const {
3637 ARM64FunctionInfo *FuncInfo =
3638 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
3642 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3643 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3644 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
3645 MachinePointerInfo(SV), false, false, 0);
3648 SDValue ARM64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
3649 SelectionDAG &DAG) const {
3650 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3651 // Standard, section B.3.
3652 MachineFunction &MF = DAG.getMachineFunction();
3653 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
3656 SDValue Chain = Op.getOperand(0);
3657 SDValue VAList = Op.getOperand(1);
3658 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3659 SmallVector<SDValue, 4> MemOps;
3661 // void *__stack at offset 0
3663 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3664 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3665 MachinePointerInfo(SV), false, false, 8));
3667 // void *__gr_top at offset 8
3668 int GPRSize = FuncInfo->getVarArgsGPRSize();
3670 SDValue GRTop, GRTopAddr;
3672 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3673 DAG.getConstant(8, getPointerTy()));
3675 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), getPointerTy());
3676 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
3677 DAG.getConstant(GPRSize, getPointerTy()));
3679 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3680 MachinePointerInfo(SV, 8), false, false, 8));
3683 // void *__vr_top at offset 16
3684 int FPRSize = FuncInfo->getVarArgsFPRSize();
3686 SDValue VRTop, VRTopAddr;
3687 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3688 DAG.getConstant(16, getPointerTy()));
3690 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), getPointerTy());
3691 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
3692 DAG.getConstant(FPRSize, getPointerTy()));
3694 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3695 MachinePointerInfo(SV, 16), false, false, 8));
3698 // int __gr_offs at offset 24
3699 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3700 DAG.getConstant(24, getPointerTy()));
3701 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
3702 GROffsAddr, MachinePointerInfo(SV, 24), false,
3705 // int __vr_offs at offset 28
3706 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3707 DAG.getConstant(28, getPointerTy()));
3708 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
3709 VROffsAddr, MachinePointerInfo(SV, 28), false,
3712 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
3716 SDValue ARM64TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
3717 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
3718 : LowerAAPCS_VASTART(Op, DAG);
3721 SDValue ARM64TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
3722 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
3724 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
3725 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3726 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
3728 return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op), Op.getOperand(1),
3729 Op.getOperand(2), DAG.getConstant(VaListSize, MVT::i32),
3730 8, false, false, MachinePointerInfo(DestSV),
3731 MachinePointerInfo(SrcSV));
3734 SDValue ARM64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
3735 assert(Subtarget->isTargetDarwin() &&
3736 "automatic va_arg instruction only works on Darwin");
3738 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3739 EVT VT = Op.getValueType();
3741 SDValue Chain = Op.getOperand(0);
3742 SDValue Addr = Op.getOperand(1);
3743 unsigned Align = Op.getConstantOperandVal(3);
3745 SDValue VAList = DAG.getLoad(getPointerTy(), DL, Chain, Addr,
3746 MachinePointerInfo(V), false, false, false, 0);
3747 Chain = VAList.getValue(1);
3750 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
3751 VAList = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3752 DAG.getConstant(Align - 1, getPointerTy()));
3753 VAList = DAG.getNode(ISD::AND, DL, getPointerTy(), VAList,
3754 DAG.getConstant(-(int64_t)Align, getPointerTy()));
3757 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
3758 uint64_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
3760 // Scalar integer and FP values smaller than 64 bits are implicitly extended
3761 // up to 64 bits. At the very least, we have to increase the striding of the
3762 // vaargs list to match this, and for FP values we need to introduce
3763 // FP_ROUND nodes as well.
3764 if (VT.isInteger() && !VT.isVector())
3766 bool NeedFPTrunc = false;
3767 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
3772 // Increment the pointer, VAList, to the next vaarg
3773 SDValue VANext = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3774 DAG.getConstant(ArgSize, getPointerTy()));
3775 // Store the incremented VAList to the legalized pointer
3776 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
3779 // Load the actual argument out of the pointer VAList
3781 // Load the value as an f64.
3782 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
3783 MachinePointerInfo(), false, false, false, 0);
3784 // Round the value down to an f32.
3785 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
3786 DAG.getIntPtrConstant(1));
3787 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
3788 // Merge the rounded value with the chain output of the load.
3789 return DAG.getMergeValues(Ops, 2, DL);
3792 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
3796 SDValue ARM64TargetLowering::LowerFRAMEADDR(SDValue Op,
3797 SelectionDAG &DAG) const {
3798 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3799 MFI->setFrameAddressIsTaken(true);
3801 EVT VT = Op.getValueType();
3803 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3804 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, ARM64::FP, VT);
3806 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
3807 MachinePointerInfo(), false, false, false, 0);
3811 SDValue ARM64TargetLowering::LowerRETURNADDR(SDValue Op,
3812 SelectionDAG &DAG) const {
3813 MachineFunction &MF = DAG.getMachineFunction();
3814 MachineFrameInfo *MFI = MF.getFrameInfo();
3815 MFI->setReturnAddressIsTaken(true);
3817 EVT VT = Op.getValueType();
3819 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3821 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3822 SDValue Offset = DAG.getConstant(8, getPointerTy());
3823 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
3824 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
3825 MachinePointerInfo(), false, false, false, 0);
3828 // Return LR, which contains the return address. Mark it an implicit live-in.
3829 unsigned Reg = MF.addLiveIn(ARM64::LR, &ARM64::GPR64RegClass);
3830 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
3833 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3834 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3835 SDValue ARM64TargetLowering::LowerShiftRightParts(SDValue Op,
3836 SelectionDAG &DAG) const {
3837 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3838 EVT VT = Op.getValueType();
3839 unsigned VTBits = VT.getSizeInBits();
3841 SDValue ShOpLo = Op.getOperand(0);
3842 SDValue ShOpHi = Op.getOperand(1);
3843 SDValue ShAmt = Op.getOperand(2);
3845 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3847 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3849 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3850 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3851 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3852 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3853 DAG.getConstant(VTBits, MVT::i64));
3854 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3856 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
3857 ISD::SETGE, dl, DAG);
3858 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3860 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3861 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3863 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3865 // ARM64 shifts larger than the register width are wrapped rather than
3866 // clamped, so we can't just emit "hi >> x".
3867 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3868 SDValue TrueValHi = Opc == ISD::SRA
3869 ? DAG.getNode(Opc, dl, VT, ShOpHi,
3870 DAG.getConstant(VTBits - 1, MVT::i64))
3871 : DAG.getConstant(0, VT);
3873 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
3875 SDValue Ops[2] = { Lo, Hi };
3876 return DAG.getMergeValues(Ops, 2, dl);
3879 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3880 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3881 SDValue ARM64TargetLowering::LowerShiftLeftParts(SDValue Op,
3882 SelectionDAG &DAG) const {
3883 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3884 EVT VT = Op.getValueType();
3885 unsigned VTBits = VT.getSizeInBits();
3887 SDValue ShOpLo = Op.getOperand(0);
3888 SDValue ShOpHi = Op.getOperand(1);
3889 SDValue ShAmt = Op.getOperand(2);
3892 assert(Op.getOpcode() == ISD::SHL_PARTS);
3893 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3894 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3895 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3896 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3897 DAG.getConstant(VTBits, MVT::i64));
3898 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3899 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3901 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3903 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
3904 ISD::SETGE, dl, DAG);
3905 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3906 SDValue Hi = DAG.getNode(ARM64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
3908 // ARM64 shifts of larger than register sizes are wrapped rather than clamped,
3909 // so we can't just emit "lo << a" if a is too big.
3910 SDValue TrueValLo = DAG.getConstant(0, VT);
3911 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3913 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3915 SDValue Ops[2] = { Lo, Hi };
3916 return DAG.getMergeValues(Ops, 2, dl);
3920 ARM64TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3921 // The ARM64 target doesn't support folding offsets into global addresses.
3925 bool ARM64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3926 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
3927 // FIXME: We should be able to handle f128 as well with a clever lowering.
3928 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
3932 return ARM64_AM::getFP64Imm(Imm) != -1;
3933 else if (VT == MVT::f32)
3934 return ARM64_AM::getFP32Imm(Imm) != -1;
3938 //===----------------------------------------------------------------------===//
3939 // ARM64 Optimization Hooks
3940 //===----------------------------------------------------------------------===//
3942 //===----------------------------------------------------------------------===//
3943 // ARM64 Inline Assembly Support
3944 //===----------------------------------------------------------------------===//
3946 // Table of Constraints
3947 // TODO: This is the current set of constraints supported by ARM for the
3948 // compiler, not all of them may make sense, e.g. S may be difficult to support.
3950 // r - A general register
3951 // w - An FP/SIMD register of some size in the range v0-v31
3952 // x - An FP/SIMD register of some size in the range v0-v15
3953 // I - Constant that can be used with an ADD instruction
3954 // J - Constant that can be used with a SUB instruction
3955 // K - Constant that can be used with a 32-bit logical instruction
3956 // L - Constant that can be used with a 64-bit logical instruction
3957 // M - Constant that can be used as a 32-bit MOV immediate
3958 // N - Constant that can be used as a 64-bit MOV immediate
3959 // Q - A memory reference with base register and no offset
3960 // S - A symbolic address
3961 // Y - Floating point constant zero
3962 // Z - Integer constant zero
3964 // Note that general register operands will be output using their 64-bit x
3965 // register name, whatever the size of the variable, unless the asm operand
3966 // is prefixed by the %w modifier. Floating-point and SIMD register operands
3967 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
3970 /// getConstraintType - Given a constraint letter, return the type of
3971 /// constraint it is for this target.
3972 ARM64TargetLowering::ConstraintType
3973 ARM64TargetLowering::getConstraintType(const std::string &Constraint) const {
3974 if (Constraint.size() == 1) {
3975 switch (Constraint[0]) {
3982 return C_RegisterClass;
3983 // An address with a single base register. Due to the way we
3984 // currently handle addresses it is the same as 'r'.
3989 return TargetLowering::getConstraintType(Constraint);
3992 /// Examine constraint type and operand type and determine a weight value.
3993 /// This object must already have been set up with the operand type
3994 /// and the current alternative constraint selected.
3995 TargetLowering::ConstraintWeight
3996 ARM64TargetLowering::getSingleConstraintMatchWeight(
3997 AsmOperandInfo &info, const char *constraint) const {
3998 ConstraintWeight weight = CW_Invalid;
3999 Value *CallOperandVal = info.CallOperandVal;
4000 // If we don't have a value, we can't do a match,
4001 // but allow it at the lowest weight.
4002 if (CallOperandVal == NULL)
4004 Type *type = CallOperandVal->getType();
4005 // Look at the constraint type.
4006 switch (*constraint) {
4008 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
4012 if (type->isFloatingPointTy() || type->isVectorTy())
4013 weight = CW_Register;
4016 weight = CW_Constant;
4022 std::pair<unsigned, const TargetRegisterClass *>
4023 ARM64TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4025 if (Constraint.size() == 1) {
4026 switch (Constraint[0]) {
4028 if (VT.getSizeInBits() == 64)
4029 return std::make_pair(0U, &ARM64::GPR64commonRegClass);
4030 return std::make_pair(0U, &ARM64::GPR32commonRegClass);
4033 return std::make_pair(0U, &ARM64::FPR32RegClass);
4034 if (VT.getSizeInBits() == 64)
4035 return std::make_pair(0U, &ARM64::FPR64RegClass);
4036 if (VT.getSizeInBits() == 128)
4037 return std::make_pair(0U, &ARM64::FPR128RegClass);
4039 // The instructions that this constraint is designed for can
4040 // only take 128-bit registers so just use that regclass.
4042 if (VT.getSizeInBits() == 128)
4043 return std::make_pair(0U, &ARM64::FPR128_loRegClass);
4047 if (StringRef("{cc}").equals_lower(Constraint))
4048 return std::make_pair(unsigned(ARM64::CPSR), &ARM64::CCRRegClass);
4050 // Use the default implementation in TargetLowering to convert the register
4051 // constraint into a member of a register class.
4052 std::pair<unsigned, const TargetRegisterClass *> Res;
4053 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4055 // Not found as a standard register?
4056 if (Res.second == 0) {
4057 unsigned Size = Constraint.size();
4058 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
4059 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
4060 const std::string Reg =
4061 std::string(&Constraint[2], &Constraint[Size - 1]);
4062 int RegNo = atoi(Reg.c_str());
4063 if (RegNo >= 0 && RegNo <= 31) {
4064 // v0 - v31 are aliases of q0 - q31.
4065 // By default we'll emit v0-v31 for this unless there's a modifier where
4066 // we'll emit the correct register as well.
4067 Res.first = ARM64::FPR128RegClass.getRegister(RegNo);
4068 Res.second = &ARM64::FPR128RegClass;
4076 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4077 /// vector. If it is invalid, don't add anything to Ops.
4078 void ARM64TargetLowering::LowerAsmOperandForConstraint(
4079 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
4080 SelectionDAG &DAG) const {
4081 SDValue Result(0, 0);
4083 // Currently only support length 1 constraints.
4084 if (Constraint.length() != 1)
4087 char ConstraintLetter = Constraint[0];
4088 switch (ConstraintLetter) {
4092 // This set of constraints deal with valid constants for various instructions.
4093 // Validate and return a target constant for them if we can.
4095 // 'z' maps to xzr or wzr so it needs an input of 0.
4096 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4097 if (!C || C->getZExtValue() != 0)
4100 if (Op.getValueType() == MVT::i64)
4101 Result = DAG.getRegister(ARM64::XZR, MVT::i64);
4103 Result = DAG.getRegister(ARM64::WZR, MVT::i32);
4113 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4117 // Grab the value and do some validation.
4118 uint64_t CVal = C->getZExtValue();
4119 switch (ConstraintLetter) {
4120 // The I constraint applies only to simple ADD or SUB immediate operands:
4121 // i.e. 0 to 4095 with optional shift by 12
4122 // The J constraint applies only to ADD or SUB immediates that would be
4123 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
4124 // instruction [or vice versa], in other words -1 to -4095 with optional
4125 // left shift by 12.
4127 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
4131 uint64_t NVal = -C->getSExtValue();
4132 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal))
4136 // The K and L constraints apply *only* to logical immediates, including
4137 // what used to be the MOVI alias for ORR (though the MOVI alias has now
4138 // been removed and MOV should be used). So these constraints have to
4139 // distinguish between bit patterns that are valid 32-bit or 64-bit
4140 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
4141 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
4144 if (ARM64_AM::isLogicalImmediate(CVal, 32))
4148 if (ARM64_AM::isLogicalImmediate(CVal, 64))
4151 // The M and N constraints are a superset of K and L respectively, for use
4152 // with the MOV (immediate) alias. As well as the logical immediates they
4153 // also match 32 or 64-bit immediates that can be loaded either using a
4154 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
4155 // (M) or 64-bit 0x1234000000000000 (N) etc.
4156 // As a note some of this code is liberally stolen from the asm parser.
4158 if (!isUInt<32>(CVal))
4160 if (ARM64_AM::isLogicalImmediate(CVal, 32))
4162 if ((CVal & 0xFFFF) == CVal)
4164 if ((CVal & 0xFFFF0000ULL) == CVal)
4166 uint64_t NCVal = ~(uint32_t)CVal;
4167 if ((NCVal & 0xFFFFULL) == NCVal)
4169 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4174 if (ARM64_AM::isLogicalImmediate(CVal, 64))
4176 if ((CVal & 0xFFFFULL) == CVal)
4178 if ((CVal & 0xFFFF0000ULL) == CVal)
4180 if ((CVal & 0xFFFF00000000ULL) == CVal)
4182 if ((CVal & 0xFFFF000000000000ULL) == CVal)
4184 uint64_t NCVal = ~CVal;
4185 if ((NCVal & 0xFFFFULL) == NCVal)
4187 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4189 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
4191 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
4199 // All assembler immediates are 64-bit integers.
4200 Result = DAG.getTargetConstant(CVal, MVT::i64);
4204 if (Result.getNode()) {
4205 Ops.push_back(Result);
4209 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4212 //===----------------------------------------------------------------------===//
4213 // ARM64 Advanced SIMD Support
4214 //===----------------------------------------------------------------------===//
4216 /// WidenVector - Given a value in the V64 register class, produce the
4217 /// equivalent value in the V128 register class.
4218 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
4219 EVT VT = V64Reg.getValueType();
4220 unsigned NarrowSize = VT.getVectorNumElements();
4221 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4222 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
4225 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
4226 V64Reg, DAG.getConstant(0, MVT::i32));
4229 /// getExtFactor - Determine the adjustment factor for the position when
4230 /// generating an "extract from vector registers" instruction.
4231 static unsigned getExtFactor(SDValue &V) {
4232 EVT EltType = V.getValueType().getVectorElementType();
4233 return EltType.getSizeInBits() / 8;
4236 /// NarrowVector - Given a value in the V128 register class, produce the
4237 /// equivalent value in the V64 register class.
4238 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
4239 EVT VT = V128Reg.getValueType();
4240 unsigned WideSize = VT.getVectorNumElements();
4241 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4242 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
4245 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, NarrowTy, V128Reg);
4248 // Gather data to see if the operation can be modelled as a
4249 // shuffle in combination with VEXTs.
4250 SDValue ARM64TargetLowering::ReconstructShuffle(SDValue Op,
4251 SelectionDAG &DAG) const {
4253 EVT VT = Op.getValueType();
4254 unsigned NumElts = VT.getVectorNumElements();
4256 SmallVector<SDValue, 2> SourceVecs;
4257 SmallVector<unsigned, 2> MinElts;
4258 SmallVector<unsigned, 2> MaxElts;
4260 for (unsigned i = 0; i < NumElts; ++i) {
4261 SDValue V = Op.getOperand(i);
4262 if (V.getOpcode() == ISD::UNDEF)
4264 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4265 // A shuffle can only come from building a vector from various
4266 // elements of other vectors.
4270 // Record this extraction against the appropriate vector if possible...
4271 SDValue SourceVec = V.getOperand(0);
4272 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4273 bool FoundSource = false;
4274 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4275 if (SourceVecs[j] == SourceVec) {
4276 if (MinElts[j] > EltNo)
4278 if (MaxElts[j] < EltNo)
4285 // Or record a new source if not...
4287 SourceVecs.push_back(SourceVec);
4288 MinElts.push_back(EltNo);
4289 MaxElts.push_back(EltNo);
4293 // Currently only do something sane when at most two source vectors
4295 if (SourceVecs.size() > 2)
4298 SDValue ShuffleSrcs[2] = { DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4299 int VEXTOffsets[2] = { 0, 0 };
4301 // This loop extracts the usage patterns of the source vectors
4302 // and prepares appropriate SDValues for a shuffle if possible.
4303 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4304 if (SourceVecs[i].getValueType() == VT) {
4305 // No VEXT necessary
4306 ShuffleSrcs[i] = SourceVecs[i];
4309 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4310 // It probably isn't worth padding out a smaller vector just to
4311 // break it down again in a shuffle.
4315 // Don't attempt to extract subvectors from BUILD_VECTOR sources
4316 // that expand or trunc the original value.
4317 // TODO: We can try to bitcast and ANY_EXTEND the result but
4318 // we need to consider the cost of vector ANY_EXTEND, and the
4319 // legality of all the types.
4320 if (SourceVecs[i].getValueType().getVectorElementType() !=
4321 VT.getVectorElementType())
4324 // Since only 64-bit and 128-bit vectors are legal on ARM and
4325 // we've eliminated the other cases...
4326 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2 * NumElts &&
4327 "unexpected vector sizes in ReconstructShuffle");
4329 if (MaxElts[i] - MinElts[i] >= NumElts) {
4330 // Span too large for a VEXT to cope
4334 if (MinElts[i] >= NumElts) {
4335 // The extraction can just take the second half
4336 VEXTOffsets[i] = NumElts;
4338 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
4339 DAG.getIntPtrConstant(NumElts));
4340 } else if (MaxElts[i] < NumElts) {
4341 // The extraction can just take the first half
4343 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4344 SourceVecs[i], DAG.getIntPtrConstant(0));
4346 // An actual VEXT is needed
4347 VEXTOffsets[i] = MinElts[i];
4348 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4349 SourceVecs[i], DAG.getIntPtrConstant(0));
4351 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
4352 DAG.getIntPtrConstant(NumElts));
4353 unsigned Imm = VEXTOffsets[i] * getExtFactor(VEXTSrc1);
4354 ShuffleSrcs[i] = DAG.getNode(ARM64ISD::EXT, dl, VT, VEXTSrc1, VEXTSrc2,
4355 DAG.getConstant(Imm, MVT::i32));
4359 SmallVector<int, 8> Mask;
4361 for (unsigned i = 0; i < NumElts; ++i) {
4362 SDValue Entry = Op.getOperand(i);
4363 if (Entry.getOpcode() == ISD::UNDEF) {
4368 SDValue ExtractVec = Entry.getOperand(0);
4370 cast<ConstantSDNode>(Op.getOperand(i).getOperand(1))->getSExtValue();
4371 if (ExtractVec == SourceVecs[0]) {
4372 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4374 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4378 // Final check before we try to produce nonsense...
4379 if (isShuffleMaskLegal(Mask, VT))
4380 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4386 // check if an EXT instruction can handle the shuffle mask when the
4387 // vector sources of the shuffle are the same.
4388 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4389 unsigned NumElts = VT.getVectorNumElements();
4391 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4397 // If this is a VEXT shuffle, the immediate value is the index of the first
4398 // element. The other shuffle indices must be the successive elements after
4400 unsigned ExpectedElt = Imm;
4401 for (unsigned i = 1; i < NumElts; ++i) {
4402 // Increment the expected index. If it wraps around, just follow it
4403 // back to index zero and keep going.
4405 if (ExpectedElt == NumElts)
4409 continue; // ignore UNDEF indices
4410 if (ExpectedElt != static_cast<unsigned>(M[i]))
4417 // check if an EXT instruction can handle the shuffle mask when the
4418 // vector sources of the shuffle are different.
4419 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
4421 unsigned NumElts = VT.getVectorNumElements();
4424 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4430 // If this is a VEXT shuffle, the immediate value is the index of the first
4431 // element. The other shuffle indices must be the successive elements after
4433 unsigned ExpectedElt = Imm;
4434 for (unsigned i = 1; i < NumElts; ++i) {
4435 // Increment the expected index. If it wraps around, it may still be
4436 // a VEXT but the source vectors must be swapped.
4438 if (ExpectedElt == NumElts * 2) {
4444 continue; // ignore UNDEF indices
4445 if (ExpectedElt != static_cast<unsigned>(M[i]))
4449 // Adjust the index value if the source operands will be swapped.
4456 /// isREVMask - Check if a vector shuffle corresponds to a REV
4457 /// instruction with the specified blocksize. (The order of the elements
4458 /// within each block of the vector is reversed.)
4459 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4460 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
4461 "Only possible block sizes for REV are: 16, 32, 64");
4463 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4467 unsigned NumElts = VT.getVectorNumElements();
4468 unsigned BlockElts = M[0] + 1;
4469 // If the first shuffle index is UNDEF, be optimistic.
4471 BlockElts = BlockSize / EltSz;
4473 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4476 for (unsigned i = 0; i < NumElts; ++i) {
4478 continue; // ignore UNDEF indices
4479 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
4486 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4487 unsigned NumElts = VT.getVectorNumElements();
4488 WhichResult = (M[0] == 0 ? 0 : 1);
4489 unsigned Idx = WhichResult * NumElts / 2;
4490 for (unsigned i = 0; i != NumElts; i += 2) {
4491 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4492 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
4500 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4501 unsigned NumElts = VT.getVectorNumElements();
4502 WhichResult = (M[0] == 0 ? 0 : 1);
4503 for (unsigned i = 0; i != NumElts; ++i) {
4505 continue; // ignore UNDEF indices
4506 if ((unsigned)M[i] != 2 * i + WhichResult)
4513 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4514 unsigned NumElts = VT.getVectorNumElements();
4515 WhichResult = (M[0] == 0 ? 0 : 1);
4516 for (unsigned i = 0; i < NumElts; i += 2) {
4517 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4518 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
4524 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
4525 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4526 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4527 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4528 unsigned NumElts = VT.getVectorNumElements();
4529 WhichResult = (M[0] == 0 ? 0 : 1);
4530 unsigned Idx = WhichResult * NumElts / 2;
4531 for (unsigned i = 0; i != NumElts; i += 2) {
4532 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4533 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
4541 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
4542 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4543 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4544 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4545 unsigned Half = VT.getVectorNumElements() / 2;
4546 WhichResult = (M[0] == 0 ? 0 : 1);
4547 for (unsigned j = 0; j != 2; ++j) {
4548 unsigned Idx = WhichResult;
4549 for (unsigned i = 0; i != Half; ++i) {
4550 int MIdx = M[i + j * Half];
4551 if (MIdx >= 0 && (unsigned)MIdx != Idx)
4560 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
4561 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4562 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4563 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4564 unsigned NumElts = VT.getVectorNumElements();
4565 WhichResult = (M[0] == 0 ? 0 : 1);
4566 for (unsigned i = 0; i < NumElts; i += 2) {
4567 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4568 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
4574 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4575 /// the specified operations to build the shuffle.
4576 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4577 SDValue RHS, SelectionDAG &DAG,
4579 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4580 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
4581 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
4584 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4593 OP_VUZPL, // VUZP, left result
4594 OP_VUZPR, // VUZP, right result
4595 OP_VZIPL, // VZIP, left result
4596 OP_VZIPR, // VZIP, right result
4597 OP_VTRNL, // VTRN, left result
4598 OP_VTRNR // VTRN, right result
4601 if (OpNum == OP_COPY) {
4602 if (LHSID == (1 * 9 + 2) * 9 + 3)
4604 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
4608 SDValue OpLHS, OpRHS;
4609 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4610 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4611 EVT VT = OpLHS.getValueType();
4615 llvm_unreachable("Unknown shuffle opcode!");
4617 // VREV divides the vector in half and swaps within the half.
4618 if (VT.getVectorElementType() == MVT::i32 ||
4619 VT.getVectorElementType() == MVT::f32)
4620 return DAG.getNode(ARM64ISD::REV64, dl, VT, OpLHS);
4621 // vrev <4 x i16> -> REV32
4622 if (VT.getVectorElementType() == MVT::i16)
4623 return DAG.getNode(ARM64ISD::REV32, dl, VT, OpLHS);
4624 // vrev <4 x i8> -> REV16
4625 assert(VT.getVectorElementType() == MVT::i8);
4626 return DAG.getNode(ARM64ISD::REV16, dl, VT, OpLHS);
4631 EVT EltTy = VT.getVectorElementType();
4633 if (EltTy == MVT::i8)
4634 Opcode = ARM64ISD::DUPLANE8;
4635 else if (EltTy == MVT::i16)
4636 Opcode = ARM64ISD::DUPLANE16;
4637 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
4638 Opcode = ARM64ISD::DUPLANE32;
4639 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
4640 Opcode = ARM64ISD::DUPLANE64;
4642 llvm_unreachable("Invalid vector element type?");
4644 if (VT.getSizeInBits() == 64)
4645 OpLHS = WidenVector(OpLHS, DAG);
4646 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, MVT::i64);
4647 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
4652 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
4653 return DAG.getNode(ARM64ISD::EXT, dl, VT, OpLHS, OpRHS,
4654 DAG.getConstant(Imm, MVT::i32));
4657 return DAG.getNode(ARM64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4659 return DAG.getNode(ARM64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4661 return DAG.getNode(ARM64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4663 return DAG.getNode(ARM64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4665 return DAG.getNode(ARM64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4667 return DAG.getNode(ARM64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4671 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
4672 SelectionDAG &DAG) {
4673 // Check to see if we can use the TBL instruction.
4674 SDValue V1 = Op.getOperand(0);
4675 SDValue V2 = Op.getOperand(1);
4678 EVT EltVT = Op.getValueType().getVectorElementType();
4679 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
4681 SmallVector<SDValue, 8> TBLMask;
4682 for (int Val : ShuffleMask) {
4683 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
4684 unsigned Offset = Byte + Val * BytesPerElt;
4685 TBLMask.push_back(DAG.getConstant(Offset, MVT::i32));
4689 MVT IndexVT = MVT::v8i8;
4690 unsigned IndexLen = 8;
4691 if (Op.getValueType().getSizeInBits() == 128) {
4692 IndexVT = MVT::v16i8;
4696 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
4697 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
4700 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
4702 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
4703 Shuffle = DAG.getNode(
4704 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4705 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4706 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4708 if (IndexLen == 8) {
4709 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
4710 Shuffle = DAG.getNode(
4711 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4712 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4713 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4715 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
4716 // cannot currently represent the register constraints on the input
4718 // Shuffle = DAG.getNode(ARM64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
4719 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4720 // &TBLMask[0], IndexLen));
4721 Shuffle = DAG.getNode(
4722 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4723 DAG.getConstant(Intrinsic::arm64_neon_tbl2, MVT::i32), V1Cst, V2Cst,
4724 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4727 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
4730 static unsigned getDUPLANEOp(EVT EltType) {
4731 if (EltType == MVT::i8)
4732 return ARM64ISD::DUPLANE8;
4733 if (EltType == MVT::i16)
4734 return ARM64ISD::DUPLANE16;
4735 if (EltType == MVT::i32 || EltType == MVT::f32)
4736 return ARM64ISD::DUPLANE32;
4737 if (EltType == MVT::i64 || EltType == MVT::f64)
4738 return ARM64ISD::DUPLANE64;
4740 llvm_unreachable("Invalid vector element type?");
4743 SDValue ARM64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4744 SelectionDAG &DAG) const {
4746 EVT VT = Op.getValueType();
4748 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4750 // Convert shuffles that are directly supported on NEON to target-specific
4751 // DAG nodes, instead of keeping them as shuffles and matching them again
4752 // during code selection. This is more efficient and avoids the possibility
4753 // of inconsistencies between legalization and selection.
4754 ArrayRef<int> ShuffleMask = SVN->getMask();
4756 SDValue V1 = Op.getOperand(0);
4757 SDValue V2 = Op.getOperand(1);
4759 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
4760 V1.getValueType().getSimpleVT())) {
4761 int Lane = SVN->getSplatIndex();
4762 // If this is undef splat, generate it via "just" vdup, if possible.
4766 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
4767 return DAG.getNode(ARM64ISD::DUP, dl, V1.getValueType(),
4769 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
4770 // constant. If so, we can just reference the lane's definition directly.
4771 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
4772 !isa<ConstantSDNode>(V1.getOperand(Lane)))
4773 return DAG.getNode(ARM64ISD::DUP, dl, VT, V1.getOperand(Lane));
4775 // Otherwise, duplicate from the lane of the input vector.
4776 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
4778 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
4779 // to make a vector of the same size as this SHUFFLE. We can ignore the
4780 // extract entirely, and canonicalise the concat using WidenVector.
4781 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
4782 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
4783 V1 = V1.getOperand(0);
4784 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
4785 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
4786 Lane -= Idx * VT.getVectorNumElements() / 2;
4787 V1 = WidenVector(V1.getOperand(Idx), DAG);
4788 } else if (VT.getSizeInBits() == 64)
4789 V1 = WidenVector(V1, DAG);
4791 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, MVT::i64));
4794 if (isREVMask(ShuffleMask, VT, 64))
4795 return DAG.getNode(ARM64ISD::REV64, dl, V1.getValueType(), V1, V2);
4796 if (isREVMask(ShuffleMask, VT, 32))
4797 return DAG.getNode(ARM64ISD::REV32, dl, V1.getValueType(), V1, V2);
4798 if (isREVMask(ShuffleMask, VT, 16))
4799 return DAG.getNode(ARM64ISD::REV16, dl, V1.getValueType(), V1, V2);
4801 bool ReverseEXT = false;
4803 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
4806 Imm *= getExtFactor(V1);
4807 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V2,
4808 DAG.getConstant(Imm, MVT::i32));
4809 } else if (V2->getOpcode() == ISD::UNDEF &&
4810 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
4811 Imm *= getExtFactor(V1);
4812 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V1,
4813 DAG.getConstant(Imm, MVT::i32));
4816 unsigned WhichResult;
4817 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
4818 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4819 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4821 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
4822 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4823 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4825 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
4826 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4827 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4830 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4831 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4832 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4834 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4835 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4836 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4838 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4839 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4840 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4843 // If the shuffle is not directly supported and it has 4 elements, use
4844 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4845 unsigned NumElts = VT.getVectorNumElements();
4847 unsigned PFIndexes[4];
4848 for (unsigned i = 0; i != 4; ++i) {
4849 if (ShuffleMask[i] < 0)
4852 PFIndexes[i] = ShuffleMask[i];
4855 // Compute the index in the perfect shuffle table.
4856 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
4857 PFIndexes[2] * 9 + PFIndexes[3];
4858 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4859 unsigned Cost = (PFEntry >> 30);
4862 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4865 return GenerateTBL(Op, ShuffleMask, DAG);
4868 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
4870 EVT VT = BVN->getValueType(0);
4871 APInt SplatBits, SplatUndef;
4872 unsigned SplatBitSize;
4874 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4875 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
4877 for (unsigned i = 0; i < NumSplats; ++i) {
4878 CnstBits <<= SplatBitSize;
4879 UndefBits <<= SplatBitSize;
4880 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
4881 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
4890 SDValue ARM64TargetLowering::LowerVectorAND(SDValue Op,
4891 SelectionDAG &DAG) const {
4892 BuildVectorSDNode *BVN =
4893 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
4894 SDValue LHS = Op.getOperand(0);
4896 EVT VT = Op.getValueType();
4901 APInt CnstBits(VT.getSizeInBits(), 0);
4902 APInt UndefBits(VT.getSizeInBits(), 0);
4903 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
4904 // We only have BIC vector immediate instruction, which is and-not.
4905 CnstBits = ~CnstBits;
4907 // We make use of a little bit of goto ickiness in order to avoid having to
4908 // duplicate the immediate matching logic for the undef toggled case.
4909 bool SecondTry = false;
4912 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
4913 CnstBits = CnstBits.zextOrTrunc(64);
4914 uint64_t CnstVal = CnstBits.getZExtValue();
4916 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
4917 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
4918 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4919 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4920 DAG.getConstant(CnstVal, MVT::i32),
4921 DAG.getConstant(0, MVT::i32));
4922 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4925 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
4926 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
4927 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4928 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4929 DAG.getConstant(CnstVal, MVT::i32),
4930 DAG.getConstant(8, MVT::i32));
4931 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4934 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
4935 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
4936 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4937 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4938 DAG.getConstant(CnstVal, MVT::i32),
4939 DAG.getConstant(16, MVT::i32));
4940 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4943 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
4944 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
4945 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4946 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4947 DAG.getConstant(CnstVal, MVT::i32),
4948 DAG.getConstant(24, MVT::i32));
4949 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4952 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
4953 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
4954 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4955 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4956 DAG.getConstant(CnstVal, MVT::i32),
4957 DAG.getConstant(0, MVT::i32));
4958 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4961 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
4962 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
4963 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4964 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4965 DAG.getConstant(CnstVal, MVT::i32),
4966 DAG.getConstant(8, MVT::i32));
4967 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4974 CnstBits = ~UndefBits;
4978 // We can always fall back to a non-immediate AND.
4983 // Specialized code to quickly find if PotentialBVec is a BuildVector that
4984 // consists of only the same constant int value, returned in reference arg
4986 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
4987 uint64_t &ConstVal) {
4988 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
4991 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
4994 EVT VT = Bvec->getValueType(0);
4995 unsigned NumElts = VT.getVectorNumElements();
4996 for (unsigned i = 1; i < NumElts; ++i)
4997 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
4999 ConstVal = FirstElt->getZExtValue();
5003 static unsigned getIntrinsicID(const SDNode *N) {
5004 unsigned Opcode = N->getOpcode();
5007 return Intrinsic::not_intrinsic;
5008 case ISD::INTRINSIC_WO_CHAIN: {
5009 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5010 if (IID < Intrinsic::num_intrinsics)
5012 return Intrinsic::not_intrinsic;
5017 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
5018 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
5019 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
5020 // Also, logical shift right -> sri, with the same structure.
5021 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
5022 EVT VT = N->getValueType(0);
5029 // Is the first op an AND?
5030 const SDValue And = N->getOperand(0);
5031 if (And.getOpcode() != ISD::AND)
5034 // Is the second op an shl or lshr?
5035 SDValue Shift = N->getOperand(1);
5036 // This will have been turned into: ARM64ISD::VSHL vector, #shift
5037 // or ARM64ISD::VLSHR vector, #shift
5038 unsigned ShiftOpc = Shift.getOpcode();
5039 if ((ShiftOpc != ARM64ISD::VSHL && ShiftOpc != ARM64ISD::VLSHR))
5041 bool IsShiftRight = ShiftOpc == ARM64ISD::VLSHR;
5043 // Is the shift amount constant?
5044 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
5048 // Is the and mask vector all constant?
5050 if (!isAllConstantBuildVector(And.getOperand(1), C1))
5053 // Is C1 == ~C2, taking into account how much one can shift elements of a
5055 uint64_t C2 = C2node->getZExtValue();
5056 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
5057 if (C2 > ElemSizeInBits)
5059 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
5060 if ((C1 & ElemMask) != (~C2 & ElemMask))
5063 SDValue X = And.getOperand(0);
5064 SDValue Y = Shift.getOperand(0);
5067 IsShiftRight ? Intrinsic::arm64_neon_vsri : Intrinsic::arm64_neon_vsli;
5069 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5070 DAG.getConstant(Intrin, MVT::i32), X, Y, Shift.getOperand(1));
5072 DEBUG(dbgs() << "arm64-lower: transformed: \n");
5073 DEBUG(N->dump(&DAG));
5074 DEBUG(dbgs() << "into: \n");
5075 DEBUG(ResultSLI->dump(&DAG));
5081 SDValue ARM64TargetLowering::LowerVectorOR(SDValue Op,
5082 SelectionDAG &DAG) const {
5083 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
5084 if (EnableARM64SlrGeneration) {
5085 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
5090 BuildVectorSDNode *BVN =
5091 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
5092 SDValue LHS = Op.getOperand(1);
5094 EVT VT = Op.getValueType();
5096 // OR commutes, so try swapping the operands.
5098 LHS = Op.getOperand(0);
5099 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5104 APInt CnstBits(VT.getSizeInBits(), 0);
5105 APInt UndefBits(VT.getSizeInBits(), 0);
5106 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5107 // We make use of a little bit of goto ickiness in order to avoid having to
5108 // duplicate the immediate matching logic for the undef toggled case.
5109 bool SecondTry = false;
5112 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5113 CnstBits = CnstBits.zextOrTrunc(64);
5114 uint64_t CnstVal = CnstBits.getZExtValue();
5116 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5117 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5118 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5119 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5120 DAG.getConstant(CnstVal, MVT::i32),
5121 DAG.getConstant(0, MVT::i32));
5122 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5125 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5126 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5127 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5128 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5129 DAG.getConstant(CnstVal, MVT::i32),
5130 DAG.getConstant(8, MVT::i32));
5131 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5134 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5135 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5136 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5137 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5138 DAG.getConstant(CnstVal, MVT::i32),
5139 DAG.getConstant(16, MVT::i32));
5140 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5143 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5144 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5145 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5146 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5147 DAG.getConstant(CnstVal, MVT::i32),
5148 DAG.getConstant(24, MVT::i32));
5149 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5152 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5153 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5154 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5155 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5156 DAG.getConstant(CnstVal, MVT::i32),
5157 DAG.getConstant(0, MVT::i32));
5158 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5161 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5162 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5163 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5164 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5165 DAG.getConstant(CnstVal, MVT::i32),
5166 DAG.getConstant(8, MVT::i32));
5167 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5174 CnstBits = UndefBits;
5178 // We can always fall back to a non-immediate OR.
5183 SDValue ARM64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
5184 SelectionDAG &DAG) const {
5185 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5187 EVT VT = Op.getValueType();
5189 APInt CnstBits(VT.getSizeInBits(), 0);
5190 APInt UndefBits(VT.getSizeInBits(), 0);
5191 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5192 // We make use of a little bit of goto ickiness in order to avoid having to
5193 // duplicate the immediate matching logic for the undef toggled case.
5194 bool SecondTry = false;
5197 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5198 CnstBits = CnstBits.zextOrTrunc(64);
5199 uint64_t CnstVal = CnstBits.getZExtValue();
5201 // Certain magic vector constants (used to express things like NOT
5202 // and NEG) are passed through unmodified. This allows codegen patterns
5203 // for these operations to match. Special-purpose patterns will lower
5204 // these immediates to MOVIs if it proves necessary.
5205 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
5208 // The many faces of MOVI...
5209 if (ARM64_AM::isAdvSIMDModImmType10(CnstVal)) {
5210 CnstVal = ARM64_AM::encodeAdvSIMDModImmType10(CnstVal);
5211 if (VT.getSizeInBits() == 128) {
5212 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::v2i64,
5213 DAG.getConstant(CnstVal, MVT::i32));
5214 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5217 // Support the V64 version via subregister insertion.
5218 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::f64,
5219 DAG.getConstant(CnstVal, MVT::i32));
5220 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5223 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5224 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5225 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5226 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5227 DAG.getConstant(CnstVal, MVT::i32),
5228 DAG.getConstant(0, MVT::i32));
5229 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5232 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5233 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5234 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5235 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5236 DAG.getConstant(CnstVal, MVT::i32),
5237 DAG.getConstant(8, MVT::i32));
5238 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5241 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5242 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5243 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5244 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5245 DAG.getConstant(CnstVal, MVT::i32),
5246 DAG.getConstant(16, MVT::i32));
5247 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5250 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5251 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5252 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5253 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5254 DAG.getConstant(CnstVal, MVT::i32),
5255 DAG.getConstant(24, MVT::i32));
5256 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5259 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5260 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5261 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5262 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5263 DAG.getConstant(CnstVal, MVT::i32),
5264 DAG.getConstant(0, MVT::i32));
5265 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5268 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5269 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5270 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5271 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5272 DAG.getConstant(CnstVal, MVT::i32),
5273 DAG.getConstant(8, MVT::i32));
5274 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5277 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
5278 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
5279 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5280 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
5281 DAG.getConstant(CnstVal, MVT::i32),
5282 DAG.getConstant(264, MVT::i32));
5283 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5286 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
5287 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
5288 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5289 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
5290 DAG.getConstant(CnstVal, MVT::i32),
5291 DAG.getConstant(272, MVT::i32));
5292 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5295 if (ARM64_AM::isAdvSIMDModImmType9(CnstVal)) {
5296 CnstVal = ARM64_AM::encodeAdvSIMDModImmType9(CnstVal);
5297 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
5298 SDValue Mov = DAG.getNode(ARM64ISD::MOVI, dl, MovTy,
5299 DAG.getConstant(CnstVal, MVT::i32));
5300 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5303 // The few faces of FMOV...
5304 if (ARM64_AM::isAdvSIMDModImmType11(CnstVal)) {
5305 CnstVal = ARM64_AM::encodeAdvSIMDModImmType11(CnstVal);
5306 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
5307 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MovTy,
5308 DAG.getConstant(CnstVal, MVT::i32));
5309 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5312 if (ARM64_AM::isAdvSIMDModImmType12(CnstVal) &&
5313 VT.getSizeInBits() == 128) {
5314 CnstVal = ARM64_AM::encodeAdvSIMDModImmType12(CnstVal);
5315 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MVT::v2f64,
5316 DAG.getConstant(CnstVal, MVT::i32));
5317 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5320 // The many faces of MVNI...
5322 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5323 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5324 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5325 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5326 DAG.getConstant(CnstVal, MVT::i32),
5327 DAG.getConstant(0, MVT::i32));
5328 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5331 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5332 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5333 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5334 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5335 DAG.getConstant(CnstVal, MVT::i32),
5336 DAG.getConstant(8, MVT::i32));
5337 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5340 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5341 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5342 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5343 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5344 DAG.getConstant(CnstVal, MVT::i32),
5345 DAG.getConstant(16, MVT::i32));
5346 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5349 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5350 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5351 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5352 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5353 DAG.getConstant(CnstVal, MVT::i32),
5354 DAG.getConstant(24, MVT::i32));
5355 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5358 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5359 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5360 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5361 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5362 DAG.getConstant(CnstVal, MVT::i32),
5363 DAG.getConstant(0, MVT::i32));
5364 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5367 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5368 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5369 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5370 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5371 DAG.getConstant(CnstVal, MVT::i32),
5372 DAG.getConstant(8, MVT::i32));
5373 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5376 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
5377 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
5378 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5379 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
5380 DAG.getConstant(CnstVal, MVT::i32),
5381 DAG.getConstant(264, MVT::i32));
5382 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5385 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
5386 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
5387 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5388 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
5389 DAG.getConstant(CnstVal, MVT::i32),
5390 DAG.getConstant(272, MVT::i32));
5391 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5398 CnstBits = UndefBits;
5403 // Scan through the operands to find some interesting properties we can
5405 // 1) If only one value is used, we can use a DUP, or
5406 // 2) if only the low element is not undef, we can just insert that, or
5407 // 3) if only one constant value is used (w/ some non-constant lanes),
5408 // we can splat the constant value into the whole vector then fill
5409 // in the non-constant lanes.
5410 // 4) FIXME: If different constant values are used, but we can intelligently
5411 // select the values we'll be overwriting for the non-constant
5412 // lanes such that we can directly materialize the vector
5413 // some other way (MOVI, e.g.), we can be sneaky.
5414 unsigned NumElts = VT.getVectorNumElements();
5415 bool isOnlyLowElement = true;
5416 bool usesOnlyOneValue = true;
5417 bool usesOnlyOneConstantValue = true;
5418 bool isConstant = true;
5419 unsigned NumConstantLanes = 0;
5421 SDValue ConstantValue;
5422 for (unsigned i = 0; i < NumElts; ++i) {
5423 SDValue V = Op.getOperand(i);
5424 if (V.getOpcode() == ISD::UNDEF)
5427 isOnlyLowElement = false;
5428 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5431 if (isa<ConstantSDNode>(V)) {
5433 if (!ConstantValue.getNode())
5435 else if (ConstantValue != V)
5436 usesOnlyOneConstantValue = false;
5439 if (!Value.getNode())
5441 else if (V != Value)
5442 usesOnlyOneValue = false;
5445 if (!Value.getNode())
5446 return DAG.getUNDEF(VT);
5448 if (isOnlyLowElement)
5449 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5451 // Use DUP for non-constant splats. For f32 constant splats, reduce to
5452 // i32 and try again.
5453 if (usesOnlyOneValue) {
5455 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5456 Value.getValueType() != VT)
5457 return DAG.getNode(ARM64ISD::DUP, dl, VT, Value);
5459 // This is actually a DUPLANExx operation, which keeps everything vectory.
5461 // DUPLANE works on 128-bit vectors, widen it if necessary.
5462 SDValue Lane = Value.getOperand(1);
5463 Value = Value.getOperand(0);
5464 if (Value.getValueType().getSizeInBits() == 64)
5465 Value = WidenVector(Value, DAG);
5467 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
5468 return DAG.getNode(Opcode, dl, VT, Value, Lane);
5471 if (VT.getVectorElementType().isFloatingPoint()) {
5472 SmallVector<SDValue, 8> Ops;
5474 (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
5475 for (unsigned i = 0; i < NumElts; ++i)
5476 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
5477 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
5478 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
5479 Val = LowerBUILD_VECTOR(Val, DAG);
5481 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5485 // If there was only one constant value used and for more than one lane,
5486 // start by splatting that value, then replace the non-constant lanes. This
5487 // is better than the default, which will perform a separate initialization
5489 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
5490 SDValue Val = DAG.getNode(ARM64ISD::DUP, dl, VT, ConstantValue);
5491 // Now insert the non-constant lanes.
5492 for (unsigned i = 0; i < NumElts; ++i) {
5493 SDValue V = Op.getOperand(i);
5494 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5495 if (!isa<ConstantSDNode>(V)) {
5496 // Note that type legalization likely mucked about with the VT of the
5497 // source operand, so we may have to convert it here before inserting.
5498 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
5504 // If all elements are constants and the case above didn't get hit, fall back
5505 // to the default expansion, which will generate a load from the constant
5510 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5512 SDValue shuffle = ReconstructShuffle(Op, DAG);
5513 if (shuffle != SDValue())
5517 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5518 // know the default expansion would otherwise fall back on something even
5519 // worse. For a vector with one or two non-undef values, that's
5520 // scalar_to_vector for the elements followed by a shuffle (provided the
5521 // shuffle is valid for the target) and materialization element by element
5522 // on the stack followed by a load for everything else.
5523 if (!isConstant && !usesOnlyOneValue) {
5524 SDValue Vec = DAG.getUNDEF(VT);
5525 SDValue Op0 = Op.getOperand(0);
5526 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
5528 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
5529 // a) Avoid a RMW dependency on the full vector register, and
5530 // b) Allow the register coalescer to fold away the copy if the
5531 // value is already in an S or D register.
5532 if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
5533 unsigned SubIdx = ElemSize == 32 ? ARM64::ssub : ARM64::dsub;
5535 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
5536 DAG.getTargetConstant(SubIdx, MVT::i32));
5537 Vec = SDValue(N, 0);
5540 for (; i < NumElts; ++i) {
5541 SDValue V = Op.getOperand(i);
5542 if (V.getOpcode() == ISD::UNDEF)
5544 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5545 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5550 // Just use the default expansion. We failed to find a better alternative.
5554 SDValue ARM64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
5555 SelectionDAG &DAG) const {
5556 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
5558 // Check for non-constant lane.
5559 if (!isa<ConstantSDNode>(Op.getOperand(2)))
5562 EVT VT = Op.getOperand(0).getValueType();
5564 // Insertion/extraction are legal for V128 types.
5565 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5566 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5569 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5570 VT != MVT::v1i64 && VT != MVT::v2f32)
5573 // For V64 types, we perform insertion by expanding the value
5574 // to a V128 type and perform the insertion on that.
5576 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5577 EVT WideTy = WideVec.getValueType();
5579 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
5580 Op.getOperand(1), Op.getOperand(2));
5581 // Re-narrow the resultant vector.
5582 return NarrowVector(Node, DAG);
5585 SDValue ARM64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5586 SelectionDAG &DAG) const {
5587 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
5589 // Check for non-constant lane.
5590 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5593 EVT VT = Op.getOperand(0).getValueType();
5595 // Insertion/extraction are legal for V128 types.
5596 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5597 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5600 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5601 VT != MVT::v1i64 && VT != MVT::v2f32)
5604 // For V64 types, we perform extraction by expanding the value
5605 // to a V128 type and perform the extraction on that.
5607 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5608 EVT WideTy = WideVec.getValueType();
5610 EVT ExtrTy = WideTy.getVectorElementType();
5611 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
5614 // For extractions, we just return the result directly.
5615 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
5619 SDValue ARM64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
5620 SelectionDAG &DAG) const {
5621 EVT VT = Op.getOperand(0).getValueType();
5627 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5630 unsigned Val = Cst->getZExtValue();
5632 unsigned Size = Op.getValueType().getSizeInBits();
5636 return DAG.getTargetExtractSubreg(ARM64::bsub, dl, Op.getValueType(),
5639 return DAG.getTargetExtractSubreg(ARM64::hsub, dl, Op.getValueType(),
5642 return DAG.getTargetExtractSubreg(ARM64::ssub, dl, Op.getValueType(),
5645 return DAG.getTargetExtractSubreg(ARM64::dsub, dl, Op.getValueType(),
5648 llvm_unreachable("Unexpected vector type in extract_subvector!");
5651 // If this is extracting the upper 64-bits of a 128-bit vector, we match
5653 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
5659 bool ARM64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5661 if (VT.getVectorNumElements() == 4 &&
5662 (VT.is128BitVector() || VT.is64BitVector())) {
5663 unsigned PFIndexes[4];
5664 for (unsigned i = 0; i != 4; ++i) {
5668 PFIndexes[i] = M[i];
5671 // Compute the index in the perfect shuffle table.
5672 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5673 PFIndexes[2] * 9 + PFIndexes[3];
5674 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5675 unsigned Cost = (PFEntry >> 30);
5682 unsigned Imm, WhichResult;
5684 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
5685 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
5686 isEXTMask(M, VT, ReverseVEXT, Imm) ||
5687 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
5688 isTRNMask(M, VT, WhichResult) || isUZPMask(M, VT, WhichResult) ||
5689 isZIPMask(M, VT, WhichResult) ||
5690 isTRN_v_undef_Mask(M, VT, WhichResult) ||
5691 isUZP_v_undef_Mask(M, VT, WhichResult) ||
5692 isZIP_v_undef_Mask(M, VT, WhichResult));
5695 /// getVShiftImm - Check if this is a valid build_vector for the immediate
5696 /// operand of a vector shift operation, where all the elements of the
5697 /// build_vector must have the same constant integer value.
5698 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
5699 // Ignore bit_converts.
5700 while (Op.getOpcode() == ISD::BITCAST)
5701 Op = Op.getOperand(0);
5702 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
5703 APInt SplatBits, SplatUndef;
5704 unsigned SplatBitSize;
5706 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
5707 HasAnyUndefs, ElementBits) ||
5708 SplatBitSize > ElementBits)
5710 Cnt = SplatBits.getSExtValue();
5714 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
5715 /// operand of a vector shift left operation. That value must be in the range:
5716 /// 0 <= Value < ElementBits for a left shift; or
5717 /// 0 <= Value <= ElementBits for a long left shift.
5718 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
5719 assert(VT.isVector() && "vector shift count is not a vector type");
5720 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5721 if (!getVShiftImm(Op, ElementBits, Cnt))
5723 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
5726 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
5727 /// operand of a vector shift right operation. For a shift opcode, the value
5728 /// is positive, but for an intrinsic the value count must be negative. The
5729 /// absolute value must be in the range:
5730 /// 1 <= |Value| <= ElementBits for a right shift; or
5731 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
5732 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
5734 assert(VT.isVector() && "vector shift count is not a vector type");
5735 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5736 if (!getVShiftImm(Op, ElementBits, Cnt))
5740 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
5743 SDValue ARM64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
5744 SelectionDAG &DAG) const {
5745 EVT VT = Op.getValueType();
5749 if (!Op.getOperand(1).getValueType().isVector())
5751 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5753 switch (Op.getOpcode()) {
5755 llvm_unreachable("unexpected shift opcode");
5758 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
5759 return DAG.getNode(ARM64ISD::VSHL, SDLoc(Op), VT, Op.getOperand(0),
5760 DAG.getConstant(Cnt, MVT::i32));
5761 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5762 DAG.getConstant(Intrinsic::arm64_neon_ushl, MVT::i32),
5763 Op.getOperand(0), Op.getOperand(1));
5766 // Right shift immediate
5767 if (isVShiftRImm(Op.getOperand(1), VT, false, false, Cnt) &&
5770 (Op.getOpcode() == ISD::SRA) ? ARM64ISD::VASHR : ARM64ISD::VLSHR;
5771 return DAG.getNode(Opc, SDLoc(Op), VT, Op.getOperand(0),
5772 DAG.getConstant(Cnt, MVT::i32));
5775 // Right shift register. Note, there is not a shift right register
5776 // instruction, but the shift left register instruction takes a signed
5777 // value, where negative numbers specify a right shift.
5778 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::arm64_neon_sshl
5779 : Intrinsic::arm64_neon_ushl;
5780 // negate the shift amount
5781 SDValue NegShift = DAG.getNode(ARM64ISD::NEG, DL, VT, Op.getOperand(1));
5782 SDValue NegShiftLeft =
5783 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5784 DAG.getConstant(Opc, MVT::i32), Op.getOperand(0), NegShift);
5785 return NegShiftLeft;
5791 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
5792 ARM64CC::CondCode CC, bool NoNans, EVT VT,
5793 SDLoc dl, SelectionDAG &DAG) {
5794 EVT SrcVT = LHS.getValueType();
5796 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
5797 APInt CnstBits(VT.getSizeInBits(), 0);
5798 APInt UndefBits(VT.getSizeInBits(), 0);
5799 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
5800 bool IsZero = IsCnst && (CnstBits == 0);
5802 if (SrcVT.getVectorElementType().isFloatingPoint()) {
5809 Fcmeq = DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5811 Fcmeq = DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5812 return DAG.getNode(ARM64ISD::NOT, dl, VT, Fcmeq);
5816 return DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5817 return DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5820 return DAG.getNode(ARM64ISD::FCMGEz, dl, VT, LHS);
5821 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, LHS, RHS);
5824 return DAG.getNode(ARM64ISD::FCMGTz, dl, VT, LHS);
5825 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, LHS, RHS);
5828 return DAG.getNode(ARM64ISD::FCMLEz, dl, VT, LHS);
5829 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, RHS, LHS);
5833 // If we ignore NaNs then we can use to the MI implementation.
5837 return DAG.getNode(ARM64ISD::FCMLTz, dl, VT, LHS);
5838 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, RHS, LHS);
5848 Cmeq = DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5850 Cmeq = DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5851 return DAG.getNode(ARM64ISD::NOT, dl, VT, Cmeq);
5855 return DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5856 return DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5859 return DAG.getNode(ARM64ISD::CMGEz, dl, VT, LHS);
5860 return DAG.getNode(ARM64ISD::CMGE, dl, VT, LHS, RHS);
5863 return DAG.getNode(ARM64ISD::CMGTz, dl, VT, LHS);
5864 return DAG.getNode(ARM64ISD::CMGT, dl, VT, LHS, RHS);
5867 return DAG.getNode(ARM64ISD::CMLEz, dl, VT, LHS);
5868 return DAG.getNode(ARM64ISD::CMGE, dl, VT, RHS, LHS);
5870 return DAG.getNode(ARM64ISD::CMHS, dl, VT, RHS, LHS);
5872 return DAG.getNode(ARM64ISD::CMHI, dl, VT, RHS, LHS);
5875 return DAG.getNode(ARM64ISD::CMLTz, dl, VT, LHS);
5876 return DAG.getNode(ARM64ISD::CMGT, dl, VT, RHS, LHS);
5878 return DAG.getNode(ARM64ISD::CMHI, dl, VT, LHS, RHS);
5880 return DAG.getNode(ARM64ISD::CMHS, dl, VT, LHS, RHS);
5884 SDValue ARM64TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
5885 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5886 SDValue LHS = Op.getOperand(0);
5887 SDValue RHS = Op.getOperand(1);
5890 if (LHS.getValueType().getVectorElementType().isInteger()) {
5891 assert(LHS.getValueType() == RHS.getValueType());
5892 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
5893 return EmitVectorComparison(LHS, RHS, ARM64CC, false, Op.getValueType(), dl,
5897 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
5898 LHS.getValueType().getVectorElementType() == MVT::f64);
5900 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
5901 // clean. Some of them require two branches to implement.
5902 ARM64CC::CondCode CC1, CC2;
5903 changeFPCCToARM64CC(CC, CC1, CC2);
5905 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
5907 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, Op.getValueType(), dl, DAG);
5908 if (!Cmp1.getNode())
5911 if (CC2 != ARM64CC::AL) {
5913 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, Op.getValueType(), dl, DAG);
5914 if (!Cmp2.getNode())
5917 return DAG.getNode(ISD::OR, dl, Cmp1.getValueType(), Cmp1, Cmp2);
5923 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
5924 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
5925 /// specified in the intrinsic calls.
5926 bool ARM64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5928 unsigned Intrinsic) const {
5929 switch (Intrinsic) {
5930 case Intrinsic::arm64_neon_ld2:
5931 case Intrinsic::arm64_neon_ld3:
5932 case Intrinsic::arm64_neon_ld4:
5933 case Intrinsic::arm64_neon_ld2lane:
5934 case Intrinsic::arm64_neon_ld3lane:
5935 case Intrinsic::arm64_neon_ld4lane:
5936 case Intrinsic::arm64_neon_ld2r:
5937 case Intrinsic::arm64_neon_ld3r:
5938 case Intrinsic::arm64_neon_ld4r: {
5939 Info.opc = ISD::INTRINSIC_W_CHAIN;
5940 // Conservatively set memVT to the entire set of vectors loaded.
5941 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
5942 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5943 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5946 Info.vol = false; // volatile loads with NEON intrinsics not supported
5947 Info.readMem = true;
5948 Info.writeMem = false;
5951 case Intrinsic::arm64_neon_st2:
5952 case Intrinsic::arm64_neon_st3:
5953 case Intrinsic::arm64_neon_st4:
5954 case Intrinsic::arm64_neon_st2lane:
5955 case Intrinsic::arm64_neon_st3lane:
5956 case Intrinsic::arm64_neon_st4lane: {
5957 Info.opc = ISD::INTRINSIC_VOID;
5958 // Conservatively set memVT to the entire set of vectors stored.
5959 unsigned NumElts = 0;
5960 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5961 Type *ArgTy = I.getArgOperand(ArgI)->getType();
5962 if (!ArgTy->isVectorTy())
5964 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
5966 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5967 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5970 Info.vol = false; // volatile stores with NEON intrinsics not supported
5971 Info.readMem = false;
5972 Info.writeMem = true;
5975 case Intrinsic::arm64_ldxr: {
5976 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
5977 Info.opc = ISD::INTRINSIC_W_CHAIN;
5978 Info.memVT = MVT::getVT(PtrTy->getElementType());
5979 Info.ptrVal = I.getArgOperand(0);
5981 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
5983 Info.readMem = true;
5984 Info.writeMem = false;
5987 case Intrinsic::arm64_stxr: {
5988 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
5989 Info.opc = ISD::INTRINSIC_W_CHAIN;
5990 Info.memVT = MVT::getVT(PtrTy->getElementType());
5991 Info.ptrVal = I.getArgOperand(1);
5993 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
5995 Info.readMem = false;
5996 Info.writeMem = true;
5999 case Intrinsic::arm64_ldxp: {
6000 Info.opc = ISD::INTRINSIC_W_CHAIN;
6001 Info.memVT = MVT::i128;
6002 Info.ptrVal = I.getArgOperand(0);
6006 Info.readMem = true;
6007 Info.writeMem = false;
6010 case Intrinsic::arm64_stxp: {
6011 Info.opc = ISD::INTRINSIC_W_CHAIN;
6012 Info.memVT = MVT::i128;
6013 Info.ptrVal = I.getArgOperand(2);
6017 Info.readMem = false;
6018 Info.writeMem = true;
6028 // Truncations from 64-bit GPR to 32-bit GPR is free.
6029 bool ARM64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
6030 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6032 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6033 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6034 if (NumBits1 <= NumBits2)
6038 bool ARM64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
6039 if (!VT1.isInteger() || !VT2.isInteger())
6041 unsigned NumBits1 = VT1.getSizeInBits();
6042 unsigned NumBits2 = VT2.getSizeInBits();
6043 if (NumBits1 <= NumBits2)
6048 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
6050 bool ARM64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
6051 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6053 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6054 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6055 if (NumBits1 == 32 && NumBits2 == 64)
6059 bool ARM64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
6060 if (!VT1.isInteger() || !VT2.isInteger())
6062 unsigned NumBits1 = VT1.getSizeInBits();
6063 unsigned NumBits2 = VT2.getSizeInBits();
6064 if (NumBits1 == 32 && NumBits2 == 64)
6069 bool ARM64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
6070 EVT VT1 = Val.getValueType();
6071 if (isZExtFree(VT1, VT2)) {
6075 if (Val.getOpcode() != ISD::LOAD)
6078 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
6079 return (VT1.isSimple() && VT1.isInteger() && VT2.isSimple() &&
6080 VT2.isInteger() && VT1.getSizeInBits() <= 32);
6083 bool ARM64TargetLowering::hasPairedLoad(Type *LoadedType,
6084 unsigned &RequiredAligment) const {
6085 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
6087 // Cyclone supports unaligned accesses.
6088 RequiredAligment = 0;
6089 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
6090 return NumBits == 32 || NumBits == 64;
6093 bool ARM64TargetLowering::hasPairedLoad(EVT LoadedType,
6094 unsigned &RequiredAligment) const {
6095 if (!LoadedType.isSimple() ||
6096 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
6098 // Cyclone supports unaligned accesses.
6099 RequiredAligment = 0;
6100 unsigned NumBits = LoadedType.getSizeInBits();
6101 return NumBits == 32 || NumBits == 64;
6104 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
6105 unsigned AlignCheck) {
6106 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
6107 (DstAlign == 0 || DstAlign % AlignCheck == 0));
6110 EVT ARM64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
6111 unsigned SrcAlign, bool IsMemset,
6112 bool ZeroMemset, bool MemcpyStrSrc,
6113 MachineFunction &MF) const {
6114 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
6115 // instruction to materialize the v2i64 zero and one store (with restrictive
6116 // addressing mode). Just do two i64 store of zero-registers.
6118 const Function *F = MF.getFunction();
6119 if (!IsMemset && Size >= 16 &&
6120 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
6121 Attribute::NoImplicitFloat) &&
6122 (memOpAlign(SrcAlign, DstAlign, 16) ||
6123 (allowsUnalignedMemoryAccesses(MVT::v2i64, 0, &Fast) && Fast)))
6126 return Size >= 8 ? MVT::i64 : MVT::i32;
6129 // 12-bit optionally shifted immediates are legal for adds.
6130 bool ARM64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
6131 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
6136 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
6137 // immediates is the same as for an add or a sub.
6138 bool ARM64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
6141 return isLegalAddImmediate(Immed);
6144 /// isLegalAddressingMode - Return true if the addressing mode represented
6145 /// by AM is legal for this target, for a load/store of the specified type.
6146 bool ARM64TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6148 // ARM64 has five basic addressing modes:
6150 // reg + 9-bit signed offset
6151 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
6153 // reg + SIZE_IN_BYTES * reg
6155 // No global is ever allowed as a base.
6159 // No reg+reg+imm addressing.
6160 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
6163 // check reg + imm case:
6164 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
6165 uint64_t NumBytes = 0;
6166 if (Ty->isSized()) {
6167 uint64_t NumBits = getDataLayout()->getTypeSizeInBits(Ty);
6168 NumBytes = NumBits / 8;
6169 if (!isPowerOf2_64(NumBits))
6174 int64_t Offset = AM.BaseOffs;
6176 // 9-bit signed offset
6177 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
6180 // 12-bit unsigned offset
6181 unsigned shift = Log2_64(NumBytes);
6182 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
6183 // Must be a multiple of NumBytes (NumBytes is a power of 2)
6184 (Offset >> shift) << shift == Offset)
6189 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
6191 if (!AM.Scale || AM.Scale == 1 ||
6192 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
6197 int ARM64TargetLowering::getScalingFactorCost(const AddrMode &AM,
6199 // Scaling factors are not free at all.
6200 // Operands | Rt Latency
6201 // -------------------------------------------
6203 // -------------------------------------------
6204 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
6205 // Rt, [Xn, Wm, <extend> #imm] |
6206 if (isLegalAddressingMode(AM, Ty))
6207 // Scale represents reg2 * scale, thus account for 1 if
6208 // it is not equal to 0 or 1.
6209 return AM.Scale != 0 && AM.Scale != 1;
6213 bool ARM64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
6214 VT = VT.getScalarType();
6219 switch (VT.getSimpleVT().SimpleTy) {
6231 ARM64TargetLowering::getScratchRegisters(CallingConv::ID) const {
6232 // LR is a callee-save register, but we must treat it as clobbered by any call
6233 // site. Hence we include LR in the scratch registers, which are in turn added
6234 // as implicit-defs for stackmaps and patchpoints.
6235 static const MCPhysReg ScratchRegs[] = {
6236 ARM64::X16, ARM64::X17, ARM64::LR, 0
6241 bool ARM64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
6243 assert(Ty->isIntegerTy());
6245 unsigned BitSize = Ty->getPrimitiveSizeInBits();
6249 int64_t Val = Imm.getSExtValue();
6250 if (Val == 0 || ARM64_AM::isLogicalImmediate(Val, BitSize))
6253 if ((int64_t)Val < 0)
6256 Val &= (1LL << 32) - 1;
6258 unsigned LZ = countLeadingZeros((uint64_t)Val);
6259 unsigned Shift = (63 - LZ) / 16;
6260 // MOVZ is free so return true for one or fewer MOVK.
6261 return (Shift < 3) ? true : false;
6264 // Generate SUBS and CSEL for integer abs.
6265 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
6266 EVT VT = N->getValueType(0);
6268 SDValue N0 = N->getOperand(0);
6269 SDValue N1 = N->getOperand(1);
6272 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
6273 // and change it to SUB and CSEL.
6274 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
6275 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
6276 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
6277 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
6278 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
6279 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
6281 // Generate SUBS & CSEL.
6283 DAG.getNode(ARM64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
6284 N0.getOperand(0), DAG.getConstant(0, VT));
6285 return DAG.getNode(ARM64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
6286 DAG.getConstant(ARM64CC::PL, MVT::i32),
6287 SDValue(Cmp.getNode(), 1));
6292 // performXorCombine - Attempts to handle integer ABS.
6293 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
6294 TargetLowering::DAGCombinerInfo &DCI,
6295 const ARM64Subtarget *Subtarget) {
6296 if (DCI.isBeforeLegalizeOps())
6299 return performIntegerAbsCombine(N, DAG);
6302 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
6303 TargetLowering::DAGCombinerInfo &DCI,
6304 const ARM64Subtarget *Subtarget) {
6305 if (DCI.isBeforeLegalizeOps())
6308 // Multiplication of a power of two plus/minus one can be done more
6309 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
6310 // future CPUs have a cheaper MADD instruction, this may need to be
6311 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
6312 // 64-bit is 5 cycles, so this is always a win.
6313 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
6314 APInt Value = C->getAPIntValue();
6315 EVT VT = N->getValueType(0);
6316 APInt VP1 = Value + 1;
6317 if (VP1.isPowerOf2()) {
6318 // Multiplying by one less than a power of two, replace with a shift
6320 SDValue ShiftedVal =
6321 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6322 DAG.getConstant(VP1.logBase2(), MVT::i64));
6323 return DAG.getNode(ISD::SUB, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6325 APInt VM1 = Value - 1;
6326 if (VM1.isPowerOf2()) {
6327 // Multiplying by one more than a power of two, replace with a shift
6329 SDValue ShiftedVal =
6330 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6331 DAG.getConstant(VM1.logBase2(), MVT::i64));
6332 return DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6338 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG) {
6339 EVT VT = N->getValueType(0);
6340 if (VT != MVT::f32 && VT != MVT::f64)
6342 // Only optimize when the source and destination types have the same width.
6343 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
6346 // If the result of an integer load is only used by an integer-to-float
6347 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
6348 // This eliminates an "integer-to-vector-move UOP and improve throughput.
6349 SDValue N0 = N->getOperand(0);
6350 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6351 // Do not change the width of a volatile load.
6352 !cast<LoadSDNode>(N0)->isVolatile()) {
6353 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6354 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
6355 LN0->getPointerInfo(), LN0->isVolatile(),
6356 LN0->isNonTemporal(), LN0->isInvariant(),
6357 LN0->getAlignment());
6359 // Make sure successors of the original load stay after it by updating them
6360 // to use the new Chain.
6361 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
6364 (N->getOpcode() == ISD::SINT_TO_FP) ? ARM64ISD::SITOF : ARM64ISD::UITOF;
6365 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
6371 /// An EXTR instruction is made up of two shifts, ORed together. This helper
6372 /// searches for and classifies those shifts.
6373 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
6375 if (N.getOpcode() == ISD::SHL)
6377 else if (N.getOpcode() == ISD::SRL)
6382 if (!isa<ConstantSDNode>(N.getOperand(1)))
6385 ShiftAmount = N->getConstantOperandVal(1);
6386 Src = N->getOperand(0);
6390 /// EXTR instruction extracts a contiguous chunk of bits from two existing
6391 /// registers viewed as a high/low pair. This function looks for the pattern:
6392 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
6393 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
6395 static SDValue tryCombineToEXTR(SDNode *N,
6396 TargetLowering::DAGCombinerInfo &DCI) {
6397 SelectionDAG &DAG = DCI.DAG;
6399 EVT VT = N->getValueType(0);
6401 assert(N->getOpcode() == ISD::OR && "Unexpected root");
6403 if (VT != MVT::i32 && VT != MVT::i64)
6407 uint32_t ShiftLHS = 0;
6409 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
6413 uint32_t ShiftRHS = 0;
6415 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
6418 // If they're both trying to come from the high part of the register, they're
6419 // not really an EXTR.
6420 if (LHSFromHi == RHSFromHi)
6423 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
6427 std::swap(LHS, RHS);
6428 std::swap(ShiftLHS, ShiftRHS);
6431 return DAG.getNode(ARM64ISD::EXTR, DL, VT, LHS, RHS,
6432 DAG.getConstant(ShiftRHS, MVT::i64));
6435 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
6436 const ARM64Subtarget *Subtarget) {
6437 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
6438 if (!EnableARM64ExtrGeneration)
6440 SelectionDAG &DAG = DCI.DAG;
6441 EVT VT = N->getValueType(0);
6443 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6446 SDValue Res = tryCombineToEXTR(N, DCI);
6453 static SDValue performBitcastCombine(SDNode *N,
6454 TargetLowering::DAGCombinerInfo &DCI,
6455 SelectionDAG &DAG) {
6456 // Wait 'til after everything is legalized to try this. That way we have
6457 // legal vector types and such.
6458 if (DCI.isBeforeLegalizeOps())
6461 // Remove extraneous bitcasts around an extract_subvector.
6463 // (v4i16 (bitconvert
6464 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
6466 // (extract_subvector ((v8i16 ...), (i64 4)))
6468 // Only interested in 64-bit vectors as the ultimate result.
6469 EVT VT = N->getValueType(0);
6472 if (VT.getSimpleVT().getSizeInBits() != 64)
6474 // Is the operand an extract_subvector starting at the beginning or halfway
6475 // point of the vector? A low half may also come through as an
6476 // EXTRACT_SUBREG, so look for that, too.
6477 SDValue Op0 = N->getOperand(0);
6478 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
6479 !(Op0->isMachineOpcode() &&
6480 Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG))
6482 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
6483 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
6484 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
6486 } else if (Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG) {
6487 if (idx != ARM64::dsub)
6489 // The dsub reference is equivalent to a lane zero subvector reference.
6492 // Look through the bitcast of the input to the extract.
6493 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
6495 SDValue Source = Op0->getOperand(0)->getOperand(0);
6496 // If the source type has twice the number of elements as our destination
6497 // type, we know this is an extract of the high or low half of the vector.
6498 EVT SVT = Source->getValueType(0);
6499 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
6502 DEBUG(dbgs() << "arm64-lower: bitcast extract_subvector simplification\n");
6504 // Create the simplified form to just extract the low or high half of the
6505 // vector directly rather than bothering with the bitcasts.
6507 unsigned NumElements = VT.getVectorNumElements();
6509 SDValue HalfIdx = DAG.getConstant(NumElements, MVT::i64);
6510 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
6512 SDValue SubReg = DAG.getTargetConstant(ARM64::dsub, MVT::i32);
6513 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
6519 static SDValue performConcatVectorsCombine(SDNode *N,
6520 TargetLowering::DAGCombinerInfo &DCI,
6521 SelectionDAG &DAG) {
6522 // Wait 'til after everything is legalized to try this. That way we have
6523 // legal vector types and such.
6524 if (DCI.isBeforeLegalizeOps())
6528 EVT VT = N->getValueType(0);
6530 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
6531 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
6532 // canonicalise to that.
6533 if (N->getOperand(0) == N->getOperand(1) && VT.getVectorNumElements() == 2) {
6534 assert(VT.getVectorElementType().getSizeInBits() == 64);
6535 return DAG.getNode(ARM64ISD::DUPLANE64, dl, VT,
6536 WidenVector(N->getOperand(0), DAG),
6537 DAG.getConstant(0, MVT::i64));
6540 // Canonicalise concat_vectors so that the right-hand vector has as few
6541 // bit-casts as possible before its real operation. The primary matching
6542 // destination for these operations will be the narrowing "2" instructions,
6543 // which depend on the operation being performed on this right-hand vector.
6545 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
6547 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
6549 SDValue Op1 = N->getOperand(1);
6550 if (Op1->getOpcode() != ISD::BITCAST)
6552 SDValue RHS = Op1->getOperand(0);
6553 MVT RHSTy = RHS.getValueType().getSimpleVT();
6554 // If the RHS is not a vector, this is not the pattern we're looking for.
6555 if (!RHSTy.isVector())
6558 DEBUG(dbgs() << "arm64-lower: concat_vectors bitcast simplification\n");
6560 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
6561 RHSTy.getVectorNumElements() * 2);
6563 ISD::BITCAST, dl, VT,
6564 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
6565 DAG.getNode(ISD::BITCAST, dl, RHSTy, N->getOperand(0)), RHS));
6568 static SDValue tryCombineFixedPointConvert(SDNode *N,
6569 TargetLowering::DAGCombinerInfo &DCI,
6570 SelectionDAG &DAG) {
6571 // Wait 'til after everything is legalized to try this. That way we have
6572 // legal vector types and such.
6573 if (DCI.isBeforeLegalizeOps())
6575 // Transform a scalar conversion of a value from a lane extract into a
6576 // lane extract of a vector conversion. E.g., from foo1 to foo2:
6577 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
6578 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
6580 // The second form interacts better with instruction selection and the
6581 // register allocator to avoid cross-class register copies that aren't
6582 // coalescable due to a lane reference.
6584 // Check the operand and see if it originates from a lane extract.
6585 SDValue Op1 = N->getOperand(1);
6586 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6587 // Yep, no additional predication needed. Perform the transform.
6588 SDValue IID = N->getOperand(0);
6589 SDValue Shift = N->getOperand(2);
6590 SDValue Vec = Op1.getOperand(0);
6591 SDValue Lane = Op1.getOperand(1);
6592 EVT ResTy = N->getValueType(0);
6596 // The vector width should be 128 bits by the time we get here, even
6597 // if it started as 64 bits (the extract_vector handling will have
6599 assert(Vec.getValueType().getSizeInBits() == 128 &&
6600 "unexpected vector size on extract_vector_elt!");
6601 if (Vec.getValueType() == MVT::v4i32)
6602 VecResTy = MVT::v4f32;
6603 else if (Vec.getValueType() == MVT::v2i64)
6604 VecResTy = MVT::v2f64;
6606 assert(0 && "unexpected vector type!");
6609 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
6610 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
6615 // AArch64 high-vector "long" operations are formed by performing the non-high
6616 // version on an extract_subvector of each operand which gets the high half:
6618 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
6620 // However, there are cases which don't have an extract_high explicitly, but
6621 // have another operation that can be made compatible with one for free. For
6624 // (dupv64 scalar) --> (extract_high (dup128 scalar))
6626 // This routine does the actual conversion of such DUPs, once outer routines
6627 // have determined that everything else is in order.
6628 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
6629 // We can handle most types of duplicate, but the lane ones have an extra
6630 // operand saying *which* lane, so we need to know.
6632 switch (N.getOpcode()) {
6636 case ARM64ISD::DUPLANE8:
6637 case ARM64ISD::DUPLANE16:
6638 case ARM64ISD::DUPLANE32:
6639 case ARM64ISD::DUPLANE64:
6646 MVT NarrowTy = N.getSimpleValueType();
6647 if (!NarrowTy.is64BitVector())
6650 MVT ElementTy = NarrowTy.getVectorElementType();
6651 unsigned NumElems = NarrowTy.getVectorNumElements();
6652 MVT NewDUPVT = MVT::getVectorVT(ElementTy, NumElems * 2);
6656 NewDUP = DAG.getNode(N.getOpcode(), SDLoc(N), NewDUPVT, N.getOperand(0),
6659 NewDUP = DAG.getNode(ARM64ISD::DUP, SDLoc(N), NewDUPVT, N.getOperand(0));
6661 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N.getNode()), NarrowTy,
6662 NewDUP, DAG.getConstant(NumElems, MVT::i64));
6665 static bool isEssentiallyExtractSubvector(SDValue N) {
6666 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
6669 return N.getOpcode() == ISD::BITCAST &&
6670 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
6673 /// \brief Helper structure to keep track of ISD::SET_CC operands.
6674 struct GenericSetCCInfo {
6675 const SDValue *Opnd0;
6676 const SDValue *Opnd1;
6680 /// \brief Helper structure to keep track of a SET_CC lowered into ARM64 code.
6681 struct ARM64SetCCInfo {
6683 ARM64CC::CondCode CC;
6686 /// \brief Helper structure to keep track of SetCC information.
6688 GenericSetCCInfo Generic;
6689 ARM64SetCCInfo ARM64;
6692 /// \brief Helper structure to be able to read SetCC information.
6693 /// If set to true, IsARM64 field, Info is a ARM64SetCCInfo, otherwise Info is
6694 /// a GenericSetCCInfo.
6695 struct SetCCInfoAndKind {
6700 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
6702 /// ARM64 lowered one.
6703 /// \p SetCCInfo is filled accordingly.
6704 /// \post SetCCInfo is meanginfull only when this function returns true.
6705 /// \return True when Op is a kind of SET_CC operation.
6706 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
6707 // If this is a setcc, this is straight forward.
6708 if (Op.getOpcode() == ISD::SETCC) {
6709 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
6710 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
6711 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6712 SetCCInfo.IsARM64 = false;
6715 // Otherwise, check if this is a matching csel instruction.
6719 if (Op.getOpcode() != ARM64ISD::CSEL)
6721 // Set the information about the operands.
6722 // TODO: we want the operands of the Cmp not the csel
6723 SetCCInfo.Info.ARM64.Cmp = &Op.getOperand(3);
6724 SetCCInfo.IsARM64 = true;
6725 SetCCInfo.Info.ARM64.CC = static_cast<ARM64CC::CondCode>(
6726 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
6728 // Check that the operands matches the constraints:
6729 // (1) Both operands must be constants.
6730 // (2) One must be 1 and the other must be 0.
6731 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
6732 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6735 if (!TValue || !FValue)
6739 if (!TValue->isOne()) {
6740 // Update the comparison when we are interested in !cc.
6741 std::swap(TValue, FValue);
6742 SetCCInfo.Info.ARM64.CC =
6743 ARM64CC::getInvertedCondCode(SetCCInfo.Info.ARM64.CC);
6745 return TValue->isOne() && FValue->isNullValue();
6748 // The folding we want to perform is:
6749 // (add x, (setcc cc ...) )
6751 // (csel x, (add x, 1), !cc ...)
6753 // The latter will get matched to a CSINC instruction.
6754 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
6755 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
6756 SDValue LHS = Op->getOperand(0);
6757 SDValue RHS = Op->getOperand(1);
6758 SetCCInfoAndKind InfoAndKind;
6760 // If neither operand is a SET_CC, give up.
6761 if (!isSetCC(LHS, InfoAndKind)) {
6762 std::swap(LHS, RHS);
6763 if (!isSetCC(LHS, InfoAndKind))
6767 // FIXME: This could be generatized to work for FP comparisons.
6768 EVT CmpVT = InfoAndKind.IsARM64
6769 ? InfoAndKind.Info.ARM64.Cmp->getOperand(0).getValueType()
6770 : InfoAndKind.Info.Generic.Opnd0->getValueType();
6771 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
6777 if (InfoAndKind.IsARM64) {
6778 CCVal = DAG.getConstant(
6779 ARM64CC::getInvertedCondCode(InfoAndKind.Info.ARM64.CC), MVT::i32);
6780 Cmp = *InfoAndKind.Info.ARM64.Cmp;
6782 Cmp = getARM64Cmp(*InfoAndKind.Info.Generic.Opnd0,
6783 *InfoAndKind.Info.Generic.Opnd1,
6784 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
6787 EVT VT = Op->getValueType(0);
6788 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, VT));
6789 return DAG.getNode(ARM64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
6792 // The basic add/sub long vector instructions have variants with "2" on the end
6793 // which act on the high-half of their inputs. They are normally matched by
6796 // (add (zeroext (extract_high LHS)),
6797 // (zeroext (extract_high RHS)))
6798 // -> uaddl2 vD, vN, vM
6800 // However, if one of the extracts is something like a duplicate, this
6801 // instruction can still be used profitably. This function puts the DAG into a
6802 // more appropriate form for those patterns to trigger.
6803 static SDValue performAddSubLongCombine(SDNode *N,
6804 TargetLowering::DAGCombinerInfo &DCI,
6805 SelectionDAG &DAG) {
6806 if (DCI.isBeforeLegalizeOps())
6809 MVT VT = N->getSimpleValueType(0);
6810 if (!VT.is128BitVector()) {
6811 if (N->getOpcode() == ISD::ADD)
6812 return performSetccAddFolding(N, DAG);
6816 // Make sure both branches are extended in the same way.
6817 SDValue LHS = N->getOperand(0);
6818 SDValue RHS = N->getOperand(1);
6819 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
6820 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
6821 LHS.getOpcode() != RHS.getOpcode())
6824 unsigned ExtType = LHS.getOpcode();
6826 // It's not worth doing if at least one of the inputs isn't already an
6827 // extract, but we don't know which it'll be so we have to try both.
6828 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
6829 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
6833 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
6834 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
6835 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
6839 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
6842 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
6845 // Massage DAGs which we can use the high-half "long" operations on into
6846 // something isel will recognize better. E.g.
6848 // (arm64_neon_umull (extract_high vec) (dupv64 scalar)) -->
6849 // (arm64_neon_umull (extract_high (v2i64 vec)))
6850 // (extract_high (v2i64 (dup128 scalar)))))
6852 static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N,
6853 TargetLowering::DAGCombinerInfo &DCI,
6854 SelectionDAG &DAG) {
6855 if (DCI.isBeforeLegalizeOps())
6858 SDValue LHS = N->getOperand(1);
6859 SDValue RHS = N->getOperand(2);
6860 assert(LHS.getValueType().is64BitVector() &&
6861 RHS.getValueType().is64BitVector() &&
6862 "unexpected shape for long operation");
6864 // Either node could be a DUP, but it's not worth doing both of them (you'd
6865 // just as well use the non-high version) so look for a corresponding extract
6866 // operation on the other "wing".
6867 if (isEssentiallyExtractSubvector(LHS)) {
6868 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
6871 } else if (isEssentiallyExtractSubvector(RHS)) {
6872 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
6877 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
6878 N->getOperand(0), LHS, RHS);
6881 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
6882 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
6883 unsigned ElemBits = ElemTy.getSizeInBits();
6885 int64_t ShiftAmount;
6886 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
6887 APInt SplatValue, SplatUndef;
6888 unsigned SplatBitSize;
6890 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
6891 HasAnyUndefs, ElemBits) ||
6892 SplatBitSize != ElemBits)
6895 ShiftAmount = SplatValue.getSExtValue();
6896 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
6897 ShiftAmount = CVN->getSExtValue();
6905 llvm_unreachable("Unknown shift intrinsic");
6906 case Intrinsic::arm64_neon_sqshl:
6907 Opcode = ARM64ISD::SQSHL_I;
6908 IsRightShift = false;
6910 case Intrinsic::arm64_neon_uqshl:
6911 Opcode = ARM64ISD::UQSHL_I;
6912 IsRightShift = false;
6914 case Intrinsic::arm64_neon_srshl:
6915 Opcode = ARM64ISD::SRSHR_I;
6916 IsRightShift = true;
6918 case Intrinsic::arm64_neon_urshl:
6919 Opcode = ARM64ISD::URSHR_I;
6920 IsRightShift = true;
6922 case Intrinsic::arm64_neon_sqshlu:
6923 Opcode = ARM64ISD::SQSHLU_I;
6924 IsRightShift = false;
6928 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits)
6929 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6930 DAG.getConstant(-ShiftAmount, MVT::i32));
6931 else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount <= ElemBits)
6932 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6933 DAG.getConstant(ShiftAmount, MVT::i32));
6938 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
6939 // the intrinsics must be legal and take an i32, this means there's almost
6940 // certainly going to be a zext in the DAG which we can eliminate.
6941 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
6942 SDValue AndN = N->getOperand(2);
6943 if (AndN.getOpcode() != ISD::AND)
6946 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
6947 if (!CMask || CMask->getZExtValue() != Mask)
6950 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
6951 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
6954 static SDValue performIntrinsicCombine(SDNode *N,
6955 TargetLowering::DAGCombinerInfo &DCI,
6956 const ARM64Subtarget *Subtarget) {
6957 SelectionDAG &DAG = DCI.DAG;
6958 unsigned IID = getIntrinsicID(N);
6962 case Intrinsic::arm64_neon_vcvtfxs2fp:
6963 case Intrinsic::arm64_neon_vcvtfxu2fp:
6964 return tryCombineFixedPointConvert(N, DCI, DAG);
6966 case Intrinsic::arm64_neon_fmax:
6967 return DAG.getNode(ARM64ISD::FMAX, SDLoc(N), N->getValueType(0),
6968 N->getOperand(1), N->getOperand(2));
6969 case Intrinsic::arm64_neon_fmin:
6970 return DAG.getNode(ARM64ISD::FMIN, SDLoc(N), N->getValueType(0),
6971 N->getOperand(1), N->getOperand(2));
6972 case Intrinsic::arm64_neon_smull:
6973 case Intrinsic::arm64_neon_umull:
6974 case Intrinsic::arm64_neon_pmull:
6975 case Intrinsic::arm64_neon_sqdmull:
6976 return tryCombineLongOpWithDup(IID, N, DCI, DAG);
6977 case Intrinsic::arm64_neon_sqshl:
6978 case Intrinsic::arm64_neon_uqshl:
6979 case Intrinsic::arm64_neon_sqshlu:
6980 case Intrinsic::arm64_neon_srshl:
6981 case Intrinsic::arm64_neon_urshl:
6982 return tryCombineShiftImm(IID, N, DAG);
6983 case Intrinsic::arm64_crc32b:
6984 case Intrinsic::arm64_crc32cb:
6985 return tryCombineCRC32(0xff, N, DAG);
6986 case Intrinsic::arm64_crc32h:
6987 case Intrinsic::arm64_crc32ch:
6988 return tryCombineCRC32(0xffff, N, DAG);
6993 static SDValue performExtendCombine(SDNode *N,
6994 TargetLowering::DAGCombinerInfo &DCI,
6995 SelectionDAG &DAG) {
6996 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
6997 // we can convert that DUP into another extract_high (of a bigger DUP), which
6998 // helps the backend to decide that an sabdl2 would be useful, saving a real
6999 // extract_high operation.
7000 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
7001 N->getOperand(0).getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
7002 SDNode *ABDNode = N->getOperand(0).getNode();
7003 unsigned IID = getIntrinsicID(ABDNode);
7004 if (IID == Intrinsic::arm64_neon_sabd ||
7005 IID == Intrinsic::arm64_neon_uabd) {
7006 SDValue NewABD = tryCombineLongOpWithDup(IID, ABDNode, DCI, DAG);
7007 if (!NewABD.getNode())
7010 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
7015 // This is effectively a custom type legalization for ARM64.
7017 // Type legalization will split an extend of a small, legal, type to a larger
7018 // illegal type by first splitting the destination type, often creating
7019 // illegal source types, which then get legalized in isel-confusing ways,
7020 // leading to really terrible codegen. E.g.,
7021 // %result = v8i32 sext v8i8 %value
7023 // %losrc = extract_subreg %value, ...
7024 // %hisrc = extract_subreg %value, ...
7025 // %lo = v4i32 sext v4i8 %losrc
7026 // %hi = v4i32 sext v4i8 %hisrc
7027 // Things go rapidly downhill from there.
7029 // For ARM64, the [sz]ext vector instructions can only go up one element
7030 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
7031 // take two instructions.
7033 // This implies that the most efficient way to do the extend from v8i8
7034 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
7035 // the normal splitting to happen for the v8i16->v8i32.
7037 // This is pre-legalization to catch some cases where the default
7038 // type legalization will create ill-tempered code.
7039 if (!DCI.isBeforeLegalizeOps())
7042 // We're only interested in cleaning things up for non-legal vector types
7043 // here. If both the source and destination are legal, things will just
7044 // work naturally without any fiddling.
7045 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7046 EVT ResVT = N->getValueType(0);
7047 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
7049 // If the vector type isn't a simple VT, it's beyond the scope of what
7050 // we're worried about here. Let legalization do its thing and hope for
7052 if (!ResVT.isSimple())
7055 SDValue Src = N->getOperand(0);
7056 MVT SrcVT = Src->getValueType(0).getSimpleVT();
7057 // If the source VT is a 64-bit vector, we can play games and get the
7058 // better results we want.
7059 if (SrcVT.getSizeInBits() != 64)
7062 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
7063 unsigned ElementCount = SrcVT.getVectorNumElements();
7064 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
7066 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
7068 // Now split the rest of the operation into two halves, each with a 64
7072 unsigned NumElements = ResVT.getVectorNumElements();
7073 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
7074 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
7075 ResVT.getVectorElementType(), NumElements / 2);
7077 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
7078 LoVT.getVectorNumElements());
7079 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7080 DAG.getIntPtrConstant(0));
7081 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7082 DAG.getIntPtrConstant(InNVT.getVectorNumElements()));
7083 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
7084 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
7086 // Now combine the parts back together so we still have a single result
7087 // like the combiner expects.
7088 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
7091 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
7092 /// value. The load store optimizer pass will merge them to store pair stores.
7093 /// This has better performance than a splat of the scalar followed by a split
7094 /// vector store. Even if the stores are not merged it is four stores vs a dup,
7095 /// followed by an ext.b and two stores.
7096 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
7097 SDValue StVal = St->getValue();
7098 EVT VT = StVal.getValueType();
7100 // Don't replace floating point stores, they possibly won't be transformed to
7101 // stp because of the store pair suppress pass.
7102 if (VT.isFloatingPoint())
7105 // Check for insert vector elements.
7106 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
7109 // We can express a splat as store pair(s) for 2 or 4 elements.
7110 unsigned NumVecElts = VT.getVectorNumElements();
7111 if (NumVecElts != 4 && NumVecElts != 2)
7113 SDValue SplatVal = StVal.getOperand(1);
7114 unsigned RemainInsertElts = NumVecElts - 1;
7116 // Check that this is a splat.
7117 while (--RemainInsertElts) {
7118 SDValue NextInsertElt = StVal.getOperand(0);
7119 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
7121 if (NextInsertElt.getOperand(1) != SplatVal)
7123 StVal = NextInsertElt;
7125 unsigned OrigAlignment = St->getAlignment();
7126 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
7127 unsigned Alignment = std::min(OrigAlignment, EltOffset);
7129 // Create scalar stores. This is at least as good as the code sequence for a
7130 // split unaligned store wich is a dup.s, ext.b, and two stores.
7131 // Most of the time the three stores should be replaced by store pair
7132 // instructions (stp).
7134 SDValue BasePtr = St->getBasePtr();
7136 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
7137 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
7139 unsigned Offset = EltOffset;
7140 while (--NumVecElts) {
7141 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7142 DAG.getConstant(Offset, MVT::i64));
7143 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
7144 St->getPointerInfo(), St->isVolatile(),
7145 St->isNonTemporal(), Alignment);
7146 Offset += EltOffset;
7151 static SDValue performSTORECombine(SDNode *N,
7152 TargetLowering::DAGCombinerInfo &DCI,
7154 const ARM64Subtarget *Subtarget) {
7155 if (!DCI.isBeforeLegalize())
7158 StoreSDNode *S = cast<StoreSDNode>(N);
7159 if (S->isVolatile())
7162 // Cyclone has bad performance on unaligned 16B stores when crossing line and
7163 // page boundries. We want to split such stores.
7164 if (!Subtarget->isCyclone())
7167 // Don't split at Oz.
7168 MachineFunction &MF = DAG.getMachineFunction();
7169 bool IsMinSize = MF.getFunction()->getAttributes().hasAttribute(
7170 AttributeSet::FunctionIndex, Attribute::MinSize);
7174 SDValue StVal = S->getValue();
7175 EVT VT = StVal.getValueType();
7177 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
7178 // those up regresses performance on micro-benchmarks and olden/bh.
7179 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
7182 // Split unaligned 16B stores. They are terrible for performance.
7183 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
7184 // extensions can use this to mark that it does not want splitting to happen
7185 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
7186 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
7187 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
7188 S->getAlignment() <= 2)
7191 // If we get a splat of a scalar convert this vector store to a store of
7192 // scalars. They will be merged into store pairs thereby removing two
7194 SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S);
7195 if (ReplacedSplat != SDValue())
7196 return ReplacedSplat;
7199 unsigned NumElts = VT.getVectorNumElements() / 2;
7200 // Split VT into two.
7202 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
7203 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7204 DAG.getIntPtrConstant(0));
7205 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7206 DAG.getIntPtrConstant(NumElts));
7207 SDValue BasePtr = S->getBasePtr();
7209 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
7210 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
7211 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7212 DAG.getConstant(8, MVT::i64));
7213 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
7214 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
7218 // Optimize compare with zero and branch.
7219 static SDValue performBRCONDCombine(SDNode *N,
7220 TargetLowering::DAGCombinerInfo &DCI,
7221 SelectionDAG &DAG) {
7222 SDValue Chain = N->getOperand(0);
7223 SDValue Dest = N->getOperand(1);
7224 SDValue CCVal = N->getOperand(2);
7225 SDValue Cmp = N->getOperand(3);
7227 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
7228 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
7229 if (CC != ARM64CC::EQ && CC != ARM64CC::NE)
7232 unsigned CmpOpc = Cmp.getOpcode();
7233 if (CmpOpc != ARM64ISD::ADDS && CmpOpc != ARM64ISD::SUBS)
7236 // Only attempt folding if there is only one use of the flag and no use of the
7238 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
7241 SDValue LHS = Cmp.getOperand(0);
7242 SDValue RHS = Cmp.getOperand(1);
7244 assert(LHS.getValueType() == RHS.getValueType() &&
7245 "Expected the value type to be the same for both operands!");
7246 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
7249 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
7250 std::swap(LHS, RHS);
7252 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
7255 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
7256 LHS.getOpcode() == ISD::SRL)
7259 // Fold the compare into the branch instruction.
7261 if (CC == ARM64CC::EQ)
7262 BR = DAG.getNode(ARM64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7264 BR = DAG.getNode(ARM64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7266 // Do not add new nodes to DAG combiner worklist.
7267 DCI.CombineTo(N, BR, false);
7272 // vselect (v1i1 setcc) ->
7273 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
7274 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
7275 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
7277 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
7278 SDValue N0 = N->getOperand(0);
7279 EVT CCVT = N0.getValueType();
7281 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
7282 CCVT.getVectorElementType() != MVT::i1)
7285 EVT ResVT = N->getValueType(0);
7286 EVT CmpVT = N0.getOperand(0).getValueType();
7287 // Only combine when the result type is of the same size as the compared
7289 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
7292 SDValue IfTrue = N->getOperand(1);
7293 SDValue IfFalse = N->getOperand(2);
7295 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
7296 N0.getOperand(0), N0.getOperand(1),
7297 cast<CondCodeSDNode>(N0.getOperand(2))->get());
7298 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
7302 SDValue ARM64TargetLowering::PerformDAGCombine(SDNode *N,
7303 DAGCombinerInfo &DCI) const {
7304 SelectionDAG &DAG = DCI.DAG;
7305 switch (N->getOpcode()) {
7310 return performAddSubLongCombine(N, DCI, DAG);
7312 return performXorCombine(N, DAG, DCI, Subtarget);
7314 return performMulCombine(N, DAG, DCI, Subtarget);
7315 case ISD::SINT_TO_FP:
7316 case ISD::UINT_TO_FP:
7317 return performIntToFpCombine(N, DAG);
7319 return performORCombine(N, DCI, Subtarget);
7320 case ISD::INTRINSIC_WO_CHAIN:
7321 return performIntrinsicCombine(N, DCI, Subtarget);
7322 case ISD::ANY_EXTEND:
7323 case ISD::ZERO_EXTEND:
7324 case ISD::SIGN_EXTEND:
7325 return performExtendCombine(N, DCI, DAG);
7327 return performBitcastCombine(N, DCI, DAG);
7328 case ISD::CONCAT_VECTORS:
7329 return performConcatVectorsCombine(N, DCI, DAG);
7331 return performVSelectCombine(N, DCI.DAG);
7333 return performSTORECombine(N, DCI, DAG, Subtarget);
7334 case ARM64ISD::BRCOND:
7335 return performBRCONDCombine(N, DCI, DAG);
7340 // Check if the return value is used as only a return value, as otherwise
7341 // we can't perform a tail-call. In particular, we need to check for
7342 // target ISD nodes that are returns and any other "odd" constructs
7343 // that the generic analysis code won't necessarily catch.
7344 bool ARM64TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
7345 if (N->getNumValues() != 1)
7347 if (!N->hasNUsesOfValue(1, 0))
7350 SDValue TCChain = Chain;
7351 SDNode *Copy = *N->use_begin();
7352 if (Copy->getOpcode() == ISD::CopyToReg) {
7353 // If the copy has a glue operand, we conservatively assume it isn't safe to
7354 // perform a tail call.
7355 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
7358 TCChain = Copy->getOperand(0);
7359 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
7362 bool HasRet = false;
7363 for (SDNode *Node : Copy->uses()) {
7364 if (Node->getOpcode() != ARM64ISD::RET_FLAG)
7376 // Return whether the an instruction can potentially be optimized to a tail
7377 // call. This will cause the optimizers to attempt to move, or duplicate,
7378 // return instructions to help enable tail call optimizations for this
7380 bool ARM64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
7381 if (!EnableARM64TailCalls)
7384 if (!CI->isTailCall())
7390 bool ARM64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
7392 ISD::MemIndexedMode &AM,
7394 SelectionDAG &DAG) const {
7395 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
7398 Base = Op->getOperand(0);
7399 // All of the indexed addressing mode instructions take a signed
7400 // 9 bit immediate offset.
7401 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
7402 int64_t RHSC = (int64_t)RHS->getZExtValue();
7403 if (RHSC >= 256 || RHSC <= -256)
7405 IsInc = (Op->getOpcode() == ISD::ADD);
7406 Offset = Op->getOperand(1);
7412 bool ARM64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
7414 ISD::MemIndexedMode &AM,
7415 SelectionDAG &DAG) const {
7418 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7419 VT = LD->getMemoryVT();
7420 Ptr = LD->getBasePtr();
7421 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7422 VT = ST->getMemoryVT();
7423 Ptr = ST->getBasePtr();
7428 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
7430 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
7434 bool ARM64TargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
7437 ISD::MemIndexedMode &AM,
7438 SelectionDAG &DAG) const {
7441 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7442 VT = LD->getMemoryVT();
7443 Ptr = LD->getBasePtr();
7444 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7445 VT = ST->getMemoryVT();
7446 Ptr = ST->getBasePtr();
7451 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
7453 // Post-indexing updates the base, so it's not a valid transform
7454 // if that's not the same as the load's pointer.
7457 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
7461 /// The only 128-bit atomic operation is an stxp that succeeds. In particular
7462 /// neither ldp nor ldxp are atomic. So the canonical sequence for an atomic
7465 /// ldxp x0, x1, [x8]
7466 /// stxp w2, x0, x1, [x8]
7468 /// If the stxp succeeds then the ldxp managed to get both halves without an
7469 /// intervening stxp from a different thread and the read was atomic.
7470 static void ReplaceATOMIC_LOAD_128(SDNode *N, SmallVectorImpl<SDValue> &Results,
7471 SelectionDAG &DAG) {
7473 AtomicSDNode *AN = cast<AtomicSDNode>(N);
7474 EVT VT = AN->getMemoryVT();
7475 SDValue Zero = DAG.getConstant(0, VT);
7477 // FIXME: Really want ATOMIC_LOAD_NOP but that doesn't fit into the existing
7478 // scheme very well. Given the complexity of what we're already generating, an
7479 // extra couple of ORRs probably won't make much difference.
7480 SDValue Result = DAG.getAtomic(ISD::ATOMIC_LOAD_OR, DL, AN->getMemoryVT(),
7481 N->getOperand(0), N->getOperand(1), Zero,
7482 AN->getMemOperand(), AN->getOrdering(),
7483 AN->getSynchScope());
7485 Results.push_back(Result.getValue(0)); // Value
7486 Results.push_back(Result.getValue(1)); // Chain
7489 static void ReplaceATOMIC_OP_128(SDNode *N, SmallVectorImpl<SDValue> &Results,
7490 SelectionDAG &DAG, unsigned NewOp) {
7492 AtomicOrdering Ordering = cast<AtomicSDNode>(N)->getOrdering();
7493 assert(N->getValueType(0) == MVT::i128 &&
7494 "Only know how to expand i128 atomics");
7496 SmallVector<SDValue, 6> Ops;
7497 Ops.push_back(N->getOperand(1)); // Ptr
7499 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7500 N->getOperand(2), DAG.getIntPtrConstant(0)));
7501 // High part of Val1
7502 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7503 N->getOperand(2), DAG.getIntPtrConstant(1)));
7504 if (NewOp == ARM64::ATOMIC_CMP_SWAP_I128) {
7506 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7507 N->getOperand(3), DAG.getIntPtrConstant(0)));
7508 // High part of Val2
7509 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7510 N->getOperand(3), DAG.getIntPtrConstant(1)));
7513 Ops.push_back(DAG.getTargetConstant(Ordering, MVT::i32));
7514 Ops.push_back(N->getOperand(0)); // Chain
7516 SDVTList Tys = DAG.getVTList(MVT::i64, MVT::i64, MVT::Other);
7517 SDNode *Result = DAG.getMachineNode(NewOp, DL, Tys, Ops);
7518 SDValue OpsF[] = { SDValue(Result, 0), SDValue(Result, 1) };
7519 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i128, OpsF, 2));
7520 Results.push_back(SDValue(Result, 2));
7523 void ARM64TargetLowering::ReplaceNodeResults(SDNode *N,
7524 SmallVectorImpl<SDValue> &Results,
7525 SelectionDAG &DAG) const {
7526 switch (N->getOpcode()) {
7528 llvm_unreachable("Don't know how to custom expand this");
7529 case ISD::ATOMIC_LOAD:
7530 ReplaceATOMIC_LOAD_128(N, Results, DAG);
7532 case ISD::ATOMIC_LOAD_ADD:
7533 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_ADD_I128);
7535 case ISD::ATOMIC_LOAD_SUB:
7536 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_SUB_I128);
7538 case ISD::ATOMIC_LOAD_AND:
7539 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_AND_I128);
7541 case ISD::ATOMIC_LOAD_OR:
7542 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_OR_I128);
7544 case ISD::ATOMIC_LOAD_XOR:
7545 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_XOR_I128);
7547 case ISD::ATOMIC_LOAD_NAND:
7548 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_NAND_I128);
7550 case ISD::ATOMIC_SWAP:
7551 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_SWAP_I128);
7553 case ISD::ATOMIC_LOAD_MIN:
7554 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_MIN_I128);
7556 case ISD::ATOMIC_LOAD_MAX:
7557 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_MAX_I128);
7559 case ISD::ATOMIC_LOAD_UMIN:
7560 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_UMIN_I128);
7562 case ISD::ATOMIC_LOAD_UMAX:
7563 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_UMAX_I128);
7565 case ISD::ATOMIC_CMP_SWAP:
7566 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_CMP_SWAP_I128);
7568 case ISD::FP_TO_UINT:
7569 case ISD::FP_TO_SINT:
7570 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
7571 // Let normal code take care of it by not adding anything to Results.