1 //===-- ARM64ISelLowering.cpp - ARM64 DAG Lowering Implementation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the ARM64TargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "arm64-lower"
16 #include "ARM64ISelLowering.h"
17 #include "ARM64PerfectShuffle.h"
18 #include "ARM64Subtarget.h"
19 #include "ARM64CallingConv.h"
20 #include "ARM64MachineFunctionInfo.h"
21 #include "ARM64TargetMachine.h"
22 #include "ARM64TargetObjectFile.h"
23 #include "MCTargetDesc/ARM64AddressingModes.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/Intrinsics.h"
31 #include "llvm/IR/Type.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetOptions.h"
39 STATISTIC(NumTailCalls, "Number of tail calls");
40 STATISTIC(NumShiftInserts, "Number of vector shift inserts");
42 // This option should go away when tail calls fully work.
44 EnableARM64TailCalls("arm64-tail-calls", cl::Hidden,
45 cl::desc("Generate ARM64 tail calls (TEMPORARY OPTION)."),
49 StrictAlign("arm64-strict-align", cl::Hidden,
50 cl::desc("Disallow all unaligned memory accesses"));
52 // Place holder until extr generation is tested fully.
54 EnableARM64ExtrGeneration("arm64-extr-generation", cl::Hidden,
55 cl::desc("Allow ARM64 (or (shift)(shift))->extract"),
59 EnableARM64SlrGeneration("arm64-shift-insert-generation", cl::Hidden,
60 cl::desc("Allow ARM64 SLI/SRI formation"),
63 //===----------------------------------------------------------------------===//
64 // ARM64 Lowering public interface.
65 //===----------------------------------------------------------------------===//
66 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
67 if (TM.getSubtarget<ARM64Subtarget>().isTargetDarwin())
68 return new ARM64_MachoTargetObjectFile();
70 return new ARM64_ELFTargetObjectFile();
73 ARM64TargetLowering::ARM64TargetLowering(ARM64TargetMachine &TM)
74 : TargetLowering(TM, createTLOF(TM)) {
75 Subtarget = &TM.getSubtarget<ARM64Subtarget>();
77 // ARM64 doesn't have comparisons which set GPRs or setcc instructions, so
78 // we have to make something up. Arbitrarily, choose ZeroOrOne.
79 setBooleanContents(ZeroOrOneBooleanContent);
80 // When comparing vectors the result sets the different elements in the
81 // vector to all-one or all-zero.
82 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
84 // Set up the register classes.
85 addRegisterClass(MVT::i32, &ARM64::GPR32allRegClass);
86 addRegisterClass(MVT::i64, &ARM64::GPR64allRegClass);
87 addRegisterClass(MVT::f16, &ARM64::FPR16RegClass);
88 addRegisterClass(MVT::f32, &ARM64::FPR32RegClass);
89 addRegisterClass(MVT::f64, &ARM64::FPR64RegClass);
90 addRegisterClass(MVT::f128, &ARM64::FPR128RegClass);
91 addRegisterClass(MVT::v16i8, &ARM64::FPR8RegClass);
92 addRegisterClass(MVT::v8i16, &ARM64::FPR16RegClass);
94 // Someone set us up the NEON.
95 addDRTypeForNEON(MVT::v2f32);
96 addDRTypeForNEON(MVT::v8i8);
97 addDRTypeForNEON(MVT::v4i16);
98 addDRTypeForNEON(MVT::v2i32);
99 addDRTypeForNEON(MVT::v1i64);
100 addDRTypeForNEON(MVT::v1f64);
102 addQRTypeForNEON(MVT::v4f32);
103 addQRTypeForNEON(MVT::v2f64);
104 addQRTypeForNEON(MVT::v16i8);
105 addQRTypeForNEON(MVT::v8i16);
106 addQRTypeForNEON(MVT::v4i32);
107 addQRTypeForNEON(MVT::v2i64);
109 // Compute derived properties from the register classes
110 computeRegisterProperties();
112 // Provide all sorts of operation actions
113 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
114 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
115 setOperationAction(ISD::SETCC, MVT::i32, Custom);
116 setOperationAction(ISD::SETCC, MVT::i64, Custom);
117 setOperationAction(ISD::SETCC, MVT::f32, Custom);
118 setOperationAction(ISD::SETCC, MVT::f64, Custom);
119 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
120 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
121 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
122 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
123 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
124 setOperationAction(ISD::SELECT, MVT::i32, Custom);
125 setOperationAction(ISD::SELECT, MVT::i64, Custom);
126 setOperationAction(ISD::SELECT, MVT::f32, Custom);
127 setOperationAction(ISD::SELECT, MVT::f64, Custom);
128 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
129 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
130 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
131 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
132 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
133 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
135 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
136 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
137 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
139 setOperationAction(ISD::FREM, MVT::f32, Expand);
140 setOperationAction(ISD::FREM, MVT::f64, Expand);
141 setOperationAction(ISD::FREM, MVT::f80, Expand);
143 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
144 // silliness like this:
145 setOperationAction(ISD::FABS, MVT::v1f64, Expand);
146 setOperationAction(ISD::FADD, MVT::v1f64, Expand);
147 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
148 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
149 setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
150 setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
151 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
152 setOperationAction(ISD::FMA, MVT::v1f64, Expand);
153 setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
154 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
155 setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
156 setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
157 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
158 setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
159 setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
160 setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
161 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
162 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
163 setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
164 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
165 setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
166 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
167 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
168 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
169 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);
171 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
172 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
173 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
174 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
175 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);
177 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
179 // Custom lowering hooks are needed for XOR
180 // to fold it into CSINC/CSINV.
181 setOperationAction(ISD::XOR, MVT::i32, Custom);
182 setOperationAction(ISD::XOR, MVT::i64, Custom);
184 // Virtually no operation on f128 is legal, but LLVM can't expand them when
185 // there's a valid register class, so we need custom operations in most cases.
186 setOperationAction(ISD::FABS, MVT::f128, Expand);
187 setOperationAction(ISD::FADD, MVT::f128, Custom);
188 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
189 setOperationAction(ISD::FCOS, MVT::f128, Expand);
190 setOperationAction(ISD::FDIV, MVT::f128, Custom);
191 setOperationAction(ISD::FMA, MVT::f128, Expand);
192 setOperationAction(ISD::FMUL, MVT::f128, Custom);
193 setOperationAction(ISD::FNEG, MVT::f128, Expand);
194 setOperationAction(ISD::FPOW, MVT::f128, Expand);
195 setOperationAction(ISD::FREM, MVT::f128, Expand);
196 setOperationAction(ISD::FRINT, MVT::f128, Expand);
197 setOperationAction(ISD::FSIN, MVT::f128, Expand);
198 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
199 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
200 setOperationAction(ISD::FSUB, MVT::f128, Custom);
201 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
202 setOperationAction(ISD::SETCC, MVT::f128, Custom);
203 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
204 setOperationAction(ISD::SELECT, MVT::f128, Custom);
205 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
206 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
208 // Lowering for many of the conversions is actually specified by the non-f128
209 // type. The LowerXXX function will be trivial when f128 isn't involved.
210 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
211 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
212 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
213 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
214 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
215 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
216 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
217 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
218 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
219 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
220 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
221 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
222 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
223 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
225 // Variable arguments.
226 setOperationAction(ISD::VASTART, MVT::Other, Custom);
227 setOperationAction(ISD::VAARG, MVT::Other, Custom);
228 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
229 setOperationAction(ISD::VAEND, MVT::Other, Expand);
231 // Variable-sized objects.
232 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
233 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
234 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
236 // Exception handling.
237 // FIXME: These are guesses. Has this been defined yet?
238 setExceptionPointerRegister(ARM64::X0);
239 setExceptionSelectorRegister(ARM64::X1);
241 // Constant pool entries
242 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
245 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
247 // Add/Sub overflow ops with MVT::Glues are lowered to CPSR dependences.
248 setOperationAction(ISD::ADDC, MVT::i32, Custom);
249 setOperationAction(ISD::ADDE, MVT::i32, Custom);
250 setOperationAction(ISD::SUBC, MVT::i32, Custom);
251 setOperationAction(ISD::SUBE, MVT::i32, Custom);
252 setOperationAction(ISD::ADDC, MVT::i64, Custom);
253 setOperationAction(ISD::ADDE, MVT::i64, Custom);
254 setOperationAction(ISD::SUBC, MVT::i64, Custom);
255 setOperationAction(ISD::SUBE, MVT::i64, Custom);
257 // ARM64 lacks both left-rotate and popcount instructions.
258 setOperationAction(ISD::ROTL, MVT::i32, Expand);
259 setOperationAction(ISD::ROTL, MVT::i64, Expand);
261 // ARM64 doesn't have a direct vector ->f32 conversion instructions for
262 // elements smaller than i32, so promote the input to i32 first.
263 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
264 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
265 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
266 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
267 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
268 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
269 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
270 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
271 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
273 // ARM64 doesn't have {U|S}MUL_LOHI.
274 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
275 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
277 // ARM64 doesn't have MUL.2d:
278 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
280 // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
281 // counterparts, which ARM64 supports directly.
282 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
283 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
284 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
285 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
287 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
288 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
290 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
291 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
292 setOperationAction(ISD::SREM, MVT::i32, Expand);
293 setOperationAction(ISD::SREM, MVT::i64, Expand);
294 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
295 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
296 setOperationAction(ISD::UREM, MVT::i32, Expand);
297 setOperationAction(ISD::UREM, MVT::i64, Expand);
299 // Custom lower Add/Sub/Mul with overflow.
300 setOperationAction(ISD::SADDO, MVT::i32, Custom);
301 setOperationAction(ISD::SADDO, MVT::i64, Custom);
302 setOperationAction(ISD::UADDO, MVT::i32, Custom);
303 setOperationAction(ISD::UADDO, MVT::i64, Custom);
304 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
305 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
306 setOperationAction(ISD::USUBO, MVT::i32, Custom);
307 setOperationAction(ISD::USUBO, MVT::i64, Custom);
308 setOperationAction(ISD::SMULO, MVT::i32, Custom);
309 setOperationAction(ISD::SMULO, MVT::i64, Custom);
310 setOperationAction(ISD::UMULO, MVT::i32, Custom);
311 setOperationAction(ISD::UMULO, MVT::i64, Custom);
313 setOperationAction(ISD::FSIN, MVT::f32, Expand);
314 setOperationAction(ISD::FSIN, MVT::f64, Expand);
315 setOperationAction(ISD::FCOS, MVT::f32, Expand);
316 setOperationAction(ISD::FCOS, MVT::f64, Expand);
317 setOperationAction(ISD::FPOW, MVT::f32, Expand);
318 setOperationAction(ISD::FPOW, MVT::f64, Expand);
319 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
320 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
322 // ARM64 has implementations of a lot of rounding-like FP operations.
323 static MVT RoundingTypes[] = { MVT::f32, MVT::f64, MVT::v2f32,
324 MVT::v4f32, MVT::v2f64 };
325 for (unsigned I = 0; I < array_lengthof(RoundingTypes); ++I) {
326 MVT Ty = RoundingTypes[I];
327 setOperationAction(ISD::FFLOOR, Ty, Legal);
328 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
329 setOperationAction(ISD::FCEIL, Ty, Legal);
330 setOperationAction(ISD::FRINT, Ty, Legal);
331 setOperationAction(ISD::FTRUNC, Ty, Legal);
332 setOperationAction(ISD::FROUND, Ty, Legal);
335 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
337 if (Subtarget->isTargetMachO()) {
338 // For iOS, we don't want to the normal expansion of a libcall to
339 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
341 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
342 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
344 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
345 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
348 // ARM64 does not have floating-point extending loads, i1 sign-extending load,
349 // floating-point truncating stores, or v2i32->v2i16 truncating store.
350 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
351 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
352 setLoadExtAction(ISD::EXTLOAD, MVT::f80, Expand);
353 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Expand);
354 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
355 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
356 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
357 setTruncStoreAction(MVT::f128, MVT::f80, Expand);
358 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
359 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
360 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
361 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
362 // Indexed loads and stores are supported.
363 for (unsigned im = (unsigned)ISD::PRE_INC;
364 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
365 setIndexedLoadAction(im, MVT::i8, Legal);
366 setIndexedLoadAction(im, MVT::i16, Legal);
367 setIndexedLoadAction(im, MVT::i32, Legal);
368 setIndexedLoadAction(im, MVT::i64, Legal);
369 setIndexedLoadAction(im, MVT::f64, Legal);
370 setIndexedLoadAction(im, MVT::f32, Legal);
371 setIndexedStoreAction(im, MVT::i8, Legal);
372 setIndexedStoreAction(im, MVT::i16, Legal);
373 setIndexedStoreAction(im, MVT::i32, Legal);
374 setIndexedStoreAction(im, MVT::i64, Legal);
375 setIndexedStoreAction(im, MVT::f64, Legal);
376 setIndexedStoreAction(im, MVT::f32, Legal);
379 // Likewise, narrowing and extending vector loads/stores aren't handled
381 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
382 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
384 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,
387 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
388 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
389 setTruncStoreAction((MVT::SimpleValueType)VT,
390 (MVT::SimpleValueType)InnerVT, Expand);
391 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
392 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
393 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
397 setOperationAction(ISD::TRAP, MVT::Other, Legal);
398 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
400 // We combine OR nodes for bitfield operations.
401 setTargetDAGCombine(ISD::OR);
403 // Vector add and sub nodes may conceal a high-half opportunity.
404 // Also, try to fold ADD into CSINC/CSINV..
405 setTargetDAGCombine(ISD::ADD);
406 setTargetDAGCombine(ISD::SUB);
408 setTargetDAGCombine(ISD::XOR);
409 setTargetDAGCombine(ISD::SINT_TO_FP);
410 setTargetDAGCombine(ISD::UINT_TO_FP);
412 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
414 setTargetDAGCombine(ISD::ANY_EXTEND);
415 setTargetDAGCombine(ISD::ZERO_EXTEND);
416 setTargetDAGCombine(ISD::SIGN_EXTEND);
417 setTargetDAGCombine(ISD::BITCAST);
418 setTargetDAGCombine(ISD::CONCAT_VECTORS);
419 setTargetDAGCombine(ISD::STORE);
421 setTargetDAGCombine(ISD::MUL);
423 setTargetDAGCombine(ISD::VSELECT);
425 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
426 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
427 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
429 setStackPointerRegisterToSaveRestore(ARM64::SP);
431 setSchedulingPreference(Sched::Hybrid);
434 MaskAndBranchFoldingIsLegal = true;
436 setMinFunctionAlignment(2);
438 RequireStrictAlign = StrictAlign;
441 void ARM64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
442 if (VT == MVT::v2f32) {
443 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
444 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
446 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
447 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
448 } else if (VT == MVT::v2f64 || VT == MVT::v4f32) {
449 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
450 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
452 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
453 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
456 // Mark vector float intrinsics as expand.
457 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
458 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
459 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
460 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
461 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
462 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
463 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
464 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
465 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
466 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
469 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
470 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
471 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
472 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
473 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
474 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
475 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
476 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
477 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
478 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
479 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
480 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
482 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
483 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
484 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
485 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
487 // CNT supports only B element sizes.
488 if (VT != MVT::v8i8 && VT != MVT::v16i8)
489 setOperationAction(ISD::CTPOP, VT.getSimpleVT(), Expand);
491 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
492 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
493 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
494 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
495 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
497 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
498 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
501 void ARM64TargetLowering::addDRTypeForNEON(MVT VT) {
502 addRegisterClass(VT, &ARM64::FPR64RegClass);
503 addTypeForNEON(VT, MVT::v2i32);
506 void ARM64TargetLowering::addQRTypeForNEON(MVT VT) {
507 addRegisterClass(VT, &ARM64::FPR128RegClass);
508 addTypeForNEON(VT, MVT::v4i32);
511 EVT ARM64TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
514 return VT.changeVectorElementTypeToInteger();
517 /// computeMaskedBitsForTargetNode - Determine which of the bits specified in
518 /// Mask are known to be either zero or one and return them in the
519 /// KnownZero/KnownOne bitsets.
520 void ARM64TargetLowering::computeMaskedBitsForTargetNode(
521 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
522 const SelectionDAG &DAG, unsigned Depth) const {
523 switch (Op.getOpcode()) {
526 case ARM64ISD::CSEL: {
527 APInt KnownZero2, KnownOne2;
528 DAG.ComputeMaskedBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
529 DAG.ComputeMaskedBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
530 KnownZero &= KnownZero2;
531 KnownOne &= KnownOne2;
534 case ISD::INTRINSIC_W_CHAIN: {
535 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
536 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
539 case Intrinsic::arm64_ldaxr:
540 case Intrinsic::arm64_ldxr: {
541 unsigned BitWidth = KnownOne.getBitWidth();
542 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
543 unsigned MemBits = VT.getScalarType().getSizeInBits();
544 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
550 case ISD::INTRINSIC_WO_CHAIN:
551 case ISD::INTRINSIC_VOID: {
552 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
556 case Intrinsic::arm64_neon_umaxv:
557 case Intrinsic::arm64_neon_uminv: {
558 // Figure out the datatype of the vector operand. The UMINV instruction
559 // will zero extend the result, so we can mark as known zero all the
560 // bits larger than the element datatype. 32-bit or larget doesn't need
561 // this as those are legal types and will be handled by isel directly.
562 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
563 unsigned BitWidth = KnownZero.getBitWidth();
564 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
565 assert(BitWidth >= 8 && "Unexpected width!");
566 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
568 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
569 assert(BitWidth >= 16 && "Unexpected width!");
570 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
580 MVT ARM64TargetLowering::getScalarShiftAmountTy(EVT LHSTy) const {
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";
690 ARM64TargetLowering::EmitF128CSEL(MachineInstr *MI,
691 MachineBasicBlock *MBB) const {
692 // We materialise the F128CSEL pseudo-instruction as some control flow and a
696 // [... previous instrs leading to comparison ...]
702 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
704 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
705 MachineFunction *MF = MBB->getParent();
706 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
707 DebugLoc DL = MI->getDebugLoc();
708 MachineFunction::iterator It = MBB;
711 unsigned DestReg = MI->getOperand(0).getReg();
712 unsigned IfTrueReg = MI->getOperand(1).getReg();
713 unsigned IfFalseReg = MI->getOperand(2).getReg();
714 unsigned CondCode = MI->getOperand(3).getImm();
715 bool CPSRKilled = MI->getOperand(4).isKill();
717 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
718 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
719 MF->insert(It, TrueBB);
720 MF->insert(It, EndBB);
722 // Transfer rest of current basic-block to EndBB
723 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
725 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
727 BuildMI(MBB, DL, TII->get(ARM64::Bcc)).addImm(CondCode).addMBB(TrueBB);
728 BuildMI(MBB, DL, TII->get(ARM64::B)).addMBB(EndBB);
729 MBB->addSuccessor(TrueBB);
730 MBB->addSuccessor(EndBB);
732 // TrueBB falls through to the end.
733 TrueBB->addSuccessor(EndBB);
736 TrueBB->addLiveIn(ARM64::CPSR);
737 EndBB->addLiveIn(ARM64::CPSR);
740 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(ARM64::PHI), DestReg)
746 MI->eraseFromParent();
751 ARM64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
752 MachineBasicBlock *BB) const {
753 switch (MI->getOpcode()) {
758 assert(0 && "Unexpected instruction for custom inserter!");
761 case ARM64::F128CSEL:
762 return EmitF128CSEL(MI, BB);
764 case TargetOpcode::STACKMAP:
765 case TargetOpcode::PATCHPOINT:
766 return emitPatchPoint(MI, BB);
768 llvm_unreachable("Unexpected instruction for custom inserter!");
771 //===----------------------------------------------------------------------===//
772 // ARM64 Lowering private implementation.
773 //===----------------------------------------------------------------------===//
775 //===----------------------------------------------------------------------===//
777 //===----------------------------------------------------------------------===//
779 /// changeIntCCToARM64CC - Convert a DAG integer condition code to an ARM64 CC
780 static ARM64CC::CondCode changeIntCCToARM64CC(ISD::CondCode CC) {
783 llvm_unreachable("Unknown condition code!");
807 /// changeFPCCToARM64CC - Convert a DAG fp condition code to an ARM64 CC.
808 static void changeFPCCToARM64CC(ISD::CondCode CC, ARM64CC::CondCode &CondCode,
809 ARM64CC::CondCode &CondCode2) {
810 CondCode2 = ARM64CC::AL;
813 llvm_unreachable("Unknown FP condition!");
816 CondCode = ARM64CC::EQ;
820 CondCode = ARM64CC::GT;
824 CondCode = ARM64CC::GE;
827 CondCode = ARM64CC::MI;
830 CondCode = ARM64CC::LS;
833 CondCode = ARM64CC::MI;
834 CondCode2 = ARM64CC::GT;
837 CondCode = ARM64CC::VC;
840 CondCode = ARM64CC::VS;
843 CondCode = ARM64CC::EQ;
844 CondCode2 = ARM64CC::VS;
847 CondCode = ARM64CC::HI;
850 CondCode = ARM64CC::PL;
854 CondCode = ARM64CC::LT;
858 CondCode = ARM64CC::LE;
862 CondCode = ARM64CC::NE;
867 static bool isLegalArithImmed(uint64_t C) {
868 // Matches ARM64DAGToDAGISel::SelectArithImmed().
869 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
872 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
873 SDLoc dl, SelectionDAG &DAG) {
874 EVT VT = LHS.getValueType();
876 if (VT.isFloatingPoint())
877 return DAG.getNode(ARM64ISD::FCMP, dl, VT, LHS, RHS);
879 // The CMP instruction is just an alias for SUBS, and representing it as
880 // SUBS means that it's possible to get CSE with subtract operations.
881 // A later phase can perform the optimization of setting the destination
882 // register to WZR/XZR if it ends up being unused.
884 // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on the
885 // grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags can be
886 // set differently by this operation. It comes down to whether "SInt(~op2)+1
887 // == SInt(~op2+1)" (and the same for UInt). If they are then everything is
888 // fine. If not then the optimization is wrong. Thus general comparisons are
889 // only valid if op2 != 0.
891 // So, finally, the only LLVM-native comparisons that don't mention C and V
892 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in the
893 // absence of information about op2.
894 unsigned Opcode = ARM64ISD::SUBS;
895 if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
896 cast<ConstantSDNode>(RHS.getOperand(0))->getZExtValue() == 0 &&
897 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
898 Opcode = ARM64ISD::ADDS;
899 RHS = RHS.getOperand(1);
902 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS)
906 static SDValue getARM64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
907 SDValue &ARM64cc, SelectionDAG &DAG, SDLoc dl) {
908 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
909 EVT VT = RHS.getValueType();
910 uint64_t C = RHSC->getZExtValue();
911 if (!isLegalArithImmed(C)) {
912 // Constant does not fit, try adjusting it by one?
918 if ((VT == MVT::i32 && C != 0x80000000 &&
919 isLegalArithImmed((uint32_t)(C - 1))) ||
920 (VT == MVT::i64 && C != 0x80000000ULL &&
921 isLegalArithImmed(C - 1ULL))) {
922 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
923 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
924 RHS = DAG.getConstant(C, VT);
929 if ((VT == MVT::i32 && C != 0 &&
930 isLegalArithImmed((uint32_t)(C - 1))) ||
931 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
932 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
933 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
934 RHS = DAG.getConstant(C, VT);
939 if ((VT == MVT::i32 && C != 0x7fffffff &&
940 isLegalArithImmed((uint32_t)(C + 1))) ||
941 (VT == MVT::i64 && C != 0x7ffffffffffffffULL &&
942 isLegalArithImmed(C + 1ULL))) {
943 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
944 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
945 RHS = DAG.getConstant(C, VT);
950 if ((VT == MVT::i32 && C != 0xffffffff &&
951 isLegalArithImmed((uint32_t)(C + 1))) ||
952 (VT == MVT::i64 && C != 0xfffffffffffffffULL &&
953 isLegalArithImmed(C + 1ULL))) {
954 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
955 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
956 RHS = DAG.getConstant(C, VT);
963 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
964 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
965 ARM64cc = DAG.getConstant(ARM64CC, MVT::i32);
969 static std::pair<SDValue, SDValue>
970 getARM64XALUOOp(ARM64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
971 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
972 "Unsupported value type");
973 SDValue Value, Overflow;
975 SDValue LHS = Op.getOperand(0);
976 SDValue RHS = Op.getOperand(1);
978 switch (Op.getOpcode()) {
980 llvm_unreachable("Unknown overflow instruction!");
982 Opc = ARM64ISD::ADDS;
986 Opc = ARM64ISD::ADDS;
990 Opc = ARM64ISD::SUBS;
994 Opc = ARM64ISD::SUBS;
997 // Multiply needs a little bit extra work.
1001 bool IsSigned = (Op.getOpcode() == ISD::SMULO) ? true : false;
1002 if (Op.getValueType() == MVT::i32) {
1003 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1004 // For a 32 bit multiply with overflow check we want the instruction
1005 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1006 // need to generate the following pattern:
1007 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1008 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1009 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1010 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1011 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1012 DAG.getConstant(0, MVT::i64));
1013 // On ARM64 the upper 32 bits are always zero extended for a 32 bit
1014 // operation. We need to clear out the upper 32 bits, because we used a
1015 // widening multiply that wrote all 64 bits. In the end this should be a
1017 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1019 // The signed overflow check requires more than just a simple check for
1020 // any bit set in the upper 32 bits of the result. These bits could be
1021 // just the sign bits of a negative number. To perform the overflow
1022 // check we have to arithmetic shift right the 32nd bit of the result by
1023 // 31 bits. Then we compare the result to the upper 32 bits.
1024 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1025 DAG.getConstant(32, MVT::i64));
1026 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1027 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1028 DAG.getConstant(31, MVT::i64));
1029 // It is important that LowerBits is last, otherwise the arithmetic
1030 // shift will not be folded into the compare (SUBS).
1031 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1032 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1035 // The overflow check for unsigned multiply is easy. We only need to
1036 // check if any of the upper 32 bits are set. This can be done with a
1037 // CMP (shifted register). For that we need to generate the following
1039 // (i64 ARM64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1040 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1041 DAG.getConstant(32, MVT::i64));
1042 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1044 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1045 UpperBits).getValue(1);
1049 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1050 // For the 64 bit multiply
1051 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1053 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1054 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1055 DAG.getConstant(63, MVT::i64));
1056 // It is important that LowerBits is last, otherwise the arithmetic
1057 // shift will not be folded into the compare (SUBS).
1058 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1059 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1062 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1063 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1065 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1066 UpperBits).getValue(1);
1073 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1075 // Emit the ARM64 operation with overflow check.
1076 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1077 Overflow = Value.getValue(1);
1079 return std::make_pair(Value, Overflow);
1082 SDValue ARM64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1083 RTLIB::Libcall Call) const {
1084 SmallVector<SDValue, 2> Ops;
1085 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1086 Ops.push_back(Op.getOperand(i));
1088 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1092 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1093 SDValue Sel = Op.getOperand(0);
1094 SDValue Other = Op.getOperand(1);
1096 // If neither operand is a SELECT_CC, give up.
1097 if (Sel.getOpcode() != ISD::SELECT_CC)
1098 std::swap(Sel, Other);
1099 if (Sel.getOpcode() != ISD::SELECT_CC)
1102 // The folding we want to perform is:
1103 // (xor x, (select_cc a, b, cc, 0, -1) )
1105 // (csel x, (xor x, -1), cc ...)
1107 // The latter will get matched to a CSINV instruction.
1109 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1110 SDValue LHS = Sel.getOperand(0);
1111 SDValue RHS = Sel.getOperand(1);
1112 SDValue TVal = Sel.getOperand(2);
1113 SDValue FVal = Sel.getOperand(3);
1116 // FIXME: This could be generalized to non-integer comparisons.
1117 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1120 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1121 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1123 // The the values aren't constants, this isn't the pattern we're looking for.
1124 if (!CFVal || !CTVal)
1127 // We can commute the SELECT_CC by inverting the condition. This
1128 // might be needed to make this fit into a CSINV pattern.
1129 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1130 std::swap(TVal, FVal);
1131 std::swap(CTVal, CFVal);
1132 CC = ISD::getSetCCInverse(CC, true);
1135 // If the constants line up, perform the transform!
1136 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1138 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1141 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1142 DAG.getConstant(-1ULL, Other.getValueType()));
1144 return DAG.getNode(ARM64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1151 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1152 EVT VT = Op.getValueType();
1154 // Let legalize expand this if it isn't a legal type yet.
1155 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1158 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1161 bool ExtraOp = false;
1162 switch (Op.getOpcode()) {
1164 assert(0 && "Invalid code");
1166 Opc = ARM64ISD::ADDS;
1169 Opc = ARM64ISD::SUBS;
1172 Opc = ARM64ISD::ADCS;
1176 Opc = ARM64ISD::SBCS;
1182 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1183 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1187 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1188 // Let legalize expand this if it isn't a legal type yet.
1189 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1192 ARM64CC::CondCode CC;
1193 // The actual operation that sets the overflow or carry flag.
1194 SDValue Value, Overflow;
1195 std::tie(Value, Overflow) = getARM64XALUOOp(CC, Op, DAG);
1197 // We use 0 and 1 as false and true values.
1198 SDValue TVal = DAG.getConstant(1, MVT::i32);
1199 SDValue FVal = DAG.getConstant(0, MVT::i32);
1201 // We use an inverted condition, because the conditional select is inverted
1202 // too. This will allow it to be selected to a single instruction:
1203 // CSINC Wd, WZR, WZR, invert(cond).
1204 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), MVT::i32);
1205 Overflow = DAG.getNode(ARM64ISD::CSEL, SDLoc(Op), MVT::i32, FVal, TVal, CCVal,
1208 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1209 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
1212 // Prefetch operands are:
1213 // 1: Address to prefetch
1215 // 3: int locality (0 = no locality ... 3 = extreme locality)
1216 // 4: bool isDataCache
1217 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1219 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1220 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1221 // The data thing is not used.
1222 // unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1224 bool IsStream = !Locality;
1225 // When the locality number is set
1227 // The front-end should have filtered out the out-of-range values
1228 assert(Locality <= 3 && "Prefetch locality out-of-range");
1229 // The locality degree is the opposite of the cache speed.
1230 // Put the number the other way around.
1231 // The encoding starts at 0 for level 1
1232 Locality = 3 - Locality;
1235 // built the mask value encoding the expected behavior.
1236 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1237 (Locality << 1) | // Cache level bits
1238 (unsigned)IsStream; // Stream bit
1239 return DAG.getNode(ARM64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1240 DAG.getConstant(PrfOp, MVT::i32), Op.getOperand(1));
1243 SDValue ARM64TargetLowering::LowerFP_EXTEND(SDValue Op,
1244 SelectionDAG &DAG) const {
1245 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1248 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1250 return LowerF128Call(Op, DAG, LC);
1253 SDValue ARM64TargetLowering::LowerFP_ROUND(SDValue Op,
1254 SelectionDAG &DAG) const {
1255 if (Op.getOperand(0).getValueType() != MVT::f128) {
1256 // It's legal except when f128 is involved
1261 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1263 // FP_ROUND node has a second operand indicating whether it is known to be
1264 // precise. That doesn't take part in the LibCall so we can't directly use
1266 SDValue SrcVal = Op.getOperand(0);
1267 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1268 /*isSigned*/ false, SDLoc(Op)).first;
1271 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1272 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1273 // Any additional optimization in this function should be recorded
1274 // in the cost tables.
1275 EVT InVT = Op.getOperand(0).getValueType();
1276 EVT VT = Op.getValueType();
1278 // FP_TO_XINT conversion from the same type are legal.
1279 if (VT.getSizeInBits() == InVT.getSizeInBits())
1282 if (InVT == MVT::v2f64) {
1284 SDValue Cv = DAG.getNode(Op.getOpcode(), dl, MVT::v2i64, Op.getOperand(0));
1285 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1288 // Type changing conversions are illegal.
1292 SDValue ARM64TargetLowering::LowerFP_TO_INT(SDValue Op,
1293 SelectionDAG &DAG) const {
1294 if (Op.getOperand(0).getValueType().isVector())
1295 return LowerVectorFP_TO_INT(Op, DAG);
1297 if (Op.getOperand(0).getValueType() != MVT::f128) {
1298 // It's legal except when f128 is involved
1303 if (Op.getOpcode() == ISD::FP_TO_SINT)
1304 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1306 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1308 SmallVector<SDValue, 2> Ops;
1309 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1310 Ops.push_back(Op.getOperand(i));
1312 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1316 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1317 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1318 // Any additional optimization in this function should be recorded
1319 // in the cost tables.
1320 EVT VT = Op.getValueType();
1322 SDValue In = Op.getOperand(0);
1323 EVT InVT = In.getValueType();
1325 // v2i32 to v2f32 is legal.
1326 if (VT == MVT::v2f32 && InVT == MVT::v2i32)
1329 // This function only handles v2f64 outputs.
1330 if (VT == MVT::v2f64) {
1331 // Extend the input argument to a v2i64 that we can feed into the
1332 // floating point conversion. Zero or sign extend based on whether
1333 // we're doing a signed or unsigned float conversion.
1335 Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
1336 assert(Op.getNumOperands() == 1 && "FP conversions take one argument");
1337 SDValue Promoted = DAG.getNode(Opc, dl, MVT::v2i64, Op.getOperand(0));
1338 return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), Promoted);
1341 // Scalarize v2i64 to v2f32 conversions.
1342 std::vector<SDValue> BuildVectorOps;
1343 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
1344 SDValue Sclr = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, In,
1345 DAG.getConstant(i, MVT::i64));
1346 Sclr = DAG.getNode(Op->getOpcode(), dl, MVT::f32, Sclr);
1347 BuildVectorOps.push_back(Sclr);
1350 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &BuildVectorOps[0],
1351 BuildVectorOps.size());
1354 SDValue ARM64TargetLowering::LowerINT_TO_FP(SDValue Op,
1355 SelectionDAG &DAG) const {
1356 if (Op.getValueType().isVector())
1357 return LowerVectorINT_TO_FP(Op, DAG);
1359 // i128 conversions are libcalls.
1360 if (Op.getOperand(0).getValueType() == MVT::i128)
1363 // Other conversions are legal, unless it's to the completely software-based
1365 if (Op.getValueType() != MVT::f128)
1369 if (Op.getOpcode() == ISD::SINT_TO_FP)
1370 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1372 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1374 return LowerF128Call(Op, DAG, LC);
1377 SDValue ARM64TargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
1378 // For iOS, we want to call an alternative entry point: __sincos_stret,
1379 // which returns the values in two S / D registers.
1381 SDValue Arg = Op.getOperand(0);
1382 EVT ArgVT = Arg.getValueType();
1383 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1390 Entry.isSExt = false;
1391 Entry.isZExt = false;
1392 Args.push_back(Entry);
1394 const char *LibcallName =
1395 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1396 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
1398 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
1399 TargetLowering::CallLoweringInfo CLI(
1400 DAG.getEntryNode(), RetTy, false, false, false, false, 0,
1401 CallingConv::Fast, /*isTaillCall=*/false,
1402 /*doesNotRet=*/false, /*isReturnValueUsed*/ true, Callee, Args, DAG, dl);
1403 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1404 return CallResult.first;
1407 SDValue ARM64TargetLowering::LowerOperation(SDValue Op,
1408 SelectionDAG &DAG) const {
1409 switch (Op.getOpcode()) {
1411 llvm_unreachable("unimplemented operand");
1413 case ISD::GlobalAddress:
1414 return LowerGlobalAddress(Op, DAG);
1415 case ISD::GlobalTLSAddress:
1416 return LowerGlobalTLSAddress(Op, DAG);
1418 return LowerSETCC(Op, DAG);
1420 return LowerBR_CC(Op, DAG);
1422 return LowerSELECT(Op, DAG);
1423 case ISD::SELECT_CC:
1424 return LowerSELECT_CC(Op, DAG);
1425 case ISD::JumpTable:
1426 return LowerJumpTable(Op, DAG);
1427 case ISD::ConstantPool:
1428 return LowerConstantPool(Op, DAG);
1429 case ISD::BlockAddress:
1430 return LowerBlockAddress(Op, DAG);
1432 return LowerVASTART(Op, DAG);
1434 return LowerVACOPY(Op, DAG);
1436 return LowerVAARG(Op, DAG);
1441 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
1448 return LowerXALUO(Op, DAG);
1450 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
1452 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
1454 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
1456 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
1458 return LowerFP_ROUND(Op, DAG);
1459 case ISD::FP_EXTEND:
1460 return LowerFP_EXTEND(Op, DAG);
1461 case ISD::FRAMEADDR:
1462 return LowerFRAMEADDR(Op, DAG);
1463 case ISD::RETURNADDR:
1464 return LowerRETURNADDR(Op, DAG);
1465 case ISD::INSERT_VECTOR_ELT:
1466 return LowerINSERT_VECTOR_ELT(Op, DAG);
1467 case ISD::EXTRACT_VECTOR_ELT:
1468 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
1469 case ISD::BUILD_VECTOR:
1470 return LowerBUILD_VECTOR(Op, DAG);
1471 case ISD::VECTOR_SHUFFLE:
1472 return LowerVECTOR_SHUFFLE(Op, DAG);
1473 case ISD::EXTRACT_SUBVECTOR:
1474 return LowerEXTRACT_SUBVECTOR(Op, DAG);
1478 return LowerVectorSRA_SRL_SHL(Op, DAG);
1479 case ISD::SHL_PARTS:
1480 return LowerShiftLeftParts(Op, DAG);
1481 case ISD::SRL_PARTS:
1482 case ISD::SRA_PARTS:
1483 return LowerShiftRightParts(Op, DAG);
1485 return LowerCTPOP(Op, DAG);
1486 case ISD::FCOPYSIGN:
1487 return LowerFCOPYSIGN(Op, DAG);
1489 return LowerVectorAND(Op, DAG);
1491 return LowerVectorOR(Op, DAG);
1493 return LowerXOR(Op, DAG);
1495 return LowerPREFETCH(Op, DAG);
1496 case ISD::SINT_TO_FP:
1497 case ISD::UINT_TO_FP:
1498 return LowerINT_TO_FP(Op, DAG);
1499 case ISD::FP_TO_SINT:
1500 case ISD::FP_TO_UINT:
1501 return LowerFP_TO_INT(Op, DAG);
1503 return LowerFSINCOS(Op, DAG);
1507 /// getFunctionAlignment - Return the Log2 alignment of this function.
1508 unsigned ARM64TargetLowering::getFunctionAlignment(const Function *F) const {
1512 //===----------------------------------------------------------------------===//
1513 // Calling Convention Implementation
1514 //===----------------------------------------------------------------------===//
1516 #include "ARM64GenCallingConv.inc"
1518 /// Selects the correct CCAssignFn for a the given CallingConvention
1520 CCAssignFn *ARM64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
1521 bool IsVarArg) const {
1524 llvm_unreachable("Unsupported calling convention.");
1525 case CallingConv::WebKit_JS:
1526 return CC_ARM64_WebKit_JS;
1527 case CallingConv::C:
1528 case CallingConv::Fast:
1529 if (!Subtarget->isTargetDarwin())
1530 return CC_ARM64_AAPCS;
1531 return IsVarArg ? CC_ARM64_DarwinPCS_VarArg : CC_ARM64_DarwinPCS;
1535 SDValue ARM64TargetLowering::LowerFormalArguments(
1536 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1537 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
1538 SmallVectorImpl<SDValue> &InVals) const {
1539 MachineFunction &MF = DAG.getMachineFunction();
1540 MachineFrameInfo *MFI = MF.getFrameInfo();
1542 // Assign locations to all of the incoming arguments.
1543 SmallVector<CCValAssign, 16> ArgLocs;
1544 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1545 getTargetMachine(), ArgLocs, *DAG.getContext());
1547 // At this point, Ins[].VT may already be promoted to i32. To correctly
1548 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
1549 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
1550 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
1551 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
1553 unsigned NumArgs = Ins.size();
1554 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
1555 unsigned CurArgIdx = 0;
1556 for (unsigned i = 0; i != NumArgs; ++i) {
1557 MVT ValVT = Ins[i].VT;
1558 std::advance(CurOrigArg, Ins[i].OrigArgIndex - CurArgIdx);
1559 CurArgIdx = Ins[i].OrigArgIndex;
1561 // Get type of the original argument.
1562 EVT ActualVT = getValueType(CurOrigArg->getType(), /*AllowUnknown*/ true);
1563 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
1564 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
1566 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
1568 else if (ActualMVT == MVT::i16)
1571 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
1573 AssignFn(i, ValVT, LocVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
1574 assert(!Res && "Call operand has unhandled type");
1577 assert(ArgLocs.size() == Ins.size());
1578 SmallVector<SDValue, 16> ArgValues;
1579 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1580 CCValAssign &VA = ArgLocs[i];
1582 if (Ins[i].Flags.isByVal()) {
1583 // Byval is used for HFAs in the PCS, but the system should work in a
1584 // non-compliant manner for larger structs.
1585 EVT PtrTy = getPointerTy();
1586 int Size = Ins[i].Flags.getByValSize();
1587 unsigned NumRegs = (Size + 7) / 8;
1590 MFI->CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false);
1591 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrTy);
1592 InVals.push_back(FrameIdxN);
1595 } if (VA.isRegLoc()) {
1596 // Arguments stored in registers.
1597 EVT RegVT = VA.getLocVT();
1600 const TargetRegisterClass *RC;
1602 if (RegVT == MVT::i32)
1603 RC = &ARM64::GPR32RegClass;
1604 else if (RegVT == MVT::i64)
1605 RC = &ARM64::GPR64RegClass;
1606 else if (RegVT == MVT::f32)
1607 RC = &ARM64::FPR32RegClass;
1608 else if (RegVT == MVT::f64 || RegVT == MVT::v1i64 ||
1609 RegVT == MVT::v1f64 || RegVT == MVT::v2i32 ||
1610 RegVT == MVT::v4i16 || RegVT == MVT::v8i8)
1611 RC = &ARM64::FPR64RegClass;
1612 else if (RegVT == MVT::v2i64 || RegVT == MVT::v4i32 ||
1613 RegVT == MVT::v8i16 || RegVT == MVT::v16i8)
1614 RC = &ARM64::FPR128RegClass;
1616 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1618 // Transform the arguments in physical registers into virtual ones.
1619 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1620 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
1622 // If this is an 8, 16 or 32-bit value, it is really passed promoted
1623 // to 64 bits. Insert an assert[sz]ext to capture this, then
1624 // truncate to the right size.
1625 switch (VA.getLocInfo()) {
1627 llvm_unreachable("Unknown loc info!");
1628 case CCValAssign::Full:
1630 case CCValAssign::BCvt:
1631 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
1633 case CCValAssign::SExt:
1634 ArgValue = DAG.getNode(ISD::AssertSext, DL, RegVT, ArgValue,
1635 DAG.getValueType(VA.getValVT()));
1636 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
1638 case CCValAssign::ZExt:
1639 ArgValue = DAG.getNode(ISD::AssertZext, DL, RegVT, ArgValue,
1640 DAG.getValueType(VA.getValVT()));
1641 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
1645 InVals.push_back(ArgValue);
1647 } else { // VA.isRegLoc()
1648 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
1649 unsigned ArgOffset = VA.getLocMemOffset();
1650 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1651 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset, true);
1653 // Create load nodes to retrieve arguments from the stack.
1654 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1655 InVals.push_back(DAG.getLoad(VA.getValVT(), DL, Chain, FIN,
1656 MachinePointerInfo::getFixedStack(FI), false,
1663 if (!Subtarget->isTargetDarwin()) {
1664 // The AAPCS variadic function ABI is identical to the non-variadic
1665 // one. As a result there may be more arguments in registers and we should
1666 // save them for future reference.
1667 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
1670 ARM64FunctionInfo *AFI = MF.getInfo<ARM64FunctionInfo>();
1671 // This will point to the next argument passed via stack.
1672 unsigned StackOffset = CCInfo.getNextStackOffset();
1673 // We currently pass all varargs at 8-byte alignment.
1674 StackOffset = ((StackOffset + 7) & ~7);
1675 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
1681 void ARM64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
1682 SelectionDAG &DAG, SDLoc DL,
1683 SDValue &Chain) const {
1684 MachineFunction &MF = DAG.getMachineFunction();
1685 MachineFrameInfo *MFI = MF.getFrameInfo();
1686 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
1688 SmallVector<SDValue, 8> MemOps;
1690 static const MCPhysReg GPRArgRegs[] = { ARM64::X0, ARM64::X1, ARM64::X2,
1691 ARM64::X3, ARM64::X4, ARM64::X5,
1692 ARM64::X6, ARM64::X7 };
1693 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
1694 unsigned FirstVariadicGPR =
1695 CCInfo.getFirstUnallocated(GPRArgRegs, NumGPRArgRegs);
1697 static const MCPhysReg FPRArgRegs[] = { ARM64::Q0, ARM64::Q1, ARM64::Q2,
1698 ARM64::Q3, ARM64::Q4, ARM64::Q5,
1699 ARM64::Q6, ARM64::Q7 };
1700 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
1701 unsigned FirstVariadicFPR =
1702 CCInfo.getFirstUnallocated(FPRArgRegs, NumFPRArgRegs);
1704 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
1706 if (GPRSaveSize != 0) {
1707 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
1709 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
1711 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
1712 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &ARM64::GPR64RegClass);
1713 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
1715 DAG.getStore(Val.getValue(1), DL, Val, FIN,
1716 MachinePointerInfo::getStack(i * 8), false, false, 0);
1717 MemOps.push_back(Store);
1718 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
1719 DAG.getConstant(8, getPointerTy()));
1723 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
1725 if (FPRSaveSize != 0) {
1726 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
1728 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
1730 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
1731 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &ARM64::FPR128RegClass);
1732 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::v2i64);
1734 DAG.getStore(Val.getValue(1), DL, Val, FIN,
1735 MachinePointerInfo::getStack(i * 16), false, false, 0);
1736 MemOps.push_back(Store);
1737 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
1738 DAG.getConstant(16, getPointerTy()));
1742 FuncInfo->setVarArgsGPRIndex(GPRIdx);
1743 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
1744 FuncInfo->setVarArgsFPRIndex(FPRIdx);
1745 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
1747 if (!MemOps.empty()) {
1748 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
1753 /// LowerCallResult - Lower the result values of a call into the
1754 /// appropriate copies out of appropriate physical registers.
1755 SDValue ARM64TargetLowering::LowerCallResult(
1756 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
1757 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
1758 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
1759 SDValue ThisVal) const {
1760 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
1761 : RetCC_ARM64_AAPCS;
1762 // Assign locations to each value returned by this call.
1763 SmallVector<CCValAssign, 16> RVLocs;
1764 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1765 getTargetMachine(), RVLocs, *DAG.getContext());
1766 CCInfo.AnalyzeCallResult(Ins, RetCC);
1768 // Copy all of the result registers out of their specified physreg.
1769 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1770 CCValAssign VA = RVLocs[i];
1772 // Pass 'this' value directly from the argument to return value, to avoid
1773 // reg unit interference
1774 if (i == 0 && isThisReturn) {
1775 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
1776 "unexpected return calling convention register assignment");
1777 InVals.push_back(ThisVal);
1782 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
1783 Chain = Val.getValue(1);
1784 InFlag = Val.getValue(2);
1786 switch (VA.getLocInfo()) {
1788 llvm_unreachable("Unknown loc info!");
1789 case CCValAssign::Full:
1791 case CCValAssign::BCvt:
1792 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
1796 InVals.push_back(Val);
1802 bool ARM64TargetLowering::isEligibleForTailCallOptimization(
1803 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
1804 bool isCalleeStructRet, bool isCallerStructRet,
1805 const SmallVectorImpl<ISD::OutputArg> &Outs,
1806 const SmallVectorImpl<SDValue> &OutVals,
1807 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
1808 // Look for obvious safe cases to perform tail call optimization that do not
1809 // require ABI changes. This is what gcc calls sibcall.
1811 // Do not sibcall optimize vararg calls unless the call site is not passing
1813 if (isVarArg && !Outs.empty())
1816 // Also avoid sibcall optimization if either caller or callee uses struct
1817 // return semantics.
1818 if (isCalleeStructRet || isCallerStructRet)
1821 // Note that currently ARM64 "C" calling convention and "Fast" calling
1822 // convention are compatible. If/when that ever changes, we'll need to
1823 // add checks here to make sure any interactions are OK.
1825 // If the callee takes no arguments then go on to check the results of the
1827 if (!Outs.empty()) {
1828 // Check if stack adjustment is needed. For now, do not do this if any
1829 // argument is passed on the stack.
1830 SmallVector<CCValAssign, 16> ArgLocs;
1831 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1832 getTargetMachine(), ArgLocs, *DAG.getContext());
1833 CCAssignFn *AssignFn = CCAssignFnForCall(CalleeCC, /*IsVarArg=*/false);
1834 CCInfo.AnalyzeCallOperands(Outs, AssignFn);
1835 if (CCInfo.getNextStackOffset()) {
1836 // Check if the arguments are already laid out in the right way as
1837 // the caller's fixed stack objects.
1838 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
1839 ++i, ++realArgIdx) {
1840 CCValAssign &VA = ArgLocs[i];
1841 if (VA.getLocInfo() == CCValAssign::Indirect)
1843 if (VA.needsCustom()) {
1844 // Just don't handle anything that needs custom adjustments for now.
1845 // If need be, we can revisit later, but we shouldn't ever end up
1848 } else if (!VA.isRegLoc()) {
1849 // Likewise, don't try to handle stack based arguments for the
1859 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
1860 /// and add input and output parameter nodes.
1861 SDValue ARM64TargetLowering::LowerCall(CallLoweringInfo &CLI,
1862 SmallVectorImpl<SDValue> &InVals) const {
1863 SelectionDAG &DAG = CLI.DAG;
1865 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
1866 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
1867 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
1868 SDValue Chain = CLI.Chain;
1869 SDValue Callee = CLI.Callee;
1870 bool &IsTailCall = CLI.IsTailCall;
1871 CallingConv::ID CallConv = CLI.CallConv;
1872 bool IsVarArg = CLI.IsVarArg;
1874 MachineFunction &MF = DAG.getMachineFunction();
1875 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1876 bool IsThisReturn = false;
1878 // If tail calls are explicitly disabled, make sure not to use them.
1879 if (!EnableARM64TailCalls)
1883 // Check if it's really possible to do a tail call.
1884 IsTailCall = isEligibleForTailCallOptimization(
1885 Callee, CallConv, IsVarArg, IsStructRet,
1886 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
1887 // We don't support GuaranteedTailCallOpt, only automatically
1888 // detected sibcalls.
1889 // FIXME: Re-evaluate. Is this true? Should it be true?
1894 // Analyze operands of the call, assigning locations to each operand.
1895 SmallVector<CCValAssign, 16> ArgLocs;
1896 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
1897 getTargetMachine(), ArgLocs, *DAG.getContext());
1900 // Handle fixed and variable vector arguments differently.
1901 // Variable vector arguments always go into memory.
1902 unsigned NumArgs = Outs.size();
1904 for (unsigned i = 0; i != NumArgs; ++i) {
1905 MVT ArgVT = Outs[i].VT;
1906 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
1907 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
1908 /*IsVarArg=*/ !Outs[i].IsFixed);
1909 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
1910 assert(!Res && "Call operand has unhandled type");
1914 // At this point, Outs[].VT may already be promoted to i32. To correctly
1915 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
1916 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
1917 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
1918 // we use a special version of AnalyzeCallOperands to pass in ValVT and
1920 unsigned NumArgs = Outs.size();
1921 for (unsigned i = 0; i != NumArgs; ++i) {
1922 MVT ValVT = Outs[i].VT;
1923 // Get type of the original argument.
1924 EVT ActualVT = getValueType(CLI.Args[Outs[i].OrigArgIndex].Ty,
1925 /*AllowUnknown*/ true);
1926 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
1927 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
1928 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
1930 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
1932 else if (ActualMVT == MVT::i16)
1935 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
1936 bool Res = AssignFn(i, ValVT, LocVT, CCValAssign::Full, ArgFlags, CCInfo);
1937 assert(!Res && "Call operand has unhandled type");
1942 // Get a count of how many bytes are to be pushed on the stack.
1943 unsigned NumBytes = CCInfo.getNextStackOffset();
1945 // Adjust the stack pointer for the new arguments...
1946 // These operations are automatically eliminated by the prolog/epilog pass
1949 DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true), DL);
1951 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, ARM64::SP, getPointerTy());
1953 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1954 SmallVector<SDValue, 8> MemOpChains;
1956 // Walk the register/memloc assignments, inserting copies/loads.
1957 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
1958 ++i, ++realArgIdx) {
1959 CCValAssign &VA = ArgLocs[i];
1960 SDValue Arg = OutVals[realArgIdx];
1961 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1963 // Promote the value if needed.
1964 switch (VA.getLocInfo()) {
1966 llvm_unreachable("Unknown loc info!");
1967 case CCValAssign::Full:
1969 case CCValAssign::SExt:
1970 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
1972 case CCValAssign::ZExt:
1973 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
1975 case CCValAssign::AExt:
1976 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
1978 case CCValAssign::BCvt:
1979 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
1981 case CCValAssign::FPExt:
1982 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
1986 if (VA.isRegLoc()) {
1987 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
1988 assert(VA.getLocVT() == MVT::i64 &&
1989 "unexpected calling convention register assignment");
1990 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
1991 "unexpected use of 'returned'");
1992 IsThisReturn = true;
1994 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1996 assert(VA.isMemLoc());
1997 // There's no reason we can't support stack args w/ tailcall, but
1998 // we currently don't, so assert if we see one.
1999 assert(!IsTailCall && "stack argument with tail call!?");
2000 unsigned LocMemOffset = VA.getLocMemOffset();
2001 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2002 PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2004 if (Outs[i].Flags.isByVal()) {
2006 DAG.getConstant(Outs[i].Flags.getByValSize(), MVT::i64);
2007 SDValue Cpy = DAG.getMemcpy(
2008 Chain, DL, PtrOff, Arg, SizeNode, Outs[i].Flags.getByValAlign(),
2009 /*isVolatile = */ false,
2010 /*alwaysInline = */ false,
2011 MachinePointerInfo::getStack(LocMemOffset), MachinePointerInfo());
2013 MemOpChains.push_back(Cpy);
2015 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
2016 // promoted to a legal register type i32, we should truncate Arg back to
2018 if (Arg.getValueType().isSimple() &&
2019 Arg.getValueType().getSimpleVT() == MVT::i32 &&
2020 (VA.getLocVT() == MVT::i1 || VA.getLocVT() == MVT::i8 ||
2021 VA.getLocVT() == MVT::i16))
2022 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getLocVT(), Arg);
2024 SDValue Store = DAG.getStore(Chain, DL, Arg, PtrOff,
2025 MachinePointerInfo::getStack(LocMemOffset),
2027 MemOpChains.push_back(Store);
2032 if (!MemOpChains.empty())
2033 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOpChains[0],
2034 MemOpChains.size());
2036 // Build a sequence of copy-to-reg nodes chained together with token chain
2037 // and flag operands which copy the outgoing args into the appropriate regs.
2039 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2040 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[i].first,
2041 RegsToPass[i].second, InFlag);
2042 InFlag = Chain.getValue(1);
2045 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2046 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2047 // node so that legalize doesn't hack it.
2048 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
2049 Subtarget->isTargetMachO()) {
2050 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2051 const GlobalValue *GV = G->getGlobal();
2052 bool InternalLinkage = GV->hasInternalLinkage();
2053 if (InternalLinkage)
2054 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2056 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0,
2058 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2060 } else if (ExternalSymbolSDNode *S =
2061 dyn_cast<ExternalSymbolSDNode>(Callee)) {
2062 const char *Sym = S->getSymbol();
2064 DAG.getTargetExternalSymbol(Sym, getPointerTy(), ARM64II::MO_GOT);
2065 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2067 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2068 const GlobalValue *GV = G->getGlobal();
2069 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2070 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2071 const char *Sym = S->getSymbol();
2072 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), 0);
2075 std::vector<SDValue> Ops;
2076 Ops.push_back(Chain);
2077 Ops.push_back(Callee);
2079 // Add argument registers to the end of the list so that they are known live
2081 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2082 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2083 RegsToPass[i].second.getValueType()));
2085 // Add a register mask operand representing the call-preserved registers.
2086 const uint32_t *Mask;
2087 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2088 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2090 // For 'this' returns, use the X0-preserving mask if applicable
2091 Mask = ARI->getThisReturnPreservedMask(CallConv);
2093 IsThisReturn = false;
2094 Mask = ARI->getCallPreservedMask(CallConv);
2097 Mask = ARI->getCallPreservedMask(CallConv);
2099 assert(Mask && "Missing call preserved mask for calling convention");
2100 Ops.push_back(DAG.getRegisterMask(Mask));
2102 if (InFlag.getNode())
2103 Ops.push_back(InFlag);
2105 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2107 // If we're doing a tall call, use a TC_RETURN here rather than an
2108 // actual call instruction.
2110 return DAG.getNode(ARM64ISD::TC_RETURN, DL, NodeTys, &Ops[0], Ops.size());
2112 // Returns a chain and a flag for retval copy to use.
2113 Chain = DAG.getNode(ARM64ISD::CALL, DL, NodeTys, &Ops[0], Ops.size());
2114 InFlag = Chain.getValue(1);
2116 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2117 DAG.getIntPtrConstant(0, true), InFlag, DL);
2119 InFlag = Chain.getValue(1);
2121 // Handle result values, copying them out of physregs into vregs that we
2123 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2124 InVals, IsThisReturn,
2125 IsThisReturn ? OutVals[0] : SDValue());
2128 bool ARM64TargetLowering::CanLowerReturn(
2129 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2130 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2131 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2132 : RetCC_ARM64_AAPCS;
2133 SmallVector<CCValAssign, 16> RVLocs;
2134 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2135 return CCInfo.CheckReturn(Outs, RetCC);
2139 ARM64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2141 const SmallVectorImpl<ISD::OutputArg> &Outs,
2142 const SmallVectorImpl<SDValue> &OutVals,
2143 SDLoc DL, SelectionDAG &DAG) const {
2144 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2145 : RetCC_ARM64_AAPCS;
2146 SmallVector<CCValAssign, 16> RVLocs;
2147 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2148 getTargetMachine(), RVLocs, *DAG.getContext());
2149 CCInfo.AnalyzeReturn(Outs, RetCC);
2151 // Copy the result values into the output registers.
2153 SmallVector<SDValue, 4> RetOps(1, Chain);
2154 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
2155 ++i, ++realRVLocIdx) {
2156 CCValAssign &VA = RVLocs[i];
2157 assert(VA.isRegLoc() && "Can only return in registers!");
2158 SDValue Arg = OutVals[realRVLocIdx];
2160 switch (VA.getLocInfo()) {
2162 llvm_unreachable("Unknown loc info!");
2163 case CCValAssign::Full:
2165 case CCValAssign::BCvt:
2166 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2170 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
2171 Flag = Chain.getValue(1);
2172 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2175 RetOps[0] = Chain; // Update chain.
2177 // Add the flag if we have it.
2179 RetOps.push_back(Flag);
2181 return DAG.getNode(ARM64ISD::RET_FLAG, DL, MVT::Other, &RetOps[0],
2185 //===----------------------------------------------------------------------===//
2186 // Other Lowering Code
2187 //===----------------------------------------------------------------------===//
2189 SDValue ARM64TargetLowering::LowerGlobalAddress(SDValue Op,
2190 SelectionDAG &DAG) const {
2191 EVT PtrVT = getPointerTy();
2193 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2194 unsigned char OpFlags =
2195 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
2197 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
2198 "unexpected offset in global node");
2200 // This also catched the large code model case for Darwin.
2201 if ((OpFlags & ARM64II::MO_GOT) != 0) {
2202 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
2203 // FIXME: Once remat is capable of dealing with instructions with register
2204 // operands, expand this into two nodes instead of using a wrapper node.
2205 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
2208 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2209 const unsigned char MO_NC = ARM64II::MO_NC;
2211 ARM64ISD::WrapperLarge, DL, PtrVT,
2212 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G3),
2213 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
2214 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
2215 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
2217 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
2218 // the only correct model on Darwin.
2219 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2220 OpFlags | ARM64II::MO_PAGE);
2221 unsigned char LoFlags = OpFlags | ARM64II::MO_PAGEOFF | ARM64II::MO_NC;
2222 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
2224 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
2225 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
2229 /// \brief Convert a TLS address reference into the correct sequence of loads
2230 /// and calls to compute the variable's address (for Darwin, currently) and
2231 /// return an SDValue containing the final node.
2233 /// Darwin only has one TLS scheme which must be capable of dealing with the
2234 /// fully general situation, in the worst case. This means:
2235 /// + "extern __thread" declaration.
2236 /// + Defined in a possibly unknown dynamic library.
2238 /// The general system is that each __thread variable has a [3 x i64] descriptor
2239 /// which contains information used by the runtime to calculate the address. The
2240 /// only part of this the compiler needs to know about is the first xword, which
2241 /// contains a function pointer that must be called with the address of the
2242 /// entire descriptor in "x0".
2244 /// Since this descriptor may be in a different unit, in general even the
2245 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
2247 /// adrp x0, _var@TLVPPAGE
2248 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
2249 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
2250 /// ; the function pointer
2251 /// blr x1 ; Uses descriptor address in x0
2252 /// ; Address of _var is now in x0.
2254 /// If the address of _var's descriptor *is* known to the linker, then it can
2255 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
2256 /// a slight efficiency gain.
2258 ARM64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
2259 SelectionDAG &DAG) const {
2260 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
2263 MVT PtrVT = getPointerTy();
2264 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2267 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2268 SDValue DescAddr = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TLVPAddr);
2270 // The first entry in the descriptor is a function pointer that we must call
2271 // to obtain the address of the variable.
2272 SDValue Chain = DAG.getEntryNode();
2273 SDValue FuncTLVGet =
2274 DAG.getLoad(MVT::i64, DL, Chain, DescAddr, MachinePointerInfo::getGOT(),
2275 false, true, true, 8);
2276 Chain = FuncTLVGet.getValue(1);
2278 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2279 MFI->setAdjustsStack(true);
2281 // TLS calls preserve all registers except those that absolutely must be
2282 // trashed: X0 (it takes an argument), LR (it's a call) and CPSR (let's not be
2284 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2285 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2286 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2288 // Finally, we can make the call. This is just a degenerate version of a
2289 // normal ARM64 call node: x0 takes the address of the descriptor, and returns
2290 // the address of the variable in this thread.
2291 Chain = DAG.getCopyToReg(Chain, DL, ARM64::X0, DescAddr, SDValue());
2292 Chain = DAG.getNode(ARM64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
2293 Chain, FuncTLVGet, DAG.getRegister(ARM64::X0, MVT::i64),
2294 DAG.getRegisterMask(Mask), Chain.getValue(1));
2295 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Chain.getValue(1));
2298 /// When accessing thread-local variables under either the general-dynamic or
2299 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
2300 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
2301 /// is a function pointer to carry out the resolution. This function takes the
2302 /// address of the descriptor in X0 and returns the TPIDR_EL0 offset in X0. All
2303 /// other registers (except LR, CPSR) are preserved.
2305 /// Thus, the ideal call sequence on AArch64 is:
2307 /// adrp x0, :tlsdesc:thread_var
2308 /// ldr x8, [x0, :tlsdesc_lo12:thread_var]
2309 /// add x0, x0, :tlsdesc_lo12:thread_var
2310 /// .tlsdesccall thread_var
2312 /// (TPIDR_EL0 offset now in x0).
2314 /// The ".tlsdesccall" directive instructs the assembler to insert a particular
2315 /// relocation to help the linker relax this sequence if it turns out to be too
2318 /// FIXME: we currently produce an extra, duplicated, ADRP instruction, but this
2320 SDValue ARM64TargetLowering::LowerELFTLSDescCall(SDValue SymAddr,
2321 SDValue DescAddr, SDLoc DL,
2322 SelectionDAG &DAG) const {
2323 EVT PtrVT = getPointerTy();
2325 // The function we need to call is simply the first entry in the GOT for this
2326 // descriptor, load it in preparation.
2327 SDValue Func = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, SymAddr);
2329 // TLS calls preserve all registers except those that absolutely must be
2330 // trashed: X0 (it takes an argument), LR (it's a call) and CPSR (let's not be
2332 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2333 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2334 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2336 // The function takes only one argument: the address of the descriptor itself
2338 SDValue Glue, Chain;
2339 Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, ARM64::X0, DescAddr, Glue);
2340 Glue = Chain.getValue(1);
2342 // We're now ready to populate the argument list, as with a normal call:
2343 SmallVector<SDValue, 6> Ops;
2344 Ops.push_back(Chain);
2345 Ops.push_back(Func);
2346 Ops.push_back(SymAddr);
2347 Ops.push_back(DAG.getRegister(ARM64::X0, PtrVT));
2348 Ops.push_back(DAG.getRegisterMask(Mask));
2349 Ops.push_back(Glue);
2351 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2352 Chain = DAG.getNode(ARM64ISD::TLSDESC_CALL, DL, NodeTys, &Ops[0], Ops.size());
2353 Glue = Chain.getValue(1);
2355 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Glue);
2358 SDValue ARM64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
2359 SelectionDAG &DAG) const {
2360 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
2361 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
2362 "ELF TLS only supported in small memory model");
2363 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2365 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
2368 EVT PtrVT = getPointerTy();
2370 const GlobalValue *GV = GA->getGlobal();
2372 SDValue ThreadBase = DAG.getNode(ARM64ISD::THREAD_POINTER, DL, PtrVT);
2374 if (Model == TLSModel::LocalExec) {
2375 SDValue HiVar = DAG.getTargetGlobalAddress(
2376 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2377 SDValue LoVar = DAG.getTargetGlobalAddress(
2378 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2380 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2381 DAG.getTargetConstant(16, MVT::i32)),
2383 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, TPOff, LoVar,
2384 DAG.getTargetConstant(0, MVT::i32)),
2386 } else if (Model == TLSModel::InitialExec) {
2387 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2388 TPOff = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TPOff);
2389 } else if (Model == TLSModel::LocalDynamic) {
2390 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
2391 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
2392 // the beginning of the module's TLS region, followed by a DTPREL offset
2395 // These accesses will need deduplicating if there's more than one.
2396 ARM64FunctionInfo *MFI =
2397 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
2398 MFI->incNumLocalDynamicTLSAccesses();
2400 // Accesses used in this sequence go via the TLS descriptor which lives in
2401 // the GOT. Prepare an address we can use to handle this.
2402 SDValue HiDesc = DAG.getTargetExternalSymbol(
2403 "_TLS_MODULE_BASE_", PtrVT, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2404 SDValue LoDesc = DAG.getTargetExternalSymbol(
2405 "_TLS_MODULE_BASE_", PtrVT,
2406 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2408 // First argument to the descriptor call is the address of the descriptor
2410 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2411 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2413 // The call needs a relocation too for linker relaxation. It doesn't make
2414 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2416 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2419 // Now we can calculate the offset from TPIDR_EL0 to this module's
2420 // thread-local area.
2421 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2423 // Now use :dtprel_whatever: operations to calculate this variable's offset
2424 // in its thread-storage area.
2425 SDValue HiVar = DAG.getTargetGlobalAddress(
2426 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2427 SDValue LoVar = DAG.getTargetGlobalAddress(
2428 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2431 SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2432 DAG.getTargetConstant(16, MVT::i32)),
2434 DTPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, DTPOff, LoVar,
2435 DAG.getTargetConstant(0, MVT::i32)),
2438 TPOff = DAG.getNode(ISD::ADD, DL, PtrVT, TPOff, DTPOff);
2439 } else if (Model == TLSModel::GeneralDynamic) {
2440 // Accesses used in this sequence go via the TLS descriptor which lives in
2441 // the GOT. Prepare an address we can use to handle this.
2442 SDValue HiDesc = DAG.getTargetGlobalAddress(
2443 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2444 SDValue LoDesc = DAG.getTargetGlobalAddress(
2446 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2448 // First argument to the descriptor call is the address of the descriptor
2450 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2451 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2453 // The call needs a relocation too for linker relaxation. It doesn't make
2454 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2457 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2459 // Finally we can make a call to calculate the offset from tpidr_el0.
2460 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2462 llvm_unreachable("Unsupported ELF TLS access model");
2464 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
2467 SDValue ARM64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
2468 SelectionDAG &DAG) const {
2469 if (Subtarget->isTargetDarwin())
2470 return LowerDarwinGlobalTLSAddress(Op, DAG);
2471 else if (Subtarget->isTargetELF())
2472 return LowerELFGlobalTLSAddress(Op, DAG);
2474 llvm_unreachable("Unexpected platform trying to use TLS");
2476 SDValue ARM64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2477 SDValue Chain = Op.getOperand(0);
2478 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2479 SDValue LHS = Op.getOperand(2);
2480 SDValue RHS = Op.getOperand(3);
2481 SDValue Dest = Op.getOperand(4);
2484 // Handle f128 first, since lowering it will result in comparing the return
2485 // value of a libcall against zero, which is just what the rest of LowerBR_CC
2486 // is expecting to deal with.
2487 if (LHS.getValueType() == MVT::f128) {
2488 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2490 // If softenSetCCOperands returned a scalar, we need to compare the result
2491 // against zero to select between true and false values.
2492 if (RHS.getNode() == 0) {
2493 RHS = DAG.getConstant(0, LHS.getValueType());
2498 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
2500 unsigned Opc = LHS.getOpcode();
2501 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
2502 cast<ConstantSDNode>(RHS)->isOne() &&
2503 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
2504 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
2505 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
2506 "Unexpected condition code.");
2507 // Only lower legal XALUO ops.
2508 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
2511 // The actual operation with overflow check.
2512 ARM64CC::CondCode OFCC;
2513 SDValue Value, Overflow;
2514 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, LHS.getValue(0), DAG);
2516 if (CC == ISD::SETNE)
2517 OFCC = getInvertedCondCode(OFCC);
2518 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
2520 return DAG.getNode(ARM64ISD::BRCOND, SDLoc(LHS), MVT::Other, Chain, Dest,
2524 if (LHS.getValueType().isInteger()) {
2525 assert((LHS.getValueType() == RHS.getValueType()) &&
2526 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
2528 // If the RHS of the comparison is zero, we can potentially fold this
2529 // to a specialized branch.
2530 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
2531 if (RHSC && RHSC->getZExtValue() == 0) {
2532 if (CC == ISD::SETEQ) {
2533 // See if we can use a TBZ to fold in an AND as well.
2534 // TBZ has a smaller branch displacement than CBZ. If the offset is
2535 // out of bounds, a late MI-layer pass rewrites branches.
2536 // 403.gcc is an example that hits this case.
2537 if (LHS.getOpcode() == ISD::AND &&
2538 isa<ConstantSDNode>(LHS.getOperand(1)) &&
2539 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
2540 SDValue Test = LHS.getOperand(0);
2541 uint64_t Mask = LHS.getConstantOperandVal(1);
2543 // TBZ only operates on i64's, but the ext should be free.
2544 if (Test.getValueType() == MVT::i32)
2545 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
2547 return DAG.getNode(ARM64ISD::TBZ, dl, MVT::Other, Chain, Test,
2548 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
2551 return DAG.getNode(ARM64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
2552 } else if (CC == ISD::SETNE) {
2553 // See if we can use a TBZ to fold in an AND as well.
2554 // TBZ has a smaller branch displacement than CBZ. If the offset is
2555 // out of bounds, a late MI-layer pass rewrites branches.
2556 // 403.gcc is an example that hits this case.
2557 if (LHS.getOpcode() == ISD::AND &&
2558 isa<ConstantSDNode>(LHS.getOperand(1)) &&
2559 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
2560 SDValue Test = LHS.getOperand(0);
2561 uint64_t Mask = LHS.getConstantOperandVal(1);
2563 // TBNZ only operates on i64's, but the ext should be free.
2564 if (Test.getValueType() == MVT::i32)
2565 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
2567 return DAG.getNode(ARM64ISD::TBNZ, dl, MVT::Other, Chain, Test,
2568 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
2571 return DAG.getNode(ARM64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
2576 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
2577 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
2581 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2583 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
2584 // clean. Some of them require two branches to implement.
2585 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
2586 ARM64CC::CondCode CC1, CC2;
2587 changeFPCCToARM64CC(CC, CC1, CC2);
2588 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
2590 DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
2591 if (CC2 != ARM64CC::AL) {
2592 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
2593 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
2600 SDValue ARM64TargetLowering::LowerFCOPYSIGN(SDValue Op,
2601 SelectionDAG &DAG) const {
2602 EVT VT = Op.getValueType();
2605 SDValue In1 = Op.getOperand(0);
2606 SDValue In2 = Op.getOperand(1);
2607 EVT SrcVT = In2.getValueType();
2609 if (SrcVT == MVT::f32 && VT == MVT::f64)
2610 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
2611 else if (SrcVT == MVT::f64 && VT == MVT::f32)
2612 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0));
2614 // FIXME: Src type is different, bail out for now. Can VT really be a
2621 SDValue EltMask, VecVal1, VecVal2;
2622 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
2625 EltMask = DAG.getConstant(0x80000000ULL, EltVT);
2627 if (!VT.isVector()) {
2628 VecVal1 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
2629 DAG.getUNDEF(VecVT), In1);
2630 VecVal2 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
2631 DAG.getUNDEF(VecVT), In2);
2633 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
2634 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
2636 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
2640 // We want to materialize a mask with the the high bit set, but the AdvSIMD
2641 // immediate moves cannot materialize that in a single instruction for
2642 // 64-bit elements. Instead, materialize zero and then negate it.
2643 EltMask = DAG.getConstant(0, EltVT);
2645 if (!VT.isVector()) {
2646 VecVal1 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
2647 DAG.getUNDEF(VecVT), In1);
2648 VecVal2 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
2649 DAG.getUNDEF(VecVT), In2);
2651 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
2652 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
2655 llvm_unreachable("Invalid type for copysign!");
2658 std::vector<SDValue> BuildVectorOps;
2659 for (unsigned i = 0; i < VecVT.getVectorNumElements(); ++i)
2660 BuildVectorOps.push_back(EltMask);
2662 SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT,
2663 &BuildVectorOps[0], BuildVectorOps.size());
2665 // If we couldn't materialize the mask above, then the mask vector will be
2666 // the zero vector, and we need to negate it here.
2667 if (VT == MVT::f64 || VT == MVT::v2f64) {
2668 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
2669 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
2670 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
2674 DAG.getNode(ARM64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
2677 return DAG.getTargetExtractSubreg(ARM64::ssub, DL, VT, Sel);
2678 else if (VT == MVT::f64)
2679 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, VT, Sel);
2681 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
2684 SDValue ARM64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
2685 if (DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
2686 AttributeSet::FunctionIndex, Attribute::NoImplicitFloat))
2689 // While there is no integer popcount instruction, it can
2690 // be more efficiently lowered to the following sequence that uses
2691 // AdvSIMD registers/instructions as long as the copies to/from
2692 // the AdvSIMD registers are cheap.
2693 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
2694 // CNT V0.8B, V0.8B // 8xbyte pop-counts
2695 // ADDV B0, V0.8B // sum 8xbyte pop-counts
2696 // UMOV X0, V0.B[0] // copy byte result back to integer reg
2697 SDValue Val = Op.getOperand(0);
2699 EVT VT = Op.getValueType();
2700 SDValue ZeroVec = DAG.getUNDEF(MVT::v8i8);
2703 if (VT == MVT::i32) {
2704 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Val);
2706 DAG.getTargetInsertSubreg(ARM64::ssub, DL, MVT::v8i8, ZeroVec, VecVal);
2708 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
2711 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, VecVal);
2712 SDValue UaddLV = DAG.getNode(
2713 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
2714 DAG.getConstant(Intrinsic::arm64_neon_uaddlv, MVT::i32), CtPop);
2717 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
2721 SDValue ARM64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
2723 if (Op.getValueType().isVector())
2724 return LowerVSETCC(Op, DAG);
2726 SDValue LHS = Op.getOperand(0);
2727 SDValue RHS = Op.getOperand(1);
2728 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
2731 // We chose ZeroOrOneBooleanContents, so use zero and one.
2732 EVT VT = Op.getValueType();
2733 SDValue TVal = DAG.getConstant(1, VT);
2734 SDValue FVal = DAG.getConstant(0, VT);
2736 // Handle f128 first, since one possible outcome is a normal integer
2737 // comparison which gets picked up by the next if statement.
2738 if (LHS.getValueType() == MVT::f128) {
2739 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2741 // If softenSetCCOperands returned a scalar, use it.
2742 if (RHS.getNode() == 0) {
2743 assert(LHS.getValueType() == Op.getValueType() &&
2744 "Unexpected setcc expansion!");
2749 if (LHS.getValueType().isInteger()) {
2752 getARM64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
2754 // Note that we inverted the condition above, so we reverse the order of
2755 // the true and false operands here. This will allow the setcc to be
2756 // matched to a single CSINC instruction.
2757 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
2760 // Now we know we're dealing with FP values.
2761 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2763 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
2764 // and do the comparison.
2765 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
2767 ARM64CC::CondCode CC1, CC2;
2768 changeFPCCToARM64CC(CC, CC1, CC2);
2769 if (CC2 == ARM64CC::AL) {
2770 changeFPCCToARM64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
2771 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
2773 // Note that we inverted the condition above, so we reverse the order of
2774 // the true and false operands here. This will allow the setcc to be
2775 // matched to a single CSINC instruction.
2776 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
2778 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
2779 // clean. Some of them require two CSELs to implement. As is in this case,
2780 // we emit the first CSEL and then emit a second using the output of the
2781 // first as the RHS. We're effectively OR'ing the two CC's together.
2783 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
2784 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
2785 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
2787 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
2788 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
2792 /// A SELECT_CC operation is really some kind of max or min if both values being
2793 /// compared are, in some sense, equal to the results in either case. However,
2794 /// it is permissible to compare f32 values and produce directly extended f64
2797 /// Extending the comparison operands would also be allowed, but is less likely
2798 /// to happen in practice since their use is right here. Note that truncate
2799 /// operations would *not* be semantically equivalent.
2800 static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) {
2804 ConstantFPSDNode *CCmp = dyn_cast<ConstantFPSDNode>(Cmp);
2805 ConstantFPSDNode *CResult = dyn_cast<ConstantFPSDNode>(Result);
2806 if (CCmp && CResult && Cmp.getValueType() == MVT::f32 &&
2807 Result.getValueType() == MVT::f64) {
2809 APFloat CmpVal = CCmp->getValueAPF();
2810 CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy);
2811 return CResult->getValueAPF().bitwiseIsEqual(CmpVal);
2814 return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp;
2817 SDValue ARM64TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2818 SDValue CC = Op->getOperand(0);
2819 SDValue TVal = Op->getOperand(1);
2820 SDValue FVal = Op->getOperand(2);
2823 unsigned Opc = CC.getOpcode();
2824 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
2826 if (CC.getResNo() == 1 &&
2827 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
2828 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
2829 // Only lower legal XALUO ops.
2830 if (!DAG.getTargetLoweringInfo().isTypeLegal(CC->getValueType(0)))
2833 ARM64CC::CondCode OFCC;
2834 SDValue Value, Overflow;
2835 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, CC.getValue(0), DAG);
2836 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
2838 return DAG.getNode(ARM64ISD::CSEL, DL, Op.getValueType(), TVal, FVal, CCVal,
2842 if (CC.getOpcode() == ISD::SETCC)
2843 return DAG.getSelectCC(DL, CC.getOperand(0), CC.getOperand(1), TVal, FVal,
2844 cast<CondCodeSDNode>(CC.getOperand(2))->get());
2846 return DAG.getSelectCC(DL, CC, DAG.getConstant(0, CC.getValueType()), TVal,
2850 SDValue ARM64TargetLowering::LowerSELECT_CC(SDValue Op,
2851 SelectionDAG &DAG) const {
2852 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2853 SDValue LHS = Op.getOperand(0);
2854 SDValue RHS = Op.getOperand(1);
2855 SDValue TVal = Op.getOperand(2);
2856 SDValue FVal = Op.getOperand(3);
2859 // Handle f128 first, because it will result in a comparison of some RTLIB
2860 // call result against zero.
2861 if (LHS.getValueType() == MVT::f128) {
2862 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2864 // If softenSetCCOperands returned a scalar, we need to compare the result
2865 // against zero to select between true and false values.
2866 if (RHS.getNode() == 0) {
2867 RHS = DAG.getConstant(0, LHS.getValueType());
2872 // Handle integers first.
2873 if (LHS.getValueType().isInteger()) {
2874 assert((LHS.getValueType() == RHS.getValueType()) &&
2875 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
2877 unsigned Opcode = ARM64ISD::CSEL;
2879 // If both the TVal and the FVal are constants, see if we can swap them in
2880 // order to for a CSINV or CSINC out of them.
2881 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
2882 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
2884 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
2885 std::swap(TVal, FVal);
2886 std::swap(CTVal, CFVal);
2887 CC = ISD::getSetCCInverse(CC, true);
2888 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
2889 std::swap(TVal, FVal);
2890 std::swap(CTVal, CFVal);
2891 CC = ISD::getSetCCInverse(CC, true);
2892 } else if (TVal.getOpcode() == ISD::XOR) {
2893 // If TVal is a NOT we want to swap TVal and FVal so that we can match
2894 // with a CSINV rather than a CSEL.
2895 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
2897 if (CVal && CVal->isAllOnesValue()) {
2898 std::swap(TVal, FVal);
2899 std::swap(CTVal, CFVal);
2900 CC = ISD::getSetCCInverse(CC, true);
2902 } else if (TVal.getOpcode() == ISD::SUB) {
2903 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
2904 // that we can match with a CSNEG rather than a CSEL.
2905 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
2907 if (CVal && CVal->isNullValue()) {
2908 std::swap(TVal, FVal);
2909 std::swap(CTVal, CFVal);
2910 CC = ISD::getSetCCInverse(CC, true);
2912 } else if (CTVal && CFVal) {
2913 const int64_t TrueVal = CTVal->getSExtValue();
2914 const int64_t FalseVal = CFVal->getSExtValue();
2917 // If both TVal and FVal are constants, see if FVal is the
2918 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
2919 // instead of a CSEL in that case.
2920 if (TrueVal == ~FalseVal) {
2921 Opcode = ARM64ISD::CSINV;
2922 } else if (TrueVal == -FalseVal) {
2923 Opcode = ARM64ISD::CSNEG;
2924 } else if (TVal.getValueType() == MVT::i32) {
2925 // If our operands are only 32-bit wide, make sure we use 32-bit
2926 // arithmetic for the check whether we can use CSINC. This ensures that
2927 // the addition in the check will wrap around properly in case there is
2928 // an overflow (which would not be the case if we do the check with
2929 // 64-bit arithmetic).
2930 const uint32_t TrueVal32 = CTVal->getZExtValue();
2931 const uint32_t FalseVal32 = CFVal->getZExtValue();
2933 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
2934 Opcode = ARM64ISD::CSINC;
2936 if (TrueVal32 > FalseVal32) {
2940 // 64-bit check whether we can use CSINC.
2941 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
2942 Opcode = ARM64ISD::CSINC;
2944 if (TrueVal > FalseVal) {
2949 // Swap TVal and FVal if necessary.
2951 std::swap(TVal, FVal);
2952 std::swap(CTVal, CFVal);
2953 CC = ISD::getSetCCInverse(CC, true);
2956 if (Opcode != ARM64ISD::CSEL) {
2957 // Drop FVal since we can get its value by simply inverting/negating
2964 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
2966 EVT VT = Op.getValueType();
2967 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
2970 // Now we know we're dealing with FP values.
2971 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2972 assert(LHS.getValueType() == RHS.getValueType());
2973 EVT VT = Op.getValueType();
2975 // Try to match this select into a max/min operation, which have dedicated
2976 // opcode in the instruction set.
2977 // NOTE: This is not correct in the presence of NaNs, so we only enable this
2979 if (getTargetMachine().Options.NoNaNsFPMath) {
2980 if (selectCCOpsAreFMaxCompatible(LHS, FVal) &&
2981 selectCCOpsAreFMaxCompatible(RHS, TVal)) {
2982 CC = ISD::getSetCCSwappedOperands(CC);
2983 std::swap(TVal, FVal);
2986 if (selectCCOpsAreFMaxCompatible(LHS, TVal) &&
2987 selectCCOpsAreFMaxCompatible(RHS, FVal)) {
2997 return DAG.getNode(ARM64ISD::FMAX, dl, VT, TVal, FVal);
3005 return DAG.getNode(ARM64ISD::FMIN, dl, VT, TVal, FVal);
3011 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3012 // and do the comparison.
3013 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3015 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3016 // clean. Some of them require two CSELs to implement.
3017 ARM64CC::CondCode CC1, CC2;
3018 changeFPCCToARM64CC(CC, CC1, CC2);
3019 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3020 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3022 // If we need a second CSEL, emit it, using the output of the first as the
3023 // RHS. We're effectively OR'ing the two CC's together.
3024 if (CC2 != ARM64CC::AL) {
3025 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3026 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3029 // Otherwise, return the output of the first CSEL.
3033 SDValue ARM64TargetLowering::LowerJumpTable(SDValue Op,
3034 SelectionDAG &DAG) const {
3035 // Jump table entries as PC relative offsets. No additional tweaking
3036 // is necessary here. Just get the address of the jump table.
3037 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3038 EVT PtrVT = getPointerTy();
3041 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3042 !Subtarget->isTargetMachO()) {
3043 const unsigned char MO_NC = ARM64II::MO_NC;
3045 ARM64ISD::WrapperLarge, DL, PtrVT,
3046 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G3),
3047 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G2 | MO_NC),
3048 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G1 | MO_NC),
3049 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G0 | MO_NC));
3052 SDValue Hi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_PAGE);
3053 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3054 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3055 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3056 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3059 SDValue ARM64TargetLowering::LowerConstantPool(SDValue Op,
3060 SelectionDAG &DAG) const {
3061 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3062 EVT PtrVT = getPointerTy();
3065 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3066 // Use the GOT for the large code model on iOS.
3067 if (Subtarget->isTargetMachO()) {
3068 SDValue GotAddr = DAG.getTargetConstantPool(
3069 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3071 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
3074 const unsigned char MO_NC = ARM64II::MO_NC;
3076 ARM64ISD::WrapperLarge, DL, PtrVT,
3077 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3078 CP->getOffset(), ARM64II::MO_G3),
3079 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3080 CP->getOffset(), ARM64II::MO_G2 | MO_NC),
3081 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3082 CP->getOffset(), ARM64II::MO_G1 | MO_NC),
3083 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3084 CP->getOffset(), ARM64II::MO_G0 | MO_NC));
3086 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
3087 // ELF, the only valid one on Darwin.
3089 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3090 CP->getOffset(), ARM64II::MO_PAGE);
3091 SDValue Lo = DAG.getTargetConstantPool(
3092 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3093 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3095 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3096 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3100 SDValue ARM64TargetLowering::LowerBlockAddress(SDValue Op,
3101 SelectionDAG &DAG) const {
3102 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
3103 EVT PtrVT = getPointerTy();
3105 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3106 !Subtarget->isTargetMachO()) {
3107 const unsigned char MO_NC = ARM64II::MO_NC;
3109 ARM64ISD::WrapperLarge, DL, PtrVT,
3110 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G3),
3111 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
3112 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
3113 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
3115 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGE);
3116 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGEOFF |
3118 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3119 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3123 SDValue ARM64TargetLowering::LowerDarwin_VASTART(SDValue Op,
3124 SelectionDAG &DAG) const {
3125 ARM64FunctionInfo *FuncInfo =
3126 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
3130 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3131 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3132 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
3133 MachinePointerInfo(SV), false, false, 0);
3136 SDValue ARM64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
3137 SelectionDAG &DAG) const {
3138 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3139 // Standard, section B.3.
3140 MachineFunction &MF = DAG.getMachineFunction();
3141 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
3144 SDValue Chain = Op.getOperand(0);
3145 SDValue VAList = Op.getOperand(1);
3146 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3147 SmallVector<SDValue, 4> MemOps;
3149 // void *__stack at offset 0
3151 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3152 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3153 MachinePointerInfo(SV), false, false, 8));
3155 // void *__gr_top at offset 8
3156 int GPRSize = FuncInfo->getVarArgsGPRSize();
3158 SDValue GRTop, GRTopAddr;
3160 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3161 DAG.getConstant(8, getPointerTy()));
3163 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), getPointerTy());
3164 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
3165 DAG.getConstant(GPRSize, getPointerTy()));
3167 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3168 MachinePointerInfo(SV, 8), false, false, 8));
3171 // void *__vr_top at offset 16
3172 int FPRSize = FuncInfo->getVarArgsFPRSize();
3174 SDValue VRTop, VRTopAddr;
3175 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3176 DAG.getConstant(16, getPointerTy()));
3178 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), getPointerTy());
3179 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
3180 DAG.getConstant(FPRSize, getPointerTy()));
3182 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3183 MachinePointerInfo(SV, 16), false, false, 8));
3186 // int __gr_offs at offset 24
3187 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3188 DAG.getConstant(24, getPointerTy()));
3189 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
3190 GROffsAddr, MachinePointerInfo(SV, 24), false,
3193 // int __vr_offs at offset 28
3194 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3195 DAG.getConstant(28, getPointerTy()));
3196 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
3197 VROffsAddr, MachinePointerInfo(SV, 28), false,
3200 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
3204 SDValue ARM64TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
3205 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
3206 : LowerAAPCS_VASTART(Op, DAG);
3209 SDValue ARM64TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
3210 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
3212 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
3213 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3214 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
3216 return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op), Op.getOperand(1),
3217 Op.getOperand(2), DAG.getConstant(VaListSize, MVT::i32),
3218 8, false, false, MachinePointerInfo(DestSV),
3219 MachinePointerInfo(SrcSV));
3222 SDValue ARM64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
3223 assert(Subtarget->isTargetDarwin() &&
3224 "automatic va_arg instruction only works on Darwin");
3226 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3227 EVT VT = Op.getValueType();
3229 SDValue Chain = Op.getOperand(0);
3230 SDValue Addr = Op.getOperand(1);
3231 unsigned Align = Op.getConstantOperandVal(3);
3233 SDValue VAList = DAG.getLoad(getPointerTy(), DL, Chain, Addr,
3234 MachinePointerInfo(V), false, false, false, 0);
3235 Chain = VAList.getValue(1);
3238 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
3239 VAList = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3240 DAG.getConstant(Align - 1, getPointerTy()));
3241 VAList = DAG.getNode(ISD::AND, DL, getPointerTy(), VAList,
3242 DAG.getConstant(-(int64_t)Align, getPointerTy()));
3245 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
3246 uint64_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
3248 // Scalar integer and FP values smaller than 64 bits are implicitly extended
3249 // up to 64 bits. At the very least, we have to increase the striding of the
3250 // vaargs list to match this, and for FP values we need to introduce
3251 // FP_ROUND nodes as well.
3252 if (VT.isInteger() && !VT.isVector())
3254 bool NeedFPTrunc = false;
3255 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
3260 // Increment the pointer, VAList, to the next vaarg
3261 SDValue VANext = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3262 DAG.getConstant(ArgSize, getPointerTy()));
3263 // Store the incremented VAList to the legalized pointer
3264 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
3267 // Load the actual argument out of the pointer VAList
3269 // Load the value as an f64.
3270 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
3271 MachinePointerInfo(), false, false, false, 0);
3272 // Round the value down to an f32.
3273 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
3274 DAG.getIntPtrConstant(1));
3275 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
3276 // Merge the rounded value with the chain output of the load.
3277 return DAG.getMergeValues(Ops, 2, DL);
3280 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
3284 SDValue ARM64TargetLowering::LowerFRAMEADDR(SDValue Op,
3285 SelectionDAG &DAG) const {
3286 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3287 MFI->setFrameAddressIsTaken(true);
3289 EVT VT = Op.getValueType();
3291 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3292 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, ARM64::FP, VT);
3294 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
3295 MachinePointerInfo(), false, false, false, 0);
3299 SDValue ARM64TargetLowering::LowerRETURNADDR(SDValue Op,
3300 SelectionDAG &DAG) const {
3301 MachineFunction &MF = DAG.getMachineFunction();
3302 MachineFrameInfo *MFI = MF.getFrameInfo();
3303 MFI->setReturnAddressIsTaken(true);
3305 EVT VT = Op.getValueType();
3307 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3309 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3310 SDValue Offset = DAG.getConstant(8, getPointerTy());
3311 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
3312 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
3313 MachinePointerInfo(), false, false, false, 0);
3316 // Return LR, which contains the return address. Mark it an implicit live-in.
3317 unsigned Reg = MF.addLiveIn(ARM64::LR, &ARM64::GPR64RegClass);
3318 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
3321 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3322 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3323 SDValue ARM64TargetLowering::LowerShiftRightParts(SDValue Op,
3324 SelectionDAG &DAG) const {
3325 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3326 EVT VT = Op.getValueType();
3327 unsigned VTBits = VT.getSizeInBits();
3329 SDValue ShOpLo = Op.getOperand(0);
3330 SDValue ShOpHi = Op.getOperand(1);
3331 SDValue ShAmt = Op.getOperand(2);
3333 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3335 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3337 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3338 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3339 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3340 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3341 DAG.getConstant(VTBits, MVT::i64));
3342 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3344 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
3345 ISD::SETGE, dl, DAG);
3346 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3348 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3349 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3351 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3353 // ARM64 shifts larger than the register width are wrapped rather than
3354 // clamped, so we can't just emit "hi >> x".
3355 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3356 SDValue TrueValHi = Opc == ISD::SRA
3357 ? DAG.getNode(Opc, dl, VT, ShOpHi,
3358 DAG.getConstant(VTBits - 1, MVT::i64))
3359 : DAG.getConstant(0, VT);
3361 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
3363 SDValue Ops[2] = { Lo, Hi };
3364 return DAG.getMergeValues(Ops, 2, dl);
3367 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3368 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3369 SDValue ARM64TargetLowering::LowerShiftLeftParts(SDValue Op,
3370 SelectionDAG &DAG) const {
3371 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3372 EVT VT = Op.getValueType();
3373 unsigned VTBits = VT.getSizeInBits();
3375 SDValue ShOpLo = Op.getOperand(0);
3376 SDValue ShOpHi = Op.getOperand(1);
3377 SDValue ShAmt = Op.getOperand(2);
3380 assert(Op.getOpcode() == ISD::SHL_PARTS);
3381 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3382 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3383 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3384 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3385 DAG.getConstant(VTBits, MVT::i64));
3386 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3387 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3389 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3391 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
3392 ISD::SETGE, dl, DAG);
3393 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3394 SDValue Hi = DAG.getNode(ARM64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
3396 // ARM64 shifts of larger than register sizes are wrapped rather than clamped,
3397 // so we can't just emit "lo << a" if a is too big.
3398 SDValue TrueValLo = DAG.getConstant(0, VT);
3399 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3401 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3403 SDValue Ops[2] = { Lo, Hi };
3404 return DAG.getMergeValues(Ops, 2, dl);
3408 ARM64TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3409 // The ARM64 target doesn't support folding offsets into global addresses.
3413 bool ARM64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3414 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
3415 // FIXME: We should be able to handle f128 as well with a clever lowering.
3416 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
3420 return ARM64_AM::getFP64Imm(Imm) != -1;
3421 else if (VT == MVT::f32)
3422 return ARM64_AM::getFP32Imm(Imm) != -1;
3426 //===----------------------------------------------------------------------===//
3427 // ARM64 Optimization Hooks
3428 //===----------------------------------------------------------------------===//
3430 //===----------------------------------------------------------------------===//
3431 // ARM64 Inline Assembly Support
3432 //===----------------------------------------------------------------------===//
3434 // Table of Constraints
3435 // TODO: This is the current set of constraints supported by ARM for the
3436 // compiler, not all of them may make sense, e.g. S may be difficult to support.
3438 // r - A general register
3439 // w - An FP/SIMD register of some size in the range v0-v31
3440 // x - An FP/SIMD register of some size in the range v0-v15
3441 // I - Constant that can be used with an ADD instruction
3442 // J - Constant that can be used with a SUB instruction
3443 // K - Constant that can be used with a 32-bit logical instruction
3444 // L - Constant that can be used with a 64-bit logical instruction
3445 // M - Constant that can be used as a 32-bit MOV immediate
3446 // N - Constant that can be used as a 64-bit MOV immediate
3447 // Q - A memory reference with base register and no offset
3448 // S - A symbolic address
3449 // Y - Floating point constant zero
3450 // Z - Integer constant zero
3452 // Note that general register operands will be output using their 64-bit x
3453 // register name, whatever the size of the variable, unless the asm operand
3454 // is prefixed by the %w modifier. Floating-point and SIMD register operands
3455 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
3458 /// getConstraintType - Given a constraint letter, return the type of
3459 /// constraint it is for this target.
3460 ARM64TargetLowering::ConstraintType
3461 ARM64TargetLowering::getConstraintType(const std::string &Constraint) const {
3462 if (Constraint.size() == 1) {
3463 switch (Constraint[0]) {
3470 return C_RegisterClass;
3471 // An address with a single base register. Due to the way we
3472 // currently handle addresses it is the same as 'r'.
3477 return TargetLowering::getConstraintType(Constraint);
3480 /// Examine constraint type and operand type and determine a weight value.
3481 /// This object must already have been set up with the operand type
3482 /// and the current alternative constraint selected.
3483 TargetLowering::ConstraintWeight
3484 ARM64TargetLowering::getSingleConstraintMatchWeight(
3485 AsmOperandInfo &info, const char *constraint) const {
3486 ConstraintWeight weight = CW_Invalid;
3487 Value *CallOperandVal = info.CallOperandVal;
3488 // If we don't have a value, we can't do a match,
3489 // but allow it at the lowest weight.
3490 if (CallOperandVal == NULL)
3492 Type *type = CallOperandVal->getType();
3493 // Look at the constraint type.
3494 switch (*constraint) {
3496 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3500 if (type->isFloatingPointTy() || type->isVectorTy())
3501 weight = CW_Register;
3504 weight = CW_Constant;
3510 std::pair<unsigned, const TargetRegisterClass *>
3511 ARM64TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3513 if (Constraint.size() == 1) {
3514 switch (Constraint[0]) {
3516 if (VT.getSizeInBits() == 64)
3517 return std::make_pair(0U, &ARM64::GPR64commonRegClass);
3518 return std::make_pair(0U, &ARM64::GPR32commonRegClass);
3521 return std::make_pair(0U, &ARM64::FPR32RegClass);
3522 if (VT.getSizeInBits() == 64)
3523 return std::make_pair(0U, &ARM64::FPR64RegClass);
3524 if (VT.getSizeInBits() == 128)
3525 return std::make_pair(0U, &ARM64::FPR128RegClass);
3527 // The instructions that this constraint is designed for can
3528 // only take 128-bit registers so just use that regclass.
3530 if (VT.getSizeInBits() == 128)
3531 return std::make_pair(0U, &ARM64::FPR128_loRegClass);
3535 if (StringRef("{cc}").equals_lower(Constraint))
3536 return std::make_pair(unsigned(ARM64::CPSR), &ARM64::CCRRegClass);
3538 // Use the default implementation in TargetLowering to convert the register
3539 // constraint into a member of a register class.
3540 std::pair<unsigned, const TargetRegisterClass *> Res;
3541 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3543 // Not found as a standard register?
3544 if (Res.second == 0) {
3545 unsigned Size = Constraint.size();
3546 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
3547 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
3548 const std::string Reg =
3549 std::string(&Constraint[2], &Constraint[Size - 1]);
3550 int RegNo = atoi(Reg.c_str());
3551 if (RegNo >= 0 && RegNo <= 31) {
3552 // v0 - v31 are aliases of q0 - q31.
3553 // By default we'll emit v0-v31 for this unless there's a modifier where
3554 // we'll emit the correct register as well.
3555 Res.first = ARM64::FPR128RegClass.getRegister(RegNo);
3556 Res.second = &ARM64::FPR128RegClass;
3564 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3565 /// vector. If it is invalid, don't add anything to Ops.
3566 void ARM64TargetLowering::LowerAsmOperandForConstraint(
3567 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
3568 SelectionDAG &DAG) const {
3569 SDValue Result(0, 0);
3571 // Currently only support length 1 constraints.
3572 if (Constraint.length() != 1)
3575 char ConstraintLetter = Constraint[0];
3576 switch (ConstraintLetter) {
3580 // This set of constraints deal with valid constants for various instructions.
3581 // Validate and return a target constant for them if we can.
3583 // 'z' maps to xzr or wzr so it needs an input of 0.
3584 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3585 if (!C || C->getZExtValue() != 0)
3588 if (Op.getValueType() == MVT::i64)
3589 Result = DAG.getRegister(ARM64::XZR, MVT::i64);
3591 Result = DAG.getRegister(ARM64::WZR, MVT::i32);
3601 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3605 // Grab the value and do some validation.
3606 uint64_t CVal = C->getZExtValue();
3607 switch (ConstraintLetter) {
3608 // The I constraint applies only to simple ADD or SUB immediate operands:
3609 // i.e. 0 to 4095 with optional shift by 12
3610 // The J constraint applies only to ADD or SUB immediates that would be
3611 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
3612 // instruction [or vice versa], in other words -1 to -4095 with optional
3613 // left shift by 12.
3615 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
3619 uint64_t NVal = -C->getSExtValue();
3620 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal))
3624 // The K and L constraints apply *only* to logical immediates, including
3625 // what used to be the MOVI alias for ORR (though the MOVI alias has now
3626 // been removed and MOV should be used). So these constraints have to
3627 // distinguish between bit patterns that are valid 32-bit or 64-bit
3628 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
3629 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
3632 if (ARM64_AM::isLogicalImmediate(CVal, 32))
3636 if (ARM64_AM::isLogicalImmediate(CVal, 64))
3639 // The M and N constraints are a superset of K and L respectively, for use
3640 // with the MOV (immediate) alias. As well as the logical immediates they
3641 // also match 32 or 64-bit immediates that can be loaded either using a
3642 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
3643 // (M) or 64-bit 0x1234000000000000 (N) etc.
3644 // As a note some of this code is liberally stolen from the asm parser.
3646 if (!isUInt<32>(CVal))
3648 if (ARM64_AM::isLogicalImmediate(CVal, 32))
3650 if ((CVal & 0xFFFF) == CVal)
3652 if ((CVal & 0xFFFF0000ULL) == CVal)
3654 uint64_t NCVal = ~(uint32_t)CVal;
3655 if ((NCVal & 0xFFFFULL) == NCVal)
3657 if ((NCVal & 0xFFFF0000ULL) == NCVal)
3662 if (ARM64_AM::isLogicalImmediate(CVal, 64))
3664 if ((CVal & 0xFFFFULL) == CVal)
3666 if ((CVal & 0xFFFF0000ULL) == CVal)
3668 if ((CVal & 0xFFFF00000000ULL) == CVal)
3670 if ((CVal & 0xFFFF000000000000ULL) == CVal)
3672 uint64_t NCVal = ~CVal;
3673 if ((NCVal & 0xFFFFULL) == NCVal)
3675 if ((NCVal & 0xFFFF0000ULL) == NCVal)
3677 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
3679 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
3687 // All assembler immediates are 64-bit integers.
3688 Result = DAG.getTargetConstant(CVal, MVT::i64);
3692 if (Result.getNode()) {
3693 Ops.push_back(Result);
3697 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
3700 //===----------------------------------------------------------------------===//
3701 // ARM64 Advanced SIMD Support
3702 //===----------------------------------------------------------------------===//
3704 /// WidenVector - Given a value in the V64 register class, produce the
3705 /// equivalent value in the V128 register class.
3706 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
3707 EVT VT = V64Reg.getValueType();
3708 unsigned NarrowSize = VT.getVectorNumElements();
3709 MVT EltTy = VT.getVectorElementType().getSimpleVT();
3710 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
3713 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
3714 V64Reg, DAG.getConstant(0, MVT::i32));
3717 /// getExtFactor - Determine the adjustment factor for the position when
3718 /// generating an "extract from vector registers" instruction.
3719 static unsigned getExtFactor(SDValue &V) {
3720 EVT EltType = V.getValueType().getVectorElementType();
3721 return EltType.getSizeInBits() / 8;
3724 /// NarrowVector - Given a value in the V128 register class, produce the
3725 /// equivalent value in the V64 register class.
3726 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
3727 EVT VT = V128Reg.getValueType();
3728 unsigned WideSize = VT.getVectorNumElements();
3729 MVT EltTy = VT.getVectorElementType().getSimpleVT();
3730 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
3733 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, NarrowTy, V128Reg);
3736 // Gather data to see if the operation can be modelled as a
3737 // shuffle in combination with VEXTs.
3738 SDValue ARM64TargetLowering::ReconstructShuffle(SDValue Op,
3739 SelectionDAG &DAG) const {
3741 EVT VT = Op.getValueType();
3742 unsigned NumElts = VT.getVectorNumElements();
3744 SmallVector<SDValue, 2> SourceVecs;
3745 SmallVector<unsigned, 2> MinElts;
3746 SmallVector<unsigned, 2> MaxElts;
3748 for (unsigned i = 0; i < NumElts; ++i) {
3749 SDValue V = Op.getOperand(i);
3750 if (V.getOpcode() == ISD::UNDEF)
3752 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
3753 // A shuffle can only come from building a vector from various
3754 // elements of other vectors.
3758 // Record this extraction against the appropriate vector if possible...
3759 SDValue SourceVec = V.getOperand(0);
3760 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
3761 bool FoundSource = false;
3762 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
3763 if (SourceVecs[j] == SourceVec) {
3764 if (MinElts[j] > EltNo)
3766 if (MaxElts[j] < EltNo)
3773 // Or record a new source if not...
3775 SourceVecs.push_back(SourceVec);
3776 MinElts.push_back(EltNo);
3777 MaxElts.push_back(EltNo);
3781 // Currently only do something sane when at most two source vectors
3783 if (SourceVecs.size() > 2)
3786 SDValue ShuffleSrcs[2] = { DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
3787 int VEXTOffsets[2] = { 0, 0 };
3789 // This loop extracts the usage patterns of the source vectors
3790 // and prepares appropriate SDValues for a shuffle if possible.
3791 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
3792 if (SourceVecs[i].getValueType() == VT) {
3793 // No VEXT necessary
3794 ShuffleSrcs[i] = SourceVecs[i];
3797 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
3798 // It probably isn't worth padding out a smaller vector just to
3799 // break it down again in a shuffle.
3803 // Don't attempt to extract subvectors from BUILD_VECTOR sources
3804 // that expand or trunc the original value.
3805 // TODO: We can try to bitcast and ANY_EXTEND the result but
3806 // we need to consider the cost of vector ANY_EXTEND, and the
3807 // legality of all the types.
3808 if (SourceVecs[i].getValueType().getVectorElementType() !=
3809 VT.getVectorElementType())
3812 // Since only 64-bit and 128-bit vectors are legal on ARM and
3813 // we've eliminated the other cases...
3814 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2 * NumElts &&
3815 "unexpected vector sizes in ReconstructShuffle");
3817 if (MaxElts[i] - MinElts[i] >= NumElts) {
3818 // Span too large for a VEXT to cope
3822 if (MinElts[i] >= NumElts) {
3823 // The extraction can just take the second half
3824 VEXTOffsets[i] = NumElts;
3826 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
3827 DAG.getIntPtrConstant(NumElts));
3828 } else if (MaxElts[i] < NumElts) {
3829 // The extraction can just take the first half
3831 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3832 SourceVecs[i], DAG.getIntPtrConstant(0));
3834 // An actual VEXT is needed
3835 VEXTOffsets[i] = MinElts[i];
3836 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3837 SourceVecs[i], DAG.getIntPtrConstant(0));
3839 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
3840 DAG.getIntPtrConstant(NumElts));
3841 unsigned Imm = VEXTOffsets[i] * getExtFactor(VEXTSrc1);
3842 ShuffleSrcs[i] = DAG.getNode(ARM64ISD::EXT, dl, VT, VEXTSrc1, VEXTSrc2,
3843 DAG.getConstant(Imm, MVT::i32));
3847 SmallVector<int, 8> Mask;
3849 for (unsigned i = 0; i < NumElts; ++i) {
3850 SDValue Entry = Op.getOperand(i);
3851 if (Entry.getOpcode() == ISD::UNDEF) {
3856 SDValue ExtractVec = Entry.getOperand(0);
3858 cast<ConstantSDNode>(Op.getOperand(i).getOperand(1))->getSExtValue();
3859 if (ExtractVec == SourceVecs[0]) {
3860 Mask.push_back(ExtractElt - VEXTOffsets[0]);
3862 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
3866 // Final check before we try to produce nonsense...
3867 if (isShuffleMaskLegal(Mask, VT))
3868 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
3874 // check if an EXT instruction can handle the shuffle mask when the
3875 // vector sources of the shuffle are the same.
3876 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
3877 unsigned NumElts = VT.getVectorNumElements();
3879 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3885 // If this is a VEXT shuffle, the immediate value is the index of the first
3886 // element. The other shuffle indices must be the successive elements after
3888 unsigned ExpectedElt = Imm;
3889 for (unsigned i = 1; i < NumElts; ++i) {
3890 // Increment the expected index. If it wraps around, just follow it
3891 // back to index zero and keep going.
3893 if (ExpectedElt == NumElts)
3897 continue; // ignore UNDEF indices
3898 if (ExpectedElt != static_cast<unsigned>(M[i]))
3905 // check if an EXT instruction can handle the shuffle mask when the
3906 // vector sources of the shuffle are different.
3907 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
3909 unsigned NumElts = VT.getVectorNumElements();
3912 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3918 // If this is a VEXT shuffle, the immediate value is the index of the first
3919 // element. The other shuffle indices must be the successive elements after
3921 unsigned ExpectedElt = Imm;
3922 for (unsigned i = 1; i < NumElts; ++i) {
3923 // Increment the expected index. If it wraps around, it may still be
3924 // a VEXT but the source vectors must be swapped.
3926 if (ExpectedElt == NumElts * 2) {
3932 continue; // ignore UNDEF indices
3933 if (ExpectedElt != static_cast<unsigned>(M[i]))
3937 // Adjust the index value if the source operands will be swapped.
3944 /// isREVMask - Check if a vector shuffle corresponds to a REV
3945 /// instruction with the specified blocksize. (The order of the elements
3946 /// within each block of the vector is reversed.)
3947 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
3948 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
3949 "Only possible block sizes for REV are: 16, 32, 64");
3951 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3955 unsigned NumElts = VT.getVectorNumElements();
3956 unsigned BlockElts = M[0] + 1;
3957 // If the first shuffle index is UNDEF, be optimistic.
3959 BlockElts = BlockSize / EltSz;
3961 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3964 for (unsigned i = 0; i < NumElts; ++i) {
3966 continue; // ignore UNDEF indices
3967 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
3974 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3975 unsigned NumElts = VT.getVectorNumElements();
3976 WhichResult = (M[0] == 0 ? 0 : 1);
3977 unsigned Idx = WhichResult * NumElts / 2;
3978 for (unsigned i = 0; i != NumElts; i += 2) {
3979 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
3980 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
3988 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3989 unsigned NumElts = VT.getVectorNumElements();
3990 WhichResult = (M[0] == 0 ? 0 : 1);
3991 for (unsigned i = 0; i != NumElts; ++i) {
3993 continue; // ignore UNDEF indices
3994 if ((unsigned)M[i] != 2 * i + WhichResult)
4001 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4002 unsigned NumElts = VT.getVectorNumElements();
4003 WhichResult = (M[0] == 0 ? 0 : 1);
4004 for (unsigned i = 0; i < NumElts; i += 2) {
4005 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4006 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
4012 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
4013 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4014 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4015 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4016 unsigned NumElts = VT.getVectorNumElements();
4017 WhichResult = (M[0] == 0 ? 0 : 1);
4018 unsigned Idx = WhichResult * NumElts / 2;
4019 for (unsigned i = 0; i != NumElts; i += 2) {
4020 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4021 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
4029 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
4030 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4031 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4032 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4033 unsigned Half = VT.getVectorNumElements() / 2;
4034 WhichResult = (M[0] == 0 ? 0 : 1);
4035 for (unsigned j = 0; j != 2; ++j) {
4036 unsigned Idx = WhichResult;
4037 for (unsigned i = 0; i != Half; ++i) {
4038 int MIdx = M[i + j * Half];
4039 if (MIdx >= 0 && (unsigned)MIdx != Idx)
4048 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
4049 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4050 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4051 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4052 unsigned NumElts = VT.getVectorNumElements();
4053 WhichResult = (M[0] == 0 ? 0 : 1);
4054 for (unsigned i = 0; i < NumElts; i += 2) {
4055 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4056 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
4062 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4063 /// the specified operations to build the shuffle.
4064 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4065 SDValue RHS, SelectionDAG &DAG,
4067 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4068 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
4069 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
4072 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4081 OP_VUZPL, // VUZP, left result
4082 OP_VUZPR, // VUZP, right result
4083 OP_VZIPL, // VZIP, left result
4084 OP_VZIPR, // VZIP, right result
4085 OP_VTRNL, // VTRN, left result
4086 OP_VTRNR // VTRN, right result
4089 if (OpNum == OP_COPY) {
4090 if (LHSID == (1 * 9 + 2) * 9 + 3)
4092 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
4096 SDValue OpLHS, OpRHS;
4097 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4098 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4099 EVT VT = OpLHS.getValueType();
4103 llvm_unreachable("Unknown shuffle opcode!");
4105 // VREV divides the vector in half and swaps within the half.
4106 if (VT.getVectorElementType() == MVT::i32 ||
4107 VT.getVectorElementType() == MVT::f32)
4108 return DAG.getNode(ARM64ISD::REV64, dl, VT, OpLHS);
4109 // vrev <4 x i16> -> REV32
4110 if (VT.getVectorElementType() == MVT::i16)
4111 return DAG.getNode(ARM64ISD::REV32, dl, VT, OpLHS);
4112 // vrev <4 x i8> -> REV16
4113 assert(VT.getVectorElementType() == MVT::i8);
4114 return DAG.getNode(ARM64ISD::REV16, dl, VT, OpLHS);
4119 EVT EltTy = VT.getVectorElementType();
4121 if (EltTy == MVT::i8)
4122 Opcode = ARM64ISD::DUPLANE8;
4123 else if (EltTy == MVT::i16)
4124 Opcode = ARM64ISD::DUPLANE16;
4125 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
4126 Opcode = ARM64ISD::DUPLANE32;
4127 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
4128 Opcode = ARM64ISD::DUPLANE64;
4130 llvm_unreachable("Invalid vector element type?");
4132 if (VT.getSizeInBits() == 64)
4133 OpLHS = WidenVector(OpLHS, DAG);
4134 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, MVT::i64);
4135 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
4140 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
4141 return DAG.getNode(ARM64ISD::EXT, dl, VT, OpLHS, OpRHS,
4142 DAG.getConstant(Imm, MVT::i32));
4145 return DAG.getNode(ARM64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4147 return DAG.getNode(ARM64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4149 return DAG.getNode(ARM64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4151 return DAG.getNode(ARM64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4153 return DAG.getNode(ARM64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4155 return DAG.getNode(ARM64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4159 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
4160 SelectionDAG &DAG) {
4161 // Check to see if we can use the TBL instruction.
4162 SDValue V1 = Op.getOperand(0);
4163 SDValue V2 = Op.getOperand(1);
4166 EVT EltVT = Op.getValueType().getVectorElementType();
4167 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
4169 SmallVector<SDValue, 8> TBLMask;
4170 for (int Val : ShuffleMask) {
4171 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
4172 unsigned Offset = Byte + Val * BytesPerElt;
4173 TBLMask.push_back(DAG.getConstant(Offset, MVT::i32));
4177 MVT IndexVT = MVT::v8i8;
4178 unsigned IndexLen = 8;
4179 if (Op.getValueType().getSizeInBits() == 128) {
4180 IndexVT = MVT::v16i8;
4184 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
4185 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
4188 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
4190 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
4191 Shuffle = DAG.getNode(
4192 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4193 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4194 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4196 if (IndexLen == 8) {
4197 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
4198 Shuffle = DAG.getNode(
4199 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4200 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4201 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4203 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
4204 // cannot currently represent the register constraints on the input
4206 // Shuffle = DAG.getNode(ARM64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
4207 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4208 // &TBLMask[0], IndexLen));
4209 Shuffle = DAG.getNode(
4210 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4211 DAG.getConstant(Intrinsic::arm64_neon_tbl2, MVT::i32), V1Cst, V2Cst,
4212 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4215 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
4218 static unsigned getDUPLANEOp(EVT EltType) {
4219 if (EltType == MVT::i8)
4220 return ARM64ISD::DUPLANE8;
4221 if (EltType == MVT::i16)
4222 return ARM64ISD::DUPLANE16;
4223 if (EltType == MVT::i32 || EltType == MVT::f32)
4224 return ARM64ISD::DUPLANE32;
4225 if (EltType == MVT::i64 || EltType == MVT::f64)
4226 return ARM64ISD::DUPLANE64;
4228 llvm_unreachable("Invalid vector element type?");
4231 SDValue ARM64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4232 SelectionDAG &DAG) const {
4234 EVT VT = Op.getValueType();
4236 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4238 // Convert shuffles that are directly supported on NEON to target-specific
4239 // DAG nodes, instead of keeping them as shuffles and matching them again
4240 // during code selection. This is more efficient and avoids the possibility
4241 // of inconsistencies between legalization and selection.
4242 ArrayRef<int> ShuffleMask = SVN->getMask();
4244 SDValue V1 = Op.getOperand(0);
4245 SDValue V2 = Op.getOperand(1);
4247 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
4248 V1.getValueType().getSimpleVT())) {
4249 int Lane = SVN->getSplatIndex();
4250 // If this is undef splat, generate it via "just" vdup, if possible.
4254 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
4255 return DAG.getNode(ARM64ISD::DUP, dl, V1.getValueType(),
4257 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
4258 // constant. If so, we can just reference the lane's definition directly.
4259 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
4260 !isa<ConstantSDNode>(V1.getOperand(Lane)))
4261 return DAG.getNode(ARM64ISD::DUP, dl, VT, V1.getOperand(Lane));
4263 // Otherwise, duplicate from the lane of the input vector.
4264 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
4266 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
4267 // to make a vector of the same size as this SHUFFLE. We can ignore the
4268 // extract entirely, and canonicalise the concat using WidenVector.
4269 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
4270 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
4271 V1 = V1.getOperand(0);
4272 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
4273 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
4274 Lane -= Idx * VT.getVectorNumElements() / 2;
4275 V1 = WidenVector(V1.getOperand(Idx), DAG);
4276 } else if (VT.getSizeInBits() == 64)
4277 V1 = WidenVector(V1, DAG);
4279 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, MVT::i64));
4282 if (isREVMask(ShuffleMask, VT, 64))
4283 return DAG.getNode(ARM64ISD::REV64, dl, V1.getValueType(), V1, V2);
4284 if (isREVMask(ShuffleMask, VT, 32))
4285 return DAG.getNode(ARM64ISD::REV32, dl, V1.getValueType(), V1, V2);
4286 if (isREVMask(ShuffleMask, VT, 16))
4287 return DAG.getNode(ARM64ISD::REV16, dl, V1.getValueType(), V1, V2);
4289 bool ReverseEXT = false;
4291 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
4294 Imm *= getExtFactor(V1);
4295 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V2,
4296 DAG.getConstant(Imm, MVT::i32));
4297 } else if (V2->getOpcode() == ISD::UNDEF &&
4298 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
4299 Imm *= getExtFactor(V1);
4300 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V1,
4301 DAG.getConstant(Imm, MVT::i32));
4304 unsigned WhichResult;
4305 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
4306 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4307 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4309 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
4310 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4311 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4313 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
4314 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4315 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4318 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4319 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4320 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4322 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4323 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4324 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4326 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4327 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4328 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4331 // If the shuffle is not directly supported and it has 4 elements, use
4332 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4333 unsigned NumElts = VT.getVectorNumElements();
4335 unsigned PFIndexes[4];
4336 for (unsigned i = 0; i != 4; ++i) {
4337 if (ShuffleMask[i] < 0)
4340 PFIndexes[i] = ShuffleMask[i];
4343 // Compute the index in the perfect shuffle table.
4344 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
4345 PFIndexes[2] * 9 + PFIndexes[3];
4346 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4347 unsigned Cost = (PFEntry >> 30);
4350 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4353 return GenerateTBL(Op, ShuffleMask, DAG);
4356 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
4358 EVT VT = BVN->getValueType(0);
4359 APInt SplatBits, SplatUndef;
4360 unsigned SplatBitSize;
4362 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4363 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
4365 for (unsigned i = 0; i < NumSplats; ++i) {
4366 CnstBits <<= SplatBitSize;
4367 UndefBits <<= SplatBitSize;
4368 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
4369 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
4378 SDValue ARM64TargetLowering::LowerVectorAND(SDValue Op,
4379 SelectionDAG &DAG) const {
4380 BuildVectorSDNode *BVN =
4381 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
4382 SDValue LHS = Op.getOperand(0);
4384 EVT VT = Op.getValueType();
4389 APInt CnstBits(VT.getSizeInBits(), 0);
4390 APInt UndefBits(VT.getSizeInBits(), 0);
4391 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
4392 // We only have BIC vector immediate instruction, which is and-not.
4393 CnstBits = ~CnstBits;
4395 // We make use of a little bit of goto ickiness in order to avoid having to
4396 // duplicate the immediate matching logic for the undef toggled case.
4397 bool SecondTry = false;
4400 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
4401 CnstBits = CnstBits.zextOrTrunc(64);
4402 uint64_t CnstVal = CnstBits.getZExtValue();
4404 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
4405 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
4406 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4407 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4408 DAG.getConstant(CnstVal, MVT::i32),
4409 DAG.getConstant(0, MVT::i32));
4410 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4413 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
4414 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
4415 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4416 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4417 DAG.getConstant(CnstVal, MVT::i32),
4418 DAG.getConstant(8, MVT::i32));
4419 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4422 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
4423 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
4424 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4425 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4426 DAG.getConstant(CnstVal, MVT::i32),
4427 DAG.getConstant(16, MVT::i32));
4428 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4431 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
4432 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
4433 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4434 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4435 DAG.getConstant(CnstVal, MVT::i32),
4436 DAG.getConstant(24, MVT::i32));
4437 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4440 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
4441 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
4442 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4443 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4444 DAG.getConstant(CnstVal, MVT::i32),
4445 DAG.getConstant(0, MVT::i32));
4446 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4449 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
4450 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
4451 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4452 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4453 DAG.getConstant(CnstVal, MVT::i32),
4454 DAG.getConstant(8, MVT::i32));
4455 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4462 CnstBits = ~UndefBits;
4466 // We can always fall back to a non-immediate AND.
4471 // Specialized code to quickly find if PotentialBVec is a BuildVector that
4472 // consists of only the same constant int value, returned in reference arg
4474 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
4475 uint64_t &ConstVal) {
4476 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
4479 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
4482 EVT VT = Bvec->getValueType(0);
4483 unsigned NumElts = VT.getVectorNumElements();
4484 for (unsigned i = 1; i < NumElts; ++i)
4485 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
4487 ConstVal = FirstElt->getZExtValue();
4491 static unsigned getIntrinsicID(const SDNode *N) {
4492 unsigned Opcode = N->getOpcode();
4495 return Intrinsic::not_intrinsic;
4496 case ISD::INTRINSIC_WO_CHAIN: {
4497 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4498 if (IID < Intrinsic::num_intrinsics)
4500 return Intrinsic::not_intrinsic;
4505 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
4506 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
4507 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
4508 // Also, logical shift right -> sri, with the same structure.
4509 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
4510 EVT VT = N->getValueType(0);
4517 // Is the first op an AND?
4518 const SDValue And = N->getOperand(0);
4519 if (And.getOpcode() != ISD::AND)
4522 // Is the second op an shl or lshr?
4523 SDValue Shift = N->getOperand(1);
4524 // This will have been turned into: ARM64ISD::VSHL vector, #shift
4525 // or ARM64ISD::VLSHR vector, #shift
4526 unsigned ShiftOpc = Shift.getOpcode();
4527 if ((ShiftOpc != ARM64ISD::VSHL && ShiftOpc != ARM64ISD::VLSHR))
4529 bool IsShiftRight = ShiftOpc == ARM64ISD::VLSHR;
4531 // Is the shift amount constant?
4532 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
4536 // Is the and mask vector all constant?
4538 if (!isAllConstantBuildVector(And.getOperand(1), C1))
4541 // Is C1 == ~C2, taking into account how much one can shift elements of a
4543 uint64_t C2 = C2node->getZExtValue();
4544 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
4545 if (C2 > ElemSizeInBits)
4547 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
4548 if ((C1 & ElemMask) != (~C2 & ElemMask))
4551 SDValue X = And.getOperand(0);
4552 SDValue Y = Shift.getOperand(0);
4555 IsShiftRight ? Intrinsic::arm64_neon_vsri : Intrinsic::arm64_neon_vsli;
4557 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
4558 DAG.getConstant(Intrin, MVT::i32), X, Y, Shift.getOperand(1));
4560 DEBUG(dbgs() << "arm64-lower: transformed: \n");
4561 DEBUG(N->dump(&DAG));
4562 DEBUG(dbgs() << "into: \n");
4563 DEBUG(ResultSLI->dump(&DAG));
4569 SDValue ARM64TargetLowering::LowerVectorOR(SDValue Op,
4570 SelectionDAG &DAG) const {
4571 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
4572 if (EnableARM64SlrGeneration) {
4573 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
4578 BuildVectorSDNode *BVN =
4579 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
4580 SDValue LHS = Op.getOperand(1);
4582 EVT VT = Op.getValueType();
4584 // OR commutes, so try swapping the operands.
4586 LHS = Op.getOperand(0);
4587 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
4592 APInt CnstBits(VT.getSizeInBits(), 0);
4593 APInt UndefBits(VT.getSizeInBits(), 0);
4594 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
4595 // We make use of a little bit of goto ickiness in order to avoid having to
4596 // duplicate the immediate matching logic for the undef toggled case.
4597 bool SecondTry = false;
4600 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
4601 CnstBits = CnstBits.zextOrTrunc(64);
4602 uint64_t CnstVal = CnstBits.getZExtValue();
4604 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
4605 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
4606 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4607 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4608 DAG.getConstant(CnstVal, MVT::i32),
4609 DAG.getConstant(0, MVT::i32));
4610 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4613 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
4614 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
4615 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4616 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4617 DAG.getConstant(CnstVal, MVT::i32),
4618 DAG.getConstant(8, MVT::i32));
4619 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4622 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
4623 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
4624 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4625 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4626 DAG.getConstant(CnstVal, MVT::i32),
4627 DAG.getConstant(16, MVT::i32));
4628 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4631 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
4632 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
4633 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4634 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4635 DAG.getConstant(CnstVal, MVT::i32),
4636 DAG.getConstant(24, MVT::i32));
4637 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4640 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
4641 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
4642 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4643 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4644 DAG.getConstant(CnstVal, MVT::i32),
4645 DAG.getConstant(0, MVT::i32));
4646 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4649 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
4650 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
4651 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4652 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4653 DAG.getConstant(CnstVal, MVT::i32),
4654 DAG.getConstant(8, MVT::i32));
4655 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4662 CnstBits = UndefBits;
4666 // We can always fall back to a non-immediate OR.
4671 SDValue ARM64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
4672 SelectionDAG &DAG) const {
4673 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4675 EVT VT = Op.getValueType();
4677 APInt CnstBits(VT.getSizeInBits(), 0);
4678 APInt UndefBits(VT.getSizeInBits(), 0);
4679 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
4680 // We make use of a little bit of goto ickiness in order to avoid having to
4681 // duplicate the immediate matching logic for the undef toggled case.
4682 bool SecondTry = false;
4685 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
4686 CnstBits = CnstBits.zextOrTrunc(64);
4687 uint64_t CnstVal = CnstBits.getZExtValue();
4689 // Certain magic vector constants (used to express things like NOT
4690 // and NEG) are passed through unmodified. This allows codegen patterns
4691 // for these operations to match. Special-purpose patterns will lower
4692 // these immediates to MOVIs if it proves necessary.
4693 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
4696 // The many faces of MOVI...
4697 if (ARM64_AM::isAdvSIMDModImmType10(CnstVal)) {
4698 CnstVal = ARM64_AM::encodeAdvSIMDModImmType10(CnstVal);
4699 if (VT.getSizeInBits() == 128) {
4700 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::v2i64,
4701 DAG.getConstant(CnstVal, MVT::i32));
4702 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4705 // Support the V64 version via subregister insertion.
4706 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::f64,
4707 DAG.getConstant(CnstVal, MVT::i32));
4708 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4711 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
4712 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
4713 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4714 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
4715 DAG.getConstant(CnstVal, MVT::i32),
4716 DAG.getConstant(0, MVT::i32));
4717 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4720 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
4721 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
4722 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4723 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
4724 DAG.getConstant(CnstVal, MVT::i32),
4725 DAG.getConstant(8, MVT::i32));
4726 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4729 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
4730 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
4731 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4732 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
4733 DAG.getConstant(CnstVal, MVT::i32),
4734 DAG.getConstant(16, MVT::i32));
4735 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4738 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
4739 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
4740 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4741 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
4742 DAG.getConstant(CnstVal, MVT::i32),
4743 DAG.getConstant(24, MVT::i32));
4744 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4747 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
4748 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
4749 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4750 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
4751 DAG.getConstant(CnstVal, MVT::i32),
4752 DAG.getConstant(0, MVT::i32));
4753 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4756 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
4757 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
4758 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4759 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
4760 DAG.getConstant(CnstVal, MVT::i32),
4761 DAG.getConstant(8, MVT::i32));
4762 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4765 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
4766 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
4767 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4768 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
4769 DAG.getConstant(CnstVal, MVT::i32),
4770 DAG.getConstant(264, MVT::i32));
4771 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4774 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
4775 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
4776 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4777 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
4778 DAG.getConstant(CnstVal, MVT::i32),
4779 DAG.getConstant(272, MVT::i32));
4780 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4783 if (ARM64_AM::isAdvSIMDModImmType9(CnstVal)) {
4784 CnstVal = ARM64_AM::encodeAdvSIMDModImmType9(CnstVal);
4785 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
4786 SDValue Mov = DAG.getNode(ARM64ISD::MOVI, dl, MovTy,
4787 DAG.getConstant(CnstVal, MVT::i32));
4788 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4791 // The few faces of FMOV...
4792 if (ARM64_AM::isAdvSIMDModImmType11(CnstVal)) {
4793 CnstVal = ARM64_AM::encodeAdvSIMDModImmType11(CnstVal);
4794 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
4795 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MovTy,
4796 DAG.getConstant(CnstVal, MVT::i32));
4797 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4800 if (ARM64_AM::isAdvSIMDModImmType12(CnstVal) &&
4801 VT.getSizeInBits() == 128) {
4802 CnstVal = ARM64_AM::encodeAdvSIMDModImmType12(CnstVal);
4803 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MVT::v2f64,
4804 DAG.getConstant(CnstVal, MVT::i32));
4805 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4808 // The many faces of MVNI...
4810 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
4811 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
4812 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4813 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
4814 DAG.getConstant(CnstVal, MVT::i32),
4815 DAG.getConstant(0, MVT::i32));
4816 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4819 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
4820 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
4821 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4822 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
4823 DAG.getConstant(CnstVal, MVT::i32),
4824 DAG.getConstant(8, MVT::i32));
4825 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4828 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
4829 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
4830 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4831 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
4832 DAG.getConstant(CnstVal, MVT::i32),
4833 DAG.getConstant(16, MVT::i32));
4834 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4837 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
4838 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
4839 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4840 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
4841 DAG.getConstant(CnstVal, MVT::i32),
4842 DAG.getConstant(24, MVT::i32));
4843 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4846 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
4847 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
4848 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4849 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
4850 DAG.getConstant(CnstVal, MVT::i32),
4851 DAG.getConstant(0, MVT::i32));
4852 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4855 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
4856 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
4857 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4858 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
4859 DAG.getConstant(CnstVal, MVT::i32),
4860 DAG.getConstant(8, MVT::i32));
4861 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4864 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
4865 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
4866 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4867 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
4868 DAG.getConstant(CnstVal, MVT::i32),
4869 DAG.getConstant(264, MVT::i32));
4870 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4873 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
4874 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
4875 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4876 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
4877 DAG.getConstant(CnstVal, MVT::i32),
4878 DAG.getConstant(272, MVT::i32));
4879 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4886 CnstBits = UndefBits;
4891 // Scan through the operands to find some interesting properties we can
4893 // 1) If only one value is used, we can use a DUP, or
4894 // 2) if only the low element is not undef, we can just insert that, or
4895 // 3) if only one constant value is used (w/ some non-constant lanes),
4896 // we can splat the constant value into the whole vector then fill
4897 // in the non-constant lanes.
4898 // 4) FIXME: If different constant values are used, but we can intelligently
4899 // select the values we'll be overwriting for the non-constant
4900 // lanes such that we can directly materialize the vector
4901 // some other way (MOVI, e.g.), we can be sneaky.
4902 unsigned NumElts = VT.getVectorNumElements();
4903 bool isOnlyLowElement = true;
4904 bool usesOnlyOneValue = true;
4905 bool usesOnlyOneConstantValue = true;
4906 bool isConstant = true;
4907 unsigned NumConstantLanes = 0;
4909 SDValue ConstantValue;
4910 for (unsigned i = 0; i < NumElts; ++i) {
4911 SDValue V = Op.getOperand(i);
4912 if (V.getOpcode() == ISD::UNDEF)
4915 isOnlyLowElement = false;
4916 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4919 if (isa<ConstantSDNode>(V) || isa<ConstantFPSDNode>(V)) {
4921 if (!ConstantValue.getNode())
4923 else if (ConstantValue != V)
4924 usesOnlyOneConstantValue = false;
4927 if (!Value.getNode())
4929 else if (V != Value)
4930 usesOnlyOneValue = false;
4933 if (!Value.getNode())
4934 return DAG.getUNDEF(VT);
4936 if (isOnlyLowElement)
4937 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4939 // Use DUP for non-constant splats. For f32 constant splats, reduce to
4940 // i32 and try again.
4941 if (usesOnlyOneValue) {
4943 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4944 Value.getValueType() != VT)
4945 return DAG.getNode(ARM64ISD::DUP, dl, VT, Value);
4947 // This is actually a DUPLANExx operation, which keeps everything vectory.
4949 // DUPLANE works on 128-bit vectors, widen it if necessary.
4950 SDValue Lane = Value.getOperand(1);
4951 Value = Value.getOperand(0);
4952 if (Value.getValueType().getSizeInBits() == 64)
4953 Value = WidenVector(Value, DAG);
4955 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
4956 return DAG.getNode(Opcode, dl, VT, Value, Lane);
4959 if (VT.getVectorElementType().isFloatingPoint()) {
4960 SmallVector<SDValue, 8> Ops;
4962 (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
4963 for (unsigned i = 0; i < NumElts; ++i)
4964 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
4965 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
4966 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4967 Val = LowerBUILD_VECTOR(Val, DAG);
4969 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4973 // If there was only one constant value used and for more than one lane,
4974 // start by splatting that value, then replace the non-constant lanes. This
4975 // is better than the default, which will perform a separate initialization
4977 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
4978 SDValue Val = DAG.getNode(ARM64ISD::DUP, dl, VT, ConstantValue);
4979 // Now insert the non-constant lanes.
4980 for (unsigned i = 0; i < NumElts; ++i) {
4981 SDValue V = Op.getOperand(i);
4982 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
4983 if (!isa<ConstantSDNode>(V) && !isa<ConstantFPSDNode>(V)) {
4984 // Note that type legalization likely mucked about with the VT of the
4985 // source operand, so we may have to convert it here before inserting.
4986 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
4992 // If all elements are constants and the case above didn't get hit, fall back
4993 // to the default expansion, which will generate a load from the constant
4998 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5000 SDValue shuffle = ReconstructShuffle(Op, DAG);
5001 if (shuffle != SDValue())
5005 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5006 // know the default expansion would otherwise fall back on something even
5007 // worse. For a vector with one or two non-undef values, that's
5008 // scalar_to_vector for the elements followed by a shuffle (provided the
5009 // shuffle is valid for the target) and materialization element by element
5010 // on the stack followed by a load for everything else.
5011 if (!isConstant && !usesOnlyOneValue) {
5012 SDValue Vec = DAG.getUNDEF(VT);
5013 SDValue Op0 = Op.getOperand(0);
5014 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
5016 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
5017 // a) Avoid a RMW dependency on the full vector register, and
5018 // b) Allow the register coalescer to fold away the copy if the
5019 // value is already in an S or D register.
5020 if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
5021 unsigned SubIdx = ElemSize == 32 ? ARM64::ssub : ARM64::dsub;
5023 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
5024 DAG.getTargetConstant(SubIdx, MVT::i32));
5025 Vec = SDValue(N, 0);
5028 for (; i < NumElts; ++i) {
5029 SDValue V = Op.getOperand(i);
5030 if (V.getOpcode() == ISD::UNDEF)
5032 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5033 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5038 // Just use the default expansion. We failed to find a better alternative.
5042 SDValue ARM64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
5043 SelectionDAG &DAG) const {
5044 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
5046 // Check for non-constant lane.
5047 if (!isa<ConstantSDNode>(Op.getOperand(2)))
5050 EVT VT = Op.getOperand(0).getValueType();
5052 // Insertion/extraction are legal for V128 types.
5053 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5054 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5057 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5058 VT != MVT::v1i64 && VT != MVT::v2f32)
5061 // For V64 types, we perform insertion by expanding the value
5062 // to a V128 type and perform the insertion on that.
5064 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5065 EVT WideTy = WideVec.getValueType();
5067 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
5068 Op.getOperand(1), Op.getOperand(2));
5069 // Re-narrow the resultant vector.
5070 return NarrowVector(Node, DAG);
5073 SDValue ARM64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5074 SelectionDAG &DAG) const {
5075 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
5077 // Check for non-constant lane.
5078 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5081 EVT VT = Op.getOperand(0).getValueType();
5083 // Insertion/extraction are legal for V128 types.
5084 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5085 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5088 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5089 VT != MVT::v1i64 && VT != MVT::v2f32)
5092 // For V64 types, we perform extraction by expanding the value
5093 // to a V128 type and perform the extraction on that.
5095 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5096 EVT WideTy = WideVec.getValueType();
5098 EVT ExtrTy = WideTy.getVectorElementType();
5099 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
5102 // For extractions, we just return the result directly.
5103 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
5107 SDValue ARM64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
5108 SelectionDAG &DAG) const {
5109 EVT VT = Op.getOperand(0).getValueType();
5115 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5118 unsigned Val = Cst->getZExtValue();
5120 unsigned Size = Op.getValueType().getSizeInBits();
5124 return DAG.getTargetExtractSubreg(ARM64::bsub, dl, Op.getValueType(),
5127 return DAG.getTargetExtractSubreg(ARM64::hsub, dl, Op.getValueType(),
5130 return DAG.getTargetExtractSubreg(ARM64::ssub, dl, Op.getValueType(),
5133 return DAG.getTargetExtractSubreg(ARM64::dsub, dl, Op.getValueType(),
5136 llvm_unreachable("Unexpected vector type in extract_subvector!");
5139 // If this is extracting the upper 64-bits of a 128-bit vector, we match
5141 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
5147 bool ARM64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5149 if (VT.getVectorNumElements() == 4 &&
5150 (VT.is128BitVector() || VT.is64BitVector())) {
5151 unsigned PFIndexes[4];
5152 for (unsigned i = 0; i != 4; ++i) {
5156 PFIndexes[i] = M[i];
5159 // Compute the index in the perfect shuffle table.
5160 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5161 PFIndexes[2] * 9 + PFIndexes[3];
5162 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5163 unsigned Cost = (PFEntry >> 30);
5170 unsigned Imm, WhichResult;
5172 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
5173 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
5174 isEXTMask(M, VT, ReverseVEXT, Imm) ||
5175 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
5176 isTRNMask(M, VT, WhichResult) || isUZPMask(M, VT, WhichResult) ||
5177 isZIPMask(M, VT, WhichResult) ||
5178 isTRN_v_undef_Mask(M, VT, WhichResult) ||
5179 isUZP_v_undef_Mask(M, VT, WhichResult) ||
5180 isZIP_v_undef_Mask(M, VT, WhichResult));
5183 /// getVShiftImm - Check if this is a valid build_vector for the immediate
5184 /// operand of a vector shift operation, where all the elements of the
5185 /// build_vector must have the same constant integer value.
5186 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
5187 // Ignore bit_converts.
5188 while (Op.getOpcode() == ISD::BITCAST)
5189 Op = Op.getOperand(0);
5190 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
5191 APInt SplatBits, SplatUndef;
5192 unsigned SplatBitSize;
5194 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
5195 HasAnyUndefs, ElementBits) ||
5196 SplatBitSize > ElementBits)
5198 Cnt = SplatBits.getSExtValue();
5202 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
5203 /// operand of a vector shift left operation. That value must be in the range:
5204 /// 0 <= Value < ElementBits for a left shift; or
5205 /// 0 <= Value <= ElementBits for a long left shift.
5206 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
5207 assert(VT.isVector() && "vector shift count is not a vector type");
5208 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5209 if (!getVShiftImm(Op, ElementBits, Cnt))
5211 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
5214 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
5215 /// operand of a vector shift right operation. For a shift opcode, the value
5216 /// is positive, but for an intrinsic the value count must be negative. The
5217 /// absolute value must be in the range:
5218 /// 1 <= |Value| <= ElementBits for a right shift; or
5219 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
5220 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
5222 assert(VT.isVector() && "vector shift count is not a vector type");
5223 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5224 if (!getVShiftImm(Op, ElementBits, Cnt))
5228 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
5231 SDValue ARM64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
5232 SelectionDAG &DAG) const {
5233 EVT VT = Op.getValueType();
5237 if (!Op.getOperand(1).getValueType().isVector())
5239 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5241 switch (Op.getOpcode()) {
5243 llvm_unreachable("unexpected shift opcode");
5246 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
5247 return DAG.getNode(ARM64ISD::VSHL, SDLoc(Op), VT, Op.getOperand(0),
5248 DAG.getConstant(Cnt, MVT::i32));
5249 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5250 DAG.getConstant(Intrinsic::arm64_neon_ushl, MVT::i32),
5251 Op.getOperand(0), Op.getOperand(1));
5254 // Right shift immediate
5255 if (isVShiftRImm(Op.getOperand(1), VT, false, false, Cnt) &&
5258 (Op.getOpcode() == ISD::SRA) ? ARM64ISD::VASHR : ARM64ISD::VLSHR;
5259 return DAG.getNode(Opc, SDLoc(Op), VT, Op.getOperand(0),
5260 DAG.getConstant(Cnt, MVT::i32));
5263 // Right shift register. Note, there is not a shift right register
5264 // instruction, but the shift left register instruction takes a signed
5265 // value, where negative numbers specify a right shift.
5266 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::arm64_neon_sshl
5267 : Intrinsic::arm64_neon_ushl;
5268 // negate the shift amount
5269 SDValue NegShift = DAG.getNode(ARM64ISD::NEG, DL, VT, Op.getOperand(1));
5270 SDValue NegShiftLeft =
5271 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5272 DAG.getConstant(Opc, MVT::i32), Op.getOperand(0), NegShift);
5273 return NegShiftLeft;
5279 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
5280 ARM64CC::CondCode CC, bool NoNans, EVT VT,
5281 SDLoc dl, SelectionDAG &DAG) {
5282 EVT SrcVT = LHS.getValueType();
5284 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
5285 APInt CnstBits(VT.getSizeInBits(), 0);
5286 APInt UndefBits(VT.getSizeInBits(), 0);
5287 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
5288 bool IsZero = IsCnst && (CnstBits == 0);
5290 if (SrcVT.getVectorElementType().isFloatingPoint()) {
5297 Fcmeq = DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5299 Fcmeq = DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5300 return DAG.getNode(ARM64ISD::NOT, dl, VT, Fcmeq);
5304 return DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5305 return DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5308 return DAG.getNode(ARM64ISD::FCMGEz, dl, VT, LHS);
5309 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, LHS, RHS);
5312 return DAG.getNode(ARM64ISD::FCMGTz, dl, VT, LHS);
5313 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, LHS, RHS);
5316 return DAG.getNode(ARM64ISD::FCMLEz, dl, VT, LHS);
5317 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, RHS, LHS);
5321 // If we ignore NaNs then we can use to the MI implementation.
5325 return DAG.getNode(ARM64ISD::FCMLTz, dl, VT, LHS);
5326 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, RHS, LHS);
5336 Cmeq = DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5338 Cmeq = DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5339 return DAG.getNode(ARM64ISD::NOT, dl, VT, Cmeq);
5343 return DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5344 return DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5347 return DAG.getNode(ARM64ISD::CMGEz, dl, VT, LHS);
5348 return DAG.getNode(ARM64ISD::CMGE, dl, VT, LHS, RHS);
5351 return DAG.getNode(ARM64ISD::CMGTz, dl, VT, LHS);
5352 return DAG.getNode(ARM64ISD::CMGT, dl, VT, LHS, RHS);
5355 return DAG.getNode(ARM64ISD::CMLEz, dl, VT, LHS);
5356 return DAG.getNode(ARM64ISD::CMGE, dl, VT, RHS, LHS);
5358 return DAG.getNode(ARM64ISD::CMHS, dl, VT, RHS, LHS);
5360 return DAG.getNode(ARM64ISD::CMHI, dl, VT, RHS, LHS);
5363 return DAG.getNode(ARM64ISD::CMLTz, dl, VT, LHS);
5364 return DAG.getNode(ARM64ISD::CMGT, dl, VT, RHS, LHS);
5366 return DAG.getNode(ARM64ISD::CMHI, dl, VT, LHS, RHS);
5368 return DAG.getNode(ARM64ISD::CMHS, dl, VT, LHS, RHS);
5372 SDValue ARM64TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
5373 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5374 SDValue LHS = Op.getOperand(0);
5375 SDValue RHS = Op.getOperand(1);
5378 if (LHS.getValueType().getVectorElementType().isInteger()) {
5379 assert(LHS.getValueType() == RHS.getValueType());
5380 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
5381 return EmitVectorComparison(LHS, RHS, ARM64CC, false, Op.getValueType(), dl,
5385 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
5386 LHS.getValueType().getVectorElementType() == MVT::f64);
5388 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
5389 // clean. Some of them require two branches to implement.
5390 ARM64CC::CondCode CC1, CC2;
5391 changeFPCCToARM64CC(CC, CC1, CC2);
5393 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
5395 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, Op.getValueType(), dl, DAG);
5396 if (!Cmp1.getNode())
5399 if (CC2 != ARM64CC::AL) {
5401 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, Op.getValueType(), dl, DAG);
5402 if (!Cmp2.getNode())
5405 return DAG.getNode(ISD::OR, dl, Cmp1.getValueType(), Cmp1, Cmp2);
5411 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
5412 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
5413 /// specified in the intrinsic calls.
5414 bool ARM64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5416 unsigned Intrinsic) const {
5417 switch (Intrinsic) {
5418 case Intrinsic::arm64_neon_ld2:
5419 case Intrinsic::arm64_neon_ld3:
5420 case Intrinsic::arm64_neon_ld4:
5421 case Intrinsic::arm64_neon_ld2lane:
5422 case Intrinsic::arm64_neon_ld3lane:
5423 case Intrinsic::arm64_neon_ld4lane:
5424 case Intrinsic::arm64_neon_ld2r:
5425 case Intrinsic::arm64_neon_ld3r:
5426 case Intrinsic::arm64_neon_ld4r: {
5427 Info.opc = ISD::INTRINSIC_W_CHAIN;
5428 // Conservatively set memVT to the entire set of vectors loaded.
5429 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
5430 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5431 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5434 Info.vol = false; // volatile loads with NEON intrinsics not supported
5435 Info.readMem = true;
5436 Info.writeMem = false;
5439 case Intrinsic::arm64_neon_st2:
5440 case Intrinsic::arm64_neon_st3:
5441 case Intrinsic::arm64_neon_st4:
5442 case Intrinsic::arm64_neon_st2lane:
5443 case Intrinsic::arm64_neon_st3lane:
5444 case Intrinsic::arm64_neon_st4lane: {
5445 Info.opc = ISD::INTRINSIC_VOID;
5446 // Conservatively set memVT to the entire set of vectors stored.
5447 unsigned NumElts = 0;
5448 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5449 Type *ArgTy = I.getArgOperand(ArgI)->getType();
5450 if (!ArgTy->isVectorTy())
5452 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
5454 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5455 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5458 Info.vol = false; // volatile stores with NEON intrinsics not supported
5459 Info.readMem = false;
5460 Info.writeMem = true;
5463 case Intrinsic::arm64_ldaxr:
5464 case Intrinsic::arm64_ldxr: {
5465 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
5466 Info.opc = ISD::INTRINSIC_W_CHAIN;
5467 Info.memVT = MVT::getVT(PtrTy->getElementType());
5468 Info.ptrVal = I.getArgOperand(0);
5470 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
5472 Info.readMem = true;
5473 Info.writeMem = false;
5476 case Intrinsic::arm64_stlxr:
5477 case Intrinsic::arm64_stxr: {
5478 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
5479 Info.opc = ISD::INTRINSIC_W_CHAIN;
5480 Info.memVT = MVT::getVT(PtrTy->getElementType());
5481 Info.ptrVal = I.getArgOperand(1);
5483 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
5485 Info.readMem = false;
5486 Info.writeMem = true;
5489 case Intrinsic::arm64_ldaxp:
5490 case Intrinsic::arm64_ldxp: {
5491 Info.opc = ISD::INTRINSIC_W_CHAIN;
5492 Info.memVT = MVT::i128;
5493 Info.ptrVal = I.getArgOperand(0);
5497 Info.readMem = true;
5498 Info.writeMem = false;
5501 case Intrinsic::arm64_stlxp:
5502 case Intrinsic::arm64_stxp: {
5503 Info.opc = ISD::INTRINSIC_W_CHAIN;
5504 Info.memVT = MVT::i128;
5505 Info.ptrVal = I.getArgOperand(2);
5509 Info.readMem = false;
5510 Info.writeMem = true;
5520 // Truncations from 64-bit GPR to 32-bit GPR is free.
5521 bool ARM64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
5522 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
5524 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5525 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5526 if (NumBits1 <= NumBits2)
5530 bool ARM64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
5531 if (!VT1.isInteger() || !VT2.isInteger())
5533 unsigned NumBits1 = VT1.getSizeInBits();
5534 unsigned NumBits2 = VT2.getSizeInBits();
5535 if (NumBits1 <= NumBits2)
5540 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
5542 bool ARM64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
5543 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
5545 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5546 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5547 if (NumBits1 == 32 && NumBits2 == 64)
5551 bool ARM64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
5552 if (!VT1.isInteger() || !VT2.isInteger())
5554 unsigned NumBits1 = VT1.getSizeInBits();
5555 unsigned NumBits2 = VT2.getSizeInBits();
5556 if (NumBits1 == 32 && NumBits2 == 64)
5561 bool ARM64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
5562 EVT VT1 = Val.getValueType();
5563 if (isZExtFree(VT1, VT2)) {
5567 if (Val.getOpcode() != ISD::LOAD)
5570 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
5571 return (VT1.isSimple() && VT1.isInteger() && VT2.isSimple() &&
5572 VT2.isInteger() && VT1.getSizeInBits() <= 32);
5575 bool ARM64TargetLowering::hasPairedLoad(Type *LoadedType,
5576 unsigned &RequiredAligment) const {
5577 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
5579 // Cyclone supports unaligned accesses.
5580 RequiredAligment = 0;
5581 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
5582 return NumBits == 32 || NumBits == 64;
5585 bool ARM64TargetLowering::hasPairedLoad(EVT LoadedType,
5586 unsigned &RequiredAligment) const {
5587 if (!LoadedType.isSimple() ||
5588 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
5590 // Cyclone supports unaligned accesses.
5591 RequiredAligment = 0;
5592 unsigned NumBits = LoadedType.getSizeInBits();
5593 return NumBits == 32 || NumBits == 64;
5596 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
5597 unsigned AlignCheck) {
5598 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
5599 (DstAlign == 0 || DstAlign % AlignCheck == 0));
5602 EVT ARM64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
5603 unsigned SrcAlign, bool IsMemset,
5604 bool ZeroMemset, bool MemcpyStrSrc,
5605 MachineFunction &MF) const {
5606 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
5607 // instruction to materialize the v2i64 zero and one store (with restrictive
5608 // addressing mode). Just do two i64 store of zero-registers.
5610 const Function *F = MF.getFunction();
5611 if (!IsMemset && Size >= 16 &&
5612 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
5613 Attribute::NoImplicitFloat) &&
5614 (memOpAlign(SrcAlign, DstAlign, 16) ||
5615 (allowsUnalignedMemoryAccesses(MVT::v2i64, 0, &Fast) && Fast)))
5618 return Size >= 8 ? MVT::i64 : MVT::i32;
5621 // 12-bit optionally shifted immediates are legal for adds.
5622 bool ARM64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
5623 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
5628 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
5629 // immediates is the same as for an add or a sub.
5630 bool ARM64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
5633 return isLegalAddImmediate(Immed);
5636 /// isLegalAddressingMode - Return true if the addressing mode represented
5637 /// by AM is legal for this target, for a load/store of the specified type.
5638 bool ARM64TargetLowering::isLegalAddressingMode(const AddrMode &AM,
5640 // ARM64 has five basic addressing modes:
5642 // reg + 9-bit signed offset
5643 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
5645 // reg + SIZE_IN_BYTES * reg
5647 // No global is ever allowed as a base.
5651 // No reg+reg+imm addressing.
5652 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
5655 // check reg + imm case:
5656 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
5657 uint64_t NumBytes = 0;
5658 if (Ty->isSized()) {
5659 uint64_t NumBits = getDataLayout()->getTypeSizeInBits(Ty);
5660 NumBytes = NumBits / 8;
5661 if (!isPowerOf2_64(NumBits))
5666 int64_t Offset = AM.BaseOffs;
5668 // 9-bit signed offset
5669 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
5672 // 12-bit unsigned offset
5673 unsigned shift = Log2_64(NumBytes);
5674 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
5675 // Must be a multiple of NumBytes (NumBytes is a power of 2)
5676 (Offset >> shift) << shift == Offset)
5681 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
5683 if (!AM.Scale || AM.Scale == 1 ||
5684 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
5689 int ARM64TargetLowering::getScalingFactorCost(const AddrMode &AM,
5691 // Scaling factors are not free at all.
5692 // Operands | Rt Latency
5693 // -------------------------------------------
5695 // -------------------------------------------
5696 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
5697 // Rt, [Xn, Wm, <extend> #imm] |
5698 if (isLegalAddressingMode(AM, Ty))
5699 // Scale represents reg2 * scale, thus account for 1 if
5700 // it is not equal to 0 or 1.
5701 return AM.Scale != 0 && AM.Scale != 1;
5705 bool ARM64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
5706 VT = VT.getScalarType();
5711 switch (VT.getSimpleVT().SimpleTy) {
5723 ARM64TargetLowering::getScratchRegisters(CallingConv::ID) const {
5724 // LR is a callee-save register, but we must treat it as clobbered by any call
5725 // site. Hence we include LR in the scratch registers, which are in turn added
5726 // as implicit-defs for stackmaps and patchpoints.
5727 static const MCPhysReg ScratchRegs[] = {
5728 ARM64::X16, ARM64::X17, ARM64::LR, 0
5733 bool ARM64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
5735 assert(Ty->isIntegerTy());
5737 unsigned BitSize = Ty->getPrimitiveSizeInBits();
5741 int64_t Val = Imm.getSExtValue();
5742 if (Val == 0 || ARM64_AM::isLogicalImmediate(Val, BitSize))
5745 if ((int64_t)Val < 0)
5748 Val &= (1LL << 32) - 1;
5750 unsigned LZ = countLeadingZeros((uint64_t)Val);
5751 unsigned Shift = (63 - LZ) / 16;
5752 // MOVZ is free so return true for one or fewer MOVK.
5753 return (Shift < 3) ? true : false;
5756 // Generate SUBS and CSEL for integer abs.
5757 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
5758 EVT VT = N->getValueType(0);
5760 SDValue N0 = N->getOperand(0);
5761 SDValue N1 = N->getOperand(1);
5764 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
5765 // and change it to SUB and CSEL.
5766 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
5767 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
5768 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
5769 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
5770 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
5771 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
5773 // Generate SUBS & CSEL.
5775 DAG.getNode(ARM64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
5776 N0.getOperand(0), DAG.getConstant(0, VT));
5777 return DAG.getNode(ARM64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
5778 DAG.getConstant(ARM64CC::PL, MVT::i32),
5779 SDValue(Cmp.getNode(), 1));
5784 // performXorCombine - Attempts to handle integer ABS.
5785 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
5786 TargetLowering::DAGCombinerInfo &DCI,
5787 const ARM64Subtarget *Subtarget) {
5788 if (DCI.isBeforeLegalizeOps())
5791 return performIntegerAbsCombine(N, DAG);
5794 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
5795 TargetLowering::DAGCombinerInfo &DCI,
5796 const ARM64Subtarget *Subtarget) {
5797 if (DCI.isBeforeLegalizeOps())
5800 // Multiplication of a power of two plus/minus one can be done more
5801 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
5802 // future CPUs have a cheaper MADD instruction, this may need to be
5803 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
5804 // 64-bit is 5 cycles, so this is always a win.
5805 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
5806 APInt Value = C->getAPIntValue();
5807 EVT VT = N->getValueType(0);
5808 APInt VP1 = Value + 1;
5809 if (VP1.isPowerOf2()) {
5810 // Multiplying by one less than a power of two, replace with a shift
5812 SDValue ShiftedVal =
5813 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
5814 DAG.getConstant(VP1.logBase2(), MVT::i64));
5815 return DAG.getNode(ISD::SUB, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
5817 APInt VM1 = Value - 1;
5818 if (VM1.isPowerOf2()) {
5819 // Multiplying by one more than a power of two, replace with a shift
5821 SDValue ShiftedVal =
5822 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
5823 DAG.getConstant(VM1.logBase2(), MVT::i64));
5824 return DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
5830 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG) {
5831 EVT VT = N->getValueType(0);
5832 if (VT != MVT::f32 && VT != MVT::f64)
5834 // Only optimize when the source and destination types have the same width.
5835 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
5838 // If the result of an integer load is only used by an integer-to-float
5839 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
5840 // This eliminates an "integer-to-vector-move UOP and improve throughput.
5841 SDValue N0 = N->getOperand(0);
5842 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
5843 // Do not change the width of a volatile load.
5844 !cast<LoadSDNode>(N0)->isVolatile()) {
5845 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5846 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
5847 LN0->getPointerInfo(), LN0->isVolatile(),
5848 LN0->isNonTemporal(), LN0->isInvariant(),
5849 LN0->getAlignment());
5851 // Make sure successors of the original load stay after it by updating them
5852 // to use the new Chain.
5853 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
5856 (N->getOpcode() == ISD::SINT_TO_FP) ? ARM64ISD::SITOF : ARM64ISD::UITOF;
5857 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
5863 /// An EXTR instruction is made up of two shifts, ORed together. This helper
5864 /// searches for and classifies those shifts.
5865 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
5867 if (N.getOpcode() == ISD::SHL)
5869 else if (N.getOpcode() == ISD::SRL)
5874 if (!isa<ConstantSDNode>(N.getOperand(1)))
5877 ShiftAmount = N->getConstantOperandVal(1);
5878 Src = N->getOperand(0);
5882 /// EXTR instruction extracts a contiguous chunk of bits from two existing
5883 /// registers viewed as a high/low pair. This function looks for the pattern:
5884 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
5885 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
5887 static SDValue tryCombineToEXTR(SDNode *N,
5888 TargetLowering::DAGCombinerInfo &DCI) {
5889 SelectionDAG &DAG = DCI.DAG;
5891 EVT VT = N->getValueType(0);
5893 assert(N->getOpcode() == ISD::OR && "Unexpected root");
5895 if (VT != MVT::i32 && VT != MVT::i64)
5899 uint32_t ShiftLHS = 0;
5901 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
5905 uint32_t ShiftRHS = 0;
5907 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
5910 // If they're both trying to come from the high part of the register, they're
5911 // not really an EXTR.
5912 if (LHSFromHi == RHSFromHi)
5915 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
5919 std::swap(LHS, RHS);
5920 std::swap(ShiftLHS, ShiftRHS);
5923 return DAG.getNode(ARM64ISD::EXTR, DL, VT, LHS, RHS,
5924 DAG.getConstant(ShiftRHS, MVT::i64));
5927 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
5928 const ARM64Subtarget *Subtarget) {
5929 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
5930 if (!EnableARM64ExtrGeneration)
5932 SelectionDAG &DAG = DCI.DAG;
5933 EVT VT = N->getValueType(0);
5935 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
5938 SDValue Res = tryCombineToEXTR(N, DCI);
5945 static SDValue performBitcastCombine(SDNode *N,
5946 TargetLowering::DAGCombinerInfo &DCI,
5947 SelectionDAG &DAG) {
5948 // Wait 'til after everything is legalized to try this. That way we have
5949 // legal vector types and such.
5950 if (DCI.isBeforeLegalizeOps())
5953 // Remove extraneous bitcasts around an extract_subvector.
5955 // (v4i16 (bitconvert
5956 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
5958 // (extract_subvector ((v8i16 ...), (i64 4)))
5960 // Only interested in 64-bit vectors as the ultimate result.
5961 EVT VT = N->getValueType(0);
5964 if (VT.getSimpleVT().getSizeInBits() != 64)
5966 // Is the operand an extract_subvector starting at the beginning or halfway
5967 // point of the vector? A low half may also come through as an
5968 // EXTRACT_SUBREG, so look for that, too.
5969 SDValue Op0 = N->getOperand(0);
5970 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
5971 !(Op0->isMachineOpcode() &&
5972 Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG))
5974 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
5975 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
5976 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
5978 } else if (Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG) {
5979 if (idx != ARM64::dsub)
5981 // The dsub reference is equivalent to a lane zero subvector reference.
5984 // Look through the bitcast of the input to the extract.
5985 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
5987 SDValue Source = Op0->getOperand(0)->getOperand(0);
5988 // If the source type has twice the number of elements as our destination
5989 // type, we know this is an extract of the high or low half of the vector.
5990 EVT SVT = Source->getValueType(0);
5991 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
5994 DEBUG(dbgs() << "arm64-lower: bitcast extract_subvector simplification\n");
5996 // Create the simplified form to just extract the low or high half of the
5997 // vector directly rather than bothering with the bitcasts.
5999 unsigned NumElements = VT.getVectorNumElements();
6001 SDValue HalfIdx = DAG.getConstant(NumElements, MVT::i64);
6002 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
6004 SDValue SubReg = DAG.getTargetConstant(ARM64::dsub, MVT::i32);
6005 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
6011 static SDValue performConcatVectorsCombine(SDNode *N,
6012 TargetLowering::DAGCombinerInfo &DCI,
6013 SelectionDAG &DAG) {
6014 // Wait 'til after everything is legalized to try this. That way we have
6015 // legal vector types and such.
6016 if (DCI.isBeforeLegalizeOps())
6020 EVT VT = N->getValueType(0);
6022 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
6023 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
6024 // canonicalise to that.
6025 if (N->getOperand(0) == N->getOperand(1) && VT.getVectorNumElements() == 2) {
6026 assert(VT.getVectorElementType().getSizeInBits() == 64);
6027 return DAG.getNode(ARM64ISD::DUPLANE64, dl, VT,
6028 WidenVector(N->getOperand(0), DAG),
6029 DAG.getConstant(0, MVT::i64));
6032 // Canonicalise concat_vectors so that the right-hand vector has as few
6033 // bit-casts as possible before its real operation. The primary matching
6034 // destination for these operations will be the narrowing "2" instructions,
6035 // which depend on the operation being performed on this right-hand vector.
6037 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
6039 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
6041 SDValue Op1 = N->getOperand(1);
6042 if (Op1->getOpcode() != ISD::BITCAST)
6044 SDValue RHS = Op1->getOperand(0);
6045 MVT RHSTy = RHS.getValueType().getSimpleVT();
6046 // If the RHS is not a vector, this is not the pattern we're looking for.
6047 if (!RHSTy.isVector())
6050 DEBUG(dbgs() << "arm64-lower: concat_vectors bitcast simplification\n");
6052 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
6053 RHSTy.getVectorNumElements() * 2);
6055 ISD::BITCAST, dl, VT,
6056 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
6057 DAG.getNode(ISD::BITCAST, dl, RHSTy, N->getOperand(0)), RHS));
6060 static SDValue tryCombineFixedPointConvert(SDNode *N,
6061 TargetLowering::DAGCombinerInfo &DCI,
6062 SelectionDAG &DAG) {
6063 // Wait 'til after everything is legalized to try this. That way we have
6064 // legal vector types and such.
6065 if (DCI.isBeforeLegalizeOps())
6067 // Transform a scalar conversion of a value from a lane extract into a
6068 // lane extract of a vector conversion. E.g., from foo1 to foo2:
6069 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
6070 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
6072 // The second form interacts better with instruction selection and the
6073 // register allocator to avoid cross-class register copies that aren't
6074 // coalescable due to a lane reference.
6076 // Check the operand and see if it originates from a lane extract.
6077 SDValue Op1 = N->getOperand(1);
6078 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6079 // Yep, no additional predication needed. Perform the transform.
6080 SDValue IID = N->getOperand(0);
6081 SDValue Shift = N->getOperand(2);
6082 SDValue Vec = Op1.getOperand(0);
6083 SDValue Lane = Op1.getOperand(1);
6084 EVT ResTy = N->getValueType(0);
6088 // The vector width should be 128 bits by the time we get here, even
6089 // if it started as 64 bits (the extract_vector handling will have
6091 assert(Vec.getValueType().getSizeInBits() == 128 &&
6092 "unexpected vector size on extract_vector_elt!");
6093 if (Vec.getValueType() == MVT::v4i32)
6094 VecResTy = MVT::v4f32;
6095 else if (Vec.getValueType() == MVT::v2i64)
6096 VecResTy = MVT::v2f64;
6098 assert(0 && "unexpected vector type!");
6101 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
6102 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
6107 // AArch64 high-vector "long" operations are formed by performing the non-high
6108 // version on an extract_subvector of each operand which gets the high half:
6110 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
6112 // However, there are cases which don't have an extract_high explicitly, but
6113 // have another operation that can be made compatible with one for free. For
6116 // (dupv64 scalar) --> (extract_high (dup128 scalar))
6118 // This routine does the actual conversion of such DUPs, once outer routines
6119 // have determined that everything else is in order.
6120 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
6121 // We can handle most types of duplicate, but the lane ones have an extra
6122 // operand saying *which* lane, so we need to know.
6124 switch (N.getOpcode()) {
6128 case ARM64ISD::DUPLANE8:
6129 case ARM64ISD::DUPLANE16:
6130 case ARM64ISD::DUPLANE32:
6131 case ARM64ISD::DUPLANE64:
6138 MVT NarrowTy = N.getSimpleValueType();
6139 if (!NarrowTy.is64BitVector())
6142 MVT ElementTy = NarrowTy.getVectorElementType();
6143 unsigned NumElems = NarrowTy.getVectorNumElements();
6144 MVT NewDUPVT = MVT::getVectorVT(ElementTy, NumElems * 2);
6148 NewDUP = DAG.getNode(N.getOpcode(), SDLoc(N), NewDUPVT, N.getOperand(0),
6151 NewDUP = DAG.getNode(ARM64ISD::DUP, SDLoc(N), NewDUPVT, N.getOperand(0));
6153 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N.getNode()), NarrowTy,
6154 NewDUP, DAG.getConstant(NumElems, MVT::i64));
6157 static bool isEssentiallyExtractSubvector(SDValue N) {
6158 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
6161 return N.getOpcode() == ISD::BITCAST &&
6162 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
6165 /// \brief Helper structure to keep track of ISD::SET_CC operands.
6166 struct GenericSetCCInfo {
6167 const SDValue *Opnd0;
6168 const SDValue *Opnd1;
6172 /// \brief Helper structure to keep track of a SET_CC lowered into ARM64 code.
6173 struct ARM64SetCCInfo {
6175 ARM64CC::CondCode CC;
6178 /// \brief Helper structure to keep track of SetCC information.
6180 GenericSetCCInfo Generic;
6181 ARM64SetCCInfo ARM64;
6184 /// \brief Helper structure to be able to read SetCC information.
6185 /// If set to true, IsARM64 field, Info is a ARM64SetCCInfo, otherwise Info is
6186 /// a GenericSetCCInfo.
6187 struct SetCCInfoAndKind {
6192 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
6194 /// ARM64 lowered one.
6195 /// \p SetCCInfo is filled accordingly.
6196 /// \post SetCCInfo is meanginfull only when this function returns true.
6197 /// \return True when Op is a kind of SET_CC operation.
6198 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
6199 // If this is a setcc, this is straight forward.
6200 if (Op.getOpcode() == ISD::SETCC) {
6201 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
6202 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
6203 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6204 SetCCInfo.IsARM64 = false;
6207 // Otherwise, check if this is a matching csel instruction.
6211 if (Op.getOpcode() != ARM64ISD::CSEL)
6213 // Set the information about the operands.
6214 // TODO: we want the operands of the Cmp not the csel
6215 SetCCInfo.Info.ARM64.Cmp = &Op.getOperand(3);
6216 SetCCInfo.IsARM64 = true;
6217 SetCCInfo.Info.ARM64.CC = static_cast<ARM64CC::CondCode>(
6218 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
6220 // Check that the operands matches the constraints:
6221 // (1) Both operands must be constants.
6222 // (2) One must be 1 and the other must be 0.
6223 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
6224 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6227 if (!TValue || !FValue)
6231 if (!TValue->isOne()) {
6232 // Update the comparison when we are interested in !cc.
6233 std::swap(TValue, FValue);
6234 SetCCInfo.Info.ARM64.CC =
6235 ARM64CC::getInvertedCondCode(SetCCInfo.Info.ARM64.CC);
6237 return TValue->isOne() && FValue->isNullValue();
6240 // The folding we want to perform is:
6241 // (add x, (setcc cc ...) )
6243 // (csel x, (add x, 1), !cc ...)
6245 // The latter will get matched to a CSINC instruction.
6246 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
6247 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
6248 SDValue LHS = Op->getOperand(0);
6249 SDValue RHS = Op->getOperand(1);
6250 SetCCInfoAndKind InfoAndKind;
6252 // If neither operand is a SET_CC, give up.
6253 if (!isSetCC(LHS, InfoAndKind)) {
6254 std::swap(LHS, RHS);
6255 if (!isSetCC(LHS, InfoAndKind))
6259 // FIXME: This could be generatized to work for FP comparisons.
6260 EVT CmpVT = InfoAndKind.IsARM64
6261 ? InfoAndKind.Info.ARM64.Cmp->getOperand(0).getValueType()
6262 : InfoAndKind.Info.Generic.Opnd0->getValueType();
6263 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
6269 if (InfoAndKind.IsARM64) {
6270 CCVal = DAG.getConstant(
6271 ARM64CC::getInvertedCondCode(InfoAndKind.Info.ARM64.CC), MVT::i32);
6272 Cmp = *InfoAndKind.Info.ARM64.Cmp;
6274 Cmp = getARM64Cmp(*InfoAndKind.Info.Generic.Opnd0,
6275 *InfoAndKind.Info.Generic.Opnd1,
6276 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
6279 EVT VT = Op->getValueType(0);
6280 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, VT));
6281 return DAG.getNode(ARM64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
6284 // The basic add/sub long vector instructions have variants with "2" on the end
6285 // which act on the high-half of their inputs. They are normally matched by
6288 // (add (zeroext (extract_high LHS)),
6289 // (zeroext (extract_high RHS)))
6290 // -> uaddl2 vD, vN, vM
6292 // However, if one of the extracts is something like a duplicate, this
6293 // instruction can still be used profitably. This function puts the DAG into a
6294 // more appropriate form for those patterns to trigger.
6295 static SDValue performAddSubLongCombine(SDNode *N,
6296 TargetLowering::DAGCombinerInfo &DCI,
6297 SelectionDAG &DAG) {
6298 if (DCI.isBeforeLegalizeOps())
6301 MVT VT = N->getSimpleValueType(0);
6302 if (!VT.is128BitVector()) {
6303 if (N->getOpcode() == ISD::ADD)
6304 return performSetccAddFolding(N, DAG);
6308 // Make sure both branches are extended in the same way.
6309 SDValue LHS = N->getOperand(0);
6310 SDValue RHS = N->getOperand(1);
6311 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
6312 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
6313 LHS.getOpcode() != RHS.getOpcode())
6316 unsigned ExtType = LHS.getOpcode();
6318 // It's not worth doing if at least one of the inputs isn't already an
6319 // extract, but we don't know which it'll be so we have to try both.
6320 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
6321 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
6325 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
6326 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
6327 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
6331 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
6334 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
6337 // Massage DAGs which we can use the high-half "long" operations on into
6338 // something isel will recognize better. E.g.
6340 // (arm64_neon_umull (extract_high vec) (dupv64 scalar)) -->
6341 // (arm64_neon_umull (extract_high (v2i64 vec)))
6342 // (extract_high (v2i64 (dup128 scalar)))))
6344 static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N,
6345 TargetLowering::DAGCombinerInfo &DCI,
6346 SelectionDAG &DAG) {
6347 if (DCI.isBeforeLegalizeOps())
6350 SDValue LHS = N->getOperand(1);
6351 SDValue RHS = N->getOperand(2);
6352 assert(LHS.getValueType().is64BitVector() &&
6353 RHS.getValueType().is64BitVector() &&
6354 "unexpected shape for long operation");
6356 // Either node could be a DUP, but it's not worth doing both of them (you'd
6357 // just as well use the non-high version) so look for a corresponding extract
6358 // operation on the other "wing".
6359 if (isEssentiallyExtractSubvector(LHS)) {
6360 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
6363 } else if (isEssentiallyExtractSubvector(RHS)) {
6364 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
6369 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
6370 N->getOperand(0), LHS, RHS);
6373 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
6374 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
6375 unsigned ElemBits = ElemTy.getSizeInBits();
6377 int64_t ShiftAmount;
6378 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
6379 APInt SplatValue, SplatUndef;
6380 unsigned SplatBitSize;
6382 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
6383 HasAnyUndefs, ElemBits) ||
6384 SplatBitSize != ElemBits)
6387 ShiftAmount = SplatValue.getSExtValue();
6388 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
6389 ShiftAmount = CVN->getSExtValue();
6397 llvm_unreachable("Unknown shift intrinsic");
6398 case Intrinsic::arm64_neon_sqshl:
6399 Opcode = ARM64ISD::SQSHL_I;
6400 IsRightShift = false;
6402 case Intrinsic::arm64_neon_uqshl:
6403 Opcode = ARM64ISD::UQSHL_I;
6404 IsRightShift = false;
6406 case Intrinsic::arm64_neon_srshl:
6407 Opcode = ARM64ISD::SRSHR_I;
6408 IsRightShift = true;
6410 case Intrinsic::arm64_neon_urshl:
6411 Opcode = ARM64ISD::URSHR_I;
6412 IsRightShift = true;
6414 case Intrinsic::arm64_neon_sqshlu:
6415 Opcode = ARM64ISD::SQSHLU_I;
6416 IsRightShift = false;
6420 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits)
6421 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6422 DAG.getConstant(-ShiftAmount, MVT::i32));
6423 else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount <= ElemBits)
6424 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6425 DAG.getConstant(ShiftAmount, MVT::i32));
6430 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
6431 // the intrinsics must be legal and take an i32, this means there's almost
6432 // certainly going to be a zext in the DAG which we can eliminate.
6433 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
6434 SDValue AndN = N->getOperand(2);
6435 if (AndN.getOpcode() != ISD::AND)
6438 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
6439 if (!CMask || CMask->getZExtValue() != Mask)
6442 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
6443 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
6446 static SDValue performIntrinsicCombine(SDNode *N,
6447 TargetLowering::DAGCombinerInfo &DCI,
6448 const ARM64Subtarget *Subtarget) {
6449 SelectionDAG &DAG = DCI.DAG;
6450 unsigned IID = getIntrinsicID(N);
6454 case Intrinsic::arm64_neon_vcvtfxs2fp:
6455 case Intrinsic::arm64_neon_vcvtfxu2fp:
6456 return tryCombineFixedPointConvert(N, DCI, DAG);
6458 case Intrinsic::arm64_neon_fmax:
6459 return DAG.getNode(ARM64ISD::FMAX, SDLoc(N), N->getValueType(0),
6460 N->getOperand(1), N->getOperand(2));
6461 case Intrinsic::arm64_neon_fmin:
6462 return DAG.getNode(ARM64ISD::FMIN, SDLoc(N), N->getValueType(0),
6463 N->getOperand(1), N->getOperand(2));
6464 case Intrinsic::arm64_neon_smull:
6465 case Intrinsic::arm64_neon_umull:
6466 case Intrinsic::arm64_neon_pmull:
6467 case Intrinsic::arm64_neon_sqdmull:
6468 return tryCombineLongOpWithDup(IID, N, DCI, DAG);
6469 case Intrinsic::arm64_neon_sqshl:
6470 case Intrinsic::arm64_neon_uqshl:
6471 case Intrinsic::arm64_neon_sqshlu:
6472 case Intrinsic::arm64_neon_srshl:
6473 case Intrinsic::arm64_neon_urshl:
6474 return tryCombineShiftImm(IID, N, DAG);
6475 case Intrinsic::arm64_crc32b:
6476 case Intrinsic::arm64_crc32cb:
6477 return tryCombineCRC32(0xff, N, DAG);
6478 case Intrinsic::arm64_crc32h:
6479 case Intrinsic::arm64_crc32ch:
6480 return tryCombineCRC32(0xffff, N, DAG);
6485 static SDValue performExtendCombine(SDNode *N,
6486 TargetLowering::DAGCombinerInfo &DCI,
6487 SelectionDAG &DAG) {
6488 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
6489 // we can convert that DUP into another extract_high (of a bigger DUP), which
6490 // helps the backend to decide that an sabdl2 would be useful, saving a real
6491 // extract_high operation.
6492 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
6493 N->getOperand(0).getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
6494 SDNode *ABDNode = N->getOperand(0).getNode();
6495 unsigned IID = getIntrinsicID(ABDNode);
6496 if (IID == Intrinsic::arm64_neon_sabd ||
6497 IID == Intrinsic::arm64_neon_uabd) {
6498 SDValue NewABD = tryCombineLongOpWithDup(IID, ABDNode, DCI, DAG);
6499 if (!NewABD.getNode())
6502 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
6507 // This is effectively a custom type legalization for ARM64.
6509 // Type legalization will split an extend of a small, legal, type to a larger
6510 // illegal type by first splitting the destination type, often creating
6511 // illegal source types, which then get legalized in isel-confusing ways,
6512 // leading to really terrible codegen. E.g.,
6513 // %result = v8i32 sext v8i8 %value
6515 // %losrc = extract_subreg %value, ...
6516 // %hisrc = extract_subreg %value, ...
6517 // %lo = v4i32 sext v4i8 %losrc
6518 // %hi = v4i32 sext v4i8 %hisrc
6519 // Things go rapidly downhill from there.
6521 // For ARM64, the [sz]ext vector instructions can only go up one element
6522 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
6523 // take two instructions.
6525 // This implies that the most efficient way to do the extend from v8i8
6526 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
6527 // the normal splitting to happen for the v8i16->v8i32.
6529 // This is pre-legalization to catch some cases where the default
6530 // type legalization will create ill-tempered code.
6531 if (!DCI.isBeforeLegalizeOps())
6534 // We're only interested in cleaning things up for non-legal vector types
6535 // here. If both the source and destination are legal, things will just
6536 // work naturally without any fiddling.
6537 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6538 EVT ResVT = N->getValueType(0);
6539 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
6541 // If the vector type isn't a simple VT, it's beyond the scope of what
6542 // we're worried about here. Let legalization do its thing and hope for
6544 if (!ResVT.isSimple())
6547 SDValue Src = N->getOperand(0);
6548 MVT SrcVT = Src->getValueType(0).getSimpleVT();
6549 // If the source VT is a 64-bit vector, we can play games and get the
6550 // better results we want.
6551 if (SrcVT.getSizeInBits() != 64)
6554 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
6555 unsigned ElementCount = SrcVT.getVectorNumElements();
6556 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
6558 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
6560 // Now split the rest of the operation into two halves, each with a 64
6564 unsigned NumElements = ResVT.getVectorNumElements();
6565 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
6566 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
6567 ResVT.getVectorElementType(), NumElements / 2);
6569 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
6570 LoVT.getVectorNumElements());
6571 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
6572 DAG.getIntPtrConstant(0));
6573 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
6574 DAG.getIntPtrConstant(InNVT.getVectorNumElements()));
6575 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
6576 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
6578 // Now combine the parts back together so we still have a single result
6579 // like the combiner expects.
6580 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
6583 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
6584 /// value. The load store optimizer pass will merge them to store pair stores.
6585 /// This has better performance than a splat of the scalar followed by a split
6586 /// vector store. Even if the stores are not merged it is four stores vs a dup,
6587 /// followed by an ext.b and two stores.
6588 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
6589 SDValue StVal = St->getValue();
6590 EVT VT = StVal.getValueType();
6592 // Don't replace floating point stores, they possibly won't be transformed to
6593 // stp because of the store pair suppress pass.
6594 if (VT.isFloatingPoint())
6597 // Check for insert vector elements.
6598 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
6601 // We can express a splat as store pair(s) for 2 or 4 elements.
6602 unsigned NumVecElts = VT.getVectorNumElements();
6603 if (NumVecElts != 4 && NumVecElts != 2)
6605 SDValue SplatVal = StVal.getOperand(1);
6606 unsigned RemainInsertElts = NumVecElts - 1;
6608 // Check that this is a splat.
6609 while (--RemainInsertElts) {
6610 SDValue NextInsertElt = StVal.getOperand(0);
6611 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
6613 if (NextInsertElt.getOperand(1) != SplatVal)
6615 StVal = NextInsertElt;
6617 unsigned OrigAlignment = St->getAlignment();
6618 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
6619 unsigned Alignment = std::min(OrigAlignment, EltOffset);
6621 // Create scalar stores. This is at least as good as the code sequence for a
6622 // split unaligned store wich is a dup.s, ext.b, and two stores.
6623 // Most of the time the three stores should be replaced by store pair
6624 // instructions (stp).
6626 SDValue BasePtr = St->getBasePtr();
6628 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
6629 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
6631 unsigned Offset = EltOffset;
6632 while (--NumVecElts) {
6633 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
6634 DAG.getConstant(Offset, MVT::i64));
6635 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
6636 St->getPointerInfo(), St->isVolatile(),
6637 St->isNonTemporal(), Alignment);
6638 Offset += EltOffset;
6643 static SDValue performSTORECombine(SDNode *N,
6644 TargetLowering::DAGCombinerInfo &DCI,
6646 const ARM64Subtarget *Subtarget) {
6647 if (!DCI.isBeforeLegalize())
6650 StoreSDNode *S = cast<StoreSDNode>(N);
6651 if (S->isVolatile())
6654 // Cyclone has bad performance on unaligned 16B stores when crossing line and
6655 // page boundries. We want to split such stores.
6656 if (!Subtarget->isCyclone())
6659 // Don't split at Oz.
6660 MachineFunction &MF = DAG.getMachineFunction();
6661 bool IsMinSize = MF.getFunction()->getAttributes().hasAttribute(
6662 AttributeSet::FunctionIndex, Attribute::MinSize);
6666 SDValue StVal = S->getValue();
6667 EVT VT = StVal.getValueType();
6669 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
6670 // those up regresses performance on micro-benchmarks and olden/bh.
6671 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
6674 // Split unaligned 16B stores. They are terrible for performance.
6675 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
6676 // extensions can use this to mark that it does not want splitting to happen
6677 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
6678 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
6679 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
6680 S->getAlignment() <= 2)
6683 // If we get a splat of a scalar convert this vector store to a store of
6684 // scalars. They will be merged into store pairs thereby removing two
6686 SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S);
6687 if (ReplacedSplat != SDValue())
6688 return ReplacedSplat;
6691 unsigned NumElts = VT.getVectorNumElements() / 2;
6692 // Split VT into two.
6694 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
6695 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
6696 DAG.getIntPtrConstant(0));
6697 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
6698 DAG.getIntPtrConstant(NumElts));
6699 SDValue BasePtr = S->getBasePtr();
6701 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
6702 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
6703 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
6704 DAG.getConstant(8, MVT::i64));
6705 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
6706 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
6710 // Optimize compare with zero and branch.
6711 static SDValue performBRCONDCombine(SDNode *N,
6712 TargetLowering::DAGCombinerInfo &DCI,
6713 SelectionDAG &DAG) {
6714 SDValue Chain = N->getOperand(0);
6715 SDValue Dest = N->getOperand(1);
6716 SDValue CCVal = N->getOperand(2);
6717 SDValue Cmp = N->getOperand(3);
6719 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
6720 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
6721 if (CC != ARM64CC::EQ && CC != ARM64CC::NE)
6724 unsigned CmpOpc = Cmp.getOpcode();
6725 if (CmpOpc != ARM64ISD::ADDS && CmpOpc != ARM64ISD::SUBS)
6728 // Only attempt folding if there is only one use of the flag and no use of the
6730 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
6733 SDValue LHS = Cmp.getOperand(0);
6734 SDValue RHS = Cmp.getOperand(1);
6736 assert(LHS.getValueType() == RHS.getValueType() &&
6737 "Expected the value type to be the same for both operands!");
6738 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
6741 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
6742 std::swap(LHS, RHS);
6744 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
6747 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
6748 LHS.getOpcode() == ISD::SRL)
6751 // Fold the compare into the branch instruction.
6753 if (CC == ARM64CC::EQ)
6754 BR = DAG.getNode(ARM64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
6756 BR = DAG.getNode(ARM64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
6758 // Do not add new nodes to DAG combiner worklist.
6759 DCI.CombineTo(N, BR, false);
6764 // vselect (v1i1 setcc) ->
6765 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
6766 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
6767 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
6769 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
6770 SDValue N0 = N->getOperand(0);
6771 EVT CCVT = N0.getValueType();
6773 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
6774 CCVT.getVectorElementType() != MVT::i1)
6777 EVT ResVT = N->getValueType(0);
6778 EVT CmpVT = N0.getOperand(0).getValueType();
6779 // Only combine when the result type is of the same size as the compared
6781 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
6784 SDValue IfTrue = N->getOperand(1);
6785 SDValue IfFalse = N->getOperand(2);
6787 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
6788 N0.getOperand(0), N0.getOperand(1),
6789 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6790 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
6794 SDValue ARM64TargetLowering::PerformDAGCombine(SDNode *N,
6795 DAGCombinerInfo &DCI) const {
6796 SelectionDAG &DAG = DCI.DAG;
6797 switch (N->getOpcode()) {
6802 return performAddSubLongCombine(N, DCI, DAG);
6804 return performXorCombine(N, DAG, DCI, Subtarget);
6806 return performMulCombine(N, DAG, DCI, Subtarget);
6807 case ISD::SINT_TO_FP:
6808 case ISD::UINT_TO_FP:
6809 return performIntToFpCombine(N, DAG);
6811 return performORCombine(N, DCI, Subtarget);
6812 case ISD::INTRINSIC_WO_CHAIN:
6813 return performIntrinsicCombine(N, DCI, Subtarget);
6814 case ISD::ANY_EXTEND:
6815 case ISD::ZERO_EXTEND:
6816 case ISD::SIGN_EXTEND:
6817 return performExtendCombine(N, DCI, DAG);
6819 return performBitcastCombine(N, DCI, DAG);
6820 case ISD::CONCAT_VECTORS:
6821 return performConcatVectorsCombine(N, DCI, DAG);
6823 return performVSelectCombine(N, DCI.DAG);
6825 return performSTORECombine(N, DCI, DAG, Subtarget);
6826 case ARM64ISD::BRCOND:
6827 return performBRCONDCombine(N, DCI, DAG);
6832 // Check if the return value is used as only a return value, as otherwise
6833 // we can't perform a tail-call. In particular, we need to check for
6834 // target ISD nodes that are returns and any other "odd" constructs
6835 // that the generic analysis code won't necessarily catch.
6836 bool ARM64TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
6837 if (N->getNumValues() != 1)
6839 if (!N->hasNUsesOfValue(1, 0))
6842 SDValue TCChain = Chain;
6843 SDNode *Copy = *N->use_begin();
6844 if (Copy->getOpcode() == ISD::CopyToReg) {
6845 // If the copy has a glue operand, we conservatively assume it isn't safe to
6846 // perform a tail call.
6847 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
6850 TCChain = Copy->getOperand(0);
6851 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
6854 bool HasRet = false;
6855 for (SDNode *Node : Copy->uses()) {
6856 if (Node->getOpcode() != ARM64ISD::RET_FLAG)
6868 // Return whether the an instruction can potentially be optimized to a tail
6869 // call. This will cause the optimizers to attempt to move, or duplicate,
6870 // return instructions to help enable tail call optimizations for this
6872 bool ARM64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
6873 if (!EnableARM64TailCalls)
6876 if (!CI->isTailCall())
6882 bool ARM64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
6884 ISD::MemIndexedMode &AM,
6886 SelectionDAG &DAG) const {
6887 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
6890 Base = Op->getOperand(0);
6891 // All of the indexed addressing mode instructions take a signed
6892 // 9 bit immediate offset.
6893 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
6894 int64_t RHSC = (int64_t)RHS->getZExtValue();
6895 if (RHSC >= 256 || RHSC <= -256)
6897 IsInc = (Op->getOpcode() == ISD::ADD);
6898 Offset = Op->getOperand(1);
6904 bool ARM64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
6906 ISD::MemIndexedMode &AM,
6907 SelectionDAG &DAG) const {
6910 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6911 VT = LD->getMemoryVT();
6912 Ptr = LD->getBasePtr();
6913 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6914 VT = ST->getMemoryVT();
6915 Ptr = ST->getBasePtr();
6920 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
6922 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
6926 bool ARM64TargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
6929 ISD::MemIndexedMode &AM,
6930 SelectionDAG &DAG) const {
6933 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6934 VT = LD->getMemoryVT();
6935 Ptr = LD->getBasePtr();
6936 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6937 VT = ST->getMemoryVT();
6938 Ptr = ST->getBasePtr();
6943 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
6945 // Post-indexing updates the base, so it's not a valid transform
6946 // if that's not the same as the load's pointer.
6949 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
6953 void ARM64TargetLowering::ReplaceNodeResults(SDNode *N,
6954 SmallVectorImpl<SDValue> &Results,
6955 SelectionDAG &DAG) const {
6956 switch (N->getOpcode()) {
6958 llvm_unreachable("Don't know how to custom expand this");
6959 case ISD::FP_TO_UINT:
6960 case ISD::FP_TO_SINT:
6961 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
6962 // Let normal code take care of it by not adding anything to Results.
6967 bool ARM64TargetLowering::shouldExpandAtomicInIR(Instruction *Inst) const {
6968 // Loads and stores less than 128-bits are already atomic; ones above that
6969 // are doomed anyway, so defer to the default libcall and blame the OS when
6971 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
6972 return SI->getValueOperand()->getType()->getPrimitiveSizeInBits() == 128;
6973 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
6974 return LI->getType()->getPrimitiveSizeInBits() == 128;
6976 // For the real atomic operations, we have ldxr/stxr up to 128 bits.
6977 return Inst->getType()->getPrimitiveSizeInBits() <= 128;
6980 Value *ARM64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
6981 AtomicOrdering Ord) const {
6982 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
6983 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
6985 Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent;
6987 // Since i128 isn't legal and intrinsics don't get type-lowered, the ldrexd
6988 // intrinsic must return {i64, i64} and we have to recombine them into a
6989 // single i128 here.
6990 if (ValTy->getPrimitiveSizeInBits() == 128) {
6992 IsAcquire ? Intrinsic::arm64_ldaxp : Intrinsic::arm64_ldxp;
6993 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int);
6995 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
6996 Value *LoHi = Builder.CreateCall(Ldxr, Addr, "lohi");
6998 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
6999 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
7000 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
7001 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
7002 return Builder.CreateOr(
7003 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64");
7006 Type *Tys[] = { Addr->getType() };
7008 IsAcquire ? Intrinsic::arm64_ldaxr : Intrinsic::arm64_ldxr;
7009 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int, Tys);
7011 return Builder.CreateTruncOrBitCast(
7012 Builder.CreateCall(Ldxr, Addr),
7013 cast<PointerType>(Addr->getType())->getElementType());
7016 Value *ARM64TargetLowering::emitStoreConditional(IRBuilder<> &Builder,
7017 Value *Val, Value *Addr,
7018 AtomicOrdering Ord) const {
7019 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
7021 Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent;
7023 // Since the intrinsics must have legal type, the i128 intrinsics take two
7024 // parameters: "i64, i64". We must marshal Val into the appropriate form
7026 if (Val->getType()->getPrimitiveSizeInBits() == 128) {
7028 IsRelease ? Intrinsic::arm64_stlxp : Intrinsic::arm64_stxp;
7029 Function *Stxr = Intrinsic::getDeclaration(M, Int);
7030 Type *Int64Ty = Type::getInt64Ty(M->getContext());
7032 Value *Lo = Builder.CreateTrunc(Val, Int64Ty, "lo");
7033 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 64), Int64Ty, "hi");
7034 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
7035 return Builder.CreateCall3(Stxr, Lo, Hi, Addr);
7039 IsRelease ? Intrinsic::arm64_stlxr : Intrinsic::arm64_stxr;
7040 Type *Tys[] = { Addr->getType() };
7041 Function *Stxr = Intrinsic::getDeclaration(M, Int, Tys);
7043 return Builder.CreateCall2(
7044 Stxr, Builder.CreateZExtOrBitCast(
7045 Val, Stxr->getFunctionType()->getParamType(0)),