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::f32, &ARM64::FPR32RegClass);
88 addRegisterClass(MVT::f64, &ARM64::FPR64RegClass);
89 addRegisterClass(MVT::f128, &ARM64::FPR128RegClass);
90 addRegisterClass(MVT::v16i8, &ARM64::FPR8RegClass);
91 addRegisterClass(MVT::v8i16, &ARM64::FPR16RegClass);
93 // Someone set us up the NEON.
94 addDRTypeForNEON(MVT::v2f32);
95 addDRTypeForNEON(MVT::v8i8);
96 addDRTypeForNEON(MVT::v4i16);
97 addDRTypeForNEON(MVT::v2i32);
98 addDRTypeForNEON(MVT::v1i64);
99 addDRTypeForNEON(MVT::v1f64);
101 addQRTypeForNEON(MVT::v4f32);
102 addQRTypeForNEON(MVT::v2f64);
103 addQRTypeForNEON(MVT::v16i8);
104 addQRTypeForNEON(MVT::v8i16);
105 addQRTypeForNEON(MVT::v4i32);
106 addQRTypeForNEON(MVT::v2i64);
108 // Compute derived properties from the register classes
109 computeRegisterProperties();
111 // Provide all sorts of operation actions
112 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
113 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
114 setOperationAction(ISD::SETCC, MVT::i32, Custom);
115 setOperationAction(ISD::SETCC, MVT::i64, Custom);
116 setOperationAction(ISD::SETCC, MVT::f32, Custom);
117 setOperationAction(ISD::SETCC, MVT::f64, Custom);
118 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
119 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
120 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
121 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
122 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
123 setOperationAction(ISD::SELECT, MVT::i32, Custom);
124 setOperationAction(ISD::SELECT, MVT::i64, Custom);
125 setOperationAction(ISD::SELECT, MVT::f32, Custom);
126 setOperationAction(ISD::SELECT, MVT::f64, Custom);
127 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
128 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
129 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
130 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
131 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
132 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
134 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
135 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
136 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
138 setOperationAction(ISD::FREM, MVT::f32, Expand);
139 setOperationAction(ISD::FREM, MVT::f64, Expand);
140 setOperationAction(ISD::FREM, MVT::f80, Expand);
142 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
143 // silliness like this:
144 setOperationAction(ISD::FABS, MVT::v1f64, Expand);
145 setOperationAction(ISD::FADD, MVT::v1f64, Expand);
146 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
147 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
148 setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
149 setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
150 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
151 setOperationAction(ISD::FMA, MVT::v1f64, Expand);
152 setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
153 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
154 setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
155 setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
156 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
157 setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
158 setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
159 setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
160 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
161 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
162 setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
163 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
164 setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
165 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
166 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
167 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
168 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);
170 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
171 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
172 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
173 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
174 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);
176 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
178 // Custom lowering hooks are needed for XOR
179 // to fold it into CSINC/CSINV.
180 setOperationAction(ISD::XOR, MVT::i32, Custom);
181 setOperationAction(ISD::XOR, MVT::i64, Custom);
183 // Virtually no operation on f128 is legal, but LLVM can't expand them when
184 // there's a valid register class, so we need custom operations in most cases.
185 setOperationAction(ISD::FABS, MVT::f128, Expand);
186 setOperationAction(ISD::FADD, MVT::f128, Custom);
187 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
188 setOperationAction(ISD::FCOS, MVT::f128, Expand);
189 setOperationAction(ISD::FDIV, MVT::f128, Custom);
190 setOperationAction(ISD::FMA, MVT::f128, Expand);
191 setOperationAction(ISD::FMUL, MVT::f128, Custom);
192 setOperationAction(ISD::FNEG, MVT::f128, Expand);
193 setOperationAction(ISD::FPOW, MVT::f128, Expand);
194 setOperationAction(ISD::FREM, MVT::f128, Expand);
195 setOperationAction(ISD::FRINT, MVT::f128, Expand);
196 setOperationAction(ISD::FSIN, MVT::f128, Expand);
197 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
198 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
199 setOperationAction(ISD::FSUB, MVT::f128, Custom);
200 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
201 setOperationAction(ISD::SETCC, MVT::f128, Custom);
202 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
203 setOperationAction(ISD::SELECT, MVT::f128, Custom);
204 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
205 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
207 // Lowering for many of the conversions is actually specified by the non-f128
208 // type. The LowerXXX function will be trivial when f128 isn't involved.
209 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
210 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
211 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
212 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
213 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
214 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
215 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
216 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
217 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
218 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
219 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
220 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
221 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
222 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
225 setOperationAction(ISD::ATOMIC_SWAP, MVT::i128, Custom);
226 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i128, Custom);
227 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i128, Custom);
228 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i128, Custom);
229 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i128, Custom);
230 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i128, Custom);
231 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i128, Custom);
232 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i128, Custom);
233 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i128, Custom);
234 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i128, Custom);
235 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i128, Custom);
236 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i128, Custom);
237 // These are surprisingly difficult. The only single-copy atomic 128-bit
238 // instruction on AArch64 is stxp (when it succeeds). So a store can safely
239 // become a simple swap, but a load can only be determined to have been atomic
240 // if storing the same value back succeeds.
241 setOperationAction(ISD::ATOMIC_LOAD, MVT::i128, Custom);
242 setOperationAction(ISD::ATOMIC_STORE, MVT::i128, Expand);
244 // Variable arguments.
245 setOperationAction(ISD::VASTART, MVT::Other, Custom);
246 setOperationAction(ISD::VAARG, MVT::Other, Custom);
247 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
248 setOperationAction(ISD::VAEND, MVT::Other, Expand);
250 // Variable-sized objects.
251 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
252 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
253 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
255 // Exception handling.
256 // FIXME: These are guesses. Has this been defined yet?
257 setExceptionPointerRegister(ARM64::X0);
258 setExceptionSelectorRegister(ARM64::X1);
260 // Constant pool entries
261 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
264 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
266 // Add/Sub overflow ops with MVT::Glues are lowered to CPSR dependences.
267 setOperationAction(ISD::ADDC, MVT::i32, Custom);
268 setOperationAction(ISD::ADDE, MVT::i32, Custom);
269 setOperationAction(ISD::SUBC, MVT::i32, Custom);
270 setOperationAction(ISD::SUBE, MVT::i32, Custom);
271 setOperationAction(ISD::ADDC, MVT::i64, Custom);
272 setOperationAction(ISD::ADDE, MVT::i64, Custom);
273 setOperationAction(ISD::SUBC, MVT::i64, Custom);
274 setOperationAction(ISD::SUBE, MVT::i64, Custom);
276 // ARM64 lacks both left-rotate and popcount instructions.
277 setOperationAction(ISD::ROTL, MVT::i32, Expand);
278 setOperationAction(ISD::ROTL, MVT::i64, Expand);
280 // ARM64 doesn't have a direct vector ->f32 conversion instructions for
281 // elements smaller than i32, so promote the input to i32 first.
282 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
283 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
284 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
285 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
286 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
287 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
288 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
289 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
290 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
292 // ARM64 doesn't have {U|S}MUL_LOHI.
293 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
294 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
296 // ARM64 doesn't have MUL.2d:
297 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
299 // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
300 // counterparts, which ARM64 supports directly.
301 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
302 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
303 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
304 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
306 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
307 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
309 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
310 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
311 setOperationAction(ISD::SREM, MVT::i32, Expand);
312 setOperationAction(ISD::SREM, MVT::i64, Expand);
313 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
314 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
315 setOperationAction(ISD::UREM, MVT::i32, Expand);
316 setOperationAction(ISD::UREM, MVT::i64, Expand);
318 // Custom lower Add/Sub/Mul with overflow.
319 setOperationAction(ISD::SADDO, MVT::i32, Custom);
320 setOperationAction(ISD::SADDO, MVT::i64, Custom);
321 setOperationAction(ISD::UADDO, MVT::i32, Custom);
322 setOperationAction(ISD::UADDO, MVT::i64, Custom);
323 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
324 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
325 setOperationAction(ISD::USUBO, MVT::i32, Custom);
326 setOperationAction(ISD::USUBO, MVT::i64, Custom);
327 setOperationAction(ISD::SMULO, MVT::i32, Custom);
328 setOperationAction(ISD::SMULO, MVT::i64, Custom);
329 setOperationAction(ISD::UMULO, MVT::i32, Custom);
330 setOperationAction(ISD::UMULO, MVT::i64, Custom);
332 setOperationAction(ISD::FSIN, MVT::f32, Expand);
333 setOperationAction(ISD::FSIN, MVT::f64, Expand);
334 setOperationAction(ISD::FCOS, MVT::f32, Expand);
335 setOperationAction(ISD::FCOS, MVT::f64, Expand);
336 setOperationAction(ISD::FPOW, MVT::f32, Expand);
337 setOperationAction(ISD::FPOW, MVT::f64, Expand);
338 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
339 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
341 // ARM64 has implementations of a lot of rounding-like FP operations.
342 static MVT RoundingTypes[] = { MVT::f32, MVT::f64, MVT::v2f32,
343 MVT::v4f32, MVT::v2f64 };
344 for (unsigned I = 0; I < array_lengthof(RoundingTypes); ++I) {
345 MVT Ty = RoundingTypes[I];
346 setOperationAction(ISD::FFLOOR, Ty, Legal);
347 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
348 setOperationAction(ISD::FCEIL, Ty, Legal);
349 setOperationAction(ISD::FRINT, Ty, Legal);
350 setOperationAction(ISD::FTRUNC, Ty, Legal);
351 setOperationAction(ISD::FROUND, Ty, Legal);
354 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
356 if (Subtarget->isTargetMachO()) {
357 // For iOS, we don't want to the normal expansion of a libcall to
358 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
360 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
361 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
363 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
364 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
367 // ARM64 does not have floating-point extending loads, i1 sign-extending load,
368 // floating-point truncating stores, or v2i32->v2i16 truncating store.
369 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
370 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
371 setLoadExtAction(ISD::EXTLOAD, MVT::f80, Expand);
372 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Expand);
373 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
374 setTruncStoreAction(MVT::f128, MVT::f80, Expand);
375 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
376 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
377 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
378 // Indexed loads and stores are supported.
379 for (unsigned im = (unsigned)ISD::PRE_INC;
380 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
381 setIndexedLoadAction(im, MVT::i8, Legal);
382 setIndexedLoadAction(im, MVT::i16, Legal);
383 setIndexedLoadAction(im, MVT::i32, Legal);
384 setIndexedLoadAction(im, MVT::i64, Legal);
385 setIndexedLoadAction(im, MVT::f64, Legal);
386 setIndexedLoadAction(im, MVT::f32, Legal);
387 setIndexedStoreAction(im, MVT::i8, Legal);
388 setIndexedStoreAction(im, MVT::i16, Legal);
389 setIndexedStoreAction(im, MVT::i32, Legal);
390 setIndexedStoreAction(im, MVT::i64, Legal);
391 setIndexedStoreAction(im, MVT::f64, Legal);
392 setIndexedStoreAction(im, MVT::f32, Legal);
395 // Likewise, narrowing and extending vector loads/stores aren't handled
397 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
398 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
400 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,
403 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
404 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
405 setTruncStoreAction((MVT::SimpleValueType)VT,
406 (MVT::SimpleValueType)InnerVT, Expand);
407 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
408 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
409 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
413 setOperationAction(ISD::TRAP, MVT::Other, Legal);
414 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
416 // We combine OR nodes for bitfield operations.
417 setTargetDAGCombine(ISD::OR);
419 // Vector add and sub nodes may conceal a high-half opportunity.
420 // Also, try to fold ADD into CSINC/CSINV..
421 setTargetDAGCombine(ISD::ADD);
422 setTargetDAGCombine(ISD::SUB);
424 setTargetDAGCombine(ISD::XOR);
425 setTargetDAGCombine(ISD::SINT_TO_FP);
426 setTargetDAGCombine(ISD::UINT_TO_FP);
428 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
430 setTargetDAGCombine(ISD::ANY_EXTEND);
431 setTargetDAGCombine(ISD::ZERO_EXTEND);
432 setTargetDAGCombine(ISD::SIGN_EXTEND);
433 setTargetDAGCombine(ISD::BITCAST);
434 setTargetDAGCombine(ISD::CONCAT_VECTORS);
435 setTargetDAGCombine(ISD::STORE);
437 setTargetDAGCombine(ISD::MUL);
439 setTargetDAGCombine(ISD::VSELECT);
441 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
442 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
443 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
445 setStackPointerRegisterToSaveRestore(ARM64::SP);
447 setSchedulingPreference(Sched::Hybrid);
450 MaskAndBranchFoldingIsLegal = true;
452 setMinFunctionAlignment(2);
454 RequireStrictAlign = StrictAlign;
457 void ARM64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
458 if (VT == MVT::v2f32) {
459 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
460 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
462 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
463 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
464 } else if (VT == MVT::v2f64 || VT == MVT::v4f32) {
465 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
466 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
468 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
469 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
472 // Mark vector float intrinsics as expand.
473 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
474 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
475 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
476 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
477 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
478 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
479 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
480 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
481 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
482 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
485 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
486 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
487 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
488 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
489 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
490 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
491 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
492 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
493 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
494 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
495 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
496 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
498 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
499 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
500 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
501 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
503 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
504 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
505 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
506 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
507 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
509 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
510 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
513 void ARM64TargetLowering::addDRTypeForNEON(MVT VT) {
514 addRegisterClass(VT, &ARM64::FPR64RegClass);
515 addTypeForNEON(VT, MVT::v2i32);
518 void ARM64TargetLowering::addQRTypeForNEON(MVT VT) {
519 addRegisterClass(VT, &ARM64::FPR128RegClass);
520 addTypeForNEON(VT, MVT::v4i32);
523 EVT ARM64TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
526 return VT.changeVectorElementTypeToInteger();
529 /// computeMaskedBitsForTargetNode - Determine which of the bits specified in
530 /// Mask are known to be either zero or one and return them in the
531 /// KnownZero/KnownOne bitsets.
532 void ARM64TargetLowering::computeMaskedBitsForTargetNode(
533 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
534 const SelectionDAG &DAG, unsigned Depth) const {
535 switch (Op.getOpcode()) {
538 case ARM64ISD::CSEL: {
539 APInt KnownZero2, KnownOne2;
540 DAG.ComputeMaskedBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
541 DAG.ComputeMaskedBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
542 KnownZero &= KnownZero2;
543 KnownOne &= KnownOne2;
546 case ISD::INTRINSIC_W_CHAIN:
548 case ISD::INTRINSIC_WO_CHAIN:
549 case ISD::INTRINSIC_VOID: {
550 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
554 case Intrinsic::arm64_neon_umaxv:
555 case Intrinsic::arm64_neon_uminv: {
556 // Figure out the datatype of the vector operand. The UMINV instruction
557 // will zero extend the result, so we can mark as known zero all the
558 // bits larger than the element datatype. 32-bit or larget doesn't need
559 // this as those are legal types and will be handled by isel directly.
560 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
561 unsigned BitWidth = KnownZero.getBitWidth();
562 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
563 assert(BitWidth >= 8 && "Unexpected width!");
564 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
566 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
567 assert(BitWidth >= 16 && "Unexpected width!");
568 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
578 MVT ARM64TargetLowering::getScalarShiftAmountTy(EVT LHSTy) const {
582 unsigned ARM64TargetLowering::getMaximalGlobalOffset() const {
583 // FIXME: On ARM64, this depends on the type.
584 // Basically, the addressable offsets are o to 4095 * Ty.getSizeInBytes().
585 // and the offset has to be a multiple of the related size in bytes.
590 ARM64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
591 const TargetLibraryInfo *libInfo) const {
592 return ARM64::createFastISel(funcInfo, libInfo);
595 const char *ARM64TargetLowering::getTargetNodeName(unsigned Opcode) const {
599 case ARM64ISD::CALL: return "ARM64ISD::CALL";
600 case ARM64ISD::ADRP: return "ARM64ISD::ADRP";
601 case ARM64ISD::ADDlow: return "ARM64ISD::ADDlow";
602 case ARM64ISD::LOADgot: return "ARM64ISD::LOADgot";
603 case ARM64ISD::RET_FLAG: return "ARM64ISD::RET_FLAG";
604 case ARM64ISD::BRCOND: return "ARM64ISD::BRCOND";
605 case ARM64ISD::CSEL: return "ARM64ISD::CSEL";
606 case ARM64ISD::FCSEL: return "ARM64ISD::FCSEL";
607 case ARM64ISD::CSINV: return "ARM64ISD::CSINV";
608 case ARM64ISD::CSNEG: return "ARM64ISD::CSNEG";
609 case ARM64ISD::CSINC: return "ARM64ISD::CSINC";
610 case ARM64ISD::THREAD_POINTER: return "ARM64ISD::THREAD_POINTER";
611 case ARM64ISD::TLSDESC_CALL: return "ARM64ISD::TLSDESC_CALL";
612 case ARM64ISD::ADC: return "ARM64ISD::ADC";
613 case ARM64ISD::SBC: return "ARM64ISD::SBC";
614 case ARM64ISD::ADDS: return "ARM64ISD::ADDS";
615 case ARM64ISD::SUBS: return "ARM64ISD::SUBS";
616 case ARM64ISD::ADCS: return "ARM64ISD::ADCS";
617 case ARM64ISD::SBCS: return "ARM64ISD::SBCS";
618 case ARM64ISD::ANDS: return "ARM64ISD::ANDS";
619 case ARM64ISD::FCMP: return "ARM64ISD::FCMP";
620 case ARM64ISD::FMIN: return "ARM64ISD::FMIN";
621 case ARM64ISD::FMAX: return "ARM64ISD::FMAX";
622 case ARM64ISD::DUP: return "ARM64ISD::DUP";
623 case ARM64ISD::DUPLANE8: return "ARM64ISD::DUPLANE8";
624 case ARM64ISD::DUPLANE16: return "ARM64ISD::DUPLANE16";
625 case ARM64ISD::DUPLANE32: return "ARM64ISD::DUPLANE32";
626 case ARM64ISD::DUPLANE64: return "ARM64ISD::DUPLANE64";
627 case ARM64ISD::MOVI: return "ARM64ISD::MOVI";
628 case ARM64ISD::MOVIshift: return "ARM64ISD::MOVIshift";
629 case ARM64ISD::MOVIedit: return "ARM64ISD::MOVIedit";
630 case ARM64ISD::MOVImsl: return "ARM64ISD::MOVImsl";
631 case ARM64ISD::FMOV: return "ARM64ISD::FMOV";
632 case ARM64ISD::MVNIshift: return "ARM64ISD::MVNIshift";
633 case ARM64ISD::MVNImsl: return "ARM64ISD::MVNImsl";
634 case ARM64ISD::BICi: return "ARM64ISD::BICi";
635 case ARM64ISD::ORRi: return "ARM64ISD::ORRi";
636 case ARM64ISD::NEG: return "ARM64ISD::NEG";
637 case ARM64ISD::EXTR: return "ARM64ISD::EXTR";
638 case ARM64ISD::ZIP1: return "ARM64ISD::ZIP1";
639 case ARM64ISD::ZIP2: return "ARM64ISD::ZIP2";
640 case ARM64ISD::UZP1: return "ARM64ISD::UZP1";
641 case ARM64ISD::UZP2: return "ARM64ISD::UZP2";
642 case ARM64ISD::TRN1: return "ARM64ISD::TRN1";
643 case ARM64ISD::TRN2: return "ARM64ISD::TRN2";
644 case ARM64ISD::REV16: return "ARM64ISD::REV16";
645 case ARM64ISD::REV32: return "ARM64ISD::REV32";
646 case ARM64ISD::REV64: return "ARM64ISD::REV64";
647 case ARM64ISD::EXT: return "ARM64ISD::EXT";
648 case ARM64ISD::VSHL: return "ARM64ISD::VSHL";
649 case ARM64ISD::VLSHR: return "ARM64ISD::VLSHR";
650 case ARM64ISD::VASHR: return "ARM64ISD::VASHR";
651 case ARM64ISD::CMEQ: return "ARM64ISD::CMEQ";
652 case ARM64ISD::CMGE: return "ARM64ISD::CMGE";
653 case ARM64ISD::CMGT: return "ARM64ISD::CMGT";
654 case ARM64ISD::CMHI: return "ARM64ISD::CMHI";
655 case ARM64ISD::CMHS: return "ARM64ISD::CMHS";
656 case ARM64ISD::FCMEQ: return "ARM64ISD::FCMEQ";
657 case ARM64ISD::FCMGE: return "ARM64ISD::FCMGE";
658 case ARM64ISD::FCMGT: return "ARM64ISD::FCMGT";
659 case ARM64ISD::CMEQz: return "ARM64ISD::CMEQz";
660 case ARM64ISD::CMGEz: return "ARM64ISD::CMGEz";
661 case ARM64ISD::CMGTz: return "ARM64ISD::CMGTz";
662 case ARM64ISD::CMLEz: return "ARM64ISD::CMLEz";
663 case ARM64ISD::CMLTz: return "ARM64ISD::CMLTz";
664 case ARM64ISD::FCMEQz: return "ARM64ISD::FCMEQz";
665 case ARM64ISD::FCMGEz: return "ARM64ISD::FCMGEz";
666 case ARM64ISD::FCMGTz: return "ARM64ISD::FCMGTz";
667 case ARM64ISD::FCMLEz: return "ARM64ISD::FCMLEz";
668 case ARM64ISD::FCMLTz: return "ARM64ISD::FCMLTz";
669 case ARM64ISD::NOT: return "ARM64ISD::NOT";
670 case ARM64ISD::BIT: return "ARM64ISD::BIT";
671 case ARM64ISD::CBZ: return "ARM64ISD::CBZ";
672 case ARM64ISD::CBNZ: return "ARM64ISD::CBNZ";
673 case ARM64ISD::TBZ: return "ARM64ISD::TBZ";
674 case ARM64ISD::TBNZ: return "ARM64ISD::TBNZ";
675 case ARM64ISD::TC_RETURN: return "ARM64ISD::TC_RETURN";
676 case ARM64ISD::SITOF: return "ARM64ISD::SITOF";
677 case ARM64ISD::UITOF: return "ARM64ISD::UITOF";
678 case ARM64ISD::SQSHL_I: return "ARM64ISD::SQSHL_I";
679 case ARM64ISD::UQSHL_I: return "ARM64ISD::UQSHL_I";
680 case ARM64ISD::SRSHR_I: return "ARM64ISD::SRSHR_I";
681 case ARM64ISD::URSHR_I: return "ARM64ISD::URSHR_I";
682 case ARM64ISD::SQSHLU_I: return "ARM64ISD::SQSHLU_I";
683 case ARM64ISD::WrapperLarge: return "ARM64ISD::WrapperLarge";
687 static void getExclusiveOperation(unsigned Size, AtomicOrdering Ord,
688 unsigned &LdrOpc, unsigned &StrOpc) {
689 static unsigned LoadBares[] = { ARM64::LDXRB, ARM64::LDXRH, ARM64::LDXRW,
690 ARM64::LDXRX, ARM64::LDXPX };
691 static unsigned LoadAcqs[] = { ARM64::LDAXRB, ARM64::LDAXRH, ARM64::LDAXRW,
692 ARM64::LDAXRX, ARM64::LDAXPX };
693 static unsigned StoreBares[] = { ARM64::STXRB, ARM64::STXRH, ARM64::STXRW,
694 ARM64::STXRX, ARM64::STXPX };
695 static unsigned StoreRels[] = { ARM64::STLXRB, ARM64::STLXRH, ARM64::STLXRW,
696 ARM64::STLXRX, ARM64::STLXPX };
698 unsigned *LoadOps, *StoreOps;
699 if (Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent)
704 if (Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent)
705 StoreOps = StoreRels;
707 StoreOps = StoreBares;
709 assert(isPowerOf2_32(Size) && Size <= 16 &&
710 "unsupported size for atomic binary op!");
712 LdrOpc = LoadOps[Log2_32(Size)];
713 StrOpc = StoreOps[Log2_32(Size)];
716 MachineBasicBlock *ARM64TargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
717 MachineBasicBlock *BB,
718 unsigned Size) const {
719 unsigned dest = MI->getOperand(0).getReg();
720 unsigned ptr = MI->getOperand(1).getReg();
721 unsigned oldval = MI->getOperand(2).getReg();
722 unsigned newval = MI->getOperand(3).getReg();
723 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(4).getImm());
724 unsigned scratch = BB->getParent()->getRegInfo().createVirtualRegister(
725 &ARM64::GPR32RegClass);
726 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
727 DebugLoc dl = MI->getDebugLoc();
729 // FIXME: We currently always generate a seq_cst operation; we should
730 // be able to relax this in some cases.
731 unsigned ldrOpc, strOpc;
732 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
734 MachineFunction *MF = BB->getParent();
735 const BasicBlock *LLVM_BB = BB->getBasicBlock();
736 MachineFunction::iterator It = BB;
737 ++It; // insert the new blocks after the current block
739 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
740 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
741 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
742 MF->insert(It, loop1MBB);
743 MF->insert(It, loop2MBB);
744 MF->insert(It, exitMBB);
746 // Transfer the remainder of BB and its successor edges to exitMBB.
747 exitMBB->splice(exitMBB->begin(), BB,
748 std::next(MachineBasicBlock::iterator(MI)), BB->end());
749 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
753 // fallthrough --> loop1MBB
754 BB->addSuccessor(loop1MBB);
761 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
762 BuildMI(BB, dl, TII->get(Size == 8 ? ARM64::SUBSXrr : ARM64::SUBSWrr))
763 .addReg(Size == 8 ? ARM64::XZR : ARM64::WZR, RegState::Define)
766 BuildMI(BB, dl, TII->get(ARM64::Bcc)).addImm(ARM64CC::NE).addMBB(exitMBB);
767 BB->addSuccessor(loop2MBB);
768 BB->addSuccessor(exitMBB);
771 // strex scratch, newval, [ptr]
775 BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
776 BuildMI(BB, dl, TII->get(ARM64::CBNZW)).addReg(scratch).addMBB(loop1MBB);
777 BB->addSuccessor(loop1MBB);
778 BB->addSuccessor(exitMBB);
784 MI->eraseFromParent(); // The instruction is gone now.
790 ARM64TargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
791 unsigned Size, unsigned BinOpcode) const {
792 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
793 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
795 const BasicBlock *LLVM_BB = BB->getBasicBlock();
796 MachineFunction *MF = BB->getParent();
797 MachineFunction::iterator It = BB;
800 unsigned dest = MI->getOperand(0).getReg();
801 unsigned ptr = MI->getOperand(1).getReg();
802 unsigned incr = MI->getOperand(2).getReg();
803 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
804 DebugLoc dl = MI->getDebugLoc();
806 unsigned ldrOpc, strOpc;
807 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
809 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
810 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
811 MF->insert(It, loopMBB);
812 MF->insert(It, exitMBB);
814 // Transfer the remainder of BB and its successor edges to exitMBB.
815 exitMBB->splice(exitMBB->begin(), BB,
816 std::next(MachineBasicBlock::iterator(MI)), BB->end());
817 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
819 MachineRegisterInfo &RegInfo = MF->getRegInfo();
820 unsigned scratch = RegInfo.createVirtualRegister(&ARM64::GPR32RegClass);
824 : RegInfo.createVirtualRegister(Size == 8 ? &ARM64::GPR64RegClass
825 : &ARM64::GPR32RegClass);
829 // fallthrough --> loopMBB
830 BB->addSuccessor(loopMBB);
834 // <binop> scratch2, dest, incr
835 // stxr scratch, scratch2, ptr
836 // cbnz scratch, loopMBB
837 // fallthrough --> exitMBB
839 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
841 // operand order needs to go the other way for NAND
842 if (BinOpcode == ARM64::BICWrr || BinOpcode == ARM64::BICXrr)
843 BuildMI(BB, dl, TII->get(BinOpcode), scratch2).addReg(incr).addReg(dest);
845 BuildMI(BB, dl, TII->get(BinOpcode), scratch2).addReg(dest).addReg(incr);
848 BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
849 BuildMI(BB, dl, TII->get(ARM64::CBNZW)).addReg(scratch).addMBB(loopMBB);
851 BB->addSuccessor(loopMBB);
852 BB->addSuccessor(exitMBB);
858 MI->eraseFromParent(); // The instruction is gone now.
863 MachineBasicBlock *ARM64TargetLowering::EmitAtomicBinary128(
864 MachineInstr *MI, MachineBasicBlock *BB, unsigned BinOpcodeLo,
865 unsigned BinOpcodeHi) const {
866 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
867 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
869 const BasicBlock *LLVM_BB = BB->getBasicBlock();
870 MachineFunction *MF = BB->getParent();
871 MachineFunction::iterator It = BB;
874 unsigned DestLo = MI->getOperand(0).getReg();
875 unsigned DestHi = MI->getOperand(1).getReg();
876 unsigned Ptr = MI->getOperand(2).getReg();
877 unsigned IncrLo = MI->getOperand(3).getReg();
878 unsigned IncrHi = MI->getOperand(4).getReg();
879 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(5).getImm());
880 DebugLoc DL = MI->getDebugLoc();
882 unsigned LdrOpc, StrOpc;
883 getExclusiveOperation(16, Ord, LdrOpc, StrOpc);
885 MachineBasicBlock *LoopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
886 MachineBasicBlock *ExitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
887 MF->insert(It, LoopMBB);
888 MF->insert(It, ExitMBB);
890 // Transfer the remainder of BB and its successor edges to exitMBB.
891 ExitMBB->splice(ExitMBB->begin(), BB,
892 std::next(MachineBasicBlock::iterator(MI)), BB->end());
893 ExitMBB->transferSuccessorsAndUpdatePHIs(BB);
895 MachineRegisterInfo &RegInfo = MF->getRegInfo();
896 unsigned ScratchRes = RegInfo.createVirtualRegister(&ARM64::GPR32RegClass);
897 unsigned ScratchLo = IncrLo, ScratchHi = IncrHi;
899 assert(BinOpcodeHi && "Expect neither or both opcodes to be defined");
900 ScratchLo = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
901 ScratchHi = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
906 // fallthrough --> LoopMBB
907 BB->addSuccessor(LoopMBB);
910 // ldxp DestLo, DestHi, Ptr
911 // <binoplo> ScratchLo, DestLo, IncrLo
912 // <binophi> ScratchHi, DestHi, IncrHi
913 // stxp ScratchRes, ScratchLo, ScratchHi, ptr
914 // cbnz ScratchRes, LoopMBB
915 // fallthrough --> ExitMBB
917 BuildMI(BB, DL, TII->get(LdrOpc), DestLo)
918 .addReg(DestHi, RegState::Define)
921 // operand order needs to go the other way for NAND
922 if (BinOpcodeLo == ARM64::BICXrr) {
923 std::swap(IncrLo, DestLo);
924 std::swap(IncrHi, DestHi);
927 BuildMI(BB, DL, TII->get(BinOpcodeLo), ScratchLo).addReg(DestLo).addReg(
929 BuildMI(BB, DL, TII->get(BinOpcodeHi), ScratchHi).addReg(DestHi).addReg(
933 BuildMI(BB, DL, TII->get(StrOpc), ScratchRes)
937 BuildMI(BB, DL, TII->get(ARM64::CBNZW)).addReg(ScratchRes).addMBB(LoopMBB);
939 BB->addSuccessor(LoopMBB);
940 BB->addSuccessor(ExitMBB);
946 MI->eraseFromParent(); // The instruction is gone now.
952 ARM64TargetLowering::EmitAtomicCmpSwap128(MachineInstr *MI,
953 MachineBasicBlock *BB) const {
954 unsigned DestLo = MI->getOperand(0).getReg();
955 unsigned DestHi = MI->getOperand(1).getReg();
956 unsigned Ptr = MI->getOperand(2).getReg();
957 unsigned OldValLo = MI->getOperand(3).getReg();
958 unsigned OldValHi = MI->getOperand(4).getReg();
959 unsigned NewValLo = MI->getOperand(5).getReg();
960 unsigned NewValHi = MI->getOperand(6).getReg();
961 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(7).getImm());
962 unsigned ScratchRes = BB->getParent()->getRegInfo().createVirtualRegister(
963 &ARM64::GPR32RegClass);
964 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
965 DebugLoc DL = MI->getDebugLoc();
967 unsigned LdrOpc, StrOpc;
968 getExclusiveOperation(16, Ord, LdrOpc, StrOpc);
970 MachineFunction *MF = BB->getParent();
971 const BasicBlock *LLVM_BB = BB->getBasicBlock();
972 MachineFunction::iterator It = BB;
973 ++It; // insert the new blocks after the current block
975 MachineBasicBlock *Loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
976 MachineBasicBlock *Loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
977 MachineBasicBlock *ExitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
978 MF->insert(It, Loop1MBB);
979 MF->insert(It, Loop2MBB);
980 MF->insert(It, ExitMBB);
982 // Transfer the remainder of BB and its successor edges to exitMBB.
983 ExitMBB->splice(ExitMBB->begin(), BB,
984 std::next(MachineBasicBlock::iterator(MI)), BB->end());
985 ExitMBB->transferSuccessorsAndUpdatePHIs(BB);
989 // fallthrough --> Loop1MBB
990 BB->addSuccessor(Loop1MBB);
993 // ldxp DestLo, DestHi, [Ptr]
994 // cmp DestLo, OldValLo
995 // sbc xzr, DestHi, OldValHi
998 BuildMI(BB, DL, TII->get(LdrOpc), DestLo)
999 .addReg(DestHi, RegState::Define)
1001 BuildMI(BB, DL, TII->get(ARM64::SUBSXrr), ARM64::XZR).addReg(DestLo).addReg(
1003 BuildMI(BB, DL, TII->get(ARM64::SBCXr), ARM64::XZR).addReg(DestHi).addReg(
1006 BuildMI(BB, DL, TII->get(ARM64::Bcc)).addImm(ARM64CC::NE).addMBB(ExitMBB);
1007 BB->addSuccessor(Loop2MBB);
1008 BB->addSuccessor(ExitMBB);
1011 // stxp ScratchRes, NewValLo, NewValHi, [Ptr]
1012 // cbnz ScratchRes, Loop1MBB
1014 BuildMI(BB, DL, TII->get(StrOpc), ScratchRes)
1018 BuildMI(BB, DL, TII->get(ARM64::CBNZW)).addReg(ScratchRes).addMBB(Loop1MBB);
1019 BB->addSuccessor(Loop1MBB);
1020 BB->addSuccessor(ExitMBB);
1026 MI->eraseFromParent(); // The instruction is gone now.
1031 MachineBasicBlock *ARM64TargetLowering::EmitAtomicMinMax128(
1032 MachineInstr *MI, MachineBasicBlock *BB, unsigned CondCode) const {
1033 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
1034 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1036 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1037 MachineFunction *MF = BB->getParent();
1038 MachineFunction::iterator It = BB;
1041 unsigned DestLo = MI->getOperand(0).getReg();
1042 unsigned DestHi = MI->getOperand(1).getReg();
1043 unsigned Ptr = MI->getOperand(2).getReg();
1044 unsigned IncrLo = MI->getOperand(3).getReg();
1045 unsigned IncrHi = MI->getOperand(4).getReg();
1046 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(5).getImm());
1047 DebugLoc DL = MI->getDebugLoc();
1049 unsigned LdrOpc, StrOpc;
1050 getExclusiveOperation(16, Ord, LdrOpc, StrOpc);
1052 MachineBasicBlock *LoopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1053 MachineBasicBlock *ExitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1054 MF->insert(It, LoopMBB);
1055 MF->insert(It, ExitMBB);
1057 // Transfer the remainder of BB and its successor edges to exitMBB.
1058 ExitMBB->splice(ExitMBB->begin(), BB,
1059 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1060 ExitMBB->transferSuccessorsAndUpdatePHIs(BB);
1062 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1063 unsigned ScratchRes = RegInfo.createVirtualRegister(&ARM64::GPR32RegClass);
1064 unsigned ScratchLo = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
1065 unsigned ScratchHi = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
1069 // fallthrough --> LoopMBB
1070 BB->addSuccessor(LoopMBB);
1073 // ldxp DestLo, DestHi, Ptr
1074 // cmp ScratchLo, DestLo, IncrLo
1075 // sbc xzr, ScratchHi, DestHi, IncrHi
1076 // csel ScratchLo, DestLo, IncrLo, <cmp-op>
1077 // csel ScratchHi, DestHi, IncrHi, <cmp-op>
1078 // stxp ScratchRes, ScratchLo, ScratchHi, ptr
1079 // cbnz ScratchRes, LoopMBB
1080 // fallthrough --> ExitMBB
1082 BuildMI(BB, DL, TII->get(LdrOpc), DestLo)
1083 .addReg(DestHi, RegState::Define)
1086 BuildMI(BB, DL, TII->get(ARM64::SUBSXrr), ARM64::XZR).addReg(DestLo).addReg(
1088 BuildMI(BB, DL, TII->get(ARM64::SBCXr), ARM64::XZR).addReg(DestHi).addReg(
1091 BuildMI(BB, DL, TII->get(ARM64::CSELXr), ScratchLo)
1095 BuildMI(BB, DL, TII->get(ARM64::CSELXr), ScratchHi)
1100 BuildMI(BB, DL, TII->get(StrOpc), ScratchRes)
1104 BuildMI(BB, DL, TII->get(ARM64::CBNZW)).addReg(ScratchRes).addMBB(LoopMBB);
1106 BB->addSuccessor(LoopMBB);
1107 BB->addSuccessor(ExitMBB);
1113 MI->eraseFromParent(); // The instruction is gone now.
1119 ARM64TargetLowering::EmitF128CSEL(MachineInstr *MI,
1120 MachineBasicBlock *MBB) const {
1121 // We materialise the F128CSEL pseudo-instruction as some control flow and a
1125 // [... previous instrs leading to comparison ...]
1131 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
1133 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1134 MachineFunction *MF = MBB->getParent();
1135 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
1136 DebugLoc DL = MI->getDebugLoc();
1137 MachineFunction::iterator It = MBB;
1140 unsigned DestReg = MI->getOperand(0).getReg();
1141 unsigned IfTrueReg = MI->getOperand(1).getReg();
1142 unsigned IfFalseReg = MI->getOperand(2).getReg();
1143 unsigned CondCode = MI->getOperand(3).getImm();
1144 bool CPSRKilled = MI->getOperand(4).isKill();
1146 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
1147 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
1148 MF->insert(It, TrueBB);
1149 MF->insert(It, EndBB);
1151 // Transfer rest of current basic-block to EndBB
1152 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
1154 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
1156 BuildMI(MBB, DL, TII->get(ARM64::Bcc)).addImm(CondCode).addMBB(TrueBB);
1157 BuildMI(MBB, DL, TII->get(ARM64::B)).addMBB(EndBB);
1158 MBB->addSuccessor(TrueBB);
1159 MBB->addSuccessor(EndBB);
1161 // TrueBB falls through to the end.
1162 TrueBB->addSuccessor(EndBB);
1165 TrueBB->addLiveIn(ARM64::CPSR);
1166 EndBB->addLiveIn(ARM64::CPSR);
1169 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(ARM64::PHI), DestReg)
1175 MI->eraseFromParent();
1180 ARM64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
1181 MachineBasicBlock *BB) const {
1182 switch (MI->getOpcode()) {
1187 assert(0 && "Unexpected instruction for custom inserter!");
1190 case ARM64::ATOMIC_LOAD_ADD_I8:
1191 return EmitAtomicBinary(MI, BB, 1, ARM64::ADDWrr);
1192 case ARM64::ATOMIC_LOAD_ADD_I16:
1193 return EmitAtomicBinary(MI, BB, 2, ARM64::ADDWrr);
1194 case ARM64::ATOMIC_LOAD_ADD_I32:
1195 return EmitAtomicBinary(MI, BB, 4, ARM64::ADDWrr);
1196 case ARM64::ATOMIC_LOAD_ADD_I64:
1197 return EmitAtomicBinary(MI, BB, 8, ARM64::ADDXrr);
1198 case ARM64::ATOMIC_LOAD_ADD_I128:
1199 return EmitAtomicBinary128(MI, BB, ARM64::ADDSXrr, ARM64::ADCXr);
1201 case ARM64::ATOMIC_LOAD_AND_I8:
1202 return EmitAtomicBinary(MI, BB, 1, ARM64::ANDWrr);
1203 case ARM64::ATOMIC_LOAD_AND_I16:
1204 return EmitAtomicBinary(MI, BB, 2, ARM64::ANDWrr);
1205 case ARM64::ATOMIC_LOAD_AND_I32:
1206 return EmitAtomicBinary(MI, BB, 4, ARM64::ANDWrr);
1207 case ARM64::ATOMIC_LOAD_AND_I64:
1208 return EmitAtomicBinary(MI, BB, 8, ARM64::ANDXrr);
1209 case ARM64::ATOMIC_LOAD_AND_I128:
1210 return EmitAtomicBinary128(MI, BB, ARM64::ANDXrr, ARM64::ANDXrr);
1212 case ARM64::ATOMIC_LOAD_OR_I8:
1213 return EmitAtomicBinary(MI, BB, 1, ARM64::ORRWrr);
1214 case ARM64::ATOMIC_LOAD_OR_I16:
1215 return EmitAtomicBinary(MI, BB, 2, ARM64::ORRWrr);
1216 case ARM64::ATOMIC_LOAD_OR_I32:
1217 return EmitAtomicBinary(MI, BB, 4, ARM64::ORRWrr);
1218 case ARM64::ATOMIC_LOAD_OR_I64:
1219 return EmitAtomicBinary(MI, BB, 8, ARM64::ORRXrr);
1220 case ARM64::ATOMIC_LOAD_OR_I128:
1221 return EmitAtomicBinary128(MI, BB, ARM64::ORRXrr, ARM64::ORRXrr);
1223 case ARM64::ATOMIC_LOAD_XOR_I8:
1224 return EmitAtomicBinary(MI, BB, 1, ARM64::EORWrr);
1225 case ARM64::ATOMIC_LOAD_XOR_I16:
1226 return EmitAtomicBinary(MI, BB, 2, ARM64::EORWrr);
1227 case ARM64::ATOMIC_LOAD_XOR_I32:
1228 return EmitAtomicBinary(MI, BB, 4, ARM64::EORWrr);
1229 case ARM64::ATOMIC_LOAD_XOR_I64:
1230 return EmitAtomicBinary(MI, BB, 8, ARM64::EORXrr);
1231 case ARM64::ATOMIC_LOAD_XOR_I128:
1232 return EmitAtomicBinary128(MI, BB, ARM64::EORXrr, ARM64::EORXrr);
1234 case ARM64::ATOMIC_LOAD_NAND_I8:
1235 return EmitAtomicBinary(MI, BB, 1, ARM64::BICWrr);
1236 case ARM64::ATOMIC_LOAD_NAND_I16:
1237 return EmitAtomicBinary(MI, BB, 2, ARM64::BICWrr);
1238 case ARM64::ATOMIC_LOAD_NAND_I32:
1239 return EmitAtomicBinary(MI, BB, 4, ARM64::BICWrr);
1240 case ARM64::ATOMIC_LOAD_NAND_I64:
1241 return EmitAtomicBinary(MI, BB, 8, ARM64::BICXrr);
1242 case ARM64::ATOMIC_LOAD_NAND_I128:
1243 return EmitAtomicBinary128(MI, BB, ARM64::BICXrr, ARM64::BICXrr);
1245 case ARM64::ATOMIC_LOAD_SUB_I8:
1246 return EmitAtomicBinary(MI, BB, 1, ARM64::SUBWrr);
1247 case ARM64::ATOMIC_LOAD_SUB_I16:
1248 return EmitAtomicBinary(MI, BB, 2, ARM64::SUBWrr);
1249 case ARM64::ATOMIC_LOAD_SUB_I32:
1250 return EmitAtomicBinary(MI, BB, 4, ARM64::SUBWrr);
1251 case ARM64::ATOMIC_LOAD_SUB_I64:
1252 return EmitAtomicBinary(MI, BB, 8, ARM64::SUBXrr);
1253 case ARM64::ATOMIC_LOAD_SUB_I128:
1254 return EmitAtomicBinary128(MI, BB, ARM64::SUBSXrr, ARM64::SBCXr);
1256 case ARM64::ATOMIC_LOAD_MIN_I128:
1257 return EmitAtomicMinMax128(MI, BB, ARM64CC::LT);
1259 case ARM64::ATOMIC_LOAD_MAX_I128:
1260 return EmitAtomicMinMax128(MI, BB, ARM64CC::GT);
1262 case ARM64::ATOMIC_LOAD_UMIN_I128:
1263 return EmitAtomicMinMax128(MI, BB, ARM64CC::CC);
1265 case ARM64::ATOMIC_LOAD_UMAX_I128:
1266 return EmitAtomicMinMax128(MI, BB, ARM64CC::HI);
1268 case ARM64::ATOMIC_SWAP_I8:
1269 return EmitAtomicBinary(MI, BB, 1, 0);
1270 case ARM64::ATOMIC_SWAP_I16:
1271 return EmitAtomicBinary(MI, BB, 2, 0);
1272 case ARM64::ATOMIC_SWAP_I32:
1273 return EmitAtomicBinary(MI, BB, 4, 0);
1274 case ARM64::ATOMIC_SWAP_I64:
1275 return EmitAtomicBinary(MI, BB, 8, 0);
1276 case ARM64::ATOMIC_SWAP_I128:
1277 return EmitAtomicBinary128(MI, BB, 0, 0);
1279 case ARM64::ATOMIC_CMP_SWAP_I8:
1280 return EmitAtomicCmpSwap(MI, BB, 1);
1281 case ARM64::ATOMIC_CMP_SWAP_I16:
1282 return EmitAtomicCmpSwap(MI, BB, 2);
1283 case ARM64::ATOMIC_CMP_SWAP_I32:
1284 return EmitAtomicCmpSwap(MI, BB, 4);
1285 case ARM64::ATOMIC_CMP_SWAP_I64:
1286 return EmitAtomicCmpSwap(MI, BB, 8);
1287 case ARM64::ATOMIC_CMP_SWAP_I128:
1288 return EmitAtomicCmpSwap128(MI, BB);
1290 case ARM64::F128CSEL:
1291 return EmitF128CSEL(MI, BB);
1293 case TargetOpcode::STACKMAP:
1294 case TargetOpcode::PATCHPOINT:
1295 return emitPatchPoint(MI, BB);
1297 llvm_unreachable("Unexpected instruction for custom inserter!");
1300 //===----------------------------------------------------------------------===//
1301 // ARM64 Lowering private implementation.
1302 //===----------------------------------------------------------------------===//
1304 //===----------------------------------------------------------------------===//
1306 //===----------------------------------------------------------------------===//
1308 /// changeIntCCToARM64CC - Convert a DAG integer condition code to an ARM64 CC
1309 static ARM64CC::CondCode changeIntCCToARM64CC(ISD::CondCode CC) {
1312 llvm_unreachable("Unknown condition code!");
1336 /// changeFPCCToARM64CC - Convert a DAG fp condition code to an ARM64 CC.
1337 static void changeFPCCToARM64CC(ISD::CondCode CC, ARM64CC::CondCode &CondCode,
1338 ARM64CC::CondCode &CondCode2) {
1339 CondCode2 = ARM64CC::AL;
1342 llvm_unreachable("Unknown FP condition!");
1345 CondCode = ARM64CC::EQ;
1349 CondCode = ARM64CC::GT;
1353 CondCode = ARM64CC::GE;
1356 CondCode = ARM64CC::MI;
1359 CondCode = ARM64CC::LS;
1362 CondCode = ARM64CC::MI;
1363 CondCode2 = ARM64CC::GT;
1366 CondCode = ARM64CC::VC;
1369 CondCode = ARM64CC::VS;
1372 CondCode = ARM64CC::EQ;
1373 CondCode2 = ARM64CC::VS;
1376 CondCode = ARM64CC::HI;
1379 CondCode = ARM64CC::PL;
1383 CondCode = ARM64CC::LT;
1387 CondCode = ARM64CC::LE;
1391 CondCode = ARM64CC::NE;
1396 static bool isLegalArithImmed(uint64_t C) {
1397 // Matches ARM64DAGToDAGISel::SelectArithImmed().
1398 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
1401 static SDValue emitComparison(SDValue LHS, SDValue RHS, SDLoc dl,
1402 SelectionDAG &DAG) {
1403 EVT VT = LHS.getValueType();
1405 if (VT.isFloatingPoint())
1406 return DAG.getNode(ARM64ISD::FCMP, dl, VT, LHS, RHS);
1408 // The CMP instruction is just an alias for SUBS, and representing it as
1409 // SUBS means that it's possible to get CSE with subtract operations.
1410 // A later phase can perform the optimization of setting the destination
1411 // register to WZR/XZR if it ends up being unused.
1412 return DAG.getNode(ARM64ISD::SUBS, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS)
1416 static SDValue getARM64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1417 SDValue &ARM64cc, SelectionDAG &DAG, SDLoc dl) {
1418 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1419 EVT VT = RHS.getValueType();
1420 uint64_t C = RHSC->getZExtValue();
1421 if (!isLegalArithImmed(C)) {
1422 // Constant does not fit, try adjusting it by one?
1428 if ((VT == MVT::i32 && C != 0x80000000 &&
1429 isLegalArithImmed((uint32_t)(C - 1))) ||
1430 (VT == MVT::i64 && C != 0x80000000ULL &&
1431 isLegalArithImmed(C - 1ULL))) {
1432 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1433 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1434 RHS = DAG.getConstant(C, VT);
1439 if ((VT == MVT::i32 && C != 0 &&
1440 isLegalArithImmed((uint32_t)(C - 1))) ||
1441 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1442 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1443 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1444 RHS = DAG.getConstant(C, VT);
1449 if ((VT == MVT::i32 && C != 0x7fffffff &&
1450 isLegalArithImmed((uint32_t)(C + 1))) ||
1451 (VT == MVT::i64 && C != 0x7ffffffffffffffULL &&
1452 isLegalArithImmed(C + 1ULL))) {
1453 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1454 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1455 RHS = DAG.getConstant(C, VT);
1460 if ((VT == MVT::i32 && C != 0xffffffff &&
1461 isLegalArithImmed((uint32_t)(C + 1))) ||
1462 (VT == MVT::i64 && C != 0xfffffffffffffffULL &&
1463 isLegalArithImmed(C + 1ULL))) {
1464 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1465 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1466 RHS = DAG.getConstant(C, VT);
1473 SDValue Cmp = emitComparison(LHS, RHS, dl, DAG);
1474 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
1475 ARM64cc = DAG.getConstant(ARM64CC, MVT::i32);
1479 static std::pair<SDValue, SDValue>
1480 getARM64XALUOOp(ARM64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1481 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1482 "Unsupported value type");
1483 SDValue Value, Overflow;
1485 SDValue LHS = Op.getOperand(0);
1486 SDValue RHS = Op.getOperand(1);
1488 switch (Op.getOpcode()) {
1490 llvm_unreachable("Unknown overflow instruction!");
1492 Opc = ARM64ISD::ADDS;
1496 Opc = ARM64ISD::ADDS;
1500 Opc = ARM64ISD::SUBS;
1504 Opc = ARM64ISD::SUBS;
1507 // Multiply needs a little bit extra work.
1511 bool IsSigned = (Op.getOpcode() == ISD::SMULO) ? true : false;
1512 if (Op.getValueType() == MVT::i32) {
1513 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1514 // For a 32 bit multiply with overflow check we want the instruction
1515 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1516 // need to generate the following pattern:
1517 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1518 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1519 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1520 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1521 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1522 DAG.getConstant(0, MVT::i64));
1523 // On ARM64 the upper 32 bits are always zero extended for a 32 bit
1524 // operation. We need to clear out the upper 32 bits, because we used a
1525 // widening multiply that wrote all 64 bits. In the end this should be a
1527 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1529 // The signed overflow check requires more than just a simple check for
1530 // any bit set in the upper 32 bits of the result. These bits could be
1531 // just the sign bits of a negative number. To perform the overflow
1532 // check we have to arithmetic shift right the 32nd bit of the result by
1533 // 31 bits. Then we compare the result to the upper 32 bits.
1534 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1535 DAG.getConstant(32, MVT::i64));
1536 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1537 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1538 DAG.getConstant(31, MVT::i64));
1539 // It is important that LowerBits is last, otherwise the arithmetic
1540 // shift will not be folded into the compare (SUBS).
1541 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1542 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1545 // The overflow check for unsigned multiply is easy. We only need to
1546 // check if any of the upper 32 bits are set. This can be done with a
1547 // CMP (shifted register). For that we need to generate the following
1549 // (i64 ARM64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1550 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1551 DAG.getConstant(32, MVT::i64));
1552 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1554 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1555 UpperBits).getValue(1);
1559 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1560 // For the 64 bit multiply
1561 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1563 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1564 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1565 DAG.getConstant(63, 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::i64, MVT::i32);
1569 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1572 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1573 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1575 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1576 UpperBits).getValue(1);
1583 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1585 // Emit the ARM64 operation with overflow check.
1586 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1587 Overflow = Value.getValue(1);
1589 return std::make_pair(Value, Overflow);
1592 SDValue ARM64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1593 RTLIB::Libcall Call) const {
1594 SmallVector<SDValue, 2> Ops;
1595 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1596 Ops.push_back(Op.getOperand(i));
1598 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1602 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1603 SDValue Sel = Op.getOperand(0);
1604 SDValue Other = Op.getOperand(1);
1606 // If neither operand is a SELECT_CC, give up.
1607 if (Sel.getOpcode() != ISD::SELECT_CC)
1608 std::swap(Sel, Other);
1609 if (Sel.getOpcode() != ISD::SELECT_CC)
1612 // The folding we want to perform is:
1613 // (xor x, (select_cc a, b, cc, 0, -1) )
1615 // (csel x, (xor x, -1), cc ...)
1617 // The latter will get matched to a CSINV instruction.
1619 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1620 SDValue LHS = Sel.getOperand(0);
1621 SDValue RHS = Sel.getOperand(1);
1622 SDValue TVal = Sel.getOperand(2);
1623 SDValue FVal = Sel.getOperand(3);
1626 // FIXME: This could be generalized to non-integer comparisons.
1627 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1630 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1631 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1633 // The the values aren't constants, this isn't the pattern we're looking for.
1634 if (!CFVal || !CTVal)
1637 // We can commute the SELECT_CC by inverting the condition. This
1638 // might be needed to make this fit into a CSINV pattern.
1639 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1640 std::swap(TVal, FVal);
1641 std::swap(CTVal, CFVal);
1642 CC = ISD::getSetCCInverse(CC, true);
1645 // If the constants line up, perform the transform!
1646 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1648 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1651 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1652 DAG.getConstant(-1ULL, Other.getValueType()));
1654 return DAG.getNode(ARM64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1661 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1662 EVT VT = Op.getValueType();
1664 // Let legalize expand this if it isn't a legal type yet.
1665 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1668 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1671 bool ExtraOp = false;
1672 switch (Op.getOpcode()) {
1674 assert(0 && "Invalid code");
1676 Opc = ARM64ISD::ADDS;
1679 Opc = ARM64ISD::SUBS;
1682 Opc = ARM64ISD::ADCS;
1686 Opc = ARM64ISD::SBCS;
1692 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1693 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1697 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1698 // Let legalize expand this if it isn't a legal type yet.
1699 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1702 ARM64CC::CondCode CC;
1703 // The actual operation that sets the overflow or carry flag.
1704 SDValue Value, Overflow;
1705 std::tie(Value, Overflow) = getARM64XALUOOp(CC, Op, DAG);
1707 // We use 0 and 1 as false and true values.
1708 SDValue TVal = DAG.getConstant(1, MVT::i32);
1709 SDValue FVal = DAG.getConstant(0, MVT::i32);
1711 // We use an inverted condition, because the conditional select is inverted
1712 // too. This will allow it to be selected to a single instruction:
1713 // CSINC Wd, WZR, WZR, invert(cond).
1714 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), MVT::i32);
1715 Overflow = DAG.getNode(ARM64ISD::CSEL, SDLoc(Op), MVT::i32, FVal, TVal, CCVal,
1718 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1719 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
1722 // Prefetch operands are:
1723 // 1: Address to prefetch
1725 // 3: int locality (0 = no locality ... 3 = extreme locality)
1726 // 4: bool isDataCache
1727 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1729 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1730 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1731 // The data thing is not used.
1732 // unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1734 bool IsStream = !Locality;
1735 // When the locality number is set
1737 // The front-end should have filtered out the out-of-range values
1738 assert(Locality <= 3 && "Prefetch locality out-of-range");
1739 // The locality degree is the opposite of the cache speed.
1740 // Put the number the other way around.
1741 // The encoding starts at 0 for level 1
1742 Locality = 3 - Locality;
1745 // built the mask value encoding the expected behavior.
1746 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1747 (Locality << 1) | // Cache level bits
1748 (unsigned)IsStream; // Stream bit
1749 return DAG.getNode(ARM64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1750 DAG.getConstant(PrfOp, MVT::i32), Op.getOperand(1));
1753 SDValue ARM64TargetLowering::LowerFP_EXTEND(SDValue Op,
1754 SelectionDAG &DAG) const {
1755 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1758 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1760 return LowerF128Call(Op, DAG, LC);
1763 SDValue ARM64TargetLowering::LowerFP_ROUND(SDValue Op,
1764 SelectionDAG &DAG) const {
1765 if (Op.getOperand(0).getValueType() != MVT::f128) {
1766 // It's legal except when f128 is involved
1771 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1773 // FP_ROUND node has a second operand indicating whether it is known to be
1774 // precise. That doesn't take part in the LibCall so we can't directly use
1776 SDValue SrcVal = Op.getOperand(0);
1777 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1778 /*isSigned*/ false, SDLoc(Op)).first;
1781 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1782 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1783 // Any additional optimization in this function should be recorded
1784 // in the cost tables.
1785 EVT InVT = Op.getOperand(0).getValueType();
1786 EVT VT = Op.getValueType();
1788 // FP_TO_XINT conversion from the same type are legal.
1789 if (VT.getSizeInBits() == InVT.getSizeInBits())
1792 if (InVT == MVT::v2f64) {
1794 SDValue Cv = DAG.getNode(Op.getOpcode(), dl, MVT::v2i64, Op.getOperand(0));
1795 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1798 // Type changing conversions are illegal.
1802 SDValue ARM64TargetLowering::LowerFP_TO_INT(SDValue Op,
1803 SelectionDAG &DAG) const {
1804 if (Op.getOperand(0).getValueType().isVector())
1805 return LowerVectorFP_TO_INT(Op, DAG);
1807 if (Op.getOperand(0).getValueType() != MVT::f128) {
1808 // It's legal except when f128 is involved
1813 if (Op.getOpcode() == ISD::FP_TO_SINT)
1814 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1816 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1818 SmallVector<SDValue, 2> Ops;
1819 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1820 Ops.push_back(Op.getOperand(i));
1822 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1826 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1827 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1828 // Any additional optimization in this function should be recorded
1829 // in the cost tables.
1830 EVT VT = Op.getValueType();
1832 SDValue In = Op.getOperand(0);
1833 EVT InVT = In.getValueType();
1835 // v2i32 to v2f32 is legal.
1836 if (VT == MVT::v2f32 && InVT == MVT::v2i32)
1839 // This function only handles v2f64 outputs.
1840 if (VT == MVT::v2f64) {
1841 // Extend the input argument to a v2i64 that we can feed into the
1842 // floating point conversion. Zero or sign extend based on whether
1843 // we're doing a signed or unsigned float conversion.
1845 Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
1846 assert(Op.getNumOperands() == 1 && "FP conversions take one argument");
1847 SDValue Promoted = DAG.getNode(Opc, dl, MVT::v2i64, Op.getOperand(0));
1848 return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), Promoted);
1851 // Scalarize v2i64 to v2f32 conversions.
1852 std::vector<SDValue> BuildVectorOps;
1853 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
1854 SDValue Sclr = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, In,
1855 DAG.getConstant(i, MVT::i64));
1856 Sclr = DAG.getNode(Op->getOpcode(), dl, MVT::f32, Sclr);
1857 BuildVectorOps.push_back(Sclr);
1860 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &BuildVectorOps[0],
1861 BuildVectorOps.size());
1864 SDValue ARM64TargetLowering::LowerINT_TO_FP(SDValue Op,
1865 SelectionDAG &DAG) const {
1866 if (Op.getValueType().isVector())
1867 return LowerVectorINT_TO_FP(Op, DAG);
1869 // i128 conversions are libcalls.
1870 if (Op.getOperand(0).getValueType() == MVT::i128)
1873 // Other conversions are legal, unless it's to the completely software-based
1875 if (Op.getValueType() != MVT::f128)
1879 if (Op.getOpcode() == ISD::SINT_TO_FP)
1880 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1882 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1884 return LowerF128Call(Op, DAG, LC);
1887 SDValue ARM64TargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
1888 // For iOS, we want to call an alternative entry point: __sincos_stret,
1889 // which returns the values in two S / D registers.
1891 SDValue Arg = Op.getOperand(0);
1892 EVT ArgVT = Arg.getValueType();
1893 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1900 Entry.isSExt = false;
1901 Entry.isZExt = false;
1902 Args.push_back(Entry);
1904 const char *LibcallName =
1905 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1906 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
1908 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
1909 TargetLowering::CallLoweringInfo CLI(
1910 DAG.getEntryNode(), RetTy, false, false, false, false, 0,
1911 CallingConv::Fast, /*isTaillCall=*/false,
1912 /*doesNotRet=*/false, /*isReturnValueUsed*/ true, Callee, Args, DAG, dl);
1913 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1914 return CallResult.first;
1917 SDValue ARM64TargetLowering::LowerOperation(SDValue Op,
1918 SelectionDAG &DAG) const {
1919 switch (Op.getOpcode()) {
1921 llvm_unreachable("unimplemented operand");
1923 case ISD::GlobalAddress:
1924 return LowerGlobalAddress(Op, DAG);
1925 case ISD::GlobalTLSAddress:
1926 return LowerGlobalTLSAddress(Op, DAG);
1928 return LowerSETCC(Op, DAG);
1930 return LowerBR_CC(Op, DAG);
1932 return LowerSELECT(Op, DAG);
1933 case ISD::SELECT_CC:
1934 return LowerSELECT_CC(Op, DAG);
1935 case ISD::JumpTable:
1936 return LowerJumpTable(Op, DAG);
1937 case ISD::ConstantPool:
1938 return LowerConstantPool(Op, DAG);
1939 case ISD::BlockAddress:
1940 return LowerBlockAddress(Op, DAG);
1942 return LowerVASTART(Op, DAG);
1944 return LowerVACOPY(Op, DAG);
1946 return LowerVAARG(Op, DAG);
1951 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
1958 return LowerXALUO(Op, DAG);
1960 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
1962 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
1964 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
1966 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
1968 return LowerFP_ROUND(Op, DAG);
1969 case ISD::FP_EXTEND:
1970 return LowerFP_EXTEND(Op, DAG);
1971 case ISD::FRAMEADDR:
1972 return LowerFRAMEADDR(Op, DAG);
1973 case ISD::RETURNADDR:
1974 return LowerRETURNADDR(Op, DAG);
1975 case ISD::INSERT_VECTOR_ELT:
1976 return LowerINSERT_VECTOR_ELT(Op, DAG);
1977 case ISD::EXTRACT_VECTOR_ELT:
1978 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
1979 case ISD::BUILD_VECTOR:
1980 return LowerBUILD_VECTOR(Op, DAG);
1981 case ISD::VECTOR_SHUFFLE:
1982 return LowerVECTOR_SHUFFLE(Op, DAG);
1983 case ISD::EXTRACT_SUBVECTOR:
1984 return LowerEXTRACT_SUBVECTOR(Op, DAG);
1988 return LowerVectorSRA_SRL_SHL(Op, DAG);
1989 case ISD::SHL_PARTS:
1990 return LowerShiftLeftParts(Op, DAG);
1991 case ISD::SRL_PARTS:
1992 case ISD::SRA_PARTS:
1993 return LowerShiftRightParts(Op, DAG);
1995 return LowerCTPOP(Op, DAG);
1996 case ISD::FCOPYSIGN:
1997 return LowerFCOPYSIGN(Op, DAG);
1999 return LowerVectorAND(Op, DAG);
2001 return LowerVectorOR(Op, DAG);
2003 return LowerXOR(Op, DAG);
2005 return LowerPREFETCH(Op, DAG);
2006 case ISD::SINT_TO_FP:
2007 case ISD::UINT_TO_FP:
2008 return LowerINT_TO_FP(Op, DAG);
2009 case ISD::FP_TO_SINT:
2010 case ISD::FP_TO_UINT:
2011 return LowerFP_TO_INT(Op, DAG);
2013 return LowerFSINCOS(Op, DAG);
2017 /// getFunctionAlignment - Return the Log2 alignment of this function.
2018 unsigned ARM64TargetLowering::getFunctionAlignment(const Function *F) const {
2022 //===----------------------------------------------------------------------===//
2023 // Calling Convention Implementation
2024 //===----------------------------------------------------------------------===//
2026 #include "ARM64GenCallingConv.inc"
2028 /// Selects the correct CCAssignFn for a the given CallingConvention
2030 CCAssignFn *ARM64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
2031 bool IsVarArg) const {
2034 llvm_unreachable("Unsupported calling convention.");
2035 case CallingConv::WebKit_JS:
2036 return CC_ARM64_WebKit_JS;
2037 case CallingConv::C:
2038 case CallingConv::Fast:
2039 if (!Subtarget->isTargetDarwin())
2040 return CC_ARM64_AAPCS;
2041 return IsVarArg ? CC_ARM64_DarwinPCS_VarArg : CC_ARM64_DarwinPCS;
2045 SDValue ARM64TargetLowering::LowerFormalArguments(
2046 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2047 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2048 SmallVectorImpl<SDValue> &InVals) const {
2049 MachineFunction &MF = DAG.getMachineFunction();
2050 MachineFrameInfo *MFI = MF.getFrameInfo();
2052 // Assign locations to all of the incoming arguments.
2053 SmallVector<CCValAssign, 16> ArgLocs;
2054 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2055 getTargetMachine(), ArgLocs, *DAG.getContext());
2057 // At this point, Ins[].VT may already be promoted to i32. To correctly
2058 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2059 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
2060 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
2061 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
2063 unsigned NumArgs = Ins.size();
2064 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2065 unsigned CurArgIdx = 0;
2066 for (unsigned i = 0; i != NumArgs; ++i) {
2067 MVT ValVT = Ins[i].VT;
2068 std::advance(CurOrigArg, Ins[i].OrigArgIndex - CurArgIdx);
2069 CurArgIdx = Ins[i].OrigArgIndex;
2071 // Get type of the original argument.
2072 EVT ActualVT = getValueType(CurOrigArg->getType(), /*AllowUnknown*/ true);
2073 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
2074 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2076 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2078 else if (ActualMVT == MVT::i16)
2081 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2083 AssignFn(i, ValVT, LocVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
2084 assert(!Res && "Call operand has unhandled type");
2088 SmallVector<SDValue, 16> ArgValues;
2089 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2090 CCValAssign &VA = ArgLocs[i];
2092 // Arguments stored in registers.
2093 if (VA.isRegLoc()) {
2094 EVT RegVT = VA.getLocVT();
2097 const TargetRegisterClass *RC;
2099 if (RegVT == MVT::i32)
2100 RC = &ARM64::GPR32RegClass;
2101 else if (RegVT == MVT::i64)
2102 RC = &ARM64::GPR64RegClass;
2103 else if (RegVT == MVT::f32)
2104 RC = &ARM64::FPR32RegClass;
2105 else if (RegVT == MVT::f64 || RegVT == MVT::v1i64 ||
2106 RegVT == MVT::v1f64 || RegVT == MVT::v2i32 ||
2107 RegVT == MVT::v4i16 || RegVT == MVT::v8i8)
2108 RC = &ARM64::FPR64RegClass;
2109 else if (RegVT == MVT::v2i64 || RegVT == MVT::v4i32 ||
2110 RegVT == MVT::v8i16 || RegVT == MVT::v16i8)
2111 RC = &ARM64::FPR128RegClass;
2113 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2115 // Transform the arguments in physical registers into virtual ones.
2116 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2117 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2119 // If this is an 8, 16 or 32-bit value, it is really passed promoted
2120 // to 64 bits. Insert an assert[sz]ext to capture this, then
2121 // truncate to the right size.
2122 switch (VA.getLocInfo()) {
2124 llvm_unreachable("Unknown loc info!");
2125 case CCValAssign::Full:
2127 case CCValAssign::BCvt:
2128 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
2130 case CCValAssign::SExt:
2131 ArgValue = DAG.getNode(ISD::AssertSext, DL, RegVT, ArgValue,
2132 DAG.getValueType(VA.getValVT()));
2133 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
2135 case CCValAssign::ZExt:
2136 ArgValue = DAG.getNode(ISD::AssertZext, DL, RegVT, ArgValue,
2137 DAG.getValueType(VA.getValVT()));
2138 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
2142 InVals.push_back(ArgValue);
2144 } else { // VA.isRegLoc()
2145 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
2146 unsigned ArgOffset = VA.getLocMemOffset();
2147 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
2148 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset, true);
2150 // Create load nodes to retrieve arguments from the stack.
2151 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2152 InVals.push_back(DAG.getLoad(VA.getValVT(), DL, Chain, FIN,
2153 MachinePointerInfo::getFixedStack(FI), false,
2160 if (!Subtarget->isTargetDarwin()) {
2161 // The AAPCS variadic function ABI is identical to the non-variadic
2162 // one. As a result there may be more arguments in registers and we should
2163 // save them for future reference.
2164 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
2167 ARM64FunctionInfo *AFI = MF.getInfo<ARM64FunctionInfo>();
2168 // This will point to the next argument passed via stack.
2169 unsigned StackOffset = CCInfo.getNextStackOffset();
2170 // We currently pass all varargs at 8-byte alignment.
2171 StackOffset = ((StackOffset + 7) & ~7);
2172 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
2178 void ARM64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
2179 SelectionDAG &DAG, SDLoc DL,
2180 SDValue &Chain) const {
2181 MachineFunction &MF = DAG.getMachineFunction();
2182 MachineFrameInfo *MFI = MF.getFrameInfo();
2183 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
2185 SmallVector<SDValue, 8> MemOps;
2187 static const MCPhysReg GPRArgRegs[] = { ARM64::X0, ARM64::X1, ARM64::X2,
2188 ARM64::X3, ARM64::X4, ARM64::X5,
2189 ARM64::X6, ARM64::X7 };
2190 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
2191 unsigned FirstVariadicGPR =
2192 CCInfo.getFirstUnallocated(GPRArgRegs, NumGPRArgRegs);
2194 static const MCPhysReg FPRArgRegs[] = { ARM64::Q0, ARM64::Q1, ARM64::Q2,
2195 ARM64::Q3, ARM64::Q4, ARM64::Q5,
2196 ARM64::Q6, ARM64::Q7 };
2197 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
2198 unsigned FirstVariadicFPR =
2199 CCInfo.getFirstUnallocated(FPRArgRegs, NumFPRArgRegs);
2201 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
2203 if (GPRSaveSize != 0) {
2204 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
2206 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
2208 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
2209 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &ARM64::GPR64RegClass);
2210 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
2212 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2213 MachinePointerInfo::getStack(i * 8), false, false, 0);
2214 MemOps.push_back(Store);
2215 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2216 DAG.getConstant(8, getPointerTy()));
2220 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
2222 if (FPRSaveSize != 0) {
2223 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
2225 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
2227 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
2228 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &ARM64::FPR128RegClass);
2229 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::v2i64);
2231 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2232 MachinePointerInfo::getStack(i * 16), false, false, 0);
2233 MemOps.push_back(Store);
2234 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2235 DAG.getConstant(16, getPointerTy()));
2239 FuncInfo->setVarArgsGPRIndex(GPRIdx);
2240 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
2241 FuncInfo->setVarArgsFPRIndex(FPRIdx);
2242 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
2244 if (!MemOps.empty()) {
2245 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
2250 /// LowerCallResult - Lower the result values of a call into the
2251 /// appropriate copies out of appropriate physical registers.
2252 SDValue ARM64TargetLowering::LowerCallResult(
2253 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
2254 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2255 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
2256 SDValue ThisVal) const {
2257 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2258 : RetCC_ARM64_AAPCS;
2259 // Assign locations to each value returned by this call.
2260 SmallVector<CCValAssign, 16> RVLocs;
2261 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2262 getTargetMachine(), RVLocs, *DAG.getContext());
2263 CCInfo.AnalyzeCallResult(Ins, RetCC);
2265 // Copy all of the result registers out of their specified physreg.
2266 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2267 CCValAssign VA = RVLocs[i];
2269 // Pass 'this' value directly from the argument to return value, to avoid
2270 // reg unit interference
2271 if (i == 0 && isThisReturn) {
2272 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
2273 "unexpected return calling convention register assignment");
2274 InVals.push_back(ThisVal);
2279 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
2280 Chain = Val.getValue(1);
2281 InFlag = Val.getValue(2);
2283 switch (VA.getLocInfo()) {
2285 llvm_unreachable("Unknown loc info!");
2286 case CCValAssign::Full:
2288 case CCValAssign::BCvt:
2289 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2293 InVals.push_back(Val);
2299 bool ARM64TargetLowering::isEligibleForTailCallOptimization(
2300 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
2301 bool isCalleeStructRet, bool isCallerStructRet,
2302 const SmallVectorImpl<ISD::OutputArg> &Outs,
2303 const SmallVectorImpl<SDValue> &OutVals,
2304 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
2305 // Look for obvious safe cases to perform tail call optimization that do not
2306 // require ABI changes. This is what gcc calls sibcall.
2308 // Do not sibcall optimize vararg calls unless the call site is not passing
2310 if (isVarArg && !Outs.empty())
2313 // Also avoid sibcall optimization if either caller or callee uses struct
2314 // return semantics.
2315 if (isCalleeStructRet || isCallerStructRet)
2318 // Note that currently ARM64 "C" calling convention and "Fast" calling
2319 // convention are compatible. If/when that ever changes, we'll need to
2320 // add checks here to make sure any interactions are OK.
2322 // If the callee takes no arguments then go on to check the results of the
2324 if (!Outs.empty()) {
2325 // Check if stack adjustment is needed. For now, do not do this if any
2326 // argument is passed on the stack.
2327 SmallVector<CCValAssign, 16> ArgLocs;
2328 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2329 getTargetMachine(), ArgLocs, *DAG.getContext());
2330 CCAssignFn *AssignFn = CCAssignFnForCall(CalleeCC, /*IsVarArg=*/false);
2331 CCInfo.AnalyzeCallOperands(Outs, AssignFn);
2332 if (CCInfo.getNextStackOffset()) {
2333 // Check if the arguments are already laid out in the right way as
2334 // the caller's fixed stack objects.
2335 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2336 ++i, ++realArgIdx) {
2337 CCValAssign &VA = ArgLocs[i];
2338 if (VA.getLocInfo() == CCValAssign::Indirect)
2340 if (VA.needsCustom()) {
2341 // Just don't handle anything that needs custom adjustments for now.
2342 // If need be, we can revisit later, but we shouldn't ever end up
2345 } else if (!VA.isRegLoc()) {
2346 // Likewise, don't try to handle stack based arguments for the
2356 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
2357 /// and add input and output parameter nodes.
2358 SDValue ARM64TargetLowering::LowerCall(CallLoweringInfo &CLI,
2359 SmallVectorImpl<SDValue> &InVals) const {
2360 SelectionDAG &DAG = CLI.DAG;
2362 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2363 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2364 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2365 SDValue Chain = CLI.Chain;
2366 SDValue Callee = CLI.Callee;
2367 bool &IsTailCall = CLI.IsTailCall;
2368 CallingConv::ID CallConv = CLI.CallConv;
2369 bool IsVarArg = CLI.IsVarArg;
2371 MachineFunction &MF = DAG.getMachineFunction();
2372 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2373 bool IsThisReturn = false;
2375 // If tail calls are explicitly disabled, make sure not to use them.
2376 if (!EnableARM64TailCalls)
2380 // Check if it's really possible to do a tail call.
2381 IsTailCall = isEligibleForTailCallOptimization(
2382 Callee, CallConv, IsVarArg, IsStructRet,
2383 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2384 // We don't support GuaranteedTailCallOpt, only automatically
2385 // detected sibcalls.
2386 // FIXME: Re-evaluate. Is this true? Should it be true?
2391 // Analyze operands of the call, assigning locations to each operand.
2392 SmallVector<CCValAssign, 16> ArgLocs;
2393 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2394 getTargetMachine(), ArgLocs, *DAG.getContext());
2397 // Handle fixed and variable vector arguments differently.
2398 // Variable vector arguments always go into memory.
2399 unsigned NumArgs = Outs.size();
2401 for (unsigned i = 0; i != NumArgs; ++i) {
2402 MVT ArgVT = Outs[i].VT;
2403 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2404 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2405 /*IsVarArg=*/ !Outs[i].IsFixed);
2406 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2407 assert(!Res && "Call operand has unhandled type");
2411 // At this point, Outs[].VT may already be promoted to i32. To correctly
2412 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2413 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
2414 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2415 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2417 unsigned NumArgs = Outs.size();
2418 for (unsigned i = 0; i != NumArgs; ++i) {
2419 MVT ValVT = Outs[i].VT;
2420 // Get type of the original argument.
2421 EVT ActualVT = getValueType(CLI.Args[Outs[i].OrigArgIndex].Ty,
2422 /*AllowUnknown*/ true);
2423 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2424 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2425 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2427 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2429 else if (ActualMVT == MVT::i16)
2432 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2433 bool Res = AssignFn(i, ValVT, LocVT, CCValAssign::Full, ArgFlags, CCInfo);
2434 assert(!Res && "Call operand has unhandled type");
2439 // Get a count of how many bytes are to be pushed on the stack.
2440 unsigned NumBytes = CCInfo.getNextStackOffset();
2442 // Adjust the stack pointer for the new arguments...
2443 // These operations are automatically eliminated by the prolog/epilog pass
2446 DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true), DL);
2448 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, ARM64::SP, getPointerTy());
2450 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2451 SmallVector<SDValue, 8> MemOpChains;
2453 // Walk the register/memloc assignments, inserting copies/loads.
2454 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2455 ++i, ++realArgIdx) {
2456 CCValAssign &VA = ArgLocs[i];
2457 SDValue Arg = OutVals[realArgIdx];
2458 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2460 // Promote the value if needed.
2461 switch (VA.getLocInfo()) {
2463 llvm_unreachable("Unknown loc info!");
2464 case CCValAssign::Full:
2466 case CCValAssign::SExt:
2467 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
2469 case CCValAssign::ZExt:
2470 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2472 case CCValAssign::AExt:
2473 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
2475 case CCValAssign::BCvt:
2476 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2478 case CCValAssign::FPExt:
2479 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
2483 if (VA.isRegLoc()) {
2484 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
2485 assert(VA.getLocVT() == MVT::i64 &&
2486 "unexpected calling convention register assignment");
2487 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
2488 "unexpected use of 'returned'");
2489 IsThisReturn = true;
2491 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2493 assert(VA.isMemLoc());
2494 // There's no reason we can't support stack args w/ tailcall, but
2495 // we currently don't, so assert if we see one.
2496 assert(!IsTailCall && "stack argument with tail call!?");
2497 unsigned LocMemOffset = VA.getLocMemOffset();
2498 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2499 PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2501 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
2502 // promoted to a legal register type i32, we should truncate Arg back to
2504 if (Arg.getValueType().isSimple() &&
2505 Arg.getValueType().getSimpleVT() == MVT::i32 &&
2506 (VA.getLocVT() == MVT::i1 || VA.getLocVT() == MVT::i8 ||
2507 VA.getLocVT() == MVT::i16))
2508 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getLocVT(), Arg);
2510 SDValue Store = DAG.getStore(Chain, DL, Arg, PtrOff,
2511 MachinePointerInfo::getStack(LocMemOffset),
2513 MemOpChains.push_back(Store);
2517 if (!MemOpChains.empty())
2518 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOpChains[0],
2519 MemOpChains.size());
2521 // Build a sequence of copy-to-reg nodes chained together with token chain
2522 // and flag operands which copy the outgoing args into the appropriate regs.
2524 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2525 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[i].first,
2526 RegsToPass[i].second, InFlag);
2527 InFlag = Chain.getValue(1);
2530 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2531 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2532 // node so that legalize doesn't hack it.
2533 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
2534 Subtarget->isTargetMachO()) {
2535 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2536 const GlobalValue *GV = G->getGlobal();
2537 bool InternalLinkage = GV->hasInternalLinkage();
2538 if (InternalLinkage)
2539 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2541 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0,
2543 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2545 } else if (ExternalSymbolSDNode *S =
2546 dyn_cast<ExternalSymbolSDNode>(Callee)) {
2547 const char *Sym = S->getSymbol();
2549 DAG.getTargetExternalSymbol(Sym, getPointerTy(), ARM64II::MO_GOT);
2550 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2552 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2553 const GlobalValue *GV = G->getGlobal();
2554 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2555 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2556 const char *Sym = S->getSymbol();
2557 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), 0);
2560 std::vector<SDValue> Ops;
2561 Ops.push_back(Chain);
2562 Ops.push_back(Callee);
2564 // Add argument registers to the end of the list so that they are known live
2566 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2567 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2568 RegsToPass[i].second.getValueType()));
2570 // Add a register mask operand representing the call-preserved registers.
2571 const uint32_t *Mask;
2572 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2573 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2575 // For 'this' returns, use the X0-preserving mask if applicable
2576 Mask = ARI->getThisReturnPreservedMask(CallConv);
2578 IsThisReturn = false;
2579 Mask = ARI->getCallPreservedMask(CallConv);
2582 Mask = ARI->getCallPreservedMask(CallConv);
2584 assert(Mask && "Missing call preserved mask for calling convention");
2585 Ops.push_back(DAG.getRegisterMask(Mask));
2587 if (InFlag.getNode())
2588 Ops.push_back(InFlag);
2590 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2592 // If we're doing a tall call, use a TC_RETURN here rather than an
2593 // actual call instruction.
2595 return DAG.getNode(ARM64ISD::TC_RETURN, DL, NodeTys, &Ops[0], Ops.size());
2597 // Returns a chain and a flag for retval copy to use.
2598 Chain = DAG.getNode(ARM64ISD::CALL, DL, NodeTys, &Ops[0], Ops.size());
2599 InFlag = Chain.getValue(1);
2601 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2602 DAG.getIntPtrConstant(0, true), InFlag, DL);
2604 InFlag = Chain.getValue(1);
2606 // Handle result values, copying them out of physregs into vregs that we
2608 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2609 InVals, IsThisReturn,
2610 IsThisReturn ? OutVals[0] : SDValue());
2613 bool ARM64TargetLowering::CanLowerReturn(
2614 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2615 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2616 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2617 : RetCC_ARM64_AAPCS;
2618 SmallVector<CCValAssign, 16> RVLocs;
2619 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2620 return CCInfo.CheckReturn(Outs, RetCC);
2624 ARM64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2626 const SmallVectorImpl<ISD::OutputArg> &Outs,
2627 const SmallVectorImpl<SDValue> &OutVals,
2628 SDLoc DL, SelectionDAG &DAG) const {
2629 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2630 : RetCC_ARM64_AAPCS;
2631 SmallVector<CCValAssign, 16> RVLocs;
2632 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2633 getTargetMachine(), RVLocs, *DAG.getContext());
2634 CCInfo.AnalyzeReturn(Outs, RetCC);
2636 // Copy the result values into the output registers.
2638 SmallVector<SDValue, 4> RetOps(1, Chain);
2639 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
2640 ++i, ++realRVLocIdx) {
2641 CCValAssign &VA = RVLocs[i];
2642 assert(VA.isRegLoc() && "Can only return in registers!");
2643 SDValue Arg = OutVals[realRVLocIdx];
2645 switch (VA.getLocInfo()) {
2647 llvm_unreachable("Unknown loc info!");
2648 case CCValAssign::Full:
2650 case CCValAssign::BCvt:
2651 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2655 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
2656 Flag = Chain.getValue(1);
2657 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2660 RetOps[0] = Chain; // Update chain.
2662 // Add the flag if we have it.
2664 RetOps.push_back(Flag);
2666 return DAG.getNode(ARM64ISD::RET_FLAG, DL, MVT::Other, &RetOps[0],
2670 //===----------------------------------------------------------------------===//
2671 // Other Lowering Code
2672 //===----------------------------------------------------------------------===//
2674 SDValue ARM64TargetLowering::LowerGlobalAddress(SDValue Op,
2675 SelectionDAG &DAG) const {
2676 EVT PtrVT = getPointerTy();
2678 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2679 unsigned char OpFlags =
2680 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
2682 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
2683 "unexpected offset in global node");
2685 // This also catched the large code model case for Darwin.
2686 if ((OpFlags & ARM64II::MO_GOT) != 0) {
2687 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
2688 // FIXME: Once remat is capable of dealing with instructions with register
2689 // operands, expand this into two nodes instead of using a wrapper node.
2690 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
2693 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2694 const unsigned char MO_NC = ARM64II::MO_NC;
2696 ARM64ISD::WrapperLarge, DL, PtrVT,
2697 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G3),
2698 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
2699 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
2700 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
2702 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
2703 // the only correct model on Darwin.
2704 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2705 OpFlags | ARM64II::MO_PAGE);
2706 unsigned char LoFlags = OpFlags | ARM64II::MO_PAGEOFF | ARM64II::MO_NC;
2707 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
2709 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
2710 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
2714 /// \brief Convert a TLS address reference into the correct sequence of loads
2715 /// and calls to compute the variable's address (for Darwin, currently) and
2716 /// return an SDValue containing the final node.
2718 /// Darwin only has one TLS scheme which must be capable of dealing with the
2719 /// fully general situation, in the worst case. This means:
2720 /// + "extern __thread" declaration.
2721 /// + Defined in a possibly unknown dynamic library.
2723 /// The general system is that each __thread variable has a [3 x i64] descriptor
2724 /// which contains information used by the runtime to calculate the address. The
2725 /// only part of this the compiler needs to know about is the first xword, which
2726 /// contains a function pointer that must be called with the address of the
2727 /// entire descriptor in "x0".
2729 /// Since this descriptor may be in a different unit, in general even the
2730 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
2732 /// adrp x0, _var@TLVPPAGE
2733 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
2734 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
2735 /// ; the function pointer
2736 /// blr x1 ; Uses descriptor address in x0
2737 /// ; Address of _var is now in x0.
2739 /// If the address of _var's descriptor *is* known to the linker, then it can
2740 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
2741 /// a slight efficiency gain.
2743 ARM64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
2744 SelectionDAG &DAG) const {
2745 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
2748 MVT PtrVT = getPointerTy();
2749 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2752 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2753 SDValue DescAddr = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TLVPAddr);
2755 // The first entry in the descriptor is a function pointer that we must call
2756 // to obtain the address of the variable.
2757 SDValue Chain = DAG.getEntryNode();
2758 SDValue FuncTLVGet =
2759 DAG.getLoad(MVT::i64, DL, Chain, DescAddr, MachinePointerInfo::getGOT(),
2760 false, true, true, 8);
2761 Chain = FuncTLVGet.getValue(1);
2763 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2764 MFI->setAdjustsStack(true);
2766 // TLS calls preserve all registers except those that absolutely must be
2767 // trashed: X0 (it takes an argument), LR (it's a call) and CPSR (let's not be
2769 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2770 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2771 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2773 // Finally, we can make the call. This is just a degenerate version of a
2774 // normal ARM64 call node: x0 takes the address of the descriptor, and returns
2775 // the address of the variable in this thread.
2776 Chain = DAG.getCopyToReg(Chain, DL, ARM64::X0, DescAddr, SDValue());
2777 Chain = DAG.getNode(ARM64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
2778 Chain, FuncTLVGet, DAG.getRegister(ARM64::X0, MVT::i64),
2779 DAG.getRegisterMask(Mask), Chain.getValue(1));
2780 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Chain.getValue(1));
2783 /// When accessing thread-local variables under either the general-dynamic or
2784 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
2785 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
2786 /// is a function pointer to carry out the resolution. This function takes the
2787 /// address of the descriptor in X0 and returns the TPIDR_EL0 offset in X0. All
2788 /// other registers (except LR, CPSR) are preserved.
2790 /// Thus, the ideal call sequence on AArch64 is:
2792 /// adrp x0, :tlsdesc:thread_var
2793 /// ldr x8, [x0, :tlsdesc_lo12:thread_var]
2794 /// add x0, x0, :tlsdesc_lo12:thread_var
2795 /// .tlsdesccall thread_var
2797 /// (TPIDR_EL0 offset now in x0).
2799 /// The ".tlsdesccall" directive instructs the assembler to insert a particular
2800 /// relocation to help the linker relax this sequence if it turns out to be too
2803 /// FIXME: we currently produce an extra, duplicated, ADRP instruction, but this
2805 SDValue ARM64TargetLowering::LowerELFTLSDescCall(SDValue SymAddr,
2806 SDValue DescAddr, SDLoc DL,
2807 SelectionDAG &DAG) const {
2808 EVT PtrVT = getPointerTy();
2810 // The function we need to call is simply the first entry in the GOT for this
2811 // descriptor, load it in preparation.
2812 SDValue Func = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, SymAddr);
2814 // TLS calls preserve all registers except those that absolutely must be
2815 // trashed: X0 (it takes an argument), LR (it's a call) and CPSR (let's not be
2817 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2818 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2819 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2821 // The function takes only one argument: the address of the descriptor itself
2823 SDValue Glue, Chain;
2824 Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, ARM64::X0, DescAddr, Glue);
2825 Glue = Chain.getValue(1);
2827 // We're now ready to populate the argument list, as with a normal call:
2828 SmallVector<SDValue, 6> Ops;
2829 Ops.push_back(Chain);
2830 Ops.push_back(Func);
2831 Ops.push_back(SymAddr);
2832 Ops.push_back(DAG.getRegister(ARM64::X0, PtrVT));
2833 Ops.push_back(DAG.getRegisterMask(Mask));
2834 Ops.push_back(Glue);
2836 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2837 Chain = DAG.getNode(ARM64ISD::TLSDESC_CALL, DL, NodeTys, &Ops[0], Ops.size());
2838 Glue = Chain.getValue(1);
2840 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Glue);
2843 SDValue ARM64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
2844 SelectionDAG &DAG) const {
2845 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
2846 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
2847 "ELF TLS only supported in small memory model");
2848 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2850 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
2853 EVT PtrVT = getPointerTy();
2855 const GlobalValue *GV = GA->getGlobal();
2857 SDValue ThreadBase = DAG.getNode(ARM64ISD::THREAD_POINTER, DL, PtrVT);
2859 if (Model == TLSModel::LocalExec) {
2860 SDValue HiVar = DAG.getTargetGlobalAddress(
2861 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2862 SDValue LoVar = DAG.getTargetGlobalAddress(
2863 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2865 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2866 DAG.getTargetConstant(16, MVT::i32)),
2868 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, TPOff, LoVar,
2869 DAG.getTargetConstant(0, MVT::i32)),
2871 } else if (Model == TLSModel::InitialExec) {
2872 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2873 TPOff = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TPOff);
2874 } else if (Model == TLSModel::LocalDynamic) {
2875 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
2876 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
2877 // the beginning of the module's TLS region, followed by a DTPREL offset
2880 // These accesses will need deduplicating if there's more than one.
2881 ARM64FunctionInfo *MFI =
2882 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
2883 MFI->incNumLocalDynamicTLSAccesses();
2885 // Accesses used in this sequence go via the TLS descriptor which lives in
2886 // the GOT. Prepare an address we can use to handle this.
2887 SDValue HiDesc = DAG.getTargetExternalSymbol(
2888 "_TLS_MODULE_BASE_", PtrVT, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2889 SDValue LoDesc = DAG.getTargetExternalSymbol(
2890 "_TLS_MODULE_BASE_", PtrVT,
2891 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2893 // First argument to the descriptor call is the address of the descriptor
2895 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2896 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2898 // The call needs a relocation too for linker relaxation. It doesn't make
2899 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2901 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2904 // Now we can calculate the offset from TPIDR_EL0 to this module's
2905 // thread-local area.
2906 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2908 // Now use :dtprel_whatever: operations to calculate this variable's offset
2909 // in its thread-storage area.
2910 SDValue HiVar = DAG.getTargetGlobalAddress(
2911 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2912 SDValue LoVar = DAG.getTargetGlobalAddress(
2913 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2916 SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2917 DAG.getTargetConstant(16, MVT::i32)),
2919 DTPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, DTPOff, LoVar,
2920 DAG.getTargetConstant(0, MVT::i32)),
2923 TPOff = DAG.getNode(ISD::ADD, DL, PtrVT, TPOff, DTPOff);
2924 } else if (Model == TLSModel::GeneralDynamic) {
2925 // Accesses used in this sequence go via the TLS descriptor which lives in
2926 // the GOT. Prepare an address we can use to handle this.
2927 SDValue HiDesc = DAG.getTargetGlobalAddress(
2928 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2929 SDValue LoDesc = DAG.getTargetGlobalAddress(
2931 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2933 // First argument to the descriptor call is the address of the descriptor
2935 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2936 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2938 // The call needs a relocation too for linker relaxation. It doesn't make
2939 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2942 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2944 // Finally we can make a call to calculate the offset from tpidr_el0.
2945 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2947 llvm_unreachable("Unsupported ELF TLS access model");
2949 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
2952 SDValue ARM64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
2953 SelectionDAG &DAG) const {
2954 if (Subtarget->isTargetDarwin())
2955 return LowerDarwinGlobalTLSAddress(Op, DAG);
2956 else if (Subtarget->isTargetELF())
2957 return LowerELFGlobalTLSAddress(Op, DAG);
2959 llvm_unreachable("Unexpected platform trying to use TLS");
2961 SDValue ARM64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2962 SDValue Chain = Op.getOperand(0);
2963 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2964 SDValue LHS = Op.getOperand(2);
2965 SDValue RHS = Op.getOperand(3);
2966 SDValue Dest = Op.getOperand(4);
2969 // Handle f128 first, since lowering it will result in comparing the return
2970 // value of a libcall against zero, which is just what the rest of LowerBR_CC
2971 // is expecting to deal with.
2972 if (LHS.getValueType() == MVT::f128) {
2973 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2975 // If softenSetCCOperands returned a scalar, we need to compare the result
2976 // against zero to select between true and false values.
2977 if (RHS.getNode() == 0) {
2978 RHS = DAG.getConstant(0, LHS.getValueType());
2983 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
2985 unsigned Opc = LHS.getOpcode();
2986 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
2987 cast<ConstantSDNode>(RHS)->isOne() &&
2988 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
2989 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
2990 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
2991 "Unexpected condition code.");
2992 // Only lower legal XALUO ops.
2993 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
2996 // The actual operation with overflow check.
2997 ARM64CC::CondCode OFCC;
2998 SDValue Value, Overflow;
2999 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, LHS.getValue(0), DAG);
3001 if (CC == ISD::SETNE)
3002 OFCC = getInvertedCondCode(OFCC);
3003 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3005 return DAG.getNode(ARM64ISD::BRCOND, SDLoc(LHS), MVT::Other, Chain, Dest,
3009 if (LHS.getValueType().isInteger()) {
3010 assert((LHS.getValueType() == RHS.getValueType()) &&
3011 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3013 // If the RHS of the comparison is zero, we can potentially fold this
3014 // to a specialized branch.
3015 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
3016 if (RHSC && RHSC->getZExtValue() == 0) {
3017 if (CC == ISD::SETEQ) {
3018 // See if we can use a TBZ to fold in an AND as well.
3019 // TBZ has a smaller branch displacement than CBZ. If the offset is
3020 // out of bounds, a late MI-layer pass rewrites branches.
3021 // 403.gcc is an example that hits this case.
3022 if (LHS.getOpcode() == ISD::AND &&
3023 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3024 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3025 SDValue Test = LHS.getOperand(0);
3026 uint64_t Mask = LHS.getConstantOperandVal(1);
3028 // TBZ only operates on i64's, but the ext should be free.
3029 if (Test.getValueType() == MVT::i32)
3030 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
3032 return DAG.getNode(ARM64ISD::TBZ, dl, MVT::Other, Chain, Test,
3033 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3036 return DAG.getNode(ARM64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
3037 } else if (CC == ISD::SETNE) {
3038 // See if we can use a TBZ to fold in an AND as well.
3039 // TBZ has a smaller branch displacement than CBZ. If the offset is
3040 // out of bounds, a late MI-layer pass rewrites branches.
3041 // 403.gcc is an example that hits this case.
3042 if (LHS.getOpcode() == ISD::AND &&
3043 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3044 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3045 SDValue Test = LHS.getOperand(0);
3046 uint64_t Mask = LHS.getConstantOperandVal(1);
3048 // TBNZ only operates on i64's, but the ext should be free.
3049 if (Test.getValueType() == MVT::i32)
3050 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
3052 return DAG.getNode(ARM64ISD::TBNZ, dl, MVT::Other, Chain, Test,
3053 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3056 return DAG.getNode(ARM64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
3061 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3062 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3066 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3068 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3069 // clean. Some of them require two branches to implement.
3070 SDValue Cmp = emitComparison(LHS, RHS, dl, DAG);
3071 ARM64CC::CondCode CC1, CC2;
3072 changeFPCCToARM64CC(CC, CC1, CC2);
3073 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3075 DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
3076 if (CC2 != ARM64CC::AL) {
3077 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3078 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
3085 SDValue ARM64TargetLowering::LowerFCOPYSIGN(SDValue Op,
3086 SelectionDAG &DAG) const {
3087 EVT VT = Op.getValueType();
3090 SDValue In1 = Op.getOperand(0);
3091 SDValue In2 = Op.getOperand(1);
3092 EVT SrcVT = In2.getValueType();
3094 if (SrcVT == MVT::f32 && VT == MVT::f64)
3095 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
3096 else if (SrcVT == MVT::f64 && VT == MVT::f32)
3097 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0));
3099 // FIXME: Src type is different, bail out for now. Can VT really be a
3106 SDValue EltMask, VecVal1, VecVal2;
3107 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
3110 EltMask = DAG.getConstant(0x80000000ULL, EltVT);
3112 if (!VT.isVector()) {
3113 VecVal1 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
3114 DAG.getUNDEF(VecVT), In1);
3115 VecVal2 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
3116 DAG.getUNDEF(VecVT), In2);
3118 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3119 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3121 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
3125 // We want to materialize a mask with the the high bit set, but the AdvSIMD
3126 // immediate moves cannot materialize that in a single instruction for
3127 // 64-bit elements. Instead, materialize zero and then negate it.
3128 EltMask = DAG.getConstant(0, EltVT);
3130 if (!VT.isVector()) {
3131 VecVal1 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
3132 DAG.getUNDEF(VecVT), In1);
3133 VecVal2 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
3134 DAG.getUNDEF(VecVT), In2);
3136 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3137 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3140 llvm_unreachable("Invalid type for copysign!");
3143 std::vector<SDValue> BuildVectorOps;
3144 for (unsigned i = 0; i < VecVT.getVectorNumElements(); ++i)
3145 BuildVectorOps.push_back(EltMask);
3147 SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT,
3148 &BuildVectorOps[0], BuildVectorOps.size());
3150 // If we couldn't materialize the mask above, then the mask vector will be
3151 // the zero vector, and we need to negate it here.
3152 if (VT == MVT::f64 || VT == MVT::v2f64) {
3153 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
3154 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
3155 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
3159 DAG.getNode(ARM64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
3162 return DAG.getTargetExtractSubreg(ARM64::ssub, DL, VT, Sel);
3163 else if (VT == MVT::f64)
3164 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, VT, Sel);
3166 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
3169 SDValue ARM64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
3170 if (DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
3171 AttributeSet::FunctionIndex, Attribute::NoImplicitFloat))
3174 // While there is no integer popcount instruction, it can
3175 // be more efficiently lowered to the following sequence that uses
3176 // AdvSIMD registers/instructions as long as the copies to/from
3177 // the AdvSIMD registers are cheap.
3178 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
3179 // CNT V0.8B, V0.8B // 8xbyte pop-counts
3180 // ADDV B0, V0.8B // sum 8xbyte pop-counts
3181 // UMOV X0, V0.B[0] // copy byte result back to integer reg
3182 SDValue Val = Op.getOperand(0);
3184 EVT VT = Op.getValueType();
3185 SDValue ZeroVec = DAG.getUNDEF(MVT::v8i8);
3188 if (VT == MVT::i32) {
3189 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Val);
3191 DAG.getTargetInsertSubreg(ARM64::ssub, DL, MVT::v8i8, ZeroVec, VecVal);
3193 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
3196 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, VecVal);
3197 SDValue UaddLV = DAG.getNode(
3198 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
3199 DAG.getConstant(Intrinsic::arm64_neon_uaddlv, MVT::i32), CtPop);
3202 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
3206 SDValue ARM64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3208 if (Op.getValueType().isVector())
3209 return LowerVSETCC(Op, DAG);
3211 SDValue LHS = Op.getOperand(0);
3212 SDValue RHS = Op.getOperand(1);
3213 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3216 // We chose ZeroOrOneBooleanContents, so use zero and one.
3217 EVT VT = Op.getValueType();
3218 SDValue TVal = DAG.getConstant(1, VT);
3219 SDValue FVal = DAG.getConstant(0, VT);
3221 // Handle f128 first, since one possible outcome is a normal integer
3222 // comparison which gets picked up by the next if statement.
3223 if (LHS.getValueType() == MVT::f128) {
3224 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3226 // If softenSetCCOperands returned a scalar, use it.
3227 if (RHS.getNode() == 0) {
3228 assert(LHS.getValueType() == Op.getValueType() &&
3229 "Unexpected setcc expansion!");
3234 if (LHS.getValueType().isInteger()) {
3237 getARM64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
3239 // Note that we inverted the condition above, so we reverse the order of
3240 // the true and false operands here. This will allow the setcc to be
3241 // matched to a single CSINC instruction.
3242 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
3245 // Now we know we're dealing with FP values.
3246 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3248 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3249 // and do the comparison.
3250 SDValue Cmp = emitComparison(LHS, RHS, dl, DAG);
3252 ARM64CC::CondCode CC1, CC2;
3253 changeFPCCToARM64CC(CC, CC1, CC2);
3254 if (CC2 == ARM64CC::AL) {
3255 changeFPCCToARM64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
3256 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3258 // Note that we inverted the condition above, so we reverse the order of
3259 // the true and false operands here. This will allow the setcc to be
3260 // matched to a single CSINC instruction.
3261 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
3263 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3264 // clean. Some of them require two CSELs to implement. As is in this case,
3265 // we emit the first CSEL and then emit a second using the output of the
3266 // first as the RHS. We're effectively OR'ing the two CC's together.
3268 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
3269 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3270 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3272 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3273 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3277 /// A SELECT_CC operation is really some kind of max or min if both values being
3278 /// compared are, in some sense, equal to the results in either case. However,
3279 /// it is permissible to compare f32 values and produce directly extended f64
3282 /// Extending the comparison operands would also be allowed, but is less likely
3283 /// to happen in practice since their use is right here. Note that truncate
3284 /// operations would *not* be semantically equivalent.
3285 static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) {
3289 ConstantFPSDNode *CCmp = dyn_cast<ConstantFPSDNode>(Cmp);
3290 ConstantFPSDNode *CResult = dyn_cast<ConstantFPSDNode>(Result);
3291 if (CCmp && CResult && Cmp.getValueType() == MVT::f32 &&
3292 Result.getValueType() == MVT::f64) {
3294 APFloat CmpVal = CCmp->getValueAPF();
3295 CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy);
3296 return CResult->getValueAPF().bitwiseIsEqual(CmpVal);
3299 return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp;
3302 SDValue ARM64TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3303 SDValue CC = Op->getOperand(0);
3304 SDValue TVal = Op->getOperand(1);
3305 SDValue FVal = Op->getOperand(2);
3308 unsigned Opc = CC.getOpcode();
3309 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
3311 if (CC.getResNo() == 1 &&
3312 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3313 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3314 // Only lower legal XALUO ops.
3315 if (!DAG.getTargetLoweringInfo().isTypeLegal(CC->getValueType(0)))
3318 ARM64CC::CondCode OFCC;
3319 SDValue Value, Overflow;
3320 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, CC.getValue(0), DAG);
3321 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3323 return DAG.getNode(ARM64ISD::CSEL, DL, Op.getValueType(), TVal, FVal, CCVal,
3327 if (CC.getOpcode() == ISD::SETCC)
3328 return DAG.getSelectCC(DL, CC.getOperand(0), CC.getOperand(1), TVal, FVal,
3329 cast<CondCodeSDNode>(CC.getOperand(2))->get());
3331 return DAG.getSelectCC(DL, CC, DAG.getConstant(0, CC.getValueType()), TVal,
3335 SDValue ARM64TargetLowering::LowerSELECT_CC(SDValue Op,
3336 SelectionDAG &DAG) const {
3337 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3338 SDValue LHS = Op.getOperand(0);
3339 SDValue RHS = Op.getOperand(1);
3340 SDValue TVal = Op.getOperand(2);
3341 SDValue FVal = Op.getOperand(3);
3344 // Handle f128 first, because it will result in a comparison of some RTLIB
3345 // call result against zero.
3346 if (LHS.getValueType() == MVT::f128) {
3347 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3349 // If softenSetCCOperands returned a scalar, we need to compare the result
3350 // against zero to select between true and false values.
3351 if (RHS.getNode() == 0) {
3352 RHS = DAG.getConstant(0, LHS.getValueType());
3357 // Handle integers first.
3358 if (LHS.getValueType().isInteger()) {
3359 assert((LHS.getValueType() == RHS.getValueType()) &&
3360 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3362 unsigned Opcode = ARM64ISD::CSEL;
3364 // If both the TVal and the FVal are constants, see if we can swap them in
3365 // order to for a CSINV or CSINC out of them.
3366 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3367 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3369 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3370 std::swap(TVal, FVal);
3371 std::swap(CTVal, CFVal);
3372 CC = ISD::getSetCCInverse(CC, true);
3373 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3374 std::swap(TVal, FVal);
3375 std::swap(CTVal, CFVal);
3376 CC = ISD::getSetCCInverse(CC, true);
3377 } else if (TVal.getOpcode() == ISD::XOR) {
3378 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3379 // with a CSINV rather than a CSEL.
3380 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
3382 if (CVal && CVal->isAllOnesValue()) {
3383 std::swap(TVal, FVal);
3384 std::swap(CTVal, CFVal);
3385 CC = ISD::getSetCCInverse(CC, true);
3387 } else if (TVal.getOpcode() == ISD::SUB) {
3388 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3389 // that we can match with a CSNEG rather than a CSEL.
3390 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
3392 if (CVal && CVal->isNullValue()) {
3393 std::swap(TVal, FVal);
3394 std::swap(CTVal, CFVal);
3395 CC = ISD::getSetCCInverse(CC, true);
3397 } else if (CTVal && CFVal) {
3398 const int64_t TrueVal = CTVal->getSExtValue();
3399 const int64_t FalseVal = CFVal->getSExtValue();
3402 // If both TVal and FVal are constants, see if FVal is the
3403 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3404 // instead of a CSEL in that case.
3405 if (TrueVal == ~FalseVal) {
3406 Opcode = ARM64ISD::CSINV;
3407 } else if (TrueVal == -FalseVal) {
3408 Opcode = ARM64ISD::CSNEG;
3409 } else if (TVal.getValueType() == MVT::i32) {
3410 // If our operands are only 32-bit wide, make sure we use 32-bit
3411 // arithmetic for the check whether we can use CSINC. This ensures that
3412 // the addition in the check will wrap around properly in case there is
3413 // an overflow (which would not be the case if we do the check with
3414 // 64-bit arithmetic).
3415 const uint32_t TrueVal32 = CTVal->getZExtValue();
3416 const uint32_t FalseVal32 = CFVal->getZExtValue();
3418 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3419 Opcode = ARM64ISD::CSINC;
3421 if (TrueVal32 > FalseVal32) {
3425 // 64-bit check whether we can use CSINC.
3426 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3427 Opcode = ARM64ISD::CSINC;
3429 if (TrueVal > FalseVal) {
3434 // Swap TVal and FVal if necessary.
3436 std::swap(TVal, FVal);
3437 std::swap(CTVal, CFVal);
3438 CC = ISD::getSetCCInverse(CC, true);
3441 if (Opcode != ARM64ISD::CSEL) {
3442 // Drop FVal since we can get its value by simply inverting/negating
3449 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3451 EVT VT = Op.getValueType();
3452 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3455 // Now we know we're dealing with FP values.
3456 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3457 assert(LHS.getValueType() == RHS.getValueType());
3458 EVT VT = Op.getValueType();
3460 // Try to match this select into a max/min operation, which have dedicated
3461 // opcode in the instruction set.
3462 // NOTE: This is not correct in the presence of NaNs, so we only enable this
3464 if (getTargetMachine().Options.NoNaNsFPMath) {
3465 if (selectCCOpsAreFMaxCompatible(LHS, FVal) &&
3466 selectCCOpsAreFMaxCompatible(RHS, TVal)) {
3467 CC = ISD::getSetCCSwappedOperands(CC);
3468 std::swap(TVal, FVal);
3471 if (selectCCOpsAreFMaxCompatible(LHS, TVal) &&
3472 selectCCOpsAreFMaxCompatible(RHS, FVal)) {
3482 return DAG.getNode(ARM64ISD::FMAX, dl, VT, TVal, FVal);
3490 return DAG.getNode(ARM64ISD::FMIN, dl, VT, TVal, FVal);
3496 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3497 // and do the comparison.
3498 SDValue Cmp = emitComparison(LHS, RHS, dl, DAG);
3500 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3501 // clean. Some of them require two CSELs to implement.
3502 ARM64CC::CondCode CC1, CC2;
3503 changeFPCCToARM64CC(CC, CC1, CC2);
3504 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3505 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3507 // If we need a second CSEL, emit it, using the output of the first as the
3508 // RHS. We're effectively OR'ing the two CC's together.
3509 if (CC2 != ARM64CC::AL) {
3510 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3511 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3514 // Otherwise, return the output of the first CSEL.
3518 SDValue ARM64TargetLowering::LowerJumpTable(SDValue Op,
3519 SelectionDAG &DAG) const {
3520 // Jump table entries as PC relative offsets. No additional tweaking
3521 // is necessary here. Just get the address of the jump table.
3522 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3523 EVT PtrVT = getPointerTy();
3526 SDValue Hi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_PAGE);
3527 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3528 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3529 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3530 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3533 SDValue ARM64TargetLowering::LowerConstantPool(SDValue Op,
3534 SelectionDAG &DAG) const {
3535 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3536 EVT PtrVT = getPointerTy();
3539 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3540 // Use the GOT for the large code model on iOS.
3541 if (Subtarget->isTargetMachO()) {
3542 SDValue GotAddr = DAG.getTargetConstantPool(
3543 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3545 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
3548 const unsigned char MO_NC = ARM64II::MO_NC;
3550 ARM64ISD::WrapperLarge, DL, PtrVT,
3551 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3552 CP->getOffset(), ARM64II::MO_G3),
3553 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3554 CP->getOffset(), ARM64II::MO_G2 | MO_NC),
3555 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3556 CP->getOffset(), ARM64II::MO_G1 | MO_NC),
3557 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3558 CP->getOffset(), ARM64II::MO_G0 | MO_NC));
3560 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
3561 // ELF, the only valid one on Darwin.
3563 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3564 CP->getOffset(), ARM64II::MO_PAGE);
3565 SDValue Lo = DAG.getTargetConstantPool(
3566 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3567 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3569 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3570 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3574 SDValue ARM64TargetLowering::LowerBlockAddress(SDValue Op,
3575 SelectionDAG &DAG) const {
3576 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
3577 EVT PtrVT = getPointerTy();
3579 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3580 !Subtarget->isTargetMachO()) {
3581 const unsigned char MO_NC = ARM64II::MO_NC;
3583 ARM64ISD::WrapperLarge, DL, PtrVT,
3584 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G3),
3585 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
3586 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
3587 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
3589 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGE);
3590 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGEOFF |
3592 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3593 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3597 SDValue ARM64TargetLowering::LowerDarwin_VASTART(SDValue Op,
3598 SelectionDAG &DAG) const {
3599 ARM64FunctionInfo *FuncInfo =
3600 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
3604 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3605 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3606 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
3607 MachinePointerInfo(SV), false, false, 0);
3610 SDValue ARM64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
3611 SelectionDAG &DAG) const {
3612 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3613 // Standard, section B.3.
3614 MachineFunction &MF = DAG.getMachineFunction();
3615 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
3618 SDValue Chain = Op.getOperand(0);
3619 SDValue VAList = Op.getOperand(1);
3620 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3621 SmallVector<SDValue, 4> MemOps;
3623 // void *__stack at offset 0
3625 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3626 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3627 MachinePointerInfo(SV), false, false, 8));
3629 // void *__gr_top at offset 8
3630 int GPRSize = FuncInfo->getVarArgsGPRSize();
3632 SDValue GRTop, GRTopAddr;
3634 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3635 DAG.getConstant(8, getPointerTy()));
3637 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), getPointerTy());
3638 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
3639 DAG.getConstant(GPRSize, getPointerTy()));
3641 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3642 MachinePointerInfo(SV, 8), false, false, 8));
3645 // void *__vr_top at offset 16
3646 int FPRSize = FuncInfo->getVarArgsFPRSize();
3648 SDValue VRTop, VRTopAddr;
3649 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3650 DAG.getConstant(16, getPointerTy()));
3652 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), getPointerTy());
3653 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
3654 DAG.getConstant(FPRSize, getPointerTy()));
3656 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3657 MachinePointerInfo(SV, 16), false, false, 8));
3660 // int __gr_offs at offset 24
3661 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3662 DAG.getConstant(24, getPointerTy()));
3663 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
3664 GROffsAddr, MachinePointerInfo(SV, 24), false,
3667 // int __vr_offs at offset 28
3668 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3669 DAG.getConstant(28, getPointerTy()));
3670 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
3671 VROffsAddr, MachinePointerInfo(SV, 28), false,
3674 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
3678 SDValue ARM64TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
3679 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
3680 : LowerAAPCS_VASTART(Op, DAG);
3683 SDValue ARM64TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
3684 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
3686 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
3687 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3688 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
3690 return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op), Op.getOperand(1),
3691 Op.getOperand(2), DAG.getConstant(VaListSize, MVT::i32),
3692 8, false, false, MachinePointerInfo(DestSV),
3693 MachinePointerInfo(SrcSV));
3696 SDValue ARM64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
3697 assert(Subtarget->isTargetDarwin() &&
3698 "automatic va_arg instruction only works on Darwin");
3700 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3701 EVT VT = Op.getValueType();
3703 SDValue Chain = Op.getOperand(0);
3704 SDValue Addr = Op.getOperand(1);
3705 unsigned Align = Op.getConstantOperandVal(3);
3707 SDValue VAList = DAG.getLoad(getPointerTy(), DL, Chain, Addr,
3708 MachinePointerInfo(V), false, false, false, 0);
3709 Chain = VAList.getValue(1);
3712 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
3713 VAList = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3714 DAG.getConstant(Align - 1, getPointerTy()));
3715 VAList = DAG.getNode(ISD::AND, DL, getPointerTy(), VAList,
3716 DAG.getConstant(-(int64_t)Align, getPointerTy()));
3719 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
3720 uint64_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
3722 // Scalar integer and FP values smaller than 64 bits are implicitly extended
3723 // up to 64 bits. At the very least, we have to increase the striding of the
3724 // vaargs list to match this, and for FP values we need to introduce
3725 // FP_ROUND nodes as well.
3726 if (VT.isInteger() && !VT.isVector())
3728 bool NeedFPTrunc = false;
3729 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
3734 // Increment the pointer, VAList, to the next vaarg
3735 SDValue VANext = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3736 DAG.getConstant(ArgSize, getPointerTy()));
3737 // Store the incremented VAList to the legalized pointer
3738 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
3741 // Load the actual argument out of the pointer VAList
3743 // Load the value as an f64.
3744 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
3745 MachinePointerInfo(), false, false, false, 0);
3746 // Round the value down to an f32.
3747 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
3748 DAG.getIntPtrConstant(1));
3749 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
3750 // Merge the rounded value with the chain output of the load.
3751 return DAG.getMergeValues(Ops, 2, DL);
3754 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
3758 SDValue ARM64TargetLowering::LowerFRAMEADDR(SDValue Op,
3759 SelectionDAG &DAG) const {
3760 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3761 MFI->setFrameAddressIsTaken(true);
3763 EVT VT = Op.getValueType();
3765 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3766 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, ARM64::FP, VT);
3768 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
3769 MachinePointerInfo(), false, false, false, 0);
3773 SDValue ARM64TargetLowering::LowerRETURNADDR(SDValue Op,
3774 SelectionDAG &DAG) const {
3775 MachineFunction &MF = DAG.getMachineFunction();
3776 MachineFrameInfo *MFI = MF.getFrameInfo();
3777 MFI->setReturnAddressIsTaken(true);
3779 EVT VT = Op.getValueType();
3781 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3783 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3784 SDValue Offset = DAG.getConstant(8, getPointerTy());
3785 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
3786 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
3787 MachinePointerInfo(), false, false, false, 0);
3790 // Return LR, which contains the return address. Mark it an implicit live-in.
3791 unsigned Reg = MF.addLiveIn(ARM64::LR, &ARM64::GPR64RegClass);
3792 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
3795 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3796 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3797 SDValue ARM64TargetLowering::LowerShiftRightParts(SDValue Op,
3798 SelectionDAG &DAG) const {
3799 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3800 EVT VT = Op.getValueType();
3801 unsigned VTBits = VT.getSizeInBits();
3803 SDValue ShOpLo = Op.getOperand(0);
3804 SDValue ShOpHi = Op.getOperand(1);
3805 SDValue ShAmt = Op.getOperand(2);
3807 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3809 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3811 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3812 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3813 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3814 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3815 DAG.getConstant(VTBits, MVT::i64));
3816 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3819 emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64), dl, DAG);
3820 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3822 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3823 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3825 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3827 // ARM64 shifts larger than the register width are wrapped rather than
3828 // clamped, so we can't just emit "hi >> x".
3829 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3830 SDValue TrueValHi = Opc == ISD::SRA
3831 ? DAG.getNode(Opc, dl, VT, ShOpHi,
3832 DAG.getConstant(VTBits - 1, MVT::i64))
3833 : DAG.getConstant(0, VT);
3835 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
3837 SDValue Ops[2] = { Lo, Hi };
3838 return DAG.getMergeValues(Ops, 2, dl);
3841 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3842 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3843 SDValue ARM64TargetLowering::LowerShiftLeftParts(SDValue Op,
3844 SelectionDAG &DAG) const {
3845 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3846 EVT VT = Op.getValueType();
3847 unsigned VTBits = VT.getSizeInBits();
3849 SDValue ShOpLo = Op.getOperand(0);
3850 SDValue ShOpHi = Op.getOperand(1);
3851 SDValue ShAmt = Op.getOperand(2);
3854 assert(Op.getOpcode() == ISD::SHL_PARTS);
3855 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3856 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3857 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3858 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3859 DAG.getConstant(VTBits, MVT::i64));
3860 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3861 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3863 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3866 emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64), dl, DAG);
3867 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3868 SDValue Hi = DAG.getNode(ARM64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
3870 // ARM64 shifts of larger than register sizes are wrapped rather than clamped,
3871 // so we can't just emit "lo << a" if a is too big.
3872 SDValue TrueValLo = DAG.getConstant(0, VT);
3873 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3875 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3877 SDValue Ops[2] = { Lo, Hi };
3878 return DAG.getMergeValues(Ops, 2, dl);
3882 ARM64TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3883 // The ARM64 target doesn't support folding offsets into global addresses.
3887 bool ARM64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3888 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
3889 // FIXME: We should be able to handle f128 as well with a clever lowering.
3890 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
3894 return ARM64_AM::getFP64Imm(Imm) != -1;
3895 else if (VT == MVT::f32)
3896 return ARM64_AM::getFP32Imm(Imm) != -1;
3900 //===----------------------------------------------------------------------===//
3901 // ARM64 Optimization Hooks
3902 //===----------------------------------------------------------------------===//
3904 //===----------------------------------------------------------------------===//
3905 // ARM64 Inline Assembly Support
3906 //===----------------------------------------------------------------------===//
3908 // Table of Constraints
3909 // TODO: This is the current set of constraints supported by ARM for the
3910 // compiler, not all of them may make sense, e.g. S may be difficult to support.
3912 // r - A general register
3913 // w - An FP/SIMD register of some size in the range v0-v31
3914 // x - An FP/SIMD register of some size in the range v0-v15
3915 // I - Constant that can be used with an ADD instruction
3916 // J - Constant that can be used with a SUB instruction
3917 // K - Constant that can be used with a 32-bit logical instruction
3918 // L - Constant that can be used with a 64-bit logical instruction
3919 // M - Constant that can be used as a 32-bit MOV immediate
3920 // N - Constant that can be used as a 64-bit MOV immediate
3921 // Q - A memory reference with base register and no offset
3922 // S - A symbolic address
3923 // Y - Floating point constant zero
3924 // Z - Integer constant zero
3926 // Note that general register operands will be output using their 64-bit x
3927 // register name, whatever the size of the variable, unless the asm operand
3928 // is prefixed by the %w modifier. Floating-point and SIMD register operands
3929 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
3932 /// getConstraintType - Given a constraint letter, return the type of
3933 /// constraint it is for this target.
3934 ARM64TargetLowering::ConstraintType
3935 ARM64TargetLowering::getConstraintType(const std::string &Constraint) const {
3936 if (Constraint.size() == 1) {
3937 switch (Constraint[0]) {
3944 return C_RegisterClass;
3945 // An address with a single base register. Due to the way we
3946 // currently handle addresses it is the same as 'r'.
3951 return TargetLowering::getConstraintType(Constraint);
3954 /// Examine constraint type and operand type and determine a weight value.
3955 /// This object must already have been set up with the operand type
3956 /// and the current alternative constraint selected.
3957 TargetLowering::ConstraintWeight
3958 ARM64TargetLowering::getSingleConstraintMatchWeight(
3959 AsmOperandInfo &info, const char *constraint) const {
3960 ConstraintWeight weight = CW_Invalid;
3961 Value *CallOperandVal = info.CallOperandVal;
3962 // If we don't have a value, we can't do a match,
3963 // but allow it at the lowest weight.
3964 if (CallOperandVal == NULL)
3966 Type *type = CallOperandVal->getType();
3967 // Look at the constraint type.
3968 switch (*constraint) {
3970 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3974 if (type->isFloatingPointTy() || type->isVectorTy())
3975 weight = CW_Register;
3978 weight = CW_Constant;
3984 std::pair<unsigned, const TargetRegisterClass *>
3985 ARM64TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3987 if (Constraint.size() == 1) {
3988 switch (Constraint[0]) {
3990 if (VT.getSizeInBits() == 64)
3991 return std::make_pair(0U, &ARM64::GPR64commonRegClass);
3992 return std::make_pair(0U, &ARM64::GPR32commonRegClass);
3995 return std::make_pair(0U, &ARM64::FPR32RegClass);
3996 if (VT.getSizeInBits() == 64)
3997 return std::make_pair(0U, &ARM64::FPR64RegClass);
3998 if (VT.getSizeInBits() == 128)
3999 return std::make_pair(0U, &ARM64::FPR128RegClass);
4001 // The instructions that this constraint is designed for can
4002 // only take 128-bit registers so just use that regclass.
4004 if (VT.getSizeInBits() == 128)
4005 return std::make_pair(0U, &ARM64::FPR128_loRegClass);
4009 if (StringRef("{cc}").equals_lower(Constraint))
4010 return std::make_pair(unsigned(ARM64::CPSR), &ARM64::CCRRegClass);
4012 // Use the default implementation in TargetLowering to convert the register
4013 // constraint into a member of a register class.
4014 std::pair<unsigned, const TargetRegisterClass *> Res;
4015 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4017 // Not found as a standard register?
4018 if (Res.second == 0) {
4019 unsigned Size = Constraint.size();
4020 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
4021 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
4022 const std::string Reg =
4023 std::string(&Constraint[2], &Constraint[Size - 1]);
4024 int RegNo = atoi(Reg.c_str());
4025 if (RegNo >= 0 && RegNo <= 31) {
4026 // v0 - v31 are aliases of q0 - q31.
4027 // By default we'll emit v0-v31 for this unless there's a modifier where
4028 // we'll emit the correct register as well.
4029 Res.first = ARM64::FPR128RegClass.getRegister(RegNo);
4030 Res.second = &ARM64::FPR128RegClass;
4038 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4039 /// vector. If it is invalid, don't add anything to Ops.
4040 void ARM64TargetLowering::LowerAsmOperandForConstraint(
4041 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
4042 SelectionDAG &DAG) const {
4043 SDValue Result(0, 0);
4045 // Currently only support length 1 constraints.
4046 if (Constraint.length() != 1)
4049 char ConstraintLetter = Constraint[0];
4050 switch (ConstraintLetter) {
4054 // This set of constraints deal with valid constants for various instructions.
4055 // Validate and return a target constant for them if we can.
4057 // 'z' maps to xzr or wzr so it needs an input of 0.
4058 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4059 if (!C || C->getZExtValue() != 0)
4062 if (Op.getValueType() == MVT::i64)
4063 Result = DAG.getRegister(ARM64::XZR, MVT::i64);
4065 Result = DAG.getRegister(ARM64::WZR, MVT::i32);
4075 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4079 // Grab the value and do some validation.
4080 uint64_t CVal = C->getZExtValue();
4081 switch (ConstraintLetter) {
4082 // The I constraint applies only to simple ADD or SUB immediate operands:
4083 // i.e. 0 to 4095 with optional shift by 12
4084 // The J constraint applies only to ADD or SUB immediates that would be
4085 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
4086 // instruction [or vice versa], in other words -1 to -4095 with optional
4087 // left shift by 12.
4089 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
4093 uint64_t NVal = -C->getSExtValue();
4094 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal))
4098 // The K and L constraints apply *only* to logical immediates, including
4099 // what used to be the MOVI alias for ORR (though the MOVI alias has now
4100 // been removed and MOV should be used). So these constraints have to
4101 // distinguish between bit patterns that are valid 32-bit or 64-bit
4102 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
4103 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
4106 if (ARM64_AM::isLogicalImmediate(CVal, 32))
4110 if (ARM64_AM::isLogicalImmediate(CVal, 64))
4113 // The M and N constraints are a superset of K and L respectively, for use
4114 // with the MOV (immediate) alias. As well as the logical immediates they
4115 // also match 32 or 64-bit immediates that can be loaded either using a
4116 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
4117 // (M) or 64-bit 0x1234000000000000 (N) etc.
4118 // As a note some of this code is liberally stolen from the asm parser.
4120 if (!isUInt<32>(CVal))
4122 if (ARM64_AM::isLogicalImmediate(CVal, 32))
4124 if ((CVal & 0xFFFF) == CVal)
4126 if ((CVal & 0xFFFF0000ULL) == CVal)
4128 uint64_t NCVal = ~(uint32_t)CVal;
4129 if ((NCVal & 0xFFFFULL) == NCVal)
4131 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4136 if (ARM64_AM::isLogicalImmediate(CVal, 64))
4138 if ((CVal & 0xFFFFULL) == CVal)
4140 if ((CVal & 0xFFFF0000ULL) == CVal)
4142 if ((CVal & 0xFFFF00000000ULL) == CVal)
4144 if ((CVal & 0xFFFF000000000000ULL) == CVal)
4146 uint64_t NCVal = ~CVal;
4147 if ((NCVal & 0xFFFFULL) == NCVal)
4149 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4151 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
4153 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
4161 // All assembler immediates are 64-bit integers.
4162 Result = DAG.getTargetConstant(CVal, MVT::i64);
4166 if (Result.getNode()) {
4167 Ops.push_back(Result);
4171 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4174 //===----------------------------------------------------------------------===//
4175 // ARM64 Advanced SIMD Support
4176 //===----------------------------------------------------------------------===//
4178 /// WidenVector - Given a value in the V64 register class, produce the
4179 /// equivalent value in the V128 register class.
4180 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
4181 EVT VT = V64Reg.getValueType();
4182 unsigned NarrowSize = VT.getVectorNumElements();
4183 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4184 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
4187 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
4188 V64Reg, DAG.getConstant(0, MVT::i32));
4191 /// getExtFactor - Determine the adjustment factor for the position when
4192 /// generating an "extract from vector registers" instruction.
4193 static unsigned getExtFactor(SDValue &V) {
4194 EVT EltType = V.getValueType().getVectorElementType();
4195 return EltType.getSizeInBits() / 8;
4198 /// NarrowVector - Given a value in the V128 register class, produce the
4199 /// equivalent value in the V64 register class.
4200 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
4201 EVT VT = V128Reg.getValueType();
4202 unsigned WideSize = VT.getVectorNumElements();
4203 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4204 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
4207 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, NarrowTy, V128Reg);
4210 // Gather data to see if the operation can be modelled as a
4211 // shuffle in combination with VEXTs.
4212 SDValue ARM64TargetLowering::ReconstructShuffle(SDValue Op,
4213 SelectionDAG &DAG) const {
4215 EVT VT = Op.getValueType();
4216 unsigned NumElts = VT.getVectorNumElements();
4218 SmallVector<SDValue, 2> SourceVecs;
4219 SmallVector<unsigned, 2> MinElts;
4220 SmallVector<unsigned, 2> MaxElts;
4222 for (unsigned i = 0; i < NumElts; ++i) {
4223 SDValue V = Op.getOperand(i);
4224 if (V.getOpcode() == ISD::UNDEF)
4226 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4227 // A shuffle can only come from building a vector from various
4228 // elements of other vectors.
4232 // Record this extraction against the appropriate vector if possible...
4233 SDValue SourceVec = V.getOperand(0);
4234 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4235 bool FoundSource = false;
4236 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4237 if (SourceVecs[j] == SourceVec) {
4238 if (MinElts[j] > EltNo)
4240 if (MaxElts[j] < EltNo)
4247 // Or record a new source if not...
4249 SourceVecs.push_back(SourceVec);
4250 MinElts.push_back(EltNo);
4251 MaxElts.push_back(EltNo);
4255 // Currently only do something sane when at most two source vectors
4257 if (SourceVecs.size() > 2)
4260 SDValue ShuffleSrcs[2] = { DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4261 int VEXTOffsets[2] = { 0, 0 };
4263 // This loop extracts the usage patterns of the source vectors
4264 // and prepares appropriate SDValues for a shuffle if possible.
4265 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4266 if (SourceVecs[i].getValueType() == VT) {
4267 // No VEXT necessary
4268 ShuffleSrcs[i] = SourceVecs[i];
4271 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4272 // It probably isn't worth padding out a smaller vector just to
4273 // break it down again in a shuffle.
4277 // Don't attempt to extract subvectors from BUILD_VECTOR sources
4278 // that expand or trunc the original value.
4279 // TODO: We can try to bitcast and ANY_EXTEND the result but
4280 // we need to consider the cost of vector ANY_EXTEND, and the
4281 // legality of all the types.
4282 if (SourceVecs[i].getValueType().getVectorElementType() !=
4283 VT.getVectorElementType())
4286 // Since only 64-bit and 128-bit vectors are legal on ARM and
4287 // we've eliminated the other cases...
4288 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2 * NumElts &&
4289 "unexpected vector sizes in ReconstructShuffle");
4291 if (MaxElts[i] - MinElts[i] >= NumElts) {
4292 // Span too large for a VEXT to cope
4296 if (MinElts[i] >= NumElts) {
4297 // The extraction can just take the second half
4298 VEXTOffsets[i] = NumElts;
4300 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
4301 DAG.getIntPtrConstant(NumElts));
4302 } else if (MaxElts[i] < NumElts) {
4303 // The extraction can just take the first half
4305 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4306 SourceVecs[i], DAG.getIntPtrConstant(0));
4308 // An actual VEXT is needed
4309 VEXTOffsets[i] = MinElts[i];
4310 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4311 SourceVecs[i], DAG.getIntPtrConstant(0));
4313 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
4314 DAG.getIntPtrConstant(NumElts));
4315 unsigned Imm = VEXTOffsets[i] * getExtFactor(VEXTSrc1);
4316 ShuffleSrcs[i] = DAG.getNode(ARM64ISD::EXT, dl, VT, VEXTSrc1, VEXTSrc2,
4317 DAG.getConstant(Imm, MVT::i32));
4321 SmallVector<int, 8> Mask;
4323 for (unsigned i = 0; i < NumElts; ++i) {
4324 SDValue Entry = Op.getOperand(i);
4325 if (Entry.getOpcode() == ISD::UNDEF) {
4330 SDValue ExtractVec = Entry.getOperand(0);
4332 cast<ConstantSDNode>(Op.getOperand(i).getOperand(1))->getSExtValue();
4333 if (ExtractVec == SourceVecs[0]) {
4334 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4336 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4340 // Final check before we try to produce nonsense...
4341 if (isShuffleMaskLegal(Mask, VT))
4342 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4348 // check if an EXT instruction can handle the shuffle mask when the
4349 // vector sources of the shuffle are the same.
4350 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4351 unsigned NumElts = VT.getVectorNumElements();
4353 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4359 // If this is a VEXT shuffle, the immediate value is the index of the first
4360 // element. The other shuffle indices must be the successive elements after
4362 unsigned ExpectedElt = Imm;
4363 for (unsigned i = 1; i < NumElts; ++i) {
4364 // Increment the expected index. If it wraps around, just follow it
4365 // back to index zero and keep going.
4367 if (ExpectedElt == NumElts)
4371 continue; // ignore UNDEF indices
4372 if (ExpectedElt != static_cast<unsigned>(M[i]))
4379 // check if an EXT instruction can handle the shuffle mask when the
4380 // vector sources of the shuffle are different.
4381 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
4383 unsigned NumElts = VT.getVectorNumElements();
4386 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4392 // If this is a VEXT shuffle, the immediate value is the index of the first
4393 // element. The other shuffle indices must be the successive elements after
4395 unsigned ExpectedElt = Imm;
4396 for (unsigned i = 1; i < NumElts; ++i) {
4397 // Increment the expected index. If it wraps around, it may still be
4398 // a VEXT but the source vectors must be swapped.
4400 if (ExpectedElt == NumElts * 2) {
4406 continue; // ignore UNDEF indices
4407 if (ExpectedElt != static_cast<unsigned>(M[i]))
4411 // Adjust the index value if the source operands will be swapped.
4418 /// isREVMask - Check if a vector shuffle corresponds to a REV
4419 /// instruction with the specified blocksize. (The order of the elements
4420 /// within each block of the vector is reversed.)
4421 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4422 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
4423 "Only possible block sizes for REV are: 16, 32, 64");
4425 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4429 unsigned NumElts = VT.getVectorNumElements();
4430 unsigned BlockElts = M[0] + 1;
4431 // If the first shuffle index is UNDEF, be optimistic.
4433 BlockElts = BlockSize / EltSz;
4435 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4438 for (unsigned i = 0; i < NumElts; ++i) {
4440 continue; // ignore UNDEF indices
4441 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
4448 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4449 unsigned NumElts = VT.getVectorNumElements();
4450 WhichResult = (M[0] == 0 ? 0 : 1);
4451 unsigned Idx = WhichResult * NumElts / 2;
4452 for (unsigned i = 0; i != NumElts; i += 2) {
4453 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4454 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
4462 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4463 unsigned NumElts = VT.getVectorNumElements();
4464 WhichResult = (M[0] == 0 ? 0 : 1);
4465 for (unsigned i = 0; i != NumElts; ++i) {
4467 continue; // ignore UNDEF indices
4468 if ((unsigned)M[i] != 2 * i + WhichResult)
4475 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4476 unsigned NumElts = VT.getVectorNumElements();
4477 WhichResult = (M[0] == 0 ? 0 : 1);
4478 for (unsigned i = 0; i < NumElts; i += 2) {
4479 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4480 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
4486 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
4487 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4488 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4489 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4490 unsigned NumElts = VT.getVectorNumElements();
4491 WhichResult = (M[0] == 0 ? 0 : 1);
4492 unsigned Idx = WhichResult * NumElts / 2;
4493 for (unsigned i = 0; i != NumElts; i += 2) {
4494 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4495 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
4503 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
4504 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4505 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4506 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4507 unsigned Half = VT.getVectorNumElements() / 2;
4508 WhichResult = (M[0] == 0 ? 0 : 1);
4509 for (unsigned j = 0; j != 2; ++j) {
4510 unsigned Idx = WhichResult;
4511 for (unsigned i = 0; i != Half; ++i) {
4512 int MIdx = M[i + j * Half];
4513 if (MIdx >= 0 && (unsigned)MIdx != Idx)
4522 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
4523 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4524 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4525 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4526 unsigned NumElts = VT.getVectorNumElements();
4527 WhichResult = (M[0] == 0 ? 0 : 1);
4528 for (unsigned i = 0; i < NumElts; i += 2) {
4529 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4530 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
4536 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4537 /// the specified operations to build the shuffle.
4538 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4539 SDValue RHS, SelectionDAG &DAG,
4541 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4542 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
4543 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
4546 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4555 OP_VUZPL, // VUZP, left result
4556 OP_VUZPR, // VUZP, right result
4557 OP_VZIPL, // VZIP, left result
4558 OP_VZIPR, // VZIP, right result
4559 OP_VTRNL, // VTRN, left result
4560 OP_VTRNR // VTRN, right result
4563 if (OpNum == OP_COPY) {
4564 if (LHSID == (1 * 9 + 2) * 9 + 3)
4566 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
4570 SDValue OpLHS, OpRHS;
4571 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4572 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4573 EVT VT = OpLHS.getValueType();
4577 llvm_unreachable("Unknown shuffle opcode!");
4579 // VREV divides the vector in half and swaps within the half.
4580 if (VT.getVectorElementType() == MVT::i32 ||
4581 VT.getVectorElementType() == MVT::f32)
4582 return DAG.getNode(ARM64ISD::REV64, dl, VT, OpLHS);
4583 // vrev <4 x i16> -> REV32
4584 if (VT.getVectorElementType() == MVT::i16)
4585 return DAG.getNode(ARM64ISD::REV32, dl, VT, OpLHS);
4586 // vrev <4 x i8> -> REV16
4587 assert(VT.getVectorElementType() == MVT::i8);
4588 return DAG.getNode(ARM64ISD::REV16, dl, VT, OpLHS);
4593 EVT EltTy = VT.getVectorElementType();
4595 if (EltTy == MVT::i8)
4596 Opcode = ARM64ISD::DUPLANE8;
4597 else if (EltTy == MVT::i16)
4598 Opcode = ARM64ISD::DUPLANE16;
4599 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
4600 Opcode = ARM64ISD::DUPLANE32;
4601 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
4602 Opcode = ARM64ISD::DUPLANE64;
4604 llvm_unreachable("Invalid vector element type?");
4606 if (VT.getSizeInBits() == 64)
4607 OpLHS = WidenVector(OpLHS, DAG);
4608 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, MVT::i64);
4609 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
4614 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
4615 return DAG.getNode(ARM64ISD::EXT, dl, VT, OpLHS, OpRHS,
4616 DAG.getConstant(Imm, MVT::i32));
4619 return DAG.getNode(ARM64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4621 return DAG.getNode(ARM64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4623 return DAG.getNode(ARM64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4625 return DAG.getNode(ARM64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4627 return DAG.getNode(ARM64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4629 return DAG.getNode(ARM64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4633 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
4634 SelectionDAG &DAG) {
4635 // Check to see if we can use the TBL instruction.
4636 SDValue V1 = Op.getOperand(0);
4637 SDValue V2 = Op.getOperand(1);
4640 EVT EltVT = Op.getValueType().getVectorElementType();
4641 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
4643 SmallVector<SDValue, 8> TBLMask;
4644 for (int Val : ShuffleMask) {
4645 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
4646 unsigned Offset = Byte + Val * BytesPerElt;
4647 TBLMask.push_back(DAG.getConstant(Offset, MVT::i32));
4651 MVT IndexVT = MVT::v8i8;
4652 unsigned IndexLen = 8;
4653 if (Op.getValueType().getSizeInBits() == 128) {
4654 IndexVT = MVT::v16i8;
4658 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
4659 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
4662 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
4664 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
4665 Shuffle = DAG.getNode(
4666 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4667 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4668 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4670 if (IndexLen == 8) {
4671 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
4672 Shuffle = DAG.getNode(
4673 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4674 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4675 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4677 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
4678 // cannot currently represent the register constraints on the input
4680 // Shuffle = DAG.getNode(ARM64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
4681 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4682 // &TBLMask[0], IndexLen));
4683 Shuffle = DAG.getNode(
4684 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4685 DAG.getConstant(Intrinsic::arm64_neon_tbl2, MVT::i32), V1Cst, V2Cst,
4686 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4689 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
4692 static unsigned getDUPLANEOp(EVT EltType) {
4693 if (EltType == MVT::i8)
4694 return ARM64ISD::DUPLANE8;
4695 if (EltType == MVT::i16)
4696 return ARM64ISD::DUPLANE16;
4697 if (EltType == MVT::i32 || EltType == MVT::f32)
4698 return ARM64ISD::DUPLANE32;
4699 if (EltType == MVT::i64 || EltType == MVT::f64)
4700 return ARM64ISD::DUPLANE64;
4702 llvm_unreachable("Invalid vector element type?");
4705 SDValue ARM64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4706 SelectionDAG &DAG) const {
4708 EVT VT = Op.getValueType();
4710 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4712 // Convert shuffles that are directly supported on NEON to target-specific
4713 // DAG nodes, instead of keeping them as shuffles and matching them again
4714 // during code selection. This is more efficient and avoids the possibility
4715 // of inconsistencies between legalization and selection.
4716 ArrayRef<int> ShuffleMask = SVN->getMask();
4718 SDValue V1 = Op.getOperand(0);
4719 SDValue V2 = Op.getOperand(1);
4721 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
4722 V1.getValueType().getSimpleVT())) {
4723 int Lane = SVN->getSplatIndex();
4724 // If this is undef splat, generate it via "just" vdup, if possible.
4728 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
4729 return DAG.getNode(ARM64ISD::DUP, dl, V1.getValueType(),
4731 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
4732 // constant. If so, we can just reference the lane's definition directly.
4733 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
4734 !isa<ConstantSDNode>(V1.getOperand(Lane)))
4735 return DAG.getNode(ARM64ISD::DUP, dl, VT, V1.getOperand(Lane));
4737 // Otherwise, duplicate from the lane of the input vector.
4738 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
4740 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
4741 // to make a vector of the same size as this SHUFFLE. We can ignore the
4742 // extract entirely, and canonicalise the concat using WidenVector.
4743 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
4744 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
4745 V1 = V1.getOperand(0);
4746 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
4747 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
4748 Lane -= Idx * VT.getVectorNumElements() / 2;
4749 V1 = WidenVector(V1.getOperand(Idx), DAG);
4750 } else if (VT.getSizeInBits() == 64)
4751 V1 = WidenVector(V1, DAG);
4753 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, MVT::i64));
4756 if (isREVMask(ShuffleMask, VT, 64))
4757 return DAG.getNode(ARM64ISD::REV64, dl, V1.getValueType(), V1, V2);
4758 if (isREVMask(ShuffleMask, VT, 32))
4759 return DAG.getNode(ARM64ISD::REV32, dl, V1.getValueType(), V1, V2);
4760 if (isREVMask(ShuffleMask, VT, 16))
4761 return DAG.getNode(ARM64ISD::REV16, dl, V1.getValueType(), V1, V2);
4763 bool ReverseEXT = false;
4765 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
4768 Imm *= getExtFactor(V1);
4769 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V2,
4770 DAG.getConstant(Imm, MVT::i32));
4771 } else if (V2->getOpcode() == ISD::UNDEF &&
4772 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
4773 Imm *= getExtFactor(V1);
4774 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V1,
4775 DAG.getConstant(Imm, MVT::i32));
4778 unsigned WhichResult;
4779 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
4780 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4781 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4783 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
4784 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4785 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4787 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
4788 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4789 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4792 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4793 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4794 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4796 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4797 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4798 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4800 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4801 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4802 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4805 // If the shuffle is not directly supported and it has 4 elements, use
4806 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4807 unsigned NumElts = VT.getVectorNumElements();
4809 unsigned PFIndexes[4];
4810 for (unsigned i = 0; i != 4; ++i) {
4811 if (ShuffleMask[i] < 0)
4814 PFIndexes[i] = ShuffleMask[i];
4817 // Compute the index in the perfect shuffle table.
4818 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
4819 PFIndexes[2] * 9 + PFIndexes[3];
4820 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4821 unsigned Cost = (PFEntry >> 30);
4824 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4827 return GenerateTBL(Op, ShuffleMask, DAG);
4830 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
4832 EVT VT = BVN->getValueType(0);
4833 APInt SplatBits, SplatUndef;
4834 unsigned SplatBitSize;
4836 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4837 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
4839 for (unsigned i = 0; i < NumSplats; ++i) {
4840 CnstBits <<= SplatBitSize;
4841 UndefBits <<= SplatBitSize;
4842 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
4843 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
4852 SDValue ARM64TargetLowering::LowerVectorAND(SDValue Op,
4853 SelectionDAG &DAG) const {
4854 BuildVectorSDNode *BVN =
4855 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
4856 SDValue LHS = Op.getOperand(0);
4858 EVT VT = Op.getValueType();
4863 APInt CnstBits(VT.getSizeInBits(), 0);
4864 APInt UndefBits(VT.getSizeInBits(), 0);
4865 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
4866 // We only have BIC vector immediate instruction, which is and-not.
4867 CnstBits = ~CnstBits;
4869 // We make use of a little bit of goto ickiness in order to avoid having to
4870 // duplicate the immediate matching logic for the undef toggled case.
4871 bool SecondTry = false;
4874 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
4875 CnstBits = CnstBits.zextOrTrunc(64);
4876 uint64_t CnstVal = CnstBits.getZExtValue();
4878 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
4879 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
4880 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4881 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4882 DAG.getConstant(CnstVal, MVT::i32),
4883 DAG.getConstant(0, MVT::i32));
4884 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4887 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
4888 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
4889 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4890 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4891 DAG.getConstant(CnstVal, MVT::i32),
4892 DAG.getConstant(8, MVT::i32));
4893 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4896 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
4897 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
4898 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4899 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4900 DAG.getConstant(CnstVal, MVT::i32),
4901 DAG.getConstant(16, MVT::i32));
4902 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4905 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
4906 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
4907 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4908 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4909 DAG.getConstant(CnstVal, MVT::i32),
4910 DAG.getConstant(24, MVT::i32));
4911 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4914 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
4915 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
4916 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4917 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4918 DAG.getConstant(CnstVal, MVT::i32),
4919 DAG.getConstant(0, MVT::i32));
4920 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4923 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
4924 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
4925 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4926 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4927 DAG.getConstant(CnstVal, MVT::i32),
4928 DAG.getConstant(8, MVT::i32));
4929 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4936 CnstBits = ~UndefBits;
4940 // We can always fall back to a non-immediate AND.
4945 // Specialized code to quickly find if PotentialBVec is a BuildVector that
4946 // consists of only the same constant int value, returned in reference arg
4948 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
4949 uint64_t &ConstVal) {
4950 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
4953 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
4956 EVT VT = Bvec->getValueType(0);
4957 unsigned NumElts = VT.getVectorNumElements();
4958 for (unsigned i = 1; i < NumElts; ++i)
4959 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
4961 ConstVal = FirstElt->getZExtValue();
4965 static unsigned getIntrinsicID(const SDNode *N) {
4966 unsigned Opcode = N->getOpcode();
4969 return Intrinsic::not_intrinsic;
4970 case ISD::INTRINSIC_WO_CHAIN: {
4971 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4972 if (IID < Intrinsic::num_intrinsics)
4974 return Intrinsic::not_intrinsic;
4979 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
4980 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
4981 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
4982 // Also, logical shift right -> sri, with the same structure.
4983 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
4984 EVT VT = N->getValueType(0);
4991 // Is the first op an AND?
4992 const SDValue And = N->getOperand(0);
4993 if (And.getOpcode() != ISD::AND)
4996 // Is the second op an shl or lshr?
4997 SDValue Shift = N->getOperand(1);
4998 // This will have been turned into: ARM64ISD::VSHL vector, #shift
4999 // or ARM64ISD::VLSHR vector, #shift
5000 unsigned ShiftOpc = Shift.getOpcode();
5001 if ((ShiftOpc != ARM64ISD::VSHL && ShiftOpc != ARM64ISD::VLSHR))
5003 bool IsShiftRight = ShiftOpc == ARM64ISD::VLSHR;
5005 // Is the shift amount constant?
5006 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
5010 // Is the and mask vector all constant?
5012 if (!isAllConstantBuildVector(And.getOperand(1), C1))
5015 // Is C1 == ~C2, taking into account how much one can shift elements of a
5017 uint64_t C2 = C2node->getZExtValue();
5018 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
5019 if (C2 > ElemSizeInBits)
5021 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
5022 if ((C1 & ElemMask) != (~C2 & ElemMask))
5025 SDValue X = And.getOperand(0);
5026 SDValue Y = Shift.getOperand(0);
5029 IsShiftRight ? Intrinsic::arm64_neon_vsri : Intrinsic::arm64_neon_vsli;
5031 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5032 DAG.getConstant(Intrin, MVT::i32), X, Y, Shift.getOperand(1));
5034 DEBUG(dbgs() << "arm64-lower: transformed: \n");
5035 DEBUG(N->dump(&DAG));
5036 DEBUG(dbgs() << "into: \n");
5037 DEBUG(ResultSLI->dump(&DAG));
5043 SDValue ARM64TargetLowering::LowerVectorOR(SDValue Op,
5044 SelectionDAG &DAG) const {
5045 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
5046 if (EnableARM64SlrGeneration) {
5047 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
5052 BuildVectorSDNode *BVN =
5053 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
5054 SDValue LHS = Op.getOperand(1);
5056 EVT VT = Op.getValueType();
5058 // OR commutes, so try swapping the operands.
5060 LHS = Op.getOperand(0);
5061 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5066 APInt CnstBits(VT.getSizeInBits(), 0);
5067 APInt UndefBits(VT.getSizeInBits(), 0);
5068 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5069 // We make use of a little bit of goto ickiness in order to avoid having to
5070 // duplicate the immediate matching logic for the undef toggled case.
5071 bool SecondTry = false;
5074 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5075 CnstBits = CnstBits.zextOrTrunc(64);
5076 uint64_t CnstVal = CnstBits.getZExtValue();
5078 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5079 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5080 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5081 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5082 DAG.getConstant(CnstVal, MVT::i32),
5083 DAG.getConstant(0, MVT::i32));
5084 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5087 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5088 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5089 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5090 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5091 DAG.getConstant(CnstVal, MVT::i32),
5092 DAG.getConstant(8, MVT::i32));
5093 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5096 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5097 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5098 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5099 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5100 DAG.getConstant(CnstVal, MVT::i32),
5101 DAG.getConstant(16, MVT::i32));
5102 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5105 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5106 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5107 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5108 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5109 DAG.getConstant(CnstVal, MVT::i32),
5110 DAG.getConstant(24, MVT::i32));
5111 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5114 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5115 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5116 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5117 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5118 DAG.getConstant(CnstVal, MVT::i32),
5119 DAG.getConstant(0, MVT::i32));
5120 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5123 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5124 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5125 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5126 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5127 DAG.getConstant(CnstVal, MVT::i32),
5128 DAG.getConstant(8, MVT::i32));
5129 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5136 CnstBits = UndefBits;
5140 // We can always fall back to a non-immediate OR.
5145 SDValue ARM64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
5146 SelectionDAG &DAG) const {
5147 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5149 EVT VT = Op.getValueType();
5151 APInt CnstBits(VT.getSizeInBits(), 0);
5152 APInt UndefBits(VT.getSizeInBits(), 0);
5153 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5154 // We make use of a little bit of goto ickiness in order to avoid having to
5155 // duplicate the immediate matching logic for the undef toggled case.
5156 bool SecondTry = false;
5159 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5160 CnstBits = CnstBits.zextOrTrunc(64);
5161 uint64_t CnstVal = CnstBits.getZExtValue();
5163 // Certain magic vector constants (used to express things like NOT
5164 // and NEG) are passed through unmodified. This allows codegen patterns
5165 // for these operations to match. Special-purpose patterns will lower
5166 // these immediates to MOVIs if it proves necessary.
5167 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
5170 // The many faces of MOVI...
5171 if (ARM64_AM::isAdvSIMDModImmType10(CnstVal)) {
5172 CnstVal = ARM64_AM::encodeAdvSIMDModImmType10(CnstVal);
5173 if (VT.getSizeInBits() == 128) {
5174 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::v2i64,
5175 DAG.getConstant(CnstVal, MVT::i32));
5176 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5179 // Support the V64 version via subregister insertion.
5180 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::f64,
5181 DAG.getConstant(CnstVal, MVT::i32));
5182 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5185 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5186 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5187 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5188 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5189 DAG.getConstant(CnstVal, MVT::i32),
5190 DAG.getConstant(0, MVT::i32));
5191 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5194 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5195 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5196 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5197 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5198 DAG.getConstant(CnstVal, MVT::i32),
5199 DAG.getConstant(8, MVT::i32));
5200 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5203 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5204 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5205 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5206 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5207 DAG.getConstant(CnstVal, MVT::i32),
5208 DAG.getConstant(16, MVT::i32));
5209 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5212 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5213 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5214 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5215 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5216 DAG.getConstant(CnstVal, MVT::i32),
5217 DAG.getConstant(24, MVT::i32));
5218 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5221 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5222 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5223 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5224 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5225 DAG.getConstant(CnstVal, MVT::i32),
5226 DAG.getConstant(0, MVT::i32));
5227 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5230 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5231 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5232 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5233 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5234 DAG.getConstant(CnstVal, MVT::i32),
5235 DAG.getConstant(8, MVT::i32));
5236 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5239 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
5240 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
5241 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5242 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
5243 DAG.getConstant(CnstVal, MVT::i32),
5244 DAG.getConstant(264, MVT::i32));
5245 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5248 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
5249 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
5250 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5251 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
5252 DAG.getConstant(CnstVal, MVT::i32),
5253 DAG.getConstant(272, MVT::i32));
5254 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5257 if (ARM64_AM::isAdvSIMDModImmType9(CnstVal)) {
5258 CnstVal = ARM64_AM::encodeAdvSIMDModImmType9(CnstVal);
5259 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
5260 SDValue Mov = DAG.getNode(ARM64ISD::MOVI, dl, MovTy,
5261 DAG.getConstant(CnstVal, MVT::i32));
5262 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5265 // The few faces of FMOV...
5266 if (ARM64_AM::isAdvSIMDModImmType11(CnstVal)) {
5267 CnstVal = ARM64_AM::encodeAdvSIMDModImmType11(CnstVal);
5268 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
5269 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MovTy,
5270 DAG.getConstant(CnstVal, MVT::i32));
5271 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5274 if (ARM64_AM::isAdvSIMDModImmType12(CnstVal) &&
5275 VT.getSizeInBits() == 128) {
5276 CnstVal = ARM64_AM::encodeAdvSIMDModImmType12(CnstVal);
5277 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MVT::v2f64,
5278 DAG.getConstant(CnstVal, MVT::i32));
5279 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5282 // The many faces of MVNI...
5284 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5285 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5286 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5287 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5288 DAG.getConstant(CnstVal, MVT::i32),
5289 DAG.getConstant(0, MVT::i32));
5290 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5293 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5294 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5295 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5296 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5297 DAG.getConstant(CnstVal, MVT::i32),
5298 DAG.getConstant(8, MVT::i32));
5299 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5302 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5303 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5304 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5305 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5306 DAG.getConstant(CnstVal, MVT::i32),
5307 DAG.getConstant(16, MVT::i32));
5308 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5311 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5312 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5313 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5314 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5315 DAG.getConstant(CnstVal, MVT::i32),
5316 DAG.getConstant(24, MVT::i32));
5317 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5320 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5321 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5322 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5323 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5324 DAG.getConstant(CnstVal, MVT::i32),
5325 DAG.getConstant(0, MVT::i32));
5326 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5329 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5330 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5331 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5332 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5333 DAG.getConstant(CnstVal, MVT::i32),
5334 DAG.getConstant(8, MVT::i32));
5335 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5338 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
5339 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
5340 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5341 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
5342 DAG.getConstant(CnstVal, MVT::i32),
5343 DAG.getConstant(264, MVT::i32));
5344 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5347 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
5348 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
5349 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5350 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
5351 DAG.getConstant(CnstVal, MVT::i32),
5352 DAG.getConstant(272, MVT::i32));
5353 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5360 CnstBits = UndefBits;
5365 // Scan through the operands to find some interesting properties we can
5367 // 1) If only one value is used, we can use a DUP, or
5368 // 2) if only the low element is not undef, we can just insert that, or
5369 // 3) if only one constant value is used (w/ some non-constant lanes),
5370 // we can splat the constant value into the whole vector then fill
5371 // in the non-constant lanes.
5372 // 4) FIXME: If different constant values are used, but we can intelligently
5373 // select the values we'll be overwriting for the non-constant
5374 // lanes such that we can directly materialize the vector
5375 // some other way (MOVI, e.g.), we can be sneaky.
5376 unsigned NumElts = VT.getVectorNumElements();
5377 bool isOnlyLowElement = true;
5378 bool usesOnlyOneValue = true;
5379 bool usesOnlyOneConstantValue = true;
5380 bool isConstant = true;
5381 unsigned NumConstantLanes = 0;
5383 SDValue ConstantValue;
5384 for (unsigned i = 0; i < NumElts; ++i) {
5385 SDValue V = Op.getOperand(i);
5386 if (V.getOpcode() == ISD::UNDEF)
5389 isOnlyLowElement = false;
5390 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5393 if (isa<ConstantSDNode>(V)) {
5395 if (!ConstantValue.getNode())
5397 else if (ConstantValue != V)
5398 usesOnlyOneConstantValue = false;
5401 if (!Value.getNode())
5403 else if (V != Value)
5404 usesOnlyOneValue = false;
5407 if (!Value.getNode())
5408 return DAG.getUNDEF(VT);
5410 if (isOnlyLowElement)
5411 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5413 // Use DUP for non-constant splats. For f32 constant splats, reduce to
5414 // i32 and try again.
5415 if (usesOnlyOneValue) {
5417 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5418 Value.getValueType() != VT)
5419 return DAG.getNode(ARM64ISD::DUP, dl, VT, Value);
5421 // This is actually a DUPLANExx operation, which keeps everything vectory.
5423 // DUPLANE works on 128-bit vectors, widen it if necessary.
5424 SDValue Lane = Value.getOperand(1);
5425 Value = Value.getOperand(0);
5426 if (Value.getValueType().getSizeInBits() == 64)
5427 Value = WidenVector(Value, DAG);
5429 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
5430 return DAG.getNode(Opcode, dl, VT, Value, Lane);
5433 if (VT.getVectorElementType().isFloatingPoint()) {
5434 SmallVector<SDValue, 8> Ops;
5436 (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
5437 for (unsigned i = 0; i < NumElts; ++i)
5438 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
5439 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
5440 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
5441 Val = LowerBUILD_VECTOR(Val, DAG);
5443 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5447 // If there was only one constant value used and for more than one lane,
5448 // start by splatting that value, then replace the non-constant lanes. This
5449 // is better than the default, which will perform a separate initialization
5451 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
5452 SDValue Val = DAG.getNode(ARM64ISD::DUP, dl, VT, ConstantValue);
5453 // Now insert the non-constant lanes.
5454 for (unsigned i = 0; i < NumElts; ++i) {
5455 SDValue V = Op.getOperand(i);
5456 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5457 if (!isa<ConstantSDNode>(V)) {
5458 // Note that type legalization likely mucked about with the VT of the
5459 // source operand, so we may have to convert it here before inserting.
5460 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
5466 // If all elements are constants and the case above didn't get hit, fall back
5467 // to the default expansion, which will generate a load from the constant
5472 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5474 SDValue shuffle = ReconstructShuffle(Op, DAG);
5475 if (shuffle != SDValue())
5479 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5480 // know the default expansion would otherwise fall back on something even
5481 // worse. For a vector with one or two non-undef values, that's
5482 // scalar_to_vector for the elements followed by a shuffle (provided the
5483 // shuffle is valid for the target) and materialization element by element
5484 // on the stack followed by a load for everything else.
5485 if (!isConstant && !usesOnlyOneValue) {
5486 SDValue Vec = DAG.getUNDEF(VT);
5487 SDValue Op0 = Op.getOperand(0);
5488 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
5490 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
5491 // a) Avoid a RMW dependency on the full vector register, and
5492 // b) Allow the register coalescer to fold away the copy if the
5493 // value is already in an S or D register.
5494 if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
5495 unsigned SubIdx = ElemSize == 32 ? ARM64::ssub : ARM64::dsub;
5497 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
5498 DAG.getTargetConstant(SubIdx, MVT::i32));
5499 Vec = SDValue(N, 0);
5502 for (; i < NumElts; ++i) {
5503 SDValue V = Op.getOperand(i);
5504 if (V.getOpcode() == ISD::UNDEF)
5506 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5507 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5512 // Just use the default expansion. We failed to find a better alternative.
5516 SDValue ARM64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
5517 SelectionDAG &DAG) const {
5518 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
5520 // Check for non-constant lane.
5521 if (!isa<ConstantSDNode>(Op.getOperand(2)))
5524 EVT VT = Op.getOperand(0).getValueType();
5526 // Insertion/extraction are legal for V128 types.
5527 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5528 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5531 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5532 VT != MVT::v1i64 && VT != MVT::v2f32)
5535 // For V64 types, we perform insertion by expanding the value
5536 // to a V128 type and perform the insertion on that.
5538 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5539 EVT WideTy = WideVec.getValueType();
5541 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
5542 Op.getOperand(1), Op.getOperand(2));
5543 // Re-narrow the resultant vector.
5544 return NarrowVector(Node, DAG);
5547 SDValue ARM64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5548 SelectionDAG &DAG) const {
5549 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
5551 // Check for non-constant lane.
5552 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5555 EVT VT = Op.getOperand(0).getValueType();
5557 // Insertion/extraction are legal for V128 types.
5558 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5559 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5562 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5563 VT != MVT::v1i64 && VT != MVT::v2f32)
5566 // For V64 types, we perform extraction by expanding the value
5567 // to a V128 type and perform the extraction on that.
5569 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5570 EVT WideTy = WideVec.getValueType();
5572 EVT ExtrTy = WideTy.getVectorElementType();
5573 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
5576 // For extractions, we just return the result directly.
5577 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
5581 SDValue ARM64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
5582 SelectionDAG &DAG) const {
5583 EVT VT = Op.getOperand(0).getValueType();
5589 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5592 unsigned Val = Cst->getZExtValue();
5594 unsigned Size = Op.getValueType().getSizeInBits();
5598 return DAG.getTargetExtractSubreg(ARM64::bsub, dl, Op.getValueType(),
5601 return DAG.getTargetExtractSubreg(ARM64::hsub, dl, Op.getValueType(),
5604 return DAG.getTargetExtractSubreg(ARM64::ssub, dl, Op.getValueType(),
5607 return DAG.getTargetExtractSubreg(ARM64::dsub, dl, Op.getValueType(),
5610 llvm_unreachable("Unexpected vector type in extract_subvector!");
5613 // If this is extracting the upper 64-bits of a 128-bit vector, we match
5615 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
5621 bool ARM64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5623 if (VT.getVectorNumElements() == 4 &&
5624 (VT.is128BitVector() || VT.is64BitVector())) {
5625 unsigned PFIndexes[4];
5626 for (unsigned i = 0; i != 4; ++i) {
5630 PFIndexes[i] = M[i];
5633 // Compute the index in the perfect shuffle table.
5634 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5635 PFIndexes[2] * 9 + PFIndexes[3];
5636 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5637 unsigned Cost = (PFEntry >> 30);
5644 unsigned Imm, WhichResult;
5646 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
5647 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
5648 isEXTMask(M, VT, ReverseVEXT, Imm) ||
5649 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
5650 isTRNMask(M, VT, WhichResult) || isUZPMask(M, VT, WhichResult) ||
5651 isZIPMask(M, VT, WhichResult) ||
5652 isTRN_v_undef_Mask(M, VT, WhichResult) ||
5653 isUZP_v_undef_Mask(M, VT, WhichResult) ||
5654 isZIP_v_undef_Mask(M, VT, WhichResult));
5657 /// getVShiftImm - Check if this is a valid build_vector for the immediate
5658 /// operand of a vector shift operation, where all the elements of the
5659 /// build_vector must have the same constant integer value.
5660 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
5661 // Ignore bit_converts.
5662 while (Op.getOpcode() == ISD::BITCAST)
5663 Op = Op.getOperand(0);
5664 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
5665 APInt SplatBits, SplatUndef;
5666 unsigned SplatBitSize;
5668 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
5669 HasAnyUndefs, ElementBits) ||
5670 SplatBitSize > ElementBits)
5672 Cnt = SplatBits.getSExtValue();
5676 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
5677 /// operand of a vector shift left operation. That value must be in the range:
5678 /// 0 <= Value < ElementBits for a left shift; or
5679 /// 0 <= Value <= ElementBits for a long left shift.
5680 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
5681 assert(VT.isVector() && "vector shift count is not a vector type");
5682 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5683 if (!getVShiftImm(Op, ElementBits, Cnt))
5685 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
5688 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
5689 /// operand of a vector shift right operation. For a shift opcode, the value
5690 /// is positive, but for an intrinsic the value count must be negative. The
5691 /// absolute value must be in the range:
5692 /// 1 <= |Value| <= ElementBits for a right shift; or
5693 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
5694 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
5696 assert(VT.isVector() && "vector shift count is not a vector type");
5697 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5698 if (!getVShiftImm(Op, ElementBits, Cnt))
5702 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
5705 SDValue ARM64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
5706 SelectionDAG &DAG) const {
5707 EVT VT = Op.getValueType();
5711 if (!Op.getOperand(1).getValueType().isVector())
5713 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5715 switch (Op.getOpcode()) {
5717 llvm_unreachable("unexpected shift opcode");
5720 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
5721 return DAG.getNode(ARM64ISD::VSHL, SDLoc(Op), VT, Op.getOperand(0),
5722 DAG.getConstant(Cnt, MVT::i32));
5723 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5724 DAG.getConstant(Intrinsic::arm64_neon_ushl, MVT::i32),
5725 Op.getOperand(0), Op.getOperand(1));
5728 // Right shift immediate
5729 if (isVShiftRImm(Op.getOperand(1), VT, false, false, Cnt) &&
5732 (Op.getOpcode() == ISD::SRA) ? ARM64ISD::VASHR : ARM64ISD::VLSHR;
5733 return DAG.getNode(Opc, SDLoc(Op), VT, Op.getOperand(0),
5734 DAG.getConstant(Cnt, MVT::i32));
5737 // Right shift register. Note, there is not a shift right register
5738 // instruction, but the shift left register instruction takes a signed
5739 // value, where negative numbers specify a right shift.
5740 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::arm64_neon_sshl
5741 : Intrinsic::arm64_neon_ushl;
5742 // negate the shift amount
5743 SDValue NegShift = DAG.getNode(ARM64ISD::NEG, DL, VT, Op.getOperand(1));
5744 SDValue NegShiftLeft =
5745 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5746 DAG.getConstant(Opc, MVT::i32), Op.getOperand(0), NegShift);
5747 return NegShiftLeft;
5753 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
5754 ARM64CC::CondCode CC, bool NoNans, EVT VT,
5755 SDLoc dl, SelectionDAG &DAG) {
5756 EVT SrcVT = LHS.getValueType();
5758 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
5759 APInt CnstBits(VT.getSizeInBits(), 0);
5760 APInt UndefBits(VT.getSizeInBits(), 0);
5761 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
5762 bool IsZero = IsCnst && (CnstBits == 0);
5764 if (SrcVT.getVectorElementType().isFloatingPoint()) {
5771 Fcmeq = DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5773 Fcmeq = DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5774 return DAG.getNode(ARM64ISD::NOT, dl, VT, Fcmeq);
5778 return DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5779 return DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5782 return DAG.getNode(ARM64ISD::FCMGEz, dl, VT, LHS);
5783 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, LHS, RHS);
5786 return DAG.getNode(ARM64ISD::FCMGTz, dl, VT, LHS);
5787 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, LHS, RHS);
5790 return DAG.getNode(ARM64ISD::FCMLEz, dl, VT, LHS);
5791 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, RHS, LHS);
5795 // If we ignore NaNs then we can use to the MI implementation.
5799 return DAG.getNode(ARM64ISD::FCMLTz, dl, VT, LHS);
5800 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, RHS, LHS);
5810 Cmeq = DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5812 Cmeq = DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5813 return DAG.getNode(ARM64ISD::NOT, dl, VT, Cmeq);
5817 return DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5818 return DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5821 return DAG.getNode(ARM64ISD::CMGEz, dl, VT, LHS);
5822 return DAG.getNode(ARM64ISD::CMGE, dl, VT, LHS, RHS);
5825 return DAG.getNode(ARM64ISD::CMGTz, dl, VT, LHS);
5826 return DAG.getNode(ARM64ISD::CMGT, dl, VT, LHS, RHS);
5829 return DAG.getNode(ARM64ISD::CMLEz, dl, VT, LHS);
5830 return DAG.getNode(ARM64ISD::CMGE, dl, VT, RHS, LHS);
5832 return DAG.getNode(ARM64ISD::CMHS, dl, VT, RHS, LHS);
5834 return DAG.getNode(ARM64ISD::CMHI, dl, VT, RHS, LHS);
5837 return DAG.getNode(ARM64ISD::CMLTz, dl, VT, LHS);
5838 return DAG.getNode(ARM64ISD::CMGT, dl, VT, RHS, LHS);
5840 return DAG.getNode(ARM64ISD::CMHI, dl, VT, LHS, RHS);
5842 return DAG.getNode(ARM64ISD::CMHS, dl, VT, LHS, RHS);
5846 SDValue ARM64TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
5847 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5848 SDValue LHS = Op.getOperand(0);
5849 SDValue RHS = Op.getOperand(1);
5852 if (LHS.getValueType().getVectorElementType().isInteger()) {
5853 assert(LHS.getValueType() == RHS.getValueType());
5854 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
5855 return EmitVectorComparison(LHS, RHS, ARM64CC, false, Op.getValueType(), dl,
5859 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
5860 LHS.getValueType().getVectorElementType() == MVT::f64);
5862 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
5863 // clean. Some of them require two branches to implement.
5864 ARM64CC::CondCode CC1, CC2;
5865 changeFPCCToARM64CC(CC, CC1, CC2);
5867 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
5869 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, Op.getValueType(), dl, DAG);
5870 if (!Cmp1.getNode())
5873 if (CC2 != ARM64CC::AL) {
5875 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, Op.getValueType(), dl, DAG);
5876 if (!Cmp2.getNode())
5879 return DAG.getNode(ISD::OR, dl, Cmp1.getValueType(), Cmp1, Cmp2);
5885 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
5886 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
5887 /// specified in the intrinsic calls.
5888 bool ARM64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5890 unsigned Intrinsic) const {
5891 switch (Intrinsic) {
5892 case Intrinsic::arm64_neon_ld2:
5893 case Intrinsic::arm64_neon_ld3:
5894 case Intrinsic::arm64_neon_ld4:
5895 case Intrinsic::arm64_neon_ld2lane:
5896 case Intrinsic::arm64_neon_ld3lane:
5897 case Intrinsic::arm64_neon_ld4lane:
5898 case Intrinsic::arm64_neon_ld2r:
5899 case Intrinsic::arm64_neon_ld3r:
5900 case Intrinsic::arm64_neon_ld4r: {
5901 Info.opc = ISD::INTRINSIC_W_CHAIN;
5902 // Conservatively set memVT to the entire set of vectors loaded.
5903 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
5904 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5905 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5908 Info.vol = false; // volatile loads with NEON intrinsics not supported
5909 Info.readMem = true;
5910 Info.writeMem = false;
5913 case Intrinsic::arm64_neon_st2:
5914 case Intrinsic::arm64_neon_st3:
5915 case Intrinsic::arm64_neon_st4:
5916 case Intrinsic::arm64_neon_st2lane:
5917 case Intrinsic::arm64_neon_st3lane:
5918 case Intrinsic::arm64_neon_st4lane: {
5919 Info.opc = ISD::INTRINSIC_VOID;
5920 // Conservatively set memVT to the entire set of vectors stored.
5921 unsigned NumElts = 0;
5922 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5923 Type *ArgTy = I.getArgOperand(ArgI)->getType();
5924 if (!ArgTy->isVectorTy())
5926 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
5928 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5929 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5932 Info.vol = false; // volatile stores with NEON intrinsics not supported
5933 Info.readMem = false;
5934 Info.writeMem = true;
5937 case Intrinsic::arm64_ldxr: {
5938 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
5939 Info.opc = ISD::INTRINSIC_W_CHAIN;
5940 Info.memVT = MVT::getVT(PtrTy->getElementType());
5941 Info.ptrVal = I.getArgOperand(0);
5943 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
5945 Info.readMem = true;
5946 Info.writeMem = false;
5949 case Intrinsic::arm64_stxr: {
5950 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
5951 Info.opc = ISD::INTRINSIC_W_CHAIN;
5952 Info.memVT = MVT::getVT(PtrTy->getElementType());
5953 Info.ptrVal = I.getArgOperand(1);
5955 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
5957 Info.readMem = false;
5958 Info.writeMem = true;
5961 case Intrinsic::arm64_ldxp: {
5962 Info.opc = ISD::INTRINSIC_W_CHAIN;
5963 Info.memVT = MVT::i128;
5964 Info.ptrVal = I.getArgOperand(0);
5968 Info.readMem = true;
5969 Info.writeMem = false;
5972 case Intrinsic::arm64_stxp: {
5973 Info.opc = ISD::INTRINSIC_W_CHAIN;
5974 Info.memVT = MVT::i128;
5975 Info.ptrVal = I.getArgOperand(2);
5979 Info.readMem = false;
5980 Info.writeMem = true;
5990 // Truncations from 64-bit GPR to 32-bit GPR is free.
5991 bool ARM64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
5992 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
5994 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5995 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5996 if (NumBits1 <= NumBits2)
6000 bool ARM64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
6001 if (!VT1.isInteger() || !VT2.isInteger())
6003 unsigned NumBits1 = VT1.getSizeInBits();
6004 unsigned NumBits2 = VT2.getSizeInBits();
6005 if (NumBits1 <= NumBits2)
6010 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
6012 bool ARM64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
6013 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6015 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6016 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6017 if (NumBits1 == 32 && NumBits2 == 64)
6021 bool ARM64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
6022 if (!VT1.isInteger() || !VT2.isInteger())
6024 unsigned NumBits1 = VT1.getSizeInBits();
6025 unsigned NumBits2 = VT2.getSizeInBits();
6026 if (NumBits1 == 32 && NumBits2 == 64)
6031 bool ARM64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
6032 EVT VT1 = Val.getValueType();
6033 if (isZExtFree(VT1, VT2)) {
6037 if (Val.getOpcode() != ISD::LOAD)
6040 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
6041 return (VT1.isSimple() && VT1.isInteger() && VT2.isSimple() &&
6042 VT2.isInteger() && VT1.getSizeInBits() <= 32);
6045 bool ARM64TargetLowering::hasPairedLoad(Type *LoadedType,
6046 unsigned &RequiredAligment) const {
6047 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
6049 // Cyclone supports unaligned accesses.
6050 RequiredAligment = 0;
6051 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
6052 return NumBits == 32 || NumBits == 64;
6055 bool ARM64TargetLowering::hasPairedLoad(EVT LoadedType,
6056 unsigned &RequiredAligment) const {
6057 if (!LoadedType.isSimple() ||
6058 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
6060 // Cyclone supports unaligned accesses.
6061 RequiredAligment = 0;
6062 unsigned NumBits = LoadedType.getSizeInBits();
6063 return NumBits == 32 || NumBits == 64;
6066 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
6067 unsigned AlignCheck) {
6068 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
6069 (DstAlign == 0 || DstAlign % AlignCheck == 0));
6072 EVT ARM64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
6073 unsigned SrcAlign, bool IsMemset,
6074 bool ZeroMemset, bool MemcpyStrSrc,
6075 MachineFunction &MF) const {
6076 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
6077 // instruction to materialize the v2i64 zero and one store (with restrictive
6078 // addressing mode). Just do two i64 store of zero-registers.
6080 const Function *F = MF.getFunction();
6081 if (!IsMemset && Size >= 16 &&
6082 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
6083 Attribute::NoImplicitFloat) &&
6084 (memOpAlign(SrcAlign, DstAlign, 16) ||
6085 (allowsUnalignedMemoryAccesses(MVT::v2i64, 0, &Fast) && Fast)))
6088 return Size >= 8 ? MVT::i64 : MVT::i32;
6091 // 12-bit optionally shifted immediates are legal for adds.
6092 bool ARM64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
6093 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
6098 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
6099 // immediates is the same as for an add or a sub.
6100 bool ARM64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
6103 return isLegalAddImmediate(Immed);
6106 /// isLegalAddressingMode - Return true if the addressing mode represented
6107 /// by AM is legal for this target, for a load/store of the specified type.
6108 bool ARM64TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6110 // ARM64 has five basic addressing modes:
6112 // reg + 9-bit signed offset
6113 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
6115 // reg + SIZE_IN_BYTES * reg
6117 // No global is ever allowed as a base.
6121 // No reg+reg+imm addressing.
6122 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
6125 // check reg + imm case:
6126 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
6127 uint64_t NumBytes = 0;
6128 if (Ty->isSized()) {
6129 uint64_t NumBits = getDataLayout()->getTypeSizeInBits(Ty);
6130 NumBytes = NumBits / 8;
6131 if (!isPowerOf2_64(NumBits))
6136 int64_t Offset = AM.BaseOffs;
6138 // 9-bit signed offset
6139 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
6142 // 12-bit unsigned offset
6143 unsigned shift = Log2_64(NumBytes);
6144 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
6145 // Must be a multiple of NumBytes (NumBytes is a power of 2)
6146 (Offset >> shift) << shift == Offset)
6151 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
6153 if (!AM.Scale || AM.Scale == 1 ||
6154 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
6159 int ARM64TargetLowering::getScalingFactorCost(const AddrMode &AM,
6161 // Scaling factors are not free at all.
6162 // Operands | Rt Latency
6163 // -------------------------------------------
6165 // -------------------------------------------
6166 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
6167 // Rt, [Xn, Wm, <extend> #imm] |
6168 if (isLegalAddressingMode(AM, Ty))
6169 // Scale represents reg2 * scale, thus account for 1 if
6170 // it is not equal to 0 or 1.
6171 return AM.Scale != 0 && AM.Scale != 1;
6175 bool ARM64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
6176 VT = VT.getScalarType();
6181 switch (VT.getSimpleVT().SimpleTy) {
6193 ARM64TargetLowering::getScratchRegisters(CallingConv::ID) const {
6194 // LR is a callee-save register, but we must treat it as clobbered by any call
6195 // site. Hence we include LR in the scratch registers, which are in turn added
6196 // as implicit-defs for stackmaps and patchpoints.
6197 static const MCPhysReg ScratchRegs[] = {
6198 ARM64::X16, ARM64::X17, ARM64::LR, 0
6203 bool ARM64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
6205 assert(Ty->isIntegerTy());
6207 unsigned BitSize = Ty->getPrimitiveSizeInBits();
6211 int64_t Val = Imm.getSExtValue();
6212 if (Val == 0 || ARM64_AM::isLogicalImmediate(Val, BitSize))
6215 if ((int64_t)Val < 0)
6218 Val &= (1LL << 32) - 1;
6220 unsigned LZ = countLeadingZeros((uint64_t)Val);
6221 unsigned Shift = (63 - LZ) / 16;
6222 // MOVZ is free so return true for one or fewer MOVK.
6223 return (Shift < 3) ? true : false;
6226 // Generate SUBS and CSEL for integer abs.
6227 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
6228 EVT VT = N->getValueType(0);
6230 SDValue N0 = N->getOperand(0);
6231 SDValue N1 = N->getOperand(1);
6234 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
6235 // and change it to SUB and CSEL.
6236 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
6237 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
6238 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
6239 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
6240 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
6241 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
6243 // Generate SUBS & CSEL.
6245 DAG.getNode(ARM64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
6246 N0.getOperand(0), DAG.getConstant(0, VT));
6247 return DAG.getNode(ARM64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
6248 DAG.getConstant(ARM64CC::PL, MVT::i32),
6249 SDValue(Cmp.getNode(), 1));
6254 // performXorCombine - Attempts to handle integer ABS.
6255 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
6256 TargetLowering::DAGCombinerInfo &DCI,
6257 const ARM64Subtarget *Subtarget) {
6258 if (DCI.isBeforeLegalizeOps())
6261 return performIntegerAbsCombine(N, DAG);
6264 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
6265 TargetLowering::DAGCombinerInfo &DCI,
6266 const ARM64Subtarget *Subtarget) {
6267 if (DCI.isBeforeLegalizeOps())
6270 // Multiplication of a power of two plus/minus one can be done more
6271 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
6272 // future CPUs have a cheaper MADD instruction, this may need to be
6273 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
6274 // 64-bit is 5 cycles, so this is always a win.
6275 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
6276 APInt Value = C->getAPIntValue();
6277 EVT VT = N->getValueType(0);
6278 APInt VP1 = Value + 1;
6279 if (VP1.isPowerOf2()) {
6280 // Multiplying by one less than a power of two, replace with a shift
6282 SDValue ShiftedVal =
6283 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6284 DAG.getConstant(VP1.logBase2(), MVT::i64));
6285 return DAG.getNode(ISD::SUB, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6287 APInt VM1 = Value - 1;
6288 if (VM1.isPowerOf2()) {
6289 // Multiplying by one more than a power of two, replace with a shift
6291 SDValue ShiftedVal =
6292 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6293 DAG.getConstant(VM1.logBase2(), MVT::i64));
6294 return DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6300 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG) {
6301 EVT VT = N->getValueType(0);
6302 if (VT != MVT::f32 && VT != MVT::f64)
6304 // Only optimize when the source and destination types have the same width.
6305 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
6308 // If the result of an integer load is only used by an integer-to-float
6309 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
6310 // This eliminates an "integer-to-vector-move UOP and improve throughput.
6311 SDValue N0 = N->getOperand(0);
6312 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6313 // Do not change the width of a volatile load.
6314 !cast<LoadSDNode>(N0)->isVolatile()) {
6315 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6316 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
6317 LN0->getPointerInfo(), LN0->isVolatile(),
6318 LN0->isNonTemporal(), LN0->isInvariant(),
6319 LN0->getAlignment());
6321 // Make sure successors of the original load stay after it by updating them
6322 // to use the new Chain.
6323 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
6326 (N->getOpcode() == ISD::SINT_TO_FP) ? ARM64ISD::SITOF : ARM64ISD::UITOF;
6327 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
6333 /// An EXTR instruction is made up of two shifts, ORed together. This helper
6334 /// searches for and classifies those shifts.
6335 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
6337 if (N.getOpcode() == ISD::SHL)
6339 else if (N.getOpcode() == ISD::SRL)
6344 if (!isa<ConstantSDNode>(N.getOperand(1)))
6347 ShiftAmount = N->getConstantOperandVal(1);
6348 Src = N->getOperand(0);
6352 /// EXTR instruction extracts a contiguous chunk of bits from two existing
6353 /// registers viewed as a high/low pair. This function looks for the pattern:
6354 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
6355 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
6357 static SDValue tryCombineToEXTR(SDNode *N,
6358 TargetLowering::DAGCombinerInfo &DCI) {
6359 SelectionDAG &DAG = DCI.DAG;
6361 EVT VT = N->getValueType(0);
6363 assert(N->getOpcode() == ISD::OR && "Unexpected root");
6365 if (VT != MVT::i32 && VT != MVT::i64)
6369 uint32_t ShiftLHS = 0;
6371 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
6375 uint32_t ShiftRHS = 0;
6377 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
6380 // If they're both trying to come from the high part of the register, they're
6381 // not really an EXTR.
6382 if (LHSFromHi == RHSFromHi)
6385 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
6389 std::swap(LHS, RHS);
6390 std::swap(ShiftLHS, ShiftRHS);
6393 return DAG.getNode(ARM64ISD::EXTR, DL, VT, LHS, RHS,
6394 DAG.getConstant(ShiftRHS, MVT::i64));
6397 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
6398 const ARM64Subtarget *Subtarget) {
6399 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
6400 if (!EnableARM64ExtrGeneration)
6402 SelectionDAG &DAG = DCI.DAG;
6403 EVT VT = N->getValueType(0);
6405 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6408 SDValue Res = tryCombineToEXTR(N, DCI);
6415 static SDValue performBitcastCombine(SDNode *N,
6416 TargetLowering::DAGCombinerInfo &DCI,
6417 SelectionDAG &DAG) {
6418 // Wait 'til after everything is legalized to try this. That way we have
6419 // legal vector types and such.
6420 if (DCI.isBeforeLegalizeOps())
6423 // Remove extraneous bitcasts around an extract_subvector.
6425 // (v4i16 (bitconvert
6426 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
6428 // (extract_subvector ((v8i16 ...), (i64 4)))
6430 // Only interested in 64-bit vectors as the ultimate result.
6431 EVT VT = N->getValueType(0);
6434 if (VT.getSimpleVT().getSizeInBits() != 64)
6436 // Is the operand an extract_subvector starting at the beginning or halfway
6437 // point of the vector? A low half may also come through as an
6438 // EXTRACT_SUBREG, so look for that, too.
6439 SDValue Op0 = N->getOperand(0);
6440 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
6441 !(Op0->isMachineOpcode() &&
6442 Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG))
6444 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
6445 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
6446 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
6448 } else if (Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG) {
6449 if (idx != ARM64::dsub)
6451 // The dsub reference is equivalent to a lane zero subvector reference.
6454 // Look through the bitcast of the input to the extract.
6455 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
6457 SDValue Source = Op0->getOperand(0)->getOperand(0);
6458 // If the source type has twice the number of elements as our destination
6459 // type, we know this is an extract of the high or low half of the vector.
6460 EVT SVT = Source->getValueType(0);
6461 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
6464 DEBUG(dbgs() << "arm64-lower: bitcast extract_subvector simplification\n");
6466 // Create the simplified form to just extract the low or high half of the
6467 // vector directly rather than bothering with the bitcasts.
6469 unsigned NumElements = VT.getVectorNumElements();
6471 SDValue HalfIdx = DAG.getConstant(NumElements, MVT::i64);
6472 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
6474 SDValue SubReg = DAG.getTargetConstant(ARM64::dsub, MVT::i32);
6475 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
6481 static SDValue performConcatVectorsCombine(SDNode *N,
6482 TargetLowering::DAGCombinerInfo &DCI,
6483 SelectionDAG &DAG) {
6484 // Wait 'til after everything is legalized to try this. That way we have
6485 // legal vector types and such.
6486 if (DCI.isBeforeLegalizeOps())
6490 EVT VT = N->getValueType(0);
6492 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
6493 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
6494 // canonicalise to that.
6495 if (N->getOperand(0) == N->getOperand(1) && VT.getVectorNumElements() == 2) {
6496 assert(VT.getVectorElementType().getSizeInBits() == 64);
6497 return DAG.getNode(ARM64ISD::DUPLANE64, dl, VT,
6498 WidenVector(N->getOperand(0), DAG),
6499 DAG.getConstant(0, MVT::i64));
6502 // Canonicalise concat_vectors so that the right-hand vector has as few
6503 // bit-casts as possible before its real operation. The primary matching
6504 // destination for these operations will be the narrowing "2" instructions,
6505 // which depend on the operation being performed on this right-hand vector.
6507 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
6509 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
6511 SDValue Op1 = N->getOperand(1);
6512 if (Op1->getOpcode() != ISD::BITCAST)
6514 SDValue RHS = Op1->getOperand(0);
6515 MVT RHSTy = RHS.getValueType().getSimpleVT();
6516 // If the RHS is not a vector, this is not the pattern we're looking for.
6517 if (!RHSTy.isVector())
6520 DEBUG(dbgs() << "arm64-lower: concat_vectors bitcast simplification\n");
6522 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
6523 RHSTy.getVectorNumElements() * 2);
6525 ISD::BITCAST, dl, VT,
6526 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
6527 DAG.getNode(ISD::BITCAST, dl, RHSTy, N->getOperand(0)), RHS));
6530 static SDValue tryCombineFixedPointConvert(SDNode *N,
6531 TargetLowering::DAGCombinerInfo &DCI,
6532 SelectionDAG &DAG) {
6533 // Wait 'til after everything is legalized to try this. That way we have
6534 // legal vector types and such.
6535 if (DCI.isBeforeLegalizeOps())
6537 // Transform a scalar conversion of a value from a lane extract into a
6538 // lane extract of a vector conversion. E.g., from foo1 to foo2:
6539 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
6540 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
6542 // The second form interacts better with instruction selection and the
6543 // register allocator to avoid cross-class register copies that aren't
6544 // coalescable due to a lane reference.
6546 // Check the operand and see if it originates from a lane extract.
6547 SDValue Op1 = N->getOperand(1);
6548 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6549 // Yep, no additional predication needed. Perform the transform.
6550 SDValue IID = N->getOperand(0);
6551 SDValue Shift = N->getOperand(2);
6552 SDValue Vec = Op1.getOperand(0);
6553 SDValue Lane = Op1.getOperand(1);
6554 EVT ResTy = N->getValueType(0);
6558 // The vector width should be 128 bits by the time we get here, even
6559 // if it started as 64 bits (the extract_vector handling will have
6561 assert(Vec.getValueType().getSizeInBits() == 128 &&
6562 "unexpected vector size on extract_vector_elt!");
6563 if (Vec.getValueType() == MVT::v4i32)
6564 VecResTy = MVT::v4f32;
6565 else if (Vec.getValueType() == MVT::v2i64)
6566 VecResTy = MVT::v2f64;
6568 assert(0 && "unexpected vector type!");
6571 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
6572 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
6577 // AArch64 high-vector "long" operations are formed by performing the non-high
6578 // version on an extract_subvector of each operand which gets the high half:
6580 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
6582 // However, there are cases which don't have an extract_high explicitly, but
6583 // have another operation that can be made compatible with one for free. For
6586 // (dupv64 scalar) --> (extract_high (dup128 scalar))
6588 // This routine does the actual conversion of such DUPs, once outer routines
6589 // have determined that everything else is in order.
6590 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
6591 // We can handle most types of duplicate, but the lane ones have an extra
6592 // operand saying *which* lane, so we need to know.
6594 switch (N.getOpcode()) {
6598 case ARM64ISD::DUPLANE8:
6599 case ARM64ISD::DUPLANE16:
6600 case ARM64ISD::DUPLANE32:
6601 case ARM64ISD::DUPLANE64:
6608 MVT NarrowTy = N.getSimpleValueType();
6609 if (!NarrowTy.is64BitVector())
6612 MVT ElementTy = NarrowTy.getVectorElementType();
6613 unsigned NumElems = NarrowTy.getVectorNumElements();
6614 MVT NewDUPVT = MVT::getVectorVT(ElementTy, NumElems * 2);
6618 NewDUP = DAG.getNode(N.getOpcode(), SDLoc(N), NewDUPVT, N.getOperand(0),
6621 NewDUP = DAG.getNode(ARM64ISD::DUP, SDLoc(N), NewDUPVT, N.getOperand(0));
6623 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N.getNode()), NarrowTy,
6624 NewDUP, DAG.getConstant(NumElems, MVT::i64));
6627 static bool isEssentiallyExtractSubvector(SDValue N) {
6628 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
6631 return N.getOpcode() == ISD::BITCAST &&
6632 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
6635 /// \brief Helper structure to keep track of ISD::SET_CC operands.
6636 struct GenericSetCCInfo {
6637 const SDValue *Opnd0;
6638 const SDValue *Opnd1;
6642 /// \brief Helper structure to keep track of a SET_CC lowered into ARM64 code.
6643 struct ARM64SetCCInfo {
6645 ARM64CC::CondCode CC;
6648 /// \brief Helper structure to keep track of SetCC information.
6650 GenericSetCCInfo Generic;
6651 ARM64SetCCInfo ARM64;
6654 /// \brief Helper structure to be able to read SetCC information.
6655 /// If set to true, IsARM64 field, Info is a ARM64SetCCInfo, otherwise Info is
6656 /// a GenericSetCCInfo.
6657 struct SetCCInfoAndKind {
6662 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
6664 /// ARM64 lowered one.
6665 /// \p SetCCInfo is filled accordingly.
6666 /// \post SetCCInfo is meanginfull only when this function returns true.
6667 /// \return True when Op is a kind of SET_CC operation.
6668 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
6669 // If this is a setcc, this is straight forward.
6670 if (Op.getOpcode() == ISD::SETCC) {
6671 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
6672 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
6673 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6674 SetCCInfo.IsARM64 = false;
6677 // Otherwise, check if this is a matching csel instruction.
6681 if (Op.getOpcode() != ARM64ISD::CSEL)
6683 // Set the information about the operands.
6684 // TODO: we want the operands of the Cmp not the csel
6685 SetCCInfo.Info.ARM64.Cmp = &Op.getOperand(3);
6686 SetCCInfo.IsARM64 = true;
6687 SetCCInfo.Info.ARM64.CC = static_cast<ARM64CC::CondCode>(
6688 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
6690 // Check that the operands matches the constraints:
6691 // (1) Both operands must be constants.
6692 // (2) One must be 1 and the other must be 0.
6693 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
6694 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6697 if (!TValue || !FValue)
6701 if (!TValue->isOne()) {
6702 // Update the comparison when we are interested in !cc.
6703 std::swap(TValue, FValue);
6704 SetCCInfo.Info.ARM64.CC =
6705 ARM64CC::getInvertedCondCode(SetCCInfo.Info.ARM64.CC);
6707 return TValue->isOne() && FValue->isNullValue();
6710 // The folding we want to perform is:
6711 // (add x, (setcc cc ...) )
6713 // (csel x, (add x, 1), !cc ...)
6715 // The latter will get matched to a CSINC instruction.
6716 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
6717 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
6718 SDValue LHS = Op->getOperand(0);
6719 SDValue RHS = Op->getOperand(1);
6720 SetCCInfoAndKind InfoAndKind;
6722 // If neither operand is a SET_CC, give up.
6723 if (!isSetCC(LHS, InfoAndKind)) {
6724 std::swap(LHS, RHS);
6725 if (!isSetCC(LHS, InfoAndKind))
6729 // FIXME: This could be generatized to work for FP comparisons.
6730 EVT CmpVT = InfoAndKind.IsARM64
6731 ? InfoAndKind.Info.ARM64.Cmp->getOperand(0).getValueType()
6732 : InfoAndKind.Info.Generic.Opnd0->getValueType();
6733 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
6739 if (InfoAndKind.IsARM64) {
6740 CCVal = DAG.getConstant(
6741 ARM64CC::getInvertedCondCode(InfoAndKind.Info.ARM64.CC), MVT::i32);
6742 Cmp = *InfoAndKind.Info.ARM64.Cmp;
6744 Cmp = getARM64Cmp(*InfoAndKind.Info.Generic.Opnd0,
6745 *InfoAndKind.Info.Generic.Opnd1,
6746 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
6749 EVT VT = Op->getValueType(0);
6750 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, VT));
6751 return DAG.getNode(ARM64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
6754 // The basic add/sub long vector instructions have variants with "2" on the end
6755 // which act on the high-half of their inputs. They are normally matched by
6758 // (add (zeroext (extract_high LHS)),
6759 // (zeroext (extract_high RHS)))
6760 // -> uaddl2 vD, vN, vM
6762 // However, if one of the extracts is something like a duplicate, this
6763 // instruction can still be used profitably. This function puts the DAG into a
6764 // more appropriate form for those patterns to trigger.
6765 static SDValue performAddSubLongCombine(SDNode *N,
6766 TargetLowering::DAGCombinerInfo &DCI,
6767 SelectionDAG &DAG) {
6768 if (DCI.isBeforeLegalizeOps())
6771 MVT VT = N->getSimpleValueType(0);
6772 if (!VT.is128BitVector()) {
6773 if (N->getOpcode() == ISD::ADD)
6774 return performSetccAddFolding(N, DAG);
6778 // Make sure both branches are extended in the same way.
6779 SDValue LHS = N->getOperand(0);
6780 SDValue RHS = N->getOperand(1);
6781 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
6782 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
6783 LHS.getOpcode() != RHS.getOpcode())
6786 unsigned ExtType = LHS.getOpcode();
6788 // It's not worth doing if at least one of the inputs isn't already an
6789 // extract, but we don't know which it'll be so we have to try both.
6790 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
6791 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
6795 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
6796 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
6797 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
6801 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
6804 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
6807 // Massage DAGs which we can use the high-half "long" operations on into
6808 // something isel will recognize better. E.g.
6810 // (arm64_neon_umull (extract_high vec) (dupv64 scalar)) -->
6811 // (arm64_neon_umull (extract_high (v2i64 vec)))
6812 // (extract_high (v2i64 (dup128 scalar)))))
6814 static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N,
6815 TargetLowering::DAGCombinerInfo &DCI,
6816 SelectionDAG &DAG) {
6817 if (DCI.isBeforeLegalizeOps())
6820 SDValue LHS = N->getOperand(1);
6821 SDValue RHS = N->getOperand(2);
6822 assert(LHS.getValueType().is64BitVector() &&
6823 RHS.getValueType().is64BitVector() &&
6824 "unexpected shape for long operation");
6826 // Either node could be a DUP, but it's not worth doing both of them (you'd
6827 // just as well use the non-high version) so look for a corresponding extract
6828 // operation on the other "wing".
6829 if (isEssentiallyExtractSubvector(LHS)) {
6830 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
6833 } else if (isEssentiallyExtractSubvector(RHS)) {
6834 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
6839 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
6840 N->getOperand(0), LHS, RHS);
6843 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
6844 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
6845 unsigned ElemBits = ElemTy.getSizeInBits();
6847 int64_t ShiftAmount;
6848 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
6849 APInt SplatValue, SplatUndef;
6850 unsigned SplatBitSize;
6852 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
6853 HasAnyUndefs, ElemBits) ||
6854 SplatBitSize != ElemBits)
6857 ShiftAmount = SplatValue.getSExtValue();
6858 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
6859 ShiftAmount = CVN->getSExtValue();
6867 llvm_unreachable("Unknown shift intrinsic");
6868 case Intrinsic::arm64_neon_sqshl:
6869 Opcode = ARM64ISD::SQSHL_I;
6870 IsRightShift = false;
6872 case Intrinsic::arm64_neon_uqshl:
6873 Opcode = ARM64ISD::UQSHL_I;
6874 IsRightShift = false;
6876 case Intrinsic::arm64_neon_srshl:
6877 Opcode = ARM64ISD::SRSHR_I;
6878 IsRightShift = true;
6880 case Intrinsic::arm64_neon_urshl:
6881 Opcode = ARM64ISD::URSHR_I;
6882 IsRightShift = true;
6884 case Intrinsic::arm64_neon_sqshlu:
6885 Opcode = ARM64ISD::SQSHLU_I;
6886 IsRightShift = false;
6890 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits)
6891 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6892 DAG.getConstant(-ShiftAmount, MVT::i32));
6893 else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount <= ElemBits)
6894 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6895 DAG.getConstant(ShiftAmount, MVT::i32));
6900 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
6901 // the intrinsics must be legal and take an i32, this means there's almost
6902 // certainly going to be a zext in the DAG which we can eliminate.
6903 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
6904 SDValue AndN = N->getOperand(2);
6905 if (AndN.getOpcode() != ISD::AND)
6908 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
6909 if (!CMask || CMask->getZExtValue() != Mask)
6912 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
6913 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
6916 static SDValue performIntrinsicCombine(SDNode *N,
6917 TargetLowering::DAGCombinerInfo &DCI,
6918 const ARM64Subtarget *Subtarget) {
6919 SelectionDAG &DAG = DCI.DAG;
6920 unsigned IID = getIntrinsicID(N);
6924 case Intrinsic::arm64_neon_vcvtfxs2fp:
6925 case Intrinsic::arm64_neon_vcvtfxu2fp:
6926 return tryCombineFixedPointConvert(N, DCI, DAG);
6928 case Intrinsic::arm64_neon_fmax:
6929 return DAG.getNode(ARM64ISD::FMAX, SDLoc(N), N->getValueType(0),
6930 N->getOperand(1), N->getOperand(2));
6931 case Intrinsic::arm64_neon_fmin:
6932 return DAG.getNode(ARM64ISD::FMIN, SDLoc(N), N->getValueType(0),
6933 N->getOperand(1), N->getOperand(2));
6934 case Intrinsic::arm64_neon_smull:
6935 case Intrinsic::arm64_neon_umull:
6936 case Intrinsic::arm64_neon_pmull:
6937 case Intrinsic::arm64_neon_sqdmull:
6938 return tryCombineLongOpWithDup(IID, N, DCI, DAG);
6939 case Intrinsic::arm64_neon_sqshl:
6940 case Intrinsic::arm64_neon_uqshl:
6941 case Intrinsic::arm64_neon_sqshlu:
6942 case Intrinsic::arm64_neon_srshl:
6943 case Intrinsic::arm64_neon_urshl:
6944 return tryCombineShiftImm(IID, N, DAG);
6945 case Intrinsic::arm64_crc32b:
6946 case Intrinsic::arm64_crc32cb:
6947 return tryCombineCRC32(0xff, N, DAG);
6948 case Intrinsic::arm64_crc32h:
6949 case Intrinsic::arm64_crc32ch:
6950 return tryCombineCRC32(0xffff, N, DAG);
6955 static SDValue performExtendCombine(SDNode *N,
6956 TargetLowering::DAGCombinerInfo &DCI,
6957 SelectionDAG &DAG) {
6958 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
6959 // we can convert that DUP into another extract_high (of a bigger DUP), which
6960 // helps the backend to decide that an sabdl2 would be useful, saving a real
6961 // extract_high operation.
6962 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
6963 N->getOperand(0).getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
6964 SDNode *ABDNode = N->getOperand(0).getNode();
6965 unsigned IID = getIntrinsicID(ABDNode);
6966 if (IID == Intrinsic::arm64_neon_sabd ||
6967 IID == Intrinsic::arm64_neon_uabd) {
6968 SDValue NewABD = tryCombineLongOpWithDup(IID, ABDNode, DCI, DAG);
6969 if (!NewABD.getNode())
6972 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
6977 // This is effectively a custom type legalization for ARM64.
6979 // Type legalization will split an extend of a small, legal, type to a larger
6980 // illegal type by first splitting the destination type, often creating
6981 // illegal source types, which then get legalized in isel-confusing ways,
6982 // leading to really terrible codegen. E.g.,
6983 // %result = v8i32 sext v8i8 %value
6985 // %losrc = extract_subreg %value, ...
6986 // %hisrc = extract_subreg %value, ...
6987 // %lo = v4i32 sext v4i8 %losrc
6988 // %hi = v4i32 sext v4i8 %hisrc
6989 // Things go rapidly downhill from there.
6991 // For ARM64, the [sz]ext vector instructions can only go up one element
6992 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
6993 // take two instructions.
6995 // This implies that the most efficient way to do the extend from v8i8
6996 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
6997 // the normal splitting to happen for the v8i16->v8i32.
6999 // This is pre-legalization to catch some cases where the default
7000 // type legalization will create ill-tempered code.
7001 if (!DCI.isBeforeLegalizeOps())
7004 // We're only interested in cleaning things up for non-legal vector types
7005 // here. If both the source and destination are legal, things will just
7006 // work naturally without any fiddling.
7007 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7008 EVT ResVT = N->getValueType(0);
7009 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
7011 // If the vector type isn't a simple VT, it's beyond the scope of what
7012 // we're worried about here. Let legalization do its thing and hope for
7014 if (!ResVT.isSimple())
7017 SDValue Src = N->getOperand(0);
7018 MVT SrcVT = Src->getValueType(0).getSimpleVT();
7019 // If the source VT is a 64-bit vector, we can play games and get the
7020 // better results we want.
7021 if (SrcVT.getSizeInBits() != 64)
7024 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
7025 unsigned ElementCount = SrcVT.getVectorNumElements();
7026 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
7028 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
7030 // Now split the rest of the operation into two halves, each with a 64
7034 unsigned NumElements = ResVT.getVectorNumElements();
7035 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
7036 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
7037 ResVT.getVectorElementType(), NumElements / 2);
7039 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
7040 LoVT.getVectorNumElements());
7041 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7042 DAG.getIntPtrConstant(0));
7043 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7044 DAG.getIntPtrConstant(InNVT.getVectorNumElements()));
7045 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
7046 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
7048 // Now combine the parts back together so we still have a single result
7049 // like the combiner expects.
7050 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
7053 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
7054 /// value. The load store optimizer pass will merge them to store pair stores.
7055 /// This has better performance than a splat of the scalar followed by a split
7056 /// vector store. Even if the stores are not merged it is four stores vs a dup,
7057 /// followed by an ext.b and two stores.
7058 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
7059 SDValue StVal = St->getValue();
7060 EVT VT = StVal.getValueType();
7062 // Don't replace floating point stores, they possibly won't be transformed to
7063 // stp because of the store pair suppress pass.
7064 if (VT.isFloatingPoint())
7067 // Check for insert vector elements.
7068 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
7071 // We can express a splat as store pair(s) for 2 or 4 elements.
7072 unsigned NumVecElts = VT.getVectorNumElements();
7073 if (NumVecElts != 4 && NumVecElts != 2)
7075 SDValue SplatVal = StVal.getOperand(1);
7076 unsigned RemainInsertElts = NumVecElts - 1;
7078 // Check that this is a splat.
7079 while (--RemainInsertElts) {
7080 SDValue NextInsertElt = StVal.getOperand(0);
7081 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
7083 if (NextInsertElt.getOperand(1) != SplatVal)
7085 StVal = NextInsertElt;
7087 unsigned OrigAlignment = St->getAlignment();
7088 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
7089 unsigned Alignment = std::min(OrigAlignment, EltOffset);
7091 // Create scalar stores. This is at least as good as the code sequence for a
7092 // split unaligned store wich is a dup.s, ext.b, and two stores.
7093 // Most of the time the three stores should be replaced by store pair
7094 // instructions (stp).
7096 SDValue BasePtr = St->getBasePtr();
7098 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
7099 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
7101 unsigned Offset = EltOffset;
7102 while (--NumVecElts) {
7103 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7104 DAG.getConstant(Offset, MVT::i64));
7105 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
7106 St->getPointerInfo(), St->isVolatile(),
7107 St->isNonTemporal(), Alignment);
7108 Offset += EltOffset;
7113 static SDValue performSTORECombine(SDNode *N,
7114 TargetLowering::DAGCombinerInfo &DCI,
7116 const ARM64Subtarget *Subtarget) {
7117 if (!DCI.isBeforeLegalize())
7120 StoreSDNode *S = cast<StoreSDNode>(N);
7121 if (S->isVolatile())
7124 // Cyclone has bad performance on unaligned 16B stores when crossing line and
7125 // page boundries. We want to split such stores.
7126 if (!Subtarget->isCyclone())
7129 // Don't split at Oz.
7130 MachineFunction &MF = DAG.getMachineFunction();
7131 bool IsMinSize = MF.getFunction()->getAttributes().hasAttribute(
7132 AttributeSet::FunctionIndex, Attribute::MinSize);
7136 SDValue StVal = S->getValue();
7137 EVT VT = StVal.getValueType();
7139 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
7140 // those up regresses performance on micro-benchmarks and olden/bh.
7141 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
7144 // Split unaligned 16B stores. They are terrible for performance.
7145 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
7146 // extensions can use this to mark that it does not want splitting to happen
7147 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
7148 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
7149 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
7150 S->getAlignment() <= 2)
7153 // If we get a splat of a scalar convert this vector store to a store of
7154 // scalars. They will be merged into store pairs thereby removing two
7156 SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S);
7157 if (ReplacedSplat != SDValue())
7158 return ReplacedSplat;
7161 unsigned NumElts = VT.getVectorNumElements() / 2;
7162 // Split VT into two.
7164 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
7165 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7166 DAG.getIntPtrConstant(0));
7167 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7168 DAG.getIntPtrConstant(NumElts));
7169 SDValue BasePtr = S->getBasePtr();
7171 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
7172 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
7173 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7174 DAG.getConstant(8, MVT::i64));
7175 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
7176 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
7180 // Optimize compare with zero and branch.
7181 static SDValue performBRCONDCombine(SDNode *N,
7182 TargetLowering::DAGCombinerInfo &DCI,
7183 SelectionDAG &DAG) {
7184 SDValue Chain = N->getOperand(0);
7185 SDValue Dest = N->getOperand(1);
7186 SDValue CCVal = N->getOperand(2);
7187 SDValue Cmp = N->getOperand(3);
7189 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
7190 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
7191 if (CC != ARM64CC::EQ && CC != ARM64CC::NE)
7194 unsigned CmpOpc = Cmp.getOpcode();
7195 if (CmpOpc != ARM64ISD::ADDS && CmpOpc != ARM64ISD::SUBS)
7198 // Only attempt folding if there is only one use of the flag and no use of the
7200 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
7203 SDValue LHS = Cmp.getOperand(0);
7204 SDValue RHS = Cmp.getOperand(1);
7206 assert(LHS.getValueType() == RHS.getValueType() &&
7207 "Expected the value type to be the same for both operands!");
7208 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
7211 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
7212 std::swap(LHS, RHS);
7214 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
7217 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
7218 LHS.getOpcode() == ISD::SRL)
7221 // Fold the compare into the branch instruction.
7223 if (CC == ARM64CC::EQ)
7224 BR = DAG.getNode(ARM64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7226 BR = DAG.getNode(ARM64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7228 // Do not add new nodes to DAG combiner worklist.
7229 DCI.CombineTo(N, BR, false);
7234 // vselect (v1i1 setcc) ->
7235 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
7236 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
7237 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
7239 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
7240 SDValue N0 = N->getOperand(0);
7241 EVT CCVT = N0.getValueType();
7243 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
7244 CCVT.getVectorElementType() != MVT::i1)
7247 EVT ResVT = N->getValueType(0);
7248 EVT CmpVT = N0.getOperand(0).getValueType();
7249 // Only combine when the result type is of the same size as the compared
7251 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
7254 SDValue IfTrue = N->getOperand(1);
7255 SDValue IfFalse = N->getOperand(2);
7257 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
7258 N0.getOperand(0), N0.getOperand(1),
7259 cast<CondCodeSDNode>(N0.getOperand(2))->get());
7260 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
7264 SDValue ARM64TargetLowering::PerformDAGCombine(SDNode *N,
7265 DAGCombinerInfo &DCI) const {
7266 SelectionDAG &DAG = DCI.DAG;
7267 switch (N->getOpcode()) {
7272 return performAddSubLongCombine(N, DCI, DAG);
7274 return performXorCombine(N, DAG, DCI, Subtarget);
7276 return performMulCombine(N, DAG, DCI, Subtarget);
7277 case ISD::SINT_TO_FP:
7278 case ISD::UINT_TO_FP:
7279 return performIntToFpCombine(N, DAG);
7281 return performORCombine(N, DCI, Subtarget);
7282 case ISD::INTRINSIC_WO_CHAIN:
7283 return performIntrinsicCombine(N, DCI, Subtarget);
7284 case ISD::ANY_EXTEND:
7285 case ISD::ZERO_EXTEND:
7286 case ISD::SIGN_EXTEND:
7287 return performExtendCombine(N, DCI, DAG);
7289 return performBitcastCombine(N, DCI, DAG);
7290 case ISD::CONCAT_VECTORS:
7291 return performConcatVectorsCombine(N, DCI, DAG);
7293 return performVSelectCombine(N, DCI.DAG);
7295 return performSTORECombine(N, DCI, DAG, Subtarget);
7296 case ARM64ISD::BRCOND:
7297 return performBRCONDCombine(N, DCI, DAG);
7302 // Check if the return value is used as only a return value, as otherwise
7303 // we can't perform a tail-call. In particular, we need to check for
7304 // target ISD nodes that are returns and any other "odd" constructs
7305 // that the generic analysis code won't necessarily catch.
7306 bool ARM64TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
7307 if (N->getNumValues() != 1)
7309 if (!N->hasNUsesOfValue(1, 0))
7312 SDValue TCChain = Chain;
7313 SDNode *Copy = *N->use_begin();
7314 if (Copy->getOpcode() == ISD::CopyToReg) {
7315 // If the copy has a glue operand, we conservatively assume it isn't safe to
7316 // perform a tail call.
7317 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
7320 TCChain = Copy->getOperand(0);
7321 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
7324 bool HasRet = false;
7325 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
7327 if (UI->getOpcode() != ARM64ISD::RET_FLAG)
7339 // Return whether the an instruction can potentially be optimized to a tail
7340 // call. This will cause the optimizers to attempt to move, or duplicate,
7341 // return instructions to help enable tail call optimizations for this
7343 bool ARM64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
7344 if (!EnableARM64TailCalls)
7347 if (!CI->isTailCall())
7353 bool ARM64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
7355 ISD::MemIndexedMode &AM,
7357 SelectionDAG &DAG) const {
7358 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
7361 Base = Op->getOperand(0);
7362 // All of the indexed addressing mode instructions take a signed
7363 // 9 bit immediate offset.
7364 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
7365 int64_t RHSC = (int64_t)RHS->getZExtValue();
7366 if (RHSC >= 256 || RHSC <= -256)
7368 IsInc = (Op->getOpcode() == ISD::ADD);
7369 Offset = Op->getOperand(1);
7375 bool ARM64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
7377 ISD::MemIndexedMode &AM,
7378 SelectionDAG &DAG) const {
7381 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7382 VT = LD->getMemoryVT();
7383 Ptr = LD->getBasePtr();
7384 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7385 VT = ST->getMemoryVT();
7386 Ptr = ST->getBasePtr();
7391 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
7393 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
7397 bool ARM64TargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
7400 ISD::MemIndexedMode &AM,
7401 SelectionDAG &DAG) const {
7404 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7405 VT = LD->getMemoryVT();
7406 Ptr = LD->getBasePtr();
7407 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7408 VT = ST->getMemoryVT();
7409 Ptr = ST->getBasePtr();
7414 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
7416 // Post-indexing updates the base, so it's not a valid transform
7417 // if that's not the same as the load's pointer.
7420 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
7424 /// The only 128-bit atomic operation is an stxp that succeeds. In particular
7425 /// neither ldp nor ldxp are atomic. So the canonical sequence for an atomic
7428 /// ldxp x0, x1, [x8]
7429 /// stxp w2, x0, x1, [x8]
7431 /// If the stxp succeeds then the ldxp managed to get both halves without an
7432 /// intervening stxp from a different thread and the read was atomic.
7433 static void ReplaceATOMIC_LOAD_128(SDNode *N, SmallVectorImpl<SDValue> &Results,
7434 SelectionDAG &DAG) {
7436 AtomicSDNode *AN = cast<AtomicSDNode>(N);
7437 EVT VT = AN->getMemoryVT();
7438 SDValue Zero = DAG.getConstant(0, VT);
7440 // FIXME: Really want ATOMIC_LOAD_NOP but that doesn't fit into the existing
7441 // scheme very well. Given the complexity of what we're already generating, an
7442 // extra couple of ORRs probably won't make much difference.
7443 SDValue Result = DAG.getAtomic(ISD::ATOMIC_LOAD_OR, DL, AN->getMemoryVT(),
7444 N->getOperand(0), N->getOperand(1), Zero,
7445 AN->getMemOperand(), AN->getOrdering(),
7446 AN->getSynchScope());
7448 Results.push_back(Result.getValue(0)); // Value
7449 Results.push_back(Result.getValue(1)); // Chain
7452 static void ReplaceATOMIC_OP_128(SDNode *N, SmallVectorImpl<SDValue> &Results,
7453 SelectionDAG &DAG, unsigned NewOp) {
7455 AtomicOrdering Ordering = cast<AtomicSDNode>(N)->getOrdering();
7456 assert(N->getValueType(0) == MVT::i128 &&
7457 "Only know how to expand i128 atomics");
7459 SmallVector<SDValue, 6> Ops;
7460 Ops.push_back(N->getOperand(1)); // Ptr
7462 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7463 N->getOperand(2), DAG.getIntPtrConstant(0)));
7464 // High part of Val1
7465 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7466 N->getOperand(2), DAG.getIntPtrConstant(1)));
7467 if (NewOp == ARM64::ATOMIC_CMP_SWAP_I128) {
7469 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7470 N->getOperand(3), DAG.getIntPtrConstant(0)));
7471 // High part of Val2
7472 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7473 N->getOperand(3), DAG.getIntPtrConstant(1)));
7476 Ops.push_back(DAG.getTargetConstant(Ordering, MVT::i32));
7477 Ops.push_back(N->getOperand(0)); // Chain
7479 SDVTList Tys = DAG.getVTList(MVT::i64, MVT::i64, MVT::Other);
7480 SDNode *Result = DAG.getMachineNode(NewOp, DL, Tys, Ops);
7481 SDValue OpsF[] = { SDValue(Result, 0), SDValue(Result, 1) };
7482 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i128, OpsF, 2));
7483 Results.push_back(SDValue(Result, 2));
7486 void ARM64TargetLowering::ReplaceNodeResults(SDNode *N,
7487 SmallVectorImpl<SDValue> &Results,
7488 SelectionDAG &DAG) const {
7489 switch (N->getOpcode()) {
7491 llvm_unreachable("Don't know how to custom expand this");
7492 case ISD::ATOMIC_LOAD:
7493 ReplaceATOMIC_LOAD_128(N, Results, DAG);
7495 case ISD::ATOMIC_LOAD_ADD:
7496 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_ADD_I128);
7498 case ISD::ATOMIC_LOAD_SUB:
7499 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_SUB_I128);
7501 case ISD::ATOMIC_LOAD_AND:
7502 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_AND_I128);
7504 case ISD::ATOMIC_LOAD_OR:
7505 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_OR_I128);
7507 case ISD::ATOMIC_LOAD_XOR:
7508 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_XOR_I128);
7510 case ISD::ATOMIC_LOAD_NAND:
7511 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_NAND_I128);
7513 case ISD::ATOMIC_SWAP:
7514 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_SWAP_I128);
7516 case ISD::ATOMIC_LOAD_MIN:
7517 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_MIN_I128);
7519 case ISD::ATOMIC_LOAD_MAX:
7520 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_MAX_I128);
7522 case ISD::ATOMIC_LOAD_UMIN:
7523 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_UMIN_I128);
7525 case ISD::ATOMIC_LOAD_UMAX:
7526 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_UMAX_I128);
7528 case ISD::ATOMIC_CMP_SWAP:
7529 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_CMP_SWAP_I128);
7531 case ISD::FP_TO_UINT:
7532 case ISD::FP_TO_SINT:
7533 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
7534 // Let normal code take care of it by not adding anything to Results.