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 #include "ARM64ISelLowering.h"
15 #include "ARM64PerfectShuffle.h"
16 #include "ARM64Subtarget.h"
17 #include "ARM64CallingConv.h"
18 #include "ARM64MachineFunctionInfo.h"
19 #include "ARM64TargetMachine.h"
20 #include "ARM64TargetObjectFile.h"
21 #include "MCTargetDesc/ARM64AddressingModes.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/Intrinsics.h"
29 #include "llvm/IR/Type.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetOptions.h"
37 #define DEBUG_TYPE "arm64-lower"
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);
88 if (Subtarget->hasFPARMv8()) {
89 addRegisterClass(MVT::f16, &ARM64::FPR16RegClass);
90 addRegisterClass(MVT::f32, &ARM64::FPR32RegClass);
91 addRegisterClass(MVT::f64, &ARM64::FPR64RegClass);
92 addRegisterClass(MVT::f128, &ARM64::FPR128RegClass);
95 if (Subtarget->hasNEON()) {
96 addRegisterClass(MVT::v16i8, &ARM64::FPR8RegClass);
97 addRegisterClass(MVT::v8i16, &ARM64::FPR16RegClass);
98 // Someone set us up the NEON.
99 addDRTypeForNEON(MVT::v2f32);
100 addDRTypeForNEON(MVT::v8i8);
101 addDRTypeForNEON(MVT::v4i16);
102 addDRTypeForNEON(MVT::v2i32);
103 addDRTypeForNEON(MVT::v1i64);
104 addDRTypeForNEON(MVT::v1f64);
106 addQRTypeForNEON(MVT::v4f32);
107 addQRTypeForNEON(MVT::v2f64);
108 addQRTypeForNEON(MVT::v16i8);
109 addQRTypeForNEON(MVT::v8i16);
110 addQRTypeForNEON(MVT::v4i32);
111 addQRTypeForNEON(MVT::v2i64);
114 // Compute derived properties from the register classes
115 computeRegisterProperties();
117 // Provide all sorts of operation actions
118 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
119 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
120 setOperationAction(ISD::SETCC, MVT::i32, Custom);
121 setOperationAction(ISD::SETCC, MVT::i64, Custom);
122 setOperationAction(ISD::SETCC, MVT::f32, Custom);
123 setOperationAction(ISD::SETCC, MVT::f64, Custom);
124 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
125 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
126 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
127 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
128 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
129 setOperationAction(ISD::SELECT, MVT::i32, Custom);
130 setOperationAction(ISD::SELECT, MVT::i64, Custom);
131 setOperationAction(ISD::SELECT, MVT::f32, Custom);
132 setOperationAction(ISD::SELECT, MVT::f64, Custom);
133 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
134 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
135 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
136 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
137 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
138 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
140 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
141 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
142 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
144 setOperationAction(ISD::FREM, MVT::f32, Expand);
145 setOperationAction(ISD::FREM, MVT::f64, Expand);
146 setOperationAction(ISD::FREM, MVT::f80, Expand);
148 // Custom lowering hooks are needed for XOR
149 // to fold it into CSINC/CSINV.
150 setOperationAction(ISD::XOR, MVT::i32, Custom);
151 setOperationAction(ISD::XOR, MVT::i64, Custom);
153 // Virtually no operation on f128 is legal, but LLVM can't expand them when
154 // there's a valid register class, so we need custom operations in most cases.
155 setOperationAction(ISD::FABS, MVT::f128, Expand);
156 setOperationAction(ISD::FADD, MVT::f128, Custom);
157 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
158 setOperationAction(ISD::FCOS, MVT::f128, Expand);
159 setOperationAction(ISD::FDIV, MVT::f128, Custom);
160 setOperationAction(ISD::FMA, MVT::f128, Expand);
161 setOperationAction(ISD::FMUL, MVT::f128, Custom);
162 setOperationAction(ISD::FNEG, MVT::f128, Expand);
163 setOperationAction(ISD::FPOW, MVT::f128, Expand);
164 setOperationAction(ISD::FREM, MVT::f128, Expand);
165 setOperationAction(ISD::FRINT, MVT::f128, Expand);
166 setOperationAction(ISD::FSIN, MVT::f128, Expand);
167 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
168 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
169 setOperationAction(ISD::FSUB, MVT::f128, Custom);
170 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
171 setOperationAction(ISD::SETCC, MVT::f128, Custom);
172 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
173 setOperationAction(ISD::SELECT, MVT::f128, Custom);
174 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
175 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
177 // Lowering for many of the conversions is actually specified by the non-f128
178 // type. The LowerXXX function will be trivial when f128 isn't involved.
179 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
180 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
181 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
182 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
183 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
184 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
185 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
186 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
187 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
188 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
189 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
190 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
191 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
192 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
194 // Variable arguments.
195 setOperationAction(ISD::VASTART, MVT::Other, Custom);
196 setOperationAction(ISD::VAARG, MVT::Other, Custom);
197 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
198 setOperationAction(ISD::VAEND, MVT::Other, Expand);
200 // Variable-sized objects.
201 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
202 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
203 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
205 // Exception handling.
206 // FIXME: These are guesses. Has this been defined yet?
207 setExceptionPointerRegister(ARM64::X0);
208 setExceptionSelectorRegister(ARM64::X1);
210 // Constant pool entries
211 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
214 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
216 // Add/Sub overflow ops with MVT::Glues are lowered to CPSR dependences.
217 setOperationAction(ISD::ADDC, MVT::i32, Custom);
218 setOperationAction(ISD::ADDE, MVT::i32, Custom);
219 setOperationAction(ISD::SUBC, MVT::i32, Custom);
220 setOperationAction(ISD::SUBE, MVT::i32, Custom);
221 setOperationAction(ISD::ADDC, MVT::i64, Custom);
222 setOperationAction(ISD::ADDE, MVT::i64, Custom);
223 setOperationAction(ISD::SUBC, MVT::i64, Custom);
224 setOperationAction(ISD::SUBE, MVT::i64, Custom);
226 // ARM64 lacks both left-rotate and popcount instructions.
227 setOperationAction(ISD::ROTL, MVT::i32, Expand);
228 setOperationAction(ISD::ROTL, MVT::i64, Expand);
230 // ARM64 doesn't have {U|S}MUL_LOHI.
231 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
232 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
235 // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
236 // counterparts, which ARM64 supports directly.
237 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
238 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
239 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
240 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
242 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
243 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
245 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
246 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
247 setOperationAction(ISD::SREM, MVT::i32, Expand);
248 setOperationAction(ISD::SREM, MVT::i64, Expand);
249 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
250 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
251 setOperationAction(ISD::UREM, MVT::i32, Expand);
252 setOperationAction(ISD::UREM, MVT::i64, Expand);
254 // Custom lower Add/Sub/Mul with overflow.
255 setOperationAction(ISD::SADDO, MVT::i32, Custom);
256 setOperationAction(ISD::SADDO, MVT::i64, Custom);
257 setOperationAction(ISD::UADDO, MVT::i32, Custom);
258 setOperationAction(ISD::UADDO, MVT::i64, Custom);
259 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
260 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
261 setOperationAction(ISD::USUBO, MVT::i32, Custom);
262 setOperationAction(ISD::USUBO, MVT::i64, Custom);
263 setOperationAction(ISD::SMULO, MVT::i32, Custom);
264 setOperationAction(ISD::SMULO, MVT::i64, Custom);
265 setOperationAction(ISD::UMULO, MVT::i32, Custom);
266 setOperationAction(ISD::UMULO, MVT::i64, Custom);
268 setOperationAction(ISD::FSIN, MVT::f32, Expand);
269 setOperationAction(ISD::FSIN, MVT::f64, Expand);
270 setOperationAction(ISD::FCOS, MVT::f32, Expand);
271 setOperationAction(ISD::FCOS, MVT::f64, Expand);
272 setOperationAction(ISD::FPOW, MVT::f32, Expand);
273 setOperationAction(ISD::FPOW, MVT::f64, Expand);
274 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
275 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
277 // ARM64 has implementations of a lot of rounding-like FP operations.
278 static MVT RoundingTypes[] = { MVT::f32, MVT::f64};
279 for (unsigned I = 0; I < array_lengthof(RoundingTypes); ++I) {
280 MVT Ty = RoundingTypes[I];
281 setOperationAction(ISD::FFLOOR, Ty, Legal);
282 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
283 setOperationAction(ISD::FCEIL, Ty, Legal);
284 setOperationAction(ISD::FRINT, Ty, Legal);
285 setOperationAction(ISD::FTRUNC, Ty, Legal);
286 setOperationAction(ISD::FROUND, Ty, Legal);
289 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
291 if (Subtarget->isTargetMachO()) {
292 // For iOS, we don't want to the normal expansion of a libcall to
293 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
295 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
296 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
298 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
299 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
302 // ARM64 does not have floating-point extending loads, i1 sign-extending load,
303 // floating-point truncating stores, or v2i32->v2i16 truncating store.
304 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
305 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
306 setLoadExtAction(ISD::EXTLOAD, MVT::f80, Expand);
307 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Expand);
308 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
309 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
310 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
311 setTruncStoreAction(MVT::f128, MVT::f80, Expand);
312 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
313 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
314 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
315 // Indexed loads and stores are supported.
316 for (unsigned im = (unsigned)ISD::PRE_INC;
317 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
318 setIndexedLoadAction(im, MVT::i8, Legal);
319 setIndexedLoadAction(im, MVT::i16, Legal);
320 setIndexedLoadAction(im, MVT::i32, Legal);
321 setIndexedLoadAction(im, MVT::i64, Legal);
322 setIndexedLoadAction(im, MVT::f64, Legal);
323 setIndexedLoadAction(im, MVT::f32, Legal);
324 setIndexedStoreAction(im, MVT::i8, Legal);
325 setIndexedStoreAction(im, MVT::i16, Legal);
326 setIndexedStoreAction(im, MVT::i32, Legal);
327 setIndexedStoreAction(im, MVT::i64, Legal);
328 setIndexedStoreAction(im, MVT::f64, Legal);
329 setIndexedStoreAction(im, MVT::f32, Legal);
333 setOperationAction(ISD::TRAP, MVT::Other, Legal);
335 // We combine OR nodes for bitfield operations.
336 setTargetDAGCombine(ISD::OR);
338 // Vector add and sub nodes may conceal a high-half opportunity.
339 // Also, try to fold ADD into CSINC/CSINV..
340 setTargetDAGCombine(ISD::ADD);
341 setTargetDAGCombine(ISD::SUB);
343 setTargetDAGCombine(ISD::XOR);
344 setTargetDAGCombine(ISD::SINT_TO_FP);
345 setTargetDAGCombine(ISD::UINT_TO_FP);
347 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
349 setTargetDAGCombine(ISD::ANY_EXTEND);
350 setTargetDAGCombine(ISD::ZERO_EXTEND);
351 setTargetDAGCombine(ISD::SIGN_EXTEND);
352 setTargetDAGCombine(ISD::BITCAST);
353 setTargetDAGCombine(ISD::CONCAT_VECTORS);
354 setTargetDAGCombine(ISD::STORE);
356 setTargetDAGCombine(ISD::MUL);
358 setTargetDAGCombine(ISD::VSELECT);
360 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
361 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
362 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
364 setStackPointerRegisterToSaveRestore(ARM64::SP);
366 setSchedulingPreference(Sched::Hybrid);
369 MaskAndBranchFoldingIsLegal = true;
371 setMinFunctionAlignment(2);
373 RequireStrictAlign = StrictAlign;
375 setHasExtractBitsInsn(true);
377 if (Subtarget->hasNEON()) {
378 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
379 // silliness like this:
380 setOperationAction(ISD::FABS, MVT::v1f64, Expand);
381 setOperationAction(ISD::FADD, MVT::v1f64, Expand);
382 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
383 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
384 setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
385 setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
386 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
387 setOperationAction(ISD::FMA, MVT::v1f64, Expand);
388 setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
389 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
390 setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
391 setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
392 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
393 setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
394 setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
395 setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
396 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
397 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
398 setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
399 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
400 setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
401 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
402 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
403 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
404 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);
406 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
407 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
408 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
409 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
410 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);
412 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
414 // ARM64 doesn't have a direct vector ->f32 conversion instructions for
415 // elements smaller than i32, so promote the input to i32 first.
416 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
417 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
418 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
419 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
420 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
421 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
422 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
423 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
424 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
426 // ARM64 doesn't have MUL.2d:
427 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
428 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
429 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
430 // Likewise, narrowing and extending vector loads/stores aren't handled
432 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
433 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
435 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,
438 setOperationAction(ISD::MULHS, (MVT::SimpleValueType)VT, Expand);
439 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
440 setOperationAction(ISD::MULHU, (MVT::SimpleValueType)VT, Expand);
441 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
443 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
444 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
445 setTruncStoreAction((MVT::SimpleValueType)VT,
446 (MVT::SimpleValueType)InnerVT, Expand);
447 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
448 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
449 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
452 // ARM64 has implementations of a lot of rounding-like FP operations.
453 static MVT RoundingVecTypes[] = {MVT::v2f32, MVT::v4f32, MVT::v2f64 };
454 for (unsigned I = 0; I < array_lengthof(RoundingVecTypes); ++I) {
455 MVT Ty = RoundingVecTypes[I];
456 setOperationAction(ISD::FFLOOR, Ty, Legal);
457 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
458 setOperationAction(ISD::FCEIL, Ty, Legal);
459 setOperationAction(ISD::FRINT, Ty, Legal);
460 setOperationAction(ISD::FTRUNC, Ty, Legal);
461 setOperationAction(ISD::FROUND, Ty, Legal);
466 void ARM64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
467 if (VT == MVT::v2f32) {
468 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
469 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
471 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
472 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
473 } else if (VT == MVT::v2f64 || VT == MVT::v4f32) {
474 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
475 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
477 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
478 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
481 // Mark vector float intrinsics as expand.
482 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
483 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
484 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
485 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
486 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
487 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
488 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
489 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
490 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
491 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
494 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
495 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
496 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
497 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
498 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
499 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
500 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
501 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
502 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
503 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
504 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
505 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
507 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
508 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
509 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
510 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
512 // CNT supports only B element sizes.
513 if (VT != MVT::v8i8 && VT != MVT::v16i8)
514 setOperationAction(ISD::CTPOP, VT.getSimpleVT(), Expand);
516 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
517 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
518 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
519 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
520 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
522 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
523 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
526 void ARM64TargetLowering::addDRTypeForNEON(MVT VT) {
527 addRegisterClass(VT, &ARM64::FPR64RegClass);
528 addTypeForNEON(VT, MVT::v2i32);
531 void ARM64TargetLowering::addQRTypeForNEON(MVT VT) {
532 addRegisterClass(VT, &ARM64::FPR128RegClass);
533 addTypeForNEON(VT, MVT::v4i32);
536 EVT ARM64TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
539 return VT.changeVectorElementTypeToInteger();
542 /// computeMaskedBitsForTargetNode - Determine which of the bits specified in
543 /// Mask are known to be either zero or one and return them in the
544 /// KnownZero/KnownOne bitsets.
545 void ARM64TargetLowering::computeMaskedBitsForTargetNode(
546 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
547 const SelectionDAG &DAG, unsigned Depth) const {
548 switch (Op.getOpcode()) {
551 case ARM64ISD::CSEL: {
552 APInt KnownZero2, KnownOne2;
553 DAG.ComputeMaskedBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
554 DAG.ComputeMaskedBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
555 KnownZero &= KnownZero2;
556 KnownOne &= KnownOne2;
559 case ISD::INTRINSIC_W_CHAIN: {
560 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
561 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
564 case Intrinsic::arm64_ldaxr:
565 case Intrinsic::arm64_ldxr: {
566 unsigned BitWidth = KnownOne.getBitWidth();
567 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
568 unsigned MemBits = VT.getScalarType().getSizeInBits();
569 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
575 case ISD::INTRINSIC_WO_CHAIN:
576 case ISD::INTRINSIC_VOID: {
577 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
581 case Intrinsic::arm64_neon_umaxv:
582 case Intrinsic::arm64_neon_uminv: {
583 // Figure out the datatype of the vector operand. The UMINV instruction
584 // will zero extend the result, so we can mark as known zero all the
585 // bits larger than the element datatype. 32-bit or larget doesn't need
586 // this as those are legal types and will be handled by isel directly.
587 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
588 unsigned BitWidth = KnownZero.getBitWidth();
589 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
590 assert(BitWidth >= 8 && "Unexpected width!");
591 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
593 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
594 assert(BitWidth >= 16 && "Unexpected width!");
595 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
605 MVT ARM64TargetLowering::getScalarShiftAmountTy(EVT LHSTy) const {
609 unsigned ARM64TargetLowering::getMaximalGlobalOffset() const {
610 // FIXME: On ARM64, this depends on the type.
611 // Basically, the addressable offsets are o to 4095 * Ty.getSizeInBytes().
612 // and the offset has to be a multiple of the related size in bytes.
617 ARM64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
618 const TargetLibraryInfo *libInfo) const {
619 return ARM64::createFastISel(funcInfo, libInfo);
622 const char *ARM64TargetLowering::getTargetNodeName(unsigned Opcode) const {
626 case ARM64ISD::CALL: return "ARM64ISD::CALL";
627 case ARM64ISD::ADRP: return "ARM64ISD::ADRP";
628 case ARM64ISD::ADDlow: return "ARM64ISD::ADDlow";
629 case ARM64ISD::LOADgot: return "ARM64ISD::LOADgot";
630 case ARM64ISD::RET_FLAG: return "ARM64ISD::RET_FLAG";
631 case ARM64ISD::BRCOND: return "ARM64ISD::BRCOND";
632 case ARM64ISD::CSEL: return "ARM64ISD::CSEL";
633 case ARM64ISD::FCSEL: return "ARM64ISD::FCSEL";
634 case ARM64ISD::CSINV: return "ARM64ISD::CSINV";
635 case ARM64ISD::CSNEG: return "ARM64ISD::CSNEG";
636 case ARM64ISD::CSINC: return "ARM64ISD::CSINC";
637 case ARM64ISD::THREAD_POINTER: return "ARM64ISD::THREAD_POINTER";
638 case ARM64ISD::TLSDESC_CALL: return "ARM64ISD::TLSDESC_CALL";
639 case ARM64ISD::ADC: return "ARM64ISD::ADC";
640 case ARM64ISD::SBC: return "ARM64ISD::SBC";
641 case ARM64ISD::ADDS: return "ARM64ISD::ADDS";
642 case ARM64ISD::SUBS: return "ARM64ISD::SUBS";
643 case ARM64ISD::ADCS: return "ARM64ISD::ADCS";
644 case ARM64ISD::SBCS: return "ARM64ISD::SBCS";
645 case ARM64ISD::ANDS: return "ARM64ISD::ANDS";
646 case ARM64ISD::FCMP: return "ARM64ISD::FCMP";
647 case ARM64ISD::FMIN: return "ARM64ISD::FMIN";
648 case ARM64ISD::FMAX: return "ARM64ISD::FMAX";
649 case ARM64ISD::DUP: return "ARM64ISD::DUP";
650 case ARM64ISD::DUPLANE8: return "ARM64ISD::DUPLANE8";
651 case ARM64ISD::DUPLANE16: return "ARM64ISD::DUPLANE16";
652 case ARM64ISD::DUPLANE32: return "ARM64ISD::DUPLANE32";
653 case ARM64ISD::DUPLANE64: return "ARM64ISD::DUPLANE64";
654 case ARM64ISD::MOVI: return "ARM64ISD::MOVI";
655 case ARM64ISD::MOVIshift: return "ARM64ISD::MOVIshift";
656 case ARM64ISD::MOVIedit: return "ARM64ISD::MOVIedit";
657 case ARM64ISD::MOVImsl: return "ARM64ISD::MOVImsl";
658 case ARM64ISD::FMOV: return "ARM64ISD::FMOV";
659 case ARM64ISD::MVNIshift: return "ARM64ISD::MVNIshift";
660 case ARM64ISD::MVNImsl: return "ARM64ISD::MVNImsl";
661 case ARM64ISD::BICi: return "ARM64ISD::BICi";
662 case ARM64ISD::ORRi: return "ARM64ISD::ORRi";
663 case ARM64ISD::BSL: return "ARM64ISD::BSL";
664 case ARM64ISD::NEG: return "ARM64ISD::NEG";
665 case ARM64ISD::EXTR: return "ARM64ISD::EXTR";
666 case ARM64ISD::ZIP1: return "ARM64ISD::ZIP1";
667 case ARM64ISD::ZIP2: return "ARM64ISD::ZIP2";
668 case ARM64ISD::UZP1: return "ARM64ISD::UZP1";
669 case ARM64ISD::UZP2: return "ARM64ISD::UZP2";
670 case ARM64ISD::TRN1: return "ARM64ISD::TRN1";
671 case ARM64ISD::TRN2: return "ARM64ISD::TRN2";
672 case ARM64ISD::REV16: return "ARM64ISD::REV16";
673 case ARM64ISD::REV32: return "ARM64ISD::REV32";
674 case ARM64ISD::REV64: return "ARM64ISD::REV64";
675 case ARM64ISD::EXT: return "ARM64ISD::EXT";
676 case ARM64ISD::VSHL: return "ARM64ISD::VSHL";
677 case ARM64ISD::VLSHR: return "ARM64ISD::VLSHR";
678 case ARM64ISD::VASHR: return "ARM64ISD::VASHR";
679 case ARM64ISD::CMEQ: return "ARM64ISD::CMEQ";
680 case ARM64ISD::CMGE: return "ARM64ISD::CMGE";
681 case ARM64ISD::CMGT: return "ARM64ISD::CMGT";
682 case ARM64ISD::CMHI: return "ARM64ISD::CMHI";
683 case ARM64ISD::CMHS: return "ARM64ISD::CMHS";
684 case ARM64ISD::FCMEQ: return "ARM64ISD::FCMEQ";
685 case ARM64ISD::FCMGE: return "ARM64ISD::FCMGE";
686 case ARM64ISD::FCMGT: return "ARM64ISD::FCMGT";
687 case ARM64ISD::CMEQz: return "ARM64ISD::CMEQz";
688 case ARM64ISD::CMGEz: return "ARM64ISD::CMGEz";
689 case ARM64ISD::CMGTz: return "ARM64ISD::CMGTz";
690 case ARM64ISD::CMLEz: return "ARM64ISD::CMLEz";
691 case ARM64ISD::CMLTz: return "ARM64ISD::CMLTz";
692 case ARM64ISD::FCMEQz: return "ARM64ISD::FCMEQz";
693 case ARM64ISD::FCMGEz: return "ARM64ISD::FCMGEz";
694 case ARM64ISD::FCMGTz: return "ARM64ISD::FCMGTz";
695 case ARM64ISD::FCMLEz: return "ARM64ISD::FCMLEz";
696 case ARM64ISD::FCMLTz: return "ARM64ISD::FCMLTz";
697 case ARM64ISD::NOT: return "ARM64ISD::NOT";
698 case ARM64ISD::BIT: return "ARM64ISD::BIT";
699 case ARM64ISD::CBZ: return "ARM64ISD::CBZ";
700 case ARM64ISD::CBNZ: return "ARM64ISD::CBNZ";
701 case ARM64ISD::TBZ: return "ARM64ISD::TBZ";
702 case ARM64ISD::TBNZ: return "ARM64ISD::TBNZ";
703 case ARM64ISD::TC_RETURN: return "ARM64ISD::TC_RETURN";
704 case ARM64ISD::SITOF: return "ARM64ISD::SITOF";
705 case ARM64ISD::UITOF: return "ARM64ISD::UITOF";
706 case ARM64ISD::SQSHL_I: return "ARM64ISD::SQSHL_I";
707 case ARM64ISD::UQSHL_I: return "ARM64ISD::UQSHL_I";
708 case ARM64ISD::SRSHR_I: return "ARM64ISD::SRSHR_I";
709 case ARM64ISD::URSHR_I: return "ARM64ISD::URSHR_I";
710 case ARM64ISD::SQSHLU_I: return "ARM64ISD::SQSHLU_I";
711 case ARM64ISD::WrapperLarge: return "ARM64ISD::WrapperLarge";
716 ARM64TargetLowering::EmitF128CSEL(MachineInstr *MI,
717 MachineBasicBlock *MBB) const {
718 // We materialise the F128CSEL pseudo-instruction as some control flow and a
722 // [... previous instrs leading to comparison ...]
728 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
730 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
731 MachineFunction *MF = MBB->getParent();
732 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
733 DebugLoc DL = MI->getDebugLoc();
734 MachineFunction::iterator It = MBB;
737 unsigned DestReg = MI->getOperand(0).getReg();
738 unsigned IfTrueReg = MI->getOperand(1).getReg();
739 unsigned IfFalseReg = MI->getOperand(2).getReg();
740 unsigned CondCode = MI->getOperand(3).getImm();
741 bool CPSRKilled = MI->getOperand(4).isKill();
743 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
744 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
745 MF->insert(It, TrueBB);
746 MF->insert(It, EndBB);
748 // Transfer rest of current basic-block to EndBB
749 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
751 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
753 BuildMI(MBB, DL, TII->get(ARM64::Bcc)).addImm(CondCode).addMBB(TrueBB);
754 BuildMI(MBB, DL, TII->get(ARM64::B)).addMBB(EndBB);
755 MBB->addSuccessor(TrueBB);
756 MBB->addSuccessor(EndBB);
758 // TrueBB falls through to the end.
759 TrueBB->addSuccessor(EndBB);
762 TrueBB->addLiveIn(ARM64::CPSR);
763 EndBB->addLiveIn(ARM64::CPSR);
766 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(ARM64::PHI), DestReg)
772 MI->eraseFromParent();
777 ARM64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
778 MachineBasicBlock *BB) const {
779 switch (MI->getOpcode()) {
784 assert(0 && "Unexpected instruction for custom inserter!");
787 case ARM64::F128CSEL:
788 return EmitF128CSEL(MI, BB);
790 case TargetOpcode::STACKMAP:
791 case TargetOpcode::PATCHPOINT:
792 return emitPatchPoint(MI, BB);
794 llvm_unreachable("Unexpected instruction for custom inserter!");
797 //===----------------------------------------------------------------------===//
798 // ARM64 Lowering private implementation.
799 //===----------------------------------------------------------------------===//
801 //===----------------------------------------------------------------------===//
803 //===----------------------------------------------------------------------===//
805 /// changeIntCCToARM64CC - Convert a DAG integer condition code to an ARM64 CC
806 static ARM64CC::CondCode changeIntCCToARM64CC(ISD::CondCode CC) {
809 llvm_unreachable("Unknown condition code!");
833 /// changeFPCCToARM64CC - Convert a DAG fp condition code to an ARM64 CC.
834 static void changeFPCCToARM64CC(ISD::CondCode CC, ARM64CC::CondCode &CondCode,
835 ARM64CC::CondCode &CondCode2) {
836 CondCode2 = ARM64CC::AL;
839 llvm_unreachable("Unknown FP condition!");
842 CondCode = ARM64CC::EQ;
846 CondCode = ARM64CC::GT;
850 CondCode = ARM64CC::GE;
853 CondCode = ARM64CC::MI;
856 CondCode = ARM64CC::LS;
859 CondCode = ARM64CC::MI;
860 CondCode2 = ARM64CC::GT;
863 CondCode = ARM64CC::VC;
866 CondCode = ARM64CC::VS;
869 CondCode = ARM64CC::EQ;
870 CondCode2 = ARM64CC::VS;
873 CondCode = ARM64CC::HI;
876 CondCode = ARM64CC::PL;
880 CondCode = ARM64CC::LT;
884 CondCode = ARM64CC::LE;
888 CondCode = ARM64CC::NE;
893 /// changeVectorFPCCToARM64CC - Convert a DAG fp condition code to an ARM64 CC
894 /// usable with the vector instructions. Fewer operations are available without
895 /// a real NZCV register, so we have to use less efficient combinations to get
897 static void changeVectorFPCCToARM64CC(ISD::CondCode CC,
898 ARM64CC::CondCode &CondCode,
899 ARM64CC::CondCode &CondCode2,
904 // Mostly the scalar mappings work fine.
905 changeFPCCToARM64CC(CC, CondCode, CondCode2);
908 Invert = true; // Fallthrough
910 CondCode = ARM64CC::MI;
911 CondCode2 = ARM64CC::GE;
918 // All of the compare-mask comparisons are ordered, but we can switch
919 // between the two by a double inversion. E.g. ULE == !OGT.
921 changeFPCCToARM64CC(getSetCCInverse(CC, false), CondCode, CondCode2);
926 static bool isLegalArithImmed(uint64_t C) {
927 // Matches ARM64DAGToDAGISel::SelectArithImmed().
928 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
931 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
932 SDLoc dl, SelectionDAG &DAG) {
933 EVT VT = LHS.getValueType();
935 if (VT.isFloatingPoint())
936 return DAG.getNode(ARM64ISD::FCMP, dl, VT, LHS, RHS);
938 // The CMP instruction is just an alias for SUBS, and representing it as
939 // SUBS means that it's possible to get CSE with subtract operations.
940 // A later phase can perform the optimization of setting the destination
941 // register to WZR/XZR if it ends up being unused.
942 unsigned Opcode = ARM64ISD::SUBS;
944 if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
945 cast<ConstantSDNode>(RHS.getOperand(0))->getZExtValue() == 0 &&
946 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
947 // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on
948 // the grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags
949 // can be set differently by this operation. It comes down to whether
950 // "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then
951 // everything is fine. If not then the optimization is wrong. Thus general
952 // comparisons are only valid if op2 != 0.
954 // So, finally, the only LLVM-native comparisons that don't mention C and V
955 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in
956 // the absence of information about op2.
957 Opcode = ARM64ISD::ADDS;
958 RHS = RHS.getOperand(1);
959 } else if (LHS.getOpcode() == ISD::AND && isa<ConstantSDNode>(RHS) &&
960 cast<ConstantSDNode>(RHS)->getZExtValue() == 0 &&
961 !isUnsignedIntSetCC(CC)) {
962 // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
963 // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
964 // of the signed comparisons.
965 Opcode = ARM64ISD::ANDS;
966 RHS = LHS.getOperand(1);
967 LHS = LHS.getOperand(0);
970 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS)
974 static SDValue getARM64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
975 SDValue &ARM64cc, SelectionDAG &DAG, SDLoc dl) {
976 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
977 EVT VT = RHS.getValueType();
978 uint64_t C = RHSC->getZExtValue();
979 if (!isLegalArithImmed(C)) {
980 // Constant does not fit, try adjusting it by one?
986 if ((VT == MVT::i32 && C != 0x80000000 &&
987 isLegalArithImmed((uint32_t)(C - 1))) ||
988 (VT == MVT::i64 && C != 0x80000000ULL &&
989 isLegalArithImmed(C - 1ULL))) {
990 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
991 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
992 RHS = DAG.getConstant(C, VT);
997 if ((VT == MVT::i32 && C != 0 &&
998 isLegalArithImmed((uint32_t)(C - 1))) ||
999 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1000 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1001 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1002 RHS = DAG.getConstant(C, VT);
1007 if ((VT == MVT::i32 && C != 0x7fffffff &&
1008 isLegalArithImmed((uint32_t)(C + 1))) ||
1009 (VT == MVT::i64 && C != 0x7ffffffffffffffULL &&
1010 isLegalArithImmed(C + 1ULL))) {
1011 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1012 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1013 RHS = DAG.getConstant(C, VT);
1018 if ((VT == MVT::i32 && C != 0xffffffff &&
1019 isLegalArithImmed((uint32_t)(C + 1))) ||
1020 (VT == MVT::i64 && C != 0xfffffffffffffffULL &&
1021 isLegalArithImmed(C + 1ULL))) {
1022 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1023 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1024 RHS = DAG.getConstant(C, VT);
1031 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
1032 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
1033 ARM64cc = DAG.getConstant(ARM64CC, MVT::i32);
1037 static std::pair<SDValue, SDValue>
1038 getARM64XALUOOp(ARM64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1039 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1040 "Unsupported value type");
1041 SDValue Value, Overflow;
1043 SDValue LHS = Op.getOperand(0);
1044 SDValue RHS = Op.getOperand(1);
1046 switch (Op.getOpcode()) {
1048 llvm_unreachable("Unknown overflow instruction!");
1050 Opc = ARM64ISD::ADDS;
1054 Opc = ARM64ISD::ADDS;
1058 Opc = ARM64ISD::SUBS;
1062 Opc = ARM64ISD::SUBS;
1065 // Multiply needs a little bit extra work.
1069 bool IsSigned = (Op.getOpcode() == ISD::SMULO) ? true : false;
1070 if (Op.getValueType() == MVT::i32) {
1071 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1072 // For a 32 bit multiply with overflow check we want the instruction
1073 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1074 // need to generate the following pattern:
1075 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1076 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1077 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1078 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1079 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1080 DAG.getConstant(0, MVT::i64));
1081 // On ARM64 the upper 32 bits are always zero extended for a 32 bit
1082 // operation. We need to clear out the upper 32 bits, because we used a
1083 // widening multiply that wrote all 64 bits. In the end this should be a
1085 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1087 // The signed overflow check requires more than just a simple check for
1088 // any bit set in the upper 32 bits of the result. These bits could be
1089 // just the sign bits of a negative number. To perform the overflow
1090 // check we have to arithmetic shift right the 32nd bit of the result by
1091 // 31 bits. Then we compare the result to the upper 32 bits.
1092 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1093 DAG.getConstant(32, MVT::i64));
1094 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1095 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1096 DAG.getConstant(31, MVT::i64));
1097 // It is important that LowerBits is last, otherwise the arithmetic
1098 // shift will not be folded into the compare (SUBS).
1099 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1100 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1103 // The overflow check for unsigned multiply is easy. We only need to
1104 // check if any of the upper 32 bits are set. This can be done with a
1105 // CMP (shifted register). For that we need to generate the following
1107 // (i64 ARM64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1108 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1109 DAG.getConstant(32, MVT::i64));
1110 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1112 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1113 UpperBits).getValue(1);
1117 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1118 // For the 64 bit multiply
1119 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1121 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1122 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1123 DAG.getConstant(63, MVT::i64));
1124 // It is important that LowerBits is last, otherwise the arithmetic
1125 // shift will not be folded into the compare (SUBS).
1126 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1127 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1130 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1131 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1133 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1134 UpperBits).getValue(1);
1141 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1143 // Emit the ARM64 operation with overflow check.
1144 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1145 Overflow = Value.getValue(1);
1147 return std::make_pair(Value, Overflow);
1150 SDValue ARM64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1151 RTLIB::Libcall Call) const {
1152 SmallVector<SDValue, 2> Ops;
1153 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1154 Ops.push_back(Op.getOperand(i));
1156 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1160 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1161 SDValue Sel = Op.getOperand(0);
1162 SDValue Other = Op.getOperand(1);
1164 // If neither operand is a SELECT_CC, give up.
1165 if (Sel.getOpcode() != ISD::SELECT_CC)
1166 std::swap(Sel, Other);
1167 if (Sel.getOpcode() != ISD::SELECT_CC)
1170 // The folding we want to perform is:
1171 // (xor x, (select_cc a, b, cc, 0, -1) )
1173 // (csel x, (xor x, -1), cc ...)
1175 // The latter will get matched to a CSINV instruction.
1177 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1178 SDValue LHS = Sel.getOperand(0);
1179 SDValue RHS = Sel.getOperand(1);
1180 SDValue TVal = Sel.getOperand(2);
1181 SDValue FVal = Sel.getOperand(3);
1184 // FIXME: This could be generalized to non-integer comparisons.
1185 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1188 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1189 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1191 // The the values aren't constants, this isn't the pattern we're looking for.
1192 if (!CFVal || !CTVal)
1195 // We can commute the SELECT_CC by inverting the condition. This
1196 // might be needed to make this fit into a CSINV pattern.
1197 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1198 std::swap(TVal, FVal);
1199 std::swap(CTVal, CFVal);
1200 CC = ISD::getSetCCInverse(CC, true);
1203 // If the constants line up, perform the transform!
1204 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1206 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1209 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1210 DAG.getConstant(-1ULL, Other.getValueType()));
1212 return DAG.getNode(ARM64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1219 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1220 EVT VT = Op.getValueType();
1222 // Let legalize expand this if it isn't a legal type yet.
1223 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1226 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1229 bool ExtraOp = false;
1230 switch (Op.getOpcode()) {
1232 assert(0 && "Invalid code");
1234 Opc = ARM64ISD::ADDS;
1237 Opc = ARM64ISD::SUBS;
1240 Opc = ARM64ISD::ADCS;
1244 Opc = ARM64ISD::SBCS;
1250 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1251 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1255 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1256 // Let legalize expand this if it isn't a legal type yet.
1257 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1260 ARM64CC::CondCode CC;
1261 // The actual operation that sets the overflow or carry flag.
1262 SDValue Value, Overflow;
1263 std::tie(Value, Overflow) = getARM64XALUOOp(CC, Op, DAG);
1265 // We use 0 and 1 as false and true values.
1266 SDValue TVal = DAG.getConstant(1, MVT::i32);
1267 SDValue FVal = DAG.getConstant(0, MVT::i32);
1269 // We use an inverted condition, because the conditional select is inverted
1270 // too. This will allow it to be selected to a single instruction:
1271 // CSINC Wd, WZR, WZR, invert(cond).
1272 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), MVT::i32);
1273 Overflow = DAG.getNode(ARM64ISD::CSEL, SDLoc(Op), MVT::i32, FVal, TVal, CCVal,
1276 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1277 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
1280 // Prefetch operands are:
1281 // 1: Address to prefetch
1283 // 3: int locality (0 = no locality ... 3 = extreme locality)
1284 // 4: bool isDataCache
1285 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1287 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1288 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1289 // The data thing is not used.
1290 // unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1292 bool IsStream = !Locality;
1293 // When the locality number is set
1295 // The front-end should have filtered out the out-of-range values
1296 assert(Locality <= 3 && "Prefetch locality out-of-range");
1297 // The locality degree is the opposite of the cache speed.
1298 // Put the number the other way around.
1299 // The encoding starts at 0 for level 1
1300 Locality = 3 - Locality;
1303 // built the mask value encoding the expected behavior.
1304 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1305 (Locality << 1) | // Cache level bits
1306 (unsigned)IsStream; // Stream bit
1307 return DAG.getNode(ARM64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1308 DAG.getConstant(PrfOp, MVT::i32), Op.getOperand(1));
1311 SDValue ARM64TargetLowering::LowerFP_EXTEND(SDValue Op,
1312 SelectionDAG &DAG) const {
1313 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1316 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1318 return LowerF128Call(Op, DAG, LC);
1321 SDValue ARM64TargetLowering::LowerFP_ROUND(SDValue Op,
1322 SelectionDAG &DAG) const {
1323 if (Op.getOperand(0).getValueType() != MVT::f128) {
1324 // It's legal except when f128 is involved
1329 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1331 // FP_ROUND node has a second operand indicating whether it is known to be
1332 // precise. That doesn't take part in the LibCall so we can't directly use
1334 SDValue SrcVal = Op.getOperand(0);
1335 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1336 /*isSigned*/ false, SDLoc(Op)).first;
1339 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1340 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1341 // Any additional optimization in this function should be recorded
1342 // in the cost tables.
1343 EVT InVT = Op.getOperand(0).getValueType();
1344 EVT VT = Op.getValueType();
1346 // FP_TO_XINT conversion from the same type are legal.
1347 if (VT.getSizeInBits() == InVT.getSizeInBits())
1350 if (InVT == MVT::v2f64 || InVT == MVT::v4f32) {
1353 DAG.getNode(Op.getOpcode(), dl, InVT.changeVectorElementTypeToInteger(),
1355 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1356 } else if (InVT == MVT::v2f32) {
1358 SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, MVT::v2f64, Op.getOperand(0));
1359 return DAG.getNode(Op.getOpcode(), dl, VT, Ext);
1362 // Type changing conversions are illegal.
1366 SDValue ARM64TargetLowering::LowerFP_TO_INT(SDValue Op,
1367 SelectionDAG &DAG) const {
1368 if (Op.getOperand(0).getValueType().isVector())
1369 return LowerVectorFP_TO_INT(Op, DAG);
1371 if (Op.getOperand(0).getValueType() != MVT::f128) {
1372 // It's legal except when f128 is involved
1377 if (Op.getOpcode() == ISD::FP_TO_SINT)
1378 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1380 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1382 SmallVector<SDValue, 2> Ops;
1383 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1384 Ops.push_back(Op.getOperand(i));
1386 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1390 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1391 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1392 // Any additional optimization in this function should be recorded
1393 // in the cost tables.
1394 EVT VT = Op.getValueType();
1396 SDValue In = Op.getOperand(0);
1397 EVT InVT = In.getValueType();
1399 // v2i32 to v2f32 is legal.
1400 if (VT == MVT::v2f32 && InVT == MVT::v2i32)
1403 // This function only handles v2f64 outputs.
1404 if (VT == MVT::v2f64) {
1405 // Extend the input argument to a v2i64 that we can feed into the
1406 // floating point conversion. Zero or sign extend based on whether
1407 // we're doing a signed or unsigned float conversion.
1409 Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
1410 assert(Op.getNumOperands() == 1 && "FP conversions take one argument");
1411 SDValue Promoted = DAG.getNode(Opc, dl, MVT::v2i64, Op.getOperand(0));
1412 return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), Promoted);
1415 // Scalarize v2i64 to v2f32 conversions.
1416 std::vector<SDValue> BuildVectorOps;
1417 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
1418 SDValue Sclr = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, In,
1419 DAG.getConstant(i, MVT::i64));
1420 Sclr = DAG.getNode(Op->getOpcode(), dl, MVT::f32, Sclr);
1421 BuildVectorOps.push_back(Sclr);
1424 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, BuildVectorOps);
1427 SDValue ARM64TargetLowering::LowerINT_TO_FP(SDValue Op,
1428 SelectionDAG &DAG) const {
1429 if (Op.getValueType().isVector())
1430 return LowerVectorINT_TO_FP(Op, DAG);
1432 // i128 conversions are libcalls.
1433 if (Op.getOperand(0).getValueType() == MVT::i128)
1436 // Other conversions are legal, unless it's to the completely software-based
1438 if (Op.getValueType() != MVT::f128)
1442 if (Op.getOpcode() == ISD::SINT_TO_FP)
1443 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1445 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1447 return LowerF128Call(Op, DAG, LC);
1450 SDValue ARM64TargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
1451 // For iOS, we want to call an alternative entry point: __sincos_stret,
1452 // which returns the values in two S / D registers.
1454 SDValue Arg = Op.getOperand(0);
1455 EVT ArgVT = Arg.getValueType();
1456 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1463 Entry.isSExt = false;
1464 Entry.isZExt = false;
1465 Args.push_back(Entry);
1467 const char *LibcallName =
1468 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1469 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
1471 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
1472 TargetLowering::CallLoweringInfo CLI(
1473 DAG.getEntryNode(), RetTy, false, false, false, false, 0,
1474 CallingConv::Fast, /*isTaillCall=*/false,
1475 /*doesNotRet=*/false, /*isReturnValueUsed*/ true, Callee, Args, DAG, dl);
1476 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1477 return CallResult.first;
1480 SDValue ARM64TargetLowering::LowerOperation(SDValue Op,
1481 SelectionDAG &DAG) const {
1482 switch (Op.getOpcode()) {
1484 llvm_unreachable("unimplemented operand");
1486 case ISD::GlobalAddress:
1487 return LowerGlobalAddress(Op, DAG);
1488 case ISD::GlobalTLSAddress:
1489 return LowerGlobalTLSAddress(Op, DAG);
1491 return LowerSETCC(Op, DAG);
1493 return LowerBR_CC(Op, DAG);
1495 return LowerSELECT(Op, DAG);
1496 case ISD::SELECT_CC:
1497 return LowerSELECT_CC(Op, DAG);
1498 case ISD::JumpTable:
1499 return LowerJumpTable(Op, DAG);
1500 case ISD::ConstantPool:
1501 return LowerConstantPool(Op, DAG);
1502 case ISD::BlockAddress:
1503 return LowerBlockAddress(Op, DAG);
1505 return LowerVASTART(Op, DAG);
1507 return LowerVACOPY(Op, DAG);
1509 return LowerVAARG(Op, DAG);
1514 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
1521 return LowerXALUO(Op, DAG);
1523 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
1525 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
1527 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
1529 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
1531 return LowerFP_ROUND(Op, DAG);
1532 case ISD::FP_EXTEND:
1533 return LowerFP_EXTEND(Op, DAG);
1534 case ISD::FRAMEADDR:
1535 return LowerFRAMEADDR(Op, DAG);
1536 case ISD::RETURNADDR:
1537 return LowerRETURNADDR(Op, DAG);
1538 case ISD::INSERT_VECTOR_ELT:
1539 return LowerINSERT_VECTOR_ELT(Op, DAG);
1540 case ISD::EXTRACT_VECTOR_ELT:
1541 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
1542 case ISD::BUILD_VECTOR:
1543 return LowerBUILD_VECTOR(Op, DAG);
1544 case ISD::VECTOR_SHUFFLE:
1545 return LowerVECTOR_SHUFFLE(Op, DAG);
1546 case ISD::EXTRACT_SUBVECTOR:
1547 return LowerEXTRACT_SUBVECTOR(Op, DAG);
1551 return LowerVectorSRA_SRL_SHL(Op, DAG);
1552 case ISD::SHL_PARTS:
1553 return LowerShiftLeftParts(Op, DAG);
1554 case ISD::SRL_PARTS:
1555 case ISD::SRA_PARTS:
1556 return LowerShiftRightParts(Op, DAG);
1558 return LowerCTPOP(Op, DAG);
1559 case ISD::FCOPYSIGN:
1560 return LowerFCOPYSIGN(Op, DAG);
1562 return LowerVectorAND(Op, DAG);
1564 return LowerVectorOR(Op, DAG);
1566 return LowerXOR(Op, DAG);
1568 return LowerPREFETCH(Op, DAG);
1569 case ISD::SINT_TO_FP:
1570 case ISD::UINT_TO_FP:
1571 return LowerINT_TO_FP(Op, DAG);
1572 case ISD::FP_TO_SINT:
1573 case ISD::FP_TO_UINT:
1574 return LowerFP_TO_INT(Op, DAG);
1576 return LowerFSINCOS(Op, DAG);
1580 /// getFunctionAlignment - Return the Log2 alignment of this function.
1581 unsigned ARM64TargetLowering::getFunctionAlignment(const Function *F) const {
1585 //===----------------------------------------------------------------------===//
1586 // Calling Convention Implementation
1587 //===----------------------------------------------------------------------===//
1589 #include "ARM64GenCallingConv.inc"
1591 /// Selects the correct CCAssignFn for a the given CallingConvention
1593 CCAssignFn *ARM64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
1594 bool IsVarArg) const {
1597 llvm_unreachable("Unsupported calling convention.");
1598 case CallingConv::WebKit_JS:
1599 return CC_ARM64_WebKit_JS;
1600 case CallingConv::C:
1601 case CallingConv::Fast:
1602 if (!Subtarget->isTargetDarwin())
1603 return CC_ARM64_AAPCS;
1604 return IsVarArg ? CC_ARM64_DarwinPCS_VarArg : CC_ARM64_DarwinPCS;
1608 SDValue ARM64TargetLowering::LowerFormalArguments(
1609 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1610 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
1611 SmallVectorImpl<SDValue> &InVals) const {
1612 MachineFunction &MF = DAG.getMachineFunction();
1613 MachineFrameInfo *MFI = MF.getFrameInfo();
1615 // Assign locations to all of the incoming arguments.
1616 SmallVector<CCValAssign, 16> ArgLocs;
1617 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1618 getTargetMachine(), ArgLocs, *DAG.getContext());
1620 // At this point, Ins[].VT may already be promoted to i32. To correctly
1621 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
1622 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
1623 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
1624 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
1626 unsigned NumArgs = Ins.size();
1627 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
1628 unsigned CurArgIdx = 0;
1629 for (unsigned i = 0; i != NumArgs; ++i) {
1630 MVT ValVT = Ins[i].VT;
1631 std::advance(CurOrigArg, Ins[i].OrigArgIndex - CurArgIdx);
1632 CurArgIdx = Ins[i].OrigArgIndex;
1634 // Get type of the original argument.
1635 EVT ActualVT = getValueType(CurOrigArg->getType(), /*AllowUnknown*/ true);
1636 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
1637 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
1639 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
1641 else if (ActualMVT == MVT::i16)
1644 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
1646 AssignFn(i, ValVT, LocVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
1647 assert(!Res && "Call operand has unhandled type");
1650 assert(ArgLocs.size() == Ins.size());
1651 SmallVector<SDValue, 16> ArgValues;
1652 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1653 CCValAssign &VA = ArgLocs[i];
1655 if (Ins[i].Flags.isByVal()) {
1656 // Byval is used for HFAs in the PCS, but the system should work in a
1657 // non-compliant manner for larger structs.
1658 EVT PtrTy = getPointerTy();
1659 int Size = Ins[i].Flags.getByValSize();
1660 unsigned NumRegs = (Size + 7) / 8;
1663 MFI->CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false);
1664 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrTy);
1665 InVals.push_back(FrameIdxN);
1668 } if (VA.isRegLoc()) {
1669 // Arguments stored in registers.
1670 EVT RegVT = VA.getLocVT();
1673 const TargetRegisterClass *RC;
1675 if (RegVT == MVT::i32)
1676 RC = &ARM64::GPR32RegClass;
1677 else if (RegVT == MVT::i64)
1678 RC = &ARM64::GPR64RegClass;
1679 else if (RegVT == MVT::f32)
1680 RC = &ARM64::FPR32RegClass;
1681 else if (RegVT == MVT::f64 || RegVT == MVT::v1i64 ||
1682 RegVT == MVT::v1f64 || RegVT == MVT::v2i32 ||
1683 RegVT == MVT::v4i16 || RegVT == MVT::v8i8)
1684 RC = &ARM64::FPR64RegClass;
1685 else if (RegVT == MVT::f128 ||RegVT == MVT::v2i64 ||
1686 RegVT == MVT::v4i32||RegVT == MVT::v8i16 ||
1687 RegVT == MVT::v16i8)
1688 RC = &ARM64::FPR128RegClass;
1690 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1692 // Transform the arguments in physical registers into virtual ones.
1693 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1694 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
1696 // If this is an 8, 16 or 32-bit value, it is really passed promoted
1697 // to 64 bits. Insert an assert[sz]ext to capture this, then
1698 // truncate to the right size.
1699 switch (VA.getLocInfo()) {
1701 llvm_unreachable("Unknown loc info!");
1702 case CCValAssign::Full:
1704 case CCValAssign::BCvt:
1705 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
1707 case CCValAssign::SExt:
1708 ArgValue = DAG.getNode(ISD::AssertSext, DL, RegVT, ArgValue,
1709 DAG.getValueType(VA.getValVT()));
1710 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
1712 case CCValAssign::ZExt:
1713 ArgValue = DAG.getNode(ISD::AssertZext, DL, RegVT, ArgValue,
1714 DAG.getValueType(VA.getValVT()));
1715 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
1719 InVals.push_back(ArgValue);
1721 } else { // VA.isRegLoc()
1722 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
1723 unsigned ArgOffset = VA.getLocMemOffset();
1724 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1725 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset, true);
1727 // Create load nodes to retrieve arguments from the stack.
1728 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1729 InVals.push_back(DAG.getLoad(VA.getValVT(), DL, Chain, FIN,
1730 MachinePointerInfo::getFixedStack(FI), false,
1737 if (!Subtarget->isTargetDarwin()) {
1738 // The AAPCS variadic function ABI is identical to the non-variadic
1739 // one. As a result there may be more arguments in registers and we should
1740 // save them for future reference.
1741 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
1744 ARM64FunctionInfo *AFI = MF.getInfo<ARM64FunctionInfo>();
1745 // This will point to the next argument passed via stack.
1746 unsigned StackOffset = CCInfo.getNextStackOffset();
1747 // We currently pass all varargs at 8-byte alignment.
1748 StackOffset = ((StackOffset + 7) & ~7);
1749 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
1755 void ARM64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
1756 SelectionDAG &DAG, SDLoc DL,
1757 SDValue &Chain) const {
1758 MachineFunction &MF = DAG.getMachineFunction();
1759 MachineFrameInfo *MFI = MF.getFrameInfo();
1760 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
1762 SmallVector<SDValue, 8> MemOps;
1764 static const MCPhysReg GPRArgRegs[] = { ARM64::X0, ARM64::X1, ARM64::X2,
1765 ARM64::X3, ARM64::X4, ARM64::X5,
1766 ARM64::X6, ARM64::X7 };
1767 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
1768 unsigned FirstVariadicGPR =
1769 CCInfo.getFirstUnallocated(GPRArgRegs, NumGPRArgRegs);
1771 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
1773 if (GPRSaveSize != 0) {
1774 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
1776 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
1778 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
1779 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &ARM64::GPR64RegClass);
1780 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
1782 DAG.getStore(Val.getValue(1), DL, Val, FIN,
1783 MachinePointerInfo::getStack(i * 8), false, false, 0);
1784 MemOps.push_back(Store);
1785 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
1786 DAG.getConstant(8, getPointerTy()));
1789 FuncInfo->setVarArgsGPRIndex(GPRIdx);
1790 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
1792 if (Subtarget->hasFPARMv8()) {
1793 static const MCPhysReg FPRArgRegs[] = { ARM64::Q0, ARM64::Q1, ARM64::Q2,
1794 ARM64::Q3, ARM64::Q4, ARM64::Q5,
1795 ARM64::Q6, ARM64::Q7 };
1796 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
1797 unsigned FirstVariadicFPR =
1798 CCInfo.getFirstUnallocated(FPRArgRegs, NumFPRArgRegs);
1800 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
1802 if (FPRSaveSize != 0) {
1803 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
1805 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
1807 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
1808 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &ARM64::FPR128RegClass);
1809 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::v2i64);
1811 DAG.getStore(Val.getValue(1), DL, Val, FIN,
1812 MachinePointerInfo::getStack(i * 16), false, false, 0);
1813 MemOps.push_back(Store);
1814 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
1815 DAG.getConstant(16, getPointerTy()));
1818 FuncInfo->setVarArgsFPRIndex(FPRIdx);
1819 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
1822 if (!MemOps.empty()) {
1823 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
1827 /// LowerCallResult - Lower the result values of a call into the
1828 /// appropriate copies out of appropriate physical registers.
1829 SDValue ARM64TargetLowering::LowerCallResult(
1830 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
1831 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
1832 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
1833 SDValue ThisVal) const {
1834 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
1835 : RetCC_ARM64_AAPCS;
1836 // Assign locations to each value returned by this call.
1837 SmallVector<CCValAssign, 16> RVLocs;
1838 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1839 getTargetMachine(), RVLocs, *DAG.getContext());
1840 CCInfo.AnalyzeCallResult(Ins, RetCC);
1842 // Copy all of the result registers out of their specified physreg.
1843 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1844 CCValAssign VA = RVLocs[i];
1846 // Pass 'this' value directly from the argument to return value, to avoid
1847 // reg unit interference
1848 if (i == 0 && isThisReturn) {
1849 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
1850 "unexpected return calling convention register assignment");
1851 InVals.push_back(ThisVal);
1856 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
1857 Chain = Val.getValue(1);
1858 InFlag = Val.getValue(2);
1860 switch (VA.getLocInfo()) {
1862 llvm_unreachable("Unknown loc info!");
1863 case CCValAssign::Full:
1865 case CCValAssign::BCvt:
1866 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
1870 InVals.push_back(Val);
1876 bool ARM64TargetLowering::isEligibleForTailCallOptimization(
1877 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
1878 bool isCalleeStructRet, bool isCallerStructRet,
1879 const SmallVectorImpl<ISD::OutputArg> &Outs,
1880 const SmallVectorImpl<SDValue> &OutVals,
1881 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
1882 // Look for obvious safe cases to perform tail call optimization that do not
1883 // require ABI changes. This is what gcc calls sibcall.
1885 // Do not sibcall optimize vararg calls unless the call site is not passing
1887 if (isVarArg && !Outs.empty())
1890 // Also avoid sibcall optimization if either caller or callee uses struct
1891 // return semantics.
1892 if (isCalleeStructRet || isCallerStructRet)
1895 // Note that currently ARM64 "C" calling convention and "Fast" calling
1896 // convention are compatible. If/when that ever changes, we'll need to
1897 // add checks here to make sure any interactions are OK.
1899 // If the callee takes no arguments then go on to check the results of the
1901 if (!Outs.empty()) {
1902 // Check if stack adjustment is needed. For now, do not do this if any
1903 // argument is passed on the stack.
1904 SmallVector<CCValAssign, 16> ArgLocs;
1905 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1906 getTargetMachine(), ArgLocs, *DAG.getContext());
1907 CCAssignFn *AssignFn = CCAssignFnForCall(CalleeCC, /*IsVarArg=*/false);
1908 CCInfo.AnalyzeCallOperands(Outs, AssignFn);
1909 if (CCInfo.getNextStackOffset()) {
1910 // Check if the arguments are already laid out in the right way as
1911 // the caller's fixed stack objects.
1912 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
1913 ++i, ++realArgIdx) {
1914 CCValAssign &VA = ArgLocs[i];
1915 if (VA.getLocInfo() == CCValAssign::Indirect)
1917 if (VA.needsCustom()) {
1918 // Just don't handle anything that needs custom adjustments for now.
1919 // If need be, we can revisit later, but we shouldn't ever end up
1922 } else if (!VA.isRegLoc()) {
1923 // Likewise, don't try to handle stack based arguments for the
1933 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
1934 /// and add input and output parameter nodes.
1935 SDValue ARM64TargetLowering::LowerCall(CallLoweringInfo &CLI,
1936 SmallVectorImpl<SDValue> &InVals) const {
1937 SelectionDAG &DAG = CLI.DAG;
1939 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
1940 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
1941 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
1942 SDValue Chain = CLI.Chain;
1943 SDValue Callee = CLI.Callee;
1944 bool &IsTailCall = CLI.IsTailCall;
1945 CallingConv::ID CallConv = CLI.CallConv;
1946 bool IsVarArg = CLI.IsVarArg;
1948 MachineFunction &MF = DAG.getMachineFunction();
1949 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1950 bool IsThisReturn = false;
1952 // If tail calls are explicitly disabled, make sure not to use them.
1953 if (!EnableARM64TailCalls)
1957 // Check if it's really possible to do a tail call.
1958 IsTailCall = isEligibleForTailCallOptimization(
1959 Callee, CallConv, IsVarArg, IsStructRet,
1960 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
1961 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
1962 report_fatal_error("failed to perform tail call elimination on a call "
1963 "site marked musttail");
1964 // We don't support GuaranteedTailCallOpt, only automatically
1965 // detected sibcalls.
1966 // FIXME: Re-evaluate. Is this true? Should it be true?
1971 // Analyze operands of the call, assigning locations to each operand.
1972 SmallVector<CCValAssign, 16> ArgLocs;
1973 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
1974 getTargetMachine(), ArgLocs, *DAG.getContext());
1977 // Handle fixed and variable vector arguments differently.
1978 // Variable vector arguments always go into memory.
1979 unsigned NumArgs = Outs.size();
1981 for (unsigned i = 0; i != NumArgs; ++i) {
1982 MVT ArgVT = Outs[i].VT;
1983 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
1984 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
1985 /*IsVarArg=*/ !Outs[i].IsFixed);
1986 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
1987 assert(!Res && "Call operand has unhandled type");
1991 // At this point, Outs[].VT may already be promoted to i32. To correctly
1992 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
1993 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
1994 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
1995 // we use a special version of AnalyzeCallOperands to pass in ValVT and
1997 unsigned NumArgs = Outs.size();
1998 for (unsigned i = 0; i != NumArgs; ++i) {
1999 MVT ValVT = Outs[i].VT;
2000 // Get type of the original argument.
2001 EVT ActualVT = getValueType(CLI.Args[Outs[i].OrigArgIndex].Ty,
2002 /*AllowUnknown*/ true);
2003 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2004 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2005 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2007 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2009 else if (ActualMVT == MVT::i16)
2012 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2013 bool Res = AssignFn(i, ValVT, LocVT, CCValAssign::Full, ArgFlags, CCInfo);
2014 assert(!Res && "Call operand has unhandled type");
2019 // Get a count of how many bytes are to be pushed on the stack.
2020 unsigned NumBytes = CCInfo.getNextStackOffset();
2022 // Adjust the stack pointer for the new arguments...
2023 // These operations are automatically eliminated by the prolog/epilog pass
2026 DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true), DL);
2028 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, ARM64::SP, getPointerTy());
2030 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2031 SmallVector<SDValue, 8> MemOpChains;
2033 // Walk the register/memloc assignments, inserting copies/loads.
2034 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2035 ++i, ++realArgIdx) {
2036 CCValAssign &VA = ArgLocs[i];
2037 SDValue Arg = OutVals[realArgIdx];
2038 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2040 // Promote the value if needed.
2041 switch (VA.getLocInfo()) {
2043 llvm_unreachable("Unknown loc info!");
2044 case CCValAssign::Full:
2046 case CCValAssign::SExt:
2047 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
2049 case CCValAssign::ZExt:
2050 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2052 case CCValAssign::AExt:
2053 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
2055 case CCValAssign::BCvt:
2056 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2058 case CCValAssign::FPExt:
2059 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
2063 if (VA.isRegLoc()) {
2064 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
2065 assert(VA.getLocVT() == MVT::i64 &&
2066 "unexpected calling convention register assignment");
2067 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
2068 "unexpected use of 'returned'");
2069 IsThisReturn = true;
2071 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2073 assert(VA.isMemLoc());
2074 // There's no reason we can't support stack args w/ tailcall, but
2075 // we currently don't, so assert if we see one.
2076 assert(!IsTailCall && "stack argument with tail call!?");
2077 unsigned LocMemOffset = VA.getLocMemOffset();
2078 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2079 PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2081 if (Outs[i].Flags.isByVal()) {
2083 DAG.getConstant(Outs[i].Flags.getByValSize(), MVT::i64);
2084 SDValue Cpy = DAG.getMemcpy(
2085 Chain, DL, PtrOff, Arg, SizeNode, Outs[i].Flags.getByValAlign(),
2086 /*isVolatile = */ false,
2087 /*alwaysInline = */ false,
2088 MachinePointerInfo::getStack(LocMemOffset), MachinePointerInfo());
2090 MemOpChains.push_back(Cpy);
2092 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
2093 // promoted to a legal register type i32, we should truncate Arg back to
2095 if (Arg.getValueType().isSimple() &&
2096 Arg.getValueType().getSimpleVT() == MVT::i32 &&
2097 (VA.getLocVT() == MVT::i1 || VA.getLocVT() == MVT::i8 ||
2098 VA.getLocVT() == MVT::i16))
2099 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getLocVT(), Arg);
2101 SDValue Store = DAG.getStore(Chain, DL, Arg, PtrOff,
2102 MachinePointerInfo::getStack(LocMemOffset),
2104 MemOpChains.push_back(Store);
2109 if (!MemOpChains.empty())
2110 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
2112 // Build a sequence of copy-to-reg nodes chained together with token chain
2113 // and flag operands which copy the outgoing args into the appropriate regs.
2115 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2116 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[i].first,
2117 RegsToPass[i].second, InFlag);
2118 InFlag = Chain.getValue(1);
2121 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2122 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2123 // node so that legalize doesn't hack it.
2124 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
2125 Subtarget->isTargetMachO()) {
2126 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2127 const GlobalValue *GV = G->getGlobal();
2128 bool InternalLinkage = GV->hasInternalLinkage();
2129 if (InternalLinkage)
2130 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2132 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0,
2134 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2136 } else if (ExternalSymbolSDNode *S =
2137 dyn_cast<ExternalSymbolSDNode>(Callee)) {
2138 const char *Sym = S->getSymbol();
2140 DAG.getTargetExternalSymbol(Sym, getPointerTy(), ARM64II::MO_GOT);
2141 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2143 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2144 const GlobalValue *GV = G->getGlobal();
2145 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2146 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2147 const char *Sym = S->getSymbol();
2148 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), 0);
2151 std::vector<SDValue> Ops;
2152 Ops.push_back(Chain);
2153 Ops.push_back(Callee);
2155 // Add argument registers to the end of the list so that they are known live
2157 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2158 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2159 RegsToPass[i].second.getValueType()));
2161 // Add a register mask operand representing the call-preserved registers.
2162 const uint32_t *Mask;
2163 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2164 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2166 // For 'this' returns, use the X0-preserving mask if applicable
2167 Mask = ARI->getThisReturnPreservedMask(CallConv);
2169 IsThisReturn = false;
2170 Mask = ARI->getCallPreservedMask(CallConv);
2173 Mask = ARI->getCallPreservedMask(CallConv);
2175 assert(Mask && "Missing call preserved mask for calling convention");
2176 Ops.push_back(DAG.getRegisterMask(Mask));
2178 if (InFlag.getNode())
2179 Ops.push_back(InFlag);
2181 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2183 // If we're doing a tall call, use a TC_RETURN here rather than an
2184 // actual call instruction.
2186 return DAG.getNode(ARM64ISD::TC_RETURN, DL, NodeTys, Ops);
2188 // Returns a chain and a flag for retval copy to use.
2189 Chain = DAG.getNode(ARM64ISD::CALL, DL, NodeTys, Ops);
2190 InFlag = Chain.getValue(1);
2192 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2193 DAG.getIntPtrConstant(0, true), InFlag, DL);
2195 InFlag = Chain.getValue(1);
2197 // Handle result values, copying them out of physregs into vregs that we
2199 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2200 InVals, IsThisReturn,
2201 IsThisReturn ? OutVals[0] : SDValue());
2204 bool ARM64TargetLowering::CanLowerReturn(
2205 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2206 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2207 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2208 : RetCC_ARM64_AAPCS;
2209 SmallVector<CCValAssign, 16> RVLocs;
2210 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2211 return CCInfo.CheckReturn(Outs, RetCC);
2215 ARM64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2217 const SmallVectorImpl<ISD::OutputArg> &Outs,
2218 const SmallVectorImpl<SDValue> &OutVals,
2219 SDLoc DL, SelectionDAG &DAG) const {
2220 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2221 : RetCC_ARM64_AAPCS;
2222 SmallVector<CCValAssign, 16> RVLocs;
2223 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2224 getTargetMachine(), RVLocs, *DAG.getContext());
2225 CCInfo.AnalyzeReturn(Outs, RetCC);
2227 // Copy the result values into the output registers.
2229 SmallVector<SDValue, 4> RetOps(1, Chain);
2230 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
2231 ++i, ++realRVLocIdx) {
2232 CCValAssign &VA = RVLocs[i];
2233 assert(VA.isRegLoc() && "Can only return in registers!");
2234 SDValue Arg = OutVals[realRVLocIdx];
2236 switch (VA.getLocInfo()) {
2238 llvm_unreachable("Unknown loc info!");
2239 case CCValAssign::Full:
2241 case CCValAssign::BCvt:
2242 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2246 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
2247 Flag = Chain.getValue(1);
2248 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2251 RetOps[0] = Chain; // Update chain.
2253 // Add the flag if we have it.
2255 RetOps.push_back(Flag);
2257 return DAG.getNode(ARM64ISD::RET_FLAG, DL, MVT::Other, RetOps);
2260 //===----------------------------------------------------------------------===//
2261 // Other Lowering Code
2262 //===----------------------------------------------------------------------===//
2264 SDValue ARM64TargetLowering::LowerGlobalAddress(SDValue Op,
2265 SelectionDAG &DAG) const {
2266 EVT PtrVT = getPointerTy();
2268 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2269 unsigned char OpFlags =
2270 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
2272 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
2273 "unexpected offset in global node");
2275 // This also catched the large code model case for Darwin.
2276 if ((OpFlags & ARM64II::MO_GOT) != 0) {
2277 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
2278 // FIXME: Once remat is capable of dealing with instructions with register
2279 // operands, expand this into two nodes instead of using a wrapper node.
2280 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
2283 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2284 const unsigned char MO_NC = ARM64II::MO_NC;
2286 ARM64ISD::WrapperLarge, DL, PtrVT,
2287 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G3),
2288 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
2289 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
2290 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
2292 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
2293 // the only correct model on Darwin.
2294 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2295 OpFlags | ARM64II::MO_PAGE);
2296 unsigned char LoFlags = OpFlags | ARM64II::MO_PAGEOFF | ARM64II::MO_NC;
2297 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
2299 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
2300 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
2304 /// \brief Convert a TLS address reference into the correct sequence of loads
2305 /// and calls to compute the variable's address (for Darwin, currently) and
2306 /// return an SDValue containing the final node.
2308 /// Darwin only has one TLS scheme which must be capable of dealing with the
2309 /// fully general situation, in the worst case. This means:
2310 /// + "extern __thread" declaration.
2311 /// + Defined in a possibly unknown dynamic library.
2313 /// The general system is that each __thread variable has a [3 x i64] descriptor
2314 /// which contains information used by the runtime to calculate the address. The
2315 /// only part of this the compiler needs to know about is the first xword, which
2316 /// contains a function pointer that must be called with the address of the
2317 /// entire descriptor in "x0".
2319 /// Since this descriptor may be in a different unit, in general even the
2320 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
2322 /// adrp x0, _var@TLVPPAGE
2323 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
2324 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
2325 /// ; the function pointer
2326 /// blr x1 ; Uses descriptor address in x0
2327 /// ; Address of _var is now in x0.
2329 /// If the address of _var's descriptor *is* known to the linker, then it can
2330 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
2331 /// a slight efficiency gain.
2333 ARM64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
2334 SelectionDAG &DAG) const {
2335 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
2338 MVT PtrVT = getPointerTy();
2339 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2342 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2343 SDValue DescAddr = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TLVPAddr);
2345 // The first entry in the descriptor is a function pointer that we must call
2346 // to obtain the address of the variable.
2347 SDValue Chain = DAG.getEntryNode();
2348 SDValue FuncTLVGet =
2349 DAG.getLoad(MVT::i64, DL, Chain, DescAddr, MachinePointerInfo::getGOT(),
2350 false, true, true, 8);
2351 Chain = FuncTLVGet.getValue(1);
2353 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2354 MFI->setAdjustsStack(true);
2356 // TLS calls preserve all registers except those that absolutely must be
2357 // trashed: X0 (it takes an argument), LR (it's a call) and CPSR (let's not be
2359 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2360 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2361 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2363 // Finally, we can make the call. This is just a degenerate version of a
2364 // normal ARM64 call node: x0 takes the address of the descriptor, and returns
2365 // the address of the variable in this thread.
2366 Chain = DAG.getCopyToReg(Chain, DL, ARM64::X0, DescAddr, SDValue());
2367 Chain = DAG.getNode(ARM64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
2368 Chain, FuncTLVGet, DAG.getRegister(ARM64::X0, MVT::i64),
2369 DAG.getRegisterMask(Mask), Chain.getValue(1));
2370 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Chain.getValue(1));
2373 /// When accessing thread-local variables under either the general-dynamic or
2374 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
2375 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
2376 /// is a function pointer to carry out the resolution. This function takes the
2377 /// address of the descriptor in X0 and returns the TPIDR_EL0 offset in X0. All
2378 /// other registers (except LR, CPSR) are preserved.
2380 /// Thus, the ideal call sequence on AArch64 is:
2382 /// adrp x0, :tlsdesc:thread_var
2383 /// ldr x8, [x0, :tlsdesc_lo12:thread_var]
2384 /// add x0, x0, :tlsdesc_lo12:thread_var
2385 /// .tlsdesccall thread_var
2387 /// (TPIDR_EL0 offset now in x0).
2389 /// The ".tlsdesccall" directive instructs the assembler to insert a particular
2390 /// relocation to help the linker relax this sequence if it turns out to be too
2393 /// FIXME: we currently produce an extra, duplicated, ADRP instruction, but this
2395 SDValue ARM64TargetLowering::LowerELFTLSDescCall(SDValue SymAddr,
2396 SDValue DescAddr, SDLoc DL,
2397 SelectionDAG &DAG) const {
2398 EVT PtrVT = getPointerTy();
2400 // The function we need to call is simply the first entry in the GOT for this
2401 // descriptor, load it in preparation.
2402 SDValue Func = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, SymAddr);
2404 // TLS calls preserve all registers except those that absolutely must be
2405 // trashed: X0 (it takes an argument), LR (it's a call) and CPSR (let's not be
2407 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2408 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2409 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2411 // The function takes only one argument: the address of the descriptor itself
2413 SDValue Glue, Chain;
2414 Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, ARM64::X0, DescAddr, Glue);
2415 Glue = Chain.getValue(1);
2417 // We're now ready to populate the argument list, as with a normal call:
2418 SmallVector<SDValue, 6> Ops;
2419 Ops.push_back(Chain);
2420 Ops.push_back(Func);
2421 Ops.push_back(SymAddr);
2422 Ops.push_back(DAG.getRegister(ARM64::X0, PtrVT));
2423 Ops.push_back(DAG.getRegisterMask(Mask));
2424 Ops.push_back(Glue);
2426 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2427 Chain = DAG.getNode(ARM64ISD::TLSDESC_CALL, DL, NodeTys, Ops);
2428 Glue = Chain.getValue(1);
2430 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Glue);
2433 SDValue ARM64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
2434 SelectionDAG &DAG) const {
2435 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
2436 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
2437 "ELF TLS only supported in small memory model");
2438 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2440 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
2443 EVT PtrVT = getPointerTy();
2445 const GlobalValue *GV = GA->getGlobal();
2447 SDValue ThreadBase = DAG.getNode(ARM64ISD::THREAD_POINTER, DL, PtrVT);
2449 if (Model == TLSModel::LocalExec) {
2450 SDValue HiVar = DAG.getTargetGlobalAddress(
2451 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2452 SDValue LoVar = DAG.getTargetGlobalAddress(
2453 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2455 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2456 DAG.getTargetConstant(16, MVT::i32)),
2458 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, TPOff, LoVar,
2459 DAG.getTargetConstant(0, MVT::i32)),
2461 } else if (Model == TLSModel::InitialExec) {
2462 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2463 TPOff = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TPOff);
2464 } else if (Model == TLSModel::LocalDynamic) {
2465 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
2466 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
2467 // the beginning of the module's TLS region, followed by a DTPREL offset
2470 // These accesses will need deduplicating if there's more than one.
2471 ARM64FunctionInfo *MFI =
2472 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
2473 MFI->incNumLocalDynamicTLSAccesses();
2475 // Accesses used in this sequence go via the TLS descriptor which lives in
2476 // the GOT. Prepare an address we can use to handle this.
2477 SDValue HiDesc = DAG.getTargetExternalSymbol(
2478 "_TLS_MODULE_BASE_", PtrVT, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2479 SDValue LoDesc = DAG.getTargetExternalSymbol(
2480 "_TLS_MODULE_BASE_", PtrVT,
2481 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2483 // First argument to the descriptor call is the address of the descriptor
2485 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2486 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2488 // The call needs a relocation too for linker relaxation. It doesn't make
2489 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2491 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2494 // Now we can calculate the offset from TPIDR_EL0 to this module's
2495 // thread-local area.
2496 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2498 // Now use :dtprel_whatever: operations to calculate this variable's offset
2499 // in its thread-storage area.
2500 SDValue HiVar = DAG.getTargetGlobalAddress(
2501 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2502 SDValue LoVar = DAG.getTargetGlobalAddress(
2503 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2506 SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2507 DAG.getTargetConstant(16, MVT::i32)),
2509 DTPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, DTPOff, LoVar,
2510 DAG.getTargetConstant(0, MVT::i32)),
2513 TPOff = DAG.getNode(ISD::ADD, DL, PtrVT, TPOff, DTPOff);
2514 } else if (Model == TLSModel::GeneralDynamic) {
2515 // Accesses used in this sequence go via the TLS descriptor which lives in
2516 // the GOT. Prepare an address we can use to handle this.
2517 SDValue HiDesc = DAG.getTargetGlobalAddress(
2518 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2519 SDValue LoDesc = DAG.getTargetGlobalAddress(
2521 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2523 // First argument to the descriptor call is the address of the descriptor
2525 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2526 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2528 // The call needs a relocation too for linker relaxation. It doesn't make
2529 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2532 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2534 // Finally we can make a call to calculate the offset from tpidr_el0.
2535 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2537 llvm_unreachable("Unsupported ELF TLS access model");
2539 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
2542 SDValue ARM64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
2543 SelectionDAG &DAG) const {
2544 if (Subtarget->isTargetDarwin())
2545 return LowerDarwinGlobalTLSAddress(Op, DAG);
2546 else if (Subtarget->isTargetELF())
2547 return LowerELFGlobalTLSAddress(Op, DAG);
2549 llvm_unreachable("Unexpected platform trying to use TLS");
2551 SDValue ARM64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2552 SDValue Chain = Op.getOperand(0);
2553 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2554 SDValue LHS = Op.getOperand(2);
2555 SDValue RHS = Op.getOperand(3);
2556 SDValue Dest = Op.getOperand(4);
2559 // Handle f128 first, since lowering it will result in comparing the return
2560 // value of a libcall against zero, which is just what the rest of LowerBR_CC
2561 // is expecting to deal with.
2562 if (LHS.getValueType() == MVT::f128) {
2563 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2565 // If softenSetCCOperands returned a scalar, we need to compare the result
2566 // against zero to select between true and false values.
2567 if (!RHS.getNode()) {
2568 RHS = DAG.getConstant(0, LHS.getValueType());
2573 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
2575 unsigned Opc = LHS.getOpcode();
2576 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
2577 cast<ConstantSDNode>(RHS)->isOne() &&
2578 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
2579 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
2580 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
2581 "Unexpected condition code.");
2582 // Only lower legal XALUO ops.
2583 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
2586 // The actual operation with overflow check.
2587 ARM64CC::CondCode OFCC;
2588 SDValue Value, Overflow;
2589 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, LHS.getValue(0), DAG);
2591 if (CC == ISD::SETNE)
2592 OFCC = getInvertedCondCode(OFCC);
2593 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
2595 return DAG.getNode(ARM64ISD::BRCOND, SDLoc(LHS), MVT::Other, Chain, Dest,
2599 if (LHS.getValueType().isInteger()) {
2600 assert((LHS.getValueType() == RHS.getValueType()) &&
2601 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
2603 // If the RHS of the comparison is zero, we can potentially fold this
2604 // to a specialized branch.
2605 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
2606 if (RHSC && RHSC->getZExtValue() == 0) {
2607 if (CC == ISD::SETEQ) {
2608 // See if we can use a TBZ to fold in an AND as well.
2609 // TBZ has a smaller branch displacement than CBZ. If the offset is
2610 // out of bounds, a late MI-layer pass rewrites branches.
2611 // 403.gcc is an example that hits this case.
2612 if (LHS.getOpcode() == ISD::AND &&
2613 isa<ConstantSDNode>(LHS.getOperand(1)) &&
2614 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
2615 SDValue Test = LHS.getOperand(0);
2616 uint64_t Mask = LHS.getConstantOperandVal(1);
2618 // TBZ only operates on i64's, but the ext should be free.
2619 if (Test.getValueType() == MVT::i32)
2620 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
2622 return DAG.getNode(ARM64ISD::TBZ, dl, MVT::Other, Chain, Test,
2623 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
2626 return DAG.getNode(ARM64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
2627 } else if (CC == ISD::SETNE) {
2628 // See if we can use a TBZ to fold in an AND as well.
2629 // TBZ has a smaller branch displacement than CBZ. If the offset is
2630 // out of bounds, a late MI-layer pass rewrites branches.
2631 // 403.gcc is an example that hits this case.
2632 if (LHS.getOpcode() == ISD::AND &&
2633 isa<ConstantSDNode>(LHS.getOperand(1)) &&
2634 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
2635 SDValue Test = LHS.getOperand(0);
2636 uint64_t Mask = LHS.getConstantOperandVal(1);
2638 // TBNZ only operates on i64's, but the ext should be free.
2639 if (Test.getValueType() == MVT::i32)
2640 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
2642 return DAG.getNode(ARM64ISD::TBNZ, dl, MVT::Other, Chain, Test,
2643 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
2646 return DAG.getNode(ARM64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
2651 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
2652 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
2656 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2658 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
2659 // clean. Some of them require two branches to implement.
2660 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
2661 ARM64CC::CondCode CC1, CC2;
2662 changeFPCCToARM64CC(CC, CC1, CC2);
2663 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
2665 DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
2666 if (CC2 != ARM64CC::AL) {
2667 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
2668 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
2675 SDValue ARM64TargetLowering::LowerFCOPYSIGN(SDValue Op,
2676 SelectionDAG &DAG) const {
2677 EVT VT = Op.getValueType();
2680 SDValue In1 = Op.getOperand(0);
2681 SDValue In2 = Op.getOperand(1);
2682 EVT SrcVT = In2.getValueType();
2684 if (SrcVT == MVT::f32 && VT == MVT::f64)
2685 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
2686 else if (SrcVT == MVT::f64 && VT == MVT::f32)
2687 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0));
2689 // FIXME: Src type is different, bail out for now. Can VT really be a
2696 SDValue EltMask, VecVal1, VecVal2;
2697 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
2700 EltMask = DAG.getConstant(0x80000000ULL, EltVT);
2702 if (!VT.isVector()) {
2703 VecVal1 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
2704 DAG.getUNDEF(VecVT), In1);
2705 VecVal2 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
2706 DAG.getUNDEF(VecVT), In2);
2708 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
2709 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
2711 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
2715 // We want to materialize a mask with the the high bit set, but the AdvSIMD
2716 // immediate moves cannot materialize that in a single instruction for
2717 // 64-bit elements. Instead, materialize zero and then negate it.
2718 EltMask = DAG.getConstant(0, EltVT);
2720 if (!VT.isVector()) {
2721 VecVal1 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
2722 DAG.getUNDEF(VecVT), In1);
2723 VecVal2 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
2724 DAG.getUNDEF(VecVT), In2);
2726 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
2727 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
2730 llvm_unreachable("Invalid type for copysign!");
2733 std::vector<SDValue> BuildVectorOps;
2734 for (unsigned i = 0; i < VecVT.getVectorNumElements(); ++i)
2735 BuildVectorOps.push_back(EltMask);
2737 SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT, BuildVectorOps);
2739 // If we couldn't materialize the mask above, then the mask vector will be
2740 // the zero vector, and we need to negate it here.
2741 if (VT == MVT::f64 || VT == MVT::v2f64) {
2742 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
2743 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
2744 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
2748 DAG.getNode(ARM64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
2751 return DAG.getTargetExtractSubreg(ARM64::ssub, DL, VT, Sel);
2752 else if (VT == MVT::f64)
2753 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, VT, Sel);
2755 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
2758 SDValue ARM64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
2759 if (DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
2760 AttributeSet::FunctionIndex, Attribute::NoImplicitFloat))
2763 // While there is no integer popcount instruction, it can
2764 // be more efficiently lowered to the following sequence that uses
2765 // AdvSIMD registers/instructions as long as the copies to/from
2766 // the AdvSIMD registers are cheap.
2767 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
2768 // CNT V0.8B, V0.8B // 8xbyte pop-counts
2769 // ADDV B0, V0.8B // sum 8xbyte pop-counts
2770 // UMOV X0, V0.B[0] // copy byte result back to integer reg
2771 SDValue Val = Op.getOperand(0);
2773 EVT VT = Op.getValueType();
2774 SDValue ZeroVec = DAG.getUNDEF(MVT::v8i8);
2777 if (VT == MVT::i32) {
2778 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Val);
2780 DAG.getTargetInsertSubreg(ARM64::ssub, DL, MVT::v8i8, ZeroVec, VecVal);
2782 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
2785 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, VecVal);
2786 SDValue UaddLV = DAG.getNode(
2787 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
2788 DAG.getConstant(Intrinsic::arm64_neon_uaddlv, MVT::i32), CtPop);
2791 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
2795 SDValue ARM64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
2797 if (Op.getValueType().isVector())
2798 return LowerVSETCC(Op, DAG);
2800 SDValue LHS = Op.getOperand(0);
2801 SDValue RHS = Op.getOperand(1);
2802 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
2805 // We chose ZeroOrOneBooleanContents, so use zero and one.
2806 EVT VT = Op.getValueType();
2807 SDValue TVal = DAG.getConstant(1, VT);
2808 SDValue FVal = DAG.getConstant(0, VT);
2810 // Handle f128 first, since one possible outcome is a normal integer
2811 // comparison which gets picked up by the next if statement.
2812 if (LHS.getValueType() == MVT::f128) {
2813 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2815 // If softenSetCCOperands returned a scalar, use it.
2816 if (!RHS.getNode()) {
2817 assert(LHS.getValueType() == Op.getValueType() &&
2818 "Unexpected setcc expansion!");
2823 if (LHS.getValueType().isInteger()) {
2826 getARM64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
2828 // Note that we inverted the condition above, so we reverse the order of
2829 // the true and false operands here. This will allow the setcc to be
2830 // matched to a single CSINC instruction.
2831 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
2834 // Now we know we're dealing with FP values.
2835 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2837 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
2838 // and do the comparison.
2839 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
2841 ARM64CC::CondCode CC1, CC2;
2842 changeFPCCToARM64CC(CC, CC1, CC2);
2843 if (CC2 == ARM64CC::AL) {
2844 changeFPCCToARM64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
2845 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
2847 // Note that we inverted the condition above, so we reverse the order of
2848 // the true and false operands here. This will allow the setcc to be
2849 // matched to a single CSINC instruction.
2850 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
2852 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
2853 // clean. Some of them require two CSELs to implement. As is in this case,
2854 // we emit the first CSEL and then emit a second using the output of the
2855 // first as the RHS. We're effectively OR'ing the two CC's together.
2857 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
2858 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
2859 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
2861 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
2862 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
2866 /// A SELECT_CC operation is really some kind of max or min if both values being
2867 /// compared are, in some sense, equal to the results in either case. However,
2868 /// it is permissible to compare f32 values and produce directly extended f64
2871 /// Extending the comparison operands would also be allowed, but is less likely
2872 /// to happen in practice since their use is right here. Note that truncate
2873 /// operations would *not* be semantically equivalent.
2874 static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) {
2878 ConstantFPSDNode *CCmp = dyn_cast<ConstantFPSDNode>(Cmp);
2879 ConstantFPSDNode *CResult = dyn_cast<ConstantFPSDNode>(Result);
2880 if (CCmp && CResult && Cmp.getValueType() == MVT::f32 &&
2881 Result.getValueType() == MVT::f64) {
2883 APFloat CmpVal = CCmp->getValueAPF();
2884 CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy);
2885 return CResult->getValueAPF().bitwiseIsEqual(CmpVal);
2888 return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp;
2891 SDValue ARM64TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2892 SDValue CC = Op->getOperand(0);
2893 SDValue TVal = Op->getOperand(1);
2894 SDValue FVal = Op->getOperand(2);
2897 unsigned Opc = CC.getOpcode();
2898 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
2900 if (CC.getResNo() == 1 &&
2901 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
2902 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
2903 // Only lower legal XALUO ops.
2904 if (!DAG.getTargetLoweringInfo().isTypeLegal(CC->getValueType(0)))
2907 ARM64CC::CondCode OFCC;
2908 SDValue Value, Overflow;
2909 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, CC.getValue(0), DAG);
2910 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
2912 return DAG.getNode(ARM64ISD::CSEL, DL, Op.getValueType(), TVal, FVal, CCVal,
2916 if (CC.getOpcode() == ISD::SETCC)
2917 return DAG.getSelectCC(DL, CC.getOperand(0), CC.getOperand(1), TVal, FVal,
2918 cast<CondCodeSDNode>(CC.getOperand(2))->get());
2920 return DAG.getSelectCC(DL, CC, DAG.getConstant(0, CC.getValueType()), TVal,
2924 SDValue ARM64TargetLowering::LowerSELECT_CC(SDValue Op,
2925 SelectionDAG &DAG) const {
2926 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2927 SDValue LHS = Op.getOperand(0);
2928 SDValue RHS = Op.getOperand(1);
2929 SDValue TVal = Op.getOperand(2);
2930 SDValue FVal = Op.getOperand(3);
2933 // Handle f128 first, because it will result in a comparison of some RTLIB
2934 // call result against zero.
2935 if (LHS.getValueType() == MVT::f128) {
2936 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2938 // If softenSetCCOperands returned a scalar, we need to compare the result
2939 // against zero to select between true and false values.
2940 if (!RHS.getNode()) {
2941 RHS = DAG.getConstant(0, LHS.getValueType());
2946 // Handle integers first.
2947 if (LHS.getValueType().isInteger()) {
2948 assert((LHS.getValueType() == RHS.getValueType()) &&
2949 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
2951 unsigned Opcode = ARM64ISD::CSEL;
2953 // If both the TVal and the FVal are constants, see if we can swap them in
2954 // order to for a CSINV or CSINC out of them.
2955 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
2956 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
2958 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
2959 std::swap(TVal, FVal);
2960 std::swap(CTVal, CFVal);
2961 CC = ISD::getSetCCInverse(CC, true);
2962 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
2963 std::swap(TVal, FVal);
2964 std::swap(CTVal, CFVal);
2965 CC = ISD::getSetCCInverse(CC, true);
2966 } else if (TVal.getOpcode() == ISD::XOR) {
2967 // If TVal is a NOT we want to swap TVal and FVal so that we can match
2968 // with a CSINV rather than a CSEL.
2969 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
2971 if (CVal && CVal->isAllOnesValue()) {
2972 std::swap(TVal, FVal);
2973 std::swap(CTVal, CFVal);
2974 CC = ISD::getSetCCInverse(CC, true);
2976 } else if (TVal.getOpcode() == ISD::SUB) {
2977 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
2978 // that we can match with a CSNEG rather than a CSEL.
2979 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
2981 if (CVal && CVal->isNullValue()) {
2982 std::swap(TVal, FVal);
2983 std::swap(CTVal, CFVal);
2984 CC = ISD::getSetCCInverse(CC, true);
2986 } else if (CTVal && CFVal) {
2987 const int64_t TrueVal = CTVal->getSExtValue();
2988 const int64_t FalseVal = CFVal->getSExtValue();
2991 // If both TVal and FVal are constants, see if FVal is the
2992 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
2993 // instead of a CSEL in that case.
2994 if (TrueVal == ~FalseVal) {
2995 Opcode = ARM64ISD::CSINV;
2996 } else if (TrueVal == -FalseVal) {
2997 Opcode = ARM64ISD::CSNEG;
2998 } else if (TVal.getValueType() == MVT::i32) {
2999 // If our operands are only 32-bit wide, make sure we use 32-bit
3000 // arithmetic for the check whether we can use CSINC. This ensures that
3001 // the addition in the check will wrap around properly in case there is
3002 // an overflow (which would not be the case if we do the check with
3003 // 64-bit arithmetic).
3004 const uint32_t TrueVal32 = CTVal->getZExtValue();
3005 const uint32_t FalseVal32 = CFVal->getZExtValue();
3007 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3008 Opcode = ARM64ISD::CSINC;
3010 if (TrueVal32 > FalseVal32) {
3014 // 64-bit check whether we can use CSINC.
3015 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3016 Opcode = ARM64ISD::CSINC;
3018 if (TrueVal > FalseVal) {
3023 // Swap TVal and FVal if necessary.
3025 std::swap(TVal, FVal);
3026 std::swap(CTVal, CFVal);
3027 CC = ISD::getSetCCInverse(CC, true);
3030 if (Opcode != ARM64ISD::CSEL) {
3031 // Drop FVal since we can get its value by simply inverting/negating
3038 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3040 EVT VT = Op.getValueType();
3041 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3044 // Now we know we're dealing with FP values.
3045 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3046 assert(LHS.getValueType() == RHS.getValueType());
3047 EVT VT = Op.getValueType();
3049 // Try to match this select into a max/min operation, which have dedicated
3050 // opcode in the instruction set.
3051 // NOTE: This is not correct in the presence of NaNs, so we only enable this
3053 if (getTargetMachine().Options.NoNaNsFPMath) {
3054 if (selectCCOpsAreFMaxCompatible(LHS, FVal) &&
3055 selectCCOpsAreFMaxCompatible(RHS, TVal)) {
3056 CC = ISD::getSetCCSwappedOperands(CC);
3057 std::swap(TVal, FVal);
3060 if (selectCCOpsAreFMaxCompatible(LHS, TVal) &&
3061 selectCCOpsAreFMaxCompatible(RHS, FVal)) {
3071 return DAG.getNode(ARM64ISD::FMAX, dl, VT, TVal, FVal);
3079 return DAG.getNode(ARM64ISD::FMIN, dl, VT, TVal, FVal);
3085 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3086 // and do the comparison.
3087 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3089 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3090 // clean. Some of them require two CSELs to implement.
3091 ARM64CC::CondCode CC1, CC2;
3092 changeFPCCToARM64CC(CC, CC1, CC2);
3093 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3094 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3096 // If we need a second CSEL, emit it, using the output of the first as the
3097 // RHS. We're effectively OR'ing the two CC's together.
3098 if (CC2 != ARM64CC::AL) {
3099 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3100 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3103 // Otherwise, return the output of the first CSEL.
3107 SDValue ARM64TargetLowering::LowerJumpTable(SDValue Op,
3108 SelectionDAG &DAG) const {
3109 // Jump table entries as PC relative offsets. No additional tweaking
3110 // is necessary here. Just get the address of the jump table.
3111 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3112 EVT PtrVT = getPointerTy();
3115 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3116 !Subtarget->isTargetMachO()) {
3117 const unsigned char MO_NC = ARM64II::MO_NC;
3119 ARM64ISD::WrapperLarge, DL, PtrVT,
3120 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G3),
3121 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G2 | MO_NC),
3122 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G1 | MO_NC),
3123 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G0 | MO_NC));
3126 SDValue Hi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_PAGE);
3127 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3128 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3129 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3130 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3133 SDValue ARM64TargetLowering::LowerConstantPool(SDValue Op,
3134 SelectionDAG &DAG) const {
3135 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3136 EVT PtrVT = getPointerTy();
3139 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3140 // Use the GOT for the large code model on iOS.
3141 if (Subtarget->isTargetMachO()) {
3142 SDValue GotAddr = DAG.getTargetConstantPool(
3143 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3145 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
3148 const unsigned char MO_NC = ARM64II::MO_NC;
3150 ARM64ISD::WrapperLarge, DL, PtrVT,
3151 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3152 CP->getOffset(), ARM64II::MO_G3),
3153 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3154 CP->getOffset(), ARM64II::MO_G2 | MO_NC),
3155 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3156 CP->getOffset(), ARM64II::MO_G1 | MO_NC),
3157 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3158 CP->getOffset(), ARM64II::MO_G0 | MO_NC));
3160 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
3161 // ELF, the only valid one on Darwin.
3163 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3164 CP->getOffset(), ARM64II::MO_PAGE);
3165 SDValue Lo = DAG.getTargetConstantPool(
3166 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3167 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3169 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3170 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3174 SDValue ARM64TargetLowering::LowerBlockAddress(SDValue Op,
3175 SelectionDAG &DAG) const {
3176 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
3177 EVT PtrVT = getPointerTy();
3179 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3180 !Subtarget->isTargetMachO()) {
3181 const unsigned char MO_NC = ARM64II::MO_NC;
3183 ARM64ISD::WrapperLarge, DL, PtrVT,
3184 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G3),
3185 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
3186 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
3187 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
3189 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGE);
3190 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGEOFF |
3192 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3193 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3197 SDValue ARM64TargetLowering::LowerDarwin_VASTART(SDValue Op,
3198 SelectionDAG &DAG) const {
3199 ARM64FunctionInfo *FuncInfo =
3200 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
3204 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3205 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3206 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
3207 MachinePointerInfo(SV), false, false, 0);
3210 SDValue ARM64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
3211 SelectionDAG &DAG) const {
3212 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3213 // Standard, section B.3.
3214 MachineFunction &MF = DAG.getMachineFunction();
3215 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
3218 SDValue Chain = Op.getOperand(0);
3219 SDValue VAList = Op.getOperand(1);
3220 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3221 SmallVector<SDValue, 4> MemOps;
3223 // void *__stack at offset 0
3225 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3226 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3227 MachinePointerInfo(SV), false, false, 8));
3229 // void *__gr_top at offset 8
3230 int GPRSize = FuncInfo->getVarArgsGPRSize();
3232 SDValue GRTop, GRTopAddr;
3234 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3235 DAG.getConstant(8, getPointerTy()));
3237 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), getPointerTy());
3238 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
3239 DAG.getConstant(GPRSize, getPointerTy()));
3241 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3242 MachinePointerInfo(SV, 8), false, false, 8));
3245 // void *__vr_top at offset 16
3246 int FPRSize = FuncInfo->getVarArgsFPRSize();
3248 SDValue VRTop, VRTopAddr;
3249 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3250 DAG.getConstant(16, getPointerTy()));
3252 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), getPointerTy());
3253 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
3254 DAG.getConstant(FPRSize, getPointerTy()));
3256 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3257 MachinePointerInfo(SV, 16), false, false, 8));
3260 // int __gr_offs at offset 24
3261 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3262 DAG.getConstant(24, getPointerTy()));
3263 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
3264 GROffsAddr, MachinePointerInfo(SV, 24), false,
3267 // int __vr_offs at offset 28
3268 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3269 DAG.getConstant(28, getPointerTy()));
3270 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
3271 VROffsAddr, MachinePointerInfo(SV, 28), false,
3274 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
3277 SDValue ARM64TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
3278 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
3279 : LowerAAPCS_VASTART(Op, DAG);
3282 SDValue ARM64TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
3283 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
3285 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
3286 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3287 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
3289 return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op), Op.getOperand(1),
3290 Op.getOperand(2), DAG.getConstant(VaListSize, MVT::i32),
3291 8, false, false, MachinePointerInfo(DestSV),
3292 MachinePointerInfo(SrcSV));
3295 SDValue ARM64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
3296 assert(Subtarget->isTargetDarwin() &&
3297 "automatic va_arg instruction only works on Darwin");
3299 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3300 EVT VT = Op.getValueType();
3302 SDValue Chain = Op.getOperand(0);
3303 SDValue Addr = Op.getOperand(1);
3304 unsigned Align = Op.getConstantOperandVal(3);
3306 SDValue VAList = DAG.getLoad(getPointerTy(), DL, Chain, Addr,
3307 MachinePointerInfo(V), false, false, false, 0);
3308 Chain = VAList.getValue(1);
3311 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
3312 VAList = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3313 DAG.getConstant(Align - 1, getPointerTy()));
3314 VAList = DAG.getNode(ISD::AND, DL, getPointerTy(), VAList,
3315 DAG.getConstant(-(int64_t)Align, getPointerTy()));
3318 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
3319 uint64_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
3321 // Scalar integer and FP values smaller than 64 bits are implicitly extended
3322 // up to 64 bits. At the very least, we have to increase the striding of the
3323 // vaargs list to match this, and for FP values we need to introduce
3324 // FP_ROUND nodes as well.
3325 if (VT.isInteger() && !VT.isVector())
3327 bool NeedFPTrunc = false;
3328 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
3333 // Increment the pointer, VAList, to the next vaarg
3334 SDValue VANext = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3335 DAG.getConstant(ArgSize, getPointerTy()));
3336 // Store the incremented VAList to the legalized pointer
3337 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
3340 // Load the actual argument out of the pointer VAList
3342 // Load the value as an f64.
3343 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
3344 MachinePointerInfo(), false, false, false, 0);
3345 // Round the value down to an f32.
3346 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
3347 DAG.getIntPtrConstant(1));
3348 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
3349 // Merge the rounded value with the chain output of the load.
3350 return DAG.getMergeValues(Ops, 2, DL);
3353 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
3357 SDValue ARM64TargetLowering::LowerFRAMEADDR(SDValue Op,
3358 SelectionDAG &DAG) const {
3359 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3360 MFI->setFrameAddressIsTaken(true);
3362 EVT VT = Op.getValueType();
3364 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3365 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, ARM64::FP, VT);
3367 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
3368 MachinePointerInfo(), false, false, false, 0);
3372 SDValue ARM64TargetLowering::LowerRETURNADDR(SDValue Op,
3373 SelectionDAG &DAG) const {
3374 MachineFunction &MF = DAG.getMachineFunction();
3375 MachineFrameInfo *MFI = MF.getFrameInfo();
3376 MFI->setReturnAddressIsTaken(true);
3378 EVT VT = Op.getValueType();
3380 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3382 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3383 SDValue Offset = DAG.getConstant(8, getPointerTy());
3384 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
3385 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
3386 MachinePointerInfo(), false, false, false, 0);
3389 // Return LR, which contains the return address. Mark it an implicit live-in.
3390 unsigned Reg = MF.addLiveIn(ARM64::LR, &ARM64::GPR64RegClass);
3391 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
3394 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3395 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3396 SDValue ARM64TargetLowering::LowerShiftRightParts(SDValue Op,
3397 SelectionDAG &DAG) const {
3398 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3399 EVT VT = Op.getValueType();
3400 unsigned VTBits = VT.getSizeInBits();
3402 SDValue ShOpLo = Op.getOperand(0);
3403 SDValue ShOpHi = Op.getOperand(1);
3404 SDValue ShAmt = Op.getOperand(2);
3406 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3408 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3410 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3411 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3412 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3413 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3414 DAG.getConstant(VTBits, MVT::i64));
3415 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3417 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
3418 ISD::SETGE, dl, DAG);
3419 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3421 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3422 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3424 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3426 // ARM64 shifts larger than the register width are wrapped rather than
3427 // clamped, so we can't just emit "hi >> x".
3428 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3429 SDValue TrueValHi = Opc == ISD::SRA
3430 ? DAG.getNode(Opc, dl, VT, ShOpHi,
3431 DAG.getConstant(VTBits - 1, MVT::i64))
3432 : DAG.getConstant(0, VT);
3434 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
3436 SDValue Ops[2] = { Lo, Hi };
3437 return DAG.getMergeValues(Ops, 2, dl);
3440 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3441 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3442 SDValue ARM64TargetLowering::LowerShiftLeftParts(SDValue Op,
3443 SelectionDAG &DAG) const {
3444 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3445 EVT VT = Op.getValueType();
3446 unsigned VTBits = VT.getSizeInBits();
3448 SDValue ShOpLo = Op.getOperand(0);
3449 SDValue ShOpHi = Op.getOperand(1);
3450 SDValue ShAmt = Op.getOperand(2);
3453 assert(Op.getOpcode() == ISD::SHL_PARTS);
3454 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3455 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3456 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3457 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3458 DAG.getConstant(VTBits, MVT::i64));
3459 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3460 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3462 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3464 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
3465 ISD::SETGE, dl, DAG);
3466 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3467 SDValue Hi = DAG.getNode(ARM64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
3469 // ARM64 shifts of larger than register sizes are wrapped rather than clamped,
3470 // so we can't just emit "lo << a" if a is too big.
3471 SDValue TrueValLo = DAG.getConstant(0, VT);
3472 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3474 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3476 SDValue Ops[2] = { Lo, Hi };
3477 return DAG.getMergeValues(Ops, 2, dl);
3481 ARM64TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3482 // The ARM64 target doesn't support folding offsets into global addresses.
3486 bool ARM64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3487 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
3488 // FIXME: We should be able to handle f128 as well with a clever lowering.
3489 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
3493 return ARM64_AM::getFP64Imm(Imm) != -1;
3494 else if (VT == MVT::f32)
3495 return ARM64_AM::getFP32Imm(Imm) != -1;
3499 //===----------------------------------------------------------------------===//
3500 // ARM64 Optimization Hooks
3501 //===----------------------------------------------------------------------===//
3503 //===----------------------------------------------------------------------===//
3504 // ARM64 Inline Assembly Support
3505 //===----------------------------------------------------------------------===//
3507 // Table of Constraints
3508 // TODO: This is the current set of constraints supported by ARM for the
3509 // compiler, not all of them may make sense, e.g. S may be difficult to support.
3511 // r - A general register
3512 // w - An FP/SIMD register of some size in the range v0-v31
3513 // x - An FP/SIMD register of some size in the range v0-v15
3514 // I - Constant that can be used with an ADD instruction
3515 // J - Constant that can be used with a SUB instruction
3516 // K - Constant that can be used with a 32-bit logical instruction
3517 // L - Constant that can be used with a 64-bit logical instruction
3518 // M - Constant that can be used as a 32-bit MOV immediate
3519 // N - Constant that can be used as a 64-bit MOV immediate
3520 // Q - A memory reference with base register and no offset
3521 // S - A symbolic address
3522 // Y - Floating point constant zero
3523 // Z - Integer constant zero
3525 // Note that general register operands will be output using their 64-bit x
3526 // register name, whatever the size of the variable, unless the asm operand
3527 // is prefixed by the %w modifier. Floating-point and SIMD register operands
3528 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
3531 /// getConstraintType - Given a constraint letter, return the type of
3532 /// constraint it is for this target.
3533 ARM64TargetLowering::ConstraintType
3534 ARM64TargetLowering::getConstraintType(const std::string &Constraint) const {
3535 if (Constraint.size() == 1) {
3536 switch (Constraint[0]) {
3543 return C_RegisterClass;
3544 // An address with a single base register. Due to the way we
3545 // currently handle addresses it is the same as 'r'.
3550 return TargetLowering::getConstraintType(Constraint);
3553 /// Examine constraint type and operand type and determine a weight value.
3554 /// This object must already have been set up with the operand type
3555 /// and the current alternative constraint selected.
3556 TargetLowering::ConstraintWeight
3557 ARM64TargetLowering::getSingleConstraintMatchWeight(
3558 AsmOperandInfo &info, const char *constraint) const {
3559 ConstraintWeight weight = CW_Invalid;
3560 Value *CallOperandVal = info.CallOperandVal;
3561 // If we don't have a value, we can't do a match,
3562 // but allow it at the lowest weight.
3563 if (!CallOperandVal)
3565 Type *type = CallOperandVal->getType();
3566 // Look at the constraint type.
3567 switch (*constraint) {
3569 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3573 if (type->isFloatingPointTy() || type->isVectorTy())
3574 weight = CW_Register;
3577 weight = CW_Constant;
3583 std::pair<unsigned, const TargetRegisterClass *>
3584 ARM64TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3586 if (Constraint.size() == 1) {
3587 switch (Constraint[0]) {
3589 if (VT.getSizeInBits() == 64)
3590 return std::make_pair(0U, &ARM64::GPR64commonRegClass);
3591 return std::make_pair(0U, &ARM64::GPR32commonRegClass);
3594 return std::make_pair(0U, &ARM64::FPR32RegClass);
3595 if (VT.getSizeInBits() == 64)
3596 return std::make_pair(0U, &ARM64::FPR64RegClass);
3597 if (VT.getSizeInBits() == 128)
3598 return std::make_pair(0U, &ARM64::FPR128RegClass);
3600 // The instructions that this constraint is designed for can
3601 // only take 128-bit registers so just use that regclass.
3603 if (VT.getSizeInBits() == 128)
3604 return std::make_pair(0U, &ARM64::FPR128_loRegClass);
3608 if (StringRef("{cc}").equals_lower(Constraint))
3609 return std::make_pair(unsigned(ARM64::CPSR), &ARM64::CCRRegClass);
3611 // Use the default implementation in TargetLowering to convert the register
3612 // constraint into a member of a register class.
3613 std::pair<unsigned, const TargetRegisterClass *> Res;
3614 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3616 // Not found as a standard register?
3618 unsigned Size = Constraint.size();
3619 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
3620 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
3621 const std::string Reg =
3622 std::string(&Constraint[2], &Constraint[Size - 1]);
3623 int RegNo = atoi(Reg.c_str());
3624 if (RegNo >= 0 && RegNo <= 31) {
3625 // v0 - v31 are aliases of q0 - q31.
3626 // By default we'll emit v0-v31 for this unless there's a modifier where
3627 // we'll emit the correct register as well.
3628 Res.first = ARM64::FPR128RegClass.getRegister(RegNo);
3629 Res.second = &ARM64::FPR128RegClass;
3637 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3638 /// vector. If it is invalid, don't add anything to Ops.
3639 void ARM64TargetLowering::LowerAsmOperandForConstraint(
3640 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
3641 SelectionDAG &DAG) const {
3644 // Currently only support length 1 constraints.
3645 if (Constraint.length() != 1)
3648 char ConstraintLetter = Constraint[0];
3649 switch (ConstraintLetter) {
3653 // This set of constraints deal with valid constants for various instructions.
3654 // Validate and return a target constant for them if we can.
3656 // 'z' maps to xzr or wzr so it needs an input of 0.
3657 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3658 if (!C || C->getZExtValue() != 0)
3661 if (Op.getValueType() == MVT::i64)
3662 Result = DAG.getRegister(ARM64::XZR, MVT::i64);
3664 Result = DAG.getRegister(ARM64::WZR, MVT::i32);
3674 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3678 // Grab the value and do some validation.
3679 uint64_t CVal = C->getZExtValue();
3680 switch (ConstraintLetter) {
3681 // The I constraint applies only to simple ADD or SUB immediate operands:
3682 // i.e. 0 to 4095 with optional shift by 12
3683 // The J constraint applies only to ADD or SUB immediates that would be
3684 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
3685 // instruction [or vice versa], in other words -1 to -4095 with optional
3686 // left shift by 12.
3688 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
3692 uint64_t NVal = -C->getSExtValue();
3693 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal))
3697 // The K and L constraints apply *only* to logical immediates, including
3698 // what used to be the MOVI alias for ORR (though the MOVI alias has now
3699 // been removed and MOV should be used). So these constraints have to
3700 // distinguish between bit patterns that are valid 32-bit or 64-bit
3701 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
3702 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
3705 if (ARM64_AM::isLogicalImmediate(CVal, 32))
3709 if (ARM64_AM::isLogicalImmediate(CVal, 64))
3712 // The M and N constraints are a superset of K and L respectively, for use
3713 // with the MOV (immediate) alias. As well as the logical immediates they
3714 // also match 32 or 64-bit immediates that can be loaded either using a
3715 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
3716 // (M) or 64-bit 0x1234000000000000 (N) etc.
3717 // As a note some of this code is liberally stolen from the asm parser.
3719 if (!isUInt<32>(CVal))
3721 if (ARM64_AM::isLogicalImmediate(CVal, 32))
3723 if ((CVal & 0xFFFF) == CVal)
3725 if ((CVal & 0xFFFF0000ULL) == CVal)
3727 uint64_t NCVal = ~(uint32_t)CVal;
3728 if ((NCVal & 0xFFFFULL) == NCVal)
3730 if ((NCVal & 0xFFFF0000ULL) == NCVal)
3735 if (ARM64_AM::isLogicalImmediate(CVal, 64))
3737 if ((CVal & 0xFFFFULL) == CVal)
3739 if ((CVal & 0xFFFF0000ULL) == CVal)
3741 if ((CVal & 0xFFFF00000000ULL) == CVal)
3743 if ((CVal & 0xFFFF000000000000ULL) == CVal)
3745 uint64_t NCVal = ~CVal;
3746 if ((NCVal & 0xFFFFULL) == NCVal)
3748 if ((NCVal & 0xFFFF0000ULL) == NCVal)
3750 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
3752 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
3760 // All assembler immediates are 64-bit integers.
3761 Result = DAG.getTargetConstant(CVal, MVT::i64);
3765 if (Result.getNode()) {
3766 Ops.push_back(Result);
3770 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
3773 //===----------------------------------------------------------------------===//
3774 // ARM64 Advanced SIMD Support
3775 //===----------------------------------------------------------------------===//
3777 /// WidenVector - Given a value in the V64 register class, produce the
3778 /// equivalent value in the V128 register class.
3779 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
3780 EVT VT = V64Reg.getValueType();
3781 unsigned NarrowSize = VT.getVectorNumElements();
3782 MVT EltTy = VT.getVectorElementType().getSimpleVT();
3783 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
3786 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
3787 V64Reg, DAG.getConstant(0, MVT::i32));
3790 /// getExtFactor - Determine the adjustment factor for the position when
3791 /// generating an "extract from vector registers" instruction.
3792 static unsigned getExtFactor(SDValue &V) {
3793 EVT EltType = V.getValueType().getVectorElementType();
3794 return EltType.getSizeInBits() / 8;
3797 /// NarrowVector - Given a value in the V128 register class, produce the
3798 /// equivalent value in the V64 register class.
3799 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
3800 EVT VT = V128Reg.getValueType();
3801 unsigned WideSize = VT.getVectorNumElements();
3802 MVT EltTy = VT.getVectorElementType().getSimpleVT();
3803 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
3806 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, NarrowTy, V128Reg);
3809 // Gather data to see if the operation can be modelled as a
3810 // shuffle in combination with VEXTs.
3811 SDValue ARM64TargetLowering::ReconstructShuffle(SDValue Op,
3812 SelectionDAG &DAG) const {
3814 EVT VT = Op.getValueType();
3815 unsigned NumElts = VT.getVectorNumElements();
3817 SmallVector<SDValue, 2> SourceVecs;
3818 SmallVector<unsigned, 2> MinElts;
3819 SmallVector<unsigned, 2> MaxElts;
3821 for (unsigned i = 0; i < NumElts; ++i) {
3822 SDValue V = Op.getOperand(i);
3823 if (V.getOpcode() == ISD::UNDEF)
3825 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
3826 // A shuffle can only come from building a vector from various
3827 // elements of other vectors.
3831 // Record this extraction against the appropriate vector if possible...
3832 SDValue SourceVec = V.getOperand(0);
3833 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
3834 bool FoundSource = false;
3835 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
3836 if (SourceVecs[j] == SourceVec) {
3837 if (MinElts[j] > EltNo)
3839 if (MaxElts[j] < EltNo)
3846 // Or record a new source if not...
3848 SourceVecs.push_back(SourceVec);
3849 MinElts.push_back(EltNo);
3850 MaxElts.push_back(EltNo);
3854 // Currently only do something sane when at most two source vectors
3856 if (SourceVecs.size() > 2)
3859 SDValue ShuffleSrcs[2] = { DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
3860 int VEXTOffsets[2] = { 0, 0 };
3862 // This loop extracts the usage patterns of the source vectors
3863 // and prepares appropriate SDValues for a shuffle if possible.
3864 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
3865 if (SourceVecs[i].getValueType() == VT) {
3866 // No VEXT necessary
3867 ShuffleSrcs[i] = SourceVecs[i];
3870 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
3871 // We can pad out the smaller vector for free, so if it's part of a
3873 ShuffleSrcs[i] = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, SourceVecs[i],
3874 DAG.getUNDEF(SourceVecs[i].getValueType()));
3878 // Don't attempt to extract subvectors from BUILD_VECTOR sources
3879 // that expand or trunc the original value.
3880 // TODO: We can try to bitcast and ANY_EXTEND the result but
3881 // we need to consider the cost of vector ANY_EXTEND, and the
3882 // legality of all the types.
3883 if (SourceVecs[i].getValueType().getVectorElementType() !=
3884 VT.getVectorElementType())
3887 // Since only 64-bit and 128-bit vectors are legal on ARM and
3888 // we've eliminated the other cases...
3889 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2 * NumElts &&
3890 "unexpected vector sizes in ReconstructShuffle");
3892 if (MaxElts[i] - MinElts[i] >= NumElts) {
3893 // Span too large for a VEXT to cope
3897 if (MinElts[i] >= NumElts) {
3898 // The extraction can just take the second half
3899 VEXTOffsets[i] = NumElts;
3901 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
3902 DAG.getIntPtrConstant(NumElts));
3903 } else if (MaxElts[i] < NumElts) {
3904 // The extraction can just take the first half
3906 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3907 SourceVecs[i], DAG.getIntPtrConstant(0));
3909 // An actual VEXT is needed
3910 VEXTOffsets[i] = MinElts[i];
3911 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3912 SourceVecs[i], DAG.getIntPtrConstant(0));
3914 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
3915 DAG.getIntPtrConstant(NumElts));
3916 unsigned Imm = VEXTOffsets[i] * getExtFactor(VEXTSrc1);
3917 ShuffleSrcs[i] = DAG.getNode(ARM64ISD::EXT, dl, VT, VEXTSrc1, VEXTSrc2,
3918 DAG.getConstant(Imm, MVT::i32));
3922 SmallVector<int, 8> Mask;
3924 for (unsigned i = 0; i < NumElts; ++i) {
3925 SDValue Entry = Op.getOperand(i);
3926 if (Entry.getOpcode() == ISD::UNDEF) {
3931 SDValue ExtractVec = Entry.getOperand(0);
3933 cast<ConstantSDNode>(Op.getOperand(i).getOperand(1))->getSExtValue();
3934 if (ExtractVec == SourceVecs[0]) {
3935 Mask.push_back(ExtractElt - VEXTOffsets[0]);
3937 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
3941 // Final check before we try to produce nonsense...
3942 if (isShuffleMaskLegal(Mask, VT))
3943 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
3949 // check if an EXT instruction can handle the shuffle mask when the
3950 // vector sources of the shuffle are the same.
3951 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
3952 unsigned NumElts = VT.getVectorNumElements();
3954 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3960 // If this is a VEXT shuffle, the immediate value is the index of the first
3961 // element. The other shuffle indices must be the successive elements after
3963 unsigned ExpectedElt = Imm;
3964 for (unsigned i = 1; i < NumElts; ++i) {
3965 // Increment the expected index. If it wraps around, just follow it
3966 // back to index zero and keep going.
3968 if (ExpectedElt == NumElts)
3972 continue; // ignore UNDEF indices
3973 if (ExpectedElt != static_cast<unsigned>(M[i]))
3980 // check if an EXT instruction can handle the shuffle mask when the
3981 // vector sources of the shuffle are different.
3982 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
3984 unsigned NumElts = VT.getVectorNumElements();
3987 // Look for the first non-undef choice and count backwards from
3988 // that. E.g. <-1, -1, 3, ...> means that an EXT must start at 3 - 2 = 1. This
3989 // guarantees that at least one index is correct.
3990 const int *FirstRealElt =
3991 std::find_if(M.begin(), M.end(), [](int Elt) { return Elt >= 0; });
3992 assert(FirstRealElt != M.end() && "Completely UNDEF shuffle? Why bother?");
3993 Imm = *FirstRealElt - (FirstRealElt - M.begin());
3995 // If this is a VEXT shuffle, the immediate value is the index of the first
3996 // element. The other shuffle indices must be the successive elements after
3998 unsigned ExpectedElt = Imm;
3999 for (unsigned i = 1; i < NumElts; ++i) {
4000 // Increment the expected index. If it wraps around, it may still be
4001 // a VEXT but the source vectors must be swapped.
4003 if (ExpectedElt == NumElts * 2) {
4009 continue; // ignore UNDEF indices
4010 if (ExpectedElt != static_cast<unsigned>(M[i]))
4014 // Adjust the index value if the source operands will be swapped.
4021 /// isREVMask - Check if a vector shuffle corresponds to a REV
4022 /// instruction with the specified blocksize. (The order of the elements
4023 /// within each block of the vector is reversed.)
4024 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4025 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
4026 "Only possible block sizes for REV are: 16, 32, 64");
4028 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4032 unsigned NumElts = VT.getVectorNumElements();
4033 unsigned BlockElts = M[0] + 1;
4034 // If the first shuffle index is UNDEF, be optimistic.
4036 BlockElts = BlockSize / EltSz;
4038 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4041 for (unsigned i = 0; i < NumElts; ++i) {
4043 continue; // ignore UNDEF indices
4044 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
4051 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4052 unsigned NumElts = VT.getVectorNumElements();
4053 WhichResult = (M[0] == 0 ? 0 : 1);
4054 unsigned Idx = WhichResult * NumElts / 2;
4055 for (unsigned i = 0; i != NumElts; i += 2) {
4056 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4057 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
4065 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4066 unsigned NumElts = VT.getVectorNumElements();
4067 WhichResult = (M[0] == 0 ? 0 : 1);
4068 for (unsigned i = 0; i != NumElts; ++i) {
4070 continue; // ignore UNDEF indices
4071 if ((unsigned)M[i] != 2 * i + WhichResult)
4078 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4079 unsigned NumElts = VT.getVectorNumElements();
4080 WhichResult = (M[0] == 0 ? 0 : 1);
4081 for (unsigned i = 0; i < NumElts; i += 2) {
4082 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4083 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
4089 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
4090 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4091 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4092 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4093 unsigned NumElts = VT.getVectorNumElements();
4094 WhichResult = (M[0] == 0 ? 0 : 1);
4095 unsigned Idx = WhichResult * NumElts / 2;
4096 for (unsigned i = 0; i != NumElts; i += 2) {
4097 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4098 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
4106 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
4107 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4108 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4109 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4110 unsigned Half = VT.getVectorNumElements() / 2;
4111 WhichResult = (M[0] == 0 ? 0 : 1);
4112 for (unsigned j = 0; j != 2; ++j) {
4113 unsigned Idx = WhichResult;
4114 for (unsigned i = 0; i != Half; ++i) {
4115 int MIdx = M[i + j * Half];
4116 if (MIdx >= 0 && (unsigned)MIdx != Idx)
4125 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
4126 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4127 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4128 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4129 unsigned NumElts = VT.getVectorNumElements();
4130 WhichResult = (M[0] == 0 ? 0 : 1);
4131 for (unsigned i = 0; i < NumElts; i += 2) {
4132 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4133 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
4139 static bool isINSMask(ArrayRef<int> M, int NumInputElements,
4140 bool &DstIsLeft, int &Anomaly) {
4141 if (M.size() != static_cast<size_t>(NumInputElements))
4144 int NumLHSMatch = 0, NumRHSMatch = 0;
4145 int LastLHSMismatch = -1, LastRHSMismatch = -1;
4147 for (int i = 0; i < NumInputElements; ++i) {
4157 LastLHSMismatch = i;
4159 if (M[i] == i + NumInputElements)
4162 LastRHSMismatch = i;
4165 if (NumLHSMatch == NumInputElements - 1) {
4167 Anomaly = LastLHSMismatch;
4169 } else if (NumRHSMatch == NumInputElements - 1) {
4171 Anomaly = LastRHSMismatch;
4178 static bool isConcatMask(ArrayRef<int> Mask, EVT VT, bool SplitLHS) {
4179 if (VT.getSizeInBits() != 128)
4182 unsigned NumElts = VT.getVectorNumElements();
4184 for (int I = 0, E = NumElts / 2; I != E; I++) {
4189 int Offset = NumElts / 2;
4190 for (int I = NumElts / 2, E = NumElts; I != E; I++) {
4191 if (Mask[I] != I + SplitLHS * Offset)
4198 static SDValue tryFormConcatFromShuffle(SDValue Op, SelectionDAG &DAG) {
4200 EVT VT = Op.getValueType();
4201 SDValue V0 = Op.getOperand(0);
4202 SDValue V1 = Op.getOperand(1);
4203 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();
4205 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
4206 VT.getVectorElementType() != V1.getValueType().getVectorElementType())
4209 bool SplitV0 = V0.getValueType().getSizeInBits() == 128;
4211 if (!isConcatMask(Mask, VT, SplitV0))
4214 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
4215 VT.getVectorNumElements() / 2);
4217 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
4218 DAG.getConstant(0, MVT::i64));
4220 if (V1.getValueType().getSizeInBits() == 128) {
4221 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
4222 DAG.getConstant(0, MVT::i64));
4224 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
4227 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4228 /// the specified operations to build the shuffle.
4229 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4230 SDValue RHS, SelectionDAG &DAG,
4232 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4233 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
4234 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
4237 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4246 OP_VUZPL, // VUZP, left result
4247 OP_VUZPR, // VUZP, right result
4248 OP_VZIPL, // VZIP, left result
4249 OP_VZIPR, // VZIP, right result
4250 OP_VTRNL, // VTRN, left result
4251 OP_VTRNR // VTRN, right result
4254 if (OpNum == OP_COPY) {
4255 if (LHSID == (1 * 9 + 2) * 9 + 3)
4257 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
4261 SDValue OpLHS, OpRHS;
4262 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4263 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4264 EVT VT = OpLHS.getValueType();
4268 llvm_unreachable("Unknown shuffle opcode!");
4270 // VREV divides the vector in half and swaps within the half.
4271 if (VT.getVectorElementType() == MVT::i32 ||
4272 VT.getVectorElementType() == MVT::f32)
4273 return DAG.getNode(ARM64ISD::REV64, dl, VT, OpLHS);
4274 // vrev <4 x i16> -> REV32
4275 if (VT.getVectorElementType() == MVT::i16)
4276 return DAG.getNode(ARM64ISD::REV32, dl, VT, OpLHS);
4277 // vrev <4 x i8> -> REV16
4278 assert(VT.getVectorElementType() == MVT::i8);
4279 return DAG.getNode(ARM64ISD::REV16, dl, VT, OpLHS);
4284 EVT EltTy = VT.getVectorElementType();
4286 if (EltTy == MVT::i8)
4287 Opcode = ARM64ISD::DUPLANE8;
4288 else if (EltTy == MVT::i16)
4289 Opcode = ARM64ISD::DUPLANE16;
4290 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
4291 Opcode = ARM64ISD::DUPLANE32;
4292 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
4293 Opcode = ARM64ISD::DUPLANE64;
4295 llvm_unreachable("Invalid vector element type?");
4297 if (VT.getSizeInBits() == 64)
4298 OpLHS = WidenVector(OpLHS, DAG);
4299 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, MVT::i64);
4300 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
4305 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
4306 return DAG.getNode(ARM64ISD::EXT, dl, VT, OpLHS, OpRHS,
4307 DAG.getConstant(Imm, MVT::i32));
4310 return DAG.getNode(ARM64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4312 return DAG.getNode(ARM64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4314 return DAG.getNode(ARM64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4316 return DAG.getNode(ARM64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4318 return DAG.getNode(ARM64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4320 return DAG.getNode(ARM64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4324 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
4325 SelectionDAG &DAG) {
4326 // Check to see if we can use the TBL instruction.
4327 SDValue V1 = Op.getOperand(0);
4328 SDValue V2 = Op.getOperand(1);
4331 EVT EltVT = Op.getValueType().getVectorElementType();
4332 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
4334 SmallVector<SDValue, 8> TBLMask;
4335 for (int Val : ShuffleMask) {
4336 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
4337 unsigned Offset = Byte + Val * BytesPerElt;
4338 TBLMask.push_back(DAG.getConstant(Offset, MVT::i32));
4342 MVT IndexVT = MVT::v8i8;
4343 unsigned IndexLen = 8;
4344 if (Op.getValueType().getSizeInBits() == 128) {
4345 IndexVT = MVT::v16i8;
4349 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
4350 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
4353 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
4355 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
4356 Shuffle = DAG.getNode(
4357 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4358 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4359 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4360 ArrayRef<SDValue>(TBLMask.data(), IndexLen)));
4362 if (IndexLen == 8) {
4363 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
4364 Shuffle = DAG.getNode(
4365 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4366 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4367 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4368 ArrayRef<SDValue>(TBLMask.data(), IndexLen)));
4370 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
4371 // cannot currently represent the register constraints on the input
4373 // Shuffle = DAG.getNode(ARM64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
4374 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4375 // &TBLMask[0], IndexLen));
4376 Shuffle = DAG.getNode(
4377 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4378 DAG.getConstant(Intrinsic::arm64_neon_tbl2, MVT::i32), V1Cst, V2Cst,
4379 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4380 ArrayRef<SDValue>(TBLMask.data(), IndexLen)));
4383 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
4386 static unsigned getDUPLANEOp(EVT EltType) {
4387 if (EltType == MVT::i8)
4388 return ARM64ISD::DUPLANE8;
4389 if (EltType == MVT::i16)
4390 return ARM64ISD::DUPLANE16;
4391 if (EltType == MVT::i32 || EltType == MVT::f32)
4392 return ARM64ISD::DUPLANE32;
4393 if (EltType == MVT::i64 || EltType == MVT::f64)
4394 return ARM64ISD::DUPLANE64;
4396 llvm_unreachable("Invalid vector element type?");
4399 SDValue ARM64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4400 SelectionDAG &DAG) const {
4402 EVT VT = Op.getValueType();
4404 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4406 // Convert shuffles that are directly supported on NEON to target-specific
4407 // DAG nodes, instead of keeping them as shuffles and matching them again
4408 // during code selection. This is more efficient and avoids the possibility
4409 // of inconsistencies between legalization and selection.
4410 ArrayRef<int> ShuffleMask = SVN->getMask();
4412 SDValue V1 = Op.getOperand(0);
4413 SDValue V2 = Op.getOperand(1);
4415 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
4416 V1.getValueType().getSimpleVT())) {
4417 int Lane = SVN->getSplatIndex();
4418 // If this is undef splat, generate it via "just" vdup, if possible.
4422 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
4423 return DAG.getNode(ARM64ISD::DUP, dl, V1.getValueType(),
4425 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
4426 // constant. If so, we can just reference the lane's definition directly.
4427 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
4428 !isa<ConstantSDNode>(V1.getOperand(Lane)))
4429 return DAG.getNode(ARM64ISD::DUP, dl, VT, V1.getOperand(Lane));
4431 // Otherwise, duplicate from the lane of the input vector.
4432 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
4434 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
4435 // to make a vector of the same size as this SHUFFLE. We can ignore the
4436 // extract entirely, and canonicalise the concat using WidenVector.
4437 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
4438 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
4439 V1 = V1.getOperand(0);
4440 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
4441 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
4442 Lane -= Idx * VT.getVectorNumElements() / 2;
4443 V1 = WidenVector(V1.getOperand(Idx), DAG);
4444 } else if (VT.getSizeInBits() == 64)
4445 V1 = WidenVector(V1, DAG);
4447 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, MVT::i64));
4450 if (isREVMask(ShuffleMask, VT, 64))
4451 return DAG.getNode(ARM64ISD::REV64, dl, V1.getValueType(), V1, V2);
4452 if (isREVMask(ShuffleMask, VT, 32))
4453 return DAG.getNode(ARM64ISD::REV32, dl, V1.getValueType(), V1, V2);
4454 if (isREVMask(ShuffleMask, VT, 16))
4455 return DAG.getNode(ARM64ISD::REV16, dl, V1.getValueType(), V1, V2);
4457 bool ReverseEXT = false;
4459 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
4462 Imm *= getExtFactor(V1);
4463 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V2,
4464 DAG.getConstant(Imm, MVT::i32));
4465 } else if (V2->getOpcode() == ISD::UNDEF &&
4466 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
4467 Imm *= getExtFactor(V1);
4468 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V1,
4469 DAG.getConstant(Imm, MVT::i32));
4472 unsigned WhichResult;
4473 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
4474 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4475 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4477 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
4478 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4479 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4481 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
4482 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4483 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4486 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4487 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4488 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4490 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4491 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4492 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4494 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4495 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4496 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4499 SDValue Concat = tryFormConcatFromShuffle(Op, DAG);
4500 if (Concat.getNode())
4505 int NumInputElements = V1.getValueType().getVectorNumElements();
4506 if (isINSMask(ShuffleMask, NumInputElements, DstIsLeft, Anomaly)) {
4507 SDValue DstVec = DstIsLeft ? V1 : V2;
4508 SDValue DstLaneV = DAG.getConstant(Anomaly, MVT::i64);
4510 SDValue SrcVec = V1;
4511 int SrcLane = ShuffleMask[Anomaly];
4512 if (SrcLane >= NumInputElements) {
4514 SrcLane -= VT.getVectorNumElements();
4516 SDValue SrcLaneV = DAG.getConstant(SrcLane, MVT::i64);
4518 EVT ScalarVT = VT.getVectorElementType();
4519 if (ScalarVT.getSizeInBits() < 32)
4520 ScalarVT = MVT::i32;
4523 ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
4524 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ScalarVT, SrcVec, SrcLaneV),
4528 // If the shuffle is not directly supported and it has 4 elements, use
4529 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4530 unsigned NumElts = VT.getVectorNumElements();
4532 unsigned PFIndexes[4];
4533 for (unsigned i = 0; i != 4; ++i) {
4534 if (ShuffleMask[i] < 0)
4537 PFIndexes[i] = ShuffleMask[i];
4540 // Compute the index in the perfect shuffle table.
4541 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
4542 PFIndexes[2] * 9 + PFIndexes[3];
4543 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4544 unsigned Cost = (PFEntry >> 30);
4547 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4550 return GenerateTBL(Op, ShuffleMask, DAG);
4553 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
4555 EVT VT = BVN->getValueType(0);
4556 APInt SplatBits, SplatUndef;
4557 unsigned SplatBitSize;
4559 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4560 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
4562 for (unsigned i = 0; i < NumSplats; ++i) {
4563 CnstBits <<= SplatBitSize;
4564 UndefBits <<= SplatBitSize;
4565 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
4566 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
4575 SDValue ARM64TargetLowering::LowerVectorAND(SDValue Op,
4576 SelectionDAG &DAG) const {
4577 BuildVectorSDNode *BVN =
4578 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
4579 SDValue LHS = Op.getOperand(0);
4581 EVT VT = Op.getValueType();
4586 APInt CnstBits(VT.getSizeInBits(), 0);
4587 APInt UndefBits(VT.getSizeInBits(), 0);
4588 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
4589 // We only have BIC vector immediate instruction, which is and-not.
4590 CnstBits = ~CnstBits;
4592 // We make use of a little bit of goto ickiness in order to avoid having to
4593 // duplicate the immediate matching logic for the undef toggled case.
4594 bool SecondTry = false;
4597 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
4598 CnstBits = CnstBits.zextOrTrunc(64);
4599 uint64_t CnstVal = CnstBits.getZExtValue();
4601 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
4602 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
4603 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4604 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4605 DAG.getConstant(CnstVal, MVT::i32),
4606 DAG.getConstant(0, MVT::i32));
4607 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4610 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
4611 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
4612 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4613 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4614 DAG.getConstant(CnstVal, MVT::i32),
4615 DAG.getConstant(8, MVT::i32));
4616 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4619 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
4620 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
4621 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4622 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4623 DAG.getConstant(CnstVal, MVT::i32),
4624 DAG.getConstant(16, MVT::i32));
4625 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4628 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
4629 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
4630 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4631 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4632 DAG.getConstant(CnstVal, MVT::i32),
4633 DAG.getConstant(24, MVT::i32));
4634 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4637 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
4638 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
4639 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4640 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4641 DAG.getConstant(CnstVal, MVT::i32),
4642 DAG.getConstant(0, MVT::i32));
4643 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4646 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
4647 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
4648 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4649 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4650 DAG.getConstant(CnstVal, MVT::i32),
4651 DAG.getConstant(8, MVT::i32));
4652 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4659 CnstBits = ~UndefBits;
4663 // We can always fall back to a non-immediate AND.
4668 // Specialized code to quickly find if PotentialBVec is a BuildVector that
4669 // consists of only the same constant int value, returned in reference arg
4671 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
4672 uint64_t &ConstVal) {
4673 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
4676 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
4679 EVT VT = Bvec->getValueType(0);
4680 unsigned NumElts = VT.getVectorNumElements();
4681 for (unsigned i = 1; i < NumElts; ++i)
4682 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
4684 ConstVal = FirstElt->getZExtValue();
4688 static unsigned getIntrinsicID(const SDNode *N) {
4689 unsigned Opcode = N->getOpcode();
4692 return Intrinsic::not_intrinsic;
4693 case ISD::INTRINSIC_WO_CHAIN: {
4694 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4695 if (IID < Intrinsic::num_intrinsics)
4697 return Intrinsic::not_intrinsic;
4702 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
4703 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
4704 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
4705 // Also, logical shift right -> sri, with the same structure.
4706 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
4707 EVT VT = N->getValueType(0);
4714 // Is the first op an AND?
4715 const SDValue And = N->getOperand(0);
4716 if (And.getOpcode() != ISD::AND)
4719 // Is the second op an shl or lshr?
4720 SDValue Shift = N->getOperand(1);
4721 // This will have been turned into: ARM64ISD::VSHL vector, #shift
4722 // or ARM64ISD::VLSHR vector, #shift
4723 unsigned ShiftOpc = Shift.getOpcode();
4724 if ((ShiftOpc != ARM64ISD::VSHL && ShiftOpc != ARM64ISD::VLSHR))
4726 bool IsShiftRight = ShiftOpc == ARM64ISD::VLSHR;
4728 // Is the shift amount constant?
4729 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
4733 // Is the and mask vector all constant?
4735 if (!isAllConstantBuildVector(And.getOperand(1), C1))
4738 // Is C1 == ~C2, taking into account how much one can shift elements of a
4740 uint64_t C2 = C2node->getZExtValue();
4741 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
4742 if (C2 > ElemSizeInBits)
4744 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
4745 if ((C1 & ElemMask) != (~C2 & ElemMask))
4748 SDValue X = And.getOperand(0);
4749 SDValue Y = Shift.getOperand(0);
4752 IsShiftRight ? Intrinsic::arm64_neon_vsri : Intrinsic::arm64_neon_vsli;
4754 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
4755 DAG.getConstant(Intrin, MVT::i32), X, Y, Shift.getOperand(1));
4757 DEBUG(dbgs() << "arm64-lower: transformed: \n");
4758 DEBUG(N->dump(&DAG));
4759 DEBUG(dbgs() << "into: \n");
4760 DEBUG(ResultSLI->dump(&DAG));
4766 SDValue ARM64TargetLowering::LowerVectorOR(SDValue Op,
4767 SelectionDAG &DAG) const {
4768 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
4769 if (EnableARM64SlrGeneration) {
4770 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
4775 BuildVectorSDNode *BVN =
4776 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
4777 SDValue LHS = Op.getOperand(1);
4779 EVT VT = Op.getValueType();
4781 // OR commutes, so try swapping the operands.
4783 LHS = Op.getOperand(0);
4784 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
4789 APInt CnstBits(VT.getSizeInBits(), 0);
4790 APInt UndefBits(VT.getSizeInBits(), 0);
4791 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
4792 // We make use of a little bit of goto ickiness in order to avoid having to
4793 // duplicate the immediate matching logic for the undef toggled case.
4794 bool SecondTry = false;
4797 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
4798 CnstBits = CnstBits.zextOrTrunc(64);
4799 uint64_t CnstVal = CnstBits.getZExtValue();
4801 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
4802 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
4803 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4804 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4805 DAG.getConstant(CnstVal, MVT::i32),
4806 DAG.getConstant(0, MVT::i32));
4807 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4810 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
4811 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
4812 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4813 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4814 DAG.getConstant(CnstVal, MVT::i32),
4815 DAG.getConstant(8, MVT::i32));
4816 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4819 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
4820 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
4821 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4822 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4823 DAG.getConstant(CnstVal, MVT::i32),
4824 DAG.getConstant(16, MVT::i32));
4825 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4828 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
4829 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
4830 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4831 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4832 DAG.getConstant(CnstVal, MVT::i32),
4833 DAG.getConstant(24, MVT::i32));
4834 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4837 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
4838 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
4839 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4840 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4841 DAG.getConstant(CnstVal, MVT::i32),
4842 DAG.getConstant(0, MVT::i32));
4843 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4846 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
4847 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
4848 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4849 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4850 DAG.getConstant(CnstVal, MVT::i32),
4851 DAG.getConstant(8, MVT::i32));
4852 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4859 CnstBits = UndefBits;
4863 // We can always fall back to a non-immediate OR.
4868 SDValue ARM64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
4869 SelectionDAG &DAG) const {
4870 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4872 EVT VT = Op.getValueType();
4874 APInt CnstBits(VT.getSizeInBits(), 0);
4875 APInt UndefBits(VT.getSizeInBits(), 0);
4876 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
4877 // We make use of a little bit of goto ickiness in order to avoid having to
4878 // duplicate the immediate matching logic for the undef toggled case.
4879 bool SecondTry = false;
4882 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
4883 CnstBits = CnstBits.zextOrTrunc(64);
4884 uint64_t CnstVal = CnstBits.getZExtValue();
4886 // Certain magic vector constants (used to express things like NOT
4887 // and NEG) are passed through unmodified. This allows codegen patterns
4888 // for these operations to match. Special-purpose patterns will lower
4889 // these immediates to MOVIs if it proves necessary.
4890 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
4893 // The many faces of MOVI...
4894 if (ARM64_AM::isAdvSIMDModImmType10(CnstVal)) {
4895 CnstVal = ARM64_AM::encodeAdvSIMDModImmType10(CnstVal);
4896 if (VT.getSizeInBits() == 128) {
4897 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::v2i64,
4898 DAG.getConstant(CnstVal, MVT::i32));
4899 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4902 // Support the V64 version via subregister insertion.
4903 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::f64,
4904 DAG.getConstant(CnstVal, MVT::i32));
4905 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4908 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
4909 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
4910 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4911 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
4912 DAG.getConstant(CnstVal, MVT::i32),
4913 DAG.getConstant(0, MVT::i32));
4914 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4917 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
4918 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
4919 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4920 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
4921 DAG.getConstant(CnstVal, MVT::i32),
4922 DAG.getConstant(8, MVT::i32));
4923 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4926 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
4927 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
4928 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4929 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
4930 DAG.getConstant(CnstVal, MVT::i32),
4931 DAG.getConstant(16, MVT::i32));
4932 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4935 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
4936 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
4937 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4938 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
4939 DAG.getConstant(CnstVal, MVT::i32),
4940 DAG.getConstant(24, MVT::i32));
4941 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4944 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
4945 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
4946 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4947 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
4948 DAG.getConstant(CnstVal, MVT::i32),
4949 DAG.getConstant(0, MVT::i32));
4950 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4953 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
4954 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
4955 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4956 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
4957 DAG.getConstant(CnstVal, MVT::i32),
4958 DAG.getConstant(8, MVT::i32));
4959 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4962 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
4963 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
4964 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4965 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
4966 DAG.getConstant(CnstVal, MVT::i32),
4967 DAG.getConstant(264, MVT::i32));
4968 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4971 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
4972 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
4973 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4974 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
4975 DAG.getConstant(CnstVal, MVT::i32),
4976 DAG.getConstant(272, MVT::i32));
4977 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4980 if (ARM64_AM::isAdvSIMDModImmType9(CnstVal)) {
4981 CnstVal = ARM64_AM::encodeAdvSIMDModImmType9(CnstVal);
4982 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
4983 SDValue Mov = DAG.getNode(ARM64ISD::MOVI, dl, MovTy,
4984 DAG.getConstant(CnstVal, MVT::i32));
4985 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4988 // The few faces of FMOV...
4989 if (ARM64_AM::isAdvSIMDModImmType11(CnstVal)) {
4990 CnstVal = ARM64_AM::encodeAdvSIMDModImmType11(CnstVal);
4991 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
4992 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MovTy,
4993 DAG.getConstant(CnstVal, MVT::i32));
4994 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4997 if (ARM64_AM::isAdvSIMDModImmType12(CnstVal) &&
4998 VT.getSizeInBits() == 128) {
4999 CnstVal = ARM64_AM::encodeAdvSIMDModImmType12(CnstVal);
5000 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MVT::v2f64,
5001 DAG.getConstant(CnstVal, MVT::i32));
5002 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5005 // The many faces of MVNI...
5007 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5008 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5009 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5010 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5011 DAG.getConstant(CnstVal, MVT::i32),
5012 DAG.getConstant(0, MVT::i32));
5013 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5016 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5017 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5018 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5019 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5020 DAG.getConstant(CnstVal, MVT::i32),
5021 DAG.getConstant(8, MVT::i32));
5022 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5025 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5026 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5027 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5028 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5029 DAG.getConstant(CnstVal, MVT::i32),
5030 DAG.getConstant(16, MVT::i32));
5031 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5034 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5035 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5036 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5037 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5038 DAG.getConstant(CnstVal, MVT::i32),
5039 DAG.getConstant(24, MVT::i32));
5040 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5043 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5044 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5045 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5046 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5047 DAG.getConstant(CnstVal, MVT::i32),
5048 DAG.getConstant(0, MVT::i32));
5049 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5052 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5053 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5054 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5055 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5056 DAG.getConstant(CnstVal, MVT::i32),
5057 DAG.getConstant(8, MVT::i32));
5058 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5061 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
5062 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
5063 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5064 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
5065 DAG.getConstant(CnstVal, MVT::i32),
5066 DAG.getConstant(264, MVT::i32));
5067 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5070 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
5071 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
5072 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5073 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
5074 DAG.getConstant(CnstVal, MVT::i32),
5075 DAG.getConstant(272, MVT::i32));
5076 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5083 CnstBits = UndefBits;
5088 // Scan through the operands to find some interesting properties we can
5090 // 1) If only one value is used, we can use a DUP, or
5091 // 2) if only the low element is not undef, we can just insert that, or
5092 // 3) if only one constant value is used (w/ some non-constant lanes),
5093 // we can splat the constant value into the whole vector then fill
5094 // in the non-constant lanes.
5095 // 4) FIXME: If different constant values are used, but we can intelligently
5096 // select the values we'll be overwriting for the non-constant
5097 // lanes such that we can directly materialize the vector
5098 // some other way (MOVI, e.g.), we can be sneaky.
5099 unsigned NumElts = VT.getVectorNumElements();
5100 bool isOnlyLowElement = true;
5101 bool usesOnlyOneValue = true;
5102 bool usesOnlyOneConstantValue = true;
5103 bool isConstant = true;
5104 unsigned NumConstantLanes = 0;
5106 SDValue ConstantValue;
5107 for (unsigned i = 0; i < NumElts; ++i) {
5108 SDValue V = Op.getOperand(i);
5109 if (V.getOpcode() == ISD::UNDEF)
5112 isOnlyLowElement = false;
5113 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5116 if (isa<ConstantSDNode>(V) || isa<ConstantFPSDNode>(V)) {
5118 if (!ConstantValue.getNode())
5120 else if (ConstantValue != V)
5121 usesOnlyOneConstantValue = false;
5124 if (!Value.getNode())
5126 else if (V != Value)
5127 usesOnlyOneValue = false;
5130 if (!Value.getNode())
5131 return DAG.getUNDEF(VT);
5133 if (isOnlyLowElement)
5134 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5136 // Use DUP for non-constant splats. For f32 constant splats, reduce to
5137 // i32 and try again.
5138 if (usesOnlyOneValue) {
5140 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5141 Value.getValueType() != VT)
5142 return DAG.getNode(ARM64ISD::DUP, dl, VT, Value);
5144 // This is actually a DUPLANExx operation, which keeps everything vectory.
5146 // DUPLANE works on 128-bit vectors, widen it if necessary.
5147 SDValue Lane = Value.getOperand(1);
5148 Value = Value.getOperand(0);
5149 if (Value.getValueType().getSizeInBits() == 64)
5150 Value = WidenVector(Value, DAG);
5152 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
5153 return DAG.getNode(Opcode, dl, VT, Value, Lane);
5156 if (VT.getVectorElementType().isFloatingPoint()) {
5157 SmallVector<SDValue, 8> Ops;
5159 (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
5160 for (unsigned i = 0; i < NumElts; ++i)
5161 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
5162 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
5163 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5164 Val = LowerBUILD_VECTOR(Val, DAG);
5166 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5170 // If there was only one constant value used and for more than one lane,
5171 // start by splatting that value, then replace the non-constant lanes. This
5172 // is better than the default, which will perform a separate initialization
5174 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
5175 SDValue Val = DAG.getNode(ARM64ISD::DUP, dl, VT, ConstantValue);
5176 // Now insert the non-constant lanes.
5177 for (unsigned i = 0; i < NumElts; ++i) {
5178 SDValue V = Op.getOperand(i);
5179 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5180 if (!isa<ConstantSDNode>(V) && !isa<ConstantFPSDNode>(V)) {
5181 // Note that type legalization likely mucked about with the VT of the
5182 // source operand, so we may have to convert it here before inserting.
5183 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
5189 // If all elements are constants and the case above didn't get hit, fall back
5190 // to the default expansion, which will generate a load from the constant
5195 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5197 SDValue shuffle = ReconstructShuffle(Op, DAG);
5198 if (shuffle != SDValue())
5202 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5203 // know the default expansion would otherwise fall back on something even
5204 // worse. For a vector with one or two non-undef values, that's
5205 // scalar_to_vector for the elements followed by a shuffle (provided the
5206 // shuffle is valid for the target) and materialization element by element
5207 // on the stack followed by a load for everything else.
5208 if (!isConstant && !usesOnlyOneValue) {
5209 SDValue Vec = DAG.getUNDEF(VT);
5210 SDValue Op0 = Op.getOperand(0);
5211 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
5213 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
5214 // a) Avoid a RMW dependency on the full vector register, and
5215 // b) Allow the register coalescer to fold away the copy if the
5216 // value is already in an S or D register.
5217 if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
5218 unsigned SubIdx = ElemSize == 32 ? ARM64::ssub : ARM64::dsub;
5220 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
5221 DAG.getTargetConstant(SubIdx, MVT::i32));
5222 Vec = SDValue(N, 0);
5225 for (; i < NumElts; ++i) {
5226 SDValue V = Op.getOperand(i);
5227 if (V.getOpcode() == ISD::UNDEF)
5229 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5230 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5235 // Just use the default expansion. We failed to find a better alternative.
5239 SDValue ARM64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
5240 SelectionDAG &DAG) const {
5241 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
5243 // Check for non-constant lane.
5244 if (!isa<ConstantSDNode>(Op.getOperand(2)))
5247 EVT VT = Op.getOperand(0).getValueType();
5249 // Insertion/extraction are legal for V128 types.
5250 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5251 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5254 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5255 VT != MVT::v1i64 && VT != MVT::v2f32)
5258 // For V64 types, we perform insertion by expanding the value
5259 // to a V128 type and perform the insertion on that.
5261 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5262 EVT WideTy = WideVec.getValueType();
5264 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
5265 Op.getOperand(1), Op.getOperand(2));
5266 // Re-narrow the resultant vector.
5267 return NarrowVector(Node, DAG);
5270 SDValue ARM64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5271 SelectionDAG &DAG) const {
5272 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
5274 // Check for non-constant lane.
5275 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5278 EVT VT = Op.getOperand(0).getValueType();
5280 // Insertion/extraction are legal for V128 types.
5281 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5282 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5285 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5286 VT != MVT::v1i64 && VT != MVT::v2f32)
5289 // For V64 types, we perform extraction by expanding the value
5290 // to a V128 type and perform the extraction on that.
5292 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5293 EVT WideTy = WideVec.getValueType();
5295 EVT ExtrTy = WideTy.getVectorElementType();
5296 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
5299 // For extractions, we just return the result directly.
5300 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
5304 SDValue ARM64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
5305 SelectionDAG &DAG) const {
5306 EVT VT = Op.getOperand(0).getValueType();
5312 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5315 unsigned Val = Cst->getZExtValue();
5317 unsigned Size = Op.getValueType().getSizeInBits();
5321 return DAG.getTargetExtractSubreg(ARM64::bsub, dl, Op.getValueType(),
5324 return DAG.getTargetExtractSubreg(ARM64::hsub, dl, Op.getValueType(),
5327 return DAG.getTargetExtractSubreg(ARM64::ssub, dl, Op.getValueType(),
5330 return DAG.getTargetExtractSubreg(ARM64::dsub, dl, Op.getValueType(),
5333 llvm_unreachable("Unexpected vector type in extract_subvector!");
5336 // If this is extracting the upper 64-bits of a 128-bit vector, we match
5338 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
5344 bool ARM64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5346 if (VT.getVectorNumElements() == 4 &&
5347 (VT.is128BitVector() || VT.is64BitVector())) {
5348 unsigned PFIndexes[4];
5349 for (unsigned i = 0; i != 4; ++i) {
5353 PFIndexes[i] = M[i];
5356 // Compute the index in the perfect shuffle table.
5357 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5358 PFIndexes[2] * 9 + PFIndexes[3];
5359 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5360 unsigned Cost = (PFEntry >> 30);
5368 unsigned DummyUnsigned;
5370 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
5371 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
5372 isEXTMask(M, VT, DummyBool, DummyUnsigned) ||
5373 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
5374 isTRNMask(M, VT, DummyUnsigned) || isUZPMask(M, VT, DummyUnsigned) ||
5375 isZIPMask(M, VT, DummyUnsigned) ||
5376 isTRN_v_undef_Mask(M, VT, DummyUnsigned) ||
5377 isUZP_v_undef_Mask(M, VT, DummyUnsigned) ||
5378 isZIP_v_undef_Mask(M, VT, DummyUnsigned) ||
5379 isINSMask(M, VT.getVectorNumElements(), DummyBool, DummyInt) ||
5380 isConcatMask(M, VT, VT.getSizeInBits() == 128));
5383 /// getVShiftImm - Check if this is a valid build_vector for the immediate
5384 /// operand of a vector shift operation, where all the elements of the
5385 /// build_vector must have the same constant integer value.
5386 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
5387 // Ignore bit_converts.
5388 while (Op.getOpcode() == ISD::BITCAST)
5389 Op = Op.getOperand(0);
5390 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
5391 APInt SplatBits, SplatUndef;
5392 unsigned SplatBitSize;
5394 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
5395 HasAnyUndefs, ElementBits) ||
5396 SplatBitSize > ElementBits)
5398 Cnt = SplatBits.getSExtValue();
5402 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
5403 /// operand of a vector shift left operation. That value must be in the range:
5404 /// 0 <= Value < ElementBits for a left shift; or
5405 /// 0 <= Value <= ElementBits for a long left shift.
5406 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
5407 assert(VT.isVector() && "vector shift count is not a vector type");
5408 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5409 if (!getVShiftImm(Op, ElementBits, Cnt))
5411 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
5414 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
5415 /// operand of a vector shift right operation. For a shift opcode, the value
5416 /// is positive, but for an intrinsic the value count must be negative. The
5417 /// absolute value must be in the range:
5418 /// 1 <= |Value| <= ElementBits for a right shift; or
5419 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
5420 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
5422 assert(VT.isVector() && "vector shift count is not a vector type");
5423 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5424 if (!getVShiftImm(Op, ElementBits, Cnt))
5428 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
5431 SDValue ARM64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
5432 SelectionDAG &DAG) const {
5433 EVT VT = Op.getValueType();
5437 if (!Op.getOperand(1).getValueType().isVector())
5439 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5441 switch (Op.getOpcode()) {
5443 llvm_unreachable("unexpected shift opcode");
5446 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
5447 return DAG.getNode(ARM64ISD::VSHL, SDLoc(Op), VT, Op.getOperand(0),
5448 DAG.getConstant(Cnt, MVT::i32));
5449 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5450 DAG.getConstant(Intrinsic::arm64_neon_ushl, MVT::i32),
5451 Op.getOperand(0), Op.getOperand(1));
5454 // Right shift immediate
5455 if (isVShiftRImm(Op.getOperand(1), VT, false, false, Cnt) &&
5458 (Op.getOpcode() == ISD::SRA) ? ARM64ISD::VASHR : ARM64ISD::VLSHR;
5459 return DAG.getNode(Opc, SDLoc(Op), VT, Op.getOperand(0),
5460 DAG.getConstant(Cnt, MVT::i32));
5463 // Right shift register. Note, there is not a shift right register
5464 // instruction, but the shift left register instruction takes a signed
5465 // value, where negative numbers specify a right shift.
5466 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::arm64_neon_sshl
5467 : Intrinsic::arm64_neon_ushl;
5468 // negate the shift amount
5469 SDValue NegShift = DAG.getNode(ARM64ISD::NEG, DL, VT, Op.getOperand(1));
5470 SDValue NegShiftLeft =
5471 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5472 DAG.getConstant(Opc, MVT::i32), Op.getOperand(0), NegShift);
5473 return NegShiftLeft;
5479 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
5480 ARM64CC::CondCode CC, bool NoNans, EVT VT,
5481 SDLoc dl, SelectionDAG &DAG) {
5482 EVT SrcVT = LHS.getValueType();
5484 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
5485 APInt CnstBits(VT.getSizeInBits(), 0);
5486 APInt UndefBits(VT.getSizeInBits(), 0);
5487 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
5488 bool IsZero = IsCnst && (CnstBits == 0);
5490 if (SrcVT.getVectorElementType().isFloatingPoint()) {
5497 Fcmeq = DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5499 Fcmeq = DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5500 return DAG.getNode(ARM64ISD::NOT, dl, VT, Fcmeq);
5504 return DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5505 return DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5508 return DAG.getNode(ARM64ISD::FCMGEz, dl, VT, LHS);
5509 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, LHS, RHS);
5512 return DAG.getNode(ARM64ISD::FCMGTz, dl, VT, LHS);
5513 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, LHS, RHS);
5516 return DAG.getNode(ARM64ISD::FCMLEz, dl, VT, LHS);
5517 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, RHS, LHS);
5521 // If we ignore NaNs then we can use to the MI implementation.
5525 return DAG.getNode(ARM64ISD::FCMLTz, dl, VT, LHS);
5526 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, RHS, LHS);
5536 Cmeq = DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5538 Cmeq = DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5539 return DAG.getNode(ARM64ISD::NOT, dl, VT, Cmeq);
5543 return DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5544 return DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5547 return DAG.getNode(ARM64ISD::CMGEz, dl, VT, LHS);
5548 return DAG.getNode(ARM64ISD::CMGE, dl, VT, LHS, RHS);
5551 return DAG.getNode(ARM64ISD::CMGTz, dl, VT, LHS);
5552 return DAG.getNode(ARM64ISD::CMGT, dl, VT, LHS, RHS);
5555 return DAG.getNode(ARM64ISD::CMLEz, dl, VT, LHS);
5556 return DAG.getNode(ARM64ISD::CMGE, dl, VT, RHS, LHS);
5558 return DAG.getNode(ARM64ISD::CMHS, dl, VT, RHS, LHS);
5560 return DAG.getNode(ARM64ISD::CMHI, dl, VT, RHS, LHS);
5563 return DAG.getNode(ARM64ISD::CMLTz, dl, VT, LHS);
5564 return DAG.getNode(ARM64ISD::CMGT, dl, VT, RHS, LHS);
5566 return DAG.getNode(ARM64ISD::CMHI, dl, VT, LHS, RHS);
5568 return DAG.getNode(ARM64ISD::CMHS, dl, VT, LHS, RHS);
5572 SDValue ARM64TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
5573 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5574 SDValue LHS = Op.getOperand(0);
5575 SDValue RHS = Op.getOperand(1);
5578 if (LHS.getValueType().getVectorElementType().isInteger()) {
5579 assert(LHS.getValueType() == RHS.getValueType());
5580 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
5581 return EmitVectorComparison(LHS, RHS, ARM64CC, false, Op.getValueType(), dl,
5585 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
5586 LHS.getValueType().getVectorElementType() == MVT::f64);
5588 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
5589 // clean. Some of them require two branches to implement.
5590 ARM64CC::CondCode CC1, CC2;
5592 changeVectorFPCCToARM64CC(CC, CC1, CC2, ShouldInvert);
5594 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
5596 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, Op.getValueType(), dl, DAG);
5600 if (CC2 != ARM64CC::AL) {
5602 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, Op.getValueType(), dl, DAG);
5603 if (!Cmp2.getNode())
5606 Cmp = DAG.getNode(ISD::OR, dl, Cmp.getValueType(), Cmp, Cmp2);
5610 return Cmp = DAG.getNOT(dl, Cmp, Cmp.getValueType());
5615 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
5616 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
5617 /// specified in the intrinsic calls.
5618 bool ARM64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5620 unsigned Intrinsic) const {
5621 switch (Intrinsic) {
5622 case Intrinsic::arm64_neon_ld2:
5623 case Intrinsic::arm64_neon_ld3:
5624 case Intrinsic::arm64_neon_ld4:
5625 case Intrinsic::arm64_neon_ld2lane:
5626 case Intrinsic::arm64_neon_ld3lane:
5627 case Intrinsic::arm64_neon_ld4lane:
5628 case Intrinsic::arm64_neon_ld2r:
5629 case Intrinsic::arm64_neon_ld3r:
5630 case Intrinsic::arm64_neon_ld4r: {
5631 Info.opc = ISD::INTRINSIC_W_CHAIN;
5632 // Conservatively set memVT to the entire set of vectors loaded.
5633 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
5634 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5635 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5638 Info.vol = false; // volatile loads with NEON intrinsics not supported
5639 Info.readMem = true;
5640 Info.writeMem = false;
5643 case Intrinsic::arm64_neon_st2:
5644 case Intrinsic::arm64_neon_st3:
5645 case Intrinsic::arm64_neon_st4:
5646 case Intrinsic::arm64_neon_st2lane:
5647 case Intrinsic::arm64_neon_st3lane:
5648 case Intrinsic::arm64_neon_st4lane: {
5649 Info.opc = ISD::INTRINSIC_VOID;
5650 // Conservatively set memVT to the entire set of vectors stored.
5651 unsigned NumElts = 0;
5652 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5653 Type *ArgTy = I.getArgOperand(ArgI)->getType();
5654 if (!ArgTy->isVectorTy())
5656 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
5658 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5659 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5662 Info.vol = false; // volatile stores with NEON intrinsics not supported
5663 Info.readMem = false;
5664 Info.writeMem = true;
5667 case Intrinsic::arm64_ldaxr:
5668 case Intrinsic::arm64_ldxr: {
5669 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
5670 Info.opc = ISD::INTRINSIC_W_CHAIN;
5671 Info.memVT = MVT::getVT(PtrTy->getElementType());
5672 Info.ptrVal = I.getArgOperand(0);
5674 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
5676 Info.readMem = true;
5677 Info.writeMem = false;
5680 case Intrinsic::arm64_stlxr:
5681 case Intrinsic::arm64_stxr: {
5682 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
5683 Info.opc = ISD::INTRINSIC_W_CHAIN;
5684 Info.memVT = MVT::getVT(PtrTy->getElementType());
5685 Info.ptrVal = I.getArgOperand(1);
5687 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
5689 Info.readMem = false;
5690 Info.writeMem = true;
5693 case Intrinsic::arm64_ldaxp:
5694 case Intrinsic::arm64_ldxp: {
5695 Info.opc = ISD::INTRINSIC_W_CHAIN;
5696 Info.memVT = MVT::i128;
5697 Info.ptrVal = I.getArgOperand(0);
5701 Info.readMem = true;
5702 Info.writeMem = false;
5705 case Intrinsic::arm64_stlxp:
5706 case Intrinsic::arm64_stxp: {
5707 Info.opc = ISD::INTRINSIC_W_CHAIN;
5708 Info.memVT = MVT::i128;
5709 Info.ptrVal = I.getArgOperand(2);
5713 Info.readMem = false;
5714 Info.writeMem = true;
5724 // Truncations from 64-bit GPR to 32-bit GPR is free.
5725 bool ARM64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
5726 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
5728 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5729 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5730 if (NumBits1 <= NumBits2)
5734 bool ARM64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
5735 if (!VT1.isInteger() || !VT2.isInteger())
5737 unsigned NumBits1 = VT1.getSizeInBits();
5738 unsigned NumBits2 = VT2.getSizeInBits();
5739 if (NumBits1 <= NumBits2)
5744 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
5746 bool ARM64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
5747 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
5749 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5750 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5751 if (NumBits1 == 32 && NumBits2 == 64)
5755 bool ARM64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
5756 if (!VT1.isInteger() || !VT2.isInteger())
5758 unsigned NumBits1 = VT1.getSizeInBits();
5759 unsigned NumBits2 = VT2.getSizeInBits();
5760 if (NumBits1 == 32 && NumBits2 == 64)
5765 bool ARM64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
5766 EVT VT1 = Val.getValueType();
5767 if (isZExtFree(VT1, VT2)) {
5771 if (Val.getOpcode() != ISD::LOAD)
5774 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
5775 return (VT1.isSimple() && VT1.isInteger() && VT2.isSimple() &&
5776 VT2.isInteger() && VT1.getSizeInBits() <= 32);
5779 bool ARM64TargetLowering::hasPairedLoad(Type *LoadedType,
5780 unsigned &RequiredAligment) const {
5781 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
5783 // Cyclone supports unaligned accesses.
5784 RequiredAligment = 0;
5785 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
5786 return NumBits == 32 || NumBits == 64;
5789 bool ARM64TargetLowering::hasPairedLoad(EVT LoadedType,
5790 unsigned &RequiredAligment) const {
5791 if (!LoadedType.isSimple() ||
5792 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
5794 // Cyclone supports unaligned accesses.
5795 RequiredAligment = 0;
5796 unsigned NumBits = LoadedType.getSizeInBits();
5797 return NumBits == 32 || NumBits == 64;
5800 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
5801 unsigned AlignCheck) {
5802 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
5803 (DstAlign == 0 || DstAlign % AlignCheck == 0));
5806 EVT ARM64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
5807 unsigned SrcAlign, bool IsMemset,
5808 bool ZeroMemset, bool MemcpyStrSrc,
5809 MachineFunction &MF) const {
5810 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
5811 // instruction to materialize the v2i64 zero and one store (with restrictive
5812 // addressing mode). Just do two i64 store of zero-registers.
5814 const Function *F = MF.getFunction();
5815 if (Subtarget->hasFPARMv8() && !IsMemset && Size >= 16 &&
5816 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
5817 Attribute::NoImplicitFloat) &&
5818 (memOpAlign(SrcAlign, DstAlign, 16) ||
5819 (allowsUnalignedMemoryAccesses(MVT::f128, 0, &Fast) && Fast)))
5822 return Size >= 8 ? MVT::i64 : MVT::i32;
5825 // 12-bit optionally shifted immediates are legal for adds.
5826 bool ARM64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
5827 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
5832 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
5833 // immediates is the same as for an add or a sub.
5834 bool ARM64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
5837 return isLegalAddImmediate(Immed);
5840 /// isLegalAddressingMode - Return true if the addressing mode represented
5841 /// by AM is legal for this target, for a load/store of the specified type.
5842 bool ARM64TargetLowering::isLegalAddressingMode(const AddrMode &AM,
5844 // ARM64 has five basic addressing modes:
5846 // reg + 9-bit signed offset
5847 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
5849 // reg + SIZE_IN_BYTES * reg
5851 // No global is ever allowed as a base.
5855 // No reg+reg+imm addressing.
5856 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
5859 // check reg + imm case:
5860 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
5861 uint64_t NumBytes = 0;
5862 if (Ty->isSized()) {
5863 uint64_t NumBits = getDataLayout()->getTypeSizeInBits(Ty);
5864 NumBytes = NumBits / 8;
5865 if (!isPowerOf2_64(NumBits))
5870 int64_t Offset = AM.BaseOffs;
5872 // 9-bit signed offset
5873 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
5876 // 12-bit unsigned offset
5877 unsigned shift = Log2_64(NumBytes);
5878 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
5879 // Must be a multiple of NumBytes (NumBytes is a power of 2)
5880 (Offset >> shift) << shift == Offset)
5885 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
5887 if (!AM.Scale || AM.Scale == 1 ||
5888 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
5893 int ARM64TargetLowering::getScalingFactorCost(const AddrMode &AM,
5895 // Scaling factors are not free at all.
5896 // Operands | Rt Latency
5897 // -------------------------------------------
5899 // -------------------------------------------
5900 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
5901 // Rt, [Xn, Wm, <extend> #imm] |
5902 if (isLegalAddressingMode(AM, Ty))
5903 // Scale represents reg2 * scale, thus account for 1 if
5904 // it is not equal to 0 or 1.
5905 return AM.Scale != 0 && AM.Scale != 1;
5909 bool ARM64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
5910 VT = VT.getScalarType();
5915 switch (VT.getSimpleVT().SimpleTy) {
5927 ARM64TargetLowering::getScratchRegisters(CallingConv::ID) const {
5928 // LR is a callee-save register, but we must treat it as clobbered by any call
5929 // site. Hence we include LR in the scratch registers, which are in turn added
5930 // as implicit-defs for stackmaps and patchpoints.
5931 static const MCPhysReg ScratchRegs[] = {
5932 ARM64::X16, ARM64::X17, ARM64::LR, 0
5937 bool ARM64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
5939 assert(Ty->isIntegerTy());
5941 unsigned BitSize = Ty->getPrimitiveSizeInBits();
5945 int64_t Val = Imm.getSExtValue();
5946 if (Val == 0 || ARM64_AM::isLogicalImmediate(Val, BitSize))
5949 if ((int64_t)Val < 0)
5952 Val &= (1LL << 32) - 1;
5954 unsigned LZ = countLeadingZeros((uint64_t)Val);
5955 unsigned Shift = (63 - LZ) / 16;
5956 // MOVZ is free so return true for one or fewer MOVK.
5957 return (Shift < 3) ? true : false;
5960 // Generate SUBS and CSEL for integer abs.
5961 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
5962 EVT VT = N->getValueType(0);
5964 SDValue N0 = N->getOperand(0);
5965 SDValue N1 = N->getOperand(1);
5968 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
5969 // and change it to SUB and CSEL.
5970 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
5971 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
5972 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
5973 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
5974 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
5975 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
5977 // Generate SUBS & CSEL.
5979 DAG.getNode(ARM64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
5980 N0.getOperand(0), DAG.getConstant(0, VT));
5981 return DAG.getNode(ARM64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
5982 DAG.getConstant(ARM64CC::PL, MVT::i32),
5983 SDValue(Cmp.getNode(), 1));
5988 // performXorCombine - Attempts to handle integer ABS.
5989 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
5990 TargetLowering::DAGCombinerInfo &DCI,
5991 const ARM64Subtarget *Subtarget) {
5992 if (DCI.isBeforeLegalizeOps())
5995 return performIntegerAbsCombine(N, DAG);
5998 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
5999 TargetLowering::DAGCombinerInfo &DCI,
6000 const ARM64Subtarget *Subtarget) {
6001 if (DCI.isBeforeLegalizeOps())
6004 // Multiplication of a power of two plus/minus one can be done more
6005 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
6006 // future CPUs have a cheaper MADD instruction, this may need to be
6007 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
6008 // 64-bit is 5 cycles, so this is always a win.
6009 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
6010 APInt Value = C->getAPIntValue();
6011 EVT VT = N->getValueType(0);
6012 APInt VP1 = Value + 1;
6013 if (VP1.isPowerOf2()) {
6014 // Multiplying by one less than a power of two, replace with a shift
6016 SDValue ShiftedVal =
6017 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6018 DAG.getConstant(VP1.logBase2(), MVT::i64));
6019 return DAG.getNode(ISD::SUB, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6021 APInt VM1 = Value - 1;
6022 if (VM1.isPowerOf2()) {
6023 // Multiplying by one more than a power of two, replace with a shift
6025 SDValue ShiftedVal =
6026 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6027 DAG.getConstant(VM1.logBase2(), MVT::i64));
6028 return DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6034 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG) {
6035 EVT VT = N->getValueType(0);
6036 if (VT != MVT::f32 && VT != MVT::f64)
6038 // Only optimize when the source and destination types have the same width.
6039 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
6042 // If the result of an integer load is only used by an integer-to-float
6043 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
6044 // This eliminates an "integer-to-vector-move UOP and improve throughput.
6045 SDValue N0 = N->getOperand(0);
6046 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6047 // Do not change the width of a volatile load.
6048 !cast<LoadSDNode>(N0)->isVolatile()) {
6049 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6050 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
6051 LN0->getPointerInfo(), LN0->isVolatile(),
6052 LN0->isNonTemporal(), LN0->isInvariant(),
6053 LN0->getAlignment());
6055 // Make sure successors of the original load stay after it by updating them
6056 // to use the new Chain.
6057 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
6060 (N->getOpcode() == ISD::SINT_TO_FP) ? ARM64ISD::SITOF : ARM64ISD::UITOF;
6061 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
6067 /// An EXTR instruction is made up of two shifts, ORed together. This helper
6068 /// searches for and classifies those shifts.
6069 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
6071 if (N.getOpcode() == ISD::SHL)
6073 else if (N.getOpcode() == ISD::SRL)
6078 if (!isa<ConstantSDNode>(N.getOperand(1)))
6081 ShiftAmount = N->getConstantOperandVal(1);
6082 Src = N->getOperand(0);
6086 /// EXTR instruction extracts a contiguous chunk of bits from two existing
6087 /// registers viewed as a high/low pair. This function looks for the pattern:
6088 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
6089 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
6091 static SDValue tryCombineToEXTR(SDNode *N,
6092 TargetLowering::DAGCombinerInfo &DCI) {
6093 SelectionDAG &DAG = DCI.DAG;
6095 EVT VT = N->getValueType(0);
6097 assert(N->getOpcode() == ISD::OR && "Unexpected root");
6099 if (VT != MVT::i32 && VT != MVT::i64)
6103 uint32_t ShiftLHS = 0;
6105 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
6109 uint32_t ShiftRHS = 0;
6111 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
6114 // If they're both trying to come from the high part of the register, they're
6115 // not really an EXTR.
6116 if (LHSFromHi == RHSFromHi)
6119 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
6123 std::swap(LHS, RHS);
6124 std::swap(ShiftLHS, ShiftRHS);
6127 return DAG.getNode(ARM64ISD::EXTR, DL, VT, LHS, RHS,
6128 DAG.getConstant(ShiftRHS, MVT::i64));
6131 static SDValue tryCombineToBSL(SDNode *N,
6132 TargetLowering::DAGCombinerInfo &DCI) {
6133 EVT VT = N->getValueType(0);
6134 SelectionDAG &DAG = DCI.DAG;
6140 SDValue N0 = N->getOperand(0);
6141 if (N0.getOpcode() != ISD::AND)
6144 SDValue N1 = N->getOperand(1);
6145 if (N1.getOpcode() != ISD::AND)
6148 // We only have to look for constant vectors here since the general, variable
6149 // case can be handled in TableGen.
6150 unsigned Bits = VT.getVectorElementType().getSizeInBits();
6151 uint64_t BitMask = Bits == 64 ? -1ULL : ((1ULL << Bits) - 1);
6152 for (int i = 1; i >= 0; --i)
6153 for (int j = 1; j >= 0; --j) {
6154 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(i));
6155 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(j));
6159 bool FoundMatch = true;
6160 for (unsigned k = 0; k < VT.getVectorNumElements(); ++k) {
6161 ConstantSDNode *CN0 = dyn_cast<ConstantSDNode>(BVN0->getOperand(k));
6162 ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(BVN1->getOperand(k));
6164 CN0->getZExtValue() != (BitMask & ~CN1->getZExtValue())) {
6171 return DAG.getNode(ARM64ISD::BSL, DL, VT, SDValue(BVN0, 0),
6172 N0->getOperand(1 - i), N1->getOperand(1 - j));
6178 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
6179 const ARM64Subtarget *Subtarget) {
6180 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
6181 if (!EnableARM64ExtrGeneration)
6183 SelectionDAG &DAG = DCI.DAG;
6184 EVT VT = N->getValueType(0);
6186 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6189 SDValue Res = tryCombineToEXTR(N, DCI);
6193 Res = tryCombineToBSL(N, DCI);
6200 static SDValue performBitcastCombine(SDNode *N,
6201 TargetLowering::DAGCombinerInfo &DCI,
6202 SelectionDAG &DAG) {
6203 // Wait 'til after everything is legalized to try this. That way we have
6204 // legal vector types and such.
6205 if (DCI.isBeforeLegalizeOps())
6208 // Remove extraneous bitcasts around an extract_subvector.
6210 // (v4i16 (bitconvert
6211 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
6213 // (extract_subvector ((v8i16 ...), (i64 4)))
6215 // Only interested in 64-bit vectors as the ultimate result.
6216 EVT VT = N->getValueType(0);
6219 if (VT.getSimpleVT().getSizeInBits() != 64)
6221 // Is the operand an extract_subvector starting at the beginning or halfway
6222 // point of the vector? A low half may also come through as an
6223 // EXTRACT_SUBREG, so look for that, too.
6224 SDValue Op0 = N->getOperand(0);
6225 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
6226 !(Op0->isMachineOpcode() &&
6227 Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG))
6229 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
6230 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
6231 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
6233 } else if (Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG) {
6234 if (idx != ARM64::dsub)
6236 // The dsub reference is equivalent to a lane zero subvector reference.
6239 // Look through the bitcast of the input to the extract.
6240 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
6242 SDValue Source = Op0->getOperand(0)->getOperand(0);
6243 // If the source type has twice the number of elements as our destination
6244 // type, we know this is an extract of the high or low half of the vector.
6245 EVT SVT = Source->getValueType(0);
6246 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
6249 DEBUG(dbgs() << "arm64-lower: bitcast extract_subvector simplification\n");
6251 // Create the simplified form to just extract the low or high half of the
6252 // vector directly rather than bothering with the bitcasts.
6254 unsigned NumElements = VT.getVectorNumElements();
6256 SDValue HalfIdx = DAG.getConstant(NumElements, MVT::i64);
6257 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
6259 SDValue SubReg = DAG.getTargetConstant(ARM64::dsub, MVT::i32);
6260 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
6266 static SDValue performConcatVectorsCombine(SDNode *N,
6267 TargetLowering::DAGCombinerInfo &DCI,
6268 SelectionDAG &DAG) {
6269 // Wait 'til after everything is legalized to try this. That way we have
6270 // legal vector types and such.
6271 if (DCI.isBeforeLegalizeOps())
6275 EVT VT = N->getValueType(0);
6277 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
6278 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
6279 // canonicalise to that.
6280 if (N->getOperand(0) == N->getOperand(1) && VT.getVectorNumElements() == 2) {
6281 assert(VT.getVectorElementType().getSizeInBits() == 64);
6282 return DAG.getNode(ARM64ISD::DUPLANE64, dl, VT,
6283 WidenVector(N->getOperand(0), DAG),
6284 DAG.getConstant(0, MVT::i64));
6287 // Canonicalise concat_vectors so that the right-hand vector has as few
6288 // bit-casts as possible before its real operation. The primary matching
6289 // destination for these operations will be the narrowing "2" instructions,
6290 // which depend on the operation being performed on this right-hand vector.
6292 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
6294 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
6296 SDValue Op1 = N->getOperand(1);
6297 if (Op1->getOpcode() != ISD::BITCAST)
6299 SDValue RHS = Op1->getOperand(0);
6300 MVT RHSTy = RHS.getValueType().getSimpleVT();
6301 // If the RHS is not a vector, this is not the pattern we're looking for.
6302 if (!RHSTy.isVector())
6305 DEBUG(dbgs() << "arm64-lower: concat_vectors bitcast simplification\n");
6307 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
6308 RHSTy.getVectorNumElements() * 2);
6310 ISD::BITCAST, dl, VT,
6311 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
6312 DAG.getNode(ISD::BITCAST, dl, RHSTy, N->getOperand(0)), RHS));
6315 static SDValue tryCombineFixedPointConvert(SDNode *N,
6316 TargetLowering::DAGCombinerInfo &DCI,
6317 SelectionDAG &DAG) {
6318 // Wait 'til after everything is legalized to try this. That way we have
6319 // legal vector types and such.
6320 if (DCI.isBeforeLegalizeOps())
6322 // Transform a scalar conversion of a value from a lane extract into a
6323 // lane extract of a vector conversion. E.g., from foo1 to foo2:
6324 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
6325 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
6327 // The second form interacts better with instruction selection and the
6328 // register allocator to avoid cross-class register copies that aren't
6329 // coalescable due to a lane reference.
6331 // Check the operand and see if it originates from a lane extract.
6332 SDValue Op1 = N->getOperand(1);
6333 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6334 // Yep, no additional predication needed. Perform the transform.
6335 SDValue IID = N->getOperand(0);
6336 SDValue Shift = N->getOperand(2);
6337 SDValue Vec = Op1.getOperand(0);
6338 SDValue Lane = Op1.getOperand(1);
6339 EVT ResTy = N->getValueType(0);
6343 // The vector width should be 128 bits by the time we get here, even
6344 // if it started as 64 bits (the extract_vector handling will have
6346 assert(Vec.getValueType().getSizeInBits() == 128 &&
6347 "unexpected vector size on extract_vector_elt!");
6348 if (Vec.getValueType() == MVT::v4i32)
6349 VecResTy = MVT::v4f32;
6350 else if (Vec.getValueType() == MVT::v2i64)
6351 VecResTy = MVT::v2f64;
6353 assert(0 && "unexpected vector type!");
6356 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
6357 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
6362 // AArch64 high-vector "long" operations are formed by performing the non-high
6363 // version on an extract_subvector of each operand which gets the high half:
6365 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
6367 // However, there are cases which don't have an extract_high explicitly, but
6368 // have another operation that can be made compatible with one for free. For
6371 // (dupv64 scalar) --> (extract_high (dup128 scalar))
6373 // This routine does the actual conversion of such DUPs, once outer routines
6374 // have determined that everything else is in order.
6375 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
6376 // We can handle most types of duplicate, but the lane ones have an extra
6377 // operand saying *which* lane, so we need to know.
6379 switch (N.getOpcode()) {
6383 case ARM64ISD::DUPLANE8:
6384 case ARM64ISD::DUPLANE16:
6385 case ARM64ISD::DUPLANE32:
6386 case ARM64ISD::DUPLANE64:
6393 MVT NarrowTy = N.getSimpleValueType();
6394 if (!NarrowTy.is64BitVector())
6397 MVT ElementTy = NarrowTy.getVectorElementType();
6398 unsigned NumElems = NarrowTy.getVectorNumElements();
6399 MVT NewDUPVT = MVT::getVectorVT(ElementTy, NumElems * 2);
6403 NewDUP = DAG.getNode(N.getOpcode(), SDLoc(N), NewDUPVT, N.getOperand(0),
6406 NewDUP = DAG.getNode(ARM64ISD::DUP, SDLoc(N), NewDUPVT, N.getOperand(0));
6408 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N.getNode()), NarrowTy,
6409 NewDUP, DAG.getConstant(NumElems, MVT::i64));
6412 static bool isEssentiallyExtractSubvector(SDValue N) {
6413 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
6416 return N.getOpcode() == ISD::BITCAST &&
6417 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
6420 /// \brief Helper structure to keep track of ISD::SET_CC operands.
6421 struct GenericSetCCInfo {
6422 const SDValue *Opnd0;
6423 const SDValue *Opnd1;
6427 /// \brief Helper structure to keep track of a SET_CC lowered into ARM64 code.
6428 struct ARM64SetCCInfo {
6430 ARM64CC::CondCode CC;
6433 /// \brief Helper structure to keep track of SetCC information.
6435 GenericSetCCInfo Generic;
6436 ARM64SetCCInfo ARM64;
6439 /// \brief Helper structure to be able to read SetCC information.
6440 /// If set to true, IsARM64 field, Info is a ARM64SetCCInfo, otherwise Info is
6441 /// a GenericSetCCInfo.
6442 struct SetCCInfoAndKind {
6447 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
6449 /// ARM64 lowered one.
6450 /// \p SetCCInfo is filled accordingly.
6451 /// \post SetCCInfo is meanginfull only when this function returns true.
6452 /// \return True when Op is a kind of SET_CC operation.
6453 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
6454 // If this is a setcc, this is straight forward.
6455 if (Op.getOpcode() == ISD::SETCC) {
6456 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
6457 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
6458 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6459 SetCCInfo.IsARM64 = false;
6462 // Otherwise, check if this is a matching csel instruction.
6466 if (Op.getOpcode() != ARM64ISD::CSEL)
6468 // Set the information about the operands.
6469 // TODO: we want the operands of the Cmp not the csel
6470 SetCCInfo.Info.ARM64.Cmp = &Op.getOperand(3);
6471 SetCCInfo.IsARM64 = true;
6472 SetCCInfo.Info.ARM64.CC = static_cast<ARM64CC::CondCode>(
6473 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
6475 // Check that the operands matches the constraints:
6476 // (1) Both operands must be constants.
6477 // (2) One must be 1 and the other must be 0.
6478 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
6479 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6482 if (!TValue || !FValue)
6486 if (!TValue->isOne()) {
6487 // Update the comparison when we are interested in !cc.
6488 std::swap(TValue, FValue);
6489 SetCCInfo.Info.ARM64.CC =
6490 ARM64CC::getInvertedCondCode(SetCCInfo.Info.ARM64.CC);
6492 return TValue->isOne() && FValue->isNullValue();
6495 // The folding we want to perform is:
6496 // (add x, (setcc cc ...) )
6498 // (csel x, (add x, 1), !cc ...)
6500 // The latter will get matched to a CSINC instruction.
6501 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
6502 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
6503 SDValue LHS = Op->getOperand(0);
6504 SDValue RHS = Op->getOperand(1);
6505 SetCCInfoAndKind InfoAndKind;
6507 // If neither operand is a SET_CC, give up.
6508 if (!isSetCC(LHS, InfoAndKind)) {
6509 std::swap(LHS, RHS);
6510 if (!isSetCC(LHS, InfoAndKind))
6514 // FIXME: This could be generatized to work for FP comparisons.
6515 EVT CmpVT = InfoAndKind.IsARM64
6516 ? InfoAndKind.Info.ARM64.Cmp->getOperand(0).getValueType()
6517 : InfoAndKind.Info.Generic.Opnd0->getValueType();
6518 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
6524 if (InfoAndKind.IsARM64) {
6525 CCVal = DAG.getConstant(
6526 ARM64CC::getInvertedCondCode(InfoAndKind.Info.ARM64.CC), MVT::i32);
6527 Cmp = *InfoAndKind.Info.ARM64.Cmp;
6529 Cmp = getARM64Cmp(*InfoAndKind.Info.Generic.Opnd0,
6530 *InfoAndKind.Info.Generic.Opnd1,
6531 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
6534 EVT VT = Op->getValueType(0);
6535 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, VT));
6536 return DAG.getNode(ARM64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
6539 // The basic add/sub long vector instructions have variants with "2" on the end
6540 // which act on the high-half of their inputs. They are normally matched by
6543 // (add (zeroext (extract_high LHS)),
6544 // (zeroext (extract_high RHS)))
6545 // -> uaddl2 vD, vN, vM
6547 // However, if one of the extracts is something like a duplicate, this
6548 // instruction can still be used profitably. This function puts the DAG into a
6549 // more appropriate form for those patterns to trigger.
6550 static SDValue performAddSubLongCombine(SDNode *N,
6551 TargetLowering::DAGCombinerInfo &DCI,
6552 SelectionDAG &DAG) {
6553 if (DCI.isBeforeLegalizeOps())
6556 MVT VT = N->getSimpleValueType(0);
6557 if (!VT.is128BitVector()) {
6558 if (N->getOpcode() == ISD::ADD)
6559 return performSetccAddFolding(N, DAG);
6563 // Make sure both branches are extended in the same way.
6564 SDValue LHS = N->getOperand(0);
6565 SDValue RHS = N->getOperand(1);
6566 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
6567 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
6568 LHS.getOpcode() != RHS.getOpcode())
6571 unsigned ExtType = LHS.getOpcode();
6573 // It's not worth doing if at least one of the inputs isn't already an
6574 // extract, but we don't know which it'll be so we have to try both.
6575 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
6576 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
6580 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
6581 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
6582 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
6586 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
6589 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
6592 // Massage DAGs which we can use the high-half "long" operations on into
6593 // something isel will recognize better. E.g.
6595 // (arm64_neon_umull (extract_high vec) (dupv64 scalar)) -->
6596 // (arm64_neon_umull (extract_high (v2i64 vec)))
6597 // (extract_high (v2i64 (dup128 scalar)))))
6599 static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N,
6600 TargetLowering::DAGCombinerInfo &DCI,
6601 SelectionDAG &DAG) {
6602 if (DCI.isBeforeLegalizeOps())
6605 SDValue LHS = N->getOperand(1);
6606 SDValue RHS = N->getOperand(2);
6607 assert(LHS.getValueType().is64BitVector() &&
6608 RHS.getValueType().is64BitVector() &&
6609 "unexpected shape for long operation");
6611 // Either node could be a DUP, but it's not worth doing both of them (you'd
6612 // just as well use the non-high version) so look for a corresponding extract
6613 // operation on the other "wing".
6614 if (isEssentiallyExtractSubvector(LHS)) {
6615 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
6618 } else if (isEssentiallyExtractSubvector(RHS)) {
6619 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
6624 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
6625 N->getOperand(0), LHS, RHS);
6628 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
6629 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
6630 unsigned ElemBits = ElemTy.getSizeInBits();
6632 int64_t ShiftAmount;
6633 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
6634 APInt SplatValue, SplatUndef;
6635 unsigned SplatBitSize;
6637 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
6638 HasAnyUndefs, ElemBits) ||
6639 SplatBitSize != ElemBits)
6642 ShiftAmount = SplatValue.getSExtValue();
6643 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
6644 ShiftAmount = CVN->getSExtValue();
6652 llvm_unreachable("Unknown shift intrinsic");
6653 case Intrinsic::arm64_neon_sqshl:
6654 Opcode = ARM64ISD::SQSHL_I;
6655 IsRightShift = false;
6657 case Intrinsic::arm64_neon_uqshl:
6658 Opcode = ARM64ISD::UQSHL_I;
6659 IsRightShift = false;
6661 case Intrinsic::arm64_neon_srshl:
6662 Opcode = ARM64ISD::SRSHR_I;
6663 IsRightShift = true;
6665 case Intrinsic::arm64_neon_urshl:
6666 Opcode = ARM64ISD::URSHR_I;
6667 IsRightShift = true;
6669 case Intrinsic::arm64_neon_sqshlu:
6670 Opcode = ARM64ISD::SQSHLU_I;
6671 IsRightShift = false;
6675 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits)
6676 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6677 DAG.getConstant(-ShiftAmount, MVT::i32));
6678 else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount <= ElemBits)
6679 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6680 DAG.getConstant(ShiftAmount, MVT::i32));
6685 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
6686 // the intrinsics must be legal and take an i32, this means there's almost
6687 // certainly going to be a zext in the DAG which we can eliminate.
6688 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
6689 SDValue AndN = N->getOperand(2);
6690 if (AndN.getOpcode() != ISD::AND)
6693 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
6694 if (!CMask || CMask->getZExtValue() != Mask)
6697 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
6698 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
6701 static SDValue performIntrinsicCombine(SDNode *N,
6702 TargetLowering::DAGCombinerInfo &DCI,
6703 const ARM64Subtarget *Subtarget) {
6704 SelectionDAG &DAG = DCI.DAG;
6705 unsigned IID = getIntrinsicID(N);
6709 case Intrinsic::arm64_neon_vcvtfxs2fp:
6710 case Intrinsic::arm64_neon_vcvtfxu2fp:
6711 return tryCombineFixedPointConvert(N, DCI, DAG);
6713 case Intrinsic::arm64_neon_fmax:
6714 return DAG.getNode(ARM64ISD::FMAX, SDLoc(N), N->getValueType(0),
6715 N->getOperand(1), N->getOperand(2));
6716 case Intrinsic::arm64_neon_fmin:
6717 return DAG.getNode(ARM64ISD::FMIN, SDLoc(N), N->getValueType(0),
6718 N->getOperand(1), N->getOperand(2));
6719 case Intrinsic::arm64_neon_smull:
6720 case Intrinsic::arm64_neon_umull:
6721 case Intrinsic::arm64_neon_pmull:
6722 case Intrinsic::arm64_neon_sqdmull:
6723 return tryCombineLongOpWithDup(IID, N, DCI, DAG);
6724 case Intrinsic::arm64_neon_sqshl:
6725 case Intrinsic::arm64_neon_uqshl:
6726 case Intrinsic::arm64_neon_sqshlu:
6727 case Intrinsic::arm64_neon_srshl:
6728 case Intrinsic::arm64_neon_urshl:
6729 return tryCombineShiftImm(IID, N, DAG);
6730 case Intrinsic::arm64_crc32b:
6731 case Intrinsic::arm64_crc32cb:
6732 return tryCombineCRC32(0xff, N, DAG);
6733 case Intrinsic::arm64_crc32h:
6734 case Intrinsic::arm64_crc32ch:
6735 return tryCombineCRC32(0xffff, N, DAG);
6740 static SDValue performExtendCombine(SDNode *N,
6741 TargetLowering::DAGCombinerInfo &DCI,
6742 SelectionDAG &DAG) {
6743 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
6744 // we can convert that DUP into another extract_high (of a bigger DUP), which
6745 // helps the backend to decide that an sabdl2 would be useful, saving a real
6746 // extract_high operation.
6747 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
6748 N->getOperand(0).getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
6749 SDNode *ABDNode = N->getOperand(0).getNode();
6750 unsigned IID = getIntrinsicID(ABDNode);
6751 if (IID == Intrinsic::arm64_neon_sabd ||
6752 IID == Intrinsic::arm64_neon_uabd) {
6753 SDValue NewABD = tryCombineLongOpWithDup(IID, ABDNode, DCI, DAG);
6754 if (!NewABD.getNode())
6757 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
6762 // This is effectively a custom type legalization for ARM64.
6764 // Type legalization will split an extend of a small, legal, type to a larger
6765 // illegal type by first splitting the destination type, often creating
6766 // illegal source types, which then get legalized in isel-confusing ways,
6767 // leading to really terrible codegen. E.g.,
6768 // %result = v8i32 sext v8i8 %value
6770 // %losrc = extract_subreg %value, ...
6771 // %hisrc = extract_subreg %value, ...
6772 // %lo = v4i32 sext v4i8 %losrc
6773 // %hi = v4i32 sext v4i8 %hisrc
6774 // Things go rapidly downhill from there.
6776 // For ARM64, the [sz]ext vector instructions can only go up one element
6777 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
6778 // take two instructions.
6780 // This implies that the most efficient way to do the extend from v8i8
6781 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
6782 // the normal splitting to happen for the v8i16->v8i32.
6784 // This is pre-legalization to catch some cases where the default
6785 // type legalization will create ill-tempered code.
6786 if (!DCI.isBeforeLegalizeOps())
6789 // We're only interested in cleaning things up for non-legal vector types
6790 // here. If both the source and destination are legal, things will just
6791 // work naturally without any fiddling.
6792 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6793 EVT ResVT = N->getValueType(0);
6794 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
6796 // If the vector type isn't a simple VT, it's beyond the scope of what
6797 // we're worried about here. Let legalization do its thing and hope for
6799 if (!ResVT.isSimple())
6802 SDValue Src = N->getOperand(0);
6803 MVT SrcVT = Src->getValueType(0).getSimpleVT();
6804 // If the source VT is a 64-bit vector, we can play games and get the
6805 // better results we want.
6806 if (SrcVT.getSizeInBits() != 64)
6809 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
6810 unsigned ElementCount = SrcVT.getVectorNumElements();
6811 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
6813 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
6815 // Now split the rest of the operation into two halves, each with a 64
6819 unsigned NumElements = ResVT.getVectorNumElements();
6820 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
6821 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
6822 ResVT.getVectorElementType(), NumElements / 2);
6824 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
6825 LoVT.getVectorNumElements());
6826 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
6827 DAG.getIntPtrConstant(0));
6828 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
6829 DAG.getIntPtrConstant(InNVT.getVectorNumElements()));
6830 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
6831 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
6833 // Now combine the parts back together so we still have a single result
6834 // like the combiner expects.
6835 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
6838 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
6839 /// value. The load store optimizer pass will merge them to store pair stores.
6840 /// This has better performance than a splat of the scalar followed by a split
6841 /// vector store. Even if the stores are not merged it is four stores vs a dup,
6842 /// followed by an ext.b and two stores.
6843 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
6844 SDValue StVal = St->getValue();
6845 EVT VT = StVal.getValueType();
6847 // Don't replace floating point stores, they possibly won't be transformed to
6848 // stp because of the store pair suppress pass.
6849 if (VT.isFloatingPoint())
6852 // Check for insert vector elements.
6853 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
6856 // We can express a splat as store pair(s) for 2 or 4 elements.
6857 unsigned NumVecElts = VT.getVectorNumElements();
6858 if (NumVecElts != 4 && NumVecElts != 2)
6860 SDValue SplatVal = StVal.getOperand(1);
6861 unsigned RemainInsertElts = NumVecElts - 1;
6863 // Check that this is a splat.
6864 while (--RemainInsertElts) {
6865 SDValue NextInsertElt = StVal.getOperand(0);
6866 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
6868 if (NextInsertElt.getOperand(1) != SplatVal)
6870 StVal = NextInsertElt;
6872 unsigned OrigAlignment = St->getAlignment();
6873 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
6874 unsigned Alignment = std::min(OrigAlignment, EltOffset);
6876 // Create scalar stores. This is at least as good as the code sequence for a
6877 // split unaligned store wich is a dup.s, ext.b, and two stores.
6878 // Most of the time the three stores should be replaced by store pair
6879 // instructions (stp).
6881 SDValue BasePtr = St->getBasePtr();
6883 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
6884 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
6886 unsigned Offset = EltOffset;
6887 while (--NumVecElts) {
6888 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
6889 DAG.getConstant(Offset, MVT::i64));
6890 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
6891 St->getPointerInfo(), St->isVolatile(),
6892 St->isNonTemporal(), Alignment);
6893 Offset += EltOffset;
6898 static SDValue performSTORECombine(SDNode *N,
6899 TargetLowering::DAGCombinerInfo &DCI,
6901 const ARM64Subtarget *Subtarget) {
6902 if (!DCI.isBeforeLegalize())
6905 StoreSDNode *S = cast<StoreSDNode>(N);
6906 if (S->isVolatile())
6909 // Cyclone has bad performance on unaligned 16B stores when crossing line and
6910 // page boundries. We want to split such stores.
6911 if (!Subtarget->isCyclone())
6914 // Don't split at Oz.
6915 MachineFunction &MF = DAG.getMachineFunction();
6916 bool IsMinSize = MF.getFunction()->getAttributes().hasAttribute(
6917 AttributeSet::FunctionIndex, Attribute::MinSize);
6921 SDValue StVal = S->getValue();
6922 EVT VT = StVal.getValueType();
6924 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
6925 // those up regresses performance on micro-benchmarks and olden/bh.
6926 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
6929 // Split unaligned 16B stores. They are terrible for performance.
6930 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
6931 // extensions can use this to mark that it does not want splitting to happen
6932 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
6933 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
6934 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
6935 S->getAlignment() <= 2)
6938 // If we get a splat of a scalar convert this vector store to a store of
6939 // scalars. They will be merged into store pairs thereby removing two
6941 SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S);
6942 if (ReplacedSplat != SDValue())
6943 return ReplacedSplat;
6946 unsigned NumElts = VT.getVectorNumElements() / 2;
6947 // Split VT into two.
6949 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
6950 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
6951 DAG.getIntPtrConstant(0));
6952 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
6953 DAG.getIntPtrConstant(NumElts));
6954 SDValue BasePtr = S->getBasePtr();
6956 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
6957 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
6958 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
6959 DAG.getConstant(8, MVT::i64));
6960 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
6961 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
6965 // Optimize compare with zero and branch.
6966 static SDValue performBRCONDCombine(SDNode *N,
6967 TargetLowering::DAGCombinerInfo &DCI,
6968 SelectionDAG &DAG) {
6969 SDValue Chain = N->getOperand(0);
6970 SDValue Dest = N->getOperand(1);
6971 SDValue CCVal = N->getOperand(2);
6972 SDValue Cmp = N->getOperand(3);
6974 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
6975 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
6976 if (CC != ARM64CC::EQ && CC != ARM64CC::NE)
6979 unsigned CmpOpc = Cmp.getOpcode();
6980 if (CmpOpc != ARM64ISD::ADDS && CmpOpc != ARM64ISD::SUBS)
6983 // Only attempt folding if there is only one use of the flag and no use of the
6985 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
6988 SDValue LHS = Cmp.getOperand(0);
6989 SDValue RHS = Cmp.getOperand(1);
6991 assert(LHS.getValueType() == RHS.getValueType() &&
6992 "Expected the value type to be the same for both operands!");
6993 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
6996 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
6997 std::swap(LHS, RHS);
6999 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
7002 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
7003 LHS.getOpcode() == ISD::SRL)
7006 // Fold the compare into the branch instruction.
7008 if (CC == ARM64CC::EQ)
7009 BR = DAG.getNode(ARM64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7011 BR = DAG.getNode(ARM64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7013 // Do not add new nodes to DAG combiner worklist.
7014 DCI.CombineTo(N, BR, false);
7019 // vselect (v1i1 setcc) ->
7020 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
7021 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
7022 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
7024 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
7025 SDValue N0 = N->getOperand(0);
7026 EVT CCVT = N0.getValueType();
7028 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
7029 CCVT.getVectorElementType() != MVT::i1)
7032 EVT ResVT = N->getValueType(0);
7033 EVT CmpVT = N0.getOperand(0).getValueType();
7034 // Only combine when the result type is of the same size as the compared
7036 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
7039 SDValue IfTrue = N->getOperand(1);
7040 SDValue IfFalse = N->getOperand(2);
7042 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
7043 N0.getOperand(0), N0.getOperand(1),
7044 cast<CondCodeSDNode>(N0.getOperand(2))->get());
7045 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
7049 SDValue ARM64TargetLowering::PerformDAGCombine(SDNode *N,
7050 DAGCombinerInfo &DCI) const {
7051 SelectionDAG &DAG = DCI.DAG;
7052 switch (N->getOpcode()) {
7057 return performAddSubLongCombine(N, DCI, DAG);
7059 return performXorCombine(N, DAG, DCI, Subtarget);
7061 return performMulCombine(N, DAG, DCI, Subtarget);
7062 case ISD::SINT_TO_FP:
7063 case ISD::UINT_TO_FP:
7064 return performIntToFpCombine(N, DAG);
7066 return performORCombine(N, DCI, Subtarget);
7067 case ISD::INTRINSIC_WO_CHAIN:
7068 return performIntrinsicCombine(N, DCI, Subtarget);
7069 case ISD::ANY_EXTEND:
7070 case ISD::ZERO_EXTEND:
7071 case ISD::SIGN_EXTEND:
7072 return performExtendCombine(N, DCI, DAG);
7074 return performBitcastCombine(N, DCI, DAG);
7075 case ISD::CONCAT_VECTORS:
7076 return performConcatVectorsCombine(N, DCI, DAG);
7078 return performVSelectCombine(N, DCI.DAG);
7080 return performSTORECombine(N, DCI, DAG, Subtarget);
7081 case ARM64ISD::BRCOND:
7082 return performBRCONDCombine(N, DCI, DAG);
7087 // Check if the return value is used as only a return value, as otherwise
7088 // we can't perform a tail-call. In particular, we need to check for
7089 // target ISD nodes that are returns and any other "odd" constructs
7090 // that the generic analysis code won't necessarily catch.
7091 bool ARM64TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
7092 if (N->getNumValues() != 1)
7094 if (!N->hasNUsesOfValue(1, 0))
7097 SDValue TCChain = Chain;
7098 SDNode *Copy = *N->use_begin();
7099 if (Copy->getOpcode() == ISD::CopyToReg) {
7100 // If the copy has a glue operand, we conservatively assume it isn't safe to
7101 // perform a tail call.
7102 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
7105 TCChain = Copy->getOperand(0);
7106 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
7109 bool HasRet = false;
7110 for (SDNode *Node : Copy->uses()) {
7111 if (Node->getOpcode() != ARM64ISD::RET_FLAG)
7123 // Return whether the an instruction can potentially be optimized to a tail
7124 // call. This will cause the optimizers to attempt to move, or duplicate,
7125 // return instructions to help enable tail call optimizations for this
7127 bool ARM64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
7128 if (!EnableARM64TailCalls)
7131 if (!CI->isTailCall())
7137 bool ARM64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
7139 ISD::MemIndexedMode &AM,
7141 SelectionDAG &DAG) const {
7142 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
7145 Base = Op->getOperand(0);
7146 // All of the indexed addressing mode instructions take a signed
7147 // 9 bit immediate offset.
7148 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
7149 int64_t RHSC = (int64_t)RHS->getZExtValue();
7150 if (RHSC >= 256 || RHSC <= -256)
7152 IsInc = (Op->getOpcode() == ISD::ADD);
7153 Offset = Op->getOperand(1);
7159 bool ARM64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
7161 ISD::MemIndexedMode &AM,
7162 SelectionDAG &DAG) const {
7165 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7166 VT = LD->getMemoryVT();
7167 Ptr = LD->getBasePtr();
7168 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7169 VT = ST->getMemoryVT();
7170 Ptr = ST->getBasePtr();
7175 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
7177 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
7181 bool ARM64TargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
7184 ISD::MemIndexedMode &AM,
7185 SelectionDAG &DAG) const {
7188 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7189 VT = LD->getMemoryVT();
7190 Ptr = LD->getBasePtr();
7191 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7192 VT = ST->getMemoryVT();
7193 Ptr = ST->getBasePtr();
7198 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
7200 // Post-indexing updates the base, so it's not a valid transform
7201 // if that's not the same as the load's pointer.
7204 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
7208 void ARM64TargetLowering::ReplaceNodeResults(SDNode *N,
7209 SmallVectorImpl<SDValue> &Results,
7210 SelectionDAG &DAG) const {
7211 switch (N->getOpcode()) {
7213 llvm_unreachable("Don't know how to custom expand this");
7214 case ISD::FP_TO_UINT:
7215 case ISD::FP_TO_SINT:
7216 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
7217 // Let normal code take care of it by not adding anything to Results.
7222 bool ARM64TargetLowering::shouldExpandAtomicInIR(Instruction *Inst) const {
7223 // Loads and stores less than 128-bits are already atomic; ones above that
7224 // are doomed anyway, so defer to the default libcall and blame the OS when
7226 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
7227 return SI->getValueOperand()->getType()->getPrimitiveSizeInBits() == 128;
7228 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
7229 return LI->getType()->getPrimitiveSizeInBits() == 128;
7231 // For the real atomic operations, we have ldxr/stxr up to 128 bits.
7232 return Inst->getType()->getPrimitiveSizeInBits() <= 128;
7235 Value *ARM64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
7236 AtomicOrdering Ord) const {
7237 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
7238 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
7240 Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent;
7242 // Since i128 isn't legal and intrinsics don't get type-lowered, the ldrexd
7243 // intrinsic must return {i64, i64} and we have to recombine them into a
7244 // single i128 here.
7245 if (ValTy->getPrimitiveSizeInBits() == 128) {
7247 IsAcquire ? Intrinsic::arm64_ldaxp : Intrinsic::arm64_ldxp;
7248 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int);
7250 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
7251 Value *LoHi = Builder.CreateCall(Ldxr, Addr, "lohi");
7253 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
7254 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
7255 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
7256 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
7257 return Builder.CreateOr(
7258 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64");
7261 Type *Tys[] = { Addr->getType() };
7263 IsAcquire ? Intrinsic::arm64_ldaxr : Intrinsic::arm64_ldxr;
7264 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int, Tys);
7266 return Builder.CreateTruncOrBitCast(
7267 Builder.CreateCall(Ldxr, Addr),
7268 cast<PointerType>(Addr->getType())->getElementType());
7271 Value *ARM64TargetLowering::emitStoreConditional(IRBuilder<> &Builder,
7272 Value *Val, Value *Addr,
7273 AtomicOrdering Ord) const {
7274 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
7276 Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent;
7278 // Since the intrinsics must have legal type, the i128 intrinsics take two
7279 // parameters: "i64, i64". We must marshal Val into the appropriate form
7281 if (Val->getType()->getPrimitiveSizeInBits() == 128) {
7283 IsRelease ? Intrinsic::arm64_stlxp : Intrinsic::arm64_stxp;
7284 Function *Stxr = Intrinsic::getDeclaration(M, Int);
7285 Type *Int64Ty = Type::getInt64Ty(M->getContext());
7287 Value *Lo = Builder.CreateTrunc(Val, Int64Ty, "lo");
7288 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 64), Int64Ty, "hi");
7289 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
7290 return Builder.CreateCall3(Stxr, Lo, Hi, Addr);
7294 IsRelease ? Intrinsic::arm64_stlxr : Intrinsic::arm64_stxr;
7295 Type *Tys[] = { Addr->getType() };
7296 Function *Stxr = Intrinsic::getDeclaration(M, Int, Tys);
7298 return Builder.CreateCall2(
7299 Stxr, Builder.CreateZExtOrBitCast(
7300 Val, Stxr->getFunctionType()->getParamType(0)),