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 // Custom lowering hooks are needed for XOR
177 // to fold it into CSINC/CSINV.
178 setOperationAction(ISD::XOR, MVT::i32, Custom);
179 setOperationAction(ISD::XOR, MVT::i64, Custom);
181 // Virtually no operation on f128 is legal, but LLVM can't expand them when
182 // there's a valid register class, so we need custom operations in most cases.
183 setOperationAction(ISD::FABS, MVT::f128, Expand);
184 setOperationAction(ISD::FADD, MVT::f128, Custom);
185 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
186 setOperationAction(ISD::FCOS, MVT::f128, Expand);
187 setOperationAction(ISD::FDIV, MVT::f128, Custom);
188 setOperationAction(ISD::FMA, MVT::f128, Expand);
189 setOperationAction(ISD::FMUL, MVT::f128, Custom);
190 setOperationAction(ISD::FNEG, MVT::f128, Expand);
191 setOperationAction(ISD::FPOW, MVT::f128, Expand);
192 setOperationAction(ISD::FREM, MVT::f128, Expand);
193 setOperationAction(ISD::FRINT, MVT::f128, Expand);
194 setOperationAction(ISD::FSIN, MVT::f128, Expand);
195 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
196 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
197 setOperationAction(ISD::FSUB, MVT::f128, Custom);
198 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
199 setOperationAction(ISD::SETCC, MVT::f128, Custom);
200 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
201 setOperationAction(ISD::SELECT, MVT::f128, Custom);
202 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
203 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
205 // Lowering for many of the conversions is actually specified by the non-f128
206 // type. The LowerXXX function will be trivial when f128 isn't involved.
207 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
208 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
209 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
210 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
211 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
212 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
213 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
214 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
215 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
216 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
217 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
218 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
219 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
220 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
223 setOperationAction(ISD::ATOMIC_SWAP, MVT::i128, Custom);
224 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i128, Custom);
225 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i128, Custom);
226 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i128, Custom);
227 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i128, Custom);
228 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i128, Custom);
229 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i128, Custom);
230 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i128, Custom);
231 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i128, Custom);
232 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i128, Custom);
233 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i128, Custom);
234 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i128, Custom);
235 // These are surprisingly difficult. The only single-copy atomic 128-bit
236 // instruction on AArch64 is stxp (when it succeeds). So a store can safely
237 // become a simple swap, but a load can only be determined to have been atomic
238 // if storing the same value back succeeds.
239 setOperationAction(ISD::ATOMIC_LOAD, MVT::i128, Custom);
240 setOperationAction(ISD::ATOMIC_STORE, MVT::i128, Expand);
242 // Variable arguments.
243 setOperationAction(ISD::VASTART, MVT::Other, Custom);
244 setOperationAction(ISD::VAARG, MVT::Other, Custom);
245 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
246 setOperationAction(ISD::VAEND, MVT::Other, Expand);
248 // Variable-sized objects.
249 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
250 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
251 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
253 // Exception handling.
254 // FIXME: These are guesses. Has this been defined yet?
255 setExceptionPointerRegister(ARM64::X0);
256 setExceptionSelectorRegister(ARM64::X1);
258 // Constant pool entries
259 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
262 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
264 // Add/Sub overflow ops with MVT::Glues are lowered to CPSR dependences.
265 setOperationAction(ISD::ADDC, MVT::i32, Custom);
266 setOperationAction(ISD::ADDE, MVT::i32, Custom);
267 setOperationAction(ISD::SUBC, MVT::i32, Custom);
268 setOperationAction(ISD::SUBE, MVT::i32, Custom);
269 setOperationAction(ISD::ADDC, MVT::i64, Custom);
270 setOperationAction(ISD::ADDE, MVT::i64, Custom);
271 setOperationAction(ISD::SUBC, MVT::i64, Custom);
272 setOperationAction(ISD::SUBE, MVT::i64, Custom);
274 // ARM64 lacks both left-rotate and popcount instructions.
275 setOperationAction(ISD::ROTL, MVT::i32, Expand);
276 setOperationAction(ISD::ROTL, MVT::i64, Expand);
278 // ARM64 doesn't have a direct vector ->f32 conversion instructions for
279 // elements smaller than i32, so promote the input to i32 first.
280 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
281 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
282 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
283 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
284 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
285 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
286 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
287 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
288 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
290 // ARM64 doesn't have {U|S}MUL_LOHI.
291 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
292 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
294 // ARM64 doesn't have MUL.2d:
295 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
297 // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
298 // counterparts, which ARM64 supports directly.
299 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
300 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
301 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
302 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
304 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
305 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
307 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
308 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
309 setOperationAction(ISD::SREM, MVT::i32, Expand);
310 setOperationAction(ISD::SREM, MVT::i64, Expand);
311 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
312 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
313 setOperationAction(ISD::UREM, MVT::i32, Expand);
314 setOperationAction(ISD::UREM, MVT::i64, Expand);
316 // Custom lower Add/Sub/Mul with overflow.
317 setOperationAction(ISD::SADDO, MVT::i32, Custom);
318 setOperationAction(ISD::SADDO, MVT::i64, Custom);
319 setOperationAction(ISD::UADDO, MVT::i32, Custom);
320 setOperationAction(ISD::UADDO, MVT::i64, Custom);
321 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
322 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
323 setOperationAction(ISD::USUBO, MVT::i32, Custom);
324 setOperationAction(ISD::USUBO, MVT::i64, Custom);
325 setOperationAction(ISD::SMULO, MVT::i32, Custom);
326 setOperationAction(ISD::SMULO, MVT::i64, Custom);
327 setOperationAction(ISD::UMULO, MVT::i32, Custom);
328 setOperationAction(ISD::UMULO, MVT::i64, Custom);
330 setOperationAction(ISD::FSIN, MVT::f32, Expand);
331 setOperationAction(ISD::FSIN, MVT::f64, Expand);
332 setOperationAction(ISD::FCOS, MVT::f32, Expand);
333 setOperationAction(ISD::FCOS, MVT::f64, Expand);
334 setOperationAction(ISD::FPOW, MVT::f32, Expand);
335 setOperationAction(ISD::FPOW, MVT::f64, Expand);
336 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
337 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
339 // ARM64 has implementations of a lot of rounding-like FP operations.
340 static MVT RoundingTypes[] = { MVT::f32, MVT::f64, MVT::v2f32,
341 MVT::v4f32, MVT::v2f64 };
342 for (unsigned I = 0; I < array_lengthof(RoundingTypes); ++I) {
343 MVT Ty = RoundingTypes[I];
344 setOperationAction(ISD::FFLOOR, Ty, Legal);
345 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
346 setOperationAction(ISD::FCEIL, Ty, Legal);
347 setOperationAction(ISD::FRINT, Ty, Legal);
348 setOperationAction(ISD::FTRUNC, Ty, Legal);
349 setOperationAction(ISD::FROUND, Ty, Legal);
352 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
354 if (Subtarget->isTargetMachO()) {
355 // For iOS, we don't want to the normal expansion of a libcall to
356 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
358 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
359 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
361 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
362 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
365 // ARM64 does not have floating-point extending loads, i1 sign-extending load,
366 // floating-point truncating stores, or v2i32->v2i16 truncating store.
367 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
368 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
369 setLoadExtAction(ISD::EXTLOAD, MVT::f80, Expand);
370 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Expand);
371 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
372 setTruncStoreAction(MVT::f128, MVT::f80, Expand);
373 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
374 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
375 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
376 // Indexed loads and stores are supported.
377 for (unsigned im = (unsigned)ISD::PRE_INC;
378 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
379 setIndexedLoadAction(im, MVT::i8, Legal);
380 setIndexedLoadAction(im, MVT::i16, Legal);
381 setIndexedLoadAction(im, MVT::i32, Legal);
382 setIndexedLoadAction(im, MVT::i64, Legal);
383 setIndexedLoadAction(im, MVT::f64, Legal);
384 setIndexedLoadAction(im, MVT::f32, Legal);
385 setIndexedStoreAction(im, MVT::i8, Legal);
386 setIndexedStoreAction(im, MVT::i16, Legal);
387 setIndexedStoreAction(im, MVT::i32, Legal);
388 setIndexedStoreAction(im, MVT::i64, Legal);
389 setIndexedStoreAction(im, MVT::f64, Legal);
390 setIndexedStoreAction(im, MVT::f32, Legal);
393 // Likewise, narrowing and extending vector loads/stores aren't handled
395 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
396 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
398 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,
401 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
402 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
403 setTruncStoreAction((MVT::SimpleValueType)VT,
404 (MVT::SimpleValueType)InnerVT, Expand);
405 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
406 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
407 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
411 setOperationAction(ISD::TRAP, MVT::Other, Legal);
412 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
414 // We combine OR nodes for bitfield operations.
415 setTargetDAGCombine(ISD::OR);
417 // Vector add and sub nodes may conceal a high-half opportunity.
418 // Also, try to fold ADD into CSINC/CSINV..
419 setTargetDAGCombine(ISD::ADD);
420 setTargetDAGCombine(ISD::SUB);
422 setTargetDAGCombine(ISD::XOR);
423 setTargetDAGCombine(ISD::SINT_TO_FP);
424 setTargetDAGCombine(ISD::UINT_TO_FP);
426 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
428 setTargetDAGCombine(ISD::ANY_EXTEND);
429 setTargetDAGCombine(ISD::ZERO_EXTEND);
430 setTargetDAGCombine(ISD::SIGN_EXTEND);
431 setTargetDAGCombine(ISD::BITCAST);
432 setTargetDAGCombine(ISD::CONCAT_VECTORS);
433 setTargetDAGCombine(ISD::STORE);
435 setTargetDAGCombine(ISD::MUL);
437 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
438 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
439 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
441 setStackPointerRegisterToSaveRestore(ARM64::SP);
443 setSchedulingPreference(Sched::Hybrid);
446 MaskAndBranchFoldingIsLegal = true;
448 setMinFunctionAlignment(2);
450 RequireStrictAlign = StrictAlign;
453 void ARM64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
454 if (VT == MVT::v2f32) {
455 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
456 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
458 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
459 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
460 } else if (VT == MVT::v2f64 || VT == MVT::v4f32) {
461 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
462 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
464 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
465 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
468 // Mark vector float intrinsics as expand.
469 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
470 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
471 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
472 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
473 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
474 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
475 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
476 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
477 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
478 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
481 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
482 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
483 setOperationAction(ISD::SCALAR_TO_VECTOR, VT.getSimpleVT(), Custom);
484 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
485 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
486 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
487 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
488 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
489 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
490 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
491 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
492 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
493 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
495 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
496 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
497 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
498 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
500 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
501 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
502 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
503 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
504 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
506 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
507 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
510 void ARM64TargetLowering::addDRTypeForNEON(MVT VT) {
511 addRegisterClass(VT, &ARM64::FPR64RegClass);
512 addTypeForNEON(VT, MVT::v2i32);
515 void ARM64TargetLowering::addQRTypeForNEON(MVT VT) {
516 addRegisterClass(VT, &ARM64::FPR128RegClass);
517 addTypeForNEON(VT, MVT::v4i32);
520 EVT ARM64TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
523 return VT.changeVectorElementTypeToInteger();
526 /// computeMaskedBitsForTargetNode - Determine which of the bits specified in
527 /// Mask are known to be either zero or one and return them in the
528 /// KnownZero/KnownOne bitsets.
529 void ARM64TargetLowering::computeMaskedBitsForTargetNode(
530 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
531 const SelectionDAG &DAG, unsigned Depth) const {
532 switch (Op.getOpcode()) {
535 case ARM64ISD::CSEL: {
536 APInt KnownZero2, KnownOne2;
537 DAG.ComputeMaskedBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
538 DAG.ComputeMaskedBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
539 KnownZero &= KnownZero2;
540 KnownOne &= KnownOne2;
543 case ISD::INTRINSIC_W_CHAIN:
545 case ISD::INTRINSIC_WO_CHAIN:
546 case ISD::INTRINSIC_VOID: {
547 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
551 case Intrinsic::arm64_neon_umaxv:
552 case Intrinsic::arm64_neon_uminv: {
553 // Figure out the datatype of the vector operand. The UMINV instruction
554 // will zero extend the result, so we can mark as known zero all the
555 // bits larger than the element datatype. 32-bit or larget doesn't need
556 // this as those are legal types and will be handled by isel directly.
557 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
558 unsigned BitWidth = KnownZero.getBitWidth();
559 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
560 assert(BitWidth >= 8 && "Unexpected width!");
561 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
563 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
564 assert(BitWidth >= 16 && "Unexpected width!");
565 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
575 MVT ARM64TargetLowering::getScalarShiftAmountTy(EVT LHSTy) const {
576 if (!LHSTy.isSimple())
578 MVT SimpleVT = LHSTy.getSimpleVT();
579 if (SimpleVT == MVT::i32)
584 unsigned ARM64TargetLowering::getMaximalGlobalOffset() const {
585 // FIXME: On ARM64, this depends on the type.
586 // Basically, the addressable offsets are o to 4095 * Ty.getSizeInBytes().
587 // and the offset has to be a multiple of the related size in bytes.
592 ARM64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
593 const TargetLibraryInfo *libInfo) const {
594 return ARM64::createFastISel(funcInfo, libInfo);
597 const char *ARM64TargetLowering::getTargetNodeName(unsigned Opcode) const {
601 case ARM64ISD::CALL: return "ARM64ISD::CALL";
602 case ARM64ISD::ADRP: return "ARM64ISD::ADRP";
603 case ARM64ISD::ADDlow: return "ARM64ISD::ADDlow";
604 case ARM64ISD::LOADgot: return "ARM64ISD::LOADgot";
605 case ARM64ISD::RET_FLAG: return "ARM64ISD::RET_FLAG";
606 case ARM64ISD::BRCOND: return "ARM64ISD::BRCOND";
607 case ARM64ISD::CSEL: return "ARM64ISD::CSEL";
608 case ARM64ISD::FCSEL: return "ARM64ISD::FCSEL";
609 case ARM64ISD::CSINV: return "ARM64ISD::CSINV";
610 case ARM64ISD::CSNEG: return "ARM64ISD::CSNEG";
611 case ARM64ISD::CSINC: return "ARM64ISD::CSINC";
612 case ARM64ISD::THREAD_POINTER: return "ARM64ISD::THREAD_POINTER";
613 case ARM64ISD::TLSDESC_CALL: return "ARM64ISD::TLSDESC_CALL";
614 case ARM64ISD::ADC: return "ARM64ISD::ADC";
615 case ARM64ISD::SBC: return "ARM64ISD::SBC";
616 case ARM64ISD::ADDS: return "ARM64ISD::ADDS";
617 case ARM64ISD::SUBS: return "ARM64ISD::SUBS";
618 case ARM64ISD::ADCS: return "ARM64ISD::ADCS";
619 case ARM64ISD::SBCS: return "ARM64ISD::SBCS";
620 case ARM64ISD::ANDS: return "ARM64ISD::ANDS";
621 case ARM64ISD::FCMP: return "ARM64ISD::FCMP";
622 case ARM64ISD::FMIN: return "ARM64ISD::FMIN";
623 case ARM64ISD::FMAX: return "ARM64ISD::FMAX";
624 case ARM64ISD::DUP: return "ARM64ISD::DUP";
625 case ARM64ISD::DUPLANE8: return "ARM64ISD::DUPLANE8";
626 case ARM64ISD::DUPLANE16: return "ARM64ISD::DUPLANE16";
627 case ARM64ISD::DUPLANE32: return "ARM64ISD::DUPLANE32";
628 case ARM64ISD::DUPLANE64: return "ARM64ISD::DUPLANE64";
629 case ARM64ISD::MOVI: return "ARM64ISD::MOVI";
630 case ARM64ISD::MOVIshift: return "ARM64ISD::MOVIshift";
631 case ARM64ISD::MOVIedit: return "ARM64ISD::MOVIedit";
632 case ARM64ISD::MOVImsl: return "ARM64ISD::MOVImsl";
633 case ARM64ISD::FMOV: return "ARM64ISD::FMOV";
634 case ARM64ISD::MVNIshift: return "ARM64ISD::MVNIshift";
635 case ARM64ISD::MVNImsl: return "ARM64ISD::MVNImsl";
636 case ARM64ISD::BICi: return "ARM64ISD::BICi";
637 case ARM64ISD::ORRi: return "ARM64ISD::ORRi";
638 case ARM64ISD::NEG: return "ARM64ISD::NEG";
639 case ARM64ISD::EXTR: return "ARM64ISD::EXTR";
640 case ARM64ISD::ZIP1: return "ARM64ISD::ZIP1";
641 case ARM64ISD::ZIP2: return "ARM64ISD::ZIP2";
642 case ARM64ISD::UZP1: return "ARM64ISD::UZP1";
643 case ARM64ISD::UZP2: return "ARM64ISD::UZP2";
644 case ARM64ISD::TRN1: return "ARM64ISD::TRN1";
645 case ARM64ISD::TRN2: return "ARM64ISD::TRN2";
646 case ARM64ISD::REV16: return "ARM64ISD::REV16";
647 case ARM64ISD::REV32: return "ARM64ISD::REV32";
648 case ARM64ISD::REV64: return "ARM64ISD::REV64";
649 case ARM64ISD::EXT: return "ARM64ISD::EXT";
650 case ARM64ISD::VSHL: return "ARM64ISD::VSHL";
651 case ARM64ISD::VLSHR: return "ARM64ISD::VLSHR";
652 case ARM64ISD::VASHR: return "ARM64ISD::VASHR";
653 case ARM64ISD::CMEQ: return "ARM64ISD::CMEQ";
654 case ARM64ISD::CMGE: return "ARM64ISD::CMGE";
655 case ARM64ISD::CMGT: return "ARM64ISD::CMGT";
656 case ARM64ISD::CMHI: return "ARM64ISD::CMHI";
657 case ARM64ISD::CMHS: return "ARM64ISD::CMHS";
658 case ARM64ISD::FCMEQ: return "ARM64ISD::FCMEQ";
659 case ARM64ISD::FCMGE: return "ARM64ISD::FCMGE";
660 case ARM64ISD::FCMGT: return "ARM64ISD::FCMGT";
661 case ARM64ISD::CMEQz: return "ARM64ISD::CMEQz";
662 case ARM64ISD::CMGEz: return "ARM64ISD::CMGEz";
663 case ARM64ISD::CMGTz: return "ARM64ISD::CMGTz";
664 case ARM64ISD::CMLEz: return "ARM64ISD::CMLEz";
665 case ARM64ISD::CMLTz: return "ARM64ISD::CMLTz";
666 case ARM64ISD::FCMEQz: return "ARM64ISD::FCMEQz";
667 case ARM64ISD::FCMGEz: return "ARM64ISD::FCMGEz";
668 case ARM64ISD::FCMGTz: return "ARM64ISD::FCMGTz";
669 case ARM64ISD::FCMLEz: return "ARM64ISD::FCMLEz";
670 case ARM64ISD::FCMLTz: return "ARM64ISD::FCMLTz";
671 case ARM64ISD::NOT: return "ARM64ISD::NOT";
672 case ARM64ISD::BIT: return "ARM64ISD::BIT";
673 case ARM64ISD::CBZ: return "ARM64ISD::CBZ";
674 case ARM64ISD::CBNZ: return "ARM64ISD::CBNZ";
675 case ARM64ISD::TBZ: return "ARM64ISD::TBZ";
676 case ARM64ISD::TBNZ: return "ARM64ISD::TBNZ";
677 case ARM64ISD::TC_RETURN: return "ARM64ISD::TC_RETURN";
678 case ARM64ISD::SITOF: return "ARM64ISD::SITOF";
679 case ARM64ISD::UITOF: return "ARM64ISD::UITOF";
680 case ARM64ISD::SQSHL_I: return "ARM64ISD::SQSHL_I";
681 case ARM64ISD::UQSHL_I: return "ARM64ISD::UQSHL_I";
682 case ARM64ISD::SRSHR_I: return "ARM64ISD::SRSHR_I";
683 case ARM64ISD::URSHR_I: return "ARM64ISD::URSHR_I";
684 case ARM64ISD::SQSHLU_I: return "ARM64ISD::SQSHLU_I";
685 case ARM64ISD::WrapperLarge: return "ARM64ISD::WrapperLarge";
689 static void getExclusiveOperation(unsigned Size, AtomicOrdering Ord,
690 unsigned &LdrOpc, unsigned &StrOpc) {
691 static unsigned LoadBares[] = { ARM64::LDXRB, ARM64::LDXRH, ARM64::LDXRW,
692 ARM64::LDXRX, ARM64::LDXPX };
693 static unsigned LoadAcqs[] = { ARM64::LDAXRB, ARM64::LDAXRH, ARM64::LDAXRW,
694 ARM64::LDAXRX, ARM64::LDAXPX };
695 static unsigned StoreBares[] = { ARM64::STXRB, ARM64::STXRH, ARM64::STXRW,
696 ARM64::STXRX, ARM64::STXPX };
697 static unsigned StoreRels[] = { ARM64::STLXRB, ARM64::STLXRH, ARM64::STLXRW,
698 ARM64::STLXRX, ARM64::STLXPX };
700 unsigned *LoadOps, *StoreOps;
701 if (Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent)
706 if (Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent)
707 StoreOps = StoreRels;
709 StoreOps = StoreBares;
711 assert(isPowerOf2_32(Size) && Size <= 16 &&
712 "unsupported size for atomic binary op!");
714 LdrOpc = LoadOps[Log2_32(Size)];
715 StrOpc = StoreOps[Log2_32(Size)];
718 MachineBasicBlock *ARM64TargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
719 MachineBasicBlock *BB,
720 unsigned Size) const {
721 unsigned dest = MI->getOperand(0).getReg();
722 unsigned ptr = MI->getOperand(1).getReg();
723 unsigned oldval = MI->getOperand(2).getReg();
724 unsigned newval = MI->getOperand(3).getReg();
725 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(4).getImm());
726 unsigned scratch = BB->getParent()->getRegInfo().createVirtualRegister(
727 &ARM64::GPR32RegClass);
728 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
729 DebugLoc dl = MI->getDebugLoc();
731 // FIXME: We currently always generate a seq_cst operation; we should
732 // be able to relax this in some cases.
733 unsigned ldrOpc, strOpc;
734 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
736 MachineFunction *MF = BB->getParent();
737 const BasicBlock *LLVM_BB = BB->getBasicBlock();
738 MachineFunction::iterator It = BB;
739 ++It; // insert the new blocks after the current block
741 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
742 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
743 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
744 MF->insert(It, loop1MBB);
745 MF->insert(It, loop2MBB);
746 MF->insert(It, exitMBB);
748 // Transfer the remainder of BB and its successor edges to exitMBB.
749 exitMBB->splice(exitMBB->begin(), BB,
750 std::next(MachineBasicBlock::iterator(MI)), BB->end());
751 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
755 // fallthrough --> loop1MBB
756 BB->addSuccessor(loop1MBB);
763 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
764 BuildMI(BB, dl, TII->get(Size == 8 ? ARM64::SUBSXrr : ARM64::SUBSWrr))
765 .addReg(Size == 8 ? ARM64::XZR : ARM64::WZR, RegState::Define)
768 BuildMI(BB, dl, TII->get(ARM64::Bcc)).addImm(ARM64CC::NE).addMBB(exitMBB);
769 BB->addSuccessor(loop2MBB);
770 BB->addSuccessor(exitMBB);
773 // strex scratch, newval, [ptr]
777 BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
778 BuildMI(BB, dl, TII->get(ARM64::CBNZW)).addReg(scratch).addMBB(loop1MBB);
779 BB->addSuccessor(loop1MBB);
780 BB->addSuccessor(exitMBB);
786 MI->eraseFromParent(); // The instruction is gone now.
792 ARM64TargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
793 unsigned Size, unsigned BinOpcode) const {
794 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
795 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
797 const BasicBlock *LLVM_BB = BB->getBasicBlock();
798 MachineFunction *MF = BB->getParent();
799 MachineFunction::iterator It = BB;
802 unsigned dest = MI->getOperand(0).getReg();
803 unsigned ptr = MI->getOperand(1).getReg();
804 unsigned incr = MI->getOperand(2).getReg();
805 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
806 DebugLoc dl = MI->getDebugLoc();
808 unsigned ldrOpc, strOpc;
809 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
811 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
812 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
813 MF->insert(It, loopMBB);
814 MF->insert(It, exitMBB);
816 // Transfer the remainder of BB and its successor edges to exitMBB.
817 exitMBB->splice(exitMBB->begin(), BB,
818 std::next(MachineBasicBlock::iterator(MI)), BB->end());
819 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
821 MachineRegisterInfo &RegInfo = MF->getRegInfo();
822 unsigned scratch = RegInfo.createVirtualRegister(&ARM64::GPR32RegClass);
826 : RegInfo.createVirtualRegister(Size == 8 ? &ARM64::GPR64RegClass
827 : &ARM64::GPR32RegClass);
831 // fallthrough --> loopMBB
832 BB->addSuccessor(loopMBB);
836 // <binop> scratch2, dest, incr
837 // stxr scratch, scratch2, ptr
838 // cbnz scratch, loopMBB
839 // fallthrough --> exitMBB
841 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
843 // operand order needs to go the other way for NAND
844 if (BinOpcode == ARM64::BICWrr || BinOpcode == ARM64::BICXrr)
845 BuildMI(BB, dl, TII->get(BinOpcode), scratch2).addReg(incr).addReg(dest);
847 BuildMI(BB, dl, TII->get(BinOpcode), scratch2).addReg(dest).addReg(incr);
850 BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
851 BuildMI(BB, dl, TII->get(ARM64::CBNZW)).addReg(scratch).addMBB(loopMBB);
853 BB->addSuccessor(loopMBB);
854 BB->addSuccessor(exitMBB);
860 MI->eraseFromParent(); // The instruction is gone now.
865 MachineBasicBlock *ARM64TargetLowering::EmitAtomicBinary128(
866 MachineInstr *MI, MachineBasicBlock *BB, unsigned BinOpcodeLo,
867 unsigned BinOpcodeHi) const {
868 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
869 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
871 const BasicBlock *LLVM_BB = BB->getBasicBlock();
872 MachineFunction *MF = BB->getParent();
873 MachineFunction::iterator It = BB;
876 unsigned DestLo = MI->getOperand(0).getReg();
877 unsigned DestHi = MI->getOperand(1).getReg();
878 unsigned Ptr = MI->getOperand(2).getReg();
879 unsigned IncrLo = MI->getOperand(3).getReg();
880 unsigned IncrHi = MI->getOperand(4).getReg();
881 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(5).getImm());
882 DebugLoc DL = MI->getDebugLoc();
884 unsigned LdrOpc, StrOpc;
885 getExclusiveOperation(16, Ord, LdrOpc, StrOpc);
887 MachineBasicBlock *LoopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
888 MachineBasicBlock *ExitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
889 MF->insert(It, LoopMBB);
890 MF->insert(It, ExitMBB);
892 // Transfer the remainder of BB and its successor edges to exitMBB.
893 ExitMBB->splice(ExitMBB->begin(), BB,
894 std::next(MachineBasicBlock::iterator(MI)), BB->end());
895 ExitMBB->transferSuccessorsAndUpdatePHIs(BB);
897 MachineRegisterInfo &RegInfo = MF->getRegInfo();
898 unsigned ScratchRes = RegInfo.createVirtualRegister(&ARM64::GPR32RegClass);
899 unsigned ScratchLo = IncrLo, ScratchHi = IncrHi;
901 assert(BinOpcodeHi && "Expect neither or both opcodes to be defined");
902 ScratchLo = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
903 ScratchHi = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
908 // fallthrough --> LoopMBB
909 BB->addSuccessor(LoopMBB);
912 // ldxp DestLo, DestHi, Ptr
913 // <binoplo> ScratchLo, DestLo, IncrLo
914 // <binophi> ScratchHi, DestHi, IncrHi
915 // stxp ScratchRes, ScratchLo, ScratchHi, ptr
916 // cbnz ScratchRes, LoopMBB
917 // fallthrough --> ExitMBB
919 BuildMI(BB, DL, TII->get(LdrOpc), DestLo)
920 .addReg(DestHi, RegState::Define)
923 // operand order needs to go the other way for NAND
924 if (BinOpcodeLo == ARM64::BICXrr) {
925 std::swap(IncrLo, DestLo);
926 std::swap(IncrHi, DestHi);
929 BuildMI(BB, DL, TII->get(BinOpcodeLo), ScratchLo).addReg(DestLo).addReg(
931 BuildMI(BB, DL, TII->get(BinOpcodeHi), ScratchHi).addReg(DestHi).addReg(
935 BuildMI(BB, DL, TII->get(StrOpc), ScratchRes)
939 BuildMI(BB, DL, TII->get(ARM64::CBNZW)).addReg(ScratchRes).addMBB(LoopMBB);
941 BB->addSuccessor(LoopMBB);
942 BB->addSuccessor(ExitMBB);
948 MI->eraseFromParent(); // The instruction is gone now.
954 ARM64TargetLowering::EmitAtomicCmpSwap128(MachineInstr *MI,
955 MachineBasicBlock *BB) const {
956 unsigned DestLo = MI->getOperand(0).getReg();
957 unsigned DestHi = MI->getOperand(1).getReg();
958 unsigned Ptr = MI->getOperand(2).getReg();
959 unsigned OldValLo = MI->getOperand(3).getReg();
960 unsigned OldValHi = MI->getOperand(4).getReg();
961 unsigned NewValLo = MI->getOperand(5).getReg();
962 unsigned NewValHi = MI->getOperand(6).getReg();
963 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(7).getImm());
964 unsigned ScratchRes = BB->getParent()->getRegInfo().createVirtualRegister(
965 &ARM64::GPR32RegClass);
966 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
967 DebugLoc DL = MI->getDebugLoc();
969 unsigned LdrOpc, StrOpc;
970 getExclusiveOperation(16, Ord, LdrOpc, StrOpc);
972 MachineFunction *MF = BB->getParent();
973 const BasicBlock *LLVM_BB = BB->getBasicBlock();
974 MachineFunction::iterator It = BB;
975 ++It; // insert the new blocks after the current block
977 MachineBasicBlock *Loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
978 MachineBasicBlock *Loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
979 MachineBasicBlock *ExitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
980 MF->insert(It, Loop1MBB);
981 MF->insert(It, Loop2MBB);
982 MF->insert(It, ExitMBB);
984 // Transfer the remainder of BB and its successor edges to exitMBB.
985 ExitMBB->splice(ExitMBB->begin(), BB,
986 std::next(MachineBasicBlock::iterator(MI)), BB->end());
987 ExitMBB->transferSuccessorsAndUpdatePHIs(BB);
991 // fallthrough --> Loop1MBB
992 BB->addSuccessor(Loop1MBB);
995 // ldxp DestLo, DestHi, [Ptr]
996 // cmp DestLo, OldValLo
997 // sbc xzr, DestHi, OldValHi
1000 BuildMI(BB, DL, TII->get(LdrOpc), DestLo)
1001 .addReg(DestHi, RegState::Define)
1003 BuildMI(BB, DL, TII->get(ARM64::SUBSXrr), ARM64::XZR).addReg(DestLo).addReg(
1005 BuildMI(BB, DL, TII->get(ARM64::SBCXr), ARM64::XZR).addReg(DestHi).addReg(
1008 BuildMI(BB, DL, TII->get(ARM64::Bcc)).addImm(ARM64CC::NE).addMBB(ExitMBB);
1009 BB->addSuccessor(Loop2MBB);
1010 BB->addSuccessor(ExitMBB);
1013 // stxp ScratchRes, NewValLo, NewValHi, [Ptr]
1014 // cbnz ScratchRes, Loop1MBB
1016 BuildMI(BB, DL, TII->get(StrOpc), ScratchRes)
1020 BuildMI(BB, DL, TII->get(ARM64::CBNZW)).addReg(ScratchRes).addMBB(Loop1MBB);
1021 BB->addSuccessor(Loop1MBB);
1022 BB->addSuccessor(ExitMBB);
1028 MI->eraseFromParent(); // The instruction is gone now.
1033 MachineBasicBlock *ARM64TargetLowering::EmitAtomicMinMax128(
1034 MachineInstr *MI, MachineBasicBlock *BB, unsigned CondCode) const {
1035 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
1036 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1038 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1039 MachineFunction *MF = BB->getParent();
1040 MachineFunction::iterator It = BB;
1043 unsigned DestLo = MI->getOperand(0).getReg();
1044 unsigned DestHi = MI->getOperand(1).getReg();
1045 unsigned Ptr = MI->getOperand(2).getReg();
1046 unsigned IncrLo = MI->getOperand(3).getReg();
1047 unsigned IncrHi = MI->getOperand(4).getReg();
1048 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(5).getImm());
1049 DebugLoc DL = MI->getDebugLoc();
1051 unsigned LdrOpc, StrOpc;
1052 getExclusiveOperation(16, Ord, LdrOpc, StrOpc);
1054 MachineBasicBlock *LoopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1055 MachineBasicBlock *ExitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1056 MF->insert(It, LoopMBB);
1057 MF->insert(It, ExitMBB);
1059 // Transfer the remainder of BB and its successor edges to exitMBB.
1060 ExitMBB->splice(ExitMBB->begin(), BB,
1061 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1062 ExitMBB->transferSuccessorsAndUpdatePHIs(BB);
1064 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1065 unsigned ScratchRes = RegInfo.createVirtualRegister(&ARM64::GPR32RegClass);
1066 unsigned ScratchLo = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
1067 unsigned ScratchHi = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
1071 // fallthrough --> LoopMBB
1072 BB->addSuccessor(LoopMBB);
1075 // ldxp DestLo, DestHi, Ptr
1076 // cmp ScratchLo, DestLo, IncrLo
1077 // sbc xzr, ScratchHi, DestHi, IncrHi
1078 // csel ScratchLo, DestLo, IncrLo, <cmp-op>
1079 // csel ScratchHi, DestHi, IncrHi, <cmp-op>
1080 // stxp ScratchRes, ScratchLo, ScratchHi, ptr
1081 // cbnz ScratchRes, LoopMBB
1082 // fallthrough --> ExitMBB
1084 BuildMI(BB, DL, TII->get(LdrOpc), DestLo)
1085 .addReg(DestHi, RegState::Define)
1088 BuildMI(BB, DL, TII->get(ARM64::SUBSXrr), ARM64::XZR).addReg(DestLo).addReg(
1090 BuildMI(BB, DL, TII->get(ARM64::SBCXr), ARM64::XZR).addReg(DestHi).addReg(
1093 BuildMI(BB, DL, TII->get(ARM64::CSELXr), ScratchLo)
1097 BuildMI(BB, DL, TII->get(ARM64::CSELXr), ScratchHi)
1102 BuildMI(BB, DL, TII->get(StrOpc), ScratchRes)
1106 BuildMI(BB, DL, TII->get(ARM64::CBNZW)).addReg(ScratchRes).addMBB(LoopMBB);
1108 BB->addSuccessor(LoopMBB);
1109 BB->addSuccessor(ExitMBB);
1115 MI->eraseFromParent(); // The instruction is gone now.
1121 ARM64TargetLowering::EmitF128CSEL(MachineInstr *MI,
1122 MachineBasicBlock *MBB) const {
1123 // We materialise the F128CSEL pseudo-instruction as some control flow and a
1127 // [... previous instrs leading to comparison ...]
1133 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
1135 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1136 MachineFunction *MF = MBB->getParent();
1137 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
1138 DebugLoc DL = MI->getDebugLoc();
1139 MachineFunction::iterator It = MBB;
1142 unsigned DestReg = MI->getOperand(0).getReg();
1143 unsigned IfTrueReg = MI->getOperand(1).getReg();
1144 unsigned IfFalseReg = MI->getOperand(2).getReg();
1145 unsigned CondCode = MI->getOperand(3).getImm();
1146 bool CPSRKilled = MI->getOperand(4).isKill();
1148 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
1149 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
1150 MF->insert(It, TrueBB);
1151 MF->insert(It, EndBB);
1153 // Transfer rest of current basic-block to EndBB
1154 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
1156 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
1158 BuildMI(MBB, DL, TII->get(ARM64::Bcc)).addImm(CondCode).addMBB(TrueBB);
1159 BuildMI(MBB, DL, TII->get(ARM64::B)).addMBB(EndBB);
1160 MBB->addSuccessor(TrueBB);
1161 MBB->addSuccessor(EndBB);
1163 // TrueBB falls through to the end.
1164 TrueBB->addSuccessor(EndBB);
1167 TrueBB->addLiveIn(ARM64::CPSR);
1168 EndBB->addLiveIn(ARM64::CPSR);
1171 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(ARM64::PHI), DestReg)
1177 MI->eraseFromParent();
1182 ARM64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
1183 MachineBasicBlock *BB) const {
1184 switch (MI->getOpcode()) {
1189 assert(0 && "Unexpected instruction for custom inserter!");
1192 case ARM64::ATOMIC_LOAD_ADD_I8:
1193 return EmitAtomicBinary(MI, BB, 1, ARM64::ADDWrr);
1194 case ARM64::ATOMIC_LOAD_ADD_I16:
1195 return EmitAtomicBinary(MI, BB, 2, ARM64::ADDWrr);
1196 case ARM64::ATOMIC_LOAD_ADD_I32:
1197 return EmitAtomicBinary(MI, BB, 4, ARM64::ADDWrr);
1198 case ARM64::ATOMIC_LOAD_ADD_I64:
1199 return EmitAtomicBinary(MI, BB, 8, ARM64::ADDXrr);
1200 case ARM64::ATOMIC_LOAD_ADD_I128:
1201 return EmitAtomicBinary128(MI, BB, ARM64::ADDSXrr, ARM64::ADCXr);
1203 case ARM64::ATOMIC_LOAD_AND_I8:
1204 return EmitAtomicBinary(MI, BB, 1, ARM64::ANDWrr);
1205 case ARM64::ATOMIC_LOAD_AND_I16:
1206 return EmitAtomicBinary(MI, BB, 2, ARM64::ANDWrr);
1207 case ARM64::ATOMIC_LOAD_AND_I32:
1208 return EmitAtomicBinary(MI, BB, 4, ARM64::ANDWrr);
1209 case ARM64::ATOMIC_LOAD_AND_I64:
1210 return EmitAtomicBinary(MI, BB, 8, ARM64::ANDXrr);
1211 case ARM64::ATOMIC_LOAD_AND_I128:
1212 return EmitAtomicBinary128(MI, BB, ARM64::ANDXrr, ARM64::ANDXrr);
1214 case ARM64::ATOMIC_LOAD_OR_I8:
1215 return EmitAtomicBinary(MI, BB, 1, ARM64::ORRWrr);
1216 case ARM64::ATOMIC_LOAD_OR_I16:
1217 return EmitAtomicBinary(MI, BB, 2, ARM64::ORRWrr);
1218 case ARM64::ATOMIC_LOAD_OR_I32:
1219 return EmitAtomicBinary(MI, BB, 4, ARM64::ORRWrr);
1220 case ARM64::ATOMIC_LOAD_OR_I64:
1221 return EmitAtomicBinary(MI, BB, 8, ARM64::ORRXrr);
1222 case ARM64::ATOMIC_LOAD_OR_I128:
1223 return EmitAtomicBinary128(MI, BB, ARM64::ORRXrr, ARM64::ORRXrr);
1225 case ARM64::ATOMIC_LOAD_XOR_I8:
1226 return EmitAtomicBinary(MI, BB, 1, ARM64::EORWrr);
1227 case ARM64::ATOMIC_LOAD_XOR_I16:
1228 return EmitAtomicBinary(MI, BB, 2, ARM64::EORWrr);
1229 case ARM64::ATOMIC_LOAD_XOR_I32:
1230 return EmitAtomicBinary(MI, BB, 4, ARM64::EORWrr);
1231 case ARM64::ATOMIC_LOAD_XOR_I64:
1232 return EmitAtomicBinary(MI, BB, 8, ARM64::EORXrr);
1233 case ARM64::ATOMIC_LOAD_XOR_I128:
1234 return EmitAtomicBinary128(MI, BB, ARM64::EORXrr, ARM64::EORXrr);
1236 case ARM64::ATOMIC_LOAD_NAND_I8:
1237 return EmitAtomicBinary(MI, BB, 1, ARM64::BICWrr);
1238 case ARM64::ATOMIC_LOAD_NAND_I16:
1239 return EmitAtomicBinary(MI, BB, 2, ARM64::BICWrr);
1240 case ARM64::ATOMIC_LOAD_NAND_I32:
1241 return EmitAtomicBinary(MI, BB, 4, ARM64::BICWrr);
1242 case ARM64::ATOMIC_LOAD_NAND_I64:
1243 return EmitAtomicBinary(MI, BB, 8, ARM64::BICXrr);
1244 case ARM64::ATOMIC_LOAD_NAND_I128:
1245 return EmitAtomicBinary128(MI, BB, ARM64::BICXrr, ARM64::BICXrr);
1247 case ARM64::ATOMIC_LOAD_SUB_I8:
1248 return EmitAtomicBinary(MI, BB, 1, ARM64::SUBWrr);
1249 case ARM64::ATOMIC_LOAD_SUB_I16:
1250 return EmitAtomicBinary(MI, BB, 2, ARM64::SUBWrr);
1251 case ARM64::ATOMIC_LOAD_SUB_I32:
1252 return EmitAtomicBinary(MI, BB, 4, ARM64::SUBWrr);
1253 case ARM64::ATOMIC_LOAD_SUB_I64:
1254 return EmitAtomicBinary(MI, BB, 8, ARM64::SUBXrr);
1255 case ARM64::ATOMIC_LOAD_SUB_I128:
1256 return EmitAtomicBinary128(MI, BB, ARM64::SUBSXrr, ARM64::SBCXr);
1258 case ARM64::ATOMIC_LOAD_MIN_I128:
1259 return EmitAtomicMinMax128(MI, BB, ARM64CC::LT);
1261 case ARM64::ATOMIC_LOAD_MAX_I128:
1262 return EmitAtomicMinMax128(MI, BB, ARM64CC::GT);
1264 case ARM64::ATOMIC_LOAD_UMIN_I128:
1265 return EmitAtomicMinMax128(MI, BB, ARM64CC::CC);
1267 case ARM64::ATOMIC_LOAD_UMAX_I128:
1268 return EmitAtomicMinMax128(MI, BB, ARM64CC::HI);
1270 case ARM64::ATOMIC_SWAP_I8:
1271 return EmitAtomicBinary(MI, BB, 1, 0);
1272 case ARM64::ATOMIC_SWAP_I16:
1273 return EmitAtomicBinary(MI, BB, 2, 0);
1274 case ARM64::ATOMIC_SWAP_I32:
1275 return EmitAtomicBinary(MI, BB, 4, 0);
1276 case ARM64::ATOMIC_SWAP_I64:
1277 return EmitAtomicBinary(MI, BB, 8, 0);
1278 case ARM64::ATOMIC_SWAP_I128:
1279 return EmitAtomicBinary128(MI, BB, 0, 0);
1281 case ARM64::ATOMIC_CMP_SWAP_I8:
1282 return EmitAtomicCmpSwap(MI, BB, 1);
1283 case ARM64::ATOMIC_CMP_SWAP_I16:
1284 return EmitAtomicCmpSwap(MI, BB, 2);
1285 case ARM64::ATOMIC_CMP_SWAP_I32:
1286 return EmitAtomicCmpSwap(MI, BB, 4);
1287 case ARM64::ATOMIC_CMP_SWAP_I64:
1288 return EmitAtomicCmpSwap(MI, BB, 8);
1289 case ARM64::ATOMIC_CMP_SWAP_I128:
1290 return EmitAtomicCmpSwap128(MI, BB);
1292 case ARM64::F128CSEL:
1293 return EmitF128CSEL(MI, BB);
1295 case TargetOpcode::STACKMAP:
1296 case TargetOpcode::PATCHPOINT:
1297 return emitPatchPoint(MI, BB);
1299 llvm_unreachable("Unexpected instruction for custom inserter!");
1302 //===----------------------------------------------------------------------===//
1303 // ARM64 Lowering private implementation.
1304 //===----------------------------------------------------------------------===//
1306 //===----------------------------------------------------------------------===//
1308 //===----------------------------------------------------------------------===//
1310 /// changeIntCCToARM64CC - Convert a DAG integer condition code to an ARM64 CC
1311 static ARM64CC::CondCode changeIntCCToARM64CC(ISD::CondCode CC) {
1314 llvm_unreachable("Unknown condition code!");
1338 /// changeFPCCToARM64CC - Convert a DAG fp condition code to an ARM64 CC.
1339 static void changeFPCCToARM64CC(ISD::CondCode CC, ARM64CC::CondCode &CondCode,
1340 ARM64CC::CondCode &CondCode2) {
1341 CondCode2 = ARM64CC::AL;
1344 llvm_unreachable("Unknown FP condition!");
1347 CondCode = ARM64CC::EQ;
1351 CondCode = ARM64CC::GT;
1355 CondCode = ARM64CC::GE;
1358 CondCode = ARM64CC::MI;
1361 CondCode = ARM64CC::LS;
1364 CondCode = ARM64CC::MI;
1365 CondCode2 = ARM64CC::GT;
1368 CondCode = ARM64CC::VC;
1371 CondCode = ARM64CC::VS;
1374 CondCode = ARM64CC::EQ;
1375 CondCode2 = ARM64CC::VS;
1378 CondCode = ARM64CC::HI;
1381 CondCode = ARM64CC::PL;
1385 CondCode = ARM64CC::LT;
1389 CondCode = ARM64CC::LE;
1393 CondCode = ARM64CC::NE;
1398 static bool isLegalArithImmed(uint64_t C) {
1399 // Matches ARM64DAGToDAGISel::SelectArithImmed().
1400 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
1403 static SDValue emitComparison(SDValue LHS, SDValue RHS, SDLoc dl,
1404 SelectionDAG &DAG) {
1405 EVT VT = LHS.getValueType();
1407 if (VT.isFloatingPoint())
1408 return DAG.getNode(ARM64ISD::FCMP, dl, VT, LHS, RHS);
1410 // The CMP instruction is just an alias for SUBS, and representing it as
1411 // SUBS means that it's possible to get CSE with subtract operations.
1412 // A later phase can perform the optimization of setting the destination
1413 // register to WZR/XZR if it ends up being unused.
1414 return DAG.getNode(ARM64ISD::SUBS, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS)
1418 static SDValue getARM64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1419 SDValue &ARM64cc, SelectionDAG &DAG, SDLoc dl) {
1420 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1421 EVT VT = RHS.getValueType();
1422 uint64_t C = RHSC->getZExtValue();
1423 if (!isLegalArithImmed(C)) {
1424 // Constant does not fit, try adjusting it by one?
1430 if ((VT == MVT::i32 && C != 0x80000000 &&
1431 isLegalArithImmed((uint32_t)(C - 1))) ||
1432 (VT == MVT::i64 && C != 0x80000000ULL &&
1433 isLegalArithImmed(C - 1ULL))) {
1434 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1435 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1436 RHS = DAG.getConstant(C, VT);
1441 if ((VT == MVT::i32 && C != 0 &&
1442 isLegalArithImmed((uint32_t)(C - 1))) ||
1443 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1444 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1445 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1446 RHS = DAG.getConstant(C, VT);
1451 if ((VT == MVT::i32 && C != 0x7fffffff &&
1452 isLegalArithImmed((uint32_t)(C + 1))) ||
1453 (VT == MVT::i64 && C != 0x7ffffffffffffffULL &&
1454 isLegalArithImmed(C + 1ULL))) {
1455 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1456 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1457 RHS = DAG.getConstant(C, VT);
1462 if ((VT == MVT::i32 && C != 0xffffffff &&
1463 isLegalArithImmed((uint32_t)(C + 1))) ||
1464 (VT == MVT::i64 && C != 0xfffffffffffffffULL &&
1465 isLegalArithImmed(C + 1ULL))) {
1466 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1467 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1468 RHS = DAG.getConstant(C, VT);
1475 SDValue Cmp = emitComparison(LHS, RHS, dl, DAG);
1476 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
1477 ARM64cc = DAG.getConstant(ARM64CC, MVT::i32);
1481 static std::pair<SDValue, SDValue>
1482 getARM64XALUOOp(ARM64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1483 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1484 "Unsupported value type");
1485 SDValue Value, Overflow;
1487 SDValue LHS = Op.getOperand(0);
1488 SDValue RHS = Op.getOperand(1);
1490 switch (Op.getOpcode()) {
1492 llvm_unreachable("Unknown overflow instruction!");
1494 Opc = ARM64ISD::ADDS;
1498 Opc = ARM64ISD::ADDS;
1502 Opc = ARM64ISD::SUBS;
1506 Opc = ARM64ISD::SUBS;
1509 // Multiply needs a little bit extra work.
1513 bool IsSigned = (Op.getOpcode() == ISD::SMULO) ? true : false;
1514 if (Op.getValueType() == MVT::i32) {
1515 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1516 // For a 32 bit multiply with overflow check we want the instruction
1517 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1518 // need to generate the following pattern:
1519 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1520 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1521 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1522 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1523 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1524 DAG.getConstant(0, MVT::i64));
1525 // On ARM64 the upper 32 bits are always zero extended for a 32 bit
1526 // operation. We need to clear out the upper 32 bits, because we used a
1527 // widening multiply that wrote all 64 bits. In the end this should be a
1529 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1531 // The signed overflow check requires more than just a simple check for
1532 // any bit set in the upper 32 bits of the result. These bits could be
1533 // just the sign bits of a negative number. To perform the overflow
1534 // check we have to arithmetic shift right the 32nd bit of the result by
1535 // 31 bits. Then we compare the result to the upper 32 bits.
1536 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1537 DAG.getConstant(32, MVT::i32));
1538 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1539 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1540 DAG.getConstant(31, MVT::i32));
1541 // It is important that LowerBits is last, otherwise the arithmetic
1542 // shift will not be folded into the compare (SUBS).
1543 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1544 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1547 // The overflow check for unsigned multiply is easy. We only need to
1548 // check if any of the upper 32 bits are set. This can be done with a
1549 // CMP (shifted register). For that we need to generate the following
1551 // (i64 ARM64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1552 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1553 DAG.getConstant(32, MVT::i32));
1554 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1556 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1557 UpperBits).getValue(1);
1561 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1562 // For the 64 bit multiply
1563 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1565 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1566 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1567 DAG.getConstant(63, MVT::i32));
1568 // It is important that LowerBits is last, otherwise the arithmetic
1569 // shift will not be folded into the compare (SUBS).
1570 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1571 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1574 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1575 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1577 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1578 UpperBits).getValue(1);
1585 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1587 // Emit the ARM64 operation with overflow check.
1588 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1589 Overflow = Value.getValue(1);
1591 return std::make_pair(Value, Overflow);
1594 SDValue ARM64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1595 RTLIB::Libcall Call) const {
1596 SmallVector<SDValue, 2> Ops;
1597 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1598 Ops.push_back(Op.getOperand(i));
1600 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1604 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1605 SDValue Sel = Op.getOperand(0);
1606 SDValue Other = Op.getOperand(1);
1608 // If neither operand is a SELECT_CC, give up.
1609 if (Sel.getOpcode() != ISD::SELECT_CC)
1610 std::swap(Sel, Other);
1611 if (Sel.getOpcode() != ISD::SELECT_CC)
1614 // The folding we want to perform is:
1615 // (xor x, (select_cc a, b, cc, 0, -1) )
1617 // (csel x, (xor x, -1), cc ...)
1619 // The latter will get matched to a CSINV instruction.
1621 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1622 SDValue LHS = Sel.getOperand(0);
1623 SDValue RHS = Sel.getOperand(1);
1624 SDValue TVal = Sel.getOperand(2);
1625 SDValue FVal = Sel.getOperand(3);
1628 // FIXME: This could be generalized to non-integer comparisons.
1629 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1632 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1633 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1635 // The the values aren't constants, this isn't the pattern we're looking for.
1636 if (!CFVal || !CTVal)
1639 // We can commute the SELECT_CC by inverting the condition. This
1640 // might be needed to make this fit into a CSINV pattern.
1641 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1642 std::swap(TVal, FVal);
1643 std::swap(CTVal, CFVal);
1644 CC = ISD::getSetCCInverse(CC, true);
1647 // If the constants line up, perform the transform!
1648 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1650 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1653 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1654 DAG.getConstant(-1ULL, Other.getValueType()));
1656 return DAG.getNode(ARM64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1663 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1664 EVT VT = Op.getValueType();
1666 // Let legalize expand this if it isn't a legal type yet.
1667 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1670 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1673 bool ExtraOp = false;
1674 switch (Op.getOpcode()) {
1676 assert(0 && "Invalid code");
1678 Opc = ARM64ISD::ADDS;
1681 Opc = ARM64ISD::SUBS;
1684 Opc = ARM64ISD::ADCS;
1688 Opc = ARM64ISD::SBCS;
1694 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1695 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1699 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1700 // Let legalize expand this if it isn't a legal type yet.
1701 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1704 ARM64CC::CondCode CC;
1705 // The actual operation that sets the overflow or carry flag.
1706 SDValue Value, Overflow;
1707 std::tie(Value, Overflow) = getARM64XALUOOp(CC, Op, DAG);
1709 // We use 0 and 1 as false and true values.
1710 SDValue TVal = DAG.getConstant(1, MVT::i32);
1711 SDValue FVal = DAG.getConstant(0, MVT::i32);
1713 // We use an inverted condition, because the conditional select is inverted
1714 // too. This will allow it to be selected to a single instruction:
1715 // CSINC Wd, WZR, WZR, invert(cond).
1716 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), MVT::i32);
1717 Overflow = DAG.getNode(ARM64ISD::CSEL, SDLoc(Op), MVT::i32, FVal, TVal, CCVal,
1720 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1721 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
1724 // Prefetch operands are:
1725 // 1: Address to prefetch
1727 // 3: int locality (0 = no locality ... 3 = extreme locality)
1728 // 4: bool isDataCache
1729 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1731 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1732 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1733 // The data thing is not used.
1734 // unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1736 bool IsStream = !Locality;
1737 // When the locality number is set
1739 // The front-end should have filtered out the out-of-range values
1740 assert(Locality <= 3 && "Prefetch locality out-of-range");
1741 // The locality degree is the opposite of the cache speed.
1742 // Put the number the other way around.
1743 // The encoding starts at 0 for level 1
1744 Locality = 3 - Locality;
1747 // built the mask value encoding the expected behavior.
1748 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1749 (Locality << 1) | // Cache level bits
1750 (unsigned)IsStream; // Stream bit
1751 return DAG.getNode(ARM64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1752 DAG.getConstant(PrfOp, MVT::i32), Op.getOperand(1));
1755 SDValue ARM64TargetLowering::LowerFP_EXTEND(SDValue Op,
1756 SelectionDAG &DAG) const {
1757 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1760 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1762 return LowerF128Call(Op, DAG, LC);
1765 SDValue ARM64TargetLowering::LowerFP_ROUND(SDValue Op,
1766 SelectionDAG &DAG) const {
1767 if (Op.getOperand(0).getValueType() != MVT::f128) {
1768 // It's legal except when f128 is involved
1773 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1775 // FP_ROUND node has a second operand indicating whether it is known to be
1776 // precise. That doesn't take part in the LibCall so we can't directly use
1778 SDValue SrcVal = Op.getOperand(0);
1779 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1780 /*isSigned*/ false, SDLoc(Op)).first;
1783 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1784 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1785 // Any additional optimization in this function should be recorded
1786 // in the cost tables.
1787 EVT InVT = Op.getOperand(0).getValueType();
1788 EVT VT = Op.getValueType();
1790 // FP_TO_XINT conversion from the same type are legal.
1791 if (VT.getSizeInBits() == InVT.getSizeInBits())
1794 if (InVT == MVT::v2f64) {
1796 SDValue Cv = DAG.getNode(Op.getOpcode(), dl, MVT::v2i64, Op.getOperand(0));
1797 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1800 // Type changing conversions are illegal.
1804 SDValue ARM64TargetLowering::LowerFP_TO_INT(SDValue Op,
1805 SelectionDAG &DAG) const {
1806 if (Op.getOperand(0).getValueType().isVector())
1807 return LowerVectorFP_TO_INT(Op, DAG);
1809 if (Op.getOperand(0).getValueType() != MVT::f128) {
1810 // It's legal except when f128 is involved
1815 if (Op.getOpcode() == ISD::FP_TO_SINT)
1816 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1818 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1820 SmallVector<SDValue, 2> Ops;
1821 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1822 Ops.push_back(Op.getOperand(i));
1824 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1828 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1829 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1830 // Any additional optimization in this function should be recorded
1831 // in the cost tables.
1832 EVT VT = Op.getValueType();
1834 SDValue In = Op.getOperand(0);
1835 EVT InVT = In.getValueType();
1837 // v2i32 to v2f32 is legal.
1838 if (VT == MVT::v2f32 && InVT == MVT::v2i32)
1841 // This function only handles v2f64 outputs.
1842 if (VT == MVT::v2f64) {
1843 // Extend the input argument to a v2i64 that we can feed into the
1844 // floating point conversion. Zero or sign extend based on whether
1845 // we're doing a signed or unsigned float conversion.
1847 Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
1848 assert(Op.getNumOperands() == 1 && "FP conversions take one argument");
1849 SDValue Promoted = DAG.getNode(Opc, dl, MVT::v2i64, Op.getOperand(0));
1850 return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), Promoted);
1853 // Scalarize v2i64 to v2f32 conversions.
1854 std::vector<SDValue> BuildVectorOps;
1855 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
1856 SDValue Sclr = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, In,
1857 DAG.getConstant(i, MVT::i64));
1858 Sclr = DAG.getNode(Op->getOpcode(), dl, MVT::f32, Sclr);
1859 BuildVectorOps.push_back(Sclr);
1862 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &BuildVectorOps[0],
1863 BuildVectorOps.size());
1866 SDValue ARM64TargetLowering::LowerINT_TO_FP(SDValue Op,
1867 SelectionDAG &DAG) const {
1868 if (Op.getValueType().isVector())
1869 return LowerVectorINT_TO_FP(Op, DAG);
1871 // i128 conversions are libcalls.
1872 if (Op.getOperand(0).getValueType() == MVT::i128)
1875 // Other conversions are legal, unless it's to the completely software-based
1877 if (Op.getValueType() != MVT::f128)
1881 if (Op.getOpcode() == ISD::SINT_TO_FP)
1882 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1884 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1886 return LowerF128Call(Op, DAG, LC);
1889 SDValue ARM64TargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
1890 // For iOS, we want to call an alternative entry point: __sincos_stret,
1891 // which returns the values in two S / D registers.
1893 SDValue Arg = Op.getOperand(0);
1894 EVT ArgVT = Arg.getValueType();
1895 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1902 Entry.isSExt = false;
1903 Entry.isZExt = false;
1904 Args.push_back(Entry);
1906 const char *LibcallName =
1907 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1908 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
1910 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
1911 TargetLowering::CallLoweringInfo CLI(
1912 DAG.getEntryNode(), RetTy, false, false, false, false, 0,
1913 CallingConv::Fast, /*isTaillCall=*/false,
1914 /*doesNotRet=*/false, /*isReturnValueUsed*/ true, Callee, Args, DAG, dl);
1915 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1916 return CallResult.first;
1919 SDValue ARM64TargetLowering::LowerOperation(SDValue Op,
1920 SelectionDAG &DAG) const {
1921 switch (Op.getOpcode()) {
1923 llvm_unreachable("unimplemented operand");
1925 case ISD::GlobalAddress:
1926 return LowerGlobalAddress(Op, DAG);
1927 case ISD::GlobalTLSAddress:
1928 return LowerGlobalTLSAddress(Op, DAG);
1930 return LowerSETCC(Op, DAG);
1932 return LowerBR_CC(Op, DAG);
1934 return LowerSELECT(Op, DAG);
1935 case ISD::SELECT_CC:
1936 return LowerSELECT_CC(Op, DAG);
1937 case ISD::JumpTable:
1938 return LowerJumpTable(Op, DAG);
1939 case ISD::ConstantPool:
1940 return LowerConstantPool(Op, DAG);
1941 case ISD::BlockAddress:
1942 return LowerBlockAddress(Op, DAG);
1944 return LowerVASTART(Op, DAG);
1946 return LowerVACOPY(Op, DAG);
1948 return LowerVAARG(Op, DAG);
1953 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
1960 return LowerXALUO(Op, DAG);
1962 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
1964 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
1966 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
1968 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
1970 return LowerFP_ROUND(Op, DAG);
1971 case ISD::FP_EXTEND:
1972 return LowerFP_EXTEND(Op, DAG);
1973 case ISD::FRAMEADDR:
1974 return LowerFRAMEADDR(Op, DAG);
1975 case ISD::RETURNADDR:
1976 return LowerRETURNADDR(Op, DAG);
1977 case ISD::INSERT_VECTOR_ELT:
1978 return LowerINSERT_VECTOR_ELT(Op, DAG);
1979 case ISD::EXTRACT_VECTOR_ELT:
1980 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
1981 case ISD::SCALAR_TO_VECTOR:
1982 return LowerSCALAR_TO_VECTOR(Op, DAG);
1983 case ISD::BUILD_VECTOR:
1984 return LowerBUILD_VECTOR(Op, DAG);
1985 case ISD::VECTOR_SHUFFLE:
1986 return LowerVECTOR_SHUFFLE(Op, DAG);
1987 case ISD::EXTRACT_SUBVECTOR:
1988 return LowerEXTRACT_SUBVECTOR(Op, DAG);
1992 return LowerVectorSRA_SRL_SHL(Op, DAG);
1993 case ISD::SHL_PARTS:
1994 return LowerShiftLeftParts(Op, DAG);
1995 case ISD::SRL_PARTS:
1996 case ISD::SRA_PARTS:
1997 return LowerShiftRightParts(Op, DAG);
1999 return LowerCTPOP(Op, DAG);
2000 case ISD::FCOPYSIGN:
2001 return LowerFCOPYSIGN(Op, DAG);
2003 return LowerVectorAND(Op, DAG);
2005 return LowerVectorOR(Op, DAG);
2007 return LowerXOR(Op, DAG);
2009 return LowerPREFETCH(Op, DAG);
2010 case ISD::SINT_TO_FP:
2011 case ISD::UINT_TO_FP:
2012 return LowerINT_TO_FP(Op, DAG);
2013 case ISD::FP_TO_SINT:
2014 case ISD::FP_TO_UINT:
2015 return LowerFP_TO_INT(Op, DAG);
2017 return LowerFSINCOS(Op, DAG);
2021 /// getFunctionAlignment - Return the Log2 alignment of this function.
2022 unsigned ARM64TargetLowering::getFunctionAlignment(const Function *F) const {
2026 //===----------------------------------------------------------------------===//
2027 // Calling Convention Implementation
2028 //===----------------------------------------------------------------------===//
2030 #include "ARM64GenCallingConv.inc"
2032 /// Selects the correct CCAssignFn for a the given CallingConvention
2034 CCAssignFn *ARM64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
2035 bool IsVarArg) const {
2038 llvm_unreachable("Unsupported calling convention.");
2039 case CallingConv::WebKit_JS:
2040 return CC_ARM64_WebKit_JS;
2041 case CallingConv::C:
2042 case CallingConv::Fast:
2043 if (!Subtarget->isTargetDarwin())
2044 return CC_ARM64_AAPCS;
2045 return IsVarArg ? CC_ARM64_DarwinPCS_VarArg : CC_ARM64_DarwinPCS;
2049 SDValue ARM64TargetLowering::LowerFormalArguments(
2050 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2051 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2052 SmallVectorImpl<SDValue> &InVals) const {
2053 MachineFunction &MF = DAG.getMachineFunction();
2054 MachineFrameInfo *MFI = MF.getFrameInfo();
2056 // Assign locations to all of the incoming arguments.
2057 SmallVector<CCValAssign, 16> ArgLocs;
2058 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2059 getTargetMachine(), ArgLocs, *DAG.getContext());
2061 // At this point, Ins[].VT may already be promoted to i32. To correctly
2062 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2063 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
2064 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
2065 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
2067 unsigned NumArgs = Ins.size();
2068 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2069 unsigned CurArgIdx = 0;
2070 for (unsigned i = 0; i != NumArgs; ++i) {
2071 MVT ValVT = Ins[i].VT;
2072 std::advance(CurOrigArg, Ins[i].OrigArgIndex - CurArgIdx);
2073 CurArgIdx = Ins[i].OrigArgIndex;
2075 // Get type of the original argument.
2076 EVT ActualVT = getValueType(CurOrigArg->getType(), /*AllowUnknown*/ true);
2077 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
2078 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2080 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2082 else if (ActualMVT == MVT::i16)
2085 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2087 AssignFn(i, ValVT, LocVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
2088 assert(!Res && "Call operand has unhandled type");
2092 SmallVector<SDValue, 16> ArgValues;
2093 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2094 CCValAssign &VA = ArgLocs[i];
2096 // Arguments stored in registers.
2097 if (VA.isRegLoc()) {
2098 EVT RegVT = VA.getLocVT();
2101 const TargetRegisterClass *RC;
2103 if (RegVT == MVT::i32)
2104 RC = &ARM64::GPR32RegClass;
2105 else if (RegVT == MVT::i64)
2106 RC = &ARM64::GPR64RegClass;
2107 else if (RegVT == MVT::f32)
2108 RC = &ARM64::FPR32RegClass;
2109 else if (RegVT == MVT::f64 || RegVT == MVT::v1i64 ||
2110 RegVT == MVT::v1f64 || RegVT == MVT::v2i32 ||
2111 RegVT == MVT::v4i16 || RegVT == MVT::v8i8)
2112 RC = &ARM64::FPR64RegClass;
2113 else if (RegVT == MVT::v2i64 || RegVT == MVT::v4i32 ||
2114 RegVT == MVT::v8i16 || RegVT == MVT::v16i8)
2115 RC = &ARM64::FPR128RegClass;
2117 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2119 // Transform the arguments in physical registers into virtual ones.
2120 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2121 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2123 // If this is an 8, 16 or 32-bit value, it is really passed promoted
2124 // to 64 bits. Insert an assert[sz]ext to capture this, then
2125 // truncate to the right size.
2126 switch (VA.getLocInfo()) {
2128 llvm_unreachable("Unknown loc info!");
2129 case CCValAssign::Full:
2131 case CCValAssign::BCvt:
2132 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
2134 case CCValAssign::SExt:
2135 ArgValue = DAG.getNode(ISD::AssertSext, DL, RegVT, ArgValue,
2136 DAG.getValueType(VA.getValVT()));
2137 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
2139 case CCValAssign::ZExt:
2140 ArgValue = DAG.getNode(ISD::AssertZext, DL, RegVT, ArgValue,
2141 DAG.getValueType(VA.getValVT()));
2142 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
2146 InVals.push_back(ArgValue);
2148 } else { // VA.isRegLoc()
2149 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
2150 unsigned ArgOffset = VA.getLocMemOffset();
2151 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
2152 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset, true);
2154 // Create load nodes to retrieve arguments from the stack.
2155 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2156 InVals.push_back(DAG.getLoad(VA.getValVT(), DL, Chain, FIN,
2157 MachinePointerInfo::getFixedStack(FI), false,
2164 if (!Subtarget->isTargetDarwin()) {
2165 // The AAPCS variadic function ABI is identical to the non-variadic
2166 // one. As a result there may be more arguments in registers and we should
2167 // save them for future reference.
2168 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
2171 ARM64FunctionInfo *AFI = MF.getInfo<ARM64FunctionInfo>();
2172 // This will point to the next argument passed via stack.
2173 unsigned StackOffset = CCInfo.getNextStackOffset();
2174 // We currently pass all varargs at 8-byte alignment.
2175 StackOffset = ((StackOffset + 7) & ~7);
2176 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
2182 void ARM64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
2183 SelectionDAG &DAG, SDLoc DL,
2184 SDValue &Chain) const {
2185 MachineFunction &MF = DAG.getMachineFunction();
2186 MachineFrameInfo *MFI = MF.getFrameInfo();
2187 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
2189 SmallVector<SDValue, 8> MemOps;
2191 static const uint16_t GPRArgRegs[] = { ARM64::X0, ARM64::X1, ARM64::X2,
2192 ARM64::X3, ARM64::X4, ARM64::X5,
2193 ARM64::X6, ARM64::X7 };
2194 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
2195 unsigned FirstVariadicGPR =
2196 CCInfo.getFirstUnallocated(GPRArgRegs, NumGPRArgRegs);
2198 static const uint16_t FPRArgRegs[] = { ARM64::Q0, ARM64::Q1, ARM64::Q2,
2199 ARM64::Q3, ARM64::Q4, ARM64::Q5,
2200 ARM64::Q6, ARM64::Q7 };
2201 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
2202 unsigned FirstVariadicFPR =
2203 CCInfo.getFirstUnallocated(FPRArgRegs, NumFPRArgRegs);
2205 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
2207 if (GPRSaveSize != 0) {
2208 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
2210 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
2212 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
2213 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &ARM64::GPR64RegClass);
2214 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
2216 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2217 MachinePointerInfo::getStack(i * 8), false, false, 0);
2218 MemOps.push_back(Store);
2219 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2220 DAG.getConstant(8, getPointerTy()));
2224 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
2226 if (FPRSaveSize != 0) {
2227 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
2229 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
2231 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
2232 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &ARM64::FPR128RegClass);
2233 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::v2i64);
2235 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2236 MachinePointerInfo::getStack(i * 16), false, false, 0);
2237 MemOps.push_back(Store);
2238 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2239 DAG.getConstant(16, getPointerTy()));
2243 FuncInfo->setVarArgsGPRIndex(GPRIdx);
2244 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
2245 FuncInfo->setVarArgsFPRIndex(FPRIdx);
2246 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
2248 if (!MemOps.empty()) {
2249 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
2254 /// LowerCallResult - Lower the result values of a call into the
2255 /// appropriate copies out of appropriate physical registers.
2256 SDValue ARM64TargetLowering::LowerCallResult(
2257 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
2258 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2259 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
2260 SDValue ThisVal) const {
2261 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2262 : RetCC_ARM64_AAPCS;
2263 // Assign locations to each value returned by this call.
2264 SmallVector<CCValAssign, 16> RVLocs;
2265 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2266 getTargetMachine(), RVLocs, *DAG.getContext());
2267 CCInfo.AnalyzeCallResult(Ins, RetCC);
2269 // Copy all of the result registers out of their specified physreg.
2270 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2271 CCValAssign VA = RVLocs[i];
2273 // Pass 'this' value directly from the argument to return value, to avoid
2274 // reg unit interference
2275 if (i == 0 && isThisReturn) {
2276 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
2277 "unexpected return calling convention register assignment");
2278 InVals.push_back(ThisVal);
2283 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
2284 Chain = Val.getValue(1);
2285 InFlag = Val.getValue(2);
2287 switch (VA.getLocInfo()) {
2289 llvm_unreachable("Unknown loc info!");
2290 case CCValAssign::Full:
2292 case CCValAssign::BCvt:
2293 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2297 InVals.push_back(Val);
2303 bool ARM64TargetLowering::isEligibleForTailCallOptimization(
2304 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
2305 bool isCalleeStructRet, bool isCallerStructRet,
2306 const SmallVectorImpl<ISD::OutputArg> &Outs,
2307 const SmallVectorImpl<SDValue> &OutVals,
2308 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
2309 // Look for obvious safe cases to perform tail call optimization that do not
2310 // require ABI changes. This is what gcc calls sibcall.
2312 // Do not sibcall optimize vararg calls unless the call site is not passing
2314 if (isVarArg && !Outs.empty())
2317 // Also avoid sibcall optimization if either caller or callee uses struct
2318 // return semantics.
2319 if (isCalleeStructRet || isCallerStructRet)
2322 // Note that currently ARM64 "C" calling convention and "Fast" calling
2323 // convention are compatible. If/when that ever changes, we'll need to
2324 // add checks here to make sure any interactions are OK.
2326 // If the callee takes no arguments then go on to check the results of the
2328 if (!Outs.empty()) {
2329 // Check if stack adjustment is needed. For now, do not do this if any
2330 // argument is passed on the stack.
2331 SmallVector<CCValAssign, 16> ArgLocs;
2332 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2333 getTargetMachine(), ArgLocs, *DAG.getContext());
2334 CCAssignFn *AssignFn = CCAssignFnForCall(CalleeCC, /*IsVarArg=*/false);
2335 CCInfo.AnalyzeCallOperands(Outs, AssignFn);
2336 if (CCInfo.getNextStackOffset()) {
2337 // Check if the arguments are already laid out in the right way as
2338 // the caller's fixed stack objects.
2339 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2340 ++i, ++realArgIdx) {
2341 CCValAssign &VA = ArgLocs[i];
2342 if (VA.getLocInfo() == CCValAssign::Indirect)
2344 if (VA.needsCustom()) {
2345 // Just don't handle anything that needs custom adjustments for now.
2346 // If need be, we can revisit later, but we shouldn't ever end up
2349 } else if (!VA.isRegLoc()) {
2350 // Likewise, don't try to handle stack based arguments for the
2360 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
2361 /// and add input and output parameter nodes.
2362 SDValue ARM64TargetLowering::LowerCall(CallLoweringInfo &CLI,
2363 SmallVectorImpl<SDValue> &InVals) const {
2364 SelectionDAG &DAG = CLI.DAG;
2366 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2367 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2368 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2369 SDValue Chain = CLI.Chain;
2370 SDValue Callee = CLI.Callee;
2371 bool &IsTailCall = CLI.IsTailCall;
2372 CallingConv::ID CallConv = CLI.CallConv;
2373 bool IsVarArg = CLI.IsVarArg;
2375 MachineFunction &MF = DAG.getMachineFunction();
2376 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2377 bool IsThisReturn = false;
2379 // If tail calls are explicitly disabled, make sure not to use them.
2380 if (!EnableARM64TailCalls)
2384 // Check if it's really possible to do a tail call.
2385 IsTailCall = isEligibleForTailCallOptimization(
2386 Callee, CallConv, IsVarArg, IsStructRet,
2387 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2388 // We don't support GuaranteedTailCallOpt, only automatically
2389 // detected sibcalls.
2390 // FIXME: Re-evaluate. Is this true? Should it be true?
2395 // Analyze operands of the call, assigning locations to each operand.
2396 SmallVector<CCValAssign, 16> ArgLocs;
2397 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2398 getTargetMachine(), ArgLocs, *DAG.getContext());
2401 // Handle fixed and variable vector arguments differently.
2402 // Variable vector arguments always go into memory.
2403 unsigned NumArgs = Outs.size();
2405 for (unsigned i = 0; i != NumArgs; ++i) {
2406 MVT ArgVT = Outs[i].VT;
2407 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2408 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2409 /*IsVarArg=*/ !Outs[i].IsFixed);
2410 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2411 assert(!Res && "Call operand has unhandled type");
2415 // At this point, Outs[].VT may already be promoted to i32. To correctly
2416 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2417 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
2418 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2419 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2421 unsigned NumArgs = Outs.size();
2422 for (unsigned i = 0; i != NumArgs; ++i) {
2423 MVT ValVT = Outs[i].VT;
2424 // Get type of the original argument.
2425 EVT ActualVT = getValueType(CLI.Args[Outs[i].OrigArgIndex].Ty,
2426 /*AllowUnknown*/ true);
2427 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2428 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2429 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2431 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2433 else if (ActualMVT == MVT::i16)
2436 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2437 bool Res = AssignFn(i, ValVT, LocVT, CCValAssign::Full, ArgFlags, CCInfo);
2438 assert(!Res && "Call operand has unhandled type");
2443 // Get a count of how many bytes are to be pushed on the stack.
2444 unsigned NumBytes = CCInfo.getNextStackOffset();
2446 // Adjust the stack pointer for the new arguments...
2447 // These operations are automatically eliminated by the prolog/epilog pass
2450 DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true), DL);
2452 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, ARM64::SP, getPointerTy());
2454 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2455 SmallVector<SDValue, 8> MemOpChains;
2457 // Walk the register/memloc assignments, inserting copies/loads.
2458 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2459 ++i, ++realArgIdx) {
2460 CCValAssign &VA = ArgLocs[i];
2461 SDValue Arg = OutVals[realArgIdx];
2462 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2464 // Promote the value if needed.
2465 switch (VA.getLocInfo()) {
2467 llvm_unreachable("Unknown loc info!");
2468 case CCValAssign::Full:
2470 case CCValAssign::SExt:
2471 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
2473 case CCValAssign::ZExt:
2474 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2476 case CCValAssign::AExt:
2477 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
2479 case CCValAssign::BCvt:
2480 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2482 case CCValAssign::FPExt:
2483 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
2487 if (VA.isRegLoc()) {
2488 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
2489 assert(VA.getLocVT() == MVT::i64 &&
2490 "unexpected calling convention register assignment");
2491 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
2492 "unexpected use of 'returned'");
2493 IsThisReturn = true;
2495 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2497 assert(VA.isMemLoc());
2498 // There's no reason we can't support stack args w/ tailcall, but
2499 // we currently don't, so assert if we see one.
2500 assert(!IsTailCall && "stack argument with tail call!?");
2501 unsigned LocMemOffset = VA.getLocMemOffset();
2502 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2503 PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2505 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
2506 // promoted to a legal register type i32, we should truncate Arg back to
2508 if (Arg.getValueType().isSimple() &&
2509 Arg.getValueType().getSimpleVT() == MVT::i32 &&
2510 (VA.getLocVT() == MVT::i1 || VA.getLocVT() == MVT::i8 ||
2511 VA.getLocVT() == MVT::i16))
2512 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getLocVT(), Arg);
2514 SDValue Store = DAG.getStore(Chain, DL, Arg, PtrOff,
2515 MachinePointerInfo::getStack(LocMemOffset),
2517 MemOpChains.push_back(Store);
2521 if (!MemOpChains.empty())
2522 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOpChains[0],
2523 MemOpChains.size());
2525 // Build a sequence of copy-to-reg nodes chained together with token chain
2526 // and flag operands which copy the outgoing args into the appropriate regs.
2528 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2529 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[i].first,
2530 RegsToPass[i].second, InFlag);
2531 InFlag = Chain.getValue(1);
2534 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2535 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2536 // node so that legalize doesn't hack it.
2537 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
2538 Subtarget->isTargetMachO()) {
2539 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2540 const GlobalValue *GV = G->getGlobal();
2541 bool InternalLinkage = GV->hasInternalLinkage();
2542 if (InternalLinkage)
2543 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2545 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0,
2547 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2549 } else if (ExternalSymbolSDNode *S =
2550 dyn_cast<ExternalSymbolSDNode>(Callee)) {
2551 const char *Sym = S->getSymbol();
2553 DAG.getTargetExternalSymbol(Sym, getPointerTy(), ARM64II::MO_GOT);
2554 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2556 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2557 const GlobalValue *GV = G->getGlobal();
2558 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2559 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2560 const char *Sym = S->getSymbol();
2561 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), 0);
2564 std::vector<SDValue> Ops;
2565 Ops.push_back(Chain);
2566 Ops.push_back(Callee);
2568 // Add argument registers to the end of the list so that they are known live
2570 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2571 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2572 RegsToPass[i].second.getValueType()));
2574 // Add a register mask operand representing the call-preserved registers.
2575 const uint32_t *Mask;
2576 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2577 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2579 // For 'this' returns, use the X0-preserving mask if applicable
2580 Mask = ARI->getThisReturnPreservedMask(CallConv);
2582 IsThisReturn = false;
2583 Mask = ARI->getCallPreservedMask(CallConv);
2586 Mask = ARI->getCallPreservedMask(CallConv);
2588 assert(Mask && "Missing call preserved mask for calling convention");
2589 Ops.push_back(DAG.getRegisterMask(Mask));
2591 if (InFlag.getNode())
2592 Ops.push_back(InFlag);
2594 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2596 // If we're doing a tall call, use a TC_RETURN here rather than an
2597 // actual call instruction.
2599 return DAG.getNode(ARM64ISD::TC_RETURN, DL, NodeTys, &Ops[0], Ops.size());
2601 // Returns a chain and a flag for retval copy to use.
2602 Chain = DAG.getNode(ARM64ISD::CALL, DL, NodeTys, &Ops[0], Ops.size());
2603 InFlag = Chain.getValue(1);
2605 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2606 DAG.getIntPtrConstant(0, true), InFlag, DL);
2608 InFlag = Chain.getValue(1);
2610 // Handle result values, copying them out of physregs into vregs that we
2612 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2613 InVals, IsThisReturn,
2614 IsThisReturn ? OutVals[0] : SDValue());
2617 bool ARM64TargetLowering::CanLowerReturn(
2618 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2619 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2620 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2621 : RetCC_ARM64_AAPCS;
2622 SmallVector<CCValAssign, 16> RVLocs;
2623 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2624 return CCInfo.CheckReturn(Outs, RetCC);
2628 ARM64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2630 const SmallVectorImpl<ISD::OutputArg> &Outs,
2631 const SmallVectorImpl<SDValue> &OutVals,
2632 SDLoc DL, SelectionDAG &DAG) const {
2633 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2634 : RetCC_ARM64_AAPCS;
2635 SmallVector<CCValAssign, 16> RVLocs;
2636 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2637 getTargetMachine(), RVLocs, *DAG.getContext());
2638 CCInfo.AnalyzeReturn(Outs, RetCC);
2640 // Copy the result values into the output registers.
2642 SmallVector<SDValue, 4> RetOps(1, Chain);
2643 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
2644 ++i, ++realRVLocIdx) {
2645 CCValAssign &VA = RVLocs[i];
2646 assert(VA.isRegLoc() && "Can only return in registers!");
2647 SDValue Arg = OutVals[realRVLocIdx];
2649 switch (VA.getLocInfo()) {
2651 llvm_unreachable("Unknown loc info!");
2652 case CCValAssign::Full:
2654 case CCValAssign::BCvt:
2655 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2659 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
2660 Flag = Chain.getValue(1);
2661 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2664 RetOps[0] = Chain; // Update chain.
2666 // Add the flag if we have it.
2668 RetOps.push_back(Flag);
2670 return DAG.getNode(ARM64ISD::RET_FLAG, DL, MVT::Other, &RetOps[0],
2674 //===----------------------------------------------------------------------===//
2675 // Other Lowering Code
2676 //===----------------------------------------------------------------------===//
2678 SDValue ARM64TargetLowering::LowerGlobalAddress(SDValue Op,
2679 SelectionDAG &DAG) const {
2680 EVT PtrVT = getPointerTy();
2682 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2683 unsigned char OpFlags =
2684 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
2686 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
2687 "unexpected offset in global node");
2689 // This also catched the large code model case for Darwin.
2690 if ((OpFlags & ARM64II::MO_GOT) != 0) {
2691 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
2692 // FIXME: Once remat is capable of dealing with instructions with register
2693 // operands, expand this into two nodes instead of using a wrapper node.
2694 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
2697 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2698 const unsigned char MO_NC = ARM64II::MO_NC;
2700 ARM64ISD::WrapperLarge, DL, PtrVT,
2701 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G3),
2702 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
2703 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
2704 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
2706 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
2707 // the only correct model on Darwin.
2708 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2709 OpFlags | ARM64II::MO_PAGE);
2710 unsigned char LoFlags = OpFlags | ARM64II::MO_PAGEOFF | ARM64II::MO_NC;
2711 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
2713 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
2714 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
2718 /// \brief Convert a TLS address reference into the correct sequence of loads
2719 /// and calls to compute the variable's address (for Darwin, currently) and
2720 /// return an SDValue containing the final node.
2722 /// Darwin only has one TLS scheme which must be capable of dealing with the
2723 /// fully general situation, in the worst case. This means:
2724 /// + "extern __thread" declaration.
2725 /// + Defined in a possibly unknown dynamic library.
2727 /// The general system is that each __thread variable has a [3 x i64] descriptor
2728 /// which contains information used by the runtime to calculate the address. The
2729 /// only part of this the compiler needs to know about is the first xword, which
2730 /// contains a function pointer that must be called with the address of the
2731 /// entire descriptor in "x0".
2733 /// Since this descriptor may be in a different unit, in general even the
2734 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
2736 /// adrp x0, _var@TLVPPAGE
2737 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
2738 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
2739 /// ; the function pointer
2740 /// blr x1 ; Uses descriptor address in x0
2741 /// ; Address of _var is now in x0.
2743 /// If the address of _var's descriptor *is* known to the linker, then it can
2744 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
2745 /// a slight efficiency gain.
2747 ARM64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
2748 SelectionDAG &DAG) const {
2749 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
2752 MVT PtrVT = getPointerTy();
2753 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2756 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2757 SDValue DescAddr = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TLVPAddr);
2759 // The first entry in the descriptor is a function pointer that we must call
2760 // to obtain the address of the variable.
2761 SDValue Chain = DAG.getEntryNode();
2762 SDValue FuncTLVGet =
2763 DAG.getLoad(MVT::i64, DL, Chain, DescAddr, MachinePointerInfo::getGOT(),
2764 false, true, true, 8);
2765 Chain = FuncTLVGet.getValue(1);
2767 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2768 MFI->setAdjustsStack(true);
2770 // TLS calls preserve all registers except those that absolutely must be
2771 // trashed: X0 (it takes an argument), LR (it's a call) and CPSR (let's not be
2773 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2774 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2775 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2777 // Finally, we can make the call. This is just a degenerate version of a
2778 // normal ARM64 call node: x0 takes the address of the descriptor, and returns
2779 // the address of the variable in this thread.
2780 Chain = DAG.getCopyToReg(Chain, DL, ARM64::X0, DescAddr, SDValue());
2781 Chain = DAG.getNode(ARM64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
2782 Chain, FuncTLVGet, DAG.getRegister(ARM64::X0, MVT::i64),
2783 DAG.getRegisterMask(Mask), Chain.getValue(1));
2784 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Chain.getValue(1));
2787 /// When accessing thread-local variables under either the general-dynamic or
2788 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
2789 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
2790 /// is a function pointer to carry out the resolution. This function takes the
2791 /// address of the descriptor in X0 and returns the TPIDR_EL0 offset in X0. All
2792 /// other registers (except LR, CPSR) are preserved.
2794 /// Thus, the ideal call sequence on AArch64 is:
2796 /// adrp x0, :tlsdesc:thread_var
2797 /// ldr x8, [x0, :tlsdesc_lo12:thread_var]
2798 /// add x0, x0, :tlsdesc_lo12:thread_var
2799 /// .tlsdesccall thread_var
2801 /// (TPIDR_EL0 offset now in x0).
2803 /// The ".tlsdesccall" directive instructs the assembler to insert a particular
2804 /// relocation to help the linker relax this sequence if it turns out to be too
2807 /// FIXME: we currently produce an extra, duplicated, ADRP instruction, but this
2809 SDValue ARM64TargetLowering::LowerELFTLSDescCall(SDValue SymAddr,
2810 SDValue DescAddr, SDLoc DL,
2811 SelectionDAG &DAG) const {
2812 EVT PtrVT = getPointerTy();
2814 // The function we need to call is simply the first entry in the GOT for this
2815 // descriptor, load it in preparation.
2816 SDValue Func = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, SymAddr);
2818 // TLS calls preserve all registers except those that absolutely must be
2819 // trashed: X0 (it takes an argument), LR (it's a call) and CPSR (let's not be
2821 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2822 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2823 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2825 // The function takes only one argument: the address of the descriptor itself
2827 SDValue Glue, Chain;
2828 Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, ARM64::X0, DescAddr, Glue);
2829 Glue = Chain.getValue(1);
2831 // We're now ready to populate the argument list, as with a normal call:
2832 SmallVector<SDValue, 6> Ops;
2833 Ops.push_back(Chain);
2834 Ops.push_back(Func);
2835 Ops.push_back(SymAddr);
2836 Ops.push_back(DAG.getRegister(ARM64::X0, PtrVT));
2837 Ops.push_back(DAG.getRegisterMask(Mask));
2838 Ops.push_back(Glue);
2840 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2841 Chain = DAG.getNode(ARM64ISD::TLSDESC_CALL, DL, NodeTys, &Ops[0], Ops.size());
2842 Glue = Chain.getValue(1);
2844 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Glue);
2847 SDValue ARM64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
2848 SelectionDAG &DAG) const {
2849 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
2850 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
2851 "ELF TLS only supported in small memory model");
2852 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2854 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
2857 EVT PtrVT = getPointerTy();
2859 const GlobalValue *GV = GA->getGlobal();
2861 SDValue ThreadBase = DAG.getNode(ARM64ISD::THREAD_POINTER, DL, PtrVT);
2863 if (Model == TLSModel::LocalExec) {
2864 SDValue HiVar = DAG.getTargetGlobalAddress(
2865 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2866 SDValue LoVar = DAG.getTargetGlobalAddress(
2867 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2869 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2870 DAG.getTargetConstant(16, MVT::i32)),
2872 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, TPOff, LoVar,
2873 DAG.getTargetConstant(0, MVT::i32)),
2875 } else if (Model == TLSModel::InitialExec) {
2876 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2877 TPOff = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TPOff);
2878 } else if (Model == TLSModel::LocalDynamic) {
2879 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
2880 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
2881 // the beginning of the module's TLS region, followed by a DTPREL offset
2884 // These accesses will need deduplicating if there's more than one.
2885 ARM64FunctionInfo *MFI =
2886 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
2887 MFI->incNumLocalDynamicTLSAccesses();
2889 // Accesses used in this sequence go via the TLS descriptor which lives in
2890 // the GOT. Prepare an address we can use to handle this.
2891 SDValue HiDesc = DAG.getTargetExternalSymbol(
2892 "_TLS_MODULE_BASE_", PtrVT, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2893 SDValue LoDesc = DAG.getTargetExternalSymbol(
2894 "_TLS_MODULE_BASE_", PtrVT,
2895 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2897 // First argument to the descriptor call is the address of the descriptor
2899 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2900 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2902 // The call needs a relocation too for linker relaxation. It doesn't make
2903 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2905 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2908 // Now we can calculate the offset from TPIDR_EL0 to this module's
2909 // thread-local area.
2910 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2912 // Now use :dtprel_whatever: operations to calculate this variable's offset
2913 // in its thread-storage area.
2914 SDValue HiVar = DAG.getTargetGlobalAddress(
2915 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2916 SDValue LoVar = DAG.getTargetGlobalAddress(
2917 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2920 SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2921 DAG.getTargetConstant(16, MVT::i32)),
2923 DTPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, DTPOff, LoVar,
2924 DAG.getTargetConstant(0, MVT::i32)),
2927 TPOff = DAG.getNode(ISD::ADD, DL, PtrVT, TPOff, DTPOff);
2928 } else if (Model == TLSModel::GeneralDynamic) {
2929 // Accesses used in this sequence go via the TLS descriptor which lives in
2930 // the GOT. Prepare an address we can use to handle this.
2931 SDValue HiDesc = DAG.getTargetGlobalAddress(
2932 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2933 SDValue LoDesc = DAG.getTargetGlobalAddress(
2935 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2937 // First argument to the descriptor call is the address of the descriptor
2939 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2940 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2942 // The call needs a relocation too for linker relaxation. It doesn't make
2943 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2946 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2948 // Finally we can make a call to calculate the offset from tpidr_el0.
2949 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2951 llvm_unreachable("Unsupported ELF TLS access model");
2953 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
2956 SDValue ARM64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
2957 SelectionDAG &DAG) const {
2958 if (Subtarget->isTargetDarwin())
2959 return LowerDarwinGlobalTLSAddress(Op, DAG);
2960 else if (Subtarget->isTargetELF())
2961 return LowerELFGlobalTLSAddress(Op, DAG);
2963 llvm_unreachable("Unexpected platform trying to use TLS");
2965 SDValue ARM64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2966 SDValue Chain = Op.getOperand(0);
2967 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2968 SDValue LHS = Op.getOperand(2);
2969 SDValue RHS = Op.getOperand(3);
2970 SDValue Dest = Op.getOperand(4);
2973 // Handle f128 first, since lowering it will result in comparing the return
2974 // value of a libcall against zero, which is just what the rest of LowerBR_CC
2975 // is expecting to deal with.
2976 if (LHS.getValueType() == MVT::f128) {
2977 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2979 // If softenSetCCOperands returned a scalar, we need to compare the result
2980 // against zero to select between true and false values.
2981 if (RHS.getNode() == 0) {
2982 RHS = DAG.getConstant(0, LHS.getValueType());
2987 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
2989 unsigned Opc = LHS.getOpcode();
2990 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
2991 cast<ConstantSDNode>(RHS)->isOne() &&
2992 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
2993 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
2994 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
2995 "Unexpected condition code.");
2996 // Only lower legal XALUO ops.
2997 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
3000 // The actual operation with overflow check.
3001 ARM64CC::CondCode OFCC;
3002 SDValue Value, Overflow;
3003 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, LHS.getValue(0), DAG);
3005 if (CC == ISD::SETNE)
3006 OFCC = getInvertedCondCode(OFCC);
3007 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3009 return DAG.getNode(ARM64ISD::BRCOND, SDLoc(LHS), MVT::Other, Chain, Dest,
3013 if (LHS.getValueType().isInteger()) {
3014 assert((LHS.getValueType() == RHS.getValueType()) &&
3015 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3017 // If the RHS of the comparison is zero, we can potentially fold this
3018 // to a specialized branch.
3019 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
3020 if (RHSC && RHSC->getZExtValue() == 0) {
3021 if (CC == ISD::SETEQ) {
3022 // See if we can use a TBZ to fold in an AND as well.
3023 // TBZ has a smaller branch displacement than CBZ. If the offset is
3024 // out of bounds, a late MI-layer pass rewrites branches.
3025 // 403.gcc is an example that hits this case.
3026 if (LHS.getOpcode() == ISD::AND &&
3027 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3028 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3029 SDValue Test = LHS.getOperand(0);
3030 uint64_t Mask = LHS.getConstantOperandVal(1);
3032 // TBZ only operates on i64's, but the ext should be free.
3033 if (Test.getValueType() == MVT::i32)
3034 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
3036 return DAG.getNode(ARM64ISD::TBZ, dl, MVT::Other, Chain, Test,
3037 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3040 return DAG.getNode(ARM64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
3041 } else if (CC == ISD::SETNE) {
3042 // See if we can use a TBZ to fold in an AND as well.
3043 // TBZ has a smaller branch displacement than CBZ. If the offset is
3044 // out of bounds, a late MI-layer pass rewrites branches.
3045 // 403.gcc is an example that hits this case.
3046 if (LHS.getOpcode() == ISD::AND &&
3047 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3048 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3049 SDValue Test = LHS.getOperand(0);
3050 uint64_t Mask = LHS.getConstantOperandVal(1);
3052 // TBNZ only operates on i64's, but the ext should be free.
3053 if (Test.getValueType() == MVT::i32)
3054 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
3056 return DAG.getNode(ARM64ISD::TBNZ, dl, MVT::Other, Chain, Test,
3057 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3060 return DAG.getNode(ARM64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
3065 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3066 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3070 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3072 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3073 // clean. Some of them require two branches to implement.
3074 SDValue Cmp = emitComparison(LHS, RHS, dl, DAG);
3075 ARM64CC::CondCode CC1, CC2;
3076 changeFPCCToARM64CC(CC, CC1, CC2);
3077 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3079 DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
3080 if (CC2 != ARM64CC::AL) {
3081 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3082 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
3089 SDValue ARM64TargetLowering::LowerFCOPYSIGN(SDValue Op,
3090 SelectionDAG &DAG) const {
3091 EVT VT = Op.getValueType();
3094 SDValue In1 = Op.getOperand(0);
3095 SDValue In2 = Op.getOperand(1);
3096 EVT SrcVT = In2.getValueType();
3098 if (SrcVT == MVT::f32 && VT == MVT::f64)
3099 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
3100 else if (SrcVT == MVT::f64 && VT == MVT::f32)
3101 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0));
3103 // FIXME: Src type is different, bail out for now. Can VT really be a
3110 SDValue EltMask, VecVal1, VecVal2;
3111 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
3114 EltMask = DAG.getConstant(0x80000000ULL, EltVT);
3116 if (!VT.isVector()) {
3117 VecVal1 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
3118 DAG.getUNDEF(VecVT), In1);
3119 VecVal2 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
3120 DAG.getUNDEF(VecVT), In2);
3122 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3123 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3125 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
3129 // We want to materialize a mask with the the high bit set, but the AdvSIMD
3130 // immediate moves cannot materialize that in a single instruction for
3131 // 64-bit elements. Instead, materialize zero and then negate it.
3132 EltMask = DAG.getConstant(0, EltVT);
3134 if (!VT.isVector()) {
3135 VecVal1 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
3136 DAG.getUNDEF(VecVT), In1);
3137 VecVal2 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
3138 DAG.getUNDEF(VecVT), In2);
3140 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3141 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3144 llvm_unreachable("Invalid type for copysign!");
3147 std::vector<SDValue> BuildVectorOps;
3148 for (unsigned i = 0; i < VecVT.getVectorNumElements(); ++i)
3149 BuildVectorOps.push_back(EltMask);
3151 SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT,
3152 &BuildVectorOps[0], BuildVectorOps.size());
3154 // If we couldn't materialize the mask above, then the mask vector will be
3155 // the zero vector, and we need to negate it here.
3156 if (VT == MVT::f64 || VT == MVT::v2f64) {
3157 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
3158 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
3159 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
3163 DAG.getNode(ARM64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
3166 return DAG.getTargetExtractSubreg(ARM64::ssub, DL, VT, Sel);
3167 else if (VT == MVT::f64)
3168 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, VT, Sel);
3170 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
3173 SDValue ARM64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
3174 if (DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
3175 AttributeSet::FunctionIndex, Attribute::NoImplicitFloat))
3178 // While there is no integer popcount instruction, it can
3179 // be more efficiently lowered to the following sequence that uses
3180 // AdvSIMD registers/instructions as long as the copies to/from
3181 // the AdvSIMD registers are cheap.
3182 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
3183 // CNT V0.8B, V0.8B // 8xbyte pop-counts
3184 // ADDV B0, V0.8B // sum 8xbyte pop-counts
3185 // UMOV X0, V0.B[0] // copy byte result back to integer reg
3186 SDValue Val = Op.getOperand(0);
3188 EVT VT = Op.getValueType();
3189 SDValue ZeroVec = DAG.getUNDEF(MVT::v8i8);
3192 if (VT == MVT::i32) {
3193 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Val);
3195 DAG.getTargetInsertSubreg(ARM64::ssub, DL, MVT::v8i8, ZeroVec, VecVal);
3197 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
3200 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, VecVal);
3201 SDValue UaddLV = DAG.getNode(
3202 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
3203 DAG.getConstant(Intrinsic::arm64_neon_uaddlv, MVT::i32), CtPop);
3206 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
3210 SDValue ARM64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3212 if (Op.getValueType().isVector())
3213 return LowerVSETCC(Op, DAG);
3215 SDValue LHS = Op.getOperand(0);
3216 SDValue RHS = Op.getOperand(1);
3217 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3220 // We chose ZeroOrOneBooleanContents, so use zero and one.
3221 EVT VT = Op.getValueType();
3222 SDValue TVal = DAG.getConstant(1, VT);
3223 SDValue FVal = DAG.getConstant(0, VT);
3225 // Handle f128 first, since one possible outcome is a normal integer
3226 // comparison which gets picked up by the next if statement.
3227 if (LHS.getValueType() == MVT::f128) {
3228 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3230 // If softenSetCCOperands returned a scalar, use it.
3231 if (RHS.getNode() == 0) {
3232 assert(LHS.getValueType() == Op.getValueType() &&
3233 "Unexpected setcc expansion!");
3238 if (LHS.getValueType().isInteger()) {
3241 getARM64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
3243 // Note that we inverted the condition above, so we reverse the order of
3244 // the true and false operands here. This will allow the setcc to be
3245 // matched to a single CSINC instruction.
3246 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
3249 // Now we know we're dealing with FP values.
3250 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3252 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3253 // and do the comparison.
3254 SDValue Cmp = emitComparison(LHS, RHS, dl, DAG);
3256 ARM64CC::CondCode CC1, CC2;
3257 changeFPCCToARM64CC(CC, CC1, CC2);
3258 if (CC2 == ARM64CC::AL) {
3259 changeFPCCToARM64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
3260 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3262 // Note that we inverted the condition above, so we reverse the order of
3263 // the true and false operands here. This will allow the setcc to be
3264 // matched to a single CSINC instruction.
3265 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
3267 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3268 // clean. Some of them require two CSELs to implement. As is in this case,
3269 // we emit the first CSEL and then emit a second using the output of the
3270 // first as the RHS. We're effectively OR'ing the two CC's together.
3272 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
3273 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3274 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3276 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3277 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3281 /// A SELECT_CC operation is really some kind of max or min if both values being
3282 /// compared are, in some sense, equal to the results in either case. However,
3283 /// it is permissible to compare f32 values and produce directly extended f64
3286 /// Extending the comparison operands would also be allowed, but is less likely
3287 /// to happen in practice since their use is right here. Note that truncate
3288 /// operations would *not* be semantically equivalent.
3289 static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) {
3293 ConstantFPSDNode *CCmp = dyn_cast<ConstantFPSDNode>(Cmp);
3294 ConstantFPSDNode *CResult = dyn_cast<ConstantFPSDNode>(Result);
3295 if (CCmp && CResult && Cmp.getValueType() == MVT::f32 &&
3296 Result.getValueType() == MVT::f64) {
3298 APFloat CmpVal = CCmp->getValueAPF();
3299 CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy);
3300 return CResult->getValueAPF().bitwiseIsEqual(CmpVal);
3303 return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp;
3306 SDValue ARM64TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3307 SDValue CC = Op->getOperand(0);
3308 SDValue TVal = Op->getOperand(1);
3309 SDValue FVal = Op->getOperand(2);
3312 unsigned Opc = CC.getOpcode();
3313 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
3315 if (CC.getResNo() == 1 &&
3316 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3317 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3318 // Only lower legal XALUO ops.
3319 if (!DAG.getTargetLoweringInfo().isTypeLegal(CC->getValueType(0)))
3322 ARM64CC::CondCode OFCC;
3323 SDValue Value, Overflow;
3324 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, CC.getValue(0), DAG);
3325 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3327 return DAG.getNode(ARM64ISD::CSEL, DL, Op.getValueType(), TVal, FVal, CCVal,
3331 if (CC.getOpcode() == ISD::SETCC)
3332 return DAG.getSelectCC(DL, CC.getOperand(0), CC.getOperand(1), TVal, FVal,
3333 cast<CondCodeSDNode>(CC.getOperand(2))->get());
3335 return DAG.getSelectCC(DL, CC, DAG.getConstant(0, CC.getValueType()), TVal,
3339 SDValue ARM64TargetLowering::LowerSELECT_CC(SDValue Op,
3340 SelectionDAG &DAG) const {
3341 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3342 SDValue LHS = Op.getOperand(0);
3343 SDValue RHS = Op.getOperand(1);
3344 SDValue TVal = Op.getOperand(2);
3345 SDValue FVal = Op.getOperand(3);
3348 // Handle f128 first, because it will result in a comparison of some RTLIB
3349 // call result against zero.
3350 if (LHS.getValueType() == MVT::f128) {
3351 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3353 // If softenSetCCOperands returned a scalar, we need to compare the result
3354 // against zero to select between true and false values.
3355 if (RHS.getNode() == 0) {
3356 RHS = DAG.getConstant(0, LHS.getValueType());
3361 // Handle integers first.
3362 if (LHS.getValueType().isInteger()) {
3363 assert((LHS.getValueType() == RHS.getValueType()) &&
3364 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3366 unsigned Opcode = ARM64ISD::CSEL;
3368 // If both the TVal and the FVal are constants, see if we can swap them in
3369 // order to for a CSINV or CSINC out of them.
3370 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3371 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3373 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3374 std::swap(TVal, FVal);
3375 std::swap(CTVal, CFVal);
3376 CC = ISD::getSetCCInverse(CC, true);
3377 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3378 std::swap(TVal, FVal);
3379 std::swap(CTVal, CFVal);
3380 CC = ISD::getSetCCInverse(CC, true);
3381 } else if (TVal.getOpcode() == ISD::XOR) {
3382 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3383 // with a CSINV rather than a CSEL.
3384 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
3386 if (CVal && CVal->isAllOnesValue()) {
3387 std::swap(TVal, FVal);
3388 std::swap(CTVal, CFVal);
3389 CC = ISD::getSetCCInverse(CC, true);
3391 } else if (TVal.getOpcode() == ISD::SUB) {
3392 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3393 // that we can match with a CSNEG rather than a CSEL.
3394 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
3396 if (CVal && CVal->isNullValue()) {
3397 std::swap(TVal, FVal);
3398 std::swap(CTVal, CFVal);
3399 CC = ISD::getSetCCInverse(CC, true);
3401 } else if (CTVal && CFVal) {
3402 const int64_t TrueVal = CTVal->getSExtValue();
3403 const int64_t FalseVal = CFVal->getSExtValue();
3406 // If both TVal and FVal are constants, see if FVal is the
3407 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3408 // instead of a CSEL in that case.
3409 if (TrueVal == ~FalseVal) {
3410 Opcode = ARM64ISD::CSINV;
3411 } else if (TrueVal == -FalseVal) {
3412 Opcode = ARM64ISD::CSNEG;
3413 } else if (TVal.getValueType() == MVT::i32) {
3414 // If our operands are only 32-bit wide, make sure we use 32-bit
3415 // arithmetic for the check whether we can use CSINC. This ensures that
3416 // the addition in the check will wrap around properly in case there is
3417 // an overflow (which would not be the case if we do the check with
3418 // 64-bit arithmetic).
3419 const uint32_t TrueVal32 = CTVal->getZExtValue();
3420 const uint32_t FalseVal32 = CFVal->getZExtValue();
3422 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3423 Opcode = ARM64ISD::CSINC;
3425 if (TrueVal32 > FalseVal32) {
3429 // 64-bit check whether we can use CSINC.
3430 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3431 Opcode = ARM64ISD::CSINC;
3433 if (TrueVal > FalseVal) {
3438 // Swap TVal and FVal if necessary.
3440 std::swap(TVal, FVal);
3441 std::swap(CTVal, CFVal);
3442 CC = ISD::getSetCCInverse(CC, true);
3445 if (Opcode != ARM64ISD::CSEL) {
3446 // Drop FVal since we can get its value by simply inverting/negating
3453 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3455 EVT VT = Op.getValueType();
3456 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3459 // Now we know we're dealing with FP values.
3460 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3461 assert(LHS.getValueType() == RHS.getValueType());
3462 EVT VT = Op.getValueType();
3464 // Try to match this select into a max/min operation, which have dedicated
3465 // opcode in the instruction set.
3466 // NOTE: This is not correct in the presence of NaNs, so we only enable this
3468 if (getTargetMachine().Options.NoNaNsFPMath) {
3469 if (selectCCOpsAreFMaxCompatible(LHS, FVal) &&
3470 selectCCOpsAreFMaxCompatible(RHS, TVal)) {
3471 CC = ISD::getSetCCSwappedOperands(CC);
3472 std::swap(TVal, FVal);
3475 if (selectCCOpsAreFMaxCompatible(LHS, TVal) &&
3476 selectCCOpsAreFMaxCompatible(RHS, FVal)) {
3486 return DAG.getNode(ARM64ISD::FMAX, dl, VT, TVal, FVal);
3494 return DAG.getNode(ARM64ISD::FMIN, dl, VT, TVal, FVal);
3500 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3501 // and do the comparison.
3502 SDValue Cmp = emitComparison(LHS, RHS, dl, DAG);
3504 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3505 // clean. Some of them require two CSELs to implement.
3506 ARM64CC::CondCode CC1, CC2;
3507 changeFPCCToARM64CC(CC, CC1, CC2);
3508 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3509 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3511 // If we need a second CSEL, emit it, using the output of the first as the
3512 // RHS. We're effectively OR'ing the two CC's together.
3513 if (CC2 != ARM64CC::AL) {
3514 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3515 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3518 // Otherwise, return the output of the first CSEL.
3522 SDValue ARM64TargetLowering::LowerJumpTable(SDValue Op,
3523 SelectionDAG &DAG) const {
3524 // Jump table entries as PC relative offsets. No additional tweaking
3525 // is necessary here. Just get the address of the jump table.
3526 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3527 EVT PtrVT = getPointerTy();
3530 SDValue Hi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_PAGE);
3531 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3532 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3533 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3534 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3537 SDValue ARM64TargetLowering::LowerConstantPool(SDValue Op,
3538 SelectionDAG &DAG) const {
3539 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3540 EVT PtrVT = getPointerTy();
3543 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3544 // Use the GOT for the large code model on iOS.
3545 if (Subtarget->isTargetMachO()) {
3546 SDValue GotAddr = DAG.getTargetConstantPool(
3547 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3549 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
3552 const unsigned char MO_NC = ARM64II::MO_NC;
3554 ARM64ISD::WrapperLarge, DL, PtrVT,
3555 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3556 CP->getOffset(), ARM64II::MO_G3),
3557 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3558 CP->getOffset(), ARM64II::MO_G2 | MO_NC),
3559 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3560 CP->getOffset(), ARM64II::MO_G1 | MO_NC),
3561 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3562 CP->getOffset(), ARM64II::MO_G0 | MO_NC));
3564 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
3565 // ELF, the only valid one on Darwin.
3567 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3568 CP->getOffset(), ARM64II::MO_PAGE);
3569 SDValue Lo = DAG.getTargetConstantPool(
3570 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3571 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3573 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3574 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3578 SDValue ARM64TargetLowering::LowerBlockAddress(SDValue Op,
3579 SelectionDAG &DAG) const {
3580 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
3581 EVT PtrVT = getPointerTy();
3583 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3584 !Subtarget->isTargetMachO()) {
3585 const unsigned char MO_NC = ARM64II::MO_NC;
3587 ARM64ISD::WrapperLarge, DL, PtrVT,
3588 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G3),
3589 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
3590 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
3591 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
3593 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGE);
3594 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGEOFF |
3596 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3597 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3601 SDValue ARM64TargetLowering::LowerDarwin_VASTART(SDValue Op,
3602 SelectionDAG &DAG) const {
3603 ARM64FunctionInfo *FuncInfo =
3604 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
3608 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3609 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3610 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
3611 MachinePointerInfo(SV), false, false, 0);
3614 SDValue ARM64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
3615 SelectionDAG &DAG) const {
3616 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3617 // Standard, section B.3.
3618 MachineFunction &MF = DAG.getMachineFunction();
3619 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
3622 SDValue Chain = Op.getOperand(0);
3623 SDValue VAList = Op.getOperand(1);
3624 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3625 SmallVector<SDValue, 4> MemOps;
3627 // void *__stack at offset 0
3629 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3630 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3631 MachinePointerInfo(SV), false, false, 8));
3633 // void *__gr_top at offset 8
3634 int GPRSize = FuncInfo->getVarArgsGPRSize();
3636 SDValue GRTop, GRTopAddr;
3638 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3639 DAG.getConstant(8, getPointerTy()));
3641 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), getPointerTy());
3642 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
3643 DAG.getConstant(GPRSize, getPointerTy()));
3645 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3646 MachinePointerInfo(SV, 8), false, false, 8));
3649 // void *__vr_top at offset 16
3650 int FPRSize = FuncInfo->getVarArgsFPRSize();
3652 SDValue VRTop, VRTopAddr;
3653 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3654 DAG.getConstant(16, getPointerTy()));
3656 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), getPointerTy());
3657 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
3658 DAG.getConstant(FPRSize, getPointerTy()));
3660 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3661 MachinePointerInfo(SV, 16), false, false, 8));
3664 // int __gr_offs at offset 24
3665 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3666 DAG.getConstant(24, getPointerTy()));
3667 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
3668 GROffsAddr, MachinePointerInfo(SV, 24), false,
3671 // int __vr_offs at offset 28
3672 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3673 DAG.getConstant(28, getPointerTy()));
3674 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
3675 VROffsAddr, MachinePointerInfo(SV, 28), false,
3678 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
3682 SDValue ARM64TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
3683 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
3684 : LowerAAPCS_VASTART(Op, DAG);
3687 SDValue ARM64TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
3688 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
3690 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
3691 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3692 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
3694 return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op), Op.getOperand(1),
3695 Op.getOperand(2), DAG.getConstant(VaListSize, MVT::i32),
3696 8, false, false, MachinePointerInfo(DestSV),
3697 MachinePointerInfo(SrcSV));
3700 SDValue ARM64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
3701 assert(Subtarget->isTargetDarwin() &&
3702 "automatic va_arg instruction only works on Darwin");
3704 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3705 EVT VT = Op.getValueType();
3707 SDValue Chain = Op.getOperand(0);
3708 SDValue Addr = Op.getOperand(1);
3709 unsigned Align = Op.getConstantOperandVal(3);
3711 SDValue VAList = DAG.getLoad(getPointerTy(), DL, Chain, Addr,
3712 MachinePointerInfo(V), false, false, false, 0);
3713 Chain = VAList.getValue(1);
3716 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
3717 VAList = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3718 DAG.getConstant(Align - 1, getPointerTy()));
3719 VAList = DAG.getNode(ISD::AND, DL, getPointerTy(), VAList,
3720 DAG.getConstant(-(int64_t)Align, getPointerTy()));
3723 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
3724 uint64_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
3726 // Scalar integer and FP values smaller than 64 bits are implicitly extended
3727 // up to 64 bits. At the very least, we have to increase the striding of the
3728 // vaargs list to match this, and for FP values we need to introduce
3729 // FP_ROUND nodes as well.
3730 if (VT.isInteger() && !VT.isVector())
3732 bool NeedFPTrunc = false;
3733 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
3738 // Increment the pointer, VAList, to the next vaarg
3739 SDValue VANext = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3740 DAG.getConstant(ArgSize, getPointerTy()));
3741 // Store the incremented VAList to the legalized pointer
3742 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
3745 // Load the actual argument out of the pointer VAList
3747 // Load the value as an f64.
3748 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
3749 MachinePointerInfo(), false, false, false, 0);
3750 // Round the value down to an f32.
3751 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
3752 DAG.getIntPtrConstant(1));
3753 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
3754 // Merge the rounded value with the chain output of the load.
3755 return DAG.getMergeValues(Ops, 2, DL);
3758 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
3762 SDValue ARM64TargetLowering::LowerFRAMEADDR(SDValue Op,
3763 SelectionDAG &DAG) const {
3764 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3765 MFI->setFrameAddressIsTaken(true);
3767 EVT VT = Op.getValueType();
3769 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3770 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, ARM64::FP, VT);
3772 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
3773 MachinePointerInfo(), false, false, false, 0);
3777 SDValue ARM64TargetLowering::LowerRETURNADDR(SDValue Op,
3778 SelectionDAG &DAG) const {
3779 MachineFunction &MF = DAG.getMachineFunction();
3780 MachineFrameInfo *MFI = MF.getFrameInfo();
3781 MFI->setReturnAddressIsTaken(true);
3783 EVT VT = Op.getValueType();
3785 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3787 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3788 SDValue Offset = DAG.getConstant(8, getPointerTy());
3789 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
3790 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
3791 MachinePointerInfo(), false, false, false, 0);
3794 // Return LR, which contains the return address. Mark it an implicit live-in.
3795 unsigned Reg = MF.addLiveIn(ARM64::LR, &ARM64::GPR64RegClass);
3796 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
3799 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3800 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3801 SDValue ARM64TargetLowering::LowerShiftRightParts(SDValue Op,
3802 SelectionDAG &DAG) const {
3803 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3804 EVT VT = Op.getValueType();
3805 unsigned VTBits = VT.getSizeInBits();
3807 SDValue ShOpLo = Op.getOperand(0);
3808 SDValue ShOpHi = Op.getOperand(1);
3809 SDValue ShAmt = Op.getOperand(2);
3811 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3813 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3815 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3816 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3817 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3818 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3819 DAG.getConstant(VTBits, MVT::i64));
3820 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3823 emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64), dl, DAG);
3824 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3826 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3827 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3829 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3831 // ARM64 shifts larger than the register width are wrapped rather than
3832 // clamped, so we can't just emit "hi >> x".
3833 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3834 SDValue TrueValHi = Opc == ISD::SRA
3835 ? DAG.getNode(Opc, dl, VT, ShOpHi,
3836 DAG.getConstant(VTBits - 1, MVT::i64))
3837 : DAG.getConstant(0, VT);
3839 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
3841 SDValue Ops[2] = { Lo, Hi };
3842 return DAG.getMergeValues(Ops, 2, dl);
3845 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3846 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3847 SDValue ARM64TargetLowering::LowerShiftLeftParts(SDValue Op,
3848 SelectionDAG &DAG) const {
3849 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3850 EVT VT = Op.getValueType();
3851 unsigned VTBits = VT.getSizeInBits();
3853 SDValue ShOpLo = Op.getOperand(0);
3854 SDValue ShOpHi = Op.getOperand(1);
3855 SDValue ShAmt = Op.getOperand(2);
3858 assert(Op.getOpcode() == ISD::SHL_PARTS);
3859 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3860 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3861 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3862 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3863 DAG.getConstant(VTBits, MVT::i64));
3864 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3865 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3867 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3870 emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64), dl, DAG);
3871 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3872 SDValue Hi = DAG.getNode(ARM64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
3874 // ARM64 shifts of larger than register sizes are wrapped rather than clamped,
3875 // so we can't just emit "lo << a" if a is too big.
3876 SDValue TrueValLo = DAG.getConstant(0, VT);
3877 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3879 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3881 SDValue Ops[2] = { Lo, Hi };
3882 return DAG.getMergeValues(Ops, 2, dl);
3886 ARM64TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3887 // The ARM64 target doesn't support folding offsets into global addresses.
3891 bool ARM64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3892 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
3893 // FIXME: We should be able to handle f128 as well with a clever lowering.
3894 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
3898 return ARM64_AM::getFP64Imm(Imm) != -1;
3899 else if (VT == MVT::f32)
3900 return ARM64_AM::getFP32Imm(Imm) != -1;
3904 //===----------------------------------------------------------------------===//
3905 // ARM64 Optimization Hooks
3906 //===----------------------------------------------------------------------===//
3908 //===----------------------------------------------------------------------===//
3909 // ARM64 Inline Assembly Support
3910 //===----------------------------------------------------------------------===//
3912 // Table of Constraints
3913 // TODO: This is the current set of constraints supported by ARM for the
3914 // compiler, not all of them may make sense, e.g. S may be difficult to support.
3916 // r - A general register
3917 // w - An FP/SIMD register of some size in the range v0-v31
3918 // x - An FP/SIMD register of some size in the range v0-v15
3919 // I - Constant that can be used with an ADD instruction
3920 // J - Constant that can be used with a SUB instruction
3921 // K - Constant that can be used with a 32-bit logical instruction
3922 // L - Constant that can be used with a 64-bit logical instruction
3923 // M - Constant that can be used as a 32-bit MOV immediate
3924 // N - Constant that can be used as a 64-bit MOV immediate
3925 // Q - A memory reference with base register and no offset
3926 // S - A symbolic address
3927 // Y - Floating point constant zero
3928 // Z - Integer constant zero
3930 // Note that general register operands will be output using their 64-bit x
3931 // register name, whatever the size of the variable, unless the asm operand
3932 // is prefixed by the %w modifier. Floating-point and SIMD register operands
3933 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
3936 /// getConstraintType - Given a constraint letter, return the type of
3937 /// constraint it is for this target.
3938 ARM64TargetLowering::ConstraintType
3939 ARM64TargetLowering::getConstraintType(const std::string &Constraint) const {
3940 if (Constraint.size() == 1) {
3941 switch (Constraint[0]) {
3948 return C_RegisterClass;
3949 // An address with a single base register. Due to the way we
3950 // currently handle addresses it is the same as 'r'.
3955 return TargetLowering::getConstraintType(Constraint);
3958 /// Examine constraint type and operand type and determine a weight value.
3959 /// This object must already have been set up with the operand type
3960 /// and the current alternative constraint selected.
3961 TargetLowering::ConstraintWeight
3962 ARM64TargetLowering::getSingleConstraintMatchWeight(
3963 AsmOperandInfo &info, const char *constraint) const {
3964 ConstraintWeight weight = CW_Invalid;
3965 Value *CallOperandVal = info.CallOperandVal;
3966 // If we don't have a value, we can't do a match,
3967 // but allow it at the lowest weight.
3968 if (CallOperandVal == NULL)
3970 Type *type = CallOperandVal->getType();
3971 // Look at the constraint type.
3972 switch (*constraint) {
3974 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3978 if (type->isFloatingPointTy() || type->isVectorTy())
3979 weight = CW_Register;
3982 weight = CW_Constant;
3988 std::pair<unsigned, const TargetRegisterClass *>
3989 ARM64TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3991 if (Constraint.size() == 1) {
3992 switch (Constraint[0]) {
3994 if (VT.getSizeInBits() == 64)
3995 return std::make_pair(0U, &ARM64::GPR64commonRegClass);
3996 return std::make_pair(0U, &ARM64::GPR32commonRegClass);
3999 return std::make_pair(0U, &ARM64::FPR32RegClass);
4000 if (VT.getSizeInBits() == 64)
4001 return std::make_pair(0U, &ARM64::FPR64RegClass);
4002 if (VT.getSizeInBits() == 128)
4003 return std::make_pair(0U, &ARM64::FPR128RegClass);
4005 // The instructions that this constraint is designed for can
4006 // only take 128-bit registers so just use that regclass.
4008 if (VT.getSizeInBits() == 128)
4009 return std::make_pair(0U, &ARM64::FPR128_loRegClass);
4013 if (StringRef("{cc}").equals_lower(Constraint))
4014 return std::make_pair(unsigned(ARM64::CPSR), &ARM64::CCRRegClass);
4016 // Use the default implementation in TargetLowering to convert the register
4017 // constraint into a member of a register class.
4018 std::pair<unsigned, const TargetRegisterClass *> Res;
4019 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4021 // Not found as a standard register?
4022 if (Res.second == 0) {
4023 unsigned Size = Constraint.size();
4024 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
4025 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
4026 const std::string Reg =
4027 std::string(&Constraint[2], &Constraint[Size - 1]);
4028 int RegNo = atoi(Reg.c_str());
4029 if (RegNo >= 0 && RegNo <= 31) {
4030 // v0 - v31 are aliases of q0 - q31.
4031 // By default we'll emit v0-v31 for this unless there's a modifier where
4032 // we'll emit the correct register as well.
4033 Res.first = ARM64::FPR128RegClass.getRegister(RegNo);
4034 Res.second = &ARM64::FPR128RegClass;
4042 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4043 /// vector. If it is invalid, don't add anything to Ops.
4044 void ARM64TargetLowering::LowerAsmOperandForConstraint(
4045 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
4046 SelectionDAG &DAG) const {
4047 SDValue Result(0, 0);
4049 // Currently only support length 1 constraints.
4050 if (Constraint.length() != 1)
4053 char ConstraintLetter = Constraint[0];
4054 switch (ConstraintLetter) {
4058 // This set of constraints deal with valid constants for various instructions.
4059 // Validate and return a target constant for them if we can.
4061 // 'z' maps to xzr or wzr so it needs an input of 0.
4062 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4063 if (!C || C->getZExtValue() != 0)
4066 if (Op.getValueType() == MVT::i64)
4067 Result = DAG.getRegister(ARM64::XZR, MVT::i64);
4069 Result = DAG.getRegister(ARM64::WZR, MVT::i32);
4079 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4083 // Grab the value and do some validation.
4084 uint64_t CVal = C->getZExtValue();
4085 switch (ConstraintLetter) {
4086 // The I constraint applies only to simple ADD or SUB immediate operands:
4087 // i.e. 0 to 4095 with optional shift by 12
4088 // The J constraint applies only to ADD or SUB immediates that would be
4089 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
4090 // instruction [or vice versa], in other words -1 to -4095 with optional
4091 // left shift by 12.
4093 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
4097 uint64_t NVal = -C->getSExtValue();
4098 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal))
4102 // The K and L constraints apply *only* to logical immediates, including
4103 // what used to be the MOVI alias for ORR (though the MOVI alias has now
4104 // been removed and MOV should be used). So these constraints have to
4105 // distinguish between bit patterns that are valid 32-bit or 64-bit
4106 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
4107 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
4110 if (ARM64_AM::isLogicalImmediate(CVal, 32))
4114 if (ARM64_AM::isLogicalImmediate(CVal, 64))
4117 // The M and N constraints are a superset of K and L respectively, for use
4118 // with the MOV (immediate) alias. As well as the logical immediates they
4119 // also match 32 or 64-bit immediates that can be loaded either using a
4120 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
4121 // (M) or 64-bit 0x1234000000000000 (N) etc.
4122 // As a note some of this code is liberally stolen from the asm parser.
4124 if (!isUInt<32>(CVal))
4126 if (ARM64_AM::isLogicalImmediate(CVal, 32))
4128 if ((CVal & 0xFFFF) == CVal)
4130 if ((CVal & 0xFFFF0000ULL) == CVal)
4132 uint64_t NCVal = ~(uint32_t)CVal;
4133 if ((NCVal & 0xFFFFULL) == NCVal)
4135 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4140 if (ARM64_AM::isLogicalImmediate(CVal, 64))
4142 if ((CVal & 0xFFFFULL) == CVal)
4144 if ((CVal & 0xFFFF0000ULL) == CVal)
4146 if ((CVal & 0xFFFF00000000ULL) == CVal)
4148 if ((CVal & 0xFFFF000000000000ULL) == CVal)
4150 uint64_t NCVal = ~CVal;
4151 if ((NCVal & 0xFFFFULL) == NCVal)
4153 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4155 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
4157 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
4165 // All assembler immediates are 64-bit integers.
4166 Result = DAG.getTargetConstant(CVal, MVT::i64);
4170 if (Result.getNode()) {
4171 Ops.push_back(Result);
4175 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4178 //===----------------------------------------------------------------------===//
4179 // ARM64 Advanced SIMD Support
4180 //===----------------------------------------------------------------------===//
4182 /// WidenVector - Given a value in the V64 register class, produce the
4183 /// equivalent value in the V128 register class.
4184 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
4185 EVT VT = V64Reg.getValueType();
4186 unsigned NarrowSize = VT.getVectorNumElements();
4187 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4188 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
4191 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
4192 V64Reg, DAG.getConstant(0, MVT::i32));
4195 /// getExtFactor - Determine the adjustment factor for the position when
4196 /// generating an "extract from vector registers" instruction.
4197 static unsigned getExtFactor(SDValue &V) {
4198 EVT EltType = V.getValueType().getVectorElementType();
4199 return EltType.getSizeInBits() / 8;
4202 /// NarrowVector - Given a value in the V128 register class, produce the
4203 /// equivalent value in the V64 register class.
4204 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
4205 EVT VT = V128Reg.getValueType();
4206 unsigned WideSize = VT.getVectorNumElements();
4207 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4208 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
4211 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, NarrowTy, V128Reg);
4214 // Gather data to see if the operation can be modelled as a
4215 // shuffle in combination with VEXTs.
4216 SDValue ARM64TargetLowering::ReconstructShuffle(SDValue Op,
4217 SelectionDAG &DAG) const {
4219 EVT VT = Op.getValueType();
4220 unsigned NumElts = VT.getVectorNumElements();
4222 SmallVector<SDValue, 2> SourceVecs;
4223 SmallVector<unsigned, 2> MinElts;
4224 SmallVector<unsigned, 2> MaxElts;
4226 for (unsigned i = 0; i < NumElts; ++i) {
4227 SDValue V = Op.getOperand(i);
4228 if (V.getOpcode() == ISD::UNDEF)
4230 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4231 // A shuffle can only come from building a vector from various
4232 // elements of other vectors.
4236 // Record this extraction against the appropriate vector if possible...
4237 SDValue SourceVec = V.getOperand(0);
4238 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4239 bool FoundSource = false;
4240 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4241 if (SourceVecs[j] == SourceVec) {
4242 if (MinElts[j] > EltNo)
4244 if (MaxElts[j] < EltNo)
4251 // Or record a new source if not...
4253 SourceVecs.push_back(SourceVec);
4254 MinElts.push_back(EltNo);
4255 MaxElts.push_back(EltNo);
4259 // Currently only do something sane when at most two source vectors
4261 if (SourceVecs.size() > 2)
4264 SDValue ShuffleSrcs[2] = { DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4265 int VEXTOffsets[2] = { 0, 0 };
4267 // This loop extracts the usage patterns of the source vectors
4268 // and prepares appropriate SDValues for a shuffle if possible.
4269 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4270 if (SourceVecs[i].getValueType() == VT) {
4271 // No VEXT necessary
4272 ShuffleSrcs[i] = SourceVecs[i];
4275 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4276 // It probably isn't worth padding out a smaller vector just to
4277 // break it down again in a shuffle.
4281 // Don't attempt to extract subvectors from BUILD_VECTOR sources
4282 // that expand or trunc the original value.
4283 // TODO: We can try to bitcast and ANY_EXTEND the result but
4284 // we need to consider the cost of vector ANY_EXTEND, and the
4285 // legality of all the types.
4286 if (SourceVecs[i].getValueType().getVectorElementType() !=
4287 VT.getVectorElementType())
4290 // Since only 64-bit and 128-bit vectors are legal on ARM and
4291 // we've eliminated the other cases...
4292 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2 * NumElts &&
4293 "unexpected vector sizes in ReconstructShuffle");
4295 if (MaxElts[i] - MinElts[i] >= NumElts) {
4296 // Span too large for a VEXT to cope
4300 if (MinElts[i] >= NumElts) {
4301 // The extraction can just take the second half
4302 VEXTOffsets[i] = NumElts;
4304 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
4305 DAG.getIntPtrConstant(NumElts));
4306 } else if (MaxElts[i] < NumElts) {
4307 // The extraction can just take the first half
4309 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4310 SourceVecs[i], DAG.getIntPtrConstant(0));
4312 // An actual VEXT is needed
4313 VEXTOffsets[i] = MinElts[i];
4314 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4315 SourceVecs[i], DAG.getIntPtrConstant(0));
4317 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
4318 DAG.getIntPtrConstant(NumElts));
4319 unsigned Imm = VEXTOffsets[i] * getExtFactor(VEXTSrc1);
4320 ShuffleSrcs[i] = DAG.getNode(ARM64ISD::EXT, dl, VT, VEXTSrc1, VEXTSrc2,
4321 DAG.getConstant(Imm, MVT::i32));
4325 SmallVector<int, 8> Mask;
4327 for (unsigned i = 0; i < NumElts; ++i) {
4328 SDValue Entry = Op.getOperand(i);
4329 if (Entry.getOpcode() == ISD::UNDEF) {
4334 SDValue ExtractVec = Entry.getOperand(0);
4336 cast<ConstantSDNode>(Op.getOperand(i).getOperand(1))->getSExtValue();
4337 if (ExtractVec == SourceVecs[0]) {
4338 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4340 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4344 // Final check before we try to produce nonsense...
4345 if (isShuffleMaskLegal(Mask, VT))
4346 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4352 // check if an EXT instruction can handle the shuffle mask when the
4353 // vector sources of the shuffle are the same.
4354 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4355 unsigned NumElts = VT.getVectorNumElements();
4357 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4363 // If this is a VEXT shuffle, the immediate value is the index of the first
4364 // element. The other shuffle indices must be the successive elements after
4366 unsigned ExpectedElt = Imm;
4367 for (unsigned i = 1; i < NumElts; ++i) {
4368 // Increment the expected index. If it wraps around, just follow it
4369 // back to index zero and keep going.
4371 if (ExpectedElt == NumElts)
4375 continue; // ignore UNDEF indices
4376 if (ExpectedElt != static_cast<unsigned>(M[i]))
4383 // check if an EXT instruction can handle the shuffle mask when the
4384 // vector sources of the shuffle are different.
4385 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
4387 unsigned NumElts = VT.getVectorNumElements();
4390 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4396 // If this is a VEXT shuffle, the immediate value is the index of the first
4397 // element. The other shuffle indices must be the successive elements after
4399 unsigned ExpectedElt = Imm;
4400 for (unsigned i = 1; i < NumElts; ++i) {
4401 // Increment the expected index. If it wraps around, it may still be
4402 // a VEXT but the source vectors must be swapped.
4404 if (ExpectedElt == NumElts * 2) {
4410 continue; // ignore UNDEF indices
4411 if (ExpectedElt != static_cast<unsigned>(M[i]))
4415 // Adjust the index value if the source operands will be swapped.
4422 /// isREVMask - Check if a vector shuffle corresponds to a REV
4423 /// instruction with the specified blocksize. (The order of the elements
4424 /// within each block of the vector is reversed.)
4425 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4426 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
4427 "Only possible block sizes for REV are: 16, 32, 64");
4429 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4433 unsigned NumElts = VT.getVectorNumElements();
4434 unsigned BlockElts = M[0] + 1;
4435 // If the first shuffle index is UNDEF, be optimistic.
4437 BlockElts = BlockSize / EltSz;
4439 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4442 for (unsigned i = 0; i < NumElts; ++i) {
4444 continue; // ignore UNDEF indices
4445 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
4452 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4453 unsigned NumElts = VT.getVectorNumElements();
4454 WhichResult = (M[0] == 0 ? 0 : 1);
4455 unsigned Idx = WhichResult * NumElts / 2;
4456 for (unsigned i = 0; i != NumElts; i += 2) {
4457 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4458 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
4466 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4467 unsigned NumElts = VT.getVectorNumElements();
4468 WhichResult = (M[0] == 0 ? 0 : 1);
4469 for (unsigned i = 0; i != NumElts; ++i) {
4471 continue; // ignore UNDEF indices
4472 if ((unsigned)M[i] != 2 * i + WhichResult)
4479 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4480 unsigned NumElts = VT.getVectorNumElements();
4481 WhichResult = (M[0] == 0 ? 0 : 1);
4482 for (unsigned i = 0; i < NumElts; i += 2) {
4483 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4484 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
4490 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
4491 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4492 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4493 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4494 unsigned NumElts = VT.getVectorNumElements();
4495 WhichResult = (M[0] == 0 ? 0 : 1);
4496 unsigned Idx = WhichResult * NumElts / 2;
4497 for (unsigned i = 0; i != NumElts; i += 2) {
4498 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4499 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
4507 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
4508 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4509 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4510 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4511 unsigned Half = VT.getVectorNumElements() / 2;
4512 WhichResult = (M[0] == 0 ? 0 : 1);
4513 for (unsigned j = 0; j != 2; ++j) {
4514 unsigned Idx = WhichResult;
4515 for (unsigned i = 0; i != Half; ++i) {
4516 int MIdx = M[i + j * Half];
4517 if (MIdx >= 0 && (unsigned)MIdx != Idx)
4526 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
4527 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4528 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4529 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4530 unsigned NumElts = VT.getVectorNumElements();
4531 WhichResult = (M[0] == 0 ? 0 : 1);
4532 for (unsigned i = 0; i < NumElts; i += 2) {
4533 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4534 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
4540 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4541 /// the specified operations to build the shuffle.
4542 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4543 SDValue RHS, SelectionDAG &DAG,
4545 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4546 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
4547 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
4550 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4559 OP_VUZPL, // VUZP, left result
4560 OP_VUZPR, // VUZP, right result
4561 OP_VZIPL, // VZIP, left result
4562 OP_VZIPR, // VZIP, right result
4563 OP_VTRNL, // VTRN, left result
4564 OP_VTRNR // VTRN, right result
4567 if (OpNum == OP_COPY) {
4568 if (LHSID == (1 * 9 + 2) * 9 + 3)
4570 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
4574 SDValue OpLHS, OpRHS;
4575 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4576 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4577 EVT VT = OpLHS.getValueType();
4581 llvm_unreachable("Unknown shuffle opcode!");
4583 // VREV divides the vector in half and swaps within the half.
4584 if (VT.getVectorElementType() == MVT::i32 ||
4585 VT.getVectorElementType() == MVT::f32)
4586 return DAG.getNode(ARM64ISD::REV64, dl, VT, OpLHS);
4587 // vrev <4 x i16> -> REV32
4588 if (VT.getVectorElementType() == MVT::i16)
4589 return DAG.getNode(ARM64ISD::REV32, dl, VT, OpLHS);
4590 // vrev <4 x i8> -> REV16
4591 assert(VT.getVectorElementType() == MVT::i8);
4592 return DAG.getNode(ARM64ISD::REV16, dl, VT, OpLHS);
4597 EVT EltTy = VT.getVectorElementType();
4599 if (EltTy == MVT::i8)
4600 Opcode = ARM64ISD::DUPLANE8;
4601 else if (EltTy == MVT::i16)
4602 Opcode = ARM64ISD::DUPLANE16;
4603 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
4604 Opcode = ARM64ISD::DUPLANE32;
4605 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
4606 Opcode = ARM64ISD::DUPLANE64;
4608 llvm_unreachable("Invalid vector element type?");
4610 if (VT.getSizeInBits() == 64)
4611 OpLHS = WidenVector(OpLHS, DAG);
4612 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, MVT::i64);
4613 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
4618 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
4619 return DAG.getNode(ARM64ISD::EXT, dl, VT, OpLHS, OpRHS,
4620 DAG.getConstant(Imm, MVT::i32));
4623 return DAG.getNode(ARM64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4625 return DAG.getNode(ARM64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4627 return DAG.getNode(ARM64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4629 return DAG.getNode(ARM64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4631 return DAG.getNode(ARM64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4633 return DAG.getNode(ARM64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4637 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
4638 SelectionDAG &DAG) {
4639 // Check to see if we can use the TBL instruction.
4640 SDValue V1 = Op.getOperand(0);
4641 SDValue V2 = Op.getOperand(1);
4644 EVT EltVT = Op.getValueType().getVectorElementType();
4645 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
4647 SmallVector<SDValue, 8> TBLMask;
4648 for (int Val : ShuffleMask) {
4649 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
4650 unsigned Offset = Byte + Val * BytesPerElt;
4651 TBLMask.push_back(DAG.getConstant(Offset, MVT::i32));
4655 MVT IndexVT = MVT::v8i8;
4656 unsigned IndexLen = 8;
4657 if (Op.getValueType().getSizeInBits() == 128) {
4658 IndexVT = MVT::v16i8;
4662 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
4663 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
4666 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
4668 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
4669 Shuffle = DAG.getNode(
4670 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4671 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4672 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4674 if (IndexLen == 8) {
4675 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
4676 Shuffle = DAG.getNode(
4677 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4678 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4679 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4681 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
4682 // cannot currently represent the register constraints on the input
4684 // Shuffle = DAG.getNode(ARM64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
4685 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4686 // &TBLMask[0], IndexLen));
4687 Shuffle = DAG.getNode(
4688 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4689 DAG.getConstant(Intrinsic::arm64_neon_tbl2, MVT::i32), V1Cst, V2Cst,
4690 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4693 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
4696 static unsigned getDUPLANEOp(EVT EltType) {
4697 if (EltType == MVT::i8)
4698 return ARM64ISD::DUPLANE8;
4699 if (EltType == MVT::i16)
4700 return ARM64ISD::DUPLANE16;
4701 if (EltType == MVT::i32 || EltType == MVT::f32)
4702 return ARM64ISD::DUPLANE32;
4703 if (EltType == MVT::i64 || EltType == MVT::f64)
4704 return ARM64ISD::DUPLANE64;
4706 llvm_unreachable("Invalid vector element type?");
4709 SDValue ARM64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4710 SelectionDAG &DAG) const {
4712 EVT VT = Op.getValueType();
4714 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4716 // Convert shuffles that are directly supported on NEON to target-specific
4717 // DAG nodes, instead of keeping them as shuffles and matching them again
4718 // during code selection. This is more efficient and avoids the possibility
4719 // of inconsistencies between legalization and selection.
4720 ArrayRef<int> ShuffleMask = SVN->getMask();
4722 SDValue V1 = Op.getOperand(0);
4723 SDValue V2 = Op.getOperand(1);
4725 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
4726 V1.getValueType().getSimpleVT())) {
4727 int Lane = SVN->getSplatIndex();
4728 // If this is undef splat, generate it via "just" vdup, if possible.
4732 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
4733 return DAG.getNode(ARM64ISD::DUP, dl, V1.getValueType(),
4735 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
4736 // constant. If so, we can just reference the lane's definition directly.
4737 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
4738 !isa<ConstantSDNode>(V1.getOperand(Lane)))
4739 return DAG.getNode(ARM64ISD::DUP, dl, VT, V1.getOperand(Lane));
4741 // Otherwise, duplicate from the lane of the input vector.
4742 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
4744 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
4745 // to make a vector of the same size as this SHUFFLE. We can ignore the
4746 // extract entirely, and canonicalise the concat using WidenVector.
4747 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
4748 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
4749 V1 = V1.getOperand(0);
4750 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
4751 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
4752 Lane -= Idx * VT.getVectorNumElements() / 2;
4753 V1 = WidenVector(V1.getOperand(Idx), DAG);
4754 } else if (VT.getSizeInBits() == 64)
4755 V1 = WidenVector(V1, DAG);
4757 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, MVT::i64));
4760 if (isREVMask(ShuffleMask, VT, 64))
4761 return DAG.getNode(ARM64ISD::REV64, dl, V1.getValueType(), V1, V2);
4762 if (isREVMask(ShuffleMask, VT, 32))
4763 return DAG.getNode(ARM64ISD::REV32, dl, V1.getValueType(), V1, V2);
4764 if (isREVMask(ShuffleMask, VT, 16))
4765 return DAG.getNode(ARM64ISD::REV16, dl, V1.getValueType(), V1, V2);
4767 bool ReverseEXT = false;
4769 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
4772 Imm *= getExtFactor(V1);
4773 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V2,
4774 DAG.getConstant(Imm, MVT::i32));
4775 } else if (V2->getOpcode() == ISD::UNDEF &&
4776 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
4777 Imm *= getExtFactor(V1);
4778 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V1,
4779 DAG.getConstant(Imm, MVT::i32));
4782 unsigned WhichResult;
4783 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
4784 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4785 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4787 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
4788 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4789 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4791 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
4792 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4793 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4796 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4797 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4798 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4800 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4801 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4802 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4804 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4805 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4806 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4809 // If the shuffle is not directly supported and it has 4 elements, use
4810 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4811 unsigned NumElts = VT.getVectorNumElements();
4813 unsigned PFIndexes[4];
4814 for (unsigned i = 0; i != 4; ++i) {
4815 if (ShuffleMask[i] < 0)
4818 PFIndexes[i] = ShuffleMask[i];
4821 // Compute the index in the perfect shuffle table.
4822 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
4823 PFIndexes[2] * 9 + PFIndexes[3];
4824 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4825 unsigned Cost = (PFEntry >> 30);
4828 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4831 return GenerateTBL(Op, ShuffleMask, DAG);
4834 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
4836 EVT VT = BVN->getValueType(0);
4837 APInt SplatBits, SplatUndef;
4838 unsigned SplatBitSize;
4840 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4841 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
4843 for (unsigned i = 0; i < NumSplats; ++i) {
4844 CnstBits <<= SplatBitSize;
4845 UndefBits <<= SplatBitSize;
4846 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
4847 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
4856 SDValue ARM64TargetLowering::LowerVectorAND(SDValue Op,
4857 SelectionDAG &DAG) const {
4858 BuildVectorSDNode *BVN =
4859 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
4860 SDValue LHS = Op.getOperand(0);
4862 EVT VT = Op.getValueType();
4867 APInt CnstBits(VT.getSizeInBits(), 0);
4868 APInt UndefBits(VT.getSizeInBits(), 0);
4869 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
4870 // We only have BIC vector immediate instruction, which is and-not.
4871 CnstBits = ~CnstBits;
4873 // We make use of a little bit of goto ickiness in order to avoid having to
4874 // duplicate the immediate matching logic for the undef toggled case.
4875 bool SecondTry = false;
4878 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
4879 CnstBits = CnstBits.zextOrTrunc(64);
4880 uint64_t CnstVal = CnstBits.getZExtValue();
4882 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
4883 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
4884 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4885 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4886 DAG.getConstant(CnstVal, MVT::i32),
4887 DAG.getConstant(0, MVT::i32));
4888 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4891 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
4892 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
4893 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4894 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4895 DAG.getConstant(CnstVal, MVT::i32),
4896 DAG.getConstant(8, MVT::i32));
4897 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4900 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
4901 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
4902 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4903 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4904 DAG.getConstant(CnstVal, MVT::i32),
4905 DAG.getConstant(16, MVT::i32));
4906 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4909 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
4910 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
4911 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4912 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4913 DAG.getConstant(CnstVal, MVT::i32),
4914 DAG.getConstant(24, MVT::i32));
4915 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4918 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
4919 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
4920 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4921 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4922 DAG.getConstant(CnstVal, MVT::i32),
4923 DAG.getConstant(0, MVT::i32));
4924 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4927 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
4928 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
4929 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4930 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4931 DAG.getConstant(CnstVal, MVT::i32),
4932 DAG.getConstant(8, MVT::i32));
4933 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4940 CnstBits = ~UndefBits;
4944 // We can always fall back to a non-immediate AND.
4949 // Specialized code to quickly find if PotentialBVec is a BuildVector that
4950 // consists of only the same constant int value, returned in reference arg
4952 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
4953 uint64_t &ConstVal) {
4954 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
4957 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
4960 EVT VT = Bvec->getValueType(0);
4961 unsigned NumElts = VT.getVectorNumElements();
4962 for (unsigned i = 1; i < NumElts; ++i)
4963 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
4965 ConstVal = FirstElt->getZExtValue();
4969 static unsigned getIntrinsicID(const SDNode *N) {
4970 unsigned Opcode = N->getOpcode();
4973 return Intrinsic::not_intrinsic;
4974 case ISD::INTRINSIC_WO_CHAIN: {
4975 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4976 if (IID < Intrinsic::num_intrinsics)
4978 return Intrinsic::not_intrinsic;
4983 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
4984 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
4985 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
4986 // Also, logical shift right -> sri, with the same structure.
4987 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
4988 EVT VT = N->getValueType(0);
4995 // Is the first op an AND?
4996 const SDValue And = N->getOperand(0);
4997 if (And.getOpcode() != ISD::AND)
5000 // Is the second op an shl or lshr?
5001 SDValue Shift = N->getOperand(1);
5002 // This will have been turned into: ARM64ISD::VSHL vector, #shift
5003 // or ARM64ISD::VLSHR vector, #shift
5004 unsigned ShiftOpc = Shift.getOpcode();
5005 if ((ShiftOpc != ARM64ISD::VSHL && ShiftOpc != ARM64ISD::VLSHR))
5007 bool IsShiftRight = ShiftOpc == ARM64ISD::VLSHR;
5009 // Is the shift amount constant?
5010 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
5014 // Is the and mask vector all constant?
5016 if (!isAllConstantBuildVector(And.getOperand(1), C1))
5019 // Is C1 == ~C2, taking into account how much one can shift elements of a
5021 uint64_t C2 = C2node->getZExtValue();
5022 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
5023 if (C2 > ElemSizeInBits)
5025 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
5026 if ((C1 & ElemMask) != (~C2 & ElemMask))
5029 SDValue X = And.getOperand(0);
5030 SDValue Y = Shift.getOperand(0);
5033 IsShiftRight ? Intrinsic::arm64_neon_vsri : Intrinsic::arm64_neon_vsli;
5035 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5036 DAG.getConstant(Intrin, MVT::i32), X, Y, Shift.getOperand(1));
5038 DEBUG(dbgs() << "arm64-lower: transformed: \n");
5039 DEBUG(N->dump(&DAG));
5040 DEBUG(dbgs() << "into: \n");
5041 DEBUG(ResultSLI->dump(&DAG));
5047 SDValue ARM64TargetLowering::LowerVectorOR(SDValue Op,
5048 SelectionDAG &DAG) const {
5049 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
5050 if (EnableARM64SlrGeneration) {
5051 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
5056 BuildVectorSDNode *BVN =
5057 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
5058 SDValue LHS = Op.getOperand(1);
5060 EVT VT = Op.getValueType();
5062 // OR commutes, so try swapping the operands.
5064 LHS = Op.getOperand(0);
5065 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5070 APInt CnstBits(VT.getSizeInBits(), 0);
5071 APInt UndefBits(VT.getSizeInBits(), 0);
5072 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5073 // We make use of a little bit of goto ickiness in order to avoid having to
5074 // duplicate the immediate matching logic for the undef toggled case.
5075 bool SecondTry = false;
5078 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5079 CnstBits = CnstBits.zextOrTrunc(64);
5080 uint64_t CnstVal = CnstBits.getZExtValue();
5082 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5083 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5084 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5085 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5086 DAG.getConstant(CnstVal, MVT::i32),
5087 DAG.getConstant(0, MVT::i32));
5088 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5091 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5092 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5093 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5094 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5095 DAG.getConstant(CnstVal, MVT::i32),
5096 DAG.getConstant(8, MVT::i32));
5097 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5100 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5101 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5102 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5103 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5104 DAG.getConstant(CnstVal, MVT::i32),
5105 DAG.getConstant(16, MVT::i32));
5106 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5109 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5110 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5111 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5112 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5113 DAG.getConstant(CnstVal, MVT::i32),
5114 DAG.getConstant(24, MVT::i32));
5115 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5118 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5119 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5120 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5121 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5122 DAG.getConstant(CnstVal, MVT::i32),
5123 DAG.getConstant(0, MVT::i32));
5124 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5127 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5128 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5129 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5130 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5131 DAG.getConstant(CnstVal, MVT::i32),
5132 DAG.getConstant(8, MVT::i32));
5133 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5140 CnstBits = UndefBits;
5144 // We can always fall back to a non-immediate OR.
5149 SDValue ARM64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
5150 SelectionDAG &DAG) const {
5151 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5153 EVT VT = Op.getValueType();
5155 APInt CnstBits(VT.getSizeInBits(), 0);
5156 APInt UndefBits(VT.getSizeInBits(), 0);
5157 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5158 // We make use of a little bit of goto ickiness in order to avoid having to
5159 // duplicate the immediate matching logic for the undef toggled case.
5160 bool SecondTry = false;
5163 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5164 CnstBits = CnstBits.zextOrTrunc(64);
5165 uint64_t CnstVal = CnstBits.getZExtValue();
5167 // Certain magic vector constants (used to express things like NOT
5168 // and NEG) are passed through unmodified. This allows codegen patterns
5169 // for these operations to match. Special-purpose patterns will lower
5170 // these immediates to MOVIs if it proves necessary.
5171 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
5174 // The many faces of MOVI...
5175 if (ARM64_AM::isAdvSIMDModImmType10(CnstVal)) {
5176 CnstVal = ARM64_AM::encodeAdvSIMDModImmType10(CnstVal);
5177 if (VT.getSizeInBits() == 128) {
5178 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::v2i64,
5179 DAG.getConstant(CnstVal, MVT::i32));
5180 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5183 // Support the V64 version via subregister insertion.
5184 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::f64,
5185 DAG.getConstant(CnstVal, MVT::i32));
5186 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5189 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5190 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5191 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5192 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5193 DAG.getConstant(CnstVal, MVT::i32),
5194 DAG.getConstant(0, MVT::i32));
5195 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5198 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5199 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5200 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5201 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5202 DAG.getConstant(CnstVal, MVT::i32),
5203 DAG.getConstant(8, MVT::i32));
5204 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5207 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5208 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5209 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5210 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5211 DAG.getConstant(CnstVal, MVT::i32),
5212 DAG.getConstant(16, MVT::i32));
5213 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5216 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5217 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5218 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5219 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5220 DAG.getConstant(CnstVal, MVT::i32),
5221 DAG.getConstant(24, MVT::i32));
5222 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5225 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5226 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5227 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5228 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5229 DAG.getConstant(CnstVal, MVT::i32),
5230 DAG.getConstant(0, MVT::i32));
5231 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5234 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5235 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5236 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5237 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5238 DAG.getConstant(CnstVal, MVT::i32),
5239 DAG.getConstant(8, MVT::i32));
5240 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5243 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
5244 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
5245 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5246 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
5247 DAG.getConstant(CnstVal, MVT::i32),
5248 DAG.getConstant(264, MVT::i32));
5249 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5252 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
5253 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
5254 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5255 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
5256 DAG.getConstant(CnstVal, MVT::i32),
5257 DAG.getConstant(272, MVT::i32));
5258 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5261 if (ARM64_AM::isAdvSIMDModImmType9(CnstVal)) {
5262 CnstVal = ARM64_AM::encodeAdvSIMDModImmType9(CnstVal);
5263 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
5264 SDValue Mov = DAG.getNode(ARM64ISD::MOVI, dl, MovTy,
5265 DAG.getConstant(CnstVal, MVT::i32));
5266 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5269 // The few faces of FMOV...
5270 if (ARM64_AM::isAdvSIMDModImmType11(CnstVal)) {
5271 CnstVal = ARM64_AM::encodeAdvSIMDModImmType11(CnstVal);
5272 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
5273 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MovTy,
5274 DAG.getConstant(CnstVal, MVT::i32));
5275 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5278 if (ARM64_AM::isAdvSIMDModImmType12(CnstVal) &&
5279 VT.getSizeInBits() == 128) {
5280 CnstVal = ARM64_AM::encodeAdvSIMDModImmType12(CnstVal);
5281 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MVT::v2f64,
5282 DAG.getConstant(CnstVal, MVT::i32));
5283 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5286 // The many faces of MVNI...
5288 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5289 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5290 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5291 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5292 DAG.getConstant(CnstVal, MVT::i32),
5293 DAG.getConstant(0, MVT::i32));
5294 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5297 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5298 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5299 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5300 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5301 DAG.getConstant(CnstVal, MVT::i32),
5302 DAG.getConstant(8, MVT::i32));
5303 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5306 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5307 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5308 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5309 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5310 DAG.getConstant(CnstVal, MVT::i32),
5311 DAG.getConstant(16, MVT::i32));
5312 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5315 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5316 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5317 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5318 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5319 DAG.getConstant(CnstVal, MVT::i32),
5320 DAG.getConstant(24, MVT::i32));
5321 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5324 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5325 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5326 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5327 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5328 DAG.getConstant(CnstVal, MVT::i32),
5329 DAG.getConstant(0, MVT::i32));
5330 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5333 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5334 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5335 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5336 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5337 DAG.getConstant(CnstVal, MVT::i32),
5338 DAG.getConstant(8, MVT::i32));
5339 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5342 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
5343 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
5344 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5345 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
5346 DAG.getConstant(CnstVal, MVT::i32),
5347 DAG.getConstant(264, MVT::i32));
5348 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5351 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
5352 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
5353 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5354 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
5355 DAG.getConstant(CnstVal, MVT::i32),
5356 DAG.getConstant(272, MVT::i32));
5357 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5364 CnstBits = UndefBits;
5369 // Scan through the operands to find some interesting properties we can
5371 // 1) If only one value is used, we can use a DUP, or
5372 // 2) if only the low element is not undef, we can just insert that, or
5373 // 3) if only one constant value is used (w/ some non-constant lanes),
5374 // we can splat the constant value into the whole vector then fill
5375 // in the non-constant lanes.
5376 // 4) FIXME: If different constant values are used, but we can intelligently
5377 // select the values we'll be overwriting for the non-constant
5378 // lanes such that we can directly materialize the vector
5379 // some other way (MOVI, e.g.), we can be sneaky.
5380 unsigned NumElts = VT.getVectorNumElements();
5381 bool isOnlyLowElement = true;
5382 bool usesOnlyOneValue = true;
5383 bool usesOnlyOneConstantValue = true;
5384 bool isConstant = true;
5385 unsigned NumConstantLanes = 0;
5387 SDValue ConstantValue;
5388 for (unsigned i = 0; i < NumElts; ++i) {
5389 SDValue V = Op.getOperand(i);
5390 if (V.getOpcode() == ISD::UNDEF)
5393 isOnlyLowElement = false;
5394 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5397 if (isa<ConstantSDNode>(V)) {
5399 if (!ConstantValue.getNode())
5401 else if (ConstantValue != V)
5402 usesOnlyOneConstantValue = false;
5405 if (!Value.getNode())
5407 else if (V != Value)
5408 usesOnlyOneValue = false;
5411 if (!Value.getNode())
5412 return DAG.getUNDEF(VT);
5414 if (isOnlyLowElement)
5415 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5417 // Use DUP for non-constant splats. For f32 constant splats, reduce to
5418 // i32 and try again.
5419 if (usesOnlyOneValue) {
5421 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5422 Value.getValueType() != VT)
5423 return DAG.getNode(ARM64ISD::DUP, dl, VT, Value);
5425 // This is actually a DUPLANExx operation, which keeps everything vectory.
5427 // DUPLANE works on 128-bit vectors, widen it if necessary.
5428 SDValue Lane = Value.getOperand(1);
5429 Value = Value.getOperand(0);
5430 if (Value.getValueType().getSizeInBits() == 64)
5431 Value = WidenVector(Value, DAG);
5433 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
5434 return DAG.getNode(Opcode, dl, VT, Value, Lane);
5437 if (VT.getVectorElementType().isFloatingPoint()) {
5438 SmallVector<SDValue, 8> Ops;
5440 (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
5441 for (unsigned i = 0; i < NumElts; ++i)
5442 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
5443 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
5444 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
5445 Val = LowerBUILD_VECTOR(Val, DAG);
5447 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5451 // If there was only one constant value used and for more than one lane,
5452 // start by splatting that value, then replace the non-constant lanes. This
5453 // is better than the default, which will perform a separate initialization
5455 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
5456 SDValue Val = DAG.getNode(ARM64ISD::DUP, dl, VT, ConstantValue);
5457 // Now insert the non-constant lanes.
5458 for (unsigned i = 0; i < NumElts; ++i) {
5459 SDValue V = Op.getOperand(i);
5460 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5461 if (!isa<ConstantSDNode>(V)) {
5462 // Note that type legalization likely mucked about with the VT of the
5463 // source operand, so we may have to convert it here before inserting.
5464 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
5470 // If all elements are constants and the case above didn't get hit, fall back
5471 // to the default expansion, which will generate a load from the constant
5476 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5478 SDValue shuffle = ReconstructShuffle(Op, DAG);
5479 if (shuffle != SDValue())
5483 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5484 // know the default expansion would otherwise fall back on something even
5485 // worse. For a vector with one or two non-undef values, that's
5486 // scalar_to_vector for the elements followed by a shuffle (provided the
5487 // shuffle is valid for the target) and materialization element by element
5488 // on the stack followed by a load for everything else.
5489 if (!isConstant && !usesOnlyOneValue) {
5490 SDValue Vec = DAG.getUNDEF(VT);
5491 SDValue Op0 = Op.getOperand(0);
5492 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
5494 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
5495 // a) Avoid a RMW dependency on the full vector register, and
5496 // b) Allow the register coalescer to fold away the copy if the
5497 // value is already in an S or D register.
5498 if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
5499 unsigned SubIdx = ElemSize == 32 ? ARM64::ssub : ARM64::dsub;
5501 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
5502 DAG.getTargetConstant(SubIdx, MVT::i32));
5503 Vec = SDValue(N, 0);
5506 for (; i < NumElts; ++i) {
5507 SDValue V = Op.getOperand(i);
5508 if (V.getOpcode() == ISD::UNDEF)
5510 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5511 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5516 // Just use the default expansion. We failed to find a better alternative.
5520 SDValue ARM64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
5521 SelectionDAG &DAG) const {
5522 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
5524 // Check for non-constant lane.
5525 if (!isa<ConstantSDNode>(Op.getOperand(2)))
5528 EVT VT = Op.getOperand(0).getValueType();
5530 // Insertion/extraction are legal for V128 types.
5531 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5532 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5535 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5536 VT != MVT::v1i64 && VT != MVT::v2f32)
5539 // For V64 types, we perform insertion by expanding the value
5540 // to a V128 type and perform the insertion on that.
5542 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5543 EVT WideTy = WideVec.getValueType();
5545 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
5546 Op.getOperand(1), Op.getOperand(2));
5547 // Re-narrow the resultant vector.
5548 return NarrowVector(Node, DAG);
5551 SDValue ARM64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5552 SelectionDAG &DAG) const {
5553 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
5555 // Check for non-constant lane.
5556 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5559 EVT VT = Op.getOperand(0).getValueType();
5561 // Insertion/extraction are legal for V128 types.
5562 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5563 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5566 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5567 VT != MVT::v1i64 && VT != MVT::v2f32)
5570 // For V64 types, we perform extraction by expanding the value
5571 // to a V128 type and perform the extraction on that.
5573 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5574 EVT WideTy = WideVec.getValueType();
5576 EVT ExtrTy = WideTy.getVectorElementType();
5577 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
5580 // For extractions, we just return the result directly.
5581 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
5585 SDValue ARM64TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
5586 SelectionDAG &DAG) const {
5587 assert(Op.getOpcode() == ISD::SCALAR_TO_VECTOR && "Unknown opcode!");
5588 // Some AdvSIMD intrinsics leave their results in the scalar B/H/S/D
5589 // registers. The default lowering will copy those to a GPR then back
5590 // to a vector register. Instead, just recognize those cases and reference
5591 // the vector register they're already a subreg of.
5592 SDValue Op0 = Op->getOperand(0);
5593 if (Op0->getOpcode() != ISD::INTRINSIC_WO_CHAIN)
5595 unsigned IID = getIntrinsicID(Op0.getNode());
5596 // The below list of intrinsics isn't exhaustive. Add cases as-needed.
5597 // FIXME: Even better would be if there were an attribute on the node
5598 // that we could query and set in the intrinsics definition or something.
5602 // Early exit if this isn't one of the intrinsics we handle.
5604 case Intrinsic::arm64_neon_uaddv:
5605 case Intrinsic::arm64_neon_saddv:
5606 case Intrinsic::arm64_neon_uaddlv:
5607 case Intrinsic::arm64_neon_saddlv:
5608 switch (Op0.getValueType().getSizeInBits()) {
5610 llvm_unreachable("Illegal result size from ARM64 vector intrinsic!");
5612 SubIdx = ARM64::bsub;
5615 SubIdx = ARM64::hsub;
5618 SubIdx = ARM64::ssub;
5621 SubIdx = ARM64::dsub;
5626 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, SDLoc(Op),
5627 Op.getValueType(), DAG.getUNDEF(Op.getValueType()),
5628 Op0, DAG.getTargetConstant(SubIdx, MVT::i32));
5629 return SDValue(N, 0);
5632 SDValue ARM64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
5633 SelectionDAG &DAG) const {
5634 EVT VT = Op.getOperand(0).getValueType();
5640 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5643 unsigned Val = Cst->getZExtValue();
5645 unsigned Size = Op.getValueType().getSizeInBits();
5649 return DAG.getTargetExtractSubreg(ARM64::bsub, dl, Op.getValueType(),
5652 return DAG.getTargetExtractSubreg(ARM64::hsub, dl, Op.getValueType(),
5655 return DAG.getTargetExtractSubreg(ARM64::ssub, dl, Op.getValueType(),
5658 return DAG.getTargetExtractSubreg(ARM64::dsub, dl, Op.getValueType(),
5661 llvm_unreachable("Unexpected vector type in extract_subvector!");
5664 // If this is extracting the upper 64-bits of a 128-bit vector, we match
5666 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
5672 bool ARM64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5674 if (VT.getVectorNumElements() == 4 &&
5675 (VT.is128BitVector() || VT.is64BitVector())) {
5676 unsigned PFIndexes[4];
5677 for (unsigned i = 0; i != 4; ++i) {
5681 PFIndexes[i] = M[i];
5684 // Compute the index in the perfect shuffle table.
5685 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5686 PFIndexes[2] * 9 + PFIndexes[3];
5687 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5688 unsigned Cost = (PFEntry >> 30);
5695 unsigned Imm, WhichResult;
5697 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
5698 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
5699 isEXTMask(M, VT, ReverseVEXT, Imm) ||
5700 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
5701 isTRNMask(M, VT, WhichResult) || isUZPMask(M, VT, WhichResult) ||
5702 isZIPMask(M, VT, WhichResult) ||
5703 isTRN_v_undef_Mask(M, VT, WhichResult) ||
5704 isUZP_v_undef_Mask(M, VT, WhichResult) ||
5705 isZIP_v_undef_Mask(M, VT, WhichResult));
5708 /// getVShiftImm - Check if this is a valid build_vector for the immediate
5709 /// operand of a vector shift operation, where all the elements of the
5710 /// build_vector must have the same constant integer value.
5711 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
5712 // Ignore bit_converts.
5713 while (Op.getOpcode() == ISD::BITCAST)
5714 Op = Op.getOperand(0);
5715 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
5716 APInt SplatBits, SplatUndef;
5717 unsigned SplatBitSize;
5719 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
5720 HasAnyUndefs, ElementBits) ||
5721 SplatBitSize > ElementBits)
5723 Cnt = SplatBits.getSExtValue();
5727 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
5728 /// operand of a vector shift left operation. That value must be in the range:
5729 /// 0 <= Value < ElementBits for a left shift; or
5730 /// 0 <= Value <= ElementBits for a long left shift.
5731 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
5732 assert(VT.isVector() && "vector shift count is not a vector type");
5733 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5734 if (!getVShiftImm(Op, ElementBits, Cnt))
5736 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
5739 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
5740 /// operand of a vector shift right operation. For a shift opcode, the value
5741 /// is positive, but for an intrinsic the value count must be negative. The
5742 /// absolute value must be in the range:
5743 /// 1 <= |Value| <= ElementBits for a right shift; or
5744 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
5745 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
5747 assert(VT.isVector() && "vector shift count is not a vector type");
5748 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5749 if (!getVShiftImm(Op, ElementBits, Cnt))
5753 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
5756 SDValue ARM64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
5757 SelectionDAG &DAG) const {
5758 EVT VT = Op.getValueType();
5762 if (!Op.getOperand(1).getValueType().isVector())
5764 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5766 switch (Op.getOpcode()) {
5768 llvm_unreachable("unexpected shift opcode");
5771 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
5772 return DAG.getNode(ARM64ISD::VSHL, SDLoc(Op), VT, Op.getOperand(0),
5773 DAG.getConstant(Cnt, MVT::i32));
5774 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5775 DAG.getConstant(Intrinsic::arm64_neon_ushl, MVT::i32),
5776 Op.getOperand(0), Op.getOperand(1));
5779 // Right shift immediate
5780 if (isVShiftRImm(Op.getOperand(1), VT, false, false, Cnt) &&
5783 (Op.getOpcode() == ISD::SRA) ? ARM64ISD::VASHR : ARM64ISD::VLSHR;
5784 return DAG.getNode(Opc, SDLoc(Op), VT, Op.getOperand(0),
5785 DAG.getConstant(Cnt, MVT::i32));
5788 // Right shift register. Note, there is not a shift right register
5789 // instruction, but the shift left register instruction takes a signed
5790 // value, where negative numbers specify a right shift.
5791 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::arm64_neon_sshl
5792 : Intrinsic::arm64_neon_ushl;
5793 // negate the shift amount
5794 SDValue NegShift = DAG.getNode(ARM64ISD::NEG, DL, VT, Op.getOperand(1));
5795 SDValue NegShiftLeft =
5796 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5797 DAG.getConstant(Opc, MVT::i32), Op.getOperand(0), NegShift);
5798 return NegShiftLeft;
5804 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
5805 ARM64CC::CondCode CC, bool NoNans, EVT VT,
5806 SDLoc dl, SelectionDAG &DAG) {
5807 EVT SrcVT = LHS.getValueType();
5809 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
5810 APInt CnstBits(VT.getSizeInBits(), 0);
5811 APInt UndefBits(VT.getSizeInBits(), 0);
5812 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
5813 bool IsZero = IsCnst && (CnstBits == 0);
5815 if (SrcVT.getVectorElementType().isFloatingPoint()) {
5822 Fcmeq = DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5824 Fcmeq = DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5825 return DAG.getNode(ARM64ISD::NOT, dl, VT, Fcmeq);
5829 return DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5830 return DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5833 return DAG.getNode(ARM64ISD::FCMGEz, dl, VT, LHS);
5834 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, LHS, RHS);
5837 return DAG.getNode(ARM64ISD::FCMGTz, dl, VT, LHS);
5838 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, LHS, RHS);
5841 return DAG.getNode(ARM64ISD::FCMLEz, dl, VT, LHS);
5842 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, RHS, LHS);
5846 // If we ignore NaNs then we can use to the MI implementation.
5850 return DAG.getNode(ARM64ISD::FCMLTz, dl, VT, LHS);
5851 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, RHS, LHS);
5861 Cmeq = DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5863 Cmeq = DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5864 return DAG.getNode(ARM64ISD::NOT, dl, VT, Cmeq);
5868 return DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5869 return DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5872 return DAG.getNode(ARM64ISD::CMGEz, dl, VT, LHS);
5873 return DAG.getNode(ARM64ISD::CMGE, dl, VT, LHS, RHS);
5876 return DAG.getNode(ARM64ISD::CMGTz, dl, VT, LHS);
5877 return DAG.getNode(ARM64ISD::CMGT, dl, VT, LHS, RHS);
5880 return DAG.getNode(ARM64ISD::CMLEz, dl, VT, LHS);
5881 return DAG.getNode(ARM64ISD::CMGE, dl, VT, RHS, LHS);
5883 return DAG.getNode(ARM64ISD::CMHS, dl, VT, RHS, LHS);
5885 return DAG.getNode(ARM64ISD::CMHI, dl, VT, RHS, LHS);
5888 return DAG.getNode(ARM64ISD::CMLTz, dl, VT, LHS);
5889 return DAG.getNode(ARM64ISD::CMGT, dl, VT, RHS, LHS);
5891 return DAG.getNode(ARM64ISD::CMHI, dl, VT, LHS, RHS);
5893 return DAG.getNode(ARM64ISD::CMHS, dl, VT, LHS, RHS);
5897 SDValue ARM64TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
5898 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5899 SDValue LHS = Op.getOperand(0);
5900 SDValue RHS = Op.getOperand(1);
5903 if (LHS.getValueType().getVectorElementType().isInteger()) {
5904 assert(LHS.getValueType() == RHS.getValueType());
5905 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
5906 return EmitVectorComparison(LHS, RHS, ARM64CC, false, Op.getValueType(), dl,
5910 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
5911 LHS.getValueType().getVectorElementType() == MVT::f64);
5913 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
5914 // clean. Some of them require two branches to implement.
5915 ARM64CC::CondCode CC1, CC2;
5916 changeFPCCToARM64CC(CC, CC1, CC2);
5918 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
5920 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, Op.getValueType(), dl, DAG);
5921 if (!Cmp1.getNode())
5924 if (CC2 != ARM64CC::AL) {
5926 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, Op.getValueType(), dl, DAG);
5927 if (!Cmp2.getNode())
5930 return DAG.getNode(ISD::OR, dl, Cmp1.getValueType(), Cmp1, Cmp2);
5936 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
5937 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
5938 /// specified in the intrinsic calls.
5939 bool ARM64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5941 unsigned Intrinsic) const {
5942 switch (Intrinsic) {
5943 case Intrinsic::arm64_neon_ld2:
5944 case Intrinsic::arm64_neon_ld3:
5945 case Intrinsic::arm64_neon_ld4:
5946 case Intrinsic::arm64_neon_ld2lane:
5947 case Intrinsic::arm64_neon_ld3lane:
5948 case Intrinsic::arm64_neon_ld4lane:
5949 case Intrinsic::arm64_neon_ld2r:
5950 case Intrinsic::arm64_neon_ld3r:
5951 case Intrinsic::arm64_neon_ld4r: {
5952 Info.opc = ISD::INTRINSIC_W_CHAIN;
5953 // Conservatively set memVT to the entire set of vectors loaded.
5954 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
5955 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5956 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5959 Info.vol = false; // volatile loads with NEON intrinsics not supported
5960 Info.readMem = true;
5961 Info.writeMem = false;
5964 case Intrinsic::arm64_neon_st2:
5965 case Intrinsic::arm64_neon_st3:
5966 case Intrinsic::arm64_neon_st4:
5967 case Intrinsic::arm64_neon_st2lane:
5968 case Intrinsic::arm64_neon_st3lane:
5969 case Intrinsic::arm64_neon_st4lane: {
5970 Info.opc = ISD::INTRINSIC_VOID;
5971 // Conservatively set memVT to the entire set of vectors stored.
5972 unsigned NumElts = 0;
5973 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5974 Type *ArgTy = I.getArgOperand(ArgI)->getType();
5975 if (!ArgTy->isVectorTy())
5977 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
5979 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5980 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5983 Info.vol = false; // volatile stores with NEON intrinsics not supported
5984 Info.readMem = false;
5985 Info.writeMem = true;
5988 case Intrinsic::arm64_ldxr: {
5989 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
5990 Info.opc = ISD::INTRINSIC_W_CHAIN;
5991 Info.memVT = MVT::getVT(PtrTy->getElementType());
5992 Info.ptrVal = I.getArgOperand(0);
5994 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
5996 Info.readMem = true;
5997 Info.writeMem = false;
6000 case Intrinsic::arm64_stxr: {
6001 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
6002 Info.opc = ISD::INTRINSIC_W_CHAIN;
6003 Info.memVT = MVT::getVT(PtrTy->getElementType());
6004 Info.ptrVal = I.getArgOperand(1);
6006 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
6008 Info.readMem = false;
6009 Info.writeMem = true;
6012 case Intrinsic::arm64_ldxp: {
6013 Info.opc = ISD::INTRINSIC_W_CHAIN;
6014 Info.memVT = MVT::i128;
6015 Info.ptrVal = I.getArgOperand(0);
6019 Info.readMem = true;
6020 Info.writeMem = false;
6023 case Intrinsic::arm64_stxp: {
6024 Info.opc = ISD::INTRINSIC_W_CHAIN;
6025 Info.memVT = MVT::i128;
6026 Info.ptrVal = I.getArgOperand(2);
6030 Info.readMem = false;
6031 Info.writeMem = true;
6041 // Truncations from 64-bit GPR to 32-bit GPR is free.
6042 bool ARM64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
6043 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6045 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6046 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6047 if (NumBits1 <= NumBits2)
6051 bool ARM64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
6052 if (!VT1.isInteger() || !VT2.isInteger())
6054 unsigned NumBits1 = VT1.getSizeInBits();
6055 unsigned NumBits2 = VT2.getSizeInBits();
6056 if (NumBits1 <= NumBits2)
6061 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
6063 bool ARM64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
6064 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6066 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6067 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6068 if (NumBits1 == 32 && NumBits2 == 64)
6072 bool ARM64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
6073 if (!VT1.isInteger() || !VT2.isInteger())
6075 unsigned NumBits1 = VT1.getSizeInBits();
6076 unsigned NumBits2 = VT2.getSizeInBits();
6077 if (NumBits1 == 32 && NumBits2 == 64)
6082 bool ARM64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
6083 EVT VT1 = Val.getValueType();
6084 if (isZExtFree(VT1, VT2)) {
6088 if (Val.getOpcode() != ISD::LOAD)
6091 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
6092 return (VT1.isSimple() && VT1.isInteger() && VT2.isSimple() &&
6093 VT2.isInteger() && VT1.getSizeInBits() <= 32);
6096 bool ARM64TargetLowering::hasPairedLoad(Type *LoadedType,
6097 unsigned &RequiredAligment) const {
6098 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
6100 // Cyclone supports unaligned accesses.
6101 RequiredAligment = 0;
6102 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
6103 return NumBits == 32 || NumBits == 64;
6106 bool ARM64TargetLowering::hasPairedLoad(EVT LoadedType,
6107 unsigned &RequiredAligment) const {
6108 if (!LoadedType.isSimple() ||
6109 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
6111 // Cyclone supports unaligned accesses.
6112 RequiredAligment = 0;
6113 unsigned NumBits = LoadedType.getSizeInBits();
6114 return NumBits == 32 || NumBits == 64;
6117 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
6118 unsigned AlignCheck) {
6119 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
6120 (DstAlign == 0 || DstAlign % AlignCheck == 0));
6123 EVT ARM64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
6124 unsigned SrcAlign, bool IsMemset,
6125 bool ZeroMemset, bool MemcpyStrSrc,
6126 MachineFunction &MF) const {
6127 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
6128 // instruction to materialize the v2i64 zero and one store (with restrictive
6129 // addressing mode). Just do two i64 store of zero-registers.
6131 const Function *F = MF.getFunction();
6132 if (!IsMemset && Size >= 16 &&
6133 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
6134 Attribute::NoImplicitFloat) &&
6135 (memOpAlign(SrcAlign, DstAlign, 16) ||
6136 (allowsUnalignedMemoryAccesses(MVT::v2i64, 0, &Fast) && Fast)))
6139 return Size >= 8 ? MVT::i64 : MVT::i32;
6142 // 12-bit optionally shifted immediates are legal for adds.
6143 bool ARM64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
6144 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
6149 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
6150 // immediates is the same as for an add or a sub.
6151 bool ARM64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
6154 return isLegalAddImmediate(Immed);
6157 /// isLegalAddressingMode - Return true if the addressing mode represented
6158 /// by AM is legal for this target, for a load/store of the specified type.
6159 bool ARM64TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6161 // ARM64 has five basic addressing modes:
6163 // reg + 9-bit signed offset
6164 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
6166 // reg + SIZE_IN_BYTES * reg
6168 // No global is ever allowed as a base.
6172 // No reg+reg+imm addressing.
6173 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
6176 // check reg + imm case:
6177 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
6178 uint64_t NumBytes = 0;
6179 if (Ty->isSized()) {
6180 uint64_t NumBits = getDataLayout()->getTypeSizeInBits(Ty);
6181 NumBytes = NumBits / 8;
6182 if (!isPowerOf2_64(NumBits))
6187 int64_t Offset = AM.BaseOffs;
6189 // 9-bit signed offset
6190 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
6193 // 12-bit unsigned offset
6194 unsigned shift = Log2_64(NumBytes);
6195 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
6196 // Must be a multiple of NumBytes (NumBytes is a power of 2)
6197 (Offset >> shift) << shift == Offset)
6202 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
6204 if (!AM.Scale || AM.Scale == 1 ||
6205 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
6210 int ARM64TargetLowering::getScalingFactorCost(const AddrMode &AM,
6212 // Scaling factors are not free at all.
6213 // Operands | Rt Latency
6214 // -------------------------------------------
6216 // -------------------------------------------
6217 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
6218 // Rt, [Xn, Wm, <extend> #imm] |
6219 if (isLegalAddressingMode(AM, Ty))
6220 // Scale represents reg2 * scale, thus account for 1 if
6221 // it is not equal to 0 or 1.
6222 return AM.Scale != 0 && AM.Scale != 1;
6226 bool ARM64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
6227 VT = VT.getScalarType();
6232 switch (VT.getSimpleVT().SimpleTy) {
6244 ARM64TargetLowering::getScratchRegisters(CallingConv::ID) const {
6245 // LR is a callee-save register, but we must treat it as clobbered by any call
6246 // site. Hence we include LR in the scratch registers, which are in turn added
6247 // as implicit-defs for stackmaps and patchpoints.
6248 static const uint16_t ScratchRegs[] = {
6249 ARM64::X16, ARM64::X17, ARM64::LR, 0
6254 bool ARM64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
6256 assert(Ty->isIntegerTy());
6258 unsigned BitSize = Ty->getPrimitiveSizeInBits();
6262 int64_t Val = Imm.getSExtValue();
6263 if (Val == 0 || ARM64_AM::isLogicalImmediate(Val, BitSize))
6266 if ((int64_t)Val < 0)
6269 Val &= (1LL << 32) - 1;
6271 unsigned LZ = countLeadingZeros((uint64_t)Val);
6272 unsigned Shift = (63 - LZ) / 16;
6273 // MOVZ is free so return true for one or fewer MOVK.
6274 return (Shift < 3) ? true : false;
6277 // Generate SUBS and CSEL for integer abs.
6278 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
6279 EVT VT = N->getValueType(0);
6281 SDValue N0 = N->getOperand(0);
6282 SDValue N1 = N->getOperand(1);
6285 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
6286 // and change it to SUB and CSEL.
6287 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
6288 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
6289 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
6290 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
6291 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
6292 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
6294 // Generate SUBS & CSEL.
6296 DAG.getNode(ARM64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
6297 N0.getOperand(0), DAG.getConstant(0, VT));
6298 return DAG.getNode(ARM64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
6299 DAG.getConstant(ARM64CC::PL, MVT::i32),
6300 SDValue(Cmp.getNode(), 1));
6305 // performXorCombine - Attempts to handle integer ABS.
6306 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
6307 TargetLowering::DAGCombinerInfo &DCI,
6308 const ARM64Subtarget *Subtarget) {
6309 if (DCI.isBeforeLegalizeOps())
6312 return performIntegerAbsCombine(N, DAG);
6315 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
6316 TargetLowering::DAGCombinerInfo &DCI,
6317 const ARM64Subtarget *Subtarget) {
6318 if (DCI.isBeforeLegalizeOps())
6321 // Multiplication of a power of two plus/minus one can be done more
6322 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
6323 // future CPUs have a cheaper MADD instruction, this may need to be
6324 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
6325 // 64-bit is 5 cycles, so this is always a win.
6326 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
6327 APInt Value = C->getAPIntValue();
6328 EVT VT = N->getValueType(0);
6329 APInt VP1 = Value + 1;
6330 if (VP1.isPowerOf2()) {
6331 // Multiplying by one less than a power of two, replace with a shift
6333 SDValue ShiftedVal = DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6334 DAG.getConstant(VP1.logBase2(), VT));
6335 return DAG.getNode(ISD::SUB, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6337 APInt VM1 = Value - 1;
6338 if (VM1.isPowerOf2()) {
6339 // Multiplying by one more than a power of two, replace with a shift
6341 SDValue ShiftedVal = DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6342 DAG.getConstant(VM1.logBase2(), VT));
6343 return DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6349 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG) {
6350 EVT VT = N->getValueType(0);
6351 if (VT != MVT::f32 && VT != MVT::f64)
6353 // Only optimize when the source and destination types have the same width.
6354 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
6357 // If the result of an integer load is only used by an integer-to-float
6358 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
6359 // This eliminates an "integer-to-vector-move UOP and improve throughput.
6360 SDValue N0 = N->getOperand(0);
6361 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6362 // Do not change the width of a volatile load.
6363 !cast<LoadSDNode>(N0)->isVolatile()) {
6364 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6365 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
6366 LN0->getPointerInfo(), LN0->isVolatile(),
6367 LN0->isNonTemporal(), LN0->isInvariant(),
6368 LN0->getAlignment());
6370 // Make sure successors of the original load stay after it by updating them
6371 // to use the new Chain.
6372 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
6375 (N->getOpcode() == ISD::SINT_TO_FP) ? ARM64ISD::SITOF : ARM64ISD::UITOF;
6376 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
6382 /// An EXTR instruction is made up of two shifts, ORed together. This helper
6383 /// searches for and classifies those shifts.
6384 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
6386 if (N.getOpcode() == ISD::SHL)
6388 else if (N.getOpcode() == ISD::SRL)
6393 if (!isa<ConstantSDNode>(N.getOperand(1)))
6396 ShiftAmount = N->getConstantOperandVal(1);
6397 Src = N->getOperand(0);
6401 /// EXTR instruction extracts a contiguous chunk of bits from two existing
6402 /// registers viewed as a high/low pair. This function looks for the pattern:
6403 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
6404 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
6406 static SDValue tryCombineToEXTR(SDNode *N,
6407 TargetLowering::DAGCombinerInfo &DCI) {
6408 SelectionDAG &DAG = DCI.DAG;
6410 EVT VT = N->getValueType(0);
6412 assert(N->getOpcode() == ISD::OR && "Unexpected root");
6414 if (VT != MVT::i32 && VT != MVT::i64)
6418 uint32_t ShiftLHS = 0;
6420 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
6424 uint32_t ShiftRHS = 0;
6426 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
6429 // If they're both trying to come from the high part of the register, they're
6430 // not really an EXTR.
6431 if (LHSFromHi == RHSFromHi)
6434 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
6438 std::swap(LHS, RHS);
6439 std::swap(ShiftLHS, ShiftRHS);
6442 return DAG.getNode(ARM64ISD::EXTR, DL, VT, LHS, RHS,
6443 DAG.getConstant(ShiftRHS, MVT::i64));
6446 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
6447 const ARM64Subtarget *Subtarget) {
6448 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
6449 if (!EnableARM64ExtrGeneration)
6451 SelectionDAG &DAG = DCI.DAG;
6452 EVT VT = N->getValueType(0);
6454 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6457 SDValue Res = tryCombineToEXTR(N, DCI);
6464 static SDValue performBitcastCombine(SDNode *N,
6465 TargetLowering::DAGCombinerInfo &DCI,
6466 SelectionDAG &DAG) {
6467 // Wait 'til after everything is legalized to try this. That way we have
6468 // legal vector types and such.
6469 if (DCI.isBeforeLegalizeOps())
6472 // Remove extraneous bitcasts around an extract_subvector.
6474 // (v4i16 (bitconvert
6475 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
6477 // (extract_subvector ((v8i16 ...), (i64 4)))
6479 // Only interested in 64-bit vectors as the ultimate result.
6480 EVT VT = N->getValueType(0);
6483 if (VT.getSimpleVT().getSizeInBits() != 64)
6485 // Is the operand an extract_subvector starting at the beginning or halfway
6486 // point of the vector? A low half may also come through as an
6487 // EXTRACT_SUBREG, so look for that, too.
6488 SDValue Op0 = N->getOperand(0);
6489 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
6490 !(Op0->isMachineOpcode() &&
6491 Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG))
6493 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
6494 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
6495 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
6497 } else if (Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG) {
6498 if (idx != ARM64::dsub)
6500 // The dsub reference is equivalent to a lane zero subvector reference.
6503 // Look through the bitcast of the input to the extract.
6504 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
6506 SDValue Source = Op0->getOperand(0)->getOperand(0);
6507 // If the source type has twice the number of elements as our destination
6508 // type, we know this is an extract of the high or low half of the vector.
6509 EVT SVT = Source->getValueType(0);
6510 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
6513 DEBUG(dbgs() << "arm64-lower: bitcast extract_subvector simplification\n");
6515 // Create the simplified form to just extract the low or high half of the
6516 // vector directly rather than bothering with the bitcasts.
6518 unsigned NumElements = VT.getVectorNumElements();
6520 SDValue HalfIdx = DAG.getConstant(NumElements, MVT::i64);
6521 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
6523 SDValue SubReg = DAG.getTargetConstant(ARM64::dsub, MVT::i32);
6524 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
6530 static SDValue performConcatVectorsCombine(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())
6539 EVT VT = N->getValueType(0);
6541 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
6542 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
6543 // canonicalise to that.
6544 if (N->getOperand(0) == N->getOperand(1) && VT.getVectorNumElements() == 2) {
6545 assert(VT.getVectorElementType().getSizeInBits() == 64);
6546 return DAG.getNode(ARM64ISD::DUPLANE64, dl, VT,
6547 WidenVector(N->getOperand(0), DAG),
6548 DAG.getConstant(0, MVT::i64));
6551 // Canonicalise concat_vectors so that the right-hand vector has as few
6552 // bit-casts as possible before its real operation. The primary matching
6553 // destination for these operations will be the narrowing "2" instructions,
6554 // which depend on the operation being performed on this right-hand vector.
6556 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
6558 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
6560 SDValue Op1 = N->getOperand(1);
6561 if (Op1->getOpcode() != ISD::BITCAST)
6563 SDValue RHS = Op1->getOperand(0);
6564 MVT RHSTy = RHS.getValueType().getSimpleVT();
6565 // If the RHS is not a vector, this is not the pattern we're looking for.
6566 if (!RHSTy.isVector())
6569 DEBUG(dbgs() << "arm64-lower: concat_vectors bitcast simplification\n");
6571 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
6572 RHSTy.getVectorNumElements() * 2);
6574 ISD::BITCAST, dl, VT,
6575 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
6576 DAG.getNode(ISD::BITCAST, dl, RHSTy, N->getOperand(0)), RHS));
6579 static SDValue tryCombineFixedPointConvert(SDNode *N,
6580 TargetLowering::DAGCombinerInfo &DCI,
6581 SelectionDAG &DAG) {
6582 // Wait 'til after everything is legalized to try this. That way we have
6583 // legal vector types and such.
6584 if (DCI.isBeforeLegalizeOps())
6586 // Transform a scalar conversion of a value from a lane extract into a
6587 // lane extract of a vector conversion. E.g., from foo1 to foo2:
6588 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
6589 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
6591 // The second form interacts better with instruction selection and the
6592 // register allocator to avoid cross-class register copies that aren't
6593 // coalescable due to a lane reference.
6595 // Check the operand and see if it originates from a lane extract.
6596 SDValue Op1 = N->getOperand(1);
6597 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6598 // Yep, no additional predication needed. Perform the transform.
6599 SDValue IID = N->getOperand(0);
6600 SDValue Shift = N->getOperand(2);
6601 SDValue Vec = Op1.getOperand(0);
6602 SDValue Lane = Op1.getOperand(1);
6603 EVT ResTy = N->getValueType(0);
6607 // The vector width should be 128 bits by the time we get here, even
6608 // if it started as 64 bits (the extract_vector handling will have
6610 assert(Vec.getValueType().getSizeInBits() == 128 &&
6611 "unexpected vector size on extract_vector_elt!");
6612 if (Vec.getValueType() == MVT::v4i32)
6613 VecResTy = MVT::v4f32;
6614 else if (Vec.getValueType() == MVT::v2i64)
6615 VecResTy = MVT::v2f64;
6617 assert(0 && "unexpected vector type!");
6620 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
6621 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
6626 // AArch64 high-vector "long" operations are formed by performing the non-high
6627 // version on an extract_subvector of each operand which gets the high half:
6629 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
6631 // However, there are cases which don't have an extract_high explicitly, but
6632 // have another operation that can be made compatible with one for free. For
6635 // (dupv64 scalar) --> (extract_high (dup128 scalar))
6637 // This routine does the actual conversion of such DUPs, once outer routines
6638 // have determined that everything else is in order.
6639 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
6640 // We can handle most types of duplicate, but the lane ones have an extra
6641 // operand saying *which* lane, so we need to know.
6643 switch (N.getOpcode()) {
6647 case ARM64ISD::DUPLANE8:
6648 case ARM64ISD::DUPLANE16:
6649 case ARM64ISD::DUPLANE32:
6650 case ARM64ISD::DUPLANE64:
6657 MVT NarrowTy = N.getSimpleValueType();
6658 if (!NarrowTy.is64BitVector())
6661 MVT ElementTy = NarrowTy.getVectorElementType();
6662 unsigned NumElems = NarrowTy.getVectorNumElements();
6663 MVT NewDUPVT = MVT::getVectorVT(ElementTy, NumElems * 2);
6667 NewDUP = DAG.getNode(N.getOpcode(), SDLoc(N), NewDUPVT, N.getOperand(0),
6670 NewDUP = DAG.getNode(ARM64ISD::DUP, SDLoc(N), NewDUPVT, N.getOperand(0));
6672 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N.getNode()), NarrowTy,
6673 NewDUP, DAG.getConstant(NumElems, MVT::i64));
6676 static bool isEssentiallyExtractSubvector(SDValue N) {
6677 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
6680 return N.getOpcode() == ISD::BITCAST &&
6681 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
6684 /// \brief Helper structure to keep track of ISD::SET_CC operands.
6685 struct GenericSetCCInfo {
6686 const SDValue *Opnd0;
6687 const SDValue *Opnd1;
6691 /// \brief Helper structure to keep track of a SET_CC lowered into ARM64 code.
6692 struct ARM64SetCCInfo {
6694 ARM64CC::CondCode CC;
6697 /// \brief Helper structure to keep track of SetCC information.
6699 GenericSetCCInfo Generic;
6700 ARM64SetCCInfo ARM64;
6703 /// \brief Helper structure to be able to read SetCC information.
6704 /// If set to true, IsARM64 field, Info is a ARM64SetCCInfo, otherwise Info is
6705 /// a GenericSetCCInfo.
6706 struct SetCCInfoAndKind {
6711 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
6713 /// ARM64 lowered one.
6714 /// \p SetCCInfo is filled accordingly.
6715 /// \post SetCCInfo is meanginfull only when this function returns true.
6716 /// \return True when Op is a kind of SET_CC operation.
6717 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
6718 // If this is a setcc, this is straight forward.
6719 if (Op.getOpcode() == ISD::SETCC) {
6720 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
6721 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
6722 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6723 SetCCInfo.IsARM64 = false;
6726 // Otherwise, check if this is a matching csel instruction.
6730 if (Op.getOpcode() != ARM64ISD::CSEL)
6732 // Set the information about the operands.
6733 // TODO: we want the operands of the Cmp not the csel
6734 SetCCInfo.Info.ARM64.Cmp = &Op.getOperand(3);
6735 SetCCInfo.IsARM64 = true;
6736 SetCCInfo.Info.ARM64.CC = static_cast<ARM64CC::CondCode>(
6737 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
6739 // Check that the operands matches the constraints:
6740 // (1) Both operands must be constants.
6741 // (2) One must be 1 and the other must be 0.
6742 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
6743 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6746 if (!TValue || !FValue)
6750 if (!TValue->isOne()) {
6751 // Update the comparison when we are interested in !cc.
6752 std::swap(TValue, FValue);
6753 SetCCInfo.Info.ARM64.CC =
6754 ARM64CC::getInvertedCondCode(SetCCInfo.Info.ARM64.CC);
6756 return TValue->isOne() && FValue->isNullValue();
6759 // The folding we want to perform is:
6760 // (add x, (setcc cc ...) )
6762 // (csel x, (add x, 1), !cc ...)
6764 // The latter will get matched to a CSINC instruction.
6765 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
6766 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
6767 SDValue LHS = Op->getOperand(0);
6768 SDValue RHS = Op->getOperand(1);
6769 SetCCInfoAndKind InfoAndKind;
6771 // If neither operand is a SET_CC, give up.
6772 if (!isSetCC(LHS, InfoAndKind)) {
6773 std::swap(LHS, RHS);
6774 if (!isSetCC(LHS, InfoAndKind))
6778 // FIXME: This could be generatized to work for FP comparisons.
6779 EVT CmpVT = InfoAndKind.IsARM64
6780 ? InfoAndKind.Info.ARM64.Cmp->getOperand(0).getValueType()
6781 : InfoAndKind.Info.Generic.Opnd0->getValueType();
6782 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
6788 if (InfoAndKind.IsARM64) {
6789 CCVal = DAG.getConstant(
6790 ARM64CC::getInvertedCondCode(InfoAndKind.Info.ARM64.CC), MVT::i32);
6791 Cmp = *InfoAndKind.Info.ARM64.Cmp;
6793 Cmp = getARM64Cmp(*InfoAndKind.Info.Generic.Opnd0,
6794 *InfoAndKind.Info.Generic.Opnd1,
6795 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
6798 EVT VT = Op->getValueType(0);
6799 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, VT));
6800 return DAG.getNode(ARM64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
6803 // The basic add/sub long vector instructions have variants with "2" on the end
6804 // which act on the high-half of their inputs. They are normally matched by
6807 // (add (zeroext (extract_high LHS)),
6808 // (zeroext (extract_high RHS)))
6809 // -> uaddl2 vD, vN, vM
6811 // However, if one of the extracts is something like a duplicate, this
6812 // instruction can still be used profitably. This function puts the DAG into a
6813 // more appropriate form for those patterns to trigger.
6814 static SDValue performAddSubLongCombine(SDNode *N,
6815 TargetLowering::DAGCombinerInfo &DCI,
6816 SelectionDAG &DAG) {
6817 if (DCI.isBeforeLegalizeOps())
6820 MVT VT = N->getSimpleValueType(0);
6821 if (!VT.is128BitVector()) {
6822 if (N->getOpcode() == ISD::ADD)
6823 return performSetccAddFolding(N, DAG);
6827 // Make sure both branches are extended in the same way.
6828 SDValue LHS = N->getOperand(0);
6829 SDValue RHS = N->getOperand(1);
6830 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
6831 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
6832 LHS.getOpcode() != RHS.getOpcode())
6835 unsigned ExtType = LHS.getOpcode();
6837 // It's not worth doing if at least one of the inputs isn't already an
6838 // extract, but we don't know which it'll be so we have to try both.
6839 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
6840 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
6844 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
6845 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
6846 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
6850 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
6853 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
6856 // Massage DAGs which we can use the high-half "long" operations on into
6857 // something isel will recognize better. E.g.
6859 // (arm64_neon_umull (extract_high vec) (dupv64 scalar)) -->
6860 // (arm64_neon_umull (extract_high (v2i64 vec)))
6861 // (extract_high (v2i64 (dup128 scalar)))))
6863 static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N,
6864 TargetLowering::DAGCombinerInfo &DCI,
6865 SelectionDAG &DAG) {
6866 if (DCI.isBeforeLegalizeOps())
6869 SDValue LHS = N->getOperand(1);
6870 SDValue RHS = N->getOperand(2);
6871 assert(LHS.getValueType().is64BitVector() &&
6872 RHS.getValueType().is64BitVector() &&
6873 "unexpected shape for long operation");
6875 // Either node could be a DUP, but it's not worth doing both of them (you'd
6876 // just as well use the non-high version) so look for a corresponding extract
6877 // operation on the other "wing".
6878 if (isEssentiallyExtractSubvector(LHS)) {
6879 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
6882 } else if (isEssentiallyExtractSubvector(RHS)) {
6883 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
6888 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
6889 N->getOperand(0), LHS, RHS);
6892 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
6893 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
6894 unsigned ElemBits = ElemTy.getSizeInBits();
6896 int64_t ShiftAmount;
6897 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
6898 APInt SplatValue, SplatUndef;
6899 unsigned SplatBitSize;
6901 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
6902 HasAnyUndefs, ElemBits) ||
6903 SplatBitSize != ElemBits)
6906 ShiftAmount = SplatValue.getSExtValue();
6907 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
6908 ShiftAmount = CVN->getSExtValue();
6916 llvm_unreachable("Unknown shift intrinsic");
6917 case Intrinsic::arm64_neon_sqshl:
6918 Opcode = ARM64ISD::SQSHL_I;
6919 IsRightShift = false;
6921 case Intrinsic::arm64_neon_uqshl:
6922 Opcode = ARM64ISD::UQSHL_I;
6923 IsRightShift = false;
6925 case Intrinsic::arm64_neon_srshl:
6926 Opcode = ARM64ISD::SRSHR_I;
6927 IsRightShift = true;
6929 case Intrinsic::arm64_neon_urshl:
6930 Opcode = ARM64ISD::URSHR_I;
6931 IsRightShift = true;
6933 case Intrinsic::arm64_neon_sqshlu:
6934 Opcode = ARM64ISD::SQSHLU_I;
6935 IsRightShift = false;
6939 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits)
6940 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6941 DAG.getConstant(-ShiftAmount, MVT::i32));
6942 else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount <= ElemBits)
6943 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6944 DAG.getConstant(ShiftAmount, MVT::i32));
6949 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
6950 // the intrinsics must be legal and take an i32, this means there's almost
6951 // certainly going to be a zext in the DAG which we can eliminate.
6952 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
6953 SDValue AndN = N->getOperand(2);
6954 if (AndN.getOpcode() != ISD::AND)
6957 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
6958 if (!CMask || CMask->getZExtValue() != Mask)
6961 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
6962 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
6965 static SDValue performIntrinsicCombine(SDNode *N,
6966 TargetLowering::DAGCombinerInfo &DCI,
6967 const ARM64Subtarget *Subtarget) {
6968 SelectionDAG &DAG = DCI.DAG;
6969 unsigned IID = getIntrinsicID(N);
6973 case Intrinsic::arm64_neon_vcvtfxs2fp:
6974 case Intrinsic::arm64_neon_vcvtfxu2fp:
6975 return tryCombineFixedPointConvert(N, DCI, DAG);
6977 case Intrinsic::arm64_neon_fmax:
6978 return DAG.getNode(ARM64ISD::FMAX, SDLoc(N), N->getValueType(0),
6979 N->getOperand(1), N->getOperand(2));
6980 case Intrinsic::arm64_neon_fmin:
6981 return DAG.getNode(ARM64ISD::FMIN, SDLoc(N), N->getValueType(0),
6982 N->getOperand(1), N->getOperand(2));
6983 case Intrinsic::arm64_neon_smull:
6984 case Intrinsic::arm64_neon_umull:
6985 case Intrinsic::arm64_neon_pmull:
6986 case Intrinsic::arm64_neon_sqdmull:
6987 return tryCombineLongOpWithDup(IID, N, DCI, DAG);
6988 case Intrinsic::arm64_neon_sqshl:
6989 case Intrinsic::arm64_neon_uqshl:
6990 case Intrinsic::arm64_neon_sqshlu:
6991 case Intrinsic::arm64_neon_srshl:
6992 case Intrinsic::arm64_neon_urshl:
6993 return tryCombineShiftImm(IID, N, DAG);
6994 case Intrinsic::arm64_crc32b:
6995 case Intrinsic::arm64_crc32cb:
6996 return tryCombineCRC32(0xff, N, DAG);
6997 case Intrinsic::arm64_crc32h:
6998 case Intrinsic::arm64_crc32ch:
6999 return tryCombineCRC32(0xffff, N, DAG);
7004 static SDValue performExtendCombine(SDNode *N,
7005 TargetLowering::DAGCombinerInfo &DCI,
7006 SelectionDAG &DAG) {
7007 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
7008 // we can convert that DUP into another extract_high (of a bigger DUP), which
7009 // helps the backend to decide that an sabdl2 would be useful, saving a real
7010 // extract_high operation.
7011 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
7012 N->getOperand(0).getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
7013 SDNode *ABDNode = N->getOperand(0).getNode();
7014 unsigned IID = getIntrinsicID(ABDNode);
7015 if (IID == Intrinsic::arm64_neon_sabd ||
7016 IID == Intrinsic::arm64_neon_uabd) {
7017 SDValue NewABD = tryCombineLongOpWithDup(IID, ABDNode, DCI, DAG);
7018 if (!NewABD.getNode())
7021 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
7026 // This is effectively a custom type legalization for ARM64.
7028 // Type legalization will split an extend of a small, legal, type to a larger
7029 // illegal type by first splitting the destination type, often creating
7030 // illegal source types, which then get legalized in isel-confusing ways,
7031 // leading to really terrible codegen. E.g.,
7032 // %result = v8i32 sext v8i8 %value
7034 // %losrc = extract_subreg %value, ...
7035 // %hisrc = extract_subreg %value, ...
7036 // %lo = v4i32 sext v4i8 %losrc
7037 // %hi = v4i32 sext v4i8 %hisrc
7038 // Things go rapidly downhill from there.
7040 // For ARM64, the [sz]ext vector instructions can only go up one element
7041 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
7042 // take two instructions.
7044 // This implies that the most efficient way to do the extend from v8i8
7045 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
7046 // the normal splitting to happen for the v8i16->v8i32.
7048 // This is pre-legalization to catch some cases where the default
7049 // type legalization will create ill-tempered code.
7050 if (!DCI.isBeforeLegalizeOps())
7053 // We're only interested in cleaning things up for non-legal vector types
7054 // here. If both the source and destination are legal, things will just
7055 // work naturally without any fiddling.
7056 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7057 EVT ResVT = N->getValueType(0);
7058 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
7060 // If the vector type isn't a simple VT, it's beyond the scope of what
7061 // we're worried about here. Let legalization do its thing and hope for
7063 if (!ResVT.isSimple())
7066 SDValue Src = N->getOperand(0);
7067 MVT SrcVT = Src->getValueType(0).getSimpleVT();
7068 // If the source VT is a 64-bit vector, we can play games and get the
7069 // better results we want.
7070 if (SrcVT.getSizeInBits() != 64)
7073 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
7074 unsigned ElementCount = SrcVT.getVectorNumElements();
7075 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
7077 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
7079 // Now split the rest of the operation into two halves, each with a 64
7083 unsigned NumElements = ResVT.getVectorNumElements();
7084 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
7085 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
7086 ResVT.getVectorElementType(), NumElements / 2);
7088 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
7089 LoVT.getVectorNumElements());
7090 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7091 DAG.getIntPtrConstant(0));
7092 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7093 DAG.getIntPtrConstant(InNVT.getVectorNumElements()));
7094 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
7095 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
7097 // Now combine the parts back together so we still have a single result
7098 // like the combiner expects.
7099 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
7102 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
7103 /// value. The load store optimizer pass will merge them to store pair stores.
7104 /// This has better performance than a splat of the scalar followed by a split
7105 /// vector store. Even if the stores are not merged it is four stores vs a dup,
7106 /// followed by an ext.b and two stores.
7107 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
7108 SDValue StVal = St->getValue();
7109 EVT VT = StVal.getValueType();
7111 // Don't replace floating point stores, they possibly won't be transformed to
7112 // stp because of the store pair suppress pass.
7113 if (VT.isFloatingPoint())
7116 // Check for insert vector elements.
7117 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
7120 // We can express a splat as store pair(s) for 2 or 4 elements.
7121 unsigned NumVecElts = VT.getVectorNumElements();
7122 if (NumVecElts != 4 && NumVecElts != 2)
7124 SDValue SplatVal = StVal.getOperand(1);
7125 unsigned RemainInsertElts = NumVecElts - 1;
7127 // Check that this is a splat.
7128 while (--RemainInsertElts) {
7129 SDValue NextInsertElt = StVal.getOperand(0);
7130 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
7132 if (NextInsertElt.getOperand(1) != SplatVal)
7134 StVal = NextInsertElt;
7136 unsigned OrigAlignment = St->getAlignment();
7137 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
7138 unsigned Alignment = std::min(OrigAlignment, EltOffset);
7140 // Create scalar stores. This is at least as good as the code sequence for a
7141 // split unaligned store wich is a dup.s, ext.b, and two stores.
7142 // Most of the time the three stores should be replaced by store pair
7143 // instructions (stp).
7145 SDValue BasePtr = St->getBasePtr();
7147 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
7148 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
7150 unsigned Offset = EltOffset;
7151 while (--NumVecElts) {
7152 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7153 DAG.getConstant(Offset, MVT::i64));
7154 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
7155 St->getPointerInfo(), St->isVolatile(),
7156 St->isNonTemporal(), Alignment);
7157 Offset += EltOffset;
7162 static SDValue performSTORECombine(SDNode *N,
7163 TargetLowering::DAGCombinerInfo &DCI,
7165 const ARM64Subtarget *Subtarget) {
7166 if (!DCI.isBeforeLegalize())
7169 StoreSDNode *S = cast<StoreSDNode>(N);
7170 if (S->isVolatile())
7173 // Cyclone has bad performance on unaligned 16B stores when crossing line and
7174 // page boundries. We want to split such stores.
7175 if (!Subtarget->isCyclone())
7178 // Don't split at Oz.
7179 MachineFunction &MF = DAG.getMachineFunction();
7180 bool IsMinSize = MF.getFunction()->getAttributes().hasAttribute(
7181 AttributeSet::FunctionIndex, Attribute::MinSize);
7185 SDValue StVal = S->getValue();
7186 EVT VT = StVal.getValueType();
7188 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
7189 // those up regresses performance on micro-benchmarks and olden/bh.
7190 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
7193 // Split unaligned 16B stores. They are terrible for performance.
7194 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
7195 // extensions can use this to mark that it does not want splitting to happen
7196 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
7197 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
7198 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
7199 S->getAlignment() <= 2)
7202 // If we get a splat of a scalar convert this vector store to a store of
7203 // scalars. They will be merged into store pairs thereby removing two
7205 SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S);
7206 if (ReplacedSplat != SDValue())
7207 return ReplacedSplat;
7210 unsigned NumElts = VT.getVectorNumElements() / 2;
7211 // Split VT into two.
7213 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
7214 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7215 DAG.getIntPtrConstant(0));
7216 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7217 DAG.getIntPtrConstant(NumElts));
7218 SDValue BasePtr = S->getBasePtr();
7220 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
7221 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
7222 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7223 DAG.getConstant(8, MVT::i64));
7224 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
7225 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
7229 // Optimize compare with zero and branch.
7230 static SDValue performBRCONDCombine(SDNode *N,
7231 TargetLowering::DAGCombinerInfo &DCI,
7232 SelectionDAG &DAG) {
7233 SDValue Chain = N->getOperand(0);
7234 SDValue Dest = N->getOperand(1);
7235 SDValue CCVal = N->getOperand(2);
7236 SDValue Cmp = N->getOperand(3);
7238 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
7239 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
7240 if (CC != ARM64CC::EQ && CC != ARM64CC::NE)
7243 unsigned CmpOpc = Cmp.getOpcode();
7244 if (CmpOpc != ARM64ISD::ADDS && CmpOpc != ARM64ISD::SUBS)
7247 // Only attempt folding if there is only one use of the flag and no use of the
7249 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
7252 SDValue LHS = Cmp.getOperand(0);
7253 SDValue RHS = Cmp.getOperand(1);
7255 assert(LHS.getValueType() == RHS.getValueType() &&
7256 "Expected the value type to be the same for both operands!");
7257 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
7260 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
7261 std::swap(LHS, RHS);
7263 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
7266 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
7267 LHS.getOpcode() == ISD::SRL)
7270 // Fold the compare into the branch instruction.
7272 if (CC == ARM64CC::EQ)
7273 BR = DAG.getNode(ARM64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7275 BR = DAG.getNode(ARM64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7277 // Do not add new nodes to DAG combiner worklist.
7278 DCI.CombineTo(N, BR, false);
7283 SDValue ARM64TargetLowering::PerformDAGCombine(SDNode *N,
7284 DAGCombinerInfo &DCI) const {
7285 SelectionDAG &DAG = DCI.DAG;
7286 switch (N->getOpcode()) {
7291 return performAddSubLongCombine(N, DCI, DAG);
7293 return performXorCombine(N, DAG, DCI, Subtarget);
7295 return performMulCombine(N, DAG, DCI, Subtarget);
7296 case ISD::SINT_TO_FP:
7297 case ISD::UINT_TO_FP:
7298 return performIntToFpCombine(N, DAG);
7300 return performORCombine(N, DCI, Subtarget);
7301 case ISD::INTRINSIC_WO_CHAIN:
7302 return performIntrinsicCombine(N, DCI, Subtarget);
7303 case ISD::ANY_EXTEND:
7304 case ISD::ZERO_EXTEND:
7305 case ISD::SIGN_EXTEND:
7306 return performExtendCombine(N, DCI, DAG);
7308 return performBitcastCombine(N, DCI, DAG);
7309 case ISD::CONCAT_VECTORS:
7310 return performConcatVectorsCombine(N, DCI, DAG);
7312 return performSTORECombine(N, DCI, DAG, Subtarget);
7313 case ARM64ISD::BRCOND:
7314 return performBRCONDCombine(N, DCI, DAG);
7319 // Check if the return value is used as only a return value, as otherwise
7320 // we can't perform a tail-call. In particular, we need to check for
7321 // target ISD nodes that are returns and any other "odd" constructs
7322 // that the generic analysis code won't necessarily catch.
7323 bool ARM64TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
7324 if (N->getNumValues() != 1)
7326 if (!N->hasNUsesOfValue(1, 0))
7329 SDValue TCChain = Chain;
7330 SDNode *Copy = *N->use_begin();
7331 if (Copy->getOpcode() == ISD::CopyToReg) {
7332 // If the copy has a glue operand, we conservatively assume it isn't safe to
7333 // perform a tail call.
7334 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
7337 TCChain = Copy->getOperand(0);
7338 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
7341 bool HasRet = false;
7342 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
7344 if (UI->getOpcode() != ARM64ISD::RET_FLAG)
7356 // Return whether the an instruction can potentially be optimized to a tail
7357 // call. This will cause the optimizers to attempt to move, or duplicate,
7358 // return instructions to help enable tail call optimizations for this
7360 bool ARM64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
7361 if (!EnableARM64TailCalls)
7364 if (!CI->isTailCall())
7370 bool ARM64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
7372 ISD::MemIndexedMode &AM,
7374 SelectionDAG &DAG) const {
7375 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
7378 Base = Op->getOperand(0);
7379 // All of the indexed addressing mode instructions take a signed
7380 // 9 bit immediate offset.
7381 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
7382 int64_t RHSC = (int64_t)RHS->getZExtValue();
7383 if (RHSC >= 256 || RHSC <= -256)
7385 IsInc = (Op->getOpcode() == ISD::ADD);
7386 Offset = Op->getOperand(1);
7392 bool ARM64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
7394 ISD::MemIndexedMode &AM,
7395 SelectionDAG &DAG) const {
7398 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7399 VT = LD->getMemoryVT();
7400 Ptr = LD->getBasePtr();
7401 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7402 VT = ST->getMemoryVT();
7403 Ptr = ST->getBasePtr();
7408 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
7410 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
7414 bool ARM64TargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
7417 ISD::MemIndexedMode &AM,
7418 SelectionDAG &DAG) const {
7421 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7422 VT = LD->getMemoryVT();
7423 Ptr = LD->getBasePtr();
7424 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7425 VT = ST->getMemoryVT();
7426 Ptr = ST->getBasePtr();
7431 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
7433 // Post-indexing updates the base, so it's not a valid transform
7434 // if that's not the same as the load's pointer.
7437 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
7441 /// The only 128-bit atomic operation is an stxp that succeeds. In particular
7442 /// neither ldp nor ldxp are atomic. So the canonical sequence for an atomic
7445 /// ldxp x0, x1, [x8]
7446 /// stxp w2, x0, x1, [x8]
7448 /// If the stxp succeeds then the ldxp managed to get both halves without an
7449 /// intervening stxp from a different thread and the read was atomic.
7450 static void ReplaceATOMIC_LOAD_128(SDNode *N, SmallVectorImpl<SDValue> &Results,
7451 SelectionDAG &DAG) {
7453 AtomicSDNode *AN = cast<AtomicSDNode>(N);
7454 EVT VT = AN->getMemoryVT();
7455 SDValue Zero = DAG.getConstant(0, VT);
7457 // FIXME: Really want ATOMIC_LOAD_NOP but that doesn't fit into the existing
7458 // scheme very well. Given the complexity of what we're already generating, an
7459 // extra couple of ORRs probably won't make much difference.
7460 SDValue Result = DAG.getAtomic(ISD::ATOMIC_LOAD_OR, DL, AN->getMemoryVT(),
7461 N->getOperand(0), N->getOperand(1), Zero,
7462 AN->getMemOperand(), AN->getOrdering(),
7463 AN->getSynchScope());
7465 Results.push_back(Result.getValue(0)); // Value
7466 Results.push_back(Result.getValue(1)); // Chain
7469 static void ReplaceATOMIC_OP_128(SDNode *N, SmallVectorImpl<SDValue> &Results,
7470 SelectionDAG &DAG, unsigned NewOp) {
7472 AtomicOrdering Ordering = cast<AtomicSDNode>(N)->getOrdering();
7473 assert(N->getValueType(0) == MVT::i128 &&
7474 "Only know how to expand i128 atomics");
7476 SmallVector<SDValue, 6> Ops;
7477 Ops.push_back(N->getOperand(1)); // Ptr
7479 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7480 N->getOperand(2), DAG.getIntPtrConstant(0)));
7481 // High part of Val1
7482 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7483 N->getOperand(2), DAG.getIntPtrConstant(1)));
7484 if (NewOp == ARM64::ATOMIC_CMP_SWAP_I128) {
7486 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7487 N->getOperand(3), DAG.getIntPtrConstant(0)));
7488 // High part of Val2
7489 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7490 N->getOperand(3), DAG.getIntPtrConstant(1)));
7493 Ops.push_back(DAG.getTargetConstant(Ordering, MVT::i32));
7494 Ops.push_back(N->getOperand(0)); // Chain
7496 SDVTList Tys = DAG.getVTList(MVT::i64, MVT::i64, MVT::Other);
7497 SDNode *Result = DAG.getMachineNode(NewOp, DL, Tys, Ops);
7498 SDValue OpsF[] = { SDValue(Result, 0), SDValue(Result, 1) };
7499 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i128, OpsF, 2));
7500 Results.push_back(SDValue(Result, 2));
7503 void ARM64TargetLowering::ReplaceNodeResults(SDNode *N,
7504 SmallVectorImpl<SDValue> &Results,
7505 SelectionDAG &DAG) const {
7506 switch (N->getOpcode()) {
7508 llvm_unreachable("Don't know how to custom expand this");
7509 case ISD::ATOMIC_LOAD:
7510 ReplaceATOMIC_LOAD_128(N, Results, DAG);
7512 case ISD::ATOMIC_LOAD_ADD:
7513 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_ADD_I128);
7515 case ISD::ATOMIC_LOAD_SUB:
7516 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_SUB_I128);
7518 case ISD::ATOMIC_LOAD_AND:
7519 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_AND_I128);
7521 case ISD::ATOMIC_LOAD_OR:
7522 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_OR_I128);
7524 case ISD::ATOMIC_LOAD_XOR:
7525 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_XOR_I128);
7527 case ISD::ATOMIC_LOAD_NAND:
7528 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_NAND_I128);
7530 case ISD::ATOMIC_SWAP:
7531 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_SWAP_I128);
7533 case ISD::ATOMIC_LOAD_MIN:
7534 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_MIN_I128);
7536 case ISD::ATOMIC_LOAD_MAX:
7537 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_MAX_I128);
7539 case ISD::ATOMIC_LOAD_UMIN:
7540 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_UMIN_I128);
7542 case ISD::ATOMIC_LOAD_UMAX:
7543 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_UMAX_I128);
7545 case ISD::ATOMIC_CMP_SWAP:
7546 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_CMP_SWAP_I128);
7548 case ISD::FP_TO_UINT:
7549 case ISD::FP_TO_SINT:
7550 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
7551 // Let normal code take care of it by not adding anything to Results.